2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <trace/events/block.h>
65 #define cpu_to_group(cpu) cpu_to_node(cpu)
66 #define ANY_GROUP NUMA_NO_NODE
68 static bool devices_handle_discard_safely
= false;
69 module_param(devices_handle_discard_safely
, bool, 0644);
70 MODULE_PARM_DESC(devices_handle_discard_safely
,
71 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
72 static struct workqueue_struct
*raid5_wq
;
77 #define NR_STRIPES 256
78 #define STRIPE_SIZE PAGE_SIZE
79 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
80 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
81 #define IO_THRESHOLD 1
82 #define BYPASS_THRESHOLD 1
83 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
84 #define HASH_MASK (NR_HASH - 1)
85 #define MAX_STRIPE_BATCH 8
87 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
89 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
90 return &conf
->stripe_hashtbl
[hash
];
93 static inline int stripe_hash_locks_hash(sector_t sect
)
95 return (sect
>> STRIPE_SHIFT
) & STRIPE_HASH_LOCKS_MASK
;
98 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
100 spin_lock_irq(conf
->hash_locks
+ hash
);
101 spin_lock(&conf
->device_lock
);
104 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
106 spin_unlock(&conf
->device_lock
);
107 spin_unlock_irq(conf
->hash_locks
+ hash
);
110 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
114 spin_lock(conf
->hash_locks
);
115 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
116 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
117 spin_lock(&conf
->device_lock
);
120 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
123 spin_unlock(&conf
->device_lock
);
124 for (i
= NR_STRIPE_HASH_LOCKS
; i
; i
--)
125 spin_unlock(conf
->hash_locks
+ i
- 1);
129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
130 * order without overlap. There may be several bio's per stripe+device, and
131 * a bio could span several devices.
132 * When walking this list for a particular stripe+device, we must never proceed
133 * beyond a bio that extends past this device, as the next bio might no longer
135 * This function is used to determine the 'next' bio in the list, given the sector
136 * of the current stripe+device
138 static inline struct bio
*r5_next_bio(struct bio
*bio
, sector_t sector
)
140 int sectors
= bio_sectors(bio
);
141 if (bio
->bi_iter
.bi_sector
+ sectors
< sector
+ STRIPE_SECTORS
)
148 * We maintain a biased count of active stripes in the bottom 16 bits of
149 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
151 static inline int raid5_bi_processed_stripes(struct bio
*bio
)
153 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
154 return (atomic_read(segments
) >> 16) & 0xffff;
157 static inline int raid5_dec_bi_active_stripes(struct bio
*bio
)
159 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
160 return atomic_sub_return(1, segments
) & 0xffff;
163 static inline void raid5_inc_bi_active_stripes(struct bio
*bio
)
165 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
166 atomic_inc(segments
);
169 static inline void raid5_set_bi_processed_stripes(struct bio
*bio
,
172 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
176 old
= atomic_read(segments
);
177 new = (old
& 0xffff) | (cnt
<< 16);
178 } while (atomic_cmpxchg(segments
, old
, new) != old
);
181 static inline void raid5_set_bi_stripes(struct bio
*bio
, unsigned int cnt
)
183 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
184 atomic_set(segments
, cnt
);
187 /* Find first data disk in a raid6 stripe */
188 static inline int raid6_d0(struct stripe_head
*sh
)
191 /* ddf always start from first device */
193 /* md starts just after Q block */
194 if (sh
->qd_idx
== sh
->disks
- 1)
197 return sh
->qd_idx
+ 1;
199 static inline int raid6_next_disk(int disk
, int raid_disks
)
202 return (disk
< raid_disks
) ? disk
: 0;
205 /* When walking through the disks in a raid5, starting at raid6_d0,
206 * We need to map each disk to a 'slot', where the data disks are slot
207 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
208 * is raid_disks-1. This help does that mapping.
210 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
211 int *count
, int syndrome_disks
)
217 if (idx
== sh
->pd_idx
)
218 return syndrome_disks
;
219 if (idx
== sh
->qd_idx
)
220 return syndrome_disks
+ 1;
226 static void return_io(struct bio
*return_bi
)
228 struct bio
*bi
= return_bi
;
231 return_bi
= bi
->bi_next
;
233 bi
->bi_iter
.bi_size
= 0;
234 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
241 static void print_raid5_conf (struct r5conf
*conf
);
243 static int stripe_operations_active(struct stripe_head
*sh
)
245 return sh
->check_state
|| sh
->reconstruct_state
||
246 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
247 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
250 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
252 struct r5conf
*conf
= sh
->raid_conf
;
253 struct r5worker_group
*group
;
255 int i
, cpu
= sh
->cpu
;
257 if (!cpu_online(cpu
)) {
258 cpu
= cpumask_any(cpu_online_mask
);
262 if (list_empty(&sh
->lru
)) {
263 struct r5worker_group
*group
;
264 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
265 list_add_tail(&sh
->lru
, &group
->handle_list
);
266 group
->stripes_cnt
++;
270 if (conf
->worker_cnt_per_group
== 0) {
271 md_wakeup_thread(conf
->mddev
->thread
);
275 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
277 group
->workers
[0].working
= true;
278 /* at least one worker should run to avoid race */
279 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
281 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
282 /* wakeup more workers */
283 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
284 if (group
->workers
[i
].working
== false) {
285 group
->workers
[i
].working
= true;
286 queue_work_on(sh
->cpu
, raid5_wq
,
287 &group
->workers
[i
].work
);
293 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
294 struct list_head
*temp_inactive_list
)
296 BUG_ON(!list_empty(&sh
->lru
));
297 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
298 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
299 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
300 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
301 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
302 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
303 sh
->bm_seq
- conf
->seq_write
> 0)
304 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
306 clear_bit(STRIPE_DELAYED
, &sh
->state
);
307 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
308 if (conf
->worker_cnt_per_group
== 0) {
309 list_add_tail(&sh
->lru
, &conf
->handle_list
);
311 raid5_wakeup_stripe_thread(sh
);
315 md_wakeup_thread(conf
->mddev
->thread
);
317 BUG_ON(stripe_operations_active(sh
));
318 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
319 if (atomic_dec_return(&conf
->preread_active_stripes
)
321 md_wakeup_thread(conf
->mddev
->thread
);
322 atomic_dec(&conf
->active_stripes
);
323 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
))
324 list_add_tail(&sh
->lru
, temp_inactive_list
);
328 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
329 struct list_head
*temp_inactive_list
)
331 if (atomic_dec_and_test(&sh
->count
))
332 do_release_stripe(conf
, sh
, temp_inactive_list
);
336 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
338 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
339 * given time. Adding stripes only takes device lock, while deleting stripes
340 * only takes hash lock.
342 static void release_inactive_stripe_list(struct r5conf
*conf
,
343 struct list_head
*temp_inactive_list
,
347 bool do_wakeup
= false;
350 if (hash
== NR_STRIPE_HASH_LOCKS
) {
351 size
= NR_STRIPE_HASH_LOCKS
;
352 hash
= NR_STRIPE_HASH_LOCKS
- 1;
356 struct list_head
*list
= &temp_inactive_list
[size
- 1];
359 * We don't hold any lock here yet, get_active_stripe() might
360 * remove stripes from the list
362 if (!list_empty_careful(list
)) {
363 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
364 if (list_empty(conf
->inactive_list
+ hash
) &&
366 atomic_dec(&conf
->empty_inactive_list_nr
);
367 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
369 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
376 wake_up(&conf
->wait_for_stripe
);
377 if (conf
->retry_read_aligned
)
378 md_wakeup_thread(conf
->mddev
->thread
);
382 /* should hold conf->device_lock already */
383 static int release_stripe_list(struct r5conf
*conf
,
384 struct list_head
*temp_inactive_list
)
386 struct stripe_head
*sh
;
388 struct llist_node
*head
;
390 head
= llist_del_all(&conf
->released_stripes
);
391 head
= llist_reverse_order(head
);
395 sh
= llist_entry(head
, struct stripe_head
, release_list
);
396 head
= llist_next(head
);
397 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
399 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
401 * Don't worry the bit is set here, because if the bit is set
402 * again, the count is always > 1. This is true for
403 * STRIPE_ON_UNPLUG_LIST bit too.
405 hash
= sh
->hash_lock_index
;
406 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
413 static void release_stripe(struct stripe_head
*sh
)
415 struct r5conf
*conf
= sh
->raid_conf
;
417 struct list_head list
;
421 /* Avoid release_list until the last reference.
423 if (atomic_add_unless(&sh
->count
, -1, 1))
426 if (unlikely(!conf
->mddev
->thread
) ||
427 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
429 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
431 md_wakeup_thread(conf
->mddev
->thread
);
434 local_irq_save(flags
);
435 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
436 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
437 INIT_LIST_HEAD(&list
);
438 hash
= sh
->hash_lock_index
;
439 do_release_stripe(conf
, sh
, &list
);
440 spin_unlock(&conf
->device_lock
);
441 release_inactive_stripe_list(conf
, &list
, hash
);
443 local_irq_restore(flags
);
446 static inline void remove_hash(struct stripe_head
*sh
)
448 pr_debug("remove_hash(), stripe %llu\n",
449 (unsigned long long)sh
->sector
);
451 hlist_del_init(&sh
->hash
);
454 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
456 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
458 pr_debug("insert_hash(), stripe %llu\n",
459 (unsigned long long)sh
->sector
);
461 hlist_add_head(&sh
->hash
, hp
);
464 /* find an idle stripe, make sure it is unhashed, and return it. */
465 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
467 struct stripe_head
*sh
= NULL
;
468 struct list_head
*first
;
470 if (list_empty(conf
->inactive_list
+ hash
))
472 first
= (conf
->inactive_list
+ hash
)->next
;
473 sh
= list_entry(first
, struct stripe_head
, lru
);
474 list_del_init(first
);
476 atomic_inc(&conf
->active_stripes
);
477 BUG_ON(hash
!= sh
->hash_lock_index
);
478 if (list_empty(conf
->inactive_list
+ hash
))
479 atomic_inc(&conf
->empty_inactive_list_nr
);
484 static void shrink_buffers(struct stripe_head
*sh
)
488 int num
= sh
->raid_conf
->pool_size
;
490 for (i
= 0; i
< num
; i
++) {
491 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
495 sh
->dev
[i
].page
= NULL
;
500 static int grow_buffers(struct stripe_head
*sh
)
503 int num
= sh
->raid_conf
->pool_size
;
505 for (i
= 0; i
< num
; i
++) {
508 if (!(page
= alloc_page(GFP_KERNEL
))) {
511 sh
->dev
[i
].page
= page
;
512 sh
->dev
[i
].orig_page
= page
;
517 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
518 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
519 struct stripe_head
*sh
);
521 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
523 struct r5conf
*conf
= sh
->raid_conf
;
526 BUG_ON(atomic_read(&sh
->count
) != 0);
527 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
528 BUG_ON(stripe_operations_active(sh
));
529 BUG_ON(sh
->batch_head
);
531 pr_debug("init_stripe called, stripe %llu\n",
532 (unsigned long long)sector
);
534 seq
= read_seqcount_begin(&conf
->gen_lock
);
535 sh
->generation
= conf
->generation
- previous
;
536 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
538 stripe_set_idx(sector
, conf
, previous
, sh
);
541 for (i
= sh
->disks
; i
--; ) {
542 struct r5dev
*dev
= &sh
->dev
[i
];
544 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
545 test_bit(R5_LOCKED
, &dev
->flags
)) {
546 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
547 (unsigned long long)sh
->sector
, i
, dev
->toread
,
548 dev
->read
, dev
->towrite
, dev
->written
,
549 test_bit(R5_LOCKED
, &dev
->flags
));
553 raid5_build_block(sh
, i
, previous
);
555 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
557 sh
->overwrite_disks
= 0;
558 insert_hash(conf
, sh
);
559 sh
->cpu
= smp_processor_id();
560 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
563 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
566 struct stripe_head
*sh
;
568 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
569 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
570 if (sh
->sector
== sector
&& sh
->generation
== generation
)
572 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
577 * Need to check if array has failed when deciding whether to:
579 * - remove non-faulty devices
582 * This determination is simple when no reshape is happening.
583 * However if there is a reshape, we need to carefully check
584 * both the before and after sections.
585 * This is because some failed devices may only affect one
586 * of the two sections, and some non-in_sync devices may
587 * be insync in the section most affected by failed devices.
589 static int calc_degraded(struct r5conf
*conf
)
591 int degraded
, degraded2
;
596 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
597 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
598 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
599 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
600 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
602 else if (test_bit(In_sync
, &rdev
->flags
))
605 /* not in-sync or faulty.
606 * If the reshape increases the number of devices,
607 * this is being recovered by the reshape, so
608 * this 'previous' section is not in_sync.
609 * If the number of devices is being reduced however,
610 * the device can only be part of the array if
611 * we are reverting a reshape, so this section will
614 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
618 if (conf
->raid_disks
== conf
->previous_raid_disks
)
622 for (i
= 0; i
< conf
->raid_disks
; i
++) {
623 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
624 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
625 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
626 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
628 else if (test_bit(In_sync
, &rdev
->flags
))
631 /* not in-sync or faulty.
632 * If reshape increases the number of devices, this
633 * section has already been recovered, else it
634 * almost certainly hasn't.
636 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
640 if (degraded2
> degraded
)
645 static int has_failed(struct r5conf
*conf
)
649 if (conf
->mddev
->reshape_position
== MaxSector
)
650 return conf
->mddev
->degraded
> conf
->max_degraded
;
652 degraded
= calc_degraded(conf
);
653 if (degraded
> conf
->max_degraded
)
658 static struct stripe_head
*
659 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
660 int previous
, int noblock
, int noquiesce
)
662 struct stripe_head
*sh
;
663 int hash
= stripe_hash_locks_hash(sector
);
665 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
667 spin_lock_irq(conf
->hash_locks
+ hash
);
670 wait_event_lock_irq(conf
->wait_for_stripe
,
671 conf
->quiesce
== 0 || noquiesce
,
672 *(conf
->hash_locks
+ hash
));
673 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
675 if (!conf
->inactive_blocked
)
676 sh
= get_free_stripe(conf
, hash
);
677 if (noblock
&& sh
== NULL
)
680 conf
->inactive_blocked
= 1;
682 conf
->wait_for_stripe
,
683 !list_empty(conf
->inactive_list
+ hash
) &&
684 (atomic_read(&conf
->active_stripes
)
685 < (conf
->max_nr_stripes
* 3 / 4)
686 || !conf
->inactive_blocked
),
687 *(conf
->hash_locks
+ hash
));
688 conf
->inactive_blocked
= 0;
690 init_stripe(sh
, sector
, previous
);
691 atomic_inc(&sh
->count
);
693 } else if (!atomic_inc_not_zero(&sh
->count
)) {
694 spin_lock(&conf
->device_lock
);
695 if (!atomic_read(&sh
->count
)) {
696 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
697 atomic_inc(&conf
->active_stripes
);
698 BUG_ON(list_empty(&sh
->lru
) &&
699 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
700 list_del_init(&sh
->lru
);
702 sh
->group
->stripes_cnt
--;
706 atomic_inc(&sh
->count
);
707 spin_unlock(&conf
->device_lock
);
709 } while (sh
== NULL
);
711 spin_unlock_irq(conf
->hash_locks
+ hash
);
715 static bool is_full_stripe_write(struct stripe_head
*sh
)
717 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
718 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
721 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
725 spin_lock(&sh2
->stripe_lock
);
726 spin_lock_nested(&sh1
->stripe_lock
, 1);
728 spin_lock(&sh1
->stripe_lock
);
729 spin_lock_nested(&sh2
->stripe_lock
, 1);
733 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
735 spin_unlock(&sh1
->stripe_lock
);
736 spin_unlock(&sh2
->stripe_lock
);
740 /* Only freshly new full stripe normal write stripe can be added to a batch list */
741 static bool stripe_can_batch(struct stripe_head
*sh
)
743 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
744 is_full_stripe_write(sh
);
747 /* we only do back search */
748 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
750 struct stripe_head
*head
;
751 sector_t head_sector
, tmp_sec
;
755 if (!stripe_can_batch(sh
))
757 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
758 tmp_sec
= sh
->sector
;
759 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
761 head_sector
= sh
->sector
- STRIPE_SECTORS
;
763 hash
= stripe_hash_locks_hash(head_sector
);
764 spin_lock_irq(conf
->hash_locks
+ hash
);
765 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
766 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
767 spin_lock(&conf
->device_lock
);
768 if (!atomic_read(&head
->count
)) {
769 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
770 atomic_inc(&conf
->active_stripes
);
771 BUG_ON(list_empty(&head
->lru
) &&
772 !test_bit(STRIPE_EXPANDING
, &head
->state
));
773 list_del_init(&head
->lru
);
775 head
->group
->stripes_cnt
--;
779 atomic_inc(&head
->count
);
780 spin_unlock(&conf
->device_lock
);
782 spin_unlock_irq(conf
->hash_locks
+ hash
);
786 if (!stripe_can_batch(head
))
789 lock_two_stripes(head
, sh
);
790 /* clear_batch_ready clear the flag */
791 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
798 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
800 if (head
->dev
[dd_idx
].towrite
->bi_rw
!= sh
->dev
[dd_idx
].towrite
->bi_rw
)
803 if (head
->batch_head
) {
804 spin_lock(&head
->batch_head
->batch_lock
);
805 /* This batch list is already running */
806 if (!stripe_can_batch(head
)) {
807 spin_unlock(&head
->batch_head
->batch_lock
);
812 * at this point, head's BATCH_READY could be cleared, but we
813 * can still add the stripe to batch list
815 list_add(&sh
->batch_list
, &head
->batch_list
);
816 spin_unlock(&head
->batch_head
->batch_lock
);
818 sh
->batch_head
= head
->batch_head
;
820 head
->batch_head
= head
;
821 sh
->batch_head
= head
->batch_head
;
822 spin_lock(&head
->batch_lock
);
823 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
824 spin_unlock(&head
->batch_lock
);
827 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
828 if (atomic_dec_return(&conf
->preread_active_stripes
)
830 md_wakeup_thread(conf
->mddev
->thread
);
832 atomic_inc(&sh
->count
);
834 unlock_two_stripes(head
, sh
);
836 release_stripe(head
);
839 /* Determine if 'data_offset' or 'new_data_offset' should be used
840 * in this stripe_head.
842 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
844 sector_t progress
= conf
->reshape_progress
;
845 /* Need a memory barrier to make sure we see the value
846 * of conf->generation, or ->data_offset that was set before
847 * reshape_progress was updated.
850 if (progress
== MaxSector
)
852 if (sh
->generation
== conf
->generation
- 1)
854 /* We are in a reshape, and this is a new-generation stripe,
855 * so use new_data_offset.
861 raid5_end_read_request(struct bio
*bi
, int error
);
863 raid5_end_write_request(struct bio
*bi
, int error
);
865 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
867 struct r5conf
*conf
= sh
->raid_conf
;
868 int i
, disks
= sh
->disks
;
869 struct stripe_head
*head_sh
= sh
;
873 for (i
= disks
; i
--; ) {
875 int replace_only
= 0;
876 struct bio
*bi
, *rbi
;
877 struct md_rdev
*rdev
, *rrdev
= NULL
;
880 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
881 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
885 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
887 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
889 else if (test_and_clear_bit(R5_WantReplace
,
890 &sh
->dev
[i
].flags
)) {
895 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
899 bi
= &sh
->dev
[i
].req
;
900 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
903 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
904 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
905 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
914 /* We raced and saw duplicates */
917 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
922 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
925 atomic_inc(&rdev
->nr_pending
);
926 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
929 atomic_inc(&rrdev
->nr_pending
);
932 /* We have already checked bad blocks for reads. Now
933 * need to check for writes. We never accept write errors
934 * on the replacement, so we don't to check rrdev.
936 while ((rw
& WRITE
) && rdev
&&
937 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
940 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
941 &first_bad
, &bad_sectors
);
946 set_bit(BlockedBadBlocks
, &rdev
->flags
);
947 if (!conf
->mddev
->external
&&
948 conf
->mddev
->flags
) {
949 /* It is very unlikely, but we might
950 * still need to write out the
951 * bad block log - better give it
953 md_check_recovery(conf
->mddev
);
956 * Because md_wait_for_blocked_rdev
957 * will dec nr_pending, we must
958 * increment it first.
960 atomic_inc(&rdev
->nr_pending
);
961 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
963 /* Acknowledged bad block - skip the write */
964 rdev_dec_pending(rdev
, conf
->mddev
);
970 if (s
->syncing
|| s
->expanding
|| s
->expanded
972 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
974 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
977 bi
->bi_bdev
= rdev
->bdev
;
979 bi
->bi_end_io
= (rw
& WRITE
)
980 ? raid5_end_write_request
981 : raid5_end_read_request
;
984 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
985 __func__
, (unsigned long long)sh
->sector
,
987 atomic_inc(&sh
->count
);
989 atomic_inc(&head_sh
->count
);
990 if (use_new_offset(conf
, sh
))
991 bi
->bi_iter
.bi_sector
= (sh
->sector
992 + rdev
->new_data_offset
);
994 bi
->bi_iter
.bi_sector
= (sh
->sector
995 + rdev
->data_offset
);
996 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
997 bi
->bi_rw
|= REQ_NOMERGE
;
999 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1000 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1001 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1003 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1004 bi
->bi_io_vec
[0].bv_offset
= 0;
1005 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1007 * If this is discard request, set bi_vcnt 0. We don't
1008 * want to confuse SCSI because SCSI will replace payload
1010 if (rw
& REQ_DISCARD
)
1013 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1015 if (conf
->mddev
->gendisk
)
1016 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
1017 bi
, disk_devt(conf
->mddev
->gendisk
),
1019 generic_make_request(bi
);
1022 if (s
->syncing
|| s
->expanding
|| s
->expanded
1024 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1026 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1029 rbi
->bi_bdev
= rrdev
->bdev
;
1031 BUG_ON(!(rw
& WRITE
));
1032 rbi
->bi_end_io
= raid5_end_write_request
;
1033 rbi
->bi_private
= sh
;
1035 pr_debug("%s: for %llu schedule op %ld on "
1036 "replacement disc %d\n",
1037 __func__
, (unsigned long long)sh
->sector
,
1039 atomic_inc(&sh
->count
);
1041 atomic_inc(&head_sh
->count
);
1042 if (use_new_offset(conf
, sh
))
1043 rbi
->bi_iter
.bi_sector
= (sh
->sector
1044 + rrdev
->new_data_offset
);
1046 rbi
->bi_iter
.bi_sector
= (sh
->sector
1047 + rrdev
->data_offset
);
1048 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1049 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1050 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1052 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1053 rbi
->bi_io_vec
[0].bv_offset
= 0;
1054 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1056 * If this is discard request, set bi_vcnt 0. We don't
1057 * want to confuse SCSI because SCSI will replace payload
1059 if (rw
& REQ_DISCARD
)
1061 if (conf
->mddev
->gendisk
)
1062 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
1063 rbi
, disk_devt(conf
->mddev
->gendisk
),
1065 generic_make_request(rbi
);
1067 if (!rdev
&& !rrdev
) {
1069 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1070 pr_debug("skip op %ld on disc %d for sector %llu\n",
1071 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
1072 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1074 set_bit(STRIPE_BATCH_ERR
,
1075 &sh
->batch_head
->state
);
1076 set_bit(STRIPE_HANDLE
, &sh
->state
);
1079 if (!head_sh
->batch_head
)
1081 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1088 static struct dma_async_tx_descriptor
*
1089 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1090 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1091 struct stripe_head
*sh
)
1094 struct bvec_iter iter
;
1095 struct page
*bio_page
;
1097 struct async_submit_ctl submit
;
1098 enum async_tx_flags flags
= 0;
1100 if (bio
->bi_iter
.bi_sector
>= sector
)
1101 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1103 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1106 flags
|= ASYNC_TX_FENCE
;
1107 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1109 bio_for_each_segment(bvl
, bio
, iter
) {
1110 int len
= bvl
.bv_len
;
1114 if (page_offset
< 0) {
1115 b_offset
= -page_offset
;
1116 page_offset
+= b_offset
;
1120 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1121 clen
= STRIPE_SIZE
- page_offset
;
1126 b_offset
+= bvl
.bv_offset
;
1127 bio_page
= bvl
.bv_page
;
1129 if (sh
->raid_conf
->skip_copy
&&
1130 b_offset
== 0 && page_offset
== 0 &&
1131 clen
== STRIPE_SIZE
)
1134 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1135 b_offset
, clen
, &submit
);
1137 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1138 page_offset
, clen
, &submit
);
1140 /* chain the operations */
1141 submit
.depend_tx
= tx
;
1143 if (clen
< len
) /* hit end of page */
1151 static void ops_complete_biofill(void *stripe_head_ref
)
1153 struct stripe_head
*sh
= stripe_head_ref
;
1154 struct bio
*return_bi
= NULL
;
1157 pr_debug("%s: stripe %llu\n", __func__
,
1158 (unsigned long long)sh
->sector
);
1160 /* clear completed biofills */
1161 for (i
= sh
->disks
; i
--; ) {
1162 struct r5dev
*dev
= &sh
->dev
[i
];
1164 /* acknowledge completion of a biofill operation */
1165 /* and check if we need to reply to a read request,
1166 * new R5_Wantfill requests are held off until
1167 * !STRIPE_BIOFILL_RUN
1169 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1170 struct bio
*rbi
, *rbi2
;
1175 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1176 dev
->sector
+ STRIPE_SECTORS
) {
1177 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1178 if (!raid5_dec_bi_active_stripes(rbi
)) {
1179 rbi
->bi_next
= return_bi
;
1186 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1188 return_io(return_bi
);
1190 set_bit(STRIPE_HANDLE
, &sh
->state
);
1194 static void ops_run_biofill(struct stripe_head
*sh
)
1196 struct dma_async_tx_descriptor
*tx
= NULL
;
1197 struct async_submit_ctl submit
;
1200 BUG_ON(sh
->batch_head
);
1201 pr_debug("%s: stripe %llu\n", __func__
,
1202 (unsigned long long)sh
->sector
);
1204 for (i
= sh
->disks
; i
--; ) {
1205 struct r5dev
*dev
= &sh
->dev
[i
];
1206 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1208 spin_lock_irq(&sh
->stripe_lock
);
1209 dev
->read
= rbi
= dev
->toread
;
1211 spin_unlock_irq(&sh
->stripe_lock
);
1212 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1213 dev
->sector
+ STRIPE_SECTORS
) {
1214 tx
= async_copy_data(0, rbi
, &dev
->page
,
1215 dev
->sector
, tx
, sh
);
1216 rbi
= r5_next_bio(rbi
, dev
->sector
);
1221 atomic_inc(&sh
->count
);
1222 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1223 async_trigger_callback(&submit
);
1226 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1233 tgt
= &sh
->dev
[target
];
1234 set_bit(R5_UPTODATE
, &tgt
->flags
);
1235 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1236 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1239 static void ops_complete_compute(void *stripe_head_ref
)
1241 struct stripe_head
*sh
= stripe_head_ref
;
1243 pr_debug("%s: stripe %llu\n", __func__
,
1244 (unsigned long long)sh
->sector
);
1246 /* mark the computed target(s) as uptodate */
1247 mark_target_uptodate(sh
, sh
->ops
.target
);
1248 mark_target_uptodate(sh
, sh
->ops
.target2
);
1250 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1251 if (sh
->check_state
== check_state_compute_run
)
1252 sh
->check_state
= check_state_compute_result
;
1253 set_bit(STRIPE_HANDLE
, &sh
->state
);
1257 /* return a pointer to the address conversion region of the scribble buffer */
1258 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1259 struct raid5_percpu
*percpu
, int i
)
1263 addr
= flex_array_get(percpu
->scribble
, i
);
1264 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1267 /* return a pointer to the address conversion region of the scribble buffer */
1268 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1272 addr
= flex_array_get(percpu
->scribble
, i
);
1276 static struct dma_async_tx_descriptor
*
1277 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1279 int disks
= sh
->disks
;
1280 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1281 int target
= sh
->ops
.target
;
1282 struct r5dev
*tgt
= &sh
->dev
[target
];
1283 struct page
*xor_dest
= tgt
->page
;
1285 struct dma_async_tx_descriptor
*tx
;
1286 struct async_submit_ctl submit
;
1289 BUG_ON(sh
->batch_head
);
1291 pr_debug("%s: stripe %llu block: %d\n",
1292 __func__
, (unsigned long long)sh
->sector
, target
);
1293 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1295 for (i
= disks
; i
--; )
1297 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1299 atomic_inc(&sh
->count
);
1301 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1302 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1303 if (unlikely(count
== 1))
1304 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1306 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1311 /* set_syndrome_sources - populate source buffers for gen_syndrome
1312 * @srcs - (struct page *) array of size sh->disks
1313 * @sh - stripe_head to parse
1315 * Populates srcs in proper layout order for the stripe and returns the
1316 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1317 * destination buffer is recorded in srcs[count] and the Q destination
1318 * is recorded in srcs[count+1]].
1320 static int set_syndrome_sources(struct page
**srcs
,
1321 struct stripe_head
*sh
,
1324 int disks
= sh
->disks
;
1325 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1326 int d0_idx
= raid6_d0(sh
);
1330 for (i
= 0; i
< disks
; i
++)
1336 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1337 struct r5dev
*dev
= &sh
->dev
[i
];
1339 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1340 (srctype
== SYNDROME_SRC_ALL
) ||
1341 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1342 test_bit(R5_Wantdrain
, &dev
->flags
)) ||
1343 (srctype
== SYNDROME_SRC_WRITTEN
&&
1345 srcs
[slot
] = sh
->dev
[i
].page
;
1346 i
= raid6_next_disk(i
, disks
);
1347 } while (i
!= d0_idx
);
1349 return syndrome_disks
;
1352 static struct dma_async_tx_descriptor
*
1353 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1355 int disks
= sh
->disks
;
1356 struct page
**blocks
= to_addr_page(percpu
, 0);
1358 int qd_idx
= sh
->qd_idx
;
1359 struct dma_async_tx_descriptor
*tx
;
1360 struct async_submit_ctl submit
;
1366 BUG_ON(sh
->batch_head
);
1367 if (sh
->ops
.target
< 0)
1368 target
= sh
->ops
.target2
;
1369 else if (sh
->ops
.target2
< 0)
1370 target
= sh
->ops
.target
;
1372 /* we should only have one valid target */
1375 pr_debug("%s: stripe %llu block: %d\n",
1376 __func__
, (unsigned long long)sh
->sector
, target
);
1378 tgt
= &sh
->dev
[target
];
1379 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1382 atomic_inc(&sh
->count
);
1384 if (target
== qd_idx
) {
1385 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1386 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1387 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1388 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1389 ops_complete_compute
, sh
,
1390 to_addr_conv(sh
, percpu
, 0));
1391 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1393 /* Compute any data- or p-drive using XOR */
1395 for (i
= disks
; i
-- ; ) {
1396 if (i
== target
|| i
== qd_idx
)
1398 blocks
[count
++] = sh
->dev
[i
].page
;
1401 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1402 NULL
, ops_complete_compute
, sh
,
1403 to_addr_conv(sh
, percpu
, 0));
1404 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1410 static struct dma_async_tx_descriptor
*
1411 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1413 int i
, count
, disks
= sh
->disks
;
1414 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1415 int d0_idx
= raid6_d0(sh
);
1416 int faila
= -1, failb
= -1;
1417 int target
= sh
->ops
.target
;
1418 int target2
= sh
->ops
.target2
;
1419 struct r5dev
*tgt
= &sh
->dev
[target
];
1420 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1421 struct dma_async_tx_descriptor
*tx
;
1422 struct page
**blocks
= to_addr_page(percpu
, 0);
1423 struct async_submit_ctl submit
;
1425 BUG_ON(sh
->batch_head
);
1426 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1427 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1428 BUG_ON(target
< 0 || target2
< 0);
1429 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1430 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1432 /* we need to open-code set_syndrome_sources to handle the
1433 * slot number conversion for 'faila' and 'failb'
1435 for (i
= 0; i
< disks
; i
++)
1440 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1442 blocks
[slot
] = sh
->dev
[i
].page
;
1448 i
= raid6_next_disk(i
, disks
);
1449 } while (i
!= d0_idx
);
1451 BUG_ON(faila
== failb
);
1454 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1455 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1457 atomic_inc(&sh
->count
);
1459 if (failb
== syndrome_disks
+1) {
1460 /* Q disk is one of the missing disks */
1461 if (faila
== syndrome_disks
) {
1462 /* Missing P+Q, just recompute */
1463 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1464 ops_complete_compute
, sh
,
1465 to_addr_conv(sh
, percpu
, 0));
1466 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1467 STRIPE_SIZE
, &submit
);
1471 int qd_idx
= sh
->qd_idx
;
1473 /* Missing D+Q: recompute D from P, then recompute Q */
1474 if (target
== qd_idx
)
1475 data_target
= target2
;
1477 data_target
= target
;
1480 for (i
= disks
; i
-- ; ) {
1481 if (i
== data_target
|| i
== qd_idx
)
1483 blocks
[count
++] = sh
->dev
[i
].page
;
1485 dest
= sh
->dev
[data_target
].page
;
1486 init_async_submit(&submit
,
1487 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1489 to_addr_conv(sh
, percpu
, 0));
1490 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1493 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1494 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1495 ops_complete_compute
, sh
,
1496 to_addr_conv(sh
, percpu
, 0));
1497 return async_gen_syndrome(blocks
, 0, count
+2,
1498 STRIPE_SIZE
, &submit
);
1501 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1502 ops_complete_compute
, sh
,
1503 to_addr_conv(sh
, percpu
, 0));
1504 if (failb
== syndrome_disks
) {
1505 /* We're missing D+P. */
1506 return async_raid6_datap_recov(syndrome_disks
+2,
1510 /* We're missing D+D. */
1511 return async_raid6_2data_recov(syndrome_disks
+2,
1512 STRIPE_SIZE
, faila
, failb
,
1518 static void ops_complete_prexor(void *stripe_head_ref
)
1520 struct stripe_head
*sh
= stripe_head_ref
;
1522 pr_debug("%s: stripe %llu\n", __func__
,
1523 (unsigned long long)sh
->sector
);
1526 static struct dma_async_tx_descriptor
*
1527 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1528 struct dma_async_tx_descriptor
*tx
)
1530 int disks
= sh
->disks
;
1531 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1532 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1533 struct async_submit_ctl submit
;
1535 /* existing parity data subtracted */
1536 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1538 BUG_ON(sh
->batch_head
);
1539 pr_debug("%s: stripe %llu\n", __func__
,
1540 (unsigned long long)sh
->sector
);
1542 for (i
= disks
; i
--; ) {
1543 struct r5dev
*dev
= &sh
->dev
[i
];
1544 /* Only process blocks that are known to be uptodate */
1545 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1546 xor_srcs
[count
++] = dev
->page
;
1549 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1550 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1551 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1556 static struct dma_async_tx_descriptor
*
1557 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1558 struct dma_async_tx_descriptor
*tx
)
1560 struct page
**blocks
= to_addr_page(percpu
, 0);
1562 struct async_submit_ctl submit
;
1564 pr_debug("%s: stripe %llu\n", __func__
,
1565 (unsigned long long)sh
->sector
);
1567 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1569 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1570 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1571 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1576 static struct dma_async_tx_descriptor
*
1577 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1579 int disks
= sh
->disks
;
1581 struct stripe_head
*head_sh
= sh
;
1583 pr_debug("%s: stripe %llu\n", __func__
,
1584 (unsigned long long)sh
->sector
);
1586 for (i
= disks
; i
--; ) {
1591 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1596 spin_lock_irq(&sh
->stripe_lock
);
1597 chosen
= dev
->towrite
;
1598 dev
->towrite
= NULL
;
1599 sh
->overwrite_disks
= 0;
1600 BUG_ON(dev
->written
);
1601 wbi
= dev
->written
= chosen
;
1602 spin_unlock_irq(&sh
->stripe_lock
);
1603 WARN_ON(dev
->page
!= dev
->orig_page
);
1605 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1606 dev
->sector
+ STRIPE_SECTORS
) {
1607 if (wbi
->bi_rw
& REQ_FUA
)
1608 set_bit(R5_WantFUA
, &dev
->flags
);
1609 if (wbi
->bi_rw
& REQ_SYNC
)
1610 set_bit(R5_SyncIO
, &dev
->flags
);
1611 if (wbi
->bi_rw
& REQ_DISCARD
)
1612 set_bit(R5_Discard
, &dev
->flags
);
1614 tx
= async_copy_data(1, wbi
, &dev
->page
,
1615 dev
->sector
, tx
, sh
);
1616 if (dev
->page
!= dev
->orig_page
) {
1617 set_bit(R5_SkipCopy
, &dev
->flags
);
1618 clear_bit(R5_UPTODATE
, &dev
->flags
);
1619 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1622 wbi
= r5_next_bio(wbi
, dev
->sector
);
1625 if (head_sh
->batch_head
) {
1626 sh
= list_first_entry(&sh
->batch_list
,
1639 static void ops_complete_reconstruct(void *stripe_head_ref
)
1641 struct stripe_head
*sh
= stripe_head_ref
;
1642 int disks
= sh
->disks
;
1643 int pd_idx
= sh
->pd_idx
;
1644 int qd_idx
= sh
->qd_idx
;
1646 bool fua
= false, sync
= false, discard
= false;
1648 pr_debug("%s: stripe %llu\n", __func__
,
1649 (unsigned long long)sh
->sector
);
1651 for (i
= disks
; i
--; ) {
1652 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1653 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1654 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1657 for (i
= disks
; i
--; ) {
1658 struct r5dev
*dev
= &sh
->dev
[i
];
1660 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1661 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1662 set_bit(R5_UPTODATE
, &dev
->flags
);
1664 set_bit(R5_WantFUA
, &dev
->flags
);
1666 set_bit(R5_SyncIO
, &dev
->flags
);
1670 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1671 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1672 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1673 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1675 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1676 sh
->reconstruct_state
= reconstruct_state_result
;
1679 set_bit(STRIPE_HANDLE
, &sh
->state
);
1684 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1685 struct dma_async_tx_descriptor
*tx
)
1687 int disks
= sh
->disks
;
1688 struct page
**xor_srcs
;
1689 struct async_submit_ctl submit
;
1690 int count
, pd_idx
= sh
->pd_idx
, i
;
1691 struct page
*xor_dest
;
1693 unsigned long flags
;
1695 struct stripe_head
*head_sh
= sh
;
1698 pr_debug("%s: stripe %llu\n", __func__
,
1699 (unsigned long long)sh
->sector
);
1701 for (i
= 0; i
< sh
->disks
; i
++) {
1704 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1707 if (i
>= sh
->disks
) {
1708 atomic_inc(&sh
->count
);
1709 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1710 ops_complete_reconstruct(sh
);
1715 xor_srcs
= to_addr_page(percpu
, j
);
1716 /* check if prexor is active which means only process blocks
1717 * that are part of a read-modify-write (written)
1719 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1721 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1722 for (i
= disks
; i
--; ) {
1723 struct r5dev
*dev
= &sh
->dev
[i
];
1724 if (head_sh
->dev
[i
].written
)
1725 xor_srcs
[count
++] = dev
->page
;
1728 xor_dest
= sh
->dev
[pd_idx
].page
;
1729 for (i
= disks
; i
--; ) {
1730 struct r5dev
*dev
= &sh
->dev
[i
];
1732 xor_srcs
[count
++] = dev
->page
;
1736 /* 1/ if we prexor'd then the dest is reused as a source
1737 * 2/ if we did not prexor then we are redoing the parity
1738 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1739 * for the synchronous xor case
1741 last_stripe
= !head_sh
->batch_head
||
1742 list_first_entry(&sh
->batch_list
,
1743 struct stripe_head
, batch_list
) == head_sh
;
1745 flags
= ASYNC_TX_ACK
|
1746 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1748 atomic_inc(&head_sh
->count
);
1749 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1750 to_addr_conv(sh
, percpu
, j
));
1752 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1753 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1754 to_addr_conv(sh
, percpu
, j
));
1757 if (unlikely(count
== 1))
1758 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1760 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1763 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1770 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1771 struct dma_async_tx_descriptor
*tx
)
1773 struct async_submit_ctl submit
;
1774 struct page
**blocks
;
1775 int count
, i
, j
= 0;
1776 struct stripe_head
*head_sh
= sh
;
1779 unsigned long txflags
;
1781 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1783 for (i
= 0; i
< sh
->disks
; i
++) {
1784 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1786 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1789 if (i
>= sh
->disks
) {
1790 atomic_inc(&sh
->count
);
1791 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1792 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1793 ops_complete_reconstruct(sh
);
1798 blocks
= to_addr_page(percpu
, j
);
1800 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1801 synflags
= SYNDROME_SRC_WRITTEN
;
1802 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1804 synflags
= SYNDROME_SRC_ALL
;
1805 txflags
= ASYNC_TX_ACK
;
1808 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1809 last_stripe
= !head_sh
->batch_head
||
1810 list_first_entry(&sh
->batch_list
,
1811 struct stripe_head
, batch_list
) == head_sh
;
1814 atomic_inc(&head_sh
->count
);
1815 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1816 head_sh
, to_addr_conv(sh
, percpu
, j
));
1818 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1819 to_addr_conv(sh
, percpu
, j
));
1820 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1823 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1829 static void ops_complete_check(void *stripe_head_ref
)
1831 struct stripe_head
*sh
= stripe_head_ref
;
1833 pr_debug("%s: stripe %llu\n", __func__
,
1834 (unsigned long long)sh
->sector
);
1836 sh
->check_state
= check_state_check_result
;
1837 set_bit(STRIPE_HANDLE
, &sh
->state
);
1841 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1843 int disks
= sh
->disks
;
1844 int pd_idx
= sh
->pd_idx
;
1845 int qd_idx
= sh
->qd_idx
;
1846 struct page
*xor_dest
;
1847 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1848 struct dma_async_tx_descriptor
*tx
;
1849 struct async_submit_ctl submit
;
1853 pr_debug("%s: stripe %llu\n", __func__
,
1854 (unsigned long long)sh
->sector
);
1856 BUG_ON(sh
->batch_head
);
1858 xor_dest
= sh
->dev
[pd_idx
].page
;
1859 xor_srcs
[count
++] = xor_dest
;
1860 for (i
= disks
; i
--; ) {
1861 if (i
== pd_idx
|| i
== qd_idx
)
1863 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1866 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1867 to_addr_conv(sh
, percpu
, 0));
1868 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1869 &sh
->ops
.zero_sum_result
, &submit
);
1871 atomic_inc(&sh
->count
);
1872 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1873 tx
= async_trigger_callback(&submit
);
1876 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1878 struct page
**srcs
= to_addr_page(percpu
, 0);
1879 struct async_submit_ctl submit
;
1882 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1883 (unsigned long long)sh
->sector
, checkp
);
1885 BUG_ON(sh
->batch_head
);
1886 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
1890 atomic_inc(&sh
->count
);
1891 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1892 sh
, to_addr_conv(sh
, percpu
, 0));
1893 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1894 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1897 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1899 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1900 struct dma_async_tx_descriptor
*tx
= NULL
;
1901 struct r5conf
*conf
= sh
->raid_conf
;
1902 int level
= conf
->level
;
1903 struct raid5_percpu
*percpu
;
1907 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1908 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1909 ops_run_biofill(sh
);
1913 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1915 tx
= ops_run_compute5(sh
, percpu
);
1917 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1918 tx
= ops_run_compute6_1(sh
, percpu
);
1920 tx
= ops_run_compute6_2(sh
, percpu
);
1922 /* terminate the chain if reconstruct is not set to be run */
1923 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1927 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
1929 tx
= ops_run_prexor5(sh
, percpu
, tx
);
1931 tx
= ops_run_prexor6(sh
, percpu
, tx
);
1934 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1935 tx
= ops_run_biodrain(sh
, tx
);
1939 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1941 ops_run_reconstruct5(sh
, percpu
, tx
);
1943 ops_run_reconstruct6(sh
, percpu
, tx
);
1946 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1947 if (sh
->check_state
== check_state_run
)
1948 ops_run_check_p(sh
, percpu
);
1949 else if (sh
->check_state
== check_state_run_q
)
1950 ops_run_check_pq(sh
, percpu
, 0);
1951 else if (sh
->check_state
== check_state_run_pq
)
1952 ops_run_check_pq(sh
, percpu
, 1);
1957 if (overlap_clear
&& !sh
->batch_head
)
1958 for (i
= disks
; i
--; ) {
1959 struct r5dev
*dev
= &sh
->dev
[i
];
1960 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1961 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1966 static int grow_one_stripe(struct r5conf
*conf
, int hash
)
1968 struct stripe_head
*sh
;
1969 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1973 sh
->raid_conf
= conf
;
1975 spin_lock_init(&sh
->stripe_lock
);
1977 if (grow_buffers(sh
)) {
1979 kmem_cache_free(conf
->slab_cache
, sh
);
1982 sh
->hash_lock_index
= hash
;
1983 /* we just created an active stripe so... */
1984 atomic_set(&sh
->count
, 1);
1985 atomic_inc(&conf
->active_stripes
);
1986 INIT_LIST_HEAD(&sh
->lru
);
1988 spin_lock_init(&sh
->batch_lock
);
1989 INIT_LIST_HEAD(&sh
->batch_list
);
1990 sh
->batch_head
= NULL
;
1995 static int grow_stripes(struct r5conf
*conf
, int num
)
1997 struct kmem_cache
*sc
;
1998 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2001 if (conf
->mddev
->gendisk
)
2002 sprintf(conf
->cache_name
[0],
2003 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2005 sprintf(conf
->cache_name
[0],
2006 "raid%d-%p", conf
->level
, conf
->mddev
);
2007 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2009 conf
->active_name
= 0;
2010 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2011 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2015 conf
->slab_cache
= sc
;
2016 conf
->pool_size
= devs
;
2017 hash
= conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2019 if (!grow_one_stripe(conf
, hash
))
2021 conf
->max_nr_stripes
++;
2022 hash
= (hash
+ 1) % NR_STRIPE_HASH_LOCKS
;
2028 * scribble_len - return the required size of the scribble region
2029 * @num - total number of disks in the array
2031 * The size must be enough to contain:
2032 * 1/ a struct page pointer for each device in the array +2
2033 * 2/ room to convert each entry in (1) to its corresponding dma
2034 * (dma_map_page()) or page (page_address()) address.
2036 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2037 * calculate over all devices (not just the data blocks), using zeros in place
2038 * of the P and Q blocks.
2040 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2042 struct flex_array
*ret
;
2045 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2046 ret
= flex_array_alloc(len
, cnt
, flags
);
2049 /* always prealloc all elements, so no locking is required */
2050 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2051 flex_array_free(ret
);
2057 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2059 /* Make all the stripes able to hold 'newsize' devices.
2060 * New slots in each stripe get 'page' set to a new page.
2062 * This happens in stages:
2063 * 1/ create a new kmem_cache and allocate the required number of
2065 * 2/ gather all the old stripe_heads and transfer the pages across
2066 * to the new stripe_heads. This will have the side effect of
2067 * freezing the array as once all stripe_heads have been collected,
2068 * no IO will be possible. Old stripe heads are freed once their
2069 * pages have been transferred over, and the old kmem_cache is
2070 * freed when all stripes are done.
2071 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2072 * we simple return a failre status - no need to clean anything up.
2073 * 4/ allocate new pages for the new slots in the new stripe_heads.
2074 * If this fails, we don't bother trying the shrink the
2075 * stripe_heads down again, we just leave them as they are.
2076 * As each stripe_head is processed the new one is released into
2079 * Once step2 is started, we cannot afford to wait for a write,
2080 * so we use GFP_NOIO allocations.
2082 struct stripe_head
*osh
, *nsh
;
2083 LIST_HEAD(newstripes
);
2084 struct disk_info
*ndisks
;
2087 struct kmem_cache
*sc
;
2091 if (newsize
<= conf
->pool_size
)
2092 return 0; /* never bother to shrink */
2094 err
= md_allow_write(conf
->mddev
);
2099 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2100 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2105 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2106 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
2110 nsh
->raid_conf
= conf
;
2111 spin_lock_init(&nsh
->stripe_lock
);
2113 list_add(&nsh
->lru
, &newstripes
);
2116 /* didn't get enough, give up */
2117 while (!list_empty(&newstripes
)) {
2118 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2119 list_del(&nsh
->lru
);
2120 kmem_cache_free(sc
, nsh
);
2122 kmem_cache_destroy(sc
);
2125 /* Step 2 - Must use GFP_NOIO now.
2126 * OK, we have enough stripes, start collecting inactive
2127 * stripes and copying them over
2131 list_for_each_entry(nsh
, &newstripes
, lru
) {
2132 lock_device_hash_lock(conf
, hash
);
2133 wait_event_cmd(conf
->wait_for_stripe
,
2134 !list_empty(conf
->inactive_list
+ hash
),
2135 unlock_device_hash_lock(conf
, hash
),
2136 lock_device_hash_lock(conf
, hash
));
2137 osh
= get_free_stripe(conf
, hash
);
2138 unlock_device_hash_lock(conf
, hash
);
2139 atomic_set(&nsh
->count
, 1);
2140 for(i
=0; i
<conf
->pool_size
; i
++) {
2141 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2142 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2144 for( ; i
<newsize
; i
++)
2145 nsh
->dev
[i
].page
= NULL
;
2146 nsh
->hash_lock_index
= hash
;
2147 kmem_cache_free(conf
->slab_cache
, osh
);
2149 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2150 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2155 kmem_cache_destroy(conf
->slab_cache
);
2158 * At this point, we are holding all the stripes so the array
2159 * is completely stalled, so now is a good time to resize
2160 * conf->disks and the scribble region
2162 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2164 for (i
=0; i
<conf
->raid_disks
; i
++)
2165 ndisks
[i
] = conf
->disks
[i
];
2167 conf
->disks
= ndisks
;
2172 for_each_present_cpu(cpu
) {
2173 struct raid5_percpu
*percpu
;
2174 struct flex_array
*scribble
;
2176 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2177 scribble
= scribble_alloc(newsize
, conf
->chunk_sectors
/
2178 STRIPE_SECTORS
, GFP_NOIO
);
2181 flex_array_free(percpu
->scribble
);
2182 percpu
->scribble
= scribble
;
2190 /* Step 4, return new stripes to service */
2191 while(!list_empty(&newstripes
)) {
2192 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2193 list_del_init(&nsh
->lru
);
2195 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2196 if (nsh
->dev
[i
].page
== NULL
) {
2197 struct page
*p
= alloc_page(GFP_NOIO
);
2198 nsh
->dev
[i
].page
= p
;
2199 nsh
->dev
[i
].orig_page
= p
;
2203 release_stripe(nsh
);
2205 /* critical section pass, GFP_NOIO no longer needed */
2207 conf
->slab_cache
= sc
;
2208 conf
->active_name
= 1-conf
->active_name
;
2209 conf
->pool_size
= newsize
;
2213 static int drop_one_stripe(struct r5conf
*conf
, int hash
)
2215 struct stripe_head
*sh
;
2217 spin_lock_irq(conf
->hash_locks
+ hash
);
2218 sh
= get_free_stripe(conf
, hash
);
2219 spin_unlock_irq(conf
->hash_locks
+ hash
);
2222 BUG_ON(atomic_read(&sh
->count
));
2224 kmem_cache_free(conf
->slab_cache
, sh
);
2225 atomic_dec(&conf
->active_stripes
);
2229 static void shrink_stripes(struct r5conf
*conf
)
2232 for (hash
= 0; hash
< NR_STRIPE_HASH_LOCKS
; hash
++)
2233 while (drop_one_stripe(conf
, hash
))
2236 if (conf
->slab_cache
)
2237 kmem_cache_destroy(conf
->slab_cache
);
2238 conf
->slab_cache
= NULL
;
2241 static void raid5_end_read_request(struct bio
* bi
, int error
)
2243 struct stripe_head
*sh
= bi
->bi_private
;
2244 struct r5conf
*conf
= sh
->raid_conf
;
2245 int disks
= sh
->disks
, i
;
2246 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2247 char b
[BDEVNAME_SIZE
];
2248 struct md_rdev
*rdev
= NULL
;
2251 for (i
=0 ; i
<disks
; i
++)
2252 if (bi
== &sh
->dev
[i
].req
)
2255 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2256 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2262 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2263 /* If replacement finished while this request was outstanding,
2264 * 'replacement' might be NULL already.
2265 * In that case it moved down to 'rdev'.
2266 * rdev is not removed until all requests are finished.
2268 rdev
= conf
->disks
[i
].replacement
;
2270 rdev
= conf
->disks
[i
].rdev
;
2272 if (use_new_offset(conf
, sh
))
2273 s
= sh
->sector
+ rdev
->new_data_offset
;
2275 s
= sh
->sector
+ rdev
->data_offset
;
2277 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2278 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2279 /* Note that this cannot happen on a
2280 * replacement device. We just fail those on
2285 "md/raid:%s: read error corrected"
2286 " (%lu sectors at %llu on %s)\n",
2287 mdname(conf
->mddev
), STRIPE_SECTORS
,
2288 (unsigned long long)s
,
2289 bdevname(rdev
->bdev
, b
));
2290 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2291 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2292 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2293 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2294 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2296 if (atomic_read(&rdev
->read_errors
))
2297 atomic_set(&rdev
->read_errors
, 0);
2299 const char *bdn
= bdevname(rdev
->bdev
, b
);
2303 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2304 atomic_inc(&rdev
->read_errors
);
2305 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2308 "md/raid:%s: read error on replacement device "
2309 "(sector %llu on %s).\n",
2310 mdname(conf
->mddev
),
2311 (unsigned long long)s
,
2313 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2317 "md/raid:%s: read error not correctable "
2318 "(sector %llu on %s).\n",
2319 mdname(conf
->mddev
),
2320 (unsigned long long)s
,
2322 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2327 "md/raid:%s: read error NOT corrected!! "
2328 "(sector %llu on %s).\n",
2329 mdname(conf
->mddev
),
2330 (unsigned long long)s
,
2332 } else if (atomic_read(&rdev
->read_errors
)
2333 > conf
->max_nr_stripes
)
2335 "md/raid:%s: Too many read errors, failing device %s.\n",
2336 mdname(conf
->mddev
), bdn
);
2339 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2340 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2343 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2344 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2345 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2347 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2349 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2350 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2352 && test_bit(In_sync
, &rdev
->flags
)
2353 && rdev_set_badblocks(
2354 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2355 md_error(conf
->mddev
, rdev
);
2358 rdev_dec_pending(rdev
, conf
->mddev
);
2359 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2360 set_bit(STRIPE_HANDLE
, &sh
->state
);
2364 static void raid5_end_write_request(struct bio
*bi
, int error
)
2366 struct stripe_head
*sh
= bi
->bi_private
;
2367 struct r5conf
*conf
= sh
->raid_conf
;
2368 int disks
= sh
->disks
, i
;
2369 struct md_rdev
*uninitialized_var(rdev
);
2370 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2373 int replacement
= 0;
2375 for (i
= 0 ; i
< disks
; i
++) {
2376 if (bi
== &sh
->dev
[i
].req
) {
2377 rdev
= conf
->disks
[i
].rdev
;
2380 if (bi
== &sh
->dev
[i
].rreq
) {
2381 rdev
= conf
->disks
[i
].replacement
;
2385 /* rdev was removed and 'replacement'
2386 * replaced it. rdev is not removed
2387 * until all requests are finished.
2389 rdev
= conf
->disks
[i
].rdev
;
2393 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2394 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2403 md_error(conf
->mddev
, rdev
);
2404 else if (is_badblock(rdev
, sh
->sector
,
2406 &first_bad
, &bad_sectors
))
2407 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2410 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2411 set_bit(WriteErrorSeen
, &rdev
->flags
);
2412 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2413 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2414 set_bit(MD_RECOVERY_NEEDED
,
2415 &rdev
->mddev
->recovery
);
2416 } else if (is_badblock(rdev
, sh
->sector
,
2418 &first_bad
, &bad_sectors
)) {
2419 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2420 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2421 /* That was a successful write so make
2422 * sure it looks like we already did
2425 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2428 rdev_dec_pending(rdev
, conf
->mddev
);
2430 if (sh
->batch_head
&& !uptodate
)
2431 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2433 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2434 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2435 set_bit(STRIPE_HANDLE
, &sh
->state
);
2438 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2439 release_stripe(sh
->batch_head
);
2442 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
2444 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
2446 struct r5dev
*dev
= &sh
->dev
[i
];
2448 bio_init(&dev
->req
);
2449 dev
->req
.bi_io_vec
= &dev
->vec
;
2450 dev
->req
.bi_max_vecs
= 1;
2451 dev
->req
.bi_private
= sh
;
2453 bio_init(&dev
->rreq
);
2454 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
2455 dev
->rreq
.bi_max_vecs
= 1;
2456 dev
->rreq
.bi_private
= sh
;
2459 dev
->sector
= compute_blocknr(sh
, i
, previous
);
2462 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2464 char b
[BDEVNAME_SIZE
];
2465 struct r5conf
*conf
= mddev
->private;
2466 unsigned long flags
;
2467 pr_debug("raid456: error called\n");
2469 spin_lock_irqsave(&conf
->device_lock
, flags
);
2470 clear_bit(In_sync
, &rdev
->flags
);
2471 mddev
->degraded
= calc_degraded(conf
);
2472 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2473 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2475 set_bit(Blocked
, &rdev
->flags
);
2476 set_bit(Faulty
, &rdev
->flags
);
2477 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2479 "md/raid:%s: Disk failure on %s, disabling device.\n"
2480 "md/raid:%s: Operation continuing on %d devices.\n",
2482 bdevname(rdev
->bdev
, b
),
2484 conf
->raid_disks
- mddev
->degraded
);
2488 * Input: a 'big' sector number,
2489 * Output: index of the data and parity disk, and the sector # in them.
2491 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2492 int previous
, int *dd_idx
,
2493 struct stripe_head
*sh
)
2495 sector_t stripe
, stripe2
;
2496 sector_t chunk_number
;
2497 unsigned int chunk_offset
;
2500 sector_t new_sector
;
2501 int algorithm
= previous
? conf
->prev_algo
2503 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2504 : conf
->chunk_sectors
;
2505 int raid_disks
= previous
? conf
->previous_raid_disks
2507 int data_disks
= raid_disks
- conf
->max_degraded
;
2509 /* First compute the information on this sector */
2512 * Compute the chunk number and the sector offset inside the chunk
2514 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2515 chunk_number
= r_sector
;
2518 * Compute the stripe number
2520 stripe
= chunk_number
;
2521 *dd_idx
= sector_div(stripe
, data_disks
);
2524 * Select the parity disk based on the user selected algorithm.
2526 pd_idx
= qd_idx
= -1;
2527 switch(conf
->level
) {
2529 pd_idx
= data_disks
;
2532 switch (algorithm
) {
2533 case ALGORITHM_LEFT_ASYMMETRIC
:
2534 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2535 if (*dd_idx
>= pd_idx
)
2538 case ALGORITHM_RIGHT_ASYMMETRIC
:
2539 pd_idx
= sector_div(stripe2
, raid_disks
);
2540 if (*dd_idx
>= pd_idx
)
2543 case ALGORITHM_LEFT_SYMMETRIC
:
2544 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2545 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2547 case ALGORITHM_RIGHT_SYMMETRIC
:
2548 pd_idx
= sector_div(stripe2
, raid_disks
);
2549 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2551 case ALGORITHM_PARITY_0
:
2555 case ALGORITHM_PARITY_N
:
2556 pd_idx
= data_disks
;
2564 switch (algorithm
) {
2565 case ALGORITHM_LEFT_ASYMMETRIC
:
2566 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2567 qd_idx
= pd_idx
+ 1;
2568 if (pd_idx
== raid_disks
-1) {
2569 (*dd_idx
)++; /* Q D D D P */
2571 } else if (*dd_idx
>= pd_idx
)
2572 (*dd_idx
) += 2; /* D D P Q D */
2574 case ALGORITHM_RIGHT_ASYMMETRIC
:
2575 pd_idx
= sector_div(stripe2
, raid_disks
);
2576 qd_idx
= pd_idx
+ 1;
2577 if (pd_idx
== raid_disks
-1) {
2578 (*dd_idx
)++; /* Q D D D P */
2580 } else if (*dd_idx
>= pd_idx
)
2581 (*dd_idx
) += 2; /* D D P Q D */
2583 case ALGORITHM_LEFT_SYMMETRIC
:
2584 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2585 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2586 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2588 case ALGORITHM_RIGHT_SYMMETRIC
:
2589 pd_idx
= sector_div(stripe2
, raid_disks
);
2590 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2591 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2594 case ALGORITHM_PARITY_0
:
2599 case ALGORITHM_PARITY_N
:
2600 pd_idx
= data_disks
;
2601 qd_idx
= data_disks
+ 1;
2604 case ALGORITHM_ROTATING_ZERO_RESTART
:
2605 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2606 * of blocks for computing Q is different.
2608 pd_idx
= sector_div(stripe2
, raid_disks
);
2609 qd_idx
= pd_idx
+ 1;
2610 if (pd_idx
== raid_disks
-1) {
2611 (*dd_idx
)++; /* Q D D D P */
2613 } else if (*dd_idx
>= pd_idx
)
2614 (*dd_idx
) += 2; /* D D P Q D */
2618 case ALGORITHM_ROTATING_N_RESTART
:
2619 /* Same a left_asymmetric, by first stripe is
2620 * D D D P Q rather than
2624 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2625 qd_idx
= pd_idx
+ 1;
2626 if (pd_idx
== raid_disks
-1) {
2627 (*dd_idx
)++; /* Q D D D P */
2629 } else if (*dd_idx
>= pd_idx
)
2630 (*dd_idx
) += 2; /* D D P Q D */
2634 case ALGORITHM_ROTATING_N_CONTINUE
:
2635 /* Same as left_symmetric but Q is before P */
2636 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2637 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2638 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2642 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2643 /* RAID5 left_asymmetric, with Q on last device */
2644 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2645 if (*dd_idx
>= pd_idx
)
2647 qd_idx
= raid_disks
- 1;
2650 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2651 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2652 if (*dd_idx
>= pd_idx
)
2654 qd_idx
= raid_disks
- 1;
2657 case ALGORITHM_LEFT_SYMMETRIC_6
:
2658 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2659 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2660 qd_idx
= raid_disks
- 1;
2663 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2664 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2665 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2666 qd_idx
= raid_disks
- 1;
2669 case ALGORITHM_PARITY_0_6
:
2672 qd_idx
= raid_disks
- 1;
2682 sh
->pd_idx
= pd_idx
;
2683 sh
->qd_idx
= qd_idx
;
2684 sh
->ddf_layout
= ddf_layout
;
2687 * Finally, compute the new sector number
2689 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2693 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2695 struct r5conf
*conf
= sh
->raid_conf
;
2696 int raid_disks
= sh
->disks
;
2697 int data_disks
= raid_disks
- conf
->max_degraded
;
2698 sector_t new_sector
= sh
->sector
, check
;
2699 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2700 : conf
->chunk_sectors
;
2701 int algorithm
= previous
? conf
->prev_algo
2705 sector_t chunk_number
;
2706 int dummy1
, dd_idx
= i
;
2708 struct stripe_head sh2
;
2710 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2711 stripe
= new_sector
;
2713 if (i
== sh
->pd_idx
)
2715 switch(conf
->level
) {
2718 switch (algorithm
) {
2719 case ALGORITHM_LEFT_ASYMMETRIC
:
2720 case ALGORITHM_RIGHT_ASYMMETRIC
:
2724 case ALGORITHM_LEFT_SYMMETRIC
:
2725 case ALGORITHM_RIGHT_SYMMETRIC
:
2728 i
-= (sh
->pd_idx
+ 1);
2730 case ALGORITHM_PARITY_0
:
2733 case ALGORITHM_PARITY_N
:
2740 if (i
== sh
->qd_idx
)
2741 return 0; /* It is the Q disk */
2742 switch (algorithm
) {
2743 case ALGORITHM_LEFT_ASYMMETRIC
:
2744 case ALGORITHM_RIGHT_ASYMMETRIC
:
2745 case ALGORITHM_ROTATING_ZERO_RESTART
:
2746 case ALGORITHM_ROTATING_N_RESTART
:
2747 if (sh
->pd_idx
== raid_disks
-1)
2748 i
--; /* Q D D D P */
2749 else if (i
> sh
->pd_idx
)
2750 i
-= 2; /* D D P Q D */
2752 case ALGORITHM_LEFT_SYMMETRIC
:
2753 case ALGORITHM_RIGHT_SYMMETRIC
:
2754 if (sh
->pd_idx
== raid_disks
-1)
2755 i
--; /* Q D D D P */
2760 i
-= (sh
->pd_idx
+ 2);
2763 case ALGORITHM_PARITY_0
:
2766 case ALGORITHM_PARITY_N
:
2768 case ALGORITHM_ROTATING_N_CONTINUE
:
2769 /* Like left_symmetric, but P is before Q */
2770 if (sh
->pd_idx
== 0)
2771 i
--; /* P D D D Q */
2776 i
-= (sh
->pd_idx
+ 1);
2779 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2780 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2784 case ALGORITHM_LEFT_SYMMETRIC_6
:
2785 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2787 i
+= data_disks
+ 1;
2788 i
-= (sh
->pd_idx
+ 1);
2790 case ALGORITHM_PARITY_0_6
:
2799 chunk_number
= stripe
* data_disks
+ i
;
2800 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2802 check
= raid5_compute_sector(conf
, r_sector
,
2803 previous
, &dummy1
, &sh2
);
2804 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2805 || sh2
.qd_idx
!= sh
->qd_idx
) {
2806 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2807 mdname(conf
->mddev
));
2814 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2815 int rcw
, int expand
)
2817 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
2818 struct r5conf
*conf
= sh
->raid_conf
;
2819 int level
= conf
->level
;
2823 for (i
= disks
; i
--; ) {
2824 struct r5dev
*dev
= &sh
->dev
[i
];
2827 set_bit(R5_LOCKED
, &dev
->flags
);
2828 set_bit(R5_Wantdrain
, &dev
->flags
);
2830 clear_bit(R5_UPTODATE
, &dev
->flags
);
2834 /* if we are not expanding this is a proper write request, and
2835 * there will be bios with new data to be drained into the
2840 /* False alarm, nothing to do */
2842 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2843 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2845 sh
->reconstruct_state
= reconstruct_state_run
;
2847 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2849 if (s
->locked
+ conf
->max_degraded
== disks
)
2850 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2851 atomic_inc(&conf
->pending_full_writes
);
2853 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2854 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2855 BUG_ON(level
== 6 &&
2856 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
2857 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
2859 for (i
= disks
; i
--; ) {
2860 struct r5dev
*dev
= &sh
->dev
[i
];
2861 if (i
== pd_idx
|| i
== qd_idx
)
2865 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2866 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2867 set_bit(R5_Wantdrain
, &dev
->flags
);
2868 set_bit(R5_LOCKED
, &dev
->flags
);
2869 clear_bit(R5_UPTODATE
, &dev
->flags
);
2874 /* False alarm - nothing to do */
2876 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2877 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2878 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2879 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2882 /* keep the parity disk(s) locked while asynchronous operations
2885 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2886 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2890 int qd_idx
= sh
->qd_idx
;
2891 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2893 set_bit(R5_LOCKED
, &dev
->flags
);
2894 clear_bit(R5_UPTODATE
, &dev
->flags
);
2898 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2899 __func__
, (unsigned long long)sh
->sector
,
2900 s
->locked
, s
->ops_request
);
2904 * Each stripe/dev can have one or more bion attached.
2905 * toread/towrite point to the first in a chain.
2906 * The bi_next chain must be in order.
2908 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
2909 int forwrite
, int previous
)
2912 struct r5conf
*conf
= sh
->raid_conf
;
2915 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2916 (unsigned long long)bi
->bi_iter
.bi_sector
,
2917 (unsigned long long)sh
->sector
);
2920 * If several bio share a stripe. The bio bi_phys_segments acts as a
2921 * reference count to avoid race. The reference count should already be
2922 * increased before this function is called (for example, in
2923 * make_request()), so other bio sharing this stripe will not free the
2924 * stripe. If a stripe is owned by one stripe, the stripe lock will
2927 spin_lock_irq(&sh
->stripe_lock
);
2928 /* Don't allow new IO added to stripes in batch list */
2932 bip
= &sh
->dev
[dd_idx
].towrite
;
2936 bip
= &sh
->dev
[dd_idx
].toread
;
2937 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
2938 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
2940 bip
= & (*bip
)->bi_next
;
2942 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
2945 if (!forwrite
|| previous
)
2946 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
2948 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2952 raid5_inc_bi_active_stripes(bi
);
2955 /* check if page is covered */
2956 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2957 for (bi
=sh
->dev
[dd_idx
].towrite
;
2958 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2959 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
2960 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2961 if (bio_end_sector(bi
) >= sector
)
2962 sector
= bio_end_sector(bi
);
2964 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2965 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
2966 sh
->overwrite_disks
++;
2969 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2970 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
2971 (unsigned long long)sh
->sector
, dd_idx
);
2972 spin_unlock_irq(&sh
->stripe_lock
);
2974 if (conf
->mddev
->bitmap
&& firstwrite
) {
2975 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2977 sh
->bm_seq
= conf
->seq_flush
+1;
2978 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2981 if (stripe_can_batch(sh
))
2982 stripe_add_to_batch_list(conf
, sh
);
2986 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2987 spin_unlock_irq(&sh
->stripe_lock
);
2991 static void end_reshape(struct r5conf
*conf
);
2993 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2994 struct stripe_head
*sh
)
2996 int sectors_per_chunk
=
2997 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2999 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3000 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3002 raid5_compute_sector(conf
,
3003 stripe
* (disks
- conf
->max_degraded
)
3004 *sectors_per_chunk
+ chunk_offset
,
3010 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3011 struct stripe_head_state
*s
, int disks
,
3012 struct bio
**return_bi
)
3015 BUG_ON(sh
->batch_head
);
3016 for (i
= disks
; i
--; ) {
3020 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3021 struct md_rdev
*rdev
;
3023 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3024 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
3025 atomic_inc(&rdev
->nr_pending
);
3030 if (!rdev_set_badblocks(
3034 md_error(conf
->mddev
, rdev
);
3035 rdev_dec_pending(rdev
, conf
->mddev
);
3038 spin_lock_irq(&sh
->stripe_lock
);
3039 /* fail all writes first */
3040 bi
= sh
->dev
[i
].towrite
;
3041 sh
->dev
[i
].towrite
= NULL
;
3042 sh
->overwrite_disks
= 0;
3043 spin_unlock_irq(&sh
->stripe_lock
);
3047 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3048 wake_up(&conf
->wait_for_overlap
);
3050 while (bi
&& bi
->bi_iter
.bi_sector
<
3051 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3052 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3053 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3054 if (!raid5_dec_bi_active_stripes(bi
)) {
3055 md_write_end(conf
->mddev
);
3056 bi
->bi_next
= *return_bi
;
3062 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3063 STRIPE_SECTORS
, 0, 0);
3065 /* and fail all 'written' */
3066 bi
= sh
->dev
[i
].written
;
3067 sh
->dev
[i
].written
= NULL
;
3068 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3069 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3070 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3073 if (bi
) bitmap_end
= 1;
3074 while (bi
&& bi
->bi_iter
.bi_sector
<
3075 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3076 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3077 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3078 if (!raid5_dec_bi_active_stripes(bi
)) {
3079 md_write_end(conf
->mddev
);
3080 bi
->bi_next
= *return_bi
;
3086 /* fail any reads if this device is non-operational and
3087 * the data has not reached the cache yet.
3089 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3090 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3091 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3092 spin_lock_irq(&sh
->stripe_lock
);
3093 bi
= sh
->dev
[i
].toread
;
3094 sh
->dev
[i
].toread
= NULL
;
3095 spin_unlock_irq(&sh
->stripe_lock
);
3096 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3097 wake_up(&conf
->wait_for_overlap
);
3098 while (bi
&& bi
->bi_iter
.bi_sector
<
3099 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3100 struct bio
*nextbi
=
3101 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3102 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3103 if (!raid5_dec_bi_active_stripes(bi
)) {
3104 bi
->bi_next
= *return_bi
;
3111 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3112 STRIPE_SECTORS
, 0, 0);
3113 /* If we were in the middle of a write the parity block might
3114 * still be locked - so just clear all R5_LOCKED flags
3116 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3119 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3120 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3121 md_wakeup_thread(conf
->mddev
->thread
);
3125 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3126 struct stripe_head_state
*s
)
3131 BUG_ON(sh
->batch_head
);
3132 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3133 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3134 wake_up(&conf
->wait_for_overlap
);
3137 /* There is nothing more to do for sync/check/repair.
3138 * Don't even need to abort as that is handled elsewhere
3139 * if needed, and not always wanted e.g. if there is a known
3141 * For recover/replace we need to record a bad block on all
3142 * non-sync devices, or abort the recovery
3144 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3145 /* During recovery devices cannot be removed, so
3146 * locking and refcounting of rdevs is not needed
3148 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3149 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
3151 && !test_bit(Faulty
, &rdev
->flags
)
3152 && !test_bit(In_sync
, &rdev
->flags
)
3153 && !rdev_set_badblocks(rdev
, sh
->sector
,
3156 rdev
= conf
->disks
[i
].replacement
;
3158 && !test_bit(Faulty
, &rdev
->flags
)
3159 && !test_bit(In_sync
, &rdev
->flags
)
3160 && !rdev_set_badblocks(rdev
, sh
->sector
,
3165 conf
->recovery_disabled
=
3166 conf
->mddev
->recovery_disabled
;
3168 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3171 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3173 struct md_rdev
*rdev
;
3175 /* Doing recovery so rcu locking not required */
3176 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
3178 && !test_bit(Faulty
, &rdev
->flags
)
3179 && !test_bit(In_sync
, &rdev
->flags
)
3180 && (rdev
->recovery_offset
<= sh
->sector
3181 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3187 /* fetch_block - checks the given member device to see if its data needs
3188 * to be read or computed to satisfy a request.
3190 * Returns 1 when no more member devices need to be checked, otherwise returns
3191 * 0 to tell the loop in handle_stripe_fill to continue
3194 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3195 int disk_idx
, int disks
)
3197 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3198 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3199 &sh
->dev
[s
->failed_num
[1]] };
3203 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3204 test_bit(R5_UPTODATE
, &dev
->flags
))
3205 /* No point reading this as we already have it or have
3206 * decided to get it.
3211 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3212 /* We need this block to directly satisfy a request */
3215 if (s
->syncing
|| s
->expanding
||
3216 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3217 /* When syncing, or expanding we read everything.
3218 * When replacing, we need the replaced block.
3222 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3223 (s
->failed
>= 2 && fdev
[1]->toread
))
3224 /* If we want to read from a failed device, then
3225 * we need to actually read every other device.
3229 /* Sometimes neither read-modify-write nor reconstruct-write
3230 * cycles can work. In those cases we read every block we
3231 * can. Then the parity-update is certain to have enough to
3233 * This can only be a problem when we need to write something,
3234 * and some device has failed. If either of those tests
3235 * fail we need look no further.
3237 if (!s
->failed
|| !s
->to_write
)
3240 if (test_bit(R5_Insync
, &dev
->flags
) &&
3241 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3242 /* Pre-reads at not permitted until after short delay
3243 * to gather multiple requests. However if this
3244 * device is no Insync, the block could only be be computed
3245 * and there is no need to delay that.
3249 for (i
= 0; i
< s
->failed
; i
++) {
3250 if (fdev
[i
]->towrite
&&
3251 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3252 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3253 /* If we have a partial write to a failed
3254 * device, then we will need to reconstruct
3255 * the content of that device, so all other
3256 * devices must be read.
3261 /* If we are forced to do a reconstruct-write, either because
3262 * the current RAID6 implementation only supports that, or
3263 * or because parity cannot be trusted and we are currently
3264 * recovering it, there is extra need to be careful.
3265 * If one of the devices that we would need to read, because
3266 * it is not being overwritten (and maybe not written at all)
3267 * is missing/faulty, then we need to read everything we can.
3269 if (sh
->raid_conf
->level
!= 6 &&
3270 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3271 /* reconstruct-write isn't being forced */
3273 for (i
= 0; i
< s
->failed
; i
++) {
3274 if (!test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3275 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3282 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3283 int disk_idx
, int disks
)
3285 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3287 /* is the data in this block needed, and can we get it? */
3288 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3289 /* we would like to get this block, possibly by computing it,
3290 * otherwise read it if the backing disk is insync
3292 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3293 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3294 if ((s
->uptodate
== disks
- 1) &&
3295 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3296 disk_idx
== s
->failed_num
[1]))) {
3297 /* have disk failed, and we're requested to fetch it;
3300 pr_debug("Computing stripe %llu block %d\n",
3301 (unsigned long long)sh
->sector
, disk_idx
);
3302 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3303 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3304 set_bit(R5_Wantcompute
, &dev
->flags
);
3305 sh
->ops
.target
= disk_idx
;
3306 sh
->ops
.target2
= -1; /* no 2nd target */
3308 /* Careful: from this point on 'uptodate' is in the eye
3309 * of raid_run_ops which services 'compute' operations
3310 * before writes. R5_Wantcompute flags a block that will
3311 * be R5_UPTODATE by the time it is needed for a
3312 * subsequent operation.
3316 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3317 /* Computing 2-failure is *very* expensive; only
3318 * do it if failed >= 2
3321 for (other
= disks
; other
--; ) {
3322 if (other
== disk_idx
)
3324 if (!test_bit(R5_UPTODATE
,
3325 &sh
->dev
[other
].flags
))
3329 pr_debug("Computing stripe %llu blocks %d,%d\n",
3330 (unsigned long long)sh
->sector
,
3332 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3333 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3334 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3335 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3336 sh
->ops
.target
= disk_idx
;
3337 sh
->ops
.target2
= other
;
3341 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3342 set_bit(R5_LOCKED
, &dev
->flags
);
3343 set_bit(R5_Wantread
, &dev
->flags
);
3345 pr_debug("Reading block %d (sync=%d)\n",
3346 disk_idx
, s
->syncing
);
3354 * handle_stripe_fill - read or compute data to satisfy pending requests.
3356 static void handle_stripe_fill(struct stripe_head
*sh
,
3357 struct stripe_head_state
*s
,
3362 BUG_ON(sh
->batch_head
);
3363 /* look for blocks to read/compute, skip this if a compute
3364 * is already in flight, or if the stripe contents are in the
3365 * midst of changing due to a write
3367 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3368 !sh
->reconstruct_state
)
3369 for (i
= disks
; i
--; )
3370 if (fetch_block(sh
, s
, i
, disks
))
3372 set_bit(STRIPE_HANDLE
, &sh
->state
);
3375 /* handle_stripe_clean_event
3376 * any written block on an uptodate or failed drive can be returned.
3377 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3378 * never LOCKED, so we don't need to test 'failed' directly.
3380 static void handle_stripe_clean_event(struct r5conf
*conf
,
3381 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
3385 int discard_pending
= 0;
3386 struct stripe_head
*head_sh
= sh
;
3387 bool do_endio
= false;
3390 for (i
= disks
; i
--; )
3391 if (sh
->dev
[i
].written
) {
3393 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3394 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3395 test_bit(R5_Discard
, &dev
->flags
) ||
3396 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3397 /* We can return any write requests */
3398 struct bio
*wbi
, *wbi2
;
3399 pr_debug("Return write for disc %d\n", i
);
3400 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3401 clear_bit(R5_UPTODATE
, &dev
->flags
);
3402 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3403 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3408 dev
->page
= dev
->orig_page
;
3410 dev
->written
= NULL
;
3411 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3412 dev
->sector
+ STRIPE_SECTORS
) {
3413 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3414 if (!raid5_dec_bi_active_stripes(wbi
)) {
3415 md_write_end(conf
->mddev
);
3416 wbi
->bi_next
= *return_bi
;
3421 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3423 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3425 if (head_sh
->batch_head
) {
3426 sh
= list_first_entry(&sh
->batch_list
,
3429 if (sh
!= head_sh
) {
3436 } else if (test_bit(R5_Discard
, &dev
->flags
))
3437 discard_pending
= 1;
3438 WARN_ON(test_bit(R5_SkipCopy
, &dev
->flags
));
3439 WARN_ON(dev
->page
!= dev
->orig_page
);
3441 if (!discard_pending
&&
3442 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3443 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3444 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3445 if (sh
->qd_idx
>= 0) {
3446 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3447 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3449 /* now that discard is done we can proceed with any sync */
3450 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3452 * SCSI discard will change some bio fields and the stripe has
3453 * no updated data, so remove it from hash list and the stripe
3454 * will be reinitialized
3456 spin_lock_irq(&conf
->device_lock
);
3459 if (head_sh
->batch_head
) {
3460 sh
= list_first_entry(&sh
->batch_list
,
3461 struct stripe_head
, batch_list
);
3465 spin_unlock_irq(&conf
->device_lock
);
3468 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3469 set_bit(STRIPE_HANDLE
, &sh
->state
);
3473 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3474 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3475 md_wakeup_thread(conf
->mddev
->thread
);
3477 if (!head_sh
->batch_head
|| !do_endio
)
3479 for (i
= 0; i
< head_sh
->disks
; i
++) {
3480 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
3483 while (!list_empty(&head_sh
->batch_list
)) {
3485 sh
= list_first_entry(&head_sh
->batch_list
,
3486 struct stripe_head
, batch_list
);
3487 list_del_init(&sh
->batch_list
);
3489 set_mask_bits(&sh
->state
, ~STRIPE_EXPAND_SYNC_FLAG
,
3490 head_sh
->state
& ~((1 << STRIPE_ACTIVE
) |
3491 (1 << STRIPE_PREREAD_ACTIVE
) |
3492 STRIPE_EXPAND_SYNC_FLAG
));
3493 sh
->check_state
= head_sh
->check_state
;
3494 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
3495 for (i
= 0; i
< sh
->disks
; i
++) {
3496 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3498 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
;
3501 spin_lock_irq(&sh
->stripe_lock
);
3502 sh
->batch_head
= NULL
;
3503 spin_unlock_irq(&sh
->stripe_lock
);
3504 if (sh
->state
& STRIPE_EXPAND_SYNC_FLAG
)
3505 set_bit(STRIPE_HANDLE
, &sh
->state
);
3509 spin_lock_irq(&head_sh
->stripe_lock
);
3510 head_sh
->batch_head
= NULL
;
3511 spin_unlock_irq(&head_sh
->stripe_lock
);
3512 wake_up_nr(&conf
->wait_for_overlap
, wakeup_nr
);
3513 if (head_sh
->state
& STRIPE_EXPAND_SYNC_FLAG
)
3514 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
3517 static void handle_stripe_dirtying(struct r5conf
*conf
,
3518 struct stripe_head
*sh
,
3519 struct stripe_head_state
*s
,
3522 int rmw
= 0, rcw
= 0, i
;
3523 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3525 /* Check whether resync is now happening or should start.
3526 * If yes, then the array is dirty (after unclean shutdown or
3527 * initial creation), so parity in some stripes might be inconsistent.
3528 * In this case, we need to always do reconstruct-write, to ensure
3529 * that in case of drive failure or read-error correction, we
3530 * generate correct data from the parity.
3532 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3533 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3535 /* Calculate the real rcw later - for now make it
3536 * look like rcw is cheaper
3539 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3540 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3541 (unsigned long long)sh
->sector
);
3542 } else for (i
= disks
; i
--; ) {
3543 /* would I have to read this buffer for read_modify_write */
3544 struct r5dev
*dev
= &sh
->dev
[i
];
3545 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3546 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3547 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3548 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3549 if (test_bit(R5_Insync
, &dev
->flags
))
3552 rmw
+= 2*disks
; /* cannot read it */
3554 /* Would I have to read this buffer for reconstruct_write */
3555 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3556 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3557 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3558 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3559 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3560 if (test_bit(R5_Insync
, &dev
->flags
))
3566 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3567 (unsigned long long)sh
->sector
, rmw
, rcw
);
3568 set_bit(STRIPE_HANDLE
, &sh
->state
);
3569 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_ENABLE_RMW
)) && rmw
> 0) {
3570 /* prefer read-modify-write, but need to get some data */
3571 if (conf
->mddev
->queue
)
3572 blk_add_trace_msg(conf
->mddev
->queue
,
3573 "raid5 rmw %llu %d",
3574 (unsigned long long)sh
->sector
, rmw
);
3575 for (i
= disks
; i
--; ) {
3576 struct r5dev
*dev
= &sh
->dev
[i
];
3577 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3578 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3579 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3580 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3581 test_bit(R5_Insync
, &dev
->flags
)) {
3582 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3584 pr_debug("Read_old block %d for r-m-w\n",
3586 set_bit(R5_LOCKED
, &dev
->flags
);
3587 set_bit(R5_Wantread
, &dev
->flags
);
3590 set_bit(STRIPE_DELAYED
, &sh
->state
);
3591 set_bit(STRIPE_HANDLE
, &sh
->state
);
3596 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_ENABLE_RMW
)) && rcw
> 0) {
3597 /* want reconstruct write, but need to get some data */
3600 for (i
= disks
; i
--; ) {
3601 struct r5dev
*dev
= &sh
->dev
[i
];
3602 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3603 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3604 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3605 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3606 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3608 if (test_bit(R5_Insync
, &dev
->flags
) &&
3609 test_bit(STRIPE_PREREAD_ACTIVE
,
3611 pr_debug("Read_old block "
3612 "%d for Reconstruct\n", i
);
3613 set_bit(R5_LOCKED
, &dev
->flags
);
3614 set_bit(R5_Wantread
, &dev
->flags
);
3618 set_bit(STRIPE_DELAYED
, &sh
->state
);
3619 set_bit(STRIPE_HANDLE
, &sh
->state
);
3623 if (rcw
&& conf
->mddev
->queue
)
3624 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3625 (unsigned long long)sh
->sector
,
3626 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3629 if (rcw
> disks
&& rmw
> disks
&&
3630 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3631 set_bit(STRIPE_DELAYED
, &sh
->state
);
3633 /* now if nothing is locked, and if we have enough data,
3634 * we can start a write request
3636 /* since handle_stripe can be called at any time we need to handle the
3637 * case where a compute block operation has been submitted and then a
3638 * subsequent call wants to start a write request. raid_run_ops only
3639 * handles the case where compute block and reconstruct are requested
3640 * simultaneously. If this is not the case then new writes need to be
3641 * held off until the compute completes.
3643 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
3644 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
3645 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
3646 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
3649 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
3650 struct stripe_head_state
*s
, int disks
)
3652 struct r5dev
*dev
= NULL
;
3654 BUG_ON(sh
->batch_head
);
3655 set_bit(STRIPE_HANDLE
, &sh
->state
);
3657 switch (sh
->check_state
) {
3658 case check_state_idle
:
3659 /* start a new check operation if there are no failures */
3660 if (s
->failed
== 0) {
3661 BUG_ON(s
->uptodate
!= disks
);
3662 sh
->check_state
= check_state_run
;
3663 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3664 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3668 dev
= &sh
->dev
[s
->failed_num
[0]];
3670 case check_state_compute_result
:
3671 sh
->check_state
= check_state_idle
;
3673 dev
= &sh
->dev
[sh
->pd_idx
];
3675 /* check that a write has not made the stripe insync */
3676 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3679 /* either failed parity check, or recovery is happening */
3680 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3681 BUG_ON(s
->uptodate
!= disks
);
3683 set_bit(R5_LOCKED
, &dev
->flags
);
3685 set_bit(R5_Wantwrite
, &dev
->flags
);
3687 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3688 set_bit(STRIPE_INSYNC
, &sh
->state
);
3690 case check_state_run
:
3691 break; /* we will be called again upon completion */
3692 case check_state_check_result
:
3693 sh
->check_state
= check_state_idle
;
3695 /* if a failure occurred during the check operation, leave
3696 * STRIPE_INSYNC not set and let the stripe be handled again
3701 /* handle a successful check operation, if parity is correct
3702 * we are done. Otherwise update the mismatch count and repair
3703 * parity if !MD_RECOVERY_CHECK
3705 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3706 /* parity is correct (on disc,
3707 * not in buffer any more)
3709 set_bit(STRIPE_INSYNC
, &sh
->state
);
3711 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3712 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3713 /* don't try to repair!! */
3714 set_bit(STRIPE_INSYNC
, &sh
->state
);
3716 sh
->check_state
= check_state_compute_run
;
3717 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3718 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3719 set_bit(R5_Wantcompute
,
3720 &sh
->dev
[sh
->pd_idx
].flags
);
3721 sh
->ops
.target
= sh
->pd_idx
;
3722 sh
->ops
.target2
= -1;
3727 case check_state_compute_run
:
3730 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3731 __func__
, sh
->check_state
,
3732 (unsigned long long) sh
->sector
);
3737 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3738 struct stripe_head_state
*s
,
3741 int pd_idx
= sh
->pd_idx
;
3742 int qd_idx
= sh
->qd_idx
;
3745 BUG_ON(sh
->batch_head
);
3746 set_bit(STRIPE_HANDLE
, &sh
->state
);
3748 BUG_ON(s
->failed
> 2);
3750 /* Want to check and possibly repair P and Q.
3751 * However there could be one 'failed' device, in which
3752 * case we can only check one of them, possibly using the
3753 * other to generate missing data
3756 switch (sh
->check_state
) {
3757 case check_state_idle
:
3758 /* start a new check operation if there are < 2 failures */
3759 if (s
->failed
== s
->q_failed
) {
3760 /* The only possible failed device holds Q, so it
3761 * makes sense to check P (If anything else were failed,
3762 * we would have used P to recreate it).
3764 sh
->check_state
= check_state_run
;
3766 if (!s
->q_failed
&& s
->failed
< 2) {
3767 /* Q is not failed, and we didn't use it to generate
3768 * anything, so it makes sense to check it
3770 if (sh
->check_state
== check_state_run
)
3771 sh
->check_state
= check_state_run_pq
;
3773 sh
->check_state
= check_state_run_q
;
3776 /* discard potentially stale zero_sum_result */
3777 sh
->ops
.zero_sum_result
= 0;
3779 if (sh
->check_state
== check_state_run
) {
3780 /* async_xor_zero_sum destroys the contents of P */
3781 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3784 if (sh
->check_state
>= check_state_run
&&
3785 sh
->check_state
<= check_state_run_pq
) {
3786 /* async_syndrome_zero_sum preserves P and Q, so
3787 * no need to mark them !uptodate here
3789 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3793 /* we have 2-disk failure */
3794 BUG_ON(s
->failed
!= 2);
3796 case check_state_compute_result
:
3797 sh
->check_state
= check_state_idle
;
3799 /* check that a write has not made the stripe insync */
3800 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3803 /* now write out any block on a failed drive,
3804 * or P or Q if they were recomputed
3806 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3807 if (s
->failed
== 2) {
3808 dev
= &sh
->dev
[s
->failed_num
[1]];
3810 set_bit(R5_LOCKED
, &dev
->flags
);
3811 set_bit(R5_Wantwrite
, &dev
->flags
);
3813 if (s
->failed
>= 1) {
3814 dev
= &sh
->dev
[s
->failed_num
[0]];
3816 set_bit(R5_LOCKED
, &dev
->flags
);
3817 set_bit(R5_Wantwrite
, &dev
->flags
);
3819 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3820 dev
= &sh
->dev
[pd_idx
];
3822 set_bit(R5_LOCKED
, &dev
->flags
);
3823 set_bit(R5_Wantwrite
, &dev
->flags
);
3825 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3826 dev
= &sh
->dev
[qd_idx
];
3828 set_bit(R5_LOCKED
, &dev
->flags
);
3829 set_bit(R5_Wantwrite
, &dev
->flags
);
3831 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3833 set_bit(STRIPE_INSYNC
, &sh
->state
);
3835 case check_state_run
:
3836 case check_state_run_q
:
3837 case check_state_run_pq
:
3838 break; /* we will be called again upon completion */
3839 case check_state_check_result
:
3840 sh
->check_state
= check_state_idle
;
3842 /* handle a successful check operation, if parity is correct
3843 * we are done. Otherwise update the mismatch count and repair
3844 * parity if !MD_RECOVERY_CHECK
3846 if (sh
->ops
.zero_sum_result
== 0) {
3847 /* both parities are correct */
3849 set_bit(STRIPE_INSYNC
, &sh
->state
);
3851 /* in contrast to the raid5 case we can validate
3852 * parity, but still have a failure to write
3855 sh
->check_state
= check_state_compute_result
;
3856 /* Returning at this point means that we may go
3857 * off and bring p and/or q uptodate again so
3858 * we make sure to check zero_sum_result again
3859 * to verify if p or q need writeback
3863 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3864 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3865 /* don't try to repair!! */
3866 set_bit(STRIPE_INSYNC
, &sh
->state
);
3868 int *target
= &sh
->ops
.target
;
3870 sh
->ops
.target
= -1;
3871 sh
->ops
.target2
= -1;
3872 sh
->check_state
= check_state_compute_run
;
3873 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3874 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3875 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3876 set_bit(R5_Wantcompute
,
3877 &sh
->dev
[pd_idx
].flags
);
3879 target
= &sh
->ops
.target2
;
3882 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3883 set_bit(R5_Wantcompute
,
3884 &sh
->dev
[qd_idx
].flags
);
3891 case check_state_compute_run
:
3894 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3895 __func__
, sh
->check_state
,
3896 (unsigned long long) sh
->sector
);
3901 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3905 /* We have read all the blocks in this stripe and now we need to
3906 * copy some of them into a target stripe for expand.
3908 struct dma_async_tx_descriptor
*tx
= NULL
;
3909 BUG_ON(sh
->batch_head
);
3910 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3911 for (i
= 0; i
< sh
->disks
; i
++)
3912 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3914 struct stripe_head
*sh2
;
3915 struct async_submit_ctl submit
;
3917 sector_t bn
= compute_blocknr(sh
, i
, 1);
3918 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3920 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3922 /* so far only the early blocks of this stripe
3923 * have been requested. When later blocks
3924 * get requested, we will try again
3927 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3928 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3929 /* must have already done this block */
3930 release_stripe(sh2
);
3934 /* place all the copies on one channel */
3935 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3936 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3937 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3940 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3941 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3942 for (j
= 0; j
< conf
->raid_disks
; j
++)
3943 if (j
!= sh2
->pd_idx
&&
3945 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3947 if (j
== conf
->raid_disks
) {
3948 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3949 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3951 release_stripe(sh2
);
3954 /* done submitting copies, wait for them to complete */
3955 async_tx_quiesce(&tx
);
3959 * handle_stripe - do things to a stripe.
3961 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3962 * state of various bits to see what needs to be done.
3964 * return some read requests which now have data
3965 * return some write requests which are safely on storage
3966 * schedule a read on some buffers
3967 * schedule a write of some buffers
3968 * return confirmation of parity correctness
3972 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3974 struct r5conf
*conf
= sh
->raid_conf
;
3975 int disks
= sh
->disks
;
3978 int do_recovery
= 0;
3980 memset(s
, 0, sizeof(*s
));
3982 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
3983 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
3984 s
->failed_num
[0] = -1;
3985 s
->failed_num
[1] = -1;
3987 /* Now to look around and see what can be done */
3989 for (i
=disks
; i
--; ) {
3990 struct md_rdev
*rdev
;
3997 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3999 dev
->toread
, dev
->towrite
, dev
->written
);
4000 /* maybe we can reply to a read
4002 * new wantfill requests are only permitted while
4003 * ops_complete_biofill is guaranteed to be inactive
4005 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4006 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4007 set_bit(R5_Wantfill
, &dev
->flags
);
4009 /* now count some things */
4010 if (test_bit(R5_LOCKED
, &dev
->flags
))
4012 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4014 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4016 BUG_ON(s
->compute
> 2);
4019 if (test_bit(R5_Wantfill
, &dev
->flags
))
4021 else if (dev
->toread
)
4025 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4030 /* Prefer to use the replacement for reads, but only
4031 * if it is recovered enough and has no bad blocks.
4033 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4034 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4035 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4036 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4037 &first_bad
, &bad_sectors
))
4038 set_bit(R5_ReadRepl
, &dev
->flags
);
4041 set_bit(R5_NeedReplace
, &dev
->flags
);
4042 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4043 clear_bit(R5_ReadRepl
, &dev
->flags
);
4045 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4048 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4049 &first_bad
, &bad_sectors
);
4050 if (s
->blocked_rdev
== NULL
4051 && (test_bit(Blocked
, &rdev
->flags
)
4054 set_bit(BlockedBadBlocks
,
4056 s
->blocked_rdev
= rdev
;
4057 atomic_inc(&rdev
->nr_pending
);
4060 clear_bit(R5_Insync
, &dev
->flags
);
4064 /* also not in-sync */
4065 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4066 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4067 /* treat as in-sync, but with a read error
4068 * which we can now try to correct
4070 set_bit(R5_Insync
, &dev
->flags
);
4071 set_bit(R5_ReadError
, &dev
->flags
);
4073 } else if (test_bit(In_sync
, &rdev
->flags
))
4074 set_bit(R5_Insync
, &dev
->flags
);
4075 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4076 /* in sync if before recovery_offset */
4077 set_bit(R5_Insync
, &dev
->flags
);
4078 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4079 test_bit(R5_Expanded
, &dev
->flags
))
4080 /* If we've reshaped into here, we assume it is Insync.
4081 * We will shortly update recovery_offset to make
4084 set_bit(R5_Insync
, &dev
->flags
);
4086 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4087 /* This flag does not apply to '.replacement'
4088 * only to .rdev, so make sure to check that*/
4089 struct md_rdev
*rdev2
= rcu_dereference(
4090 conf
->disks
[i
].rdev
);
4092 clear_bit(R5_Insync
, &dev
->flags
);
4093 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4094 s
->handle_bad_blocks
= 1;
4095 atomic_inc(&rdev2
->nr_pending
);
4097 clear_bit(R5_WriteError
, &dev
->flags
);
4099 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4100 /* This flag does not apply to '.replacement'
4101 * only to .rdev, so make sure to check that*/
4102 struct md_rdev
*rdev2
= rcu_dereference(
4103 conf
->disks
[i
].rdev
);
4104 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4105 s
->handle_bad_blocks
= 1;
4106 atomic_inc(&rdev2
->nr_pending
);
4108 clear_bit(R5_MadeGood
, &dev
->flags
);
4110 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4111 struct md_rdev
*rdev2
= rcu_dereference(
4112 conf
->disks
[i
].replacement
);
4113 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4114 s
->handle_bad_blocks
= 1;
4115 atomic_inc(&rdev2
->nr_pending
);
4117 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4119 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4120 /* The ReadError flag will just be confusing now */
4121 clear_bit(R5_ReadError
, &dev
->flags
);
4122 clear_bit(R5_ReWrite
, &dev
->flags
);
4124 if (test_bit(R5_ReadError
, &dev
->flags
))
4125 clear_bit(R5_Insync
, &dev
->flags
);
4126 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4128 s
->failed_num
[s
->failed
] = i
;
4130 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4134 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4135 /* If there is a failed device being replaced,
4136 * we must be recovering.
4137 * else if we are after recovery_cp, we must be syncing
4138 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4139 * else we can only be replacing
4140 * sync and recovery both need to read all devices, and so
4141 * use the same flag.
4144 sh
->sector
>= conf
->mddev
->recovery_cp
||
4145 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4153 static int clear_batch_ready(struct stripe_head
*sh
)
4155 struct stripe_head
*tmp
;
4156 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4158 spin_lock(&sh
->stripe_lock
);
4159 if (!sh
->batch_head
) {
4160 spin_unlock(&sh
->stripe_lock
);
4165 * this stripe could be added to a batch list before we check
4166 * BATCH_READY, skips it
4168 if (sh
->batch_head
!= sh
) {
4169 spin_unlock(&sh
->stripe_lock
);
4172 spin_lock(&sh
->batch_lock
);
4173 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4174 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4175 spin_unlock(&sh
->batch_lock
);
4176 spin_unlock(&sh
->stripe_lock
);
4179 * BATCH_READY is cleared, no new stripes can be added.
4180 * batch_list can be accessed without lock
4185 static void check_break_stripe_batch_list(struct stripe_head
*sh
)
4187 struct stripe_head
*head_sh
, *next
;
4190 if (!test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4195 sh
= list_first_entry(&sh
->batch_list
,
4196 struct stripe_head
, batch_list
);
4197 BUG_ON(sh
== head_sh
);
4198 } while (!test_bit(STRIPE_DEGRADED
, &sh
->state
));
4200 while (sh
!= head_sh
) {
4201 next
= list_first_entry(&sh
->batch_list
,
4202 struct stripe_head
, batch_list
);
4203 list_del_init(&sh
->batch_list
);
4205 set_mask_bits(&sh
->state
, ~STRIPE_EXPAND_SYNC_FLAG
,
4206 head_sh
->state
& ~((1 << STRIPE_ACTIVE
) |
4207 (1 << STRIPE_PREREAD_ACTIVE
) |
4208 (1 << STRIPE_DEGRADED
) |
4209 STRIPE_EXPAND_SYNC_FLAG
));
4210 sh
->check_state
= head_sh
->check_state
;
4211 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4212 for (i
= 0; i
< sh
->disks
; i
++)
4213 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4214 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4216 spin_lock_irq(&sh
->stripe_lock
);
4217 sh
->batch_head
= NULL
;
4218 spin_unlock_irq(&sh
->stripe_lock
);
4220 set_bit(STRIPE_HANDLE
, &sh
->state
);
4227 static void handle_stripe(struct stripe_head
*sh
)
4229 struct stripe_head_state s
;
4230 struct r5conf
*conf
= sh
->raid_conf
;
4233 int disks
= sh
->disks
;
4234 struct r5dev
*pdev
, *qdev
;
4236 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4237 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4238 /* already being handled, ensure it gets handled
4239 * again when current action finishes */
4240 set_bit(STRIPE_HANDLE
, &sh
->state
);
4244 if (clear_batch_ready(sh
) ) {
4245 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4249 check_break_stripe_batch_list(sh
);
4251 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4252 spin_lock(&sh
->stripe_lock
);
4253 /* Cannot process 'sync' concurrently with 'discard' */
4254 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4255 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4256 set_bit(STRIPE_SYNCING
, &sh
->state
);
4257 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4258 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4260 spin_unlock(&sh
->stripe_lock
);
4262 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4264 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4265 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4266 (unsigned long long)sh
->sector
, sh
->state
,
4267 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4268 sh
->check_state
, sh
->reconstruct_state
);
4270 analyse_stripe(sh
, &s
);
4272 if (s
.handle_bad_blocks
) {
4273 set_bit(STRIPE_HANDLE
, &sh
->state
);
4277 if (unlikely(s
.blocked_rdev
)) {
4278 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4279 s
.replacing
|| s
.to_write
|| s
.written
) {
4280 set_bit(STRIPE_HANDLE
, &sh
->state
);
4283 /* There is nothing for the blocked_rdev to block */
4284 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4285 s
.blocked_rdev
= NULL
;
4288 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4289 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4290 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4293 pr_debug("locked=%d uptodate=%d to_read=%d"
4294 " to_write=%d failed=%d failed_num=%d,%d\n",
4295 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4296 s
.failed_num
[0], s
.failed_num
[1]);
4297 /* check if the array has lost more than max_degraded devices and,
4298 * if so, some requests might need to be failed.
4300 if (s
.failed
> conf
->max_degraded
) {
4301 sh
->check_state
= 0;
4302 sh
->reconstruct_state
= 0;
4303 if (s
.to_read
+s
.to_write
+s
.written
)
4304 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
4305 if (s
.syncing
+ s
.replacing
)
4306 handle_failed_sync(conf
, sh
, &s
);
4309 /* Now we check to see if any write operations have recently
4313 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4315 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4316 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4317 sh
->reconstruct_state
= reconstruct_state_idle
;
4319 /* All the 'written' buffers and the parity block are ready to
4320 * be written back to disk
4322 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4323 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4324 BUG_ON(sh
->qd_idx
>= 0 &&
4325 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4326 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4327 for (i
= disks
; i
--; ) {
4328 struct r5dev
*dev
= &sh
->dev
[i
];
4329 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4330 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4332 pr_debug("Writing block %d\n", i
);
4333 set_bit(R5_Wantwrite
, &dev
->flags
);
4338 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4339 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4341 set_bit(STRIPE_INSYNC
, &sh
->state
);
4344 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4345 s
.dec_preread_active
= 1;
4349 * might be able to return some write requests if the parity blocks
4350 * are safe, or on a failed drive
4352 pdev
= &sh
->dev
[sh
->pd_idx
];
4353 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4354 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4355 qdev
= &sh
->dev
[sh
->qd_idx
];
4356 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4357 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4361 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4362 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4363 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4364 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4365 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4366 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4367 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4368 test_bit(R5_Discard
, &qdev
->flags
))))))
4369 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
4371 /* Now we might consider reading some blocks, either to check/generate
4372 * parity, or to satisfy requests
4373 * or to load a block that is being partially written.
4375 if (s
.to_read
|| s
.non_overwrite
4376 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4377 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4380 handle_stripe_fill(sh
, &s
, disks
);
4382 /* Now to consider new write requests and what else, if anything
4383 * should be read. We do not handle new writes when:
4384 * 1/ A 'write' operation (copy+xor) is already in flight.
4385 * 2/ A 'check' operation is in flight, as it may clobber the parity
4388 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
4389 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4391 /* maybe we need to check and possibly fix the parity for this stripe
4392 * Any reads will already have been scheduled, so we just see if enough
4393 * data is available. The parity check is held off while parity
4394 * dependent operations are in flight.
4396 if (sh
->check_state
||
4397 (s
.syncing
&& s
.locked
== 0 &&
4398 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4399 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4400 if (conf
->level
== 6)
4401 handle_parity_checks6(conf
, sh
, &s
, disks
);
4403 handle_parity_checks5(conf
, sh
, &s
, disks
);
4406 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4407 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4408 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4409 /* Write out to replacement devices where possible */
4410 for (i
= 0; i
< conf
->raid_disks
; i
++)
4411 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4412 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4413 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4414 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4418 set_bit(STRIPE_INSYNC
, &sh
->state
);
4419 set_bit(STRIPE_REPLACED
, &sh
->state
);
4421 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4422 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4423 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4424 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4425 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4426 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4427 wake_up(&conf
->wait_for_overlap
);
4430 /* If the failed drives are just a ReadError, then we might need
4431 * to progress the repair/check process
4433 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4434 for (i
= 0; i
< s
.failed
; i
++) {
4435 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4436 if (test_bit(R5_ReadError
, &dev
->flags
)
4437 && !test_bit(R5_LOCKED
, &dev
->flags
)
4438 && test_bit(R5_UPTODATE
, &dev
->flags
)
4440 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4441 set_bit(R5_Wantwrite
, &dev
->flags
);
4442 set_bit(R5_ReWrite
, &dev
->flags
);
4443 set_bit(R5_LOCKED
, &dev
->flags
);
4446 /* let's read it back */
4447 set_bit(R5_Wantread
, &dev
->flags
);
4448 set_bit(R5_LOCKED
, &dev
->flags
);
4454 /* Finish reconstruct operations initiated by the expansion process */
4455 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4456 struct stripe_head
*sh_src
4457 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4458 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4459 /* sh cannot be written until sh_src has been read.
4460 * so arrange for sh to be delayed a little
4462 set_bit(STRIPE_DELAYED
, &sh
->state
);
4463 set_bit(STRIPE_HANDLE
, &sh
->state
);
4464 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4466 atomic_inc(&conf
->preread_active_stripes
);
4467 release_stripe(sh_src
);
4471 release_stripe(sh_src
);
4473 sh
->reconstruct_state
= reconstruct_state_idle
;
4474 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4475 for (i
= conf
->raid_disks
; i
--; ) {
4476 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4477 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4482 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4483 !sh
->reconstruct_state
) {
4484 /* Need to write out all blocks after computing parity */
4485 sh
->disks
= conf
->raid_disks
;
4486 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4487 schedule_reconstruction(sh
, &s
, 1, 1);
4488 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4489 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4490 atomic_dec(&conf
->reshape_stripes
);
4491 wake_up(&conf
->wait_for_overlap
);
4492 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4495 if (s
.expanding
&& s
.locked
== 0 &&
4496 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4497 handle_stripe_expansion(conf
, sh
);
4500 /* wait for this device to become unblocked */
4501 if (unlikely(s
.blocked_rdev
)) {
4502 if (conf
->mddev
->external
)
4503 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4506 /* Internal metadata will immediately
4507 * be written by raid5d, so we don't
4508 * need to wait here.
4510 rdev_dec_pending(s
.blocked_rdev
,
4514 if (s
.handle_bad_blocks
)
4515 for (i
= disks
; i
--; ) {
4516 struct md_rdev
*rdev
;
4517 struct r5dev
*dev
= &sh
->dev
[i
];
4518 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4519 /* We own a safe reference to the rdev */
4520 rdev
= conf
->disks
[i
].rdev
;
4521 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4523 md_error(conf
->mddev
, rdev
);
4524 rdev_dec_pending(rdev
, conf
->mddev
);
4526 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4527 rdev
= conf
->disks
[i
].rdev
;
4528 rdev_clear_badblocks(rdev
, sh
->sector
,
4530 rdev_dec_pending(rdev
, conf
->mddev
);
4532 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4533 rdev
= conf
->disks
[i
].replacement
;
4535 /* rdev have been moved down */
4536 rdev
= conf
->disks
[i
].rdev
;
4537 rdev_clear_badblocks(rdev
, sh
->sector
,
4539 rdev_dec_pending(rdev
, conf
->mddev
);
4544 raid_run_ops(sh
, s
.ops_request
);
4548 if (s
.dec_preread_active
) {
4549 /* We delay this until after ops_run_io so that if make_request
4550 * is waiting on a flush, it won't continue until the writes
4551 * have actually been submitted.
4553 atomic_dec(&conf
->preread_active_stripes
);
4554 if (atomic_read(&conf
->preread_active_stripes
) <
4556 md_wakeup_thread(conf
->mddev
->thread
);
4559 return_io(s
.return_bi
);
4561 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4564 static void raid5_activate_delayed(struct r5conf
*conf
)
4566 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
4567 while (!list_empty(&conf
->delayed_list
)) {
4568 struct list_head
*l
= conf
->delayed_list
.next
;
4569 struct stripe_head
*sh
;
4570 sh
= list_entry(l
, struct stripe_head
, lru
);
4572 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4573 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4574 atomic_inc(&conf
->preread_active_stripes
);
4575 list_add_tail(&sh
->lru
, &conf
->hold_list
);
4576 raid5_wakeup_stripe_thread(sh
);
4581 static void activate_bit_delay(struct r5conf
*conf
,
4582 struct list_head
*temp_inactive_list
)
4584 /* device_lock is held */
4585 struct list_head head
;
4586 list_add(&head
, &conf
->bitmap_list
);
4587 list_del_init(&conf
->bitmap_list
);
4588 while (!list_empty(&head
)) {
4589 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
4591 list_del_init(&sh
->lru
);
4592 atomic_inc(&sh
->count
);
4593 hash
= sh
->hash_lock_index
;
4594 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
4598 static int raid5_congested(struct mddev
*mddev
, int bits
)
4600 struct r5conf
*conf
= mddev
->private;
4602 /* No difference between reads and writes. Just check
4603 * how busy the stripe_cache is
4606 if (conf
->inactive_blocked
)
4610 if (atomic_read(&conf
->empty_inactive_list_nr
))
4616 /* We want read requests to align with chunks where possible,
4617 * but write requests don't need to.
4619 static int raid5_mergeable_bvec(struct mddev
*mddev
,
4620 struct bvec_merge_data
*bvm
,
4621 struct bio_vec
*biovec
)
4623 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
4625 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4626 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
4628 if ((bvm
->bi_rw
& 1) == WRITE
)
4629 return biovec
->bv_len
; /* always allow writes to be mergeable */
4631 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4632 chunk_sectors
= mddev
->new_chunk_sectors
;
4633 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
4634 if (max
< 0) max
= 0;
4635 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
4636 return biovec
->bv_len
;
4641 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
4643 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
4644 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4645 unsigned int bio_sectors
= bio_sectors(bio
);
4647 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4648 chunk_sectors
= mddev
->new_chunk_sectors
;
4649 return chunk_sectors
>=
4650 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
4654 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4655 * later sampled by raid5d.
4657 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
4659 unsigned long flags
;
4661 spin_lock_irqsave(&conf
->device_lock
, flags
);
4663 bi
->bi_next
= conf
->retry_read_aligned_list
;
4664 conf
->retry_read_aligned_list
= bi
;
4666 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4667 md_wakeup_thread(conf
->mddev
->thread
);
4670 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
4674 bi
= conf
->retry_read_aligned
;
4676 conf
->retry_read_aligned
= NULL
;
4679 bi
= conf
->retry_read_aligned_list
;
4681 conf
->retry_read_aligned_list
= bi
->bi_next
;
4684 * this sets the active strip count to 1 and the processed
4685 * strip count to zero (upper 8 bits)
4687 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
4694 * The "raid5_align_endio" should check if the read succeeded and if it
4695 * did, call bio_endio on the original bio (having bio_put the new bio
4697 * If the read failed..
4699 static void raid5_align_endio(struct bio
*bi
, int error
)
4701 struct bio
* raid_bi
= bi
->bi_private
;
4702 struct mddev
*mddev
;
4703 struct r5conf
*conf
;
4704 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4705 struct md_rdev
*rdev
;
4709 rdev
= (void*)raid_bi
->bi_next
;
4710 raid_bi
->bi_next
= NULL
;
4711 mddev
= rdev
->mddev
;
4712 conf
= mddev
->private;
4714 rdev_dec_pending(rdev
, conf
->mddev
);
4716 if (!error
&& uptodate
) {
4717 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
4719 bio_endio(raid_bi
, 0);
4720 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4721 wake_up(&conf
->wait_for_stripe
);
4725 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4727 add_bio_to_retry(raid_bi
, conf
);
4730 static int bio_fits_rdev(struct bio
*bi
)
4732 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
4734 if (bio_sectors(bi
) > queue_max_sectors(q
))
4736 blk_recount_segments(q
, bi
);
4737 if (bi
->bi_phys_segments
> queue_max_segments(q
))
4740 if (q
->merge_bvec_fn
)
4741 /* it's too hard to apply the merge_bvec_fn at this stage,
4749 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
4751 struct r5conf
*conf
= mddev
->private;
4753 struct bio
* align_bi
;
4754 struct md_rdev
*rdev
;
4755 sector_t end_sector
;
4757 if (!in_chunk_boundary(mddev
, raid_bio
)) {
4758 pr_debug("chunk_aligned_read : non aligned\n");
4762 * use bio_clone_mddev to make a copy of the bio
4764 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
4768 * set bi_end_io to a new function, and set bi_private to the
4771 align_bi
->bi_end_io
= raid5_align_endio
;
4772 align_bi
->bi_private
= raid_bio
;
4776 align_bi
->bi_iter
.bi_sector
=
4777 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
4780 end_sector
= bio_end_sector(align_bi
);
4782 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
4783 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
4784 rdev
->recovery_offset
< end_sector
) {
4785 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
4787 (test_bit(Faulty
, &rdev
->flags
) ||
4788 !(test_bit(In_sync
, &rdev
->flags
) ||
4789 rdev
->recovery_offset
>= end_sector
)))
4796 atomic_inc(&rdev
->nr_pending
);
4798 raid_bio
->bi_next
= (void*)rdev
;
4799 align_bi
->bi_bdev
= rdev
->bdev
;
4800 __clear_bit(BIO_SEG_VALID
, &align_bi
->bi_flags
);
4802 if (!bio_fits_rdev(align_bi
) ||
4803 is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
4804 bio_sectors(align_bi
),
4805 &first_bad
, &bad_sectors
)) {
4806 /* too big in some way, or has a known bad block */
4808 rdev_dec_pending(rdev
, mddev
);
4812 /* No reshape active, so we can trust rdev->data_offset */
4813 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
4815 spin_lock_irq(&conf
->device_lock
);
4816 wait_event_lock_irq(conf
->wait_for_stripe
,
4819 atomic_inc(&conf
->active_aligned_reads
);
4820 spin_unlock_irq(&conf
->device_lock
);
4823 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4824 align_bi
, disk_devt(mddev
->gendisk
),
4825 raid_bio
->bi_iter
.bi_sector
);
4826 generic_make_request(align_bi
);
4835 /* __get_priority_stripe - get the next stripe to process
4837 * Full stripe writes are allowed to pass preread active stripes up until
4838 * the bypass_threshold is exceeded. In general the bypass_count
4839 * increments when the handle_list is handled before the hold_list; however, it
4840 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4841 * stripe with in flight i/o. The bypass_count will be reset when the
4842 * head of the hold_list has changed, i.e. the head was promoted to the
4845 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
4847 struct stripe_head
*sh
= NULL
, *tmp
;
4848 struct list_head
*handle_list
= NULL
;
4849 struct r5worker_group
*wg
= NULL
;
4851 if (conf
->worker_cnt_per_group
== 0) {
4852 handle_list
= &conf
->handle_list
;
4853 } else if (group
!= ANY_GROUP
) {
4854 handle_list
= &conf
->worker_groups
[group
].handle_list
;
4855 wg
= &conf
->worker_groups
[group
];
4858 for (i
= 0; i
< conf
->group_cnt
; i
++) {
4859 handle_list
= &conf
->worker_groups
[i
].handle_list
;
4860 wg
= &conf
->worker_groups
[i
];
4861 if (!list_empty(handle_list
))
4866 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4868 list_empty(handle_list
) ? "empty" : "busy",
4869 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4870 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4872 if (!list_empty(handle_list
)) {
4873 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
4875 if (list_empty(&conf
->hold_list
))
4876 conf
->bypass_count
= 0;
4877 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4878 if (conf
->hold_list
.next
== conf
->last_hold
)
4879 conf
->bypass_count
++;
4881 conf
->last_hold
= conf
->hold_list
.next
;
4882 conf
->bypass_count
-= conf
->bypass_threshold
;
4883 if (conf
->bypass_count
< 0)
4884 conf
->bypass_count
= 0;
4887 } else if (!list_empty(&conf
->hold_list
) &&
4888 ((conf
->bypass_threshold
&&
4889 conf
->bypass_count
> conf
->bypass_threshold
) ||
4890 atomic_read(&conf
->pending_full_writes
) == 0)) {
4892 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
4893 if (conf
->worker_cnt_per_group
== 0 ||
4894 group
== ANY_GROUP
||
4895 !cpu_online(tmp
->cpu
) ||
4896 cpu_to_group(tmp
->cpu
) == group
) {
4903 conf
->bypass_count
-= conf
->bypass_threshold
;
4904 if (conf
->bypass_count
< 0)
4905 conf
->bypass_count
= 0;
4917 list_del_init(&sh
->lru
);
4918 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
4922 struct raid5_plug_cb
{
4923 struct blk_plug_cb cb
;
4924 struct list_head list
;
4925 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
4928 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4930 struct raid5_plug_cb
*cb
= container_of(
4931 blk_cb
, struct raid5_plug_cb
, cb
);
4932 struct stripe_head
*sh
;
4933 struct mddev
*mddev
= cb
->cb
.data
;
4934 struct r5conf
*conf
= mddev
->private;
4938 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4939 spin_lock_irq(&conf
->device_lock
);
4940 while (!list_empty(&cb
->list
)) {
4941 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4942 list_del_init(&sh
->lru
);
4944 * avoid race release_stripe_plug() sees
4945 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4946 * is still in our list
4948 smp_mb__before_atomic();
4949 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4951 * STRIPE_ON_RELEASE_LIST could be set here. In that
4952 * case, the count is always > 1 here
4954 hash
= sh
->hash_lock_index
;
4955 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
4958 spin_unlock_irq(&conf
->device_lock
);
4960 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
4961 NR_STRIPE_HASH_LOCKS
);
4963 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
4967 static void release_stripe_plug(struct mddev
*mddev
,
4968 struct stripe_head
*sh
)
4970 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4971 raid5_unplug
, mddev
,
4972 sizeof(struct raid5_plug_cb
));
4973 struct raid5_plug_cb
*cb
;
4980 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4982 if (cb
->list
.next
== NULL
) {
4984 INIT_LIST_HEAD(&cb
->list
);
4985 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
4986 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
4989 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
4990 list_add_tail(&sh
->lru
, &cb
->list
);
4995 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
4997 struct r5conf
*conf
= mddev
->private;
4998 sector_t logical_sector
, last_sector
;
4999 struct stripe_head
*sh
;
5003 if (mddev
->reshape_position
!= MaxSector
)
5004 /* Skip discard while reshape is happening */
5007 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5008 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5011 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5013 stripe_sectors
= conf
->chunk_sectors
*
5014 (conf
->raid_disks
- conf
->max_degraded
);
5015 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5017 sector_div(last_sector
, stripe_sectors
);
5019 logical_sector
*= conf
->chunk_sectors
;
5020 last_sector
*= conf
->chunk_sectors
;
5022 for (; logical_sector
< last_sector
;
5023 logical_sector
+= STRIPE_SECTORS
) {
5027 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5028 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5029 TASK_UNINTERRUPTIBLE
);
5030 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5031 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5036 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5037 spin_lock_irq(&sh
->stripe_lock
);
5038 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5039 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5041 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5042 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5043 spin_unlock_irq(&sh
->stripe_lock
);
5049 set_bit(STRIPE_DISCARD
, &sh
->state
);
5050 finish_wait(&conf
->wait_for_overlap
, &w
);
5051 sh
->overwrite_disks
= 0;
5052 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5053 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5055 sh
->dev
[d
].towrite
= bi
;
5056 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5057 raid5_inc_bi_active_stripes(bi
);
5058 sh
->overwrite_disks
++;
5060 spin_unlock_irq(&sh
->stripe_lock
);
5061 if (conf
->mddev
->bitmap
) {
5063 d
< conf
->raid_disks
- conf
->max_degraded
;
5065 bitmap_startwrite(mddev
->bitmap
,
5069 sh
->bm_seq
= conf
->seq_flush
+ 1;
5070 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5073 set_bit(STRIPE_HANDLE
, &sh
->state
);
5074 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5075 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5076 atomic_inc(&conf
->preread_active_stripes
);
5077 release_stripe_plug(mddev
, sh
);
5080 remaining
= raid5_dec_bi_active_stripes(bi
);
5081 if (remaining
== 0) {
5082 md_write_end(mddev
);
5087 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
5089 struct r5conf
*conf
= mddev
->private;
5091 sector_t new_sector
;
5092 sector_t logical_sector
, last_sector
;
5093 struct stripe_head
*sh
;
5094 const int rw
= bio_data_dir(bi
);
5099 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
5100 md_flush_request(mddev
, bi
);
5104 md_write_start(mddev
, bi
);
5107 mddev
->reshape_position
== MaxSector
&&
5108 chunk_aligned_read(mddev
,bi
))
5111 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
5112 make_discard_request(mddev
, bi
);
5116 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5117 last_sector
= bio_end_sector(bi
);
5119 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5121 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5122 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5128 seq
= read_seqcount_begin(&conf
->gen_lock
);
5131 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5132 TASK_UNINTERRUPTIBLE
);
5133 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5134 /* spinlock is needed as reshape_progress may be
5135 * 64bit on a 32bit platform, and so it might be
5136 * possible to see a half-updated value
5137 * Of course reshape_progress could change after
5138 * the lock is dropped, so once we get a reference
5139 * to the stripe that we think it is, we will have
5142 spin_lock_irq(&conf
->device_lock
);
5143 if (mddev
->reshape_backwards
5144 ? logical_sector
< conf
->reshape_progress
5145 : logical_sector
>= conf
->reshape_progress
) {
5148 if (mddev
->reshape_backwards
5149 ? logical_sector
< conf
->reshape_safe
5150 : logical_sector
>= conf
->reshape_safe
) {
5151 spin_unlock_irq(&conf
->device_lock
);
5157 spin_unlock_irq(&conf
->device_lock
);
5160 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5163 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5164 (unsigned long long)new_sector
,
5165 (unsigned long long)logical_sector
);
5167 sh
= get_active_stripe(conf
, new_sector
, previous
,
5168 (bi
->bi_rw
&RWA_MASK
), 0);
5170 if (unlikely(previous
)) {
5171 /* expansion might have moved on while waiting for a
5172 * stripe, so we must do the range check again.
5173 * Expansion could still move past after this
5174 * test, but as we are holding a reference to
5175 * 'sh', we know that if that happens,
5176 * STRIPE_EXPANDING will get set and the expansion
5177 * won't proceed until we finish with the stripe.
5180 spin_lock_irq(&conf
->device_lock
);
5181 if (mddev
->reshape_backwards
5182 ? logical_sector
>= conf
->reshape_progress
5183 : logical_sector
< conf
->reshape_progress
)
5184 /* mismatch, need to try again */
5186 spin_unlock_irq(&conf
->device_lock
);
5194 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5195 /* Might have got the wrong stripe_head
5203 logical_sector
>= mddev
->suspend_lo
&&
5204 logical_sector
< mddev
->suspend_hi
) {
5206 /* As the suspend_* range is controlled by
5207 * userspace, we want an interruptible
5210 flush_signals(current
);
5211 prepare_to_wait(&conf
->wait_for_overlap
,
5212 &w
, TASK_INTERRUPTIBLE
);
5213 if (logical_sector
>= mddev
->suspend_lo
&&
5214 logical_sector
< mddev
->suspend_hi
) {
5221 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5222 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5223 /* Stripe is busy expanding or
5224 * add failed due to overlap. Flush everything
5227 md_wakeup_thread(mddev
->thread
);
5233 set_bit(STRIPE_HANDLE
, &sh
->state
);
5234 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5235 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5236 (bi
->bi_rw
& REQ_SYNC
) &&
5237 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5238 atomic_inc(&conf
->preread_active_stripes
);
5239 release_stripe_plug(mddev
, sh
);
5241 /* cannot get stripe for read-ahead, just give-up */
5242 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
5246 finish_wait(&conf
->wait_for_overlap
, &w
);
5248 remaining
= raid5_dec_bi_active_stripes(bi
);
5249 if (remaining
== 0) {
5252 md_write_end(mddev
);
5254 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
5260 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5262 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5264 /* reshaping is quite different to recovery/resync so it is
5265 * handled quite separately ... here.
5267 * On each call to sync_request, we gather one chunk worth of
5268 * destination stripes and flag them as expanding.
5269 * Then we find all the source stripes and request reads.
5270 * As the reads complete, handle_stripe will copy the data
5271 * into the destination stripe and release that stripe.
5273 struct r5conf
*conf
= mddev
->private;
5274 struct stripe_head
*sh
;
5275 sector_t first_sector
, last_sector
;
5276 int raid_disks
= conf
->previous_raid_disks
;
5277 int data_disks
= raid_disks
- conf
->max_degraded
;
5278 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5281 sector_t writepos
, readpos
, safepos
;
5282 sector_t stripe_addr
;
5283 int reshape_sectors
;
5284 struct list_head stripes
;
5286 if (sector_nr
== 0) {
5287 /* If restarting in the middle, skip the initial sectors */
5288 if (mddev
->reshape_backwards
&&
5289 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5290 sector_nr
= raid5_size(mddev
, 0, 0)
5291 - conf
->reshape_progress
;
5292 } else if (!mddev
->reshape_backwards
&&
5293 conf
->reshape_progress
> 0)
5294 sector_nr
= conf
->reshape_progress
;
5295 sector_div(sector_nr
, new_data_disks
);
5297 mddev
->curr_resync_completed
= sector_nr
;
5298 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5304 /* We need to process a full chunk at a time.
5305 * If old and new chunk sizes differ, we need to process the
5308 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
5309 reshape_sectors
= mddev
->new_chunk_sectors
;
5311 reshape_sectors
= mddev
->chunk_sectors
;
5313 /* We update the metadata at least every 10 seconds, or when
5314 * the data about to be copied would over-write the source of
5315 * the data at the front of the range. i.e. one new_stripe
5316 * along from reshape_progress new_maps to after where
5317 * reshape_safe old_maps to
5319 writepos
= conf
->reshape_progress
;
5320 sector_div(writepos
, new_data_disks
);
5321 readpos
= conf
->reshape_progress
;
5322 sector_div(readpos
, data_disks
);
5323 safepos
= conf
->reshape_safe
;
5324 sector_div(safepos
, data_disks
);
5325 if (mddev
->reshape_backwards
) {
5326 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
5327 readpos
+= reshape_sectors
;
5328 safepos
+= reshape_sectors
;
5330 writepos
+= reshape_sectors
;
5331 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5332 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5335 /* Having calculated the 'writepos' possibly use it
5336 * to set 'stripe_addr' which is where we will write to.
5338 if (mddev
->reshape_backwards
) {
5339 BUG_ON(conf
->reshape_progress
== 0);
5340 stripe_addr
= writepos
;
5341 BUG_ON((mddev
->dev_sectors
&
5342 ~((sector_t
)reshape_sectors
- 1))
5343 - reshape_sectors
- stripe_addr
5346 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5347 stripe_addr
= sector_nr
;
5350 /* 'writepos' is the most advanced device address we might write.
5351 * 'readpos' is the least advanced device address we might read.
5352 * 'safepos' is the least address recorded in the metadata as having
5354 * If there is a min_offset_diff, these are adjusted either by
5355 * increasing the safepos/readpos if diff is negative, or
5356 * increasing writepos if diff is positive.
5357 * If 'readpos' is then behind 'writepos', there is no way that we can
5358 * ensure safety in the face of a crash - that must be done by userspace
5359 * making a backup of the data. So in that case there is no particular
5360 * rush to update metadata.
5361 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5362 * update the metadata to advance 'safepos' to match 'readpos' so that
5363 * we can be safe in the event of a crash.
5364 * So we insist on updating metadata if safepos is behind writepos and
5365 * readpos is beyond writepos.
5366 * In any case, update the metadata every 10 seconds.
5367 * Maybe that number should be configurable, but I'm not sure it is
5368 * worth it.... maybe it could be a multiple of safemode_delay???
5370 if (conf
->min_offset_diff
< 0) {
5371 safepos
+= -conf
->min_offset_diff
;
5372 readpos
+= -conf
->min_offset_diff
;
5374 writepos
+= conf
->min_offset_diff
;
5376 if ((mddev
->reshape_backwards
5377 ? (safepos
> writepos
&& readpos
< writepos
)
5378 : (safepos
< writepos
&& readpos
> writepos
)) ||
5379 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5380 /* Cannot proceed until we've updated the superblock... */
5381 wait_event(conf
->wait_for_overlap
,
5382 atomic_read(&conf
->reshape_stripes
)==0
5383 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5384 if (atomic_read(&conf
->reshape_stripes
) != 0)
5386 mddev
->reshape_position
= conf
->reshape_progress
;
5387 mddev
->curr_resync_completed
= sector_nr
;
5388 conf
->reshape_checkpoint
= jiffies
;
5389 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5390 md_wakeup_thread(mddev
->thread
);
5391 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
5392 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5393 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5395 spin_lock_irq(&conf
->device_lock
);
5396 conf
->reshape_safe
= mddev
->reshape_position
;
5397 spin_unlock_irq(&conf
->device_lock
);
5398 wake_up(&conf
->wait_for_overlap
);
5399 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5402 INIT_LIST_HEAD(&stripes
);
5403 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5405 int skipped_disk
= 0;
5406 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5407 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5408 atomic_inc(&conf
->reshape_stripes
);
5409 /* If any of this stripe is beyond the end of the old
5410 * array, then we need to zero those blocks
5412 for (j
=sh
->disks
; j
--;) {
5414 if (j
== sh
->pd_idx
)
5416 if (conf
->level
== 6 &&
5419 s
= compute_blocknr(sh
, j
, 0);
5420 if (s
< raid5_size(mddev
, 0, 0)) {
5424 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5425 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5426 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5428 if (!skipped_disk
) {
5429 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5430 set_bit(STRIPE_HANDLE
, &sh
->state
);
5432 list_add(&sh
->lru
, &stripes
);
5434 spin_lock_irq(&conf
->device_lock
);
5435 if (mddev
->reshape_backwards
)
5436 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5438 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5439 spin_unlock_irq(&conf
->device_lock
);
5440 /* Ok, those stripe are ready. We can start scheduling
5441 * reads on the source stripes.
5442 * The source stripes are determined by mapping the first and last
5443 * block on the destination stripes.
5446 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5449 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5450 * new_data_disks
- 1),
5452 if (last_sector
>= mddev
->dev_sectors
)
5453 last_sector
= mddev
->dev_sectors
- 1;
5454 while (first_sector
<= last_sector
) {
5455 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
5456 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5457 set_bit(STRIPE_HANDLE
, &sh
->state
);
5459 first_sector
+= STRIPE_SECTORS
;
5461 /* Now that the sources are clearly marked, we can release
5462 * the destination stripes
5464 while (!list_empty(&stripes
)) {
5465 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5466 list_del_init(&sh
->lru
);
5469 /* If this takes us to the resync_max point where we have to pause,
5470 * then we need to write out the superblock.
5472 sector_nr
+= reshape_sectors
;
5473 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
5474 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5475 /* Cannot proceed until we've updated the superblock... */
5476 wait_event(conf
->wait_for_overlap
,
5477 atomic_read(&conf
->reshape_stripes
) == 0
5478 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5479 if (atomic_read(&conf
->reshape_stripes
) != 0)
5481 mddev
->reshape_position
= conf
->reshape_progress
;
5482 mddev
->curr_resync_completed
= sector_nr
;
5483 conf
->reshape_checkpoint
= jiffies
;
5484 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5485 md_wakeup_thread(mddev
->thread
);
5486 wait_event(mddev
->sb_wait
,
5487 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
5488 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5489 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5491 spin_lock_irq(&conf
->device_lock
);
5492 conf
->reshape_safe
= mddev
->reshape_position
;
5493 spin_unlock_irq(&conf
->device_lock
);
5494 wake_up(&conf
->wait_for_overlap
);
5495 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5498 return reshape_sectors
;
5501 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5503 struct r5conf
*conf
= mddev
->private;
5504 struct stripe_head
*sh
;
5505 sector_t max_sector
= mddev
->dev_sectors
;
5506 sector_t sync_blocks
;
5507 int still_degraded
= 0;
5510 if (sector_nr
>= max_sector
) {
5511 /* just being told to finish up .. nothing much to do */
5513 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
5518 if (mddev
->curr_resync
< max_sector
) /* aborted */
5519 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
5521 else /* completed sync */
5523 bitmap_close_sync(mddev
->bitmap
);
5528 /* Allow raid5_quiesce to complete */
5529 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
5531 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
5532 return reshape_request(mddev
, sector_nr
, skipped
);
5534 /* No need to check resync_max as we never do more than one
5535 * stripe, and as resync_max will always be on a chunk boundary,
5536 * if the check in md_do_sync didn't fire, there is no chance
5537 * of overstepping resync_max here
5540 /* if there is too many failed drives and we are trying
5541 * to resync, then assert that we are finished, because there is
5542 * nothing we can do.
5544 if (mddev
->degraded
>= conf
->max_degraded
&&
5545 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
5546 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
5550 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
5552 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
5553 sync_blocks
>= STRIPE_SECTORS
) {
5554 /* we can skip this block, and probably more */
5555 sync_blocks
/= STRIPE_SECTORS
;
5557 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
5560 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
5562 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
5564 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
5565 /* make sure we don't swamp the stripe cache if someone else
5566 * is trying to get access
5568 schedule_timeout_uninterruptible(1);
5570 /* Need to check if array will still be degraded after recovery/resync
5571 * Note in case of > 1 drive failures it's possible we're rebuilding
5572 * one drive while leaving another faulty drive in array.
5575 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5576 struct md_rdev
*rdev
= ACCESS_ONCE(conf
->disks
[i
].rdev
);
5578 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
5583 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
5585 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
5586 set_bit(STRIPE_HANDLE
, &sh
->state
);
5590 return STRIPE_SECTORS
;
5593 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
5595 /* We may not be able to submit a whole bio at once as there
5596 * may not be enough stripe_heads available.
5597 * We cannot pre-allocate enough stripe_heads as we may need
5598 * more than exist in the cache (if we allow ever large chunks).
5599 * So we do one stripe head at a time and record in
5600 * ->bi_hw_segments how many have been done.
5602 * We *know* that this entire raid_bio is in one chunk, so
5603 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5605 struct stripe_head
*sh
;
5607 sector_t sector
, logical_sector
, last_sector
;
5612 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
5613 ~((sector_t
)STRIPE_SECTORS
-1);
5614 sector
= raid5_compute_sector(conf
, logical_sector
,
5616 last_sector
= bio_end_sector(raid_bio
);
5618 for (; logical_sector
< last_sector
;
5619 logical_sector
+= STRIPE_SECTORS
,
5620 sector
+= STRIPE_SECTORS
,
5623 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
5624 /* already done this stripe */
5627 sh
= get_active_stripe(conf
, sector
, 0, 1, 1);
5630 /* failed to get a stripe - must wait */
5631 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5632 conf
->retry_read_aligned
= raid_bio
;
5636 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
5638 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5639 conf
->retry_read_aligned
= raid_bio
;
5643 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
5648 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
5649 if (remaining
== 0) {
5650 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
5652 bio_endio(raid_bio
, 0);
5654 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5655 wake_up(&conf
->wait_for_stripe
);
5659 static int handle_active_stripes(struct r5conf
*conf
, int group
,
5660 struct r5worker
*worker
,
5661 struct list_head
*temp_inactive_list
)
5663 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
5664 int i
, batch_size
= 0, hash
;
5665 bool release_inactive
= false;
5667 while (batch_size
< MAX_STRIPE_BATCH
&&
5668 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
5669 batch
[batch_size
++] = sh
;
5671 if (batch_size
== 0) {
5672 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5673 if (!list_empty(temp_inactive_list
+ i
))
5675 if (i
== NR_STRIPE_HASH_LOCKS
)
5677 release_inactive
= true;
5679 spin_unlock_irq(&conf
->device_lock
);
5681 release_inactive_stripe_list(conf
, temp_inactive_list
,
5682 NR_STRIPE_HASH_LOCKS
);
5684 if (release_inactive
) {
5685 spin_lock_irq(&conf
->device_lock
);
5689 for (i
= 0; i
< batch_size
; i
++)
5690 handle_stripe(batch
[i
]);
5694 spin_lock_irq(&conf
->device_lock
);
5695 for (i
= 0; i
< batch_size
; i
++) {
5696 hash
= batch
[i
]->hash_lock_index
;
5697 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
5702 static void raid5_do_work(struct work_struct
*work
)
5704 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
5705 struct r5worker_group
*group
= worker
->group
;
5706 struct r5conf
*conf
= group
->conf
;
5707 int group_id
= group
- conf
->worker_groups
;
5709 struct blk_plug plug
;
5711 pr_debug("+++ raid5worker active\n");
5713 blk_start_plug(&plug
);
5715 spin_lock_irq(&conf
->device_lock
);
5717 int batch_size
, released
;
5719 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
5721 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
5722 worker
->temp_inactive_list
);
5723 worker
->working
= false;
5724 if (!batch_size
&& !released
)
5726 handled
+= batch_size
;
5728 pr_debug("%d stripes handled\n", handled
);
5730 spin_unlock_irq(&conf
->device_lock
);
5731 blk_finish_plug(&plug
);
5733 pr_debug("--- raid5worker inactive\n");
5737 * This is our raid5 kernel thread.
5739 * We scan the hash table for stripes which can be handled now.
5740 * During the scan, completed stripes are saved for us by the interrupt
5741 * handler, so that they will not have to wait for our next wakeup.
5743 static void raid5d(struct md_thread
*thread
)
5745 struct mddev
*mddev
= thread
->mddev
;
5746 struct r5conf
*conf
= mddev
->private;
5748 struct blk_plug plug
;
5750 pr_debug("+++ raid5d active\n");
5752 md_check_recovery(mddev
);
5754 blk_start_plug(&plug
);
5756 spin_lock_irq(&conf
->device_lock
);
5759 int batch_size
, released
;
5761 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
5764 !list_empty(&conf
->bitmap_list
)) {
5765 /* Now is a good time to flush some bitmap updates */
5767 spin_unlock_irq(&conf
->device_lock
);
5768 bitmap_unplug(mddev
->bitmap
);
5769 spin_lock_irq(&conf
->device_lock
);
5770 conf
->seq_write
= conf
->seq_flush
;
5771 activate_bit_delay(conf
, conf
->temp_inactive_list
);
5773 raid5_activate_delayed(conf
);
5775 while ((bio
= remove_bio_from_retry(conf
))) {
5777 spin_unlock_irq(&conf
->device_lock
);
5778 ok
= retry_aligned_read(conf
, bio
);
5779 spin_lock_irq(&conf
->device_lock
);
5785 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
5786 conf
->temp_inactive_list
);
5787 if (!batch_size
&& !released
)
5789 handled
+= batch_size
;
5791 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
5792 spin_unlock_irq(&conf
->device_lock
);
5793 md_check_recovery(mddev
);
5794 spin_lock_irq(&conf
->device_lock
);
5797 pr_debug("%d stripes handled\n", handled
);
5799 spin_unlock_irq(&conf
->device_lock
);
5801 async_tx_issue_pending_all();
5802 blk_finish_plug(&plug
);
5804 pr_debug("--- raid5d inactive\n");
5808 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
5810 struct r5conf
*conf
;
5812 spin_lock(&mddev
->lock
);
5813 conf
= mddev
->private;
5815 ret
= sprintf(page
, "%d\n", conf
->max_nr_stripes
);
5816 spin_unlock(&mddev
->lock
);
5821 raid5_set_cache_size(struct mddev
*mddev
, int size
)
5823 struct r5conf
*conf
= mddev
->private;
5827 if (size
<= 16 || size
> 32768)
5829 hash
= (conf
->max_nr_stripes
- 1) % NR_STRIPE_HASH_LOCKS
;
5830 while (size
< conf
->max_nr_stripes
) {
5831 if (drop_one_stripe(conf
, hash
))
5832 conf
->max_nr_stripes
--;
5837 hash
= NR_STRIPE_HASH_LOCKS
- 1;
5839 err
= md_allow_write(mddev
);
5842 hash
= conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
5843 while (size
> conf
->max_nr_stripes
) {
5844 if (grow_one_stripe(conf
, hash
))
5845 conf
->max_nr_stripes
++;
5847 hash
= (hash
+ 1) % NR_STRIPE_HASH_LOCKS
;
5851 EXPORT_SYMBOL(raid5_set_cache_size
);
5854 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
5856 struct r5conf
*conf
;
5860 if (len
>= PAGE_SIZE
)
5862 if (kstrtoul(page
, 10, &new))
5864 err
= mddev_lock(mddev
);
5867 conf
= mddev
->private;
5871 err
= raid5_set_cache_size(mddev
, new);
5872 mddev_unlock(mddev
);
5877 static struct md_sysfs_entry
5878 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
5879 raid5_show_stripe_cache_size
,
5880 raid5_store_stripe_cache_size
);
5883 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
5885 struct r5conf
*conf
;
5887 spin_lock(&mddev
->lock
);
5888 conf
= mddev
->private;
5890 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
5891 spin_unlock(&mddev
->lock
);
5896 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
5898 struct r5conf
*conf
;
5902 if (len
>= PAGE_SIZE
)
5904 if (kstrtoul(page
, 10, &new))
5907 err
= mddev_lock(mddev
);
5910 conf
= mddev
->private;
5913 else if (new > conf
->max_nr_stripes
)
5916 conf
->bypass_threshold
= new;
5917 mddev_unlock(mddev
);
5921 static struct md_sysfs_entry
5922 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
5924 raid5_show_preread_threshold
,
5925 raid5_store_preread_threshold
);
5928 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
5930 struct r5conf
*conf
;
5932 spin_lock(&mddev
->lock
);
5933 conf
= mddev
->private;
5935 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
5936 spin_unlock(&mddev
->lock
);
5941 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
5943 struct r5conf
*conf
;
5947 if (len
>= PAGE_SIZE
)
5949 if (kstrtoul(page
, 10, &new))
5953 err
= mddev_lock(mddev
);
5956 conf
= mddev
->private;
5959 else if (new != conf
->skip_copy
) {
5960 mddev_suspend(mddev
);
5961 conf
->skip_copy
= new;
5963 mddev
->queue
->backing_dev_info
.capabilities
|=
5964 BDI_CAP_STABLE_WRITES
;
5966 mddev
->queue
->backing_dev_info
.capabilities
&=
5967 ~BDI_CAP_STABLE_WRITES
;
5968 mddev_resume(mddev
);
5970 mddev_unlock(mddev
);
5974 static struct md_sysfs_entry
5975 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
5976 raid5_show_skip_copy
,
5977 raid5_store_skip_copy
);
5980 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
5982 struct r5conf
*conf
= mddev
->private;
5984 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
5989 static struct md_sysfs_entry
5990 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
5993 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
5995 struct r5conf
*conf
;
5997 spin_lock(&mddev
->lock
);
5998 conf
= mddev
->private;
6000 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6001 spin_unlock(&mddev
->lock
);
6005 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6007 int *worker_cnt_per_group
,
6008 struct r5worker_group
**worker_groups
);
6010 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6012 struct r5conf
*conf
;
6015 struct r5worker_group
*new_groups
, *old_groups
;
6016 int group_cnt
, worker_cnt_per_group
;
6018 if (len
>= PAGE_SIZE
)
6020 if (kstrtoul(page
, 10, &new))
6023 err
= mddev_lock(mddev
);
6026 conf
= mddev
->private;
6029 else if (new != conf
->worker_cnt_per_group
) {
6030 mddev_suspend(mddev
);
6032 old_groups
= conf
->worker_groups
;
6034 flush_workqueue(raid5_wq
);
6036 err
= alloc_thread_groups(conf
, new,
6037 &group_cnt
, &worker_cnt_per_group
,
6040 spin_lock_irq(&conf
->device_lock
);
6041 conf
->group_cnt
= group_cnt
;
6042 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6043 conf
->worker_groups
= new_groups
;
6044 spin_unlock_irq(&conf
->device_lock
);
6047 kfree(old_groups
[0].workers
);
6050 mddev_resume(mddev
);
6052 mddev_unlock(mddev
);
6057 static struct md_sysfs_entry
6058 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6059 raid5_show_group_thread_cnt
,
6060 raid5_store_group_thread_cnt
);
6062 static struct attribute
*raid5_attrs
[] = {
6063 &raid5_stripecache_size
.attr
,
6064 &raid5_stripecache_active
.attr
,
6065 &raid5_preread_bypass_threshold
.attr
,
6066 &raid5_group_thread_cnt
.attr
,
6067 &raid5_skip_copy
.attr
,
6070 static struct attribute_group raid5_attrs_group
= {
6072 .attrs
= raid5_attrs
,
6075 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6077 int *worker_cnt_per_group
,
6078 struct r5worker_group
**worker_groups
)
6082 struct r5worker
*workers
;
6084 *worker_cnt_per_group
= cnt
;
6087 *worker_groups
= NULL
;
6090 *group_cnt
= num_possible_nodes();
6091 size
= sizeof(struct r5worker
) * cnt
;
6092 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6093 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6094 *group_cnt
, GFP_NOIO
);
6095 if (!*worker_groups
|| !workers
) {
6097 kfree(*worker_groups
);
6101 for (i
= 0; i
< *group_cnt
; i
++) {
6102 struct r5worker_group
*group
;
6104 group
= &(*worker_groups
)[i
];
6105 INIT_LIST_HEAD(&group
->handle_list
);
6107 group
->workers
= workers
+ i
* cnt
;
6109 for (j
= 0; j
< cnt
; j
++) {
6110 struct r5worker
*worker
= group
->workers
+ j
;
6111 worker
->group
= group
;
6112 INIT_WORK(&worker
->work
, raid5_do_work
);
6114 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6115 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6122 static void free_thread_groups(struct r5conf
*conf
)
6124 if (conf
->worker_groups
)
6125 kfree(conf
->worker_groups
[0].workers
);
6126 kfree(conf
->worker_groups
);
6127 conf
->worker_groups
= NULL
;
6131 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6133 struct r5conf
*conf
= mddev
->private;
6136 sectors
= mddev
->dev_sectors
;
6138 /* size is defined by the smallest of previous and new size */
6139 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6141 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
6142 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
6143 return sectors
* (raid_disks
- conf
->max_degraded
);
6146 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6148 safe_put_page(percpu
->spare_page
);
6149 if (percpu
->scribble
)
6150 flex_array_free(percpu
->scribble
);
6151 percpu
->spare_page
= NULL
;
6152 percpu
->scribble
= NULL
;
6155 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6157 if (conf
->level
== 6 && !percpu
->spare_page
)
6158 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6159 if (!percpu
->scribble
)
6160 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6161 conf
->previous_raid_disks
), conf
->chunk_sectors
/
6162 STRIPE_SECTORS
, GFP_KERNEL
);
6164 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6165 free_scratch_buffer(conf
, percpu
);
6172 static void raid5_free_percpu(struct r5conf
*conf
)
6179 #ifdef CONFIG_HOTPLUG_CPU
6180 unregister_cpu_notifier(&conf
->cpu_notify
);
6184 for_each_possible_cpu(cpu
)
6185 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6188 free_percpu(conf
->percpu
);
6191 static void free_conf(struct r5conf
*conf
)
6193 free_thread_groups(conf
);
6194 shrink_stripes(conf
);
6195 raid5_free_percpu(conf
);
6197 kfree(conf
->stripe_hashtbl
);
6201 #ifdef CONFIG_HOTPLUG_CPU
6202 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
6205 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
6206 long cpu
= (long)hcpu
;
6207 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6210 case CPU_UP_PREPARE
:
6211 case CPU_UP_PREPARE_FROZEN
:
6212 if (alloc_scratch_buffer(conf
, percpu
)) {
6213 pr_err("%s: failed memory allocation for cpu%ld\n",
6215 return notifier_from_errno(-ENOMEM
);
6219 case CPU_DEAD_FROZEN
:
6220 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6229 static int raid5_alloc_percpu(struct r5conf
*conf
)
6234 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6238 #ifdef CONFIG_HOTPLUG_CPU
6239 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
6240 conf
->cpu_notify
.priority
= 0;
6241 err
= register_cpu_notifier(&conf
->cpu_notify
);
6247 for_each_present_cpu(cpu
) {
6248 err
= alloc_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6250 pr_err("%s: failed memory allocation for cpu%ld\n",
6260 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6262 struct r5conf
*conf
;
6263 int raid_disk
, memory
, max_disks
;
6264 struct md_rdev
*rdev
;
6265 struct disk_info
*disk
;
6268 int group_cnt
, worker_cnt_per_group
;
6269 struct r5worker_group
*new_group
;
6271 if (mddev
->new_level
!= 5
6272 && mddev
->new_level
!= 4
6273 && mddev
->new_level
!= 6) {
6274 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6275 mdname(mddev
), mddev
->new_level
);
6276 return ERR_PTR(-EIO
);
6278 if ((mddev
->new_level
== 5
6279 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6280 (mddev
->new_level
== 6
6281 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6282 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
6283 mdname(mddev
), mddev
->new_layout
);
6284 return ERR_PTR(-EIO
);
6286 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6287 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6288 mdname(mddev
), mddev
->raid_disks
);
6289 return ERR_PTR(-EINVAL
);
6292 if (!mddev
->new_chunk_sectors
||
6293 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6294 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6295 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
6296 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6297 return ERR_PTR(-EINVAL
);
6300 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6303 /* Don't enable multi-threading by default*/
6304 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6306 conf
->group_cnt
= group_cnt
;
6307 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6308 conf
->worker_groups
= new_group
;
6311 spin_lock_init(&conf
->device_lock
);
6312 seqcount_init(&conf
->gen_lock
);
6313 init_waitqueue_head(&conf
->wait_for_stripe
);
6314 init_waitqueue_head(&conf
->wait_for_overlap
);
6315 INIT_LIST_HEAD(&conf
->handle_list
);
6316 INIT_LIST_HEAD(&conf
->hold_list
);
6317 INIT_LIST_HEAD(&conf
->delayed_list
);
6318 INIT_LIST_HEAD(&conf
->bitmap_list
);
6319 init_llist_head(&conf
->released_stripes
);
6320 atomic_set(&conf
->active_stripes
, 0);
6321 atomic_set(&conf
->preread_active_stripes
, 0);
6322 atomic_set(&conf
->active_aligned_reads
, 0);
6323 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6324 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6326 conf
->raid_disks
= mddev
->raid_disks
;
6327 if (mddev
->reshape_position
== MaxSector
)
6328 conf
->previous_raid_disks
= mddev
->raid_disks
;
6330 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6331 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6333 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6338 conf
->mddev
= mddev
;
6340 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6343 /* We init hash_locks[0] separately to that it can be used
6344 * as the reference lock in the spin_lock_nest_lock() call
6345 * in lock_all_device_hash_locks_irq in order to convince
6346 * lockdep that we know what we are doing.
6348 spin_lock_init(conf
->hash_locks
);
6349 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6350 spin_lock_init(conf
->hash_locks
+ i
);
6352 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6353 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6355 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6356 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6358 conf
->level
= mddev
->new_level
;
6359 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6360 if (raid5_alloc_percpu(conf
) != 0)
6363 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6365 rdev_for_each(rdev
, mddev
) {
6366 raid_disk
= rdev
->raid_disk
;
6367 if (raid_disk
>= max_disks
6370 disk
= conf
->disks
+ raid_disk
;
6372 if (test_bit(Replacement
, &rdev
->flags
)) {
6373 if (disk
->replacement
)
6375 disk
->replacement
= rdev
;
6382 if (test_bit(In_sync
, &rdev
->flags
)) {
6383 char b
[BDEVNAME_SIZE
];
6384 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
6386 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
6387 } else if (rdev
->saved_raid_disk
!= raid_disk
)
6388 /* Cannot rely on bitmap to complete recovery */
6392 conf
->level
= mddev
->new_level
;
6393 if (conf
->level
== 6) {
6394 conf
->max_degraded
= 2;
6395 if (raid6_call
.xor_syndrome
)
6396 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6398 conf
->rmw_level
= PARITY_DISABLE_RMW
;
6400 conf
->max_degraded
= 1;
6401 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6403 conf
->algorithm
= mddev
->new_layout
;
6404 conf
->reshape_progress
= mddev
->reshape_position
;
6405 if (conf
->reshape_progress
!= MaxSector
) {
6406 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
6407 conf
->prev_algo
= mddev
->layout
;
6410 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
6411 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
6412 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
6413 if (grow_stripes(conf
, NR_STRIPES
)) {
6415 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6416 mdname(mddev
), memory
);
6419 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
6420 mdname(mddev
), memory
);
6422 sprintf(pers_name
, "raid%d", mddev
->new_level
);
6423 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
6424 if (!conf
->thread
) {
6426 "md/raid:%s: couldn't allocate thread.\n",
6436 return ERR_PTR(-EIO
);
6438 return ERR_PTR(-ENOMEM
);
6441 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
6444 case ALGORITHM_PARITY_0
:
6445 if (raid_disk
< max_degraded
)
6448 case ALGORITHM_PARITY_N
:
6449 if (raid_disk
>= raid_disks
- max_degraded
)
6452 case ALGORITHM_PARITY_0_6
:
6453 if (raid_disk
== 0 ||
6454 raid_disk
== raid_disks
- 1)
6457 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6458 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6459 case ALGORITHM_LEFT_SYMMETRIC_6
:
6460 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6461 if (raid_disk
== raid_disks
- 1)
6467 static int run(struct mddev
*mddev
)
6469 struct r5conf
*conf
;
6470 int working_disks
= 0;
6471 int dirty_parity_disks
= 0;
6472 struct md_rdev
*rdev
;
6473 sector_t reshape_offset
= 0;
6475 long long min_offset_diff
= 0;
6478 if (mddev
->recovery_cp
!= MaxSector
)
6479 printk(KERN_NOTICE
"md/raid:%s: not clean"
6480 " -- starting background reconstruction\n",
6483 rdev_for_each(rdev
, mddev
) {
6485 if (rdev
->raid_disk
< 0)
6487 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
6489 min_offset_diff
= diff
;
6491 } else if (mddev
->reshape_backwards
&&
6492 diff
< min_offset_diff
)
6493 min_offset_diff
= diff
;
6494 else if (!mddev
->reshape_backwards
&&
6495 diff
> min_offset_diff
)
6496 min_offset_diff
= diff
;
6499 if (mddev
->reshape_position
!= MaxSector
) {
6500 /* Check that we can continue the reshape.
6501 * Difficulties arise if the stripe we would write to
6502 * next is at or after the stripe we would read from next.
6503 * For a reshape that changes the number of devices, this
6504 * is only possible for a very short time, and mdadm makes
6505 * sure that time appears to have past before assembling
6506 * the array. So we fail if that time hasn't passed.
6507 * For a reshape that keeps the number of devices the same
6508 * mdadm must be monitoring the reshape can keeping the
6509 * critical areas read-only and backed up. It will start
6510 * the array in read-only mode, so we check for that.
6512 sector_t here_new
, here_old
;
6514 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
6516 if (mddev
->new_level
!= mddev
->level
) {
6517 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
6518 "required - aborting.\n",
6522 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6523 /* reshape_position must be on a new-stripe boundary, and one
6524 * further up in new geometry must map after here in old
6527 here_new
= mddev
->reshape_position
;
6528 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
6529 (mddev
->raid_disks
- max_degraded
))) {
6530 printk(KERN_ERR
"md/raid:%s: reshape_position not "
6531 "on a stripe boundary\n", mdname(mddev
));
6534 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
6535 /* here_new is the stripe we will write to */
6536 here_old
= mddev
->reshape_position
;
6537 sector_div(here_old
, mddev
->chunk_sectors
*
6538 (old_disks
-max_degraded
));
6539 /* here_old is the first stripe that we might need to read
6541 if (mddev
->delta_disks
== 0) {
6542 if ((here_new
* mddev
->new_chunk_sectors
!=
6543 here_old
* mddev
->chunk_sectors
)) {
6544 printk(KERN_ERR
"md/raid:%s: reshape position is"
6545 " confused - aborting\n", mdname(mddev
));
6548 /* We cannot be sure it is safe to start an in-place
6549 * reshape. It is only safe if user-space is monitoring
6550 * and taking constant backups.
6551 * mdadm always starts a situation like this in
6552 * readonly mode so it can take control before
6553 * allowing any writes. So just check for that.
6555 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
6556 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
6557 /* not really in-place - so OK */;
6558 else if (mddev
->ro
== 0) {
6559 printk(KERN_ERR
"md/raid:%s: in-place reshape "
6560 "must be started in read-only mode "
6565 } else if (mddev
->reshape_backwards
6566 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
6567 here_old
* mddev
->chunk_sectors
)
6568 : (here_new
* mddev
->new_chunk_sectors
>=
6569 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
6570 /* Reading from the same stripe as writing to - bad */
6571 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
6572 "auto-recovery - aborting.\n",
6576 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
6578 /* OK, we should be able to continue; */
6580 BUG_ON(mddev
->level
!= mddev
->new_level
);
6581 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
6582 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
6583 BUG_ON(mddev
->delta_disks
!= 0);
6586 if (mddev
->private == NULL
)
6587 conf
= setup_conf(mddev
);
6589 conf
= mddev
->private;
6592 return PTR_ERR(conf
);
6594 conf
->min_offset_diff
= min_offset_diff
;
6595 mddev
->thread
= conf
->thread
;
6596 conf
->thread
= NULL
;
6597 mddev
->private = conf
;
6599 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
6601 rdev
= conf
->disks
[i
].rdev
;
6602 if (!rdev
&& conf
->disks
[i
].replacement
) {
6603 /* The replacement is all we have yet */
6604 rdev
= conf
->disks
[i
].replacement
;
6605 conf
->disks
[i
].replacement
= NULL
;
6606 clear_bit(Replacement
, &rdev
->flags
);
6607 conf
->disks
[i
].rdev
= rdev
;
6611 if (conf
->disks
[i
].replacement
&&
6612 conf
->reshape_progress
!= MaxSector
) {
6613 /* replacements and reshape simply do not mix. */
6614 printk(KERN_ERR
"md: cannot handle concurrent "
6615 "replacement and reshape.\n");
6618 if (test_bit(In_sync
, &rdev
->flags
)) {
6622 /* This disc is not fully in-sync. However if it
6623 * just stored parity (beyond the recovery_offset),
6624 * when we don't need to be concerned about the
6625 * array being dirty.
6626 * When reshape goes 'backwards', we never have
6627 * partially completed devices, so we only need
6628 * to worry about reshape going forwards.
6630 /* Hack because v0.91 doesn't store recovery_offset properly. */
6631 if (mddev
->major_version
== 0 &&
6632 mddev
->minor_version
> 90)
6633 rdev
->recovery_offset
= reshape_offset
;
6635 if (rdev
->recovery_offset
< reshape_offset
) {
6636 /* We need to check old and new layout */
6637 if (!only_parity(rdev
->raid_disk
,
6640 conf
->max_degraded
))
6643 if (!only_parity(rdev
->raid_disk
,
6645 conf
->previous_raid_disks
,
6646 conf
->max_degraded
))
6648 dirty_parity_disks
++;
6652 * 0 for a fully functional array, 1 or 2 for a degraded array.
6654 mddev
->degraded
= calc_degraded(conf
);
6656 if (has_failed(conf
)) {
6657 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
6658 " (%d/%d failed)\n",
6659 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
6663 /* device size must be a multiple of chunk size */
6664 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
6665 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
6667 if (mddev
->degraded
> dirty_parity_disks
&&
6668 mddev
->recovery_cp
!= MaxSector
) {
6669 if (mddev
->ok_start_degraded
)
6671 "md/raid:%s: starting dirty degraded array"
6672 " - data corruption possible.\n",
6676 "md/raid:%s: cannot start dirty degraded array.\n",
6682 if (mddev
->degraded
== 0)
6683 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
6684 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
6685 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
6688 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
6689 " out of %d devices, algorithm %d\n",
6690 mdname(mddev
), conf
->level
,
6691 mddev
->raid_disks
- mddev
->degraded
,
6692 mddev
->raid_disks
, mddev
->new_layout
);
6694 print_raid5_conf(conf
);
6696 if (conf
->reshape_progress
!= MaxSector
) {
6697 conf
->reshape_safe
= conf
->reshape_progress
;
6698 atomic_set(&conf
->reshape_stripes
, 0);
6699 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6700 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6701 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6702 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6703 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6707 /* Ok, everything is just fine now */
6708 if (mddev
->to_remove
== &raid5_attrs_group
)
6709 mddev
->to_remove
= NULL
;
6710 else if (mddev
->kobj
.sd
&&
6711 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
6713 "raid5: failed to create sysfs attributes for %s\n",
6715 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6719 bool discard_supported
= true;
6720 /* read-ahead size must cover two whole stripes, which
6721 * is 2 * (datadisks) * chunksize where 'n' is the
6722 * number of raid devices
6724 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
6725 int stripe
= data_disks
*
6726 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
6727 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6728 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6730 chunk_size
= mddev
->chunk_sectors
<< 9;
6731 blk_queue_io_min(mddev
->queue
, chunk_size
);
6732 blk_queue_io_opt(mddev
->queue
, chunk_size
*
6733 (conf
->raid_disks
- conf
->max_degraded
));
6734 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
6736 * We can only discard a whole stripe. It doesn't make sense to
6737 * discard data disk but write parity disk
6739 stripe
= stripe
* PAGE_SIZE
;
6740 /* Round up to power of 2, as discard handling
6741 * currently assumes that */
6742 while ((stripe
-1) & stripe
)
6743 stripe
= (stripe
| (stripe
-1)) + 1;
6744 mddev
->queue
->limits
.discard_alignment
= stripe
;
6745 mddev
->queue
->limits
.discard_granularity
= stripe
;
6747 * unaligned part of discard request will be ignored, so can't
6748 * guarantee discard_zeroes_data
6750 mddev
->queue
->limits
.discard_zeroes_data
= 0;
6752 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
6754 rdev_for_each(rdev
, mddev
) {
6755 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6756 rdev
->data_offset
<< 9);
6757 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6758 rdev
->new_data_offset
<< 9);
6760 * discard_zeroes_data is required, otherwise data
6761 * could be lost. Consider a scenario: discard a stripe
6762 * (the stripe could be inconsistent if
6763 * discard_zeroes_data is 0); write one disk of the
6764 * stripe (the stripe could be inconsistent again
6765 * depending on which disks are used to calculate
6766 * parity); the disk is broken; The stripe data of this
6769 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
6770 !bdev_get_queue(rdev
->bdev
)->
6771 limits
.discard_zeroes_data
)
6772 discard_supported
= false;
6773 /* Unfortunately, discard_zeroes_data is not currently
6774 * a guarantee - just a hint. So we only allow DISCARD
6775 * if the sysadmin has confirmed that only safe devices
6776 * are in use by setting a module parameter.
6778 if (!devices_handle_discard_safely
) {
6779 if (discard_supported
) {
6780 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6781 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6783 discard_supported
= false;
6787 if (discard_supported
&&
6788 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
6789 mddev
->queue
->limits
.discard_granularity
>= stripe
)
6790 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
6793 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
6799 md_unregister_thread(&mddev
->thread
);
6800 print_raid5_conf(conf
);
6802 mddev
->private = NULL
;
6803 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
6807 static void raid5_free(struct mddev
*mddev
, void *priv
)
6809 struct r5conf
*conf
= priv
;
6812 mddev
->to_remove
= &raid5_attrs_group
;
6815 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
6817 struct r5conf
*conf
= mddev
->private;
6820 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
6821 mddev
->chunk_sectors
/ 2, mddev
->layout
);
6822 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
6823 for (i
= 0; i
< conf
->raid_disks
; i
++)
6824 seq_printf (seq
, "%s",
6825 conf
->disks
[i
].rdev
&&
6826 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
6827 seq_printf (seq
, "]");
6830 static void print_raid5_conf (struct r5conf
*conf
)
6833 struct disk_info
*tmp
;
6835 printk(KERN_DEBUG
"RAID conf printout:\n");
6837 printk("(conf==NULL)\n");
6840 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
6842 conf
->raid_disks
- conf
->mddev
->degraded
);
6844 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6845 char b
[BDEVNAME_SIZE
];
6846 tmp
= conf
->disks
+ i
;
6848 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
6849 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
6850 bdevname(tmp
->rdev
->bdev
, b
));
6854 static int raid5_spare_active(struct mddev
*mddev
)
6857 struct r5conf
*conf
= mddev
->private;
6858 struct disk_info
*tmp
;
6860 unsigned long flags
;
6862 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6863 tmp
= conf
->disks
+ i
;
6864 if (tmp
->replacement
6865 && tmp
->replacement
->recovery_offset
== MaxSector
6866 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
6867 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
6868 /* Replacement has just become active. */
6870 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
6873 /* Replaced device not technically faulty,
6874 * but we need to be sure it gets removed
6875 * and never re-added.
6877 set_bit(Faulty
, &tmp
->rdev
->flags
);
6878 sysfs_notify_dirent_safe(
6879 tmp
->rdev
->sysfs_state
);
6881 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
6882 } else if (tmp
->rdev
6883 && tmp
->rdev
->recovery_offset
== MaxSector
6884 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
6885 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
6887 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
6890 spin_lock_irqsave(&conf
->device_lock
, flags
);
6891 mddev
->degraded
= calc_degraded(conf
);
6892 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
6893 print_raid5_conf(conf
);
6897 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
6899 struct r5conf
*conf
= mddev
->private;
6901 int number
= rdev
->raid_disk
;
6902 struct md_rdev
**rdevp
;
6903 struct disk_info
*p
= conf
->disks
+ number
;
6905 print_raid5_conf(conf
);
6906 if (rdev
== p
->rdev
)
6908 else if (rdev
== p
->replacement
)
6909 rdevp
= &p
->replacement
;
6913 if (number
>= conf
->raid_disks
&&
6914 conf
->reshape_progress
== MaxSector
)
6915 clear_bit(In_sync
, &rdev
->flags
);
6917 if (test_bit(In_sync
, &rdev
->flags
) ||
6918 atomic_read(&rdev
->nr_pending
)) {
6922 /* Only remove non-faulty devices if recovery
6925 if (!test_bit(Faulty
, &rdev
->flags
) &&
6926 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
6927 !has_failed(conf
) &&
6928 (!p
->replacement
|| p
->replacement
== rdev
) &&
6929 number
< conf
->raid_disks
) {
6935 if (atomic_read(&rdev
->nr_pending
)) {
6936 /* lost the race, try later */
6939 } else if (p
->replacement
) {
6940 /* We must have just cleared 'rdev' */
6941 p
->rdev
= p
->replacement
;
6942 clear_bit(Replacement
, &p
->replacement
->flags
);
6943 smp_mb(); /* Make sure other CPUs may see both as identical
6944 * but will never see neither - if they are careful
6946 p
->replacement
= NULL
;
6947 clear_bit(WantReplacement
, &rdev
->flags
);
6949 /* We might have just removed the Replacement as faulty-
6950 * clear the bit just in case
6952 clear_bit(WantReplacement
, &rdev
->flags
);
6955 print_raid5_conf(conf
);
6959 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
6961 struct r5conf
*conf
= mddev
->private;
6964 struct disk_info
*p
;
6966 int last
= conf
->raid_disks
- 1;
6968 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
6971 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
6972 /* no point adding a device */
6975 if (rdev
->raid_disk
>= 0)
6976 first
= last
= rdev
->raid_disk
;
6979 * find the disk ... but prefer rdev->saved_raid_disk
6982 if (rdev
->saved_raid_disk
>= 0 &&
6983 rdev
->saved_raid_disk
>= first
&&
6984 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
6985 first
= rdev
->saved_raid_disk
;
6987 for (disk
= first
; disk
<= last
; disk
++) {
6988 p
= conf
->disks
+ disk
;
6989 if (p
->rdev
== NULL
) {
6990 clear_bit(In_sync
, &rdev
->flags
);
6991 rdev
->raid_disk
= disk
;
6993 if (rdev
->saved_raid_disk
!= disk
)
6995 rcu_assign_pointer(p
->rdev
, rdev
);
6999 for (disk
= first
; disk
<= last
; disk
++) {
7000 p
= conf
->disks
+ disk
;
7001 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7002 p
->replacement
== NULL
) {
7003 clear_bit(In_sync
, &rdev
->flags
);
7004 set_bit(Replacement
, &rdev
->flags
);
7005 rdev
->raid_disk
= disk
;
7008 rcu_assign_pointer(p
->replacement
, rdev
);
7013 print_raid5_conf(conf
);
7017 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7019 /* no resync is happening, and there is enough space
7020 * on all devices, so we can resize.
7021 * We need to make sure resync covers any new space.
7022 * If the array is shrinking we should possibly wait until
7023 * any io in the removed space completes, but it hardly seems
7027 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
7028 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7029 if (mddev
->external_size
&&
7030 mddev
->array_sectors
> newsize
)
7032 if (mddev
->bitmap
) {
7033 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7037 md_set_array_sectors(mddev
, newsize
);
7038 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7039 revalidate_disk(mddev
->gendisk
);
7040 if (sectors
> mddev
->dev_sectors
&&
7041 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7042 mddev
->recovery_cp
= mddev
->dev_sectors
;
7043 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7045 mddev
->dev_sectors
= sectors
;
7046 mddev
->resync_max_sectors
= sectors
;
7050 static int check_stripe_cache(struct mddev
*mddev
)
7052 /* Can only proceed if there are plenty of stripe_heads.
7053 * We need a minimum of one full stripe,, and for sensible progress
7054 * it is best to have about 4 times that.
7055 * If we require 4 times, then the default 256 4K stripe_heads will
7056 * allow for chunk sizes up to 256K, which is probably OK.
7057 * If the chunk size is greater, user-space should request more
7058 * stripe_heads first.
7060 struct r5conf
*conf
= mddev
->private;
7061 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7062 > conf
->max_nr_stripes
||
7063 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7064 > conf
->max_nr_stripes
) {
7065 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7067 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7074 static int check_reshape(struct mddev
*mddev
)
7076 struct r5conf
*conf
= mddev
->private;
7078 if (mddev
->delta_disks
== 0 &&
7079 mddev
->new_layout
== mddev
->layout
&&
7080 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7081 return 0; /* nothing to do */
7082 if (has_failed(conf
))
7084 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7085 /* We might be able to shrink, but the devices must
7086 * be made bigger first.
7087 * For raid6, 4 is the minimum size.
7088 * Otherwise 2 is the minimum
7091 if (mddev
->level
== 6)
7093 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7097 if (!check_stripe_cache(mddev
))
7100 return resize_stripes(conf
, (conf
->previous_raid_disks
7101 + mddev
->delta_disks
));
7104 static int raid5_start_reshape(struct mddev
*mddev
)
7106 struct r5conf
*conf
= mddev
->private;
7107 struct md_rdev
*rdev
;
7109 unsigned long flags
;
7111 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7114 if (!check_stripe_cache(mddev
))
7117 if (has_failed(conf
))
7120 rdev_for_each(rdev
, mddev
) {
7121 if (!test_bit(In_sync
, &rdev
->flags
)
7122 && !test_bit(Faulty
, &rdev
->flags
))
7126 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7127 /* Not enough devices even to make a degraded array
7132 /* Refuse to reduce size of the array. Any reductions in
7133 * array size must be through explicit setting of array_size
7136 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7137 < mddev
->array_sectors
) {
7138 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
7139 "before number of disks\n", mdname(mddev
));
7143 atomic_set(&conf
->reshape_stripes
, 0);
7144 spin_lock_irq(&conf
->device_lock
);
7145 write_seqcount_begin(&conf
->gen_lock
);
7146 conf
->previous_raid_disks
= conf
->raid_disks
;
7147 conf
->raid_disks
+= mddev
->delta_disks
;
7148 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7149 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7150 conf
->prev_algo
= conf
->algorithm
;
7151 conf
->algorithm
= mddev
->new_layout
;
7153 /* Code that selects data_offset needs to see the generation update
7154 * if reshape_progress has been set - so a memory barrier needed.
7157 if (mddev
->reshape_backwards
)
7158 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7160 conf
->reshape_progress
= 0;
7161 conf
->reshape_safe
= conf
->reshape_progress
;
7162 write_seqcount_end(&conf
->gen_lock
);
7163 spin_unlock_irq(&conf
->device_lock
);
7165 /* Now make sure any requests that proceeded on the assumption
7166 * the reshape wasn't running - like Discard or Read - have
7169 mddev_suspend(mddev
);
7170 mddev_resume(mddev
);
7172 /* Add some new drives, as many as will fit.
7173 * We know there are enough to make the newly sized array work.
7174 * Don't add devices if we are reducing the number of
7175 * devices in the array. This is because it is not possible
7176 * to correctly record the "partially reconstructed" state of
7177 * such devices during the reshape and confusion could result.
7179 if (mddev
->delta_disks
>= 0) {
7180 rdev_for_each(rdev
, mddev
)
7181 if (rdev
->raid_disk
< 0 &&
7182 !test_bit(Faulty
, &rdev
->flags
)) {
7183 if (raid5_add_disk(mddev
, rdev
) == 0) {
7185 >= conf
->previous_raid_disks
)
7186 set_bit(In_sync
, &rdev
->flags
);
7188 rdev
->recovery_offset
= 0;
7190 if (sysfs_link_rdev(mddev
, rdev
))
7191 /* Failure here is OK */;
7193 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7194 && !test_bit(Faulty
, &rdev
->flags
)) {
7195 /* This is a spare that was manually added */
7196 set_bit(In_sync
, &rdev
->flags
);
7199 /* When a reshape changes the number of devices,
7200 * ->degraded is measured against the larger of the
7201 * pre and post number of devices.
7203 spin_lock_irqsave(&conf
->device_lock
, flags
);
7204 mddev
->degraded
= calc_degraded(conf
);
7205 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7207 mddev
->raid_disks
= conf
->raid_disks
;
7208 mddev
->reshape_position
= conf
->reshape_progress
;
7209 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7211 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7212 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7213 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7214 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7215 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7217 if (!mddev
->sync_thread
) {
7218 mddev
->recovery
= 0;
7219 spin_lock_irq(&conf
->device_lock
);
7220 write_seqcount_begin(&conf
->gen_lock
);
7221 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7222 mddev
->new_chunk_sectors
=
7223 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7224 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7225 rdev_for_each(rdev
, mddev
)
7226 rdev
->new_data_offset
= rdev
->data_offset
;
7228 conf
->generation
--;
7229 conf
->reshape_progress
= MaxSector
;
7230 mddev
->reshape_position
= MaxSector
;
7231 write_seqcount_end(&conf
->gen_lock
);
7232 spin_unlock_irq(&conf
->device_lock
);
7235 conf
->reshape_checkpoint
= jiffies
;
7236 md_wakeup_thread(mddev
->sync_thread
);
7237 md_new_event(mddev
);
7241 /* This is called from the reshape thread and should make any
7242 * changes needed in 'conf'
7244 static void end_reshape(struct r5conf
*conf
)
7247 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7248 struct md_rdev
*rdev
;
7250 spin_lock_irq(&conf
->device_lock
);
7251 conf
->previous_raid_disks
= conf
->raid_disks
;
7252 rdev_for_each(rdev
, conf
->mddev
)
7253 rdev
->data_offset
= rdev
->new_data_offset
;
7255 conf
->reshape_progress
= MaxSector
;
7256 spin_unlock_irq(&conf
->device_lock
);
7257 wake_up(&conf
->wait_for_overlap
);
7259 /* read-ahead size must cover two whole stripes, which is
7260 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7262 if (conf
->mddev
->queue
) {
7263 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7264 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7266 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
7267 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
7272 /* This is called from the raid5d thread with mddev_lock held.
7273 * It makes config changes to the device.
7275 static void raid5_finish_reshape(struct mddev
*mddev
)
7277 struct r5conf
*conf
= mddev
->private;
7279 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7281 if (mddev
->delta_disks
> 0) {
7282 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7283 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7284 revalidate_disk(mddev
->gendisk
);
7287 spin_lock_irq(&conf
->device_lock
);
7288 mddev
->degraded
= calc_degraded(conf
);
7289 spin_unlock_irq(&conf
->device_lock
);
7290 for (d
= conf
->raid_disks
;
7291 d
< conf
->raid_disks
- mddev
->delta_disks
;
7293 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
7295 clear_bit(In_sync
, &rdev
->flags
);
7296 rdev
= conf
->disks
[d
].replacement
;
7298 clear_bit(In_sync
, &rdev
->flags
);
7301 mddev
->layout
= conf
->algorithm
;
7302 mddev
->chunk_sectors
= conf
->chunk_sectors
;
7303 mddev
->reshape_position
= MaxSector
;
7304 mddev
->delta_disks
= 0;
7305 mddev
->reshape_backwards
= 0;
7309 static void raid5_quiesce(struct mddev
*mddev
, int state
)
7311 struct r5conf
*conf
= mddev
->private;
7314 case 2: /* resume for a suspend */
7315 wake_up(&conf
->wait_for_overlap
);
7318 case 1: /* stop all writes */
7319 lock_all_device_hash_locks_irq(conf
);
7320 /* '2' tells resync/reshape to pause so that all
7321 * active stripes can drain
7324 wait_event_cmd(conf
->wait_for_stripe
,
7325 atomic_read(&conf
->active_stripes
) == 0 &&
7326 atomic_read(&conf
->active_aligned_reads
) == 0,
7327 unlock_all_device_hash_locks_irq(conf
),
7328 lock_all_device_hash_locks_irq(conf
));
7330 unlock_all_device_hash_locks_irq(conf
);
7331 /* allow reshape to continue */
7332 wake_up(&conf
->wait_for_overlap
);
7335 case 0: /* re-enable writes */
7336 lock_all_device_hash_locks_irq(conf
);
7338 wake_up(&conf
->wait_for_stripe
);
7339 wake_up(&conf
->wait_for_overlap
);
7340 unlock_all_device_hash_locks_irq(conf
);
7345 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
7347 struct r0conf
*raid0_conf
= mddev
->private;
7350 /* for raid0 takeover only one zone is supported */
7351 if (raid0_conf
->nr_strip_zones
> 1) {
7352 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7354 return ERR_PTR(-EINVAL
);
7357 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
7358 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
7359 mddev
->dev_sectors
= sectors
;
7360 mddev
->new_level
= level
;
7361 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7362 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
7363 mddev
->raid_disks
+= 1;
7364 mddev
->delta_disks
= 1;
7365 /* make sure it will be not marked as dirty */
7366 mddev
->recovery_cp
= MaxSector
;
7368 return setup_conf(mddev
);
7371 static void *raid5_takeover_raid1(struct mddev
*mddev
)
7375 if (mddev
->raid_disks
!= 2 ||
7376 mddev
->degraded
> 1)
7377 return ERR_PTR(-EINVAL
);
7379 /* Should check if there are write-behind devices? */
7381 chunksect
= 64*2; /* 64K by default */
7383 /* The array must be an exact multiple of chunksize */
7384 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
7387 if ((chunksect
<<9) < STRIPE_SIZE
)
7388 /* array size does not allow a suitable chunk size */
7389 return ERR_PTR(-EINVAL
);
7391 mddev
->new_level
= 5;
7392 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7393 mddev
->new_chunk_sectors
= chunksect
;
7395 return setup_conf(mddev
);
7398 static void *raid5_takeover_raid6(struct mddev
*mddev
)
7402 switch (mddev
->layout
) {
7403 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7404 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
7406 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7407 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
7409 case ALGORITHM_LEFT_SYMMETRIC_6
:
7410 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7412 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7413 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
7415 case ALGORITHM_PARITY_0_6
:
7416 new_layout
= ALGORITHM_PARITY_0
;
7418 case ALGORITHM_PARITY_N
:
7419 new_layout
= ALGORITHM_PARITY_N
;
7422 return ERR_PTR(-EINVAL
);
7424 mddev
->new_level
= 5;
7425 mddev
->new_layout
= new_layout
;
7426 mddev
->delta_disks
= -1;
7427 mddev
->raid_disks
-= 1;
7428 return setup_conf(mddev
);
7431 static int raid5_check_reshape(struct mddev
*mddev
)
7433 /* For a 2-drive array, the layout and chunk size can be changed
7434 * immediately as not restriping is needed.
7435 * For larger arrays we record the new value - after validation
7436 * to be used by a reshape pass.
7438 struct r5conf
*conf
= mddev
->private;
7439 int new_chunk
= mddev
->new_chunk_sectors
;
7441 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
7443 if (new_chunk
> 0) {
7444 if (!is_power_of_2(new_chunk
))
7446 if (new_chunk
< (PAGE_SIZE
>>9))
7448 if (mddev
->array_sectors
& (new_chunk
-1))
7449 /* not factor of array size */
7453 /* They look valid */
7455 if (mddev
->raid_disks
== 2) {
7456 /* can make the change immediately */
7457 if (mddev
->new_layout
>= 0) {
7458 conf
->algorithm
= mddev
->new_layout
;
7459 mddev
->layout
= mddev
->new_layout
;
7461 if (new_chunk
> 0) {
7462 conf
->chunk_sectors
= new_chunk
;
7463 mddev
->chunk_sectors
= new_chunk
;
7465 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7466 md_wakeup_thread(mddev
->thread
);
7468 return check_reshape(mddev
);
7471 static int raid6_check_reshape(struct mddev
*mddev
)
7473 int new_chunk
= mddev
->new_chunk_sectors
;
7475 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
7477 if (new_chunk
> 0) {
7478 if (!is_power_of_2(new_chunk
))
7480 if (new_chunk
< (PAGE_SIZE
>> 9))
7482 if (mddev
->array_sectors
& (new_chunk
-1))
7483 /* not factor of array size */
7487 /* They look valid */
7488 return check_reshape(mddev
);
7491 static void *raid5_takeover(struct mddev
*mddev
)
7493 /* raid5 can take over:
7494 * raid0 - if there is only one strip zone - make it a raid4 layout
7495 * raid1 - if there are two drives. We need to know the chunk size
7496 * raid4 - trivial - just use a raid4 layout.
7497 * raid6 - Providing it is a *_6 layout
7499 if (mddev
->level
== 0)
7500 return raid45_takeover_raid0(mddev
, 5);
7501 if (mddev
->level
== 1)
7502 return raid5_takeover_raid1(mddev
);
7503 if (mddev
->level
== 4) {
7504 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7505 mddev
->new_level
= 5;
7506 return setup_conf(mddev
);
7508 if (mddev
->level
== 6)
7509 return raid5_takeover_raid6(mddev
);
7511 return ERR_PTR(-EINVAL
);
7514 static void *raid4_takeover(struct mddev
*mddev
)
7516 /* raid4 can take over:
7517 * raid0 - if there is only one strip zone
7518 * raid5 - if layout is right
7520 if (mddev
->level
== 0)
7521 return raid45_takeover_raid0(mddev
, 4);
7522 if (mddev
->level
== 5 &&
7523 mddev
->layout
== ALGORITHM_PARITY_N
) {
7524 mddev
->new_layout
= 0;
7525 mddev
->new_level
= 4;
7526 return setup_conf(mddev
);
7528 return ERR_PTR(-EINVAL
);
7531 static struct md_personality raid5_personality
;
7533 static void *raid6_takeover(struct mddev
*mddev
)
7535 /* Currently can only take over a raid5. We map the
7536 * personality to an equivalent raid6 personality
7537 * with the Q block at the end.
7541 if (mddev
->pers
!= &raid5_personality
)
7542 return ERR_PTR(-EINVAL
);
7543 if (mddev
->degraded
> 1)
7544 return ERR_PTR(-EINVAL
);
7545 if (mddev
->raid_disks
> 253)
7546 return ERR_PTR(-EINVAL
);
7547 if (mddev
->raid_disks
< 3)
7548 return ERR_PTR(-EINVAL
);
7550 switch (mddev
->layout
) {
7551 case ALGORITHM_LEFT_ASYMMETRIC
:
7552 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
7554 case ALGORITHM_RIGHT_ASYMMETRIC
:
7555 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
7557 case ALGORITHM_LEFT_SYMMETRIC
:
7558 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
7560 case ALGORITHM_RIGHT_SYMMETRIC
:
7561 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
7563 case ALGORITHM_PARITY_0
:
7564 new_layout
= ALGORITHM_PARITY_0_6
;
7566 case ALGORITHM_PARITY_N
:
7567 new_layout
= ALGORITHM_PARITY_N
;
7570 return ERR_PTR(-EINVAL
);
7572 mddev
->new_level
= 6;
7573 mddev
->new_layout
= new_layout
;
7574 mddev
->delta_disks
= 1;
7575 mddev
->raid_disks
+= 1;
7576 return setup_conf(mddev
);
7579 static struct md_personality raid6_personality
=
7583 .owner
= THIS_MODULE
,
7584 .make_request
= make_request
,
7588 .error_handler
= error
,
7589 .hot_add_disk
= raid5_add_disk
,
7590 .hot_remove_disk
= raid5_remove_disk
,
7591 .spare_active
= raid5_spare_active
,
7592 .sync_request
= sync_request
,
7593 .resize
= raid5_resize
,
7595 .check_reshape
= raid6_check_reshape
,
7596 .start_reshape
= raid5_start_reshape
,
7597 .finish_reshape
= raid5_finish_reshape
,
7598 .quiesce
= raid5_quiesce
,
7599 .takeover
= raid6_takeover
,
7600 .congested
= raid5_congested
,
7601 .mergeable_bvec
= raid5_mergeable_bvec
,
7603 static struct md_personality raid5_personality
=
7607 .owner
= THIS_MODULE
,
7608 .make_request
= make_request
,
7612 .error_handler
= error
,
7613 .hot_add_disk
= raid5_add_disk
,
7614 .hot_remove_disk
= raid5_remove_disk
,
7615 .spare_active
= raid5_spare_active
,
7616 .sync_request
= sync_request
,
7617 .resize
= raid5_resize
,
7619 .check_reshape
= raid5_check_reshape
,
7620 .start_reshape
= raid5_start_reshape
,
7621 .finish_reshape
= raid5_finish_reshape
,
7622 .quiesce
= raid5_quiesce
,
7623 .takeover
= raid5_takeover
,
7624 .congested
= raid5_congested
,
7625 .mergeable_bvec
= raid5_mergeable_bvec
,
7628 static struct md_personality raid4_personality
=
7632 .owner
= THIS_MODULE
,
7633 .make_request
= make_request
,
7637 .error_handler
= error
,
7638 .hot_add_disk
= raid5_add_disk
,
7639 .hot_remove_disk
= raid5_remove_disk
,
7640 .spare_active
= raid5_spare_active
,
7641 .sync_request
= sync_request
,
7642 .resize
= raid5_resize
,
7644 .check_reshape
= raid5_check_reshape
,
7645 .start_reshape
= raid5_start_reshape
,
7646 .finish_reshape
= raid5_finish_reshape
,
7647 .quiesce
= raid5_quiesce
,
7648 .takeover
= raid4_takeover
,
7649 .congested
= raid5_congested
,
7650 .mergeable_bvec
= raid5_mergeable_bvec
,
7653 static int __init
raid5_init(void)
7655 raid5_wq
= alloc_workqueue("raid5wq",
7656 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
7659 register_md_personality(&raid6_personality
);
7660 register_md_personality(&raid5_personality
);
7661 register_md_personality(&raid4_personality
);
7665 static void raid5_exit(void)
7667 unregister_md_personality(&raid6_personality
);
7668 unregister_md_personality(&raid5_personality
);
7669 unregister_md_personality(&raid4_personality
);
7670 destroy_workqueue(raid5_wq
);
7673 module_init(raid5_init
);
7674 module_exit(raid5_exit
);
7675 MODULE_LICENSE("GPL");
7676 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7677 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7678 MODULE_ALIAS("md-raid5");
7679 MODULE_ALIAS("md-raid4");
7680 MODULE_ALIAS("md-level-5");
7681 MODULE_ALIAS("md-level-4");
7682 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7683 MODULE_ALIAS("md-raid6");
7684 MODULE_ALIAS("md-level-6");
7686 /* This used to be two separate modules, they were: */
7687 MODULE_ALIAS("raid5");
7688 MODULE_ALIAS("raid6");