2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <linux/sched/signal.h>
60 #include <trace/events/block.h>
67 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
69 #define cpu_to_group(cpu) cpu_to_node(cpu)
70 #define ANY_GROUP NUMA_NO_NODE
72 static bool devices_handle_discard_safely
= false;
73 module_param(devices_handle_discard_safely
, bool, 0644);
74 MODULE_PARM_DESC(devices_handle_discard_safely
,
75 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
76 static struct workqueue_struct
*raid5_wq
;
78 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
80 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
81 return &conf
->stripe_hashtbl
[hash
];
84 static inline int stripe_hash_locks_hash(sector_t sect
)
86 return (sect
>> STRIPE_SHIFT
) & STRIPE_HASH_LOCKS_MASK
;
89 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
91 spin_lock_irq(conf
->hash_locks
+ hash
);
92 spin_lock(&conf
->device_lock
);
95 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
97 spin_unlock(&conf
->device_lock
);
98 spin_unlock_irq(conf
->hash_locks
+ hash
);
101 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
105 spin_lock(conf
->hash_locks
);
106 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
107 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
108 spin_lock(&conf
->device_lock
);
111 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
114 spin_unlock(&conf
->device_lock
);
115 for (i
= NR_STRIPE_HASH_LOCKS
; i
; i
--)
116 spin_unlock(conf
->hash_locks
+ i
- 1);
120 /* Find first data disk in a raid6 stripe */
121 static inline int raid6_d0(struct stripe_head
*sh
)
124 /* ddf always start from first device */
126 /* md starts just after Q block */
127 if (sh
->qd_idx
== sh
->disks
- 1)
130 return sh
->qd_idx
+ 1;
132 static inline int raid6_next_disk(int disk
, int raid_disks
)
135 return (disk
< raid_disks
) ? disk
: 0;
138 /* When walking through the disks in a raid5, starting at raid6_d0,
139 * We need to map each disk to a 'slot', where the data disks are slot
140 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
141 * is raid_disks-1. This help does that mapping.
143 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
144 int *count
, int syndrome_disks
)
150 if (idx
== sh
->pd_idx
)
151 return syndrome_disks
;
152 if (idx
== sh
->qd_idx
)
153 return syndrome_disks
+ 1;
159 static void return_io(struct bio_list
*return_bi
)
162 while ((bi
= bio_list_pop(return_bi
)) != NULL
) {
163 bi
->bi_iter
.bi_size
= 0;
164 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
170 static void print_raid5_conf (struct r5conf
*conf
);
172 static int stripe_operations_active(struct stripe_head
*sh
)
174 return sh
->check_state
|| sh
->reconstruct_state
||
175 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
176 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
179 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
181 struct r5conf
*conf
= sh
->raid_conf
;
182 struct r5worker_group
*group
;
184 int i
, cpu
= sh
->cpu
;
186 if (!cpu_online(cpu
)) {
187 cpu
= cpumask_any(cpu_online_mask
);
191 if (list_empty(&sh
->lru
)) {
192 struct r5worker_group
*group
;
193 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
194 list_add_tail(&sh
->lru
, &group
->handle_list
);
195 group
->stripes_cnt
++;
199 if (conf
->worker_cnt_per_group
== 0) {
200 md_wakeup_thread(conf
->mddev
->thread
);
204 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
206 group
->workers
[0].working
= true;
207 /* at least one worker should run to avoid race */
208 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
210 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
211 /* wakeup more workers */
212 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
213 if (group
->workers
[i
].working
== false) {
214 group
->workers
[i
].working
= true;
215 queue_work_on(sh
->cpu
, raid5_wq
,
216 &group
->workers
[i
].work
);
222 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
223 struct list_head
*temp_inactive_list
)
226 int injournal
= 0; /* number of date pages with R5_InJournal */
228 BUG_ON(!list_empty(&sh
->lru
));
229 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
231 if (r5c_is_writeback(conf
->log
))
232 for (i
= sh
->disks
; i
--; )
233 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
236 * When quiesce in r5c write back, set STRIPE_HANDLE for stripes with
237 * data in journal, so they are not released to cached lists
239 if (conf
->quiesce
&& r5c_is_writeback(conf
->log
) &&
240 !test_bit(STRIPE_HANDLE
, &sh
->state
) && injournal
!= 0) {
241 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
242 r5c_make_stripe_write_out(sh
);
243 set_bit(STRIPE_HANDLE
, &sh
->state
);
246 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
247 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
248 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
249 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
250 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
251 sh
->bm_seq
- conf
->seq_write
> 0)
252 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
254 clear_bit(STRIPE_DELAYED
, &sh
->state
);
255 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
256 if (conf
->worker_cnt_per_group
== 0) {
257 list_add_tail(&sh
->lru
, &conf
->handle_list
);
259 raid5_wakeup_stripe_thread(sh
);
263 md_wakeup_thread(conf
->mddev
->thread
);
265 BUG_ON(stripe_operations_active(sh
));
266 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
267 if (atomic_dec_return(&conf
->preread_active_stripes
)
269 md_wakeup_thread(conf
->mddev
->thread
);
270 atomic_dec(&conf
->active_stripes
);
271 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
272 if (!r5c_is_writeback(conf
->log
))
273 list_add_tail(&sh
->lru
, temp_inactive_list
);
275 WARN_ON(test_bit(R5_InJournal
, &sh
->dev
[sh
->pd_idx
].flags
));
277 list_add_tail(&sh
->lru
, temp_inactive_list
);
278 else if (injournal
== conf
->raid_disks
- conf
->max_degraded
) {
280 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
))
281 atomic_inc(&conf
->r5c_cached_full_stripes
);
282 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
))
283 atomic_dec(&conf
->r5c_cached_partial_stripes
);
284 list_add_tail(&sh
->lru
, &conf
->r5c_full_stripe_list
);
285 r5c_check_cached_full_stripe(conf
);
288 * STRIPE_R5C_PARTIAL_STRIPE is set in
289 * r5c_try_caching_write(). No need to
292 list_add_tail(&sh
->lru
, &conf
->r5c_partial_stripe_list
);
298 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
299 struct list_head
*temp_inactive_list
)
301 if (atomic_dec_and_test(&sh
->count
))
302 do_release_stripe(conf
, sh
, temp_inactive_list
);
306 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
308 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
309 * given time. Adding stripes only takes device lock, while deleting stripes
310 * only takes hash lock.
312 static void release_inactive_stripe_list(struct r5conf
*conf
,
313 struct list_head
*temp_inactive_list
,
317 bool do_wakeup
= false;
320 if (hash
== NR_STRIPE_HASH_LOCKS
) {
321 size
= NR_STRIPE_HASH_LOCKS
;
322 hash
= NR_STRIPE_HASH_LOCKS
- 1;
326 struct list_head
*list
= &temp_inactive_list
[size
- 1];
329 * We don't hold any lock here yet, raid5_get_active_stripe() might
330 * remove stripes from the list
332 if (!list_empty_careful(list
)) {
333 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
334 if (list_empty(conf
->inactive_list
+ hash
) &&
336 atomic_dec(&conf
->empty_inactive_list_nr
);
337 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
339 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
346 wake_up(&conf
->wait_for_stripe
);
347 if (atomic_read(&conf
->active_stripes
) == 0)
348 wake_up(&conf
->wait_for_quiescent
);
349 if (conf
->retry_read_aligned
)
350 md_wakeup_thread(conf
->mddev
->thread
);
354 /* should hold conf->device_lock already */
355 static int release_stripe_list(struct r5conf
*conf
,
356 struct list_head
*temp_inactive_list
)
358 struct stripe_head
*sh
, *t
;
360 struct llist_node
*head
;
362 head
= llist_del_all(&conf
->released_stripes
);
363 head
= llist_reverse_order(head
);
364 llist_for_each_entry_safe(sh
, t
, head
, release_list
) {
367 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
369 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
371 * Don't worry the bit is set here, because if the bit is set
372 * again, the count is always > 1. This is true for
373 * STRIPE_ON_UNPLUG_LIST bit too.
375 hash
= sh
->hash_lock_index
;
376 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
383 void raid5_release_stripe(struct stripe_head
*sh
)
385 struct r5conf
*conf
= sh
->raid_conf
;
387 struct list_head list
;
391 /* Avoid release_list until the last reference.
393 if (atomic_add_unless(&sh
->count
, -1, 1))
396 if (unlikely(!conf
->mddev
->thread
) ||
397 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
399 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
401 md_wakeup_thread(conf
->mddev
->thread
);
404 local_irq_save(flags
);
405 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
406 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
407 INIT_LIST_HEAD(&list
);
408 hash
= sh
->hash_lock_index
;
409 do_release_stripe(conf
, sh
, &list
);
410 spin_unlock(&conf
->device_lock
);
411 release_inactive_stripe_list(conf
, &list
, hash
);
413 local_irq_restore(flags
);
416 static inline void remove_hash(struct stripe_head
*sh
)
418 pr_debug("remove_hash(), stripe %llu\n",
419 (unsigned long long)sh
->sector
);
421 hlist_del_init(&sh
->hash
);
424 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
426 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
428 pr_debug("insert_hash(), stripe %llu\n",
429 (unsigned long long)sh
->sector
);
431 hlist_add_head(&sh
->hash
, hp
);
434 /* find an idle stripe, make sure it is unhashed, and return it. */
435 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
437 struct stripe_head
*sh
= NULL
;
438 struct list_head
*first
;
440 if (list_empty(conf
->inactive_list
+ hash
))
442 first
= (conf
->inactive_list
+ hash
)->next
;
443 sh
= list_entry(first
, struct stripe_head
, lru
);
444 list_del_init(first
);
446 atomic_inc(&conf
->active_stripes
);
447 BUG_ON(hash
!= sh
->hash_lock_index
);
448 if (list_empty(conf
->inactive_list
+ hash
))
449 atomic_inc(&conf
->empty_inactive_list_nr
);
454 static void shrink_buffers(struct stripe_head
*sh
)
458 int num
= sh
->raid_conf
->pool_size
;
460 for (i
= 0; i
< num
; i
++) {
461 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
465 sh
->dev
[i
].page
= NULL
;
470 static int grow_buffers(struct stripe_head
*sh
, gfp_t gfp
)
473 int num
= sh
->raid_conf
->pool_size
;
475 for (i
= 0; i
< num
; i
++) {
478 if (!(page
= alloc_page(gfp
))) {
481 sh
->dev
[i
].page
= page
;
482 sh
->dev
[i
].orig_page
= page
;
487 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
488 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
489 struct stripe_head
*sh
);
491 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
493 struct r5conf
*conf
= sh
->raid_conf
;
496 BUG_ON(atomic_read(&sh
->count
) != 0);
497 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
498 BUG_ON(stripe_operations_active(sh
));
499 BUG_ON(sh
->batch_head
);
501 pr_debug("init_stripe called, stripe %llu\n",
502 (unsigned long long)sector
);
504 seq
= read_seqcount_begin(&conf
->gen_lock
);
505 sh
->generation
= conf
->generation
- previous
;
506 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
508 stripe_set_idx(sector
, conf
, previous
, sh
);
511 for (i
= sh
->disks
; i
--; ) {
512 struct r5dev
*dev
= &sh
->dev
[i
];
514 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
515 test_bit(R5_LOCKED
, &dev
->flags
)) {
516 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
517 (unsigned long long)sh
->sector
, i
, dev
->toread
,
518 dev
->read
, dev
->towrite
, dev
->written
,
519 test_bit(R5_LOCKED
, &dev
->flags
));
523 raid5_build_block(sh
, i
, previous
);
525 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
527 sh
->overwrite_disks
= 0;
528 insert_hash(conf
, sh
);
529 sh
->cpu
= smp_processor_id();
530 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
533 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
536 struct stripe_head
*sh
;
538 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
539 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
540 if (sh
->sector
== sector
&& sh
->generation
== generation
)
542 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
547 * Need to check if array has failed when deciding whether to:
549 * - remove non-faulty devices
552 * This determination is simple when no reshape is happening.
553 * However if there is a reshape, we need to carefully check
554 * both the before and after sections.
555 * This is because some failed devices may only affect one
556 * of the two sections, and some non-in_sync devices may
557 * be insync in the section most affected by failed devices.
559 int raid5_calc_degraded(struct r5conf
*conf
)
561 int degraded
, degraded2
;
566 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
567 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
568 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
569 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
570 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
572 else if (test_bit(In_sync
, &rdev
->flags
))
575 /* not in-sync or faulty.
576 * If the reshape increases the number of devices,
577 * this is being recovered by the reshape, so
578 * this 'previous' section is not in_sync.
579 * If the number of devices is being reduced however,
580 * the device can only be part of the array if
581 * we are reverting a reshape, so this section will
584 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
588 if (conf
->raid_disks
== conf
->previous_raid_disks
)
592 for (i
= 0; i
< conf
->raid_disks
; i
++) {
593 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
594 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
595 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
596 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
598 else if (test_bit(In_sync
, &rdev
->flags
))
601 /* not in-sync or faulty.
602 * If reshape increases the number of devices, this
603 * section has already been recovered, else it
604 * almost certainly hasn't.
606 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
610 if (degraded2
> degraded
)
615 static int has_failed(struct r5conf
*conf
)
619 if (conf
->mddev
->reshape_position
== MaxSector
)
620 return conf
->mddev
->degraded
> conf
->max_degraded
;
622 degraded
= raid5_calc_degraded(conf
);
623 if (degraded
> conf
->max_degraded
)
629 raid5_get_active_stripe(struct r5conf
*conf
, sector_t sector
,
630 int previous
, int noblock
, int noquiesce
)
632 struct stripe_head
*sh
;
633 int hash
= stripe_hash_locks_hash(sector
);
634 int inc_empty_inactive_list_flag
;
636 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
638 spin_lock_irq(conf
->hash_locks
+ hash
);
641 wait_event_lock_irq(conf
->wait_for_quiescent
,
642 conf
->quiesce
== 0 || noquiesce
,
643 *(conf
->hash_locks
+ hash
));
644 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
646 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
647 sh
= get_free_stripe(conf
, hash
);
648 if (!sh
&& !test_bit(R5_DID_ALLOC
,
650 set_bit(R5_ALLOC_MORE
,
653 if (noblock
&& sh
== NULL
)
656 r5c_check_stripe_cache_usage(conf
);
658 set_bit(R5_INACTIVE_BLOCKED
,
660 r5l_wake_reclaim(conf
->log
, 0);
662 conf
->wait_for_stripe
,
663 !list_empty(conf
->inactive_list
+ hash
) &&
664 (atomic_read(&conf
->active_stripes
)
665 < (conf
->max_nr_stripes
* 3 / 4)
666 || !test_bit(R5_INACTIVE_BLOCKED
,
667 &conf
->cache_state
)),
668 *(conf
->hash_locks
+ hash
));
669 clear_bit(R5_INACTIVE_BLOCKED
,
672 init_stripe(sh
, sector
, previous
);
673 atomic_inc(&sh
->count
);
675 } else if (!atomic_inc_not_zero(&sh
->count
)) {
676 spin_lock(&conf
->device_lock
);
677 if (!atomic_read(&sh
->count
)) {
678 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
679 atomic_inc(&conf
->active_stripes
);
680 BUG_ON(list_empty(&sh
->lru
) &&
681 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
682 inc_empty_inactive_list_flag
= 0;
683 if (!list_empty(conf
->inactive_list
+ hash
))
684 inc_empty_inactive_list_flag
= 1;
685 list_del_init(&sh
->lru
);
686 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
687 atomic_inc(&conf
->empty_inactive_list_nr
);
689 sh
->group
->stripes_cnt
--;
693 atomic_inc(&sh
->count
);
694 spin_unlock(&conf
->device_lock
);
696 } while (sh
== NULL
);
698 spin_unlock_irq(conf
->hash_locks
+ hash
);
702 static bool is_full_stripe_write(struct stripe_head
*sh
)
704 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
705 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
708 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
712 spin_lock(&sh2
->stripe_lock
);
713 spin_lock_nested(&sh1
->stripe_lock
, 1);
715 spin_lock(&sh1
->stripe_lock
);
716 spin_lock_nested(&sh2
->stripe_lock
, 1);
720 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
722 spin_unlock(&sh1
->stripe_lock
);
723 spin_unlock(&sh2
->stripe_lock
);
727 /* Only freshly new full stripe normal write stripe can be added to a batch list */
728 static bool stripe_can_batch(struct stripe_head
*sh
)
730 struct r5conf
*conf
= sh
->raid_conf
;
734 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
735 !test_bit(STRIPE_BITMAP_PENDING
, &sh
->state
) &&
736 is_full_stripe_write(sh
);
739 /* we only do back search */
740 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
742 struct stripe_head
*head
;
743 sector_t head_sector
, tmp_sec
;
746 int inc_empty_inactive_list_flag
;
748 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
749 tmp_sec
= sh
->sector
;
750 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
752 head_sector
= sh
->sector
- STRIPE_SECTORS
;
754 hash
= stripe_hash_locks_hash(head_sector
);
755 spin_lock_irq(conf
->hash_locks
+ hash
);
756 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
757 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
758 spin_lock(&conf
->device_lock
);
759 if (!atomic_read(&head
->count
)) {
760 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
761 atomic_inc(&conf
->active_stripes
);
762 BUG_ON(list_empty(&head
->lru
) &&
763 !test_bit(STRIPE_EXPANDING
, &head
->state
));
764 inc_empty_inactive_list_flag
= 0;
765 if (!list_empty(conf
->inactive_list
+ hash
))
766 inc_empty_inactive_list_flag
= 1;
767 list_del_init(&head
->lru
);
768 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
769 atomic_inc(&conf
->empty_inactive_list_nr
);
771 head
->group
->stripes_cnt
--;
775 atomic_inc(&head
->count
);
776 spin_unlock(&conf
->device_lock
);
778 spin_unlock_irq(conf
->hash_locks
+ hash
);
782 if (!stripe_can_batch(head
))
785 lock_two_stripes(head
, sh
);
786 /* clear_batch_ready clear the flag */
787 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
794 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
796 if (head
->dev
[dd_idx
].towrite
->bi_opf
!= sh
->dev
[dd_idx
].towrite
->bi_opf
||
797 bio_op(head
->dev
[dd_idx
].towrite
) != bio_op(sh
->dev
[dd_idx
].towrite
))
800 if (head
->batch_head
) {
801 spin_lock(&head
->batch_head
->batch_lock
);
802 /* This batch list is already running */
803 if (!stripe_can_batch(head
)) {
804 spin_unlock(&head
->batch_head
->batch_lock
);
809 * at this point, head's BATCH_READY could be cleared, but we
810 * can still add the stripe to batch list
812 list_add(&sh
->batch_list
, &head
->batch_list
);
813 spin_unlock(&head
->batch_head
->batch_lock
);
815 sh
->batch_head
= head
->batch_head
;
817 head
->batch_head
= head
;
818 sh
->batch_head
= head
->batch_head
;
819 spin_lock(&head
->batch_lock
);
820 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
821 spin_unlock(&head
->batch_lock
);
824 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
825 if (atomic_dec_return(&conf
->preread_active_stripes
)
827 md_wakeup_thread(conf
->mddev
->thread
);
829 if (test_and_clear_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
830 int seq
= sh
->bm_seq
;
831 if (test_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
) &&
832 sh
->batch_head
->bm_seq
> seq
)
833 seq
= sh
->batch_head
->bm_seq
;
834 set_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
);
835 sh
->batch_head
->bm_seq
= seq
;
838 atomic_inc(&sh
->count
);
840 unlock_two_stripes(head
, sh
);
842 raid5_release_stripe(head
);
845 /* Determine if 'data_offset' or 'new_data_offset' should be used
846 * in this stripe_head.
848 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
850 sector_t progress
= conf
->reshape_progress
;
851 /* Need a memory barrier to make sure we see the value
852 * of conf->generation, or ->data_offset that was set before
853 * reshape_progress was updated.
856 if (progress
== MaxSector
)
858 if (sh
->generation
== conf
->generation
- 1)
860 /* We are in a reshape, and this is a new-generation stripe,
861 * so use new_data_offset.
866 static void flush_deferred_bios(struct r5conf
*conf
)
871 if (!conf
->batch_bio_dispatch
|| !conf
->group_cnt
)
875 spin_lock(&conf
->pending_bios_lock
);
876 bio_list_merge(&tmp
, &conf
->pending_bios
);
877 bio_list_init(&conf
->pending_bios
);
878 spin_unlock(&conf
->pending_bios_lock
);
880 while ((bio
= bio_list_pop(&tmp
)))
881 generic_make_request(bio
);
884 static void defer_bio_issue(struct r5conf
*conf
, struct bio
*bio
)
887 * change group_cnt will drain all bios, so this is safe
889 * A read generally means a read-modify-write, which usually means a
890 * randwrite, so we don't delay it
892 if (!conf
->batch_bio_dispatch
|| !conf
->group_cnt
||
893 bio_op(bio
) == REQ_OP_READ
) {
894 generic_make_request(bio
);
897 spin_lock(&conf
->pending_bios_lock
);
898 bio_list_add(&conf
->pending_bios
, bio
);
899 spin_unlock(&conf
->pending_bios_lock
);
900 md_wakeup_thread(conf
->mddev
->thread
);
904 raid5_end_read_request(struct bio
*bi
);
906 raid5_end_write_request(struct bio
*bi
);
908 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
910 struct r5conf
*conf
= sh
->raid_conf
;
911 int i
, disks
= sh
->disks
;
912 struct stripe_head
*head_sh
= sh
;
916 if (!test_bit(STRIPE_R5C_CACHING
, &sh
->state
)) {
917 /* writing out phase */
918 if (s
->waiting_extra_page
)
920 if (r5l_write_stripe(conf
->log
, sh
) == 0)
922 } else { /* caching phase */
923 if (test_bit(STRIPE_LOG_TRAPPED
, &sh
->state
)) {
924 r5c_cache_data(conf
->log
, sh
, s
);
929 for (i
= disks
; i
--; ) {
930 int op
, op_flags
= 0;
931 int replace_only
= 0;
932 struct bio
*bi
, *rbi
;
933 struct md_rdev
*rdev
, *rrdev
= NULL
;
936 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
938 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
940 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
942 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
944 else if (test_and_clear_bit(R5_WantReplace
,
945 &sh
->dev
[i
].flags
)) {
950 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
951 op_flags
|= REQ_SYNC
;
954 bi
= &sh
->dev
[i
].req
;
955 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
958 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
959 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
960 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
965 if (op_is_write(op
)) {
969 /* We raced and saw duplicates */
972 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
977 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
980 atomic_inc(&rdev
->nr_pending
);
981 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
984 atomic_inc(&rrdev
->nr_pending
);
987 /* We have already checked bad blocks for reads. Now
988 * need to check for writes. We never accept write errors
989 * on the replacement, so we don't to check rrdev.
991 while (op_is_write(op
) && rdev
&&
992 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
995 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
996 &first_bad
, &bad_sectors
);
1001 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1002 if (!conf
->mddev
->external
&&
1003 conf
->mddev
->sb_flags
) {
1004 /* It is very unlikely, but we might
1005 * still need to write out the
1006 * bad block log - better give it
1008 md_check_recovery(conf
->mddev
);
1011 * Because md_wait_for_blocked_rdev
1012 * will dec nr_pending, we must
1013 * increment it first.
1015 atomic_inc(&rdev
->nr_pending
);
1016 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
1018 /* Acknowledged bad block - skip the write */
1019 rdev_dec_pending(rdev
, conf
->mddev
);
1025 if (s
->syncing
|| s
->expanding
|| s
->expanded
1027 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
1029 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1031 bi
->bi_bdev
= rdev
->bdev
;
1032 bio_set_op_attrs(bi
, op
, op_flags
);
1033 bi
->bi_end_io
= op_is_write(op
)
1034 ? raid5_end_write_request
1035 : raid5_end_read_request
;
1036 bi
->bi_private
= sh
;
1038 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1039 __func__
, (unsigned long long)sh
->sector
,
1041 atomic_inc(&sh
->count
);
1043 atomic_inc(&head_sh
->count
);
1044 if (use_new_offset(conf
, sh
))
1045 bi
->bi_iter
.bi_sector
= (sh
->sector
1046 + rdev
->new_data_offset
);
1048 bi
->bi_iter
.bi_sector
= (sh
->sector
1049 + rdev
->data_offset
);
1050 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1051 bi
->bi_opf
|= REQ_NOMERGE
;
1053 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1054 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1056 if (!op_is_write(op
) &&
1057 test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
1059 * issuing read for a page in journal, this
1060 * must be preparing for prexor in rmw; read
1061 * the data into orig_page
1063 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].orig_page
;
1065 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1067 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1068 bi
->bi_io_vec
[0].bv_offset
= 0;
1069 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1071 * If this is discard request, set bi_vcnt 0. We don't
1072 * want to confuse SCSI because SCSI will replace payload
1074 if (op
== REQ_OP_DISCARD
)
1077 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1079 if (conf
->mddev
->gendisk
)
1080 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
1081 bi
, disk_devt(conf
->mddev
->gendisk
),
1083 defer_bio_issue(conf
, bi
);
1086 if (s
->syncing
|| s
->expanding
|| s
->expanded
1088 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1090 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1092 rbi
->bi_bdev
= rrdev
->bdev
;
1093 bio_set_op_attrs(rbi
, op
, op_flags
);
1094 BUG_ON(!op_is_write(op
));
1095 rbi
->bi_end_io
= raid5_end_write_request
;
1096 rbi
->bi_private
= sh
;
1098 pr_debug("%s: for %llu schedule op %d on "
1099 "replacement disc %d\n",
1100 __func__
, (unsigned long long)sh
->sector
,
1102 atomic_inc(&sh
->count
);
1104 atomic_inc(&head_sh
->count
);
1105 if (use_new_offset(conf
, sh
))
1106 rbi
->bi_iter
.bi_sector
= (sh
->sector
1107 + rrdev
->new_data_offset
);
1109 rbi
->bi_iter
.bi_sector
= (sh
->sector
1110 + rrdev
->data_offset
);
1111 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1112 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1113 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1115 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1116 rbi
->bi_io_vec
[0].bv_offset
= 0;
1117 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1119 * If this is discard request, set bi_vcnt 0. We don't
1120 * want to confuse SCSI because SCSI will replace payload
1122 if (op
== REQ_OP_DISCARD
)
1124 if (conf
->mddev
->gendisk
)
1125 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
1126 rbi
, disk_devt(conf
->mddev
->gendisk
),
1128 defer_bio_issue(conf
, rbi
);
1130 if (!rdev
&& !rrdev
) {
1131 if (op_is_write(op
))
1132 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1133 pr_debug("skip op %d on disc %d for sector %llu\n",
1134 bi
->bi_opf
, i
, (unsigned long long)sh
->sector
);
1135 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1136 set_bit(STRIPE_HANDLE
, &sh
->state
);
1139 if (!head_sh
->batch_head
)
1141 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1148 static struct dma_async_tx_descriptor
*
1149 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1150 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1151 struct stripe_head
*sh
, int no_skipcopy
)
1154 struct bvec_iter iter
;
1155 struct page
*bio_page
;
1157 struct async_submit_ctl submit
;
1158 enum async_tx_flags flags
= 0;
1160 if (bio
->bi_iter
.bi_sector
>= sector
)
1161 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1163 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1166 flags
|= ASYNC_TX_FENCE
;
1167 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1169 bio_for_each_segment(bvl
, bio
, iter
) {
1170 int len
= bvl
.bv_len
;
1174 if (page_offset
< 0) {
1175 b_offset
= -page_offset
;
1176 page_offset
+= b_offset
;
1180 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1181 clen
= STRIPE_SIZE
- page_offset
;
1186 b_offset
+= bvl
.bv_offset
;
1187 bio_page
= bvl
.bv_page
;
1189 if (sh
->raid_conf
->skip_copy
&&
1190 b_offset
== 0 && page_offset
== 0 &&
1191 clen
== STRIPE_SIZE
&&
1195 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1196 b_offset
, clen
, &submit
);
1198 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1199 page_offset
, clen
, &submit
);
1201 /* chain the operations */
1202 submit
.depend_tx
= tx
;
1204 if (clen
< len
) /* hit end of page */
1212 static void ops_complete_biofill(void *stripe_head_ref
)
1214 struct stripe_head
*sh
= stripe_head_ref
;
1215 struct bio_list return_bi
= BIO_EMPTY_LIST
;
1218 pr_debug("%s: stripe %llu\n", __func__
,
1219 (unsigned long long)sh
->sector
);
1221 /* clear completed biofills */
1222 for (i
= sh
->disks
; i
--; ) {
1223 struct r5dev
*dev
= &sh
->dev
[i
];
1225 /* acknowledge completion of a biofill operation */
1226 /* and check if we need to reply to a read request,
1227 * new R5_Wantfill requests are held off until
1228 * !STRIPE_BIOFILL_RUN
1230 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1231 struct bio
*rbi
, *rbi2
;
1236 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1237 dev
->sector
+ STRIPE_SECTORS
) {
1238 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1239 if (!raid5_dec_bi_active_stripes(rbi
))
1240 bio_list_add(&return_bi
, rbi
);
1245 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1247 return_io(&return_bi
);
1249 set_bit(STRIPE_HANDLE
, &sh
->state
);
1250 raid5_release_stripe(sh
);
1253 static void ops_run_biofill(struct stripe_head
*sh
)
1255 struct dma_async_tx_descriptor
*tx
= NULL
;
1256 struct async_submit_ctl submit
;
1259 BUG_ON(sh
->batch_head
);
1260 pr_debug("%s: stripe %llu\n", __func__
,
1261 (unsigned long long)sh
->sector
);
1263 for (i
= sh
->disks
; i
--; ) {
1264 struct r5dev
*dev
= &sh
->dev
[i
];
1265 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1267 spin_lock_irq(&sh
->stripe_lock
);
1268 dev
->read
= rbi
= dev
->toread
;
1270 spin_unlock_irq(&sh
->stripe_lock
);
1271 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1272 dev
->sector
+ STRIPE_SECTORS
) {
1273 tx
= async_copy_data(0, rbi
, &dev
->page
,
1274 dev
->sector
, tx
, sh
, 0);
1275 rbi
= r5_next_bio(rbi
, dev
->sector
);
1280 atomic_inc(&sh
->count
);
1281 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1282 async_trigger_callback(&submit
);
1285 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1292 tgt
= &sh
->dev
[target
];
1293 set_bit(R5_UPTODATE
, &tgt
->flags
);
1294 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1295 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1298 static void ops_complete_compute(void *stripe_head_ref
)
1300 struct stripe_head
*sh
= stripe_head_ref
;
1302 pr_debug("%s: stripe %llu\n", __func__
,
1303 (unsigned long long)sh
->sector
);
1305 /* mark the computed target(s) as uptodate */
1306 mark_target_uptodate(sh
, sh
->ops
.target
);
1307 mark_target_uptodate(sh
, sh
->ops
.target2
);
1309 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1310 if (sh
->check_state
== check_state_compute_run
)
1311 sh
->check_state
= check_state_compute_result
;
1312 set_bit(STRIPE_HANDLE
, &sh
->state
);
1313 raid5_release_stripe(sh
);
1316 /* return a pointer to the address conversion region of the scribble buffer */
1317 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1318 struct raid5_percpu
*percpu
, int i
)
1322 addr
= flex_array_get(percpu
->scribble
, i
);
1323 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1326 /* return a pointer to the address conversion region of the scribble buffer */
1327 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1331 addr
= flex_array_get(percpu
->scribble
, i
);
1335 static struct dma_async_tx_descriptor
*
1336 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1338 int disks
= sh
->disks
;
1339 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1340 int target
= sh
->ops
.target
;
1341 struct r5dev
*tgt
= &sh
->dev
[target
];
1342 struct page
*xor_dest
= tgt
->page
;
1344 struct dma_async_tx_descriptor
*tx
;
1345 struct async_submit_ctl submit
;
1348 BUG_ON(sh
->batch_head
);
1350 pr_debug("%s: stripe %llu block: %d\n",
1351 __func__
, (unsigned long long)sh
->sector
, target
);
1352 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1354 for (i
= disks
; i
--; )
1356 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1358 atomic_inc(&sh
->count
);
1360 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1361 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1362 if (unlikely(count
== 1))
1363 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1365 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1370 /* set_syndrome_sources - populate source buffers for gen_syndrome
1371 * @srcs - (struct page *) array of size sh->disks
1372 * @sh - stripe_head to parse
1374 * Populates srcs in proper layout order for the stripe and returns the
1375 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1376 * destination buffer is recorded in srcs[count] and the Q destination
1377 * is recorded in srcs[count+1]].
1379 static int set_syndrome_sources(struct page
**srcs
,
1380 struct stripe_head
*sh
,
1383 int disks
= sh
->disks
;
1384 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1385 int d0_idx
= raid6_d0(sh
);
1389 for (i
= 0; i
< disks
; i
++)
1395 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1396 struct r5dev
*dev
= &sh
->dev
[i
];
1398 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1399 (srctype
== SYNDROME_SRC_ALL
) ||
1400 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1401 (test_bit(R5_Wantdrain
, &dev
->flags
) ||
1402 test_bit(R5_InJournal
, &dev
->flags
))) ||
1403 (srctype
== SYNDROME_SRC_WRITTEN
&&
1405 if (test_bit(R5_InJournal
, &dev
->flags
))
1406 srcs
[slot
] = sh
->dev
[i
].orig_page
;
1408 srcs
[slot
] = sh
->dev
[i
].page
;
1410 i
= raid6_next_disk(i
, disks
);
1411 } while (i
!= d0_idx
);
1413 return syndrome_disks
;
1416 static struct dma_async_tx_descriptor
*
1417 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1419 int disks
= sh
->disks
;
1420 struct page
**blocks
= to_addr_page(percpu
, 0);
1422 int qd_idx
= sh
->qd_idx
;
1423 struct dma_async_tx_descriptor
*tx
;
1424 struct async_submit_ctl submit
;
1430 BUG_ON(sh
->batch_head
);
1431 if (sh
->ops
.target
< 0)
1432 target
= sh
->ops
.target2
;
1433 else if (sh
->ops
.target2
< 0)
1434 target
= sh
->ops
.target
;
1436 /* we should only have one valid target */
1439 pr_debug("%s: stripe %llu block: %d\n",
1440 __func__
, (unsigned long long)sh
->sector
, target
);
1442 tgt
= &sh
->dev
[target
];
1443 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1446 atomic_inc(&sh
->count
);
1448 if (target
== qd_idx
) {
1449 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1450 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1451 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1452 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1453 ops_complete_compute
, sh
,
1454 to_addr_conv(sh
, percpu
, 0));
1455 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1457 /* Compute any data- or p-drive using XOR */
1459 for (i
= disks
; i
-- ; ) {
1460 if (i
== target
|| i
== qd_idx
)
1462 blocks
[count
++] = sh
->dev
[i
].page
;
1465 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1466 NULL
, ops_complete_compute
, sh
,
1467 to_addr_conv(sh
, percpu
, 0));
1468 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1474 static struct dma_async_tx_descriptor
*
1475 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1477 int i
, count
, disks
= sh
->disks
;
1478 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1479 int d0_idx
= raid6_d0(sh
);
1480 int faila
= -1, failb
= -1;
1481 int target
= sh
->ops
.target
;
1482 int target2
= sh
->ops
.target2
;
1483 struct r5dev
*tgt
= &sh
->dev
[target
];
1484 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1485 struct dma_async_tx_descriptor
*tx
;
1486 struct page
**blocks
= to_addr_page(percpu
, 0);
1487 struct async_submit_ctl submit
;
1489 BUG_ON(sh
->batch_head
);
1490 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1491 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1492 BUG_ON(target
< 0 || target2
< 0);
1493 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1494 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1496 /* we need to open-code set_syndrome_sources to handle the
1497 * slot number conversion for 'faila' and 'failb'
1499 for (i
= 0; i
< disks
; i
++)
1504 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1506 blocks
[slot
] = sh
->dev
[i
].page
;
1512 i
= raid6_next_disk(i
, disks
);
1513 } while (i
!= d0_idx
);
1515 BUG_ON(faila
== failb
);
1518 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1519 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1521 atomic_inc(&sh
->count
);
1523 if (failb
== syndrome_disks
+1) {
1524 /* Q disk is one of the missing disks */
1525 if (faila
== syndrome_disks
) {
1526 /* Missing P+Q, just recompute */
1527 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1528 ops_complete_compute
, sh
,
1529 to_addr_conv(sh
, percpu
, 0));
1530 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1531 STRIPE_SIZE
, &submit
);
1535 int qd_idx
= sh
->qd_idx
;
1537 /* Missing D+Q: recompute D from P, then recompute Q */
1538 if (target
== qd_idx
)
1539 data_target
= target2
;
1541 data_target
= target
;
1544 for (i
= disks
; i
-- ; ) {
1545 if (i
== data_target
|| i
== qd_idx
)
1547 blocks
[count
++] = sh
->dev
[i
].page
;
1549 dest
= sh
->dev
[data_target
].page
;
1550 init_async_submit(&submit
,
1551 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1553 to_addr_conv(sh
, percpu
, 0));
1554 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1557 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1558 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1559 ops_complete_compute
, sh
,
1560 to_addr_conv(sh
, percpu
, 0));
1561 return async_gen_syndrome(blocks
, 0, count
+2,
1562 STRIPE_SIZE
, &submit
);
1565 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1566 ops_complete_compute
, sh
,
1567 to_addr_conv(sh
, percpu
, 0));
1568 if (failb
== syndrome_disks
) {
1569 /* We're missing D+P. */
1570 return async_raid6_datap_recov(syndrome_disks
+2,
1574 /* We're missing D+D. */
1575 return async_raid6_2data_recov(syndrome_disks
+2,
1576 STRIPE_SIZE
, faila
, failb
,
1582 static void ops_complete_prexor(void *stripe_head_ref
)
1584 struct stripe_head
*sh
= stripe_head_ref
;
1586 pr_debug("%s: stripe %llu\n", __func__
,
1587 (unsigned long long)sh
->sector
);
1589 if (r5c_is_writeback(sh
->raid_conf
->log
))
1591 * raid5-cache write back uses orig_page during prexor.
1592 * After prexor, it is time to free orig_page
1594 r5c_release_extra_page(sh
);
1597 static struct dma_async_tx_descriptor
*
1598 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1599 struct dma_async_tx_descriptor
*tx
)
1601 int disks
= sh
->disks
;
1602 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1603 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1604 struct async_submit_ctl submit
;
1606 /* existing parity data subtracted */
1607 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1609 BUG_ON(sh
->batch_head
);
1610 pr_debug("%s: stripe %llu\n", __func__
,
1611 (unsigned long long)sh
->sector
);
1613 for (i
= disks
; i
--; ) {
1614 struct r5dev
*dev
= &sh
->dev
[i
];
1615 /* Only process blocks that are known to be uptodate */
1616 if (test_bit(R5_InJournal
, &dev
->flags
))
1617 xor_srcs
[count
++] = dev
->orig_page
;
1618 else if (test_bit(R5_Wantdrain
, &dev
->flags
))
1619 xor_srcs
[count
++] = dev
->page
;
1622 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1623 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1624 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1629 static struct dma_async_tx_descriptor
*
1630 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1631 struct dma_async_tx_descriptor
*tx
)
1633 struct page
**blocks
= to_addr_page(percpu
, 0);
1635 struct async_submit_ctl submit
;
1637 pr_debug("%s: stripe %llu\n", __func__
,
1638 (unsigned long long)sh
->sector
);
1640 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1642 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1643 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1644 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1649 static struct dma_async_tx_descriptor
*
1650 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1652 struct r5conf
*conf
= sh
->raid_conf
;
1653 int disks
= sh
->disks
;
1655 struct stripe_head
*head_sh
= sh
;
1657 pr_debug("%s: stripe %llu\n", __func__
,
1658 (unsigned long long)sh
->sector
);
1660 for (i
= disks
; i
--; ) {
1665 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1671 * clear R5_InJournal, so when rewriting a page in
1672 * journal, it is not skipped by r5l_log_stripe()
1674 clear_bit(R5_InJournal
, &dev
->flags
);
1675 spin_lock_irq(&sh
->stripe_lock
);
1676 chosen
= dev
->towrite
;
1677 dev
->towrite
= NULL
;
1678 sh
->overwrite_disks
= 0;
1679 BUG_ON(dev
->written
);
1680 wbi
= dev
->written
= chosen
;
1681 spin_unlock_irq(&sh
->stripe_lock
);
1682 WARN_ON(dev
->page
!= dev
->orig_page
);
1684 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1685 dev
->sector
+ STRIPE_SECTORS
) {
1686 if (wbi
->bi_opf
& REQ_FUA
)
1687 set_bit(R5_WantFUA
, &dev
->flags
);
1688 if (wbi
->bi_opf
& REQ_SYNC
)
1689 set_bit(R5_SyncIO
, &dev
->flags
);
1690 if (bio_op(wbi
) == REQ_OP_DISCARD
)
1691 set_bit(R5_Discard
, &dev
->flags
);
1693 tx
= async_copy_data(1, wbi
, &dev
->page
,
1694 dev
->sector
, tx
, sh
,
1695 r5c_is_writeback(conf
->log
));
1696 if (dev
->page
!= dev
->orig_page
&&
1697 !r5c_is_writeback(conf
->log
)) {
1698 set_bit(R5_SkipCopy
, &dev
->flags
);
1699 clear_bit(R5_UPTODATE
, &dev
->flags
);
1700 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1703 wbi
= r5_next_bio(wbi
, dev
->sector
);
1706 if (head_sh
->batch_head
) {
1707 sh
= list_first_entry(&sh
->batch_list
,
1720 static void ops_complete_reconstruct(void *stripe_head_ref
)
1722 struct stripe_head
*sh
= stripe_head_ref
;
1723 int disks
= sh
->disks
;
1724 int pd_idx
= sh
->pd_idx
;
1725 int qd_idx
= sh
->qd_idx
;
1727 bool fua
= false, sync
= false, discard
= false;
1729 pr_debug("%s: stripe %llu\n", __func__
,
1730 (unsigned long long)sh
->sector
);
1732 for (i
= disks
; i
--; ) {
1733 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1734 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1735 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1738 for (i
= disks
; i
--; ) {
1739 struct r5dev
*dev
= &sh
->dev
[i
];
1741 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1742 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1743 set_bit(R5_UPTODATE
, &dev
->flags
);
1745 set_bit(R5_WantFUA
, &dev
->flags
);
1747 set_bit(R5_SyncIO
, &dev
->flags
);
1751 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1752 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1753 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1754 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1756 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1757 sh
->reconstruct_state
= reconstruct_state_result
;
1760 set_bit(STRIPE_HANDLE
, &sh
->state
);
1761 raid5_release_stripe(sh
);
1765 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1766 struct dma_async_tx_descriptor
*tx
)
1768 int disks
= sh
->disks
;
1769 struct page
**xor_srcs
;
1770 struct async_submit_ctl submit
;
1771 int count
, pd_idx
= sh
->pd_idx
, i
;
1772 struct page
*xor_dest
;
1774 unsigned long flags
;
1776 struct stripe_head
*head_sh
= sh
;
1779 pr_debug("%s: stripe %llu\n", __func__
,
1780 (unsigned long long)sh
->sector
);
1782 for (i
= 0; i
< sh
->disks
; i
++) {
1785 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1788 if (i
>= sh
->disks
) {
1789 atomic_inc(&sh
->count
);
1790 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1791 ops_complete_reconstruct(sh
);
1796 xor_srcs
= to_addr_page(percpu
, j
);
1797 /* check if prexor is active which means only process blocks
1798 * that are part of a read-modify-write (written)
1800 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1802 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1803 for (i
= disks
; i
--; ) {
1804 struct r5dev
*dev
= &sh
->dev
[i
];
1805 if (head_sh
->dev
[i
].written
||
1806 test_bit(R5_InJournal
, &head_sh
->dev
[i
].flags
))
1807 xor_srcs
[count
++] = dev
->page
;
1810 xor_dest
= sh
->dev
[pd_idx
].page
;
1811 for (i
= disks
; i
--; ) {
1812 struct r5dev
*dev
= &sh
->dev
[i
];
1814 xor_srcs
[count
++] = dev
->page
;
1818 /* 1/ if we prexor'd then the dest is reused as a source
1819 * 2/ if we did not prexor then we are redoing the parity
1820 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1821 * for the synchronous xor case
1823 last_stripe
= !head_sh
->batch_head
||
1824 list_first_entry(&sh
->batch_list
,
1825 struct stripe_head
, batch_list
) == head_sh
;
1827 flags
= ASYNC_TX_ACK
|
1828 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1830 atomic_inc(&head_sh
->count
);
1831 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1832 to_addr_conv(sh
, percpu
, j
));
1834 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1835 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1836 to_addr_conv(sh
, percpu
, j
));
1839 if (unlikely(count
== 1))
1840 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1842 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1845 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1852 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1853 struct dma_async_tx_descriptor
*tx
)
1855 struct async_submit_ctl submit
;
1856 struct page
**blocks
;
1857 int count
, i
, j
= 0;
1858 struct stripe_head
*head_sh
= sh
;
1861 unsigned long txflags
;
1863 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1865 for (i
= 0; i
< sh
->disks
; i
++) {
1866 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1868 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1871 if (i
>= sh
->disks
) {
1872 atomic_inc(&sh
->count
);
1873 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1874 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1875 ops_complete_reconstruct(sh
);
1880 blocks
= to_addr_page(percpu
, j
);
1882 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1883 synflags
= SYNDROME_SRC_WRITTEN
;
1884 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1886 synflags
= SYNDROME_SRC_ALL
;
1887 txflags
= ASYNC_TX_ACK
;
1890 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1891 last_stripe
= !head_sh
->batch_head
||
1892 list_first_entry(&sh
->batch_list
,
1893 struct stripe_head
, batch_list
) == head_sh
;
1896 atomic_inc(&head_sh
->count
);
1897 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1898 head_sh
, to_addr_conv(sh
, percpu
, j
));
1900 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1901 to_addr_conv(sh
, percpu
, j
));
1902 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1905 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1911 static void ops_complete_check(void *stripe_head_ref
)
1913 struct stripe_head
*sh
= stripe_head_ref
;
1915 pr_debug("%s: stripe %llu\n", __func__
,
1916 (unsigned long long)sh
->sector
);
1918 sh
->check_state
= check_state_check_result
;
1919 set_bit(STRIPE_HANDLE
, &sh
->state
);
1920 raid5_release_stripe(sh
);
1923 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1925 int disks
= sh
->disks
;
1926 int pd_idx
= sh
->pd_idx
;
1927 int qd_idx
= sh
->qd_idx
;
1928 struct page
*xor_dest
;
1929 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1930 struct dma_async_tx_descriptor
*tx
;
1931 struct async_submit_ctl submit
;
1935 pr_debug("%s: stripe %llu\n", __func__
,
1936 (unsigned long long)sh
->sector
);
1938 BUG_ON(sh
->batch_head
);
1940 xor_dest
= sh
->dev
[pd_idx
].page
;
1941 xor_srcs
[count
++] = xor_dest
;
1942 for (i
= disks
; i
--; ) {
1943 if (i
== pd_idx
|| i
== qd_idx
)
1945 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1948 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1949 to_addr_conv(sh
, percpu
, 0));
1950 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1951 &sh
->ops
.zero_sum_result
, &submit
);
1953 atomic_inc(&sh
->count
);
1954 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1955 tx
= async_trigger_callback(&submit
);
1958 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1960 struct page
**srcs
= to_addr_page(percpu
, 0);
1961 struct async_submit_ctl submit
;
1964 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1965 (unsigned long long)sh
->sector
, checkp
);
1967 BUG_ON(sh
->batch_head
);
1968 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
1972 atomic_inc(&sh
->count
);
1973 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1974 sh
, to_addr_conv(sh
, percpu
, 0));
1975 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1976 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1979 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1981 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1982 struct dma_async_tx_descriptor
*tx
= NULL
;
1983 struct r5conf
*conf
= sh
->raid_conf
;
1984 int level
= conf
->level
;
1985 struct raid5_percpu
*percpu
;
1989 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1990 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1991 ops_run_biofill(sh
);
1995 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1997 tx
= ops_run_compute5(sh
, percpu
);
1999 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
2000 tx
= ops_run_compute6_1(sh
, percpu
);
2002 tx
= ops_run_compute6_2(sh
, percpu
);
2004 /* terminate the chain if reconstruct is not set to be run */
2005 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
2009 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
2011 tx
= ops_run_prexor5(sh
, percpu
, tx
);
2013 tx
= ops_run_prexor6(sh
, percpu
, tx
);
2016 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
2017 tx
= ops_run_biodrain(sh
, tx
);
2021 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
2023 ops_run_reconstruct5(sh
, percpu
, tx
);
2025 ops_run_reconstruct6(sh
, percpu
, tx
);
2028 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
2029 if (sh
->check_state
== check_state_run
)
2030 ops_run_check_p(sh
, percpu
);
2031 else if (sh
->check_state
== check_state_run_q
)
2032 ops_run_check_pq(sh
, percpu
, 0);
2033 else if (sh
->check_state
== check_state_run_pq
)
2034 ops_run_check_pq(sh
, percpu
, 1);
2039 if (overlap_clear
&& !sh
->batch_head
)
2040 for (i
= disks
; i
--; ) {
2041 struct r5dev
*dev
= &sh
->dev
[i
];
2042 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
2043 wake_up(&sh
->raid_conf
->wait_for_overlap
);
2048 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
,
2051 struct stripe_head
*sh
;
2054 sh
= kmem_cache_zalloc(sc
, gfp
);
2056 spin_lock_init(&sh
->stripe_lock
);
2057 spin_lock_init(&sh
->batch_lock
);
2058 INIT_LIST_HEAD(&sh
->batch_list
);
2059 INIT_LIST_HEAD(&sh
->lru
);
2060 INIT_LIST_HEAD(&sh
->r5c
);
2061 INIT_LIST_HEAD(&sh
->log_list
);
2062 atomic_set(&sh
->count
, 1);
2063 sh
->log_start
= MaxSector
;
2064 for (i
= 0; i
< disks
; i
++) {
2065 struct r5dev
*dev
= &sh
->dev
[i
];
2067 bio_init(&dev
->req
, &dev
->vec
, 1);
2068 bio_init(&dev
->rreq
, &dev
->rvec
, 1);
2073 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2075 struct stripe_head
*sh
;
2077 sh
= alloc_stripe(conf
->slab_cache
, gfp
, conf
->pool_size
);
2081 sh
->raid_conf
= conf
;
2083 if (grow_buffers(sh
, gfp
)) {
2085 kmem_cache_free(conf
->slab_cache
, sh
);
2088 sh
->hash_lock_index
=
2089 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2090 /* we just created an active stripe so... */
2091 atomic_inc(&conf
->active_stripes
);
2093 raid5_release_stripe(sh
);
2094 conf
->max_nr_stripes
++;
2098 static int grow_stripes(struct r5conf
*conf
, int num
)
2100 struct kmem_cache
*sc
;
2101 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2103 if (conf
->mddev
->gendisk
)
2104 sprintf(conf
->cache_name
[0],
2105 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2107 sprintf(conf
->cache_name
[0],
2108 "raid%d-%p", conf
->level
, conf
->mddev
);
2109 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2111 conf
->active_name
= 0;
2112 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2113 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2117 conf
->slab_cache
= sc
;
2118 conf
->pool_size
= devs
;
2120 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2127 * scribble_len - return the required size of the scribble region
2128 * @num - total number of disks in the array
2130 * The size must be enough to contain:
2131 * 1/ a struct page pointer for each device in the array +2
2132 * 2/ room to convert each entry in (1) to its corresponding dma
2133 * (dma_map_page()) or page (page_address()) address.
2135 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2136 * calculate over all devices (not just the data blocks), using zeros in place
2137 * of the P and Q blocks.
2139 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2141 struct flex_array
*ret
;
2144 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2145 ret
= flex_array_alloc(len
, cnt
, flags
);
2148 /* always prealloc all elements, so no locking is required */
2149 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2150 flex_array_free(ret
);
2156 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2162 * Never shrink. And mddev_suspend() could deadlock if this is called
2163 * from raid5d. In that case, scribble_disks and scribble_sectors
2164 * should equal to new_disks and new_sectors
2166 if (conf
->scribble_disks
>= new_disks
&&
2167 conf
->scribble_sectors
>= new_sectors
)
2169 mddev_suspend(conf
->mddev
);
2171 for_each_present_cpu(cpu
) {
2172 struct raid5_percpu
*percpu
;
2173 struct flex_array
*scribble
;
2175 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2176 scribble
= scribble_alloc(new_disks
,
2177 new_sectors
/ STRIPE_SECTORS
,
2181 flex_array_free(percpu
->scribble
);
2182 percpu
->scribble
= scribble
;
2189 mddev_resume(conf
->mddev
);
2191 conf
->scribble_disks
= new_disks
;
2192 conf
->scribble_sectors
= new_sectors
;
2197 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2199 /* Make all the stripes able to hold 'newsize' devices.
2200 * New slots in each stripe get 'page' set to a new page.
2202 * This happens in stages:
2203 * 1/ create a new kmem_cache and allocate the required number of
2205 * 2/ gather all the old stripe_heads and transfer the pages across
2206 * to the new stripe_heads. This will have the side effect of
2207 * freezing the array as once all stripe_heads have been collected,
2208 * no IO will be possible. Old stripe heads are freed once their
2209 * pages have been transferred over, and the old kmem_cache is
2210 * freed when all stripes are done.
2211 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2212 * we simple return a failre status - no need to clean anything up.
2213 * 4/ allocate new pages for the new slots in the new stripe_heads.
2214 * If this fails, we don't bother trying the shrink the
2215 * stripe_heads down again, we just leave them as they are.
2216 * As each stripe_head is processed the new one is released into
2219 * Once step2 is started, we cannot afford to wait for a write,
2220 * so we use GFP_NOIO allocations.
2222 struct stripe_head
*osh
, *nsh
;
2223 LIST_HEAD(newstripes
);
2224 struct disk_info
*ndisks
;
2226 struct kmem_cache
*sc
;
2230 if (newsize
<= conf
->pool_size
)
2231 return 0; /* never bother to shrink */
2233 err
= md_allow_write(conf
->mddev
);
2238 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2239 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2244 /* Need to ensure auto-resizing doesn't interfere */
2245 mutex_lock(&conf
->cache_size_mutex
);
2247 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2248 nsh
= alloc_stripe(sc
, GFP_KERNEL
, newsize
);
2252 nsh
->raid_conf
= conf
;
2253 list_add(&nsh
->lru
, &newstripes
);
2256 /* didn't get enough, give up */
2257 while (!list_empty(&newstripes
)) {
2258 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2259 list_del(&nsh
->lru
);
2260 kmem_cache_free(sc
, nsh
);
2262 kmem_cache_destroy(sc
);
2263 mutex_unlock(&conf
->cache_size_mutex
);
2266 /* Step 2 - Must use GFP_NOIO now.
2267 * OK, we have enough stripes, start collecting inactive
2268 * stripes and copying them over
2272 list_for_each_entry(nsh
, &newstripes
, lru
) {
2273 lock_device_hash_lock(conf
, hash
);
2274 wait_event_cmd(conf
->wait_for_stripe
,
2275 !list_empty(conf
->inactive_list
+ hash
),
2276 unlock_device_hash_lock(conf
, hash
),
2277 lock_device_hash_lock(conf
, hash
));
2278 osh
= get_free_stripe(conf
, hash
);
2279 unlock_device_hash_lock(conf
, hash
);
2281 for(i
=0; i
<conf
->pool_size
; i
++) {
2282 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2283 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2285 nsh
->hash_lock_index
= hash
;
2286 kmem_cache_free(conf
->slab_cache
, osh
);
2288 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2289 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2294 kmem_cache_destroy(conf
->slab_cache
);
2297 * At this point, we are holding all the stripes so the array
2298 * is completely stalled, so now is a good time to resize
2299 * conf->disks and the scribble region
2301 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2303 for (i
= 0; i
< conf
->pool_size
; i
++)
2304 ndisks
[i
] = conf
->disks
[i
];
2306 for (i
= conf
->pool_size
; i
< newsize
; i
++) {
2307 ndisks
[i
].extra_page
= alloc_page(GFP_NOIO
);
2308 if (!ndisks
[i
].extra_page
)
2313 for (i
= conf
->pool_size
; i
< newsize
; i
++)
2314 if (ndisks
[i
].extra_page
)
2315 put_page(ndisks
[i
].extra_page
);
2319 conf
->disks
= ndisks
;
2324 mutex_unlock(&conf
->cache_size_mutex
);
2325 /* Step 4, return new stripes to service */
2326 while(!list_empty(&newstripes
)) {
2327 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2328 list_del_init(&nsh
->lru
);
2330 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2331 if (nsh
->dev
[i
].page
== NULL
) {
2332 struct page
*p
= alloc_page(GFP_NOIO
);
2333 nsh
->dev
[i
].page
= p
;
2334 nsh
->dev
[i
].orig_page
= p
;
2338 raid5_release_stripe(nsh
);
2340 /* critical section pass, GFP_NOIO no longer needed */
2342 conf
->slab_cache
= sc
;
2343 conf
->active_name
= 1-conf
->active_name
;
2345 conf
->pool_size
= newsize
;
2349 static int drop_one_stripe(struct r5conf
*conf
)
2351 struct stripe_head
*sh
;
2352 int hash
= (conf
->max_nr_stripes
- 1) & STRIPE_HASH_LOCKS_MASK
;
2354 spin_lock_irq(conf
->hash_locks
+ hash
);
2355 sh
= get_free_stripe(conf
, hash
);
2356 spin_unlock_irq(conf
->hash_locks
+ hash
);
2359 BUG_ON(atomic_read(&sh
->count
));
2361 kmem_cache_free(conf
->slab_cache
, sh
);
2362 atomic_dec(&conf
->active_stripes
);
2363 conf
->max_nr_stripes
--;
2367 static void shrink_stripes(struct r5conf
*conf
)
2369 while (conf
->max_nr_stripes
&&
2370 drop_one_stripe(conf
))
2373 kmem_cache_destroy(conf
->slab_cache
);
2374 conf
->slab_cache
= NULL
;
2377 static void raid5_end_read_request(struct bio
* bi
)
2379 struct stripe_head
*sh
= bi
->bi_private
;
2380 struct r5conf
*conf
= sh
->raid_conf
;
2381 int disks
= sh
->disks
, i
;
2382 char b
[BDEVNAME_SIZE
];
2383 struct md_rdev
*rdev
= NULL
;
2386 for (i
=0 ; i
<disks
; i
++)
2387 if (bi
== &sh
->dev
[i
].req
)
2390 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2391 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2398 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2399 /* If replacement finished while this request was outstanding,
2400 * 'replacement' might be NULL already.
2401 * In that case it moved down to 'rdev'.
2402 * rdev is not removed until all requests are finished.
2404 rdev
= conf
->disks
[i
].replacement
;
2406 rdev
= conf
->disks
[i
].rdev
;
2408 if (use_new_offset(conf
, sh
))
2409 s
= sh
->sector
+ rdev
->new_data_offset
;
2411 s
= sh
->sector
+ rdev
->data_offset
;
2412 if (!bi
->bi_error
) {
2413 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2414 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2415 /* Note that this cannot happen on a
2416 * replacement device. We just fail those on
2419 pr_info_ratelimited(
2420 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2421 mdname(conf
->mddev
), STRIPE_SECTORS
,
2422 (unsigned long long)s
,
2423 bdevname(rdev
->bdev
, b
));
2424 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2425 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2426 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2427 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2428 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2430 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
2432 * end read for a page in journal, this
2433 * must be preparing for prexor in rmw
2435 set_bit(R5_OrigPageUPTDODATE
, &sh
->dev
[i
].flags
);
2437 if (atomic_read(&rdev
->read_errors
))
2438 atomic_set(&rdev
->read_errors
, 0);
2440 const char *bdn
= bdevname(rdev
->bdev
, b
);
2444 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2445 atomic_inc(&rdev
->read_errors
);
2446 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2447 pr_warn_ratelimited(
2448 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2449 mdname(conf
->mddev
),
2450 (unsigned long long)s
,
2452 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2454 pr_warn_ratelimited(
2455 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2456 mdname(conf
->mddev
),
2457 (unsigned long long)s
,
2459 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2462 pr_warn_ratelimited(
2463 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2464 mdname(conf
->mddev
),
2465 (unsigned long long)s
,
2467 } else if (atomic_read(&rdev
->read_errors
)
2468 > conf
->max_nr_stripes
)
2469 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2470 mdname(conf
->mddev
), bdn
);
2473 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2474 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2477 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2478 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2479 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2481 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2483 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2484 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2486 && test_bit(In_sync
, &rdev
->flags
)
2487 && rdev_set_badblocks(
2488 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2489 md_error(conf
->mddev
, rdev
);
2492 rdev_dec_pending(rdev
, conf
->mddev
);
2494 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2495 set_bit(STRIPE_HANDLE
, &sh
->state
);
2496 raid5_release_stripe(sh
);
2499 static void raid5_end_write_request(struct bio
*bi
)
2501 struct stripe_head
*sh
= bi
->bi_private
;
2502 struct r5conf
*conf
= sh
->raid_conf
;
2503 int disks
= sh
->disks
, i
;
2504 struct md_rdev
*uninitialized_var(rdev
);
2507 int replacement
= 0;
2509 for (i
= 0 ; i
< disks
; i
++) {
2510 if (bi
== &sh
->dev
[i
].req
) {
2511 rdev
= conf
->disks
[i
].rdev
;
2514 if (bi
== &sh
->dev
[i
].rreq
) {
2515 rdev
= conf
->disks
[i
].replacement
;
2519 /* rdev was removed and 'replacement'
2520 * replaced it. rdev is not removed
2521 * until all requests are finished.
2523 rdev
= conf
->disks
[i
].rdev
;
2527 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2528 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2538 md_error(conf
->mddev
, rdev
);
2539 else if (is_badblock(rdev
, sh
->sector
,
2541 &first_bad
, &bad_sectors
))
2542 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2545 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2546 set_bit(WriteErrorSeen
, &rdev
->flags
);
2547 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2548 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2549 set_bit(MD_RECOVERY_NEEDED
,
2550 &rdev
->mddev
->recovery
);
2551 } else if (is_badblock(rdev
, sh
->sector
,
2553 &first_bad
, &bad_sectors
)) {
2554 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2555 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2556 /* That was a successful write so make
2557 * sure it looks like we already did
2560 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2563 rdev_dec_pending(rdev
, conf
->mddev
);
2565 if (sh
->batch_head
&& bi
->bi_error
&& !replacement
)
2566 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2569 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2570 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2571 set_bit(STRIPE_HANDLE
, &sh
->state
);
2572 raid5_release_stripe(sh
);
2574 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2575 raid5_release_stripe(sh
->batch_head
);
2578 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
2580 struct r5dev
*dev
= &sh
->dev
[i
];
2583 dev
->sector
= raid5_compute_blocknr(sh
, i
, previous
);
2586 static void raid5_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2588 char b
[BDEVNAME_SIZE
];
2589 struct r5conf
*conf
= mddev
->private;
2590 unsigned long flags
;
2591 pr_debug("raid456: error called\n");
2593 spin_lock_irqsave(&conf
->device_lock
, flags
);
2594 clear_bit(In_sync
, &rdev
->flags
);
2595 mddev
->degraded
= raid5_calc_degraded(conf
);
2596 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2597 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2599 set_bit(Blocked
, &rdev
->flags
);
2600 set_bit(Faulty
, &rdev
->flags
);
2601 set_mask_bits(&mddev
->sb_flags
, 0,
2602 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
2603 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2604 "md/raid:%s: Operation continuing on %d devices.\n",
2606 bdevname(rdev
->bdev
, b
),
2608 conf
->raid_disks
- mddev
->degraded
);
2609 r5c_update_on_rdev_error(mddev
);
2613 * Input: a 'big' sector number,
2614 * Output: index of the data and parity disk, and the sector # in them.
2616 sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2617 int previous
, int *dd_idx
,
2618 struct stripe_head
*sh
)
2620 sector_t stripe
, stripe2
;
2621 sector_t chunk_number
;
2622 unsigned int chunk_offset
;
2625 sector_t new_sector
;
2626 int algorithm
= previous
? conf
->prev_algo
2628 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2629 : conf
->chunk_sectors
;
2630 int raid_disks
= previous
? conf
->previous_raid_disks
2632 int data_disks
= raid_disks
- conf
->max_degraded
;
2634 /* First compute the information on this sector */
2637 * Compute the chunk number and the sector offset inside the chunk
2639 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2640 chunk_number
= r_sector
;
2643 * Compute the stripe number
2645 stripe
= chunk_number
;
2646 *dd_idx
= sector_div(stripe
, data_disks
);
2649 * Select the parity disk based on the user selected algorithm.
2651 pd_idx
= qd_idx
= -1;
2652 switch(conf
->level
) {
2654 pd_idx
= data_disks
;
2657 switch (algorithm
) {
2658 case ALGORITHM_LEFT_ASYMMETRIC
:
2659 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2660 if (*dd_idx
>= pd_idx
)
2663 case ALGORITHM_RIGHT_ASYMMETRIC
:
2664 pd_idx
= sector_div(stripe2
, raid_disks
);
2665 if (*dd_idx
>= pd_idx
)
2668 case ALGORITHM_LEFT_SYMMETRIC
:
2669 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2670 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2672 case ALGORITHM_RIGHT_SYMMETRIC
:
2673 pd_idx
= sector_div(stripe2
, raid_disks
);
2674 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2676 case ALGORITHM_PARITY_0
:
2680 case ALGORITHM_PARITY_N
:
2681 pd_idx
= data_disks
;
2689 switch (algorithm
) {
2690 case ALGORITHM_LEFT_ASYMMETRIC
:
2691 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2692 qd_idx
= pd_idx
+ 1;
2693 if (pd_idx
== raid_disks
-1) {
2694 (*dd_idx
)++; /* Q D D D P */
2696 } else if (*dd_idx
>= pd_idx
)
2697 (*dd_idx
) += 2; /* D D P Q D */
2699 case ALGORITHM_RIGHT_ASYMMETRIC
:
2700 pd_idx
= sector_div(stripe2
, raid_disks
);
2701 qd_idx
= pd_idx
+ 1;
2702 if (pd_idx
== raid_disks
-1) {
2703 (*dd_idx
)++; /* Q D D D P */
2705 } else if (*dd_idx
>= pd_idx
)
2706 (*dd_idx
) += 2; /* D D P Q D */
2708 case ALGORITHM_LEFT_SYMMETRIC
:
2709 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2710 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2711 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2713 case ALGORITHM_RIGHT_SYMMETRIC
:
2714 pd_idx
= sector_div(stripe2
, raid_disks
);
2715 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2716 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2719 case ALGORITHM_PARITY_0
:
2724 case ALGORITHM_PARITY_N
:
2725 pd_idx
= data_disks
;
2726 qd_idx
= data_disks
+ 1;
2729 case ALGORITHM_ROTATING_ZERO_RESTART
:
2730 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2731 * of blocks for computing Q is different.
2733 pd_idx
= sector_div(stripe2
, raid_disks
);
2734 qd_idx
= pd_idx
+ 1;
2735 if (pd_idx
== raid_disks
-1) {
2736 (*dd_idx
)++; /* Q D D D P */
2738 } else if (*dd_idx
>= pd_idx
)
2739 (*dd_idx
) += 2; /* D D P Q D */
2743 case ALGORITHM_ROTATING_N_RESTART
:
2744 /* Same a left_asymmetric, by first stripe is
2745 * D D D P Q rather than
2749 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2750 qd_idx
= pd_idx
+ 1;
2751 if (pd_idx
== raid_disks
-1) {
2752 (*dd_idx
)++; /* Q D D D P */
2754 } else if (*dd_idx
>= pd_idx
)
2755 (*dd_idx
) += 2; /* D D P Q D */
2759 case ALGORITHM_ROTATING_N_CONTINUE
:
2760 /* Same as left_symmetric but Q is before P */
2761 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2762 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2763 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2767 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2768 /* RAID5 left_asymmetric, with Q on last device */
2769 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2770 if (*dd_idx
>= pd_idx
)
2772 qd_idx
= raid_disks
- 1;
2775 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2776 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2777 if (*dd_idx
>= pd_idx
)
2779 qd_idx
= raid_disks
- 1;
2782 case ALGORITHM_LEFT_SYMMETRIC_6
:
2783 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2784 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2785 qd_idx
= raid_disks
- 1;
2788 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2789 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2790 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2791 qd_idx
= raid_disks
- 1;
2794 case ALGORITHM_PARITY_0_6
:
2797 qd_idx
= raid_disks
- 1;
2807 sh
->pd_idx
= pd_idx
;
2808 sh
->qd_idx
= qd_idx
;
2809 sh
->ddf_layout
= ddf_layout
;
2812 * Finally, compute the new sector number
2814 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2818 sector_t
raid5_compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2820 struct r5conf
*conf
= sh
->raid_conf
;
2821 int raid_disks
= sh
->disks
;
2822 int data_disks
= raid_disks
- conf
->max_degraded
;
2823 sector_t new_sector
= sh
->sector
, check
;
2824 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2825 : conf
->chunk_sectors
;
2826 int algorithm
= previous
? conf
->prev_algo
2830 sector_t chunk_number
;
2831 int dummy1
, dd_idx
= i
;
2833 struct stripe_head sh2
;
2835 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2836 stripe
= new_sector
;
2838 if (i
== sh
->pd_idx
)
2840 switch(conf
->level
) {
2843 switch (algorithm
) {
2844 case ALGORITHM_LEFT_ASYMMETRIC
:
2845 case ALGORITHM_RIGHT_ASYMMETRIC
:
2849 case ALGORITHM_LEFT_SYMMETRIC
:
2850 case ALGORITHM_RIGHT_SYMMETRIC
:
2853 i
-= (sh
->pd_idx
+ 1);
2855 case ALGORITHM_PARITY_0
:
2858 case ALGORITHM_PARITY_N
:
2865 if (i
== sh
->qd_idx
)
2866 return 0; /* It is the Q disk */
2867 switch (algorithm
) {
2868 case ALGORITHM_LEFT_ASYMMETRIC
:
2869 case ALGORITHM_RIGHT_ASYMMETRIC
:
2870 case ALGORITHM_ROTATING_ZERO_RESTART
:
2871 case ALGORITHM_ROTATING_N_RESTART
:
2872 if (sh
->pd_idx
== raid_disks
-1)
2873 i
--; /* Q D D D P */
2874 else if (i
> sh
->pd_idx
)
2875 i
-= 2; /* D D P Q D */
2877 case ALGORITHM_LEFT_SYMMETRIC
:
2878 case ALGORITHM_RIGHT_SYMMETRIC
:
2879 if (sh
->pd_idx
== raid_disks
-1)
2880 i
--; /* Q D D D P */
2885 i
-= (sh
->pd_idx
+ 2);
2888 case ALGORITHM_PARITY_0
:
2891 case ALGORITHM_PARITY_N
:
2893 case ALGORITHM_ROTATING_N_CONTINUE
:
2894 /* Like left_symmetric, but P is before Q */
2895 if (sh
->pd_idx
== 0)
2896 i
--; /* P D D D Q */
2901 i
-= (sh
->pd_idx
+ 1);
2904 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2905 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2909 case ALGORITHM_LEFT_SYMMETRIC_6
:
2910 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2912 i
+= data_disks
+ 1;
2913 i
-= (sh
->pd_idx
+ 1);
2915 case ALGORITHM_PARITY_0_6
:
2924 chunk_number
= stripe
* data_disks
+ i
;
2925 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2927 check
= raid5_compute_sector(conf
, r_sector
,
2928 previous
, &dummy1
, &sh2
);
2929 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2930 || sh2
.qd_idx
!= sh
->qd_idx
) {
2931 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
2932 mdname(conf
->mddev
));
2939 * There are cases where we want handle_stripe_dirtying() and
2940 * schedule_reconstruction() to delay towrite to some dev of a stripe.
2942 * This function checks whether we want to delay the towrite. Specifically,
2943 * we delay the towrite when:
2945 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
2946 * stripe has data in journal (for other devices).
2948 * In this case, when reading data for the non-overwrite dev, it is
2949 * necessary to handle complex rmw of write back cache (prexor with
2950 * orig_page, and xor with page). To keep read path simple, we would
2951 * like to flush data in journal to RAID disks first, so complex rmw
2952 * is handled in the write patch (handle_stripe_dirtying).
2954 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
2956 * It is important to be able to flush all stripes in raid5-cache.
2957 * Therefore, we need reserve some space on the journal device for
2958 * these flushes. If flush operation includes pending writes to the
2959 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
2960 * for the flush out. If we exclude these pending writes from flush
2961 * operation, we only need (conf->max_degraded + 1) pages per stripe.
2962 * Therefore, excluding pending writes in these cases enables more
2963 * efficient use of the journal device.
2965 * Note: To make sure the stripe makes progress, we only delay
2966 * towrite for stripes with data already in journal (injournal > 0).
2967 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
2968 * no_space_stripes list.
2971 static inline bool delay_towrite(struct r5conf
*conf
,
2973 struct stripe_head_state
*s
)
2976 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2977 !test_bit(R5_Insync
, &dev
->flags
) && s
->injournal
)
2980 if (test_bit(R5C_LOG_CRITICAL
, &conf
->cache_state
) &&
2987 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2988 int rcw
, int expand
)
2990 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
2991 struct r5conf
*conf
= sh
->raid_conf
;
2992 int level
= conf
->level
;
2996 * In some cases, handle_stripe_dirtying initially decided to
2997 * run rmw and allocates extra page for prexor. However, rcw is
2998 * cheaper later on. We need to free the extra page now,
2999 * because we won't be able to do that in ops_complete_prexor().
3001 r5c_release_extra_page(sh
);
3003 for (i
= disks
; i
--; ) {
3004 struct r5dev
*dev
= &sh
->dev
[i
];
3006 if (dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) {
3007 set_bit(R5_LOCKED
, &dev
->flags
);
3008 set_bit(R5_Wantdrain
, &dev
->flags
);
3010 clear_bit(R5_UPTODATE
, &dev
->flags
);
3012 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3013 set_bit(R5_LOCKED
, &dev
->flags
);
3017 /* if we are not expanding this is a proper write request, and
3018 * there will be bios with new data to be drained into the
3023 /* False alarm, nothing to do */
3025 sh
->reconstruct_state
= reconstruct_state_drain_run
;
3026 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3028 sh
->reconstruct_state
= reconstruct_state_run
;
3030 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3032 if (s
->locked
+ conf
->max_degraded
== disks
)
3033 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3034 atomic_inc(&conf
->pending_full_writes
);
3036 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
3037 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
3038 BUG_ON(level
== 6 &&
3039 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
3040 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
3042 for (i
= disks
; i
--; ) {
3043 struct r5dev
*dev
= &sh
->dev
[i
];
3044 if (i
== pd_idx
|| i
== qd_idx
)
3048 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3049 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3050 set_bit(R5_Wantdrain
, &dev
->flags
);
3051 set_bit(R5_LOCKED
, &dev
->flags
);
3052 clear_bit(R5_UPTODATE
, &dev
->flags
);
3054 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3055 set_bit(R5_LOCKED
, &dev
->flags
);
3060 /* False alarm - nothing to do */
3062 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
3063 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
3064 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3065 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3068 /* keep the parity disk(s) locked while asynchronous operations
3071 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
3072 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3076 int qd_idx
= sh
->qd_idx
;
3077 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
3079 set_bit(R5_LOCKED
, &dev
->flags
);
3080 clear_bit(R5_UPTODATE
, &dev
->flags
);
3084 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3085 __func__
, (unsigned long long)sh
->sector
,
3086 s
->locked
, s
->ops_request
);
3090 * Each stripe/dev can have one or more bion attached.
3091 * toread/towrite point to the first in a chain.
3092 * The bi_next chain must be in order.
3094 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
3095 int forwrite
, int previous
)
3098 struct r5conf
*conf
= sh
->raid_conf
;
3101 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3102 (unsigned long long)bi
->bi_iter
.bi_sector
,
3103 (unsigned long long)sh
->sector
);
3106 * If several bio share a stripe. The bio bi_phys_segments acts as a
3107 * reference count to avoid race. The reference count should already be
3108 * increased before this function is called (for example, in
3109 * raid5_make_request()), so other bio sharing this stripe will not free the
3110 * stripe. If a stripe is owned by one stripe, the stripe lock will
3113 spin_lock_irq(&sh
->stripe_lock
);
3114 /* Don't allow new IO added to stripes in batch list */
3118 bip
= &sh
->dev
[dd_idx
].towrite
;
3122 bip
= &sh
->dev
[dd_idx
].toread
;
3123 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
3124 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
3126 bip
= & (*bip
)->bi_next
;
3128 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
3131 if (!forwrite
|| previous
)
3132 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
3134 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
3138 raid5_inc_bi_active_stripes(bi
);
3141 /* check if page is covered */
3142 sector_t sector
= sh
->dev
[dd_idx
].sector
;
3143 for (bi
=sh
->dev
[dd_idx
].towrite
;
3144 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
3145 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
3146 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
3147 if (bio_end_sector(bi
) >= sector
)
3148 sector
= bio_end_sector(bi
);
3150 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
3151 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
3152 sh
->overwrite_disks
++;
3155 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3156 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
3157 (unsigned long long)sh
->sector
, dd_idx
);
3159 if (conf
->mddev
->bitmap
&& firstwrite
) {
3160 /* Cannot hold spinlock over bitmap_startwrite,
3161 * but must ensure this isn't added to a batch until
3162 * we have added to the bitmap and set bm_seq.
3163 * So set STRIPE_BITMAP_PENDING to prevent
3165 * If multiple add_stripe_bio() calls race here they
3166 * much all set STRIPE_BITMAP_PENDING. So only the first one
3167 * to complete "bitmap_startwrite" gets to set
3168 * STRIPE_BIT_DELAY. This is important as once a stripe
3169 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3172 set_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3173 spin_unlock_irq(&sh
->stripe_lock
);
3174 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
3176 spin_lock_irq(&sh
->stripe_lock
);
3177 clear_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3178 if (!sh
->batch_head
) {
3179 sh
->bm_seq
= conf
->seq_flush
+1;
3180 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
3183 spin_unlock_irq(&sh
->stripe_lock
);
3185 if (stripe_can_batch(sh
))
3186 stripe_add_to_batch_list(conf
, sh
);
3190 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
3191 spin_unlock_irq(&sh
->stripe_lock
);
3195 static void end_reshape(struct r5conf
*conf
);
3197 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3198 struct stripe_head
*sh
)
3200 int sectors_per_chunk
=
3201 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3203 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3204 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3206 raid5_compute_sector(conf
,
3207 stripe
* (disks
- conf
->max_degraded
)
3208 *sectors_per_chunk
+ chunk_offset
,
3214 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3215 struct stripe_head_state
*s
, int disks
,
3216 struct bio_list
*return_bi
)
3219 BUG_ON(sh
->batch_head
);
3220 for (i
= disks
; i
--; ) {
3224 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3225 struct md_rdev
*rdev
;
3227 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3228 if (rdev
&& test_bit(In_sync
, &rdev
->flags
) &&
3229 !test_bit(Faulty
, &rdev
->flags
))
3230 atomic_inc(&rdev
->nr_pending
);
3235 if (!rdev_set_badblocks(
3239 md_error(conf
->mddev
, rdev
);
3240 rdev_dec_pending(rdev
, conf
->mddev
);
3243 spin_lock_irq(&sh
->stripe_lock
);
3244 /* fail all writes first */
3245 bi
= sh
->dev
[i
].towrite
;
3246 sh
->dev
[i
].towrite
= NULL
;
3247 sh
->overwrite_disks
= 0;
3248 spin_unlock_irq(&sh
->stripe_lock
);
3252 r5l_stripe_write_finished(sh
);
3254 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3255 wake_up(&conf
->wait_for_overlap
);
3257 while (bi
&& bi
->bi_iter
.bi_sector
<
3258 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3259 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3261 bi
->bi_error
= -EIO
;
3262 if (!raid5_dec_bi_active_stripes(bi
)) {
3263 md_write_end(conf
->mddev
);
3264 bio_list_add(return_bi
, bi
);
3269 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3270 STRIPE_SECTORS
, 0, 0);
3272 /* and fail all 'written' */
3273 bi
= sh
->dev
[i
].written
;
3274 sh
->dev
[i
].written
= NULL
;
3275 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3276 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3277 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3280 if (bi
) bitmap_end
= 1;
3281 while (bi
&& bi
->bi_iter
.bi_sector
<
3282 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3283 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3285 bi
->bi_error
= -EIO
;
3286 if (!raid5_dec_bi_active_stripes(bi
)) {
3287 md_write_end(conf
->mddev
);
3288 bio_list_add(return_bi
, bi
);
3293 /* fail any reads if this device is non-operational and
3294 * the data has not reached the cache yet.
3296 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3297 s
->failed
> conf
->max_degraded
&&
3298 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3299 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3300 spin_lock_irq(&sh
->stripe_lock
);
3301 bi
= sh
->dev
[i
].toread
;
3302 sh
->dev
[i
].toread
= NULL
;
3303 spin_unlock_irq(&sh
->stripe_lock
);
3304 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3305 wake_up(&conf
->wait_for_overlap
);
3308 while (bi
&& bi
->bi_iter
.bi_sector
<
3309 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3310 struct bio
*nextbi
=
3311 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3313 bi
->bi_error
= -EIO
;
3314 if (!raid5_dec_bi_active_stripes(bi
))
3315 bio_list_add(return_bi
, bi
);
3320 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3321 STRIPE_SECTORS
, 0, 0);
3322 /* If we were in the middle of a write the parity block might
3323 * still be locked - so just clear all R5_LOCKED flags
3325 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3330 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3331 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3332 md_wakeup_thread(conf
->mddev
->thread
);
3336 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3337 struct stripe_head_state
*s
)
3342 BUG_ON(sh
->batch_head
);
3343 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3344 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3345 wake_up(&conf
->wait_for_overlap
);
3348 /* There is nothing more to do for sync/check/repair.
3349 * Don't even need to abort as that is handled elsewhere
3350 * if needed, and not always wanted e.g. if there is a known
3352 * For recover/replace we need to record a bad block on all
3353 * non-sync devices, or abort the recovery
3355 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3356 /* During recovery devices cannot be removed, so
3357 * locking and refcounting of rdevs is not needed
3360 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3361 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3363 && !test_bit(Faulty
, &rdev
->flags
)
3364 && !test_bit(In_sync
, &rdev
->flags
)
3365 && !rdev_set_badblocks(rdev
, sh
->sector
,
3368 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3370 && !test_bit(Faulty
, &rdev
->flags
)
3371 && !test_bit(In_sync
, &rdev
->flags
)
3372 && !rdev_set_badblocks(rdev
, sh
->sector
,
3378 conf
->recovery_disabled
=
3379 conf
->mddev
->recovery_disabled
;
3381 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3384 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3386 struct md_rdev
*rdev
;
3390 rdev
= rcu_dereference(sh
->raid_conf
->disks
[disk_idx
].replacement
);
3392 && !test_bit(Faulty
, &rdev
->flags
)
3393 && !test_bit(In_sync
, &rdev
->flags
)
3394 && (rdev
->recovery_offset
<= sh
->sector
3395 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3401 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3402 int disk_idx
, int disks
)
3404 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3405 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3406 &sh
->dev
[s
->failed_num
[1]] };
3410 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3411 test_bit(R5_UPTODATE
, &dev
->flags
))
3412 /* No point reading this as we already have it or have
3413 * decided to get it.
3418 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3419 /* We need this block to directly satisfy a request */
3422 if (s
->syncing
|| s
->expanding
||
3423 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3424 /* When syncing, or expanding we read everything.
3425 * When replacing, we need the replaced block.
3429 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3430 (s
->failed
>= 2 && fdev
[1]->toread
))
3431 /* If we want to read from a failed device, then
3432 * we need to actually read every other device.
3436 /* Sometimes neither read-modify-write nor reconstruct-write
3437 * cycles can work. In those cases we read every block we
3438 * can. Then the parity-update is certain to have enough to
3440 * This can only be a problem when we need to write something,
3441 * and some device has failed. If either of those tests
3442 * fail we need look no further.
3444 if (!s
->failed
|| !s
->to_write
)
3447 if (test_bit(R5_Insync
, &dev
->flags
) &&
3448 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3449 /* Pre-reads at not permitted until after short delay
3450 * to gather multiple requests. However if this
3451 * device is no Insync, the block could only be be computed
3452 * and there is no need to delay that.
3456 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3457 if (fdev
[i
]->towrite
&&
3458 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3459 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3460 /* If we have a partial write to a failed
3461 * device, then we will need to reconstruct
3462 * the content of that device, so all other
3463 * devices must be read.
3468 /* If we are forced to do a reconstruct-write, either because
3469 * the current RAID6 implementation only supports that, or
3470 * or because parity cannot be trusted and we are currently
3471 * recovering it, there is extra need to be careful.
3472 * If one of the devices that we would need to read, because
3473 * it is not being overwritten (and maybe not written at all)
3474 * is missing/faulty, then we need to read everything we can.
3476 if (sh
->raid_conf
->level
!= 6 &&
3477 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3478 /* reconstruct-write isn't being forced */
3480 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3481 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3482 s
->failed_num
[i
] != sh
->qd_idx
&&
3483 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3484 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3491 /* fetch_block - checks the given member device to see if its data needs
3492 * to be read or computed to satisfy a request.
3494 * Returns 1 when no more member devices need to be checked, otherwise returns
3495 * 0 to tell the loop in handle_stripe_fill to continue
3497 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3498 int disk_idx
, int disks
)
3500 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3502 /* is the data in this block needed, and can we get it? */
3503 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3504 /* we would like to get this block, possibly by computing it,
3505 * otherwise read it if the backing disk is insync
3507 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3508 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3509 BUG_ON(sh
->batch_head
);
3510 if ((s
->uptodate
== disks
- 1) &&
3511 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3512 disk_idx
== s
->failed_num
[1]))) {
3513 /* have disk failed, and we're requested to fetch it;
3516 pr_debug("Computing stripe %llu block %d\n",
3517 (unsigned long long)sh
->sector
, disk_idx
);
3518 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3519 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3520 set_bit(R5_Wantcompute
, &dev
->flags
);
3521 sh
->ops
.target
= disk_idx
;
3522 sh
->ops
.target2
= -1; /* no 2nd target */
3524 /* Careful: from this point on 'uptodate' is in the eye
3525 * of raid_run_ops which services 'compute' operations
3526 * before writes. R5_Wantcompute flags a block that will
3527 * be R5_UPTODATE by the time it is needed for a
3528 * subsequent operation.
3532 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3533 /* Computing 2-failure is *very* expensive; only
3534 * do it if failed >= 2
3537 for (other
= disks
; other
--; ) {
3538 if (other
== disk_idx
)
3540 if (!test_bit(R5_UPTODATE
,
3541 &sh
->dev
[other
].flags
))
3545 pr_debug("Computing stripe %llu blocks %d,%d\n",
3546 (unsigned long long)sh
->sector
,
3548 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3549 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3550 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3551 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3552 sh
->ops
.target
= disk_idx
;
3553 sh
->ops
.target2
= other
;
3557 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3558 set_bit(R5_LOCKED
, &dev
->flags
);
3559 set_bit(R5_Wantread
, &dev
->flags
);
3561 pr_debug("Reading block %d (sync=%d)\n",
3562 disk_idx
, s
->syncing
);
3570 * handle_stripe_fill - read or compute data to satisfy pending requests.
3572 static void handle_stripe_fill(struct stripe_head
*sh
,
3573 struct stripe_head_state
*s
,
3578 /* look for blocks to read/compute, skip this if a compute
3579 * is already in flight, or if the stripe contents are in the
3580 * midst of changing due to a write
3582 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3583 !sh
->reconstruct_state
) {
3586 * For degraded stripe with data in journal, do not handle
3587 * read requests yet, instead, flush the stripe to raid
3588 * disks first, this avoids handling complex rmw of write
3589 * back cache (prexor with orig_page, and then xor with
3590 * page) in the read path
3592 if (s
->injournal
&& s
->failed
) {
3593 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
3594 r5c_make_stripe_write_out(sh
);
3598 for (i
= disks
; i
--; )
3599 if (fetch_block(sh
, s
, i
, disks
))
3603 set_bit(STRIPE_HANDLE
, &sh
->state
);
3606 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
3607 unsigned long handle_flags
);
3608 /* handle_stripe_clean_event
3609 * any written block on an uptodate or failed drive can be returned.
3610 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3611 * never LOCKED, so we don't need to test 'failed' directly.
3613 static void handle_stripe_clean_event(struct r5conf
*conf
,
3614 struct stripe_head
*sh
, int disks
, struct bio_list
*return_bi
)
3618 int discard_pending
= 0;
3619 struct stripe_head
*head_sh
= sh
;
3620 bool do_endio
= false;
3622 for (i
= disks
; i
--; )
3623 if (sh
->dev
[i
].written
) {
3625 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3626 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3627 test_bit(R5_Discard
, &dev
->flags
) ||
3628 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3629 /* We can return any write requests */
3630 struct bio
*wbi
, *wbi2
;
3631 pr_debug("Return write for disc %d\n", i
);
3632 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3633 clear_bit(R5_UPTODATE
, &dev
->flags
);
3634 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3635 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3640 dev
->page
= dev
->orig_page
;
3642 dev
->written
= NULL
;
3643 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3644 dev
->sector
+ STRIPE_SECTORS
) {
3645 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3646 if (!raid5_dec_bi_active_stripes(wbi
)) {
3647 md_write_end(conf
->mddev
);
3648 bio_list_add(return_bi
, wbi
);
3652 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3654 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3656 if (head_sh
->batch_head
) {
3657 sh
= list_first_entry(&sh
->batch_list
,
3660 if (sh
!= head_sh
) {
3667 } else if (test_bit(R5_Discard
, &dev
->flags
))
3668 discard_pending
= 1;
3671 r5l_stripe_write_finished(sh
);
3673 if (!discard_pending
&&
3674 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3676 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3677 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3678 if (sh
->qd_idx
>= 0) {
3679 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3680 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3682 /* now that discard is done we can proceed with any sync */
3683 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3685 * SCSI discard will change some bio fields and the stripe has
3686 * no updated data, so remove it from hash list and the stripe
3687 * will be reinitialized
3690 hash
= sh
->hash_lock_index
;
3691 spin_lock_irq(conf
->hash_locks
+ hash
);
3693 spin_unlock_irq(conf
->hash_locks
+ hash
);
3694 if (head_sh
->batch_head
) {
3695 sh
= list_first_entry(&sh
->batch_list
,
3696 struct stripe_head
, batch_list
);
3702 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3703 set_bit(STRIPE_HANDLE
, &sh
->state
);
3707 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3708 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3709 md_wakeup_thread(conf
->mddev
->thread
);
3711 if (head_sh
->batch_head
&& do_endio
)
3712 break_stripe_batch_list(head_sh
, STRIPE_EXPAND_SYNC_FLAGS
);
3716 * For RMW in write back cache, we need extra page in prexor to store the
3717 * old data. This page is stored in dev->orig_page.
3719 * This function checks whether we have data for prexor. The exact logic
3721 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3723 static inline bool uptodate_for_rmw(struct r5dev
*dev
)
3725 return (test_bit(R5_UPTODATE
, &dev
->flags
)) &&
3726 (!test_bit(R5_InJournal
, &dev
->flags
) ||
3727 test_bit(R5_OrigPageUPTDODATE
, &dev
->flags
));
3730 static int handle_stripe_dirtying(struct r5conf
*conf
,
3731 struct stripe_head
*sh
,
3732 struct stripe_head_state
*s
,
3735 int rmw
= 0, rcw
= 0, i
;
3736 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3738 /* Check whether resync is now happening or should start.
3739 * If yes, then the array is dirty (after unclean shutdown or
3740 * initial creation), so parity in some stripes might be inconsistent.
3741 * In this case, we need to always do reconstruct-write, to ensure
3742 * that in case of drive failure or read-error correction, we
3743 * generate correct data from the parity.
3745 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3746 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3748 /* Calculate the real rcw later - for now make it
3749 * look like rcw is cheaper
3752 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3753 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3754 (unsigned long long)sh
->sector
);
3755 } else for (i
= disks
; i
--; ) {
3756 /* would I have to read this buffer for read_modify_write */
3757 struct r5dev
*dev
= &sh
->dev
[i
];
3758 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3759 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3760 test_bit(R5_InJournal
, &dev
->flags
)) &&
3761 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3762 !(uptodate_for_rmw(dev
) ||
3763 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3764 if (test_bit(R5_Insync
, &dev
->flags
))
3767 rmw
+= 2*disks
; /* cannot read it */
3769 /* Would I have to read this buffer for reconstruct_write */
3770 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3771 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3772 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3773 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3774 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3775 if (test_bit(R5_Insync
, &dev
->flags
))
3782 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3783 (unsigned long long)sh
->sector
, sh
->state
, rmw
, rcw
);
3784 set_bit(STRIPE_HANDLE
, &sh
->state
);
3785 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_PREFER_RMW
)) && rmw
> 0) {
3786 /* prefer read-modify-write, but need to get some data */
3787 if (conf
->mddev
->queue
)
3788 blk_add_trace_msg(conf
->mddev
->queue
,
3789 "raid5 rmw %llu %d",
3790 (unsigned long long)sh
->sector
, rmw
);
3791 for (i
= disks
; i
--; ) {
3792 struct r5dev
*dev
= &sh
->dev
[i
];
3793 if (test_bit(R5_InJournal
, &dev
->flags
) &&
3794 dev
->page
== dev
->orig_page
&&
3795 !test_bit(R5_LOCKED
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3796 /* alloc page for prexor */
3797 struct page
*p
= alloc_page(GFP_NOIO
);
3805 * alloc_page() failed, try use
3806 * disk_info->extra_page
3808 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE
,
3809 &conf
->cache_state
)) {
3810 r5c_use_extra_page(sh
);
3814 /* extra_page in use, add to delayed_list */
3815 set_bit(STRIPE_DELAYED
, &sh
->state
);
3816 s
->waiting_extra_page
= 1;
3821 for (i
= disks
; i
--; ) {
3822 struct r5dev
*dev
= &sh
->dev
[i
];
3823 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3824 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3825 test_bit(R5_InJournal
, &dev
->flags
)) &&
3826 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3827 !(uptodate_for_rmw(dev
) ||
3828 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3829 test_bit(R5_Insync
, &dev
->flags
)) {
3830 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3832 pr_debug("Read_old block %d for r-m-w\n",
3834 set_bit(R5_LOCKED
, &dev
->flags
);
3835 set_bit(R5_Wantread
, &dev
->flags
);
3838 set_bit(STRIPE_DELAYED
, &sh
->state
);
3839 set_bit(STRIPE_HANDLE
, &sh
->state
);
3844 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_PREFER_RMW
)) && rcw
> 0) {
3845 /* want reconstruct write, but need to get some data */
3848 for (i
= disks
; i
--; ) {
3849 struct r5dev
*dev
= &sh
->dev
[i
];
3850 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3851 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3852 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3853 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3854 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3856 if (test_bit(R5_Insync
, &dev
->flags
) &&
3857 test_bit(STRIPE_PREREAD_ACTIVE
,
3859 pr_debug("Read_old block "
3860 "%d for Reconstruct\n", i
);
3861 set_bit(R5_LOCKED
, &dev
->flags
);
3862 set_bit(R5_Wantread
, &dev
->flags
);
3866 set_bit(STRIPE_DELAYED
, &sh
->state
);
3867 set_bit(STRIPE_HANDLE
, &sh
->state
);
3871 if (rcw
&& conf
->mddev
->queue
)
3872 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3873 (unsigned long long)sh
->sector
,
3874 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3877 if (rcw
> disks
&& rmw
> disks
&&
3878 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3879 set_bit(STRIPE_DELAYED
, &sh
->state
);
3881 /* now if nothing is locked, and if we have enough data,
3882 * we can start a write request
3884 /* since handle_stripe can be called at any time we need to handle the
3885 * case where a compute block operation has been submitted and then a
3886 * subsequent call wants to start a write request. raid_run_ops only
3887 * handles the case where compute block and reconstruct are requested
3888 * simultaneously. If this is not the case then new writes need to be
3889 * held off until the compute completes.
3891 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
3892 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
3893 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
3894 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
3898 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
3899 struct stripe_head_state
*s
, int disks
)
3901 struct r5dev
*dev
= NULL
;
3903 BUG_ON(sh
->batch_head
);
3904 set_bit(STRIPE_HANDLE
, &sh
->state
);
3906 switch (sh
->check_state
) {
3907 case check_state_idle
:
3908 /* start a new check operation if there are no failures */
3909 if (s
->failed
== 0) {
3910 BUG_ON(s
->uptodate
!= disks
);
3911 sh
->check_state
= check_state_run
;
3912 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3913 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3917 dev
= &sh
->dev
[s
->failed_num
[0]];
3919 case check_state_compute_result
:
3920 sh
->check_state
= check_state_idle
;
3922 dev
= &sh
->dev
[sh
->pd_idx
];
3924 /* check that a write has not made the stripe insync */
3925 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3928 /* either failed parity check, or recovery is happening */
3929 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3930 BUG_ON(s
->uptodate
!= disks
);
3932 set_bit(R5_LOCKED
, &dev
->flags
);
3934 set_bit(R5_Wantwrite
, &dev
->flags
);
3936 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3937 set_bit(STRIPE_INSYNC
, &sh
->state
);
3939 case check_state_run
:
3940 break; /* we will be called again upon completion */
3941 case check_state_check_result
:
3942 sh
->check_state
= check_state_idle
;
3944 /* if a failure occurred during the check operation, leave
3945 * STRIPE_INSYNC not set and let the stripe be handled again
3950 /* handle a successful check operation, if parity is correct
3951 * we are done. Otherwise update the mismatch count and repair
3952 * parity if !MD_RECOVERY_CHECK
3954 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3955 /* parity is correct (on disc,
3956 * not in buffer any more)
3958 set_bit(STRIPE_INSYNC
, &sh
->state
);
3960 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3961 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3962 /* don't try to repair!! */
3963 set_bit(STRIPE_INSYNC
, &sh
->state
);
3965 sh
->check_state
= check_state_compute_run
;
3966 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3967 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3968 set_bit(R5_Wantcompute
,
3969 &sh
->dev
[sh
->pd_idx
].flags
);
3970 sh
->ops
.target
= sh
->pd_idx
;
3971 sh
->ops
.target2
= -1;
3976 case check_state_compute_run
:
3979 pr_err("%s: unknown check_state: %d sector: %llu\n",
3980 __func__
, sh
->check_state
,
3981 (unsigned long long) sh
->sector
);
3986 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3987 struct stripe_head_state
*s
,
3990 int pd_idx
= sh
->pd_idx
;
3991 int qd_idx
= sh
->qd_idx
;
3994 BUG_ON(sh
->batch_head
);
3995 set_bit(STRIPE_HANDLE
, &sh
->state
);
3997 BUG_ON(s
->failed
> 2);
3999 /* Want to check and possibly repair P and Q.
4000 * However there could be one 'failed' device, in which
4001 * case we can only check one of them, possibly using the
4002 * other to generate missing data
4005 switch (sh
->check_state
) {
4006 case check_state_idle
:
4007 /* start a new check operation if there are < 2 failures */
4008 if (s
->failed
== s
->q_failed
) {
4009 /* The only possible failed device holds Q, so it
4010 * makes sense to check P (If anything else were failed,
4011 * we would have used P to recreate it).
4013 sh
->check_state
= check_state_run
;
4015 if (!s
->q_failed
&& s
->failed
< 2) {
4016 /* Q is not failed, and we didn't use it to generate
4017 * anything, so it makes sense to check it
4019 if (sh
->check_state
== check_state_run
)
4020 sh
->check_state
= check_state_run_pq
;
4022 sh
->check_state
= check_state_run_q
;
4025 /* discard potentially stale zero_sum_result */
4026 sh
->ops
.zero_sum_result
= 0;
4028 if (sh
->check_state
== check_state_run
) {
4029 /* async_xor_zero_sum destroys the contents of P */
4030 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
4033 if (sh
->check_state
>= check_state_run
&&
4034 sh
->check_state
<= check_state_run_pq
) {
4035 /* async_syndrome_zero_sum preserves P and Q, so
4036 * no need to mark them !uptodate here
4038 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4042 /* we have 2-disk failure */
4043 BUG_ON(s
->failed
!= 2);
4045 case check_state_compute_result
:
4046 sh
->check_state
= check_state_idle
;
4048 /* check that a write has not made the stripe insync */
4049 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4052 /* now write out any block on a failed drive,
4053 * or P or Q if they were recomputed
4055 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
4056 if (s
->failed
== 2) {
4057 dev
= &sh
->dev
[s
->failed_num
[1]];
4059 set_bit(R5_LOCKED
, &dev
->flags
);
4060 set_bit(R5_Wantwrite
, &dev
->flags
);
4062 if (s
->failed
>= 1) {
4063 dev
= &sh
->dev
[s
->failed_num
[0]];
4065 set_bit(R5_LOCKED
, &dev
->flags
);
4066 set_bit(R5_Wantwrite
, &dev
->flags
);
4068 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4069 dev
= &sh
->dev
[pd_idx
];
4071 set_bit(R5_LOCKED
, &dev
->flags
);
4072 set_bit(R5_Wantwrite
, &dev
->flags
);
4074 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4075 dev
= &sh
->dev
[qd_idx
];
4077 set_bit(R5_LOCKED
, &dev
->flags
);
4078 set_bit(R5_Wantwrite
, &dev
->flags
);
4080 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4082 set_bit(STRIPE_INSYNC
, &sh
->state
);
4084 case check_state_run
:
4085 case check_state_run_q
:
4086 case check_state_run_pq
:
4087 break; /* we will be called again upon completion */
4088 case check_state_check_result
:
4089 sh
->check_state
= check_state_idle
;
4091 /* handle a successful check operation, if parity is correct
4092 * we are done. Otherwise update the mismatch count and repair
4093 * parity if !MD_RECOVERY_CHECK
4095 if (sh
->ops
.zero_sum_result
== 0) {
4096 /* both parities are correct */
4098 set_bit(STRIPE_INSYNC
, &sh
->state
);
4100 /* in contrast to the raid5 case we can validate
4101 * parity, but still have a failure to write
4104 sh
->check_state
= check_state_compute_result
;
4105 /* Returning at this point means that we may go
4106 * off and bring p and/or q uptodate again so
4107 * we make sure to check zero_sum_result again
4108 * to verify if p or q need writeback
4112 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
4113 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
4114 /* don't try to repair!! */
4115 set_bit(STRIPE_INSYNC
, &sh
->state
);
4117 int *target
= &sh
->ops
.target
;
4119 sh
->ops
.target
= -1;
4120 sh
->ops
.target2
= -1;
4121 sh
->check_state
= check_state_compute_run
;
4122 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4123 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4124 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4125 set_bit(R5_Wantcompute
,
4126 &sh
->dev
[pd_idx
].flags
);
4128 target
= &sh
->ops
.target2
;
4131 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4132 set_bit(R5_Wantcompute
,
4133 &sh
->dev
[qd_idx
].flags
);
4140 case check_state_compute_run
:
4143 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4144 __func__
, sh
->check_state
,
4145 (unsigned long long) sh
->sector
);
4150 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
4154 /* We have read all the blocks in this stripe and now we need to
4155 * copy some of them into a target stripe for expand.
4157 struct dma_async_tx_descriptor
*tx
= NULL
;
4158 BUG_ON(sh
->batch_head
);
4159 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4160 for (i
= 0; i
< sh
->disks
; i
++)
4161 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
4163 struct stripe_head
*sh2
;
4164 struct async_submit_ctl submit
;
4166 sector_t bn
= raid5_compute_blocknr(sh
, i
, 1);
4167 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
4169 sh2
= raid5_get_active_stripe(conf
, s
, 0, 1, 1);
4171 /* so far only the early blocks of this stripe
4172 * have been requested. When later blocks
4173 * get requested, we will try again
4176 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
4177 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
4178 /* must have already done this block */
4179 raid5_release_stripe(sh2
);
4183 /* place all the copies on one channel */
4184 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
4185 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
4186 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
4189 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
4190 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
4191 for (j
= 0; j
< conf
->raid_disks
; j
++)
4192 if (j
!= sh2
->pd_idx
&&
4194 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
4196 if (j
== conf
->raid_disks
) {
4197 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
4198 set_bit(STRIPE_HANDLE
, &sh2
->state
);
4200 raid5_release_stripe(sh2
);
4203 /* done submitting copies, wait for them to complete */
4204 async_tx_quiesce(&tx
);
4208 * handle_stripe - do things to a stripe.
4210 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4211 * state of various bits to see what needs to be done.
4213 * return some read requests which now have data
4214 * return some write requests which are safely on storage
4215 * schedule a read on some buffers
4216 * schedule a write of some buffers
4217 * return confirmation of parity correctness
4221 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
4223 struct r5conf
*conf
= sh
->raid_conf
;
4224 int disks
= sh
->disks
;
4227 int do_recovery
= 0;
4229 memset(s
, 0, sizeof(*s
));
4231 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
4232 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
4233 s
->failed_num
[0] = -1;
4234 s
->failed_num
[1] = -1;
4235 s
->log_failed
= r5l_log_disk_error(conf
);
4237 /* Now to look around and see what can be done */
4239 for (i
=disks
; i
--; ) {
4240 struct md_rdev
*rdev
;
4247 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4249 dev
->toread
, dev
->towrite
, dev
->written
);
4250 /* maybe we can reply to a read
4252 * new wantfill requests are only permitted while
4253 * ops_complete_biofill is guaranteed to be inactive
4255 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4256 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4257 set_bit(R5_Wantfill
, &dev
->flags
);
4259 /* now count some things */
4260 if (test_bit(R5_LOCKED
, &dev
->flags
))
4262 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4264 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4266 BUG_ON(s
->compute
> 2);
4269 if (test_bit(R5_Wantfill
, &dev
->flags
))
4271 else if (dev
->toread
)
4275 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4280 /* Prefer to use the replacement for reads, but only
4281 * if it is recovered enough and has no bad blocks.
4283 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4284 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4285 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4286 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4287 &first_bad
, &bad_sectors
))
4288 set_bit(R5_ReadRepl
, &dev
->flags
);
4290 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4291 set_bit(R5_NeedReplace
, &dev
->flags
);
4293 clear_bit(R5_NeedReplace
, &dev
->flags
);
4294 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4295 clear_bit(R5_ReadRepl
, &dev
->flags
);
4297 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4300 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4301 &first_bad
, &bad_sectors
);
4302 if (s
->blocked_rdev
== NULL
4303 && (test_bit(Blocked
, &rdev
->flags
)
4306 set_bit(BlockedBadBlocks
,
4308 s
->blocked_rdev
= rdev
;
4309 atomic_inc(&rdev
->nr_pending
);
4312 clear_bit(R5_Insync
, &dev
->flags
);
4316 /* also not in-sync */
4317 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4318 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4319 /* treat as in-sync, but with a read error
4320 * which we can now try to correct
4322 set_bit(R5_Insync
, &dev
->flags
);
4323 set_bit(R5_ReadError
, &dev
->flags
);
4325 } else if (test_bit(In_sync
, &rdev
->flags
))
4326 set_bit(R5_Insync
, &dev
->flags
);
4327 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4328 /* in sync if before recovery_offset */
4329 set_bit(R5_Insync
, &dev
->flags
);
4330 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4331 test_bit(R5_Expanded
, &dev
->flags
))
4332 /* If we've reshaped into here, we assume it is Insync.
4333 * We will shortly update recovery_offset to make
4336 set_bit(R5_Insync
, &dev
->flags
);
4338 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4339 /* This flag does not apply to '.replacement'
4340 * only to .rdev, so make sure to check that*/
4341 struct md_rdev
*rdev2
= rcu_dereference(
4342 conf
->disks
[i
].rdev
);
4344 clear_bit(R5_Insync
, &dev
->flags
);
4345 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4346 s
->handle_bad_blocks
= 1;
4347 atomic_inc(&rdev2
->nr_pending
);
4349 clear_bit(R5_WriteError
, &dev
->flags
);
4351 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4352 /* This flag does not apply to '.replacement'
4353 * only to .rdev, so make sure to check that*/
4354 struct md_rdev
*rdev2
= rcu_dereference(
4355 conf
->disks
[i
].rdev
);
4356 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4357 s
->handle_bad_blocks
= 1;
4358 atomic_inc(&rdev2
->nr_pending
);
4360 clear_bit(R5_MadeGood
, &dev
->flags
);
4362 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4363 struct md_rdev
*rdev2
= rcu_dereference(
4364 conf
->disks
[i
].replacement
);
4365 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4366 s
->handle_bad_blocks
= 1;
4367 atomic_inc(&rdev2
->nr_pending
);
4369 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4371 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4372 /* The ReadError flag will just be confusing now */
4373 clear_bit(R5_ReadError
, &dev
->flags
);
4374 clear_bit(R5_ReWrite
, &dev
->flags
);
4376 if (test_bit(R5_ReadError
, &dev
->flags
))
4377 clear_bit(R5_Insync
, &dev
->flags
);
4378 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4380 s
->failed_num
[s
->failed
] = i
;
4382 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4386 if (test_bit(R5_InJournal
, &dev
->flags
))
4388 if (test_bit(R5_InJournal
, &dev
->flags
) && dev
->written
)
4391 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4392 /* If there is a failed device being replaced,
4393 * we must be recovering.
4394 * else if we are after recovery_cp, we must be syncing
4395 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4396 * else we can only be replacing
4397 * sync and recovery both need to read all devices, and so
4398 * use the same flag.
4401 sh
->sector
>= conf
->mddev
->recovery_cp
||
4402 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4410 static int clear_batch_ready(struct stripe_head
*sh
)
4412 /* Return '1' if this is a member of batch, or
4413 * '0' if it is a lone stripe or a head which can now be
4416 struct stripe_head
*tmp
;
4417 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4418 return (sh
->batch_head
&& sh
->batch_head
!= sh
);
4419 spin_lock(&sh
->stripe_lock
);
4420 if (!sh
->batch_head
) {
4421 spin_unlock(&sh
->stripe_lock
);
4426 * this stripe could be added to a batch list before we check
4427 * BATCH_READY, skips it
4429 if (sh
->batch_head
!= sh
) {
4430 spin_unlock(&sh
->stripe_lock
);
4433 spin_lock(&sh
->batch_lock
);
4434 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4435 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4436 spin_unlock(&sh
->batch_lock
);
4437 spin_unlock(&sh
->stripe_lock
);
4440 * BATCH_READY is cleared, no new stripes can be added.
4441 * batch_list can be accessed without lock
4446 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4447 unsigned long handle_flags
)
4449 struct stripe_head
*sh
, *next
;
4453 list_for_each_entry_safe(sh
, next
, &head_sh
->batch_list
, batch_list
) {
4455 list_del_init(&sh
->batch_list
);
4457 WARN_ONCE(sh
->state
& ((1 << STRIPE_ACTIVE
) |
4458 (1 << STRIPE_SYNCING
) |
4459 (1 << STRIPE_REPLACED
) |
4460 (1 << STRIPE_DELAYED
) |
4461 (1 << STRIPE_BIT_DELAY
) |
4462 (1 << STRIPE_FULL_WRITE
) |
4463 (1 << STRIPE_BIOFILL_RUN
) |
4464 (1 << STRIPE_COMPUTE_RUN
) |
4465 (1 << STRIPE_OPS_REQ_PENDING
) |
4466 (1 << STRIPE_DISCARD
) |
4467 (1 << STRIPE_BATCH_READY
) |
4468 (1 << STRIPE_BATCH_ERR
) |
4469 (1 << STRIPE_BITMAP_PENDING
)),
4470 "stripe state: %lx\n", sh
->state
);
4471 WARN_ONCE(head_sh
->state
& ((1 << STRIPE_DISCARD
) |
4472 (1 << STRIPE_REPLACED
)),
4473 "head stripe state: %lx\n", head_sh
->state
);
4475 set_mask_bits(&sh
->state
, ~(STRIPE_EXPAND_SYNC_FLAGS
|
4476 (1 << STRIPE_PREREAD_ACTIVE
) |
4477 (1 << STRIPE_DEGRADED
)),
4478 head_sh
->state
& (1 << STRIPE_INSYNC
));
4480 sh
->check_state
= head_sh
->check_state
;
4481 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4482 for (i
= 0; i
< sh
->disks
; i
++) {
4483 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
4485 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4486 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4488 spin_lock_irq(&sh
->stripe_lock
);
4489 sh
->batch_head
= NULL
;
4490 spin_unlock_irq(&sh
->stripe_lock
);
4491 if (handle_flags
== 0 ||
4492 sh
->state
& handle_flags
)
4493 set_bit(STRIPE_HANDLE
, &sh
->state
);
4494 raid5_release_stripe(sh
);
4496 spin_lock_irq(&head_sh
->stripe_lock
);
4497 head_sh
->batch_head
= NULL
;
4498 spin_unlock_irq(&head_sh
->stripe_lock
);
4499 for (i
= 0; i
< head_sh
->disks
; i
++)
4500 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
4502 if (head_sh
->state
& handle_flags
)
4503 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
4506 wake_up(&head_sh
->raid_conf
->wait_for_overlap
);
4509 static void handle_stripe(struct stripe_head
*sh
)
4511 struct stripe_head_state s
;
4512 struct r5conf
*conf
= sh
->raid_conf
;
4515 int disks
= sh
->disks
;
4516 struct r5dev
*pdev
, *qdev
;
4518 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4519 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4520 /* already being handled, ensure it gets handled
4521 * again when current action finishes */
4522 set_bit(STRIPE_HANDLE
, &sh
->state
);
4526 if (clear_batch_ready(sh
) ) {
4527 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4531 if (test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4532 break_stripe_batch_list(sh
, 0);
4534 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4535 spin_lock(&sh
->stripe_lock
);
4536 /* Cannot process 'sync' concurrently with 'discard' */
4537 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4538 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4539 set_bit(STRIPE_SYNCING
, &sh
->state
);
4540 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4541 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4543 spin_unlock(&sh
->stripe_lock
);
4545 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4547 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4548 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4549 (unsigned long long)sh
->sector
, sh
->state
,
4550 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4551 sh
->check_state
, sh
->reconstruct_state
);
4553 analyse_stripe(sh
, &s
);
4555 if (test_bit(STRIPE_LOG_TRAPPED
, &sh
->state
))
4558 if (s
.handle_bad_blocks
) {
4559 set_bit(STRIPE_HANDLE
, &sh
->state
);
4563 if (unlikely(s
.blocked_rdev
)) {
4564 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4565 s
.replacing
|| s
.to_write
|| s
.written
) {
4566 set_bit(STRIPE_HANDLE
, &sh
->state
);
4569 /* There is nothing for the blocked_rdev to block */
4570 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4571 s
.blocked_rdev
= NULL
;
4574 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4575 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4576 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4579 pr_debug("locked=%d uptodate=%d to_read=%d"
4580 " to_write=%d failed=%d failed_num=%d,%d\n",
4581 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4582 s
.failed_num
[0], s
.failed_num
[1]);
4583 /* check if the array has lost more than max_degraded devices and,
4584 * if so, some requests might need to be failed.
4586 if (s
.failed
> conf
->max_degraded
|| s
.log_failed
) {
4587 sh
->check_state
= 0;
4588 sh
->reconstruct_state
= 0;
4589 break_stripe_batch_list(sh
, 0);
4590 if (s
.to_read
+s
.to_write
+s
.written
)
4591 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
4592 if (s
.syncing
+ s
.replacing
)
4593 handle_failed_sync(conf
, sh
, &s
);
4596 /* Now we check to see if any write operations have recently
4600 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4602 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4603 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4604 sh
->reconstruct_state
= reconstruct_state_idle
;
4606 /* All the 'written' buffers and the parity block are ready to
4607 * be written back to disk
4609 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4610 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4611 BUG_ON(sh
->qd_idx
>= 0 &&
4612 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4613 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4614 for (i
= disks
; i
--; ) {
4615 struct r5dev
*dev
= &sh
->dev
[i
];
4616 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4617 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4618 dev
->written
|| test_bit(R5_InJournal
,
4620 pr_debug("Writing block %d\n", i
);
4621 set_bit(R5_Wantwrite
, &dev
->flags
);
4626 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4627 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4629 set_bit(STRIPE_INSYNC
, &sh
->state
);
4632 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4633 s
.dec_preread_active
= 1;
4637 * might be able to return some write requests if the parity blocks
4638 * are safe, or on a failed drive
4640 pdev
= &sh
->dev
[sh
->pd_idx
];
4641 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4642 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4643 qdev
= &sh
->dev
[sh
->qd_idx
];
4644 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4645 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4649 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4650 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4651 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4652 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4653 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4654 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4655 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4656 test_bit(R5_Discard
, &qdev
->flags
))))))
4657 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
4660 r5c_handle_cached_data_endio(conf
, sh
, disks
, &s
.return_bi
);
4661 r5l_stripe_write_finished(sh
);
4663 /* Now we might consider reading some blocks, either to check/generate
4664 * parity, or to satisfy requests
4665 * or to load a block that is being partially written.
4667 if (s
.to_read
|| s
.non_overwrite
4668 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4669 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4672 handle_stripe_fill(sh
, &s
, disks
);
4675 * When the stripe finishes full journal write cycle (write to journal
4676 * and raid disk), this is the clean up procedure so it is ready for
4679 r5c_finish_stripe_write_out(conf
, sh
, &s
);
4682 * Now to consider new write requests, cache write back and what else,
4683 * if anything should be read. We do not handle new writes when:
4684 * 1/ A 'write' operation (copy+xor) is already in flight.
4685 * 2/ A 'check' operation is in flight, as it may clobber the parity
4687 * 3/ A r5c cache log write is in flight.
4690 if (!sh
->reconstruct_state
&& !sh
->check_state
&& !sh
->log_io
) {
4691 if (!r5c_is_writeback(conf
->log
)) {
4693 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4694 } else { /* write back cache */
4697 /* First, try handle writes in caching phase */
4699 ret
= r5c_try_caching_write(conf
, sh
, &s
,
4702 * If caching phase failed: ret == -EAGAIN
4704 * stripe under reclaim: !caching && injournal
4706 * fall back to handle_stripe_dirtying()
4708 if (ret
== -EAGAIN
||
4709 /* stripe under reclaim: !caching && injournal */
4710 (!test_bit(STRIPE_R5C_CACHING
, &sh
->state
) &&
4712 ret
= handle_stripe_dirtying(conf
, sh
, &s
,
4720 /* maybe we need to check and possibly fix the parity for this stripe
4721 * Any reads will already have been scheduled, so we just see if enough
4722 * data is available. The parity check is held off while parity
4723 * dependent operations are in flight.
4725 if (sh
->check_state
||
4726 (s
.syncing
&& s
.locked
== 0 &&
4727 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4728 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4729 if (conf
->level
== 6)
4730 handle_parity_checks6(conf
, sh
, &s
, disks
);
4732 handle_parity_checks5(conf
, sh
, &s
, disks
);
4735 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4736 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4737 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4738 /* Write out to replacement devices where possible */
4739 for (i
= 0; i
< conf
->raid_disks
; i
++)
4740 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4741 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4742 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4743 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4747 set_bit(STRIPE_INSYNC
, &sh
->state
);
4748 set_bit(STRIPE_REPLACED
, &sh
->state
);
4750 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4751 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4752 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4753 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4754 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4755 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4756 wake_up(&conf
->wait_for_overlap
);
4759 /* If the failed drives are just a ReadError, then we might need
4760 * to progress the repair/check process
4762 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4763 for (i
= 0; i
< s
.failed
; i
++) {
4764 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4765 if (test_bit(R5_ReadError
, &dev
->flags
)
4766 && !test_bit(R5_LOCKED
, &dev
->flags
)
4767 && test_bit(R5_UPTODATE
, &dev
->flags
)
4769 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4770 set_bit(R5_Wantwrite
, &dev
->flags
);
4771 set_bit(R5_ReWrite
, &dev
->flags
);
4772 set_bit(R5_LOCKED
, &dev
->flags
);
4775 /* let's read it back */
4776 set_bit(R5_Wantread
, &dev
->flags
);
4777 set_bit(R5_LOCKED
, &dev
->flags
);
4783 /* Finish reconstruct operations initiated by the expansion process */
4784 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4785 struct stripe_head
*sh_src
4786 = raid5_get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4787 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4788 /* sh cannot be written until sh_src has been read.
4789 * so arrange for sh to be delayed a little
4791 set_bit(STRIPE_DELAYED
, &sh
->state
);
4792 set_bit(STRIPE_HANDLE
, &sh
->state
);
4793 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4795 atomic_inc(&conf
->preread_active_stripes
);
4796 raid5_release_stripe(sh_src
);
4800 raid5_release_stripe(sh_src
);
4802 sh
->reconstruct_state
= reconstruct_state_idle
;
4803 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4804 for (i
= conf
->raid_disks
; i
--; ) {
4805 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4806 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4811 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4812 !sh
->reconstruct_state
) {
4813 /* Need to write out all blocks after computing parity */
4814 sh
->disks
= conf
->raid_disks
;
4815 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4816 schedule_reconstruction(sh
, &s
, 1, 1);
4817 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4818 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4819 atomic_dec(&conf
->reshape_stripes
);
4820 wake_up(&conf
->wait_for_overlap
);
4821 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4824 if (s
.expanding
&& s
.locked
== 0 &&
4825 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4826 handle_stripe_expansion(conf
, sh
);
4829 /* wait for this device to become unblocked */
4830 if (unlikely(s
.blocked_rdev
)) {
4831 if (conf
->mddev
->external
)
4832 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4835 /* Internal metadata will immediately
4836 * be written by raid5d, so we don't
4837 * need to wait here.
4839 rdev_dec_pending(s
.blocked_rdev
,
4843 if (s
.handle_bad_blocks
)
4844 for (i
= disks
; i
--; ) {
4845 struct md_rdev
*rdev
;
4846 struct r5dev
*dev
= &sh
->dev
[i
];
4847 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4848 /* We own a safe reference to the rdev */
4849 rdev
= conf
->disks
[i
].rdev
;
4850 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4852 md_error(conf
->mddev
, rdev
);
4853 rdev_dec_pending(rdev
, conf
->mddev
);
4855 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4856 rdev
= conf
->disks
[i
].rdev
;
4857 rdev_clear_badblocks(rdev
, sh
->sector
,
4859 rdev_dec_pending(rdev
, conf
->mddev
);
4861 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4862 rdev
= conf
->disks
[i
].replacement
;
4864 /* rdev have been moved down */
4865 rdev
= conf
->disks
[i
].rdev
;
4866 rdev_clear_badblocks(rdev
, sh
->sector
,
4868 rdev_dec_pending(rdev
, conf
->mddev
);
4873 raid_run_ops(sh
, s
.ops_request
);
4877 if (s
.dec_preread_active
) {
4878 /* We delay this until after ops_run_io so that if make_request
4879 * is waiting on a flush, it won't continue until the writes
4880 * have actually been submitted.
4882 atomic_dec(&conf
->preread_active_stripes
);
4883 if (atomic_read(&conf
->preread_active_stripes
) <
4885 md_wakeup_thread(conf
->mddev
->thread
);
4888 if (!bio_list_empty(&s
.return_bi
)) {
4889 if (test_bit(MD_SB_CHANGE_PENDING
, &conf
->mddev
->sb_flags
)) {
4890 spin_lock_irq(&conf
->device_lock
);
4891 bio_list_merge(&conf
->return_bi
, &s
.return_bi
);
4892 spin_unlock_irq(&conf
->device_lock
);
4893 md_wakeup_thread(conf
->mddev
->thread
);
4895 return_io(&s
.return_bi
);
4898 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4901 static void raid5_activate_delayed(struct r5conf
*conf
)
4903 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
4904 while (!list_empty(&conf
->delayed_list
)) {
4905 struct list_head
*l
= conf
->delayed_list
.next
;
4906 struct stripe_head
*sh
;
4907 sh
= list_entry(l
, struct stripe_head
, lru
);
4909 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4910 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4911 atomic_inc(&conf
->preread_active_stripes
);
4912 list_add_tail(&sh
->lru
, &conf
->hold_list
);
4913 raid5_wakeup_stripe_thread(sh
);
4918 static void activate_bit_delay(struct r5conf
*conf
,
4919 struct list_head
*temp_inactive_list
)
4921 /* device_lock is held */
4922 struct list_head head
;
4923 list_add(&head
, &conf
->bitmap_list
);
4924 list_del_init(&conf
->bitmap_list
);
4925 while (!list_empty(&head
)) {
4926 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
4928 list_del_init(&sh
->lru
);
4929 atomic_inc(&sh
->count
);
4930 hash
= sh
->hash_lock_index
;
4931 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
4935 static int raid5_congested(struct mddev
*mddev
, int bits
)
4937 struct r5conf
*conf
= mddev
->private;
4939 /* No difference between reads and writes. Just check
4940 * how busy the stripe_cache is
4943 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
4946 /* Also checks whether there is pressure on r5cache log space */
4947 if (test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
))
4951 if (atomic_read(&conf
->empty_inactive_list_nr
))
4957 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
4959 struct r5conf
*conf
= mddev
->private;
4960 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
4961 unsigned int chunk_sectors
;
4962 unsigned int bio_sectors
= bio_sectors(bio
);
4964 chunk_sectors
= min(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
4965 return chunk_sectors
>=
4966 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
4970 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4971 * later sampled by raid5d.
4973 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
4975 unsigned long flags
;
4977 spin_lock_irqsave(&conf
->device_lock
, flags
);
4979 bi
->bi_next
= conf
->retry_read_aligned_list
;
4980 conf
->retry_read_aligned_list
= bi
;
4982 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4983 md_wakeup_thread(conf
->mddev
->thread
);
4986 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
4990 bi
= conf
->retry_read_aligned
;
4992 conf
->retry_read_aligned
= NULL
;
4995 bi
= conf
->retry_read_aligned_list
;
4997 conf
->retry_read_aligned_list
= bi
->bi_next
;
5000 * this sets the active strip count to 1 and the processed
5001 * strip count to zero (upper 8 bits)
5003 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
5010 * The "raid5_align_endio" should check if the read succeeded and if it
5011 * did, call bio_endio on the original bio (having bio_put the new bio
5013 * If the read failed..
5015 static void raid5_align_endio(struct bio
*bi
)
5017 struct bio
* raid_bi
= bi
->bi_private
;
5018 struct mddev
*mddev
;
5019 struct r5conf
*conf
;
5020 struct md_rdev
*rdev
;
5021 int error
= bi
->bi_error
;
5025 rdev
= (void*)raid_bi
->bi_next
;
5026 raid_bi
->bi_next
= NULL
;
5027 mddev
= rdev
->mddev
;
5028 conf
= mddev
->private;
5030 rdev_dec_pending(rdev
, conf
->mddev
);
5033 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
5036 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5037 wake_up(&conf
->wait_for_quiescent
);
5041 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5043 add_bio_to_retry(raid_bi
, conf
);
5046 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
5048 struct r5conf
*conf
= mddev
->private;
5050 struct bio
* align_bi
;
5051 struct md_rdev
*rdev
;
5052 sector_t end_sector
;
5054 if (!in_chunk_boundary(mddev
, raid_bio
)) {
5055 pr_debug("%s: non aligned\n", __func__
);
5059 * use bio_clone_fast to make a copy of the bio
5061 align_bi
= bio_clone_fast(raid_bio
, GFP_NOIO
, mddev
->bio_set
);
5065 * set bi_end_io to a new function, and set bi_private to the
5068 align_bi
->bi_end_io
= raid5_align_endio
;
5069 align_bi
->bi_private
= raid_bio
;
5073 align_bi
->bi_iter
.bi_sector
=
5074 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
5077 end_sector
= bio_end_sector(align_bi
);
5079 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
5080 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
5081 rdev
->recovery_offset
< end_sector
) {
5082 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
5084 (test_bit(Faulty
, &rdev
->flags
) ||
5085 !(test_bit(In_sync
, &rdev
->flags
) ||
5086 rdev
->recovery_offset
>= end_sector
)))
5090 if (r5c_big_stripe_cached(conf
, align_bi
->bi_iter
.bi_sector
)) {
5100 atomic_inc(&rdev
->nr_pending
);
5102 raid_bio
->bi_next
= (void*)rdev
;
5103 align_bi
->bi_bdev
= rdev
->bdev
;
5104 bio_clear_flag(align_bi
, BIO_SEG_VALID
);
5106 if (is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
5107 bio_sectors(align_bi
),
5108 &first_bad
, &bad_sectors
)) {
5110 rdev_dec_pending(rdev
, mddev
);
5114 /* No reshape active, so we can trust rdev->data_offset */
5115 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
5117 spin_lock_irq(&conf
->device_lock
);
5118 wait_event_lock_irq(conf
->wait_for_quiescent
,
5121 atomic_inc(&conf
->active_aligned_reads
);
5122 spin_unlock_irq(&conf
->device_lock
);
5125 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
5126 align_bi
, disk_devt(mddev
->gendisk
),
5127 raid_bio
->bi_iter
.bi_sector
);
5128 generic_make_request(align_bi
);
5137 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
5142 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
5143 unsigned chunk_sects
= mddev
->chunk_sectors
;
5144 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
5146 if (sectors
< bio_sectors(raid_bio
)) {
5147 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, fs_bio_set
);
5148 bio_chain(split
, raid_bio
);
5152 if (!raid5_read_one_chunk(mddev
, split
)) {
5153 if (split
!= raid_bio
)
5154 generic_make_request(raid_bio
);
5157 } while (split
!= raid_bio
);
5162 /* __get_priority_stripe - get the next stripe to process
5164 * Full stripe writes are allowed to pass preread active stripes up until
5165 * the bypass_threshold is exceeded. In general the bypass_count
5166 * increments when the handle_list is handled before the hold_list; however, it
5167 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5168 * stripe with in flight i/o. The bypass_count will be reset when the
5169 * head of the hold_list has changed, i.e. the head was promoted to the
5172 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
5174 struct stripe_head
*sh
= NULL
, *tmp
;
5175 struct list_head
*handle_list
= NULL
;
5176 struct r5worker_group
*wg
= NULL
;
5178 if (conf
->worker_cnt_per_group
== 0) {
5179 handle_list
= &conf
->handle_list
;
5180 } else if (group
!= ANY_GROUP
) {
5181 handle_list
= &conf
->worker_groups
[group
].handle_list
;
5182 wg
= &conf
->worker_groups
[group
];
5185 for (i
= 0; i
< conf
->group_cnt
; i
++) {
5186 handle_list
= &conf
->worker_groups
[i
].handle_list
;
5187 wg
= &conf
->worker_groups
[i
];
5188 if (!list_empty(handle_list
))
5193 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5195 list_empty(handle_list
) ? "empty" : "busy",
5196 list_empty(&conf
->hold_list
) ? "empty" : "busy",
5197 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
5199 if (!list_empty(handle_list
)) {
5200 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
5202 if (list_empty(&conf
->hold_list
))
5203 conf
->bypass_count
= 0;
5204 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
5205 if (conf
->hold_list
.next
== conf
->last_hold
)
5206 conf
->bypass_count
++;
5208 conf
->last_hold
= conf
->hold_list
.next
;
5209 conf
->bypass_count
-= conf
->bypass_threshold
;
5210 if (conf
->bypass_count
< 0)
5211 conf
->bypass_count
= 0;
5214 } else if (!list_empty(&conf
->hold_list
) &&
5215 ((conf
->bypass_threshold
&&
5216 conf
->bypass_count
> conf
->bypass_threshold
) ||
5217 atomic_read(&conf
->pending_full_writes
) == 0)) {
5219 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
5220 if (conf
->worker_cnt_per_group
== 0 ||
5221 group
== ANY_GROUP
||
5222 !cpu_online(tmp
->cpu
) ||
5223 cpu_to_group(tmp
->cpu
) == group
) {
5230 conf
->bypass_count
-= conf
->bypass_threshold
;
5231 if (conf
->bypass_count
< 0)
5232 conf
->bypass_count
= 0;
5244 list_del_init(&sh
->lru
);
5245 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
5249 struct raid5_plug_cb
{
5250 struct blk_plug_cb cb
;
5251 struct list_head list
;
5252 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
5255 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
5257 struct raid5_plug_cb
*cb
= container_of(
5258 blk_cb
, struct raid5_plug_cb
, cb
);
5259 struct stripe_head
*sh
;
5260 struct mddev
*mddev
= cb
->cb
.data
;
5261 struct r5conf
*conf
= mddev
->private;
5265 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
5266 spin_lock_irq(&conf
->device_lock
);
5267 while (!list_empty(&cb
->list
)) {
5268 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
5269 list_del_init(&sh
->lru
);
5271 * avoid race release_stripe_plug() sees
5272 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5273 * is still in our list
5275 smp_mb__before_atomic();
5276 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
5278 * STRIPE_ON_RELEASE_LIST could be set here. In that
5279 * case, the count is always > 1 here
5281 hash
= sh
->hash_lock_index
;
5282 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
5285 spin_unlock_irq(&conf
->device_lock
);
5287 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
5288 NR_STRIPE_HASH_LOCKS
);
5290 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5294 static void release_stripe_plug(struct mddev
*mddev
,
5295 struct stripe_head
*sh
)
5297 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5298 raid5_unplug
, mddev
,
5299 sizeof(struct raid5_plug_cb
));
5300 struct raid5_plug_cb
*cb
;
5303 raid5_release_stripe(sh
);
5307 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5309 if (cb
->list
.next
== NULL
) {
5311 INIT_LIST_HEAD(&cb
->list
);
5312 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5313 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5316 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5317 list_add_tail(&sh
->lru
, &cb
->list
);
5319 raid5_release_stripe(sh
);
5322 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5324 struct r5conf
*conf
= mddev
->private;
5325 sector_t logical_sector
, last_sector
;
5326 struct stripe_head
*sh
;
5330 if (mddev
->reshape_position
!= MaxSector
)
5331 /* Skip discard while reshape is happening */
5334 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5335 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5338 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5340 stripe_sectors
= conf
->chunk_sectors
*
5341 (conf
->raid_disks
- conf
->max_degraded
);
5342 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5344 sector_div(last_sector
, stripe_sectors
);
5346 logical_sector
*= conf
->chunk_sectors
;
5347 last_sector
*= conf
->chunk_sectors
;
5349 for (; logical_sector
< last_sector
;
5350 logical_sector
+= STRIPE_SECTORS
) {
5354 sh
= raid5_get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5355 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5356 TASK_UNINTERRUPTIBLE
);
5357 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5358 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5359 raid5_release_stripe(sh
);
5363 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5364 spin_lock_irq(&sh
->stripe_lock
);
5365 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5366 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5368 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5369 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5370 spin_unlock_irq(&sh
->stripe_lock
);
5371 raid5_release_stripe(sh
);
5376 set_bit(STRIPE_DISCARD
, &sh
->state
);
5377 finish_wait(&conf
->wait_for_overlap
, &w
);
5378 sh
->overwrite_disks
= 0;
5379 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5380 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5382 sh
->dev
[d
].towrite
= bi
;
5383 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5384 raid5_inc_bi_active_stripes(bi
);
5385 sh
->overwrite_disks
++;
5387 spin_unlock_irq(&sh
->stripe_lock
);
5388 if (conf
->mddev
->bitmap
) {
5390 d
< conf
->raid_disks
- conf
->max_degraded
;
5392 bitmap_startwrite(mddev
->bitmap
,
5396 sh
->bm_seq
= conf
->seq_flush
+ 1;
5397 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5400 set_bit(STRIPE_HANDLE
, &sh
->state
);
5401 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5402 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5403 atomic_inc(&conf
->preread_active_stripes
);
5404 release_stripe_plug(mddev
, sh
);
5407 remaining
= raid5_dec_bi_active_stripes(bi
);
5408 if (remaining
== 0) {
5409 md_write_end(mddev
);
5414 static void raid5_make_request(struct mddev
*mddev
, struct bio
* bi
)
5416 struct r5conf
*conf
= mddev
->private;
5418 sector_t new_sector
;
5419 sector_t logical_sector
, last_sector
;
5420 struct stripe_head
*sh
;
5421 const int rw
= bio_data_dir(bi
);
5425 bool do_flush
= false;
5427 if (unlikely(bi
->bi_opf
& REQ_PREFLUSH
)) {
5428 int ret
= r5l_handle_flush_request(conf
->log
, bi
);
5432 if (ret
== -ENODEV
) {
5433 md_flush_request(mddev
, bi
);
5436 /* ret == -EAGAIN, fallback */
5438 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5439 * we need to flush journal device
5441 do_flush
= bi
->bi_opf
& REQ_PREFLUSH
;
5444 md_write_start(mddev
, bi
);
5447 * If array is degraded, better not do chunk aligned read because
5448 * later we might have to read it again in order to reconstruct
5449 * data on failed drives.
5451 if (rw
== READ
&& mddev
->degraded
== 0 &&
5452 mddev
->reshape_position
== MaxSector
) {
5453 bi
= chunk_aligned_read(mddev
, bi
);
5458 if (unlikely(bio_op(bi
) == REQ_OP_DISCARD
)) {
5459 make_discard_request(mddev
, bi
);
5463 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5464 last_sector
= bio_end_sector(bi
);
5466 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5468 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5469 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5475 seq
= read_seqcount_begin(&conf
->gen_lock
);
5478 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5479 TASK_UNINTERRUPTIBLE
);
5480 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5481 /* spinlock is needed as reshape_progress may be
5482 * 64bit on a 32bit platform, and so it might be
5483 * possible to see a half-updated value
5484 * Of course reshape_progress could change after
5485 * the lock is dropped, so once we get a reference
5486 * to the stripe that we think it is, we will have
5489 spin_lock_irq(&conf
->device_lock
);
5490 if (mddev
->reshape_backwards
5491 ? logical_sector
< conf
->reshape_progress
5492 : logical_sector
>= conf
->reshape_progress
) {
5495 if (mddev
->reshape_backwards
5496 ? logical_sector
< conf
->reshape_safe
5497 : logical_sector
>= conf
->reshape_safe
) {
5498 spin_unlock_irq(&conf
->device_lock
);
5504 spin_unlock_irq(&conf
->device_lock
);
5507 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5510 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5511 (unsigned long long)new_sector
,
5512 (unsigned long long)logical_sector
);
5514 sh
= raid5_get_active_stripe(conf
, new_sector
, previous
,
5515 (bi
->bi_opf
& REQ_RAHEAD
), 0);
5517 if (unlikely(previous
)) {
5518 /* expansion might have moved on while waiting for a
5519 * stripe, so we must do the range check again.
5520 * Expansion could still move past after this
5521 * test, but as we are holding a reference to
5522 * 'sh', we know that if that happens,
5523 * STRIPE_EXPANDING will get set and the expansion
5524 * won't proceed until we finish with the stripe.
5527 spin_lock_irq(&conf
->device_lock
);
5528 if (mddev
->reshape_backwards
5529 ? logical_sector
>= conf
->reshape_progress
5530 : logical_sector
< conf
->reshape_progress
)
5531 /* mismatch, need to try again */
5533 spin_unlock_irq(&conf
->device_lock
);
5535 raid5_release_stripe(sh
);
5541 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5542 /* Might have got the wrong stripe_head
5545 raid5_release_stripe(sh
);
5550 logical_sector
>= mddev
->suspend_lo
&&
5551 logical_sector
< mddev
->suspend_hi
) {
5552 raid5_release_stripe(sh
);
5553 /* As the suspend_* range is controlled by
5554 * userspace, we want an interruptible
5557 flush_signals(current
);
5558 prepare_to_wait(&conf
->wait_for_overlap
,
5559 &w
, TASK_INTERRUPTIBLE
);
5560 if (logical_sector
>= mddev
->suspend_lo
&&
5561 logical_sector
< mddev
->suspend_hi
) {
5568 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5569 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5570 /* Stripe is busy expanding or
5571 * add failed due to overlap. Flush everything
5574 md_wakeup_thread(mddev
->thread
);
5575 raid5_release_stripe(sh
);
5581 set_bit(STRIPE_R5C_PREFLUSH
, &sh
->state
);
5582 /* we only need flush for one stripe */
5586 set_bit(STRIPE_HANDLE
, &sh
->state
);
5587 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5588 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5589 (bi
->bi_opf
& REQ_SYNC
) &&
5590 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5591 atomic_inc(&conf
->preread_active_stripes
);
5592 release_stripe_plug(mddev
, sh
);
5594 /* cannot get stripe for read-ahead, just give-up */
5595 bi
->bi_error
= -EIO
;
5599 finish_wait(&conf
->wait_for_overlap
, &w
);
5601 remaining
= raid5_dec_bi_active_stripes(bi
);
5602 if (remaining
== 0) {
5605 md_write_end(mddev
);
5607 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
5613 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5615 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5617 /* reshaping is quite different to recovery/resync so it is
5618 * handled quite separately ... here.
5620 * On each call to sync_request, we gather one chunk worth of
5621 * destination stripes and flag them as expanding.
5622 * Then we find all the source stripes and request reads.
5623 * As the reads complete, handle_stripe will copy the data
5624 * into the destination stripe and release that stripe.
5626 struct r5conf
*conf
= mddev
->private;
5627 struct stripe_head
*sh
;
5628 sector_t first_sector
, last_sector
;
5629 int raid_disks
= conf
->previous_raid_disks
;
5630 int data_disks
= raid_disks
- conf
->max_degraded
;
5631 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5634 sector_t writepos
, readpos
, safepos
;
5635 sector_t stripe_addr
;
5636 int reshape_sectors
;
5637 struct list_head stripes
;
5640 if (sector_nr
== 0) {
5641 /* If restarting in the middle, skip the initial sectors */
5642 if (mddev
->reshape_backwards
&&
5643 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5644 sector_nr
= raid5_size(mddev
, 0, 0)
5645 - conf
->reshape_progress
;
5646 } else if (mddev
->reshape_backwards
&&
5647 conf
->reshape_progress
== MaxSector
) {
5648 /* shouldn't happen, but just in case, finish up.*/
5649 sector_nr
= MaxSector
;
5650 } else if (!mddev
->reshape_backwards
&&
5651 conf
->reshape_progress
> 0)
5652 sector_nr
= conf
->reshape_progress
;
5653 sector_div(sector_nr
, new_data_disks
);
5655 mddev
->curr_resync_completed
= sector_nr
;
5656 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5663 /* We need to process a full chunk at a time.
5664 * If old and new chunk sizes differ, we need to process the
5668 reshape_sectors
= max(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5670 /* We update the metadata at least every 10 seconds, or when
5671 * the data about to be copied would over-write the source of
5672 * the data at the front of the range. i.e. one new_stripe
5673 * along from reshape_progress new_maps to after where
5674 * reshape_safe old_maps to
5676 writepos
= conf
->reshape_progress
;
5677 sector_div(writepos
, new_data_disks
);
5678 readpos
= conf
->reshape_progress
;
5679 sector_div(readpos
, data_disks
);
5680 safepos
= conf
->reshape_safe
;
5681 sector_div(safepos
, data_disks
);
5682 if (mddev
->reshape_backwards
) {
5683 BUG_ON(writepos
< reshape_sectors
);
5684 writepos
-= reshape_sectors
;
5685 readpos
+= reshape_sectors
;
5686 safepos
+= reshape_sectors
;
5688 writepos
+= reshape_sectors
;
5689 /* readpos and safepos are worst-case calculations.
5690 * A negative number is overly pessimistic, and causes
5691 * obvious problems for unsigned storage. So clip to 0.
5693 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5694 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5697 /* Having calculated the 'writepos' possibly use it
5698 * to set 'stripe_addr' which is where we will write to.
5700 if (mddev
->reshape_backwards
) {
5701 BUG_ON(conf
->reshape_progress
== 0);
5702 stripe_addr
= writepos
;
5703 BUG_ON((mddev
->dev_sectors
&
5704 ~((sector_t
)reshape_sectors
- 1))
5705 - reshape_sectors
- stripe_addr
5708 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5709 stripe_addr
= sector_nr
;
5712 /* 'writepos' is the most advanced device address we might write.
5713 * 'readpos' is the least advanced device address we might read.
5714 * 'safepos' is the least address recorded in the metadata as having
5716 * If there is a min_offset_diff, these are adjusted either by
5717 * increasing the safepos/readpos if diff is negative, or
5718 * increasing writepos if diff is positive.
5719 * If 'readpos' is then behind 'writepos', there is no way that we can
5720 * ensure safety in the face of a crash - that must be done by userspace
5721 * making a backup of the data. So in that case there is no particular
5722 * rush to update metadata.
5723 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5724 * update the metadata to advance 'safepos' to match 'readpos' so that
5725 * we can be safe in the event of a crash.
5726 * So we insist on updating metadata if safepos is behind writepos and
5727 * readpos is beyond writepos.
5728 * In any case, update the metadata every 10 seconds.
5729 * Maybe that number should be configurable, but I'm not sure it is
5730 * worth it.... maybe it could be a multiple of safemode_delay???
5732 if (conf
->min_offset_diff
< 0) {
5733 safepos
+= -conf
->min_offset_diff
;
5734 readpos
+= -conf
->min_offset_diff
;
5736 writepos
+= conf
->min_offset_diff
;
5738 if ((mddev
->reshape_backwards
5739 ? (safepos
> writepos
&& readpos
< writepos
)
5740 : (safepos
< writepos
&& readpos
> writepos
)) ||
5741 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5742 /* Cannot proceed until we've updated the superblock... */
5743 wait_event(conf
->wait_for_overlap
,
5744 atomic_read(&conf
->reshape_stripes
)==0
5745 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5746 if (atomic_read(&conf
->reshape_stripes
) != 0)
5748 mddev
->reshape_position
= conf
->reshape_progress
;
5749 mddev
->curr_resync_completed
= sector_nr
;
5750 conf
->reshape_checkpoint
= jiffies
;
5751 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5752 md_wakeup_thread(mddev
->thread
);
5753 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
5754 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5755 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5757 spin_lock_irq(&conf
->device_lock
);
5758 conf
->reshape_safe
= mddev
->reshape_position
;
5759 spin_unlock_irq(&conf
->device_lock
);
5760 wake_up(&conf
->wait_for_overlap
);
5761 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5764 INIT_LIST_HEAD(&stripes
);
5765 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5767 int skipped_disk
= 0;
5768 sh
= raid5_get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5769 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5770 atomic_inc(&conf
->reshape_stripes
);
5771 /* If any of this stripe is beyond the end of the old
5772 * array, then we need to zero those blocks
5774 for (j
=sh
->disks
; j
--;) {
5776 if (j
== sh
->pd_idx
)
5778 if (conf
->level
== 6 &&
5781 s
= raid5_compute_blocknr(sh
, j
, 0);
5782 if (s
< raid5_size(mddev
, 0, 0)) {
5786 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5787 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5788 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5790 if (!skipped_disk
) {
5791 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5792 set_bit(STRIPE_HANDLE
, &sh
->state
);
5794 list_add(&sh
->lru
, &stripes
);
5796 spin_lock_irq(&conf
->device_lock
);
5797 if (mddev
->reshape_backwards
)
5798 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5800 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5801 spin_unlock_irq(&conf
->device_lock
);
5802 /* Ok, those stripe are ready. We can start scheduling
5803 * reads on the source stripes.
5804 * The source stripes are determined by mapping the first and last
5805 * block on the destination stripes.
5808 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5811 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5812 * new_data_disks
- 1),
5814 if (last_sector
>= mddev
->dev_sectors
)
5815 last_sector
= mddev
->dev_sectors
- 1;
5816 while (first_sector
<= last_sector
) {
5817 sh
= raid5_get_active_stripe(conf
, first_sector
, 1, 0, 1);
5818 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5819 set_bit(STRIPE_HANDLE
, &sh
->state
);
5820 raid5_release_stripe(sh
);
5821 first_sector
+= STRIPE_SECTORS
;
5823 /* Now that the sources are clearly marked, we can release
5824 * the destination stripes
5826 while (!list_empty(&stripes
)) {
5827 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5828 list_del_init(&sh
->lru
);
5829 raid5_release_stripe(sh
);
5831 /* If this takes us to the resync_max point where we have to pause,
5832 * then we need to write out the superblock.
5834 sector_nr
+= reshape_sectors
;
5835 retn
= reshape_sectors
;
5837 if (mddev
->curr_resync_completed
> mddev
->resync_max
||
5838 (sector_nr
- mddev
->curr_resync_completed
) * 2
5839 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5840 /* Cannot proceed until we've updated the superblock... */
5841 wait_event(conf
->wait_for_overlap
,
5842 atomic_read(&conf
->reshape_stripes
) == 0
5843 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5844 if (atomic_read(&conf
->reshape_stripes
) != 0)
5846 mddev
->reshape_position
= conf
->reshape_progress
;
5847 mddev
->curr_resync_completed
= sector_nr
;
5848 conf
->reshape_checkpoint
= jiffies
;
5849 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5850 md_wakeup_thread(mddev
->thread
);
5851 wait_event(mddev
->sb_wait
,
5852 !test_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
)
5853 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5854 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5856 spin_lock_irq(&conf
->device_lock
);
5857 conf
->reshape_safe
= mddev
->reshape_position
;
5858 spin_unlock_irq(&conf
->device_lock
);
5859 wake_up(&conf
->wait_for_overlap
);
5860 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5866 static inline sector_t
raid5_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
5869 struct r5conf
*conf
= mddev
->private;
5870 struct stripe_head
*sh
;
5871 sector_t max_sector
= mddev
->dev_sectors
;
5872 sector_t sync_blocks
;
5873 int still_degraded
= 0;
5876 if (sector_nr
>= max_sector
) {
5877 /* just being told to finish up .. nothing much to do */
5879 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
5884 if (mddev
->curr_resync
< max_sector
) /* aborted */
5885 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
5887 else /* completed sync */
5889 bitmap_close_sync(mddev
->bitmap
);
5894 /* Allow raid5_quiesce to complete */
5895 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
5897 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
5898 return reshape_request(mddev
, sector_nr
, skipped
);
5900 /* No need to check resync_max as we never do more than one
5901 * stripe, and as resync_max will always be on a chunk boundary,
5902 * if the check in md_do_sync didn't fire, there is no chance
5903 * of overstepping resync_max here
5906 /* if there is too many failed drives and we are trying
5907 * to resync, then assert that we are finished, because there is
5908 * nothing we can do.
5910 if (mddev
->degraded
>= conf
->max_degraded
&&
5911 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
5912 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
5916 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
5918 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
5919 sync_blocks
>= STRIPE_SECTORS
) {
5920 /* we can skip this block, and probably more */
5921 sync_blocks
/= STRIPE_SECTORS
;
5923 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
5926 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, false);
5928 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 1, 0);
5930 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 0, 0);
5931 /* make sure we don't swamp the stripe cache if someone else
5932 * is trying to get access
5934 schedule_timeout_uninterruptible(1);
5936 /* Need to check if array will still be degraded after recovery/resync
5937 * Note in case of > 1 drive failures it's possible we're rebuilding
5938 * one drive while leaving another faulty drive in array.
5941 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5942 struct md_rdev
*rdev
= ACCESS_ONCE(conf
->disks
[i
].rdev
);
5944 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
5949 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
5951 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
5952 set_bit(STRIPE_HANDLE
, &sh
->state
);
5954 raid5_release_stripe(sh
);
5956 return STRIPE_SECTORS
;
5959 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
5961 /* We may not be able to submit a whole bio at once as there
5962 * may not be enough stripe_heads available.
5963 * We cannot pre-allocate enough stripe_heads as we may need
5964 * more than exist in the cache (if we allow ever large chunks).
5965 * So we do one stripe head at a time and record in
5966 * ->bi_hw_segments how many have been done.
5968 * We *know* that this entire raid_bio is in one chunk, so
5969 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5971 struct stripe_head
*sh
;
5973 sector_t sector
, logical_sector
, last_sector
;
5978 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
5979 ~((sector_t
)STRIPE_SECTORS
-1);
5980 sector
= raid5_compute_sector(conf
, logical_sector
,
5982 last_sector
= bio_end_sector(raid_bio
);
5984 for (; logical_sector
< last_sector
;
5985 logical_sector
+= STRIPE_SECTORS
,
5986 sector
+= STRIPE_SECTORS
,
5989 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
5990 /* already done this stripe */
5993 sh
= raid5_get_active_stripe(conf
, sector
, 0, 1, 1);
5996 /* failed to get a stripe - must wait */
5997 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5998 conf
->retry_read_aligned
= raid_bio
;
6002 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
6003 raid5_release_stripe(sh
);
6004 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
6005 conf
->retry_read_aligned
= raid_bio
;
6009 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
6011 raid5_release_stripe(sh
);
6014 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
6015 if (remaining
== 0) {
6016 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
6018 bio_endio(raid_bio
);
6020 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
6021 wake_up(&conf
->wait_for_quiescent
);
6025 static int handle_active_stripes(struct r5conf
*conf
, int group
,
6026 struct r5worker
*worker
,
6027 struct list_head
*temp_inactive_list
)
6029 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
6030 int i
, batch_size
= 0, hash
;
6031 bool release_inactive
= false;
6033 while (batch_size
< MAX_STRIPE_BATCH
&&
6034 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
6035 batch
[batch_size
++] = sh
;
6037 if (batch_size
== 0) {
6038 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6039 if (!list_empty(temp_inactive_list
+ i
))
6041 if (i
== NR_STRIPE_HASH_LOCKS
) {
6042 spin_unlock_irq(&conf
->device_lock
);
6043 r5l_flush_stripe_to_raid(conf
->log
);
6044 spin_lock_irq(&conf
->device_lock
);
6047 release_inactive
= true;
6049 spin_unlock_irq(&conf
->device_lock
);
6051 release_inactive_stripe_list(conf
, temp_inactive_list
,
6052 NR_STRIPE_HASH_LOCKS
);
6054 r5l_flush_stripe_to_raid(conf
->log
);
6055 if (release_inactive
) {
6056 spin_lock_irq(&conf
->device_lock
);
6060 for (i
= 0; i
< batch_size
; i
++)
6061 handle_stripe(batch
[i
]);
6062 r5l_write_stripe_run(conf
->log
);
6066 spin_lock_irq(&conf
->device_lock
);
6067 for (i
= 0; i
< batch_size
; i
++) {
6068 hash
= batch
[i
]->hash_lock_index
;
6069 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
6074 static void raid5_do_work(struct work_struct
*work
)
6076 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
6077 struct r5worker_group
*group
= worker
->group
;
6078 struct r5conf
*conf
= group
->conf
;
6079 int group_id
= group
- conf
->worker_groups
;
6081 struct blk_plug plug
;
6083 pr_debug("+++ raid5worker active\n");
6085 blk_start_plug(&plug
);
6087 spin_lock_irq(&conf
->device_lock
);
6089 int batch_size
, released
;
6091 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
6093 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
6094 worker
->temp_inactive_list
);
6095 worker
->working
= false;
6096 if (!batch_size
&& !released
)
6098 handled
+= batch_size
;
6100 pr_debug("%d stripes handled\n", handled
);
6102 spin_unlock_irq(&conf
->device_lock
);
6103 blk_finish_plug(&plug
);
6105 pr_debug("--- raid5worker inactive\n");
6109 * This is our raid5 kernel thread.
6111 * We scan the hash table for stripes which can be handled now.
6112 * During the scan, completed stripes are saved for us by the interrupt
6113 * handler, so that they will not have to wait for our next wakeup.
6115 static void raid5d(struct md_thread
*thread
)
6117 struct mddev
*mddev
= thread
->mddev
;
6118 struct r5conf
*conf
= mddev
->private;
6120 struct blk_plug plug
;
6122 pr_debug("+++ raid5d active\n");
6124 md_check_recovery(mddev
);
6126 if (!bio_list_empty(&conf
->return_bi
) &&
6127 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
6128 struct bio_list tmp
= BIO_EMPTY_LIST
;
6129 spin_lock_irq(&conf
->device_lock
);
6130 if (!test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
6131 bio_list_merge(&tmp
, &conf
->return_bi
);
6132 bio_list_init(&conf
->return_bi
);
6134 spin_unlock_irq(&conf
->device_lock
);
6138 blk_start_plug(&plug
);
6140 spin_lock_irq(&conf
->device_lock
);
6143 int batch_size
, released
;
6145 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
6147 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6150 !list_empty(&conf
->bitmap_list
)) {
6151 /* Now is a good time to flush some bitmap updates */
6153 spin_unlock_irq(&conf
->device_lock
);
6154 bitmap_unplug(mddev
->bitmap
);
6155 spin_lock_irq(&conf
->device_lock
);
6156 conf
->seq_write
= conf
->seq_flush
;
6157 activate_bit_delay(conf
, conf
->temp_inactive_list
);
6159 raid5_activate_delayed(conf
);
6161 while ((bio
= remove_bio_from_retry(conf
))) {
6163 spin_unlock_irq(&conf
->device_lock
);
6164 ok
= retry_aligned_read(conf
, bio
);
6165 spin_lock_irq(&conf
->device_lock
);
6171 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
6172 conf
->temp_inactive_list
);
6173 if (!batch_size
&& !released
)
6175 handled
+= batch_size
;
6177 if (mddev
->sb_flags
& ~(1 << MD_SB_CHANGE_PENDING
)) {
6178 spin_unlock_irq(&conf
->device_lock
);
6179 md_check_recovery(mddev
);
6180 spin_lock_irq(&conf
->device_lock
);
6183 pr_debug("%d stripes handled\n", handled
);
6185 spin_unlock_irq(&conf
->device_lock
);
6186 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
) &&
6187 mutex_trylock(&conf
->cache_size_mutex
)) {
6188 grow_one_stripe(conf
, __GFP_NOWARN
);
6189 /* Set flag even if allocation failed. This helps
6190 * slow down allocation requests when mem is short
6192 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6193 mutex_unlock(&conf
->cache_size_mutex
);
6196 flush_deferred_bios(conf
);
6198 r5l_flush_stripe_to_raid(conf
->log
);
6200 async_tx_issue_pending_all();
6201 blk_finish_plug(&plug
);
6203 pr_debug("--- raid5d inactive\n");
6207 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
6209 struct r5conf
*conf
;
6211 spin_lock(&mddev
->lock
);
6212 conf
= mddev
->private;
6214 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
6215 spin_unlock(&mddev
->lock
);
6220 raid5_set_cache_size(struct mddev
*mddev
, int size
)
6222 struct r5conf
*conf
= mddev
->private;
6225 if (size
<= 16 || size
> 32768)
6228 conf
->min_nr_stripes
= size
;
6229 mutex_lock(&conf
->cache_size_mutex
);
6230 while (size
< conf
->max_nr_stripes
&&
6231 drop_one_stripe(conf
))
6233 mutex_unlock(&conf
->cache_size_mutex
);
6236 err
= md_allow_write(mddev
);
6240 mutex_lock(&conf
->cache_size_mutex
);
6241 while (size
> conf
->max_nr_stripes
)
6242 if (!grow_one_stripe(conf
, GFP_KERNEL
))
6244 mutex_unlock(&conf
->cache_size_mutex
);
6248 EXPORT_SYMBOL(raid5_set_cache_size
);
6251 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
6253 struct r5conf
*conf
;
6257 if (len
>= PAGE_SIZE
)
6259 if (kstrtoul(page
, 10, &new))
6261 err
= mddev_lock(mddev
);
6264 conf
= mddev
->private;
6268 err
= raid5_set_cache_size(mddev
, new);
6269 mddev_unlock(mddev
);
6274 static struct md_sysfs_entry
6275 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
6276 raid5_show_stripe_cache_size
,
6277 raid5_store_stripe_cache_size
);
6280 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
6282 struct r5conf
*conf
= mddev
->private;
6284 return sprintf(page
, "%d\n", conf
->rmw_level
);
6290 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
6292 struct r5conf
*conf
= mddev
->private;
6298 if (len
>= PAGE_SIZE
)
6301 if (kstrtoul(page
, 10, &new))
6304 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
6307 if (new != PARITY_DISABLE_RMW
&&
6308 new != PARITY_ENABLE_RMW
&&
6309 new != PARITY_PREFER_RMW
)
6312 conf
->rmw_level
= new;
6316 static struct md_sysfs_entry
6317 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
6318 raid5_show_rmw_level
,
6319 raid5_store_rmw_level
);
6323 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
6325 struct r5conf
*conf
;
6327 spin_lock(&mddev
->lock
);
6328 conf
= mddev
->private;
6330 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
6331 spin_unlock(&mddev
->lock
);
6336 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
6338 struct r5conf
*conf
;
6342 if (len
>= PAGE_SIZE
)
6344 if (kstrtoul(page
, 10, &new))
6347 err
= mddev_lock(mddev
);
6350 conf
= mddev
->private;
6353 else if (new > conf
->min_nr_stripes
)
6356 conf
->bypass_threshold
= new;
6357 mddev_unlock(mddev
);
6361 static struct md_sysfs_entry
6362 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
6364 raid5_show_preread_threshold
,
6365 raid5_store_preread_threshold
);
6368 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
6370 struct r5conf
*conf
;
6372 spin_lock(&mddev
->lock
);
6373 conf
= mddev
->private;
6375 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6376 spin_unlock(&mddev
->lock
);
6381 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6383 struct r5conf
*conf
;
6387 if (len
>= PAGE_SIZE
)
6389 if (kstrtoul(page
, 10, &new))
6393 err
= mddev_lock(mddev
);
6396 conf
= mddev
->private;
6399 else if (new != conf
->skip_copy
) {
6400 mddev_suspend(mddev
);
6401 conf
->skip_copy
= new;
6403 mddev
->queue
->backing_dev_info
->capabilities
|=
6404 BDI_CAP_STABLE_WRITES
;
6406 mddev
->queue
->backing_dev_info
->capabilities
&=
6407 ~BDI_CAP_STABLE_WRITES
;
6408 mddev_resume(mddev
);
6410 mddev_unlock(mddev
);
6414 static struct md_sysfs_entry
6415 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6416 raid5_show_skip_copy
,
6417 raid5_store_skip_copy
);
6420 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6422 struct r5conf
*conf
= mddev
->private;
6424 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6429 static struct md_sysfs_entry
6430 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6433 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6435 struct r5conf
*conf
;
6437 spin_lock(&mddev
->lock
);
6438 conf
= mddev
->private;
6440 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6441 spin_unlock(&mddev
->lock
);
6445 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6447 int *worker_cnt_per_group
,
6448 struct r5worker_group
**worker_groups
);
6450 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6452 struct r5conf
*conf
;
6455 struct r5worker_group
*new_groups
, *old_groups
;
6456 int group_cnt
, worker_cnt_per_group
;
6458 if (len
>= PAGE_SIZE
)
6460 if (kstrtoul(page
, 10, &new))
6463 err
= mddev_lock(mddev
);
6466 conf
= mddev
->private;
6469 else if (new != conf
->worker_cnt_per_group
) {
6470 mddev_suspend(mddev
);
6472 old_groups
= conf
->worker_groups
;
6474 flush_workqueue(raid5_wq
);
6476 err
= alloc_thread_groups(conf
, new,
6477 &group_cnt
, &worker_cnt_per_group
,
6480 spin_lock_irq(&conf
->device_lock
);
6481 conf
->group_cnt
= group_cnt
;
6482 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6483 conf
->worker_groups
= new_groups
;
6484 spin_unlock_irq(&conf
->device_lock
);
6487 kfree(old_groups
[0].workers
);
6490 mddev_resume(mddev
);
6492 mddev_unlock(mddev
);
6497 static struct md_sysfs_entry
6498 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6499 raid5_show_group_thread_cnt
,
6500 raid5_store_group_thread_cnt
);
6502 static struct attribute
*raid5_attrs
[] = {
6503 &raid5_stripecache_size
.attr
,
6504 &raid5_stripecache_active
.attr
,
6505 &raid5_preread_bypass_threshold
.attr
,
6506 &raid5_group_thread_cnt
.attr
,
6507 &raid5_skip_copy
.attr
,
6508 &raid5_rmw_level
.attr
,
6509 &r5c_journal_mode
.attr
,
6512 static struct attribute_group raid5_attrs_group
= {
6514 .attrs
= raid5_attrs
,
6517 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6519 int *worker_cnt_per_group
,
6520 struct r5worker_group
**worker_groups
)
6524 struct r5worker
*workers
;
6526 *worker_cnt_per_group
= cnt
;
6529 *worker_groups
= NULL
;
6532 *group_cnt
= num_possible_nodes();
6533 size
= sizeof(struct r5worker
) * cnt
;
6534 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6535 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6536 *group_cnt
, GFP_NOIO
);
6537 if (!*worker_groups
|| !workers
) {
6539 kfree(*worker_groups
);
6543 for (i
= 0; i
< *group_cnt
; i
++) {
6544 struct r5worker_group
*group
;
6546 group
= &(*worker_groups
)[i
];
6547 INIT_LIST_HEAD(&group
->handle_list
);
6549 group
->workers
= workers
+ i
* cnt
;
6551 for (j
= 0; j
< cnt
; j
++) {
6552 struct r5worker
*worker
= group
->workers
+ j
;
6553 worker
->group
= group
;
6554 INIT_WORK(&worker
->work
, raid5_do_work
);
6556 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6557 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6564 static void free_thread_groups(struct r5conf
*conf
)
6566 if (conf
->worker_groups
)
6567 kfree(conf
->worker_groups
[0].workers
);
6568 kfree(conf
->worker_groups
);
6569 conf
->worker_groups
= NULL
;
6573 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6575 struct r5conf
*conf
= mddev
->private;
6578 sectors
= mddev
->dev_sectors
;
6580 /* size is defined by the smallest of previous and new size */
6581 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6583 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
6584 sectors
&= ~((sector_t
)conf
->prev_chunk_sectors
- 1);
6585 return sectors
* (raid_disks
- conf
->max_degraded
);
6588 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6590 safe_put_page(percpu
->spare_page
);
6591 if (percpu
->scribble
)
6592 flex_array_free(percpu
->scribble
);
6593 percpu
->spare_page
= NULL
;
6594 percpu
->scribble
= NULL
;
6597 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6599 if (conf
->level
== 6 && !percpu
->spare_page
)
6600 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6601 if (!percpu
->scribble
)
6602 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6603 conf
->previous_raid_disks
),
6604 max(conf
->chunk_sectors
,
6605 conf
->prev_chunk_sectors
)
6609 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6610 free_scratch_buffer(conf
, percpu
);
6617 static int raid456_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
6619 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6621 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6625 static void raid5_free_percpu(struct r5conf
*conf
)
6630 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6631 free_percpu(conf
->percpu
);
6634 static void free_conf(struct r5conf
*conf
)
6639 r5l_exit_log(conf
->log
);
6640 if (conf
->shrinker
.nr_deferred
)
6641 unregister_shrinker(&conf
->shrinker
);
6643 free_thread_groups(conf
);
6644 shrink_stripes(conf
);
6645 raid5_free_percpu(conf
);
6646 for (i
= 0; i
< conf
->pool_size
; i
++)
6647 if (conf
->disks
[i
].extra_page
)
6648 put_page(conf
->disks
[i
].extra_page
);
6650 kfree(conf
->stripe_hashtbl
);
6654 static int raid456_cpu_up_prepare(unsigned int cpu
, struct hlist_node
*node
)
6656 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6657 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6659 if (alloc_scratch_buffer(conf
, percpu
)) {
6660 pr_warn("%s: failed memory allocation for cpu%u\n",
6667 static int raid5_alloc_percpu(struct r5conf
*conf
)
6671 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6675 err
= cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6677 conf
->scribble_disks
= max(conf
->raid_disks
,
6678 conf
->previous_raid_disks
);
6679 conf
->scribble_sectors
= max(conf
->chunk_sectors
,
6680 conf
->prev_chunk_sectors
);
6685 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6686 struct shrink_control
*sc
)
6688 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6689 unsigned long ret
= SHRINK_STOP
;
6691 if (mutex_trylock(&conf
->cache_size_mutex
)) {
6693 while (ret
< sc
->nr_to_scan
&&
6694 conf
->max_nr_stripes
> conf
->min_nr_stripes
) {
6695 if (drop_one_stripe(conf
) == 0) {
6701 mutex_unlock(&conf
->cache_size_mutex
);
6706 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6707 struct shrink_control
*sc
)
6709 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6711 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6712 /* unlikely, but not impossible */
6714 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6717 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6719 struct r5conf
*conf
;
6720 int raid_disk
, memory
, max_disks
;
6721 struct md_rdev
*rdev
;
6722 struct disk_info
*disk
;
6725 int group_cnt
, worker_cnt_per_group
;
6726 struct r5worker_group
*new_group
;
6728 if (mddev
->new_level
!= 5
6729 && mddev
->new_level
!= 4
6730 && mddev
->new_level
!= 6) {
6731 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6732 mdname(mddev
), mddev
->new_level
);
6733 return ERR_PTR(-EIO
);
6735 if ((mddev
->new_level
== 5
6736 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6737 (mddev
->new_level
== 6
6738 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6739 pr_warn("md/raid:%s: layout %d not supported\n",
6740 mdname(mddev
), mddev
->new_layout
);
6741 return ERR_PTR(-EIO
);
6743 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6744 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6745 mdname(mddev
), mddev
->raid_disks
);
6746 return ERR_PTR(-EINVAL
);
6749 if (!mddev
->new_chunk_sectors
||
6750 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6751 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6752 pr_warn("md/raid:%s: invalid chunk size %d\n",
6753 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6754 return ERR_PTR(-EINVAL
);
6757 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6760 /* Don't enable multi-threading by default*/
6761 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6763 conf
->group_cnt
= group_cnt
;
6764 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6765 conf
->worker_groups
= new_group
;
6768 spin_lock_init(&conf
->device_lock
);
6769 seqcount_init(&conf
->gen_lock
);
6770 mutex_init(&conf
->cache_size_mutex
);
6771 init_waitqueue_head(&conf
->wait_for_quiescent
);
6772 init_waitqueue_head(&conf
->wait_for_stripe
);
6773 init_waitqueue_head(&conf
->wait_for_overlap
);
6774 INIT_LIST_HEAD(&conf
->handle_list
);
6775 INIT_LIST_HEAD(&conf
->hold_list
);
6776 INIT_LIST_HEAD(&conf
->delayed_list
);
6777 INIT_LIST_HEAD(&conf
->bitmap_list
);
6778 bio_list_init(&conf
->return_bi
);
6779 init_llist_head(&conf
->released_stripes
);
6780 atomic_set(&conf
->active_stripes
, 0);
6781 atomic_set(&conf
->preread_active_stripes
, 0);
6782 atomic_set(&conf
->active_aligned_reads
, 0);
6783 bio_list_init(&conf
->pending_bios
);
6784 spin_lock_init(&conf
->pending_bios_lock
);
6785 conf
->batch_bio_dispatch
= true;
6786 rdev_for_each(rdev
, mddev
) {
6787 if (test_bit(Journal
, &rdev
->flags
))
6789 if (blk_queue_nonrot(bdev_get_queue(rdev
->bdev
))) {
6790 conf
->batch_bio_dispatch
= false;
6795 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6796 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6798 conf
->raid_disks
= mddev
->raid_disks
;
6799 if (mddev
->reshape_position
== MaxSector
)
6800 conf
->previous_raid_disks
= mddev
->raid_disks
;
6802 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6803 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6805 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6811 for (i
= 0; i
< max_disks
; i
++) {
6812 conf
->disks
[i
].extra_page
= alloc_page(GFP_KERNEL
);
6813 if (!conf
->disks
[i
].extra_page
)
6817 conf
->mddev
= mddev
;
6819 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6822 /* We init hash_locks[0] separately to that it can be used
6823 * as the reference lock in the spin_lock_nest_lock() call
6824 * in lock_all_device_hash_locks_irq in order to convince
6825 * lockdep that we know what we are doing.
6827 spin_lock_init(conf
->hash_locks
);
6828 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6829 spin_lock_init(conf
->hash_locks
+ i
);
6831 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6832 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6834 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6835 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6837 atomic_set(&conf
->r5c_cached_full_stripes
, 0);
6838 INIT_LIST_HEAD(&conf
->r5c_full_stripe_list
);
6839 atomic_set(&conf
->r5c_cached_partial_stripes
, 0);
6840 INIT_LIST_HEAD(&conf
->r5c_partial_stripe_list
);
6841 atomic_set(&conf
->r5c_flushing_full_stripes
, 0);
6842 atomic_set(&conf
->r5c_flushing_partial_stripes
, 0);
6844 conf
->level
= mddev
->new_level
;
6845 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6846 if (raid5_alloc_percpu(conf
) != 0)
6849 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6851 rdev_for_each(rdev
, mddev
) {
6852 raid_disk
= rdev
->raid_disk
;
6853 if (raid_disk
>= max_disks
6854 || raid_disk
< 0 || test_bit(Journal
, &rdev
->flags
))
6856 disk
= conf
->disks
+ raid_disk
;
6858 if (test_bit(Replacement
, &rdev
->flags
)) {
6859 if (disk
->replacement
)
6861 disk
->replacement
= rdev
;
6868 if (test_bit(In_sync
, &rdev
->flags
)) {
6869 char b
[BDEVNAME_SIZE
];
6870 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
6871 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
6872 } else if (rdev
->saved_raid_disk
!= raid_disk
)
6873 /* Cannot rely on bitmap to complete recovery */
6877 conf
->level
= mddev
->new_level
;
6878 if (conf
->level
== 6) {
6879 conf
->max_degraded
= 2;
6880 if (raid6_call
.xor_syndrome
)
6881 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6883 conf
->rmw_level
= PARITY_DISABLE_RMW
;
6885 conf
->max_degraded
= 1;
6886 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6888 conf
->algorithm
= mddev
->new_layout
;
6889 conf
->reshape_progress
= mddev
->reshape_position
;
6890 if (conf
->reshape_progress
!= MaxSector
) {
6891 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
6892 conf
->prev_algo
= mddev
->layout
;
6894 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
6895 conf
->prev_algo
= conf
->algorithm
;
6898 conf
->min_nr_stripes
= NR_STRIPES
;
6899 if (mddev
->reshape_position
!= MaxSector
) {
6900 int stripes
= max_t(int,
6901 ((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4,
6902 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4);
6903 conf
->min_nr_stripes
= max(NR_STRIPES
, stripes
);
6904 if (conf
->min_nr_stripes
!= NR_STRIPES
)
6905 pr_info("md/raid:%s: force stripe size %d for reshape\n",
6906 mdname(mddev
), conf
->min_nr_stripes
);
6908 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
6909 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
6910 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
6911 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
6912 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
6913 mdname(mddev
), memory
);
6916 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev
), memory
);
6918 * Losing a stripe head costs more than the time to refill it,
6919 * it reduces the queue depth and so can hurt throughput.
6920 * So set it rather large, scaled by number of devices.
6922 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
6923 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
6924 conf
->shrinker
.count_objects
= raid5_cache_count
;
6925 conf
->shrinker
.batch
= 128;
6926 conf
->shrinker
.flags
= 0;
6927 if (register_shrinker(&conf
->shrinker
)) {
6928 pr_warn("md/raid:%s: couldn't register shrinker.\n",
6933 sprintf(pers_name
, "raid%d", mddev
->new_level
);
6934 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
6935 if (!conf
->thread
) {
6936 pr_warn("md/raid:%s: couldn't allocate thread.\n",
6946 return ERR_PTR(-EIO
);
6948 return ERR_PTR(-ENOMEM
);
6951 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
6954 case ALGORITHM_PARITY_0
:
6955 if (raid_disk
< max_degraded
)
6958 case ALGORITHM_PARITY_N
:
6959 if (raid_disk
>= raid_disks
- max_degraded
)
6962 case ALGORITHM_PARITY_0_6
:
6963 if (raid_disk
== 0 ||
6964 raid_disk
== raid_disks
- 1)
6967 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6968 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6969 case ALGORITHM_LEFT_SYMMETRIC_6
:
6970 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6971 if (raid_disk
== raid_disks
- 1)
6977 static int raid5_run(struct mddev
*mddev
)
6979 struct r5conf
*conf
;
6980 int working_disks
= 0;
6981 int dirty_parity_disks
= 0;
6982 struct md_rdev
*rdev
;
6983 struct md_rdev
*journal_dev
= NULL
;
6984 sector_t reshape_offset
= 0;
6986 long long min_offset_diff
= 0;
6989 if (mddev
->recovery_cp
!= MaxSector
)
6990 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
6993 rdev_for_each(rdev
, mddev
) {
6996 if (test_bit(Journal
, &rdev
->flags
)) {
7000 if (rdev
->raid_disk
< 0)
7002 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
7004 min_offset_diff
= diff
;
7006 } else if (mddev
->reshape_backwards
&&
7007 diff
< min_offset_diff
)
7008 min_offset_diff
= diff
;
7009 else if (!mddev
->reshape_backwards
&&
7010 diff
> min_offset_diff
)
7011 min_offset_diff
= diff
;
7014 if (mddev
->reshape_position
!= MaxSector
) {
7015 /* Check that we can continue the reshape.
7016 * Difficulties arise if the stripe we would write to
7017 * next is at or after the stripe we would read from next.
7018 * For a reshape that changes the number of devices, this
7019 * is only possible for a very short time, and mdadm makes
7020 * sure that time appears to have past before assembling
7021 * the array. So we fail if that time hasn't passed.
7022 * For a reshape that keeps the number of devices the same
7023 * mdadm must be monitoring the reshape can keeping the
7024 * critical areas read-only and backed up. It will start
7025 * the array in read-only mode, so we check for that.
7027 sector_t here_new
, here_old
;
7029 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
7034 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7039 if (mddev
->new_level
!= mddev
->level
) {
7040 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7044 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
7045 /* reshape_position must be on a new-stripe boundary, and one
7046 * further up in new geometry must map after here in old
7048 * If the chunk sizes are different, then as we perform reshape
7049 * in units of the largest of the two, reshape_position needs
7050 * be a multiple of the largest chunk size times new data disks.
7052 here_new
= mddev
->reshape_position
;
7053 chunk_sectors
= max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
);
7054 new_data_disks
= mddev
->raid_disks
- max_degraded
;
7055 if (sector_div(here_new
, chunk_sectors
* new_data_disks
)) {
7056 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7060 reshape_offset
= here_new
* chunk_sectors
;
7061 /* here_new is the stripe we will write to */
7062 here_old
= mddev
->reshape_position
;
7063 sector_div(here_old
, chunk_sectors
* (old_disks
-max_degraded
));
7064 /* here_old is the first stripe that we might need to read
7066 if (mddev
->delta_disks
== 0) {
7067 /* We cannot be sure it is safe to start an in-place
7068 * reshape. It is only safe if user-space is monitoring
7069 * and taking constant backups.
7070 * mdadm always starts a situation like this in
7071 * readonly mode so it can take control before
7072 * allowing any writes. So just check for that.
7074 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
7075 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
7076 /* not really in-place - so OK */;
7077 else if (mddev
->ro
== 0) {
7078 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7082 } else if (mddev
->reshape_backwards
7083 ? (here_new
* chunk_sectors
+ min_offset_diff
<=
7084 here_old
* chunk_sectors
)
7085 : (here_new
* chunk_sectors
>=
7086 here_old
* chunk_sectors
+ (-min_offset_diff
))) {
7087 /* Reading from the same stripe as writing to - bad */
7088 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7092 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev
));
7093 /* OK, we should be able to continue; */
7095 BUG_ON(mddev
->level
!= mddev
->new_level
);
7096 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
7097 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
7098 BUG_ON(mddev
->delta_disks
!= 0);
7101 if (mddev
->private == NULL
)
7102 conf
= setup_conf(mddev
);
7104 conf
= mddev
->private;
7107 return PTR_ERR(conf
);
7109 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
)) {
7111 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7114 set_disk_ro(mddev
->gendisk
, 1);
7115 } else if (mddev
->recovery_cp
== MaxSector
)
7116 set_bit(MD_JOURNAL_CLEAN
, &mddev
->flags
);
7119 conf
->min_offset_diff
= min_offset_diff
;
7120 mddev
->thread
= conf
->thread
;
7121 conf
->thread
= NULL
;
7122 mddev
->private = conf
;
7124 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
7126 rdev
= conf
->disks
[i
].rdev
;
7127 if (!rdev
&& conf
->disks
[i
].replacement
) {
7128 /* The replacement is all we have yet */
7129 rdev
= conf
->disks
[i
].replacement
;
7130 conf
->disks
[i
].replacement
= NULL
;
7131 clear_bit(Replacement
, &rdev
->flags
);
7132 conf
->disks
[i
].rdev
= rdev
;
7136 if (conf
->disks
[i
].replacement
&&
7137 conf
->reshape_progress
!= MaxSector
) {
7138 /* replacements and reshape simply do not mix. */
7139 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7142 if (test_bit(In_sync
, &rdev
->flags
)) {
7146 /* This disc is not fully in-sync. However if it
7147 * just stored parity (beyond the recovery_offset),
7148 * when we don't need to be concerned about the
7149 * array being dirty.
7150 * When reshape goes 'backwards', we never have
7151 * partially completed devices, so we only need
7152 * to worry about reshape going forwards.
7154 /* Hack because v0.91 doesn't store recovery_offset properly. */
7155 if (mddev
->major_version
== 0 &&
7156 mddev
->minor_version
> 90)
7157 rdev
->recovery_offset
= reshape_offset
;
7159 if (rdev
->recovery_offset
< reshape_offset
) {
7160 /* We need to check old and new layout */
7161 if (!only_parity(rdev
->raid_disk
,
7164 conf
->max_degraded
))
7167 if (!only_parity(rdev
->raid_disk
,
7169 conf
->previous_raid_disks
,
7170 conf
->max_degraded
))
7172 dirty_parity_disks
++;
7176 * 0 for a fully functional array, 1 or 2 for a degraded array.
7178 mddev
->degraded
= raid5_calc_degraded(conf
);
7180 if (has_failed(conf
)) {
7181 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7182 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
7186 /* device size must be a multiple of chunk size */
7187 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
7188 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
7190 if (mddev
->degraded
> dirty_parity_disks
&&
7191 mddev
->recovery_cp
!= MaxSector
) {
7192 if (mddev
->ok_start_degraded
)
7193 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7196 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7202 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7203 mdname(mddev
), conf
->level
,
7204 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
7207 print_raid5_conf(conf
);
7209 if (conf
->reshape_progress
!= MaxSector
) {
7210 conf
->reshape_safe
= conf
->reshape_progress
;
7211 atomic_set(&conf
->reshape_stripes
, 0);
7212 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7213 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7214 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7215 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7216 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7220 /* Ok, everything is just fine now */
7221 if (mddev
->to_remove
== &raid5_attrs_group
)
7222 mddev
->to_remove
= NULL
;
7223 else if (mddev
->kobj
.sd
&&
7224 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
7225 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7227 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7231 bool discard_supported
= true;
7232 /* read-ahead size must cover two whole stripes, which
7233 * is 2 * (datadisks) * chunksize where 'n' is the
7234 * number of raid devices
7236 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
7237 int stripe
= data_disks
*
7238 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
7239 if (mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7240 mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
7242 chunk_size
= mddev
->chunk_sectors
<< 9;
7243 blk_queue_io_min(mddev
->queue
, chunk_size
);
7244 blk_queue_io_opt(mddev
->queue
, chunk_size
*
7245 (conf
->raid_disks
- conf
->max_degraded
));
7246 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
7248 * We can only discard a whole stripe. It doesn't make sense to
7249 * discard data disk but write parity disk
7251 stripe
= stripe
* PAGE_SIZE
;
7252 /* Round up to power of 2, as discard handling
7253 * currently assumes that */
7254 while ((stripe
-1) & stripe
)
7255 stripe
= (stripe
| (stripe
-1)) + 1;
7256 mddev
->queue
->limits
.discard_alignment
= stripe
;
7257 mddev
->queue
->limits
.discard_granularity
= stripe
;
7260 * We use 16-bit counter of active stripes in bi_phys_segments
7261 * (minus one for over-loaded initialization)
7263 blk_queue_max_hw_sectors(mddev
->queue
, 0xfffe * STRIPE_SECTORS
);
7264 blk_queue_max_discard_sectors(mddev
->queue
,
7265 0xfffe * STRIPE_SECTORS
);
7268 * unaligned part of discard request will be ignored, so can't
7269 * guarantee discard_zeroes_data
7271 mddev
->queue
->limits
.discard_zeroes_data
= 0;
7273 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
7275 rdev_for_each(rdev
, mddev
) {
7276 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7277 rdev
->data_offset
<< 9);
7278 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7279 rdev
->new_data_offset
<< 9);
7281 * discard_zeroes_data is required, otherwise data
7282 * could be lost. Consider a scenario: discard a stripe
7283 * (the stripe could be inconsistent if
7284 * discard_zeroes_data is 0); write one disk of the
7285 * stripe (the stripe could be inconsistent again
7286 * depending on which disks are used to calculate
7287 * parity); the disk is broken; The stripe data of this
7290 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
7291 !bdev_get_queue(rdev
->bdev
)->
7292 limits
.discard_zeroes_data
)
7293 discard_supported
= false;
7294 /* Unfortunately, discard_zeroes_data is not currently
7295 * a guarantee - just a hint. So we only allow DISCARD
7296 * if the sysadmin has confirmed that only safe devices
7297 * are in use by setting a module parameter.
7299 if (!devices_handle_discard_safely
) {
7300 if (discard_supported
) {
7301 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7302 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7304 discard_supported
= false;
7308 if (discard_supported
&&
7309 mddev
->queue
->limits
.max_discard_sectors
>= (stripe
>> 9) &&
7310 mddev
->queue
->limits
.discard_granularity
>= stripe
)
7311 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
7314 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
7317 blk_queue_max_hw_sectors(mddev
->queue
, UINT_MAX
);
7321 char b
[BDEVNAME_SIZE
];
7323 pr_debug("md/raid:%s: using device %s as journal\n",
7324 mdname(mddev
), bdevname(journal_dev
->bdev
, b
));
7325 if (r5l_init_log(conf
, journal_dev
))
7331 md_unregister_thread(&mddev
->thread
);
7332 print_raid5_conf(conf
);
7334 mddev
->private = NULL
;
7335 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev
));
7339 static void raid5_free(struct mddev
*mddev
, void *priv
)
7341 struct r5conf
*conf
= priv
;
7344 mddev
->to_remove
= &raid5_attrs_group
;
7347 static void raid5_status(struct seq_file
*seq
, struct mddev
*mddev
)
7349 struct r5conf
*conf
= mddev
->private;
7352 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
7353 conf
->chunk_sectors
/ 2, mddev
->layout
);
7354 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
7356 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7357 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
7358 seq_printf (seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
7361 seq_printf (seq
, "]");
7364 static void print_raid5_conf (struct r5conf
*conf
)
7367 struct disk_info
*tmp
;
7369 pr_debug("RAID conf printout:\n");
7371 pr_debug("(conf==NULL)\n");
7374 pr_debug(" --- level:%d rd:%d wd:%d\n", conf
->level
,
7376 conf
->raid_disks
- conf
->mddev
->degraded
);
7378 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7379 char b
[BDEVNAME_SIZE
];
7380 tmp
= conf
->disks
+ i
;
7382 pr_debug(" disk %d, o:%d, dev:%s\n",
7383 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
7384 bdevname(tmp
->rdev
->bdev
, b
));
7388 static int raid5_spare_active(struct mddev
*mddev
)
7391 struct r5conf
*conf
= mddev
->private;
7392 struct disk_info
*tmp
;
7394 unsigned long flags
;
7396 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7397 tmp
= conf
->disks
+ i
;
7398 if (tmp
->replacement
7399 && tmp
->replacement
->recovery_offset
== MaxSector
7400 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
7401 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
7402 /* Replacement has just become active. */
7404 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
7407 /* Replaced device not technically faulty,
7408 * but we need to be sure it gets removed
7409 * and never re-added.
7411 set_bit(Faulty
, &tmp
->rdev
->flags
);
7412 sysfs_notify_dirent_safe(
7413 tmp
->rdev
->sysfs_state
);
7415 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
7416 } else if (tmp
->rdev
7417 && tmp
->rdev
->recovery_offset
== MaxSector
7418 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
7419 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
7421 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
7424 spin_lock_irqsave(&conf
->device_lock
, flags
);
7425 mddev
->degraded
= raid5_calc_degraded(conf
);
7426 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7427 print_raid5_conf(conf
);
7431 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7433 struct r5conf
*conf
= mddev
->private;
7435 int number
= rdev
->raid_disk
;
7436 struct md_rdev
**rdevp
;
7437 struct disk_info
*p
= conf
->disks
+ number
;
7439 print_raid5_conf(conf
);
7440 if (test_bit(Journal
, &rdev
->flags
) && conf
->log
) {
7441 struct r5l_log
*log
;
7443 * we can't wait pending write here, as this is called in
7444 * raid5d, wait will deadlock.
7446 if (atomic_read(&mddev
->writes_pending
))
7454 if (rdev
== p
->rdev
)
7456 else if (rdev
== p
->replacement
)
7457 rdevp
= &p
->replacement
;
7461 if (number
>= conf
->raid_disks
&&
7462 conf
->reshape_progress
== MaxSector
)
7463 clear_bit(In_sync
, &rdev
->flags
);
7465 if (test_bit(In_sync
, &rdev
->flags
) ||
7466 atomic_read(&rdev
->nr_pending
)) {
7470 /* Only remove non-faulty devices if recovery
7473 if (!test_bit(Faulty
, &rdev
->flags
) &&
7474 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7475 !has_failed(conf
) &&
7476 (!p
->replacement
|| p
->replacement
== rdev
) &&
7477 number
< conf
->raid_disks
) {
7482 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
7484 if (atomic_read(&rdev
->nr_pending
)) {
7485 /* lost the race, try later */
7490 if (p
->replacement
) {
7491 /* We must have just cleared 'rdev' */
7492 p
->rdev
= p
->replacement
;
7493 clear_bit(Replacement
, &p
->replacement
->flags
);
7494 smp_mb(); /* Make sure other CPUs may see both as identical
7495 * but will never see neither - if they are careful
7497 p
->replacement
= NULL
;
7498 clear_bit(WantReplacement
, &rdev
->flags
);
7500 /* We might have just removed the Replacement as faulty-
7501 * clear the bit just in case
7503 clear_bit(WantReplacement
, &rdev
->flags
);
7506 print_raid5_conf(conf
);
7510 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7512 struct r5conf
*conf
= mddev
->private;
7515 struct disk_info
*p
;
7517 int last
= conf
->raid_disks
- 1;
7519 if (test_bit(Journal
, &rdev
->flags
)) {
7520 char b
[BDEVNAME_SIZE
];
7524 rdev
->raid_disk
= 0;
7526 * The array is in readonly mode if journal is missing, so no
7527 * write requests running. We should be safe
7529 r5l_init_log(conf
, rdev
);
7530 pr_debug("md/raid:%s: using device %s as journal\n",
7531 mdname(mddev
), bdevname(rdev
->bdev
, b
));
7534 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7537 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7538 /* no point adding a device */
7541 if (rdev
->raid_disk
>= 0)
7542 first
= last
= rdev
->raid_disk
;
7545 * find the disk ... but prefer rdev->saved_raid_disk
7548 if (rdev
->saved_raid_disk
>= 0 &&
7549 rdev
->saved_raid_disk
>= first
&&
7550 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7551 first
= rdev
->saved_raid_disk
;
7553 for (disk
= first
; disk
<= last
; disk
++) {
7554 p
= conf
->disks
+ disk
;
7555 if (p
->rdev
== NULL
) {
7556 clear_bit(In_sync
, &rdev
->flags
);
7557 rdev
->raid_disk
= disk
;
7559 if (rdev
->saved_raid_disk
!= disk
)
7561 rcu_assign_pointer(p
->rdev
, rdev
);
7565 for (disk
= first
; disk
<= last
; disk
++) {
7566 p
= conf
->disks
+ disk
;
7567 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7568 p
->replacement
== NULL
) {
7569 clear_bit(In_sync
, &rdev
->flags
);
7570 set_bit(Replacement
, &rdev
->flags
);
7571 rdev
->raid_disk
= disk
;
7574 rcu_assign_pointer(p
->replacement
, rdev
);
7579 print_raid5_conf(conf
);
7583 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7585 /* no resync is happening, and there is enough space
7586 * on all devices, so we can resize.
7587 * We need to make sure resync covers any new space.
7588 * If the array is shrinking we should possibly wait until
7589 * any io in the removed space completes, but it hardly seems
7593 struct r5conf
*conf
= mddev
->private;
7597 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
7598 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7599 if (mddev
->external_size
&&
7600 mddev
->array_sectors
> newsize
)
7602 if (mddev
->bitmap
) {
7603 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7607 md_set_array_sectors(mddev
, newsize
);
7608 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7609 revalidate_disk(mddev
->gendisk
);
7610 if (sectors
> mddev
->dev_sectors
&&
7611 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7612 mddev
->recovery_cp
= mddev
->dev_sectors
;
7613 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7615 mddev
->dev_sectors
= sectors
;
7616 mddev
->resync_max_sectors
= sectors
;
7620 static int check_stripe_cache(struct mddev
*mddev
)
7622 /* Can only proceed if there are plenty of stripe_heads.
7623 * We need a minimum of one full stripe,, and for sensible progress
7624 * it is best to have about 4 times that.
7625 * If we require 4 times, then the default 256 4K stripe_heads will
7626 * allow for chunk sizes up to 256K, which is probably OK.
7627 * If the chunk size is greater, user-space should request more
7628 * stripe_heads first.
7630 struct r5conf
*conf
= mddev
->private;
7631 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7632 > conf
->min_nr_stripes
||
7633 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7634 > conf
->min_nr_stripes
) {
7635 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7637 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7644 static int check_reshape(struct mddev
*mddev
)
7646 struct r5conf
*conf
= mddev
->private;
7650 if (mddev
->delta_disks
== 0 &&
7651 mddev
->new_layout
== mddev
->layout
&&
7652 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7653 return 0; /* nothing to do */
7654 if (has_failed(conf
))
7656 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7657 /* We might be able to shrink, but the devices must
7658 * be made bigger first.
7659 * For raid6, 4 is the minimum size.
7660 * Otherwise 2 is the minimum
7663 if (mddev
->level
== 6)
7665 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7669 if (!check_stripe_cache(mddev
))
7672 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
7673 mddev
->delta_disks
> 0)
7674 if (resize_chunks(conf
,
7675 conf
->previous_raid_disks
7676 + max(0, mddev
->delta_disks
),
7677 max(mddev
->new_chunk_sectors
,
7678 mddev
->chunk_sectors
)
7681 return resize_stripes(conf
, (conf
->previous_raid_disks
7682 + mddev
->delta_disks
));
7685 static int raid5_start_reshape(struct mddev
*mddev
)
7687 struct r5conf
*conf
= mddev
->private;
7688 struct md_rdev
*rdev
;
7690 unsigned long flags
;
7692 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7695 if (!check_stripe_cache(mddev
))
7698 if (has_failed(conf
))
7701 rdev_for_each(rdev
, mddev
) {
7702 if (!test_bit(In_sync
, &rdev
->flags
)
7703 && !test_bit(Faulty
, &rdev
->flags
))
7707 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7708 /* Not enough devices even to make a degraded array
7713 /* Refuse to reduce size of the array. Any reductions in
7714 * array size must be through explicit setting of array_size
7717 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7718 < mddev
->array_sectors
) {
7719 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7724 atomic_set(&conf
->reshape_stripes
, 0);
7725 spin_lock_irq(&conf
->device_lock
);
7726 write_seqcount_begin(&conf
->gen_lock
);
7727 conf
->previous_raid_disks
= conf
->raid_disks
;
7728 conf
->raid_disks
+= mddev
->delta_disks
;
7729 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7730 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7731 conf
->prev_algo
= conf
->algorithm
;
7732 conf
->algorithm
= mddev
->new_layout
;
7734 /* Code that selects data_offset needs to see the generation update
7735 * if reshape_progress has been set - so a memory barrier needed.
7738 if (mddev
->reshape_backwards
)
7739 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7741 conf
->reshape_progress
= 0;
7742 conf
->reshape_safe
= conf
->reshape_progress
;
7743 write_seqcount_end(&conf
->gen_lock
);
7744 spin_unlock_irq(&conf
->device_lock
);
7746 /* Now make sure any requests that proceeded on the assumption
7747 * the reshape wasn't running - like Discard or Read - have
7750 mddev_suspend(mddev
);
7751 mddev_resume(mddev
);
7753 /* Add some new drives, as many as will fit.
7754 * We know there are enough to make the newly sized array work.
7755 * Don't add devices if we are reducing the number of
7756 * devices in the array. This is because it is not possible
7757 * to correctly record the "partially reconstructed" state of
7758 * such devices during the reshape and confusion could result.
7760 if (mddev
->delta_disks
>= 0) {
7761 rdev_for_each(rdev
, mddev
)
7762 if (rdev
->raid_disk
< 0 &&
7763 !test_bit(Faulty
, &rdev
->flags
)) {
7764 if (raid5_add_disk(mddev
, rdev
) == 0) {
7766 >= conf
->previous_raid_disks
)
7767 set_bit(In_sync
, &rdev
->flags
);
7769 rdev
->recovery_offset
= 0;
7771 if (sysfs_link_rdev(mddev
, rdev
))
7772 /* Failure here is OK */;
7774 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7775 && !test_bit(Faulty
, &rdev
->flags
)) {
7776 /* This is a spare that was manually added */
7777 set_bit(In_sync
, &rdev
->flags
);
7780 /* When a reshape changes the number of devices,
7781 * ->degraded is measured against the larger of the
7782 * pre and post number of devices.
7784 spin_lock_irqsave(&conf
->device_lock
, flags
);
7785 mddev
->degraded
= raid5_calc_degraded(conf
);
7786 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7788 mddev
->raid_disks
= conf
->raid_disks
;
7789 mddev
->reshape_position
= conf
->reshape_progress
;
7790 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
7792 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7793 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7794 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
7795 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7796 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7797 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7799 if (!mddev
->sync_thread
) {
7800 mddev
->recovery
= 0;
7801 spin_lock_irq(&conf
->device_lock
);
7802 write_seqcount_begin(&conf
->gen_lock
);
7803 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7804 mddev
->new_chunk_sectors
=
7805 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7806 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7807 rdev_for_each(rdev
, mddev
)
7808 rdev
->new_data_offset
= rdev
->data_offset
;
7810 conf
->generation
--;
7811 conf
->reshape_progress
= MaxSector
;
7812 mddev
->reshape_position
= MaxSector
;
7813 write_seqcount_end(&conf
->gen_lock
);
7814 spin_unlock_irq(&conf
->device_lock
);
7817 conf
->reshape_checkpoint
= jiffies
;
7818 md_wakeup_thread(mddev
->sync_thread
);
7819 md_new_event(mddev
);
7823 /* This is called from the reshape thread and should make any
7824 * changes needed in 'conf'
7826 static void end_reshape(struct r5conf
*conf
)
7829 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7830 struct md_rdev
*rdev
;
7832 spin_lock_irq(&conf
->device_lock
);
7833 conf
->previous_raid_disks
= conf
->raid_disks
;
7834 rdev_for_each(rdev
, conf
->mddev
)
7835 rdev
->data_offset
= rdev
->new_data_offset
;
7837 conf
->reshape_progress
= MaxSector
;
7838 conf
->mddev
->reshape_position
= MaxSector
;
7839 spin_unlock_irq(&conf
->device_lock
);
7840 wake_up(&conf
->wait_for_overlap
);
7842 /* read-ahead size must cover two whole stripes, which is
7843 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7845 if (conf
->mddev
->queue
) {
7846 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7847 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7849 if (conf
->mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7850 conf
->mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
7855 /* This is called from the raid5d thread with mddev_lock held.
7856 * It makes config changes to the device.
7858 static void raid5_finish_reshape(struct mddev
*mddev
)
7860 struct r5conf
*conf
= mddev
->private;
7862 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7864 if (mddev
->delta_disks
> 0) {
7865 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7867 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7868 revalidate_disk(mddev
->gendisk
);
7872 spin_lock_irq(&conf
->device_lock
);
7873 mddev
->degraded
= raid5_calc_degraded(conf
);
7874 spin_unlock_irq(&conf
->device_lock
);
7875 for (d
= conf
->raid_disks
;
7876 d
< conf
->raid_disks
- mddev
->delta_disks
;
7878 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
7880 clear_bit(In_sync
, &rdev
->flags
);
7881 rdev
= conf
->disks
[d
].replacement
;
7883 clear_bit(In_sync
, &rdev
->flags
);
7886 mddev
->layout
= conf
->algorithm
;
7887 mddev
->chunk_sectors
= conf
->chunk_sectors
;
7888 mddev
->reshape_position
= MaxSector
;
7889 mddev
->delta_disks
= 0;
7890 mddev
->reshape_backwards
= 0;
7894 static void raid5_quiesce(struct mddev
*mddev
, int state
)
7896 struct r5conf
*conf
= mddev
->private;
7899 case 2: /* resume for a suspend */
7900 wake_up(&conf
->wait_for_overlap
);
7903 case 1: /* stop all writes */
7904 lock_all_device_hash_locks_irq(conf
);
7905 /* '2' tells resync/reshape to pause so that all
7906 * active stripes can drain
7908 r5c_flush_cache(conf
, INT_MAX
);
7910 wait_event_cmd(conf
->wait_for_quiescent
,
7911 atomic_read(&conf
->active_stripes
) == 0 &&
7912 atomic_read(&conf
->active_aligned_reads
) == 0,
7913 unlock_all_device_hash_locks_irq(conf
),
7914 lock_all_device_hash_locks_irq(conf
));
7916 unlock_all_device_hash_locks_irq(conf
);
7917 /* allow reshape to continue */
7918 wake_up(&conf
->wait_for_overlap
);
7921 case 0: /* re-enable writes */
7922 lock_all_device_hash_locks_irq(conf
);
7924 wake_up(&conf
->wait_for_quiescent
);
7925 wake_up(&conf
->wait_for_overlap
);
7926 unlock_all_device_hash_locks_irq(conf
);
7929 r5l_quiesce(conf
->log
, state
);
7932 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
7934 struct r0conf
*raid0_conf
= mddev
->private;
7937 /* for raid0 takeover only one zone is supported */
7938 if (raid0_conf
->nr_strip_zones
> 1) {
7939 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7941 return ERR_PTR(-EINVAL
);
7944 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
7945 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
7946 mddev
->dev_sectors
= sectors
;
7947 mddev
->new_level
= level
;
7948 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7949 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
7950 mddev
->raid_disks
+= 1;
7951 mddev
->delta_disks
= 1;
7952 /* make sure it will be not marked as dirty */
7953 mddev
->recovery_cp
= MaxSector
;
7955 return setup_conf(mddev
);
7958 static void *raid5_takeover_raid1(struct mddev
*mddev
)
7963 if (mddev
->raid_disks
!= 2 ||
7964 mddev
->degraded
> 1)
7965 return ERR_PTR(-EINVAL
);
7967 /* Should check if there are write-behind devices? */
7969 chunksect
= 64*2; /* 64K by default */
7971 /* The array must be an exact multiple of chunksize */
7972 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
7975 if ((chunksect
<<9) < STRIPE_SIZE
)
7976 /* array size does not allow a suitable chunk size */
7977 return ERR_PTR(-EINVAL
);
7979 mddev
->new_level
= 5;
7980 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7981 mddev
->new_chunk_sectors
= chunksect
;
7983 ret
= setup_conf(mddev
);
7985 mddev_clear_unsupported_flags(mddev
,
7986 UNSUPPORTED_MDDEV_FLAGS
);
7990 static void *raid5_takeover_raid6(struct mddev
*mddev
)
7994 switch (mddev
->layout
) {
7995 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7996 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
7998 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7999 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
8001 case ALGORITHM_LEFT_SYMMETRIC_6
:
8002 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8004 case ALGORITHM_RIGHT_SYMMETRIC_6
:
8005 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
8007 case ALGORITHM_PARITY_0_6
:
8008 new_layout
= ALGORITHM_PARITY_0
;
8010 case ALGORITHM_PARITY_N
:
8011 new_layout
= ALGORITHM_PARITY_N
;
8014 return ERR_PTR(-EINVAL
);
8016 mddev
->new_level
= 5;
8017 mddev
->new_layout
= new_layout
;
8018 mddev
->delta_disks
= -1;
8019 mddev
->raid_disks
-= 1;
8020 return setup_conf(mddev
);
8023 static int raid5_check_reshape(struct mddev
*mddev
)
8025 /* For a 2-drive array, the layout and chunk size can be changed
8026 * immediately as not restriping is needed.
8027 * For larger arrays we record the new value - after validation
8028 * to be used by a reshape pass.
8030 struct r5conf
*conf
= mddev
->private;
8031 int new_chunk
= mddev
->new_chunk_sectors
;
8033 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
8035 if (new_chunk
> 0) {
8036 if (!is_power_of_2(new_chunk
))
8038 if (new_chunk
< (PAGE_SIZE
>>9))
8040 if (mddev
->array_sectors
& (new_chunk
-1))
8041 /* not factor of array size */
8045 /* They look valid */
8047 if (mddev
->raid_disks
== 2) {
8048 /* can make the change immediately */
8049 if (mddev
->new_layout
>= 0) {
8050 conf
->algorithm
= mddev
->new_layout
;
8051 mddev
->layout
= mddev
->new_layout
;
8053 if (new_chunk
> 0) {
8054 conf
->chunk_sectors
= new_chunk
;
8055 mddev
->chunk_sectors
= new_chunk
;
8057 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
8058 md_wakeup_thread(mddev
->thread
);
8060 return check_reshape(mddev
);
8063 static int raid6_check_reshape(struct mddev
*mddev
)
8065 int new_chunk
= mddev
->new_chunk_sectors
;
8067 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
8069 if (new_chunk
> 0) {
8070 if (!is_power_of_2(new_chunk
))
8072 if (new_chunk
< (PAGE_SIZE
>> 9))
8074 if (mddev
->array_sectors
& (new_chunk
-1))
8075 /* not factor of array size */
8079 /* They look valid */
8080 return check_reshape(mddev
);
8083 static void *raid5_takeover(struct mddev
*mddev
)
8085 /* raid5 can take over:
8086 * raid0 - if there is only one strip zone - make it a raid4 layout
8087 * raid1 - if there are two drives. We need to know the chunk size
8088 * raid4 - trivial - just use a raid4 layout.
8089 * raid6 - Providing it is a *_6 layout
8091 if (mddev
->level
== 0)
8092 return raid45_takeover_raid0(mddev
, 5);
8093 if (mddev
->level
== 1)
8094 return raid5_takeover_raid1(mddev
);
8095 if (mddev
->level
== 4) {
8096 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8097 mddev
->new_level
= 5;
8098 return setup_conf(mddev
);
8100 if (mddev
->level
== 6)
8101 return raid5_takeover_raid6(mddev
);
8103 return ERR_PTR(-EINVAL
);
8106 static void *raid4_takeover(struct mddev
*mddev
)
8108 /* raid4 can take over:
8109 * raid0 - if there is only one strip zone
8110 * raid5 - if layout is right
8112 if (mddev
->level
== 0)
8113 return raid45_takeover_raid0(mddev
, 4);
8114 if (mddev
->level
== 5 &&
8115 mddev
->layout
== ALGORITHM_PARITY_N
) {
8116 mddev
->new_layout
= 0;
8117 mddev
->new_level
= 4;
8118 return setup_conf(mddev
);
8120 return ERR_PTR(-EINVAL
);
8123 static struct md_personality raid5_personality
;
8125 static void *raid6_takeover(struct mddev
*mddev
)
8127 /* Currently can only take over a raid5. We map the
8128 * personality to an equivalent raid6 personality
8129 * with the Q block at the end.
8133 if (mddev
->pers
!= &raid5_personality
)
8134 return ERR_PTR(-EINVAL
);
8135 if (mddev
->degraded
> 1)
8136 return ERR_PTR(-EINVAL
);
8137 if (mddev
->raid_disks
> 253)
8138 return ERR_PTR(-EINVAL
);
8139 if (mddev
->raid_disks
< 3)
8140 return ERR_PTR(-EINVAL
);
8142 switch (mddev
->layout
) {
8143 case ALGORITHM_LEFT_ASYMMETRIC
:
8144 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
8146 case ALGORITHM_RIGHT_ASYMMETRIC
:
8147 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
8149 case ALGORITHM_LEFT_SYMMETRIC
:
8150 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
8152 case ALGORITHM_RIGHT_SYMMETRIC
:
8153 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
8155 case ALGORITHM_PARITY_0
:
8156 new_layout
= ALGORITHM_PARITY_0_6
;
8158 case ALGORITHM_PARITY_N
:
8159 new_layout
= ALGORITHM_PARITY_N
;
8162 return ERR_PTR(-EINVAL
);
8164 mddev
->new_level
= 6;
8165 mddev
->new_layout
= new_layout
;
8166 mddev
->delta_disks
= 1;
8167 mddev
->raid_disks
+= 1;
8168 return setup_conf(mddev
);
8171 static struct md_personality raid6_personality
=
8175 .owner
= THIS_MODULE
,
8176 .make_request
= raid5_make_request
,
8179 .status
= raid5_status
,
8180 .error_handler
= raid5_error
,
8181 .hot_add_disk
= raid5_add_disk
,
8182 .hot_remove_disk
= raid5_remove_disk
,
8183 .spare_active
= raid5_spare_active
,
8184 .sync_request
= raid5_sync_request
,
8185 .resize
= raid5_resize
,
8187 .check_reshape
= raid6_check_reshape
,
8188 .start_reshape
= raid5_start_reshape
,
8189 .finish_reshape
= raid5_finish_reshape
,
8190 .quiesce
= raid5_quiesce
,
8191 .takeover
= raid6_takeover
,
8192 .congested
= raid5_congested
,
8194 static struct md_personality raid5_personality
=
8198 .owner
= THIS_MODULE
,
8199 .make_request
= raid5_make_request
,
8202 .status
= raid5_status
,
8203 .error_handler
= raid5_error
,
8204 .hot_add_disk
= raid5_add_disk
,
8205 .hot_remove_disk
= raid5_remove_disk
,
8206 .spare_active
= raid5_spare_active
,
8207 .sync_request
= raid5_sync_request
,
8208 .resize
= raid5_resize
,
8210 .check_reshape
= raid5_check_reshape
,
8211 .start_reshape
= raid5_start_reshape
,
8212 .finish_reshape
= raid5_finish_reshape
,
8213 .quiesce
= raid5_quiesce
,
8214 .takeover
= raid5_takeover
,
8215 .congested
= raid5_congested
,
8218 static struct md_personality raid4_personality
=
8222 .owner
= THIS_MODULE
,
8223 .make_request
= raid5_make_request
,
8226 .status
= raid5_status
,
8227 .error_handler
= raid5_error
,
8228 .hot_add_disk
= raid5_add_disk
,
8229 .hot_remove_disk
= raid5_remove_disk
,
8230 .spare_active
= raid5_spare_active
,
8231 .sync_request
= raid5_sync_request
,
8232 .resize
= raid5_resize
,
8234 .check_reshape
= raid5_check_reshape
,
8235 .start_reshape
= raid5_start_reshape
,
8236 .finish_reshape
= raid5_finish_reshape
,
8237 .quiesce
= raid5_quiesce
,
8238 .takeover
= raid4_takeover
,
8239 .congested
= raid5_congested
,
8242 static int __init
raid5_init(void)
8246 raid5_wq
= alloc_workqueue("raid5wq",
8247 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
8251 ret
= cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE
,
8253 raid456_cpu_up_prepare
,
8256 destroy_workqueue(raid5_wq
);
8259 register_md_personality(&raid6_personality
);
8260 register_md_personality(&raid5_personality
);
8261 register_md_personality(&raid4_personality
);
8265 static void raid5_exit(void)
8267 unregister_md_personality(&raid6_personality
);
8268 unregister_md_personality(&raid5_personality
);
8269 unregister_md_personality(&raid4_personality
);
8270 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE
);
8271 destroy_workqueue(raid5_wq
);
8274 module_init(raid5_init
);
8275 module_exit(raid5_exit
);
8276 MODULE_LICENSE("GPL");
8277 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8278 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8279 MODULE_ALIAS("md-raid5");
8280 MODULE_ALIAS("md-raid4");
8281 MODULE_ALIAS("md-level-5");
8282 MODULE_ALIAS("md-level-4");
8283 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8284 MODULE_ALIAS("md-raid6");
8285 MODULE_ALIAS("md-level-6");
8287 /* This used to be two separate modules, they were: */
8288 MODULE_ALIAS("raid5");
8289 MODULE_ALIAS("raid6");