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_list
*return_bi
)
229 while ((bi
= bio_list_pop(return_bi
)) != NULL
) {
230 bi
->bi_iter
.bi_size
= 0;
231 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
237 static void print_raid5_conf (struct r5conf
*conf
);
239 static int stripe_operations_active(struct stripe_head
*sh
)
241 return sh
->check_state
|| sh
->reconstruct_state
||
242 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
243 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
246 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
248 struct r5conf
*conf
= sh
->raid_conf
;
249 struct r5worker_group
*group
;
251 int i
, cpu
= sh
->cpu
;
253 if (!cpu_online(cpu
)) {
254 cpu
= cpumask_any(cpu_online_mask
);
258 if (list_empty(&sh
->lru
)) {
259 struct r5worker_group
*group
;
260 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
261 list_add_tail(&sh
->lru
, &group
->handle_list
);
262 group
->stripes_cnt
++;
266 if (conf
->worker_cnt_per_group
== 0) {
267 md_wakeup_thread(conf
->mddev
->thread
);
271 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
273 group
->workers
[0].working
= true;
274 /* at least one worker should run to avoid race */
275 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
277 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
278 /* wakeup more workers */
279 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
280 if (group
->workers
[i
].working
== false) {
281 group
->workers
[i
].working
= true;
282 queue_work_on(sh
->cpu
, raid5_wq
,
283 &group
->workers
[i
].work
);
289 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
290 struct list_head
*temp_inactive_list
)
292 BUG_ON(!list_empty(&sh
->lru
));
293 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
294 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
295 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
296 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
297 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
298 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
299 sh
->bm_seq
- conf
->seq_write
> 0)
300 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
302 clear_bit(STRIPE_DELAYED
, &sh
->state
);
303 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
304 if (conf
->worker_cnt_per_group
== 0) {
305 list_add_tail(&sh
->lru
, &conf
->handle_list
);
307 raid5_wakeup_stripe_thread(sh
);
311 md_wakeup_thread(conf
->mddev
->thread
);
313 BUG_ON(stripe_operations_active(sh
));
314 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
315 if (atomic_dec_return(&conf
->preread_active_stripes
)
317 md_wakeup_thread(conf
->mddev
->thread
);
318 atomic_dec(&conf
->active_stripes
);
319 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
))
320 list_add_tail(&sh
->lru
, temp_inactive_list
);
324 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
325 struct list_head
*temp_inactive_list
)
327 if (atomic_dec_and_test(&sh
->count
))
328 do_release_stripe(conf
, sh
, temp_inactive_list
);
332 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
334 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
335 * given time. Adding stripes only takes device lock, while deleting stripes
336 * only takes hash lock.
338 static void release_inactive_stripe_list(struct r5conf
*conf
,
339 struct list_head
*temp_inactive_list
,
343 unsigned long do_wakeup
= 0;
347 if (hash
== NR_STRIPE_HASH_LOCKS
) {
348 size
= NR_STRIPE_HASH_LOCKS
;
349 hash
= NR_STRIPE_HASH_LOCKS
- 1;
353 struct list_head
*list
= &temp_inactive_list
[size
- 1];
356 * We don't hold any lock here yet, raid5_get_active_stripe() might
357 * remove stripes from the list
359 if (!list_empty_careful(list
)) {
360 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
361 if (list_empty(conf
->inactive_list
+ hash
) &&
363 atomic_dec(&conf
->empty_inactive_list_nr
);
364 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
365 do_wakeup
|= 1 << hash
;
366 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
372 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++) {
373 if (do_wakeup
& (1 << i
))
374 wake_up(&conf
->wait_for_stripe
[i
]);
378 if (atomic_read(&conf
->active_stripes
) == 0)
379 wake_up(&conf
->wait_for_quiescent
);
380 if (conf
->retry_read_aligned
)
381 md_wakeup_thread(conf
->mddev
->thread
);
385 /* should hold conf->device_lock already */
386 static int release_stripe_list(struct r5conf
*conf
,
387 struct list_head
*temp_inactive_list
)
389 struct stripe_head
*sh
;
391 struct llist_node
*head
;
393 head
= llist_del_all(&conf
->released_stripes
);
394 head
= llist_reverse_order(head
);
398 sh
= llist_entry(head
, struct stripe_head
, release_list
);
399 head
= llist_next(head
);
400 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
402 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
404 * Don't worry the bit is set here, because if the bit is set
405 * again, the count is always > 1. This is true for
406 * STRIPE_ON_UNPLUG_LIST bit too.
408 hash
= sh
->hash_lock_index
;
409 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
416 void raid5_release_stripe(struct stripe_head
*sh
)
418 struct r5conf
*conf
= sh
->raid_conf
;
420 struct list_head list
;
424 /* Avoid release_list until the last reference.
426 if (atomic_add_unless(&sh
->count
, -1, 1))
429 if (unlikely(!conf
->mddev
->thread
) ||
430 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
432 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
434 md_wakeup_thread(conf
->mddev
->thread
);
437 local_irq_save(flags
);
438 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
439 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
440 INIT_LIST_HEAD(&list
);
441 hash
= sh
->hash_lock_index
;
442 do_release_stripe(conf
, sh
, &list
);
443 spin_unlock(&conf
->device_lock
);
444 release_inactive_stripe_list(conf
, &list
, hash
);
446 local_irq_restore(flags
);
449 static inline void remove_hash(struct stripe_head
*sh
)
451 pr_debug("remove_hash(), stripe %llu\n",
452 (unsigned long long)sh
->sector
);
454 hlist_del_init(&sh
->hash
);
457 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
459 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
461 pr_debug("insert_hash(), stripe %llu\n",
462 (unsigned long long)sh
->sector
);
464 hlist_add_head(&sh
->hash
, hp
);
467 /* find an idle stripe, make sure it is unhashed, and return it. */
468 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
470 struct stripe_head
*sh
= NULL
;
471 struct list_head
*first
;
473 if (list_empty(conf
->inactive_list
+ hash
))
475 first
= (conf
->inactive_list
+ hash
)->next
;
476 sh
= list_entry(first
, struct stripe_head
, lru
);
477 list_del_init(first
);
479 atomic_inc(&conf
->active_stripes
);
480 BUG_ON(hash
!= sh
->hash_lock_index
);
481 if (list_empty(conf
->inactive_list
+ hash
))
482 atomic_inc(&conf
->empty_inactive_list_nr
);
487 static void shrink_buffers(struct stripe_head
*sh
)
491 int num
= sh
->raid_conf
->pool_size
;
493 for (i
= 0; i
< num
; i
++) {
494 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
498 sh
->dev
[i
].page
= NULL
;
503 static int grow_buffers(struct stripe_head
*sh
, gfp_t gfp
)
506 int num
= sh
->raid_conf
->pool_size
;
508 for (i
= 0; i
< num
; i
++) {
511 if (!(page
= alloc_page(gfp
))) {
514 sh
->dev
[i
].page
= page
;
515 sh
->dev
[i
].orig_page
= page
;
520 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
521 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
522 struct stripe_head
*sh
);
524 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
526 struct r5conf
*conf
= sh
->raid_conf
;
529 BUG_ON(atomic_read(&sh
->count
) != 0);
530 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
531 BUG_ON(stripe_operations_active(sh
));
532 BUG_ON(sh
->batch_head
);
534 pr_debug("init_stripe called, stripe %llu\n",
535 (unsigned long long)sector
);
537 seq
= read_seqcount_begin(&conf
->gen_lock
);
538 sh
->generation
= conf
->generation
- previous
;
539 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
541 stripe_set_idx(sector
, conf
, previous
, sh
);
544 for (i
= sh
->disks
; i
--; ) {
545 struct r5dev
*dev
= &sh
->dev
[i
];
547 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
548 test_bit(R5_LOCKED
, &dev
->flags
)) {
549 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
550 (unsigned long long)sh
->sector
, i
, dev
->toread
,
551 dev
->read
, dev
->towrite
, dev
->written
,
552 test_bit(R5_LOCKED
, &dev
->flags
));
556 raid5_build_block(sh
, i
, previous
);
558 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
560 sh
->overwrite_disks
= 0;
561 insert_hash(conf
, sh
);
562 sh
->cpu
= smp_processor_id();
563 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
566 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
569 struct stripe_head
*sh
;
571 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
572 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
573 if (sh
->sector
== sector
&& sh
->generation
== generation
)
575 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
580 * Need to check if array has failed when deciding whether to:
582 * - remove non-faulty devices
585 * This determination is simple when no reshape is happening.
586 * However if there is a reshape, we need to carefully check
587 * both the before and after sections.
588 * This is because some failed devices may only affect one
589 * of the two sections, and some non-in_sync devices may
590 * be insync in the section most affected by failed devices.
592 static int calc_degraded(struct r5conf
*conf
)
594 int degraded
, degraded2
;
599 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
600 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
601 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
602 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
603 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
605 else if (test_bit(In_sync
, &rdev
->flags
))
608 /* not in-sync or faulty.
609 * If the reshape increases the number of devices,
610 * this is being recovered by the reshape, so
611 * this 'previous' section is not in_sync.
612 * If the number of devices is being reduced however,
613 * the device can only be part of the array if
614 * we are reverting a reshape, so this section will
617 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
621 if (conf
->raid_disks
== conf
->previous_raid_disks
)
625 for (i
= 0; i
< conf
->raid_disks
; i
++) {
626 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
627 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
628 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
629 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
631 else if (test_bit(In_sync
, &rdev
->flags
))
634 /* not in-sync or faulty.
635 * If reshape increases the number of devices, this
636 * section has already been recovered, else it
637 * almost certainly hasn't.
639 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
643 if (degraded2
> degraded
)
648 static int has_failed(struct r5conf
*conf
)
652 if (conf
->mddev
->reshape_position
== MaxSector
)
653 return conf
->mddev
->degraded
> conf
->max_degraded
;
655 degraded
= calc_degraded(conf
);
656 if (degraded
> conf
->max_degraded
)
662 raid5_get_active_stripe(struct r5conf
*conf
, sector_t sector
,
663 int previous
, int noblock
, int noquiesce
)
665 struct stripe_head
*sh
;
666 int hash
= stripe_hash_locks_hash(sector
);
668 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
670 spin_lock_irq(conf
->hash_locks
+ hash
);
673 wait_event_lock_irq(conf
->wait_for_quiescent
,
674 conf
->quiesce
== 0 || noquiesce
,
675 *(conf
->hash_locks
+ hash
));
676 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
678 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
679 sh
= get_free_stripe(conf
, hash
);
680 if (!sh
&& !test_bit(R5_DID_ALLOC
,
682 set_bit(R5_ALLOC_MORE
,
685 if (noblock
&& sh
== NULL
)
688 set_bit(R5_INACTIVE_BLOCKED
,
690 wait_event_exclusive_cmd(
691 conf
->wait_for_stripe
[hash
],
692 !list_empty(conf
->inactive_list
+ hash
) &&
693 (atomic_read(&conf
->active_stripes
)
694 < (conf
->max_nr_stripes
* 3 / 4)
695 || !test_bit(R5_INACTIVE_BLOCKED
,
696 &conf
->cache_state
)),
697 spin_unlock_irq(conf
->hash_locks
+ hash
),
698 spin_lock_irq(conf
->hash_locks
+ hash
));
699 clear_bit(R5_INACTIVE_BLOCKED
,
702 init_stripe(sh
, sector
, previous
);
703 atomic_inc(&sh
->count
);
705 } else if (!atomic_inc_not_zero(&sh
->count
)) {
706 spin_lock(&conf
->device_lock
);
707 if (!atomic_read(&sh
->count
)) {
708 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
709 atomic_inc(&conf
->active_stripes
);
710 BUG_ON(list_empty(&sh
->lru
) &&
711 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
712 list_del_init(&sh
->lru
);
714 sh
->group
->stripes_cnt
--;
718 atomic_inc(&sh
->count
);
719 spin_unlock(&conf
->device_lock
);
721 } while (sh
== NULL
);
723 if (!list_empty(conf
->inactive_list
+ hash
))
724 wake_up(&conf
->wait_for_stripe
[hash
]);
726 spin_unlock_irq(conf
->hash_locks
+ hash
);
730 static bool is_full_stripe_write(struct stripe_head
*sh
)
732 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
733 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
736 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
740 spin_lock(&sh2
->stripe_lock
);
741 spin_lock_nested(&sh1
->stripe_lock
, 1);
743 spin_lock(&sh1
->stripe_lock
);
744 spin_lock_nested(&sh2
->stripe_lock
, 1);
748 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
750 spin_unlock(&sh1
->stripe_lock
);
751 spin_unlock(&sh2
->stripe_lock
);
755 /* Only freshly new full stripe normal write stripe can be added to a batch list */
756 static bool stripe_can_batch(struct stripe_head
*sh
)
758 struct r5conf
*conf
= sh
->raid_conf
;
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 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
776 tmp_sec
= sh
->sector
;
777 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
779 head_sector
= sh
->sector
- STRIPE_SECTORS
;
781 hash
= stripe_hash_locks_hash(head_sector
);
782 spin_lock_irq(conf
->hash_locks
+ hash
);
783 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
784 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
785 spin_lock(&conf
->device_lock
);
786 if (!atomic_read(&head
->count
)) {
787 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
788 atomic_inc(&conf
->active_stripes
);
789 BUG_ON(list_empty(&head
->lru
) &&
790 !test_bit(STRIPE_EXPANDING
, &head
->state
));
791 list_del_init(&head
->lru
);
793 head
->group
->stripes_cnt
--;
797 atomic_inc(&head
->count
);
798 spin_unlock(&conf
->device_lock
);
800 spin_unlock_irq(conf
->hash_locks
+ hash
);
804 if (!stripe_can_batch(head
))
807 lock_two_stripes(head
, sh
);
808 /* clear_batch_ready clear the flag */
809 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
816 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
818 if (head
->dev
[dd_idx
].towrite
->bi_rw
!= sh
->dev
[dd_idx
].towrite
->bi_rw
)
821 if (head
->batch_head
) {
822 spin_lock(&head
->batch_head
->batch_lock
);
823 /* This batch list is already running */
824 if (!stripe_can_batch(head
)) {
825 spin_unlock(&head
->batch_head
->batch_lock
);
830 * at this point, head's BATCH_READY could be cleared, but we
831 * can still add the stripe to batch list
833 list_add(&sh
->batch_list
, &head
->batch_list
);
834 spin_unlock(&head
->batch_head
->batch_lock
);
836 sh
->batch_head
= head
->batch_head
;
838 head
->batch_head
= head
;
839 sh
->batch_head
= head
->batch_head
;
840 spin_lock(&head
->batch_lock
);
841 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
842 spin_unlock(&head
->batch_lock
);
845 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
846 if (atomic_dec_return(&conf
->preread_active_stripes
)
848 md_wakeup_thread(conf
->mddev
->thread
);
850 if (test_and_clear_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
851 int seq
= sh
->bm_seq
;
852 if (test_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
) &&
853 sh
->batch_head
->bm_seq
> seq
)
854 seq
= sh
->batch_head
->bm_seq
;
855 set_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
);
856 sh
->batch_head
->bm_seq
= seq
;
859 atomic_inc(&sh
->count
);
861 unlock_two_stripes(head
, sh
);
863 raid5_release_stripe(head
);
866 /* Determine if 'data_offset' or 'new_data_offset' should be used
867 * in this stripe_head.
869 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
871 sector_t progress
= conf
->reshape_progress
;
872 /* Need a memory barrier to make sure we see the value
873 * of conf->generation, or ->data_offset that was set before
874 * reshape_progress was updated.
877 if (progress
== MaxSector
)
879 if (sh
->generation
== conf
->generation
- 1)
881 /* We are in a reshape, and this is a new-generation stripe,
882 * so use new_data_offset.
888 raid5_end_read_request(struct bio
*bi
);
890 raid5_end_write_request(struct bio
*bi
);
892 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
894 struct r5conf
*conf
= sh
->raid_conf
;
895 int i
, disks
= sh
->disks
;
896 struct stripe_head
*head_sh
= sh
;
900 if (r5l_write_stripe(conf
->log
, sh
) == 0)
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_list return_bi
= BIO_EMPTY_LIST
;
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 bio_list_add(&return_bi
, rbi
);
1210 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1212 return_io(&return_bi
);
1214 set_bit(STRIPE_HANDLE
, &sh
->state
);
1215 raid5_release_stripe(sh
);
1218 static void ops_run_biofill(struct stripe_head
*sh
)
1220 struct dma_async_tx_descriptor
*tx
= NULL
;
1221 struct async_submit_ctl submit
;
1224 BUG_ON(sh
->batch_head
);
1225 pr_debug("%s: stripe %llu\n", __func__
,
1226 (unsigned long long)sh
->sector
);
1228 for (i
= sh
->disks
; i
--; ) {
1229 struct r5dev
*dev
= &sh
->dev
[i
];
1230 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1232 spin_lock_irq(&sh
->stripe_lock
);
1233 dev
->read
= rbi
= dev
->toread
;
1235 spin_unlock_irq(&sh
->stripe_lock
);
1236 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1237 dev
->sector
+ STRIPE_SECTORS
) {
1238 tx
= async_copy_data(0, rbi
, &dev
->page
,
1239 dev
->sector
, tx
, sh
);
1240 rbi
= r5_next_bio(rbi
, dev
->sector
);
1245 atomic_inc(&sh
->count
);
1246 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1247 async_trigger_callback(&submit
);
1250 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1257 tgt
= &sh
->dev
[target
];
1258 set_bit(R5_UPTODATE
, &tgt
->flags
);
1259 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1260 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1263 static void ops_complete_compute(void *stripe_head_ref
)
1265 struct stripe_head
*sh
= stripe_head_ref
;
1267 pr_debug("%s: stripe %llu\n", __func__
,
1268 (unsigned long long)sh
->sector
);
1270 /* mark the computed target(s) as uptodate */
1271 mark_target_uptodate(sh
, sh
->ops
.target
);
1272 mark_target_uptodate(sh
, sh
->ops
.target2
);
1274 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1275 if (sh
->check_state
== check_state_compute_run
)
1276 sh
->check_state
= check_state_compute_result
;
1277 set_bit(STRIPE_HANDLE
, &sh
->state
);
1278 raid5_release_stripe(sh
);
1281 /* return a pointer to the address conversion region of the scribble buffer */
1282 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1283 struct raid5_percpu
*percpu
, int i
)
1287 addr
= flex_array_get(percpu
->scribble
, i
);
1288 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1291 /* return a pointer to the address conversion region of the scribble buffer */
1292 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1296 addr
= flex_array_get(percpu
->scribble
, i
);
1300 static struct dma_async_tx_descriptor
*
1301 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1303 int disks
= sh
->disks
;
1304 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1305 int target
= sh
->ops
.target
;
1306 struct r5dev
*tgt
= &sh
->dev
[target
];
1307 struct page
*xor_dest
= tgt
->page
;
1309 struct dma_async_tx_descriptor
*tx
;
1310 struct async_submit_ctl submit
;
1313 BUG_ON(sh
->batch_head
);
1315 pr_debug("%s: stripe %llu block: %d\n",
1316 __func__
, (unsigned long long)sh
->sector
, target
);
1317 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1319 for (i
= disks
; i
--; )
1321 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1323 atomic_inc(&sh
->count
);
1325 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1326 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1327 if (unlikely(count
== 1))
1328 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1330 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1335 /* set_syndrome_sources - populate source buffers for gen_syndrome
1336 * @srcs - (struct page *) array of size sh->disks
1337 * @sh - stripe_head to parse
1339 * Populates srcs in proper layout order for the stripe and returns the
1340 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1341 * destination buffer is recorded in srcs[count] and the Q destination
1342 * is recorded in srcs[count+1]].
1344 static int set_syndrome_sources(struct page
**srcs
,
1345 struct stripe_head
*sh
,
1348 int disks
= sh
->disks
;
1349 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1350 int d0_idx
= raid6_d0(sh
);
1354 for (i
= 0; i
< disks
; i
++)
1360 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1361 struct r5dev
*dev
= &sh
->dev
[i
];
1363 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1364 (srctype
== SYNDROME_SRC_ALL
) ||
1365 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1366 test_bit(R5_Wantdrain
, &dev
->flags
)) ||
1367 (srctype
== SYNDROME_SRC_WRITTEN
&&
1369 srcs
[slot
] = sh
->dev
[i
].page
;
1370 i
= raid6_next_disk(i
, disks
);
1371 } while (i
!= d0_idx
);
1373 return syndrome_disks
;
1376 static struct dma_async_tx_descriptor
*
1377 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1379 int disks
= sh
->disks
;
1380 struct page
**blocks
= to_addr_page(percpu
, 0);
1382 int qd_idx
= sh
->qd_idx
;
1383 struct dma_async_tx_descriptor
*tx
;
1384 struct async_submit_ctl submit
;
1390 BUG_ON(sh
->batch_head
);
1391 if (sh
->ops
.target
< 0)
1392 target
= sh
->ops
.target2
;
1393 else if (sh
->ops
.target2
< 0)
1394 target
= sh
->ops
.target
;
1396 /* we should only have one valid target */
1399 pr_debug("%s: stripe %llu block: %d\n",
1400 __func__
, (unsigned long long)sh
->sector
, target
);
1402 tgt
= &sh
->dev
[target
];
1403 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1406 atomic_inc(&sh
->count
);
1408 if (target
== qd_idx
) {
1409 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1410 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1411 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1412 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1413 ops_complete_compute
, sh
,
1414 to_addr_conv(sh
, percpu
, 0));
1415 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1417 /* Compute any data- or p-drive using XOR */
1419 for (i
= disks
; i
-- ; ) {
1420 if (i
== target
|| i
== qd_idx
)
1422 blocks
[count
++] = sh
->dev
[i
].page
;
1425 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1426 NULL
, ops_complete_compute
, sh
,
1427 to_addr_conv(sh
, percpu
, 0));
1428 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1434 static struct dma_async_tx_descriptor
*
1435 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1437 int i
, count
, disks
= sh
->disks
;
1438 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1439 int d0_idx
= raid6_d0(sh
);
1440 int faila
= -1, failb
= -1;
1441 int target
= sh
->ops
.target
;
1442 int target2
= sh
->ops
.target2
;
1443 struct r5dev
*tgt
= &sh
->dev
[target
];
1444 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1445 struct dma_async_tx_descriptor
*tx
;
1446 struct page
**blocks
= to_addr_page(percpu
, 0);
1447 struct async_submit_ctl submit
;
1449 BUG_ON(sh
->batch_head
);
1450 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1451 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1452 BUG_ON(target
< 0 || target2
< 0);
1453 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1454 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1456 /* we need to open-code set_syndrome_sources to handle the
1457 * slot number conversion for 'faila' and 'failb'
1459 for (i
= 0; i
< disks
; i
++)
1464 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1466 blocks
[slot
] = sh
->dev
[i
].page
;
1472 i
= raid6_next_disk(i
, disks
);
1473 } while (i
!= d0_idx
);
1475 BUG_ON(faila
== failb
);
1478 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1479 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1481 atomic_inc(&sh
->count
);
1483 if (failb
== syndrome_disks
+1) {
1484 /* Q disk is one of the missing disks */
1485 if (faila
== syndrome_disks
) {
1486 /* Missing P+Q, just recompute */
1487 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1488 ops_complete_compute
, sh
,
1489 to_addr_conv(sh
, percpu
, 0));
1490 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1491 STRIPE_SIZE
, &submit
);
1495 int qd_idx
= sh
->qd_idx
;
1497 /* Missing D+Q: recompute D from P, then recompute Q */
1498 if (target
== qd_idx
)
1499 data_target
= target2
;
1501 data_target
= target
;
1504 for (i
= disks
; i
-- ; ) {
1505 if (i
== data_target
|| i
== qd_idx
)
1507 blocks
[count
++] = sh
->dev
[i
].page
;
1509 dest
= sh
->dev
[data_target
].page
;
1510 init_async_submit(&submit
,
1511 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1513 to_addr_conv(sh
, percpu
, 0));
1514 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1517 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1518 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1519 ops_complete_compute
, sh
,
1520 to_addr_conv(sh
, percpu
, 0));
1521 return async_gen_syndrome(blocks
, 0, count
+2,
1522 STRIPE_SIZE
, &submit
);
1525 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1526 ops_complete_compute
, sh
,
1527 to_addr_conv(sh
, percpu
, 0));
1528 if (failb
== syndrome_disks
) {
1529 /* We're missing D+P. */
1530 return async_raid6_datap_recov(syndrome_disks
+2,
1534 /* We're missing D+D. */
1535 return async_raid6_2data_recov(syndrome_disks
+2,
1536 STRIPE_SIZE
, faila
, failb
,
1542 static void ops_complete_prexor(void *stripe_head_ref
)
1544 struct stripe_head
*sh
= stripe_head_ref
;
1546 pr_debug("%s: stripe %llu\n", __func__
,
1547 (unsigned long long)sh
->sector
);
1550 static struct dma_async_tx_descriptor
*
1551 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1552 struct dma_async_tx_descriptor
*tx
)
1554 int disks
= sh
->disks
;
1555 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1556 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1557 struct async_submit_ctl submit
;
1559 /* existing parity data subtracted */
1560 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1562 BUG_ON(sh
->batch_head
);
1563 pr_debug("%s: stripe %llu\n", __func__
,
1564 (unsigned long long)sh
->sector
);
1566 for (i
= disks
; i
--; ) {
1567 struct r5dev
*dev
= &sh
->dev
[i
];
1568 /* Only process blocks that are known to be uptodate */
1569 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1570 xor_srcs
[count
++] = dev
->page
;
1573 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1574 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1575 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1580 static struct dma_async_tx_descriptor
*
1581 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1582 struct dma_async_tx_descriptor
*tx
)
1584 struct page
**blocks
= to_addr_page(percpu
, 0);
1586 struct async_submit_ctl submit
;
1588 pr_debug("%s: stripe %llu\n", __func__
,
1589 (unsigned long long)sh
->sector
);
1591 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1593 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1594 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1595 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1600 static struct dma_async_tx_descriptor
*
1601 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1603 int disks
= sh
->disks
;
1605 struct stripe_head
*head_sh
= sh
;
1607 pr_debug("%s: stripe %llu\n", __func__
,
1608 (unsigned long long)sh
->sector
);
1610 for (i
= disks
; i
--; ) {
1615 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1620 spin_lock_irq(&sh
->stripe_lock
);
1621 chosen
= dev
->towrite
;
1622 dev
->towrite
= NULL
;
1623 sh
->overwrite_disks
= 0;
1624 BUG_ON(dev
->written
);
1625 wbi
= dev
->written
= chosen
;
1626 spin_unlock_irq(&sh
->stripe_lock
);
1627 WARN_ON(dev
->page
!= dev
->orig_page
);
1629 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1630 dev
->sector
+ STRIPE_SECTORS
) {
1631 if (wbi
->bi_rw
& REQ_FUA
)
1632 set_bit(R5_WantFUA
, &dev
->flags
);
1633 if (wbi
->bi_rw
& REQ_SYNC
)
1634 set_bit(R5_SyncIO
, &dev
->flags
);
1635 if (wbi
->bi_rw
& REQ_DISCARD
)
1636 set_bit(R5_Discard
, &dev
->flags
);
1638 tx
= async_copy_data(1, wbi
, &dev
->page
,
1639 dev
->sector
, tx
, sh
);
1640 if (dev
->page
!= dev
->orig_page
) {
1641 set_bit(R5_SkipCopy
, &dev
->flags
);
1642 clear_bit(R5_UPTODATE
, &dev
->flags
);
1643 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1646 wbi
= r5_next_bio(wbi
, dev
->sector
);
1649 if (head_sh
->batch_head
) {
1650 sh
= list_first_entry(&sh
->batch_list
,
1663 static void ops_complete_reconstruct(void *stripe_head_ref
)
1665 struct stripe_head
*sh
= stripe_head_ref
;
1666 int disks
= sh
->disks
;
1667 int pd_idx
= sh
->pd_idx
;
1668 int qd_idx
= sh
->qd_idx
;
1670 bool fua
= false, sync
= false, discard
= false;
1672 pr_debug("%s: stripe %llu\n", __func__
,
1673 (unsigned long long)sh
->sector
);
1675 for (i
= disks
; i
--; ) {
1676 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1677 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1678 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1681 for (i
= disks
; i
--; ) {
1682 struct r5dev
*dev
= &sh
->dev
[i
];
1684 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1685 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1686 set_bit(R5_UPTODATE
, &dev
->flags
);
1688 set_bit(R5_WantFUA
, &dev
->flags
);
1690 set_bit(R5_SyncIO
, &dev
->flags
);
1694 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1695 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1696 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1697 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1699 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1700 sh
->reconstruct_state
= reconstruct_state_result
;
1703 set_bit(STRIPE_HANDLE
, &sh
->state
);
1704 raid5_release_stripe(sh
);
1708 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1709 struct dma_async_tx_descriptor
*tx
)
1711 int disks
= sh
->disks
;
1712 struct page
**xor_srcs
;
1713 struct async_submit_ctl submit
;
1714 int count
, pd_idx
= sh
->pd_idx
, i
;
1715 struct page
*xor_dest
;
1717 unsigned long flags
;
1719 struct stripe_head
*head_sh
= sh
;
1722 pr_debug("%s: stripe %llu\n", __func__
,
1723 (unsigned long long)sh
->sector
);
1725 for (i
= 0; i
< sh
->disks
; i
++) {
1728 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1731 if (i
>= sh
->disks
) {
1732 atomic_inc(&sh
->count
);
1733 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1734 ops_complete_reconstruct(sh
);
1739 xor_srcs
= to_addr_page(percpu
, j
);
1740 /* check if prexor is active which means only process blocks
1741 * that are part of a read-modify-write (written)
1743 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1745 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1746 for (i
= disks
; i
--; ) {
1747 struct r5dev
*dev
= &sh
->dev
[i
];
1748 if (head_sh
->dev
[i
].written
)
1749 xor_srcs
[count
++] = dev
->page
;
1752 xor_dest
= sh
->dev
[pd_idx
].page
;
1753 for (i
= disks
; i
--; ) {
1754 struct r5dev
*dev
= &sh
->dev
[i
];
1756 xor_srcs
[count
++] = dev
->page
;
1760 /* 1/ if we prexor'd then the dest is reused as a source
1761 * 2/ if we did not prexor then we are redoing the parity
1762 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1763 * for the synchronous xor case
1765 last_stripe
= !head_sh
->batch_head
||
1766 list_first_entry(&sh
->batch_list
,
1767 struct stripe_head
, batch_list
) == head_sh
;
1769 flags
= ASYNC_TX_ACK
|
1770 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1772 atomic_inc(&head_sh
->count
);
1773 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1774 to_addr_conv(sh
, percpu
, j
));
1776 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1777 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1778 to_addr_conv(sh
, percpu
, j
));
1781 if (unlikely(count
== 1))
1782 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1784 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1787 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1794 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1795 struct dma_async_tx_descriptor
*tx
)
1797 struct async_submit_ctl submit
;
1798 struct page
**blocks
;
1799 int count
, i
, j
= 0;
1800 struct stripe_head
*head_sh
= sh
;
1803 unsigned long txflags
;
1805 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1807 for (i
= 0; i
< sh
->disks
; i
++) {
1808 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1810 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1813 if (i
>= sh
->disks
) {
1814 atomic_inc(&sh
->count
);
1815 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1816 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1817 ops_complete_reconstruct(sh
);
1822 blocks
= to_addr_page(percpu
, j
);
1824 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1825 synflags
= SYNDROME_SRC_WRITTEN
;
1826 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1828 synflags
= SYNDROME_SRC_ALL
;
1829 txflags
= ASYNC_TX_ACK
;
1832 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1833 last_stripe
= !head_sh
->batch_head
||
1834 list_first_entry(&sh
->batch_list
,
1835 struct stripe_head
, batch_list
) == head_sh
;
1838 atomic_inc(&head_sh
->count
);
1839 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1840 head_sh
, to_addr_conv(sh
, percpu
, j
));
1842 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1843 to_addr_conv(sh
, percpu
, j
));
1844 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1847 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1853 static void ops_complete_check(void *stripe_head_ref
)
1855 struct stripe_head
*sh
= stripe_head_ref
;
1857 pr_debug("%s: stripe %llu\n", __func__
,
1858 (unsigned long long)sh
->sector
);
1860 sh
->check_state
= check_state_check_result
;
1861 set_bit(STRIPE_HANDLE
, &sh
->state
);
1862 raid5_release_stripe(sh
);
1865 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1867 int disks
= sh
->disks
;
1868 int pd_idx
= sh
->pd_idx
;
1869 int qd_idx
= sh
->qd_idx
;
1870 struct page
*xor_dest
;
1871 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1872 struct dma_async_tx_descriptor
*tx
;
1873 struct async_submit_ctl submit
;
1877 pr_debug("%s: stripe %llu\n", __func__
,
1878 (unsigned long long)sh
->sector
);
1880 BUG_ON(sh
->batch_head
);
1882 xor_dest
= sh
->dev
[pd_idx
].page
;
1883 xor_srcs
[count
++] = xor_dest
;
1884 for (i
= disks
; i
--; ) {
1885 if (i
== pd_idx
|| i
== qd_idx
)
1887 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1890 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1891 to_addr_conv(sh
, percpu
, 0));
1892 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1893 &sh
->ops
.zero_sum_result
, &submit
);
1895 atomic_inc(&sh
->count
);
1896 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1897 tx
= async_trigger_callback(&submit
);
1900 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1902 struct page
**srcs
= to_addr_page(percpu
, 0);
1903 struct async_submit_ctl submit
;
1906 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1907 (unsigned long long)sh
->sector
, checkp
);
1909 BUG_ON(sh
->batch_head
);
1910 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
1914 atomic_inc(&sh
->count
);
1915 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1916 sh
, to_addr_conv(sh
, percpu
, 0));
1917 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1918 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1921 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1923 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1924 struct dma_async_tx_descriptor
*tx
= NULL
;
1925 struct r5conf
*conf
= sh
->raid_conf
;
1926 int level
= conf
->level
;
1927 struct raid5_percpu
*percpu
;
1931 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1932 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1933 ops_run_biofill(sh
);
1937 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1939 tx
= ops_run_compute5(sh
, percpu
);
1941 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1942 tx
= ops_run_compute6_1(sh
, percpu
);
1944 tx
= ops_run_compute6_2(sh
, percpu
);
1946 /* terminate the chain if reconstruct is not set to be run */
1947 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1951 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
1953 tx
= ops_run_prexor5(sh
, percpu
, tx
);
1955 tx
= ops_run_prexor6(sh
, percpu
, tx
);
1958 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1959 tx
= ops_run_biodrain(sh
, tx
);
1963 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1965 ops_run_reconstruct5(sh
, percpu
, tx
);
1967 ops_run_reconstruct6(sh
, percpu
, tx
);
1970 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1971 if (sh
->check_state
== check_state_run
)
1972 ops_run_check_p(sh
, percpu
);
1973 else if (sh
->check_state
== check_state_run_q
)
1974 ops_run_check_pq(sh
, percpu
, 0);
1975 else if (sh
->check_state
== check_state_run_pq
)
1976 ops_run_check_pq(sh
, percpu
, 1);
1981 if (overlap_clear
&& !sh
->batch_head
)
1982 for (i
= disks
; i
--; ) {
1983 struct r5dev
*dev
= &sh
->dev
[i
];
1984 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1985 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1990 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
)
1992 struct stripe_head
*sh
;
1994 sh
= kmem_cache_zalloc(sc
, gfp
);
1996 spin_lock_init(&sh
->stripe_lock
);
1997 spin_lock_init(&sh
->batch_lock
);
1998 INIT_LIST_HEAD(&sh
->batch_list
);
1999 INIT_LIST_HEAD(&sh
->lru
);
2000 atomic_set(&sh
->count
, 1);
2004 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2006 struct stripe_head
*sh
;
2008 sh
= alloc_stripe(conf
->slab_cache
, gfp
);
2012 sh
->raid_conf
= conf
;
2014 if (grow_buffers(sh
, gfp
)) {
2016 kmem_cache_free(conf
->slab_cache
, sh
);
2019 sh
->hash_lock_index
=
2020 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2021 /* we just created an active stripe so... */
2022 atomic_inc(&conf
->active_stripes
);
2024 raid5_release_stripe(sh
);
2025 conf
->max_nr_stripes
++;
2029 static int grow_stripes(struct r5conf
*conf
, int num
)
2031 struct kmem_cache
*sc
;
2032 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2034 if (conf
->mddev
->gendisk
)
2035 sprintf(conf
->cache_name
[0],
2036 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2038 sprintf(conf
->cache_name
[0],
2039 "raid%d-%p", conf
->level
, conf
->mddev
);
2040 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2042 conf
->active_name
= 0;
2043 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2044 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2048 conf
->slab_cache
= sc
;
2049 conf
->pool_size
= devs
;
2051 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2058 * scribble_len - return the required size of the scribble region
2059 * @num - total number of disks in the array
2061 * The size must be enough to contain:
2062 * 1/ a struct page pointer for each device in the array +2
2063 * 2/ room to convert each entry in (1) to its corresponding dma
2064 * (dma_map_page()) or page (page_address()) address.
2066 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2067 * calculate over all devices (not just the data blocks), using zeros in place
2068 * of the P and Q blocks.
2070 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2072 struct flex_array
*ret
;
2075 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2076 ret
= flex_array_alloc(len
, cnt
, flags
);
2079 /* always prealloc all elements, so no locking is required */
2080 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2081 flex_array_free(ret
);
2087 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2092 mddev_suspend(conf
->mddev
);
2094 for_each_present_cpu(cpu
) {
2095 struct raid5_percpu
*percpu
;
2096 struct flex_array
*scribble
;
2098 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2099 scribble
= scribble_alloc(new_disks
,
2100 new_sectors
/ STRIPE_SECTORS
,
2104 flex_array_free(percpu
->scribble
);
2105 percpu
->scribble
= scribble
;
2112 mddev_resume(conf
->mddev
);
2116 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2118 /* Make all the stripes able to hold 'newsize' devices.
2119 * New slots in each stripe get 'page' set to a new page.
2121 * This happens in stages:
2122 * 1/ create a new kmem_cache and allocate the required number of
2124 * 2/ gather all the old stripe_heads and transfer the pages across
2125 * to the new stripe_heads. This will have the side effect of
2126 * freezing the array as once all stripe_heads have been collected,
2127 * no IO will be possible. Old stripe heads are freed once their
2128 * pages have been transferred over, and the old kmem_cache is
2129 * freed when all stripes are done.
2130 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2131 * we simple return a failre status - no need to clean anything up.
2132 * 4/ allocate new pages for the new slots in the new stripe_heads.
2133 * If this fails, we don't bother trying the shrink the
2134 * stripe_heads down again, we just leave them as they are.
2135 * As each stripe_head is processed the new one is released into
2138 * Once step2 is started, we cannot afford to wait for a write,
2139 * so we use GFP_NOIO allocations.
2141 struct stripe_head
*osh
, *nsh
;
2142 LIST_HEAD(newstripes
);
2143 struct disk_info
*ndisks
;
2145 struct kmem_cache
*sc
;
2149 if (newsize
<= conf
->pool_size
)
2150 return 0; /* never bother to shrink */
2152 err
= md_allow_write(conf
->mddev
);
2157 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2158 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2163 /* Need to ensure auto-resizing doesn't interfere */
2164 mutex_lock(&conf
->cache_size_mutex
);
2166 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2167 nsh
= alloc_stripe(sc
, GFP_KERNEL
);
2171 nsh
->raid_conf
= conf
;
2172 list_add(&nsh
->lru
, &newstripes
);
2175 /* didn't get enough, give up */
2176 while (!list_empty(&newstripes
)) {
2177 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2178 list_del(&nsh
->lru
);
2179 kmem_cache_free(sc
, nsh
);
2181 kmem_cache_destroy(sc
);
2182 mutex_unlock(&conf
->cache_size_mutex
);
2185 /* Step 2 - Must use GFP_NOIO now.
2186 * OK, we have enough stripes, start collecting inactive
2187 * stripes and copying them over
2191 list_for_each_entry(nsh
, &newstripes
, lru
) {
2192 lock_device_hash_lock(conf
, hash
);
2193 wait_event_exclusive_cmd(conf
->wait_for_stripe
[hash
],
2194 !list_empty(conf
->inactive_list
+ hash
),
2195 unlock_device_hash_lock(conf
, hash
),
2196 lock_device_hash_lock(conf
, hash
));
2197 osh
= get_free_stripe(conf
, hash
);
2198 unlock_device_hash_lock(conf
, hash
);
2200 for(i
=0; i
<conf
->pool_size
; i
++) {
2201 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2202 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2204 nsh
->hash_lock_index
= hash
;
2205 kmem_cache_free(conf
->slab_cache
, osh
);
2207 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2208 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2213 kmem_cache_destroy(conf
->slab_cache
);
2216 * At this point, we are holding all the stripes so the array
2217 * is completely stalled, so now is a good time to resize
2218 * conf->disks and the scribble region
2220 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2222 for (i
=0; i
<conf
->raid_disks
; i
++)
2223 ndisks
[i
] = conf
->disks
[i
];
2225 conf
->disks
= ndisks
;
2229 mutex_unlock(&conf
->cache_size_mutex
);
2230 /* Step 4, return new stripes to service */
2231 while(!list_empty(&newstripes
)) {
2232 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2233 list_del_init(&nsh
->lru
);
2235 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2236 if (nsh
->dev
[i
].page
== NULL
) {
2237 struct page
*p
= alloc_page(GFP_NOIO
);
2238 nsh
->dev
[i
].page
= p
;
2239 nsh
->dev
[i
].orig_page
= p
;
2243 raid5_release_stripe(nsh
);
2245 /* critical section pass, GFP_NOIO no longer needed */
2247 conf
->slab_cache
= sc
;
2248 conf
->active_name
= 1-conf
->active_name
;
2250 conf
->pool_size
= newsize
;
2254 static int drop_one_stripe(struct r5conf
*conf
)
2256 struct stripe_head
*sh
;
2257 int hash
= (conf
->max_nr_stripes
- 1) & STRIPE_HASH_LOCKS_MASK
;
2259 spin_lock_irq(conf
->hash_locks
+ hash
);
2260 sh
= get_free_stripe(conf
, hash
);
2261 spin_unlock_irq(conf
->hash_locks
+ hash
);
2264 BUG_ON(atomic_read(&sh
->count
));
2266 kmem_cache_free(conf
->slab_cache
, sh
);
2267 atomic_dec(&conf
->active_stripes
);
2268 conf
->max_nr_stripes
--;
2272 static void shrink_stripes(struct r5conf
*conf
)
2274 while (conf
->max_nr_stripes
&&
2275 drop_one_stripe(conf
))
2278 kmem_cache_destroy(conf
->slab_cache
);
2279 conf
->slab_cache
= NULL
;
2282 static void raid5_end_read_request(struct bio
* bi
)
2284 struct stripe_head
*sh
= bi
->bi_private
;
2285 struct r5conf
*conf
= sh
->raid_conf
;
2286 int disks
= sh
->disks
, i
;
2287 char b
[BDEVNAME_SIZE
];
2288 struct md_rdev
*rdev
= NULL
;
2291 for (i
=0 ; i
<disks
; i
++)
2292 if (bi
== &sh
->dev
[i
].req
)
2295 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2296 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2302 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2303 /* If replacement finished while this request was outstanding,
2304 * 'replacement' might be NULL already.
2305 * In that case it moved down to 'rdev'.
2306 * rdev is not removed until all requests are finished.
2308 rdev
= conf
->disks
[i
].replacement
;
2310 rdev
= conf
->disks
[i
].rdev
;
2312 if (use_new_offset(conf
, sh
))
2313 s
= sh
->sector
+ rdev
->new_data_offset
;
2315 s
= sh
->sector
+ rdev
->data_offset
;
2316 if (!bi
->bi_error
) {
2317 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2318 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2319 /* Note that this cannot happen on a
2320 * replacement device. We just fail those on
2325 "md/raid:%s: read error corrected"
2326 " (%lu sectors at %llu on %s)\n",
2327 mdname(conf
->mddev
), STRIPE_SECTORS
,
2328 (unsigned long long)s
,
2329 bdevname(rdev
->bdev
, b
));
2330 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2331 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2332 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2333 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2334 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2336 if (atomic_read(&rdev
->read_errors
))
2337 atomic_set(&rdev
->read_errors
, 0);
2339 const char *bdn
= bdevname(rdev
->bdev
, b
);
2343 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2344 atomic_inc(&rdev
->read_errors
);
2345 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2348 "md/raid:%s: read error on replacement device "
2349 "(sector %llu on %s).\n",
2350 mdname(conf
->mddev
),
2351 (unsigned long long)s
,
2353 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2357 "md/raid:%s: read error not correctable "
2358 "(sector %llu on %s).\n",
2359 mdname(conf
->mddev
),
2360 (unsigned long long)s
,
2362 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2367 "md/raid:%s: read error NOT corrected!! "
2368 "(sector %llu on %s).\n",
2369 mdname(conf
->mddev
),
2370 (unsigned long long)s
,
2372 } else if (atomic_read(&rdev
->read_errors
)
2373 > conf
->max_nr_stripes
)
2375 "md/raid:%s: Too many read errors, failing device %s.\n",
2376 mdname(conf
->mddev
), bdn
);
2379 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2380 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2383 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2384 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2385 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2387 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2389 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2390 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2392 && test_bit(In_sync
, &rdev
->flags
)
2393 && rdev_set_badblocks(
2394 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2395 md_error(conf
->mddev
, rdev
);
2398 rdev_dec_pending(rdev
, conf
->mddev
);
2399 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2400 set_bit(STRIPE_HANDLE
, &sh
->state
);
2401 raid5_release_stripe(sh
);
2404 static void raid5_end_write_request(struct bio
*bi
)
2406 struct stripe_head
*sh
= bi
->bi_private
;
2407 struct r5conf
*conf
= sh
->raid_conf
;
2408 int disks
= sh
->disks
, i
;
2409 struct md_rdev
*uninitialized_var(rdev
);
2412 int replacement
= 0;
2414 for (i
= 0 ; i
< disks
; i
++) {
2415 if (bi
== &sh
->dev
[i
].req
) {
2416 rdev
= conf
->disks
[i
].rdev
;
2419 if (bi
== &sh
->dev
[i
].rreq
) {
2420 rdev
= conf
->disks
[i
].replacement
;
2424 /* rdev was removed and 'replacement'
2425 * replaced it. rdev is not removed
2426 * until all requests are finished.
2428 rdev
= conf
->disks
[i
].rdev
;
2432 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2433 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2442 md_error(conf
->mddev
, rdev
);
2443 else if (is_badblock(rdev
, sh
->sector
,
2445 &first_bad
, &bad_sectors
))
2446 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2449 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2450 set_bit(WriteErrorSeen
, &rdev
->flags
);
2451 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2452 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2453 set_bit(MD_RECOVERY_NEEDED
,
2454 &rdev
->mddev
->recovery
);
2455 } else if (is_badblock(rdev
, sh
->sector
,
2457 &first_bad
, &bad_sectors
)) {
2458 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2459 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2460 /* That was a successful write so make
2461 * sure it looks like we already did
2464 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2467 rdev_dec_pending(rdev
, conf
->mddev
);
2469 if (sh
->batch_head
&& bi
->bi_error
&& !replacement
)
2470 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2472 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2473 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2474 set_bit(STRIPE_HANDLE
, &sh
->state
);
2475 raid5_release_stripe(sh
);
2477 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2478 raid5_release_stripe(sh
->batch_head
);
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
= raid5_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
);
2515 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
2517 "md/raid:%s: Disk failure on %s, disabling device.\n"
2518 "md/raid:%s: Operation continuing on %d devices.\n",
2520 bdevname(rdev
->bdev
, b
),
2522 conf
->raid_disks
- mddev
->degraded
);
2526 * Input: a 'big' sector number,
2527 * Output: index of the data and parity disk, and the sector # in them.
2529 sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2530 int previous
, int *dd_idx
,
2531 struct stripe_head
*sh
)
2533 sector_t stripe
, stripe2
;
2534 sector_t chunk_number
;
2535 unsigned int chunk_offset
;
2538 sector_t new_sector
;
2539 int algorithm
= previous
? conf
->prev_algo
2541 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2542 : conf
->chunk_sectors
;
2543 int raid_disks
= previous
? conf
->previous_raid_disks
2545 int data_disks
= raid_disks
- conf
->max_degraded
;
2547 /* First compute the information on this sector */
2550 * Compute the chunk number and the sector offset inside the chunk
2552 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2553 chunk_number
= r_sector
;
2556 * Compute the stripe number
2558 stripe
= chunk_number
;
2559 *dd_idx
= sector_div(stripe
, data_disks
);
2562 * Select the parity disk based on the user selected algorithm.
2564 pd_idx
= qd_idx
= -1;
2565 switch(conf
->level
) {
2567 pd_idx
= data_disks
;
2570 switch (algorithm
) {
2571 case ALGORITHM_LEFT_ASYMMETRIC
:
2572 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2573 if (*dd_idx
>= pd_idx
)
2576 case ALGORITHM_RIGHT_ASYMMETRIC
:
2577 pd_idx
= sector_div(stripe2
, raid_disks
);
2578 if (*dd_idx
>= pd_idx
)
2581 case ALGORITHM_LEFT_SYMMETRIC
:
2582 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2583 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2585 case ALGORITHM_RIGHT_SYMMETRIC
:
2586 pd_idx
= sector_div(stripe2
, raid_disks
);
2587 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2589 case ALGORITHM_PARITY_0
:
2593 case ALGORITHM_PARITY_N
:
2594 pd_idx
= data_disks
;
2602 switch (algorithm
) {
2603 case ALGORITHM_LEFT_ASYMMETRIC
:
2604 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2605 qd_idx
= pd_idx
+ 1;
2606 if (pd_idx
== raid_disks
-1) {
2607 (*dd_idx
)++; /* Q D D D P */
2609 } else if (*dd_idx
>= pd_idx
)
2610 (*dd_idx
) += 2; /* D D P Q D */
2612 case ALGORITHM_RIGHT_ASYMMETRIC
:
2613 pd_idx
= sector_div(stripe2
, raid_disks
);
2614 qd_idx
= pd_idx
+ 1;
2615 if (pd_idx
== raid_disks
-1) {
2616 (*dd_idx
)++; /* Q D D D P */
2618 } else if (*dd_idx
>= pd_idx
)
2619 (*dd_idx
) += 2; /* D D P Q D */
2621 case ALGORITHM_LEFT_SYMMETRIC
:
2622 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2623 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2624 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2626 case ALGORITHM_RIGHT_SYMMETRIC
:
2627 pd_idx
= sector_div(stripe2
, raid_disks
);
2628 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2629 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2632 case ALGORITHM_PARITY_0
:
2637 case ALGORITHM_PARITY_N
:
2638 pd_idx
= data_disks
;
2639 qd_idx
= data_disks
+ 1;
2642 case ALGORITHM_ROTATING_ZERO_RESTART
:
2643 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2644 * of blocks for computing Q is different.
2646 pd_idx
= sector_div(stripe2
, raid_disks
);
2647 qd_idx
= pd_idx
+ 1;
2648 if (pd_idx
== raid_disks
-1) {
2649 (*dd_idx
)++; /* Q D D D P */
2651 } else if (*dd_idx
>= pd_idx
)
2652 (*dd_idx
) += 2; /* D D P Q D */
2656 case ALGORITHM_ROTATING_N_RESTART
:
2657 /* Same a left_asymmetric, by first stripe is
2658 * D D D P Q rather than
2662 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2663 qd_idx
= pd_idx
+ 1;
2664 if (pd_idx
== raid_disks
-1) {
2665 (*dd_idx
)++; /* Q D D D P */
2667 } else if (*dd_idx
>= pd_idx
)
2668 (*dd_idx
) += 2; /* D D P Q D */
2672 case ALGORITHM_ROTATING_N_CONTINUE
:
2673 /* Same as left_symmetric but Q is before P */
2674 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2675 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2676 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2680 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2681 /* RAID5 left_asymmetric, with Q on last device */
2682 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2683 if (*dd_idx
>= pd_idx
)
2685 qd_idx
= raid_disks
- 1;
2688 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2689 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2690 if (*dd_idx
>= pd_idx
)
2692 qd_idx
= raid_disks
- 1;
2695 case ALGORITHM_LEFT_SYMMETRIC_6
:
2696 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2697 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2698 qd_idx
= raid_disks
- 1;
2701 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2702 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2703 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2704 qd_idx
= raid_disks
- 1;
2707 case ALGORITHM_PARITY_0_6
:
2710 qd_idx
= raid_disks
- 1;
2720 sh
->pd_idx
= pd_idx
;
2721 sh
->qd_idx
= qd_idx
;
2722 sh
->ddf_layout
= ddf_layout
;
2725 * Finally, compute the new sector number
2727 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2731 sector_t
raid5_compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2733 struct r5conf
*conf
= sh
->raid_conf
;
2734 int raid_disks
= sh
->disks
;
2735 int data_disks
= raid_disks
- conf
->max_degraded
;
2736 sector_t new_sector
= sh
->sector
, check
;
2737 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2738 : conf
->chunk_sectors
;
2739 int algorithm
= previous
? conf
->prev_algo
2743 sector_t chunk_number
;
2744 int dummy1
, dd_idx
= i
;
2746 struct stripe_head sh2
;
2748 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2749 stripe
= new_sector
;
2751 if (i
== sh
->pd_idx
)
2753 switch(conf
->level
) {
2756 switch (algorithm
) {
2757 case ALGORITHM_LEFT_ASYMMETRIC
:
2758 case ALGORITHM_RIGHT_ASYMMETRIC
:
2762 case ALGORITHM_LEFT_SYMMETRIC
:
2763 case ALGORITHM_RIGHT_SYMMETRIC
:
2766 i
-= (sh
->pd_idx
+ 1);
2768 case ALGORITHM_PARITY_0
:
2771 case ALGORITHM_PARITY_N
:
2778 if (i
== sh
->qd_idx
)
2779 return 0; /* It is the Q disk */
2780 switch (algorithm
) {
2781 case ALGORITHM_LEFT_ASYMMETRIC
:
2782 case ALGORITHM_RIGHT_ASYMMETRIC
:
2783 case ALGORITHM_ROTATING_ZERO_RESTART
:
2784 case ALGORITHM_ROTATING_N_RESTART
:
2785 if (sh
->pd_idx
== raid_disks
-1)
2786 i
--; /* Q D D D P */
2787 else if (i
> sh
->pd_idx
)
2788 i
-= 2; /* D D P Q D */
2790 case ALGORITHM_LEFT_SYMMETRIC
:
2791 case ALGORITHM_RIGHT_SYMMETRIC
:
2792 if (sh
->pd_idx
== raid_disks
-1)
2793 i
--; /* Q D D D P */
2798 i
-= (sh
->pd_idx
+ 2);
2801 case ALGORITHM_PARITY_0
:
2804 case ALGORITHM_PARITY_N
:
2806 case ALGORITHM_ROTATING_N_CONTINUE
:
2807 /* Like left_symmetric, but P is before Q */
2808 if (sh
->pd_idx
== 0)
2809 i
--; /* P D D D Q */
2814 i
-= (sh
->pd_idx
+ 1);
2817 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2818 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2822 case ALGORITHM_LEFT_SYMMETRIC_6
:
2823 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2825 i
+= data_disks
+ 1;
2826 i
-= (sh
->pd_idx
+ 1);
2828 case ALGORITHM_PARITY_0_6
:
2837 chunk_number
= stripe
* data_disks
+ i
;
2838 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2840 check
= raid5_compute_sector(conf
, r_sector
,
2841 previous
, &dummy1
, &sh2
);
2842 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2843 || sh2
.qd_idx
!= sh
->qd_idx
) {
2844 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2845 mdname(conf
->mddev
));
2852 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2853 int rcw
, int expand
)
2855 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
2856 struct r5conf
*conf
= sh
->raid_conf
;
2857 int level
= conf
->level
;
2861 for (i
= disks
; i
--; ) {
2862 struct r5dev
*dev
= &sh
->dev
[i
];
2865 set_bit(R5_LOCKED
, &dev
->flags
);
2866 set_bit(R5_Wantdrain
, &dev
->flags
);
2868 clear_bit(R5_UPTODATE
, &dev
->flags
);
2872 /* if we are not expanding this is a proper write request, and
2873 * there will be bios with new data to be drained into the
2878 /* False alarm, nothing to do */
2880 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2881 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2883 sh
->reconstruct_state
= reconstruct_state_run
;
2885 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2887 if (s
->locked
+ conf
->max_degraded
== disks
)
2888 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2889 atomic_inc(&conf
->pending_full_writes
);
2891 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2892 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2893 BUG_ON(level
== 6 &&
2894 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
2895 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
2897 for (i
= disks
; i
--; ) {
2898 struct r5dev
*dev
= &sh
->dev
[i
];
2899 if (i
== pd_idx
|| i
== qd_idx
)
2903 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2904 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2905 set_bit(R5_Wantdrain
, &dev
->flags
);
2906 set_bit(R5_LOCKED
, &dev
->flags
);
2907 clear_bit(R5_UPTODATE
, &dev
->flags
);
2912 /* False alarm - nothing to do */
2914 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2915 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2916 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2917 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2920 /* keep the parity disk(s) locked while asynchronous operations
2923 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2924 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2928 int qd_idx
= sh
->qd_idx
;
2929 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2931 set_bit(R5_LOCKED
, &dev
->flags
);
2932 clear_bit(R5_UPTODATE
, &dev
->flags
);
2936 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2937 __func__
, (unsigned long long)sh
->sector
,
2938 s
->locked
, s
->ops_request
);
2942 * Each stripe/dev can have one or more bion attached.
2943 * toread/towrite point to the first in a chain.
2944 * The bi_next chain must be in order.
2946 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
2947 int forwrite
, int previous
)
2950 struct r5conf
*conf
= sh
->raid_conf
;
2953 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2954 (unsigned long long)bi
->bi_iter
.bi_sector
,
2955 (unsigned long long)sh
->sector
);
2958 * If several bio share a stripe. The bio bi_phys_segments acts as a
2959 * reference count to avoid race. The reference count should already be
2960 * increased before this function is called (for example, in
2961 * make_request()), so other bio sharing this stripe will not free the
2962 * stripe. If a stripe is owned by one stripe, the stripe lock will
2965 spin_lock_irq(&sh
->stripe_lock
);
2966 /* Don't allow new IO added to stripes in batch list */
2970 bip
= &sh
->dev
[dd_idx
].towrite
;
2974 bip
= &sh
->dev
[dd_idx
].toread
;
2975 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
2976 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
2978 bip
= & (*bip
)->bi_next
;
2980 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
2983 if (!forwrite
|| previous
)
2984 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
2986 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2990 raid5_inc_bi_active_stripes(bi
);
2993 /* check if page is covered */
2994 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2995 for (bi
=sh
->dev
[dd_idx
].towrite
;
2996 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2997 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
2998 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2999 if (bio_end_sector(bi
) >= sector
)
3000 sector
= bio_end_sector(bi
);
3002 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
3003 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
3004 sh
->overwrite_disks
++;
3007 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3008 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
3009 (unsigned long long)sh
->sector
, dd_idx
);
3011 if (conf
->mddev
->bitmap
&& firstwrite
) {
3012 /* Cannot hold spinlock over bitmap_startwrite,
3013 * but must ensure this isn't added to a batch until
3014 * we have added to the bitmap and set bm_seq.
3015 * So set STRIPE_BITMAP_PENDING to prevent
3017 * If multiple add_stripe_bio() calls race here they
3018 * much all set STRIPE_BITMAP_PENDING. So only the first one
3019 * to complete "bitmap_startwrite" gets to set
3020 * STRIPE_BIT_DELAY. This is important as once a stripe
3021 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3024 set_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3025 spin_unlock_irq(&sh
->stripe_lock
);
3026 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
3028 spin_lock_irq(&sh
->stripe_lock
);
3029 clear_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3030 if (!sh
->batch_head
) {
3031 sh
->bm_seq
= conf
->seq_flush
+1;
3032 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
3035 spin_unlock_irq(&sh
->stripe_lock
);
3037 if (stripe_can_batch(sh
))
3038 stripe_add_to_batch_list(conf
, sh
);
3042 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
3043 spin_unlock_irq(&sh
->stripe_lock
);
3047 static void end_reshape(struct r5conf
*conf
);
3049 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3050 struct stripe_head
*sh
)
3052 int sectors_per_chunk
=
3053 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3055 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3056 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3058 raid5_compute_sector(conf
,
3059 stripe
* (disks
- conf
->max_degraded
)
3060 *sectors_per_chunk
+ chunk_offset
,
3066 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3067 struct stripe_head_state
*s
, int disks
,
3068 struct bio_list
*return_bi
)
3071 BUG_ON(sh
->batch_head
);
3072 for (i
= disks
; i
--; ) {
3076 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3077 struct md_rdev
*rdev
;
3079 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3080 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
3081 atomic_inc(&rdev
->nr_pending
);
3086 if (!rdev_set_badblocks(
3090 md_error(conf
->mddev
, rdev
);
3091 rdev_dec_pending(rdev
, conf
->mddev
);
3094 spin_lock_irq(&sh
->stripe_lock
);
3095 /* fail all writes first */
3096 bi
= sh
->dev
[i
].towrite
;
3097 sh
->dev
[i
].towrite
= NULL
;
3098 sh
->overwrite_disks
= 0;
3099 spin_unlock_irq(&sh
->stripe_lock
);
3103 r5l_stripe_write_finished(sh
);
3105 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3106 wake_up(&conf
->wait_for_overlap
);
3108 while (bi
&& bi
->bi_iter
.bi_sector
<
3109 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3110 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3112 bi
->bi_error
= -EIO
;
3113 if (!raid5_dec_bi_active_stripes(bi
)) {
3114 md_write_end(conf
->mddev
);
3115 bio_list_add(return_bi
, bi
);
3120 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3121 STRIPE_SECTORS
, 0, 0);
3123 /* and fail all 'written' */
3124 bi
= sh
->dev
[i
].written
;
3125 sh
->dev
[i
].written
= NULL
;
3126 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3127 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3128 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3131 if (bi
) bitmap_end
= 1;
3132 while (bi
&& bi
->bi_iter
.bi_sector
<
3133 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3134 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3136 bi
->bi_error
= -EIO
;
3137 if (!raid5_dec_bi_active_stripes(bi
)) {
3138 md_write_end(conf
->mddev
);
3139 bio_list_add(return_bi
, bi
);
3144 /* fail any reads if this device is non-operational and
3145 * the data has not reached the cache yet.
3147 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3148 s
->failed
> conf
->max_degraded
&&
3149 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3150 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3151 spin_lock_irq(&sh
->stripe_lock
);
3152 bi
= sh
->dev
[i
].toread
;
3153 sh
->dev
[i
].toread
= NULL
;
3154 spin_unlock_irq(&sh
->stripe_lock
);
3155 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3156 wake_up(&conf
->wait_for_overlap
);
3159 while (bi
&& bi
->bi_iter
.bi_sector
<
3160 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3161 struct bio
*nextbi
=
3162 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3164 bi
->bi_error
= -EIO
;
3165 if (!raid5_dec_bi_active_stripes(bi
))
3166 bio_list_add(return_bi
, bi
);
3171 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3172 STRIPE_SECTORS
, 0, 0);
3173 /* If we were in the middle of a write the parity block might
3174 * still be locked - so just clear all R5_LOCKED flags
3176 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3181 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3182 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3183 md_wakeup_thread(conf
->mddev
->thread
);
3187 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3188 struct stripe_head_state
*s
)
3193 BUG_ON(sh
->batch_head
);
3194 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3195 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3196 wake_up(&conf
->wait_for_overlap
);
3199 /* There is nothing more to do for sync/check/repair.
3200 * Don't even need to abort as that is handled elsewhere
3201 * if needed, and not always wanted e.g. if there is a known
3203 * For recover/replace we need to record a bad block on all
3204 * non-sync devices, or abort the recovery
3206 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3207 /* During recovery devices cannot be removed, so
3208 * locking and refcounting of rdevs is not needed
3210 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3211 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
3213 && !test_bit(Faulty
, &rdev
->flags
)
3214 && !test_bit(In_sync
, &rdev
->flags
)
3215 && !rdev_set_badblocks(rdev
, sh
->sector
,
3218 rdev
= conf
->disks
[i
].replacement
;
3220 && !test_bit(Faulty
, &rdev
->flags
)
3221 && !test_bit(In_sync
, &rdev
->flags
)
3222 && !rdev_set_badblocks(rdev
, sh
->sector
,
3227 conf
->recovery_disabled
=
3228 conf
->mddev
->recovery_disabled
;
3230 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3233 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3235 struct md_rdev
*rdev
;
3237 /* Doing recovery so rcu locking not required */
3238 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
3240 && !test_bit(Faulty
, &rdev
->flags
)
3241 && !test_bit(In_sync
, &rdev
->flags
)
3242 && (rdev
->recovery_offset
<= sh
->sector
3243 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3249 /* fetch_block - checks the given member device to see if its data needs
3250 * to be read or computed to satisfy a request.
3252 * Returns 1 when no more member devices need to be checked, otherwise returns
3253 * 0 to tell the loop in handle_stripe_fill to continue
3256 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3257 int disk_idx
, int disks
)
3259 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3260 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3261 &sh
->dev
[s
->failed_num
[1]] };
3265 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3266 test_bit(R5_UPTODATE
, &dev
->flags
))
3267 /* No point reading this as we already have it or have
3268 * decided to get it.
3273 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3274 /* We need this block to directly satisfy a request */
3277 if (s
->syncing
|| s
->expanding
||
3278 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3279 /* When syncing, or expanding we read everything.
3280 * When replacing, we need the replaced block.
3284 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3285 (s
->failed
>= 2 && fdev
[1]->toread
))
3286 /* If we want to read from a failed device, then
3287 * we need to actually read every other device.
3291 /* Sometimes neither read-modify-write nor reconstruct-write
3292 * cycles can work. In those cases we read every block we
3293 * can. Then the parity-update is certain to have enough to
3295 * This can only be a problem when we need to write something,
3296 * and some device has failed. If either of those tests
3297 * fail we need look no further.
3299 if (!s
->failed
|| !s
->to_write
)
3302 if (test_bit(R5_Insync
, &dev
->flags
) &&
3303 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3304 /* Pre-reads at not permitted until after short delay
3305 * to gather multiple requests. However if this
3306 * device is no Insync, the block could only be be computed
3307 * and there is no need to delay that.
3311 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3312 if (fdev
[i
]->towrite
&&
3313 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3314 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3315 /* If we have a partial write to a failed
3316 * device, then we will need to reconstruct
3317 * the content of that device, so all other
3318 * devices must be read.
3323 /* If we are forced to do a reconstruct-write, either because
3324 * the current RAID6 implementation only supports that, or
3325 * or because parity cannot be trusted and we are currently
3326 * recovering it, there is extra need to be careful.
3327 * If one of the devices that we would need to read, because
3328 * it is not being overwritten (and maybe not written at all)
3329 * is missing/faulty, then we need to read everything we can.
3331 if (sh
->raid_conf
->level
!= 6 &&
3332 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3333 /* reconstruct-write isn't being forced */
3335 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3336 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3337 s
->failed_num
[i
] != sh
->qd_idx
&&
3338 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3339 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3346 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3347 int disk_idx
, int disks
)
3349 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3351 /* is the data in this block needed, and can we get it? */
3352 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3353 /* we would like to get this block, possibly by computing it,
3354 * otherwise read it if the backing disk is insync
3356 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3357 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3358 BUG_ON(sh
->batch_head
);
3359 if ((s
->uptodate
== disks
- 1) &&
3360 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3361 disk_idx
== s
->failed_num
[1]))) {
3362 /* have disk failed, and we're requested to fetch it;
3365 pr_debug("Computing stripe %llu block %d\n",
3366 (unsigned long long)sh
->sector
, disk_idx
);
3367 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3368 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3369 set_bit(R5_Wantcompute
, &dev
->flags
);
3370 sh
->ops
.target
= disk_idx
;
3371 sh
->ops
.target2
= -1; /* no 2nd target */
3373 /* Careful: from this point on 'uptodate' is in the eye
3374 * of raid_run_ops which services 'compute' operations
3375 * before writes. R5_Wantcompute flags a block that will
3376 * be R5_UPTODATE by the time it is needed for a
3377 * subsequent operation.
3381 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3382 /* Computing 2-failure is *very* expensive; only
3383 * do it if failed >= 2
3386 for (other
= disks
; other
--; ) {
3387 if (other
== disk_idx
)
3389 if (!test_bit(R5_UPTODATE
,
3390 &sh
->dev
[other
].flags
))
3394 pr_debug("Computing stripe %llu blocks %d,%d\n",
3395 (unsigned long long)sh
->sector
,
3397 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3398 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3399 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3400 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3401 sh
->ops
.target
= disk_idx
;
3402 sh
->ops
.target2
= other
;
3406 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3407 set_bit(R5_LOCKED
, &dev
->flags
);
3408 set_bit(R5_Wantread
, &dev
->flags
);
3410 pr_debug("Reading block %d (sync=%d)\n",
3411 disk_idx
, s
->syncing
);
3419 * handle_stripe_fill - read or compute data to satisfy pending requests.
3421 static void handle_stripe_fill(struct stripe_head
*sh
,
3422 struct stripe_head_state
*s
,
3427 /* look for blocks to read/compute, skip this if a compute
3428 * is already in flight, or if the stripe contents are in the
3429 * midst of changing due to a write
3431 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3432 !sh
->reconstruct_state
)
3433 for (i
= disks
; i
--; )
3434 if (fetch_block(sh
, s
, i
, disks
))
3436 set_bit(STRIPE_HANDLE
, &sh
->state
);
3439 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
3440 unsigned long handle_flags
);
3441 /* handle_stripe_clean_event
3442 * any written block on an uptodate or failed drive can be returned.
3443 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3444 * never LOCKED, so we don't need to test 'failed' directly.
3446 static void handle_stripe_clean_event(struct r5conf
*conf
,
3447 struct stripe_head
*sh
, int disks
, struct bio_list
*return_bi
)
3451 int discard_pending
= 0;
3452 struct stripe_head
*head_sh
= sh
;
3453 bool do_endio
= false;
3455 for (i
= disks
; i
--; )
3456 if (sh
->dev
[i
].written
) {
3458 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3459 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3460 test_bit(R5_Discard
, &dev
->flags
) ||
3461 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3462 /* We can return any write requests */
3463 struct bio
*wbi
, *wbi2
;
3464 pr_debug("Return write for disc %d\n", i
);
3465 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3466 clear_bit(R5_UPTODATE
, &dev
->flags
);
3467 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3468 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3473 dev
->page
= dev
->orig_page
;
3475 dev
->written
= NULL
;
3476 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3477 dev
->sector
+ STRIPE_SECTORS
) {
3478 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3479 if (!raid5_dec_bi_active_stripes(wbi
)) {
3480 md_write_end(conf
->mddev
);
3481 bio_list_add(return_bi
, wbi
);
3485 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3487 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3489 if (head_sh
->batch_head
) {
3490 sh
= list_first_entry(&sh
->batch_list
,
3493 if (sh
!= head_sh
) {
3500 } else if (test_bit(R5_Discard
, &dev
->flags
))
3501 discard_pending
= 1;
3502 WARN_ON(test_bit(R5_SkipCopy
, &dev
->flags
));
3503 WARN_ON(dev
->page
!= dev
->orig_page
);
3506 r5l_stripe_write_finished(sh
);
3508 if (!discard_pending
&&
3509 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3511 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3512 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3513 if (sh
->qd_idx
>= 0) {
3514 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3515 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3517 /* now that discard is done we can proceed with any sync */
3518 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3520 * SCSI discard will change some bio fields and the stripe has
3521 * no updated data, so remove it from hash list and the stripe
3522 * will be reinitialized
3525 hash
= sh
->hash_lock_index
;
3526 spin_lock_irq(conf
->hash_locks
+ hash
);
3528 spin_unlock_irq(conf
->hash_locks
+ hash
);
3529 if (head_sh
->batch_head
) {
3530 sh
= list_first_entry(&sh
->batch_list
,
3531 struct stripe_head
, batch_list
);
3537 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3538 set_bit(STRIPE_HANDLE
, &sh
->state
);
3542 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3543 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3544 md_wakeup_thread(conf
->mddev
->thread
);
3546 if (head_sh
->batch_head
&& do_endio
)
3547 break_stripe_batch_list(head_sh
, STRIPE_EXPAND_SYNC_FLAGS
);
3550 static void handle_stripe_dirtying(struct r5conf
*conf
,
3551 struct stripe_head
*sh
,
3552 struct stripe_head_state
*s
,
3555 int rmw
= 0, rcw
= 0, i
;
3556 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3558 /* Check whether resync is now happening or should start.
3559 * If yes, then the array is dirty (after unclean shutdown or
3560 * initial creation), so parity in some stripes might be inconsistent.
3561 * In this case, we need to always do reconstruct-write, to ensure
3562 * that in case of drive failure or read-error correction, we
3563 * generate correct data from the parity.
3565 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3566 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3568 /* Calculate the real rcw later - for now make it
3569 * look like rcw is cheaper
3572 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3573 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3574 (unsigned long long)sh
->sector
);
3575 } else for (i
= disks
; i
--; ) {
3576 /* would I have to read this buffer for read_modify_write */
3577 struct r5dev
*dev
= &sh
->dev
[i
];
3578 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3579 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3580 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3581 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3582 if (test_bit(R5_Insync
, &dev
->flags
))
3585 rmw
+= 2*disks
; /* cannot read it */
3587 /* Would I have to read this buffer for reconstruct_write */
3588 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3589 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3590 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3591 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3592 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3593 if (test_bit(R5_Insync
, &dev
->flags
))
3599 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3600 (unsigned long long)sh
->sector
, rmw
, rcw
);
3601 set_bit(STRIPE_HANDLE
, &sh
->state
);
3602 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_ENABLE_RMW
)) && rmw
> 0) {
3603 /* prefer read-modify-write, but need to get some data */
3604 if (conf
->mddev
->queue
)
3605 blk_add_trace_msg(conf
->mddev
->queue
,
3606 "raid5 rmw %llu %d",
3607 (unsigned long long)sh
->sector
, rmw
);
3608 for (i
= disks
; i
--; ) {
3609 struct r5dev
*dev
= &sh
->dev
[i
];
3610 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3611 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3612 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3613 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3614 test_bit(R5_Insync
, &dev
->flags
)) {
3615 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3617 pr_debug("Read_old block %d for r-m-w\n",
3619 set_bit(R5_LOCKED
, &dev
->flags
);
3620 set_bit(R5_Wantread
, &dev
->flags
);
3623 set_bit(STRIPE_DELAYED
, &sh
->state
);
3624 set_bit(STRIPE_HANDLE
, &sh
->state
);
3629 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_ENABLE_RMW
)) && rcw
> 0) {
3630 /* want reconstruct write, but need to get some data */
3633 for (i
= disks
; i
--; ) {
3634 struct r5dev
*dev
= &sh
->dev
[i
];
3635 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3636 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3637 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3638 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3639 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3641 if (test_bit(R5_Insync
, &dev
->flags
) &&
3642 test_bit(STRIPE_PREREAD_ACTIVE
,
3644 pr_debug("Read_old block "
3645 "%d for Reconstruct\n", i
);
3646 set_bit(R5_LOCKED
, &dev
->flags
);
3647 set_bit(R5_Wantread
, &dev
->flags
);
3651 set_bit(STRIPE_DELAYED
, &sh
->state
);
3652 set_bit(STRIPE_HANDLE
, &sh
->state
);
3656 if (rcw
&& conf
->mddev
->queue
)
3657 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3658 (unsigned long long)sh
->sector
,
3659 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3662 if (rcw
> disks
&& rmw
> disks
&&
3663 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3664 set_bit(STRIPE_DELAYED
, &sh
->state
);
3666 /* now if nothing is locked, and if we have enough data,
3667 * we can start a write request
3669 /* since handle_stripe can be called at any time we need to handle the
3670 * case where a compute block operation has been submitted and then a
3671 * subsequent call wants to start a write request. raid_run_ops only
3672 * handles the case where compute block and reconstruct are requested
3673 * simultaneously. If this is not the case then new writes need to be
3674 * held off until the compute completes.
3676 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
3677 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
3678 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
3679 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
3682 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
3683 struct stripe_head_state
*s
, int disks
)
3685 struct r5dev
*dev
= NULL
;
3687 BUG_ON(sh
->batch_head
);
3688 set_bit(STRIPE_HANDLE
, &sh
->state
);
3690 switch (sh
->check_state
) {
3691 case check_state_idle
:
3692 /* start a new check operation if there are no failures */
3693 if (s
->failed
== 0) {
3694 BUG_ON(s
->uptodate
!= disks
);
3695 sh
->check_state
= check_state_run
;
3696 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3697 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3701 dev
= &sh
->dev
[s
->failed_num
[0]];
3703 case check_state_compute_result
:
3704 sh
->check_state
= check_state_idle
;
3706 dev
= &sh
->dev
[sh
->pd_idx
];
3708 /* check that a write has not made the stripe insync */
3709 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3712 /* either failed parity check, or recovery is happening */
3713 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3714 BUG_ON(s
->uptodate
!= disks
);
3716 set_bit(R5_LOCKED
, &dev
->flags
);
3718 set_bit(R5_Wantwrite
, &dev
->flags
);
3720 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3721 set_bit(STRIPE_INSYNC
, &sh
->state
);
3723 case check_state_run
:
3724 break; /* we will be called again upon completion */
3725 case check_state_check_result
:
3726 sh
->check_state
= check_state_idle
;
3728 /* if a failure occurred during the check operation, leave
3729 * STRIPE_INSYNC not set and let the stripe be handled again
3734 /* handle a successful check operation, if parity is correct
3735 * we are done. Otherwise update the mismatch count and repair
3736 * parity if !MD_RECOVERY_CHECK
3738 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3739 /* parity is correct (on disc,
3740 * not in buffer any more)
3742 set_bit(STRIPE_INSYNC
, &sh
->state
);
3744 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3745 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3746 /* don't try to repair!! */
3747 set_bit(STRIPE_INSYNC
, &sh
->state
);
3749 sh
->check_state
= check_state_compute_run
;
3750 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3751 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3752 set_bit(R5_Wantcompute
,
3753 &sh
->dev
[sh
->pd_idx
].flags
);
3754 sh
->ops
.target
= sh
->pd_idx
;
3755 sh
->ops
.target2
= -1;
3760 case check_state_compute_run
:
3763 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3764 __func__
, sh
->check_state
,
3765 (unsigned long long) sh
->sector
);
3770 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3771 struct stripe_head_state
*s
,
3774 int pd_idx
= sh
->pd_idx
;
3775 int qd_idx
= sh
->qd_idx
;
3778 BUG_ON(sh
->batch_head
);
3779 set_bit(STRIPE_HANDLE
, &sh
->state
);
3781 BUG_ON(s
->failed
> 2);
3783 /* Want to check and possibly repair P and Q.
3784 * However there could be one 'failed' device, in which
3785 * case we can only check one of them, possibly using the
3786 * other to generate missing data
3789 switch (sh
->check_state
) {
3790 case check_state_idle
:
3791 /* start a new check operation if there are < 2 failures */
3792 if (s
->failed
== s
->q_failed
) {
3793 /* The only possible failed device holds Q, so it
3794 * makes sense to check P (If anything else were failed,
3795 * we would have used P to recreate it).
3797 sh
->check_state
= check_state_run
;
3799 if (!s
->q_failed
&& s
->failed
< 2) {
3800 /* Q is not failed, and we didn't use it to generate
3801 * anything, so it makes sense to check it
3803 if (sh
->check_state
== check_state_run
)
3804 sh
->check_state
= check_state_run_pq
;
3806 sh
->check_state
= check_state_run_q
;
3809 /* discard potentially stale zero_sum_result */
3810 sh
->ops
.zero_sum_result
= 0;
3812 if (sh
->check_state
== check_state_run
) {
3813 /* async_xor_zero_sum destroys the contents of P */
3814 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3817 if (sh
->check_state
>= check_state_run
&&
3818 sh
->check_state
<= check_state_run_pq
) {
3819 /* async_syndrome_zero_sum preserves P and Q, so
3820 * no need to mark them !uptodate here
3822 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3826 /* we have 2-disk failure */
3827 BUG_ON(s
->failed
!= 2);
3829 case check_state_compute_result
:
3830 sh
->check_state
= check_state_idle
;
3832 /* check that a write has not made the stripe insync */
3833 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3836 /* now write out any block on a failed drive,
3837 * or P or Q if they were recomputed
3839 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3840 if (s
->failed
== 2) {
3841 dev
= &sh
->dev
[s
->failed_num
[1]];
3843 set_bit(R5_LOCKED
, &dev
->flags
);
3844 set_bit(R5_Wantwrite
, &dev
->flags
);
3846 if (s
->failed
>= 1) {
3847 dev
= &sh
->dev
[s
->failed_num
[0]];
3849 set_bit(R5_LOCKED
, &dev
->flags
);
3850 set_bit(R5_Wantwrite
, &dev
->flags
);
3852 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3853 dev
= &sh
->dev
[pd_idx
];
3855 set_bit(R5_LOCKED
, &dev
->flags
);
3856 set_bit(R5_Wantwrite
, &dev
->flags
);
3858 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3859 dev
= &sh
->dev
[qd_idx
];
3861 set_bit(R5_LOCKED
, &dev
->flags
);
3862 set_bit(R5_Wantwrite
, &dev
->flags
);
3864 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3866 set_bit(STRIPE_INSYNC
, &sh
->state
);
3868 case check_state_run
:
3869 case check_state_run_q
:
3870 case check_state_run_pq
:
3871 break; /* we will be called again upon completion */
3872 case check_state_check_result
:
3873 sh
->check_state
= check_state_idle
;
3875 /* handle a successful check operation, if parity is correct
3876 * we are done. Otherwise update the mismatch count and repair
3877 * parity if !MD_RECOVERY_CHECK
3879 if (sh
->ops
.zero_sum_result
== 0) {
3880 /* both parities are correct */
3882 set_bit(STRIPE_INSYNC
, &sh
->state
);
3884 /* in contrast to the raid5 case we can validate
3885 * parity, but still have a failure to write
3888 sh
->check_state
= check_state_compute_result
;
3889 /* Returning at this point means that we may go
3890 * off and bring p and/or q uptodate again so
3891 * we make sure to check zero_sum_result again
3892 * to verify if p or q need writeback
3896 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3897 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3898 /* don't try to repair!! */
3899 set_bit(STRIPE_INSYNC
, &sh
->state
);
3901 int *target
= &sh
->ops
.target
;
3903 sh
->ops
.target
= -1;
3904 sh
->ops
.target2
= -1;
3905 sh
->check_state
= check_state_compute_run
;
3906 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3907 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3908 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3909 set_bit(R5_Wantcompute
,
3910 &sh
->dev
[pd_idx
].flags
);
3912 target
= &sh
->ops
.target2
;
3915 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3916 set_bit(R5_Wantcompute
,
3917 &sh
->dev
[qd_idx
].flags
);
3924 case check_state_compute_run
:
3927 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3928 __func__
, sh
->check_state
,
3929 (unsigned long long) sh
->sector
);
3934 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3938 /* We have read all the blocks in this stripe and now we need to
3939 * copy some of them into a target stripe for expand.
3941 struct dma_async_tx_descriptor
*tx
= NULL
;
3942 BUG_ON(sh
->batch_head
);
3943 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3944 for (i
= 0; i
< sh
->disks
; i
++)
3945 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3947 struct stripe_head
*sh2
;
3948 struct async_submit_ctl submit
;
3950 sector_t bn
= raid5_compute_blocknr(sh
, i
, 1);
3951 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3953 sh2
= raid5_get_active_stripe(conf
, s
, 0, 1, 1);
3955 /* so far only the early blocks of this stripe
3956 * have been requested. When later blocks
3957 * get requested, we will try again
3960 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3961 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3962 /* must have already done this block */
3963 raid5_release_stripe(sh2
);
3967 /* place all the copies on one channel */
3968 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3969 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3970 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3973 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3974 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3975 for (j
= 0; j
< conf
->raid_disks
; j
++)
3976 if (j
!= sh2
->pd_idx
&&
3978 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3980 if (j
== conf
->raid_disks
) {
3981 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3982 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3984 raid5_release_stripe(sh2
);
3987 /* done submitting copies, wait for them to complete */
3988 async_tx_quiesce(&tx
);
3992 * handle_stripe - do things to a stripe.
3994 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3995 * state of various bits to see what needs to be done.
3997 * return some read requests which now have data
3998 * return some write requests which are safely on storage
3999 * schedule a read on some buffers
4000 * schedule a write of some buffers
4001 * return confirmation of parity correctness
4005 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
4007 struct r5conf
*conf
= sh
->raid_conf
;
4008 int disks
= sh
->disks
;
4011 int do_recovery
= 0;
4013 memset(s
, 0, sizeof(*s
));
4015 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
4016 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
4017 s
->failed_num
[0] = -1;
4018 s
->failed_num
[1] = -1;
4019 s
->log_failed
= r5l_log_disk_error(conf
);
4021 /* Now to look around and see what can be done */
4023 for (i
=disks
; i
--; ) {
4024 struct md_rdev
*rdev
;
4031 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4033 dev
->toread
, dev
->towrite
, dev
->written
);
4034 /* maybe we can reply to a read
4036 * new wantfill requests are only permitted while
4037 * ops_complete_biofill is guaranteed to be inactive
4039 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4040 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4041 set_bit(R5_Wantfill
, &dev
->flags
);
4043 /* now count some things */
4044 if (test_bit(R5_LOCKED
, &dev
->flags
))
4046 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4048 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4050 BUG_ON(s
->compute
> 2);
4053 if (test_bit(R5_Wantfill
, &dev
->flags
))
4055 else if (dev
->toread
)
4059 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4064 /* Prefer to use the replacement for reads, but only
4065 * if it is recovered enough and has no bad blocks.
4067 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4068 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4069 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4070 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4071 &first_bad
, &bad_sectors
))
4072 set_bit(R5_ReadRepl
, &dev
->flags
);
4074 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4075 set_bit(R5_NeedReplace
, &dev
->flags
);
4077 clear_bit(R5_NeedReplace
, &dev
->flags
);
4078 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4079 clear_bit(R5_ReadRepl
, &dev
->flags
);
4081 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4084 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4085 &first_bad
, &bad_sectors
);
4086 if (s
->blocked_rdev
== NULL
4087 && (test_bit(Blocked
, &rdev
->flags
)
4090 set_bit(BlockedBadBlocks
,
4092 s
->blocked_rdev
= rdev
;
4093 atomic_inc(&rdev
->nr_pending
);
4096 clear_bit(R5_Insync
, &dev
->flags
);
4100 /* also not in-sync */
4101 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4102 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4103 /* treat as in-sync, but with a read error
4104 * which we can now try to correct
4106 set_bit(R5_Insync
, &dev
->flags
);
4107 set_bit(R5_ReadError
, &dev
->flags
);
4109 } else if (test_bit(In_sync
, &rdev
->flags
))
4110 set_bit(R5_Insync
, &dev
->flags
);
4111 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4112 /* in sync if before recovery_offset */
4113 set_bit(R5_Insync
, &dev
->flags
);
4114 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4115 test_bit(R5_Expanded
, &dev
->flags
))
4116 /* If we've reshaped into here, we assume it is Insync.
4117 * We will shortly update recovery_offset to make
4120 set_bit(R5_Insync
, &dev
->flags
);
4122 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4123 /* This flag does not apply to '.replacement'
4124 * only to .rdev, so make sure to check that*/
4125 struct md_rdev
*rdev2
= rcu_dereference(
4126 conf
->disks
[i
].rdev
);
4128 clear_bit(R5_Insync
, &dev
->flags
);
4129 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4130 s
->handle_bad_blocks
= 1;
4131 atomic_inc(&rdev2
->nr_pending
);
4133 clear_bit(R5_WriteError
, &dev
->flags
);
4135 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4136 /* This flag does not apply to '.replacement'
4137 * only to .rdev, so make sure to check that*/
4138 struct md_rdev
*rdev2
= rcu_dereference(
4139 conf
->disks
[i
].rdev
);
4140 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4141 s
->handle_bad_blocks
= 1;
4142 atomic_inc(&rdev2
->nr_pending
);
4144 clear_bit(R5_MadeGood
, &dev
->flags
);
4146 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4147 struct md_rdev
*rdev2
= rcu_dereference(
4148 conf
->disks
[i
].replacement
);
4149 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4150 s
->handle_bad_blocks
= 1;
4151 atomic_inc(&rdev2
->nr_pending
);
4153 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4155 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4156 /* The ReadError flag will just be confusing now */
4157 clear_bit(R5_ReadError
, &dev
->flags
);
4158 clear_bit(R5_ReWrite
, &dev
->flags
);
4160 if (test_bit(R5_ReadError
, &dev
->flags
))
4161 clear_bit(R5_Insync
, &dev
->flags
);
4162 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4164 s
->failed_num
[s
->failed
] = i
;
4166 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4170 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4171 /* If there is a failed device being replaced,
4172 * we must be recovering.
4173 * else if we are after recovery_cp, we must be syncing
4174 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4175 * else we can only be replacing
4176 * sync and recovery both need to read all devices, and so
4177 * use the same flag.
4180 sh
->sector
>= conf
->mddev
->recovery_cp
||
4181 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4189 static int clear_batch_ready(struct stripe_head
*sh
)
4191 /* Return '1' if this is a member of batch, or
4192 * '0' if it is a lone stripe or a head which can now be
4195 struct stripe_head
*tmp
;
4196 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4197 return (sh
->batch_head
&& sh
->batch_head
!= sh
);
4198 spin_lock(&sh
->stripe_lock
);
4199 if (!sh
->batch_head
) {
4200 spin_unlock(&sh
->stripe_lock
);
4205 * this stripe could be added to a batch list before we check
4206 * BATCH_READY, skips it
4208 if (sh
->batch_head
!= sh
) {
4209 spin_unlock(&sh
->stripe_lock
);
4212 spin_lock(&sh
->batch_lock
);
4213 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4214 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4215 spin_unlock(&sh
->batch_lock
);
4216 spin_unlock(&sh
->stripe_lock
);
4219 * BATCH_READY is cleared, no new stripes can be added.
4220 * batch_list can be accessed without lock
4225 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4226 unsigned long handle_flags
)
4228 struct stripe_head
*sh
, *next
;
4232 list_for_each_entry_safe(sh
, next
, &head_sh
->batch_list
, batch_list
) {
4234 list_del_init(&sh
->batch_list
);
4236 WARN_ON_ONCE(sh
->state
& ((1 << STRIPE_ACTIVE
) |
4237 (1 << STRIPE_SYNCING
) |
4238 (1 << STRIPE_REPLACED
) |
4239 (1 << STRIPE_PREREAD_ACTIVE
) |
4240 (1 << STRIPE_DELAYED
) |
4241 (1 << STRIPE_BIT_DELAY
) |
4242 (1 << STRIPE_FULL_WRITE
) |
4243 (1 << STRIPE_BIOFILL_RUN
) |
4244 (1 << STRIPE_COMPUTE_RUN
) |
4245 (1 << STRIPE_OPS_REQ_PENDING
) |
4246 (1 << STRIPE_DISCARD
) |
4247 (1 << STRIPE_BATCH_READY
) |
4248 (1 << STRIPE_BATCH_ERR
) |
4249 (1 << STRIPE_BITMAP_PENDING
)));
4250 WARN_ON_ONCE(head_sh
->state
& ((1 << STRIPE_DISCARD
) |
4251 (1 << STRIPE_REPLACED
)));
4253 set_mask_bits(&sh
->state
, ~(STRIPE_EXPAND_SYNC_FLAGS
|
4254 (1 << STRIPE_DEGRADED
)),
4255 head_sh
->state
& (1 << STRIPE_INSYNC
));
4257 sh
->check_state
= head_sh
->check_state
;
4258 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4259 for (i
= 0; i
< sh
->disks
; i
++) {
4260 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
4262 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4263 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4265 spin_lock_irq(&sh
->stripe_lock
);
4266 sh
->batch_head
= NULL
;
4267 spin_unlock_irq(&sh
->stripe_lock
);
4268 if (handle_flags
== 0 ||
4269 sh
->state
& handle_flags
)
4270 set_bit(STRIPE_HANDLE
, &sh
->state
);
4271 raid5_release_stripe(sh
);
4273 spin_lock_irq(&head_sh
->stripe_lock
);
4274 head_sh
->batch_head
= NULL
;
4275 spin_unlock_irq(&head_sh
->stripe_lock
);
4276 for (i
= 0; i
< head_sh
->disks
; i
++)
4277 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
4279 if (head_sh
->state
& handle_flags
)
4280 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
4283 wake_up(&head_sh
->raid_conf
->wait_for_overlap
);
4286 static void handle_stripe(struct stripe_head
*sh
)
4288 struct stripe_head_state s
;
4289 struct r5conf
*conf
= sh
->raid_conf
;
4292 int disks
= sh
->disks
;
4293 struct r5dev
*pdev
, *qdev
;
4295 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4296 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4297 /* already being handled, ensure it gets handled
4298 * again when current action finishes */
4299 set_bit(STRIPE_HANDLE
, &sh
->state
);
4303 if (clear_batch_ready(sh
) ) {
4304 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4308 if (test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4309 break_stripe_batch_list(sh
, 0);
4311 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4312 spin_lock(&sh
->stripe_lock
);
4313 /* Cannot process 'sync' concurrently with 'discard' */
4314 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4315 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4316 set_bit(STRIPE_SYNCING
, &sh
->state
);
4317 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4318 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4320 spin_unlock(&sh
->stripe_lock
);
4322 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4324 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4325 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4326 (unsigned long long)sh
->sector
, sh
->state
,
4327 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4328 sh
->check_state
, sh
->reconstruct_state
);
4330 analyse_stripe(sh
, &s
);
4332 if (test_bit(STRIPE_LOG_TRAPPED
, &sh
->state
))
4335 if (s
.handle_bad_blocks
) {
4336 set_bit(STRIPE_HANDLE
, &sh
->state
);
4340 if (unlikely(s
.blocked_rdev
)) {
4341 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4342 s
.replacing
|| s
.to_write
|| s
.written
) {
4343 set_bit(STRIPE_HANDLE
, &sh
->state
);
4346 /* There is nothing for the blocked_rdev to block */
4347 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4348 s
.blocked_rdev
= NULL
;
4351 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4352 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4353 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4356 pr_debug("locked=%d uptodate=%d to_read=%d"
4357 " to_write=%d failed=%d failed_num=%d,%d\n",
4358 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4359 s
.failed_num
[0], s
.failed_num
[1]);
4360 /* check if the array has lost more than max_degraded devices and,
4361 * if so, some requests might need to be failed.
4363 if (s
.failed
> conf
->max_degraded
|| s
.log_failed
) {
4364 sh
->check_state
= 0;
4365 sh
->reconstruct_state
= 0;
4366 break_stripe_batch_list(sh
, 0);
4367 if (s
.to_read
+s
.to_write
+s
.written
)
4368 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
4369 if (s
.syncing
+ s
.replacing
)
4370 handle_failed_sync(conf
, sh
, &s
);
4373 /* Now we check to see if any write operations have recently
4377 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4379 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4380 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4381 sh
->reconstruct_state
= reconstruct_state_idle
;
4383 /* All the 'written' buffers and the parity block are ready to
4384 * be written back to disk
4386 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4387 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4388 BUG_ON(sh
->qd_idx
>= 0 &&
4389 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4390 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4391 for (i
= disks
; i
--; ) {
4392 struct r5dev
*dev
= &sh
->dev
[i
];
4393 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4394 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4396 pr_debug("Writing block %d\n", i
);
4397 set_bit(R5_Wantwrite
, &dev
->flags
);
4402 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4403 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4405 set_bit(STRIPE_INSYNC
, &sh
->state
);
4408 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4409 s
.dec_preread_active
= 1;
4413 * might be able to return some write requests if the parity blocks
4414 * are safe, or on a failed drive
4416 pdev
= &sh
->dev
[sh
->pd_idx
];
4417 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4418 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4419 qdev
= &sh
->dev
[sh
->qd_idx
];
4420 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4421 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4425 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4426 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4427 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4428 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4429 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4430 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4431 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4432 test_bit(R5_Discard
, &qdev
->flags
))))))
4433 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
4435 /* Now we might consider reading some blocks, either to check/generate
4436 * parity, or to satisfy requests
4437 * or to load a block that is being partially written.
4439 if (s
.to_read
|| s
.non_overwrite
4440 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4441 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4444 handle_stripe_fill(sh
, &s
, disks
);
4446 /* Now to consider new write requests and what else, if anything
4447 * should be read. We do not handle new writes when:
4448 * 1/ A 'write' operation (copy+xor) is already in flight.
4449 * 2/ A 'check' operation is in flight, as it may clobber the parity
4452 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
4453 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4455 /* maybe we need to check and possibly fix the parity for this stripe
4456 * Any reads will already have been scheduled, so we just see if enough
4457 * data is available. The parity check is held off while parity
4458 * dependent operations are in flight.
4460 if (sh
->check_state
||
4461 (s
.syncing
&& s
.locked
== 0 &&
4462 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4463 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4464 if (conf
->level
== 6)
4465 handle_parity_checks6(conf
, sh
, &s
, disks
);
4467 handle_parity_checks5(conf
, sh
, &s
, disks
);
4470 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4471 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4472 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4473 /* Write out to replacement devices where possible */
4474 for (i
= 0; i
< conf
->raid_disks
; i
++)
4475 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4476 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4477 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4478 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4482 set_bit(STRIPE_INSYNC
, &sh
->state
);
4483 set_bit(STRIPE_REPLACED
, &sh
->state
);
4485 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4486 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4487 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4488 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4489 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4490 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4491 wake_up(&conf
->wait_for_overlap
);
4494 /* If the failed drives are just a ReadError, then we might need
4495 * to progress the repair/check process
4497 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4498 for (i
= 0; i
< s
.failed
; i
++) {
4499 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4500 if (test_bit(R5_ReadError
, &dev
->flags
)
4501 && !test_bit(R5_LOCKED
, &dev
->flags
)
4502 && test_bit(R5_UPTODATE
, &dev
->flags
)
4504 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4505 set_bit(R5_Wantwrite
, &dev
->flags
);
4506 set_bit(R5_ReWrite
, &dev
->flags
);
4507 set_bit(R5_LOCKED
, &dev
->flags
);
4510 /* let's read it back */
4511 set_bit(R5_Wantread
, &dev
->flags
);
4512 set_bit(R5_LOCKED
, &dev
->flags
);
4518 /* Finish reconstruct operations initiated by the expansion process */
4519 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4520 struct stripe_head
*sh_src
4521 = raid5_get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4522 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4523 /* sh cannot be written until sh_src has been read.
4524 * so arrange for sh to be delayed a little
4526 set_bit(STRIPE_DELAYED
, &sh
->state
);
4527 set_bit(STRIPE_HANDLE
, &sh
->state
);
4528 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4530 atomic_inc(&conf
->preread_active_stripes
);
4531 raid5_release_stripe(sh_src
);
4535 raid5_release_stripe(sh_src
);
4537 sh
->reconstruct_state
= reconstruct_state_idle
;
4538 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4539 for (i
= conf
->raid_disks
; i
--; ) {
4540 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4541 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4546 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4547 !sh
->reconstruct_state
) {
4548 /* Need to write out all blocks after computing parity */
4549 sh
->disks
= conf
->raid_disks
;
4550 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4551 schedule_reconstruction(sh
, &s
, 1, 1);
4552 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4553 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4554 atomic_dec(&conf
->reshape_stripes
);
4555 wake_up(&conf
->wait_for_overlap
);
4556 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4559 if (s
.expanding
&& s
.locked
== 0 &&
4560 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4561 handle_stripe_expansion(conf
, sh
);
4564 /* wait for this device to become unblocked */
4565 if (unlikely(s
.blocked_rdev
)) {
4566 if (conf
->mddev
->external
)
4567 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4570 /* Internal metadata will immediately
4571 * be written by raid5d, so we don't
4572 * need to wait here.
4574 rdev_dec_pending(s
.blocked_rdev
,
4578 if (s
.handle_bad_blocks
)
4579 for (i
= disks
; i
--; ) {
4580 struct md_rdev
*rdev
;
4581 struct r5dev
*dev
= &sh
->dev
[i
];
4582 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4583 /* We own a safe reference to the rdev */
4584 rdev
= conf
->disks
[i
].rdev
;
4585 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4587 md_error(conf
->mddev
, rdev
);
4588 rdev_dec_pending(rdev
, conf
->mddev
);
4590 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4591 rdev
= conf
->disks
[i
].rdev
;
4592 rdev_clear_badblocks(rdev
, sh
->sector
,
4594 rdev_dec_pending(rdev
, conf
->mddev
);
4596 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4597 rdev
= conf
->disks
[i
].replacement
;
4599 /* rdev have been moved down */
4600 rdev
= conf
->disks
[i
].rdev
;
4601 rdev_clear_badblocks(rdev
, sh
->sector
,
4603 rdev_dec_pending(rdev
, conf
->mddev
);
4608 raid_run_ops(sh
, s
.ops_request
);
4612 if (s
.dec_preread_active
) {
4613 /* We delay this until after ops_run_io so that if make_request
4614 * is waiting on a flush, it won't continue until the writes
4615 * have actually been submitted.
4617 atomic_dec(&conf
->preread_active_stripes
);
4618 if (atomic_read(&conf
->preread_active_stripes
) <
4620 md_wakeup_thread(conf
->mddev
->thread
);
4623 if (!bio_list_empty(&s
.return_bi
)) {
4624 if (test_bit(MD_CHANGE_PENDING
, &conf
->mddev
->flags
)) {
4625 spin_lock_irq(&conf
->device_lock
);
4626 bio_list_merge(&conf
->return_bi
, &s
.return_bi
);
4627 spin_unlock_irq(&conf
->device_lock
);
4628 md_wakeup_thread(conf
->mddev
->thread
);
4630 return_io(&s
.return_bi
);
4633 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4636 static void raid5_activate_delayed(struct r5conf
*conf
)
4638 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
4639 while (!list_empty(&conf
->delayed_list
)) {
4640 struct list_head
*l
= conf
->delayed_list
.next
;
4641 struct stripe_head
*sh
;
4642 sh
= list_entry(l
, struct stripe_head
, lru
);
4644 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4645 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4646 atomic_inc(&conf
->preread_active_stripes
);
4647 list_add_tail(&sh
->lru
, &conf
->hold_list
);
4648 raid5_wakeup_stripe_thread(sh
);
4653 static void activate_bit_delay(struct r5conf
*conf
,
4654 struct list_head
*temp_inactive_list
)
4656 /* device_lock is held */
4657 struct list_head head
;
4658 list_add(&head
, &conf
->bitmap_list
);
4659 list_del_init(&conf
->bitmap_list
);
4660 while (!list_empty(&head
)) {
4661 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
4663 list_del_init(&sh
->lru
);
4664 atomic_inc(&sh
->count
);
4665 hash
= sh
->hash_lock_index
;
4666 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
4670 static int raid5_congested(struct mddev
*mddev
, int bits
)
4672 struct r5conf
*conf
= mddev
->private;
4674 /* No difference between reads and writes. Just check
4675 * how busy the stripe_cache is
4678 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
4682 if (atomic_read(&conf
->empty_inactive_list_nr
))
4688 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
4690 struct r5conf
*conf
= mddev
->private;
4691 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
4692 unsigned int chunk_sectors
;
4693 unsigned int bio_sectors
= bio_sectors(bio
);
4695 chunk_sectors
= min(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
4696 return chunk_sectors
>=
4697 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
4701 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4702 * later sampled by raid5d.
4704 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
4706 unsigned long flags
;
4708 spin_lock_irqsave(&conf
->device_lock
, flags
);
4710 bi
->bi_next
= conf
->retry_read_aligned_list
;
4711 conf
->retry_read_aligned_list
= bi
;
4713 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4714 md_wakeup_thread(conf
->mddev
->thread
);
4717 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
4721 bi
= conf
->retry_read_aligned
;
4723 conf
->retry_read_aligned
= NULL
;
4726 bi
= conf
->retry_read_aligned_list
;
4728 conf
->retry_read_aligned_list
= bi
->bi_next
;
4731 * this sets the active strip count to 1 and the processed
4732 * strip count to zero (upper 8 bits)
4734 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
4741 * The "raid5_align_endio" should check if the read succeeded and if it
4742 * did, call bio_endio on the original bio (having bio_put the new bio
4744 * If the read failed..
4746 static void raid5_align_endio(struct bio
*bi
)
4748 struct bio
* raid_bi
= bi
->bi_private
;
4749 struct mddev
*mddev
;
4750 struct r5conf
*conf
;
4751 struct md_rdev
*rdev
;
4752 int error
= bi
->bi_error
;
4756 rdev
= (void*)raid_bi
->bi_next
;
4757 raid_bi
->bi_next
= NULL
;
4758 mddev
= rdev
->mddev
;
4759 conf
= mddev
->private;
4761 rdev_dec_pending(rdev
, conf
->mddev
);
4764 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
4767 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4768 wake_up(&conf
->wait_for_quiescent
);
4772 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4774 add_bio_to_retry(raid_bi
, conf
);
4777 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
4779 struct r5conf
*conf
= mddev
->private;
4781 struct bio
* align_bi
;
4782 struct md_rdev
*rdev
;
4783 sector_t end_sector
;
4785 if (!in_chunk_boundary(mddev
, raid_bio
)) {
4786 pr_debug("%s: non aligned\n", __func__
);
4790 * use bio_clone_mddev to make a copy of the bio
4792 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
4796 * set bi_end_io to a new function, and set bi_private to the
4799 align_bi
->bi_end_io
= raid5_align_endio
;
4800 align_bi
->bi_private
= raid_bio
;
4804 align_bi
->bi_iter
.bi_sector
=
4805 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
4808 end_sector
= bio_end_sector(align_bi
);
4810 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
4811 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
4812 rdev
->recovery_offset
< end_sector
) {
4813 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
4815 (test_bit(Faulty
, &rdev
->flags
) ||
4816 !(test_bit(In_sync
, &rdev
->flags
) ||
4817 rdev
->recovery_offset
>= end_sector
)))
4824 atomic_inc(&rdev
->nr_pending
);
4826 raid_bio
->bi_next
= (void*)rdev
;
4827 align_bi
->bi_bdev
= rdev
->bdev
;
4828 bio_clear_flag(align_bi
, BIO_SEG_VALID
);
4830 if (is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
4831 bio_sectors(align_bi
),
4832 &first_bad
, &bad_sectors
)) {
4834 rdev_dec_pending(rdev
, mddev
);
4838 /* No reshape active, so we can trust rdev->data_offset */
4839 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
4841 spin_lock_irq(&conf
->device_lock
);
4842 wait_event_lock_irq(conf
->wait_for_quiescent
,
4845 atomic_inc(&conf
->active_aligned_reads
);
4846 spin_unlock_irq(&conf
->device_lock
);
4849 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4850 align_bi
, disk_devt(mddev
->gendisk
),
4851 raid_bio
->bi_iter
.bi_sector
);
4852 generic_make_request(align_bi
);
4861 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
4866 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
4867 unsigned chunk_sects
= mddev
->chunk_sectors
;
4868 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
4870 if (sectors
< bio_sectors(raid_bio
)) {
4871 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, fs_bio_set
);
4872 bio_chain(split
, raid_bio
);
4876 if (!raid5_read_one_chunk(mddev
, split
)) {
4877 if (split
!= raid_bio
)
4878 generic_make_request(raid_bio
);
4881 } while (split
!= raid_bio
);
4886 /* __get_priority_stripe - get the next stripe to process
4888 * Full stripe writes are allowed to pass preread active stripes up until
4889 * the bypass_threshold is exceeded. In general the bypass_count
4890 * increments when the handle_list is handled before the hold_list; however, it
4891 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4892 * stripe with in flight i/o. The bypass_count will be reset when the
4893 * head of the hold_list has changed, i.e. the head was promoted to the
4896 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
4898 struct stripe_head
*sh
= NULL
, *tmp
;
4899 struct list_head
*handle_list
= NULL
;
4900 struct r5worker_group
*wg
= NULL
;
4902 if (conf
->worker_cnt_per_group
== 0) {
4903 handle_list
= &conf
->handle_list
;
4904 } else if (group
!= ANY_GROUP
) {
4905 handle_list
= &conf
->worker_groups
[group
].handle_list
;
4906 wg
= &conf
->worker_groups
[group
];
4909 for (i
= 0; i
< conf
->group_cnt
; i
++) {
4910 handle_list
= &conf
->worker_groups
[i
].handle_list
;
4911 wg
= &conf
->worker_groups
[i
];
4912 if (!list_empty(handle_list
))
4917 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4919 list_empty(handle_list
) ? "empty" : "busy",
4920 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4921 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4923 if (!list_empty(handle_list
)) {
4924 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
4926 if (list_empty(&conf
->hold_list
))
4927 conf
->bypass_count
= 0;
4928 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4929 if (conf
->hold_list
.next
== conf
->last_hold
)
4930 conf
->bypass_count
++;
4932 conf
->last_hold
= conf
->hold_list
.next
;
4933 conf
->bypass_count
-= conf
->bypass_threshold
;
4934 if (conf
->bypass_count
< 0)
4935 conf
->bypass_count
= 0;
4938 } else if (!list_empty(&conf
->hold_list
) &&
4939 ((conf
->bypass_threshold
&&
4940 conf
->bypass_count
> conf
->bypass_threshold
) ||
4941 atomic_read(&conf
->pending_full_writes
) == 0)) {
4943 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
4944 if (conf
->worker_cnt_per_group
== 0 ||
4945 group
== ANY_GROUP
||
4946 !cpu_online(tmp
->cpu
) ||
4947 cpu_to_group(tmp
->cpu
) == group
) {
4954 conf
->bypass_count
-= conf
->bypass_threshold
;
4955 if (conf
->bypass_count
< 0)
4956 conf
->bypass_count
= 0;
4968 list_del_init(&sh
->lru
);
4969 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
4973 struct raid5_plug_cb
{
4974 struct blk_plug_cb cb
;
4975 struct list_head list
;
4976 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
4979 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4981 struct raid5_plug_cb
*cb
= container_of(
4982 blk_cb
, struct raid5_plug_cb
, cb
);
4983 struct stripe_head
*sh
;
4984 struct mddev
*mddev
= cb
->cb
.data
;
4985 struct r5conf
*conf
= mddev
->private;
4989 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4990 spin_lock_irq(&conf
->device_lock
);
4991 while (!list_empty(&cb
->list
)) {
4992 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4993 list_del_init(&sh
->lru
);
4995 * avoid race release_stripe_plug() sees
4996 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4997 * is still in our list
4999 smp_mb__before_atomic();
5000 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
5002 * STRIPE_ON_RELEASE_LIST could be set here. In that
5003 * case, the count is always > 1 here
5005 hash
= sh
->hash_lock_index
;
5006 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
5009 spin_unlock_irq(&conf
->device_lock
);
5011 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
5012 NR_STRIPE_HASH_LOCKS
);
5014 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5018 static void release_stripe_plug(struct mddev
*mddev
,
5019 struct stripe_head
*sh
)
5021 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5022 raid5_unplug
, mddev
,
5023 sizeof(struct raid5_plug_cb
));
5024 struct raid5_plug_cb
*cb
;
5027 raid5_release_stripe(sh
);
5031 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5033 if (cb
->list
.next
== NULL
) {
5035 INIT_LIST_HEAD(&cb
->list
);
5036 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5037 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5040 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5041 list_add_tail(&sh
->lru
, &cb
->list
);
5043 raid5_release_stripe(sh
);
5046 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5048 struct r5conf
*conf
= mddev
->private;
5049 sector_t logical_sector
, last_sector
;
5050 struct stripe_head
*sh
;
5054 if (mddev
->reshape_position
!= MaxSector
)
5055 /* Skip discard while reshape is happening */
5058 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5059 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5062 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5064 stripe_sectors
= conf
->chunk_sectors
*
5065 (conf
->raid_disks
- conf
->max_degraded
);
5066 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5068 sector_div(last_sector
, stripe_sectors
);
5070 logical_sector
*= conf
->chunk_sectors
;
5071 last_sector
*= conf
->chunk_sectors
;
5073 for (; logical_sector
< last_sector
;
5074 logical_sector
+= STRIPE_SECTORS
) {
5078 sh
= raid5_get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5079 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5080 TASK_UNINTERRUPTIBLE
);
5081 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5082 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5083 raid5_release_stripe(sh
);
5087 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5088 spin_lock_irq(&sh
->stripe_lock
);
5089 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5090 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5092 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5093 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5094 spin_unlock_irq(&sh
->stripe_lock
);
5095 raid5_release_stripe(sh
);
5100 set_bit(STRIPE_DISCARD
, &sh
->state
);
5101 finish_wait(&conf
->wait_for_overlap
, &w
);
5102 sh
->overwrite_disks
= 0;
5103 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5104 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5106 sh
->dev
[d
].towrite
= bi
;
5107 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5108 raid5_inc_bi_active_stripes(bi
);
5109 sh
->overwrite_disks
++;
5111 spin_unlock_irq(&sh
->stripe_lock
);
5112 if (conf
->mddev
->bitmap
) {
5114 d
< conf
->raid_disks
- conf
->max_degraded
;
5116 bitmap_startwrite(mddev
->bitmap
,
5120 sh
->bm_seq
= conf
->seq_flush
+ 1;
5121 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5124 set_bit(STRIPE_HANDLE
, &sh
->state
);
5125 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5126 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5127 atomic_inc(&conf
->preread_active_stripes
);
5128 release_stripe_plug(mddev
, sh
);
5131 remaining
= raid5_dec_bi_active_stripes(bi
);
5132 if (remaining
== 0) {
5133 md_write_end(mddev
);
5138 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
5140 struct r5conf
*conf
= mddev
->private;
5142 sector_t new_sector
;
5143 sector_t logical_sector
, last_sector
;
5144 struct stripe_head
*sh
;
5145 const int rw
= bio_data_dir(bi
);
5150 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
5151 int ret
= r5l_handle_flush_request(conf
->log
, bi
);
5155 if (ret
== -ENODEV
) {
5156 md_flush_request(mddev
, bi
);
5159 /* ret == -EAGAIN, fallback */
5162 md_write_start(mddev
, bi
);
5165 * If array is degraded, better not do chunk aligned read because
5166 * later we might have to read it again in order to reconstruct
5167 * data on failed drives.
5169 if (rw
== READ
&& mddev
->degraded
== 0 &&
5170 mddev
->reshape_position
== MaxSector
) {
5171 bi
= chunk_aligned_read(mddev
, bi
);
5176 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
5177 make_discard_request(mddev
, bi
);
5181 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5182 last_sector
= bio_end_sector(bi
);
5184 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5186 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5187 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5193 seq
= read_seqcount_begin(&conf
->gen_lock
);
5196 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5197 TASK_UNINTERRUPTIBLE
);
5198 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5199 /* spinlock is needed as reshape_progress may be
5200 * 64bit on a 32bit platform, and so it might be
5201 * possible to see a half-updated value
5202 * Of course reshape_progress could change after
5203 * the lock is dropped, so once we get a reference
5204 * to the stripe that we think it is, we will have
5207 spin_lock_irq(&conf
->device_lock
);
5208 if (mddev
->reshape_backwards
5209 ? logical_sector
< conf
->reshape_progress
5210 : logical_sector
>= conf
->reshape_progress
) {
5213 if (mddev
->reshape_backwards
5214 ? logical_sector
< conf
->reshape_safe
5215 : logical_sector
>= conf
->reshape_safe
) {
5216 spin_unlock_irq(&conf
->device_lock
);
5222 spin_unlock_irq(&conf
->device_lock
);
5225 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5228 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5229 (unsigned long long)new_sector
,
5230 (unsigned long long)logical_sector
);
5232 sh
= raid5_get_active_stripe(conf
, new_sector
, previous
,
5233 (bi
->bi_rw
&RWA_MASK
), 0);
5235 if (unlikely(previous
)) {
5236 /* expansion might have moved on while waiting for a
5237 * stripe, so we must do the range check again.
5238 * Expansion could still move past after this
5239 * test, but as we are holding a reference to
5240 * 'sh', we know that if that happens,
5241 * STRIPE_EXPANDING will get set and the expansion
5242 * won't proceed until we finish with the stripe.
5245 spin_lock_irq(&conf
->device_lock
);
5246 if (mddev
->reshape_backwards
5247 ? logical_sector
>= conf
->reshape_progress
5248 : logical_sector
< conf
->reshape_progress
)
5249 /* mismatch, need to try again */
5251 spin_unlock_irq(&conf
->device_lock
);
5253 raid5_release_stripe(sh
);
5259 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5260 /* Might have got the wrong stripe_head
5263 raid5_release_stripe(sh
);
5268 logical_sector
>= mddev
->suspend_lo
&&
5269 logical_sector
< mddev
->suspend_hi
) {
5270 raid5_release_stripe(sh
);
5271 /* As the suspend_* range is controlled by
5272 * userspace, we want an interruptible
5275 flush_signals(current
);
5276 prepare_to_wait(&conf
->wait_for_overlap
,
5277 &w
, TASK_INTERRUPTIBLE
);
5278 if (logical_sector
>= mddev
->suspend_lo
&&
5279 logical_sector
< mddev
->suspend_hi
) {
5286 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5287 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5288 /* Stripe is busy expanding or
5289 * add failed due to overlap. Flush everything
5292 md_wakeup_thread(mddev
->thread
);
5293 raid5_release_stripe(sh
);
5298 set_bit(STRIPE_HANDLE
, &sh
->state
);
5299 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5300 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5301 (bi
->bi_rw
& REQ_SYNC
) &&
5302 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5303 atomic_inc(&conf
->preread_active_stripes
);
5304 release_stripe_plug(mddev
, sh
);
5306 /* cannot get stripe for read-ahead, just give-up */
5307 bi
->bi_error
= -EIO
;
5311 finish_wait(&conf
->wait_for_overlap
, &w
);
5313 remaining
= raid5_dec_bi_active_stripes(bi
);
5314 if (remaining
== 0) {
5317 md_write_end(mddev
);
5319 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
5325 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5327 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5329 /* reshaping is quite different to recovery/resync so it is
5330 * handled quite separately ... here.
5332 * On each call to sync_request, we gather one chunk worth of
5333 * destination stripes and flag them as expanding.
5334 * Then we find all the source stripes and request reads.
5335 * As the reads complete, handle_stripe will copy the data
5336 * into the destination stripe and release that stripe.
5338 struct r5conf
*conf
= mddev
->private;
5339 struct stripe_head
*sh
;
5340 sector_t first_sector
, last_sector
;
5341 int raid_disks
= conf
->previous_raid_disks
;
5342 int data_disks
= raid_disks
- conf
->max_degraded
;
5343 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5346 sector_t writepos
, readpos
, safepos
;
5347 sector_t stripe_addr
;
5348 int reshape_sectors
;
5349 struct list_head stripes
;
5352 if (sector_nr
== 0) {
5353 /* If restarting in the middle, skip the initial sectors */
5354 if (mddev
->reshape_backwards
&&
5355 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5356 sector_nr
= raid5_size(mddev
, 0, 0)
5357 - conf
->reshape_progress
;
5358 } else if (mddev
->reshape_backwards
&&
5359 conf
->reshape_progress
== MaxSector
) {
5360 /* shouldn't happen, but just in case, finish up.*/
5361 sector_nr
= MaxSector
;
5362 } else if (!mddev
->reshape_backwards
&&
5363 conf
->reshape_progress
> 0)
5364 sector_nr
= conf
->reshape_progress
;
5365 sector_div(sector_nr
, new_data_disks
);
5367 mddev
->curr_resync_completed
= sector_nr
;
5368 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5375 /* We need to process a full chunk at a time.
5376 * If old and new chunk sizes differ, we need to process the
5380 reshape_sectors
= max(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5382 /* We update the metadata at least every 10 seconds, or when
5383 * the data about to be copied would over-write the source of
5384 * the data at the front of the range. i.e. one new_stripe
5385 * along from reshape_progress new_maps to after where
5386 * reshape_safe old_maps to
5388 writepos
= conf
->reshape_progress
;
5389 sector_div(writepos
, new_data_disks
);
5390 readpos
= conf
->reshape_progress
;
5391 sector_div(readpos
, data_disks
);
5392 safepos
= conf
->reshape_safe
;
5393 sector_div(safepos
, data_disks
);
5394 if (mddev
->reshape_backwards
) {
5395 BUG_ON(writepos
< reshape_sectors
);
5396 writepos
-= reshape_sectors
;
5397 readpos
+= reshape_sectors
;
5398 safepos
+= reshape_sectors
;
5400 writepos
+= reshape_sectors
;
5401 /* readpos and safepos are worst-case calculations.
5402 * A negative number is overly pessimistic, and causes
5403 * obvious problems for unsigned storage. So clip to 0.
5405 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5406 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5409 /* Having calculated the 'writepos' possibly use it
5410 * to set 'stripe_addr' which is where we will write to.
5412 if (mddev
->reshape_backwards
) {
5413 BUG_ON(conf
->reshape_progress
== 0);
5414 stripe_addr
= writepos
;
5415 BUG_ON((mddev
->dev_sectors
&
5416 ~((sector_t
)reshape_sectors
- 1))
5417 - reshape_sectors
- stripe_addr
5420 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5421 stripe_addr
= sector_nr
;
5424 /* 'writepos' is the most advanced device address we might write.
5425 * 'readpos' is the least advanced device address we might read.
5426 * 'safepos' is the least address recorded in the metadata as having
5428 * If there is a min_offset_diff, these are adjusted either by
5429 * increasing the safepos/readpos if diff is negative, or
5430 * increasing writepos if diff is positive.
5431 * If 'readpos' is then behind 'writepos', there is no way that we can
5432 * ensure safety in the face of a crash - that must be done by userspace
5433 * making a backup of the data. So in that case there is no particular
5434 * rush to update metadata.
5435 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5436 * update the metadata to advance 'safepos' to match 'readpos' so that
5437 * we can be safe in the event of a crash.
5438 * So we insist on updating metadata if safepos is behind writepos and
5439 * readpos is beyond writepos.
5440 * In any case, update the metadata every 10 seconds.
5441 * Maybe that number should be configurable, but I'm not sure it is
5442 * worth it.... maybe it could be a multiple of safemode_delay???
5444 if (conf
->min_offset_diff
< 0) {
5445 safepos
+= -conf
->min_offset_diff
;
5446 readpos
+= -conf
->min_offset_diff
;
5448 writepos
+= conf
->min_offset_diff
;
5450 if ((mddev
->reshape_backwards
5451 ? (safepos
> writepos
&& readpos
< writepos
)
5452 : (safepos
< writepos
&& readpos
> writepos
)) ||
5453 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5454 /* Cannot proceed until we've updated the superblock... */
5455 wait_event(conf
->wait_for_overlap
,
5456 atomic_read(&conf
->reshape_stripes
)==0
5457 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5458 if (atomic_read(&conf
->reshape_stripes
) != 0)
5460 mddev
->reshape_position
= conf
->reshape_progress
;
5461 mddev
->curr_resync_completed
= sector_nr
;
5462 conf
->reshape_checkpoint
= jiffies
;
5463 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5464 md_wakeup_thread(mddev
->thread
);
5465 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
5466 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5467 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5469 spin_lock_irq(&conf
->device_lock
);
5470 conf
->reshape_safe
= mddev
->reshape_position
;
5471 spin_unlock_irq(&conf
->device_lock
);
5472 wake_up(&conf
->wait_for_overlap
);
5473 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5476 INIT_LIST_HEAD(&stripes
);
5477 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5479 int skipped_disk
= 0;
5480 sh
= raid5_get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5481 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5482 atomic_inc(&conf
->reshape_stripes
);
5483 /* If any of this stripe is beyond the end of the old
5484 * array, then we need to zero those blocks
5486 for (j
=sh
->disks
; j
--;) {
5488 if (j
== sh
->pd_idx
)
5490 if (conf
->level
== 6 &&
5493 s
= raid5_compute_blocknr(sh
, j
, 0);
5494 if (s
< raid5_size(mddev
, 0, 0)) {
5498 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5499 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5500 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5502 if (!skipped_disk
) {
5503 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5504 set_bit(STRIPE_HANDLE
, &sh
->state
);
5506 list_add(&sh
->lru
, &stripes
);
5508 spin_lock_irq(&conf
->device_lock
);
5509 if (mddev
->reshape_backwards
)
5510 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5512 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5513 spin_unlock_irq(&conf
->device_lock
);
5514 /* Ok, those stripe are ready. We can start scheduling
5515 * reads on the source stripes.
5516 * The source stripes are determined by mapping the first and last
5517 * block on the destination stripes.
5520 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5523 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5524 * new_data_disks
- 1),
5526 if (last_sector
>= mddev
->dev_sectors
)
5527 last_sector
= mddev
->dev_sectors
- 1;
5528 while (first_sector
<= last_sector
) {
5529 sh
= raid5_get_active_stripe(conf
, first_sector
, 1, 0, 1);
5530 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5531 set_bit(STRIPE_HANDLE
, &sh
->state
);
5532 raid5_release_stripe(sh
);
5533 first_sector
+= STRIPE_SECTORS
;
5535 /* Now that the sources are clearly marked, we can release
5536 * the destination stripes
5538 while (!list_empty(&stripes
)) {
5539 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5540 list_del_init(&sh
->lru
);
5541 raid5_release_stripe(sh
);
5543 /* If this takes us to the resync_max point where we have to pause,
5544 * then we need to write out the superblock.
5546 sector_nr
+= reshape_sectors
;
5547 retn
= reshape_sectors
;
5549 if (mddev
->curr_resync_completed
> mddev
->resync_max
||
5550 (sector_nr
- mddev
->curr_resync_completed
) * 2
5551 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5552 /* Cannot proceed until we've updated the superblock... */
5553 wait_event(conf
->wait_for_overlap
,
5554 atomic_read(&conf
->reshape_stripes
) == 0
5555 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5556 if (atomic_read(&conf
->reshape_stripes
) != 0)
5558 mddev
->reshape_position
= conf
->reshape_progress
;
5559 mddev
->curr_resync_completed
= sector_nr
;
5560 conf
->reshape_checkpoint
= jiffies
;
5561 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5562 md_wakeup_thread(mddev
->thread
);
5563 wait_event(mddev
->sb_wait
,
5564 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
5565 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5566 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5568 spin_lock_irq(&conf
->device_lock
);
5569 conf
->reshape_safe
= mddev
->reshape_position
;
5570 spin_unlock_irq(&conf
->device_lock
);
5571 wake_up(&conf
->wait_for_overlap
);
5572 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5578 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5580 struct r5conf
*conf
= mddev
->private;
5581 struct stripe_head
*sh
;
5582 sector_t max_sector
= mddev
->dev_sectors
;
5583 sector_t sync_blocks
;
5584 int still_degraded
= 0;
5587 if (sector_nr
>= max_sector
) {
5588 /* just being told to finish up .. nothing much to do */
5590 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
5595 if (mddev
->curr_resync
< max_sector
) /* aborted */
5596 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
5598 else /* completed sync */
5600 bitmap_close_sync(mddev
->bitmap
);
5605 /* Allow raid5_quiesce to complete */
5606 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
5608 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
5609 return reshape_request(mddev
, sector_nr
, skipped
);
5611 /* No need to check resync_max as we never do more than one
5612 * stripe, and as resync_max will always be on a chunk boundary,
5613 * if the check in md_do_sync didn't fire, there is no chance
5614 * of overstepping resync_max here
5617 /* if there is too many failed drives and we are trying
5618 * to resync, then assert that we are finished, because there is
5619 * nothing we can do.
5621 if (mddev
->degraded
>= conf
->max_degraded
&&
5622 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
5623 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
5627 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
5629 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
5630 sync_blocks
>= STRIPE_SECTORS
) {
5631 /* we can skip this block, and probably more */
5632 sync_blocks
/= STRIPE_SECTORS
;
5634 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
5637 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, false);
5639 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 1, 0);
5641 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 0, 0);
5642 /* make sure we don't swamp the stripe cache if someone else
5643 * is trying to get access
5645 schedule_timeout_uninterruptible(1);
5647 /* Need to check if array will still be degraded after recovery/resync
5648 * Note in case of > 1 drive failures it's possible we're rebuilding
5649 * one drive while leaving another faulty drive in array.
5652 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5653 struct md_rdev
*rdev
= ACCESS_ONCE(conf
->disks
[i
].rdev
);
5655 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
5660 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
5662 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
5663 set_bit(STRIPE_HANDLE
, &sh
->state
);
5665 raid5_release_stripe(sh
);
5667 return STRIPE_SECTORS
;
5670 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
5672 /* We may not be able to submit a whole bio at once as there
5673 * may not be enough stripe_heads available.
5674 * We cannot pre-allocate enough stripe_heads as we may need
5675 * more than exist in the cache (if we allow ever large chunks).
5676 * So we do one stripe head at a time and record in
5677 * ->bi_hw_segments how many have been done.
5679 * We *know* that this entire raid_bio is in one chunk, so
5680 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5682 struct stripe_head
*sh
;
5684 sector_t sector
, logical_sector
, last_sector
;
5689 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
5690 ~((sector_t
)STRIPE_SECTORS
-1);
5691 sector
= raid5_compute_sector(conf
, logical_sector
,
5693 last_sector
= bio_end_sector(raid_bio
);
5695 for (; logical_sector
< last_sector
;
5696 logical_sector
+= STRIPE_SECTORS
,
5697 sector
+= STRIPE_SECTORS
,
5700 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
5701 /* already done this stripe */
5704 sh
= raid5_get_active_stripe(conf
, sector
, 0, 1, 1);
5707 /* failed to get a stripe - must wait */
5708 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5709 conf
->retry_read_aligned
= raid_bio
;
5713 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
5714 raid5_release_stripe(sh
);
5715 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5716 conf
->retry_read_aligned
= raid_bio
;
5720 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
5722 raid5_release_stripe(sh
);
5725 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
5726 if (remaining
== 0) {
5727 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
5729 bio_endio(raid_bio
);
5731 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5732 wake_up(&conf
->wait_for_quiescent
);
5736 static int handle_active_stripes(struct r5conf
*conf
, int group
,
5737 struct r5worker
*worker
,
5738 struct list_head
*temp_inactive_list
)
5740 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
5741 int i
, batch_size
= 0, hash
;
5742 bool release_inactive
= false;
5744 while (batch_size
< MAX_STRIPE_BATCH
&&
5745 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
5746 batch
[batch_size
++] = sh
;
5748 if (batch_size
== 0) {
5749 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5750 if (!list_empty(temp_inactive_list
+ i
))
5752 if (i
== NR_STRIPE_HASH_LOCKS
) {
5753 spin_unlock_irq(&conf
->device_lock
);
5754 r5l_flush_stripe_to_raid(conf
->log
);
5755 spin_lock_irq(&conf
->device_lock
);
5758 release_inactive
= true;
5760 spin_unlock_irq(&conf
->device_lock
);
5762 release_inactive_stripe_list(conf
, temp_inactive_list
,
5763 NR_STRIPE_HASH_LOCKS
);
5765 r5l_flush_stripe_to_raid(conf
->log
);
5766 if (release_inactive
) {
5767 spin_lock_irq(&conf
->device_lock
);
5771 for (i
= 0; i
< batch_size
; i
++)
5772 handle_stripe(batch
[i
]);
5773 r5l_write_stripe_run(conf
->log
);
5777 spin_lock_irq(&conf
->device_lock
);
5778 for (i
= 0; i
< batch_size
; i
++) {
5779 hash
= batch
[i
]->hash_lock_index
;
5780 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
5785 static void raid5_do_work(struct work_struct
*work
)
5787 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
5788 struct r5worker_group
*group
= worker
->group
;
5789 struct r5conf
*conf
= group
->conf
;
5790 int group_id
= group
- conf
->worker_groups
;
5792 struct blk_plug plug
;
5794 pr_debug("+++ raid5worker active\n");
5796 blk_start_plug(&plug
);
5798 spin_lock_irq(&conf
->device_lock
);
5800 int batch_size
, released
;
5802 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
5804 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
5805 worker
->temp_inactive_list
);
5806 worker
->working
= false;
5807 if (!batch_size
&& !released
)
5809 handled
+= batch_size
;
5811 pr_debug("%d stripes handled\n", handled
);
5813 spin_unlock_irq(&conf
->device_lock
);
5814 blk_finish_plug(&plug
);
5816 pr_debug("--- raid5worker inactive\n");
5820 * This is our raid5 kernel thread.
5822 * We scan the hash table for stripes which can be handled now.
5823 * During the scan, completed stripes are saved for us by the interrupt
5824 * handler, so that they will not have to wait for our next wakeup.
5826 static void raid5d(struct md_thread
*thread
)
5828 struct mddev
*mddev
= thread
->mddev
;
5829 struct r5conf
*conf
= mddev
->private;
5831 struct blk_plug plug
;
5833 pr_debug("+++ raid5d active\n");
5835 md_check_recovery(mddev
);
5837 if (!bio_list_empty(&conf
->return_bi
) &&
5838 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
5839 struct bio_list tmp
= BIO_EMPTY_LIST
;
5840 spin_lock_irq(&conf
->device_lock
);
5841 if (!test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
5842 bio_list_merge(&tmp
, &conf
->return_bi
);
5843 bio_list_init(&conf
->return_bi
);
5845 spin_unlock_irq(&conf
->device_lock
);
5849 blk_start_plug(&plug
);
5851 spin_lock_irq(&conf
->device_lock
);
5854 int batch_size
, released
;
5856 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
5858 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5861 !list_empty(&conf
->bitmap_list
)) {
5862 /* Now is a good time to flush some bitmap updates */
5864 spin_unlock_irq(&conf
->device_lock
);
5865 bitmap_unplug(mddev
->bitmap
);
5866 spin_lock_irq(&conf
->device_lock
);
5867 conf
->seq_write
= conf
->seq_flush
;
5868 activate_bit_delay(conf
, conf
->temp_inactive_list
);
5870 raid5_activate_delayed(conf
);
5872 while ((bio
= remove_bio_from_retry(conf
))) {
5874 spin_unlock_irq(&conf
->device_lock
);
5875 ok
= retry_aligned_read(conf
, bio
);
5876 spin_lock_irq(&conf
->device_lock
);
5882 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
5883 conf
->temp_inactive_list
);
5884 if (!batch_size
&& !released
)
5886 handled
+= batch_size
;
5888 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
5889 spin_unlock_irq(&conf
->device_lock
);
5890 md_check_recovery(mddev
);
5891 spin_lock_irq(&conf
->device_lock
);
5894 pr_debug("%d stripes handled\n", handled
);
5896 spin_unlock_irq(&conf
->device_lock
);
5897 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
) &&
5898 mutex_trylock(&conf
->cache_size_mutex
)) {
5899 grow_one_stripe(conf
, __GFP_NOWARN
);
5900 /* Set flag even if allocation failed. This helps
5901 * slow down allocation requests when mem is short
5903 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5904 mutex_unlock(&conf
->cache_size_mutex
);
5907 r5l_flush_stripe_to_raid(conf
->log
);
5909 async_tx_issue_pending_all();
5910 blk_finish_plug(&plug
);
5912 pr_debug("--- raid5d inactive\n");
5916 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
5918 struct r5conf
*conf
;
5920 spin_lock(&mddev
->lock
);
5921 conf
= mddev
->private;
5923 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
5924 spin_unlock(&mddev
->lock
);
5929 raid5_set_cache_size(struct mddev
*mddev
, int size
)
5931 struct r5conf
*conf
= mddev
->private;
5934 if (size
<= 16 || size
> 32768)
5937 conf
->min_nr_stripes
= size
;
5938 mutex_lock(&conf
->cache_size_mutex
);
5939 while (size
< conf
->max_nr_stripes
&&
5940 drop_one_stripe(conf
))
5942 mutex_unlock(&conf
->cache_size_mutex
);
5945 err
= md_allow_write(mddev
);
5949 mutex_lock(&conf
->cache_size_mutex
);
5950 while (size
> conf
->max_nr_stripes
)
5951 if (!grow_one_stripe(conf
, GFP_KERNEL
))
5953 mutex_unlock(&conf
->cache_size_mutex
);
5957 EXPORT_SYMBOL(raid5_set_cache_size
);
5960 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
5962 struct r5conf
*conf
;
5966 if (len
>= PAGE_SIZE
)
5968 if (kstrtoul(page
, 10, &new))
5970 err
= mddev_lock(mddev
);
5973 conf
= mddev
->private;
5977 err
= raid5_set_cache_size(mddev
, new);
5978 mddev_unlock(mddev
);
5983 static struct md_sysfs_entry
5984 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
5985 raid5_show_stripe_cache_size
,
5986 raid5_store_stripe_cache_size
);
5989 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
5991 struct r5conf
*conf
= mddev
->private;
5993 return sprintf(page
, "%d\n", conf
->rmw_level
);
5999 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
6001 struct r5conf
*conf
= mddev
->private;
6007 if (len
>= PAGE_SIZE
)
6010 if (kstrtoul(page
, 10, &new))
6013 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
6016 if (new != PARITY_DISABLE_RMW
&&
6017 new != PARITY_ENABLE_RMW
&&
6018 new != PARITY_PREFER_RMW
)
6021 conf
->rmw_level
= new;
6025 static struct md_sysfs_entry
6026 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
6027 raid5_show_rmw_level
,
6028 raid5_store_rmw_level
);
6032 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
6034 struct r5conf
*conf
;
6036 spin_lock(&mddev
->lock
);
6037 conf
= mddev
->private;
6039 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
6040 spin_unlock(&mddev
->lock
);
6045 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
6047 struct r5conf
*conf
;
6051 if (len
>= PAGE_SIZE
)
6053 if (kstrtoul(page
, 10, &new))
6056 err
= mddev_lock(mddev
);
6059 conf
= mddev
->private;
6062 else if (new > conf
->min_nr_stripes
)
6065 conf
->bypass_threshold
= new;
6066 mddev_unlock(mddev
);
6070 static struct md_sysfs_entry
6071 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
6073 raid5_show_preread_threshold
,
6074 raid5_store_preread_threshold
);
6077 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
6079 struct r5conf
*conf
;
6081 spin_lock(&mddev
->lock
);
6082 conf
= mddev
->private;
6084 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6085 spin_unlock(&mddev
->lock
);
6090 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6092 struct r5conf
*conf
;
6096 if (len
>= PAGE_SIZE
)
6098 if (kstrtoul(page
, 10, &new))
6102 err
= mddev_lock(mddev
);
6105 conf
= mddev
->private;
6108 else if (new != conf
->skip_copy
) {
6109 mddev_suspend(mddev
);
6110 conf
->skip_copy
= new;
6112 mddev
->queue
->backing_dev_info
.capabilities
|=
6113 BDI_CAP_STABLE_WRITES
;
6115 mddev
->queue
->backing_dev_info
.capabilities
&=
6116 ~BDI_CAP_STABLE_WRITES
;
6117 mddev_resume(mddev
);
6119 mddev_unlock(mddev
);
6123 static struct md_sysfs_entry
6124 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6125 raid5_show_skip_copy
,
6126 raid5_store_skip_copy
);
6129 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6131 struct r5conf
*conf
= mddev
->private;
6133 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6138 static struct md_sysfs_entry
6139 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6142 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6144 struct r5conf
*conf
;
6146 spin_lock(&mddev
->lock
);
6147 conf
= mddev
->private;
6149 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6150 spin_unlock(&mddev
->lock
);
6154 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6156 int *worker_cnt_per_group
,
6157 struct r5worker_group
**worker_groups
);
6159 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6161 struct r5conf
*conf
;
6164 struct r5worker_group
*new_groups
, *old_groups
;
6165 int group_cnt
, worker_cnt_per_group
;
6167 if (len
>= PAGE_SIZE
)
6169 if (kstrtoul(page
, 10, &new))
6172 err
= mddev_lock(mddev
);
6175 conf
= mddev
->private;
6178 else if (new != conf
->worker_cnt_per_group
) {
6179 mddev_suspend(mddev
);
6181 old_groups
= conf
->worker_groups
;
6183 flush_workqueue(raid5_wq
);
6185 err
= alloc_thread_groups(conf
, new,
6186 &group_cnt
, &worker_cnt_per_group
,
6189 spin_lock_irq(&conf
->device_lock
);
6190 conf
->group_cnt
= group_cnt
;
6191 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6192 conf
->worker_groups
= new_groups
;
6193 spin_unlock_irq(&conf
->device_lock
);
6196 kfree(old_groups
[0].workers
);
6199 mddev_resume(mddev
);
6201 mddev_unlock(mddev
);
6206 static struct md_sysfs_entry
6207 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6208 raid5_show_group_thread_cnt
,
6209 raid5_store_group_thread_cnt
);
6211 static struct attribute
*raid5_attrs
[] = {
6212 &raid5_stripecache_size
.attr
,
6213 &raid5_stripecache_active
.attr
,
6214 &raid5_preread_bypass_threshold
.attr
,
6215 &raid5_group_thread_cnt
.attr
,
6216 &raid5_skip_copy
.attr
,
6217 &raid5_rmw_level
.attr
,
6220 static struct attribute_group raid5_attrs_group
= {
6222 .attrs
= raid5_attrs
,
6225 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6227 int *worker_cnt_per_group
,
6228 struct r5worker_group
**worker_groups
)
6232 struct r5worker
*workers
;
6234 *worker_cnt_per_group
= cnt
;
6237 *worker_groups
= NULL
;
6240 *group_cnt
= num_possible_nodes();
6241 size
= sizeof(struct r5worker
) * cnt
;
6242 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6243 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6244 *group_cnt
, GFP_NOIO
);
6245 if (!*worker_groups
|| !workers
) {
6247 kfree(*worker_groups
);
6251 for (i
= 0; i
< *group_cnt
; i
++) {
6252 struct r5worker_group
*group
;
6254 group
= &(*worker_groups
)[i
];
6255 INIT_LIST_HEAD(&group
->handle_list
);
6257 group
->workers
= workers
+ i
* cnt
;
6259 for (j
= 0; j
< cnt
; j
++) {
6260 struct r5worker
*worker
= group
->workers
+ j
;
6261 worker
->group
= group
;
6262 INIT_WORK(&worker
->work
, raid5_do_work
);
6264 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6265 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6272 static void free_thread_groups(struct r5conf
*conf
)
6274 if (conf
->worker_groups
)
6275 kfree(conf
->worker_groups
[0].workers
);
6276 kfree(conf
->worker_groups
);
6277 conf
->worker_groups
= NULL
;
6281 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6283 struct r5conf
*conf
= mddev
->private;
6286 sectors
= mddev
->dev_sectors
;
6288 /* size is defined by the smallest of previous and new size */
6289 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6291 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
6292 sectors
&= ~((sector_t
)conf
->prev_chunk_sectors
- 1);
6293 return sectors
* (raid_disks
- conf
->max_degraded
);
6296 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6298 safe_put_page(percpu
->spare_page
);
6299 if (percpu
->scribble
)
6300 flex_array_free(percpu
->scribble
);
6301 percpu
->spare_page
= NULL
;
6302 percpu
->scribble
= NULL
;
6305 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6307 if (conf
->level
== 6 && !percpu
->spare_page
)
6308 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6309 if (!percpu
->scribble
)
6310 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6311 conf
->previous_raid_disks
),
6312 max(conf
->chunk_sectors
,
6313 conf
->prev_chunk_sectors
)
6317 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6318 free_scratch_buffer(conf
, percpu
);
6325 static void raid5_free_percpu(struct r5conf
*conf
)
6332 #ifdef CONFIG_HOTPLUG_CPU
6333 unregister_cpu_notifier(&conf
->cpu_notify
);
6337 for_each_possible_cpu(cpu
)
6338 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6341 free_percpu(conf
->percpu
);
6344 static void free_conf(struct r5conf
*conf
)
6347 r5l_exit_log(conf
->log
);
6348 if (conf
->shrinker
.seeks
)
6349 unregister_shrinker(&conf
->shrinker
);
6351 free_thread_groups(conf
);
6352 shrink_stripes(conf
);
6353 raid5_free_percpu(conf
);
6355 kfree(conf
->stripe_hashtbl
);
6359 #ifdef CONFIG_HOTPLUG_CPU
6360 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
6363 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
6364 long cpu
= (long)hcpu
;
6365 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6368 case CPU_UP_PREPARE
:
6369 case CPU_UP_PREPARE_FROZEN
:
6370 if (alloc_scratch_buffer(conf
, percpu
)) {
6371 pr_err("%s: failed memory allocation for cpu%ld\n",
6373 return notifier_from_errno(-ENOMEM
);
6377 case CPU_DEAD_FROZEN
:
6378 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6387 static int raid5_alloc_percpu(struct r5conf
*conf
)
6392 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6396 #ifdef CONFIG_HOTPLUG_CPU
6397 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
6398 conf
->cpu_notify
.priority
= 0;
6399 err
= register_cpu_notifier(&conf
->cpu_notify
);
6405 for_each_present_cpu(cpu
) {
6406 err
= alloc_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6408 pr_err("%s: failed memory allocation for cpu%ld\n",
6418 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6419 struct shrink_control
*sc
)
6421 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6422 unsigned long ret
= SHRINK_STOP
;
6424 if (mutex_trylock(&conf
->cache_size_mutex
)) {
6426 while (ret
< sc
->nr_to_scan
&&
6427 conf
->max_nr_stripes
> conf
->min_nr_stripes
) {
6428 if (drop_one_stripe(conf
) == 0) {
6434 mutex_unlock(&conf
->cache_size_mutex
);
6439 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6440 struct shrink_control
*sc
)
6442 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6444 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6445 /* unlikely, but not impossible */
6447 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6450 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6452 struct r5conf
*conf
;
6453 int raid_disk
, memory
, max_disks
;
6454 struct md_rdev
*rdev
;
6455 struct disk_info
*disk
;
6458 int group_cnt
, worker_cnt_per_group
;
6459 struct r5worker_group
*new_group
;
6461 if (mddev
->new_level
!= 5
6462 && mddev
->new_level
!= 4
6463 && mddev
->new_level
!= 6) {
6464 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6465 mdname(mddev
), mddev
->new_level
);
6466 return ERR_PTR(-EIO
);
6468 if ((mddev
->new_level
== 5
6469 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6470 (mddev
->new_level
== 6
6471 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6472 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
6473 mdname(mddev
), mddev
->new_layout
);
6474 return ERR_PTR(-EIO
);
6476 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6477 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6478 mdname(mddev
), mddev
->raid_disks
);
6479 return ERR_PTR(-EINVAL
);
6482 if (!mddev
->new_chunk_sectors
||
6483 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6484 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6485 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
6486 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6487 return ERR_PTR(-EINVAL
);
6490 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6493 /* Don't enable multi-threading by default*/
6494 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6496 conf
->group_cnt
= group_cnt
;
6497 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6498 conf
->worker_groups
= new_group
;
6501 spin_lock_init(&conf
->device_lock
);
6502 seqcount_init(&conf
->gen_lock
);
6503 mutex_init(&conf
->cache_size_mutex
);
6504 init_waitqueue_head(&conf
->wait_for_quiescent
);
6505 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++) {
6506 init_waitqueue_head(&conf
->wait_for_stripe
[i
]);
6508 init_waitqueue_head(&conf
->wait_for_overlap
);
6509 INIT_LIST_HEAD(&conf
->handle_list
);
6510 INIT_LIST_HEAD(&conf
->hold_list
);
6511 INIT_LIST_HEAD(&conf
->delayed_list
);
6512 INIT_LIST_HEAD(&conf
->bitmap_list
);
6513 bio_list_init(&conf
->return_bi
);
6514 init_llist_head(&conf
->released_stripes
);
6515 atomic_set(&conf
->active_stripes
, 0);
6516 atomic_set(&conf
->preread_active_stripes
, 0);
6517 atomic_set(&conf
->active_aligned_reads
, 0);
6518 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6519 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6521 conf
->raid_disks
= mddev
->raid_disks
;
6522 if (mddev
->reshape_position
== MaxSector
)
6523 conf
->previous_raid_disks
= mddev
->raid_disks
;
6525 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6526 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6528 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6533 conf
->mddev
= mddev
;
6535 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6538 /* We init hash_locks[0] separately to that it can be used
6539 * as the reference lock in the spin_lock_nest_lock() call
6540 * in lock_all_device_hash_locks_irq in order to convince
6541 * lockdep that we know what we are doing.
6543 spin_lock_init(conf
->hash_locks
);
6544 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6545 spin_lock_init(conf
->hash_locks
+ i
);
6547 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6548 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6550 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6551 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6553 conf
->level
= mddev
->new_level
;
6554 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6555 if (raid5_alloc_percpu(conf
) != 0)
6558 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6560 rdev_for_each(rdev
, mddev
) {
6561 raid_disk
= rdev
->raid_disk
;
6562 if (raid_disk
>= max_disks
6563 || raid_disk
< 0 || test_bit(Journal
, &rdev
->flags
))
6565 disk
= conf
->disks
+ raid_disk
;
6567 if (test_bit(Replacement
, &rdev
->flags
)) {
6568 if (disk
->replacement
)
6570 disk
->replacement
= rdev
;
6577 if (test_bit(In_sync
, &rdev
->flags
)) {
6578 char b
[BDEVNAME_SIZE
];
6579 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
6581 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
6582 } else if (rdev
->saved_raid_disk
!= raid_disk
)
6583 /* Cannot rely on bitmap to complete recovery */
6587 conf
->level
= mddev
->new_level
;
6588 if (conf
->level
== 6) {
6589 conf
->max_degraded
= 2;
6590 if (raid6_call
.xor_syndrome
)
6591 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6593 conf
->rmw_level
= PARITY_DISABLE_RMW
;
6595 conf
->max_degraded
= 1;
6596 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6598 conf
->algorithm
= mddev
->new_layout
;
6599 conf
->reshape_progress
= mddev
->reshape_position
;
6600 if (conf
->reshape_progress
!= MaxSector
) {
6601 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
6602 conf
->prev_algo
= mddev
->layout
;
6604 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
6605 conf
->prev_algo
= conf
->algorithm
;
6608 conf
->min_nr_stripes
= NR_STRIPES
;
6609 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
6610 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
6611 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
6612 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
6614 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6615 mdname(mddev
), memory
);
6618 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
6619 mdname(mddev
), memory
);
6621 * Losing a stripe head costs more than the time to refill it,
6622 * it reduces the queue depth and so can hurt throughput.
6623 * So set it rather large, scaled by number of devices.
6625 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
6626 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
6627 conf
->shrinker
.count_objects
= raid5_cache_count
;
6628 conf
->shrinker
.batch
= 128;
6629 conf
->shrinker
.flags
= 0;
6630 register_shrinker(&conf
->shrinker
);
6632 sprintf(pers_name
, "raid%d", mddev
->new_level
);
6633 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
6634 if (!conf
->thread
) {
6636 "md/raid:%s: couldn't allocate thread.\n",
6646 return ERR_PTR(-EIO
);
6648 return ERR_PTR(-ENOMEM
);
6651 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
6654 case ALGORITHM_PARITY_0
:
6655 if (raid_disk
< max_degraded
)
6658 case ALGORITHM_PARITY_N
:
6659 if (raid_disk
>= raid_disks
- max_degraded
)
6662 case ALGORITHM_PARITY_0_6
:
6663 if (raid_disk
== 0 ||
6664 raid_disk
== raid_disks
- 1)
6667 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6668 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6669 case ALGORITHM_LEFT_SYMMETRIC_6
:
6670 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6671 if (raid_disk
== raid_disks
- 1)
6677 static int run(struct mddev
*mddev
)
6679 struct r5conf
*conf
;
6680 int working_disks
= 0;
6681 int dirty_parity_disks
= 0;
6682 struct md_rdev
*rdev
;
6683 struct md_rdev
*journal_dev
= NULL
;
6684 sector_t reshape_offset
= 0;
6686 long long min_offset_diff
= 0;
6689 if (mddev
->recovery_cp
!= MaxSector
)
6690 printk(KERN_NOTICE
"md/raid:%s: not clean"
6691 " -- starting background reconstruction\n",
6694 rdev_for_each(rdev
, mddev
) {
6697 if (test_bit(Journal
, &rdev
->flags
)) {
6701 if (rdev
->raid_disk
< 0)
6703 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
6705 min_offset_diff
= diff
;
6707 } else if (mddev
->reshape_backwards
&&
6708 diff
< min_offset_diff
)
6709 min_offset_diff
= diff
;
6710 else if (!mddev
->reshape_backwards
&&
6711 diff
> min_offset_diff
)
6712 min_offset_diff
= diff
;
6715 if (mddev
->reshape_position
!= MaxSector
) {
6716 /* Check that we can continue the reshape.
6717 * Difficulties arise if the stripe we would write to
6718 * next is at or after the stripe we would read from next.
6719 * For a reshape that changes the number of devices, this
6720 * is only possible for a very short time, and mdadm makes
6721 * sure that time appears to have past before assembling
6722 * the array. So we fail if that time hasn't passed.
6723 * For a reshape that keeps the number of devices the same
6724 * mdadm must be monitoring the reshape can keeping the
6725 * critical areas read-only and backed up. It will start
6726 * the array in read-only mode, so we check for that.
6728 sector_t here_new
, here_old
;
6730 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
6735 printk(KERN_ERR
"md/raid:%s: don't support reshape with journal - aborting.\n",
6740 if (mddev
->new_level
!= mddev
->level
) {
6741 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
6742 "required - aborting.\n",
6746 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6747 /* reshape_position must be on a new-stripe boundary, and one
6748 * further up in new geometry must map after here in old
6750 * If the chunk sizes are different, then as we perform reshape
6751 * in units of the largest of the two, reshape_position needs
6752 * be a multiple of the largest chunk size times new data disks.
6754 here_new
= mddev
->reshape_position
;
6755 chunk_sectors
= max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
);
6756 new_data_disks
= mddev
->raid_disks
- max_degraded
;
6757 if (sector_div(here_new
, chunk_sectors
* new_data_disks
)) {
6758 printk(KERN_ERR
"md/raid:%s: reshape_position not "
6759 "on a stripe boundary\n", mdname(mddev
));
6762 reshape_offset
= here_new
* chunk_sectors
;
6763 /* here_new is the stripe we will write to */
6764 here_old
= mddev
->reshape_position
;
6765 sector_div(here_old
, chunk_sectors
* (old_disks
-max_degraded
));
6766 /* here_old is the first stripe that we might need to read
6768 if (mddev
->delta_disks
== 0) {
6769 /* We cannot be sure it is safe to start an in-place
6770 * reshape. It is only safe if user-space is monitoring
6771 * and taking constant backups.
6772 * mdadm always starts a situation like this in
6773 * readonly mode so it can take control before
6774 * allowing any writes. So just check for that.
6776 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
6777 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
6778 /* not really in-place - so OK */;
6779 else if (mddev
->ro
== 0) {
6780 printk(KERN_ERR
"md/raid:%s: in-place reshape "
6781 "must be started in read-only mode "
6786 } else if (mddev
->reshape_backwards
6787 ? (here_new
* chunk_sectors
+ min_offset_diff
<=
6788 here_old
* chunk_sectors
)
6789 : (here_new
* chunk_sectors
>=
6790 here_old
* chunk_sectors
+ (-min_offset_diff
))) {
6791 /* Reading from the same stripe as writing to - bad */
6792 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
6793 "auto-recovery - aborting.\n",
6797 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
6799 /* OK, we should be able to continue; */
6801 BUG_ON(mddev
->level
!= mddev
->new_level
);
6802 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
6803 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
6804 BUG_ON(mddev
->delta_disks
!= 0);
6807 if (mddev
->private == NULL
)
6808 conf
= setup_conf(mddev
);
6810 conf
= mddev
->private;
6813 return PTR_ERR(conf
);
6815 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) && !journal_dev
) {
6816 printk(KERN_ERR
"md/raid:%s: journal disk is missing, force array readonly\n",
6819 set_disk_ro(mddev
->gendisk
, 1);
6822 conf
->min_offset_diff
= min_offset_diff
;
6823 mddev
->thread
= conf
->thread
;
6824 conf
->thread
= NULL
;
6825 mddev
->private = conf
;
6827 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
6829 rdev
= conf
->disks
[i
].rdev
;
6830 if (!rdev
&& conf
->disks
[i
].replacement
) {
6831 /* The replacement is all we have yet */
6832 rdev
= conf
->disks
[i
].replacement
;
6833 conf
->disks
[i
].replacement
= NULL
;
6834 clear_bit(Replacement
, &rdev
->flags
);
6835 conf
->disks
[i
].rdev
= rdev
;
6839 if (conf
->disks
[i
].replacement
&&
6840 conf
->reshape_progress
!= MaxSector
) {
6841 /* replacements and reshape simply do not mix. */
6842 printk(KERN_ERR
"md: cannot handle concurrent "
6843 "replacement and reshape.\n");
6846 if (test_bit(In_sync
, &rdev
->flags
)) {
6850 /* This disc is not fully in-sync. However if it
6851 * just stored parity (beyond the recovery_offset),
6852 * when we don't need to be concerned about the
6853 * array being dirty.
6854 * When reshape goes 'backwards', we never have
6855 * partially completed devices, so we only need
6856 * to worry about reshape going forwards.
6858 /* Hack because v0.91 doesn't store recovery_offset properly. */
6859 if (mddev
->major_version
== 0 &&
6860 mddev
->minor_version
> 90)
6861 rdev
->recovery_offset
= reshape_offset
;
6863 if (rdev
->recovery_offset
< reshape_offset
) {
6864 /* We need to check old and new layout */
6865 if (!only_parity(rdev
->raid_disk
,
6868 conf
->max_degraded
))
6871 if (!only_parity(rdev
->raid_disk
,
6873 conf
->previous_raid_disks
,
6874 conf
->max_degraded
))
6876 dirty_parity_disks
++;
6880 * 0 for a fully functional array, 1 or 2 for a degraded array.
6882 mddev
->degraded
= calc_degraded(conf
);
6884 if (has_failed(conf
)) {
6885 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
6886 " (%d/%d failed)\n",
6887 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
6891 /* device size must be a multiple of chunk size */
6892 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
6893 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
6895 if (mddev
->degraded
> dirty_parity_disks
&&
6896 mddev
->recovery_cp
!= MaxSector
) {
6897 if (mddev
->ok_start_degraded
)
6899 "md/raid:%s: starting dirty degraded array"
6900 " - data corruption possible.\n",
6904 "md/raid:%s: cannot start dirty degraded array.\n",
6910 if (mddev
->degraded
== 0)
6911 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
6912 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
6913 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
6916 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
6917 " out of %d devices, algorithm %d\n",
6918 mdname(mddev
), conf
->level
,
6919 mddev
->raid_disks
- mddev
->degraded
,
6920 mddev
->raid_disks
, mddev
->new_layout
);
6922 print_raid5_conf(conf
);
6924 if (conf
->reshape_progress
!= MaxSector
) {
6925 conf
->reshape_safe
= conf
->reshape_progress
;
6926 atomic_set(&conf
->reshape_stripes
, 0);
6927 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6928 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6929 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6930 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6931 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6935 /* Ok, everything is just fine now */
6936 if (mddev
->to_remove
== &raid5_attrs_group
)
6937 mddev
->to_remove
= NULL
;
6938 else if (mddev
->kobj
.sd
&&
6939 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
6941 "raid5: failed to create sysfs attributes for %s\n",
6943 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6947 bool discard_supported
= true;
6948 /* read-ahead size must cover two whole stripes, which
6949 * is 2 * (datadisks) * chunksize where 'n' is the
6950 * number of raid devices
6952 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
6953 int stripe
= data_disks
*
6954 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
6955 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6956 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6958 chunk_size
= mddev
->chunk_sectors
<< 9;
6959 blk_queue_io_min(mddev
->queue
, chunk_size
);
6960 blk_queue_io_opt(mddev
->queue
, chunk_size
*
6961 (conf
->raid_disks
- conf
->max_degraded
));
6962 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
6964 * We can only discard a whole stripe. It doesn't make sense to
6965 * discard data disk but write parity disk
6967 stripe
= stripe
* PAGE_SIZE
;
6968 /* Round up to power of 2, as discard handling
6969 * currently assumes that */
6970 while ((stripe
-1) & stripe
)
6971 stripe
= (stripe
| (stripe
-1)) + 1;
6972 mddev
->queue
->limits
.discard_alignment
= stripe
;
6973 mddev
->queue
->limits
.discard_granularity
= stripe
;
6975 * unaligned part of discard request will be ignored, so can't
6976 * guarantee discard_zeroes_data
6978 mddev
->queue
->limits
.discard_zeroes_data
= 0;
6980 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
6982 rdev_for_each(rdev
, mddev
) {
6983 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6984 rdev
->data_offset
<< 9);
6985 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6986 rdev
->new_data_offset
<< 9);
6988 * discard_zeroes_data is required, otherwise data
6989 * could be lost. Consider a scenario: discard a stripe
6990 * (the stripe could be inconsistent if
6991 * discard_zeroes_data is 0); write one disk of the
6992 * stripe (the stripe could be inconsistent again
6993 * depending on which disks are used to calculate
6994 * parity); the disk is broken; The stripe data of this
6997 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
6998 !bdev_get_queue(rdev
->bdev
)->
6999 limits
.discard_zeroes_data
)
7000 discard_supported
= false;
7001 /* Unfortunately, discard_zeroes_data is not currently
7002 * a guarantee - just a hint. So we only allow DISCARD
7003 * if the sysadmin has confirmed that only safe devices
7004 * are in use by setting a module parameter.
7006 if (!devices_handle_discard_safely
) {
7007 if (discard_supported
) {
7008 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7009 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7011 discard_supported
= false;
7015 if (discard_supported
&&
7016 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
7017 mddev
->queue
->limits
.discard_granularity
>= stripe
)
7018 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
7021 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
7026 char b
[BDEVNAME_SIZE
];
7028 printk(KERN_INFO
"md/raid:%s: using device %s as journal\n",
7029 mdname(mddev
), bdevname(journal_dev
->bdev
, b
));
7030 r5l_init_log(conf
, journal_dev
);
7035 md_unregister_thread(&mddev
->thread
);
7036 print_raid5_conf(conf
);
7038 mddev
->private = NULL
;
7039 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
7043 static void raid5_free(struct mddev
*mddev
, void *priv
)
7045 struct r5conf
*conf
= priv
;
7048 mddev
->to_remove
= &raid5_attrs_group
;
7051 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
7053 struct r5conf
*conf
= mddev
->private;
7056 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
7057 conf
->chunk_sectors
/ 2, mddev
->layout
);
7058 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
7059 for (i
= 0; i
< conf
->raid_disks
; i
++)
7060 seq_printf (seq
, "%s",
7061 conf
->disks
[i
].rdev
&&
7062 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
7063 seq_printf (seq
, "]");
7066 static void print_raid5_conf (struct r5conf
*conf
)
7069 struct disk_info
*tmp
;
7071 printk(KERN_DEBUG
"RAID conf printout:\n");
7073 printk("(conf==NULL)\n");
7076 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
7078 conf
->raid_disks
- conf
->mddev
->degraded
);
7080 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7081 char b
[BDEVNAME_SIZE
];
7082 tmp
= conf
->disks
+ i
;
7084 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
7085 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
7086 bdevname(tmp
->rdev
->bdev
, b
));
7090 static int raid5_spare_active(struct mddev
*mddev
)
7093 struct r5conf
*conf
= mddev
->private;
7094 struct disk_info
*tmp
;
7096 unsigned long flags
;
7098 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7099 tmp
= conf
->disks
+ i
;
7100 if (tmp
->replacement
7101 && tmp
->replacement
->recovery_offset
== MaxSector
7102 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
7103 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
7104 /* Replacement has just become active. */
7106 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
7109 /* Replaced device not technically faulty,
7110 * but we need to be sure it gets removed
7111 * and never re-added.
7113 set_bit(Faulty
, &tmp
->rdev
->flags
);
7114 sysfs_notify_dirent_safe(
7115 tmp
->rdev
->sysfs_state
);
7117 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
7118 } else if (tmp
->rdev
7119 && tmp
->rdev
->recovery_offset
== MaxSector
7120 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
7121 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
7123 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
7126 spin_lock_irqsave(&conf
->device_lock
, flags
);
7127 mddev
->degraded
= calc_degraded(conf
);
7128 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7129 print_raid5_conf(conf
);
7133 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7135 struct r5conf
*conf
= mddev
->private;
7137 int number
= rdev
->raid_disk
;
7138 struct md_rdev
**rdevp
;
7139 struct disk_info
*p
= conf
->disks
+ number
;
7141 print_raid5_conf(conf
);
7142 if (test_bit(Journal
, &rdev
->flags
) && conf
->log
) {
7143 struct r5l_log
*log
;
7145 * we can't wait pending write here, as this is called in
7146 * raid5d, wait will deadlock.
7148 if (atomic_read(&mddev
->writes_pending
))
7156 if (rdev
== p
->rdev
)
7158 else if (rdev
== p
->replacement
)
7159 rdevp
= &p
->replacement
;
7163 if (number
>= conf
->raid_disks
&&
7164 conf
->reshape_progress
== MaxSector
)
7165 clear_bit(In_sync
, &rdev
->flags
);
7167 if (test_bit(In_sync
, &rdev
->flags
) ||
7168 atomic_read(&rdev
->nr_pending
)) {
7172 /* Only remove non-faulty devices if recovery
7175 if (!test_bit(Faulty
, &rdev
->flags
) &&
7176 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7177 !has_failed(conf
) &&
7178 (!p
->replacement
|| p
->replacement
== rdev
) &&
7179 number
< conf
->raid_disks
) {
7185 if (atomic_read(&rdev
->nr_pending
)) {
7186 /* lost the race, try later */
7189 } else if (p
->replacement
) {
7190 /* We must have just cleared 'rdev' */
7191 p
->rdev
= p
->replacement
;
7192 clear_bit(Replacement
, &p
->replacement
->flags
);
7193 smp_mb(); /* Make sure other CPUs may see both as identical
7194 * but will never see neither - if they are careful
7196 p
->replacement
= NULL
;
7197 clear_bit(WantReplacement
, &rdev
->flags
);
7199 /* We might have just removed the Replacement as faulty-
7200 * clear the bit just in case
7202 clear_bit(WantReplacement
, &rdev
->flags
);
7205 print_raid5_conf(conf
);
7209 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7211 struct r5conf
*conf
= mddev
->private;
7214 struct disk_info
*p
;
7216 int last
= conf
->raid_disks
- 1;
7218 if (test_bit(Journal
, &rdev
->flags
)) {
7219 char b
[BDEVNAME_SIZE
];
7223 rdev
->raid_disk
= 0;
7225 * The array is in readonly mode if journal is missing, so no
7226 * write requests running. We should be safe
7228 r5l_init_log(conf
, rdev
);
7229 printk(KERN_INFO
"md/raid:%s: using device %s as journal\n",
7230 mdname(mddev
), bdevname(rdev
->bdev
, b
));
7233 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7236 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7237 /* no point adding a device */
7240 if (rdev
->raid_disk
>= 0)
7241 first
= last
= rdev
->raid_disk
;
7244 * find the disk ... but prefer rdev->saved_raid_disk
7247 if (rdev
->saved_raid_disk
>= 0 &&
7248 rdev
->saved_raid_disk
>= first
&&
7249 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7250 first
= rdev
->saved_raid_disk
;
7252 for (disk
= first
; disk
<= last
; disk
++) {
7253 p
= conf
->disks
+ disk
;
7254 if (p
->rdev
== NULL
) {
7255 clear_bit(In_sync
, &rdev
->flags
);
7256 rdev
->raid_disk
= disk
;
7258 if (rdev
->saved_raid_disk
!= disk
)
7260 rcu_assign_pointer(p
->rdev
, rdev
);
7264 for (disk
= first
; disk
<= last
; disk
++) {
7265 p
= conf
->disks
+ disk
;
7266 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7267 p
->replacement
== NULL
) {
7268 clear_bit(In_sync
, &rdev
->flags
);
7269 set_bit(Replacement
, &rdev
->flags
);
7270 rdev
->raid_disk
= disk
;
7273 rcu_assign_pointer(p
->replacement
, rdev
);
7278 print_raid5_conf(conf
);
7282 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7284 /* no resync is happening, and there is enough space
7285 * on all devices, so we can resize.
7286 * We need to make sure resync covers any new space.
7287 * If the array is shrinking we should possibly wait until
7288 * any io in the removed space completes, but it hardly seems
7292 struct r5conf
*conf
= mddev
->private;
7296 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
7297 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7298 if (mddev
->external_size
&&
7299 mddev
->array_sectors
> newsize
)
7301 if (mddev
->bitmap
) {
7302 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7306 md_set_array_sectors(mddev
, newsize
);
7307 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7308 revalidate_disk(mddev
->gendisk
);
7309 if (sectors
> mddev
->dev_sectors
&&
7310 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7311 mddev
->recovery_cp
= mddev
->dev_sectors
;
7312 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7314 mddev
->dev_sectors
= sectors
;
7315 mddev
->resync_max_sectors
= sectors
;
7319 static int check_stripe_cache(struct mddev
*mddev
)
7321 /* Can only proceed if there are plenty of stripe_heads.
7322 * We need a minimum of one full stripe,, and for sensible progress
7323 * it is best to have about 4 times that.
7324 * If we require 4 times, then the default 256 4K stripe_heads will
7325 * allow for chunk sizes up to 256K, which is probably OK.
7326 * If the chunk size is greater, user-space should request more
7327 * stripe_heads first.
7329 struct r5conf
*conf
= mddev
->private;
7330 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7331 > conf
->min_nr_stripes
||
7332 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7333 > conf
->min_nr_stripes
) {
7334 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7336 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7343 static int check_reshape(struct mddev
*mddev
)
7345 struct r5conf
*conf
= mddev
->private;
7349 if (mddev
->delta_disks
== 0 &&
7350 mddev
->new_layout
== mddev
->layout
&&
7351 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7352 return 0; /* nothing to do */
7353 if (has_failed(conf
))
7355 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7356 /* We might be able to shrink, but the devices must
7357 * be made bigger first.
7358 * For raid6, 4 is the minimum size.
7359 * Otherwise 2 is the minimum
7362 if (mddev
->level
== 6)
7364 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7368 if (!check_stripe_cache(mddev
))
7371 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
7372 mddev
->delta_disks
> 0)
7373 if (resize_chunks(conf
,
7374 conf
->previous_raid_disks
7375 + max(0, mddev
->delta_disks
),
7376 max(mddev
->new_chunk_sectors
,
7377 mddev
->chunk_sectors
)
7380 return resize_stripes(conf
, (conf
->previous_raid_disks
7381 + mddev
->delta_disks
));
7384 static int raid5_start_reshape(struct mddev
*mddev
)
7386 struct r5conf
*conf
= mddev
->private;
7387 struct md_rdev
*rdev
;
7389 unsigned long flags
;
7391 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7394 if (!check_stripe_cache(mddev
))
7397 if (has_failed(conf
))
7400 rdev_for_each(rdev
, mddev
) {
7401 if (!test_bit(In_sync
, &rdev
->flags
)
7402 && !test_bit(Faulty
, &rdev
->flags
))
7406 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7407 /* Not enough devices even to make a degraded array
7412 /* Refuse to reduce size of the array. Any reductions in
7413 * array size must be through explicit setting of array_size
7416 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7417 < mddev
->array_sectors
) {
7418 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
7419 "before number of disks\n", mdname(mddev
));
7423 atomic_set(&conf
->reshape_stripes
, 0);
7424 spin_lock_irq(&conf
->device_lock
);
7425 write_seqcount_begin(&conf
->gen_lock
);
7426 conf
->previous_raid_disks
= conf
->raid_disks
;
7427 conf
->raid_disks
+= mddev
->delta_disks
;
7428 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7429 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7430 conf
->prev_algo
= conf
->algorithm
;
7431 conf
->algorithm
= mddev
->new_layout
;
7433 /* Code that selects data_offset needs to see the generation update
7434 * if reshape_progress has been set - so a memory barrier needed.
7437 if (mddev
->reshape_backwards
)
7438 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7440 conf
->reshape_progress
= 0;
7441 conf
->reshape_safe
= conf
->reshape_progress
;
7442 write_seqcount_end(&conf
->gen_lock
);
7443 spin_unlock_irq(&conf
->device_lock
);
7445 /* Now make sure any requests that proceeded on the assumption
7446 * the reshape wasn't running - like Discard or Read - have
7449 mddev_suspend(mddev
);
7450 mddev_resume(mddev
);
7452 /* Add some new drives, as many as will fit.
7453 * We know there are enough to make the newly sized array work.
7454 * Don't add devices if we are reducing the number of
7455 * devices in the array. This is because it is not possible
7456 * to correctly record the "partially reconstructed" state of
7457 * such devices during the reshape and confusion could result.
7459 if (mddev
->delta_disks
>= 0) {
7460 rdev_for_each(rdev
, mddev
)
7461 if (rdev
->raid_disk
< 0 &&
7462 !test_bit(Faulty
, &rdev
->flags
)) {
7463 if (raid5_add_disk(mddev
, rdev
) == 0) {
7465 >= conf
->previous_raid_disks
)
7466 set_bit(In_sync
, &rdev
->flags
);
7468 rdev
->recovery_offset
= 0;
7470 if (sysfs_link_rdev(mddev
, rdev
))
7471 /* Failure here is OK */;
7473 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7474 && !test_bit(Faulty
, &rdev
->flags
)) {
7475 /* This is a spare that was manually added */
7476 set_bit(In_sync
, &rdev
->flags
);
7479 /* When a reshape changes the number of devices,
7480 * ->degraded is measured against the larger of the
7481 * pre and post number of devices.
7483 spin_lock_irqsave(&conf
->device_lock
, flags
);
7484 mddev
->degraded
= calc_degraded(conf
);
7485 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7487 mddev
->raid_disks
= conf
->raid_disks
;
7488 mddev
->reshape_position
= conf
->reshape_progress
;
7489 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7491 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7492 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7493 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
7494 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7495 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7496 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7498 if (!mddev
->sync_thread
) {
7499 mddev
->recovery
= 0;
7500 spin_lock_irq(&conf
->device_lock
);
7501 write_seqcount_begin(&conf
->gen_lock
);
7502 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7503 mddev
->new_chunk_sectors
=
7504 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7505 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7506 rdev_for_each(rdev
, mddev
)
7507 rdev
->new_data_offset
= rdev
->data_offset
;
7509 conf
->generation
--;
7510 conf
->reshape_progress
= MaxSector
;
7511 mddev
->reshape_position
= MaxSector
;
7512 write_seqcount_end(&conf
->gen_lock
);
7513 spin_unlock_irq(&conf
->device_lock
);
7516 conf
->reshape_checkpoint
= jiffies
;
7517 md_wakeup_thread(mddev
->sync_thread
);
7518 md_new_event(mddev
);
7522 /* This is called from the reshape thread and should make any
7523 * changes needed in 'conf'
7525 static void end_reshape(struct r5conf
*conf
)
7528 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7529 struct md_rdev
*rdev
;
7531 spin_lock_irq(&conf
->device_lock
);
7532 conf
->previous_raid_disks
= conf
->raid_disks
;
7533 rdev_for_each(rdev
, conf
->mddev
)
7534 rdev
->data_offset
= rdev
->new_data_offset
;
7536 conf
->reshape_progress
= MaxSector
;
7537 conf
->mddev
->reshape_position
= MaxSector
;
7538 spin_unlock_irq(&conf
->device_lock
);
7539 wake_up(&conf
->wait_for_overlap
);
7541 /* read-ahead size must cover two whole stripes, which is
7542 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7544 if (conf
->mddev
->queue
) {
7545 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7546 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7548 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
7549 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
7554 /* This is called from the raid5d thread with mddev_lock held.
7555 * It makes config changes to the device.
7557 static void raid5_finish_reshape(struct mddev
*mddev
)
7559 struct r5conf
*conf
= mddev
->private;
7561 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7563 if (mddev
->delta_disks
> 0) {
7564 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7565 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7566 revalidate_disk(mddev
->gendisk
);
7569 spin_lock_irq(&conf
->device_lock
);
7570 mddev
->degraded
= calc_degraded(conf
);
7571 spin_unlock_irq(&conf
->device_lock
);
7572 for (d
= conf
->raid_disks
;
7573 d
< conf
->raid_disks
- mddev
->delta_disks
;
7575 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
7577 clear_bit(In_sync
, &rdev
->flags
);
7578 rdev
= conf
->disks
[d
].replacement
;
7580 clear_bit(In_sync
, &rdev
->flags
);
7583 mddev
->layout
= conf
->algorithm
;
7584 mddev
->chunk_sectors
= conf
->chunk_sectors
;
7585 mddev
->reshape_position
= MaxSector
;
7586 mddev
->delta_disks
= 0;
7587 mddev
->reshape_backwards
= 0;
7591 static void raid5_quiesce(struct mddev
*mddev
, int state
)
7593 struct r5conf
*conf
= mddev
->private;
7596 case 2: /* resume for a suspend */
7597 wake_up(&conf
->wait_for_overlap
);
7600 case 1: /* stop all writes */
7601 lock_all_device_hash_locks_irq(conf
);
7602 /* '2' tells resync/reshape to pause so that all
7603 * active stripes can drain
7606 wait_event_cmd(conf
->wait_for_quiescent
,
7607 atomic_read(&conf
->active_stripes
) == 0 &&
7608 atomic_read(&conf
->active_aligned_reads
) == 0,
7609 unlock_all_device_hash_locks_irq(conf
),
7610 lock_all_device_hash_locks_irq(conf
));
7612 unlock_all_device_hash_locks_irq(conf
);
7613 /* allow reshape to continue */
7614 wake_up(&conf
->wait_for_overlap
);
7617 case 0: /* re-enable writes */
7618 lock_all_device_hash_locks_irq(conf
);
7620 wake_up(&conf
->wait_for_quiescent
);
7621 wake_up(&conf
->wait_for_overlap
);
7622 unlock_all_device_hash_locks_irq(conf
);
7625 r5l_quiesce(conf
->log
, state
);
7628 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
7630 struct r0conf
*raid0_conf
= mddev
->private;
7633 /* for raid0 takeover only one zone is supported */
7634 if (raid0_conf
->nr_strip_zones
> 1) {
7635 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7637 return ERR_PTR(-EINVAL
);
7640 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
7641 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
7642 mddev
->dev_sectors
= sectors
;
7643 mddev
->new_level
= level
;
7644 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7645 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
7646 mddev
->raid_disks
+= 1;
7647 mddev
->delta_disks
= 1;
7648 /* make sure it will be not marked as dirty */
7649 mddev
->recovery_cp
= MaxSector
;
7651 return setup_conf(mddev
);
7654 static void *raid5_takeover_raid1(struct mddev
*mddev
)
7658 if (mddev
->raid_disks
!= 2 ||
7659 mddev
->degraded
> 1)
7660 return ERR_PTR(-EINVAL
);
7662 /* Should check if there are write-behind devices? */
7664 chunksect
= 64*2; /* 64K by default */
7666 /* The array must be an exact multiple of chunksize */
7667 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
7670 if ((chunksect
<<9) < STRIPE_SIZE
)
7671 /* array size does not allow a suitable chunk size */
7672 return ERR_PTR(-EINVAL
);
7674 mddev
->new_level
= 5;
7675 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7676 mddev
->new_chunk_sectors
= chunksect
;
7678 return setup_conf(mddev
);
7681 static void *raid5_takeover_raid6(struct mddev
*mddev
)
7685 switch (mddev
->layout
) {
7686 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7687 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
7689 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7690 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
7692 case ALGORITHM_LEFT_SYMMETRIC_6
:
7693 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7695 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7696 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
7698 case ALGORITHM_PARITY_0_6
:
7699 new_layout
= ALGORITHM_PARITY_0
;
7701 case ALGORITHM_PARITY_N
:
7702 new_layout
= ALGORITHM_PARITY_N
;
7705 return ERR_PTR(-EINVAL
);
7707 mddev
->new_level
= 5;
7708 mddev
->new_layout
= new_layout
;
7709 mddev
->delta_disks
= -1;
7710 mddev
->raid_disks
-= 1;
7711 return setup_conf(mddev
);
7714 static int raid5_check_reshape(struct mddev
*mddev
)
7716 /* For a 2-drive array, the layout and chunk size can be changed
7717 * immediately as not restriping is needed.
7718 * For larger arrays we record the new value - after validation
7719 * to be used by a reshape pass.
7721 struct r5conf
*conf
= mddev
->private;
7722 int new_chunk
= mddev
->new_chunk_sectors
;
7724 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
7726 if (new_chunk
> 0) {
7727 if (!is_power_of_2(new_chunk
))
7729 if (new_chunk
< (PAGE_SIZE
>>9))
7731 if (mddev
->array_sectors
& (new_chunk
-1))
7732 /* not factor of array size */
7736 /* They look valid */
7738 if (mddev
->raid_disks
== 2) {
7739 /* can make the change immediately */
7740 if (mddev
->new_layout
>= 0) {
7741 conf
->algorithm
= mddev
->new_layout
;
7742 mddev
->layout
= mddev
->new_layout
;
7744 if (new_chunk
> 0) {
7745 conf
->chunk_sectors
= new_chunk
;
7746 mddev
->chunk_sectors
= new_chunk
;
7748 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7749 md_wakeup_thread(mddev
->thread
);
7751 return check_reshape(mddev
);
7754 static int raid6_check_reshape(struct mddev
*mddev
)
7756 int new_chunk
= mddev
->new_chunk_sectors
;
7758 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
7760 if (new_chunk
> 0) {
7761 if (!is_power_of_2(new_chunk
))
7763 if (new_chunk
< (PAGE_SIZE
>> 9))
7765 if (mddev
->array_sectors
& (new_chunk
-1))
7766 /* not factor of array size */
7770 /* They look valid */
7771 return check_reshape(mddev
);
7774 static void *raid5_takeover(struct mddev
*mddev
)
7776 /* raid5 can take over:
7777 * raid0 - if there is only one strip zone - make it a raid4 layout
7778 * raid1 - if there are two drives. We need to know the chunk size
7779 * raid4 - trivial - just use a raid4 layout.
7780 * raid6 - Providing it is a *_6 layout
7782 if (mddev
->level
== 0)
7783 return raid45_takeover_raid0(mddev
, 5);
7784 if (mddev
->level
== 1)
7785 return raid5_takeover_raid1(mddev
);
7786 if (mddev
->level
== 4) {
7787 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7788 mddev
->new_level
= 5;
7789 return setup_conf(mddev
);
7791 if (mddev
->level
== 6)
7792 return raid5_takeover_raid6(mddev
);
7794 return ERR_PTR(-EINVAL
);
7797 static void *raid4_takeover(struct mddev
*mddev
)
7799 /* raid4 can take over:
7800 * raid0 - if there is only one strip zone
7801 * raid5 - if layout is right
7803 if (mddev
->level
== 0)
7804 return raid45_takeover_raid0(mddev
, 4);
7805 if (mddev
->level
== 5 &&
7806 mddev
->layout
== ALGORITHM_PARITY_N
) {
7807 mddev
->new_layout
= 0;
7808 mddev
->new_level
= 4;
7809 return setup_conf(mddev
);
7811 return ERR_PTR(-EINVAL
);
7814 static struct md_personality raid5_personality
;
7816 static void *raid6_takeover(struct mddev
*mddev
)
7818 /* Currently can only take over a raid5. We map the
7819 * personality to an equivalent raid6 personality
7820 * with the Q block at the end.
7824 if (mddev
->pers
!= &raid5_personality
)
7825 return ERR_PTR(-EINVAL
);
7826 if (mddev
->degraded
> 1)
7827 return ERR_PTR(-EINVAL
);
7828 if (mddev
->raid_disks
> 253)
7829 return ERR_PTR(-EINVAL
);
7830 if (mddev
->raid_disks
< 3)
7831 return ERR_PTR(-EINVAL
);
7833 switch (mddev
->layout
) {
7834 case ALGORITHM_LEFT_ASYMMETRIC
:
7835 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
7837 case ALGORITHM_RIGHT_ASYMMETRIC
:
7838 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
7840 case ALGORITHM_LEFT_SYMMETRIC
:
7841 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
7843 case ALGORITHM_RIGHT_SYMMETRIC
:
7844 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
7846 case ALGORITHM_PARITY_0
:
7847 new_layout
= ALGORITHM_PARITY_0_6
;
7849 case ALGORITHM_PARITY_N
:
7850 new_layout
= ALGORITHM_PARITY_N
;
7853 return ERR_PTR(-EINVAL
);
7855 mddev
->new_level
= 6;
7856 mddev
->new_layout
= new_layout
;
7857 mddev
->delta_disks
= 1;
7858 mddev
->raid_disks
+= 1;
7859 return setup_conf(mddev
);
7862 static struct md_personality raid6_personality
=
7866 .owner
= THIS_MODULE
,
7867 .make_request
= make_request
,
7871 .error_handler
= error
,
7872 .hot_add_disk
= raid5_add_disk
,
7873 .hot_remove_disk
= raid5_remove_disk
,
7874 .spare_active
= raid5_spare_active
,
7875 .sync_request
= sync_request
,
7876 .resize
= raid5_resize
,
7878 .check_reshape
= raid6_check_reshape
,
7879 .start_reshape
= raid5_start_reshape
,
7880 .finish_reshape
= raid5_finish_reshape
,
7881 .quiesce
= raid5_quiesce
,
7882 .takeover
= raid6_takeover
,
7883 .congested
= raid5_congested
,
7885 static struct md_personality raid5_personality
=
7889 .owner
= THIS_MODULE
,
7890 .make_request
= make_request
,
7894 .error_handler
= error
,
7895 .hot_add_disk
= raid5_add_disk
,
7896 .hot_remove_disk
= raid5_remove_disk
,
7897 .spare_active
= raid5_spare_active
,
7898 .sync_request
= sync_request
,
7899 .resize
= raid5_resize
,
7901 .check_reshape
= raid5_check_reshape
,
7902 .start_reshape
= raid5_start_reshape
,
7903 .finish_reshape
= raid5_finish_reshape
,
7904 .quiesce
= raid5_quiesce
,
7905 .takeover
= raid5_takeover
,
7906 .congested
= raid5_congested
,
7909 static struct md_personality raid4_personality
=
7913 .owner
= THIS_MODULE
,
7914 .make_request
= make_request
,
7918 .error_handler
= error
,
7919 .hot_add_disk
= raid5_add_disk
,
7920 .hot_remove_disk
= raid5_remove_disk
,
7921 .spare_active
= raid5_spare_active
,
7922 .sync_request
= sync_request
,
7923 .resize
= raid5_resize
,
7925 .check_reshape
= raid5_check_reshape
,
7926 .start_reshape
= raid5_start_reshape
,
7927 .finish_reshape
= raid5_finish_reshape
,
7928 .quiesce
= raid5_quiesce
,
7929 .takeover
= raid4_takeover
,
7930 .congested
= raid5_congested
,
7933 static int __init
raid5_init(void)
7935 raid5_wq
= alloc_workqueue("raid5wq",
7936 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
7939 register_md_personality(&raid6_personality
);
7940 register_md_personality(&raid5_personality
);
7941 register_md_personality(&raid4_personality
);
7945 static void raid5_exit(void)
7947 unregister_md_personality(&raid6_personality
);
7948 unregister_md_personality(&raid5_personality
);
7949 unregister_md_personality(&raid4_personality
);
7950 destroy_workqueue(raid5_wq
);
7953 module_init(raid5_init
);
7954 module_exit(raid5_exit
);
7955 MODULE_LICENSE("GPL");
7956 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7957 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7958 MODULE_ALIAS("md-raid5");
7959 MODULE_ALIAS("md-raid4");
7960 MODULE_ALIAS("md-level-5");
7961 MODULE_ALIAS("md-level-4");
7962 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7963 MODULE_ALIAS("md-raid6");
7964 MODULE_ALIAS("md-level-6");
7966 /* This used to be two separate modules, they were: */
7967 MODULE_ALIAS("raid5");
7968 MODULE_ALIAS("raid6");