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 raid5_build_block(struct stripe_head *sh, int i, int previous);
498 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
499 struct stripe_head *sh);
501 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
503 struct r5conf *conf = sh->raid_conf;
506 BUG_ON(atomic_read(&sh->count) != 0);
507 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
508 BUG_ON(stripe_operations_active(sh));
509 BUG_ON(sh->batch_head);
511 pr_debug("init_stripe called, stripe %llu\n",
512 (unsigned long long)sector);
514 seq = read_seqcount_begin(&conf->gen_lock);
515 sh->generation = conf->generation - previous;
516 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
518 stripe_set_idx(sector, conf, previous, sh);
521 for (i = sh->disks; i--; ) {
522 struct r5dev *dev = &sh->dev[i];
524 if (dev->toread || dev->read || dev->towrite || dev->written ||
525 test_bit(R5_LOCKED, &dev->flags)) {
526 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
527 (unsigned long long)sh->sector, i, dev->toread,
528 dev->read, dev->towrite, dev->written,
529 test_bit(R5_LOCKED, &dev->flags));
533 raid5_build_block(sh, i, previous);
535 if (read_seqcount_retry(&conf->gen_lock, seq))
537 sh->overwrite_disks = 0;
538 insert_hash(conf, sh);
539 sh->cpu = smp_processor_id();
540 set_bit(STRIPE_BATCH_READY, &sh->state);
543 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
546 struct stripe_head *sh;
548 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
549 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
550 if (sh->sector == sector && sh->generation == generation)
552 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
557 * Need to check if array has failed when deciding whether to:
559 * - remove non-faulty devices
562 * This determination is simple when no reshape is happening.
563 * However if there is a reshape, we need to carefully check
564 * both the before and after sections.
565 * This is because some failed devices may only affect one
566 * of the two sections, and some non-in_sync devices may
567 * be insync in the section most affected by failed devices.
569 int raid5_calc_degraded(struct r5conf *conf)
571 int degraded, degraded2;
576 for (i = 0; i < conf->previous_raid_disks; i++) {
577 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
578 if (rdev && test_bit(Faulty, &rdev->flags))
579 rdev = rcu_dereference(conf->disks[i].replacement);
580 if (!rdev || test_bit(Faulty, &rdev->flags))
582 else if (test_bit(In_sync, &rdev->flags))
585 /* not in-sync or faulty.
586 * If the reshape increases the number of devices,
587 * this is being recovered by the reshape, so
588 * this 'previous' section is not in_sync.
589 * If the number of devices is being reduced however,
590 * the device can only be part of the array if
591 * we are reverting a reshape, so this section will
594 if (conf->raid_disks >= conf->previous_raid_disks)
598 if (conf->raid_disks == conf->previous_raid_disks)
602 for (i = 0; i < conf->raid_disks; i++) {
603 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
604 if (rdev && test_bit(Faulty, &rdev->flags))
605 rdev = rcu_dereference(conf->disks[i].replacement);
606 if (!rdev || test_bit(Faulty, &rdev->flags))
608 else if (test_bit(In_sync, &rdev->flags))
611 /* not in-sync or faulty.
612 * If reshape increases the number of devices, this
613 * section has already been recovered, else it
614 * almost certainly hasn't.
616 if (conf->raid_disks <= conf->previous_raid_disks)
620 if (degraded2 > degraded)
625 static int has_failed(struct r5conf *conf)
629 if (conf->mddev->reshape_position == MaxSector)
630 return conf->mddev->degraded > conf->max_degraded;
632 degraded = raid5_calc_degraded(conf);
633 if (degraded > conf->max_degraded)
639 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
640 int previous, int noblock, int noquiesce)
642 struct stripe_head *sh;
643 int hash = stripe_hash_locks_hash(sector);
644 int inc_empty_inactive_list_flag;
646 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
648 spin_lock_irq(conf->hash_locks + hash);
651 wait_event_lock_irq(conf->wait_for_quiescent,
652 conf->quiesce == 0 || noquiesce,
653 *(conf->hash_locks + hash));
654 sh = __find_stripe(conf, sector, conf->generation - previous);
656 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
657 sh = get_free_stripe(conf, hash);
658 if (!sh && !test_bit(R5_DID_ALLOC,
660 set_bit(R5_ALLOC_MORE,
663 if (noblock && sh == NULL)
666 r5c_check_stripe_cache_usage(conf);
668 set_bit(R5_INACTIVE_BLOCKED,
670 r5l_wake_reclaim(conf->log, 0);
672 conf->wait_for_stripe,
673 !list_empty(conf->inactive_list + hash) &&
674 (atomic_read(&conf->active_stripes)
675 < (conf->max_nr_stripes * 3 / 4)
676 || !test_bit(R5_INACTIVE_BLOCKED,
677 &conf->cache_state)),
678 *(conf->hash_locks + hash));
679 clear_bit(R5_INACTIVE_BLOCKED,
682 init_stripe(sh, sector, previous);
683 atomic_inc(&sh->count);
685 } else if (!atomic_inc_not_zero(&sh->count)) {
686 spin_lock(&conf->device_lock);
687 if (!atomic_read(&sh->count)) {
688 if (!test_bit(STRIPE_HANDLE, &sh->state))
689 atomic_inc(&conf->active_stripes);
690 BUG_ON(list_empty(&sh->lru) &&
691 !test_bit(STRIPE_EXPANDING, &sh->state));
692 inc_empty_inactive_list_flag = 0;
693 if (!list_empty(conf->inactive_list + hash))
694 inc_empty_inactive_list_flag = 1;
695 list_del_init(&sh->lru);
696 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
697 atomic_inc(&conf->empty_inactive_list_nr);
699 sh->group->stripes_cnt--;
703 atomic_inc(&sh->count);
704 spin_unlock(&conf->device_lock);
706 } while (sh == NULL);
708 spin_unlock_irq(conf->hash_locks + hash);
712 static bool is_full_stripe_write(struct stripe_head *sh)
714 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
715 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
718 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
721 spin_lock_irq(&sh2->stripe_lock);
722 spin_lock_nested(&sh1->stripe_lock, 1);
724 spin_lock_irq(&sh1->stripe_lock);
725 spin_lock_nested(&sh2->stripe_lock, 1);
729 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
731 spin_unlock(&sh1->stripe_lock);
732 spin_unlock_irq(&sh2->stripe_lock);
735 /* Only freshly new full stripe normal write stripe can be added to a batch list */
736 static bool stripe_can_batch(struct stripe_head *sh)
738 struct r5conf *conf = sh->raid_conf;
740 if (conf->log || raid5_has_ppl(conf))
742 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
743 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
744 is_full_stripe_write(sh);
747 /* we only do back search */
748 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
750 struct stripe_head *head;
751 sector_t head_sector, tmp_sec;
754 int inc_empty_inactive_list_flag;
756 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
757 tmp_sec = sh->sector;
758 if (!sector_div(tmp_sec, conf->chunk_sectors))
760 head_sector = sh->sector - STRIPE_SECTORS;
762 hash = stripe_hash_locks_hash(head_sector);
763 spin_lock_irq(conf->hash_locks + hash);
764 head = __find_stripe(conf, head_sector, conf->generation);
765 if (head && !atomic_inc_not_zero(&head->count)) {
766 spin_lock(&conf->device_lock);
767 if (!atomic_read(&head->count)) {
768 if (!test_bit(STRIPE_HANDLE, &head->state))
769 atomic_inc(&conf->active_stripes);
770 BUG_ON(list_empty(&head->lru) &&
771 !test_bit(STRIPE_EXPANDING, &head->state));
772 inc_empty_inactive_list_flag = 0;
773 if (!list_empty(conf->inactive_list + hash))
774 inc_empty_inactive_list_flag = 1;
775 list_del_init(&head->lru);
776 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
777 atomic_inc(&conf->empty_inactive_list_nr);
779 head->group->stripes_cnt--;
783 atomic_inc(&head->count);
784 spin_unlock(&conf->device_lock);
786 spin_unlock_irq(conf->hash_locks + hash);
790 if (!stripe_can_batch(head))
793 lock_two_stripes(head, sh);
794 /* clear_batch_ready clear the flag */
795 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
802 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
804 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
805 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
808 if (head->batch_head) {
809 spin_lock(&head->batch_head->batch_lock);
810 /* This batch list is already running */
811 if (!stripe_can_batch(head)) {
812 spin_unlock(&head->batch_head->batch_lock);
817 * at this point, head's BATCH_READY could be cleared, but we
818 * can still add the stripe to batch list
820 list_add(&sh->batch_list, &head->batch_list);
821 spin_unlock(&head->batch_head->batch_lock);
823 sh->batch_head = head->batch_head;
825 head->batch_head = head;
826 sh->batch_head = head->batch_head;
827 spin_lock(&head->batch_lock);
828 list_add_tail(&sh->batch_list, &head->batch_list);
829 spin_unlock(&head->batch_lock);
832 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
833 if (atomic_dec_return(&conf->preread_active_stripes)
835 md_wakeup_thread(conf->mddev->thread);
837 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
838 int seq = sh->bm_seq;
839 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
840 sh->batch_head->bm_seq > seq)
841 seq = sh->batch_head->bm_seq;
842 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
843 sh->batch_head->bm_seq = seq;
846 atomic_inc(&sh->count);
848 unlock_two_stripes(head, sh);
850 raid5_release_stripe(head);
853 /* Determine if 'data_offset' or 'new_data_offset' should be used
854 * in this stripe_head.
856 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
858 sector_t progress = conf->reshape_progress;
859 /* Need a memory barrier to make sure we see the value
860 * of conf->generation, or ->data_offset that was set before
861 * reshape_progress was updated.
864 if (progress == MaxSector)
866 if (sh->generation == conf->generation - 1)
868 /* We are in a reshape, and this is a new-generation stripe,
869 * so use new_data_offset.
874 static void dispatch_bio_list(struct bio_list *tmp)
878 while ((bio = bio_list_pop(tmp)))
879 generic_make_request(bio);
882 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
884 const struct r5pending_data *da = list_entry(a,
885 struct r5pending_data, sibling);
886 const struct r5pending_data *db = list_entry(b,
887 struct r5pending_data, sibling);
888 if (da->sector > db->sector)
890 if (da->sector < db->sector)
895 static void dispatch_defer_bios(struct r5conf *conf, int target,
896 struct bio_list *list)
898 struct r5pending_data *data;
899 struct list_head *first, *next = NULL;
902 if (conf->pending_data_cnt == 0)
905 list_sort(NULL, &conf->pending_list, cmp_stripe);
907 first = conf->pending_list.next;
909 /* temporarily move the head */
910 if (conf->next_pending_data)
911 list_move_tail(&conf->pending_list,
912 &conf->next_pending_data->sibling);
914 while (!list_empty(&conf->pending_list)) {
915 data = list_first_entry(&conf->pending_list,
916 struct r5pending_data, sibling);
917 if (&data->sibling == first)
918 first = data->sibling.next;
919 next = data->sibling.next;
921 bio_list_merge(list, &data->bios);
922 list_move(&data->sibling, &conf->free_list);
927 conf->pending_data_cnt -= cnt;
928 BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
930 if (next != &conf->pending_list)
931 conf->next_pending_data = list_entry(next,
932 struct r5pending_data, sibling);
934 conf->next_pending_data = NULL;
935 /* list isn't empty */
936 if (first != &conf->pending_list)
937 list_move_tail(&conf->pending_list, first);
940 static void flush_deferred_bios(struct r5conf *conf)
942 struct bio_list tmp = BIO_EMPTY_LIST;
944 if (conf->pending_data_cnt == 0)
947 spin_lock(&conf->pending_bios_lock);
948 dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
949 BUG_ON(conf->pending_data_cnt != 0);
950 spin_unlock(&conf->pending_bios_lock);
952 dispatch_bio_list(&tmp);
955 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
956 struct bio_list *bios)
958 struct bio_list tmp = BIO_EMPTY_LIST;
959 struct r5pending_data *ent;
961 spin_lock(&conf->pending_bios_lock);
962 ent = list_first_entry(&conf->free_list, struct r5pending_data,
964 list_move_tail(&ent->sibling, &conf->pending_list);
965 ent->sector = sector;
966 bio_list_init(&ent->bios);
967 bio_list_merge(&ent->bios, bios);
968 conf->pending_data_cnt++;
969 if (conf->pending_data_cnt >= PENDING_IO_MAX)
970 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
972 spin_unlock(&conf->pending_bios_lock);
974 dispatch_bio_list(&tmp);
978 raid5_end_read_request(struct bio *bi);
980 raid5_end_write_request(struct bio *bi);
982 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
984 struct r5conf *conf = sh->raid_conf;
985 int i, disks = sh->disks;
986 struct stripe_head *head_sh = sh;
987 struct bio_list pending_bios = BIO_EMPTY_LIST;
992 if (log_stripe(sh, s) == 0)
995 should_defer = conf->batch_bio_dispatch && conf->group_cnt;
997 for (i = disks; i--; ) {
998 int op, op_flags = 0;
999 int replace_only = 0;
1000 struct bio *bi, *rbi;
1001 struct md_rdev *rdev, *rrdev = NULL;
1004 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1006 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1008 if (test_bit(R5_Discard, &sh->dev[i].flags))
1009 op = REQ_OP_DISCARD;
1010 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1012 else if (test_and_clear_bit(R5_WantReplace,
1013 &sh->dev[i].flags)) {
1018 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1019 op_flags |= REQ_SYNC;
1022 bi = &sh->dev[i].req;
1023 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1026 rrdev = rcu_dereference(conf->disks[i].replacement);
1027 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1028 rdev = rcu_dereference(conf->disks[i].rdev);
1033 if (op_is_write(op)) {
1037 /* We raced and saw duplicates */
1040 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1045 if (rdev && test_bit(Faulty, &rdev->flags))
1048 atomic_inc(&rdev->nr_pending);
1049 if (rrdev && test_bit(Faulty, &rrdev->flags))
1052 atomic_inc(&rrdev->nr_pending);
1055 /* We have already checked bad blocks for reads. Now
1056 * need to check for writes. We never accept write errors
1057 * on the replacement, so we don't to check rrdev.
1059 while (op_is_write(op) && rdev &&
1060 test_bit(WriteErrorSeen, &rdev->flags)) {
1063 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1064 &first_bad, &bad_sectors);
1069 set_bit(BlockedBadBlocks, &rdev->flags);
1070 if (!conf->mddev->external &&
1071 conf->mddev->sb_flags) {
1072 /* It is very unlikely, but we might
1073 * still need to write out the
1074 * bad block log - better give it
1076 md_check_recovery(conf->mddev);
1079 * Because md_wait_for_blocked_rdev
1080 * will dec nr_pending, we must
1081 * increment it first.
1083 atomic_inc(&rdev->nr_pending);
1084 md_wait_for_blocked_rdev(rdev, conf->mddev);
1086 /* Acknowledged bad block - skip the write */
1087 rdev_dec_pending(rdev, conf->mddev);
1093 if (s->syncing || s->expanding || s->expanded
1095 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1097 set_bit(STRIPE_IO_STARTED, &sh->state);
1099 bio_set_dev(bi, rdev->bdev);
1100 bio_set_op_attrs(bi, op, op_flags);
1101 bi->bi_end_io = op_is_write(op)
1102 ? raid5_end_write_request
1103 : raid5_end_read_request;
1104 bi->bi_private = sh;
1106 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1107 __func__, (unsigned long long)sh->sector,
1109 atomic_inc(&sh->count);
1111 atomic_inc(&head_sh->count);
1112 if (use_new_offset(conf, sh))
1113 bi->bi_iter.bi_sector = (sh->sector
1114 + rdev->new_data_offset);
1116 bi->bi_iter.bi_sector = (sh->sector
1117 + rdev->data_offset);
1118 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1119 bi->bi_opf |= REQ_NOMERGE;
1121 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1122 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1124 if (!op_is_write(op) &&
1125 test_bit(R5_InJournal, &sh->dev[i].flags))
1127 * issuing read for a page in journal, this
1128 * must be preparing for prexor in rmw; read
1129 * the data into orig_page
1131 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1133 sh->dev[i].vec.bv_page = sh->dev[i].page;
1135 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1136 bi->bi_io_vec[0].bv_offset = 0;
1137 bi->bi_iter.bi_size = STRIPE_SIZE;
1139 * If this is discard request, set bi_vcnt 0. We don't
1140 * want to confuse SCSI because SCSI will replace payload
1142 if (op == REQ_OP_DISCARD)
1145 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1147 if (conf->mddev->gendisk)
1148 trace_block_bio_remap(bi->bi_disk->queue,
1149 bi, disk_devt(conf->mddev->gendisk),
1151 if (should_defer && op_is_write(op))
1152 bio_list_add(&pending_bios, bi);
1154 generic_make_request(bi);
1157 if (s->syncing || s->expanding || s->expanded
1159 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1161 set_bit(STRIPE_IO_STARTED, &sh->state);
1163 bio_set_dev(rbi, rrdev->bdev);
1164 bio_set_op_attrs(rbi, op, op_flags);
1165 BUG_ON(!op_is_write(op));
1166 rbi->bi_end_io = raid5_end_write_request;
1167 rbi->bi_private = sh;
1169 pr_debug("%s: for %llu schedule op %d on "
1170 "replacement disc %d\n",
1171 __func__, (unsigned long long)sh->sector,
1173 atomic_inc(&sh->count);
1175 atomic_inc(&head_sh->count);
1176 if (use_new_offset(conf, sh))
1177 rbi->bi_iter.bi_sector = (sh->sector
1178 + rrdev->new_data_offset);
1180 rbi->bi_iter.bi_sector = (sh->sector
1181 + rrdev->data_offset);
1182 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1183 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1184 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1186 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1187 rbi->bi_io_vec[0].bv_offset = 0;
1188 rbi->bi_iter.bi_size = STRIPE_SIZE;
1190 * If this is discard request, set bi_vcnt 0. We don't
1191 * want to confuse SCSI because SCSI will replace payload
1193 if (op == REQ_OP_DISCARD)
1195 if (conf->mddev->gendisk)
1196 trace_block_bio_remap(rbi->bi_disk->queue,
1197 rbi, disk_devt(conf->mddev->gendisk),
1199 if (should_defer && op_is_write(op))
1200 bio_list_add(&pending_bios, rbi);
1202 generic_make_request(rbi);
1204 if (!rdev && !rrdev) {
1205 if (op_is_write(op))
1206 set_bit(STRIPE_DEGRADED, &sh->state);
1207 pr_debug("skip op %d on disc %d for sector %llu\n",
1208 bi->bi_opf, i, (unsigned long long)sh->sector);
1209 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1210 set_bit(STRIPE_HANDLE, &sh->state);
1213 if (!head_sh->batch_head)
1215 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1221 if (should_defer && !bio_list_empty(&pending_bios))
1222 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1225 static struct dma_async_tx_descriptor *
1226 async_copy_data(int frombio, struct bio *bio, struct page **page,
1227 sector_t sector, struct dma_async_tx_descriptor *tx,
1228 struct stripe_head *sh, int no_skipcopy)
1231 struct bvec_iter iter;
1232 struct page *bio_page;
1234 struct async_submit_ctl submit;
1235 enum async_tx_flags flags = 0;
1237 if (bio->bi_iter.bi_sector >= sector)
1238 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1240 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1243 flags |= ASYNC_TX_FENCE;
1244 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1246 bio_for_each_segment(bvl, bio, iter) {
1247 int len = bvl.bv_len;
1251 if (page_offset < 0) {
1252 b_offset = -page_offset;
1253 page_offset += b_offset;
1257 if (len > 0 && page_offset + len > STRIPE_SIZE)
1258 clen = STRIPE_SIZE - page_offset;
1263 b_offset += bvl.bv_offset;
1264 bio_page = bvl.bv_page;
1266 if (sh->raid_conf->skip_copy &&
1267 b_offset == 0 && page_offset == 0 &&
1268 clen == STRIPE_SIZE &&
1272 tx = async_memcpy(*page, bio_page, page_offset,
1273 b_offset, clen, &submit);
1275 tx = async_memcpy(bio_page, *page, b_offset,
1276 page_offset, clen, &submit);
1278 /* chain the operations */
1279 submit.depend_tx = tx;
1281 if (clen < len) /* hit end of page */
1289 static void ops_complete_biofill(void *stripe_head_ref)
1291 struct stripe_head *sh = stripe_head_ref;
1294 pr_debug("%s: stripe %llu\n", __func__,
1295 (unsigned long long)sh->sector);
1297 /* clear completed biofills */
1298 for (i = sh->disks; i--; ) {
1299 struct r5dev *dev = &sh->dev[i];
1301 /* acknowledge completion of a biofill operation */
1302 /* and check if we need to reply to a read request,
1303 * new R5_Wantfill requests are held off until
1304 * !STRIPE_BIOFILL_RUN
1306 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1307 struct bio *rbi, *rbi2;
1312 while (rbi && rbi->bi_iter.bi_sector <
1313 dev->sector + STRIPE_SECTORS) {
1314 rbi2 = r5_next_bio(rbi, dev->sector);
1320 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1322 set_bit(STRIPE_HANDLE, &sh->state);
1323 raid5_release_stripe(sh);
1326 static void ops_run_biofill(struct stripe_head *sh)
1328 struct dma_async_tx_descriptor *tx = NULL;
1329 struct async_submit_ctl submit;
1332 BUG_ON(sh->batch_head);
1333 pr_debug("%s: stripe %llu\n", __func__,
1334 (unsigned long long)sh->sector);
1336 for (i = sh->disks; i--; ) {
1337 struct r5dev *dev = &sh->dev[i];
1338 if (test_bit(R5_Wantfill, &dev->flags)) {
1340 spin_lock_irq(&sh->stripe_lock);
1341 dev->read = rbi = dev->toread;
1343 spin_unlock_irq(&sh->stripe_lock);
1344 while (rbi && rbi->bi_iter.bi_sector <
1345 dev->sector + STRIPE_SECTORS) {
1346 tx = async_copy_data(0, rbi, &dev->page,
1347 dev->sector, tx, sh, 0);
1348 rbi = r5_next_bio(rbi, dev->sector);
1353 atomic_inc(&sh->count);
1354 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1355 async_trigger_callback(&submit);
1358 static void mark_target_uptodate(struct stripe_head *sh, int target)
1365 tgt = &sh->dev[target];
1366 set_bit(R5_UPTODATE, &tgt->flags);
1367 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1368 clear_bit(R5_Wantcompute, &tgt->flags);
1371 static void ops_complete_compute(void *stripe_head_ref)
1373 struct stripe_head *sh = stripe_head_ref;
1375 pr_debug("%s: stripe %llu\n", __func__,
1376 (unsigned long long)sh->sector);
1378 /* mark the computed target(s) as uptodate */
1379 mark_target_uptodate(sh, sh->ops.target);
1380 mark_target_uptodate(sh, sh->ops.target2);
1382 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1383 if (sh->check_state == check_state_compute_run)
1384 sh->check_state = check_state_compute_result;
1385 set_bit(STRIPE_HANDLE, &sh->state);
1386 raid5_release_stripe(sh);
1389 /* return a pointer to the address conversion region of the scribble buffer */
1390 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1391 struct raid5_percpu *percpu, int i)
1395 addr = flex_array_get(percpu->scribble, i);
1396 return addr + sizeof(struct page *) * (sh->disks + 2);
1399 /* return a pointer to the address conversion region of the scribble buffer */
1400 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1404 addr = flex_array_get(percpu->scribble, i);
1408 static struct dma_async_tx_descriptor *
1409 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1411 int disks = sh->disks;
1412 struct page **xor_srcs = to_addr_page(percpu, 0);
1413 int target = sh->ops.target;
1414 struct r5dev *tgt = &sh->dev[target];
1415 struct page *xor_dest = tgt->page;
1417 struct dma_async_tx_descriptor *tx;
1418 struct async_submit_ctl submit;
1421 BUG_ON(sh->batch_head);
1423 pr_debug("%s: stripe %llu block: %d\n",
1424 __func__, (unsigned long long)sh->sector, target);
1425 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1427 for (i = disks; i--; )
1429 xor_srcs[count++] = sh->dev[i].page;
1431 atomic_inc(&sh->count);
1433 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1434 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1435 if (unlikely(count == 1))
1436 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1438 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1443 /* set_syndrome_sources - populate source buffers for gen_syndrome
1444 * @srcs - (struct page *) array of size sh->disks
1445 * @sh - stripe_head to parse
1447 * Populates srcs in proper layout order for the stripe and returns the
1448 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1449 * destination buffer is recorded in srcs[count] and the Q destination
1450 * is recorded in srcs[count+1]].
1452 static int set_syndrome_sources(struct page **srcs,
1453 struct stripe_head *sh,
1456 int disks = sh->disks;
1457 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1458 int d0_idx = raid6_d0(sh);
1462 for (i = 0; i < disks; i++)
1468 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1469 struct r5dev *dev = &sh->dev[i];
1471 if (i == sh->qd_idx || i == sh->pd_idx ||
1472 (srctype == SYNDROME_SRC_ALL) ||
1473 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1474 (test_bit(R5_Wantdrain, &dev->flags) ||
1475 test_bit(R5_InJournal, &dev->flags))) ||
1476 (srctype == SYNDROME_SRC_WRITTEN &&
1478 test_bit(R5_InJournal, &dev->flags)))) {
1479 if (test_bit(R5_InJournal, &dev->flags))
1480 srcs[slot] = sh->dev[i].orig_page;
1482 srcs[slot] = sh->dev[i].page;
1484 i = raid6_next_disk(i, disks);
1485 } while (i != d0_idx);
1487 return syndrome_disks;
1490 static struct dma_async_tx_descriptor *
1491 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1493 int disks = sh->disks;
1494 struct page **blocks = to_addr_page(percpu, 0);
1496 int qd_idx = sh->qd_idx;
1497 struct dma_async_tx_descriptor *tx;
1498 struct async_submit_ctl submit;
1504 BUG_ON(sh->batch_head);
1505 if (sh->ops.target < 0)
1506 target = sh->ops.target2;
1507 else if (sh->ops.target2 < 0)
1508 target = sh->ops.target;
1510 /* we should only have one valid target */
1513 pr_debug("%s: stripe %llu block: %d\n",
1514 __func__, (unsigned long long)sh->sector, target);
1516 tgt = &sh->dev[target];
1517 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1520 atomic_inc(&sh->count);
1522 if (target == qd_idx) {
1523 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1524 blocks[count] = NULL; /* regenerating p is not necessary */
1525 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1526 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1527 ops_complete_compute, sh,
1528 to_addr_conv(sh, percpu, 0));
1529 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1531 /* Compute any data- or p-drive using XOR */
1533 for (i = disks; i-- ; ) {
1534 if (i == target || i == qd_idx)
1536 blocks[count++] = sh->dev[i].page;
1539 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1540 NULL, ops_complete_compute, sh,
1541 to_addr_conv(sh, percpu, 0));
1542 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1548 static struct dma_async_tx_descriptor *
1549 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1551 int i, count, disks = sh->disks;
1552 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1553 int d0_idx = raid6_d0(sh);
1554 int faila = -1, failb = -1;
1555 int target = sh->ops.target;
1556 int target2 = sh->ops.target2;
1557 struct r5dev *tgt = &sh->dev[target];
1558 struct r5dev *tgt2 = &sh->dev[target2];
1559 struct dma_async_tx_descriptor *tx;
1560 struct page **blocks = to_addr_page(percpu, 0);
1561 struct async_submit_ctl submit;
1563 BUG_ON(sh->batch_head);
1564 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1565 __func__, (unsigned long long)sh->sector, target, target2);
1566 BUG_ON(target < 0 || target2 < 0);
1567 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1568 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1570 /* we need to open-code set_syndrome_sources to handle the
1571 * slot number conversion for 'faila' and 'failb'
1573 for (i = 0; i < disks ; i++)
1578 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1580 blocks[slot] = sh->dev[i].page;
1586 i = raid6_next_disk(i, disks);
1587 } while (i != d0_idx);
1589 BUG_ON(faila == failb);
1592 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1593 __func__, (unsigned long long)sh->sector, faila, failb);
1595 atomic_inc(&sh->count);
1597 if (failb == syndrome_disks+1) {
1598 /* Q disk is one of the missing disks */
1599 if (faila == syndrome_disks) {
1600 /* Missing P+Q, just recompute */
1601 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1602 ops_complete_compute, sh,
1603 to_addr_conv(sh, percpu, 0));
1604 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1605 STRIPE_SIZE, &submit);
1609 int qd_idx = sh->qd_idx;
1611 /* Missing D+Q: recompute D from P, then recompute Q */
1612 if (target == qd_idx)
1613 data_target = target2;
1615 data_target = target;
1618 for (i = disks; i-- ; ) {
1619 if (i == data_target || i == qd_idx)
1621 blocks[count++] = sh->dev[i].page;
1623 dest = sh->dev[data_target].page;
1624 init_async_submit(&submit,
1625 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1627 to_addr_conv(sh, percpu, 0));
1628 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1631 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1632 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1633 ops_complete_compute, sh,
1634 to_addr_conv(sh, percpu, 0));
1635 return async_gen_syndrome(blocks, 0, count+2,
1636 STRIPE_SIZE, &submit);
1639 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1640 ops_complete_compute, sh,
1641 to_addr_conv(sh, percpu, 0));
1642 if (failb == syndrome_disks) {
1643 /* We're missing D+P. */
1644 return async_raid6_datap_recov(syndrome_disks+2,
1648 /* We're missing D+D. */
1649 return async_raid6_2data_recov(syndrome_disks+2,
1650 STRIPE_SIZE, faila, failb,
1656 static void ops_complete_prexor(void *stripe_head_ref)
1658 struct stripe_head *sh = stripe_head_ref;
1660 pr_debug("%s: stripe %llu\n", __func__,
1661 (unsigned long long)sh->sector);
1663 if (r5c_is_writeback(sh->raid_conf->log))
1665 * raid5-cache write back uses orig_page during prexor.
1666 * After prexor, it is time to free orig_page
1668 r5c_release_extra_page(sh);
1671 static struct dma_async_tx_descriptor *
1672 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1673 struct dma_async_tx_descriptor *tx)
1675 int disks = sh->disks;
1676 struct page **xor_srcs = to_addr_page(percpu, 0);
1677 int count = 0, pd_idx = sh->pd_idx, i;
1678 struct async_submit_ctl submit;
1680 /* existing parity data subtracted */
1681 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1683 BUG_ON(sh->batch_head);
1684 pr_debug("%s: stripe %llu\n", __func__,
1685 (unsigned long long)sh->sector);
1687 for (i = disks; i--; ) {
1688 struct r5dev *dev = &sh->dev[i];
1689 /* Only process blocks that are known to be uptodate */
1690 if (test_bit(R5_InJournal, &dev->flags))
1691 xor_srcs[count++] = dev->orig_page;
1692 else if (test_bit(R5_Wantdrain, &dev->flags))
1693 xor_srcs[count++] = dev->page;
1696 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1697 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1698 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1703 static struct dma_async_tx_descriptor *
1704 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1705 struct dma_async_tx_descriptor *tx)
1707 struct page **blocks = to_addr_page(percpu, 0);
1709 struct async_submit_ctl submit;
1711 pr_debug("%s: stripe %llu\n", __func__,
1712 (unsigned long long)sh->sector);
1714 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1716 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1717 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1718 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1723 static struct dma_async_tx_descriptor *
1724 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1726 struct r5conf *conf = sh->raid_conf;
1727 int disks = sh->disks;
1729 struct stripe_head *head_sh = sh;
1731 pr_debug("%s: stripe %llu\n", __func__,
1732 (unsigned long long)sh->sector);
1734 for (i = disks; i--; ) {
1739 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1745 * clear R5_InJournal, so when rewriting a page in
1746 * journal, it is not skipped by r5l_log_stripe()
1748 clear_bit(R5_InJournal, &dev->flags);
1749 spin_lock_irq(&sh->stripe_lock);
1750 chosen = dev->towrite;
1751 dev->towrite = NULL;
1752 sh->overwrite_disks = 0;
1753 BUG_ON(dev->written);
1754 wbi = dev->written = chosen;
1755 spin_unlock_irq(&sh->stripe_lock);
1756 WARN_ON(dev->page != dev->orig_page);
1758 while (wbi && wbi->bi_iter.bi_sector <
1759 dev->sector + STRIPE_SECTORS) {
1760 if (wbi->bi_opf & REQ_FUA)
1761 set_bit(R5_WantFUA, &dev->flags);
1762 if (wbi->bi_opf & REQ_SYNC)
1763 set_bit(R5_SyncIO, &dev->flags);
1764 if (bio_op(wbi) == REQ_OP_DISCARD)
1765 set_bit(R5_Discard, &dev->flags);
1767 tx = async_copy_data(1, wbi, &dev->page,
1768 dev->sector, tx, sh,
1769 r5c_is_writeback(conf->log));
1770 if (dev->page != dev->orig_page &&
1771 !r5c_is_writeback(conf->log)) {
1772 set_bit(R5_SkipCopy, &dev->flags);
1773 clear_bit(R5_UPTODATE, &dev->flags);
1774 clear_bit(R5_OVERWRITE, &dev->flags);
1777 wbi = r5_next_bio(wbi, dev->sector);
1780 if (head_sh->batch_head) {
1781 sh = list_first_entry(&sh->batch_list,
1794 static void ops_complete_reconstruct(void *stripe_head_ref)
1796 struct stripe_head *sh = stripe_head_ref;
1797 int disks = sh->disks;
1798 int pd_idx = sh->pd_idx;
1799 int qd_idx = sh->qd_idx;
1801 bool fua = false, sync = false, discard = false;
1803 pr_debug("%s: stripe %llu\n", __func__,
1804 (unsigned long long)sh->sector);
1806 for (i = disks; i--; ) {
1807 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1808 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1809 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1812 for (i = disks; i--; ) {
1813 struct r5dev *dev = &sh->dev[i];
1815 if (dev->written || i == pd_idx || i == qd_idx) {
1816 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1817 set_bit(R5_UPTODATE, &dev->flags);
1819 set_bit(R5_WantFUA, &dev->flags);
1821 set_bit(R5_SyncIO, &dev->flags);
1825 if (sh->reconstruct_state == reconstruct_state_drain_run)
1826 sh->reconstruct_state = reconstruct_state_drain_result;
1827 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1828 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1830 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1831 sh->reconstruct_state = reconstruct_state_result;
1834 set_bit(STRIPE_HANDLE, &sh->state);
1835 raid5_release_stripe(sh);
1839 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1840 struct dma_async_tx_descriptor *tx)
1842 int disks = sh->disks;
1843 struct page **xor_srcs;
1844 struct async_submit_ctl submit;
1845 int count, pd_idx = sh->pd_idx, i;
1846 struct page *xor_dest;
1848 unsigned long flags;
1850 struct stripe_head *head_sh = sh;
1853 pr_debug("%s: stripe %llu\n", __func__,
1854 (unsigned long long)sh->sector);
1856 for (i = 0; i < sh->disks; i++) {
1859 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1862 if (i >= sh->disks) {
1863 atomic_inc(&sh->count);
1864 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1865 ops_complete_reconstruct(sh);
1870 xor_srcs = to_addr_page(percpu, j);
1871 /* check if prexor is active which means only process blocks
1872 * that are part of a read-modify-write (written)
1874 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1876 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1877 for (i = disks; i--; ) {
1878 struct r5dev *dev = &sh->dev[i];
1879 if (head_sh->dev[i].written ||
1880 test_bit(R5_InJournal, &head_sh->dev[i].flags))
1881 xor_srcs[count++] = dev->page;
1884 xor_dest = sh->dev[pd_idx].page;
1885 for (i = disks; i--; ) {
1886 struct r5dev *dev = &sh->dev[i];
1888 xor_srcs[count++] = dev->page;
1892 /* 1/ if we prexor'd then the dest is reused as a source
1893 * 2/ if we did not prexor then we are redoing the parity
1894 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1895 * for the synchronous xor case
1897 last_stripe = !head_sh->batch_head ||
1898 list_first_entry(&sh->batch_list,
1899 struct stripe_head, batch_list) == head_sh;
1901 flags = ASYNC_TX_ACK |
1902 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1904 atomic_inc(&head_sh->count);
1905 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1906 to_addr_conv(sh, percpu, j));
1908 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1909 init_async_submit(&submit, flags, tx, NULL, NULL,
1910 to_addr_conv(sh, percpu, j));
1913 if (unlikely(count == 1))
1914 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1916 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1919 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1926 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1927 struct dma_async_tx_descriptor *tx)
1929 struct async_submit_ctl submit;
1930 struct page **blocks;
1931 int count, i, j = 0;
1932 struct stripe_head *head_sh = sh;
1935 unsigned long txflags;
1937 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1939 for (i = 0; i < sh->disks; i++) {
1940 if (sh->pd_idx == i || sh->qd_idx == i)
1942 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1945 if (i >= sh->disks) {
1946 atomic_inc(&sh->count);
1947 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1948 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1949 ops_complete_reconstruct(sh);
1954 blocks = to_addr_page(percpu, j);
1956 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1957 synflags = SYNDROME_SRC_WRITTEN;
1958 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1960 synflags = SYNDROME_SRC_ALL;
1961 txflags = ASYNC_TX_ACK;
1964 count = set_syndrome_sources(blocks, sh, synflags);
1965 last_stripe = !head_sh->batch_head ||
1966 list_first_entry(&sh->batch_list,
1967 struct stripe_head, batch_list) == head_sh;
1970 atomic_inc(&head_sh->count);
1971 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1972 head_sh, to_addr_conv(sh, percpu, j));
1974 init_async_submit(&submit, 0, tx, NULL, NULL,
1975 to_addr_conv(sh, percpu, j));
1976 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1979 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1985 static void ops_complete_check(void *stripe_head_ref)
1987 struct stripe_head *sh = stripe_head_ref;
1989 pr_debug("%s: stripe %llu\n", __func__,
1990 (unsigned long long)sh->sector);
1992 sh->check_state = check_state_check_result;
1993 set_bit(STRIPE_HANDLE, &sh->state);
1994 raid5_release_stripe(sh);
1997 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1999 int disks = sh->disks;
2000 int pd_idx = sh->pd_idx;
2001 int qd_idx = sh->qd_idx;
2002 struct page *xor_dest;
2003 struct page **xor_srcs = to_addr_page(percpu, 0);
2004 struct dma_async_tx_descriptor *tx;
2005 struct async_submit_ctl submit;
2009 pr_debug("%s: stripe %llu\n", __func__,
2010 (unsigned long long)sh->sector);
2012 BUG_ON(sh->batch_head);
2014 xor_dest = sh->dev[pd_idx].page;
2015 xor_srcs[count++] = xor_dest;
2016 for (i = disks; i--; ) {
2017 if (i == pd_idx || i == qd_idx)
2019 xor_srcs[count++] = sh->dev[i].page;
2022 init_async_submit(&submit, 0, NULL, NULL, NULL,
2023 to_addr_conv(sh, percpu, 0));
2024 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2025 &sh->ops.zero_sum_result, &submit);
2027 atomic_inc(&sh->count);
2028 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2029 tx = async_trigger_callback(&submit);
2032 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2034 struct page **srcs = to_addr_page(percpu, 0);
2035 struct async_submit_ctl submit;
2038 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2039 (unsigned long long)sh->sector, checkp);
2041 BUG_ON(sh->batch_head);
2042 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2046 atomic_inc(&sh->count);
2047 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2048 sh, to_addr_conv(sh, percpu, 0));
2049 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2050 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2053 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2055 int overlap_clear = 0, i, disks = sh->disks;
2056 struct dma_async_tx_descriptor *tx = NULL;
2057 struct r5conf *conf = sh->raid_conf;
2058 int level = conf->level;
2059 struct raid5_percpu *percpu;
2063 percpu = per_cpu_ptr(conf->percpu, cpu);
2064 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2065 ops_run_biofill(sh);
2069 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2071 tx = ops_run_compute5(sh, percpu);
2073 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2074 tx = ops_run_compute6_1(sh, percpu);
2076 tx = ops_run_compute6_2(sh, percpu);
2078 /* terminate the chain if reconstruct is not set to be run */
2079 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2083 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2085 tx = ops_run_prexor5(sh, percpu, tx);
2087 tx = ops_run_prexor6(sh, percpu, tx);
2090 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2091 tx = ops_run_partial_parity(sh, percpu, tx);
2093 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2094 tx = ops_run_biodrain(sh, tx);
2098 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2100 ops_run_reconstruct5(sh, percpu, tx);
2102 ops_run_reconstruct6(sh, percpu, tx);
2105 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2106 if (sh->check_state == check_state_run)
2107 ops_run_check_p(sh, percpu);
2108 else if (sh->check_state == check_state_run_q)
2109 ops_run_check_pq(sh, percpu, 0);
2110 else if (sh->check_state == check_state_run_pq)
2111 ops_run_check_pq(sh, percpu, 1);
2116 if (overlap_clear && !sh->batch_head)
2117 for (i = disks; i--; ) {
2118 struct r5dev *dev = &sh->dev[i];
2119 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2120 wake_up(&sh->raid_conf->wait_for_overlap);
2125 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2128 __free_page(sh->ppl_page);
2129 kmem_cache_free(sc, sh);
2132 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2133 int disks, struct r5conf *conf)
2135 struct stripe_head *sh;
2138 sh = kmem_cache_zalloc(sc, gfp);
2140 spin_lock_init(&sh->stripe_lock);
2141 spin_lock_init(&sh->batch_lock);
2142 INIT_LIST_HEAD(&sh->batch_list);
2143 INIT_LIST_HEAD(&sh->lru);
2144 INIT_LIST_HEAD(&sh->r5c);
2145 INIT_LIST_HEAD(&sh->log_list);
2146 atomic_set(&sh->count, 1);
2147 sh->raid_conf = conf;
2148 sh->log_start = MaxSector;
2149 for (i = 0; i < disks; i++) {
2150 struct r5dev *dev = &sh->dev[i];
2152 bio_init(&dev->req, &dev->vec, 1);
2153 bio_init(&dev->rreq, &dev->rvec, 1);
2156 if (raid5_has_ppl(conf)) {
2157 sh->ppl_page = alloc_page(gfp);
2158 if (!sh->ppl_page) {
2159 free_stripe(sc, sh);
2166 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2168 struct stripe_head *sh;
2170 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2174 if (grow_buffers(sh, gfp)) {
2176 free_stripe(conf->slab_cache, sh);
2179 sh->hash_lock_index =
2180 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2181 /* we just created an active stripe so... */
2182 atomic_inc(&conf->active_stripes);
2184 raid5_release_stripe(sh);
2185 conf->max_nr_stripes++;
2189 static int grow_stripes(struct r5conf *conf, int num)
2191 struct kmem_cache *sc;
2192 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2194 if (conf->mddev->gendisk)
2195 sprintf(conf->cache_name[0],
2196 "raid%d-%s", conf->level, mdname(conf->mddev));
2198 sprintf(conf->cache_name[0],
2199 "raid%d-%p", conf->level, conf->mddev);
2200 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2202 conf->active_name = 0;
2203 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2204 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2208 conf->slab_cache = sc;
2209 conf->pool_size = devs;
2211 if (!grow_one_stripe(conf, GFP_KERNEL))
2218 * scribble_len - return the required size of the scribble region
2219 * @num - total number of disks in the array
2221 * The size must be enough to contain:
2222 * 1/ a struct page pointer for each device in the array +2
2223 * 2/ room to convert each entry in (1) to its corresponding dma
2224 * (dma_map_page()) or page (page_address()) address.
2226 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2227 * calculate over all devices (not just the data blocks), using zeros in place
2228 * of the P and Q blocks.
2230 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2232 struct flex_array *ret;
2235 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2236 ret = flex_array_alloc(len, cnt, flags);
2239 /* always prealloc all elements, so no locking is required */
2240 if (flex_array_prealloc(ret, 0, cnt, flags)) {
2241 flex_array_free(ret);
2247 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2253 * Never shrink. And mddev_suspend() could deadlock if this is called
2254 * from raid5d. In that case, scribble_disks and scribble_sectors
2255 * should equal to new_disks and new_sectors
2257 if (conf->scribble_disks >= new_disks &&
2258 conf->scribble_sectors >= new_sectors)
2260 mddev_suspend(conf->mddev);
2262 for_each_present_cpu(cpu) {
2263 struct raid5_percpu *percpu;
2264 struct flex_array *scribble;
2266 percpu = per_cpu_ptr(conf->percpu, cpu);
2267 scribble = scribble_alloc(new_disks,
2268 new_sectors / STRIPE_SECTORS,
2272 flex_array_free(percpu->scribble);
2273 percpu->scribble = scribble;
2280 mddev_resume(conf->mddev);
2282 conf->scribble_disks = new_disks;
2283 conf->scribble_sectors = new_sectors;
2288 static int resize_stripes(struct r5conf *conf, int newsize)
2290 /* Make all the stripes able to hold 'newsize' devices.
2291 * New slots in each stripe get 'page' set to a new page.
2293 * This happens in stages:
2294 * 1/ create a new kmem_cache and allocate the required number of
2296 * 2/ gather all the old stripe_heads and transfer the pages across
2297 * to the new stripe_heads. This will have the side effect of
2298 * freezing the array as once all stripe_heads have been collected,
2299 * no IO will be possible. Old stripe heads are freed once their
2300 * pages have been transferred over, and the old kmem_cache is
2301 * freed when all stripes are done.
2302 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2303 * we simple return a failure status - no need to clean anything up.
2304 * 4/ allocate new pages for the new slots in the new stripe_heads.
2305 * If this fails, we don't bother trying the shrink the
2306 * stripe_heads down again, we just leave them as they are.
2307 * As each stripe_head is processed the new one is released into
2310 * Once step2 is started, we cannot afford to wait for a write,
2311 * so we use GFP_NOIO allocations.
2313 struct stripe_head *osh, *nsh;
2314 LIST_HEAD(newstripes);
2315 struct disk_info *ndisks;
2317 struct kmem_cache *sc;
2321 md_allow_write(conf->mddev);
2324 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2325 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2330 /* Need to ensure auto-resizing doesn't interfere */
2331 mutex_lock(&conf->cache_size_mutex);
2333 for (i = conf->max_nr_stripes; i; i--) {
2334 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2338 list_add(&nsh->lru, &newstripes);
2341 /* didn't get enough, give up */
2342 while (!list_empty(&newstripes)) {
2343 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2344 list_del(&nsh->lru);
2345 free_stripe(sc, nsh);
2347 kmem_cache_destroy(sc);
2348 mutex_unlock(&conf->cache_size_mutex);
2351 /* Step 2 - Must use GFP_NOIO now.
2352 * OK, we have enough stripes, start collecting inactive
2353 * stripes and copying them over
2357 list_for_each_entry(nsh, &newstripes, lru) {
2358 lock_device_hash_lock(conf, hash);
2359 wait_event_cmd(conf->wait_for_stripe,
2360 !list_empty(conf->inactive_list + hash),
2361 unlock_device_hash_lock(conf, hash),
2362 lock_device_hash_lock(conf, hash));
2363 osh = get_free_stripe(conf, hash);
2364 unlock_device_hash_lock(conf, hash);
2366 for(i=0; i<conf->pool_size; i++) {
2367 nsh->dev[i].page = osh->dev[i].page;
2368 nsh->dev[i].orig_page = osh->dev[i].page;
2370 nsh->hash_lock_index = hash;
2371 free_stripe(conf->slab_cache, osh);
2373 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2374 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2379 kmem_cache_destroy(conf->slab_cache);
2382 * At this point, we are holding all the stripes so the array
2383 * is completely stalled, so now is a good time to resize
2384 * conf->disks and the scribble region
2386 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2388 for (i = 0; i < conf->pool_size; i++)
2389 ndisks[i] = conf->disks[i];
2391 for (i = conf->pool_size; i < newsize; i++) {
2392 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2393 if (!ndisks[i].extra_page)
2398 for (i = conf->pool_size; i < newsize; i++)
2399 if (ndisks[i].extra_page)
2400 put_page(ndisks[i].extra_page);
2404 conf->disks = ndisks;
2409 mutex_unlock(&conf->cache_size_mutex);
2411 conf->slab_cache = sc;
2412 conf->active_name = 1-conf->active_name;
2414 /* Step 4, return new stripes to service */
2415 while(!list_empty(&newstripes)) {
2416 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2417 list_del_init(&nsh->lru);
2419 for (i=conf->raid_disks; i < newsize; i++)
2420 if (nsh->dev[i].page == NULL) {
2421 struct page *p = alloc_page(GFP_NOIO);
2422 nsh->dev[i].page = p;
2423 nsh->dev[i].orig_page = p;
2427 raid5_release_stripe(nsh);
2429 /* critical section pass, GFP_NOIO no longer needed */
2432 conf->pool_size = newsize;
2436 static int drop_one_stripe(struct r5conf *conf)
2438 struct stripe_head *sh;
2439 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2441 spin_lock_irq(conf->hash_locks + hash);
2442 sh = get_free_stripe(conf, hash);
2443 spin_unlock_irq(conf->hash_locks + hash);
2446 BUG_ON(atomic_read(&sh->count));
2448 free_stripe(conf->slab_cache, sh);
2449 atomic_dec(&conf->active_stripes);
2450 conf->max_nr_stripes--;
2454 static void shrink_stripes(struct r5conf *conf)
2456 while (conf->max_nr_stripes &&
2457 drop_one_stripe(conf))
2460 kmem_cache_destroy(conf->slab_cache);
2461 conf->slab_cache = NULL;
2464 static void raid5_end_read_request(struct bio * bi)
2466 struct stripe_head *sh = bi->bi_private;
2467 struct r5conf *conf = sh->raid_conf;
2468 int disks = sh->disks, i;
2469 char b[BDEVNAME_SIZE];
2470 struct md_rdev *rdev = NULL;
2473 for (i=0 ; i<disks; i++)
2474 if (bi == &sh->dev[i].req)
2477 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2478 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2485 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2486 /* If replacement finished while this request was outstanding,
2487 * 'replacement' might be NULL already.
2488 * In that case it moved down to 'rdev'.
2489 * rdev is not removed until all requests are finished.
2491 rdev = conf->disks[i].replacement;
2493 rdev = conf->disks[i].rdev;
2495 if (use_new_offset(conf, sh))
2496 s = sh->sector + rdev->new_data_offset;
2498 s = sh->sector + rdev->data_offset;
2499 if (!bi->bi_status) {
2500 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2501 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2502 /* Note that this cannot happen on a
2503 * replacement device. We just fail those on
2506 pr_info_ratelimited(
2507 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2508 mdname(conf->mddev), STRIPE_SECTORS,
2509 (unsigned long long)s,
2510 bdevname(rdev->bdev, b));
2511 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2512 clear_bit(R5_ReadError, &sh->dev[i].flags);
2513 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2514 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2515 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2517 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2519 * end read for a page in journal, this
2520 * must be preparing for prexor in rmw
2522 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2524 if (atomic_read(&rdev->read_errors))
2525 atomic_set(&rdev->read_errors, 0);
2527 const char *bdn = bdevname(rdev->bdev, b);
2531 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2532 atomic_inc(&rdev->read_errors);
2533 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2534 pr_warn_ratelimited(
2535 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2536 mdname(conf->mddev),
2537 (unsigned long long)s,
2539 else if (conf->mddev->degraded >= conf->max_degraded) {
2541 pr_warn_ratelimited(
2542 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2543 mdname(conf->mddev),
2544 (unsigned long long)s,
2546 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2549 pr_warn_ratelimited(
2550 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2551 mdname(conf->mddev),
2552 (unsigned long long)s,
2554 } else if (atomic_read(&rdev->read_errors)
2555 > conf->max_nr_stripes)
2556 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2557 mdname(conf->mddev), bdn);
2560 if (set_bad && test_bit(In_sync, &rdev->flags)
2561 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2564 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2565 set_bit(R5_ReadError, &sh->dev[i].flags);
2566 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2568 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2570 clear_bit(R5_ReadError, &sh->dev[i].flags);
2571 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2573 && test_bit(In_sync, &rdev->flags)
2574 && rdev_set_badblocks(
2575 rdev, sh->sector, STRIPE_SECTORS, 0)))
2576 md_error(conf->mddev, rdev);
2579 rdev_dec_pending(rdev, conf->mddev);
2581 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2582 set_bit(STRIPE_HANDLE, &sh->state);
2583 raid5_release_stripe(sh);
2586 static void raid5_end_write_request(struct bio *bi)
2588 struct stripe_head *sh = bi->bi_private;
2589 struct r5conf *conf = sh->raid_conf;
2590 int disks = sh->disks, i;
2591 struct md_rdev *uninitialized_var(rdev);
2594 int replacement = 0;
2596 for (i = 0 ; i < disks; i++) {
2597 if (bi == &sh->dev[i].req) {
2598 rdev = conf->disks[i].rdev;
2601 if (bi == &sh->dev[i].rreq) {
2602 rdev = conf->disks[i].replacement;
2606 /* rdev was removed and 'replacement'
2607 * replaced it. rdev is not removed
2608 * until all requests are finished.
2610 rdev = conf->disks[i].rdev;
2614 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2615 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2625 md_error(conf->mddev, rdev);
2626 else if (is_badblock(rdev, sh->sector,
2628 &first_bad, &bad_sectors))
2629 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2631 if (bi->bi_status) {
2632 set_bit(STRIPE_DEGRADED, &sh->state);
2633 set_bit(WriteErrorSeen, &rdev->flags);
2634 set_bit(R5_WriteError, &sh->dev[i].flags);
2635 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2636 set_bit(MD_RECOVERY_NEEDED,
2637 &rdev->mddev->recovery);
2638 } else if (is_badblock(rdev, sh->sector,
2640 &first_bad, &bad_sectors)) {
2641 set_bit(R5_MadeGood, &sh->dev[i].flags);
2642 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2643 /* That was a successful write so make
2644 * sure it looks like we already did
2647 set_bit(R5_ReWrite, &sh->dev[i].flags);
2650 rdev_dec_pending(rdev, conf->mddev);
2652 if (sh->batch_head && bi->bi_status && !replacement)
2653 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2656 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2657 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2658 set_bit(STRIPE_HANDLE, &sh->state);
2659 raid5_release_stripe(sh);
2661 if (sh->batch_head && sh != sh->batch_head)
2662 raid5_release_stripe(sh->batch_head);
2665 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2667 struct r5dev *dev = &sh->dev[i];
2670 dev->sector = raid5_compute_blocknr(sh, i, previous);
2673 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2675 char b[BDEVNAME_SIZE];
2676 struct r5conf *conf = mddev->private;
2677 unsigned long flags;
2678 pr_debug("raid456: error called\n");
2680 spin_lock_irqsave(&conf->device_lock, flags);
2681 clear_bit(In_sync, &rdev->flags);
2682 mddev->degraded = raid5_calc_degraded(conf);
2683 spin_unlock_irqrestore(&conf->device_lock, flags);
2684 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2686 set_bit(Blocked, &rdev->flags);
2687 set_bit(Faulty, &rdev->flags);
2688 set_mask_bits(&mddev->sb_flags, 0,
2689 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2690 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2691 "md/raid:%s: Operation continuing on %d devices.\n",
2693 bdevname(rdev->bdev, b),
2695 conf->raid_disks - mddev->degraded);
2696 r5c_update_on_rdev_error(mddev, rdev);
2700 * Input: a 'big' sector number,
2701 * Output: index of the data and parity disk, and the sector # in them.
2703 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2704 int previous, int *dd_idx,
2705 struct stripe_head *sh)
2707 sector_t stripe, stripe2;
2708 sector_t chunk_number;
2709 unsigned int chunk_offset;
2712 sector_t new_sector;
2713 int algorithm = previous ? conf->prev_algo
2715 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2716 : conf->chunk_sectors;
2717 int raid_disks = previous ? conf->previous_raid_disks
2719 int data_disks = raid_disks - conf->max_degraded;
2721 /* First compute the information on this sector */
2724 * Compute the chunk number and the sector offset inside the chunk
2726 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2727 chunk_number = r_sector;
2730 * Compute the stripe number
2732 stripe = chunk_number;
2733 *dd_idx = sector_div(stripe, data_disks);
2736 * Select the parity disk based on the user selected algorithm.
2738 pd_idx = qd_idx = -1;
2739 switch(conf->level) {
2741 pd_idx = data_disks;
2744 switch (algorithm) {
2745 case ALGORITHM_LEFT_ASYMMETRIC:
2746 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2747 if (*dd_idx >= pd_idx)
2750 case ALGORITHM_RIGHT_ASYMMETRIC:
2751 pd_idx = sector_div(stripe2, raid_disks);
2752 if (*dd_idx >= pd_idx)
2755 case ALGORITHM_LEFT_SYMMETRIC:
2756 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2757 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2759 case ALGORITHM_RIGHT_SYMMETRIC:
2760 pd_idx = sector_div(stripe2, raid_disks);
2761 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2763 case ALGORITHM_PARITY_0:
2767 case ALGORITHM_PARITY_N:
2768 pd_idx = data_disks;
2776 switch (algorithm) {
2777 case ALGORITHM_LEFT_ASYMMETRIC:
2778 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2779 qd_idx = pd_idx + 1;
2780 if (pd_idx == raid_disks-1) {
2781 (*dd_idx)++; /* Q D D D P */
2783 } else if (*dd_idx >= pd_idx)
2784 (*dd_idx) += 2; /* D D P Q D */
2786 case ALGORITHM_RIGHT_ASYMMETRIC:
2787 pd_idx = sector_div(stripe2, raid_disks);
2788 qd_idx = pd_idx + 1;
2789 if (pd_idx == raid_disks-1) {
2790 (*dd_idx)++; /* Q D D D P */
2792 } else if (*dd_idx >= pd_idx)
2793 (*dd_idx) += 2; /* D D P Q D */
2795 case ALGORITHM_LEFT_SYMMETRIC:
2796 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2797 qd_idx = (pd_idx + 1) % raid_disks;
2798 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2800 case ALGORITHM_RIGHT_SYMMETRIC:
2801 pd_idx = sector_div(stripe2, raid_disks);
2802 qd_idx = (pd_idx + 1) % raid_disks;
2803 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2806 case ALGORITHM_PARITY_0:
2811 case ALGORITHM_PARITY_N:
2812 pd_idx = data_disks;
2813 qd_idx = data_disks + 1;
2816 case ALGORITHM_ROTATING_ZERO_RESTART:
2817 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2818 * of blocks for computing Q is different.
2820 pd_idx = sector_div(stripe2, raid_disks);
2821 qd_idx = pd_idx + 1;
2822 if (pd_idx == raid_disks-1) {
2823 (*dd_idx)++; /* Q D D D P */
2825 } else if (*dd_idx >= pd_idx)
2826 (*dd_idx) += 2; /* D D P Q D */
2830 case ALGORITHM_ROTATING_N_RESTART:
2831 /* Same a left_asymmetric, by first stripe is
2832 * D D D P Q rather than
2836 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2837 qd_idx = pd_idx + 1;
2838 if (pd_idx == raid_disks-1) {
2839 (*dd_idx)++; /* Q D D D P */
2841 } else if (*dd_idx >= pd_idx)
2842 (*dd_idx) += 2; /* D D P Q D */
2846 case ALGORITHM_ROTATING_N_CONTINUE:
2847 /* Same as left_symmetric but Q is before P */
2848 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2849 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2850 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2854 case ALGORITHM_LEFT_ASYMMETRIC_6:
2855 /* RAID5 left_asymmetric, with Q on last device */
2856 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2857 if (*dd_idx >= pd_idx)
2859 qd_idx = raid_disks - 1;
2862 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2863 pd_idx = sector_div(stripe2, raid_disks-1);
2864 if (*dd_idx >= pd_idx)
2866 qd_idx = raid_disks - 1;
2869 case ALGORITHM_LEFT_SYMMETRIC_6:
2870 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2871 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2872 qd_idx = raid_disks - 1;
2875 case ALGORITHM_RIGHT_SYMMETRIC_6:
2876 pd_idx = sector_div(stripe2, raid_disks-1);
2877 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2878 qd_idx = raid_disks - 1;
2881 case ALGORITHM_PARITY_0_6:
2884 qd_idx = raid_disks - 1;
2894 sh->pd_idx = pd_idx;
2895 sh->qd_idx = qd_idx;
2896 sh->ddf_layout = ddf_layout;
2899 * Finally, compute the new sector number
2901 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2905 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2907 struct r5conf *conf = sh->raid_conf;
2908 int raid_disks = sh->disks;
2909 int data_disks = raid_disks - conf->max_degraded;
2910 sector_t new_sector = sh->sector, check;
2911 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2912 : conf->chunk_sectors;
2913 int algorithm = previous ? conf->prev_algo
2917 sector_t chunk_number;
2918 int dummy1, dd_idx = i;
2920 struct stripe_head sh2;
2922 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2923 stripe = new_sector;
2925 if (i == sh->pd_idx)
2927 switch(conf->level) {
2930 switch (algorithm) {
2931 case ALGORITHM_LEFT_ASYMMETRIC:
2932 case ALGORITHM_RIGHT_ASYMMETRIC:
2936 case ALGORITHM_LEFT_SYMMETRIC:
2937 case ALGORITHM_RIGHT_SYMMETRIC:
2940 i -= (sh->pd_idx + 1);
2942 case ALGORITHM_PARITY_0:
2945 case ALGORITHM_PARITY_N:
2952 if (i == sh->qd_idx)
2953 return 0; /* It is the Q disk */
2954 switch (algorithm) {
2955 case ALGORITHM_LEFT_ASYMMETRIC:
2956 case ALGORITHM_RIGHT_ASYMMETRIC:
2957 case ALGORITHM_ROTATING_ZERO_RESTART:
2958 case ALGORITHM_ROTATING_N_RESTART:
2959 if (sh->pd_idx == raid_disks-1)
2960 i--; /* Q D D D P */
2961 else if (i > sh->pd_idx)
2962 i -= 2; /* D D P Q D */
2964 case ALGORITHM_LEFT_SYMMETRIC:
2965 case ALGORITHM_RIGHT_SYMMETRIC:
2966 if (sh->pd_idx == raid_disks-1)
2967 i--; /* Q D D D P */
2972 i -= (sh->pd_idx + 2);
2975 case ALGORITHM_PARITY_0:
2978 case ALGORITHM_PARITY_N:
2980 case ALGORITHM_ROTATING_N_CONTINUE:
2981 /* Like left_symmetric, but P is before Q */
2982 if (sh->pd_idx == 0)
2983 i--; /* P D D D Q */
2988 i -= (sh->pd_idx + 1);
2991 case ALGORITHM_LEFT_ASYMMETRIC_6:
2992 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2996 case ALGORITHM_LEFT_SYMMETRIC_6:
2997 case ALGORITHM_RIGHT_SYMMETRIC_6:
2999 i += data_disks + 1;
3000 i -= (sh->pd_idx + 1);
3002 case ALGORITHM_PARITY_0_6:
3011 chunk_number = stripe * data_disks + i;
3012 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3014 check = raid5_compute_sector(conf, r_sector,
3015 previous, &dummy1, &sh2);
3016 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3017 || sh2.qd_idx != sh->qd_idx) {
3018 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3019 mdname(conf->mddev));
3026 * There are cases where we want handle_stripe_dirtying() and
3027 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3029 * This function checks whether we want to delay the towrite. Specifically,
3030 * we delay the towrite when:
3032 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3033 * stripe has data in journal (for other devices).
3035 * In this case, when reading data for the non-overwrite dev, it is
3036 * necessary to handle complex rmw of write back cache (prexor with
3037 * orig_page, and xor with page). To keep read path simple, we would
3038 * like to flush data in journal to RAID disks first, so complex rmw
3039 * is handled in the write patch (handle_stripe_dirtying).
3041 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3043 * It is important to be able to flush all stripes in raid5-cache.
3044 * Therefore, we need reserve some space on the journal device for
3045 * these flushes. If flush operation includes pending writes to the
3046 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3047 * for the flush out. If we exclude these pending writes from flush
3048 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3049 * Therefore, excluding pending writes in these cases enables more
3050 * efficient use of the journal device.
3052 * Note: To make sure the stripe makes progress, we only delay
3053 * towrite for stripes with data already in journal (injournal > 0).
3054 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3055 * no_space_stripes list.
3057 * 3. during journal failure
3058 * In journal failure, we try to flush all cached data to raid disks
3059 * based on data in stripe cache. The array is read-only to upper
3060 * layers, so we would skip all pending writes.
3063 static inline bool delay_towrite(struct r5conf *conf,
3065 struct stripe_head_state *s)
3068 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3069 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3072 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3076 if (s->log_failed && s->injournal)
3082 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3083 int rcw, int expand)
3085 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3086 struct r5conf *conf = sh->raid_conf;
3087 int level = conf->level;
3091 * In some cases, handle_stripe_dirtying initially decided to
3092 * run rmw and allocates extra page for prexor. However, rcw is
3093 * cheaper later on. We need to free the extra page now,
3094 * because we won't be able to do that in ops_complete_prexor().
3096 r5c_release_extra_page(sh);
3098 for (i = disks; i--; ) {
3099 struct r5dev *dev = &sh->dev[i];
3101 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3102 set_bit(R5_LOCKED, &dev->flags);
3103 set_bit(R5_Wantdrain, &dev->flags);
3105 clear_bit(R5_UPTODATE, &dev->flags);
3107 } else if (test_bit(R5_InJournal, &dev->flags)) {
3108 set_bit(R5_LOCKED, &dev->flags);
3112 /* if we are not expanding this is a proper write request, and
3113 * there will be bios with new data to be drained into the
3118 /* False alarm, nothing to do */
3120 sh->reconstruct_state = reconstruct_state_drain_run;
3121 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3123 sh->reconstruct_state = reconstruct_state_run;
3125 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3127 if (s->locked + conf->max_degraded == disks)
3128 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3129 atomic_inc(&conf->pending_full_writes);
3131 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3132 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3133 BUG_ON(level == 6 &&
3134 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3135 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3137 for (i = disks; i--; ) {
3138 struct r5dev *dev = &sh->dev[i];
3139 if (i == pd_idx || i == qd_idx)
3143 (test_bit(R5_UPTODATE, &dev->flags) ||
3144 test_bit(R5_Wantcompute, &dev->flags))) {
3145 set_bit(R5_Wantdrain, &dev->flags);
3146 set_bit(R5_LOCKED, &dev->flags);
3147 clear_bit(R5_UPTODATE, &dev->flags);
3149 } else if (test_bit(R5_InJournal, &dev->flags)) {
3150 set_bit(R5_LOCKED, &dev->flags);
3155 /* False alarm - nothing to do */
3157 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3158 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3159 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3160 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3163 /* keep the parity disk(s) locked while asynchronous operations
3166 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3167 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3171 int qd_idx = sh->qd_idx;
3172 struct r5dev *dev = &sh->dev[qd_idx];
3174 set_bit(R5_LOCKED, &dev->flags);
3175 clear_bit(R5_UPTODATE, &dev->flags);
3179 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3180 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3181 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3182 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3183 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3185 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3186 __func__, (unsigned long long)sh->sector,
3187 s->locked, s->ops_request);
3191 * Each stripe/dev can have one or more bion attached.
3192 * toread/towrite point to the first in a chain.
3193 * The bi_next chain must be in order.
3195 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3196 int forwrite, int previous)
3199 struct r5conf *conf = sh->raid_conf;
3202 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3203 (unsigned long long)bi->bi_iter.bi_sector,
3204 (unsigned long long)sh->sector);
3206 spin_lock_irq(&sh->stripe_lock);
3207 /* Don't allow new IO added to stripes in batch list */
3211 bip = &sh->dev[dd_idx].towrite;
3215 bip = &sh->dev[dd_idx].toread;
3216 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3217 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3219 bip = & (*bip)->bi_next;
3221 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3224 if (forwrite && raid5_has_ppl(conf)) {
3226 * With PPL only writes to consecutive data chunks within a
3227 * stripe are allowed because for a single stripe_head we can
3228 * only have one PPL entry at a time, which describes one data
3229 * range. Not really an overlap, but wait_for_overlap can be
3230 * used to handle this.
3238 for (i = 0; i < sh->disks; i++) {
3239 if (i != sh->pd_idx &&
3240 (i == dd_idx || sh->dev[i].towrite)) {
3241 sector = sh->dev[i].sector;
3242 if (count == 0 || sector < first)
3250 if (first + conf->chunk_sectors * (count - 1) != last)
3254 if (!forwrite || previous)
3255 clear_bit(STRIPE_BATCH_READY, &sh->state);
3257 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3261 bio_inc_remaining(bi);
3262 md_write_inc(conf->mddev, bi);
3265 /* check if page is covered */
3266 sector_t sector = sh->dev[dd_idx].sector;
3267 for (bi=sh->dev[dd_idx].towrite;
3268 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3269 bi && bi->bi_iter.bi_sector <= sector;
3270 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3271 if (bio_end_sector(bi) >= sector)
3272 sector = bio_end_sector(bi);
3274 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3275 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3276 sh->overwrite_disks++;
3279 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3280 (unsigned long long)(*bip)->bi_iter.bi_sector,
3281 (unsigned long long)sh->sector, dd_idx);
3283 if (conf->mddev->bitmap && firstwrite) {
3284 /* Cannot hold spinlock over bitmap_startwrite,
3285 * but must ensure this isn't added to a batch until
3286 * we have added to the bitmap and set bm_seq.
3287 * So set STRIPE_BITMAP_PENDING to prevent
3289 * If multiple add_stripe_bio() calls race here they
3290 * much all set STRIPE_BITMAP_PENDING. So only the first one
3291 * to complete "bitmap_startwrite" gets to set
3292 * STRIPE_BIT_DELAY. This is important as once a stripe
3293 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3296 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3297 spin_unlock_irq(&sh->stripe_lock);
3298 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3300 spin_lock_irq(&sh->stripe_lock);
3301 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3302 if (!sh->batch_head) {
3303 sh->bm_seq = conf->seq_flush+1;
3304 set_bit(STRIPE_BIT_DELAY, &sh->state);
3307 spin_unlock_irq(&sh->stripe_lock);
3309 if (stripe_can_batch(sh))
3310 stripe_add_to_batch_list(conf, sh);
3314 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3315 spin_unlock_irq(&sh->stripe_lock);
3319 static void end_reshape(struct r5conf *conf);
3321 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3322 struct stripe_head *sh)
3324 int sectors_per_chunk =
3325 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3327 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3328 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3330 raid5_compute_sector(conf,
3331 stripe * (disks - conf->max_degraded)
3332 *sectors_per_chunk + chunk_offset,
3338 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3339 struct stripe_head_state *s, int disks)
3342 BUG_ON(sh->batch_head);
3343 for (i = disks; i--; ) {
3347 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3348 struct md_rdev *rdev;
3350 rdev = rcu_dereference(conf->disks[i].rdev);
3351 if (rdev && test_bit(In_sync, &rdev->flags) &&
3352 !test_bit(Faulty, &rdev->flags))
3353 atomic_inc(&rdev->nr_pending);
3358 if (!rdev_set_badblocks(
3362 md_error(conf->mddev, rdev);
3363 rdev_dec_pending(rdev, conf->mddev);
3366 spin_lock_irq(&sh->stripe_lock);
3367 /* fail all writes first */
3368 bi = sh->dev[i].towrite;
3369 sh->dev[i].towrite = NULL;
3370 sh->overwrite_disks = 0;
3371 spin_unlock_irq(&sh->stripe_lock);
3375 log_stripe_write_finished(sh);
3377 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3378 wake_up(&conf->wait_for_overlap);
3380 while (bi && bi->bi_iter.bi_sector <
3381 sh->dev[i].sector + STRIPE_SECTORS) {
3382 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3384 md_write_end(conf->mddev);
3389 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3390 STRIPE_SECTORS, 0, 0);
3392 /* and fail all 'written' */
3393 bi = sh->dev[i].written;
3394 sh->dev[i].written = NULL;
3395 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3396 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3397 sh->dev[i].page = sh->dev[i].orig_page;
3400 if (bi) bitmap_end = 1;
3401 while (bi && bi->bi_iter.bi_sector <
3402 sh->dev[i].sector + STRIPE_SECTORS) {
3403 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3405 md_write_end(conf->mddev);
3410 /* fail any reads if this device is non-operational and
3411 * the data has not reached the cache yet.
3413 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3414 s->failed > conf->max_degraded &&
3415 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3416 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3417 spin_lock_irq(&sh->stripe_lock);
3418 bi = sh->dev[i].toread;
3419 sh->dev[i].toread = NULL;
3420 spin_unlock_irq(&sh->stripe_lock);
3421 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3422 wake_up(&conf->wait_for_overlap);
3425 while (bi && bi->bi_iter.bi_sector <
3426 sh->dev[i].sector + STRIPE_SECTORS) {
3427 struct bio *nextbi =
3428 r5_next_bio(bi, sh->dev[i].sector);
3435 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3436 STRIPE_SECTORS, 0, 0);
3437 /* If we were in the middle of a write the parity block might
3438 * still be locked - so just clear all R5_LOCKED flags
3440 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3445 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3446 if (atomic_dec_and_test(&conf->pending_full_writes))
3447 md_wakeup_thread(conf->mddev->thread);
3451 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3452 struct stripe_head_state *s)
3457 BUG_ON(sh->batch_head);
3458 clear_bit(STRIPE_SYNCING, &sh->state);
3459 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3460 wake_up(&conf->wait_for_overlap);
3463 /* There is nothing more to do for sync/check/repair.
3464 * Don't even need to abort as that is handled elsewhere
3465 * if needed, and not always wanted e.g. if there is a known
3467 * For recover/replace we need to record a bad block on all
3468 * non-sync devices, or abort the recovery
3470 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3471 /* During recovery devices cannot be removed, so
3472 * locking and refcounting of rdevs is not needed
3475 for (i = 0; i < conf->raid_disks; i++) {
3476 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3478 && !test_bit(Faulty, &rdev->flags)
3479 && !test_bit(In_sync, &rdev->flags)
3480 && !rdev_set_badblocks(rdev, sh->sector,
3483 rdev = rcu_dereference(conf->disks[i].replacement);
3485 && !test_bit(Faulty, &rdev->flags)
3486 && !test_bit(In_sync, &rdev->flags)
3487 && !rdev_set_badblocks(rdev, sh->sector,
3493 conf->recovery_disabled =
3494 conf->mddev->recovery_disabled;
3496 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3499 static int want_replace(struct stripe_head *sh, int disk_idx)
3501 struct md_rdev *rdev;
3505 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3507 && !test_bit(Faulty, &rdev->flags)
3508 && !test_bit(In_sync, &rdev->flags)
3509 && (rdev->recovery_offset <= sh->sector
3510 || rdev->mddev->recovery_cp <= sh->sector))
3516 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3517 int disk_idx, int disks)
3519 struct r5dev *dev = &sh->dev[disk_idx];
3520 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3521 &sh->dev[s->failed_num[1]] };
3525 if (test_bit(R5_LOCKED, &dev->flags) ||
3526 test_bit(R5_UPTODATE, &dev->flags))
3527 /* No point reading this as we already have it or have
3528 * decided to get it.
3533 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3534 /* We need this block to directly satisfy a request */
3537 if (s->syncing || s->expanding ||
3538 (s->replacing && want_replace(sh, disk_idx)))
3539 /* When syncing, or expanding we read everything.
3540 * When replacing, we need the replaced block.
3544 if ((s->failed >= 1 && fdev[0]->toread) ||
3545 (s->failed >= 2 && fdev[1]->toread))
3546 /* If we want to read from a failed device, then
3547 * we need to actually read every other device.
3551 /* Sometimes neither read-modify-write nor reconstruct-write
3552 * cycles can work. In those cases we read every block we
3553 * can. Then the parity-update is certain to have enough to
3555 * This can only be a problem when we need to write something,
3556 * and some device has failed. If either of those tests
3557 * fail we need look no further.
3559 if (!s->failed || !s->to_write)
3562 if (test_bit(R5_Insync, &dev->flags) &&
3563 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3564 /* Pre-reads at not permitted until after short delay
3565 * to gather multiple requests. However if this
3566 * device is no Insync, the block could only be computed
3567 * and there is no need to delay that.
3571 for (i = 0; i < s->failed && i < 2; i++) {
3572 if (fdev[i]->towrite &&
3573 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3574 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3575 /* If we have a partial write to a failed
3576 * device, then we will need to reconstruct
3577 * the content of that device, so all other
3578 * devices must be read.
3583 /* If we are forced to do a reconstruct-write, either because
3584 * the current RAID6 implementation only supports that, or
3585 * because parity cannot be trusted and we are currently
3586 * recovering it, there is extra need to be careful.
3587 * If one of the devices that we would need to read, because
3588 * it is not being overwritten (and maybe not written at all)
3589 * is missing/faulty, then we need to read everything we can.
3591 if (sh->raid_conf->level != 6 &&
3592 sh->sector < sh->raid_conf->mddev->recovery_cp)
3593 /* reconstruct-write isn't being forced */
3595 for (i = 0; i < s->failed && i < 2; i++) {
3596 if (s->failed_num[i] != sh->pd_idx &&
3597 s->failed_num[i] != sh->qd_idx &&
3598 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3599 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3606 /* fetch_block - checks the given member device to see if its data needs
3607 * to be read or computed to satisfy a request.
3609 * Returns 1 when no more member devices need to be checked, otherwise returns
3610 * 0 to tell the loop in handle_stripe_fill to continue
3612 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3613 int disk_idx, int disks)
3615 struct r5dev *dev = &sh->dev[disk_idx];
3617 /* is the data in this block needed, and can we get it? */
3618 if (need_this_block(sh, s, disk_idx, disks)) {
3619 /* we would like to get this block, possibly by computing it,
3620 * otherwise read it if the backing disk is insync
3622 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3623 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3624 BUG_ON(sh->batch_head);
3627 * In the raid6 case if the only non-uptodate disk is P
3628 * then we already trusted P to compute the other failed
3629 * drives. It is safe to compute rather than re-read P.
3630 * In other cases we only compute blocks from failed
3631 * devices, otherwise check/repair might fail to detect
3632 * a real inconsistency.
3635 if ((s->uptodate == disks - 1) &&
3636 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3637 (s->failed && (disk_idx == s->failed_num[0] ||
3638 disk_idx == s->failed_num[1])))) {
3639 /* have disk failed, and we're requested to fetch it;
3642 pr_debug("Computing stripe %llu block %d\n",
3643 (unsigned long long)sh->sector, disk_idx);
3644 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3645 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3646 set_bit(R5_Wantcompute, &dev->flags);
3647 sh->ops.target = disk_idx;
3648 sh->ops.target2 = -1; /* no 2nd target */
3650 /* Careful: from this point on 'uptodate' is in the eye
3651 * of raid_run_ops which services 'compute' operations
3652 * before writes. R5_Wantcompute flags a block that will
3653 * be R5_UPTODATE by the time it is needed for a
3654 * subsequent operation.
3658 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3659 /* Computing 2-failure is *very* expensive; only
3660 * do it if failed >= 2
3663 for (other = disks; other--; ) {
3664 if (other == disk_idx)
3666 if (!test_bit(R5_UPTODATE,
3667 &sh->dev[other].flags))
3671 pr_debug("Computing stripe %llu blocks %d,%d\n",
3672 (unsigned long long)sh->sector,
3674 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3675 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3676 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3677 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3678 sh->ops.target = disk_idx;
3679 sh->ops.target2 = other;
3683 } else if (test_bit(R5_Insync, &dev->flags)) {
3684 set_bit(R5_LOCKED, &dev->flags);
3685 set_bit(R5_Wantread, &dev->flags);
3687 pr_debug("Reading block %d (sync=%d)\n",
3688 disk_idx, s->syncing);
3696 * handle_stripe_fill - read or compute data to satisfy pending requests.
3698 static void handle_stripe_fill(struct stripe_head *sh,
3699 struct stripe_head_state *s,
3704 /* look for blocks to read/compute, skip this if a compute
3705 * is already in flight, or if the stripe contents are in the
3706 * midst of changing due to a write
3708 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3709 !sh->reconstruct_state) {
3712 * For degraded stripe with data in journal, do not handle
3713 * read requests yet, instead, flush the stripe to raid
3714 * disks first, this avoids handling complex rmw of write
3715 * back cache (prexor with orig_page, and then xor with
3716 * page) in the read path
3718 if (s->injournal && s->failed) {
3719 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3720 r5c_make_stripe_write_out(sh);
3724 for (i = disks; i--; )
3725 if (fetch_block(sh, s, i, disks))
3729 set_bit(STRIPE_HANDLE, &sh->state);
3732 static void break_stripe_batch_list(struct stripe_head *head_sh,
3733 unsigned long handle_flags);
3734 /* handle_stripe_clean_event
3735 * any written block on an uptodate or failed drive can be returned.
3736 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3737 * never LOCKED, so we don't need to test 'failed' directly.
3739 static void handle_stripe_clean_event(struct r5conf *conf,
3740 struct stripe_head *sh, int disks)
3744 int discard_pending = 0;
3745 struct stripe_head *head_sh = sh;
3746 bool do_endio = false;
3748 for (i = disks; i--; )
3749 if (sh->dev[i].written) {
3751 if (!test_bit(R5_LOCKED, &dev->flags) &&
3752 (test_bit(R5_UPTODATE, &dev->flags) ||
3753 test_bit(R5_Discard, &dev->flags) ||
3754 test_bit(R5_SkipCopy, &dev->flags))) {
3755 /* We can return any write requests */
3756 struct bio *wbi, *wbi2;
3757 pr_debug("Return write for disc %d\n", i);
3758 if (test_and_clear_bit(R5_Discard, &dev->flags))
3759 clear_bit(R5_UPTODATE, &dev->flags);
3760 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3761 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3766 dev->page = dev->orig_page;
3768 dev->written = NULL;
3769 while (wbi && wbi->bi_iter.bi_sector <
3770 dev->sector + STRIPE_SECTORS) {
3771 wbi2 = r5_next_bio(wbi, dev->sector);
3772 md_write_end(conf->mddev);
3776 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3778 !test_bit(STRIPE_DEGRADED, &sh->state),
3780 if (head_sh->batch_head) {
3781 sh = list_first_entry(&sh->batch_list,
3784 if (sh != head_sh) {
3791 } else if (test_bit(R5_Discard, &dev->flags))
3792 discard_pending = 1;
3795 log_stripe_write_finished(sh);
3797 if (!discard_pending &&
3798 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3800 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3801 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3802 if (sh->qd_idx >= 0) {
3803 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3804 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3806 /* now that discard is done we can proceed with any sync */
3807 clear_bit(STRIPE_DISCARD, &sh->state);
3809 * SCSI discard will change some bio fields and the stripe has
3810 * no updated data, so remove it from hash list and the stripe
3811 * will be reinitialized
3814 hash = sh->hash_lock_index;
3815 spin_lock_irq(conf->hash_locks + hash);
3817 spin_unlock_irq(conf->hash_locks + hash);
3818 if (head_sh->batch_head) {
3819 sh = list_first_entry(&sh->batch_list,
3820 struct stripe_head, batch_list);
3826 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3827 set_bit(STRIPE_HANDLE, &sh->state);
3831 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3832 if (atomic_dec_and_test(&conf->pending_full_writes))
3833 md_wakeup_thread(conf->mddev->thread);
3835 if (head_sh->batch_head && do_endio)
3836 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3840 * For RMW in write back cache, we need extra page in prexor to store the
3841 * old data. This page is stored in dev->orig_page.
3843 * This function checks whether we have data for prexor. The exact logic
3845 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3847 static inline bool uptodate_for_rmw(struct r5dev *dev)
3849 return (test_bit(R5_UPTODATE, &dev->flags)) &&
3850 (!test_bit(R5_InJournal, &dev->flags) ||
3851 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3854 static int handle_stripe_dirtying(struct r5conf *conf,
3855 struct stripe_head *sh,
3856 struct stripe_head_state *s,
3859 int rmw = 0, rcw = 0, i;
3860 sector_t recovery_cp = conf->mddev->recovery_cp;
3862 /* Check whether resync is now happening or should start.
3863 * If yes, then the array is dirty (after unclean shutdown or
3864 * initial creation), so parity in some stripes might be inconsistent.
3865 * In this case, we need to always do reconstruct-write, to ensure
3866 * that in case of drive failure or read-error correction, we
3867 * generate correct data from the parity.
3869 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3870 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3872 /* Calculate the real rcw later - for now make it
3873 * look like rcw is cheaper
3876 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3877 conf->rmw_level, (unsigned long long)recovery_cp,
3878 (unsigned long long)sh->sector);
3879 } else for (i = disks; i--; ) {
3880 /* would I have to read this buffer for read_modify_write */
3881 struct r5dev *dev = &sh->dev[i];
3882 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3883 i == sh->pd_idx || i == sh->qd_idx ||
3884 test_bit(R5_InJournal, &dev->flags)) &&
3885 !test_bit(R5_LOCKED, &dev->flags) &&
3886 !(uptodate_for_rmw(dev) ||
3887 test_bit(R5_Wantcompute, &dev->flags))) {
3888 if (test_bit(R5_Insync, &dev->flags))
3891 rmw += 2*disks; /* cannot read it */
3893 /* Would I have to read this buffer for reconstruct_write */
3894 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3895 i != sh->pd_idx && i != sh->qd_idx &&
3896 !test_bit(R5_LOCKED, &dev->flags) &&
3897 !(test_bit(R5_UPTODATE, &dev->flags) ||
3898 test_bit(R5_Wantcompute, &dev->flags))) {
3899 if (test_bit(R5_Insync, &dev->flags))
3906 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3907 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3908 set_bit(STRIPE_HANDLE, &sh->state);
3909 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3910 /* prefer read-modify-write, but need to get some data */
3911 if (conf->mddev->queue)
3912 blk_add_trace_msg(conf->mddev->queue,
3913 "raid5 rmw %llu %d",
3914 (unsigned long long)sh->sector, rmw);
3915 for (i = disks; i--; ) {
3916 struct r5dev *dev = &sh->dev[i];
3917 if (test_bit(R5_InJournal, &dev->flags) &&
3918 dev->page == dev->orig_page &&
3919 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3920 /* alloc page for prexor */
3921 struct page *p = alloc_page(GFP_NOIO);
3929 * alloc_page() failed, try use
3930 * disk_info->extra_page
3932 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3933 &conf->cache_state)) {
3934 r5c_use_extra_page(sh);
3938 /* extra_page in use, add to delayed_list */
3939 set_bit(STRIPE_DELAYED, &sh->state);
3940 s->waiting_extra_page = 1;
3945 for (i = disks; i--; ) {
3946 struct r5dev *dev = &sh->dev[i];
3947 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3948 i == sh->pd_idx || i == sh->qd_idx ||
3949 test_bit(R5_InJournal, &dev->flags)) &&
3950 !test_bit(R5_LOCKED, &dev->flags) &&
3951 !(uptodate_for_rmw(dev) ||
3952 test_bit(R5_Wantcompute, &dev->flags)) &&
3953 test_bit(R5_Insync, &dev->flags)) {
3954 if (test_bit(STRIPE_PREREAD_ACTIVE,
3956 pr_debug("Read_old block %d for r-m-w\n",
3958 set_bit(R5_LOCKED, &dev->flags);
3959 set_bit(R5_Wantread, &dev->flags);
3962 set_bit(STRIPE_DELAYED, &sh->state);
3963 set_bit(STRIPE_HANDLE, &sh->state);
3968 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3969 /* want reconstruct write, but need to get some data */
3972 for (i = disks; i--; ) {
3973 struct r5dev *dev = &sh->dev[i];
3974 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3975 i != sh->pd_idx && i != sh->qd_idx &&
3976 !test_bit(R5_LOCKED, &dev->flags) &&
3977 !(test_bit(R5_UPTODATE, &dev->flags) ||
3978 test_bit(R5_Wantcompute, &dev->flags))) {
3980 if (test_bit(R5_Insync, &dev->flags) &&
3981 test_bit(STRIPE_PREREAD_ACTIVE,
3983 pr_debug("Read_old block "
3984 "%d for Reconstruct\n", i);
3985 set_bit(R5_LOCKED, &dev->flags);
3986 set_bit(R5_Wantread, &dev->flags);
3990 set_bit(STRIPE_DELAYED, &sh->state);
3991 set_bit(STRIPE_HANDLE, &sh->state);
3995 if (rcw && conf->mddev->queue)
3996 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3997 (unsigned long long)sh->sector,
3998 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4001 if (rcw > disks && rmw > disks &&
4002 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4003 set_bit(STRIPE_DELAYED, &sh->state);
4005 /* now if nothing is locked, and if we have enough data,
4006 * we can start a write request
4008 /* since handle_stripe can be called at any time we need to handle the
4009 * case where a compute block operation has been submitted and then a
4010 * subsequent call wants to start a write request. raid_run_ops only
4011 * handles the case where compute block and reconstruct are requested
4012 * simultaneously. If this is not the case then new writes need to be
4013 * held off until the compute completes.
4015 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4016 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4017 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4018 schedule_reconstruction(sh, s, rcw == 0, 0);
4022 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4023 struct stripe_head_state *s, int disks)
4025 struct r5dev *dev = NULL;
4027 BUG_ON(sh->batch_head);
4028 set_bit(STRIPE_HANDLE, &sh->state);
4030 switch (sh->check_state) {
4031 case check_state_idle:
4032 /* start a new check operation if there are no failures */
4033 if (s->failed == 0) {
4034 BUG_ON(s->uptodate != disks);
4035 sh->check_state = check_state_run;
4036 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4037 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4041 dev = &sh->dev[s->failed_num[0]];
4043 case check_state_compute_result:
4044 sh->check_state = check_state_idle;
4046 dev = &sh->dev[sh->pd_idx];
4048 /* check that a write has not made the stripe insync */
4049 if (test_bit(STRIPE_INSYNC, &sh->state))
4052 /* either failed parity check, or recovery is happening */
4053 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4054 BUG_ON(s->uptodate != disks);
4056 set_bit(R5_LOCKED, &dev->flags);
4058 set_bit(R5_Wantwrite, &dev->flags);
4060 clear_bit(STRIPE_DEGRADED, &sh->state);
4061 set_bit(STRIPE_INSYNC, &sh->state);
4063 case check_state_run:
4064 break; /* we will be called again upon completion */
4065 case check_state_check_result:
4066 sh->check_state = check_state_idle;
4068 /* if a failure occurred during the check operation, leave
4069 * STRIPE_INSYNC not set and let the stripe be handled again
4074 /* handle a successful check operation, if parity is correct
4075 * we are done. Otherwise update the mismatch count and repair
4076 * parity if !MD_RECOVERY_CHECK
4078 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4079 /* parity is correct (on disc,
4080 * not in buffer any more)
4082 set_bit(STRIPE_INSYNC, &sh->state);
4084 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4085 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4086 /* don't try to repair!! */
4087 set_bit(STRIPE_INSYNC, &sh->state);
4088 pr_warn_ratelimited("%s: mismatch sector in range "
4089 "%llu-%llu\n", mdname(conf->mddev),
4090 (unsigned long long) sh->sector,
4091 (unsigned long long) sh->sector +
4094 sh->check_state = check_state_compute_run;
4095 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4096 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4097 set_bit(R5_Wantcompute,
4098 &sh->dev[sh->pd_idx].flags);
4099 sh->ops.target = sh->pd_idx;
4100 sh->ops.target2 = -1;
4105 case check_state_compute_run:
4108 pr_err("%s: unknown check_state: %d sector: %llu\n",
4109 __func__, sh->check_state,
4110 (unsigned long long) sh->sector);
4115 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4116 struct stripe_head_state *s,
4119 int pd_idx = sh->pd_idx;
4120 int qd_idx = sh->qd_idx;
4123 BUG_ON(sh->batch_head);
4124 set_bit(STRIPE_HANDLE, &sh->state);
4126 BUG_ON(s->failed > 2);
4128 /* Want to check and possibly repair P and Q.
4129 * However there could be one 'failed' device, in which
4130 * case we can only check one of them, possibly using the
4131 * other to generate missing data
4134 switch (sh->check_state) {
4135 case check_state_idle:
4136 /* start a new check operation if there are < 2 failures */
4137 if (s->failed == s->q_failed) {
4138 /* The only possible failed device holds Q, so it
4139 * makes sense to check P (If anything else were failed,
4140 * we would have used P to recreate it).
4142 sh->check_state = check_state_run;
4144 if (!s->q_failed && s->failed < 2) {
4145 /* Q is not failed, and we didn't use it to generate
4146 * anything, so it makes sense to check it
4148 if (sh->check_state == check_state_run)
4149 sh->check_state = check_state_run_pq;
4151 sh->check_state = check_state_run_q;
4154 /* discard potentially stale zero_sum_result */
4155 sh->ops.zero_sum_result = 0;
4157 if (sh->check_state == check_state_run) {
4158 /* async_xor_zero_sum destroys the contents of P */
4159 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4162 if (sh->check_state >= check_state_run &&
4163 sh->check_state <= check_state_run_pq) {
4164 /* async_syndrome_zero_sum preserves P and Q, so
4165 * no need to mark them !uptodate here
4167 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4171 /* we have 2-disk failure */
4172 BUG_ON(s->failed != 2);
4174 case check_state_compute_result:
4175 sh->check_state = check_state_idle;
4177 /* check that a write has not made the stripe insync */
4178 if (test_bit(STRIPE_INSYNC, &sh->state))
4181 /* now write out any block on a failed drive,
4182 * or P or Q if they were recomputed
4184 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
4185 if (s->failed == 2) {
4186 dev = &sh->dev[s->failed_num[1]];
4188 set_bit(R5_LOCKED, &dev->flags);
4189 set_bit(R5_Wantwrite, &dev->flags);
4191 if (s->failed >= 1) {
4192 dev = &sh->dev[s->failed_num[0]];
4194 set_bit(R5_LOCKED, &dev->flags);
4195 set_bit(R5_Wantwrite, &dev->flags);
4197 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4198 dev = &sh->dev[pd_idx];
4200 set_bit(R5_LOCKED, &dev->flags);
4201 set_bit(R5_Wantwrite, &dev->flags);
4203 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4204 dev = &sh->dev[qd_idx];
4206 set_bit(R5_LOCKED, &dev->flags);
4207 set_bit(R5_Wantwrite, &dev->flags);
4209 clear_bit(STRIPE_DEGRADED, &sh->state);
4211 set_bit(STRIPE_INSYNC, &sh->state);
4213 case check_state_run:
4214 case check_state_run_q:
4215 case check_state_run_pq:
4216 break; /* we will be called again upon completion */
4217 case check_state_check_result:
4218 sh->check_state = check_state_idle;
4220 /* handle a successful check operation, if parity is correct
4221 * we are done. Otherwise update the mismatch count and repair
4222 * parity if !MD_RECOVERY_CHECK
4224 if (sh->ops.zero_sum_result == 0) {
4225 /* both parities are correct */
4227 set_bit(STRIPE_INSYNC, &sh->state);
4229 /* in contrast to the raid5 case we can validate
4230 * parity, but still have a failure to write
4233 sh->check_state = check_state_compute_result;
4234 /* Returning at this point means that we may go
4235 * off and bring p and/or q uptodate again so
4236 * we make sure to check zero_sum_result again
4237 * to verify if p or q need writeback
4241 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4242 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4243 /* don't try to repair!! */
4244 set_bit(STRIPE_INSYNC, &sh->state);
4245 pr_warn_ratelimited("%s: mismatch sector in range "
4246 "%llu-%llu\n", mdname(conf->mddev),
4247 (unsigned long long) sh->sector,
4248 (unsigned long long) sh->sector +
4251 int *target = &sh->ops.target;
4253 sh->ops.target = -1;
4254 sh->ops.target2 = -1;
4255 sh->check_state = check_state_compute_run;
4256 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4257 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4258 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4259 set_bit(R5_Wantcompute,
4260 &sh->dev[pd_idx].flags);
4262 target = &sh->ops.target2;
4265 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4266 set_bit(R5_Wantcompute,
4267 &sh->dev[qd_idx].flags);
4274 case check_state_compute_run:
4277 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4278 __func__, sh->check_state,
4279 (unsigned long long) sh->sector);
4284 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4288 /* We have read all the blocks in this stripe and now we need to
4289 * copy some of them into a target stripe for expand.
4291 struct dma_async_tx_descriptor *tx = NULL;
4292 BUG_ON(sh->batch_head);
4293 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4294 for (i = 0; i < sh->disks; i++)
4295 if (i != sh->pd_idx && i != sh->qd_idx) {
4297 struct stripe_head *sh2;
4298 struct async_submit_ctl submit;
4300 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4301 sector_t s = raid5_compute_sector(conf, bn, 0,
4303 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4305 /* so far only the early blocks of this stripe
4306 * have been requested. When later blocks
4307 * get requested, we will try again
4310 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4311 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4312 /* must have already done this block */
4313 raid5_release_stripe(sh2);
4317 /* place all the copies on one channel */
4318 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4319 tx = async_memcpy(sh2->dev[dd_idx].page,
4320 sh->dev[i].page, 0, 0, STRIPE_SIZE,
4323 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4324 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4325 for (j = 0; j < conf->raid_disks; j++)
4326 if (j != sh2->pd_idx &&
4328 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4330 if (j == conf->raid_disks) {
4331 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4332 set_bit(STRIPE_HANDLE, &sh2->state);
4334 raid5_release_stripe(sh2);
4337 /* done submitting copies, wait for them to complete */
4338 async_tx_quiesce(&tx);
4342 * handle_stripe - do things to a stripe.
4344 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4345 * state of various bits to see what needs to be done.
4347 * return some read requests which now have data
4348 * return some write requests which are safely on storage
4349 * schedule a read on some buffers
4350 * schedule a write of some buffers
4351 * return confirmation of parity correctness
4355 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4357 struct r5conf *conf = sh->raid_conf;
4358 int disks = sh->disks;
4361 int do_recovery = 0;
4363 memset(s, 0, sizeof(*s));
4365 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4366 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4367 s->failed_num[0] = -1;
4368 s->failed_num[1] = -1;
4369 s->log_failed = r5l_log_disk_error(conf);
4371 /* Now to look around and see what can be done */
4373 for (i=disks; i--; ) {
4374 struct md_rdev *rdev;
4381 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4383 dev->toread, dev->towrite, dev->written);
4384 /* maybe we can reply to a read
4386 * new wantfill requests are only permitted while
4387 * ops_complete_biofill is guaranteed to be inactive
4389 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4390 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4391 set_bit(R5_Wantfill, &dev->flags);
4393 /* now count some things */
4394 if (test_bit(R5_LOCKED, &dev->flags))
4396 if (test_bit(R5_UPTODATE, &dev->flags))
4398 if (test_bit(R5_Wantcompute, &dev->flags)) {
4400 BUG_ON(s->compute > 2);
4403 if (test_bit(R5_Wantfill, &dev->flags))
4405 else if (dev->toread)
4409 if (!test_bit(R5_OVERWRITE, &dev->flags))
4414 /* Prefer to use the replacement for reads, but only
4415 * if it is recovered enough and has no bad blocks.
4417 rdev = rcu_dereference(conf->disks[i].replacement);
4418 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4419 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4420 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4421 &first_bad, &bad_sectors))
4422 set_bit(R5_ReadRepl, &dev->flags);
4424 if (rdev && !test_bit(Faulty, &rdev->flags))
4425 set_bit(R5_NeedReplace, &dev->flags);
4427 clear_bit(R5_NeedReplace, &dev->flags);
4428 rdev = rcu_dereference(conf->disks[i].rdev);
4429 clear_bit(R5_ReadRepl, &dev->flags);
4431 if (rdev && test_bit(Faulty, &rdev->flags))
4434 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4435 &first_bad, &bad_sectors);
4436 if (s->blocked_rdev == NULL
4437 && (test_bit(Blocked, &rdev->flags)
4440 set_bit(BlockedBadBlocks,
4442 s->blocked_rdev = rdev;
4443 atomic_inc(&rdev->nr_pending);
4446 clear_bit(R5_Insync, &dev->flags);
4450 /* also not in-sync */
4451 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4452 test_bit(R5_UPTODATE, &dev->flags)) {
4453 /* treat as in-sync, but with a read error
4454 * which we can now try to correct
4456 set_bit(R5_Insync, &dev->flags);
4457 set_bit(R5_ReadError, &dev->flags);
4459 } else if (test_bit(In_sync, &rdev->flags))
4460 set_bit(R5_Insync, &dev->flags);
4461 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4462 /* in sync if before recovery_offset */
4463 set_bit(R5_Insync, &dev->flags);
4464 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4465 test_bit(R5_Expanded, &dev->flags))
4466 /* If we've reshaped into here, we assume it is Insync.
4467 * We will shortly update recovery_offset to make
4470 set_bit(R5_Insync, &dev->flags);
4472 if (test_bit(R5_WriteError, &dev->flags)) {
4473 /* This flag does not apply to '.replacement'
4474 * only to .rdev, so make sure to check that*/
4475 struct md_rdev *rdev2 = rcu_dereference(
4476 conf->disks[i].rdev);
4478 clear_bit(R5_Insync, &dev->flags);
4479 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4480 s->handle_bad_blocks = 1;
4481 atomic_inc(&rdev2->nr_pending);
4483 clear_bit(R5_WriteError, &dev->flags);
4485 if (test_bit(R5_MadeGood, &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);
4490 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4491 s->handle_bad_blocks = 1;
4492 atomic_inc(&rdev2->nr_pending);
4494 clear_bit(R5_MadeGood, &dev->flags);
4496 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4497 struct md_rdev *rdev2 = rcu_dereference(
4498 conf->disks[i].replacement);
4499 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4500 s->handle_bad_blocks = 1;
4501 atomic_inc(&rdev2->nr_pending);
4503 clear_bit(R5_MadeGoodRepl, &dev->flags);
4505 if (!test_bit(R5_Insync, &dev->flags)) {
4506 /* The ReadError flag will just be confusing now */
4507 clear_bit(R5_ReadError, &dev->flags);
4508 clear_bit(R5_ReWrite, &dev->flags);
4510 if (test_bit(R5_ReadError, &dev->flags))
4511 clear_bit(R5_Insync, &dev->flags);
4512 if (!test_bit(R5_Insync, &dev->flags)) {
4514 s->failed_num[s->failed] = i;
4516 if (rdev && !test_bit(Faulty, &rdev->flags))
4520 if (test_bit(R5_InJournal, &dev->flags))
4522 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4525 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4526 /* If there is a failed device being replaced,
4527 * we must be recovering.
4528 * else if we are after recovery_cp, we must be syncing
4529 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4530 * else we can only be replacing
4531 * sync and recovery both need to read all devices, and so
4532 * use the same flag.
4535 sh->sector >= conf->mddev->recovery_cp ||
4536 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4544 static int clear_batch_ready(struct stripe_head *sh)
4546 /* Return '1' if this is a member of batch, or
4547 * '0' if it is a lone stripe or a head which can now be
4550 struct stripe_head *tmp;
4551 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4552 return (sh->batch_head && sh->batch_head != sh);
4553 spin_lock(&sh->stripe_lock);
4554 if (!sh->batch_head) {
4555 spin_unlock(&sh->stripe_lock);
4560 * this stripe could be added to a batch list before we check
4561 * BATCH_READY, skips it
4563 if (sh->batch_head != sh) {
4564 spin_unlock(&sh->stripe_lock);
4567 spin_lock(&sh->batch_lock);
4568 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4569 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4570 spin_unlock(&sh->batch_lock);
4571 spin_unlock(&sh->stripe_lock);
4574 * BATCH_READY is cleared, no new stripes can be added.
4575 * batch_list can be accessed without lock
4580 static void break_stripe_batch_list(struct stripe_head *head_sh,
4581 unsigned long handle_flags)
4583 struct stripe_head *sh, *next;
4587 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4589 list_del_init(&sh->batch_list);
4591 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4592 (1 << STRIPE_SYNCING) |
4593 (1 << STRIPE_REPLACED) |
4594 (1 << STRIPE_DELAYED) |
4595 (1 << STRIPE_BIT_DELAY) |
4596 (1 << STRIPE_FULL_WRITE) |
4597 (1 << STRIPE_BIOFILL_RUN) |
4598 (1 << STRIPE_COMPUTE_RUN) |
4599 (1 << STRIPE_OPS_REQ_PENDING) |
4600 (1 << STRIPE_DISCARD) |
4601 (1 << STRIPE_BATCH_READY) |
4602 (1 << STRIPE_BATCH_ERR) |
4603 (1 << STRIPE_BITMAP_PENDING)),
4604 "stripe state: %lx\n", sh->state);
4605 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4606 (1 << STRIPE_REPLACED)),
4607 "head stripe state: %lx\n", head_sh->state);
4609 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4610 (1 << STRIPE_PREREAD_ACTIVE) |
4611 (1 << STRIPE_DEGRADED)),
4612 head_sh->state & (1 << STRIPE_INSYNC));
4614 sh->check_state = head_sh->check_state;
4615 sh->reconstruct_state = head_sh->reconstruct_state;
4616 for (i = 0; i < sh->disks; i++) {
4617 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4619 sh->dev[i].flags = head_sh->dev[i].flags &
4620 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4622 spin_lock_irq(&sh->stripe_lock);
4623 sh->batch_head = NULL;
4624 spin_unlock_irq(&sh->stripe_lock);
4625 if (handle_flags == 0 ||
4626 sh->state & handle_flags)
4627 set_bit(STRIPE_HANDLE, &sh->state);
4628 raid5_release_stripe(sh);
4630 spin_lock_irq(&head_sh->stripe_lock);
4631 head_sh->batch_head = NULL;
4632 spin_unlock_irq(&head_sh->stripe_lock);
4633 for (i = 0; i < head_sh->disks; i++)
4634 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4636 if (head_sh->state & handle_flags)
4637 set_bit(STRIPE_HANDLE, &head_sh->state);
4640 wake_up(&head_sh->raid_conf->wait_for_overlap);
4643 static void handle_stripe(struct stripe_head *sh)
4645 struct stripe_head_state s;
4646 struct r5conf *conf = sh->raid_conf;
4649 int disks = sh->disks;
4650 struct r5dev *pdev, *qdev;
4652 clear_bit(STRIPE_HANDLE, &sh->state);
4653 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4654 /* already being handled, ensure it gets handled
4655 * again when current action finishes */
4656 set_bit(STRIPE_HANDLE, &sh->state);
4660 if (clear_batch_ready(sh) ) {
4661 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4665 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4666 break_stripe_batch_list(sh, 0);
4668 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4669 spin_lock(&sh->stripe_lock);
4671 * Cannot process 'sync' concurrently with 'discard'.
4672 * Flush data in r5cache before 'sync'.
4674 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4675 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4676 !test_bit(STRIPE_DISCARD, &sh->state) &&
4677 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4678 set_bit(STRIPE_SYNCING, &sh->state);
4679 clear_bit(STRIPE_INSYNC, &sh->state);
4680 clear_bit(STRIPE_REPLACED, &sh->state);
4682 spin_unlock(&sh->stripe_lock);
4684 clear_bit(STRIPE_DELAYED, &sh->state);
4686 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4687 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4688 (unsigned long long)sh->sector, sh->state,
4689 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4690 sh->check_state, sh->reconstruct_state);
4692 analyse_stripe(sh, &s);
4694 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4697 if (s.handle_bad_blocks ||
4698 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4699 set_bit(STRIPE_HANDLE, &sh->state);
4703 if (unlikely(s.blocked_rdev)) {
4704 if (s.syncing || s.expanding || s.expanded ||
4705 s.replacing || s.to_write || s.written) {
4706 set_bit(STRIPE_HANDLE, &sh->state);
4709 /* There is nothing for the blocked_rdev to block */
4710 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4711 s.blocked_rdev = NULL;
4714 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4715 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4716 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4719 pr_debug("locked=%d uptodate=%d to_read=%d"
4720 " to_write=%d failed=%d failed_num=%d,%d\n",
4721 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4722 s.failed_num[0], s.failed_num[1]);
4724 * check if the array has lost more than max_degraded devices and,
4725 * if so, some requests might need to be failed.
4727 * When journal device failed (log_failed), we will only process
4728 * the stripe if there is data need write to raid disks
4730 if (s.failed > conf->max_degraded ||
4731 (s.log_failed && s.injournal == 0)) {
4732 sh->check_state = 0;
4733 sh->reconstruct_state = 0;
4734 break_stripe_batch_list(sh, 0);
4735 if (s.to_read+s.to_write+s.written)
4736 handle_failed_stripe(conf, sh, &s, disks);
4737 if (s.syncing + s.replacing)
4738 handle_failed_sync(conf, sh, &s);
4741 /* Now we check to see if any write operations have recently
4745 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4747 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4748 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4749 sh->reconstruct_state = reconstruct_state_idle;
4751 /* All the 'written' buffers and the parity block are ready to
4752 * be written back to disk
4754 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4755 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4756 BUG_ON(sh->qd_idx >= 0 &&
4757 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4758 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4759 for (i = disks; i--; ) {
4760 struct r5dev *dev = &sh->dev[i];
4761 if (test_bit(R5_LOCKED, &dev->flags) &&
4762 (i == sh->pd_idx || i == sh->qd_idx ||
4763 dev->written || test_bit(R5_InJournal,
4765 pr_debug("Writing block %d\n", i);
4766 set_bit(R5_Wantwrite, &dev->flags);
4771 if (!test_bit(R5_Insync, &dev->flags) ||
4772 ((i == sh->pd_idx || i == sh->qd_idx) &&
4774 set_bit(STRIPE_INSYNC, &sh->state);
4777 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4778 s.dec_preread_active = 1;
4782 * might be able to return some write requests if the parity blocks
4783 * are safe, or on a failed drive
4785 pdev = &sh->dev[sh->pd_idx];
4786 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4787 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4788 qdev = &sh->dev[sh->qd_idx];
4789 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4790 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4794 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4795 && !test_bit(R5_LOCKED, &pdev->flags)
4796 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4797 test_bit(R5_Discard, &pdev->flags))))) &&
4798 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4799 && !test_bit(R5_LOCKED, &qdev->flags)
4800 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4801 test_bit(R5_Discard, &qdev->flags))))))
4802 handle_stripe_clean_event(conf, sh, disks);
4805 r5c_handle_cached_data_endio(conf, sh, disks);
4806 log_stripe_write_finished(sh);
4808 /* Now we might consider reading some blocks, either to check/generate
4809 * parity, or to satisfy requests
4810 * or to load a block that is being partially written.
4812 if (s.to_read || s.non_overwrite
4813 || (conf->level == 6 && s.to_write && s.failed)
4814 || (s.syncing && (s.uptodate + s.compute < disks))
4817 handle_stripe_fill(sh, &s, disks);
4820 * When the stripe finishes full journal write cycle (write to journal
4821 * and raid disk), this is the clean up procedure so it is ready for
4824 r5c_finish_stripe_write_out(conf, sh, &s);
4827 * Now to consider new write requests, cache write back and what else,
4828 * if anything should be read. We do not handle new writes when:
4829 * 1/ A 'write' operation (copy+xor) is already in flight.
4830 * 2/ A 'check' operation is in flight, as it may clobber the parity
4832 * 3/ A r5c cache log write is in flight.
4835 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4836 if (!r5c_is_writeback(conf->log)) {
4838 handle_stripe_dirtying(conf, sh, &s, disks);
4839 } else { /* write back cache */
4842 /* First, try handle writes in caching phase */
4844 ret = r5c_try_caching_write(conf, sh, &s,
4847 * If caching phase failed: ret == -EAGAIN
4849 * stripe under reclaim: !caching && injournal
4851 * fall back to handle_stripe_dirtying()
4853 if (ret == -EAGAIN ||
4854 /* stripe under reclaim: !caching && injournal */
4855 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4857 ret = handle_stripe_dirtying(conf, sh, &s,
4865 /* maybe we need to check and possibly fix the parity for this stripe
4866 * Any reads will already have been scheduled, so we just see if enough
4867 * data is available. The parity check is held off while parity
4868 * dependent operations are in flight.
4870 if (sh->check_state ||
4871 (s.syncing && s.locked == 0 &&
4872 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4873 !test_bit(STRIPE_INSYNC, &sh->state))) {
4874 if (conf->level == 6)
4875 handle_parity_checks6(conf, sh, &s, disks);
4877 handle_parity_checks5(conf, sh, &s, disks);
4880 if ((s.replacing || s.syncing) && s.locked == 0
4881 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4882 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4883 /* Write out to replacement devices where possible */
4884 for (i = 0; i < conf->raid_disks; i++)
4885 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4886 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4887 set_bit(R5_WantReplace, &sh->dev[i].flags);
4888 set_bit(R5_LOCKED, &sh->dev[i].flags);
4892 set_bit(STRIPE_INSYNC, &sh->state);
4893 set_bit(STRIPE_REPLACED, &sh->state);
4895 if ((s.syncing || s.replacing) && s.locked == 0 &&
4896 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4897 test_bit(STRIPE_INSYNC, &sh->state)) {
4898 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4899 clear_bit(STRIPE_SYNCING, &sh->state);
4900 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4901 wake_up(&conf->wait_for_overlap);
4904 /* If the failed drives are just a ReadError, then we might need
4905 * to progress the repair/check process
4907 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4908 for (i = 0; i < s.failed; i++) {
4909 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4910 if (test_bit(R5_ReadError, &dev->flags)
4911 && !test_bit(R5_LOCKED, &dev->flags)
4912 && test_bit(R5_UPTODATE, &dev->flags)
4914 if (!test_bit(R5_ReWrite, &dev->flags)) {
4915 set_bit(R5_Wantwrite, &dev->flags);
4916 set_bit(R5_ReWrite, &dev->flags);
4917 set_bit(R5_LOCKED, &dev->flags);
4920 /* let's read it back */
4921 set_bit(R5_Wantread, &dev->flags);
4922 set_bit(R5_LOCKED, &dev->flags);
4928 /* Finish reconstruct operations initiated by the expansion process */
4929 if (sh->reconstruct_state == reconstruct_state_result) {
4930 struct stripe_head *sh_src
4931 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4932 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4933 /* sh cannot be written until sh_src has been read.
4934 * so arrange for sh to be delayed a little
4936 set_bit(STRIPE_DELAYED, &sh->state);
4937 set_bit(STRIPE_HANDLE, &sh->state);
4938 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4940 atomic_inc(&conf->preread_active_stripes);
4941 raid5_release_stripe(sh_src);
4945 raid5_release_stripe(sh_src);
4947 sh->reconstruct_state = reconstruct_state_idle;
4948 clear_bit(STRIPE_EXPANDING, &sh->state);
4949 for (i = conf->raid_disks; i--; ) {
4950 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4951 set_bit(R5_LOCKED, &sh->dev[i].flags);
4956 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4957 !sh->reconstruct_state) {
4958 /* Need to write out all blocks after computing parity */
4959 sh->disks = conf->raid_disks;
4960 stripe_set_idx(sh->sector, conf, 0, sh);
4961 schedule_reconstruction(sh, &s, 1, 1);
4962 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4963 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4964 atomic_dec(&conf->reshape_stripes);
4965 wake_up(&conf->wait_for_overlap);
4966 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4969 if (s.expanding && s.locked == 0 &&
4970 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4971 handle_stripe_expansion(conf, sh);
4974 /* wait for this device to become unblocked */
4975 if (unlikely(s.blocked_rdev)) {
4976 if (conf->mddev->external)
4977 md_wait_for_blocked_rdev(s.blocked_rdev,
4980 /* Internal metadata will immediately
4981 * be written by raid5d, so we don't
4982 * need to wait here.
4984 rdev_dec_pending(s.blocked_rdev,
4988 if (s.handle_bad_blocks)
4989 for (i = disks; i--; ) {
4990 struct md_rdev *rdev;
4991 struct r5dev *dev = &sh->dev[i];
4992 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4993 /* We own a safe reference to the rdev */
4994 rdev = conf->disks[i].rdev;
4995 if (!rdev_set_badblocks(rdev, sh->sector,
4997 md_error(conf->mddev, rdev);
4998 rdev_dec_pending(rdev, conf->mddev);
5000 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5001 rdev = conf->disks[i].rdev;
5002 rdev_clear_badblocks(rdev, sh->sector,
5004 rdev_dec_pending(rdev, conf->mddev);
5006 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5007 rdev = conf->disks[i].replacement;
5009 /* rdev have been moved down */
5010 rdev = conf->disks[i].rdev;
5011 rdev_clear_badblocks(rdev, sh->sector,
5013 rdev_dec_pending(rdev, conf->mddev);
5018 raid_run_ops(sh, s.ops_request);
5022 if (s.dec_preread_active) {
5023 /* We delay this until after ops_run_io so that if make_request
5024 * is waiting on a flush, it won't continue until the writes
5025 * have actually been submitted.
5027 atomic_dec(&conf->preread_active_stripes);
5028 if (atomic_read(&conf->preread_active_stripes) <
5030 md_wakeup_thread(conf->mddev->thread);
5033 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5036 static void raid5_activate_delayed(struct r5conf *conf)
5038 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5039 while (!list_empty(&conf->delayed_list)) {
5040 struct list_head *l = conf->delayed_list.next;
5041 struct stripe_head *sh;
5042 sh = list_entry(l, struct stripe_head, lru);
5044 clear_bit(STRIPE_DELAYED, &sh->state);
5045 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5046 atomic_inc(&conf->preread_active_stripes);
5047 list_add_tail(&sh->lru, &conf->hold_list);
5048 raid5_wakeup_stripe_thread(sh);
5053 static void activate_bit_delay(struct r5conf *conf,
5054 struct list_head *temp_inactive_list)
5056 /* device_lock is held */
5057 struct list_head head;
5058 list_add(&head, &conf->bitmap_list);
5059 list_del_init(&conf->bitmap_list);
5060 while (!list_empty(&head)) {
5061 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5063 list_del_init(&sh->lru);
5064 atomic_inc(&sh->count);
5065 hash = sh->hash_lock_index;
5066 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5070 static int raid5_congested(struct mddev *mddev, int bits)
5072 struct r5conf *conf = mddev->private;
5074 /* No difference between reads and writes. Just check
5075 * how busy the stripe_cache is
5078 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5081 /* Also checks whether there is pressure on r5cache log space */
5082 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5086 if (atomic_read(&conf->empty_inactive_list_nr))
5092 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5094 struct r5conf *conf = mddev->private;
5095 sector_t sector = bio->bi_iter.bi_sector;
5096 unsigned int chunk_sectors;
5097 unsigned int bio_sectors = bio_sectors(bio);
5099 WARN_ON_ONCE(bio->bi_partno);
5101 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5102 return chunk_sectors >=
5103 ((sector & (chunk_sectors - 1)) + bio_sectors);
5107 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5108 * later sampled by raid5d.
5110 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5112 unsigned long flags;
5114 spin_lock_irqsave(&conf->device_lock, flags);
5116 bi->bi_next = conf->retry_read_aligned_list;
5117 conf->retry_read_aligned_list = bi;
5119 spin_unlock_irqrestore(&conf->device_lock, flags);
5120 md_wakeup_thread(conf->mddev->thread);
5123 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5124 unsigned int *offset)
5128 bi = conf->retry_read_aligned;
5130 *offset = conf->retry_read_offset;
5131 conf->retry_read_aligned = NULL;
5134 bi = conf->retry_read_aligned_list;
5136 conf->retry_read_aligned_list = bi->bi_next;
5145 * The "raid5_align_endio" should check if the read succeeded and if it
5146 * did, call bio_endio on the original bio (having bio_put the new bio
5148 * If the read failed..
5150 static void raid5_align_endio(struct bio *bi)
5152 struct bio* raid_bi = bi->bi_private;
5153 struct mddev *mddev;
5154 struct r5conf *conf;
5155 struct md_rdev *rdev;
5156 blk_status_t error = bi->bi_status;
5160 rdev = (void*)raid_bi->bi_next;
5161 raid_bi->bi_next = NULL;
5162 mddev = rdev->mddev;
5163 conf = mddev->private;
5165 rdev_dec_pending(rdev, conf->mddev);
5169 if (atomic_dec_and_test(&conf->active_aligned_reads))
5170 wake_up(&conf->wait_for_quiescent);
5174 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5176 add_bio_to_retry(raid_bi, conf);
5179 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5181 struct r5conf *conf = mddev->private;
5183 struct bio* align_bi;
5184 struct md_rdev *rdev;
5185 sector_t end_sector;
5187 if (!in_chunk_boundary(mddev, raid_bio)) {
5188 pr_debug("%s: non aligned\n", __func__);
5192 * use bio_clone_fast to make a copy of the bio
5194 align_bi = bio_clone_fast(raid_bio, GFP_NOIO, mddev->bio_set);
5198 * set bi_end_io to a new function, and set bi_private to the
5201 align_bi->bi_end_io = raid5_align_endio;
5202 align_bi->bi_private = raid_bio;
5206 align_bi->bi_iter.bi_sector =
5207 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5210 end_sector = bio_end_sector(align_bi);
5212 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5213 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5214 rdev->recovery_offset < end_sector) {
5215 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5217 (test_bit(Faulty, &rdev->flags) ||
5218 !(test_bit(In_sync, &rdev->flags) ||
5219 rdev->recovery_offset >= end_sector)))
5223 if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5233 atomic_inc(&rdev->nr_pending);
5235 raid_bio->bi_next = (void*)rdev;
5236 bio_set_dev(align_bi, rdev->bdev);
5237 bio_clear_flag(align_bi, BIO_SEG_VALID);
5239 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5240 bio_sectors(align_bi),
5241 &first_bad, &bad_sectors)) {
5243 rdev_dec_pending(rdev, mddev);
5247 /* No reshape active, so we can trust rdev->data_offset */
5248 align_bi->bi_iter.bi_sector += rdev->data_offset;
5250 spin_lock_irq(&conf->device_lock);
5251 wait_event_lock_irq(conf->wait_for_quiescent,
5254 atomic_inc(&conf->active_aligned_reads);
5255 spin_unlock_irq(&conf->device_lock);
5258 trace_block_bio_remap(align_bi->bi_disk->queue,
5259 align_bi, disk_devt(mddev->gendisk),
5260 raid_bio->bi_iter.bi_sector);
5261 generic_make_request(align_bi);
5270 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5273 sector_t sector = raid_bio->bi_iter.bi_sector;
5274 unsigned chunk_sects = mddev->chunk_sectors;
5275 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5277 if (sectors < bio_sectors(raid_bio)) {
5278 struct r5conf *conf = mddev->private;
5279 split = bio_split(raid_bio, sectors, GFP_NOIO, conf->bio_split);
5280 bio_chain(split, raid_bio);
5281 generic_make_request(raid_bio);
5285 if (!raid5_read_one_chunk(mddev, raid_bio))
5291 /* __get_priority_stripe - get the next stripe to process
5293 * Full stripe writes are allowed to pass preread active stripes up until
5294 * the bypass_threshold is exceeded. In general the bypass_count
5295 * increments when the handle_list is handled before the hold_list; however, it
5296 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5297 * stripe with in flight i/o. The bypass_count will be reset when the
5298 * head of the hold_list has changed, i.e. the head was promoted to the
5301 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5303 struct stripe_head *sh, *tmp;
5304 struct list_head *handle_list = NULL;
5305 struct r5worker_group *wg;
5306 bool second_try = !r5c_is_writeback(conf->log) &&
5307 !r5l_log_disk_error(conf);
5308 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5309 r5l_log_disk_error(conf);
5314 if (conf->worker_cnt_per_group == 0) {
5315 handle_list = try_loprio ? &conf->loprio_list :
5317 } else if (group != ANY_GROUP) {
5318 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5319 &conf->worker_groups[group].handle_list;
5320 wg = &conf->worker_groups[group];
5323 for (i = 0; i < conf->group_cnt; i++) {
5324 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5325 &conf->worker_groups[i].handle_list;
5326 wg = &conf->worker_groups[i];
5327 if (!list_empty(handle_list))
5332 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5334 list_empty(handle_list) ? "empty" : "busy",
5335 list_empty(&conf->hold_list) ? "empty" : "busy",
5336 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5338 if (!list_empty(handle_list)) {
5339 sh = list_entry(handle_list->next, typeof(*sh), lru);
5341 if (list_empty(&conf->hold_list))
5342 conf->bypass_count = 0;
5343 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5344 if (conf->hold_list.next == conf->last_hold)
5345 conf->bypass_count++;
5347 conf->last_hold = conf->hold_list.next;
5348 conf->bypass_count -= conf->bypass_threshold;
5349 if (conf->bypass_count < 0)
5350 conf->bypass_count = 0;
5353 } else if (!list_empty(&conf->hold_list) &&
5354 ((conf->bypass_threshold &&
5355 conf->bypass_count > conf->bypass_threshold) ||
5356 atomic_read(&conf->pending_full_writes) == 0)) {
5358 list_for_each_entry(tmp, &conf->hold_list, lru) {
5359 if (conf->worker_cnt_per_group == 0 ||
5360 group == ANY_GROUP ||
5361 !cpu_online(tmp->cpu) ||
5362 cpu_to_group(tmp->cpu) == group) {
5369 conf->bypass_count -= conf->bypass_threshold;
5370 if (conf->bypass_count < 0)
5371 conf->bypass_count = 0;
5380 try_loprio = !try_loprio;
5388 list_del_init(&sh->lru);
5389 BUG_ON(atomic_inc_return(&sh->count) != 1);
5393 struct raid5_plug_cb {
5394 struct blk_plug_cb cb;
5395 struct list_head list;
5396 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5399 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5401 struct raid5_plug_cb *cb = container_of(
5402 blk_cb, struct raid5_plug_cb, cb);
5403 struct stripe_head *sh;
5404 struct mddev *mddev = cb->cb.data;
5405 struct r5conf *conf = mddev->private;
5409 if (cb->list.next && !list_empty(&cb->list)) {
5410 spin_lock_irq(&conf->device_lock);
5411 while (!list_empty(&cb->list)) {
5412 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5413 list_del_init(&sh->lru);
5415 * avoid race release_stripe_plug() sees
5416 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5417 * is still in our list
5419 smp_mb__before_atomic();
5420 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5422 * STRIPE_ON_RELEASE_LIST could be set here. In that
5423 * case, the count is always > 1 here
5425 hash = sh->hash_lock_index;
5426 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5429 spin_unlock_irq(&conf->device_lock);
5431 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5432 NR_STRIPE_HASH_LOCKS);
5434 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5438 static void release_stripe_plug(struct mddev *mddev,
5439 struct stripe_head *sh)
5441 struct blk_plug_cb *blk_cb = blk_check_plugged(
5442 raid5_unplug, mddev,
5443 sizeof(struct raid5_plug_cb));
5444 struct raid5_plug_cb *cb;
5447 raid5_release_stripe(sh);
5451 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5453 if (cb->list.next == NULL) {
5455 INIT_LIST_HEAD(&cb->list);
5456 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5457 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5460 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5461 list_add_tail(&sh->lru, &cb->list);
5463 raid5_release_stripe(sh);
5466 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5468 struct r5conf *conf = mddev->private;
5469 sector_t logical_sector, last_sector;
5470 struct stripe_head *sh;
5473 if (mddev->reshape_position != MaxSector)
5474 /* Skip discard while reshape is happening */
5477 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5478 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5482 stripe_sectors = conf->chunk_sectors *
5483 (conf->raid_disks - conf->max_degraded);
5484 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5486 sector_div(last_sector, stripe_sectors);
5488 logical_sector *= conf->chunk_sectors;
5489 last_sector *= conf->chunk_sectors;
5491 for (; logical_sector < last_sector;
5492 logical_sector += STRIPE_SECTORS) {
5496 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5497 prepare_to_wait(&conf->wait_for_overlap, &w,
5498 TASK_UNINTERRUPTIBLE);
5499 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5500 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5501 raid5_release_stripe(sh);
5505 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5506 spin_lock_irq(&sh->stripe_lock);
5507 for (d = 0; d < conf->raid_disks; d++) {
5508 if (d == sh->pd_idx || d == sh->qd_idx)
5510 if (sh->dev[d].towrite || sh->dev[d].toread) {
5511 set_bit(R5_Overlap, &sh->dev[d].flags);
5512 spin_unlock_irq(&sh->stripe_lock);
5513 raid5_release_stripe(sh);
5518 set_bit(STRIPE_DISCARD, &sh->state);
5519 finish_wait(&conf->wait_for_overlap, &w);
5520 sh->overwrite_disks = 0;
5521 for (d = 0; d < conf->raid_disks; d++) {
5522 if (d == sh->pd_idx || d == sh->qd_idx)
5524 sh->dev[d].towrite = bi;
5525 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5526 bio_inc_remaining(bi);
5527 md_write_inc(mddev, bi);
5528 sh->overwrite_disks++;
5530 spin_unlock_irq(&sh->stripe_lock);
5531 if (conf->mddev->bitmap) {
5533 d < conf->raid_disks - conf->max_degraded;
5535 bitmap_startwrite(mddev->bitmap,
5539 sh->bm_seq = conf->seq_flush + 1;
5540 set_bit(STRIPE_BIT_DELAY, &sh->state);
5543 set_bit(STRIPE_HANDLE, &sh->state);
5544 clear_bit(STRIPE_DELAYED, &sh->state);
5545 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5546 atomic_inc(&conf->preread_active_stripes);
5547 release_stripe_plug(mddev, sh);
5553 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5555 struct r5conf *conf = mddev->private;
5557 sector_t new_sector;
5558 sector_t logical_sector, last_sector;
5559 struct stripe_head *sh;
5560 const int rw = bio_data_dir(bi);
5563 bool do_flush = false;
5565 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5566 int ret = r5l_handle_flush_request(conf->log, bi);
5570 if (ret == -ENODEV) {
5571 md_flush_request(mddev, bi);
5574 /* ret == -EAGAIN, fallback */
5576 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5577 * we need to flush journal device
5579 do_flush = bi->bi_opf & REQ_PREFLUSH;
5582 if (!md_write_start(mddev, bi))
5585 * If array is degraded, better not do chunk aligned read because
5586 * later we might have to read it again in order to reconstruct
5587 * data on failed drives.
5589 if (rw == READ && mddev->degraded == 0 &&
5590 mddev->reshape_position == MaxSector) {
5591 bi = chunk_aligned_read(mddev, bi);
5596 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5597 make_discard_request(mddev, bi);
5598 md_write_end(mddev);
5602 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5603 last_sector = bio_end_sector(bi);
5606 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5607 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5613 seq = read_seqcount_begin(&conf->gen_lock);
5616 prepare_to_wait(&conf->wait_for_overlap, &w,
5617 TASK_UNINTERRUPTIBLE);
5618 if (unlikely(conf->reshape_progress != MaxSector)) {
5619 /* spinlock is needed as reshape_progress may be
5620 * 64bit on a 32bit platform, and so it might be
5621 * possible to see a half-updated value
5622 * Of course reshape_progress could change after
5623 * the lock is dropped, so once we get a reference
5624 * to the stripe that we think it is, we will have
5627 spin_lock_irq(&conf->device_lock);
5628 if (mddev->reshape_backwards
5629 ? logical_sector < conf->reshape_progress
5630 : logical_sector >= conf->reshape_progress) {
5633 if (mddev->reshape_backwards
5634 ? logical_sector < conf->reshape_safe
5635 : logical_sector >= conf->reshape_safe) {
5636 spin_unlock_irq(&conf->device_lock);
5642 spin_unlock_irq(&conf->device_lock);
5645 new_sector = raid5_compute_sector(conf, logical_sector,
5648 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5649 (unsigned long long)new_sector,
5650 (unsigned long long)logical_sector);
5652 sh = raid5_get_active_stripe(conf, new_sector, previous,
5653 (bi->bi_opf & REQ_RAHEAD), 0);
5655 if (unlikely(previous)) {
5656 /* expansion might have moved on while waiting for a
5657 * stripe, so we must do the range check again.
5658 * Expansion could still move past after this
5659 * test, but as we are holding a reference to
5660 * 'sh', we know that if that happens,
5661 * STRIPE_EXPANDING will get set and the expansion
5662 * won't proceed until we finish with the stripe.
5665 spin_lock_irq(&conf->device_lock);
5666 if (mddev->reshape_backwards
5667 ? logical_sector >= conf->reshape_progress
5668 : logical_sector < conf->reshape_progress)
5669 /* mismatch, need to try again */
5671 spin_unlock_irq(&conf->device_lock);
5673 raid5_release_stripe(sh);
5679 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5680 /* Might have got the wrong stripe_head
5683 raid5_release_stripe(sh);
5688 logical_sector >= mddev->suspend_lo &&
5689 logical_sector < mddev->suspend_hi) {
5690 raid5_release_stripe(sh);
5691 /* As the suspend_* range is controlled by
5692 * userspace, we want an interruptible
5695 prepare_to_wait(&conf->wait_for_overlap,
5696 &w, TASK_INTERRUPTIBLE);
5697 if (logical_sector >= mddev->suspend_lo &&
5698 logical_sector < mddev->suspend_hi) {
5701 sigprocmask(SIG_BLOCK, &full, &old);
5703 sigprocmask(SIG_SETMASK, &old, NULL);
5709 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5710 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5711 /* Stripe is busy expanding or
5712 * add failed due to overlap. Flush everything
5715 md_wakeup_thread(mddev->thread);
5716 raid5_release_stripe(sh);
5722 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5723 /* we only need flush for one stripe */
5727 set_bit(STRIPE_HANDLE, &sh->state);
5728 clear_bit(STRIPE_DELAYED, &sh->state);
5729 if ((!sh->batch_head || sh == sh->batch_head) &&
5730 (bi->bi_opf & REQ_SYNC) &&
5731 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5732 atomic_inc(&conf->preread_active_stripes);
5733 release_stripe_plug(mddev, sh);
5735 /* cannot get stripe for read-ahead, just give-up */
5736 bi->bi_status = BLK_STS_IOERR;
5740 finish_wait(&conf->wait_for_overlap, &w);
5743 md_write_end(mddev);
5748 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5750 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5752 /* reshaping is quite different to recovery/resync so it is
5753 * handled quite separately ... here.
5755 * On each call to sync_request, we gather one chunk worth of
5756 * destination stripes and flag them as expanding.
5757 * Then we find all the source stripes and request reads.
5758 * As the reads complete, handle_stripe will copy the data
5759 * into the destination stripe and release that stripe.
5761 struct r5conf *conf = mddev->private;
5762 struct stripe_head *sh;
5763 sector_t first_sector, last_sector;
5764 int raid_disks = conf->previous_raid_disks;
5765 int data_disks = raid_disks - conf->max_degraded;
5766 int new_data_disks = conf->raid_disks - conf->max_degraded;
5769 sector_t writepos, readpos, safepos;
5770 sector_t stripe_addr;
5771 int reshape_sectors;
5772 struct list_head stripes;
5775 if (sector_nr == 0) {
5776 /* If restarting in the middle, skip the initial sectors */
5777 if (mddev->reshape_backwards &&
5778 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5779 sector_nr = raid5_size(mddev, 0, 0)
5780 - conf->reshape_progress;
5781 } else if (mddev->reshape_backwards &&
5782 conf->reshape_progress == MaxSector) {
5783 /* shouldn't happen, but just in case, finish up.*/
5784 sector_nr = MaxSector;
5785 } else if (!mddev->reshape_backwards &&
5786 conf->reshape_progress > 0)
5787 sector_nr = conf->reshape_progress;
5788 sector_div(sector_nr, new_data_disks);
5790 mddev->curr_resync_completed = sector_nr;
5791 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5798 /* We need to process a full chunk at a time.
5799 * If old and new chunk sizes differ, we need to process the
5803 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5805 /* We update the metadata at least every 10 seconds, or when
5806 * the data about to be copied would over-write the source of
5807 * the data at the front of the range. i.e. one new_stripe
5808 * along from reshape_progress new_maps to after where
5809 * reshape_safe old_maps to
5811 writepos = conf->reshape_progress;
5812 sector_div(writepos, new_data_disks);
5813 readpos = conf->reshape_progress;
5814 sector_div(readpos, data_disks);
5815 safepos = conf->reshape_safe;
5816 sector_div(safepos, data_disks);
5817 if (mddev->reshape_backwards) {
5818 BUG_ON(writepos < reshape_sectors);
5819 writepos -= reshape_sectors;
5820 readpos += reshape_sectors;
5821 safepos += reshape_sectors;
5823 writepos += reshape_sectors;
5824 /* readpos and safepos are worst-case calculations.
5825 * A negative number is overly pessimistic, and causes
5826 * obvious problems for unsigned storage. So clip to 0.
5828 readpos -= min_t(sector_t, reshape_sectors, readpos);
5829 safepos -= min_t(sector_t, reshape_sectors, safepos);
5832 /* Having calculated the 'writepos' possibly use it
5833 * to set 'stripe_addr' which is where we will write to.
5835 if (mddev->reshape_backwards) {
5836 BUG_ON(conf->reshape_progress == 0);
5837 stripe_addr = writepos;
5838 BUG_ON((mddev->dev_sectors &
5839 ~((sector_t)reshape_sectors - 1))
5840 - reshape_sectors - stripe_addr
5843 BUG_ON(writepos != sector_nr + reshape_sectors);
5844 stripe_addr = sector_nr;
5847 /* 'writepos' is the most advanced device address we might write.
5848 * 'readpos' is the least advanced device address we might read.
5849 * 'safepos' is the least address recorded in the metadata as having
5851 * If there is a min_offset_diff, these are adjusted either by
5852 * increasing the safepos/readpos if diff is negative, or
5853 * increasing writepos if diff is positive.
5854 * If 'readpos' is then behind 'writepos', there is no way that we can
5855 * ensure safety in the face of a crash - that must be done by userspace
5856 * making a backup of the data. So in that case there is no particular
5857 * rush to update metadata.
5858 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5859 * update the metadata to advance 'safepos' to match 'readpos' so that
5860 * we can be safe in the event of a crash.
5861 * So we insist on updating metadata if safepos is behind writepos and
5862 * readpos is beyond writepos.
5863 * In any case, update the metadata every 10 seconds.
5864 * Maybe that number should be configurable, but I'm not sure it is
5865 * worth it.... maybe it could be a multiple of safemode_delay???
5867 if (conf->min_offset_diff < 0) {
5868 safepos += -conf->min_offset_diff;
5869 readpos += -conf->min_offset_diff;
5871 writepos += conf->min_offset_diff;
5873 if ((mddev->reshape_backwards
5874 ? (safepos > writepos && readpos < writepos)
5875 : (safepos < writepos && readpos > writepos)) ||
5876 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5877 /* Cannot proceed until we've updated the superblock... */
5878 wait_event(conf->wait_for_overlap,
5879 atomic_read(&conf->reshape_stripes)==0
5880 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5881 if (atomic_read(&conf->reshape_stripes) != 0)
5883 mddev->reshape_position = conf->reshape_progress;
5884 mddev->curr_resync_completed = sector_nr;
5885 conf->reshape_checkpoint = jiffies;
5886 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5887 md_wakeup_thread(mddev->thread);
5888 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5889 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5890 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5892 spin_lock_irq(&conf->device_lock);
5893 conf->reshape_safe = mddev->reshape_position;
5894 spin_unlock_irq(&conf->device_lock);
5895 wake_up(&conf->wait_for_overlap);
5896 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5899 INIT_LIST_HEAD(&stripes);
5900 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5902 int skipped_disk = 0;
5903 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5904 set_bit(STRIPE_EXPANDING, &sh->state);
5905 atomic_inc(&conf->reshape_stripes);
5906 /* If any of this stripe is beyond the end of the old
5907 * array, then we need to zero those blocks
5909 for (j=sh->disks; j--;) {
5911 if (j == sh->pd_idx)
5913 if (conf->level == 6 &&
5916 s = raid5_compute_blocknr(sh, j, 0);
5917 if (s < raid5_size(mddev, 0, 0)) {
5921 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5922 set_bit(R5_Expanded, &sh->dev[j].flags);
5923 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5925 if (!skipped_disk) {
5926 set_bit(STRIPE_EXPAND_READY, &sh->state);
5927 set_bit(STRIPE_HANDLE, &sh->state);
5929 list_add(&sh->lru, &stripes);
5931 spin_lock_irq(&conf->device_lock);
5932 if (mddev->reshape_backwards)
5933 conf->reshape_progress -= reshape_sectors * new_data_disks;
5935 conf->reshape_progress += reshape_sectors * new_data_disks;
5936 spin_unlock_irq(&conf->device_lock);
5937 /* Ok, those stripe are ready. We can start scheduling
5938 * reads on the source stripes.
5939 * The source stripes are determined by mapping the first and last
5940 * block on the destination stripes.
5943 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5946 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5947 * new_data_disks - 1),
5949 if (last_sector >= mddev->dev_sectors)
5950 last_sector = mddev->dev_sectors - 1;
5951 while (first_sector <= last_sector) {
5952 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5953 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5954 set_bit(STRIPE_HANDLE, &sh->state);
5955 raid5_release_stripe(sh);
5956 first_sector += STRIPE_SECTORS;
5958 /* Now that the sources are clearly marked, we can release
5959 * the destination stripes
5961 while (!list_empty(&stripes)) {
5962 sh = list_entry(stripes.next, struct stripe_head, lru);
5963 list_del_init(&sh->lru);
5964 raid5_release_stripe(sh);
5966 /* If this takes us to the resync_max point where we have to pause,
5967 * then we need to write out the superblock.
5969 sector_nr += reshape_sectors;
5970 retn = reshape_sectors;
5972 if (mddev->curr_resync_completed > mddev->resync_max ||
5973 (sector_nr - mddev->curr_resync_completed) * 2
5974 >= mddev->resync_max - mddev->curr_resync_completed) {
5975 /* Cannot proceed until we've updated the superblock... */
5976 wait_event(conf->wait_for_overlap,
5977 atomic_read(&conf->reshape_stripes) == 0
5978 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5979 if (atomic_read(&conf->reshape_stripes) != 0)
5981 mddev->reshape_position = conf->reshape_progress;
5982 mddev->curr_resync_completed = sector_nr;
5983 conf->reshape_checkpoint = jiffies;
5984 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5985 md_wakeup_thread(mddev->thread);
5986 wait_event(mddev->sb_wait,
5987 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5988 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5989 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5991 spin_lock_irq(&conf->device_lock);
5992 conf->reshape_safe = mddev->reshape_position;
5993 spin_unlock_irq(&conf->device_lock);
5994 wake_up(&conf->wait_for_overlap);
5995 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6001 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6004 struct r5conf *conf = mddev->private;
6005 struct stripe_head *sh;
6006 sector_t max_sector = mddev->dev_sectors;
6007 sector_t sync_blocks;
6008 int still_degraded = 0;
6011 if (sector_nr >= max_sector) {
6012 /* just being told to finish up .. nothing much to do */
6014 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6019 if (mddev->curr_resync < max_sector) /* aborted */
6020 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6022 else /* completed sync */
6024 bitmap_close_sync(mddev->bitmap);
6029 /* Allow raid5_quiesce to complete */
6030 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6032 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6033 return reshape_request(mddev, sector_nr, skipped);
6035 /* No need to check resync_max as we never do more than one
6036 * stripe, and as resync_max will always be on a chunk boundary,
6037 * if the check in md_do_sync didn't fire, there is no chance
6038 * of overstepping resync_max here
6041 /* if there is too many failed drives and we are trying
6042 * to resync, then assert that we are finished, because there is
6043 * nothing we can do.
6045 if (mddev->degraded >= conf->max_degraded &&
6046 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6047 sector_t rv = mddev->dev_sectors - sector_nr;
6051 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6053 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6054 sync_blocks >= STRIPE_SECTORS) {
6055 /* we can skip this block, and probably more */
6056 sync_blocks /= STRIPE_SECTORS;
6058 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6061 bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6063 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6065 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6066 /* make sure we don't swamp the stripe cache if someone else
6067 * is trying to get access
6069 schedule_timeout_uninterruptible(1);
6071 /* Need to check if array will still be degraded after recovery/resync
6072 * Note in case of > 1 drive failures it's possible we're rebuilding
6073 * one drive while leaving another faulty drive in array.
6076 for (i = 0; i < conf->raid_disks; i++) {
6077 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
6079 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6084 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6086 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6087 set_bit(STRIPE_HANDLE, &sh->state);
6089 raid5_release_stripe(sh);
6091 return STRIPE_SECTORS;
6094 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6095 unsigned int offset)
6097 /* We may not be able to submit a whole bio at once as there
6098 * may not be enough stripe_heads available.
6099 * We cannot pre-allocate enough stripe_heads as we may need
6100 * more than exist in the cache (if we allow ever large chunks).
6101 * So we do one stripe head at a time and record in
6102 * ->bi_hw_segments how many have been done.
6104 * We *know* that this entire raid_bio is in one chunk, so
6105 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6107 struct stripe_head *sh;
6109 sector_t sector, logical_sector, last_sector;
6113 logical_sector = raid_bio->bi_iter.bi_sector &
6114 ~((sector_t)STRIPE_SECTORS-1);
6115 sector = raid5_compute_sector(conf, logical_sector,
6117 last_sector = bio_end_sector(raid_bio);
6119 for (; logical_sector < last_sector;
6120 logical_sector += STRIPE_SECTORS,
6121 sector += STRIPE_SECTORS,
6125 /* already done this stripe */
6128 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6131 /* failed to get a stripe - must wait */
6132 conf->retry_read_aligned = raid_bio;
6133 conf->retry_read_offset = scnt;
6137 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6138 raid5_release_stripe(sh);
6139 conf->retry_read_aligned = raid_bio;
6140 conf->retry_read_offset = scnt;
6144 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6146 raid5_release_stripe(sh);
6150 bio_endio(raid_bio);
6152 if (atomic_dec_and_test(&conf->active_aligned_reads))
6153 wake_up(&conf->wait_for_quiescent);
6157 static int handle_active_stripes(struct r5conf *conf, int group,
6158 struct r5worker *worker,
6159 struct list_head *temp_inactive_list)
6161 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6162 int i, batch_size = 0, hash;
6163 bool release_inactive = false;
6165 while (batch_size < MAX_STRIPE_BATCH &&
6166 (sh = __get_priority_stripe(conf, group)) != NULL)
6167 batch[batch_size++] = sh;
6169 if (batch_size == 0) {
6170 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6171 if (!list_empty(temp_inactive_list + i))
6173 if (i == NR_STRIPE_HASH_LOCKS) {
6174 spin_unlock_irq(&conf->device_lock);
6175 r5l_flush_stripe_to_raid(conf->log);
6176 spin_lock_irq(&conf->device_lock);
6179 release_inactive = true;
6181 spin_unlock_irq(&conf->device_lock);
6183 release_inactive_stripe_list(conf, temp_inactive_list,
6184 NR_STRIPE_HASH_LOCKS);
6186 r5l_flush_stripe_to_raid(conf->log);
6187 if (release_inactive) {
6188 spin_lock_irq(&conf->device_lock);
6192 for (i = 0; i < batch_size; i++)
6193 handle_stripe(batch[i]);
6194 log_write_stripe_run(conf);
6198 spin_lock_irq(&conf->device_lock);
6199 for (i = 0; i < batch_size; i++) {
6200 hash = batch[i]->hash_lock_index;
6201 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6206 static void raid5_do_work(struct work_struct *work)
6208 struct r5worker *worker = container_of(work, struct r5worker, work);
6209 struct r5worker_group *group = worker->group;
6210 struct r5conf *conf = group->conf;
6211 struct mddev *mddev = conf->mddev;
6212 int group_id = group - conf->worker_groups;
6214 struct blk_plug plug;
6216 pr_debug("+++ raid5worker active\n");
6218 blk_start_plug(&plug);
6220 spin_lock_irq(&conf->device_lock);
6222 int batch_size, released;
6224 released = release_stripe_list(conf, worker->temp_inactive_list);
6226 batch_size = handle_active_stripes(conf, group_id, worker,
6227 worker->temp_inactive_list);
6228 worker->working = false;
6229 if (!batch_size && !released)
6231 handled += batch_size;
6232 wait_event_lock_irq(mddev->sb_wait,
6233 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6236 pr_debug("%d stripes handled\n", handled);
6238 spin_unlock_irq(&conf->device_lock);
6240 async_tx_issue_pending_all();
6241 blk_finish_plug(&plug);
6243 pr_debug("--- raid5worker inactive\n");
6247 * This is our raid5 kernel thread.
6249 * We scan the hash table for stripes which can be handled now.
6250 * During the scan, completed stripes are saved for us by the interrupt
6251 * handler, so that they will not have to wait for our next wakeup.
6253 static void raid5d(struct md_thread *thread)
6255 struct mddev *mddev = thread->mddev;
6256 struct r5conf *conf = mddev->private;
6258 struct blk_plug plug;
6260 pr_debug("+++ raid5d active\n");
6262 md_check_recovery(mddev);
6264 blk_start_plug(&plug);
6266 spin_lock_irq(&conf->device_lock);
6269 int batch_size, released;
6270 unsigned int offset;
6272 released = release_stripe_list(conf, conf->temp_inactive_list);
6274 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6277 !list_empty(&conf->bitmap_list)) {
6278 /* Now is a good time to flush some bitmap updates */
6280 spin_unlock_irq(&conf->device_lock);
6281 bitmap_unplug(mddev->bitmap);
6282 spin_lock_irq(&conf->device_lock);
6283 conf->seq_write = conf->seq_flush;
6284 activate_bit_delay(conf, conf->temp_inactive_list);
6286 raid5_activate_delayed(conf);
6288 while ((bio = remove_bio_from_retry(conf, &offset))) {
6290 spin_unlock_irq(&conf->device_lock);
6291 ok = retry_aligned_read(conf, bio, offset);
6292 spin_lock_irq(&conf->device_lock);
6298 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6299 conf->temp_inactive_list);
6300 if (!batch_size && !released)
6302 handled += batch_size;
6304 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6305 spin_unlock_irq(&conf->device_lock);
6306 md_check_recovery(mddev);
6307 spin_lock_irq(&conf->device_lock);
6310 pr_debug("%d stripes handled\n", handled);
6312 spin_unlock_irq(&conf->device_lock);
6313 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6314 mutex_trylock(&conf->cache_size_mutex)) {
6315 grow_one_stripe(conf, __GFP_NOWARN);
6316 /* Set flag even if allocation failed. This helps
6317 * slow down allocation requests when mem is short
6319 set_bit(R5_DID_ALLOC, &conf->cache_state);
6320 mutex_unlock(&conf->cache_size_mutex);
6323 flush_deferred_bios(conf);
6325 r5l_flush_stripe_to_raid(conf->log);
6327 async_tx_issue_pending_all();
6328 blk_finish_plug(&plug);
6330 pr_debug("--- raid5d inactive\n");
6334 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6336 struct r5conf *conf;
6338 spin_lock(&mddev->lock);
6339 conf = mddev->private;
6341 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6342 spin_unlock(&mddev->lock);
6347 raid5_set_cache_size(struct mddev *mddev, int size)
6349 struct r5conf *conf = mddev->private;
6351 if (size <= 16 || size > 32768)
6354 conf->min_nr_stripes = size;
6355 mutex_lock(&conf->cache_size_mutex);
6356 while (size < conf->max_nr_stripes &&
6357 drop_one_stripe(conf))
6359 mutex_unlock(&conf->cache_size_mutex);
6361 md_allow_write(mddev);
6363 mutex_lock(&conf->cache_size_mutex);
6364 while (size > conf->max_nr_stripes)
6365 if (!grow_one_stripe(conf, GFP_KERNEL))
6367 mutex_unlock(&conf->cache_size_mutex);
6371 EXPORT_SYMBOL(raid5_set_cache_size);
6374 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6376 struct r5conf *conf;
6380 if (len >= PAGE_SIZE)
6382 if (kstrtoul(page, 10, &new))
6384 err = mddev_lock(mddev);
6387 conf = mddev->private;
6391 err = raid5_set_cache_size(mddev, new);
6392 mddev_unlock(mddev);
6397 static struct md_sysfs_entry
6398 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6399 raid5_show_stripe_cache_size,
6400 raid5_store_stripe_cache_size);
6403 raid5_show_rmw_level(struct mddev *mddev, char *page)
6405 struct r5conf *conf = mddev->private;
6407 return sprintf(page, "%d\n", conf->rmw_level);
6413 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6415 struct r5conf *conf = mddev->private;
6421 if (len >= PAGE_SIZE)
6424 if (kstrtoul(page, 10, &new))
6427 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6430 if (new != PARITY_DISABLE_RMW &&
6431 new != PARITY_ENABLE_RMW &&
6432 new != PARITY_PREFER_RMW)
6435 conf->rmw_level = new;
6439 static struct md_sysfs_entry
6440 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6441 raid5_show_rmw_level,
6442 raid5_store_rmw_level);
6446 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6448 struct r5conf *conf;
6450 spin_lock(&mddev->lock);
6451 conf = mddev->private;
6453 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6454 spin_unlock(&mddev->lock);
6459 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6461 struct r5conf *conf;
6465 if (len >= PAGE_SIZE)
6467 if (kstrtoul(page, 10, &new))
6470 err = mddev_lock(mddev);
6473 conf = mddev->private;
6476 else if (new > conf->min_nr_stripes)
6479 conf->bypass_threshold = new;
6480 mddev_unlock(mddev);
6484 static struct md_sysfs_entry
6485 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6487 raid5_show_preread_threshold,
6488 raid5_store_preread_threshold);
6491 raid5_show_skip_copy(struct mddev *mddev, char *page)
6493 struct r5conf *conf;
6495 spin_lock(&mddev->lock);
6496 conf = mddev->private;
6498 ret = sprintf(page, "%d\n", conf->skip_copy);
6499 spin_unlock(&mddev->lock);
6504 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6506 struct r5conf *conf;
6510 if (len >= PAGE_SIZE)
6512 if (kstrtoul(page, 10, &new))
6516 err = mddev_lock(mddev);
6519 conf = mddev->private;
6522 else if (new != conf->skip_copy) {
6523 mddev_suspend(mddev);
6524 conf->skip_copy = new;
6526 mddev->queue->backing_dev_info->capabilities |=
6527 BDI_CAP_STABLE_WRITES;
6529 mddev->queue->backing_dev_info->capabilities &=
6530 ~BDI_CAP_STABLE_WRITES;
6531 mddev_resume(mddev);
6533 mddev_unlock(mddev);
6537 static struct md_sysfs_entry
6538 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6539 raid5_show_skip_copy,
6540 raid5_store_skip_copy);
6543 stripe_cache_active_show(struct mddev *mddev, char *page)
6545 struct r5conf *conf = mddev->private;
6547 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6552 static struct md_sysfs_entry
6553 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6556 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6558 struct r5conf *conf;
6560 spin_lock(&mddev->lock);
6561 conf = mddev->private;
6563 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6564 spin_unlock(&mddev->lock);
6568 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6570 int *worker_cnt_per_group,
6571 struct r5worker_group **worker_groups);
6573 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6575 struct r5conf *conf;
6578 struct r5worker_group *new_groups, *old_groups;
6579 int group_cnt, worker_cnt_per_group;
6581 if (len >= PAGE_SIZE)
6583 if (kstrtoul(page, 10, &new))
6586 err = mddev_lock(mddev);
6589 conf = mddev->private;
6592 else if (new != conf->worker_cnt_per_group) {
6593 mddev_suspend(mddev);
6595 old_groups = conf->worker_groups;
6597 flush_workqueue(raid5_wq);
6599 err = alloc_thread_groups(conf, new,
6600 &group_cnt, &worker_cnt_per_group,
6603 spin_lock_irq(&conf->device_lock);
6604 conf->group_cnt = group_cnt;
6605 conf->worker_cnt_per_group = worker_cnt_per_group;
6606 conf->worker_groups = new_groups;
6607 spin_unlock_irq(&conf->device_lock);
6610 kfree(old_groups[0].workers);
6613 mddev_resume(mddev);
6615 mddev_unlock(mddev);
6620 static struct md_sysfs_entry
6621 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6622 raid5_show_group_thread_cnt,
6623 raid5_store_group_thread_cnt);
6625 static struct attribute *raid5_attrs[] = {
6626 &raid5_stripecache_size.attr,
6627 &raid5_stripecache_active.attr,
6628 &raid5_preread_bypass_threshold.attr,
6629 &raid5_group_thread_cnt.attr,
6630 &raid5_skip_copy.attr,
6631 &raid5_rmw_level.attr,
6632 &r5c_journal_mode.attr,
6635 static struct attribute_group raid5_attrs_group = {
6637 .attrs = raid5_attrs,
6640 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6642 int *worker_cnt_per_group,
6643 struct r5worker_group **worker_groups)
6647 struct r5worker *workers;
6649 *worker_cnt_per_group = cnt;
6652 *worker_groups = NULL;
6655 *group_cnt = num_possible_nodes();
6656 size = sizeof(struct r5worker) * cnt;
6657 workers = kzalloc(size * *group_cnt, GFP_NOIO);
6658 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6659 *group_cnt, GFP_NOIO);
6660 if (!*worker_groups || !workers) {
6662 kfree(*worker_groups);
6666 for (i = 0; i < *group_cnt; i++) {
6667 struct r5worker_group *group;
6669 group = &(*worker_groups)[i];
6670 INIT_LIST_HEAD(&group->handle_list);
6671 INIT_LIST_HEAD(&group->loprio_list);
6673 group->workers = workers + i * cnt;
6675 for (j = 0; j < cnt; j++) {
6676 struct r5worker *worker = group->workers + j;
6677 worker->group = group;
6678 INIT_WORK(&worker->work, raid5_do_work);
6680 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6681 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6688 static void free_thread_groups(struct r5conf *conf)
6690 if (conf->worker_groups)
6691 kfree(conf->worker_groups[0].workers);
6692 kfree(conf->worker_groups);
6693 conf->worker_groups = NULL;
6697 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6699 struct r5conf *conf = mddev->private;
6702 sectors = mddev->dev_sectors;
6704 /* size is defined by the smallest of previous and new size */
6705 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6707 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6708 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6709 return sectors * (raid_disks - conf->max_degraded);
6712 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6714 safe_put_page(percpu->spare_page);
6715 if (percpu->scribble)
6716 flex_array_free(percpu->scribble);
6717 percpu->spare_page = NULL;
6718 percpu->scribble = NULL;
6721 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6723 if (conf->level == 6 && !percpu->spare_page)
6724 percpu->spare_page = alloc_page(GFP_KERNEL);
6725 if (!percpu->scribble)
6726 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6727 conf->previous_raid_disks),
6728 max(conf->chunk_sectors,
6729 conf->prev_chunk_sectors)
6733 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6734 free_scratch_buffer(conf, percpu);
6741 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6743 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6745 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6749 static void raid5_free_percpu(struct r5conf *conf)
6754 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6755 free_percpu(conf->percpu);
6758 static void free_conf(struct r5conf *conf)
6764 if (conf->shrinker.nr_deferred)
6765 unregister_shrinker(&conf->shrinker);
6767 free_thread_groups(conf);
6768 shrink_stripes(conf);
6769 raid5_free_percpu(conf);
6770 for (i = 0; i < conf->pool_size; i++)
6771 if (conf->disks[i].extra_page)
6772 put_page(conf->disks[i].extra_page);
6774 if (conf->bio_split)
6775 bioset_free(conf->bio_split);
6776 kfree(conf->stripe_hashtbl);
6777 kfree(conf->pending_data);
6781 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6783 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6784 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6786 if (alloc_scratch_buffer(conf, percpu)) {
6787 pr_warn("%s: failed memory allocation for cpu%u\n",
6794 static int raid5_alloc_percpu(struct r5conf *conf)
6798 conf->percpu = alloc_percpu(struct raid5_percpu);
6802 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6804 conf->scribble_disks = max(conf->raid_disks,
6805 conf->previous_raid_disks);
6806 conf->scribble_sectors = max(conf->chunk_sectors,
6807 conf->prev_chunk_sectors);
6812 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6813 struct shrink_control *sc)
6815 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6816 unsigned long ret = SHRINK_STOP;
6818 if (mutex_trylock(&conf->cache_size_mutex)) {
6820 while (ret < sc->nr_to_scan &&
6821 conf->max_nr_stripes > conf->min_nr_stripes) {
6822 if (drop_one_stripe(conf) == 0) {
6828 mutex_unlock(&conf->cache_size_mutex);
6833 static unsigned long raid5_cache_count(struct shrinker *shrink,
6834 struct shrink_control *sc)
6836 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6838 if (conf->max_nr_stripes < conf->min_nr_stripes)
6839 /* unlikely, but not impossible */
6841 return conf->max_nr_stripes - conf->min_nr_stripes;
6844 static struct r5conf *setup_conf(struct mddev *mddev)
6846 struct r5conf *conf;
6847 int raid_disk, memory, max_disks;
6848 struct md_rdev *rdev;
6849 struct disk_info *disk;
6852 int group_cnt, worker_cnt_per_group;
6853 struct r5worker_group *new_group;
6855 if (mddev->new_level != 5
6856 && mddev->new_level != 4
6857 && mddev->new_level != 6) {
6858 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6859 mdname(mddev), mddev->new_level);
6860 return ERR_PTR(-EIO);
6862 if ((mddev->new_level == 5
6863 && !algorithm_valid_raid5(mddev->new_layout)) ||
6864 (mddev->new_level == 6
6865 && !algorithm_valid_raid6(mddev->new_layout))) {
6866 pr_warn("md/raid:%s: layout %d not supported\n",
6867 mdname(mddev), mddev->new_layout);
6868 return ERR_PTR(-EIO);
6870 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6871 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6872 mdname(mddev), mddev->raid_disks);
6873 return ERR_PTR(-EINVAL);
6876 if (!mddev->new_chunk_sectors ||
6877 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6878 !is_power_of_2(mddev->new_chunk_sectors)) {
6879 pr_warn("md/raid:%s: invalid chunk size %d\n",
6880 mdname(mddev), mddev->new_chunk_sectors << 9);
6881 return ERR_PTR(-EINVAL);
6884 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6887 INIT_LIST_HEAD(&conf->free_list);
6888 INIT_LIST_HEAD(&conf->pending_list);
6889 conf->pending_data = kzalloc(sizeof(struct r5pending_data) *
6890 PENDING_IO_MAX, GFP_KERNEL);
6891 if (!conf->pending_data)
6893 for (i = 0; i < PENDING_IO_MAX; i++)
6894 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6895 /* Don't enable multi-threading by default*/
6896 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6898 conf->group_cnt = group_cnt;
6899 conf->worker_cnt_per_group = worker_cnt_per_group;
6900 conf->worker_groups = new_group;
6903 spin_lock_init(&conf->device_lock);
6904 seqcount_init(&conf->gen_lock);
6905 mutex_init(&conf->cache_size_mutex);
6906 init_waitqueue_head(&conf->wait_for_quiescent);
6907 init_waitqueue_head(&conf->wait_for_stripe);
6908 init_waitqueue_head(&conf->wait_for_overlap);
6909 INIT_LIST_HEAD(&conf->handle_list);
6910 INIT_LIST_HEAD(&conf->loprio_list);
6911 INIT_LIST_HEAD(&conf->hold_list);
6912 INIT_LIST_HEAD(&conf->delayed_list);
6913 INIT_LIST_HEAD(&conf->bitmap_list);
6914 init_llist_head(&conf->released_stripes);
6915 atomic_set(&conf->active_stripes, 0);
6916 atomic_set(&conf->preread_active_stripes, 0);
6917 atomic_set(&conf->active_aligned_reads, 0);
6918 spin_lock_init(&conf->pending_bios_lock);
6919 conf->batch_bio_dispatch = true;
6920 rdev_for_each(rdev, mddev) {
6921 if (test_bit(Journal, &rdev->flags))
6923 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6924 conf->batch_bio_dispatch = false;
6929 conf->bypass_threshold = BYPASS_THRESHOLD;
6930 conf->recovery_disabled = mddev->recovery_disabled - 1;
6932 conf->raid_disks = mddev->raid_disks;
6933 if (mddev->reshape_position == MaxSector)
6934 conf->previous_raid_disks = mddev->raid_disks;
6936 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6937 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6939 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6945 for (i = 0; i < max_disks; i++) {
6946 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6947 if (!conf->disks[i].extra_page)
6951 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
6952 if (!conf->bio_split)
6954 conf->mddev = mddev;
6956 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6959 /* We init hash_locks[0] separately to that it can be used
6960 * as the reference lock in the spin_lock_nest_lock() call
6961 * in lock_all_device_hash_locks_irq in order to convince
6962 * lockdep that we know what we are doing.
6964 spin_lock_init(conf->hash_locks);
6965 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6966 spin_lock_init(conf->hash_locks + i);
6968 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6969 INIT_LIST_HEAD(conf->inactive_list + i);
6971 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6972 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6974 atomic_set(&conf->r5c_cached_full_stripes, 0);
6975 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6976 atomic_set(&conf->r5c_cached_partial_stripes, 0);
6977 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6978 atomic_set(&conf->r5c_flushing_full_stripes, 0);
6979 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
6981 conf->level = mddev->new_level;
6982 conf->chunk_sectors = mddev->new_chunk_sectors;
6983 if (raid5_alloc_percpu(conf) != 0)
6986 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6988 rdev_for_each(rdev, mddev) {
6989 raid_disk = rdev->raid_disk;
6990 if (raid_disk >= max_disks
6991 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6993 disk = conf->disks + raid_disk;
6995 if (test_bit(Replacement, &rdev->flags)) {
6996 if (disk->replacement)
6998 disk->replacement = rdev;
7005 if (test_bit(In_sync, &rdev->flags)) {
7006 char b[BDEVNAME_SIZE];
7007 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7008 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7009 } else if (rdev->saved_raid_disk != raid_disk)
7010 /* Cannot rely on bitmap to complete recovery */
7014 conf->level = mddev->new_level;
7015 if (conf->level == 6) {
7016 conf->max_degraded = 2;
7017 if (raid6_call.xor_syndrome)
7018 conf->rmw_level = PARITY_ENABLE_RMW;
7020 conf->rmw_level = PARITY_DISABLE_RMW;
7022 conf->max_degraded = 1;
7023 conf->rmw_level = PARITY_ENABLE_RMW;
7025 conf->algorithm = mddev->new_layout;
7026 conf->reshape_progress = mddev->reshape_position;
7027 if (conf->reshape_progress != MaxSector) {
7028 conf->prev_chunk_sectors = mddev->chunk_sectors;
7029 conf->prev_algo = mddev->layout;
7031 conf->prev_chunk_sectors = conf->chunk_sectors;
7032 conf->prev_algo = conf->algorithm;
7035 conf->min_nr_stripes = NR_STRIPES;
7036 if (mddev->reshape_position != MaxSector) {
7037 int stripes = max_t(int,
7038 ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7039 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7040 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7041 if (conf->min_nr_stripes != NR_STRIPES)
7042 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7043 mdname(mddev), conf->min_nr_stripes);
7045 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7046 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7047 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7048 if (grow_stripes(conf, conf->min_nr_stripes)) {
7049 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7050 mdname(mddev), memory);
7053 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7055 * Losing a stripe head costs more than the time to refill it,
7056 * it reduces the queue depth and so can hurt throughput.
7057 * So set it rather large, scaled by number of devices.
7059 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7060 conf->shrinker.scan_objects = raid5_cache_scan;
7061 conf->shrinker.count_objects = raid5_cache_count;
7062 conf->shrinker.batch = 128;
7063 conf->shrinker.flags = 0;
7064 if (register_shrinker(&conf->shrinker)) {
7065 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7070 sprintf(pers_name, "raid%d", mddev->new_level);
7071 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7072 if (!conf->thread) {
7073 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7083 return ERR_PTR(-EIO);
7085 return ERR_PTR(-ENOMEM);
7088 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7091 case ALGORITHM_PARITY_0:
7092 if (raid_disk < max_degraded)
7095 case ALGORITHM_PARITY_N:
7096 if (raid_disk >= raid_disks - max_degraded)
7099 case ALGORITHM_PARITY_0_6:
7100 if (raid_disk == 0 ||
7101 raid_disk == raid_disks - 1)
7104 case ALGORITHM_LEFT_ASYMMETRIC_6:
7105 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7106 case ALGORITHM_LEFT_SYMMETRIC_6:
7107 case ALGORITHM_RIGHT_SYMMETRIC_6:
7108 if (raid_disk == raid_disks - 1)
7114 static int raid5_run(struct mddev *mddev)
7116 struct r5conf *conf;
7117 int working_disks = 0;
7118 int dirty_parity_disks = 0;
7119 struct md_rdev *rdev;
7120 struct md_rdev *journal_dev = NULL;
7121 sector_t reshape_offset = 0;
7123 long long min_offset_diff = 0;
7126 if (mddev_init_writes_pending(mddev) < 0)
7129 if (mddev->recovery_cp != MaxSector)
7130 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7133 rdev_for_each(rdev, mddev) {
7136 if (test_bit(Journal, &rdev->flags)) {
7140 if (rdev->raid_disk < 0)
7142 diff = (rdev->new_data_offset - rdev->data_offset);
7144 min_offset_diff = diff;
7146 } else if (mddev->reshape_backwards &&
7147 diff < min_offset_diff)
7148 min_offset_diff = diff;
7149 else if (!mddev->reshape_backwards &&
7150 diff > min_offset_diff)
7151 min_offset_diff = diff;
7154 if (mddev->reshape_position != MaxSector) {
7155 /* Check that we can continue the reshape.
7156 * Difficulties arise if the stripe we would write to
7157 * next is at or after the stripe we would read from next.
7158 * For a reshape that changes the number of devices, this
7159 * is only possible for a very short time, and mdadm makes
7160 * sure that time appears to have past before assembling
7161 * the array. So we fail if that time hasn't passed.
7162 * For a reshape that keeps the number of devices the same
7163 * mdadm must be monitoring the reshape can keeping the
7164 * critical areas read-only and backed up. It will start
7165 * the array in read-only mode, so we check for that.
7167 sector_t here_new, here_old;
7169 int max_degraded = (mddev->level == 6 ? 2 : 1);
7174 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7179 if (mddev->new_level != mddev->level) {
7180 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7184 old_disks = mddev->raid_disks - mddev->delta_disks;
7185 /* reshape_position must be on a new-stripe boundary, and one
7186 * further up in new geometry must map after here in old
7188 * If the chunk sizes are different, then as we perform reshape
7189 * in units of the largest of the two, reshape_position needs
7190 * be a multiple of the largest chunk size times new data disks.
7192 here_new = mddev->reshape_position;
7193 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7194 new_data_disks = mddev->raid_disks - max_degraded;
7195 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7196 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7200 reshape_offset = here_new * chunk_sectors;
7201 /* here_new is the stripe we will write to */
7202 here_old = mddev->reshape_position;
7203 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7204 /* here_old is the first stripe that we might need to read
7206 if (mddev->delta_disks == 0) {
7207 /* We cannot be sure it is safe to start an in-place
7208 * reshape. It is only safe if user-space is monitoring
7209 * and taking constant backups.
7210 * mdadm always starts a situation like this in
7211 * readonly mode so it can take control before
7212 * allowing any writes. So just check for that.
7214 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7215 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7216 /* not really in-place - so OK */;
7217 else if (mddev->ro == 0) {
7218 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7222 } else if (mddev->reshape_backwards
7223 ? (here_new * chunk_sectors + min_offset_diff <=
7224 here_old * chunk_sectors)
7225 : (here_new * chunk_sectors >=
7226 here_old * chunk_sectors + (-min_offset_diff))) {
7227 /* Reading from the same stripe as writing to - bad */
7228 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7232 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7233 /* OK, we should be able to continue; */
7235 BUG_ON(mddev->level != mddev->new_level);
7236 BUG_ON(mddev->layout != mddev->new_layout);
7237 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7238 BUG_ON(mddev->delta_disks != 0);
7241 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7242 test_bit(MD_HAS_PPL, &mddev->flags)) {
7243 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7245 clear_bit(MD_HAS_PPL, &mddev->flags);
7248 if (mddev->private == NULL)
7249 conf = setup_conf(mddev);
7251 conf = mddev->private;
7254 return PTR_ERR(conf);
7256 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7258 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7261 set_disk_ro(mddev->gendisk, 1);
7262 } else if (mddev->recovery_cp == MaxSector)
7263 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7266 conf->min_offset_diff = min_offset_diff;
7267 mddev->thread = conf->thread;
7268 conf->thread = NULL;
7269 mddev->private = conf;
7271 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7273 rdev = conf->disks[i].rdev;
7274 if (!rdev && conf->disks[i].replacement) {
7275 /* The replacement is all we have yet */
7276 rdev = conf->disks[i].replacement;
7277 conf->disks[i].replacement = NULL;
7278 clear_bit(Replacement, &rdev->flags);
7279 conf->disks[i].rdev = rdev;
7283 if (conf->disks[i].replacement &&
7284 conf->reshape_progress != MaxSector) {
7285 /* replacements and reshape simply do not mix. */
7286 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7289 if (test_bit(In_sync, &rdev->flags)) {
7293 /* This disc is not fully in-sync. However if it
7294 * just stored parity (beyond the recovery_offset),
7295 * when we don't need to be concerned about the
7296 * array being dirty.
7297 * When reshape goes 'backwards', we never have
7298 * partially completed devices, so we only need
7299 * to worry about reshape going forwards.
7301 /* Hack because v0.91 doesn't store recovery_offset properly. */
7302 if (mddev->major_version == 0 &&
7303 mddev->minor_version > 90)
7304 rdev->recovery_offset = reshape_offset;
7306 if (rdev->recovery_offset < reshape_offset) {
7307 /* We need to check old and new layout */
7308 if (!only_parity(rdev->raid_disk,
7311 conf->max_degraded))
7314 if (!only_parity(rdev->raid_disk,
7316 conf->previous_raid_disks,
7317 conf->max_degraded))
7319 dirty_parity_disks++;
7323 * 0 for a fully functional array, 1 or 2 for a degraded array.
7325 mddev->degraded = raid5_calc_degraded(conf);
7327 if (has_failed(conf)) {
7328 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7329 mdname(mddev), mddev->degraded, conf->raid_disks);
7333 /* device size must be a multiple of chunk size */
7334 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7335 mddev->resync_max_sectors = mddev->dev_sectors;
7337 if (mddev->degraded > dirty_parity_disks &&
7338 mddev->recovery_cp != MaxSector) {
7339 if (test_bit(MD_HAS_PPL, &mddev->flags))
7340 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7342 else if (mddev->ok_start_degraded)
7343 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7346 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7352 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7353 mdname(mddev), conf->level,
7354 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7357 print_raid5_conf(conf);
7359 if (conf->reshape_progress != MaxSector) {
7360 conf->reshape_safe = conf->reshape_progress;
7361 atomic_set(&conf->reshape_stripes, 0);
7362 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7363 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7364 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7365 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7366 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7370 /* Ok, everything is just fine now */
7371 if (mddev->to_remove == &raid5_attrs_group)
7372 mddev->to_remove = NULL;
7373 else if (mddev->kobj.sd &&
7374 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7375 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7377 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7381 /* read-ahead size must cover two whole stripes, which
7382 * is 2 * (datadisks) * chunksize where 'n' is the
7383 * number of raid devices
7385 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7386 int stripe = data_disks *
7387 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7388 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7389 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7391 chunk_size = mddev->chunk_sectors << 9;
7392 blk_queue_io_min(mddev->queue, chunk_size);
7393 blk_queue_io_opt(mddev->queue, chunk_size *
7394 (conf->raid_disks - conf->max_degraded));
7395 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7397 * We can only discard a whole stripe. It doesn't make sense to
7398 * discard data disk but write parity disk
7400 stripe = stripe * PAGE_SIZE;
7401 /* Round up to power of 2, as discard handling
7402 * currently assumes that */
7403 while ((stripe-1) & stripe)
7404 stripe = (stripe | (stripe-1)) + 1;
7405 mddev->queue->limits.discard_alignment = stripe;
7406 mddev->queue->limits.discard_granularity = stripe;
7408 blk_queue_max_write_same_sectors(mddev->queue, 0);
7409 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7411 rdev_for_each(rdev, mddev) {
7412 disk_stack_limits(mddev->gendisk, rdev->bdev,
7413 rdev->data_offset << 9);
7414 disk_stack_limits(mddev->gendisk, rdev->bdev,
7415 rdev->new_data_offset << 9);
7419 * zeroing is required, otherwise data
7420 * could be lost. Consider a scenario: discard a stripe
7421 * (the stripe could be inconsistent if
7422 * discard_zeroes_data is 0); write one disk of the
7423 * stripe (the stripe could be inconsistent again
7424 * depending on which disks are used to calculate
7425 * parity); the disk is broken; The stripe data of this
7428 * We only allow DISCARD if the sysadmin has confirmed that
7429 * only safe devices are in use by setting a module parameter.
7430 * A better idea might be to turn DISCARD into WRITE_ZEROES
7431 * requests, as that is required to be safe.
7433 if (devices_handle_discard_safely &&
7434 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7435 mddev->queue->limits.discard_granularity >= stripe)
7436 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7439 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7442 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7445 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7450 md_unregister_thread(&mddev->thread);
7451 print_raid5_conf(conf);
7453 mddev->private = NULL;
7454 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7458 static void raid5_free(struct mddev *mddev, void *priv)
7460 struct r5conf *conf = priv;
7463 mddev->to_remove = &raid5_attrs_group;
7466 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7468 struct r5conf *conf = mddev->private;
7471 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7472 conf->chunk_sectors / 2, mddev->layout);
7473 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7475 for (i = 0; i < conf->raid_disks; i++) {
7476 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7477 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7480 seq_printf (seq, "]");
7483 static void print_raid5_conf (struct r5conf *conf)
7486 struct disk_info *tmp;
7488 pr_debug("RAID conf printout:\n");
7490 pr_debug("(conf==NULL)\n");
7493 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7495 conf->raid_disks - conf->mddev->degraded);
7497 for (i = 0; i < conf->raid_disks; i++) {
7498 char b[BDEVNAME_SIZE];
7499 tmp = conf->disks + i;
7501 pr_debug(" disk %d, o:%d, dev:%s\n",
7502 i, !test_bit(Faulty, &tmp->rdev->flags),
7503 bdevname(tmp->rdev->bdev, b));
7507 static int raid5_spare_active(struct mddev *mddev)
7510 struct r5conf *conf = mddev->private;
7511 struct disk_info *tmp;
7513 unsigned long flags;
7515 for (i = 0; i < conf->raid_disks; i++) {
7516 tmp = conf->disks + i;
7517 if (tmp->replacement
7518 && tmp->replacement->recovery_offset == MaxSector
7519 && !test_bit(Faulty, &tmp->replacement->flags)
7520 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7521 /* Replacement has just become active. */
7523 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7526 /* Replaced device not technically faulty,
7527 * but we need to be sure it gets removed
7528 * and never re-added.
7530 set_bit(Faulty, &tmp->rdev->flags);
7531 sysfs_notify_dirent_safe(
7532 tmp->rdev->sysfs_state);
7534 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7535 } else if (tmp->rdev
7536 && tmp->rdev->recovery_offset == MaxSector
7537 && !test_bit(Faulty, &tmp->rdev->flags)
7538 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7540 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7543 spin_lock_irqsave(&conf->device_lock, flags);
7544 mddev->degraded = raid5_calc_degraded(conf);
7545 spin_unlock_irqrestore(&conf->device_lock, flags);
7546 print_raid5_conf(conf);
7550 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7552 struct r5conf *conf = mddev->private;
7554 int number = rdev->raid_disk;
7555 struct md_rdev **rdevp;
7556 struct disk_info *p = conf->disks + number;
7558 print_raid5_conf(conf);
7559 if (test_bit(Journal, &rdev->flags) && conf->log) {
7561 * we can't wait pending write here, as this is called in
7562 * raid5d, wait will deadlock.
7563 * neilb: there is no locking about new writes here,
7564 * so this cannot be safe.
7566 if (atomic_read(&conf->active_stripes) ||
7567 atomic_read(&conf->r5c_cached_full_stripes) ||
7568 atomic_read(&conf->r5c_cached_partial_stripes)) {
7574 if (rdev == p->rdev)
7576 else if (rdev == p->replacement)
7577 rdevp = &p->replacement;
7581 if (number >= conf->raid_disks &&
7582 conf->reshape_progress == MaxSector)
7583 clear_bit(In_sync, &rdev->flags);
7585 if (test_bit(In_sync, &rdev->flags) ||
7586 atomic_read(&rdev->nr_pending)) {
7590 /* Only remove non-faulty devices if recovery
7593 if (!test_bit(Faulty, &rdev->flags) &&
7594 mddev->recovery_disabled != conf->recovery_disabled &&
7595 !has_failed(conf) &&
7596 (!p->replacement || p->replacement == rdev) &&
7597 number < conf->raid_disks) {
7602 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7604 if (atomic_read(&rdev->nr_pending)) {
7605 /* lost the race, try later */
7611 err = log_modify(conf, rdev, false);
7615 if (p->replacement) {
7616 /* We must have just cleared 'rdev' */
7617 p->rdev = p->replacement;
7618 clear_bit(Replacement, &p->replacement->flags);
7619 smp_mb(); /* Make sure other CPUs may see both as identical
7620 * but will never see neither - if they are careful
7622 p->replacement = NULL;
7625 err = log_modify(conf, p->rdev, true);
7628 clear_bit(WantReplacement, &rdev->flags);
7631 print_raid5_conf(conf);
7635 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7637 struct r5conf *conf = mddev->private;
7640 struct disk_info *p;
7642 int last = conf->raid_disks - 1;
7644 if (test_bit(Journal, &rdev->flags)) {
7648 rdev->raid_disk = 0;
7650 * The array is in readonly mode if journal is missing, so no
7651 * write requests running. We should be safe
7653 log_init(conf, rdev, false);
7656 if (mddev->recovery_disabled == conf->recovery_disabled)
7659 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7660 /* no point adding a device */
7663 if (rdev->raid_disk >= 0)
7664 first = last = rdev->raid_disk;
7667 * find the disk ... but prefer rdev->saved_raid_disk
7670 if (rdev->saved_raid_disk >= 0 &&
7671 rdev->saved_raid_disk >= first &&
7672 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7673 first = rdev->saved_raid_disk;
7675 for (disk = first; disk <= last; disk++) {
7676 p = conf->disks + disk;
7677 if (p->rdev == NULL) {
7678 clear_bit(In_sync, &rdev->flags);
7679 rdev->raid_disk = disk;
7680 if (rdev->saved_raid_disk != disk)
7682 rcu_assign_pointer(p->rdev, rdev);
7684 err = log_modify(conf, rdev, true);
7689 for (disk = first; disk <= last; disk++) {
7690 p = conf->disks + disk;
7691 if (test_bit(WantReplacement, &p->rdev->flags) &&
7692 p->replacement == NULL) {
7693 clear_bit(In_sync, &rdev->flags);
7694 set_bit(Replacement, &rdev->flags);
7695 rdev->raid_disk = disk;
7698 rcu_assign_pointer(p->replacement, rdev);
7703 print_raid5_conf(conf);
7707 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7709 /* no resync is happening, and there is enough space
7710 * on all devices, so we can resize.
7711 * We need to make sure resync covers any new space.
7712 * If the array is shrinking we should possibly wait until
7713 * any io in the removed space completes, but it hardly seems
7717 struct r5conf *conf = mddev->private;
7719 if (conf->log || raid5_has_ppl(conf))
7721 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7722 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7723 if (mddev->external_size &&
7724 mddev->array_sectors > newsize)
7726 if (mddev->bitmap) {
7727 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7731 md_set_array_sectors(mddev, newsize);
7732 if (sectors > mddev->dev_sectors &&
7733 mddev->recovery_cp > mddev->dev_sectors) {
7734 mddev->recovery_cp = mddev->dev_sectors;
7735 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7737 mddev->dev_sectors = sectors;
7738 mddev->resync_max_sectors = sectors;
7742 static int check_stripe_cache(struct mddev *mddev)
7744 /* Can only proceed if there are plenty of stripe_heads.
7745 * We need a minimum of one full stripe,, and for sensible progress
7746 * it is best to have about 4 times that.
7747 * If we require 4 times, then the default 256 4K stripe_heads will
7748 * allow for chunk sizes up to 256K, which is probably OK.
7749 * If the chunk size is greater, user-space should request more
7750 * stripe_heads first.
7752 struct r5conf *conf = mddev->private;
7753 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7754 > conf->min_nr_stripes ||
7755 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7756 > conf->min_nr_stripes) {
7757 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7759 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7766 static int check_reshape(struct mddev *mddev)
7768 struct r5conf *conf = mddev->private;
7770 if (conf->log || raid5_has_ppl(conf))
7772 if (mddev->delta_disks == 0 &&
7773 mddev->new_layout == mddev->layout &&
7774 mddev->new_chunk_sectors == mddev->chunk_sectors)
7775 return 0; /* nothing to do */
7776 if (has_failed(conf))
7778 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7779 /* We might be able to shrink, but the devices must
7780 * be made bigger first.
7781 * For raid6, 4 is the minimum size.
7782 * Otherwise 2 is the minimum
7785 if (mddev->level == 6)
7787 if (mddev->raid_disks + mddev->delta_disks < min)
7791 if (!check_stripe_cache(mddev))
7794 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7795 mddev->delta_disks > 0)
7796 if (resize_chunks(conf,
7797 conf->previous_raid_disks
7798 + max(0, mddev->delta_disks),
7799 max(mddev->new_chunk_sectors,
7800 mddev->chunk_sectors)
7804 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7805 return 0; /* never bother to shrink */
7806 return resize_stripes(conf, (conf->previous_raid_disks
7807 + mddev->delta_disks));
7810 static int raid5_start_reshape(struct mddev *mddev)
7812 struct r5conf *conf = mddev->private;
7813 struct md_rdev *rdev;
7815 unsigned long flags;
7817 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7820 if (!check_stripe_cache(mddev))
7823 if (has_failed(conf))
7826 rdev_for_each(rdev, mddev) {
7827 if (!test_bit(In_sync, &rdev->flags)
7828 && !test_bit(Faulty, &rdev->flags))
7832 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7833 /* Not enough devices even to make a degraded array
7838 /* Refuse to reduce size of the array. Any reductions in
7839 * array size must be through explicit setting of array_size
7842 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7843 < mddev->array_sectors) {
7844 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7849 atomic_set(&conf->reshape_stripes, 0);
7850 spin_lock_irq(&conf->device_lock);
7851 write_seqcount_begin(&conf->gen_lock);
7852 conf->previous_raid_disks = conf->raid_disks;
7853 conf->raid_disks += mddev->delta_disks;
7854 conf->prev_chunk_sectors = conf->chunk_sectors;
7855 conf->chunk_sectors = mddev->new_chunk_sectors;
7856 conf->prev_algo = conf->algorithm;
7857 conf->algorithm = mddev->new_layout;
7859 /* Code that selects data_offset needs to see the generation update
7860 * if reshape_progress has been set - so a memory barrier needed.
7863 if (mddev->reshape_backwards)
7864 conf->reshape_progress = raid5_size(mddev, 0, 0);
7866 conf->reshape_progress = 0;
7867 conf->reshape_safe = conf->reshape_progress;
7868 write_seqcount_end(&conf->gen_lock);
7869 spin_unlock_irq(&conf->device_lock);
7871 /* Now make sure any requests that proceeded on the assumption
7872 * the reshape wasn't running - like Discard or Read - have
7875 mddev_suspend(mddev);
7876 mddev_resume(mddev);
7878 /* Add some new drives, as many as will fit.
7879 * We know there are enough to make the newly sized array work.
7880 * Don't add devices if we are reducing the number of
7881 * devices in the array. This is because it is not possible
7882 * to correctly record the "partially reconstructed" state of
7883 * such devices during the reshape and confusion could result.
7885 if (mddev->delta_disks >= 0) {
7886 rdev_for_each(rdev, mddev)
7887 if (rdev->raid_disk < 0 &&
7888 !test_bit(Faulty, &rdev->flags)) {
7889 if (raid5_add_disk(mddev, rdev) == 0) {
7891 >= conf->previous_raid_disks)
7892 set_bit(In_sync, &rdev->flags);
7894 rdev->recovery_offset = 0;
7896 if (sysfs_link_rdev(mddev, rdev))
7897 /* Failure here is OK */;
7899 } else if (rdev->raid_disk >= conf->previous_raid_disks
7900 && !test_bit(Faulty, &rdev->flags)) {
7901 /* This is a spare that was manually added */
7902 set_bit(In_sync, &rdev->flags);
7905 /* When a reshape changes the number of devices,
7906 * ->degraded is measured against the larger of the
7907 * pre and post number of devices.
7909 spin_lock_irqsave(&conf->device_lock, flags);
7910 mddev->degraded = raid5_calc_degraded(conf);
7911 spin_unlock_irqrestore(&conf->device_lock, flags);
7913 mddev->raid_disks = conf->raid_disks;
7914 mddev->reshape_position = conf->reshape_progress;
7915 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7917 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7918 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7919 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7920 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7921 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7922 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7924 if (!mddev->sync_thread) {
7925 mddev->recovery = 0;
7926 spin_lock_irq(&conf->device_lock);
7927 write_seqcount_begin(&conf->gen_lock);
7928 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7929 mddev->new_chunk_sectors =
7930 conf->chunk_sectors = conf->prev_chunk_sectors;
7931 mddev->new_layout = conf->algorithm = conf->prev_algo;
7932 rdev_for_each(rdev, mddev)
7933 rdev->new_data_offset = rdev->data_offset;
7935 conf->generation --;
7936 conf->reshape_progress = MaxSector;
7937 mddev->reshape_position = MaxSector;
7938 write_seqcount_end(&conf->gen_lock);
7939 spin_unlock_irq(&conf->device_lock);
7942 conf->reshape_checkpoint = jiffies;
7943 md_wakeup_thread(mddev->sync_thread);
7944 md_new_event(mddev);
7948 /* This is called from the reshape thread and should make any
7949 * changes needed in 'conf'
7951 static void end_reshape(struct r5conf *conf)
7954 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7956 spin_lock_irq(&conf->device_lock);
7957 conf->previous_raid_disks = conf->raid_disks;
7958 md_finish_reshape(conf->mddev);
7960 conf->reshape_progress = MaxSector;
7961 conf->mddev->reshape_position = MaxSector;
7962 spin_unlock_irq(&conf->device_lock);
7963 wake_up(&conf->wait_for_overlap);
7965 /* read-ahead size must cover two whole stripes, which is
7966 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7968 if (conf->mddev->queue) {
7969 int data_disks = conf->raid_disks - conf->max_degraded;
7970 int stripe = data_disks * ((conf->chunk_sectors << 9)
7972 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7973 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7978 /* This is called from the raid5d thread with mddev_lock held.
7979 * It makes config changes to the device.
7981 static void raid5_finish_reshape(struct mddev *mddev)
7983 struct r5conf *conf = mddev->private;
7985 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7987 if (mddev->delta_disks > 0) {
7988 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7990 set_capacity(mddev->gendisk, mddev->array_sectors);
7991 revalidate_disk(mddev->gendisk);
7995 spin_lock_irq(&conf->device_lock);
7996 mddev->degraded = raid5_calc_degraded(conf);
7997 spin_unlock_irq(&conf->device_lock);
7998 for (d = conf->raid_disks ;
7999 d < conf->raid_disks - mddev->delta_disks;
8001 struct md_rdev *rdev = conf->disks[d].rdev;
8003 clear_bit(In_sync, &rdev->flags);
8004 rdev = conf->disks[d].replacement;
8006 clear_bit(In_sync, &rdev->flags);
8009 mddev->layout = conf->algorithm;
8010 mddev->chunk_sectors = conf->chunk_sectors;
8011 mddev->reshape_position = MaxSector;
8012 mddev->delta_disks = 0;
8013 mddev->reshape_backwards = 0;
8017 static void raid5_quiesce(struct mddev *mddev, int state)
8019 struct r5conf *conf = mddev->private;
8022 case 2: /* resume for a suspend */
8023 wake_up(&conf->wait_for_overlap);
8026 case 1: /* stop all writes */
8027 lock_all_device_hash_locks_irq(conf);
8028 /* '2' tells resync/reshape to pause so that all
8029 * active stripes can drain
8031 r5c_flush_cache(conf, INT_MAX);
8033 wait_event_cmd(conf->wait_for_quiescent,
8034 atomic_read(&conf->active_stripes) == 0 &&
8035 atomic_read(&conf->active_aligned_reads) == 0,
8036 unlock_all_device_hash_locks_irq(conf),
8037 lock_all_device_hash_locks_irq(conf));
8039 unlock_all_device_hash_locks_irq(conf);
8040 /* allow reshape to continue */
8041 wake_up(&conf->wait_for_overlap);
8044 case 0: /* re-enable writes */
8045 lock_all_device_hash_locks_irq(conf);
8047 wake_up(&conf->wait_for_quiescent);
8048 wake_up(&conf->wait_for_overlap);
8049 unlock_all_device_hash_locks_irq(conf);
8052 r5l_quiesce(conf->log, state);
8055 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8057 struct r0conf *raid0_conf = mddev->private;
8060 /* for raid0 takeover only one zone is supported */
8061 if (raid0_conf->nr_strip_zones > 1) {
8062 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8064 return ERR_PTR(-EINVAL);
8067 sectors = raid0_conf->strip_zone[0].zone_end;
8068 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8069 mddev->dev_sectors = sectors;
8070 mddev->new_level = level;
8071 mddev->new_layout = ALGORITHM_PARITY_N;
8072 mddev->new_chunk_sectors = mddev->chunk_sectors;
8073 mddev->raid_disks += 1;
8074 mddev->delta_disks = 1;
8075 /* make sure it will be not marked as dirty */
8076 mddev->recovery_cp = MaxSector;
8078 return setup_conf(mddev);
8081 static void *raid5_takeover_raid1(struct mddev *mddev)
8086 if (mddev->raid_disks != 2 ||
8087 mddev->degraded > 1)
8088 return ERR_PTR(-EINVAL);
8090 /* Should check if there are write-behind devices? */
8092 chunksect = 64*2; /* 64K by default */
8094 /* The array must be an exact multiple of chunksize */
8095 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8098 if ((chunksect<<9) < STRIPE_SIZE)
8099 /* array size does not allow a suitable chunk size */
8100 return ERR_PTR(-EINVAL);
8102 mddev->new_level = 5;
8103 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8104 mddev->new_chunk_sectors = chunksect;
8106 ret = setup_conf(mddev);
8108 mddev_clear_unsupported_flags(mddev,
8109 UNSUPPORTED_MDDEV_FLAGS);
8113 static void *raid5_takeover_raid6(struct mddev *mddev)
8117 switch (mddev->layout) {
8118 case ALGORITHM_LEFT_ASYMMETRIC_6:
8119 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8121 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8122 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8124 case ALGORITHM_LEFT_SYMMETRIC_6:
8125 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8127 case ALGORITHM_RIGHT_SYMMETRIC_6:
8128 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8130 case ALGORITHM_PARITY_0_6:
8131 new_layout = ALGORITHM_PARITY_0;
8133 case ALGORITHM_PARITY_N:
8134 new_layout = ALGORITHM_PARITY_N;
8137 return ERR_PTR(-EINVAL);
8139 mddev->new_level = 5;
8140 mddev->new_layout = new_layout;
8141 mddev->delta_disks = -1;
8142 mddev->raid_disks -= 1;
8143 return setup_conf(mddev);
8146 static int raid5_check_reshape(struct mddev *mddev)
8148 /* For a 2-drive array, the layout and chunk size can be changed
8149 * immediately as not restriping is needed.
8150 * For larger arrays we record the new value - after validation
8151 * to be used by a reshape pass.
8153 struct r5conf *conf = mddev->private;
8154 int new_chunk = mddev->new_chunk_sectors;
8156 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8158 if (new_chunk > 0) {
8159 if (!is_power_of_2(new_chunk))
8161 if (new_chunk < (PAGE_SIZE>>9))
8163 if (mddev->array_sectors & (new_chunk-1))
8164 /* not factor of array size */
8168 /* They look valid */
8170 if (mddev->raid_disks == 2) {
8171 /* can make the change immediately */
8172 if (mddev->new_layout >= 0) {
8173 conf->algorithm = mddev->new_layout;
8174 mddev->layout = mddev->new_layout;
8176 if (new_chunk > 0) {
8177 conf->chunk_sectors = new_chunk ;
8178 mddev->chunk_sectors = new_chunk;
8180 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8181 md_wakeup_thread(mddev->thread);
8183 return check_reshape(mddev);
8186 static int raid6_check_reshape(struct mddev *mddev)
8188 int new_chunk = mddev->new_chunk_sectors;
8190 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8192 if (new_chunk > 0) {
8193 if (!is_power_of_2(new_chunk))
8195 if (new_chunk < (PAGE_SIZE >> 9))
8197 if (mddev->array_sectors & (new_chunk-1))
8198 /* not factor of array size */
8202 /* They look valid */
8203 return check_reshape(mddev);
8206 static void *raid5_takeover(struct mddev *mddev)
8208 /* raid5 can take over:
8209 * raid0 - if there is only one strip zone - make it a raid4 layout
8210 * raid1 - if there are two drives. We need to know the chunk size
8211 * raid4 - trivial - just use a raid4 layout.
8212 * raid6 - Providing it is a *_6 layout
8214 if (mddev->level == 0)
8215 return raid45_takeover_raid0(mddev, 5);
8216 if (mddev->level == 1)
8217 return raid5_takeover_raid1(mddev);
8218 if (mddev->level == 4) {
8219 mddev->new_layout = ALGORITHM_PARITY_N;
8220 mddev->new_level = 5;
8221 return setup_conf(mddev);
8223 if (mddev->level == 6)
8224 return raid5_takeover_raid6(mddev);
8226 return ERR_PTR(-EINVAL);
8229 static void *raid4_takeover(struct mddev *mddev)
8231 /* raid4 can take over:
8232 * raid0 - if there is only one strip zone
8233 * raid5 - if layout is right
8235 if (mddev->level == 0)
8236 return raid45_takeover_raid0(mddev, 4);
8237 if (mddev->level == 5 &&
8238 mddev->layout == ALGORITHM_PARITY_N) {
8239 mddev->new_layout = 0;
8240 mddev->new_level = 4;
8241 return setup_conf(mddev);
8243 return ERR_PTR(-EINVAL);
8246 static struct md_personality raid5_personality;
8248 static void *raid6_takeover(struct mddev *mddev)
8250 /* Currently can only take over a raid5. We map the
8251 * personality to an equivalent raid6 personality
8252 * with the Q block at the end.
8256 if (mddev->pers != &raid5_personality)
8257 return ERR_PTR(-EINVAL);
8258 if (mddev->degraded > 1)
8259 return ERR_PTR(-EINVAL);
8260 if (mddev->raid_disks > 253)
8261 return ERR_PTR(-EINVAL);
8262 if (mddev->raid_disks < 3)
8263 return ERR_PTR(-EINVAL);
8265 switch (mddev->layout) {
8266 case ALGORITHM_LEFT_ASYMMETRIC:
8267 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8269 case ALGORITHM_RIGHT_ASYMMETRIC:
8270 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8272 case ALGORITHM_LEFT_SYMMETRIC:
8273 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8275 case ALGORITHM_RIGHT_SYMMETRIC:
8276 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8278 case ALGORITHM_PARITY_0:
8279 new_layout = ALGORITHM_PARITY_0_6;
8281 case ALGORITHM_PARITY_N:
8282 new_layout = ALGORITHM_PARITY_N;
8285 return ERR_PTR(-EINVAL);
8287 mddev->new_level = 6;
8288 mddev->new_layout = new_layout;
8289 mddev->delta_disks = 1;
8290 mddev->raid_disks += 1;
8291 return setup_conf(mddev);
8294 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8296 struct r5conf *conf;
8299 err = mddev_lock(mddev);
8302 conf = mddev->private;
8304 mddev_unlock(mddev);
8308 if (strncmp(buf, "ppl", 3) == 0) {
8309 /* ppl only works with RAID 5 */
8310 if (!raid5_has_ppl(conf) && conf->level == 5) {
8311 err = log_init(conf, NULL, true);
8313 err = resize_stripes(conf, conf->pool_size);
8319 } else if (strncmp(buf, "resync", 6) == 0) {
8320 if (raid5_has_ppl(conf)) {
8321 mddev_suspend(mddev);
8323 mddev_resume(mddev);
8324 err = resize_stripes(conf, conf->pool_size);
8325 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8326 r5l_log_disk_error(conf)) {
8327 bool journal_dev_exists = false;
8328 struct md_rdev *rdev;
8330 rdev_for_each(rdev, mddev)
8331 if (test_bit(Journal, &rdev->flags)) {
8332 journal_dev_exists = true;
8336 if (!journal_dev_exists) {
8337 mddev_suspend(mddev);
8338 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8339 mddev_resume(mddev);
8340 } else /* need remove journal device first */
8349 md_update_sb(mddev, 1);
8351 mddev_unlock(mddev);
8356 static struct md_personality raid6_personality =
8360 .owner = THIS_MODULE,
8361 .make_request = raid5_make_request,
8364 .status = raid5_status,
8365 .error_handler = raid5_error,
8366 .hot_add_disk = raid5_add_disk,
8367 .hot_remove_disk= raid5_remove_disk,
8368 .spare_active = raid5_spare_active,
8369 .sync_request = raid5_sync_request,
8370 .resize = raid5_resize,
8372 .check_reshape = raid6_check_reshape,
8373 .start_reshape = raid5_start_reshape,
8374 .finish_reshape = raid5_finish_reshape,
8375 .quiesce = raid5_quiesce,
8376 .takeover = raid6_takeover,
8377 .congested = raid5_congested,
8378 .change_consistency_policy = raid5_change_consistency_policy,
8380 static struct md_personality raid5_personality =
8384 .owner = THIS_MODULE,
8385 .make_request = raid5_make_request,
8388 .status = raid5_status,
8389 .error_handler = raid5_error,
8390 .hot_add_disk = raid5_add_disk,
8391 .hot_remove_disk= raid5_remove_disk,
8392 .spare_active = raid5_spare_active,
8393 .sync_request = raid5_sync_request,
8394 .resize = raid5_resize,
8396 .check_reshape = raid5_check_reshape,
8397 .start_reshape = raid5_start_reshape,
8398 .finish_reshape = raid5_finish_reshape,
8399 .quiesce = raid5_quiesce,
8400 .takeover = raid5_takeover,
8401 .congested = raid5_congested,
8402 .change_consistency_policy = raid5_change_consistency_policy,
8405 static struct md_personality raid4_personality =
8409 .owner = THIS_MODULE,
8410 .make_request = raid5_make_request,
8413 .status = raid5_status,
8414 .error_handler = raid5_error,
8415 .hot_add_disk = raid5_add_disk,
8416 .hot_remove_disk= raid5_remove_disk,
8417 .spare_active = raid5_spare_active,
8418 .sync_request = raid5_sync_request,
8419 .resize = raid5_resize,
8421 .check_reshape = raid5_check_reshape,
8422 .start_reshape = raid5_start_reshape,
8423 .finish_reshape = raid5_finish_reshape,
8424 .quiesce = raid5_quiesce,
8425 .takeover = raid4_takeover,
8426 .congested = raid5_congested,
8427 .change_consistency_policy = raid5_change_consistency_policy,
8430 static int __init raid5_init(void)
8434 raid5_wq = alloc_workqueue("raid5wq",
8435 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8439 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8441 raid456_cpu_up_prepare,
8444 destroy_workqueue(raid5_wq);
8447 register_md_personality(&raid6_personality);
8448 register_md_personality(&raid5_personality);
8449 register_md_personality(&raid4_personality);
8453 static void raid5_exit(void)
8455 unregister_md_personality(&raid6_personality);
8456 unregister_md_personality(&raid5_personality);
8457 unregister_md_personality(&raid4_personality);
8458 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8459 destroy_workqueue(raid5_wq);
8462 module_init(raid5_init);
8463 module_exit(raid5_exit);
8464 MODULE_LICENSE("GPL");
8465 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8466 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8467 MODULE_ALIAS("md-raid5");
8468 MODULE_ALIAS("md-raid4");
8469 MODULE_ALIAS("md-level-5");
8470 MODULE_ALIAS("md-level-4");
8471 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8472 MODULE_ALIAS("md-raid6");
8473 MODULE_ALIAS("md-level-6");
8475 /* This used to be two separate modules, they were: */
8476 MODULE_ALIAS("raid5");
8477 MODULE_ALIAS("raid6");