2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <trace/events/block.h>
65 #define cpu_to_group(cpu) cpu_to_node(cpu)
66 #define ANY_GROUP NUMA_NO_NODE
68 static bool devices_handle_discard_safely
= false;
69 module_param(devices_handle_discard_safely
, bool, 0644);
70 MODULE_PARM_DESC(devices_handle_discard_safely
,
71 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
72 static struct workqueue_struct
*raid5_wq
;
77 #define NR_STRIPES 256
78 #define STRIPE_SIZE PAGE_SIZE
79 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
80 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
81 #define IO_THRESHOLD 1
82 #define BYPASS_THRESHOLD 1
83 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
84 #define HASH_MASK (NR_HASH - 1)
85 #define MAX_STRIPE_BATCH 8
87 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
89 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
90 return &conf
->stripe_hashtbl
[hash
];
93 static inline int stripe_hash_locks_hash(sector_t sect
)
95 return (sect
>> STRIPE_SHIFT
) & STRIPE_HASH_LOCKS_MASK
;
98 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
100 spin_lock_irq(conf
->hash_locks
+ hash
);
101 spin_lock(&conf
->device_lock
);
104 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
106 spin_unlock(&conf
->device_lock
);
107 spin_unlock_irq(conf
->hash_locks
+ hash
);
110 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
114 spin_lock(conf
->hash_locks
);
115 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
116 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
117 spin_lock(&conf
->device_lock
);
120 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
123 spin_unlock(&conf
->device_lock
);
124 for (i
= NR_STRIPE_HASH_LOCKS
; i
; i
--)
125 spin_unlock(conf
->hash_locks
+ i
- 1);
129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
130 * order without overlap. There may be several bio's per stripe+device, and
131 * a bio could span several devices.
132 * When walking this list for a particular stripe+device, we must never proceed
133 * beyond a bio that extends past this device, as the next bio might no longer
135 * This function is used to determine the 'next' bio in the list, given the sector
136 * of the current stripe+device
138 static inline struct bio
*r5_next_bio(struct bio
*bio
, sector_t sector
)
140 int sectors
= bio_sectors(bio
);
141 if (bio
->bi_iter
.bi_sector
+ sectors
< sector
+ STRIPE_SECTORS
)
148 * We maintain a biased count of active stripes in the bottom 16 bits of
149 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
151 static inline int raid5_bi_processed_stripes(struct bio
*bio
)
153 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
154 return (atomic_read(segments
) >> 16) & 0xffff;
157 static inline int raid5_dec_bi_active_stripes(struct bio
*bio
)
159 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
160 return atomic_sub_return(1, segments
) & 0xffff;
163 static inline void raid5_inc_bi_active_stripes(struct bio
*bio
)
165 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
166 atomic_inc(segments
);
169 static inline void raid5_set_bi_processed_stripes(struct bio
*bio
,
172 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
176 old
= atomic_read(segments
);
177 new = (old
& 0xffff) | (cnt
<< 16);
178 } while (atomic_cmpxchg(segments
, old
, new) != old
);
181 static inline void raid5_set_bi_stripes(struct bio
*bio
, unsigned int cnt
)
183 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
184 atomic_set(segments
, cnt
);
187 /* Find first data disk in a raid6 stripe */
188 static inline int raid6_d0(struct stripe_head
*sh
)
191 /* ddf always start from first device */
193 /* md starts just after Q block */
194 if (sh
->qd_idx
== sh
->disks
- 1)
197 return sh
->qd_idx
+ 1;
199 static inline int raid6_next_disk(int disk
, int raid_disks
)
202 return (disk
< raid_disks
) ? disk
: 0;
205 /* When walking through the disks in a raid5, starting at raid6_d0,
206 * We need to map each disk to a 'slot', where the data disks are slot
207 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
208 * is raid_disks-1. This help does that mapping.
210 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
211 int *count
, int syndrome_disks
)
217 if (idx
== sh
->pd_idx
)
218 return syndrome_disks
;
219 if (idx
== sh
->qd_idx
)
220 return syndrome_disks
+ 1;
226 static void return_io(struct bio
*return_bi
)
228 struct bio
*bi
= return_bi
;
231 return_bi
= bi
->bi_next
;
233 bi
->bi_iter
.bi_size
= 0;
234 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
241 static void print_raid5_conf (struct r5conf
*conf
);
243 static int stripe_operations_active(struct stripe_head
*sh
)
245 return sh
->check_state
|| sh
->reconstruct_state
||
246 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
247 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
250 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
252 struct r5conf
*conf
= sh
->raid_conf
;
253 struct r5worker_group
*group
;
255 int i
, cpu
= sh
->cpu
;
257 if (!cpu_online(cpu
)) {
258 cpu
= cpumask_any(cpu_online_mask
);
262 if (list_empty(&sh
->lru
)) {
263 struct r5worker_group
*group
;
264 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
265 list_add_tail(&sh
->lru
, &group
->handle_list
);
266 group
->stripes_cnt
++;
270 if (conf
->worker_cnt_per_group
== 0) {
271 md_wakeup_thread(conf
->mddev
->thread
);
275 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
277 group
->workers
[0].working
= true;
278 /* at least one worker should run to avoid race */
279 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
281 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
282 /* wakeup more workers */
283 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
284 if (group
->workers
[i
].working
== false) {
285 group
->workers
[i
].working
= true;
286 queue_work_on(sh
->cpu
, raid5_wq
,
287 &group
->workers
[i
].work
);
293 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
294 struct list_head
*temp_inactive_list
)
296 BUG_ON(!list_empty(&sh
->lru
));
297 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
298 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
299 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
300 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
301 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
302 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
303 sh
->bm_seq
- conf
->seq_write
> 0)
304 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
306 clear_bit(STRIPE_DELAYED
, &sh
->state
);
307 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
308 if (conf
->worker_cnt_per_group
== 0) {
309 list_add_tail(&sh
->lru
, &conf
->handle_list
);
311 raid5_wakeup_stripe_thread(sh
);
315 md_wakeup_thread(conf
->mddev
->thread
);
317 BUG_ON(stripe_operations_active(sh
));
318 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
319 if (atomic_dec_return(&conf
->preread_active_stripes
)
321 md_wakeup_thread(conf
->mddev
->thread
);
322 atomic_dec(&conf
->active_stripes
);
323 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
))
324 list_add_tail(&sh
->lru
, temp_inactive_list
);
328 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
329 struct list_head
*temp_inactive_list
)
331 if (atomic_dec_and_test(&sh
->count
))
332 do_release_stripe(conf
, sh
, temp_inactive_list
);
336 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
338 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
339 * given time. Adding stripes only takes device lock, while deleting stripes
340 * only takes hash lock.
342 static void release_inactive_stripe_list(struct r5conf
*conf
,
343 struct list_head
*temp_inactive_list
,
347 bool do_wakeup
= false;
350 if (hash
== NR_STRIPE_HASH_LOCKS
) {
351 size
= NR_STRIPE_HASH_LOCKS
;
352 hash
= NR_STRIPE_HASH_LOCKS
- 1;
356 struct list_head
*list
= &temp_inactive_list
[size
- 1];
359 * We don't hold any lock here yet, get_active_stripe() might
360 * remove stripes from the list
362 if (!list_empty_careful(list
)) {
363 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
364 if (list_empty(conf
->inactive_list
+ hash
) &&
366 atomic_dec(&conf
->empty_inactive_list_nr
);
367 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
369 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
376 wake_up(&conf
->wait_for_stripe
);
377 if (conf
->retry_read_aligned
)
378 md_wakeup_thread(conf
->mddev
->thread
);
382 /* should hold conf->device_lock already */
383 static int release_stripe_list(struct r5conf
*conf
,
384 struct list_head
*temp_inactive_list
)
386 struct stripe_head
*sh
;
388 struct llist_node
*head
;
390 head
= llist_del_all(&conf
->released_stripes
);
391 head
= llist_reverse_order(head
);
395 sh
= llist_entry(head
, struct stripe_head
, release_list
);
396 head
= llist_next(head
);
397 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
399 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
401 * Don't worry the bit is set here, because if the bit is set
402 * again, the count is always > 1. This is true for
403 * STRIPE_ON_UNPLUG_LIST bit too.
405 hash
= sh
->hash_lock_index
;
406 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
413 static void release_stripe(struct stripe_head
*sh
)
415 struct r5conf
*conf
= sh
->raid_conf
;
417 struct list_head list
;
421 /* Avoid release_list until the last reference.
423 if (atomic_add_unless(&sh
->count
, -1, 1))
426 if (unlikely(!conf
->mddev
->thread
) ||
427 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
429 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
431 md_wakeup_thread(conf
->mddev
->thread
);
434 local_irq_save(flags
);
435 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
436 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
437 INIT_LIST_HEAD(&list
);
438 hash
= sh
->hash_lock_index
;
439 do_release_stripe(conf
, sh
, &list
);
440 spin_unlock(&conf
->device_lock
);
441 release_inactive_stripe_list(conf
, &list
, hash
);
443 local_irq_restore(flags
);
446 static inline void remove_hash(struct stripe_head
*sh
)
448 pr_debug("remove_hash(), stripe %llu\n",
449 (unsigned long long)sh
->sector
);
451 hlist_del_init(&sh
->hash
);
454 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
456 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
458 pr_debug("insert_hash(), stripe %llu\n",
459 (unsigned long long)sh
->sector
);
461 hlist_add_head(&sh
->hash
, hp
);
464 /* find an idle stripe, make sure it is unhashed, and return it. */
465 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
467 struct stripe_head
*sh
= NULL
;
468 struct list_head
*first
;
470 if (list_empty(conf
->inactive_list
+ hash
))
472 first
= (conf
->inactive_list
+ hash
)->next
;
473 sh
= list_entry(first
, struct stripe_head
, lru
);
474 list_del_init(first
);
476 atomic_inc(&conf
->active_stripes
);
477 BUG_ON(hash
!= sh
->hash_lock_index
);
478 if (list_empty(conf
->inactive_list
+ hash
))
479 atomic_inc(&conf
->empty_inactive_list_nr
);
484 static void shrink_buffers(struct stripe_head
*sh
)
488 int num
= sh
->raid_conf
->pool_size
;
490 for (i
= 0; i
< num
; i
++) {
491 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
495 sh
->dev
[i
].page
= NULL
;
500 static int grow_buffers(struct stripe_head
*sh
, gfp_t gfp
)
503 int num
= sh
->raid_conf
->pool_size
;
505 for (i
= 0; i
< num
; i
++) {
508 if (!(page
= alloc_page(gfp
))) {
511 sh
->dev
[i
].page
= page
;
512 sh
->dev
[i
].orig_page
= page
;
517 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
518 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
519 struct stripe_head
*sh
);
521 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
523 struct r5conf
*conf
= sh
->raid_conf
;
526 BUG_ON(atomic_read(&sh
->count
) != 0);
527 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
528 BUG_ON(stripe_operations_active(sh
));
529 BUG_ON(sh
->batch_head
);
531 pr_debug("init_stripe called, stripe %llu\n",
532 (unsigned long long)sector
);
534 seq
= read_seqcount_begin(&conf
->gen_lock
);
535 sh
->generation
= conf
->generation
- previous
;
536 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
538 stripe_set_idx(sector
, conf
, previous
, sh
);
541 for (i
= sh
->disks
; i
--; ) {
542 struct r5dev
*dev
= &sh
->dev
[i
];
544 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
545 test_bit(R5_LOCKED
, &dev
->flags
)) {
546 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
547 (unsigned long long)sh
->sector
, i
, dev
->toread
,
548 dev
->read
, dev
->towrite
, dev
->written
,
549 test_bit(R5_LOCKED
, &dev
->flags
));
553 raid5_build_block(sh
, i
, previous
);
555 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
557 sh
->overwrite_disks
= 0;
558 insert_hash(conf
, sh
);
559 sh
->cpu
= smp_processor_id();
560 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
563 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
566 struct stripe_head
*sh
;
568 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
569 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
570 if (sh
->sector
== sector
&& sh
->generation
== generation
)
572 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
577 * Need to check if array has failed when deciding whether to:
579 * - remove non-faulty devices
582 * This determination is simple when no reshape is happening.
583 * However if there is a reshape, we need to carefully check
584 * both the before and after sections.
585 * This is because some failed devices may only affect one
586 * of the two sections, and some non-in_sync devices may
587 * be insync in the section most affected by failed devices.
589 static int calc_degraded(struct r5conf
*conf
)
591 int degraded
, degraded2
;
596 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
597 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
598 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
599 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
600 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
602 else if (test_bit(In_sync
, &rdev
->flags
))
605 /* not in-sync or faulty.
606 * If the reshape increases the number of devices,
607 * this is being recovered by the reshape, so
608 * this 'previous' section is not in_sync.
609 * If the number of devices is being reduced however,
610 * the device can only be part of the array if
611 * we are reverting a reshape, so this section will
614 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
618 if (conf
->raid_disks
== conf
->previous_raid_disks
)
622 for (i
= 0; i
< conf
->raid_disks
; i
++) {
623 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
624 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
625 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
626 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
628 else if (test_bit(In_sync
, &rdev
->flags
))
631 /* not in-sync or faulty.
632 * If reshape increases the number of devices, this
633 * section has already been recovered, else it
634 * almost certainly hasn't.
636 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
640 if (degraded2
> degraded
)
645 static int has_failed(struct r5conf
*conf
)
649 if (conf
->mddev
->reshape_position
== MaxSector
)
650 return conf
->mddev
->degraded
> conf
->max_degraded
;
652 degraded
= calc_degraded(conf
);
653 if (degraded
> conf
->max_degraded
)
658 static struct stripe_head
*
659 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
660 int previous
, int noblock
, int noquiesce
)
662 struct stripe_head
*sh
;
663 int hash
= stripe_hash_locks_hash(sector
);
665 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
667 spin_lock_irq(conf
->hash_locks
+ hash
);
670 wait_event_lock_irq(conf
->wait_for_stripe
,
671 conf
->quiesce
== 0 || noquiesce
,
672 *(conf
->hash_locks
+ hash
));
673 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
675 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
676 sh
= get_free_stripe(conf
, hash
);
677 if (!sh
&& llist_empty(&conf
->released_stripes
) &&
678 !test_bit(R5_DID_ALLOC
, &conf
->cache_state
))
679 set_bit(R5_ALLOC_MORE
,
682 if (noblock
&& sh
== NULL
)
685 set_bit(R5_INACTIVE_BLOCKED
,
688 conf
->wait_for_stripe
,
689 !list_empty(conf
->inactive_list
+ hash
) &&
690 (atomic_read(&conf
->active_stripes
)
691 < (conf
->max_nr_stripes
* 3 / 4)
692 || !test_bit(R5_INACTIVE_BLOCKED
,
693 &conf
->cache_state
)),
694 *(conf
->hash_locks
+ hash
));
695 clear_bit(R5_INACTIVE_BLOCKED
,
698 init_stripe(sh
, sector
, previous
);
699 atomic_inc(&sh
->count
);
701 } else if (!atomic_inc_not_zero(&sh
->count
)) {
702 spin_lock(&conf
->device_lock
);
703 if (!atomic_read(&sh
->count
)) {
704 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
705 atomic_inc(&conf
->active_stripes
);
706 BUG_ON(list_empty(&sh
->lru
) &&
707 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
708 list_del_init(&sh
->lru
);
710 sh
->group
->stripes_cnt
--;
714 atomic_inc(&sh
->count
);
715 spin_unlock(&conf
->device_lock
);
717 } while (sh
== NULL
);
719 spin_unlock_irq(conf
->hash_locks
+ hash
);
723 static bool is_full_stripe_write(struct stripe_head
*sh
)
725 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
726 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
729 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
733 spin_lock(&sh2
->stripe_lock
);
734 spin_lock_nested(&sh1
->stripe_lock
, 1);
736 spin_lock(&sh1
->stripe_lock
);
737 spin_lock_nested(&sh2
->stripe_lock
, 1);
741 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
743 spin_unlock(&sh1
->stripe_lock
);
744 spin_unlock(&sh2
->stripe_lock
);
748 /* Only freshly new full stripe normal write stripe can be added to a batch list */
749 static bool stripe_can_batch(struct stripe_head
*sh
)
751 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
752 is_full_stripe_write(sh
);
755 /* we only do back search */
756 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
758 struct stripe_head
*head
;
759 sector_t head_sector
, tmp_sec
;
763 if (!stripe_can_batch(sh
))
765 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
766 tmp_sec
= sh
->sector
;
767 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
769 head_sector
= sh
->sector
- STRIPE_SECTORS
;
771 hash
= stripe_hash_locks_hash(head_sector
);
772 spin_lock_irq(conf
->hash_locks
+ hash
);
773 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
774 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
775 spin_lock(&conf
->device_lock
);
776 if (!atomic_read(&head
->count
)) {
777 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
778 atomic_inc(&conf
->active_stripes
);
779 BUG_ON(list_empty(&head
->lru
) &&
780 !test_bit(STRIPE_EXPANDING
, &head
->state
));
781 list_del_init(&head
->lru
);
783 head
->group
->stripes_cnt
--;
787 atomic_inc(&head
->count
);
788 spin_unlock(&conf
->device_lock
);
790 spin_unlock_irq(conf
->hash_locks
+ hash
);
794 if (!stripe_can_batch(head
))
797 lock_two_stripes(head
, sh
);
798 /* clear_batch_ready clear the flag */
799 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
806 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
808 if (head
->dev
[dd_idx
].towrite
->bi_rw
!= sh
->dev
[dd_idx
].towrite
->bi_rw
)
811 if (head
->batch_head
) {
812 spin_lock(&head
->batch_head
->batch_lock
);
813 /* This batch list is already running */
814 if (!stripe_can_batch(head
)) {
815 spin_unlock(&head
->batch_head
->batch_lock
);
820 * at this point, head's BATCH_READY could be cleared, but we
821 * can still add the stripe to batch list
823 list_add(&sh
->batch_list
, &head
->batch_list
);
824 spin_unlock(&head
->batch_head
->batch_lock
);
826 sh
->batch_head
= head
->batch_head
;
828 head
->batch_head
= head
;
829 sh
->batch_head
= head
->batch_head
;
830 spin_lock(&head
->batch_lock
);
831 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
832 spin_unlock(&head
->batch_lock
);
835 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
836 if (atomic_dec_return(&conf
->preread_active_stripes
)
838 md_wakeup_thread(conf
->mddev
->thread
);
840 atomic_inc(&sh
->count
);
842 unlock_two_stripes(head
, sh
);
844 release_stripe(head
);
847 /* Determine if 'data_offset' or 'new_data_offset' should be used
848 * in this stripe_head.
850 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
852 sector_t progress
= conf
->reshape_progress
;
853 /* Need a memory barrier to make sure we see the value
854 * of conf->generation, or ->data_offset that was set before
855 * reshape_progress was updated.
858 if (progress
== MaxSector
)
860 if (sh
->generation
== conf
->generation
- 1)
862 /* We are in a reshape, and this is a new-generation stripe,
863 * so use new_data_offset.
869 raid5_end_read_request(struct bio
*bi
, int error
);
871 raid5_end_write_request(struct bio
*bi
, int error
);
873 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
875 struct r5conf
*conf
= sh
->raid_conf
;
876 int i
, disks
= sh
->disks
;
877 struct stripe_head
*head_sh
= sh
;
881 for (i
= disks
; i
--; ) {
883 int replace_only
= 0;
884 struct bio
*bi
, *rbi
;
885 struct md_rdev
*rdev
, *rrdev
= NULL
;
888 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
889 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
893 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
895 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
897 else if (test_and_clear_bit(R5_WantReplace
,
898 &sh
->dev
[i
].flags
)) {
903 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
907 bi
= &sh
->dev
[i
].req
;
908 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
911 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
912 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
913 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
922 /* We raced and saw duplicates */
925 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
930 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
933 atomic_inc(&rdev
->nr_pending
);
934 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
937 atomic_inc(&rrdev
->nr_pending
);
940 /* We have already checked bad blocks for reads. Now
941 * need to check for writes. We never accept write errors
942 * on the replacement, so we don't to check rrdev.
944 while ((rw
& WRITE
) && rdev
&&
945 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
948 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
949 &first_bad
, &bad_sectors
);
954 set_bit(BlockedBadBlocks
, &rdev
->flags
);
955 if (!conf
->mddev
->external
&&
956 conf
->mddev
->flags
) {
957 /* It is very unlikely, but we might
958 * still need to write out the
959 * bad block log - better give it
961 md_check_recovery(conf
->mddev
);
964 * Because md_wait_for_blocked_rdev
965 * will dec nr_pending, we must
966 * increment it first.
968 atomic_inc(&rdev
->nr_pending
);
969 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
971 /* Acknowledged bad block - skip the write */
972 rdev_dec_pending(rdev
, conf
->mddev
);
978 if (s
->syncing
|| s
->expanding
|| s
->expanded
980 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
982 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
985 bi
->bi_bdev
= rdev
->bdev
;
987 bi
->bi_end_io
= (rw
& WRITE
)
988 ? raid5_end_write_request
989 : raid5_end_read_request
;
992 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
993 __func__
, (unsigned long long)sh
->sector
,
995 atomic_inc(&sh
->count
);
997 atomic_inc(&head_sh
->count
);
998 if (use_new_offset(conf
, sh
))
999 bi
->bi_iter
.bi_sector
= (sh
->sector
1000 + rdev
->new_data_offset
);
1002 bi
->bi_iter
.bi_sector
= (sh
->sector
1003 + rdev
->data_offset
);
1004 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1005 bi
->bi_rw
|= REQ_NOMERGE
;
1007 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1008 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1009 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1011 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1012 bi
->bi_io_vec
[0].bv_offset
= 0;
1013 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1015 * If this is discard request, set bi_vcnt 0. We don't
1016 * want to confuse SCSI because SCSI will replace payload
1018 if (rw
& REQ_DISCARD
)
1021 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1023 if (conf
->mddev
->gendisk
)
1024 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
1025 bi
, disk_devt(conf
->mddev
->gendisk
),
1027 generic_make_request(bi
);
1030 if (s
->syncing
|| s
->expanding
|| s
->expanded
1032 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1034 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1037 rbi
->bi_bdev
= rrdev
->bdev
;
1039 BUG_ON(!(rw
& WRITE
));
1040 rbi
->bi_end_io
= raid5_end_write_request
;
1041 rbi
->bi_private
= sh
;
1043 pr_debug("%s: for %llu schedule op %ld on "
1044 "replacement disc %d\n",
1045 __func__
, (unsigned long long)sh
->sector
,
1047 atomic_inc(&sh
->count
);
1049 atomic_inc(&head_sh
->count
);
1050 if (use_new_offset(conf
, sh
))
1051 rbi
->bi_iter
.bi_sector
= (sh
->sector
1052 + rrdev
->new_data_offset
);
1054 rbi
->bi_iter
.bi_sector
= (sh
->sector
1055 + rrdev
->data_offset
);
1056 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1057 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1058 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1060 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1061 rbi
->bi_io_vec
[0].bv_offset
= 0;
1062 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1064 * If this is discard request, set bi_vcnt 0. We don't
1065 * want to confuse SCSI because SCSI will replace payload
1067 if (rw
& REQ_DISCARD
)
1069 if (conf
->mddev
->gendisk
)
1070 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
1071 rbi
, disk_devt(conf
->mddev
->gendisk
),
1073 generic_make_request(rbi
);
1075 if (!rdev
&& !rrdev
) {
1077 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1078 pr_debug("skip op %ld on disc %d for sector %llu\n",
1079 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
1080 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1082 set_bit(STRIPE_BATCH_ERR
,
1083 &sh
->batch_head
->state
);
1084 set_bit(STRIPE_HANDLE
, &sh
->state
);
1087 if (!head_sh
->batch_head
)
1089 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1096 static struct dma_async_tx_descriptor
*
1097 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1098 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1099 struct stripe_head
*sh
)
1102 struct bvec_iter iter
;
1103 struct page
*bio_page
;
1105 struct async_submit_ctl submit
;
1106 enum async_tx_flags flags
= 0;
1108 if (bio
->bi_iter
.bi_sector
>= sector
)
1109 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1111 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1114 flags
|= ASYNC_TX_FENCE
;
1115 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1117 bio_for_each_segment(bvl
, bio
, iter
) {
1118 int len
= bvl
.bv_len
;
1122 if (page_offset
< 0) {
1123 b_offset
= -page_offset
;
1124 page_offset
+= b_offset
;
1128 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1129 clen
= STRIPE_SIZE
- page_offset
;
1134 b_offset
+= bvl
.bv_offset
;
1135 bio_page
= bvl
.bv_page
;
1137 if (sh
->raid_conf
->skip_copy
&&
1138 b_offset
== 0 && page_offset
== 0 &&
1139 clen
== STRIPE_SIZE
)
1142 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1143 b_offset
, clen
, &submit
);
1145 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1146 page_offset
, clen
, &submit
);
1148 /* chain the operations */
1149 submit
.depend_tx
= tx
;
1151 if (clen
< len
) /* hit end of page */
1159 static void ops_complete_biofill(void *stripe_head_ref
)
1161 struct stripe_head
*sh
= stripe_head_ref
;
1162 struct bio
*return_bi
= NULL
;
1165 pr_debug("%s: stripe %llu\n", __func__
,
1166 (unsigned long long)sh
->sector
);
1168 /* clear completed biofills */
1169 for (i
= sh
->disks
; i
--; ) {
1170 struct r5dev
*dev
= &sh
->dev
[i
];
1172 /* acknowledge completion of a biofill operation */
1173 /* and check if we need to reply to a read request,
1174 * new R5_Wantfill requests are held off until
1175 * !STRIPE_BIOFILL_RUN
1177 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1178 struct bio
*rbi
, *rbi2
;
1183 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1184 dev
->sector
+ STRIPE_SECTORS
) {
1185 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1186 if (!raid5_dec_bi_active_stripes(rbi
)) {
1187 rbi
->bi_next
= return_bi
;
1194 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1196 return_io(return_bi
);
1198 set_bit(STRIPE_HANDLE
, &sh
->state
);
1202 static void ops_run_biofill(struct stripe_head
*sh
)
1204 struct dma_async_tx_descriptor
*tx
= NULL
;
1205 struct async_submit_ctl submit
;
1208 BUG_ON(sh
->batch_head
);
1209 pr_debug("%s: stripe %llu\n", __func__
,
1210 (unsigned long long)sh
->sector
);
1212 for (i
= sh
->disks
; i
--; ) {
1213 struct r5dev
*dev
= &sh
->dev
[i
];
1214 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1216 spin_lock_irq(&sh
->stripe_lock
);
1217 dev
->read
= rbi
= dev
->toread
;
1219 spin_unlock_irq(&sh
->stripe_lock
);
1220 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1221 dev
->sector
+ STRIPE_SECTORS
) {
1222 tx
= async_copy_data(0, rbi
, &dev
->page
,
1223 dev
->sector
, tx
, sh
);
1224 rbi
= r5_next_bio(rbi
, dev
->sector
);
1229 atomic_inc(&sh
->count
);
1230 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1231 async_trigger_callback(&submit
);
1234 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1241 tgt
= &sh
->dev
[target
];
1242 set_bit(R5_UPTODATE
, &tgt
->flags
);
1243 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1244 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1247 static void ops_complete_compute(void *stripe_head_ref
)
1249 struct stripe_head
*sh
= stripe_head_ref
;
1251 pr_debug("%s: stripe %llu\n", __func__
,
1252 (unsigned long long)sh
->sector
);
1254 /* mark the computed target(s) as uptodate */
1255 mark_target_uptodate(sh
, sh
->ops
.target
);
1256 mark_target_uptodate(sh
, sh
->ops
.target2
);
1258 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1259 if (sh
->check_state
== check_state_compute_run
)
1260 sh
->check_state
= check_state_compute_result
;
1261 set_bit(STRIPE_HANDLE
, &sh
->state
);
1265 /* return a pointer to the address conversion region of the scribble buffer */
1266 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1267 struct raid5_percpu
*percpu
, int i
)
1271 addr
= flex_array_get(percpu
->scribble
, i
);
1272 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1275 /* return a pointer to the address conversion region of the scribble buffer */
1276 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1280 addr
= flex_array_get(percpu
->scribble
, i
);
1284 static struct dma_async_tx_descriptor
*
1285 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1287 int disks
= sh
->disks
;
1288 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1289 int target
= sh
->ops
.target
;
1290 struct r5dev
*tgt
= &sh
->dev
[target
];
1291 struct page
*xor_dest
= tgt
->page
;
1293 struct dma_async_tx_descriptor
*tx
;
1294 struct async_submit_ctl submit
;
1297 BUG_ON(sh
->batch_head
);
1299 pr_debug("%s: stripe %llu block: %d\n",
1300 __func__
, (unsigned long long)sh
->sector
, target
);
1301 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1303 for (i
= disks
; i
--; )
1305 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1307 atomic_inc(&sh
->count
);
1309 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1310 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1311 if (unlikely(count
== 1))
1312 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1314 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1319 /* set_syndrome_sources - populate source buffers for gen_syndrome
1320 * @srcs - (struct page *) array of size sh->disks
1321 * @sh - stripe_head to parse
1323 * Populates srcs in proper layout order for the stripe and returns the
1324 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1325 * destination buffer is recorded in srcs[count] and the Q destination
1326 * is recorded in srcs[count+1]].
1328 static int set_syndrome_sources(struct page
**srcs
,
1329 struct stripe_head
*sh
,
1332 int disks
= sh
->disks
;
1333 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1334 int d0_idx
= raid6_d0(sh
);
1338 for (i
= 0; i
< disks
; i
++)
1344 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1345 struct r5dev
*dev
= &sh
->dev
[i
];
1347 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1348 (srctype
== SYNDROME_SRC_ALL
) ||
1349 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1350 test_bit(R5_Wantdrain
, &dev
->flags
)) ||
1351 (srctype
== SYNDROME_SRC_WRITTEN
&&
1353 srcs
[slot
] = sh
->dev
[i
].page
;
1354 i
= raid6_next_disk(i
, disks
);
1355 } while (i
!= d0_idx
);
1357 return syndrome_disks
;
1360 static struct dma_async_tx_descriptor
*
1361 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1363 int disks
= sh
->disks
;
1364 struct page
**blocks
= to_addr_page(percpu
, 0);
1366 int qd_idx
= sh
->qd_idx
;
1367 struct dma_async_tx_descriptor
*tx
;
1368 struct async_submit_ctl submit
;
1374 BUG_ON(sh
->batch_head
);
1375 if (sh
->ops
.target
< 0)
1376 target
= sh
->ops
.target2
;
1377 else if (sh
->ops
.target2
< 0)
1378 target
= sh
->ops
.target
;
1380 /* we should only have one valid target */
1383 pr_debug("%s: stripe %llu block: %d\n",
1384 __func__
, (unsigned long long)sh
->sector
, target
);
1386 tgt
= &sh
->dev
[target
];
1387 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1390 atomic_inc(&sh
->count
);
1392 if (target
== qd_idx
) {
1393 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1394 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1395 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1396 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1397 ops_complete_compute
, sh
,
1398 to_addr_conv(sh
, percpu
, 0));
1399 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1401 /* Compute any data- or p-drive using XOR */
1403 for (i
= disks
; i
-- ; ) {
1404 if (i
== target
|| i
== qd_idx
)
1406 blocks
[count
++] = sh
->dev
[i
].page
;
1409 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1410 NULL
, ops_complete_compute
, sh
,
1411 to_addr_conv(sh
, percpu
, 0));
1412 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1418 static struct dma_async_tx_descriptor
*
1419 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1421 int i
, count
, disks
= sh
->disks
;
1422 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1423 int d0_idx
= raid6_d0(sh
);
1424 int faila
= -1, failb
= -1;
1425 int target
= sh
->ops
.target
;
1426 int target2
= sh
->ops
.target2
;
1427 struct r5dev
*tgt
= &sh
->dev
[target
];
1428 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1429 struct dma_async_tx_descriptor
*tx
;
1430 struct page
**blocks
= to_addr_page(percpu
, 0);
1431 struct async_submit_ctl submit
;
1433 BUG_ON(sh
->batch_head
);
1434 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1435 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1436 BUG_ON(target
< 0 || target2
< 0);
1437 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1438 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1440 /* we need to open-code set_syndrome_sources to handle the
1441 * slot number conversion for 'faila' and 'failb'
1443 for (i
= 0; i
< disks
; i
++)
1448 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1450 blocks
[slot
] = sh
->dev
[i
].page
;
1456 i
= raid6_next_disk(i
, disks
);
1457 } while (i
!= d0_idx
);
1459 BUG_ON(faila
== failb
);
1462 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1463 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1465 atomic_inc(&sh
->count
);
1467 if (failb
== syndrome_disks
+1) {
1468 /* Q disk is one of the missing disks */
1469 if (faila
== syndrome_disks
) {
1470 /* Missing P+Q, just recompute */
1471 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1472 ops_complete_compute
, sh
,
1473 to_addr_conv(sh
, percpu
, 0));
1474 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1475 STRIPE_SIZE
, &submit
);
1479 int qd_idx
= sh
->qd_idx
;
1481 /* Missing D+Q: recompute D from P, then recompute Q */
1482 if (target
== qd_idx
)
1483 data_target
= target2
;
1485 data_target
= target
;
1488 for (i
= disks
; i
-- ; ) {
1489 if (i
== data_target
|| i
== qd_idx
)
1491 blocks
[count
++] = sh
->dev
[i
].page
;
1493 dest
= sh
->dev
[data_target
].page
;
1494 init_async_submit(&submit
,
1495 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1497 to_addr_conv(sh
, percpu
, 0));
1498 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1501 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1502 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1503 ops_complete_compute
, sh
,
1504 to_addr_conv(sh
, percpu
, 0));
1505 return async_gen_syndrome(blocks
, 0, count
+2,
1506 STRIPE_SIZE
, &submit
);
1509 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1510 ops_complete_compute
, sh
,
1511 to_addr_conv(sh
, percpu
, 0));
1512 if (failb
== syndrome_disks
) {
1513 /* We're missing D+P. */
1514 return async_raid6_datap_recov(syndrome_disks
+2,
1518 /* We're missing D+D. */
1519 return async_raid6_2data_recov(syndrome_disks
+2,
1520 STRIPE_SIZE
, faila
, failb
,
1526 static void ops_complete_prexor(void *stripe_head_ref
)
1528 struct stripe_head
*sh
= stripe_head_ref
;
1530 pr_debug("%s: stripe %llu\n", __func__
,
1531 (unsigned long long)sh
->sector
);
1534 static struct dma_async_tx_descriptor
*
1535 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1536 struct dma_async_tx_descriptor
*tx
)
1538 int disks
= sh
->disks
;
1539 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1540 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1541 struct async_submit_ctl submit
;
1543 /* existing parity data subtracted */
1544 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1546 BUG_ON(sh
->batch_head
);
1547 pr_debug("%s: stripe %llu\n", __func__
,
1548 (unsigned long long)sh
->sector
);
1550 for (i
= disks
; i
--; ) {
1551 struct r5dev
*dev
= &sh
->dev
[i
];
1552 /* Only process blocks that are known to be uptodate */
1553 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1554 xor_srcs
[count
++] = dev
->page
;
1557 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1558 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1559 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1564 static struct dma_async_tx_descriptor
*
1565 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1566 struct dma_async_tx_descriptor
*tx
)
1568 struct page
**blocks
= to_addr_page(percpu
, 0);
1570 struct async_submit_ctl submit
;
1572 pr_debug("%s: stripe %llu\n", __func__
,
1573 (unsigned long long)sh
->sector
);
1575 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1577 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1578 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1579 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1584 static struct dma_async_tx_descriptor
*
1585 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1587 int disks
= sh
->disks
;
1589 struct stripe_head
*head_sh
= sh
;
1591 pr_debug("%s: stripe %llu\n", __func__
,
1592 (unsigned long long)sh
->sector
);
1594 for (i
= disks
; i
--; ) {
1599 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1604 spin_lock_irq(&sh
->stripe_lock
);
1605 chosen
= dev
->towrite
;
1606 dev
->towrite
= NULL
;
1607 sh
->overwrite_disks
= 0;
1608 BUG_ON(dev
->written
);
1609 wbi
= dev
->written
= chosen
;
1610 spin_unlock_irq(&sh
->stripe_lock
);
1611 WARN_ON(dev
->page
!= dev
->orig_page
);
1613 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1614 dev
->sector
+ STRIPE_SECTORS
) {
1615 if (wbi
->bi_rw
& REQ_FUA
)
1616 set_bit(R5_WantFUA
, &dev
->flags
);
1617 if (wbi
->bi_rw
& REQ_SYNC
)
1618 set_bit(R5_SyncIO
, &dev
->flags
);
1619 if (wbi
->bi_rw
& REQ_DISCARD
)
1620 set_bit(R5_Discard
, &dev
->flags
);
1622 tx
= async_copy_data(1, wbi
, &dev
->page
,
1623 dev
->sector
, tx
, sh
);
1624 if (dev
->page
!= dev
->orig_page
) {
1625 set_bit(R5_SkipCopy
, &dev
->flags
);
1626 clear_bit(R5_UPTODATE
, &dev
->flags
);
1627 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1630 wbi
= r5_next_bio(wbi
, dev
->sector
);
1633 if (head_sh
->batch_head
) {
1634 sh
= list_first_entry(&sh
->batch_list
,
1647 static void ops_complete_reconstruct(void *stripe_head_ref
)
1649 struct stripe_head
*sh
= stripe_head_ref
;
1650 int disks
= sh
->disks
;
1651 int pd_idx
= sh
->pd_idx
;
1652 int qd_idx
= sh
->qd_idx
;
1654 bool fua
= false, sync
= false, discard
= false;
1656 pr_debug("%s: stripe %llu\n", __func__
,
1657 (unsigned long long)sh
->sector
);
1659 for (i
= disks
; i
--; ) {
1660 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1661 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1662 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1665 for (i
= disks
; i
--; ) {
1666 struct r5dev
*dev
= &sh
->dev
[i
];
1668 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1669 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1670 set_bit(R5_UPTODATE
, &dev
->flags
);
1672 set_bit(R5_WantFUA
, &dev
->flags
);
1674 set_bit(R5_SyncIO
, &dev
->flags
);
1678 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1679 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1680 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1681 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1683 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1684 sh
->reconstruct_state
= reconstruct_state_result
;
1687 set_bit(STRIPE_HANDLE
, &sh
->state
);
1692 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1693 struct dma_async_tx_descriptor
*tx
)
1695 int disks
= sh
->disks
;
1696 struct page
**xor_srcs
;
1697 struct async_submit_ctl submit
;
1698 int count
, pd_idx
= sh
->pd_idx
, i
;
1699 struct page
*xor_dest
;
1701 unsigned long flags
;
1703 struct stripe_head
*head_sh
= sh
;
1706 pr_debug("%s: stripe %llu\n", __func__
,
1707 (unsigned long long)sh
->sector
);
1709 for (i
= 0; i
< sh
->disks
; i
++) {
1712 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1715 if (i
>= sh
->disks
) {
1716 atomic_inc(&sh
->count
);
1717 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1718 ops_complete_reconstruct(sh
);
1723 xor_srcs
= to_addr_page(percpu
, j
);
1724 /* check if prexor is active which means only process blocks
1725 * that are part of a read-modify-write (written)
1727 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1729 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1730 for (i
= disks
; i
--; ) {
1731 struct r5dev
*dev
= &sh
->dev
[i
];
1732 if (head_sh
->dev
[i
].written
)
1733 xor_srcs
[count
++] = dev
->page
;
1736 xor_dest
= sh
->dev
[pd_idx
].page
;
1737 for (i
= disks
; i
--; ) {
1738 struct r5dev
*dev
= &sh
->dev
[i
];
1740 xor_srcs
[count
++] = dev
->page
;
1744 /* 1/ if we prexor'd then the dest is reused as a source
1745 * 2/ if we did not prexor then we are redoing the parity
1746 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1747 * for the synchronous xor case
1749 last_stripe
= !head_sh
->batch_head
||
1750 list_first_entry(&sh
->batch_list
,
1751 struct stripe_head
, batch_list
) == head_sh
;
1753 flags
= ASYNC_TX_ACK
|
1754 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1756 atomic_inc(&head_sh
->count
);
1757 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1758 to_addr_conv(sh
, percpu
, j
));
1760 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1761 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1762 to_addr_conv(sh
, percpu
, j
));
1765 if (unlikely(count
== 1))
1766 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1768 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1771 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1778 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1779 struct dma_async_tx_descriptor
*tx
)
1781 struct async_submit_ctl submit
;
1782 struct page
**blocks
;
1783 int count
, i
, j
= 0;
1784 struct stripe_head
*head_sh
= sh
;
1787 unsigned long txflags
;
1789 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1791 for (i
= 0; i
< sh
->disks
; i
++) {
1792 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1794 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1797 if (i
>= sh
->disks
) {
1798 atomic_inc(&sh
->count
);
1799 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1800 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1801 ops_complete_reconstruct(sh
);
1806 blocks
= to_addr_page(percpu
, j
);
1808 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1809 synflags
= SYNDROME_SRC_WRITTEN
;
1810 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1812 synflags
= SYNDROME_SRC_ALL
;
1813 txflags
= ASYNC_TX_ACK
;
1816 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1817 last_stripe
= !head_sh
->batch_head
||
1818 list_first_entry(&sh
->batch_list
,
1819 struct stripe_head
, batch_list
) == head_sh
;
1822 atomic_inc(&head_sh
->count
);
1823 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1824 head_sh
, to_addr_conv(sh
, percpu
, j
));
1826 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1827 to_addr_conv(sh
, percpu
, j
));
1828 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1831 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1837 static void ops_complete_check(void *stripe_head_ref
)
1839 struct stripe_head
*sh
= stripe_head_ref
;
1841 pr_debug("%s: stripe %llu\n", __func__
,
1842 (unsigned long long)sh
->sector
);
1844 sh
->check_state
= check_state_check_result
;
1845 set_bit(STRIPE_HANDLE
, &sh
->state
);
1849 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1851 int disks
= sh
->disks
;
1852 int pd_idx
= sh
->pd_idx
;
1853 int qd_idx
= sh
->qd_idx
;
1854 struct page
*xor_dest
;
1855 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1856 struct dma_async_tx_descriptor
*tx
;
1857 struct async_submit_ctl submit
;
1861 pr_debug("%s: stripe %llu\n", __func__
,
1862 (unsigned long long)sh
->sector
);
1864 BUG_ON(sh
->batch_head
);
1866 xor_dest
= sh
->dev
[pd_idx
].page
;
1867 xor_srcs
[count
++] = xor_dest
;
1868 for (i
= disks
; i
--; ) {
1869 if (i
== pd_idx
|| i
== qd_idx
)
1871 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1874 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1875 to_addr_conv(sh
, percpu
, 0));
1876 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1877 &sh
->ops
.zero_sum_result
, &submit
);
1879 atomic_inc(&sh
->count
);
1880 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1881 tx
= async_trigger_callback(&submit
);
1884 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1886 struct page
**srcs
= to_addr_page(percpu
, 0);
1887 struct async_submit_ctl submit
;
1890 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1891 (unsigned long long)sh
->sector
, checkp
);
1893 BUG_ON(sh
->batch_head
);
1894 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
1898 atomic_inc(&sh
->count
);
1899 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1900 sh
, to_addr_conv(sh
, percpu
, 0));
1901 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1902 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1905 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1907 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1908 struct dma_async_tx_descriptor
*tx
= NULL
;
1909 struct r5conf
*conf
= sh
->raid_conf
;
1910 int level
= conf
->level
;
1911 struct raid5_percpu
*percpu
;
1915 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1916 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1917 ops_run_biofill(sh
);
1921 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1923 tx
= ops_run_compute5(sh
, percpu
);
1925 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1926 tx
= ops_run_compute6_1(sh
, percpu
);
1928 tx
= ops_run_compute6_2(sh
, percpu
);
1930 /* terminate the chain if reconstruct is not set to be run */
1931 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1935 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
1937 tx
= ops_run_prexor5(sh
, percpu
, tx
);
1939 tx
= ops_run_prexor6(sh
, percpu
, tx
);
1942 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1943 tx
= ops_run_biodrain(sh
, tx
);
1947 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1949 ops_run_reconstruct5(sh
, percpu
, tx
);
1951 ops_run_reconstruct6(sh
, percpu
, tx
);
1954 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1955 if (sh
->check_state
== check_state_run
)
1956 ops_run_check_p(sh
, percpu
);
1957 else if (sh
->check_state
== check_state_run_q
)
1958 ops_run_check_pq(sh
, percpu
, 0);
1959 else if (sh
->check_state
== check_state_run_pq
)
1960 ops_run_check_pq(sh
, percpu
, 1);
1965 if (overlap_clear
&& !sh
->batch_head
)
1966 for (i
= disks
; i
--; ) {
1967 struct r5dev
*dev
= &sh
->dev
[i
];
1968 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1969 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1974 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
)
1976 struct stripe_head
*sh
;
1978 sh
= kmem_cache_zalloc(sc
, gfp
);
1980 spin_lock_init(&sh
->stripe_lock
);
1981 spin_lock_init(&sh
->batch_lock
);
1982 INIT_LIST_HEAD(&sh
->batch_list
);
1983 INIT_LIST_HEAD(&sh
->lru
);
1984 atomic_set(&sh
->count
, 1);
1988 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
1990 struct stripe_head
*sh
;
1992 sh
= alloc_stripe(conf
->slab_cache
, gfp
);
1996 sh
->raid_conf
= conf
;
1998 if (grow_buffers(sh
, gfp
)) {
2000 kmem_cache_free(conf
->slab_cache
, sh
);
2003 sh
->hash_lock_index
=
2004 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2005 /* we just created an active stripe so... */
2006 atomic_inc(&conf
->active_stripes
);
2009 conf
->max_nr_stripes
++;
2013 static int grow_stripes(struct r5conf
*conf
, int num
)
2015 struct kmem_cache
*sc
;
2016 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2018 if (conf
->mddev
->gendisk
)
2019 sprintf(conf
->cache_name
[0],
2020 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2022 sprintf(conf
->cache_name
[0],
2023 "raid%d-%p", conf
->level
, conf
->mddev
);
2024 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2026 conf
->active_name
= 0;
2027 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2028 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2032 conf
->slab_cache
= sc
;
2033 conf
->pool_size
= devs
;
2035 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2042 * scribble_len - return the required size of the scribble region
2043 * @num - total number of disks in the array
2045 * The size must be enough to contain:
2046 * 1/ a struct page pointer for each device in the array +2
2047 * 2/ room to convert each entry in (1) to its corresponding dma
2048 * (dma_map_page()) or page (page_address()) address.
2050 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2051 * calculate over all devices (not just the data blocks), using zeros in place
2052 * of the P and Q blocks.
2054 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2056 struct flex_array
*ret
;
2059 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2060 ret
= flex_array_alloc(len
, cnt
, flags
);
2063 /* always prealloc all elements, so no locking is required */
2064 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2065 flex_array_free(ret
);
2071 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2073 /* Make all the stripes able to hold 'newsize' devices.
2074 * New slots in each stripe get 'page' set to a new page.
2076 * This happens in stages:
2077 * 1/ create a new kmem_cache and allocate the required number of
2079 * 2/ gather all the old stripe_heads and transfer the pages across
2080 * to the new stripe_heads. This will have the side effect of
2081 * freezing the array as once all stripe_heads have been collected,
2082 * no IO will be possible. Old stripe heads are freed once their
2083 * pages have been transferred over, and the old kmem_cache is
2084 * freed when all stripes are done.
2085 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2086 * we simple return a failre status - no need to clean anything up.
2087 * 4/ allocate new pages for the new slots in the new stripe_heads.
2088 * If this fails, we don't bother trying the shrink the
2089 * stripe_heads down again, we just leave them as they are.
2090 * As each stripe_head is processed the new one is released into
2093 * Once step2 is started, we cannot afford to wait for a write,
2094 * so we use GFP_NOIO allocations.
2096 struct stripe_head
*osh
, *nsh
;
2097 LIST_HEAD(newstripes
);
2098 struct disk_info
*ndisks
;
2101 struct kmem_cache
*sc
;
2105 if (newsize
<= conf
->pool_size
)
2106 return 0; /* never bother to shrink */
2108 err
= md_allow_write(conf
->mddev
);
2113 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2114 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2119 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2120 nsh
= alloc_stripe(sc
, GFP_KERNEL
);
2124 nsh
->raid_conf
= conf
;
2125 list_add(&nsh
->lru
, &newstripes
);
2128 /* didn't get enough, give up */
2129 while (!list_empty(&newstripes
)) {
2130 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2131 list_del(&nsh
->lru
);
2132 kmem_cache_free(sc
, nsh
);
2134 kmem_cache_destroy(sc
);
2137 /* Step 2 - Must use GFP_NOIO now.
2138 * OK, we have enough stripes, start collecting inactive
2139 * stripes and copying them over
2143 list_for_each_entry(nsh
, &newstripes
, lru
) {
2144 lock_device_hash_lock(conf
, hash
);
2145 wait_event_cmd(conf
->wait_for_stripe
,
2146 !list_empty(conf
->inactive_list
+ hash
),
2147 unlock_device_hash_lock(conf
, hash
),
2148 lock_device_hash_lock(conf
, hash
));
2149 osh
= get_free_stripe(conf
, hash
);
2150 unlock_device_hash_lock(conf
, hash
);
2152 for(i
=0; i
<conf
->pool_size
; i
++) {
2153 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2154 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2156 nsh
->hash_lock_index
= hash
;
2157 kmem_cache_free(conf
->slab_cache
, osh
);
2159 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2160 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2165 kmem_cache_destroy(conf
->slab_cache
);
2168 * At this point, we are holding all the stripes so the array
2169 * is completely stalled, so now is a good time to resize
2170 * conf->disks and the scribble region
2172 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2174 for (i
=0; i
<conf
->raid_disks
; i
++)
2175 ndisks
[i
] = conf
->disks
[i
];
2177 conf
->disks
= ndisks
;
2182 for_each_present_cpu(cpu
) {
2183 struct raid5_percpu
*percpu
;
2184 struct flex_array
*scribble
;
2186 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2187 scribble
= scribble_alloc(newsize
, conf
->chunk_sectors
/
2188 STRIPE_SECTORS
, GFP_NOIO
);
2191 flex_array_free(percpu
->scribble
);
2192 percpu
->scribble
= scribble
;
2200 /* Step 4, return new stripes to service */
2201 while(!list_empty(&newstripes
)) {
2202 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2203 list_del_init(&nsh
->lru
);
2205 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2206 if (nsh
->dev
[i
].page
== NULL
) {
2207 struct page
*p
= alloc_page(GFP_NOIO
);
2208 nsh
->dev
[i
].page
= p
;
2209 nsh
->dev
[i
].orig_page
= p
;
2213 release_stripe(nsh
);
2215 /* critical section pass, GFP_NOIO no longer needed */
2217 conf
->slab_cache
= sc
;
2218 conf
->active_name
= 1-conf
->active_name
;
2220 conf
->pool_size
= newsize
;
2224 static int drop_one_stripe(struct r5conf
*conf
)
2226 struct stripe_head
*sh
;
2227 int hash
= (conf
->max_nr_stripes
- 1) % NR_STRIPE_HASH_LOCKS
;
2229 spin_lock_irq(conf
->hash_locks
+ hash
);
2230 sh
= get_free_stripe(conf
, hash
);
2231 spin_unlock_irq(conf
->hash_locks
+ hash
);
2234 BUG_ON(atomic_read(&sh
->count
));
2236 kmem_cache_free(conf
->slab_cache
, sh
);
2237 atomic_dec(&conf
->active_stripes
);
2238 conf
->max_nr_stripes
--;
2242 static void shrink_stripes(struct r5conf
*conf
)
2244 while (conf
->max_nr_stripes
&&
2245 drop_one_stripe(conf
))
2248 if (conf
->slab_cache
)
2249 kmem_cache_destroy(conf
->slab_cache
);
2250 conf
->slab_cache
= NULL
;
2253 static void raid5_end_read_request(struct bio
* bi
, int error
)
2255 struct stripe_head
*sh
= bi
->bi_private
;
2256 struct r5conf
*conf
= sh
->raid_conf
;
2257 int disks
= sh
->disks
, i
;
2258 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2259 char b
[BDEVNAME_SIZE
];
2260 struct md_rdev
*rdev
= NULL
;
2263 for (i
=0 ; i
<disks
; i
++)
2264 if (bi
== &sh
->dev
[i
].req
)
2267 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2268 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2274 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2275 /* If replacement finished while this request was outstanding,
2276 * 'replacement' might be NULL already.
2277 * In that case it moved down to 'rdev'.
2278 * rdev is not removed until all requests are finished.
2280 rdev
= conf
->disks
[i
].replacement
;
2282 rdev
= conf
->disks
[i
].rdev
;
2284 if (use_new_offset(conf
, sh
))
2285 s
= sh
->sector
+ rdev
->new_data_offset
;
2287 s
= sh
->sector
+ rdev
->data_offset
;
2289 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2290 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2291 /* Note that this cannot happen on a
2292 * replacement device. We just fail those on
2297 "md/raid:%s: read error corrected"
2298 " (%lu sectors at %llu on %s)\n",
2299 mdname(conf
->mddev
), STRIPE_SECTORS
,
2300 (unsigned long long)s
,
2301 bdevname(rdev
->bdev
, b
));
2302 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2303 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2304 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2305 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2306 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2308 if (atomic_read(&rdev
->read_errors
))
2309 atomic_set(&rdev
->read_errors
, 0);
2311 const char *bdn
= bdevname(rdev
->bdev
, b
);
2315 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2316 atomic_inc(&rdev
->read_errors
);
2317 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2320 "md/raid:%s: read error on replacement device "
2321 "(sector %llu on %s).\n",
2322 mdname(conf
->mddev
),
2323 (unsigned long long)s
,
2325 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2329 "md/raid:%s: read error not correctable "
2330 "(sector %llu on %s).\n",
2331 mdname(conf
->mddev
),
2332 (unsigned long long)s
,
2334 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2339 "md/raid:%s: read error NOT corrected!! "
2340 "(sector %llu on %s).\n",
2341 mdname(conf
->mddev
),
2342 (unsigned long long)s
,
2344 } else if (atomic_read(&rdev
->read_errors
)
2345 > conf
->max_nr_stripes
)
2347 "md/raid:%s: Too many read errors, failing device %s.\n",
2348 mdname(conf
->mddev
), bdn
);
2351 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2352 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2355 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2356 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2357 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2359 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2361 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2362 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2364 && test_bit(In_sync
, &rdev
->flags
)
2365 && rdev_set_badblocks(
2366 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2367 md_error(conf
->mddev
, rdev
);
2370 rdev_dec_pending(rdev
, conf
->mddev
);
2371 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2372 set_bit(STRIPE_HANDLE
, &sh
->state
);
2376 static void raid5_end_write_request(struct bio
*bi
, int error
)
2378 struct stripe_head
*sh
= bi
->bi_private
;
2379 struct r5conf
*conf
= sh
->raid_conf
;
2380 int disks
= sh
->disks
, i
;
2381 struct md_rdev
*uninitialized_var(rdev
);
2382 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2385 int replacement
= 0;
2387 for (i
= 0 ; i
< disks
; i
++) {
2388 if (bi
== &sh
->dev
[i
].req
) {
2389 rdev
= conf
->disks
[i
].rdev
;
2392 if (bi
== &sh
->dev
[i
].rreq
) {
2393 rdev
= conf
->disks
[i
].replacement
;
2397 /* rdev was removed and 'replacement'
2398 * replaced it. rdev is not removed
2399 * until all requests are finished.
2401 rdev
= conf
->disks
[i
].rdev
;
2405 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2406 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2415 md_error(conf
->mddev
, rdev
);
2416 else if (is_badblock(rdev
, sh
->sector
,
2418 &first_bad
, &bad_sectors
))
2419 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2422 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2423 set_bit(WriteErrorSeen
, &rdev
->flags
);
2424 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2425 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2426 set_bit(MD_RECOVERY_NEEDED
,
2427 &rdev
->mddev
->recovery
);
2428 } else if (is_badblock(rdev
, sh
->sector
,
2430 &first_bad
, &bad_sectors
)) {
2431 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2432 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2433 /* That was a successful write so make
2434 * sure it looks like we already did
2437 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2440 rdev_dec_pending(rdev
, conf
->mddev
);
2442 if (sh
->batch_head
&& !uptodate
)
2443 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2445 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2446 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2447 set_bit(STRIPE_HANDLE
, &sh
->state
);
2450 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2451 release_stripe(sh
->batch_head
);
2454 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
2456 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
2458 struct r5dev
*dev
= &sh
->dev
[i
];
2460 bio_init(&dev
->req
);
2461 dev
->req
.bi_io_vec
= &dev
->vec
;
2462 dev
->req
.bi_max_vecs
= 1;
2463 dev
->req
.bi_private
= sh
;
2465 bio_init(&dev
->rreq
);
2466 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
2467 dev
->rreq
.bi_max_vecs
= 1;
2468 dev
->rreq
.bi_private
= sh
;
2471 dev
->sector
= compute_blocknr(sh
, i
, previous
);
2474 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2476 char b
[BDEVNAME_SIZE
];
2477 struct r5conf
*conf
= mddev
->private;
2478 unsigned long flags
;
2479 pr_debug("raid456: error called\n");
2481 spin_lock_irqsave(&conf
->device_lock
, flags
);
2482 clear_bit(In_sync
, &rdev
->flags
);
2483 mddev
->degraded
= calc_degraded(conf
);
2484 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2485 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2487 set_bit(Blocked
, &rdev
->flags
);
2488 set_bit(Faulty
, &rdev
->flags
);
2489 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2491 "md/raid:%s: Disk failure on %s, disabling device.\n"
2492 "md/raid:%s: Operation continuing on %d devices.\n",
2494 bdevname(rdev
->bdev
, b
),
2496 conf
->raid_disks
- mddev
->degraded
);
2500 * Input: a 'big' sector number,
2501 * Output: index of the data and parity disk, and the sector # in them.
2503 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2504 int previous
, int *dd_idx
,
2505 struct stripe_head
*sh
)
2507 sector_t stripe
, stripe2
;
2508 sector_t chunk_number
;
2509 unsigned int chunk_offset
;
2512 sector_t new_sector
;
2513 int algorithm
= previous
? conf
->prev_algo
2515 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2516 : conf
->chunk_sectors
;
2517 int raid_disks
= previous
? conf
->previous_raid_disks
2519 int data_disks
= raid_disks
- conf
->max_degraded
;
2521 /* First compute the information on this sector */
2524 * Compute the chunk number and the sector offset inside the chunk
2526 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2527 chunk_number
= r_sector
;
2530 * Compute the stripe number
2532 stripe
= chunk_number
;
2533 *dd_idx
= sector_div(stripe
, data_disks
);
2536 * Select the parity disk based on the user selected algorithm.
2538 pd_idx
= qd_idx
= -1;
2539 switch(conf
->level
) {
2541 pd_idx
= data_disks
;
2544 switch (algorithm
) {
2545 case ALGORITHM_LEFT_ASYMMETRIC
:
2546 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2547 if (*dd_idx
>= pd_idx
)
2550 case ALGORITHM_RIGHT_ASYMMETRIC
:
2551 pd_idx
= sector_div(stripe2
, raid_disks
);
2552 if (*dd_idx
>= pd_idx
)
2555 case ALGORITHM_LEFT_SYMMETRIC
:
2556 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2557 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2559 case ALGORITHM_RIGHT_SYMMETRIC
:
2560 pd_idx
= sector_div(stripe2
, raid_disks
);
2561 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2563 case ALGORITHM_PARITY_0
:
2567 case ALGORITHM_PARITY_N
:
2568 pd_idx
= data_disks
;
2576 switch (algorithm
) {
2577 case ALGORITHM_LEFT_ASYMMETRIC
:
2578 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2579 qd_idx
= pd_idx
+ 1;
2580 if (pd_idx
== raid_disks
-1) {
2581 (*dd_idx
)++; /* Q D D D P */
2583 } else if (*dd_idx
>= pd_idx
)
2584 (*dd_idx
) += 2; /* D D P Q D */
2586 case ALGORITHM_RIGHT_ASYMMETRIC
:
2587 pd_idx
= sector_div(stripe2
, raid_disks
);
2588 qd_idx
= pd_idx
+ 1;
2589 if (pd_idx
== raid_disks
-1) {
2590 (*dd_idx
)++; /* Q D D D P */
2592 } else if (*dd_idx
>= pd_idx
)
2593 (*dd_idx
) += 2; /* D D P Q D */
2595 case ALGORITHM_LEFT_SYMMETRIC
:
2596 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2597 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2598 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2600 case ALGORITHM_RIGHT_SYMMETRIC
:
2601 pd_idx
= sector_div(stripe2
, raid_disks
);
2602 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2603 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2606 case ALGORITHM_PARITY_0
:
2611 case ALGORITHM_PARITY_N
:
2612 pd_idx
= data_disks
;
2613 qd_idx
= data_disks
+ 1;
2616 case ALGORITHM_ROTATING_ZERO_RESTART
:
2617 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2618 * of blocks for computing Q is different.
2620 pd_idx
= sector_div(stripe2
, raid_disks
);
2621 qd_idx
= pd_idx
+ 1;
2622 if (pd_idx
== raid_disks
-1) {
2623 (*dd_idx
)++; /* Q D D D P */
2625 } else if (*dd_idx
>= pd_idx
)
2626 (*dd_idx
) += 2; /* D D P Q D */
2630 case ALGORITHM_ROTATING_N_RESTART
:
2631 /* Same a left_asymmetric, by first stripe is
2632 * D D D P Q rather than
2636 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2637 qd_idx
= pd_idx
+ 1;
2638 if (pd_idx
== raid_disks
-1) {
2639 (*dd_idx
)++; /* Q D D D P */
2641 } else if (*dd_idx
>= pd_idx
)
2642 (*dd_idx
) += 2; /* D D P Q D */
2646 case ALGORITHM_ROTATING_N_CONTINUE
:
2647 /* Same as left_symmetric but Q is before P */
2648 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2649 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2650 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2654 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2655 /* RAID5 left_asymmetric, with Q on last device */
2656 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2657 if (*dd_idx
>= pd_idx
)
2659 qd_idx
= raid_disks
- 1;
2662 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2663 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2664 if (*dd_idx
>= pd_idx
)
2666 qd_idx
= raid_disks
- 1;
2669 case ALGORITHM_LEFT_SYMMETRIC_6
:
2670 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2671 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2672 qd_idx
= raid_disks
- 1;
2675 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2676 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2677 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2678 qd_idx
= raid_disks
- 1;
2681 case ALGORITHM_PARITY_0_6
:
2684 qd_idx
= raid_disks
- 1;
2694 sh
->pd_idx
= pd_idx
;
2695 sh
->qd_idx
= qd_idx
;
2696 sh
->ddf_layout
= ddf_layout
;
2699 * Finally, compute the new sector number
2701 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2705 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2707 struct r5conf
*conf
= sh
->raid_conf
;
2708 int raid_disks
= sh
->disks
;
2709 int data_disks
= raid_disks
- conf
->max_degraded
;
2710 sector_t new_sector
= sh
->sector
, check
;
2711 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2712 : conf
->chunk_sectors
;
2713 int algorithm
= previous
? conf
->prev_algo
2717 sector_t chunk_number
;
2718 int dummy1
, dd_idx
= i
;
2720 struct stripe_head sh2
;
2722 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2723 stripe
= new_sector
;
2725 if (i
== sh
->pd_idx
)
2727 switch(conf
->level
) {
2730 switch (algorithm
) {
2731 case ALGORITHM_LEFT_ASYMMETRIC
:
2732 case ALGORITHM_RIGHT_ASYMMETRIC
:
2736 case ALGORITHM_LEFT_SYMMETRIC
:
2737 case ALGORITHM_RIGHT_SYMMETRIC
:
2740 i
-= (sh
->pd_idx
+ 1);
2742 case ALGORITHM_PARITY_0
:
2745 case ALGORITHM_PARITY_N
:
2752 if (i
== sh
->qd_idx
)
2753 return 0; /* It is the Q disk */
2754 switch (algorithm
) {
2755 case ALGORITHM_LEFT_ASYMMETRIC
:
2756 case ALGORITHM_RIGHT_ASYMMETRIC
:
2757 case ALGORITHM_ROTATING_ZERO_RESTART
:
2758 case ALGORITHM_ROTATING_N_RESTART
:
2759 if (sh
->pd_idx
== raid_disks
-1)
2760 i
--; /* Q D D D P */
2761 else if (i
> sh
->pd_idx
)
2762 i
-= 2; /* D D P Q D */
2764 case ALGORITHM_LEFT_SYMMETRIC
:
2765 case ALGORITHM_RIGHT_SYMMETRIC
:
2766 if (sh
->pd_idx
== raid_disks
-1)
2767 i
--; /* Q D D D P */
2772 i
-= (sh
->pd_idx
+ 2);
2775 case ALGORITHM_PARITY_0
:
2778 case ALGORITHM_PARITY_N
:
2780 case ALGORITHM_ROTATING_N_CONTINUE
:
2781 /* Like left_symmetric, but P is before Q */
2782 if (sh
->pd_idx
== 0)
2783 i
--; /* P D D D Q */
2788 i
-= (sh
->pd_idx
+ 1);
2791 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2792 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2796 case ALGORITHM_LEFT_SYMMETRIC_6
:
2797 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2799 i
+= data_disks
+ 1;
2800 i
-= (sh
->pd_idx
+ 1);
2802 case ALGORITHM_PARITY_0_6
:
2811 chunk_number
= stripe
* data_disks
+ i
;
2812 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2814 check
= raid5_compute_sector(conf
, r_sector
,
2815 previous
, &dummy1
, &sh2
);
2816 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2817 || sh2
.qd_idx
!= sh
->qd_idx
) {
2818 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2819 mdname(conf
->mddev
));
2826 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2827 int rcw
, int expand
)
2829 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
2830 struct r5conf
*conf
= sh
->raid_conf
;
2831 int level
= conf
->level
;
2835 for (i
= disks
; i
--; ) {
2836 struct r5dev
*dev
= &sh
->dev
[i
];
2839 set_bit(R5_LOCKED
, &dev
->flags
);
2840 set_bit(R5_Wantdrain
, &dev
->flags
);
2842 clear_bit(R5_UPTODATE
, &dev
->flags
);
2846 /* if we are not expanding this is a proper write request, and
2847 * there will be bios with new data to be drained into the
2852 /* False alarm, nothing to do */
2854 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2855 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2857 sh
->reconstruct_state
= reconstruct_state_run
;
2859 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2861 if (s
->locked
+ conf
->max_degraded
== disks
)
2862 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2863 atomic_inc(&conf
->pending_full_writes
);
2865 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2866 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2867 BUG_ON(level
== 6 &&
2868 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
2869 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
2871 for (i
= disks
; i
--; ) {
2872 struct r5dev
*dev
= &sh
->dev
[i
];
2873 if (i
== pd_idx
|| i
== qd_idx
)
2877 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2878 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2879 set_bit(R5_Wantdrain
, &dev
->flags
);
2880 set_bit(R5_LOCKED
, &dev
->flags
);
2881 clear_bit(R5_UPTODATE
, &dev
->flags
);
2886 /* False alarm - nothing to do */
2888 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2889 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2890 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2891 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2894 /* keep the parity disk(s) locked while asynchronous operations
2897 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2898 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2902 int qd_idx
= sh
->qd_idx
;
2903 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2905 set_bit(R5_LOCKED
, &dev
->flags
);
2906 clear_bit(R5_UPTODATE
, &dev
->flags
);
2910 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2911 __func__
, (unsigned long long)sh
->sector
,
2912 s
->locked
, s
->ops_request
);
2916 * Each stripe/dev can have one or more bion attached.
2917 * toread/towrite point to the first in a chain.
2918 * The bi_next chain must be in order.
2920 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
2921 int forwrite
, int previous
)
2924 struct r5conf
*conf
= sh
->raid_conf
;
2927 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2928 (unsigned long long)bi
->bi_iter
.bi_sector
,
2929 (unsigned long long)sh
->sector
);
2932 * If several bio share a stripe. The bio bi_phys_segments acts as a
2933 * reference count to avoid race. The reference count should already be
2934 * increased before this function is called (for example, in
2935 * make_request()), so other bio sharing this stripe will not free the
2936 * stripe. If a stripe is owned by one stripe, the stripe lock will
2939 spin_lock_irq(&sh
->stripe_lock
);
2940 /* Don't allow new IO added to stripes in batch list */
2944 bip
= &sh
->dev
[dd_idx
].towrite
;
2948 bip
= &sh
->dev
[dd_idx
].toread
;
2949 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
2950 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
2952 bip
= & (*bip
)->bi_next
;
2954 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
2957 if (!forwrite
|| previous
)
2958 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
2960 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2964 raid5_inc_bi_active_stripes(bi
);
2967 /* check if page is covered */
2968 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2969 for (bi
=sh
->dev
[dd_idx
].towrite
;
2970 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2971 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
2972 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2973 if (bio_end_sector(bi
) >= sector
)
2974 sector
= bio_end_sector(bi
);
2976 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2977 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
2978 sh
->overwrite_disks
++;
2981 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2982 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
2983 (unsigned long long)sh
->sector
, dd_idx
);
2984 spin_unlock_irq(&sh
->stripe_lock
);
2986 if (conf
->mddev
->bitmap
&& firstwrite
) {
2987 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2989 sh
->bm_seq
= conf
->seq_flush
+1;
2990 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2993 if (stripe_can_batch(sh
))
2994 stripe_add_to_batch_list(conf
, sh
);
2998 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2999 spin_unlock_irq(&sh
->stripe_lock
);
3003 static void end_reshape(struct r5conf
*conf
);
3005 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3006 struct stripe_head
*sh
)
3008 int sectors_per_chunk
=
3009 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3011 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3012 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3014 raid5_compute_sector(conf
,
3015 stripe
* (disks
- conf
->max_degraded
)
3016 *sectors_per_chunk
+ chunk_offset
,
3022 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3023 struct stripe_head_state
*s
, int disks
,
3024 struct bio
**return_bi
)
3027 BUG_ON(sh
->batch_head
);
3028 for (i
= disks
; i
--; ) {
3032 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3033 struct md_rdev
*rdev
;
3035 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3036 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
3037 atomic_inc(&rdev
->nr_pending
);
3042 if (!rdev_set_badblocks(
3046 md_error(conf
->mddev
, rdev
);
3047 rdev_dec_pending(rdev
, conf
->mddev
);
3050 spin_lock_irq(&sh
->stripe_lock
);
3051 /* fail all writes first */
3052 bi
= sh
->dev
[i
].towrite
;
3053 sh
->dev
[i
].towrite
= NULL
;
3054 sh
->overwrite_disks
= 0;
3055 spin_unlock_irq(&sh
->stripe_lock
);
3059 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3060 wake_up(&conf
->wait_for_overlap
);
3062 while (bi
&& bi
->bi_iter
.bi_sector
<
3063 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3064 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3065 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3066 if (!raid5_dec_bi_active_stripes(bi
)) {
3067 md_write_end(conf
->mddev
);
3068 bi
->bi_next
= *return_bi
;
3074 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3075 STRIPE_SECTORS
, 0, 0);
3077 /* and fail all 'written' */
3078 bi
= sh
->dev
[i
].written
;
3079 sh
->dev
[i
].written
= NULL
;
3080 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3081 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3082 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3085 if (bi
) bitmap_end
= 1;
3086 while (bi
&& bi
->bi_iter
.bi_sector
<
3087 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3088 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3089 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3090 if (!raid5_dec_bi_active_stripes(bi
)) {
3091 md_write_end(conf
->mddev
);
3092 bi
->bi_next
= *return_bi
;
3098 /* fail any reads if this device is non-operational and
3099 * the data has not reached the cache yet.
3101 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3102 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3103 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3104 spin_lock_irq(&sh
->stripe_lock
);
3105 bi
= sh
->dev
[i
].toread
;
3106 sh
->dev
[i
].toread
= NULL
;
3107 spin_unlock_irq(&sh
->stripe_lock
);
3108 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3109 wake_up(&conf
->wait_for_overlap
);
3110 while (bi
&& bi
->bi_iter
.bi_sector
<
3111 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3112 struct bio
*nextbi
=
3113 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3114 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3115 if (!raid5_dec_bi_active_stripes(bi
)) {
3116 bi
->bi_next
= *return_bi
;
3123 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3124 STRIPE_SECTORS
, 0, 0);
3125 /* If we were in the middle of a write the parity block might
3126 * still be locked - so just clear all R5_LOCKED flags
3128 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3131 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3132 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3133 md_wakeup_thread(conf
->mddev
->thread
);
3137 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3138 struct stripe_head_state
*s
)
3143 BUG_ON(sh
->batch_head
);
3144 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3145 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3146 wake_up(&conf
->wait_for_overlap
);
3149 /* There is nothing more to do for sync/check/repair.
3150 * Don't even need to abort as that is handled elsewhere
3151 * if needed, and not always wanted e.g. if there is a known
3153 * For recover/replace we need to record a bad block on all
3154 * non-sync devices, or abort the recovery
3156 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3157 /* During recovery devices cannot be removed, so
3158 * locking and refcounting of rdevs is not needed
3160 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3161 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
3163 && !test_bit(Faulty
, &rdev
->flags
)
3164 && !test_bit(In_sync
, &rdev
->flags
)
3165 && !rdev_set_badblocks(rdev
, sh
->sector
,
3168 rdev
= conf
->disks
[i
].replacement
;
3170 && !test_bit(Faulty
, &rdev
->flags
)
3171 && !test_bit(In_sync
, &rdev
->flags
)
3172 && !rdev_set_badblocks(rdev
, sh
->sector
,
3177 conf
->recovery_disabled
=
3178 conf
->mddev
->recovery_disabled
;
3180 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3183 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3185 struct md_rdev
*rdev
;
3187 /* Doing recovery so rcu locking not required */
3188 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
3190 && !test_bit(Faulty
, &rdev
->flags
)
3191 && !test_bit(In_sync
, &rdev
->flags
)
3192 && (rdev
->recovery_offset
<= sh
->sector
3193 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3199 /* fetch_block - checks the given member device to see if its data needs
3200 * to be read or computed to satisfy a request.
3202 * Returns 1 when no more member devices need to be checked, otherwise returns
3203 * 0 to tell the loop in handle_stripe_fill to continue
3206 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3207 int disk_idx
, int disks
)
3209 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3210 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3211 &sh
->dev
[s
->failed_num
[1]] };
3215 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3216 test_bit(R5_UPTODATE
, &dev
->flags
))
3217 /* No point reading this as we already have it or have
3218 * decided to get it.
3223 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3224 /* We need this block to directly satisfy a request */
3227 if (s
->syncing
|| s
->expanding
||
3228 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3229 /* When syncing, or expanding we read everything.
3230 * When replacing, we need the replaced block.
3234 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3235 (s
->failed
>= 2 && fdev
[1]->toread
))
3236 /* If we want to read from a failed device, then
3237 * we need to actually read every other device.
3241 /* Sometimes neither read-modify-write nor reconstruct-write
3242 * cycles can work. In those cases we read every block we
3243 * can. Then the parity-update is certain to have enough to
3245 * This can only be a problem when we need to write something,
3246 * and some device has failed. If either of those tests
3247 * fail we need look no further.
3249 if (!s
->failed
|| !s
->to_write
)
3252 if (test_bit(R5_Insync
, &dev
->flags
) &&
3253 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3254 /* Pre-reads at not permitted until after short delay
3255 * to gather multiple requests. However if this
3256 * device is no Insync, the block could only be be computed
3257 * and there is no need to delay that.
3261 for (i
= 0; i
< s
->failed
; i
++) {
3262 if (fdev
[i
]->towrite
&&
3263 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3264 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3265 /* If we have a partial write to a failed
3266 * device, then we will need to reconstruct
3267 * the content of that device, so all other
3268 * devices must be read.
3273 /* If we are forced to do a reconstruct-write, either because
3274 * the current RAID6 implementation only supports that, or
3275 * or because parity cannot be trusted and we are currently
3276 * recovering it, there is extra need to be careful.
3277 * If one of the devices that we would need to read, because
3278 * it is not being overwritten (and maybe not written at all)
3279 * is missing/faulty, then we need to read everything we can.
3281 if (sh
->raid_conf
->level
!= 6 &&
3282 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3283 /* reconstruct-write isn't being forced */
3285 for (i
= 0; i
< s
->failed
; i
++) {
3286 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3287 s
->failed_num
[i
] != sh
->qd_idx
&&
3288 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3289 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3296 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3297 int disk_idx
, int disks
)
3299 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3301 /* is the data in this block needed, and can we get it? */
3302 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3303 /* we would like to get this block, possibly by computing it,
3304 * otherwise read it if the backing disk is insync
3306 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3307 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3308 BUG_ON(sh
->batch_head
);
3309 if ((s
->uptodate
== disks
- 1) &&
3310 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3311 disk_idx
== s
->failed_num
[1]))) {
3312 /* have disk failed, and we're requested to fetch it;
3315 pr_debug("Computing stripe %llu block %d\n",
3316 (unsigned long long)sh
->sector
, disk_idx
);
3317 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3318 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3319 set_bit(R5_Wantcompute
, &dev
->flags
);
3320 sh
->ops
.target
= disk_idx
;
3321 sh
->ops
.target2
= -1; /* no 2nd target */
3323 /* Careful: from this point on 'uptodate' is in the eye
3324 * of raid_run_ops which services 'compute' operations
3325 * before writes. R5_Wantcompute flags a block that will
3326 * be R5_UPTODATE by the time it is needed for a
3327 * subsequent operation.
3331 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3332 /* Computing 2-failure is *very* expensive; only
3333 * do it if failed >= 2
3336 for (other
= disks
; other
--; ) {
3337 if (other
== disk_idx
)
3339 if (!test_bit(R5_UPTODATE
,
3340 &sh
->dev
[other
].flags
))
3344 pr_debug("Computing stripe %llu blocks %d,%d\n",
3345 (unsigned long long)sh
->sector
,
3347 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3348 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3349 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3350 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3351 sh
->ops
.target
= disk_idx
;
3352 sh
->ops
.target2
= other
;
3356 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3357 set_bit(R5_LOCKED
, &dev
->flags
);
3358 set_bit(R5_Wantread
, &dev
->flags
);
3360 pr_debug("Reading block %d (sync=%d)\n",
3361 disk_idx
, s
->syncing
);
3369 * handle_stripe_fill - read or compute data to satisfy pending requests.
3371 static void handle_stripe_fill(struct stripe_head
*sh
,
3372 struct stripe_head_state
*s
,
3377 /* look for blocks to read/compute, skip this if a compute
3378 * is already in flight, or if the stripe contents are in the
3379 * midst of changing due to a write
3381 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3382 !sh
->reconstruct_state
)
3383 for (i
= disks
; i
--; )
3384 if (fetch_block(sh
, s
, i
, disks
))
3386 set_bit(STRIPE_HANDLE
, &sh
->state
);
3389 /* handle_stripe_clean_event
3390 * any written block on an uptodate or failed drive can be returned.
3391 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3392 * never LOCKED, so we don't need to test 'failed' directly.
3394 static void handle_stripe_clean_event(struct r5conf
*conf
,
3395 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
3399 int discard_pending
= 0;
3400 struct stripe_head
*head_sh
= sh
;
3401 bool do_endio
= false;
3404 for (i
= disks
; i
--; )
3405 if (sh
->dev
[i
].written
) {
3407 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3408 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3409 test_bit(R5_Discard
, &dev
->flags
) ||
3410 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3411 /* We can return any write requests */
3412 struct bio
*wbi
, *wbi2
;
3413 pr_debug("Return write for disc %d\n", i
);
3414 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3415 clear_bit(R5_UPTODATE
, &dev
->flags
);
3416 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3417 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3422 dev
->page
= dev
->orig_page
;
3424 dev
->written
= NULL
;
3425 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3426 dev
->sector
+ STRIPE_SECTORS
) {
3427 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3428 if (!raid5_dec_bi_active_stripes(wbi
)) {
3429 md_write_end(conf
->mddev
);
3430 wbi
->bi_next
= *return_bi
;
3435 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3437 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3439 if (head_sh
->batch_head
) {
3440 sh
= list_first_entry(&sh
->batch_list
,
3443 if (sh
!= head_sh
) {
3450 } else if (test_bit(R5_Discard
, &dev
->flags
))
3451 discard_pending
= 1;
3452 WARN_ON(test_bit(R5_SkipCopy
, &dev
->flags
));
3453 WARN_ON(dev
->page
!= dev
->orig_page
);
3455 if (!discard_pending
&&
3456 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3457 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3458 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3459 if (sh
->qd_idx
>= 0) {
3460 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3461 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3463 /* now that discard is done we can proceed with any sync */
3464 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3466 * SCSI discard will change some bio fields and the stripe has
3467 * no updated data, so remove it from hash list and the stripe
3468 * will be reinitialized
3470 spin_lock_irq(&conf
->device_lock
);
3473 if (head_sh
->batch_head
) {
3474 sh
= list_first_entry(&sh
->batch_list
,
3475 struct stripe_head
, batch_list
);
3479 spin_unlock_irq(&conf
->device_lock
);
3482 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3483 set_bit(STRIPE_HANDLE
, &sh
->state
);
3487 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3488 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3489 md_wakeup_thread(conf
->mddev
->thread
);
3491 if (!head_sh
->batch_head
|| !do_endio
)
3493 for (i
= 0; i
< head_sh
->disks
; i
++) {
3494 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
3497 while (!list_empty(&head_sh
->batch_list
)) {
3499 sh
= list_first_entry(&head_sh
->batch_list
,
3500 struct stripe_head
, batch_list
);
3501 list_del_init(&sh
->batch_list
);
3503 set_mask_bits(&sh
->state
, ~STRIPE_EXPAND_SYNC_FLAG
,
3504 head_sh
->state
& ~((1 << STRIPE_ACTIVE
) |
3505 (1 << STRIPE_PREREAD_ACTIVE
) |
3506 STRIPE_EXPAND_SYNC_FLAG
));
3507 sh
->check_state
= head_sh
->check_state
;
3508 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
3509 for (i
= 0; i
< sh
->disks
; i
++) {
3510 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3512 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
;
3515 spin_lock_irq(&sh
->stripe_lock
);
3516 sh
->batch_head
= NULL
;
3517 spin_unlock_irq(&sh
->stripe_lock
);
3518 if (sh
->state
& STRIPE_EXPAND_SYNC_FLAG
)
3519 set_bit(STRIPE_HANDLE
, &sh
->state
);
3523 spin_lock_irq(&head_sh
->stripe_lock
);
3524 head_sh
->batch_head
= NULL
;
3525 spin_unlock_irq(&head_sh
->stripe_lock
);
3526 wake_up_nr(&conf
->wait_for_overlap
, wakeup_nr
);
3527 if (head_sh
->state
& STRIPE_EXPAND_SYNC_FLAG
)
3528 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
3531 static void handle_stripe_dirtying(struct r5conf
*conf
,
3532 struct stripe_head
*sh
,
3533 struct stripe_head_state
*s
,
3536 int rmw
= 0, rcw
= 0, i
;
3537 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3539 /* Check whether resync is now happening or should start.
3540 * If yes, then the array is dirty (after unclean shutdown or
3541 * initial creation), so parity in some stripes might be inconsistent.
3542 * In this case, we need to always do reconstruct-write, to ensure
3543 * that in case of drive failure or read-error correction, we
3544 * generate correct data from the parity.
3546 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3547 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3549 /* Calculate the real rcw later - for now make it
3550 * look like rcw is cheaper
3553 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3554 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3555 (unsigned long long)sh
->sector
);
3556 } else for (i
= disks
; i
--; ) {
3557 /* would I have to read this buffer for read_modify_write */
3558 struct r5dev
*dev
= &sh
->dev
[i
];
3559 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3560 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3561 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3562 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3563 if (test_bit(R5_Insync
, &dev
->flags
))
3566 rmw
+= 2*disks
; /* cannot read it */
3568 /* Would I have to read this buffer for reconstruct_write */
3569 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3570 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3571 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3572 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3573 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3574 if (test_bit(R5_Insync
, &dev
->flags
))
3580 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3581 (unsigned long long)sh
->sector
, rmw
, rcw
);
3582 set_bit(STRIPE_HANDLE
, &sh
->state
);
3583 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_ENABLE_RMW
)) && rmw
> 0) {
3584 /* prefer read-modify-write, but need to get some data */
3585 if (conf
->mddev
->queue
)
3586 blk_add_trace_msg(conf
->mddev
->queue
,
3587 "raid5 rmw %llu %d",
3588 (unsigned long long)sh
->sector
, rmw
);
3589 for (i
= disks
; i
--; ) {
3590 struct r5dev
*dev
= &sh
->dev
[i
];
3591 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3592 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3593 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3594 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3595 test_bit(R5_Insync
, &dev
->flags
)) {
3596 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3598 pr_debug("Read_old block %d for r-m-w\n",
3600 set_bit(R5_LOCKED
, &dev
->flags
);
3601 set_bit(R5_Wantread
, &dev
->flags
);
3604 set_bit(STRIPE_DELAYED
, &sh
->state
);
3605 set_bit(STRIPE_HANDLE
, &sh
->state
);
3610 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_ENABLE_RMW
)) && rcw
> 0) {
3611 /* want reconstruct write, but need to get some data */
3614 for (i
= disks
; i
--; ) {
3615 struct r5dev
*dev
= &sh
->dev
[i
];
3616 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3617 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3618 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3619 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3620 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3622 if (test_bit(R5_Insync
, &dev
->flags
) &&
3623 test_bit(STRIPE_PREREAD_ACTIVE
,
3625 pr_debug("Read_old block "
3626 "%d for Reconstruct\n", i
);
3627 set_bit(R5_LOCKED
, &dev
->flags
);
3628 set_bit(R5_Wantread
, &dev
->flags
);
3632 set_bit(STRIPE_DELAYED
, &sh
->state
);
3633 set_bit(STRIPE_HANDLE
, &sh
->state
);
3637 if (rcw
&& conf
->mddev
->queue
)
3638 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3639 (unsigned long long)sh
->sector
,
3640 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3643 if (rcw
> disks
&& rmw
> disks
&&
3644 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3645 set_bit(STRIPE_DELAYED
, &sh
->state
);
3647 /* now if nothing is locked, and if we have enough data,
3648 * we can start a write request
3650 /* since handle_stripe can be called at any time we need to handle the
3651 * case where a compute block operation has been submitted and then a
3652 * subsequent call wants to start a write request. raid_run_ops only
3653 * handles the case where compute block and reconstruct are requested
3654 * simultaneously. If this is not the case then new writes need to be
3655 * held off until the compute completes.
3657 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
3658 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
3659 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
3660 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
3663 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
3664 struct stripe_head_state
*s
, int disks
)
3666 struct r5dev
*dev
= NULL
;
3668 BUG_ON(sh
->batch_head
);
3669 set_bit(STRIPE_HANDLE
, &sh
->state
);
3671 switch (sh
->check_state
) {
3672 case check_state_idle
:
3673 /* start a new check operation if there are no failures */
3674 if (s
->failed
== 0) {
3675 BUG_ON(s
->uptodate
!= disks
);
3676 sh
->check_state
= check_state_run
;
3677 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3678 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3682 dev
= &sh
->dev
[s
->failed_num
[0]];
3684 case check_state_compute_result
:
3685 sh
->check_state
= check_state_idle
;
3687 dev
= &sh
->dev
[sh
->pd_idx
];
3689 /* check that a write has not made the stripe insync */
3690 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3693 /* either failed parity check, or recovery is happening */
3694 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3695 BUG_ON(s
->uptodate
!= disks
);
3697 set_bit(R5_LOCKED
, &dev
->flags
);
3699 set_bit(R5_Wantwrite
, &dev
->flags
);
3701 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3702 set_bit(STRIPE_INSYNC
, &sh
->state
);
3704 case check_state_run
:
3705 break; /* we will be called again upon completion */
3706 case check_state_check_result
:
3707 sh
->check_state
= check_state_idle
;
3709 /* if a failure occurred during the check operation, leave
3710 * STRIPE_INSYNC not set and let the stripe be handled again
3715 /* handle a successful check operation, if parity is correct
3716 * we are done. Otherwise update the mismatch count and repair
3717 * parity if !MD_RECOVERY_CHECK
3719 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3720 /* parity is correct (on disc,
3721 * not in buffer any more)
3723 set_bit(STRIPE_INSYNC
, &sh
->state
);
3725 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3726 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3727 /* don't try to repair!! */
3728 set_bit(STRIPE_INSYNC
, &sh
->state
);
3730 sh
->check_state
= check_state_compute_run
;
3731 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3732 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3733 set_bit(R5_Wantcompute
,
3734 &sh
->dev
[sh
->pd_idx
].flags
);
3735 sh
->ops
.target
= sh
->pd_idx
;
3736 sh
->ops
.target2
= -1;
3741 case check_state_compute_run
:
3744 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3745 __func__
, sh
->check_state
,
3746 (unsigned long long) sh
->sector
);
3751 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3752 struct stripe_head_state
*s
,
3755 int pd_idx
= sh
->pd_idx
;
3756 int qd_idx
= sh
->qd_idx
;
3759 BUG_ON(sh
->batch_head
);
3760 set_bit(STRIPE_HANDLE
, &sh
->state
);
3762 BUG_ON(s
->failed
> 2);
3764 /* Want to check and possibly repair P and Q.
3765 * However there could be one 'failed' device, in which
3766 * case we can only check one of them, possibly using the
3767 * other to generate missing data
3770 switch (sh
->check_state
) {
3771 case check_state_idle
:
3772 /* start a new check operation if there are < 2 failures */
3773 if (s
->failed
== s
->q_failed
) {
3774 /* The only possible failed device holds Q, so it
3775 * makes sense to check P (If anything else were failed,
3776 * we would have used P to recreate it).
3778 sh
->check_state
= check_state_run
;
3780 if (!s
->q_failed
&& s
->failed
< 2) {
3781 /* Q is not failed, and we didn't use it to generate
3782 * anything, so it makes sense to check it
3784 if (sh
->check_state
== check_state_run
)
3785 sh
->check_state
= check_state_run_pq
;
3787 sh
->check_state
= check_state_run_q
;
3790 /* discard potentially stale zero_sum_result */
3791 sh
->ops
.zero_sum_result
= 0;
3793 if (sh
->check_state
== check_state_run
) {
3794 /* async_xor_zero_sum destroys the contents of P */
3795 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3798 if (sh
->check_state
>= check_state_run
&&
3799 sh
->check_state
<= check_state_run_pq
) {
3800 /* async_syndrome_zero_sum preserves P and Q, so
3801 * no need to mark them !uptodate here
3803 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3807 /* we have 2-disk failure */
3808 BUG_ON(s
->failed
!= 2);
3810 case check_state_compute_result
:
3811 sh
->check_state
= check_state_idle
;
3813 /* check that a write has not made the stripe insync */
3814 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3817 /* now write out any block on a failed drive,
3818 * or P or Q if they were recomputed
3820 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3821 if (s
->failed
== 2) {
3822 dev
= &sh
->dev
[s
->failed_num
[1]];
3824 set_bit(R5_LOCKED
, &dev
->flags
);
3825 set_bit(R5_Wantwrite
, &dev
->flags
);
3827 if (s
->failed
>= 1) {
3828 dev
= &sh
->dev
[s
->failed_num
[0]];
3830 set_bit(R5_LOCKED
, &dev
->flags
);
3831 set_bit(R5_Wantwrite
, &dev
->flags
);
3833 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3834 dev
= &sh
->dev
[pd_idx
];
3836 set_bit(R5_LOCKED
, &dev
->flags
);
3837 set_bit(R5_Wantwrite
, &dev
->flags
);
3839 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3840 dev
= &sh
->dev
[qd_idx
];
3842 set_bit(R5_LOCKED
, &dev
->flags
);
3843 set_bit(R5_Wantwrite
, &dev
->flags
);
3845 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3847 set_bit(STRIPE_INSYNC
, &sh
->state
);
3849 case check_state_run
:
3850 case check_state_run_q
:
3851 case check_state_run_pq
:
3852 break; /* we will be called again upon completion */
3853 case check_state_check_result
:
3854 sh
->check_state
= check_state_idle
;
3856 /* handle a successful check operation, if parity is correct
3857 * we are done. Otherwise update the mismatch count and repair
3858 * parity if !MD_RECOVERY_CHECK
3860 if (sh
->ops
.zero_sum_result
== 0) {
3861 /* both parities are correct */
3863 set_bit(STRIPE_INSYNC
, &sh
->state
);
3865 /* in contrast to the raid5 case we can validate
3866 * parity, but still have a failure to write
3869 sh
->check_state
= check_state_compute_result
;
3870 /* Returning at this point means that we may go
3871 * off and bring p and/or q uptodate again so
3872 * we make sure to check zero_sum_result again
3873 * to verify if p or q need writeback
3877 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3878 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3879 /* don't try to repair!! */
3880 set_bit(STRIPE_INSYNC
, &sh
->state
);
3882 int *target
= &sh
->ops
.target
;
3884 sh
->ops
.target
= -1;
3885 sh
->ops
.target2
= -1;
3886 sh
->check_state
= check_state_compute_run
;
3887 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3888 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3889 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3890 set_bit(R5_Wantcompute
,
3891 &sh
->dev
[pd_idx
].flags
);
3893 target
= &sh
->ops
.target2
;
3896 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3897 set_bit(R5_Wantcompute
,
3898 &sh
->dev
[qd_idx
].flags
);
3905 case check_state_compute_run
:
3908 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3909 __func__
, sh
->check_state
,
3910 (unsigned long long) sh
->sector
);
3915 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3919 /* We have read all the blocks in this stripe and now we need to
3920 * copy some of them into a target stripe for expand.
3922 struct dma_async_tx_descriptor
*tx
= NULL
;
3923 BUG_ON(sh
->batch_head
);
3924 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3925 for (i
= 0; i
< sh
->disks
; i
++)
3926 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3928 struct stripe_head
*sh2
;
3929 struct async_submit_ctl submit
;
3931 sector_t bn
= compute_blocknr(sh
, i
, 1);
3932 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3934 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3936 /* so far only the early blocks of this stripe
3937 * have been requested. When later blocks
3938 * get requested, we will try again
3941 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3942 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3943 /* must have already done this block */
3944 release_stripe(sh2
);
3948 /* place all the copies on one channel */
3949 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3950 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3951 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3954 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3955 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3956 for (j
= 0; j
< conf
->raid_disks
; j
++)
3957 if (j
!= sh2
->pd_idx
&&
3959 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3961 if (j
== conf
->raid_disks
) {
3962 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3963 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3965 release_stripe(sh2
);
3968 /* done submitting copies, wait for them to complete */
3969 async_tx_quiesce(&tx
);
3973 * handle_stripe - do things to a stripe.
3975 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3976 * state of various bits to see what needs to be done.
3978 * return some read requests which now have data
3979 * return some write requests which are safely on storage
3980 * schedule a read on some buffers
3981 * schedule a write of some buffers
3982 * return confirmation of parity correctness
3986 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3988 struct r5conf
*conf
= sh
->raid_conf
;
3989 int disks
= sh
->disks
;
3992 int do_recovery
= 0;
3994 memset(s
, 0, sizeof(*s
));
3996 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
3997 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
3998 s
->failed_num
[0] = -1;
3999 s
->failed_num
[1] = -1;
4001 /* Now to look around and see what can be done */
4003 for (i
=disks
; i
--; ) {
4004 struct md_rdev
*rdev
;
4011 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4013 dev
->toread
, dev
->towrite
, dev
->written
);
4014 /* maybe we can reply to a read
4016 * new wantfill requests are only permitted while
4017 * ops_complete_biofill is guaranteed to be inactive
4019 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4020 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4021 set_bit(R5_Wantfill
, &dev
->flags
);
4023 /* now count some things */
4024 if (test_bit(R5_LOCKED
, &dev
->flags
))
4026 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4028 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4030 BUG_ON(s
->compute
> 2);
4033 if (test_bit(R5_Wantfill
, &dev
->flags
))
4035 else if (dev
->toread
)
4039 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4044 /* Prefer to use the replacement for reads, but only
4045 * if it is recovered enough and has no bad blocks.
4047 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4048 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4049 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4050 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4051 &first_bad
, &bad_sectors
))
4052 set_bit(R5_ReadRepl
, &dev
->flags
);
4055 set_bit(R5_NeedReplace
, &dev
->flags
);
4056 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4057 clear_bit(R5_ReadRepl
, &dev
->flags
);
4059 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4062 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4063 &first_bad
, &bad_sectors
);
4064 if (s
->blocked_rdev
== NULL
4065 && (test_bit(Blocked
, &rdev
->flags
)
4068 set_bit(BlockedBadBlocks
,
4070 s
->blocked_rdev
= rdev
;
4071 atomic_inc(&rdev
->nr_pending
);
4074 clear_bit(R5_Insync
, &dev
->flags
);
4078 /* also not in-sync */
4079 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4080 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4081 /* treat as in-sync, but with a read error
4082 * which we can now try to correct
4084 set_bit(R5_Insync
, &dev
->flags
);
4085 set_bit(R5_ReadError
, &dev
->flags
);
4087 } else if (test_bit(In_sync
, &rdev
->flags
))
4088 set_bit(R5_Insync
, &dev
->flags
);
4089 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4090 /* in sync if before recovery_offset */
4091 set_bit(R5_Insync
, &dev
->flags
);
4092 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4093 test_bit(R5_Expanded
, &dev
->flags
))
4094 /* If we've reshaped into here, we assume it is Insync.
4095 * We will shortly update recovery_offset to make
4098 set_bit(R5_Insync
, &dev
->flags
);
4100 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4101 /* This flag does not apply to '.replacement'
4102 * only to .rdev, so make sure to check that*/
4103 struct md_rdev
*rdev2
= rcu_dereference(
4104 conf
->disks
[i
].rdev
);
4106 clear_bit(R5_Insync
, &dev
->flags
);
4107 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4108 s
->handle_bad_blocks
= 1;
4109 atomic_inc(&rdev2
->nr_pending
);
4111 clear_bit(R5_WriteError
, &dev
->flags
);
4113 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4114 /* This flag does not apply to '.replacement'
4115 * only to .rdev, so make sure to check that*/
4116 struct md_rdev
*rdev2
= rcu_dereference(
4117 conf
->disks
[i
].rdev
);
4118 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4119 s
->handle_bad_blocks
= 1;
4120 atomic_inc(&rdev2
->nr_pending
);
4122 clear_bit(R5_MadeGood
, &dev
->flags
);
4124 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4125 struct md_rdev
*rdev2
= rcu_dereference(
4126 conf
->disks
[i
].replacement
);
4127 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4128 s
->handle_bad_blocks
= 1;
4129 atomic_inc(&rdev2
->nr_pending
);
4131 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4133 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4134 /* The ReadError flag will just be confusing now */
4135 clear_bit(R5_ReadError
, &dev
->flags
);
4136 clear_bit(R5_ReWrite
, &dev
->flags
);
4138 if (test_bit(R5_ReadError
, &dev
->flags
))
4139 clear_bit(R5_Insync
, &dev
->flags
);
4140 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4142 s
->failed_num
[s
->failed
] = i
;
4144 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4148 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4149 /* If there is a failed device being replaced,
4150 * we must be recovering.
4151 * else if we are after recovery_cp, we must be syncing
4152 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4153 * else we can only be replacing
4154 * sync and recovery both need to read all devices, and so
4155 * use the same flag.
4158 sh
->sector
>= conf
->mddev
->recovery_cp
||
4159 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4167 static int clear_batch_ready(struct stripe_head
*sh
)
4169 struct stripe_head
*tmp
;
4170 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4172 spin_lock(&sh
->stripe_lock
);
4173 if (!sh
->batch_head
) {
4174 spin_unlock(&sh
->stripe_lock
);
4179 * this stripe could be added to a batch list before we check
4180 * BATCH_READY, skips it
4182 if (sh
->batch_head
!= sh
) {
4183 spin_unlock(&sh
->stripe_lock
);
4186 spin_lock(&sh
->batch_lock
);
4187 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4188 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4189 spin_unlock(&sh
->batch_lock
);
4190 spin_unlock(&sh
->stripe_lock
);
4193 * BATCH_READY is cleared, no new stripes can be added.
4194 * batch_list can be accessed without lock
4199 static void check_break_stripe_batch_list(struct stripe_head
*sh
)
4201 struct stripe_head
*head_sh
, *next
;
4204 if (!test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4209 sh
= list_first_entry(&sh
->batch_list
,
4210 struct stripe_head
, batch_list
);
4211 BUG_ON(sh
== head_sh
);
4212 } while (!test_bit(STRIPE_DEGRADED
, &sh
->state
));
4214 while (sh
!= head_sh
) {
4215 next
= list_first_entry(&sh
->batch_list
,
4216 struct stripe_head
, batch_list
);
4217 list_del_init(&sh
->batch_list
);
4219 set_mask_bits(&sh
->state
, ~STRIPE_EXPAND_SYNC_FLAG
,
4220 head_sh
->state
& ~((1 << STRIPE_ACTIVE
) |
4221 (1 << STRIPE_PREREAD_ACTIVE
) |
4222 (1 << STRIPE_DEGRADED
) |
4223 STRIPE_EXPAND_SYNC_FLAG
));
4224 sh
->check_state
= head_sh
->check_state
;
4225 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4226 for (i
= 0; i
< sh
->disks
; i
++)
4227 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4228 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4230 spin_lock_irq(&sh
->stripe_lock
);
4231 sh
->batch_head
= NULL
;
4232 spin_unlock_irq(&sh
->stripe_lock
);
4234 set_bit(STRIPE_HANDLE
, &sh
->state
);
4241 static void handle_stripe(struct stripe_head
*sh
)
4243 struct stripe_head_state s
;
4244 struct r5conf
*conf
= sh
->raid_conf
;
4247 int disks
= sh
->disks
;
4248 struct r5dev
*pdev
, *qdev
;
4250 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4251 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4252 /* already being handled, ensure it gets handled
4253 * again when current action finishes */
4254 set_bit(STRIPE_HANDLE
, &sh
->state
);
4258 if (clear_batch_ready(sh
) ) {
4259 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4263 check_break_stripe_batch_list(sh
);
4265 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4266 spin_lock(&sh
->stripe_lock
);
4267 /* Cannot process 'sync' concurrently with 'discard' */
4268 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4269 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4270 set_bit(STRIPE_SYNCING
, &sh
->state
);
4271 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4272 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4274 spin_unlock(&sh
->stripe_lock
);
4276 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4278 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4279 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4280 (unsigned long long)sh
->sector
, sh
->state
,
4281 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4282 sh
->check_state
, sh
->reconstruct_state
);
4284 analyse_stripe(sh
, &s
);
4286 if (s
.handle_bad_blocks
) {
4287 set_bit(STRIPE_HANDLE
, &sh
->state
);
4291 if (unlikely(s
.blocked_rdev
)) {
4292 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4293 s
.replacing
|| s
.to_write
|| s
.written
) {
4294 set_bit(STRIPE_HANDLE
, &sh
->state
);
4297 /* There is nothing for the blocked_rdev to block */
4298 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4299 s
.blocked_rdev
= NULL
;
4302 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4303 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4304 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4307 pr_debug("locked=%d uptodate=%d to_read=%d"
4308 " to_write=%d failed=%d failed_num=%d,%d\n",
4309 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4310 s
.failed_num
[0], s
.failed_num
[1]);
4311 /* check if the array has lost more than max_degraded devices and,
4312 * if so, some requests might need to be failed.
4314 if (s
.failed
> conf
->max_degraded
) {
4315 sh
->check_state
= 0;
4316 sh
->reconstruct_state
= 0;
4317 if (s
.to_read
+s
.to_write
+s
.written
)
4318 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
4319 if (s
.syncing
+ s
.replacing
)
4320 handle_failed_sync(conf
, sh
, &s
);
4323 /* Now we check to see if any write operations have recently
4327 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4329 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4330 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4331 sh
->reconstruct_state
= reconstruct_state_idle
;
4333 /* All the 'written' buffers and the parity block are ready to
4334 * be written back to disk
4336 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4337 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4338 BUG_ON(sh
->qd_idx
>= 0 &&
4339 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4340 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4341 for (i
= disks
; i
--; ) {
4342 struct r5dev
*dev
= &sh
->dev
[i
];
4343 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4344 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4346 pr_debug("Writing block %d\n", i
);
4347 set_bit(R5_Wantwrite
, &dev
->flags
);
4352 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4353 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4355 set_bit(STRIPE_INSYNC
, &sh
->state
);
4358 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4359 s
.dec_preread_active
= 1;
4363 * might be able to return some write requests if the parity blocks
4364 * are safe, or on a failed drive
4366 pdev
= &sh
->dev
[sh
->pd_idx
];
4367 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4368 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4369 qdev
= &sh
->dev
[sh
->qd_idx
];
4370 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4371 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4375 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4376 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4377 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4378 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4379 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4380 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4381 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4382 test_bit(R5_Discard
, &qdev
->flags
))))))
4383 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
4385 /* Now we might consider reading some blocks, either to check/generate
4386 * parity, or to satisfy requests
4387 * or to load a block that is being partially written.
4389 if (s
.to_read
|| s
.non_overwrite
4390 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4391 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4394 handle_stripe_fill(sh
, &s
, disks
);
4396 /* Now to consider new write requests and what else, if anything
4397 * should be read. We do not handle new writes when:
4398 * 1/ A 'write' operation (copy+xor) is already in flight.
4399 * 2/ A 'check' operation is in flight, as it may clobber the parity
4402 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
4403 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4405 /* maybe we need to check and possibly fix the parity for this stripe
4406 * Any reads will already have been scheduled, so we just see if enough
4407 * data is available. The parity check is held off while parity
4408 * dependent operations are in flight.
4410 if (sh
->check_state
||
4411 (s
.syncing
&& s
.locked
== 0 &&
4412 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4413 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4414 if (conf
->level
== 6)
4415 handle_parity_checks6(conf
, sh
, &s
, disks
);
4417 handle_parity_checks5(conf
, sh
, &s
, disks
);
4420 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4421 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4422 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4423 /* Write out to replacement devices where possible */
4424 for (i
= 0; i
< conf
->raid_disks
; i
++)
4425 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4426 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4427 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4428 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4432 set_bit(STRIPE_INSYNC
, &sh
->state
);
4433 set_bit(STRIPE_REPLACED
, &sh
->state
);
4435 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4436 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4437 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4438 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4439 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4440 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4441 wake_up(&conf
->wait_for_overlap
);
4444 /* If the failed drives are just a ReadError, then we might need
4445 * to progress the repair/check process
4447 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4448 for (i
= 0; i
< s
.failed
; i
++) {
4449 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4450 if (test_bit(R5_ReadError
, &dev
->flags
)
4451 && !test_bit(R5_LOCKED
, &dev
->flags
)
4452 && test_bit(R5_UPTODATE
, &dev
->flags
)
4454 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4455 set_bit(R5_Wantwrite
, &dev
->flags
);
4456 set_bit(R5_ReWrite
, &dev
->flags
);
4457 set_bit(R5_LOCKED
, &dev
->flags
);
4460 /* let's read it back */
4461 set_bit(R5_Wantread
, &dev
->flags
);
4462 set_bit(R5_LOCKED
, &dev
->flags
);
4468 /* Finish reconstruct operations initiated by the expansion process */
4469 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4470 struct stripe_head
*sh_src
4471 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4472 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4473 /* sh cannot be written until sh_src has been read.
4474 * so arrange for sh to be delayed a little
4476 set_bit(STRIPE_DELAYED
, &sh
->state
);
4477 set_bit(STRIPE_HANDLE
, &sh
->state
);
4478 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4480 atomic_inc(&conf
->preread_active_stripes
);
4481 release_stripe(sh_src
);
4485 release_stripe(sh_src
);
4487 sh
->reconstruct_state
= reconstruct_state_idle
;
4488 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4489 for (i
= conf
->raid_disks
; i
--; ) {
4490 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4491 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4496 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4497 !sh
->reconstruct_state
) {
4498 /* Need to write out all blocks after computing parity */
4499 sh
->disks
= conf
->raid_disks
;
4500 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4501 schedule_reconstruction(sh
, &s
, 1, 1);
4502 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4503 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4504 atomic_dec(&conf
->reshape_stripes
);
4505 wake_up(&conf
->wait_for_overlap
);
4506 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4509 if (s
.expanding
&& s
.locked
== 0 &&
4510 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4511 handle_stripe_expansion(conf
, sh
);
4514 /* wait for this device to become unblocked */
4515 if (unlikely(s
.blocked_rdev
)) {
4516 if (conf
->mddev
->external
)
4517 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4520 /* Internal metadata will immediately
4521 * be written by raid5d, so we don't
4522 * need to wait here.
4524 rdev_dec_pending(s
.blocked_rdev
,
4528 if (s
.handle_bad_blocks
)
4529 for (i
= disks
; i
--; ) {
4530 struct md_rdev
*rdev
;
4531 struct r5dev
*dev
= &sh
->dev
[i
];
4532 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4533 /* We own a safe reference to the rdev */
4534 rdev
= conf
->disks
[i
].rdev
;
4535 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4537 md_error(conf
->mddev
, rdev
);
4538 rdev_dec_pending(rdev
, conf
->mddev
);
4540 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4541 rdev
= conf
->disks
[i
].rdev
;
4542 rdev_clear_badblocks(rdev
, sh
->sector
,
4544 rdev_dec_pending(rdev
, conf
->mddev
);
4546 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4547 rdev
= conf
->disks
[i
].replacement
;
4549 /* rdev have been moved down */
4550 rdev
= conf
->disks
[i
].rdev
;
4551 rdev_clear_badblocks(rdev
, sh
->sector
,
4553 rdev_dec_pending(rdev
, conf
->mddev
);
4558 raid_run_ops(sh
, s
.ops_request
);
4562 if (s
.dec_preread_active
) {
4563 /* We delay this until after ops_run_io so that if make_request
4564 * is waiting on a flush, it won't continue until the writes
4565 * have actually been submitted.
4567 atomic_dec(&conf
->preread_active_stripes
);
4568 if (atomic_read(&conf
->preread_active_stripes
) <
4570 md_wakeup_thread(conf
->mddev
->thread
);
4573 return_io(s
.return_bi
);
4575 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4578 static void raid5_activate_delayed(struct r5conf
*conf
)
4580 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
4581 while (!list_empty(&conf
->delayed_list
)) {
4582 struct list_head
*l
= conf
->delayed_list
.next
;
4583 struct stripe_head
*sh
;
4584 sh
= list_entry(l
, struct stripe_head
, lru
);
4586 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4587 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4588 atomic_inc(&conf
->preread_active_stripes
);
4589 list_add_tail(&sh
->lru
, &conf
->hold_list
);
4590 raid5_wakeup_stripe_thread(sh
);
4595 static void activate_bit_delay(struct r5conf
*conf
,
4596 struct list_head
*temp_inactive_list
)
4598 /* device_lock is held */
4599 struct list_head head
;
4600 list_add(&head
, &conf
->bitmap_list
);
4601 list_del_init(&conf
->bitmap_list
);
4602 while (!list_empty(&head
)) {
4603 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
4605 list_del_init(&sh
->lru
);
4606 atomic_inc(&sh
->count
);
4607 hash
= sh
->hash_lock_index
;
4608 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
4612 static int raid5_congested(struct mddev
*mddev
, int bits
)
4614 struct r5conf
*conf
= mddev
->private;
4616 /* No difference between reads and writes. Just check
4617 * how busy the stripe_cache is
4620 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
4624 if (atomic_read(&conf
->empty_inactive_list_nr
))
4630 /* We want read requests to align with chunks where possible,
4631 * but write requests don't need to.
4633 static int raid5_mergeable_bvec(struct mddev
*mddev
,
4634 struct bvec_merge_data
*bvm
,
4635 struct bio_vec
*biovec
)
4637 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
4639 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4640 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
4643 * always allow writes to be mergeable, read as well if array
4644 * is degraded as we'll go through stripe cache anyway.
4646 if ((bvm
->bi_rw
& 1) == WRITE
|| mddev
->degraded
)
4647 return biovec
->bv_len
;
4649 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4650 chunk_sectors
= mddev
->new_chunk_sectors
;
4651 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
4652 if (max
< 0) max
= 0;
4653 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
4654 return biovec
->bv_len
;
4659 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
4661 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
4662 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4663 unsigned int bio_sectors
= bio_sectors(bio
);
4665 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4666 chunk_sectors
= mddev
->new_chunk_sectors
;
4667 return chunk_sectors
>=
4668 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
4672 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4673 * later sampled by raid5d.
4675 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
4677 unsigned long flags
;
4679 spin_lock_irqsave(&conf
->device_lock
, flags
);
4681 bi
->bi_next
= conf
->retry_read_aligned_list
;
4682 conf
->retry_read_aligned_list
= bi
;
4684 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4685 md_wakeup_thread(conf
->mddev
->thread
);
4688 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
4692 bi
= conf
->retry_read_aligned
;
4694 conf
->retry_read_aligned
= NULL
;
4697 bi
= conf
->retry_read_aligned_list
;
4699 conf
->retry_read_aligned_list
= bi
->bi_next
;
4702 * this sets the active strip count to 1 and the processed
4703 * strip count to zero (upper 8 bits)
4705 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
4712 * The "raid5_align_endio" should check if the read succeeded and if it
4713 * did, call bio_endio on the original bio (having bio_put the new bio
4715 * If the read failed..
4717 static void raid5_align_endio(struct bio
*bi
, int error
)
4719 struct bio
* raid_bi
= bi
->bi_private
;
4720 struct mddev
*mddev
;
4721 struct r5conf
*conf
;
4722 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4723 struct md_rdev
*rdev
;
4727 rdev
= (void*)raid_bi
->bi_next
;
4728 raid_bi
->bi_next
= NULL
;
4729 mddev
= rdev
->mddev
;
4730 conf
= mddev
->private;
4732 rdev_dec_pending(rdev
, conf
->mddev
);
4734 if (!error
&& uptodate
) {
4735 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
4737 bio_endio(raid_bi
, 0);
4738 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4739 wake_up(&conf
->wait_for_stripe
);
4743 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4745 add_bio_to_retry(raid_bi
, conf
);
4748 static int bio_fits_rdev(struct bio
*bi
)
4750 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
4752 if (bio_sectors(bi
) > queue_max_sectors(q
))
4754 blk_recount_segments(q
, bi
);
4755 if (bi
->bi_phys_segments
> queue_max_segments(q
))
4758 if (q
->merge_bvec_fn
)
4759 /* it's too hard to apply the merge_bvec_fn at this stage,
4767 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
4769 struct r5conf
*conf
= mddev
->private;
4771 struct bio
* align_bi
;
4772 struct md_rdev
*rdev
;
4773 sector_t end_sector
;
4775 if (!in_chunk_boundary(mddev
, raid_bio
)) {
4776 pr_debug("chunk_aligned_read : non aligned\n");
4780 * use bio_clone_mddev to make a copy of the bio
4782 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
4786 * set bi_end_io to a new function, and set bi_private to the
4789 align_bi
->bi_end_io
= raid5_align_endio
;
4790 align_bi
->bi_private
= raid_bio
;
4794 align_bi
->bi_iter
.bi_sector
=
4795 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
4798 end_sector
= bio_end_sector(align_bi
);
4800 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
4801 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
4802 rdev
->recovery_offset
< end_sector
) {
4803 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
4805 (test_bit(Faulty
, &rdev
->flags
) ||
4806 !(test_bit(In_sync
, &rdev
->flags
) ||
4807 rdev
->recovery_offset
>= end_sector
)))
4814 atomic_inc(&rdev
->nr_pending
);
4816 raid_bio
->bi_next
= (void*)rdev
;
4817 align_bi
->bi_bdev
= rdev
->bdev
;
4818 __clear_bit(BIO_SEG_VALID
, &align_bi
->bi_flags
);
4820 if (!bio_fits_rdev(align_bi
) ||
4821 is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
4822 bio_sectors(align_bi
),
4823 &first_bad
, &bad_sectors
)) {
4824 /* too big in some way, or has a known bad block */
4826 rdev_dec_pending(rdev
, mddev
);
4830 /* No reshape active, so we can trust rdev->data_offset */
4831 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
4833 spin_lock_irq(&conf
->device_lock
);
4834 wait_event_lock_irq(conf
->wait_for_stripe
,
4837 atomic_inc(&conf
->active_aligned_reads
);
4838 spin_unlock_irq(&conf
->device_lock
);
4841 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4842 align_bi
, disk_devt(mddev
->gendisk
),
4843 raid_bio
->bi_iter
.bi_sector
);
4844 generic_make_request(align_bi
);
4853 /* __get_priority_stripe - get the next stripe to process
4855 * Full stripe writes are allowed to pass preread active stripes up until
4856 * the bypass_threshold is exceeded. In general the bypass_count
4857 * increments when the handle_list is handled before the hold_list; however, it
4858 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4859 * stripe with in flight i/o. The bypass_count will be reset when the
4860 * head of the hold_list has changed, i.e. the head was promoted to the
4863 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
4865 struct stripe_head
*sh
= NULL
, *tmp
;
4866 struct list_head
*handle_list
= NULL
;
4867 struct r5worker_group
*wg
= NULL
;
4869 if (conf
->worker_cnt_per_group
== 0) {
4870 handle_list
= &conf
->handle_list
;
4871 } else if (group
!= ANY_GROUP
) {
4872 handle_list
= &conf
->worker_groups
[group
].handle_list
;
4873 wg
= &conf
->worker_groups
[group
];
4876 for (i
= 0; i
< conf
->group_cnt
; i
++) {
4877 handle_list
= &conf
->worker_groups
[i
].handle_list
;
4878 wg
= &conf
->worker_groups
[i
];
4879 if (!list_empty(handle_list
))
4884 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4886 list_empty(handle_list
) ? "empty" : "busy",
4887 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4888 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4890 if (!list_empty(handle_list
)) {
4891 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
4893 if (list_empty(&conf
->hold_list
))
4894 conf
->bypass_count
= 0;
4895 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4896 if (conf
->hold_list
.next
== conf
->last_hold
)
4897 conf
->bypass_count
++;
4899 conf
->last_hold
= conf
->hold_list
.next
;
4900 conf
->bypass_count
-= conf
->bypass_threshold
;
4901 if (conf
->bypass_count
< 0)
4902 conf
->bypass_count
= 0;
4905 } else if (!list_empty(&conf
->hold_list
) &&
4906 ((conf
->bypass_threshold
&&
4907 conf
->bypass_count
> conf
->bypass_threshold
) ||
4908 atomic_read(&conf
->pending_full_writes
) == 0)) {
4910 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
4911 if (conf
->worker_cnt_per_group
== 0 ||
4912 group
== ANY_GROUP
||
4913 !cpu_online(tmp
->cpu
) ||
4914 cpu_to_group(tmp
->cpu
) == group
) {
4921 conf
->bypass_count
-= conf
->bypass_threshold
;
4922 if (conf
->bypass_count
< 0)
4923 conf
->bypass_count
= 0;
4935 list_del_init(&sh
->lru
);
4936 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
4940 struct raid5_plug_cb
{
4941 struct blk_plug_cb cb
;
4942 struct list_head list
;
4943 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
4946 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4948 struct raid5_plug_cb
*cb
= container_of(
4949 blk_cb
, struct raid5_plug_cb
, cb
);
4950 struct stripe_head
*sh
;
4951 struct mddev
*mddev
= cb
->cb
.data
;
4952 struct r5conf
*conf
= mddev
->private;
4956 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4957 spin_lock_irq(&conf
->device_lock
);
4958 while (!list_empty(&cb
->list
)) {
4959 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4960 list_del_init(&sh
->lru
);
4962 * avoid race release_stripe_plug() sees
4963 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4964 * is still in our list
4966 smp_mb__before_atomic();
4967 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4969 * STRIPE_ON_RELEASE_LIST could be set here. In that
4970 * case, the count is always > 1 here
4972 hash
= sh
->hash_lock_index
;
4973 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
4976 spin_unlock_irq(&conf
->device_lock
);
4978 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
4979 NR_STRIPE_HASH_LOCKS
);
4981 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
4985 static void release_stripe_plug(struct mddev
*mddev
,
4986 struct stripe_head
*sh
)
4988 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4989 raid5_unplug
, mddev
,
4990 sizeof(struct raid5_plug_cb
));
4991 struct raid5_plug_cb
*cb
;
4998 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5000 if (cb
->list
.next
== NULL
) {
5002 INIT_LIST_HEAD(&cb
->list
);
5003 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5004 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5007 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5008 list_add_tail(&sh
->lru
, &cb
->list
);
5013 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5015 struct r5conf
*conf
= mddev
->private;
5016 sector_t logical_sector
, last_sector
;
5017 struct stripe_head
*sh
;
5021 if (mddev
->reshape_position
!= MaxSector
)
5022 /* Skip discard while reshape is happening */
5025 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5026 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5029 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5031 stripe_sectors
= conf
->chunk_sectors
*
5032 (conf
->raid_disks
- conf
->max_degraded
);
5033 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5035 sector_div(last_sector
, stripe_sectors
);
5037 logical_sector
*= conf
->chunk_sectors
;
5038 last_sector
*= conf
->chunk_sectors
;
5040 for (; logical_sector
< last_sector
;
5041 logical_sector
+= STRIPE_SECTORS
) {
5045 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5046 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5047 TASK_UNINTERRUPTIBLE
);
5048 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5049 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5054 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5055 spin_lock_irq(&sh
->stripe_lock
);
5056 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5057 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5059 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5060 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5061 spin_unlock_irq(&sh
->stripe_lock
);
5067 set_bit(STRIPE_DISCARD
, &sh
->state
);
5068 finish_wait(&conf
->wait_for_overlap
, &w
);
5069 sh
->overwrite_disks
= 0;
5070 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5071 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5073 sh
->dev
[d
].towrite
= bi
;
5074 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5075 raid5_inc_bi_active_stripes(bi
);
5076 sh
->overwrite_disks
++;
5078 spin_unlock_irq(&sh
->stripe_lock
);
5079 if (conf
->mddev
->bitmap
) {
5081 d
< conf
->raid_disks
- conf
->max_degraded
;
5083 bitmap_startwrite(mddev
->bitmap
,
5087 sh
->bm_seq
= conf
->seq_flush
+ 1;
5088 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5091 set_bit(STRIPE_HANDLE
, &sh
->state
);
5092 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5093 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5094 atomic_inc(&conf
->preread_active_stripes
);
5095 release_stripe_plug(mddev
, sh
);
5098 remaining
= raid5_dec_bi_active_stripes(bi
);
5099 if (remaining
== 0) {
5100 md_write_end(mddev
);
5105 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
5107 struct r5conf
*conf
= mddev
->private;
5109 sector_t new_sector
;
5110 sector_t logical_sector
, last_sector
;
5111 struct stripe_head
*sh
;
5112 const int rw
= bio_data_dir(bi
);
5117 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
5118 md_flush_request(mddev
, bi
);
5122 md_write_start(mddev
, bi
);
5125 * If array is degraded, better not do chunk aligned read because
5126 * later we might have to read it again in order to reconstruct
5127 * data on failed drives.
5129 if (rw
== READ
&& mddev
->degraded
== 0 &&
5130 mddev
->reshape_position
== MaxSector
&&
5131 chunk_aligned_read(mddev
,bi
))
5134 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
5135 make_discard_request(mddev
, bi
);
5139 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5140 last_sector
= bio_end_sector(bi
);
5142 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5144 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5145 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5151 seq
= read_seqcount_begin(&conf
->gen_lock
);
5154 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5155 TASK_UNINTERRUPTIBLE
);
5156 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5157 /* spinlock is needed as reshape_progress may be
5158 * 64bit on a 32bit platform, and so it might be
5159 * possible to see a half-updated value
5160 * Of course reshape_progress could change after
5161 * the lock is dropped, so once we get a reference
5162 * to the stripe that we think it is, we will have
5165 spin_lock_irq(&conf
->device_lock
);
5166 if (mddev
->reshape_backwards
5167 ? logical_sector
< conf
->reshape_progress
5168 : logical_sector
>= conf
->reshape_progress
) {
5171 if (mddev
->reshape_backwards
5172 ? logical_sector
< conf
->reshape_safe
5173 : logical_sector
>= conf
->reshape_safe
) {
5174 spin_unlock_irq(&conf
->device_lock
);
5180 spin_unlock_irq(&conf
->device_lock
);
5183 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5186 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5187 (unsigned long long)new_sector
,
5188 (unsigned long long)logical_sector
);
5190 sh
= get_active_stripe(conf
, new_sector
, previous
,
5191 (bi
->bi_rw
&RWA_MASK
), 0);
5193 if (unlikely(previous
)) {
5194 /* expansion might have moved on while waiting for a
5195 * stripe, so we must do the range check again.
5196 * Expansion could still move past after this
5197 * test, but as we are holding a reference to
5198 * 'sh', we know that if that happens,
5199 * STRIPE_EXPANDING will get set and the expansion
5200 * won't proceed until we finish with the stripe.
5203 spin_lock_irq(&conf
->device_lock
);
5204 if (mddev
->reshape_backwards
5205 ? logical_sector
>= conf
->reshape_progress
5206 : logical_sector
< conf
->reshape_progress
)
5207 /* mismatch, need to try again */
5209 spin_unlock_irq(&conf
->device_lock
);
5217 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5218 /* Might have got the wrong stripe_head
5226 logical_sector
>= mddev
->suspend_lo
&&
5227 logical_sector
< mddev
->suspend_hi
) {
5229 /* As the suspend_* range is controlled by
5230 * userspace, we want an interruptible
5233 flush_signals(current
);
5234 prepare_to_wait(&conf
->wait_for_overlap
,
5235 &w
, TASK_INTERRUPTIBLE
);
5236 if (logical_sector
>= mddev
->suspend_lo
&&
5237 logical_sector
< mddev
->suspend_hi
) {
5244 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5245 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5246 /* Stripe is busy expanding or
5247 * add failed due to overlap. Flush everything
5250 md_wakeup_thread(mddev
->thread
);
5256 set_bit(STRIPE_HANDLE
, &sh
->state
);
5257 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5258 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5259 (bi
->bi_rw
& REQ_SYNC
) &&
5260 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5261 atomic_inc(&conf
->preread_active_stripes
);
5262 release_stripe_plug(mddev
, sh
);
5264 /* cannot get stripe for read-ahead, just give-up */
5265 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
5269 finish_wait(&conf
->wait_for_overlap
, &w
);
5271 remaining
= raid5_dec_bi_active_stripes(bi
);
5272 if (remaining
== 0) {
5275 md_write_end(mddev
);
5277 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
5283 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5285 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5287 /* reshaping is quite different to recovery/resync so it is
5288 * handled quite separately ... here.
5290 * On each call to sync_request, we gather one chunk worth of
5291 * destination stripes and flag them as expanding.
5292 * Then we find all the source stripes and request reads.
5293 * As the reads complete, handle_stripe will copy the data
5294 * into the destination stripe and release that stripe.
5296 struct r5conf
*conf
= mddev
->private;
5297 struct stripe_head
*sh
;
5298 sector_t first_sector
, last_sector
;
5299 int raid_disks
= conf
->previous_raid_disks
;
5300 int data_disks
= raid_disks
- conf
->max_degraded
;
5301 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5304 sector_t writepos
, readpos
, safepos
;
5305 sector_t stripe_addr
;
5306 int reshape_sectors
;
5307 struct list_head stripes
;
5309 if (sector_nr
== 0) {
5310 /* If restarting in the middle, skip the initial sectors */
5311 if (mddev
->reshape_backwards
&&
5312 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5313 sector_nr
= raid5_size(mddev
, 0, 0)
5314 - conf
->reshape_progress
;
5315 } else if (!mddev
->reshape_backwards
&&
5316 conf
->reshape_progress
> 0)
5317 sector_nr
= conf
->reshape_progress
;
5318 sector_div(sector_nr
, new_data_disks
);
5320 mddev
->curr_resync_completed
= sector_nr
;
5321 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5327 /* We need to process a full chunk at a time.
5328 * If old and new chunk sizes differ, we need to process the
5331 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
5332 reshape_sectors
= mddev
->new_chunk_sectors
;
5334 reshape_sectors
= mddev
->chunk_sectors
;
5336 /* We update the metadata at least every 10 seconds, or when
5337 * the data about to be copied would over-write the source of
5338 * the data at the front of the range. i.e. one new_stripe
5339 * along from reshape_progress new_maps to after where
5340 * reshape_safe old_maps to
5342 writepos
= conf
->reshape_progress
;
5343 sector_div(writepos
, new_data_disks
);
5344 readpos
= conf
->reshape_progress
;
5345 sector_div(readpos
, data_disks
);
5346 safepos
= conf
->reshape_safe
;
5347 sector_div(safepos
, data_disks
);
5348 if (mddev
->reshape_backwards
) {
5349 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
5350 readpos
+= reshape_sectors
;
5351 safepos
+= reshape_sectors
;
5353 writepos
+= reshape_sectors
;
5354 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5355 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5358 /* Having calculated the 'writepos' possibly use it
5359 * to set 'stripe_addr' which is where we will write to.
5361 if (mddev
->reshape_backwards
) {
5362 BUG_ON(conf
->reshape_progress
== 0);
5363 stripe_addr
= writepos
;
5364 BUG_ON((mddev
->dev_sectors
&
5365 ~((sector_t
)reshape_sectors
- 1))
5366 - reshape_sectors
- stripe_addr
5369 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5370 stripe_addr
= sector_nr
;
5373 /* 'writepos' is the most advanced device address we might write.
5374 * 'readpos' is the least advanced device address we might read.
5375 * 'safepos' is the least address recorded in the metadata as having
5377 * If there is a min_offset_diff, these are adjusted either by
5378 * increasing the safepos/readpos if diff is negative, or
5379 * increasing writepos if diff is positive.
5380 * If 'readpos' is then behind 'writepos', there is no way that we can
5381 * ensure safety in the face of a crash - that must be done by userspace
5382 * making a backup of the data. So in that case there is no particular
5383 * rush to update metadata.
5384 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5385 * update the metadata to advance 'safepos' to match 'readpos' so that
5386 * we can be safe in the event of a crash.
5387 * So we insist on updating metadata if safepos is behind writepos and
5388 * readpos is beyond writepos.
5389 * In any case, update the metadata every 10 seconds.
5390 * Maybe that number should be configurable, but I'm not sure it is
5391 * worth it.... maybe it could be a multiple of safemode_delay???
5393 if (conf
->min_offset_diff
< 0) {
5394 safepos
+= -conf
->min_offset_diff
;
5395 readpos
+= -conf
->min_offset_diff
;
5397 writepos
+= conf
->min_offset_diff
;
5399 if ((mddev
->reshape_backwards
5400 ? (safepos
> writepos
&& readpos
< writepos
)
5401 : (safepos
< writepos
&& readpos
> writepos
)) ||
5402 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5403 /* Cannot proceed until we've updated the superblock... */
5404 wait_event(conf
->wait_for_overlap
,
5405 atomic_read(&conf
->reshape_stripes
)==0
5406 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5407 if (atomic_read(&conf
->reshape_stripes
) != 0)
5409 mddev
->reshape_position
= conf
->reshape_progress
;
5410 mddev
->curr_resync_completed
= sector_nr
;
5411 conf
->reshape_checkpoint
= jiffies
;
5412 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5413 md_wakeup_thread(mddev
->thread
);
5414 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
5415 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5416 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5418 spin_lock_irq(&conf
->device_lock
);
5419 conf
->reshape_safe
= mddev
->reshape_position
;
5420 spin_unlock_irq(&conf
->device_lock
);
5421 wake_up(&conf
->wait_for_overlap
);
5422 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5425 INIT_LIST_HEAD(&stripes
);
5426 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5428 int skipped_disk
= 0;
5429 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5430 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5431 atomic_inc(&conf
->reshape_stripes
);
5432 /* If any of this stripe is beyond the end of the old
5433 * array, then we need to zero those blocks
5435 for (j
=sh
->disks
; j
--;) {
5437 if (j
== sh
->pd_idx
)
5439 if (conf
->level
== 6 &&
5442 s
= compute_blocknr(sh
, j
, 0);
5443 if (s
< raid5_size(mddev
, 0, 0)) {
5447 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5448 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5449 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5451 if (!skipped_disk
) {
5452 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5453 set_bit(STRIPE_HANDLE
, &sh
->state
);
5455 list_add(&sh
->lru
, &stripes
);
5457 spin_lock_irq(&conf
->device_lock
);
5458 if (mddev
->reshape_backwards
)
5459 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5461 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5462 spin_unlock_irq(&conf
->device_lock
);
5463 /* Ok, those stripe are ready. We can start scheduling
5464 * reads on the source stripes.
5465 * The source stripes are determined by mapping the first and last
5466 * block on the destination stripes.
5469 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5472 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5473 * new_data_disks
- 1),
5475 if (last_sector
>= mddev
->dev_sectors
)
5476 last_sector
= mddev
->dev_sectors
- 1;
5477 while (first_sector
<= last_sector
) {
5478 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
5479 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5480 set_bit(STRIPE_HANDLE
, &sh
->state
);
5482 first_sector
+= STRIPE_SECTORS
;
5484 /* Now that the sources are clearly marked, we can release
5485 * the destination stripes
5487 while (!list_empty(&stripes
)) {
5488 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5489 list_del_init(&sh
->lru
);
5492 /* If this takes us to the resync_max point where we have to pause,
5493 * then we need to write out the superblock.
5495 sector_nr
+= reshape_sectors
;
5496 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
5497 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5498 /* Cannot proceed until we've updated the superblock... */
5499 wait_event(conf
->wait_for_overlap
,
5500 atomic_read(&conf
->reshape_stripes
) == 0
5501 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5502 if (atomic_read(&conf
->reshape_stripes
) != 0)
5504 mddev
->reshape_position
= conf
->reshape_progress
;
5505 mddev
->curr_resync_completed
= sector_nr
;
5506 conf
->reshape_checkpoint
= jiffies
;
5507 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5508 md_wakeup_thread(mddev
->thread
);
5509 wait_event(mddev
->sb_wait
,
5510 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
5511 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5512 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5514 spin_lock_irq(&conf
->device_lock
);
5515 conf
->reshape_safe
= mddev
->reshape_position
;
5516 spin_unlock_irq(&conf
->device_lock
);
5517 wake_up(&conf
->wait_for_overlap
);
5518 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5521 return reshape_sectors
;
5524 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5526 struct r5conf
*conf
= mddev
->private;
5527 struct stripe_head
*sh
;
5528 sector_t max_sector
= mddev
->dev_sectors
;
5529 sector_t sync_blocks
;
5530 int still_degraded
= 0;
5533 if (sector_nr
>= max_sector
) {
5534 /* just being told to finish up .. nothing much to do */
5536 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
5541 if (mddev
->curr_resync
< max_sector
) /* aborted */
5542 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
5544 else /* completed sync */
5546 bitmap_close_sync(mddev
->bitmap
);
5551 /* Allow raid5_quiesce to complete */
5552 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
5554 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
5555 return reshape_request(mddev
, sector_nr
, skipped
);
5557 /* No need to check resync_max as we never do more than one
5558 * stripe, and as resync_max will always be on a chunk boundary,
5559 * if the check in md_do_sync didn't fire, there is no chance
5560 * of overstepping resync_max here
5563 /* if there is too many failed drives and we are trying
5564 * to resync, then assert that we are finished, because there is
5565 * nothing we can do.
5567 if (mddev
->degraded
>= conf
->max_degraded
&&
5568 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
5569 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
5573 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
5575 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
5576 sync_blocks
>= STRIPE_SECTORS
) {
5577 /* we can skip this block, and probably more */
5578 sync_blocks
/= STRIPE_SECTORS
;
5580 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
5583 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
5585 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
5587 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
5588 /* make sure we don't swamp the stripe cache if someone else
5589 * is trying to get access
5591 schedule_timeout_uninterruptible(1);
5593 /* Need to check if array will still be degraded after recovery/resync
5594 * Note in case of > 1 drive failures it's possible we're rebuilding
5595 * one drive while leaving another faulty drive in array.
5598 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5599 struct md_rdev
*rdev
= ACCESS_ONCE(conf
->disks
[i
].rdev
);
5601 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
5606 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
5608 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
5609 set_bit(STRIPE_HANDLE
, &sh
->state
);
5613 return STRIPE_SECTORS
;
5616 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
5618 /* We may not be able to submit a whole bio at once as there
5619 * may not be enough stripe_heads available.
5620 * We cannot pre-allocate enough stripe_heads as we may need
5621 * more than exist in the cache (if we allow ever large chunks).
5622 * So we do one stripe head at a time and record in
5623 * ->bi_hw_segments how many have been done.
5625 * We *know* that this entire raid_bio is in one chunk, so
5626 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5628 struct stripe_head
*sh
;
5630 sector_t sector
, logical_sector
, last_sector
;
5635 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
5636 ~((sector_t
)STRIPE_SECTORS
-1);
5637 sector
= raid5_compute_sector(conf
, logical_sector
,
5639 last_sector
= bio_end_sector(raid_bio
);
5641 for (; logical_sector
< last_sector
;
5642 logical_sector
+= STRIPE_SECTORS
,
5643 sector
+= STRIPE_SECTORS
,
5646 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
5647 /* already done this stripe */
5650 sh
= get_active_stripe(conf
, sector
, 0, 1, 1);
5653 /* failed to get a stripe - must wait */
5654 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5655 conf
->retry_read_aligned
= raid_bio
;
5659 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
5661 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5662 conf
->retry_read_aligned
= raid_bio
;
5666 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
5671 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
5672 if (remaining
== 0) {
5673 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
5675 bio_endio(raid_bio
, 0);
5677 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5678 wake_up(&conf
->wait_for_stripe
);
5682 static int handle_active_stripes(struct r5conf
*conf
, int group
,
5683 struct r5worker
*worker
,
5684 struct list_head
*temp_inactive_list
)
5686 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
5687 int i
, batch_size
= 0, hash
;
5688 bool release_inactive
= false;
5690 while (batch_size
< MAX_STRIPE_BATCH
&&
5691 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
5692 batch
[batch_size
++] = sh
;
5694 if (batch_size
== 0) {
5695 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5696 if (!list_empty(temp_inactive_list
+ i
))
5698 if (i
== NR_STRIPE_HASH_LOCKS
)
5700 release_inactive
= true;
5702 spin_unlock_irq(&conf
->device_lock
);
5704 release_inactive_stripe_list(conf
, temp_inactive_list
,
5705 NR_STRIPE_HASH_LOCKS
);
5707 if (release_inactive
) {
5708 spin_lock_irq(&conf
->device_lock
);
5712 for (i
= 0; i
< batch_size
; i
++)
5713 handle_stripe(batch
[i
]);
5717 spin_lock_irq(&conf
->device_lock
);
5718 for (i
= 0; i
< batch_size
; i
++) {
5719 hash
= batch
[i
]->hash_lock_index
;
5720 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
5725 static void raid5_do_work(struct work_struct
*work
)
5727 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
5728 struct r5worker_group
*group
= worker
->group
;
5729 struct r5conf
*conf
= group
->conf
;
5730 int group_id
= group
- conf
->worker_groups
;
5732 struct blk_plug plug
;
5734 pr_debug("+++ raid5worker active\n");
5736 blk_start_plug(&plug
);
5738 spin_lock_irq(&conf
->device_lock
);
5740 int batch_size
, released
;
5742 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
5744 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
5745 worker
->temp_inactive_list
);
5746 worker
->working
= false;
5747 if (!batch_size
&& !released
)
5749 handled
+= batch_size
;
5751 pr_debug("%d stripes handled\n", handled
);
5753 spin_unlock_irq(&conf
->device_lock
);
5754 blk_finish_plug(&plug
);
5756 pr_debug("--- raid5worker inactive\n");
5760 * This is our raid5 kernel thread.
5762 * We scan the hash table for stripes which can be handled now.
5763 * During the scan, completed stripes are saved for us by the interrupt
5764 * handler, so that they will not have to wait for our next wakeup.
5766 static void raid5d(struct md_thread
*thread
)
5768 struct mddev
*mddev
= thread
->mddev
;
5769 struct r5conf
*conf
= mddev
->private;
5771 struct blk_plug plug
;
5773 pr_debug("+++ raid5d active\n");
5775 md_check_recovery(mddev
);
5777 blk_start_plug(&plug
);
5779 spin_lock_irq(&conf
->device_lock
);
5782 int batch_size
, released
;
5784 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
5786 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5789 !list_empty(&conf
->bitmap_list
)) {
5790 /* Now is a good time to flush some bitmap updates */
5792 spin_unlock_irq(&conf
->device_lock
);
5793 bitmap_unplug(mddev
->bitmap
);
5794 spin_lock_irq(&conf
->device_lock
);
5795 conf
->seq_write
= conf
->seq_flush
;
5796 activate_bit_delay(conf
, conf
->temp_inactive_list
);
5798 raid5_activate_delayed(conf
);
5800 while ((bio
= remove_bio_from_retry(conf
))) {
5802 spin_unlock_irq(&conf
->device_lock
);
5803 ok
= retry_aligned_read(conf
, bio
);
5804 spin_lock_irq(&conf
->device_lock
);
5810 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
5811 conf
->temp_inactive_list
);
5812 if (!batch_size
&& !released
)
5814 handled
+= batch_size
;
5816 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
5817 spin_unlock_irq(&conf
->device_lock
);
5818 md_check_recovery(mddev
);
5819 spin_lock_irq(&conf
->device_lock
);
5822 pr_debug("%d stripes handled\n", handled
);
5824 spin_unlock_irq(&conf
->device_lock
);
5825 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
)) {
5826 grow_one_stripe(conf
, __GFP_NOWARN
);
5827 /* Set flag even if allocation failed. This helps
5828 * slow down allocation requests when mem is short
5830 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5833 async_tx_issue_pending_all();
5834 blk_finish_plug(&plug
);
5836 pr_debug("--- raid5d inactive\n");
5840 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
5842 struct r5conf
*conf
;
5844 spin_lock(&mddev
->lock
);
5845 conf
= mddev
->private;
5847 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
5848 spin_unlock(&mddev
->lock
);
5853 raid5_set_cache_size(struct mddev
*mddev
, int size
)
5855 struct r5conf
*conf
= mddev
->private;
5858 if (size
<= 16 || size
> 32768)
5861 conf
->min_nr_stripes
= size
;
5862 while (size
< conf
->max_nr_stripes
&&
5863 drop_one_stripe(conf
))
5867 err
= md_allow_write(mddev
);
5871 while (size
> conf
->max_nr_stripes
)
5872 if (!grow_one_stripe(conf
, GFP_KERNEL
))
5877 EXPORT_SYMBOL(raid5_set_cache_size
);
5880 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
5882 struct r5conf
*conf
;
5886 if (len
>= PAGE_SIZE
)
5888 if (kstrtoul(page
, 10, &new))
5890 err
= mddev_lock(mddev
);
5893 conf
= mddev
->private;
5897 err
= raid5_set_cache_size(mddev
, new);
5898 mddev_unlock(mddev
);
5903 static struct md_sysfs_entry
5904 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
5905 raid5_show_stripe_cache_size
,
5906 raid5_store_stripe_cache_size
);
5909 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
5911 struct r5conf
*conf
= mddev
->private;
5913 return sprintf(page
, "%d\n", conf
->rmw_level
);
5919 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
5921 struct r5conf
*conf
= mddev
->private;
5927 if (len
>= PAGE_SIZE
)
5930 if (kstrtoul(page
, 10, &new))
5933 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
5936 if (new != PARITY_DISABLE_RMW
&&
5937 new != PARITY_ENABLE_RMW
&&
5938 new != PARITY_PREFER_RMW
)
5941 conf
->rmw_level
= new;
5945 static struct md_sysfs_entry
5946 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
5947 raid5_show_rmw_level
,
5948 raid5_store_rmw_level
);
5952 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
5954 struct r5conf
*conf
;
5956 spin_lock(&mddev
->lock
);
5957 conf
= mddev
->private;
5959 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
5960 spin_unlock(&mddev
->lock
);
5965 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
5967 struct r5conf
*conf
;
5971 if (len
>= PAGE_SIZE
)
5973 if (kstrtoul(page
, 10, &new))
5976 err
= mddev_lock(mddev
);
5979 conf
= mddev
->private;
5982 else if (new > conf
->min_nr_stripes
)
5985 conf
->bypass_threshold
= new;
5986 mddev_unlock(mddev
);
5990 static struct md_sysfs_entry
5991 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
5993 raid5_show_preread_threshold
,
5994 raid5_store_preread_threshold
);
5997 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
5999 struct r5conf
*conf
;
6001 spin_lock(&mddev
->lock
);
6002 conf
= mddev
->private;
6004 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6005 spin_unlock(&mddev
->lock
);
6010 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6012 struct r5conf
*conf
;
6016 if (len
>= PAGE_SIZE
)
6018 if (kstrtoul(page
, 10, &new))
6022 err
= mddev_lock(mddev
);
6025 conf
= mddev
->private;
6028 else if (new != conf
->skip_copy
) {
6029 mddev_suspend(mddev
);
6030 conf
->skip_copy
= new;
6032 mddev
->queue
->backing_dev_info
.capabilities
|=
6033 BDI_CAP_STABLE_WRITES
;
6035 mddev
->queue
->backing_dev_info
.capabilities
&=
6036 ~BDI_CAP_STABLE_WRITES
;
6037 mddev_resume(mddev
);
6039 mddev_unlock(mddev
);
6043 static struct md_sysfs_entry
6044 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6045 raid5_show_skip_copy
,
6046 raid5_store_skip_copy
);
6049 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6051 struct r5conf
*conf
= mddev
->private;
6053 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6058 static struct md_sysfs_entry
6059 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6062 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6064 struct r5conf
*conf
;
6066 spin_lock(&mddev
->lock
);
6067 conf
= mddev
->private;
6069 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6070 spin_unlock(&mddev
->lock
);
6074 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6076 int *worker_cnt_per_group
,
6077 struct r5worker_group
**worker_groups
);
6079 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6081 struct r5conf
*conf
;
6084 struct r5worker_group
*new_groups
, *old_groups
;
6085 int group_cnt
, worker_cnt_per_group
;
6087 if (len
>= PAGE_SIZE
)
6089 if (kstrtoul(page
, 10, &new))
6092 err
= mddev_lock(mddev
);
6095 conf
= mddev
->private;
6098 else if (new != conf
->worker_cnt_per_group
) {
6099 mddev_suspend(mddev
);
6101 old_groups
= conf
->worker_groups
;
6103 flush_workqueue(raid5_wq
);
6105 err
= alloc_thread_groups(conf
, new,
6106 &group_cnt
, &worker_cnt_per_group
,
6109 spin_lock_irq(&conf
->device_lock
);
6110 conf
->group_cnt
= group_cnt
;
6111 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6112 conf
->worker_groups
= new_groups
;
6113 spin_unlock_irq(&conf
->device_lock
);
6116 kfree(old_groups
[0].workers
);
6119 mddev_resume(mddev
);
6121 mddev_unlock(mddev
);
6126 static struct md_sysfs_entry
6127 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6128 raid5_show_group_thread_cnt
,
6129 raid5_store_group_thread_cnt
);
6131 static struct attribute
*raid5_attrs
[] = {
6132 &raid5_stripecache_size
.attr
,
6133 &raid5_stripecache_active
.attr
,
6134 &raid5_preread_bypass_threshold
.attr
,
6135 &raid5_group_thread_cnt
.attr
,
6136 &raid5_skip_copy
.attr
,
6137 &raid5_rmw_level
.attr
,
6140 static struct attribute_group raid5_attrs_group
= {
6142 .attrs
= raid5_attrs
,
6145 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6147 int *worker_cnt_per_group
,
6148 struct r5worker_group
**worker_groups
)
6152 struct r5worker
*workers
;
6154 *worker_cnt_per_group
= cnt
;
6157 *worker_groups
= NULL
;
6160 *group_cnt
= num_possible_nodes();
6161 size
= sizeof(struct r5worker
) * cnt
;
6162 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6163 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6164 *group_cnt
, GFP_NOIO
);
6165 if (!*worker_groups
|| !workers
) {
6167 kfree(*worker_groups
);
6171 for (i
= 0; i
< *group_cnt
; i
++) {
6172 struct r5worker_group
*group
;
6174 group
= &(*worker_groups
)[i
];
6175 INIT_LIST_HEAD(&group
->handle_list
);
6177 group
->workers
= workers
+ i
* cnt
;
6179 for (j
= 0; j
< cnt
; j
++) {
6180 struct r5worker
*worker
= group
->workers
+ j
;
6181 worker
->group
= group
;
6182 INIT_WORK(&worker
->work
, raid5_do_work
);
6184 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6185 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6192 static void free_thread_groups(struct r5conf
*conf
)
6194 if (conf
->worker_groups
)
6195 kfree(conf
->worker_groups
[0].workers
);
6196 kfree(conf
->worker_groups
);
6197 conf
->worker_groups
= NULL
;
6201 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6203 struct r5conf
*conf
= mddev
->private;
6206 sectors
= mddev
->dev_sectors
;
6208 /* size is defined by the smallest of previous and new size */
6209 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6211 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
6212 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
6213 return sectors
* (raid_disks
- conf
->max_degraded
);
6216 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6218 safe_put_page(percpu
->spare_page
);
6219 if (percpu
->scribble
)
6220 flex_array_free(percpu
->scribble
);
6221 percpu
->spare_page
= NULL
;
6222 percpu
->scribble
= NULL
;
6225 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6227 if (conf
->level
== 6 && !percpu
->spare_page
)
6228 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6229 if (!percpu
->scribble
)
6230 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6231 conf
->previous_raid_disks
), conf
->chunk_sectors
/
6232 STRIPE_SECTORS
, GFP_KERNEL
);
6234 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6235 free_scratch_buffer(conf
, percpu
);
6242 static void raid5_free_percpu(struct r5conf
*conf
)
6249 #ifdef CONFIG_HOTPLUG_CPU
6250 unregister_cpu_notifier(&conf
->cpu_notify
);
6254 for_each_possible_cpu(cpu
)
6255 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6258 free_percpu(conf
->percpu
);
6261 static void free_conf(struct r5conf
*conf
)
6263 if (conf
->shrinker
.seeks
)
6264 unregister_shrinker(&conf
->shrinker
);
6265 free_thread_groups(conf
);
6266 shrink_stripes(conf
);
6267 raid5_free_percpu(conf
);
6269 kfree(conf
->stripe_hashtbl
);
6273 #ifdef CONFIG_HOTPLUG_CPU
6274 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
6277 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
6278 long cpu
= (long)hcpu
;
6279 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6282 case CPU_UP_PREPARE
:
6283 case CPU_UP_PREPARE_FROZEN
:
6284 if (alloc_scratch_buffer(conf
, percpu
)) {
6285 pr_err("%s: failed memory allocation for cpu%ld\n",
6287 return notifier_from_errno(-ENOMEM
);
6291 case CPU_DEAD_FROZEN
:
6292 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6301 static int raid5_alloc_percpu(struct r5conf
*conf
)
6306 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6310 #ifdef CONFIG_HOTPLUG_CPU
6311 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
6312 conf
->cpu_notify
.priority
= 0;
6313 err
= register_cpu_notifier(&conf
->cpu_notify
);
6319 for_each_present_cpu(cpu
) {
6320 err
= alloc_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6322 pr_err("%s: failed memory allocation for cpu%ld\n",
6332 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6333 struct shrink_control
*sc
)
6335 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6337 while (ret
< sc
->nr_to_scan
) {
6338 if (drop_one_stripe(conf
) == 0)
6345 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6346 struct shrink_control
*sc
)
6348 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6350 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6351 /* unlikely, but not impossible */
6353 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6356 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6358 struct r5conf
*conf
;
6359 int raid_disk
, memory
, max_disks
;
6360 struct md_rdev
*rdev
;
6361 struct disk_info
*disk
;
6364 int group_cnt
, worker_cnt_per_group
;
6365 struct r5worker_group
*new_group
;
6367 if (mddev
->new_level
!= 5
6368 && mddev
->new_level
!= 4
6369 && mddev
->new_level
!= 6) {
6370 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6371 mdname(mddev
), mddev
->new_level
);
6372 return ERR_PTR(-EIO
);
6374 if ((mddev
->new_level
== 5
6375 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6376 (mddev
->new_level
== 6
6377 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6378 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
6379 mdname(mddev
), mddev
->new_layout
);
6380 return ERR_PTR(-EIO
);
6382 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6383 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6384 mdname(mddev
), mddev
->raid_disks
);
6385 return ERR_PTR(-EINVAL
);
6388 if (!mddev
->new_chunk_sectors
||
6389 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6390 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6391 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
6392 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6393 return ERR_PTR(-EINVAL
);
6396 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6399 /* Don't enable multi-threading by default*/
6400 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6402 conf
->group_cnt
= group_cnt
;
6403 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6404 conf
->worker_groups
= new_group
;
6407 spin_lock_init(&conf
->device_lock
);
6408 seqcount_init(&conf
->gen_lock
);
6409 init_waitqueue_head(&conf
->wait_for_stripe
);
6410 init_waitqueue_head(&conf
->wait_for_overlap
);
6411 INIT_LIST_HEAD(&conf
->handle_list
);
6412 INIT_LIST_HEAD(&conf
->hold_list
);
6413 INIT_LIST_HEAD(&conf
->delayed_list
);
6414 INIT_LIST_HEAD(&conf
->bitmap_list
);
6415 init_llist_head(&conf
->released_stripes
);
6416 atomic_set(&conf
->active_stripes
, 0);
6417 atomic_set(&conf
->preread_active_stripes
, 0);
6418 atomic_set(&conf
->active_aligned_reads
, 0);
6419 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6420 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6422 conf
->raid_disks
= mddev
->raid_disks
;
6423 if (mddev
->reshape_position
== MaxSector
)
6424 conf
->previous_raid_disks
= mddev
->raid_disks
;
6426 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6427 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6429 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6434 conf
->mddev
= mddev
;
6436 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6439 /* We init hash_locks[0] separately to that it can be used
6440 * as the reference lock in the spin_lock_nest_lock() call
6441 * in lock_all_device_hash_locks_irq in order to convince
6442 * lockdep that we know what we are doing.
6444 spin_lock_init(conf
->hash_locks
);
6445 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6446 spin_lock_init(conf
->hash_locks
+ i
);
6448 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6449 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6451 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6452 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6454 conf
->level
= mddev
->new_level
;
6455 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6456 if (raid5_alloc_percpu(conf
) != 0)
6459 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6461 rdev_for_each(rdev
, mddev
) {
6462 raid_disk
= rdev
->raid_disk
;
6463 if (raid_disk
>= max_disks
6466 disk
= conf
->disks
+ raid_disk
;
6468 if (test_bit(Replacement
, &rdev
->flags
)) {
6469 if (disk
->replacement
)
6471 disk
->replacement
= rdev
;
6478 if (test_bit(In_sync
, &rdev
->flags
)) {
6479 char b
[BDEVNAME_SIZE
];
6480 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
6482 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
6483 } else if (rdev
->saved_raid_disk
!= raid_disk
)
6484 /* Cannot rely on bitmap to complete recovery */
6488 conf
->level
= mddev
->new_level
;
6489 if (conf
->level
== 6) {
6490 conf
->max_degraded
= 2;
6491 if (raid6_call
.xor_syndrome
)
6492 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6494 conf
->rmw_level
= PARITY_DISABLE_RMW
;
6496 conf
->max_degraded
= 1;
6497 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6499 conf
->algorithm
= mddev
->new_layout
;
6500 conf
->reshape_progress
= mddev
->reshape_position
;
6501 if (conf
->reshape_progress
!= MaxSector
) {
6502 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
6503 conf
->prev_algo
= mddev
->layout
;
6506 conf
->min_nr_stripes
= NR_STRIPES
;
6507 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
6508 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
6509 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
6510 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
6512 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6513 mdname(mddev
), memory
);
6516 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
6517 mdname(mddev
), memory
);
6519 * Losing a stripe head costs more than the time to refill it,
6520 * it reduces the queue depth and so can hurt throughput.
6521 * So set it rather large, scaled by number of devices.
6523 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
6524 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
6525 conf
->shrinker
.count_objects
= raid5_cache_count
;
6526 conf
->shrinker
.batch
= 128;
6527 conf
->shrinker
.flags
= 0;
6528 register_shrinker(&conf
->shrinker
);
6530 sprintf(pers_name
, "raid%d", mddev
->new_level
);
6531 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
6532 if (!conf
->thread
) {
6534 "md/raid:%s: couldn't allocate thread.\n",
6544 return ERR_PTR(-EIO
);
6546 return ERR_PTR(-ENOMEM
);
6549 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
6552 case ALGORITHM_PARITY_0
:
6553 if (raid_disk
< max_degraded
)
6556 case ALGORITHM_PARITY_N
:
6557 if (raid_disk
>= raid_disks
- max_degraded
)
6560 case ALGORITHM_PARITY_0_6
:
6561 if (raid_disk
== 0 ||
6562 raid_disk
== raid_disks
- 1)
6565 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6566 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6567 case ALGORITHM_LEFT_SYMMETRIC_6
:
6568 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6569 if (raid_disk
== raid_disks
- 1)
6575 static int run(struct mddev
*mddev
)
6577 struct r5conf
*conf
;
6578 int working_disks
= 0;
6579 int dirty_parity_disks
= 0;
6580 struct md_rdev
*rdev
;
6581 sector_t reshape_offset
= 0;
6583 long long min_offset_diff
= 0;
6586 if (mddev
->recovery_cp
!= MaxSector
)
6587 printk(KERN_NOTICE
"md/raid:%s: not clean"
6588 " -- starting background reconstruction\n",
6591 rdev_for_each(rdev
, mddev
) {
6593 if (rdev
->raid_disk
< 0)
6595 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
6597 min_offset_diff
= diff
;
6599 } else if (mddev
->reshape_backwards
&&
6600 diff
< min_offset_diff
)
6601 min_offset_diff
= diff
;
6602 else if (!mddev
->reshape_backwards
&&
6603 diff
> min_offset_diff
)
6604 min_offset_diff
= diff
;
6607 if (mddev
->reshape_position
!= MaxSector
) {
6608 /* Check that we can continue the reshape.
6609 * Difficulties arise if the stripe we would write to
6610 * next is at or after the stripe we would read from next.
6611 * For a reshape that changes the number of devices, this
6612 * is only possible for a very short time, and mdadm makes
6613 * sure that time appears to have past before assembling
6614 * the array. So we fail if that time hasn't passed.
6615 * For a reshape that keeps the number of devices the same
6616 * mdadm must be monitoring the reshape can keeping the
6617 * critical areas read-only and backed up. It will start
6618 * the array in read-only mode, so we check for that.
6620 sector_t here_new
, here_old
;
6622 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
6624 if (mddev
->new_level
!= mddev
->level
) {
6625 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
6626 "required - aborting.\n",
6630 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6631 /* reshape_position must be on a new-stripe boundary, and one
6632 * further up in new geometry must map after here in old
6635 here_new
= mddev
->reshape_position
;
6636 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
6637 (mddev
->raid_disks
- max_degraded
))) {
6638 printk(KERN_ERR
"md/raid:%s: reshape_position not "
6639 "on a stripe boundary\n", mdname(mddev
));
6642 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
6643 /* here_new is the stripe we will write to */
6644 here_old
= mddev
->reshape_position
;
6645 sector_div(here_old
, mddev
->chunk_sectors
*
6646 (old_disks
-max_degraded
));
6647 /* here_old is the first stripe that we might need to read
6649 if (mddev
->delta_disks
== 0) {
6650 if ((here_new
* mddev
->new_chunk_sectors
!=
6651 here_old
* mddev
->chunk_sectors
)) {
6652 printk(KERN_ERR
"md/raid:%s: reshape position is"
6653 " confused - aborting\n", mdname(mddev
));
6656 /* We cannot be sure it is safe to start an in-place
6657 * reshape. It is only safe if user-space is monitoring
6658 * and taking constant backups.
6659 * mdadm always starts a situation like this in
6660 * readonly mode so it can take control before
6661 * allowing any writes. So just check for that.
6663 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
6664 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
6665 /* not really in-place - so OK */;
6666 else if (mddev
->ro
== 0) {
6667 printk(KERN_ERR
"md/raid:%s: in-place reshape "
6668 "must be started in read-only mode "
6673 } else if (mddev
->reshape_backwards
6674 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
6675 here_old
* mddev
->chunk_sectors
)
6676 : (here_new
* mddev
->new_chunk_sectors
>=
6677 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
6678 /* Reading from the same stripe as writing to - bad */
6679 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
6680 "auto-recovery - aborting.\n",
6684 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
6686 /* OK, we should be able to continue; */
6688 BUG_ON(mddev
->level
!= mddev
->new_level
);
6689 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
6690 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
6691 BUG_ON(mddev
->delta_disks
!= 0);
6694 if (mddev
->private == NULL
)
6695 conf
= setup_conf(mddev
);
6697 conf
= mddev
->private;
6700 return PTR_ERR(conf
);
6702 conf
->min_offset_diff
= min_offset_diff
;
6703 mddev
->thread
= conf
->thread
;
6704 conf
->thread
= NULL
;
6705 mddev
->private = conf
;
6707 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
6709 rdev
= conf
->disks
[i
].rdev
;
6710 if (!rdev
&& conf
->disks
[i
].replacement
) {
6711 /* The replacement is all we have yet */
6712 rdev
= conf
->disks
[i
].replacement
;
6713 conf
->disks
[i
].replacement
= NULL
;
6714 clear_bit(Replacement
, &rdev
->flags
);
6715 conf
->disks
[i
].rdev
= rdev
;
6719 if (conf
->disks
[i
].replacement
&&
6720 conf
->reshape_progress
!= MaxSector
) {
6721 /* replacements and reshape simply do not mix. */
6722 printk(KERN_ERR
"md: cannot handle concurrent "
6723 "replacement and reshape.\n");
6726 if (test_bit(In_sync
, &rdev
->flags
)) {
6730 /* This disc is not fully in-sync. However if it
6731 * just stored parity (beyond the recovery_offset),
6732 * when we don't need to be concerned about the
6733 * array being dirty.
6734 * When reshape goes 'backwards', we never have
6735 * partially completed devices, so we only need
6736 * to worry about reshape going forwards.
6738 /* Hack because v0.91 doesn't store recovery_offset properly. */
6739 if (mddev
->major_version
== 0 &&
6740 mddev
->minor_version
> 90)
6741 rdev
->recovery_offset
= reshape_offset
;
6743 if (rdev
->recovery_offset
< reshape_offset
) {
6744 /* We need to check old and new layout */
6745 if (!only_parity(rdev
->raid_disk
,
6748 conf
->max_degraded
))
6751 if (!only_parity(rdev
->raid_disk
,
6753 conf
->previous_raid_disks
,
6754 conf
->max_degraded
))
6756 dirty_parity_disks
++;
6760 * 0 for a fully functional array, 1 or 2 for a degraded array.
6762 mddev
->degraded
= calc_degraded(conf
);
6764 if (has_failed(conf
)) {
6765 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
6766 " (%d/%d failed)\n",
6767 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
6771 /* device size must be a multiple of chunk size */
6772 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
6773 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
6775 if (mddev
->degraded
> dirty_parity_disks
&&
6776 mddev
->recovery_cp
!= MaxSector
) {
6777 if (mddev
->ok_start_degraded
)
6779 "md/raid:%s: starting dirty degraded array"
6780 " - data corruption possible.\n",
6784 "md/raid:%s: cannot start dirty degraded array.\n",
6790 if (mddev
->degraded
== 0)
6791 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
6792 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
6793 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
6796 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
6797 " out of %d devices, algorithm %d\n",
6798 mdname(mddev
), conf
->level
,
6799 mddev
->raid_disks
- mddev
->degraded
,
6800 mddev
->raid_disks
, mddev
->new_layout
);
6802 print_raid5_conf(conf
);
6804 if (conf
->reshape_progress
!= MaxSector
) {
6805 conf
->reshape_safe
= conf
->reshape_progress
;
6806 atomic_set(&conf
->reshape_stripes
, 0);
6807 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6808 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6809 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6810 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6811 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6815 /* Ok, everything is just fine now */
6816 if (mddev
->to_remove
== &raid5_attrs_group
)
6817 mddev
->to_remove
= NULL
;
6818 else if (mddev
->kobj
.sd
&&
6819 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
6821 "raid5: failed to create sysfs attributes for %s\n",
6823 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6827 bool discard_supported
= true;
6828 /* read-ahead size must cover two whole stripes, which
6829 * is 2 * (datadisks) * chunksize where 'n' is the
6830 * number of raid devices
6832 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
6833 int stripe
= data_disks
*
6834 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
6835 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6836 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6838 chunk_size
= mddev
->chunk_sectors
<< 9;
6839 blk_queue_io_min(mddev
->queue
, chunk_size
);
6840 blk_queue_io_opt(mddev
->queue
, chunk_size
*
6841 (conf
->raid_disks
- conf
->max_degraded
));
6842 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
6844 * We can only discard a whole stripe. It doesn't make sense to
6845 * discard data disk but write parity disk
6847 stripe
= stripe
* PAGE_SIZE
;
6848 /* Round up to power of 2, as discard handling
6849 * currently assumes that */
6850 while ((stripe
-1) & stripe
)
6851 stripe
= (stripe
| (stripe
-1)) + 1;
6852 mddev
->queue
->limits
.discard_alignment
= stripe
;
6853 mddev
->queue
->limits
.discard_granularity
= stripe
;
6855 * unaligned part of discard request will be ignored, so can't
6856 * guarantee discard_zeroes_data
6858 mddev
->queue
->limits
.discard_zeroes_data
= 0;
6860 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
6862 rdev_for_each(rdev
, mddev
) {
6863 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6864 rdev
->data_offset
<< 9);
6865 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6866 rdev
->new_data_offset
<< 9);
6868 * discard_zeroes_data is required, otherwise data
6869 * could be lost. Consider a scenario: discard a stripe
6870 * (the stripe could be inconsistent if
6871 * discard_zeroes_data is 0); write one disk of the
6872 * stripe (the stripe could be inconsistent again
6873 * depending on which disks are used to calculate
6874 * parity); the disk is broken; The stripe data of this
6877 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
6878 !bdev_get_queue(rdev
->bdev
)->
6879 limits
.discard_zeroes_data
)
6880 discard_supported
= false;
6881 /* Unfortunately, discard_zeroes_data is not currently
6882 * a guarantee - just a hint. So we only allow DISCARD
6883 * if the sysadmin has confirmed that only safe devices
6884 * are in use by setting a module parameter.
6886 if (!devices_handle_discard_safely
) {
6887 if (discard_supported
) {
6888 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6889 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6891 discard_supported
= false;
6895 if (discard_supported
&&
6896 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
6897 mddev
->queue
->limits
.discard_granularity
>= stripe
)
6898 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
6901 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
6907 md_unregister_thread(&mddev
->thread
);
6908 print_raid5_conf(conf
);
6910 mddev
->private = NULL
;
6911 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
6915 static void raid5_free(struct mddev
*mddev
, void *priv
)
6917 struct r5conf
*conf
= priv
;
6920 mddev
->to_remove
= &raid5_attrs_group
;
6923 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
6925 struct r5conf
*conf
= mddev
->private;
6928 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
6929 mddev
->chunk_sectors
/ 2, mddev
->layout
);
6930 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
6931 for (i
= 0; i
< conf
->raid_disks
; i
++)
6932 seq_printf (seq
, "%s",
6933 conf
->disks
[i
].rdev
&&
6934 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
6935 seq_printf (seq
, "]");
6938 static void print_raid5_conf (struct r5conf
*conf
)
6941 struct disk_info
*tmp
;
6943 printk(KERN_DEBUG
"RAID conf printout:\n");
6945 printk("(conf==NULL)\n");
6948 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
6950 conf
->raid_disks
- conf
->mddev
->degraded
);
6952 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6953 char b
[BDEVNAME_SIZE
];
6954 tmp
= conf
->disks
+ i
;
6956 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
6957 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
6958 bdevname(tmp
->rdev
->bdev
, b
));
6962 static int raid5_spare_active(struct mddev
*mddev
)
6965 struct r5conf
*conf
= mddev
->private;
6966 struct disk_info
*tmp
;
6968 unsigned long flags
;
6970 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6971 tmp
= conf
->disks
+ i
;
6972 if (tmp
->replacement
6973 && tmp
->replacement
->recovery_offset
== MaxSector
6974 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
6975 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
6976 /* Replacement has just become active. */
6978 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
6981 /* Replaced device not technically faulty,
6982 * but we need to be sure it gets removed
6983 * and never re-added.
6985 set_bit(Faulty
, &tmp
->rdev
->flags
);
6986 sysfs_notify_dirent_safe(
6987 tmp
->rdev
->sysfs_state
);
6989 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
6990 } else if (tmp
->rdev
6991 && tmp
->rdev
->recovery_offset
== MaxSector
6992 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
6993 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
6995 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
6998 spin_lock_irqsave(&conf
->device_lock
, flags
);
6999 mddev
->degraded
= calc_degraded(conf
);
7000 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7001 print_raid5_conf(conf
);
7005 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7007 struct r5conf
*conf
= mddev
->private;
7009 int number
= rdev
->raid_disk
;
7010 struct md_rdev
**rdevp
;
7011 struct disk_info
*p
= conf
->disks
+ number
;
7013 print_raid5_conf(conf
);
7014 if (rdev
== p
->rdev
)
7016 else if (rdev
== p
->replacement
)
7017 rdevp
= &p
->replacement
;
7021 if (number
>= conf
->raid_disks
&&
7022 conf
->reshape_progress
== MaxSector
)
7023 clear_bit(In_sync
, &rdev
->flags
);
7025 if (test_bit(In_sync
, &rdev
->flags
) ||
7026 atomic_read(&rdev
->nr_pending
)) {
7030 /* Only remove non-faulty devices if recovery
7033 if (!test_bit(Faulty
, &rdev
->flags
) &&
7034 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7035 !has_failed(conf
) &&
7036 (!p
->replacement
|| p
->replacement
== rdev
) &&
7037 number
< conf
->raid_disks
) {
7043 if (atomic_read(&rdev
->nr_pending
)) {
7044 /* lost the race, try later */
7047 } else if (p
->replacement
) {
7048 /* We must have just cleared 'rdev' */
7049 p
->rdev
= p
->replacement
;
7050 clear_bit(Replacement
, &p
->replacement
->flags
);
7051 smp_mb(); /* Make sure other CPUs may see both as identical
7052 * but will never see neither - if they are careful
7054 p
->replacement
= NULL
;
7055 clear_bit(WantReplacement
, &rdev
->flags
);
7057 /* We might have just removed the Replacement as faulty-
7058 * clear the bit just in case
7060 clear_bit(WantReplacement
, &rdev
->flags
);
7063 print_raid5_conf(conf
);
7067 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7069 struct r5conf
*conf
= mddev
->private;
7072 struct disk_info
*p
;
7074 int last
= conf
->raid_disks
- 1;
7076 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7079 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7080 /* no point adding a device */
7083 if (rdev
->raid_disk
>= 0)
7084 first
= last
= rdev
->raid_disk
;
7087 * find the disk ... but prefer rdev->saved_raid_disk
7090 if (rdev
->saved_raid_disk
>= 0 &&
7091 rdev
->saved_raid_disk
>= first
&&
7092 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7093 first
= rdev
->saved_raid_disk
;
7095 for (disk
= first
; disk
<= last
; disk
++) {
7096 p
= conf
->disks
+ disk
;
7097 if (p
->rdev
== NULL
) {
7098 clear_bit(In_sync
, &rdev
->flags
);
7099 rdev
->raid_disk
= disk
;
7101 if (rdev
->saved_raid_disk
!= disk
)
7103 rcu_assign_pointer(p
->rdev
, rdev
);
7107 for (disk
= first
; disk
<= last
; disk
++) {
7108 p
= conf
->disks
+ disk
;
7109 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7110 p
->replacement
== NULL
) {
7111 clear_bit(In_sync
, &rdev
->flags
);
7112 set_bit(Replacement
, &rdev
->flags
);
7113 rdev
->raid_disk
= disk
;
7116 rcu_assign_pointer(p
->replacement
, rdev
);
7121 print_raid5_conf(conf
);
7125 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7127 /* no resync is happening, and there is enough space
7128 * on all devices, so we can resize.
7129 * We need to make sure resync covers any new space.
7130 * If the array is shrinking we should possibly wait until
7131 * any io in the removed space completes, but it hardly seems
7135 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
7136 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7137 if (mddev
->external_size
&&
7138 mddev
->array_sectors
> newsize
)
7140 if (mddev
->bitmap
) {
7141 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7145 md_set_array_sectors(mddev
, newsize
);
7146 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7147 revalidate_disk(mddev
->gendisk
);
7148 if (sectors
> mddev
->dev_sectors
&&
7149 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7150 mddev
->recovery_cp
= mddev
->dev_sectors
;
7151 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7153 mddev
->dev_sectors
= sectors
;
7154 mddev
->resync_max_sectors
= sectors
;
7158 static int check_stripe_cache(struct mddev
*mddev
)
7160 /* Can only proceed if there are plenty of stripe_heads.
7161 * We need a minimum of one full stripe,, and for sensible progress
7162 * it is best to have about 4 times that.
7163 * If we require 4 times, then the default 256 4K stripe_heads will
7164 * allow for chunk sizes up to 256K, which is probably OK.
7165 * If the chunk size is greater, user-space should request more
7166 * stripe_heads first.
7168 struct r5conf
*conf
= mddev
->private;
7169 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7170 > conf
->min_nr_stripes
||
7171 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7172 > conf
->min_nr_stripes
) {
7173 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7175 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7182 static int check_reshape(struct mddev
*mddev
)
7184 struct r5conf
*conf
= mddev
->private;
7186 if (mddev
->delta_disks
== 0 &&
7187 mddev
->new_layout
== mddev
->layout
&&
7188 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7189 return 0; /* nothing to do */
7190 if (has_failed(conf
))
7192 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7193 /* We might be able to shrink, but the devices must
7194 * be made bigger first.
7195 * For raid6, 4 is the minimum size.
7196 * Otherwise 2 is the minimum
7199 if (mddev
->level
== 6)
7201 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7205 if (!check_stripe_cache(mddev
))
7208 return resize_stripes(conf
, (conf
->previous_raid_disks
7209 + mddev
->delta_disks
));
7212 static int raid5_start_reshape(struct mddev
*mddev
)
7214 struct r5conf
*conf
= mddev
->private;
7215 struct md_rdev
*rdev
;
7217 unsigned long flags
;
7219 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7222 if (!check_stripe_cache(mddev
))
7225 if (has_failed(conf
))
7228 rdev_for_each(rdev
, mddev
) {
7229 if (!test_bit(In_sync
, &rdev
->flags
)
7230 && !test_bit(Faulty
, &rdev
->flags
))
7234 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7235 /* Not enough devices even to make a degraded array
7240 /* Refuse to reduce size of the array. Any reductions in
7241 * array size must be through explicit setting of array_size
7244 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7245 < mddev
->array_sectors
) {
7246 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
7247 "before number of disks\n", mdname(mddev
));
7251 atomic_set(&conf
->reshape_stripes
, 0);
7252 spin_lock_irq(&conf
->device_lock
);
7253 write_seqcount_begin(&conf
->gen_lock
);
7254 conf
->previous_raid_disks
= conf
->raid_disks
;
7255 conf
->raid_disks
+= mddev
->delta_disks
;
7256 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7257 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7258 conf
->prev_algo
= conf
->algorithm
;
7259 conf
->algorithm
= mddev
->new_layout
;
7261 /* Code that selects data_offset needs to see the generation update
7262 * if reshape_progress has been set - so a memory barrier needed.
7265 if (mddev
->reshape_backwards
)
7266 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7268 conf
->reshape_progress
= 0;
7269 conf
->reshape_safe
= conf
->reshape_progress
;
7270 write_seqcount_end(&conf
->gen_lock
);
7271 spin_unlock_irq(&conf
->device_lock
);
7273 /* Now make sure any requests that proceeded on the assumption
7274 * the reshape wasn't running - like Discard or Read - have
7277 mddev_suspend(mddev
);
7278 mddev_resume(mddev
);
7280 /* Add some new drives, as many as will fit.
7281 * We know there are enough to make the newly sized array work.
7282 * Don't add devices if we are reducing the number of
7283 * devices in the array. This is because it is not possible
7284 * to correctly record the "partially reconstructed" state of
7285 * such devices during the reshape and confusion could result.
7287 if (mddev
->delta_disks
>= 0) {
7288 rdev_for_each(rdev
, mddev
)
7289 if (rdev
->raid_disk
< 0 &&
7290 !test_bit(Faulty
, &rdev
->flags
)) {
7291 if (raid5_add_disk(mddev
, rdev
) == 0) {
7293 >= conf
->previous_raid_disks
)
7294 set_bit(In_sync
, &rdev
->flags
);
7296 rdev
->recovery_offset
= 0;
7298 if (sysfs_link_rdev(mddev
, rdev
))
7299 /* Failure here is OK */;
7301 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7302 && !test_bit(Faulty
, &rdev
->flags
)) {
7303 /* This is a spare that was manually added */
7304 set_bit(In_sync
, &rdev
->flags
);
7307 /* When a reshape changes the number of devices,
7308 * ->degraded is measured against the larger of the
7309 * pre and post number of devices.
7311 spin_lock_irqsave(&conf
->device_lock
, flags
);
7312 mddev
->degraded
= calc_degraded(conf
);
7313 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7315 mddev
->raid_disks
= conf
->raid_disks
;
7316 mddev
->reshape_position
= conf
->reshape_progress
;
7317 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7319 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7320 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7321 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7322 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7323 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7325 if (!mddev
->sync_thread
) {
7326 mddev
->recovery
= 0;
7327 spin_lock_irq(&conf
->device_lock
);
7328 write_seqcount_begin(&conf
->gen_lock
);
7329 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7330 mddev
->new_chunk_sectors
=
7331 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7332 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7333 rdev_for_each(rdev
, mddev
)
7334 rdev
->new_data_offset
= rdev
->data_offset
;
7336 conf
->generation
--;
7337 conf
->reshape_progress
= MaxSector
;
7338 mddev
->reshape_position
= MaxSector
;
7339 write_seqcount_end(&conf
->gen_lock
);
7340 spin_unlock_irq(&conf
->device_lock
);
7343 conf
->reshape_checkpoint
= jiffies
;
7344 md_wakeup_thread(mddev
->sync_thread
);
7345 md_new_event(mddev
);
7349 /* This is called from the reshape thread and should make any
7350 * changes needed in 'conf'
7352 static void end_reshape(struct r5conf
*conf
)
7355 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7356 struct md_rdev
*rdev
;
7358 spin_lock_irq(&conf
->device_lock
);
7359 conf
->previous_raid_disks
= conf
->raid_disks
;
7360 rdev_for_each(rdev
, conf
->mddev
)
7361 rdev
->data_offset
= rdev
->new_data_offset
;
7363 conf
->reshape_progress
= MaxSector
;
7364 spin_unlock_irq(&conf
->device_lock
);
7365 wake_up(&conf
->wait_for_overlap
);
7367 /* read-ahead size must cover two whole stripes, which is
7368 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7370 if (conf
->mddev
->queue
) {
7371 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7372 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7374 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
7375 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
7380 /* This is called from the raid5d thread with mddev_lock held.
7381 * It makes config changes to the device.
7383 static void raid5_finish_reshape(struct mddev
*mddev
)
7385 struct r5conf
*conf
= mddev
->private;
7387 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7389 if (mddev
->delta_disks
> 0) {
7390 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7391 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7392 revalidate_disk(mddev
->gendisk
);
7395 spin_lock_irq(&conf
->device_lock
);
7396 mddev
->degraded
= calc_degraded(conf
);
7397 spin_unlock_irq(&conf
->device_lock
);
7398 for (d
= conf
->raid_disks
;
7399 d
< conf
->raid_disks
- mddev
->delta_disks
;
7401 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
7403 clear_bit(In_sync
, &rdev
->flags
);
7404 rdev
= conf
->disks
[d
].replacement
;
7406 clear_bit(In_sync
, &rdev
->flags
);
7409 mddev
->layout
= conf
->algorithm
;
7410 mddev
->chunk_sectors
= conf
->chunk_sectors
;
7411 mddev
->reshape_position
= MaxSector
;
7412 mddev
->delta_disks
= 0;
7413 mddev
->reshape_backwards
= 0;
7417 static void raid5_quiesce(struct mddev
*mddev
, int state
)
7419 struct r5conf
*conf
= mddev
->private;
7422 case 2: /* resume for a suspend */
7423 wake_up(&conf
->wait_for_overlap
);
7426 case 1: /* stop all writes */
7427 lock_all_device_hash_locks_irq(conf
);
7428 /* '2' tells resync/reshape to pause so that all
7429 * active stripes can drain
7432 wait_event_cmd(conf
->wait_for_stripe
,
7433 atomic_read(&conf
->active_stripes
) == 0 &&
7434 atomic_read(&conf
->active_aligned_reads
) == 0,
7435 unlock_all_device_hash_locks_irq(conf
),
7436 lock_all_device_hash_locks_irq(conf
));
7438 unlock_all_device_hash_locks_irq(conf
);
7439 /* allow reshape to continue */
7440 wake_up(&conf
->wait_for_overlap
);
7443 case 0: /* re-enable writes */
7444 lock_all_device_hash_locks_irq(conf
);
7446 wake_up(&conf
->wait_for_stripe
);
7447 wake_up(&conf
->wait_for_overlap
);
7448 unlock_all_device_hash_locks_irq(conf
);
7453 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
7455 struct r0conf
*raid0_conf
= mddev
->private;
7458 /* for raid0 takeover only one zone is supported */
7459 if (raid0_conf
->nr_strip_zones
> 1) {
7460 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7462 return ERR_PTR(-EINVAL
);
7465 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
7466 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
7467 mddev
->dev_sectors
= sectors
;
7468 mddev
->new_level
= level
;
7469 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7470 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
7471 mddev
->raid_disks
+= 1;
7472 mddev
->delta_disks
= 1;
7473 /* make sure it will be not marked as dirty */
7474 mddev
->recovery_cp
= MaxSector
;
7476 return setup_conf(mddev
);
7479 static void *raid5_takeover_raid1(struct mddev
*mddev
)
7483 if (mddev
->raid_disks
!= 2 ||
7484 mddev
->degraded
> 1)
7485 return ERR_PTR(-EINVAL
);
7487 /* Should check if there are write-behind devices? */
7489 chunksect
= 64*2; /* 64K by default */
7491 /* The array must be an exact multiple of chunksize */
7492 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
7495 if ((chunksect
<<9) < STRIPE_SIZE
)
7496 /* array size does not allow a suitable chunk size */
7497 return ERR_PTR(-EINVAL
);
7499 mddev
->new_level
= 5;
7500 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7501 mddev
->new_chunk_sectors
= chunksect
;
7503 return setup_conf(mddev
);
7506 static void *raid5_takeover_raid6(struct mddev
*mddev
)
7510 switch (mddev
->layout
) {
7511 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7512 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
7514 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7515 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
7517 case ALGORITHM_LEFT_SYMMETRIC_6
:
7518 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7520 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7521 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
7523 case ALGORITHM_PARITY_0_6
:
7524 new_layout
= ALGORITHM_PARITY_0
;
7526 case ALGORITHM_PARITY_N
:
7527 new_layout
= ALGORITHM_PARITY_N
;
7530 return ERR_PTR(-EINVAL
);
7532 mddev
->new_level
= 5;
7533 mddev
->new_layout
= new_layout
;
7534 mddev
->delta_disks
= -1;
7535 mddev
->raid_disks
-= 1;
7536 return setup_conf(mddev
);
7539 static int raid5_check_reshape(struct mddev
*mddev
)
7541 /* For a 2-drive array, the layout and chunk size can be changed
7542 * immediately as not restriping is needed.
7543 * For larger arrays we record the new value - after validation
7544 * to be used by a reshape pass.
7546 struct r5conf
*conf
= mddev
->private;
7547 int new_chunk
= mddev
->new_chunk_sectors
;
7549 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
7551 if (new_chunk
> 0) {
7552 if (!is_power_of_2(new_chunk
))
7554 if (new_chunk
< (PAGE_SIZE
>>9))
7556 if (mddev
->array_sectors
& (new_chunk
-1))
7557 /* not factor of array size */
7561 /* They look valid */
7563 if (mddev
->raid_disks
== 2) {
7564 /* can make the change immediately */
7565 if (mddev
->new_layout
>= 0) {
7566 conf
->algorithm
= mddev
->new_layout
;
7567 mddev
->layout
= mddev
->new_layout
;
7569 if (new_chunk
> 0) {
7570 conf
->chunk_sectors
= new_chunk
;
7571 mddev
->chunk_sectors
= new_chunk
;
7573 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7574 md_wakeup_thread(mddev
->thread
);
7576 return check_reshape(mddev
);
7579 static int raid6_check_reshape(struct mddev
*mddev
)
7581 int new_chunk
= mddev
->new_chunk_sectors
;
7583 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
7585 if (new_chunk
> 0) {
7586 if (!is_power_of_2(new_chunk
))
7588 if (new_chunk
< (PAGE_SIZE
>> 9))
7590 if (mddev
->array_sectors
& (new_chunk
-1))
7591 /* not factor of array size */
7595 /* They look valid */
7596 return check_reshape(mddev
);
7599 static void *raid5_takeover(struct mddev
*mddev
)
7601 /* raid5 can take over:
7602 * raid0 - if there is only one strip zone - make it a raid4 layout
7603 * raid1 - if there are two drives. We need to know the chunk size
7604 * raid4 - trivial - just use a raid4 layout.
7605 * raid6 - Providing it is a *_6 layout
7607 if (mddev
->level
== 0)
7608 return raid45_takeover_raid0(mddev
, 5);
7609 if (mddev
->level
== 1)
7610 return raid5_takeover_raid1(mddev
);
7611 if (mddev
->level
== 4) {
7612 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7613 mddev
->new_level
= 5;
7614 return setup_conf(mddev
);
7616 if (mddev
->level
== 6)
7617 return raid5_takeover_raid6(mddev
);
7619 return ERR_PTR(-EINVAL
);
7622 static void *raid4_takeover(struct mddev
*mddev
)
7624 /* raid4 can take over:
7625 * raid0 - if there is only one strip zone
7626 * raid5 - if layout is right
7628 if (mddev
->level
== 0)
7629 return raid45_takeover_raid0(mddev
, 4);
7630 if (mddev
->level
== 5 &&
7631 mddev
->layout
== ALGORITHM_PARITY_N
) {
7632 mddev
->new_layout
= 0;
7633 mddev
->new_level
= 4;
7634 return setup_conf(mddev
);
7636 return ERR_PTR(-EINVAL
);
7639 static struct md_personality raid5_personality
;
7641 static void *raid6_takeover(struct mddev
*mddev
)
7643 /* Currently can only take over a raid5. We map the
7644 * personality to an equivalent raid6 personality
7645 * with the Q block at the end.
7649 if (mddev
->pers
!= &raid5_personality
)
7650 return ERR_PTR(-EINVAL
);
7651 if (mddev
->degraded
> 1)
7652 return ERR_PTR(-EINVAL
);
7653 if (mddev
->raid_disks
> 253)
7654 return ERR_PTR(-EINVAL
);
7655 if (mddev
->raid_disks
< 3)
7656 return ERR_PTR(-EINVAL
);
7658 switch (mddev
->layout
) {
7659 case ALGORITHM_LEFT_ASYMMETRIC
:
7660 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
7662 case ALGORITHM_RIGHT_ASYMMETRIC
:
7663 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
7665 case ALGORITHM_LEFT_SYMMETRIC
:
7666 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
7668 case ALGORITHM_RIGHT_SYMMETRIC
:
7669 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
7671 case ALGORITHM_PARITY_0
:
7672 new_layout
= ALGORITHM_PARITY_0_6
;
7674 case ALGORITHM_PARITY_N
:
7675 new_layout
= ALGORITHM_PARITY_N
;
7678 return ERR_PTR(-EINVAL
);
7680 mddev
->new_level
= 6;
7681 mddev
->new_layout
= new_layout
;
7682 mddev
->delta_disks
= 1;
7683 mddev
->raid_disks
+= 1;
7684 return setup_conf(mddev
);
7687 static struct md_personality raid6_personality
=
7691 .owner
= THIS_MODULE
,
7692 .make_request
= make_request
,
7696 .error_handler
= error
,
7697 .hot_add_disk
= raid5_add_disk
,
7698 .hot_remove_disk
= raid5_remove_disk
,
7699 .spare_active
= raid5_spare_active
,
7700 .sync_request
= sync_request
,
7701 .resize
= raid5_resize
,
7703 .check_reshape
= raid6_check_reshape
,
7704 .start_reshape
= raid5_start_reshape
,
7705 .finish_reshape
= raid5_finish_reshape
,
7706 .quiesce
= raid5_quiesce
,
7707 .takeover
= raid6_takeover
,
7708 .congested
= raid5_congested
,
7709 .mergeable_bvec
= raid5_mergeable_bvec
,
7711 static struct md_personality raid5_personality
=
7715 .owner
= THIS_MODULE
,
7716 .make_request
= make_request
,
7720 .error_handler
= error
,
7721 .hot_add_disk
= raid5_add_disk
,
7722 .hot_remove_disk
= raid5_remove_disk
,
7723 .spare_active
= raid5_spare_active
,
7724 .sync_request
= sync_request
,
7725 .resize
= raid5_resize
,
7727 .check_reshape
= raid5_check_reshape
,
7728 .start_reshape
= raid5_start_reshape
,
7729 .finish_reshape
= raid5_finish_reshape
,
7730 .quiesce
= raid5_quiesce
,
7731 .takeover
= raid5_takeover
,
7732 .congested
= raid5_congested
,
7733 .mergeable_bvec
= raid5_mergeable_bvec
,
7736 static struct md_personality raid4_personality
=
7740 .owner
= THIS_MODULE
,
7741 .make_request
= make_request
,
7745 .error_handler
= error
,
7746 .hot_add_disk
= raid5_add_disk
,
7747 .hot_remove_disk
= raid5_remove_disk
,
7748 .spare_active
= raid5_spare_active
,
7749 .sync_request
= sync_request
,
7750 .resize
= raid5_resize
,
7752 .check_reshape
= raid5_check_reshape
,
7753 .start_reshape
= raid5_start_reshape
,
7754 .finish_reshape
= raid5_finish_reshape
,
7755 .quiesce
= raid5_quiesce
,
7756 .takeover
= raid4_takeover
,
7757 .congested
= raid5_congested
,
7758 .mergeable_bvec
= raid5_mergeable_bvec
,
7761 static int __init
raid5_init(void)
7763 raid5_wq
= alloc_workqueue("raid5wq",
7764 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
7767 register_md_personality(&raid6_personality
);
7768 register_md_personality(&raid5_personality
);
7769 register_md_personality(&raid4_personality
);
7773 static void raid5_exit(void)
7775 unregister_md_personality(&raid6_personality
);
7776 unregister_md_personality(&raid5_personality
);
7777 unregister_md_personality(&raid4_personality
);
7778 destroy_workqueue(raid5_wq
);
7781 module_init(raid5_init
);
7782 module_exit(raid5_exit
);
7783 MODULE_LICENSE("GPL");
7784 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7785 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7786 MODULE_ALIAS("md-raid5");
7787 MODULE_ALIAS("md-raid4");
7788 MODULE_ALIAS("md-level-5");
7789 MODULE_ALIAS("md-level-4");
7790 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7791 MODULE_ALIAS("md-raid6");
7792 MODULE_ALIAS("md-level-6");
7794 /* This used to be two separate modules, they were: */
7795 MODULE_ALIAS("raid5");
7796 MODULE_ALIAS("raid6");