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>
58 #include <trace/events/block.h>
59 #include <linux/list_sort.h>
64 #include "md-bitmap.h"
65 #include "raid5-log.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)
104 spin_lock_irq(conf->hash_locks);
105 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
106 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
107 spin_lock(&conf->device_lock);
110 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
113 spin_unlock(&conf->device_lock);
114 for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
115 spin_unlock(conf->hash_locks + i);
116 spin_unlock_irq(conf->hash_locks);
119 /* Find first data disk in a raid6 stripe */
120 static inline int raid6_d0(struct stripe_head *sh)
123 /* ddf always start from first device */
125 /* md starts just after Q block */
126 if (sh->qd_idx == sh->disks - 1)
129 return sh->qd_idx + 1;
131 static inline int raid6_next_disk(int disk, int raid_disks)
134 return (disk < raid_disks) ? disk : 0;
137 /* When walking through the disks in a raid5, starting at raid6_d0,
138 * We need to map each disk to a 'slot', where the data disks are slot
139 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
140 * is raid_disks-1. This help does that mapping.
142 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
143 int *count, int syndrome_disks)
149 if (idx == sh->pd_idx)
150 return syndrome_disks;
151 if (idx == sh->qd_idx)
152 return syndrome_disks + 1;
158 static void print_raid5_conf (struct r5conf *conf);
160 static int stripe_operations_active(struct stripe_head *sh)
162 return sh->check_state || sh->reconstruct_state ||
163 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
164 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
167 static bool stripe_is_lowprio(struct stripe_head *sh)
169 return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
170 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
171 !test_bit(STRIPE_R5C_CACHING, &sh->state);
174 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
176 struct r5conf *conf = sh->raid_conf;
177 struct r5worker_group *group;
179 int i, cpu = sh->cpu;
181 if (!cpu_online(cpu)) {
182 cpu = cpumask_any(cpu_online_mask);
186 if (list_empty(&sh->lru)) {
187 struct r5worker_group *group;
188 group = conf->worker_groups + cpu_to_group(cpu);
189 if (stripe_is_lowprio(sh))
190 list_add_tail(&sh->lru, &group->loprio_list);
192 list_add_tail(&sh->lru, &group->handle_list);
193 group->stripes_cnt++;
197 if (conf->worker_cnt_per_group == 0) {
198 md_wakeup_thread(conf->mddev->thread);
202 group = conf->worker_groups + cpu_to_group(sh->cpu);
204 group->workers[0].working = true;
205 /* at least one worker should run to avoid race */
206 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
208 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
209 /* wakeup more workers */
210 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
211 if (group->workers[i].working == false) {
212 group->workers[i].working = true;
213 queue_work_on(sh->cpu, raid5_wq,
214 &group->workers[i].work);
220 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
221 struct list_head *temp_inactive_list)
224 int injournal = 0; /* number of date pages with R5_InJournal */
226 BUG_ON(!list_empty(&sh->lru));
227 BUG_ON(atomic_read(&conf->active_stripes)==0);
229 if (r5c_is_writeback(conf->log))
230 for (i = sh->disks; i--; )
231 if (test_bit(R5_InJournal, &sh->dev[i].flags))
234 * In the following cases, the stripe cannot be released to cached
235 * lists. Therefore, we make the stripe write out and set
237 * 1. when quiesce in r5c write back;
238 * 2. when resync is requested fot the stripe.
240 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
241 (conf->quiesce && r5c_is_writeback(conf->log) &&
242 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
243 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
244 r5c_make_stripe_write_out(sh);
245 set_bit(STRIPE_HANDLE, &sh->state);
248 if (test_bit(STRIPE_HANDLE, &sh->state)) {
249 if (test_bit(STRIPE_DELAYED, &sh->state) &&
250 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
251 list_add_tail(&sh->lru, &conf->delayed_list);
252 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
253 sh->bm_seq - conf->seq_write > 0)
254 list_add_tail(&sh->lru, &conf->bitmap_list);
256 clear_bit(STRIPE_DELAYED, &sh->state);
257 clear_bit(STRIPE_BIT_DELAY, &sh->state);
258 if (conf->worker_cnt_per_group == 0) {
259 if (stripe_is_lowprio(sh))
260 list_add_tail(&sh->lru,
263 list_add_tail(&sh->lru,
266 raid5_wakeup_stripe_thread(sh);
270 md_wakeup_thread(conf->mddev->thread);
272 BUG_ON(stripe_operations_active(sh));
273 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
274 if (atomic_dec_return(&conf->preread_active_stripes)
276 md_wakeup_thread(conf->mddev->thread);
277 atomic_dec(&conf->active_stripes);
278 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
279 if (!r5c_is_writeback(conf->log))
280 list_add_tail(&sh->lru, temp_inactive_list);
282 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
284 list_add_tail(&sh->lru, temp_inactive_list);
285 else if (injournal == conf->raid_disks - conf->max_degraded) {
287 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
288 atomic_inc(&conf->r5c_cached_full_stripes);
289 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
290 atomic_dec(&conf->r5c_cached_partial_stripes);
291 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
292 r5c_check_cached_full_stripe(conf);
295 * STRIPE_R5C_PARTIAL_STRIPE is set in
296 * r5c_try_caching_write(). No need to
299 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
305 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
306 struct list_head *temp_inactive_list)
308 if (atomic_dec_and_test(&sh->count))
309 do_release_stripe(conf, sh, temp_inactive_list);
313 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
315 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
316 * given time. Adding stripes only takes device lock, while deleting stripes
317 * only takes hash lock.
319 static void release_inactive_stripe_list(struct r5conf *conf,
320 struct list_head *temp_inactive_list,
324 bool do_wakeup = false;
327 if (hash == NR_STRIPE_HASH_LOCKS) {
328 size = NR_STRIPE_HASH_LOCKS;
329 hash = NR_STRIPE_HASH_LOCKS - 1;
333 struct list_head *list = &temp_inactive_list[size - 1];
336 * We don't hold any lock here yet, raid5_get_active_stripe() might
337 * remove stripes from the list
339 if (!list_empty_careful(list)) {
340 spin_lock_irqsave(conf->hash_locks + hash, flags);
341 if (list_empty(conf->inactive_list + hash) &&
343 atomic_dec(&conf->empty_inactive_list_nr);
344 list_splice_tail_init(list, conf->inactive_list + hash);
346 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
353 wake_up(&conf->wait_for_stripe);
354 if (atomic_read(&conf->active_stripes) == 0)
355 wake_up(&conf->wait_for_quiescent);
356 if (conf->retry_read_aligned)
357 md_wakeup_thread(conf->mddev->thread);
361 /* should hold conf->device_lock already */
362 static int release_stripe_list(struct r5conf *conf,
363 struct list_head *temp_inactive_list)
365 struct stripe_head *sh, *t;
367 struct llist_node *head;
369 head = llist_del_all(&conf->released_stripes);
370 head = llist_reverse_order(head);
371 llist_for_each_entry_safe(sh, t, head, release_list) {
374 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
376 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
378 * Don't worry the bit is set here, because if the bit is set
379 * again, the count is always > 1. This is true for
380 * STRIPE_ON_UNPLUG_LIST bit too.
382 hash = sh->hash_lock_index;
383 __release_stripe(conf, sh, &temp_inactive_list[hash]);
390 void raid5_release_stripe(struct stripe_head *sh)
392 struct r5conf *conf = sh->raid_conf;
394 struct list_head list;
398 /* Avoid release_list until the last reference.
400 if (atomic_add_unless(&sh->count, -1, 1))
403 if (unlikely(!conf->mddev->thread) ||
404 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
406 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
408 md_wakeup_thread(conf->mddev->thread);
411 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
412 if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
413 INIT_LIST_HEAD(&list);
414 hash = sh->hash_lock_index;
415 do_release_stripe(conf, sh, &list);
416 spin_unlock_irqrestore(&conf->device_lock, flags);
417 release_inactive_stripe_list(conf, &list, hash);
421 static inline void remove_hash(struct stripe_head *sh)
423 pr_debug("remove_hash(), stripe %llu\n",
424 (unsigned long long)sh->sector);
426 hlist_del_init(&sh->hash);
429 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
431 struct hlist_head *hp = stripe_hash(conf, sh->sector);
433 pr_debug("insert_hash(), stripe %llu\n",
434 (unsigned long long)sh->sector);
436 hlist_add_head(&sh->hash, hp);
439 /* find an idle stripe, make sure it is unhashed, and return it. */
440 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
442 struct stripe_head *sh = NULL;
443 struct list_head *first;
445 if (list_empty(conf->inactive_list + hash))
447 first = (conf->inactive_list + hash)->next;
448 sh = list_entry(first, struct stripe_head, lru);
449 list_del_init(first);
451 atomic_inc(&conf->active_stripes);
452 BUG_ON(hash != sh->hash_lock_index);
453 if (list_empty(conf->inactive_list + hash))
454 atomic_inc(&conf->empty_inactive_list_nr);
459 static void shrink_buffers(struct stripe_head *sh)
463 int num = sh->raid_conf->pool_size;
465 for (i = 0; i < num ; i++) {
466 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
470 sh->dev[i].page = NULL;
475 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
478 int num = sh->raid_conf->pool_size;
480 for (i = 0; i < num; i++) {
483 if (!(page = alloc_page(gfp))) {
486 sh->dev[i].page = page;
487 sh->dev[i].orig_page = page;
493 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
494 struct stripe_head *sh);
496 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
498 struct r5conf *conf = sh->raid_conf;
501 BUG_ON(atomic_read(&sh->count) != 0);
502 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
503 BUG_ON(stripe_operations_active(sh));
504 BUG_ON(sh->batch_head);
506 pr_debug("init_stripe called, stripe %llu\n",
507 (unsigned long long)sector);
509 seq = read_seqcount_begin(&conf->gen_lock);
510 sh->generation = conf->generation - previous;
511 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
513 stripe_set_idx(sector, conf, previous, sh);
516 for (i = sh->disks; i--; ) {
517 struct r5dev *dev = &sh->dev[i];
519 if (dev->toread || dev->read || dev->towrite || dev->written ||
520 test_bit(R5_LOCKED, &dev->flags)) {
521 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
522 (unsigned long long)sh->sector, i, dev->toread,
523 dev->read, dev->towrite, dev->written,
524 test_bit(R5_LOCKED, &dev->flags));
528 dev->sector = raid5_compute_blocknr(sh, i, previous);
530 if (read_seqcount_retry(&conf->gen_lock, seq))
532 sh->overwrite_disks = 0;
533 insert_hash(conf, sh);
534 sh->cpu = smp_processor_id();
535 set_bit(STRIPE_BATCH_READY, &sh->state);
538 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
541 struct stripe_head *sh;
543 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
544 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
545 if (sh->sector == sector && sh->generation == generation)
547 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
552 * Need to check if array has failed when deciding whether to:
554 * - remove non-faulty devices
557 * This determination is simple when no reshape is happening.
558 * However if there is a reshape, we need to carefully check
559 * both the before and after sections.
560 * This is because some failed devices may only affect one
561 * of the two sections, and some non-in_sync devices may
562 * be insync in the section most affected by failed devices.
564 int raid5_calc_degraded(struct r5conf *conf)
566 int degraded, degraded2;
571 for (i = 0; i < conf->previous_raid_disks; i++) {
572 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
573 if (rdev && test_bit(Faulty, &rdev->flags))
574 rdev = rcu_dereference(conf->disks[i].replacement);
575 if (!rdev || test_bit(Faulty, &rdev->flags))
577 else if (test_bit(In_sync, &rdev->flags))
580 /* not in-sync or faulty.
581 * If the reshape increases the number of devices,
582 * this is being recovered by the reshape, so
583 * this 'previous' section is not in_sync.
584 * If the number of devices is being reduced however,
585 * the device can only be part of the array if
586 * we are reverting a reshape, so this section will
589 if (conf->raid_disks >= conf->previous_raid_disks)
593 if (conf->raid_disks == conf->previous_raid_disks)
597 for (i = 0; i < conf->raid_disks; i++) {
598 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
599 if (rdev && test_bit(Faulty, &rdev->flags))
600 rdev = rcu_dereference(conf->disks[i].replacement);
601 if (!rdev || test_bit(Faulty, &rdev->flags))
603 else if (test_bit(In_sync, &rdev->flags))
606 /* not in-sync or faulty.
607 * If reshape increases the number of devices, this
608 * section has already been recovered, else it
609 * almost certainly hasn't.
611 if (conf->raid_disks <= conf->previous_raid_disks)
615 if (degraded2 > degraded)
620 static int has_failed(struct r5conf *conf)
624 if (conf->mddev->reshape_position == MaxSector)
625 return conf->mddev->degraded > conf->max_degraded;
627 degraded = raid5_calc_degraded(conf);
628 if (degraded > conf->max_degraded)
634 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
635 int previous, int noblock, int noquiesce)
637 struct stripe_head *sh;
638 int hash = stripe_hash_locks_hash(sector);
639 int inc_empty_inactive_list_flag;
641 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
643 spin_lock_irq(conf->hash_locks + hash);
646 wait_event_lock_irq(conf->wait_for_quiescent,
647 conf->quiesce == 0 || noquiesce,
648 *(conf->hash_locks + hash));
649 sh = __find_stripe(conf, sector, conf->generation - previous);
651 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
652 sh = get_free_stripe(conf, hash);
653 if (!sh && !test_bit(R5_DID_ALLOC,
655 set_bit(R5_ALLOC_MORE,
658 if (noblock && sh == NULL)
661 r5c_check_stripe_cache_usage(conf);
663 set_bit(R5_INACTIVE_BLOCKED,
665 r5l_wake_reclaim(conf->log, 0);
667 conf->wait_for_stripe,
668 !list_empty(conf->inactive_list + hash) &&
669 (atomic_read(&conf->active_stripes)
670 < (conf->max_nr_stripes * 3 / 4)
671 || !test_bit(R5_INACTIVE_BLOCKED,
672 &conf->cache_state)),
673 *(conf->hash_locks + hash));
674 clear_bit(R5_INACTIVE_BLOCKED,
677 init_stripe(sh, sector, previous);
678 atomic_inc(&sh->count);
680 } else if (!atomic_inc_not_zero(&sh->count)) {
681 spin_lock(&conf->device_lock);
682 if (!atomic_read(&sh->count)) {
683 if (!test_bit(STRIPE_HANDLE, &sh->state))
684 atomic_inc(&conf->active_stripes);
685 BUG_ON(list_empty(&sh->lru) &&
686 !test_bit(STRIPE_EXPANDING, &sh->state));
687 inc_empty_inactive_list_flag = 0;
688 if (!list_empty(conf->inactive_list + hash))
689 inc_empty_inactive_list_flag = 1;
690 list_del_init(&sh->lru);
691 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
692 atomic_inc(&conf->empty_inactive_list_nr);
694 sh->group->stripes_cnt--;
698 atomic_inc(&sh->count);
699 spin_unlock(&conf->device_lock);
701 } while (sh == NULL);
703 spin_unlock_irq(conf->hash_locks + hash);
707 static bool is_full_stripe_write(struct stripe_head *sh)
709 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
710 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
713 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
714 __acquires(&sh1->stripe_lock)
715 __acquires(&sh2->stripe_lock)
718 spin_lock_irq(&sh2->stripe_lock);
719 spin_lock_nested(&sh1->stripe_lock, 1);
721 spin_lock_irq(&sh1->stripe_lock);
722 spin_lock_nested(&sh2->stripe_lock, 1);
726 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
727 __releases(&sh1->stripe_lock)
728 __releases(&sh2->stripe_lock)
730 spin_unlock(&sh1->stripe_lock);
731 spin_unlock_irq(&sh2->stripe_lock);
734 /* Only freshly new full stripe normal write stripe can be added to a batch list */
735 static bool stripe_can_batch(struct stripe_head *sh)
737 struct r5conf *conf = sh->raid_conf;
739 if (raid5_has_log(conf) || raid5_has_ppl(conf))
741 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
742 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
743 is_full_stripe_write(sh);
746 /* we only do back search */
747 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
749 struct stripe_head *head;
750 sector_t head_sector, tmp_sec;
753 int inc_empty_inactive_list_flag;
755 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
756 tmp_sec = sh->sector;
757 if (!sector_div(tmp_sec, conf->chunk_sectors))
759 head_sector = sh->sector - STRIPE_SECTORS;
761 hash = stripe_hash_locks_hash(head_sector);
762 spin_lock_irq(conf->hash_locks + hash);
763 head = __find_stripe(conf, head_sector, conf->generation);
764 if (head && !atomic_inc_not_zero(&head->count)) {
765 spin_lock(&conf->device_lock);
766 if (!atomic_read(&head->count)) {
767 if (!test_bit(STRIPE_HANDLE, &head->state))
768 atomic_inc(&conf->active_stripes);
769 BUG_ON(list_empty(&head->lru) &&
770 !test_bit(STRIPE_EXPANDING, &head->state));
771 inc_empty_inactive_list_flag = 0;
772 if (!list_empty(conf->inactive_list + hash))
773 inc_empty_inactive_list_flag = 1;
774 list_del_init(&head->lru);
775 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
776 atomic_inc(&conf->empty_inactive_list_nr);
778 head->group->stripes_cnt--;
782 atomic_inc(&head->count);
783 spin_unlock(&conf->device_lock);
785 spin_unlock_irq(conf->hash_locks + hash);
789 if (!stripe_can_batch(head))
792 lock_two_stripes(head, sh);
793 /* clear_batch_ready clear the flag */
794 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
801 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
803 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
804 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
807 if (head->batch_head) {
808 spin_lock(&head->batch_head->batch_lock);
809 /* This batch list is already running */
810 if (!stripe_can_batch(head)) {
811 spin_unlock(&head->batch_head->batch_lock);
815 * We must assign batch_head of this stripe within the
816 * batch_lock, otherwise clear_batch_ready of batch head
817 * stripe could clear BATCH_READY bit of this stripe and
818 * this stripe->batch_head doesn't get assigned, which
819 * could confuse clear_batch_ready for this stripe
821 sh->batch_head = head->batch_head;
824 * at this point, head's BATCH_READY could be cleared, but we
825 * can still add the stripe to batch list
827 list_add(&sh->batch_list, &head->batch_list);
828 spin_unlock(&head->batch_head->batch_lock);
830 head->batch_head = head;
831 sh->batch_head = head->batch_head;
832 spin_lock(&head->batch_lock);
833 list_add_tail(&sh->batch_list, &head->batch_list);
834 spin_unlock(&head->batch_lock);
837 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
838 if (atomic_dec_return(&conf->preread_active_stripes)
840 md_wakeup_thread(conf->mddev->thread);
842 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
843 int seq = sh->bm_seq;
844 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
845 sh->batch_head->bm_seq > seq)
846 seq = sh->batch_head->bm_seq;
847 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
848 sh->batch_head->bm_seq = seq;
851 atomic_inc(&sh->count);
853 unlock_two_stripes(head, sh);
855 raid5_release_stripe(head);
858 /* Determine if 'data_offset' or 'new_data_offset' should be used
859 * in this stripe_head.
861 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
863 sector_t progress = conf->reshape_progress;
864 /* Need a memory barrier to make sure we see the value
865 * of conf->generation, or ->data_offset that was set before
866 * reshape_progress was updated.
869 if (progress == MaxSector)
871 if (sh->generation == conf->generation - 1)
873 /* We are in a reshape, and this is a new-generation stripe,
874 * so use new_data_offset.
879 static void dispatch_bio_list(struct bio_list *tmp)
883 while ((bio = bio_list_pop(tmp)))
884 generic_make_request(bio);
887 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
889 const struct r5pending_data *da = list_entry(a,
890 struct r5pending_data, sibling);
891 const struct r5pending_data *db = list_entry(b,
892 struct r5pending_data, sibling);
893 if (da->sector > db->sector)
895 if (da->sector < db->sector)
900 static void dispatch_defer_bios(struct r5conf *conf, int target,
901 struct bio_list *list)
903 struct r5pending_data *data;
904 struct list_head *first, *next = NULL;
907 if (conf->pending_data_cnt == 0)
910 list_sort(NULL, &conf->pending_list, cmp_stripe);
912 first = conf->pending_list.next;
914 /* temporarily move the head */
915 if (conf->next_pending_data)
916 list_move_tail(&conf->pending_list,
917 &conf->next_pending_data->sibling);
919 while (!list_empty(&conf->pending_list)) {
920 data = list_first_entry(&conf->pending_list,
921 struct r5pending_data, sibling);
922 if (&data->sibling == first)
923 first = data->sibling.next;
924 next = data->sibling.next;
926 bio_list_merge(list, &data->bios);
927 list_move(&data->sibling, &conf->free_list);
932 conf->pending_data_cnt -= cnt;
933 BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
935 if (next != &conf->pending_list)
936 conf->next_pending_data = list_entry(next,
937 struct r5pending_data, sibling);
939 conf->next_pending_data = NULL;
940 /* list isn't empty */
941 if (first != &conf->pending_list)
942 list_move_tail(&conf->pending_list, first);
945 static void flush_deferred_bios(struct r5conf *conf)
947 struct bio_list tmp = BIO_EMPTY_LIST;
949 if (conf->pending_data_cnt == 0)
952 spin_lock(&conf->pending_bios_lock);
953 dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
954 BUG_ON(conf->pending_data_cnt != 0);
955 spin_unlock(&conf->pending_bios_lock);
957 dispatch_bio_list(&tmp);
960 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
961 struct bio_list *bios)
963 struct bio_list tmp = BIO_EMPTY_LIST;
964 struct r5pending_data *ent;
966 spin_lock(&conf->pending_bios_lock);
967 ent = list_first_entry(&conf->free_list, struct r5pending_data,
969 list_move_tail(&ent->sibling, &conf->pending_list);
970 ent->sector = sector;
971 bio_list_init(&ent->bios);
972 bio_list_merge(&ent->bios, bios);
973 conf->pending_data_cnt++;
974 if (conf->pending_data_cnt >= PENDING_IO_MAX)
975 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
977 spin_unlock(&conf->pending_bios_lock);
979 dispatch_bio_list(&tmp);
983 raid5_end_read_request(struct bio *bi);
985 raid5_end_write_request(struct bio *bi);
987 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
989 struct r5conf *conf = sh->raid_conf;
990 int i, disks = sh->disks;
991 struct stripe_head *head_sh = sh;
992 struct bio_list pending_bios = BIO_EMPTY_LIST;
997 if (log_stripe(sh, s) == 0)
1000 should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1002 for (i = disks; i--; ) {
1003 int op, op_flags = 0;
1004 int replace_only = 0;
1005 struct bio *bi, *rbi;
1006 struct md_rdev *rdev, *rrdev = NULL;
1009 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1011 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1013 if (test_bit(R5_Discard, &sh->dev[i].flags))
1014 op = REQ_OP_DISCARD;
1015 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1017 else if (test_and_clear_bit(R5_WantReplace,
1018 &sh->dev[i].flags)) {
1023 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1024 op_flags |= REQ_SYNC;
1027 bi = &sh->dev[i].req;
1028 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1031 rrdev = rcu_dereference(conf->disks[i].replacement);
1032 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1033 rdev = rcu_dereference(conf->disks[i].rdev);
1038 if (op_is_write(op)) {
1042 /* We raced and saw duplicates */
1045 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1050 if (rdev && test_bit(Faulty, &rdev->flags))
1053 atomic_inc(&rdev->nr_pending);
1054 if (rrdev && test_bit(Faulty, &rrdev->flags))
1057 atomic_inc(&rrdev->nr_pending);
1060 /* We have already checked bad blocks for reads. Now
1061 * need to check for writes. We never accept write errors
1062 * on the replacement, so we don't to check rrdev.
1064 while (op_is_write(op) && rdev &&
1065 test_bit(WriteErrorSeen, &rdev->flags)) {
1068 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1069 &first_bad, &bad_sectors);
1074 set_bit(BlockedBadBlocks, &rdev->flags);
1075 if (!conf->mddev->external &&
1076 conf->mddev->sb_flags) {
1077 /* It is very unlikely, but we might
1078 * still need to write out the
1079 * bad block log - better give it
1081 md_check_recovery(conf->mddev);
1084 * Because md_wait_for_blocked_rdev
1085 * will dec nr_pending, we must
1086 * increment it first.
1088 atomic_inc(&rdev->nr_pending);
1089 md_wait_for_blocked_rdev(rdev, conf->mddev);
1091 /* Acknowledged bad block - skip the write */
1092 rdev_dec_pending(rdev, conf->mddev);
1098 if (s->syncing || s->expanding || s->expanded
1100 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1102 set_bit(STRIPE_IO_STARTED, &sh->state);
1104 bio_set_dev(bi, rdev->bdev);
1105 bio_set_op_attrs(bi, op, op_flags);
1106 bi->bi_end_io = op_is_write(op)
1107 ? raid5_end_write_request
1108 : raid5_end_read_request;
1109 bi->bi_private = sh;
1111 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1112 __func__, (unsigned long long)sh->sector,
1114 atomic_inc(&sh->count);
1116 atomic_inc(&head_sh->count);
1117 if (use_new_offset(conf, sh))
1118 bi->bi_iter.bi_sector = (sh->sector
1119 + rdev->new_data_offset);
1121 bi->bi_iter.bi_sector = (sh->sector
1122 + rdev->data_offset);
1123 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1124 bi->bi_opf |= REQ_NOMERGE;
1126 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1127 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1129 if (!op_is_write(op) &&
1130 test_bit(R5_InJournal, &sh->dev[i].flags))
1132 * issuing read for a page in journal, this
1133 * must be preparing for prexor in rmw; read
1134 * the data into orig_page
1136 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1138 sh->dev[i].vec.bv_page = sh->dev[i].page;
1140 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1141 bi->bi_io_vec[0].bv_offset = 0;
1142 bi->bi_iter.bi_size = STRIPE_SIZE;
1143 bi->bi_write_hint = sh->dev[i].write_hint;
1145 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1147 * If this is discard request, set bi_vcnt 0. We don't
1148 * want to confuse SCSI because SCSI will replace payload
1150 if (op == REQ_OP_DISCARD)
1153 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1155 if (conf->mddev->gendisk)
1156 trace_block_bio_remap(bi->bi_disk->queue,
1157 bi, disk_devt(conf->mddev->gendisk),
1159 if (should_defer && op_is_write(op))
1160 bio_list_add(&pending_bios, bi);
1162 generic_make_request(bi);
1165 if (s->syncing || s->expanding || s->expanded
1167 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1169 set_bit(STRIPE_IO_STARTED, &sh->state);
1171 bio_set_dev(rbi, rrdev->bdev);
1172 bio_set_op_attrs(rbi, op, op_flags);
1173 BUG_ON(!op_is_write(op));
1174 rbi->bi_end_io = raid5_end_write_request;
1175 rbi->bi_private = sh;
1177 pr_debug("%s: for %llu schedule op %d on "
1178 "replacement disc %d\n",
1179 __func__, (unsigned long long)sh->sector,
1181 atomic_inc(&sh->count);
1183 atomic_inc(&head_sh->count);
1184 if (use_new_offset(conf, sh))
1185 rbi->bi_iter.bi_sector = (sh->sector
1186 + rrdev->new_data_offset);
1188 rbi->bi_iter.bi_sector = (sh->sector
1189 + rrdev->data_offset);
1190 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1191 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1192 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1194 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1195 rbi->bi_io_vec[0].bv_offset = 0;
1196 rbi->bi_iter.bi_size = STRIPE_SIZE;
1197 rbi->bi_write_hint = sh->dev[i].write_hint;
1198 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1200 * If this is discard request, set bi_vcnt 0. We don't
1201 * want to confuse SCSI because SCSI will replace payload
1203 if (op == REQ_OP_DISCARD)
1205 if (conf->mddev->gendisk)
1206 trace_block_bio_remap(rbi->bi_disk->queue,
1207 rbi, disk_devt(conf->mddev->gendisk),
1209 if (should_defer && op_is_write(op))
1210 bio_list_add(&pending_bios, rbi);
1212 generic_make_request(rbi);
1214 if (!rdev && !rrdev) {
1215 if (op_is_write(op))
1216 set_bit(STRIPE_DEGRADED, &sh->state);
1217 pr_debug("skip op %d on disc %d for sector %llu\n",
1218 bi->bi_opf, i, (unsigned long long)sh->sector);
1219 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1220 set_bit(STRIPE_HANDLE, &sh->state);
1223 if (!head_sh->batch_head)
1225 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1231 if (should_defer && !bio_list_empty(&pending_bios))
1232 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1235 static struct dma_async_tx_descriptor *
1236 async_copy_data(int frombio, struct bio *bio, struct page **page,
1237 sector_t sector, struct dma_async_tx_descriptor *tx,
1238 struct stripe_head *sh, int no_skipcopy)
1241 struct bvec_iter iter;
1242 struct page *bio_page;
1244 struct async_submit_ctl submit;
1245 enum async_tx_flags flags = 0;
1247 if (bio->bi_iter.bi_sector >= sector)
1248 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1250 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1253 flags |= ASYNC_TX_FENCE;
1254 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1256 bio_for_each_segment(bvl, bio, iter) {
1257 int len = bvl.bv_len;
1261 if (page_offset < 0) {
1262 b_offset = -page_offset;
1263 page_offset += b_offset;
1267 if (len > 0 && page_offset + len > STRIPE_SIZE)
1268 clen = STRIPE_SIZE - page_offset;
1273 b_offset += bvl.bv_offset;
1274 bio_page = bvl.bv_page;
1276 if (sh->raid_conf->skip_copy &&
1277 b_offset == 0 && page_offset == 0 &&
1278 clen == STRIPE_SIZE &&
1282 tx = async_memcpy(*page, bio_page, page_offset,
1283 b_offset, clen, &submit);
1285 tx = async_memcpy(bio_page, *page, b_offset,
1286 page_offset, clen, &submit);
1288 /* chain the operations */
1289 submit.depend_tx = tx;
1291 if (clen < len) /* hit end of page */
1299 static void ops_complete_biofill(void *stripe_head_ref)
1301 struct stripe_head *sh = stripe_head_ref;
1304 pr_debug("%s: stripe %llu\n", __func__,
1305 (unsigned long long)sh->sector);
1307 /* clear completed biofills */
1308 for (i = sh->disks; i--; ) {
1309 struct r5dev *dev = &sh->dev[i];
1311 /* acknowledge completion of a biofill operation */
1312 /* and check if we need to reply to a read request,
1313 * new R5_Wantfill requests are held off until
1314 * !STRIPE_BIOFILL_RUN
1316 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1317 struct bio *rbi, *rbi2;
1322 while (rbi && rbi->bi_iter.bi_sector <
1323 dev->sector + STRIPE_SECTORS) {
1324 rbi2 = r5_next_bio(rbi, dev->sector);
1330 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1332 set_bit(STRIPE_HANDLE, &sh->state);
1333 raid5_release_stripe(sh);
1336 static void ops_run_biofill(struct stripe_head *sh)
1338 struct dma_async_tx_descriptor *tx = NULL;
1339 struct async_submit_ctl submit;
1342 BUG_ON(sh->batch_head);
1343 pr_debug("%s: stripe %llu\n", __func__,
1344 (unsigned long long)sh->sector);
1346 for (i = sh->disks; i--; ) {
1347 struct r5dev *dev = &sh->dev[i];
1348 if (test_bit(R5_Wantfill, &dev->flags)) {
1350 spin_lock_irq(&sh->stripe_lock);
1351 dev->read = rbi = dev->toread;
1353 spin_unlock_irq(&sh->stripe_lock);
1354 while (rbi && rbi->bi_iter.bi_sector <
1355 dev->sector + STRIPE_SECTORS) {
1356 tx = async_copy_data(0, rbi, &dev->page,
1357 dev->sector, tx, sh, 0);
1358 rbi = r5_next_bio(rbi, dev->sector);
1363 atomic_inc(&sh->count);
1364 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1365 async_trigger_callback(&submit);
1368 static void mark_target_uptodate(struct stripe_head *sh, int target)
1375 tgt = &sh->dev[target];
1376 set_bit(R5_UPTODATE, &tgt->flags);
1377 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1378 clear_bit(R5_Wantcompute, &tgt->flags);
1381 static void ops_complete_compute(void *stripe_head_ref)
1383 struct stripe_head *sh = stripe_head_ref;
1385 pr_debug("%s: stripe %llu\n", __func__,
1386 (unsigned long long)sh->sector);
1388 /* mark the computed target(s) as uptodate */
1389 mark_target_uptodate(sh, sh->ops.target);
1390 mark_target_uptodate(sh, sh->ops.target2);
1392 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1393 if (sh->check_state == check_state_compute_run)
1394 sh->check_state = check_state_compute_result;
1395 set_bit(STRIPE_HANDLE, &sh->state);
1396 raid5_release_stripe(sh);
1399 /* return a pointer to the address conversion region of the scribble buffer */
1400 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1402 return percpu->scribble + i * percpu->scribble_obj_size;
1405 /* return a pointer to the address conversion region of the scribble buffer */
1406 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1407 struct raid5_percpu *percpu, int i)
1409 return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1412 static struct dma_async_tx_descriptor *
1413 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1415 int disks = sh->disks;
1416 struct page **xor_srcs = to_addr_page(percpu, 0);
1417 int target = sh->ops.target;
1418 struct r5dev *tgt = &sh->dev[target];
1419 struct page *xor_dest = tgt->page;
1421 struct dma_async_tx_descriptor *tx;
1422 struct async_submit_ctl submit;
1425 BUG_ON(sh->batch_head);
1427 pr_debug("%s: stripe %llu block: %d\n",
1428 __func__, (unsigned long long)sh->sector, target);
1429 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1431 for (i = disks; i--; )
1433 xor_srcs[count++] = sh->dev[i].page;
1435 atomic_inc(&sh->count);
1437 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1438 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1439 if (unlikely(count == 1))
1440 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1442 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1447 /* set_syndrome_sources - populate source buffers for gen_syndrome
1448 * @srcs - (struct page *) array of size sh->disks
1449 * @sh - stripe_head to parse
1451 * Populates srcs in proper layout order for the stripe and returns the
1452 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1453 * destination buffer is recorded in srcs[count] and the Q destination
1454 * is recorded in srcs[count+1]].
1456 static int set_syndrome_sources(struct page **srcs,
1457 struct stripe_head *sh,
1460 int disks = sh->disks;
1461 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1462 int d0_idx = raid6_d0(sh);
1466 for (i = 0; i < disks; i++)
1472 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1473 struct r5dev *dev = &sh->dev[i];
1475 if (i == sh->qd_idx || i == sh->pd_idx ||
1476 (srctype == SYNDROME_SRC_ALL) ||
1477 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1478 (test_bit(R5_Wantdrain, &dev->flags) ||
1479 test_bit(R5_InJournal, &dev->flags))) ||
1480 (srctype == SYNDROME_SRC_WRITTEN &&
1482 test_bit(R5_InJournal, &dev->flags)))) {
1483 if (test_bit(R5_InJournal, &dev->flags))
1484 srcs[slot] = sh->dev[i].orig_page;
1486 srcs[slot] = sh->dev[i].page;
1488 i = raid6_next_disk(i, disks);
1489 } while (i != d0_idx);
1491 return syndrome_disks;
1494 static struct dma_async_tx_descriptor *
1495 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1497 int disks = sh->disks;
1498 struct page **blocks = to_addr_page(percpu, 0);
1500 int qd_idx = sh->qd_idx;
1501 struct dma_async_tx_descriptor *tx;
1502 struct async_submit_ctl submit;
1508 BUG_ON(sh->batch_head);
1509 if (sh->ops.target < 0)
1510 target = sh->ops.target2;
1511 else if (sh->ops.target2 < 0)
1512 target = sh->ops.target;
1514 /* we should only have one valid target */
1517 pr_debug("%s: stripe %llu block: %d\n",
1518 __func__, (unsigned long long)sh->sector, target);
1520 tgt = &sh->dev[target];
1521 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1524 atomic_inc(&sh->count);
1526 if (target == qd_idx) {
1527 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1528 blocks[count] = NULL; /* regenerating p is not necessary */
1529 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1530 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1531 ops_complete_compute, sh,
1532 to_addr_conv(sh, percpu, 0));
1533 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1535 /* Compute any data- or p-drive using XOR */
1537 for (i = disks; i-- ; ) {
1538 if (i == target || i == qd_idx)
1540 blocks[count++] = sh->dev[i].page;
1543 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1544 NULL, ops_complete_compute, sh,
1545 to_addr_conv(sh, percpu, 0));
1546 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1552 static struct dma_async_tx_descriptor *
1553 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1555 int i, count, disks = sh->disks;
1556 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1557 int d0_idx = raid6_d0(sh);
1558 int faila = -1, failb = -1;
1559 int target = sh->ops.target;
1560 int target2 = sh->ops.target2;
1561 struct r5dev *tgt = &sh->dev[target];
1562 struct r5dev *tgt2 = &sh->dev[target2];
1563 struct dma_async_tx_descriptor *tx;
1564 struct page **blocks = to_addr_page(percpu, 0);
1565 struct async_submit_ctl submit;
1567 BUG_ON(sh->batch_head);
1568 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1569 __func__, (unsigned long long)sh->sector, target, target2);
1570 BUG_ON(target < 0 || target2 < 0);
1571 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1572 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1574 /* we need to open-code set_syndrome_sources to handle the
1575 * slot number conversion for 'faila' and 'failb'
1577 for (i = 0; i < disks ; i++)
1582 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1584 blocks[slot] = sh->dev[i].page;
1590 i = raid6_next_disk(i, disks);
1591 } while (i != d0_idx);
1593 BUG_ON(faila == failb);
1596 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1597 __func__, (unsigned long long)sh->sector, faila, failb);
1599 atomic_inc(&sh->count);
1601 if (failb == syndrome_disks+1) {
1602 /* Q disk is one of the missing disks */
1603 if (faila == syndrome_disks) {
1604 /* Missing P+Q, just recompute */
1605 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1606 ops_complete_compute, sh,
1607 to_addr_conv(sh, percpu, 0));
1608 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1609 STRIPE_SIZE, &submit);
1613 int qd_idx = sh->qd_idx;
1615 /* Missing D+Q: recompute D from P, then recompute Q */
1616 if (target == qd_idx)
1617 data_target = target2;
1619 data_target = target;
1622 for (i = disks; i-- ; ) {
1623 if (i == data_target || i == qd_idx)
1625 blocks[count++] = sh->dev[i].page;
1627 dest = sh->dev[data_target].page;
1628 init_async_submit(&submit,
1629 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1631 to_addr_conv(sh, percpu, 0));
1632 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1635 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1636 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1637 ops_complete_compute, sh,
1638 to_addr_conv(sh, percpu, 0));
1639 return async_gen_syndrome(blocks, 0, count+2,
1640 STRIPE_SIZE, &submit);
1643 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1644 ops_complete_compute, sh,
1645 to_addr_conv(sh, percpu, 0));
1646 if (failb == syndrome_disks) {
1647 /* We're missing D+P. */
1648 return async_raid6_datap_recov(syndrome_disks+2,
1652 /* We're missing D+D. */
1653 return async_raid6_2data_recov(syndrome_disks+2,
1654 STRIPE_SIZE, faila, failb,
1660 static void ops_complete_prexor(void *stripe_head_ref)
1662 struct stripe_head *sh = stripe_head_ref;
1664 pr_debug("%s: stripe %llu\n", __func__,
1665 (unsigned long long)sh->sector);
1667 if (r5c_is_writeback(sh->raid_conf->log))
1669 * raid5-cache write back uses orig_page during prexor.
1670 * After prexor, it is time to free orig_page
1672 r5c_release_extra_page(sh);
1675 static struct dma_async_tx_descriptor *
1676 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1677 struct dma_async_tx_descriptor *tx)
1679 int disks = sh->disks;
1680 struct page **xor_srcs = to_addr_page(percpu, 0);
1681 int count = 0, pd_idx = sh->pd_idx, i;
1682 struct async_submit_ctl submit;
1684 /* existing parity data subtracted */
1685 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1687 BUG_ON(sh->batch_head);
1688 pr_debug("%s: stripe %llu\n", __func__,
1689 (unsigned long long)sh->sector);
1691 for (i = disks; i--; ) {
1692 struct r5dev *dev = &sh->dev[i];
1693 /* Only process blocks that are known to be uptodate */
1694 if (test_bit(R5_InJournal, &dev->flags))
1695 xor_srcs[count++] = dev->orig_page;
1696 else if (test_bit(R5_Wantdrain, &dev->flags))
1697 xor_srcs[count++] = dev->page;
1700 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1701 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1702 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1707 static struct dma_async_tx_descriptor *
1708 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1709 struct dma_async_tx_descriptor *tx)
1711 struct page **blocks = to_addr_page(percpu, 0);
1713 struct async_submit_ctl submit;
1715 pr_debug("%s: stripe %llu\n", __func__,
1716 (unsigned long long)sh->sector);
1718 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1720 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1721 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1722 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1727 static struct dma_async_tx_descriptor *
1728 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1730 struct r5conf *conf = sh->raid_conf;
1731 int disks = sh->disks;
1733 struct stripe_head *head_sh = sh;
1735 pr_debug("%s: stripe %llu\n", __func__,
1736 (unsigned long long)sh->sector);
1738 for (i = disks; i--; ) {
1743 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1749 * clear R5_InJournal, so when rewriting a page in
1750 * journal, it is not skipped by r5l_log_stripe()
1752 clear_bit(R5_InJournal, &dev->flags);
1753 spin_lock_irq(&sh->stripe_lock);
1754 chosen = dev->towrite;
1755 dev->towrite = NULL;
1756 sh->overwrite_disks = 0;
1757 BUG_ON(dev->written);
1758 wbi = dev->written = chosen;
1759 spin_unlock_irq(&sh->stripe_lock);
1760 WARN_ON(dev->page != dev->orig_page);
1762 while (wbi && wbi->bi_iter.bi_sector <
1763 dev->sector + STRIPE_SECTORS) {
1764 if (wbi->bi_opf & REQ_FUA)
1765 set_bit(R5_WantFUA, &dev->flags);
1766 if (wbi->bi_opf & REQ_SYNC)
1767 set_bit(R5_SyncIO, &dev->flags);
1768 if (bio_op(wbi) == REQ_OP_DISCARD)
1769 set_bit(R5_Discard, &dev->flags);
1771 tx = async_copy_data(1, wbi, &dev->page,
1772 dev->sector, tx, sh,
1773 r5c_is_writeback(conf->log));
1774 if (dev->page != dev->orig_page &&
1775 !r5c_is_writeback(conf->log)) {
1776 set_bit(R5_SkipCopy, &dev->flags);
1777 clear_bit(R5_UPTODATE, &dev->flags);
1778 clear_bit(R5_OVERWRITE, &dev->flags);
1781 wbi = r5_next_bio(wbi, dev->sector);
1784 if (head_sh->batch_head) {
1785 sh = list_first_entry(&sh->batch_list,
1798 static void ops_complete_reconstruct(void *stripe_head_ref)
1800 struct stripe_head *sh = stripe_head_ref;
1801 int disks = sh->disks;
1802 int pd_idx = sh->pd_idx;
1803 int qd_idx = sh->qd_idx;
1805 bool fua = false, sync = false, discard = false;
1807 pr_debug("%s: stripe %llu\n", __func__,
1808 (unsigned long long)sh->sector);
1810 for (i = disks; i--; ) {
1811 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1812 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1813 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1816 for (i = disks; i--; ) {
1817 struct r5dev *dev = &sh->dev[i];
1819 if (dev->written || i == pd_idx || i == qd_idx) {
1820 if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1821 set_bit(R5_UPTODATE, &dev->flags);
1822 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1823 set_bit(R5_Expanded, &dev->flags);
1826 set_bit(R5_WantFUA, &dev->flags);
1828 set_bit(R5_SyncIO, &dev->flags);
1832 if (sh->reconstruct_state == reconstruct_state_drain_run)
1833 sh->reconstruct_state = reconstruct_state_drain_result;
1834 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1835 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1837 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1838 sh->reconstruct_state = reconstruct_state_result;
1841 set_bit(STRIPE_HANDLE, &sh->state);
1842 raid5_release_stripe(sh);
1846 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1847 struct dma_async_tx_descriptor *tx)
1849 int disks = sh->disks;
1850 struct page **xor_srcs;
1851 struct async_submit_ctl submit;
1852 int count, pd_idx = sh->pd_idx, i;
1853 struct page *xor_dest;
1855 unsigned long flags;
1857 struct stripe_head *head_sh = sh;
1860 pr_debug("%s: stripe %llu\n", __func__,
1861 (unsigned long long)sh->sector);
1863 for (i = 0; i < sh->disks; i++) {
1866 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1869 if (i >= sh->disks) {
1870 atomic_inc(&sh->count);
1871 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1872 ops_complete_reconstruct(sh);
1877 xor_srcs = to_addr_page(percpu, j);
1878 /* check if prexor is active which means only process blocks
1879 * that are part of a read-modify-write (written)
1881 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1883 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1884 for (i = disks; i--; ) {
1885 struct r5dev *dev = &sh->dev[i];
1886 if (head_sh->dev[i].written ||
1887 test_bit(R5_InJournal, &head_sh->dev[i].flags))
1888 xor_srcs[count++] = dev->page;
1891 xor_dest = sh->dev[pd_idx].page;
1892 for (i = disks; i--; ) {
1893 struct r5dev *dev = &sh->dev[i];
1895 xor_srcs[count++] = dev->page;
1899 /* 1/ if we prexor'd then the dest is reused as a source
1900 * 2/ if we did not prexor then we are redoing the parity
1901 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1902 * for the synchronous xor case
1904 last_stripe = !head_sh->batch_head ||
1905 list_first_entry(&sh->batch_list,
1906 struct stripe_head, batch_list) == head_sh;
1908 flags = ASYNC_TX_ACK |
1909 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1911 atomic_inc(&head_sh->count);
1912 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1913 to_addr_conv(sh, percpu, j));
1915 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1916 init_async_submit(&submit, flags, tx, NULL, NULL,
1917 to_addr_conv(sh, percpu, j));
1920 if (unlikely(count == 1))
1921 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1923 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1926 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1933 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1934 struct dma_async_tx_descriptor *tx)
1936 struct async_submit_ctl submit;
1937 struct page **blocks;
1938 int count, i, j = 0;
1939 struct stripe_head *head_sh = sh;
1942 unsigned long txflags;
1944 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1946 for (i = 0; i < sh->disks; i++) {
1947 if (sh->pd_idx == i || sh->qd_idx == i)
1949 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1952 if (i >= sh->disks) {
1953 atomic_inc(&sh->count);
1954 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1955 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1956 ops_complete_reconstruct(sh);
1961 blocks = to_addr_page(percpu, j);
1963 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1964 synflags = SYNDROME_SRC_WRITTEN;
1965 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1967 synflags = SYNDROME_SRC_ALL;
1968 txflags = ASYNC_TX_ACK;
1971 count = set_syndrome_sources(blocks, sh, synflags);
1972 last_stripe = !head_sh->batch_head ||
1973 list_first_entry(&sh->batch_list,
1974 struct stripe_head, batch_list) == head_sh;
1977 atomic_inc(&head_sh->count);
1978 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1979 head_sh, to_addr_conv(sh, percpu, j));
1981 init_async_submit(&submit, 0, tx, NULL, NULL,
1982 to_addr_conv(sh, percpu, j));
1983 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1986 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1992 static void ops_complete_check(void *stripe_head_ref)
1994 struct stripe_head *sh = stripe_head_ref;
1996 pr_debug("%s: stripe %llu\n", __func__,
1997 (unsigned long long)sh->sector);
1999 sh->check_state = check_state_check_result;
2000 set_bit(STRIPE_HANDLE, &sh->state);
2001 raid5_release_stripe(sh);
2004 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2006 int disks = sh->disks;
2007 int pd_idx = sh->pd_idx;
2008 int qd_idx = sh->qd_idx;
2009 struct page *xor_dest;
2010 struct page **xor_srcs = to_addr_page(percpu, 0);
2011 struct dma_async_tx_descriptor *tx;
2012 struct async_submit_ctl submit;
2016 pr_debug("%s: stripe %llu\n", __func__,
2017 (unsigned long long)sh->sector);
2019 BUG_ON(sh->batch_head);
2021 xor_dest = sh->dev[pd_idx].page;
2022 xor_srcs[count++] = xor_dest;
2023 for (i = disks; i--; ) {
2024 if (i == pd_idx || i == qd_idx)
2026 xor_srcs[count++] = sh->dev[i].page;
2029 init_async_submit(&submit, 0, NULL, NULL, NULL,
2030 to_addr_conv(sh, percpu, 0));
2031 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2032 &sh->ops.zero_sum_result, &submit);
2034 atomic_inc(&sh->count);
2035 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2036 tx = async_trigger_callback(&submit);
2039 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2041 struct page **srcs = to_addr_page(percpu, 0);
2042 struct async_submit_ctl submit;
2045 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2046 (unsigned long long)sh->sector, checkp);
2048 BUG_ON(sh->batch_head);
2049 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2053 atomic_inc(&sh->count);
2054 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2055 sh, to_addr_conv(sh, percpu, 0));
2056 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2057 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2060 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2062 int overlap_clear = 0, i, disks = sh->disks;
2063 struct dma_async_tx_descriptor *tx = NULL;
2064 struct r5conf *conf = sh->raid_conf;
2065 int level = conf->level;
2066 struct raid5_percpu *percpu;
2070 percpu = per_cpu_ptr(conf->percpu, cpu);
2071 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2072 ops_run_biofill(sh);
2076 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2078 tx = ops_run_compute5(sh, percpu);
2080 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2081 tx = ops_run_compute6_1(sh, percpu);
2083 tx = ops_run_compute6_2(sh, percpu);
2085 /* terminate the chain if reconstruct is not set to be run */
2086 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2090 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2092 tx = ops_run_prexor5(sh, percpu, tx);
2094 tx = ops_run_prexor6(sh, percpu, tx);
2097 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2098 tx = ops_run_partial_parity(sh, percpu, tx);
2100 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2101 tx = ops_run_biodrain(sh, tx);
2105 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2107 ops_run_reconstruct5(sh, percpu, tx);
2109 ops_run_reconstruct6(sh, percpu, tx);
2112 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2113 if (sh->check_state == check_state_run)
2114 ops_run_check_p(sh, percpu);
2115 else if (sh->check_state == check_state_run_q)
2116 ops_run_check_pq(sh, percpu, 0);
2117 else if (sh->check_state == check_state_run_pq)
2118 ops_run_check_pq(sh, percpu, 1);
2123 if (overlap_clear && !sh->batch_head)
2124 for (i = disks; i--; ) {
2125 struct r5dev *dev = &sh->dev[i];
2126 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2127 wake_up(&sh->raid_conf->wait_for_overlap);
2132 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2135 __free_page(sh->ppl_page);
2136 kmem_cache_free(sc, sh);
2139 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2140 int disks, struct r5conf *conf)
2142 struct stripe_head *sh;
2145 sh = kmem_cache_zalloc(sc, gfp);
2147 spin_lock_init(&sh->stripe_lock);
2148 spin_lock_init(&sh->batch_lock);
2149 INIT_LIST_HEAD(&sh->batch_list);
2150 INIT_LIST_HEAD(&sh->lru);
2151 INIT_LIST_HEAD(&sh->r5c);
2152 INIT_LIST_HEAD(&sh->log_list);
2153 atomic_set(&sh->count, 1);
2154 sh->raid_conf = conf;
2155 sh->log_start = MaxSector;
2156 for (i = 0; i < disks; i++) {
2157 struct r5dev *dev = &sh->dev[i];
2159 bio_init(&dev->req, &dev->vec, 1);
2160 bio_init(&dev->rreq, &dev->rvec, 1);
2163 if (raid5_has_ppl(conf)) {
2164 sh->ppl_page = alloc_page(gfp);
2165 if (!sh->ppl_page) {
2166 free_stripe(sc, sh);
2173 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2175 struct stripe_head *sh;
2177 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2181 if (grow_buffers(sh, gfp)) {
2183 free_stripe(conf->slab_cache, sh);
2186 sh->hash_lock_index =
2187 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2188 /* we just created an active stripe so... */
2189 atomic_inc(&conf->active_stripes);
2191 raid5_release_stripe(sh);
2192 conf->max_nr_stripes++;
2196 static int grow_stripes(struct r5conf *conf, int num)
2198 struct kmem_cache *sc;
2199 size_t namelen = sizeof(conf->cache_name[0]);
2200 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2202 if (conf->mddev->gendisk)
2203 snprintf(conf->cache_name[0], namelen,
2204 "raid%d-%s", conf->level, mdname(conf->mddev));
2206 snprintf(conf->cache_name[0], namelen,
2207 "raid%d-%p", conf->level, conf->mddev);
2208 snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2210 conf->active_name = 0;
2211 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2212 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2216 conf->slab_cache = sc;
2217 conf->pool_size = devs;
2219 if (!grow_one_stripe(conf, GFP_KERNEL))
2226 * scribble_len - return the required size of the scribble region
2227 * @num - total number of disks in the array
2229 * The size must be enough to contain:
2230 * 1/ a struct page pointer for each device in the array +2
2231 * 2/ room to convert each entry in (1) to its corresponding dma
2232 * (dma_map_page()) or page (page_address()) address.
2234 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2235 * calculate over all devices (not just the data blocks), using zeros in place
2236 * of the P and Q blocks.
2238 static int scribble_alloc(struct raid5_percpu *percpu,
2239 int num, int cnt, gfp_t flags)
2242 sizeof(struct page *) * (num+2) +
2243 sizeof(addr_conv_t) * (num+2);
2246 scribble = kvmalloc_array(cnt, obj_size, flags);
2250 kvfree(percpu->scribble);
2252 percpu->scribble = scribble;
2253 percpu->scribble_obj_size = obj_size;
2257 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2263 * Never shrink. And mddev_suspend() could deadlock if this is called
2264 * from raid5d. In that case, scribble_disks and scribble_sectors
2265 * should equal to new_disks and new_sectors
2267 if (conf->scribble_disks >= new_disks &&
2268 conf->scribble_sectors >= new_sectors)
2270 mddev_suspend(conf->mddev);
2273 for_each_present_cpu(cpu) {
2274 struct raid5_percpu *percpu;
2276 percpu = per_cpu_ptr(conf->percpu, cpu);
2277 err = scribble_alloc(percpu, new_disks,
2278 new_sectors / STRIPE_SECTORS,
2285 mddev_resume(conf->mddev);
2287 conf->scribble_disks = new_disks;
2288 conf->scribble_sectors = new_sectors;
2293 static int resize_stripes(struct r5conf *conf, int newsize)
2295 /* Make all the stripes able to hold 'newsize' devices.
2296 * New slots in each stripe get 'page' set to a new page.
2298 * This happens in stages:
2299 * 1/ create a new kmem_cache and allocate the required number of
2301 * 2/ gather all the old stripe_heads and transfer the pages across
2302 * to the new stripe_heads. This will have the side effect of
2303 * freezing the array as once all stripe_heads have been collected,
2304 * no IO will be possible. Old stripe heads are freed once their
2305 * pages have been transferred over, and the old kmem_cache is
2306 * freed when all stripes are done.
2307 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2308 * we simple return a failure status - no need to clean anything up.
2309 * 4/ allocate new pages for the new slots in the new stripe_heads.
2310 * If this fails, we don't bother trying the shrink the
2311 * stripe_heads down again, we just leave them as they are.
2312 * As each stripe_head is processed the new one is released into
2315 * Once step2 is started, we cannot afford to wait for a write,
2316 * so we use GFP_NOIO allocations.
2318 struct stripe_head *osh, *nsh;
2319 LIST_HEAD(newstripes);
2320 struct disk_info *ndisks;
2322 struct kmem_cache *sc;
2326 md_allow_write(conf->mddev);
2329 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2330 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2335 /* Need to ensure auto-resizing doesn't interfere */
2336 mutex_lock(&conf->cache_size_mutex);
2338 for (i = conf->max_nr_stripes; i; i--) {
2339 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2343 list_add(&nsh->lru, &newstripes);
2346 /* didn't get enough, give up */
2347 while (!list_empty(&newstripes)) {
2348 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2349 list_del(&nsh->lru);
2350 free_stripe(sc, nsh);
2352 kmem_cache_destroy(sc);
2353 mutex_unlock(&conf->cache_size_mutex);
2356 /* Step 2 - Must use GFP_NOIO now.
2357 * OK, we have enough stripes, start collecting inactive
2358 * stripes and copying them over
2362 list_for_each_entry(nsh, &newstripes, lru) {
2363 lock_device_hash_lock(conf, hash);
2364 wait_event_cmd(conf->wait_for_stripe,
2365 !list_empty(conf->inactive_list + hash),
2366 unlock_device_hash_lock(conf, hash),
2367 lock_device_hash_lock(conf, hash));
2368 osh = get_free_stripe(conf, hash);
2369 unlock_device_hash_lock(conf, hash);
2371 for(i=0; i<conf->pool_size; i++) {
2372 nsh->dev[i].page = osh->dev[i].page;
2373 nsh->dev[i].orig_page = osh->dev[i].page;
2375 nsh->hash_lock_index = hash;
2376 free_stripe(conf->slab_cache, osh);
2378 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2379 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2384 kmem_cache_destroy(conf->slab_cache);
2387 * At this point, we are holding all the stripes so the array
2388 * is completely stalled, so now is a good time to resize
2389 * conf->disks and the scribble region
2391 ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2393 for (i = 0; i < conf->pool_size; i++)
2394 ndisks[i] = conf->disks[i];
2396 for (i = conf->pool_size; i < newsize; i++) {
2397 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2398 if (!ndisks[i].extra_page)
2403 for (i = conf->pool_size; i < newsize; i++)
2404 if (ndisks[i].extra_page)
2405 put_page(ndisks[i].extra_page);
2409 conf->disks = ndisks;
2414 mutex_unlock(&conf->cache_size_mutex);
2416 conf->slab_cache = sc;
2417 conf->active_name = 1-conf->active_name;
2419 /* Step 4, return new stripes to service */
2420 while(!list_empty(&newstripes)) {
2421 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2422 list_del_init(&nsh->lru);
2424 for (i=conf->raid_disks; i < newsize; i++)
2425 if (nsh->dev[i].page == NULL) {
2426 struct page *p = alloc_page(GFP_NOIO);
2427 nsh->dev[i].page = p;
2428 nsh->dev[i].orig_page = p;
2432 raid5_release_stripe(nsh);
2434 /* critical section pass, GFP_NOIO no longer needed */
2437 conf->pool_size = newsize;
2441 static int drop_one_stripe(struct r5conf *conf)
2443 struct stripe_head *sh;
2444 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2446 spin_lock_irq(conf->hash_locks + hash);
2447 sh = get_free_stripe(conf, hash);
2448 spin_unlock_irq(conf->hash_locks + hash);
2451 BUG_ON(atomic_read(&sh->count));
2453 free_stripe(conf->slab_cache, sh);
2454 atomic_dec(&conf->active_stripes);
2455 conf->max_nr_stripes--;
2459 static void shrink_stripes(struct r5conf *conf)
2461 while (conf->max_nr_stripes &&
2462 drop_one_stripe(conf))
2465 kmem_cache_destroy(conf->slab_cache);
2466 conf->slab_cache = NULL;
2469 static void raid5_end_read_request(struct bio * bi)
2471 struct stripe_head *sh = bi->bi_private;
2472 struct r5conf *conf = sh->raid_conf;
2473 int disks = sh->disks, i;
2474 char b[BDEVNAME_SIZE];
2475 struct md_rdev *rdev = NULL;
2478 for (i=0 ; i<disks; i++)
2479 if (bi == &sh->dev[i].req)
2482 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2483 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2490 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2491 /* If replacement finished while this request was outstanding,
2492 * 'replacement' might be NULL already.
2493 * In that case it moved down to 'rdev'.
2494 * rdev is not removed until all requests are finished.
2496 rdev = conf->disks[i].replacement;
2498 rdev = conf->disks[i].rdev;
2500 if (use_new_offset(conf, sh))
2501 s = sh->sector + rdev->new_data_offset;
2503 s = sh->sector + rdev->data_offset;
2504 if (!bi->bi_status) {
2505 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2506 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2507 /* Note that this cannot happen on a
2508 * replacement device. We just fail those on
2511 pr_info_ratelimited(
2512 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2513 mdname(conf->mddev), STRIPE_SECTORS,
2514 (unsigned long long)s,
2515 bdevname(rdev->bdev, b));
2516 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2517 clear_bit(R5_ReadError, &sh->dev[i].flags);
2518 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2519 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2520 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2522 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2524 * end read for a page in journal, this
2525 * must be preparing for prexor in rmw
2527 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2529 if (atomic_read(&rdev->read_errors))
2530 atomic_set(&rdev->read_errors, 0);
2532 const char *bdn = bdevname(rdev->bdev, b);
2536 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2537 atomic_inc(&rdev->read_errors);
2538 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2539 pr_warn_ratelimited(
2540 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2541 mdname(conf->mddev),
2542 (unsigned long long)s,
2544 else if (conf->mddev->degraded >= conf->max_degraded) {
2546 pr_warn_ratelimited(
2547 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2548 mdname(conf->mddev),
2549 (unsigned long long)s,
2551 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2554 pr_warn_ratelimited(
2555 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2556 mdname(conf->mddev),
2557 (unsigned long long)s,
2559 } else if (atomic_read(&rdev->read_errors)
2560 > conf->max_nr_stripes)
2561 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2562 mdname(conf->mddev), bdn);
2565 if (set_bad && test_bit(In_sync, &rdev->flags)
2566 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2569 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2570 set_bit(R5_ReadError, &sh->dev[i].flags);
2571 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2573 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2575 clear_bit(R5_ReadError, &sh->dev[i].flags);
2576 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2578 && test_bit(In_sync, &rdev->flags)
2579 && rdev_set_badblocks(
2580 rdev, sh->sector, STRIPE_SECTORS, 0)))
2581 md_error(conf->mddev, rdev);
2584 rdev_dec_pending(rdev, conf->mddev);
2586 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2587 set_bit(STRIPE_HANDLE, &sh->state);
2588 raid5_release_stripe(sh);
2591 static void raid5_end_write_request(struct bio *bi)
2593 struct stripe_head *sh = bi->bi_private;
2594 struct r5conf *conf = sh->raid_conf;
2595 int disks = sh->disks, i;
2596 struct md_rdev *uninitialized_var(rdev);
2599 int replacement = 0;
2601 for (i = 0 ; i < disks; i++) {
2602 if (bi == &sh->dev[i].req) {
2603 rdev = conf->disks[i].rdev;
2606 if (bi == &sh->dev[i].rreq) {
2607 rdev = conf->disks[i].replacement;
2611 /* rdev was removed and 'replacement'
2612 * replaced it. rdev is not removed
2613 * until all requests are finished.
2615 rdev = conf->disks[i].rdev;
2619 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2620 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2630 md_error(conf->mddev, rdev);
2631 else if (is_badblock(rdev, sh->sector,
2633 &first_bad, &bad_sectors))
2634 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2636 if (bi->bi_status) {
2637 set_bit(STRIPE_DEGRADED, &sh->state);
2638 set_bit(WriteErrorSeen, &rdev->flags);
2639 set_bit(R5_WriteError, &sh->dev[i].flags);
2640 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2641 set_bit(MD_RECOVERY_NEEDED,
2642 &rdev->mddev->recovery);
2643 } else if (is_badblock(rdev, sh->sector,
2645 &first_bad, &bad_sectors)) {
2646 set_bit(R5_MadeGood, &sh->dev[i].flags);
2647 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2648 /* That was a successful write so make
2649 * sure it looks like we already did
2652 set_bit(R5_ReWrite, &sh->dev[i].flags);
2655 rdev_dec_pending(rdev, conf->mddev);
2657 if (sh->batch_head && bi->bi_status && !replacement)
2658 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2661 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2662 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2663 set_bit(STRIPE_HANDLE, &sh->state);
2664 raid5_release_stripe(sh);
2666 if (sh->batch_head && sh != sh->batch_head)
2667 raid5_release_stripe(sh->batch_head);
2670 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2672 char b[BDEVNAME_SIZE];
2673 struct r5conf *conf = mddev->private;
2674 unsigned long flags;
2675 pr_debug("raid456: error called\n");
2677 spin_lock_irqsave(&conf->device_lock, flags);
2679 if (test_bit(In_sync, &rdev->flags) &&
2680 mddev->degraded == conf->max_degraded) {
2682 * Don't allow to achieve failed state
2683 * Don't try to recover this device
2685 conf->recovery_disabled = mddev->recovery_disabled;
2686 spin_unlock_irqrestore(&conf->device_lock, flags);
2690 set_bit(Faulty, &rdev->flags);
2691 clear_bit(In_sync, &rdev->flags);
2692 mddev->degraded = raid5_calc_degraded(conf);
2693 spin_unlock_irqrestore(&conf->device_lock, flags);
2694 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2696 set_bit(Blocked, &rdev->flags);
2697 set_mask_bits(&mddev->sb_flags, 0,
2698 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2699 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2700 "md/raid:%s: Operation continuing on %d devices.\n",
2702 bdevname(rdev->bdev, b),
2704 conf->raid_disks - mddev->degraded);
2705 r5c_update_on_rdev_error(mddev, rdev);
2709 * Input: a 'big' sector number,
2710 * Output: index of the data and parity disk, and the sector # in them.
2712 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2713 int previous, int *dd_idx,
2714 struct stripe_head *sh)
2716 sector_t stripe, stripe2;
2717 sector_t chunk_number;
2718 unsigned int chunk_offset;
2721 sector_t new_sector;
2722 int algorithm = previous ? conf->prev_algo
2724 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2725 : conf->chunk_sectors;
2726 int raid_disks = previous ? conf->previous_raid_disks
2728 int data_disks = raid_disks - conf->max_degraded;
2730 /* First compute the information on this sector */
2733 * Compute the chunk number and the sector offset inside the chunk
2735 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2736 chunk_number = r_sector;
2739 * Compute the stripe number
2741 stripe = chunk_number;
2742 *dd_idx = sector_div(stripe, data_disks);
2745 * Select the parity disk based on the user selected algorithm.
2747 pd_idx = qd_idx = -1;
2748 switch(conf->level) {
2750 pd_idx = data_disks;
2753 switch (algorithm) {
2754 case ALGORITHM_LEFT_ASYMMETRIC:
2755 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2756 if (*dd_idx >= pd_idx)
2759 case ALGORITHM_RIGHT_ASYMMETRIC:
2760 pd_idx = sector_div(stripe2, raid_disks);
2761 if (*dd_idx >= pd_idx)
2764 case ALGORITHM_LEFT_SYMMETRIC:
2765 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2766 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2768 case ALGORITHM_RIGHT_SYMMETRIC:
2769 pd_idx = sector_div(stripe2, raid_disks);
2770 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2772 case ALGORITHM_PARITY_0:
2776 case ALGORITHM_PARITY_N:
2777 pd_idx = data_disks;
2785 switch (algorithm) {
2786 case ALGORITHM_LEFT_ASYMMETRIC:
2787 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2788 qd_idx = pd_idx + 1;
2789 if (pd_idx == raid_disks-1) {
2790 (*dd_idx)++; /* Q D D D P */
2792 } else if (*dd_idx >= pd_idx)
2793 (*dd_idx) += 2; /* D D P Q D */
2795 case ALGORITHM_RIGHT_ASYMMETRIC:
2796 pd_idx = sector_div(stripe2, raid_disks);
2797 qd_idx = pd_idx + 1;
2798 if (pd_idx == raid_disks-1) {
2799 (*dd_idx)++; /* Q D D D P */
2801 } else if (*dd_idx >= pd_idx)
2802 (*dd_idx) += 2; /* D D P Q D */
2804 case ALGORITHM_LEFT_SYMMETRIC:
2805 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2806 qd_idx = (pd_idx + 1) % raid_disks;
2807 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2809 case ALGORITHM_RIGHT_SYMMETRIC:
2810 pd_idx = sector_div(stripe2, raid_disks);
2811 qd_idx = (pd_idx + 1) % raid_disks;
2812 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2815 case ALGORITHM_PARITY_0:
2820 case ALGORITHM_PARITY_N:
2821 pd_idx = data_disks;
2822 qd_idx = data_disks + 1;
2825 case ALGORITHM_ROTATING_ZERO_RESTART:
2826 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2827 * of blocks for computing Q is different.
2829 pd_idx = sector_div(stripe2, raid_disks);
2830 qd_idx = pd_idx + 1;
2831 if (pd_idx == raid_disks-1) {
2832 (*dd_idx)++; /* Q D D D P */
2834 } else if (*dd_idx >= pd_idx)
2835 (*dd_idx) += 2; /* D D P Q D */
2839 case ALGORITHM_ROTATING_N_RESTART:
2840 /* Same a left_asymmetric, by first stripe is
2841 * D D D P Q rather than
2845 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2846 qd_idx = pd_idx + 1;
2847 if (pd_idx == raid_disks-1) {
2848 (*dd_idx)++; /* Q D D D P */
2850 } else if (*dd_idx >= pd_idx)
2851 (*dd_idx) += 2; /* D D P Q D */
2855 case ALGORITHM_ROTATING_N_CONTINUE:
2856 /* Same as left_symmetric but Q is before P */
2857 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2858 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2859 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2863 case ALGORITHM_LEFT_ASYMMETRIC_6:
2864 /* RAID5 left_asymmetric, with Q on last device */
2865 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2866 if (*dd_idx >= pd_idx)
2868 qd_idx = raid_disks - 1;
2871 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2872 pd_idx = sector_div(stripe2, raid_disks-1);
2873 if (*dd_idx >= pd_idx)
2875 qd_idx = raid_disks - 1;
2878 case ALGORITHM_LEFT_SYMMETRIC_6:
2879 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2880 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2881 qd_idx = raid_disks - 1;
2884 case ALGORITHM_RIGHT_SYMMETRIC_6:
2885 pd_idx = sector_div(stripe2, raid_disks-1);
2886 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2887 qd_idx = raid_disks - 1;
2890 case ALGORITHM_PARITY_0_6:
2893 qd_idx = raid_disks - 1;
2903 sh->pd_idx = pd_idx;
2904 sh->qd_idx = qd_idx;
2905 sh->ddf_layout = ddf_layout;
2908 * Finally, compute the new sector number
2910 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2914 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2916 struct r5conf *conf = sh->raid_conf;
2917 int raid_disks = sh->disks;
2918 int data_disks = raid_disks - conf->max_degraded;
2919 sector_t new_sector = sh->sector, check;
2920 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2921 : conf->chunk_sectors;
2922 int algorithm = previous ? conf->prev_algo
2926 sector_t chunk_number;
2927 int dummy1, dd_idx = i;
2929 struct stripe_head sh2;
2931 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2932 stripe = new_sector;
2934 if (i == sh->pd_idx)
2936 switch(conf->level) {
2939 switch (algorithm) {
2940 case ALGORITHM_LEFT_ASYMMETRIC:
2941 case ALGORITHM_RIGHT_ASYMMETRIC:
2945 case ALGORITHM_LEFT_SYMMETRIC:
2946 case ALGORITHM_RIGHT_SYMMETRIC:
2949 i -= (sh->pd_idx + 1);
2951 case ALGORITHM_PARITY_0:
2954 case ALGORITHM_PARITY_N:
2961 if (i == sh->qd_idx)
2962 return 0; /* It is the Q disk */
2963 switch (algorithm) {
2964 case ALGORITHM_LEFT_ASYMMETRIC:
2965 case ALGORITHM_RIGHT_ASYMMETRIC:
2966 case ALGORITHM_ROTATING_ZERO_RESTART:
2967 case ALGORITHM_ROTATING_N_RESTART:
2968 if (sh->pd_idx == raid_disks-1)
2969 i--; /* Q D D D P */
2970 else if (i > sh->pd_idx)
2971 i -= 2; /* D D P Q D */
2973 case ALGORITHM_LEFT_SYMMETRIC:
2974 case ALGORITHM_RIGHT_SYMMETRIC:
2975 if (sh->pd_idx == raid_disks-1)
2976 i--; /* Q D D D P */
2981 i -= (sh->pd_idx + 2);
2984 case ALGORITHM_PARITY_0:
2987 case ALGORITHM_PARITY_N:
2989 case ALGORITHM_ROTATING_N_CONTINUE:
2990 /* Like left_symmetric, but P is before Q */
2991 if (sh->pd_idx == 0)
2992 i--; /* P D D D Q */
2997 i -= (sh->pd_idx + 1);
3000 case ALGORITHM_LEFT_ASYMMETRIC_6:
3001 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3005 case ALGORITHM_LEFT_SYMMETRIC_6:
3006 case ALGORITHM_RIGHT_SYMMETRIC_6:
3008 i += data_disks + 1;
3009 i -= (sh->pd_idx + 1);
3011 case ALGORITHM_PARITY_0_6:
3020 chunk_number = stripe * data_disks + i;
3021 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3023 check = raid5_compute_sector(conf, r_sector,
3024 previous, &dummy1, &sh2);
3025 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3026 || sh2.qd_idx != sh->qd_idx) {
3027 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3028 mdname(conf->mddev));
3035 * There are cases where we want handle_stripe_dirtying() and
3036 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3038 * This function checks whether we want to delay the towrite. Specifically,
3039 * we delay the towrite when:
3041 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3042 * stripe has data in journal (for other devices).
3044 * In this case, when reading data for the non-overwrite dev, it is
3045 * necessary to handle complex rmw of write back cache (prexor with
3046 * orig_page, and xor with page). To keep read path simple, we would
3047 * like to flush data in journal to RAID disks first, so complex rmw
3048 * is handled in the write patch (handle_stripe_dirtying).
3050 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3052 * It is important to be able to flush all stripes in raid5-cache.
3053 * Therefore, we need reserve some space on the journal device for
3054 * these flushes. If flush operation includes pending writes to the
3055 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3056 * for the flush out. If we exclude these pending writes from flush
3057 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3058 * Therefore, excluding pending writes in these cases enables more
3059 * efficient use of the journal device.
3061 * Note: To make sure the stripe makes progress, we only delay
3062 * towrite for stripes with data already in journal (injournal > 0).
3063 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3064 * no_space_stripes list.
3066 * 3. during journal failure
3067 * In journal failure, we try to flush all cached data to raid disks
3068 * based on data in stripe cache. The array is read-only to upper
3069 * layers, so we would skip all pending writes.
3072 static inline bool delay_towrite(struct r5conf *conf,
3074 struct stripe_head_state *s)
3077 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3078 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3081 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3085 if (s->log_failed && s->injournal)
3091 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3092 int rcw, int expand)
3094 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3095 struct r5conf *conf = sh->raid_conf;
3096 int level = conf->level;
3100 * In some cases, handle_stripe_dirtying initially decided to
3101 * run rmw and allocates extra page for prexor. However, rcw is
3102 * cheaper later on. We need to free the extra page now,
3103 * because we won't be able to do that in ops_complete_prexor().
3105 r5c_release_extra_page(sh);
3107 for (i = disks; i--; ) {
3108 struct r5dev *dev = &sh->dev[i];
3110 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3111 set_bit(R5_LOCKED, &dev->flags);
3112 set_bit(R5_Wantdrain, &dev->flags);
3114 clear_bit(R5_UPTODATE, &dev->flags);
3116 } else if (test_bit(R5_InJournal, &dev->flags)) {
3117 set_bit(R5_LOCKED, &dev->flags);
3121 /* if we are not expanding this is a proper write request, and
3122 * there will be bios with new data to be drained into the
3127 /* False alarm, nothing to do */
3129 sh->reconstruct_state = reconstruct_state_drain_run;
3130 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3132 sh->reconstruct_state = reconstruct_state_run;
3134 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3136 if (s->locked + conf->max_degraded == disks)
3137 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3138 atomic_inc(&conf->pending_full_writes);
3140 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3141 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3142 BUG_ON(level == 6 &&
3143 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3144 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3146 for (i = disks; i--; ) {
3147 struct r5dev *dev = &sh->dev[i];
3148 if (i == pd_idx || i == qd_idx)
3152 (test_bit(R5_UPTODATE, &dev->flags) ||
3153 test_bit(R5_Wantcompute, &dev->flags))) {
3154 set_bit(R5_Wantdrain, &dev->flags);
3155 set_bit(R5_LOCKED, &dev->flags);
3156 clear_bit(R5_UPTODATE, &dev->flags);
3158 } else if (test_bit(R5_InJournal, &dev->flags)) {
3159 set_bit(R5_LOCKED, &dev->flags);
3164 /* False alarm - nothing to do */
3166 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3167 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3168 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3169 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3172 /* keep the parity disk(s) locked while asynchronous operations
3175 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3176 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3180 int qd_idx = sh->qd_idx;
3181 struct r5dev *dev = &sh->dev[qd_idx];
3183 set_bit(R5_LOCKED, &dev->flags);
3184 clear_bit(R5_UPTODATE, &dev->flags);
3188 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3189 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3190 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3191 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3192 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3194 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3195 __func__, (unsigned long long)sh->sector,
3196 s->locked, s->ops_request);
3200 * Each stripe/dev can have one or more bion attached.
3201 * toread/towrite point to the first in a chain.
3202 * The bi_next chain must be in order.
3204 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3205 int forwrite, int previous)
3208 struct r5conf *conf = sh->raid_conf;
3211 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3212 (unsigned long long)bi->bi_iter.bi_sector,
3213 (unsigned long long)sh->sector);
3215 spin_lock_irq(&sh->stripe_lock);
3216 sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3217 /* Don't allow new IO added to stripes in batch list */
3221 bip = &sh->dev[dd_idx].towrite;
3225 bip = &sh->dev[dd_idx].toread;
3226 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3227 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3229 bip = & (*bip)->bi_next;
3231 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3234 if (forwrite && raid5_has_ppl(conf)) {
3236 * With PPL only writes to consecutive data chunks within a
3237 * stripe are allowed because for a single stripe_head we can
3238 * only have one PPL entry at a time, which describes one data
3239 * range. Not really an overlap, but wait_for_overlap can be
3240 * used to handle this.
3248 for (i = 0; i < sh->disks; i++) {
3249 if (i != sh->pd_idx &&
3250 (i == dd_idx || sh->dev[i].towrite)) {
3251 sector = sh->dev[i].sector;
3252 if (count == 0 || sector < first)
3260 if (first + conf->chunk_sectors * (count - 1) != last)
3264 if (!forwrite || previous)
3265 clear_bit(STRIPE_BATCH_READY, &sh->state);
3267 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3271 bio_inc_remaining(bi);
3272 md_write_inc(conf->mddev, bi);
3275 /* check if page is covered */
3276 sector_t sector = sh->dev[dd_idx].sector;
3277 for (bi=sh->dev[dd_idx].towrite;
3278 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3279 bi && bi->bi_iter.bi_sector <= sector;
3280 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3281 if (bio_end_sector(bi) >= sector)
3282 sector = bio_end_sector(bi);
3284 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3285 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3286 sh->overwrite_disks++;
3289 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3290 (unsigned long long)(*bip)->bi_iter.bi_sector,
3291 (unsigned long long)sh->sector, dd_idx);
3293 if (conf->mddev->bitmap && firstwrite) {
3294 /* Cannot hold spinlock over bitmap_startwrite,
3295 * but must ensure this isn't added to a batch until
3296 * we have added to the bitmap and set bm_seq.
3297 * So set STRIPE_BITMAP_PENDING to prevent
3299 * If multiple add_stripe_bio() calls race here they
3300 * much all set STRIPE_BITMAP_PENDING. So only the first one
3301 * to complete "bitmap_startwrite" gets to set
3302 * STRIPE_BIT_DELAY. This is important as once a stripe
3303 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3306 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3307 spin_unlock_irq(&sh->stripe_lock);
3308 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3310 spin_lock_irq(&sh->stripe_lock);
3311 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3312 if (!sh->batch_head) {
3313 sh->bm_seq = conf->seq_flush+1;
3314 set_bit(STRIPE_BIT_DELAY, &sh->state);
3317 spin_unlock_irq(&sh->stripe_lock);
3319 if (stripe_can_batch(sh))
3320 stripe_add_to_batch_list(conf, sh);
3324 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3325 spin_unlock_irq(&sh->stripe_lock);
3329 static void end_reshape(struct r5conf *conf);
3331 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3332 struct stripe_head *sh)
3334 int sectors_per_chunk =
3335 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3337 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3338 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3340 raid5_compute_sector(conf,
3341 stripe * (disks - conf->max_degraded)
3342 *sectors_per_chunk + chunk_offset,
3348 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3349 struct stripe_head_state *s, int disks)
3352 BUG_ON(sh->batch_head);
3353 for (i = disks; i--; ) {
3357 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3358 struct md_rdev *rdev;
3360 rdev = rcu_dereference(conf->disks[i].rdev);
3361 if (rdev && test_bit(In_sync, &rdev->flags) &&
3362 !test_bit(Faulty, &rdev->flags))
3363 atomic_inc(&rdev->nr_pending);
3368 if (!rdev_set_badblocks(
3372 md_error(conf->mddev, rdev);
3373 rdev_dec_pending(rdev, conf->mddev);
3376 spin_lock_irq(&sh->stripe_lock);
3377 /* fail all writes first */
3378 bi = sh->dev[i].towrite;
3379 sh->dev[i].towrite = NULL;
3380 sh->overwrite_disks = 0;
3381 spin_unlock_irq(&sh->stripe_lock);
3385 log_stripe_write_finished(sh);
3387 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3388 wake_up(&conf->wait_for_overlap);
3390 while (bi && bi->bi_iter.bi_sector <
3391 sh->dev[i].sector + STRIPE_SECTORS) {
3392 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3394 md_write_end(conf->mddev);
3399 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3400 STRIPE_SECTORS, 0, 0);
3402 /* and fail all 'written' */
3403 bi = sh->dev[i].written;
3404 sh->dev[i].written = NULL;
3405 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3406 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3407 sh->dev[i].page = sh->dev[i].orig_page;
3410 if (bi) bitmap_end = 1;
3411 while (bi && bi->bi_iter.bi_sector <
3412 sh->dev[i].sector + STRIPE_SECTORS) {
3413 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3415 md_write_end(conf->mddev);
3420 /* fail any reads if this device is non-operational and
3421 * the data has not reached the cache yet.
3423 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3424 s->failed > conf->max_degraded &&
3425 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3426 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3427 spin_lock_irq(&sh->stripe_lock);
3428 bi = sh->dev[i].toread;
3429 sh->dev[i].toread = NULL;
3430 spin_unlock_irq(&sh->stripe_lock);
3431 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3432 wake_up(&conf->wait_for_overlap);
3435 while (bi && bi->bi_iter.bi_sector <
3436 sh->dev[i].sector + STRIPE_SECTORS) {
3437 struct bio *nextbi =
3438 r5_next_bio(bi, sh->dev[i].sector);
3445 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3446 STRIPE_SECTORS, 0, 0);
3447 /* If we were in the middle of a write the parity block might
3448 * still be locked - so just clear all R5_LOCKED flags
3450 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3455 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3456 if (atomic_dec_and_test(&conf->pending_full_writes))
3457 md_wakeup_thread(conf->mddev->thread);
3461 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3462 struct stripe_head_state *s)
3467 BUG_ON(sh->batch_head);
3468 clear_bit(STRIPE_SYNCING, &sh->state);
3469 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3470 wake_up(&conf->wait_for_overlap);
3473 /* There is nothing more to do for sync/check/repair.
3474 * Don't even need to abort as that is handled elsewhere
3475 * if needed, and not always wanted e.g. if there is a known
3477 * For recover/replace we need to record a bad block on all
3478 * non-sync devices, or abort the recovery
3480 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3481 /* During recovery devices cannot be removed, so
3482 * locking and refcounting of rdevs is not needed
3485 for (i = 0; i < conf->raid_disks; i++) {
3486 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3488 && !test_bit(Faulty, &rdev->flags)
3489 && !test_bit(In_sync, &rdev->flags)
3490 && !rdev_set_badblocks(rdev, sh->sector,
3493 rdev = rcu_dereference(conf->disks[i].replacement);
3495 && !test_bit(Faulty, &rdev->flags)
3496 && !test_bit(In_sync, &rdev->flags)
3497 && !rdev_set_badblocks(rdev, sh->sector,
3503 conf->recovery_disabled =
3504 conf->mddev->recovery_disabled;
3506 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3509 static int want_replace(struct stripe_head *sh, int disk_idx)
3511 struct md_rdev *rdev;
3515 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3517 && !test_bit(Faulty, &rdev->flags)
3518 && !test_bit(In_sync, &rdev->flags)
3519 && (rdev->recovery_offset <= sh->sector
3520 || rdev->mddev->recovery_cp <= sh->sector))
3526 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3527 int disk_idx, int disks)
3529 struct r5dev *dev = &sh->dev[disk_idx];
3530 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3531 &sh->dev[s->failed_num[1]] };
3535 if (test_bit(R5_LOCKED, &dev->flags) ||
3536 test_bit(R5_UPTODATE, &dev->flags))
3537 /* No point reading this as we already have it or have
3538 * decided to get it.
3543 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3544 /* We need this block to directly satisfy a request */
3547 if (s->syncing || s->expanding ||
3548 (s->replacing && want_replace(sh, disk_idx)))
3549 /* When syncing, or expanding we read everything.
3550 * When replacing, we need the replaced block.
3554 if ((s->failed >= 1 && fdev[0]->toread) ||
3555 (s->failed >= 2 && fdev[1]->toread))
3556 /* If we want to read from a failed device, then
3557 * we need to actually read every other device.
3561 /* Sometimes neither read-modify-write nor reconstruct-write
3562 * cycles can work. In those cases we read every block we
3563 * can. Then the parity-update is certain to have enough to
3565 * This can only be a problem when we need to write something,
3566 * and some device has failed. If either of those tests
3567 * fail we need look no further.
3569 if (!s->failed || !s->to_write)
3572 if (test_bit(R5_Insync, &dev->flags) &&
3573 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3574 /* Pre-reads at not permitted until after short delay
3575 * to gather multiple requests. However if this
3576 * device is no Insync, the block could only be computed
3577 * and there is no need to delay that.
3581 for (i = 0; i < s->failed && i < 2; i++) {
3582 if (fdev[i]->towrite &&
3583 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3584 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3585 /* If we have a partial write to a failed
3586 * device, then we will need to reconstruct
3587 * the content of that device, so all other
3588 * devices must be read.
3593 /* If we are forced to do a reconstruct-write, either because
3594 * the current RAID6 implementation only supports that, or
3595 * because parity cannot be trusted and we are currently
3596 * recovering it, there is extra need to be careful.
3597 * If one of the devices that we would need to read, because
3598 * it is not being overwritten (and maybe not written at all)
3599 * is missing/faulty, then we need to read everything we can.
3601 if (sh->raid_conf->level != 6 &&
3602 sh->sector < sh->raid_conf->mddev->recovery_cp)
3603 /* reconstruct-write isn't being forced */
3605 for (i = 0; i < s->failed && i < 2; i++) {
3606 if (s->failed_num[i] != sh->pd_idx &&
3607 s->failed_num[i] != sh->qd_idx &&
3608 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3609 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3616 /* fetch_block - checks the given member device to see if its data needs
3617 * to be read or computed to satisfy a request.
3619 * Returns 1 when no more member devices need to be checked, otherwise returns
3620 * 0 to tell the loop in handle_stripe_fill to continue
3622 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3623 int disk_idx, int disks)
3625 struct r5dev *dev = &sh->dev[disk_idx];
3627 /* is the data in this block needed, and can we get it? */
3628 if (need_this_block(sh, s, disk_idx, disks)) {
3629 /* we would like to get this block, possibly by computing it,
3630 * otherwise read it if the backing disk is insync
3632 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3633 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3634 BUG_ON(sh->batch_head);
3637 * In the raid6 case if the only non-uptodate disk is P
3638 * then we already trusted P to compute the other failed
3639 * drives. It is safe to compute rather than re-read P.
3640 * In other cases we only compute blocks from failed
3641 * devices, otherwise check/repair might fail to detect
3642 * a real inconsistency.
3645 if ((s->uptodate == disks - 1) &&
3646 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3647 (s->failed && (disk_idx == s->failed_num[0] ||
3648 disk_idx == s->failed_num[1])))) {
3649 /* have disk failed, and we're requested to fetch it;
3652 pr_debug("Computing stripe %llu block %d\n",
3653 (unsigned long long)sh->sector, disk_idx);
3654 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3655 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3656 set_bit(R5_Wantcompute, &dev->flags);
3657 sh->ops.target = disk_idx;
3658 sh->ops.target2 = -1; /* no 2nd target */
3660 /* Careful: from this point on 'uptodate' is in the eye
3661 * of raid_run_ops which services 'compute' operations
3662 * before writes. R5_Wantcompute flags a block that will
3663 * be R5_UPTODATE by the time it is needed for a
3664 * subsequent operation.
3668 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3669 /* Computing 2-failure is *very* expensive; only
3670 * do it if failed >= 2
3673 for (other = disks; other--; ) {
3674 if (other == disk_idx)
3676 if (!test_bit(R5_UPTODATE,
3677 &sh->dev[other].flags))
3681 pr_debug("Computing stripe %llu blocks %d,%d\n",
3682 (unsigned long long)sh->sector,
3684 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3685 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3686 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3687 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3688 sh->ops.target = disk_idx;
3689 sh->ops.target2 = other;
3693 } else if (test_bit(R5_Insync, &dev->flags)) {
3694 set_bit(R5_LOCKED, &dev->flags);
3695 set_bit(R5_Wantread, &dev->flags);
3697 pr_debug("Reading block %d (sync=%d)\n",
3698 disk_idx, s->syncing);
3706 * handle_stripe_fill - read or compute data to satisfy pending requests.
3708 static void handle_stripe_fill(struct stripe_head *sh,
3709 struct stripe_head_state *s,
3714 /* look for blocks to read/compute, skip this if a compute
3715 * is already in flight, or if the stripe contents are in the
3716 * midst of changing due to a write
3718 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3719 !sh->reconstruct_state) {
3722 * For degraded stripe with data in journal, do not handle
3723 * read requests yet, instead, flush the stripe to raid
3724 * disks first, this avoids handling complex rmw of write
3725 * back cache (prexor with orig_page, and then xor with
3726 * page) in the read path
3728 if (s->injournal && s->failed) {
3729 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3730 r5c_make_stripe_write_out(sh);
3734 for (i = disks; i--; )
3735 if (fetch_block(sh, s, i, disks))
3739 set_bit(STRIPE_HANDLE, &sh->state);
3742 static void break_stripe_batch_list(struct stripe_head *head_sh,
3743 unsigned long handle_flags);
3744 /* handle_stripe_clean_event
3745 * any written block on an uptodate or failed drive can be returned.
3746 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3747 * never LOCKED, so we don't need to test 'failed' directly.
3749 static void handle_stripe_clean_event(struct r5conf *conf,
3750 struct stripe_head *sh, int disks)
3754 int discard_pending = 0;
3755 struct stripe_head *head_sh = sh;
3756 bool do_endio = false;
3758 for (i = disks; i--; )
3759 if (sh->dev[i].written) {
3761 if (!test_bit(R5_LOCKED, &dev->flags) &&
3762 (test_bit(R5_UPTODATE, &dev->flags) ||
3763 test_bit(R5_Discard, &dev->flags) ||
3764 test_bit(R5_SkipCopy, &dev->flags))) {
3765 /* We can return any write requests */
3766 struct bio *wbi, *wbi2;
3767 pr_debug("Return write for disc %d\n", i);
3768 if (test_and_clear_bit(R5_Discard, &dev->flags))
3769 clear_bit(R5_UPTODATE, &dev->flags);
3770 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3771 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3776 dev->page = dev->orig_page;
3778 dev->written = NULL;
3779 while (wbi && wbi->bi_iter.bi_sector <
3780 dev->sector + STRIPE_SECTORS) {
3781 wbi2 = r5_next_bio(wbi, dev->sector);
3782 md_write_end(conf->mddev);
3786 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3788 !test_bit(STRIPE_DEGRADED, &sh->state),
3790 if (head_sh->batch_head) {
3791 sh = list_first_entry(&sh->batch_list,
3794 if (sh != head_sh) {
3801 } else if (test_bit(R5_Discard, &dev->flags))
3802 discard_pending = 1;
3805 log_stripe_write_finished(sh);
3807 if (!discard_pending &&
3808 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3810 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3811 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3812 if (sh->qd_idx >= 0) {
3813 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3814 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3816 /* now that discard is done we can proceed with any sync */
3817 clear_bit(STRIPE_DISCARD, &sh->state);
3819 * SCSI discard will change some bio fields and the stripe has
3820 * no updated data, so remove it from hash list and the stripe
3821 * will be reinitialized
3824 hash = sh->hash_lock_index;
3825 spin_lock_irq(conf->hash_locks + hash);
3827 spin_unlock_irq(conf->hash_locks + hash);
3828 if (head_sh->batch_head) {
3829 sh = list_first_entry(&sh->batch_list,
3830 struct stripe_head, batch_list);
3836 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3837 set_bit(STRIPE_HANDLE, &sh->state);
3841 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3842 if (atomic_dec_and_test(&conf->pending_full_writes))
3843 md_wakeup_thread(conf->mddev->thread);
3845 if (head_sh->batch_head && do_endio)
3846 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3850 * For RMW in write back cache, we need extra page in prexor to store the
3851 * old data. This page is stored in dev->orig_page.
3853 * This function checks whether we have data for prexor. The exact logic
3855 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3857 static inline bool uptodate_for_rmw(struct r5dev *dev)
3859 return (test_bit(R5_UPTODATE, &dev->flags)) &&
3860 (!test_bit(R5_InJournal, &dev->flags) ||
3861 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3864 static int handle_stripe_dirtying(struct r5conf *conf,
3865 struct stripe_head *sh,
3866 struct stripe_head_state *s,
3869 int rmw = 0, rcw = 0, i;
3870 sector_t recovery_cp = conf->mddev->recovery_cp;
3872 /* Check whether resync is now happening or should start.
3873 * If yes, then the array is dirty (after unclean shutdown or
3874 * initial creation), so parity in some stripes might be inconsistent.
3875 * In this case, we need to always do reconstruct-write, to ensure
3876 * that in case of drive failure or read-error correction, we
3877 * generate correct data from the parity.
3879 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3880 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3882 /* Calculate the real rcw later - for now make it
3883 * look like rcw is cheaper
3886 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3887 conf->rmw_level, (unsigned long long)recovery_cp,
3888 (unsigned long long)sh->sector);
3889 } else for (i = disks; i--; ) {
3890 /* would I have to read this buffer for read_modify_write */
3891 struct r5dev *dev = &sh->dev[i];
3892 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3893 i == sh->pd_idx || i == sh->qd_idx ||
3894 test_bit(R5_InJournal, &dev->flags)) &&
3895 !test_bit(R5_LOCKED, &dev->flags) &&
3896 !(uptodate_for_rmw(dev) ||
3897 test_bit(R5_Wantcompute, &dev->flags))) {
3898 if (test_bit(R5_Insync, &dev->flags))
3901 rmw += 2*disks; /* cannot read it */
3903 /* Would I have to read this buffer for reconstruct_write */
3904 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3905 i != sh->pd_idx && i != sh->qd_idx &&
3906 !test_bit(R5_LOCKED, &dev->flags) &&
3907 !(test_bit(R5_UPTODATE, &dev->flags) ||
3908 test_bit(R5_Wantcompute, &dev->flags))) {
3909 if (test_bit(R5_Insync, &dev->flags))
3916 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3917 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3918 set_bit(STRIPE_HANDLE, &sh->state);
3919 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3920 /* prefer read-modify-write, but need to get some data */
3921 if (conf->mddev->queue)
3922 blk_add_trace_msg(conf->mddev->queue,
3923 "raid5 rmw %llu %d",
3924 (unsigned long long)sh->sector, rmw);
3925 for (i = disks; i--; ) {
3926 struct r5dev *dev = &sh->dev[i];
3927 if (test_bit(R5_InJournal, &dev->flags) &&
3928 dev->page == dev->orig_page &&
3929 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3930 /* alloc page for prexor */
3931 struct page *p = alloc_page(GFP_NOIO);
3939 * alloc_page() failed, try use
3940 * disk_info->extra_page
3942 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3943 &conf->cache_state)) {
3944 r5c_use_extra_page(sh);
3948 /* extra_page in use, add to delayed_list */
3949 set_bit(STRIPE_DELAYED, &sh->state);
3950 s->waiting_extra_page = 1;
3955 for (i = disks; i--; ) {
3956 struct r5dev *dev = &sh->dev[i];
3957 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3958 i == sh->pd_idx || i == sh->qd_idx ||
3959 test_bit(R5_InJournal, &dev->flags)) &&
3960 !test_bit(R5_LOCKED, &dev->flags) &&
3961 !(uptodate_for_rmw(dev) ||
3962 test_bit(R5_Wantcompute, &dev->flags)) &&
3963 test_bit(R5_Insync, &dev->flags)) {
3964 if (test_bit(STRIPE_PREREAD_ACTIVE,
3966 pr_debug("Read_old block %d for r-m-w\n",
3968 set_bit(R5_LOCKED, &dev->flags);
3969 set_bit(R5_Wantread, &dev->flags);
3972 set_bit(STRIPE_DELAYED, &sh->state);
3973 set_bit(STRIPE_HANDLE, &sh->state);
3978 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3979 /* want reconstruct write, but need to get some data */
3982 for (i = disks; i--; ) {
3983 struct r5dev *dev = &sh->dev[i];
3984 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3985 i != sh->pd_idx && i != sh->qd_idx &&
3986 !test_bit(R5_LOCKED, &dev->flags) &&
3987 !(test_bit(R5_UPTODATE, &dev->flags) ||
3988 test_bit(R5_Wantcompute, &dev->flags))) {
3990 if (test_bit(R5_Insync, &dev->flags) &&
3991 test_bit(STRIPE_PREREAD_ACTIVE,
3993 pr_debug("Read_old block "
3994 "%d for Reconstruct\n", i);
3995 set_bit(R5_LOCKED, &dev->flags);
3996 set_bit(R5_Wantread, &dev->flags);
4000 set_bit(STRIPE_DELAYED, &sh->state);
4001 set_bit(STRIPE_HANDLE, &sh->state);
4005 if (rcw && conf->mddev->queue)
4006 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4007 (unsigned long long)sh->sector,
4008 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4011 if (rcw > disks && rmw > disks &&
4012 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4013 set_bit(STRIPE_DELAYED, &sh->state);
4015 /* now if nothing is locked, and if we have enough data,
4016 * we can start a write request
4018 /* since handle_stripe can be called at any time we need to handle the
4019 * case where a compute block operation has been submitted and then a
4020 * subsequent call wants to start a write request. raid_run_ops only
4021 * handles the case where compute block and reconstruct are requested
4022 * simultaneously. If this is not the case then new writes need to be
4023 * held off until the compute completes.
4025 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4026 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4027 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4028 schedule_reconstruction(sh, s, rcw == 0, 0);
4032 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4033 struct stripe_head_state *s, int disks)
4035 struct r5dev *dev = NULL;
4037 BUG_ON(sh->batch_head);
4038 set_bit(STRIPE_HANDLE, &sh->state);
4040 switch (sh->check_state) {
4041 case check_state_idle:
4042 /* start a new check operation if there are no failures */
4043 if (s->failed == 0) {
4044 BUG_ON(s->uptodate != disks);
4045 sh->check_state = check_state_run;
4046 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4047 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4051 dev = &sh->dev[s->failed_num[0]];
4053 case check_state_compute_result:
4054 sh->check_state = check_state_idle;
4056 dev = &sh->dev[sh->pd_idx];
4058 /* check that a write has not made the stripe insync */
4059 if (test_bit(STRIPE_INSYNC, &sh->state))
4062 /* either failed parity check, or recovery is happening */
4063 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4064 BUG_ON(s->uptodate != disks);
4066 set_bit(R5_LOCKED, &dev->flags);
4068 set_bit(R5_Wantwrite, &dev->flags);
4070 clear_bit(STRIPE_DEGRADED, &sh->state);
4071 set_bit(STRIPE_INSYNC, &sh->state);
4073 case check_state_run:
4074 break; /* we will be called again upon completion */
4075 case check_state_check_result:
4076 sh->check_state = check_state_idle;
4078 /* if a failure occurred during the check operation, leave
4079 * STRIPE_INSYNC not set and let the stripe be handled again
4084 /* handle a successful check operation, if parity is correct
4085 * we are done. Otherwise update the mismatch count and repair
4086 * parity if !MD_RECOVERY_CHECK
4088 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4089 /* parity is correct (on disc,
4090 * not in buffer any more)
4092 set_bit(STRIPE_INSYNC, &sh->state);
4094 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4095 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4096 /* don't try to repair!! */
4097 set_bit(STRIPE_INSYNC, &sh->state);
4098 pr_warn_ratelimited("%s: mismatch sector in range "
4099 "%llu-%llu\n", mdname(conf->mddev),
4100 (unsigned long long) sh->sector,
4101 (unsigned long long) sh->sector +
4104 sh->check_state = check_state_compute_run;
4105 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4106 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4107 set_bit(R5_Wantcompute,
4108 &sh->dev[sh->pd_idx].flags);
4109 sh->ops.target = sh->pd_idx;
4110 sh->ops.target2 = -1;
4115 case check_state_compute_run:
4118 pr_err("%s: unknown check_state: %d sector: %llu\n",
4119 __func__, sh->check_state,
4120 (unsigned long long) sh->sector);
4125 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4126 struct stripe_head_state *s,
4129 int pd_idx = sh->pd_idx;
4130 int qd_idx = sh->qd_idx;
4133 BUG_ON(sh->batch_head);
4134 set_bit(STRIPE_HANDLE, &sh->state);
4136 BUG_ON(s->failed > 2);
4138 /* Want to check and possibly repair P and Q.
4139 * However there could be one 'failed' device, in which
4140 * case we can only check one of them, possibly using the
4141 * other to generate missing data
4144 switch (sh->check_state) {
4145 case check_state_idle:
4146 /* start a new check operation if there are < 2 failures */
4147 if (s->failed == s->q_failed) {
4148 /* The only possible failed device holds Q, so it
4149 * makes sense to check P (If anything else were failed,
4150 * we would have used P to recreate it).
4152 sh->check_state = check_state_run;
4154 if (!s->q_failed && s->failed < 2) {
4155 /* Q is not failed, and we didn't use it to generate
4156 * anything, so it makes sense to check it
4158 if (sh->check_state == check_state_run)
4159 sh->check_state = check_state_run_pq;
4161 sh->check_state = check_state_run_q;
4164 /* discard potentially stale zero_sum_result */
4165 sh->ops.zero_sum_result = 0;
4167 if (sh->check_state == check_state_run) {
4168 /* async_xor_zero_sum destroys the contents of P */
4169 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4172 if (sh->check_state >= check_state_run &&
4173 sh->check_state <= check_state_run_pq) {
4174 /* async_syndrome_zero_sum preserves P and Q, so
4175 * no need to mark them !uptodate here
4177 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4181 /* we have 2-disk failure */
4182 BUG_ON(s->failed != 2);
4184 case check_state_compute_result:
4185 sh->check_state = check_state_idle;
4187 /* check that a write has not made the stripe insync */
4188 if (test_bit(STRIPE_INSYNC, &sh->state))
4191 /* now write out any block on a failed drive,
4192 * or P or Q if they were recomputed
4195 if (s->failed == 2) {
4196 dev = &sh->dev[s->failed_num[1]];
4198 set_bit(R5_LOCKED, &dev->flags);
4199 set_bit(R5_Wantwrite, &dev->flags);
4201 if (s->failed >= 1) {
4202 dev = &sh->dev[s->failed_num[0]];
4204 set_bit(R5_LOCKED, &dev->flags);
4205 set_bit(R5_Wantwrite, &dev->flags);
4207 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4208 dev = &sh->dev[pd_idx];
4210 set_bit(R5_LOCKED, &dev->flags);
4211 set_bit(R5_Wantwrite, &dev->flags);
4213 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4214 dev = &sh->dev[qd_idx];
4216 set_bit(R5_LOCKED, &dev->flags);
4217 set_bit(R5_Wantwrite, &dev->flags);
4219 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4220 "%s: disk%td not up to date\n",
4221 mdname(conf->mddev),
4222 dev - (struct r5dev *) &sh->dev)) {
4223 clear_bit(R5_LOCKED, &dev->flags);
4224 clear_bit(R5_Wantwrite, &dev->flags);
4227 clear_bit(STRIPE_DEGRADED, &sh->state);
4229 set_bit(STRIPE_INSYNC, &sh->state);
4231 case check_state_run:
4232 case check_state_run_q:
4233 case check_state_run_pq:
4234 break; /* we will be called again upon completion */
4235 case check_state_check_result:
4236 sh->check_state = check_state_idle;
4238 /* handle a successful check operation, if parity is correct
4239 * we are done. Otherwise update the mismatch count and repair
4240 * parity if !MD_RECOVERY_CHECK
4242 if (sh->ops.zero_sum_result == 0) {
4243 /* both parities are correct */
4245 set_bit(STRIPE_INSYNC, &sh->state);
4247 /* in contrast to the raid5 case we can validate
4248 * parity, but still have a failure to write
4251 sh->check_state = check_state_compute_result;
4252 /* Returning at this point means that we may go
4253 * off and bring p and/or q uptodate again so
4254 * we make sure to check zero_sum_result again
4255 * to verify if p or q need writeback
4259 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4260 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4261 /* don't try to repair!! */
4262 set_bit(STRIPE_INSYNC, &sh->state);
4263 pr_warn_ratelimited("%s: mismatch sector in range "
4264 "%llu-%llu\n", mdname(conf->mddev),
4265 (unsigned long long) sh->sector,
4266 (unsigned long long) sh->sector +
4269 int *target = &sh->ops.target;
4271 sh->ops.target = -1;
4272 sh->ops.target2 = -1;
4273 sh->check_state = check_state_compute_run;
4274 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4275 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4276 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4277 set_bit(R5_Wantcompute,
4278 &sh->dev[pd_idx].flags);
4280 target = &sh->ops.target2;
4283 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4284 set_bit(R5_Wantcompute,
4285 &sh->dev[qd_idx].flags);
4292 case check_state_compute_run:
4295 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4296 __func__, sh->check_state,
4297 (unsigned long long) sh->sector);
4302 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4306 /* We have read all the blocks in this stripe and now we need to
4307 * copy some of them into a target stripe for expand.
4309 struct dma_async_tx_descriptor *tx = NULL;
4310 BUG_ON(sh->batch_head);
4311 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4312 for (i = 0; i < sh->disks; i++)
4313 if (i != sh->pd_idx && i != sh->qd_idx) {
4315 struct stripe_head *sh2;
4316 struct async_submit_ctl submit;
4318 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4319 sector_t s = raid5_compute_sector(conf, bn, 0,
4321 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4323 /* so far only the early blocks of this stripe
4324 * have been requested. When later blocks
4325 * get requested, we will try again
4328 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4329 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4330 /* must have already done this block */
4331 raid5_release_stripe(sh2);
4335 /* place all the copies on one channel */
4336 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4337 tx = async_memcpy(sh2->dev[dd_idx].page,
4338 sh->dev[i].page, 0, 0, STRIPE_SIZE,
4341 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4342 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4343 for (j = 0; j < conf->raid_disks; j++)
4344 if (j != sh2->pd_idx &&
4346 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4348 if (j == conf->raid_disks) {
4349 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4350 set_bit(STRIPE_HANDLE, &sh2->state);
4352 raid5_release_stripe(sh2);
4355 /* done submitting copies, wait for them to complete */
4356 async_tx_quiesce(&tx);
4360 * handle_stripe - do things to a stripe.
4362 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4363 * state of various bits to see what needs to be done.
4365 * return some read requests which now have data
4366 * return some write requests which are safely on storage
4367 * schedule a read on some buffers
4368 * schedule a write of some buffers
4369 * return confirmation of parity correctness
4373 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4375 struct r5conf *conf = sh->raid_conf;
4376 int disks = sh->disks;
4379 int do_recovery = 0;
4381 memset(s, 0, sizeof(*s));
4383 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4384 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4385 s->failed_num[0] = -1;
4386 s->failed_num[1] = -1;
4387 s->log_failed = r5l_log_disk_error(conf);
4389 /* Now to look around and see what can be done */
4391 for (i=disks; i--; ) {
4392 struct md_rdev *rdev;
4399 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4401 dev->toread, dev->towrite, dev->written);
4402 /* maybe we can reply to a read
4404 * new wantfill requests are only permitted while
4405 * ops_complete_biofill is guaranteed to be inactive
4407 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4408 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4409 set_bit(R5_Wantfill, &dev->flags);
4411 /* now count some things */
4412 if (test_bit(R5_LOCKED, &dev->flags))
4414 if (test_bit(R5_UPTODATE, &dev->flags))
4416 if (test_bit(R5_Wantcompute, &dev->flags)) {
4418 BUG_ON(s->compute > 2);
4421 if (test_bit(R5_Wantfill, &dev->flags))
4423 else if (dev->toread)
4427 if (!test_bit(R5_OVERWRITE, &dev->flags))
4432 /* Prefer to use the replacement for reads, but only
4433 * if it is recovered enough and has no bad blocks.
4435 rdev = rcu_dereference(conf->disks[i].replacement);
4436 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4437 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4438 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4439 &first_bad, &bad_sectors))
4440 set_bit(R5_ReadRepl, &dev->flags);
4442 if (rdev && !test_bit(Faulty, &rdev->flags))
4443 set_bit(R5_NeedReplace, &dev->flags);
4445 clear_bit(R5_NeedReplace, &dev->flags);
4446 rdev = rcu_dereference(conf->disks[i].rdev);
4447 clear_bit(R5_ReadRepl, &dev->flags);
4449 if (rdev && test_bit(Faulty, &rdev->flags))
4452 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4453 &first_bad, &bad_sectors);
4454 if (s->blocked_rdev == NULL
4455 && (test_bit(Blocked, &rdev->flags)
4458 set_bit(BlockedBadBlocks,
4460 s->blocked_rdev = rdev;
4461 atomic_inc(&rdev->nr_pending);
4464 clear_bit(R5_Insync, &dev->flags);
4468 /* also not in-sync */
4469 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4470 test_bit(R5_UPTODATE, &dev->flags)) {
4471 /* treat as in-sync, but with a read error
4472 * which we can now try to correct
4474 set_bit(R5_Insync, &dev->flags);
4475 set_bit(R5_ReadError, &dev->flags);
4477 } else if (test_bit(In_sync, &rdev->flags))
4478 set_bit(R5_Insync, &dev->flags);
4479 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4480 /* in sync if before recovery_offset */
4481 set_bit(R5_Insync, &dev->flags);
4482 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4483 test_bit(R5_Expanded, &dev->flags))
4484 /* If we've reshaped into here, we assume it is Insync.
4485 * We will shortly update recovery_offset to make
4488 set_bit(R5_Insync, &dev->flags);
4490 if (test_bit(R5_WriteError, &dev->flags)) {
4491 /* This flag does not apply to '.replacement'
4492 * only to .rdev, so make sure to check that*/
4493 struct md_rdev *rdev2 = rcu_dereference(
4494 conf->disks[i].rdev);
4496 clear_bit(R5_Insync, &dev->flags);
4497 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4498 s->handle_bad_blocks = 1;
4499 atomic_inc(&rdev2->nr_pending);
4501 clear_bit(R5_WriteError, &dev->flags);
4503 if (test_bit(R5_MadeGood, &dev->flags)) {
4504 /* This flag does not apply to '.replacement'
4505 * only to .rdev, so make sure to check that*/
4506 struct md_rdev *rdev2 = rcu_dereference(
4507 conf->disks[i].rdev);
4508 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4509 s->handle_bad_blocks = 1;
4510 atomic_inc(&rdev2->nr_pending);
4512 clear_bit(R5_MadeGood, &dev->flags);
4514 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4515 struct md_rdev *rdev2 = rcu_dereference(
4516 conf->disks[i].replacement);
4517 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4518 s->handle_bad_blocks = 1;
4519 atomic_inc(&rdev2->nr_pending);
4521 clear_bit(R5_MadeGoodRepl, &dev->flags);
4523 if (!test_bit(R5_Insync, &dev->flags)) {
4524 /* The ReadError flag will just be confusing now */
4525 clear_bit(R5_ReadError, &dev->flags);
4526 clear_bit(R5_ReWrite, &dev->flags);
4528 if (test_bit(R5_ReadError, &dev->flags))
4529 clear_bit(R5_Insync, &dev->flags);
4530 if (!test_bit(R5_Insync, &dev->flags)) {
4532 s->failed_num[s->failed] = i;
4534 if (rdev && !test_bit(Faulty, &rdev->flags))
4537 rdev = rcu_dereference(
4538 conf->disks[i].replacement);
4539 if (rdev && !test_bit(Faulty, &rdev->flags))
4544 if (test_bit(R5_InJournal, &dev->flags))
4546 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4549 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4550 /* If there is a failed device being replaced,
4551 * we must be recovering.
4552 * else if we are after recovery_cp, we must be syncing
4553 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4554 * else we can only be replacing
4555 * sync and recovery both need to read all devices, and so
4556 * use the same flag.
4559 sh->sector >= conf->mddev->recovery_cp ||
4560 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4568 static int clear_batch_ready(struct stripe_head *sh)
4570 /* Return '1' if this is a member of batch, or
4571 * '0' if it is a lone stripe or a head which can now be
4574 struct stripe_head *tmp;
4575 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4576 return (sh->batch_head && sh->batch_head != sh);
4577 spin_lock(&sh->stripe_lock);
4578 if (!sh->batch_head) {
4579 spin_unlock(&sh->stripe_lock);
4584 * this stripe could be added to a batch list before we check
4585 * BATCH_READY, skips it
4587 if (sh->batch_head != sh) {
4588 spin_unlock(&sh->stripe_lock);
4591 spin_lock(&sh->batch_lock);
4592 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4593 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4594 spin_unlock(&sh->batch_lock);
4595 spin_unlock(&sh->stripe_lock);
4598 * BATCH_READY is cleared, no new stripes can be added.
4599 * batch_list can be accessed without lock
4604 static void break_stripe_batch_list(struct stripe_head *head_sh,
4605 unsigned long handle_flags)
4607 struct stripe_head *sh, *next;
4611 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4613 list_del_init(&sh->batch_list);
4615 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4616 (1 << STRIPE_SYNCING) |
4617 (1 << STRIPE_REPLACED) |
4618 (1 << STRIPE_DELAYED) |
4619 (1 << STRIPE_BIT_DELAY) |
4620 (1 << STRIPE_FULL_WRITE) |
4621 (1 << STRIPE_BIOFILL_RUN) |
4622 (1 << STRIPE_COMPUTE_RUN) |
4623 (1 << STRIPE_OPS_REQ_PENDING) |
4624 (1 << STRIPE_DISCARD) |
4625 (1 << STRIPE_BATCH_READY) |
4626 (1 << STRIPE_BATCH_ERR) |
4627 (1 << STRIPE_BITMAP_PENDING)),
4628 "stripe state: %lx\n", sh->state);
4629 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4630 (1 << STRIPE_REPLACED)),
4631 "head stripe state: %lx\n", head_sh->state);
4633 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4634 (1 << STRIPE_PREREAD_ACTIVE) |
4635 (1 << STRIPE_DEGRADED) |
4636 (1 << STRIPE_ON_UNPLUG_LIST)),
4637 head_sh->state & (1 << STRIPE_INSYNC));
4639 sh->check_state = head_sh->check_state;
4640 sh->reconstruct_state = head_sh->reconstruct_state;
4641 spin_lock_irq(&sh->stripe_lock);
4642 sh->batch_head = NULL;
4643 spin_unlock_irq(&sh->stripe_lock);
4644 for (i = 0; i < sh->disks; i++) {
4645 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4647 sh->dev[i].flags = head_sh->dev[i].flags &
4648 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4650 if (handle_flags == 0 ||
4651 sh->state & handle_flags)
4652 set_bit(STRIPE_HANDLE, &sh->state);
4653 raid5_release_stripe(sh);
4655 spin_lock_irq(&head_sh->stripe_lock);
4656 head_sh->batch_head = NULL;
4657 spin_unlock_irq(&head_sh->stripe_lock);
4658 for (i = 0; i < head_sh->disks; i++)
4659 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4661 if (head_sh->state & handle_flags)
4662 set_bit(STRIPE_HANDLE, &head_sh->state);
4665 wake_up(&head_sh->raid_conf->wait_for_overlap);
4668 static void handle_stripe(struct stripe_head *sh)
4670 struct stripe_head_state s;
4671 struct r5conf *conf = sh->raid_conf;
4674 int disks = sh->disks;
4675 struct r5dev *pdev, *qdev;
4677 clear_bit(STRIPE_HANDLE, &sh->state);
4678 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4679 /* already being handled, ensure it gets handled
4680 * again when current action finishes */
4681 set_bit(STRIPE_HANDLE, &sh->state);
4685 if (clear_batch_ready(sh) ) {
4686 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4690 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4691 break_stripe_batch_list(sh, 0);
4693 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4694 spin_lock(&sh->stripe_lock);
4696 * Cannot process 'sync' concurrently with 'discard'.
4697 * Flush data in r5cache before 'sync'.
4699 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4700 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4701 !test_bit(STRIPE_DISCARD, &sh->state) &&
4702 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4703 set_bit(STRIPE_SYNCING, &sh->state);
4704 clear_bit(STRIPE_INSYNC, &sh->state);
4705 clear_bit(STRIPE_REPLACED, &sh->state);
4707 spin_unlock(&sh->stripe_lock);
4709 clear_bit(STRIPE_DELAYED, &sh->state);
4711 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4712 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4713 (unsigned long long)sh->sector, sh->state,
4714 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4715 sh->check_state, sh->reconstruct_state);
4717 analyse_stripe(sh, &s);
4719 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4722 if (s.handle_bad_blocks ||
4723 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4724 set_bit(STRIPE_HANDLE, &sh->state);
4728 if (unlikely(s.blocked_rdev)) {
4729 if (s.syncing || s.expanding || s.expanded ||
4730 s.replacing || s.to_write || s.written) {
4731 set_bit(STRIPE_HANDLE, &sh->state);
4734 /* There is nothing for the blocked_rdev to block */
4735 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4736 s.blocked_rdev = NULL;
4739 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4740 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4741 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4744 pr_debug("locked=%d uptodate=%d to_read=%d"
4745 " to_write=%d failed=%d failed_num=%d,%d\n",
4746 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4747 s.failed_num[0], s.failed_num[1]);
4749 * check if the array has lost more than max_degraded devices and,
4750 * if so, some requests might need to be failed.
4752 * When journal device failed (log_failed), we will only process
4753 * the stripe if there is data need write to raid disks
4755 if (s.failed > conf->max_degraded ||
4756 (s.log_failed && s.injournal == 0)) {
4757 sh->check_state = 0;
4758 sh->reconstruct_state = 0;
4759 break_stripe_batch_list(sh, 0);
4760 if (s.to_read+s.to_write+s.written)
4761 handle_failed_stripe(conf, sh, &s, disks);
4762 if (s.syncing + s.replacing)
4763 handle_failed_sync(conf, sh, &s);
4766 /* Now we check to see if any write operations have recently
4770 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4772 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4773 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4774 sh->reconstruct_state = reconstruct_state_idle;
4776 /* All the 'written' buffers and the parity block are ready to
4777 * be written back to disk
4779 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4780 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4781 BUG_ON(sh->qd_idx >= 0 &&
4782 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4783 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4784 for (i = disks; i--; ) {
4785 struct r5dev *dev = &sh->dev[i];
4786 if (test_bit(R5_LOCKED, &dev->flags) &&
4787 (i == sh->pd_idx || i == sh->qd_idx ||
4788 dev->written || test_bit(R5_InJournal,
4790 pr_debug("Writing block %d\n", i);
4791 set_bit(R5_Wantwrite, &dev->flags);
4796 if (!test_bit(R5_Insync, &dev->flags) ||
4797 ((i == sh->pd_idx || i == sh->qd_idx) &&
4799 set_bit(STRIPE_INSYNC, &sh->state);
4802 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4803 s.dec_preread_active = 1;
4807 * might be able to return some write requests if the parity blocks
4808 * are safe, or on a failed drive
4810 pdev = &sh->dev[sh->pd_idx];
4811 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4812 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4813 qdev = &sh->dev[sh->qd_idx];
4814 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4815 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4819 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4820 && !test_bit(R5_LOCKED, &pdev->flags)
4821 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4822 test_bit(R5_Discard, &pdev->flags))))) &&
4823 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4824 && !test_bit(R5_LOCKED, &qdev->flags)
4825 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4826 test_bit(R5_Discard, &qdev->flags))))))
4827 handle_stripe_clean_event(conf, sh, disks);
4830 r5c_handle_cached_data_endio(conf, sh, disks);
4831 log_stripe_write_finished(sh);
4833 /* Now we might consider reading some blocks, either to check/generate
4834 * parity, or to satisfy requests
4835 * or to load a block that is being partially written.
4837 if (s.to_read || s.non_overwrite
4838 || (conf->level == 6 && s.to_write && s.failed)
4839 || (s.syncing && (s.uptodate + s.compute < disks))
4842 handle_stripe_fill(sh, &s, disks);
4845 * When the stripe finishes full journal write cycle (write to journal
4846 * and raid disk), this is the clean up procedure so it is ready for
4849 r5c_finish_stripe_write_out(conf, sh, &s);
4852 * Now to consider new write requests, cache write back and what else,
4853 * if anything should be read. We do not handle new writes when:
4854 * 1/ A 'write' operation (copy+xor) is already in flight.
4855 * 2/ A 'check' operation is in flight, as it may clobber the parity
4857 * 3/ A r5c cache log write is in flight.
4860 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4861 if (!r5c_is_writeback(conf->log)) {
4863 handle_stripe_dirtying(conf, sh, &s, disks);
4864 } else { /* write back cache */
4867 /* First, try handle writes in caching phase */
4869 ret = r5c_try_caching_write(conf, sh, &s,
4872 * If caching phase failed: ret == -EAGAIN
4874 * stripe under reclaim: !caching && injournal
4876 * fall back to handle_stripe_dirtying()
4878 if (ret == -EAGAIN ||
4879 /* stripe under reclaim: !caching && injournal */
4880 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4882 ret = handle_stripe_dirtying(conf, sh, &s,
4890 /* maybe we need to check and possibly fix the parity for this stripe
4891 * Any reads will already have been scheduled, so we just see if enough
4892 * data is available. The parity check is held off while parity
4893 * dependent operations are in flight.
4895 if (sh->check_state ||
4896 (s.syncing && s.locked == 0 &&
4897 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4898 !test_bit(STRIPE_INSYNC, &sh->state))) {
4899 if (conf->level == 6)
4900 handle_parity_checks6(conf, sh, &s, disks);
4902 handle_parity_checks5(conf, sh, &s, disks);
4905 if ((s.replacing || s.syncing) && s.locked == 0
4906 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4907 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4908 /* Write out to replacement devices where possible */
4909 for (i = 0; i < conf->raid_disks; i++)
4910 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4911 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4912 set_bit(R5_WantReplace, &sh->dev[i].flags);
4913 set_bit(R5_LOCKED, &sh->dev[i].flags);
4917 set_bit(STRIPE_INSYNC, &sh->state);
4918 set_bit(STRIPE_REPLACED, &sh->state);
4920 if ((s.syncing || s.replacing) && s.locked == 0 &&
4921 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4922 test_bit(STRIPE_INSYNC, &sh->state)) {
4923 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4924 clear_bit(STRIPE_SYNCING, &sh->state);
4925 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4926 wake_up(&conf->wait_for_overlap);
4929 /* If the failed drives are just a ReadError, then we might need
4930 * to progress the repair/check process
4932 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4933 for (i = 0; i < s.failed; i++) {
4934 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4935 if (test_bit(R5_ReadError, &dev->flags)
4936 && !test_bit(R5_LOCKED, &dev->flags)
4937 && test_bit(R5_UPTODATE, &dev->flags)
4939 if (!test_bit(R5_ReWrite, &dev->flags)) {
4940 set_bit(R5_Wantwrite, &dev->flags);
4941 set_bit(R5_ReWrite, &dev->flags);
4942 set_bit(R5_LOCKED, &dev->flags);
4945 /* let's read it back */
4946 set_bit(R5_Wantread, &dev->flags);
4947 set_bit(R5_LOCKED, &dev->flags);
4953 /* Finish reconstruct operations initiated by the expansion process */
4954 if (sh->reconstruct_state == reconstruct_state_result) {
4955 struct stripe_head *sh_src
4956 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4957 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4958 /* sh cannot be written until sh_src has been read.
4959 * so arrange for sh to be delayed a little
4961 set_bit(STRIPE_DELAYED, &sh->state);
4962 set_bit(STRIPE_HANDLE, &sh->state);
4963 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4965 atomic_inc(&conf->preread_active_stripes);
4966 raid5_release_stripe(sh_src);
4970 raid5_release_stripe(sh_src);
4972 sh->reconstruct_state = reconstruct_state_idle;
4973 clear_bit(STRIPE_EXPANDING, &sh->state);
4974 for (i = conf->raid_disks; i--; ) {
4975 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4976 set_bit(R5_LOCKED, &sh->dev[i].flags);
4981 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4982 !sh->reconstruct_state) {
4983 /* Need to write out all blocks after computing parity */
4984 sh->disks = conf->raid_disks;
4985 stripe_set_idx(sh->sector, conf, 0, sh);
4986 schedule_reconstruction(sh, &s, 1, 1);
4987 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4988 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4989 atomic_dec(&conf->reshape_stripes);
4990 wake_up(&conf->wait_for_overlap);
4991 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4994 if (s.expanding && s.locked == 0 &&
4995 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4996 handle_stripe_expansion(conf, sh);
4999 /* wait for this device to become unblocked */
5000 if (unlikely(s.blocked_rdev)) {
5001 if (conf->mddev->external)
5002 md_wait_for_blocked_rdev(s.blocked_rdev,
5005 /* Internal metadata will immediately
5006 * be written by raid5d, so we don't
5007 * need to wait here.
5009 rdev_dec_pending(s.blocked_rdev,
5013 if (s.handle_bad_blocks)
5014 for (i = disks; i--; ) {
5015 struct md_rdev *rdev;
5016 struct r5dev *dev = &sh->dev[i];
5017 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5018 /* We own a safe reference to the rdev */
5019 rdev = conf->disks[i].rdev;
5020 if (!rdev_set_badblocks(rdev, sh->sector,
5022 md_error(conf->mddev, rdev);
5023 rdev_dec_pending(rdev, conf->mddev);
5025 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5026 rdev = conf->disks[i].rdev;
5027 rdev_clear_badblocks(rdev, sh->sector,
5029 rdev_dec_pending(rdev, conf->mddev);
5031 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5032 rdev = conf->disks[i].replacement;
5034 /* rdev have been moved down */
5035 rdev = conf->disks[i].rdev;
5036 rdev_clear_badblocks(rdev, sh->sector,
5038 rdev_dec_pending(rdev, conf->mddev);
5043 raid_run_ops(sh, s.ops_request);
5047 if (s.dec_preread_active) {
5048 /* We delay this until after ops_run_io so that if make_request
5049 * is waiting on a flush, it won't continue until the writes
5050 * have actually been submitted.
5052 atomic_dec(&conf->preread_active_stripes);
5053 if (atomic_read(&conf->preread_active_stripes) <
5055 md_wakeup_thread(conf->mddev->thread);
5058 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5061 static void raid5_activate_delayed(struct r5conf *conf)
5063 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5064 while (!list_empty(&conf->delayed_list)) {
5065 struct list_head *l = conf->delayed_list.next;
5066 struct stripe_head *sh;
5067 sh = list_entry(l, struct stripe_head, lru);
5069 clear_bit(STRIPE_DELAYED, &sh->state);
5070 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5071 atomic_inc(&conf->preread_active_stripes);
5072 list_add_tail(&sh->lru, &conf->hold_list);
5073 raid5_wakeup_stripe_thread(sh);
5078 static void activate_bit_delay(struct r5conf *conf,
5079 struct list_head *temp_inactive_list)
5081 /* device_lock is held */
5082 struct list_head head;
5083 list_add(&head, &conf->bitmap_list);
5084 list_del_init(&conf->bitmap_list);
5085 while (!list_empty(&head)) {
5086 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5088 list_del_init(&sh->lru);
5089 atomic_inc(&sh->count);
5090 hash = sh->hash_lock_index;
5091 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5095 static int raid5_congested(struct mddev *mddev, int bits)
5097 struct r5conf *conf = mddev->private;
5099 /* No difference between reads and writes. Just check
5100 * how busy the stripe_cache is
5103 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5106 /* Also checks whether there is pressure on r5cache log space */
5107 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5111 if (atomic_read(&conf->empty_inactive_list_nr))
5117 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5119 struct r5conf *conf = mddev->private;
5120 sector_t sector = bio->bi_iter.bi_sector;
5121 unsigned int chunk_sectors;
5122 unsigned int bio_sectors = bio_sectors(bio);
5124 WARN_ON_ONCE(bio->bi_partno);
5126 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5127 return chunk_sectors >=
5128 ((sector & (chunk_sectors - 1)) + bio_sectors);
5132 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5133 * later sampled by raid5d.
5135 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5137 unsigned long flags;
5139 spin_lock_irqsave(&conf->device_lock, flags);
5141 bi->bi_next = conf->retry_read_aligned_list;
5142 conf->retry_read_aligned_list = bi;
5144 spin_unlock_irqrestore(&conf->device_lock, flags);
5145 md_wakeup_thread(conf->mddev->thread);
5148 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5149 unsigned int *offset)
5153 bi = conf->retry_read_aligned;
5155 *offset = conf->retry_read_offset;
5156 conf->retry_read_aligned = NULL;
5159 bi = conf->retry_read_aligned_list;
5161 conf->retry_read_aligned_list = bi->bi_next;
5170 * The "raid5_align_endio" should check if the read succeeded and if it
5171 * did, call bio_endio on the original bio (having bio_put the new bio
5173 * If the read failed..
5175 static void raid5_align_endio(struct bio *bi)
5177 struct bio* raid_bi = bi->bi_private;
5178 struct mddev *mddev;
5179 struct r5conf *conf;
5180 struct md_rdev *rdev;
5181 blk_status_t error = bi->bi_status;
5185 rdev = (void*)raid_bi->bi_next;
5186 raid_bi->bi_next = NULL;
5187 mddev = rdev->mddev;
5188 conf = mddev->private;
5190 rdev_dec_pending(rdev, conf->mddev);
5194 if (atomic_dec_and_test(&conf->active_aligned_reads))
5195 wake_up(&conf->wait_for_quiescent);
5199 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5201 add_bio_to_retry(raid_bi, conf);
5204 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5206 struct r5conf *conf = mddev->private;
5208 struct bio* align_bi;
5209 struct md_rdev *rdev;
5210 sector_t end_sector;
5212 if (!in_chunk_boundary(mddev, raid_bio)) {
5213 pr_debug("%s: non aligned\n", __func__);
5217 * use bio_clone_fast to make a copy of the bio
5219 align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5223 * set bi_end_io to a new function, and set bi_private to the
5226 align_bi->bi_end_io = raid5_align_endio;
5227 align_bi->bi_private = raid_bio;
5231 align_bi->bi_iter.bi_sector =
5232 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5235 end_sector = bio_end_sector(align_bi);
5237 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5238 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5239 rdev->recovery_offset < end_sector) {
5240 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5242 (test_bit(Faulty, &rdev->flags) ||
5243 !(test_bit(In_sync, &rdev->flags) ||
5244 rdev->recovery_offset >= end_sector)))
5248 if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5258 atomic_inc(&rdev->nr_pending);
5260 raid_bio->bi_next = (void*)rdev;
5261 bio_set_dev(align_bi, rdev->bdev);
5262 bio_clear_flag(align_bi, BIO_SEG_VALID);
5264 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5265 bio_sectors(align_bi),
5266 &first_bad, &bad_sectors)) {
5268 rdev_dec_pending(rdev, mddev);
5272 /* No reshape active, so we can trust rdev->data_offset */
5273 align_bi->bi_iter.bi_sector += rdev->data_offset;
5275 spin_lock_irq(&conf->device_lock);
5276 wait_event_lock_irq(conf->wait_for_quiescent,
5279 atomic_inc(&conf->active_aligned_reads);
5280 spin_unlock_irq(&conf->device_lock);
5283 trace_block_bio_remap(align_bi->bi_disk->queue,
5284 align_bi, disk_devt(mddev->gendisk),
5285 raid_bio->bi_iter.bi_sector);
5286 generic_make_request(align_bi);
5295 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5298 sector_t sector = raid_bio->bi_iter.bi_sector;
5299 unsigned chunk_sects = mddev->chunk_sectors;
5300 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5302 if (sectors < bio_sectors(raid_bio)) {
5303 struct r5conf *conf = mddev->private;
5304 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5305 bio_chain(split, raid_bio);
5306 generic_make_request(raid_bio);
5310 if (!raid5_read_one_chunk(mddev, raid_bio))
5316 /* __get_priority_stripe - get the next stripe to process
5318 * Full stripe writes are allowed to pass preread active stripes up until
5319 * the bypass_threshold is exceeded. In general the bypass_count
5320 * increments when the handle_list is handled before the hold_list; however, it
5321 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5322 * stripe with in flight i/o. The bypass_count will be reset when the
5323 * head of the hold_list has changed, i.e. the head was promoted to the
5326 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5328 struct stripe_head *sh, *tmp;
5329 struct list_head *handle_list = NULL;
5330 struct r5worker_group *wg;
5331 bool second_try = !r5c_is_writeback(conf->log) &&
5332 !r5l_log_disk_error(conf);
5333 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5334 r5l_log_disk_error(conf);
5339 if (conf->worker_cnt_per_group == 0) {
5340 handle_list = try_loprio ? &conf->loprio_list :
5342 } else if (group != ANY_GROUP) {
5343 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5344 &conf->worker_groups[group].handle_list;
5345 wg = &conf->worker_groups[group];
5348 for (i = 0; i < conf->group_cnt; i++) {
5349 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5350 &conf->worker_groups[i].handle_list;
5351 wg = &conf->worker_groups[i];
5352 if (!list_empty(handle_list))
5357 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5359 list_empty(handle_list) ? "empty" : "busy",
5360 list_empty(&conf->hold_list) ? "empty" : "busy",
5361 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5363 if (!list_empty(handle_list)) {
5364 sh = list_entry(handle_list->next, typeof(*sh), lru);
5366 if (list_empty(&conf->hold_list))
5367 conf->bypass_count = 0;
5368 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5369 if (conf->hold_list.next == conf->last_hold)
5370 conf->bypass_count++;
5372 conf->last_hold = conf->hold_list.next;
5373 conf->bypass_count -= conf->bypass_threshold;
5374 if (conf->bypass_count < 0)
5375 conf->bypass_count = 0;
5378 } else if (!list_empty(&conf->hold_list) &&
5379 ((conf->bypass_threshold &&
5380 conf->bypass_count > conf->bypass_threshold) ||
5381 atomic_read(&conf->pending_full_writes) == 0)) {
5383 list_for_each_entry(tmp, &conf->hold_list, lru) {
5384 if (conf->worker_cnt_per_group == 0 ||
5385 group == ANY_GROUP ||
5386 !cpu_online(tmp->cpu) ||
5387 cpu_to_group(tmp->cpu) == group) {
5394 conf->bypass_count -= conf->bypass_threshold;
5395 if (conf->bypass_count < 0)
5396 conf->bypass_count = 0;
5405 try_loprio = !try_loprio;
5413 list_del_init(&sh->lru);
5414 BUG_ON(atomic_inc_return(&sh->count) != 1);
5418 struct raid5_plug_cb {
5419 struct blk_plug_cb cb;
5420 struct list_head list;
5421 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5424 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5426 struct raid5_plug_cb *cb = container_of(
5427 blk_cb, struct raid5_plug_cb, cb);
5428 struct stripe_head *sh;
5429 struct mddev *mddev = cb->cb.data;
5430 struct r5conf *conf = mddev->private;
5434 if (cb->list.next && !list_empty(&cb->list)) {
5435 spin_lock_irq(&conf->device_lock);
5436 while (!list_empty(&cb->list)) {
5437 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5438 list_del_init(&sh->lru);
5440 * avoid race release_stripe_plug() sees
5441 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5442 * is still in our list
5444 smp_mb__before_atomic();
5445 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5447 * STRIPE_ON_RELEASE_LIST could be set here. In that
5448 * case, the count is always > 1 here
5450 hash = sh->hash_lock_index;
5451 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5454 spin_unlock_irq(&conf->device_lock);
5456 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5457 NR_STRIPE_HASH_LOCKS);
5459 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5463 static void release_stripe_plug(struct mddev *mddev,
5464 struct stripe_head *sh)
5466 struct blk_plug_cb *blk_cb = blk_check_plugged(
5467 raid5_unplug, mddev,
5468 sizeof(struct raid5_plug_cb));
5469 struct raid5_plug_cb *cb;
5472 raid5_release_stripe(sh);
5476 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5478 if (cb->list.next == NULL) {
5480 INIT_LIST_HEAD(&cb->list);
5481 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5482 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5485 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5486 list_add_tail(&sh->lru, &cb->list);
5488 raid5_release_stripe(sh);
5491 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5493 struct r5conf *conf = mddev->private;
5494 sector_t logical_sector, last_sector;
5495 struct stripe_head *sh;
5498 if (mddev->reshape_position != MaxSector)
5499 /* Skip discard while reshape is happening */
5502 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5503 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5507 stripe_sectors = conf->chunk_sectors *
5508 (conf->raid_disks - conf->max_degraded);
5509 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5511 sector_div(last_sector, stripe_sectors);
5513 logical_sector *= conf->chunk_sectors;
5514 last_sector *= conf->chunk_sectors;
5516 for (; logical_sector < last_sector;
5517 logical_sector += STRIPE_SECTORS) {
5521 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5522 prepare_to_wait(&conf->wait_for_overlap, &w,
5523 TASK_UNINTERRUPTIBLE);
5524 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5525 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5526 raid5_release_stripe(sh);
5530 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5531 spin_lock_irq(&sh->stripe_lock);
5532 for (d = 0; d < conf->raid_disks; d++) {
5533 if (d == sh->pd_idx || d == sh->qd_idx)
5535 if (sh->dev[d].towrite || sh->dev[d].toread) {
5536 set_bit(R5_Overlap, &sh->dev[d].flags);
5537 spin_unlock_irq(&sh->stripe_lock);
5538 raid5_release_stripe(sh);
5543 set_bit(STRIPE_DISCARD, &sh->state);
5544 finish_wait(&conf->wait_for_overlap, &w);
5545 sh->overwrite_disks = 0;
5546 for (d = 0; d < conf->raid_disks; d++) {
5547 if (d == sh->pd_idx || d == sh->qd_idx)
5549 sh->dev[d].towrite = bi;
5550 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5551 bio_inc_remaining(bi);
5552 md_write_inc(mddev, bi);
5553 sh->overwrite_disks++;
5555 spin_unlock_irq(&sh->stripe_lock);
5556 if (conf->mddev->bitmap) {
5558 d < conf->raid_disks - conf->max_degraded;
5560 md_bitmap_startwrite(mddev->bitmap,
5564 sh->bm_seq = conf->seq_flush + 1;
5565 set_bit(STRIPE_BIT_DELAY, &sh->state);
5568 set_bit(STRIPE_HANDLE, &sh->state);
5569 clear_bit(STRIPE_DELAYED, &sh->state);
5570 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5571 atomic_inc(&conf->preread_active_stripes);
5572 release_stripe_plug(mddev, sh);
5578 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5580 struct r5conf *conf = mddev->private;
5582 sector_t new_sector;
5583 sector_t logical_sector, last_sector;
5584 struct stripe_head *sh;
5585 const int rw = bio_data_dir(bi);
5588 bool do_flush = false;
5590 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5591 int ret = log_handle_flush_request(conf, bi);
5595 if (ret == -ENODEV) {
5596 md_flush_request(mddev, bi);
5599 /* ret == -EAGAIN, fallback */
5601 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5602 * we need to flush journal device
5604 do_flush = bi->bi_opf & REQ_PREFLUSH;
5607 if (!md_write_start(mddev, bi))
5610 * If array is degraded, better not do chunk aligned read because
5611 * later we might have to read it again in order to reconstruct
5612 * data on failed drives.
5614 if (rw == READ && mddev->degraded == 0 &&
5615 mddev->reshape_position == MaxSector) {
5616 bi = chunk_aligned_read(mddev, bi);
5621 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5622 make_discard_request(mddev, bi);
5623 md_write_end(mddev);
5627 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5628 last_sector = bio_end_sector(bi);
5631 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5632 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5638 seq = read_seqcount_begin(&conf->gen_lock);
5641 prepare_to_wait(&conf->wait_for_overlap, &w,
5642 TASK_UNINTERRUPTIBLE);
5643 if (unlikely(conf->reshape_progress != MaxSector)) {
5644 /* spinlock is needed as reshape_progress may be
5645 * 64bit on a 32bit platform, and so it might be
5646 * possible to see a half-updated value
5647 * Of course reshape_progress could change after
5648 * the lock is dropped, so once we get a reference
5649 * to the stripe that we think it is, we will have
5652 spin_lock_irq(&conf->device_lock);
5653 if (mddev->reshape_backwards
5654 ? logical_sector < conf->reshape_progress
5655 : logical_sector >= conf->reshape_progress) {
5658 if (mddev->reshape_backwards
5659 ? logical_sector < conf->reshape_safe
5660 : logical_sector >= conf->reshape_safe) {
5661 spin_unlock_irq(&conf->device_lock);
5667 spin_unlock_irq(&conf->device_lock);
5670 new_sector = raid5_compute_sector(conf, logical_sector,
5673 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5674 (unsigned long long)new_sector,
5675 (unsigned long long)logical_sector);
5677 sh = raid5_get_active_stripe(conf, new_sector, previous,
5678 (bi->bi_opf & REQ_RAHEAD), 0);
5680 if (unlikely(previous)) {
5681 /* expansion might have moved on while waiting for a
5682 * stripe, so we must do the range check again.
5683 * Expansion could still move past after this
5684 * test, but as we are holding a reference to
5685 * 'sh', we know that if that happens,
5686 * STRIPE_EXPANDING will get set and the expansion
5687 * won't proceed until we finish with the stripe.
5690 spin_lock_irq(&conf->device_lock);
5691 if (mddev->reshape_backwards
5692 ? logical_sector >= conf->reshape_progress
5693 : logical_sector < conf->reshape_progress)
5694 /* mismatch, need to try again */
5696 spin_unlock_irq(&conf->device_lock);
5698 raid5_release_stripe(sh);
5704 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5705 /* Might have got the wrong stripe_head
5708 raid5_release_stripe(sh);
5712 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5713 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5714 /* Stripe is busy expanding or
5715 * add failed due to overlap. Flush everything
5718 md_wakeup_thread(mddev->thread);
5719 raid5_release_stripe(sh);
5725 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5726 /* we only need flush for one stripe */
5730 set_bit(STRIPE_HANDLE, &sh->state);
5731 clear_bit(STRIPE_DELAYED, &sh->state);
5732 if ((!sh->batch_head || sh == sh->batch_head) &&
5733 (bi->bi_opf & REQ_SYNC) &&
5734 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5735 atomic_inc(&conf->preread_active_stripes);
5736 release_stripe_plug(mddev, sh);
5738 /* cannot get stripe for read-ahead, just give-up */
5739 bi->bi_status = BLK_STS_IOERR;
5743 finish_wait(&conf->wait_for_overlap, &w);
5746 md_write_end(mddev);
5751 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5753 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5755 /* reshaping is quite different to recovery/resync so it is
5756 * handled quite separately ... here.
5758 * On each call to sync_request, we gather one chunk worth of
5759 * destination stripes and flag them as expanding.
5760 * Then we find all the source stripes and request reads.
5761 * As the reads complete, handle_stripe will copy the data
5762 * into the destination stripe and release that stripe.
5764 struct r5conf *conf = mddev->private;
5765 struct stripe_head *sh;
5766 struct md_rdev *rdev;
5767 sector_t first_sector, last_sector;
5768 int raid_disks = conf->previous_raid_disks;
5769 int data_disks = raid_disks - conf->max_degraded;
5770 int new_data_disks = conf->raid_disks - conf->max_degraded;
5773 sector_t writepos, readpos, safepos;
5774 sector_t stripe_addr;
5775 int reshape_sectors;
5776 struct list_head stripes;
5779 if (sector_nr == 0) {
5780 /* If restarting in the middle, skip the initial sectors */
5781 if (mddev->reshape_backwards &&
5782 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5783 sector_nr = raid5_size(mddev, 0, 0)
5784 - conf->reshape_progress;
5785 } else if (mddev->reshape_backwards &&
5786 conf->reshape_progress == MaxSector) {
5787 /* shouldn't happen, but just in case, finish up.*/
5788 sector_nr = MaxSector;
5789 } else if (!mddev->reshape_backwards &&
5790 conf->reshape_progress > 0)
5791 sector_nr = conf->reshape_progress;
5792 sector_div(sector_nr, new_data_disks);
5794 mddev->curr_resync_completed = sector_nr;
5795 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5802 /* We need to process a full chunk at a time.
5803 * If old and new chunk sizes differ, we need to process the
5807 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5809 /* We update the metadata at least every 10 seconds, or when
5810 * the data about to be copied would over-write the source of
5811 * the data at the front of the range. i.e. one new_stripe
5812 * along from reshape_progress new_maps to after where
5813 * reshape_safe old_maps to
5815 writepos = conf->reshape_progress;
5816 sector_div(writepos, new_data_disks);
5817 readpos = conf->reshape_progress;
5818 sector_div(readpos, data_disks);
5819 safepos = conf->reshape_safe;
5820 sector_div(safepos, data_disks);
5821 if (mddev->reshape_backwards) {
5822 BUG_ON(writepos < reshape_sectors);
5823 writepos -= reshape_sectors;
5824 readpos += reshape_sectors;
5825 safepos += reshape_sectors;
5827 writepos += reshape_sectors;
5828 /* readpos and safepos are worst-case calculations.
5829 * A negative number is overly pessimistic, and causes
5830 * obvious problems for unsigned storage. So clip to 0.
5832 readpos -= min_t(sector_t, reshape_sectors, readpos);
5833 safepos -= min_t(sector_t, reshape_sectors, safepos);
5836 /* Having calculated the 'writepos' possibly use it
5837 * to set 'stripe_addr' which is where we will write to.
5839 if (mddev->reshape_backwards) {
5840 BUG_ON(conf->reshape_progress == 0);
5841 stripe_addr = writepos;
5842 BUG_ON((mddev->dev_sectors &
5843 ~((sector_t)reshape_sectors - 1))
5844 - reshape_sectors - stripe_addr
5847 BUG_ON(writepos != sector_nr + reshape_sectors);
5848 stripe_addr = sector_nr;
5851 /* 'writepos' is the most advanced device address we might write.
5852 * 'readpos' is the least advanced device address we might read.
5853 * 'safepos' is the least address recorded in the metadata as having
5855 * If there is a min_offset_diff, these are adjusted either by
5856 * increasing the safepos/readpos if diff is negative, or
5857 * increasing writepos if diff is positive.
5858 * If 'readpos' is then behind 'writepos', there is no way that we can
5859 * ensure safety in the face of a crash - that must be done by userspace
5860 * making a backup of the data. So in that case there is no particular
5861 * rush to update metadata.
5862 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5863 * update the metadata to advance 'safepos' to match 'readpos' so that
5864 * we can be safe in the event of a crash.
5865 * So we insist on updating metadata if safepos is behind writepos and
5866 * readpos is beyond writepos.
5867 * In any case, update the metadata every 10 seconds.
5868 * Maybe that number should be configurable, but I'm not sure it is
5869 * worth it.... maybe it could be a multiple of safemode_delay???
5871 if (conf->min_offset_diff < 0) {
5872 safepos += -conf->min_offset_diff;
5873 readpos += -conf->min_offset_diff;
5875 writepos += conf->min_offset_diff;
5877 if ((mddev->reshape_backwards
5878 ? (safepos > writepos && readpos < writepos)
5879 : (safepos < writepos && readpos > writepos)) ||
5880 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5881 /* Cannot proceed until we've updated the superblock... */
5882 wait_event(conf->wait_for_overlap,
5883 atomic_read(&conf->reshape_stripes)==0
5884 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5885 if (atomic_read(&conf->reshape_stripes) != 0)
5887 mddev->reshape_position = conf->reshape_progress;
5888 mddev->curr_resync_completed = sector_nr;
5889 if (!mddev->reshape_backwards)
5890 /* Can update recovery_offset */
5891 rdev_for_each(rdev, mddev)
5892 if (rdev->raid_disk >= 0 &&
5893 !test_bit(Journal, &rdev->flags) &&
5894 !test_bit(In_sync, &rdev->flags) &&
5895 rdev->recovery_offset < sector_nr)
5896 rdev->recovery_offset = sector_nr;
5898 conf->reshape_checkpoint = jiffies;
5899 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5900 md_wakeup_thread(mddev->thread);
5901 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5902 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5903 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5905 spin_lock_irq(&conf->device_lock);
5906 conf->reshape_safe = mddev->reshape_position;
5907 spin_unlock_irq(&conf->device_lock);
5908 wake_up(&conf->wait_for_overlap);
5909 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5912 INIT_LIST_HEAD(&stripes);
5913 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5915 int skipped_disk = 0;
5916 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5917 set_bit(STRIPE_EXPANDING, &sh->state);
5918 atomic_inc(&conf->reshape_stripes);
5919 /* If any of this stripe is beyond the end of the old
5920 * array, then we need to zero those blocks
5922 for (j=sh->disks; j--;) {
5924 if (j == sh->pd_idx)
5926 if (conf->level == 6 &&
5929 s = raid5_compute_blocknr(sh, j, 0);
5930 if (s < raid5_size(mddev, 0, 0)) {
5934 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5935 set_bit(R5_Expanded, &sh->dev[j].flags);
5936 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5938 if (!skipped_disk) {
5939 set_bit(STRIPE_EXPAND_READY, &sh->state);
5940 set_bit(STRIPE_HANDLE, &sh->state);
5942 list_add(&sh->lru, &stripes);
5944 spin_lock_irq(&conf->device_lock);
5945 if (mddev->reshape_backwards)
5946 conf->reshape_progress -= reshape_sectors * new_data_disks;
5948 conf->reshape_progress += reshape_sectors * new_data_disks;
5949 spin_unlock_irq(&conf->device_lock);
5950 /* Ok, those stripe are ready. We can start scheduling
5951 * reads on the source stripes.
5952 * The source stripes are determined by mapping the first and last
5953 * block on the destination stripes.
5956 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5959 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5960 * new_data_disks - 1),
5962 if (last_sector >= mddev->dev_sectors)
5963 last_sector = mddev->dev_sectors - 1;
5964 while (first_sector <= last_sector) {
5965 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5966 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5967 set_bit(STRIPE_HANDLE, &sh->state);
5968 raid5_release_stripe(sh);
5969 first_sector += STRIPE_SECTORS;
5971 /* Now that the sources are clearly marked, we can release
5972 * the destination stripes
5974 while (!list_empty(&stripes)) {
5975 sh = list_entry(stripes.next, struct stripe_head, lru);
5976 list_del_init(&sh->lru);
5977 raid5_release_stripe(sh);
5979 /* If this takes us to the resync_max point where we have to pause,
5980 * then we need to write out the superblock.
5982 sector_nr += reshape_sectors;
5983 retn = reshape_sectors;
5985 if (mddev->curr_resync_completed > mddev->resync_max ||
5986 (sector_nr - mddev->curr_resync_completed) * 2
5987 >= mddev->resync_max - mddev->curr_resync_completed) {
5988 /* Cannot proceed until we've updated the superblock... */
5989 wait_event(conf->wait_for_overlap,
5990 atomic_read(&conf->reshape_stripes) == 0
5991 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5992 if (atomic_read(&conf->reshape_stripes) != 0)
5994 mddev->reshape_position = conf->reshape_progress;
5995 mddev->curr_resync_completed = sector_nr;
5996 if (!mddev->reshape_backwards)
5997 /* Can update recovery_offset */
5998 rdev_for_each(rdev, mddev)
5999 if (rdev->raid_disk >= 0 &&
6000 !test_bit(Journal, &rdev->flags) &&
6001 !test_bit(In_sync, &rdev->flags) &&
6002 rdev->recovery_offset < sector_nr)
6003 rdev->recovery_offset = sector_nr;
6004 conf->reshape_checkpoint = jiffies;
6005 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6006 md_wakeup_thread(mddev->thread);
6007 wait_event(mddev->sb_wait,
6008 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6009 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6010 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6012 spin_lock_irq(&conf->device_lock);
6013 conf->reshape_safe = mddev->reshape_position;
6014 spin_unlock_irq(&conf->device_lock);
6015 wake_up(&conf->wait_for_overlap);
6016 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6022 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6025 struct r5conf *conf = mddev->private;
6026 struct stripe_head *sh;
6027 sector_t max_sector = mddev->dev_sectors;
6028 sector_t sync_blocks;
6029 int still_degraded = 0;
6032 if (sector_nr >= max_sector) {
6033 /* just being told to finish up .. nothing much to do */
6035 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6040 if (mddev->curr_resync < max_sector) /* aborted */
6041 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6043 else /* completed sync */
6045 md_bitmap_close_sync(mddev->bitmap);
6050 /* Allow raid5_quiesce to complete */
6051 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6053 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6054 return reshape_request(mddev, sector_nr, skipped);
6056 /* No need to check resync_max as we never do more than one
6057 * stripe, and as resync_max will always be on a chunk boundary,
6058 * if the check in md_do_sync didn't fire, there is no chance
6059 * of overstepping resync_max here
6062 /* if there is too many failed drives and we are trying
6063 * to resync, then assert that we are finished, because there is
6064 * nothing we can do.
6066 if (mddev->degraded >= conf->max_degraded &&
6067 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6068 sector_t rv = mddev->dev_sectors - sector_nr;
6072 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6074 !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6075 sync_blocks >= STRIPE_SECTORS) {
6076 /* we can skip this block, and probably more */
6077 sync_blocks /= STRIPE_SECTORS;
6079 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6082 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6084 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6086 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6087 /* make sure we don't swamp the stripe cache if someone else
6088 * is trying to get access
6090 schedule_timeout_uninterruptible(1);
6092 /* Need to check if array will still be degraded after recovery/resync
6093 * Note in case of > 1 drive failures it's possible we're rebuilding
6094 * one drive while leaving another faulty drive in array.
6097 for (i = 0; i < conf->raid_disks; i++) {
6098 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6100 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6105 md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6107 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6108 set_bit(STRIPE_HANDLE, &sh->state);
6110 raid5_release_stripe(sh);
6112 return STRIPE_SECTORS;
6115 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6116 unsigned int offset)
6118 /* We may not be able to submit a whole bio at once as there
6119 * may not be enough stripe_heads available.
6120 * We cannot pre-allocate enough stripe_heads as we may need
6121 * more than exist in the cache (if we allow ever large chunks).
6122 * So we do one stripe head at a time and record in
6123 * ->bi_hw_segments how many have been done.
6125 * We *know* that this entire raid_bio is in one chunk, so
6126 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6128 struct stripe_head *sh;
6130 sector_t sector, logical_sector, last_sector;
6134 logical_sector = raid_bio->bi_iter.bi_sector &
6135 ~((sector_t)STRIPE_SECTORS-1);
6136 sector = raid5_compute_sector(conf, logical_sector,
6138 last_sector = bio_end_sector(raid_bio);
6140 for (; logical_sector < last_sector;
6141 logical_sector += STRIPE_SECTORS,
6142 sector += STRIPE_SECTORS,
6146 /* already done this stripe */
6149 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6152 /* failed to get a stripe - must wait */
6153 conf->retry_read_aligned = raid_bio;
6154 conf->retry_read_offset = scnt;
6158 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6159 raid5_release_stripe(sh);
6160 conf->retry_read_aligned = raid_bio;
6161 conf->retry_read_offset = scnt;
6165 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6167 raid5_release_stripe(sh);
6171 bio_endio(raid_bio);
6173 if (atomic_dec_and_test(&conf->active_aligned_reads))
6174 wake_up(&conf->wait_for_quiescent);
6178 static int handle_active_stripes(struct r5conf *conf, int group,
6179 struct r5worker *worker,
6180 struct list_head *temp_inactive_list)
6181 __releases(&conf->device_lock)
6182 __acquires(&conf->device_lock)
6184 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6185 int i, batch_size = 0, hash;
6186 bool release_inactive = false;
6188 while (batch_size < MAX_STRIPE_BATCH &&
6189 (sh = __get_priority_stripe(conf, group)) != NULL)
6190 batch[batch_size++] = sh;
6192 if (batch_size == 0) {
6193 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6194 if (!list_empty(temp_inactive_list + i))
6196 if (i == NR_STRIPE_HASH_LOCKS) {
6197 spin_unlock_irq(&conf->device_lock);
6198 log_flush_stripe_to_raid(conf);
6199 spin_lock_irq(&conf->device_lock);
6202 release_inactive = true;
6204 spin_unlock_irq(&conf->device_lock);
6206 release_inactive_stripe_list(conf, temp_inactive_list,
6207 NR_STRIPE_HASH_LOCKS);
6209 r5l_flush_stripe_to_raid(conf->log);
6210 if (release_inactive) {
6211 spin_lock_irq(&conf->device_lock);
6215 for (i = 0; i < batch_size; i++)
6216 handle_stripe(batch[i]);
6217 log_write_stripe_run(conf);
6221 spin_lock_irq(&conf->device_lock);
6222 for (i = 0; i < batch_size; i++) {
6223 hash = batch[i]->hash_lock_index;
6224 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6229 static void raid5_do_work(struct work_struct *work)
6231 struct r5worker *worker = container_of(work, struct r5worker, work);
6232 struct r5worker_group *group = worker->group;
6233 struct r5conf *conf = group->conf;
6234 struct mddev *mddev = conf->mddev;
6235 int group_id = group - conf->worker_groups;
6237 struct blk_plug plug;
6239 pr_debug("+++ raid5worker active\n");
6241 blk_start_plug(&plug);
6243 spin_lock_irq(&conf->device_lock);
6245 int batch_size, released;
6247 released = release_stripe_list(conf, worker->temp_inactive_list);
6249 batch_size = handle_active_stripes(conf, group_id, worker,
6250 worker->temp_inactive_list);
6251 worker->working = false;
6252 if (!batch_size && !released)
6254 handled += batch_size;
6255 wait_event_lock_irq(mddev->sb_wait,
6256 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6259 pr_debug("%d stripes handled\n", handled);
6261 spin_unlock_irq(&conf->device_lock);
6263 flush_deferred_bios(conf);
6265 r5l_flush_stripe_to_raid(conf->log);
6267 async_tx_issue_pending_all();
6268 blk_finish_plug(&plug);
6270 pr_debug("--- raid5worker inactive\n");
6274 * This is our raid5 kernel thread.
6276 * We scan the hash table for stripes which can be handled now.
6277 * During the scan, completed stripes are saved for us by the interrupt
6278 * handler, so that they will not have to wait for our next wakeup.
6280 static void raid5d(struct md_thread *thread)
6282 struct mddev *mddev = thread->mddev;
6283 struct r5conf *conf = mddev->private;
6285 struct blk_plug plug;
6287 pr_debug("+++ raid5d active\n");
6289 md_check_recovery(mddev);
6291 blk_start_plug(&plug);
6293 spin_lock_irq(&conf->device_lock);
6296 int batch_size, released;
6297 unsigned int offset;
6299 released = release_stripe_list(conf, conf->temp_inactive_list);
6301 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6304 !list_empty(&conf->bitmap_list)) {
6305 /* Now is a good time to flush some bitmap updates */
6307 spin_unlock_irq(&conf->device_lock);
6308 md_bitmap_unplug(mddev->bitmap);
6309 spin_lock_irq(&conf->device_lock);
6310 conf->seq_write = conf->seq_flush;
6311 activate_bit_delay(conf, conf->temp_inactive_list);
6313 raid5_activate_delayed(conf);
6315 while ((bio = remove_bio_from_retry(conf, &offset))) {
6317 spin_unlock_irq(&conf->device_lock);
6318 ok = retry_aligned_read(conf, bio, offset);
6319 spin_lock_irq(&conf->device_lock);
6325 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6326 conf->temp_inactive_list);
6327 if (!batch_size && !released)
6329 handled += batch_size;
6331 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6332 spin_unlock_irq(&conf->device_lock);
6333 md_check_recovery(mddev);
6334 spin_lock_irq(&conf->device_lock);
6337 pr_debug("%d stripes handled\n", handled);
6339 spin_unlock_irq(&conf->device_lock);
6340 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6341 mutex_trylock(&conf->cache_size_mutex)) {
6342 grow_one_stripe(conf, __GFP_NOWARN);
6343 /* Set flag even if allocation failed. This helps
6344 * slow down allocation requests when mem is short
6346 set_bit(R5_DID_ALLOC, &conf->cache_state);
6347 mutex_unlock(&conf->cache_size_mutex);
6350 flush_deferred_bios(conf);
6352 r5l_flush_stripe_to_raid(conf->log);
6354 async_tx_issue_pending_all();
6355 blk_finish_plug(&plug);
6357 pr_debug("--- raid5d inactive\n");
6361 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6363 struct r5conf *conf;
6365 spin_lock(&mddev->lock);
6366 conf = mddev->private;
6368 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6369 spin_unlock(&mddev->lock);
6374 raid5_set_cache_size(struct mddev *mddev, int size)
6377 struct r5conf *conf = mddev->private;
6379 if (size <= 16 || size > 32768)
6382 conf->min_nr_stripes = size;
6383 mutex_lock(&conf->cache_size_mutex);
6384 while (size < conf->max_nr_stripes &&
6385 drop_one_stripe(conf))
6387 mutex_unlock(&conf->cache_size_mutex);
6389 md_allow_write(mddev);
6391 mutex_lock(&conf->cache_size_mutex);
6392 while (size > conf->max_nr_stripes)
6393 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6394 conf->min_nr_stripes = conf->max_nr_stripes;
6398 mutex_unlock(&conf->cache_size_mutex);
6402 EXPORT_SYMBOL(raid5_set_cache_size);
6405 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6407 struct r5conf *conf;
6411 if (len >= PAGE_SIZE)
6413 if (kstrtoul(page, 10, &new))
6415 err = mddev_lock(mddev);
6418 conf = mddev->private;
6422 err = raid5_set_cache_size(mddev, new);
6423 mddev_unlock(mddev);
6428 static struct md_sysfs_entry
6429 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6430 raid5_show_stripe_cache_size,
6431 raid5_store_stripe_cache_size);
6434 raid5_show_rmw_level(struct mddev *mddev, char *page)
6436 struct r5conf *conf = mddev->private;
6438 return sprintf(page, "%d\n", conf->rmw_level);
6444 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6446 struct r5conf *conf = mddev->private;
6452 if (len >= PAGE_SIZE)
6455 if (kstrtoul(page, 10, &new))
6458 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6461 if (new != PARITY_DISABLE_RMW &&
6462 new != PARITY_ENABLE_RMW &&
6463 new != PARITY_PREFER_RMW)
6466 conf->rmw_level = new;
6470 static struct md_sysfs_entry
6471 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6472 raid5_show_rmw_level,
6473 raid5_store_rmw_level);
6477 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6479 struct r5conf *conf;
6481 spin_lock(&mddev->lock);
6482 conf = mddev->private;
6484 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6485 spin_unlock(&mddev->lock);
6490 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6492 struct r5conf *conf;
6496 if (len >= PAGE_SIZE)
6498 if (kstrtoul(page, 10, &new))
6501 err = mddev_lock(mddev);
6504 conf = mddev->private;
6507 else if (new > conf->min_nr_stripes)
6510 conf->bypass_threshold = new;
6511 mddev_unlock(mddev);
6515 static struct md_sysfs_entry
6516 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6518 raid5_show_preread_threshold,
6519 raid5_store_preread_threshold);
6522 raid5_show_skip_copy(struct mddev *mddev, char *page)
6524 struct r5conf *conf;
6526 spin_lock(&mddev->lock);
6527 conf = mddev->private;
6529 ret = sprintf(page, "%d\n", conf->skip_copy);
6530 spin_unlock(&mddev->lock);
6535 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6537 struct r5conf *conf;
6541 if (len >= PAGE_SIZE)
6543 if (kstrtoul(page, 10, &new))
6547 err = mddev_lock(mddev);
6550 conf = mddev->private;
6553 else if (new != conf->skip_copy) {
6554 mddev_suspend(mddev);
6555 conf->skip_copy = new;
6557 mddev->queue->backing_dev_info->capabilities |=
6558 BDI_CAP_STABLE_WRITES;
6560 mddev->queue->backing_dev_info->capabilities &=
6561 ~BDI_CAP_STABLE_WRITES;
6562 mddev_resume(mddev);
6564 mddev_unlock(mddev);
6568 static struct md_sysfs_entry
6569 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6570 raid5_show_skip_copy,
6571 raid5_store_skip_copy);
6574 stripe_cache_active_show(struct mddev *mddev, char *page)
6576 struct r5conf *conf = mddev->private;
6578 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6583 static struct md_sysfs_entry
6584 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6587 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6589 struct r5conf *conf;
6591 spin_lock(&mddev->lock);
6592 conf = mddev->private;
6594 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6595 spin_unlock(&mddev->lock);
6599 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6601 int *worker_cnt_per_group,
6602 struct r5worker_group **worker_groups);
6604 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6606 struct r5conf *conf;
6609 struct r5worker_group *new_groups, *old_groups;
6610 int group_cnt, worker_cnt_per_group;
6612 if (len >= PAGE_SIZE)
6614 if (kstrtouint(page, 10, &new))
6616 /* 8192 should be big enough */
6620 err = mddev_lock(mddev);
6623 conf = mddev->private;
6626 else if (new != conf->worker_cnt_per_group) {
6627 mddev_suspend(mddev);
6629 old_groups = conf->worker_groups;
6631 flush_workqueue(raid5_wq);
6633 err = alloc_thread_groups(conf, new,
6634 &group_cnt, &worker_cnt_per_group,
6637 spin_lock_irq(&conf->device_lock);
6638 conf->group_cnt = group_cnt;
6639 conf->worker_cnt_per_group = worker_cnt_per_group;
6640 conf->worker_groups = new_groups;
6641 spin_unlock_irq(&conf->device_lock);
6644 kfree(old_groups[0].workers);
6647 mddev_resume(mddev);
6649 mddev_unlock(mddev);
6654 static struct md_sysfs_entry
6655 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6656 raid5_show_group_thread_cnt,
6657 raid5_store_group_thread_cnt);
6659 static struct attribute *raid5_attrs[] = {
6660 &raid5_stripecache_size.attr,
6661 &raid5_stripecache_active.attr,
6662 &raid5_preread_bypass_threshold.attr,
6663 &raid5_group_thread_cnt.attr,
6664 &raid5_skip_copy.attr,
6665 &raid5_rmw_level.attr,
6666 &r5c_journal_mode.attr,
6667 &ppl_write_hint.attr,
6670 static struct attribute_group raid5_attrs_group = {
6672 .attrs = raid5_attrs,
6675 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6677 int *worker_cnt_per_group,
6678 struct r5worker_group **worker_groups)
6682 struct r5worker *workers;
6684 *worker_cnt_per_group = cnt;
6687 *worker_groups = NULL;
6690 *group_cnt = num_possible_nodes();
6691 size = sizeof(struct r5worker) * cnt;
6692 workers = kcalloc(size, *group_cnt, GFP_NOIO);
6693 *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6695 if (!*worker_groups || !workers) {
6697 kfree(*worker_groups);
6701 for (i = 0; i < *group_cnt; i++) {
6702 struct r5worker_group *group;
6704 group = &(*worker_groups)[i];
6705 INIT_LIST_HEAD(&group->handle_list);
6706 INIT_LIST_HEAD(&group->loprio_list);
6708 group->workers = workers + i * cnt;
6710 for (j = 0; j < cnt; j++) {
6711 struct r5worker *worker = group->workers + j;
6712 worker->group = group;
6713 INIT_WORK(&worker->work, raid5_do_work);
6715 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6716 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6723 static void free_thread_groups(struct r5conf *conf)
6725 if (conf->worker_groups)
6726 kfree(conf->worker_groups[0].workers);
6727 kfree(conf->worker_groups);
6728 conf->worker_groups = NULL;
6732 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6734 struct r5conf *conf = mddev->private;
6737 sectors = mddev->dev_sectors;
6739 /* size is defined by the smallest of previous and new size */
6740 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6742 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6743 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6744 return sectors * (raid_disks - conf->max_degraded);
6747 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6749 safe_put_page(percpu->spare_page);
6750 percpu->spare_page = NULL;
6751 kvfree(percpu->scribble);
6752 percpu->scribble = NULL;
6755 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6757 if (conf->level == 6 && !percpu->spare_page) {
6758 percpu->spare_page = alloc_page(GFP_KERNEL);
6759 if (!percpu->spare_page)
6763 if (scribble_alloc(percpu,
6764 max(conf->raid_disks,
6765 conf->previous_raid_disks),
6766 max(conf->chunk_sectors,
6767 conf->prev_chunk_sectors)
6770 free_scratch_buffer(conf, percpu);
6777 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6779 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6781 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6785 static void raid5_free_percpu(struct r5conf *conf)
6790 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6791 free_percpu(conf->percpu);
6794 static void free_conf(struct r5conf *conf)
6800 unregister_shrinker(&conf->shrinker);
6801 free_thread_groups(conf);
6802 shrink_stripes(conf);
6803 raid5_free_percpu(conf);
6804 for (i = 0; i < conf->pool_size; i++)
6805 if (conf->disks[i].extra_page)
6806 put_page(conf->disks[i].extra_page);
6808 bioset_exit(&conf->bio_split);
6809 kfree(conf->stripe_hashtbl);
6810 kfree(conf->pending_data);
6814 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6816 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6817 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6819 if (alloc_scratch_buffer(conf, percpu)) {
6820 pr_warn("%s: failed memory allocation for cpu%u\n",
6827 static int raid5_alloc_percpu(struct r5conf *conf)
6831 conf->percpu = alloc_percpu(struct raid5_percpu);
6835 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6837 conf->scribble_disks = max(conf->raid_disks,
6838 conf->previous_raid_disks);
6839 conf->scribble_sectors = max(conf->chunk_sectors,
6840 conf->prev_chunk_sectors);
6845 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6846 struct shrink_control *sc)
6848 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6849 unsigned long ret = SHRINK_STOP;
6851 if (mutex_trylock(&conf->cache_size_mutex)) {
6853 while (ret < sc->nr_to_scan &&
6854 conf->max_nr_stripes > conf->min_nr_stripes) {
6855 if (drop_one_stripe(conf) == 0) {
6861 mutex_unlock(&conf->cache_size_mutex);
6866 static unsigned long raid5_cache_count(struct shrinker *shrink,
6867 struct shrink_control *sc)
6869 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6871 if (conf->max_nr_stripes < conf->min_nr_stripes)
6872 /* unlikely, but not impossible */
6874 return conf->max_nr_stripes - conf->min_nr_stripes;
6877 static struct r5conf *setup_conf(struct mddev *mddev)
6879 struct r5conf *conf;
6880 int raid_disk, memory, max_disks;
6881 struct md_rdev *rdev;
6882 struct disk_info *disk;
6885 int group_cnt, worker_cnt_per_group;
6886 struct r5worker_group *new_group;
6889 if (mddev->new_level != 5
6890 && mddev->new_level != 4
6891 && mddev->new_level != 6) {
6892 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6893 mdname(mddev), mddev->new_level);
6894 return ERR_PTR(-EIO);
6896 if ((mddev->new_level == 5
6897 && !algorithm_valid_raid5(mddev->new_layout)) ||
6898 (mddev->new_level == 6
6899 && !algorithm_valid_raid6(mddev->new_layout))) {
6900 pr_warn("md/raid:%s: layout %d not supported\n",
6901 mdname(mddev), mddev->new_layout);
6902 return ERR_PTR(-EIO);
6904 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6905 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6906 mdname(mddev), mddev->raid_disks);
6907 return ERR_PTR(-EINVAL);
6910 if (!mddev->new_chunk_sectors ||
6911 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6912 !is_power_of_2(mddev->new_chunk_sectors)) {
6913 pr_warn("md/raid:%s: invalid chunk size %d\n",
6914 mdname(mddev), mddev->new_chunk_sectors << 9);
6915 return ERR_PTR(-EINVAL);
6918 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6921 INIT_LIST_HEAD(&conf->free_list);
6922 INIT_LIST_HEAD(&conf->pending_list);
6923 conf->pending_data = kcalloc(PENDING_IO_MAX,
6924 sizeof(struct r5pending_data),
6926 if (!conf->pending_data)
6928 for (i = 0; i < PENDING_IO_MAX; i++)
6929 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6930 /* Don't enable multi-threading by default*/
6931 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6933 conf->group_cnt = group_cnt;
6934 conf->worker_cnt_per_group = worker_cnt_per_group;
6935 conf->worker_groups = new_group;
6938 spin_lock_init(&conf->device_lock);
6939 seqcount_init(&conf->gen_lock);
6940 mutex_init(&conf->cache_size_mutex);
6941 init_waitqueue_head(&conf->wait_for_quiescent);
6942 init_waitqueue_head(&conf->wait_for_stripe);
6943 init_waitqueue_head(&conf->wait_for_overlap);
6944 INIT_LIST_HEAD(&conf->handle_list);
6945 INIT_LIST_HEAD(&conf->loprio_list);
6946 INIT_LIST_HEAD(&conf->hold_list);
6947 INIT_LIST_HEAD(&conf->delayed_list);
6948 INIT_LIST_HEAD(&conf->bitmap_list);
6949 init_llist_head(&conf->released_stripes);
6950 atomic_set(&conf->active_stripes, 0);
6951 atomic_set(&conf->preread_active_stripes, 0);
6952 atomic_set(&conf->active_aligned_reads, 0);
6953 spin_lock_init(&conf->pending_bios_lock);
6954 conf->batch_bio_dispatch = true;
6955 rdev_for_each(rdev, mddev) {
6956 if (test_bit(Journal, &rdev->flags))
6958 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6959 conf->batch_bio_dispatch = false;
6964 conf->bypass_threshold = BYPASS_THRESHOLD;
6965 conf->recovery_disabled = mddev->recovery_disabled - 1;
6967 conf->raid_disks = mddev->raid_disks;
6968 if (mddev->reshape_position == MaxSector)
6969 conf->previous_raid_disks = mddev->raid_disks;
6971 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6972 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6974 conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6980 for (i = 0; i < max_disks; i++) {
6981 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6982 if (!conf->disks[i].extra_page)
6986 ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6989 conf->mddev = mddev;
6991 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6994 /* We init hash_locks[0] separately to that it can be used
6995 * as the reference lock in the spin_lock_nest_lock() call
6996 * in lock_all_device_hash_locks_irq in order to convince
6997 * lockdep that we know what we are doing.
6999 spin_lock_init(conf->hash_locks);
7000 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7001 spin_lock_init(conf->hash_locks + i);
7003 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7004 INIT_LIST_HEAD(conf->inactive_list + i);
7006 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7007 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7009 atomic_set(&conf->r5c_cached_full_stripes, 0);
7010 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7011 atomic_set(&conf->r5c_cached_partial_stripes, 0);
7012 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7013 atomic_set(&conf->r5c_flushing_full_stripes, 0);
7014 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7016 conf->level = mddev->new_level;
7017 conf->chunk_sectors = mddev->new_chunk_sectors;
7018 if (raid5_alloc_percpu(conf) != 0)
7021 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7023 rdev_for_each(rdev, mddev) {
7024 raid_disk = rdev->raid_disk;
7025 if (raid_disk >= max_disks
7026 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7028 disk = conf->disks + raid_disk;
7030 if (test_bit(Replacement, &rdev->flags)) {
7031 if (disk->replacement)
7033 disk->replacement = rdev;
7040 if (test_bit(In_sync, &rdev->flags)) {
7041 char b[BDEVNAME_SIZE];
7042 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7043 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7044 } else if (rdev->saved_raid_disk != raid_disk)
7045 /* Cannot rely on bitmap to complete recovery */
7049 conf->level = mddev->new_level;
7050 if (conf->level == 6) {
7051 conf->max_degraded = 2;
7052 if (raid6_call.xor_syndrome)
7053 conf->rmw_level = PARITY_ENABLE_RMW;
7055 conf->rmw_level = PARITY_DISABLE_RMW;
7057 conf->max_degraded = 1;
7058 conf->rmw_level = PARITY_ENABLE_RMW;
7060 conf->algorithm = mddev->new_layout;
7061 conf->reshape_progress = mddev->reshape_position;
7062 if (conf->reshape_progress != MaxSector) {
7063 conf->prev_chunk_sectors = mddev->chunk_sectors;
7064 conf->prev_algo = mddev->layout;
7066 conf->prev_chunk_sectors = conf->chunk_sectors;
7067 conf->prev_algo = conf->algorithm;
7070 conf->min_nr_stripes = NR_STRIPES;
7071 if (mddev->reshape_position != MaxSector) {
7072 int stripes = max_t(int,
7073 ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7074 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7075 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7076 if (conf->min_nr_stripes != NR_STRIPES)
7077 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7078 mdname(mddev), conf->min_nr_stripes);
7080 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7081 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7082 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7083 if (grow_stripes(conf, conf->min_nr_stripes)) {
7084 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7085 mdname(mddev), memory);
7088 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7090 * Losing a stripe head costs more than the time to refill it,
7091 * it reduces the queue depth and so can hurt throughput.
7092 * So set it rather large, scaled by number of devices.
7094 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7095 conf->shrinker.scan_objects = raid5_cache_scan;
7096 conf->shrinker.count_objects = raid5_cache_count;
7097 conf->shrinker.batch = 128;
7098 conf->shrinker.flags = 0;
7099 if (register_shrinker(&conf->shrinker)) {
7100 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7105 sprintf(pers_name, "raid%d", mddev->new_level);
7106 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7107 if (!conf->thread) {
7108 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7118 return ERR_PTR(-EIO);
7120 return ERR_PTR(-ENOMEM);
7123 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7126 case ALGORITHM_PARITY_0:
7127 if (raid_disk < max_degraded)
7130 case ALGORITHM_PARITY_N:
7131 if (raid_disk >= raid_disks - max_degraded)
7134 case ALGORITHM_PARITY_0_6:
7135 if (raid_disk == 0 ||
7136 raid_disk == raid_disks - 1)
7139 case ALGORITHM_LEFT_ASYMMETRIC_6:
7140 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7141 case ALGORITHM_LEFT_SYMMETRIC_6:
7142 case ALGORITHM_RIGHT_SYMMETRIC_6:
7143 if (raid_disk == raid_disks - 1)
7149 static int raid5_run(struct mddev *mddev)
7151 struct r5conf *conf;
7152 int working_disks = 0;
7153 int dirty_parity_disks = 0;
7154 struct md_rdev *rdev;
7155 struct md_rdev *journal_dev = NULL;
7156 sector_t reshape_offset = 0;
7158 long long min_offset_diff = 0;
7161 if (mddev_init_writes_pending(mddev) < 0)
7164 if (mddev->recovery_cp != MaxSector)
7165 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7168 rdev_for_each(rdev, mddev) {
7171 if (test_bit(Journal, &rdev->flags)) {
7175 if (rdev->raid_disk < 0)
7177 diff = (rdev->new_data_offset - rdev->data_offset);
7179 min_offset_diff = diff;
7181 } else if (mddev->reshape_backwards &&
7182 diff < min_offset_diff)
7183 min_offset_diff = diff;
7184 else if (!mddev->reshape_backwards &&
7185 diff > min_offset_diff)
7186 min_offset_diff = diff;
7189 if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7190 (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7191 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7196 if (mddev->reshape_position != MaxSector) {
7197 /* Check that we can continue the reshape.
7198 * Difficulties arise if the stripe we would write to
7199 * next is at or after the stripe we would read from next.
7200 * For a reshape that changes the number of devices, this
7201 * is only possible for a very short time, and mdadm makes
7202 * sure that time appears to have past before assembling
7203 * the array. So we fail if that time hasn't passed.
7204 * For a reshape that keeps the number of devices the same
7205 * mdadm must be monitoring the reshape can keeping the
7206 * critical areas read-only and backed up. It will start
7207 * the array in read-only mode, so we check for that.
7209 sector_t here_new, here_old;
7211 int max_degraded = (mddev->level == 6 ? 2 : 1);
7216 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7221 if (mddev->new_level != mddev->level) {
7222 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7226 old_disks = mddev->raid_disks - mddev->delta_disks;
7227 /* reshape_position must be on a new-stripe boundary, and one
7228 * further up in new geometry must map after here in old
7230 * If the chunk sizes are different, then as we perform reshape
7231 * in units of the largest of the two, reshape_position needs
7232 * be a multiple of the largest chunk size times new data disks.
7234 here_new = mddev->reshape_position;
7235 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7236 new_data_disks = mddev->raid_disks - max_degraded;
7237 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7238 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7242 reshape_offset = here_new * chunk_sectors;
7243 /* here_new is the stripe we will write to */
7244 here_old = mddev->reshape_position;
7245 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7246 /* here_old is the first stripe that we might need to read
7248 if (mddev->delta_disks == 0) {
7249 /* We cannot be sure it is safe to start an in-place
7250 * reshape. It is only safe if user-space is monitoring
7251 * and taking constant backups.
7252 * mdadm always starts a situation like this in
7253 * readonly mode so it can take control before
7254 * allowing any writes. So just check for that.
7256 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7257 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7258 /* not really in-place - so OK */;
7259 else if (mddev->ro == 0) {
7260 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7264 } else if (mddev->reshape_backwards
7265 ? (here_new * chunk_sectors + min_offset_diff <=
7266 here_old * chunk_sectors)
7267 : (here_new * chunk_sectors >=
7268 here_old * chunk_sectors + (-min_offset_diff))) {
7269 /* Reading from the same stripe as writing to - bad */
7270 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7274 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7275 /* OK, we should be able to continue; */
7277 BUG_ON(mddev->level != mddev->new_level);
7278 BUG_ON(mddev->layout != mddev->new_layout);
7279 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7280 BUG_ON(mddev->delta_disks != 0);
7283 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7284 test_bit(MD_HAS_PPL, &mddev->flags)) {
7285 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7287 clear_bit(MD_HAS_PPL, &mddev->flags);
7288 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7291 if (mddev->private == NULL)
7292 conf = setup_conf(mddev);
7294 conf = mddev->private;
7297 return PTR_ERR(conf);
7299 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7301 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7304 set_disk_ro(mddev->gendisk, 1);
7305 } else if (mddev->recovery_cp == MaxSector)
7306 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7309 conf->min_offset_diff = min_offset_diff;
7310 mddev->thread = conf->thread;
7311 conf->thread = NULL;
7312 mddev->private = conf;
7314 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7316 rdev = conf->disks[i].rdev;
7317 if (!rdev && conf->disks[i].replacement) {
7318 /* The replacement is all we have yet */
7319 rdev = conf->disks[i].replacement;
7320 conf->disks[i].replacement = NULL;
7321 clear_bit(Replacement, &rdev->flags);
7322 conf->disks[i].rdev = rdev;
7326 if (conf->disks[i].replacement &&
7327 conf->reshape_progress != MaxSector) {
7328 /* replacements and reshape simply do not mix. */
7329 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7332 if (test_bit(In_sync, &rdev->flags)) {
7336 /* This disc is not fully in-sync. However if it
7337 * just stored parity (beyond the recovery_offset),
7338 * when we don't need to be concerned about the
7339 * array being dirty.
7340 * When reshape goes 'backwards', we never have
7341 * partially completed devices, so we only need
7342 * to worry about reshape going forwards.
7344 /* Hack because v0.91 doesn't store recovery_offset properly. */
7345 if (mddev->major_version == 0 &&
7346 mddev->minor_version > 90)
7347 rdev->recovery_offset = reshape_offset;
7349 if (rdev->recovery_offset < reshape_offset) {
7350 /* We need to check old and new layout */
7351 if (!only_parity(rdev->raid_disk,
7354 conf->max_degraded))
7357 if (!only_parity(rdev->raid_disk,
7359 conf->previous_raid_disks,
7360 conf->max_degraded))
7362 dirty_parity_disks++;
7366 * 0 for a fully functional array, 1 or 2 for a degraded array.
7368 mddev->degraded = raid5_calc_degraded(conf);
7370 if (has_failed(conf)) {
7371 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7372 mdname(mddev), mddev->degraded, conf->raid_disks);
7376 /* device size must be a multiple of chunk size */
7377 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7378 mddev->resync_max_sectors = mddev->dev_sectors;
7380 if (mddev->degraded > dirty_parity_disks &&
7381 mddev->recovery_cp != MaxSector) {
7382 if (test_bit(MD_HAS_PPL, &mddev->flags))
7383 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7385 else if (mddev->ok_start_degraded)
7386 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7389 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7395 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7396 mdname(mddev), conf->level,
7397 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7400 print_raid5_conf(conf);
7402 if (conf->reshape_progress != MaxSector) {
7403 conf->reshape_safe = conf->reshape_progress;
7404 atomic_set(&conf->reshape_stripes, 0);
7405 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7406 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7407 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7408 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7409 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7411 if (!mddev->sync_thread)
7415 /* Ok, everything is just fine now */
7416 if (mddev->to_remove == &raid5_attrs_group)
7417 mddev->to_remove = NULL;
7418 else if (mddev->kobj.sd &&
7419 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7420 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7422 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7426 /* read-ahead size must cover two whole stripes, which
7427 * is 2 * (datadisks) * chunksize where 'n' is the
7428 * number of raid devices
7430 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7431 int stripe = data_disks *
7432 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7433 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7434 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7436 chunk_size = mddev->chunk_sectors << 9;
7437 blk_queue_io_min(mddev->queue, chunk_size);
7438 blk_queue_io_opt(mddev->queue, chunk_size *
7439 (conf->raid_disks - conf->max_degraded));
7440 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7442 * We can only discard a whole stripe. It doesn't make sense to
7443 * discard data disk but write parity disk
7445 stripe = stripe * PAGE_SIZE;
7446 /* Round up to power of 2, as discard handling
7447 * currently assumes that */
7448 while ((stripe-1) & stripe)
7449 stripe = (stripe | (stripe-1)) + 1;
7450 mddev->queue->limits.discard_alignment = stripe;
7451 mddev->queue->limits.discard_granularity = stripe;
7453 blk_queue_max_write_same_sectors(mddev->queue, 0);
7454 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7456 rdev_for_each(rdev, mddev) {
7457 disk_stack_limits(mddev->gendisk, rdev->bdev,
7458 rdev->data_offset << 9);
7459 disk_stack_limits(mddev->gendisk, rdev->bdev,
7460 rdev->new_data_offset << 9);
7464 * zeroing is required, otherwise data
7465 * could be lost. Consider a scenario: discard a stripe
7466 * (the stripe could be inconsistent if
7467 * discard_zeroes_data is 0); write one disk of the
7468 * stripe (the stripe could be inconsistent again
7469 * depending on which disks are used to calculate
7470 * parity); the disk is broken; The stripe data of this
7473 * We only allow DISCARD if the sysadmin has confirmed that
7474 * only safe devices are in use by setting a module parameter.
7475 * A better idea might be to turn DISCARD into WRITE_ZEROES
7476 * requests, as that is required to be safe.
7478 if (devices_handle_discard_safely &&
7479 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7480 mddev->queue->limits.discard_granularity >= stripe)
7481 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7484 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7487 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7490 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7495 md_unregister_thread(&mddev->thread);
7496 print_raid5_conf(conf);
7498 mddev->private = NULL;
7499 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7503 static void raid5_free(struct mddev *mddev, void *priv)
7505 struct r5conf *conf = priv;
7508 mddev->to_remove = &raid5_attrs_group;
7511 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7513 struct r5conf *conf = mddev->private;
7516 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7517 conf->chunk_sectors / 2, mddev->layout);
7518 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7520 for (i = 0; i < conf->raid_disks; i++) {
7521 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7522 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7525 seq_printf (seq, "]");
7528 static void print_raid5_conf (struct r5conf *conf)
7531 struct disk_info *tmp;
7533 pr_debug("RAID conf printout:\n");
7535 pr_debug("(conf==NULL)\n");
7538 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7540 conf->raid_disks - conf->mddev->degraded);
7542 for (i = 0; i < conf->raid_disks; i++) {
7543 char b[BDEVNAME_SIZE];
7544 tmp = conf->disks + i;
7546 pr_debug(" disk %d, o:%d, dev:%s\n",
7547 i, !test_bit(Faulty, &tmp->rdev->flags),
7548 bdevname(tmp->rdev->bdev, b));
7552 static int raid5_spare_active(struct mddev *mddev)
7555 struct r5conf *conf = mddev->private;
7556 struct disk_info *tmp;
7558 unsigned long flags;
7560 for (i = 0; i < conf->raid_disks; i++) {
7561 tmp = conf->disks + i;
7562 if (tmp->replacement
7563 && tmp->replacement->recovery_offset == MaxSector
7564 && !test_bit(Faulty, &tmp->replacement->flags)
7565 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7566 /* Replacement has just become active. */
7568 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7571 /* Replaced device not technically faulty,
7572 * but we need to be sure it gets removed
7573 * and never re-added.
7575 set_bit(Faulty, &tmp->rdev->flags);
7576 sysfs_notify_dirent_safe(
7577 tmp->rdev->sysfs_state);
7579 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7580 } else if (tmp->rdev
7581 && tmp->rdev->recovery_offset == MaxSector
7582 && !test_bit(Faulty, &tmp->rdev->flags)
7583 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7585 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7588 spin_lock_irqsave(&conf->device_lock, flags);
7589 mddev->degraded = raid5_calc_degraded(conf);
7590 spin_unlock_irqrestore(&conf->device_lock, flags);
7591 print_raid5_conf(conf);
7595 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7597 struct r5conf *conf = mddev->private;
7599 int number = rdev->raid_disk;
7600 struct md_rdev **rdevp;
7601 struct disk_info *p = conf->disks + number;
7603 print_raid5_conf(conf);
7604 if (test_bit(Journal, &rdev->flags) && conf->log) {
7606 * we can't wait pending write here, as this is called in
7607 * raid5d, wait will deadlock.
7608 * neilb: there is no locking about new writes here,
7609 * so this cannot be safe.
7611 if (atomic_read(&conf->active_stripes) ||
7612 atomic_read(&conf->r5c_cached_full_stripes) ||
7613 atomic_read(&conf->r5c_cached_partial_stripes)) {
7619 if (rdev == p->rdev)
7621 else if (rdev == p->replacement)
7622 rdevp = &p->replacement;
7626 if (number >= conf->raid_disks &&
7627 conf->reshape_progress == MaxSector)
7628 clear_bit(In_sync, &rdev->flags);
7630 if (test_bit(In_sync, &rdev->flags) ||
7631 atomic_read(&rdev->nr_pending)) {
7635 /* Only remove non-faulty devices if recovery
7638 if (!test_bit(Faulty, &rdev->flags) &&
7639 mddev->recovery_disabled != conf->recovery_disabled &&
7640 !has_failed(conf) &&
7641 (!p->replacement || p->replacement == rdev) &&
7642 number < conf->raid_disks) {
7647 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7649 if (atomic_read(&rdev->nr_pending)) {
7650 /* lost the race, try later */
7656 err = log_modify(conf, rdev, false);
7660 if (p->replacement) {
7661 /* We must have just cleared 'rdev' */
7662 p->rdev = p->replacement;
7663 clear_bit(Replacement, &p->replacement->flags);
7664 smp_mb(); /* Make sure other CPUs may see both as identical
7665 * but will never see neither - if they are careful
7667 p->replacement = NULL;
7670 err = log_modify(conf, p->rdev, true);
7673 clear_bit(WantReplacement, &rdev->flags);
7676 print_raid5_conf(conf);
7680 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7682 struct r5conf *conf = mddev->private;
7685 struct disk_info *p;
7687 int last = conf->raid_disks - 1;
7689 if (test_bit(Journal, &rdev->flags)) {
7693 rdev->raid_disk = 0;
7695 * The array is in readonly mode if journal is missing, so no
7696 * write requests running. We should be safe
7698 log_init(conf, rdev, false);
7701 if (mddev->recovery_disabled == conf->recovery_disabled)
7704 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7705 /* no point adding a device */
7708 if (rdev->raid_disk >= 0)
7709 first = last = rdev->raid_disk;
7712 * find the disk ... but prefer rdev->saved_raid_disk
7715 if (rdev->saved_raid_disk >= 0 &&
7716 rdev->saved_raid_disk >= first &&
7717 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7718 first = rdev->saved_raid_disk;
7720 for (disk = first; disk <= last; disk++) {
7721 p = conf->disks + disk;
7722 if (p->rdev == NULL) {
7723 clear_bit(In_sync, &rdev->flags);
7724 rdev->raid_disk = disk;
7725 if (rdev->saved_raid_disk != disk)
7727 rcu_assign_pointer(p->rdev, rdev);
7729 err = log_modify(conf, rdev, true);
7734 for (disk = first; disk <= last; disk++) {
7735 p = conf->disks + disk;
7736 if (test_bit(WantReplacement, &p->rdev->flags) &&
7737 p->replacement == NULL) {
7738 clear_bit(In_sync, &rdev->flags);
7739 set_bit(Replacement, &rdev->flags);
7740 rdev->raid_disk = disk;
7743 rcu_assign_pointer(p->replacement, rdev);
7748 print_raid5_conf(conf);
7752 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7754 /* no resync is happening, and there is enough space
7755 * on all devices, so we can resize.
7756 * We need to make sure resync covers any new space.
7757 * If the array is shrinking we should possibly wait until
7758 * any io in the removed space completes, but it hardly seems
7762 struct r5conf *conf = mddev->private;
7764 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7766 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7767 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7768 if (mddev->external_size &&
7769 mddev->array_sectors > newsize)
7771 if (mddev->bitmap) {
7772 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7776 md_set_array_sectors(mddev, newsize);
7777 if (sectors > mddev->dev_sectors &&
7778 mddev->recovery_cp > mddev->dev_sectors) {
7779 mddev->recovery_cp = mddev->dev_sectors;
7780 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7782 mddev->dev_sectors = sectors;
7783 mddev->resync_max_sectors = sectors;
7787 static int check_stripe_cache(struct mddev *mddev)
7789 /* Can only proceed if there are plenty of stripe_heads.
7790 * We need a minimum of one full stripe,, and for sensible progress
7791 * it is best to have about 4 times that.
7792 * If we require 4 times, then the default 256 4K stripe_heads will
7793 * allow for chunk sizes up to 256K, which is probably OK.
7794 * If the chunk size is greater, user-space should request more
7795 * stripe_heads first.
7797 struct r5conf *conf = mddev->private;
7798 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7799 > conf->min_nr_stripes ||
7800 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7801 > conf->min_nr_stripes) {
7802 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7804 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7811 static int check_reshape(struct mddev *mddev)
7813 struct r5conf *conf = mddev->private;
7815 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7817 if (mddev->delta_disks == 0 &&
7818 mddev->new_layout == mddev->layout &&
7819 mddev->new_chunk_sectors == mddev->chunk_sectors)
7820 return 0; /* nothing to do */
7821 if (has_failed(conf))
7823 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7824 /* We might be able to shrink, but the devices must
7825 * be made bigger first.
7826 * For raid6, 4 is the minimum size.
7827 * Otherwise 2 is the minimum
7830 if (mddev->level == 6)
7832 if (mddev->raid_disks + mddev->delta_disks < min)
7836 if (!check_stripe_cache(mddev))
7839 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7840 mddev->delta_disks > 0)
7841 if (resize_chunks(conf,
7842 conf->previous_raid_disks
7843 + max(0, mddev->delta_disks),
7844 max(mddev->new_chunk_sectors,
7845 mddev->chunk_sectors)
7849 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7850 return 0; /* never bother to shrink */
7851 return resize_stripes(conf, (conf->previous_raid_disks
7852 + mddev->delta_disks));
7855 static int raid5_start_reshape(struct mddev *mddev)
7857 struct r5conf *conf = mddev->private;
7858 struct md_rdev *rdev;
7860 unsigned long flags;
7862 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7865 if (!check_stripe_cache(mddev))
7868 if (has_failed(conf))
7871 rdev_for_each(rdev, mddev) {
7872 if (!test_bit(In_sync, &rdev->flags)
7873 && !test_bit(Faulty, &rdev->flags))
7877 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7878 /* Not enough devices even to make a degraded array
7883 /* Refuse to reduce size of the array. Any reductions in
7884 * array size must be through explicit setting of array_size
7887 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7888 < mddev->array_sectors) {
7889 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7894 atomic_set(&conf->reshape_stripes, 0);
7895 spin_lock_irq(&conf->device_lock);
7896 write_seqcount_begin(&conf->gen_lock);
7897 conf->previous_raid_disks = conf->raid_disks;
7898 conf->raid_disks += mddev->delta_disks;
7899 conf->prev_chunk_sectors = conf->chunk_sectors;
7900 conf->chunk_sectors = mddev->new_chunk_sectors;
7901 conf->prev_algo = conf->algorithm;
7902 conf->algorithm = mddev->new_layout;
7904 /* Code that selects data_offset needs to see the generation update
7905 * if reshape_progress has been set - so a memory barrier needed.
7908 if (mddev->reshape_backwards)
7909 conf->reshape_progress = raid5_size(mddev, 0, 0);
7911 conf->reshape_progress = 0;
7912 conf->reshape_safe = conf->reshape_progress;
7913 write_seqcount_end(&conf->gen_lock);
7914 spin_unlock_irq(&conf->device_lock);
7916 /* Now make sure any requests that proceeded on the assumption
7917 * the reshape wasn't running - like Discard or Read - have
7920 mddev_suspend(mddev);
7921 mddev_resume(mddev);
7923 /* Add some new drives, as many as will fit.
7924 * We know there are enough to make the newly sized array work.
7925 * Don't add devices if we are reducing the number of
7926 * devices in the array. This is because it is not possible
7927 * to correctly record the "partially reconstructed" state of
7928 * such devices during the reshape and confusion could result.
7930 if (mddev->delta_disks >= 0) {
7931 rdev_for_each(rdev, mddev)
7932 if (rdev->raid_disk < 0 &&
7933 !test_bit(Faulty, &rdev->flags)) {
7934 if (raid5_add_disk(mddev, rdev) == 0) {
7936 >= conf->previous_raid_disks)
7937 set_bit(In_sync, &rdev->flags);
7939 rdev->recovery_offset = 0;
7941 if (sysfs_link_rdev(mddev, rdev))
7942 /* Failure here is OK */;
7944 } else if (rdev->raid_disk >= conf->previous_raid_disks
7945 && !test_bit(Faulty, &rdev->flags)) {
7946 /* This is a spare that was manually added */
7947 set_bit(In_sync, &rdev->flags);
7950 /* When a reshape changes the number of devices,
7951 * ->degraded is measured against the larger of the
7952 * pre and post number of devices.
7954 spin_lock_irqsave(&conf->device_lock, flags);
7955 mddev->degraded = raid5_calc_degraded(conf);
7956 spin_unlock_irqrestore(&conf->device_lock, flags);
7958 mddev->raid_disks = conf->raid_disks;
7959 mddev->reshape_position = conf->reshape_progress;
7960 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7962 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7963 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7964 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7965 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7966 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7967 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7969 if (!mddev->sync_thread) {
7970 mddev->recovery = 0;
7971 spin_lock_irq(&conf->device_lock);
7972 write_seqcount_begin(&conf->gen_lock);
7973 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7974 mddev->new_chunk_sectors =
7975 conf->chunk_sectors = conf->prev_chunk_sectors;
7976 mddev->new_layout = conf->algorithm = conf->prev_algo;
7977 rdev_for_each(rdev, mddev)
7978 rdev->new_data_offset = rdev->data_offset;
7980 conf->generation --;
7981 conf->reshape_progress = MaxSector;
7982 mddev->reshape_position = MaxSector;
7983 write_seqcount_end(&conf->gen_lock);
7984 spin_unlock_irq(&conf->device_lock);
7987 conf->reshape_checkpoint = jiffies;
7988 md_wakeup_thread(mddev->sync_thread);
7989 md_new_event(mddev);
7993 /* This is called from the reshape thread and should make any
7994 * changes needed in 'conf'
7996 static void end_reshape(struct r5conf *conf)
7999 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8000 struct md_rdev *rdev;
8002 spin_lock_irq(&conf->device_lock);
8003 conf->previous_raid_disks = conf->raid_disks;
8004 md_finish_reshape(conf->mddev);
8006 conf->reshape_progress = MaxSector;
8007 conf->mddev->reshape_position = MaxSector;
8008 rdev_for_each(rdev, conf->mddev)
8009 if (rdev->raid_disk >= 0 &&
8010 !test_bit(Journal, &rdev->flags) &&
8011 !test_bit(In_sync, &rdev->flags))
8012 rdev->recovery_offset = MaxSector;
8013 spin_unlock_irq(&conf->device_lock);
8014 wake_up(&conf->wait_for_overlap);
8016 /* read-ahead size must cover two whole stripes, which is
8017 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8019 if (conf->mddev->queue) {
8020 int data_disks = conf->raid_disks - conf->max_degraded;
8021 int stripe = data_disks * ((conf->chunk_sectors << 9)
8023 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8024 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8029 /* This is called from the raid5d thread with mddev_lock held.
8030 * It makes config changes to the device.
8032 static void raid5_finish_reshape(struct mddev *mddev)
8034 struct r5conf *conf = mddev->private;
8036 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8038 if (mddev->delta_disks <= 0) {
8040 spin_lock_irq(&conf->device_lock);
8041 mddev->degraded = raid5_calc_degraded(conf);
8042 spin_unlock_irq(&conf->device_lock);
8043 for (d = conf->raid_disks ;
8044 d < conf->raid_disks - mddev->delta_disks;
8046 struct md_rdev *rdev = conf->disks[d].rdev;
8048 clear_bit(In_sync, &rdev->flags);
8049 rdev = conf->disks[d].replacement;
8051 clear_bit(In_sync, &rdev->flags);
8054 mddev->layout = conf->algorithm;
8055 mddev->chunk_sectors = conf->chunk_sectors;
8056 mddev->reshape_position = MaxSector;
8057 mddev->delta_disks = 0;
8058 mddev->reshape_backwards = 0;
8062 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8064 struct r5conf *conf = mddev->private;
8067 /* stop all writes */
8068 lock_all_device_hash_locks_irq(conf);
8069 /* '2' tells resync/reshape to pause so that all
8070 * active stripes can drain
8072 r5c_flush_cache(conf, INT_MAX);
8074 wait_event_cmd(conf->wait_for_quiescent,
8075 atomic_read(&conf->active_stripes) == 0 &&
8076 atomic_read(&conf->active_aligned_reads) == 0,
8077 unlock_all_device_hash_locks_irq(conf),
8078 lock_all_device_hash_locks_irq(conf));
8080 unlock_all_device_hash_locks_irq(conf);
8081 /* allow reshape to continue */
8082 wake_up(&conf->wait_for_overlap);
8084 /* re-enable writes */
8085 lock_all_device_hash_locks_irq(conf);
8087 wake_up(&conf->wait_for_quiescent);
8088 wake_up(&conf->wait_for_overlap);
8089 unlock_all_device_hash_locks_irq(conf);
8091 log_quiesce(conf, quiesce);
8094 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8096 struct r0conf *raid0_conf = mddev->private;
8099 /* for raid0 takeover only one zone is supported */
8100 if (raid0_conf->nr_strip_zones > 1) {
8101 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8103 return ERR_PTR(-EINVAL);
8106 sectors = raid0_conf->strip_zone[0].zone_end;
8107 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8108 mddev->dev_sectors = sectors;
8109 mddev->new_level = level;
8110 mddev->new_layout = ALGORITHM_PARITY_N;
8111 mddev->new_chunk_sectors = mddev->chunk_sectors;
8112 mddev->raid_disks += 1;
8113 mddev->delta_disks = 1;
8114 /* make sure it will be not marked as dirty */
8115 mddev->recovery_cp = MaxSector;
8117 return setup_conf(mddev);
8120 static void *raid5_takeover_raid1(struct mddev *mddev)
8125 if (mddev->raid_disks != 2 ||
8126 mddev->degraded > 1)
8127 return ERR_PTR(-EINVAL);
8129 /* Should check if there are write-behind devices? */
8131 chunksect = 64*2; /* 64K by default */
8133 /* The array must be an exact multiple of chunksize */
8134 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8137 if ((chunksect<<9) < STRIPE_SIZE)
8138 /* array size does not allow a suitable chunk size */
8139 return ERR_PTR(-EINVAL);
8141 mddev->new_level = 5;
8142 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8143 mddev->new_chunk_sectors = chunksect;
8145 ret = setup_conf(mddev);
8147 mddev_clear_unsupported_flags(mddev,
8148 UNSUPPORTED_MDDEV_FLAGS);
8152 static void *raid5_takeover_raid6(struct mddev *mddev)
8156 switch (mddev->layout) {
8157 case ALGORITHM_LEFT_ASYMMETRIC_6:
8158 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8160 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8161 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8163 case ALGORITHM_LEFT_SYMMETRIC_6:
8164 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8166 case ALGORITHM_RIGHT_SYMMETRIC_6:
8167 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8169 case ALGORITHM_PARITY_0_6:
8170 new_layout = ALGORITHM_PARITY_0;
8172 case ALGORITHM_PARITY_N:
8173 new_layout = ALGORITHM_PARITY_N;
8176 return ERR_PTR(-EINVAL);
8178 mddev->new_level = 5;
8179 mddev->new_layout = new_layout;
8180 mddev->delta_disks = -1;
8181 mddev->raid_disks -= 1;
8182 return setup_conf(mddev);
8185 static int raid5_check_reshape(struct mddev *mddev)
8187 /* For a 2-drive array, the layout and chunk size can be changed
8188 * immediately as not restriping is needed.
8189 * For larger arrays we record the new value - after validation
8190 * to be used by a reshape pass.
8192 struct r5conf *conf = mddev->private;
8193 int new_chunk = mddev->new_chunk_sectors;
8195 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8197 if (new_chunk > 0) {
8198 if (!is_power_of_2(new_chunk))
8200 if (new_chunk < (PAGE_SIZE>>9))
8202 if (mddev->array_sectors & (new_chunk-1))
8203 /* not factor of array size */
8207 /* They look valid */
8209 if (mddev->raid_disks == 2) {
8210 /* can make the change immediately */
8211 if (mddev->new_layout >= 0) {
8212 conf->algorithm = mddev->new_layout;
8213 mddev->layout = mddev->new_layout;
8215 if (new_chunk > 0) {
8216 conf->chunk_sectors = new_chunk ;
8217 mddev->chunk_sectors = new_chunk;
8219 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8220 md_wakeup_thread(mddev->thread);
8222 return check_reshape(mddev);
8225 static int raid6_check_reshape(struct mddev *mddev)
8227 int new_chunk = mddev->new_chunk_sectors;
8229 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8231 if (new_chunk > 0) {
8232 if (!is_power_of_2(new_chunk))
8234 if (new_chunk < (PAGE_SIZE >> 9))
8236 if (mddev->array_sectors & (new_chunk-1))
8237 /* not factor of array size */
8241 /* They look valid */
8242 return check_reshape(mddev);
8245 static void *raid5_takeover(struct mddev *mddev)
8247 /* raid5 can take over:
8248 * raid0 - if there is only one strip zone - make it a raid4 layout
8249 * raid1 - if there are two drives. We need to know the chunk size
8250 * raid4 - trivial - just use a raid4 layout.
8251 * raid6 - Providing it is a *_6 layout
8253 if (mddev->level == 0)
8254 return raid45_takeover_raid0(mddev, 5);
8255 if (mddev->level == 1)
8256 return raid5_takeover_raid1(mddev);
8257 if (mddev->level == 4) {
8258 mddev->new_layout = ALGORITHM_PARITY_N;
8259 mddev->new_level = 5;
8260 return setup_conf(mddev);
8262 if (mddev->level == 6)
8263 return raid5_takeover_raid6(mddev);
8265 return ERR_PTR(-EINVAL);
8268 static void *raid4_takeover(struct mddev *mddev)
8270 /* raid4 can take over:
8271 * raid0 - if there is only one strip zone
8272 * raid5 - if layout is right
8274 if (mddev->level == 0)
8275 return raid45_takeover_raid0(mddev, 4);
8276 if (mddev->level == 5 &&
8277 mddev->layout == ALGORITHM_PARITY_N) {
8278 mddev->new_layout = 0;
8279 mddev->new_level = 4;
8280 return setup_conf(mddev);
8282 return ERR_PTR(-EINVAL);
8285 static struct md_personality raid5_personality;
8287 static void *raid6_takeover(struct mddev *mddev)
8289 /* Currently can only take over a raid5. We map the
8290 * personality to an equivalent raid6 personality
8291 * with the Q block at the end.
8295 if (mddev->pers != &raid5_personality)
8296 return ERR_PTR(-EINVAL);
8297 if (mddev->degraded > 1)
8298 return ERR_PTR(-EINVAL);
8299 if (mddev->raid_disks > 253)
8300 return ERR_PTR(-EINVAL);
8301 if (mddev->raid_disks < 3)
8302 return ERR_PTR(-EINVAL);
8304 switch (mddev->layout) {
8305 case ALGORITHM_LEFT_ASYMMETRIC:
8306 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8308 case ALGORITHM_RIGHT_ASYMMETRIC:
8309 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8311 case ALGORITHM_LEFT_SYMMETRIC:
8312 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8314 case ALGORITHM_RIGHT_SYMMETRIC:
8315 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8317 case ALGORITHM_PARITY_0:
8318 new_layout = ALGORITHM_PARITY_0_6;
8320 case ALGORITHM_PARITY_N:
8321 new_layout = ALGORITHM_PARITY_N;
8324 return ERR_PTR(-EINVAL);
8326 mddev->new_level = 6;
8327 mddev->new_layout = new_layout;
8328 mddev->delta_disks = 1;
8329 mddev->raid_disks += 1;
8330 return setup_conf(mddev);
8333 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8335 struct r5conf *conf;
8338 err = mddev_lock(mddev);
8341 conf = mddev->private;
8343 mddev_unlock(mddev);
8347 if (strncmp(buf, "ppl", 3) == 0) {
8348 /* ppl only works with RAID 5 */
8349 if (!raid5_has_ppl(conf) && conf->level == 5) {
8350 err = log_init(conf, NULL, true);
8352 err = resize_stripes(conf, conf->pool_size);
8358 } else if (strncmp(buf, "resync", 6) == 0) {
8359 if (raid5_has_ppl(conf)) {
8360 mddev_suspend(mddev);
8362 mddev_resume(mddev);
8363 err = resize_stripes(conf, conf->pool_size);
8364 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8365 r5l_log_disk_error(conf)) {
8366 bool journal_dev_exists = false;
8367 struct md_rdev *rdev;
8369 rdev_for_each(rdev, mddev)
8370 if (test_bit(Journal, &rdev->flags)) {
8371 journal_dev_exists = true;
8375 if (!journal_dev_exists) {
8376 mddev_suspend(mddev);
8377 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8378 mddev_resume(mddev);
8379 } else /* need remove journal device first */
8388 md_update_sb(mddev, 1);
8390 mddev_unlock(mddev);
8395 static int raid5_start(struct mddev *mddev)
8397 struct r5conf *conf = mddev->private;
8399 return r5l_start(conf->log);
8402 static struct md_personality raid6_personality =
8406 .owner = THIS_MODULE,
8407 .make_request = raid5_make_request,
8409 .start = raid5_start,
8411 .status = raid5_status,
8412 .error_handler = raid5_error,
8413 .hot_add_disk = raid5_add_disk,
8414 .hot_remove_disk= raid5_remove_disk,
8415 .spare_active = raid5_spare_active,
8416 .sync_request = raid5_sync_request,
8417 .resize = raid5_resize,
8419 .check_reshape = raid6_check_reshape,
8420 .start_reshape = raid5_start_reshape,
8421 .finish_reshape = raid5_finish_reshape,
8422 .quiesce = raid5_quiesce,
8423 .takeover = raid6_takeover,
8424 .congested = raid5_congested,
8425 .change_consistency_policy = raid5_change_consistency_policy,
8427 static struct md_personality raid5_personality =
8431 .owner = THIS_MODULE,
8432 .make_request = raid5_make_request,
8434 .start = raid5_start,
8436 .status = raid5_status,
8437 .error_handler = raid5_error,
8438 .hot_add_disk = raid5_add_disk,
8439 .hot_remove_disk= raid5_remove_disk,
8440 .spare_active = raid5_spare_active,
8441 .sync_request = raid5_sync_request,
8442 .resize = raid5_resize,
8444 .check_reshape = raid5_check_reshape,
8445 .start_reshape = raid5_start_reshape,
8446 .finish_reshape = raid5_finish_reshape,
8447 .quiesce = raid5_quiesce,
8448 .takeover = raid5_takeover,
8449 .congested = raid5_congested,
8450 .change_consistency_policy = raid5_change_consistency_policy,
8453 static struct md_personality raid4_personality =
8457 .owner = THIS_MODULE,
8458 .make_request = raid5_make_request,
8460 .start = raid5_start,
8462 .status = raid5_status,
8463 .error_handler = raid5_error,
8464 .hot_add_disk = raid5_add_disk,
8465 .hot_remove_disk= raid5_remove_disk,
8466 .spare_active = raid5_spare_active,
8467 .sync_request = raid5_sync_request,
8468 .resize = raid5_resize,
8470 .check_reshape = raid5_check_reshape,
8471 .start_reshape = raid5_start_reshape,
8472 .finish_reshape = raid5_finish_reshape,
8473 .quiesce = raid5_quiesce,
8474 .takeover = raid4_takeover,
8475 .congested = raid5_congested,
8476 .change_consistency_policy = raid5_change_consistency_policy,
8479 static int __init raid5_init(void)
8483 raid5_wq = alloc_workqueue("raid5wq",
8484 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8488 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8490 raid456_cpu_up_prepare,
8493 destroy_workqueue(raid5_wq);
8496 register_md_personality(&raid6_personality);
8497 register_md_personality(&raid5_personality);
8498 register_md_personality(&raid4_personality);
8502 static void raid5_exit(void)
8504 unregister_md_personality(&raid6_personality);
8505 unregister_md_personality(&raid5_personality);
8506 unregister_md_personality(&raid4_personality);
8507 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8508 destroy_workqueue(raid5_wq);
8511 module_init(raid5_init);
8512 module_exit(raid5_exit);
8513 MODULE_LICENSE("GPL");
8514 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8515 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8516 MODULE_ALIAS("md-raid5");
8517 MODULE_ALIAS("md-raid4");
8518 MODULE_ALIAS("md-level-5");
8519 MODULE_ALIAS("md-level-4");
8520 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8521 MODULE_ALIAS("md-raid6");
8522 MODULE_ALIAS("md-level-6");
8524 /* This used to be two separate modules, they were: */
8525 MODULE_ALIAS("raid5");
8526 MODULE_ALIAS("raid6");