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/kthread.h>
40 #include <linux/raid/pq.h>
41 #include <linux/async_tx.h>
42 #include <linux/module.h>
43 #include <linux/async.h>
44 #include <linux/seq_file.h>
45 #include <linux/cpu.h>
46 #include <linux/slab.h>
47 #include <linux/ratelimit.h>
48 #include <linux/nodemask.h>
50 #include <trace/events/block.h>
51 #include <linux/list_sort.h>
56 #include "md-bitmap.h"
57 #include "raid5-log.h"
59 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
61 #define cpu_to_group(cpu) cpu_to_node(cpu)
62 #define ANY_GROUP NUMA_NO_NODE
64 static bool devices_handle_discard_safely
= false;
65 module_param(devices_handle_discard_safely
, bool, 0644);
66 MODULE_PARM_DESC(devices_handle_discard_safely
,
67 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
68 static struct workqueue_struct
*raid5_wq
;
70 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
72 int hash
= (sect
>> RAID5_STRIPE_SHIFT(conf
)) & HASH_MASK
;
73 return &conf
->stripe_hashtbl
[hash
];
76 static inline int stripe_hash_locks_hash(struct r5conf
*conf
, sector_t sect
)
78 return (sect
>> RAID5_STRIPE_SHIFT(conf
)) & STRIPE_HASH_LOCKS_MASK
;
81 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
83 spin_lock_irq(conf
->hash_locks
+ hash
);
84 spin_lock(&conf
->device_lock
);
87 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
89 spin_unlock(&conf
->device_lock
);
90 spin_unlock_irq(conf
->hash_locks
+ hash
);
93 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
96 spin_lock_irq(conf
->hash_locks
);
97 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
98 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
99 spin_lock(&conf
->device_lock
);
102 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
105 spin_unlock(&conf
->device_lock
);
106 for (i
= NR_STRIPE_HASH_LOCKS
- 1; i
; i
--)
107 spin_unlock(conf
->hash_locks
+ i
);
108 spin_unlock_irq(conf
->hash_locks
);
111 /* Find first data disk in a raid6 stripe */
112 static inline int raid6_d0(struct stripe_head
*sh
)
115 /* ddf always start from first device */
117 /* md starts just after Q block */
118 if (sh
->qd_idx
== sh
->disks
- 1)
121 return sh
->qd_idx
+ 1;
123 static inline int raid6_next_disk(int disk
, int raid_disks
)
126 return (disk
< raid_disks
) ? disk
: 0;
129 /* When walking through the disks in a raid5, starting at raid6_d0,
130 * We need to map each disk to a 'slot', where the data disks are slot
131 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
132 * is raid_disks-1. This help does that mapping.
134 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
135 int *count
, int syndrome_disks
)
141 if (idx
== sh
->pd_idx
)
142 return syndrome_disks
;
143 if (idx
== sh
->qd_idx
)
144 return syndrome_disks
+ 1;
150 static void print_raid5_conf (struct r5conf
*conf
);
152 static int stripe_operations_active(struct stripe_head
*sh
)
154 return sh
->check_state
|| sh
->reconstruct_state
||
155 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
156 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
159 static bool stripe_is_lowprio(struct stripe_head
*sh
)
161 return (test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) ||
162 test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
)) &&
163 !test_bit(STRIPE_R5C_CACHING
, &sh
->state
);
166 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
168 struct r5conf
*conf
= sh
->raid_conf
;
169 struct r5worker_group
*group
;
171 int i
, cpu
= sh
->cpu
;
173 if (!cpu_online(cpu
)) {
174 cpu
= cpumask_any(cpu_online_mask
);
178 if (list_empty(&sh
->lru
)) {
179 struct r5worker_group
*group
;
180 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
181 if (stripe_is_lowprio(sh
))
182 list_add_tail(&sh
->lru
, &group
->loprio_list
);
184 list_add_tail(&sh
->lru
, &group
->handle_list
);
185 group
->stripes_cnt
++;
189 if (conf
->worker_cnt_per_group
== 0) {
190 md_wakeup_thread(conf
->mddev
->thread
);
194 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
196 group
->workers
[0].working
= true;
197 /* at least one worker should run to avoid race */
198 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
200 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
201 /* wakeup more workers */
202 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
203 if (group
->workers
[i
].working
== false) {
204 group
->workers
[i
].working
= true;
205 queue_work_on(sh
->cpu
, raid5_wq
,
206 &group
->workers
[i
].work
);
212 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
213 struct list_head
*temp_inactive_list
)
216 int injournal
= 0; /* number of date pages with R5_InJournal */
218 BUG_ON(!list_empty(&sh
->lru
));
219 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
221 if (r5c_is_writeback(conf
->log
))
222 for (i
= sh
->disks
; i
--; )
223 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
226 * In the following cases, the stripe cannot be released to cached
227 * lists. Therefore, we make the stripe write out and set
229 * 1. when quiesce in r5c write back;
230 * 2. when resync is requested fot the stripe.
232 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) ||
233 (conf
->quiesce
&& r5c_is_writeback(conf
->log
) &&
234 !test_bit(STRIPE_HANDLE
, &sh
->state
) && injournal
!= 0)) {
235 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
236 r5c_make_stripe_write_out(sh
);
237 set_bit(STRIPE_HANDLE
, &sh
->state
);
240 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
241 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
242 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
243 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
244 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
245 sh
->bm_seq
- conf
->seq_write
> 0)
246 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
248 clear_bit(STRIPE_DELAYED
, &sh
->state
);
249 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
250 if (conf
->worker_cnt_per_group
== 0) {
251 if (stripe_is_lowprio(sh
))
252 list_add_tail(&sh
->lru
,
255 list_add_tail(&sh
->lru
,
258 raid5_wakeup_stripe_thread(sh
);
262 md_wakeup_thread(conf
->mddev
->thread
);
264 BUG_ON(stripe_operations_active(sh
));
265 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
266 if (atomic_dec_return(&conf
->preread_active_stripes
)
268 md_wakeup_thread(conf
->mddev
->thread
);
269 atomic_dec(&conf
->active_stripes
);
270 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
271 if (!r5c_is_writeback(conf
->log
))
272 list_add_tail(&sh
->lru
, temp_inactive_list
);
274 WARN_ON(test_bit(R5_InJournal
, &sh
->dev
[sh
->pd_idx
].flags
));
276 list_add_tail(&sh
->lru
, temp_inactive_list
);
277 else if (injournal
== conf
->raid_disks
- conf
->max_degraded
) {
279 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
))
280 atomic_inc(&conf
->r5c_cached_full_stripes
);
281 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
))
282 atomic_dec(&conf
->r5c_cached_partial_stripes
);
283 list_add_tail(&sh
->lru
, &conf
->r5c_full_stripe_list
);
284 r5c_check_cached_full_stripe(conf
);
287 * STRIPE_R5C_PARTIAL_STRIPE is set in
288 * r5c_try_caching_write(). No need to
291 list_add_tail(&sh
->lru
, &conf
->r5c_partial_stripe_list
);
297 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
298 struct list_head
*temp_inactive_list
)
300 if (atomic_dec_and_test(&sh
->count
))
301 do_release_stripe(conf
, sh
, temp_inactive_list
);
305 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
307 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
308 * given time. Adding stripes only takes device lock, while deleting stripes
309 * only takes hash lock.
311 static void release_inactive_stripe_list(struct r5conf
*conf
,
312 struct list_head
*temp_inactive_list
,
316 bool do_wakeup
= false;
319 if (hash
== NR_STRIPE_HASH_LOCKS
) {
320 size
= NR_STRIPE_HASH_LOCKS
;
321 hash
= NR_STRIPE_HASH_LOCKS
- 1;
325 struct list_head
*list
= &temp_inactive_list
[size
- 1];
328 * We don't hold any lock here yet, raid5_get_active_stripe() might
329 * remove stripes from the list
331 if (!list_empty_careful(list
)) {
332 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
333 if (list_empty(conf
->inactive_list
+ hash
) &&
335 atomic_dec(&conf
->empty_inactive_list_nr
);
336 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
338 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
345 wake_up(&conf
->wait_for_stripe
);
346 if (atomic_read(&conf
->active_stripes
) == 0)
347 wake_up(&conf
->wait_for_quiescent
);
348 if (conf
->retry_read_aligned
)
349 md_wakeup_thread(conf
->mddev
->thread
);
353 /* should hold conf->device_lock already */
354 static int release_stripe_list(struct r5conf
*conf
,
355 struct list_head
*temp_inactive_list
)
357 struct stripe_head
*sh
, *t
;
359 struct llist_node
*head
;
361 head
= llist_del_all(&conf
->released_stripes
);
362 head
= llist_reverse_order(head
);
363 llist_for_each_entry_safe(sh
, t
, head
, release_list
) {
366 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
368 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
370 * Don't worry the bit is set here, because if the bit is set
371 * again, the count is always > 1. This is true for
372 * STRIPE_ON_UNPLUG_LIST bit too.
374 hash
= sh
->hash_lock_index
;
375 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
382 void raid5_release_stripe(struct stripe_head
*sh
)
384 struct r5conf
*conf
= sh
->raid_conf
;
386 struct list_head list
;
390 /* Avoid release_list until the last reference.
392 if (atomic_add_unless(&sh
->count
, -1, 1))
395 if (unlikely(!conf
->mddev
->thread
) ||
396 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
398 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
400 md_wakeup_thread(conf
->mddev
->thread
);
403 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
404 if (atomic_dec_and_lock_irqsave(&sh
->count
, &conf
->device_lock
, flags
)) {
405 INIT_LIST_HEAD(&list
);
406 hash
= sh
->hash_lock_index
;
407 do_release_stripe(conf
, sh
, &list
);
408 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
409 release_inactive_stripe_list(conf
, &list
, hash
);
413 static inline void remove_hash(struct stripe_head
*sh
)
415 pr_debug("remove_hash(), stripe %llu\n",
416 (unsigned long long)sh
->sector
);
418 hlist_del_init(&sh
->hash
);
421 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
423 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
425 pr_debug("insert_hash(), stripe %llu\n",
426 (unsigned long long)sh
->sector
);
428 hlist_add_head(&sh
->hash
, hp
);
431 /* find an idle stripe, make sure it is unhashed, and return it. */
432 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
434 struct stripe_head
*sh
= NULL
;
435 struct list_head
*first
;
437 if (list_empty(conf
->inactive_list
+ hash
))
439 first
= (conf
->inactive_list
+ hash
)->next
;
440 sh
= list_entry(first
, struct stripe_head
, lru
);
441 list_del_init(first
);
443 atomic_inc(&conf
->active_stripes
);
444 BUG_ON(hash
!= sh
->hash_lock_index
);
445 if (list_empty(conf
->inactive_list
+ hash
))
446 atomic_inc(&conf
->empty_inactive_list_nr
);
451 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
452 static void free_stripe_pages(struct stripe_head
*sh
)
457 /* Have not allocate page pool */
461 for (i
= 0; i
< sh
->nr_pages
; i
++) {
469 static int alloc_stripe_pages(struct stripe_head
*sh
, gfp_t gfp
)
474 for (i
= 0; i
< sh
->nr_pages
; i
++) {
475 /* The page have allocated. */
481 free_stripe_pages(sh
);
490 init_stripe_shared_pages(struct stripe_head
*sh
, struct r5conf
*conf
, int disks
)
497 /* Each of the sh->dev[i] need one conf->stripe_size */
498 cnt
= PAGE_SIZE
/ conf
->stripe_size
;
499 nr_pages
= (disks
+ cnt
- 1) / cnt
;
501 sh
->pages
= kcalloc(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
504 sh
->nr_pages
= nr_pages
;
505 sh
->stripes_per_page
= cnt
;
510 static void shrink_buffers(struct stripe_head
*sh
)
513 int num
= sh
->raid_conf
->pool_size
;
515 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
516 for (i
= 0; i
< num
; i
++) {
519 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
523 sh
->dev
[i
].page
= NULL
;
527 for (i
= 0; i
< num
; i
++)
528 sh
->dev
[i
].page
= NULL
;
529 free_stripe_pages(sh
); /* Free pages */
533 static int grow_buffers(struct stripe_head
*sh
, gfp_t gfp
)
536 int num
= sh
->raid_conf
->pool_size
;
538 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
539 for (i
= 0; i
< num
; i
++) {
542 if (!(page
= alloc_page(gfp
))) {
545 sh
->dev
[i
].page
= page
;
546 sh
->dev
[i
].orig_page
= page
;
547 sh
->dev
[i
].offset
= 0;
550 if (alloc_stripe_pages(sh
, gfp
))
553 for (i
= 0; i
< num
; i
++) {
554 sh
->dev
[i
].page
= raid5_get_dev_page(sh
, i
);
555 sh
->dev
[i
].orig_page
= sh
->dev
[i
].page
;
556 sh
->dev
[i
].offset
= raid5_get_page_offset(sh
, i
);
562 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
563 struct stripe_head
*sh
);
565 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
567 struct r5conf
*conf
= sh
->raid_conf
;
570 BUG_ON(atomic_read(&sh
->count
) != 0);
571 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
572 BUG_ON(stripe_operations_active(sh
));
573 BUG_ON(sh
->batch_head
);
575 pr_debug("init_stripe called, stripe %llu\n",
576 (unsigned long long)sector
);
578 seq
= read_seqcount_begin(&conf
->gen_lock
);
579 sh
->generation
= conf
->generation
- previous
;
580 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
582 stripe_set_idx(sector
, conf
, previous
, sh
);
585 for (i
= sh
->disks
; i
--; ) {
586 struct r5dev
*dev
= &sh
->dev
[i
];
588 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
589 test_bit(R5_LOCKED
, &dev
->flags
)) {
590 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
591 (unsigned long long)sh
->sector
, i
, dev
->toread
,
592 dev
->read
, dev
->towrite
, dev
->written
,
593 test_bit(R5_LOCKED
, &dev
->flags
));
597 dev
->sector
= raid5_compute_blocknr(sh
, i
, previous
);
599 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
601 sh
->overwrite_disks
= 0;
602 insert_hash(conf
, sh
);
603 sh
->cpu
= smp_processor_id();
604 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
607 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
610 struct stripe_head
*sh
;
612 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
613 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
614 if (sh
->sector
== sector
&& sh
->generation
== generation
)
616 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
621 * Need to check if array has failed when deciding whether to:
623 * - remove non-faulty devices
626 * This determination is simple when no reshape is happening.
627 * However if there is a reshape, we need to carefully check
628 * both the before and after sections.
629 * This is because some failed devices may only affect one
630 * of the two sections, and some non-in_sync devices may
631 * be insync in the section most affected by failed devices.
633 int raid5_calc_degraded(struct r5conf
*conf
)
635 int degraded
, degraded2
;
640 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
641 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
642 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
643 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
644 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
646 else if (test_bit(In_sync
, &rdev
->flags
))
649 /* not in-sync or faulty.
650 * If the reshape increases the number of devices,
651 * this is being recovered by the reshape, so
652 * this 'previous' section is not in_sync.
653 * If the number of devices is being reduced however,
654 * the device can only be part of the array if
655 * we are reverting a reshape, so this section will
658 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
662 if (conf
->raid_disks
== conf
->previous_raid_disks
)
666 for (i
= 0; i
< conf
->raid_disks
; i
++) {
667 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
668 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
669 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
670 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
672 else if (test_bit(In_sync
, &rdev
->flags
))
675 /* not in-sync or faulty.
676 * If reshape increases the number of devices, this
677 * section has already been recovered, else it
678 * almost certainly hasn't.
680 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
684 if (degraded2
> degraded
)
689 static int has_failed(struct r5conf
*conf
)
693 if (conf
->mddev
->reshape_position
== MaxSector
)
694 return conf
->mddev
->degraded
> conf
->max_degraded
;
696 degraded
= raid5_calc_degraded(conf
);
697 if (degraded
> conf
->max_degraded
)
703 raid5_get_active_stripe(struct r5conf
*conf
, sector_t sector
,
704 int previous
, int noblock
, int noquiesce
)
706 struct stripe_head
*sh
;
707 int hash
= stripe_hash_locks_hash(conf
, sector
);
708 int inc_empty_inactive_list_flag
;
710 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
712 spin_lock_irq(conf
->hash_locks
+ hash
);
715 wait_event_lock_irq(conf
->wait_for_quiescent
,
716 conf
->quiesce
== 0 || noquiesce
,
717 *(conf
->hash_locks
+ hash
));
718 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
720 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
721 sh
= get_free_stripe(conf
, hash
);
722 if (!sh
&& !test_bit(R5_DID_ALLOC
,
724 set_bit(R5_ALLOC_MORE
,
727 if (noblock
&& sh
== NULL
)
730 r5c_check_stripe_cache_usage(conf
);
732 set_bit(R5_INACTIVE_BLOCKED
,
734 r5l_wake_reclaim(conf
->log
, 0);
736 conf
->wait_for_stripe
,
737 !list_empty(conf
->inactive_list
+ hash
) &&
738 (atomic_read(&conf
->active_stripes
)
739 < (conf
->max_nr_stripes
* 3 / 4)
740 || !test_bit(R5_INACTIVE_BLOCKED
,
741 &conf
->cache_state
)),
742 *(conf
->hash_locks
+ hash
));
743 clear_bit(R5_INACTIVE_BLOCKED
,
746 init_stripe(sh
, sector
, previous
);
747 atomic_inc(&sh
->count
);
749 } else if (!atomic_inc_not_zero(&sh
->count
)) {
750 spin_lock(&conf
->device_lock
);
751 if (!atomic_read(&sh
->count
)) {
752 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
753 atomic_inc(&conf
->active_stripes
);
754 BUG_ON(list_empty(&sh
->lru
) &&
755 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
756 inc_empty_inactive_list_flag
= 0;
757 if (!list_empty(conf
->inactive_list
+ hash
))
758 inc_empty_inactive_list_flag
= 1;
759 list_del_init(&sh
->lru
);
760 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
761 atomic_inc(&conf
->empty_inactive_list_nr
);
763 sh
->group
->stripes_cnt
--;
767 atomic_inc(&sh
->count
);
768 spin_unlock(&conf
->device_lock
);
770 } while (sh
== NULL
);
772 spin_unlock_irq(conf
->hash_locks
+ hash
);
776 static bool is_full_stripe_write(struct stripe_head
*sh
)
778 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
779 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
782 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
783 __acquires(&sh1
->stripe_lock
)
784 __acquires(&sh2
->stripe_lock
)
787 spin_lock_irq(&sh2
->stripe_lock
);
788 spin_lock_nested(&sh1
->stripe_lock
, 1);
790 spin_lock_irq(&sh1
->stripe_lock
);
791 spin_lock_nested(&sh2
->stripe_lock
, 1);
795 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
796 __releases(&sh1
->stripe_lock
)
797 __releases(&sh2
->stripe_lock
)
799 spin_unlock(&sh1
->stripe_lock
);
800 spin_unlock_irq(&sh2
->stripe_lock
);
803 /* Only freshly new full stripe normal write stripe can be added to a batch list */
804 static bool stripe_can_batch(struct stripe_head
*sh
)
806 struct r5conf
*conf
= sh
->raid_conf
;
808 if (raid5_has_log(conf
) || raid5_has_ppl(conf
))
810 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
811 !test_bit(STRIPE_BITMAP_PENDING
, &sh
->state
) &&
812 is_full_stripe_write(sh
);
815 /* we only do back search */
816 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
818 struct stripe_head
*head
;
819 sector_t head_sector
, tmp_sec
;
822 int inc_empty_inactive_list_flag
;
824 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
825 tmp_sec
= sh
->sector
;
826 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
828 head_sector
= sh
->sector
- RAID5_STRIPE_SECTORS(conf
);
830 hash
= stripe_hash_locks_hash(conf
, head_sector
);
831 spin_lock_irq(conf
->hash_locks
+ hash
);
832 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
833 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
834 spin_lock(&conf
->device_lock
);
835 if (!atomic_read(&head
->count
)) {
836 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
837 atomic_inc(&conf
->active_stripes
);
838 BUG_ON(list_empty(&head
->lru
) &&
839 !test_bit(STRIPE_EXPANDING
, &head
->state
));
840 inc_empty_inactive_list_flag
= 0;
841 if (!list_empty(conf
->inactive_list
+ hash
))
842 inc_empty_inactive_list_flag
= 1;
843 list_del_init(&head
->lru
);
844 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
845 atomic_inc(&conf
->empty_inactive_list_nr
);
847 head
->group
->stripes_cnt
--;
851 atomic_inc(&head
->count
);
852 spin_unlock(&conf
->device_lock
);
854 spin_unlock_irq(conf
->hash_locks
+ hash
);
858 if (!stripe_can_batch(head
))
861 lock_two_stripes(head
, sh
);
862 /* clear_batch_ready clear the flag */
863 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
870 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
872 if (head
->dev
[dd_idx
].towrite
->bi_opf
!= sh
->dev
[dd_idx
].towrite
->bi_opf
||
873 bio_op(head
->dev
[dd_idx
].towrite
) != bio_op(sh
->dev
[dd_idx
].towrite
))
876 if (head
->batch_head
) {
877 spin_lock(&head
->batch_head
->batch_lock
);
878 /* This batch list is already running */
879 if (!stripe_can_batch(head
)) {
880 spin_unlock(&head
->batch_head
->batch_lock
);
884 * We must assign batch_head of this stripe within the
885 * batch_lock, otherwise clear_batch_ready of batch head
886 * stripe could clear BATCH_READY bit of this stripe and
887 * this stripe->batch_head doesn't get assigned, which
888 * could confuse clear_batch_ready for this stripe
890 sh
->batch_head
= head
->batch_head
;
893 * at this point, head's BATCH_READY could be cleared, but we
894 * can still add the stripe to batch list
896 list_add(&sh
->batch_list
, &head
->batch_list
);
897 spin_unlock(&head
->batch_head
->batch_lock
);
899 head
->batch_head
= head
;
900 sh
->batch_head
= head
->batch_head
;
901 spin_lock(&head
->batch_lock
);
902 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
903 spin_unlock(&head
->batch_lock
);
906 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
907 if (atomic_dec_return(&conf
->preread_active_stripes
)
909 md_wakeup_thread(conf
->mddev
->thread
);
911 if (test_and_clear_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
912 int seq
= sh
->bm_seq
;
913 if (test_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
) &&
914 sh
->batch_head
->bm_seq
> seq
)
915 seq
= sh
->batch_head
->bm_seq
;
916 set_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
);
917 sh
->batch_head
->bm_seq
= seq
;
920 atomic_inc(&sh
->count
);
922 unlock_two_stripes(head
, sh
);
924 raid5_release_stripe(head
);
927 /* Determine if 'data_offset' or 'new_data_offset' should be used
928 * in this stripe_head.
930 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
932 sector_t progress
= conf
->reshape_progress
;
933 /* Need a memory barrier to make sure we see the value
934 * of conf->generation, or ->data_offset that was set before
935 * reshape_progress was updated.
938 if (progress
== MaxSector
)
940 if (sh
->generation
== conf
->generation
- 1)
942 /* We are in a reshape, and this is a new-generation stripe,
943 * so use new_data_offset.
948 static void dispatch_bio_list(struct bio_list
*tmp
)
952 while ((bio
= bio_list_pop(tmp
)))
953 submit_bio_noacct(bio
);
956 static int cmp_stripe(void *priv
, const struct list_head
*a
,
957 const struct list_head
*b
)
959 const struct r5pending_data
*da
= list_entry(a
,
960 struct r5pending_data
, sibling
);
961 const struct r5pending_data
*db
= list_entry(b
,
962 struct r5pending_data
, sibling
);
963 if (da
->sector
> db
->sector
)
965 if (da
->sector
< db
->sector
)
970 static void dispatch_defer_bios(struct r5conf
*conf
, int target
,
971 struct bio_list
*list
)
973 struct r5pending_data
*data
;
974 struct list_head
*first
, *next
= NULL
;
977 if (conf
->pending_data_cnt
== 0)
980 list_sort(NULL
, &conf
->pending_list
, cmp_stripe
);
982 first
= conf
->pending_list
.next
;
984 /* temporarily move the head */
985 if (conf
->next_pending_data
)
986 list_move_tail(&conf
->pending_list
,
987 &conf
->next_pending_data
->sibling
);
989 while (!list_empty(&conf
->pending_list
)) {
990 data
= list_first_entry(&conf
->pending_list
,
991 struct r5pending_data
, sibling
);
992 if (&data
->sibling
== first
)
993 first
= data
->sibling
.next
;
994 next
= data
->sibling
.next
;
996 bio_list_merge(list
, &data
->bios
);
997 list_move(&data
->sibling
, &conf
->free_list
);
1002 conf
->pending_data_cnt
-= cnt
;
1003 BUG_ON(conf
->pending_data_cnt
< 0 || cnt
< target
);
1005 if (next
!= &conf
->pending_list
)
1006 conf
->next_pending_data
= list_entry(next
,
1007 struct r5pending_data
, sibling
);
1009 conf
->next_pending_data
= NULL
;
1010 /* list isn't empty */
1011 if (first
!= &conf
->pending_list
)
1012 list_move_tail(&conf
->pending_list
, first
);
1015 static void flush_deferred_bios(struct r5conf
*conf
)
1017 struct bio_list tmp
= BIO_EMPTY_LIST
;
1019 if (conf
->pending_data_cnt
== 0)
1022 spin_lock(&conf
->pending_bios_lock
);
1023 dispatch_defer_bios(conf
, conf
->pending_data_cnt
, &tmp
);
1024 BUG_ON(conf
->pending_data_cnt
!= 0);
1025 spin_unlock(&conf
->pending_bios_lock
);
1027 dispatch_bio_list(&tmp
);
1030 static void defer_issue_bios(struct r5conf
*conf
, sector_t sector
,
1031 struct bio_list
*bios
)
1033 struct bio_list tmp
= BIO_EMPTY_LIST
;
1034 struct r5pending_data
*ent
;
1036 spin_lock(&conf
->pending_bios_lock
);
1037 ent
= list_first_entry(&conf
->free_list
, struct r5pending_data
,
1039 list_move_tail(&ent
->sibling
, &conf
->pending_list
);
1040 ent
->sector
= sector
;
1041 bio_list_init(&ent
->bios
);
1042 bio_list_merge(&ent
->bios
, bios
);
1043 conf
->pending_data_cnt
++;
1044 if (conf
->pending_data_cnt
>= PENDING_IO_MAX
)
1045 dispatch_defer_bios(conf
, PENDING_IO_ONE_FLUSH
, &tmp
);
1047 spin_unlock(&conf
->pending_bios_lock
);
1049 dispatch_bio_list(&tmp
);
1053 raid5_end_read_request(struct bio
*bi
);
1055 raid5_end_write_request(struct bio
*bi
);
1057 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
1059 struct r5conf
*conf
= sh
->raid_conf
;
1060 int i
, disks
= sh
->disks
;
1061 struct stripe_head
*head_sh
= sh
;
1062 struct bio_list pending_bios
= BIO_EMPTY_LIST
;
1067 if (log_stripe(sh
, s
) == 0)
1070 should_defer
= conf
->batch_bio_dispatch
&& conf
->group_cnt
;
1072 for (i
= disks
; i
--; ) {
1073 int op
, op_flags
= 0;
1074 int replace_only
= 0;
1075 struct bio
*bi
, *rbi
;
1076 struct md_rdev
*rdev
, *rrdev
= NULL
;
1079 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
1081 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
1083 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1084 op
= REQ_OP_DISCARD
;
1085 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
1087 else if (test_and_clear_bit(R5_WantReplace
,
1088 &sh
->dev
[i
].flags
)) {
1093 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
1094 op_flags
|= REQ_SYNC
;
1097 bi
= &sh
->dev
[i
].req
;
1098 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
1101 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
1102 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1103 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
1108 if (op_is_write(op
)) {
1112 /* We raced and saw duplicates */
1115 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
1120 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
1123 atomic_inc(&rdev
->nr_pending
);
1124 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
1127 atomic_inc(&rrdev
->nr_pending
);
1130 /* We have already checked bad blocks for reads. Now
1131 * need to check for writes. We never accept write errors
1132 * on the replacement, so we don't to check rrdev.
1134 while (op_is_write(op
) && rdev
&&
1135 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1138 int bad
= is_badblock(rdev
, sh
->sector
, RAID5_STRIPE_SECTORS(conf
),
1139 &first_bad
, &bad_sectors
);
1144 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1145 if (!conf
->mddev
->external
&&
1146 conf
->mddev
->sb_flags
) {
1147 /* It is very unlikely, but we might
1148 * still need to write out the
1149 * bad block log - better give it
1151 md_check_recovery(conf
->mddev
);
1154 * Because md_wait_for_blocked_rdev
1155 * will dec nr_pending, we must
1156 * increment it first.
1158 atomic_inc(&rdev
->nr_pending
);
1159 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
1161 /* Acknowledged bad block - skip the write */
1162 rdev_dec_pending(rdev
, conf
->mddev
);
1168 if (s
->syncing
|| s
->expanding
|| s
->expanded
1170 md_sync_acct(rdev
->bdev
, RAID5_STRIPE_SECTORS(conf
));
1172 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1174 bio_set_dev(bi
, rdev
->bdev
);
1175 bio_set_op_attrs(bi
, op
, op_flags
);
1176 bi
->bi_end_io
= op_is_write(op
)
1177 ? raid5_end_write_request
1178 : raid5_end_read_request
;
1179 bi
->bi_private
= sh
;
1181 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1182 __func__
, (unsigned long long)sh
->sector
,
1184 atomic_inc(&sh
->count
);
1186 atomic_inc(&head_sh
->count
);
1187 if (use_new_offset(conf
, sh
))
1188 bi
->bi_iter
.bi_sector
= (sh
->sector
1189 + rdev
->new_data_offset
);
1191 bi
->bi_iter
.bi_sector
= (sh
->sector
1192 + rdev
->data_offset
);
1193 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1194 bi
->bi_opf
|= REQ_NOMERGE
;
1196 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1197 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1199 if (!op_is_write(op
) &&
1200 test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
1202 * issuing read for a page in journal, this
1203 * must be preparing for prexor in rmw; read
1204 * the data into orig_page
1206 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].orig_page
;
1208 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1210 bi
->bi_io_vec
[0].bv_len
= RAID5_STRIPE_SIZE(conf
);
1211 bi
->bi_io_vec
[0].bv_offset
= sh
->dev
[i
].offset
;
1212 bi
->bi_iter
.bi_size
= RAID5_STRIPE_SIZE(conf
);
1213 bi
->bi_write_hint
= sh
->dev
[i
].write_hint
;
1215 sh
->dev
[i
].write_hint
= RWH_WRITE_LIFE_NOT_SET
;
1217 * If this is discard request, set bi_vcnt 0. We don't
1218 * want to confuse SCSI because SCSI will replace payload
1220 if (op
== REQ_OP_DISCARD
)
1223 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1225 if (conf
->mddev
->gendisk
)
1226 trace_block_bio_remap(bi
,
1227 disk_devt(conf
->mddev
->gendisk
),
1229 if (should_defer
&& op_is_write(op
))
1230 bio_list_add(&pending_bios
, bi
);
1232 submit_bio_noacct(bi
);
1235 if (s
->syncing
|| s
->expanding
|| s
->expanded
1237 md_sync_acct(rrdev
->bdev
, RAID5_STRIPE_SECTORS(conf
));
1239 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1241 bio_set_dev(rbi
, rrdev
->bdev
);
1242 bio_set_op_attrs(rbi
, op
, op_flags
);
1243 BUG_ON(!op_is_write(op
));
1244 rbi
->bi_end_io
= raid5_end_write_request
;
1245 rbi
->bi_private
= sh
;
1247 pr_debug("%s: for %llu schedule op %d on "
1248 "replacement disc %d\n",
1249 __func__
, (unsigned long long)sh
->sector
,
1251 atomic_inc(&sh
->count
);
1253 atomic_inc(&head_sh
->count
);
1254 if (use_new_offset(conf
, sh
))
1255 rbi
->bi_iter
.bi_sector
= (sh
->sector
1256 + rrdev
->new_data_offset
);
1258 rbi
->bi_iter
.bi_sector
= (sh
->sector
1259 + rrdev
->data_offset
);
1260 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1261 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1262 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1264 rbi
->bi_io_vec
[0].bv_len
= RAID5_STRIPE_SIZE(conf
);
1265 rbi
->bi_io_vec
[0].bv_offset
= sh
->dev
[i
].offset
;
1266 rbi
->bi_iter
.bi_size
= RAID5_STRIPE_SIZE(conf
);
1267 rbi
->bi_write_hint
= sh
->dev
[i
].write_hint
;
1268 sh
->dev
[i
].write_hint
= RWH_WRITE_LIFE_NOT_SET
;
1270 * If this is discard request, set bi_vcnt 0. We don't
1271 * want to confuse SCSI because SCSI will replace payload
1273 if (op
== REQ_OP_DISCARD
)
1275 if (conf
->mddev
->gendisk
)
1276 trace_block_bio_remap(rbi
,
1277 disk_devt(conf
->mddev
->gendisk
),
1279 if (should_defer
&& op_is_write(op
))
1280 bio_list_add(&pending_bios
, rbi
);
1282 submit_bio_noacct(rbi
);
1284 if (!rdev
&& !rrdev
) {
1285 if (op_is_write(op
))
1286 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1287 pr_debug("skip op %d on disc %d for sector %llu\n",
1288 bi
->bi_opf
, i
, (unsigned long long)sh
->sector
);
1289 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1290 set_bit(STRIPE_HANDLE
, &sh
->state
);
1293 if (!head_sh
->batch_head
)
1295 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1301 if (should_defer
&& !bio_list_empty(&pending_bios
))
1302 defer_issue_bios(conf
, head_sh
->sector
, &pending_bios
);
1305 static struct dma_async_tx_descriptor
*
1306 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1307 unsigned int poff
, sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1308 struct stripe_head
*sh
, int no_skipcopy
)
1311 struct bvec_iter iter
;
1312 struct page
*bio_page
;
1314 struct async_submit_ctl submit
;
1315 enum async_tx_flags flags
= 0;
1316 struct r5conf
*conf
= sh
->raid_conf
;
1318 if (bio
->bi_iter
.bi_sector
>= sector
)
1319 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1321 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1324 flags
|= ASYNC_TX_FENCE
;
1325 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1327 bio_for_each_segment(bvl
, bio
, iter
) {
1328 int len
= bvl
.bv_len
;
1332 if (page_offset
< 0) {
1333 b_offset
= -page_offset
;
1334 page_offset
+= b_offset
;
1338 if (len
> 0 && page_offset
+ len
> RAID5_STRIPE_SIZE(conf
))
1339 clen
= RAID5_STRIPE_SIZE(conf
) - page_offset
;
1344 b_offset
+= bvl
.bv_offset
;
1345 bio_page
= bvl
.bv_page
;
1347 if (conf
->skip_copy
&&
1348 b_offset
== 0 && page_offset
== 0 &&
1349 clen
== RAID5_STRIPE_SIZE(conf
) &&
1353 tx
= async_memcpy(*page
, bio_page
, page_offset
+ poff
,
1354 b_offset
, clen
, &submit
);
1356 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1357 page_offset
+ poff
, clen
, &submit
);
1359 /* chain the operations */
1360 submit
.depend_tx
= tx
;
1362 if (clen
< len
) /* hit end of page */
1370 static void ops_complete_biofill(void *stripe_head_ref
)
1372 struct stripe_head
*sh
= stripe_head_ref
;
1374 struct r5conf
*conf
= sh
->raid_conf
;
1376 pr_debug("%s: stripe %llu\n", __func__
,
1377 (unsigned long long)sh
->sector
);
1379 /* clear completed biofills */
1380 for (i
= sh
->disks
; i
--; ) {
1381 struct r5dev
*dev
= &sh
->dev
[i
];
1383 /* acknowledge completion of a biofill operation */
1384 /* and check if we need to reply to a read request,
1385 * new R5_Wantfill requests are held off until
1386 * !STRIPE_BIOFILL_RUN
1388 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1389 struct bio
*rbi
, *rbi2
;
1394 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1395 dev
->sector
+ RAID5_STRIPE_SECTORS(conf
)) {
1396 rbi2
= r5_next_bio(conf
, rbi
, dev
->sector
);
1402 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1404 set_bit(STRIPE_HANDLE
, &sh
->state
);
1405 raid5_release_stripe(sh
);
1408 static void ops_run_biofill(struct stripe_head
*sh
)
1410 struct dma_async_tx_descriptor
*tx
= NULL
;
1411 struct async_submit_ctl submit
;
1413 struct r5conf
*conf
= sh
->raid_conf
;
1415 BUG_ON(sh
->batch_head
);
1416 pr_debug("%s: stripe %llu\n", __func__
,
1417 (unsigned long long)sh
->sector
);
1419 for (i
= sh
->disks
; i
--; ) {
1420 struct r5dev
*dev
= &sh
->dev
[i
];
1421 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1423 spin_lock_irq(&sh
->stripe_lock
);
1424 dev
->read
= rbi
= dev
->toread
;
1426 spin_unlock_irq(&sh
->stripe_lock
);
1427 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1428 dev
->sector
+ RAID5_STRIPE_SECTORS(conf
)) {
1429 tx
= async_copy_data(0, rbi
, &dev
->page
,
1431 dev
->sector
, tx
, sh
, 0);
1432 rbi
= r5_next_bio(conf
, rbi
, dev
->sector
);
1437 atomic_inc(&sh
->count
);
1438 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1439 async_trigger_callback(&submit
);
1442 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1449 tgt
= &sh
->dev
[target
];
1450 set_bit(R5_UPTODATE
, &tgt
->flags
);
1451 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1452 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1455 static void ops_complete_compute(void *stripe_head_ref
)
1457 struct stripe_head
*sh
= stripe_head_ref
;
1459 pr_debug("%s: stripe %llu\n", __func__
,
1460 (unsigned long long)sh
->sector
);
1462 /* mark the computed target(s) as uptodate */
1463 mark_target_uptodate(sh
, sh
->ops
.target
);
1464 mark_target_uptodate(sh
, sh
->ops
.target2
);
1466 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1467 if (sh
->check_state
== check_state_compute_run
)
1468 sh
->check_state
= check_state_compute_result
;
1469 set_bit(STRIPE_HANDLE
, &sh
->state
);
1470 raid5_release_stripe(sh
);
1473 /* return a pointer to the address conversion region of the scribble buffer */
1474 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1476 return percpu
->scribble
+ i
* percpu
->scribble_obj_size
;
1479 /* return a pointer to the address conversion region of the scribble buffer */
1480 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1481 struct raid5_percpu
*percpu
, int i
)
1483 return (void *) (to_addr_page(percpu
, i
) + sh
->disks
+ 2);
1487 * Return a pointer to record offset address.
1489 static unsigned int *
1490 to_addr_offs(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1492 return (unsigned int *) (to_addr_conv(sh
, percpu
, 0) + sh
->disks
+ 2);
1495 static struct dma_async_tx_descriptor
*
1496 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1498 int disks
= sh
->disks
;
1499 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1500 unsigned int *off_srcs
= to_addr_offs(sh
, percpu
);
1501 int target
= sh
->ops
.target
;
1502 struct r5dev
*tgt
= &sh
->dev
[target
];
1503 struct page
*xor_dest
= tgt
->page
;
1504 unsigned int off_dest
= tgt
->offset
;
1506 struct dma_async_tx_descriptor
*tx
;
1507 struct async_submit_ctl submit
;
1510 BUG_ON(sh
->batch_head
);
1512 pr_debug("%s: stripe %llu block: %d\n",
1513 __func__
, (unsigned long long)sh
->sector
, target
);
1514 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1516 for (i
= disks
; i
--; ) {
1518 off_srcs
[count
] = sh
->dev
[i
].offset
;
1519 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1523 atomic_inc(&sh
->count
);
1525 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1526 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1527 if (unlikely(count
== 1))
1528 tx
= async_memcpy(xor_dest
, xor_srcs
[0], off_dest
, off_srcs
[0],
1529 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1531 tx
= async_xor_offs(xor_dest
, off_dest
, xor_srcs
, off_srcs
, count
,
1532 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1537 /* set_syndrome_sources - populate source buffers for gen_syndrome
1538 * @srcs - (struct page *) array of size sh->disks
1539 * @offs - (unsigned int) array of offset for each page
1540 * @sh - stripe_head to parse
1542 * Populates srcs in proper layout order for the stripe and returns the
1543 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1544 * destination buffer is recorded in srcs[count] and the Q destination
1545 * is recorded in srcs[count+1]].
1547 static int set_syndrome_sources(struct page
**srcs
,
1549 struct stripe_head
*sh
,
1552 int disks
= sh
->disks
;
1553 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1554 int d0_idx
= raid6_d0(sh
);
1558 for (i
= 0; i
< disks
; i
++)
1564 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1565 struct r5dev
*dev
= &sh
->dev
[i
];
1567 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1568 (srctype
== SYNDROME_SRC_ALL
) ||
1569 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1570 (test_bit(R5_Wantdrain
, &dev
->flags
) ||
1571 test_bit(R5_InJournal
, &dev
->flags
))) ||
1572 (srctype
== SYNDROME_SRC_WRITTEN
&&
1574 test_bit(R5_InJournal
, &dev
->flags
)))) {
1575 if (test_bit(R5_InJournal
, &dev
->flags
))
1576 srcs
[slot
] = sh
->dev
[i
].orig_page
;
1578 srcs
[slot
] = sh
->dev
[i
].page
;
1580 * For R5_InJournal, PAGE_SIZE must be 4KB and will
1581 * not shared page. In that case, dev[i].offset
1584 offs
[slot
] = sh
->dev
[i
].offset
;
1586 i
= raid6_next_disk(i
, disks
);
1587 } while (i
!= d0_idx
);
1589 return syndrome_disks
;
1592 static struct dma_async_tx_descriptor
*
1593 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1595 int disks
= sh
->disks
;
1596 struct page
**blocks
= to_addr_page(percpu
, 0);
1597 unsigned int *offs
= to_addr_offs(sh
, percpu
);
1599 int qd_idx
= sh
->qd_idx
;
1600 struct dma_async_tx_descriptor
*tx
;
1601 struct async_submit_ctl submit
;
1604 unsigned int dest_off
;
1608 BUG_ON(sh
->batch_head
);
1609 if (sh
->ops
.target
< 0)
1610 target
= sh
->ops
.target2
;
1611 else if (sh
->ops
.target2
< 0)
1612 target
= sh
->ops
.target
;
1614 /* we should only have one valid target */
1617 pr_debug("%s: stripe %llu block: %d\n",
1618 __func__
, (unsigned long long)sh
->sector
, target
);
1620 tgt
= &sh
->dev
[target
];
1621 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1623 dest_off
= tgt
->offset
;
1625 atomic_inc(&sh
->count
);
1627 if (target
== qd_idx
) {
1628 count
= set_syndrome_sources(blocks
, offs
, sh
, SYNDROME_SRC_ALL
);
1629 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1630 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1631 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1632 ops_complete_compute
, sh
,
1633 to_addr_conv(sh
, percpu
, 0));
1634 tx
= async_gen_syndrome(blocks
, offs
, count
+2,
1635 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1637 /* Compute any data- or p-drive using XOR */
1639 for (i
= disks
; i
-- ; ) {
1640 if (i
== target
|| i
== qd_idx
)
1642 offs
[count
] = sh
->dev
[i
].offset
;
1643 blocks
[count
++] = sh
->dev
[i
].page
;
1646 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1647 NULL
, ops_complete_compute
, sh
,
1648 to_addr_conv(sh
, percpu
, 0));
1649 tx
= async_xor_offs(dest
, dest_off
, blocks
, offs
, count
,
1650 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1656 static struct dma_async_tx_descriptor
*
1657 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1659 int i
, count
, disks
= sh
->disks
;
1660 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1661 int d0_idx
= raid6_d0(sh
);
1662 int faila
= -1, failb
= -1;
1663 int target
= sh
->ops
.target
;
1664 int target2
= sh
->ops
.target2
;
1665 struct r5dev
*tgt
= &sh
->dev
[target
];
1666 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1667 struct dma_async_tx_descriptor
*tx
;
1668 struct page
**blocks
= to_addr_page(percpu
, 0);
1669 unsigned int *offs
= to_addr_offs(sh
, percpu
);
1670 struct async_submit_ctl submit
;
1672 BUG_ON(sh
->batch_head
);
1673 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1674 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1675 BUG_ON(target
< 0 || target2
< 0);
1676 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1677 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1679 /* we need to open-code set_syndrome_sources to handle the
1680 * slot number conversion for 'faila' and 'failb'
1682 for (i
= 0; i
< disks
; i
++) {
1689 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1691 offs
[slot
] = sh
->dev
[i
].offset
;
1692 blocks
[slot
] = sh
->dev
[i
].page
;
1698 i
= raid6_next_disk(i
, disks
);
1699 } while (i
!= d0_idx
);
1701 BUG_ON(faila
== failb
);
1704 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1705 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1707 atomic_inc(&sh
->count
);
1709 if (failb
== syndrome_disks
+1) {
1710 /* Q disk is one of the missing disks */
1711 if (faila
== syndrome_disks
) {
1712 /* Missing P+Q, just recompute */
1713 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1714 ops_complete_compute
, sh
,
1715 to_addr_conv(sh
, percpu
, 0));
1716 return async_gen_syndrome(blocks
, offs
, syndrome_disks
+2,
1717 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1721 unsigned int dest_off
;
1723 int qd_idx
= sh
->qd_idx
;
1725 /* Missing D+Q: recompute D from P, then recompute Q */
1726 if (target
== qd_idx
)
1727 data_target
= target2
;
1729 data_target
= target
;
1732 for (i
= disks
; i
-- ; ) {
1733 if (i
== data_target
|| i
== qd_idx
)
1735 offs
[count
] = sh
->dev
[i
].offset
;
1736 blocks
[count
++] = sh
->dev
[i
].page
;
1738 dest
= sh
->dev
[data_target
].page
;
1739 dest_off
= sh
->dev
[data_target
].offset
;
1740 init_async_submit(&submit
,
1741 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1743 to_addr_conv(sh
, percpu
, 0));
1744 tx
= async_xor_offs(dest
, dest_off
, blocks
, offs
, count
,
1745 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1748 count
= set_syndrome_sources(blocks
, offs
, sh
, SYNDROME_SRC_ALL
);
1749 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1750 ops_complete_compute
, sh
,
1751 to_addr_conv(sh
, percpu
, 0));
1752 return async_gen_syndrome(blocks
, offs
, count
+2,
1753 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1757 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1758 ops_complete_compute
, sh
,
1759 to_addr_conv(sh
, percpu
, 0));
1760 if (failb
== syndrome_disks
) {
1761 /* We're missing D+P. */
1762 return async_raid6_datap_recov(syndrome_disks
+2,
1763 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1765 blocks
, offs
, &submit
);
1767 /* We're missing D+D. */
1768 return async_raid6_2data_recov(syndrome_disks
+2,
1769 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1771 blocks
, offs
, &submit
);
1776 static void ops_complete_prexor(void *stripe_head_ref
)
1778 struct stripe_head
*sh
= stripe_head_ref
;
1780 pr_debug("%s: stripe %llu\n", __func__
,
1781 (unsigned long long)sh
->sector
);
1783 if (r5c_is_writeback(sh
->raid_conf
->log
))
1785 * raid5-cache write back uses orig_page during prexor.
1786 * After prexor, it is time to free orig_page
1788 r5c_release_extra_page(sh
);
1791 static struct dma_async_tx_descriptor
*
1792 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1793 struct dma_async_tx_descriptor
*tx
)
1795 int disks
= sh
->disks
;
1796 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1797 unsigned int *off_srcs
= to_addr_offs(sh
, percpu
);
1798 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1799 struct async_submit_ctl submit
;
1801 /* existing parity data subtracted */
1802 unsigned int off_dest
= off_srcs
[count
] = sh
->dev
[pd_idx
].offset
;
1803 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1805 BUG_ON(sh
->batch_head
);
1806 pr_debug("%s: stripe %llu\n", __func__
,
1807 (unsigned long long)sh
->sector
);
1809 for (i
= disks
; i
--; ) {
1810 struct r5dev
*dev
= &sh
->dev
[i
];
1811 /* Only process blocks that are known to be uptodate */
1812 if (test_bit(R5_InJournal
, &dev
->flags
)) {
1814 * For this case, PAGE_SIZE must be equal to 4KB and
1815 * page offset is zero.
1817 off_srcs
[count
] = dev
->offset
;
1818 xor_srcs
[count
++] = dev
->orig_page
;
1819 } else if (test_bit(R5_Wantdrain
, &dev
->flags
)) {
1820 off_srcs
[count
] = dev
->offset
;
1821 xor_srcs
[count
++] = dev
->page
;
1825 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1826 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1827 tx
= async_xor_offs(xor_dest
, off_dest
, xor_srcs
, off_srcs
, count
,
1828 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1833 static struct dma_async_tx_descriptor
*
1834 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1835 struct dma_async_tx_descriptor
*tx
)
1837 struct page
**blocks
= to_addr_page(percpu
, 0);
1838 unsigned int *offs
= to_addr_offs(sh
, percpu
);
1840 struct async_submit_ctl submit
;
1842 pr_debug("%s: stripe %llu\n", __func__
,
1843 (unsigned long long)sh
->sector
);
1845 count
= set_syndrome_sources(blocks
, offs
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1847 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1848 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1849 tx
= async_gen_syndrome(blocks
, offs
, count
+2,
1850 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1855 static struct dma_async_tx_descriptor
*
1856 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1858 struct r5conf
*conf
= sh
->raid_conf
;
1859 int disks
= sh
->disks
;
1861 struct stripe_head
*head_sh
= sh
;
1863 pr_debug("%s: stripe %llu\n", __func__
,
1864 (unsigned long long)sh
->sector
);
1866 for (i
= disks
; i
--; ) {
1871 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1877 * clear R5_InJournal, so when rewriting a page in
1878 * journal, it is not skipped by r5l_log_stripe()
1880 clear_bit(R5_InJournal
, &dev
->flags
);
1881 spin_lock_irq(&sh
->stripe_lock
);
1882 chosen
= dev
->towrite
;
1883 dev
->towrite
= NULL
;
1884 sh
->overwrite_disks
= 0;
1885 BUG_ON(dev
->written
);
1886 wbi
= dev
->written
= chosen
;
1887 spin_unlock_irq(&sh
->stripe_lock
);
1888 WARN_ON(dev
->page
!= dev
->orig_page
);
1890 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1891 dev
->sector
+ RAID5_STRIPE_SECTORS(conf
)) {
1892 if (wbi
->bi_opf
& REQ_FUA
)
1893 set_bit(R5_WantFUA
, &dev
->flags
);
1894 if (wbi
->bi_opf
& REQ_SYNC
)
1895 set_bit(R5_SyncIO
, &dev
->flags
);
1896 if (bio_op(wbi
) == REQ_OP_DISCARD
)
1897 set_bit(R5_Discard
, &dev
->flags
);
1899 tx
= async_copy_data(1, wbi
, &dev
->page
,
1901 dev
->sector
, tx
, sh
,
1902 r5c_is_writeback(conf
->log
));
1903 if (dev
->page
!= dev
->orig_page
&&
1904 !r5c_is_writeback(conf
->log
)) {
1905 set_bit(R5_SkipCopy
, &dev
->flags
);
1906 clear_bit(R5_UPTODATE
, &dev
->flags
);
1907 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1910 wbi
= r5_next_bio(conf
, wbi
, dev
->sector
);
1913 if (head_sh
->batch_head
) {
1914 sh
= list_first_entry(&sh
->batch_list
,
1927 static void ops_complete_reconstruct(void *stripe_head_ref
)
1929 struct stripe_head
*sh
= stripe_head_ref
;
1930 int disks
= sh
->disks
;
1931 int pd_idx
= sh
->pd_idx
;
1932 int qd_idx
= sh
->qd_idx
;
1934 bool fua
= false, sync
= false, discard
= false;
1936 pr_debug("%s: stripe %llu\n", __func__
,
1937 (unsigned long long)sh
->sector
);
1939 for (i
= disks
; i
--; ) {
1940 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1941 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1942 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1945 for (i
= disks
; i
--; ) {
1946 struct r5dev
*dev
= &sh
->dev
[i
];
1948 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1949 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
)) {
1950 set_bit(R5_UPTODATE
, &dev
->flags
);
1951 if (test_bit(STRIPE_EXPAND_READY
, &sh
->state
))
1952 set_bit(R5_Expanded
, &dev
->flags
);
1955 set_bit(R5_WantFUA
, &dev
->flags
);
1957 set_bit(R5_SyncIO
, &dev
->flags
);
1961 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1962 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1963 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1964 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1966 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1967 sh
->reconstruct_state
= reconstruct_state_result
;
1970 set_bit(STRIPE_HANDLE
, &sh
->state
);
1971 raid5_release_stripe(sh
);
1975 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1976 struct dma_async_tx_descriptor
*tx
)
1978 int disks
= sh
->disks
;
1979 struct page
**xor_srcs
;
1980 unsigned int *off_srcs
;
1981 struct async_submit_ctl submit
;
1982 int count
, pd_idx
= sh
->pd_idx
, i
;
1983 struct page
*xor_dest
;
1984 unsigned int off_dest
;
1986 unsigned long flags
;
1988 struct stripe_head
*head_sh
= sh
;
1991 pr_debug("%s: stripe %llu\n", __func__
,
1992 (unsigned long long)sh
->sector
);
1994 for (i
= 0; i
< sh
->disks
; i
++) {
1997 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
2000 if (i
>= sh
->disks
) {
2001 atomic_inc(&sh
->count
);
2002 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
2003 ops_complete_reconstruct(sh
);
2008 xor_srcs
= to_addr_page(percpu
, j
);
2009 off_srcs
= to_addr_offs(sh
, percpu
);
2010 /* check if prexor is active which means only process blocks
2011 * that are part of a read-modify-write (written)
2013 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
2015 off_dest
= off_srcs
[count
] = sh
->dev
[pd_idx
].offset
;
2016 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
2017 for (i
= disks
; i
--; ) {
2018 struct r5dev
*dev
= &sh
->dev
[i
];
2019 if (head_sh
->dev
[i
].written
||
2020 test_bit(R5_InJournal
, &head_sh
->dev
[i
].flags
)) {
2021 off_srcs
[count
] = dev
->offset
;
2022 xor_srcs
[count
++] = dev
->page
;
2026 xor_dest
= sh
->dev
[pd_idx
].page
;
2027 off_dest
= sh
->dev
[pd_idx
].offset
;
2028 for (i
= disks
; i
--; ) {
2029 struct r5dev
*dev
= &sh
->dev
[i
];
2031 off_srcs
[count
] = dev
->offset
;
2032 xor_srcs
[count
++] = dev
->page
;
2037 /* 1/ if we prexor'd then the dest is reused as a source
2038 * 2/ if we did not prexor then we are redoing the parity
2039 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2040 * for the synchronous xor case
2042 last_stripe
= !head_sh
->batch_head
||
2043 list_first_entry(&sh
->batch_list
,
2044 struct stripe_head
, batch_list
) == head_sh
;
2046 flags
= ASYNC_TX_ACK
|
2047 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
2049 atomic_inc(&head_sh
->count
);
2050 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
2051 to_addr_conv(sh
, percpu
, j
));
2053 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
2054 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
2055 to_addr_conv(sh
, percpu
, j
));
2058 if (unlikely(count
== 1))
2059 tx
= async_memcpy(xor_dest
, xor_srcs
[0], off_dest
, off_srcs
[0],
2060 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
2062 tx
= async_xor_offs(xor_dest
, off_dest
, xor_srcs
, off_srcs
, count
,
2063 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
2066 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
2073 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
2074 struct dma_async_tx_descriptor
*tx
)
2076 struct async_submit_ctl submit
;
2077 struct page
**blocks
;
2079 int count
, i
, j
= 0;
2080 struct stripe_head
*head_sh
= sh
;
2083 unsigned long txflags
;
2085 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
2087 for (i
= 0; i
< sh
->disks
; i
++) {
2088 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
2090 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
2093 if (i
>= sh
->disks
) {
2094 atomic_inc(&sh
->count
);
2095 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
2096 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
2097 ops_complete_reconstruct(sh
);
2102 blocks
= to_addr_page(percpu
, j
);
2103 offs
= to_addr_offs(sh
, percpu
);
2105 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
2106 synflags
= SYNDROME_SRC_WRITTEN
;
2107 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
2109 synflags
= SYNDROME_SRC_ALL
;
2110 txflags
= ASYNC_TX_ACK
;
2113 count
= set_syndrome_sources(blocks
, offs
, sh
, synflags
);
2114 last_stripe
= !head_sh
->batch_head
||
2115 list_first_entry(&sh
->batch_list
,
2116 struct stripe_head
, batch_list
) == head_sh
;
2119 atomic_inc(&head_sh
->count
);
2120 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
2121 head_sh
, to_addr_conv(sh
, percpu
, j
));
2123 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
2124 to_addr_conv(sh
, percpu
, j
));
2125 tx
= async_gen_syndrome(blocks
, offs
, count
+2,
2126 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
2129 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
2135 static void ops_complete_check(void *stripe_head_ref
)
2137 struct stripe_head
*sh
= stripe_head_ref
;
2139 pr_debug("%s: stripe %llu\n", __func__
,
2140 (unsigned long long)sh
->sector
);
2142 sh
->check_state
= check_state_check_result
;
2143 set_bit(STRIPE_HANDLE
, &sh
->state
);
2144 raid5_release_stripe(sh
);
2147 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
2149 int disks
= sh
->disks
;
2150 int pd_idx
= sh
->pd_idx
;
2151 int qd_idx
= sh
->qd_idx
;
2152 struct page
*xor_dest
;
2153 unsigned int off_dest
;
2154 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
2155 unsigned int *off_srcs
= to_addr_offs(sh
, percpu
);
2156 struct dma_async_tx_descriptor
*tx
;
2157 struct async_submit_ctl submit
;
2161 pr_debug("%s: stripe %llu\n", __func__
,
2162 (unsigned long long)sh
->sector
);
2164 BUG_ON(sh
->batch_head
);
2166 xor_dest
= sh
->dev
[pd_idx
].page
;
2167 off_dest
= sh
->dev
[pd_idx
].offset
;
2168 off_srcs
[count
] = off_dest
;
2169 xor_srcs
[count
++] = xor_dest
;
2170 for (i
= disks
; i
--; ) {
2171 if (i
== pd_idx
|| i
== qd_idx
)
2173 off_srcs
[count
] = sh
->dev
[i
].offset
;
2174 xor_srcs
[count
++] = sh
->dev
[i
].page
;
2177 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
2178 to_addr_conv(sh
, percpu
, 0));
2179 tx
= async_xor_val_offs(xor_dest
, off_dest
, xor_srcs
, off_srcs
, count
,
2180 RAID5_STRIPE_SIZE(sh
->raid_conf
),
2181 &sh
->ops
.zero_sum_result
, &submit
);
2183 atomic_inc(&sh
->count
);
2184 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
2185 tx
= async_trigger_callback(&submit
);
2188 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
2190 struct page
**srcs
= to_addr_page(percpu
, 0);
2191 unsigned int *offs
= to_addr_offs(sh
, percpu
);
2192 struct async_submit_ctl submit
;
2195 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
2196 (unsigned long long)sh
->sector
, checkp
);
2198 BUG_ON(sh
->batch_head
);
2199 count
= set_syndrome_sources(srcs
, offs
, sh
, SYNDROME_SRC_ALL
);
2203 atomic_inc(&sh
->count
);
2204 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
2205 sh
, to_addr_conv(sh
, percpu
, 0));
2206 async_syndrome_val(srcs
, offs
, count
+2,
2207 RAID5_STRIPE_SIZE(sh
->raid_conf
),
2208 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, 0, &submit
);
2211 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
2213 int overlap_clear
= 0, i
, disks
= sh
->disks
;
2214 struct dma_async_tx_descriptor
*tx
= NULL
;
2215 struct r5conf
*conf
= sh
->raid_conf
;
2216 int level
= conf
->level
;
2217 struct raid5_percpu
*percpu
;
2221 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2222 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
2223 ops_run_biofill(sh
);
2227 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
2229 tx
= ops_run_compute5(sh
, percpu
);
2231 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
2232 tx
= ops_run_compute6_1(sh
, percpu
);
2234 tx
= ops_run_compute6_2(sh
, percpu
);
2236 /* terminate the chain if reconstruct is not set to be run */
2237 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
2241 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
2243 tx
= ops_run_prexor5(sh
, percpu
, tx
);
2245 tx
= ops_run_prexor6(sh
, percpu
, tx
);
2248 if (test_bit(STRIPE_OP_PARTIAL_PARITY
, &ops_request
))
2249 tx
= ops_run_partial_parity(sh
, percpu
, tx
);
2251 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
2252 tx
= ops_run_biodrain(sh
, tx
);
2256 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
2258 ops_run_reconstruct5(sh
, percpu
, tx
);
2260 ops_run_reconstruct6(sh
, percpu
, tx
);
2263 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
2264 if (sh
->check_state
== check_state_run
)
2265 ops_run_check_p(sh
, percpu
);
2266 else if (sh
->check_state
== check_state_run_q
)
2267 ops_run_check_pq(sh
, percpu
, 0);
2268 else if (sh
->check_state
== check_state_run_pq
)
2269 ops_run_check_pq(sh
, percpu
, 1);
2274 if (overlap_clear
&& !sh
->batch_head
)
2275 for (i
= disks
; i
--; ) {
2276 struct r5dev
*dev
= &sh
->dev
[i
];
2277 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
2278 wake_up(&sh
->raid_conf
->wait_for_overlap
);
2283 static void free_stripe(struct kmem_cache
*sc
, struct stripe_head
*sh
)
2285 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2289 __free_page(sh
->ppl_page
);
2290 kmem_cache_free(sc
, sh
);
2293 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
,
2294 int disks
, struct r5conf
*conf
)
2296 struct stripe_head
*sh
;
2299 sh
= kmem_cache_zalloc(sc
, gfp
);
2301 spin_lock_init(&sh
->stripe_lock
);
2302 spin_lock_init(&sh
->batch_lock
);
2303 INIT_LIST_HEAD(&sh
->batch_list
);
2304 INIT_LIST_HEAD(&sh
->lru
);
2305 INIT_LIST_HEAD(&sh
->r5c
);
2306 INIT_LIST_HEAD(&sh
->log_list
);
2307 atomic_set(&sh
->count
, 1);
2308 sh
->raid_conf
= conf
;
2309 sh
->log_start
= MaxSector
;
2310 for (i
= 0; i
< disks
; i
++) {
2311 struct r5dev
*dev
= &sh
->dev
[i
];
2313 bio_init(&dev
->req
, &dev
->vec
, 1);
2314 bio_init(&dev
->rreq
, &dev
->rvec
, 1);
2317 if (raid5_has_ppl(conf
)) {
2318 sh
->ppl_page
= alloc_page(gfp
);
2319 if (!sh
->ppl_page
) {
2320 free_stripe(sc
, sh
);
2324 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2325 if (init_stripe_shared_pages(sh
, conf
, disks
)) {
2326 free_stripe(sc
, sh
);
2333 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2335 struct stripe_head
*sh
;
2337 sh
= alloc_stripe(conf
->slab_cache
, gfp
, conf
->pool_size
, conf
);
2341 if (grow_buffers(sh
, gfp
)) {
2343 free_stripe(conf
->slab_cache
, sh
);
2346 sh
->hash_lock_index
=
2347 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2348 /* we just created an active stripe so... */
2349 atomic_inc(&conf
->active_stripes
);
2351 raid5_release_stripe(sh
);
2352 conf
->max_nr_stripes
++;
2356 static int grow_stripes(struct r5conf
*conf
, int num
)
2358 struct kmem_cache
*sc
;
2359 size_t namelen
= sizeof(conf
->cache_name
[0]);
2360 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2362 if (conf
->mddev
->gendisk
)
2363 snprintf(conf
->cache_name
[0], namelen
,
2364 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2366 snprintf(conf
->cache_name
[0], namelen
,
2367 "raid%d-%p", conf
->level
, conf
->mddev
);
2368 snprintf(conf
->cache_name
[1], namelen
, "%.27s-alt", conf
->cache_name
[0]);
2370 conf
->active_name
= 0;
2371 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2372 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2376 conf
->slab_cache
= sc
;
2377 conf
->pool_size
= devs
;
2379 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2386 * scribble_alloc - allocate percpu scribble buffer for required size
2387 * of the scribble region
2388 * @percpu: from for_each_present_cpu() of the caller
2389 * @num: total number of disks in the array
2390 * @cnt: scribble objs count for required size of the scribble region
2392 * The scribble buffer size must be enough to contain:
2393 * 1/ a struct page pointer for each device in the array +2
2394 * 2/ room to convert each entry in (1) to its corresponding dma
2395 * (dma_map_page()) or page (page_address()) address.
2397 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2398 * calculate over all devices (not just the data blocks), using zeros in place
2399 * of the P and Q blocks.
2401 static int scribble_alloc(struct raid5_percpu
*percpu
,
2405 sizeof(struct page
*) * (num
+ 2) +
2406 sizeof(addr_conv_t
) * (num
+ 2) +
2407 sizeof(unsigned int) * (num
+ 2);
2411 * If here is in raid array suspend context, it is in memalloc noio
2412 * context as well, there is no potential recursive memory reclaim
2413 * I/Os with the GFP_KERNEL flag.
2415 scribble
= kvmalloc_array(cnt
, obj_size
, GFP_KERNEL
);
2419 kvfree(percpu
->scribble
);
2421 percpu
->scribble
= scribble
;
2422 percpu
->scribble_obj_size
= obj_size
;
2426 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2432 * Never shrink. And mddev_suspend() could deadlock if this is called
2433 * from raid5d. In that case, scribble_disks and scribble_sectors
2434 * should equal to new_disks and new_sectors
2436 if (conf
->scribble_disks
>= new_disks
&&
2437 conf
->scribble_sectors
>= new_sectors
)
2439 mddev_suspend(conf
->mddev
);
2442 for_each_present_cpu(cpu
) {
2443 struct raid5_percpu
*percpu
;
2445 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2446 err
= scribble_alloc(percpu
, new_disks
,
2447 new_sectors
/ RAID5_STRIPE_SECTORS(conf
));
2453 mddev_resume(conf
->mddev
);
2455 conf
->scribble_disks
= new_disks
;
2456 conf
->scribble_sectors
= new_sectors
;
2461 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2463 /* Make all the stripes able to hold 'newsize' devices.
2464 * New slots in each stripe get 'page' set to a new page.
2466 * This happens in stages:
2467 * 1/ create a new kmem_cache and allocate the required number of
2469 * 2/ gather all the old stripe_heads and transfer the pages across
2470 * to the new stripe_heads. This will have the side effect of
2471 * freezing the array as once all stripe_heads have been collected,
2472 * no IO will be possible. Old stripe heads are freed once their
2473 * pages have been transferred over, and the old kmem_cache is
2474 * freed when all stripes are done.
2475 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2476 * we simple return a failure status - no need to clean anything up.
2477 * 4/ allocate new pages for the new slots in the new stripe_heads.
2478 * If this fails, we don't bother trying the shrink the
2479 * stripe_heads down again, we just leave them as they are.
2480 * As each stripe_head is processed the new one is released into
2483 * Once step2 is started, we cannot afford to wait for a write,
2484 * so we use GFP_NOIO allocations.
2486 struct stripe_head
*osh
, *nsh
;
2487 LIST_HEAD(newstripes
);
2488 struct disk_info
*ndisks
;
2490 struct kmem_cache
*sc
;
2494 md_allow_write(conf
->mddev
);
2497 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2498 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2503 /* Need to ensure auto-resizing doesn't interfere */
2504 mutex_lock(&conf
->cache_size_mutex
);
2506 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2507 nsh
= alloc_stripe(sc
, GFP_KERNEL
, newsize
, conf
);
2511 list_add(&nsh
->lru
, &newstripes
);
2514 /* didn't get enough, give up */
2515 while (!list_empty(&newstripes
)) {
2516 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2517 list_del(&nsh
->lru
);
2518 free_stripe(sc
, nsh
);
2520 kmem_cache_destroy(sc
);
2521 mutex_unlock(&conf
->cache_size_mutex
);
2524 /* Step 2 - Must use GFP_NOIO now.
2525 * OK, we have enough stripes, start collecting inactive
2526 * stripes and copying them over
2530 list_for_each_entry(nsh
, &newstripes
, lru
) {
2531 lock_device_hash_lock(conf
, hash
);
2532 wait_event_cmd(conf
->wait_for_stripe
,
2533 !list_empty(conf
->inactive_list
+ hash
),
2534 unlock_device_hash_lock(conf
, hash
),
2535 lock_device_hash_lock(conf
, hash
));
2536 osh
= get_free_stripe(conf
, hash
);
2537 unlock_device_hash_lock(conf
, hash
);
2539 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2540 for (i
= 0; i
< osh
->nr_pages
; i
++) {
2541 nsh
->pages
[i
] = osh
->pages
[i
];
2542 osh
->pages
[i
] = NULL
;
2545 for(i
=0; i
<conf
->pool_size
; i
++) {
2546 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2547 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2548 nsh
->dev
[i
].offset
= osh
->dev
[i
].offset
;
2550 nsh
->hash_lock_index
= hash
;
2551 free_stripe(conf
->slab_cache
, osh
);
2553 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2554 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2559 kmem_cache_destroy(conf
->slab_cache
);
2562 * At this point, we are holding all the stripes so the array
2563 * is completely stalled, so now is a good time to resize
2564 * conf->disks and the scribble region
2566 ndisks
= kcalloc(newsize
, sizeof(struct disk_info
), GFP_NOIO
);
2568 for (i
= 0; i
< conf
->pool_size
; i
++)
2569 ndisks
[i
] = conf
->disks
[i
];
2571 for (i
= conf
->pool_size
; i
< newsize
; i
++) {
2572 ndisks
[i
].extra_page
= alloc_page(GFP_NOIO
);
2573 if (!ndisks
[i
].extra_page
)
2578 for (i
= conf
->pool_size
; i
< newsize
; i
++)
2579 if (ndisks
[i
].extra_page
)
2580 put_page(ndisks
[i
].extra_page
);
2584 conf
->disks
= ndisks
;
2589 conf
->slab_cache
= sc
;
2590 conf
->active_name
= 1-conf
->active_name
;
2592 /* Step 4, return new stripes to service */
2593 while(!list_empty(&newstripes
)) {
2594 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2595 list_del_init(&nsh
->lru
);
2597 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2598 for (i
= 0; i
< nsh
->nr_pages
; i
++) {
2601 nsh
->pages
[i
] = alloc_page(GFP_NOIO
);
2606 for (i
= conf
->raid_disks
; i
< newsize
; i
++) {
2607 if (nsh
->dev
[i
].page
)
2609 nsh
->dev
[i
].page
= raid5_get_dev_page(nsh
, i
);
2610 nsh
->dev
[i
].orig_page
= nsh
->dev
[i
].page
;
2611 nsh
->dev
[i
].offset
= raid5_get_page_offset(nsh
, i
);
2614 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2615 if (nsh
->dev
[i
].page
== NULL
) {
2616 struct page
*p
= alloc_page(GFP_NOIO
);
2617 nsh
->dev
[i
].page
= p
;
2618 nsh
->dev
[i
].orig_page
= p
;
2619 nsh
->dev
[i
].offset
= 0;
2624 raid5_release_stripe(nsh
);
2626 /* critical section pass, GFP_NOIO no longer needed */
2629 conf
->pool_size
= newsize
;
2630 mutex_unlock(&conf
->cache_size_mutex
);
2635 static int drop_one_stripe(struct r5conf
*conf
)
2637 struct stripe_head
*sh
;
2638 int hash
= (conf
->max_nr_stripes
- 1) & STRIPE_HASH_LOCKS_MASK
;
2640 spin_lock_irq(conf
->hash_locks
+ hash
);
2641 sh
= get_free_stripe(conf
, hash
);
2642 spin_unlock_irq(conf
->hash_locks
+ hash
);
2645 BUG_ON(atomic_read(&sh
->count
));
2647 free_stripe(conf
->slab_cache
, sh
);
2648 atomic_dec(&conf
->active_stripes
);
2649 conf
->max_nr_stripes
--;
2653 static void shrink_stripes(struct r5conf
*conf
)
2655 while (conf
->max_nr_stripes
&&
2656 drop_one_stripe(conf
))
2659 kmem_cache_destroy(conf
->slab_cache
);
2660 conf
->slab_cache
= NULL
;
2663 static void raid5_end_read_request(struct bio
* bi
)
2665 struct stripe_head
*sh
= bi
->bi_private
;
2666 struct r5conf
*conf
= sh
->raid_conf
;
2667 int disks
= sh
->disks
, i
;
2668 char b
[BDEVNAME_SIZE
];
2669 struct md_rdev
*rdev
= NULL
;
2672 for (i
=0 ; i
<disks
; i
++)
2673 if (bi
== &sh
->dev
[i
].req
)
2676 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2677 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2684 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2685 /* If replacement finished while this request was outstanding,
2686 * 'replacement' might be NULL already.
2687 * In that case it moved down to 'rdev'.
2688 * rdev is not removed until all requests are finished.
2690 rdev
= conf
->disks
[i
].replacement
;
2692 rdev
= conf
->disks
[i
].rdev
;
2694 if (use_new_offset(conf
, sh
))
2695 s
= sh
->sector
+ rdev
->new_data_offset
;
2697 s
= sh
->sector
+ rdev
->data_offset
;
2698 if (!bi
->bi_status
) {
2699 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2700 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2701 /* Note that this cannot happen on a
2702 * replacement device. We just fail those on
2705 pr_info_ratelimited(
2706 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2707 mdname(conf
->mddev
), RAID5_STRIPE_SECTORS(conf
),
2708 (unsigned long long)s
,
2709 bdevname(rdev
->bdev
, b
));
2710 atomic_add(RAID5_STRIPE_SECTORS(conf
), &rdev
->corrected_errors
);
2711 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2712 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2713 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2714 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2716 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
2718 * end read for a page in journal, this
2719 * must be preparing for prexor in rmw
2721 set_bit(R5_OrigPageUPTDODATE
, &sh
->dev
[i
].flags
);
2723 if (atomic_read(&rdev
->read_errors
))
2724 atomic_set(&rdev
->read_errors
, 0);
2726 const char *bdn
= bdevname(rdev
->bdev
, b
);
2730 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2731 if (!(bi
->bi_status
== BLK_STS_PROTECTION
))
2732 atomic_inc(&rdev
->read_errors
);
2733 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2734 pr_warn_ratelimited(
2735 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2736 mdname(conf
->mddev
),
2737 (unsigned long long)s
,
2739 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2741 pr_warn_ratelimited(
2742 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2743 mdname(conf
->mddev
),
2744 (unsigned long long)s
,
2746 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2749 pr_warn_ratelimited(
2750 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2751 mdname(conf
->mddev
),
2752 (unsigned long long)s
,
2754 } else if (atomic_read(&rdev
->read_errors
)
2755 > conf
->max_nr_stripes
) {
2756 if (!test_bit(Faulty
, &rdev
->flags
)) {
2757 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2758 mdname(conf
->mddev
),
2759 atomic_read(&rdev
->read_errors
),
2760 conf
->max_nr_stripes
);
2761 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2762 mdname(conf
->mddev
), bdn
);
2766 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2767 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2770 if (sh
->qd_idx
>= 0 && sh
->pd_idx
== i
)
2771 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2772 else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2773 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2774 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2776 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2778 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2779 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2781 && test_bit(In_sync
, &rdev
->flags
)
2782 && rdev_set_badblocks(
2783 rdev
, sh
->sector
, RAID5_STRIPE_SECTORS(conf
), 0)))
2784 md_error(conf
->mddev
, rdev
);
2787 rdev_dec_pending(rdev
, conf
->mddev
);
2789 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2790 set_bit(STRIPE_HANDLE
, &sh
->state
);
2791 raid5_release_stripe(sh
);
2794 static void raid5_end_write_request(struct bio
*bi
)
2796 struct stripe_head
*sh
= bi
->bi_private
;
2797 struct r5conf
*conf
= sh
->raid_conf
;
2798 int disks
= sh
->disks
, i
;
2799 struct md_rdev
*rdev
;
2802 int replacement
= 0;
2804 for (i
= 0 ; i
< disks
; i
++) {
2805 if (bi
== &sh
->dev
[i
].req
) {
2806 rdev
= conf
->disks
[i
].rdev
;
2809 if (bi
== &sh
->dev
[i
].rreq
) {
2810 rdev
= conf
->disks
[i
].replacement
;
2814 /* rdev was removed and 'replacement'
2815 * replaced it. rdev is not removed
2816 * until all requests are finished.
2818 rdev
= conf
->disks
[i
].rdev
;
2822 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2823 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2833 md_error(conf
->mddev
, rdev
);
2834 else if (is_badblock(rdev
, sh
->sector
,
2835 RAID5_STRIPE_SECTORS(conf
),
2836 &first_bad
, &bad_sectors
))
2837 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2839 if (bi
->bi_status
) {
2840 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2841 set_bit(WriteErrorSeen
, &rdev
->flags
);
2842 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2843 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2844 set_bit(MD_RECOVERY_NEEDED
,
2845 &rdev
->mddev
->recovery
);
2846 } else if (is_badblock(rdev
, sh
->sector
,
2847 RAID5_STRIPE_SECTORS(conf
),
2848 &first_bad
, &bad_sectors
)) {
2849 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2850 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2851 /* That was a successful write so make
2852 * sure it looks like we already did
2855 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2858 rdev_dec_pending(rdev
, conf
->mddev
);
2860 if (sh
->batch_head
&& bi
->bi_status
&& !replacement
)
2861 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2864 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2865 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2866 set_bit(STRIPE_HANDLE
, &sh
->state
);
2867 raid5_release_stripe(sh
);
2869 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2870 raid5_release_stripe(sh
->batch_head
);
2873 static void raid5_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2875 char b
[BDEVNAME_SIZE
];
2876 struct r5conf
*conf
= mddev
->private;
2877 unsigned long flags
;
2878 pr_debug("raid456: error called\n");
2880 spin_lock_irqsave(&conf
->device_lock
, flags
);
2882 if (test_bit(In_sync
, &rdev
->flags
) &&
2883 mddev
->degraded
== conf
->max_degraded
) {
2885 * Don't allow to achieve failed state
2886 * Don't try to recover this device
2888 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2889 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2893 set_bit(Faulty
, &rdev
->flags
);
2894 clear_bit(In_sync
, &rdev
->flags
);
2895 mddev
->degraded
= raid5_calc_degraded(conf
);
2896 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2897 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2899 set_bit(Blocked
, &rdev
->flags
);
2900 set_mask_bits(&mddev
->sb_flags
, 0,
2901 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
2902 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2903 "md/raid:%s: Operation continuing on %d devices.\n",
2905 bdevname(rdev
->bdev
, b
),
2907 conf
->raid_disks
- mddev
->degraded
);
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
->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 WARN_ON_ONCE(bio
->bi_bdev
->bd_partno
);
5316 chunk_sectors
= min(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5317 return chunk_sectors
>=
5318 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
5322 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5323 * later sampled by raid5d.
5325 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
5327 unsigned long flags
;
5329 spin_lock_irqsave(&conf
->device_lock
, flags
);
5331 bi
->bi_next
= conf
->retry_read_aligned_list
;
5332 conf
->retry_read_aligned_list
= bi
;
5334 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5335 md_wakeup_thread(conf
->mddev
->thread
);
5338 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
,
5339 unsigned int *offset
)
5343 bi
= conf
->retry_read_aligned
;
5345 *offset
= conf
->retry_read_offset
;
5346 conf
->retry_read_aligned
= NULL
;
5349 bi
= conf
->retry_read_aligned_list
;
5351 conf
->retry_read_aligned_list
= bi
->bi_next
;
5360 * The "raid5_align_endio" should check if the read succeeded and if it
5361 * did, call bio_endio on the original bio (having bio_put the new bio
5363 * If the read failed..
5365 static void raid5_align_endio(struct bio
*bi
)
5367 struct bio
* raid_bi
= bi
->bi_private
;
5368 struct mddev
*mddev
;
5369 struct r5conf
*conf
;
5370 struct md_rdev
*rdev
;
5371 blk_status_t error
= bi
->bi_status
;
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
);
5384 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5385 wake_up(&conf
->wait_for_quiescent
);
5389 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5391 add_bio_to_retry(raid_bi
, conf
);
5394 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
5396 struct r5conf
*conf
= mddev
->private;
5397 struct bio
*align_bio
;
5398 struct md_rdev
*rdev
;
5399 sector_t sector
, end_sector
, first_bad
;
5400 int bad_sectors
, dd_idx
;
5402 if (!in_chunk_boundary(mddev
, raid_bio
)) {
5403 pr_debug("%s: non aligned\n", __func__
);
5407 sector
= raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
, 0,
5409 end_sector
= bio_end_sector(raid_bio
);
5412 if (r5c_big_stripe_cached(conf
, sector
))
5413 goto out_rcu_unlock
;
5415 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
5416 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
5417 rdev
->recovery_offset
< end_sector
) {
5418 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
5420 goto out_rcu_unlock
;
5421 if (test_bit(Faulty
, &rdev
->flags
) ||
5422 !(test_bit(In_sync
, &rdev
->flags
) ||
5423 rdev
->recovery_offset
>= end_sector
))
5424 goto out_rcu_unlock
;
5427 atomic_inc(&rdev
->nr_pending
);
5430 align_bio
= bio_clone_fast(raid_bio
, GFP_NOIO
, &mddev
->bio_set
);
5431 bio_set_dev(align_bio
, rdev
->bdev
);
5432 align_bio
->bi_end_io
= raid5_align_endio
;
5433 align_bio
->bi_private
= raid_bio
;
5434 align_bio
->bi_iter
.bi_sector
= sector
;
5436 raid_bio
->bi_next
= (void *)rdev
;
5438 if (is_badblock(rdev
, sector
, bio_sectors(align_bio
), &first_bad
,
5441 rdev_dec_pending(rdev
, mddev
);
5445 /* No reshape active, so we can trust rdev->data_offset */
5446 align_bio
->bi_iter
.bi_sector
+= rdev
->data_offset
;
5448 spin_lock_irq(&conf
->device_lock
);
5449 wait_event_lock_irq(conf
->wait_for_quiescent
, conf
->quiesce
== 0,
5451 atomic_inc(&conf
->active_aligned_reads
);
5452 spin_unlock_irq(&conf
->device_lock
);
5455 trace_block_bio_remap(align_bio
, disk_devt(mddev
->gendisk
),
5456 raid_bio
->bi_iter
.bi_sector
);
5457 submit_bio_noacct(align_bio
);
5465 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
5468 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
5469 unsigned chunk_sects
= mddev
->chunk_sectors
;
5470 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
5472 if (sectors
< bio_sectors(raid_bio
)) {
5473 struct r5conf
*conf
= mddev
->private;
5474 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, &conf
->bio_split
);
5475 bio_chain(split
, raid_bio
);
5476 submit_bio_noacct(raid_bio
);
5480 if (!raid5_read_one_chunk(mddev
, raid_bio
))
5486 /* __get_priority_stripe - get the next stripe to process
5488 * Full stripe writes are allowed to pass preread active stripes up until
5489 * the bypass_threshold is exceeded. In general the bypass_count
5490 * increments when the handle_list is handled before the hold_list; however, it
5491 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5492 * stripe with in flight i/o. The bypass_count will be reset when the
5493 * head of the hold_list has changed, i.e. the head was promoted to the
5496 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
5498 struct stripe_head
*sh
, *tmp
;
5499 struct list_head
*handle_list
= NULL
;
5500 struct r5worker_group
*wg
;
5501 bool second_try
= !r5c_is_writeback(conf
->log
) &&
5502 !r5l_log_disk_error(conf
);
5503 bool try_loprio
= test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
) ||
5504 r5l_log_disk_error(conf
);
5509 if (conf
->worker_cnt_per_group
== 0) {
5510 handle_list
= try_loprio
? &conf
->loprio_list
:
5512 } else if (group
!= ANY_GROUP
) {
5513 handle_list
= try_loprio
? &conf
->worker_groups
[group
].loprio_list
:
5514 &conf
->worker_groups
[group
].handle_list
;
5515 wg
= &conf
->worker_groups
[group
];
5518 for (i
= 0; i
< conf
->group_cnt
; i
++) {
5519 handle_list
= try_loprio
? &conf
->worker_groups
[i
].loprio_list
:
5520 &conf
->worker_groups
[i
].handle_list
;
5521 wg
= &conf
->worker_groups
[i
];
5522 if (!list_empty(handle_list
))
5527 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5529 list_empty(handle_list
) ? "empty" : "busy",
5530 list_empty(&conf
->hold_list
) ? "empty" : "busy",
5531 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
5533 if (!list_empty(handle_list
)) {
5534 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
5536 if (list_empty(&conf
->hold_list
))
5537 conf
->bypass_count
= 0;
5538 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
5539 if (conf
->hold_list
.next
== conf
->last_hold
)
5540 conf
->bypass_count
++;
5542 conf
->last_hold
= conf
->hold_list
.next
;
5543 conf
->bypass_count
-= conf
->bypass_threshold
;
5544 if (conf
->bypass_count
< 0)
5545 conf
->bypass_count
= 0;
5548 } else if (!list_empty(&conf
->hold_list
) &&
5549 ((conf
->bypass_threshold
&&
5550 conf
->bypass_count
> conf
->bypass_threshold
) ||
5551 atomic_read(&conf
->pending_full_writes
) == 0)) {
5553 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
5554 if (conf
->worker_cnt_per_group
== 0 ||
5555 group
== ANY_GROUP
||
5556 !cpu_online(tmp
->cpu
) ||
5557 cpu_to_group(tmp
->cpu
) == group
) {
5564 conf
->bypass_count
-= conf
->bypass_threshold
;
5565 if (conf
->bypass_count
< 0)
5566 conf
->bypass_count
= 0;
5575 try_loprio
= !try_loprio
;
5583 list_del_init(&sh
->lru
);
5584 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
5588 struct raid5_plug_cb
{
5589 struct blk_plug_cb cb
;
5590 struct list_head list
;
5591 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
5594 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
5596 struct raid5_plug_cb
*cb
= container_of(
5597 blk_cb
, struct raid5_plug_cb
, cb
);
5598 struct stripe_head
*sh
;
5599 struct mddev
*mddev
= cb
->cb
.data
;
5600 struct r5conf
*conf
= mddev
->private;
5604 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
5605 spin_lock_irq(&conf
->device_lock
);
5606 while (!list_empty(&cb
->list
)) {
5607 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
5608 list_del_init(&sh
->lru
);
5610 * avoid race release_stripe_plug() sees
5611 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5612 * is still in our list
5614 smp_mb__before_atomic();
5615 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
5617 * STRIPE_ON_RELEASE_LIST could be set here. In that
5618 * case, the count is always > 1 here
5620 hash
= sh
->hash_lock_index
;
5621 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
5624 spin_unlock_irq(&conf
->device_lock
);
5626 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
5627 NR_STRIPE_HASH_LOCKS
);
5629 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5633 static void release_stripe_plug(struct mddev
*mddev
,
5634 struct stripe_head
*sh
)
5636 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5637 raid5_unplug
, mddev
,
5638 sizeof(struct raid5_plug_cb
));
5639 struct raid5_plug_cb
*cb
;
5642 raid5_release_stripe(sh
);
5646 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5648 if (cb
->list
.next
== NULL
) {
5650 INIT_LIST_HEAD(&cb
->list
);
5651 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5652 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5655 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5656 list_add_tail(&sh
->lru
, &cb
->list
);
5658 raid5_release_stripe(sh
);
5661 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5663 struct r5conf
*conf
= mddev
->private;
5664 sector_t logical_sector
, last_sector
;
5665 struct stripe_head
*sh
;
5668 if (mddev
->reshape_position
!= MaxSector
)
5669 /* Skip discard while reshape is happening */
5672 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)RAID5_STRIPE_SECTORS(conf
)-1);
5673 last_sector
= bio_end_sector(bi
);
5677 stripe_sectors
= conf
->chunk_sectors
*
5678 (conf
->raid_disks
- conf
->max_degraded
);
5679 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5681 sector_div(last_sector
, stripe_sectors
);
5683 logical_sector
*= conf
->chunk_sectors
;
5684 last_sector
*= conf
->chunk_sectors
;
5686 for (; logical_sector
< last_sector
;
5687 logical_sector
+= RAID5_STRIPE_SECTORS(conf
)) {
5691 sh
= raid5_get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5692 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5693 TASK_UNINTERRUPTIBLE
);
5694 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5695 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5696 raid5_release_stripe(sh
);
5700 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5701 spin_lock_irq(&sh
->stripe_lock
);
5702 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5703 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5705 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5706 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5707 spin_unlock_irq(&sh
->stripe_lock
);
5708 raid5_release_stripe(sh
);
5713 set_bit(STRIPE_DISCARD
, &sh
->state
);
5714 finish_wait(&conf
->wait_for_overlap
, &w
);
5715 sh
->overwrite_disks
= 0;
5716 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5717 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5719 sh
->dev
[d
].towrite
= bi
;
5720 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5721 bio_inc_remaining(bi
);
5722 md_write_inc(mddev
, bi
);
5723 sh
->overwrite_disks
++;
5725 spin_unlock_irq(&sh
->stripe_lock
);
5726 if (conf
->mddev
->bitmap
) {
5728 d
< conf
->raid_disks
- conf
->max_degraded
;
5730 md_bitmap_startwrite(mddev
->bitmap
,
5732 RAID5_STRIPE_SECTORS(conf
),
5734 sh
->bm_seq
= conf
->seq_flush
+ 1;
5735 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5738 set_bit(STRIPE_HANDLE
, &sh
->state
);
5739 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5740 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5741 atomic_inc(&conf
->preread_active_stripes
);
5742 release_stripe_plug(mddev
, sh
);
5748 static bool raid5_make_request(struct mddev
*mddev
, struct bio
* bi
)
5750 struct r5conf
*conf
= mddev
->private;
5752 sector_t new_sector
;
5753 sector_t logical_sector
, last_sector
;
5754 struct stripe_head
*sh
;
5755 const int rw
= bio_data_dir(bi
);
5758 bool do_flush
= false;
5760 if (unlikely(bi
->bi_opf
& REQ_PREFLUSH
)) {
5761 int ret
= log_handle_flush_request(conf
, bi
);
5765 if (ret
== -ENODEV
) {
5766 if (md_flush_request(mddev
, bi
))
5769 /* ret == -EAGAIN, fallback */
5771 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5772 * we need to flush journal device
5774 do_flush
= bi
->bi_opf
& REQ_PREFLUSH
;
5777 if (!md_write_start(mddev
, bi
))
5780 * If array is degraded, better not do chunk aligned read because
5781 * later we might have to read it again in order to reconstruct
5782 * data on failed drives.
5784 if (rw
== READ
&& mddev
->degraded
== 0 &&
5785 mddev
->reshape_position
== MaxSector
) {
5786 bi
= chunk_aligned_read(mddev
, bi
);
5791 if (unlikely(bio_op(bi
) == REQ_OP_DISCARD
)) {
5792 make_discard_request(mddev
, bi
);
5793 md_write_end(mddev
);
5797 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)RAID5_STRIPE_SECTORS(conf
)-1);
5798 last_sector
= bio_end_sector(bi
);
5801 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5802 for (; logical_sector
< last_sector
; logical_sector
+= RAID5_STRIPE_SECTORS(conf
)) {
5808 seq
= read_seqcount_begin(&conf
->gen_lock
);
5811 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5812 TASK_UNINTERRUPTIBLE
);
5813 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5814 /* spinlock is needed as reshape_progress may be
5815 * 64bit on a 32bit platform, and so it might be
5816 * possible to see a half-updated value
5817 * Of course reshape_progress could change after
5818 * the lock is dropped, so once we get a reference
5819 * to the stripe that we think it is, we will have
5822 spin_lock_irq(&conf
->device_lock
);
5823 if (mddev
->reshape_backwards
5824 ? logical_sector
< conf
->reshape_progress
5825 : logical_sector
>= conf
->reshape_progress
) {
5828 if (mddev
->reshape_backwards
5829 ? logical_sector
< conf
->reshape_safe
5830 : logical_sector
>= conf
->reshape_safe
) {
5831 spin_unlock_irq(&conf
->device_lock
);
5837 spin_unlock_irq(&conf
->device_lock
);
5840 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5843 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5844 (unsigned long long)new_sector
,
5845 (unsigned long long)logical_sector
);
5847 sh
= raid5_get_active_stripe(conf
, new_sector
, previous
,
5848 (bi
->bi_opf
& REQ_RAHEAD
), 0);
5850 if (unlikely(previous
)) {
5851 /* expansion might have moved on while waiting for a
5852 * stripe, so we must do the range check again.
5853 * Expansion could still move past after this
5854 * test, but as we are holding a reference to
5855 * 'sh', we know that if that happens,
5856 * STRIPE_EXPANDING will get set and the expansion
5857 * won't proceed until we finish with the stripe.
5860 spin_lock_irq(&conf
->device_lock
);
5861 if (mddev
->reshape_backwards
5862 ? logical_sector
>= conf
->reshape_progress
5863 : logical_sector
< conf
->reshape_progress
)
5864 /* mismatch, need to try again */
5866 spin_unlock_irq(&conf
->device_lock
);
5868 raid5_release_stripe(sh
);
5874 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5875 /* Might have got the wrong stripe_head
5878 raid5_release_stripe(sh
);
5882 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5883 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5884 /* Stripe is busy expanding or
5885 * add failed due to overlap. Flush everything
5888 md_wakeup_thread(mddev
->thread
);
5889 raid5_release_stripe(sh
);
5895 set_bit(STRIPE_R5C_PREFLUSH
, &sh
->state
);
5896 /* we only need flush for one stripe */
5900 set_bit(STRIPE_HANDLE
, &sh
->state
);
5901 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5902 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5903 (bi
->bi_opf
& REQ_SYNC
) &&
5904 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5905 atomic_inc(&conf
->preread_active_stripes
);
5906 release_stripe_plug(mddev
, sh
);
5908 /* cannot get stripe for read-ahead, just give-up */
5909 bi
->bi_status
= BLK_STS_IOERR
;
5913 finish_wait(&conf
->wait_for_overlap
, &w
);
5916 md_write_end(mddev
);
5921 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5923 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5925 /* reshaping is quite different to recovery/resync so it is
5926 * handled quite separately ... here.
5928 * On each call to sync_request, we gather one chunk worth of
5929 * destination stripes and flag them as expanding.
5930 * Then we find all the source stripes and request reads.
5931 * As the reads complete, handle_stripe will copy the data
5932 * into the destination stripe and release that stripe.
5934 struct r5conf
*conf
= mddev
->private;
5935 struct stripe_head
*sh
;
5936 struct md_rdev
*rdev
;
5937 sector_t first_sector
, last_sector
;
5938 int raid_disks
= conf
->previous_raid_disks
;
5939 int data_disks
= raid_disks
- conf
->max_degraded
;
5940 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5943 sector_t writepos
, readpos
, safepos
;
5944 sector_t stripe_addr
;
5945 int reshape_sectors
;
5946 struct list_head stripes
;
5949 if (sector_nr
== 0) {
5950 /* If restarting in the middle, skip the initial sectors */
5951 if (mddev
->reshape_backwards
&&
5952 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5953 sector_nr
= raid5_size(mddev
, 0, 0)
5954 - conf
->reshape_progress
;
5955 } else if (mddev
->reshape_backwards
&&
5956 conf
->reshape_progress
== MaxSector
) {
5957 /* shouldn't happen, but just in case, finish up.*/
5958 sector_nr
= MaxSector
;
5959 } else if (!mddev
->reshape_backwards
&&
5960 conf
->reshape_progress
> 0)
5961 sector_nr
= conf
->reshape_progress
;
5962 sector_div(sector_nr
, new_data_disks
);
5964 mddev
->curr_resync_completed
= sector_nr
;
5965 sysfs_notify_dirent_safe(mddev
->sysfs_completed
);
5972 /* We need to process a full chunk at a time.
5973 * If old and new chunk sizes differ, we need to process the
5977 reshape_sectors
= max(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5979 /* We update the metadata at least every 10 seconds, or when
5980 * the data about to be copied would over-write the source of
5981 * the data at the front of the range. i.e. one new_stripe
5982 * along from reshape_progress new_maps to after where
5983 * reshape_safe old_maps to
5985 writepos
= conf
->reshape_progress
;
5986 sector_div(writepos
, new_data_disks
);
5987 readpos
= conf
->reshape_progress
;
5988 sector_div(readpos
, data_disks
);
5989 safepos
= conf
->reshape_safe
;
5990 sector_div(safepos
, data_disks
);
5991 if (mddev
->reshape_backwards
) {
5992 BUG_ON(writepos
< reshape_sectors
);
5993 writepos
-= reshape_sectors
;
5994 readpos
+= reshape_sectors
;
5995 safepos
+= reshape_sectors
;
5997 writepos
+= reshape_sectors
;
5998 /* readpos and safepos are worst-case calculations.
5999 * A negative number is overly pessimistic, and causes
6000 * obvious problems for unsigned storage. So clip to 0.
6002 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
6003 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
6006 /* Having calculated the 'writepos' possibly use it
6007 * to set 'stripe_addr' which is where we will write to.
6009 if (mddev
->reshape_backwards
) {
6010 BUG_ON(conf
->reshape_progress
== 0);
6011 stripe_addr
= writepos
;
6012 BUG_ON((mddev
->dev_sectors
&
6013 ~((sector_t
)reshape_sectors
- 1))
6014 - reshape_sectors
- stripe_addr
6017 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
6018 stripe_addr
= sector_nr
;
6021 /* 'writepos' is the most advanced device address we might write.
6022 * 'readpos' is the least advanced device address we might read.
6023 * 'safepos' is the least address recorded in the metadata as having
6025 * If there is a min_offset_diff, these are adjusted either by
6026 * increasing the safepos/readpos if diff is negative, or
6027 * increasing writepos if diff is positive.
6028 * If 'readpos' is then behind 'writepos', there is no way that we can
6029 * ensure safety in the face of a crash - that must be done by userspace
6030 * making a backup of the data. So in that case there is no particular
6031 * rush to update metadata.
6032 * Otherwise if 'safepos' is behind 'writepos', then we really need to
6033 * update the metadata to advance 'safepos' to match 'readpos' so that
6034 * we can be safe in the event of a crash.
6035 * So we insist on updating metadata if safepos is behind writepos and
6036 * readpos is beyond writepos.
6037 * In any case, update the metadata every 10 seconds.
6038 * Maybe that number should be configurable, but I'm not sure it is
6039 * worth it.... maybe it could be a multiple of safemode_delay???
6041 if (conf
->min_offset_diff
< 0) {
6042 safepos
+= -conf
->min_offset_diff
;
6043 readpos
+= -conf
->min_offset_diff
;
6045 writepos
+= conf
->min_offset_diff
;
6047 if ((mddev
->reshape_backwards
6048 ? (safepos
> writepos
&& readpos
< writepos
)
6049 : (safepos
< writepos
&& readpos
> writepos
)) ||
6050 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
6051 /* Cannot proceed until we've updated the superblock... */
6052 wait_event(conf
->wait_for_overlap
,
6053 atomic_read(&conf
->reshape_stripes
)==0
6054 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
6055 if (atomic_read(&conf
->reshape_stripes
) != 0)
6057 mddev
->reshape_position
= conf
->reshape_progress
;
6058 mddev
->curr_resync_completed
= sector_nr
;
6059 if (!mddev
->reshape_backwards
)
6060 /* Can update recovery_offset */
6061 rdev_for_each(rdev
, mddev
)
6062 if (rdev
->raid_disk
>= 0 &&
6063 !test_bit(Journal
, &rdev
->flags
) &&
6064 !test_bit(In_sync
, &rdev
->flags
) &&
6065 rdev
->recovery_offset
< sector_nr
)
6066 rdev
->recovery_offset
= sector_nr
;
6068 conf
->reshape_checkpoint
= jiffies
;
6069 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
6070 md_wakeup_thread(mddev
->thread
);
6071 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
6072 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
6073 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
6075 spin_lock_irq(&conf
->device_lock
);
6076 conf
->reshape_safe
= mddev
->reshape_position
;
6077 spin_unlock_irq(&conf
->device_lock
);
6078 wake_up(&conf
->wait_for_overlap
);
6079 sysfs_notify_dirent_safe(mddev
->sysfs_completed
);
6082 INIT_LIST_HEAD(&stripes
);
6083 for (i
= 0; i
< reshape_sectors
; i
+= RAID5_STRIPE_SECTORS(conf
)) {
6085 int skipped_disk
= 0;
6086 sh
= raid5_get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
6087 set_bit(STRIPE_EXPANDING
, &sh
->state
);
6088 atomic_inc(&conf
->reshape_stripes
);
6089 /* If any of this stripe is beyond the end of the old
6090 * array, then we need to zero those blocks
6092 for (j
=sh
->disks
; j
--;) {
6094 if (j
== sh
->pd_idx
)
6096 if (conf
->level
== 6 &&
6099 s
= raid5_compute_blocknr(sh
, j
, 0);
6100 if (s
< raid5_size(mddev
, 0, 0)) {
6104 memset(page_address(sh
->dev
[j
].page
), 0, RAID5_STRIPE_SIZE(conf
));
6105 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
6106 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
6108 if (!skipped_disk
) {
6109 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
6110 set_bit(STRIPE_HANDLE
, &sh
->state
);
6112 list_add(&sh
->lru
, &stripes
);
6114 spin_lock_irq(&conf
->device_lock
);
6115 if (mddev
->reshape_backwards
)
6116 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
6118 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
6119 spin_unlock_irq(&conf
->device_lock
);
6120 /* Ok, those stripe are ready. We can start scheduling
6121 * reads on the source stripes.
6122 * The source stripes are determined by mapping the first and last
6123 * block on the destination stripes.
6126 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
6129 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
6130 * new_data_disks
- 1),
6132 if (last_sector
>= mddev
->dev_sectors
)
6133 last_sector
= mddev
->dev_sectors
- 1;
6134 while (first_sector
<= last_sector
) {
6135 sh
= raid5_get_active_stripe(conf
, first_sector
, 1, 0, 1);
6136 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
6137 set_bit(STRIPE_HANDLE
, &sh
->state
);
6138 raid5_release_stripe(sh
);
6139 first_sector
+= RAID5_STRIPE_SECTORS(conf
);
6141 /* Now that the sources are clearly marked, we can release
6142 * the destination stripes
6144 while (!list_empty(&stripes
)) {
6145 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
6146 list_del_init(&sh
->lru
);
6147 raid5_release_stripe(sh
);
6149 /* If this takes us to the resync_max point where we have to pause,
6150 * then we need to write out the superblock.
6152 sector_nr
+= reshape_sectors
;
6153 retn
= reshape_sectors
;
6155 if (mddev
->curr_resync_completed
> mddev
->resync_max
||
6156 (sector_nr
- mddev
->curr_resync_completed
) * 2
6157 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
6158 /* Cannot proceed until we've updated the superblock... */
6159 wait_event(conf
->wait_for_overlap
,
6160 atomic_read(&conf
->reshape_stripes
) == 0
6161 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
6162 if (atomic_read(&conf
->reshape_stripes
) != 0)
6164 mddev
->reshape_position
= conf
->reshape_progress
;
6165 mddev
->curr_resync_completed
= sector_nr
;
6166 if (!mddev
->reshape_backwards
)
6167 /* Can update recovery_offset */
6168 rdev_for_each(rdev
, mddev
)
6169 if (rdev
->raid_disk
>= 0 &&
6170 !test_bit(Journal
, &rdev
->flags
) &&
6171 !test_bit(In_sync
, &rdev
->flags
) &&
6172 rdev
->recovery_offset
< sector_nr
)
6173 rdev
->recovery_offset
= sector_nr
;
6174 conf
->reshape_checkpoint
= jiffies
;
6175 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
6176 md_wakeup_thread(mddev
->thread
);
6177 wait_event(mddev
->sb_wait
,
6178 !test_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
)
6179 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
6180 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
6182 spin_lock_irq(&conf
->device_lock
);
6183 conf
->reshape_safe
= mddev
->reshape_position
;
6184 spin_unlock_irq(&conf
->device_lock
);
6185 wake_up(&conf
->wait_for_overlap
);
6186 sysfs_notify_dirent_safe(mddev
->sysfs_completed
);
6192 static inline sector_t
raid5_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
6195 struct r5conf
*conf
= mddev
->private;
6196 struct stripe_head
*sh
;
6197 sector_t max_sector
= mddev
->dev_sectors
;
6198 sector_t sync_blocks
;
6199 int still_degraded
= 0;
6202 if (sector_nr
>= max_sector
) {
6203 /* just being told to finish up .. nothing much to do */
6205 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
6210 if (mddev
->curr_resync
< max_sector
) /* aborted */
6211 md_bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
6213 else /* completed sync */
6215 md_bitmap_close_sync(mddev
->bitmap
);
6220 /* Allow raid5_quiesce to complete */
6221 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
6223 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
6224 return reshape_request(mddev
, sector_nr
, skipped
);
6226 /* No need to check resync_max as we never do more than one
6227 * stripe, and as resync_max will always be on a chunk boundary,
6228 * if the check in md_do_sync didn't fire, there is no chance
6229 * of overstepping resync_max here
6232 /* if there is too many failed drives and we are trying
6233 * to resync, then assert that we are finished, because there is
6234 * nothing we can do.
6236 if (mddev
->degraded
>= conf
->max_degraded
&&
6237 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
6238 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
6242 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
6244 !md_bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
6245 sync_blocks
>= RAID5_STRIPE_SECTORS(conf
)) {
6246 /* we can skip this block, and probably more */
6247 do_div(sync_blocks
, RAID5_STRIPE_SECTORS(conf
));
6249 /* keep things rounded to whole stripes */
6250 return sync_blocks
* RAID5_STRIPE_SECTORS(conf
);
6253 md_bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, false);
6255 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 1, 0);
6257 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 0, 0);
6258 /* make sure we don't swamp the stripe cache if someone else
6259 * is trying to get access
6261 schedule_timeout_uninterruptible(1);
6263 /* Need to check if array will still be degraded after recovery/resync
6264 * Note in case of > 1 drive failures it's possible we're rebuilding
6265 * one drive while leaving another faulty drive in array.
6268 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6269 struct md_rdev
*rdev
= READ_ONCE(conf
->disks
[i
].rdev
);
6271 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
6276 md_bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
6278 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
6279 set_bit(STRIPE_HANDLE
, &sh
->state
);
6281 raid5_release_stripe(sh
);
6283 return RAID5_STRIPE_SECTORS(conf
);
6286 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
,
6287 unsigned int offset
)
6289 /* We may not be able to submit a whole bio at once as there
6290 * may not be enough stripe_heads available.
6291 * We cannot pre-allocate enough stripe_heads as we may need
6292 * more than exist in the cache (if we allow ever large chunks).
6293 * So we do one stripe head at a time and record in
6294 * ->bi_hw_segments how many have been done.
6296 * We *know* that this entire raid_bio is in one chunk, so
6297 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6299 struct stripe_head
*sh
;
6301 sector_t sector
, logical_sector
, last_sector
;
6305 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
6306 ~((sector_t
)RAID5_STRIPE_SECTORS(conf
)-1);
6307 sector
= raid5_compute_sector(conf
, logical_sector
,
6309 last_sector
= bio_end_sector(raid_bio
);
6311 for (; logical_sector
< last_sector
;
6312 logical_sector
+= RAID5_STRIPE_SECTORS(conf
),
6313 sector
+= RAID5_STRIPE_SECTORS(conf
),
6317 /* already done this stripe */
6320 sh
= raid5_get_active_stripe(conf
, sector
, 0, 1, 1);
6323 /* failed to get a stripe - must wait */
6324 conf
->retry_read_aligned
= raid_bio
;
6325 conf
->retry_read_offset
= scnt
;
6329 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
6330 raid5_release_stripe(sh
);
6331 conf
->retry_read_aligned
= raid_bio
;
6332 conf
->retry_read_offset
= scnt
;
6336 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
6338 raid5_release_stripe(sh
);
6342 bio_endio(raid_bio
);
6344 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
6345 wake_up(&conf
->wait_for_quiescent
);
6349 static int handle_active_stripes(struct r5conf
*conf
, int group
,
6350 struct r5worker
*worker
,
6351 struct list_head
*temp_inactive_list
)
6352 __releases(&conf
->device_lock
)
6353 __acquires(&conf
->device_lock
)
6355 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
6356 int i
, batch_size
= 0, hash
;
6357 bool release_inactive
= false;
6359 while (batch_size
< MAX_STRIPE_BATCH
&&
6360 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
6361 batch
[batch_size
++] = sh
;
6363 if (batch_size
== 0) {
6364 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6365 if (!list_empty(temp_inactive_list
+ i
))
6367 if (i
== NR_STRIPE_HASH_LOCKS
) {
6368 spin_unlock_irq(&conf
->device_lock
);
6369 log_flush_stripe_to_raid(conf
);
6370 spin_lock_irq(&conf
->device_lock
);
6373 release_inactive
= true;
6375 spin_unlock_irq(&conf
->device_lock
);
6377 release_inactive_stripe_list(conf
, temp_inactive_list
,
6378 NR_STRIPE_HASH_LOCKS
);
6380 r5l_flush_stripe_to_raid(conf
->log
);
6381 if (release_inactive
) {
6382 spin_lock_irq(&conf
->device_lock
);
6386 for (i
= 0; i
< batch_size
; i
++)
6387 handle_stripe(batch
[i
]);
6388 log_write_stripe_run(conf
);
6392 spin_lock_irq(&conf
->device_lock
);
6393 for (i
= 0; i
< batch_size
; i
++) {
6394 hash
= batch
[i
]->hash_lock_index
;
6395 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
6400 static void raid5_do_work(struct work_struct
*work
)
6402 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
6403 struct r5worker_group
*group
= worker
->group
;
6404 struct r5conf
*conf
= group
->conf
;
6405 struct mddev
*mddev
= conf
->mddev
;
6406 int group_id
= group
- conf
->worker_groups
;
6408 struct blk_plug plug
;
6410 pr_debug("+++ raid5worker active\n");
6412 blk_start_plug(&plug
);
6414 spin_lock_irq(&conf
->device_lock
);
6416 int batch_size
, released
;
6418 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
6420 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
6421 worker
->temp_inactive_list
);
6422 worker
->working
= false;
6423 if (!batch_size
&& !released
)
6425 handled
+= batch_size
;
6426 wait_event_lock_irq(mddev
->sb_wait
,
6427 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
),
6430 pr_debug("%d stripes handled\n", handled
);
6432 spin_unlock_irq(&conf
->device_lock
);
6434 flush_deferred_bios(conf
);
6436 r5l_flush_stripe_to_raid(conf
->log
);
6438 async_tx_issue_pending_all();
6439 blk_finish_plug(&plug
);
6441 pr_debug("--- raid5worker inactive\n");
6445 * This is our raid5 kernel thread.
6447 * We scan the hash table for stripes which can be handled now.
6448 * During the scan, completed stripes are saved for us by the interrupt
6449 * handler, so that they will not have to wait for our next wakeup.
6451 static void raid5d(struct md_thread
*thread
)
6453 struct mddev
*mddev
= thread
->mddev
;
6454 struct r5conf
*conf
= mddev
->private;
6456 struct blk_plug plug
;
6458 pr_debug("+++ raid5d active\n");
6460 md_check_recovery(mddev
);
6462 blk_start_plug(&plug
);
6464 spin_lock_irq(&conf
->device_lock
);
6467 int batch_size
, released
;
6468 unsigned int offset
;
6470 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
6472 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6475 !list_empty(&conf
->bitmap_list
)) {
6476 /* Now is a good time to flush some bitmap updates */
6478 spin_unlock_irq(&conf
->device_lock
);
6479 md_bitmap_unplug(mddev
->bitmap
);
6480 spin_lock_irq(&conf
->device_lock
);
6481 conf
->seq_write
= conf
->seq_flush
;
6482 activate_bit_delay(conf
, conf
->temp_inactive_list
);
6484 raid5_activate_delayed(conf
);
6486 while ((bio
= remove_bio_from_retry(conf
, &offset
))) {
6488 spin_unlock_irq(&conf
->device_lock
);
6489 ok
= retry_aligned_read(conf
, bio
, offset
);
6490 spin_lock_irq(&conf
->device_lock
);
6496 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
6497 conf
->temp_inactive_list
);
6498 if (!batch_size
&& !released
)
6500 handled
+= batch_size
;
6502 if (mddev
->sb_flags
& ~(1 << MD_SB_CHANGE_PENDING
)) {
6503 spin_unlock_irq(&conf
->device_lock
);
6504 md_check_recovery(mddev
);
6505 spin_lock_irq(&conf
->device_lock
);
6508 pr_debug("%d stripes handled\n", handled
);
6510 spin_unlock_irq(&conf
->device_lock
);
6511 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
) &&
6512 mutex_trylock(&conf
->cache_size_mutex
)) {
6513 grow_one_stripe(conf
, __GFP_NOWARN
);
6514 /* Set flag even if allocation failed. This helps
6515 * slow down allocation requests when mem is short
6517 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6518 mutex_unlock(&conf
->cache_size_mutex
);
6521 flush_deferred_bios(conf
);
6523 r5l_flush_stripe_to_raid(conf
->log
);
6525 async_tx_issue_pending_all();
6526 blk_finish_plug(&plug
);
6528 pr_debug("--- raid5d inactive\n");
6532 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
6534 struct r5conf
*conf
;
6536 spin_lock(&mddev
->lock
);
6537 conf
= mddev
->private;
6539 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
6540 spin_unlock(&mddev
->lock
);
6545 raid5_set_cache_size(struct mddev
*mddev
, int size
)
6548 struct r5conf
*conf
= mddev
->private;
6550 if (size
<= 16 || size
> 32768)
6553 conf
->min_nr_stripes
= size
;
6554 mutex_lock(&conf
->cache_size_mutex
);
6555 while (size
< conf
->max_nr_stripes
&&
6556 drop_one_stripe(conf
))
6558 mutex_unlock(&conf
->cache_size_mutex
);
6560 md_allow_write(mddev
);
6562 mutex_lock(&conf
->cache_size_mutex
);
6563 while (size
> conf
->max_nr_stripes
)
6564 if (!grow_one_stripe(conf
, GFP_KERNEL
)) {
6565 conf
->min_nr_stripes
= conf
->max_nr_stripes
;
6569 mutex_unlock(&conf
->cache_size_mutex
);
6573 EXPORT_SYMBOL(raid5_set_cache_size
);
6576 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
6578 struct r5conf
*conf
;
6582 if (len
>= PAGE_SIZE
)
6584 if (kstrtoul(page
, 10, &new))
6586 err
= mddev_lock(mddev
);
6589 conf
= mddev
->private;
6593 err
= raid5_set_cache_size(mddev
, new);
6594 mddev_unlock(mddev
);
6599 static struct md_sysfs_entry
6600 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
6601 raid5_show_stripe_cache_size
,
6602 raid5_store_stripe_cache_size
);
6605 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
6607 struct r5conf
*conf
= mddev
->private;
6609 return sprintf(page
, "%d\n", conf
->rmw_level
);
6615 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
6617 struct r5conf
*conf
= mddev
->private;
6623 if (len
>= PAGE_SIZE
)
6626 if (kstrtoul(page
, 10, &new))
6629 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
6632 if (new != PARITY_DISABLE_RMW
&&
6633 new != PARITY_ENABLE_RMW
&&
6634 new != PARITY_PREFER_RMW
)
6637 conf
->rmw_level
= new;
6641 static struct md_sysfs_entry
6642 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
6643 raid5_show_rmw_level
,
6644 raid5_store_rmw_level
);
6647 raid5_show_stripe_size(struct mddev
*mddev
, char *page
)
6649 struct r5conf
*conf
;
6652 spin_lock(&mddev
->lock
);
6653 conf
= mddev
->private;
6655 ret
= sprintf(page
, "%lu\n", RAID5_STRIPE_SIZE(conf
));
6656 spin_unlock(&mddev
->lock
);
6660 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6662 raid5_store_stripe_size(struct mddev
*mddev
, const char *page
, size_t len
)
6664 struct r5conf
*conf
;
6669 if (len
>= PAGE_SIZE
)
6671 if (kstrtoul(page
, 10, &new))
6675 * The value should not be bigger than PAGE_SIZE. It requires to
6676 * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6679 if (new % DEFAULT_STRIPE_SIZE
!= 0 ||
6680 new > PAGE_SIZE
|| new == 0 ||
6681 new != roundup_pow_of_two(new))
6684 err
= mddev_lock(mddev
);
6688 conf
= mddev
->private;
6694 if (new == conf
->stripe_size
)
6697 pr_debug("md/raid: change stripe_size from %lu to %lu\n",
6698 conf
->stripe_size
, new);
6700 if (mddev
->sync_thread
||
6701 test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
) ||
6702 mddev
->reshape_position
!= MaxSector
||
6703 mddev
->sysfs_active
) {
6708 mddev_suspend(mddev
);
6709 mutex_lock(&conf
->cache_size_mutex
);
6710 size
= conf
->max_nr_stripes
;
6712 shrink_stripes(conf
);
6714 conf
->stripe_size
= new;
6715 conf
->stripe_shift
= ilog2(new) - 9;
6716 conf
->stripe_sectors
= new >> 9;
6717 if (grow_stripes(conf
, size
)) {
6718 pr_warn("md/raid:%s: couldn't allocate buffers\n",
6722 mutex_unlock(&conf
->cache_size_mutex
);
6723 mddev_resume(mddev
);
6726 mddev_unlock(mddev
);
6730 static struct md_sysfs_entry
6731 raid5_stripe_size
= __ATTR(stripe_size
, 0644,
6732 raid5_show_stripe_size
,
6733 raid5_store_stripe_size
);
6735 static struct md_sysfs_entry
6736 raid5_stripe_size
= __ATTR(stripe_size
, 0444,
6737 raid5_show_stripe_size
,
6742 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
6744 struct r5conf
*conf
;
6746 spin_lock(&mddev
->lock
);
6747 conf
= mddev
->private;
6749 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
6750 spin_unlock(&mddev
->lock
);
6755 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
6757 struct r5conf
*conf
;
6761 if (len
>= PAGE_SIZE
)
6763 if (kstrtoul(page
, 10, &new))
6766 err
= mddev_lock(mddev
);
6769 conf
= mddev
->private;
6772 else if (new > conf
->min_nr_stripes
)
6775 conf
->bypass_threshold
= new;
6776 mddev_unlock(mddev
);
6780 static struct md_sysfs_entry
6781 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
6783 raid5_show_preread_threshold
,
6784 raid5_store_preread_threshold
);
6787 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
6789 struct r5conf
*conf
;
6791 spin_lock(&mddev
->lock
);
6792 conf
= mddev
->private;
6794 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6795 spin_unlock(&mddev
->lock
);
6800 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6802 struct r5conf
*conf
;
6806 if (len
>= PAGE_SIZE
)
6808 if (kstrtoul(page
, 10, &new))
6812 err
= mddev_lock(mddev
);
6815 conf
= mddev
->private;
6818 else if (new != conf
->skip_copy
) {
6819 struct request_queue
*q
= mddev
->queue
;
6821 mddev_suspend(mddev
);
6822 conf
->skip_copy
= new;
6824 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES
, q
);
6826 blk_queue_flag_clear(QUEUE_FLAG_STABLE_WRITES
, q
);
6827 mddev_resume(mddev
);
6829 mddev_unlock(mddev
);
6833 static struct md_sysfs_entry
6834 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6835 raid5_show_skip_copy
,
6836 raid5_store_skip_copy
);
6839 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6841 struct r5conf
*conf
= mddev
->private;
6843 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6848 static struct md_sysfs_entry
6849 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6852 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6854 struct r5conf
*conf
;
6856 spin_lock(&mddev
->lock
);
6857 conf
= mddev
->private;
6859 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6860 spin_unlock(&mddev
->lock
);
6864 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6866 struct r5worker_group
**worker_groups
);
6868 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6870 struct r5conf
*conf
;
6873 struct r5worker_group
*new_groups
, *old_groups
;
6876 if (len
>= PAGE_SIZE
)
6878 if (kstrtouint(page
, 10, &new))
6880 /* 8192 should be big enough */
6884 err
= mddev_lock(mddev
);
6887 conf
= mddev
->private;
6890 else if (new != conf
->worker_cnt_per_group
) {
6891 mddev_suspend(mddev
);
6893 old_groups
= conf
->worker_groups
;
6895 flush_workqueue(raid5_wq
);
6897 err
= alloc_thread_groups(conf
, new, &group_cnt
, &new_groups
);
6899 spin_lock_irq(&conf
->device_lock
);
6900 conf
->group_cnt
= group_cnt
;
6901 conf
->worker_cnt_per_group
= new;
6902 conf
->worker_groups
= new_groups
;
6903 spin_unlock_irq(&conf
->device_lock
);
6906 kfree(old_groups
[0].workers
);
6909 mddev_resume(mddev
);
6911 mddev_unlock(mddev
);
6916 static struct md_sysfs_entry
6917 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6918 raid5_show_group_thread_cnt
,
6919 raid5_store_group_thread_cnt
);
6921 static struct attribute
*raid5_attrs
[] = {
6922 &raid5_stripecache_size
.attr
,
6923 &raid5_stripecache_active
.attr
,
6924 &raid5_preread_bypass_threshold
.attr
,
6925 &raid5_group_thread_cnt
.attr
,
6926 &raid5_skip_copy
.attr
,
6927 &raid5_rmw_level
.attr
,
6928 &raid5_stripe_size
.attr
,
6929 &r5c_journal_mode
.attr
,
6930 &ppl_write_hint
.attr
,
6933 static struct attribute_group raid5_attrs_group
= {
6935 .attrs
= raid5_attrs
,
6938 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
, int *group_cnt
,
6939 struct r5worker_group
**worker_groups
)
6943 struct r5worker
*workers
;
6947 *worker_groups
= NULL
;
6950 *group_cnt
= num_possible_nodes();
6951 size
= sizeof(struct r5worker
) * cnt
;
6952 workers
= kcalloc(size
, *group_cnt
, GFP_NOIO
);
6953 *worker_groups
= kcalloc(*group_cnt
, sizeof(struct r5worker_group
),
6955 if (!*worker_groups
|| !workers
) {
6957 kfree(*worker_groups
);
6961 for (i
= 0; i
< *group_cnt
; i
++) {
6962 struct r5worker_group
*group
;
6964 group
= &(*worker_groups
)[i
];
6965 INIT_LIST_HEAD(&group
->handle_list
);
6966 INIT_LIST_HEAD(&group
->loprio_list
);
6968 group
->workers
= workers
+ i
* cnt
;
6970 for (j
= 0; j
< cnt
; j
++) {
6971 struct r5worker
*worker
= group
->workers
+ j
;
6972 worker
->group
= group
;
6973 INIT_WORK(&worker
->work
, raid5_do_work
);
6975 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6976 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6983 static void free_thread_groups(struct r5conf
*conf
)
6985 if (conf
->worker_groups
)
6986 kfree(conf
->worker_groups
[0].workers
);
6987 kfree(conf
->worker_groups
);
6988 conf
->worker_groups
= NULL
;
6992 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6994 struct r5conf
*conf
= mddev
->private;
6997 sectors
= mddev
->dev_sectors
;
6999 /* size is defined by the smallest of previous and new size */
7000 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
7002 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
7003 sectors
&= ~((sector_t
)conf
->prev_chunk_sectors
- 1);
7004 return sectors
* (raid_disks
- conf
->max_degraded
);
7007 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
7009 safe_put_page(percpu
->spare_page
);
7010 percpu
->spare_page
= NULL
;
7011 kvfree(percpu
->scribble
);
7012 percpu
->scribble
= NULL
;
7015 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
7017 if (conf
->level
== 6 && !percpu
->spare_page
) {
7018 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
7019 if (!percpu
->spare_page
)
7023 if (scribble_alloc(percpu
,
7024 max(conf
->raid_disks
,
7025 conf
->previous_raid_disks
),
7026 max(conf
->chunk_sectors
,
7027 conf
->prev_chunk_sectors
)
7028 / RAID5_STRIPE_SECTORS(conf
))) {
7029 free_scratch_buffer(conf
, percpu
);
7036 static int raid456_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
7038 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
7040 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
7044 static void raid5_free_percpu(struct r5conf
*conf
)
7049 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
7050 free_percpu(conf
->percpu
);
7053 static void free_conf(struct r5conf
*conf
)
7059 unregister_shrinker(&conf
->shrinker
);
7060 free_thread_groups(conf
);
7061 shrink_stripes(conf
);
7062 raid5_free_percpu(conf
);
7063 for (i
= 0; i
< conf
->pool_size
; i
++)
7064 if (conf
->disks
[i
].extra_page
)
7065 put_page(conf
->disks
[i
].extra_page
);
7067 bioset_exit(&conf
->bio_split
);
7068 kfree(conf
->stripe_hashtbl
);
7069 kfree(conf
->pending_data
);
7073 static int raid456_cpu_up_prepare(unsigned int cpu
, struct hlist_node
*node
)
7075 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
7076 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
7078 if (alloc_scratch_buffer(conf
, percpu
)) {
7079 pr_warn("%s: failed memory allocation for cpu%u\n",
7086 static int raid5_alloc_percpu(struct r5conf
*conf
)
7090 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
7094 err
= cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
7096 conf
->scribble_disks
= max(conf
->raid_disks
,
7097 conf
->previous_raid_disks
);
7098 conf
->scribble_sectors
= max(conf
->chunk_sectors
,
7099 conf
->prev_chunk_sectors
);
7104 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
7105 struct shrink_control
*sc
)
7107 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
7108 unsigned long ret
= SHRINK_STOP
;
7110 if (mutex_trylock(&conf
->cache_size_mutex
)) {
7112 while (ret
< sc
->nr_to_scan
&&
7113 conf
->max_nr_stripes
> conf
->min_nr_stripes
) {
7114 if (drop_one_stripe(conf
) == 0) {
7120 mutex_unlock(&conf
->cache_size_mutex
);
7125 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
7126 struct shrink_control
*sc
)
7128 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
7130 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
7131 /* unlikely, but not impossible */
7133 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
7136 static struct r5conf
*setup_conf(struct mddev
*mddev
)
7138 struct r5conf
*conf
;
7139 int raid_disk
, memory
, max_disks
;
7140 struct md_rdev
*rdev
;
7141 struct disk_info
*disk
;
7145 struct r5worker_group
*new_group
;
7148 if (mddev
->new_level
!= 5
7149 && mddev
->new_level
!= 4
7150 && mddev
->new_level
!= 6) {
7151 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7152 mdname(mddev
), mddev
->new_level
);
7153 return ERR_PTR(-EIO
);
7155 if ((mddev
->new_level
== 5
7156 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
7157 (mddev
->new_level
== 6
7158 && !algorithm_valid_raid6(mddev
->new_layout
))) {
7159 pr_warn("md/raid:%s: layout %d not supported\n",
7160 mdname(mddev
), mddev
->new_layout
);
7161 return ERR_PTR(-EIO
);
7163 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
7164 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7165 mdname(mddev
), mddev
->raid_disks
);
7166 return ERR_PTR(-EINVAL
);
7169 if (!mddev
->new_chunk_sectors
||
7170 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
7171 !is_power_of_2(mddev
->new_chunk_sectors
)) {
7172 pr_warn("md/raid:%s: invalid chunk size %d\n",
7173 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
7174 return ERR_PTR(-EINVAL
);
7177 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
7181 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7182 conf
->stripe_size
= DEFAULT_STRIPE_SIZE
;
7183 conf
->stripe_shift
= ilog2(DEFAULT_STRIPE_SIZE
) - 9;
7184 conf
->stripe_sectors
= DEFAULT_STRIPE_SIZE
>> 9;
7186 INIT_LIST_HEAD(&conf
->free_list
);
7187 INIT_LIST_HEAD(&conf
->pending_list
);
7188 conf
->pending_data
= kcalloc(PENDING_IO_MAX
,
7189 sizeof(struct r5pending_data
),
7191 if (!conf
->pending_data
)
7193 for (i
= 0; i
< PENDING_IO_MAX
; i
++)
7194 list_add(&conf
->pending_data
[i
].sibling
, &conf
->free_list
);
7195 /* Don't enable multi-threading by default*/
7196 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &new_group
)) {
7197 conf
->group_cnt
= group_cnt
;
7198 conf
->worker_cnt_per_group
= 0;
7199 conf
->worker_groups
= new_group
;
7202 spin_lock_init(&conf
->device_lock
);
7203 seqcount_spinlock_init(&conf
->gen_lock
, &conf
->device_lock
);
7204 mutex_init(&conf
->cache_size_mutex
);
7205 init_waitqueue_head(&conf
->wait_for_quiescent
);
7206 init_waitqueue_head(&conf
->wait_for_stripe
);
7207 init_waitqueue_head(&conf
->wait_for_overlap
);
7208 INIT_LIST_HEAD(&conf
->handle_list
);
7209 INIT_LIST_HEAD(&conf
->loprio_list
);
7210 INIT_LIST_HEAD(&conf
->hold_list
);
7211 INIT_LIST_HEAD(&conf
->delayed_list
);
7212 INIT_LIST_HEAD(&conf
->bitmap_list
);
7213 init_llist_head(&conf
->released_stripes
);
7214 atomic_set(&conf
->active_stripes
, 0);
7215 atomic_set(&conf
->preread_active_stripes
, 0);
7216 atomic_set(&conf
->active_aligned_reads
, 0);
7217 spin_lock_init(&conf
->pending_bios_lock
);
7218 conf
->batch_bio_dispatch
= true;
7219 rdev_for_each(rdev
, mddev
) {
7220 if (test_bit(Journal
, &rdev
->flags
))
7222 if (blk_queue_nonrot(bdev_get_queue(rdev
->bdev
))) {
7223 conf
->batch_bio_dispatch
= false;
7228 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
7229 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
7231 conf
->raid_disks
= mddev
->raid_disks
;
7232 if (mddev
->reshape_position
== MaxSector
)
7233 conf
->previous_raid_disks
= mddev
->raid_disks
;
7235 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
7236 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
7238 conf
->disks
= kcalloc(max_disks
, sizeof(struct disk_info
),
7244 for (i
= 0; i
< max_disks
; i
++) {
7245 conf
->disks
[i
].extra_page
= alloc_page(GFP_KERNEL
);
7246 if (!conf
->disks
[i
].extra_page
)
7250 ret
= bioset_init(&conf
->bio_split
, BIO_POOL_SIZE
, 0, 0);
7253 conf
->mddev
= mddev
;
7255 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
7258 /* We init hash_locks[0] separately to that it can be used
7259 * as the reference lock in the spin_lock_nest_lock() call
7260 * in lock_all_device_hash_locks_irq in order to convince
7261 * lockdep that we know what we are doing.
7263 spin_lock_init(conf
->hash_locks
);
7264 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
7265 spin_lock_init(conf
->hash_locks
+ i
);
7267 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
7268 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
7270 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
7271 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
7273 atomic_set(&conf
->r5c_cached_full_stripes
, 0);
7274 INIT_LIST_HEAD(&conf
->r5c_full_stripe_list
);
7275 atomic_set(&conf
->r5c_cached_partial_stripes
, 0);
7276 INIT_LIST_HEAD(&conf
->r5c_partial_stripe_list
);
7277 atomic_set(&conf
->r5c_flushing_full_stripes
, 0);
7278 atomic_set(&conf
->r5c_flushing_partial_stripes
, 0);
7280 conf
->level
= mddev
->new_level
;
7281 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7282 if (raid5_alloc_percpu(conf
) != 0)
7285 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
7287 rdev_for_each(rdev
, mddev
) {
7288 raid_disk
= rdev
->raid_disk
;
7289 if (raid_disk
>= max_disks
7290 || raid_disk
< 0 || test_bit(Journal
, &rdev
->flags
))
7292 disk
= conf
->disks
+ raid_disk
;
7294 if (test_bit(Replacement
, &rdev
->flags
)) {
7295 if (disk
->replacement
)
7297 disk
->replacement
= rdev
;
7304 if (test_bit(In_sync
, &rdev
->flags
)) {
7305 char b
[BDEVNAME_SIZE
];
7306 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7307 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
7308 } else if (rdev
->saved_raid_disk
!= raid_disk
)
7309 /* Cannot rely on bitmap to complete recovery */
7313 conf
->level
= mddev
->new_level
;
7314 if (conf
->level
== 6) {
7315 conf
->max_degraded
= 2;
7316 if (raid6_call
.xor_syndrome
)
7317 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7319 conf
->rmw_level
= PARITY_DISABLE_RMW
;
7321 conf
->max_degraded
= 1;
7322 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7324 conf
->algorithm
= mddev
->new_layout
;
7325 conf
->reshape_progress
= mddev
->reshape_position
;
7326 if (conf
->reshape_progress
!= MaxSector
) {
7327 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
7328 conf
->prev_algo
= mddev
->layout
;
7330 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7331 conf
->prev_algo
= conf
->algorithm
;
7334 conf
->min_nr_stripes
= NR_STRIPES
;
7335 if (mddev
->reshape_position
!= MaxSector
) {
7336 int stripes
= max_t(int,
7337 ((mddev
->chunk_sectors
<< 9) / RAID5_STRIPE_SIZE(conf
)) * 4,
7338 ((mddev
->new_chunk_sectors
<< 9) / RAID5_STRIPE_SIZE(conf
)) * 4);
7339 conf
->min_nr_stripes
= max(NR_STRIPES
, stripes
);
7340 if (conf
->min_nr_stripes
!= NR_STRIPES
)
7341 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7342 mdname(mddev
), conf
->min_nr_stripes
);
7344 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
7345 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
7346 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
7347 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
7348 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7349 mdname(mddev
), memory
);
7352 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev
), memory
);
7354 * Losing a stripe head costs more than the time to refill it,
7355 * it reduces the queue depth and so can hurt throughput.
7356 * So set it rather large, scaled by number of devices.
7358 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
7359 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
7360 conf
->shrinker
.count_objects
= raid5_cache_count
;
7361 conf
->shrinker
.batch
= 128;
7362 conf
->shrinker
.flags
= 0;
7363 if (register_shrinker(&conf
->shrinker
)) {
7364 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7369 sprintf(pers_name
, "raid%d", mddev
->new_level
);
7370 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
7371 if (!conf
->thread
) {
7372 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7382 return ERR_PTR(-EIO
);
7384 return ERR_PTR(-ENOMEM
);
7387 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
7390 case ALGORITHM_PARITY_0
:
7391 if (raid_disk
< max_degraded
)
7394 case ALGORITHM_PARITY_N
:
7395 if (raid_disk
>= raid_disks
- max_degraded
)
7398 case ALGORITHM_PARITY_0_6
:
7399 if (raid_disk
== 0 ||
7400 raid_disk
== raid_disks
- 1)
7403 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7404 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7405 case ALGORITHM_LEFT_SYMMETRIC_6
:
7406 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7407 if (raid_disk
== raid_disks
- 1)
7413 static void raid5_set_io_opt(struct r5conf
*conf
)
7415 blk_queue_io_opt(conf
->mddev
->queue
, (conf
->chunk_sectors
<< 9) *
7416 (conf
->raid_disks
- conf
->max_degraded
));
7419 static int raid5_run(struct mddev
*mddev
)
7421 struct r5conf
*conf
;
7422 int working_disks
= 0;
7423 int dirty_parity_disks
= 0;
7424 struct md_rdev
*rdev
;
7425 struct md_rdev
*journal_dev
= NULL
;
7426 sector_t reshape_offset
= 0;
7428 long long min_offset_diff
= 0;
7431 if (mddev_init_writes_pending(mddev
) < 0)
7434 if (mddev
->recovery_cp
!= MaxSector
)
7435 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7438 rdev_for_each(rdev
, mddev
) {
7441 if (test_bit(Journal
, &rdev
->flags
)) {
7445 if (rdev
->raid_disk
< 0)
7447 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
7449 min_offset_diff
= diff
;
7451 } else if (mddev
->reshape_backwards
&&
7452 diff
< min_offset_diff
)
7453 min_offset_diff
= diff
;
7454 else if (!mddev
->reshape_backwards
&&
7455 diff
> min_offset_diff
)
7456 min_offset_diff
= diff
;
7459 if ((test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) || journal_dev
) &&
7460 (mddev
->bitmap_info
.offset
|| mddev
->bitmap_info
.file
)) {
7461 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7466 if (mddev
->reshape_position
!= MaxSector
) {
7467 /* Check that we can continue the reshape.
7468 * Difficulties arise if the stripe we would write to
7469 * next is at or after the stripe we would read from next.
7470 * For a reshape that changes the number of devices, this
7471 * is only possible for a very short time, and mdadm makes
7472 * sure that time appears to have past before assembling
7473 * the array. So we fail if that time hasn't passed.
7474 * For a reshape that keeps the number of devices the same
7475 * mdadm must be monitoring the reshape can keeping the
7476 * critical areas read-only and backed up. It will start
7477 * the array in read-only mode, so we check for that.
7479 sector_t here_new
, here_old
;
7481 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
7486 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7491 if (mddev
->new_level
!= mddev
->level
) {
7492 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7496 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
7497 /* reshape_position must be on a new-stripe boundary, and one
7498 * further up in new geometry must map after here in old
7500 * If the chunk sizes are different, then as we perform reshape
7501 * in units of the largest of the two, reshape_position needs
7502 * be a multiple of the largest chunk size times new data disks.
7504 here_new
= mddev
->reshape_position
;
7505 chunk_sectors
= max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
);
7506 new_data_disks
= mddev
->raid_disks
- max_degraded
;
7507 if (sector_div(here_new
, chunk_sectors
* new_data_disks
)) {
7508 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7512 reshape_offset
= here_new
* chunk_sectors
;
7513 /* here_new is the stripe we will write to */
7514 here_old
= mddev
->reshape_position
;
7515 sector_div(here_old
, chunk_sectors
* (old_disks
-max_degraded
));
7516 /* here_old is the first stripe that we might need to read
7518 if (mddev
->delta_disks
== 0) {
7519 /* We cannot be sure it is safe to start an in-place
7520 * reshape. It is only safe if user-space is monitoring
7521 * and taking constant backups.
7522 * mdadm always starts a situation like this in
7523 * readonly mode so it can take control before
7524 * allowing any writes. So just check for that.
7526 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
7527 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
7528 /* not really in-place - so OK */;
7529 else if (mddev
->ro
== 0) {
7530 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7534 } else if (mddev
->reshape_backwards
7535 ? (here_new
* chunk_sectors
+ min_offset_diff
<=
7536 here_old
* chunk_sectors
)
7537 : (here_new
* chunk_sectors
>=
7538 here_old
* chunk_sectors
+ (-min_offset_diff
))) {
7539 /* Reading from the same stripe as writing to - bad */
7540 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7544 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev
));
7545 /* OK, we should be able to continue; */
7547 BUG_ON(mddev
->level
!= mddev
->new_level
);
7548 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
7549 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
7550 BUG_ON(mddev
->delta_disks
!= 0);
7553 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) &&
7554 test_bit(MD_HAS_PPL
, &mddev
->flags
)) {
7555 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7557 clear_bit(MD_HAS_PPL
, &mddev
->flags
);
7558 clear_bit(MD_HAS_MULTIPLE_PPLS
, &mddev
->flags
);
7561 if (mddev
->private == NULL
)
7562 conf
= setup_conf(mddev
);
7564 conf
= mddev
->private;
7567 return PTR_ERR(conf
);
7569 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
)) {
7571 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7574 set_disk_ro(mddev
->gendisk
, 1);
7575 } else if (mddev
->recovery_cp
== MaxSector
)
7576 set_bit(MD_JOURNAL_CLEAN
, &mddev
->flags
);
7579 conf
->min_offset_diff
= min_offset_diff
;
7580 mddev
->thread
= conf
->thread
;
7581 conf
->thread
= NULL
;
7582 mddev
->private = conf
;
7584 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
7586 rdev
= conf
->disks
[i
].rdev
;
7587 if (!rdev
&& conf
->disks
[i
].replacement
) {
7588 /* The replacement is all we have yet */
7589 rdev
= conf
->disks
[i
].replacement
;
7590 conf
->disks
[i
].replacement
= NULL
;
7591 clear_bit(Replacement
, &rdev
->flags
);
7592 conf
->disks
[i
].rdev
= rdev
;
7596 if (conf
->disks
[i
].replacement
&&
7597 conf
->reshape_progress
!= MaxSector
) {
7598 /* replacements and reshape simply do not mix. */
7599 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7602 if (test_bit(In_sync
, &rdev
->flags
)) {
7606 /* This disc is not fully in-sync. However if it
7607 * just stored parity (beyond the recovery_offset),
7608 * when we don't need to be concerned about the
7609 * array being dirty.
7610 * When reshape goes 'backwards', we never have
7611 * partially completed devices, so we only need
7612 * to worry about reshape going forwards.
7614 /* Hack because v0.91 doesn't store recovery_offset properly. */
7615 if (mddev
->major_version
== 0 &&
7616 mddev
->minor_version
> 90)
7617 rdev
->recovery_offset
= reshape_offset
;
7619 if (rdev
->recovery_offset
< reshape_offset
) {
7620 /* We need to check old and new layout */
7621 if (!only_parity(rdev
->raid_disk
,
7624 conf
->max_degraded
))
7627 if (!only_parity(rdev
->raid_disk
,
7629 conf
->previous_raid_disks
,
7630 conf
->max_degraded
))
7632 dirty_parity_disks
++;
7636 * 0 for a fully functional array, 1 or 2 for a degraded array.
7638 mddev
->degraded
= raid5_calc_degraded(conf
);
7640 if (has_failed(conf
)) {
7641 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7642 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
7646 /* device size must be a multiple of chunk size */
7647 mddev
->dev_sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
7648 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
7650 if (mddev
->degraded
> dirty_parity_disks
&&
7651 mddev
->recovery_cp
!= MaxSector
) {
7652 if (test_bit(MD_HAS_PPL
, &mddev
->flags
))
7653 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7655 else if (mddev
->ok_start_degraded
)
7656 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7659 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7665 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7666 mdname(mddev
), conf
->level
,
7667 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
7670 print_raid5_conf(conf
);
7672 if (conf
->reshape_progress
!= MaxSector
) {
7673 conf
->reshape_safe
= conf
->reshape_progress
;
7674 atomic_set(&conf
->reshape_stripes
, 0);
7675 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7676 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7677 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7678 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7679 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7681 if (!mddev
->sync_thread
)
7685 /* Ok, everything is just fine now */
7686 if (mddev
->to_remove
== &raid5_attrs_group
)
7687 mddev
->to_remove
= NULL
;
7688 else if (mddev
->kobj
.sd
&&
7689 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
7690 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7692 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7696 /* read-ahead size must cover two whole stripes, which
7697 * is 2 * (datadisks) * chunksize where 'n' is the
7698 * number of raid devices
7700 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
7701 int stripe
= data_disks
*
7702 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
7704 chunk_size
= mddev
->chunk_sectors
<< 9;
7705 blk_queue_io_min(mddev
->queue
, chunk_size
);
7706 raid5_set_io_opt(conf
);
7707 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
7709 * We can only discard a whole stripe. It doesn't make sense to
7710 * discard data disk but write parity disk
7712 stripe
= stripe
* PAGE_SIZE
;
7713 /* Round up to power of 2, as discard handling
7714 * currently assumes that */
7715 while ((stripe
-1) & stripe
)
7716 stripe
= (stripe
| (stripe
-1)) + 1;
7717 mddev
->queue
->limits
.discard_alignment
= stripe
;
7718 mddev
->queue
->limits
.discard_granularity
= stripe
;
7720 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
7721 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
7723 rdev_for_each(rdev
, mddev
) {
7724 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7725 rdev
->data_offset
<< 9);
7726 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7727 rdev
->new_data_offset
<< 9);
7731 * zeroing is required, otherwise data
7732 * could be lost. Consider a scenario: discard a stripe
7733 * (the stripe could be inconsistent if
7734 * discard_zeroes_data is 0); write one disk of the
7735 * stripe (the stripe could be inconsistent again
7736 * depending on which disks are used to calculate
7737 * parity); the disk is broken; The stripe data of this
7740 * We only allow DISCARD if the sysadmin has confirmed that
7741 * only safe devices are in use by setting a module parameter.
7742 * A better idea might be to turn DISCARD into WRITE_ZEROES
7743 * requests, as that is required to be safe.
7745 if (devices_handle_discard_safely
&&
7746 mddev
->queue
->limits
.max_discard_sectors
>= (stripe
>> 9) &&
7747 mddev
->queue
->limits
.discard_granularity
>= stripe
)
7748 blk_queue_flag_set(QUEUE_FLAG_DISCARD
,
7751 blk_queue_flag_clear(QUEUE_FLAG_DISCARD
,
7754 blk_queue_max_hw_sectors(mddev
->queue
, UINT_MAX
);
7757 if (log_init(conf
, journal_dev
, raid5_has_ppl(conf
)))
7762 md_unregister_thread(&mddev
->thread
);
7763 print_raid5_conf(conf
);
7765 mddev
->private = NULL
;
7766 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev
));
7770 static void raid5_free(struct mddev
*mddev
, void *priv
)
7772 struct r5conf
*conf
= priv
;
7775 mddev
->to_remove
= &raid5_attrs_group
;
7778 static void raid5_status(struct seq_file
*seq
, struct mddev
*mddev
)
7780 struct r5conf
*conf
= mddev
->private;
7783 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
7784 conf
->chunk_sectors
/ 2, mddev
->layout
);
7785 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
7787 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7788 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
7789 seq_printf (seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
7792 seq_printf (seq
, "]");
7795 static void print_raid5_conf (struct r5conf
*conf
)
7798 struct disk_info
*tmp
;
7800 pr_debug("RAID conf printout:\n");
7802 pr_debug("(conf==NULL)\n");
7805 pr_debug(" --- level:%d rd:%d wd:%d\n", conf
->level
,
7807 conf
->raid_disks
- conf
->mddev
->degraded
);
7809 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7810 char b
[BDEVNAME_SIZE
];
7811 tmp
= conf
->disks
+ i
;
7813 pr_debug(" disk %d, o:%d, dev:%s\n",
7814 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
7815 bdevname(tmp
->rdev
->bdev
, b
));
7819 static int raid5_spare_active(struct mddev
*mddev
)
7822 struct r5conf
*conf
= mddev
->private;
7823 struct disk_info
*tmp
;
7825 unsigned long flags
;
7827 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7828 tmp
= conf
->disks
+ i
;
7829 if (tmp
->replacement
7830 && tmp
->replacement
->recovery_offset
== MaxSector
7831 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
7832 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
7833 /* Replacement has just become active. */
7835 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
7838 /* Replaced device not technically faulty,
7839 * but we need to be sure it gets removed
7840 * and never re-added.
7842 set_bit(Faulty
, &tmp
->rdev
->flags
);
7843 sysfs_notify_dirent_safe(
7844 tmp
->rdev
->sysfs_state
);
7846 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
7847 } else if (tmp
->rdev
7848 && tmp
->rdev
->recovery_offset
== MaxSector
7849 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
7850 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
7852 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
7855 spin_lock_irqsave(&conf
->device_lock
, flags
);
7856 mddev
->degraded
= raid5_calc_degraded(conf
);
7857 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7858 print_raid5_conf(conf
);
7862 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7864 struct r5conf
*conf
= mddev
->private;
7866 int number
= rdev
->raid_disk
;
7867 struct md_rdev
**rdevp
;
7868 struct disk_info
*p
= conf
->disks
+ number
;
7870 print_raid5_conf(conf
);
7871 if (test_bit(Journal
, &rdev
->flags
) && conf
->log
) {
7873 * we can't wait pending write here, as this is called in
7874 * raid5d, wait will deadlock.
7875 * neilb: there is no locking about new writes here,
7876 * so this cannot be safe.
7878 if (atomic_read(&conf
->active_stripes
) ||
7879 atomic_read(&conf
->r5c_cached_full_stripes
) ||
7880 atomic_read(&conf
->r5c_cached_partial_stripes
)) {
7886 if (rdev
== p
->rdev
)
7888 else if (rdev
== p
->replacement
)
7889 rdevp
= &p
->replacement
;
7893 if (number
>= conf
->raid_disks
&&
7894 conf
->reshape_progress
== MaxSector
)
7895 clear_bit(In_sync
, &rdev
->flags
);
7897 if (test_bit(In_sync
, &rdev
->flags
) ||
7898 atomic_read(&rdev
->nr_pending
)) {
7902 /* Only remove non-faulty devices if recovery
7905 if (!test_bit(Faulty
, &rdev
->flags
) &&
7906 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7907 !has_failed(conf
) &&
7908 (!p
->replacement
|| p
->replacement
== rdev
) &&
7909 number
< conf
->raid_disks
) {
7914 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
7916 if (atomic_read(&rdev
->nr_pending
)) {
7917 /* lost the race, try later */
7923 err
= log_modify(conf
, rdev
, false);
7927 if (p
->replacement
) {
7928 /* We must have just cleared 'rdev' */
7929 p
->rdev
= p
->replacement
;
7930 clear_bit(Replacement
, &p
->replacement
->flags
);
7931 smp_mb(); /* Make sure other CPUs may see both as identical
7932 * but will never see neither - if they are careful
7934 p
->replacement
= NULL
;
7937 err
= log_modify(conf
, p
->rdev
, true);
7940 clear_bit(WantReplacement
, &rdev
->flags
);
7943 print_raid5_conf(conf
);
7947 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7949 struct r5conf
*conf
= mddev
->private;
7950 int ret
, err
= -EEXIST
;
7952 struct disk_info
*p
;
7954 int last
= conf
->raid_disks
- 1;
7956 if (test_bit(Journal
, &rdev
->flags
)) {
7960 rdev
->raid_disk
= 0;
7962 * The array is in readonly mode if journal is missing, so no
7963 * write requests running. We should be safe
7965 ret
= log_init(conf
, rdev
, false);
7969 ret
= r5l_start(conf
->log
);
7975 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7978 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7979 /* no point adding a device */
7982 if (rdev
->raid_disk
>= 0)
7983 first
= last
= rdev
->raid_disk
;
7986 * find the disk ... but prefer rdev->saved_raid_disk
7989 if (rdev
->saved_raid_disk
>= 0 &&
7990 rdev
->saved_raid_disk
>= first
&&
7991 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7992 first
= rdev
->saved_raid_disk
;
7994 for (disk
= first
; disk
<= last
; disk
++) {
7995 p
= conf
->disks
+ disk
;
7996 if (p
->rdev
== NULL
) {
7997 clear_bit(In_sync
, &rdev
->flags
);
7998 rdev
->raid_disk
= disk
;
7999 if (rdev
->saved_raid_disk
!= disk
)
8001 rcu_assign_pointer(p
->rdev
, rdev
);
8003 err
= log_modify(conf
, rdev
, true);
8008 for (disk
= first
; disk
<= last
; disk
++) {
8009 p
= conf
->disks
+ disk
;
8010 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
8011 p
->replacement
== NULL
) {
8012 clear_bit(In_sync
, &rdev
->flags
);
8013 set_bit(Replacement
, &rdev
->flags
);
8014 rdev
->raid_disk
= disk
;
8017 rcu_assign_pointer(p
->replacement
, rdev
);
8022 print_raid5_conf(conf
);
8026 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
8028 /* no resync is happening, and there is enough space
8029 * on all devices, so we can resize.
8030 * We need to make sure resync covers any new space.
8031 * If the array is shrinking we should possibly wait until
8032 * any io in the removed space completes, but it hardly seems
8036 struct r5conf
*conf
= mddev
->private;
8038 if (raid5_has_log(conf
) || raid5_has_ppl(conf
))
8040 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
8041 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
8042 if (mddev
->external_size
&&
8043 mddev
->array_sectors
> newsize
)
8045 if (mddev
->bitmap
) {
8046 int ret
= md_bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
8050 md_set_array_sectors(mddev
, newsize
);
8051 if (sectors
> mddev
->dev_sectors
&&
8052 mddev
->recovery_cp
> mddev
->dev_sectors
) {
8053 mddev
->recovery_cp
= mddev
->dev_sectors
;
8054 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
8056 mddev
->dev_sectors
= sectors
;
8057 mddev
->resync_max_sectors
= sectors
;
8061 static int check_stripe_cache(struct mddev
*mddev
)
8063 /* Can only proceed if there are plenty of stripe_heads.
8064 * We need a minimum of one full stripe,, and for sensible progress
8065 * it is best to have about 4 times that.
8066 * If we require 4 times, then the default 256 4K stripe_heads will
8067 * allow for chunk sizes up to 256K, which is probably OK.
8068 * If the chunk size is greater, user-space should request more
8069 * stripe_heads first.
8071 struct r5conf
*conf
= mddev
->private;
8072 if (((mddev
->chunk_sectors
<< 9) / RAID5_STRIPE_SIZE(conf
)) * 4
8073 > conf
->min_nr_stripes
||
8074 ((mddev
->new_chunk_sectors
<< 9) / RAID5_STRIPE_SIZE(conf
)) * 4
8075 > conf
->min_nr_stripes
) {
8076 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
8078 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
8079 / RAID5_STRIPE_SIZE(conf
))*4);
8085 static int check_reshape(struct mddev
*mddev
)
8087 struct r5conf
*conf
= mddev
->private;
8089 if (raid5_has_log(conf
) || raid5_has_ppl(conf
))
8091 if (mddev
->delta_disks
== 0 &&
8092 mddev
->new_layout
== mddev
->layout
&&
8093 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
8094 return 0; /* nothing to do */
8095 if (has_failed(conf
))
8097 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
8098 /* We might be able to shrink, but the devices must
8099 * be made bigger first.
8100 * For raid6, 4 is the minimum size.
8101 * Otherwise 2 is the minimum
8104 if (mddev
->level
== 6)
8106 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
8110 if (!check_stripe_cache(mddev
))
8113 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
8114 mddev
->delta_disks
> 0)
8115 if (resize_chunks(conf
,
8116 conf
->previous_raid_disks
8117 + max(0, mddev
->delta_disks
),
8118 max(mddev
->new_chunk_sectors
,
8119 mddev
->chunk_sectors
)
8123 if (conf
->previous_raid_disks
+ mddev
->delta_disks
<= conf
->pool_size
)
8124 return 0; /* never bother to shrink */
8125 return resize_stripes(conf
, (conf
->previous_raid_disks
8126 + mddev
->delta_disks
));
8129 static int raid5_start_reshape(struct mddev
*mddev
)
8131 struct r5conf
*conf
= mddev
->private;
8132 struct md_rdev
*rdev
;
8134 unsigned long flags
;
8136 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
8139 if (!check_stripe_cache(mddev
))
8142 if (has_failed(conf
))
8145 rdev_for_each(rdev
, mddev
) {
8146 if (!test_bit(In_sync
, &rdev
->flags
)
8147 && !test_bit(Faulty
, &rdev
->flags
))
8151 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
8152 /* Not enough devices even to make a degraded array
8157 /* Refuse to reduce size of the array. Any reductions in
8158 * array size must be through explicit setting of array_size
8161 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
8162 < mddev
->array_sectors
) {
8163 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8168 atomic_set(&conf
->reshape_stripes
, 0);
8169 spin_lock_irq(&conf
->device_lock
);
8170 write_seqcount_begin(&conf
->gen_lock
);
8171 conf
->previous_raid_disks
= conf
->raid_disks
;
8172 conf
->raid_disks
+= mddev
->delta_disks
;
8173 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
8174 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
8175 conf
->prev_algo
= conf
->algorithm
;
8176 conf
->algorithm
= mddev
->new_layout
;
8178 /* Code that selects data_offset needs to see the generation update
8179 * if reshape_progress has been set - so a memory barrier needed.
8182 if (mddev
->reshape_backwards
)
8183 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
8185 conf
->reshape_progress
= 0;
8186 conf
->reshape_safe
= conf
->reshape_progress
;
8187 write_seqcount_end(&conf
->gen_lock
);
8188 spin_unlock_irq(&conf
->device_lock
);
8190 /* Now make sure any requests that proceeded on the assumption
8191 * the reshape wasn't running - like Discard or Read - have
8194 mddev_suspend(mddev
);
8195 mddev_resume(mddev
);
8197 /* Add some new drives, as many as will fit.
8198 * We know there are enough to make the newly sized array work.
8199 * Don't add devices if we are reducing the number of
8200 * devices in the array. This is because it is not possible
8201 * to correctly record the "partially reconstructed" state of
8202 * such devices during the reshape and confusion could result.
8204 if (mddev
->delta_disks
>= 0) {
8205 rdev_for_each(rdev
, mddev
)
8206 if (rdev
->raid_disk
< 0 &&
8207 !test_bit(Faulty
, &rdev
->flags
)) {
8208 if (raid5_add_disk(mddev
, rdev
) == 0) {
8210 >= conf
->previous_raid_disks
)
8211 set_bit(In_sync
, &rdev
->flags
);
8213 rdev
->recovery_offset
= 0;
8215 /* Failure here is OK */
8216 sysfs_link_rdev(mddev
, rdev
);
8218 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
8219 && !test_bit(Faulty
, &rdev
->flags
)) {
8220 /* This is a spare that was manually added */
8221 set_bit(In_sync
, &rdev
->flags
);
8224 /* When a reshape changes the number of devices,
8225 * ->degraded is measured against the larger of the
8226 * pre and post number of devices.
8228 spin_lock_irqsave(&conf
->device_lock
, flags
);
8229 mddev
->degraded
= raid5_calc_degraded(conf
);
8230 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
8232 mddev
->raid_disks
= conf
->raid_disks
;
8233 mddev
->reshape_position
= conf
->reshape_progress
;
8234 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
8236 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
8237 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
8238 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
8239 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
8240 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
8241 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
8243 if (!mddev
->sync_thread
) {
8244 mddev
->recovery
= 0;
8245 spin_lock_irq(&conf
->device_lock
);
8246 write_seqcount_begin(&conf
->gen_lock
);
8247 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
8248 mddev
->new_chunk_sectors
=
8249 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
8250 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
8251 rdev_for_each(rdev
, mddev
)
8252 rdev
->new_data_offset
= rdev
->data_offset
;
8254 conf
->generation
--;
8255 conf
->reshape_progress
= MaxSector
;
8256 mddev
->reshape_position
= MaxSector
;
8257 write_seqcount_end(&conf
->gen_lock
);
8258 spin_unlock_irq(&conf
->device_lock
);
8261 conf
->reshape_checkpoint
= jiffies
;
8262 md_wakeup_thread(mddev
->sync_thread
);
8263 md_new_event(mddev
);
8267 /* This is called from the reshape thread and should make any
8268 * changes needed in 'conf'
8270 static void end_reshape(struct r5conf
*conf
)
8273 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
8274 struct md_rdev
*rdev
;
8276 spin_lock_irq(&conf
->device_lock
);
8277 conf
->previous_raid_disks
= conf
->raid_disks
;
8278 md_finish_reshape(conf
->mddev
);
8280 conf
->reshape_progress
= MaxSector
;
8281 conf
->mddev
->reshape_position
= MaxSector
;
8282 rdev_for_each(rdev
, conf
->mddev
)
8283 if (rdev
->raid_disk
>= 0 &&
8284 !test_bit(Journal
, &rdev
->flags
) &&
8285 !test_bit(In_sync
, &rdev
->flags
))
8286 rdev
->recovery_offset
= MaxSector
;
8287 spin_unlock_irq(&conf
->device_lock
);
8288 wake_up(&conf
->wait_for_overlap
);
8290 if (conf
->mddev
->queue
)
8291 raid5_set_io_opt(conf
);
8295 /* This is called from the raid5d thread with mddev_lock held.
8296 * It makes config changes to the device.
8298 static void raid5_finish_reshape(struct mddev
*mddev
)
8300 struct r5conf
*conf
= mddev
->private;
8302 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
8304 if (mddev
->delta_disks
<= 0) {
8306 spin_lock_irq(&conf
->device_lock
);
8307 mddev
->degraded
= raid5_calc_degraded(conf
);
8308 spin_unlock_irq(&conf
->device_lock
);
8309 for (d
= conf
->raid_disks
;
8310 d
< conf
->raid_disks
- mddev
->delta_disks
;
8312 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
8314 clear_bit(In_sync
, &rdev
->flags
);
8315 rdev
= conf
->disks
[d
].replacement
;
8317 clear_bit(In_sync
, &rdev
->flags
);
8320 mddev
->layout
= conf
->algorithm
;
8321 mddev
->chunk_sectors
= conf
->chunk_sectors
;
8322 mddev
->reshape_position
= MaxSector
;
8323 mddev
->delta_disks
= 0;
8324 mddev
->reshape_backwards
= 0;
8328 static void raid5_quiesce(struct mddev
*mddev
, int quiesce
)
8330 struct r5conf
*conf
= mddev
->private;
8333 /* stop all writes */
8334 lock_all_device_hash_locks_irq(conf
);
8335 /* '2' tells resync/reshape to pause so that all
8336 * active stripes can drain
8338 r5c_flush_cache(conf
, INT_MAX
);
8340 wait_event_cmd(conf
->wait_for_quiescent
,
8341 atomic_read(&conf
->active_stripes
) == 0 &&
8342 atomic_read(&conf
->active_aligned_reads
) == 0,
8343 unlock_all_device_hash_locks_irq(conf
),
8344 lock_all_device_hash_locks_irq(conf
));
8346 unlock_all_device_hash_locks_irq(conf
);
8347 /* allow reshape to continue */
8348 wake_up(&conf
->wait_for_overlap
);
8350 /* re-enable writes */
8351 lock_all_device_hash_locks_irq(conf
);
8353 wake_up(&conf
->wait_for_quiescent
);
8354 wake_up(&conf
->wait_for_overlap
);
8355 unlock_all_device_hash_locks_irq(conf
);
8357 log_quiesce(conf
, quiesce
);
8360 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
8362 struct r0conf
*raid0_conf
= mddev
->private;
8365 /* for raid0 takeover only one zone is supported */
8366 if (raid0_conf
->nr_strip_zones
> 1) {
8367 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8369 return ERR_PTR(-EINVAL
);
8372 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
8373 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
8374 mddev
->dev_sectors
= sectors
;
8375 mddev
->new_level
= level
;
8376 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8377 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
8378 mddev
->raid_disks
+= 1;
8379 mddev
->delta_disks
= 1;
8380 /* make sure it will be not marked as dirty */
8381 mddev
->recovery_cp
= MaxSector
;
8383 return setup_conf(mddev
);
8386 static void *raid5_takeover_raid1(struct mddev
*mddev
)
8391 if (mddev
->raid_disks
!= 2 ||
8392 mddev
->degraded
> 1)
8393 return ERR_PTR(-EINVAL
);
8395 /* Should check if there are write-behind devices? */
8397 chunksect
= 64*2; /* 64K by default */
8399 /* The array must be an exact multiple of chunksize */
8400 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
8403 if ((chunksect
<<9) < RAID5_STRIPE_SIZE((struct r5conf
*)mddev
->private))
8404 /* array size does not allow a suitable chunk size */
8405 return ERR_PTR(-EINVAL
);
8407 mddev
->new_level
= 5;
8408 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8409 mddev
->new_chunk_sectors
= chunksect
;
8411 ret
= setup_conf(mddev
);
8413 mddev_clear_unsupported_flags(mddev
,
8414 UNSUPPORTED_MDDEV_FLAGS
);
8418 static void *raid5_takeover_raid6(struct mddev
*mddev
)
8422 switch (mddev
->layout
) {
8423 case ALGORITHM_LEFT_ASYMMETRIC_6
:
8424 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
8426 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
8427 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
8429 case ALGORITHM_LEFT_SYMMETRIC_6
:
8430 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8432 case ALGORITHM_RIGHT_SYMMETRIC_6
:
8433 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
8435 case ALGORITHM_PARITY_0_6
:
8436 new_layout
= ALGORITHM_PARITY_0
;
8438 case ALGORITHM_PARITY_N
:
8439 new_layout
= ALGORITHM_PARITY_N
;
8442 return ERR_PTR(-EINVAL
);
8444 mddev
->new_level
= 5;
8445 mddev
->new_layout
= new_layout
;
8446 mddev
->delta_disks
= -1;
8447 mddev
->raid_disks
-= 1;
8448 return setup_conf(mddev
);
8451 static int raid5_check_reshape(struct mddev
*mddev
)
8453 /* For a 2-drive array, the layout and chunk size can be changed
8454 * immediately as not restriping is needed.
8455 * For larger arrays we record the new value - after validation
8456 * to be used by a reshape pass.
8458 struct r5conf
*conf
= mddev
->private;
8459 int new_chunk
= mddev
->new_chunk_sectors
;
8461 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
8463 if (new_chunk
> 0) {
8464 if (!is_power_of_2(new_chunk
))
8466 if (new_chunk
< (PAGE_SIZE
>>9))
8468 if (mddev
->array_sectors
& (new_chunk
-1))
8469 /* not factor of array size */
8473 /* They look valid */
8475 if (mddev
->raid_disks
== 2) {
8476 /* can make the change immediately */
8477 if (mddev
->new_layout
>= 0) {
8478 conf
->algorithm
= mddev
->new_layout
;
8479 mddev
->layout
= mddev
->new_layout
;
8481 if (new_chunk
> 0) {
8482 conf
->chunk_sectors
= new_chunk
;
8483 mddev
->chunk_sectors
= new_chunk
;
8485 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
8486 md_wakeup_thread(mddev
->thread
);
8488 return check_reshape(mddev
);
8491 static int raid6_check_reshape(struct mddev
*mddev
)
8493 int new_chunk
= mddev
->new_chunk_sectors
;
8495 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
8497 if (new_chunk
> 0) {
8498 if (!is_power_of_2(new_chunk
))
8500 if (new_chunk
< (PAGE_SIZE
>> 9))
8502 if (mddev
->array_sectors
& (new_chunk
-1))
8503 /* not factor of array size */
8507 /* They look valid */
8508 return check_reshape(mddev
);
8511 static void *raid5_takeover(struct mddev
*mddev
)
8513 /* raid5 can take over:
8514 * raid0 - if there is only one strip zone - make it a raid4 layout
8515 * raid1 - if there are two drives. We need to know the chunk size
8516 * raid4 - trivial - just use a raid4 layout.
8517 * raid6 - Providing it is a *_6 layout
8519 if (mddev
->level
== 0)
8520 return raid45_takeover_raid0(mddev
, 5);
8521 if (mddev
->level
== 1)
8522 return raid5_takeover_raid1(mddev
);
8523 if (mddev
->level
== 4) {
8524 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8525 mddev
->new_level
= 5;
8526 return setup_conf(mddev
);
8528 if (mddev
->level
== 6)
8529 return raid5_takeover_raid6(mddev
);
8531 return ERR_PTR(-EINVAL
);
8534 static void *raid4_takeover(struct mddev
*mddev
)
8536 /* raid4 can take over:
8537 * raid0 - if there is only one strip zone
8538 * raid5 - if layout is right
8540 if (mddev
->level
== 0)
8541 return raid45_takeover_raid0(mddev
, 4);
8542 if (mddev
->level
== 5 &&
8543 mddev
->layout
== ALGORITHM_PARITY_N
) {
8544 mddev
->new_layout
= 0;
8545 mddev
->new_level
= 4;
8546 return setup_conf(mddev
);
8548 return ERR_PTR(-EINVAL
);
8551 static struct md_personality raid5_personality
;
8553 static void *raid6_takeover(struct mddev
*mddev
)
8555 /* Currently can only take over a raid5. We map the
8556 * personality to an equivalent raid6 personality
8557 * with the Q block at the end.
8561 if (mddev
->pers
!= &raid5_personality
)
8562 return ERR_PTR(-EINVAL
);
8563 if (mddev
->degraded
> 1)
8564 return ERR_PTR(-EINVAL
);
8565 if (mddev
->raid_disks
> 253)
8566 return ERR_PTR(-EINVAL
);
8567 if (mddev
->raid_disks
< 3)
8568 return ERR_PTR(-EINVAL
);
8570 switch (mddev
->layout
) {
8571 case ALGORITHM_LEFT_ASYMMETRIC
:
8572 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
8574 case ALGORITHM_RIGHT_ASYMMETRIC
:
8575 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
8577 case ALGORITHM_LEFT_SYMMETRIC
:
8578 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
8580 case ALGORITHM_RIGHT_SYMMETRIC
:
8581 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
8583 case ALGORITHM_PARITY_0
:
8584 new_layout
= ALGORITHM_PARITY_0_6
;
8586 case ALGORITHM_PARITY_N
:
8587 new_layout
= ALGORITHM_PARITY_N
;
8590 return ERR_PTR(-EINVAL
);
8592 mddev
->new_level
= 6;
8593 mddev
->new_layout
= new_layout
;
8594 mddev
->delta_disks
= 1;
8595 mddev
->raid_disks
+= 1;
8596 return setup_conf(mddev
);
8599 static int raid5_change_consistency_policy(struct mddev
*mddev
, const char *buf
)
8601 struct r5conf
*conf
;
8604 err
= mddev_lock(mddev
);
8607 conf
= mddev
->private;
8609 mddev_unlock(mddev
);
8613 if (strncmp(buf
, "ppl", 3) == 0) {
8614 /* ppl only works with RAID 5 */
8615 if (!raid5_has_ppl(conf
) && conf
->level
== 5) {
8616 err
= log_init(conf
, NULL
, true);
8618 err
= resize_stripes(conf
, conf
->pool_size
);
8624 } else if (strncmp(buf
, "resync", 6) == 0) {
8625 if (raid5_has_ppl(conf
)) {
8626 mddev_suspend(mddev
);
8628 mddev_resume(mddev
);
8629 err
= resize_stripes(conf
, conf
->pool_size
);
8630 } else if (test_bit(MD_HAS_JOURNAL
, &conf
->mddev
->flags
) &&
8631 r5l_log_disk_error(conf
)) {
8632 bool journal_dev_exists
= false;
8633 struct md_rdev
*rdev
;
8635 rdev_for_each(rdev
, mddev
)
8636 if (test_bit(Journal
, &rdev
->flags
)) {
8637 journal_dev_exists
= true;
8641 if (!journal_dev_exists
) {
8642 mddev_suspend(mddev
);
8643 clear_bit(MD_HAS_JOURNAL
, &mddev
->flags
);
8644 mddev_resume(mddev
);
8645 } else /* need remove journal device first */
8654 md_update_sb(mddev
, 1);
8656 mddev_unlock(mddev
);
8661 static int raid5_start(struct mddev
*mddev
)
8663 struct r5conf
*conf
= mddev
->private;
8665 return r5l_start(conf
->log
);
8668 static struct md_personality raid6_personality
=
8672 .owner
= THIS_MODULE
,
8673 .make_request
= raid5_make_request
,
8675 .start
= raid5_start
,
8677 .status
= raid5_status
,
8678 .error_handler
= raid5_error
,
8679 .hot_add_disk
= raid5_add_disk
,
8680 .hot_remove_disk
= raid5_remove_disk
,
8681 .spare_active
= raid5_spare_active
,
8682 .sync_request
= raid5_sync_request
,
8683 .resize
= raid5_resize
,
8685 .check_reshape
= raid6_check_reshape
,
8686 .start_reshape
= raid5_start_reshape
,
8687 .finish_reshape
= raid5_finish_reshape
,
8688 .quiesce
= raid5_quiesce
,
8689 .takeover
= raid6_takeover
,
8690 .change_consistency_policy
= raid5_change_consistency_policy
,
8692 static struct md_personality raid5_personality
=
8696 .owner
= THIS_MODULE
,
8697 .make_request
= raid5_make_request
,
8699 .start
= raid5_start
,
8701 .status
= raid5_status
,
8702 .error_handler
= raid5_error
,
8703 .hot_add_disk
= raid5_add_disk
,
8704 .hot_remove_disk
= raid5_remove_disk
,
8705 .spare_active
= raid5_spare_active
,
8706 .sync_request
= raid5_sync_request
,
8707 .resize
= raid5_resize
,
8709 .check_reshape
= raid5_check_reshape
,
8710 .start_reshape
= raid5_start_reshape
,
8711 .finish_reshape
= raid5_finish_reshape
,
8712 .quiesce
= raid5_quiesce
,
8713 .takeover
= raid5_takeover
,
8714 .change_consistency_policy
= raid5_change_consistency_policy
,
8717 static struct md_personality raid4_personality
=
8721 .owner
= THIS_MODULE
,
8722 .make_request
= raid5_make_request
,
8724 .start
= raid5_start
,
8726 .status
= raid5_status
,
8727 .error_handler
= raid5_error
,
8728 .hot_add_disk
= raid5_add_disk
,
8729 .hot_remove_disk
= raid5_remove_disk
,
8730 .spare_active
= raid5_spare_active
,
8731 .sync_request
= raid5_sync_request
,
8732 .resize
= raid5_resize
,
8734 .check_reshape
= raid5_check_reshape
,
8735 .start_reshape
= raid5_start_reshape
,
8736 .finish_reshape
= raid5_finish_reshape
,
8737 .quiesce
= raid5_quiesce
,
8738 .takeover
= raid4_takeover
,
8739 .change_consistency_policy
= raid5_change_consistency_policy
,
8742 static int __init
raid5_init(void)
8746 raid5_wq
= alloc_workqueue("raid5wq",
8747 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
8751 ret
= cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE
,
8753 raid456_cpu_up_prepare
,
8756 destroy_workqueue(raid5_wq
);
8759 register_md_personality(&raid6_personality
);
8760 register_md_personality(&raid5_personality
);
8761 register_md_personality(&raid4_personality
);
8765 static void raid5_exit(void)
8767 unregister_md_personality(&raid6_personality
);
8768 unregister_md_personality(&raid5_personality
);
8769 unregister_md_personality(&raid4_personality
);
8770 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE
);
8771 destroy_workqueue(raid5_wq
);
8774 module_init(raid5_init
);
8775 module_exit(raid5_exit
);
8776 MODULE_LICENSE("GPL");
8777 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8778 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8779 MODULE_ALIAS("md-raid5");
8780 MODULE_ALIAS("md-raid4");
8781 MODULE_ALIAS("md-level-5");
8782 MODULE_ALIAS("md-level-4");
8783 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8784 MODULE_ALIAS("md-raid6");
8785 MODULE_ALIAS("md-level-6");
8787 /* This used to be two separate modules, they were: */
8788 MODULE_ALIAS("raid5");
8789 MODULE_ALIAS("raid6");