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
, gfp_t gfp
)
503 int num
= sh
->raid_conf
->pool_size
;
505 for (i
= 0; i
< num
; i
++) {
508 if (!(page
= alloc_page(gfp
))) {
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 (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
676 sh
= get_free_stripe(conf
, hash
);
677 if (!sh
&& llist_empty(&conf
->released_stripes
) &&
678 !test_bit(R5_DID_ALLOC
, &conf
->cache_state
))
679 set_bit(R5_ALLOC_MORE
,
682 if (noblock
&& sh
== NULL
)
685 set_bit(R5_INACTIVE_BLOCKED
,
688 conf
->wait_for_stripe
,
689 !list_empty(conf
->inactive_list
+ hash
) &&
690 (atomic_read(&conf
->active_stripes
)
691 < (conf
->max_nr_stripes
* 3 / 4)
692 || !test_bit(R5_INACTIVE_BLOCKED
,
693 &conf
->cache_state
)),
694 *(conf
->hash_locks
+ hash
));
695 clear_bit(R5_INACTIVE_BLOCKED
,
698 init_stripe(sh
, sector
, previous
);
699 atomic_inc(&sh
->count
);
701 } else if (!atomic_inc_not_zero(&sh
->count
)) {
702 spin_lock(&conf
->device_lock
);
703 if (!atomic_read(&sh
->count
)) {
704 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
705 atomic_inc(&conf
->active_stripes
);
706 BUG_ON(list_empty(&sh
->lru
) &&
707 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
708 list_del_init(&sh
->lru
);
710 sh
->group
->stripes_cnt
--;
714 atomic_inc(&sh
->count
);
715 spin_unlock(&conf
->device_lock
);
717 } while (sh
== NULL
);
719 spin_unlock_irq(conf
->hash_locks
+ hash
);
723 static bool is_full_stripe_write(struct stripe_head
*sh
)
725 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
726 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
729 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
733 spin_lock(&sh2
->stripe_lock
);
734 spin_lock_nested(&sh1
->stripe_lock
, 1);
736 spin_lock(&sh1
->stripe_lock
);
737 spin_lock_nested(&sh2
->stripe_lock
, 1);
741 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
743 spin_unlock(&sh1
->stripe_lock
);
744 spin_unlock(&sh2
->stripe_lock
);
748 /* Only freshly new full stripe normal write stripe can be added to a batch list */
749 static bool stripe_can_batch(struct stripe_head
*sh
)
751 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
752 is_full_stripe_write(sh
);
755 /* we only do back search */
756 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
758 struct stripe_head
*head
;
759 sector_t head_sector
, tmp_sec
;
763 if (!stripe_can_batch(sh
))
765 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
766 tmp_sec
= sh
->sector
;
767 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
769 head_sector
= sh
->sector
- STRIPE_SECTORS
;
771 hash
= stripe_hash_locks_hash(head_sector
);
772 spin_lock_irq(conf
->hash_locks
+ hash
);
773 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
774 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
775 spin_lock(&conf
->device_lock
);
776 if (!atomic_read(&head
->count
)) {
777 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
778 atomic_inc(&conf
->active_stripes
);
779 BUG_ON(list_empty(&head
->lru
) &&
780 !test_bit(STRIPE_EXPANDING
, &head
->state
));
781 list_del_init(&head
->lru
);
783 head
->group
->stripes_cnt
--;
787 atomic_inc(&head
->count
);
788 spin_unlock(&conf
->device_lock
);
790 spin_unlock_irq(conf
->hash_locks
+ hash
);
794 if (!stripe_can_batch(head
))
797 lock_two_stripes(head
, sh
);
798 /* clear_batch_ready clear the flag */
799 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
806 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
808 if (head
->dev
[dd_idx
].towrite
->bi_rw
!= sh
->dev
[dd_idx
].towrite
->bi_rw
)
811 if (head
->batch_head
) {
812 spin_lock(&head
->batch_head
->batch_lock
);
813 /* This batch list is already running */
814 if (!stripe_can_batch(head
)) {
815 spin_unlock(&head
->batch_head
->batch_lock
);
820 * at this point, head's BATCH_READY could be cleared, but we
821 * can still add the stripe to batch list
823 list_add(&sh
->batch_list
, &head
->batch_list
);
824 spin_unlock(&head
->batch_head
->batch_lock
);
826 sh
->batch_head
= head
->batch_head
;
828 head
->batch_head
= head
;
829 sh
->batch_head
= head
->batch_head
;
830 spin_lock(&head
->batch_lock
);
831 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
832 spin_unlock(&head
->batch_lock
);
835 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
836 if (atomic_dec_return(&conf
->preread_active_stripes
)
838 md_wakeup_thread(conf
->mddev
->thread
);
840 atomic_inc(&sh
->count
);
842 unlock_two_stripes(head
, sh
);
844 release_stripe(head
);
847 /* Determine if 'data_offset' or 'new_data_offset' should be used
848 * in this stripe_head.
850 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
852 sector_t progress
= conf
->reshape_progress
;
853 /* Need a memory barrier to make sure we see the value
854 * of conf->generation, or ->data_offset that was set before
855 * reshape_progress was updated.
858 if (progress
== MaxSector
)
860 if (sh
->generation
== conf
->generation
- 1)
862 /* We are in a reshape, and this is a new-generation stripe,
863 * so use new_data_offset.
869 raid5_end_read_request(struct bio
*bi
, int error
);
871 raid5_end_write_request(struct bio
*bi
, int error
);
873 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
875 struct r5conf
*conf
= sh
->raid_conf
;
876 int i
, disks
= sh
->disks
;
877 struct stripe_head
*head_sh
= sh
;
881 for (i
= disks
; i
--; ) {
883 int replace_only
= 0;
884 struct bio
*bi
, *rbi
;
885 struct md_rdev
*rdev
, *rrdev
= NULL
;
888 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
889 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
893 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
895 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
897 else if (test_and_clear_bit(R5_WantReplace
,
898 &sh
->dev
[i
].flags
)) {
903 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
907 bi
= &sh
->dev
[i
].req
;
908 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
911 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
912 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
913 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
922 /* We raced and saw duplicates */
925 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
930 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
933 atomic_inc(&rdev
->nr_pending
);
934 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
937 atomic_inc(&rrdev
->nr_pending
);
940 /* We have already checked bad blocks for reads. Now
941 * need to check for writes. We never accept write errors
942 * on the replacement, so we don't to check rrdev.
944 while ((rw
& WRITE
) && rdev
&&
945 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
948 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
949 &first_bad
, &bad_sectors
);
954 set_bit(BlockedBadBlocks
, &rdev
->flags
);
955 if (!conf
->mddev
->external
&&
956 conf
->mddev
->flags
) {
957 /* It is very unlikely, but we might
958 * still need to write out the
959 * bad block log - better give it
961 md_check_recovery(conf
->mddev
);
964 * Because md_wait_for_blocked_rdev
965 * will dec nr_pending, we must
966 * increment it first.
968 atomic_inc(&rdev
->nr_pending
);
969 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
971 /* Acknowledged bad block - skip the write */
972 rdev_dec_pending(rdev
, conf
->mddev
);
978 if (s
->syncing
|| s
->expanding
|| s
->expanded
980 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
982 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
985 bi
->bi_bdev
= rdev
->bdev
;
987 bi
->bi_end_io
= (rw
& WRITE
)
988 ? raid5_end_write_request
989 : raid5_end_read_request
;
992 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
993 __func__
, (unsigned long long)sh
->sector
,
995 atomic_inc(&sh
->count
);
997 atomic_inc(&head_sh
->count
);
998 if (use_new_offset(conf
, sh
))
999 bi
->bi_iter
.bi_sector
= (sh
->sector
1000 + rdev
->new_data_offset
);
1002 bi
->bi_iter
.bi_sector
= (sh
->sector
1003 + rdev
->data_offset
);
1004 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1005 bi
->bi_rw
|= REQ_NOMERGE
;
1007 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1008 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1009 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1011 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1012 bi
->bi_io_vec
[0].bv_offset
= 0;
1013 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1015 * If this is discard request, set bi_vcnt 0. We don't
1016 * want to confuse SCSI because SCSI will replace payload
1018 if (rw
& REQ_DISCARD
)
1021 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1023 if (conf
->mddev
->gendisk
)
1024 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
1025 bi
, disk_devt(conf
->mddev
->gendisk
),
1027 generic_make_request(bi
);
1030 if (s
->syncing
|| s
->expanding
|| s
->expanded
1032 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1034 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1037 rbi
->bi_bdev
= rrdev
->bdev
;
1039 BUG_ON(!(rw
& WRITE
));
1040 rbi
->bi_end_io
= raid5_end_write_request
;
1041 rbi
->bi_private
= sh
;
1043 pr_debug("%s: for %llu schedule op %ld on "
1044 "replacement disc %d\n",
1045 __func__
, (unsigned long long)sh
->sector
,
1047 atomic_inc(&sh
->count
);
1049 atomic_inc(&head_sh
->count
);
1050 if (use_new_offset(conf
, sh
))
1051 rbi
->bi_iter
.bi_sector
= (sh
->sector
1052 + rrdev
->new_data_offset
);
1054 rbi
->bi_iter
.bi_sector
= (sh
->sector
1055 + rrdev
->data_offset
);
1056 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1057 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1058 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1060 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1061 rbi
->bi_io_vec
[0].bv_offset
= 0;
1062 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1064 * If this is discard request, set bi_vcnt 0. We don't
1065 * want to confuse SCSI because SCSI will replace payload
1067 if (rw
& REQ_DISCARD
)
1069 if (conf
->mddev
->gendisk
)
1070 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
1071 rbi
, disk_devt(conf
->mddev
->gendisk
),
1073 generic_make_request(rbi
);
1075 if (!rdev
&& !rrdev
) {
1077 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1078 pr_debug("skip op %ld on disc %d for sector %llu\n",
1079 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
1080 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1082 set_bit(STRIPE_BATCH_ERR
,
1083 &sh
->batch_head
->state
);
1084 set_bit(STRIPE_HANDLE
, &sh
->state
);
1087 if (!head_sh
->batch_head
)
1089 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1096 static struct dma_async_tx_descriptor
*
1097 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1098 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1099 struct stripe_head
*sh
)
1102 struct bvec_iter iter
;
1103 struct page
*bio_page
;
1105 struct async_submit_ctl submit
;
1106 enum async_tx_flags flags
= 0;
1108 if (bio
->bi_iter
.bi_sector
>= sector
)
1109 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1111 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1114 flags
|= ASYNC_TX_FENCE
;
1115 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1117 bio_for_each_segment(bvl
, bio
, iter
) {
1118 int len
= bvl
.bv_len
;
1122 if (page_offset
< 0) {
1123 b_offset
= -page_offset
;
1124 page_offset
+= b_offset
;
1128 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1129 clen
= STRIPE_SIZE
- page_offset
;
1134 b_offset
+= bvl
.bv_offset
;
1135 bio_page
= bvl
.bv_page
;
1137 if (sh
->raid_conf
->skip_copy
&&
1138 b_offset
== 0 && page_offset
== 0 &&
1139 clen
== STRIPE_SIZE
)
1142 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1143 b_offset
, clen
, &submit
);
1145 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1146 page_offset
, clen
, &submit
);
1148 /* chain the operations */
1149 submit
.depend_tx
= tx
;
1151 if (clen
< len
) /* hit end of page */
1159 static void ops_complete_biofill(void *stripe_head_ref
)
1161 struct stripe_head
*sh
= stripe_head_ref
;
1162 struct bio
*return_bi
= NULL
;
1165 pr_debug("%s: stripe %llu\n", __func__
,
1166 (unsigned long long)sh
->sector
);
1168 /* clear completed biofills */
1169 for (i
= sh
->disks
; i
--; ) {
1170 struct r5dev
*dev
= &sh
->dev
[i
];
1172 /* acknowledge completion of a biofill operation */
1173 /* and check if we need to reply to a read request,
1174 * new R5_Wantfill requests are held off until
1175 * !STRIPE_BIOFILL_RUN
1177 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1178 struct bio
*rbi
, *rbi2
;
1183 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1184 dev
->sector
+ STRIPE_SECTORS
) {
1185 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1186 if (!raid5_dec_bi_active_stripes(rbi
)) {
1187 rbi
->bi_next
= return_bi
;
1194 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1196 return_io(return_bi
);
1198 set_bit(STRIPE_HANDLE
, &sh
->state
);
1202 static void ops_run_biofill(struct stripe_head
*sh
)
1204 struct dma_async_tx_descriptor
*tx
= NULL
;
1205 struct async_submit_ctl submit
;
1208 BUG_ON(sh
->batch_head
);
1209 pr_debug("%s: stripe %llu\n", __func__
,
1210 (unsigned long long)sh
->sector
);
1212 for (i
= sh
->disks
; i
--; ) {
1213 struct r5dev
*dev
= &sh
->dev
[i
];
1214 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1216 spin_lock_irq(&sh
->stripe_lock
);
1217 dev
->read
= rbi
= dev
->toread
;
1219 spin_unlock_irq(&sh
->stripe_lock
);
1220 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1221 dev
->sector
+ STRIPE_SECTORS
) {
1222 tx
= async_copy_data(0, rbi
, &dev
->page
,
1223 dev
->sector
, tx
, sh
);
1224 rbi
= r5_next_bio(rbi
, dev
->sector
);
1229 atomic_inc(&sh
->count
);
1230 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1231 async_trigger_callback(&submit
);
1234 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1241 tgt
= &sh
->dev
[target
];
1242 set_bit(R5_UPTODATE
, &tgt
->flags
);
1243 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1244 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1247 static void ops_complete_compute(void *stripe_head_ref
)
1249 struct stripe_head
*sh
= stripe_head_ref
;
1251 pr_debug("%s: stripe %llu\n", __func__
,
1252 (unsigned long long)sh
->sector
);
1254 /* mark the computed target(s) as uptodate */
1255 mark_target_uptodate(sh
, sh
->ops
.target
);
1256 mark_target_uptodate(sh
, sh
->ops
.target2
);
1258 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1259 if (sh
->check_state
== check_state_compute_run
)
1260 sh
->check_state
= check_state_compute_result
;
1261 set_bit(STRIPE_HANDLE
, &sh
->state
);
1265 /* return a pointer to the address conversion region of the scribble buffer */
1266 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1267 struct raid5_percpu
*percpu
, int i
)
1271 addr
= flex_array_get(percpu
->scribble
, i
);
1272 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1275 /* return a pointer to the address conversion region of the scribble buffer */
1276 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1280 addr
= flex_array_get(percpu
->scribble
, i
);
1284 static struct dma_async_tx_descriptor
*
1285 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1287 int disks
= sh
->disks
;
1288 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1289 int target
= sh
->ops
.target
;
1290 struct r5dev
*tgt
= &sh
->dev
[target
];
1291 struct page
*xor_dest
= tgt
->page
;
1293 struct dma_async_tx_descriptor
*tx
;
1294 struct async_submit_ctl submit
;
1297 BUG_ON(sh
->batch_head
);
1299 pr_debug("%s: stripe %llu block: %d\n",
1300 __func__
, (unsigned long long)sh
->sector
, target
);
1301 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1303 for (i
= disks
; i
--; )
1305 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1307 atomic_inc(&sh
->count
);
1309 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1310 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1311 if (unlikely(count
== 1))
1312 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1314 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1319 /* set_syndrome_sources - populate source buffers for gen_syndrome
1320 * @srcs - (struct page *) array of size sh->disks
1321 * @sh - stripe_head to parse
1323 * Populates srcs in proper layout order for the stripe and returns the
1324 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1325 * destination buffer is recorded in srcs[count] and the Q destination
1326 * is recorded in srcs[count+1]].
1328 static int set_syndrome_sources(struct page
**srcs
,
1329 struct stripe_head
*sh
,
1332 int disks
= sh
->disks
;
1333 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1334 int d0_idx
= raid6_d0(sh
);
1338 for (i
= 0; i
< disks
; i
++)
1344 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1345 struct r5dev
*dev
= &sh
->dev
[i
];
1347 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1348 (srctype
== SYNDROME_SRC_ALL
) ||
1349 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1350 test_bit(R5_Wantdrain
, &dev
->flags
)) ||
1351 (srctype
== SYNDROME_SRC_WRITTEN
&&
1353 srcs
[slot
] = sh
->dev
[i
].page
;
1354 i
= raid6_next_disk(i
, disks
);
1355 } while (i
!= d0_idx
);
1357 return syndrome_disks
;
1360 static struct dma_async_tx_descriptor
*
1361 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1363 int disks
= sh
->disks
;
1364 struct page
**blocks
= to_addr_page(percpu
, 0);
1366 int qd_idx
= sh
->qd_idx
;
1367 struct dma_async_tx_descriptor
*tx
;
1368 struct async_submit_ctl submit
;
1374 BUG_ON(sh
->batch_head
);
1375 if (sh
->ops
.target
< 0)
1376 target
= sh
->ops
.target2
;
1377 else if (sh
->ops
.target2
< 0)
1378 target
= sh
->ops
.target
;
1380 /* we should only have one valid target */
1383 pr_debug("%s: stripe %llu block: %d\n",
1384 __func__
, (unsigned long long)sh
->sector
, target
);
1386 tgt
= &sh
->dev
[target
];
1387 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1390 atomic_inc(&sh
->count
);
1392 if (target
== qd_idx
) {
1393 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1394 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1395 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1396 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1397 ops_complete_compute
, sh
,
1398 to_addr_conv(sh
, percpu
, 0));
1399 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1401 /* Compute any data- or p-drive using XOR */
1403 for (i
= disks
; i
-- ; ) {
1404 if (i
== target
|| i
== qd_idx
)
1406 blocks
[count
++] = sh
->dev
[i
].page
;
1409 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1410 NULL
, ops_complete_compute
, sh
,
1411 to_addr_conv(sh
, percpu
, 0));
1412 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1418 static struct dma_async_tx_descriptor
*
1419 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1421 int i
, count
, disks
= sh
->disks
;
1422 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1423 int d0_idx
= raid6_d0(sh
);
1424 int faila
= -1, failb
= -1;
1425 int target
= sh
->ops
.target
;
1426 int target2
= sh
->ops
.target2
;
1427 struct r5dev
*tgt
= &sh
->dev
[target
];
1428 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1429 struct dma_async_tx_descriptor
*tx
;
1430 struct page
**blocks
= to_addr_page(percpu
, 0);
1431 struct async_submit_ctl submit
;
1433 BUG_ON(sh
->batch_head
);
1434 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1435 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1436 BUG_ON(target
< 0 || target2
< 0);
1437 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1438 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1440 /* we need to open-code set_syndrome_sources to handle the
1441 * slot number conversion for 'faila' and 'failb'
1443 for (i
= 0; i
< disks
; i
++)
1448 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1450 blocks
[slot
] = sh
->dev
[i
].page
;
1456 i
= raid6_next_disk(i
, disks
);
1457 } while (i
!= d0_idx
);
1459 BUG_ON(faila
== failb
);
1462 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1463 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1465 atomic_inc(&sh
->count
);
1467 if (failb
== syndrome_disks
+1) {
1468 /* Q disk is one of the missing disks */
1469 if (faila
== syndrome_disks
) {
1470 /* Missing P+Q, just recompute */
1471 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1472 ops_complete_compute
, sh
,
1473 to_addr_conv(sh
, percpu
, 0));
1474 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1475 STRIPE_SIZE
, &submit
);
1479 int qd_idx
= sh
->qd_idx
;
1481 /* Missing D+Q: recompute D from P, then recompute Q */
1482 if (target
== qd_idx
)
1483 data_target
= target2
;
1485 data_target
= target
;
1488 for (i
= disks
; i
-- ; ) {
1489 if (i
== data_target
|| i
== qd_idx
)
1491 blocks
[count
++] = sh
->dev
[i
].page
;
1493 dest
= sh
->dev
[data_target
].page
;
1494 init_async_submit(&submit
,
1495 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1497 to_addr_conv(sh
, percpu
, 0));
1498 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1501 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1502 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1503 ops_complete_compute
, sh
,
1504 to_addr_conv(sh
, percpu
, 0));
1505 return async_gen_syndrome(blocks
, 0, count
+2,
1506 STRIPE_SIZE
, &submit
);
1509 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1510 ops_complete_compute
, sh
,
1511 to_addr_conv(sh
, percpu
, 0));
1512 if (failb
== syndrome_disks
) {
1513 /* We're missing D+P. */
1514 return async_raid6_datap_recov(syndrome_disks
+2,
1518 /* We're missing D+D. */
1519 return async_raid6_2data_recov(syndrome_disks
+2,
1520 STRIPE_SIZE
, faila
, failb
,
1526 static void ops_complete_prexor(void *stripe_head_ref
)
1528 struct stripe_head
*sh
= stripe_head_ref
;
1530 pr_debug("%s: stripe %llu\n", __func__
,
1531 (unsigned long long)sh
->sector
);
1534 static struct dma_async_tx_descriptor
*
1535 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1536 struct dma_async_tx_descriptor
*tx
)
1538 int disks
= sh
->disks
;
1539 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1540 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1541 struct async_submit_ctl submit
;
1543 /* existing parity data subtracted */
1544 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1546 BUG_ON(sh
->batch_head
);
1547 pr_debug("%s: stripe %llu\n", __func__
,
1548 (unsigned long long)sh
->sector
);
1550 for (i
= disks
; i
--; ) {
1551 struct r5dev
*dev
= &sh
->dev
[i
];
1552 /* Only process blocks that are known to be uptodate */
1553 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1554 xor_srcs
[count
++] = dev
->page
;
1557 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1558 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1559 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1564 static struct dma_async_tx_descriptor
*
1565 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1566 struct dma_async_tx_descriptor
*tx
)
1568 struct page
**blocks
= to_addr_page(percpu
, 0);
1570 struct async_submit_ctl submit
;
1572 pr_debug("%s: stripe %llu\n", __func__
,
1573 (unsigned long long)sh
->sector
);
1575 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1577 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1578 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1579 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1584 static struct dma_async_tx_descriptor
*
1585 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1587 int disks
= sh
->disks
;
1589 struct stripe_head
*head_sh
= sh
;
1591 pr_debug("%s: stripe %llu\n", __func__
,
1592 (unsigned long long)sh
->sector
);
1594 for (i
= disks
; i
--; ) {
1599 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1604 spin_lock_irq(&sh
->stripe_lock
);
1605 chosen
= dev
->towrite
;
1606 dev
->towrite
= NULL
;
1607 sh
->overwrite_disks
= 0;
1608 BUG_ON(dev
->written
);
1609 wbi
= dev
->written
= chosen
;
1610 spin_unlock_irq(&sh
->stripe_lock
);
1611 WARN_ON(dev
->page
!= dev
->orig_page
);
1613 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1614 dev
->sector
+ STRIPE_SECTORS
) {
1615 if (wbi
->bi_rw
& REQ_FUA
)
1616 set_bit(R5_WantFUA
, &dev
->flags
);
1617 if (wbi
->bi_rw
& REQ_SYNC
)
1618 set_bit(R5_SyncIO
, &dev
->flags
);
1619 if (wbi
->bi_rw
& REQ_DISCARD
)
1620 set_bit(R5_Discard
, &dev
->flags
);
1622 tx
= async_copy_data(1, wbi
, &dev
->page
,
1623 dev
->sector
, tx
, sh
);
1624 if (dev
->page
!= dev
->orig_page
) {
1625 set_bit(R5_SkipCopy
, &dev
->flags
);
1626 clear_bit(R5_UPTODATE
, &dev
->flags
);
1627 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1630 wbi
= r5_next_bio(wbi
, dev
->sector
);
1633 if (head_sh
->batch_head
) {
1634 sh
= list_first_entry(&sh
->batch_list
,
1647 static void ops_complete_reconstruct(void *stripe_head_ref
)
1649 struct stripe_head
*sh
= stripe_head_ref
;
1650 int disks
= sh
->disks
;
1651 int pd_idx
= sh
->pd_idx
;
1652 int qd_idx
= sh
->qd_idx
;
1654 bool fua
= false, sync
= false, discard
= false;
1656 pr_debug("%s: stripe %llu\n", __func__
,
1657 (unsigned long long)sh
->sector
);
1659 for (i
= disks
; i
--; ) {
1660 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1661 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1662 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1665 for (i
= disks
; i
--; ) {
1666 struct r5dev
*dev
= &sh
->dev
[i
];
1668 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1669 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1670 set_bit(R5_UPTODATE
, &dev
->flags
);
1672 set_bit(R5_WantFUA
, &dev
->flags
);
1674 set_bit(R5_SyncIO
, &dev
->flags
);
1678 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1679 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1680 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1681 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1683 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1684 sh
->reconstruct_state
= reconstruct_state_result
;
1687 set_bit(STRIPE_HANDLE
, &sh
->state
);
1692 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1693 struct dma_async_tx_descriptor
*tx
)
1695 int disks
= sh
->disks
;
1696 struct page
**xor_srcs
;
1697 struct async_submit_ctl submit
;
1698 int count
, pd_idx
= sh
->pd_idx
, i
;
1699 struct page
*xor_dest
;
1701 unsigned long flags
;
1703 struct stripe_head
*head_sh
= sh
;
1706 pr_debug("%s: stripe %llu\n", __func__
,
1707 (unsigned long long)sh
->sector
);
1709 for (i
= 0; i
< sh
->disks
; i
++) {
1712 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1715 if (i
>= sh
->disks
) {
1716 atomic_inc(&sh
->count
);
1717 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1718 ops_complete_reconstruct(sh
);
1723 xor_srcs
= to_addr_page(percpu
, j
);
1724 /* check if prexor is active which means only process blocks
1725 * that are part of a read-modify-write (written)
1727 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1729 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1730 for (i
= disks
; i
--; ) {
1731 struct r5dev
*dev
= &sh
->dev
[i
];
1732 if (head_sh
->dev
[i
].written
)
1733 xor_srcs
[count
++] = dev
->page
;
1736 xor_dest
= sh
->dev
[pd_idx
].page
;
1737 for (i
= disks
; i
--; ) {
1738 struct r5dev
*dev
= &sh
->dev
[i
];
1740 xor_srcs
[count
++] = dev
->page
;
1744 /* 1/ if we prexor'd then the dest is reused as a source
1745 * 2/ if we did not prexor then we are redoing the parity
1746 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1747 * for the synchronous xor case
1749 last_stripe
= !head_sh
->batch_head
||
1750 list_first_entry(&sh
->batch_list
,
1751 struct stripe_head
, batch_list
) == head_sh
;
1753 flags
= ASYNC_TX_ACK
|
1754 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1756 atomic_inc(&head_sh
->count
);
1757 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1758 to_addr_conv(sh
, percpu
, j
));
1760 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1761 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1762 to_addr_conv(sh
, percpu
, j
));
1765 if (unlikely(count
== 1))
1766 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1768 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1771 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1778 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1779 struct dma_async_tx_descriptor
*tx
)
1781 struct async_submit_ctl submit
;
1782 struct page
**blocks
;
1783 int count
, i
, j
= 0;
1784 struct stripe_head
*head_sh
= sh
;
1787 unsigned long txflags
;
1789 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1791 for (i
= 0; i
< sh
->disks
; i
++) {
1792 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1794 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1797 if (i
>= sh
->disks
) {
1798 atomic_inc(&sh
->count
);
1799 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1800 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1801 ops_complete_reconstruct(sh
);
1806 blocks
= to_addr_page(percpu
, j
);
1808 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1809 synflags
= SYNDROME_SRC_WRITTEN
;
1810 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1812 synflags
= SYNDROME_SRC_ALL
;
1813 txflags
= ASYNC_TX_ACK
;
1816 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1817 last_stripe
= !head_sh
->batch_head
||
1818 list_first_entry(&sh
->batch_list
,
1819 struct stripe_head
, batch_list
) == head_sh
;
1822 atomic_inc(&head_sh
->count
);
1823 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1824 head_sh
, to_addr_conv(sh
, percpu
, j
));
1826 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1827 to_addr_conv(sh
, percpu
, j
));
1828 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1831 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1837 static void ops_complete_check(void *stripe_head_ref
)
1839 struct stripe_head
*sh
= stripe_head_ref
;
1841 pr_debug("%s: stripe %llu\n", __func__
,
1842 (unsigned long long)sh
->sector
);
1844 sh
->check_state
= check_state_check_result
;
1845 set_bit(STRIPE_HANDLE
, &sh
->state
);
1849 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1851 int disks
= sh
->disks
;
1852 int pd_idx
= sh
->pd_idx
;
1853 int qd_idx
= sh
->qd_idx
;
1854 struct page
*xor_dest
;
1855 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1856 struct dma_async_tx_descriptor
*tx
;
1857 struct async_submit_ctl submit
;
1861 pr_debug("%s: stripe %llu\n", __func__
,
1862 (unsigned long long)sh
->sector
);
1864 BUG_ON(sh
->batch_head
);
1866 xor_dest
= sh
->dev
[pd_idx
].page
;
1867 xor_srcs
[count
++] = xor_dest
;
1868 for (i
= disks
; i
--; ) {
1869 if (i
== pd_idx
|| i
== qd_idx
)
1871 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1874 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1875 to_addr_conv(sh
, percpu
, 0));
1876 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1877 &sh
->ops
.zero_sum_result
, &submit
);
1879 atomic_inc(&sh
->count
);
1880 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1881 tx
= async_trigger_callback(&submit
);
1884 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1886 struct page
**srcs
= to_addr_page(percpu
, 0);
1887 struct async_submit_ctl submit
;
1890 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1891 (unsigned long long)sh
->sector
, checkp
);
1893 BUG_ON(sh
->batch_head
);
1894 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
1898 atomic_inc(&sh
->count
);
1899 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1900 sh
, to_addr_conv(sh
, percpu
, 0));
1901 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1902 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1905 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1907 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1908 struct dma_async_tx_descriptor
*tx
= NULL
;
1909 struct r5conf
*conf
= sh
->raid_conf
;
1910 int level
= conf
->level
;
1911 struct raid5_percpu
*percpu
;
1915 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1916 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1917 ops_run_biofill(sh
);
1921 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1923 tx
= ops_run_compute5(sh
, percpu
);
1925 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1926 tx
= ops_run_compute6_1(sh
, percpu
);
1928 tx
= ops_run_compute6_2(sh
, percpu
);
1930 /* terminate the chain if reconstruct is not set to be run */
1931 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1935 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
1937 tx
= ops_run_prexor5(sh
, percpu
, tx
);
1939 tx
= ops_run_prexor6(sh
, percpu
, tx
);
1942 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1943 tx
= ops_run_biodrain(sh
, tx
);
1947 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1949 ops_run_reconstruct5(sh
, percpu
, tx
);
1951 ops_run_reconstruct6(sh
, percpu
, tx
);
1954 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1955 if (sh
->check_state
== check_state_run
)
1956 ops_run_check_p(sh
, percpu
);
1957 else if (sh
->check_state
== check_state_run_q
)
1958 ops_run_check_pq(sh
, percpu
, 0);
1959 else if (sh
->check_state
== check_state_run_pq
)
1960 ops_run_check_pq(sh
, percpu
, 1);
1965 if (overlap_clear
&& !sh
->batch_head
)
1966 for (i
= disks
; i
--; ) {
1967 struct r5dev
*dev
= &sh
->dev
[i
];
1968 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1969 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1974 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
1976 struct stripe_head
*sh
;
1977 sh
= kmem_cache_zalloc(conf
->slab_cache
, gfp
);
1981 sh
->raid_conf
= conf
;
1983 spin_lock_init(&sh
->stripe_lock
);
1985 if (grow_buffers(sh
, gfp
)) {
1987 kmem_cache_free(conf
->slab_cache
, sh
);
1990 sh
->hash_lock_index
=
1991 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
1992 /* we just created an active stripe so... */
1993 atomic_set(&sh
->count
, 1);
1994 atomic_inc(&conf
->active_stripes
);
1995 INIT_LIST_HEAD(&sh
->lru
);
1997 spin_lock_init(&sh
->batch_lock
);
1998 INIT_LIST_HEAD(&sh
->batch_list
);
1999 sh
->batch_head
= NULL
;
2001 conf
->max_nr_stripes
++;
2005 static int grow_stripes(struct r5conf
*conf
, int num
)
2007 struct kmem_cache
*sc
;
2008 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2010 if (conf
->mddev
->gendisk
)
2011 sprintf(conf
->cache_name
[0],
2012 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2014 sprintf(conf
->cache_name
[0],
2015 "raid%d-%p", conf
->level
, conf
->mddev
);
2016 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2018 conf
->active_name
= 0;
2019 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2020 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2024 conf
->slab_cache
= sc
;
2025 conf
->pool_size
= devs
;
2027 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2034 * scribble_len - return the required size of the scribble region
2035 * @num - total number of disks in the array
2037 * The size must be enough to contain:
2038 * 1/ a struct page pointer for each device in the array +2
2039 * 2/ room to convert each entry in (1) to its corresponding dma
2040 * (dma_map_page()) or page (page_address()) address.
2042 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2043 * calculate over all devices (not just the data blocks), using zeros in place
2044 * of the P and Q blocks.
2046 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2048 struct flex_array
*ret
;
2051 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2052 ret
= flex_array_alloc(len
, cnt
, flags
);
2055 /* always prealloc all elements, so no locking is required */
2056 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2057 flex_array_free(ret
);
2063 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2065 /* Make all the stripes able to hold 'newsize' devices.
2066 * New slots in each stripe get 'page' set to a new page.
2068 * This happens in stages:
2069 * 1/ create a new kmem_cache and allocate the required number of
2071 * 2/ gather all the old stripe_heads and transfer the pages across
2072 * to the new stripe_heads. This will have the side effect of
2073 * freezing the array as once all stripe_heads have been collected,
2074 * no IO will be possible. Old stripe heads are freed once their
2075 * pages have been transferred over, and the old kmem_cache is
2076 * freed when all stripes are done.
2077 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2078 * we simple return a failre status - no need to clean anything up.
2079 * 4/ allocate new pages for the new slots in the new stripe_heads.
2080 * If this fails, we don't bother trying the shrink the
2081 * stripe_heads down again, we just leave them as they are.
2082 * As each stripe_head is processed the new one is released into
2085 * Once step2 is started, we cannot afford to wait for a write,
2086 * so we use GFP_NOIO allocations.
2088 struct stripe_head
*osh
, *nsh
;
2089 LIST_HEAD(newstripes
);
2090 struct disk_info
*ndisks
;
2093 struct kmem_cache
*sc
;
2097 if (newsize
<= conf
->pool_size
)
2098 return 0; /* never bother to shrink */
2100 err
= md_allow_write(conf
->mddev
);
2105 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2106 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2111 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2112 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
2116 nsh
->raid_conf
= conf
;
2117 spin_lock_init(&nsh
->stripe_lock
);
2119 list_add(&nsh
->lru
, &newstripes
);
2122 /* didn't get enough, give up */
2123 while (!list_empty(&newstripes
)) {
2124 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2125 list_del(&nsh
->lru
);
2126 kmem_cache_free(sc
, nsh
);
2128 kmem_cache_destroy(sc
);
2131 /* Step 2 - Must use GFP_NOIO now.
2132 * OK, we have enough stripes, start collecting inactive
2133 * stripes and copying them over
2137 list_for_each_entry(nsh
, &newstripes
, lru
) {
2138 lock_device_hash_lock(conf
, hash
);
2139 wait_event_cmd(conf
->wait_for_stripe
,
2140 !list_empty(conf
->inactive_list
+ hash
),
2141 unlock_device_hash_lock(conf
, hash
),
2142 lock_device_hash_lock(conf
, hash
));
2143 osh
= get_free_stripe(conf
, hash
);
2144 unlock_device_hash_lock(conf
, hash
);
2145 atomic_set(&nsh
->count
, 1);
2146 for(i
=0; i
<conf
->pool_size
; i
++) {
2147 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2148 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2150 for( ; i
<newsize
; i
++)
2151 nsh
->dev
[i
].page
= NULL
;
2152 nsh
->hash_lock_index
= hash
;
2153 kmem_cache_free(conf
->slab_cache
, osh
);
2155 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2156 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2161 kmem_cache_destroy(conf
->slab_cache
);
2164 * At this point, we are holding all the stripes so the array
2165 * is completely stalled, so now is a good time to resize
2166 * conf->disks and the scribble region
2168 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2170 for (i
=0; i
<conf
->raid_disks
; i
++)
2171 ndisks
[i
] = conf
->disks
[i
];
2173 conf
->disks
= ndisks
;
2178 for_each_present_cpu(cpu
) {
2179 struct raid5_percpu
*percpu
;
2180 struct flex_array
*scribble
;
2182 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2183 scribble
= scribble_alloc(newsize
, conf
->chunk_sectors
/
2184 STRIPE_SECTORS
, GFP_NOIO
);
2187 flex_array_free(percpu
->scribble
);
2188 percpu
->scribble
= scribble
;
2196 /* Step 4, return new stripes to service */
2197 while(!list_empty(&newstripes
)) {
2198 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2199 list_del_init(&nsh
->lru
);
2201 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2202 if (nsh
->dev
[i
].page
== NULL
) {
2203 struct page
*p
= alloc_page(GFP_NOIO
);
2204 nsh
->dev
[i
].page
= p
;
2205 nsh
->dev
[i
].orig_page
= p
;
2209 release_stripe(nsh
);
2211 /* critical section pass, GFP_NOIO no longer needed */
2213 conf
->slab_cache
= sc
;
2214 conf
->active_name
= 1-conf
->active_name
;
2215 conf
->pool_size
= newsize
;
2219 static int drop_one_stripe(struct r5conf
*conf
)
2221 struct stripe_head
*sh
;
2222 int hash
= (conf
->max_nr_stripes
- 1) % NR_STRIPE_HASH_LOCKS
;
2224 spin_lock_irq(conf
->hash_locks
+ hash
);
2225 sh
= get_free_stripe(conf
, hash
);
2226 spin_unlock_irq(conf
->hash_locks
+ hash
);
2229 BUG_ON(atomic_read(&sh
->count
));
2231 kmem_cache_free(conf
->slab_cache
, sh
);
2232 atomic_dec(&conf
->active_stripes
);
2233 conf
->max_nr_stripes
--;
2237 static void shrink_stripes(struct r5conf
*conf
)
2239 while (conf
->max_nr_stripes
&&
2240 drop_one_stripe(conf
))
2243 if (conf
->slab_cache
)
2244 kmem_cache_destroy(conf
->slab_cache
);
2245 conf
->slab_cache
= NULL
;
2248 static void raid5_end_read_request(struct bio
* bi
, int error
)
2250 struct stripe_head
*sh
= bi
->bi_private
;
2251 struct r5conf
*conf
= sh
->raid_conf
;
2252 int disks
= sh
->disks
, i
;
2253 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2254 char b
[BDEVNAME_SIZE
];
2255 struct md_rdev
*rdev
= NULL
;
2258 for (i
=0 ; i
<disks
; i
++)
2259 if (bi
== &sh
->dev
[i
].req
)
2262 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2263 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2269 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2270 /* If replacement finished while this request was outstanding,
2271 * 'replacement' might be NULL already.
2272 * In that case it moved down to 'rdev'.
2273 * rdev is not removed until all requests are finished.
2275 rdev
= conf
->disks
[i
].replacement
;
2277 rdev
= conf
->disks
[i
].rdev
;
2279 if (use_new_offset(conf
, sh
))
2280 s
= sh
->sector
+ rdev
->new_data_offset
;
2282 s
= sh
->sector
+ rdev
->data_offset
;
2284 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2285 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2286 /* Note that this cannot happen on a
2287 * replacement device. We just fail those on
2292 "md/raid:%s: read error corrected"
2293 " (%lu sectors at %llu on %s)\n",
2294 mdname(conf
->mddev
), STRIPE_SECTORS
,
2295 (unsigned long long)s
,
2296 bdevname(rdev
->bdev
, b
));
2297 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2298 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2299 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2300 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2301 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2303 if (atomic_read(&rdev
->read_errors
))
2304 atomic_set(&rdev
->read_errors
, 0);
2306 const char *bdn
= bdevname(rdev
->bdev
, b
);
2310 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2311 atomic_inc(&rdev
->read_errors
);
2312 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2315 "md/raid:%s: read error on replacement device "
2316 "(sector %llu on %s).\n",
2317 mdname(conf
->mddev
),
2318 (unsigned long long)s
,
2320 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2324 "md/raid:%s: read error not correctable "
2325 "(sector %llu on %s).\n",
2326 mdname(conf
->mddev
),
2327 (unsigned long long)s
,
2329 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2334 "md/raid:%s: read error NOT corrected!! "
2335 "(sector %llu on %s).\n",
2336 mdname(conf
->mddev
),
2337 (unsigned long long)s
,
2339 } else if (atomic_read(&rdev
->read_errors
)
2340 > conf
->max_nr_stripes
)
2342 "md/raid:%s: Too many read errors, failing device %s.\n",
2343 mdname(conf
->mddev
), bdn
);
2346 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2347 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2350 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2351 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2352 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2354 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2356 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2357 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2359 && test_bit(In_sync
, &rdev
->flags
)
2360 && rdev_set_badblocks(
2361 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2362 md_error(conf
->mddev
, rdev
);
2365 rdev_dec_pending(rdev
, conf
->mddev
);
2366 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2367 set_bit(STRIPE_HANDLE
, &sh
->state
);
2371 static void raid5_end_write_request(struct bio
*bi
, int error
)
2373 struct stripe_head
*sh
= bi
->bi_private
;
2374 struct r5conf
*conf
= sh
->raid_conf
;
2375 int disks
= sh
->disks
, i
;
2376 struct md_rdev
*uninitialized_var(rdev
);
2377 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2380 int replacement
= 0;
2382 for (i
= 0 ; i
< disks
; i
++) {
2383 if (bi
== &sh
->dev
[i
].req
) {
2384 rdev
= conf
->disks
[i
].rdev
;
2387 if (bi
== &sh
->dev
[i
].rreq
) {
2388 rdev
= conf
->disks
[i
].replacement
;
2392 /* rdev was removed and 'replacement'
2393 * replaced it. rdev is not removed
2394 * until all requests are finished.
2396 rdev
= conf
->disks
[i
].rdev
;
2400 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2401 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2410 md_error(conf
->mddev
, rdev
);
2411 else if (is_badblock(rdev
, sh
->sector
,
2413 &first_bad
, &bad_sectors
))
2414 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2417 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2418 set_bit(WriteErrorSeen
, &rdev
->flags
);
2419 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2420 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2421 set_bit(MD_RECOVERY_NEEDED
,
2422 &rdev
->mddev
->recovery
);
2423 } else if (is_badblock(rdev
, sh
->sector
,
2425 &first_bad
, &bad_sectors
)) {
2426 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2427 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2428 /* That was a successful write so make
2429 * sure it looks like we already did
2432 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2435 rdev_dec_pending(rdev
, conf
->mddev
);
2437 if (sh
->batch_head
&& !uptodate
)
2438 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2440 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2441 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2442 set_bit(STRIPE_HANDLE
, &sh
->state
);
2445 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2446 release_stripe(sh
->batch_head
);
2449 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
2451 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
2453 struct r5dev
*dev
= &sh
->dev
[i
];
2455 bio_init(&dev
->req
);
2456 dev
->req
.bi_io_vec
= &dev
->vec
;
2457 dev
->req
.bi_max_vecs
= 1;
2458 dev
->req
.bi_private
= sh
;
2460 bio_init(&dev
->rreq
);
2461 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
2462 dev
->rreq
.bi_max_vecs
= 1;
2463 dev
->rreq
.bi_private
= sh
;
2466 dev
->sector
= compute_blocknr(sh
, i
, previous
);
2469 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2471 char b
[BDEVNAME_SIZE
];
2472 struct r5conf
*conf
= mddev
->private;
2473 unsigned long flags
;
2474 pr_debug("raid456: error called\n");
2476 spin_lock_irqsave(&conf
->device_lock
, flags
);
2477 clear_bit(In_sync
, &rdev
->flags
);
2478 mddev
->degraded
= calc_degraded(conf
);
2479 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2480 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2482 set_bit(Blocked
, &rdev
->flags
);
2483 set_bit(Faulty
, &rdev
->flags
);
2484 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2486 "md/raid:%s: Disk failure on %s, disabling device.\n"
2487 "md/raid:%s: Operation continuing on %d devices.\n",
2489 bdevname(rdev
->bdev
, b
),
2491 conf
->raid_disks
- mddev
->degraded
);
2495 * Input: a 'big' sector number,
2496 * Output: index of the data and parity disk, and the sector # in them.
2498 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2499 int previous
, int *dd_idx
,
2500 struct stripe_head
*sh
)
2502 sector_t stripe
, stripe2
;
2503 sector_t chunk_number
;
2504 unsigned int chunk_offset
;
2507 sector_t new_sector
;
2508 int algorithm
= previous
? conf
->prev_algo
2510 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2511 : conf
->chunk_sectors
;
2512 int raid_disks
= previous
? conf
->previous_raid_disks
2514 int data_disks
= raid_disks
- conf
->max_degraded
;
2516 /* First compute the information on this sector */
2519 * Compute the chunk number and the sector offset inside the chunk
2521 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2522 chunk_number
= r_sector
;
2525 * Compute the stripe number
2527 stripe
= chunk_number
;
2528 *dd_idx
= sector_div(stripe
, data_disks
);
2531 * Select the parity disk based on the user selected algorithm.
2533 pd_idx
= qd_idx
= -1;
2534 switch(conf
->level
) {
2536 pd_idx
= data_disks
;
2539 switch (algorithm
) {
2540 case ALGORITHM_LEFT_ASYMMETRIC
:
2541 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2542 if (*dd_idx
>= pd_idx
)
2545 case ALGORITHM_RIGHT_ASYMMETRIC
:
2546 pd_idx
= sector_div(stripe2
, raid_disks
);
2547 if (*dd_idx
>= pd_idx
)
2550 case ALGORITHM_LEFT_SYMMETRIC
:
2551 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2552 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2554 case ALGORITHM_RIGHT_SYMMETRIC
:
2555 pd_idx
= sector_div(stripe2
, raid_disks
);
2556 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2558 case ALGORITHM_PARITY_0
:
2562 case ALGORITHM_PARITY_N
:
2563 pd_idx
= data_disks
;
2571 switch (algorithm
) {
2572 case ALGORITHM_LEFT_ASYMMETRIC
:
2573 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2574 qd_idx
= pd_idx
+ 1;
2575 if (pd_idx
== raid_disks
-1) {
2576 (*dd_idx
)++; /* Q D D D P */
2578 } else if (*dd_idx
>= pd_idx
)
2579 (*dd_idx
) += 2; /* D D P Q D */
2581 case ALGORITHM_RIGHT_ASYMMETRIC
:
2582 pd_idx
= sector_div(stripe2
, raid_disks
);
2583 qd_idx
= pd_idx
+ 1;
2584 if (pd_idx
== raid_disks
-1) {
2585 (*dd_idx
)++; /* Q D D D P */
2587 } else if (*dd_idx
>= pd_idx
)
2588 (*dd_idx
) += 2; /* D D P Q D */
2590 case ALGORITHM_LEFT_SYMMETRIC
:
2591 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2592 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2593 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2595 case ALGORITHM_RIGHT_SYMMETRIC
:
2596 pd_idx
= sector_div(stripe2
, raid_disks
);
2597 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2598 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2601 case ALGORITHM_PARITY_0
:
2606 case ALGORITHM_PARITY_N
:
2607 pd_idx
= data_disks
;
2608 qd_idx
= data_disks
+ 1;
2611 case ALGORITHM_ROTATING_ZERO_RESTART
:
2612 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2613 * of blocks for computing Q is different.
2615 pd_idx
= sector_div(stripe2
, raid_disks
);
2616 qd_idx
= pd_idx
+ 1;
2617 if (pd_idx
== raid_disks
-1) {
2618 (*dd_idx
)++; /* Q D D D P */
2620 } else if (*dd_idx
>= pd_idx
)
2621 (*dd_idx
) += 2; /* D D P Q D */
2625 case ALGORITHM_ROTATING_N_RESTART
:
2626 /* Same a left_asymmetric, by first stripe is
2627 * D D D P Q rather than
2631 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2632 qd_idx
= pd_idx
+ 1;
2633 if (pd_idx
== raid_disks
-1) {
2634 (*dd_idx
)++; /* Q D D D P */
2636 } else if (*dd_idx
>= pd_idx
)
2637 (*dd_idx
) += 2; /* D D P Q D */
2641 case ALGORITHM_ROTATING_N_CONTINUE
:
2642 /* Same as left_symmetric but Q is before P */
2643 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2644 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2645 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2649 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2650 /* RAID5 left_asymmetric, with Q on last device */
2651 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2652 if (*dd_idx
>= pd_idx
)
2654 qd_idx
= raid_disks
- 1;
2657 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2658 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2659 if (*dd_idx
>= pd_idx
)
2661 qd_idx
= raid_disks
- 1;
2664 case ALGORITHM_LEFT_SYMMETRIC_6
:
2665 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2666 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2667 qd_idx
= raid_disks
- 1;
2670 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2671 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2672 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2673 qd_idx
= raid_disks
- 1;
2676 case ALGORITHM_PARITY_0_6
:
2679 qd_idx
= raid_disks
- 1;
2689 sh
->pd_idx
= pd_idx
;
2690 sh
->qd_idx
= qd_idx
;
2691 sh
->ddf_layout
= ddf_layout
;
2694 * Finally, compute the new sector number
2696 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2700 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2702 struct r5conf
*conf
= sh
->raid_conf
;
2703 int raid_disks
= sh
->disks
;
2704 int data_disks
= raid_disks
- conf
->max_degraded
;
2705 sector_t new_sector
= sh
->sector
, check
;
2706 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2707 : conf
->chunk_sectors
;
2708 int algorithm
= previous
? conf
->prev_algo
2712 sector_t chunk_number
;
2713 int dummy1
, dd_idx
= i
;
2715 struct stripe_head sh2
;
2717 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2718 stripe
= new_sector
;
2720 if (i
== sh
->pd_idx
)
2722 switch(conf
->level
) {
2725 switch (algorithm
) {
2726 case ALGORITHM_LEFT_ASYMMETRIC
:
2727 case ALGORITHM_RIGHT_ASYMMETRIC
:
2731 case ALGORITHM_LEFT_SYMMETRIC
:
2732 case ALGORITHM_RIGHT_SYMMETRIC
:
2735 i
-= (sh
->pd_idx
+ 1);
2737 case ALGORITHM_PARITY_0
:
2740 case ALGORITHM_PARITY_N
:
2747 if (i
== sh
->qd_idx
)
2748 return 0; /* It is the Q disk */
2749 switch (algorithm
) {
2750 case ALGORITHM_LEFT_ASYMMETRIC
:
2751 case ALGORITHM_RIGHT_ASYMMETRIC
:
2752 case ALGORITHM_ROTATING_ZERO_RESTART
:
2753 case ALGORITHM_ROTATING_N_RESTART
:
2754 if (sh
->pd_idx
== raid_disks
-1)
2755 i
--; /* Q D D D P */
2756 else if (i
> sh
->pd_idx
)
2757 i
-= 2; /* D D P Q D */
2759 case ALGORITHM_LEFT_SYMMETRIC
:
2760 case ALGORITHM_RIGHT_SYMMETRIC
:
2761 if (sh
->pd_idx
== raid_disks
-1)
2762 i
--; /* Q D D D P */
2767 i
-= (sh
->pd_idx
+ 2);
2770 case ALGORITHM_PARITY_0
:
2773 case ALGORITHM_PARITY_N
:
2775 case ALGORITHM_ROTATING_N_CONTINUE
:
2776 /* Like left_symmetric, but P is before Q */
2777 if (sh
->pd_idx
== 0)
2778 i
--; /* P D D D Q */
2783 i
-= (sh
->pd_idx
+ 1);
2786 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2787 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2791 case ALGORITHM_LEFT_SYMMETRIC_6
:
2792 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2794 i
+= data_disks
+ 1;
2795 i
-= (sh
->pd_idx
+ 1);
2797 case ALGORITHM_PARITY_0_6
:
2806 chunk_number
= stripe
* data_disks
+ i
;
2807 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2809 check
= raid5_compute_sector(conf
, r_sector
,
2810 previous
, &dummy1
, &sh2
);
2811 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2812 || sh2
.qd_idx
!= sh
->qd_idx
) {
2813 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2814 mdname(conf
->mddev
));
2821 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2822 int rcw
, int expand
)
2824 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
2825 struct r5conf
*conf
= sh
->raid_conf
;
2826 int level
= conf
->level
;
2830 for (i
= disks
; i
--; ) {
2831 struct r5dev
*dev
= &sh
->dev
[i
];
2834 set_bit(R5_LOCKED
, &dev
->flags
);
2835 set_bit(R5_Wantdrain
, &dev
->flags
);
2837 clear_bit(R5_UPTODATE
, &dev
->flags
);
2841 /* if we are not expanding this is a proper write request, and
2842 * there will be bios with new data to be drained into the
2847 /* False alarm, nothing to do */
2849 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2850 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2852 sh
->reconstruct_state
= reconstruct_state_run
;
2854 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2856 if (s
->locked
+ conf
->max_degraded
== disks
)
2857 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2858 atomic_inc(&conf
->pending_full_writes
);
2860 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2861 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2862 BUG_ON(level
== 6 &&
2863 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
2864 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
2866 for (i
= disks
; i
--; ) {
2867 struct r5dev
*dev
= &sh
->dev
[i
];
2868 if (i
== pd_idx
|| i
== qd_idx
)
2872 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2873 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2874 set_bit(R5_Wantdrain
, &dev
->flags
);
2875 set_bit(R5_LOCKED
, &dev
->flags
);
2876 clear_bit(R5_UPTODATE
, &dev
->flags
);
2881 /* False alarm - nothing to do */
2883 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2884 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2885 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2886 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2889 /* keep the parity disk(s) locked while asynchronous operations
2892 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2893 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2897 int qd_idx
= sh
->qd_idx
;
2898 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2900 set_bit(R5_LOCKED
, &dev
->flags
);
2901 clear_bit(R5_UPTODATE
, &dev
->flags
);
2905 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2906 __func__
, (unsigned long long)sh
->sector
,
2907 s
->locked
, s
->ops_request
);
2911 * Each stripe/dev can have one or more bion attached.
2912 * toread/towrite point to the first in a chain.
2913 * The bi_next chain must be in order.
2915 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
2916 int forwrite
, int previous
)
2919 struct r5conf
*conf
= sh
->raid_conf
;
2922 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2923 (unsigned long long)bi
->bi_iter
.bi_sector
,
2924 (unsigned long long)sh
->sector
);
2927 * If several bio share a stripe. The bio bi_phys_segments acts as a
2928 * reference count to avoid race. The reference count should already be
2929 * increased before this function is called (for example, in
2930 * make_request()), so other bio sharing this stripe will not free the
2931 * stripe. If a stripe is owned by one stripe, the stripe lock will
2934 spin_lock_irq(&sh
->stripe_lock
);
2935 /* Don't allow new IO added to stripes in batch list */
2939 bip
= &sh
->dev
[dd_idx
].towrite
;
2943 bip
= &sh
->dev
[dd_idx
].toread
;
2944 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
2945 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
2947 bip
= & (*bip
)->bi_next
;
2949 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
2952 if (!forwrite
|| previous
)
2953 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
2955 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2959 raid5_inc_bi_active_stripes(bi
);
2962 /* check if page is covered */
2963 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2964 for (bi
=sh
->dev
[dd_idx
].towrite
;
2965 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2966 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
2967 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2968 if (bio_end_sector(bi
) >= sector
)
2969 sector
= bio_end_sector(bi
);
2971 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2972 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
2973 sh
->overwrite_disks
++;
2976 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2977 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
2978 (unsigned long long)sh
->sector
, dd_idx
);
2979 spin_unlock_irq(&sh
->stripe_lock
);
2981 if (conf
->mddev
->bitmap
&& firstwrite
) {
2982 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2984 sh
->bm_seq
= conf
->seq_flush
+1;
2985 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2988 if (stripe_can_batch(sh
))
2989 stripe_add_to_batch_list(conf
, sh
);
2993 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2994 spin_unlock_irq(&sh
->stripe_lock
);
2998 static void end_reshape(struct r5conf
*conf
);
3000 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3001 struct stripe_head
*sh
)
3003 int sectors_per_chunk
=
3004 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3006 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3007 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3009 raid5_compute_sector(conf
,
3010 stripe
* (disks
- conf
->max_degraded
)
3011 *sectors_per_chunk
+ chunk_offset
,
3017 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3018 struct stripe_head_state
*s
, int disks
,
3019 struct bio
**return_bi
)
3022 BUG_ON(sh
->batch_head
);
3023 for (i
= disks
; i
--; ) {
3027 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3028 struct md_rdev
*rdev
;
3030 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3031 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
3032 atomic_inc(&rdev
->nr_pending
);
3037 if (!rdev_set_badblocks(
3041 md_error(conf
->mddev
, rdev
);
3042 rdev_dec_pending(rdev
, conf
->mddev
);
3045 spin_lock_irq(&sh
->stripe_lock
);
3046 /* fail all writes first */
3047 bi
= sh
->dev
[i
].towrite
;
3048 sh
->dev
[i
].towrite
= NULL
;
3049 sh
->overwrite_disks
= 0;
3050 spin_unlock_irq(&sh
->stripe_lock
);
3054 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3055 wake_up(&conf
->wait_for_overlap
);
3057 while (bi
&& bi
->bi_iter
.bi_sector
<
3058 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3059 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3060 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3061 if (!raid5_dec_bi_active_stripes(bi
)) {
3062 md_write_end(conf
->mddev
);
3063 bi
->bi_next
= *return_bi
;
3069 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3070 STRIPE_SECTORS
, 0, 0);
3072 /* and fail all 'written' */
3073 bi
= sh
->dev
[i
].written
;
3074 sh
->dev
[i
].written
= NULL
;
3075 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3076 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3077 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3080 if (bi
) bitmap_end
= 1;
3081 while (bi
&& bi
->bi_iter
.bi_sector
<
3082 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3083 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3084 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3085 if (!raid5_dec_bi_active_stripes(bi
)) {
3086 md_write_end(conf
->mddev
);
3087 bi
->bi_next
= *return_bi
;
3093 /* fail any reads if this device is non-operational and
3094 * the data has not reached the cache yet.
3096 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3097 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3098 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3099 spin_lock_irq(&sh
->stripe_lock
);
3100 bi
= sh
->dev
[i
].toread
;
3101 sh
->dev
[i
].toread
= NULL
;
3102 spin_unlock_irq(&sh
->stripe_lock
);
3103 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3104 wake_up(&conf
->wait_for_overlap
);
3105 while (bi
&& bi
->bi_iter
.bi_sector
<
3106 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3107 struct bio
*nextbi
=
3108 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3109 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3110 if (!raid5_dec_bi_active_stripes(bi
)) {
3111 bi
->bi_next
= *return_bi
;
3118 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3119 STRIPE_SECTORS
, 0, 0);
3120 /* If we were in the middle of a write the parity block might
3121 * still be locked - so just clear all R5_LOCKED flags
3123 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3126 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3127 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3128 md_wakeup_thread(conf
->mddev
->thread
);
3132 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3133 struct stripe_head_state
*s
)
3138 BUG_ON(sh
->batch_head
);
3139 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3140 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3141 wake_up(&conf
->wait_for_overlap
);
3144 /* There is nothing more to do for sync/check/repair.
3145 * Don't even need to abort as that is handled elsewhere
3146 * if needed, and not always wanted e.g. if there is a known
3148 * For recover/replace we need to record a bad block on all
3149 * non-sync devices, or abort the recovery
3151 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3152 /* During recovery devices cannot be removed, so
3153 * locking and refcounting of rdevs is not needed
3155 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3156 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
3158 && !test_bit(Faulty
, &rdev
->flags
)
3159 && !test_bit(In_sync
, &rdev
->flags
)
3160 && !rdev_set_badblocks(rdev
, sh
->sector
,
3163 rdev
= conf
->disks
[i
].replacement
;
3165 && !test_bit(Faulty
, &rdev
->flags
)
3166 && !test_bit(In_sync
, &rdev
->flags
)
3167 && !rdev_set_badblocks(rdev
, sh
->sector
,
3172 conf
->recovery_disabled
=
3173 conf
->mddev
->recovery_disabled
;
3175 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3178 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3180 struct md_rdev
*rdev
;
3182 /* Doing recovery so rcu locking not required */
3183 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
3185 && !test_bit(Faulty
, &rdev
->flags
)
3186 && !test_bit(In_sync
, &rdev
->flags
)
3187 && (rdev
->recovery_offset
<= sh
->sector
3188 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3194 /* fetch_block - checks the given member device to see if its data needs
3195 * to be read or computed to satisfy a request.
3197 * Returns 1 when no more member devices need to be checked, otherwise returns
3198 * 0 to tell the loop in handle_stripe_fill to continue
3201 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3202 int disk_idx
, int disks
)
3204 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3205 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3206 &sh
->dev
[s
->failed_num
[1]] };
3210 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3211 test_bit(R5_UPTODATE
, &dev
->flags
))
3212 /* No point reading this as we already have it or have
3213 * decided to get it.
3218 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3219 /* We need this block to directly satisfy a request */
3222 if (s
->syncing
|| s
->expanding
||
3223 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3224 /* When syncing, or expanding we read everything.
3225 * When replacing, we need the replaced block.
3229 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3230 (s
->failed
>= 2 && fdev
[1]->toread
))
3231 /* If we want to read from a failed device, then
3232 * we need to actually read every other device.
3236 /* Sometimes neither read-modify-write nor reconstruct-write
3237 * cycles can work. In those cases we read every block we
3238 * can. Then the parity-update is certain to have enough to
3240 * This can only be a problem when we need to write something,
3241 * and some device has failed. If either of those tests
3242 * fail we need look no further.
3244 if (!s
->failed
|| !s
->to_write
)
3247 if (test_bit(R5_Insync
, &dev
->flags
) &&
3248 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3249 /* Pre-reads at not permitted until after short delay
3250 * to gather multiple requests. However if this
3251 * device is no Insync, the block could only be be computed
3252 * and there is no need to delay that.
3256 for (i
= 0; i
< s
->failed
; i
++) {
3257 if (fdev
[i
]->towrite
&&
3258 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3259 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3260 /* If we have a partial write to a failed
3261 * device, then we will need to reconstruct
3262 * the content of that device, so all other
3263 * devices must be read.
3268 /* If we are forced to do a reconstruct-write, either because
3269 * the current RAID6 implementation only supports that, or
3270 * or because parity cannot be trusted and we are currently
3271 * recovering it, there is extra need to be careful.
3272 * If one of the devices that we would need to read, because
3273 * it is not being overwritten (and maybe not written at all)
3274 * is missing/faulty, then we need to read everything we can.
3276 if (sh
->raid_conf
->level
!= 6 &&
3277 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3278 /* reconstruct-write isn't being forced */
3280 for (i
= 0; i
< s
->failed
; i
++) {
3281 if (!test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3282 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3289 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3290 int disk_idx
, int disks
)
3292 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3294 /* is the data in this block needed, and can we get it? */
3295 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3296 /* we would like to get this block, possibly by computing it,
3297 * otherwise read it if the backing disk is insync
3299 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3300 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3301 if ((s
->uptodate
== disks
- 1) &&
3302 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3303 disk_idx
== s
->failed_num
[1]))) {
3304 /* have disk failed, and we're requested to fetch it;
3307 pr_debug("Computing stripe %llu block %d\n",
3308 (unsigned long long)sh
->sector
, disk_idx
);
3309 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3310 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3311 set_bit(R5_Wantcompute
, &dev
->flags
);
3312 sh
->ops
.target
= disk_idx
;
3313 sh
->ops
.target2
= -1; /* no 2nd target */
3315 /* Careful: from this point on 'uptodate' is in the eye
3316 * of raid_run_ops which services 'compute' operations
3317 * before writes. R5_Wantcompute flags a block that will
3318 * be R5_UPTODATE by the time it is needed for a
3319 * subsequent operation.
3323 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3324 /* Computing 2-failure is *very* expensive; only
3325 * do it if failed >= 2
3328 for (other
= disks
; other
--; ) {
3329 if (other
== disk_idx
)
3331 if (!test_bit(R5_UPTODATE
,
3332 &sh
->dev
[other
].flags
))
3336 pr_debug("Computing stripe %llu blocks %d,%d\n",
3337 (unsigned long long)sh
->sector
,
3339 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3340 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3341 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3342 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3343 sh
->ops
.target
= disk_idx
;
3344 sh
->ops
.target2
= other
;
3348 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3349 set_bit(R5_LOCKED
, &dev
->flags
);
3350 set_bit(R5_Wantread
, &dev
->flags
);
3352 pr_debug("Reading block %d (sync=%d)\n",
3353 disk_idx
, s
->syncing
);
3361 * handle_stripe_fill - read or compute data to satisfy pending requests.
3363 static void handle_stripe_fill(struct stripe_head
*sh
,
3364 struct stripe_head_state
*s
,
3369 BUG_ON(sh
->batch_head
);
3370 /* look for blocks to read/compute, skip this if a compute
3371 * is already in flight, or if the stripe contents are in the
3372 * midst of changing due to a write
3374 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3375 !sh
->reconstruct_state
)
3376 for (i
= disks
; i
--; )
3377 if (fetch_block(sh
, s
, i
, disks
))
3379 set_bit(STRIPE_HANDLE
, &sh
->state
);
3382 /* handle_stripe_clean_event
3383 * any written block on an uptodate or failed drive can be returned.
3384 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3385 * never LOCKED, so we don't need to test 'failed' directly.
3387 static void handle_stripe_clean_event(struct r5conf
*conf
,
3388 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
3392 int discard_pending
= 0;
3393 struct stripe_head
*head_sh
= sh
;
3394 bool do_endio
= false;
3397 for (i
= disks
; i
--; )
3398 if (sh
->dev
[i
].written
) {
3400 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3401 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3402 test_bit(R5_Discard
, &dev
->flags
) ||
3403 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3404 /* We can return any write requests */
3405 struct bio
*wbi
, *wbi2
;
3406 pr_debug("Return write for disc %d\n", i
);
3407 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3408 clear_bit(R5_UPTODATE
, &dev
->flags
);
3409 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3410 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3415 dev
->page
= dev
->orig_page
;
3417 dev
->written
= NULL
;
3418 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3419 dev
->sector
+ STRIPE_SECTORS
) {
3420 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3421 if (!raid5_dec_bi_active_stripes(wbi
)) {
3422 md_write_end(conf
->mddev
);
3423 wbi
->bi_next
= *return_bi
;
3428 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3430 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3432 if (head_sh
->batch_head
) {
3433 sh
= list_first_entry(&sh
->batch_list
,
3436 if (sh
!= head_sh
) {
3443 } else if (test_bit(R5_Discard
, &dev
->flags
))
3444 discard_pending
= 1;
3445 WARN_ON(test_bit(R5_SkipCopy
, &dev
->flags
));
3446 WARN_ON(dev
->page
!= dev
->orig_page
);
3448 if (!discard_pending
&&
3449 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3450 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3451 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3452 if (sh
->qd_idx
>= 0) {
3453 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3454 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3456 /* now that discard is done we can proceed with any sync */
3457 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3459 * SCSI discard will change some bio fields and the stripe has
3460 * no updated data, so remove it from hash list and the stripe
3461 * will be reinitialized
3463 spin_lock_irq(&conf
->device_lock
);
3466 if (head_sh
->batch_head
) {
3467 sh
= list_first_entry(&sh
->batch_list
,
3468 struct stripe_head
, batch_list
);
3472 spin_unlock_irq(&conf
->device_lock
);
3475 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3476 set_bit(STRIPE_HANDLE
, &sh
->state
);
3480 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3481 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3482 md_wakeup_thread(conf
->mddev
->thread
);
3484 if (!head_sh
->batch_head
|| !do_endio
)
3486 for (i
= 0; i
< head_sh
->disks
; i
++) {
3487 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
3490 while (!list_empty(&head_sh
->batch_list
)) {
3492 sh
= list_first_entry(&head_sh
->batch_list
,
3493 struct stripe_head
, batch_list
);
3494 list_del_init(&sh
->batch_list
);
3496 set_mask_bits(&sh
->state
, ~STRIPE_EXPAND_SYNC_FLAG
,
3497 head_sh
->state
& ~((1 << STRIPE_ACTIVE
) |
3498 (1 << STRIPE_PREREAD_ACTIVE
) |
3499 STRIPE_EXPAND_SYNC_FLAG
));
3500 sh
->check_state
= head_sh
->check_state
;
3501 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
3502 for (i
= 0; i
< sh
->disks
; i
++) {
3503 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3505 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
;
3508 spin_lock_irq(&sh
->stripe_lock
);
3509 sh
->batch_head
= NULL
;
3510 spin_unlock_irq(&sh
->stripe_lock
);
3511 if (sh
->state
& STRIPE_EXPAND_SYNC_FLAG
)
3512 set_bit(STRIPE_HANDLE
, &sh
->state
);
3516 spin_lock_irq(&head_sh
->stripe_lock
);
3517 head_sh
->batch_head
= NULL
;
3518 spin_unlock_irq(&head_sh
->stripe_lock
);
3519 wake_up_nr(&conf
->wait_for_overlap
, wakeup_nr
);
3520 if (head_sh
->state
& STRIPE_EXPAND_SYNC_FLAG
)
3521 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
3524 static void handle_stripe_dirtying(struct r5conf
*conf
,
3525 struct stripe_head
*sh
,
3526 struct stripe_head_state
*s
,
3529 int rmw
= 0, rcw
= 0, i
;
3530 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3532 /* Check whether resync is now happening or should start.
3533 * If yes, then the array is dirty (after unclean shutdown or
3534 * initial creation), so parity in some stripes might be inconsistent.
3535 * In this case, we need to always do reconstruct-write, to ensure
3536 * that in case of drive failure or read-error correction, we
3537 * generate correct data from the parity.
3539 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3540 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3542 /* Calculate the real rcw later - for now make it
3543 * look like rcw is cheaper
3546 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3547 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3548 (unsigned long long)sh
->sector
);
3549 } else for (i
= disks
; i
--; ) {
3550 /* would I have to read this buffer for read_modify_write */
3551 struct r5dev
*dev
= &sh
->dev
[i
];
3552 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3553 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3554 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3555 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3556 if (test_bit(R5_Insync
, &dev
->flags
))
3559 rmw
+= 2*disks
; /* cannot read it */
3561 /* Would I have to read this buffer for reconstruct_write */
3562 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3563 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3564 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3565 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3566 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3567 if (test_bit(R5_Insync
, &dev
->flags
))
3573 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3574 (unsigned long long)sh
->sector
, rmw
, rcw
);
3575 set_bit(STRIPE_HANDLE
, &sh
->state
);
3576 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_ENABLE_RMW
)) && rmw
> 0) {
3577 /* prefer read-modify-write, but need to get some data */
3578 if (conf
->mddev
->queue
)
3579 blk_add_trace_msg(conf
->mddev
->queue
,
3580 "raid5 rmw %llu %d",
3581 (unsigned long long)sh
->sector
, rmw
);
3582 for (i
= disks
; i
--; ) {
3583 struct r5dev
*dev
= &sh
->dev
[i
];
3584 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3585 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3586 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3587 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3588 test_bit(R5_Insync
, &dev
->flags
)) {
3589 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3591 pr_debug("Read_old block %d for r-m-w\n",
3593 set_bit(R5_LOCKED
, &dev
->flags
);
3594 set_bit(R5_Wantread
, &dev
->flags
);
3597 set_bit(STRIPE_DELAYED
, &sh
->state
);
3598 set_bit(STRIPE_HANDLE
, &sh
->state
);
3603 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_ENABLE_RMW
)) && rcw
> 0) {
3604 /* want reconstruct write, but need to get some data */
3607 for (i
= disks
; i
--; ) {
3608 struct r5dev
*dev
= &sh
->dev
[i
];
3609 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3610 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3611 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3612 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3613 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3615 if (test_bit(R5_Insync
, &dev
->flags
) &&
3616 test_bit(STRIPE_PREREAD_ACTIVE
,
3618 pr_debug("Read_old block "
3619 "%d for Reconstruct\n", i
);
3620 set_bit(R5_LOCKED
, &dev
->flags
);
3621 set_bit(R5_Wantread
, &dev
->flags
);
3625 set_bit(STRIPE_DELAYED
, &sh
->state
);
3626 set_bit(STRIPE_HANDLE
, &sh
->state
);
3630 if (rcw
&& conf
->mddev
->queue
)
3631 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3632 (unsigned long long)sh
->sector
,
3633 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3636 if (rcw
> disks
&& rmw
> disks
&&
3637 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3638 set_bit(STRIPE_DELAYED
, &sh
->state
);
3640 /* now if nothing is locked, and if we have enough data,
3641 * we can start a write request
3643 /* since handle_stripe can be called at any time we need to handle the
3644 * case where a compute block operation has been submitted and then a
3645 * subsequent call wants to start a write request. raid_run_ops only
3646 * handles the case where compute block and reconstruct are requested
3647 * simultaneously. If this is not the case then new writes need to be
3648 * held off until the compute completes.
3650 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
3651 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
3652 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
3653 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
3656 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
3657 struct stripe_head_state
*s
, int disks
)
3659 struct r5dev
*dev
= NULL
;
3661 BUG_ON(sh
->batch_head
);
3662 set_bit(STRIPE_HANDLE
, &sh
->state
);
3664 switch (sh
->check_state
) {
3665 case check_state_idle
:
3666 /* start a new check operation if there are no failures */
3667 if (s
->failed
== 0) {
3668 BUG_ON(s
->uptodate
!= disks
);
3669 sh
->check_state
= check_state_run
;
3670 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3671 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3675 dev
= &sh
->dev
[s
->failed_num
[0]];
3677 case check_state_compute_result
:
3678 sh
->check_state
= check_state_idle
;
3680 dev
= &sh
->dev
[sh
->pd_idx
];
3682 /* check that a write has not made the stripe insync */
3683 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3686 /* either failed parity check, or recovery is happening */
3687 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3688 BUG_ON(s
->uptodate
!= disks
);
3690 set_bit(R5_LOCKED
, &dev
->flags
);
3692 set_bit(R5_Wantwrite
, &dev
->flags
);
3694 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3695 set_bit(STRIPE_INSYNC
, &sh
->state
);
3697 case check_state_run
:
3698 break; /* we will be called again upon completion */
3699 case check_state_check_result
:
3700 sh
->check_state
= check_state_idle
;
3702 /* if a failure occurred during the check operation, leave
3703 * STRIPE_INSYNC not set and let the stripe be handled again
3708 /* handle a successful check operation, if parity is correct
3709 * we are done. Otherwise update the mismatch count and repair
3710 * parity if !MD_RECOVERY_CHECK
3712 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3713 /* parity is correct (on disc,
3714 * not in buffer any more)
3716 set_bit(STRIPE_INSYNC
, &sh
->state
);
3718 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3719 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3720 /* don't try to repair!! */
3721 set_bit(STRIPE_INSYNC
, &sh
->state
);
3723 sh
->check_state
= check_state_compute_run
;
3724 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3725 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3726 set_bit(R5_Wantcompute
,
3727 &sh
->dev
[sh
->pd_idx
].flags
);
3728 sh
->ops
.target
= sh
->pd_idx
;
3729 sh
->ops
.target2
= -1;
3734 case check_state_compute_run
:
3737 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3738 __func__
, sh
->check_state
,
3739 (unsigned long long) sh
->sector
);
3744 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3745 struct stripe_head_state
*s
,
3748 int pd_idx
= sh
->pd_idx
;
3749 int qd_idx
= sh
->qd_idx
;
3752 BUG_ON(sh
->batch_head
);
3753 set_bit(STRIPE_HANDLE
, &sh
->state
);
3755 BUG_ON(s
->failed
> 2);
3757 /* Want to check and possibly repair P and Q.
3758 * However there could be one 'failed' device, in which
3759 * case we can only check one of them, possibly using the
3760 * other to generate missing data
3763 switch (sh
->check_state
) {
3764 case check_state_idle
:
3765 /* start a new check operation if there are < 2 failures */
3766 if (s
->failed
== s
->q_failed
) {
3767 /* The only possible failed device holds Q, so it
3768 * makes sense to check P (If anything else were failed,
3769 * we would have used P to recreate it).
3771 sh
->check_state
= check_state_run
;
3773 if (!s
->q_failed
&& s
->failed
< 2) {
3774 /* Q is not failed, and we didn't use it to generate
3775 * anything, so it makes sense to check it
3777 if (sh
->check_state
== check_state_run
)
3778 sh
->check_state
= check_state_run_pq
;
3780 sh
->check_state
= check_state_run_q
;
3783 /* discard potentially stale zero_sum_result */
3784 sh
->ops
.zero_sum_result
= 0;
3786 if (sh
->check_state
== check_state_run
) {
3787 /* async_xor_zero_sum destroys the contents of P */
3788 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3791 if (sh
->check_state
>= check_state_run
&&
3792 sh
->check_state
<= check_state_run_pq
) {
3793 /* async_syndrome_zero_sum preserves P and Q, so
3794 * no need to mark them !uptodate here
3796 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3800 /* we have 2-disk failure */
3801 BUG_ON(s
->failed
!= 2);
3803 case check_state_compute_result
:
3804 sh
->check_state
= check_state_idle
;
3806 /* check that a write has not made the stripe insync */
3807 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3810 /* now write out any block on a failed drive,
3811 * or P or Q if they were recomputed
3813 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3814 if (s
->failed
== 2) {
3815 dev
= &sh
->dev
[s
->failed_num
[1]];
3817 set_bit(R5_LOCKED
, &dev
->flags
);
3818 set_bit(R5_Wantwrite
, &dev
->flags
);
3820 if (s
->failed
>= 1) {
3821 dev
= &sh
->dev
[s
->failed_num
[0]];
3823 set_bit(R5_LOCKED
, &dev
->flags
);
3824 set_bit(R5_Wantwrite
, &dev
->flags
);
3826 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3827 dev
= &sh
->dev
[pd_idx
];
3829 set_bit(R5_LOCKED
, &dev
->flags
);
3830 set_bit(R5_Wantwrite
, &dev
->flags
);
3832 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3833 dev
= &sh
->dev
[qd_idx
];
3835 set_bit(R5_LOCKED
, &dev
->flags
);
3836 set_bit(R5_Wantwrite
, &dev
->flags
);
3838 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3840 set_bit(STRIPE_INSYNC
, &sh
->state
);
3842 case check_state_run
:
3843 case check_state_run_q
:
3844 case check_state_run_pq
:
3845 break; /* we will be called again upon completion */
3846 case check_state_check_result
:
3847 sh
->check_state
= check_state_idle
;
3849 /* handle a successful check operation, if parity is correct
3850 * we are done. Otherwise update the mismatch count and repair
3851 * parity if !MD_RECOVERY_CHECK
3853 if (sh
->ops
.zero_sum_result
== 0) {
3854 /* both parities are correct */
3856 set_bit(STRIPE_INSYNC
, &sh
->state
);
3858 /* in contrast to the raid5 case we can validate
3859 * parity, but still have a failure to write
3862 sh
->check_state
= check_state_compute_result
;
3863 /* Returning at this point means that we may go
3864 * off and bring p and/or q uptodate again so
3865 * we make sure to check zero_sum_result again
3866 * to verify if p or q need writeback
3870 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3871 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3872 /* don't try to repair!! */
3873 set_bit(STRIPE_INSYNC
, &sh
->state
);
3875 int *target
= &sh
->ops
.target
;
3877 sh
->ops
.target
= -1;
3878 sh
->ops
.target2
= -1;
3879 sh
->check_state
= check_state_compute_run
;
3880 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3881 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3882 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3883 set_bit(R5_Wantcompute
,
3884 &sh
->dev
[pd_idx
].flags
);
3886 target
= &sh
->ops
.target2
;
3889 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3890 set_bit(R5_Wantcompute
,
3891 &sh
->dev
[qd_idx
].flags
);
3898 case check_state_compute_run
:
3901 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3902 __func__
, sh
->check_state
,
3903 (unsigned long long) sh
->sector
);
3908 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3912 /* We have read all the blocks in this stripe and now we need to
3913 * copy some of them into a target stripe for expand.
3915 struct dma_async_tx_descriptor
*tx
= NULL
;
3916 BUG_ON(sh
->batch_head
);
3917 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3918 for (i
= 0; i
< sh
->disks
; i
++)
3919 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3921 struct stripe_head
*sh2
;
3922 struct async_submit_ctl submit
;
3924 sector_t bn
= compute_blocknr(sh
, i
, 1);
3925 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3927 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3929 /* so far only the early blocks of this stripe
3930 * have been requested. When later blocks
3931 * get requested, we will try again
3934 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3935 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3936 /* must have already done this block */
3937 release_stripe(sh2
);
3941 /* place all the copies on one channel */
3942 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3943 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3944 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3947 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3948 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3949 for (j
= 0; j
< conf
->raid_disks
; j
++)
3950 if (j
!= sh2
->pd_idx
&&
3952 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3954 if (j
== conf
->raid_disks
) {
3955 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3956 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3958 release_stripe(sh2
);
3961 /* done submitting copies, wait for them to complete */
3962 async_tx_quiesce(&tx
);
3966 * handle_stripe - do things to a stripe.
3968 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3969 * state of various bits to see what needs to be done.
3971 * return some read requests which now have data
3972 * return some write requests which are safely on storage
3973 * schedule a read on some buffers
3974 * schedule a write of some buffers
3975 * return confirmation of parity correctness
3979 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3981 struct r5conf
*conf
= sh
->raid_conf
;
3982 int disks
= sh
->disks
;
3985 int do_recovery
= 0;
3987 memset(s
, 0, sizeof(*s
));
3989 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
3990 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
3991 s
->failed_num
[0] = -1;
3992 s
->failed_num
[1] = -1;
3994 /* Now to look around and see what can be done */
3996 for (i
=disks
; i
--; ) {
3997 struct md_rdev
*rdev
;
4004 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4006 dev
->toread
, dev
->towrite
, dev
->written
);
4007 /* maybe we can reply to a read
4009 * new wantfill requests are only permitted while
4010 * ops_complete_biofill is guaranteed to be inactive
4012 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4013 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4014 set_bit(R5_Wantfill
, &dev
->flags
);
4016 /* now count some things */
4017 if (test_bit(R5_LOCKED
, &dev
->flags
))
4019 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4021 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4023 BUG_ON(s
->compute
> 2);
4026 if (test_bit(R5_Wantfill
, &dev
->flags
))
4028 else if (dev
->toread
)
4032 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4037 /* Prefer to use the replacement for reads, but only
4038 * if it is recovered enough and has no bad blocks.
4040 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4041 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4042 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4043 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4044 &first_bad
, &bad_sectors
))
4045 set_bit(R5_ReadRepl
, &dev
->flags
);
4048 set_bit(R5_NeedReplace
, &dev
->flags
);
4049 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4050 clear_bit(R5_ReadRepl
, &dev
->flags
);
4052 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4055 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4056 &first_bad
, &bad_sectors
);
4057 if (s
->blocked_rdev
== NULL
4058 && (test_bit(Blocked
, &rdev
->flags
)
4061 set_bit(BlockedBadBlocks
,
4063 s
->blocked_rdev
= rdev
;
4064 atomic_inc(&rdev
->nr_pending
);
4067 clear_bit(R5_Insync
, &dev
->flags
);
4071 /* also not in-sync */
4072 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4073 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4074 /* treat as in-sync, but with a read error
4075 * which we can now try to correct
4077 set_bit(R5_Insync
, &dev
->flags
);
4078 set_bit(R5_ReadError
, &dev
->flags
);
4080 } else if (test_bit(In_sync
, &rdev
->flags
))
4081 set_bit(R5_Insync
, &dev
->flags
);
4082 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4083 /* in sync if before recovery_offset */
4084 set_bit(R5_Insync
, &dev
->flags
);
4085 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4086 test_bit(R5_Expanded
, &dev
->flags
))
4087 /* If we've reshaped into here, we assume it is Insync.
4088 * We will shortly update recovery_offset to make
4091 set_bit(R5_Insync
, &dev
->flags
);
4093 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4094 /* This flag does not apply to '.replacement'
4095 * only to .rdev, so make sure to check that*/
4096 struct md_rdev
*rdev2
= rcu_dereference(
4097 conf
->disks
[i
].rdev
);
4099 clear_bit(R5_Insync
, &dev
->flags
);
4100 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4101 s
->handle_bad_blocks
= 1;
4102 atomic_inc(&rdev2
->nr_pending
);
4104 clear_bit(R5_WriteError
, &dev
->flags
);
4106 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4107 /* This flag does not apply to '.replacement'
4108 * only to .rdev, so make sure to check that*/
4109 struct md_rdev
*rdev2
= rcu_dereference(
4110 conf
->disks
[i
].rdev
);
4111 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4112 s
->handle_bad_blocks
= 1;
4113 atomic_inc(&rdev2
->nr_pending
);
4115 clear_bit(R5_MadeGood
, &dev
->flags
);
4117 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4118 struct md_rdev
*rdev2
= rcu_dereference(
4119 conf
->disks
[i
].replacement
);
4120 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4121 s
->handle_bad_blocks
= 1;
4122 atomic_inc(&rdev2
->nr_pending
);
4124 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4126 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4127 /* The ReadError flag will just be confusing now */
4128 clear_bit(R5_ReadError
, &dev
->flags
);
4129 clear_bit(R5_ReWrite
, &dev
->flags
);
4131 if (test_bit(R5_ReadError
, &dev
->flags
))
4132 clear_bit(R5_Insync
, &dev
->flags
);
4133 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4135 s
->failed_num
[s
->failed
] = i
;
4137 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4141 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4142 /* If there is a failed device being replaced,
4143 * we must be recovering.
4144 * else if we are after recovery_cp, we must be syncing
4145 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4146 * else we can only be replacing
4147 * sync and recovery both need to read all devices, and so
4148 * use the same flag.
4151 sh
->sector
>= conf
->mddev
->recovery_cp
||
4152 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4160 static int clear_batch_ready(struct stripe_head
*sh
)
4162 struct stripe_head
*tmp
;
4163 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4165 spin_lock(&sh
->stripe_lock
);
4166 if (!sh
->batch_head
) {
4167 spin_unlock(&sh
->stripe_lock
);
4172 * this stripe could be added to a batch list before we check
4173 * BATCH_READY, skips it
4175 if (sh
->batch_head
!= sh
) {
4176 spin_unlock(&sh
->stripe_lock
);
4179 spin_lock(&sh
->batch_lock
);
4180 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4181 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4182 spin_unlock(&sh
->batch_lock
);
4183 spin_unlock(&sh
->stripe_lock
);
4186 * BATCH_READY is cleared, no new stripes can be added.
4187 * batch_list can be accessed without lock
4192 static void check_break_stripe_batch_list(struct stripe_head
*sh
)
4194 struct stripe_head
*head_sh
, *next
;
4197 if (!test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4202 sh
= list_first_entry(&sh
->batch_list
,
4203 struct stripe_head
, batch_list
);
4204 BUG_ON(sh
== head_sh
);
4205 } while (!test_bit(STRIPE_DEGRADED
, &sh
->state
));
4207 while (sh
!= head_sh
) {
4208 next
= list_first_entry(&sh
->batch_list
,
4209 struct stripe_head
, batch_list
);
4210 list_del_init(&sh
->batch_list
);
4212 set_mask_bits(&sh
->state
, ~STRIPE_EXPAND_SYNC_FLAG
,
4213 head_sh
->state
& ~((1 << STRIPE_ACTIVE
) |
4214 (1 << STRIPE_PREREAD_ACTIVE
) |
4215 (1 << STRIPE_DEGRADED
) |
4216 STRIPE_EXPAND_SYNC_FLAG
));
4217 sh
->check_state
= head_sh
->check_state
;
4218 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4219 for (i
= 0; i
< sh
->disks
; i
++)
4220 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4221 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4223 spin_lock_irq(&sh
->stripe_lock
);
4224 sh
->batch_head
= NULL
;
4225 spin_unlock_irq(&sh
->stripe_lock
);
4227 set_bit(STRIPE_HANDLE
, &sh
->state
);
4234 static void handle_stripe(struct stripe_head
*sh
)
4236 struct stripe_head_state s
;
4237 struct r5conf
*conf
= sh
->raid_conf
;
4240 int disks
= sh
->disks
;
4241 struct r5dev
*pdev
, *qdev
;
4243 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4244 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4245 /* already being handled, ensure it gets handled
4246 * again when current action finishes */
4247 set_bit(STRIPE_HANDLE
, &sh
->state
);
4251 if (clear_batch_ready(sh
) ) {
4252 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4256 check_break_stripe_batch_list(sh
);
4258 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4259 spin_lock(&sh
->stripe_lock
);
4260 /* Cannot process 'sync' concurrently with 'discard' */
4261 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4262 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4263 set_bit(STRIPE_SYNCING
, &sh
->state
);
4264 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4265 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4267 spin_unlock(&sh
->stripe_lock
);
4269 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4271 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4272 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4273 (unsigned long long)sh
->sector
, sh
->state
,
4274 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4275 sh
->check_state
, sh
->reconstruct_state
);
4277 analyse_stripe(sh
, &s
);
4279 if (s
.handle_bad_blocks
) {
4280 set_bit(STRIPE_HANDLE
, &sh
->state
);
4284 if (unlikely(s
.blocked_rdev
)) {
4285 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4286 s
.replacing
|| s
.to_write
|| s
.written
) {
4287 set_bit(STRIPE_HANDLE
, &sh
->state
);
4290 /* There is nothing for the blocked_rdev to block */
4291 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4292 s
.blocked_rdev
= NULL
;
4295 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4296 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4297 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4300 pr_debug("locked=%d uptodate=%d to_read=%d"
4301 " to_write=%d failed=%d failed_num=%d,%d\n",
4302 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4303 s
.failed_num
[0], s
.failed_num
[1]);
4304 /* check if the array has lost more than max_degraded devices and,
4305 * if so, some requests might need to be failed.
4307 if (s
.failed
> conf
->max_degraded
) {
4308 sh
->check_state
= 0;
4309 sh
->reconstruct_state
= 0;
4310 if (s
.to_read
+s
.to_write
+s
.written
)
4311 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
4312 if (s
.syncing
+ s
.replacing
)
4313 handle_failed_sync(conf
, sh
, &s
);
4316 /* Now we check to see if any write operations have recently
4320 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4322 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4323 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4324 sh
->reconstruct_state
= reconstruct_state_idle
;
4326 /* All the 'written' buffers and the parity block are ready to
4327 * be written back to disk
4329 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4330 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4331 BUG_ON(sh
->qd_idx
>= 0 &&
4332 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4333 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4334 for (i
= disks
; i
--; ) {
4335 struct r5dev
*dev
= &sh
->dev
[i
];
4336 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4337 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4339 pr_debug("Writing block %d\n", i
);
4340 set_bit(R5_Wantwrite
, &dev
->flags
);
4345 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4346 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4348 set_bit(STRIPE_INSYNC
, &sh
->state
);
4351 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4352 s
.dec_preread_active
= 1;
4356 * might be able to return some write requests if the parity blocks
4357 * are safe, or on a failed drive
4359 pdev
= &sh
->dev
[sh
->pd_idx
];
4360 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4361 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4362 qdev
= &sh
->dev
[sh
->qd_idx
];
4363 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4364 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4368 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4369 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4370 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4371 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4372 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4373 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4374 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4375 test_bit(R5_Discard
, &qdev
->flags
))))))
4376 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
4378 /* Now we might consider reading some blocks, either to check/generate
4379 * parity, or to satisfy requests
4380 * or to load a block that is being partially written.
4382 if (s
.to_read
|| s
.non_overwrite
4383 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4384 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4387 handle_stripe_fill(sh
, &s
, disks
);
4389 /* Now to consider new write requests and what else, if anything
4390 * should be read. We do not handle new writes when:
4391 * 1/ A 'write' operation (copy+xor) is already in flight.
4392 * 2/ A 'check' operation is in flight, as it may clobber the parity
4395 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
4396 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4398 /* maybe we need to check and possibly fix the parity for this stripe
4399 * Any reads will already have been scheduled, so we just see if enough
4400 * data is available. The parity check is held off while parity
4401 * dependent operations are in flight.
4403 if (sh
->check_state
||
4404 (s
.syncing
&& s
.locked
== 0 &&
4405 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4406 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4407 if (conf
->level
== 6)
4408 handle_parity_checks6(conf
, sh
, &s
, disks
);
4410 handle_parity_checks5(conf
, sh
, &s
, disks
);
4413 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4414 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4415 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4416 /* Write out to replacement devices where possible */
4417 for (i
= 0; i
< conf
->raid_disks
; i
++)
4418 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4419 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4420 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4421 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4425 set_bit(STRIPE_INSYNC
, &sh
->state
);
4426 set_bit(STRIPE_REPLACED
, &sh
->state
);
4428 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4429 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4430 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4431 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4432 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4433 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4434 wake_up(&conf
->wait_for_overlap
);
4437 /* If the failed drives are just a ReadError, then we might need
4438 * to progress the repair/check process
4440 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4441 for (i
= 0; i
< s
.failed
; i
++) {
4442 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4443 if (test_bit(R5_ReadError
, &dev
->flags
)
4444 && !test_bit(R5_LOCKED
, &dev
->flags
)
4445 && test_bit(R5_UPTODATE
, &dev
->flags
)
4447 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4448 set_bit(R5_Wantwrite
, &dev
->flags
);
4449 set_bit(R5_ReWrite
, &dev
->flags
);
4450 set_bit(R5_LOCKED
, &dev
->flags
);
4453 /* let's read it back */
4454 set_bit(R5_Wantread
, &dev
->flags
);
4455 set_bit(R5_LOCKED
, &dev
->flags
);
4461 /* Finish reconstruct operations initiated by the expansion process */
4462 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4463 struct stripe_head
*sh_src
4464 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4465 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4466 /* sh cannot be written until sh_src has been read.
4467 * so arrange for sh to be delayed a little
4469 set_bit(STRIPE_DELAYED
, &sh
->state
);
4470 set_bit(STRIPE_HANDLE
, &sh
->state
);
4471 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4473 atomic_inc(&conf
->preread_active_stripes
);
4474 release_stripe(sh_src
);
4478 release_stripe(sh_src
);
4480 sh
->reconstruct_state
= reconstruct_state_idle
;
4481 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4482 for (i
= conf
->raid_disks
; i
--; ) {
4483 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4484 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4489 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4490 !sh
->reconstruct_state
) {
4491 /* Need to write out all blocks after computing parity */
4492 sh
->disks
= conf
->raid_disks
;
4493 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4494 schedule_reconstruction(sh
, &s
, 1, 1);
4495 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4496 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4497 atomic_dec(&conf
->reshape_stripes
);
4498 wake_up(&conf
->wait_for_overlap
);
4499 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4502 if (s
.expanding
&& s
.locked
== 0 &&
4503 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4504 handle_stripe_expansion(conf
, sh
);
4507 /* wait for this device to become unblocked */
4508 if (unlikely(s
.blocked_rdev
)) {
4509 if (conf
->mddev
->external
)
4510 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4513 /* Internal metadata will immediately
4514 * be written by raid5d, so we don't
4515 * need to wait here.
4517 rdev_dec_pending(s
.blocked_rdev
,
4521 if (s
.handle_bad_blocks
)
4522 for (i
= disks
; i
--; ) {
4523 struct md_rdev
*rdev
;
4524 struct r5dev
*dev
= &sh
->dev
[i
];
4525 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4526 /* We own a safe reference to the rdev */
4527 rdev
= conf
->disks
[i
].rdev
;
4528 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4530 md_error(conf
->mddev
, rdev
);
4531 rdev_dec_pending(rdev
, conf
->mddev
);
4533 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4534 rdev
= conf
->disks
[i
].rdev
;
4535 rdev_clear_badblocks(rdev
, sh
->sector
,
4537 rdev_dec_pending(rdev
, conf
->mddev
);
4539 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4540 rdev
= conf
->disks
[i
].replacement
;
4542 /* rdev have been moved down */
4543 rdev
= conf
->disks
[i
].rdev
;
4544 rdev_clear_badblocks(rdev
, sh
->sector
,
4546 rdev_dec_pending(rdev
, conf
->mddev
);
4551 raid_run_ops(sh
, s
.ops_request
);
4555 if (s
.dec_preread_active
) {
4556 /* We delay this until after ops_run_io so that if make_request
4557 * is waiting on a flush, it won't continue until the writes
4558 * have actually been submitted.
4560 atomic_dec(&conf
->preread_active_stripes
);
4561 if (atomic_read(&conf
->preread_active_stripes
) <
4563 md_wakeup_thread(conf
->mddev
->thread
);
4566 return_io(s
.return_bi
);
4568 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4571 static void raid5_activate_delayed(struct r5conf
*conf
)
4573 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
4574 while (!list_empty(&conf
->delayed_list
)) {
4575 struct list_head
*l
= conf
->delayed_list
.next
;
4576 struct stripe_head
*sh
;
4577 sh
= list_entry(l
, struct stripe_head
, lru
);
4579 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4580 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4581 atomic_inc(&conf
->preread_active_stripes
);
4582 list_add_tail(&sh
->lru
, &conf
->hold_list
);
4583 raid5_wakeup_stripe_thread(sh
);
4588 static void activate_bit_delay(struct r5conf
*conf
,
4589 struct list_head
*temp_inactive_list
)
4591 /* device_lock is held */
4592 struct list_head head
;
4593 list_add(&head
, &conf
->bitmap_list
);
4594 list_del_init(&conf
->bitmap_list
);
4595 while (!list_empty(&head
)) {
4596 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
4598 list_del_init(&sh
->lru
);
4599 atomic_inc(&sh
->count
);
4600 hash
= sh
->hash_lock_index
;
4601 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
4605 static int raid5_congested(struct mddev
*mddev
, int bits
)
4607 struct r5conf
*conf
= mddev
->private;
4609 /* No difference between reads and writes. Just check
4610 * how busy the stripe_cache is
4613 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
4617 if (atomic_read(&conf
->empty_inactive_list_nr
))
4623 /* We want read requests to align with chunks where possible,
4624 * but write requests don't need to.
4626 static int raid5_mergeable_bvec(struct mddev
*mddev
,
4627 struct bvec_merge_data
*bvm
,
4628 struct bio_vec
*biovec
)
4630 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
4632 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4633 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
4636 * always allow writes to be mergeable, read as well if array
4637 * is degraded as we'll go through stripe cache anyway.
4639 if ((bvm
->bi_rw
& 1) == WRITE
|| mddev
->degraded
)
4640 return biovec
->bv_len
;
4642 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4643 chunk_sectors
= mddev
->new_chunk_sectors
;
4644 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
4645 if (max
< 0) max
= 0;
4646 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
4647 return biovec
->bv_len
;
4652 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
4654 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
4655 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4656 unsigned int bio_sectors
= bio_sectors(bio
);
4658 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4659 chunk_sectors
= mddev
->new_chunk_sectors
;
4660 return chunk_sectors
>=
4661 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
4665 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4666 * later sampled by raid5d.
4668 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
4670 unsigned long flags
;
4672 spin_lock_irqsave(&conf
->device_lock
, flags
);
4674 bi
->bi_next
= conf
->retry_read_aligned_list
;
4675 conf
->retry_read_aligned_list
= bi
;
4677 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4678 md_wakeup_thread(conf
->mddev
->thread
);
4681 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
4685 bi
= conf
->retry_read_aligned
;
4687 conf
->retry_read_aligned
= NULL
;
4690 bi
= conf
->retry_read_aligned_list
;
4692 conf
->retry_read_aligned_list
= bi
->bi_next
;
4695 * this sets the active strip count to 1 and the processed
4696 * strip count to zero (upper 8 bits)
4698 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
4705 * The "raid5_align_endio" should check if the read succeeded and if it
4706 * did, call bio_endio on the original bio (having bio_put the new bio
4708 * If the read failed..
4710 static void raid5_align_endio(struct bio
*bi
, int error
)
4712 struct bio
* raid_bi
= bi
->bi_private
;
4713 struct mddev
*mddev
;
4714 struct r5conf
*conf
;
4715 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4716 struct md_rdev
*rdev
;
4720 rdev
= (void*)raid_bi
->bi_next
;
4721 raid_bi
->bi_next
= NULL
;
4722 mddev
= rdev
->mddev
;
4723 conf
= mddev
->private;
4725 rdev_dec_pending(rdev
, conf
->mddev
);
4727 if (!error
&& uptodate
) {
4728 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
4730 bio_endio(raid_bi
, 0);
4731 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4732 wake_up(&conf
->wait_for_stripe
);
4736 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4738 add_bio_to_retry(raid_bi
, conf
);
4741 static int bio_fits_rdev(struct bio
*bi
)
4743 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
4745 if (bio_sectors(bi
) > queue_max_sectors(q
))
4747 blk_recount_segments(q
, bi
);
4748 if (bi
->bi_phys_segments
> queue_max_segments(q
))
4751 if (q
->merge_bvec_fn
)
4752 /* it's too hard to apply the merge_bvec_fn at this stage,
4760 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
4762 struct r5conf
*conf
= mddev
->private;
4764 struct bio
* align_bi
;
4765 struct md_rdev
*rdev
;
4766 sector_t end_sector
;
4768 if (!in_chunk_boundary(mddev
, raid_bio
)) {
4769 pr_debug("chunk_aligned_read : non aligned\n");
4773 * use bio_clone_mddev to make a copy of the bio
4775 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
4779 * set bi_end_io to a new function, and set bi_private to the
4782 align_bi
->bi_end_io
= raid5_align_endio
;
4783 align_bi
->bi_private
= raid_bio
;
4787 align_bi
->bi_iter
.bi_sector
=
4788 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
4791 end_sector
= bio_end_sector(align_bi
);
4793 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
4794 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
4795 rdev
->recovery_offset
< end_sector
) {
4796 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
4798 (test_bit(Faulty
, &rdev
->flags
) ||
4799 !(test_bit(In_sync
, &rdev
->flags
) ||
4800 rdev
->recovery_offset
>= end_sector
)))
4807 atomic_inc(&rdev
->nr_pending
);
4809 raid_bio
->bi_next
= (void*)rdev
;
4810 align_bi
->bi_bdev
= rdev
->bdev
;
4811 __clear_bit(BIO_SEG_VALID
, &align_bi
->bi_flags
);
4813 if (!bio_fits_rdev(align_bi
) ||
4814 is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
4815 bio_sectors(align_bi
),
4816 &first_bad
, &bad_sectors
)) {
4817 /* too big in some way, or has a known bad block */
4819 rdev_dec_pending(rdev
, mddev
);
4823 /* No reshape active, so we can trust rdev->data_offset */
4824 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
4826 spin_lock_irq(&conf
->device_lock
);
4827 wait_event_lock_irq(conf
->wait_for_stripe
,
4830 atomic_inc(&conf
->active_aligned_reads
);
4831 spin_unlock_irq(&conf
->device_lock
);
4834 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4835 align_bi
, disk_devt(mddev
->gendisk
),
4836 raid_bio
->bi_iter
.bi_sector
);
4837 generic_make_request(align_bi
);
4846 /* __get_priority_stripe - get the next stripe to process
4848 * Full stripe writes are allowed to pass preread active stripes up until
4849 * the bypass_threshold is exceeded. In general the bypass_count
4850 * increments when the handle_list is handled before the hold_list; however, it
4851 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4852 * stripe with in flight i/o. The bypass_count will be reset when the
4853 * head of the hold_list has changed, i.e. the head was promoted to the
4856 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
4858 struct stripe_head
*sh
= NULL
, *tmp
;
4859 struct list_head
*handle_list
= NULL
;
4860 struct r5worker_group
*wg
= NULL
;
4862 if (conf
->worker_cnt_per_group
== 0) {
4863 handle_list
= &conf
->handle_list
;
4864 } else if (group
!= ANY_GROUP
) {
4865 handle_list
= &conf
->worker_groups
[group
].handle_list
;
4866 wg
= &conf
->worker_groups
[group
];
4869 for (i
= 0; i
< conf
->group_cnt
; i
++) {
4870 handle_list
= &conf
->worker_groups
[i
].handle_list
;
4871 wg
= &conf
->worker_groups
[i
];
4872 if (!list_empty(handle_list
))
4877 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4879 list_empty(handle_list
) ? "empty" : "busy",
4880 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4881 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4883 if (!list_empty(handle_list
)) {
4884 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
4886 if (list_empty(&conf
->hold_list
))
4887 conf
->bypass_count
= 0;
4888 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4889 if (conf
->hold_list
.next
== conf
->last_hold
)
4890 conf
->bypass_count
++;
4892 conf
->last_hold
= conf
->hold_list
.next
;
4893 conf
->bypass_count
-= conf
->bypass_threshold
;
4894 if (conf
->bypass_count
< 0)
4895 conf
->bypass_count
= 0;
4898 } else if (!list_empty(&conf
->hold_list
) &&
4899 ((conf
->bypass_threshold
&&
4900 conf
->bypass_count
> conf
->bypass_threshold
) ||
4901 atomic_read(&conf
->pending_full_writes
) == 0)) {
4903 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
4904 if (conf
->worker_cnt_per_group
== 0 ||
4905 group
== ANY_GROUP
||
4906 !cpu_online(tmp
->cpu
) ||
4907 cpu_to_group(tmp
->cpu
) == group
) {
4914 conf
->bypass_count
-= conf
->bypass_threshold
;
4915 if (conf
->bypass_count
< 0)
4916 conf
->bypass_count
= 0;
4928 list_del_init(&sh
->lru
);
4929 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
4933 struct raid5_plug_cb
{
4934 struct blk_plug_cb cb
;
4935 struct list_head list
;
4936 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
4939 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4941 struct raid5_plug_cb
*cb
= container_of(
4942 blk_cb
, struct raid5_plug_cb
, cb
);
4943 struct stripe_head
*sh
;
4944 struct mddev
*mddev
= cb
->cb
.data
;
4945 struct r5conf
*conf
= mddev
->private;
4949 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4950 spin_lock_irq(&conf
->device_lock
);
4951 while (!list_empty(&cb
->list
)) {
4952 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4953 list_del_init(&sh
->lru
);
4955 * avoid race release_stripe_plug() sees
4956 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4957 * is still in our list
4959 smp_mb__before_atomic();
4960 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4962 * STRIPE_ON_RELEASE_LIST could be set here. In that
4963 * case, the count is always > 1 here
4965 hash
= sh
->hash_lock_index
;
4966 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
4969 spin_unlock_irq(&conf
->device_lock
);
4971 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
4972 NR_STRIPE_HASH_LOCKS
);
4974 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
4978 static void release_stripe_plug(struct mddev
*mddev
,
4979 struct stripe_head
*sh
)
4981 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4982 raid5_unplug
, mddev
,
4983 sizeof(struct raid5_plug_cb
));
4984 struct raid5_plug_cb
*cb
;
4991 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4993 if (cb
->list
.next
== NULL
) {
4995 INIT_LIST_HEAD(&cb
->list
);
4996 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
4997 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5000 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5001 list_add_tail(&sh
->lru
, &cb
->list
);
5006 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5008 struct r5conf
*conf
= mddev
->private;
5009 sector_t logical_sector
, last_sector
;
5010 struct stripe_head
*sh
;
5014 if (mddev
->reshape_position
!= MaxSector
)
5015 /* Skip discard while reshape is happening */
5018 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5019 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5022 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5024 stripe_sectors
= conf
->chunk_sectors
*
5025 (conf
->raid_disks
- conf
->max_degraded
);
5026 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5028 sector_div(last_sector
, stripe_sectors
);
5030 logical_sector
*= conf
->chunk_sectors
;
5031 last_sector
*= conf
->chunk_sectors
;
5033 for (; logical_sector
< last_sector
;
5034 logical_sector
+= STRIPE_SECTORS
) {
5038 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5039 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5040 TASK_UNINTERRUPTIBLE
);
5041 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5042 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5047 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5048 spin_lock_irq(&sh
->stripe_lock
);
5049 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5050 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5052 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5053 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5054 spin_unlock_irq(&sh
->stripe_lock
);
5060 set_bit(STRIPE_DISCARD
, &sh
->state
);
5061 finish_wait(&conf
->wait_for_overlap
, &w
);
5062 sh
->overwrite_disks
= 0;
5063 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5064 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5066 sh
->dev
[d
].towrite
= bi
;
5067 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5068 raid5_inc_bi_active_stripes(bi
);
5069 sh
->overwrite_disks
++;
5071 spin_unlock_irq(&sh
->stripe_lock
);
5072 if (conf
->mddev
->bitmap
) {
5074 d
< conf
->raid_disks
- conf
->max_degraded
;
5076 bitmap_startwrite(mddev
->bitmap
,
5080 sh
->bm_seq
= conf
->seq_flush
+ 1;
5081 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5084 set_bit(STRIPE_HANDLE
, &sh
->state
);
5085 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5086 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5087 atomic_inc(&conf
->preread_active_stripes
);
5088 release_stripe_plug(mddev
, sh
);
5091 remaining
= raid5_dec_bi_active_stripes(bi
);
5092 if (remaining
== 0) {
5093 md_write_end(mddev
);
5098 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
5100 struct r5conf
*conf
= mddev
->private;
5102 sector_t new_sector
;
5103 sector_t logical_sector
, last_sector
;
5104 struct stripe_head
*sh
;
5105 const int rw
= bio_data_dir(bi
);
5110 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
5111 md_flush_request(mddev
, bi
);
5115 md_write_start(mddev
, bi
);
5118 * If array is degraded, better not do chunk aligned read because
5119 * later we might have to read it again in order to reconstruct
5120 * data on failed drives.
5122 if (rw
== READ
&& mddev
->degraded
== 0 &&
5123 mddev
->reshape_position
== MaxSector
&&
5124 chunk_aligned_read(mddev
,bi
))
5127 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
5128 make_discard_request(mddev
, bi
);
5132 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5133 last_sector
= bio_end_sector(bi
);
5135 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5137 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5138 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5144 seq
= read_seqcount_begin(&conf
->gen_lock
);
5147 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5148 TASK_UNINTERRUPTIBLE
);
5149 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5150 /* spinlock is needed as reshape_progress may be
5151 * 64bit on a 32bit platform, and so it might be
5152 * possible to see a half-updated value
5153 * Of course reshape_progress could change after
5154 * the lock is dropped, so once we get a reference
5155 * to the stripe that we think it is, we will have
5158 spin_lock_irq(&conf
->device_lock
);
5159 if (mddev
->reshape_backwards
5160 ? logical_sector
< conf
->reshape_progress
5161 : logical_sector
>= conf
->reshape_progress
) {
5164 if (mddev
->reshape_backwards
5165 ? logical_sector
< conf
->reshape_safe
5166 : logical_sector
>= conf
->reshape_safe
) {
5167 spin_unlock_irq(&conf
->device_lock
);
5173 spin_unlock_irq(&conf
->device_lock
);
5176 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5179 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5180 (unsigned long long)new_sector
,
5181 (unsigned long long)logical_sector
);
5183 sh
= get_active_stripe(conf
, new_sector
, previous
,
5184 (bi
->bi_rw
&RWA_MASK
), 0);
5186 if (unlikely(previous
)) {
5187 /* expansion might have moved on while waiting for a
5188 * stripe, so we must do the range check again.
5189 * Expansion could still move past after this
5190 * test, but as we are holding a reference to
5191 * 'sh', we know that if that happens,
5192 * STRIPE_EXPANDING will get set and the expansion
5193 * won't proceed until we finish with the stripe.
5196 spin_lock_irq(&conf
->device_lock
);
5197 if (mddev
->reshape_backwards
5198 ? logical_sector
>= conf
->reshape_progress
5199 : logical_sector
< conf
->reshape_progress
)
5200 /* mismatch, need to try again */
5202 spin_unlock_irq(&conf
->device_lock
);
5210 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5211 /* Might have got the wrong stripe_head
5219 logical_sector
>= mddev
->suspend_lo
&&
5220 logical_sector
< mddev
->suspend_hi
) {
5222 /* As the suspend_* range is controlled by
5223 * userspace, we want an interruptible
5226 flush_signals(current
);
5227 prepare_to_wait(&conf
->wait_for_overlap
,
5228 &w
, TASK_INTERRUPTIBLE
);
5229 if (logical_sector
>= mddev
->suspend_lo
&&
5230 logical_sector
< mddev
->suspend_hi
) {
5237 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5238 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5239 /* Stripe is busy expanding or
5240 * add failed due to overlap. Flush everything
5243 md_wakeup_thread(mddev
->thread
);
5249 set_bit(STRIPE_HANDLE
, &sh
->state
);
5250 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5251 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5252 (bi
->bi_rw
& REQ_SYNC
) &&
5253 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5254 atomic_inc(&conf
->preread_active_stripes
);
5255 release_stripe_plug(mddev
, sh
);
5257 /* cannot get stripe for read-ahead, just give-up */
5258 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
5262 finish_wait(&conf
->wait_for_overlap
, &w
);
5264 remaining
= raid5_dec_bi_active_stripes(bi
);
5265 if (remaining
== 0) {
5268 md_write_end(mddev
);
5270 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
5276 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5278 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5280 /* reshaping is quite different to recovery/resync so it is
5281 * handled quite separately ... here.
5283 * On each call to sync_request, we gather one chunk worth of
5284 * destination stripes and flag them as expanding.
5285 * Then we find all the source stripes and request reads.
5286 * As the reads complete, handle_stripe will copy the data
5287 * into the destination stripe and release that stripe.
5289 struct r5conf
*conf
= mddev
->private;
5290 struct stripe_head
*sh
;
5291 sector_t first_sector
, last_sector
;
5292 int raid_disks
= conf
->previous_raid_disks
;
5293 int data_disks
= raid_disks
- conf
->max_degraded
;
5294 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5297 sector_t writepos
, readpos
, safepos
;
5298 sector_t stripe_addr
;
5299 int reshape_sectors
;
5300 struct list_head stripes
;
5302 if (sector_nr
== 0) {
5303 /* If restarting in the middle, skip the initial sectors */
5304 if (mddev
->reshape_backwards
&&
5305 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5306 sector_nr
= raid5_size(mddev
, 0, 0)
5307 - conf
->reshape_progress
;
5308 } else if (!mddev
->reshape_backwards
&&
5309 conf
->reshape_progress
> 0)
5310 sector_nr
= conf
->reshape_progress
;
5311 sector_div(sector_nr
, new_data_disks
);
5313 mddev
->curr_resync_completed
= sector_nr
;
5314 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5320 /* We need to process a full chunk at a time.
5321 * If old and new chunk sizes differ, we need to process the
5324 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
5325 reshape_sectors
= mddev
->new_chunk_sectors
;
5327 reshape_sectors
= mddev
->chunk_sectors
;
5329 /* We update the metadata at least every 10 seconds, or when
5330 * the data about to be copied would over-write the source of
5331 * the data at the front of the range. i.e. one new_stripe
5332 * along from reshape_progress new_maps to after where
5333 * reshape_safe old_maps to
5335 writepos
= conf
->reshape_progress
;
5336 sector_div(writepos
, new_data_disks
);
5337 readpos
= conf
->reshape_progress
;
5338 sector_div(readpos
, data_disks
);
5339 safepos
= conf
->reshape_safe
;
5340 sector_div(safepos
, data_disks
);
5341 if (mddev
->reshape_backwards
) {
5342 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
5343 readpos
+= reshape_sectors
;
5344 safepos
+= reshape_sectors
;
5346 writepos
+= reshape_sectors
;
5347 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5348 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5351 /* Having calculated the 'writepos' possibly use it
5352 * to set 'stripe_addr' which is where we will write to.
5354 if (mddev
->reshape_backwards
) {
5355 BUG_ON(conf
->reshape_progress
== 0);
5356 stripe_addr
= writepos
;
5357 BUG_ON((mddev
->dev_sectors
&
5358 ~((sector_t
)reshape_sectors
- 1))
5359 - reshape_sectors
- stripe_addr
5362 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5363 stripe_addr
= sector_nr
;
5366 /* 'writepos' is the most advanced device address we might write.
5367 * 'readpos' is the least advanced device address we might read.
5368 * 'safepos' is the least address recorded in the metadata as having
5370 * If there is a min_offset_diff, these are adjusted either by
5371 * increasing the safepos/readpos if diff is negative, or
5372 * increasing writepos if diff is positive.
5373 * If 'readpos' is then behind 'writepos', there is no way that we can
5374 * ensure safety in the face of a crash - that must be done by userspace
5375 * making a backup of the data. So in that case there is no particular
5376 * rush to update metadata.
5377 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5378 * update the metadata to advance 'safepos' to match 'readpos' so that
5379 * we can be safe in the event of a crash.
5380 * So we insist on updating metadata if safepos is behind writepos and
5381 * readpos is beyond writepos.
5382 * In any case, update the metadata every 10 seconds.
5383 * Maybe that number should be configurable, but I'm not sure it is
5384 * worth it.... maybe it could be a multiple of safemode_delay???
5386 if (conf
->min_offset_diff
< 0) {
5387 safepos
+= -conf
->min_offset_diff
;
5388 readpos
+= -conf
->min_offset_diff
;
5390 writepos
+= conf
->min_offset_diff
;
5392 if ((mddev
->reshape_backwards
5393 ? (safepos
> writepos
&& readpos
< writepos
)
5394 : (safepos
< writepos
&& readpos
> writepos
)) ||
5395 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5396 /* Cannot proceed until we've updated the superblock... */
5397 wait_event(conf
->wait_for_overlap
,
5398 atomic_read(&conf
->reshape_stripes
)==0
5399 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5400 if (atomic_read(&conf
->reshape_stripes
) != 0)
5402 mddev
->reshape_position
= conf
->reshape_progress
;
5403 mddev
->curr_resync_completed
= sector_nr
;
5404 conf
->reshape_checkpoint
= jiffies
;
5405 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5406 md_wakeup_thread(mddev
->thread
);
5407 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
5408 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5409 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5411 spin_lock_irq(&conf
->device_lock
);
5412 conf
->reshape_safe
= mddev
->reshape_position
;
5413 spin_unlock_irq(&conf
->device_lock
);
5414 wake_up(&conf
->wait_for_overlap
);
5415 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5418 INIT_LIST_HEAD(&stripes
);
5419 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5421 int skipped_disk
= 0;
5422 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5423 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5424 atomic_inc(&conf
->reshape_stripes
);
5425 /* If any of this stripe is beyond the end of the old
5426 * array, then we need to zero those blocks
5428 for (j
=sh
->disks
; j
--;) {
5430 if (j
== sh
->pd_idx
)
5432 if (conf
->level
== 6 &&
5435 s
= compute_blocknr(sh
, j
, 0);
5436 if (s
< raid5_size(mddev
, 0, 0)) {
5440 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5441 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5442 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5444 if (!skipped_disk
) {
5445 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5446 set_bit(STRIPE_HANDLE
, &sh
->state
);
5448 list_add(&sh
->lru
, &stripes
);
5450 spin_lock_irq(&conf
->device_lock
);
5451 if (mddev
->reshape_backwards
)
5452 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5454 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5455 spin_unlock_irq(&conf
->device_lock
);
5456 /* Ok, those stripe are ready. We can start scheduling
5457 * reads on the source stripes.
5458 * The source stripes are determined by mapping the first and last
5459 * block on the destination stripes.
5462 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5465 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5466 * new_data_disks
- 1),
5468 if (last_sector
>= mddev
->dev_sectors
)
5469 last_sector
= mddev
->dev_sectors
- 1;
5470 while (first_sector
<= last_sector
) {
5471 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
5472 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5473 set_bit(STRIPE_HANDLE
, &sh
->state
);
5475 first_sector
+= STRIPE_SECTORS
;
5477 /* Now that the sources are clearly marked, we can release
5478 * the destination stripes
5480 while (!list_empty(&stripes
)) {
5481 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5482 list_del_init(&sh
->lru
);
5485 /* If this takes us to the resync_max point where we have to pause,
5486 * then we need to write out the superblock.
5488 sector_nr
+= reshape_sectors
;
5489 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
5490 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5491 /* Cannot proceed until we've updated the superblock... */
5492 wait_event(conf
->wait_for_overlap
,
5493 atomic_read(&conf
->reshape_stripes
) == 0
5494 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5495 if (atomic_read(&conf
->reshape_stripes
) != 0)
5497 mddev
->reshape_position
= conf
->reshape_progress
;
5498 mddev
->curr_resync_completed
= sector_nr
;
5499 conf
->reshape_checkpoint
= jiffies
;
5500 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5501 md_wakeup_thread(mddev
->thread
);
5502 wait_event(mddev
->sb_wait
,
5503 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
5504 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5505 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5507 spin_lock_irq(&conf
->device_lock
);
5508 conf
->reshape_safe
= mddev
->reshape_position
;
5509 spin_unlock_irq(&conf
->device_lock
);
5510 wake_up(&conf
->wait_for_overlap
);
5511 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5514 return reshape_sectors
;
5517 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5519 struct r5conf
*conf
= mddev
->private;
5520 struct stripe_head
*sh
;
5521 sector_t max_sector
= mddev
->dev_sectors
;
5522 sector_t sync_blocks
;
5523 int still_degraded
= 0;
5526 if (sector_nr
>= max_sector
) {
5527 /* just being told to finish up .. nothing much to do */
5529 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
5534 if (mddev
->curr_resync
< max_sector
) /* aborted */
5535 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
5537 else /* completed sync */
5539 bitmap_close_sync(mddev
->bitmap
);
5544 /* Allow raid5_quiesce to complete */
5545 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
5547 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
5548 return reshape_request(mddev
, sector_nr
, skipped
);
5550 /* No need to check resync_max as we never do more than one
5551 * stripe, and as resync_max will always be on a chunk boundary,
5552 * if the check in md_do_sync didn't fire, there is no chance
5553 * of overstepping resync_max here
5556 /* if there is too many failed drives and we are trying
5557 * to resync, then assert that we are finished, because there is
5558 * nothing we can do.
5560 if (mddev
->degraded
>= conf
->max_degraded
&&
5561 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
5562 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
5566 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
5568 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
5569 sync_blocks
>= STRIPE_SECTORS
) {
5570 /* we can skip this block, and probably more */
5571 sync_blocks
/= STRIPE_SECTORS
;
5573 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
5576 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
5578 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
5580 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
5581 /* make sure we don't swamp the stripe cache if someone else
5582 * is trying to get access
5584 schedule_timeout_uninterruptible(1);
5586 /* Need to check if array will still be degraded after recovery/resync
5587 * Note in case of > 1 drive failures it's possible we're rebuilding
5588 * one drive while leaving another faulty drive in array.
5591 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5592 struct md_rdev
*rdev
= ACCESS_ONCE(conf
->disks
[i
].rdev
);
5594 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
5599 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
5601 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
5602 set_bit(STRIPE_HANDLE
, &sh
->state
);
5606 return STRIPE_SECTORS
;
5609 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
5611 /* We may not be able to submit a whole bio at once as there
5612 * may not be enough stripe_heads available.
5613 * We cannot pre-allocate enough stripe_heads as we may need
5614 * more than exist in the cache (if we allow ever large chunks).
5615 * So we do one stripe head at a time and record in
5616 * ->bi_hw_segments how many have been done.
5618 * We *know* that this entire raid_bio is in one chunk, so
5619 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5621 struct stripe_head
*sh
;
5623 sector_t sector
, logical_sector
, last_sector
;
5628 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
5629 ~((sector_t
)STRIPE_SECTORS
-1);
5630 sector
= raid5_compute_sector(conf
, logical_sector
,
5632 last_sector
= bio_end_sector(raid_bio
);
5634 for (; logical_sector
< last_sector
;
5635 logical_sector
+= STRIPE_SECTORS
,
5636 sector
+= STRIPE_SECTORS
,
5639 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
5640 /* already done this stripe */
5643 sh
= get_active_stripe(conf
, sector
, 0, 1, 1);
5646 /* failed to get a stripe - must wait */
5647 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5648 conf
->retry_read_aligned
= raid_bio
;
5652 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
5654 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5655 conf
->retry_read_aligned
= raid_bio
;
5659 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
5664 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
5665 if (remaining
== 0) {
5666 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
5668 bio_endio(raid_bio
, 0);
5670 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5671 wake_up(&conf
->wait_for_stripe
);
5675 static int handle_active_stripes(struct r5conf
*conf
, int group
,
5676 struct r5worker
*worker
,
5677 struct list_head
*temp_inactive_list
)
5679 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
5680 int i
, batch_size
= 0, hash
;
5681 bool release_inactive
= false;
5683 while (batch_size
< MAX_STRIPE_BATCH
&&
5684 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
5685 batch
[batch_size
++] = sh
;
5687 if (batch_size
== 0) {
5688 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5689 if (!list_empty(temp_inactive_list
+ i
))
5691 if (i
== NR_STRIPE_HASH_LOCKS
)
5693 release_inactive
= true;
5695 spin_unlock_irq(&conf
->device_lock
);
5697 release_inactive_stripe_list(conf
, temp_inactive_list
,
5698 NR_STRIPE_HASH_LOCKS
);
5700 if (release_inactive
) {
5701 spin_lock_irq(&conf
->device_lock
);
5705 for (i
= 0; i
< batch_size
; i
++)
5706 handle_stripe(batch
[i
]);
5710 spin_lock_irq(&conf
->device_lock
);
5711 for (i
= 0; i
< batch_size
; i
++) {
5712 hash
= batch
[i
]->hash_lock_index
;
5713 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
5718 static void raid5_do_work(struct work_struct
*work
)
5720 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
5721 struct r5worker_group
*group
= worker
->group
;
5722 struct r5conf
*conf
= group
->conf
;
5723 int group_id
= group
- conf
->worker_groups
;
5725 struct blk_plug plug
;
5727 pr_debug("+++ raid5worker active\n");
5729 blk_start_plug(&plug
);
5731 spin_lock_irq(&conf
->device_lock
);
5733 int batch_size
, released
;
5735 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
5737 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
5738 worker
->temp_inactive_list
);
5739 worker
->working
= false;
5740 if (!batch_size
&& !released
)
5742 handled
+= batch_size
;
5744 pr_debug("%d stripes handled\n", handled
);
5746 spin_unlock_irq(&conf
->device_lock
);
5747 blk_finish_plug(&plug
);
5749 pr_debug("--- raid5worker inactive\n");
5753 * This is our raid5 kernel thread.
5755 * We scan the hash table for stripes which can be handled now.
5756 * During the scan, completed stripes are saved for us by the interrupt
5757 * handler, so that they will not have to wait for our next wakeup.
5759 static void raid5d(struct md_thread
*thread
)
5761 struct mddev
*mddev
= thread
->mddev
;
5762 struct r5conf
*conf
= mddev
->private;
5764 struct blk_plug plug
;
5766 pr_debug("+++ raid5d active\n");
5768 md_check_recovery(mddev
);
5770 blk_start_plug(&plug
);
5772 spin_lock_irq(&conf
->device_lock
);
5775 int batch_size
, released
;
5777 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
5779 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5782 !list_empty(&conf
->bitmap_list
)) {
5783 /* Now is a good time to flush some bitmap updates */
5785 spin_unlock_irq(&conf
->device_lock
);
5786 bitmap_unplug(mddev
->bitmap
);
5787 spin_lock_irq(&conf
->device_lock
);
5788 conf
->seq_write
= conf
->seq_flush
;
5789 activate_bit_delay(conf
, conf
->temp_inactive_list
);
5791 raid5_activate_delayed(conf
);
5793 while ((bio
= remove_bio_from_retry(conf
))) {
5795 spin_unlock_irq(&conf
->device_lock
);
5796 ok
= retry_aligned_read(conf
, bio
);
5797 spin_lock_irq(&conf
->device_lock
);
5803 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
5804 conf
->temp_inactive_list
);
5805 if (!batch_size
&& !released
)
5807 handled
+= batch_size
;
5809 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
5810 spin_unlock_irq(&conf
->device_lock
);
5811 md_check_recovery(mddev
);
5812 spin_lock_irq(&conf
->device_lock
);
5815 pr_debug("%d stripes handled\n", handled
);
5817 spin_unlock_irq(&conf
->device_lock
);
5818 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
)) {
5819 grow_one_stripe(conf
, __GFP_NOWARN
);
5820 /* Set flag even if allocation failed. This helps
5821 * slow down allocation requests when mem is short
5823 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5826 async_tx_issue_pending_all();
5827 blk_finish_plug(&plug
);
5829 pr_debug("--- raid5d inactive\n");
5833 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
5835 struct r5conf
*conf
;
5837 spin_lock(&mddev
->lock
);
5838 conf
= mddev
->private;
5840 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
5841 spin_unlock(&mddev
->lock
);
5846 raid5_set_cache_size(struct mddev
*mddev
, int size
)
5848 struct r5conf
*conf
= mddev
->private;
5851 if (size
<= 16 || size
> 32768)
5854 conf
->min_nr_stripes
= size
;
5855 while (size
< conf
->max_nr_stripes
&&
5856 drop_one_stripe(conf
))
5860 err
= md_allow_write(mddev
);
5864 while (size
> conf
->max_nr_stripes
)
5865 if (!grow_one_stripe(conf
, GFP_KERNEL
))
5870 EXPORT_SYMBOL(raid5_set_cache_size
);
5873 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
5875 struct r5conf
*conf
;
5879 if (len
>= PAGE_SIZE
)
5881 if (kstrtoul(page
, 10, &new))
5883 err
= mddev_lock(mddev
);
5886 conf
= mddev
->private;
5890 err
= raid5_set_cache_size(mddev
, new);
5891 mddev_unlock(mddev
);
5896 static struct md_sysfs_entry
5897 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
5898 raid5_show_stripe_cache_size
,
5899 raid5_store_stripe_cache_size
);
5902 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
5904 struct r5conf
*conf
= mddev
->private;
5906 return sprintf(page
, "%d\n", conf
->rmw_level
);
5912 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
5914 struct r5conf
*conf
= mddev
->private;
5920 if (len
>= PAGE_SIZE
)
5923 if (kstrtoul(page
, 10, &new))
5926 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
5929 if (new != PARITY_DISABLE_RMW
&&
5930 new != PARITY_ENABLE_RMW
&&
5931 new != PARITY_PREFER_RMW
)
5934 conf
->rmw_level
= new;
5938 static struct md_sysfs_entry
5939 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
5940 raid5_show_rmw_level
,
5941 raid5_store_rmw_level
);
5945 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
5947 struct r5conf
*conf
;
5949 spin_lock(&mddev
->lock
);
5950 conf
= mddev
->private;
5952 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
5953 spin_unlock(&mddev
->lock
);
5958 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
5960 struct r5conf
*conf
;
5964 if (len
>= PAGE_SIZE
)
5966 if (kstrtoul(page
, 10, &new))
5969 err
= mddev_lock(mddev
);
5972 conf
= mddev
->private;
5975 else if (new > conf
->min_nr_stripes
)
5978 conf
->bypass_threshold
= new;
5979 mddev_unlock(mddev
);
5983 static struct md_sysfs_entry
5984 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
5986 raid5_show_preread_threshold
,
5987 raid5_store_preread_threshold
);
5990 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
5992 struct r5conf
*conf
;
5994 spin_lock(&mddev
->lock
);
5995 conf
= mddev
->private;
5997 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
5998 spin_unlock(&mddev
->lock
);
6003 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6005 struct r5conf
*conf
;
6009 if (len
>= PAGE_SIZE
)
6011 if (kstrtoul(page
, 10, &new))
6015 err
= mddev_lock(mddev
);
6018 conf
= mddev
->private;
6021 else if (new != conf
->skip_copy
) {
6022 mddev_suspend(mddev
);
6023 conf
->skip_copy
= new;
6025 mddev
->queue
->backing_dev_info
.capabilities
|=
6026 BDI_CAP_STABLE_WRITES
;
6028 mddev
->queue
->backing_dev_info
.capabilities
&=
6029 ~BDI_CAP_STABLE_WRITES
;
6030 mddev_resume(mddev
);
6032 mddev_unlock(mddev
);
6036 static struct md_sysfs_entry
6037 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6038 raid5_show_skip_copy
,
6039 raid5_store_skip_copy
);
6042 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6044 struct r5conf
*conf
= mddev
->private;
6046 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6051 static struct md_sysfs_entry
6052 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6055 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6057 struct r5conf
*conf
;
6059 spin_lock(&mddev
->lock
);
6060 conf
= mddev
->private;
6062 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6063 spin_unlock(&mddev
->lock
);
6067 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6069 int *worker_cnt_per_group
,
6070 struct r5worker_group
**worker_groups
);
6072 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6074 struct r5conf
*conf
;
6077 struct r5worker_group
*new_groups
, *old_groups
;
6078 int group_cnt
, worker_cnt_per_group
;
6080 if (len
>= PAGE_SIZE
)
6082 if (kstrtoul(page
, 10, &new))
6085 err
= mddev_lock(mddev
);
6088 conf
= mddev
->private;
6091 else if (new != conf
->worker_cnt_per_group
) {
6092 mddev_suspend(mddev
);
6094 old_groups
= conf
->worker_groups
;
6096 flush_workqueue(raid5_wq
);
6098 err
= alloc_thread_groups(conf
, new,
6099 &group_cnt
, &worker_cnt_per_group
,
6102 spin_lock_irq(&conf
->device_lock
);
6103 conf
->group_cnt
= group_cnt
;
6104 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6105 conf
->worker_groups
= new_groups
;
6106 spin_unlock_irq(&conf
->device_lock
);
6109 kfree(old_groups
[0].workers
);
6112 mddev_resume(mddev
);
6114 mddev_unlock(mddev
);
6119 static struct md_sysfs_entry
6120 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6121 raid5_show_group_thread_cnt
,
6122 raid5_store_group_thread_cnt
);
6124 static struct attribute
*raid5_attrs
[] = {
6125 &raid5_stripecache_size
.attr
,
6126 &raid5_stripecache_active
.attr
,
6127 &raid5_preread_bypass_threshold
.attr
,
6128 &raid5_group_thread_cnt
.attr
,
6129 &raid5_skip_copy
.attr
,
6130 &raid5_rmw_level
.attr
,
6133 static struct attribute_group raid5_attrs_group
= {
6135 .attrs
= raid5_attrs
,
6138 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6140 int *worker_cnt_per_group
,
6141 struct r5worker_group
**worker_groups
)
6145 struct r5worker
*workers
;
6147 *worker_cnt_per_group
= cnt
;
6150 *worker_groups
= NULL
;
6153 *group_cnt
= num_possible_nodes();
6154 size
= sizeof(struct r5worker
) * cnt
;
6155 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6156 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6157 *group_cnt
, GFP_NOIO
);
6158 if (!*worker_groups
|| !workers
) {
6160 kfree(*worker_groups
);
6164 for (i
= 0; i
< *group_cnt
; i
++) {
6165 struct r5worker_group
*group
;
6167 group
= &(*worker_groups
)[i
];
6168 INIT_LIST_HEAD(&group
->handle_list
);
6170 group
->workers
= workers
+ i
* cnt
;
6172 for (j
= 0; j
< cnt
; j
++) {
6173 struct r5worker
*worker
= group
->workers
+ j
;
6174 worker
->group
= group
;
6175 INIT_WORK(&worker
->work
, raid5_do_work
);
6177 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6178 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6185 static void free_thread_groups(struct r5conf
*conf
)
6187 if (conf
->worker_groups
)
6188 kfree(conf
->worker_groups
[0].workers
);
6189 kfree(conf
->worker_groups
);
6190 conf
->worker_groups
= NULL
;
6194 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6196 struct r5conf
*conf
= mddev
->private;
6199 sectors
= mddev
->dev_sectors
;
6201 /* size is defined by the smallest of previous and new size */
6202 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6204 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
6205 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
6206 return sectors
* (raid_disks
- conf
->max_degraded
);
6209 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6211 safe_put_page(percpu
->spare_page
);
6212 if (percpu
->scribble
)
6213 flex_array_free(percpu
->scribble
);
6214 percpu
->spare_page
= NULL
;
6215 percpu
->scribble
= NULL
;
6218 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6220 if (conf
->level
== 6 && !percpu
->spare_page
)
6221 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6222 if (!percpu
->scribble
)
6223 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6224 conf
->previous_raid_disks
), conf
->chunk_sectors
/
6225 STRIPE_SECTORS
, GFP_KERNEL
);
6227 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6228 free_scratch_buffer(conf
, percpu
);
6235 static void raid5_free_percpu(struct r5conf
*conf
)
6242 #ifdef CONFIG_HOTPLUG_CPU
6243 unregister_cpu_notifier(&conf
->cpu_notify
);
6247 for_each_possible_cpu(cpu
)
6248 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6251 free_percpu(conf
->percpu
);
6254 static void free_conf(struct r5conf
*conf
)
6256 if (conf
->shrinker
.seeks
)
6257 unregister_shrinker(&conf
->shrinker
);
6258 free_thread_groups(conf
);
6259 shrink_stripes(conf
);
6260 raid5_free_percpu(conf
);
6262 kfree(conf
->stripe_hashtbl
);
6266 #ifdef CONFIG_HOTPLUG_CPU
6267 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
6270 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
6271 long cpu
= (long)hcpu
;
6272 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6275 case CPU_UP_PREPARE
:
6276 case CPU_UP_PREPARE_FROZEN
:
6277 if (alloc_scratch_buffer(conf
, percpu
)) {
6278 pr_err("%s: failed memory allocation for cpu%ld\n",
6280 return notifier_from_errno(-ENOMEM
);
6284 case CPU_DEAD_FROZEN
:
6285 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6294 static int raid5_alloc_percpu(struct r5conf
*conf
)
6299 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6303 #ifdef CONFIG_HOTPLUG_CPU
6304 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
6305 conf
->cpu_notify
.priority
= 0;
6306 err
= register_cpu_notifier(&conf
->cpu_notify
);
6312 for_each_present_cpu(cpu
) {
6313 err
= alloc_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6315 pr_err("%s: failed memory allocation for cpu%ld\n",
6325 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6326 struct shrink_control
*sc
)
6328 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6330 while (ret
< sc
->nr_to_scan
) {
6331 if (drop_one_stripe(conf
) == 0)
6338 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6339 struct shrink_control
*sc
)
6341 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6343 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6344 /* unlikely, but not impossible */
6346 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6349 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6351 struct r5conf
*conf
;
6352 int raid_disk
, memory
, max_disks
;
6353 struct md_rdev
*rdev
;
6354 struct disk_info
*disk
;
6357 int group_cnt
, worker_cnt_per_group
;
6358 struct r5worker_group
*new_group
;
6360 if (mddev
->new_level
!= 5
6361 && mddev
->new_level
!= 4
6362 && mddev
->new_level
!= 6) {
6363 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6364 mdname(mddev
), mddev
->new_level
);
6365 return ERR_PTR(-EIO
);
6367 if ((mddev
->new_level
== 5
6368 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6369 (mddev
->new_level
== 6
6370 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6371 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
6372 mdname(mddev
), mddev
->new_layout
);
6373 return ERR_PTR(-EIO
);
6375 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6376 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6377 mdname(mddev
), mddev
->raid_disks
);
6378 return ERR_PTR(-EINVAL
);
6381 if (!mddev
->new_chunk_sectors
||
6382 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6383 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6384 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
6385 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6386 return ERR_PTR(-EINVAL
);
6389 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6392 /* Don't enable multi-threading by default*/
6393 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6395 conf
->group_cnt
= group_cnt
;
6396 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6397 conf
->worker_groups
= new_group
;
6400 spin_lock_init(&conf
->device_lock
);
6401 seqcount_init(&conf
->gen_lock
);
6402 init_waitqueue_head(&conf
->wait_for_stripe
);
6403 init_waitqueue_head(&conf
->wait_for_overlap
);
6404 INIT_LIST_HEAD(&conf
->handle_list
);
6405 INIT_LIST_HEAD(&conf
->hold_list
);
6406 INIT_LIST_HEAD(&conf
->delayed_list
);
6407 INIT_LIST_HEAD(&conf
->bitmap_list
);
6408 init_llist_head(&conf
->released_stripes
);
6409 atomic_set(&conf
->active_stripes
, 0);
6410 atomic_set(&conf
->preread_active_stripes
, 0);
6411 atomic_set(&conf
->active_aligned_reads
, 0);
6412 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6413 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6415 conf
->raid_disks
= mddev
->raid_disks
;
6416 if (mddev
->reshape_position
== MaxSector
)
6417 conf
->previous_raid_disks
= mddev
->raid_disks
;
6419 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6420 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6422 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6427 conf
->mddev
= mddev
;
6429 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6432 /* We init hash_locks[0] separately to that it can be used
6433 * as the reference lock in the spin_lock_nest_lock() call
6434 * in lock_all_device_hash_locks_irq in order to convince
6435 * lockdep that we know what we are doing.
6437 spin_lock_init(conf
->hash_locks
);
6438 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6439 spin_lock_init(conf
->hash_locks
+ i
);
6441 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6442 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6444 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6445 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6447 conf
->level
= mddev
->new_level
;
6448 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6449 if (raid5_alloc_percpu(conf
) != 0)
6452 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6454 rdev_for_each(rdev
, mddev
) {
6455 raid_disk
= rdev
->raid_disk
;
6456 if (raid_disk
>= max_disks
6459 disk
= conf
->disks
+ raid_disk
;
6461 if (test_bit(Replacement
, &rdev
->flags
)) {
6462 if (disk
->replacement
)
6464 disk
->replacement
= rdev
;
6471 if (test_bit(In_sync
, &rdev
->flags
)) {
6472 char b
[BDEVNAME_SIZE
];
6473 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
6475 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
6476 } else if (rdev
->saved_raid_disk
!= raid_disk
)
6477 /* Cannot rely on bitmap to complete recovery */
6481 conf
->level
= mddev
->new_level
;
6482 if (conf
->level
== 6) {
6483 conf
->max_degraded
= 2;
6484 if (raid6_call
.xor_syndrome
)
6485 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6487 conf
->rmw_level
= PARITY_DISABLE_RMW
;
6489 conf
->max_degraded
= 1;
6490 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6492 conf
->algorithm
= mddev
->new_layout
;
6493 conf
->reshape_progress
= mddev
->reshape_position
;
6494 if (conf
->reshape_progress
!= MaxSector
) {
6495 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
6496 conf
->prev_algo
= mddev
->layout
;
6499 conf
->min_nr_stripes
= NR_STRIPES
;
6500 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
6501 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
6502 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
6503 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
6505 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6506 mdname(mddev
), memory
);
6509 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
6510 mdname(mddev
), memory
);
6512 * Losing a stripe head costs more than the time to refill it,
6513 * it reduces the queue depth and so can hurt throughput.
6514 * So set it rather large, scaled by number of devices.
6516 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
6517 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
6518 conf
->shrinker
.count_objects
= raid5_cache_count
;
6519 conf
->shrinker
.batch
= 128;
6520 conf
->shrinker
.flags
= 0;
6521 register_shrinker(&conf
->shrinker
);
6523 sprintf(pers_name
, "raid%d", mddev
->new_level
);
6524 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
6525 if (!conf
->thread
) {
6527 "md/raid:%s: couldn't allocate thread.\n",
6537 return ERR_PTR(-EIO
);
6539 return ERR_PTR(-ENOMEM
);
6542 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
6545 case ALGORITHM_PARITY_0
:
6546 if (raid_disk
< max_degraded
)
6549 case ALGORITHM_PARITY_N
:
6550 if (raid_disk
>= raid_disks
- max_degraded
)
6553 case ALGORITHM_PARITY_0_6
:
6554 if (raid_disk
== 0 ||
6555 raid_disk
== raid_disks
- 1)
6558 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6559 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6560 case ALGORITHM_LEFT_SYMMETRIC_6
:
6561 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6562 if (raid_disk
== raid_disks
- 1)
6568 static int run(struct mddev
*mddev
)
6570 struct r5conf
*conf
;
6571 int working_disks
= 0;
6572 int dirty_parity_disks
= 0;
6573 struct md_rdev
*rdev
;
6574 sector_t reshape_offset
= 0;
6576 long long min_offset_diff
= 0;
6579 if (mddev
->recovery_cp
!= MaxSector
)
6580 printk(KERN_NOTICE
"md/raid:%s: not clean"
6581 " -- starting background reconstruction\n",
6584 rdev_for_each(rdev
, mddev
) {
6586 if (rdev
->raid_disk
< 0)
6588 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
6590 min_offset_diff
= diff
;
6592 } else if (mddev
->reshape_backwards
&&
6593 diff
< min_offset_diff
)
6594 min_offset_diff
= diff
;
6595 else if (!mddev
->reshape_backwards
&&
6596 diff
> min_offset_diff
)
6597 min_offset_diff
= diff
;
6600 if (mddev
->reshape_position
!= MaxSector
) {
6601 /* Check that we can continue the reshape.
6602 * Difficulties arise if the stripe we would write to
6603 * next is at or after the stripe we would read from next.
6604 * For a reshape that changes the number of devices, this
6605 * is only possible for a very short time, and mdadm makes
6606 * sure that time appears to have past before assembling
6607 * the array. So we fail if that time hasn't passed.
6608 * For a reshape that keeps the number of devices the same
6609 * mdadm must be monitoring the reshape can keeping the
6610 * critical areas read-only and backed up. It will start
6611 * the array in read-only mode, so we check for that.
6613 sector_t here_new
, here_old
;
6615 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
6617 if (mddev
->new_level
!= mddev
->level
) {
6618 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
6619 "required - aborting.\n",
6623 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6624 /* reshape_position must be on a new-stripe boundary, and one
6625 * further up in new geometry must map after here in old
6628 here_new
= mddev
->reshape_position
;
6629 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
6630 (mddev
->raid_disks
- max_degraded
))) {
6631 printk(KERN_ERR
"md/raid:%s: reshape_position not "
6632 "on a stripe boundary\n", mdname(mddev
));
6635 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
6636 /* here_new is the stripe we will write to */
6637 here_old
= mddev
->reshape_position
;
6638 sector_div(here_old
, mddev
->chunk_sectors
*
6639 (old_disks
-max_degraded
));
6640 /* here_old is the first stripe that we might need to read
6642 if (mddev
->delta_disks
== 0) {
6643 if ((here_new
* mddev
->new_chunk_sectors
!=
6644 here_old
* mddev
->chunk_sectors
)) {
6645 printk(KERN_ERR
"md/raid:%s: reshape position is"
6646 " confused - aborting\n", mdname(mddev
));
6649 /* We cannot be sure it is safe to start an in-place
6650 * reshape. It is only safe if user-space is monitoring
6651 * and taking constant backups.
6652 * mdadm always starts a situation like this in
6653 * readonly mode so it can take control before
6654 * allowing any writes. So just check for that.
6656 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
6657 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
6658 /* not really in-place - so OK */;
6659 else if (mddev
->ro
== 0) {
6660 printk(KERN_ERR
"md/raid:%s: in-place reshape "
6661 "must be started in read-only mode "
6666 } else if (mddev
->reshape_backwards
6667 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
6668 here_old
* mddev
->chunk_sectors
)
6669 : (here_new
* mddev
->new_chunk_sectors
>=
6670 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
6671 /* Reading from the same stripe as writing to - bad */
6672 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
6673 "auto-recovery - aborting.\n",
6677 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
6679 /* OK, we should be able to continue; */
6681 BUG_ON(mddev
->level
!= mddev
->new_level
);
6682 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
6683 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
6684 BUG_ON(mddev
->delta_disks
!= 0);
6687 if (mddev
->private == NULL
)
6688 conf
= setup_conf(mddev
);
6690 conf
= mddev
->private;
6693 return PTR_ERR(conf
);
6695 conf
->min_offset_diff
= min_offset_diff
;
6696 mddev
->thread
= conf
->thread
;
6697 conf
->thread
= NULL
;
6698 mddev
->private = conf
;
6700 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
6702 rdev
= conf
->disks
[i
].rdev
;
6703 if (!rdev
&& conf
->disks
[i
].replacement
) {
6704 /* The replacement is all we have yet */
6705 rdev
= conf
->disks
[i
].replacement
;
6706 conf
->disks
[i
].replacement
= NULL
;
6707 clear_bit(Replacement
, &rdev
->flags
);
6708 conf
->disks
[i
].rdev
= rdev
;
6712 if (conf
->disks
[i
].replacement
&&
6713 conf
->reshape_progress
!= MaxSector
) {
6714 /* replacements and reshape simply do not mix. */
6715 printk(KERN_ERR
"md: cannot handle concurrent "
6716 "replacement and reshape.\n");
6719 if (test_bit(In_sync
, &rdev
->flags
)) {
6723 /* This disc is not fully in-sync. However if it
6724 * just stored parity (beyond the recovery_offset),
6725 * when we don't need to be concerned about the
6726 * array being dirty.
6727 * When reshape goes 'backwards', we never have
6728 * partially completed devices, so we only need
6729 * to worry about reshape going forwards.
6731 /* Hack because v0.91 doesn't store recovery_offset properly. */
6732 if (mddev
->major_version
== 0 &&
6733 mddev
->minor_version
> 90)
6734 rdev
->recovery_offset
= reshape_offset
;
6736 if (rdev
->recovery_offset
< reshape_offset
) {
6737 /* We need to check old and new layout */
6738 if (!only_parity(rdev
->raid_disk
,
6741 conf
->max_degraded
))
6744 if (!only_parity(rdev
->raid_disk
,
6746 conf
->previous_raid_disks
,
6747 conf
->max_degraded
))
6749 dirty_parity_disks
++;
6753 * 0 for a fully functional array, 1 or 2 for a degraded array.
6755 mddev
->degraded
= calc_degraded(conf
);
6757 if (has_failed(conf
)) {
6758 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
6759 " (%d/%d failed)\n",
6760 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
6764 /* device size must be a multiple of chunk size */
6765 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
6766 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
6768 if (mddev
->degraded
> dirty_parity_disks
&&
6769 mddev
->recovery_cp
!= MaxSector
) {
6770 if (mddev
->ok_start_degraded
)
6772 "md/raid:%s: starting dirty degraded array"
6773 " - data corruption possible.\n",
6777 "md/raid:%s: cannot start dirty degraded array.\n",
6783 if (mddev
->degraded
== 0)
6784 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
6785 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
6786 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
6789 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
6790 " out of %d devices, algorithm %d\n",
6791 mdname(mddev
), conf
->level
,
6792 mddev
->raid_disks
- mddev
->degraded
,
6793 mddev
->raid_disks
, mddev
->new_layout
);
6795 print_raid5_conf(conf
);
6797 if (conf
->reshape_progress
!= MaxSector
) {
6798 conf
->reshape_safe
= conf
->reshape_progress
;
6799 atomic_set(&conf
->reshape_stripes
, 0);
6800 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6801 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6802 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6803 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6804 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6808 /* Ok, everything is just fine now */
6809 if (mddev
->to_remove
== &raid5_attrs_group
)
6810 mddev
->to_remove
= NULL
;
6811 else if (mddev
->kobj
.sd
&&
6812 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
6814 "raid5: failed to create sysfs attributes for %s\n",
6816 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6820 bool discard_supported
= true;
6821 /* read-ahead size must cover two whole stripes, which
6822 * is 2 * (datadisks) * chunksize where 'n' is the
6823 * number of raid devices
6825 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
6826 int stripe
= data_disks
*
6827 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
6828 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6829 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6831 chunk_size
= mddev
->chunk_sectors
<< 9;
6832 blk_queue_io_min(mddev
->queue
, chunk_size
);
6833 blk_queue_io_opt(mddev
->queue
, chunk_size
*
6834 (conf
->raid_disks
- conf
->max_degraded
));
6835 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
6837 * We can only discard a whole stripe. It doesn't make sense to
6838 * discard data disk but write parity disk
6840 stripe
= stripe
* PAGE_SIZE
;
6841 /* Round up to power of 2, as discard handling
6842 * currently assumes that */
6843 while ((stripe
-1) & stripe
)
6844 stripe
= (stripe
| (stripe
-1)) + 1;
6845 mddev
->queue
->limits
.discard_alignment
= stripe
;
6846 mddev
->queue
->limits
.discard_granularity
= stripe
;
6848 * unaligned part of discard request will be ignored, so can't
6849 * guarantee discard_zeroes_data
6851 mddev
->queue
->limits
.discard_zeroes_data
= 0;
6853 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
6855 rdev_for_each(rdev
, mddev
) {
6856 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6857 rdev
->data_offset
<< 9);
6858 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6859 rdev
->new_data_offset
<< 9);
6861 * discard_zeroes_data is required, otherwise data
6862 * could be lost. Consider a scenario: discard a stripe
6863 * (the stripe could be inconsistent if
6864 * discard_zeroes_data is 0); write one disk of the
6865 * stripe (the stripe could be inconsistent again
6866 * depending on which disks are used to calculate
6867 * parity); the disk is broken; The stripe data of this
6870 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
6871 !bdev_get_queue(rdev
->bdev
)->
6872 limits
.discard_zeroes_data
)
6873 discard_supported
= false;
6874 /* Unfortunately, discard_zeroes_data is not currently
6875 * a guarantee - just a hint. So we only allow DISCARD
6876 * if the sysadmin has confirmed that only safe devices
6877 * are in use by setting a module parameter.
6879 if (!devices_handle_discard_safely
) {
6880 if (discard_supported
) {
6881 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6882 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6884 discard_supported
= false;
6888 if (discard_supported
&&
6889 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
6890 mddev
->queue
->limits
.discard_granularity
>= stripe
)
6891 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
6894 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
6900 md_unregister_thread(&mddev
->thread
);
6901 print_raid5_conf(conf
);
6903 mddev
->private = NULL
;
6904 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
6908 static void raid5_free(struct mddev
*mddev
, void *priv
)
6910 struct r5conf
*conf
= priv
;
6913 mddev
->to_remove
= &raid5_attrs_group
;
6916 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
6918 struct r5conf
*conf
= mddev
->private;
6921 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
6922 mddev
->chunk_sectors
/ 2, mddev
->layout
);
6923 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
6924 for (i
= 0; i
< conf
->raid_disks
; i
++)
6925 seq_printf (seq
, "%s",
6926 conf
->disks
[i
].rdev
&&
6927 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
6928 seq_printf (seq
, "]");
6931 static void print_raid5_conf (struct r5conf
*conf
)
6934 struct disk_info
*tmp
;
6936 printk(KERN_DEBUG
"RAID conf printout:\n");
6938 printk("(conf==NULL)\n");
6941 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
6943 conf
->raid_disks
- conf
->mddev
->degraded
);
6945 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6946 char b
[BDEVNAME_SIZE
];
6947 tmp
= conf
->disks
+ i
;
6949 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
6950 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
6951 bdevname(tmp
->rdev
->bdev
, b
));
6955 static int raid5_spare_active(struct mddev
*mddev
)
6958 struct r5conf
*conf
= mddev
->private;
6959 struct disk_info
*tmp
;
6961 unsigned long flags
;
6963 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6964 tmp
= conf
->disks
+ i
;
6965 if (tmp
->replacement
6966 && tmp
->replacement
->recovery_offset
== MaxSector
6967 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
6968 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
6969 /* Replacement has just become active. */
6971 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
6974 /* Replaced device not technically faulty,
6975 * but we need to be sure it gets removed
6976 * and never re-added.
6978 set_bit(Faulty
, &tmp
->rdev
->flags
);
6979 sysfs_notify_dirent_safe(
6980 tmp
->rdev
->sysfs_state
);
6982 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
6983 } else if (tmp
->rdev
6984 && tmp
->rdev
->recovery_offset
== MaxSector
6985 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
6986 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
6988 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
6991 spin_lock_irqsave(&conf
->device_lock
, flags
);
6992 mddev
->degraded
= calc_degraded(conf
);
6993 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
6994 print_raid5_conf(conf
);
6998 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7000 struct r5conf
*conf
= mddev
->private;
7002 int number
= rdev
->raid_disk
;
7003 struct md_rdev
**rdevp
;
7004 struct disk_info
*p
= conf
->disks
+ number
;
7006 print_raid5_conf(conf
);
7007 if (rdev
== p
->rdev
)
7009 else if (rdev
== p
->replacement
)
7010 rdevp
= &p
->replacement
;
7014 if (number
>= conf
->raid_disks
&&
7015 conf
->reshape_progress
== MaxSector
)
7016 clear_bit(In_sync
, &rdev
->flags
);
7018 if (test_bit(In_sync
, &rdev
->flags
) ||
7019 atomic_read(&rdev
->nr_pending
)) {
7023 /* Only remove non-faulty devices if recovery
7026 if (!test_bit(Faulty
, &rdev
->flags
) &&
7027 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7028 !has_failed(conf
) &&
7029 (!p
->replacement
|| p
->replacement
== rdev
) &&
7030 number
< conf
->raid_disks
) {
7036 if (atomic_read(&rdev
->nr_pending
)) {
7037 /* lost the race, try later */
7040 } else if (p
->replacement
) {
7041 /* We must have just cleared 'rdev' */
7042 p
->rdev
= p
->replacement
;
7043 clear_bit(Replacement
, &p
->replacement
->flags
);
7044 smp_mb(); /* Make sure other CPUs may see both as identical
7045 * but will never see neither - if they are careful
7047 p
->replacement
= NULL
;
7048 clear_bit(WantReplacement
, &rdev
->flags
);
7050 /* We might have just removed the Replacement as faulty-
7051 * clear the bit just in case
7053 clear_bit(WantReplacement
, &rdev
->flags
);
7056 print_raid5_conf(conf
);
7060 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7062 struct r5conf
*conf
= mddev
->private;
7065 struct disk_info
*p
;
7067 int last
= conf
->raid_disks
- 1;
7069 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7072 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7073 /* no point adding a device */
7076 if (rdev
->raid_disk
>= 0)
7077 first
= last
= rdev
->raid_disk
;
7080 * find the disk ... but prefer rdev->saved_raid_disk
7083 if (rdev
->saved_raid_disk
>= 0 &&
7084 rdev
->saved_raid_disk
>= first
&&
7085 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7086 first
= rdev
->saved_raid_disk
;
7088 for (disk
= first
; disk
<= last
; disk
++) {
7089 p
= conf
->disks
+ disk
;
7090 if (p
->rdev
== NULL
) {
7091 clear_bit(In_sync
, &rdev
->flags
);
7092 rdev
->raid_disk
= disk
;
7094 if (rdev
->saved_raid_disk
!= disk
)
7096 rcu_assign_pointer(p
->rdev
, rdev
);
7100 for (disk
= first
; disk
<= last
; disk
++) {
7101 p
= conf
->disks
+ disk
;
7102 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7103 p
->replacement
== NULL
) {
7104 clear_bit(In_sync
, &rdev
->flags
);
7105 set_bit(Replacement
, &rdev
->flags
);
7106 rdev
->raid_disk
= disk
;
7109 rcu_assign_pointer(p
->replacement
, rdev
);
7114 print_raid5_conf(conf
);
7118 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7120 /* no resync is happening, and there is enough space
7121 * on all devices, so we can resize.
7122 * We need to make sure resync covers any new space.
7123 * If the array is shrinking we should possibly wait until
7124 * any io in the removed space completes, but it hardly seems
7128 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
7129 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7130 if (mddev
->external_size
&&
7131 mddev
->array_sectors
> newsize
)
7133 if (mddev
->bitmap
) {
7134 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7138 md_set_array_sectors(mddev
, newsize
);
7139 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7140 revalidate_disk(mddev
->gendisk
);
7141 if (sectors
> mddev
->dev_sectors
&&
7142 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7143 mddev
->recovery_cp
= mddev
->dev_sectors
;
7144 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7146 mddev
->dev_sectors
= sectors
;
7147 mddev
->resync_max_sectors
= sectors
;
7151 static int check_stripe_cache(struct mddev
*mddev
)
7153 /* Can only proceed if there are plenty of stripe_heads.
7154 * We need a minimum of one full stripe,, and for sensible progress
7155 * it is best to have about 4 times that.
7156 * If we require 4 times, then the default 256 4K stripe_heads will
7157 * allow for chunk sizes up to 256K, which is probably OK.
7158 * If the chunk size is greater, user-space should request more
7159 * stripe_heads first.
7161 struct r5conf
*conf
= mddev
->private;
7162 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7163 > conf
->min_nr_stripes
||
7164 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7165 > conf
->min_nr_stripes
) {
7166 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7168 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7175 static int check_reshape(struct mddev
*mddev
)
7177 struct r5conf
*conf
= mddev
->private;
7179 if (mddev
->delta_disks
== 0 &&
7180 mddev
->new_layout
== mddev
->layout
&&
7181 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7182 return 0; /* nothing to do */
7183 if (has_failed(conf
))
7185 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7186 /* We might be able to shrink, but the devices must
7187 * be made bigger first.
7188 * For raid6, 4 is the minimum size.
7189 * Otherwise 2 is the minimum
7192 if (mddev
->level
== 6)
7194 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7198 if (!check_stripe_cache(mddev
))
7201 return resize_stripes(conf
, (conf
->previous_raid_disks
7202 + mddev
->delta_disks
));
7205 static int raid5_start_reshape(struct mddev
*mddev
)
7207 struct r5conf
*conf
= mddev
->private;
7208 struct md_rdev
*rdev
;
7210 unsigned long flags
;
7212 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7215 if (!check_stripe_cache(mddev
))
7218 if (has_failed(conf
))
7221 rdev_for_each(rdev
, mddev
) {
7222 if (!test_bit(In_sync
, &rdev
->flags
)
7223 && !test_bit(Faulty
, &rdev
->flags
))
7227 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7228 /* Not enough devices even to make a degraded array
7233 /* Refuse to reduce size of the array. Any reductions in
7234 * array size must be through explicit setting of array_size
7237 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7238 < mddev
->array_sectors
) {
7239 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
7240 "before number of disks\n", mdname(mddev
));
7244 atomic_set(&conf
->reshape_stripes
, 0);
7245 spin_lock_irq(&conf
->device_lock
);
7246 write_seqcount_begin(&conf
->gen_lock
);
7247 conf
->previous_raid_disks
= conf
->raid_disks
;
7248 conf
->raid_disks
+= mddev
->delta_disks
;
7249 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7250 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7251 conf
->prev_algo
= conf
->algorithm
;
7252 conf
->algorithm
= mddev
->new_layout
;
7254 /* Code that selects data_offset needs to see the generation update
7255 * if reshape_progress has been set - so a memory barrier needed.
7258 if (mddev
->reshape_backwards
)
7259 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7261 conf
->reshape_progress
= 0;
7262 conf
->reshape_safe
= conf
->reshape_progress
;
7263 write_seqcount_end(&conf
->gen_lock
);
7264 spin_unlock_irq(&conf
->device_lock
);
7266 /* Now make sure any requests that proceeded on the assumption
7267 * the reshape wasn't running - like Discard or Read - have
7270 mddev_suspend(mddev
);
7271 mddev_resume(mddev
);
7273 /* Add some new drives, as many as will fit.
7274 * We know there are enough to make the newly sized array work.
7275 * Don't add devices if we are reducing the number of
7276 * devices in the array. This is because it is not possible
7277 * to correctly record the "partially reconstructed" state of
7278 * such devices during the reshape and confusion could result.
7280 if (mddev
->delta_disks
>= 0) {
7281 rdev_for_each(rdev
, mddev
)
7282 if (rdev
->raid_disk
< 0 &&
7283 !test_bit(Faulty
, &rdev
->flags
)) {
7284 if (raid5_add_disk(mddev
, rdev
) == 0) {
7286 >= conf
->previous_raid_disks
)
7287 set_bit(In_sync
, &rdev
->flags
);
7289 rdev
->recovery_offset
= 0;
7291 if (sysfs_link_rdev(mddev
, rdev
))
7292 /* Failure here is OK */;
7294 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7295 && !test_bit(Faulty
, &rdev
->flags
)) {
7296 /* This is a spare that was manually added */
7297 set_bit(In_sync
, &rdev
->flags
);
7300 /* When a reshape changes the number of devices,
7301 * ->degraded is measured against the larger of the
7302 * pre and post number of devices.
7304 spin_lock_irqsave(&conf
->device_lock
, flags
);
7305 mddev
->degraded
= calc_degraded(conf
);
7306 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7308 mddev
->raid_disks
= conf
->raid_disks
;
7309 mddev
->reshape_position
= conf
->reshape_progress
;
7310 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7312 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7313 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7314 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7315 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7316 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7318 if (!mddev
->sync_thread
) {
7319 mddev
->recovery
= 0;
7320 spin_lock_irq(&conf
->device_lock
);
7321 write_seqcount_begin(&conf
->gen_lock
);
7322 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7323 mddev
->new_chunk_sectors
=
7324 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7325 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7326 rdev_for_each(rdev
, mddev
)
7327 rdev
->new_data_offset
= rdev
->data_offset
;
7329 conf
->generation
--;
7330 conf
->reshape_progress
= MaxSector
;
7331 mddev
->reshape_position
= MaxSector
;
7332 write_seqcount_end(&conf
->gen_lock
);
7333 spin_unlock_irq(&conf
->device_lock
);
7336 conf
->reshape_checkpoint
= jiffies
;
7337 md_wakeup_thread(mddev
->sync_thread
);
7338 md_new_event(mddev
);
7342 /* This is called from the reshape thread and should make any
7343 * changes needed in 'conf'
7345 static void end_reshape(struct r5conf
*conf
)
7348 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7349 struct md_rdev
*rdev
;
7351 spin_lock_irq(&conf
->device_lock
);
7352 conf
->previous_raid_disks
= conf
->raid_disks
;
7353 rdev_for_each(rdev
, conf
->mddev
)
7354 rdev
->data_offset
= rdev
->new_data_offset
;
7356 conf
->reshape_progress
= MaxSector
;
7357 spin_unlock_irq(&conf
->device_lock
);
7358 wake_up(&conf
->wait_for_overlap
);
7360 /* read-ahead size must cover two whole stripes, which is
7361 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7363 if (conf
->mddev
->queue
) {
7364 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7365 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7367 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
7368 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
7373 /* This is called from the raid5d thread with mddev_lock held.
7374 * It makes config changes to the device.
7376 static void raid5_finish_reshape(struct mddev
*mddev
)
7378 struct r5conf
*conf
= mddev
->private;
7380 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7382 if (mddev
->delta_disks
> 0) {
7383 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7384 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7385 revalidate_disk(mddev
->gendisk
);
7388 spin_lock_irq(&conf
->device_lock
);
7389 mddev
->degraded
= calc_degraded(conf
);
7390 spin_unlock_irq(&conf
->device_lock
);
7391 for (d
= conf
->raid_disks
;
7392 d
< conf
->raid_disks
- mddev
->delta_disks
;
7394 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
7396 clear_bit(In_sync
, &rdev
->flags
);
7397 rdev
= conf
->disks
[d
].replacement
;
7399 clear_bit(In_sync
, &rdev
->flags
);
7402 mddev
->layout
= conf
->algorithm
;
7403 mddev
->chunk_sectors
= conf
->chunk_sectors
;
7404 mddev
->reshape_position
= MaxSector
;
7405 mddev
->delta_disks
= 0;
7406 mddev
->reshape_backwards
= 0;
7410 static void raid5_quiesce(struct mddev
*mddev
, int state
)
7412 struct r5conf
*conf
= mddev
->private;
7415 case 2: /* resume for a suspend */
7416 wake_up(&conf
->wait_for_overlap
);
7419 case 1: /* stop all writes */
7420 lock_all_device_hash_locks_irq(conf
);
7421 /* '2' tells resync/reshape to pause so that all
7422 * active stripes can drain
7425 wait_event_cmd(conf
->wait_for_stripe
,
7426 atomic_read(&conf
->active_stripes
) == 0 &&
7427 atomic_read(&conf
->active_aligned_reads
) == 0,
7428 unlock_all_device_hash_locks_irq(conf
),
7429 lock_all_device_hash_locks_irq(conf
));
7431 unlock_all_device_hash_locks_irq(conf
);
7432 /* allow reshape to continue */
7433 wake_up(&conf
->wait_for_overlap
);
7436 case 0: /* re-enable writes */
7437 lock_all_device_hash_locks_irq(conf
);
7439 wake_up(&conf
->wait_for_stripe
);
7440 wake_up(&conf
->wait_for_overlap
);
7441 unlock_all_device_hash_locks_irq(conf
);
7446 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
7448 struct r0conf
*raid0_conf
= mddev
->private;
7451 /* for raid0 takeover only one zone is supported */
7452 if (raid0_conf
->nr_strip_zones
> 1) {
7453 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7455 return ERR_PTR(-EINVAL
);
7458 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
7459 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
7460 mddev
->dev_sectors
= sectors
;
7461 mddev
->new_level
= level
;
7462 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7463 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
7464 mddev
->raid_disks
+= 1;
7465 mddev
->delta_disks
= 1;
7466 /* make sure it will be not marked as dirty */
7467 mddev
->recovery_cp
= MaxSector
;
7469 return setup_conf(mddev
);
7472 static void *raid5_takeover_raid1(struct mddev
*mddev
)
7476 if (mddev
->raid_disks
!= 2 ||
7477 mddev
->degraded
> 1)
7478 return ERR_PTR(-EINVAL
);
7480 /* Should check if there are write-behind devices? */
7482 chunksect
= 64*2; /* 64K by default */
7484 /* The array must be an exact multiple of chunksize */
7485 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
7488 if ((chunksect
<<9) < STRIPE_SIZE
)
7489 /* array size does not allow a suitable chunk size */
7490 return ERR_PTR(-EINVAL
);
7492 mddev
->new_level
= 5;
7493 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7494 mddev
->new_chunk_sectors
= chunksect
;
7496 return setup_conf(mddev
);
7499 static void *raid5_takeover_raid6(struct mddev
*mddev
)
7503 switch (mddev
->layout
) {
7504 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7505 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
7507 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7508 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
7510 case ALGORITHM_LEFT_SYMMETRIC_6
:
7511 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7513 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7514 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
7516 case ALGORITHM_PARITY_0_6
:
7517 new_layout
= ALGORITHM_PARITY_0
;
7519 case ALGORITHM_PARITY_N
:
7520 new_layout
= ALGORITHM_PARITY_N
;
7523 return ERR_PTR(-EINVAL
);
7525 mddev
->new_level
= 5;
7526 mddev
->new_layout
= new_layout
;
7527 mddev
->delta_disks
= -1;
7528 mddev
->raid_disks
-= 1;
7529 return setup_conf(mddev
);
7532 static int raid5_check_reshape(struct mddev
*mddev
)
7534 /* For a 2-drive array, the layout and chunk size can be changed
7535 * immediately as not restriping is needed.
7536 * For larger arrays we record the new value - after validation
7537 * to be used by a reshape pass.
7539 struct r5conf
*conf
= mddev
->private;
7540 int new_chunk
= mddev
->new_chunk_sectors
;
7542 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
7544 if (new_chunk
> 0) {
7545 if (!is_power_of_2(new_chunk
))
7547 if (new_chunk
< (PAGE_SIZE
>>9))
7549 if (mddev
->array_sectors
& (new_chunk
-1))
7550 /* not factor of array size */
7554 /* They look valid */
7556 if (mddev
->raid_disks
== 2) {
7557 /* can make the change immediately */
7558 if (mddev
->new_layout
>= 0) {
7559 conf
->algorithm
= mddev
->new_layout
;
7560 mddev
->layout
= mddev
->new_layout
;
7562 if (new_chunk
> 0) {
7563 conf
->chunk_sectors
= new_chunk
;
7564 mddev
->chunk_sectors
= new_chunk
;
7566 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7567 md_wakeup_thread(mddev
->thread
);
7569 return check_reshape(mddev
);
7572 static int raid6_check_reshape(struct mddev
*mddev
)
7574 int new_chunk
= mddev
->new_chunk_sectors
;
7576 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
7578 if (new_chunk
> 0) {
7579 if (!is_power_of_2(new_chunk
))
7581 if (new_chunk
< (PAGE_SIZE
>> 9))
7583 if (mddev
->array_sectors
& (new_chunk
-1))
7584 /* not factor of array size */
7588 /* They look valid */
7589 return check_reshape(mddev
);
7592 static void *raid5_takeover(struct mddev
*mddev
)
7594 /* raid5 can take over:
7595 * raid0 - if there is only one strip zone - make it a raid4 layout
7596 * raid1 - if there are two drives. We need to know the chunk size
7597 * raid4 - trivial - just use a raid4 layout.
7598 * raid6 - Providing it is a *_6 layout
7600 if (mddev
->level
== 0)
7601 return raid45_takeover_raid0(mddev
, 5);
7602 if (mddev
->level
== 1)
7603 return raid5_takeover_raid1(mddev
);
7604 if (mddev
->level
== 4) {
7605 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7606 mddev
->new_level
= 5;
7607 return setup_conf(mddev
);
7609 if (mddev
->level
== 6)
7610 return raid5_takeover_raid6(mddev
);
7612 return ERR_PTR(-EINVAL
);
7615 static void *raid4_takeover(struct mddev
*mddev
)
7617 /* raid4 can take over:
7618 * raid0 - if there is only one strip zone
7619 * raid5 - if layout is right
7621 if (mddev
->level
== 0)
7622 return raid45_takeover_raid0(mddev
, 4);
7623 if (mddev
->level
== 5 &&
7624 mddev
->layout
== ALGORITHM_PARITY_N
) {
7625 mddev
->new_layout
= 0;
7626 mddev
->new_level
= 4;
7627 return setup_conf(mddev
);
7629 return ERR_PTR(-EINVAL
);
7632 static struct md_personality raid5_personality
;
7634 static void *raid6_takeover(struct mddev
*mddev
)
7636 /* Currently can only take over a raid5. We map the
7637 * personality to an equivalent raid6 personality
7638 * with the Q block at the end.
7642 if (mddev
->pers
!= &raid5_personality
)
7643 return ERR_PTR(-EINVAL
);
7644 if (mddev
->degraded
> 1)
7645 return ERR_PTR(-EINVAL
);
7646 if (mddev
->raid_disks
> 253)
7647 return ERR_PTR(-EINVAL
);
7648 if (mddev
->raid_disks
< 3)
7649 return ERR_PTR(-EINVAL
);
7651 switch (mddev
->layout
) {
7652 case ALGORITHM_LEFT_ASYMMETRIC
:
7653 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
7655 case ALGORITHM_RIGHT_ASYMMETRIC
:
7656 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
7658 case ALGORITHM_LEFT_SYMMETRIC
:
7659 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
7661 case ALGORITHM_RIGHT_SYMMETRIC
:
7662 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
7664 case ALGORITHM_PARITY_0
:
7665 new_layout
= ALGORITHM_PARITY_0_6
;
7667 case ALGORITHM_PARITY_N
:
7668 new_layout
= ALGORITHM_PARITY_N
;
7671 return ERR_PTR(-EINVAL
);
7673 mddev
->new_level
= 6;
7674 mddev
->new_layout
= new_layout
;
7675 mddev
->delta_disks
= 1;
7676 mddev
->raid_disks
+= 1;
7677 return setup_conf(mddev
);
7680 static struct md_personality raid6_personality
=
7684 .owner
= THIS_MODULE
,
7685 .make_request
= make_request
,
7689 .error_handler
= error
,
7690 .hot_add_disk
= raid5_add_disk
,
7691 .hot_remove_disk
= raid5_remove_disk
,
7692 .spare_active
= raid5_spare_active
,
7693 .sync_request
= sync_request
,
7694 .resize
= raid5_resize
,
7696 .check_reshape
= raid6_check_reshape
,
7697 .start_reshape
= raid5_start_reshape
,
7698 .finish_reshape
= raid5_finish_reshape
,
7699 .quiesce
= raid5_quiesce
,
7700 .takeover
= raid6_takeover
,
7701 .congested
= raid5_congested
,
7702 .mergeable_bvec
= raid5_mergeable_bvec
,
7704 static struct md_personality raid5_personality
=
7708 .owner
= THIS_MODULE
,
7709 .make_request
= make_request
,
7713 .error_handler
= error
,
7714 .hot_add_disk
= raid5_add_disk
,
7715 .hot_remove_disk
= raid5_remove_disk
,
7716 .spare_active
= raid5_spare_active
,
7717 .sync_request
= sync_request
,
7718 .resize
= raid5_resize
,
7720 .check_reshape
= raid5_check_reshape
,
7721 .start_reshape
= raid5_start_reshape
,
7722 .finish_reshape
= raid5_finish_reshape
,
7723 .quiesce
= raid5_quiesce
,
7724 .takeover
= raid5_takeover
,
7725 .congested
= raid5_congested
,
7726 .mergeable_bvec
= raid5_mergeable_bvec
,
7729 static struct md_personality raid4_personality
=
7733 .owner
= THIS_MODULE
,
7734 .make_request
= make_request
,
7738 .error_handler
= error
,
7739 .hot_add_disk
= raid5_add_disk
,
7740 .hot_remove_disk
= raid5_remove_disk
,
7741 .spare_active
= raid5_spare_active
,
7742 .sync_request
= sync_request
,
7743 .resize
= raid5_resize
,
7745 .check_reshape
= raid5_check_reshape
,
7746 .start_reshape
= raid5_start_reshape
,
7747 .finish_reshape
= raid5_finish_reshape
,
7748 .quiesce
= raid5_quiesce
,
7749 .takeover
= raid4_takeover
,
7750 .congested
= raid5_congested
,
7751 .mergeable_bvec
= raid5_mergeable_bvec
,
7754 static int __init
raid5_init(void)
7756 raid5_wq
= alloc_workqueue("raid5wq",
7757 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
7760 register_md_personality(&raid6_personality
);
7761 register_md_personality(&raid5_personality
);
7762 register_md_personality(&raid4_personality
);
7766 static void raid5_exit(void)
7768 unregister_md_personality(&raid6_personality
);
7769 unregister_md_personality(&raid5_personality
);
7770 unregister_md_personality(&raid4_personality
);
7771 destroy_workqueue(raid5_wq
);
7774 module_init(raid5_init
);
7775 module_exit(raid5_exit
);
7776 MODULE_LICENSE("GPL");
7777 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7778 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7779 MODULE_ALIAS("md-raid5");
7780 MODULE_ALIAS("md-raid4");
7781 MODULE_ALIAS("md-level-5");
7782 MODULE_ALIAS("md-level-4");
7783 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7784 MODULE_ALIAS("md-raid6");
7785 MODULE_ALIAS("md-level-6");
7787 /* This used to be two separate modules, they were: */
7788 MODULE_ALIAS("raid5");
7789 MODULE_ALIAS("raid6");