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 unsigned long do_wakeup
= 0;
351 if (hash
== NR_STRIPE_HASH_LOCKS
) {
352 size
= NR_STRIPE_HASH_LOCKS
;
353 hash
= NR_STRIPE_HASH_LOCKS
- 1;
357 struct list_head
*list
= &temp_inactive_list
[size
- 1];
360 * We don't hold any lock here yet, get_active_stripe() might
361 * remove stripes from the list
363 if (!list_empty_careful(list
)) {
364 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
365 if (list_empty(conf
->inactive_list
+ hash
) &&
367 atomic_dec(&conf
->empty_inactive_list_nr
);
368 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
369 do_wakeup
|= 1 << hash
;
370 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
376 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++) {
377 if (do_wakeup
& (1 << i
))
378 wake_up(&conf
->wait_for_stripe
[i
]);
382 if (atomic_read(&conf
->active_stripes
) == 0)
383 wake_up(&conf
->wait_for_quiescent
);
384 if (conf
->retry_read_aligned
)
385 md_wakeup_thread(conf
->mddev
->thread
);
389 /* should hold conf->device_lock already */
390 static int release_stripe_list(struct r5conf
*conf
,
391 struct list_head
*temp_inactive_list
)
393 struct stripe_head
*sh
;
395 struct llist_node
*head
;
397 head
= llist_del_all(&conf
->released_stripes
);
398 head
= llist_reverse_order(head
);
402 sh
= llist_entry(head
, struct stripe_head
, release_list
);
403 head
= llist_next(head
);
404 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
406 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
408 * Don't worry the bit is set here, because if the bit is set
409 * again, the count is always > 1. This is true for
410 * STRIPE_ON_UNPLUG_LIST bit too.
412 hash
= sh
->hash_lock_index
;
413 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
420 static void release_stripe(struct stripe_head
*sh
)
422 struct r5conf
*conf
= sh
->raid_conf
;
424 struct list_head list
;
428 /* Avoid release_list until the last reference.
430 if (atomic_add_unless(&sh
->count
, -1, 1))
433 if (unlikely(!conf
->mddev
->thread
) ||
434 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
436 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
438 md_wakeup_thread(conf
->mddev
->thread
);
441 local_irq_save(flags
);
442 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
443 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
444 INIT_LIST_HEAD(&list
);
445 hash
= sh
->hash_lock_index
;
446 do_release_stripe(conf
, sh
, &list
);
447 spin_unlock(&conf
->device_lock
);
448 release_inactive_stripe_list(conf
, &list
, hash
);
450 local_irq_restore(flags
);
453 static inline void remove_hash(struct stripe_head
*sh
)
455 pr_debug("remove_hash(), stripe %llu\n",
456 (unsigned long long)sh
->sector
);
458 hlist_del_init(&sh
->hash
);
461 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
463 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
465 pr_debug("insert_hash(), stripe %llu\n",
466 (unsigned long long)sh
->sector
);
468 hlist_add_head(&sh
->hash
, hp
);
471 /* find an idle stripe, make sure it is unhashed, and return it. */
472 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
474 struct stripe_head
*sh
= NULL
;
475 struct list_head
*first
;
477 if (list_empty(conf
->inactive_list
+ hash
))
479 first
= (conf
->inactive_list
+ hash
)->next
;
480 sh
= list_entry(first
, struct stripe_head
, lru
);
481 list_del_init(first
);
483 atomic_inc(&conf
->active_stripes
);
484 BUG_ON(hash
!= sh
->hash_lock_index
);
485 if (list_empty(conf
->inactive_list
+ hash
))
486 atomic_inc(&conf
->empty_inactive_list_nr
);
491 static void shrink_buffers(struct stripe_head
*sh
)
495 int num
= sh
->raid_conf
->pool_size
;
497 for (i
= 0; i
< num
; i
++) {
498 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
502 sh
->dev
[i
].page
= NULL
;
507 static int grow_buffers(struct stripe_head
*sh
, gfp_t gfp
)
510 int num
= sh
->raid_conf
->pool_size
;
512 for (i
= 0; i
< num
; i
++) {
515 if (!(page
= alloc_page(gfp
))) {
518 sh
->dev
[i
].page
= page
;
519 sh
->dev
[i
].orig_page
= page
;
524 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
525 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
526 struct stripe_head
*sh
);
528 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
530 struct r5conf
*conf
= sh
->raid_conf
;
533 BUG_ON(atomic_read(&sh
->count
) != 0);
534 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
535 BUG_ON(stripe_operations_active(sh
));
536 BUG_ON(sh
->batch_head
);
538 pr_debug("init_stripe called, stripe %llu\n",
539 (unsigned long long)sector
);
541 seq
= read_seqcount_begin(&conf
->gen_lock
);
542 sh
->generation
= conf
->generation
- previous
;
543 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
545 stripe_set_idx(sector
, conf
, previous
, sh
);
548 for (i
= sh
->disks
; i
--; ) {
549 struct r5dev
*dev
= &sh
->dev
[i
];
551 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
552 test_bit(R5_LOCKED
, &dev
->flags
)) {
553 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
554 (unsigned long long)sh
->sector
, i
, dev
->toread
,
555 dev
->read
, dev
->towrite
, dev
->written
,
556 test_bit(R5_LOCKED
, &dev
->flags
));
560 raid5_build_block(sh
, i
, previous
);
562 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
564 sh
->overwrite_disks
= 0;
565 insert_hash(conf
, sh
);
566 sh
->cpu
= smp_processor_id();
567 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
570 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
573 struct stripe_head
*sh
;
575 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
576 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
577 if (sh
->sector
== sector
&& sh
->generation
== generation
)
579 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
584 * Need to check if array has failed when deciding whether to:
586 * - remove non-faulty devices
589 * This determination is simple when no reshape is happening.
590 * However if there is a reshape, we need to carefully check
591 * both the before and after sections.
592 * This is because some failed devices may only affect one
593 * of the two sections, and some non-in_sync devices may
594 * be insync in the section most affected by failed devices.
596 static int calc_degraded(struct r5conf
*conf
)
598 int degraded
, degraded2
;
603 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
604 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
605 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
606 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
607 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
609 else if (test_bit(In_sync
, &rdev
->flags
))
612 /* not in-sync or faulty.
613 * If the reshape increases the number of devices,
614 * this is being recovered by the reshape, so
615 * this 'previous' section is not in_sync.
616 * If the number of devices is being reduced however,
617 * the device can only be part of the array if
618 * we are reverting a reshape, so this section will
621 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
625 if (conf
->raid_disks
== conf
->previous_raid_disks
)
629 for (i
= 0; i
< conf
->raid_disks
; i
++) {
630 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
631 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
632 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
633 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
635 else if (test_bit(In_sync
, &rdev
->flags
))
638 /* not in-sync or faulty.
639 * If reshape increases the number of devices, this
640 * section has already been recovered, else it
641 * almost certainly hasn't.
643 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
647 if (degraded2
> degraded
)
652 static int has_failed(struct r5conf
*conf
)
656 if (conf
->mddev
->reshape_position
== MaxSector
)
657 return conf
->mddev
->degraded
> conf
->max_degraded
;
659 degraded
= calc_degraded(conf
);
660 if (degraded
> conf
->max_degraded
)
665 static struct stripe_head
*
666 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
667 int previous
, int noblock
, int noquiesce
)
669 struct stripe_head
*sh
;
670 int hash
= stripe_hash_locks_hash(sector
);
672 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
674 spin_lock_irq(conf
->hash_locks
+ hash
);
677 wait_event_lock_irq(conf
->wait_for_quiescent
,
678 conf
->quiesce
== 0 || noquiesce
,
679 *(conf
->hash_locks
+ hash
));
680 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
682 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
683 sh
= get_free_stripe(conf
, hash
);
684 if (!sh
&& !test_bit(R5_DID_ALLOC
,
686 set_bit(R5_ALLOC_MORE
,
689 if (noblock
&& sh
== NULL
)
692 set_bit(R5_INACTIVE_BLOCKED
,
694 wait_event_exclusive_cmd(
695 conf
->wait_for_stripe
[hash
],
696 !list_empty(conf
->inactive_list
+ hash
) &&
697 (atomic_read(&conf
->active_stripes
)
698 < (conf
->max_nr_stripes
* 3 / 4)
699 || !test_bit(R5_INACTIVE_BLOCKED
,
700 &conf
->cache_state
)),
701 spin_unlock_irq(conf
->hash_locks
+ hash
),
702 spin_lock_irq(conf
->hash_locks
+ hash
));
703 clear_bit(R5_INACTIVE_BLOCKED
,
706 init_stripe(sh
, sector
, previous
);
707 atomic_inc(&sh
->count
);
709 } else if (!atomic_inc_not_zero(&sh
->count
)) {
710 spin_lock(&conf
->device_lock
);
711 if (!atomic_read(&sh
->count
)) {
712 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
713 atomic_inc(&conf
->active_stripes
);
714 BUG_ON(list_empty(&sh
->lru
) &&
715 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
716 list_del_init(&sh
->lru
);
718 sh
->group
->stripes_cnt
--;
722 atomic_inc(&sh
->count
);
723 spin_unlock(&conf
->device_lock
);
725 } while (sh
== NULL
);
727 if (!list_empty(conf
->inactive_list
+ hash
))
728 wake_up(&conf
->wait_for_stripe
[hash
]);
730 spin_unlock_irq(conf
->hash_locks
+ hash
);
734 static bool is_full_stripe_write(struct stripe_head
*sh
)
736 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
737 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
740 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
744 spin_lock(&sh2
->stripe_lock
);
745 spin_lock_nested(&sh1
->stripe_lock
, 1);
747 spin_lock(&sh1
->stripe_lock
);
748 spin_lock_nested(&sh2
->stripe_lock
, 1);
752 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
754 spin_unlock(&sh1
->stripe_lock
);
755 spin_unlock(&sh2
->stripe_lock
);
759 /* Only freshly new full stripe normal write stripe can be added to a batch list */
760 static bool stripe_can_batch(struct stripe_head
*sh
)
762 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
763 !test_bit(STRIPE_BITMAP_PENDING
, &sh
->state
) &&
764 is_full_stripe_write(sh
);
767 /* we only do back search */
768 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
770 struct stripe_head
*head
;
771 sector_t head_sector
, tmp_sec
;
775 if (!stripe_can_batch(sh
))
777 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
778 tmp_sec
= sh
->sector
;
779 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
781 head_sector
= sh
->sector
- STRIPE_SECTORS
;
783 hash
= stripe_hash_locks_hash(head_sector
);
784 spin_lock_irq(conf
->hash_locks
+ hash
);
785 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
786 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
787 spin_lock(&conf
->device_lock
);
788 if (!atomic_read(&head
->count
)) {
789 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
790 atomic_inc(&conf
->active_stripes
);
791 BUG_ON(list_empty(&head
->lru
) &&
792 !test_bit(STRIPE_EXPANDING
, &head
->state
));
793 list_del_init(&head
->lru
);
795 head
->group
->stripes_cnt
--;
799 atomic_inc(&head
->count
);
800 spin_unlock(&conf
->device_lock
);
802 spin_unlock_irq(conf
->hash_locks
+ hash
);
806 if (!stripe_can_batch(head
))
809 lock_two_stripes(head
, sh
);
810 /* clear_batch_ready clear the flag */
811 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
818 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
820 if (head
->dev
[dd_idx
].towrite
->bi_rw
!= sh
->dev
[dd_idx
].towrite
->bi_rw
)
823 if (head
->batch_head
) {
824 spin_lock(&head
->batch_head
->batch_lock
);
825 /* This batch list is already running */
826 if (!stripe_can_batch(head
)) {
827 spin_unlock(&head
->batch_head
->batch_lock
);
832 * at this point, head's BATCH_READY could be cleared, but we
833 * can still add the stripe to batch list
835 list_add(&sh
->batch_list
, &head
->batch_list
);
836 spin_unlock(&head
->batch_head
->batch_lock
);
838 sh
->batch_head
= head
->batch_head
;
840 head
->batch_head
= head
;
841 sh
->batch_head
= head
->batch_head
;
842 spin_lock(&head
->batch_lock
);
843 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
844 spin_unlock(&head
->batch_lock
);
847 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
848 if (atomic_dec_return(&conf
->preread_active_stripes
)
850 md_wakeup_thread(conf
->mddev
->thread
);
852 if (test_and_clear_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
853 int seq
= sh
->bm_seq
;
854 if (test_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
) &&
855 sh
->batch_head
->bm_seq
> seq
)
856 seq
= sh
->batch_head
->bm_seq
;
857 set_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
);
858 sh
->batch_head
->bm_seq
= seq
;
861 atomic_inc(&sh
->count
);
863 unlock_two_stripes(head
, sh
);
865 release_stripe(head
);
868 /* Determine if 'data_offset' or 'new_data_offset' should be used
869 * in this stripe_head.
871 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
873 sector_t progress
= conf
->reshape_progress
;
874 /* Need a memory barrier to make sure we see the value
875 * of conf->generation, or ->data_offset that was set before
876 * reshape_progress was updated.
879 if (progress
== MaxSector
)
881 if (sh
->generation
== conf
->generation
- 1)
883 /* We are in a reshape, and this is a new-generation stripe,
884 * so use new_data_offset.
890 raid5_end_read_request(struct bio
*bi
);
892 raid5_end_write_request(struct bio
*bi
);
894 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
896 struct r5conf
*conf
= sh
->raid_conf
;
897 int i
, disks
= sh
->disks
;
898 struct stripe_head
*head_sh
= sh
;
902 for (i
= disks
; i
--; ) {
904 int replace_only
= 0;
905 struct bio
*bi
, *rbi
;
906 struct md_rdev
*rdev
, *rrdev
= NULL
;
909 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
910 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
914 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
916 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
918 else if (test_and_clear_bit(R5_WantReplace
,
919 &sh
->dev
[i
].flags
)) {
924 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
928 bi
= &sh
->dev
[i
].req
;
929 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
932 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
933 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
934 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
943 /* We raced and saw duplicates */
946 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
951 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
954 atomic_inc(&rdev
->nr_pending
);
955 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
958 atomic_inc(&rrdev
->nr_pending
);
961 /* We have already checked bad blocks for reads. Now
962 * need to check for writes. We never accept write errors
963 * on the replacement, so we don't to check rrdev.
965 while ((rw
& WRITE
) && rdev
&&
966 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
969 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
970 &first_bad
, &bad_sectors
);
975 set_bit(BlockedBadBlocks
, &rdev
->flags
);
976 if (!conf
->mddev
->external
&&
977 conf
->mddev
->flags
) {
978 /* It is very unlikely, but we might
979 * still need to write out the
980 * bad block log - better give it
982 md_check_recovery(conf
->mddev
);
985 * Because md_wait_for_blocked_rdev
986 * will dec nr_pending, we must
987 * increment it first.
989 atomic_inc(&rdev
->nr_pending
);
990 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
992 /* Acknowledged bad block - skip the write */
993 rdev_dec_pending(rdev
, conf
->mddev
);
999 if (s
->syncing
|| s
->expanding
|| s
->expanded
1001 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
1003 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1006 bi
->bi_bdev
= rdev
->bdev
;
1008 bi
->bi_end_io
= (rw
& WRITE
)
1009 ? raid5_end_write_request
1010 : raid5_end_read_request
;
1011 bi
->bi_private
= sh
;
1013 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1014 __func__
, (unsigned long long)sh
->sector
,
1016 atomic_inc(&sh
->count
);
1018 atomic_inc(&head_sh
->count
);
1019 if (use_new_offset(conf
, sh
))
1020 bi
->bi_iter
.bi_sector
= (sh
->sector
1021 + rdev
->new_data_offset
);
1023 bi
->bi_iter
.bi_sector
= (sh
->sector
1024 + rdev
->data_offset
);
1025 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1026 bi
->bi_rw
|= REQ_NOMERGE
;
1028 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1029 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1030 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1032 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1033 bi
->bi_io_vec
[0].bv_offset
= 0;
1034 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1036 * If this is discard request, set bi_vcnt 0. We don't
1037 * want to confuse SCSI because SCSI will replace payload
1039 if (rw
& REQ_DISCARD
)
1042 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1044 if (conf
->mddev
->gendisk
)
1045 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
1046 bi
, disk_devt(conf
->mddev
->gendisk
),
1048 generic_make_request(bi
);
1051 if (s
->syncing
|| s
->expanding
|| s
->expanded
1053 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1055 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1058 rbi
->bi_bdev
= rrdev
->bdev
;
1060 BUG_ON(!(rw
& WRITE
));
1061 rbi
->bi_end_io
= raid5_end_write_request
;
1062 rbi
->bi_private
= sh
;
1064 pr_debug("%s: for %llu schedule op %ld on "
1065 "replacement disc %d\n",
1066 __func__
, (unsigned long long)sh
->sector
,
1068 atomic_inc(&sh
->count
);
1070 atomic_inc(&head_sh
->count
);
1071 if (use_new_offset(conf
, sh
))
1072 rbi
->bi_iter
.bi_sector
= (sh
->sector
1073 + rrdev
->new_data_offset
);
1075 rbi
->bi_iter
.bi_sector
= (sh
->sector
1076 + rrdev
->data_offset
);
1077 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1078 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1079 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1081 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1082 rbi
->bi_io_vec
[0].bv_offset
= 0;
1083 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1085 * If this is discard request, set bi_vcnt 0. We don't
1086 * want to confuse SCSI because SCSI will replace payload
1088 if (rw
& REQ_DISCARD
)
1090 if (conf
->mddev
->gendisk
)
1091 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
1092 rbi
, disk_devt(conf
->mddev
->gendisk
),
1094 generic_make_request(rbi
);
1096 if (!rdev
&& !rrdev
) {
1098 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1099 pr_debug("skip op %ld on disc %d for sector %llu\n",
1100 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
1101 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1102 set_bit(STRIPE_HANDLE
, &sh
->state
);
1105 if (!head_sh
->batch_head
)
1107 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1114 static struct dma_async_tx_descriptor
*
1115 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1116 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1117 struct stripe_head
*sh
)
1120 struct bvec_iter iter
;
1121 struct page
*bio_page
;
1123 struct async_submit_ctl submit
;
1124 enum async_tx_flags flags
= 0;
1126 if (bio
->bi_iter
.bi_sector
>= sector
)
1127 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1129 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1132 flags
|= ASYNC_TX_FENCE
;
1133 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1135 bio_for_each_segment(bvl
, bio
, iter
) {
1136 int len
= bvl
.bv_len
;
1140 if (page_offset
< 0) {
1141 b_offset
= -page_offset
;
1142 page_offset
+= b_offset
;
1146 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1147 clen
= STRIPE_SIZE
- page_offset
;
1152 b_offset
+= bvl
.bv_offset
;
1153 bio_page
= bvl
.bv_page
;
1155 if (sh
->raid_conf
->skip_copy
&&
1156 b_offset
== 0 && page_offset
== 0 &&
1157 clen
== STRIPE_SIZE
)
1160 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1161 b_offset
, clen
, &submit
);
1163 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1164 page_offset
, clen
, &submit
);
1166 /* chain the operations */
1167 submit
.depend_tx
= tx
;
1169 if (clen
< len
) /* hit end of page */
1177 static void ops_complete_biofill(void *stripe_head_ref
)
1179 struct stripe_head
*sh
= stripe_head_ref
;
1180 struct bio
*return_bi
= NULL
;
1183 pr_debug("%s: stripe %llu\n", __func__
,
1184 (unsigned long long)sh
->sector
);
1186 /* clear completed biofills */
1187 for (i
= sh
->disks
; i
--; ) {
1188 struct r5dev
*dev
= &sh
->dev
[i
];
1190 /* acknowledge completion of a biofill operation */
1191 /* and check if we need to reply to a read request,
1192 * new R5_Wantfill requests are held off until
1193 * !STRIPE_BIOFILL_RUN
1195 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1196 struct bio
*rbi
, *rbi2
;
1201 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1202 dev
->sector
+ STRIPE_SECTORS
) {
1203 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1204 if (!raid5_dec_bi_active_stripes(rbi
)) {
1205 rbi
->bi_next
= return_bi
;
1212 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1214 return_io(return_bi
);
1216 set_bit(STRIPE_HANDLE
, &sh
->state
);
1220 static void ops_run_biofill(struct stripe_head
*sh
)
1222 struct dma_async_tx_descriptor
*tx
= NULL
;
1223 struct async_submit_ctl submit
;
1226 BUG_ON(sh
->batch_head
);
1227 pr_debug("%s: stripe %llu\n", __func__
,
1228 (unsigned long long)sh
->sector
);
1230 for (i
= sh
->disks
; i
--; ) {
1231 struct r5dev
*dev
= &sh
->dev
[i
];
1232 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1234 spin_lock_irq(&sh
->stripe_lock
);
1235 dev
->read
= rbi
= dev
->toread
;
1237 spin_unlock_irq(&sh
->stripe_lock
);
1238 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1239 dev
->sector
+ STRIPE_SECTORS
) {
1240 tx
= async_copy_data(0, rbi
, &dev
->page
,
1241 dev
->sector
, tx
, sh
);
1242 rbi
= r5_next_bio(rbi
, dev
->sector
);
1247 atomic_inc(&sh
->count
);
1248 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1249 async_trigger_callback(&submit
);
1252 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1259 tgt
= &sh
->dev
[target
];
1260 set_bit(R5_UPTODATE
, &tgt
->flags
);
1261 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1262 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1265 static void ops_complete_compute(void *stripe_head_ref
)
1267 struct stripe_head
*sh
= stripe_head_ref
;
1269 pr_debug("%s: stripe %llu\n", __func__
,
1270 (unsigned long long)sh
->sector
);
1272 /* mark the computed target(s) as uptodate */
1273 mark_target_uptodate(sh
, sh
->ops
.target
);
1274 mark_target_uptodate(sh
, sh
->ops
.target2
);
1276 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1277 if (sh
->check_state
== check_state_compute_run
)
1278 sh
->check_state
= check_state_compute_result
;
1279 set_bit(STRIPE_HANDLE
, &sh
->state
);
1283 /* return a pointer to the address conversion region of the scribble buffer */
1284 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1285 struct raid5_percpu
*percpu
, int i
)
1289 addr
= flex_array_get(percpu
->scribble
, i
);
1290 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1293 /* return a pointer to the address conversion region of the scribble buffer */
1294 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1298 addr
= flex_array_get(percpu
->scribble
, i
);
1302 static struct dma_async_tx_descriptor
*
1303 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1305 int disks
= sh
->disks
;
1306 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1307 int target
= sh
->ops
.target
;
1308 struct r5dev
*tgt
= &sh
->dev
[target
];
1309 struct page
*xor_dest
= tgt
->page
;
1311 struct dma_async_tx_descriptor
*tx
;
1312 struct async_submit_ctl submit
;
1315 BUG_ON(sh
->batch_head
);
1317 pr_debug("%s: stripe %llu block: %d\n",
1318 __func__
, (unsigned long long)sh
->sector
, target
);
1319 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1321 for (i
= disks
; i
--; )
1323 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1325 atomic_inc(&sh
->count
);
1327 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1328 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1329 if (unlikely(count
== 1))
1330 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1332 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1337 /* set_syndrome_sources - populate source buffers for gen_syndrome
1338 * @srcs - (struct page *) array of size sh->disks
1339 * @sh - stripe_head to parse
1341 * Populates srcs in proper layout order for the stripe and returns the
1342 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1343 * destination buffer is recorded in srcs[count] and the Q destination
1344 * is recorded in srcs[count+1]].
1346 static int set_syndrome_sources(struct page
**srcs
,
1347 struct stripe_head
*sh
,
1350 int disks
= sh
->disks
;
1351 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1352 int d0_idx
= raid6_d0(sh
);
1356 for (i
= 0; i
< disks
; i
++)
1362 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1363 struct r5dev
*dev
= &sh
->dev
[i
];
1365 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1366 (srctype
== SYNDROME_SRC_ALL
) ||
1367 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1368 test_bit(R5_Wantdrain
, &dev
->flags
)) ||
1369 (srctype
== SYNDROME_SRC_WRITTEN
&&
1371 srcs
[slot
] = sh
->dev
[i
].page
;
1372 i
= raid6_next_disk(i
, disks
);
1373 } while (i
!= d0_idx
);
1375 return syndrome_disks
;
1378 static struct dma_async_tx_descriptor
*
1379 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1381 int disks
= sh
->disks
;
1382 struct page
**blocks
= to_addr_page(percpu
, 0);
1384 int qd_idx
= sh
->qd_idx
;
1385 struct dma_async_tx_descriptor
*tx
;
1386 struct async_submit_ctl submit
;
1392 BUG_ON(sh
->batch_head
);
1393 if (sh
->ops
.target
< 0)
1394 target
= sh
->ops
.target2
;
1395 else if (sh
->ops
.target2
< 0)
1396 target
= sh
->ops
.target
;
1398 /* we should only have one valid target */
1401 pr_debug("%s: stripe %llu block: %d\n",
1402 __func__
, (unsigned long long)sh
->sector
, target
);
1404 tgt
= &sh
->dev
[target
];
1405 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1408 atomic_inc(&sh
->count
);
1410 if (target
== qd_idx
) {
1411 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1412 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1413 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1414 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1415 ops_complete_compute
, sh
,
1416 to_addr_conv(sh
, percpu
, 0));
1417 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1419 /* Compute any data- or p-drive using XOR */
1421 for (i
= disks
; i
-- ; ) {
1422 if (i
== target
|| i
== qd_idx
)
1424 blocks
[count
++] = sh
->dev
[i
].page
;
1427 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1428 NULL
, ops_complete_compute
, sh
,
1429 to_addr_conv(sh
, percpu
, 0));
1430 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1436 static struct dma_async_tx_descriptor
*
1437 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1439 int i
, count
, disks
= sh
->disks
;
1440 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1441 int d0_idx
= raid6_d0(sh
);
1442 int faila
= -1, failb
= -1;
1443 int target
= sh
->ops
.target
;
1444 int target2
= sh
->ops
.target2
;
1445 struct r5dev
*tgt
= &sh
->dev
[target
];
1446 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1447 struct dma_async_tx_descriptor
*tx
;
1448 struct page
**blocks
= to_addr_page(percpu
, 0);
1449 struct async_submit_ctl submit
;
1451 BUG_ON(sh
->batch_head
);
1452 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1453 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1454 BUG_ON(target
< 0 || target2
< 0);
1455 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1456 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1458 /* we need to open-code set_syndrome_sources to handle the
1459 * slot number conversion for 'faila' and 'failb'
1461 for (i
= 0; i
< disks
; i
++)
1466 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1468 blocks
[slot
] = sh
->dev
[i
].page
;
1474 i
= raid6_next_disk(i
, disks
);
1475 } while (i
!= d0_idx
);
1477 BUG_ON(faila
== failb
);
1480 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1481 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1483 atomic_inc(&sh
->count
);
1485 if (failb
== syndrome_disks
+1) {
1486 /* Q disk is one of the missing disks */
1487 if (faila
== syndrome_disks
) {
1488 /* Missing P+Q, just recompute */
1489 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1490 ops_complete_compute
, sh
,
1491 to_addr_conv(sh
, percpu
, 0));
1492 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1493 STRIPE_SIZE
, &submit
);
1497 int qd_idx
= sh
->qd_idx
;
1499 /* Missing D+Q: recompute D from P, then recompute Q */
1500 if (target
== qd_idx
)
1501 data_target
= target2
;
1503 data_target
= target
;
1506 for (i
= disks
; i
-- ; ) {
1507 if (i
== data_target
|| i
== qd_idx
)
1509 blocks
[count
++] = sh
->dev
[i
].page
;
1511 dest
= sh
->dev
[data_target
].page
;
1512 init_async_submit(&submit
,
1513 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1515 to_addr_conv(sh
, percpu
, 0));
1516 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1519 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1520 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1521 ops_complete_compute
, sh
,
1522 to_addr_conv(sh
, percpu
, 0));
1523 return async_gen_syndrome(blocks
, 0, count
+2,
1524 STRIPE_SIZE
, &submit
);
1527 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1528 ops_complete_compute
, sh
,
1529 to_addr_conv(sh
, percpu
, 0));
1530 if (failb
== syndrome_disks
) {
1531 /* We're missing D+P. */
1532 return async_raid6_datap_recov(syndrome_disks
+2,
1536 /* We're missing D+D. */
1537 return async_raid6_2data_recov(syndrome_disks
+2,
1538 STRIPE_SIZE
, faila
, failb
,
1544 static void ops_complete_prexor(void *stripe_head_ref
)
1546 struct stripe_head
*sh
= stripe_head_ref
;
1548 pr_debug("%s: stripe %llu\n", __func__
,
1549 (unsigned long long)sh
->sector
);
1552 static struct dma_async_tx_descriptor
*
1553 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1554 struct dma_async_tx_descriptor
*tx
)
1556 int disks
= sh
->disks
;
1557 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1558 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1559 struct async_submit_ctl submit
;
1561 /* existing parity data subtracted */
1562 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1564 BUG_ON(sh
->batch_head
);
1565 pr_debug("%s: stripe %llu\n", __func__
,
1566 (unsigned long long)sh
->sector
);
1568 for (i
= disks
; i
--; ) {
1569 struct r5dev
*dev
= &sh
->dev
[i
];
1570 /* Only process blocks that are known to be uptodate */
1571 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1572 xor_srcs
[count
++] = dev
->page
;
1575 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1576 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1577 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1582 static struct dma_async_tx_descriptor
*
1583 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1584 struct dma_async_tx_descriptor
*tx
)
1586 struct page
**blocks
= to_addr_page(percpu
, 0);
1588 struct async_submit_ctl submit
;
1590 pr_debug("%s: stripe %llu\n", __func__
,
1591 (unsigned long long)sh
->sector
);
1593 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1595 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1596 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1597 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1602 static struct dma_async_tx_descriptor
*
1603 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1605 int disks
= sh
->disks
;
1607 struct stripe_head
*head_sh
= sh
;
1609 pr_debug("%s: stripe %llu\n", __func__
,
1610 (unsigned long long)sh
->sector
);
1612 for (i
= disks
; i
--; ) {
1617 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1622 spin_lock_irq(&sh
->stripe_lock
);
1623 chosen
= dev
->towrite
;
1624 dev
->towrite
= NULL
;
1625 sh
->overwrite_disks
= 0;
1626 BUG_ON(dev
->written
);
1627 wbi
= dev
->written
= chosen
;
1628 spin_unlock_irq(&sh
->stripe_lock
);
1629 WARN_ON(dev
->page
!= dev
->orig_page
);
1631 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1632 dev
->sector
+ STRIPE_SECTORS
) {
1633 if (wbi
->bi_rw
& REQ_FUA
)
1634 set_bit(R5_WantFUA
, &dev
->flags
);
1635 if (wbi
->bi_rw
& REQ_SYNC
)
1636 set_bit(R5_SyncIO
, &dev
->flags
);
1637 if (wbi
->bi_rw
& REQ_DISCARD
)
1638 set_bit(R5_Discard
, &dev
->flags
);
1640 tx
= async_copy_data(1, wbi
, &dev
->page
,
1641 dev
->sector
, tx
, sh
);
1642 if (dev
->page
!= dev
->orig_page
) {
1643 set_bit(R5_SkipCopy
, &dev
->flags
);
1644 clear_bit(R5_UPTODATE
, &dev
->flags
);
1645 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1648 wbi
= r5_next_bio(wbi
, dev
->sector
);
1651 if (head_sh
->batch_head
) {
1652 sh
= list_first_entry(&sh
->batch_list
,
1665 static void ops_complete_reconstruct(void *stripe_head_ref
)
1667 struct stripe_head
*sh
= stripe_head_ref
;
1668 int disks
= sh
->disks
;
1669 int pd_idx
= sh
->pd_idx
;
1670 int qd_idx
= sh
->qd_idx
;
1672 bool fua
= false, sync
= false, discard
= false;
1674 pr_debug("%s: stripe %llu\n", __func__
,
1675 (unsigned long long)sh
->sector
);
1677 for (i
= disks
; i
--; ) {
1678 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1679 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1680 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1683 for (i
= disks
; i
--; ) {
1684 struct r5dev
*dev
= &sh
->dev
[i
];
1686 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1687 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1688 set_bit(R5_UPTODATE
, &dev
->flags
);
1690 set_bit(R5_WantFUA
, &dev
->flags
);
1692 set_bit(R5_SyncIO
, &dev
->flags
);
1696 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1697 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1698 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1699 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1701 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1702 sh
->reconstruct_state
= reconstruct_state_result
;
1705 set_bit(STRIPE_HANDLE
, &sh
->state
);
1710 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1711 struct dma_async_tx_descriptor
*tx
)
1713 int disks
= sh
->disks
;
1714 struct page
**xor_srcs
;
1715 struct async_submit_ctl submit
;
1716 int count
, pd_idx
= sh
->pd_idx
, i
;
1717 struct page
*xor_dest
;
1719 unsigned long flags
;
1721 struct stripe_head
*head_sh
= sh
;
1724 pr_debug("%s: stripe %llu\n", __func__
,
1725 (unsigned long long)sh
->sector
);
1727 for (i
= 0; i
< sh
->disks
; i
++) {
1730 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1733 if (i
>= sh
->disks
) {
1734 atomic_inc(&sh
->count
);
1735 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1736 ops_complete_reconstruct(sh
);
1741 xor_srcs
= to_addr_page(percpu
, j
);
1742 /* check if prexor is active which means only process blocks
1743 * that are part of a read-modify-write (written)
1745 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1747 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1748 for (i
= disks
; i
--; ) {
1749 struct r5dev
*dev
= &sh
->dev
[i
];
1750 if (head_sh
->dev
[i
].written
)
1751 xor_srcs
[count
++] = dev
->page
;
1754 xor_dest
= sh
->dev
[pd_idx
].page
;
1755 for (i
= disks
; i
--; ) {
1756 struct r5dev
*dev
= &sh
->dev
[i
];
1758 xor_srcs
[count
++] = dev
->page
;
1762 /* 1/ if we prexor'd then the dest is reused as a source
1763 * 2/ if we did not prexor then we are redoing the parity
1764 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1765 * for the synchronous xor case
1767 last_stripe
= !head_sh
->batch_head
||
1768 list_first_entry(&sh
->batch_list
,
1769 struct stripe_head
, batch_list
) == head_sh
;
1771 flags
= ASYNC_TX_ACK
|
1772 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1774 atomic_inc(&head_sh
->count
);
1775 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1776 to_addr_conv(sh
, percpu
, j
));
1778 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1779 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1780 to_addr_conv(sh
, percpu
, j
));
1783 if (unlikely(count
== 1))
1784 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1786 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1789 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1796 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1797 struct dma_async_tx_descriptor
*tx
)
1799 struct async_submit_ctl submit
;
1800 struct page
**blocks
;
1801 int count
, i
, j
= 0;
1802 struct stripe_head
*head_sh
= sh
;
1805 unsigned long txflags
;
1807 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1809 for (i
= 0; i
< sh
->disks
; i
++) {
1810 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1812 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1815 if (i
>= sh
->disks
) {
1816 atomic_inc(&sh
->count
);
1817 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1818 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1819 ops_complete_reconstruct(sh
);
1824 blocks
= to_addr_page(percpu
, j
);
1826 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1827 synflags
= SYNDROME_SRC_WRITTEN
;
1828 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1830 synflags
= SYNDROME_SRC_ALL
;
1831 txflags
= ASYNC_TX_ACK
;
1834 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1835 last_stripe
= !head_sh
->batch_head
||
1836 list_first_entry(&sh
->batch_list
,
1837 struct stripe_head
, batch_list
) == head_sh
;
1840 atomic_inc(&head_sh
->count
);
1841 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1842 head_sh
, to_addr_conv(sh
, percpu
, j
));
1844 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1845 to_addr_conv(sh
, percpu
, j
));
1846 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1849 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1855 static void ops_complete_check(void *stripe_head_ref
)
1857 struct stripe_head
*sh
= stripe_head_ref
;
1859 pr_debug("%s: stripe %llu\n", __func__
,
1860 (unsigned long long)sh
->sector
);
1862 sh
->check_state
= check_state_check_result
;
1863 set_bit(STRIPE_HANDLE
, &sh
->state
);
1867 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1869 int disks
= sh
->disks
;
1870 int pd_idx
= sh
->pd_idx
;
1871 int qd_idx
= sh
->qd_idx
;
1872 struct page
*xor_dest
;
1873 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1874 struct dma_async_tx_descriptor
*tx
;
1875 struct async_submit_ctl submit
;
1879 pr_debug("%s: stripe %llu\n", __func__
,
1880 (unsigned long long)sh
->sector
);
1882 BUG_ON(sh
->batch_head
);
1884 xor_dest
= sh
->dev
[pd_idx
].page
;
1885 xor_srcs
[count
++] = xor_dest
;
1886 for (i
= disks
; i
--; ) {
1887 if (i
== pd_idx
|| i
== qd_idx
)
1889 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1892 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1893 to_addr_conv(sh
, percpu
, 0));
1894 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1895 &sh
->ops
.zero_sum_result
, &submit
);
1897 atomic_inc(&sh
->count
);
1898 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1899 tx
= async_trigger_callback(&submit
);
1902 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1904 struct page
**srcs
= to_addr_page(percpu
, 0);
1905 struct async_submit_ctl submit
;
1908 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1909 (unsigned long long)sh
->sector
, checkp
);
1911 BUG_ON(sh
->batch_head
);
1912 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
1916 atomic_inc(&sh
->count
);
1917 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1918 sh
, to_addr_conv(sh
, percpu
, 0));
1919 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1920 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1923 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1925 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1926 struct dma_async_tx_descriptor
*tx
= NULL
;
1927 struct r5conf
*conf
= sh
->raid_conf
;
1928 int level
= conf
->level
;
1929 struct raid5_percpu
*percpu
;
1933 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1934 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1935 ops_run_biofill(sh
);
1939 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1941 tx
= ops_run_compute5(sh
, percpu
);
1943 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1944 tx
= ops_run_compute6_1(sh
, percpu
);
1946 tx
= ops_run_compute6_2(sh
, percpu
);
1948 /* terminate the chain if reconstruct is not set to be run */
1949 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1953 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
1955 tx
= ops_run_prexor5(sh
, percpu
, tx
);
1957 tx
= ops_run_prexor6(sh
, percpu
, tx
);
1960 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1961 tx
= ops_run_biodrain(sh
, tx
);
1965 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1967 ops_run_reconstruct5(sh
, percpu
, tx
);
1969 ops_run_reconstruct6(sh
, percpu
, tx
);
1972 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1973 if (sh
->check_state
== check_state_run
)
1974 ops_run_check_p(sh
, percpu
);
1975 else if (sh
->check_state
== check_state_run_q
)
1976 ops_run_check_pq(sh
, percpu
, 0);
1977 else if (sh
->check_state
== check_state_run_pq
)
1978 ops_run_check_pq(sh
, percpu
, 1);
1983 if (overlap_clear
&& !sh
->batch_head
)
1984 for (i
= disks
; i
--; ) {
1985 struct r5dev
*dev
= &sh
->dev
[i
];
1986 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1987 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1992 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
)
1994 struct stripe_head
*sh
;
1996 sh
= kmem_cache_zalloc(sc
, gfp
);
1998 spin_lock_init(&sh
->stripe_lock
);
1999 spin_lock_init(&sh
->batch_lock
);
2000 INIT_LIST_HEAD(&sh
->batch_list
);
2001 INIT_LIST_HEAD(&sh
->lru
);
2002 atomic_set(&sh
->count
, 1);
2006 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2008 struct stripe_head
*sh
;
2010 sh
= alloc_stripe(conf
->slab_cache
, gfp
);
2014 sh
->raid_conf
= conf
;
2016 if (grow_buffers(sh
, gfp
)) {
2018 kmem_cache_free(conf
->slab_cache
, sh
);
2021 sh
->hash_lock_index
=
2022 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2023 /* we just created an active stripe so... */
2024 atomic_inc(&conf
->active_stripes
);
2027 conf
->max_nr_stripes
++;
2031 static int grow_stripes(struct r5conf
*conf
, int num
)
2033 struct kmem_cache
*sc
;
2034 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2036 if (conf
->mddev
->gendisk
)
2037 sprintf(conf
->cache_name
[0],
2038 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2040 sprintf(conf
->cache_name
[0],
2041 "raid%d-%p", conf
->level
, conf
->mddev
);
2042 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2044 conf
->active_name
= 0;
2045 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2046 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2050 conf
->slab_cache
= sc
;
2051 conf
->pool_size
= devs
;
2053 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2060 * scribble_len - return the required size of the scribble region
2061 * @num - total number of disks in the array
2063 * The size must be enough to contain:
2064 * 1/ a struct page pointer for each device in the array +2
2065 * 2/ room to convert each entry in (1) to its corresponding dma
2066 * (dma_map_page()) or page (page_address()) address.
2068 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2069 * calculate over all devices (not just the data blocks), using zeros in place
2070 * of the P and Q blocks.
2072 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2074 struct flex_array
*ret
;
2077 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2078 ret
= flex_array_alloc(len
, cnt
, flags
);
2081 /* always prealloc all elements, so no locking is required */
2082 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2083 flex_array_free(ret
);
2089 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2094 mddev_suspend(conf
->mddev
);
2096 for_each_present_cpu(cpu
) {
2097 struct raid5_percpu
*percpu
;
2098 struct flex_array
*scribble
;
2100 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2101 scribble
= scribble_alloc(new_disks
,
2102 new_sectors
/ STRIPE_SECTORS
,
2106 flex_array_free(percpu
->scribble
);
2107 percpu
->scribble
= scribble
;
2114 mddev_resume(conf
->mddev
);
2118 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2120 /* Make all the stripes able to hold 'newsize' devices.
2121 * New slots in each stripe get 'page' set to a new page.
2123 * This happens in stages:
2124 * 1/ create a new kmem_cache and allocate the required number of
2126 * 2/ gather all the old stripe_heads and transfer the pages across
2127 * to the new stripe_heads. This will have the side effect of
2128 * freezing the array as once all stripe_heads have been collected,
2129 * no IO will be possible. Old stripe heads are freed once their
2130 * pages have been transferred over, and the old kmem_cache is
2131 * freed when all stripes are done.
2132 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2133 * we simple return a failre status - no need to clean anything up.
2134 * 4/ allocate new pages for the new slots in the new stripe_heads.
2135 * If this fails, we don't bother trying the shrink the
2136 * stripe_heads down again, we just leave them as they are.
2137 * As each stripe_head is processed the new one is released into
2140 * Once step2 is started, we cannot afford to wait for a write,
2141 * so we use GFP_NOIO allocations.
2143 struct stripe_head
*osh
, *nsh
;
2144 LIST_HEAD(newstripes
);
2145 struct disk_info
*ndisks
;
2147 struct kmem_cache
*sc
;
2151 if (newsize
<= conf
->pool_size
)
2152 return 0; /* never bother to shrink */
2154 err
= md_allow_write(conf
->mddev
);
2159 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2160 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2165 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2166 nsh
= alloc_stripe(sc
, GFP_KERNEL
);
2170 nsh
->raid_conf
= conf
;
2171 list_add(&nsh
->lru
, &newstripes
);
2174 /* didn't get enough, give up */
2175 while (!list_empty(&newstripes
)) {
2176 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2177 list_del(&nsh
->lru
);
2178 kmem_cache_free(sc
, nsh
);
2180 kmem_cache_destroy(sc
);
2183 /* Step 2 - Must use GFP_NOIO now.
2184 * OK, we have enough stripes, start collecting inactive
2185 * stripes and copying them over
2189 list_for_each_entry(nsh
, &newstripes
, lru
) {
2190 lock_device_hash_lock(conf
, hash
);
2191 wait_event_exclusive_cmd(conf
->wait_for_stripe
[hash
],
2192 !list_empty(conf
->inactive_list
+ hash
),
2193 unlock_device_hash_lock(conf
, hash
),
2194 lock_device_hash_lock(conf
, hash
));
2195 osh
= get_free_stripe(conf
, hash
);
2196 unlock_device_hash_lock(conf
, hash
);
2198 for(i
=0; i
<conf
->pool_size
; i
++) {
2199 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2200 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2202 nsh
->hash_lock_index
= hash
;
2203 kmem_cache_free(conf
->slab_cache
, osh
);
2205 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2206 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2211 kmem_cache_destroy(conf
->slab_cache
);
2214 * At this point, we are holding all the stripes so the array
2215 * is completely stalled, so now is a good time to resize
2216 * conf->disks and the scribble region
2218 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2220 for (i
=0; i
<conf
->raid_disks
; i
++)
2221 ndisks
[i
] = conf
->disks
[i
];
2223 conf
->disks
= ndisks
;
2227 /* Step 4, return new stripes to service */
2228 while(!list_empty(&newstripes
)) {
2229 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2230 list_del_init(&nsh
->lru
);
2232 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2233 if (nsh
->dev
[i
].page
== NULL
) {
2234 struct page
*p
= alloc_page(GFP_NOIO
);
2235 nsh
->dev
[i
].page
= p
;
2236 nsh
->dev
[i
].orig_page
= p
;
2240 release_stripe(nsh
);
2242 /* critical section pass, GFP_NOIO no longer needed */
2244 conf
->slab_cache
= sc
;
2245 conf
->active_name
= 1-conf
->active_name
;
2247 conf
->pool_size
= newsize
;
2251 static int drop_one_stripe(struct r5conf
*conf
)
2253 struct stripe_head
*sh
;
2254 int hash
= (conf
->max_nr_stripes
- 1) % NR_STRIPE_HASH_LOCKS
;
2256 spin_lock_irq(conf
->hash_locks
+ hash
);
2257 sh
= get_free_stripe(conf
, hash
);
2258 spin_unlock_irq(conf
->hash_locks
+ hash
);
2261 BUG_ON(atomic_read(&sh
->count
));
2263 kmem_cache_free(conf
->slab_cache
, sh
);
2264 atomic_dec(&conf
->active_stripes
);
2265 conf
->max_nr_stripes
--;
2269 static void shrink_stripes(struct r5conf
*conf
)
2271 while (conf
->max_nr_stripes
&&
2272 drop_one_stripe(conf
))
2275 if (conf
->slab_cache
)
2276 kmem_cache_destroy(conf
->slab_cache
);
2277 conf
->slab_cache
= NULL
;
2280 static void raid5_end_read_request(struct bio
* bi
)
2282 struct stripe_head
*sh
= bi
->bi_private
;
2283 struct r5conf
*conf
= sh
->raid_conf
;
2284 int disks
= sh
->disks
, i
;
2285 char b
[BDEVNAME_SIZE
];
2286 struct md_rdev
*rdev
= NULL
;
2289 for (i
=0 ; i
<disks
; i
++)
2290 if (bi
== &sh
->dev
[i
].req
)
2293 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2294 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2300 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2301 /* If replacement finished while this request was outstanding,
2302 * 'replacement' might be NULL already.
2303 * In that case it moved down to 'rdev'.
2304 * rdev is not removed until all requests are finished.
2306 rdev
= conf
->disks
[i
].replacement
;
2308 rdev
= conf
->disks
[i
].rdev
;
2310 if (use_new_offset(conf
, sh
))
2311 s
= sh
->sector
+ rdev
->new_data_offset
;
2313 s
= sh
->sector
+ rdev
->data_offset
;
2314 if (!bi
->bi_error
) {
2315 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2316 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2317 /* Note that this cannot happen on a
2318 * replacement device. We just fail those on
2323 "md/raid:%s: read error corrected"
2324 " (%lu sectors at %llu on %s)\n",
2325 mdname(conf
->mddev
), STRIPE_SECTORS
,
2326 (unsigned long long)s
,
2327 bdevname(rdev
->bdev
, b
));
2328 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2329 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2330 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2331 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2332 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2334 if (atomic_read(&rdev
->read_errors
))
2335 atomic_set(&rdev
->read_errors
, 0);
2337 const char *bdn
= bdevname(rdev
->bdev
, b
);
2341 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2342 atomic_inc(&rdev
->read_errors
);
2343 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2346 "md/raid:%s: read error on replacement device "
2347 "(sector %llu on %s).\n",
2348 mdname(conf
->mddev
),
2349 (unsigned long long)s
,
2351 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2355 "md/raid:%s: read error not correctable "
2356 "(sector %llu on %s).\n",
2357 mdname(conf
->mddev
),
2358 (unsigned long long)s
,
2360 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2365 "md/raid:%s: read error NOT corrected!! "
2366 "(sector %llu on %s).\n",
2367 mdname(conf
->mddev
),
2368 (unsigned long long)s
,
2370 } else if (atomic_read(&rdev
->read_errors
)
2371 > conf
->max_nr_stripes
)
2373 "md/raid:%s: Too many read errors, failing device %s.\n",
2374 mdname(conf
->mddev
), bdn
);
2377 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2378 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2381 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2382 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2383 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2385 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2387 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2388 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2390 && test_bit(In_sync
, &rdev
->flags
)
2391 && rdev_set_badblocks(
2392 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2393 md_error(conf
->mddev
, rdev
);
2396 rdev_dec_pending(rdev
, conf
->mddev
);
2397 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2398 set_bit(STRIPE_HANDLE
, &sh
->state
);
2402 static void raid5_end_write_request(struct bio
*bi
)
2404 struct stripe_head
*sh
= bi
->bi_private
;
2405 struct r5conf
*conf
= sh
->raid_conf
;
2406 int disks
= sh
->disks
, i
;
2407 struct md_rdev
*uninitialized_var(rdev
);
2410 int replacement
= 0;
2412 for (i
= 0 ; i
< disks
; i
++) {
2413 if (bi
== &sh
->dev
[i
].req
) {
2414 rdev
= conf
->disks
[i
].rdev
;
2417 if (bi
== &sh
->dev
[i
].rreq
) {
2418 rdev
= conf
->disks
[i
].replacement
;
2422 /* rdev was removed and 'replacement'
2423 * replaced it. rdev is not removed
2424 * until all requests are finished.
2426 rdev
= conf
->disks
[i
].rdev
;
2430 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2431 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2440 md_error(conf
->mddev
, rdev
);
2441 else if (is_badblock(rdev
, sh
->sector
,
2443 &first_bad
, &bad_sectors
))
2444 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2447 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2448 set_bit(WriteErrorSeen
, &rdev
->flags
);
2449 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2450 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2451 set_bit(MD_RECOVERY_NEEDED
,
2452 &rdev
->mddev
->recovery
);
2453 } else if (is_badblock(rdev
, sh
->sector
,
2455 &first_bad
, &bad_sectors
)) {
2456 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2457 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2458 /* That was a successful write so make
2459 * sure it looks like we already did
2462 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2465 rdev_dec_pending(rdev
, conf
->mddev
);
2467 if (sh
->batch_head
&& bi
->bi_error
&& !replacement
)
2468 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2470 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2471 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2472 set_bit(STRIPE_HANDLE
, &sh
->state
);
2475 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2476 release_stripe(sh
->batch_head
);
2479 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
2481 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
2483 struct r5dev
*dev
= &sh
->dev
[i
];
2485 bio_init(&dev
->req
);
2486 dev
->req
.bi_io_vec
= &dev
->vec
;
2487 dev
->req
.bi_max_vecs
= 1;
2488 dev
->req
.bi_private
= sh
;
2490 bio_init(&dev
->rreq
);
2491 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
2492 dev
->rreq
.bi_max_vecs
= 1;
2493 dev
->rreq
.bi_private
= sh
;
2496 dev
->sector
= compute_blocknr(sh
, i
, previous
);
2499 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2501 char b
[BDEVNAME_SIZE
];
2502 struct r5conf
*conf
= mddev
->private;
2503 unsigned long flags
;
2504 pr_debug("raid456: error called\n");
2506 spin_lock_irqsave(&conf
->device_lock
, flags
);
2507 clear_bit(In_sync
, &rdev
->flags
);
2508 mddev
->degraded
= calc_degraded(conf
);
2509 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2510 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2512 set_bit(Blocked
, &rdev
->flags
);
2513 set_bit(Faulty
, &rdev
->flags
);
2514 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2516 "md/raid:%s: Disk failure on %s, disabling device.\n"
2517 "md/raid:%s: Operation continuing on %d devices.\n",
2519 bdevname(rdev
->bdev
, b
),
2521 conf
->raid_disks
- mddev
->degraded
);
2525 * Input: a 'big' sector number,
2526 * Output: index of the data and parity disk, and the sector # in them.
2528 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2529 int previous
, int *dd_idx
,
2530 struct stripe_head
*sh
)
2532 sector_t stripe
, stripe2
;
2533 sector_t chunk_number
;
2534 unsigned int chunk_offset
;
2537 sector_t new_sector
;
2538 int algorithm
= previous
? conf
->prev_algo
2540 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2541 : conf
->chunk_sectors
;
2542 int raid_disks
= previous
? conf
->previous_raid_disks
2544 int data_disks
= raid_disks
- conf
->max_degraded
;
2546 /* First compute the information on this sector */
2549 * Compute the chunk number and the sector offset inside the chunk
2551 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2552 chunk_number
= r_sector
;
2555 * Compute the stripe number
2557 stripe
= chunk_number
;
2558 *dd_idx
= sector_div(stripe
, data_disks
);
2561 * Select the parity disk based on the user selected algorithm.
2563 pd_idx
= qd_idx
= -1;
2564 switch(conf
->level
) {
2566 pd_idx
= data_disks
;
2569 switch (algorithm
) {
2570 case ALGORITHM_LEFT_ASYMMETRIC
:
2571 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2572 if (*dd_idx
>= pd_idx
)
2575 case ALGORITHM_RIGHT_ASYMMETRIC
:
2576 pd_idx
= sector_div(stripe2
, raid_disks
);
2577 if (*dd_idx
>= pd_idx
)
2580 case ALGORITHM_LEFT_SYMMETRIC
:
2581 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2582 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2584 case ALGORITHM_RIGHT_SYMMETRIC
:
2585 pd_idx
= sector_div(stripe2
, raid_disks
);
2586 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2588 case ALGORITHM_PARITY_0
:
2592 case ALGORITHM_PARITY_N
:
2593 pd_idx
= data_disks
;
2601 switch (algorithm
) {
2602 case ALGORITHM_LEFT_ASYMMETRIC
:
2603 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2604 qd_idx
= pd_idx
+ 1;
2605 if (pd_idx
== raid_disks
-1) {
2606 (*dd_idx
)++; /* Q D D D P */
2608 } else if (*dd_idx
>= pd_idx
)
2609 (*dd_idx
) += 2; /* D D P Q D */
2611 case ALGORITHM_RIGHT_ASYMMETRIC
:
2612 pd_idx
= sector_div(stripe2
, raid_disks
);
2613 qd_idx
= pd_idx
+ 1;
2614 if (pd_idx
== raid_disks
-1) {
2615 (*dd_idx
)++; /* Q D D D P */
2617 } else if (*dd_idx
>= pd_idx
)
2618 (*dd_idx
) += 2; /* D D P Q D */
2620 case ALGORITHM_LEFT_SYMMETRIC
:
2621 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2622 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2623 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2625 case ALGORITHM_RIGHT_SYMMETRIC
:
2626 pd_idx
= sector_div(stripe2
, raid_disks
);
2627 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2628 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2631 case ALGORITHM_PARITY_0
:
2636 case ALGORITHM_PARITY_N
:
2637 pd_idx
= data_disks
;
2638 qd_idx
= data_disks
+ 1;
2641 case ALGORITHM_ROTATING_ZERO_RESTART
:
2642 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2643 * of blocks for computing Q is different.
2645 pd_idx
= sector_div(stripe2
, raid_disks
);
2646 qd_idx
= pd_idx
+ 1;
2647 if (pd_idx
== raid_disks
-1) {
2648 (*dd_idx
)++; /* Q D D D P */
2650 } else if (*dd_idx
>= pd_idx
)
2651 (*dd_idx
) += 2; /* D D P Q D */
2655 case ALGORITHM_ROTATING_N_RESTART
:
2656 /* Same a left_asymmetric, by first stripe is
2657 * D D D P Q rather than
2661 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2662 qd_idx
= pd_idx
+ 1;
2663 if (pd_idx
== raid_disks
-1) {
2664 (*dd_idx
)++; /* Q D D D P */
2666 } else if (*dd_idx
>= pd_idx
)
2667 (*dd_idx
) += 2; /* D D P Q D */
2671 case ALGORITHM_ROTATING_N_CONTINUE
:
2672 /* Same as left_symmetric but Q is before P */
2673 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2674 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2675 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2679 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2680 /* RAID5 left_asymmetric, with Q on last device */
2681 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2682 if (*dd_idx
>= pd_idx
)
2684 qd_idx
= raid_disks
- 1;
2687 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2688 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2689 if (*dd_idx
>= pd_idx
)
2691 qd_idx
= raid_disks
- 1;
2694 case ALGORITHM_LEFT_SYMMETRIC_6
:
2695 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2696 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2697 qd_idx
= raid_disks
- 1;
2700 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2701 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2702 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2703 qd_idx
= raid_disks
- 1;
2706 case ALGORITHM_PARITY_0_6
:
2709 qd_idx
= raid_disks
- 1;
2719 sh
->pd_idx
= pd_idx
;
2720 sh
->qd_idx
= qd_idx
;
2721 sh
->ddf_layout
= ddf_layout
;
2724 * Finally, compute the new sector number
2726 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2730 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2732 struct r5conf
*conf
= sh
->raid_conf
;
2733 int raid_disks
= sh
->disks
;
2734 int data_disks
= raid_disks
- conf
->max_degraded
;
2735 sector_t new_sector
= sh
->sector
, check
;
2736 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2737 : conf
->chunk_sectors
;
2738 int algorithm
= previous
? conf
->prev_algo
2742 sector_t chunk_number
;
2743 int dummy1
, dd_idx
= i
;
2745 struct stripe_head sh2
;
2747 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2748 stripe
= new_sector
;
2750 if (i
== sh
->pd_idx
)
2752 switch(conf
->level
) {
2755 switch (algorithm
) {
2756 case ALGORITHM_LEFT_ASYMMETRIC
:
2757 case ALGORITHM_RIGHT_ASYMMETRIC
:
2761 case ALGORITHM_LEFT_SYMMETRIC
:
2762 case ALGORITHM_RIGHT_SYMMETRIC
:
2765 i
-= (sh
->pd_idx
+ 1);
2767 case ALGORITHM_PARITY_0
:
2770 case ALGORITHM_PARITY_N
:
2777 if (i
== sh
->qd_idx
)
2778 return 0; /* It is the Q disk */
2779 switch (algorithm
) {
2780 case ALGORITHM_LEFT_ASYMMETRIC
:
2781 case ALGORITHM_RIGHT_ASYMMETRIC
:
2782 case ALGORITHM_ROTATING_ZERO_RESTART
:
2783 case ALGORITHM_ROTATING_N_RESTART
:
2784 if (sh
->pd_idx
== raid_disks
-1)
2785 i
--; /* Q D D D P */
2786 else if (i
> sh
->pd_idx
)
2787 i
-= 2; /* D D P Q D */
2789 case ALGORITHM_LEFT_SYMMETRIC
:
2790 case ALGORITHM_RIGHT_SYMMETRIC
:
2791 if (sh
->pd_idx
== raid_disks
-1)
2792 i
--; /* Q D D D P */
2797 i
-= (sh
->pd_idx
+ 2);
2800 case ALGORITHM_PARITY_0
:
2803 case ALGORITHM_PARITY_N
:
2805 case ALGORITHM_ROTATING_N_CONTINUE
:
2806 /* Like left_symmetric, but P is before Q */
2807 if (sh
->pd_idx
== 0)
2808 i
--; /* P D D D Q */
2813 i
-= (sh
->pd_idx
+ 1);
2816 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2817 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2821 case ALGORITHM_LEFT_SYMMETRIC_6
:
2822 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2824 i
+= data_disks
+ 1;
2825 i
-= (sh
->pd_idx
+ 1);
2827 case ALGORITHM_PARITY_0_6
:
2836 chunk_number
= stripe
* data_disks
+ i
;
2837 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2839 check
= raid5_compute_sector(conf
, r_sector
,
2840 previous
, &dummy1
, &sh2
);
2841 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2842 || sh2
.qd_idx
!= sh
->qd_idx
) {
2843 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2844 mdname(conf
->mddev
));
2851 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2852 int rcw
, int expand
)
2854 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
2855 struct r5conf
*conf
= sh
->raid_conf
;
2856 int level
= conf
->level
;
2860 for (i
= disks
; i
--; ) {
2861 struct r5dev
*dev
= &sh
->dev
[i
];
2864 set_bit(R5_LOCKED
, &dev
->flags
);
2865 set_bit(R5_Wantdrain
, &dev
->flags
);
2867 clear_bit(R5_UPTODATE
, &dev
->flags
);
2871 /* if we are not expanding this is a proper write request, and
2872 * there will be bios with new data to be drained into the
2877 /* False alarm, nothing to do */
2879 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2880 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2882 sh
->reconstruct_state
= reconstruct_state_run
;
2884 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2886 if (s
->locked
+ conf
->max_degraded
== disks
)
2887 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2888 atomic_inc(&conf
->pending_full_writes
);
2890 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2891 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2892 BUG_ON(level
== 6 &&
2893 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
2894 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
2896 for (i
= disks
; i
--; ) {
2897 struct r5dev
*dev
= &sh
->dev
[i
];
2898 if (i
== pd_idx
|| i
== qd_idx
)
2902 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2903 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2904 set_bit(R5_Wantdrain
, &dev
->flags
);
2905 set_bit(R5_LOCKED
, &dev
->flags
);
2906 clear_bit(R5_UPTODATE
, &dev
->flags
);
2911 /* False alarm - nothing to do */
2913 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2914 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2915 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2916 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2919 /* keep the parity disk(s) locked while asynchronous operations
2922 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2923 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2927 int qd_idx
= sh
->qd_idx
;
2928 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2930 set_bit(R5_LOCKED
, &dev
->flags
);
2931 clear_bit(R5_UPTODATE
, &dev
->flags
);
2935 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2936 __func__
, (unsigned long long)sh
->sector
,
2937 s
->locked
, s
->ops_request
);
2941 * Each stripe/dev can have one or more bion attached.
2942 * toread/towrite point to the first in a chain.
2943 * The bi_next chain must be in order.
2945 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
2946 int forwrite
, int previous
)
2949 struct r5conf
*conf
= sh
->raid_conf
;
2952 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2953 (unsigned long long)bi
->bi_iter
.bi_sector
,
2954 (unsigned long long)sh
->sector
);
2957 * If several bio share a stripe. The bio bi_phys_segments acts as a
2958 * reference count to avoid race. The reference count should already be
2959 * increased before this function is called (for example, in
2960 * make_request()), so other bio sharing this stripe will not free the
2961 * stripe. If a stripe is owned by one stripe, the stripe lock will
2964 spin_lock_irq(&sh
->stripe_lock
);
2965 /* Don't allow new IO added to stripes in batch list */
2969 bip
= &sh
->dev
[dd_idx
].towrite
;
2973 bip
= &sh
->dev
[dd_idx
].toread
;
2974 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
2975 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
2977 bip
= & (*bip
)->bi_next
;
2979 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
2982 if (!forwrite
|| previous
)
2983 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
2985 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2989 raid5_inc_bi_active_stripes(bi
);
2992 /* check if page is covered */
2993 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2994 for (bi
=sh
->dev
[dd_idx
].towrite
;
2995 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2996 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
2997 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2998 if (bio_end_sector(bi
) >= sector
)
2999 sector
= bio_end_sector(bi
);
3001 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
3002 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
3003 sh
->overwrite_disks
++;
3006 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3007 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
3008 (unsigned long long)sh
->sector
, dd_idx
);
3010 if (conf
->mddev
->bitmap
&& firstwrite
) {
3011 /* Cannot hold spinlock over bitmap_startwrite,
3012 * but must ensure this isn't added to a batch until
3013 * we have added to the bitmap and set bm_seq.
3014 * So set STRIPE_BITMAP_PENDING to prevent
3016 * If multiple add_stripe_bio() calls race here they
3017 * much all set STRIPE_BITMAP_PENDING. So only the first one
3018 * to complete "bitmap_startwrite" gets to set
3019 * STRIPE_BIT_DELAY. This is important as once a stripe
3020 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3023 set_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3024 spin_unlock_irq(&sh
->stripe_lock
);
3025 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
3027 spin_lock_irq(&sh
->stripe_lock
);
3028 clear_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3029 if (!sh
->batch_head
) {
3030 sh
->bm_seq
= conf
->seq_flush
+1;
3031 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
3034 spin_unlock_irq(&sh
->stripe_lock
);
3036 if (stripe_can_batch(sh
))
3037 stripe_add_to_batch_list(conf
, sh
);
3041 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
3042 spin_unlock_irq(&sh
->stripe_lock
);
3046 static void end_reshape(struct r5conf
*conf
);
3048 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3049 struct stripe_head
*sh
)
3051 int sectors_per_chunk
=
3052 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3054 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3055 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3057 raid5_compute_sector(conf
,
3058 stripe
* (disks
- conf
->max_degraded
)
3059 *sectors_per_chunk
+ chunk_offset
,
3065 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3066 struct stripe_head_state
*s
, int disks
,
3067 struct bio
**return_bi
)
3070 BUG_ON(sh
->batch_head
);
3071 for (i
= disks
; i
--; ) {
3075 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3076 struct md_rdev
*rdev
;
3078 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3079 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
3080 atomic_inc(&rdev
->nr_pending
);
3085 if (!rdev_set_badblocks(
3089 md_error(conf
->mddev
, rdev
);
3090 rdev_dec_pending(rdev
, conf
->mddev
);
3093 spin_lock_irq(&sh
->stripe_lock
);
3094 /* fail all writes first */
3095 bi
= sh
->dev
[i
].towrite
;
3096 sh
->dev
[i
].towrite
= NULL
;
3097 sh
->overwrite_disks
= 0;
3098 spin_unlock_irq(&sh
->stripe_lock
);
3102 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3103 wake_up(&conf
->wait_for_overlap
);
3105 while (bi
&& bi
->bi_iter
.bi_sector
<
3106 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3107 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3109 bi
->bi_error
= -EIO
;
3110 if (!raid5_dec_bi_active_stripes(bi
)) {
3111 md_write_end(conf
->mddev
);
3112 bi
->bi_next
= *return_bi
;
3118 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3119 STRIPE_SECTORS
, 0, 0);
3121 /* and fail all 'written' */
3122 bi
= sh
->dev
[i
].written
;
3123 sh
->dev
[i
].written
= NULL
;
3124 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3125 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3126 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3129 if (bi
) bitmap_end
= 1;
3130 while (bi
&& bi
->bi_iter
.bi_sector
<
3131 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3132 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3134 bi
->bi_error
= -EIO
;
3135 if (!raid5_dec_bi_active_stripes(bi
)) {
3136 md_write_end(conf
->mddev
);
3137 bi
->bi_next
= *return_bi
;
3143 /* fail any reads if this device is non-operational and
3144 * the data has not reached the cache yet.
3146 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3147 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3148 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3149 spin_lock_irq(&sh
->stripe_lock
);
3150 bi
= sh
->dev
[i
].toread
;
3151 sh
->dev
[i
].toread
= NULL
;
3152 spin_unlock_irq(&sh
->stripe_lock
);
3153 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3154 wake_up(&conf
->wait_for_overlap
);
3155 while (bi
&& bi
->bi_iter
.bi_sector
<
3156 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3157 struct bio
*nextbi
=
3158 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3160 bi
->bi_error
= -EIO
;
3161 if (!raid5_dec_bi_active_stripes(bi
)) {
3162 bi
->bi_next
= *return_bi
;
3169 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3170 STRIPE_SECTORS
, 0, 0);
3171 /* If we were in the middle of a write the parity block might
3172 * still be locked - so just clear all R5_LOCKED flags
3174 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3177 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3178 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3179 md_wakeup_thread(conf
->mddev
->thread
);
3183 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3184 struct stripe_head_state
*s
)
3189 BUG_ON(sh
->batch_head
);
3190 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3191 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3192 wake_up(&conf
->wait_for_overlap
);
3195 /* There is nothing more to do for sync/check/repair.
3196 * Don't even need to abort as that is handled elsewhere
3197 * if needed, and not always wanted e.g. if there is a known
3199 * For recover/replace we need to record a bad block on all
3200 * non-sync devices, or abort the recovery
3202 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3203 /* During recovery devices cannot be removed, so
3204 * locking and refcounting of rdevs is not needed
3206 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3207 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
3209 && !test_bit(Faulty
, &rdev
->flags
)
3210 && !test_bit(In_sync
, &rdev
->flags
)
3211 && !rdev_set_badblocks(rdev
, sh
->sector
,
3214 rdev
= conf
->disks
[i
].replacement
;
3216 && !test_bit(Faulty
, &rdev
->flags
)
3217 && !test_bit(In_sync
, &rdev
->flags
)
3218 && !rdev_set_badblocks(rdev
, sh
->sector
,
3223 conf
->recovery_disabled
=
3224 conf
->mddev
->recovery_disabled
;
3226 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3229 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3231 struct md_rdev
*rdev
;
3233 /* Doing recovery so rcu locking not required */
3234 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
3236 && !test_bit(Faulty
, &rdev
->flags
)
3237 && !test_bit(In_sync
, &rdev
->flags
)
3238 && (rdev
->recovery_offset
<= sh
->sector
3239 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3245 /* fetch_block - checks the given member device to see if its data needs
3246 * to be read or computed to satisfy a request.
3248 * Returns 1 when no more member devices need to be checked, otherwise returns
3249 * 0 to tell the loop in handle_stripe_fill to continue
3252 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3253 int disk_idx
, int disks
)
3255 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3256 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3257 &sh
->dev
[s
->failed_num
[1]] };
3261 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3262 test_bit(R5_UPTODATE
, &dev
->flags
))
3263 /* No point reading this as we already have it or have
3264 * decided to get it.
3269 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3270 /* We need this block to directly satisfy a request */
3273 if (s
->syncing
|| s
->expanding
||
3274 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3275 /* When syncing, or expanding we read everything.
3276 * When replacing, we need the replaced block.
3280 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3281 (s
->failed
>= 2 && fdev
[1]->toread
))
3282 /* If we want to read from a failed device, then
3283 * we need to actually read every other device.
3287 /* Sometimes neither read-modify-write nor reconstruct-write
3288 * cycles can work. In those cases we read every block we
3289 * can. Then the parity-update is certain to have enough to
3291 * This can only be a problem when we need to write something,
3292 * and some device has failed. If either of those tests
3293 * fail we need look no further.
3295 if (!s
->failed
|| !s
->to_write
)
3298 if (test_bit(R5_Insync
, &dev
->flags
) &&
3299 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3300 /* Pre-reads at not permitted until after short delay
3301 * to gather multiple requests. However if this
3302 * device is no Insync, the block could only be be computed
3303 * and there is no need to delay that.
3307 for (i
= 0; i
< s
->failed
; i
++) {
3308 if (fdev
[i
]->towrite
&&
3309 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3310 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3311 /* If we have a partial write to a failed
3312 * device, then we will need to reconstruct
3313 * the content of that device, so all other
3314 * devices must be read.
3319 /* If we are forced to do a reconstruct-write, either because
3320 * the current RAID6 implementation only supports that, or
3321 * or because parity cannot be trusted and we are currently
3322 * recovering it, there is extra need to be careful.
3323 * If one of the devices that we would need to read, because
3324 * it is not being overwritten (and maybe not written at all)
3325 * is missing/faulty, then we need to read everything we can.
3327 if (sh
->raid_conf
->level
!= 6 &&
3328 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3329 /* reconstruct-write isn't being forced */
3331 for (i
= 0; i
< s
->failed
; i
++) {
3332 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3333 s
->failed_num
[i
] != sh
->qd_idx
&&
3334 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3335 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3342 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3343 int disk_idx
, int disks
)
3345 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3347 /* is the data in this block needed, and can we get it? */
3348 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3349 /* we would like to get this block, possibly by computing it,
3350 * otherwise read it if the backing disk is insync
3352 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3353 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3354 BUG_ON(sh
->batch_head
);
3355 if ((s
->uptodate
== disks
- 1) &&
3356 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3357 disk_idx
== s
->failed_num
[1]))) {
3358 /* have disk failed, and we're requested to fetch it;
3361 pr_debug("Computing stripe %llu block %d\n",
3362 (unsigned long long)sh
->sector
, disk_idx
);
3363 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3364 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3365 set_bit(R5_Wantcompute
, &dev
->flags
);
3366 sh
->ops
.target
= disk_idx
;
3367 sh
->ops
.target2
= -1; /* no 2nd target */
3369 /* Careful: from this point on 'uptodate' is in the eye
3370 * of raid_run_ops which services 'compute' operations
3371 * before writes. R5_Wantcompute flags a block that will
3372 * be R5_UPTODATE by the time it is needed for a
3373 * subsequent operation.
3377 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3378 /* Computing 2-failure is *very* expensive; only
3379 * do it if failed >= 2
3382 for (other
= disks
; other
--; ) {
3383 if (other
== disk_idx
)
3385 if (!test_bit(R5_UPTODATE
,
3386 &sh
->dev
[other
].flags
))
3390 pr_debug("Computing stripe %llu blocks %d,%d\n",
3391 (unsigned long long)sh
->sector
,
3393 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3394 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3395 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3396 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3397 sh
->ops
.target
= disk_idx
;
3398 sh
->ops
.target2
= other
;
3402 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3403 set_bit(R5_LOCKED
, &dev
->flags
);
3404 set_bit(R5_Wantread
, &dev
->flags
);
3406 pr_debug("Reading block %d (sync=%d)\n",
3407 disk_idx
, s
->syncing
);
3415 * handle_stripe_fill - read or compute data to satisfy pending requests.
3417 static void handle_stripe_fill(struct stripe_head
*sh
,
3418 struct stripe_head_state
*s
,
3423 /* look for blocks to read/compute, skip this if a compute
3424 * is already in flight, or if the stripe contents are in the
3425 * midst of changing due to a write
3427 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3428 !sh
->reconstruct_state
)
3429 for (i
= disks
; i
--; )
3430 if (fetch_block(sh
, s
, i
, disks
))
3432 set_bit(STRIPE_HANDLE
, &sh
->state
);
3435 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
3436 unsigned long handle_flags
);
3437 /* handle_stripe_clean_event
3438 * any written block on an uptodate or failed drive can be returned.
3439 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3440 * never LOCKED, so we don't need to test 'failed' directly.
3442 static void handle_stripe_clean_event(struct r5conf
*conf
,
3443 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
3447 int discard_pending
= 0;
3448 struct stripe_head
*head_sh
= sh
;
3449 bool do_endio
= false;
3451 for (i
= disks
; i
--; )
3452 if (sh
->dev
[i
].written
) {
3454 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3455 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3456 test_bit(R5_Discard
, &dev
->flags
) ||
3457 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3458 /* We can return any write requests */
3459 struct bio
*wbi
, *wbi2
;
3460 pr_debug("Return write for disc %d\n", i
);
3461 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3462 clear_bit(R5_UPTODATE
, &dev
->flags
);
3463 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3464 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3469 dev
->page
= dev
->orig_page
;
3471 dev
->written
= NULL
;
3472 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3473 dev
->sector
+ STRIPE_SECTORS
) {
3474 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3475 if (!raid5_dec_bi_active_stripes(wbi
)) {
3476 md_write_end(conf
->mddev
);
3477 wbi
->bi_next
= *return_bi
;
3482 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3484 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3486 if (head_sh
->batch_head
) {
3487 sh
= list_first_entry(&sh
->batch_list
,
3490 if (sh
!= head_sh
) {
3497 } else if (test_bit(R5_Discard
, &dev
->flags
))
3498 discard_pending
= 1;
3499 WARN_ON(test_bit(R5_SkipCopy
, &dev
->flags
));
3500 WARN_ON(dev
->page
!= dev
->orig_page
);
3502 if (!discard_pending
&&
3503 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3504 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3505 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3506 if (sh
->qd_idx
>= 0) {
3507 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3508 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3510 /* now that discard is done we can proceed with any sync */
3511 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3513 * SCSI discard will change some bio fields and the stripe has
3514 * no updated data, so remove it from hash list and the stripe
3515 * will be reinitialized
3517 spin_lock_irq(&conf
->device_lock
);
3520 if (head_sh
->batch_head
) {
3521 sh
= list_first_entry(&sh
->batch_list
,
3522 struct stripe_head
, batch_list
);
3526 spin_unlock_irq(&conf
->device_lock
);
3529 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3530 set_bit(STRIPE_HANDLE
, &sh
->state
);
3534 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3535 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3536 md_wakeup_thread(conf
->mddev
->thread
);
3538 if (head_sh
->batch_head
&& do_endio
)
3539 break_stripe_batch_list(head_sh
, STRIPE_EXPAND_SYNC_FLAGS
);
3542 static void handle_stripe_dirtying(struct r5conf
*conf
,
3543 struct stripe_head
*sh
,
3544 struct stripe_head_state
*s
,
3547 int rmw
= 0, rcw
= 0, i
;
3548 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3550 /* Check whether resync is now happening or should start.
3551 * If yes, then the array is dirty (after unclean shutdown or
3552 * initial creation), so parity in some stripes might be inconsistent.
3553 * In this case, we need to always do reconstruct-write, to ensure
3554 * that in case of drive failure or read-error correction, we
3555 * generate correct data from the parity.
3557 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3558 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3560 /* Calculate the real rcw later - for now make it
3561 * look like rcw is cheaper
3564 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3565 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3566 (unsigned long long)sh
->sector
);
3567 } else for (i
= disks
; i
--; ) {
3568 /* would I have to read this buffer for read_modify_write */
3569 struct r5dev
*dev
= &sh
->dev
[i
];
3570 if ((dev
->towrite
|| 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
))
3577 rmw
+= 2*disks
; /* cannot read it */
3579 /* Would I have to read this buffer for reconstruct_write */
3580 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3581 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3582 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3583 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3584 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3585 if (test_bit(R5_Insync
, &dev
->flags
))
3591 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3592 (unsigned long long)sh
->sector
, rmw
, rcw
);
3593 set_bit(STRIPE_HANDLE
, &sh
->state
);
3594 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_ENABLE_RMW
)) && rmw
> 0) {
3595 /* prefer read-modify-write, but need to get some data */
3596 if (conf
->mddev
->queue
)
3597 blk_add_trace_msg(conf
->mddev
->queue
,
3598 "raid5 rmw %llu %d",
3599 (unsigned long long)sh
->sector
, rmw
);
3600 for (i
= disks
; i
--; ) {
3601 struct r5dev
*dev
= &sh
->dev
[i
];
3602 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3603 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3604 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3605 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3606 test_bit(R5_Insync
, &dev
->flags
)) {
3607 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3609 pr_debug("Read_old block %d for r-m-w\n",
3611 set_bit(R5_LOCKED
, &dev
->flags
);
3612 set_bit(R5_Wantread
, &dev
->flags
);
3615 set_bit(STRIPE_DELAYED
, &sh
->state
);
3616 set_bit(STRIPE_HANDLE
, &sh
->state
);
3621 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_ENABLE_RMW
)) && rcw
> 0) {
3622 /* want reconstruct write, but need to get some data */
3625 for (i
= disks
; i
--; ) {
3626 struct r5dev
*dev
= &sh
->dev
[i
];
3627 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3628 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3629 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3630 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3631 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3633 if (test_bit(R5_Insync
, &dev
->flags
) &&
3634 test_bit(STRIPE_PREREAD_ACTIVE
,
3636 pr_debug("Read_old block "
3637 "%d for Reconstruct\n", i
);
3638 set_bit(R5_LOCKED
, &dev
->flags
);
3639 set_bit(R5_Wantread
, &dev
->flags
);
3643 set_bit(STRIPE_DELAYED
, &sh
->state
);
3644 set_bit(STRIPE_HANDLE
, &sh
->state
);
3648 if (rcw
&& conf
->mddev
->queue
)
3649 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3650 (unsigned long long)sh
->sector
,
3651 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3654 if (rcw
> disks
&& rmw
> disks
&&
3655 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3656 set_bit(STRIPE_DELAYED
, &sh
->state
);
3658 /* now if nothing is locked, and if we have enough data,
3659 * we can start a write request
3661 /* since handle_stripe can be called at any time we need to handle the
3662 * case where a compute block operation has been submitted and then a
3663 * subsequent call wants to start a write request. raid_run_ops only
3664 * handles the case where compute block and reconstruct are requested
3665 * simultaneously. If this is not the case then new writes need to be
3666 * held off until the compute completes.
3668 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
3669 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
3670 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
3671 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
3674 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
3675 struct stripe_head_state
*s
, int disks
)
3677 struct r5dev
*dev
= NULL
;
3679 BUG_ON(sh
->batch_head
);
3680 set_bit(STRIPE_HANDLE
, &sh
->state
);
3682 switch (sh
->check_state
) {
3683 case check_state_idle
:
3684 /* start a new check operation if there are no failures */
3685 if (s
->failed
== 0) {
3686 BUG_ON(s
->uptodate
!= disks
);
3687 sh
->check_state
= check_state_run
;
3688 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3689 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3693 dev
= &sh
->dev
[s
->failed_num
[0]];
3695 case check_state_compute_result
:
3696 sh
->check_state
= check_state_idle
;
3698 dev
= &sh
->dev
[sh
->pd_idx
];
3700 /* check that a write has not made the stripe insync */
3701 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3704 /* either failed parity check, or recovery is happening */
3705 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3706 BUG_ON(s
->uptodate
!= disks
);
3708 set_bit(R5_LOCKED
, &dev
->flags
);
3710 set_bit(R5_Wantwrite
, &dev
->flags
);
3712 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3713 set_bit(STRIPE_INSYNC
, &sh
->state
);
3715 case check_state_run
:
3716 break; /* we will be called again upon completion */
3717 case check_state_check_result
:
3718 sh
->check_state
= check_state_idle
;
3720 /* if a failure occurred during the check operation, leave
3721 * STRIPE_INSYNC not set and let the stripe be handled again
3726 /* handle a successful check operation, if parity is correct
3727 * we are done. Otherwise update the mismatch count and repair
3728 * parity if !MD_RECOVERY_CHECK
3730 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3731 /* parity is correct (on disc,
3732 * not in buffer any more)
3734 set_bit(STRIPE_INSYNC
, &sh
->state
);
3736 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3737 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3738 /* don't try to repair!! */
3739 set_bit(STRIPE_INSYNC
, &sh
->state
);
3741 sh
->check_state
= check_state_compute_run
;
3742 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3743 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3744 set_bit(R5_Wantcompute
,
3745 &sh
->dev
[sh
->pd_idx
].flags
);
3746 sh
->ops
.target
= sh
->pd_idx
;
3747 sh
->ops
.target2
= -1;
3752 case check_state_compute_run
:
3755 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3756 __func__
, sh
->check_state
,
3757 (unsigned long long) sh
->sector
);
3762 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3763 struct stripe_head_state
*s
,
3766 int pd_idx
= sh
->pd_idx
;
3767 int qd_idx
= sh
->qd_idx
;
3770 BUG_ON(sh
->batch_head
);
3771 set_bit(STRIPE_HANDLE
, &sh
->state
);
3773 BUG_ON(s
->failed
> 2);
3775 /* Want to check and possibly repair P and Q.
3776 * However there could be one 'failed' device, in which
3777 * case we can only check one of them, possibly using the
3778 * other to generate missing data
3781 switch (sh
->check_state
) {
3782 case check_state_idle
:
3783 /* start a new check operation if there are < 2 failures */
3784 if (s
->failed
== s
->q_failed
) {
3785 /* The only possible failed device holds Q, so it
3786 * makes sense to check P (If anything else were failed,
3787 * we would have used P to recreate it).
3789 sh
->check_state
= check_state_run
;
3791 if (!s
->q_failed
&& s
->failed
< 2) {
3792 /* Q is not failed, and we didn't use it to generate
3793 * anything, so it makes sense to check it
3795 if (sh
->check_state
== check_state_run
)
3796 sh
->check_state
= check_state_run_pq
;
3798 sh
->check_state
= check_state_run_q
;
3801 /* discard potentially stale zero_sum_result */
3802 sh
->ops
.zero_sum_result
= 0;
3804 if (sh
->check_state
== check_state_run
) {
3805 /* async_xor_zero_sum destroys the contents of P */
3806 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3809 if (sh
->check_state
>= check_state_run
&&
3810 sh
->check_state
<= check_state_run_pq
) {
3811 /* async_syndrome_zero_sum preserves P and Q, so
3812 * no need to mark them !uptodate here
3814 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3818 /* we have 2-disk failure */
3819 BUG_ON(s
->failed
!= 2);
3821 case check_state_compute_result
:
3822 sh
->check_state
= check_state_idle
;
3824 /* check that a write has not made the stripe insync */
3825 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3828 /* now write out any block on a failed drive,
3829 * or P or Q if they were recomputed
3831 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3832 if (s
->failed
== 2) {
3833 dev
= &sh
->dev
[s
->failed_num
[1]];
3835 set_bit(R5_LOCKED
, &dev
->flags
);
3836 set_bit(R5_Wantwrite
, &dev
->flags
);
3838 if (s
->failed
>= 1) {
3839 dev
= &sh
->dev
[s
->failed_num
[0]];
3841 set_bit(R5_LOCKED
, &dev
->flags
);
3842 set_bit(R5_Wantwrite
, &dev
->flags
);
3844 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3845 dev
= &sh
->dev
[pd_idx
];
3847 set_bit(R5_LOCKED
, &dev
->flags
);
3848 set_bit(R5_Wantwrite
, &dev
->flags
);
3850 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3851 dev
= &sh
->dev
[qd_idx
];
3853 set_bit(R5_LOCKED
, &dev
->flags
);
3854 set_bit(R5_Wantwrite
, &dev
->flags
);
3856 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3858 set_bit(STRIPE_INSYNC
, &sh
->state
);
3860 case check_state_run
:
3861 case check_state_run_q
:
3862 case check_state_run_pq
:
3863 break; /* we will be called again upon completion */
3864 case check_state_check_result
:
3865 sh
->check_state
= check_state_idle
;
3867 /* handle a successful check operation, if parity is correct
3868 * we are done. Otherwise update the mismatch count and repair
3869 * parity if !MD_RECOVERY_CHECK
3871 if (sh
->ops
.zero_sum_result
== 0) {
3872 /* both parities are correct */
3874 set_bit(STRIPE_INSYNC
, &sh
->state
);
3876 /* in contrast to the raid5 case we can validate
3877 * parity, but still have a failure to write
3880 sh
->check_state
= check_state_compute_result
;
3881 /* Returning at this point means that we may go
3882 * off and bring p and/or q uptodate again so
3883 * we make sure to check zero_sum_result again
3884 * to verify if p or q need writeback
3888 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3889 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3890 /* don't try to repair!! */
3891 set_bit(STRIPE_INSYNC
, &sh
->state
);
3893 int *target
= &sh
->ops
.target
;
3895 sh
->ops
.target
= -1;
3896 sh
->ops
.target2
= -1;
3897 sh
->check_state
= check_state_compute_run
;
3898 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3899 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3900 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3901 set_bit(R5_Wantcompute
,
3902 &sh
->dev
[pd_idx
].flags
);
3904 target
= &sh
->ops
.target2
;
3907 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3908 set_bit(R5_Wantcompute
,
3909 &sh
->dev
[qd_idx
].flags
);
3916 case check_state_compute_run
:
3919 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3920 __func__
, sh
->check_state
,
3921 (unsigned long long) sh
->sector
);
3926 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3930 /* We have read all the blocks in this stripe and now we need to
3931 * copy some of them into a target stripe for expand.
3933 struct dma_async_tx_descriptor
*tx
= NULL
;
3934 BUG_ON(sh
->batch_head
);
3935 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3936 for (i
= 0; i
< sh
->disks
; i
++)
3937 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3939 struct stripe_head
*sh2
;
3940 struct async_submit_ctl submit
;
3942 sector_t bn
= compute_blocknr(sh
, i
, 1);
3943 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3945 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3947 /* so far only the early blocks of this stripe
3948 * have been requested. When later blocks
3949 * get requested, we will try again
3952 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3953 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3954 /* must have already done this block */
3955 release_stripe(sh2
);
3959 /* place all the copies on one channel */
3960 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3961 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3962 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3965 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3966 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3967 for (j
= 0; j
< conf
->raid_disks
; j
++)
3968 if (j
!= sh2
->pd_idx
&&
3970 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3972 if (j
== conf
->raid_disks
) {
3973 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3974 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3976 release_stripe(sh2
);
3979 /* done submitting copies, wait for them to complete */
3980 async_tx_quiesce(&tx
);
3984 * handle_stripe - do things to a stripe.
3986 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3987 * state of various bits to see what needs to be done.
3989 * return some read requests which now have data
3990 * return some write requests which are safely on storage
3991 * schedule a read on some buffers
3992 * schedule a write of some buffers
3993 * return confirmation of parity correctness
3997 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3999 struct r5conf
*conf
= sh
->raid_conf
;
4000 int disks
= sh
->disks
;
4003 int do_recovery
= 0;
4005 memset(s
, 0, sizeof(*s
));
4007 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
4008 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
4009 s
->failed_num
[0] = -1;
4010 s
->failed_num
[1] = -1;
4012 /* Now to look around and see what can be done */
4014 for (i
=disks
; i
--; ) {
4015 struct md_rdev
*rdev
;
4022 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4024 dev
->toread
, dev
->towrite
, dev
->written
);
4025 /* maybe we can reply to a read
4027 * new wantfill requests are only permitted while
4028 * ops_complete_biofill is guaranteed to be inactive
4030 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4031 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4032 set_bit(R5_Wantfill
, &dev
->flags
);
4034 /* now count some things */
4035 if (test_bit(R5_LOCKED
, &dev
->flags
))
4037 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4039 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4041 BUG_ON(s
->compute
> 2);
4044 if (test_bit(R5_Wantfill
, &dev
->flags
))
4046 else if (dev
->toread
)
4050 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4055 /* Prefer to use the replacement for reads, but only
4056 * if it is recovered enough and has no bad blocks.
4058 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4059 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4060 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4061 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4062 &first_bad
, &bad_sectors
))
4063 set_bit(R5_ReadRepl
, &dev
->flags
);
4066 set_bit(R5_NeedReplace
, &dev
->flags
);
4067 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4068 clear_bit(R5_ReadRepl
, &dev
->flags
);
4070 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4073 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4074 &first_bad
, &bad_sectors
);
4075 if (s
->blocked_rdev
== NULL
4076 && (test_bit(Blocked
, &rdev
->flags
)
4079 set_bit(BlockedBadBlocks
,
4081 s
->blocked_rdev
= rdev
;
4082 atomic_inc(&rdev
->nr_pending
);
4085 clear_bit(R5_Insync
, &dev
->flags
);
4089 /* also not in-sync */
4090 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4091 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4092 /* treat as in-sync, but with a read error
4093 * which we can now try to correct
4095 set_bit(R5_Insync
, &dev
->flags
);
4096 set_bit(R5_ReadError
, &dev
->flags
);
4098 } else if (test_bit(In_sync
, &rdev
->flags
))
4099 set_bit(R5_Insync
, &dev
->flags
);
4100 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4101 /* in sync if before recovery_offset */
4102 set_bit(R5_Insync
, &dev
->flags
);
4103 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4104 test_bit(R5_Expanded
, &dev
->flags
))
4105 /* If we've reshaped into here, we assume it is Insync.
4106 * We will shortly update recovery_offset to make
4109 set_bit(R5_Insync
, &dev
->flags
);
4111 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4112 /* This flag does not apply to '.replacement'
4113 * only to .rdev, so make sure to check that*/
4114 struct md_rdev
*rdev2
= rcu_dereference(
4115 conf
->disks
[i
].rdev
);
4117 clear_bit(R5_Insync
, &dev
->flags
);
4118 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4119 s
->handle_bad_blocks
= 1;
4120 atomic_inc(&rdev2
->nr_pending
);
4122 clear_bit(R5_WriteError
, &dev
->flags
);
4124 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4125 /* This flag does not apply to '.replacement'
4126 * only to .rdev, so make sure to check that*/
4127 struct md_rdev
*rdev2
= rcu_dereference(
4128 conf
->disks
[i
].rdev
);
4129 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4130 s
->handle_bad_blocks
= 1;
4131 atomic_inc(&rdev2
->nr_pending
);
4133 clear_bit(R5_MadeGood
, &dev
->flags
);
4135 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4136 struct md_rdev
*rdev2
= rcu_dereference(
4137 conf
->disks
[i
].replacement
);
4138 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4139 s
->handle_bad_blocks
= 1;
4140 atomic_inc(&rdev2
->nr_pending
);
4142 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4144 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4145 /* The ReadError flag will just be confusing now */
4146 clear_bit(R5_ReadError
, &dev
->flags
);
4147 clear_bit(R5_ReWrite
, &dev
->flags
);
4149 if (test_bit(R5_ReadError
, &dev
->flags
))
4150 clear_bit(R5_Insync
, &dev
->flags
);
4151 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4153 s
->failed_num
[s
->failed
] = i
;
4155 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4159 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4160 /* If there is a failed device being replaced,
4161 * we must be recovering.
4162 * else if we are after recovery_cp, we must be syncing
4163 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4164 * else we can only be replacing
4165 * sync and recovery both need to read all devices, and so
4166 * use the same flag.
4169 sh
->sector
>= conf
->mddev
->recovery_cp
||
4170 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4178 static int clear_batch_ready(struct stripe_head
*sh
)
4180 /* Return '1' if this is a member of batch, or
4181 * '0' if it is a lone stripe or a head which can now be
4184 struct stripe_head
*tmp
;
4185 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4186 return (sh
->batch_head
&& sh
->batch_head
!= sh
);
4187 spin_lock(&sh
->stripe_lock
);
4188 if (!sh
->batch_head
) {
4189 spin_unlock(&sh
->stripe_lock
);
4194 * this stripe could be added to a batch list before we check
4195 * BATCH_READY, skips it
4197 if (sh
->batch_head
!= sh
) {
4198 spin_unlock(&sh
->stripe_lock
);
4201 spin_lock(&sh
->batch_lock
);
4202 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4203 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4204 spin_unlock(&sh
->batch_lock
);
4205 spin_unlock(&sh
->stripe_lock
);
4208 * BATCH_READY is cleared, no new stripes can be added.
4209 * batch_list can be accessed without lock
4214 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4215 unsigned long handle_flags
)
4217 struct stripe_head
*sh
, *next
;
4221 list_for_each_entry_safe(sh
, next
, &head_sh
->batch_list
, batch_list
) {
4223 list_del_init(&sh
->batch_list
);
4225 WARN_ON_ONCE(sh
->state
& ((1 << STRIPE_ACTIVE
) |
4226 (1 << STRIPE_SYNCING
) |
4227 (1 << STRIPE_REPLACED
) |
4228 (1 << STRIPE_PREREAD_ACTIVE
) |
4229 (1 << STRIPE_DELAYED
) |
4230 (1 << STRIPE_BIT_DELAY
) |
4231 (1 << STRIPE_FULL_WRITE
) |
4232 (1 << STRIPE_BIOFILL_RUN
) |
4233 (1 << STRIPE_COMPUTE_RUN
) |
4234 (1 << STRIPE_OPS_REQ_PENDING
) |
4235 (1 << STRIPE_DISCARD
) |
4236 (1 << STRIPE_BATCH_READY
) |
4237 (1 << STRIPE_BATCH_ERR
) |
4238 (1 << STRIPE_BITMAP_PENDING
)));
4239 WARN_ON_ONCE(head_sh
->state
& ((1 << STRIPE_DISCARD
) |
4240 (1 << STRIPE_REPLACED
)));
4242 set_mask_bits(&sh
->state
, ~(STRIPE_EXPAND_SYNC_FLAGS
|
4243 (1 << STRIPE_DEGRADED
)),
4244 head_sh
->state
& (1 << STRIPE_INSYNC
));
4246 sh
->check_state
= head_sh
->check_state
;
4247 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4248 for (i
= 0; i
< sh
->disks
; i
++) {
4249 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
4251 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4252 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4254 spin_lock_irq(&sh
->stripe_lock
);
4255 sh
->batch_head
= NULL
;
4256 spin_unlock_irq(&sh
->stripe_lock
);
4257 if (handle_flags
== 0 ||
4258 sh
->state
& handle_flags
)
4259 set_bit(STRIPE_HANDLE
, &sh
->state
);
4262 spin_lock_irq(&head_sh
->stripe_lock
);
4263 head_sh
->batch_head
= NULL
;
4264 spin_unlock_irq(&head_sh
->stripe_lock
);
4265 for (i
= 0; i
< head_sh
->disks
; i
++)
4266 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
4268 if (head_sh
->state
& handle_flags
)
4269 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
4272 wake_up(&head_sh
->raid_conf
->wait_for_overlap
);
4275 static void handle_stripe(struct stripe_head
*sh
)
4277 struct stripe_head_state s
;
4278 struct r5conf
*conf
= sh
->raid_conf
;
4281 int disks
= sh
->disks
;
4282 struct r5dev
*pdev
, *qdev
;
4284 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4285 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4286 /* already being handled, ensure it gets handled
4287 * again when current action finishes */
4288 set_bit(STRIPE_HANDLE
, &sh
->state
);
4292 if (clear_batch_ready(sh
) ) {
4293 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4297 if (test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4298 break_stripe_batch_list(sh
, 0);
4300 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4301 spin_lock(&sh
->stripe_lock
);
4302 /* Cannot process 'sync' concurrently with 'discard' */
4303 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4304 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4305 set_bit(STRIPE_SYNCING
, &sh
->state
);
4306 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4307 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4309 spin_unlock(&sh
->stripe_lock
);
4311 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4313 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4314 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4315 (unsigned long long)sh
->sector
, sh
->state
,
4316 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4317 sh
->check_state
, sh
->reconstruct_state
);
4319 analyse_stripe(sh
, &s
);
4321 if (s
.handle_bad_blocks
) {
4322 set_bit(STRIPE_HANDLE
, &sh
->state
);
4326 if (unlikely(s
.blocked_rdev
)) {
4327 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4328 s
.replacing
|| s
.to_write
|| s
.written
) {
4329 set_bit(STRIPE_HANDLE
, &sh
->state
);
4332 /* There is nothing for the blocked_rdev to block */
4333 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4334 s
.blocked_rdev
= NULL
;
4337 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4338 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4339 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4342 pr_debug("locked=%d uptodate=%d to_read=%d"
4343 " to_write=%d failed=%d failed_num=%d,%d\n",
4344 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4345 s
.failed_num
[0], s
.failed_num
[1]);
4346 /* check if the array has lost more than max_degraded devices and,
4347 * if so, some requests might need to be failed.
4349 if (s
.failed
> conf
->max_degraded
) {
4350 sh
->check_state
= 0;
4351 sh
->reconstruct_state
= 0;
4352 break_stripe_batch_list(sh
, 0);
4353 if (s
.to_read
+s
.to_write
+s
.written
)
4354 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
4355 if (s
.syncing
+ s
.replacing
)
4356 handle_failed_sync(conf
, sh
, &s
);
4359 /* Now we check to see if any write operations have recently
4363 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4365 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4366 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4367 sh
->reconstruct_state
= reconstruct_state_idle
;
4369 /* All the 'written' buffers and the parity block are ready to
4370 * be written back to disk
4372 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4373 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4374 BUG_ON(sh
->qd_idx
>= 0 &&
4375 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4376 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4377 for (i
= disks
; i
--; ) {
4378 struct r5dev
*dev
= &sh
->dev
[i
];
4379 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4380 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4382 pr_debug("Writing block %d\n", i
);
4383 set_bit(R5_Wantwrite
, &dev
->flags
);
4388 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4389 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4391 set_bit(STRIPE_INSYNC
, &sh
->state
);
4394 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4395 s
.dec_preread_active
= 1;
4399 * might be able to return some write requests if the parity blocks
4400 * are safe, or on a failed drive
4402 pdev
= &sh
->dev
[sh
->pd_idx
];
4403 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4404 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4405 qdev
= &sh
->dev
[sh
->qd_idx
];
4406 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4407 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4411 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4412 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4413 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4414 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4415 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4416 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4417 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4418 test_bit(R5_Discard
, &qdev
->flags
))))))
4419 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
4421 /* Now we might consider reading some blocks, either to check/generate
4422 * parity, or to satisfy requests
4423 * or to load a block that is being partially written.
4425 if (s
.to_read
|| s
.non_overwrite
4426 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4427 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4430 handle_stripe_fill(sh
, &s
, disks
);
4432 /* Now to consider new write requests and what else, if anything
4433 * should be read. We do not handle new writes when:
4434 * 1/ A 'write' operation (copy+xor) is already in flight.
4435 * 2/ A 'check' operation is in flight, as it may clobber the parity
4438 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
4439 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4441 /* maybe we need to check and possibly fix the parity for this stripe
4442 * Any reads will already have been scheduled, so we just see if enough
4443 * data is available. The parity check is held off while parity
4444 * dependent operations are in flight.
4446 if (sh
->check_state
||
4447 (s
.syncing
&& s
.locked
== 0 &&
4448 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4449 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4450 if (conf
->level
== 6)
4451 handle_parity_checks6(conf
, sh
, &s
, disks
);
4453 handle_parity_checks5(conf
, sh
, &s
, disks
);
4456 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4457 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4458 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4459 /* Write out to replacement devices where possible */
4460 for (i
= 0; i
< conf
->raid_disks
; i
++)
4461 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4462 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4463 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4464 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4468 set_bit(STRIPE_INSYNC
, &sh
->state
);
4469 set_bit(STRIPE_REPLACED
, &sh
->state
);
4471 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4472 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4473 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4474 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4475 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4476 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4477 wake_up(&conf
->wait_for_overlap
);
4480 /* If the failed drives are just a ReadError, then we might need
4481 * to progress the repair/check process
4483 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4484 for (i
= 0; i
< s
.failed
; i
++) {
4485 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4486 if (test_bit(R5_ReadError
, &dev
->flags
)
4487 && !test_bit(R5_LOCKED
, &dev
->flags
)
4488 && test_bit(R5_UPTODATE
, &dev
->flags
)
4490 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4491 set_bit(R5_Wantwrite
, &dev
->flags
);
4492 set_bit(R5_ReWrite
, &dev
->flags
);
4493 set_bit(R5_LOCKED
, &dev
->flags
);
4496 /* let's read it back */
4497 set_bit(R5_Wantread
, &dev
->flags
);
4498 set_bit(R5_LOCKED
, &dev
->flags
);
4504 /* Finish reconstruct operations initiated by the expansion process */
4505 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4506 struct stripe_head
*sh_src
4507 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4508 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4509 /* sh cannot be written until sh_src has been read.
4510 * so arrange for sh to be delayed a little
4512 set_bit(STRIPE_DELAYED
, &sh
->state
);
4513 set_bit(STRIPE_HANDLE
, &sh
->state
);
4514 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4516 atomic_inc(&conf
->preread_active_stripes
);
4517 release_stripe(sh_src
);
4521 release_stripe(sh_src
);
4523 sh
->reconstruct_state
= reconstruct_state_idle
;
4524 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4525 for (i
= conf
->raid_disks
; i
--; ) {
4526 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4527 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4532 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4533 !sh
->reconstruct_state
) {
4534 /* Need to write out all blocks after computing parity */
4535 sh
->disks
= conf
->raid_disks
;
4536 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4537 schedule_reconstruction(sh
, &s
, 1, 1);
4538 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4539 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4540 atomic_dec(&conf
->reshape_stripes
);
4541 wake_up(&conf
->wait_for_overlap
);
4542 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4545 if (s
.expanding
&& s
.locked
== 0 &&
4546 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4547 handle_stripe_expansion(conf
, sh
);
4550 /* wait for this device to become unblocked */
4551 if (unlikely(s
.blocked_rdev
)) {
4552 if (conf
->mddev
->external
)
4553 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4556 /* Internal metadata will immediately
4557 * be written by raid5d, so we don't
4558 * need to wait here.
4560 rdev_dec_pending(s
.blocked_rdev
,
4564 if (s
.handle_bad_blocks
)
4565 for (i
= disks
; i
--; ) {
4566 struct md_rdev
*rdev
;
4567 struct r5dev
*dev
= &sh
->dev
[i
];
4568 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4569 /* We own a safe reference to the rdev */
4570 rdev
= conf
->disks
[i
].rdev
;
4571 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4573 md_error(conf
->mddev
, rdev
);
4574 rdev_dec_pending(rdev
, conf
->mddev
);
4576 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4577 rdev
= conf
->disks
[i
].rdev
;
4578 rdev_clear_badblocks(rdev
, sh
->sector
,
4580 rdev_dec_pending(rdev
, conf
->mddev
);
4582 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4583 rdev
= conf
->disks
[i
].replacement
;
4585 /* rdev have been moved down */
4586 rdev
= conf
->disks
[i
].rdev
;
4587 rdev_clear_badblocks(rdev
, sh
->sector
,
4589 rdev_dec_pending(rdev
, conf
->mddev
);
4594 raid_run_ops(sh
, s
.ops_request
);
4598 if (s
.dec_preread_active
) {
4599 /* We delay this until after ops_run_io so that if make_request
4600 * is waiting on a flush, it won't continue until the writes
4601 * have actually been submitted.
4603 atomic_dec(&conf
->preread_active_stripes
);
4604 if (atomic_read(&conf
->preread_active_stripes
) <
4606 md_wakeup_thread(conf
->mddev
->thread
);
4609 return_io(s
.return_bi
);
4611 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4614 static void raid5_activate_delayed(struct r5conf
*conf
)
4616 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
4617 while (!list_empty(&conf
->delayed_list
)) {
4618 struct list_head
*l
= conf
->delayed_list
.next
;
4619 struct stripe_head
*sh
;
4620 sh
= list_entry(l
, struct stripe_head
, lru
);
4622 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4623 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4624 atomic_inc(&conf
->preread_active_stripes
);
4625 list_add_tail(&sh
->lru
, &conf
->hold_list
);
4626 raid5_wakeup_stripe_thread(sh
);
4631 static void activate_bit_delay(struct r5conf
*conf
,
4632 struct list_head
*temp_inactive_list
)
4634 /* device_lock is held */
4635 struct list_head head
;
4636 list_add(&head
, &conf
->bitmap_list
);
4637 list_del_init(&conf
->bitmap_list
);
4638 while (!list_empty(&head
)) {
4639 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
4641 list_del_init(&sh
->lru
);
4642 atomic_inc(&sh
->count
);
4643 hash
= sh
->hash_lock_index
;
4644 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
4648 static int raid5_congested(struct mddev
*mddev
, int bits
)
4650 struct r5conf
*conf
= mddev
->private;
4652 /* No difference between reads and writes. Just check
4653 * how busy the stripe_cache is
4656 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
4660 if (atomic_read(&conf
->empty_inactive_list_nr
))
4666 /* We want read requests to align with chunks where possible,
4667 * but write requests don't need to.
4669 static int raid5_mergeable_bvec(struct mddev
*mddev
,
4670 struct bvec_merge_data
*bvm
,
4671 struct bio_vec
*biovec
)
4673 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
4675 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4676 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
4679 * always allow writes to be mergeable, read as well if array
4680 * is degraded as we'll go through stripe cache anyway.
4682 if ((bvm
->bi_rw
& 1) == WRITE
|| mddev
->degraded
)
4683 return biovec
->bv_len
;
4685 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4686 chunk_sectors
= mddev
->new_chunk_sectors
;
4687 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
4688 if (max
< 0) max
= 0;
4689 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
4690 return biovec
->bv_len
;
4695 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
4697 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
4698 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4699 unsigned int bio_sectors
= bio_sectors(bio
);
4701 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4702 chunk_sectors
= mddev
->new_chunk_sectors
;
4703 return chunk_sectors
>=
4704 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
4708 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4709 * later sampled by raid5d.
4711 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
4713 unsigned long flags
;
4715 spin_lock_irqsave(&conf
->device_lock
, flags
);
4717 bi
->bi_next
= conf
->retry_read_aligned_list
;
4718 conf
->retry_read_aligned_list
= bi
;
4720 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4721 md_wakeup_thread(conf
->mddev
->thread
);
4724 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
4728 bi
= conf
->retry_read_aligned
;
4730 conf
->retry_read_aligned
= NULL
;
4733 bi
= conf
->retry_read_aligned_list
;
4735 conf
->retry_read_aligned_list
= bi
->bi_next
;
4738 * this sets the active strip count to 1 and the processed
4739 * strip count to zero (upper 8 bits)
4741 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
4748 * The "raid5_align_endio" should check if the read succeeded and if it
4749 * did, call bio_endio on the original bio (having bio_put the new bio
4751 * If the read failed..
4753 static void raid5_align_endio(struct bio
*bi
)
4755 struct bio
* raid_bi
= bi
->bi_private
;
4756 struct mddev
*mddev
;
4757 struct r5conf
*conf
;
4758 struct md_rdev
*rdev
;
4759 int error
= bi
->bi_error
;
4763 rdev
= (void*)raid_bi
->bi_next
;
4764 raid_bi
->bi_next
= NULL
;
4765 mddev
= rdev
->mddev
;
4766 conf
= mddev
->private;
4768 rdev_dec_pending(rdev
, conf
->mddev
);
4771 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
4774 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4775 wake_up(&conf
->wait_for_quiescent
);
4779 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4781 add_bio_to_retry(raid_bi
, conf
);
4784 static int bio_fits_rdev(struct bio
*bi
)
4786 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
4788 if (bio_sectors(bi
) > queue_max_sectors(q
))
4790 blk_recount_segments(q
, bi
);
4791 if (bi
->bi_phys_segments
> queue_max_segments(q
))
4794 if (q
->merge_bvec_fn
)
4795 /* it's too hard to apply the merge_bvec_fn at this stage,
4803 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
4805 struct r5conf
*conf
= mddev
->private;
4807 struct bio
* align_bi
;
4808 struct md_rdev
*rdev
;
4809 sector_t end_sector
;
4811 if (!in_chunk_boundary(mddev
, raid_bio
)) {
4812 pr_debug("%s: non aligned\n", __func__
);
4816 * use bio_clone_mddev to make a copy of the bio
4818 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
4822 * set bi_end_io to a new function, and set bi_private to the
4825 align_bi
->bi_end_io
= raid5_align_endio
;
4826 align_bi
->bi_private
= raid_bio
;
4830 align_bi
->bi_iter
.bi_sector
=
4831 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
4834 end_sector
= bio_end_sector(align_bi
);
4836 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
4837 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
4838 rdev
->recovery_offset
< end_sector
) {
4839 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
4841 (test_bit(Faulty
, &rdev
->flags
) ||
4842 !(test_bit(In_sync
, &rdev
->flags
) ||
4843 rdev
->recovery_offset
>= end_sector
)))
4850 atomic_inc(&rdev
->nr_pending
);
4852 raid_bio
->bi_next
= (void*)rdev
;
4853 align_bi
->bi_bdev
= rdev
->bdev
;
4854 bio_clear_flag(align_bi
, BIO_SEG_VALID
);
4856 if (!bio_fits_rdev(align_bi
) ||
4857 is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
4858 bio_sectors(align_bi
),
4859 &first_bad
, &bad_sectors
)) {
4860 /* too big in some way, or has a known bad block */
4862 rdev_dec_pending(rdev
, mddev
);
4866 /* No reshape active, so we can trust rdev->data_offset */
4867 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
4869 spin_lock_irq(&conf
->device_lock
);
4870 wait_event_lock_irq(conf
->wait_for_quiescent
,
4873 atomic_inc(&conf
->active_aligned_reads
);
4874 spin_unlock_irq(&conf
->device_lock
);
4877 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4878 align_bi
, disk_devt(mddev
->gendisk
),
4879 raid_bio
->bi_iter
.bi_sector
);
4880 generic_make_request(align_bi
);
4889 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
4894 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
4895 unsigned chunk_sects
= mddev
->chunk_sectors
;
4896 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
4898 if (sectors
< bio_sectors(raid_bio
)) {
4899 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, fs_bio_set
);
4900 bio_chain(split
, raid_bio
);
4904 if (!raid5_read_one_chunk(mddev
, split
)) {
4905 if (split
!= raid_bio
)
4906 generic_make_request(raid_bio
);
4909 } while (split
!= raid_bio
);
4914 /* __get_priority_stripe - get the next stripe to process
4916 * Full stripe writes are allowed to pass preread active stripes up until
4917 * the bypass_threshold is exceeded. In general the bypass_count
4918 * increments when the handle_list is handled before the hold_list; however, it
4919 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4920 * stripe with in flight i/o. The bypass_count will be reset when the
4921 * head of the hold_list has changed, i.e. the head was promoted to the
4924 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
4926 struct stripe_head
*sh
= NULL
, *tmp
;
4927 struct list_head
*handle_list
= NULL
;
4928 struct r5worker_group
*wg
= NULL
;
4930 if (conf
->worker_cnt_per_group
== 0) {
4931 handle_list
= &conf
->handle_list
;
4932 } else if (group
!= ANY_GROUP
) {
4933 handle_list
= &conf
->worker_groups
[group
].handle_list
;
4934 wg
= &conf
->worker_groups
[group
];
4937 for (i
= 0; i
< conf
->group_cnt
; i
++) {
4938 handle_list
= &conf
->worker_groups
[i
].handle_list
;
4939 wg
= &conf
->worker_groups
[i
];
4940 if (!list_empty(handle_list
))
4945 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4947 list_empty(handle_list
) ? "empty" : "busy",
4948 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4949 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4951 if (!list_empty(handle_list
)) {
4952 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
4954 if (list_empty(&conf
->hold_list
))
4955 conf
->bypass_count
= 0;
4956 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4957 if (conf
->hold_list
.next
== conf
->last_hold
)
4958 conf
->bypass_count
++;
4960 conf
->last_hold
= conf
->hold_list
.next
;
4961 conf
->bypass_count
-= conf
->bypass_threshold
;
4962 if (conf
->bypass_count
< 0)
4963 conf
->bypass_count
= 0;
4966 } else if (!list_empty(&conf
->hold_list
) &&
4967 ((conf
->bypass_threshold
&&
4968 conf
->bypass_count
> conf
->bypass_threshold
) ||
4969 atomic_read(&conf
->pending_full_writes
) == 0)) {
4971 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
4972 if (conf
->worker_cnt_per_group
== 0 ||
4973 group
== ANY_GROUP
||
4974 !cpu_online(tmp
->cpu
) ||
4975 cpu_to_group(tmp
->cpu
) == group
) {
4982 conf
->bypass_count
-= conf
->bypass_threshold
;
4983 if (conf
->bypass_count
< 0)
4984 conf
->bypass_count
= 0;
4996 list_del_init(&sh
->lru
);
4997 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
5001 struct raid5_plug_cb
{
5002 struct blk_plug_cb cb
;
5003 struct list_head list
;
5004 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
5007 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
5009 struct raid5_plug_cb
*cb
= container_of(
5010 blk_cb
, struct raid5_plug_cb
, cb
);
5011 struct stripe_head
*sh
;
5012 struct mddev
*mddev
= cb
->cb
.data
;
5013 struct r5conf
*conf
= mddev
->private;
5017 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
5018 spin_lock_irq(&conf
->device_lock
);
5019 while (!list_empty(&cb
->list
)) {
5020 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
5021 list_del_init(&sh
->lru
);
5023 * avoid race release_stripe_plug() sees
5024 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5025 * is still in our list
5027 smp_mb__before_atomic();
5028 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
5030 * STRIPE_ON_RELEASE_LIST could be set here. In that
5031 * case, the count is always > 1 here
5033 hash
= sh
->hash_lock_index
;
5034 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
5037 spin_unlock_irq(&conf
->device_lock
);
5039 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
5040 NR_STRIPE_HASH_LOCKS
);
5042 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5046 static void release_stripe_plug(struct mddev
*mddev
,
5047 struct stripe_head
*sh
)
5049 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5050 raid5_unplug
, mddev
,
5051 sizeof(struct raid5_plug_cb
));
5052 struct raid5_plug_cb
*cb
;
5059 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5061 if (cb
->list
.next
== NULL
) {
5063 INIT_LIST_HEAD(&cb
->list
);
5064 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5065 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5068 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5069 list_add_tail(&sh
->lru
, &cb
->list
);
5074 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5076 struct r5conf
*conf
= mddev
->private;
5077 sector_t logical_sector
, last_sector
;
5078 struct stripe_head
*sh
;
5082 if (mddev
->reshape_position
!= MaxSector
)
5083 /* Skip discard while reshape is happening */
5086 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5087 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5090 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5092 stripe_sectors
= conf
->chunk_sectors
*
5093 (conf
->raid_disks
- conf
->max_degraded
);
5094 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5096 sector_div(last_sector
, stripe_sectors
);
5098 logical_sector
*= conf
->chunk_sectors
;
5099 last_sector
*= conf
->chunk_sectors
;
5101 for (; logical_sector
< last_sector
;
5102 logical_sector
+= STRIPE_SECTORS
) {
5106 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5107 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5108 TASK_UNINTERRUPTIBLE
);
5109 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5110 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5115 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5116 spin_lock_irq(&sh
->stripe_lock
);
5117 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5118 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5120 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5121 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5122 spin_unlock_irq(&sh
->stripe_lock
);
5128 set_bit(STRIPE_DISCARD
, &sh
->state
);
5129 finish_wait(&conf
->wait_for_overlap
, &w
);
5130 sh
->overwrite_disks
= 0;
5131 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5132 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5134 sh
->dev
[d
].towrite
= bi
;
5135 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5136 raid5_inc_bi_active_stripes(bi
);
5137 sh
->overwrite_disks
++;
5139 spin_unlock_irq(&sh
->stripe_lock
);
5140 if (conf
->mddev
->bitmap
) {
5142 d
< conf
->raid_disks
- conf
->max_degraded
;
5144 bitmap_startwrite(mddev
->bitmap
,
5148 sh
->bm_seq
= conf
->seq_flush
+ 1;
5149 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5152 set_bit(STRIPE_HANDLE
, &sh
->state
);
5153 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5154 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5155 atomic_inc(&conf
->preread_active_stripes
);
5156 release_stripe_plug(mddev
, sh
);
5159 remaining
= raid5_dec_bi_active_stripes(bi
);
5160 if (remaining
== 0) {
5161 md_write_end(mddev
);
5166 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
5168 struct r5conf
*conf
= mddev
->private;
5170 sector_t new_sector
;
5171 sector_t logical_sector
, last_sector
;
5172 struct stripe_head
*sh
;
5173 const int rw
= bio_data_dir(bi
);
5178 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
5179 md_flush_request(mddev
, bi
);
5183 md_write_start(mddev
, bi
);
5186 * If array is degraded, better not do chunk aligned read because
5187 * later we might have to read it again in order to reconstruct
5188 * data on failed drives.
5190 if (rw
== READ
&& mddev
->degraded
== 0 &&
5191 mddev
->reshape_position
== MaxSector
) {
5192 bi
= chunk_aligned_read(mddev
, bi
);
5197 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
5198 make_discard_request(mddev
, bi
);
5202 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5203 last_sector
= bio_end_sector(bi
);
5205 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5207 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5208 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5214 seq
= read_seqcount_begin(&conf
->gen_lock
);
5217 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5218 TASK_UNINTERRUPTIBLE
);
5219 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5220 /* spinlock is needed as reshape_progress may be
5221 * 64bit on a 32bit platform, and so it might be
5222 * possible to see a half-updated value
5223 * Of course reshape_progress could change after
5224 * the lock is dropped, so once we get a reference
5225 * to the stripe that we think it is, we will have
5228 spin_lock_irq(&conf
->device_lock
);
5229 if (mddev
->reshape_backwards
5230 ? logical_sector
< conf
->reshape_progress
5231 : logical_sector
>= conf
->reshape_progress
) {
5234 if (mddev
->reshape_backwards
5235 ? logical_sector
< conf
->reshape_safe
5236 : logical_sector
>= conf
->reshape_safe
) {
5237 spin_unlock_irq(&conf
->device_lock
);
5243 spin_unlock_irq(&conf
->device_lock
);
5246 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5249 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5250 (unsigned long long)new_sector
,
5251 (unsigned long long)logical_sector
);
5253 sh
= get_active_stripe(conf
, new_sector
, previous
,
5254 (bi
->bi_rw
&RWA_MASK
), 0);
5256 if (unlikely(previous
)) {
5257 /* expansion might have moved on while waiting for a
5258 * stripe, so we must do the range check again.
5259 * Expansion could still move past after this
5260 * test, but as we are holding a reference to
5261 * 'sh', we know that if that happens,
5262 * STRIPE_EXPANDING will get set and the expansion
5263 * won't proceed until we finish with the stripe.
5266 spin_lock_irq(&conf
->device_lock
);
5267 if (mddev
->reshape_backwards
5268 ? logical_sector
>= conf
->reshape_progress
5269 : logical_sector
< conf
->reshape_progress
)
5270 /* mismatch, need to try again */
5272 spin_unlock_irq(&conf
->device_lock
);
5280 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5281 /* Might have got the wrong stripe_head
5289 logical_sector
>= mddev
->suspend_lo
&&
5290 logical_sector
< mddev
->suspend_hi
) {
5292 /* As the suspend_* range is controlled by
5293 * userspace, we want an interruptible
5296 flush_signals(current
);
5297 prepare_to_wait(&conf
->wait_for_overlap
,
5298 &w
, TASK_INTERRUPTIBLE
);
5299 if (logical_sector
>= mddev
->suspend_lo
&&
5300 logical_sector
< mddev
->suspend_hi
) {
5307 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5308 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5309 /* Stripe is busy expanding or
5310 * add failed due to overlap. Flush everything
5313 md_wakeup_thread(mddev
->thread
);
5319 set_bit(STRIPE_HANDLE
, &sh
->state
);
5320 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5321 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5322 (bi
->bi_rw
& REQ_SYNC
) &&
5323 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5324 atomic_inc(&conf
->preread_active_stripes
);
5325 release_stripe_plug(mddev
, sh
);
5327 /* cannot get stripe for read-ahead, just give-up */
5328 bi
->bi_error
= -EIO
;
5332 finish_wait(&conf
->wait_for_overlap
, &w
);
5334 remaining
= raid5_dec_bi_active_stripes(bi
);
5335 if (remaining
== 0) {
5338 md_write_end(mddev
);
5340 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
5346 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5348 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5350 /* reshaping is quite different to recovery/resync so it is
5351 * handled quite separately ... here.
5353 * On each call to sync_request, we gather one chunk worth of
5354 * destination stripes and flag them as expanding.
5355 * Then we find all the source stripes and request reads.
5356 * As the reads complete, handle_stripe will copy the data
5357 * into the destination stripe and release that stripe.
5359 struct r5conf
*conf
= mddev
->private;
5360 struct stripe_head
*sh
;
5361 sector_t first_sector
, last_sector
;
5362 int raid_disks
= conf
->previous_raid_disks
;
5363 int data_disks
= raid_disks
- conf
->max_degraded
;
5364 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5367 sector_t writepos
, readpos
, safepos
;
5368 sector_t stripe_addr
;
5369 int reshape_sectors
;
5370 struct list_head stripes
;
5372 if (sector_nr
== 0) {
5373 /* If restarting in the middle, skip the initial sectors */
5374 if (mddev
->reshape_backwards
&&
5375 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5376 sector_nr
= raid5_size(mddev
, 0, 0)
5377 - conf
->reshape_progress
;
5378 } else if (!mddev
->reshape_backwards
&&
5379 conf
->reshape_progress
> 0)
5380 sector_nr
= conf
->reshape_progress
;
5381 sector_div(sector_nr
, new_data_disks
);
5383 mddev
->curr_resync_completed
= sector_nr
;
5384 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5390 /* We need to process a full chunk at a time.
5391 * If old and new chunk sizes differ, we need to process the
5394 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
5395 reshape_sectors
= mddev
->new_chunk_sectors
;
5397 reshape_sectors
= mddev
->chunk_sectors
;
5399 /* We update the metadata at least every 10 seconds, or when
5400 * the data about to be copied would over-write the source of
5401 * the data at the front of the range. i.e. one new_stripe
5402 * along from reshape_progress new_maps to after where
5403 * reshape_safe old_maps to
5405 writepos
= conf
->reshape_progress
;
5406 sector_div(writepos
, new_data_disks
);
5407 readpos
= conf
->reshape_progress
;
5408 sector_div(readpos
, data_disks
);
5409 safepos
= conf
->reshape_safe
;
5410 sector_div(safepos
, data_disks
);
5411 if (mddev
->reshape_backwards
) {
5412 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
5413 readpos
+= reshape_sectors
;
5414 safepos
+= reshape_sectors
;
5416 writepos
+= reshape_sectors
;
5417 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5418 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5421 /* Having calculated the 'writepos' possibly use it
5422 * to set 'stripe_addr' which is where we will write to.
5424 if (mddev
->reshape_backwards
) {
5425 BUG_ON(conf
->reshape_progress
== 0);
5426 stripe_addr
= writepos
;
5427 BUG_ON((mddev
->dev_sectors
&
5428 ~((sector_t
)reshape_sectors
- 1))
5429 - reshape_sectors
- stripe_addr
5432 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5433 stripe_addr
= sector_nr
;
5436 /* 'writepos' is the most advanced device address we might write.
5437 * 'readpos' is the least advanced device address we might read.
5438 * 'safepos' is the least address recorded in the metadata as having
5440 * If there is a min_offset_diff, these are adjusted either by
5441 * increasing the safepos/readpos if diff is negative, or
5442 * increasing writepos if diff is positive.
5443 * If 'readpos' is then behind 'writepos', there is no way that we can
5444 * ensure safety in the face of a crash - that must be done by userspace
5445 * making a backup of the data. So in that case there is no particular
5446 * rush to update metadata.
5447 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5448 * update the metadata to advance 'safepos' to match 'readpos' so that
5449 * we can be safe in the event of a crash.
5450 * So we insist on updating metadata if safepos is behind writepos and
5451 * readpos is beyond writepos.
5452 * In any case, update the metadata every 10 seconds.
5453 * Maybe that number should be configurable, but I'm not sure it is
5454 * worth it.... maybe it could be a multiple of safemode_delay???
5456 if (conf
->min_offset_diff
< 0) {
5457 safepos
+= -conf
->min_offset_diff
;
5458 readpos
+= -conf
->min_offset_diff
;
5460 writepos
+= conf
->min_offset_diff
;
5462 if ((mddev
->reshape_backwards
5463 ? (safepos
> writepos
&& readpos
< writepos
)
5464 : (safepos
< writepos
&& readpos
> writepos
)) ||
5465 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5466 /* Cannot proceed until we've updated the superblock... */
5467 wait_event(conf
->wait_for_overlap
,
5468 atomic_read(&conf
->reshape_stripes
)==0
5469 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5470 if (atomic_read(&conf
->reshape_stripes
) != 0)
5472 mddev
->reshape_position
= conf
->reshape_progress
;
5473 mddev
->curr_resync_completed
= sector_nr
;
5474 conf
->reshape_checkpoint
= jiffies
;
5475 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5476 md_wakeup_thread(mddev
->thread
);
5477 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
5478 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5479 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5481 spin_lock_irq(&conf
->device_lock
);
5482 conf
->reshape_safe
= mddev
->reshape_position
;
5483 spin_unlock_irq(&conf
->device_lock
);
5484 wake_up(&conf
->wait_for_overlap
);
5485 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5488 INIT_LIST_HEAD(&stripes
);
5489 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5491 int skipped_disk
= 0;
5492 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5493 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5494 atomic_inc(&conf
->reshape_stripes
);
5495 /* If any of this stripe is beyond the end of the old
5496 * array, then we need to zero those blocks
5498 for (j
=sh
->disks
; j
--;) {
5500 if (j
== sh
->pd_idx
)
5502 if (conf
->level
== 6 &&
5505 s
= compute_blocknr(sh
, j
, 0);
5506 if (s
< raid5_size(mddev
, 0, 0)) {
5510 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5511 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5512 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5514 if (!skipped_disk
) {
5515 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5516 set_bit(STRIPE_HANDLE
, &sh
->state
);
5518 list_add(&sh
->lru
, &stripes
);
5520 spin_lock_irq(&conf
->device_lock
);
5521 if (mddev
->reshape_backwards
)
5522 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5524 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5525 spin_unlock_irq(&conf
->device_lock
);
5526 /* Ok, those stripe are ready. We can start scheduling
5527 * reads on the source stripes.
5528 * The source stripes are determined by mapping the first and last
5529 * block on the destination stripes.
5532 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5535 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5536 * new_data_disks
- 1),
5538 if (last_sector
>= mddev
->dev_sectors
)
5539 last_sector
= mddev
->dev_sectors
- 1;
5540 while (first_sector
<= last_sector
) {
5541 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
5542 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5543 set_bit(STRIPE_HANDLE
, &sh
->state
);
5545 first_sector
+= STRIPE_SECTORS
;
5547 /* Now that the sources are clearly marked, we can release
5548 * the destination stripes
5550 while (!list_empty(&stripes
)) {
5551 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5552 list_del_init(&sh
->lru
);
5555 /* If this takes us to the resync_max point where we have to pause,
5556 * then we need to write out the superblock.
5558 sector_nr
+= reshape_sectors
;
5559 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
5560 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5561 /* Cannot proceed until we've updated the superblock... */
5562 wait_event(conf
->wait_for_overlap
,
5563 atomic_read(&conf
->reshape_stripes
) == 0
5564 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5565 if (atomic_read(&conf
->reshape_stripes
) != 0)
5567 mddev
->reshape_position
= conf
->reshape_progress
;
5568 mddev
->curr_resync_completed
= sector_nr
;
5569 conf
->reshape_checkpoint
= jiffies
;
5570 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5571 md_wakeup_thread(mddev
->thread
);
5572 wait_event(mddev
->sb_wait
,
5573 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
5574 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5575 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5577 spin_lock_irq(&conf
->device_lock
);
5578 conf
->reshape_safe
= mddev
->reshape_position
;
5579 spin_unlock_irq(&conf
->device_lock
);
5580 wake_up(&conf
->wait_for_overlap
);
5581 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5584 return reshape_sectors
;
5587 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5589 struct r5conf
*conf
= mddev
->private;
5590 struct stripe_head
*sh
;
5591 sector_t max_sector
= mddev
->dev_sectors
;
5592 sector_t sync_blocks
;
5593 int still_degraded
= 0;
5596 if (sector_nr
>= max_sector
) {
5597 /* just being told to finish up .. nothing much to do */
5599 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
5604 if (mddev
->curr_resync
< max_sector
) /* aborted */
5605 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
5607 else /* completed sync */
5609 bitmap_close_sync(mddev
->bitmap
);
5614 /* Allow raid5_quiesce to complete */
5615 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
5617 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
5618 return reshape_request(mddev
, sector_nr
, skipped
);
5620 /* No need to check resync_max as we never do more than one
5621 * stripe, and as resync_max will always be on a chunk boundary,
5622 * if the check in md_do_sync didn't fire, there is no chance
5623 * of overstepping resync_max here
5626 /* if there is too many failed drives and we are trying
5627 * to resync, then assert that we are finished, because there is
5628 * nothing we can do.
5630 if (mddev
->degraded
>= conf
->max_degraded
&&
5631 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
5632 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
5636 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
5638 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
5639 sync_blocks
>= STRIPE_SECTORS
) {
5640 /* we can skip this block, and probably more */
5641 sync_blocks
/= STRIPE_SECTORS
;
5643 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
5646 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
5648 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
5650 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
5651 /* make sure we don't swamp the stripe cache if someone else
5652 * is trying to get access
5654 schedule_timeout_uninterruptible(1);
5656 /* Need to check if array will still be degraded after recovery/resync
5657 * Note in case of > 1 drive failures it's possible we're rebuilding
5658 * one drive while leaving another faulty drive in array.
5661 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5662 struct md_rdev
*rdev
= ACCESS_ONCE(conf
->disks
[i
].rdev
);
5664 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
5669 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
5671 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
5672 set_bit(STRIPE_HANDLE
, &sh
->state
);
5676 return STRIPE_SECTORS
;
5679 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
5681 /* We may not be able to submit a whole bio at once as there
5682 * may not be enough stripe_heads available.
5683 * We cannot pre-allocate enough stripe_heads as we may need
5684 * more than exist in the cache (if we allow ever large chunks).
5685 * So we do one stripe head at a time and record in
5686 * ->bi_hw_segments how many have been done.
5688 * We *know* that this entire raid_bio is in one chunk, so
5689 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5691 struct stripe_head
*sh
;
5693 sector_t sector
, logical_sector
, last_sector
;
5698 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
5699 ~((sector_t
)STRIPE_SECTORS
-1);
5700 sector
= raid5_compute_sector(conf
, logical_sector
,
5702 last_sector
= bio_end_sector(raid_bio
);
5704 for (; logical_sector
< last_sector
;
5705 logical_sector
+= STRIPE_SECTORS
,
5706 sector
+= STRIPE_SECTORS
,
5709 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
5710 /* already done this stripe */
5713 sh
= get_active_stripe(conf
, sector
, 0, 1, 1);
5716 /* failed to get a stripe - must wait */
5717 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5718 conf
->retry_read_aligned
= raid_bio
;
5722 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
5724 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5725 conf
->retry_read_aligned
= raid_bio
;
5729 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
5734 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
5735 if (remaining
== 0) {
5736 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
5738 bio_endio(raid_bio
);
5740 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5741 wake_up(&conf
->wait_for_quiescent
);
5745 static int handle_active_stripes(struct r5conf
*conf
, int group
,
5746 struct r5worker
*worker
,
5747 struct list_head
*temp_inactive_list
)
5749 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
5750 int i
, batch_size
= 0, hash
;
5751 bool release_inactive
= false;
5753 while (batch_size
< MAX_STRIPE_BATCH
&&
5754 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
5755 batch
[batch_size
++] = sh
;
5757 if (batch_size
== 0) {
5758 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5759 if (!list_empty(temp_inactive_list
+ i
))
5761 if (i
== NR_STRIPE_HASH_LOCKS
)
5763 release_inactive
= true;
5765 spin_unlock_irq(&conf
->device_lock
);
5767 release_inactive_stripe_list(conf
, temp_inactive_list
,
5768 NR_STRIPE_HASH_LOCKS
);
5770 if (release_inactive
) {
5771 spin_lock_irq(&conf
->device_lock
);
5775 for (i
= 0; i
< batch_size
; i
++)
5776 handle_stripe(batch
[i
]);
5780 spin_lock_irq(&conf
->device_lock
);
5781 for (i
= 0; i
< batch_size
; i
++) {
5782 hash
= batch
[i
]->hash_lock_index
;
5783 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
5788 static void raid5_do_work(struct work_struct
*work
)
5790 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
5791 struct r5worker_group
*group
= worker
->group
;
5792 struct r5conf
*conf
= group
->conf
;
5793 int group_id
= group
- conf
->worker_groups
;
5795 struct blk_plug plug
;
5797 pr_debug("+++ raid5worker active\n");
5799 blk_start_plug(&plug
);
5801 spin_lock_irq(&conf
->device_lock
);
5803 int batch_size
, released
;
5805 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
5807 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
5808 worker
->temp_inactive_list
);
5809 worker
->working
= false;
5810 if (!batch_size
&& !released
)
5812 handled
+= batch_size
;
5814 pr_debug("%d stripes handled\n", handled
);
5816 spin_unlock_irq(&conf
->device_lock
);
5817 blk_finish_plug(&plug
);
5819 pr_debug("--- raid5worker inactive\n");
5823 * This is our raid5 kernel thread.
5825 * We scan the hash table for stripes which can be handled now.
5826 * During the scan, completed stripes are saved for us by the interrupt
5827 * handler, so that they will not have to wait for our next wakeup.
5829 static void raid5d(struct md_thread
*thread
)
5831 struct mddev
*mddev
= thread
->mddev
;
5832 struct r5conf
*conf
= mddev
->private;
5834 struct blk_plug plug
;
5836 pr_debug("+++ raid5d active\n");
5838 md_check_recovery(mddev
);
5840 blk_start_plug(&plug
);
5842 spin_lock_irq(&conf
->device_lock
);
5845 int batch_size
, released
;
5847 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
5849 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5852 !list_empty(&conf
->bitmap_list
)) {
5853 /* Now is a good time to flush some bitmap updates */
5855 spin_unlock_irq(&conf
->device_lock
);
5856 bitmap_unplug(mddev
->bitmap
);
5857 spin_lock_irq(&conf
->device_lock
);
5858 conf
->seq_write
= conf
->seq_flush
;
5859 activate_bit_delay(conf
, conf
->temp_inactive_list
);
5861 raid5_activate_delayed(conf
);
5863 while ((bio
= remove_bio_from_retry(conf
))) {
5865 spin_unlock_irq(&conf
->device_lock
);
5866 ok
= retry_aligned_read(conf
, bio
);
5867 spin_lock_irq(&conf
->device_lock
);
5873 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
5874 conf
->temp_inactive_list
);
5875 if (!batch_size
&& !released
)
5877 handled
+= batch_size
;
5879 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
5880 spin_unlock_irq(&conf
->device_lock
);
5881 md_check_recovery(mddev
);
5882 spin_lock_irq(&conf
->device_lock
);
5885 pr_debug("%d stripes handled\n", handled
);
5887 spin_unlock_irq(&conf
->device_lock
);
5888 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
)) {
5889 grow_one_stripe(conf
, __GFP_NOWARN
);
5890 /* Set flag even if allocation failed. This helps
5891 * slow down allocation requests when mem is short
5893 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5896 async_tx_issue_pending_all();
5897 blk_finish_plug(&plug
);
5899 pr_debug("--- raid5d inactive\n");
5903 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
5905 struct r5conf
*conf
;
5907 spin_lock(&mddev
->lock
);
5908 conf
= mddev
->private;
5910 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
5911 spin_unlock(&mddev
->lock
);
5916 raid5_set_cache_size(struct mddev
*mddev
, int size
)
5918 struct r5conf
*conf
= mddev
->private;
5921 if (size
<= 16 || size
> 32768)
5924 conf
->min_nr_stripes
= size
;
5925 while (size
< conf
->max_nr_stripes
&&
5926 drop_one_stripe(conf
))
5930 err
= md_allow_write(mddev
);
5934 while (size
> conf
->max_nr_stripes
)
5935 if (!grow_one_stripe(conf
, GFP_KERNEL
))
5940 EXPORT_SYMBOL(raid5_set_cache_size
);
5943 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
5945 struct r5conf
*conf
;
5949 if (len
>= PAGE_SIZE
)
5951 if (kstrtoul(page
, 10, &new))
5953 err
= mddev_lock(mddev
);
5956 conf
= mddev
->private;
5960 err
= raid5_set_cache_size(mddev
, new);
5961 mddev_unlock(mddev
);
5966 static struct md_sysfs_entry
5967 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
5968 raid5_show_stripe_cache_size
,
5969 raid5_store_stripe_cache_size
);
5972 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
5974 struct r5conf
*conf
= mddev
->private;
5976 return sprintf(page
, "%d\n", conf
->rmw_level
);
5982 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
5984 struct r5conf
*conf
= mddev
->private;
5990 if (len
>= PAGE_SIZE
)
5993 if (kstrtoul(page
, 10, &new))
5996 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
5999 if (new != PARITY_DISABLE_RMW
&&
6000 new != PARITY_ENABLE_RMW
&&
6001 new != PARITY_PREFER_RMW
)
6004 conf
->rmw_level
= new;
6008 static struct md_sysfs_entry
6009 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
6010 raid5_show_rmw_level
,
6011 raid5_store_rmw_level
);
6015 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
6017 struct r5conf
*conf
;
6019 spin_lock(&mddev
->lock
);
6020 conf
= mddev
->private;
6022 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
6023 spin_unlock(&mddev
->lock
);
6028 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
6030 struct r5conf
*conf
;
6034 if (len
>= PAGE_SIZE
)
6036 if (kstrtoul(page
, 10, &new))
6039 err
= mddev_lock(mddev
);
6042 conf
= mddev
->private;
6045 else if (new > conf
->min_nr_stripes
)
6048 conf
->bypass_threshold
= new;
6049 mddev_unlock(mddev
);
6053 static struct md_sysfs_entry
6054 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
6056 raid5_show_preread_threshold
,
6057 raid5_store_preread_threshold
);
6060 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
6062 struct r5conf
*conf
;
6064 spin_lock(&mddev
->lock
);
6065 conf
= mddev
->private;
6067 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6068 spin_unlock(&mddev
->lock
);
6073 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6075 struct r5conf
*conf
;
6079 if (len
>= PAGE_SIZE
)
6081 if (kstrtoul(page
, 10, &new))
6085 err
= mddev_lock(mddev
);
6088 conf
= mddev
->private;
6091 else if (new != conf
->skip_copy
) {
6092 mddev_suspend(mddev
);
6093 conf
->skip_copy
= new;
6095 mddev
->queue
->backing_dev_info
.capabilities
|=
6096 BDI_CAP_STABLE_WRITES
;
6098 mddev
->queue
->backing_dev_info
.capabilities
&=
6099 ~BDI_CAP_STABLE_WRITES
;
6100 mddev_resume(mddev
);
6102 mddev_unlock(mddev
);
6106 static struct md_sysfs_entry
6107 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6108 raid5_show_skip_copy
,
6109 raid5_store_skip_copy
);
6112 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6114 struct r5conf
*conf
= mddev
->private;
6116 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6121 static struct md_sysfs_entry
6122 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6125 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6127 struct r5conf
*conf
;
6129 spin_lock(&mddev
->lock
);
6130 conf
= mddev
->private;
6132 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6133 spin_unlock(&mddev
->lock
);
6137 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6139 int *worker_cnt_per_group
,
6140 struct r5worker_group
**worker_groups
);
6142 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6144 struct r5conf
*conf
;
6147 struct r5worker_group
*new_groups
, *old_groups
;
6148 int group_cnt
, worker_cnt_per_group
;
6150 if (len
>= PAGE_SIZE
)
6152 if (kstrtoul(page
, 10, &new))
6155 err
= mddev_lock(mddev
);
6158 conf
= mddev
->private;
6161 else if (new != conf
->worker_cnt_per_group
) {
6162 mddev_suspend(mddev
);
6164 old_groups
= conf
->worker_groups
;
6166 flush_workqueue(raid5_wq
);
6168 err
= alloc_thread_groups(conf
, new,
6169 &group_cnt
, &worker_cnt_per_group
,
6172 spin_lock_irq(&conf
->device_lock
);
6173 conf
->group_cnt
= group_cnt
;
6174 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6175 conf
->worker_groups
= new_groups
;
6176 spin_unlock_irq(&conf
->device_lock
);
6179 kfree(old_groups
[0].workers
);
6182 mddev_resume(mddev
);
6184 mddev_unlock(mddev
);
6189 static struct md_sysfs_entry
6190 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6191 raid5_show_group_thread_cnt
,
6192 raid5_store_group_thread_cnt
);
6194 static struct attribute
*raid5_attrs
[] = {
6195 &raid5_stripecache_size
.attr
,
6196 &raid5_stripecache_active
.attr
,
6197 &raid5_preread_bypass_threshold
.attr
,
6198 &raid5_group_thread_cnt
.attr
,
6199 &raid5_skip_copy
.attr
,
6200 &raid5_rmw_level
.attr
,
6203 static struct attribute_group raid5_attrs_group
= {
6205 .attrs
= raid5_attrs
,
6208 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6210 int *worker_cnt_per_group
,
6211 struct r5worker_group
**worker_groups
)
6215 struct r5worker
*workers
;
6217 *worker_cnt_per_group
= cnt
;
6220 *worker_groups
= NULL
;
6223 *group_cnt
= num_possible_nodes();
6224 size
= sizeof(struct r5worker
) * cnt
;
6225 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6226 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6227 *group_cnt
, GFP_NOIO
);
6228 if (!*worker_groups
|| !workers
) {
6230 kfree(*worker_groups
);
6234 for (i
= 0; i
< *group_cnt
; i
++) {
6235 struct r5worker_group
*group
;
6237 group
= &(*worker_groups
)[i
];
6238 INIT_LIST_HEAD(&group
->handle_list
);
6240 group
->workers
= workers
+ i
* cnt
;
6242 for (j
= 0; j
< cnt
; j
++) {
6243 struct r5worker
*worker
= group
->workers
+ j
;
6244 worker
->group
= group
;
6245 INIT_WORK(&worker
->work
, raid5_do_work
);
6247 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6248 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6255 static void free_thread_groups(struct r5conf
*conf
)
6257 if (conf
->worker_groups
)
6258 kfree(conf
->worker_groups
[0].workers
);
6259 kfree(conf
->worker_groups
);
6260 conf
->worker_groups
= NULL
;
6264 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6266 struct r5conf
*conf
= mddev
->private;
6269 sectors
= mddev
->dev_sectors
;
6271 /* size is defined by the smallest of previous and new size */
6272 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6274 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
6275 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
6276 return sectors
* (raid_disks
- conf
->max_degraded
);
6279 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6281 safe_put_page(percpu
->spare_page
);
6282 if (percpu
->scribble
)
6283 flex_array_free(percpu
->scribble
);
6284 percpu
->spare_page
= NULL
;
6285 percpu
->scribble
= NULL
;
6288 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6290 if (conf
->level
== 6 && !percpu
->spare_page
)
6291 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6292 if (!percpu
->scribble
)
6293 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6294 conf
->previous_raid_disks
),
6295 max(conf
->chunk_sectors
,
6296 conf
->prev_chunk_sectors
)
6300 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6301 free_scratch_buffer(conf
, percpu
);
6308 static void raid5_free_percpu(struct r5conf
*conf
)
6315 #ifdef CONFIG_HOTPLUG_CPU
6316 unregister_cpu_notifier(&conf
->cpu_notify
);
6320 for_each_possible_cpu(cpu
)
6321 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6324 free_percpu(conf
->percpu
);
6327 static void free_conf(struct r5conf
*conf
)
6329 if (conf
->shrinker
.seeks
)
6330 unregister_shrinker(&conf
->shrinker
);
6331 free_thread_groups(conf
);
6332 shrink_stripes(conf
);
6333 raid5_free_percpu(conf
);
6335 kfree(conf
->stripe_hashtbl
);
6339 #ifdef CONFIG_HOTPLUG_CPU
6340 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
6343 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
6344 long cpu
= (long)hcpu
;
6345 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6348 case CPU_UP_PREPARE
:
6349 case CPU_UP_PREPARE_FROZEN
:
6350 if (alloc_scratch_buffer(conf
, percpu
)) {
6351 pr_err("%s: failed memory allocation for cpu%ld\n",
6353 return notifier_from_errno(-ENOMEM
);
6357 case CPU_DEAD_FROZEN
:
6358 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6367 static int raid5_alloc_percpu(struct r5conf
*conf
)
6372 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6376 #ifdef CONFIG_HOTPLUG_CPU
6377 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
6378 conf
->cpu_notify
.priority
= 0;
6379 err
= register_cpu_notifier(&conf
->cpu_notify
);
6385 for_each_present_cpu(cpu
) {
6386 err
= alloc_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6388 pr_err("%s: failed memory allocation for cpu%ld\n",
6398 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6399 struct shrink_control
*sc
)
6401 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6403 while (ret
< sc
->nr_to_scan
) {
6404 if (drop_one_stripe(conf
) == 0)
6411 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6412 struct shrink_control
*sc
)
6414 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6416 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6417 /* unlikely, but not impossible */
6419 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6422 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6424 struct r5conf
*conf
;
6425 int raid_disk
, memory
, max_disks
;
6426 struct md_rdev
*rdev
;
6427 struct disk_info
*disk
;
6430 int group_cnt
, worker_cnt_per_group
;
6431 struct r5worker_group
*new_group
;
6433 if (mddev
->new_level
!= 5
6434 && mddev
->new_level
!= 4
6435 && mddev
->new_level
!= 6) {
6436 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6437 mdname(mddev
), mddev
->new_level
);
6438 return ERR_PTR(-EIO
);
6440 if ((mddev
->new_level
== 5
6441 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6442 (mddev
->new_level
== 6
6443 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6444 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
6445 mdname(mddev
), mddev
->new_layout
);
6446 return ERR_PTR(-EIO
);
6448 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6449 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6450 mdname(mddev
), mddev
->raid_disks
);
6451 return ERR_PTR(-EINVAL
);
6454 if (!mddev
->new_chunk_sectors
||
6455 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6456 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6457 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
6458 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6459 return ERR_PTR(-EINVAL
);
6462 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6465 /* Don't enable multi-threading by default*/
6466 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6468 conf
->group_cnt
= group_cnt
;
6469 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6470 conf
->worker_groups
= new_group
;
6473 spin_lock_init(&conf
->device_lock
);
6474 seqcount_init(&conf
->gen_lock
);
6475 init_waitqueue_head(&conf
->wait_for_quiescent
);
6476 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++) {
6477 init_waitqueue_head(&conf
->wait_for_stripe
[i
]);
6479 init_waitqueue_head(&conf
->wait_for_overlap
);
6480 INIT_LIST_HEAD(&conf
->handle_list
);
6481 INIT_LIST_HEAD(&conf
->hold_list
);
6482 INIT_LIST_HEAD(&conf
->delayed_list
);
6483 INIT_LIST_HEAD(&conf
->bitmap_list
);
6484 init_llist_head(&conf
->released_stripes
);
6485 atomic_set(&conf
->active_stripes
, 0);
6486 atomic_set(&conf
->preread_active_stripes
, 0);
6487 atomic_set(&conf
->active_aligned_reads
, 0);
6488 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6489 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6491 conf
->raid_disks
= mddev
->raid_disks
;
6492 if (mddev
->reshape_position
== MaxSector
)
6493 conf
->previous_raid_disks
= mddev
->raid_disks
;
6495 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6496 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6498 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6503 conf
->mddev
= mddev
;
6505 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6508 /* We init hash_locks[0] separately to that it can be used
6509 * as the reference lock in the spin_lock_nest_lock() call
6510 * in lock_all_device_hash_locks_irq in order to convince
6511 * lockdep that we know what we are doing.
6513 spin_lock_init(conf
->hash_locks
);
6514 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6515 spin_lock_init(conf
->hash_locks
+ i
);
6517 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6518 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6520 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6521 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6523 conf
->level
= mddev
->new_level
;
6524 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6525 if (raid5_alloc_percpu(conf
) != 0)
6528 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6530 rdev_for_each(rdev
, mddev
) {
6531 raid_disk
= rdev
->raid_disk
;
6532 if (raid_disk
>= max_disks
6535 disk
= conf
->disks
+ raid_disk
;
6537 if (test_bit(Replacement
, &rdev
->flags
)) {
6538 if (disk
->replacement
)
6540 disk
->replacement
= rdev
;
6547 if (test_bit(In_sync
, &rdev
->flags
)) {
6548 char b
[BDEVNAME_SIZE
];
6549 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
6551 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
6552 } else if (rdev
->saved_raid_disk
!= raid_disk
)
6553 /* Cannot rely on bitmap to complete recovery */
6557 conf
->level
= mddev
->new_level
;
6558 if (conf
->level
== 6) {
6559 conf
->max_degraded
= 2;
6560 if (raid6_call
.xor_syndrome
)
6561 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6563 conf
->rmw_level
= PARITY_DISABLE_RMW
;
6565 conf
->max_degraded
= 1;
6566 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6568 conf
->algorithm
= mddev
->new_layout
;
6569 conf
->reshape_progress
= mddev
->reshape_position
;
6570 if (conf
->reshape_progress
!= MaxSector
) {
6571 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
6572 conf
->prev_algo
= mddev
->layout
;
6575 conf
->min_nr_stripes
= NR_STRIPES
;
6576 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
6577 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
6578 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
6579 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
6581 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6582 mdname(mddev
), memory
);
6585 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
6586 mdname(mddev
), memory
);
6588 * Losing a stripe head costs more than the time to refill it,
6589 * it reduces the queue depth and so can hurt throughput.
6590 * So set it rather large, scaled by number of devices.
6592 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
6593 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
6594 conf
->shrinker
.count_objects
= raid5_cache_count
;
6595 conf
->shrinker
.batch
= 128;
6596 conf
->shrinker
.flags
= 0;
6597 register_shrinker(&conf
->shrinker
);
6599 sprintf(pers_name
, "raid%d", mddev
->new_level
);
6600 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
6601 if (!conf
->thread
) {
6603 "md/raid:%s: couldn't allocate thread.\n",
6613 return ERR_PTR(-EIO
);
6615 return ERR_PTR(-ENOMEM
);
6618 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
6621 case ALGORITHM_PARITY_0
:
6622 if (raid_disk
< max_degraded
)
6625 case ALGORITHM_PARITY_N
:
6626 if (raid_disk
>= raid_disks
- max_degraded
)
6629 case ALGORITHM_PARITY_0_6
:
6630 if (raid_disk
== 0 ||
6631 raid_disk
== raid_disks
- 1)
6634 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6635 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6636 case ALGORITHM_LEFT_SYMMETRIC_6
:
6637 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6638 if (raid_disk
== raid_disks
- 1)
6644 static int run(struct mddev
*mddev
)
6646 struct r5conf
*conf
;
6647 int working_disks
= 0;
6648 int dirty_parity_disks
= 0;
6649 struct md_rdev
*rdev
;
6650 sector_t reshape_offset
= 0;
6652 long long min_offset_diff
= 0;
6655 if (mddev
->recovery_cp
!= MaxSector
)
6656 printk(KERN_NOTICE
"md/raid:%s: not clean"
6657 " -- starting background reconstruction\n",
6660 rdev_for_each(rdev
, mddev
) {
6662 if (rdev
->raid_disk
< 0)
6664 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
6666 min_offset_diff
= diff
;
6668 } else if (mddev
->reshape_backwards
&&
6669 diff
< min_offset_diff
)
6670 min_offset_diff
= diff
;
6671 else if (!mddev
->reshape_backwards
&&
6672 diff
> min_offset_diff
)
6673 min_offset_diff
= diff
;
6676 if (mddev
->reshape_position
!= MaxSector
) {
6677 /* Check that we can continue the reshape.
6678 * Difficulties arise if the stripe we would write to
6679 * next is at or after the stripe we would read from next.
6680 * For a reshape that changes the number of devices, this
6681 * is only possible for a very short time, and mdadm makes
6682 * sure that time appears to have past before assembling
6683 * the array. So we fail if that time hasn't passed.
6684 * For a reshape that keeps the number of devices the same
6685 * mdadm must be monitoring the reshape can keeping the
6686 * critical areas read-only and backed up. It will start
6687 * the array in read-only mode, so we check for that.
6689 sector_t here_new
, here_old
;
6691 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
6693 if (mddev
->new_level
!= mddev
->level
) {
6694 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
6695 "required - aborting.\n",
6699 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6700 /* reshape_position must be on a new-stripe boundary, and one
6701 * further up in new geometry must map after here in old
6704 here_new
= mddev
->reshape_position
;
6705 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
6706 (mddev
->raid_disks
- max_degraded
))) {
6707 printk(KERN_ERR
"md/raid:%s: reshape_position not "
6708 "on a stripe boundary\n", mdname(mddev
));
6711 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
6712 /* here_new is the stripe we will write to */
6713 here_old
= mddev
->reshape_position
;
6714 sector_div(here_old
, mddev
->chunk_sectors
*
6715 (old_disks
-max_degraded
));
6716 /* here_old is the first stripe that we might need to read
6718 if (mddev
->delta_disks
== 0) {
6719 if ((here_new
* mddev
->new_chunk_sectors
!=
6720 here_old
* mddev
->chunk_sectors
)) {
6721 printk(KERN_ERR
"md/raid:%s: reshape position is"
6722 " confused - aborting\n", mdname(mddev
));
6725 /* We cannot be sure it is safe to start an in-place
6726 * reshape. It is only safe if user-space is monitoring
6727 * and taking constant backups.
6728 * mdadm always starts a situation like this in
6729 * readonly mode so it can take control before
6730 * allowing any writes. So just check for that.
6732 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
6733 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
6734 /* not really in-place - so OK */;
6735 else if (mddev
->ro
== 0) {
6736 printk(KERN_ERR
"md/raid:%s: in-place reshape "
6737 "must be started in read-only mode "
6742 } else if (mddev
->reshape_backwards
6743 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
6744 here_old
* mddev
->chunk_sectors
)
6745 : (here_new
* mddev
->new_chunk_sectors
>=
6746 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
6747 /* Reading from the same stripe as writing to - bad */
6748 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
6749 "auto-recovery - aborting.\n",
6753 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
6755 /* OK, we should be able to continue; */
6757 BUG_ON(mddev
->level
!= mddev
->new_level
);
6758 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
6759 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
6760 BUG_ON(mddev
->delta_disks
!= 0);
6763 if (mddev
->private == NULL
)
6764 conf
= setup_conf(mddev
);
6766 conf
= mddev
->private;
6769 return PTR_ERR(conf
);
6771 conf
->min_offset_diff
= min_offset_diff
;
6772 mddev
->thread
= conf
->thread
;
6773 conf
->thread
= NULL
;
6774 mddev
->private = conf
;
6776 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
6778 rdev
= conf
->disks
[i
].rdev
;
6779 if (!rdev
&& conf
->disks
[i
].replacement
) {
6780 /* The replacement is all we have yet */
6781 rdev
= conf
->disks
[i
].replacement
;
6782 conf
->disks
[i
].replacement
= NULL
;
6783 clear_bit(Replacement
, &rdev
->flags
);
6784 conf
->disks
[i
].rdev
= rdev
;
6788 if (conf
->disks
[i
].replacement
&&
6789 conf
->reshape_progress
!= MaxSector
) {
6790 /* replacements and reshape simply do not mix. */
6791 printk(KERN_ERR
"md: cannot handle concurrent "
6792 "replacement and reshape.\n");
6795 if (test_bit(In_sync
, &rdev
->flags
)) {
6799 /* This disc is not fully in-sync. However if it
6800 * just stored parity (beyond the recovery_offset),
6801 * when we don't need to be concerned about the
6802 * array being dirty.
6803 * When reshape goes 'backwards', we never have
6804 * partially completed devices, so we only need
6805 * to worry about reshape going forwards.
6807 /* Hack because v0.91 doesn't store recovery_offset properly. */
6808 if (mddev
->major_version
== 0 &&
6809 mddev
->minor_version
> 90)
6810 rdev
->recovery_offset
= reshape_offset
;
6812 if (rdev
->recovery_offset
< reshape_offset
) {
6813 /* We need to check old and new layout */
6814 if (!only_parity(rdev
->raid_disk
,
6817 conf
->max_degraded
))
6820 if (!only_parity(rdev
->raid_disk
,
6822 conf
->previous_raid_disks
,
6823 conf
->max_degraded
))
6825 dirty_parity_disks
++;
6829 * 0 for a fully functional array, 1 or 2 for a degraded array.
6831 mddev
->degraded
= calc_degraded(conf
);
6833 if (has_failed(conf
)) {
6834 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
6835 " (%d/%d failed)\n",
6836 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
6840 /* device size must be a multiple of chunk size */
6841 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
6842 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
6844 if (mddev
->degraded
> dirty_parity_disks
&&
6845 mddev
->recovery_cp
!= MaxSector
) {
6846 if (mddev
->ok_start_degraded
)
6848 "md/raid:%s: starting dirty degraded array"
6849 " - data corruption possible.\n",
6853 "md/raid:%s: cannot start dirty degraded array.\n",
6859 if (mddev
->degraded
== 0)
6860 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
6861 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
6862 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
6865 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
6866 " out of %d devices, algorithm %d\n",
6867 mdname(mddev
), conf
->level
,
6868 mddev
->raid_disks
- mddev
->degraded
,
6869 mddev
->raid_disks
, mddev
->new_layout
);
6871 print_raid5_conf(conf
);
6873 if (conf
->reshape_progress
!= MaxSector
) {
6874 conf
->reshape_safe
= conf
->reshape_progress
;
6875 atomic_set(&conf
->reshape_stripes
, 0);
6876 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6877 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6878 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6879 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6880 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6884 /* Ok, everything is just fine now */
6885 if (mddev
->to_remove
== &raid5_attrs_group
)
6886 mddev
->to_remove
= NULL
;
6887 else if (mddev
->kobj
.sd
&&
6888 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
6890 "raid5: failed to create sysfs attributes for %s\n",
6892 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6896 bool discard_supported
= true;
6897 /* read-ahead size must cover two whole stripes, which
6898 * is 2 * (datadisks) * chunksize where 'n' is the
6899 * number of raid devices
6901 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
6902 int stripe
= data_disks
*
6903 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
6904 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6905 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6907 chunk_size
= mddev
->chunk_sectors
<< 9;
6908 blk_queue_io_min(mddev
->queue
, chunk_size
);
6909 blk_queue_io_opt(mddev
->queue
, chunk_size
*
6910 (conf
->raid_disks
- conf
->max_degraded
));
6911 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
6913 * We can only discard a whole stripe. It doesn't make sense to
6914 * discard data disk but write parity disk
6916 stripe
= stripe
* PAGE_SIZE
;
6917 /* Round up to power of 2, as discard handling
6918 * currently assumes that */
6919 while ((stripe
-1) & stripe
)
6920 stripe
= (stripe
| (stripe
-1)) + 1;
6921 mddev
->queue
->limits
.discard_alignment
= stripe
;
6922 mddev
->queue
->limits
.discard_granularity
= stripe
;
6924 * unaligned part of discard request will be ignored, so can't
6925 * guarantee discard_zeroes_data
6927 mddev
->queue
->limits
.discard_zeroes_data
= 0;
6929 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
6931 rdev_for_each(rdev
, mddev
) {
6932 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6933 rdev
->data_offset
<< 9);
6934 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6935 rdev
->new_data_offset
<< 9);
6937 * discard_zeroes_data is required, otherwise data
6938 * could be lost. Consider a scenario: discard a stripe
6939 * (the stripe could be inconsistent if
6940 * discard_zeroes_data is 0); write one disk of the
6941 * stripe (the stripe could be inconsistent again
6942 * depending on which disks are used to calculate
6943 * parity); the disk is broken; The stripe data of this
6946 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
6947 !bdev_get_queue(rdev
->bdev
)->
6948 limits
.discard_zeroes_data
)
6949 discard_supported
= false;
6950 /* Unfortunately, discard_zeroes_data is not currently
6951 * a guarantee - just a hint. So we only allow DISCARD
6952 * if the sysadmin has confirmed that only safe devices
6953 * are in use by setting a module parameter.
6955 if (!devices_handle_discard_safely
) {
6956 if (discard_supported
) {
6957 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6958 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6960 discard_supported
= false;
6964 if (discard_supported
&&
6965 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
6966 mddev
->queue
->limits
.discard_granularity
>= stripe
)
6967 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
6970 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
6976 md_unregister_thread(&mddev
->thread
);
6977 print_raid5_conf(conf
);
6979 mddev
->private = NULL
;
6980 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
6984 static void raid5_free(struct mddev
*mddev
, void *priv
)
6986 struct r5conf
*conf
= priv
;
6989 mddev
->to_remove
= &raid5_attrs_group
;
6992 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
6994 struct r5conf
*conf
= mddev
->private;
6997 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
6998 mddev
->chunk_sectors
/ 2, mddev
->layout
);
6999 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
7000 for (i
= 0; i
< conf
->raid_disks
; i
++)
7001 seq_printf (seq
, "%s",
7002 conf
->disks
[i
].rdev
&&
7003 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
7004 seq_printf (seq
, "]");
7007 static void print_raid5_conf (struct r5conf
*conf
)
7010 struct disk_info
*tmp
;
7012 printk(KERN_DEBUG
"RAID conf printout:\n");
7014 printk("(conf==NULL)\n");
7017 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
7019 conf
->raid_disks
- conf
->mddev
->degraded
);
7021 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7022 char b
[BDEVNAME_SIZE
];
7023 tmp
= conf
->disks
+ i
;
7025 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
7026 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
7027 bdevname(tmp
->rdev
->bdev
, b
));
7031 static int raid5_spare_active(struct mddev
*mddev
)
7034 struct r5conf
*conf
= mddev
->private;
7035 struct disk_info
*tmp
;
7037 unsigned long flags
;
7039 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7040 tmp
= conf
->disks
+ i
;
7041 if (tmp
->replacement
7042 && tmp
->replacement
->recovery_offset
== MaxSector
7043 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
7044 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
7045 /* Replacement has just become active. */
7047 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
7050 /* Replaced device not technically faulty,
7051 * but we need to be sure it gets removed
7052 * and never re-added.
7054 set_bit(Faulty
, &tmp
->rdev
->flags
);
7055 sysfs_notify_dirent_safe(
7056 tmp
->rdev
->sysfs_state
);
7058 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
7059 } else if (tmp
->rdev
7060 && tmp
->rdev
->recovery_offset
== MaxSector
7061 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
7062 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
7064 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
7067 spin_lock_irqsave(&conf
->device_lock
, flags
);
7068 mddev
->degraded
= calc_degraded(conf
);
7069 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7070 print_raid5_conf(conf
);
7074 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7076 struct r5conf
*conf
= mddev
->private;
7078 int number
= rdev
->raid_disk
;
7079 struct md_rdev
**rdevp
;
7080 struct disk_info
*p
= conf
->disks
+ number
;
7082 print_raid5_conf(conf
);
7083 if (rdev
== p
->rdev
)
7085 else if (rdev
== p
->replacement
)
7086 rdevp
= &p
->replacement
;
7090 if (number
>= conf
->raid_disks
&&
7091 conf
->reshape_progress
== MaxSector
)
7092 clear_bit(In_sync
, &rdev
->flags
);
7094 if (test_bit(In_sync
, &rdev
->flags
) ||
7095 atomic_read(&rdev
->nr_pending
)) {
7099 /* Only remove non-faulty devices if recovery
7102 if (!test_bit(Faulty
, &rdev
->flags
) &&
7103 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7104 !has_failed(conf
) &&
7105 (!p
->replacement
|| p
->replacement
== rdev
) &&
7106 number
< conf
->raid_disks
) {
7112 if (atomic_read(&rdev
->nr_pending
)) {
7113 /* lost the race, try later */
7116 } else if (p
->replacement
) {
7117 /* We must have just cleared 'rdev' */
7118 p
->rdev
= p
->replacement
;
7119 clear_bit(Replacement
, &p
->replacement
->flags
);
7120 smp_mb(); /* Make sure other CPUs may see both as identical
7121 * but will never see neither - if they are careful
7123 p
->replacement
= NULL
;
7124 clear_bit(WantReplacement
, &rdev
->flags
);
7126 /* We might have just removed the Replacement as faulty-
7127 * clear the bit just in case
7129 clear_bit(WantReplacement
, &rdev
->flags
);
7132 print_raid5_conf(conf
);
7136 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7138 struct r5conf
*conf
= mddev
->private;
7141 struct disk_info
*p
;
7143 int last
= conf
->raid_disks
- 1;
7145 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7148 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7149 /* no point adding a device */
7152 if (rdev
->raid_disk
>= 0)
7153 first
= last
= rdev
->raid_disk
;
7156 * find the disk ... but prefer rdev->saved_raid_disk
7159 if (rdev
->saved_raid_disk
>= 0 &&
7160 rdev
->saved_raid_disk
>= first
&&
7161 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7162 first
= rdev
->saved_raid_disk
;
7164 for (disk
= first
; disk
<= last
; disk
++) {
7165 p
= conf
->disks
+ disk
;
7166 if (p
->rdev
== NULL
) {
7167 clear_bit(In_sync
, &rdev
->flags
);
7168 rdev
->raid_disk
= disk
;
7170 if (rdev
->saved_raid_disk
!= disk
)
7172 rcu_assign_pointer(p
->rdev
, rdev
);
7176 for (disk
= first
; disk
<= last
; disk
++) {
7177 p
= conf
->disks
+ disk
;
7178 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7179 p
->replacement
== NULL
) {
7180 clear_bit(In_sync
, &rdev
->flags
);
7181 set_bit(Replacement
, &rdev
->flags
);
7182 rdev
->raid_disk
= disk
;
7185 rcu_assign_pointer(p
->replacement
, rdev
);
7190 print_raid5_conf(conf
);
7194 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7196 /* no resync is happening, and there is enough space
7197 * on all devices, so we can resize.
7198 * We need to make sure resync covers any new space.
7199 * If the array is shrinking we should possibly wait until
7200 * any io in the removed space completes, but it hardly seems
7204 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
7205 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7206 if (mddev
->external_size
&&
7207 mddev
->array_sectors
> newsize
)
7209 if (mddev
->bitmap
) {
7210 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7214 md_set_array_sectors(mddev
, newsize
);
7215 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7216 revalidate_disk(mddev
->gendisk
);
7217 if (sectors
> mddev
->dev_sectors
&&
7218 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7219 mddev
->recovery_cp
= mddev
->dev_sectors
;
7220 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7222 mddev
->dev_sectors
= sectors
;
7223 mddev
->resync_max_sectors
= sectors
;
7227 static int check_stripe_cache(struct mddev
*mddev
)
7229 /* Can only proceed if there are plenty of stripe_heads.
7230 * We need a minimum of one full stripe,, and for sensible progress
7231 * it is best to have about 4 times that.
7232 * If we require 4 times, then the default 256 4K stripe_heads will
7233 * allow for chunk sizes up to 256K, which is probably OK.
7234 * If the chunk size is greater, user-space should request more
7235 * stripe_heads first.
7237 struct r5conf
*conf
= mddev
->private;
7238 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7239 > conf
->min_nr_stripes
||
7240 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7241 > conf
->min_nr_stripes
) {
7242 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7244 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7251 static int check_reshape(struct mddev
*mddev
)
7253 struct r5conf
*conf
= mddev
->private;
7255 if (mddev
->delta_disks
== 0 &&
7256 mddev
->new_layout
== mddev
->layout
&&
7257 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7258 return 0; /* nothing to do */
7259 if (has_failed(conf
))
7261 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7262 /* We might be able to shrink, but the devices must
7263 * be made bigger first.
7264 * For raid6, 4 is the minimum size.
7265 * Otherwise 2 is the minimum
7268 if (mddev
->level
== 6)
7270 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7274 if (!check_stripe_cache(mddev
))
7277 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
7278 mddev
->delta_disks
> 0)
7279 if (resize_chunks(conf
,
7280 conf
->previous_raid_disks
7281 + max(0, mddev
->delta_disks
),
7282 max(mddev
->new_chunk_sectors
,
7283 mddev
->chunk_sectors
)
7286 return resize_stripes(conf
, (conf
->previous_raid_disks
7287 + mddev
->delta_disks
));
7290 static int raid5_start_reshape(struct mddev
*mddev
)
7292 struct r5conf
*conf
= mddev
->private;
7293 struct md_rdev
*rdev
;
7295 unsigned long flags
;
7297 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7300 if (!check_stripe_cache(mddev
))
7303 if (has_failed(conf
))
7306 rdev_for_each(rdev
, mddev
) {
7307 if (!test_bit(In_sync
, &rdev
->flags
)
7308 && !test_bit(Faulty
, &rdev
->flags
))
7312 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7313 /* Not enough devices even to make a degraded array
7318 /* Refuse to reduce size of the array. Any reductions in
7319 * array size must be through explicit setting of array_size
7322 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7323 < mddev
->array_sectors
) {
7324 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
7325 "before number of disks\n", mdname(mddev
));
7329 atomic_set(&conf
->reshape_stripes
, 0);
7330 spin_lock_irq(&conf
->device_lock
);
7331 write_seqcount_begin(&conf
->gen_lock
);
7332 conf
->previous_raid_disks
= conf
->raid_disks
;
7333 conf
->raid_disks
+= mddev
->delta_disks
;
7334 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7335 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7336 conf
->prev_algo
= conf
->algorithm
;
7337 conf
->algorithm
= mddev
->new_layout
;
7339 /* Code that selects data_offset needs to see the generation update
7340 * if reshape_progress has been set - so a memory barrier needed.
7343 if (mddev
->reshape_backwards
)
7344 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7346 conf
->reshape_progress
= 0;
7347 conf
->reshape_safe
= conf
->reshape_progress
;
7348 write_seqcount_end(&conf
->gen_lock
);
7349 spin_unlock_irq(&conf
->device_lock
);
7351 /* Now make sure any requests that proceeded on the assumption
7352 * the reshape wasn't running - like Discard or Read - have
7355 mddev_suspend(mddev
);
7356 mddev_resume(mddev
);
7358 /* Add some new drives, as many as will fit.
7359 * We know there are enough to make the newly sized array work.
7360 * Don't add devices if we are reducing the number of
7361 * devices in the array. This is because it is not possible
7362 * to correctly record the "partially reconstructed" state of
7363 * such devices during the reshape and confusion could result.
7365 if (mddev
->delta_disks
>= 0) {
7366 rdev_for_each(rdev
, mddev
)
7367 if (rdev
->raid_disk
< 0 &&
7368 !test_bit(Faulty
, &rdev
->flags
)) {
7369 if (raid5_add_disk(mddev
, rdev
) == 0) {
7371 >= conf
->previous_raid_disks
)
7372 set_bit(In_sync
, &rdev
->flags
);
7374 rdev
->recovery_offset
= 0;
7376 if (sysfs_link_rdev(mddev
, rdev
))
7377 /* Failure here is OK */;
7379 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7380 && !test_bit(Faulty
, &rdev
->flags
)) {
7381 /* This is a spare that was manually added */
7382 set_bit(In_sync
, &rdev
->flags
);
7385 /* When a reshape changes the number of devices,
7386 * ->degraded is measured against the larger of the
7387 * pre and post number of devices.
7389 spin_lock_irqsave(&conf
->device_lock
, flags
);
7390 mddev
->degraded
= calc_degraded(conf
);
7391 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7393 mddev
->raid_disks
= conf
->raid_disks
;
7394 mddev
->reshape_position
= conf
->reshape_progress
;
7395 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7397 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7398 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7399 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
7400 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7401 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7402 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7404 if (!mddev
->sync_thread
) {
7405 mddev
->recovery
= 0;
7406 spin_lock_irq(&conf
->device_lock
);
7407 write_seqcount_begin(&conf
->gen_lock
);
7408 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7409 mddev
->new_chunk_sectors
=
7410 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7411 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7412 rdev_for_each(rdev
, mddev
)
7413 rdev
->new_data_offset
= rdev
->data_offset
;
7415 conf
->generation
--;
7416 conf
->reshape_progress
= MaxSector
;
7417 mddev
->reshape_position
= MaxSector
;
7418 write_seqcount_end(&conf
->gen_lock
);
7419 spin_unlock_irq(&conf
->device_lock
);
7422 conf
->reshape_checkpoint
= jiffies
;
7423 md_wakeup_thread(mddev
->sync_thread
);
7424 md_new_event(mddev
);
7428 /* This is called from the reshape thread and should make any
7429 * changes needed in 'conf'
7431 static void end_reshape(struct r5conf
*conf
)
7434 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7435 struct md_rdev
*rdev
;
7437 spin_lock_irq(&conf
->device_lock
);
7438 conf
->previous_raid_disks
= conf
->raid_disks
;
7439 rdev_for_each(rdev
, conf
->mddev
)
7440 rdev
->data_offset
= rdev
->new_data_offset
;
7442 conf
->reshape_progress
= MaxSector
;
7443 spin_unlock_irq(&conf
->device_lock
);
7444 wake_up(&conf
->wait_for_overlap
);
7446 /* read-ahead size must cover two whole stripes, which is
7447 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7449 if (conf
->mddev
->queue
) {
7450 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7451 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7453 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
7454 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
7459 /* This is called from the raid5d thread with mddev_lock held.
7460 * It makes config changes to the device.
7462 static void raid5_finish_reshape(struct mddev
*mddev
)
7464 struct r5conf
*conf
= mddev
->private;
7466 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7468 if (mddev
->delta_disks
> 0) {
7469 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7470 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7471 revalidate_disk(mddev
->gendisk
);
7474 spin_lock_irq(&conf
->device_lock
);
7475 mddev
->degraded
= calc_degraded(conf
);
7476 spin_unlock_irq(&conf
->device_lock
);
7477 for (d
= conf
->raid_disks
;
7478 d
< conf
->raid_disks
- mddev
->delta_disks
;
7480 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
7482 clear_bit(In_sync
, &rdev
->flags
);
7483 rdev
= conf
->disks
[d
].replacement
;
7485 clear_bit(In_sync
, &rdev
->flags
);
7488 mddev
->layout
= conf
->algorithm
;
7489 mddev
->chunk_sectors
= conf
->chunk_sectors
;
7490 mddev
->reshape_position
= MaxSector
;
7491 mddev
->delta_disks
= 0;
7492 mddev
->reshape_backwards
= 0;
7496 static void raid5_quiesce(struct mddev
*mddev
, int state
)
7498 struct r5conf
*conf
= mddev
->private;
7501 case 2: /* resume for a suspend */
7502 wake_up(&conf
->wait_for_overlap
);
7505 case 1: /* stop all writes */
7506 lock_all_device_hash_locks_irq(conf
);
7507 /* '2' tells resync/reshape to pause so that all
7508 * active stripes can drain
7511 wait_event_cmd(conf
->wait_for_quiescent
,
7512 atomic_read(&conf
->active_stripes
) == 0 &&
7513 atomic_read(&conf
->active_aligned_reads
) == 0,
7514 unlock_all_device_hash_locks_irq(conf
),
7515 lock_all_device_hash_locks_irq(conf
));
7517 unlock_all_device_hash_locks_irq(conf
);
7518 /* allow reshape to continue */
7519 wake_up(&conf
->wait_for_overlap
);
7522 case 0: /* re-enable writes */
7523 lock_all_device_hash_locks_irq(conf
);
7525 wake_up(&conf
->wait_for_quiescent
);
7526 wake_up(&conf
->wait_for_overlap
);
7527 unlock_all_device_hash_locks_irq(conf
);
7532 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
7534 struct r0conf
*raid0_conf
= mddev
->private;
7537 /* for raid0 takeover only one zone is supported */
7538 if (raid0_conf
->nr_strip_zones
> 1) {
7539 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7541 return ERR_PTR(-EINVAL
);
7544 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
7545 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
7546 mddev
->dev_sectors
= sectors
;
7547 mddev
->new_level
= level
;
7548 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7549 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
7550 mddev
->raid_disks
+= 1;
7551 mddev
->delta_disks
= 1;
7552 /* make sure it will be not marked as dirty */
7553 mddev
->recovery_cp
= MaxSector
;
7555 return setup_conf(mddev
);
7558 static void *raid5_takeover_raid1(struct mddev
*mddev
)
7562 if (mddev
->raid_disks
!= 2 ||
7563 mddev
->degraded
> 1)
7564 return ERR_PTR(-EINVAL
);
7566 /* Should check if there are write-behind devices? */
7568 chunksect
= 64*2; /* 64K by default */
7570 /* The array must be an exact multiple of chunksize */
7571 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
7574 if ((chunksect
<<9) < STRIPE_SIZE
)
7575 /* array size does not allow a suitable chunk size */
7576 return ERR_PTR(-EINVAL
);
7578 mddev
->new_level
= 5;
7579 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7580 mddev
->new_chunk_sectors
= chunksect
;
7582 return setup_conf(mddev
);
7585 static void *raid5_takeover_raid6(struct mddev
*mddev
)
7589 switch (mddev
->layout
) {
7590 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7591 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
7593 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7594 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
7596 case ALGORITHM_LEFT_SYMMETRIC_6
:
7597 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7599 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7600 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
7602 case ALGORITHM_PARITY_0_6
:
7603 new_layout
= ALGORITHM_PARITY_0
;
7605 case ALGORITHM_PARITY_N
:
7606 new_layout
= ALGORITHM_PARITY_N
;
7609 return ERR_PTR(-EINVAL
);
7611 mddev
->new_level
= 5;
7612 mddev
->new_layout
= new_layout
;
7613 mddev
->delta_disks
= -1;
7614 mddev
->raid_disks
-= 1;
7615 return setup_conf(mddev
);
7618 static int raid5_check_reshape(struct mddev
*mddev
)
7620 /* For a 2-drive array, the layout and chunk size can be changed
7621 * immediately as not restriping is needed.
7622 * For larger arrays we record the new value - after validation
7623 * to be used by a reshape pass.
7625 struct r5conf
*conf
= mddev
->private;
7626 int new_chunk
= mddev
->new_chunk_sectors
;
7628 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
7630 if (new_chunk
> 0) {
7631 if (!is_power_of_2(new_chunk
))
7633 if (new_chunk
< (PAGE_SIZE
>>9))
7635 if (mddev
->array_sectors
& (new_chunk
-1))
7636 /* not factor of array size */
7640 /* They look valid */
7642 if (mddev
->raid_disks
== 2) {
7643 /* can make the change immediately */
7644 if (mddev
->new_layout
>= 0) {
7645 conf
->algorithm
= mddev
->new_layout
;
7646 mddev
->layout
= mddev
->new_layout
;
7648 if (new_chunk
> 0) {
7649 conf
->chunk_sectors
= new_chunk
;
7650 mddev
->chunk_sectors
= new_chunk
;
7652 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7653 md_wakeup_thread(mddev
->thread
);
7655 return check_reshape(mddev
);
7658 static int raid6_check_reshape(struct mddev
*mddev
)
7660 int new_chunk
= mddev
->new_chunk_sectors
;
7662 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
7664 if (new_chunk
> 0) {
7665 if (!is_power_of_2(new_chunk
))
7667 if (new_chunk
< (PAGE_SIZE
>> 9))
7669 if (mddev
->array_sectors
& (new_chunk
-1))
7670 /* not factor of array size */
7674 /* They look valid */
7675 return check_reshape(mddev
);
7678 static void *raid5_takeover(struct mddev
*mddev
)
7680 /* raid5 can take over:
7681 * raid0 - if there is only one strip zone - make it a raid4 layout
7682 * raid1 - if there are two drives. We need to know the chunk size
7683 * raid4 - trivial - just use a raid4 layout.
7684 * raid6 - Providing it is a *_6 layout
7686 if (mddev
->level
== 0)
7687 return raid45_takeover_raid0(mddev
, 5);
7688 if (mddev
->level
== 1)
7689 return raid5_takeover_raid1(mddev
);
7690 if (mddev
->level
== 4) {
7691 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7692 mddev
->new_level
= 5;
7693 return setup_conf(mddev
);
7695 if (mddev
->level
== 6)
7696 return raid5_takeover_raid6(mddev
);
7698 return ERR_PTR(-EINVAL
);
7701 static void *raid4_takeover(struct mddev
*mddev
)
7703 /* raid4 can take over:
7704 * raid0 - if there is only one strip zone
7705 * raid5 - if layout is right
7707 if (mddev
->level
== 0)
7708 return raid45_takeover_raid0(mddev
, 4);
7709 if (mddev
->level
== 5 &&
7710 mddev
->layout
== ALGORITHM_PARITY_N
) {
7711 mddev
->new_layout
= 0;
7712 mddev
->new_level
= 4;
7713 return setup_conf(mddev
);
7715 return ERR_PTR(-EINVAL
);
7718 static struct md_personality raid5_personality
;
7720 static void *raid6_takeover(struct mddev
*mddev
)
7722 /* Currently can only take over a raid5. We map the
7723 * personality to an equivalent raid6 personality
7724 * with the Q block at the end.
7728 if (mddev
->pers
!= &raid5_personality
)
7729 return ERR_PTR(-EINVAL
);
7730 if (mddev
->degraded
> 1)
7731 return ERR_PTR(-EINVAL
);
7732 if (mddev
->raid_disks
> 253)
7733 return ERR_PTR(-EINVAL
);
7734 if (mddev
->raid_disks
< 3)
7735 return ERR_PTR(-EINVAL
);
7737 switch (mddev
->layout
) {
7738 case ALGORITHM_LEFT_ASYMMETRIC
:
7739 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
7741 case ALGORITHM_RIGHT_ASYMMETRIC
:
7742 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
7744 case ALGORITHM_LEFT_SYMMETRIC
:
7745 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
7747 case ALGORITHM_RIGHT_SYMMETRIC
:
7748 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
7750 case ALGORITHM_PARITY_0
:
7751 new_layout
= ALGORITHM_PARITY_0_6
;
7753 case ALGORITHM_PARITY_N
:
7754 new_layout
= ALGORITHM_PARITY_N
;
7757 return ERR_PTR(-EINVAL
);
7759 mddev
->new_level
= 6;
7760 mddev
->new_layout
= new_layout
;
7761 mddev
->delta_disks
= 1;
7762 mddev
->raid_disks
+= 1;
7763 return setup_conf(mddev
);
7766 static struct md_personality raid6_personality
=
7770 .owner
= THIS_MODULE
,
7771 .make_request
= make_request
,
7775 .error_handler
= error
,
7776 .hot_add_disk
= raid5_add_disk
,
7777 .hot_remove_disk
= raid5_remove_disk
,
7778 .spare_active
= raid5_spare_active
,
7779 .sync_request
= sync_request
,
7780 .resize
= raid5_resize
,
7782 .check_reshape
= raid6_check_reshape
,
7783 .start_reshape
= raid5_start_reshape
,
7784 .finish_reshape
= raid5_finish_reshape
,
7785 .quiesce
= raid5_quiesce
,
7786 .takeover
= raid6_takeover
,
7787 .congested
= raid5_congested
,
7788 .mergeable_bvec
= raid5_mergeable_bvec
,
7790 static struct md_personality raid5_personality
=
7794 .owner
= THIS_MODULE
,
7795 .make_request
= make_request
,
7799 .error_handler
= error
,
7800 .hot_add_disk
= raid5_add_disk
,
7801 .hot_remove_disk
= raid5_remove_disk
,
7802 .spare_active
= raid5_spare_active
,
7803 .sync_request
= sync_request
,
7804 .resize
= raid5_resize
,
7806 .check_reshape
= raid5_check_reshape
,
7807 .start_reshape
= raid5_start_reshape
,
7808 .finish_reshape
= raid5_finish_reshape
,
7809 .quiesce
= raid5_quiesce
,
7810 .takeover
= raid5_takeover
,
7811 .congested
= raid5_congested
,
7812 .mergeable_bvec
= raid5_mergeable_bvec
,
7815 static struct md_personality raid4_personality
=
7819 .owner
= THIS_MODULE
,
7820 .make_request
= make_request
,
7824 .error_handler
= error
,
7825 .hot_add_disk
= raid5_add_disk
,
7826 .hot_remove_disk
= raid5_remove_disk
,
7827 .spare_active
= raid5_spare_active
,
7828 .sync_request
= sync_request
,
7829 .resize
= raid5_resize
,
7831 .check_reshape
= raid5_check_reshape
,
7832 .start_reshape
= raid5_start_reshape
,
7833 .finish_reshape
= raid5_finish_reshape
,
7834 .quiesce
= raid5_quiesce
,
7835 .takeover
= raid4_takeover
,
7836 .congested
= raid5_congested
,
7837 .mergeable_bvec
= raid5_mergeable_bvec
,
7840 static int __init
raid5_init(void)
7842 raid5_wq
= alloc_workqueue("raid5wq",
7843 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
7846 register_md_personality(&raid6_personality
);
7847 register_md_personality(&raid5_personality
);
7848 register_md_personality(&raid4_personality
);
7852 static void raid5_exit(void)
7854 unregister_md_personality(&raid6_personality
);
7855 unregister_md_personality(&raid5_personality
);
7856 unregister_md_personality(&raid4_personality
);
7857 destroy_workqueue(raid5_wq
);
7860 module_init(raid5_init
);
7861 module_exit(raid5_exit
);
7862 MODULE_LICENSE("GPL");
7863 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7864 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7865 MODULE_ALIAS("md-raid5");
7866 MODULE_ALIAS("md-raid4");
7867 MODULE_ALIAS("md-level-5");
7868 MODULE_ALIAS("md-level-4");
7869 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7870 MODULE_ALIAS("md-raid6");
7871 MODULE_ALIAS("md-level-6");
7873 /* This used to be two separate modules, they were: */
7874 MODULE_ALIAS("raid5");
7875 MODULE_ALIAS("raid6");