1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * raid5.c : Multiple Devices driver for Linux
4 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
5 * Copyright (C) 1999, 2000 Ingo Molnar
6 * Copyright (C) 2002, 2003 H. Peter Anvin
8 * RAID-4/5/6 management functions.
9 * Thanks to Penguin Computing for making the RAID-6 development possible
10 * by donating a test server!
16 * The sequencing for updating the bitmap reliably is a little
17 * subtle (and I got it wrong the first time) so it deserves some
20 * We group bitmap updates into batches. Each batch has a number.
21 * We may write out several batches at once, but that isn't very important.
22 * conf->seq_write is the number of the last batch successfully written.
23 * conf->seq_flush is the number of the last batch that was closed to
25 * When we discover that we will need to write to any block in a stripe
26 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
27 * the number of the batch it will be in. This is seq_flush+1.
28 * When we are ready to do a write, if that batch hasn't been written yet,
29 * we plug the array and queue the stripe for later.
30 * When an unplug happens, we increment bm_flush, thus closing the current
32 * When we notice that bm_flush > bm_write, we write out all pending updates
33 * to the bitmap, and advance bm_write to where bm_flush was.
34 * This may occasionally write a bit out twice, but is sure never to
38 #include <linux/blkdev.h>
39 #include <linux/delay.h>
40 #include <linux/kthread.h>
41 #include <linux/raid/pq.h>
42 #include <linux/async_tx.h>
43 #include <linux/module.h>
44 #include <linux/async.h>
45 #include <linux/seq_file.h>
46 #include <linux/cpu.h>
47 #include <linux/slab.h>
48 #include <linux/ratelimit.h>
49 #include <linux/nodemask.h>
51 #include <trace/events/block.h>
52 #include <linux/list_sort.h>
57 #include "md-bitmap.h"
58 #include "raid5-log.h"
60 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
62 #define cpu_to_group(cpu) cpu_to_node(cpu)
63 #define ANY_GROUP NUMA_NO_NODE
65 static bool devices_handle_discard_safely = false;
66 module_param(devices_handle_discard_safely, bool, 0644);
67 MODULE_PARM_DESC(devices_handle_discard_safely,
68 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
69 static struct workqueue_struct *raid5_wq;
71 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
73 int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
74 return &conf->stripe_hashtbl[hash];
77 static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
79 return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
82 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
84 spin_lock_irq(conf->hash_locks + hash);
85 spin_lock(&conf->device_lock);
88 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
90 spin_unlock(&conf->device_lock);
91 spin_unlock_irq(conf->hash_locks + hash);
94 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
97 spin_lock_irq(conf->hash_locks);
98 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
99 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
100 spin_lock(&conf->device_lock);
103 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
106 spin_unlock(&conf->device_lock);
107 for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
108 spin_unlock(conf->hash_locks + i);
109 spin_unlock_irq(conf->hash_locks);
112 /* Find first data disk in a raid6 stripe */
113 static inline int raid6_d0(struct stripe_head *sh)
116 /* ddf always start from first device */
118 /* md starts just after Q block */
119 if (sh->qd_idx == sh->disks - 1)
122 return sh->qd_idx + 1;
124 static inline int raid6_next_disk(int disk, int raid_disks)
127 return (disk < raid_disks) ? disk : 0;
130 /* When walking through the disks in a raid5, starting at raid6_d0,
131 * We need to map each disk to a 'slot', where the data disks are slot
132 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
133 * is raid_disks-1. This help does that mapping.
135 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
136 int *count, int syndrome_disks)
142 if (idx == sh->pd_idx)
143 return syndrome_disks;
144 if (idx == sh->qd_idx)
145 return syndrome_disks + 1;
151 static void print_raid5_conf (struct r5conf *conf);
153 static int stripe_operations_active(struct stripe_head *sh)
155 return sh->check_state || sh->reconstruct_state ||
156 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
157 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
160 static bool stripe_is_lowprio(struct stripe_head *sh)
162 return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
163 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
164 !test_bit(STRIPE_R5C_CACHING, &sh->state);
167 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
169 struct r5conf *conf = sh->raid_conf;
170 struct r5worker_group *group;
172 int i, cpu = sh->cpu;
174 if (!cpu_online(cpu)) {
175 cpu = cpumask_any(cpu_online_mask);
179 if (list_empty(&sh->lru)) {
180 struct r5worker_group *group;
181 group = conf->worker_groups + cpu_to_group(cpu);
182 if (stripe_is_lowprio(sh))
183 list_add_tail(&sh->lru, &group->loprio_list);
185 list_add_tail(&sh->lru, &group->handle_list);
186 group->stripes_cnt++;
190 if (conf->worker_cnt_per_group == 0) {
191 md_wakeup_thread(conf->mddev->thread);
195 group = conf->worker_groups + cpu_to_group(sh->cpu);
197 group->workers[0].working = true;
198 /* at least one worker should run to avoid race */
199 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
201 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
202 /* wakeup more workers */
203 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
204 if (group->workers[i].working == false) {
205 group->workers[i].working = true;
206 queue_work_on(sh->cpu, raid5_wq,
207 &group->workers[i].work);
213 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
214 struct list_head *temp_inactive_list)
217 int injournal = 0; /* number of date pages with R5_InJournal */
219 BUG_ON(!list_empty(&sh->lru));
220 BUG_ON(atomic_read(&conf->active_stripes)==0);
222 if (r5c_is_writeback(conf->log))
223 for (i = sh->disks; i--; )
224 if (test_bit(R5_InJournal, &sh->dev[i].flags))
227 * In the following cases, the stripe cannot be released to cached
228 * lists. Therefore, we make the stripe write out and set
230 * 1. when quiesce in r5c write back;
231 * 2. when resync is requested fot the stripe.
233 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
234 (conf->quiesce && r5c_is_writeback(conf->log) &&
235 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
236 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
237 r5c_make_stripe_write_out(sh);
238 set_bit(STRIPE_HANDLE, &sh->state);
241 if (test_bit(STRIPE_HANDLE, &sh->state)) {
242 if (test_bit(STRIPE_DELAYED, &sh->state) &&
243 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
244 list_add_tail(&sh->lru, &conf->delayed_list);
245 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
246 sh->bm_seq - conf->seq_write > 0)
247 list_add_tail(&sh->lru, &conf->bitmap_list);
249 clear_bit(STRIPE_DELAYED, &sh->state);
250 clear_bit(STRIPE_BIT_DELAY, &sh->state);
251 if (conf->worker_cnt_per_group == 0) {
252 if (stripe_is_lowprio(sh))
253 list_add_tail(&sh->lru,
256 list_add_tail(&sh->lru,
259 raid5_wakeup_stripe_thread(sh);
263 md_wakeup_thread(conf->mddev->thread);
265 BUG_ON(stripe_operations_active(sh));
266 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
267 if (atomic_dec_return(&conf->preread_active_stripes)
269 md_wakeup_thread(conf->mddev->thread);
270 atomic_dec(&conf->active_stripes);
271 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
272 if (!r5c_is_writeback(conf->log))
273 list_add_tail(&sh->lru, temp_inactive_list);
275 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
277 list_add_tail(&sh->lru, temp_inactive_list);
278 else if (injournal == conf->raid_disks - conf->max_degraded) {
280 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
281 atomic_inc(&conf->r5c_cached_full_stripes);
282 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
283 atomic_dec(&conf->r5c_cached_partial_stripes);
284 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
285 r5c_check_cached_full_stripe(conf);
288 * STRIPE_R5C_PARTIAL_STRIPE is set in
289 * r5c_try_caching_write(). No need to
292 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
298 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
299 struct list_head *temp_inactive_list)
301 if (atomic_dec_and_test(&sh->count))
302 do_release_stripe(conf, sh, temp_inactive_list);
306 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
308 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
309 * given time. Adding stripes only takes device lock, while deleting stripes
310 * only takes hash lock.
312 static void release_inactive_stripe_list(struct r5conf *conf,
313 struct list_head *temp_inactive_list,
317 bool do_wakeup = false;
320 if (hash == NR_STRIPE_HASH_LOCKS) {
321 size = NR_STRIPE_HASH_LOCKS;
322 hash = NR_STRIPE_HASH_LOCKS - 1;
326 struct list_head *list = &temp_inactive_list[size - 1];
329 * We don't hold any lock here yet, raid5_get_active_stripe() might
330 * remove stripes from the list
332 if (!list_empty_careful(list)) {
333 spin_lock_irqsave(conf->hash_locks + hash, flags);
334 if (list_empty(conf->inactive_list + hash) &&
336 atomic_dec(&conf->empty_inactive_list_nr);
337 list_splice_tail_init(list, conf->inactive_list + hash);
339 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
346 wake_up(&conf->wait_for_stripe);
347 if (atomic_read(&conf->active_stripes) == 0)
348 wake_up(&conf->wait_for_quiescent);
349 if (conf->retry_read_aligned)
350 md_wakeup_thread(conf->mddev->thread);
354 /* should hold conf->device_lock already */
355 static int release_stripe_list(struct r5conf *conf,
356 struct list_head *temp_inactive_list)
358 struct stripe_head *sh, *t;
360 struct llist_node *head;
362 head = llist_del_all(&conf->released_stripes);
363 head = llist_reverse_order(head);
364 llist_for_each_entry_safe(sh, t, head, release_list) {
367 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
369 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
371 * Don't worry the bit is set here, because if the bit is set
372 * again, the count is always > 1. This is true for
373 * STRIPE_ON_UNPLUG_LIST bit too.
375 hash = sh->hash_lock_index;
376 __release_stripe(conf, sh, &temp_inactive_list[hash]);
383 void raid5_release_stripe(struct stripe_head *sh)
385 struct r5conf *conf = sh->raid_conf;
387 struct list_head list;
391 /* Avoid release_list until the last reference.
393 if (atomic_add_unless(&sh->count, -1, 1))
396 if (unlikely(!conf->mddev->thread) ||
397 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
399 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
401 md_wakeup_thread(conf->mddev->thread);
404 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
405 if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
406 INIT_LIST_HEAD(&list);
407 hash = sh->hash_lock_index;
408 do_release_stripe(conf, sh, &list);
409 spin_unlock_irqrestore(&conf->device_lock, flags);
410 release_inactive_stripe_list(conf, &list, hash);
414 static inline void remove_hash(struct stripe_head *sh)
416 pr_debug("remove_hash(), stripe %llu\n",
417 (unsigned long long)sh->sector);
419 hlist_del_init(&sh->hash);
422 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
424 struct hlist_head *hp = stripe_hash(conf, sh->sector);
426 pr_debug("insert_hash(), stripe %llu\n",
427 (unsigned long long)sh->sector);
429 hlist_add_head(&sh->hash, hp);
432 /* find an idle stripe, make sure it is unhashed, and return it. */
433 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
435 struct stripe_head *sh = NULL;
436 struct list_head *first;
438 if (list_empty(conf->inactive_list + hash))
440 first = (conf->inactive_list + hash)->next;
441 sh = list_entry(first, struct stripe_head, lru);
442 list_del_init(first);
444 atomic_inc(&conf->active_stripes);
445 BUG_ON(hash != sh->hash_lock_index);
446 if (list_empty(conf->inactive_list + hash))
447 atomic_inc(&conf->empty_inactive_list_nr);
452 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
453 static void free_stripe_pages(struct stripe_head *sh)
458 /* Have not allocate page pool */
462 for (i = 0; i < sh->nr_pages; i++) {
470 static int alloc_stripe_pages(struct stripe_head *sh, gfp_t gfp)
475 for (i = 0; i < sh->nr_pages; i++) {
476 /* The page have allocated. */
482 free_stripe_pages(sh);
491 init_stripe_shared_pages(struct stripe_head *sh, struct r5conf *conf, int disks)
498 /* Each of the sh->dev[i] need one conf->stripe_size */
499 cnt = PAGE_SIZE / conf->stripe_size;
500 nr_pages = (disks + cnt - 1) / cnt;
502 sh->pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
505 sh->nr_pages = nr_pages;
506 sh->stripes_per_page = cnt;
511 static void shrink_buffers(struct stripe_head *sh)
514 int num = sh->raid_conf->pool_size;
516 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
517 for (i = 0; i < num ; i++) {
520 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
524 sh->dev[i].page = NULL;
528 for (i = 0; i < num; i++)
529 sh->dev[i].page = NULL;
530 free_stripe_pages(sh); /* Free pages */
534 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
537 int num = sh->raid_conf->pool_size;
539 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
540 for (i = 0; i < num; i++) {
543 if (!(page = alloc_page(gfp))) {
546 sh->dev[i].page = page;
547 sh->dev[i].orig_page = page;
548 sh->dev[i].offset = 0;
551 if (alloc_stripe_pages(sh, gfp))
554 for (i = 0; i < num; i++) {
555 sh->dev[i].page = raid5_get_dev_page(sh, i);
556 sh->dev[i].orig_page = sh->dev[i].page;
557 sh->dev[i].offset = raid5_get_page_offset(sh, i);
563 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
564 struct stripe_head *sh);
566 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
568 struct r5conf *conf = sh->raid_conf;
571 BUG_ON(atomic_read(&sh->count) != 0);
572 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
573 BUG_ON(stripe_operations_active(sh));
574 BUG_ON(sh->batch_head);
576 pr_debug("init_stripe called, stripe %llu\n",
577 (unsigned long long)sector);
579 seq = read_seqcount_begin(&conf->gen_lock);
580 sh->generation = conf->generation - previous;
581 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
583 stripe_set_idx(sector, conf, previous, sh);
586 for (i = sh->disks; i--; ) {
587 struct r5dev *dev = &sh->dev[i];
589 if (dev->toread || dev->read || dev->towrite || dev->written ||
590 test_bit(R5_LOCKED, &dev->flags)) {
591 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
592 (unsigned long long)sh->sector, i, dev->toread,
593 dev->read, dev->towrite, dev->written,
594 test_bit(R5_LOCKED, &dev->flags));
598 dev->sector = raid5_compute_blocknr(sh, i, previous);
600 if (read_seqcount_retry(&conf->gen_lock, seq))
602 sh->overwrite_disks = 0;
603 insert_hash(conf, sh);
604 sh->cpu = smp_processor_id();
605 set_bit(STRIPE_BATCH_READY, &sh->state);
608 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
611 struct stripe_head *sh;
613 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
614 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
615 if (sh->sector == sector && sh->generation == generation)
617 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
622 * Need to check if array has failed when deciding whether to:
624 * - remove non-faulty devices
627 * This determination is simple when no reshape is happening.
628 * However if there is a reshape, we need to carefully check
629 * both the before and after sections.
630 * This is because some failed devices may only affect one
631 * of the two sections, and some non-in_sync devices may
632 * be insync in the section most affected by failed devices.
634 int raid5_calc_degraded(struct r5conf *conf)
636 int degraded, degraded2;
641 for (i = 0; i < conf->previous_raid_disks; i++) {
642 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
643 if (rdev && test_bit(Faulty, &rdev->flags))
644 rdev = rcu_dereference(conf->disks[i].replacement);
645 if (!rdev || test_bit(Faulty, &rdev->flags))
647 else if (test_bit(In_sync, &rdev->flags))
650 /* not in-sync or faulty.
651 * If the reshape increases the number of devices,
652 * this is being recovered by the reshape, so
653 * this 'previous' section is not in_sync.
654 * If the number of devices is being reduced however,
655 * the device can only be part of the array if
656 * we are reverting a reshape, so this section will
659 if (conf->raid_disks >= conf->previous_raid_disks)
663 if (conf->raid_disks == conf->previous_raid_disks)
667 for (i = 0; i < conf->raid_disks; i++) {
668 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
669 if (rdev && test_bit(Faulty, &rdev->flags))
670 rdev = rcu_dereference(conf->disks[i].replacement);
671 if (!rdev || test_bit(Faulty, &rdev->flags))
673 else if (test_bit(In_sync, &rdev->flags))
676 /* not in-sync or faulty.
677 * If reshape increases the number of devices, this
678 * section has already been recovered, else it
679 * almost certainly hasn't.
681 if (conf->raid_disks <= conf->previous_raid_disks)
685 if (degraded2 > degraded)
690 static bool has_failed(struct r5conf *conf)
692 int degraded = conf->mddev->degraded;
694 if (test_bit(MD_BROKEN, &conf->mddev->flags))
697 if (conf->mddev->reshape_position != MaxSector)
698 degraded = raid5_calc_degraded(conf);
700 return degraded > conf->max_degraded;
704 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
705 int previous, int noblock, int noquiesce)
707 struct stripe_head *sh;
708 int hash = stripe_hash_locks_hash(conf, sector);
709 int inc_empty_inactive_list_flag;
711 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
713 spin_lock_irq(conf->hash_locks + hash);
716 wait_event_lock_irq(conf->wait_for_quiescent,
717 conf->quiesce == 0 || noquiesce,
718 *(conf->hash_locks + hash));
719 sh = __find_stripe(conf, sector, conf->generation - previous);
721 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
722 sh = get_free_stripe(conf, hash);
723 if (!sh && !test_bit(R5_DID_ALLOC,
725 set_bit(R5_ALLOC_MORE,
728 if (noblock && sh == NULL)
731 r5c_check_stripe_cache_usage(conf);
733 set_bit(R5_INACTIVE_BLOCKED,
735 r5l_wake_reclaim(conf->log, 0);
737 conf->wait_for_stripe,
738 !list_empty(conf->inactive_list + hash) &&
739 (atomic_read(&conf->active_stripes)
740 < (conf->max_nr_stripes * 3 / 4)
741 || !test_bit(R5_INACTIVE_BLOCKED,
742 &conf->cache_state)),
743 *(conf->hash_locks + hash));
744 clear_bit(R5_INACTIVE_BLOCKED,
747 init_stripe(sh, sector, previous);
748 atomic_inc(&sh->count);
750 } else if (!atomic_inc_not_zero(&sh->count)) {
751 spin_lock(&conf->device_lock);
752 if (!atomic_read(&sh->count)) {
753 if (!test_bit(STRIPE_HANDLE, &sh->state))
754 atomic_inc(&conf->active_stripes);
755 BUG_ON(list_empty(&sh->lru) &&
756 !test_bit(STRIPE_EXPANDING, &sh->state));
757 inc_empty_inactive_list_flag = 0;
758 if (!list_empty(conf->inactive_list + hash))
759 inc_empty_inactive_list_flag = 1;
760 list_del_init(&sh->lru);
761 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
762 atomic_inc(&conf->empty_inactive_list_nr);
764 sh->group->stripes_cnt--;
768 atomic_inc(&sh->count);
769 spin_unlock(&conf->device_lock);
771 } while (sh == NULL);
773 spin_unlock_irq(conf->hash_locks + hash);
777 static bool is_full_stripe_write(struct stripe_head *sh)
779 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
780 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
783 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
784 __acquires(&sh1->stripe_lock)
785 __acquires(&sh2->stripe_lock)
788 spin_lock_irq(&sh2->stripe_lock);
789 spin_lock_nested(&sh1->stripe_lock, 1);
791 spin_lock_irq(&sh1->stripe_lock);
792 spin_lock_nested(&sh2->stripe_lock, 1);
796 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
797 __releases(&sh1->stripe_lock)
798 __releases(&sh2->stripe_lock)
800 spin_unlock(&sh1->stripe_lock);
801 spin_unlock_irq(&sh2->stripe_lock);
804 /* Only freshly new full stripe normal write stripe can be added to a batch list */
805 static bool stripe_can_batch(struct stripe_head *sh)
807 struct r5conf *conf = sh->raid_conf;
809 if (raid5_has_log(conf) || raid5_has_ppl(conf))
811 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
812 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
813 is_full_stripe_write(sh);
816 /* we only do back search */
817 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
819 struct stripe_head *head;
820 sector_t head_sector, tmp_sec;
823 int inc_empty_inactive_list_flag;
825 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
826 tmp_sec = sh->sector;
827 if (!sector_div(tmp_sec, conf->chunk_sectors))
829 head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
831 hash = stripe_hash_locks_hash(conf, head_sector);
832 spin_lock_irq(conf->hash_locks + hash);
833 head = __find_stripe(conf, head_sector, conf->generation);
834 if (head && !atomic_inc_not_zero(&head->count)) {
835 spin_lock(&conf->device_lock);
836 if (!atomic_read(&head->count)) {
837 if (!test_bit(STRIPE_HANDLE, &head->state))
838 atomic_inc(&conf->active_stripes);
839 BUG_ON(list_empty(&head->lru) &&
840 !test_bit(STRIPE_EXPANDING, &head->state));
841 inc_empty_inactive_list_flag = 0;
842 if (!list_empty(conf->inactive_list + hash))
843 inc_empty_inactive_list_flag = 1;
844 list_del_init(&head->lru);
845 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
846 atomic_inc(&conf->empty_inactive_list_nr);
848 head->group->stripes_cnt--;
852 atomic_inc(&head->count);
853 spin_unlock(&conf->device_lock);
855 spin_unlock_irq(conf->hash_locks + hash);
859 if (!stripe_can_batch(head))
862 lock_two_stripes(head, sh);
863 /* clear_batch_ready clear the flag */
864 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
871 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
873 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
874 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
877 if (head->batch_head) {
878 spin_lock(&head->batch_head->batch_lock);
879 /* This batch list is already running */
880 if (!stripe_can_batch(head)) {
881 spin_unlock(&head->batch_head->batch_lock);
885 * We must assign batch_head of this stripe within the
886 * batch_lock, otherwise clear_batch_ready of batch head
887 * stripe could clear BATCH_READY bit of this stripe and
888 * this stripe->batch_head doesn't get assigned, which
889 * could confuse clear_batch_ready for this stripe
891 sh->batch_head = head->batch_head;
894 * at this point, head's BATCH_READY could be cleared, but we
895 * can still add the stripe to batch list
897 list_add(&sh->batch_list, &head->batch_list);
898 spin_unlock(&head->batch_head->batch_lock);
900 head->batch_head = head;
901 sh->batch_head = head->batch_head;
902 spin_lock(&head->batch_lock);
903 list_add_tail(&sh->batch_list, &head->batch_list);
904 spin_unlock(&head->batch_lock);
907 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
908 if (atomic_dec_return(&conf->preread_active_stripes)
910 md_wakeup_thread(conf->mddev->thread);
912 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
913 int seq = sh->bm_seq;
914 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
915 sh->batch_head->bm_seq > seq)
916 seq = sh->batch_head->bm_seq;
917 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
918 sh->batch_head->bm_seq = seq;
921 atomic_inc(&sh->count);
923 unlock_two_stripes(head, sh);
925 raid5_release_stripe(head);
928 /* Determine if 'data_offset' or 'new_data_offset' should be used
929 * in this stripe_head.
931 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
933 sector_t progress = conf->reshape_progress;
934 /* Need a memory barrier to make sure we see the value
935 * of conf->generation, or ->data_offset that was set before
936 * reshape_progress was updated.
939 if (progress == MaxSector)
941 if (sh->generation == conf->generation - 1)
943 /* We are in a reshape, and this is a new-generation stripe,
944 * so use new_data_offset.
949 static void dispatch_bio_list(struct bio_list *tmp)
953 while ((bio = bio_list_pop(tmp)))
954 submit_bio_noacct(bio);
957 static int cmp_stripe(void *priv, const struct list_head *a,
958 const struct list_head *b)
960 const struct r5pending_data *da = list_entry(a,
961 struct r5pending_data, sibling);
962 const struct r5pending_data *db = list_entry(b,
963 struct r5pending_data, sibling);
964 if (da->sector > db->sector)
966 if (da->sector < db->sector)
971 static void dispatch_defer_bios(struct r5conf *conf, int target,
972 struct bio_list *list)
974 struct r5pending_data *data;
975 struct list_head *first, *next = NULL;
978 if (conf->pending_data_cnt == 0)
981 list_sort(NULL, &conf->pending_list, cmp_stripe);
983 first = conf->pending_list.next;
985 /* temporarily move the head */
986 if (conf->next_pending_data)
987 list_move_tail(&conf->pending_list,
988 &conf->next_pending_data->sibling);
990 while (!list_empty(&conf->pending_list)) {
991 data = list_first_entry(&conf->pending_list,
992 struct r5pending_data, sibling);
993 if (&data->sibling == first)
994 first = data->sibling.next;
995 next = data->sibling.next;
997 bio_list_merge(list, &data->bios);
998 list_move(&data->sibling, &conf->free_list);
1003 conf->pending_data_cnt -= cnt;
1004 BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
1006 if (next != &conf->pending_list)
1007 conf->next_pending_data = list_entry(next,
1008 struct r5pending_data, sibling);
1010 conf->next_pending_data = NULL;
1011 /* list isn't empty */
1012 if (first != &conf->pending_list)
1013 list_move_tail(&conf->pending_list, first);
1016 static void flush_deferred_bios(struct r5conf *conf)
1018 struct bio_list tmp = BIO_EMPTY_LIST;
1020 if (conf->pending_data_cnt == 0)
1023 spin_lock(&conf->pending_bios_lock);
1024 dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
1025 BUG_ON(conf->pending_data_cnt != 0);
1026 spin_unlock(&conf->pending_bios_lock);
1028 dispatch_bio_list(&tmp);
1031 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
1032 struct bio_list *bios)
1034 struct bio_list tmp = BIO_EMPTY_LIST;
1035 struct r5pending_data *ent;
1037 spin_lock(&conf->pending_bios_lock);
1038 ent = list_first_entry(&conf->free_list, struct r5pending_data,
1040 list_move_tail(&ent->sibling, &conf->pending_list);
1041 ent->sector = sector;
1042 bio_list_init(&ent->bios);
1043 bio_list_merge(&ent->bios, bios);
1044 conf->pending_data_cnt++;
1045 if (conf->pending_data_cnt >= PENDING_IO_MAX)
1046 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
1048 spin_unlock(&conf->pending_bios_lock);
1050 dispatch_bio_list(&tmp);
1054 raid5_end_read_request(struct bio *bi);
1056 raid5_end_write_request(struct bio *bi);
1058 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
1060 struct r5conf *conf = sh->raid_conf;
1061 int i, disks = sh->disks;
1062 struct stripe_head *head_sh = sh;
1063 struct bio_list pending_bios = BIO_EMPTY_LIST;
1068 if (log_stripe(sh, s) == 0)
1071 should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1073 for (i = disks; i--; ) {
1074 int op, op_flags = 0;
1075 int replace_only = 0;
1076 struct bio *bi, *rbi;
1077 struct md_rdev *rdev, *rrdev = NULL;
1080 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1082 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1084 if (test_bit(R5_Discard, &sh->dev[i].flags))
1085 op = REQ_OP_DISCARD;
1086 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1088 else if (test_and_clear_bit(R5_WantReplace,
1089 &sh->dev[i].flags)) {
1094 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1095 op_flags |= REQ_SYNC;
1098 bi = &sh->dev[i].req;
1099 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1102 rrdev = rcu_dereference(conf->disks[i].replacement);
1103 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1104 rdev = rcu_dereference(conf->disks[i].rdev);
1109 if (op_is_write(op)) {
1113 /* We raced and saw duplicates */
1116 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1121 if (rdev && test_bit(Faulty, &rdev->flags))
1124 atomic_inc(&rdev->nr_pending);
1125 if (rrdev && test_bit(Faulty, &rrdev->flags))
1128 atomic_inc(&rrdev->nr_pending);
1131 /* We have already checked bad blocks for reads. Now
1132 * need to check for writes. We never accept write errors
1133 * on the replacement, so we don't to check rrdev.
1135 while (op_is_write(op) && rdev &&
1136 test_bit(WriteErrorSeen, &rdev->flags)) {
1139 int bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
1140 &first_bad, &bad_sectors);
1145 set_bit(BlockedBadBlocks, &rdev->flags);
1146 if (!conf->mddev->external &&
1147 conf->mddev->sb_flags) {
1148 /* It is very unlikely, but we might
1149 * still need to write out the
1150 * bad block log - better give it
1152 md_check_recovery(conf->mddev);
1155 * Because md_wait_for_blocked_rdev
1156 * will dec nr_pending, we must
1157 * increment it first.
1159 atomic_inc(&rdev->nr_pending);
1160 md_wait_for_blocked_rdev(rdev, conf->mddev);
1162 /* Acknowledged bad block - skip the write */
1163 rdev_dec_pending(rdev, conf->mddev);
1169 if (s->syncing || s->expanding || s->expanded
1171 md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
1173 set_bit(STRIPE_IO_STARTED, &sh->state);
1175 bio_set_dev(bi, rdev->bdev);
1176 bio_set_op_attrs(bi, op, op_flags);
1177 bi->bi_end_io = op_is_write(op)
1178 ? raid5_end_write_request
1179 : raid5_end_read_request;
1180 bi->bi_private = sh;
1182 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1183 __func__, (unsigned long long)sh->sector,
1185 atomic_inc(&sh->count);
1187 atomic_inc(&head_sh->count);
1188 if (use_new_offset(conf, sh))
1189 bi->bi_iter.bi_sector = (sh->sector
1190 + rdev->new_data_offset);
1192 bi->bi_iter.bi_sector = (sh->sector
1193 + rdev->data_offset);
1194 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1195 bi->bi_opf |= REQ_NOMERGE;
1197 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1198 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1200 if (!op_is_write(op) &&
1201 test_bit(R5_InJournal, &sh->dev[i].flags))
1203 * issuing read for a page in journal, this
1204 * must be preparing for prexor in rmw; read
1205 * the data into orig_page
1207 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1209 sh->dev[i].vec.bv_page = sh->dev[i].page;
1211 bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1212 bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1213 bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1214 bi->bi_write_hint = sh->dev[i].write_hint;
1216 sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1218 * If this is discard request, set bi_vcnt 0. We don't
1219 * want to confuse SCSI because SCSI will replace payload
1221 if (op == REQ_OP_DISCARD)
1224 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1226 if (conf->mddev->gendisk)
1227 trace_block_bio_remap(bi,
1228 disk_devt(conf->mddev->gendisk),
1230 if (should_defer && op_is_write(op))
1231 bio_list_add(&pending_bios, bi);
1233 submit_bio_noacct(bi);
1236 if (s->syncing || s->expanding || s->expanded
1238 md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
1240 set_bit(STRIPE_IO_STARTED, &sh->state);
1242 bio_set_dev(rbi, rrdev->bdev);
1243 bio_set_op_attrs(rbi, op, op_flags);
1244 BUG_ON(!op_is_write(op));
1245 rbi->bi_end_io = raid5_end_write_request;
1246 rbi->bi_private = sh;
1248 pr_debug("%s: for %llu schedule op %d on "
1249 "replacement disc %d\n",
1250 __func__, (unsigned long long)sh->sector,
1252 atomic_inc(&sh->count);
1254 atomic_inc(&head_sh->count);
1255 if (use_new_offset(conf, sh))
1256 rbi->bi_iter.bi_sector = (sh->sector
1257 + rrdev->new_data_offset);
1259 rbi->bi_iter.bi_sector = (sh->sector
1260 + rrdev->data_offset);
1261 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1262 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1263 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1265 rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1266 rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1267 rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1268 rbi->bi_write_hint = sh->dev[i].write_hint;
1269 sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1271 * If this is discard request, set bi_vcnt 0. We don't
1272 * want to confuse SCSI because SCSI will replace payload
1274 if (op == REQ_OP_DISCARD)
1276 if (conf->mddev->gendisk)
1277 trace_block_bio_remap(rbi,
1278 disk_devt(conf->mddev->gendisk),
1280 if (should_defer && op_is_write(op))
1281 bio_list_add(&pending_bios, rbi);
1283 submit_bio_noacct(rbi);
1285 if (!rdev && !rrdev) {
1286 if (op_is_write(op))
1287 set_bit(STRIPE_DEGRADED, &sh->state);
1288 pr_debug("skip op %d on disc %d for sector %llu\n",
1289 bi->bi_opf, i, (unsigned long long)sh->sector);
1290 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1291 set_bit(STRIPE_HANDLE, &sh->state);
1294 if (!head_sh->batch_head)
1296 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1302 if (should_defer && !bio_list_empty(&pending_bios))
1303 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1306 static struct dma_async_tx_descriptor *
1307 async_copy_data(int frombio, struct bio *bio, struct page **page,
1308 unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1309 struct stripe_head *sh, int no_skipcopy)
1312 struct bvec_iter iter;
1313 struct page *bio_page;
1315 struct async_submit_ctl submit;
1316 enum async_tx_flags flags = 0;
1317 struct r5conf *conf = sh->raid_conf;
1319 if (bio->bi_iter.bi_sector >= sector)
1320 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1322 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1325 flags |= ASYNC_TX_FENCE;
1326 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1328 bio_for_each_segment(bvl, bio, iter) {
1329 int len = bvl.bv_len;
1333 if (page_offset < 0) {
1334 b_offset = -page_offset;
1335 page_offset += b_offset;
1339 if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1340 clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1345 b_offset += bvl.bv_offset;
1346 bio_page = bvl.bv_page;
1348 if (conf->skip_copy &&
1349 b_offset == 0 && page_offset == 0 &&
1350 clen == RAID5_STRIPE_SIZE(conf) &&
1354 tx = async_memcpy(*page, bio_page, page_offset + poff,
1355 b_offset, clen, &submit);
1357 tx = async_memcpy(bio_page, *page, b_offset,
1358 page_offset + poff, clen, &submit);
1360 /* chain the operations */
1361 submit.depend_tx = tx;
1363 if (clen < len) /* hit end of page */
1371 static void ops_complete_biofill(void *stripe_head_ref)
1373 struct stripe_head *sh = stripe_head_ref;
1375 struct r5conf *conf = sh->raid_conf;
1377 pr_debug("%s: stripe %llu\n", __func__,
1378 (unsigned long long)sh->sector);
1380 /* clear completed biofills */
1381 for (i = sh->disks; i--; ) {
1382 struct r5dev *dev = &sh->dev[i];
1384 /* acknowledge completion of a biofill operation */
1385 /* and check if we need to reply to a read request,
1386 * new R5_Wantfill requests are held off until
1387 * !STRIPE_BIOFILL_RUN
1389 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1390 struct bio *rbi, *rbi2;
1395 while (rbi && rbi->bi_iter.bi_sector <
1396 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1397 rbi2 = r5_next_bio(conf, rbi, dev->sector);
1403 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1405 set_bit(STRIPE_HANDLE, &sh->state);
1406 raid5_release_stripe(sh);
1409 static void ops_run_biofill(struct stripe_head *sh)
1411 struct dma_async_tx_descriptor *tx = NULL;
1412 struct async_submit_ctl submit;
1414 struct r5conf *conf = sh->raid_conf;
1416 BUG_ON(sh->batch_head);
1417 pr_debug("%s: stripe %llu\n", __func__,
1418 (unsigned long long)sh->sector);
1420 for (i = sh->disks; i--; ) {
1421 struct r5dev *dev = &sh->dev[i];
1422 if (test_bit(R5_Wantfill, &dev->flags)) {
1424 spin_lock_irq(&sh->stripe_lock);
1425 dev->read = rbi = dev->toread;
1427 spin_unlock_irq(&sh->stripe_lock);
1428 while (rbi && rbi->bi_iter.bi_sector <
1429 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1430 tx = async_copy_data(0, rbi, &dev->page,
1432 dev->sector, tx, sh, 0);
1433 rbi = r5_next_bio(conf, rbi, dev->sector);
1438 atomic_inc(&sh->count);
1439 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1440 async_trigger_callback(&submit);
1443 static void mark_target_uptodate(struct stripe_head *sh, int target)
1450 tgt = &sh->dev[target];
1451 set_bit(R5_UPTODATE, &tgt->flags);
1452 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1453 clear_bit(R5_Wantcompute, &tgt->flags);
1456 static void ops_complete_compute(void *stripe_head_ref)
1458 struct stripe_head *sh = stripe_head_ref;
1460 pr_debug("%s: stripe %llu\n", __func__,
1461 (unsigned long long)sh->sector);
1463 /* mark the computed target(s) as uptodate */
1464 mark_target_uptodate(sh, sh->ops.target);
1465 mark_target_uptodate(sh, sh->ops.target2);
1467 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1468 if (sh->check_state == check_state_compute_run)
1469 sh->check_state = check_state_compute_result;
1470 set_bit(STRIPE_HANDLE, &sh->state);
1471 raid5_release_stripe(sh);
1474 /* return a pointer to the address conversion region of the scribble buffer */
1475 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1477 return percpu->scribble + i * percpu->scribble_obj_size;
1480 /* return a pointer to the address conversion region of the scribble buffer */
1481 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1482 struct raid5_percpu *percpu, int i)
1484 return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1488 * Return a pointer to record offset address.
1490 static unsigned int *
1491 to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1493 return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1496 static struct dma_async_tx_descriptor *
1497 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1499 int disks = sh->disks;
1500 struct page **xor_srcs = to_addr_page(percpu, 0);
1501 unsigned int *off_srcs = to_addr_offs(sh, percpu);
1502 int target = sh->ops.target;
1503 struct r5dev *tgt = &sh->dev[target];
1504 struct page *xor_dest = tgt->page;
1505 unsigned int off_dest = tgt->offset;
1507 struct dma_async_tx_descriptor *tx;
1508 struct async_submit_ctl submit;
1511 BUG_ON(sh->batch_head);
1513 pr_debug("%s: stripe %llu block: %d\n",
1514 __func__, (unsigned long long)sh->sector, target);
1515 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1517 for (i = disks; i--; ) {
1519 off_srcs[count] = sh->dev[i].offset;
1520 xor_srcs[count++] = sh->dev[i].page;
1524 atomic_inc(&sh->count);
1526 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1527 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1528 if (unlikely(count == 1))
1529 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1530 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1532 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1533 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1538 /* set_syndrome_sources - populate source buffers for gen_syndrome
1539 * @srcs - (struct page *) array of size sh->disks
1540 * @offs - (unsigned int) array of offset for each page
1541 * @sh - stripe_head to parse
1543 * Populates srcs in proper layout order for the stripe and returns the
1544 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1545 * destination buffer is recorded in srcs[count] and the Q destination
1546 * is recorded in srcs[count+1]].
1548 static int set_syndrome_sources(struct page **srcs,
1550 struct stripe_head *sh,
1553 int disks = sh->disks;
1554 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1555 int d0_idx = raid6_d0(sh);
1559 for (i = 0; i < disks; i++)
1565 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1566 struct r5dev *dev = &sh->dev[i];
1568 if (i == sh->qd_idx || i == sh->pd_idx ||
1569 (srctype == SYNDROME_SRC_ALL) ||
1570 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1571 (test_bit(R5_Wantdrain, &dev->flags) ||
1572 test_bit(R5_InJournal, &dev->flags))) ||
1573 (srctype == SYNDROME_SRC_WRITTEN &&
1575 test_bit(R5_InJournal, &dev->flags)))) {
1576 if (test_bit(R5_InJournal, &dev->flags))
1577 srcs[slot] = sh->dev[i].orig_page;
1579 srcs[slot] = sh->dev[i].page;
1581 * For R5_InJournal, PAGE_SIZE must be 4KB and will
1582 * not shared page. In that case, dev[i].offset
1585 offs[slot] = sh->dev[i].offset;
1587 i = raid6_next_disk(i, disks);
1588 } while (i != d0_idx);
1590 return syndrome_disks;
1593 static struct dma_async_tx_descriptor *
1594 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1596 int disks = sh->disks;
1597 struct page **blocks = to_addr_page(percpu, 0);
1598 unsigned int *offs = to_addr_offs(sh, percpu);
1600 int qd_idx = sh->qd_idx;
1601 struct dma_async_tx_descriptor *tx;
1602 struct async_submit_ctl submit;
1605 unsigned int dest_off;
1609 BUG_ON(sh->batch_head);
1610 if (sh->ops.target < 0)
1611 target = sh->ops.target2;
1612 else if (sh->ops.target2 < 0)
1613 target = sh->ops.target;
1615 /* we should only have one valid target */
1618 pr_debug("%s: stripe %llu block: %d\n",
1619 __func__, (unsigned long long)sh->sector, target);
1621 tgt = &sh->dev[target];
1622 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1624 dest_off = tgt->offset;
1626 atomic_inc(&sh->count);
1628 if (target == qd_idx) {
1629 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1630 blocks[count] = NULL; /* regenerating p is not necessary */
1631 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1632 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1633 ops_complete_compute, sh,
1634 to_addr_conv(sh, percpu, 0));
1635 tx = async_gen_syndrome(blocks, offs, count+2,
1636 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1638 /* Compute any data- or p-drive using XOR */
1640 for (i = disks; i-- ; ) {
1641 if (i == target || i == qd_idx)
1643 offs[count] = sh->dev[i].offset;
1644 blocks[count++] = sh->dev[i].page;
1647 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1648 NULL, ops_complete_compute, sh,
1649 to_addr_conv(sh, percpu, 0));
1650 tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1651 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1657 static struct dma_async_tx_descriptor *
1658 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1660 int i, count, disks = sh->disks;
1661 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1662 int d0_idx = raid6_d0(sh);
1663 int faila = -1, failb = -1;
1664 int target = sh->ops.target;
1665 int target2 = sh->ops.target2;
1666 struct r5dev *tgt = &sh->dev[target];
1667 struct r5dev *tgt2 = &sh->dev[target2];
1668 struct dma_async_tx_descriptor *tx;
1669 struct page **blocks = to_addr_page(percpu, 0);
1670 unsigned int *offs = to_addr_offs(sh, percpu);
1671 struct async_submit_ctl submit;
1673 BUG_ON(sh->batch_head);
1674 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1675 __func__, (unsigned long long)sh->sector, target, target2);
1676 BUG_ON(target < 0 || target2 < 0);
1677 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1678 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1680 /* we need to open-code set_syndrome_sources to handle the
1681 * slot number conversion for 'faila' and 'failb'
1683 for (i = 0; i < disks ; i++) {
1690 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1692 offs[slot] = sh->dev[i].offset;
1693 blocks[slot] = sh->dev[i].page;
1699 i = raid6_next_disk(i, disks);
1700 } while (i != d0_idx);
1702 BUG_ON(faila == failb);
1705 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1706 __func__, (unsigned long long)sh->sector, faila, failb);
1708 atomic_inc(&sh->count);
1710 if (failb == syndrome_disks+1) {
1711 /* Q disk is one of the missing disks */
1712 if (faila == syndrome_disks) {
1713 /* Missing P+Q, just recompute */
1714 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1715 ops_complete_compute, sh,
1716 to_addr_conv(sh, percpu, 0));
1717 return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1718 RAID5_STRIPE_SIZE(sh->raid_conf),
1722 unsigned int dest_off;
1724 int qd_idx = sh->qd_idx;
1726 /* Missing D+Q: recompute D from P, then recompute Q */
1727 if (target == qd_idx)
1728 data_target = target2;
1730 data_target = target;
1733 for (i = disks; i-- ; ) {
1734 if (i == data_target || i == qd_idx)
1736 offs[count] = sh->dev[i].offset;
1737 blocks[count++] = sh->dev[i].page;
1739 dest = sh->dev[data_target].page;
1740 dest_off = sh->dev[data_target].offset;
1741 init_async_submit(&submit,
1742 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1744 to_addr_conv(sh, percpu, 0));
1745 tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1746 RAID5_STRIPE_SIZE(sh->raid_conf),
1749 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1750 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1751 ops_complete_compute, sh,
1752 to_addr_conv(sh, percpu, 0));
1753 return async_gen_syndrome(blocks, offs, count+2,
1754 RAID5_STRIPE_SIZE(sh->raid_conf),
1758 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1759 ops_complete_compute, sh,
1760 to_addr_conv(sh, percpu, 0));
1761 if (failb == syndrome_disks) {
1762 /* We're missing D+P. */
1763 return async_raid6_datap_recov(syndrome_disks+2,
1764 RAID5_STRIPE_SIZE(sh->raid_conf),
1766 blocks, offs, &submit);
1768 /* We're missing D+D. */
1769 return async_raid6_2data_recov(syndrome_disks+2,
1770 RAID5_STRIPE_SIZE(sh->raid_conf),
1772 blocks, offs, &submit);
1777 static void ops_complete_prexor(void *stripe_head_ref)
1779 struct stripe_head *sh = stripe_head_ref;
1781 pr_debug("%s: stripe %llu\n", __func__,
1782 (unsigned long long)sh->sector);
1784 if (r5c_is_writeback(sh->raid_conf->log))
1786 * raid5-cache write back uses orig_page during prexor.
1787 * After prexor, it is time to free orig_page
1789 r5c_release_extra_page(sh);
1792 static struct dma_async_tx_descriptor *
1793 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1794 struct dma_async_tx_descriptor *tx)
1796 int disks = sh->disks;
1797 struct page **xor_srcs = to_addr_page(percpu, 0);
1798 unsigned int *off_srcs = to_addr_offs(sh, percpu);
1799 int count = 0, pd_idx = sh->pd_idx, i;
1800 struct async_submit_ctl submit;
1802 /* existing parity data subtracted */
1803 unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1804 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1806 BUG_ON(sh->batch_head);
1807 pr_debug("%s: stripe %llu\n", __func__,
1808 (unsigned long long)sh->sector);
1810 for (i = disks; i--; ) {
1811 struct r5dev *dev = &sh->dev[i];
1812 /* Only process blocks that are known to be uptodate */
1813 if (test_bit(R5_InJournal, &dev->flags)) {
1815 * For this case, PAGE_SIZE must be equal to 4KB and
1816 * page offset is zero.
1818 off_srcs[count] = dev->offset;
1819 xor_srcs[count++] = dev->orig_page;
1820 } else if (test_bit(R5_Wantdrain, &dev->flags)) {
1821 off_srcs[count] = dev->offset;
1822 xor_srcs[count++] = dev->page;
1826 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1827 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1828 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1829 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1834 static struct dma_async_tx_descriptor *
1835 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1836 struct dma_async_tx_descriptor *tx)
1838 struct page **blocks = to_addr_page(percpu, 0);
1839 unsigned int *offs = to_addr_offs(sh, percpu);
1841 struct async_submit_ctl submit;
1843 pr_debug("%s: stripe %llu\n", __func__,
1844 (unsigned long long)sh->sector);
1846 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1848 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1849 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1850 tx = async_gen_syndrome(blocks, offs, count+2,
1851 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1856 static struct dma_async_tx_descriptor *
1857 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1859 struct r5conf *conf = sh->raid_conf;
1860 int disks = sh->disks;
1862 struct stripe_head *head_sh = sh;
1864 pr_debug("%s: stripe %llu\n", __func__,
1865 (unsigned long long)sh->sector);
1867 for (i = disks; i--; ) {
1872 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1878 * clear R5_InJournal, so when rewriting a page in
1879 * journal, it is not skipped by r5l_log_stripe()
1881 clear_bit(R5_InJournal, &dev->flags);
1882 spin_lock_irq(&sh->stripe_lock);
1883 chosen = dev->towrite;
1884 dev->towrite = NULL;
1885 sh->overwrite_disks = 0;
1886 BUG_ON(dev->written);
1887 wbi = dev->written = chosen;
1888 spin_unlock_irq(&sh->stripe_lock);
1889 WARN_ON(dev->page != dev->orig_page);
1891 while (wbi && wbi->bi_iter.bi_sector <
1892 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1893 if (wbi->bi_opf & REQ_FUA)
1894 set_bit(R5_WantFUA, &dev->flags);
1895 if (wbi->bi_opf & REQ_SYNC)
1896 set_bit(R5_SyncIO, &dev->flags);
1897 if (bio_op(wbi) == REQ_OP_DISCARD)
1898 set_bit(R5_Discard, &dev->flags);
1900 tx = async_copy_data(1, wbi, &dev->page,
1902 dev->sector, tx, sh,
1903 r5c_is_writeback(conf->log));
1904 if (dev->page != dev->orig_page &&
1905 !r5c_is_writeback(conf->log)) {
1906 set_bit(R5_SkipCopy, &dev->flags);
1907 clear_bit(R5_UPTODATE, &dev->flags);
1908 clear_bit(R5_OVERWRITE, &dev->flags);
1911 wbi = r5_next_bio(conf, wbi, dev->sector);
1914 if (head_sh->batch_head) {
1915 sh = list_first_entry(&sh->batch_list,
1928 static void ops_complete_reconstruct(void *stripe_head_ref)
1930 struct stripe_head *sh = stripe_head_ref;
1931 int disks = sh->disks;
1932 int pd_idx = sh->pd_idx;
1933 int qd_idx = sh->qd_idx;
1935 bool fua = false, sync = false, discard = false;
1937 pr_debug("%s: stripe %llu\n", __func__,
1938 (unsigned long long)sh->sector);
1940 for (i = disks; i--; ) {
1941 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1942 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1943 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1946 for (i = disks; i--; ) {
1947 struct r5dev *dev = &sh->dev[i];
1949 if (dev->written || i == pd_idx || i == qd_idx) {
1950 if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1951 set_bit(R5_UPTODATE, &dev->flags);
1952 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1953 set_bit(R5_Expanded, &dev->flags);
1956 set_bit(R5_WantFUA, &dev->flags);
1958 set_bit(R5_SyncIO, &dev->flags);
1962 if (sh->reconstruct_state == reconstruct_state_drain_run)
1963 sh->reconstruct_state = reconstruct_state_drain_result;
1964 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1965 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1967 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1968 sh->reconstruct_state = reconstruct_state_result;
1971 set_bit(STRIPE_HANDLE, &sh->state);
1972 raid5_release_stripe(sh);
1976 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1977 struct dma_async_tx_descriptor *tx)
1979 int disks = sh->disks;
1980 struct page **xor_srcs;
1981 unsigned int *off_srcs;
1982 struct async_submit_ctl submit;
1983 int count, pd_idx = sh->pd_idx, i;
1984 struct page *xor_dest;
1985 unsigned int off_dest;
1987 unsigned long flags;
1989 struct stripe_head *head_sh = sh;
1992 pr_debug("%s: stripe %llu\n", __func__,
1993 (unsigned long long)sh->sector);
1995 for (i = 0; i < sh->disks; i++) {
1998 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2001 if (i >= sh->disks) {
2002 atomic_inc(&sh->count);
2003 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2004 ops_complete_reconstruct(sh);
2009 xor_srcs = to_addr_page(percpu, j);
2010 off_srcs = to_addr_offs(sh, percpu);
2011 /* check if prexor is active which means only process blocks
2012 * that are part of a read-modify-write (written)
2014 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2016 off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2017 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2018 for (i = disks; i--; ) {
2019 struct r5dev *dev = &sh->dev[i];
2020 if (head_sh->dev[i].written ||
2021 test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2022 off_srcs[count] = dev->offset;
2023 xor_srcs[count++] = dev->page;
2027 xor_dest = sh->dev[pd_idx].page;
2028 off_dest = sh->dev[pd_idx].offset;
2029 for (i = disks; i--; ) {
2030 struct r5dev *dev = &sh->dev[i];
2032 off_srcs[count] = dev->offset;
2033 xor_srcs[count++] = dev->page;
2038 /* 1/ if we prexor'd then the dest is reused as a source
2039 * 2/ if we did not prexor then we are redoing the parity
2040 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2041 * for the synchronous xor case
2043 last_stripe = !head_sh->batch_head ||
2044 list_first_entry(&sh->batch_list,
2045 struct stripe_head, batch_list) == head_sh;
2047 flags = ASYNC_TX_ACK |
2048 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2050 atomic_inc(&head_sh->count);
2051 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2052 to_addr_conv(sh, percpu, j));
2054 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2055 init_async_submit(&submit, flags, tx, NULL, NULL,
2056 to_addr_conv(sh, percpu, j));
2059 if (unlikely(count == 1))
2060 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2061 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2063 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2064 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2067 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2074 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2075 struct dma_async_tx_descriptor *tx)
2077 struct async_submit_ctl submit;
2078 struct page **blocks;
2080 int count, i, j = 0;
2081 struct stripe_head *head_sh = sh;
2084 unsigned long txflags;
2086 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2088 for (i = 0; i < sh->disks; i++) {
2089 if (sh->pd_idx == i || sh->qd_idx == i)
2091 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2094 if (i >= sh->disks) {
2095 atomic_inc(&sh->count);
2096 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2097 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2098 ops_complete_reconstruct(sh);
2103 blocks = to_addr_page(percpu, j);
2104 offs = to_addr_offs(sh, percpu);
2106 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2107 synflags = SYNDROME_SRC_WRITTEN;
2108 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2110 synflags = SYNDROME_SRC_ALL;
2111 txflags = ASYNC_TX_ACK;
2114 count = set_syndrome_sources(blocks, offs, sh, synflags);
2115 last_stripe = !head_sh->batch_head ||
2116 list_first_entry(&sh->batch_list,
2117 struct stripe_head, batch_list) == head_sh;
2120 atomic_inc(&head_sh->count);
2121 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2122 head_sh, to_addr_conv(sh, percpu, j));
2124 init_async_submit(&submit, 0, tx, NULL, NULL,
2125 to_addr_conv(sh, percpu, j));
2126 tx = async_gen_syndrome(blocks, offs, count+2,
2127 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2130 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2136 static void ops_complete_check(void *stripe_head_ref)
2138 struct stripe_head *sh = stripe_head_ref;
2140 pr_debug("%s: stripe %llu\n", __func__,
2141 (unsigned long long)sh->sector);
2143 sh->check_state = check_state_check_result;
2144 set_bit(STRIPE_HANDLE, &sh->state);
2145 raid5_release_stripe(sh);
2148 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2150 int disks = sh->disks;
2151 int pd_idx = sh->pd_idx;
2152 int qd_idx = sh->qd_idx;
2153 struct page *xor_dest;
2154 unsigned int off_dest;
2155 struct page **xor_srcs = to_addr_page(percpu, 0);
2156 unsigned int *off_srcs = to_addr_offs(sh, percpu);
2157 struct dma_async_tx_descriptor *tx;
2158 struct async_submit_ctl submit;
2162 pr_debug("%s: stripe %llu\n", __func__,
2163 (unsigned long long)sh->sector);
2165 BUG_ON(sh->batch_head);
2167 xor_dest = sh->dev[pd_idx].page;
2168 off_dest = sh->dev[pd_idx].offset;
2169 off_srcs[count] = off_dest;
2170 xor_srcs[count++] = xor_dest;
2171 for (i = disks; i--; ) {
2172 if (i == pd_idx || i == qd_idx)
2174 off_srcs[count] = sh->dev[i].offset;
2175 xor_srcs[count++] = sh->dev[i].page;
2178 init_async_submit(&submit, 0, NULL, NULL, NULL,
2179 to_addr_conv(sh, percpu, 0));
2180 tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2181 RAID5_STRIPE_SIZE(sh->raid_conf),
2182 &sh->ops.zero_sum_result, &submit);
2184 atomic_inc(&sh->count);
2185 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2186 tx = async_trigger_callback(&submit);
2189 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2191 struct page **srcs = to_addr_page(percpu, 0);
2192 unsigned int *offs = to_addr_offs(sh, percpu);
2193 struct async_submit_ctl submit;
2196 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2197 (unsigned long long)sh->sector, checkp);
2199 BUG_ON(sh->batch_head);
2200 count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2204 atomic_inc(&sh->count);
2205 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2206 sh, to_addr_conv(sh, percpu, 0));
2207 async_syndrome_val(srcs, offs, count+2,
2208 RAID5_STRIPE_SIZE(sh->raid_conf),
2209 &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2212 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2214 int overlap_clear = 0, i, disks = sh->disks;
2215 struct dma_async_tx_descriptor *tx = NULL;
2216 struct r5conf *conf = sh->raid_conf;
2217 int level = conf->level;
2218 struct raid5_percpu *percpu;
2221 cpu = get_cpu_light();
2222 percpu = per_cpu_ptr(conf->percpu, cpu);
2223 spin_lock(&percpu->lock);
2224 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2225 ops_run_biofill(sh);
2229 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2231 tx = ops_run_compute5(sh, percpu);
2233 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2234 tx = ops_run_compute6_1(sh, percpu);
2236 tx = ops_run_compute6_2(sh, percpu);
2238 /* terminate the chain if reconstruct is not set to be run */
2239 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2243 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2245 tx = ops_run_prexor5(sh, percpu, tx);
2247 tx = ops_run_prexor6(sh, percpu, tx);
2250 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2251 tx = ops_run_partial_parity(sh, percpu, tx);
2253 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2254 tx = ops_run_biodrain(sh, tx);
2258 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2260 ops_run_reconstruct5(sh, percpu, tx);
2262 ops_run_reconstruct6(sh, percpu, tx);
2265 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2266 if (sh->check_state == check_state_run)
2267 ops_run_check_p(sh, percpu);
2268 else if (sh->check_state == check_state_run_q)
2269 ops_run_check_pq(sh, percpu, 0);
2270 else if (sh->check_state == check_state_run_pq)
2271 ops_run_check_pq(sh, percpu, 1);
2276 if (overlap_clear && !sh->batch_head)
2277 for (i = disks; i--; ) {
2278 struct r5dev *dev = &sh->dev[i];
2279 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2280 wake_up(&sh->raid_conf->wait_for_overlap);
2282 spin_unlock(&percpu->lock);
2286 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2288 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2292 __free_page(sh->ppl_page);
2293 kmem_cache_free(sc, sh);
2296 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2297 int disks, struct r5conf *conf)
2299 struct stripe_head *sh;
2302 sh = kmem_cache_zalloc(sc, gfp);
2304 spin_lock_init(&sh->stripe_lock);
2305 spin_lock_init(&sh->batch_lock);
2306 INIT_LIST_HEAD(&sh->batch_list);
2307 INIT_LIST_HEAD(&sh->lru);
2308 INIT_LIST_HEAD(&sh->r5c);
2309 INIT_LIST_HEAD(&sh->log_list);
2310 atomic_set(&sh->count, 1);
2311 sh->raid_conf = conf;
2312 sh->log_start = MaxSector;
2313 for (i = 0; i < disks; i++) {
2314 struct r5dev *dev = &sh->dev[i];
2316 bio_init(&dev->req, &dev->vec, 1);
2317 bio_init(&dev->rreq, &dev->rvec, 1);
2320 if (raid5_has_ppl(conf)) {
2321 sh->ppl_page = alloc_page(gfp);
2322 if (!sh->ppl_page) {
2323 free_stripe(sc, sh);
2327 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2328 if (init_stripe_shared_pages(sh, conf, disks)) {
2329 free_stripe(sc, sh);
2336 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2338 struct stripe_head *sh;
2340 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2344 if (grow_buffers(sh, gfp)) {
2346 free_stripe(conf->slab_cache, sh);
2349 sh->hash_lock_index =
2350 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2351 /* we just created an active stripe so... */
2352 atomic_inc(&conf->active_stripes);
2354 raid5_release_stripe(sh);
2355 conf->max_nr_stripes++;
2359 static int grow_stripes(struct r5conf *conf, int num)
2361 struct kmem_cache *sc;
2362 size_t namelen = sizeof(conf->cache_name[0]);
2363 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2365 if (conf->mddev->gendisk)
2366 snprintf(conf->cache_name[0], namelen,
2367 "raid%d-%s", conf->level, mdname(conf->mddev));
2369 snprintf(conf->cache_name[0], namelen,
2370 "raid%d-%p", conf->level, conf->mddev);
2371 snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2373 conf->active_name = 0;
2374 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2375 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2379 conf->slab_cache = sc;
2380 conf->pool_size = devs;
2382 if (!grow_one_stripe(conf, GFP_KERNEL))
2389 * scribble_alloc - allocate percpu scribble buffer for required size
2390 * of the scribble region
2391 * @percpu: from for_each_present_cpu() of the caller
2392 * @num: total number of disks in the array
2393 * @cnt: scribble objs count for required size of the scribble region
2395 * The scribble buffer size must be enough to contain:
2396 * 1/ a struct page pointer for each device in the array +2
2397 * 2/ room to convert each entry in (1) to its corresponding dma
2398 * (dma_map_page()) or page (page_address()) address.
2400 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2401 * calculate over all devices (not just the data blocks), using zeros in place
2402 * of the P and Q blocks.
2404 static int scribble_alloc(struct raid5_percpu *percpu,
2408 sizeof(struct page *) * (num + 2) +
2409 sizeof(addr_conv_t) * (num + 2) +
2410 sizeof(unsigned int) * (num + 2);
2414 * If here is in raid array suspend context, it is in memalloc noio
2415 * context as well, there is no potential recursive memory reclaim
2416 * I/Os with the GFP_KERNEL flag.
2418 scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2422 kvfree(percpu->scribble);
2424 percpu->scribble = scribble;
2425 percpu->scribble_obj_size = obj_size;
2429 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2435 * Never shrink. And mddev_suspend() could deadlock if this is called
2436 * from raid5d. In that case, scribble_disks and scribble_sectors
2437 * should equal to new_disks and new_sectors
2439 if (conf->scribble_disks >= new_disks &&
2440 conf->scribble_sectors >= new_sectors)
2442 mddev_suspend(conf->mddev);
2445 for_each_present_cpu(cpu) {
2446 struct raid5_percpu *percpu;
2448 percpu = per_cpu_ptr(conf->percpu, cpu);
2449 err = scribble_alloc(percpu, new_disks,
2450 new_sectors / RAID5_STRIPE_SECTORS(conf));
2456 mddev_resume(conf->mddev);
2458 conf->scribble_disks = new_disks;
2459 conf->scribble_sectors = new_sectors;
2464 static int resize_stripes(struct r5conf *conf, int newsize)
2466 /* Make all the stripes able to hold 'newsize' devices.
2467 * New slots in each stripe get 'page' set to a new page.
2469 * This happens in stages:
2470 * 1/ create a new kmem_cache and allocate the required number of
2472 * 2/ gather all the old stripe_heads and transfer the pages across
2473 * to the new stripe_heads. This will have the side effect of
2474 * freezing the array as once all stripe_heads have been collected,
2475 * no IO will be possible. Old stripe heads are freed once their
2476 * pages have been transferred over, and the old kmem_cache is
2477 * freed when all stripes are done.
2478 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2479 * we simple return a failure status - no need to clean anything up.
2480 * 4/ allocate new pages for the new slots in the new stripe_heads.
2481 * If this fails, we don't bother trying the shrink the
2482 * stripe_heads down again, we just leave them as they are.
2483 * As each stripe_head is processed the new one is released into
2486 * Once step2 is started, we cannot afford to wait for a write,
2487 * so we use GFP_NOIO allocations.
2489 struct stripe_head *osh, *nsh;
2490 LIST_HEAD(newstripes);
2491 struct disk_info *ndisks;
2493 struct kmem_cache *sc;
2497 md_allow_write(conf->mddev);
2500 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2501 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2506 /* Need to ensure auto-resizing doesn't interfere */
2507 mutex_lock(&conf->cache_size_mutex);
2509 for (i = conf->max_nr_stripes; i; i--) {
2510 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2514 list_add(&nsh->lru, &newstripes);
2517 /* didn't get enough, give up */
2518 while (!list_empty(&newstripes)) {
2519 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2520 list_del(&nsh->lru);
2521 free_stripe(sc, nsh);
2523 kmem_cache_destroy(sc);
2524 mutex_unlock(&conf->cache_size_mutex);
2527 /* Step 2 - Must use GFP_NOIO now.
2528 * OK, we have enough stripes, start collecting inactive
2529 * stripes and copying them over
2533 list_for_each_entry(nsh, &newstripes, lru) {
2534 lock_device_hash_lock(conf, hash);
2535 wait_event_cmd(conf->wait_for_stripe,
2536 !list_empty(conf->inactive_list + hash),
2537 unlock_device_hash_lock(conf, hash),
2538 lock_device_hash_lock(conf, hash));
2539 osh = get_free_stripe(conf, hash);
2540 unlock_device_hash_lock(conf, hash);
2542 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2543 for (i = 0; i < osh->nr_pages; i++) {
2544 nsh->pages[i] = osh->pages[i];
2545 osh->pages[i] = NULL;
2548 for(i=0; i<conf->pool_size; i++) {
2549 nsh->dev[i].page = osh->dev[i].page;
2550 nsh->dev[i].orig_page = osh->dev[i].page;
2551 nsh->dev[i].offset = osh->dev[i].offset;
2553 nsh->hash_lock_index = hash;
2554 free_stripe(conf->slab_cache, osh);
2556 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2557 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2562 kmem_cache_destroy(conf->slab_cache);
2565 * At this point, we are holding all the stripes so the array
2566 * is completely stalled, so now is a good time to resize
2567 * conf->disks and the scribble region
2569 ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2571 for (i = 0; i < conf->pool_size; i++)
2572 ndisks[i] = conf->disks[i];
2574 for (i = conf->pool_size; i < newsize; i++) {
2575 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2576 if (!ndisks[i].extra_page)
2581 for (i = conf->pool_size; i < newsize; i++)
2582 if (ndisks[i].extra_page)
2583 put_page(ndisks[i].extra_page);
2587 conf->disks = ndisks;
2592 conf->slab_cache = sc;
2593 conf->active_name = 1-conf->active_name;
2595 /* Step 4, return new stripes to service */
2596 while(!list_empty(&newstripes)) {
2597 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2598 list_del_init(&nsh->lru);
2600 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2601 for (i = 0; i < nsh->nr_pages; i++) {
2604 nsh->pages[i] = alloc_page(GFP_NOIO);
2609 for (i = conf->raid_disks; i < newsize; i++) {
2610 if (nsh->dev[i].page)
2612 nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2613 nsh->dev[i].orig_page = nsh->dev[i].page;
2614 nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2617 for (i=conf->raid_disks; i < newsize; i++)
2618 if (nsh->dev[i].page == NULL) {
2619 struct page *p = alloc_page(GFP_NOIO);
2620 nsh->dev[i].page = p;
2621 nsh->dev[i].orig_page = p;
2622 nsh->dev[i].offset = 0;
2627 raid5_release_stripe(nsh);
2629 /* critical section pass, GFP_NOIO no longer needed */
2632 conf->pool_size = newsize;
2633 mutex_unlock(&conf->cache_size_mutex);
2638 static int drop_one_stripe(struct r5conf *conf)
2640 struct stripe_head *sh;
2641 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2643 spin_lock_irq(conf->hash_locks + hash);
2644 sh = get_free_stripe(conf, hash);
2645 spin_unlock_irq(conf->hash_locks + hash);
2648 BUG_ON(atomic_read(&sh->count));
2650 free_stripe(conf->slab_cache, sh);
2651 atomic_dec(&conf->active_stripes);
2652 conf->max_nr_stripes--;
2656 static void shrink_stripes(struct r5conf *conf)
2658 while (conf->max_nr_stripes &&
2659 drop_one_stripe(conf))
2662 kmem_cache_destroy(conf->slab_cache);
2663 conf->slab_cache = NULL;
2666 static void raid5_end_read_request(struct bio * bi)
2668 struct stripe_head *sh = bi->bi_private;
2669 struct r5conf *conf = sh->raid_conf;
2670 int disks = sh->disks, i;
2671 char b[BDEVNAME_SIZE];
2672 struct md_rdev *rdev = NULL;
2675 for (i=0 ; i<disks; i++)
2676 if (bi == &sh->dev[i].req)
2679 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2680 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2687 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2688 /* If replacement finished while this request was outstanding,
2689 * 'replacement' might be NULL already.
2690 * In that case it moved down to 'rdev'.
2691 * rdev is not removed until all requests are finished.
2693 rdev = conf->disks[i].replacement;
2695 rdev = conf->disks[i].rdev;
2697 if (use_new_offset(conf, sh))
2698 s = sh->sector + rdev->new_data_offset;
2700 s = sh->sector + rdev->data_offset;
2701 if (!bi->bi_status) {
2702 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2703 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2704 /* Note that this cannot happen on a
2705 * replacement device. We just fail those on
2708 pr_info_ratelimited(
2709 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2710 mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2711 (unsigned long long)s,
2712 bdevname(rdev->bdev, b));
2713 atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2714 clear_bit(R5_ReadError, &sh->dev[i].flags);
2715 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2716 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2717 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2719 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2721 * end read for a page in journal, this
2722 * must be preparing for prexor in rmw
2724 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2726 if (atomic_read(&rdev->read_errors))
2727 atomic_set(&rdev->read_errors, 0);
2729 const char *bdn = bdevname(rdev->bdev, b);
2733 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2734 if (!(bi->bi_status == BLK_STS_PROTECTION))
2735 atomic_inc(&rdev->read_errors);
2736 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2737 pr_warn_ratelimited(
2738 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2739 mdname(conf->mddev),
2740 (unsigned long long)s,
2742 else if (conf->mddev->degraded >= conf->max_degraded) {
2744 pr_warn_ratelimited(
2745 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2746 mdname(conf->mddev),
2747 (unsigned long long)s,
2749 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2752 pr_warn_ratelimited(
2753 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2754 mdname(conf->mddev),
2755 (unsigned long long)s,
2757 } else if (atomic_read(&rdev->read_errors)
2758 > conf->max_nr_stripes) {
2759 if (!test_bit(Faulty, &rdev->flags)) {
2760 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2761 mdname(conf->mddev),
2762 atomic_read(&rdev->read_errors),
2763 conf->max_nr_stripes);
2764 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2765 mdname(conf->mddev), bdn);
2769 if (set_bad && test_bit(In_sync, &rdev->flags)
2770 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2773 if (sh->qd_idx >= 0 && sh->pd_idx == i)
2774 set_bit(R5_ReadError, &sh->dev[i].flags);
2775 else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2776 set_bit(R5_ReadError, &sh->dev[i].flags);
2777 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2779 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2781 clear_bit(R5_ReadError, &sh->dev[i].flags);
2782 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2784 && test_bit(In_sync, &rdev->flags)
2785 && rdev_set_badblocks(
2786 rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2787 md_error(conf->mddev, rdev);
2790 rdev_dec_pending(rdev, conf->mddev);
2792 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2793 set_bit(STRIPE_HANDLE, &sh->state);
2794 raid5_release_stripe(sh);
2797 static void raid5_end_write_request(struct bio *bi)
2799 struct stripe_head *sh = bi->bi_private;
2800 struct r5conf *conf = sh->raid_conf;
2801 int disks = sh->disks, i;
2802 struct md_rdev *rdev;
2805 int replacement = 0;
2807 for (i = 0 ; i < disks; i++) {
2808 if (bi == &sh->dev[i].req) {
2809 rdev = conf->disks[i].rdev;
2812 if (bi == &sh->dev[i].rreq) {
2813 rdev = conf->disks[i].replacement;
2817 /* rdev was removed and 'replacement'
2818 * replaced it. rdev is not removed
2819 * until all requests are finished.
2821 rdev = conf->disks[i].rdev;
2825 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2826 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2836 md_error(conf->mddev, rdev);
2837 else if (is_badblock(rdev, sh->sector,
2838 RAID5_STRIPE_SECTORS(conf),
2839 &first_bad, &bad_sectors))
2840 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2842 if (bi->bi_status) {
2843 set_bit(STRIPE_DEGRADED, &sh->state);
2844 set_bit(WriteErrorSeen, &rdev->flags);
2845 set_bit(R5_WriteError, &sh->dev[i].flags);
2846 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2847 set_bit(MD_RECOVERY_NEEDED,
2848 &rdev->mddev->recovery);
2849 } else if (is_badblock(rdev, sh->sector,
2850 RAID5_STRIPE_SECTORS(conf),
2851 &first_bad, &bad_sectors)) {
2852 set_bit(R5_MadeGood, &sh->dev[i].flags);
2853 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2854 /* That was a successful write so make
2855 * sure it looks like we already did
2858 set_bit(R5_ReWrite, &sh->dev[i].flags);
2861 rdev_dec_pending(rdev, conf->mddev);
2863 if (sh->batch_head && bi->bi_status && !replacement)
2864 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2867 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2868 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2869 set_bit(STRIPE_HANDLE, &sh->state);
2871 if (sh->batch_head && sh != sh->batch_head)
2872 raid5_release_stripe(sh->batch_head);
2873 raid5_release_stripe(sh);
2876 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2878 char b[BDEVNAME_SIZE];
2879 struct r5conf *conf = mddev->private;
2880 unsigned long flags;
2881 pr_debug("raid456: error called\n");
2883 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n",
2884 mdname(mddev), bdevname(rdev->bdev, b));
2886 spin_lock_irqsave(&conf->device_lock, flags);
2887 set_bit(Faulty, &rdev->flags);
2888 clear_bit(In_sync, &rdev->flags);
2889 mddev->degraded = raid5_calc_degraded(conf);
2891 if (has_failed(conf)) {
2892 set_bit(MD_BROKEN, &conf->mddev->flags);
2893 conf->recovery_disabled = mddev->recovery_disabled;
2895 pr_crit("md/raid:%s: Cannot continue operation (%d/%d failed).\n",
2896 mdname(mddev), mddev->degraded, conf->raid_disks);
2898 pr_crit("md/raid:%s: Operation continuing on %d devices.\n",
2899 mdname(mddev), conf->raid_disks - mddev->degraded);
2902 spin_unlock_irqrestore(&conf->device_lock, flags);
2903 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2905 set_bit(Blocked, &rdev->flags);
2906 set_mask_bits(&mddev->sb_flags, 0,
2907 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2908 r5c_update_on_rdev_error(mddev, rdev);
2912 * Input: a 'big' sector number,
2913 * Output: index of the data and parity disk, and the sector # in them.
2915 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2916 int previous, int *dd_idx,
2917 struct stripe_head *sh)
2919 sector_t stripe, stripe2;
2920 sector_t chunk_number;
2921 unsigned int chunk_offset;
2924 sector_t new_sector;
2925 int algorithm = previous ? conf->prev_algo
2927 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2928 : conf->chunk_sectors;
2929 int raid_disks = previous ? conf->previous_raid_disks
2931 int data_disks = raid_disks - conf->max_degraded;
2933 /* First compute the information on this sector */
2936 * Compute the chunk number and the sector offset inside the chunk
2938 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2939 chunk_number = r_sector;
2942 * Compute the stripe number
2944 stripe = chunk_number;
2945 *dd_idx = sector_div(stripe, data_disks);
2948 * Select the parity disk based on the user selected algorithm.
2950 pd_idx = qd_idx = -1;
2951 switch(conf->level) {
2953 pd_idx = data_disks;
2956 switch (algorithm) {
2957 case ALGORITHM_LEFT_ASYMMETRIC:
2958 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2959 if (*dd_idx >= pd_idx)
2962 case ALGORITHM_RIGHT_ASYMMETRIC:
2963 pd_idx = sector_div(stripe2, raid_disks);
2964 if (*dd_idx >= pd_idx)
2967 case ALGORITHM_LEFT_SYMMETRIC:
2968 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2969 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2971 case ALGORITHM_RIGHT_SYMMETRIC:
2972 pd_idx = sector_div(stripe2, raid_disks);
2973 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2975 case ALGORITHM_PARITY_0:
2979 case ALGORITHM_PARITY_N:
2980 pd_idx = data_disks;
2988 switch (algorithm) {
2989 case ALGORITHM_LEFT_ASYMMETRIC:
2990 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2991 qd_idx = pd_idx + 1;
2992 if (pd_idx == raid_disks-1) {
2993 (*dd_idx)++; /* Q D D D P */
2995 } else if (*dd_idx >= pd_idx)
2996 (*dd_idx) += 2; /* D D P Q D */
2998 case ALGORITHM_RIGHT_ASYMMETRIC:
2999 pd_idx = sector_div(stripe2, raid_disks);
3000 qd_idx = pd_idx + 1;
3001 if (pd_idx == raid_disks-1) {
3002 (*dd_idx)++; /* Q D D D P */
3004 } else if (*dd_idx >= pd_idx)
3005 (*dd_idx) += 2; /* D D P Q D */
3007 case ALGORITHM_LEFT_SYMMETRIC:
3008 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3009 qd_idx = (pd_idx + 1) % raid_disks;
3010 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3012 case ALGORITHM_RIGHT_SYMMETRIC:
3013 pd_idx = sector_div(stripe2, raid_disks);
3014 qd_idx = (pd_idx + 1) % raid_disks;
3015 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3018 case ALGORITHM_PARITY_0:
3023 case ALGORITHM_PARITY_N:
3024 pd_idx = data_disks;
3025 qd_idx = data_disks + 1;
3028 case ALGORITHM_ROTATING_ZERO_RESTART:
3029 /* Exactly the same as RIGHT_ASYMMETRIC, but or
3030 * of blocks for computing Q is different.
3032 pd_idx = sector_div(stripe2, raid_disks);
3033 qd_idx = pd_idx + 1;
3034 if (pd_idx == raid_disks-1) {
3035 (*dd_idx)++; /* Q D D D P */
3037 } else if (*dd_idx >= pd_idx)
3038 (*dd_idx) += 2; /* D D P Q D */
3042 case ALGORITHM_ROTATING_N_RESTART:
3043 /* Same a left_asymmetric, by first stripe is
3044 * D D D P Q rather than
3048 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3049 qd_idx = pd_idx + 1;
3050 if (pd_idx == raid_disks-1) {
3051 (*dd_idx)++; /* Q D D D P */
3053 } else if (*dd_idx >= pd_idx)
3054 (*dd_idx) += 2; /* D D P Q D */
3058 case ALGORITHM_ROTATING_N_CONTINUE:
3059 /* Same as left_symmetric but Q is before P */
3060 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3061 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3062 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3066 case ALGORITHM_LEFT_ASYMMETRIC_6:
3067 /* RAID5 left_asymmetric, with Q on last device */
3068 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3069 if (*dd_idx >= pd_idx)
3071 qd_idx = raid_disks - 1;
3074 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3075 pd_idx = sector_div(stripe2, raid_disks-1);
3076 if (*dd_idx >= pd_idx)
3078 qd_idx = raid_disks - 1;
3081 case ALGORITHM_LEFT_SYMMETRIC_6:
3082 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3083 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3084 qd_idx = raid_disks - 1;
3087 case ALGORITHM_RIGHT_SYMMETRIC_6:
3088 pd_idx = sector_div(stripe2, raid_disks-1);
3089 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3090 qd_idx = raid_disks - 1;
3093 case ALGORITHM_PARITY_0_6:
3096 qd_idx = raid_disks - 1;
3106 sh->pd_idx = pd_idx;
3107 sh->qd_idx = qd_idx;
3108 sh->ddf_layout = ddf_layout;
3111 * Finally, compute the new sector number
3113 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3117 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
3119 struct r5conf *conf = sh->raid_conf;
3120 int raid_disks = sh->disks;
3121 int data_disks = raid_disks - conf->max_degraded;
3122 sector_t new_sector = sh->sector, check;
3123 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3124 : conf->chunk_sectors;
3125 int algorithm = previous ? conf->prev_algo
3129 sector_t chunk_number;
3130 int dummy1, dd_idx = i;
3132 struct stripe_head sh2;
3134 chunk_offset = sector_div(new_sector, sectors_per_chunk);
3135 stripe = new_sector;
3137 if (i == sh->pd_idx)
3139 switch(conf->level) {
3142 switch (algorithm) {
3143 case ALGORITHM_LEFT_ASYMMETRIC:
3144 case ALGORITHM_RIGHT_ASYMMETRIC:
3148 case ALGORITHM_LEFT_SYMMETRIC:
3149 case ALGORITHM_RIGHT_SYMMETRIC:
3152 i -= (sh->pd_idx + 1);
3154 case ALGORITHM_PARITY_0:
3157 case ALGORITHM_PARITY_N:
3164 if (i == sh->qd_idx)
3165 return 0; /* It is the Q disk */
3166 switch (algorithm) {
3167 case ALGORITHM_LEFT_ASYMMETRIC:
3168 case ALGORITHM_RIGHT_ASYMMETRIC:
3169 case ALGORITHM_ROTATING_ZERO_RESTART:
3170 case ALGORITHM_ROTATING_N_RESTART:
3171 if (sh->pd_idx == raid_disks-1)
3172 i--; /* Q D D D P */
3173 else if (i > sh->pd_idx)
3174 i -= 2; /* D D P Q D */
3176 case ALGORITHM_LEFT_SYMMETRIC:
3177 case ALGORITHM_RIGHT_SYMMETRIC:
3178 if (sh->pd_idx == raid_disks-1)
3179 i--; /* Q D D D P */
3184 i -= (sh->pd_idx + 2);
3187 case ALGORITHM_PARITY_0:
3190 case ALGORITHM_PARITY_N:
3192 case ALGORITHM_ROTATING_N_CONTINUE:
3193 /* Like left_symmetric, but P is before Q */
3194 if (sh->pd_idx == 0)
3195 i--; /* P D D D Q */
3200 i -= (sh->pd_idx + 1);
3203 case ALGORITHM_LEFT_ASYMMETRIC_6:
3204 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3208 case ALGORITHM_LEFT_SYMMETRIC_6:
3209 case ALGORITHM_RIGHT_SYMMETRIC_6:
3211 i += data_disks + 1;
3212 i -= (sh->pd_idx + 1);
3214 case ALGORITHM_PARITY_0_6:
3223 chunk_number = stripe * data_disks + i;
3224 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3226 check = raid5_compute_sector(conf, r_sector,
3227 previous, &dummy1, &sh2);
3228 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3229 || sh2.qd_idx != sh->qd_idx) {
3230 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3231 mdname(conf->mddev));
3238 * There are cases where we want handle_stripe_dirtying() and
3239 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3241 * This function checks whether we want to delay the towrite. Specifically,
3242 * we delay the towrite when:
3244 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3245 * stripe has data in journal (for other devices).
3247 * In this case, when reading data for the non-overwrite dev, it is
3248 * necessary to handle complex rmw of write back cache (prexor with
3249 * orig_page, and xor with page). To keep read path simple, we would
3250 * like to flush data in journal to RAID disks first, so complex rmw
3251 * is handled in the write patch (handle_stripe_dirtying).
3253 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3255 * It is important to be able to flush all stripes in raid5-cache.
3256 * Therefore, we need reserve some space on the journal device for
3257 * these flushes. If flush operation includes pending writes to the
3258 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3259 * for the flush out. If we exclude these pending writes from flush
3260 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3261 * Therefore, excluding pending writes in these cases enables more
3262 * efficient use of the journal device.
3264 * Note: To make sure the stripe makes progress, we only delay
3265 * towrite for stripes with data already in journal (injournal > 0).
3266 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3267 * no_space_stripes list.
3269 * 3. during journal failure
3270 * In journal failure, we try to flush all cached data to raid disks
3271 * based on data in stripe cache. The array is read-only to upper
3272 * layers, so we would skip all pending writes.
3275 static inline bool delay_towrite(struct r5conf *conf,
3277 struct stripe_head_state *s)
3280 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3281 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3284 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3288 if (s->log_failed && s->injournal)
3294 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3295 int rcw, int expand)
3297 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3298 struct r5conf *conf = sh->raid_conf;
3299 int level = conf->level;
3303 * In some cases, handle_stripe_dirtying initially decided to
3304 * run rmw and allocates extra page for prexor. However, rcw is
3305 * cheaper later on. We need to free the extra page now,
3306 * because we won't be able to do that in ops_complete_prexor().
3308 r5c_release_extra_page(sh);
3310 for (i = disks; i--; ) {
3311 struct r5dev *dev = &sh->dev[i];
3313 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3314 set_bit(R5_LOCKED, &dev->flags);
3315 set_bit(R5_Wantdrain, &dev->flags);
3317 clear_bit(R5_UPTODATE, &dev->flags);
3319 } else if (test_bit(R5_InJournal, &dev->flags)) {
3320 set_bit(R5_LOCKED, &dev->flags);
3324 /* if we are not expanding this is a proper write request, and
3325 * there will be bios with new data to be drained into the
3330 /* False alarm, nothing to do */
3332 sh->reconstruct_state = reconstruct_state_drain_run;
3333 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3335 sh->reconstruct_state = reconstruct_state_run;
3337 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3339 if (s->locked + conf->max_degraded == disks)
3340 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3341 atomic_inc(&conf->pending_full_writes);
3343 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3344 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3345 BUG_ON(level == 6 &&
3346 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3347 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3349 for (i = disks; i--; ) {
3350 struct r5dev *dev = &sh->dev[i];
3351 if (i == pd_idx || i == qd_idx)
3355 (test_bit(R5_UPTODATE, &dev->flags) ||
3356 test_bit(R5_Wantcompute, &dev->flags))) {
3357 set_bit(R5_Wantdrain, &dev->flags);
3358 set_bit(R5_LOCKED, &dev->flags);
3359 clear_bit(R5_UPTODATE, &dev->flags);
3361 } else if (test_bit(R5_InJournal, &dev->flags)) {
3362 set_bit(R5_LOCKED, &dev->flags);
3367 /* False alarm - nothing to do */
3369 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3370 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3371 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3372 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3375 /* keep the parity disk(s) locked while asynchronous operations
3378 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3379 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3383 int qd_idx = sh->qd_idx;
3384 struct r5dev *dev = &sh->dev[qd_idx];
3386 set_bit(R5_LOCKED, &dev->flags);
3387 clear_bit(R5_UPTODATE, &dev->flags);
3391 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3392 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3393 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3394 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3395 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3397 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3398 __func__, (unsigned long long)sh->sector,
3399 s->locked, s->ops_request);
3403 * Each stripe/dev can have one or more bion attached.
3404 * toread/towrite point to the first in a chain.
3405 * The bi_next chain must be in order.
3407 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3408 int forwrite, int previous)
3411 struct r5conf *conf = sh->raid_conf;
3414 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3415 (unsigned long long)bi->bi_iter.bi_sector,
3416 (unsigned long long)sh->sector);
3418 spin_lock_irq(&sh->stripe_lock);
3419 sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3420 /* Don't allow new IO added to stripes in batch list */
3424 bip = &sh->dev[dd_idx].towrite;
3428 bip = &sh->dev[dd_idx].toread;
3429 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3430 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3432 bip = & (*bip)->bi_next;
3434 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3437 if (forwrite && raid5_has_ppl(conf)) {
3439 * With PPL only writes to consecutive data chunks within a
3440 * stripe are allowed because for a single stripe_head we can
3441 * only have one PPL entry at a time, which describes one data
3442 * range. Not really an overlap, but wait_for_overlap can be
3443 * used to handle this.
3451 for (i = 0; i < sh->disks; i++) {
3452 if (i != sh->pd_idx &&
3453 (i == dd_idx || sh->dev[i].towrite)) {
3454 sector = sh->dev[i].sector;
3455 if (count == 0 || sector < first)
3463 if (first + conf->chunk_sectors * (count - 1) != last)
3467 if (!forwrite || previous)
3468 clear_bit(STRIPE_BATCH_READY, &sh->state);
3470 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3474 bio_inc_remaining(bi);
3475 md_write_inc(conf->mddev, bi);
3478 /* check if page is covered */
3479 sector_t sector = sh->dev[dd_idx].sector;
3480 for (bi=sh->dev[dd_idx].towrite;
3481 sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3482 bi && bi->bi_iter.bi_sector <= sector;
3483 bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3484 if (bio_end_sector(bi) >= sector)
3485 sector = bio_end_sector(bi);
3487 if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3488 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3489 sh->overwrite_disks++;
3492 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3493 (unsigned long long)(*bip)->bi_iter.bi_sector,
3494 (unsigned long long)sh->sector, dd_idx);
3496 if (conf->mddev->bitmap && firstwrite) {
3497 /* Cannot hold spinlock over bitmap_startwrite,
3498 * but must ensure this isn't added to a batch until
3499 * we have added to the bitmap and set bm_seq.
3500 * So set STRIPE_BITMAP_PENDING to prevent
3502 * If multiple add_stripe_bio() calls race here they
3503 * much all set STRIPE_BITMAP_PENDING. So only the first one
3504 * to complete "bitmap_startwrite" gets to set
3505 * STRIPE_BIT_DELAY. This is important as once a stripe
3506 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3509 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3510 spin_unlock_irq(&sh->stripe_lock);
3511 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3512 RAID5_STRIPE_SECTORS(conf), 0);
3513 spin_lock_irq(&sh->stripe_lock);
3514 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3515 if (!sh->batch_head) {
3516 sh->bm_seq = conf->seq_flush+1;
3517 set_bit(STRIPE_BIT_DELAY, &sh->state);
3520 spin_unlock_irq(&sh->stripe_lock);
3522 if (stripe_can_batch(sh))
3523 stripe_add_to_batch_list(conf, sh);
3527 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3528 spin_unlock_irq(&sh->stripe_lock);
3532 static void end_reshape(struct r5conf *conf);
3534 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3535 struct stripe_head *sh)
3537 int sectors_per_chunk =
3538 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3540 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3541 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3543 raid5_compute_sector(conf,
3544 stripe * (disks - conf->max_degraded)
3545 *sectors_per_chunk + chunk_offset,
3551 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3552 struct stripe_head_state *s, int disks)
3555 BUG_ON(sh->batch_head);
3556 for (i = disks; i--; ) {
3560 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3561 struct md_rdev *rdev;
3563 rdev = rcu_dereference(conf->disks[i].rdev);
3564 if (rdev && test_bit(In_sync, &rdev->flags) &&
3565 !test_bit(Faulty, &rdev->flags))
3566 atomic_inc(&rdev->nr_pending);
3571 if (!rdev_set_badblocks(
3574 RAID5_STRIPE_SECTORS(conf), 0))
3575 md_error(conf->mddev, rdev);
3576 rdev_dec_pending(rdev, conf->mddev);
3579 spin_lock_irq(&sh->stripe_lock);
3580 /* fail all writes first */
3581 bi = sh->dev[i].towrite;
3582 sh->dev[i].towrite = NULL;
3583 sh->overwrite_disks = 0;
3584 spin_unlock_irq(&sh->stripe_lock);
3588 log_stripe_write_finished(sh);
3590 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3591 wake_up(&conf->wait_for_overlap);
3593 while (bi && bi->bi_iter.bi_sector <
3594 sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3595 struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3597 md_write_end(conf->mddev);
3602 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3603 RAID5_STRIPE_SECTORS(conf), 0, 0);
3605 /* and fail all 'written' */
3606 bi = sh->dev[i].written;
3607 sh->dev[i].written = NULL;
3608 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3609 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3610 sh->dev[i].page = sh->dev[i].orig_page;
3613 if (bi) bitmap_end = 1;
3614 while (bi && bi->bi_iter.bi_sector <
3615 sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3616 struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3618 md_write_end(conf->mddev);
3623 /* fail any reads if this device is non-operational and
3624 * the data has not reached the cache yet.
3626 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3627 s->failed > conf->max_degraded &&
3628 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3629 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3630 spin_lock_irq(&sh->stripe_lock);
3631 bi = sh->dev[i].toread;
3632 sh->dev[i].toread = NULL;
3633 spin_unlock_irq(&sh->stripe_lock);
3634 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3635 wake_up(&conf->wait_for_overlap);
3638 while (bi && bi->bi_iter.bi_sector <
3639 sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3640 struct bio *nextbi =
3641 r5_next_bio(conf, bi, sh->dev[i].sector);
3648 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3649 RAID5_STRIPE_SECTORS(conf), 0, 0);
3650 /* If we were in the middle of a write the parity block might
3651 * still be locked - so just clear all R5_LOCKED flags
3653 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3658 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3659 if (atomic_dec_and_test(&conf->pending_full_writes))
3660 md_wakeup_thread(conf->mddev->thread);
3664 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3665 struct stripe_head_state *s)
3670 BUG_ON(sh->batch_head);
3671 clear_bit(STRIPE_SYNCING, &sh->state);
3672 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3673 wake_up(&conf->wait_for_overlap);
3676 /* There is nothing more to do for sync/check/repair.
3677 * Don't even need to abort as that is handled elsewhere
3678 * if needed, and not always wanted e.g. if there is a known
3680 * For recover/replace we need to record a bad block on all
3681 * non-sync devices, or abort the recovery
3683 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3684 /* During recovery devices cannot be removed, so
3685 * locking and refcounting of rdevs is not needed
3688 for (i = 0; i < conf->raid_disks; i++) {
3689 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3691 && !test_bit(Faulty, &rdev->flags)
3692 && !test_bit(In_sync, &rdev->flags)
3693 && !rdev_set_badblocks(rdev, sh->sector,
3694 RAID5_STRIPE_SECTORS(conf), 0))
3696 rdev = rcu_dereference(conf->disks[i].replacement);
3698 && !test_bit(Faulty, &rdev->flags)
3699 && !test_bit(In_sync, &rdev->flags)
3700 && !rdev_set_badblocks(rdev, sh->sector,
3701 RAID5_STRIPE_SECTORS(conf), 0))
3706 conf->recovery_disabled =
3707 conf->mddev->recovery_disabled;
3709 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3712 static int want_replace(struct stripe_head *sh, int disk_idx)
3714 struct md_rdev *rdev;
3718 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3720 && !test_bit(Faulty, &rdev->flags)
3721 && !test_bit(In_sync, &rdev->flags)
3722 && (rdev->recovery_offset <= sh->sector
3723 || rdev->mddev->recovery_cp <= sh->sector))
3729 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3730 int disk_idx, int disks)
3732 struct r5dev *dev = &sh->dev[disk_idx];
3733 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3734 &sh->dev[s->failed_num[1]] };
3736 bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3739 if (test_bit(R5_LOCKED, &dev->flags) ||
3740 test_bit(R5_UPTODATE, &dev->flags))
3741 /* No point reading this as we already have it or have
3742 * decided to get it.
3747 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3748 /* We need this block to directly satisfy a request */
3751 if (s->syncing || s->expanding ||
3752 (s->replacing && want_replace(sh, disk_idx)))
3753 /* When syncing, or expanding we read everything.
3754 * When replacing, we need the replaced block.
3758 if ((s->failed >= 1 && fdev[0]->toread) ||
3759 (s->failed >= 2 && fdev[1]->toread))
3760 /* If we want to read from a failed device, then
3761 * we need to actually read every other device.
3765 /* Sometimes neither read-modify-write nor reconstruct-write
3766 * cycles can work. In those cases we read every block we
3767 * can. Then the parity-update is certain to have enough to
3769 * This can only be a problem when we need to write something,
3770 * and some device has failed. If either of those tests
3771 * fail we need look no further.
3773 if (!s->failed || !s->to_write)
3776 if (test_bit(R5_Insync, &dev->flags) &&
3777 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3778 /* Pre-reads at not permitted until after short delay
3779 * to gather multiple requests. However if this
3780 * device is no Insync, the block could only be computed
3781 * and there is no need to delay that.
3785 for (i = 0; i < s->failed && i < 2; i++) {
3786 if (fdev[i]->towrite &&
3787 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3788 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3789 /* If we have a partial write to a failed
3790 * device, then we will need to reconstruct
3791 * the content of that device, so all other
3792 * devices must be read.
3796 if (s->failed >= 2 &&
3797 (fdev[i]->towrite ||
3798 s->failed_num[i] == sh->pd_idx ||
3799 s->failed_num[i] == sh->qd_idx) &&
3800 !test_bit(R5_UPTODATE, &fdev[i]->flags))
3801 /* In max degraded raid6, If the failed disk is P, Q,
3802 * or we want to read the failed disk, we need to do
3803 * reconstruct-write.
3808 /* If we are forced to do a reconstruct-write, because parity
3809 * cannot be trusted and we are currently recovering it, there
3810 * is extra need to be careful.
3811 * If one of the devices that we would need to read, because
3812 * it is not being overwritten (and maybe not written at all)
3813 * is missing/faulty, then we need to read everything we can.
3816 sh->sector < sh->raid_conf->mddev->recovery_cp)
3817 /* reconstruct-write isn't being forced */
3819 for (i = 0; i < s->failed && i < 2; i++) {
3820 if (s->failed_num[i] != sh->pd_idx &&
3821 s->failed_num[i] != sh->qd_idx &&
3822 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3823 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3830 /* fetch_block - checks the given member device to see if its data needs
3831 * to be read or computed to satisfy a request.
3833 * Returns 1 when no more member devices need to be checked, otherwise returns
3834 * 0 to tell the loop in handle_stripe_fill to continue
3836 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3837 int disk_idx, int disks)
3839 struct r5dev *dev = &sh->dev[disk_idx];
3841 /* is the data in this block needed, and can we get it? */
3842 if (need_this_block(sh, s, disk_idx, disks)) {
3843 /* we would like to get this block, possibly by computing it,
3844 * otherwise read it if the backing disk is insync
3846 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3847 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3848 BUG_ON(sh->batch_head);
3851 * In the raid6 case if the only non-uptodate disk is P
3852 * then we already trusted P to compute the other failed
3853 * drives. It is safe to compute rather than re-read P.
3854 * In other cases we only compute blocks from failed
3855 * devices, otherwise check/repair might fail to detect
3856 * a real inconsistency.
3859 if ((s->uptodate == disks - 1) &&
3860 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3861 (s->failed && (disk_idx == s->failed_num[0] ||
3862 disk_idx == s->failed_num[1])))) {
3863 /* have disk failed, and we're requested to fetch it;
3866 pr_debug("Computing stripe %llu block %d\n",
3867 (unsigned long long)sh->sector, disk_idx);
3868 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3869 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3870 set_bit(R5_Wantcompute, &dev->flags);
3871 sh->ops.target = disk_idx;
3872 sh->ops.target2 = -1; /* no 2nd target */
3874 /* Careful: from this point on 'uptodate' is in the eye
3875 * of raid_run_ops which services 'compute' operations
3876 * before writes. R5_Wantcompute flags a block that will
3877 * be R5_UPTODATE by the time it is needed for a
3878 * subsequent operation.
3882 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3883 /* Computing 2-failure is *very* expensive; only
3884 * do it if failed >= 2
3887 for (other = disks; other--; ) {
3888 if (other == disk_idx)
3890 if (!test_bit(R5_UPTODATE,
3891 &sh->dev[other].flags))
3895 pr_debug("Computing stripe %llu blocks %d,%d\n",
3896 (unsigned long long)sh->sector,
3898 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3899 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3900 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3901 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3902 sh->ops.target = disk_idx;
3903 sh->ops.target2 = other;
3907 } else if (test_bit(R5_Insync, &dev->flags)) {
3908 set_bit(R5_LOCKED, &dev->flags);
3909 set_bit(R5_Wantread, &dev->flags);
3911 pr_debug("Reading block %d (sync=%d)\n",
3912 disk_idx, s->syncing);
3920 * handle_stripe_fill - read or compute data to satisfy pending requests.
3922 static void handle_stripe_fill(struct stripe_head *sh,
3923 struct stripe_head_state *s,
3928 /* look for blocks to read/compute, skip this if a compute
3929 * is already in flight, or if the stripe contents are in the
3930 * midst of changing due to a write
3932 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3933 !sh->reconstruct_state) {
3936 * For degraded stripe with data in journal, do not handle
3937 * read requests yet, instead, flush the stripe to raid
3938 * disks first, this avoids handling complex rmw of write
3939 * back cache (prexor with orig_page, and then xor with
3940 * page) in the read path
3942 if (s->to_read && s->injournal && s->failed) {
3943 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3944 r5c_make_stripe_write_out(sh);
3948 for (i = disks; i--; )
3949 if (fetch_block(sh, s, i, disks))
3953 set_bit(STRIPE_HANDLE, &sh->state);
3956 static void break_stripe_batch_list(struct stripe_head *head_sh,
3957 unsigned long handle_flags);
3958 /* handle_stripe_clean_event
3959 * any written block on an uptodate or failed drive can be returned.
3960 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3961 * never LOCKED, so we don't need to test 'failed' directly.
3963 static void handle_stripe_clean_event(struct r5conf *conf,
3964 struct stripe_head *sh, int disks)
3968 int discard_pending = 0;
3969 struct stripe_head *head_sh = sh;
3970 bool do_endio = false;
3972 for (i = disks; i--; )
3973 if (sh->dev[i].written) {
3975 if (!test_bit(R5_LOCKED, &dev->flags) &&
3976 (test_bit(R5_UPTODATE, &dev->flags) ||
3977 test_bit(R5_Discard, &dev->flags) ||
3978 test_bit(R5_SkipCopy, &dev->flags))) {
3979 /* We can return any write requests */
3980 struct bio *wbi, *wbi2;
3981 pr_debug("Return write for disc %d\n", i);
3982 if (test_and_clear_bit(R5_Discard, &dev->flags))
3983 clear_bit(R5_UPTODATE, &dev->flags);
3984 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3985 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3990 dev->page = dev->orig_page;
3992 dev->written = NULL;
3993 while (wbi && wbi->bi_iter.bi_sector <
3994 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
3995 wbi2 = r5_next_bio(conf, wbi, dev->sector);
3996 md_write_end(conf->mddev);
4000 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
4001 RAID5_STRIPE_SECTORS(conf),
4002 !test_bit(STRIPE_DEGRADED, &sh->state),
4004 if (head_sh->batch_head) {
4005 sh = list_first_entry(&sh->batch_list,
4008 if (sh != head_sh) {
4015 } else if (test_bit(R5_Discard, &dev->flags))
4016 discard_pending = 1;
4019 log_stripe_write_finished(sh);
4021 if (!discard_pending &&
4022 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4024 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4025 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4026 if (sh->qd_idx >= 0) {
4027 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4028 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4030 /* now that discard is done we can proceed with any sync */
4031 clear_bit(STRIPE_DISCARD, &sh->state);
4033 * SCSI discard will change some bio fields and the stripe has
4034 * no updated data, so remove it from hash list and the stripe
4035 * will be reinitialized
4038 hash = sh->hash_lock_index;
4039 spin_lock_irq(conf->hash_locks + hash);
4041 spin_unlock_irq(conf->hash_locks + hash);
4042 if (head_sh->batch_head) {
4043 sh = list_first_entry(&sh->batch_list,
4044 struct stripe_head, batch_list);
4050 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4051 set_bit(STRIPE_HANDLE, &sh->state);
4055 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4056 if (atomic_dec_and_test(&conf->pending_full_writes))
4057 md_wakeup_thread(conf->mddev->thread);
4059 if (head_sh->batch_head && do_endio)
4060 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4064 * For RMW in write back cache, we need extra page in prexor to store the
4065 * old data. This page is stored in dev->orig_page.
4067 * This function checks whether we have data for prexor. The exact logic
4069 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4071 static inline bool uptodate_for_rmw(struct r5dev *dev)
4073 return (test_bit(R5_UPTODATE, &dev->flags)) &&
4074 (!test_bit(R5_InJournal, &dev->flags) ||
4075 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4078 static int handle_stripe_dirtying(struct r5conf *conf,
4079 struct stripe_head *sh,
4080 struct stripe_head_state *s,
4083 int rmw = 0, rcw = 0, i;
4084 sector_t recovery_cp = conf->mddev->recovery_cp;
4086 /* Check whether resync is now happening or should start.
4087 * If yes, then the array is dirty (after unclean shutdown or
4088 * initial creation), so parity in some stripes might be inconsistent.
4089 * In this case, we need to always do reconstruct-write, to ensure
4090 * that in case of drive failure or read-error correction, we
4091 * generate correct data from the parity.
4093 if (conf->rmw_level == PARITY_DISABLE_RMW ||
4094 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
4096 /* Calculate the real rcw later - for now make it
4097 * look like rcw is cheaper
4100 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4101 conf->rmw_level, (unsigned long long)recovery_cp,
4102 (unsigned long long)sh->sector);
4103 } else for (i = disks; i--; ) {
4104 /* would I have to read this buffer for read_modify_write */
4105 struct r5dev *dev = &sh->dev[i];
4106 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4107 i == sh->pd_idx || i == sh->qd_idx ||
4108 test_bit(R5_InJournal, &dev->flags)) &&
4109 !test_bit(R5_LOCKED, &dev->flags) &&
4110 !(uptodate_for_rmw(dev) ||
4111 test_bit(R5_Wantcompute, &dev->flags))) {
4112 if (test_bit(R5_Insync, &dev->flags))
4115 rmw += 2*disks; /* cannot read it */
4117 /* Would I have to read this buffer for reconstruct_write */
4118 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4119 i != sh->pd_idx && i != sh->qd_idx &&
4120 !test_bit(R5_LOCKED, &dev->flags) &&
4121 !(test_bit(R5_UPTODATE, &dev->flags) ||
4122 test_bit(R5_Wantcompute, &dev->flags))) {
4123 if (test_bit(R5_Insync, &dev->flags))
4130 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4131 (unsigned long long)sh->sector, sh->state, rmw, rcw);
4132 set_bit(STRIPE_HANDLE, &sh->state);
4133 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4134 /* prefer read-modify-write, but need to get some data */
4135 if (conf->mddev->queue)
4136 blk_add_trace_msg(conf->mddev->queue,
4137 "raid5 rmw %llu %d",
4138 (unsigned long long)sh->sector, rmw);
4139 for (i = disks; i--; ) {
4140 struct r5dev *dev = &sh->dev[i];
4141 if (test_bit(R5_InJournal, &dev->flags) &&
4142 dev->page == dev->orig_page &&
4143 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4144 /* alloc page for prexor */
4145 struct page *p = alloc_page(GFP_NOIO);
4153 * alloc_page() failed, try use
4154 * disk_info->extra_page
4156 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4157 &conf->cache_state)) {
4158 r5c_use_extra_page(sh);
4162 /* extra_page in use, add to delayed_list */
4163 set_bit(STRIPE_DELAYED, &sh->state);
4164 s->waiting_extra_page = 1;
4169 for (i = disks; i--; ) {
4170 struct r5dev *dev = &sh->dev[i];
4171 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4172 i == sh->pd_idx || i == sh->qd_idx ||
4173 test_bit(R5_InJournal, &dev->flags)) &&
4174 !test_bit(R5_LOCKED, &dev->flags) &&
4175 !(uptodate_for_rmw(dev) ||
4176 test_bit(R5_Wantcompute, &dev->flags)) &&
4177 test_bit(R5_Insync, &dev->flags)) {
4178 if (test_bit(STRIPE_PREREAD_ACTIVE,
4180 pr_debug("Read_old block %d for r-m-w\n",
4182 set_bit(R5_LOCKED, &dev->flags);
4183 set_bit(R5_Wantread, &dev->flags);
4186 set_bit(STRIPE_DELAYED, &sh->state);
4190 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4191 /* want reconstruct write, but need to get some data */
4194 for (i = disks; i--; ) {
4195 struct r5dev *dev = &sh->dev[i];
4196 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4197 i != sh->pd_idx && i != sh->qd_idx &&
4198 !test_bit(R5_LOCKED, &dev->flags) &&
4199 !(test_bit(R5_UPTODATE, &dev->flags) ||
4200 test_bit(R5_Wantcompute, &dev->flags))) {
4202 if (test_bit(R5_Insync, &dev->flags) &&
4203 test_bit(STRIPE_PREREAD_ACTIVE,
4205 pr_debug("Read_old block "
4206 "%d for Reconstruct\n", i);
4207 set_bit(R5_LOCKED, &dev->flags);
4208 set_bit(R5_Wantread, &dev->flags);
4212 set_bit(STRIPE_DELAYED, &sh->state);
4215 if (rcw && conf->mddev->queue)
4216 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4217 (unsigned long long)sh->sector,
4218 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4221 if (rcw > disks && rmw > disks &&
4222 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4223 set_bit(STRIPE_DELAYED, &sh->state);
4225 /* now if nothing is locked, and if we have enough data,
4226 * we can start a write request
4228 /* since handle_stripe can be called at any time we need to handle the
4229 * case where a compute block operation has been submitted and then a
4230 * subsequent call wants to start a write request. raid_run_ops only
4231 * handles the case where compute block and reconstruct are requested
4232 * simultaneously. If this is not the case then new writes need to be
4233 * held off until the compute completes.
4235 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4236 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4237 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4238 schedule_reconstruction(sh, s, rcw == 0, 0);
4242 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4243 struct stripe_head_state *s, int disks)
4245 struct r5dev *dev = NULL;
4247 BUG_ON(sh->batch_head);
4248 set_bit(STRIPE_HANDLE, &sh->state);
4250 switch (sh->check_state) {
4251 case check_state_idle:
4252 /* start a new check operation if there are no failures */
4253 if (s->failed == 0) {
4254 BUG_ON(s->uptodate != disks);
4255 sh->check_state = check_state_run;
4256 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4257 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4261 dev = &sh->dev[s->failed_num[0]];
4263 case check_state_compute_result:
4264 sh->check_state = check_state_idle;
4266 dev = &sh->dev[sh->pd_idx];
4268 /* check that a write has not made the stripe insync */
4269 if (test_bit(STRIPE_INSYNC, &sh->state))
4272 /* either failed parity check, or recovery is happening */
4273 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4274 BUG_ON(s->uptodate != disks);
4276 set_bit(R5_LOCKED, &dev->flags);
4278 set_bit(R5_Wantwrite, &dev->flags);
4280 clear_bit(STRIPE_DEGRADED, &sh->state);
4281 set_bit(STRIPE_INSYNC, &sh->state);
4283 case check_state_run:
4284 break; /* we will be called again upon completion */
4285 case check_state_check_result:
4286 sh->check_state = check_state_idle;
4288 /* if a failure occurred during the check operation, leave
4289 * STRIPE_INSYNC not set and let the stripe be handled again
4294 /* handle a successful check operation, if parity is correct
4295 * we are done. Otherwise update the mismatch count and repair
4296 * parity if !MD_RECOVERY_CHECK
4298 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4299 /* parity is correct (on disc,
4300 * not in buffer any more)
4302 set_bit(STRIPE_INSYNC, &sh->state);
4304 atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4305 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4306 /* don't try to repair!! */
4307 set_bit(STRIPE_INSYNC, &sh->state);
4308 pr_warn_ratelimited("%s: mismatch sector in range "
4309 "%llu-%llu\n", mdname(conf->mddev),
4310 (unsigned long long) sh->sector,
4311 (unsigned long long) sh->sector +
4312 RAID5_STRIPE_SECTORS(conf));
4314 sh->check_state = check_state_compute_run;
4315 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4316 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4317 set_bit(R5_Wantcompute,
4318 &sh->dev[sh->pd_idx].flags);
4319 sh->ops.target = sh->pd_idx;
4320 sh->ops.target2 = -1;
4325 case check_state_compute_run:
4328 pr_err("%s: unknown check_state: %d sector: %llu\n",
4329 __func__, sh->check_state,
4330 (unsigned long long) sh->sector);
4335 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4336 struct stripe_head_state *s,
4339 int pd_idx = sh->pd_idx;
4340 int qd_idx = sh->qd_idx;
4343 BUG_ON(sh->batch_head);
4344 set_bit(STRIPE_HANDLE, &sh->state);
4346 BUG_ON(s->failed > 2);
4348 /* Want to check and possibly repair P and Q.
4349 * However there could be one 'failed' device, in which
4350 * case we can only check one of them, possibly using the
4351 * other to generate missing data
4354 switch (sh->check_state) {
4355 case check_state_idle:
4356 /* start a new check operation if there are < 2 failures */
4357 if (s->failed == s->q_failed) {
4358 /* The only possible failed device holds Q, so it
4359 * makes sense to check P (If anything else were failed,
4360 * we would have used P to recreate it).
4362 sh->check_state = check_state_run;
4364 if (!s->q_failed && s->failed < 2) {
4365 /* Q is not failed, and we didn't use it to generate
4366 * anything, so it makes sense to check it
4368 if (sh->check_state == check_state_run)
4369 sh->check_state = check_state_run_pq;
4371 sh->check_state = check_state_run_q;
4374 /* discard potentially stale zero_sum_result */
4375 sh->ops.zero_sum_result = 0;
4377 if (sh->check_state == check_state_run) {
4378 /* async_xor_zero_sum destroys the contents of P */
4379 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4382 if (sh->check_state >= check_state_run &&
4383 sh->check_state <= check_state_run_pq) {
4384 /* async_syndrome_zero_sum preserves P and Q, so
4385 * no need to mark them !uptodate here
4387 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4391 /* we have 2-disk failure */
4392 BUG_ON(s->failed != 2);
4394 case check_state_compute_result:
4395 sh->check_state = check_state_idle;
4397 /* check that a write has not made the stripe insync */
4398 if (test_bit(STRIPE_INSYNC, &sh->state))
4401 /* now write out any block on a failed drive,
4402 * or P or Q if they were recomputed
4405 if (s->failed == 2) {
4406 dev = &sh->dev[s->failed_num[1]];
4408 set_bit(R5_LOCKED, &dev->flags);
4409 set_bit(R5_Wantwrite, &dev->flags);
4411 if (s->failed >= 1) {
4412 dev = &sh->dev[s->failed_num[0]];
4414 set_bit(R5_LOCKED, &dev->flags);
4415 set_bit(R5_Wantwrite, &dev->flags);
4417 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4418 dev = &sh->dev[pd_idx];
4420 set_bit(R5_LOCKED, &dev->flags);
4421 set_bit(R5_Wantwrite, &dev->flags);
4423 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4424 dev = &sh->dev[qd_idx];
4426 set_bit(R5_LOCKED, &dev->flags);
4427 set_bit(R5_Wantwrite, &dev->flags);
4429 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4430 "%s: disk%td not up to date\n",
4431 mdname(conf->mddev),
4432 dev - (struct r5dev *) &sh->dev)) {
4433 clear_bit(R5_LOCKED, &dev->flags);
4434 clear_bit(R5_Wantwrite, &dev->flags);
4437 clear_bit(STRIPE_DEGRADED, &sh->state);
4439 set_bit(STRIPE_INSYNC, &sh->state);
4441 case check_state_run:
4442 case check_state_run_q:
4443 case check_state_run_pq:
4444 break; /* we will be called again upon completion */
4445 case check_state_check_result:
4446 sh->check_state = check_state_idle;
4448 /* handle a successful check operation, if parity is correct
4449 * we are done. Otherwise update the mismatch count and repair
4450 * parity if !MD_RECOVERY_CHECK
4452 if (sh->ops.zero_sum_result == 0) {
4453 /* both parities are correct */
4455 set_bit(STRIPE_INSYNC, &sh->state);
4457 /* in contrast to the raid5 case we can validate
4458 * parity, but still have a failure to write
4461 sh->check_state = check_state_compute_result;
4462 /* Returning at this point means that we may go
4463 * off and bring p and/or q uptodate again so
4464 * we make sure to check zero_sum_result again
4465 * to verify if p or q need writeback
4469 atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4470 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4471 /* don't try to repair!! */
4472 set_bit(STRIPE_INSYNC, &sh->state);
4473 pr_warn_ratelimited("%s: mismatch sector in range "
4474 "%llu-%llu\n", mdname(conf->mddev),
4475 (unsigned long long) sh->sector,
4476 (unsigned long long) sh->sector +
4477 RAID5_STRIPE_SECTORS(conf));
4479 int *target = &sh->ops.target;
4481 sh->ops.target = -1;
4482 sh->ops.target2 = -1;
4483 sh->check_state = check_state_compute_run;
4484 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4485 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4486 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4487 set_bit(R5_Wantcompute,
4488 &sh->dev[pd_idx].flags);
4490 target = &sh->ops.target2;
4493 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4494 set_bit(R5_Wantcompute,
4495 &sh->dev[qd_idx].flags);
4502 case check_state_compute_run:
4505 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4506 __func__, sh->check_state,
4507 (unsigned long long) sh->sector);
4512 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4516 /* We have read all the blocks in this stripe and now we need to
4517 * copy some of them into a target stripe for expand.
4519 struct dma_async_tx_descriptor *tx = NULL;
4520 BUG_ON(sh->batch_head);
4521 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4522 for (i = 0; i < sh->disks; i++)
4523 if (i != sh->pd_idx && i != sh->qd_idx) {
4525 struct stripe_head *sh2;
4526 struct async_submit_ctl submit;
4528 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4529 sector_t s = raid5_compute_sector(conf, bn, 0,
4531 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4533 /* so far only the early blocks of this stripe
4534 * have been requested. When later blocks
4535 * get requested, we will try again
4538 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4539 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4540 /* must have already done this block */
4541 raid5_release_stripe(sh2);
4545 /* place all the copies on one channel */
4546 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4547 tx = async_memcpy(sh2->dev[dd_idx].page,
4548 sh->dev[i].page, sh2->dev[dd_idx].offset,
4549 sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4552 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4553 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4554 for (j = 0; j < conf->raid_disks; j++)
4555 if (j != sh2->pd_idx &&
4557 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4559 if (j == conf->raid_disks) {
4560 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4561 set_bit(STRIPE_HANDLE, &sh2->state);
4563 raid5_release_stripe(sh2);
4566 /* done submitting copies, wait for them to complete */
4567 async_tx_quiesce(&tx);
4571 * handle_stripe - do things to a stripe.
4573 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4574 * state of various bits to see what needs to be done.
4576 * return some read requests which now have data
4577 * return some write requests which are safely on storage
4578 * schedule a read on some buffers
4579 * schedule a write of some buffers
4580 * return confirmation of parity correctness
4584 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4586 struct r5conf *conf = sh->raid_conf;
4587 int disks = sh->disks;
4590 int do_recovery = 0;
4592 memset(s, 0, sizeof(*s));
4594 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4595 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4596 s->failed_num[0] = -1;
4597 s->failed_num[1] = -1;
4598 s->log_failed = r5l_log_disk_error(conf);
4600 /* Now to look around and see what can be done */
4602 for (i=disks; i--; ) {
4603 struct md_rdev *rdev;
4610 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4612 dev->toread, dev->towrite, dev->written);
4613 /* maybe we can reply to a read
4615 * new wantfill requests are only permitted while
4616 * ops_complete_biofill is guaranteed to be inactive
4618 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4619 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4620 set_bit(R5_Wantfill, &dev->flags);
4622 /* now count some things */
4623 if (test_bit(R5_LOCKED, &dev->flags))
4625 if (test_bit(R5_UPTODATE, &dev->flags))
4627 if (test_bit(R5_Wantcompute, &dev->flags)) {
4629 BUG_ON(s->compute > 2);
4632 if (test_bit(R5_Wantfill, &dev->flags))
4634 else if (dev->toread)
4638 if (!test_bit(R5_OVERWRITE, &dev->flags))
4643 /* Prefer to use the replacement for reads, but only
4644 * if it is recovered enough and has no bad blocks.
4646 rdev = rcu_dereference(conf->disks[i].replacement);
4647 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4648 rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4649 !is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4650 &first_bad, &bad_sectors))
4651 set_bit(R5_ReadRepl, &dev->flags);
4653 if (rdev && !test_bit(Faulty, &rdev->flags))
4654 set_bit(R5_NeedReplace, &dev->flags);
4656 clear_bit(R5_NeedReplace, &dev->flags);
4657 rdev = rcu_dereference(conf->disks[i].rdev);
4658 clear_bit(R5_ReadRepl, &dev->flags);
4660 if (rdev && test_bit(Faulty, &rdev->flags))
4663 is_bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4664 &first_bad, &bad_sectors);
4665 if (s->blocked_rdev == NULL
4666 && (test_bit(Blocked, &rdev->flags)
4669 set_bit(BlockedBadBlocks,
4671 s->blocked_rdev = rdev;
4672 atomic_inc(&rdev->nr_pending);
4675 clear_bit(R5_Insync, &dev->flags);
4679 /* also not in-sync */
4680 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4681 test_bit(R5_UPTODATE, &dev->flags)) {
4682 /* treat as in-sync, but with a read error
4683 * which we can now try to correct
4685 set_bit(R5_Insync, &dev->flags);
4686 set_bit(R5_ReadError, &dev->flags);
4688 } else if (test_bit(In_sync, &rdev->flags))
4689 set_bit(R5_Insync, &dev->flags);
4690 else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4691 /* in sync if before recovery_offset */
4692 set_bit(R5_Insync, &dev->flags);
4693 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4694 test_bit(R5_Expanded, &dev->flags))
4695 /* If we've reshaped into here, we assume it is Insync.
4696 * We will shortly update recovery_offset to make
4699 set_bit(R5_Insync, &dev->flags);
4701 if (test_bit(R5_WriteError, &dev->flags)) {
4702 /* This flag does not apply to '.replacement'
4703 * only to .rdev, so make sure to check that*/
4704 struct md_rdev *rdev2 = rcu_dereference(
4705 conf->disks[i].rdev);
4707 clear_bit(R5_Insync, &dev->flags);
4708 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4709 s->handle_bad_blocks = 1;
4710 atomic_inc(&rdev2->nr_pending);
4712 clear_bit(R5_WriteError, &dev->flags);
4714 if (test_bit(R5_MadeGood, &dev->flags)) {
4715 /* This flag does not apply to '.replacement'
4716 * only to .rdev, so make sure to check that*/
4717 struct md_rdev *rdev2 = rcu_dereference(
4718 conf->disks[i].rdev);
4719 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4720 s->handle_bad_blocks = 1;
4721 atomic_inc(&rdev2->nr_pending);
4723 clear_bit(R5_MadeGood, &dev->flags);
4725 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4726 struct md_rdev *rdev2 = rcu_dereference(
4727 conf->disks[i].replacement);
4728 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4729 s->handle_bad_blocks = 1;
4730 atomic_inc(&rdev2->nr_pending);
4732 clear_bit(R5_MadeGoodRepl, &dev->flags);
4734 if (!test_bit(R5_Insync, &dev->flags)) {
4735 /* The ReadError flag will just be confusing now */
4736 clear_bit(R5_ReadError, &dev->flags);
4737 clear_bit(R5_ReWrite, &dev->flags);
4739 if (test_bit(R5_ReadError, &dev->flags))
4740 clear_bit(R5_Insync, &dev->flags);
4741 if (!test_bit(R5_Insync, &dev->flags)) {
4743 s->failed_num[s->failed] = i;
4745 if (rdev && !test_bit(Faulty, &rdev->flags))
4748 rdev = rcu_dereference(
4749 conf->disks[i].replacement);
4750 if (rdev && !test_bit(Faulty, &rdev->flags))
4755 if (test_bit(R5_InJournal, &dev->flags))
4757 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4760 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4761 /* If there is a failed device being replaced,
4762 * we must be recovering.
4763 * else if we are after recovery_cp, we must be syncing
4764 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4765 * else we can only be replacing
4766 * sync and recovery both need to read all devices, and so
4767 * use the same flag.
4770 sh->sector >= conf->mddev->recovery_cp ||
4771 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4780 * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4781 * a head which can now be handled.
4783 static int clear_batch_ready(struct stripe_head *sh)
4785 struct stripe_head *tmp;
4786 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4787 return (sh->batch_head && sh->batch_head != sh);
4788 spin_lock(&sh->stripe_lock);
4789 if (!sh->batch_head) {
4790 spin_unlock(&sh->stripe_lock);
4795 * this stripe could be added to a batch list before we check
4796 * BATCH_READY, skips it
4798 if (sh->batch_head != sh) {
4799 spin_unlock(&sh->stripe_lock);
4802 spin_lock(&sh->batch_lock);
4803 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4804 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4805 spin_unlock(&sh->batch_lock);
4806 spin_unlock(&sh->stripe_lock);
4809 * BATCH_READY is cleared, no new stripes can be added.
4810 * batch_list can be accessed without lock
4815 static void break_stripe_batch_list(struct stripe_head *head_sh,
4816 unsigned long handle_flags)
4818 struct stripe_head *sh, *next;
4822 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4824 list_del_init(&sh->batch_list);
4826 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4827 (1 << STRIPE_SYNCING) |
4828 (1 << STRIPE_REPLACED) |
4829 (1 << STRIPE_DELAYED) |
4830 (1 << STRIPE_BIT_DELAY) |
4831 (1 << STRIPE_FULL_WRITE) |
4832 (1 << STRIPE_BIOFILL_RUN) |
4833 (1 << STRIPE_COMPUTE_RUN) |
4834 (1 << STRIPE_DISCARD) |
4835 (1 << STRIPE_BATCH_READY) |
4836 (1 << STRIPE_BATCH_ERR) |
4837 (1 << STRIPE_BITMAP_PENDING)),
4838 "stripe state: %lx\n", sh->state);
4839 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4840 (1 << STRIPE_REPLACED)),
4841 "head stripe state: %lx\n", head_sh->state);
4843 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4844 (1 << STRIPE_PREREAD_ACTIVE) |
4845 (1 << STRIPE_DEGRADED) |
4846 (1 << STRIPE_ON_UNPLUG_LIST)),
4847 head_sh->state & (1 << STRIPE_INSYNC));
4849 sh->check_state = head_sh->check_state;
4850 sh->reconstruct_state = head_sh->reconstruct_state;
4851 spin_lock_irq(&sh->stripe_lock);
4852 sh->batch_head = NULL;
4853 spin_unlock_irq(&sh->stripe_lock);
4854 for (i = 0; i < sh->disks; i++) {
4855 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4857 sh->dev[i].flags = head_sh->dev[i].flags &
4858 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4860 if (handle_flags == 0 ||
4861 sh->state & handle_flags)
4862 set_bit(STRIPE_HANDLE, &sh->state);
4863 raid5_release_stripe(sh);
4865 spin_lock_irq(&head_sh->stripe_lock);
4866 head_sh->batch_head = NULL;
4867 spin_unlock_irq(&head_sh->stripe_lock);
4868 for (i = 0; i < head_sh->disks; i++)
4869 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4871 if (head_sh->state & handle_flags)
4872 set_bit(STRIPE_HANDLE, &head_sh->state);
4875 wake_up(&head_sh->raid_conf->wait_for_overlap);
4878 static void handle_stripe(struct stripe_head *sh)
4880 struct stripe_head_state s;
4881 struct r5conf *conf = sh->raid_conf;
4884 int disks = sh->disks;
4885 struct r5dev *pdev, *qdev;
4887 clear_bit(STRIPE_HANDLE, &sh->state);
4890 * handle_stripe should not continue handle the batched stripe, only
4891 * the head of batch list or lone stripe can continue. Otherwise we
4892 * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4893 * is set for the batched stripe.
4895 if (clear_batch_ready(sh))
4898 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4899 /* already being handled, ensure it gets handled
4900 * again when current action finishes */
4901 set_bit(STRIPE_HANDLE, &sh->state);
4905 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4906 break_stripe_batch_list(sh, 0);
4908 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4909 spin_lock(&sh->stripe_lock);
4911 * Cannot process 'sync' concurrently with 'discard'.
4912 * Flush data in r5cache before 'sync'.
4914 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4915 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4916 !test_bit(STRIPE_DISCARD, &sh->state) &&
4917 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4918 set_bit(STRIPE_SYNCING, &sh->state);
4919 clear_bit(STRIPE_INSYNC, &sh->state);
4920 clear_bit(STRIPE_REPLACED, &sh->state);
4922 spin_unlock(&sh->stripe_lock);
4924 clear_bit(STRIPE_DELAYED, &sh->state);
4926 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4927 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4928 (unsigned long long)sh->sector, sh->state,
4929 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4930 sh->check_state, sh->reconstruct_state);
4932 analyse_stripe(sh, &s);
4934 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4937 if (s.handle_bad_blocks ||
4938 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4939 set_bit(STRIPE_HANDLE, &sh->state);
4943 if (unlikely(s.blocked_rdev)) {
4944 if (s.syncing || s.expanding || s.expanded ||
4945 s.replacing || s.to_write || s.written) {
4946 set_bit(STRIPE_HANDLE, &sh->state);
4949 /* There is nothing for the blocked_rdev to block */
4950 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4951 s.blocked_rdev = NULL;
4954 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4955 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4956 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4959 pr_debug("locked=%d uptodate=%d to_read=%d"
4960 " to_write=%d failed=%d failed_num=%d,%d\n",
4961 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4962 s.failed_num[0], s.failed_num[1]);
4964 * check if the array has lost more than max_degraded devices and,
4965 * if so, some requests might need to be failed.
4967 * When journal device failed (log_failed), we will only process
4968 * the stripe if there is data need write to raid disks
4970 if (s.failed > conf->max_degraded ||
4971 (s.log_failed && s.injournal == 0)) {
4972 sh->check_state = 0;
4973 sh->reconstruct_state = 0;
4974 break_stripe_batch_list(sh, 0);
4975 if (s.to_read+s.to_write+s.written)
4976 handle_failed_stripe(conf, sh, &s, disks);
4977 if (s.syncing + s.replacing)
4978 handle_failed_sync(conf, sh, &s);
4981 /* Now we check to see if any write operations have recently
4985 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4987 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4988 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4989 sh->reconstruct_state = reconstruct_state_idle;
4991 /* All the 'written' buffers and the parity block are ready to
4992 * be written back to disk
4994 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4995 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4996 BUG_ON(sh->qd_idx >= 0 &&
4997 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4998 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4999 for (i = disks; i--; ) {
5000 struct r5dev *dev = &sh->dev[i];
5001 if (test_bit(R5_LOCKED, &dev->flags) &&
5002 (i == sh->pd_idx || i == sh->qd_idx ||
5003 dev->written || test_bit(R5_InJournal,
5005 pr_debug("Writing block %d\n", i);
5006 set_bit(R5_Wantwrite, &dev->flags);
5011 if (!test_bit(R5_Insync, &dev->flags) ||
5012 ((i == sh->pd_idx || i == sh->qd_idx) &&
5014 set_bit(STRIPE_INSYNC, &sh->state);
5017 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5018 s.dec_preread_active = 1;
5022 * might be able to return some write requests if the parity blocks
5023 * are safe, or on a failed drive
5025 pdev = &sh->dev[sh->pd_idx];
5026 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
5027 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
5028 qdev = &sh->dev[sh->qd_idx];
5029 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
5030 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
5034 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
5035 && !test_bit(R5_LOCKED, &pdev->flags)
5036 && (test_bit(R5_UPTODATE, &pdev->flags) ||
5037 test_bit(R5_Discard, &pdev->flags))))) &&
5038 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
5039 && !test_bit(R5_LOCKED, &qdev->flags)
5040 && (test_bit(R5_UPTODATE, &qdev->flags) ||
5041 test_bit(R5_Discard, &qdev->flags))))))
5042 handle_stripe_clean_event(conf, sh, disks);
5045 r5c_handle_cached_data_endio(conf, sh, disks);
5046 log_stripe_write_finished(sh);
5048 /* Now we might consider reading some blocks, either to check/generate
5049 * parity, or to satisfy requests
5050 * or to load a block that is being partially written.
5052 if (s.to_read || s.non_overwrite
5053 || (s.to_write && s.failed)
5054 || (s.syncing && (s.uptodate + s.compute < disks))
5057 handle_stripe_fill(sh, &s, disks);
5060 * When the stripe finishes full journal write cycle (write to journal
5061 * and raid disk), this is the clean up procedure so it is ready for
5064 r5c_finish_stripe_write_out(conf, sh, &s);
5067 * Now to consider new write requests, cache write back and what else,
5068 * if anything should be read. We do not handle new writes when:
5069 * 1/ A 'write' operation (copy+xor) is already in flight.
5070 * 2/ A 'check' operation is in flight, as it may clobber the parity
5072 * 3/ A r5c cache log write is in flight.
5075 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
5076 if (!r5c_is_writeback(conf->log)) {
5078 handle_stripe_dirtying(conf, sh, &s, disks);
5079 } else { /* write back cache */
5082 /* First, try handle writes in caching phase */
5084 ret = r5c_try_caching_write(conf, sh, &s,
5087 * If caching phase failed: ret == -EAGAIN
5089 * stripe under reclaim: !caching && injournal
5091 * fall back to handle_stripe_dirtying()
5093 if (ret == -EAGAIN ||
5094 /* stripe under reclaim: !caching && injournal */
5095 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
5097 ret = handle_stripe_dirtying(conf, sh, &s,
5105 /* maybe we need to check and possibly fix the parity for this stripe
5106 * Any reads will already have been scheduled, so we just see if enough
5107 * data is available. The parity check is held off while parity
5108 * dependent operations are in flight.
5110 if (sh->check_state ||
5111 (s.syncing && s.locked == 0 &&
5112 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5113 !test_bit(STRIPE_INSYNC, &sh->state))) {
5114 if (conf->level == 6)
5115 handle_parity_checks6(conf, sh, &s, disks);
5117 handle_parity_checks5(conf, sh, &s, disks);
5120 if ((s.replacing || s.syncing) && s.locked == 0
5121 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
5122 && !test_bit(STRIPE_REPLACED, &sh->state)) {
5123 /* Write out to replacement devices where possible */
5124 for (i = 0; i < conf->raid_disks; i++)
5125 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
5126 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
5127 set_bit(R5_WantReplace, &sh->dev[i].flags);
5128 set_bit(R5_LOCKED, &sh->dev[i].flags);
5132 set_bit(STRIPE_INSYNC, &sh->state);
5133 set_bit(STRIPE_REPLACED, &sh->state);
5135 if ((s.syncing || s.replacing) && s.locked == 0 &&
5136 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5137 test_bit(STRIPE_INSYNC, &sh->state)) {
5138 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5139 clear_bit(STRIPE_SYNCING, &sh->state);
5140 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
5141 wake_up(&conf->wait_for_overlap);
5144 /* If the failed drives are just a ReadError, then we might need
5145 * to progress the repair/check process
5147 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
5148 for (i = 0; i < s.failed; i++) {
5149 struct r5dev *dev = &sh->dev[s.failed_num[i]];
5150 if (test_bit(R5_ReadError, &dev->flags)
5151 && !test_bit(R5_LOCKED, &dev->flags)
5152 && test_bit(R5_UPTODATE, &dev->flags)
5154 if (!test_bit(R5_ReWrite, &dev->flags)) {
5155 set_bit(R5_Wantwrite, &dev->flags);
5156 set_bit(R5_ReWrite, &dev->flags);
5158 /* let's read it back */
5159 set_bit(R5_Wantread, &dev->flags);
5160 set_bit(R5_LOCKED, &dev->flags);
5165 /* Finish reconstruct operations initiated by the expansion process */
5166 if (sh->reconstruct_state == reconstruct_state_result) {
5167 struct stripe_head *sh_src
5168 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
5169 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
5170 /* sh cannot be written until sh_src has been read.
5171 * so arrange for sh to be delayed a little
5173 set_bit(STRIPE_DELAYED, &sh->state);
5174 set_bit(STRIPE_HANDLE, &sh->state);
5175 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5177 atomic_inc(&conf->preread_active_stripes);
5178 raid5_release_stripe(sh_src);
5182 raid5_release_stripe(sh_src);
5184 sh->reconstruct_state = reconstruct_state_idle;
5185 clear_bit(STRIPE_EXPANDING, &sh->state);
5186 for (i = conf->raid_disks; i--; ) {
5187 set_bit(R5_Wantwrite, &sh->dev[i].flags);
5188 set_bit(R5_LOCKED, &sh->dev[i].flags);
5193 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5194 !sh->reconstruct_state) {
5195 /* Need to write out all blocks after computing parity */
5196 sh->disks = conf->raid_disks;
5197 stripe_set_idx(sh->sector, conf, 0, sh);
5198 schedule_reconstruction(sh, &s, 1, 1);
5199 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5200 clear_bit(STRIPE_EXPAND_READY, &sh->state);
5201 atomic_dec(&conf->reshape_stripes);
5202 wake_up(&conf->wait_for_overlap);
5203 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5206 if (s.expanding && s.locked == 0 &&
5207 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5208 handle_stripe_expansion(conf, sh);
5211 /* wait for this device to become unblocked */
5212 if (unlikely(s.blocked_rdev)) {
5213 if (conf->mddev->external)
5214 md_wait_for_blocked_rdev(s.blocked_rdev,
5217 /* Internal metadata will immediately
5218 * be written by raid5d, so we don't
5219 * need to wait here.
5221 rdev_dec_pending(s.blocked_rdev,
5225 if (s.handle_bad_blocks)
5226 for (i = disks; i--; ) {
5227 struct md_rdev *rdev;
5228 struct r5dev *dev = &sh->dev[i];
5229 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5230 /* We own a safe reference to the rdev */
5231 rdev = conf->disks[i].rdev;
5232 if (!rdev_set_badblocks(rdev, sh->sector,
5233 RAID5_STRIPE_SECTORS(conf), 0))
5234 md_error(conf->mddev, rdev);
5235 rdev_dec_pending(rdev, conf->mddev);
5237 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5238 rdev = conf->disks[i].rdev;
5239 rdev_clear_badblocks(rdev, sh->sector,
5240 RAID5_STRIPE_SECTORS(conf), 0);
5241 rdev_dec_pending(rdev, conf->mddev);
5243 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5244 rdev = conf->disks[i].replacement;
5246 /* rdev have been moved down */
5247 rdev = conf->disks[i].rdev;
5248 rdev_clear_badblocks(rdev, sh->sector,
5249 RAID5_STRIPE_SECTORS(conf), 0);
5250 rdev_dec_pending(rdev, conf->mddev);
5255 raid_run_ops(sh, s.ops_request);
5259 if (s.dec_preread_active) {
5260 /* We delay this until after ops_run_io so that if make_request
5261 * is waiting on a flush, it won't continue until the writes
5262 * have actually been submitted.
5264 atomic_dec(&conf->preread_active_stripes);
5265 if (atomic_read(&conf->preread_active_stripes) <
5267 md_wakeup_thread(conf->mddev->thread);
5270 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5273 static void raid5_activate_delayed(struct r5conf *conf)
5275 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5276 while (!list_empty(&conf->delayed_list)) {
5277 struct list_head *l = conf->delayed_list.next;
5278 struct stripe_head *sh;
5279 sh = list_entry(l, struct stripe_head, lru);
5281 clear_bit(STRIPE_DELAYED, &sh->state);
5282 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5283 atomic_inc(&conf->preread_active_stripes);
5284 list_add_tail(&sh->lru, &conf->hold_list);
5285 raid5_wakeup_stripe_thread(sh);
5290 static void activate_bit_delay(struct r5conf *conf,
5291 struct list_head *temp_inactive_list)
5293 /* device_lock is held */
5294 struct list_head head;
5295 list_add(&head, &conf->bitmap_list);
5296 list_del_init(&conf->bitmap_list);
5297 while (!list_empty(&head)) {
5298 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5300 list_del_init(&sh->lru);
5301 atomic_inc(&sh->count);
5302 hash = sh->hash_lock_index;
5303 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5307 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5309 struct r5conf *conf = mddev->private;
5310 sector_t sector = bio->bi_iter.bi_sector;
5311 unsigned int chunk_sectors;
5312 unsigned int bio_sectors = bio_sectors(bio);
5314 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5315 return chunk_sectors >=
5316 ((sector & (chunk_sectors - 1)) + bio_sectors);
5320 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5321 * later sampled by raid5d.
5323 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5325 unsigned long flags;
5327 spin_lock_irqsave(&conf->device_lock, flags);
5329 bi->bi_next = conf->retry_read_aligned_list;
5330 conf->retry_read_aligned_list = bi;
5332 spin_unlock_irqrestore(&conf->device_lock, flags);
5333 md_wakeup_thread(conf->mddev->thread);
5336 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5337 unsigned int *offset)
5341 bi = conf->retry_read_aligned;
5343 *offset = conf->retry_read_offset;
5344 conf->retry_read_aligned = NULL;
5347 bi = conf->retry_read_aligned_list;
5349 conf->retry_read_aligned_list = bi->bi_next;
5358 * The "raid5_align_endio" should check if the read succeeded and if it
5359 * did, call bio_endio on the original bio (having bio_put the new bio
5361 * If the read failed..
5363 static void raid5_align_endio(struct bio *bi)
5365 struct md_io_acct *md_io_acct = bi->bi_private;
5366 struct bio *raid_bi = md_io_acct->orig_bio;
5367 struct mddev *mddev;
5368 struct r5conf *conf;
5369 struct md_rdev *rdev;
5370 blk_status_t error = bi->bi_status;
5371 unsigned long start_time = md_io_acct->start_time;
5375 rdev = (void*)raid_bi->bi_next;
5376 raid_bi->bi_next = NULL;
5377 mddev = rdev->mddev;
5378 conf = mddev->private;
5380 rdev_dec_pending(rdev, conf->mddev);
5383 if (blk_queue_io_stat(raid_bi->bi_bdev->bd_disk->queue))
5384 bio_end_io_acct(raid_bi, start_time);
5386 if (atomic_dec_and_test(&conf->active_aligned_reads))
5387 wake_up(&conf->wait_for_quiescent);
5391 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5393 add_bio_to_retry(raid_bi, conf);
5396 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5398 struct r5conf *conf = mddev->private;
5399 struct bio *align_bio;
5400 struct md_rdev *rdev;
5401 sector_t sector, end_sector, first_bad;
5402 int bad_sectors, dd_idx;
5403 struct md_io_acct *md_io_acct;
5406 if (!in_chunk_boundary(mddev, raid_bio)) {
5407 pr_debug("%s: non aligned\n", __func__);
5411 sector = raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector, 0,
5413 end_sector = bio_end_sector(raid_bio);
5416 if (r5c_big_stripe_cached(conf, sector))
5417 goto out_rcu_unlock;
5419 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5420 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5421 rdev->recovery_offset < end_sector) {
5422 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5424 goto out_rcu_unlock;
5425 if (test_bit(Faulty, &rdev->flags) ||
5426 !(test_bit(In_sync, &rdev->flags) ||
5427 rdev->recovery_offset >= end_sector))
5428 goto out_rcu_unlock;
5431 atomic_inc(&rdev->nr_pending);
5434 if (is_badblock(rdev, sector, bio_sectors(raid_bio), &first_bad,
5436 rdev_dec_pending(rdev, mddev);
5440 align_bio = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->io_acct_set);
5441 md_io_acct = container_of(align_bio, struct md_io_acct, bio_clone);
5442 raid_bio->bi_next = (void *)rdev;
5443 if (blk_queue_io_stat(raid_bio->bi_bdev->bd_disk->queue))
5444 md_io_acct->start_time = bio_start_io_acct(raid_bio);
5445 md_io_acct->orig_bio = raid_bio;
5447 bio_set_dev(align_bio, rdev->bdev);
5448 align_bio->bi_end_io = raid5_align_endio;
5449 align_bio->bi_private = md_io_acct;
5450 align_bio->bi_iter.bi_sector = sector;
5452 /* No reshape active, so we can trust rdev->data_offset */
5453 align_bio->bi_iter.bi_sector += rdev->data_offset;
5456 if (conf->quiesce == 0) {
5457 atomic_inc(&conf->active_aligned_reads);
5460 /* need a memory barrier to detect the race with raid5_quiesce() */
5461 if (!did_inc || smp_load_acquire(&conf->quiesce) != 0) {
5462 /* quiesce is in progress, so we need to undo io activation and wait
5465 if (did_inc && atomic_dec_and_test(&conf->active_aligned_reads))
5466 wake_up(&conf->wait_for_quiescent);
5467 spin_lock_irq(&conf->device_lock);
5468 wait_event_lock_irq(conf->wait_for_quiescent, conf->quiesce == 0,
5470 atomic_inc(&conf->active_aligned_reads);
5471 spin_unlock_irq(&conf->device_lock);
5475 trace_block_bio_remap(align_bio, disk_devt(mddev->gendisk),
5476 raid_bio->bi_iter.bi_sector);
5477 submit_bio_noacct(align_bio);
5485 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5488 sector_t sector = raid_bio->bi_iter.bi_sector;
5489 unsigned chunk_sects = mddev->chunk_sectors;
5490 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5492 if (sectors < bio_sectors(raid_bio)) {
5493 struct r5conf *conf = mddev->private;
5494 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5495 bio_chain(split, raid_bio);
5496 submit_bio_noacct(raid_bio);
5500 if (!raid5_read_one_chunk(mddev, raid_bio))
5506 /* __get_priority_stripe - get the next stripe to process
5508 * Full stripe writes are allowed to pass preread active stripes up until
5509 * the bypass_threshold is exceeded. In general the bypass_count
5510 * increments when the handle_list is handled before the hold_list; however, it
5511 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5512 * stripe with in flight i/o. The bypass_count will be reset when the
5513 * head of the hold_list has changed, i.e. the head was promoted to the
5516 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5518 struct stripe_head *sh, *tmp;
5519 struct list_head *handle_list = NULL;
5520 struct r5worker_group *wg;
5521 bool second_try = !r5c_is_writeback(conf->log) &&
5522 !r5l_log_disk_error(conf);
5523 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5524 r5l_log_disk_error(conf);
5529 if (conf->worker_cnt_per_group == 0) {
5530 handle_list = try_loprio ? &conf->loprio_list :
5532 } else if (group != ANY_GROUP) {
5533 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5534 &conf->worker_groups[group].handle_list;
5535 wg = &conf->worker_groups[group];
5538 for (i = 0; i < conf->group_cnt; i++) {
5539 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5540 &conf->worker_groups[i].handle_list;
5541 wg = &conf->worker_groups[i];
5542 if (!list_empty(handle_list))
5547 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5549 list_empty(handle_list) ? "empty" : "busy",
5550 list_empty(&conf->hold_list) ? "empty" : "busy",
5551 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5553 if (!list_empty(handle_list)) {
5554 sh = list_entry(handle_list->next, typeof(*sh), lru);
5556 if (list_empty(&conf->hold_list))
5557 conf->bypass_count = 0;
5558 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5559 if (conf->hold_list.next == conf->last_hold)
5560 conf->bypass_count++;
5562 conf->last_hold = conf->hold_list.next;
5563 conf->bypass_count -= conf->bypass_threshold;
5564 if (conf->bypass_count < 0)
5565 conf->bypass_count = 0;
5568 } else if (!list_empty(&conf->hold_list) &&
5569 ((conf->bypass_threshold &&
5570 conf->bypass_count > conf->bypass_threshold) ||
5571 atomic_read(&conf->pending_full_writes) == 0)) {
5573 list_for_each_entry(tmp, &conf->hold_list, lru) {
5574 if (conf->worker_cnt_per_group == 0 ||
5575 group == ANY_GROUP ||
5576 !cpu_online(tmp->cpu) ||
5577 cpu_to_group(tmp->cpu) == group) {
5584 conf->bypass_count -= conf->bypass_threshold;
5585 if (conf->bypass_count < 0)
5586 conf->bypass_count = 0;
5595 try_loprio = !try_loprio;
5603 list_del_init(&sh->lru);
5604 BUG_ON(atomic_inc_return(&sh->count) != 1);
5608 struct raid5_plug_cb {
5609 struct blk_plug_cb cb;
5610 struct list_head list;
5611 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5614 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5616 struct raid5_plug_cb *cb = container_of(
5617 blk_cb, struct raid5_plug_cb, cb);
5618 struct stripe_head *sh;
5619 struct mddev *mddev = cb->cb.data;
5620 struct r5conf *conf = mddev->private;
5624 if (cb->list.next && !list_empty(&cb->list)) {
5625 spin_lock_irq(&conf->device_lock);
5626 while (!list_empty(&cb->list)) {
5627 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5628 list_del_init(&sh->lru);
5630 * avoid race release_stripe_plug() sees
5631 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5632 * is still in our list
5634 smp_mb__before_atomic();
5635 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5637 * STRIPE_ON_RELEASE_LIST could be set here. In that
5638 * case, the count is always > 1 here
5640 hash = sh->hash_lock_index;
5641 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5644 spin_unlock_irq(&conf->device_lock);
5646 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5647 NR_STRIPE_HASH_LOCKS);
5649 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5653 static void release_stripe_plug(struct mddev *mddev,
5654 struct stripe_head *sh)
5656 struct blk_plug_cb *blk_cb = blk_check_plugged(
5657 raid5_unplug, mddev,
5658 sizeof(struct raid5_plug_cb));
5659 struct raid5_plug_cb *cb;
5662 raid5_release_stripe(sh);
5666 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5668 if (cb->list.next == NULL) {
5670 INIT_LIST_HEAD(&cb->list);
5671 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5672 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5675 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5676 list_add_tail(&sh->lru, &cb->list);
5678 raid5_release_stripe(sh);
5681 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5683 struct r5conf *conf = mddev->private;
5684 sector_t logical_sector, last_sector;
5685 struct stripe_head *sh;
5688 if (mddev->reshape_position != MaxSector)
5689 /* Skip discard while reshape is happening */
5692 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5693 last_sector = bio_end_sector(bi);
5697 stripe_sectors = conf->chunk_sectors *
5698 (conf->raid_disks - conf->max_degraded);
5699 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5701 sector_div(last_sector, stripe_sectors);
5703 logical_sector *= conf->chunk_sectors;
5704 last_sector *= conf->chunk_sectors;
5706 for (; logical_sector < last_sector;
5707 logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5711 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5712 prepare_to_wait(&conf->wait_for_overlap, &w,
5713 TASK_UNINTERRUPTIBLE);
5714 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5715 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5716 raid5_release_stripe(sh);
5720 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5721 spin_lock_irq(&sh->stripe_lock);
5722 for (d = 0; d < conf->raid_disks; d++) {
5723 if (d == sh->pd_idx || d == sh->qd_idx)
5725 if (sh->dev[d].towrite || sh->dev[d].toread) {
5726 set_bit(R5_Overlap, &sh->dev[d].flags);
5727 spin_unlock_irq(&sh->stripe_lock);
5728 raid5_release_stripe(sh);
5733 set_bit(STRIPE_DISCARD, &sh->state);
5734 finish_wait(&conf->wait_for_overlap, &w);
5735 sh->overwrite_disks = 0;
5736 for (d = 0; d < conf->raid_disks; d++) {
5737 if (d == sh->pd_idx || d == sh->qd_idx)
5739 sh->dev[d].towrite = bi;
5740 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5741 bio_inc_remaining(bi);
5742 md_write_inc(mddev, bi);
5743 sh->overwrite_disks++;
5745 spin_unlock_irq(&sh->stripe_lock);
5746 if (conf->mddev->bitmap) {
5748 d < conf->raid_disks - conf->max_degraded;
5750 md_bitmap_startwrite(mddev->bitmap,
5752 RAID5_STRIPE_SECTORS(conf),
5754 sh->bm_seq = conf->seq_flush + 1;
5755 set_bit(STRIPE_BIT_DELAY, &sh->state);
5758 set_bit(STRIPE_HANDLE, &sh->state);
5759 clear_bit(STRIPE_DELAYED, &sh->state);
5760 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5761 atomic_inc(&conf->preread_active_stripes);
5762 release_stripe_plug(mddev, sh);
5768 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5770 struct r5conf *conf = mddev->private;
5772 sector_t new_sector;
5773 sector_t logical_sector, last_sector;
5774 struct stripe_head *sh;
5775 const int rw = bio_data_dir(bi);
5778 bool do_flush = false;
5780 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5781 int ret = log_handle_flush_request(conf, bi);
5785 if (ret == -ENODEV) {
5786 if (md_flush_request(mddev, bi))
5789 /* ret == -EAGAIN, fallback */
5791 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5792 * we need to flush journal device
5794 do_flush = bi->bi_opf & REQ_PREFLUSH;
5797 if (!md_write_start(mddev, bi))
5800 * If array is degraded, better not do chunk aligned read because
5801 * later we might have to read it again in order to reconstruct
5802 * data on failed drives.
5804 if (rw == READ && mddev->degraded == 0 &&
5805 mddev->reshape_position == MaxSector) {
5806 bi = chunk_aligned_read(mddev, bi);
5811 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5812 make_discard_request(mddev, bi);
5813 md_write_end(mddev);
5817 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5818 last_sector = bio_end_sector(bi);
5821 md_account_bio(mddev, &bi);
5822 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5823 for (; logical_sector < last_sector; logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5829 seq = read_seqcount_begin(&conf->gen_lock);
5832 prepare_to_wait(&conf->wait_for_overlap, &w,
5833 TASK_UNINTERRUPTIBLE);
5834 if (unlikely(conf->reshape_progress != MaxSector)) {
5835 /* spinlock is needed as reshape_progress may be
5836 * 64bit on a 32bit platform, and so it might be
5837 * possible to see a half-updated value
5838 * Of course reshape_progress could change after
5839 * the lock is dropped, so once we get a reference
5840 * to the stripe that we think it is, we will have
5843 spin_lock_irq(&conf->device_lock);
5844 if (mddev->reshape_backwards
5845 ? logical_sector < conf->reshape_progress
5846 : logical_sector >= conf->reshape_progress) {
5849 if (mddev->reshape_backwards
5850 ? logical_sector < conf->reshape_safe
5851 : logical_sector >= conf->reshape_safe) {
5852 spin_unlock_irq(&conf->device_lock);
5858 spin_unlock_irq(&conf->device_lock);
5861 new_sector = raid5_compute_sector(conf, logical_sector,
5864 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5865 (unsigned long long)new_sector,
5866 (unsigned long long)logical_sector);
5868 sh = raid5_get_active_stripe(conf, new_sector, previous,
5869 (bi->bi_opf & REQ_RAHEAD), 0);
5871 if (unlikely(previous)) {
5872 /* expansion might have moved on while waiting for a
5873 * stripe, so we must do the range check again.
5874 * Expansion could still move past after this
5875 * test, but as we are holding a reference to
5876 * 'sh', we know that if that happens,
5877 * STRIPE_EXPANDING will get set and the expansion
5878 * won't proceed until we finish with the stripe.
5881 spin_lock_irq(&conf->device_lock);
5882 if (mddev->reshape_backwards
5883 ? logical_sector >= conf->reshape_progress
5884 : logical_sector < conf->reshape_progress)
5885 /* mismatch, need to try again */
5887 spin_unlock_irq(&conf->device_lock);
5889 raid5_release_stripe(sh);
5895 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5896 /* Might have got the wrong stripe_head
5899 raid5_release_stripe(sh);
5903 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5904 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5905 /* Stripe is busy expanding or
5906 * add failed due to overlap. Flush everything
5909 md_wakeup_thread(mddev->thread);
5910 raid5_release_stripe(sh);
5916 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5917 /* we only need flush for one stripe */
5921 set_bit(STRIPE_HANDLE, &sh->state);
5922 clear_bit(STRIPE_DELAYED, &sh->state);
5923 if ((!sh->batch_head || sh == sh->batch_head) &&
5924 (bi->bi_opf & REQ_SYNC) &&
5925 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5926 atomic_inc(&conf->preread_active_stripes);
5927 release_stripe_plug(mddev, sh);
5929 /* cannot get stripe for read-ahead, just give-up */
5930 bi->bi_status = BLK_STS_IOERR;
5934 finish_wait(&conf->wait_for_overlap, &w);
5937 md_write_end(mddev);
5942 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5944 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5946 /* reshaping is quite different to recovery/resync so it is
5947 * handled quite separately ... here.
5949 * On each call to sync_request, we gather one chunk worth of
5950 * destination stripes and flag them as expanding.
5951 * Then we find all the source stripes and request reads.
5952 * As the reads complete, handle_stripe will copy the data
5953 * into the destination stripe and release that stripe.
5955 struct r5conf *conf = mddev->private;
5956 struct stripe_head *sh;
5957 struct md_rdev *rdev;
5958 sector_t first_sector, last_sector;
5959 int raid_disks = conf->previous_raid_disks;
5960 int data_disks = raid_disks - conf->max_degraded;
5961 int new_data_disks = conf->raid_disks - conf->max_degraded;
5964 sector_t writepos, readpos, safepos;
5965 sector_t stripe_addr;
5966 int reshape_sectors;
5967 struct list_head stripes;
5970 if (sector_nr == 0) {
5971 /* If restarting in the middle, skip the initial sectors */
5972 if (mddev->reshape_backwards &&
5973 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5974 sector_nr = raid5_size(mddev, 0, 0)
5975 - conf->reshape_progress;
5976 } else if (mddev->reshape_backwards &&
5977 conf->reshape_progress == MaxSector) {
5978 /* shouldn't happen, but just in case, finish up.*/
5979 sector_nr = MaxSector;
5980 } else if (!mddev->reshape_backwards &&
5981 conf->reshape_progress > 0)
5982 sector_nr = conf->reshape_progress;
5983 sector_div(sector_nr, new_data_disks);
5985 mddev->curr_resync_completed = sector_nr;
5986 sysfs_notify_dirent_safe(mddev->sysfs_completed);
5993 /* We need to process a full chunk at a time.
5994 * If old and new chunk sizes differ, we need to process the
5998 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
6000 /* We update the metadata at least every 10 seconds, or when
6001 * the data about to be copied would over-write the source of
6002 * the data at the front of the range. i.e. one new_stripe
6003 * along from reshape_progress new_maps to after where
6004 * reshape_safe old_maps to
6006 writepos = conf->reshape_progress;
6007 sector_div(writepos, new_data_disks);
6008 readpos = conf->reshape_progress;
6009 sector_div(readpos, data_disks);
6010 safepos = conf->reshape_safe;
6011 sector_div(safepos, data_disks);
6012 if (mddev->reshape_backwards) {
6013 BUG_ON(writepos < reshape_sectors);
6014 writepos -= reshape_sectors;
6015 readpos += reshape_sectors;
6016 safepos += reshape_sectors;
6018 writepos += reshape_sectors;
6019 /* readpos and safepos are worst-case calculations.
6020 * A negative number is overly pessimistic, and causes
6021 * obvious problems for unsigned storage. So clip to 0.
6023 readpos -= min_t(sector_t, reshape_sectors, readpos);
6024 safepos -= min_t(sector_t, reshape_sectors, safepos);
6027 /* Having calculated the 'writepos' possibly use it
6028 * to set 'stripe_addr' which is where we will write to.
6030 if (mddev->reshape_backwards) {
6031 BUG_ON(conf->reshape_progress == 0);
6032 stripe_addr = writepos;
6033 BUG_ON((mddev->dev_sectors &
6034 ~((sector_t)reshape_sectors - 1))
6035 - reshape_sectors - stripe_addr
6038 BUG_ON(writepos != sector_nr + reshape_sectors);
6039 stripe_addr = sector_nr;
6042 /* 'writepos' is the most advanced device address we might write.
6043 * 'readpos' is the least advanced device address we might read.
6044 * 'safepos' is the least address recorded in the metadata as having
6046 * If there is a min_offset_diff, these are adjusted either by
6047 * increasing the safepos/readpos if diff is negative, or
6048 * increasing writepos if diff is positive.
6049 * If 'readpos' is then behind 'writepos', there is no way that we can
6050 * ensure safety in the face of a crash - that must be done by userspace
6051 * making a backup of the data. So in that case there is no particular
6052 * rush to update metadata.
6053 * Otherwise if 'safepos' is behind 'writepos', then we really need to
6054 * update the metadata to advance 'safepos' to match 'readpos' so that
6055 * we can be safe in the event of a crash.
6056 * So we insist on updating metadata if safepos is behind writepos and
6057 * readpos is beyond writepos.
6058 * In any case, update the metadata every 10 seconds.
6059 * Maybe that number should be configurable, but I'm not sure it is
6060 * worth it.... maybe it could be a multiple of safemode_delay???
6062 if (conf->min_offset_diff < 0) {
6063 safepos += -conf->min_offset_diff;
6064 readpos += -conf->min_offset_diff;
6066 writepos += conf->min_offset_diff;
6068 if ((mddev->reshape_backwards
6069 ? (safepos > writepos && readpos < writepos)
6070 : (safepos < writepos && readpos > writepos)) ||
6071 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
6072 /* Cannot proceed until we've updated the superblock... */
6073 wait_event(conf->wait_for_overlap,
6074 atomic_read(&conf->reshape_stripes)==0
6075 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6076 if (atomic_read(&conf->reshape_stripes) != 0)
6078 mddev->reshape_position = conf->reshape_progress;
6079 mddev->curr_resync_completed = sector_nr;
6080 if (!mddev->reshape_backwards)
6081 /* Can update recovery_offset */
6082 rdev_for_each(rdev, mddev)
6083 if (rdev->raid_disk >= 0 &&
6084 !test_bit(Journal, &rdev->flags) &&
6085 !test_bit(In_sync, &rdev->flags) &&
6086 rdev->recovery_offset < sector_nr)
6087 rdev->recovery_offset = sector_nr;
6089 conf->reshape_checkpoint = jiffies;
6090 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6091 md_wakeup_thread(mddev->thread);
6092 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
6093 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6094 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6096 spin_lock_irq(&conf->device_lock);
6097 conf->reshape_safe = mddev->reshape_position;
6098 spin_unlock_irq(&conf->device_lock);
6099 wake_up(&conf->wait_for_overlap);
6100 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6103 INIT_LIST_HEAD(&stripes);
6104 for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
6106 int skipped_disk = 0;
6107 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
6108 set_bit(STRIPE_EXPANDING, &sh->state);
6109 atomic_inc(&conf->reshape_stripes);
6110 /* If any of this stripe is beyond the end of the old
6111 * array, then we need to zero those blocks
6113 for (j=sh->disks; j--;) {
6115 if (j == sh->pd_idx)
6117 if (conf->level == 6 &&
6120 s = raid5_compute_blocknr(sh, j, 0);
6121 if (s < raid5_size(mddev, 0, 0)) {
6125 memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
6126 set_bit(R5_Expanded, &sh->dev[j].flags);
6127 set_bit(R5_UPTODATE, &sh->dev[j].flags);
6129 if (!skipped_disk) {
6130 set_bit(STRIPE_EXPAND_READY, &sh->state);
6131 set_bit(STRIPE_HANDLE, &sh->state);
6133 list_add(&sh->lru, &stripes);
6135 spin_lock_irq(&conf->device_lock);
6136 if (mddev->reshape_backwards)
6137 conf->reshape_progress -= reshape_sectors * new_data_disks;
6139 conf->reshape_progress += reshape_sectors * new_data_disks;
6140 spin_unlock_irq(&conf->device_lock);
6141 /* Ok, those stripe are ready. We can start scheduling
6142 * reads on the source stripes.
6143 * The source stripes are determined by mapping the first and last
6144 * block on the destination stripes.
6147 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
6150 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
6151 * new_data_disks - 1),
6153 if (last_sector >= mddev->dev_sectors)
6154 last_sector = mddev->dev_sectors - 1;
6155 while (first_sector <= last_sector) {
6156 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
6157 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
6158 set_bit(STRIPE_HANDLE, &sh->state);
6159 raid5_release_stripe(sh);
6160 first_sector += RAID5_STRIPE_SECTORS(conf);
6162 /* Now that the sources are clearly marked, we can release
6163 * the destination stripes
6165 while (!list_empty(&stripes)) {
6166 sh = list_entry(stripes.next, struct stripe_head, lru);
6167 list_del_init(&sh->lru);
6168 raid5_release_stripe(sh);
6170 /* If this takes us to the resync_max point where we have to pause,
6171 * then we need to write out the superblock.
6173 sector_nr += reshape_sectors;
6174 retn = reshape_sectors;
6176 if (mddev->curr_resync_completed > mddev->resync_max ||
6177 (sector_nr - mddev->curr_resync_completed) * 2
6178 >= mddev->resync_max - mddev->curr_resync_completed) {
6179 /* Cannot proceed until we've updated the superblock... */
6180 wait_event(conf->wait_for_overlap,
6181 atomic_read(&conf->reshape_stripes) == 0
6182 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6183 if (atomic_read(&conf->reshape_stripes) != 0)
6185 mddev->reshape_position = conf->reshape_progress;
6186 mddev->curr_resync_completed = sector_nr;
6187 if (!mddev->reshape_backwards)
6188 /* Can update recovery_offset */
6189 rdev_for_each(rdev, mddev)
6190 if (rdev->raid_disk >= 0 &&
6191 !test_bit(Journal, &rdev->flags) &&
6192 !test_bit(In_sync, &rdev->flags) &&
6193 rdev->recovery_offset < sector_nr)
6194 rdev->recovery_offset = sector_nr;
6195 conf->reshape_checkpoint = jiffies;
6196 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6197 md_wakeup_thread(mddev->thread);
6198 wait_event(mddev->sb_wait,
6199 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6200 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6201 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6203 spin_lock_irq(&conf->device_lock);
6204 conf->reshape_safe = mddev->reshape_position;
6205 spin_unlock_irq(&conf->device_lock);
6206 wake_up(&conf->wait_for_overlap);
6207 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6213 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6216 struct r5conf *conf = mddev->private;
6217 struct stripe_head *sh;
6218 sector_t max_sector = mddev->dev_sectors;
6219 sector_t sync_blocks;
6220 int still_degraded = 0;
6223 if (sector_nr >= max_sector) {
6224 /* just being told to finish up .. nothing much to do */
6226 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6231 if (mddev->curr_resync < max_sector) /* aborted */
6232 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6234 else /* completed sync */
6236 md_bitmap_close_sync(mddev->bitmap);
6241 /* Allow raid5_quiesce to complete */
6242 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6244 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6245 return reshape_request(mddev, sector_nr, skipped);
6247 /* No need to check resync_max as we never do more than one
6248 * stripe, and as resync_max will always be on a chunk boundary,
6249 * if the check in md_do_sync didn't fire, there is no chance
6250 * of overstepping resync_max here
6253 /* if there is too many failed drives and we are trying
6254 * to resync, then assert that we are finished, because there is
6255 * nothing we can do.
6257 if (mddev->degraded >= conf->max_degraded &&
6258 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6259 sector_t rv = mddev->dev_sectors - sector_nr;
6263 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6265 !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6266 sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6267 /* we can skip this block, and probably more */
6268 do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6270 /* keep things rounded to whole stripes */
6271 return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6274 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6276 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6278 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6279 /* make sure we don't swamp the stripe cache if someone else
6280 * is trying to get access
6282 schedule_timeout_uninterruptible(1);
6284 /* Need to check if array will still be degraded after recovery/resync
6285 * Note in case of > 1 drive failures it's possible we're rebuilding
6286 * one drive while leaving another faulty drive in array.
6289 for (i = 0; i < conf->raid_disks; i++) {
6290 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6292 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6297 md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6299 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6300 set_bit(STRIPE_HANDLE, &sh->state);
6302 raid5_release_stripe(sh);
6304 return RAID5_STRIPE_SECTORS(conf);
6307 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6308 unsigned int offset)
6310 /* We may not be able to submit a whole bio at once as there
6311 * may not be enough stripe_heads available.
6312 * We cannot pre-allocate enough stripe_heads as we may need
6313 * more than exist in the cache (if we allow ever large chunks).
6314 * So we do one stripe head at a time and record in
6315 * ->bi_hw_segments how many have been done.
6317 * We *know* that this entire raid_bio is in one chunk, so
6318 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6320 struct stripe_head *sh;
6322 sector_t sector, logical_sector, last_sector;
6326 logical_sector = raid_bio->bi_iter.bi_sector &
6327 ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6328 sector = raid5_compute_sector(conf, logical_sector,
6330 last_sector = bio_end_sector(raid_bio);
6332 for (; logical_sector < last_sector;
6333 logical_sector += RAID5_STRIPE_SECTORS(conf),
6334 sector += RAID5_STRIPE_SECTORS(conf),
6338 /* already done this stripe */
6341 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6344 /* failed to get a stripe - must wait */
6345 conf->retry_read_aligned = raid_bio;
6346 conf->retry_read_offset = scnt;
6350 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6351 raid5_release_stripe(sh);
6352 conf->retry_read_aligned = raid_bio;
6353 conf->retry_read_offset = scnt;
6357 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6359 raid5_release_stripe(sh);
6363 bio_endio(raid_bio);
6365 if (atomic_dec_and_test(&conf->active_aligned_reads))
6366 wake_up(&conf->wait_for_quiescent);
6370 static int handle_active_stripes(struct r5conf *conf, int group,
6371 struct r5worker *worker,
6372 struct list_head *temp_inactive_list)
6373 __releases(&conf->device_lock)
6374 __acquires(&conf->device_lock)
6376 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6377 int i, batch_size = 0, hash;
6378 bool release_inactive = false;
6380 while (batch_size < MAX_STRIPE_BATCH &&
6381 (sh = __get_priority_stripe(conf, group)) != NULL)
6382 batch[batch_size++] = sh;
6384 if (batch_size == 0) {
6385 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6386 if (!list_empty(temp_inactive_list + i))
6388 if (i == NR_STRIPE_HASH_LOCKS) {
6389 spin_unlock_irq(&conf->device_lock);
6390 log_flush_stripe_to_raid(conf);
6391 spin_lock_irq(&conf->device_lock);
6394 release_inactive = true;
6396 spin_unlock_irq(&conf->device_lock);
6398 release_inactive_stripe_list(conf, temp_inactive_list,
6399 NR_STRIPE_HASH_LOCKS);
6401 r5l_flush_stripe_to_raid(conf->log);
6402 if (release_inactive) {
6403 spin_lock_irq(&conf->device_lock);
6407 for (i = 0; i < batch_size; i++)
6408 handle_stripe(batch[i]);
6409 log_write_stripe_run(conf);
6413 spin_lock_irq(&conf->device_lock);
6414 for (i = 0; i < batch_size; i++) {
6415 hash = batch[i]->hash_lock_index;
6416 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6421 static void raid5_do_work(struct work_struct *work)
6423 struct r5worker *worker = container_of(work, struct r5worker, work);
6424 struct r5worker_group *group = worker->group;
6425 struct r5conf *conf = group->conf;
6426 struct mddev *mddev = conf->mddev;
6427 int group_id = group - conf->worker_groups;
6429 struct blk_plug plug;
6431 pr_debug("+++ raid5worker active\n");
6433 blk_start_plug(&plug);
6435 spin_lock_irq(&conf->device_lock);
6437 int batch_size, released;
6439 released = release_stripe_list(conf, worker->temp_inactive_list);
6441 batch_size = handle_active_stripes(conf, group_id, worker,
6442 worker->temp_inactive_list);
6443 worker->working = false;
6444 if (!batch_size && !released)
6446 handled += batch_size;
6447 wait_event_lock_irq(mddev->sb_wait,
6448 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6451 pr_debug("%d stripes handled\n", handled);
6453 spin_unlock_irq(&conf->device_lock);
6455 flush_deferred_bios(conf);
6457 r5l_flush_stripe_to_raid(conf->log);
6459 async_tx_issue_pending_all();
6460 blk_finish_plug(&plug);
6462 pr_debug("--- raid5worker inactive\n");
6466 * This is our raid5 kernel thread.
6468 * We scan the hash table for stripes which can be handled now.
6469 * During the scan, completed stripes are saved for us by the interrupt
6470 * handler, so that they will not have to wait for our next wakeup.
6472 static void raid5d(struct md_thread *thread)
6474 struct mddev *mddev = thread->mddev;
6475 struct r5conf *conf = mddev->private;
6477 struct blk_plug plug;
6479 pr_debug("+++ raid5d active\n");
6481 md_check_recovery(mddev);
6483 blk_start_plug(&plug);
6485 spin_lock_irq(&conf->device_lock);
6488 int batch_size, released;
6489 unsigned int offset;
6491 released = release_stripe_list(conf, conf->temp_inactive_list);
6493 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6496 !list_empty(&conf->bitmap_list)) {
6497 /* Now is a good time to flush some bitmap updates */
6499 spin_unlock_irq(&conf->device_lock);
6500 md_bitmap_unplug(mddev->bitmap);
6501 spin_lock_irq(&conf->device_lock);
6502 conf->seq_write = conf->seq_flush;
6503 activate_bit_delay(conf, conf->temp_inactive_list);
6505 raid5_activate_delayed(conf);
6507 while ((bio = remove_bio_from_retry(conf, &offset))) {
6509 spin_unlock_irq(&conf->device_lock);
6510 ok = retry_aligned_read(conf, bio, offset);
6511 spin_lock_irq(&conf->device_lock);
6517 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6518 conf->temp_inactive_list);
6519 if (!batch_size && !released)
6521 handled += batch_size;
6523 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6524 spin_unlock_irq(&conf->device_lock);
6525 md_check_recovery(mddev);
6526 spin_lock_irq(&conf->device_lock);
6529 * Waiting on MD_SB_CHANGE_PENDING below may deadlock
6530 * seeing md_check_recovery() is needed to clear
6531 * the flag when using mdmon.
6536 wait_event_lock_irq(mddev->sb_wait,
6537 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6540 pr_debug("%d stripes handled\n", handled);
6542 spin_unlock_irq(&conf->device_lock);
6543 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6544 mutex_trylock(&conf->cache_size_mutex)) {
6545 grow_one_stripe(conf, __GFP_NOWARN);
6546 /* Set flag even if allocation failed. This helps
6547 * slow down allocation requests when mem is short
6549 set_bit(R5_DID_ALLOC, &conf->cache_state);
6550 mutex_unlock(&conf->cache_size_mutex);
6553 flush_deferred_bios(conf);
6555 r5l_flush_stripe_to_raid(conf->log);
6557 async_tx_issue_pending_all();
6558 blk_finish_plug(&plug);
6560 pr_debug("--- raid5d inactive\n");
6564 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6566 struct r5conf *conf;
6568 spin_lock(&mddev->lock);
6569 conf = mddev->private;
6571 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6572 spin_unlock(&mddev->lock);
6577 raid5_set_cache_size(struct mddev *mddev, int size)
6580 struct r5conf *conf = mddev->private;
6582 if (size <= 16 || size > 32768)
6585 conf->min_nr_stripes = size;
6586 mutex_lock(&conf->cache_size_mutex);
6587 while (size < conf->max_nr_stripes &&
6588 drop_one_stripe(conf))
6590 mutex_unlock(&conf->cache_size_mutex);
6592 md_allow_write(mddev);
6594 mutex_lock(&conf->cache_size_mutex);
6595 while (size > conf->max_nr_stripes)
6596 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6597 conf->min_nr_stripes = conf->max_nr_stripes;
6601 mutex_unlock(&conf->cache_size_mutex);
6605 EXPORT_SYMBOL(raid5_set_cache_size);
6608 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6610 struct r5conf *conf;
6614 if (len >= PAGE_SIZE)
6616 if (kstrtoul(page, 10, &new))
6618 err = mddev_lock(mddev);
6621 conf = mddev->private;
6625 err = raid5_set_cache_size(mddev, new);
6626 mddev_unlock(mddev);
6631 static struct md_sysfs_entry
6632 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6633 raid5_show_stripe_cache_size,
6634 raid5_store_stripe_cache_size);
6637 raid5_show_rmw_level(struct mddev *mddev, char *page)
6639 struct r5conf *conf = mddev->private;
6641 return sprintf(page, "%d\n", conf->rmw_level);
6647 raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6649 struct r5conf *conf = mddev->private;
6655 if (len >= PAGE_SIZE)
6658 if (kstrtoul(page, 10, &new))
6661 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6664 if (new != PARITY_DISABLE_RMW &&
6665 new != PARITY_ENABLE_RMW &&
6666 new != PARITY_PREFER_RMW)
6669 conf->rmw_level = new;
6673 static struct md_sysfs_entry
6674 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6675 raid5_show_rmw_level,
6676 raid5_store_rmw_level);
6679 raid5_show_stripe_size(struct mddev *mddev, char *page)
6681 struct r5conf *conf;
6684 spin_lock(&mddev->lock);
6685 conf = mddev->private;
6687 ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6688 spin_unlock(&mddev->lock);
6692 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6694 raid5_store_stripe_size(struct mddev *mddev, const char *page, size_t len)
6696 struct r5conf *conf;
6701 if (len >= PAGE_SIZE)
6703 if (kstrtoul(page, 10, &new))
6707 * The value should not be bigger than PAGE_SIZE. It requires to
6708 * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6711 if (new % DEFAULT_STRIPE_SIZE != 0 ||
6712 new > PAGE_SIZE || new == 0 ||
6713 new != roundup_pow_of_two(new))
6716 err = mddev_lock(mddev);
6720 conf = mddev->private;
6726 if (new == conf->stripe_size)
6729 pr_debug("md/raid: change stripe_size from %lu to %lu\n",
6730 conf->stripe_size, new);
6732 if (mddev->sync_thread ||
6733 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
6734 mddev->reshape_position != MaxSector ||
6735 mddev->sysfs_active) {
6740 mddev_suspend(mddev);
6741 mutex_lock(&conf->cache_size_mutex);
6742 size = conf->max_nr_stripes;
6744 shrink_stripes(conf);
6746 conf->stripe_size = new;
6747 conf->stripe_shift = ilog2(new) - 9;
6748 conf->stripe_sectors = new >> 9;
6749 if (grow_stripes(conf, size)) {
6750 pr_warn("md/raid:%s: couldn't allocate buffers\n",
6754 mutex_unlock(&conf->cache_size_mutex);
6755 mddev_resume(mddev);
6758 mddev_unlock(mddev);
6762 static struct md_sysfs_entry
6763 raid5_stripe_size = __ATTR(stripe_size, 0644,
6764 raid5_show_stripe_size,
6765 raid5_store_stripe_size);
6767 static struct md_sysfs_entry
6768 raid5_stripe_size = __ATTR(stripe_size, 0444,
6769 raid5_show_stripe_size,
6774 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6776 struct r5conf *conf;
6778 spin_lock(&mddev->lock);
6779 conf = mddev->private;
6781 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6782 spin_unlock(&mddev->lock);
6787 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6789 struct r5conf *conf;
6793 if (len >= PAGE_SIZE)
6795 if (kstrtoul(page, 10, &new))
6798 err = mddev_lock(mddev);
6801 conf = mddev->private;
6804 else if (new > conf->min_nr_stripes)
6807 conf->bypass_threshold = new;
6808 mddev_unlock(mddev);
6812 static struct md_sysfs_entry
6813 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6815 raid5_show_preread_threshold,
6816 raid5_store_preread_threshold);
6819 raid5_show_skip_copy(struct mddev *mddev, char *page)
6821 struct r5conf *conf;
6823 spin_lock(&mddev->lock);
6824 conf = mddev->private;
6826 ret = sprintf(page, "%d\n", conf->skip_copy);
6827 spin_unlock(&mddev->lock);
6832 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6834 struct r5conf *conf;
6838 if (len >= PAGE_SIZE)
6840 if (kstrtoul(page, 10, &new))
6844 err = mddev_lock(mddev);
6847 conf = mddev->private;
6850 else if (new != conf->skip_copy) {
6851 struct request_queue *q = mddev->queue;
6853 mddev_suspend(mddev);
6854 conf->skip_copy = new;
6856 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, q);
6858 blk_queue_flag_clear(QUEUE_FLAG_STABLE_WRITES, q);
6859 mddev_resume(mddev);
6861 mddev_unlock(mddev);
6865 static struct md_sysfs_entry
6866 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6867 raid5_show_skip_copy,
6868 raid5_store_skip_copy);
6871 stripe_cache_active_show(struct mddev *mddev, char *page)
6873 struct r5conf *conf = mddev->private;
6875 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6880 static struct md_sysfs_entry
6881 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6884 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6886 struct r5conf *conf;
6888 spin_lock(&mddev->lock);
6889 conf = mddev->private;
6891 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6892 spin_unlock(&mddev->lock);
6896 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6898 struct r5worker_group **worker_groups);
6900 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6902 struct r5conf *conf;
6905 struct r5worker_group *new_groups, *old_groups;
6908 if (len >= PAGE_SIZE)
6910 if (kstrtouint(page, 10, &new))
6912 /* 8192 should be big enough */
6916 err = mddev_lock(mddev);
6919 conf = mddev->private;
6922 else if (new != conf->worker_cnt_per_group) {
6923 mddev_suspend(mddev);
6925 old_groups = conf->worker_groups;
6927 flush_workqueue(raid5_wq);
6929 err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
6931 spin_lock_irq(&conf->device_lock);
6932 conf->group_cnt = group_cnt;
6933 conf->worker_cnt_per_group = new;
6934 conf->worker_groups = new_groups;
6935 spin_unlock_irq(&conf->device_lock);
6938 kfree(old_groups[0].workers);
6941 mddev_resume(mddev);
6943 mddev_unlock(mddev);
6948 static struct md_sysfs_entry
6949 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6950 raid5_show_group_thread_cnt,
6951 raid5_store_group_thread_cnt);
6953 static struct attribute *raid5_attrs[] = {
6954 &raid5_stripecache_size.attr,
6955 &raid5_stripecache_active.attr,
6956 &raid5_preread_bypass_threshold.attr,
6957 &raid5_group_thread_cnt.attr,
6958 &raid5_skip_copy.attr,
6959 &raid5_rmw_level.attr,
6960 &raid5_stripe_size.attr,
6961 &r5c_journal_mode.attr,
6962 &ppl_write_hint.attr,
6965 static const struct attribute_group raid5_attrs_group = {
6967 .attrs = raid5_attrs,
6970 static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
6971 struct r5worker_group **worker_groups)
6975 struct r5worker *workers;
6979 *worker_groups = NULL;
6982 *group_cnt = num_possible_nodes();
6983 size = sizeof(struct r5worker) * cnt;
6984 workers = kcalloc(size, *group_cnt, GFP_NOIO);
6985 *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6987 if (!*worker_groups || !workers) {
6989 kfree(*worker_groups);
6993 for (i = 0; i < *group_cnt; i++) {
6994 struct r5worker_group *group;
6996 group = &(*worker_groups)[i];
6997 INIT_LIST_HEAD(&group->handle_list);
6998 INIT_LIST_HEAD(&group->loprio_list);
7000 group->workers = workers + i * cnt;
7002 for (j = 0; j < cnt; j++) {
7003 struct r5worker *worker = group->workers + j;
7004 worker->group = group;
7005 INIT_WORK(&worker->work, raid5_do_work);
7007 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
7008 INIT_LIST_HEAD(worker->temp_inactive_list + k);
7015 static void free_thread_groups(struct r5conf *conf)
7017 if (conf->worker_groups)
7018 kfree(conf->worker_groups[0].workers);
7019 kfree(conf->worker_groups);
7020 conf->worker_groups = NULL;
7024 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
7026 struct r5conf *conf = mddev->private;
7029 sectors = mddev->dev_sectors;
7031 /* size is defined by the smallest of previous and new size */
7032 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
7034 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7035 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
7036 return sectors * (raid_disks - conf->max_degraded);
7039 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7041 safe_put_page(percpu->spare_page);
7042 percpu->spare_page = NULL;
7043 kvfree(percpu->scribble);
7044 percpu->scribble = NULL;
7047 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7049 if (conf->level == 6 && !percpu->spare_page) {
7050 percpu->spare_page = alloc_page(GFP_KERNEL);
7051 if (!percpu->spare_page)
7055 if (scribble_alloc(percpu,
7056 max(conf->raid_disks,
7057 conf->previous_raid_disks),
7058 max(conf->chunk_sectors,
7059 conf->prev_chunk_sectors)
7060 / RAID5_STRIPE_SECTORS(conf))) {
7061 free_scratch_buffer(conf, percpu);
7068 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
7070 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7072 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
7076 static void raid5_free_percpu(struct r5conf *conf)
7081 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7082 free_percpu(conf->percpu);
7085 static void free_conf(struct r5conf *conf)
7091 unregister_shrinker(&conf->shrinker);
7092 free_thread_groups(conf);
7093 shrink_stripes(conf);
7094 raid5_free_percpu(conf);
7095 for (i = 0; i < conf->pool_size; i++)
7096 if (conf->disks[i].extra_page)
7097 put_page(conf->disks[i].extra_page);
7099 bioset_exit(&conf->bio_split);
7100 kfree(conf->stripe_hashtbl);
7101 kfree(conf->pending_data);
7105 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
7107 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7108 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
7110 if (alloc_scratch_buffer(conf, percpu)) {
7111 pr_warn("%s: failed memory allocation for cpu%u\n",
7115 spin_lock_init(&per_cpu_ptr(conf->percpu, cpu)->lock);
7119 static int raid5_alloc_percpu(struct r5conf *conf)
7123 conf->percpu = alloc_percpu(struct raid5_percpu);
7127 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7129 conf->scribble_disks = max(conf->raid_disks,
7130 conf->previous_raid_disks);
7131 conf->scribble_sectors = max(conf->chunk_sectors,
7132 conf->prev_chunk_sectors);
7137 static unsigned long raid5_cache_scan(struct shrinker *shrink,
7138 struct shrink_control *sc)
7140 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
7141 unsigned long ret = SHRINK_STOP;
7143 if (mutex_trylock(&conf->cache_size_mutex)) {
7145 while (ret < sc->nr_to_scan &&
7146 conf->max_nr_stripes > conf->min_nr_stripes) {
7147 if (drop_one_stripe(conf) == 0) {
7153 mutex_unlock(&conf->cache_size_mutex);
7158 static unsigned long raid5_cache_count(struct shrinker *shrink,
7159 struct shrink_control *sc)
7161 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
7163 if (conf->max_nr_stripes < conf->min_nr_stripes)
7164 /* unlikely, but not impossible */
7166 return conf->max_nr_stripes - conf->min_nr_stripes;
7169 static struct r5conf *setup_conf(struct mddev *mddev)
7171 struct r5conf *conf;
7172 int raid_disk, memory, max_disks;
7173 struct md_rdev *rdev;
7174 struct disk_info *disk;
7178 struct r5worker_group *new_group;
7181 if (mddev->new_level != 5
7182 && mddev->new_level != 4
7183 && mddev->new_level != 6) {
7184 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7185 mdname(mddev), mddev->new_level);
7186 return ERR_PTR(-EIO);
7188 if ((mddev->new_level == 5
7189 && !algorithm_valid_raid5(mddev->new_layout)) ||
7190 (mddev->new_level == 6
7191 && !algorithm_valid_raid6(mddev->new_layout))) {
7192 pr_warn("md/raid:%s: layout %d not supported\n",
7193 mdname(mddev), mddev->new_layout);
7194 return ERR_PTR(-EIO);
7196 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
7197 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7198 mdname(mddev), mddev->raid_disks);
7199 return ERR_PTR(-EINVAL);
7202 if (!mddev->new_chunk_sectors ||
7203 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
7204 !is_power_of_2(mddev->new_chunk_sectors)) {
7205 pr_warn("md/raid:%s: invalid chunk size %d\n",
7206 mdname(mddev), mddev->new_chunk_sectors << 9);
7207 return ERR_PTR(-EINVAL);
7210 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7214 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7215 conf->stripe_size = DEFAULT_STRIPE_SIZE;
7216 conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7217 conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7219 INIT_LIST_HEAD(&conf->free_list);
7220 INIT_LIST_HEAD(&conf->pending_list);
7221 conf->pending_data = kcalloc(PENDING_IO_MAX,
7222 sizeof(struct r5pending_data),
7224 if (!conf->pending_data)
7226 for (i = 0; i < PENDING_IO_MAX; i++)
7227 list_add(&conf->pending_data[i].sibling, &conf->free_list);
7228 /* Don't enable multi-threading by default*/
7229 if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7230 conf->group_cnt = group_cnt;
7231 conf->worker_cnt_per_group = 0;
7232 conf->worker_groups = new_group;
7235 spin_lock_init(&conf->device_lock);
7236 seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7237 mutex_init(&conf->cache_size_mutex);
7238 init_waitqueue_head(&conf->wait_for_quiescent);
7239 init_waitqueue_head(&conf->wait_for_stripe);
7240 init_waitqueue_head(&conf->wait_for_overlap);
7241 INIT_LIST_HEAD(&conf->handle_list);
7242 INIT_LIST_HEAD(&conf->loprio_list);
7243 INIT_LIST_HEAD(&conf->hold_list);
7244 INIT_LIST_HEAD(&conf->delayed_list);
7245 INIT_LIST_HEAD(&conf->bitmap_list);
7246 init_llist_head(&conf->released_stripes);
7247 atomic_set(&conf->active_stripes, 0);
7248 atomic_set(&conf->preread_active_stripes, 0);
7249 atomic_set(&conf->active_aligned_reads, 0);
7250 spin_lock_init(&conf->pending_bios_lock);
7251 conf->batch_bio_dispatch = true;
7252 rdev_for_each(rdev, mddev) {
7253 if (test_bit(Journal, &rdev->flags))
7255 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
7256 conf->batch_bio_dispatch = false;
7261 conf->bypass_threshold = BYPASS_THRESHOLD;
7262 conf->recovery_disabled = mddev->recovery_disabled - 1;
7264 conf->raid_disks = mddev->raid_disks;
7265 if (mddev->reshape_position == MaxSector)
7266 conf->previous_raid_disks = mddev->raid_disks;
7268 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7269 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
7271 conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7277 for (i = 0; i < max_disks; i++) {
7278 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7279 if (!conf->disks[i].extra_page)
7283 ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7286 conf->mddev = mddev;
7288 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
7291 /* We init hash_locks[0] separately to that it can be used
7292 * as the reference lock in the spin_lock_nest_lock() call
7293 * in lock_all_device_hash_locks_irq in order to convince
7294 * lockdep that we know what we are doing.
7296 spin_lock_init(conf->hash_locks);
7297 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7298 spin_lock_init(conf->hash_locks + i);
7300 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7301 INIT_LIST_HEAD(conf->inactive_list + i);
7303 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7304 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7306 atomic_set(&conf->r5c_cached_full_stripes, 0);
7307 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7308 atomic_set(&conf->r5c_cached_partial_stripes, 0);
7309 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7310 atomic_set(&conf->r5c_flushing_full_stripes, 0);
7311 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7313 conf->level = mddev->new_level;
7314 conf->chunk_sectors = mddev->new_chunk_sectors;
7315 if (raid5_alloc_percpu(conf) != 0)
7318 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7320 rdev_for_each(rdev, mddev) {
7321 raid_disk = rdev->raid_disk;
7322 if (raid_disk >= max_disks
7323 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7325 disk = conf->disks + raid_disk;
7327 if (test_bit(Replacement, &rdev->flags)) {
7328 if (disk->replacement)
7330 disk->replacement = rdev;
7337 if (test_bit(In_sync, &rdev->flags)) {
7338 char b[BDEVNAME_SIZE];
7339 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7340 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7341 } else if (rdev->saved_raid_disk != raid_disk)
7342 /* Cannot rely on bitmap to complete recovery */
7346 conf->level = mddev->new_level;
7347 if (conf->level == 6) {
7348 conf->max_degraded = 2;
7349 if (raid6_call.xor_syndrome)
7350 conf->rmw_level = PARITY_ENABLE_RMW;
7352 conf->rmw_level = PARITY_DISABLE_RMW;
7354 conf->max_degraded = 1;
7355 conf->rmw_level = PARITY_ENABLE_RMW;
7357 conf->algorithm = mddev->new_layout;
7358 conf->reshape_progress = mddev->reshape_position;
7359 if (conf->reshape_progress != MaxSector) {
7360 conf->prev_chunk_sectors = mddev->chunk_sectors;
7361 conf->prev_algo = mddev->layout;
7363 conf->prev_chunk_sectors = conf->chunk_sectors;
7364 conf->prev_algo = conf->algorithm;
7367 conf->min_nr_stripes = NR_STRIPES;
7368 if (mddev->reshape_position != MaxSector) {
7369 int stripes = max_t(int,
7370 ((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7371 ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7372 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7373 if (conf->min_nr_stripes != NR_STRIPES)
7374 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7375 mdname(mddev), conf->min_nr_stripes);
7377 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7378 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7379 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7380 if (grow_stripes(conf, conf->min_nr_stripes)) {
7381 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7382 mdname(mddev), memory);
7385 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7387 * Losing a stripe head costs more than the time to refill it,
7388 * it reduces the queue depth and so can hurt throughput.
7389 * So set it rather large, scaled by number of devices.
7391 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7392 conf->shrinker.scan_objects = raid5_cache_scan;
7393 conf->shrinker.count_objects = raid5_cache_count;
7394 conf->shrinker.batch = 128;
7395 conf->shrinker.flags = 0;
7396 if (register_shrinker(&conf->shrinker)) {
7397 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7402 sprintf(pers_name, "raid%d", mddev->new_level);
7403 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7404 if (!conf->thread) {
7405 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7415 return ERR_PTR(-EIO);
7417 return ERR_PTR(-ENOMEM);
7420 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7423 case ALGORITHM_PARITY_0:
7424 if (raid_disk < max_degraded)
7427 case ALGORITHM_PARITY_N:
7428 if (raid_disk >= raid_disks - max_degraded)
7431 case ALGORITHM_PARITY_0_6:
7432 if (raid_disk == 0 ||
7433 raid_disk == raid_disks - 1)
7436 case ALGORITHM_LEFT_ASYMMETRIC_6:
7437 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7438 case ALGORITHM_LEFT_SYMMETRIC_6:
7439 case ALGORITHM_RIGHT_SYMMETRIC_6:
7440 if (raid_disk == raid_disks - 1)
7446 static void raid5_set_io_opt(struct r5conf *conf)
7448 blk_queue_io_opt(conf->mddev->queue, (conf->chunk_sectors << 9) *
7449 (conf->raid_disks - conf->max_degraded));
7452 static int raid5_run(struct mddev *mddev)
7454 struct r5conf *conf;
7455 int working_disks = 0;
7456 int dirty_parity_disks = 0;
7457 struct md_rdev *rdev;
7458 struct md_rdev *journal_dev = NULL;
7459 sector_t reshape_offset = 0;
7461 long long min_offset_diff = 0;
7464 if (acct_bioset_init(mddev)) {
7465 pr_err("md/raid456:%s: alloc acct bioset failed.\n", mdname(mddev));
7469 if (mddev_init_writes_pending(mddev) < 0) {
7474 if (mddev->recovery_cp != MaxSector)
7475 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7478 rdev_for_each(rdev, mddev) {
7481 if (test_bit(Journal, &rdev->flags)) {
7485 if (rdev->raid_disk < 0)
7487 diff = (rdev->new_data_offset - rdev->data_offset);
7489 min_offset_diff = diff;
7491 } else if (mddev->reshape_backwards &&
7492 diff < min_offset_diff)
7493 min_offset_diff = diff;
7494 else if (!mddev->reshape_backwards &&
7495 diff > min_offset_diff)
7496 min_offset_diff = diff;
7499 if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7500 (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7501 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7507 if (mddev->reshape_position != MaxSector) {
7508 /* Check that we can continue the reshape.
7509 * Difficulties arise if the stripe we would write to
7510 * next is at or after the stripe we would read from next.
7511 * For a reshape that changes the number of devices, this
7512 * is only possible for a very short time, and mdadm makes
7513 * sure that time appears to have past before assembling
7514 * the array. So we fail if that time hasn't passed.
7515 * For a reshape that keeps the number of devices the same
7516 * mdadm must be monitoring the reshape can keeping the
7517 * critical areas read-only and backed up. It will start
7518 * the array in read-only mode, so we check for that.
7520 sector_t here_new, here_old;
7522 int max_degraded = (mddev->level == 6 ? 2 : 1);
7527 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7533 if (mddev->new_level != mddev->level) {
7534 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7539 old_disks = mddev->raid_disks - mddev->delta_disks;
7540 /* reshape_position must be on a new-stripe boundary, and one
7541 * further up in new geometry must map after here in old
7543 * If the chunk sizes are different, then as we perform reshape
7544 * in units of the largest of the two, reshape_position needs
7545 * be a multiple of the largest chunk size times new data disks.
7547 here_new = mddev->reshape_position;
7548 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7549 new_data_disks = mddev->raid_disks - max_degraded;
7550 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7551 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7556 reshape_offset = here_new * chunk_sectors;
7557 /* here_new is the stripe we will write to */
7558 here_old = mddev->reshape_position;
7559 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7560 /* here_old is the first stripe that we might need to read
7562 if (mddev->delta_disks == 0) {
7563 /* We cannot be sure it is safe to start an in-place
7564 * reshape. It is only safe if user-space is monitoring
7565 * and taking constant backups.
7566 * mdadm always starts a situation like this in
7567 * readonly mode so it can take control before
7568 * allowing any writes. So just check for that.
7570 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7571 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7572 /* not really in-place - so OK */;
7573 else if (mddev->ro == 0) {
7574 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7579 } else if (mddev->reshape_backwards
7580 ? (here_new * chunk_sectors + min_offset_diff <=
7581 here_old * chunk_sectors)
7582 : (here_new * chunk_sectors >=
7583 here_old * chunk_sectors + (-min_offset_diff))) {
7584 /* Reading from the same stripe as writing to - bad */
7585 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7590 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7591 /* OK, we should be able to continue; */
7593 BUG_ON(mddev->level != mddev->new_level);
7594 BUG_ON(mddev->layout != mddev->new_layout);
7595 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7596 BUG_ON(mddev->delta_disks != 0);
7599 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7600 test_bit(MD_HAS_PPL, &mddev->flags)) {
7601 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7603 clear_bit(MD_HAS_PPL, &mddev->flags);
7604 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7607 if (mddev->private == NULL)
7608 conf = setup_conf(mddev);
7610 conf = mddev->private;
7613 ret = PTR_ERR(conf);
7617 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7619 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7622 set_disk_ro(mddev->gendisk, 1);
7623 } else if (mddev->recovery_cp == MaxSector)
7624 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7627 conf->min_offset_diff = min_offset_diff;
7628 mddev->thread = conf->thread;
7629 conf->thread = NULL;
7630 mddev->private = conf;
7632 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7634 rdev = conf->disks[i].rdev;
7635 if (!rdev && conf->disks[i].replacement) {
7636 /* The replacement is all we have yet */
7637 rdev = conf->disks[i].replacement;
7638 conf->disks[i].replacement = NULL;
7639 clear_bit(Replacement, &rdev->flags);
7640 conf->disks[i].rdev = rdev;
7644 if (conf->disks[i].replacement &&
7645 conf->reshape_progress != MaxSector) {
7646 /* replacements and reshape simply do not mix. */
7647 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7650 if (test_bit(In_sync, &rdev->flags)) {
7654 /* This disc is not fully in-sync. However if it
7655 * just stored parity (beyond the recovery_offset),
7656 * when we don't need to be concerned about the
7657 * array being dirty.
7658 * When reshape goes 'backwards', we never have
7659 * partially completed devices, so we only need
7660 * to worry about reshape going forwards.
7662 /* Hack because v0.91 doesn't store recovery_offset properly. */
7663 if (mddev->major_version == 0 &&
7664 mddev->minor_version > 90)
7665 rdev->recovery_offset = reshape_offset;
7667 if (rdev->recovery_offset < reshape_offset) {
7668 /* We need to check old and new layout */
7669 if (!only_parity(rdev->raid_disk,
7672 conf->max_degraded))
7675 if (!only_parity(rdev->raid_disk,
7677 conf->previous_raid_disks,
7678 conf->max_degraded))
7680 dirty_parity_disks++;
7684 * 0 for a fully functional array, 1 or 2 for a degraded array.
7686 mddev->degraded = raid5_calc_degraded(conf);
7688 if (has_failed(conf)) {
7689 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7690 mdname(mddev), mddev->degraded, conf->raid_disks);
7694 /* device size must be a multiple of chunk size */
7695 mddev->dev_sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7696 mddev->resync_max_sectors = mddev->dev_sectors;
7698 if (mddev->degraded > dirty_parity_disks &&
7699 mddev->recovery_cp != MaxSector) {
7700 if (test_bit(MD_HAS_PPL, &mddev->flags))
7701 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7703 else if (mddev->ok_start_degraded)
7704 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7707 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7713 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7714 mdname(mddev), conf->level,
7715 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7718 print_raid5_conf(conf);
7720 if (conf->reshape_progress != MaxSector) {
7721 conf->reshape_safe = conf->reshape_progress;
7722 atomic_set(&conf->reshape_stripes, 0);
7723 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7724 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7725 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7726 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7727 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7729 if (!mddev->sync_thread)
7733 /* Ok, everything is just fine now */
7734 if (mddev->to_remove == &raid5_attrs_group)
7735 mddev->to_remove = NULL;
7736 else if (mddev->kobj.sd &&
7737 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7738 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7740 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7744 /* read-ahead size must cover two whole stripes, which
7745 * is 2 * (datadisks) * chunksize where 'n' is the
7746 * number of raid devices
7748 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7749 int stripe = data_disks *
7750 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7752 chunk_size = mddev->chunk_sectors << 9;
7753 blk_queue_io_min(mddev->queue, chunk_size);
7754 raid5_set_io_opt(conf);
7755 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7757 * We can only discard a whole stripe. It doesn't make sense to
7758 * discard data disk but write parity disk
7760 stripe = stripe * PAGE_SIZE;
7761 /* Round up to power of 2, as discard handling
7762 * currently assumes that */
7763 while ((stripe-1) & stripe)
7764 stripe = (stripe | (stripe-1)) + 1;
7765 mddev->queue->limits.discard_alignment = stripe;
7766 mddev->queue->limits.discard_granularity = stripe;
7768 blk_queue_max_write_same_sectors(mddev->queue, 0);
7769 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7771 rdev_for_each(rdev, mddev) {
7772 disk_stack_limits(mddev->gendisk, rdev->bdev,
7773 rdev->data_offset << 9);
7774 disk_stack_limits(mddev->gendisk, rdev->bdev,
7775 rdev->new_data_offset << 9);
7779 * zeroing is required, otherwise data
7780 * could be lost. Consider a scenario: discard a stripe
7781 * (the stripe could be inconsistent if
7782 * discard_zeroes_data is 0); write one disk of the
7783 * stripe (the stripe could be inconsistent again
7784 * depending on which disks are used to calculate
7785 * parity); the disk is broken; The stripe data of this
7788 * We only allow DISCARD if the sysadmin has confirmed that
7789 * only safe devices are in use by setting a module parameter.
7790 * A better idea might be to turn DISCARD into WRITE_ZEROES
7791 * requests, as that is required to be safe.
7793 if (devices_handle_discard_safely &&
7794 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7795 mddev->queue->limits.discard_granularity >= stripe)
7796 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7799 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7802 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7805 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7810 md_unregister_thread(&mddev->thread);
7811 print_raid5_conf(conf);
7813 mddev->private = NULL;
7814 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7817 acct_bioset_exit(mddev);
7821 static void raid5_free(struct mddev *mddev, void *priv)
7823 struct r5conf *conf = priv;
7826 acct_bioset_exit(mddev);
7827 mddev->to_remove = &raid5_attrs_group;
7830 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7832 struct r5conf *conf = mddev->private;
7835 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7836 conf->chunk_sectors / 2, mddev->layout);
7837 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7839 for (i = 0; i < conf->raid_disks; i++) {
7840 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7841 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7844 seq_printf (seq, "]");
7847 static void print_raid5_conf (struct r5conf *conf)
7850 struct disk_info *tmp;
7852 pr_debug("RAID conf printout:\n");
7854 pr_debug("(conf==NULL)\n");
7857 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7859 conf->raid_disks - conf->mddev->degraded);
7861 for (i = 0; i < conf->raid_disks; i++) {
7862 char b[BDEVNAME_SIZE];
7863 tmp = conf->disks + i;
7865 pr_debug(" disk %d, o:%d, dev:%s\n",
7866 i, !test_bit(Faulty, &tmp->rdev->flags),
7867 bdevname(tmp->rdev->bdev, b));
7871 static int raid5_spare_active(struct mddev *mddev)
7874 struct r5conf *conf = mddev->private;
7875 struct disk_info *tmp;
7877 unsigned long flags;
7879 for (i = 0; i < conf->raid_disks; i++) {
7880 tmp = conf->disks + i;
7881 if (tmp->replacement
7882 && tmp->replacement->recovery_offset == MaxSector
7883 && !test_bit(Faulty, &tmp->replacement->flags)
7884 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7885 /* Replacement has just become active. */
7887 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7890 /* Replaced device not technically faulty,
7891 * but we need to be sure it gets removed
7892 * and never re-added.
7894 set_bit(Faulty, &tmp->rdev->flags);
7895 sysfs_notify_dirent_safe(
7896 tmp->rdev->sysfs_state);
7898 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7899 } else if (tmp->rdev
7900 && tmp->rdev->recovery_offset == MaxSector
7901 && !test_bit(Faulty, &tmp->rdev->flags)
7902 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7904 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7907 spin_lock_irqsave(&conf->device_lock, flags);
7908 mddev->degraded = raid5_calc_degraded(conf);
7909 spin_unlock_irqrestore(&conf->device_lock, flags);
7910 print_raid5_conf(conf);
7914 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7916 struct r5conf *conf = mddev->private;
7918 int number = rdev->raid_disk;
7919 struct md_rdev **rdevp;
7920 struct disk_info *p = conf->disks + number;
7922 print_raid5_conf(conf);
7923 if (test_bit(Journal, &rdev->flags) && conf->log) {
7925 * we can't wait pending write here, as this is called in
7926 * raid5d, wait will deadlock.
7927 * neilb: there is no locking about new writes here,
7928 * so this cannot be safe.
7930 if (atomic_read(&conf->active_stripes) ||
7931 atomic_read(&conf->r5c_cached_full_stripes) ||
7932 atomic_read(&conf->r5c_cached_partial_stripes)) {
7938 if (rdev == p->rdev)
7940 else if (rdev == p->replacement)
7941 rdevp = &p->replacement;
7945 if (number >= conf->raid_disks &&
7946 conf->reshape_progress == MaxSector)
7947 clear_bit(In_sync, &rdev->flags);
7949 if (test_bit(In_sync, &rdev->flags) ||
7950 atomic_read(&rdev->nr_pending)) {
7954 /* Only remove non-faulty devices if recovery
7957 if (!test_bit(Faulty, &rdev->flags) &&
7958 mddev->recovery_disabled != conf->recovery_disabled &&
7959 !has_failed(conf) &&
7960 (!p->replacement || p->replacement == rdev) &&
7961 number < conf->raid_disks) {
7966 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7968 if (atomic_read(&rdev->nr_pending)) {
7969 /* lost the race, try later */
7975 err = log_modify(conf, rdev, false);
7979 if (p->replacement) {
7980 /* We must have just cleared 'rdev' */
7981 p->rdev = p->replacement;
7982 clear_bit(Replacement, &p->replacement->flags);
7983 smp_mb(); /* Make sure other CPUs may see both as identical
7984 * but will never see neither - if they are careful
7986 p->replacement = NULL;
7989 err = log_modify(conf, p->rdev, true);
7992 clear_bit(WantReplacement, &rdev->flags);
7995 print_raid5_conf(conf);
7999 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
8001 struct r5conf *conf = mddev->private;
8002 int ret, err = -EEXIST;
8004 struct disk_info *p;
8006 int last = conf->raid_disks - 1;
8008 if (test_bit(Journal, &rdev->flags)) {
8012 rdev->raid_disk = 0;
8014 * The array is in readonly mode if journal is missing, so no
8015 * write requests running. We should be safe
8017 ret = log_init(conf, rdev, false);
8021 ret = r5l_start(conf->log);
8027 if (mddev->recovery_disabled == conf->recovery_disabled)
8030 if (rdev->saved_raid_disk < 0 && has_failed(conf))
8031 /* no point adding a device */
8034 if (rdev->raid_disk >= 0)
8035 first = last = rdev->raid_disk;
8038 * find the disk ... but prefer rdev->saved_raid_disk
8041 if (rdev->saved_raid_disk >= 0 &&
8042 rdev->saved_raid_disk >= first &&
8043 rdev->saved_raid_disk <= last &&
8044 conf->disks[rdev->saved_raid_disk].rdev == NULL)
8045 first = rdev->saved_raid_disk;
8047 for (disk = first; disk <= last; disk++) {
8048 p = conf->disks + disk;
8049 if (p->rdev == NULL) {
8050 clear_bit(In_sync, &rdev->flags);
8051 rdev->raid_disk = disk;
8052 if (rdev->saved_raid_disk != disk)
8054 rcu_assign_pointer(p->rdev, rdev);
8056 err = log_modify(conf, rdev, true);
8061 for (disk = first; disk <= last; disk++) {
8062 p = conf->disks + disk;
8063 if (test_bit(WantReplacement, &p->rdev->flags) &&
8064 p->replacement == NULL) {
8065 clear_bit(In_sync, &rdev->flags);
8066 set_bit(Replacement, &rdev->flags);
8067 rdev->raid_disk = disk;
8070 rcu_assign_pointer(p->replacement, rdev);
8075 print_raid5_conf(conf);
8079 static int raid5_resize(struct mddev *mddev, sector_t sectors)
8081 /* no resync is happening, and there is enough space
8082 * on all devices, so we can resize.
8083 * We need to make sure resync covers any new space.
8084 * If the array is shrinking we should possibly wait until
8085 * any io in the removed space completes, but it hardly seems
8089 struct r5conf *conf = mddev->private;
8091 if (raid5_has_log(conf) || raid5_has_ppl(conf))
8093 sectors &= ~((sector_t)conf->chunk_sectors - 1);
8094 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
8095 if (mddev->external_size &&
8096 mddev->array_sectors > newsize)
8098 if (mddev->bitmap) {
8099 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
8103 md_set_array_sectors(mddev, newsize);
8104 if (sectors > mddev->dev_sectors &&
8105 mddev->recovery_cp > mddev->dev_sectors) {
8106 mddev->recovery_cp = mddev->dev_sectors;
8107 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8109 mddev->dev_sectors = sectors;
8110 mddev->resync_max_sectors = sectors;
8114 static int check_stripe_cache(struct mddev *mddev)
8116 /* Can only proceed if there are plenty of stripe_heads.
8117 * We need a minimum of one full stripe,, and for sensible progress
8118 * it is best to have about 4 times that.
8119 * If we require 4 times, then the default 256 4K stripe_heads will
8120 * allow for chunk sizes up to 256K, which is probably OK.
8121 * If the chunk size is greater, user-space should request more
8122 * stripe_heads first.
8124 struct r5conf *conf = mddev->private;
8125 if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8126 > conf->min_nr_stripes ||
8127 ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8128 > conf->min_nr_stripes) {
8129 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
8131 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
8132 / RAID5_STRIPE_SIZE(conf))*4);
8138 static int check_reshape(struct mddev *mddev)
8140 struct r5conf *conf = mddev->private;
8142 if (raid5_has_log(conf) || raid5_has_ppl(conf))
8144 if (mddev->delta_disks == 0 &&
8145 mddev->new_layout == mddev->layout &&
8146 mddev->new_chunk_sectors == mddev->chunk_sectors)
8147 return 0; /* nothing to do */
8148 if (has_failed(conf))
8150 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
8151 /* We might be able to shrink, but the devices must
8152 * be made bigger first.
8153 * For raid6, 4 is the minimum size.
8154 * Otherwise 2 is the minimum
8157 if (mddev->level == 6)
8159 if (mddev->raid_disks + mddev->delta_disks < min)
8163 if (!check_stripe_cache(mddev))
8166 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
8167 mddev->delta_disks > 0)
8168 if (resize_chunks(conf,
8169 conf->previous_raid_disks
8170 + max(0, mddev->delta_disks),
8171 max(mddev->new_chunk_sectors,
8172 mddev->chunk_sectors)
8176 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
8177 return 0; /* never bother to shrink */
8178 return resize_stripes(conf, (conf->previous_raid_disks
8179 + mddev->delta_disks));
8182 static int raid5_start_reshape(struct mddev *mddev)
8184 struct r5conf *conf = mddev->private;
8185 struct md_rdev *rdev;
8187 unsigned long flags;
8189 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
8192 if (!check_stripe_cache(mddev))
8195 if (has_failed(conf))
8198 rdev_for_each(rdev, mddev) {
8199 if (!test_bit(In_sync, &rdev->flags)
8200 && !test_bit(Faulty, &rdev->flags))
8204 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
8205 /* Not enough devices even to make a degraded array
8210 /* Refuse to reduce size of the array. Any reductions in
8211 * array size must be through explicit setting of array_size
8214 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
8215 < mddev->array_sectors) {
8216 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8221 atomic_set(&conf->reshape_stripes, 0);
8222 spin_lock_irq(&conf->device_lock);
8223 write_seqcount_begin(&conf->gen_lock);
8224 conf->previous_raid_disks = conf->raid_disks;
8225 conf->raid_disks += mddev->delta_disks;
8226 conf->prev_chunk_sectors = conf->chunk_sectors;
8227 conf->chunk_sectors = mddev->new_chunk_sectors;
8228 conf->prev_algo = conf->algorithm;
8229 conf->algorithm = mddev->new_layout;
8231 /* Code that selects data_offset needs to see the generation update
8232 * if reshape_progress has been set - so a memory barrier needed.
8235 if (mddev->reshape_backwards)
8236 conf->reshape_progress = raid5_size(mddev, 0, 0);
8238 conf->reshape_progress = 0;
8239 conf->reshape_safe = conf->reshape_progress;
8240 write_seqcount_end(&conf->gen_lock);
8241 spin_unlock_irq(&conf->device_lock);
8243 /* Now make sure any requests that proceeded on the assumption
8244 * the reshape wasn't running - like Discard or Read - have
8247 mddev_suspend(mddev);
8248 mddev_resume(mddev);
8250 /* Add some new drives, as many as will fit.
8251 * We know there are enough to make the newly sized array work.
8252 * Don't add devices if we are reducing the number of
8253 * devices in the array. This is because it is not possible
8254 * to correctly record the "partially reconstructed" state of
8255 * such devices during the reshape and confusion could result.
8257 if (mddev->delta_disks >= 0) {
8258 rdev_for_each(rdev, mddev)
8259 if (rdev->raid_disk < 0 &&
8260 !test_bit(Faulty, &rdev->flags)) {
8261 if (raid5_add_disk(mddev, rdev) == 0) {
8263 >= conf->previous_raid_disks)
8264 set_bit(In_sync, &rdev->flags);
8266 rdev->recovery_offset = 0;
8268 /* Failure here is OK */
8269 sysfs_link_rdev(mddev, rdev);
8271 } else if (rdev->raid_disk >= conf->previous_raid_disks
8272 && !test_bit(Faulty, &rdev->flags)) {
8273 /* This is a spare that was manually added */
8274 set_bit(In_sync, &rdev->flags);
8277 /* When a reshape changes the number of devices,
8278 * ->degraded is measured against the larger of the
8279 * pre and post number of devices.
8281 spin_lock_irqsave(&conf->device_lock, flags);
8282 mddev->degraded = raid5_calc_degraded(conf);
8283 spin_unlock_irqrestore(&conf->device_lock, flags);
8285 mddev->raid_disks = conf->raid_disks;
8286 mddev->reshape_position = conf->reshape_progress;
8287 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8289 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8290 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8291 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8292 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8293 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
8294 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
8296 if (!mddev->sync_thread) {
8297 mddev->recovery = 0;
8298 spin_lock_irq(&conf->device_lock);
8299 write_seqcount_begin(&conf->gen_lock);
8300 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
8301 mddev->new_chunk_sectors =
8302 conf->chunk_sectors = conf->prev_chunk_sectors;
8303 mddev->new_layout = conf->algorithm = conf->prev_algo;
8304 rdev_for_each(rdev, mddev)
8305 rdev->new_data_offset = rdev->data_offset;
8307 conf->generation --;
8308 conf->reshape_progress = MaxSector;
8309 mddev->reshape_position = MaxSector;
8310 write_seqcount_end(&conf->gen_lock);
8311 spin_unlock_irq(&conf->device_lock);
8314 conf->reshape_checkpoint = jiffies;
8315 md_wakeup_thread(mddev->sync_thread);
8316 md_new_event(mddev);
8320 /* This is called from the reshape thread and should make any
8321 * changes needed in 'conf'
8323 static void end_reshape(struct r5conf *conf)
8326 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8327 struct md_rdev *rdev;
8329 spin_lock_irq(&conf->device_lock);
8330 conf->previous_raid_disks = conf->raid_disks;
8331 md_finish_reshape(conf->mddev);
8333 conf->reshape_progress = MaxSector;
8334 conf->mddev->reshape_position = MaxSector;
8335 rdev_for_each(rdev, conf->mddev)
8336 if (rdev->raid_disk >= 0 &&
8337 !test_bit(Journal, &rdev->flags) &&
8338 !test_bit(In_sync, &rdev->flags))
8339 rdev->recovery_offset = MaxSector;
8340 spin_unlock_irq(&conf->device_lock);
8341 wake_up(&conf->wait_for_overlap);
8343 if (conf->mddev->queue)
8344 raid5_set_io_opt(conf);
8348 /* This is called from the raid5d thread with mddev_lock held.
8349 * It makes config changes to the device.
8351 static void raid5_finish_reshape(struct mddev *mddev)
8353 struct r5conf *conf = mddev->private;
8355 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8357 if (mddev->delta_disks <= 0) {
8359 spin_lock_irq(&conf->device_lock);
8360 mddev->degraded = raid5_calc_degraded(conf);
8361 spin_unlock_irq(&conf->device_lock);
8362 for (d = conf->raid_disks ;
8363 d < conf->raid_disks - mddev->delta_disks;
8365 struct md_rdev *rdev = conf->disks[d].rdev;
8367 clear_bit(In_sync, &rdev->flags);
8368 rdev = conf->disks[d].replacement;
8370 clear_bit(In_sync, &rdev->flags);
8373 mddev->layout = conf->algorithm;
8374 mddev->chunk_sectors = conf->chunk_sectors;
8375 mddev->reshape_position = MaxSector;
8376 mddev->delta_disks = 0;
8377 mddev->reshape_backwards = 0;
8381 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8383 struct r5conf *conf = mddev->private;
8386 /* stop all writes */
8387 lock_all_device_hash_locks_irq(conf);
8388 /* '2' tells resync/reshape to pause so that all
8389 * active stripes can drain
8391 r5c_flush_cache(conf, INT_MAX);
8392 /* need a memory barrier to make sure read_one_chunk() sees
8393 * quiesce started and reverts to slow (locked) path.
8395 smp_store_release(&conf->quiesce, 2);
8396 wait_event_cmd(conf->wait_for_quiescent,
8397 atomic_read(&conf->active_stripes) == 0 &&
8398 atomic_read(&conf->active_aligned_reads) == 0,
8399 unlock_all_device_hash_locks_irq(conf),
8400 lock_all_device_hash_locks_irq(conf));
8402 unlock_all_device_hash_locks_irq(conf);
8403 /* allow reshape to continue */
8404 wake_up(&conf->wait_for_overlap);
8406 /* re-enable writes */
8407 lock_all_device_hash_locks_irq(conf);
8409 wake_up(&conf->wait_for_quiescent);
8410 wake_up(&conf->wait_for_overlap);
8411 unlock_all_device_hash_locks_irq(conf);
8413 log_quiesce(conf, quiesce);
8416 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8418 struct r0conf *raid0_conf = mddev->private;
8421 /* for raid0 takeover only one zone is supported */
8422 if (raid0_conf->nr_strip_zones > 1) {
8423 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8425 return ERR_PTR(-EINVAL);
8428 sectors = raid0_conf->strip_zone[0].zone_end;
8429 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8430 mddev->dev_sectors = sectors;
8431 mddev->new_level = level;
8432 mddev->new_layout = ALGORITHM_PARITY_N;
8433 mddev->new_chunk_sectors = mddev->chunk_sectors;
8434 mddev->raid_disks += 1;
8435 mddev->delta_disks = 1;
8436 /* make sure it will be not marked as dirty */
8437 mddev->recovery_cp = MaxSector;
8439 return setup_conf(mddev);
8442 static void *raid5_takeover_raid1(struct mddev *mddev)
8447 if (mddev->raid_disks != 2 ||
8448 mddev->degraded > 1)
8449 return ERR_PTR(-EINVAL);
8451 /* Should check if there are write-behind devices? */
8453 chunksect = 64*2; /* 64K by default */
8455 /* The array must be an exact multiple of chunksize */
8456 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8459 if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8460 /* array size does not allow a suitable chunk size */
8461 return ERR_PTR(-EINVAL);
8463 mddev->new_level = 5;
8464 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8465 mddev->new_chunk_sectors = chunksect;
8467 ret = setup_conf(mddev);
8469 mddev_clear_unsupported_flags(mddev,
8470 UNSUPPORTED_MDDEV_FLAGS);
8474 static void *raid5_takeover_raid6(struct mddev *mddev)
8478 switch (mddev->layout) {
8479 case ALGORITHM_LEFT_ASYMMETRIC_6:
8480 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8482 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8483 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8485 case ALGORITHM_LEFT_SYMMETRIC_6:
8486 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8488 case ALGORITHM_RIGHT_SYMMETRIC_6:
8489 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8491 case ALGORITHM_PARITY_0_6:
8492 new_layout = ALGORITHM_PARITY_0;
8494 case ALGORITHM_PARITY_N:
8495 new_layout = ALGORITHM_PARITY_N;
8498 return ERR_PTR(-EINVAL);
8500 mddev->new_level = 5;
8501 mddev->new_layout = new_layout;
8502 mddev->delta_disks = -1;
8503 mddev->raid_disks -= 1;
8504 return setup_conf(mddev);
8507 static int raid5_check_reshape(struct mddev *mddev)
8509 /* For a 2-drive array, the layout and chunk size can be changed
8510 * immediately as not restriping is needed.
8511 * For larger arrays we record the new value - after validation
8512 * to be used by a reshape pass.
8514 struct r5conf *conf = mddev->private;
8515 int new_chunk = mddev->new_chunk_sectors;
8517 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8519 if (new_chunk > 0) {
8520 if (!is_power_of_2(new_chunk))
8522 if (new_chunk < (PAGE_SIZE>>9))
8524 if (mddev->array_sectors & (new_chunk-1))
8525 /* not factor of array size */
8529 /* They look valid */
8531 if (mddev->raid_disks == 2) {
8532 /* can make the change immediately */
8533 if (mddev->new_layout >= 0) {
8534 conf->algorithm = mddev->new_layout;
8535 mddev->layout = mddev->new_layout;
8537 if (new_chunk > 0) {
8538 conf->chunk_sectors = new_chunk ;
8539 mddev->chunk_sectors = new_chunk;
8541 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8542 md_wakeup_thread(mddev->thread);
8544 return check_reshape(mddev);
8547 static int raid6_check_reshape(struct mddev *mddev)
8549 int new_chunk = mddev->new_chunk_sectors;
8551 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8553 if (new_chunk > 0) {
8554 if (!is_power_of_2(new_chunk))
8556 if (new_chunk < (PAGE_SIZE >> 9))
8558 if (mddev->array_sectors & (new_chunk-1))
8559 /* not factor of array size */
8563 /* They look valid */
8564 return check_reshape(mddev);
8567 static void *raid5_takeover(struct mddev *mddev)
8569 /* raid5 can take over:
8570 * raid0 - if there is only one strip zone - make it a raid4 layout
8571 * raid1 - if there are two drives. We need to know the chunk size
8572 * raid4 - trivial - just use a raid4 layout.
8573 * raid6 - Providing it is a *_6 layout
8575 if (mddev->level == 0)
8576 return raid45_takeover_raid0(mddev, 5);
8577 if (mddev->level == 1)
8578 return raid5_takeover_raid1(mddev);
8579 if (mddev->level == 4) {
8580 mddev->new_layout = ALGORITHM_PARITY_N;
8581 mddev->new_level = 5;
8582 return setup_conf(mddev);
8584 if (mddev->level == 6)
8585 return raid5_takeover_raid6(mddev);
8587 return ERR_PTR(-EINVAL);
8590 static void *raid4_takeover(struct mddev *mddev)
8592 /* raid4 can take over:
8593 * raid0 - if there is only one strip zone
8594 * raid5 - if layout is right
8596 if (mddev->level == 0)
8597 return raid45_takeover_raid0(mddev, 4);
8598 if (mddev->level == 5 &&
8599 mddev->layout == ALGORITHM_PARITY_N) {
8600 mddev->new_layout = 0;
8601 mddev->new_level = 4;
8602 return setup_conf(mddev);
8604 return ERR_PTR(-EINVAL);
8607 static struct md_personality raid5_personality;
8609 static void *raid6_takeover(struct mddev *mddev)
8611 /* Currently can only take over a raid5. We map the
8612 * personality to an equivalent raid6 personality
8613 * with the Q block at the end.
8617 if (mddev->pers != &raid5_personality)
8618 return ERR_PTR(-EINVAL);
8619 if (mddev->degraded > 1)
8620 return ERR_PTR(-EINVAL);
8621 if (mddev->raid_disks > 253)
8622 return ERR_PTR(-EINVAL);
8623 if (mddev->raid_disks < 3)
8624 return ERR_PTR(-EINVAL);
8626 switch (mddev->layout) {
8627 case ALGORITHM_LEFT_ASYMMETRIC:
8628 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8630 case ALGORITHM_RIGHT_ASYMMETRIC:
8631 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8633 case ALGORITHM_LEFT_SYMMETRIC:
8634 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8636 case ALGORITHM_RIGHT_SYMMETRIC:
8637 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8639 case ALGORITHM_PARITY_0:
8640 new_layout = ALGORITHM_PARITY_0_6;
8642 case ALGORITHM_PARITY_N:
8643 new_layout = ALGORITHM_PARITY_N;
8646 return ERR_PTR(-EINVAL);
8648 mddev->new_level = 6;
8649 mddev->new_layout = new_layout;
8650 mddev->delta_disks = 1;
8651 mddev->raid_disks += 1;
8652 return setup_conf(mddev);
8655 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8657 struct r5conf *conf;
8660 err = mddev_lock(mddev);
8663 conf = mddev->private;
8665 mddev_unlock(mddev);
8669 if (strncmp(buf, "ppl", 3) == 0) {
8670 /* ppl only works with RAID 5 */
8671 if (!raid5_has_ppl(conf) && conf->level == 5) {
8672 err = log_init(conf, NULL, true);
8674 err = resize_stripes(conf, conf->pool_size);
8680 } else if (strncmp(buf, "resync", 6) == 0) {
8681 if (raid5_has_ppl(conf)) {
8682 mddev_suspend(mddev);
8684 mddev_resume(mddev);
8685 err = resize_stripes(conf, conf->pool_size);
8686 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8687 r5l_log_disk_error(conf)) {
8688 bool journal_dev_exists = false;
8689 struct md_rdev *rdev;
8691 rdev_for_each(rdev, mddev)
8692 if (test_bit(Journal, &rdev->flags)) {
8693 journal_dev_exists = true;
8697 if (!journal_dev_exists) {
8698 mddev_suspend(mddev);
8699 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8700 mddev_resume(mddev);
8701 } else /* need remove journal device first */
8710 md_update_sb(mddev, 1);
8712 mddev_unlock(mddev);
8717 static int raid5_start(struct mddev *mddev)
8719 struct r5conf *conf = mddev->private;
8721 return r5l_start(conf->log);
8724 static struct md_personality raid6_personality =
8728 .owner = THIS_MODULE,
8729 .make_request = raid5_make_request,
8731 .start = raid5_start,
8733 .status = raid5_status,
8734 .error_handler = raid5_error,
8735 .hot_add_disk = raid5_add_disk,
8736 .hot_remove_disk= raid5_remove_disk,
8737 .spare_active = raid5_spare_active,
8738 .sync_request = raid5_sync_request,
8739 .resize = raid5_resize,
8741 .check_reshape = raid6_check_reshape,
8742 .start_reshape = raid5_start_reshape,
8743 .finish_reshape = raid5_finish_reshape,
8744 .quiesce = raid5_quiesce,
8745 .takeover = raid6_takeover,
8746 .change_consistency_policy = raid5_change_consistency_policy,
8748 static struct md_personality raid5_personality =
8752 .owner = THIS_MODULE,
8753 .make_request = raid5_make_request,
8755 .start = raid5_start,
8757 .status = raid5_status,
8758 .error_handler = raid5_error,
8759 .hot_add_disk = raid5_add_disk,
8760 .hot_remove_disk= raid5_remove_disk,
8761 .spare_active = raid5_spare_active,
8762 .sync_request = raid5_sync_request,
8763 .resize = raid5_resize,
8765 .check_reshape = raid5_check_reshape,
8766 .start_reshape = raid5_start_reshape,
8767 .finish_reshape = raid5_finish_reshape,
8768 .quiesce = raid5_quiesce,
8769 .takeover = raid5_takeover,
8770 .change_consistency_policy = raid5_change_consistency_policy,
8773 static struct md_personality raid4_personality =
8777 .owner = THIS_MODULE,
8778 .make_request = raid5_make_request,
8780 .start = raid5_start,
8782 .status = raid5_status,
8783 .error_handler = raid5_error,
8784 .hot_add_disk = raid5_add_disk,
8785 .hot_remove_disk= raid5_remove_disk,
8786 .spare_active = raid5_spare_active,
8787 .sync_request = raid5_sync_request,
8788 .resize = raid5_resize,
8790 .check_reshape = raid5_check_reshape,
8791 .start_reshape = raid5_start_reshape,
8792 .finish_reshape = raid5_finish_reshape,
8793 .quiesce = raid5_quiesce,
8794 .takeover = raid4_takeover,
8795 .change_consistency_policy = raid5_change_consistency_policy,
8798 static int __init raid5_init(void)
8802 raid5_wq = alloc_workqueue("raid5wq",
8803 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8807 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8809 raid456_cpu_up_prepare,
8812 destroy_workqueue(raid5_wq);
8815 register_md_personality(&raid6_personality);
8816 register_md_personality(&raid5_personality);
8817 register_md_personality(&raid4_personality);
8821 static void raid5_exit(void)
8823 unregister_md_personality(&raid6_personality);
8824 unregister_md_personality(&raid5_personality);
8825 unregister_md_personality(&raid4_personality);
8826 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8827 destroy_workqueue(raid5_wq);
8830 module_init(raid5_init);
8831 module_exit(raid5_exit);
8832 MODULE_LICENSE("GPL");
8833 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8834 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8835 MODULE_ALIAS("md-raid5");
8836 MODULE_ALIAS("md-raid4");
8837 MODULE_ALIAS("md-level-5");
8838 MODULE_ALIAS("md-level-4");
8839 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8840 MODULE_ALIAS("md-raid6");
8841 MODULE_ALIAS("md-level-6");
8843 /* This used to be two separate modules, they were: */
8844 MODULE_ALIAS("raid5");
8845 MODULE_ALIAS("raid6");