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 <trace/events/block.h>
65 #define cpu_to_group(cpu) cpu_to_node(cpu)
66 #define ANY_GROUP NUMA_NO_NODE
68 static bool devices_handle_discard_safely
= false;
69 module_param(devices_handle_discard_safely
, bool, 0644);
70 MODULE_PARM_DESC(devices_handle_discard_safely
,
71 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
72 static struct workqueue_struct
*raid5_wq
;
77 #define NR_STRIPES 256
78 #define STRIPE_SIZE PAGE_SIZE
79 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
80 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
81 #define IO_THRESHOLD 1
82 #define BYPASS_THRESHOLD 1
83 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
84 #define HASH_MASK (NR_HASH - 1)
85 #define MAX_STRIPE_BATCH 8
87 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
89 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
90 return &conf
->stripe_hashtbl
[hash
];
93 static inline int stripe_hash_locks_hash(sector_t sect
)
95 return (sect
>> STRIPE_SHIFT
) & STRIPE_HASH_LOCKS_MASK
;
98 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
100 spin_lock_irq(conf
->hash_locks
+ hash
);
101 spin_lock(&conf
->device_lock
);
104 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
106 spin_unlock(&conf
->device_lock
);
107 spin_unlock_irq(conf
->hash_locks
+ hash
);
110 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
114 spin_lock(conf
->hash_locks
);
115 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
116 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
117 spin_lock(&conf
->device_lock
);
120 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
123 spin_unlock(&conf
->device_lock
);
124 for (i
= NR_STRIPE_HASH_LOCKS
; i
; i
--)
125 spin_unlock(conf
->hash_locks
+ i
- 1);
129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
130 * order without overlap. There may be several bio's per stripe+device, and
131 * a bio could span several devices.
132 * When walking this list for a particular stripe+device, we must never proceed
133 * beyond a bio that extends past this device, as the next bio might no longer
135 * This function is used to determine the 'next' bio in the list, given the sector
136 * of the current stripe+device
138 static inline struct bio
*r5_next_bio(struct bio
*bio
, sector_t sector
)
140 int sectors
= bio_sectors(bio
);
141 if (bio
->bi_iter
.bi_sector
+ sectors
< sector
+ STRIPE_SECTORS
)
148 * We maintain a biased count of active stripes in the bottom 16 bits of
149 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
151 static inline int raid5_bi_processed_stripes(struct bio
*bio
)
153 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
154 return (atomic_read(segments
) >> 16) & 0xffff;
157 static inline int raid5_dec_bi_active_stripes(struct bio
*bio
)
159 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
160 return atomic_sub_return(1, segments
) & 0xffff;
163 static inline void raid5_inc_bi_active_stripes(struct bio
*bio
)
165 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
166 atomic_inc(segments
);
169 static inline void raid5_set_bi_processed_stripes(struct bio
*bio
,
172 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
176 old
= atomic_read(segments
);
177 new = (old
& 0xffff) | (cnt
<< 16);
178 } while (atomic_cmpxchg(segments
, old
, new) != old
);
181 static inline void raid5_set_bi_stripes(struct bio
*bio
, unsigned int cnt
)
183 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
184 atomic_set(segments
, cnt
);
187 /* Find first data disk in a raid6 stripe */
188 static inline int raid6_d0(struct stripe_head
*sh
)
191 /* ddf always start from first device */
193 /* md starts just after Q block */
194 if (sh
->qd_idx
== sh
->disks
- 1)
197 return sh
->qd_idx
+ 1;
199 static inline int raid6_next_disk(int disk
, int raid_disks
)
202 return (disk
< raid_disks
) ? disk
: 0;
205 /* When walking through the disks in a raid5, starting at raid6_d0,
206 * We need to map each disk to a 'slot', where the data disks are slot
207 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
208 * is raid_disks-1. This help does that mapping.
210 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
211 int *count
, int syndrome_disks
)
217 if (idx
== sh
->pd_idx
)
218 return syndrome_disks
;
219 if (idx
== sh
->qd_idx
)
220 return syndrome_disks
+ 1;
226 static void return_io(struct bio
*return_bi
)
228 struct bio
*bi
= return_bi
;
231 return_bi
= bi
->bi_next
;
233 bi
->bi_iter
.bi_size
= 0;
234 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
241 static void print_raid5_conf (struct r5conf
*conf
);
243 static int stripe_operations_active(struct stripe_head
*sh
)
245 return sh
->check_state
|| sh
->reconstruct_state
||
246 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
247 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
250 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
252 struct r5conf
*conf
= sh
->raid_conf
;
253 struct r5worker_group
*group
;
255 int i
, cpu
= sh
->cpu
;
257 if (!cpu_online(cpu
)) {
258 cpu
= cpumask_any(cpu_online_mask
);
262 if (list_empty(&sh
->lru
)) {
263 struct r5worker_group
*group
;
264 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
265 list_add_tail(&sh
->lru
, &group
->handle_list
);
266 group
->stripes_cnt
++;
270 if (conf
->worker_cnt_per_group
== 0) {
271 md_wakeup_thread(conf
->mddev
->thread
);
275 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
277 group
->workers
[0].working
= true;
278 /* at least one worker should run to avoid race */
279 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
281 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
282 /* wakeup more workers */
283 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
284 if (group
->workers
[i
].working
== false) {
285 group
->workers
[i
].working
= true;
286 queue_work_on(sh
->cpu
, raid5_wq
,
287 &group
->workers
[i
].work
);
293 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
294 struct list_head
*temp_inactive_list
)
296 BUG_ON(!list_empty(&sh
->lru
));
297 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
298 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
299 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
300 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
301 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
302 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
303 sh
->bm_seq
- conf
->seq_write
> 0)
304 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
306 clear_bit(STRIPE_DELAYED
, &sh
->state
);
307 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
308 if (conf
->worker_cnt_per_group
== 0) {
309 list_add_tail(&sh
->lru
, &conf
->handle_list
);
311 raid5_wakeup_stripe_thread(sh
);
315 md_wakeup_thread(conf
->mddev
->thread
);
317 BUG_ON(stripe_operations_active(sh
));
318 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
319 if (atomic_dec_return(&conf
->preread_active_stripes
)
321 md_wakeup_thread(conf
->mddev
->thread
);
322 atomic_dec(&conf
->active_stripes
);
323 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
))
324 list_add_tail(&sh
->lru
, temp_inactive_list
);
328 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
329 struct list_head
*temp_inactive_list
)
331 if (atomic_dec_and_test(&sh
->count
))
332 do_release_stripe(conf
, sh
, temp_inactive_list
);
336 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
338 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
339 * given time. Adding stripes only takes device lock, while deleting stripes
340 * only takes hash lock.
342 static void release_inactive_stripe_list(struct r5conf
*conf
,
343 struct list_head
*temp_inactive_list
,
347 bool do_wakeup
= false;
350 if (hash
== NR_STRIPE_HASH_LOCKS
) {
351 size
= NR_STRIPE_HASH_LOCKS
;
352 hash
= NR_STRIPE_HASH_LOCKS
- 1;
356 struct list_head
*list
= &temp_inactive_list
[size
- 1];
359 * We don't hold any lock here yet, get_active_stripe() might
360 * remove stripes from the list
362 if (!list_empty_careful(list
)) {
363 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
364 if (list_empty(conf
->inactive_list
+ hash
) &&
366 atomic_dec(&conf
->empty_inactive_list_nr
);
367 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
369 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
376 wake_up(&conf
->wait_for_stripe
);
377 if (conf
->retry_read_aligned
)
378 md_wakeup_thread(conf
->mddev
->thread
);
382 /* should hold conf->device_lock already */
383 static int release_stripe_list(struct r5conf
*conf
,
384 struct list_head
*temp_inactive_list
)
386 struct stripe_head
*sh
;
388 struct llist_node
*head
;
390 head
= llist_del_all(&conf
->released_stripes
);
391 head
= llist_reverse_order(head
);
395 sh
= llist_entry(head
, struct stripe_head
, release_list
);
396 head
= llist_next(head
);
397 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
399 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
401 * Don't worry the bit is set here, because if the bit is set
402 * again, the count is always > 1. This is true for
403 * STRIPE_ON_UNPLUG_LIST bit too.
405 hash
= sh
->hash_lock_index
;
406 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
413 static void release_stripe(struct stripe_head
*sh
)
415 struct r5conf
*conf
= sh
->raid_conf
;
417 struct list_head list
;
421 /* Avoid release_list until the last reference.
423 if (atomic_add_unless(&sh
->count
, -1, 1))
426 if (unlikely(!conf
->mddev
->thread
) ||
427 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
429 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
431 md_wakeup_thread(conf
->mddev
->thread
);
434 local_irq_save(flags
);
435 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
436 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
437 INIT_LIST_HEAD(&list
);
438 hash
= sh
->hash_lock_index
;
439 do_release_stripe(conf
, sh
, &list
);
440 spin_unlock(&conf
->device_lock
);
441 release_inactive_stripe_list(conf
, &list
, hash
);
443 local_irq_restore(flags
);
446 static inline void remove_hash(struct stripe_head
*sh
)
448 pr_debug("remove_hash(), stripe %llu\n",
449 (unsigned long long)sh
->sector
);
451 hlist_del_init(&sh
->hash
);
454 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
456 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
458 pr_debug("insert_hash(), stripe %llu\n",
459 (unsigned long long)sh
->sector
);
461 hlist_add_head(&sh
->hash
, hp
);
464 /* find an idle stripe, make sure it is unhashed, and return it. */
465 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
467 struct stripe_head
*sh
= NULL
;
468 struct list_head
*first
;
470 if (list_empty(conf
->inactive_list
+ hash
))
472 first
= (conf
->inactive_list
+ hash
)->next
;
473 sh
= list_entry(first
, struct stripe_head
, lru
);
474 list_del_init(first
);
476 atomic_inc(&conf
->active_stripes
);
477 BUG_ON(hash
!= sh
->hash_lock_index
);
478 if (list_empty(conf
->inactive_list
+ hash
))
479 atomic_inc(&conf
->empty_inactive_list_nr
);
484 static void shrink_buffers(struct stripe_head
*sh
)
488 int num
= sh
->raid_conf
->pool_size
;
490 for (i
= 0; i
< num
; i
++) {
491 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
495 sh
->dev
[i
].page
= NULL
;
500 static int grow_buffers(struct stripe_head
*sh
)
503 int num
= sh
->raid_conf
->pool_size
;
505 for (i
= 0; i
< num
; i
++) {
508 if (!(page
= alloc_page(GFP_KERNEL
))) {
511 sh
->dev
[i
].page
= page
;
512 sh
->dev
[i
].orig_page
= page
;
517 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
518 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
519 struct stripe_head
*sh
);
521 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
523 struct r5conf
*conf
= sh
->raid_conf
;
526 BUG_ON(atomic_read(&sh
->count
) != 0);
527 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
528 BUG_ON(stripe_operations_active(sh
));
529 BUG_ON(sh
->batch_head
);
531 pr_debug("init_stripe called, stripe %llu\n",
532 (unsigned long long)sector
);
534 seq
= read_seqcount_begin(&conf
->gen_lock
);
535 sh
->generation
= conf
->generation
- previous
;
536 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
538 stripe_set_idx(sector
, conf
, previous
, sh
);
541 for (i
= sh
->disks
; i
--; ) {
542 struct r5dev
*dev
= &sh
->dev
[i
];
544 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
545 test_bit(R5_LOCKED
, &dev
->flags
)) {
546 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
547 (unsigned long long)sh
->sector
, i
, dev
->toread
,
548 dev
->read
, dev
->towrite
, dev
->written
,
549 test_bit(R5_LOCKED
, &dev
->flags
));
553 raid5_build_block(sh
, i
, previous
);
555 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
557 sh
->overwrite_disks
= 0;
558 insert_hash(conf
, sh
);
559 sh
->cpu
= smp_processor_id();
560 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
563 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
566 struct stripe_head
*sh
;
568 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
569 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
570 if (sh
->sector
== sector
&& sh
->generation
== generation
)
572 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
577 * Need to check if array has failed when deciding whether to:
579 * - remove non-faulty devices
582 * This determination is simple when no reshape is happening.
583 * However if there is a reshape, we need to carefully check
584 * both the before and after sections.
585 * This is because some failed devices may only affect one
586 * of the two sections, and some non-in_sync devices may
587 * be insync in the section most affected by failed devices.
589 static int calc_degraded(struct r5conf
*conf
)
591 int degraded
, degraded2
;
596 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
597 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
598 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
599 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
600 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
602 else if (test_bit(In_sync
, &rdev
->flags
))
605 /* not in-sync or faulty.
606 * If the reshape increases the number of devices,
607 * this is being recovered by the reshape, so
608 * this 'previous' section is not in_sync.
609 * If the number of devices is being reduced however,
610 * the device can only be part of the array if
611 * we are reverting a reshape, so this section will
614 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
618 if (conf
->raid_disks
== conf
->previous_raid_disks
)
622 for (i
= 0; i
< conf
->raid_disks
; i
++) {
623 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
624 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
625 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
626 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
628 else if (test_bit(In_sync
, &rdev
->flags
))
631 /* not in-sync or faulty.
632 * If reshape increases the number of devices, this
633 * section has already been recovered, else it
634 * almost certainly hasn't.
636 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
640 if (degraded2
> degraded
)
645 static int has_failed(struct r5conf
*conf
)
649 if (conf
->mddev
->reshape_position
== MaxSector
)
650 return conf
->mddev
->degraded
> conf
->max_degraded
;
652 degraded
= calc_degraded(conf
);
653 if (degraded
> conf
->max_degraded
)
658 static struct stripe_head
*
659 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
660 int previous
, int noblock
, int noquiesce
)
662 struct stripe_head
*sh
;
663 int hash
= stripe_hash_locks_hash(sector
);
665 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
667 spin_lock_irq(conf
->hash_locks
+ hash
);
670 wait_event_lock_irq(conf
->wait_for_stripe
,
671 conf
->quiesce
== 0 || noquiesce
,
672 *(conf
->hash_locks
+ hash
));
673 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
675 if (!conf
->inactive_blocked
)
676 sh
= get_free_stripe(conf
, hash
);
677 if (noblock
&& sh
== NULL
)
680 conf
->inactive_blocked
= 1;
682 conf
->wait_for_stripe
,
683 !list_empty(conf
->inactive_list
+ hash
) &&
684 (atomic_read(&conf
->active_stripes
)
685 < (conf
->max_nr_stripes
* 3 / 4)
686 || !conf
->inactive_blocked
),
687 *(conf
->hash_locks
+ hash
));
688 conf
->inactive_blocked
= 0;
690 init_stripe(sh
, sector
, previous
);
691 atomic_inc(&sh
->count
);
693 } else if (!atomic_inc_not_zero(&sh
->count
)) {
694 spin_lock(&conf
->device_lock
);
695 if (!atomic_read(&sh
->count
)) {
696 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
697 atomic_inc(&conf
->active_stripes
);
698 BUG_ON(list_empty(&sh
->lru
) &&
699 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
700 list_del_init(&sh
->lru
);
702 sh
->group
->stripes_cnt
--;
706 atomic_inc(&sh
->count
);
707 spin_unlock(&conf
->device_lock
);
709 } while (sh
== NULL
);
711 spin_unlock_irq(conf
->hash_locks
+ hash
);
715 static bool is_full_stripe_write(struct stripe_head
*sh
)
717 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
718 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
721 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
725 spin_lock(&sh2
->stripe_lock
);
726 spin_lock_nested(&sh1
->stripe_lock
, 1);
728 spin_lock(&sh1
->stripe_lock
);
729 spin_lock_nested(&sh2
->stripe_lock
, 1);
733 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
735 spin_unlock(&sh1
->stripe_lock
);
736 spin_unlock(&sh2
->stripe_lock
);
740 /* Only freshly new full stripe normal write stripe can be added to a batch list */
741 static bool stripe_can_batch(struct stripe_head
*sh
)
743 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
744 is_full_stripe_write(sh
);
747 /* we only do back search */
748 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
750 struct stripe_head
*head
;
751 sector_t head_sector
, tmp_sec
;
755 if (!stripe_can_batch(sh
))
757 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
758 tmp_sec
= sh
->sector
;
759 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
761 head_sector
= sh
->sector
- STRIPE_SECTORS
;
763 hash
= stripe_hash_locks_hash(head_sector
);
764 spin_lock_irq(conf
->hash_locks
+ hash
);
765 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
766 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
767 spin_lock(&conf
->device_lock
);
768 if (!atomic_read(&head
->count
)) {
769 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
770 atomic_inc(&conf
->active_stripes
);
771 BUG_ON(list_empty(&head
->lru
) &&
772 !test_bit(STRIPE_EXPANDING
, &head
->state
));
773 list_del_init(&head
->lru
);
775 head
->group
->stripes_cnt
--;
779 atomic_inc(&head
->count
);
780 spin_unlock(&conf
->device_lock
);
782 spin_unlock_irq(conf
->hash_locks
+ hash
);
786 if (!stripe_can_batch(head
))
789 lock_two_stripes(head
, sh
);
790 /* clear_batch_ready clear the flag */
791 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
798 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
800 if (head
->dev
[dd_idx
].towrite
->bi_rw
!= sh
->dev
[dd_idx
].towrite
->bi_rw
)
803 if (head
->batch_head
) {
804 spin_lock(&head
->batch_head
->batch_lock
);
805 /* This batch list is already running */
806 if (!stripe_can_batch(head
)) {
807 spin_unlock(&head
->batch_head
->batch_lock
);
812 * at this point, head's BATCH_READY could be cleared, but we
813 * can still add the stripe to batch list
815 list_add(&sh
->batch_list
, &head
->batch_list
);
816 spin_unlock(&head
->batch_head
->batch_lock
);
818 sh
->batch_head
= head
->batch_head
;
820 head
->batch_head
= head
;
821 sh
->batch_head
= head
->batch_head
;
822 spin_lock(&head
->batch_lock
);
823 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
824 spin_unlock(&head
->batch_lock
);
827 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
828 if (atomic_dec_return(&conf
->preread_active_stripes
)
830 md_wakeup_thread(conf
->mddev
->thread
);
832 atomic_inc(&sh
->count
);
834 unlock_two_stripes(head
, sh
);
836 release_stripe(head
);
839 /* Determine if 'data_offset' or 'new_data_offset' should be used
840 * in this stripe_head.
842 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
844 sector_t progress
= conf
->reshape_progress
;
845 /* Need a memory barrier to make sure we see the value
846 * of conf->generation, or ->data_offset that was set before
847 * reshape_progress was updated.
850 if (progress
== MaxSector
)
852 if (sh
->generation
== conf
->generation
- 1)
854 /* We are in a reshape, and this is a new-generation stripe,
855 * so use new_data_offset.
861 raid5_end_read_request(struct bio
*bi
, int error
);
863 raid5_end_write_request(struct bio
*bi
, int error
);
865 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
867 struct r5conf
*conf
= sh
->raid_conf
;
868 int i
, disks
= sh
->disks
;
869 struct stripe_head
*head_sh
= sh
;
873 for (i
= disks
; i
--; ) {
875 int replace_only
= 0;
876 struct bio
*bi
, *rbi
;
877 struct md_rdev
*rdev
, *rrdev
= NULL
;
880 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
881 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
885 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
887 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
889 else if (test_and_clear_bit(R5_WantReplace
,
890 &sh
->dev
[i
].flags
)) {
895 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
899 bi
= &sh
->dev
[i
].req
;
900 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
903 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
904 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
905 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
914 /* We raced and saw duplicates */
917 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
922 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
925 atomic_inc(&rdev
->nr_pending
);
926 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
929 atomic_inc(&rrdev
->nr_pending
);
932 /* We have already checked bad blocks for reads. Now
933 * need to check for writes. We never accept write errors
934 * on the replacement, so we don't to check rrdev.
936 while ((rw
& WRITE
) && rdev
&&
937 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
940 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
941 &first_bad
, &bad_sectors
);
946 set_bit(BlockedBadBlocks
, &rdev
->flags
);
947 if (!conf
->mddev
->external
&&
948 conf
->mddev
->flags
) {
949 /* It is very unlikely, but we might
950 * still need to write out the
951 * bad block log - better give it
953 md_check_recovery(conf
->mddev
);
956 * Because md_wait_for_blocked_rdev
957 * will dec nr_pending, we must
958 * increment it first.
960 atomic_inc(&rdev
->nr_pending
);
961 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
963 /* Acknowledged bad block - skip the write */
964 rdev_dec_pending(rdev
, conf
->mddev
);
970 if (s
->syncing
|| s
->expanding
|| s
->expanded
972 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
974 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
977 bi
->bi_bdev
= rdev
->bdev
;
979 bi
->bi_end_io
= (rw
& WRITE
)
980 ? raid5_end_write_request
981 : raid5_end_read_request
;
984 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
985 __func__
, (unsigned long long)sh
->sector
,
987 atomic_inc(&sh
->count
);
989 atomic_inc(&head_sh
->count
);
990 if (use_new_offset(conf
, sh
))
991 bi
->bi_iter
.bi_sector
= (sh
->sector
992 + rdev
->new_data_offset
);
994 bi
->bi_iter
.bi_sector
= (sh
->sector
995 + rdev
->data_offset
);
996 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
997 bi
->bi_rw
|= REQ_NOMERGE
;
999 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1000 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1001 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1003 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1004 bi
->bi_io_vec
[0].bv_offset
= 0;
1005 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1007 * If this is discard request, set bi_vcnt 0. We don't
1008 * want to confuse SCSI because SCSI will replace payload
1010 if (rw
& REQ_DISCARD
)
1013 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1015 if (conf
->mddev
->gendisk
)
1016 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
1017 bi
, disk_devt(conf
->mddev
->gendisk
),
1019 generic_make_request(bi
);
1022 if (s
->syncing
|| s
->expanding
|| s
->expanded
1024 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1026 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1029 rbi
->bi_bdev
= rrdev
->bdev
;
1031 BUG_ON(!(rw
& WRITE
));
1032 rbi
->bi_end_io
= raid5_end_write_request
;
1033 rbi
->bi_private
= sh
;
1035 pr_debug("%s: for %llu schedule op %ld on "
1036 "replacement disc %d\n",
1037 __func__
, (unsigned long long)sh
->sector
,
1039 atomic_inc(&sh
->count
);
1041 atomic_inc(&head_sh
->count
);
1042 if (use_new_offset(conf
, sh
))
1043 rbi
->bi_iter
.bi_sector
= (sh
->sector
1044 + rrdev
->new_data_offset
);
1046 rbi
->bi_iter
.bi_sector
= (sh
->sector
1047 + rrdev
->data_offset
);
1048 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1049 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1050 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1052 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1053 rbi
->bi_io_vec
[0].bv_offset
= 0;
1054 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1056 * If this is discard request, set bi_vcnt 0. We don't
1057 * want to confuse SCSI because SCSI will replace payload
1059 if (rw
& REQ_DISCARD
)
1061 if (conf
->mddev
->gendisk
)
1062 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
1063 rbi
, disk_devt(conf
->mddev
->gendisk
),
1065 generic_make_request(rbi
);
1067 if (!rdev
&& !rrdev
) {
1069 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1070 pr_debug("skip op %ld on disc %d for sector %llu\n",
1071 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
1072 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1074 set_bit(STRIPE_BATCH_ERR
,
1075 &sh
->batch_head
->state
);
1076 set_bit(STRIPE_HANDLE
, &sh
->state
);
1079 if (!head_sh
->batch_head
)
1081 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1088 static struct dma_async_tx_descriptor
*
1089 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1090 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1091 struct stripe_head
*sh
)
1094 struct bvec_iter iter
;
1095 struct page
*bio_page
;
1097 struct async_submit_ctl submit
;
1098 enum async_tx_flags flags
= 0;
1100 if (bio
->bi_iter
.bi_sector
>= sector
)
1101 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1103 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1106 flags
|= ASYNC_TX_FENCE
;
1107 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1109 bio_for_each_segment(bvl
, bio
, iter
) {
1110 int len
= bvl
.bv_len
;
1114 if (page_offset
< 0) {
1115 b_offset
= -page_offset
;
1116 page_offset
+= b_offset
;
1120 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1121 clen
= STRIPE_SIZE
- page_offset
;
1126 b_offset
+= bvl
.bv_offset
;
1127 bio_page
= bvl
.bv_page
;
1129 if (sh
->raid_conf
->skip_copy
&&
1130 b_offset
== 0 && page_offset
== 0 &&
1131 clen
== STRIPE_SIZE
)
1134 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1135 b_offset
, clen
, &submit
);
1137 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1138 page_offset
, clen
, &submit
);
1140 /* chain the operations */
1141 submit
.depend_tx
= tx
;
1143 if (clen
< len
) /* hit end of page */
1151 static void ops_complete_biofill(void *stripe_head_ref
)
1153 struct stripe_head
*sh
= stripe_head_ref
;
1154 struct bio
*return_bi
= NULL
;
1157 pr_debug("%s: stripe %llu\n", __func__
,
1158 (unsigned long long)sh
->sector
);
1160 /* clear completed biofills */
1161 for (i
= sh
->disks
; i
--; ) {
1162 struct r5dev
*dev
= &sh
->dev
[i
];
1164 /* acknowledge completion of a biofill operation */
1165 /* and check if we need to reply to a read request,
1166 * new R5_Wantfill requests are held off until
1167 * !STRIPE_BIOFILL_RUN
1169 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1170 struct bio
*rbi
, *rbi2
;
1175 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1176 dev
->sector
+ STRIPE_SECTORS
) {
1177 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1178 if (!raid5_dec_bi_active_stripes(rbi
)) {
1179 rbi
->bi_next
= return_bi
;
1186 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1188 return_io(return_bi
);
1190 set_bit(STRIPE_HANDLE
, &sh
->state
);
1194 static void ops_run_biofill(struct stripe_head
*sh
)
1196 struct dma_async_tx_descriptor
*tx
= NULL
;
1197 struct async_submit_ctl submit
;
1200 BUG_ON(sh
->batch_head
);
1201 pr_debug("%s: stripe %llu\n", __func__
,
1202 (unsigned long long)sh
->sector
);
1204 for (i
= sh
->disks
; i
--; ) {
1205 struct r5dev
*dev
= &sh
->dev
[i
];
1206 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1208 spin_lock_irq(&sh
->stripe_lock
);
1209 dev
->read
= rbi
= dev
->toread
;
1211 spin_unlock_irq(&sh
->stripe_lock
);
1212 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1213 dev
->sector
+ STRIPE_SECTORS
) {
1214 tx
= async_copy_data(0, rbi
, &dev
->page
,
1215 dev
->sector
, tx
, sh
);
1216 rbi
= r5_next_bio(rbi
, dev
->sector
);
1221 atomic_inc(&sh
->count
);
1222 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1223 async_trigger_callback(&submit
);
1226 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1233 tgt
= &sh
->dev
[target
];
1234 set_bit(R5_UPTODATE
, &tgt
->flags
);
1235 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1236 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1239 static void ops_complete_compute(void *stripe_head_ref
)
1241 struct stripe_head
*sh
= stripe_head_ref
;
1243 pr_debug("%s: stripe %llu\n", __func__
,
1244 (unsigned long long)sh
->sector
);
1246 /* mark the computed target(s) as uptodate */
1247 mark_target_uptodate(sh
, sh
->ops
.target
);
1248 mark_target_uptodate(sh
, sh
->ops
.target2
);
1250 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1251 if (sh
->check_state
== check_state_compute_run
)
1252 sh
->check_state
= check_state_compute_result
;
1253 set_bit(STRIPE_HANDLE
, &sh
->state
);
1257 /* return a pointer to the address conversion region of the scribble buffer */
1258 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1259 struct raid5_percpu
*percpu
, int i
)
1263 addr
= flex_array_get(percpu
->scribble
, i
);
1264 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1267 /* return a pointer to the address conversion region of the scribble buffer */
1268 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1272 addr
= flex_array_get(percpu
->scribble
, i
);
1276 static struct dma_async_tx_descriptor
*
1277 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1279 int disks
= sh
->disks
;
1280 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1281 int target
= sh
->ops
.target
;
1282 struct r5dev
*tgt
= &sh
->dev
[target
];
1283 struct page
*xor_dest
= tgt
->page
;
1285 struct dma_async_tx_descriptor
*tx
;
1286 struct async_submit_ctl submit
;
1289 BUG_ON(sh
->batch_head
);
1291 pr_debug("%s: stripe %llu block: %d\n",
1292 __func__
, (unsigned long long)sh
->sector
, target
);
1293 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1295 for (i
= disks
; i
--; )
1297 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1299 atomic_inc(&sh
->count
);
1301 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1302 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1303 if (unlikely(count
== 1))
1304 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1306 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1311 /* set_syndrome_sources - populate source buffers for gen_syndrome
1312 * @srcs - (struct page *) array of size sh->disks
1313 * @sh - stripe_head to parse
1315 * Populates srcs in proper layout order for the stripe and returns the
1316 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1317 * destination buffer is recorded in srcs[count] and the Q destination
1318 * is recorded in srcs[count+1]].
1320 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
1322 int disks
= sh
->disks
;
1323 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1324 int d0_idx
= raid6_d0(sh
);
1328 for (i
= 0; i
< disks
; i
++)
1334 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1336 srcs
[slot
] = sh
->dev
[i
].page
;
1337 i
= raid6_next_disk(i
, disks
);
1338 } while (i
!= d0_idx
);
1340 return syndrome_disks
;
1343 static struct dma_async_tx_descriptor
*
1344 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1346 int disks
= sh
->disks
;
1347 struct page
**blocks
= to_addr_page(percpu
, 0);
1349 int qd_idx
= sh
->qd_idx
;
1350 struct dma_async_tx_descriptor
*tx
;
1351 struct async_submit_ctl submit
;
1357 BUG_ON(sh
->batch_head
);
1358 if (sh
->ops
.target
< 0)
1359 target
= sh
->ops
.target2
;
1360 else if (sh
->ops
.target2
< 0)
1361 target
= sh
->ops
.target
;
1363 /* we should only have one valid target */
1366 pr_debug("%s: stripe %llu block: %d\n",
1367 __func__
, (unsigned long long)sh
->sector
, target
);
1369 tgt
= &sh
->dev
[target
];
1370 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1373 atomic_inc(&sh
->count
);
1375 if (target
== qd_idx
) {
1376 count
= set_syndrome_sources(blocks
, sh
);
1377 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1378 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1379 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1380 ops_complete_compute
, sh
,
1381 to_addr_conv(sh
, percpu
, 0));
1382 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1384 /* Compute any data- or p-drive using XOR */
1386 for (i
= disks
; i
-- ; ) {
1387 if (i
== target
|| i
== qd_idx
)
1389 blocks
[count
++] = sh
->dev
[i
].page
;
1392 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1393 NULL
, ops_complete_compute
, sh
,
1394 to_addr_conv(sh
, percpu
, 0));
1395 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1401 static struct dma_async_tx_descriptor
*
1402 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1404 int i
, count
, disks
= sh
->disks
;
1405 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1406 int d0_idx
= raid6_d0(sh
);
1407 int faila
= -1, failb
= -1;
1408 int target
= sh
->ops
.target
;
1409 int target2
= sh
->ops
.target2
;
1410 struct r5dev
*tgt
= &sh
->dev
[target
];
1411 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1412 struct dma_async_tx_descriptor
*tx
;
1413 struct page
**blocks
= to_addr_page(percpu
, 0);
1414 struct async_submit_ctl submit
;
1416 BUG_ON(sh
->batch_head
);
1417 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1418 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1419 BUG_ON(target
< 0 || target2
< 0);
1420 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1421 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1423 /* we need to open-code set_syndrome_sources to handle the
1424 * slot number conversion for 'faila' and 'failb'
1426 for (i
= 0; i
< disks
; i
++)
1431 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1433 blocks
[slot
] = sh
->dev
[i
].page
;
1439 i
= raid6_next_disk(i
, disks
);
1440 } while (i
!= d0_idx
);
1442 BUG_ON(faila
== failb
);
1445 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1446 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1448 atomic_inc(&sh
->count
);
1450 if (failb
== syndrome_disks
+1) {
1451 /* Q disk is one of the missing disks */
1452 if (faila
== syndrome_disks
) {
1453 /* Missing P+Q, just recompute */
1454 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1455 ops_complete_compute
, sh
,
1456 to_addr_conv(sh
, percpu
, 0));
1457 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1458 STRIPE_SIZE
, &submit
);
1462 int qd_idx
= sh
->qd_idx
;
1464 /* Missing D+Q: recompute D from P, then recompute Q */
1465 if (target
== qd_idx
)
1466 data_target
= target2
;
1468 data_target
= target
;
1471 for (i
= disks
; i
-- ; ) {
1472 if (i
== data_target
|| i
== qd_idx
)
1474 blocks
[count
++] = sh
->dev
[i
].page
;
1476 dest
= sh
->dev
[data_target
].page
;
1477 init_async_submit(&submit
,
1478 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1480 to_addr_conv(sh
, percpu
, 0));
1481 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1484 count
= set_syndrome_sources(blocks
, sh
);
1485 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1486 ops_complete_compute
, sh
,
1487 to_addr_conv(sh
, percpu
, 0));
1488 return async_gen_syndrome(blocks
, 0, count
+2,
1489 STRIPE_SIZE
, &submit
);
1492 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1493 ops_complete_compute
, sh
,
1494 to_addr_conv(sh
, percpu
, 0));
1495 if (failb
== syndrome_disks
) {
1496 /* We're missing D+P. */
1497 return async_raid6_datap_recov(syndrome_disks
+2,
1501 /* We're missing D+D. */
1502 return async_raid6_2data_recov(syndrome_disks
+2,
1503 STRIPE_SIZE
, faila
, failb
,
1509 static void ops_complete_prexor(void *stripe_head_ref
)
1511 struct stripe_head
*sh
= stripe_head_ref
;
1513 pr_debug("%s: stripe %llu\n", __func__
,
1514 (unsigned long long)sh
->sector
);
1517 static struct dma_async_tx_descriptor
*
1518 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1519 struct dma_async_tx_descriptor
*tx
)
1521 int disks
= sh
->disks
;
1522 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1523 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1524 struct async_submit_ctl submit
;
1526 /* existing parity data subtracted */
1527 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1529 BUG_ON(sh
->batch_head
);
1530 pr_debug("%s: stripe %llu\n", __func__
,
1531 (unsigned long long)sh
->sector
);
1533 for (i
= disks
; i
--; ) {
1534 struct r5dev
*dev
= &sh
->dev
[i
];
1535 /* Only process blocks that are known to be uptodate */
1536 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1537 xor_srcs
[count
++] = dev
->page
;
1540 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1541 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1542 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1547 static struct dma_async_tx_descriptor
*
1548 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1550 int disks
= sh
->disks
;
1552 struct stripe_head
*head_sh
= sh
;
1554 pr_debug("%s: stripe %llu\n", __func__
,
1555 (unsigned long long)sh
->sector
);
1557 for (i
= disks
; i
--; ) {
1562 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1567 spin_lock_irq(&sh
->stripe_lock
);
1568 chosen
= dev
->towrite
;
1569 dev
->towrite
= NULL
;
1570 sh
->overwrite_disks
= 0;
1571 BUG_ON(dev
->written
);
1572 wbi
= dev
->written
= chosen
;
1573 spin_unlock_irq(&sh
->stripe_lock
);
1574 WARN_ON(dev
->page
!= dev
->orig_page
);
1576 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1577 dev
->sector
+ STRIPE_SECTORS
) {
1578 if (wbi
->bi_rw
& REQ_FUA
)
1579 set_bit(R5_WantFUA
, &dev
->flags
);
1580 if (wbi
->bi_rw
& REQ_SYNC
)
1581 set_bit(R5_SyncIO
, &dev
->flags
);
1582 if (wbi
->bi_rw
& REQ_DISCARD
)
1583 set_bit(R5_Discard
, &dev
->flags
);
1585 tx
= async_copy_data(1, wbi
, &dev
->page
,
1586 dev
->sector
, tx
, sh
);
1587 if (dev
->page
!= dev
->orig_page
) {
1588 set_bit(R5_SkipCopy
, &dev
->flags
);
1589 clear_bit(R5_UPTODATE
, &dev
->flags
);
1590 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1593 wbi
= r5_next_bio(wbi
, dev
->sector
);
1596 if (head_sh
->batch_head
) {
1597 sh
= list_first_entry(&sh
->batch_list
,
1610 static void ops_complete_reconstruct(void *stripe_head_ref
)
1612 struct stripe_head
*sh
= stripe_head_ref
;
1613 int disks
= sh
->disks
;
1614 int pd_idx
= sh
->pd_idx
;
1615 int qd_idx
= sh
->qd_idx
;
1617 bool fua
= false, sync
= false, discard
= false;
1619 pr_debug("%s: stripe %llu\n", __func__
,
1620 (unsigned long long)sh
->sector
);
1622 for (i
= disks
; i
--; ) {
1623 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1624 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1625 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1628 for (i
= disks
; i
--; ) {
1629 struct r5dev
*dev
= &sh
->dev
[i
];
1631 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1632 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1633 set_bit(R5_UPTODATE
, &dev
->flags
);
1635 set_bit(R5_WantFUA
, &dev
->flags
);
1637 set_bit(R5_SyncIO
, &dev
->flags
);
1641 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1642 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1643 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1644 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1646 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1647 sh
->reconstruct_state
= reconstruct_state_result
;
1650 set_bit(STRIPE_HANDLE
, &sh
->state
);
1655 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1656 struct dma_async_tx_descriptor
*tx
)
1658 int disks
= sh
->disks
;
1659 struct page
**xor_srcs
;
1660 struct async_submit_ctl submit
;
1661 int count
, pd_idx
= sh
->pd_idx
, i
;
1662 struct page
*xor_dest
;
1664 unsigned long flags
;
1666 struct stripe_head
*head_sh
= sh
;
1669 pr_debug("%s: stripe %llu\n", __func__
,
1670 (unsigned long long)sh
->sector
);
1672 for (i
= 0; i
< sh
->disks
; i
++) {
1675 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1678 if (i
>= sh
->disks
) {
1679 atomic_inc(&sh
->count
);
1680 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1681 ops_complete_reconstruct(sh
);
1686 xor_srcs
= to_addr_page(percpu
, j
);
1687 /* check if prexor is active which means only process blocks
1688 * that are part of a read-modify-write (written)
1690 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1692 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1693 for (i
= disks
; i
--; ) {
1694 struct r5dev
*dev
= &sh
->dev
[i
];
1695 if (head_sh
->dev
[i
].written
)
1696 xor_srcs
[count
++] = dev
->page
;
1699 xor_dest
= sh
->dev
[pd_idx
].page
;
1700 for (i
= disks
; i
--; ) {
1701 struct r5dev
*dev
= &sh
->dev
[i
];
1703 xor_srcs
[count
++] = dev
->page
;
1707 /* 1/ if we prexor'd then the dest is reused as a source
1708 * 2/ if we did not prexor then we are redoing the parity
1709 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1710 * for the synchronous xor case
1712 last_stripe
= !head_sh
->batch_head
||
1713 list_first_entry(&sh
->batch_list
,
1714 struct stripe_head
, batch_list
) == head_sh
;
1716 flags
= ASYNC_TX_ACK
|
1717 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1719 atomic_inc(&head_sh
->count
);
1720 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1721 to_addr_conv(sh
, percpu
, j
));
1723 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1724 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1725 to_addr_conv(sh
, percpu
, j
));
1728 if (unlikely(count
== 1))
1729 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1731 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1734 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1741 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1742 struct dma_async_tx_descriptor
*tx
)
1744 struct async_submit_ctl submit
;
1745 struct page
**blocks
;
1746 int count
, i
, j
= 0;
1747 struct stripe_head
*head_sh
= sh
;
1750 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1752 for (i
= 0; i
< sh
->disks
; i
++) {
1753 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1755 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1758 if (i
>= sh
->disks
) {
1759 atomic_inc(&sh
->count
);
1760 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1761 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1762 ops_complete_reconstruct(sh
);
1767 blocks
= to_addr_page(percpu
, j
);
1768 count
= set_syndrome_sources(blocks
, sh
);
1769 last_stripe
= !head_sh
->batch_head
||
1770 list_first_entry(&sh
->batch_list
,
1771 struct stripe_head
, batch_list
) == head_sh
;
1774 atomic_inc(&head_sh
->count
);
1775 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1776 head_sh
, to_addr_conv(sh
, percpu
, j
));
1778 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1779 to_addr_conv(sh
, percpu
, j
));
1780 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1783 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1789 static void ops_complete_check(void *stripe_head_ref
)
1791 struct stripe_head
*sh
= stripe_head_ref
;
1793 pr_debug("%s: stripe %llu\n", __func__
,
1794 (unsigned long long)sh
->sector
);
1796 sh
->check_state
= check_state_check_result
;
1797 set_bit(STRIPE_HANDLE
, &sh
->state
);
1801 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1803 int disks
= sh
->disks
;
1804 int pd_idx
= sh
->pd_idx
;
1805 int qd_idx
= sh
->qd_idx
;
1806 struct page
*xor_dest
;
1807 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1808 struct dma_async_tx_descriptor
*tx
;
1809 struct async_submit_ctl submit
;
1813 pr_debug("%s: stripe %llu\n", __func__
,
1814 (unsigned long long)sh
->sector
);
1816 BUG_ON(sh
->batch_head
);
1818 xor_dest
= sh
->dev
[pd_idx
].page
;
1819 xor_srcs
[count
++] = xor_dest
;
1820 for (i
= disks
; i
--; ) {
1821 if (i
== pd_idx
|| i
== qd_idx
)
1823 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1826 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1827 to_addr_conv(sh
, percpu
, 0));
1828 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1829 &sh
->ops
.zero_sum_result
, &submit
);
1831 atomic_inc(&sh
->count
);
1832 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1833 tx
= async_trigger_callback(&submit
);
1836 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1838 struct page
**srcs
= to_addr_page(percpu
, 0);
1839 struct async_submit_ctl submit
;
1842 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1843 (unsigned long long)sh
->sector
, checkp
);
1845 BUG_ON(sh
->batch_head
);
1846 count
= set_syndrome_sources(srcs
, sh
);
1850 atomic_inc(&sh
->count
);
1851 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1852 sh
, to_addr_conv(sh
, percpu
, 0));
1853 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1854 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1857 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1859 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1860 struct dma_async_tx_descriptor
*tx
= NULL
;
1861 struct r5conf
*conf
= sh
->raid_conf
;
1862 int level
= conf
->level
;
1863 struct raid5_percpu
*percpu
;
1867 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1868 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1869 ops_run_biofill(sh
);
1873 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1875 tx
= ops_run_compute5(sh
, percpu
);
1877 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1878 tx
= ops_run_compute6_1(sh
, percpu
);
1880 tx
= ops_run_compute6_2(sh
, percpu
);
1882 /* terminate the chain if reconstruct is not set to be run */
1883 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1887 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1888 tx
= ops_run_prexor(sh
, percpu
, tx
);
1890 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1891 tx
= ops_run_biodrain(sh
, tx
);
1895 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1897 ops_run_reconstruct5(sh
, percpu
, tx
);
1899 ops_run_reconstruct6(sh
, percpu
, tx
);
1902 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1903 if (sh
->check_state
== check_state_run
)
1904 ops_run_check_p(sh
, percpu
);
1905 else if (sh
->check_state
== check_state_run_q
)
1906 ops_run_check_pq(sh
, percpu
, 0);
1907 else if (sh
->check_state
== check_state_run_pq
)
1908 ops_run_check_pq(sh
, percpu
, 1);
1913 if (overlap_clear
&& !sh
->batch_head
)
1914 for (i
= disks
; i
--; ) {
1915 struct r5dev
*dev
= &sh
->dev
[i
];
1916 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1917 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1922 static int grow_one_stripe(struct r5conf
*conf
, int hash
)
1924 struct stripe_head
*sh
;
1925 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1929 sh
->raid_conf
= conf
;
1931 spin_lock_init(&sh
->stripe_lock
);
1933 if (grow_buffers(sh
)) {
1935 kmem_cache_free(conf
->slab_cache
, sh
);
1938 sh
->hash_lock_index
= hash
;
1939 /* we just created an active stripe so... */
1940 atomic_set(&sh
->count
, 1);
1941 atomic_inc(&conf
->active_stripes
);
1942 INIT_LIST_HEAD(&sh
->lru
);
1944 spin_lock_init(&sh
->batch_lock
);
1945 INIT_LIST_HEAD(&sh
->batch_list
);
1946 sh
->batch_head
= NULL
;
1951 static int grow_stripes(struct r5conf
*conf
, int num
)
1953 struct kmem_cache
*sc
;
1954 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1957 if (conf
->mddev
->gendisk
)
1958 sprintf(conf
->cache_name
[0],
1959 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1961 sprintf(conf
->cache_name
[0],
1962 "raid%d-%p", conf
->level
, conf
->mddev
);
1963 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1965 conf
->active_name
= 0;
1966 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1967 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1971 conf
->slab_cache
= sc
;
1972 conf
->pool_size
= devs
;
1973 hash
= conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
1975 if (!grow_one_stripe(conf
, hash
))
1977 conf
->max_nr_stripes
++;
1978 hash
= (hash
+ 1) % NR_STRIPE_HASH_LOCKS
;
1984 * scribble_len - return the required size of the scribble region
1985 * @num - total number of disks in the array
1987 * The size must be enough to contain:
1988 * 1/ a struct page pointer for each device in the array +2
1989 * 2/ room to convert each entry in (1) to its corresponding dma
1990 * (dma_map_page()) or page (page_address()) address.
1992 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1993 * calculate over all devices (not just the data blocks), using zeros in place
1994 * of the P and Q blocks.
1996 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
1998 struct flex_array
*ret
;
2001 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2002 ret
= flex_array_alloc(len
, cnt
, flags
);
2005 /* always prealloc all elements, so no locking is required */
2006 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2007 flex_array_free(ret
);
2013 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2015 /* Make all the stripes able to hold 'newsize' devices.
2016 * New slots in each stripe get 'page' set to a new page.
2018 * This happens in stages:
2019 * 1/ create a new kmem_cache and allocate the required number of
2021 * 2/ gather all the old stripe_heads and transfer the pages across
2022 * to the new stripe_heads. This will have the side effect of
2023 * freezing the array as once all stripe_heads have been collected,
2024 * no IO will be possible. Old stripe heads are freed once their
2025 * pages have been transferred over, and the old kmem_cache is
2026 * freed when all stripes are done.
2027 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2028 * we simple return a failre status - no need to clean anything up.
2029 * 4/ allocate new pages for the new slots in the new stripe_heads.
2030 * If this fails, we don't bother trying the shrink the
2031 * stripe_heads down again, we just leave them as they are.
2032 * As each stripe_head is processed the new one is released into
2035 * Once step2 is started, we cannot afford to wait for a write,
2036 * so we use GFP_NOIO allocations.
2038 struct stripe_head
*osh
, *nsh
;
2039 LIST_HEAD(newstripes
);
2040 struct disk_info
*ndisks
;
2043 struct kmem_cache
*sc
;
2047 if (newsize
<= conf
->pool_size
)
2048 return 0; /* never bother to shrink */
2050 err
= md_allow_write(conf
->mddev
);
2055 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2056 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2061 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2062 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
2066 nsh
->raid_conf
= conf
;
2067 spin_lock_init(&nsh
->stripe_lock
);
2069 list_add(&nsh
->lru
, &newstripes
);
2072 /* didn't get enough, give up */
2073 while (!list_empty(&newstripes
)) {
2074 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2075 list_del(&nsh
->lru
);
2076 kmem_cache_free(sc
, nsh
);
2078 kmem_cache_destroy(sc
);
2081 /* Step 2 - Must use GFP_NOIO now.
2082 * OK, we have enough stripes, start collecting inactive
2083 * stripes and copying them over
2087 list_for_each_entry(nsh
, &newstripes
, lru
) {
2088 lock_device_hash_lock(conf
, hash
);
2089 wait_event_cmd(conf
->wait_for_stripe
,
2090 !list_empty(conf
->inactive_list
+ hash
),
2091 unlock_device_hash_lock(conf
, hash
),
2092 lock_device_hash_lock(conf
, hash
));
2093 osh
= get_free_stripe(conf
, hash
);
2094 unlock_device_hash_lock(conf
, hash
);
2095 atomic_set(&nsh
->count
, 1);
2096 for(i
=0; i
<conf
->pool_size
; i
++) {
2097 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2098 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2100 for( ; i
<newsize
; i
++)
2101 nsh
->dev
[i
].page
= NULL
;
2102 nsh
->hash_lock_index
= hash
;
2103 kmem_cache_free(conf
->slab_cache
, osh
);
2105 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2106 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2111 kmem_cache_destroy(conf
->slab_cache
);
2114 * At this point, we are holding all the stripes so the array
2115 * is completely stalled, so now is a good time to resize
2116 * conf->disks and the scribble region
2118 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2120 for (i
=0; i
<conf
->raid_disks
; i
++)
2121 ndisks
[i
] = conf
->disks
[i
];
2123 conf
->disks
= ndisks
;
2128 for_each_present_cpu(cpu
) {
2129 struct raid5_percpu
*percpu
;
2130 struct flex_array
*scribble
;
2132 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2133 scribble
= scribble_alloc(newsize
, conf
->chunk_sectors
/
2134 STRIPE_SECTORS
, GFP_NOIO
);
2137 flex_array_free(percpu
->scribble
);
2138 percpu
->scribble
= scribble
;
2146 /* Step 4, return new stripes to service */
2147 while(!list_empty(&newstripes
)) {
2148 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2149 list_del_init(&nsh
->lru
);
2151 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2152 if (nsh
->dev
[i
].page
== NULL
) {
2153 struct page
*p
= alloc_page(GFP_NOIO
);
2154 nsh
->dev
[i
].page
= p
;
2155 nsh
->dev
[i
].orig_page
= p
;
2159 release_stripe(nsh
);
2161 /* critical section pass, GFP_NOIO no longer needed */
2163 conf
->slab_cache
= sc
;
2164 conf
->active_name
= 1-conf
->active_name
;
2165 conf
->pool_size
= newsize
;
2169 static int drop_one_stripe(struct r5conf
*conf
, int hash
)
2171 struct stripe_head
*sh
;
2173 spin_lock_irq(conf
->hash_locks
+ hash
);
2174 sh
= get_free_stripe(conf
, hash
);
2175 spin_unlock_irq(conf
->hash_locks
+ hash
);
2178 BUG_ON(atomic_read(&sh
->count
));
2180 kmem_cache_free(conf
->slab_cache
, sh
);
2181 atomic_dec(&conf
->active_stripes
);
2185 static void shrink_stripes(struct r5conf
*conf
)
2188 for (hash
= 0; hash
< NR_STRIPE_HASH_LOCKS
; hash
++)
2189 while (drop_one_stripe(conf
, hash
))
2192 if (conf
->slab_cache
)
2193 kmem_cache_destroy(conf
->slab_cache
);
2194 conf
->slab_cache
= NULL
;
2197 static void raid5_end_read_request(struct bio
* bi
, int error
)
2199 struct stripe_head
*sh
= bi
->bi_private
;
2200 struct r5conf
*conf
= sh
->raid_conf
;
2201 int disks
= sh
->disks
, i
;
2202 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2203 char b
[BDEVNAME_SIZE
];
2204 struct md_rdev
*rdev
= NULL
;
2207 for (i
=0 ; i
<disks
; i
++)
2208 if (bi
== &sh
->dev
[i
].req
)
2211 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2212 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2218 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2219 /* If replacement finished while this request was outstanding,
2220 * 'replacement' might be NULL already.
2221 * In that case it moved down to 'rdev'.
2222 * rdev is not removed until all requests are finished.
2224 rdev
= conf
->disks
[i
].replacement
;
2226 rdev
= conf
->disks
[i
].rdev
;
2228 if (use_new_offset(conf
, sh
))
2229 s
= sh
->sector
+ rdev
->new_data_offset
;
2231 s
= sh
->sector
+ rdev
->data_offset
;
2233 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2234 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2235 /* Note that this cannot happen on a
2236 * replacement device. We just fail those on
2241 "md/raid:%s: read error corrected"
2242 " (%lu sectors at %llu on %s)\n",
2243 mdname(conf
->mddev
), STRIPE_SECTORS
,
2244 (unsigned long long)s
,
2245 bdevname(rdev
->bdev
, b
));
2246 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2247 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2248 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2249 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2250 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2252 if (atomic_read(&rdev
->read_errors
))
2253 atomic_set(&rdev
->read_errors
, 0);
2255 const char *bdn
= bdevname(rdev
->bdev
, b
);
2259 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2260 atomic_inc(&rdev
->read_errors
);
2261 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2264 "md/raid:%s: read error on replacement device "
2265 "(sector %llu on %s).\n",
2266 mdname(conf
->mddev
),
2267 (unsigned long long)s
,
2269 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2273 "md/raid:%s: read error not correctable "
2274 "(sector %llu on %s).\n",
2275 mdname(conf
->mddev
),
2276 (unsigned long long)s
,
2278 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2283 "md/raid:%s: read error NOT corrected!! "
2284 "(sector %llu on %s).\n",
2285 mdname(conf
->mddev
),
2286 (unsigned long long)s
,
2288 } else if (atomic_read(&rdev
->read_errors
)
2289 > conf
->max_nr_stripes
)
2291 "md/raid:%s: Too many read errors, failing device %s.\n",
2292 mdname(conf
->mddev
), bdn
);
2295 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2296 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2299 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2300 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2301 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2303 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2305 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2306 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2308 && test_bit(In_sync
, &rdev
->flags
)
2309 && rdev_set_badblocks(
2310 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2311 md_error(conf
->mddev
, rdev
);
2314 rdev_dec_pending(rdev
, conf
->mddev
);
2315 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2316 set_bit(STRIPE_HANDLE
, &sh
->state
);
2320 static void raid5_end_write_request(struct bio
*bi
, int error
)
2322 struct stripe_head
*sh
= bi
->bi_private
;
2323 struct r5conf
*conf
= sh
->raid_conf
;
2324 int disks
= sh
->disks
, i
;
2325 struct md_rdev
*uninitialized_var(rdev
);
2326 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2329 int replacement
= 0;
2331 for (i
= 0 ; i
< disks
; i
++) {
2332 if (bi
== &sh
->dev
[i
].req
) {
2333 rdev
= conf
->disks
[i
].rdev
;
2336 if (bi
== &sh
->dev
[i
].rreq
) {
2337 rdev
= conf
->disks
[i
].replacement
;
2341 /* rdev was removed and 'replacement'
2342 * replaced it. rdev is not removed
2343 * until all requests are finished.
2345 rdev
= conf
->disks
[i
].rdev
;
2349 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2350 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2359 md_error(conf
->mddev
, rdev
);
2360 else if (is_badblock(rdev
, sh
->sector
,
2362 &first_bad
, &bad_sectors
))
2363 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2366 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2367 set_bit(WriteErrorSeen
, &rdev
->flags
);
2368 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2369 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2370 set_bit(MD_RECOVERY_NEEDED
,
2371 &rdev
->mddev
->recovery
);
2372 } else if (is_badblock(rdev
, sh
->sector
,
2374 &first_bad
, &bad_sectors
)) {
2375 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2376 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2377 /* That was a successful write so make
2378 * sure it looks like we already did
2381 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2384 rdev_dec_pending(rdev
, conf
->mddev
);
2386 if (sh
->batch_head
&& !uptodate
)
2387 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2389 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2390 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2391 set_bit(STRIPE_HANDLE
, &sh
->state
);
2394 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2395 release_stripe(sh
->batch_head
);
2398 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
2400 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
2402 struct r5dev
*dev
= &sh
->dev
[i
];
2404 bio_init(&dev
->req
);
2405 dev
->req
.bi_io_vec
= &dev
->vec
;
2406 dev
->req
.bi_max_vecs
= 1;
2407 dev
->req
.bi_private
= sh
;
2409 bio_init(&dev
->rreq
);
2410 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
2411 dev
->rreq
.bi_max_vecs
= 1;
2412 dev
->rreq
.bi_private
= sh
;
2415 dev
->sector
= compute_blocknr(sh
, i
, previous
);
2418 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2420 char b
[BDEVNAME_SIZE
];
2421 struct r5conf
*conf
= mddev
->private;
2422 unsigned long flags
;
2423 pr_debug("raid456: error called\n");
2425 spin_lock_irqsave(&conf
->device_lock
, flags
);
2426 clear_bit(In_sync
, &rdev
->flags
);
2427 mddev
->degraded
= calc_degraded(conf
);
2428 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2429 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2431 set_bit(Blocked
, &rdev
->flags
);
2432 set_bit(Faulty
, &rdev
->flags
);
2433 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2435 "md/raid:%s: Disk failure on %s, disabling device.\n"
2436 "md/raid:%s: Operation continuing on %d devices.\n",
2438 bdevname(rdev
->bdev
, b
),
2440 conf
->raid_disks
- mddev
->degraded
);
2444 * Input: a 'big' sector number,
2445 * Output: index of the data and parity disk, and the sector # in them.
2447 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2448 int previous
, int *dd_idx
,
2449 struct stripe_head
*sh
)
2451 sector_t stripe
, stripe2
;
2452 sector_t chunk_number
;
2453 unsigned int chunk_offset
;
2456 sector_t new_sector
;
2457 int algorithm
= previous
? conf
->prev_algo
2459 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2460 : conf
->chunk_sectors
;
2461 int raid_disks
= previous
? conf
->previous_raid_disks
2463 int data_disks
= raid_disks
- conf
->max_degraded
;
2465 /* First compute the information on this sector */
2468 * Compute the chunk number and the sector offset inside the chunk
2470 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2471 chunk_number
= r_sector
;
2474 * Compute the stripe number
2476 stripe
= chunk_number
;
2477 *dd_idx
= sector_div(stripe
, data_disks
);
2480 * Select the parity disk based on the user selected algorithm.
2482 pd_idx
= qd_idx
= -1;
2483 switch(conf
->level
) {
2485 pd_idx
= data_disks
;
2488 switch (algorithm
) {
2489 case ALGORITHM_LEFT_ASYMMETRIC
:
2490 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2491 if (*dd_idx
>= pd_idx
)
2494 case ALGORITHM_RIGHT_ASYMMETRIC
:
2495 pd_idx
= sector_div(stripe2
, raid_disks
);
2496 if (*dd_idx
>= pd_idx
)
2499 case ALGORITHM_LEFT_SYMMETRIC
:
2500 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2501 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2503 case ALGORITHM_RIGHT_SYMMETRIC
:
2504 pd_idx
= sector_div(stripe2
, raid_disks
);
2505 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2507 case ALGORITHM_PARITY_0
:
2511 case ALGORITHM_PARITY_N
:
2512 pd_idx
= data_disks
;
2520 switch (algorithm
) {
2521 case ALGORITHM_LEFT_ASYMMETRIC
:
2522 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2523 qd_idx
= pd_idx
+ 1;
2524 if (pd_idx
== raid_disks
-1) {
2525 (*dd_idx
)++; /* Q D D D P */
2527 } else if (*dd_idx
>= pd_idx
)
2528 (*dd_idx
) += 2; /* D D P Q D */
2530 case ALGORITHM_RIGHT_ASYMMETRIC
:
2531 pd_idx
= sector_div(stripe2
, raid_disks
);
2532 qd_idx
= pd_idx
+ 1;
2533 if (pd_idx
== raid_disks
-1) {
2534 (*dd_idx
)++; /* Q D D D P */
2536 } else if (*dd_idx
>= pd_idx
)
2537 (*dd_idx
) += 2; /* D D P Q D */
2539 case ALGORITHM_LEFT_SYMMETRIC
:
2540 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2541 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2542 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2544 case ALGORITHM_RIGHT_SYMMETRIC
:
2545 pd_idx
= sector_div(stripe2
, raid_disks
);
2546 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2547 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2550 case ALGORITHM_PARITY_0
:
2555 case ALGORITHM_PARITY_N
:
2556 pd_idx
= data_disks
;
2557 qd_idx
= data_disks
+ 1;
2560 case ALGORITHM_ROTATING_ZERO_RESTART
:
2561 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2562 * of blocks for computing Q is different.
2564 pd_idx
= sector_div(stripe2
, raid_disks
);
2565 qd_idx
= pd_idx
+ 1;
2566 if (pd_idx
== raid_disks
-1) {
2567 (*dd_idx
)++; /* Q D D D P */
2569 } else if (*dd_idx
>= pd_idx
)
2570 (*dd_idx
) += 2; /* D D P Q D */
2574 case ALGORITHM_ROTATING_N_RESTART
:
2575 /* Same a left_asymmetric, by first stripe is
2576 * D D D P Q rather than
2580 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2581 qd_idx
= pd_idx
+ 1;
2582 if (pd_idx
== raid_disks
-1) {
2583 (*dd_idx
)++; /* Q D D D P */
2585 } else if (*dd_idx
>= pd_idx
)
2586 (*dd_idx
) += 2; /* D D P Q D */
2590 case ALGORITHM_ROTATING_N_CONTINUE
:
2591 /* Same as left_symmetric but Q is before P */
2592 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2593 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2594 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2598 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2599 /* RAID5 left_asymmetric, with Q on last device */
2600 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2601 if (*dd_idx
>= pd_idx
)
2603 qd_idx
= raid_disks
- 1;
2606 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2607 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2608 if (*dd_idx
>= pd_idx
)
2610 qd_idx
= raid_disks
- 1;
2613 case ALGORITHM_LEFT_SYMMETRIC_6
:
2614 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2615 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2616 qd_idx
= raid_disks
- 1;
2619 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2620 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2621 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2622 qd_idx
= raid_disks
- 1;
2625 case ALGORITHM_PARITY_0_6
:
2628 qd_idx
= raid_disks
- 1;
2638 sh
->pd_idx
= pd_idx
;
2639 sh
->qd_idx
= qd_idx
;
2640 sh
->ddf_layout
= ddf_layout
;
2643 * Finally, compute the new sector number
2645 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2649 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2651 struct r5conf
*conf
= sh
->raid_conf
;
2652 int raid_disks
= sh
->disks
;
2653 int data_disks
= raid_disks
- conf
->max_degraded
;
2654 sector_t new_sector
= sh
->sector
, check
;
2655 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2656 : conf
->chunk_sectors
;
2657 int algorithm
= previous
? conf
->prev_algo
2661 sector_t chunk_number
;
2662 int dummy1
, dd_idx
= i
;
2664 struct stripe_head sh2
;
2666 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2667 stripe
= new_sector
;
2669 if (i
== sh
->pd_idx
)
2671 switch(conf
->level
) {
2674 switch (algorithm
) {
2675 case ALGORITHM_LEFT_ASYMMETRIC
:
2676 case ALGORITHM_RIGHT_ASYMMETRIC
:
2680 case ALGORITHM_LEFT_SYMMETRIC
:
2681 case ALGORITHM_RIGHT_SYMMETRIC
:
2684 i
-= (sh
->pd_idx
+ 1);
2686 case ALGORITHM_PARITY_0
:
2689 case ALGORITHM_PARITY_N
:
2696 if (i
== sh
->qd_idx
)
2697 return 0; /* It is the Q disk */
2698 switch (algorithm
) {
2699 case ALGORITHM_LEFT_ASYMMETRIC
:
2700 case ALGORITHM_RIGHT_ASYMMETRIC
:
2701 case ALGORITHM_ROTATING_ZERO_RESTART
:
2702 case ALGORITHM_ROTATING_N_RESTART
:
2703 if (sh
->pd_idx
== raid_disks
-1)
2704 i
--; /* Q D D D P */
2705 else if (i
> sh
->pd_idx
)
2706 i
-= 2; /* D D P Q D */
2708 case ALGORITHM_LEFT_SYMMETRIC
:
2709 case ALGORITHM_RIGHT_SYMMETRIC
:
2710 if (sh
->pd_idx
== raid_disks
-1)
2711 i
--; /* Q D D D P */
2716 i
-= (sh
->pd_idx
+ 2);
2719 case ALGORITHM_PARITY_0
:
2722 case ALGORITHM_PARITY_N
:
2724 case ALGORITHM_ROTATING_N_CONTINUE
:
2725 /* Like left_symmetric, but P is before Q */
2726 if (sh
->pd_idx
== 0)
2727 i
--; /* P D D D Q */
2732 i
-= (sh
->pd_idx
+ 1);
2735 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2736 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2740 case ALGORITHM_LEFT_SYMMETRIC_6
:
2741 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2743 i
+= data_disks
+ 1;
2744 i
-= (sh
->pd_idx
+ 1);
2746 case ALGORITHM_PARITY_0_6
:
2755 chunk_number
= stripe
* data_disks
+ i
;
2756 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2758 check
= raid5_compute_sector(conf
, r_sector
,
2759 previous
, &dummy1
, &sh2
);
2760 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2761 || sh2
.qd_idx
!= sh
->qd_idx
) {
2762 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2763 mdname(conf
->mddev
));
2770 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2771 int rcw
, int expand
)
2773 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2774 struct r5conf
*conf
= sh
->raid_conf
;
2775 int level
= conf
->level
;
2779 for (i
= disks
; i
--; ) {
2780 struct r5dev
*dev
= &sh
->dev
[i
];
2783 set_bit(R5_LOCKED
, &dev
->flags
);
2784 set_bit(R5_Wantdrain
, &dev
->flags
);
2786 clear_bit(R5_UPTODATE
, &dev
->flags
);
2790 /* if we are not expanding this is a proper write request, and
2791 * there will be bios with new data to be drained into the
2796 /* False alarm, nothing to do */
2798 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2799 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2801 sh
->reconstruct_state
= reconstruct_state_run
;
2803 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2805 if (s
->locked
+ conf
->max_degraded
== disks
)
2806 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2807 atomic_inc(&conf
->pending_full_writes
);
2810 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2811 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2813 for (i
= disks
; i
--; ) {
2814 struct r5dev
*dev
= &sh
->dev
[i
];
2819 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2820 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2821 set_bit(R5_Wantdrain
, &dev
->flags
);
2822 set_bit(R5_LOCKED
, &dev
->flags
);
2823 clear_bit(R5_UPTODATE
, &dev
->flags
);
2828 /* False alarm - nothing to do */
2830 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2831 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2832 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2833 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2836 /* keep the parity disk(s) locked while asynchronous operations
2839 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2840 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2844 int qd_idx
= sh
->qd_idx
;
2845 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2847 set_bit(R5_LOCKED
, &dev
->flags
);
2848 clear_bit(R5_UPTODATE
, &dev
->flags
);
2852 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2853 __func__
, (unsigned long long)sh
->sector
,
2854 s
->locked
, s
->ops_request
);
2858 * Each stripe/dev can have one or more bion attached.
2859 * toread/towrite point to the first in a chain.
2860 * The bi_next chain must be in order.
2862 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
2863 int forwrite
, int previous
)
2866 struct r5conf
*conf
= sh
->raid_conf
;
2869 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2870 (unsigned long long)bi
->bi_iter
.bi_sector
,
2871 (unsigned long long)sh
->sector
);
2874 * If several bio share a stripe. The bio bi_phys_segments acts as a
2875 * reference count to avoid race. The reference count should already be
2876 * increased before this function is called (for example, in
2877 * make_request()), so other bio sharing this stripe will not free the
2878 * stripe. If a stripe is owned by one stripe, the stripe lock will
2881 spin_lock_irq(&sh
->stripe_lock
);
2882 /* Don't allow new IO added to stripes in batch list */
2886 bip
= &sh
->dev
[dd_idx
].towrite
;
2890 bip
= &sh
->dev
[dd_idx
].toread
;
2891 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
2892 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
2894 bip
= & (*bip
)->bi_next
;
2896 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
2899 if (!forwrite
|| previous
)
2900 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
2902 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2906 raid5_inc_bi_active_stripes(bi
);
2909 /* check if page is covered */
2910 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2911 for (bi
=sh
->dev
[dd_idx
].towrite
;
2912 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2913 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
2914 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2915 if (bio_end_sector(bi
) >= sector
)
2916 sector
= bio_end_sector(bi
);
2918 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2919 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
2920 sh
->overwrite_disks
++;
2923 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2924 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
2925 (unsigned long long)sh
->sector
, dd_idx
);
2926 spin_unlock_irq(&sh
->stripe_lock
);
2928 if (conf
->mddev
->bitmap
&& firstwrite
) {
2929 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2931 sh
->bm_seq
= conf
->seq_flush
+1;
2932 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2935 if (stripe_can_batch(sh
))
2936 stripe_add_to_batch_list(conf
, sh
);
2940 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2941 spin_unlock_irq(&sh
->stripe_lock
);
2945 static void end_reshape(struct r5conf
*conf
);
2947 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2948 struct stripe_head
*sh
)
2950 int sectors_per_chunk
=
2951 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2953 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2954 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2956 raid5_compute_sector(conf
,
2957 stripe
* (disks
- conf
->max_degraded
)
2958 *sectors_per_chunk
+ chunk_offset
,
2964 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2965 struct stripe_head_state
*s
, int disks
,
2966 struct bio
**return_bi
)
2969 BUG_ON(sh
->batch_head
);
2970 for (i
= disks
; i
--; ) {
2974 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2975 struct md_rdev
*rdev
;
2977 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2978 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2979 atomic_inc(&rdev
->nr_pending
);
2984 if (!rdev_set_badblocks(
2988 md_error(conf
->mddev
, rdev
);
2989 rdev_dec_pending(rdev
, conf
->mddev
);
2992 spin_lock_irq(&sh
->stripe_lock
);
2993 /* fail all writes first */
2994 bi
= sh
->dev
[i
].towrite
;
2995 sh
->dev
[i
].towrite
= NULL
;
2996 sh
->overwrite_disks
= 0;
2997 spin_unlock_irq(&sh
->stripe_lock
);
3001 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3002 wake_up(&conf
->wait_for_overlap
);
3004 while (bi
&& bi
->bi_iter
.bi_sector
<
3005 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3006 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3007 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3008 if (!raid5_dec_bi_active_stripes(bi
)) {
3009 md_write_end(conf
->mddev
);
3010 bi
->bi_next
= *return_bi
;
3016 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3017 STRIPE_SECTORS
, 0, 0);
3019 /* and fail all 'written' */
3020 bi
= sh
->dev
[i
].written
;
3021 sh
->dev
[i
].written
= NULL
;
3022 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3023 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3024 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3027 if (bi
) bitmap_end
= 1;
3028 while (bi
&& bi
->bi_iter
.bi_sector
<
3029 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3030 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3031 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3032 if (!raid5_dec_bi_active_stripes(bi
)) {
3033 md_write_end(conf
->mddev
);
3034 bi
->bi_next
= *return_bi
;
3040 /* fail any reads if this device is non-operational and
3041 * the data has not reached the cache yet.
3043 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3044 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3045 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3046 spin_lock_irq(&sh
->stripe_lock
);
3047 bi
= sh
->dev
[i
].toread
;
3048 sh
->dev
[i
].toread
= NULL
;
3049 spin_unlock_irq(&sh
->stripe_lock
);
3050 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3051 wake_up(&conf
->wait_for_overlap
);
3052 while (bi
&& bi
->bi_iter
.bi_sector
<
3053 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3054 struct bio
*nextbi
=
3055 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3056 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3057 if (!raid5_dec_bi_active_stripes(bi
)) {
3058 bi
->bi_next
= *return_bi
;
3065 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3066 STRIPE_SECTORS
, 0, 0);
3067 /* If we were in the middle of a write the parity block might
3068 * still be locked - so just clear all R5_LOCKED flags
3070 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3073 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3074 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3075 md_wakeup_thread(conf
->mddev
->thread
);
3079 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3080 struct stripe_head_state
*s
)
3085 BUG_ON(sh
->batch_head
);
3086 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3087 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3088 wake_up(&conf
->wait_for_overlap
);
3091 /* There is nothing more to do for sync/check/repair.
3092 * Don't even need to abort as that is handled elsewhere
3093 * if needed, and not always wanted e.g. if there is a known
3095 * For recover/replace we need to record a bad block on all
3096 * non-sync devices, or abort the recovery
3098 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3099 /* During recovery devices cannot be removed, so
3100 * locking and refcounting of rdevs is not needed
3102 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3103 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
3105 && !test_bit(Faulty
, &rdev
->flags
)
3106 && !test_bit(In_sync
, &rdev
->flags
)
3107 && !rdev_set_badblocks(rdev
, sh
->sector
,
3110 rdev
= conf
->disks
[i
].replacement
;
3112 && !test_bit(Faulty
, &rdev
->flags
)
3113 && !test_bit(In_sync
, &rdev
->flags
)
3114 && !rdev_set_badblocks(rdev
, sh
->sector
,
3119 conf
->recovery_disabled
=
3120 conf
->mddev
->recovery_disabled
;
3122 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3125 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3127 struct md_rdev
*rdev
;
3129 /* Doing recovery so rcu locking not required */
3130 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
3132 && !test_bit(Faulty
, &rdev
->flags
)
3133 && !test_bit(In_sync
, &rdev
->flags
)
3134 && (rdev
->recovery_offset
<= sh
->sector
3135 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3141 /* fetch_block - checks the given member device to see if its data needs
3142 * to be read or computed to satisfy a request.
3144 * Returns 1 when no more member devices need to be checked, otherwise returns
3145 * 0 to tell the loop in handle_stripe_fill to continue
3148 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3149 int disk_idx
, int disks
)
3151 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3152 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3153 &sh
->dev
[s
->failed_num
[1]] };
3157 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3158 test_bit(R5_UPTODATE
, &dev
->flags
))
3159 /* No point reading this as we already have it or have
3160 * decided to get it.
3165 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3166 /* We need this block to directly satisfy a request */
3169 if (s
->syncing
|| s
->expanding
||
3170 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3171 /* When syncing, or expanding we read everything.
3172 * When replacing, we need the replaced block.
3176 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3177 (s
->failed
>= 2 && fdev
[1]->toread
))
3178 /* If we want to read from a failed device, then
3179 * we need to actually read every other device.
3183 /* Sometimes neither read-modify-write nor reconstruct-write
3184 * cycles can work. In those cases we read every block we
3185 * can. Then the parity-update is certain to have enough to
3187 * This can only be a problem when we need to write something,
3188 * and some device has failed. If either of those tests
3189 * fail we need look no further.
3191 if (!s
->failed
|| !s
->to_write
)
3194 if (test_bit(R5_Insync
, &dev
->flags
) &&
3195 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3196 /* Pre-reads at not permitted until after short delay
3197 * to gather multiple requests. However if this
3198 * device is no Insync, the block could only be be computed
3199 * and there is no need to delay that.
3203 for (i
= 0; i
< s
->failed
; i
++) {
3204 if (fdev
[i
]->towrite
&&
3205 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3206 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3207 /* If we have a partial write to a failed
3208 * device, then we will need to reconstruct
3209 * the content of that device, so all other
3210 * devices must be read.
3215 /* If we are forced to do a reconstruct-write, either because
3216 * the current RAID6 implementation only supports that, or
3217 * or because parity cannot be trusted and we are currently
3218 * recovering it, there is extra need to be careful.
3219 * If one of the devices that we would need to read, because
3220 * it is not being overwritten (and maybe not written at all)
3221 * is missing/faulty, then we need to read everything we can.
3223 if (sh
->raid_conf
->level
!= 6 &&
3224 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3225 /* reconstruct-write isn't being forced */
3227 for (i
= 0; i
< s
->failed
; i
++) {
3228 if (!test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3229 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3236 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3237 int disk_idx
, int disks
)
3239 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3241 /* is the data in this block needed, and can we get it? */
3242 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3243 /* we would like to get this block, possibly by computing it,
3244 * otherwise read it if the backing disk is insync
3246 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3247 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3248 if ((s
->uptodate
== disks
- 1) &&
3249 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3250 disk_idx
== s
->failed_num
[1]))) {
3251 /* have disk failed, and we're requested to fetch it;
3254 pr_debug("Computing stripe %llu block %d\n",
3255 (unsigned long long)sh
->sector
, disk_idx
);
3256 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3257 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3258 set_bit(R5_Wantcompute
, &dev
->flags
);
3259 sh
->ops
.target
= disk_idx
;
3260 sh
->ops
.target2
= -1; /* no 2nd target */
3262 /* Careful: from this point on 'uptodate' is in the eye
3263 * of raid_run_ops which services 'compute' operations
3264 * before writes. R5_Wantcompute flags a block that will
3265 * be R5_UPTODATE by the time it is needed for a
3266 * subsequent operation.
3270 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3271 /* Computing 2-failure is *very* expensive; only
3272 * do it if failed >= 2
3275 for (other
= disks
; other
--; ) {
3276 if (other
== disk_idx
)
3278 if (!test_bit(R5_UPTODATE
,
3279 &sh
->dev
[other
].flags
))
3283 pr_debug("Computing stripe %llu blocks %d,%d\n",
3284 (unsigned long long)sh
->sector
,
3286 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3287 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3288 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3289 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3290 sh
->ops
.target
= disk_idx
;
3291 sh
->ops
.target2
= other
;
3295 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3296 set_bit(R5_LOCKED
, &dev
->flags
);
3297 set_bit(R5_Wantread
, &dev
->flags
);
3299 pr_debug("Reading block %d (sync=%d)\n",
3300 disk_idx
, s
->syncing
);
3308 * handle_stripe_fill - read or compute data to satisfy pending requests.
3310 static void handle_stripe_fill(struct stripe_head
*sh
,
3311 struct stripe_head_state
*s
,
3316 BUG_ON(sh
->batch_head
);
3317 /* look for blocks to read/compute, skip this if a compute
3318 * is already in flight, or if the stripe contents are in the
3319 * midst of changing due to a write
3321 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3322 !sh
->reconstruct_state
)
3323 for (i
= disks
; i
--; )
3324 if (fetch_block(sh
, s
, i
, disks
))
3326 set_bit(STRIPE_HANDLE
, &sh
->state
);
3329 /* handle_stripe_clean_event
3330 * any written block on an uptodate or failed drive can be returned.
3331 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3332 * never LOCKED, so we don't need to test 'failed' directly.
3334 static void handle_stripe_clean_event(struct r5conf
*conf
,
3335 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
3339 int discard_pending
= 0;
3340 struct stripe_head
*head_sh
= sh
;
3341 bool do_endio
= false;
3344 for (i
= disks
; i
--; )
3345 if (sh
->dev
[i
].written
) {
3347 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3348 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3349 test_bit(R5_Discard
, &dev
->flags
) ||
3350 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3351 /* We can return any write requests */
3352 struct bio
*wbi
, *wbi2
;
3353 pr_debug("Return write for disc %d\n", i
);
3354 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3355 clear_bit(R5_UPTODATE
, &dev
->flags
);
3356 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3357 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3362 dev
->page
= dev
->orig_page
;
3364 dev
->written
= NULL
;
3365 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3366 dev
->sector
+ STRIPE_SECTORS
) {
3367 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3368 if (!raid5_dec_bi_active_stripes(wbi
)) {
3369 md_write_end(conf
->mddev
);
3370 wbi
->bi_next
= *return_bi
;
3375 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3377 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3379 if (head_sh
->batch_head
) {
3380 sh
= list_first_entry(&sh
->batch_list
,
3383 if (sh
!= head_sh
) {
3390 } else if (test_bit(R5_Discard
, &dev
->flags
))
3391 discard_pending
= 1;
3392 WARN_ON(test_bit(R5_SkipCopy
, &dev
->flags
));
3393 WARN_ON(dev
->page
!= dev
->orig_page
);
3395 if (!discard_pending
&&
3396 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3397 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3398 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3399 if (sh
->qd_idx
>= 0) {
3400 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3401 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3403 /* now that discard is done we can proceed with any sync */
3404 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3406 * SCSI discard will change some bio fields and the stripe has
3407 * no updated data, so remove it from hash list and the stripe
3408 * will be reinitialized
3410 spin_lock_irq(&conf
->device_lock
);
3413 if (head_sh
->batch_head
) {
3414 sh
= list_first_entry(&sh
->batch_list
,
3415 struct stripe_head
, batch_list
);
3419 spin_unlock_irq(&conf
->device_lock
);
3422 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3423 set_bit(STRIPE_HANDLE
, &sh
->state
);
3427 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3428 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3429 md_wakeup_thread(conf
->mddev
->thread
);
3431 if (!head_sh
->batch_head
|| !do_endio
)
3433 for (i
= 0; i
< head_sh
->disks
; i
++) {
3434 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
3437 while (!list_empty(&head_sh
->batch_list
)) {
3439 sh
= list_first_entry(&head_sh
->batch_list
,
3440 struct stripe_head
, batch_list
);
3441 list_del_init(&sh
->batch_list
);
3443 set_mask_bits(&sh
->state
, ~STRIPE_EXPAND_SYNC_FLAG
,
3444 head_sh
->state
& ~((1 << STRIPE_ACTIVE
) |
3445 (1 << STRIPE_PREREAD_ACTIVE
) |
3446 STRIPE_EXPAND_SYNC_FLAG
));
3447 sh
->check_state
= head_sh
->check_state
;
3448 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
3449 for (i
= 0; i
< sh
->disks
; i
++) {
3450 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3452 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
;
3455 spin_lock_irq(&sh
->stripe_lock
);
3456 sh
->batch_head
= NULL
;
3457 spin_unlock_irq(&sh
->stripe_lock
);
3458 if (sh
->state
& STRIPE_EXPAND_SYNC_FLAG
)
3459 set_bit(STRIPE_HANDLE
, &sh
->state
);
3463 spin_lock_irq(&head_sh
->stripe_lock
);
3464 head_sh
->batch_head
= NULL
;
3465 spin_unlock_irq(&head_sh
->stripe_lock
);
3466 wake_up_nr(&conf
->wait_for_overlap
, wakeup_nr
);
3467 if (head_sh
->state
& STRIPE_EXPAND_SYNC_FLAG
)
3468 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
3471 static void handle_stripe_dirtying(struct r5conf
*conf
,
3472 struct stripe_head
*sh
,
3473 struct stripe_head_state
*s
,
3476 int rmw
= 0, rcw
= 0, i
;
3477 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3479 /* RAID6 requires 'rcw' in current implementation.
3480 * Otherwise, check whether resync is now happening or should start.
3481 * If yes, then the array is dirty (after unclean shutdown or
3482 * initial creation), so parity in some stripes might be inconsistent.
3483 * In this case, we need to always do reconstruct-write, to ensure
3484 * that in case of drive failure or read-error correction, we
3485 * generate correct data from the parity.
3487 if (conf
->max_degraded
== 2 ||
3488 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3490 /* Calculate the real rcw later - for now make it
3491 * look like rcw is cheaper
3494 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3495 conf
->max_degraded
, (unsigned long long)recovery_cp
,
3496 (unsigned long long)sh
->sector
);
3497 } else for (i
= disks
; i
--; ) {
3498 /* would I have to read this buffer for read_modify_write */
3499 struct r5dev
*dev
= &sh
->dev
[i
];
3500 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
3501 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3502 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3503 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3504 if (test_bit(R5_Insync
, &dev
->flags
))
3507 rmw
+= 2*disks
; /* cannot read it */
3509 /* Would I have to read this buffer for reconstruct_write */
3510 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
3511 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3512 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3513 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3514 if (test_bit(R5_Insync
, &dev
->flags
))
3520 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3521 (unsigned long long)sh
->sector
, rmw
, rcw
);
3522 set_bit(STRIPE_HANDLE
, &sh
->state
);
3523 if (rmw
< rcw
&& rmw
> 0) {
3524 /* prefer read-modify-write, but need to get some data */
3525 if (conf
->mddev
->queue
)
3526 blk_add_trace_msg(conf
->mddev
->queue
,
3527 "raid5 rmw %llu %d",
3528 (unsigned long long)sh
->sector
, rmw
);
3529 for (i
= disks
; i
--; ) {
3530 struct r5dev
*dev
= &sh
->dev
[i
];
3531 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
3532 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3533 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3534 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3535 test_bit(R5_Insync
, &dev
->flags
)) {
3536 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3538 pr_debug("Read_old block %d for r-m-w\n",
3540 set_bit(R5_LOCKED
, &dev
->flags
);
3541 set_bit(R5_Wantread
, &dev
->flags
);
3544 set_bit(STRIPE_DELAYED
, &sh
->state
);
3545 set_bit(STRIPE_HANDLE
, &sh
->state
);
3550 if (rcw
<= rmw
&& rcw
> 0) {
3551 /* want reconstruct write, but need to get some data */
3554 for (i
= disks
; i
--; ) {
3555 struct r5dev
*dev
= &sh
->dev
[i
];
3556 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3557 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3558 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3559 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3560 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3562 if (test_bit(R5_Insync
, &dev
->flags
) &&
3563 test_bit(STRIPE_PREREAD_ACTIVE
,
3565 pr_debug("Read_old block "
3566 "%d for Reconstruct\n", i
);
3567 set_bit(R5_LOCKED
, &dev
->flags
);
3568 set_bit(R5_Wantread
, &dev
->flags
);
3572 set_bit(STRIPE_DELAYED
, &sh
->state
);
3573 set_bit(STRIPE_HANDLE
, &sh
->state
);
3577 if (rcw
&& conf
->mddev
->queue
)
3578 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3579 (unsigned long long)sh
->sector
,
3580 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3583 if (rcw
> disks
&& rmw
> disks
&&
3584 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3585 set_bit(STRIPE_DELAYED
, &sh
->state
);
3587 /* now if nothing is locked, and if we have enough data,
3588 * we can start a write request
3590 /* since handle_stripe can be called at any time we need to handle the
3591 * case where a compute block operation has been submitted and then a
3592 * subsequent call wants to start a write request. raid_run_ops only
3593 * handles the case where compute block and reconstruct are requested
3594 * simultaneously. If this is not the case then new writes need to be
3595 * held off until the compute completes.
3597 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
3598 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
3599 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
3600 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
3603 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
3604 struct stripe_head_state
*s
, int disks
)
3606 struct r5dev
*dev
= NULL
;
3608 BUG_ON(sh
->batch_head
);
3609 set_bit(STRIPE_HANDLE
, &sh
->state
);
3611 switch (sh
->check_state
) {
3612 case check_state_idle
:
3613 /* start a new check operation if there are no failures */
3614 if (s
->failed
== 0) {
3615 BUG_ON(s
->uptodate
!= disks
);
3616 sh
->check_state
= check_state_run
;
3617 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3618 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3622 dev
= &sh
->dev
[s
->failed_num
[0]];
3624 case check_state_compute_result
:
3625 sh
->check_state
= check_state_idle
;
3627 dev
= &sh
->dev
[sh
->pd_idx
];
3629 /* check that a write has not made the stripe insync */
3630 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3633 /* either failed parity check, or recovery is happening */
3634 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3635 BUG_ON(s
->uptodate
!= disks
);
3637 set_bit(R5_LOCKED
, &dev
->flags
);
3639 set_bit(R5_Wantwrite
, &dev
->flags
);
3641 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3642 set_bit(STRIPE_INSYNC
, &sh
->state
);
3644 case check_state_run
:
3645 break; /* we will be called again upon completion */
3646 case check_state_check_result
:
3647 sh
->check_state
= check_state_idle
;
3649 /* if a failure occurred during the check operation, leave
3650 * STRIPE_INSYNC not set and let the stripe be handled again
3655 /* handle a successful check operation, if parity is correct
3656 * we are done. Otherwise update the mismatch count and repair
3657 * parity if !MD_RECOVERY_CHECK
3659 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3660 /* parity is correct (on disc,
3661 * not in buffer any more)
3663 set_bit(STRIPE_INSYNC
, &sh
->state
);
3665 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3666 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3667 /* don't try to repair!! */
3668 set_bit(STRIPE_INSYNC
, &sh
->state
);
3670 sh
->check_state
= check_state_compute_run
;
3671 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3672 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3673 set_bit(R5_Wantcompute
,
3674 &sh
->dev
[sh
->pd_idx
].flags
);
3675 sh
->ops
.target
= sh
->pd_idx
;
3676 sh
->ops
.target2
= -1;
3681 case check_state_compute_run
:
3684 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3685 __func__
, sh
->check_state
,
3686 (unsigned long long) sh
->sector
);
3691 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3692 struct stripe_head_state
*s
,
3695 int pd_idx
= sh
->pd_idx
;
3696 int qd_idx
= sh
->qd_idx
;
3699 BUG_ON(sh
->batch_head
);
3700 set_bit(STRIPE_HANDLE
, &sh
->state
);
3702 BUG_ON(s
->failed
> 2);
3704 /* Want to check and possibly repair P and Q.
3705 * However there could be one 'failed' device, in which
3706 * case we can only check one of them, possibly using the
3707 * other to generate missing data
3710 switch (sh
->check_state
) {
3711 case check_state_idle
:
3712 /* start a new check operation if there are < 2 failures */
3713 if (s
->failed
== s
->q_failed
) {
3714 /* The only possible failed device holds Q, so it
3715 * makes sense to check P (If anything else were failed,
3716 * we would have used P to recreate it).
3718 sh
->check_state
= check_state_run
;
3720 if (!s
->q_failed
&& s
->failed
< 2) {
3721 /* Q is not failed, and we didn't use it to generate
3722 * anything, so it makes sense to check it
3724 if (sh
->check_state
== check_state_run
)
3725 sh
->check_state
= check_state_run_pq
;
3727 sh
->check_state
= check_state_run_q
;
3730 /* discard potentially stale zero_sum_result */
3731 sh
->ops
.zero_sum_result
= 0;
3733 if (sh
->check_state
== check_state_run
) {
3734 /* async_xor_zero_sum destroys the contents of P */
3735 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3738 if (sh
->check_state
>= check_state_run
&&
3739 sh
->check_state
<= check_state_run_pq
) {
3740 /* async_syndrome_zero_sum preserves P and Q, so
3741 * no need to mark them !uptodate here
3743 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3747 /* we have 2-disk failure */
3748 BUG_ON(s
->failed
!= 2);
3750 case check_state_compute_result
:
3751 sh
->check_state
= check_state_idle
;
3753 /* check that a write has not made the stripe insync */
3754 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3757 /* now write out any block on a failed drive,
3758 * or P or Q if they were recomputed
3760 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3761 if (s
->failed
== 2) {
3762 dev
= &sh
->dev
[s
->failed_num
[1]];
3764 set_bit(R5_LOCKED
, &dev
->flags
);
3765 set_bit(R5_Wantwrite
, &dev
->flags
);
3767 if (s
->failed
>= 1) {
3768 dev
= &sh
->dev
[s
->failed_num
[0]];
3770 set_bit(R5_LOCKED
, &dev
->flags
);
3771 set_bit(R5_Wantwrite
, &dev
->flags
);
3773 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3774 dev
= &sh
->dev
[pd_idx
];
3776 set_bit(R5_LOCKED
, &dev
->flags
);
3777 set_bit(R5_Wantwrite
, &dev
->flags
);
3779 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3780 dev
= &sh
->dev
[qd_idx
];
3782 set_bit(R5_LOCKED
, &dev
->flags
);
3783 set_bit(R5_Wantwrite
, &dev
->flags
);
3785 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3787 set_bit(STRIPE_INSYNC
, &sh
->state
);
3789 case check_state_run
:
3790 case check_state_run_q
:
3791 case check_state_run_pq
:
3792 break; /* we will be called again upon completion */
3793 case check_state_check_result
:
3794 sh
->check_state
= check_state_idle
;
3796 /* handle a successful check operation, if parity is correct
3797 * we are done. Otherwise update the mismatch count and repair
3798 * parity if !MD_RECOVERY_CHECK
3800 if (sh
->ops
.zero_sum_result
== 0) {
3801 /* both parities are correct */
3803 set_bit(STRIPE_INSYNC
, &sh
->state
);
3805 /* in contrast to the raid5 case we can validate
3806 * parity, but still have a failure to write
3809 sh
->check_state
= check_state_compute_result
;
3810 /* Returning at this point means that we may go
3811 * off and bring p and/or q uptodate again so
3812 * we make sure to check zero_sum_result again
3813 * to verify if p or q need writeback
3817 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3818 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3819 /* don't try to repair!! */
3820 set_bit(STRIPE_INSYNC
, &sh
->state
);
3822 int *target
= &sh
->ops
.target
;
3824 sh
->ops
.target
= -1;
3825 sh
->ops
.target2
= -1;
3826 sh
->check_state
= check_state_compute_run
;
3827 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3828 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3829 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3830 set_bit(R5_Wantcompute
,
3831 &sh
->dev
[pd_idx
].flags
);
3833 target
= &sh
->ops
.target2
;
3836 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3837 set_bit(R5_Wantcompute
,
3838 &sh
->dev
[qd_idx
].flags
);
3845 case check_state_compute_run
:
3848 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3849 __func__
, sh
->check_state
,
3850 (unsigned long long) sh
->sector
);
3855 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3859 /* We have read all the blocks in this stripe and now we need to
3860 * copy some of them into a target stripe for expand.
3862 struct dma_async_tx_descriptor
*tx
= NULL
;
3863 BUG_ON(sh
->batch_head
);
3864 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3865 for (i
= 0; i
< sh
->disks
; i
++)
3866 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3868 struct stripe_head
*sh2
;
3869 struct async_submit_ctl submit
;
3871 sector_t bn
= compute_blocknr(sh
, i
, 1);
3872 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3874 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3876 /* so far only the early blocks of this stripe
3877 * have been requested. When later blocks
3878 * get requested, we will try again
3881 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3882 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3883 /* must have already done this block */
3884 release_stripe(sh2
);
3888 /* place all the copies on one channel */
3889 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3890 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3891 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3894 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3895 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3896 for (j
= 0; j
< conf
->raid_disks
; j
++)
3897 if (j
!= sh2
->pd_idx
&&
3899 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3901 if (j
== conf
->raid_disks
) {
3902 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3903 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3905 release_stripe(sh2
);
3908 /* done submitting copies, wait for them to complete */
3909 async_tx_quiesce(&tx
);
3913 * handle_stripe - do things to a stripe.
3915 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3916 * state of various bits to see what needs to be done.
3918 * return some read requests which now have data
3919 * return some write requests which are safely on storage
3920 * schedule a read on some buffers
3921 * schedule a write of some buffers
3922 * return confirmation of parity correctness
3926 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3928 struct r5conf
*conf
= sh
->raid_conf
;
3929 int disks
= sh
->disks
;
3932 int do_recovery
= 0;
3934 memset(s
, 0, sizeof(*s
));
3936 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
3937 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
3938 s
->failed_num
[0] = -1;
3939 s
->failed_num
[1] = -1;
3941 /* Now to look around and see what can be done */
3943 for (i
=disks
; i
--; ) {
3944 struct md_rdev
*rdev
;
3951 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3953 dev
->toread
, dev
->towrite
, dev
->written
);
3954 /* maybe we can reply to a read
3956 * new wantfill requests are only permitted while
3957 * ops_complete_biofill is guaranteed to be inactive
3959 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3960 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3961 set_bit(R5_Wantfill
, &dev
->flags
);
3963 /* now count some things */
3964 if (test_bit(R5_LOCKED
, &dev
->flags
))
3966 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3968 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3970 BUG_ON(s
->compute
> 2);
3973 if (test_bit(R5_Wantfill
, &dev
->flags
))
3975 else if (dev
->toread
)
3979 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3984 /* Prefer to use the replacement for reads, but only
3985 * if it is recovered enough and has no bad blocks.
3987 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3988 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3989 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3990 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3991 &first_bad
, &bad_sectors
))
3992 set_bit(R5_ReadRepl
, &dev
->flags
);
3995 set_bit(R5_NeedReplace
, &dev
->flags
);
3996 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3997 clear_bit(R5_ReadRepl
, &dev
->flags
);
3999 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4002 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4003 &first_bad
, &bad_sectors
);
4004 if (s
->blocked_rdev
== NULL
4005 && (test_bit(Blocked
, &rdev
->flags
)
4008 set_bit(BlockedBadBlocks
,
4010 s
->blocked_rdev
= rdev
;
4011 atomic_inc(&rdev
->nr_pending
);
4014 clear_bit(R5_Insync
, &dev
->flags
);
4018 /* also not in-sync */
4019 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4020 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4021 /* treat as in-sync, but with a read error
4022 * which we can now try to correct
4024 set_bit(R5_Insync
, &dev
->flags
);
4025 set_bit(R5_ReadError
, &dev
->flags
);
4027 } else if (test_bit(In_sync
, &rdev
->flags
))
4028 set_bit(R5_Insync
, &dev
->flags
);
4029 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4030 /* in sync if before recovery_offset */
4031 set_bit(R5_Insync
, &dev
->flags
);
4032 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4033 test_bit(R5_Expanded
, &dev
->flags
))
4034 /* If we've reshaped into here, we assume it is Insync.
4035 * We will shortly update recovery_offset to make
4038 set_bit(R5_Insync
, &dev
->flags
);
4040 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4041 /* This flag does not apply to '.replacement'
4042 * only to .rdev, so make sure to check that*/
4043 struct md_rdev
*rdev2
= rcu_dereference(
4044 conf
->disks
[i
].rdev
);
4046 clear_bit(R5_Insync
, &dev
->flags
);
4047 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4048 s
->handle_bad_blocks
= 1;
4049 atomic_inc(&rdev2
->nr_pending
);
4051 clear_bit(R5_WriteError
, &dev
->flags
);
4053 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4054 /* This flag does not apply to '.replacement'
4055 * only to .rdev, so make sure to check that*/
4056 struct md_rdev
*rdev2
= rcu_dereference(
4057 conf
->disks
[i
].rdev
);
4058 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4059 s
->handle_bad_blocks
= 1;
4060 atomic_inc(&rdev2
->nr_pending
);
4062 clear_bit(R5_MadeGood
, &dev
->flags
);
4064 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4065 struct md_rdev
*rdev2
= rcu_dereference(
4066 conf
->disks
[i
].replacement
);
4067 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4068 s
->handle_bad_blocks
= 1;
4069 atomic_inc(&rdev2
->nr_pending
);
4071 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4073 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4074 /* The ReadError flag will just be confusing now */
4075 clear_bit(R5_ReadError
, &dev
->flags
);
4076 clear_bit(R5_ReWrite
, &dev
->flags
);
4078 if (test_bit(R5_ReadError
, &dev
->flags
))
4079 clear_bit(R5_Insync
, &dev
->flags
);
4080 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4082 s
->failed_num
[s
->failed
] = i
;
4084 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4088 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4089 /* If there is a failed device being replaced,
4090 * we must be recovering.
4091 * else if we are after recovery_cp, we must be syncing
4092 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4093 * else we can only be replacing
4094 * sync and recovery both need to read all devices, and so
4095 * use the same flag.
4098 sh
->sector
>= conf
->mddev
->recovery_cp
||
4099 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4107 static int clear_batch_ready(struct stripe_head
*sh
)
4109 struct stripe_head
*tmp
;
4110 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4112 spin_lock(&sh
->stripe_lock
);
4113 if (!sh
->batch_head
) {
4114 spin_unlock(&sh
->stripe_lock
);
4119 * this stripe could be added to a batch list before we check
4120 * BATCH_READY, skips it
4122 if (sh
->batch_head
!= sh
) {
4123 spin_unlock(&sh
->stripe_lock
);
4126 spin_lock(&sh
->batch_lock
);
4127 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4128 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4129 spin_unlock(&sh
->batch_lock
);
4130 spin_unlock(&sh
->stripe_lock
);
4133 * BATCH_READY is cleared, no new stripes can be added.
4134 * batch_list can be accessed without lock
4139 static void check_break_stripe_batch_list(struct stripe_head
*sh
)
4141 struct stripe_head
*head_sh
, *next
;
4144 if (!test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4149 sh
= list_first_entry(&sh
->batch_list
,
4150 struct stripe_head
, batch_list
);
4151 BUG_ON(sh
== head_sh
);
4152 } while (!test_bit(STRIPE_DEGRADED
, &sh
->state
));
4154 while (sh
!= head_sh
) {
4155 next
= list_first_entry(&sh
->batch_list
,
4156 struct stripe_head
, batch_list
);
4157 list_del_init(&sh
->batch_list
);
4159 set_mask_bits(&sh
->state
, ~STRIPE_EXPAND_SYNC_FLAG
,
4160 head_sh
->state
& ~((1 << STRIPE_ACTIVE
) |
4161 (1 << STRIPE_PREREAD_ACTIVE
) |
4162 (1 << STRIPE_DEGRADED
) |
4163 STRIPE_EXPAND_SYNC_FLAG
));
4164 sh
->check_state
= head_sh
->check_state
;
4165 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4166 for (i
= 0; i
< sh
->disks
; i
++)
4167 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4168 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4170 spin_lock_irq(&sh
->stripe_lock
);
4171 sh
->batch_head
= NULL
;
4172 spin_unlock_irq(&sh
->stripe_lock
);
4174 set_bit(STRIPE_HANDLE
, &sh
->state
);
4181 static void handle_stripe(struct stripe_head
*sh
)
4183 struct stripe_head_state s
;
4184 struct r5conf
*conf
= sh
->raid_conf
;
4187 int disks
= sh
->disks
;
4188 struct r5dev
*pdev
, *qdev
;
4190 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4191 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4192 /* already being handled, ensure it gets handled
4193 * again when current action finishes */
4194 set_bit(STRIPE_HANDLE
, &sh
->state
);
4198 if (clear_batch_ready(sh
) ) {
4199 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4203 check_break_stripe_batch_list(sh
);
4205 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4206 spin_lock(&sh
->stripe_lock
);
4207 /* Cannot process 'sync' concurrently with 'discard' */
4208 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4209 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4210 set_bit(STRIPE_SYNCING
, &sh
->state
);
4211 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4212 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4214 spin_unlock(&sh
->stripe_lock
);
4216 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4218 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4219 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4220 (unsigned long long)sh
->sector
, sh
->state
,
4221 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4222 sh
->check_state
, sh
->reconstruct_state
);
4224 analyse_stripe(sh
, &s
);
4226 if (s
.handle_bad_blocks
) {
4227 set_bit(STRIPE_HANDLE
, &sh
->state
);
4231 if (unlikely(s
.blocked_rdev
)) {
4232 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4233 s
.replacing
|| s
.to_write
|| s
.written
) {
4234 set_bit(STRIPE_HANDLE
, &sh
->state
);
4237 /* There is nothing for the blocked_rdev to block */
4238 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4239 s
.blocked_rdev
= NULL
;
4242 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4243 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4244 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4247 pr_debug("locked=%d uptodate=%d to_read=%d"
4248 " to_write=%d failed=%d failed_num=%d,%d\n",
4249 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4250 s
.failed_num
[0], s
.failed_num
[1]);
4251 /* check if the array has lost more than max_degraded devices and,
4252 * if so, some requests might need to be failed.
4254 if (s
.failed
> conf
->max_degraded
) {
4255 sh
->check_state
= 0;
4256 sh
->reconstruct_state
= 0;
4257 if (s
.to_read
+s
.to_write
+s
.written
)
4258 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
4259 if (s
.syncing
+ s
.replacing
)
4260 handle_failed_sync(conf
, sh
, &s
);
4263 /* Now we check to see if any write operations have recently
4267 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4269 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4270 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4271 sh
->reconstruct_state
= reconstruct_state_idle
;
4273 /* All the 'written' buffers and the parity block are ready to
4274 * be written back to disk
4276 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4277 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4278 BUG_ON(sh
->qd_idx
>= 0 &&
4279 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4280 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4281 for (i
= disks
; i
--; ) {
4282 struct r5dev
*dev
= &sh
->dev
[i
];
4283 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4284 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4286 pr_debug("Writing block %d\n", i
);
4287 set_bit(R5_Wantwrite
, &dev
->flags
);
4292 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4293 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4295 set_bit(STRIPE_INSYNC
, &sh
->state
);
4298 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4299 s
.dec_preread_active
= 1;
4303 * might be able to return some write requests if the parity blocks
4304 * are safe, or on a failed drive
4306 pdev
= &sh
->dev
[sh
->pd_idx
];
4307 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4308 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4309 qdev
= &sh
->dev
[sh
->qd_idx
];
4310 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4311 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4315 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4316 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4317 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4318 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4319 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4320 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4321 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4322 test_bit(R5_Discard
, &qdev
->flags
))))))
4323 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
4325 /* Now we might consider reading some blocks, either to check/generate
4326 * parity, or to satisfy requests
4327 * or to load a block that is being partially written.
4329 if (s
.to_read
|| s
.non_overwrite
4330 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4331 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4334 handle_stripe_fill(sh
, &s
, disks
);
4336 /* Now to consider new write requests and what else, if anything
4337 * should be read. We do not handle new writes when:
4338 * 1/ A 'write' operation (copy+xor) is already in flight.
4339 * 2/ A 'check' operation is in flight, as it may clobber the parity
4342 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
4343 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4345 /* maybe we need to check and possibly fix the parity for this stripe
4346 * Any reads will already have been scheduled, so we just see if enough
4347 * data is available. The parity check is held off while parity
4348 * dependent operations are in flight.
4350 if (sh
->check_state
||
4351 (s
.syncing
&& s
.locked
== 0 &&
4352 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4353 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4354 if (conf
->level
== 6)
4355 handle_parity_checks6(conf
, sh
, &s
, disks
);
4357 handle_parity_checks5(conf
, sh
, &s
, disks
);
4360 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4361 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4362 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4363 /* Write out to replacement devices where possible */
4364 for (i
= 0; i
< conf
->raid_disks
; i
++)
4365 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4366 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4367 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4368 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4372 set_bit(STRIPE_INSYNC
, &sh
->state
);
4373 set_bit(STRIPE_REPLACED
, &sh
->state
);
4375 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4376 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4377 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4378 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4379 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4380 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4381 wake_up(&conf
->wait_for_overlap
);
4384 /* If the failed drives are just a ReadError, then we might need
4385 * to progress the repair/check process
4387 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4388 for (i
= 0; i
< s
.failed
; i
++) {
4389 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4390 if (test_bit(R5_ReadError
, &dev
->flags
)
4391 && !test_bit(R5_LOCKED
, &dev
->flags
)
4392 && test_bit(R5_UPTODATE
, &dev
->flags
)
4394 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4395 set_bit(R5_Wantwrite
, &dev
->flags
);
4396 set_bit(R5_ReWrite
, &dev
->flags
);
4397 set_bit(R5_LOCKED
, &dev
->flags
);
4400 /* let's read it back */
4401 set_bit(R5_Wantread
, &dev
->flags
);
4402 set_bit(R5_LOCKED
, &dev
->flags
);
4408 /* Finish reconstruct operations initiated by the expansion process */
4409 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4410 struct stripe_head
*sh_src
4411 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4412 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4413 /* sh cannot be written until sh_src has been read.
4414 * so arrange for sh to be delayed a little
4416 set_bit(STRIPE_DELAYED
, &sh
->state
);
4417 set_bit(STRIPE_HANDLE
, &sh
->state
);
4418 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4420 atomic_inc(&conf
->preread_active_stripes
);
4421 release_stripe(sh_src
);
4425 release_stripe(sh_src
);
4427 sh
->reconstruct_state
= reconstruct_state_idle
;
4428 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4429 for (i
= conf
->raid_disks
; i
--; ) {
4430 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4431 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4436 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4437 !sh
->reconstruct_state
) {
4438 /* Need to write out all blocks after computing parity */
4439 sh
->disks
= conf
->raid_disks
;
4440 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4441 schedule_reconstruction(sh
, &s
, 1, 1);
4442 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4443 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4444 atomic_dec(&conf
->reshape_stripes
);
4445 wake_up(&conf
->wait_for_overlap
);
4446 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4449 if (s
.expanding
&& s
.locked
== 0 &&
4450 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4451 handle_stripe_expansion(conf
, sh
);
4454 /* wait for this device to become unblocked */
4455 if (unlikely(s
.blocked_rdev
)) {
4456 if (conf
->mddev
->external
)
4457 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4460 /* Internal metadata will immediately
4461 * be written by raid5d, so we don't
4462 * need to wait here.
4464 rdev_dec_pending(s
.blocked_rdev
,
4468 if (s
.handle_bad_blocks
)
4469 for (i
= disks
; i
--; ) {
4470 struct md_rdev
*rdev
;
4471 struct r5dev
*dev
= &sh
->dev
[i
];
4472 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4473 /* We own a safe reference to the rdev */
4474 rdev
= conf
->disks
[i
].rdev
;
4475 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4477 md_error(conf
->mddev
, rdev
);
4478 rdev_dec_pending(rdev
, conf
->mddev
);
4480 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4481 rdev
= conf
->disks
[i
].rdev
;
4482 rdev_clear_badblocks(rdev
, sh
->sector
,
4484 rdev_dec_pending(rdev
, conf
->mddev
);
4486 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4487 rdev
= conf
->disks
[i
].replacement
;
4489 /* rdev have been moved down */
4490 rdev
= conf
->disks
[i
].rdev
;
4491 rdev_clear_badblocks(rdev
, sh
->sector
,
4493 rdev_dec_pending(rdev
, conf
->mddev
);
4498 raid_run_ops(sh
, s
.ops_request
);
4502 if (s
.dec_preread_active
) {
4503 /* We delay this until after ops_run_io so that if make_request
4504 * is waiting on a flush, it won't continue until the writes
4505 * have actually been submitted.
4507 atomic_dec(&conf
->preread_active_stripes
);
4508 if (atomic_read(&conf
->preread_active_stripes
) <
4510 md_wakeup_thread(conf
->mddev
->thread
);
4513 return_io(s
.return_bi
);
4515 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4518 static void raid5_activate_delayed(struct r5conf
*conf
)
4520 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
4521 while (!list_empty(&conf
->delayed_list
)) {
4522 struct list_head
*l
= conf
->delayed_list
.next
;
4523 struct stripe_head
*sh
;
4524 sh
= list_entry(l
, struct stripe_head
, lru
);
4526 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4527 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4528 atomic_inc(&conf
->preread_active_stripes
);
4529 list_add_tail(&sh
->lru
, &conf
->hold_list
);
4530 raid5_wakeup_stripe_thread(sh
);
4535 static void activate_bit_delay(struct r5conf
*conf
,
4536 struct list_head
*temp_inactive_list
)
4538 /* device_lock is held */
4539 struct list_head head
;
4540 list_add(&head
, &conf
->bitmap_list
);
4541 list_del_init(&conf
->bitmap_list
);
4542 while (!list_empty(&head
)) {
4543 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
4545 list_del_init(&sh
->lru
);
4546 atomic_inc(&sh
->count
);
4547 hash
= sh
->hash_lock_index
;
4548 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
4552 static int raid5_congested(struct mddev
*mddev
, int bits
)
4554 struct r5conf
*conf
= mddev
->private;
4556 /* No difference between reads and writes. Just check
4557 * how busy the stripe_cache is
4560 if (conf
->inactive_blocked
)
4564 if (atomic_read(&conf
->empty_inactive_list_nr
))
4570 /* We want read requests to align with chunks where possible,
4571 * but write requests don't need to.
4573 static int raid5_mergeable_bvec(struct mddev
*mddev
,
4574 struct bvec_merge_data
*bvm
,
4575 struct bio_vec
*biovec
)
4577 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
4579 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4580 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
4582 if ((bvm
->bi_rw
& 1) == WRITE
)
4583 return biovec
->bv_len
; /* always allow writes to be mergeable */
4585 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4586 chunk_sectors
= mddev
->new_chunk_sectors
;
4587 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
4588 if (max
< 0) max
= 0;
4589 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
4590 return biovec
->bv_len
;
4595 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
4597 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
4598 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4599 unsigned int bio_sectors
= bio_sectors(bio
);
4601 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4602 chunk_sectors
= mddev
->new_chunk_sectors
;
4603 return chunk_sectors
>=
4604 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
4608 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4609 * later sampled by raid5d.
4611 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
4613 unsigned long flags
;
4615 spin_lock_irqsave(&conf
->device_lock
, flags
);
4617 bi
->bi_next
= conf
->retry_read_aligned_list
;
4618 conf
->retry_read_aligned_list
= bi
;
4620 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4621 md_wakeup_thread(conf
->mddev
->thread
);
4624 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
4628 bi
= conf
->retry_read_aligned
;
4630 conf
->retry_read_aligned
= NULL
;
4633 bi
= conf
->retry_read_aligned_list
;
4635 conf
->retry_read_aligned_list
= bi
->bi_next
;
4638 * this sets the active strip count to 1 and the processed
4639 * strip count to zero (upper 8 bits)
4641 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
4648 * The "raid5_align_endio" should check if the read succeeded and if it
4649 * did, call bio_endio on the original bio (having bio_put the new bio
4651 * If the read failed..
4653 static void raid5_align_endio(struct bio
*bi
, int error
)
4655 struct bio
* raid_bi
= bi
->bi_private
;
4656 struct mddev
*mddev
;
4657 struct r5conf
*conf
;
4658 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4659 struct md_rdev
*rdev
;
4663 rdev
= (void*)raid_bi
->bi_next
;
4664 raid_bi
->bi_next
= NULL
;
4665 mddev
= rdev
->mddev
;
4666 conf
= mddev
->private;
4668 rdev_dec_pending(rdev
, conf
->mddev
);
4670 if (!error
&& uptodate
) {
4671 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
4673 bio_endio(raid_bi
, 0);
4674 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4675 wake_up(&conf
->wait_for_stripe
);
4679 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4681 add_bio_to_retry(raid_bi
, conf
);
4684 static int bio_fits_rdev(struct bio
*bi
)
4686 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
4688 if (bio_sectors(bi
) > queue_max_sectors(q
))
4690 blk_recount_segments(q
, bi
);
4691 if (bi
->bi_phys_segments
> queue_max_segments(q
))
4694 if (q
->merge_bvec_fn
)
4695 /* it's too hard to apply the merge_bvec_fn at this stage,
4703 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
4705 struct r5conf
*conf
= mddev
->private;
4707 struct bio
* align_bi
;
4708 struct md_rdev
*rdev
;
4709 sector_t end_sector
;
4711 if (!in_chunk_boundary(mddev
, raid_bio
)) {
4712 pr_debug("chunk_aligned_read : non aligned\n");
4716 * use bio_clone_mddev to make a copy of the bio
4718 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
4722 * set bi_end_io to a new function, and set bi_private to the
4725 align_bi
->bi_end_io
= raid5_align_endio
;
4726 align_bi
->bi_private
= raid_bio
;
4730 align_bi
->bi_iter
.bi_sector
=
4731 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
4734 end_sector
= bio_end_sector(align_bi
);
4736 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
4737 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
4738 rdev
->recovery_offset
< end_sector
) {
4739 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
4741 (test_bit(Faulty
, &rdev
->flags
) ||
4742 !(test_bit(In_sync
, &rdev
->flags
) ||
4743 rdev
->recovery_offset
>= end_sector
)))
4750 atomic_inc(&rdev
->nr_pending
);
4752 raid_bio
->bi_next
= (void*)rdev
;
4753 align_bi
->bi_bdev
= rdev
->bdev
;
4754 __clear_bit(BIO_SEG_VALID
, &align_bi
->bi_flags
);
4756 if (!bio_fits_rdev(align_bi
) ||
4757 is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
4758 bio_sectors(align_bi
),
4759 &first_bad
, &bad_sectors
)) {
4760 /* too big in some way, or has a known bad block */
4762 rdev_dec_pending(rdev
, mddev
);
4766 /* No reshape active, so we can trust rdev->data_offset */
4767 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
4769 spin_lock_irq(&conf
->device_lock
);
4770 wait_event_lock_irq(conf
->wait_for_stripe
,
4773 atomic_inc(&conf
->active_aligned_reads
);
4774 spin_unlock_irq(&conf
->device_lock
);
4777 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4778 align_bi
, disk_devt(mddev
->gendisk
),
4779 raid_bio
->bi_iter
.bi_sector
);
4780 generic_make_request(align_bi
);
4789 /* __get_priority_stripe - get the next stripe to process
4791 * Full stripe writes are allowed to pass preread active stripes up until
4792 * the bypass_threshold is exceeded. In general the bypass_count
4793 * increments when the handle_list is handled before the hold_list; however, it
4794 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4795 * stripe with in flight i/o. The bypass_count will be reset when the
4796 * head of the hold_list has changed, i.e. the head was promoted to the
4799 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
4801 struct stripe_head
*sh
= NULL
, *tmp
;
4802 struct list_head
*handle_list
= NULL
;
4803 struct r5worker_group
*wg
= NULL
;
4805 if (conf
->worker_cnt_per_group
== 0) {
4806 handle_list
= &conf
->handle_list
;
4807 } else if (group
!= ANY_GROUP
) {
4808 handle_list
= &conf
->worker_groups
[group
].handle_list
;
4809 wg
= &conf
->worker_groups
[group
];
4812 for (i
= 0; i
< conf
->group_cnt
; i
++) {
4813 handle_list
= &conf
->worker_groups
[i
].handle_list
;
4814 wg
= &conf
->worker_groups
[i
];
4815 if (!list_empty(handle_list
))
4820 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4822 list_empty(handle_list
) ? "empty" : "busy",
4823 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4824 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4826 if (!list_empty(handle_list
)) {
4827 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
4829 if (list_empty(&conf
->hold_list
))
4830 conf
->bypass_count
= 0;
4831 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4832 if (conf
->hold_list
.next
== conf
->last_hold
)
4833 conf
->bypass_count
++;
4835 conf
->last_hold
= conf
->hold_list
.next
;
4836 conf
->bypass_count
-= conf
->bypass_threshold
;
4837 if (conf
->bypass_count
< 0)
4838 conf
->bypass_count
= 0;
4841 } else if (!list_empty(&conf
->hold_list
) &&
4842 ((conf
->bypass_threshold
&&
4843 conf
->bypass_count
> conf
->bypass_threshold
) ||
4844 atomic_read(&conf
->pending_full_writes
) == 0)) {
4846 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
4847 if (conf
->worker_cnt_per_group
== 0 ||
4848 group
== ANY_GROUP
||
4849 !cpu_online(tmp
->cpu
) ||
4850 cpu_to_group(tmp
->cpu
) == group
) {
4857 conf
->bypass_count
-= conf
->bypass_threshold
;
4858 if (conf
->bypass_count
< 0)
4859 conf
->bypass_count
= 0;
4871 list_del_init(&sh
->lru
);
4872 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
4876 struct raid5_plug_cb
{
4877 struct blk_plug_cb cb
;
4878 struct list_head list
;
4879 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
4882 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4884 struct raid5_plug_cb
*cb
= container_of(
4885 blk_cb
, struct raid5_plug_cb
, cb
);
4886 struct stripe_head
*sh
;
4887 struct mddev
*mddev
= cb
->cb
.data
;
4888 struct r5conf
*conf
= mddev
->private;
4892 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4893 spin_lock_irq(&conf
->device_lock
);
4894 while (!list_empty(&cb
->list
)) {
4895 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4896 list_del_init(&sh
->lru
);
4898 * avoid race release_stripe_plug() sees
4899 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4900 * is still in our list
4902 smp_mb__before_atomic();
4903 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4905 * STRIPE_ON_RELEASE_LIST could be set here. In that
4906 * case, the count is always > 1 here
4908 hash
= sh
->hash_lock_index
;
4909 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
4912 spin_unlock_irq(&conf
->device_lock
);
4914 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
4915 NR_STRIPE_HASH_LOCKS
);
4917 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
4921 static void release_stripe_plug(struct mddev
*mddev
,
4922 struct stripe_head
*sh
)
4924 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4925 raid5_unplug
, mddev
,
4926 sizeof(struct raid5_plug_cb
));
4927 struct raid5_plug_cb
*cb
;
4934 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4936 if (cb
->list
.next
== NULL
) {
4938 INIT_LIST_HEAD(&cb
->list
);
4939 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
4940 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
4943 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
4944 list_add_tail(&sh
->lru
, &cb
->list
);
4949 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
4951 struct r5conf
*conf
= mddev
->private;
4952 sector_t logical_sector
, last_sector
;
4953 struct stripe_head
*sh
;
4957 if (mddev
->reshape_position
!= MaxSector
)
4958 /* Skip discard while reshape is happening */
4961 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4962 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
4965 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4967 stripe_sectors
= conf
->chunk_sectors
*
4968 (conf
->raid_disks
- conf
->max_degraded
);
4969 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
4971 sector_div(last_sector
, stripe_sectors
);
4973 logical_sector
*= conf
->chunk_sectors
;
4974 last_sector
*= conf
->chunk_sectors
;
4976 for (; logical_sector
< last_sector
;
4977 logical_sector
+= STRIPE_SECTORS
) {
4981 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
4982 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
4983 TASK_UNINTERRUPTIBLE
);
4984 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4985 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4990 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4991 spin_lock_irq(&sh
->stripe_lock
);
4992 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4993 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4995 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
4996 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
4997 spin_unlock_irq(&sh
->stripe_lock
);
5003 set_bit(STRIPE_DISCARD
, &sh
->state
);
5004 finish_wait(&conf
->wait_for_overlap
, &w
);
5005 sh
->overwrite_disks
= 0;
5006 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5007 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5009 sh
->dev
[d
].towrite
= bi
;
5010 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5011 raid5_inc_bi_active_stripes(bi
);
5012 sh
->overwrite_disks
++;
5014 spin_unlock_irq(&sh
->stripe_lock
);
5015 if (conf
->mddev
->bitmap
) {
5017 d
< conf
->raid_disks
- conf
->max_degraded
;
5019 bitmap_startwrite(mddev
->bitmap
,
5023 sh
->bm_seq
= conf
->seq_flush
+ 1;
5024 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5027 set_bit(STRIPE_HANDLE
, &sh
->state
);
5028 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5029 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5030 atomic_inc(&conf
->preread_active_stripes
);
5031 release_stripe_plug(mddev
, sh
);
5034 remaining
= raid5_dec_bi_active_stripes(bi
);
5035 if (remaining
== 0) {
5036 md_write_end(mddev
);
5041 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
5043 struct r5conf
*conf
= mddev
->private;
5045 sector_t new_sector
;
5046 sector_t logical_sector
, last_sector
;
5047 struct stripe_head
*sh
;
5048 const int rw
= bio_data_dir(bi
);
5053 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
5054 md_flush_request(mddev
, bi
);
5058 md_write_start(mddev
, bi
);
5061 mddev
->reshape_position
== MaxSector
&&
5062 chunk_aligned_read(mddev
,bi
))
5065 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
5066 make_discard_request(mddev
, bi
);
5070 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5071 last_sector
= bio_end_sector(bi
);
5073 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5075 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5076 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5082 seq
= read_seqcount_begin(&conf
->gen_lock
);
5085 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5086 TASK_UNINTERRUPTIBLE
);
5087 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5088 /* spinlock is needed as reshape_progress may be
5089 * 64bit on a 32bit platform, and so it might be
5090 * possible to see a half-updated value
5091 * Of course reshape_progress could change after
5092 * the lock is dropped, so once we get a reference
5093 * to the stripe that we think it is, we will have
5096 spin_lock_irq(&conf
->device_lock
);
5097 if (mddev
->reshape_backwards
5098 ? logical_sector
< conf
->reshape_progress
5099 : logical_sector
>= conf
->reshape_progress
) {
5102 if (mddev
->reshape_backwards
5103 ? logical_sector
< conf
->reshape_safe
5104 : logical_sector
>= conf
->reshape_safe
) {
5105 spin_unlock_irq(&conf
->device_lock
);
5111 spin_unlock_irq(&conf
->device_lock
);
5114 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5117 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5118 (unsigned long long)new_sector
,
5119 (unsigned long long)logical_sector
);
5121 sh
= get_active_stripe(conf
, new_sector
, previous
,
5122 (bi
->bi_rw
&RWA_MASK
), 0);
5124 if (unlikely(previous
)) {
5125 /* expansion might have moved on while waiting for a
5126 * stripe, so we must do the range check again.
5127 * Expansion could still move past after this
5128 * test, but as we are holding a reference to
5129 * 'sh', we know that if that happens,
5130 * STRIPE_EXPANDING will get set and the expansion
5131 * won't proceed until we finish with the stripe.
5134 spin_lock_irq(&conf
->device_lock
);
5135 if (mddev
->reshape_backwards
5136 ? logical_sector
>= conf
->reshape_progress
5137 : logical_sector
< conf
->reshape_progress
)
5138 /* mismatch, need to try again */
5140 spin_unlock_irq(&conf
->device_lock
);
5148 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5149 /* Might have got the wrong stripe_head
5157 logical_sector
>= mddev
->suspend_lo
&&
5158 logical_sector
< mddev
->suspend_hi
) {
5160 /* As the suspend_* range is controlled by
5161 * userspace, we want an interruptible
5164 flush_signals(current
);
5165 prepare_to_wait(&conf
->wait_for_overlap
,
5166 &w
, TASK_INTERRUPTIBLE
);
5167 if (logical_sector
>= mddev
->suspend_lo
&&
5168 logical_sector
< mddev
->suspend_hi
) {
5175 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5176 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5177 /* Stripe is busy expanding or
5178 * add failed due to overlap. Flush everything
5181 md_wakeup_thread(mddev
->thread
);
5187 set_bit(STRIPE_HANDLE
, &sh
->state
);
5188 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5189 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5190 (bi
->bi_rw
& REQ_SYNC
) &&
5191 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5192 atomic_inc(&conf
->preread_active_stripes
);
5193 release_stripe_plug(mddev
, sh
);
5195 /* cannot get stripe for read-ahead, just give-up */
5196 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
5200 finish_wait(&conf
->wait_for_overlap
, &w
);
5202 remaining
= raid5_dec_bi_active_stripes(bi
);
5203 if (remaining
== 0) {
5206 md_write_end(mddev
);
5208 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
5214 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5216 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5218 /* reshaping is quite different to recovery/resync so it is
5219 * handled quite separately ... here.
5221 * On each call to sync_request, we gather one chunk worth of
5222 * destination stripes and flag them as expanding.
5223 * Then we find all the source stripes and request reads.
5224 * As the reads complete, handle_stripe will copy the data
5225 * into the destination stripe and release that stripe.
5227 struct r5conf
*conf
= mddev
->private;
5228 struct stripe_head
*sh
;
5229 sector_t first_sector
, last_sector
;
5230 int raid_disks
= conf
->previous_raid_disks
;
5231 int data_disks
= raid_disks
- conf
->max_degraded
;
5232 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5235 sector_t writepos
, readpos
, safepos
;
5236 sector_t stripe_addr
;
5237 int reshape_sectors
;
5238 struct list_head stripes
;
5240 if (sector_nr
== 0) {
5241 /* If restarting in the middle, skip the initial sectors */
5242 if (mddev
->reshape_backwards
&&
5243 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5244 sector_nr
= raid5_size(mddev
, 0, 0)
5245 - conf
->reshape_progress
;
5246 } else if (!mddev
->reshape_backwards
&&
5247 conf
->reshape_progress
> 0)
5248 sector_nr
= conf
->reshape_progress
;
5249 sector_div(sector_nr
, new_data_disks
);
5251 mddev
->curr_resync_completed
= sector_nr
;
5252 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5258 /* We need to process a full chunk at a time.
5259 * If old and new chunk sizes differ, we need to process the
5262 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
5263 reshape_sectors
= mddev
->new_chunk_sectors
;
5265 reshape_sectors
= mddev
->chunk_sectors
;
5267 /* We update the metadata at least every 10 seconds, or when
5268 * the data about to be copied would over-write the source of
5269 * the data at the front of the range. i.e. one new_stripe
5270 * along from reshape_progress new_maps to after where
5271 * reshape_safe old_maps to
5273 writepos
= conf
->reshape_progress
;
5274 sector_div(writepos
, new_data_disks
);
5275 readpos
= conf
->reshape_progress
;
5276 sector_div(readpos
, data_disks
);
5277 safepos
= conf
->reshape_safe
;
5278 sector_div(safepos
, data_disks
);
5279 if (mddev
->reshape_backwards
) {
5280 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
5281 readpos
+= reshape_sectors
;
5282 safepos
+= reshape_sectors
;
5284 writepos
+= reshape_sectors
;
5285 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5286 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5289 /* Having calculated the 'writepos' possibly use it
5290 * to set 'stripe_addr' which is where we will write to.
5292 if (mddev
->reshape_backwards
) {
5293 BUG_ON(conf
->reshape_progress
== 0);
5294 stripe_addr
= writepos
;
5295 BUG_ON((mddev
->dev_sectors
&
5296 ~((sector_t
)reshape_sectors
- 1))
5297 - reshape_sectors
- stripe_addr
5300 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5301 stripe_addr
= sector_nr
;
5304 /* 'writepos' is the most advanced device address we might write.
5305 * 'readpos' is the least advanced device address we might read.
5306 * 'safepos' is the least address recorded in the metadata as having
5308 * If there is a min_offset_diff, these are adjusted either by
5309 * increasing the safepos/readpos if diff is negative, or
5310 * increasing writepos if diff is positive.
5311 * If 'readpos' is then behind 'writepos', there is no way that we can
5312 * ensure safety in the face of a crash - that must be done by userspace
5313 * making a backup of the data. So in that case there is no particular
5314 * rush to update metadata.
5315 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5316 * update the metadata to advance 'safepos' to match 'readpos' so that
5317 * we can be safe in the event of a crash.
5318 * So we insist on updating metadata if safepos is behind writepos and
5319 * readpos is beyond writepos.
5320 * In any case, update the metadata every 10 seconds.
5321 * Maybe that number should be configurable, but I'm not sure it is
5322 * worth it.... maybe it could be a multiple of safemode_delay???
5324 if (conf
->min_offset_diff
< 0) {
5325 safepos
+= -conf
->min_offset_diff
;
5326 readpos
+= -conf
->min_offset_diff
;
5328 writepos
+= conf
->min_offset_diff
;
5330 if ((mddev
->reshape_backwards
5331 ? (safepos
> writepos
&& readpos
< writepos
)
5332 : (safepos
< writepos
&& readpos
> writepos
)) ||
5333 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5334 /* Cannot proceed until we've updated the superblock... */
5335 wait_event(conf
->wait_for_overlap
,
5336 atomic_read(&conf
->reshape_stripes
)==0
5337 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5338 if (atomic_read(&conf
->reshape_stripes
) != 0)
5340 mddev
->reshape_position
= conf
->reshape_progress
;
5341 mddev
->curr_resync_completed
= sector_nr
;
5342 conf
->reshape_checkpoint
= jiffies
;
5343 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5344 md_wakeup_thread(mddev
->thread
);
5345 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
5346 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5347 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5349 spin_lock_irq(&conf
->device_lock
);
5350 conf
->reshape_safe
= mddev
->reshape_position
;
5351 spin_unlock_irq(&conf
->device_lock
);
5352 wake_up(&conf
->wait_for_overlap
);
5353 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5356 INIT_LIST_HEAD(&stripes
);
5357 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5359 int skipped_disk
= 0;
5360 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5361 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5362 atomic_inc(&conf
->reshape_stripes
);
5363 /* If any of this stripe is beyond the end of the old
5364 * array, then we need to zero those blocks
5366 for (j
=sh
->disks
; j
--;) {
5368 if (j
== sh
->pd_idx
)
5370 if (conf
->level
== 6 &&
5373 s
= compute_blocknr(sh
, j
, 0);
5374 if (s
< raid5_size(mddev
, 0, 0)) {
5378 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5379 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5380 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5382 if (!skipped_disk
) {
5383 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5384 set_bit(STRIPE_HANDLE
, &sh
->state
);
5386 list_add(&sh
->lru
, &stripes
);
5388 spin_lock_irq(&conf
->device_lock
);
5389 if (mddev
->reshape_backwards
)
5390 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5392 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5393 spin_unlock_irq(&conf
->device_lock
);
5394 /* Ok, those stripe are ready. We can start scheduling
5395 * reads on the source stripes.
5396 * The source stripes are determined by mapping the first and last
5397 * block on the destination stripes.
5400 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5403 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5404 * new_data_disks
- 1),
5406 if (last_sector
>= mddev
->dev_sectors
)
5407 last_sector
= mddev
->dev_sectors
- 1;
5408 while (first_sector
<= last_sector
) {
5409 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
5410 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5411 set_bit(STRIPE_HANDLE
, &sh
->state
);
5413 first_sector
+= STRIPE_SECTORS
;
5415 /* Now that the sources are clearly marked, we can release
5416 * the destination stripes
5418 while (!list_empty(&stripes
)) {
5419 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5420 list_del_init(&sh
->lru
);
5423 /* If this takes us to the resync_max point where we have to pause,
5424 * then we need to write out the superblock.
5426 sector_nr
+= reshape_sectors
;
5427 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
5428 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5429 /* Cannot proceed until we've updated the superblock... */
5430 wait_event(conf
->wait_for_overlap
,
5431 atomic_read(&conf
->reshape_stripes
) == 0
5432 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5433 if (atomic_read(&conf
->reshape_stripes
) != 0)
5435 mddev
->reshape_position
= conf
->reshape_progress
;
5436 mddev
->curr_resync_completed
= sector_nr
;
5437 conf
->reshape_checkpoint
= jiffies
;
5438 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5439 md_wakeup_thread(mddev
->thread
);
5440 wait_event(mddev
->sb_wait
,
5441 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
5442 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5443 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5445 spin_lock_irq(&conf
->device_lock
);
5446 conf
->reshape_safe
= mddev
->reshape_position
;
5447 spin_unlock_irq(&conf
->device_lock
);
5448 wake_up(&conf
->wait_for_overlap
);
5449 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5452 return reshape_sectors
;
5455 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5457 struct r5conf
*conf
= mddev
->private;
5458 struct stripe_head
*sh
;
5459 sector_t max_sector
= mddev
->dev_sectors
;
5460 sector_t sync_blocks
;
5461 int still_degraded
= 0;
5464 if (sector_nr
>= max_sector
) {
5465 /* just being told to finish up .. nothing much to do */
5467 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
5472 if (mddev
->curr_resync
< max_sector
) /* aborted */
5473 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
5475 else /* completed sync */
5477 bitmap_close_sync(mddev
->bitmap
);
5482 /* Allow raid5_quiesce to complete */
5483 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
5485 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
5486 return reshape_request(mddev
, sector_nr
, skipped
);
5488 /* No need to check resync_max as we never do more than one
5489 * stripe, and as resync_max will always be on a chunk boundary,
5490 * if the check in md_do_sync didn't fire, there is no chance
5491 * of overstepping resync_max here
5494 /* if there is too many failed drives and we are trying
5495 * to resync, then assert that we are finished, because there is
5496 * nothing we can do.
5498 if (mddev
->degraded
>= conf
->max_degraded
&&
5499 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
5500 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
5504 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
5506 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
5507 sync_blocks
>= STRIPE_SECTORS
) {
5508 /* we can skip this block, and probably more */
5509 sync_blocks
/= STRIPE_SECTORS
;
5511 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
5514 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
5516 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
5518 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
5519 /* make sure we don't swamp the stripe cache if someone else
5520 * is trying to get access
5522 schedule_timeout_uninterruptible(1);
5524 /* Need to check if array will still be degraded after recovery/resync
5525 * Note in case of > 1 drive failures it's possible we're rebuilding
5526 * one drive while leaving another faulty drive in array.
5529 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5530 struct md_rdev
*rdev
= ACCESS_ONCE(conf
->disks
[i
].rdev
);
5532 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
5537 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
5539 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
5540 set_bit(STRIPE_HANDLE
, &sh
->state
);
5544 return STRIPE_SECTORS
;
5547 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
5549 /* We may not be able to submit a whole bio at once as there
5550 * may not be enough stripe_heads available.
5551 * We cannot pre-allocate enough stripe_heads as we may need
5552 * more than exist in the cache (if we allow ever large chunks).
5553 * So we do one stripe head at a time and record in
5554 * ->bi_hw_segments how many have been done.
5556 * We *know* that this entire raid_bio is in one chunk, so
5557 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5559 struct stripe_head
*sh
;
5561 sector_t sector
, logical_sector
, last_sector
;
5566 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
5567 ~((sector_t
)STRIPE_SECTORS
-1);
5568 sector
= raid5_compute_sector(conf
, logical_sector
,
5570 last_sector
= bio_end_sector(raid_bio
);
5572 for (; logical_sector
< last_sector
;
5573 logical_sector
+= STRIPE_SECTORS
,
5574 sector
+= STRIPE_SECTORS
,
5577 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
5578 /* already done this stripe */
5581 sh
= get_active_stripe(conf
, sector
, 0, 1, 1);
5584 /* failed to get a stripe - must wait */
5585 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5586 conf
->retry_read_aligned
= raid_bio
;
5590 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
5592 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5593 conf
->retry_read_aligned
= raid_bio
;
5597 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
5602 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
5603 if (remaining
== 0) {
5604 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
5606 bio_endio(raid_bio
, 0);
5608 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5609 wake_up(&conf
->wait_for_stripe
);
5613 static int handle_active_stripes(struct r5conf
*conf
, int group
,
5614 struct r5worker
*worker
,
5615 struct list_head
*temp_inactive_list
)
5617 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
5618 int i
, batch_size
= 0, hash
;
5619 bool release_inactive
= false;
5621 while (batch_size
< MAX_STRIPE_BATCH
&&
5622 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
5623 batch
[batch_size
++] = sh
;
5625 if (batch_size
== 0) {
5626 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5627 if (!list_empty(temp_inactive_list
+ i
))
5629 if (i
== NR_STRIPE_HASH_LOCKS
)
5631 release_inactive
= true;
5633 spin_unlock_irq(&conf
->device_lock
);
5635 release_inactive_stripe_list(conf
, temp_inactive_list
,
5636 NR_STRIPE_HASH_LOCKS
);
5638 if (release_inactive
) {
5639 spin_lock_irq(&conf
->device_lock
);
5643 for (i
= 0; i
< batch_size
; i
++)
5644 handle_stripe(batch
[i
]);
5648 spin_lock_irq(&conf
->device_lock
);
5649 for (i
= 0; i
< batch_size
; i
++) {
5650 hash
= batch
[i
]->hash_lock_index
;
5651 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
5656 static void raid5_do_work(struct work_struct
*work
)
5658 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
5659 struct r5worker_group
*group
= worker
->group
;
5660 struct r5conf
*conf
= group
->conf
;
5661 int group_id
= group
- conf
->worker_groups
;
5663 struct blk_plug plug
;
5665 pr_debug("+++ raid5worker active\n");
5667 blk_start_plug(&plug
);
5669 spin_lock_irq(&conf
->device_lock
);
5671 int batch_size
, released
;
5673 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
5675 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
5676 worker
->temp_inactive_list
);
5677 worker
->working
= false;
5678 if (!batch_size
&& !released
)
5680 handled
+= batch_size
;
5682 pr_debug("%d stripes handled\n", handled
);
5684 spin_unlock_irq(&conf
->device_lock
);
5685 blk_finish_plug(&plug
);
5687 pr_debug("--- raid5worker inactive\n");
5691 * This is our raid5 kernel thread.
5693 * We scan the hash table for stripes which can be handled now.
5694 * During the scan, completed stripes are saved for us by the interrupt
5695 * handler, so that they will not have to wait for our next wakeup.
5697 static void raid5d(struct md_thread
*thread
)
5699 struct mddev
*mddev
= thread
->mddev
;
5700 struct r5conf
*conf
= mddev
->private;
5702 struct blk_plug plug
;
5704 pr_debug("+++ raid5d active\n");
5706 md_check_recovery(mddev
);
5708 blk_start_plug(&plug
);
5710 spin_lock_irq(&conf
->device_lock
);
5713 int batch_size
, released
;
5715 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
5718 !list_empty(&conf
->bitmap_list
)) {
5719 /* Now is a good time to flush some bitmap updates */
5721 spin_unlock_irq(&conf
->device_lock
);
5722 bitmap_unplug(mddev
->bitmap
);
5723 spin_lock_irq(&conf
->device_lock
);
5724 conf
->seq_write
= conf
->seq_flush
;
5725 activate_bit_delay(conf
, conf
->temp_inactive_list
);
5727 raid5_activate_delayed(conf
);
5729 while ((bio
= remove_bio_from_retry(conf
))) {
5731 spin_unlock_irq(&conf
->device_lock
);
5732 ok
= retry_aligned_read(conf
, bio
);
5733 spin_lock_irq(&conf
->device_lock
);
5739 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
5740 conf
->temp_inactive_list
);
5741 if (!batch_size
&& !released
)
5743 handled
+= batch_size
;
5745 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
5746 spin_unlock_irq(&conf
->device_lock
);
5747 md_check_recovery(mddev
);
5748 spin_lock_irq(&conf
->device_lock
);
5751 pr_debug("%d stripes handled\n", handled
);
5753 spin_unlock_irq(&conf
->device_lock
);
5755 async_tx_issue_pending_all();
5756 blk_finish_plug(&plug
);
5758 pr_debug("--- raid5d inactive\n");
5762 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
5764 struct r5conf
*conf
;
5766 spin_lock(&mddev
->lock
);
5767 conf
= mddev
->private;
5769 ret
= sprintf(page
, "%d\n", conf
->max_nr_stripes
);
5770 spin_unlock(&mddev
->lock
);
5775 raid5_set_cache_size(struct mddev
*mddev
, int size
)
5777 struct r5conf
*conf
= mddev
->private;
5781 if (size
<= 16 || size
> 32768)
5783 hash
= (conf
->max_nr_stripes
- 1) % NR_STRIPE_HASH_LOCKS
;
5784 while (size
< conf
->max_nr_stripes
) {
5785 if (drop_one_stripe(conf
, hash
))
5786 conf
->max_nr_stripes
--;
5791 hash
= NR_STRIPE_HASH_LOCKS
- 1;
5793 err
= md_allow_write(mddev
);
5796 hash
= conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
5797 while (size
> conf
->max_nr_stripes
) {
5798 if (grow_one_stripe(conf
, hash
))
5799 conf
->max_nr_stripes
++;
5801 hash
= (hash
+ 1) % NR_STRIPE_HASH_LOCKS
;
5805 EXPORT_SYMBOL(raid5_set_cache_size
);
5808 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
5810 struct r5conf
*conf
;
5814 if (len
>= PAGE_SIZE
)
5816 if (kstrtoul(page
, 10, &new))
5818 err
= mddev_lock(mddev
);
5821 conf
= mddev
->private;
5825 err
= raid5_set_cache_size(mddev
, new);
5826 mddev_unlock(mddev
);
5831 static struct md_sysfs_entry
5832 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
5833 raid5_show_stripe_cache_size
,
5834 raid5_store_stripe_cache_size
);
5837 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
5839 struct r5conf
*conf
;
5841 spin_lock(&mddev
->lock
);
5842 conf
= mddev
->private;
5844 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
5845 spin_unlock(&mddev
->lock
);
5850 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
5852 struct r5conf
*conf
;
5856 if (len
>= PAGE_SIZE
)
5858 if (kstrtoul(page
, 10, &new))
5861 err
= mddev_lock(mddev
);
5864 conf
= mddev
->private;
5867 else if (new > conf
->max_nr_stripes
)
5870 conf
->bypass_threshold
= new;
5871 mddev_unlock(mddev
);
5875 static struct md_sysfs_entry
5876 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
5878 raid5_show_preread_threshold
,
5879 raid5_store_preread_threshold
);
5882 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
5884 struct r5conf
*conf
;
5886 spin_lock(&mddev
->lock
);
5887 conf
= mddev
->private;
5889 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
5890 spin_unlock(&mddev
->lock
);
5895 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
5897 struct r5conf
*conf
;
5901 if (len
>= PAGE_SIZE
)
5903 if (kstrtoul(page
, 10, &new))
5907 err
= mddev_lock(mddev
);
5910 conf
= mddev
->private;
5913 else if (new != conf
->skip_copy
) {
5914 mddev_suspend(mddev
);
5915 conf
->skip_copy
= new;
5917 mddev
->queue
->backing_dev_info
.capabilities
|=
5918 BDI_CAP_STABLE_WRITES
;
5920 mddev
->queue
->backing_dev_info
.capabilities
&=
5921 ~BDI_CAP_STABLE_WRITES
;
5922 mddev_resume(mddev
);
5924 mddev_unlock(mddev
);
5928 static struct md_sysfs_entry
5929 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
5930 raid5_show_skip_copy
,
5931 raid5_store_skip_copy
);
5934 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
5936 struct r5conf
*conf
= mddev
->private;
5938 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
5943 static struct md_sysfs_entry
5944 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
5947 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
5949 struct r5conf
*conf
;
5951 spin_lock(&mddev
->lock
);
5952 conf
= mddev
->private;
5954 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
5955 spin_unlock(&mddev
->lock
);
5959 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
5961 int *worker_cnt_per_group
,
5962 struct r5worker_group
**worker_groups
);
5964 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
5966 struct r5conf
*conf
;
5969 struct r5worker_group
*new_groups
, *old_groups
;
5970 int group_cnt
, worker_cnt_per_group
;
5972 if (len
>= PAGE_SIZE
)
5974 if (kstrtoul(page
, 10, &new))
5977 err
= mddev_lock(mddev
);
5980 conf
= mddev
->private;
5983 else if (new != conf
->worker_cnt_per_group
) {
5984 mddev_suspend(mddev
);
5986 old_groups
= conf
->worker_groups
;
5988 flush_workqueue(raid5_wq
);
5990 err
= alloc_thread_groups(conf
, new,
5991 &group_cnt
, &worker_cnt_per_group
,
5994 spin_lock_irq(&conf
->device_lock
);
5995 conf
->group_cnt
= group_cnt
;
5996 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
5997 conf
->worker_groups
= new_groups
;
5998 spin_unlock_irq(&conf
->device_lock
);
6001 kfree(old_groups
[0].workers
);
6004 mddev_resume(mddev
);
6006 mddev_unlock(mddev
);
6011 static struct md_sysfs_entry
6012 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6013 raid5_show_group_thread_cnt
,
6014 raid5_store_group_thread_cnt
);
6016 static struct attribute
*raid5_attrs
[] = {
6017 &raid5_stripecache_size
.attr
,
6018 &raid5_stripecache_active
.attr
,
6019 &raid5_preread_bypass_threshold
.attr
,
6020 &raid5_group_thread_cnt
.attr
,
6021 &raid5_skip_copy
.attr
,
6024 static struct attribute_group raid5_attrs_group
= {
6026 .attrs
= raid5_attrs
,
6029 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6031 int *worker_cnt_per_group
,
6032 struct r5worker_group
**worker_groups
)
6036 struct r5worker
*workers
;
6038 *worker_cnt_per_group
= cnt
;
6041 *worker_groups
= NULL
;
6044 *group_cnt
= num_possible_nodes();
6045 size
= sizeof(struct r5worker
) * cnt
;
6046 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6047 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6048 *group_cnt
, GFP_NOIO
);
6049 if (!*worker_groups
|| !workers
) {
6051 kfree(*worker_groups
);
6055 for (i
= 0; i
< *group_cnt
; i
++) {
6056 struct r5worker_group
*group
;
6058 group
= &(*worker_groups
)[i
];
6059 INIT_LIST_HEAD(&group
->handle_list
);
6061 group
->workers
= workers
+ i
* cnt
;
6063 for (j
= 0; j
< cnt
; j
++) {
6064 struct r5worker
*worker
= group
->workers
+ j
;
6065 worker
->group
= group
;
6066 INIT_WORK(&worker
->work
, raid5_do_work
);
6068 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6069 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6076 static void free_thread_groups(struct r5conf
*conf
)
6078 if (conf
->worker_groups
)
6079 kfree(conf
->worker_groups
[0].workers
);
6080 kfree(conf
->worker_groups
);
6081 conf
->worker_groups
= NULL
;
6085 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6087 struct r5conf
*conf
= mddev
->private;
6090 sectors
= mddev
->dev_sectors
;
6092 /* size is defined by the smallest of previous and new size */
6093 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6095 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
6096 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
6097 return sectors
* (raid_disks
- conf
->max_degraded
);
6100 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6102 safe_put_page(percpu
->spare_page
);
6103 if (percpu
->scribble
)
6104 flex_array_free(percpu
->scribble
);
6105 percpu
->spare_page
= NULL
;
6106 percpu
->scribble
= NULL
;
6109 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6111 if (conf
->level
== 6 && !percpu
->spare_page
)
6112 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6113 if (!percpu
->scribble
)
6114 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6115 conf
->previous_raid_disks
), conf
->chunk_sectors
/
6116 STRIPE_SECTORS
, GFP_KERNEL
);
6118 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6119 free_scratch_buffer(conf
, percpu
);
6126 static void raid5_free_percpu(struct r5conf
*conf
)
6133 #ifdef CONFIG_HOTPLUG_CPU
6134 unregister_cpu_notifier(&conf
->cpu_notify
);
6138 for_each_possible_cpu(cpu
)
6139 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6142 free_percpu(conf
->percpu
);
6145 static void free_conf(struct r5conf
*conf
)
6147 free_thread_groups(conf
);
6148 shrink_stripes(conf
);
6149 raid5_free_percpu(conf
);
6151 kfree(conf
->stripe_hashtbl
);
6155 #ifdef CONFIG_HOTPLUG_CPU
6156 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
6159 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
6160 long cpu
= (long)hcpu
;
6161 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6164 case CPU_UP_PREPARE
:
6165 case CPU_UP_PREPARE_FROZEN
:
6166 if (alloc_scratch_buffer(conf
, percpu
)) {
6167 pr_err("%s: failed memory allocation for cpu%ld\n",
6169 return notifier_from_errno(-ENOMEM
);
6173 case CPU_DEAD_FROZEN
:
6174 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6183 static int raid5_alloc_percpu(struct r5conf
*conf
)
6188 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6192 #ifdef CONFIG_HOTPLUG_CPU
6193 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
6194 conf
->cpu_notify
.priority
= 0;
6195 err
= register_cpu_notifier(&conf
->cpu_notify
);
6201 for_each_present_cpu(cpu
) {
6202 err
= alloc_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6204 pr_err("%s: failed memory allocation for cpu%ld\n",
6214 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6216 struct r5conf
*conf
;
6217 int raid_disk
, memory
, max_disks
;
6218 struct md_rdev
*rdev
;
6219 struct disk_info
*disk
;
6222 int group_cnt
, worker_cnt_per_group
;
6223 struct r5worker_group
*new_group
;
6225 if (mddev
->new_level
!= 5
6226 && mddev
->new_level
!= 4
6227 && mddev
->new_level
!= 6) {
6228 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6229 mdname(mddev
), mddev
->new_level
);
6230 return ERR_PTR(-EIO
);
6232 if ((mddev
->new_level
== 5
6233 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6234 (mddev
->new_level
== 6
6235 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6236 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
6237 mdname(mddev
), mddev
->new_layout
);
6238 return ERR_PTR(-EIO
);
6240 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6241 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6242 mdname(mddev
), mddev
->raid_disks
);
6243 return ERR_PTR(-EINVAL
);
6246 if (!mddev
->new_chunk_sectors
||
6247 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6248 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6249 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
6250 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6251 return ERR_PTR(-EINVAL
);
6254 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6257 /* Don't enable multi-threading by default*/
6258 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6260 conf
->group_cnt
= group_cnt
;
6261 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6262 conf
->worker_groups
= new_group
;
6265 spin_lock_init(&conf
->device_lock
);
6266 seqcount_init(&conf
->gen_lock
);
6267 init_waitqueue_head(&conf
->wait_for_stripe
);
6268 init_waitqueue_head(&conf
->wait_for_overlap
);
6269 INIT_LIST_HEAD(&conf
->handle_list
);
6270 INIT_LIST_HEAD(&conf
->hold_list
);
6271 INIT_LIST_HEAD(&conf
->delayed_list
);
6272 INIT_LIST_HEAD(&conf
->bitmap_list
);
6273 init_llist_head(&conf
->released_stripes
);
6274 atomic_set(&conf
->active_stripes
, 0);
6275 atomic_set(&conf
->preread_active_stripes
, 0);
6276 atomic_set(&conf
->active_aligned_reads
, 0);
6277 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6278 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6280 conf
->raid_disks
= mddev
->raid_disks
;
6281 if (mddev
->reshape_position
== MaxSector
)
6282 conf
->previous_raid_disks
= mddev
->raid_disks
;
6284 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6285 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6287 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6292 conf
->mddev
= mddev
;
6294 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6297 /* We init hash_locks[0] separately to that it can be used
6298 * as the reference lock in the spin_lock_nest_lock() call
6299 * in lock_all_device_hash_locks_irq in order to convince
6300 * lockdep that we know what we are doing.
6302 spin_lock_init(conf
->hash_locks
);
6303 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6304 spin_lock_init(conf
->hash_locks
+ i
);
6306 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6307 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6309 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6310 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6312 conf
->level
= mddev
->new_level
;
6313 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6314 if (raid5_alloc_percpu(conf
) != 0)
6317 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6319 rdev_for_each(rdev
, mddev
) {
6320 raid_disk
= rdev
->raid_disk
;
6321 if (raid_disk
>= max_disks
6324 disk
= conf
->disks
+ raid_disk
;
6326 if (test_bit(Replacement
, &rdev
->flags
)) {
6327 if (disk
->replacement
)
6329 disk
->replacement
= rdev
;
6336 if (test_bit(In_sync
, &rdev
->flags
)) {
6337 char b
[BDEVNAME_SIZE
];
6338 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
6340 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
6341 } else if (rdev
->saved_raid_disk
!= raid_disk
)
6342 /* Cannot rely on bitmap to complete recovery */
6346 conf
->level
= mddev
->new_level
;
6347 if (conf
->level
== 6)
6348 conf
->max_degraded
= 2;
6350 conf
->max_degraded
= 1;
6351 conf
->algorithm
= mddev
->new_layout
;
6352 conf
->reshape_progress
= mddev
->reshape_position
;
6353 if (conf
->reshape_progress
!= MaxSector
) {
6354 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
6355 conf
->prev_algo
= mddev
->layout
;
6358 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
6359 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
6360 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
6361 if (grow_stripes(conf
, NR_STRIPES
)) {
6363 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6364 mdname(mddev
), memory
);
6367 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
6368 mdname(mddev
), memory
);
6370 sprintf(pers_name
, "raid%d", mddev
->new_level
);
6371 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
6372 if (!conf
->thread
) {
6374 "md/raid:%s: couldn't allocate thread.\n",
6384 return ERR_PTR(-EIO
);
6386 return ERR_PTR(-ENOMEM
);
6389 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
6392 case ALGORITHM_PARITY_0
:
6393 if (raid_disk
< max_degraded
)
6396 case ALGORITHM_PARITY_N
:
6397 if (raid_disk
>= raid_disks
- max_degraded
)
6400 case ALGORITHM_PARITY_0_6
:
6401 if (raid_disk
== 0 ||
6402 raid_disk
== raid_disks
- 1)
6405 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6406 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6407 case ALGORITHM_LEFT_SYMMETRIC_6
:
6408 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6409 if (raid_disk
== raid_disks
- 1)
6415 static int run(struct mddev
*mddev
)
6417 struct r5conf
*conf
;
6418 int working_disks
= 0;
6419 int dirty_parity_disks
= 0;
6420 struct md_rdev
*rdev
;
6421 sector_t reshape_offset
= 0;
6423 long long min_offset_diff
= 0;
6426 if (mddev
->recovery_cp
!= MaxSector
)
6427 printk(KERN_NOTICE
"md/raid:%s: not clean"
6428 " -- starting background reconstruction\n",
6431 rdev_for_each(rdev
, mddev
) {
6433 if (rdev
->raid_disk
< 0)
6435 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
6437 min_offset_diff
= diff
;
6439 } else if (mddev
->reshape_backwards
&&
6440 diff
< min_offset_diff
)
6441 min_offset_diff
= diff
;
6442 else if (!mddev
->reshape_backwards
&&
6443 diff
> min_offset_diff
)
6444 min_offset_diff
= diff
;
6447 if (mddev
->reshape_position
!= MaxSector
) {
6448 /* Check that we can continue the reshape.
6449 * Difficulties arise if the stripe we would write to
6450 * next is at or after the stripe we would read from next.
6451 * For a reshape that changes the number of devices, this
6452 * is only possible for a very short time, and mdadm makes
6453 * sure that time appears to have past before assembling
6454 * the array. So we fail if that time hasn't passed.
6455 * For a reshape that keeps the number of devices the same
6456 * mdadm must be monitoring the reshape can keeping the
6457 * critical areas read-only and backed up. It will start
6458 * the array in read-only mode, so we check for that.
6460 sector_t here_new
, here_old
;
6462 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
6464 if (mddev
->new_level
!= mddev
->level
) {
6465 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
6466 "required - aborting.\n",
6470 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6471 /* reshape_position must be on a new-stripe boundary, and one
6472 * further up in new geometry must map after here in old
6475 here_new
= mddev
->reshape_position
;
6476 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
6477 (mddev
->raid_disks
- max_degraded
))) {
6478 printk(KERN_ERR
"md/raid:%s: reshape_position not "
6479 "on a stripe boundary\n", mdname(mddev
));
6482 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
6483 /* here_new is the stripe we will write to */
6484 here_old
= mddev
->reshape_position
;
6485 sector_div(here_old
, mddev
->chunk_sectors
*
6486 (old_disks
-max_degraded
));
6487 /* here_old is the first stripe that we might need to read
6489 if (mddev
->delta_disks
== 0) {
6490 if ((here_new
* mddev
->new_chunk_sectors
!=
6491 here_old
* mddev
->chunk_sectors
)) {
6492 printk(KERN_ERR
"md/raid:%s: reshape position is"
6493 " confused - aborting\n", mdname(mddev
));
6496 /* We cannot be sure it is safe to start an in-place
6497 * reshape. It is only safe if user-space is monitoring
6498 * and taking constant backups.
6499 * mdadm always starts a situation like this in
6500 * readonly mode so it can take control before
6501 * allowing any writes. So just check for that.
6503 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
6504 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
6505 /* not really in-place - so OK */;
6506 else if (mddev
->ro
== 0) {
6507 printk(KERN_ERR
"md/raid:%s: in-place reshape "
6508 "must be started in read-only mode "
6513 } else if (mddev
->reshape_backwards
6514 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
6515 here_old
* mddev
->chunk_sectors
)
6516 : (here_new
* mddev
->new_chunk_sectors
>=
6517 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
6518 /* Reading from the same stripe as writing to - bad */
6519 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
6520 "auto-recovery - aborting.\n",
6524 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
6526 /* OK, we should be able to continue; */
6528 BUG_ON(mddev
->level
!= mddev
->new_level
);
6529 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
6530 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
6531 BUG_ON(mddev
->delta_disks
!= 0);
6534 if (mddev
->private == NULL
)
6535 conf
= setup_conf(mddev
);
6537 conf
= mddev
->private;
6540 return PTR_ERR(conf
);
6542 conf
->min_offset_diff
= min_offset_diff
;
6543 mddev
->thread
= conf
->thread
;
6544 conf
->thread
= NULL
;
6545 mddev
->private = conf
;
6547 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
6549 rdev
= conf
->disks
[i
].rdev
;
6550 if (!rdev
&& conf
->disks
[i
].replacement
) {
6551 /* The replacement is all we have yet */
6552 rdev
= conf
->disks
[i
].replacement
;
6553 conf
->disks
[i
].replacement
= NULL
;
6554 clear_bit(Replacement
, &rdev
->flags
);
6555 conf
->disks
[i
].rdev
= rdev
;
6559 if (conf
->disks
[i
].replacement
&&
6560 conf
->reshape_progress
!= MaxSector
) {
6561 /* replacements and reshape simply do not mix. */
6562 printk(KERN_ERR
"md: cannot handle concurrent "
6563 "replacement and reshape.\n");
6566 if (test_bit(In_sync
, &rdev
->flags
)) {
6570 /* This disc is not fully in-sync. However if it
6571 * just stored parity (beyond the recovery_offset),
6572 * when we don't need to be concerned about the
6573 * array being dirty.
6574 * When reshape goes 'backwards', we never have
6575 * partially completed devices, so we only need
6576 * to worry about reshape going forwards.
6578 /* Hack because v0.91 doesn't store recovery_offset properly. */
6579 if (mddev
->major_version
== 0 &&
6580 mddev
->minor_version
> 90)
6581 rdev
->recovery_offset
= reshape_offset
;
6583 if (rdev
->recovery_offset
< reshape_offset
) {
6584 /* We need to check old and new layout */
6585 if (!only_parity(rdev
->raid_disk
,
6588 conf
->max_degraded
))
6591 if (!only_parity(rdev
->raid_disk
,
6593 conf
->previous_raid_disks
,
6594 conf
->max_degraded
))
6596 dirty_parity_disks
++;
6600 * 0 for a fully functional array, 1 or 2 for a degraded array.
6602 mddev
->degraded
= calc_degraded(conf
);
6604 if (has_failed(conf
)) {
6605 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
6606 " (%d/%d failed)\n",
6607 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
6611 /* device size must be a multiple of chunk size */
6612 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
6613 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
6615 if (mddev
->degraded
> dirty_parity_disks
&&
6616 mddev
->recovery_cp
!= MaxSector
) {
6617 if (mddev
->ok_start_degraded
)
6619 "md/raid:%s: starting dirty degraded array"
6620 " - data corruption possible.\n",
6624 "md/raid:%s: cannot start dirty degraded array.\n",
6630 if (mddev
->degraded
== 0)
6631 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
6632 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
6633 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
6636 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
6637 " out of %d devices, algorithm %d\n",
6638 mdname(mddev
), conf
->level
,
6639 mddev
->raid_disks
- mddev
->degraded
,
6640 mddev
->raid_disks
, mddev
->new_layout
);
6642 print_raid5_conf(conf
);
6644 if (conf
->reshape_progress
!= MaxSector
) {
6645 conf
->reshape_safe
= conf
->reshape_progress
;
6646 atomic_set(&conf
->reshape_stripes
, 0);
6647 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6648 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6649 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6650 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6651 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6655 /* Ok, everything is just fine now */
6656 if (mddev
->to_remove
== &raid5_attrs_group
)
6657 mddev
->to_remove
= NULL
;
6658 else if (mddev
->kobj
.sd
&&
6659 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
6661 "raid5: failed to create sysfs attributes for %s\n",
6663 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6667 bool discard_supported
= true;
6668 /* read-ahead size must cover two whole stripes, which
6669 * is 2 * (datadisks) * chunksize where 'n' is the
6670 * number of raid devices
6672 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
6673 int stripe
= data_disks
*
6674 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
6675 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6676 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6678 chunk_size
= mddev
->chunk_sectors
<< 9;
6679 blk_queue_io_min(mddev
->queue
, chunk_size
);
6680 blk_queue_io_opt(mddev
->queue
, chunk_size
*
6681 (conf
->raid_disks
- conf
->max_degraded
));
6682 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
6684 * We can only discard a whole stripe. It doesn't make sense to
6685 * discard data disk but write parity disk
6687 stripe
= stripe
* PAGE_SIZE
;
6688 /* Round up to power of 2, as discard handling
6689 * currently assumes that */
6690 while ((stripe
-1) & stripe
)
6691 stripe
= (stripe
| (stripe
-1)) + 1;
6692 mddev
->queue
->limits
.discard_alignment
= stripe
;
6693 mddev
->queue
->limits
.discard_granularity
= stripe
;
6695 * unaligned part of discard request will be ignored, so can't
6696 * guarantee discard_zeroes_data
6698 mddev
->queue
->limits
.discard_zeroes_data
= 0;
6700 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
6702 rdev_for_each(rdev
, mddev
) {
6703 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6704 rdev
->data_offset
<< 9);
6705 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6706 rdev
->new_data_offset
<< 9);
6708 * discard_zeroes_data is required, otherwise data
6709 * could be lost. Consider a scenario: discard a stripe
6710 * (the stripe could be inconsistent if
6711 * discard_zeroes_data is 0); write one disk of the
6712 * stripe (the stripe could be inconsistent again
6713 * depending on which disks are used to calculate
6714 * parity); the disk is broken; The stripe data of this
6717 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
6718 !bdev_get_queue(rdev
->bdev
)->
6719 limits
.discard_zeroes_data
)
6720 discard_supported
= false;
6721 /* Unfortunately, discard_zeroes_data is not currently
6722 * a guarantee - just a hint. So we only allow DISCARD
6723 * if the sysadmin has confirmed that only safe devices
6724 * are in use by setting a module parameter.
6726 if (!devices_handle_discard_safely
) {
6727 if (discard_supported
) {
6728 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6729 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6731 discard_supported
= false;
6735 if (discard_supported
&&
6736 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
6737 mddev
->queue
->limits
.discard_granularity
>= stripe
)
6738 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
6741 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
6747 md_unregister_thread(&mddev
->thread
);
6748 print_raid5_conf(conf
);
6750 mddev
->private = NULL
;
6751 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
6755 static void raid5_free(struct mddev
*mddev
, void *priv
)
6757 struct r5conf
*conf
= priv
;
6760 mddev
->to_remove
= &raid5_attrs_group
;
6763 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
6765 struct r5conf
*conf
= mddev
->private;
6768 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
6769 mddev
->chunk_sectors
/ 2, mddev
->layout
);
6770 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
6771 for (i
= 0; i
< conf
->raid_disks
; i
++)
6772 seq_printf (seq
, "%s",
6773 conf
->disks
[i
].rdev
&&
6774 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
6775 seq_printf (seq
, "]");
6778 static void print_raid5_conf (struct r5conf
*conf
)
6781 struct disk_info
*tmp
;
6783 printk(KERN_DEBUG
"RAID conf printout:\n");
6785 printk("(conf==NULL)\n");
6788 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
6790 conf
->raid_disks
- conf
->mddev
->degraded
);
6792 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6793 char b
[BDEVNAME_SIZE
];
6794 tmp
= conf
->disks
+ i
;
6796 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
6797 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
6798 bdevname(tmp
->rdev
->bdev
, b
));
6802 static int raid5_spare_active(struct mddev
*mddev
)
6805 struct r5conf
*conf
= mddev
->private;
6806 struct disk_info
*tmp
;
6808 unsigned long flags
;
6810 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6811 tmp
= conf
->disks
+ i
;
6812 if (tmp
->replacement
6813 && tmp
->replacement
->recovery_offset
== MaxSector
6814 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
6815 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
6816 /* Replacement has just become active. */
6818 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
6821 /* Replaced device not technically faulty,
6822 * but we need to be sure it gets removed
6823 * and never re-added.
6825 set_bit(Faulty
, &tmp
->rdev
->flags
);
6826 sysfs_notify_dirent_safe(
6827 tmp
->rdev
->sysfs_state
);
6829 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
6830 } else if (tmp
->rdev
6831 && tmp
->rdev
->recovery_offset
== MaxSector
6832 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
6833 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
6835 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
6838 spin_lock_irqsave(&conf
->device_lock
, flags
);
6839 mddev
->degraded
= calc_degraded(conf
);
6840 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
6841 print_raid5_conf(conf
);
6845 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
6847 struct r5conf
*conf
= mddev
->private;
6849 int number
= rdev
->raid_disk
;
6850 struct md_rdev
**rdevp
;
6851 struct disk_info
*p
= conf
->disks
+ number
;
6853 print_raid5_conf(conf
);
6854 if (rdev
== p
->rdev
)
6856 else if (rdev
== p
->replacement
)
6857 rdevp
= &p
->replacement
;
6861 if (number
>= conf
->raid_disks
&&
6862 conf
->reshape_progress
== MaxSector
)
6863 clear_bit(In_sync
, &rdev
->flags
);
6865 if (test_bit(In_sync
, &rdev
->flags
) ||
6866 atomic_read(&rdev
->nr_pending
)) {
6870 /* Only remove non-faulty devices if recovery
6873 if (!test_bit(Faulty
, &rdev
->flags
) &&
6874 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
6875 !has_failed(conf
) &&
6876 (!p
->replacement
|| p
->replacement
== rdev
) &&
6877 number
< conf
->raid_disks
) {
6883 if (atomic_read(&rdev
->nr_pending
)) {
6884 /* lost the race, try later */
6887 } else if (p
->replacement
) {
6888 /* We must have just cleared 'rdev' */
6889 p
->rdev
= p
->replacement
;
6890 clear_bit(Replacement
, &p
->replacement
->flags
);
6891 smp_mb(); /* Make sure other CPUs may see both as identical
6892 * but will never see neither - if they are careful
6894 p
->replacement
= NULL
;
6895 clear_bit(WantReplacement
, &rdev
->flags
);
6897 /* We might have just removed the Replacement as faulty-
6898 * clear the bit just in case
6900 clear_bit(WantReplacement
, &rdev
->flags
);
6903 print_raid5_conf(conf
);
6907 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
6909 struct r5conf
*conf
= mddev
->private;
6912 struct disk_info
*p
;
6914 int last
= conf
->raid_disks
- 1;
6916 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
6919 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
6920 /* no point adding a device */
6923 if (rdev
->raid_disk
>= 0)
6924 first
= last
= rdev
->raid_disk
;
6927 * find the disk ... but prefer rdev->saved_raid_disk
6930 if (rdev
->saved_raid_disk
>= 0 &&
6931 rdev
->saved_raid_disk
>= first
&&
6932 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
6933 first
= rdev
->saved_raid_disk
;
6935 for (disk
= first
; disk
<= last
; disk
++) {
6936 p
= conf
->disks
+ disk
;
6937 if (p
->rdev
== NULL
) {
6938 clear_bit(In_sync
, &rdev
->flags
);
6939 rdev
->raid_disk
= disk
;
6941 if (rdev
->saved_raid_disk
!= disk
)
6943 rcu_assign_pointer(p
->rdev
, rdev
);
6947 for (disk
= first
; disk
<= last
; disk
++) {
6948 p
= conf
->disks
+ disk
;
6949 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
6950 p
->replacement
== NULL
) {
6951 clear_bit(In_sync
, &rdev
->flags
);
6952 set_bit(Replacement
, &rdev
->flags
);
6953 rdev
->raid_disk
= disk
;
6956 rcu_assign_pointer(p
->replacement
, rdev
);
6961 print_raid5_conf(conf
);
6965 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
6967 /* no resync is happening, and there is enough space
6968 * on all devices, so we can resize.
6969 * We need to make sure resync covers any new space.
6970 * If the array is shrinking we should possibly wait until
6971 * any io in the removed space completes, but it hardly seems
6975 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
6976 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
6977 if (mddev
->external_size
&&
6978 mddev
->array_sectors
> newsize
)
6980 if (mddev
->bitmap
) {
6981 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
6985 md_set_array_sectors(mddev
, newsize
);
6986 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
6987 revalidate_disk(mddev
->gendisk
);
6988 if (sectors
> mddev
->dev_sectors
&&
6989 mddev
->recovery_cp
> mddev
->dev_sectors
) {
6990 mddev
->recovery_cp
= mddev
->dev_sectors
;
6991 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
6993 mddev
->dev_sectors
= sectors
;
6994 mddev
->resync_max_sectors
= sectors
;
6998 static int check_stripe_cache(struct mddev
*mddev
)
7000 /* Can only proceed if there are plenty of stripe_heads.
7001 * We need a minimum of one full stripe,, and for sensible progress
7002 * it is best to have about 4 times that.
7003 * If we require 4 times, then the default 256 4K stripe_heads will
7004 * allow for chunk sizes up to 256K, which is probably OK.
7005 * If the chunk size is greater, user-space should request more
7006 * stripe_heads first.
7008 struct r5conf
*conf
= mddev
->private;
7009 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7010 > conf
->max_nr_stripes
||
7011 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7012 > conf
->max_nr_stripes
) {
7013 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7015 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7022 static int check_reshape(struct mddev
*mddev
)
7024 struct r5conf
*conf
= mddev
->private;
7026 if (mddev
->delta_disks
== 0 &&
7027 mddev
->new_layout
== mddev
->layout
&&
7028 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7029 return 0; /* nothing to do */
7030 if (has_failed(conf
))
7032 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7033 /* We might be able to shrink, but the devices must
7034 * be made bigger first.
7035 * For raid6, 4 is the minimum size.
7036 * Otherwise 2 is the minimum
7039 if (mddev
->level
== 6)
7041 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7045 if (!check_stripe_cache(mddev
))
7048 return resize_stripes(conf
, (conf
->previous_raid_disks
7049 + mddev
->delta_disks
));
7052 static int raid5_start_reshape(struct mddev
*mddev
)
7054 struct r5conf
*conf
= mddev
->private;
7055 struct md_rdev
*rdev
;
7057 unsigned long flags
;
7059 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7062 if (!check_stripe_cache(mddev
))
7065 if (has_failed(conf
))
7068 rdev_for_each(rdev
, mddev
) {
7069 if (!test_bit(In_sync
, &rdev
->flags
)
7070 && !test_bit(Faulty
, &rdev
->flags
))
7074 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7075 /* Not enough devices even to make a degraded array
7080 /* Refuse to reduce size of the array. Any reductions in
7081 * array size must be through explicit setting of array_size
7084 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7085 < mddev
->array_sectors
) {
7086 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
7087 "before number of disks\n", mdname(mddev
));
7091 atomic_set(&conf
->reshape_stripes
, 0);
7092 spin_lock_irq(&conf
->device_lock
);
7093 write_seqcount_begin(&conf
->gen_lock
);
7094 conf
->previous_raid_disks
= conf
->raid_disks
;
7095 conf
->raid_disks
+= mddev
->delta_disks
;
7096 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7097 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7098 conf
->prev_algo
= conf
->algorithm
;
7099 conf
->algorithm
= mddev
->new_layout
;
7101 /* Code that selects data_offset needs to see the generation update
7102 * if reshape_progress has been set - so a memory barrier needed.
7105 if (mddev
->reshape_backwards
)
7106 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7108 conf
->reshape_progress
= 0;
7109 conf
->reshape_safe
= conf
->reshape_progress
;
7110 write_seqcount_end(&conf
->gen_lock
);
7111 spin_unlock_irq(&conf
->device_lock
);
7113 /* Now make sure any requests that proceeded on the assumption
7114 * the reshape wasn't running - like Discard or Read - have
7117 mddev_suspend(mddev
);
7118 mddev_resume(mddev
);
7120 /* Add some new drives, as many as will fit.
7121 * We know there are enough to make the newly sized array work.
7122 * Don't add devices if we are reducing the number of
7123 * devices in the array. This is because it is not possible
7124 * to correctly record the "partially reconstructed" state of
7125 * such devices during the reshape and confusion could result.
7127 if (mddev
->delta_disks
>= 0) {
7128 rdev_for_each(rdev
, mddev
)
7129 if (rdev
->raid_disk
< 0 &&
7130 !test_bit(Faulty
, &rdev
->flags
)) {
7131 if (raid5_add_disk(mddev
, rdev
) == 0) {
7133 >= conf
->previous_raid_disks
)
7134 set_bit(In_sync
, &rdev
->flags
);
7136 rdev
->recovery_offset
= 0;
7138 if (sysfs_link_rdev(mddev
, rdev
))
7139 /* Failure here is OK */;
7141 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7142 && !test_bit(Faulty
, &rdev
->flags
)) {
7143 /* This is a spare that was manually added */
7144 set_bit(In_sync
, &rdev
->flags
);
7147 /* When a reshape changes the number of devices,
7148 * ->degraded is measured against the larger of the
7149 * pre and post number of devices.
7151 spin_lock_irqsave(&conf
->device_lock
, flags
);
7152 mddev
->degraded
= calc_degraded(conf
);
7153 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7155 mddev
->raid_disks
= conf
->raid_disks
;
7156 mddev
->reshape_position
= conf
->reshape_progress
;
7157 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7159 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7160 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7161 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7162 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7163 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7165 if (!mddev
->sync_thread
) {
7166 mddev
->recovery
= 0;
7167 spin_lock_irq(&conf
->device_lock
);
7168 write_seqcount_begin(&conf
->gen_lock
);
7169 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7170 mddev
->new_chunk_sectors
=
7171 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7172 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7173 rdev_for_each(rdev
, mddev
)
7174 rdev
->new_data_offset
= rdev
->data_offset
;
7176 conf
->generation
--;
7177 conf
->reshape_progress
= MaxSector
;
7178 mddev
->reshape_position
= MaxSector
;
7179 write_seqcount_end(&conf
->gen_lock
);
7180 spin_unlock_irq(&conf
->device_lock
);
7183 conf
->reshape_checkpoint
= jiffies
;
7184 md_wakeup_thread(mddev
->sync_thread
);
7185 md_new_event(mddev
);
7189 /* This is called from the reshape thread and should make any
7190 * changes needed in 'conf'
7192 static void end_reshape(struct r5conf
*conf
)
7195 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7196 struct md_rdev
*rdev
;
7198 spin_lock_irq(&conf
->device_lock
);
7199 conf
->previous_raid_disks
= conf
->raid_disks
;
7200 rdev_for_each(rdev
, conf
->mddev
)
7201 rdev
->data_offset
= rdev
->new_data_offset
;
7203 conf
->reshape_progress
= MaxSector
;
7204 spin_unlock_irq(&conf
->device_lock
);
7205 wake_up(&conf
->wait_for_overlap
);
7207 /* read-ahead size must cover two whole stripes, which is
7208 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7210 if (conf
->mddev
->queue
) {
7211 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7212 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7214 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
7215 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
7220 /* This is called from the raid5d thread with mddev_lock held.
7221 * It makes config changes to the device.
7223 static void raid5_finish_reshape(struct mddev
*mddev
)
7225 struct r5conf
*conf
= mddev
->private;
7227 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7229 if (mddev
->delta_disks
> 0) {
7230 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7231 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7232 revalidate_disk(mddev
->gendisk
);
7235 spin_lock_irq(&conf
->device_lock
);
7236 mddev
->degraded
= calc_degraded(conf
);
7237 spin_unlock_irq(&conf
->device_lock
);
7238 for (d
= conf
->raid_disks
;
7239 d
< conf
->raid_disks
- mddev
->delta_disks
;
7241 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
7243 clear_bit(In_sync
, &rdev
->flags
);
7244 rdev
= conf
->disks
[d
].replacement
;
7246 clear_bit(In_sync
, &rdev
->flags
);
7249 mddev
->layout
= conf
->algorithm
;
7250 mddev
->chunk_sectors
= conf
->chunk_sectors
;
7251 mddev
->reshape_position
= MaxSector
;
7252 mddev
->delta_disks
= 0;
7253 mddev
->reshape_backwards
= 0;
7257 static void raid5_quiesce(struct mddev
*mddev
, int state
)
7259 struct r5conf
*conf
= mddev
->private;
7262 case 2: /* resume for a suspend */
7263 wake_up(&conf
->wait_for_overlap
);
7266 case 1: /* stop all writes */
7267 lock_all_device_hash_locks_irq(conf
);
7268 /* '2' tells resync/reshape to pause so that all
7269 * active stripes can drain
7272 wait_event_cmd(conf
->wait_for_stripe
,
7273 atomic_read(&conf
->active_stripes
) == 0 &&
7274 atomic_read(&conf
->active_aligned_reads
) == 0,
7275 unlock_all_device_hash_locks_irq(conf
),
7276 lock_all_device_hash_locks_irq(conf
));
7278 unlock_all_device_hash_locks_irq(conf
);
7279 /* allow reshape to continue */
7280 wake_up(&conf
->wait_for_overlap
);
7283 case 0: /* re-enable writes */
7284 lock_all_device_hash_locks_irq(conf
);
7286 wake_up(&conf
->wait_for_stripe
);
7287 wake_up(&conf
->wait_for_overlap
);
7288 unlock_all_device_hash_locks_irq(conf
);
7293 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
7295 struct r0conf
*raid0_conf
= mddev
->private;
7298 /* for raid0 takeover only one zone is supported */
7299 if (raid0_conf
->nr_strip_zones
> 1) {
7300 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7302 return ERR_PTR(-EINVAL
);
7305 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
7306 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
7307 mddev
->dev_sectors
= sectors
;
7308 mddev
->new_level
= level
;
7309 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7310 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
7311 mddev
->raid_disks
+= 1;
7312 mddev
->delta_disks
= 1;
7313 /* make sure it will be not marked as dirty */
7314 mddev
->recovery_cp
= MaxSector
;
7316 return setup_conf(mddev
);
7319 static void *raid5_takeover_raid1(struct mddev
*mddev
)
7323 if (mddev
->raid_disks
!= 2 ||
7324 mddev
->degraded
> 1)
7325 return ERR_PTR(-EINVAL
);
7327 /* Should check if there are write-behind devices? */
7329 chunksect
= 64*2; /* 64K by default */
7331 /* The array must be an exact multiple of chunksize */
7332 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
7335 if ((chunksect
<<9) < STRIPE_SIZE
)
7336 /* array size does not allow a suitable chunk size */
7337 return ERR_PTR(-EINVAL
);
7339 mddev
->new_level
= 5;
7340 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7341 mddev
->new_chunk_sectors
= chunksect
;
7343 return setup_conf(mddev
);
7346 static void *raid5_takeover_raid6(struct mddev
*mddev
)
7350 switch (mddev
->layout
) {
7351 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7352 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
7354 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7355 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
7357 case ALGORITHM_LEFT_SYMMETRIC_6
:
7358 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7360 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7361 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
7363 case ALGORITHM_PARITY_0_6
:
7364 new_layout
= ALGORITHM_PARITY_0
;
7366 case ALGORITHM_PARITY_N
:
7367 new_layout
= ALGORITHM_PARITY_N
;
7370 return ERR_PTR(-EINVAL
);
7372 mddev
->new_level
= 5;
7373 mddev
->new_layout
= new_layout
;
7374 mddev
->delta_disks
= -1;
7375 mddev
->raid_disks
-= 1;
7376 return setup_conf(mddev
);
7379 static int raid5_check_reshape(struct mddev
*mddev
)
7381 /* For a 2-drive array, the layout and chunk size can be changed
7382 * immediately as not restriping is needed.
7383 * For larger arrays we record the new value - after validation
7384 * to be used by a reshape pass.
7386 struct r5conf
*conf
= mddev
->private;
7387 int new_chunk
= mddev
->new_chunk_sectors
;
7389 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
7391 if (new_chunk
> 0) {
7392 if (!is_power_of_2(new_chunk
))
7394 if (new_chunk
< (PAGE_SIZE
>>9))
7396 if (mddev
->array_sectors
& (new_chunk
-1))
7397 /* not factor of array size */
7401 /* They look valid */
7403 if (mddev
->raid_disks
== 2) {
7404 /* can make the change immediately */
7405 if (mddev
->new_layout
>= 0) {
7406 conf
->algorithm
= mddev
->new_layout
;
7407 mddev
->layout
= mddev
->new_layout
;
7409 if (new_chunk
> 0) {
7410 conf
->chunk_sectors
= new_chunk
;
7411 mddev
->chunk_sectors
= new_chunk
;
7413 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7414 md_wakeup_thread(mddev
->thread
);
7416 return check_reshape(mddev
);
7419 static int raid6_check_reshape(struct mddev
*mddev
)
7421 int new_chunk
= mddev
->new_chunk_sectors
;
7423 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
7425 if (new_chunk
> 0) {
7426 if (!is_power_of_2(new_chunk
))
7428 if (new_chunk
< (PAGE_SIZE
>> 9))
7430 if (mddev
->array_sectors
& (new_chunk
-1))
7431 /* not factor of array size */
7435 /* They look valid */
7436 return check_reshape(mddev
);
7439 static void *raid5_takeover(struct mddev
*mddev
)
7441 /* raid5 can take over:
7442 * raid0 - if there is only one strip zone - make it a raid4 layout
7443 * raid1 - if there are two drives. We need to know the chunk size
7444 * raid4 - trivial - just use a raid4 layout.
7445 * raid6 - Providing it is a *_6 layout
7447 if (mddev
->level
== 0)
7448 return raid45_takeover_raid0(mddev
, 5);
7449 if (mddev
->level
== 1)
7450 return raid5_takeover_raid1(mddev
);
7451 if (mddev
->level
== 4) {
7452 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7453 mddev
->new_level
= 5;
7454 return setup_conf(mddev
);
7456 if (mddev
->level
== 6)
7457 return raid5_takeover_raid6(mddev
);
7459 return ERR_PTR(-EINVAL
);
7462 static void *raid4_takeover(struct mddev
*mddev
)
7464 /* raid4 can take over:
7465 * raid0 - if there is only one strip zone
7466 * raid5 - if layout is right
7468 if (mddev
->level
== 0)
7469 return raid45_takeover_raid0(mddev
, 4);
7470 if (mddev
->level
== 5 &&
7471 mddev
->layout
== ALGORITHM_PARITY_N
) {
7472 mddev
->new_layout
= 0;
7473 mddev
->new_level
= 4;
7474 return setup_conf(mddev
);
7476 return ERR_PTR(-EINVAL
);
7479 static struct md_personality raid5_personality
;
7481 static void *raid6_takeover(struct mddev
*mddev
)
7483 /* Currently can only take over a raid5. We map the
7484 * personality to an equivalent raid6 personality
7485 * with the Q block at the end.
7489 if (mddev
->pers
!= &raid5_personality
)
7490 return ERR_PTR(-EINVAL
);
7491 if (mddev
->degraded
> 1)
7492 return ERR_PTR(-EINVAL
);
7493 if (mddev
->raid_disks
> 253)
7494 return ERR_PTR(-EINVAL
);
7495 if (mddev
->raid_disks
< 3)
7496 return ERR_PTR(-EINVAL
);
7498 switch (mddev
->layout
) {
7499 case ALGORITHM_LEFT_ASYMMETRIC
:
7500 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
7502 case ALGORITHM_RIGHT_ASYMMETRIC
:
7503 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
7505 case ALGORITHM_LEFT_SYMMETRIC
:
7506 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
7508 case ALGORITHM_RIGHT_SYMMETRIC
:
7509 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
7511 case ALGORITHM_PARITY_0
:
7512 new_layout
= ALGORITHM_PARITY_0_6
;
7514 case ALGORITHM_PARITY_N
:
7515 new_layout
= ALGORITHM_PARITY_N
;
7518 return ERR_PTR(-EINVAL
);
7520 mddev
->new_level
= 6;
7521 mddev
->new_layout
= new_layout
;
7522 mddev
->delta_disks
= 1;
7523 mddev
->raid_disks
+= 1;
7524 return setup_conf(mddev
);
7527 static struct md_personality raid6_personality
=
7531 .owner
= THIS_MODULE
,
7532 .make_request
= make_request
,
7536 .error_handler
= error
,
7537 .hot_add_disk
= raid5_add_disk
,
7538 .hot_remove_disk
= raid5_remove_disk
,
7539 .spare_active
= raid5_spare_active
,
7540 .sync_request
= sync_request
,
7541 .resize
= raid5_resize
,
7543 .check_reshape
= raid6_check_reshape
,
7544 .start_reshape
= raid5_start_reshape
,
7545 .finish_reshape
= raid5_finish_reshape
,
7546 .quiesce
= raid5_quiesce
,
7547 .takeover
= raid6_takeover
,
7548 .congested
= raid5_congested
,
7549 .mergeable_bvec
= raid5_mergeable_bvec
,
7551 static struct md_personality raid5_personality
=
7555 .owner
= THIS_MODULE
,
7556 .make_request
= make_request
,
7560 .error_handler
= error
,
7561 .hot_add_disk
= raid5_add_disk
,
7562 .hot_remove_disk
= raid5_remove_disk
,
7563 .spare_active
= raid5_spare_active
,
7564 .sync_request
= sync_request
,
7565 .resize
= raid5_resize
,
7567 .check_reshape
= raid5_check_reshape
,
7568 .start_reshape
= raid5_start_reshape
,
7569 .finish_reshape
= raid5_finish_reshape
,
7570 .quiesce
= raid5_quiesce
,
7571 .takeover
= raid5_takeover
,
7572 .congested
= raid5_congested
,
7573 .mergeable_bvec
= raid5_mergeable_bvec
,
7576 static struct md_personality raid4_personality
=
7580 .owner
= THIS_MODULE
,
7581 .make_request
= make_request
,
7585 .error_handler
= error
,
7586 .hot_add_disk
= raid5_add_disk
,
7587 .hot_remove_disk
= raid5_remove_disk
,
7588 .spare_active
= raid5_spare_active
,
7589 .sync_request
= sync_request
,
7590 .resize
= raid5_resize
,
7592 .check_reshape
= raid5_check_reshape
,
7593 .start_reshape
= raid5_start_reshape
,
7594 .finish_reshape
= raid5_finish_reshape
,
7595 .quiesce
= raid5_quiesce
,
7596 .takeover
= raid4_takeover
,
7597 .congested
= raid5_congested
,
7598 .mergeable_bvec
= raid5_mergeable_bvec
,
7601 static int __init
raid5_init(void)
7603 raid5_wq
= alloc_workqueue("raid5wq",
7604 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
7607 register_md_personality(&raid6_personality
);
7608 register_md_personality(&raid5_personality
);
7609 register_md_personality(&raid4_personality
);
7613 static void raid5_exit(void)
7615 unregister_md_personality(&raid6_personality
);
7616 unregister_md_personality(&raid5_personality
);
7617 unregister_md_personality(&raid4_personality
);
7618 destroy_workqueue(raid5_wq
);
7621 module_init(raid5_init
);
7622 module_exit(raid5_exit
);
7623 MODULE_LICENSE("GPL");
7624 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7625 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7626 MODULE_ALIAS("md-raid5");
7627 MODULE_ALIAS("md-raid4");
7628 MODULE_ALIAS("md-level-5");
7629 MODULE_ALIAS("md-level-4");
7630 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7631 MODULE_ALIAS("md-raid6");
7632 MODULE_ALIAS("md-level-6");
7634 /* This used to be two separate modules, they were: */
7635 MODULE_ALIAS("raid5");
7636 MODULE_ALIAS("raid6");