2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <linux/sched/signal.h>
60 #include <trace/events/block.h>
61 #include <linux/list_sort.h>
67 #include "raid5-log.h"
69 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
71 #define cpu_to_group(cpu) cpu_to_node(cpu)
72 #define ANY_GROUP NUMA_NO_NODE
74 static bool devices_handle_discard_safely = false;
75 module_param(devices_handle_discard_safely, bool, 0644);
76 MODULE_PARM_DESC(devices_handle_discard_safely,
77 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
78 static struct workqueue_struct *raid5_wq;
80 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
82 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
83 return &conf->stripe_hashtbl[hash];
86 static inline int stripe_hash_locks_hash(sector_t sect)
88 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
91 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
93 spin_lock_irq(conf->hash_locks + hash);
94 spin_lock(&conf->device_lock);
97 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
99 spin_unlock(&conf->device_lock);
100 spin_unlock_irq(conf->hash_locks + hash);
103 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
106 spin_lock_irq(conf->hash_locks);
107 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
108 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
109 spin_lock(&conf->device_lock);
112 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
115 spin_unlock(&conf->device_lock);
116 for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
117 spin_unlock(conf->hash_locks + i);
118 spin_unlock_irq(conf->hash_locks);
121 /* Find first data disk in a raid6 stripe */
122 static inline int raid6_d0(struct stripe_head *sh)
125 /* ddf always start from first device */
127 /* md starts just after Q block */
128 if (sh->qd_idx == sh->disks - 1)
131 return sh->qd_idx + 1;
133 static inline int raid6_next_disk(int disk, int raid_disks)
136 return (disk < raid_disks) ? disk : 0;
139 /* When walking through the disks in a raid5, starting at raid6_d0,
140 * We need to map each disk to a 'slot', where the data disks are slot
141 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
142 * is raid_disks-1. This help does that mapping.
144 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
145 int *count, int syndrome_disks)
151 if (idx == sh->pd_idx)
152 return syndrome_disks;
153 if (idx == sh->qd_idx)
154 return syndrome_disks + 1;
160 static void print_raid5_conf (struct r5conf *conf);
162 static int stripe_operations_active(struct stripe_head *sh)
164 return sh->check_state || sh->reconstruct_state ||
165 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
166 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
169 static bool stripe_is_lowprio(struct stripe_head *sh)
171 return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
172 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
173 !test_bit(STRIPE_R5C_CACHING, &sh->state);
176 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
178 struct r5conf *conf = sh->raid_conf;
179 struct r5worker_group *group;
181 int i, cpu = sh->cpu;
183 if (!cpu_online(cpu)) {
184 cpu = cpumask_any(cpu_online_mask);
188 if (list_empty(&sh->lru)) {
189 struct r5worker_group *group;
190 group = conf->worker_groups + cpu_to_group(cpu);
191 if (stripe_is_lowprio(sh))
192 list_add_tail(&sh->lru, &group->loprio_list);
194 list_add_tail(&sh->lru, &group->handle_list);
195 group->stripes_cnt++;
199 if (conf->worker_cnt_per_group == 0) {
200 md_wakeup_thread(conf->mddev->thread);
204 group = conf->worker_groups + cpu_to_group(sh->cpu);
206 group->workers[0].working = true;
207 /* at least one worker should run to avoid race */
208 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
210 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
211 /* wakeup more workers */
212 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
213 if (group->workers[i].working == false) {
214 group->workers[i].working = true;
215 queue_work_on(sh->cpu, raid5_wq,
216 &group->workers[i].work);
222 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
223 struct list_head *temp_inactive_list)
226 int injournal = 0; /* number of date pages with R5_InJournal */
228 BUG_ON(!list_empty(&sh->lru));
229 BUG_ON(atomic_read(&conf->active_stripes)==0);
231 if (r5c_is_writeback(conf->log))
232 for (i = sh->disks; i--; )
233 if (test_bit(R5_InJournal, &sh->dev[i].flags))
236 * In the following cases, the stripe cannot be released to cached
237 * lists. Therefore, we make the stripe write out and set
239 * 1. when quiesce in r5c write back;
240 * 2. when resync is requested fot the stripe.
242 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
243 (conf->quiesce && r5c_is_writeback(conf->log) &&
244 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
245 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
246 r5c_make_stripe_write_out(sh);
247 set_bit(STRIPE_HANDLE, &sh->state);
250 if (test_bit(STRIPE_HANDLE, &sh->state)) {
251 if (test_bit(STRIPE_DELAYED, &sh->state) &&
252 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
253 list_add_tail(&sh->lru, &conf->delayed_list);
254 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
255 sh->bm_seq - conf->seq_write > 0)
256 list_add_tail(&sh->lru, &conf->bitmap_list);
258 clear_bit(STRIPE_DELAYED, &sh->state);
259 clear_bit(STRIPE_BIT_DELAY, &sh->state);
260 if (conf->worker_cnt_per_group == 0) {
261 if (stripe_is_lowprio(sh))
262 list_add_tail(&sh->lru,
265 list_add_tail(&sh->lru,
268 raid5_wakeup_stripe_thread(sh);
272 md_wakeup_thread(conf->mddev->thread);
274 BUG_ON(stripe_operations_active(sh));
275 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
276 if (atomic_dec_return(&conf->preread_active_stripes)
278 md_wakeup_thread(conf->mddev->thread);
279 atomic_dec(&conf->active_stripes);
280 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
281 if (!r5c_is_writeback(conf->log))
282 list_add_tail(&sh->lru, temp_inactive_list);
284 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
286 list_add_tail(&sh->lru, temp_inactive_list);
287 else if (injournal == conf->raid_disks - conf->max_degraded) {
289 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
290 atomic_inc(&conf->r5c_cached_full_stripes);
291 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
292 atomic_dec(&conf->r5c_cached_partial_stripes);
293 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
294 r5c_check_cached_full_stripe(conf);
297 * STRIPE_R5C_PARTIAL_STRIPE is set in
298 * r5c_try_caching_write(). No need to
301 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
307 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
308 struct list_head *temp_inactive_list)
310 if (atomic_dec_and_test(&sh->count))
311 do_release_stripe(conf, sh, temp_inactive_list);
315 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
317 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
318 * given time. Adding stripes only takes device lock, while deleting stripes
319 * only takes hash lock.
321 static void release_inactive_stripe_list(struct r5conf *conf,
322 struct list_head *temp_inactive_list,
326 bool do_wakeup = false;
329 if (hash == NR_STRIPE_HASH_LOCKS) {
330 size = NR_STRIPE_HASH_LOCKS;
331 hash = NR_STRIPE_HASH_LOCKS - 1;
335 struct list_head *list = &temp_inactive_list[size - 1];
338 * We don't hold any lock here yet, raid5_get_active_stripe() might
339 * remove stripes from the list
341 if (!list_empty_careful(list)) {
342 spin_lock_irqsave(conf->hash_locks + hash, flags);
343 if (list_empty(conf->inactive_list + hash) &&
345 atomic_dec(&conf->empty_inactive_list_nr);
346 list_splice_tail_init(list, conf->inactive_list + hash);
348 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
355 wake_up(&conf->wait_for_stripe);
356 if (atomic_read(&conf->active_stripes) == 0)
357 wake_up(&conf->wait_for_quiescent);
358 if (conf->retry_read_aligned)
359 md_wakeup_thread(conf->mddev->thread);
363 /* should hold conf->device_lock already */
364 static int release_stripe_list(struct r5conf *conf,
365 struct list_head *temp_inactive_list)
367 struct stripe_head *sh, *t;
369 struct llist_node *head;
371 head = llist_del_all(&conf->released_stripes);
372 head = llist_reverse_order(head);
373 llist_for_each_entry_safe(sh, t, head, release_list) {
376 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
378 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
380 * Don't worry the bit is set here, because if the bit is set
381 * again, the count is always > 1. This is true for
382 * STRIPE_ON_UNPLUG_LIST bit too.
384 hash = sh->hash_lock_index;
385 __release_stripe(conf, sh, &temp_inactive_list[hash]);
392 void raid5_release_stripe(struct stripe_head *sh)
394 struct r5conf *conf = sh->raid_conf;
396 struct list_head list;
400 /* Avoid release_list until the last reference.
402 if (atomic_add_unless(&sh->count, -1, 1))
405 if (unlikely(!conf->mddev->thread) ||
406 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
408 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
410 md_wakeup_thread(conf->mddev->thread);
413 local_irq_save(flags);
414 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
415 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
416 INIT_LIST_HEAD(&list);
417 hash = sh->hash_lock_index;
418 do_release_stripe(conf, sh, &list);
419 spin_unlock(&conf->device_lock);
420 release_inactive_stripe_list(conf, &list, hash);
422 local_irq_restore(flags);
425 static inline void remove_hash(struct stripe_head *sh)
427 pr_debug("remove_hash(), stripe %llu\n",
428 (unsigned long long)sh->sector);
430 hlist_del_init(&sh->hash);
433 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
435 struct hlist_head *hp = stripe_hash(conf, sh->sector);
437 pr_debug("insert_hash(), stripe %llu\n",
438 (unsigned long long)sh->sector);
440 hlist_add_head(&sh->hash, hp);
443 /* find an idle stripe, make sure it is unhashed, and return it. */
444 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
446 struct stripe_head *sh = NULL;
447 struct list_head *first;
449 if (list_empty(conf->inactive_list + hash))
451 first = (conf->inactive_list + hash)->next;
452 sh = list_entry(first, struct stripe_head, lru);
453 list_del_init(first);
455 atomic_inc(&conf->active_stripes);
456 BUG_ON(hash != sh->hash_lock_index);
457 if (list_empty(conf->inactive_list + hash))
458 atomic_inc(&conf->empty_inactive_list_nr);
463 static void shrink_buffers(struct stripe_head *sh)
467 int num = sh->raid_conf->pool_size;
469 for (i = 0; i < num ; i++) {
470 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
474 sh->dev[i].page = NULL;
479 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
482 int num = sh->raid_conf->pool_size;
484 for (i = 0; i < num; i++) {
487 if (!(page = alloc_page(gfp))) {
490 sh->dev[i].page = page;
491 sh->dev[i].orig_page = page;
497 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
498 struct stripe_head *sh);
500 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
502 struct r5conf *conf = sh->raid_conf;
505 BUG_ON(atomic_read(&sh->count) != 0);
506 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
507 BUG_ON(stripe_operations_active(sh));
508 BUG_ON(sh->batch_head);
510 pr_debug("init_stripe called, stripe %llu\n",
511 (unsigned long long)sector);
513 seq = read_seqcount_begin(&conf->gen_lock);
514 sh->generation = conf->generation - previous;
515 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
517 stripe_set_idx(sector, conf, previous, sh);
520 for (i = sh->disks; i--; ) {
521 struct r5dev *dev = &sh->dev[i];
523 if (dev->toread || dev->read || dev->towrite || dev->written ||
524 test_bit(R5_LOCKED, &dev->flags)) {
525 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
526 (unsigned long long)sh->sector, i, dev->toread,
527 dev->read, dev->towrite, dev->written,
528 test_bit(R5_LOCKED, &dev->flags));
532 dev->sector = raid5_compute_blocknr(sh, i, previous);
534 if (read_seqcount_retry(&conf->gen_lock, seq))
536 sh->overwrite_disks = 0;
537 insert_hash(conf, sh);
538 sh->cpu = smp_processor_id();
539 set_bit(STRIPE_BATCH_READY, &sh->state);
542 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
545 struct stripe_head *sh;
547 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
548 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
549 if (sh->sector == sector && sh->generation == generation)
551 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
556 * Need to check if array has failed when deciding whether to:
558 * - remove non-faulty devices
561 * This determination is simple when no reshape is happening.
562 * However if there is a reshape, we need to carefully check
563 * both the before and after sections.
564 * This is because some failed devices may only affect one
565 * of the two sections, and some non-in_sync devices may
566 * be insync in the section most affected by failed devices.
568 int raid5_calc_degraded(struct r5conf *conf)
570 int degraded, degraded2;
575 for (i = 0; i < conf->previous_raid_disks; i++) {
576 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
577 if (rdev && test_bit(Faulty, &rdev->flags))
578 rdev = rcu_dereference(conf->disks[i].replacement);
579 if (!rdev || test_bit(Faulty, &rdev->flags))
581 else if (test_bit(In_sync, &rdev->flags))
584 /* not in-sync or faulty.
585 * If the reshape increases the number of devices,
586 * this is being recovered by the reshape, so
587 * this 'previous' section is not in_sync.
588 * If the number of devices is being reduced however,
589 * the device can only be part of the array if
590 * we are reverting a reshape, so this section will
593 if (conf->raid_disks >= conf->previous_raid_disks)
597 if (conf->raid_disks == conf->previous_raid_disks)
601 for (i = 0; i < conf->raid_disks; i++) {
602 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
603 if (rdev && test_bit(Faulty, &rdev->flags))
604 rdev = rcu_dereference(conf->disks[i].replacement);
605 if (!rdev || test_bit(Faulty, &rdev->flags))
607 else if (test_bit(In_sync, &rdev->flags))
610 /* not in-sync or faulty.
611 * If reshape increases the number of devices, this
612 * section has already been recovered, else it
613 * almost certainly hasn't.
615 if (conf->raid_disks <= conf->previous_raid_disks)
619 if (degraded2 > degraded)
624 static int has_failed(struct r5conf *conf)
628 if (conf->mddev->reshape_position == MaxSector)
629 return conf->mddev->degraded > conf->max_degraded;
631 degraded = raid5_calc_degraded(conf);
632 if (degraded > conf->max_degraded)
638 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
639 int previous, int noblock, int noquiesce)
641 struct stripe_head *sh;
642 int hash = stripe_hash_locks_hash(sector);
643 int inc_empty_inactive_list_flag;
645 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
647 spin_lock_irq(conf->hash_locks + hash);
650 wait_event_lock_irq(conf->wait_for_quiescent,
651 conf->quiesce == 0 || noquiesce,
652 *(conf->hash_locks + hash));
653 sh = __find_stripe(conf, sector, conf->generation - previous);
655 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
656 sh = get_free_stripe(conf, hash);
657 if (!sh && !test_bit(R5_DID_ALLOC,
659 set_bit(R5_ALLOC_MORE,
662 if (noblock && sh == NULL)
665 r5c_check_stripe_cache_usage(conf);
667 set_bit(R5_INACTIVE_BLOCKED,
669 r5l_wake_reclaim(conf->log, 0);
671 conf->wait_for_stripe,
672 !list_empty(conf->inactive_list + hash) &&
673 (atomic_read(&conf->active_stripes)
674 < (conf->max_nr_stripes * 3 / 4)
675 || !test_bit(R5_INACTIVE_BLOCKED,
676 &conf->cache_state)),
677 *(conf->hash_locks + hash));
678 clear_bit(R5_INACTIVE_BLOCKED,
681 init_stripe(sh, sector, previous);
682 atomic_inc(&sh->count);
684 } else if (!atomic_inc_not_zero(&sh->count)) {
685 spin_lock(&conf->device_lock);
686 if (!atomic_read(&sh->count)) {
687 if (!test_bit(STRIPE_HANDLE, &sh->state))
688 atomic_inc(&conf->active_stripes);
689 BUG_ON(list_empty(&sh->lru) &&
690 !test_bit(STRIPE_EXPANDING, &sh->state));
691 inc_empty_inactive_list_flag = 0;
692 if (!list_empty(conf->inactive_list + hash))
693 inc_empty_inactive_list_flag = 1;
694 list_del_init(&sh->lru);
695 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
696 atomic_inc(&conf->empty_inactive_list_nr);
698 sh->group->stripes_cnt--;
702 atomic_inc(&sh->count);
703 spin_unlock(&conf->device_lock);
705 } while (sh == NULL);
707 spin_unlock_irq(conf->hash_locks + hash);
711 static bool is_full_stripe_write(struct stripe_head *sh)
713 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
714 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
717 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
720 spin_lock_irq(&sh2->stripe_lock);
721 spin_lock_nested(&sh1->stripe_lock, 1);
723 spin_lock_irq(&sh1->stripe_lock);
724 spin_lock_nested(&sh2->stripe_lock, 1);
728 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
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;
1144 * If this is discard request, set bi_vcnt 0. We don't
1145 * want to confuse SCSI because SCSI will replace payload
1147 if (op == REQ_OP_DISCARD)
1150 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1152 if (conf->mddev->gendisk)
1153 trace_block_bio_remap(bi->bi_disk->queue,
1154 bi, disk_devt(conf->mddev->gendisk),
1156 if (should_defer && op_is_write(op))
1157 bio_list_add(&pending_bios, bi);
1159 generic_make_request(bi);
1162 if (s->syncing || s->expanding || s->expanded
1164 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1166 set_bit(STRIPE_IO_STARTED, &sh->state);
1168 bio_set_dev(rbi, rrdev->bdev);
1169 bio_set_op_attrs(rbi, op, op_flags);
1170 BUG_ON(!op_is_write(op));
1171 rbi->bi_end_io = raid5_end_write_request;
1172 rbi->bi_private = sh;
1174 pr_debug("%s: for %llu schedule op %d on "
1175 "replacement disc %d\n",
1176 __func__, (unsigned long long)sh->sector,
1178 atomic_inc(&sh->count);
1180 atomic_inc(&head_sh->count);
1181 if (use_new_offset(conf, sh))
1182 rbi->bi_iter.bi_sector = (sh->sector
1183 + rrdev->new_data_offset);
1185 rbi->bi_iter.bi_sector = (sh->sector
1186 + rrdev->data_offset);
1187 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1188 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1189 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1191 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1192 rbi->bi_io_vec[0].bv_offset = 0;
1193 rbi->bi_iter.bi_size = STRIPE_SIZE;
1195 * If this is discard request, set bi_vcnt 0. We don't
1196 * want to confuse SCSI because SCSI will replace payload
1198 if (op == REQ_OP_DISCARD)
1200 if (conf->mddev->gendisk)
1201 trace_block_bio_remap(rbi->bi_disk->queue,
1202 rbi, disk_devt(conf->mddev->gendisk),
1204 if (should_defer && op_is_write(op))
1205 bio_list_add(&pending_bios, rbi);
1207 generic_make_request(rbi);
1209 if (!rdev && !rrdev) {
1210 if (op_is_write(op))
1211 set_bit(STRIPE_DEGRADED, &sh->state);
1212 pr_debug("skip op %d on disc %d for sector %llu\n",
1213 bi->bi_opf, i, (unsigned long long)sh->sector);
1214 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1215 set_bit(STRIPE_HANDLE, &sh->state);
1218 if (!head_sh->batch_head)
1220 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1226 if (should_defer && !bio_list_empty(&pending_bios))
1227 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1230 static struct dma_async_tx_descriptor *
1231 async_copy_data(int frombio, struct bio *bio, struct page **page,
1232 sector_t sector, struct dma_async_tx_descriptor *tx,
1233 struct stripe_head *sh, int no_skipcopy)
1236 struct bvec_iter iter;
1237 struct page *bio_page;
1239 struct async_submit_ctl submit;
1240 enum async_tx_flags flags = 0;
1242 if (bio->bi_iter.bi_sector >= sector)
1243 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1245 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1248 flags |= ASYNC_TX_FENCE;
1249 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1251 bio_for_each_segment(bvl, bio, iter) {
1252 int len = bvl.bv_len;
1256 if (page_offset < 0) {
1257 b_offset = -page_offset;
1258 page_offset += b_offset;
1262 if (len > 0 && page_offset + len > STRIPE_SIZE)
1263 clen = STRIPE_SIZE - page_offset;
1268 b_offset += bvl.bv_offset;
1269 bio_page = bvl.bv_page;
1271 if (sh->raid_conf->skip_copy &&
1272 b_offset == 0 && page_offset == 0 &&
1273 clen == STRIPE_SIZE &&
1277 tx = async_memcpy(*page, bio_page, page_offset,
1278 b_offset, clen, &submit);
1280 tx = async_memcpy(bio_page, *page, b_offset,
1281 page_offset, clen, &submit);
1283 /* chain the operations */
1284 submit.depend_tx = tx;
1286 if (clen < len) /* hit end of page */
1294 static void ops_complete_biofill(void *stripe_head_ref)
1296 struct stripe_head *sh = stripe_head_ref;
1299 pr_debug("%s: stripe %llu\n", __func__,
1300 (unsigned long long)sh->sector);
1302 /* clear completed biofills */
1303 for (i = sh->disks; i--; ) {
1304 struct r5dev *dev = &sh->dev[i];
1306 /* acknowledge completion of a biofill operation */
1307 /* and check if we need to reply to a read request,
1308 * new R5_Wantfill requests are held off until
1309 * !STRIPE_BIOFILL_RUN
1311 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1312 struct bio *rbi, *rbi2;
1317 while (rbi && rbi->bi_iter.bi_sector <
1318 dev->sector + STRIPE_SECTORS) {
1319 rbi2 = r5_next_bio(rbi, dev->sector);
1325 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1327 set_bit(STRIPE_HANDLE, &sh->state);
1328 raid5_release_stripe(sh);
1331 static void ops_run_biofill(struct stripe_head *sh)
1333 struct dma_async_tx_descriptor *tx = NULL;
1334 struct async_submit_ctl submit;
1337 BUG_ON(sh->batch_head);
1338 pr_debug("%s: stripe %llu\n", __func__,
1339 (unsigned long long)sh->sector);
1341 for (i = sh->disks; i--; ) {
1342 struct r5dev *dev = &sh->dev[i];
1343 if (test_bit(R5_Wantfill, &dev->flags)) {
1345 spin_lock_irq(&sh->stripe_lock);
1346 dev->read = rbi = dev->toread;
1348 spin_unlock_irq(&sh->stripe_lock);
1349 while (rbi && rbi->bi_iter.bi_sector <
1350 dev->sector + STRIPE_SECTORS) {
1351 tx = async_copy_data(0, rbi, &dev->page,
1352 dev->sector, tx, sh, 0);
1353 rbi = r5_next_bio(rbi, dev->sector);
1358 atomic_inc(&sh->count);
1359 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1360 async_trigger_callback(&submit);
1363 static void mark_target_uptodate(struct stripe_head *sh, int target)
1370 tgt = &sh->dev[target];
1371 set_bit(R5_UPTODATE, &tgt->flags);
1372 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1373 clear_bit(R5_Wantcompute, &tgt->flags);
1376 static void ops_complete_compute(void *stripe_head_ref)
1378 struct stripe_head *sh = stripe_head_ref;
1380 pr_debug("%s: stripe %llu\n", __func__,
1381 (unsigned long long)sh->sector);
1383 /* mark the computed target(s) as uptodate */
1384 mark_target_uptodate(sh, sh->ops.target);
1385 mark_target_uptodate(sh, sh->ops.target2);
1387 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1388 if (sh->check_state == check_state_compute_run)
1389 sh->check_state = check_state_compute_result;
1390 set_bit(STRIPE_HANDLE, &sh->state);
1391 raid5_release_stripe(sh);
1394 /* return a pointer to the address conversion region of the scribble buffer */
1395 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1396 struct raid5_percpu *percpu, int i)
1400 addr = flex_array_get(percpu->scribble, i);
1401 return addr + sizeof(struct page *) * (sh->disks + 2);
1404 /* return a pointer to the address conversion region of the scribble buffer */
1405 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1409 addr = flex_array_get(percpu->scribble, i);
1413 static struct dma_async_tx_descriptor *
1414 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1416 int disks = sh->disks;
1417 struct page **xor_srcs = to_addr_page(percpu, 0);
1418 int target = sh->ops.target;
1419 struct r5dev *tgt = &sh->dev[target];
1420 struct page *xor_dest = tgt->page;
1422 struct dma_async_tx_descriptor *tx;
1423 struct async_submit_ctl submit;
1426 BUG_ON(sh->batch_head);
1428 pr_debug("%s: stripe %llu block: %d\n",
1429 __func__, (unsigned long long)sh->sector, target);
1430 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1432 for (i = disks; i--; )
1434 xor_srcs[count++] = sh->dev[i].page;
1436 atomic_inc(&sh->count);
1438 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1439 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1440 if (unlikely(count == 1))
1441 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1443 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1448 /* set_syndrome_sources - populate source buffers for gen_syndrome
1449 * @srcs - (struct page *) array of size sh->disks
1450 * @sh - stripe_head to parse
1452 * Populates srcs in proper layout order for the stripe and returns the
1453 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1454 * destination buffer is recorded in srcs[count] and the Q destination
1455 * is recorded in srcs[count+1]].
1457 static int set_syndrome_sources(struct page **srcs,
1458 struct stripe_head *sh,
1461 int disks = sh->disks;
1462 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1463 int d0_idx = raid6_d0(sh);
1467 for (i = 0; i < disks; i++)
1473 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1474 struct r5dev *dev = &sh->dev[i];
1476 if (i == sh->qd_idx || i == sh->pd_idx ||
1477 (srctype == SYNDROME_SRC_ALL) ||
1478 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1479 (test_bit(R5_Wantdrain, &dev->flags) ||
1480 test_bit(R5_InJournal, &dev->flags))) ||
1481 (srctype == SYNDROME_SRC_WRITTEN &&
1483 test_bit(R5_InJournal, &dev->flags)))) {
1484 if (test_bit(R5_InJournal, &dev->flags))
1485 srcs[slot] = sh->dev[i].orig_page;
1487 srcs[slot] = sh->dev[i].page;
1489 i = raid6_next_disk(i, disks);
1490 } while (i != d0_idx);
1492 return syndrome_disks;
1495 static struct dma_async_tx_descriptor *
1496 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1498 int disks = sh->disks;
1499 struct page **blocks = to_addr_page(percpu, 0);
1501 int qd_idx = sh->qd_idx;
1502 struct dma_async_tx_descriptor *tx;
1503 struct async_submit_ctl submit;
1509 BUG_ON(sh->batch_head);
1510 if (sh->ops.target < 0)
1511 target = sh->ops.target2;
1512 else if (sh->ops.target2 < 0)
1513 target = sh->ops.target;
1515 /* we should only have one valid target */
1518 pr_debug("%s: stripe %llu block: %d\n",
1519 __func__, (unsigned long long)sh->sector, target);
1521 tgt = &sh->dev[target];
1522 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1525 atomic_inc(&sh->count);
1527 if (target == qd_idx) {
1528 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1529 blocks[count] = NULL; /* regenerating p is not necessary */
1530 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1531 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1532 ops_complete_compute, sh,
1533 to_addr_conv(sh, percpu, 0));
1534 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1536 /* Compute any data- or p-drive using XOR */
1538 for (i = disks; i-- ; ) {
1539 if (i == target || i == qd_idx)
1541 blocks[count++] = sh->dev[i].page;
1544 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1545 NULL, ops_complete_compute, sh,
1546 to_addr_conv(sh, percpu, 0));
1547 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1553 static struct dma_async_tx_descriptor *
1554 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1556 int i, count, disks = sh->disks;
1557 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1558 int d0_idx = raid6_d0(sh);
1559 int faila = -1, failb = -1;
1560 int target = sh->ops.target;
1561 int target2 = sh->ops.target2;
1562 struct r5dev *tgt = &sh->dev[target];
1563 struct r5dev *tgt2 = &sh->dev[target2];
1564 struct dma_async_tx_descriptor *tx;
1565 struct page **blocks = to_addr_page(percpu, 0);
1566 struct async_submit_ctl submit;
1568 BUG_ON(sh->batch_head);
1569 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1570 __func__, (unsigned long long)sh->sector, target, target2);
1571 BUG_ON(target < 0 || target2 < 0);
1572 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1573 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1575 /* we need to open-code set_syndrome_sources to handle the
1576 * slot number conversion for 'faila' and 'failb'
1578 for (i = 0; i < disks ; i++)
1583 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1585 blocks[slot] = sh->dev[i].page;
1591 i = raid6_next_disk(i, disks);
1592 } while (i != d0_idx);
1594 BUG_ON(faila == failb);
1597 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1598 __func__, (unsigned long long)sh->sector, faila, failb);
1600 atomic_inc(&sh->count);
1602 if (failb == syndrome_disks+1) {
1603 /* Q disk is one of the missing disks */
1604 if (faila == syndrome_disks) {
1605 /* Missing P+Q, just recompute */
1606 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1607 ops_complete_compute, sh,
1608 to_addr_conv(sh, percpu, 0));
1609 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1610 STRIPE_SIZE, &submit);
1614 int qd_idx = sh->qd_idx;
1616 /* Missing D+Q: recompute D from P, then recompute Q */
1617 if (target == qd_idx)
1618 data_target = target2;
1620 data_target = target;
1623 for (i = disks; i-- ; ) {
1624 if (i == data_target || i == qd_idx)
1626 blocks[count++] = sh->dev[i].page;
1628 dest = sh->dev[data_target].page;
1629 init_async_submit(&submit,
1630 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1632 to_addr_conv(sh, percpu, 0));
1633 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1636 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1637 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1638 ops_complete_compute, sh,
1639 to_addr_conv(sh, percpu, 0));
1640 return async_gen_syndrome(blocks, 0, count+2,
1641 STRIPE_SIZE, &submit);
1644 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1645 ops_complete_compute, sh,
1646 to_addr_conv(sh, percpu, 0));
1647 if (failb == syndrome_disks) {
1648 /* We're missing D+P. */
1649 return async_raid6_datap_recov(syndrome_disks+2,
1653 /* We're missing D+D. */
1654 return async_raid6_2data_recov(syndrome_disks+2,
1655 STRIPE_SIZE, faila, failb,
1661 static void ops_complete_prexor(void *stripe_head_ref)
1663 struct stripe_head *sh = stripe_head_ref;
1665 pr_debug("%s: stripe %llu\n", __func__,
1666 (unsigned long long)sh->sector);
1668 if (r5c_is_writeback(sh->raid_conf->log))
1670 * raid5-cache write back uses orig_page during prexor.
1671 * After prexor, it is time to free orig_page
1673 r5c_release_extra_page(sh);
1676 static struct dma_async_tx_descriptor *
1677 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1678 struct dma_async_tx_descriptor *tx)
1680 int disks = sh->disks;
1681 struct page **xor_srcs = to_addr_page(percpu, 0);
1682 int count = 0, pd_idx = sh->pd_idx, i;
1683 struct async_submit_ctl submit;
1685 /* existing parity data subtracted */
1686 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1688 BUG_ON(sh->batch_head);
1689 pr_debug("%s: stripe %llu\n", __func__,
1690 (unsigned long long)sh->sector);
1692 for (i = disks; i--; ) {
1693 struct r5dev *dev = &sh->dev[i];
1694 /* Only process blocks that are known to be uptodate */
1695 if (test_bit(R5_InJournal, &dev->flags))
1696 xor_srcs[count++] = dev->orig_page;
1697 else if (test_bit(R5_Wantdrain, &dev->flags))
1698 xor_srcs[count++] = dev->page;
1701 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1702 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1703 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1708 static struct dma_async_tx_descriptor *
1709 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1710 struct dma_async_tx_descriptor *tx)
1712 struct page **blocks = to_addr_page(percpu, 0);
1714 struct async_submit_ctl submit;
1716 pr_debug("%s: stripe %llu\n", __func__,
1717 (unsigned long long)sh->sector);
1719 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1721 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1722 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1723 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1728 static struct dma_async_tx_descriptor *
1729 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1731 struct r5conf *conf = sh->raid_conf;
1732 int disks = sh->disks;
1734 struct stripe_head *head_sh = sh;
1736 pr_debug("%s: stripe %llu\n", __func__,
1737 (unsigned long long)sh->sector);
1739 for (i = disks; i--; ) {
1744 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1750 * clear R5_InJournal, so when rewriting a page in
1751 * journal, it is not skipped by r5l_log_stripe()
1753 clear_bit(R5_InJournal, &dev->flags);
1754 spin_lock_irq(&sh->stripe_lock);
1755 chosen = dev->towrite;
1756 dev->towrite = NULL;
1757 sh->overwrite_disks = 0;
1758 BUG_ON(dev->written);
1759 wbi = dev->written = chosen;
1760 spin_unlock_irq(&sh->stripe_lock);
1761 WARN_ON(dev->page != dev->orig_page);
1763 while (wbi && wbi->bi_iter.bi_sector <
1764 dev->sector + STRIPE_SECTORS) {
1765 if (wbi->bi_opf & REQ_FUA)
1766 set_bit(R5_WantFUA, &dev->flags);
1767 if (wbi->bi_opf & REQ_SYNC)
1768 set_bit(R5_SyncIO, &dev->flags);
1769 if (bio_op(wbi) == REQ_OP_DISCARD)
1770 set_bit(R5_Discard, &dev->flags);
1772 tx = async_copy_data(1, wbi, &dev->page,
1773 dev->sector, tx, sh,
1774 r5c_is_writeback(conf->log));
1775 if (dev->page != dev->orig_page &&
1776 !r5c_is_writeback(conf->log)) {
1777 set_bit(R5_SkipCopy, &dev->flags);
1778 clear_bit(R5_UPTODATE, &dev->flags);
1779 clear_bit(R5_OVERWRITE, &dev->flags);
1782 wbi = r5_next_bio(wbi, dev->sector);
1785 if (head_sh->batch_head) {
1786 sh = list_first_entry(&sh->batch_list,
1799 static void ops_complete_reconstruct(void *stripe_head_ref)
1801 struct stripe_head *sh = stripe_head_ref;
1802 int disks = sh->disks;
1803 int pd_idx = sh->pd_idx;
1804 int qd_idx = sh->qd_idx;
1806 bool fua = false, sync = false, discard = false;
1808 pr_debug("%s: stripe %llu\n", __func__,
1809 (unsigned long long)sh->sector);
1811 for (i = disks; i--; ) {
1812 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1813 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1814 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1817 for (i = disks; i--; ) {
1818 struct r5dev *dev = &sh->dev[i];
1820 if (dev->written || i == pd_idx || i == qd_idx) {
1821 if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1822 set_bit(R5_UPTODATE, &dev->flags);
1823 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1824 set_bit(R5_Expanded, &dev->flags);
1827 set_bit(R5_WantFUA, &dev->flags);
1829 set_bit(R5_SyncIO, &dev->flags);
1833 if (sh->reconstruct_state == reconstruct_state_drain_run)
1834 sh->reconstruct_state = reconstruct_state_drain_result;
1835 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1836 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1838 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1839 sh->reconstruct_state = reconstruct_state_result;
1842 set_bit(STRIPE_HANDLE, &sh->state);
1843 raid5_release_stripe(sh);
1847 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1848 struct dma_async_tx_descriptor *tx)
1850 int disks = sh->disks;
1851 struct page **xor_srcs;
1852 struct async_submit_ctl submit;
1853 int count, pd_idx = sh->pd_idx, i;
1854 struct page *xor_dest;
1856 unsigned long flags;
1858 struct stripe_head *head_sh = sh;
1861 pr_debug("%s: stripe %llu\n", __func__,
1862 (unsigned long long)sh->sector);
1864 for (i = 0; i < sh->disks; i++) {
1867 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1870 if (i >= sh->disks) {
1871 atomic_inc(&sh->count);
1872 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1873 ops_complete_reconstruct(sh);
1878 xor_srcs = to_addr_page(percpu, j);
1879 /* check if prexor is active which means only process blocks
1880 * that are part of a read-modify-write (written)
1882 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1884 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1885 for (i = disks; i--; ) {
1886 struct r5dev *dev = &sh->dev[i];
1887 if (head_sh->dev[i].written ||
1888 test_bit(R5_InJournal, &head_sh->dev[i].flags))
1889 xor_srcs[count++] = dev->page;
1892 xor_dest = sh->dev[pd_idx].page;
1893 for (i = disks; i--; ) {
1894 struct r5dev *dev = &sh->dev[i];
1896 xor_srcs[count++] = dev->page;
1900 /* 1/ if we prexor'd then the dest is reused as a source
1901 * 2/ if we did not prexor then we are redoing the parity
1902 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1903 * for the synchronous xor case
1905 last_stripe = !head_sh->batch_head ||
1906 list_first_entry(&sh->batch_list,
1907 struct stripe_head, batch_list) == head_sh;
1909 flags = ASYNC_TX_ACK |
1910 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1912 atomic_inc(&head_sh->count);
1913 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1914 to_addr_conv(sh, percpu, j));
1916 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1917 init_async_submit(&submit, flags, tx, NULL, NULL,
1918 to_addr_conv(sh, percpu, j));
1921 if (unlikely(count == 1))
1922 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1924 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1927 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1934 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1935 struct dma_async_tx_descriptor *tx)
1937 struct async_submit_ctl submit;
1938 struct page **blocks;
1939 int count, i, j = 0;
1940 struct stripe_head *head_sh = sh;
1943 unsigned long txflags;
1945 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1947 for (i = 0; i < sh->disks; i++) {
1948 if (sh->pd_idx == i || sh->qd_idx == i)
1950 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1953 if (i >= sh->disks) {
1954 atomic_inc(&sh->count);
1955 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1956 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1957 ops_complete_reconstruct(sh);
1962 blocks = to_addr_page(percpu, j);
1964 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1965 synflags = SYNDROME_SRC_WRITTEN;
1966 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1968 synflags = SYNDROME_SRC_ALL;
1969 txflags = ASYNC_TX_ACK;
1972 count = set_syndrome_sources(blocks, sh, synflags);
1973 last_stripe = !head_sh->batch_head ||
1974 list_first_entry(&sh->batch_list,
1975 struct stripe_head, batch_list) == head_sh;
1978 atomic_inc(&head_sh->count);
1979 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1980 head_sh, to_addr_conv(sh, percpu, j));
1982 init_async_submit(&submit, 0, tx, NULL, NULL,
1983 to_addr_conv(sh, percpu, j));
1984 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1987 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1993 static void ops_complete_check(void *stripe_head_ref)
1995 struct stripe_head *sh = stripe_head_ref;
1997 pr_debug("%s: stripe %llu\n", __func__,
1998 (unsigned long long)sh->sector);
2000 sh->check_state = check_state_check_result;
2001 set_bit(STRIPE_HANDLE, &sh->state);
2002 raid5_release_stripe(sh);
2005 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2007 int disks = sh->disks;
2008 int pd_idx = sh->pd_idx;
2009 int qd_idx = sh->qd_idx;
2010 struct page *xor_dest;
2011 struct page **xor_srcs = to_addr_page(percpu, 0);
2012 struct dma_async_tx_descriptor *tx;
2013 struct async_submit_ctl submit;
2017 pr_debug("%s: stripe %llu\n", __func__,
2018 (unsigned long long)sh->sector);
2020 BUG_ON(sh->batch_head);
2022 xor_dest = sh->dev[pd_idx].page;
2023 xor_srcs[count++] = xor_dest;
2024 for (i = disks; i--; ) {
2025 if (i == pd_idx || i == qd_idx)
2027 xor_srcs[count++] = sh->dev[i].page;
2030 init_async_submit(&submit, 0, NULL, NULL, NULL,
2031 to_addr_conv(sh, percpu, 0));
2032 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2033 &sh->ops.zero_sum_result, &submit);
2035 atomic_inc(&sh->count);
2036 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2037 tx = async_trigger_callback(&submit);
2040 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2042 struct page **srcs = to_addr_page(percpu, 0);
2043 struct async_submit_ctl submit;
2046 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2047 (unsigned long long)sh->sector, checkp);
2049 BUG_ON(sh->batch_head);
2050 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2054 atomic_inc(&sh->count);
2055 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2056 sh, to_addr_conv(sh, percpu, 0));
2057 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2058 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2061 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2063 int overlap_clear = 0, i, disks = sh->disks;
2064 struct dma_async_tx_descriptor *tx = NULL;
2065 struct r5conf *conf = sh->raid_conf;
2066 int level = conf->level;
2067 struct raid5_percpu *percpu;
2071 percpu = per_cpu_ptr(conf->percpu, cpu);
2072 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2073 ops_run_biofill(sh);
2077 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2079 tx = ops_run_compute5(sh, percpu);
2081 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2082 tx = ops_run_compute6_1(sh, percpu);
2084 tx = ops_run_compute6_2(sh, percpu);
2086 /* terminate the chain if reconstruct is not set to be run */
2087 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2091 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2093 tx = ops_run_prexor5(sh, percpu, tx);
2095 tx = ops_run_prexor6(sh, percpu, tx);
2098 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2099 tx = ops_run_partial_parity(sh, percpu, tx);
2101 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2102 tx = ops_run_biodrain(sh, tx);
2106 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2108 ops_run_reconstruct5(sh, percpu, tx);
2110 ops_run_reconstruct6(sh, percpu, tx);
2113 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2114 if (sh->check_state == check_state_run)
2115 ops_run_check_p(sh, percpu);
2116 else if (sh->check_state == check_state_run_q)
2117 ops_run_check_pq(sh, percpu, 0);
2118 else if (sh->check_state == check_state_run_pq)
2119 ops_run_check_pq(sh, percpu, 1);
2124 if (overlap_clear && !sh->batch_head)
2125 for (i = disks; i--; ) {
2126 struct r5dev *dev = &sh->dev[i];
2127 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2128 wake_up(&sh->raid_conf->wait_for_overlap);
2133 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2136 __free_page(sh->ppl_page);
2137 kmem_cache_free(sc, sh);
2140 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2141 int disks, struct r5conf *conf)
2143 struct stripe_head *sh;
2146 sh = kmem_cache_zalloc(sc, gfp);
2148 spin_lock_init(&sh->stripe_lock);
2149 spin_lock_init(&sh->batch_lock);
2150 INIT_LIST_HEAD(&sh->batch_list);
2151 INIT_LIST_HEAD(&sh->lru);
2152 INIT_LIST_HEAD(&sh->r5c);
2153 INIT_LIST_HEAD(&sh->log_list);
2154 atomic_set(&sh->count, 1);
2155 sh->raid_conf = conf;
2156 sh->log_start = MaxSector;
2157 for (i = 0; i < disks; i++) {
2158 struct r5dev *dev = &sh->dev[i];
2160 bio_init(&dev->req, &dev->vec, 1);
2161 bio_init(&dev->rreq, &dev->rvec, 1);
2164 if (raid5_has_ppl(conf)) {
2165 sh->ppl_page = alloc_page(gfp);
2166 if (!sh->ppl_page) {
2167 free_stripe(sc, sh);
2174 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2176 struct stripe_head *sh;
2178 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2182 if (grow_buffers(sh, gfp)) {
2184 free_stripe(conf->slab_cache, sh);
2187 sh->hash_lock_index =
2188 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2189 /* we just created an active stripe so... */
2190 atomic_inc(&conf->active_stripes);
2192 raid5_release_stripe(sh);
2193 conf->max_nr_stripes++;
2197 static int grow_stripes(struct r5conf *conf, int num)
2199 struct kmem_cache *sc;
2200 size_t namelen = sizeof(conf->cache_name[0]);
2201 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2203 if (conf->mddev->gendisk)
2204 snprintf(conf->cache_name[0], namelen,
2205 "raid%d-%s", conf->level, mdname(conf->mddev));
2207 snprintf(conf->cache_name[0], namelen,
2208 "raid%d-%p", conf->level, conf->mddev);
2209 snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2211 conf->active_name = 0;
2212 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2213 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2217 conf->slab_cache = sc;
2218 conf->pool_size = devs;
2220 if (!grow_one_stripe(conf, GFP_KERNEL))
2227 * scribble_len - return the required size of the scribble region
2228 * @num - total number of disks in the array
2230 * The size must be enough to contain:
2231 * 1/ a struct page pointer for each device in the array +2
2232 * 2/ room to convert each entry in (1) to its corresponding dma
2233 * (dma_map_page()) or page (page_address()) address.
2235 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2236 * calculate over all devices (not just the data blocks), using zeros in place
2237 * of the P and Q blocks.
2239 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2241 struct flex_array *ret;
2244 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2245 ret = flex_array_alloc(len, cnt, flags);
2248 /* always prealloc all elements, so no locking is required */
2249 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2250 flex_array_free(ret);
2256 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2262 * Never shrink. And mddev_suspend() could deadlock if this is called
2263 * from raid5d. In that case, scribble_disks and scribble_sectors
2264 * should equal to new_disks and new_sectors
2266 if (conf->scribble_disks >= new_disks &&
2267 conf->scribble_sectors >= new_sectors)
2269 mddev_suspend(conf->mddev);
2271 for_each_present_cpu(cpu) {
2272 struct raid5_percpu *percpu;
2273 struct flex_array *scribble;
2275 percpu = per_cpu_ptr(conf->percpu, cpu);
2276 scribble = scribble_alloc(new_disks,
2277 new_sectors / STRIPE_SECTORS,
2281 flex_array_free(percpu->scribble);
2282 percpu->scribble = scribble;
2289 mddev_resume(conf->mddev);
2291 conf->scribble_disks = new_disks;
2292 conf->scribble_sectors = new_sectors;
2297 static int resize_stripes(struct r5conf *conf, int newsize)
2299 /* Make all the stripes able to hold 'newsize' devices.
2300 * New slots in each stripe get 'page' set to a new page.
2302 * This happens in stages:
2303 * 1/ create a new kmem_cache and allocate the required number of
2305 * 2/ gather all the old stripe_heads and transfer the pages across
2306 * to the new stripe_heads. This will have the side effect of
2307 * freezing the array as once all stripe_heads have been collected,
2308 * no IO will be possible. Old stripe heads are freed once their
2309 * pages have been transferred over, and the old kmem_cache is
2310 * freed when all stripes are done.
2311 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2312 * we simple return a failure status - no need to clean anything up.
2313 * 4/ allocate new pages for the new slots in the new stripe_heads.
2314 * If this fails, we don't bother trying the shrink the
2315 * stripe_heads down again, we just leave them as they are.
2316 * As each stripe_head is processed the new one is released into
2319 * Once step2 is started, we cannot afford to wait for a write,
2320 * so we use GFP_NOIO allocations.
2322 struct stripe_head *osh, *nsh;
2323 LIST_HEAD(newstripes);
2324 struct disk_info *ndisks;
2326 struct kmem_cache *sc;
2330 md_allow_write(conf->mddev);
2333 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2334 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2339 /* Need to ensure auto-resizing doesn't interfere */
2340 mutex_lock(&conf->cache_size_mutex);
2342 for (i = conf->max_nr_stripes; i; i--) {
2343 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2347 list_add(&nsh->lru, &newstripes);
2350 /* didn't get enough, give up */
2351 while (!list_empty(&newstripes)) {
2352 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2353 list_del(&nsh->lru);
2354 free_stripe(sc, nsh);
2356 kmem_cache_destroy(sc);
2357 mutex_unlock(&conf->cache_size_mutex);
2360 /* Step 2 - Must use GFP_NOIO now.
2361 * OK, we have enough stripes, start collecting inactive
2362 * stripes and copying them over
2366 list_for_each_entry(nsh, &newstripes, lru) {
2367 lock_device_hash_lock(conf, hash);
2368 wait_event_cmd(conf->wait_for_stripe,
2369 !list_empty(conf->inactive_list + hash),
2370 unlock_device_hash_lock(conf, hash),
2371 lock_device_hash_lock(conf, hash));
2372 osh = get_free_stripe(conf, hash);
2373 unlock_device_hash_lock(conf, hash);
2375 for(i=0; i<conf->pool_size; i++) {
2376 nsh->dev[i].page = osh->dev[i].page;
2377 nsh->dev[i].orig_page = osh->dev[i].page;
2379 nsh->hash_lock_index = hash;
2380 free_stripe(conf->slab_cache, osh);
2382 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2383 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2388 kmem_cache_destroy(conf->slab_cache);
2391 * At this point, we are holding all the stripes so the array
2392 * is completely stalled, so now is a good time to resize
2393 * conf->disks and the scribble region
2395 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2397 for (i = 0; i < conf->pool_size; i++)
2398 ndisks[i] = conf->disks[i];
2400 for (i = conf->pool_size; i < newsize; i++) {
2401 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2402 if (!ndisks[i].extra_page)
2407 for (i = conf->pool_size; i < newsize; i++)
2408 if (ndisks[i].extra_page)
2409 put_page(ndisks[i].extra_page);
2413 conf->disks = ndisks;
2418 conf->slab_cache = sc;
2419 conf->active_name = 1-conf->active_name;
2421 /* Step 4, return new stripes to service */
2422 while(!list_empty(&newstripes)) {
2423 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2424 list_del_init(&nsh->lru);
2426 for (i=conf->raid_disks; i < newsize; i++)
2427 if (nsh->dev[i].page == NULL) {
2428 struct page *p = alloc_page(GFP_NOIO);
2429 nsh->dev[i].page = p;
2430 nsh->dev[i].orig_page = p;
2434 raid5_release_stripe(nsh);
2436 /* critical section pass, GFP_NOIO no longer needed */
2439 conf->pool_size = newsize;
2440 mutex_unlock(&conf->cache_size_mutex);
2445 static int drop_one_stripe(struct r5conf *conf)
2447 struct stripe_head *sh;
2448 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2450 spin_lock_irq(conf->hash_locks + hash);
2451 sh = get_free_stripe(conf, hash);
2452 spin_unlock_irq(conf->hash_locks + hash);
2455 BUG_ON(atomic_read(&sh->count));
2457 free_stripe(conf->slab_cache, sh);
2458 atomic_dec(&conf->active_stripes);
2459 conf->max_nr_stripes--;
2463 static void shrink_stripes(struct r5conf *conf)
2465 while (conf->max_nr_stripes &&
2466 drop_one_stripe(conf))
2469 kmem_cache_destroy(conf->slab_cache);
2470 conf->slab_cache = NULL;
2473 static void raid5_end_read_request(struct bio * bi)
2475 struct stripe_head *sh = bi->bi_private;
2476 struct r5conf *conf = sh->raid_conf;
2477 int disks = sh->disks, i;
2478 char b[BDEVNAME_SIZE];
2479 struct md_rdev *rdev = NULL;
2482 for (i=0 ; i<disks; i++)
2483 if (bi == &sh->dev[i].req)
2486 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2487 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2494 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2495 /* If replacement finished while this request was outstanding,
2496 * 'replacement' might be NULL already.
2497 * In that case it moved down to 'rdev'.
2498 * rdev is not removed until all requests are finished.
2500 rdev = conf->disks[i].replacement;
2502 rdev = conf->disks[i].rdev;
2504 if (use_new_offset(conf, sh))
2505 s = sh->sector + rdev->new_data_offset;
2507 s = sh->sector + rdev->data_offset;
2508 if (!bi->bi_status) {
2509 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2510 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2511 /* Note that this cannot happen on a
2512 * replacement device. We just fail those on
2515 pr_info_ratelimited(
2516 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2517 mdname(conf->mddev), STRIPE_SECTORS,
2518 (unsigned long long)s,
2519 bdevname(rdev->bdev, b));
2520 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2521 clear_bit(R5_ReadError, &sh->dev[i].flags);
2522 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2523 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2524 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2526 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2528 * end read for a page in journal, this
2529 * must be preparing for prexor in rmw
2531 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2533 if (atomic_read(&rdev->read_errors))
2534 atomic_set(&rdev->read_errors, 0);
2536 const char *bdn = bdevname(rdev->bdev, b);
2540 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2541 if (!(bi->bi_status == BLK_STS_PROTECTION))
2542 atomic_inc(&rdev->read_errors);
2543 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2544 pr_warn_ratelimited(
2545 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2546 mdname(conf->mddev),
2547 (unsigned long long)s,
2549 else if (conf->mddev->degraded >= conf->max_degraded) {
2551 pr_warn_ratelimited(
2552 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2553 mdname(conf->mddev),
2554 (unsigned long long)s,
2556 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2559 pr_warn_ratelimited(
2560 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2561 mdname(conf->mddev),
2562 (unsigned long long)s,
2564 } else if (atomic_read(&rdev->read_errors)
2565 > conf->max_nr_stripes)
2566 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2567 mdname(conf->mddev), bdn);
2570 if (set_bad && test_bit(In_sync, &rdev->flags)
2571 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2574 if (sh->qd_idx >= 0 && sh->pd_idx == i)
2575 set_bit(R5_ReadError, &sh->dev[i].flags);
2576 else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2577 set_bit(R5_ReadError, &sh->dev[i].flags);
2578 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2580 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2582 clear_bit(R5_ReadError, &sh->dev[i].flags);
2583 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2585 && test_bit(In_sync, &rdev->flags)
2586 && rdev_set_badblocks(
2587 rdev, sh->sector, STRIPE_SECTORS, 0)))
2588 md_error(conf->mddev, rdev);
2591 rdev_dec_pending(rdev, conf->mddev);
2593 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2594 set_bit(STRIPE_HANDLE, &sh->state);
2595 raid5_release_stripe(sh);
2598 static void raid5_end_write_request(struct bio *bi)
2600 struct stripe_head *sh = bi->bi_private;
2601 struct r5conf *conf = sh->raid_conf;
2602 int disks = sh->disks, i;
2603 struct md_rdev *uninitialized_var(rdev);
2606 int replacement = 0;
2608 for (i = 0 ; i < disks; i++) {
2609 if (bi == &sh->dev[i].req) {
2610 rdev = conf->disks[i].rdev;
2613 if (bi == &sh->dev[i].rreq) {
2614 rdev = conf->disks[i].replacement;
2618 /* rdev was removed and 'replacement'
2619 * replaced it. rdev is not removed
2620 * until all requests are finished.
2622 rdev = conf->disks[i].rdev;
2626 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2627 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2637 md_error(conf->mddev, rdev);
2638 else if (is_badblock(rdev, sh->sector,
2640 &first_bad, &bad_sectors))
2641 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2643 if (bi->bi_status) {
2644 set_bit(STRIPE_DEGRADED, &sh->state);
2645 set_bit(WriteErrorSeen, &rdev->flags);
2646 set_bit(R5_WriteError, &sh->dev[i].flags);
2647 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2648 set_bit(MD_RECOVERY_NEEDED,
2649 &rdev->mddev->recovery);
2650 } else if (is_badblock(rdev, sh->sector,
2652 &first_bad, &bad_sectors)) {
2653 set_bit(R5_MadeGood, &sh->dev[i].flags);
2654 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2655 /* That was a successful write so make
2656 * sure it looks like we already did
2659 set_bit(R5_ReWrite, &sh->dev[i].flags);
2662 rdev_dec_pending(rdev, conf->mddev);
2664 if (sh->batch_head && bi->bi_status && !replacement)
2665 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2668 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2669 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2670 set_bit(STRIPE_HANDLE, &sh->state);
2671 raid5_release_stripe(sh);
2673 if (sh->batch_head && sh != sh->batch_head)
2674 raid5_release_stripe(sh->batch_head);
2677 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2679 char b[BDEVNAME_SIZE];
2680 struct r5conf *conf = mddev->private;
2681 unsigned long flags;
2682 pr_debug("raid456: error called\n");
2684 spin_lock_irqsave(&conf->device_lock, flags);
2685 set_bit(Faulty, &rdev->flags);
2686 clear_bit(In_sync, &rdev->flags);
2687 mddev->degraded = raid5_calc_degraded(conf);
2688 spin_unlock_irqrestore(&conf->device_lock, flags);
2689 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2691 set_bit(Blocked, &rdev->flags);
2692 set_mask_bits(&mddev->sb_flags, 0,
2693 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2694 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2695 "md/raid:%s: Operation continuing on %d devices.\n",
2697 bdevname(rdev->bdev, b),
2699 conf->raid_disks - mddev->degraded);
2700 r5c_update_on_rdev_error(mddev, rdev);
2704 * Input: a 'big' sector number,
2705 * Output: index of the data and parity disk, and the sector # in them.
2707 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2708 int previous, int *dd_idx,
2709 struct stripe_head *sh)
2711 sector_t stripe, stripe2;
2712 sector_t chunk_number;
2713 unsigned int chunk_offset;
2716 sector_t new_sector;
2717 int algorithm = previous ? conf->prev_algo
2719 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2720 : conf->chunk_sectors;
2721 int raid_disks = previous ? conf->previous_raid_disks
2723 int data_disks = raid_disks - conf->max_degraded;
2725 /* First compute the information on this sector */
2728 * Compute the chunk number and the sector offset inside the chunk
2730 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2731 chunk_number = r_sector;
2734 * Compute the stripe number
2736 stripe = chunk_number;
2737 *dd_idx = sector_div(stripe, data_disks);
2740 * Select the parity disk based on the user selected algorithm.
2742 pd_idx = qd_idx = -1;
2743 switch(conf->level) {
2745 pd_idx = data_disks;
2748 switch (algorithm) {
2749 case ALGORITHM_LEFT_ASYMMETRIC:
2750 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2751 if (*dd_idx >= pd_idx)
2754 case ALGORITHM_RIGHT_ASYMMETRIC:
2755 pd_idx = sector_div(stripe2, raid_disks);
2756 if (*dd_idx >= pd_idx)
2759 case ALGORITHM_LEFT_SYMMETRIC:
2760 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2761 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2763 case ALGORITHM_RIGHT_SYMMETRIC:
2764 pd_idx = sector_div(stripe2, raid_disks);
2765 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2767 case ALGORITHM_PARITY_0:
2771 case ALGORITHM_PARITY_N:
2772 pd_idx = data_disks;
2780 switch (algorithm) {
2781 case ALGORITHM_LEFT_ASYMMETRIC:
2782 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2783 qd_idx = pd_idx + 1;
2784 if (pd_idx == raid_disks-1) {
2785 (*dd_idx)++; /* Q D D D P */
2787 } else if (*dd_idx >= pd_idx)
2788 (*dd_idx) += 2; /* D D P Q D */
2790 case ALGORITHM_RIGHT_ASYMMETRIC:
2791 pd_idx = sector_div(stripe2, raid_disks);
2792 qd_idx = pd_idx + 1;
2793 if (pd_idx == raid_disks-1) {
2794 (*dd_idx)++; /* Q D D D P */
2796 } else if (*dd_idx >= pd_idx)
2797 (*dd_idx) += 2; /* D D P Q D */
2799 case ALGORITHM_LEFT_SYMMETRIC:
2800 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2801 qd_idx = (pd_idx + 1) % raid_disks;
2802 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2804 case ALGORITHM_RIGHT_SYMMETRIC:
2805 pd_idx = sector_div(stripe2, raid_disks);
2806 qd_idx = (pd_idx + 1) % raid_disks;
2807 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2810 case ALGORITHM_PARITY_0:
2815 case ALGORITHM_PARITY_N:
2816 pd_idx = data_disks;
2817 qd_idx = data_disks + 1;
2820 case ALGORITHM_ROTATING_ZERO_RESTART:
2821 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2822 * of blocks for computing Q is different.
2824 pd_idx = sector_div(stripe2, raid_disks);
2825 qd_idx = pd_idx + 1;
2826 if (pd_idx == raid_disks-1) {
2827 (*dd_idx)++; /* Q D D D P */
2829 } else if (*dd_idx >= pd_idx)
2830 (*dd_idx) += 2; /* D D P Q D */
2834 case ALGORITHM_ROTATING_N_RESTART:
2835 /* Same a left_asymmetric, by first stripe is
2836 * D D D P Q rather than
2840 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2841 qd_idx = pd_idx + 1;
2842 if (pd_idx == raid_disks-1) {
2843 (*dd_idx)++; /* Q D D D P */
2845 } else if (*dd_idx >= pd_idx)
2846 (*dd_idx) += 2; /* D D P Q D */
2850 case ALGORITHM_ROTATING_N_CONTINUE:
2851 /* Same as left_symmetric but Q is before P */
2852 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2853 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2854 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2858 case ALGORITHM_LEFT_ASYMMETRIC_6:
2859 /* RAID5 left_asymmetric, with Q on last device */
2860 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2861 if (*dd_idx >= pd_idx)
2863 qd_idx = raid_disks - 1;
2866 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2867 pd_idx = sector_div(stripe2, raid_disks-1);
2868 if (*dd_idx >= pd_idx)
2870 qd_idx = raid_disks - 1;
2873 case ALGORITHM_LEFT_SYMMETRIC_6:
2874 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2875 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2876 qd_idx = raid_disks - 1;
2879 case ALGORITHM_RIGHT_SYMMETRIC_6:
2880 pd_idx = sector_div(stripe2, raid_disks-1);
2881 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2882 qd_idx = raid_disks - 1;
2885 case ALGORITHM_PARITY_0_6:
2888 qd_idx = raid_disks - 1;
2898 sh->pd_idx = pd_idx;
2899 sh->qd_idx = qd_idx;
2900 sh->ddf_layout = ddf_layout;
2903 * Finally, compute the new sector number
2905 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2909 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2911 struct r5conf *conf = sh->raid_conf;
2912 int raid_disks = sh->disks;
2913 int data_disks = raid_disks - conf->max_degraded;
2914 sector_t new_sector = sh->sector, check;
2915 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2916 : conf->chunk_sectors;
2917 int algorithm = previous ? conf->prev_algo
2921 sector_t chunk_number;
2922 int dummy1, dd_idx = i;
2924 struct stripe_head sh2;
2926 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2927 stripe = new_sector;
2929 if (i == sh->pd_idx)
2931 switch(conf->level) {
2934 switch (algorithm) {
2935 case ALGORITHM_LEFT_ASYMMETRIC:
2936 case ALGORITHM_RIGHT_ASYMMETRIC:
2940 case ALGORITHM_LEFT_SYMMETRIC:
2941 case ALGORITHM_RIGHT_SYMMETRIC:
2944 i -= (sh->pd_idx + 1);
2946 case ALGORITHM_PARITY_0:
2949 case ALGORITHM_PARITY_N:
2956 if (i == sh->qd_idx)
2957 return 0; /* It is the Q disk */
2958 switch (algorithm) {
2959 case ALGORITHM_LEFT_ASYMMETRIC:
2960 case ALGORITHM_RIGHT_ASYMMETRIC:
2961 case ALGORITHM_ROTATING_ZERO_RESTART:
2962 case ALGORITHM_ROTATING_N_RESTART:
2963 if (sh->pd_idx == raid_disks-1)
2964 i--; /* Q D D D P */
2965 else if (i > sh->pd_idx)
2966 i -= 2; /* D D P Q D */
2968 case ALGORITHM_LEFT_SYMMETRIC:
2969 case ALGORITHM_RIGHT_SYMMETRIC:
2970 if (sh->pd_idx == raid_disks-1)
2971 i--; /* Q D D D P */
2976 i -= (sh->pd_idx + 2);
2979 case ALGORITHM_PARITY_0:
2982 case ALGORITHM_PARITY_N:
2984 case ALGORITHM_ROTATING_N_CONTINUE:
2985 /* Like left_symmetric, but P is before Q */
2986 if (sh->pd_idx == 0)
2987 i--; /* P D D D Q */
2992 i -= (sh->pd_idx + 1);
2995 case ALGORITHM_LEFT_ASYMMETRIC_6:
2996 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3000 case ALGORITHM_LEFT_SYMMETRIC_6:
3001 case ALGORITHM_RIGHT_SYMMETRIC_6:
3003 i += data_disks + 1;
3004 i -= (sh->pd_idx + 1);
3006 case ALGORITHM_PARITY_0_6:
3015 chunk_number = stripe * data_disks + i;
3016 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3018 check = raid5_compute_sector(conf, r_sector,
3019 previous, &dummy1, &sh2);
3020 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3021 || sh2.qd_idx != sh->qd_idx) {
3022 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3023 mdname(conf->mddev));
3030 * There are cases where we want handle_stripe_dirtying() and
3031 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3033 * This function checks whether we want to delay the towrite. Specifically,
3034 * we delay the towrite when:
3036 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3037 * stripe has data in journal (for other devices).
3039 * In this case, when reading data for the non-overwrite dev, it is
3040 * necessary to handle complex rmw of write back cache (prexor with
3041 * orig_page, and xor with page). To keep read path simple, we would
3042 * like to flush data in journal to RAID disks first, so complex rmw
3043 * is handled in the write patch (handle_stripe_dirtying).
3045 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3047 * It is important to be able to flush all stripes in raid5-cache.
3048 * Therefore, we need reserve some space on the journal device for
3049 * these flushes. If flush operation includes pending writes to the
3050 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3051 * for the flush out. If we exclude these pending writes from flush
3052 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3053 * Therefore, excluding pending writes in these cases enables more
3054 * efficient use of the journal device.
3056 * Note: To make sure the stripe makes progress, we only delay
3057 * towrite for stripes with data already in journal (injournal > 0).
3058 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3059 * no_space_stripes list.
3061 * 3. during journal failure
3062 * In journal failure, we try to flush all cached data to raid disks
3063 * based on data in stripe cache. The array is read-only to upper
3064 * layers, so we would skip all pending writes.
3067 static inline bool delay_towrite(struct r5conf *conf,
3069 struct stripe_head_state *s)
3072 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3073 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3076 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3080 if (s->log_failed && s->injournal)
3086 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3087 int rcw, int expand)
3089 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3090 struct r5conf *conf = sh->raid_conf;
3091 int level = conf->level;
3095 * In some cases, handle_stripe_dirtying initially decided to
3096 * run rmw and allocates extra page for prexor. However, rcw is
3097 * cheaper later on. We need to free the extra page now,
3098 * because we won't be able to do that in ops_complete_prexor().
3100 r5c_release_extra_page(sh);
3102 for (i = disks; i--; ) {
3103 struct r5dev *dev = &sh->dev[i];
3105 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3106 set_bit(R5_LOCKED, &dev->flags);
3107 set_bit(R5_Wantdrain, &dev->flags);
3109 clear_bit(R5_UPTODATE, &dev->flags);
3111 } else if (test_bit(R5_InJournal, &dev->flags)) {
3112 set_bit(R5_LOCKED, &dev->flags);
3116 /* if we are not expanding this is a proper write request, and
3117 * there will be bios with new data to be drained into the
3122 /* False alarm, nothing to do */
3124 sh->reconstruct_state = reconstruct_state_drain_run;
3125 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3127 sh->reconstruct_state = reconstruct_state_run;
3129 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3131 if (s->locked + conf->max_degraded == disks)
3132 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3133 atomic_inc(&conf->pending_full_writes);
3135 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3136 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3137 BUG_ON(level == 6 &&
3138 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3139 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3141 for (i = disks; i--; ) {
3142 struct r5dev *dev = &sh->dev[i];
3143 if (i == pd_idx || i == qd_idx)
3147 (test_bit(R5_UPTODATE, &dev->flags) ||
3148 test_bit(R5_Wantcompute, &dev->flags))) {
3149 set_bit(R5_Wantdrain, &dev->flags);
3150 set_bit(R5_LOCKED, &dev->flags);
3151 clear_bit(R5_UPTODATE, &dev->flags);
3153 } else if (test_bit(R5_InJournal, &dev->flags)) {
3154 set_bit(R5_LOCKED, &dev->flags);
3159 /* False alarm - nothing to do */
3161 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3162 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3163 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3164 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3167 /* keep the parity disk(s) locked while asynchronous operations
3170 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3171 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3175 int qd_idx = sh->qd_idx;
3176 struct r5dev *dev = &sh->dev[qd_idx];
3178 set_bit(R5_LOCKED, &dev->flags);
3179 clear_bit(R5_UPTODATE, &dev->flags);
3183 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3184 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3185 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3186 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3187 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3189 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3190 __func__, (unsigned long long)sh->sector,
3191 s->locked, s->ops_request);
3195 * Each stripe/dev can have one or more bion attached.
3196 * toread/towrite point to the first in a chain.
3197 * The bi_next chain must be in order.
3199 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3200 int forwrite, int previous)
3203 struct r5conf *conf = sh->raid_conf;
3206 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3207 (unsigned long long)bi->bi_iter.bi_sector,
3208 (unsigned long long)sh->sector);
3210 spin_lock_irq(&sh->stripe_lock);
3211 /* Don't allow new IO added to stripes in batch list */
3215 bip = &sh->dev[dd_idx].towrite;
3219 bip = &sh->dev[dd_idx].toread;
3220 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3221 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3223 bip = & (*bip)->bi_next;
3225 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3228 if (forwrite && raid5_has_ppl(conf)) {
3230 * With PPL only writes to consecutive data chunks within a
3231 * stripe are allowed because for a single stripe_head we can
3232 * only have one PPL entry at a time, which describes one data
3233 * range. Not really an overlap, but wait_for_overlap can be
3234 * used to handle this.
3242 for (i = 0; i < sh->disks; i++) {
3243 if (i != sh->pd_idx &&
3244 (i == dd_idx || sh->dev[i].towrite)) {
3245 sector = sh->dev[i].sector;
3246 if (count == 0 || sector < first)
3254 if (first + conf->chunk_sectors * (count - 1) != last)
3258 if (!forwrite || previous)
3259 clear_bit(STRIPE_BATCH_READY, &sh->state);
3261 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3265 bio_inc_remaining(bi);
3266 md_write_inc(conf->mddev, bi);
3269 /* check if page is covered */
3270 sector_t sector = sh->dev[dd_idx].sector;
3271 for (bi=sh->dev[dd_idx].towrite;
3272 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3273 bi && bi->bi_iter.bi_sector <= sector;
3274 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3275 if (bio_end_sector(bi) >= sector)
3276 sector = bio_end_sector(bi);
3278 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3279 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3280 sh->overwrite_disks++;
3283 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3284 (unsigned long long)(*bip)->bi_iter.bi_sector,
3285 (unsigned long long)sh->sector, dd_idx);
3287 if (conf->mddev->bitmap && firstwrite) {
3288 /* Cannot hold spinlock over bitmap_startwrite,
3289 * but must ensure this isn't added to a batch until
3290 * we have added to the bitmap and set bm_seq.
3291 * So set STRIPE_BITMAP_PENDING to prevent
3293 * If multiple add_stripe_bio() calls race here they
3294 * much all set STRIPE_BITMAP_PENDING. So only the first one
3295 * to complete "bitmap_startwrite" gets to set
3296 * STRIPE_BIT_DELAY. This is important as once a stripe
3297 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3300 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3301 spin_unlock_irq(&sh->stripe_lock);
3302 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3304 spin_lock_irq(&sh->stripe_lock);
3305 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3306 if (!sh->batch_head) {
3307 sh->bm_seq = conf->seq_flush+1;
3308 set_bit(STRIPE_BIT_DELAY, &sh->state);
3311 spin_unlock_irq(&sh->stripe_lock);
3313 if (stripe_can_batch(sh))
3314 stripe_add_to_batch_list(conf, sh);
3318 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3319 spin_unlock_irq(&sh->stripe_lock);
3323 static void end_reshape(struct r5conf *conf);
3325 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3326 struct stripe_head *sh)
3328 int sectors_per_chunk =
3329 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3331 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3332 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3334 raid5_compute_sector(conf,
3335 stripe * (disks - conf->max_degraded)
3336 *sectors_per_chunk + chunk_offset,
3342 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3343 struct stripe_head_state *s, int disks)
3346 BUG_ON(sh->batch_head);
3347 for (i = disks; i--; ) {
3351 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3352 struct md_rdev *rdev;
3354 rdev = rcu_dereference(conf->disks[i].rdev);
3355 if (rdev && test_bit(In_sync, &rdev->flags) &&
3356 !test_bit(Faulty, &rdev->flags))
3357 atomic_inc(&rdev->nr_pending);
3362 if (!rdev_set_badblocks(
3366 md_error(conf->mddev, rdev);
3367 rdev_dec_pending(rdev, conf->mddev);
3370 spin_lock_irq(&sh->stripe_lock);
3371 /* fail all writes first */
3372 bi = sh->dev[i].towrite;
3373 sh->dev[i].towrite = NULL;
3374 sh->overwrite_disks = 0;
3375 spin_unlock_irq(&sh->stripe_lock);
3379 log_stripe_write_finished(sh);
3381 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3382 wake_up(&conf->wait_for_overlap);
3384 while (bi && bi->bi_iter.bi_sector <
3385 sh->dev[i].sector + STRIPE_SECTORS) {
3386 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3388 md_write_end(conf->mddev);
3393 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3394 STRIPE_SECTORS, 0, 0);
3396 /* and fail all 'written' */
3397 bi = sh->dev[i].written;
3398 sh->dev[i].written = NULL;
3399 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3400 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3401 sh->dev[i].page = sh->dev[i].orig_page;
3404 if (bi) bitmap_end = 1;
3405 while (bi && bi->bi_iter.bi_sector <
3406 sh->dev[i].sector + STRIPE_SECTORS) {
3407 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3409 md_write_end(conf->mddev);
3414 /* fail any reads if this device is non-operational and
3415 * the data has not reached the cache yet.
3417 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3418 s->failed > conf->max_degraded &&
3419 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3420 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3421 spin_lock_irq(&sh->stripe_lock);
3422 bi = sh->dev[i].toread;
3423 sh->dev[i].toread = NULL;
3424 spin_unlock_irq(&sh->stripe_lock);
3425 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3426 wake_up(&conf->wait_for_overlap);
3429 while (bi && bi->bi_iter.bi_sector <
3430 sh->dev[i].sector + STRIPE_SECTORS) {
3431 struct bio *nextbi =
3432 r5_next_bio(bi, sh->dev[i].sector);
3439 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3440 STRIPE_SECTORS, 0, 0);
3441 /* If we were in the middle of a write the parity block might
3442 * still be locked - so just clear all R5_LOCKED flags
3444 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3449 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3450 if (atomic_dec_and_test(&conf->pending_full_writes))
3451 md_wakeup_thread(conf->mddev->thread);
3455 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3456 struct stripe_head_state *s)
3461 BUG_ON(sh->batch_head);
3462 clear_bit(STRIPE_SYNCING, &sh->state);
3463 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3464 wake_up(&conf->wait_for_overlap);
3467 /* There is nothing more to do for sync/check/repair.
3468 * Don't even need to abort as that is handled elsewhere
3469 * if needed, and not always wanted e.g. if there is a known
3471 * For recover/replace we need to record a bad block on all
3472 * non-sync devices, or abort the recovery
3474 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3475 /* During recovery devices cannot be removed, so
3476 * locking and refcounting of rdevs is not needed
3479 for (i = 0; i < conf->raid_disks; i++) {
3480 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3482 && !test_bit(Faulty, &rdev->flags)
3483 && !test_bit(In_sync, &rdev->flags)
3484 && !rdev_set_badblocks(rdev, sh->sector,
3487 rdev = rcu_dereference(conf->disks[i].replacement);
3489 && !test_bit(Faulty, &rdev->flags)
3490 && !test_bit(In_sync, &rdev->flags)
3491 && !rdev_set_badblocks(rdev, sh->sector,
3497 conf->recovery_disabled =
3498 conf->mddev->recovery_disabled;
3500 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3503 static int want_replace(struct stripe_head *sh, int disk_idx)
3505 struct md_rdev *rdev;
3509 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3511 && !test_bit(Faulty, &rdev->flags)
3512 && !test_bit(In_sync, &rdev->flags)
3513 && (rdev->recovery_offset <= sh->sector
3514 || rdev->mddev->recovery_cp <= sh->sector))
3520 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3521 int disk_idx, int disks)
3523 struct r5dev *dev = &sh->dev[disk_idx];
3524 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3525 &sh->dev[s->failed_num[1]] };
3529 if (test_bit(R5_LOCKED, &dev->flags) ||
3530 test_bit(R5_UPTODATE, &dev->flags))
3531 /* No point reading this as we already have it or have
3532 * decided to get it.
3537 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3538 /* We need this block to directly satisfy a request */
3541 if (s->syncing || s->expanding ||
3542 (s->replacing && want_replace(sh, disk_idx)))
3543 /* When syncing, or expanding we read everything.
3544 * When replacing, we need the replaced block.
3548 if ((s->failed >= 1 && fdev[0]->toread) ||
3549 (s->failed >= 2 && fdev[1]->toread))
3550 /* If we want to read from a failed device, then
3551 * we need to actually read every other device.
3555 /* Sometimes neither read-modify-write nor reconstruct-write
3556 * cycles can work. In those cases we read every block we
3557 * can. Then the parity-update is certain to have enough to
3559 * This can only be a problem when we need to write something,
3560 * and some device has failed. If either of those tests
3561 * fail we need look no further.
3563 if (!s->failed || !s->to_write)
3566 if (test_bit(R5_Insync, &dev->flags) &&
3567 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3568 /* Pre-reads at not permitted until after short delay
3569 * to gather multiple requests. However if this
3570 * device is no Insync, the block could only be computed
3571 * and there is no need to delay that.
3575 for (i = 0; i < s->failed && i < 2; i++) {
3576 if (fdev[i]->towrite &&
3577 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3578 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3579 /* If we have a partial write to a failed
3580 * device, then we will need to reconstruct
3581 * the content of that device, so all other
3582 * devices must be read.
3587 /* If we are forced to do a reconstruct-write, either because
3588 * the current RAID6 implementation only supports that, or
3589 * because parity cannot be trusted and we are currently
3590 * recovering it, there is extra need to be careful.
3591 * If one of the devices that we would need to read, because
3592 * it is not being overwritten (and maybe not written at all)
3593 * is missing/faulty, then we need to read everything we can.
3595 if (sh->raid_conf->level != 6 &&
3596 sh->raid_conf->rmw_level != PARITY_DISABLE_RMW &&
3597 sh->sector < sh->raid_conf->mddev->recovery_cp)
3598 /* reconstruct-write isn't being forced */
3600 for (i = 0; i < s->failed && i < 2; i++) {
3601 if (s->failed_num[i] != sh->pd_idx &&
3602 s->failed_num[i] != sh->qd_idx &&
3603 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3604 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3611 /* fetch_block - checks the given member device to see if its data needs
3612 * to be read or computed to satisfy a request.
3614 * Returns 1 when no more member devices need to be checked, otherwise returns
3615 * 0 to tell the loop in handle_stripe_fill to continue
3617 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3618 int disk_idx, int disks)
3620 struct r5dev *dev = &sh->dev[disk_idx];
3622 /* is the data in this block needed, and can we get it? */
3623 if (need_this_block(sh, s, disk_idx, disks)) {
3624 /* we would like to get this block, possibly by computing it,
3625 * otherwise read it if the backing disk is insync
3627 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3628 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3629 BUG_ON(sh->batch_head);
3632 * In the raid6 case if the only non-uptodate disk is P
3633 * then we already trusted P to compute the other failed
3634 * drives. It is safe to compute rather than re-read P.
3635 * In other cases we only compute blocks from failed
3636 * devices, otherwise check/repair might fail to detect
3637 * a real inconsistency.
3640 if ((s->uptodate == disks - 1) &&
3641 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3642 (s->failed && (disk_idx == s->failed_num[0] ||
3643 disk_idx == s->failed_num[1])))) {
3644 /* have disk failed, and we're requested to fetch it;
3647 pr_debug("Computing stripe %llu block %d\n",
3648 (unsigned long long)sh->sector, disk_idx);
3649 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3650 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3651 set_bit(R5_Wantcompute, &dev->flags);
3652 sh->ops.target = disk_idx;
3653 sh->ops.target2 = -1; /* no 2nd target */
3655 /* Careful: from this point on 'uptodate' is in the eye
3656 * of raid_run_ops which services 'compute' operations
3657 * before writes. R5_Wantcompute flags a block that will
3658 * be R5_UPTODATE by the time it is needed for a
3659 * subsequent operation.
3663 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3664 /* Computing 2-failure is *very* expensive; only
3665 * do it if failed >= 2
3668 for (other = disks; other--; ) {
3669 if (other == disk_idx)
3671 if (!test_bit(R5_UPTODATE,
3672 &sh->dev[other].flags))
3676 pr_debug("Computing stripe %llu blocks %d,%d\n",
3677 (unsigned long long)sh->sector,
3679 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3680 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3681 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3682 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3683 sh->ops.target = disk_idx;
3684 sh->ops.target2 = other;
3688 } else if (test_bit(R5_Insync, &dev->flags)) {
3689 set_bit(R5_LOCKED, &dev->flags);
3690 set_bit(R5_Wantread, &dev->flags);
3692 pr_debug("Reading block %d (sync=%d)\n",
3693 disk_idx, s->syncing);
3701 * handle_stripe_fill - read or compute data to satisfy pending requests.
3703 static void handle_stripe_fill(struct stripe_head *sh,
3704 struct stripe_head_state *s,
3709 /* look for blocks to read/compute, skip this if a compute
3710 * is already in flight, or if the stripe contents are in the
3711 * midst of changing due to a write
3713 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3714 !sh->reconstruct_state) {
3717 * For degraded stripe with data in journal, do not handle
3718 * read requests yet, instead, flush the stripe to raid
3719 * disks first, this avoids handling complex rmw of write
3720 * back cache (prexor with orig_page, and then xor with
3721 * page) in the read path
3723 if (s->injournal && s->failed) {
3724 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3725 r5c_make_stripe_write_out(sh);
3729 for (i = disks; i--; )
3730 if (fetch_block(sh, s, i, disks))
3734 set_bit(STRIPE_HANDLE, &sh->state);
3737 static void break_stripe_batch_list(struct stripe_head *head_sh,
3738 unsigned long handle_flags);
3739 /* handle_stripe_clean_event
3740 * any written block on an uptodate or failed drive can be returned.
3741 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3742 * never LOCKED, so we don't need to test 'failed' directly.
3744 static void handle_stripe_clean_event(struct r5conf *conf,
3745 struct stripe_head *sh, int disks)
3749 int discard_pending = 0;
3750 struct stripe_head *head_sh = sh;
3751 bool do_endio = false;
3753 for (i = disks; i--; )
3754 if (sh->dev[i].written) {
3756 if (!test_bit(R5_LOCKED, &dev->flags) &&
3757 (test_bit(R5_UPTODATE, &dev->flags) ||
3758 test_bit(R5_Discard, &dev->flags) ||
3759 test_bit(R5_SkipCopy, &dev->flags))) {
3760 /* We can return any write requests */
3761 struct bio *wbi, *wbi2;
3762 pr_debug("Return write for disc %d\n", i);
3763 if (test_and_clear_bit(R5_Discard, &dev->flags))
3764 clear_bit(R5_UPTODATE, &dev->flags);
3765 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3766 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3771 dev->page = dev->orig_page;
3773 dev->written = NULL;
3774 while (wbi && wbi->bi_iter.bi_sector <
3775 dev->sector + STRIPE_SECTORS) {
3776 wbi2 = r5_next_bio(wbi, dev->sector);
3777 md_write_end(conf->mddev);
3781 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3783 !test_bit(STRIPE_DEGRADED, &sh->state),
3785 if (head_sh->batch_head) {
3786 sh = list_first_entry(&sh->batch_list,
3789 if (sh != head_sh) {
3796 } else if (test_bit(R5_Discard, &dev->flags))
3797 discard_pending = 1;
3800 log_stripe_write_finished(sh);
3802 if (!discard_pending &&
3803 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3805 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3806 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3807 if (sh->qd_idx >= 0) {
3808 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3809 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3811 /* now that discard is done we can proceed with any sync */
3812 clear_bit(STRIPE_DISCARD, &sh->state);
3814 * SCSI discard will change some bio fields and the stripe has
3815 * no updated data, so remove it from hash list and the stripe
3816 * will be reinitialized
3819 hash = sh->hash_lock_index;
3820 spin_lock_irq(conf->hash_locks + hash);
3822 spin_unlock_irq(conf->hash_locks + hash);
3823 if (head_sh->batch_head) {
3824 sh = list_first_entry(&sh->batch_list,
3825 struct stripe_head, batch_list);
3831 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3832 set_bit(STRIPE_HANDLE, &sh->state);
3836 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3837 if (atomic_dec_and_test(&conf->pending_full_writes))
3838 md_wakeup_thread(conf->mddev->thread);
3840 if (head_sh->batch_head && do_endio)
3841 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3845 * For RMW in write back cache, we need extra page in prexor to store the
3846 * old data. This page is stored in dev->orig_page.
3848 * This function checks whether we have data for prexor. The exact logic
3850 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3852 static inline bool uptodate_for_rmw(struct r5dev *dev)
3854 return (test_bit(R5_UPTODATE, &dev->flags)) &&
3855 (!test_bit(R5_InJournal, &dev->flags) ||
3856 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3859 static int handle_stripe_dirtying(struct r5conf *conf,
3860 struct stripe_head *sh,
3861 struct stripe_head_state *s,
3864 int rmw = 0, rcw = 0, i;
3865 sector_t recovery_cp = conf->mddev->recovery_cp;
3867 /* Check whether resync is now happening or should start.
3868 * If yes, then the array is dirty (after unclean shutdown or
3869 * initial creation), so parity in some stripes might be inconsistent.
3870 * In this case, we need to always do reconstruct-write, to ensure
3871 * that in case of drive failure or read-error correction, we
3872 * generate correct data from the parity.
3874 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3875 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3877 /* Calculate the real rcw later - for now make it
3878 * look like rcw is cheaper
3881 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3882 conf->rmw_level, (unsigned long long)recovery_cp,
3883 (unsigned long long)sh->sector);
3884 } else for (i = disks; i--; ) {
3885 /* would I have to read this buffer for read_modify_write */
3886 struct r5dev *dev = &sh->dev[i];
3887 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3888 i == sh->pd_idx || i == sh->qd_idx ||
3889 test_bit(R5_InJournal, &dev->flags)) &&
3890 !test_bit(R5_LOCKED, &dev->flags) &&
3891 !(uptodate_for_rmw(dev) ||
3892 test_bit(R5_Wantcompute, &dev->flags))) {
3893 if (test_bit(R5_Insync, &dev->flags))
3896 rmw += 2*disks; /* cannot read it */
3898 /* Would I have to read this buffer for reconstruct_write */
3899 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3900 i != sh->pd_idx && i != sh->qd_idx &&
3901 !test_bit(R5_LOCKED, &dev->flags) &&
3902 !(test_bit(R5_UPTODATE, &dev->flags) ||
3903 test_bit(R5_Wantcompute, &dev->flags))) {
3904 if (test_bit(R5_Insync, &dev->flags))
3911 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3912 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3913 set_bit(STRIPE_HANDLE, &sh->state);
3914 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3915 /* prefer read-modify-write, but need to get some data */
3916 if (conf->mddev->queue)
3917 blk_add_trace_msg(conf->mddev->queue,
3918 "raid5 rmw %llu %d",
3919 (unsigned long long)sh->sector, rmw);
3920 for (i = disks; i--; ) {
3921 struct r5dev *dev = &sh->dev[i];
3922 if (test_bit(R5_InJournal, &dev->flags) &&
3923 dev->page == dev->orig_page &&
3924 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3925 /* alloc page for prexor */
3926 struct page *p = alloc_page(GFP_NOIO);
3934 * alloc_page() failed, try use
3935 * disk_info->extra_page
3937 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3938 &conf->cache_state)) {
3939 r5c_use_extra_page(sh);
3943 /* extra_page in use, add to delayed_list */
3944 set_bit(STRIPE_DELAYED, &sh->state);
3945 s->waiting_extra_page = 1;
3950 for (i = disks; i--; ) {
3951 struct r5dev *dev = &sh->dev[i];
3952 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3953 i == sh->pd_idx || i == sh->qd_idx ||
3954 test_bit(R5_InJournal, &dev->flags)) &&
3955 !test_bit(R5_LOCKED, &dev->flags) &&
3956 !(uptodate_for_rmw(dev) ||
3957 test_bit(R5_Wantcompute, &dev->flags)) &&
3958 test_bit(R5_Insync, &dev->flags)) {
3959 if (test_bit(STRIPE_PREREAD_ACTIVE,
3961 pr_debug("Read_old block %d for r-m-w\n",
3963 set_bit(R5_LOCKED, &dev->flags);
3964 set_bit(R5_Wantread, &dev->flags);
3967 set_bit(STRIPE_DELAYED, &sh->state);
3968 set_bit(STRIPE_HANDLE, &sh->state);
3973 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3974 /* want reconstruct write, but need to get some data */
3977 for (i = disks; i--; ) {
3978 struct r5dev *dev = &sh->dev[i];
3979 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3980 i != sh->pd_idx && i != sh->qd_idx &&
3981 !test_bit(R5_LOCKED, &dev->flags) &&
3982 !(test_bit(R5_UPTODATE, &dev->flags) ||
3983 test_bit(R5_Wantcompute, &dev->flags))) {
3985 if (test_bit(R5_Insync, &dev->flags) &&
3986 test_bit(STRIPE_PREREAD_ACTIVE,
3988 pr_debug("Read_old block "
3989 "%d for Reconstruct\n", i);
3990 set_bit(R5_LOCKED, &dev->flags);
3991 set_bit(R5_Wantread, &dev->flags);
3995 set_bit(STRIPE_DELAYED, &sh->state);
3996 set_bit(STRIPE_HANDLE, &sh->state);
4000 if (rcw && conf->mddev->queue)
4001 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4002 (unsigned long long)sh->sector,
4003 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4006 if (rcw > disks && rmw > disks &&
4007 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4008 set_bit(STRIPE_DELAYED, &sh->state);
4010 /* now if nothing is locked, and if we have enough data,
4011 * we can start a write request
4013 /* since handle_stripe can be called at any time we need to handle the
4014 * case where a compute block operation has been submitted and then a
4015 * subsequent call wants to start a write request. raid_run_ops only
4016 * handles the case where compute block and reconstruct are requested
4017 * simultaneously. If this is not the case then new writes need to be
4018 * held off until the compute completes.
4020 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4021 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4022 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4023 schedule_reconstruction(sh, s, rcw == 0, 0);
4027 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4028 struct stripe_head_state *s, int disks)
4030 struct r5dev *dev = NULL;
4032 BUG_ON(sh->batch_head);
4033 set_bit(STRIPE_HANDLE, &sh->state);
4035 switch (sh->check_state) {
4036 case check_state_idle:
4037 /* start a new check operation if there are no failures */
4038 if (s->failed == 0) {
4039 BUG_ON(s->uptodate != disks);
4040 sh->check_state = check_state_run;
4041 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4042 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4046 dev = &sh->dev[s->failed_num[0]];
4048 case check_state_compute_result:
4049 sh->check_state = check_state_idle;
4051 dev = &sh->dev[sh->pd_idx];
4053 /* check that a write has not made the stripe insync */
4054 if (test_bit(STRIPE_INSYNC, &sh->state))
4057 /* either failed parity check, or recovery is happening */
4058 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4059 BUG_ON(s->uptodate != disks);
4061 set_bit(R5_LOCKED, &dev->flags);
4063 set_bit(R5_Wantwrite, &dev->flags);
4065 clear_bit(STRIPE_DEGRADED, &sh->state);
4066 set_bit(STRIPE_INSYNC, &sh->state);
4068 case check_state_run:
4069 break; /* we will be called again upon completion */
4070 case check_state_check_result:
4071 sh->check_state = check_state_idle;
4073 /* if a failure occurred during the check operation, leave
4074 * STRIPE_INSYNC not set and let the stripe be handled again
4079 /* handle a successful check operation, if parity is correct
4080 * we are done. Otherwise update the mismatch count and repair
4081 * parity if !MD_RECOVERY_CHECK
4083 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4084 /* parity is correct (on disc,
4085 * not in buffer any more)
4087 set_bit(STRIPE_INSYNC, &sh->state);
4089 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4090 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4091 /* don't try to repair!! */
4092 set_bit(STRIPE_INSYNC, &sh->state);
4093 pr_warn_ratelimited("%s: mismatch sector in range "
4094 "%llu-%llu\n", mdname(conf->mddev),
4095 (unsigned long long) sh->sector,
4096 (unsigned long long) sh->sector +
4099 sh->check_state = check_state_compute_run;
4100 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4101 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4102 set_bit(R5_Wantcompute,
4103 &sh->dev[sh->pd_idx].flags);
4104 sh->ops.target = sh->pd_idx;
4105 sh->ops.target2 = -1;
4110 case check_state_compute_run:
4113 pr_err("%s: unknown check_state: %d sector: %llu\n",
4114 __func__, sh->check_state,
4115 (unsigned long long) sh->sector);
4120 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4121 struct stripe_head_state *s,
4124 int pd_idx = sh->pd_idx;
4125 int qd_idx = sh->qd_idx;
4128 BUG_ON(sh->batch_head);
4129 set_bit(STRIPE_HANDLE, &sh->state);
4131 BUG_ON(s->failed > 2);
4133 /* Want to check and possibly repair P and Q.
4134 * However there could be one 'failed' device, in which
4135 * case we can only check one of them, possibly using the
4136 * other to generate missing data
4139 switch (sh->check_state) {
4140 case check_state_idle:
4141 /* start a new check operation if there are < 2 failures */
4142 if (s->failed == s->q_failed) {
4143 /* The only possible failed device holds Q, so it
4144 * makes sense to check P (If anything else were failed,
4145 * we would have used P to recreate it).
4147 sh->check_state = check_state_run;
4149 if (!s->q_failed && s->failed < 2) {
4150 /* Q is not failed, and we didn't use it to generate
4151 * anything, so it makes sense to check it
4153 if (sh->check_state == check_state_run)
4154 sh->check_state = check_state_run_pq;
4156 sh->check_state = check_state_run_q;
4159 /* discard potentially stale zero_sum_result */
4160 sh->ops.zero_sum_result = 0;
4162 if (sh->check_state == check_state_run) {
4163 /* async_xor_zero_sum destroys the contents of P */
4164 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4167 if (sh->check_state >= check_state_run &&
4168 sh->check_state <= check_state_run_pq) {
4169 /* async_syndrome_zero_sum preserves P and Q, so
4170 * no need to mark them !uptodate here
4172 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4176 /* we have 2-disk failure */
4177 BUG_ON(s->failed != 2);
4179 case check_state_compute_result:
4180 sh->check_state = check_state_idle;
4182 /* check that a write has not made the stripe insync */
4183 if (test_bit(STRIPE_INSYNC, &sh->state))
4186 /* now write out any block on a failed drive,
4187 * or P or Q if they were recomputed
4190 if (s->failed == 2) {
4191 dev = &sh->dev[s->failed_num[1]];
4193 set_bit(R5_LOCKED, &dev->flags);
4194 set_bit(R5_Wantwrite, &dev->flags);
4196 if (s->failed >= 1) {
4197 dev = &sh->dev[s->failed_num[0]];
4199 set_bit(R5_LOCKED, &dev->flags);
4200 set_bit(R5_Wantwrite, &dev->flags);
4202 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4203 dev = &sh->dev[pd_idx];
4205 set_bit(R5_LOCKED, &dev->flags);
4206 set_bit(R5_Wantwrite, &dev->flags);
4208 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4209 dev = &sh->dev[qd_idx];
4211 set_bit(R5_LOCKED, &dev->flags);
4212 set_bit(R5_Wantwrite, &dev->flags);
4214 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4215 "%s: disk%td not up to date\n",
4216 mdname(conf->mddev),
4217 dev - (struct r5dev *) &sh->dev)) {
4218 clear_bit(R5_LOCKED, &dev->flags);
4219 clear_bit(R5_Wantwrite, &dev->flags);
4222 clear_bit(STRIPE_DEGRADED, &sh->state);
4224 set_bit(STRIPE_INSYNC, &sh->state);
4226 case check_state_run:
4227 case check_state_run_q:
4228 case check_state_run_pq:
4229 break; /* we will be called again upon completion */
4230 case check_state_check_result:
4231 sh->check_state = check_state_idle;
4233 /* handle a successful check operation, if parity is correct
4234 * we are done. Otherwise update the mismatch count and repair
4235 * parity if !MD_RECOVERY_CHECK
4237 if (sh->ops.zero_sum_result == 0) {
4238 /* both parities are correct */
4240 set_bit(STRIPE_INSYNC, &sh->state);
4242 /* in contrast to the raid5 case we can validate
4243 * parity, but still have a failure to write
4246 sh->check_state = check_state_compute_result;
4247 /* Returning at this point means that we may go
4248 * off and bring p and/or q uptodate again so
4249 * we make sure to check zero_sum_result again
4250 * to verify if p or q need writeback
4254 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4255 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4256 /* don't try to repair!! */
4257 set_bit(STRIPE_INSYNC, &sh->state);
4258 pr_warn_ratelimited("%s: mismatch sector in range "
4259 "%llu-%llu\n", mdname(conf->mddev),
4260 (unsigned long long) sh->sector,
4261 (unsigned long long) sh->sector +
4264 int *target = &sh->ops.target;
4266 sh->ops.target = -1;
4267 sh->ops.target2 = -1;
4268 sh->check_state = check_state_compute_run;
4269 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4270 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4271 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4272 set_bit(R5_Wantcompute,
4273 &sh->dev[pd_idx].flags);
4275 target = &sh->ops.target2;
4278 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4279 set_bit(R5_Wantcompute,
4280 &sh->dev[qd_idx].flags);
4287 case check_state_compute_run:
4290 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4291 __func__, sh->check_state,
4292 (unsigned long long) sh->sector);
4297 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4301 /* We have read all the blocks in this stripe and now we need to
4302 * copy some of them into a target stripe for expand.
4304 struct dma_async_tx_descriptor *tx = NULL;
4305 BUG_ON(sh->batch_head);
4306 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4307 for (i = 0; i < sh->disks; i++)
4308 if (i != sh->pd_idx && i != sh->qd_idx) {
4310 struct stripe_head *sh2;
4311 struct async_submit_ctl submit;
4313 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4314 sector_t s = raid5_compute_sector(conf, bn, 0,
4316 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4318 /* so far only the early blocks of this stripe
4319 * have been requested. When later blocks
4320 * get requested, we will try again
4323 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4324 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4325 /* must have already done this block */
4326 raid5_release_stripe(sh2);
4330 /* place all the copies on one channel */
4331 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4332 tx = async_memcpy(sh2->dev[dd_idx].page,
4333 sh->dev[i].page, 0, 0, STRIPE_SIZE,
4336 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4337 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4338 for (j = 0; j < conf->raid_disks; j++)
4339 if (j != sh2->pd_idx &&
4341 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4343 if (j == conf->raid_disks) {
4344 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4345 set_bit(STRIPE_HANDLE, &sh2->state);
4347 raid5_release_stripe(sh2);
4350 /* done submitting copies, wait for them to complete */
4351 async_tx_quiesce(&tx);
4355 * handle_stripe - do things to a stripe.
4357 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4358 * state of various bits to see what needs to be done.
4360 * return some read requests which now have data
4361 * return some write requests which are safely on storage
4362 * schedule a read on some buffers
4363 * schedule a write of some buffers
4364 * return confirmation of parity correctness
4368 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4370 struct r5conf *conf = sh->raid_conf;
4371 int disks = sh->disks;
4374 int do_recovery = 0;
4376 memset(s, 0, sizeof(*s));
4378 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4379 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4380 s->failed_num[0] = -1;
4381 s->failed_num[1] = -1;
4382 s->log_failed = r5l_log_disk_error(conf);
4384 /* Now to look around and see what can be done */
4386 for (i=disks; i--; ) {
4387 struct md_rdev *rdev;
4394 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4396 dev->toread, dev->towrite, dev->written);
4397 /* maybe we can reply to a read
4399 * new wantfill requests are only permitted while
4400 * ops_complete_biofill is guaranteed to be inactive
4402 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4403 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4404 set_bit(R5_Wantfill, &dev->flags);
4406 /* now count some things */
4407 if (test_bit(R5_LOCKED, &dev->flags))
4409 if (test_bit(R5_UPTODATE, &dev->flags))
4411 if (test_bit(R5_Wantcompute, &dev->flags)) {
4413 BUG_ON(s->compute > 2);
4416 if (test_bit(R5_Wantfill, &dev->flags))
4418 else if (dev->toread)
4422 if (!test_bit(R5_OVERWRITE, &dev->flags))
4427 /* Prefer to use the replacement for reads, but only
4428 * if it is recovered enough and has no bad blocks.
4430 rdev = rcu_dereference(conf->disks[i].replacement);
4431 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4432 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4433 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4434 &first_bad, &bad_sectors))
4435 set_bit(R5_ReadRepl, &dev->flags);
4437 if (rdev && !test_bit(Faulty, &rdev->flags))
4438 set_bit(R5_NeedReplace, &dev->flags);
4440 clear_bit(R5_NeedReplace, &dev->flags);
4441 rdev = rcu_dereference(conf->disks[i].rdev);
4442 clear_bit(R5_ReadRepl, &dev->flags);
4444 if (rdev && test_bit(Faulty, &rdev->flags))
4447 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4448 &first_bad, &bad_sectors);
4449 if (s->blocked_rdev == NULL
4450 && (test_bit(Blocked, &rdev->flags)
4453 set_bit(BlockedBadBlocks,
4455 s->blocked_rdev = rdev;
4456 atomic_inc(&rdev->nr_pending);
4459 clear_bit(R5_Insync, &dev->flags);
4463 /* also not in-sync */
4464 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4465 test_bit(R5_UPTODATE, &dev->flags)) {
4466 /* treat as in-sync, but with a read error
4467 * which we can now try to correct
4469 set_bit(R5_Insync, &dev->flags);
4470 set_bit(R5_ReadError, &dev->flags);
4472 } else if (test_bit(In_sync, &rdev->flags))
4473 set_bit(R5_Insync, &dev->flags);
4474 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4475 /* in sync if before recovery_offset */
4476 set_bit(R5_Insync, &dev->flags);
4477 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4478 test_bit(R5_Expanded, &dev->flags))
4479 /* If we've reshaped into here, we assume it is Insync.
4480 * We will shortly update recovery_offset to make
4483 set_bit(R5_Insync, &dev->flags);
4485 if (test_bit(R5_WriteError, &dev->flags)) {
4486 /* This flag does not apply to '.replacement'
4487 * only to .rdev, so make sure to check that*/
4488 struct md_rdev *rdev2 = rcu_dereference(
4489 conf->disks[i].rdev);
4491 clear_bit(R5_Insync, &dev->flags);
4492 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4493 s->handle_bad_blocks = 1;
4494 atomic_inc(&rdev2->nr_pending);
4496 clear_bit(R5_WriteError, &dev->flags);
4498 if (test_bit(R5_MadeGood, &dev->flags)) {
4499 /* This flag does not apply to '.replacement'
4500 * only to .rdev, so make sure to check that*/
4501 struct md_rdev *rdev2 = rcu_dereference(
4502 conf->disks[i].rdev);
4503 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4504 s->handle_bad_blocks = 1;
4505 atomic_inc(&rdev2->nr_pending);
4507 clear_bit(R5_MadeGood, &dev->flags);
4509 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4510 struct md_rdev *rdev2 = rcu_dereference(
4511 conf->disks[i].replacement);
4512 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4513 s->handle_bad_blocks = 1;
4514 atomic_inc(&rdev2->nr_pending);
4516 clear_bit(R5_MadeGoodRepl, &dev->flags);
4518 if (!test_bit(R5_Insync, &dev->flags)) {
4519 /* The ReadError flag will just be confusing now */
4520 clear_bit(R5_ReadError, &dev->flags);
4521 clear_bit(R5_ReWrite, &dev->flags);
4523 if (test_bit(R5_ReadError, &dev->flags))
4524 clear_bit(R5_Insync, &dev->flags);
4525 if (!test_bit(R5_Insync, &dev->flags)) {
4527 s->failed_num[s->failed] = i;
4529 if (rdev && !test_bit(Faulty, &rdev->flags))
4532 rdev = rcu_dereference(
4533 conf->disks[i].replacement);
4534 if (rdev && !test_bit(Faulty, &rdev->flags))
4539 if (test_bit(R5_InJournal, &dev->flags))
4541 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4544 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4545 /* If there is a failed device being replaced,
4546 * we must be recovering.
4547 * else if we are after recovery_cp, we must be syncing
4548 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4549 * else we can only be replacing
4550 * sync and recovery both need to read all devices, and so
4551 * use the same flag.
4554 sh->sector >= conf->mddev->recovery_cp ||
4555 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4563 static int clear_batch_ready(struct stripe_head *sh)
4565 /* Return '1' if this is a member of batch, or
4566 * '0' if it is a lone stripe or a head which can now be
4569 struct stripe_head *tmp;
4570 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4571 return (sh->batch_head && sh->batch_head != sh);
4572 spin_lock(&sh->stripe_lock);
4573 if (!sh->batch_head) {
4574 spin_unlock(&sh->stripe_lock);
4579 * this stripe could be added to a batch list before we check
4580 * BATCH_READY, skips it
4582 if (sh->batch_head != sh) {
4583 spin_unlock(&sh->stripe_lock);
4586 spin_lock(&sh->batch_lock);
4587 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4588 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4589 spin_unlock(&sh->batch_lock);
4590 spin_unlock(&sh->stripe_lock);
4593 * BATCH_READY is cleared, no new stripes can be added.
4594 * batch_list can be accessed without lock
4599 static void break_stripe_batch_list(struct stripe_head *head_sh,
4600 unsigned long handle_flags)
4602 struct stripe_head *sh, *next;
4606 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4608 list_del_init(&sh->batch_list);
4610 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4611 (1 << STRIPE_SYNCING) |
4612 (1 << STRIPE_REPLACED) |
4613 (1 << STRIPE_DELAYED) |
4614 (1 << STRIPE_BIT_DELAY) |
4615 (1 << STRIPE_FULL_WRITE) |
4616 (1 << STRIPE_BIOFILL_RUN) |
4617 (1 << STRIPE_COMPUTE_RUN) |
4618 (1 << STRIPE_OPS_REQ_PENDING) |
4619 (1 << STRIPE_DISCARD) |
4620 (1 << STRIPE_BATCH_READY) |
4621 (1 << STRIPE_BATCH_ERR) |
4622 (1 << STRIPE_BITMAP_PENDING)),
4623 "stripe state: %lx\n", sh->state);
4624 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4625 (1 << STRIPE_REPLACED)),
4626 "head stripe state: %lx\n", head_sh->state);
4628 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4629 (1 << STRIPE_PREREAD_ACTIVE) |
4630 (1 << STRIPE_DEGRADED) |
4631 (1 << STRIPE_ON_UNPLUG_LIST)),
4632 head_sh->state & (1 << STRIPE_INSYNC));
4634 sh->check_state = head_sh->check_state;
4635 sh->reconstruct_state = head_sh->reconstruct_state;
4636 for (i = 0; i < sh->disks; i++) {
4637 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4639 sh->dev[i].flags = head_sh->dev[i].flags &
4640 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4642 spin_lock_irq(&sh->stripe_lock);
4643 sh->batch_head = NULL;
4644 spin_unlock_irq(&sh->stripe_lock);
4645 if (handle_flags == 0 ||
4646 sh->state & handle_flags)
4647 set_bit(STRIPE_HANDLE, &sh->state);
4648 raid5_release_stripe(sh);
4650 spin_lock_irq(&head_sh->stripe_lock);
4651 head_sh->batch_head = NULL;
4652 spin_unlock_irq(&head_sh->stripe_lock);
4653 for (i = 0; i < head_sh->disks; i++)
4654 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4656 if (head_sh->state & handle_flags)
4657 set_bit(STRIPE_HANDLE, &head_sh->state);
4660 wake_up(&head_sh->raid_conf->wait_for_overlap);
4663 static void handle_stripe(struct stripe_head *sh)
4665 struct stripe_head_state s;
4666 struct r5conf *conf = sh->raid_conf;
4669 int disks = sh->disks;
4670 struct r5dev *pdev, *qdev;
4672 clear_bit(STRIPE_HANDLE, &sh->state);
4673 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4674 /* already being handled, ensure it gets handled
4675 * again when current action finishes */
4676 set_bit(STRIPE_HANDLE, &sh->state);
4680 if (clear_batch_ready(sh) ) {
4681 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4685 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4686 break_stripe_batch_list(sh, 0);
4688 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4689 spin_lock(&sh->stripe_lock);
4691 * Cannot process 'sync' concurrently with 'discard'.
4692 * Flush data in r5cache before 'sync'.
4694 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4695 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4696 !test_bit(STRIPE_DISCARD, &sh->state) &&
4697 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4698 set_bit(STRIPE_SYNCING, &sh->state);
4699 clear_bit(STRIPE_INSYNC, &sh->state);
4700 clear_bit(STRIPE_REPLACED, &sh->state);
4702 spin_unlock(&sh->stripe_lock);
4704 clear_bit(STRIPE_DELAYED, &sh->state);
4706 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4707 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4708 (unsigned long long)sh->sector, sh->state,
4709 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4710 sh->check_state, sh->reconstruct_state);
4712 analyse_stripe(sh, &s);
4714 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4717 if (s.handle_bad_blocks ||
4718 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4719 set_bit(STRIPE_HANDLE, &sh->state);
4723 if (unlikely(s.blocked_rdev)) {
4724 if (s.syncing || s.expanding || s.expanded ||
4725 s.replacing || s.to_write || s.written) {
4726 set_bit(STRIPE_HANDLE, &sh->state);
4729 /* There is nothing for the blocked_rdev to block */
4730 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4731 s.blocked_rdev = NULL;
4734 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4735 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4736 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4739 pr_debug("locked=%d uptodate=%d to_read=%d"
4740 " to_write=%d failed=%d failed_num=%d,%d\n",
4741 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4742 s.failed_num[0], s.failed_num[1]);
4744 * check if the array has lost more than max_degraded devices and,
4745 * if so, some requests might need to be failed.
4747 * When journal device failed (log_failed), we will only process
4748 * the stripe if there is data need write to raid disks
4750 if (s.failed > conf->max_degraded ||
4751 (s.log_failed && s.injournal == 0)) {
4752 sh->check_state = 0;
4753 sh->reconstruct_state = 0;
4754 break_stripe_batch_list(sh, 0);
4755 if (s.to_read+s.to_write+s.written)
4756 handle_failed_stripe(conf, sh, &s, disks);
4757 if (s.syncing + s.replacing)
4758 handle_failed_sync(conf, sh, &s);
4761 /* Now we check to see if any write operations have recently
4765 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4767 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4768 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4769 sh->reconstruct_state = reconstruct_state_idle;
4771 /* All the 'written' buffers and the parity block are ready to
4772 * be written back to disk
4774 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4775 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4776 BUG_ON(sh->qd_idx >= 0 &&
4777 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4778 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4779 for (i = disks; i--; ) {
4780 struct r5dev *dev = &sh->dev[i];
4781 if (test_bit(R5_LOCKED, &dev->flags) &&
4782 (i == sh->pd_idx || i == sh->qd_idx ||
4783 dev->written || test_bit(R5_InJournal,
4785 pr_debug("Writing block %d\n", i);
4786 set_bit(R5_Wantwrite, &dev->flags);
4791 if (!test_bit(R5_Insync, &dev->flags) ||
4792 ((i == sh->pd_idx || i == sh->qd_idx) &&
4794 set_bit(STRIPE_INSYNC, &sh->state);
4797 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4798 s.dec_preread_active = 1;
4802 * might be able to return some write requests if the parity blocks
4803 * are safe, or on a failed drive
4805 pdev = &sh->dev[sh->pd_idx];
4806 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4807 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4808 qdev = &sh->dev[sh->qd_idx];
4809 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4810 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4814 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4815 && !test_bit(R5_LOCKED, &pdev->flags)
4816 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4817 test_bit(R5_Discard, &pdev->flags))))) &&
4818 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4819 && !test_bit(R5_LOCKED, &qdev->flags)
4820 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4821 test_bit(R5_Discard, &qdev->flags))))))
4822 handle_stripe_clean_event(conf, sh, disks);
4825 r5c_handle_cached_data_endio(conf, sh, disks);
4826 log_stripe_write_finished(sh);
4828 /* Now we might consider reading some blocks, either to check/generate
4829 * parity, or to satisfy requests
4830 * or to load a block that is being partially written.
4832 if (s.to_read || s.non_overwrite
4833 || (s.to_write && s.failed)
4834 || (s.syncing && (s.uptodate + s.compute < disks))
4837 handle_stripe_fill(sh, &s, disks);
4840 * When the stripe finishes full journal write cycle (write to journal
4841 * and raid disk), this is the clean up procedure so it is ready for
4844 r5c_finish_stripe_write_out(conf, sh, &s);
4847 * Now to consider new write requests, cache write back and what else,
4848 * if anything should be read. We do not handle new writes when:
4849 * 1/ A 'write' operation (copy+xor) is already in flight.
4850 * 2/ A 'check' operation is in flight, as it may clobber the parity
4852 * 3/ A r5c cache log write is in flight.
4855 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4856 if (!r5c_is_writeback(conf->log)) {
4858 handle_stripe_dirtying(conf, sh, &s, disks);
4859 } else { /* write back cache */
4862 /* First, try handle writes in caching phase */
4864 ret = r5c_try_caching_write(conf, sh, &s,
4867 * If caching phase failed: ret == -EAGAIN
4869 * stripe under reclaim: !caching && injournal
4871 * fall back to handle_stripe_dirtying()
4873 if (ret == -EAGAIN ||
4874 /* stripe under reclaim: !caching && injournal */
4875 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4877 ret = handle_stripe_dirtying(conf, sh, &s,
4885 /* maybe we need to check and possibly fix the parity for this stripe
4886 * Any reads will already have been scheduled, so we just see if enough
4887 * data is available. The parity check is held off while parity
4888 * dependent operations are in flight.
4890 if (sh->check_state ||
4891 (s.syncing && s.locked == 0 &&
4892 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4893 !test_bit(STRIPE_INSYNC, &sh->state))) {
4894 if (conf->level == 6)
4895 handle_parity_checks6(conf, sh, &s, disks);
4897 handle_parity_checks5(conf, sh, &s, disks);
4900 if ((s.replacing || s.syncing) && s.locked == 0
4901 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4902 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4903 /* Write out to replacement devices where possible */
4904 for (i = 0; i < conf->raid_disks; i++)
4905 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4906 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4907 set_bit(R5_WantReplace, &sh->dev[i].flags);
4908 set_bit(R5_LOCKED, &sh->dev[i].flags);
4912 set_bit(STRIPE_INSYNC, &sh->state);
4913 set_bit(STRIPE_REPLACED, &sh->state);
4915 if ((s.syncing || s.replacing) && s.locked == 0 &&
4916 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4917 test_bit(STRIPE_INSYNC, &sh->state)) {
4918 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4919 clear_bit(STRIPE_SYNCING, &sh->state);
4920 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4921 wake_up(&conf->wait_for_overlap);
4924 /* If the failed drives are just a ReadError, then we might need
4925 * to progress the repair/check process
4927 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4928 for (i = 0; i < s.failed; i++) {
4929 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4930 if (test_bit(R5_ReadError, &dev->flags)
4931 && !test_bit(R5_LOCKED, &dev->flags)
4932 && test_bit(R5_UPTODATE, &dev->flags)
4934 if (!test_bit(R5_ReWrite, &dev->flags)) {
4935 set_bit(R5_Wantwrite, &dev->flags);
4936 set_bit(R5_ReWrite, &dev->flags);
4937 set_bit(R5_LOCKED, &dev->flags);
4940 /* let's read it back */
4941 set_bit(R5_Wantread, &dev->flags);
4942 set_bit(R5_LOCKED, &dev->flags);
4948 /* Finish reconstruct operations initiated by the expansion process */
4949 if (sh->reconstruct_state == reconstruct_state_result) {
4950 struct stripe_head *sh_src
4951 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4952 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4953 /* sh cannot be written until sh_src has been read.
4954 * so arrange for sh to be delayed a little
4956 set_bit(STRIPE_DELAYED, &sh->state);
4957 set_bit(STRIPE_HANDLE, &sh->state);
4958 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4960 atomic_inc(&conf->preread_active_stripes);
4961 raid5_release_stripe(sh_src);
4965 raid5_release_stripe(sh_src);
4967 sh->reconstruct_state = reconstruct_state_idle;
4968 clear_bit(STRIPE_EXPANDING, &sh->state);
4969 for (i = conf->raid_disks; i--; ) {
4970 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4971 set_bit(R5_LOCKED, &sh->dev[i].flags);
4976 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4977 !sh->reconstruct_state) {
4978 /* Need to write out all blocks after computing parity */
4979 sh->disks = conf->raid_disks;
4980 stripe_set_idx(sh->sector, conf, 0, sh);
4981 schedule_reconstruction(sh, &s, 1, 1);
4982 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4983 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4984 atomic_dec(&conf->reshape_stripes);
4985 wake_up(&conf->wait_for_overlap);
4986 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4989 if (s.expanding && s.locked == 0 &&
4990 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4991 handle_stripe_expansion(conf, sh);
4994 /* wait for this device to become unblocked */
4995 if (unlikely(s.blocked_rdev)) {
4996 if (conf->mddev->external)
4997 md_wait_for_blocked_rdev(s.blocked_rdev,
5000 /* Internal metadata will immediately
5001 * be written by raid5d, so we don't
5002 * need to wait here.
5004 rdev_dec_pending(s.blocked_rdev,
5008 if (s.handle_bad_blocks)
5009 for (i = disks; i--; ) {
5010 struct md_rdev *rdev;
5011 struct r5dev *dev = &sh->dev[i];
5012 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5013 /* We own a safe reference to the rdev */
5014 rdev = conf->disks[i].rdev;
5015 if (!rdev_set_badblocks(rdev, sh->sector,
5017 md_error(conf->mddev, rdev);
5018 rdev_dec_pending(rdev, conf->mddev);
5020 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5021 rdev = conf->disks[i].rdev;
5022 rdev_clear_badblocks(rdev, sh->sector,
5024 rdev_dec_pending(rdev, conf->mddev);
5026 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5027 rdev = conf->disks[i].replacement;
5029 /* rdev have been moved down */
5030 rdev = conf->disks[i].rdev;
5031 rdev_clear_badblocks(rdev, sh->sector,
5033 rdev_dec_pending(rdev, conf->mddev);
5038 raid_run_ops(sh, s.ops_request);
5042 if (s.dec_preread_active) {
5043 /* We delay this until after ops_run_io so that if make_request
5044 * is waiting on a flush, it won't continue until the writes
5045 * have actually been submitted.
5047 atomic_dec(&conf->preread_active_stripes);
5048 if (atomic_read(&conf->preread_active_stripes) <
5050 md_wakeup_thread(conf->mddev->thread);
5053 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5056 static void raid5_activate_delayed(struct r5conf *conf)
5058 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5059 while (!list_empty(&conf->delayed_list)) {
5060 struct list_head *l = conf->delayed_list.next;
5061 struct stripe_head *sh;
5062 sh = list_entry(l, struct stripe_head, lru);
5064 clear_bit(STRIPE_DELAYED, &sh->state);
5065 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5066 atomic_inc(&conf->preread_active_stripes);
5067 list_add_tail(&sh->lru, &conf->hold_list);
5068 raid5_wakeup_stripe_thread(sh);
5073 static void activate_bit_delay(struct r5conf *conf,
5074 struct list_head *temp_inactive_list)
5076 /* device_lock is held */
5077 struct list_head head;
5078 list_add(&head, &conf->bitmap_list);
5079 list_del_init(&conf->bitmap_list);
5080 while (!list_empty(&head)) {
5081 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5083 list_del_init(&sh->lru);
5084 atomic_inc(&sh->count);
5085 hash = sh->hash_lock_index;
5086 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5090 static int raid5_congested(struct mddev *mddev, int bits)
5092 struct r5conf *conf = mddev->private;
5094 /* No difference between reads and writes. Just check
5095 * how busy the stripe_cache is
5098 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5101 /* Also checks whether there is pressure on r5cache log space */
5102 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5106 if (atomic_read(&conf->empty_inactive_list_nr))
5112 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5114 struct r5conf *conf = mddev->private;
5115 sector_t sector = bio->bi_iter.bi_sector;
5116 unsigned int chunk_sectors;
5117 unsigned int bio_sectors = bio_sectors(bio);
5119 WARN_ON_ONCE(bio->bi_partno);
5121 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5122 return chunk_sectors >=
5123 ((sector & (chunk_sectors - 1)) + bio_sectors);
5127 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5128 * later sampled by raid5d.
5130 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5132 unsigned long flags;
5134 spin_lock_irqsave(&conf->device_lock, flags);
5136 bi->bi_next = conf->retry_read_aligned_list;
5137 conf->retry_read_aligned_list = bi;
5139 spin_unlock_irqrestore(&conf->device_lock, flags);
5140 md_wakeup_thread(conf->mddev->thread);
5143 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5144 unsigned int *offset)
5148 bi = conf->retry_read_aligned;
5150 *offset = conf->retry_read_offset;
5151 conf->retry_read_aligned = NULL;
5154 bi = conf->retry_read_aligned_list;
5156 conf->retry_read_aligned_list = bi->bi_next;
5165 * The "raid5_align_endio" should check if the read succeeded and if it
5166 * did, call bio_endio on the original bio (having bio_put the new bio
5168 * If the read failed..
5170 static void raid5_align_endio(struct bio *bi)
5172 struct bio* raid_bi = bi->bi_private;
5173 struct mddev *mddev;
5174 struct r5conf *conf;
5175 struct md_rdev *rdev;
5176 blk_status_t error = bi->bi_status;
5180 rdev = (void*)raid_bi->bi_next;
5181 raid_bi->bi_next = NULL;
5182 mddev = rdev->mddev;
5183 conf = mddev->private;
5185 rdev_dec_pending(rdev, conf->mddev);
5189 if (atomic_dec_and_test(&conf->active_aligned_reads))
5190 wake_up(&conf->wait_for_quiescent);
5194 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5196 add_bio_to_retry(raid_bi, conf);
5199 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5201 struct r5conf *conf = mddev->private;
5203 struct bio* align_bi;
5204 struct md_rdev *rdev;
5205 sector_t end_sector;
5207 if (!in_chunk_boundary(mddev, raid_bio)) {
5208 pr_debug("%s: non aligned\n", __func__);
5212 * use bio_clone_fast to make a copy of the bio
5214 align_bi = bio_clone_fast(raid_bio, GFP_NOIO, mddev->bio_set);
5218 * set bi_end_io to a new function, and set bi_private to the
5221 align_bi->bi_end_io = raid5_align_endio;
5222 align_bi->bi_private = raid_bio;
5226 align_bi->bi_iter.bi_sector =
5227 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5230 end_sector = bio_end_sector(align_bi);
5232 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5233 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5234 rdev->recovery_offset < end_sector) {
5235 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5237 (test_bit(Faulty, &rdev->flags) ||
5238 !(test_bit(In_sync, &rdev->flags) ||
5239 rdev->recovery_offset >= end_sector)))
5243 if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5253 atomic_inc(&rdev->nr_pending);
5255 raid_bio->bi_next = (void*)rdev;
5256 bio_set_dev(align_bi, rdev->bdev);
5257 bio_clear_flag(align_bi, BIO_SEG_VALID);
5259 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5260 bio_sectors(align_bi),
5261 &first_bad, &bad_sectors)) {
5263 rdev_dec_pending(rdev, mddev);
5267 /* No reshape active, so we can trust rdev->data_offset */
5268 align_bi->bi_iter.bi_sector += rdev->data_offset;
5270 spin_lock_irq(&conf->device_lock);
5271 wait_event_lock_irq(conf->wait_for_quiescent,
5274 atomic_inc(&conf->active_aligned_reads);
5275 spin_unlock_irq(&conf->device_lock);
5278 trace_block_bio_remap(align_bi->bi_disk->queue,
5279 align_bi, disk_devt(mddev->gendisk),
5280 raid_bio->bi_iter.bi_sector);
5281 generic_make_request(align_bi);
5290 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5293 sector_t sector = raid_bio->bi_iter.bi_sector;
5294 unsigned chunk_sects = mddev->chunk_sectors;
5295 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5297 if (sectors < bio_sectors(raid_bio)) {
5298 struct r5conf *conf = mddev->private;
5299 split = bio_split(raid_bio, sectors, GFP_NOIO, conf->bio_split);
5300 bio_chain(split, raid_bio);
5301 generic_make_request(raid_bio);
5305 if (!raid5_read_one_chunk(mddev, raid_bio))
5311 /* __get_priority_stripe - get the next stripe to process
5313 * Full stripe writes are allowed to pass preread active stripes up until
5314 * the bypass_threshold is exceeded. In general the bypass_count
5315 * increments when the handle_list is handled before the hold_list; however, it
5316 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5317 * stripe with in flight i/o. The bypass_count will be reset when the
5318 * head of the hold_list has changed, i.e. the head was promoted to the
5321 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5323 struct stripe_head *sh, *tmp;
5324 struct list_head *handle_list = NULL;
5325 struct r5worker_group *wg;
5326 bool second_try = !r5c_is_writeback(conf->log) &&
5327 !r5l_log_disk_error(conf);
5328 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5329 r5l_log_disk_error(conf);
5334 if (conf->worker_cnt_per_group == 0) {
5335 handle_list = try_loprio ? &conf->loprio_list :
5337 } else if (group != ANY_GROUP) {
5338 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5339 &conf->worker_groups[group].handle_list;
5340 wg = &conf->worker_groups[group];
5343 for (i = 0; i < conf->group_cnt; i++) {
5344 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5345 &conf->worker_groups[i].handle_list;
5346 wg = &conf->worker_groups[i];
5347 if (!list_empty(handle_list))
5352 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5354 list_empty(handle_list) ? "empty" : "busy",
5355 list_empty(&conf->hold_list) ? "empty" : "busy",
5356 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5358 if (!list_empty(handle_list)) {
5359 sh = list_entry(handle_list->next, typeof(*sh), lru);
5361 if (list_empty(&conf->hold_list))
5362 conf->bypass_count = 0;
5363 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5364 if (conf->hold_list.next == conf->last_hold)
5365 conf->bypass_count++;
5367 conf->last_hold = conf->hold_list.next;
5368 conf->bypass_count -= conf->bypass_threshold;
5369 if (conf->bypass_count < 0)
5370 conf->bypass_count = 0;
5373 } else if (!list_empty(&conf->hold_list) &&
5374 ((conf->bypass_threshold &&
5375 conf->bypass_count > conf->bypass_threshold) ||
5376 atomic_read(&conf->pending_full_writes) == 0)) {
5378 list_for_each_entry(tmp, &conf->hold_list, lru) {
5379 if (conf->worker_cnt_per_group == 0 ||
5380 group == ANY_GROUP ||
5381 !cpu_online(tmp->cpu) ||
5382 cpu_to_group(tmp->cpu) == group) {
5389 conf->bypass_count -= conf->bypass_threshold;
5390 if (conf->bypass_count < 0)
5391 conf->bypass_count = 0;
5400 try_loprio = !try_loprio;
5408 list_del_init(&sh->lru);
5409 BUG_ON(atomic_inc_return(&sh->count) != 1);
5413 struct raid5_plug_cb {
5414 struct blk_plug_cb cb;
5415 struct list_head list;
5416 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5419 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5421 struct raid5_plug_cb *cb = container_of(
5422 blk_cb, struct raid5_plug_cb, cb);
5423 struct stripe_head *sh;
5424 struct mddev *mddev = cb->cb.data;
5425 struct r5conf *conf = mddev->private;
5429 if (cb->list.next && !list_empty(&cb->list)) {
5430 spin_lock_irq(&conf->device_lock);
5431 while (!list_empty(&cb->list)) {
5432 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5433 list_del_init(&sh->lru);
5435 * avoid race release_stripe_plug() sees
5436 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5437 * is still in our list
5439 smp_mb__before_atomic();
5440 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5442 * STRIPE_ON_RELEASE_LIST could be set here. In that
5443 * case, the count is always > 1 here
5445 hash = sh->hash_lock_index;
5446 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5449 spin_unlock_irq(&conf->device_lock);
5451 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5452 NR_STRIPE_HASH_LOCKS);
5454 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5458 static void release_stripe_plug(struct mddev *mddev,
5459 struct stripe_head *sh)
5461 struct blk_plug_cb *blk_cb = blk_check_plugged(
5462 raid5_unplug, mddev,
5463 sizeof(struct raid5_plug_cb));
5464 struct raid5_plug_cb *cb;
5467 raid5_release_stripe(sh);
5471 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5473 if (cb->list.next == NULL) {
5475 INIT_LIST_HEAD(&cb->list);
5476 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5477 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5480 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5481 list_add_tail(&sh->lru, &cb->list);
5483 raid5_release_stripe(sh);
5486 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5488 struct r5conf *conf = mddev->private;
5489 sector_t logical_sector, last_sector;
5490 struct stripe_head *sh;
5493 if (mddev->reshape_position != MaxSector)
5494 /* Skip discard while reshape is happening */
5497 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5498 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5502 stripe_sectors = conf->chunk_sectors *
5503 (conf->raid_disks - conf->max_degraded);
5504 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5506 sector_div(last_sector, stripe_sectors);
5508 logical_sector *= conf->chunk_sectors;
5509 last_sector *= conf->chunk_sectors;
5511 for (; logical_sector < last_sector;
5512 logical_sector += STRIPE_SECTORS) {
5516 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5517 prepare_to_wait(&conf->wait_for_overlap, &w,
5518 TASK_UNINTERRUPTIBLE);
5519 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5520 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5521 raid5_release_stripe(sh);
5525 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5526 spin_lock_irq(&sh->stripe_lock);
5527 for (d = 0; d < conf->raid_disks; d++) {
5528 if (d == sh->pd_idx || d == sh->qd_idx)
5530 if (sh->dev[d].towrite || sh->dev[d].toread) {
5531 set_bit(R5_Overlap, &sh->dev[d].flags);
5532 spin_unlock_irq(&sh->stripe_lock);
5533 raid5_release_stripe(sh);
5538 set_bit(STRIPE_DISCARD, &sh->state);
5539 finish_wait(&conf->wait_for_overlap, &w);
5540 sh->overwrite_disks = 0;
5541 for (d = 0; d < conf->raid_disks; d++) {
5542 if (d == sh->pd_idx || d == sh->qd_idx)
5544 sh->dev[d].towrite = bi;
5545 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5546 bio_inc_remaining(bi);
5547 md_write_inc(mddev, bi);
5548 sh->overwrite_disks++;
5550 spin_unlock_irq(&sh->stripe_lock);
5551 if (conf->mddev->bitmap) {
5553 d < conf->raid_disks - conf->max_degraded;
5555 bitmap_startwrite(mddev->bitmap,
5559 sh->bm_seq = conf->seq_flush + 1;
5560 set_bit(STRIPE_BIT_DELAY, &sh->state);
5563 set_bit(STRIPE_HANDLE, &sh->state);
5564 clear_bit(STRIPE_DELAYED, &sh->state);
5565 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5566 atomic_inc(&conf->preread_active_stripes);
5567 release_stripe_plug(mddev, sh);
5573 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5575 struct r5conf *conf = mddev->private;
5577 sector_t new_sector;
5578 sector_t logical_sector, last_sector;
5579 struct stripe_head *sh;
5580 const int rw = bio_data_dir(bi);
5583 bool do_flush = false;
5585 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5586 int ret = r5l_handle_flush_request(conf->log, bi);
5590 if (ret == -ENODEV) {
5591 md_flush_request(mddev, bi);
5594 /* ret == -EAGAIN, fallback */
5596 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5597 * we need to flush journal device
5599 do_flush = bi->bi_opf & REQ_PREFLUSH;
5602 if (!md_write_start(mddev, bi))
5605 * If array is degraded, better not do chunk aligned read because
5606 * later we might have to read it again in order to reconstruct
5607 * data on failed drives.
5609 if (rw == READ && mddev->degraded == 0 &&
5610 mddev->reshape_position == MaxSector) {
5611 bi = chunk_aligned_read(mddev, bi);
5616 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5617 make_discard_request(mddev, bi);
5618 md_write_end(mddev);
5622 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5623 last_sector = bio_end_sector(bi);
5626 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5627 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5633 seq = read_seqcount_begin(&conf->gen_lock);
5636 prepare_to_wait(&conf->wait_for_overlap, &w,
5637 TASK_UNINTERRUPTIBLE);
5638 if (unlikely(conf->reshape_progress != MaxSector)) {
5639 /* spinlock is needed as reshape_progress may be
5640 * 64bit on a 32bit platform, and so it might be
5641 * possible to see a half-updated value
5642 * Of course reshape_progress could change after
5643 * the lock is dropped, so once we get a reference
5644 * to the stripe that we think it is, we will have
5647 spin_lock_irq(&conf->device_lock);
5648 if (mddev->reshape_backwards
5649 ? logical_sector < conf->reshape_progress
5650 : logical_sector >= conf->reshape_progress) {
5653 if (mddev->reshape_backwards
5654 ? logical_sector < conf->reshape_safe
5655 : logical_sector >= conf->reshape_safe) {
5656 spin_unlock_irq(&conf->device_lock);
5662 spin_unlock_irq(&conf->device_lock);
5665 new_sector = raid5_compute_sector(conf, logical_sector,
5668 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5669 (unsigned long long)new_sector,
5670 (unsigned long long)logical_sector);
5672 sh = raid5_get_active_stripe(conf, new_sector, previous,
5673 (bi->bi_opf & REQ_RAHEAD), 0);
5675 if (unlikely(previous)) {
5676 /* expansion might have moved on while waiting for a
5677 * stripe, so we must do the range check again.
5678 * Expansion could still move past after this
5679 * test, but as we are holding a reference to
5680 * 'sh', we know that if that happens,
5681 * STRIPE_EXPANDING will get set and the expansion
5682 * won't proceed until we finish with the stripe.
5685 spin_lock_irq(&conf->device_lock);
5686 if (mddev->reshape_backwards
5687 ? logical_sector >= conf->reshape_progress
5688 : logical_sector < conf->reshape_progress)
5689 /* mismatch, need to try again */
5691 spin_unlock_irq(&conf->device_lock);
5693 raid5_release_stripe(sh);
5699 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5700 /* Might have got the wrong stripe_head
5703 raid5_release_stripe(sh);
5707 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5708 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5709 /* Stripe is busy expanding or
5710 * add failed due to overlap. Flush everything
5713 md_wakeup_thread(mddev->thread);
5714 raid5_release_stripe(sh);
5720 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5721 /* we only need flush for one stripe */
5725 if (!sh->batch_head || sh == sh->batch_head)
5726 set_bit(STRIPE_HANDLE, &sh->state);
5727 clear_bit(STRIPE_DELAYED, &sh->state);
5728 if ((!sh->batch_head || sh == sh->batch_head) &&
5729 (bi->bi_opf & REQ_SYNC) &&
5730 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5731 atomic_inc(&conf->preread_active_stripes);
5732 release_stripe_plug(mddev, sh);
5734 /* cannot get stripe for read-ahead, just give-up */
5735 bi->bi_status = BLK_STS_IOERR;
5739 finish_wait(&conf->wait_for_overlap, &w);
5742 md_write_end(mddev);
5747 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5749 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5751 /* reshaping is quite different to recovery/resync so it is
5752 * handled quite separately ... here.
5754 * On each call to sync_request, we gather one chunk worth of
5755 * destination stripes and flag them as expanding.
5756 * Then we find all the source stripes and request reads.
5757 * As the reads complete, handle_stripe will copy the data
5758 * into the destination stripe and release that stripe.
5760 struct r5conf *conf = mddev->private;
5761 struct stripe_head *sh;
5762 sector_t first_sector, last_sector;
5763 int raid_disks = conf->previous_raid_disks;
5764 int data_disks = raid_disks - conf->max_degraded;
5765 int new_data_disks = conf->raid_disks - conf->max_degraded;
5768 sector_t writepos, readpos, safepos;
5769 sector_t stripe_addr;
5770 int reshape_sectors;
5771 struct list_head stripes;
5774 if (sector_nr == 0) {
5775 /* If restarting in the middle, skip the initial sectors */
5776 if (mddev->reshape_backwards &&
5777 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5778 sector_nr = raid5_size(mddev, 0, 0)
5779 - conf->reshape_progress;
5780 } else if (mddev->reshape_backwards &&
5781 conf->reshape_progress == MaxSector) {
5782 /* shouldn't happen, but just in case, finish up.*/
5783 sector_nr = MaxSector;
5784 } else if (!mddev->reshape_backwards &&
5785 conf->reshape_progress > 0)
5786 sector_nr = conf->reshape_progress;
5787 sector_div(sector_nr, new_data_disks);
5789 mddev->curr_resync_completed = sector_nr;
5790 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5797 /* We need to process a full chunk at a time.
5798 * If old and new chunk sizes differ, we need to process the
5802 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5804 /* We update the metadata at least every 10 seconds, or when
5805 * the data about to be copied would over-write the source of
5806 * the data at the front of the range. i.e. one new_stripe
5807 * along from reshape_progress new_maps to after where
5808 * reshape_safe old_maps to
5810 writepos = conf->reshape_progress;
5811 sector_div(writepos, new_data_disks);
5812 readpos = conf->reshape_progress;
5813 sector_div(readpos, data_disks);
5814 safepos = conf->reshape_safe;
5815 sector_div(safepos, data_disks);
5816 if (mddev->reshape_backwards) {
5817 BUG_ON(writepos < reshape_sectors);
5818 writepos -= reshape_sectors;
5819 readpos += reshape_sectors;
5820 safepos += reshape_sectors;
5822 writepos += reshape_sectors;
5823 /* readpos and safepos are worst-case calculations.
5824 * A negative number is overly pessimistic, and causes
5825 * obvious problems for unsigned storage. So clip to 0.
5827 readpos -= min_t(sector_t, reshape_sectors, readpos);
5828 safepos -= min_t(sector_t, reshape_sectors, safepos);
5831 /* Having calculated the 'writepos' possibly use it
5832 * to set 'stripe_addr' which is where we will write to.
5834 if (mddev->reshape_backwards) {
5835 BUG_ON(conf->reshape_progress == 0);
5836 stripe_addr = writepos;
5837 BUG_ON((mddev->dev_sectors &
5838 ~((sector_t)reshape_sectors - 1))
5839 - reshape_sectors - stripe_addr
5842 BUG_ON(writepos != sector_nr + reshape_sectors);
5843 stripe_addr = sector_nr;
5846 /* 'writepos' is the most advanced device address we might write.
5847 * 'readpos' is the least advanced device address we might read.
5848 * 'safepos' is the least address recorded in the metadata as having
5850 * If there is a min_offset_diff, these are adjusted either by
5851 * increasing the safepos/readpos if diff is negative, or
5852 * increasing writepos if diff is positive.
5853 * If 'readpos' is then behind 'writepos', there is no way that we can
5854 * ensure safety in the face of a crash - that must be done by userspace
5855 * making a backup of the data. So in that case there is no particular
5856 * rush to update metadata.
5857 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5858 * update the metadata to advance 'safepos' to match 'readpos' so that
5859 * we can be safe in the event of a crash.
5860 * So we insist on updating metadata if safepos is behind writepos and
5861 * readpos is beyond writepos.
5862 * In any case, update the metadata every 10 seconds.
5863 * Maybe that number should be configurable, but I'm not sure it is
5864 * worth it.... maybe it could be a multiple of safemode_delay???
5866 if (conf->min_offset_diff < 0) {
5867 safepos += -conf->min_offset_diff;
5868 readpos += -conf->min_offset_diff;
5870 writepos += conf->min_offset_diff;
5872 if ((mddev->reshape_backwards
5873 ? (safepos > writepos && readpos < writepos)
5874 : (safepos < writepos && readpos > writepos)) ||
5875 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5876 /* Cannot proceed until we've updated the superblock... */
5877 wait_event(conf->wait_for_overlap,
5878 atomic_read(&conf->reshape_stripes)==0
5879 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5880 if (atomic_read(&conf->reshape_stripes) != 0)
5882 mddev->reshape_position = conf->reshape_progress;
5883 mddev->curr_resync_completed = sector_nr;
5884 conf->reshape_checkpoint = jiffies;
5885 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5886 md_wakeup_thread(mddev->thread);
5887 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5888 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5889 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5891 spin_lock_irq(&conf->device_lock);
5892 conf->reshape_safe = mddev->reshape_position;
5893 spin_unlock_irq(&conf->device_lock);
5894 wake_up(&conf->wait_for_overlap);
5895 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5898 INIT_LIST_HEAD(&stripes);
5899 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5901 int skipped_disk = 0;
5902 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5903 set_bit(STRIPE_EXPANDING, &sh->state);
5904 atomic_inc(&conf->reshape_stripes);
5905 /* If any of this stripe is beyond the end of the old
5906 * array, then we need to zero those blocks
5908 for (j=sh->disks; j--;) {
5910 if (j == sh->pd_idx)
5912 if (conf->level == 6 &&
5915 s = raid5_compute_blocknr(sh, j, 0);
5916 if (s < raid5_size(mddev, 0, 0)) {
5920 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5921 set_bit(R5_Expanded, &sh->dev[j].flags);
5922 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5924 if (!skipped_disk) {
5925 set_bit(STRIPE_EXPAND_READY, &sh->state);
5926 set_bit(STRIPE_HANDLE, &sh->state);
5928 list_add(&sh->lru, &stripes);
5930 spin_lock_irq(&conf->device_lock);
5931 if (mddev->reshape_backwards)
5932 conf->reshape_progress -= reshape_sectors * new_data_disks;
5934 conf->reshape_progress += reshape_sectors * new_data_disks;
5935 spin_unlock_irq(&conf->device_lock);
5936 /* Ok, those stripe are ready. We can start scheduling
5937 * reads on the source stripes.
5938 * The source stripes are determined by mapping the first and last
5939 * block on the destination stripes.
5942 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5945 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5946 * new_data_disks - 1),
5948 if (last_sector >= mddev->dev_sectors)
5949 last_sector = mddev->dev_sectors - 1;
5950 while (first_sector <= last_sector) {
5951 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5952 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5953 set_bit(STRIPE_HANDLE, &sh->state);
5954 raid5_release_stripe(sh);
5955 first_sector += STRIPE_SECTORS;
5957 /* Now that the sources are clearly marked, we can release
5958 * the destination stripes
5960 while (!list_empty(&stripes)) {
5961 sh = list_entry(stripes.next, struct stripe_head, lru);
5962 list_del_init(&sh->lru);
5963 raid5_release_stripe(sh);
5965 /* If this takes us to the resync_max point where we have to pause,
5966 * then we need to write out the superblock.
5968 sector_nr += reshape_sectors;
5969 retn = reshape_sectors;
5971 if (mddev->curr_resync_completed > mddev->resync_max ||
5972 (sector_nr - mddev->curr_resync_completed) * 2
5973 >= mddev->resync_max - mddev->curr_resync_completed) {
5974 /* Cannot proceed until we've updated the superblock... */
5975 wait_event(conf->wait_for_overlap,
5976 atomic_read(&conf->reshape_stripes) == 0
5977 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5978 if (atomic_read(&conf->reshape_stripes) != 0)
5980 mddev->reshape_position = conf->reshape_progress;
5981 mddev->curr_resync_completed = sector_nr;
5982 conf->reshape_checkpoint = jiffies;
5983 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5984 md_wakeup_thread(mddev->thread);
5985 wait_event(mddev->sb_wait,
5986 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5987 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5988 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5990 spin_lock_irq(&conf->device_lock);
5991 conf->reshape_safe = mddev->reshape_position;
5992 spin_unlock_irq(&conf->device_lock);
5993 wake_up(&conf->wait_for_overlap);
5994 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6000 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6003 struct r5conf *conf = mddev->private;
6004 struct stripe_head *sh;
6005 sector_t max_sector = mddev->dev_sectors;
6006 sector_t sync_blocks;
6007 int still_degraded = 0;
6010 if (sector_nr >= max_sector) {
6011 /* just being told to finish up .. nothing much to do */
6013 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6018 if (mddev->curr_resync < max_sector) /* aborted */
6019 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6021 else /* completed sync */
6023 bitmap_close_sync(mddev->bitmap);
6028 /* Allow raid5_quiesce to complete */
6029 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6031 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6032 return reshape_request(mddev, sector_nr, skipped);
6034 /* No need to check resync_max as we never do more than one
6035 * stripe, and as resync_max will always be on a chunk boundary,
6036 * if the check in md_do_sync didn't fire, there is no chance
6037 * of overstepping resync_max here
6040 /* if there is too many failed drives and we are trying
6041 * to resync, then assert that we are finished, because there is
6042 * nothing we can do.
6044 if (mddev->degraded >= conf->max_degraded &&
6045 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6046 sector_t rv = mddev->dev_sectors - sector_nr;
6050 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6052 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6053 sync_blocks >= STRIPE_SECTORS) {
6054 /* we can skip this block, and probably more */
6055 sync_blocks /= STRIPE_SECTORS;
6057 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6060 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6062 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6064 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6065 /* make sure we don't swamp the stripe cache if someone else
6066 * is trying to get access
6068 schedule_timeout_uninterruptible(1);
6070 /* Need to check if array will still be degraded after recovery/resync
6071 * Note in case of > 1 drive failures it's possible we're rebuilding
6072 * one drive while leaving another faulty drive in array.
6075 for (i = 0; i < conf->raid_disks; i++) {
6076 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
6078 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6083 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6085 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6086 set_bit(STRIPE_HANDLE, &sh->state);
6088 raid5_release_stripe(sh);
6090 return STRIPE_SECTORS;
6093 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6094 unsigned int offset)
6096 /* We may not be able to submit a whole bio at once as there
6097 * may not be enough stripe_heads available.
6098 * We cannot pre-allocate enough stripe_heads as we may need
6099 * more than exist in the cache (if we allow ever large chunks).
6100 * So we do one stripe head at a time and record in
6101 * ->bi_hw_segments how many have been done.
6103 * We *know* that this entire raid_bio is in one chunk, so
6104 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6106 struct stripe_head *sh;
6108 sector_t sector, logical_sector, last_sector;
6112 logical_sector = raid_bio->bi_iter.bi_sector &
6113 ~((sector_t)STRIPE_SECTORS-1);
6114 sector = raid5_compute_sector(conf, logical_sector,
6116 last_sector = bio_end_sector(raid_bio);
6118 for (; logical_sector < last_sector;
6119 logical_sector += STRIPE_SECTORS,
6120 sector += STRIPE_SECTORS,
6124 /* already done this stripe */
6127 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6130 /* failed to get a stripe - must wait */
6131 conf->retry_read_aligned = raid_bio;
6132 conf->retry_read_offset = scnt;
6136 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6137 raid5_release_stripe(sh);
6138 conf->retry_read_aligned = raid_bio;
6139 conf->retry_read_offset = scnt;
6143 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6145 raid5_release_stripe(sh);
6149 bio_endio(raid_bio);
6151 if (atomic_dec_and_test(&conf->active_aligned_reads))
6152 wake_up(&conf->wait_for_quiescent);
6156 static int handle_active_stripes(struct r5conf *conf, int group,
6157 struct r5worker *worker,
6158 struct list_head *temp_inactive_list)
6160 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6161 int i, batch_size = 0, hash;
6162 bool release_inactive = false;
6164 while (batch_size < MAX_STRIPE_BATCH &&
6165 (sh = __get_priority_stripe(conf, group)) != NULL)
6166 batch[batch_size++] = sh;
6168 if (batch_size == 0) {
6169 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6170 if (!list_empty(temp_inactive_list + i))
6172 if (i == NR_STRIPE_HASH_LOCKS) {
6173 spin_unlock_irq(&conf->device_lock);
6174 r5l_flush_stripe_to_raid(conf->log);
6175 spin_lock_irq(&conf->device_lock);
6178 release_inactive = true;
6180 spin_unlock_irq(&conf->device_lock);
6182 release_inactive_stripe_list(conf, temp_inactive_list,
6183 NR_STRIPE_HASH_LOCKS);
6185 r5l_flush_stripe_to_raid(conf->log);
6186 if (release_inactive) {
6187 spin_lock_irq(&conf->device_lock);
6191 for (i = 0; i < batch_size; i++)
6192 handle_stripe(batch[i]);
6193 log_write_stripe_run(conf);
6197 spin_lock_irq(&conf->device_lock);
6198 for (i = 0; i < batch_size; i++) {
6199 hash = batch[i]->hash_lock_index;
6200 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6205 static void raid5_do_work(struct work_struct *work)
6207 struct r5worker *worker = container_of(work, struct r5worker, work);
6208 struct r5worker_group *group = worker->group;
6209 struct r5conf *conf = group->conf;
6210 struct mddev *mddev = conf->mddev;
6211 int group_id = group - conf->worker_groups;
6213 struct blk_plug plug;
6215 pr_debug("+++ raid5worker active\n");
6217 blk_start_plug(&plug);
6219 spin_lock_irq(&conf->device_lock);
6221 int batch_size, released;
6223 released = release_stripe_list(conf, worker->temp_inactive_list);
6225 batch_size = handle_active_stripes(conf, group_id, worker,
6226 worker->temp_inactive_list);
6227 worker->working = false;
6228 if (!batch_size && !released)
6230 handled += batch_size;
6231 wait_event_lock_irq(mddev->sb_wait,
6232 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6235 pr_debug("%d stripes handled\n", handled);
6237 spin_unlock_irq(&conf->device_lock);
6239 flush_deferred_bios(conf);
6241 r5l_flush_stripe_to_raid(conf->log);
6243 async_tx_issue_pending_all();
6244 blk_finish_plug(&plug);
6246 pr_debug("--- raid5worker inactive\n");
6250 * This is our raid5 kernel thread.
6252 * We scan the hash table for stripes which can be handled now.
6253 * During the scan, completed stripes are saved for us by the interrupt
6254 * handler, so that they will not have to wait for our next wakeup.
6256 static void raid5d(struct md_thread *thread)
6258 struct mddev *mddev = thread->mddev;
6259 struct r5conf *conf = mddev->private;
6261 struct blk_plug plug;
6263 pr_debug("+++ raid5d active\n");
6265 md_check_recovery(mddev);
6267 blk_start_plug(&plug);
6269 spin_lock_irq(&conf->device_lock);
6272 int batch_size, released;
6273 unsigned int offset;
6275 released = release_stripe_list(conf, conf->temp_inactive_list);
6277 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6280 !list_empty(&conf->bitmap_list)) {
6281 /* Now is a good time to flush some bitmap updates */
6283 spin_unlock_irq(&conf->device_lock);
6284 bitmap_unplug(mddev->bitmap);
6285 spin_lock_irq(&conf->device_lock);
6286 conf->seq_write = conf->seq_flush;
6287 activate_bit_delay(conf, conf->temp_inactive_list);
6289 raid5_activate_delayed(conf);
6291 while ((bio = remove_bio_from_retry(conf, &offset))) {
6293 spin_unlock_irq(&conf->device_lock);
6294 ok = retry_aligned_read(conf, bio, offset);
6295 spin_lock_irq(&conf->device_lock);
6301 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6302 conf->temp_inactive_list);
6303 if (!batch_size && !released)
6305 handled += batch_size;
6307 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6308 spin_unlock_irq(&conf->device_lock);
6309 md_check_recovery(mddev);
6310 spin_lock_irq(&conf->device_lock);
6313 pr_debug("%d stripes handled\n", handled);
6315 spin_unlock_irq(&conf->device_lock);
6316 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6317 mutex_trylock(&conf->cache_size_mutex)) {
6318 grow_one_stripe(conf, __GFP_NOWARN);
6319 /* Set flag even if allocation failed. This helps
6320 * slow down allocation requests when mem is short
6322 set_bit(R5_DID_ALLOC, &conf->cache_state);
6323 mutex_unlock(&conf->cache_size_mutex);
6326 flush_deferred_bios(conf);
6328 r5l_flush_stripe_to_raid(conf->log);
6330 async_tx_issue_pending_all();
6331 blk_finish_plug(&plug);
6333 pr_debug("--- raid5d inactive\n");
6337 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6339 struct r5conf *conf;
6341 spin_lock(&mddev->lock);
6342 conf = mddev->private;
6344 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6345 spin_unlock(&mddev->lock);
6350 raid5_set_cache_size(struct mddev *mddev, int size)
6353 struct r5conf *conf = mddev->private;
6355 if (size <= 16 || size > 32768)
6358 conf->min_nr_stripes = size;
6359 mutex_lock(&conf->cache_size_mutex);
6360 while (size < conf->max_nr_stripes &&
6361 drop_one_stripe(conf))
6363 mutex_unlock(&conf->cache_size_mutex);
6365 md_allow_write(mddev);
6367 mutex_lock(&conf->cache_size_mutex);
6368 while (size > conf->max_nr_stripes)
6369 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6370 conf->min_nr_stripes = conf->max_nr_stripes;
6374 mutex_unlock(&conf->cache_size_mutex);
6378 EXPORT_SYMBOL(raid5_set_cache_size);
6381 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6383 struct r5conf *conf;
6387 if (len >= PAGE_SIZE)
6389 if (kstrtoul(page, 10, &new))
6391 err = mddev_lock(mddev);
6394 conf = mddev->private;
6398 err = raid5_set_cache_size(mddev, new);
6399 mddev_unlock(mddev);
6404 static struct md_sysfs_entry
6405 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6406 raid5_show_stripe_cache_size,
6407 raid5_store_stripe_cache_size);
6410 raid5_show_rmw_level(struct mddev *mddev, char *page)
6412 struct r5conf *conf = mddev->private;
6414 return sprintf(page, "%d\n", conf->rmw_level);
6420 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6422 struct r5conf *conf = mddev->private;
6428 if (len >= PAGE_SIZE)
6431 if (kstrtoul(page, 10, &new))
6434 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6437 if (new != PARITY_DISABLE_RMW &&
6438 new != PARITY_ENABLE_RMW &&
6439 new != PARITY_PREFER_RMW)
6442 conf->rmw_level = new;
6446 static struct md_sysfs_entry
6447 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6448 raid5_show_rmw_level,
6449 raid5_store_rmw_level);
6453 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6455 struct r5conf *conf;
6457 spin_lock(&mddev->lock);
6458 conf = mddev->private;
6460 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6461 spin_unlock(&mddev->lock);
6466 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6468 struct r5conf *conf;
6472 if (len >= PAGE_SIZE)
6474 if (kstrtoul(page, 10, &new))
6477 err = mddev_lock(mddev);
6480 conf = mddev->private;
6483 else if (new > conf->min_nr_stripes)
6486 conf->bypass_threshold = new;
6487 mddev_unlock(mddev);
6491 static struct md_sysfs_entry
6492 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6494 raid5_show_preread_threshold,
6495 raid5_store_preread_threshold);
6498 raid5_show_skip_copy(struct mddev *mddev, char *page)
6500 struct r5conf *conf;
6502 spin_lock(&mddev->lock);
6503 conf = mddev->private;
6505 ret = sprintf(page, "%d\n", conf->skip_copy);
6506 spin_unlock(&mddev->lock);
6511 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6513 struct r5conf *conf;
6517 if (len >= PAGE_SIZE)
6519 if (kstrtoul(page, 10, &new))
6523 err = mddev_lock(mddev);
6526 conf = mddev->private;
6529 else if (new != conf->skip_copy) {
6530 mddev_suspend(mddev);
6531 conf->skip_copy = new;
6533 mddev->queue->backing_dev_info->capabilities |=
6534 BDI_CAP_STABLE_WRITES;
6536 mddev->queue->backing_dev_info->capabilities &=
6537 ~BDI_CAP_STABLE_WRITES;
6538 mddev_resume(mddev);
6540 mddev_unlock(mddev);
6544 static struct md_sysfs_entry
6545 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6546 raid5_show_skip_copy,
6547 raid5_store_skip_copy);
6550 stripe_cache_active_show(struct mddev *mddev, char *page)
6552 struct r5conf *conf = mddev->private;
6554 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6559 static struct md_sysfs_entry
6560 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6563 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6565 struct r5conf *conf;
6567 spin_lock(&mddev->lock);
6568 conf = mddev->private;
6570 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6571 spin_unlock(&mddev->lock);
6575 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6577 int *worker_cnt_per_group,
6578 struct r5worker_group **worker_groups);
6580 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6582 struct r5conf *conf;
6585 struct r5worker_group *new_groups, *old_groups;
6586 int group_cnt, worker_cnt_per_group;
6588 if (len >= PAGE_SIZE)
6590 if (kstrtouint(page, 10, &new))
6592 /* 8192 should be big enough */
6596 err = mddev_lock(mddev);
6599 conf = mddev->private;
6602 else if (new != conf->worker_cnt_per_group) {
6603 mddev_suspend(mddev);
6605 old_groups = conf->worker_groups;
6607 flush_workqueue(raid5_wq);
6609 err = alloc_thread_groups(conf, new,
6610 &group_cnt, &worker_cnt_per_group,
6613 spin_lock_irq(&conf->device_lock);
6614 conf->group_cnt = group_cnt;
6615 conf->worker_cnt_per_group = worker_cnt_per_group;
6616 conf->worker_groups = new_groups;
6617 spin_unlock_irq(&conf->device_lock);
6620 kfree(old_groups[0].workers);
6623 mddev_resume(mddev);
6625 mddev_unlock(mddev);
6630 static struct md_sysfs_entry
6631 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6632 raid5_show_group_thread_cnt,
6633 raid5_store_group_thread_cnt);
6635 static struct attribute *raid5_attrs[] = {
6636 &raid5_stripecache_size.attr,
6637 &raid5_stripecache_active.attr,
6638 &raid5_preread_bypass_threshold.attr,
6639 &raid5_group_thread_cnt.attr,
6640 &raid5_skip_copy.attr,
6641 &raid5_rmw_level.attr,
6642 &r5c_journal_mode.attr,
6645 static struct attribute_group raid5_attrs_group = {
6647 .attrs = raid5_attrs,
6650 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6652 int *worker_cnt_per_group,
6653 struct r5worker_group **worker_groups)
6657 struct r5worker *workers;
6659 *worker_cnt_per_group = cnt;
6662 *worker_groups = NULL;
6665 *group_cnt = num_possible_nodes();
6666 size = sizeof(struct r5worker) * cnt;
6667 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6668 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6669 *group_cnt, GFP_NOIO);
6670 if (!*worker_groups || !workers) {
6672 kfree(*worker_groups);
6676 for (i = 0; i < *group_cnt; i++) {
6677 struct r5worker_group *group;
6679 group = &(*worker_groups)[i];
6680 INIT_LIST_HEAD(&group->handle_list);
6681 INIT_LIST_HEAD(&group->loprio_list);
6683 group->workers = workers + i * cnt;
6685 for (j = 0; j < cnt; j++) {
6686 struct r5worker *worker = group->workers + j;
6687 worker->group = group;
6688 INIT_WORK(&worker->work, raid5_do_work);
6690 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6691 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6698 static void free_thread_groups(struct r5conf *conf)
6700 if (conf->worker_groups)
6701 kfree(conf->worker_groups[0].workers);
6702 kfree(conf->worker_groups);
6703 conf->worker_groups = NULL;
6707 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6709 struct r5conf *conf = mddev->private;
6712 sectors = mddev->dev_sectors;
6714 /* size is defined by the smallest of previous and new size */
6715 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6717 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6718 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6719 return sectors * (raid_disks - conf->max_degraded);
6722 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6724 safe_put_page(percpu->spare_page);
6725 if (percpu->scribble)
6726 flex_array_free(percpu->scribble);
6727 percpu->spare_page = NULL;
6728 percpu->scribble = NULL;
6731 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6733 if (conf->level == 6 && !percpu->spare_page)
6734 percpu->spare_page = alloc_page(GFP_KERNEL);
6735 if (!percpu->scribble)
6736 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6737 conf->previous_raid_disks),
6738 max(conf->chunk_sectors,
6739 conf->prev_chunk_sectors)
6743 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6744 free_scratch_buffer(conf, percpu);
6751 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6753 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6755 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6759 static void raid5_free_percpu(struct r5conf *conf)
6764 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6765 free_percpu(conf->percpu);
6768 static void free_conf(struct r5conf *conf)
6774 if (conf->shrinker.nr_deferred)
6775 unregister_shrinker(&conf->shrinker);
6777 free_thread_groups(conf);
6778 shrink_stripes(conf);
6779 raid5_free_percpu(conf);
6780 for (i = 0; i < conf->pool_size; i++)
6781 if (conf->disks[i].extra_page)
6782 put_page(conf->disks[i].extra_page);
6784 if (conf->bio_split)
6785 bioset_free(conf->bio_split);
6786 kfree(conf->stripe_hashtbl);
6787 kfree(conf->pending_data);
6791 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6793 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6794 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6796 if (alloc_scratch_buffer(conf, percpu)) {
6797 pr_warn("%s: failed memory allocation for cpu%u\n",
6804 static int raid5_alloc_percpu(struct r5conf *conf)
6808 conf->percpu = alloc_percpu(struct raid5_percpu);
6812 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6814 conf->scribble_disks = max(conf->raid_disks,
6815 conf->previous_raid_disks);
6816 conf->scribble_sectors = max(conf->chunk_sectors,
6817 conf->prev_chunk_sectors);
6822 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6823 struct shrink_control *sc)
6825 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6826 unsigned long ret = SHRINK_STOP;
6828 if (mutex_trylock(&conf->cache_size_mutex)) {
6830 while (ret < sc->nr_to_scan &&
6831 conf->max_nr_stripes > conf->min_nr_stripes) {
6832 if (drop_one_stripe(conf) == 0) {
6838 mutex_unlock(&conf->cache_size_mutex);
6843 static unsigned long raid5_cache_count(struct shrinker *shrink,
6844 struct shrink_control *sc)
6846 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6848 if (conf->max_nr_stripes < conf->min_nr_stripes)
6849 /* unlikely, but not impossible */
6851 return conf->max_nr_stripes - conf->min_nr_stripes;
6854 static struct r5conf *setup_conf(struct mddev *mddev)
6856 struct r5conf *conf;
6857 int raid_disk, memory, max_disks;
6858 struct md_rdev *rdev;
6859 struct disk_info *disk;
6862 int group_cnt, worker_cnt_per_group;
6863 struct r5worker_group *new_group;
6865 if (mddev->new_level != 5
6866 && mddev->new_level != 4
6867 && mddev->new_level != 6) {
6868 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6869 mdname(mddev), mddev->new_level);
6870 return ERR_PTR(-EIO);
6872 if ((mddev->new_level == 5
6873 && !algorithm_valid_raid5(mddev->new_layout)) ||
6874 (mddev->new_level == 6
6875 && !algorithm_valid_raid6(mddev->new_layout))) {
6876 pr_warn("md/raid:%s: layout %d not supported\n",
6877 mdname(mddev), mddev->new_layout);
6878 return ERR_PTR(-EIO);
6880 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6881 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6882 mdname(mddev), mddev->raid_disks);
6883 return ERR_PTR(-EINVAL);
6886 if (!mddev->new_chunk_sectors ||
6887 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6888 !is_power_of_2(mddev->new_chunk_sectors)) {
6889 pr_warn("md/raid:%s: invalid chunk size %d\n",
6890 mdname(mddev), mddev->new_chunk_sectors << 9);
6891 return ERR_PTR(-EINVAL);
6894 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6897 INIT_LIST_HEAD(&conf->free_list);
6898 INIT_LIST_HEAD(&conf->pending_list);
6899 conf->pending_data = kzalloc(sizeof(struct r5pending_data) *
6900 PENDING_IO_MAX, GFP_KERNEL);
6901 if (!conf->pending_data)
6903 for (i = 0; i < PENDING_IO_MAX; i++)
6904 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6905 /* Don't enable multi-threading by default*/
6906 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6908 conf->group_cnt = group_cnt;
6909 conf->worker_cnt_per_group = worker_cnt_per_group;
6910 conf->worker_groups = new_group;
6913 spin_lock_init(&conf->device_lock);
6914 seqcount_init(&conf->gen_lock);
6915 mutex_init(&conf->cache_size_mutex);
6916 init_waitqueue_head(&conf->wait_for_quiescent);
6917 init_waitqueue_head(&conf->wait_for_stripe);
6918 init_waitqueue_head(&conf->wait_for_overlap);
6919 INIT_LIST_HEAD(&conf->handle_list);
6920 INIT_LIST_HEAD(&conf->loprio_list);
6921 INIT_LIST_HEAD(&conf->hold_list);
6922 INIT_LIST_HEAD(&conf->delayed_list);
6923 INIT_LIST_HEAD(&conf->bitmap_list);
6924 init_llist_head(&conf->released_stripes);
6925 atomic_set(&conf->active_stripes, 0);
6926 atomic_set(&conf->preread_active_stripes, 0);
6927 atomic_set(&conf->active_aligned_reads, 0);
6928 spin_lock_init(&conf->pending_bios_lock);
6929 conf->batch_bio_dispatch = true;
6930 rdev_for_each(rdev, mddev) {
6931 if (test_bit(Journal, &rdev->flags))
6933 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6934 conf->batch_bio_dispatch = false;
6939 conf->bypass_threshold = BYPASS_THRESHOLD;
6940 conf->recovery_disabled = mddev->recovery_disabled - 1;
6942 conf->raid_disks = mddev->raid_disks;
6943 if (mddev->reshape_position == MaxSector)
6944 conf->previous_raid_disks = mddev->raid_disks;
6946 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6947 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6949 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6955 for (i = 0; i < max_disks; i++) {
6956 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6957 if (!conf->disks[i].extra_page)
6961 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
6962 if (!conf->bio_split)
6964 conf->mddev = mddev;
6966 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6969 /* We init hash_locks[0] separately to that it can be used
6970 * as the reference lock in the spin_lock_nest_lock() call
6971 * in lock_all_device_hash_locks_irq in order to convince
6972 * lockdep that we know what we are doing.
6974 spin_lock_init(conf->hash_locks);
6975 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6976 spin_lock_init(conf->hash_locks + i);
6978 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6979 INIT_LIST_HEAD(conf->inactive_list + i);
6981 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6982 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6984 atomic_set(&conf->r5c_cached_full_stripes, 0);
6985 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6986 atomic_set(&conf->r5c_cached_partial_stripes, 0);
6987 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6988 atomic_set(&conf->r5c_flushing_full_stripes, 0);
6989 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
6991 conf->level = mddev->new_level;
6992 conf->chunk_sectors = mddev->new_chunk_sectors;
6993 if (raid5_alloc_percpu(conf) != 0)
6996 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6998 rdev_for_each(rdev, mddev) {
6999 raid_disk = rdev->raid_disk;
7000 if (raid_disk >= max_disks
7001 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7003 disk = conf->disks + raid_disk;
7005 if (test_bit(Replacement, &rdev->flags)) {
7006 if (disk->replacement)
7008 disk->replacement = rdev;
7015 if (test_bit(In_sync, &rdev->flags)) {
7016 char b[BDEVNAME_SIZE];
7017 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7018 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7019 } else if (rdev->saved_raid_disk != raid_disk)
7020 /* Cannot rely on bitmap to complete recovery */
7024 conf->level = mddev->new_level;
7025 if (conf->level == 6) {
7026 conf->max_degraded = 2;
7027 if (raid6_call.xor_syndrome)
7028 conf->rmw_level = PARITY_ENABLE_RMW;
7030 conf->rmw_level = PARITY_DISABLE_RMW;
7032 conf->max_degraded = 1;
7033 conf->rmw_level = PARITY_ENABLE_RMW;
7035 conf->algorithm = mddev->new_layout;
7036 conf->reshape_progress = mddev->reshape_position;
7037 if (conf->reshape_progress != MaxSector) {
7038 conf->prev_chunk_sectors = mddev->chunk_sectors;
7039 conf->prev_algo = mddev->layout;
7041 conf->prev_chunk_sectors = conf->chunk_sectors;
7042 conf->prev_algo = conf->algorithm;
7045 conf->min_nr_stripes = NR_STRIPES;
7046 if (mddev->reshape_position != MaxSector) {
7047 int stripes = max_t(int,
7048 ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7049 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7050 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7051 if (conf->min_nr_stripes != NR_STRIPES)
7052 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7053 mdname(mddev), conf->min_nr_stripes);
7055 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7056 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7057 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7058 if (grow_stripes(conf, conf->min_nr_stripes)) {
7059 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7060 mdname(mddev), memory);
7063 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7065 * Losing a stripe head costs more than the time to refill it,
7066 * it reduces the queue depth and so can hurt throughput.
7067 * So set it rather large, scaled by number of devices.
7069 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7070 conf->shrinker.scan_objects = raid5_cache_scan;
7071 conf->shrinker.count_objects = raid5_cache_count;
7072 conf->shrinker.batch = 128;
7073 conf->shrinker.flags = 0;
7074 if (register_shrinker(&conf->shrinker)) {
7075 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7080 sprintf(pers_name, "raid%d", mddev->new_level);
7081 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7082 if (!conf->thread) {
7083 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7093 return ERR_PTR(-EIO);
7095 return ERR_PTR(-ENOMEM);
7098 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7101 case ALGORITHM_PARITY_0:
7102 if (raid_disk < max_degraded)
7105 case ALGORITHM_PARITY_N:
7106 if (raid_disk >= raid_disks - max_degraded)
7109 case ALGORITHM_PARITY_0_6:
7110 if (raid_disk == 0 ||
7111 raid_disk == raid_disks - 1)
7114 case ALGORITHM_LEFT_ASYMMETRIC_6:
7115 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7116 case ALGORITHM_LEFT_SYMMETRIC_6:
7117 case ALGORITHM_RIGHT_SYMMETRIC_6:
7118 if (raid_disk == raid_disks - 1)
7124 static int raid5_run(struct mddev *mddev)
7126 struct r5conf *conf;
7127 int working_disks = 0;
7128 int dirty_parity_disks = 0;
7129 struct md_rdev *rdev;
7130 struct md_rdev *journal_dev = NULL;
7131 sector_t reshape_offset = 0;
7133 long long min_offset_diff = 0;
7136 if (mddev_init_writes_pending(mddev) < 0)
7139 if (mddev->recovery_cp != MaxSector)
7140 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7143 rdev_for_each(rdev, mddev) {
7146 if (test_bit(Journal, &rdev->flags)) {
7150 if (rdev->raid_disk < 0)
7152 diff = (rdev->new_data_offset - rdev->data_offset);
7154 min_offset_diff = diff;
7156 } else if (mddev->reshape_backwards &&
7157 diff < min_offset_diff)
7158 min_offset_diff = diff;
7159 else if (!mddev->reshape_backwards &&
7160 diff > min_offset_diff)
7161 min_offset_diff = diff;
7164 if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7165 (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7166 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7171 if (mddev->reshape_position != MaxSector) {
7172 /* Check that we can continue the reshape.
7173 * Difficulties arise if the stripe we would write to
7174 * next is at or after the stripe we would read from next.
7175 * For a reshape that changes the number of devices, this
7176 * is only possible for a very short time, and mdadm makes
7177 * sure that time appears to have past before assembling
7178 * the array. So we fail if that time hasn't passed.
7179 * For a reshape that keeps the number of devices the same
7180 * mdadm must be monitoring the reshape can keeping the
7181 * critical areas read-only and backed up. It will start
7182 * the array in read-only mode, so we check for that.
7184 sector_t here_new, here_old;
7186 int max_degraded = (mddev->level == 6 ? 2 : 1);
7191 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7196 if (mddev->new_level != mddev->level) {
7197 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7201 old_disks = mddev->raid_disks - mddev->delta_disks;
7202 /* reshape_position must be on a new-stripe boundary, and one
7203 * further up in new geometry must map after here in old
7205 * If the chunk sizes are different, then as we perform reshape
7206 * in units of the largest of the two, reshape_position needs
7207 * be a multiple of the largest chunk size times new data disks.
7209 here_new = mddev->reshape_position;
7210 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7211 new_data_disks = mddev->raid_disks - max_degraded;
7212 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7213 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7217 reshape_offset = here_new * chunk_sectors;
7218 /* here_new is the stripe we will write to */
7219 here_old = mddev->reshape_position;
7220 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7221 /* here_old is the first stripe that we might need to read
7223 if (mddev->delta_disks == 0) {
7224 /* We cannot be sure it is safe to start an in-place
7225 * reshape. It is only safe if user-space is monitoring
7226 * and taking constant backups.
7227 * mdadm always starts a situation like this in
7228 * readonly mode so it can take control before
7229 * allowing any writes. So just check for that.
7231 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7232 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7233 /* not really in-place - so OK */;
7234 else if (mddev->ro == 0) {
7235 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7239 } else if (mddev->reshape_backwards
7240 ? (here_new * chunk_sectors + min_offset_diff <=
7241 here_old * chunk_sectors)
7242 : (here_new * chunk_sectors >=
7243 here_old * chunk_sectors + (-min_offset_diff))) {
7244 /* Reading from the same stripe as writing to - bad */
7245 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7249 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7250 /* OK, we should be able to continue; */
7252 BUG_ON(mddev->level != mddev->new_level);
7253 BUG_ON(mddev->layout != mddev->new_layout);
7254 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7255 BUG_ON(mddev->delta_disks != 0);
7258 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7259 test_bit(MD_HAS_PPL, &mddev->flags)) {
7260 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7262 clear_bit(MD_HAS_PPL, &mddev->flags);
7263 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7266 if (mddev->private == NULL)
7267 conf = setup_conf(mddev);
7269 conf = mddev->private;
7272 return PTR_ERR(conf);
7274 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7276 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7279 set_disk_ro(mddev->gendisk, 1);
7280 } else if (mddev->recovery_cp == MaxSector)
7281 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7284 conf->min_offset_diff = min_offset_diff;
7285 mddev->thread = conf->thread;
7286 conf->thread = NULL;
7287 mddev->private = conf;
7289 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7291 rdev = conf->disks[i].rdev;
7292 if (!rdev && conf->disks[i].replacement) {
7293 /* The replacement is all we have yet */
7294 rdev = conf->disks[i].replacement;
7295 conf->disks[i].replacement = NULL;
7296 clear_bit(Replacement, &rdev->flags);
7297 conf->disks[i].rdev = rdev;
7301 if (conf->disks[i].replacement &&
7302 conf->reshape_progress != MaxSector) {
7303 /* replacements and reshape simply do not mix. */
7304 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7307 if (test_bit(In_sync, &rdev->flags)) {
7311 /* This disc is not fully in-sync. However if it
7312 * just stored parity (beyond the recovery_offset),
7313 * when we don't need to be concerned about the
7314 * array being dirty.
7315 * When reshape goes 'backwards', we never have
7316 * partially completed devices, so we only need
7317 * to worry about reshape going forwards.
7319 /* Hack because v0.91 doesn't store recovery_offset properly. */
7320 if (mddev->major_version == 0 &&
7321 mddev->minor_version > 90)
7322 rdev->recovery_offset = reshape_offset;
7324 if (rdev->recovery_offset < reshape_offset) {
7325 /* We need to check old and new layout */
7326 if (!only_parity(rdev->raid_disk,
7329 conf->max_degraded))
7332 if (!only_parity(rdev->raid_disk,
7334 conf->previous_raid_disks,
7335 conf->max_degraded))
7337 dirty_parity_disks++;
7341 * 0 for a fully functional array, 1 or 2 for a degraded array.
7343 mddev->degraded = raid5_calc_degraded(conf);
7345 if (has_failed(conf)) {
7346 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7347 mdname(mddev), mddev->degraded, conf->raid_disks);
7351 /* device size must be a multiple of chunk size */
7352 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7353 mddev->resync_max_sectors = mddev->dev_sectors;
7355 if (mddev->degraded > dirty_parity_disks &&
7356 mddev->recovery_cp != MaxSector) {
7357 if (test_bit(MD_HAS_PPL, &mddev->flags))
7358 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7360 else if (mddev->ok_start_degraded)
7361 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7364 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7370 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7371 mdname(mddev), conf->level,
7372 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7375 print_raid5_conf(conf);
7377 if (conf->reshape_progress != MaxSector) {
7378 conf->reshape_safe = conf->reshape_progress;
7379 atomic_set(&conf->reshape_stripes, 0);
7380 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7381 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7382 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7383 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7384 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7386 if (!mddev->sync_thread)
7390 /* Ok, everything is just fine now */
7391 if (mddev->to_remove == &raid5_attrs_group)
7392 mddev->to_remove = NULL;
7393 else if (mddev->kobj.sd &&
7394 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7395 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7397 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7401 /* read-ahead size must cover two whole stripes, which
7402 * is 2 * (datadisks) * chunksize where 'n' is the
7403 * number of raid devices
7405 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7406 int stripe = data_disks *
7407 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7408 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7409 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7411 chunk_size = mddev->chunk_sectors << 9;
7412 blk_queue_io_min(mddev->queue, chunk_size);
7413 blk_queue_io_opt(mddev->queue, chunk_size *
7414 (conf->raid_disks - conf->max_degraded));
7415 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7417 * We can only discard a whole stripe. It doesn't make sense to
7418 * discard data disk but write parity disk
7420 stripe = stripe * PAGE_SIZE;
7421 /* Round up to power of 2, as discard handling
7422 * currently assumes that */
7423 while ((stripe-1) & stripe)
7424 stripe = (stripe | (stripe-1)) + 1;
7425 mddev->queue->limits.discard_alignment = stripe;
7426 mddev->queue->limits.discard_granularity = stripe;
7428 blk_queue_max_write_same_sectors(mddev->queue, 0);
7429 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7431 rdev_for_each(rdev, mddev) {
7432 disk_stack_limits(mddev->gendisk, rdev->bdev,
7433 rdev->data_offset << 9);
7434 disk_stack_limits(mddev->gendisk, rdev->bdev,
7435 rdev->new_data_offset << 9);
7439 * zeroing is required, otherwise data
7440 * could be lost. Consider a scenario: discard a stripe
7441 * (the stripe could be inconsistent if
7442 * discard_zeroes_data is 0); write one disk of the
7443 * stripe (the stripe could be inconsistent again
7444 * depending on which disks are used to calculate
7445 * parity); the disk is broken; The stripe data of this
7448 * We only allow DISCARD if the sysadmin has confirmed that
7449 * only safe devices are in use by setting a module parameter.
7450 * A better idea might be to turn DISCARD into WRITE_ZEROES
7451 * requests, as that is required to be safe.
7453 if (devices_handle_discard_safely &&
7454 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7455 mddev->queue->limits.discard_granularity >= stripe)
7456 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7459 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7462 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7465 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7470 md_unregister_thread(&mddev->thread);
7471 print_raid5_conf(conf);
7473 mddev->private = NULL;
7474 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7478 static void raid5_free(struct mddev *mddev, void *priv)
7480 struct r5conf *conf = priv;
7483 mddev->to_remove = &raid5_attrs_group;
7486 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7488 struct r5conf *conf = mddev->private;
7491 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7492 conf->chunk_sectors / 2, mddev->layout);
7493 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7495 for (i = 0; i < conf->raid_disks; i++) {
7496 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7497 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7500 seq_printf (seq, "]");
7503 static void print_raid5_conf (struct r5conf *conf)
7506 struct disk_info *tmp;
7508 pr_debug("RAID conf printout:\n");
7510 pr_debug("(conf==NULL)\n");
7513 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7515 conf->raid_disks - conf->mddev->degraded);
7517 for (i = 0; i < conf->raid_disks; i++) {
7518 char b[BDEVNAME_SIZE];
7519 tmp = conf->disks + i;
7521 pr_debug(" disk %d, o:%d, dev:%s\n",
7522 i, !test_bit(Faulty, &tmp->rdev->flags),
7523 bdevname(tmp->rdev->bdev, b));
7527 static int raid5_spare_active(struct mddev *mddev)
7530 struct r5conf *conf = mddev->private;
7531 struct disk_info *tmp;
7533 unsigned long flags;
7535 for (i = 0; i < conf->raid_disks; i++) {
7536 tmp = conf->disks + i;
7537 if (tmp->replacement
7538 && tmp->replacement->recovery_offset == MaxSector
7539 && !test_bit(Faulty, &tmp->replacement->flags)
7540 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7541 /* Replacement has just become active. */
7543 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7546 /* Replaced device not technically faulty,
7547 * but we need to be sure it gets removed
7548 * and never re-added.
7550 set_bit(Faulty, &tmp->rdev->flags);
7551 sysfs_notify_dirent_safe(
7552 tmp->rdev->sysfs_state);
7554 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7555 } else if (tmp->rdev
7556 && tmp->rdev->recovery_offset == MaxSector
7557 && !test_bit(Faulty, &tmp->rdev->flags)
7558 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7560 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7563 spin_lock_irqsave(&conf->device_lock, flags);
7564 mddev->degraded = raid5_calc_degraded(conf);
7565 spin_unlock_irqrestore(&conf->device_lock, flags);
7566 print_raid5_conf(conf);
7570 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7572 struct r5conf *conf = mddev->private;
7574 int number = rdev->raid_disk;
7575 struct md_rdev **rdevp;
7576 struct disk_info *p = conf->disks + number;
7578 print_raid5_conf(conf);
7579 if (test_bit(Journal, &rdev->flags) && conf->log) {
7581 * we can't wait pending write here, as this is called in
7582 * raid5d, wait will deadlock.
7583 * neilb: there is no locking about new writes here,
7584 * so this cannot be safe.
7586 if (atomic_read(&conf->active_stripes) ||
7587 atomic_read(&conf->r5c_cached_full_stripes) ||
7588 atomic_read(&conf->r5c_cached_partial_stripes)) {
7594 if (rdev == p->rdev)
7596 else if (rdev == p->replacement)
7597 rdevp = &p->replacement;
7601 if (number >= conf->raid_disks &&
7602 conf->reshape_progress == MaxSector)
7603 clear_bit(In_sync, &rdev->flags);
7605 if (test_bit(In_sync, &rdev->flags) ||
7606 atomic_read(&rdev->nr_pending)) {
7610 /* Only remove non-faulty devices if recovery
7613 if (!test_bit(Faulty, &rdev->flags) &&
7614 mddev->recovery_disabled != conf->recovery_disabled &&
7615 !has_failed(conf) &&
7616 (!p->replacement || p->replacement == rdev) &&
7617 number < conf->raid_disks) {
7622 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7624 if (atomic_read(&rdev->nr_pending)) {
7625 /* lost the race, try later */
7631 err = log_modify(conf, rdev, false);
7635 if (p->replacement) {
7636 /* We must have just cleared 'rdev' */
7637 p->rdev = p->replacement;
7638 clear_bit(Replacement, &p->replacement->flags);
7639 smp_mb(); /* Make sure other CPUs may see both as identical
7640 * but will never see neither - if they are careful
7642 p->replacement = NULL;
7645 err = log_modify(conf, p->rdev, true);
7648 clear_bit(WantReplacement, &rdev->flags);
7651 print_raid5_conf(conf);
7655 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7657 struct r5conf *conf = mddev->private;
7660 struct disk_info *p;
7662 int last = conf->raid_disks - 1;
7664 if (test_bit(Journal, &rdev->flags)) {
7668 rdev->raid_disk = 0;
7670 * The array is in readonly mode if journal is missing, so no
7671 * write requests running. We should be safe
7673 log_init(conf, rdev, false);
7676 if (mddev->recovery_disabled == conf->recovery_disabled)
7679 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7680 /* no point adding a device */
7683 if (rdev->raid_disk >= 0)
7684 first = last = rdev->raid_disk;
7687 * find the disk ... but prefer rdev->saved_raid_disk
7690 if (rdev->saved_raid_disk >= 0 &&
7691 rdev->saved_raid_disk >= first &&
7692 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7693 first = rdev->saved_raid_disk;
7695 for (disk = first; disk <= last; disk++) {
7696 p = conf->disks + disk;
7697 if (p->rdev == NULL) {
7698 clear_bit(In_sync, &rdev->flags);
7699 rdev->raid_disk = disk;
7700 if (rdev->saved_raid_disk != disk)
7702 rcu_assign_pointer(p->rdev, rdev);
7704 err = log_modify(conf, rdev, true);
7709 for (disk = first; disk <= last; disk++) {
7710 p = conf->disks + disk;
7711 if (test_bit(WantReplacement, &p->rdev->flags) &&
7712 p->replacement == NULL) {
7713 clear_bit(In_sync, &rdev->flags);
7714 set_bit(Replacement, &rdev->flags);
7715 rdev->raid_disk = disk;
7718 rcu_assign_pointer(p->replacement, rdev);
7723 print_raid5_conf(conf);
7727 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7729 /* no resync is happening, and there is enough space
7730 * on all devices, so we can resize.
7731 * We need to make sure resync covers any new space.
7732 * If the array is shrinking we should possibly wait until
7733 * any io in the removed space completes, but it hardly seems
7737 struct r5conf *conf = mddev->private;
7739 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7741 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7742 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7743 if (mddev->external_size &&
7744 mddev->array_sectors > newsize)
7746 if (mddev->bitmap) {
7747 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7751 md_set_array_sectors(mddev, newsize);
7752 if (sectors > mddev->dev_sectors &&
7753 mddev->recovery_cp > mddev->dev_sectors) {
7754 mddev->recovery_cp = mddev->dev_sectors;
7755 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7757 mddev->dev_sectors = sectors;
7758 mddev->resync_max_sectors = sectors;
7762 static int check_stripe_cache(struct mddev *mddev)
7764 /* Can only proceed if there are plenty of stripe_heads.
7765 * We need a minimum of one full stripe,, and for sensible progress
7766 * it is best to have about 4 times that.
7767 * If we require 4 times, then the default 256 4K stripe_heads will
7768 * allow for chunk sizes up to 256K, which is probably OK.
7769 * If the chunk size is greater, user-space should request more
7770 * stripe_heads first.
7772 struct r5conf *conf = mddev->private;
7773 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7774 > conf->min_nr_stripes ||
7775 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7776 > conf->min_nr_stripes) {
7777 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7779 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7786 static int check_reshape(struct mddev *mddev)
7788 struct r5conf *conf = mddev->private;
7790 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7792 if (mddev->delta_disks == 0 &&
7793 mddev->new_layout == mddev->layout &&
7794 mddev->new_chunk_sectors == mddev->chunk_sectors)
7795 return 0; /* nothing to do */
7796 if (has_failed(conf))
7798 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7799 /* We might be able to shrink, but the devices must
7800 * be made bigger first.
7801 * For raid6, 4 is the minimum size.
7802 * Otherwise 2 is the minimum
7805 if (mddev->level == 6)
7807 if (mddev->raid_disks + mddev->delta_disks < min)
7811 if (!check_stripe_cache(mddev))
7814 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7815 mddev->delta_disks > 0)
7816 if (resize_chunks(conf,
7817 conf->previous_raid_disks
7818 + max(0, mddev->delta_disks),
7819 max(mddev->new_chunk_sectors,
7820 mddev->chunk_sectors)
7824 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7825 return 0; /* never bother to shrink */
7826 return resize_stripes(conf, (conf->previous_raid_disks
7827 + mddev->delta_disks));
7830 static int raid5_start_reshape(struct mddev *mddev)
7832 struct r5conf *conf = mddev->private;
7833 struct md_rdev *rdev;
7835 unsigned long flags;
7837 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7840 if (!check_stripe_cache(mddev))
7843 if (has_failed(conf))
7846 rdev_for_each(rdev, mddev) {
7847 if (!test_bit(In_sync, &rdev->flags)
7848 && !test_bit(Faulty, &rdev->flags))
7852 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7853 /* Not enough devices even to make a degraded array
7858 /* Refuse to reduce size of the array. Any reductions in
7859 * array size must be through explicit setting of array_size
7862 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7863 < mddev->array_sectors) {
7864 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7869 atomic_set(&conf->reshape_stripes, 0);
7870 spin_lock_irq(&conf->device_lock);
7871 write_seqcount_begin(&conf->gen_lock);
7872 conf->previous_raid_disks = conf->raid_disks;
7873 conf->raid_disks += mddev->delta_disks;
7874 conf->prev_chunk_sectors = conf->chunk_sectors;
7875 conf->chunk_sectors = mddev->new_chunk_sectors;
7876 conf->prev_algo = conf->algorithm;
7877 conf->algorithm = mddev->new_layout;
7879 /* Code that selects data_offset needs to see the generation update
7880 * if reshape_progress has been set - so a memory barrier needed.
7883 if (mddev->reshape_backwards)
7884 conf->reshape_progress = raid5_size(mddev, 0, 0);
7886 conf->reshape_progress = 0;
7887 conf->reshape_safe = conf->reshape_progress;
7888 write_seqcount_end(&conf->gen_lock);
7889 spin_unlock_irq(&conf->device_lock);
7891 /* Now make sure any requests that proceeded on the assumption
7892 * the reshape wasn't running - like Discard or Read - have
7895 mddev_suspend(mddev);
7896 mddev_resume(mddev);
7898 /* Add some new drives, as many as will fit.
7899 * We know there are enough to make the newly sized array work.
7900 * Don't add devices if we are reducing the number of
7901 * devices in the array. This is because it is not possible
7902 * to correctly record the "partially reconstructed" state of
7903 * such devices during the reshape and confusion could result.
7905 if (mddev->delta_disks >= 0) {
7906 rdev_for_each(rdev, mddev)
7907 if (rdev->raid_disk < 0 &&
7908 !test_bit(Faulty, &rdev->flags)) {
7909 if (raid5_add_disk(mddev, rdev) == 0) {
7911 >= conf->previous_raid_disks)
7912 set_bit(In_sync, &rdev->flags);
7914 rdev->recovery_offset = 0;
7916 if (sysfs_link_rdev(mddev, rdev))
7917 /* Failure here is OK */;
7919 } else if (rdev->raid_disk >= conf->previous_raid_disks
7920 && !test_bit(Faulty, &rdev->flags)) {
7921 /* This is a spare that was manually added */
7922 set_bit(In_sync, &rdev->flags);
7925 /* When a reshape changes the number of devices,
7926 * ->degraded is measured against the larger of the
7927 * pre and post number of devices.
7929 spin_lock_irqsave(&conf->device_lock, flags);
7930 mddev->degraded = raid5_calc_degraded(conf);
7931 spin_unlock_irqrestore(&conf->device_lock, flags);
7933 mddev->raid_disks = conf->raid_disks;
7934 mddev->reshape_position = conf->reshape_progress;
7935 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7937 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7938 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7939 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7940 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7941 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7942 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7944 if (!mddev->sync_thread) {
7945 mddev->recovery = 0;
7946 spin_lock_irq(&conf->device_lock);
7947 write_seqcount_begin(&conf->gen_lock);
7948 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7949 mddev->new_chunk_sectors =
7950 conf->chunk_sectors = conf->prev_chunk_sectors;
7951 mddev->new_layout = conf->algorithm = conf->prev_algo;
7952 rdev_for_each(rdev, mddev)
7953 rdev->new_data_offset = rdev->data_offset;
7955 conf->generation --;
7956 conf->reshape_progress = MaxSector;
7957 mddev->reshape_position = MaxSector;
7958 write_seqcount_end(&conf->gen_lock);
7959 spin_unlock_irq(&conf->device_lock);
7962 conf->reshape_checkpoint = jiffies;
7963 md_wakeup_thread(mddev->sync_thread);
7964 md_new_event(mddev);
7968 /* This is called from the reshape thread and should make any
7969 * changes needed in 'conf'
7971 static void end_reshape(struct r5conf *conf)
7974 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7976 spin_lock_irq(&conf->device_lock);
7977 conf->previous_raid_disks = conf->raid_disks;
7978 md_finish_reshape(conf->mddev);
7980 conf->reshape_progress = MaxSector;
7981 conf->mddev->reshape_position = MaxSector;
7982 spin_unlock_irq(&conf->device_lock);
7983 wake_up(&conf->wait_for_overlap);
7985 /* read-ahead size must cover two whole stripes, which is
7986 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7988 if (conf->mddev->queue) {
7989 int data_disks = conf->raid_disks - conf->max_degraded;
7990 int stripe = data_disks * ((conf->chunk_sectors << 9)
7992 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7993 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7998 /* This is called from the raid5d thread with mddev_lock held.
7999 * It makes config changes to the device.
8001 static void raid5_finish_reshape(struct mddev *mddev)
8003 struct r5conf *conf = mddev->private;
8005 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8007 if (mddev->delta_disks <= 0) {
8009 spin_lock_irq(&conf->device_lock);
8010 mddev->degraded = raid5_calc_degraded(conf);
8011 spin_unlock_irq(&conf->device_lock);
8012 for (d = conf->raid_disks ;
8013 d < conf->raid_disks - mddev->delta_disks;
8015 struct md_rdev *rdev = conf->disks[d].rdev;
8017 clear_bit(In_sync, &rdev->flags);
8018 rdev = conf->disks[d].replacement;
8020 clear_bit(In_sync, &rdev->flags);
8023 mddev->layout = conf->algorithm;
8024 mddev->chunk_sectors = conf->chunk_sectors;
8025 mddev->reshape_position = MaxSector;
8026 mddev->delta_disks = 0;
8027 mddev->reshape_backwards = 0;
8031 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8033 struct r5conf *conf = mddev->private;
8036 /* stop all writes */
8037 lock_all_device_hash_locks_irq(conf);
8038 /* '2' tells resync/reshape to pause so that all
8039 * active stripes can drain
8041 r5c_flush_cache(conf, INT_MAX);
8043 wait_event_cmd(conf->wait_for_quiescent,
8044 atomic_read(&conf->active_stripes) == 0 &&
8045 atomic_read(&conf->active_aligned_reads) == 0,
8046 unlock_all_device_hash_locks_irq(conf),
8047 lock_all_device_hash_locks_irq(conf));
8049 unlock_all_device_hash_locks_irq(conf);
8050 /* allow reshape to continue */
8051 wake_up(&conf->wait_for_overlap);
8053 /* re-enable writes */
8054 lock_all_device_hash_locks_irq(conf);
8056 wake_up(&conf->wait_for_quiescent);
8057 wake_up(&conf->wait_for_overlap);
8058 unlock_all_device_hash_locks_irq(conf);
8060 r5l_quiesce(conf->log, quiesce);
8063 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8065 struct r0conf *raid0_conf = mddev->private;
8068 /* for raid0 takeover only one zone is supported */
8069 if (raid0_conf->nr_strip_zones > 1) {
8070 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8072 return ERR_PTR(-EINVAL);
8075 sectors = raid0_conf->strip_zone[0].zone_end;
8076 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8077 mddev->dev_sectors = sectors;
8078 mddev->new_level = level;
8079 mddev->new_layout = ALGORITHM_PARITY_N;
8080 mddev->new_chunk_sectors = mddev->chunk_sectors;
8081 mddev->raid_disks += 1;
8082 mddev->delta_disks = 1;
8083 /* make sure it will be not marked as dirty */
8084 mddev->recovery_cp = MaxSector;
8086 return setup_conf(mddev);
8089 static void *raid5_takeover_raid1(struct mddev *mddev)
8094 if (mddev->raid_disks != 2 ||
8095 mddev->degraded > 1)
8096 return ERR_PTR(-EINVAL);
8098 /* Should check if there are write-behind devices? */
8100 chunksect = 64*2; /* 64K by default */
8102 /* The array must be an exact multiple of chunksize */
8103 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8106 if ((chunksect<<9) < STRIPE_SIZE)
8107 /* array size does not allow a suitable chunk size */
8108 return ERR_PTR(-EINVAL);
8110 mddev->new_level = 5;
8111 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8112 mddev->new_chunk_sectors = chunksect;
8114 ret = setup_conf(mddev);
8116 mddev_clear_unsupported_flags(mddev,
8117 UNSUPPORTED_MDDEV_FLAGS);
8121 static void *raid5_takeover_raid6(struct mddev *mddev)
8125 switch (mddev->layout) {
8126 case ALGORITHM_LEFT_ASYMMETRIC_6:
8127 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8129 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8130 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8132 case ALGORITHM_LEFT_SYMMETRIC_6:
8133 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8135 case ALGORITHM_RIGHT_SYMMETRIC_6:
8136 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8138 case ALGORITHM_PARITY_0_6:
8139 new_layout = ALGORITHM_PARITY_0;
8141 case ALGORITHM_PARITY_N:
8142 new_layout = ALGORITHM_PARITY_N;
8145 return ERR_PTR(-EINVAL);
8147 mddev->new_level = 5;
8148 mddev->new_layout = new_layout;
8149 mddev->delta_disks = -1;
8150 mddev->raid_disks -= 1;
8151 return setup_conf(mddev);
8154 static int raid5_check_reshape(struct mddev *mddev)
8156 /* For a 2-drive array, the layout and chunk size can be changed
8157 * immediately as not restriping is needed.
8158 * For larger arrays we record the new value - after validation
8159 * to be used by a reshape pass.
8161 struct r5conf *conf = mddev->private;
8162 int new_chunk = mddev->new_chunk_sectors;
8164 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8166 if (new_chunk > 0) {
8167 if (!is_power_of_2(new_chunk))
8169 if (new_chunk < (PAGE_SIZE>>9))
8171 if (mddev->array_sectors & (new_chunk-1))
8172 /* not factor of array size */
8176 /* They look valid */
8178 if (mddev->raid_disks == 2) {
8179 /* can make the change immediately */
8180 if (mddev->new_layout >= 0) {
8181 conf->algorithm = mddev->new_layout;
8182 mddev->layout = mddev->new_layout;
8184 if (new_chunk > 0) {
8185 conf->chunk_sectors = new_chunk ;
8186 mddev->chunk_sectors = new_chunk;
8188 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8189 md_wakeup_thread(mddev->thread);
8191 return check_reshape(mddev);
8194 static int raid6_check_reshape(struct mddev *mddev)
8196 int new_chunk = mddev->new_chunk_sectors;
8198 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8200 if (new_chunk > 0) {
8201 if (!is_power_of_2(new_chunk))
8203 if (new_chunk < (PAGE_SIZE >> 9))
8205 if (mddev->array_sectors & (new_chunk-1))
8206 /* not factor of array size */
8210 /* They look valid */
8211 return check_reshape(mddev);
8214 static void *raid5_takeover(struct mddev *mddev)
8216 /* raid5 can take over:
8217 * raid0 - if there is only one strip zone - make it a raid4 layout
8218 * raid1 - if there are two drives. We need to know the chunk size
8219 * raid4 - trivial - just use a raid4 layout.
8220 * raid6 - Providing it is a *_6 layout
8222 if (mddev->level == 0)
8223 return raid45_takeover_raid0(mddev, 5);
8224 if (mddev->level == 1)
8225 return raid5_takeover_raid1(mddev);
8226 if (mddev->level == 4) {
8227 mddev->new_layout = ALGORITHM_PARITY_N;
8228 mddev->new_level = 5;
8229 return setup_conf(mddev);
8231 if (mddev->level == 6)
8232 return raid5_takeover_raid6(mddev);
8234 return ERR_PTR(-EINVAL);
8237 static void *raid4_takeover(struct mddev *mddev)
8239 /* raid4 can take over:
8240 * raid0 - if there is only one strip zone
8241 * raid5 - if layout is right
8243 if (mddev->level == 0)
8244 return raid45_takeover_raid0(mddev, 4);
8245 if (mddev->level == 5 &&
8246 mddev->layout == ALGORITHM_PARITY_N) {
8247 mddev->new_layout = 0;
8248 mddev->new_level = 4;
8249 return setup_conf(mddev);
8251 return ERR_PTR(-EINVAL);
8254 static struct md_personality raid5_personality;
8256 static void *raid6_takeover(struct mddev *mddev)
8258 /* Currently can only take over a raid5. We map the
8259 * personality to an equivalent raid6 personality
8260 * with the Q block at the end.
8264 if (mddev->pers != &raid5_personality)
8265 return ERR_PTR(-EINVAL);
8266 if (mddev->degraded > 1)
8267 return ERR_PTR(-EINVAL);
8268 if (mddev->raid_disks > 253)
8269 return ERR_PTR(-EINVAL);
8270 if (mddev->raid_disks < 3)
8271 return ERR_PTR(-EINVAL);
8273 switch (mddev->layout) {
8274 case ALGORITHM_LEFT_ASYMMETRIC:
8275 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8277 case ALGORITHM_RIGHT_ASYMMETRIC:
8278 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8280 case ALGORITHM_LEFT_SYMMETRIC:
8281 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8283 case ALGORITHM_RIGHT_SYMMETRIC:
8284 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8286 case ALGORITHM_PARITY_0:
8287 new_layout = ALGORITHM_PARITY_0_6;
8289 case ALGORITHM_PARITY_N:
8290 new_layout = ALGORITHM_PARITY_N;
8293 return ERR_PTR(-EINVAL);
8295 mddev->new_level = 6;
8296 mddev->new_layout = new_layout;
8297 mddev->delta_disks = 1;
8298 mddev->raid_disks += 1;
8299 return setup_conf(mddev);
8302 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8304 struct r5conf *conf;
8307 err = mddev_lock(mddev);
8310 conf = mddev->private;
8312 mddev_unlock(mddev);
8316 if (strncmp(buf, "ppl", 3) == 0) {
8317 /* ppl only works with RAID 5 */
8318 if (!raid5_has_ppl(conf) && conf->level == 5) {
8319 err = log_init(conf, NULL, true);
8321 err = resize_stripes(conf, conf->pool_size);
8327 } else if (strncmp(buf, "resync", 6) == 0) {
8328 if (raid5_has_ppl(conf)) {
8329 mddev_suspend(mddev);
8331 mddev_resume(mddev);
8332 err = resize_stripes(conf, conf->pool_size);
8333 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8334 r5l_log_disk_error(conf)) {
8335 bool journal_dev_exists = false;
8336 struct md_rdev *rdev;
8338 rdev_for_each(rdev, mddev)
8339 if (test_bit(Journal, &rdev->flags)) {
8340 journal_dev_exists = true;
8344 if (!journal_dev_exists) {
8345 mddev_suspend(mddev);
8346 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8347 mddev_resume(mddev);
8348 } else /* need remove journal device first */
8357 md_update_sb(mddev, 1);
8359 mddev_unlock(mddev);
8364 static struct md_personality raid6_personality =
8368 .owner = THIS_MODULE,
8369 .make_request = raid5_make_request,
8372 .status = raid5_status,
8373 .error_handler = raid5_error,
8374 .hot_add_disk = raid5_add_disk,
8375 .hot_remove_disk= raid5_remove_disk,
8376 .spare_active = raid5_spare_active,
8377 .sync_request = raid5_sync_request,
8378 .resize = raid5_resize,
8380 .check_reshape = raid6_check_reshape,
8381 .start_reshape = raid5_start_reshape,
8382 .finish_reshape = raid5_finish_reshape,
8383 .quiesce = raid5_quiesce,
8384 .takeover = raid6_takeover,
8385 .congested = raid5_congested,
8386 .change_consistency_policy = raid5_change_consistency_policy,
8388 static struct md_personality raid5_personality =
8392 .owner = THIS_MODULE,
8393 .make_request = raid5_make_request,
8396 .status = raid5_status,
8397 .error_handler = raid5_error,
8398 .hot_add_disk = raid5_add_disk,
8399 .hot_remove_disk= raid5_remove_disk,
8400 .spare_active = raid5_spare_active,
8401 .sync_request = raid5_sync_request,
8402 .resize = raid5_resize,
8404 .check_reshape = raid5_check_reshape,
8405 .start_reshape = raid5_start_reshape,
8406 .finish_reshape = raid5_finish_reshape,
8407 .quiesce = raid5_quiesce,
8408 .takeover = raid5_takeover,
8409 .congested = raid5_congested,
8410 .change_consistency_policy = raid5_change_consistency_policy,
8413 static struct md_personality raid4_personality =
8417 .owner = THIS_MODULE,
8418 .make_request = raid5_make_request,
8421 .status = raid5_status,
8422 .error_handler = raid5_error,
8423 .hot_add_disk = raid5_add_disk,
8424 .hot_remove_disk= raid5_remove_disk,
8425 .spare_active = raid5_spare_active,
8426 .sync_request = raid5_sync_request,
8427 .resize = raid5_resize,
8429 .check_reshape = raid5_check_reshape,
8430 .start_reshape = raid5_start_reshape,
8431 .finish_reshape = raid5_finish_reshape,
8432 .quiesce = raid5_quiesce,
8433 .takeover = raid4_takeover,
8434 .congested = raid5_congested,
8435 .change_consistency_policy = raid5_change_consistency_policy,
8438 static int __init raid5_init(void)
8442 raid5_wq = alloc_workqueue("raid5wq",
8443 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8447 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8449 raid456_cpu_up_prepare,
8452 destroy_workqueue(raid5_wq);
8455 register_md_personality(&raid6_personality);
8456 register_md_personality(&raid5_personality);
8457 register_md_personality(&raid4_personality);
8461 static void raid5_exit(void)
8463 unregister_md_personality(&raid6_personality);
8464 unregister_md_personality(&raid5_personality);
8465 unregister_md_personality(&raid4_personality);
8466 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8467 destroy_workqueue(raid5_wq);
8470 module_init(raid5_init);
8471 module_exit(raid5_exit);
8472 MODULE_LICENSE("GPL");
8473 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8474 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8475 MODULE_ALIAS("md-raid5");
8476 MODULE_ALIAS("md-raid4");
8477 MODULE_ALIAS("md-level-5");
8478 MODULE_ALIAS("md-level-4");
8479 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8480 MODULE_ALIAS("md-raid6");
8481 MODULE_ALIAS("md-level-6");
8483 /* This used to be two separate modules, they were: */
8484 MODULE_ALIAS("raid5");
8485 MODULE_ALIAS("raid6");