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, 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 static void 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
)
661 static struct stripe_head
*
662 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 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
759 !test_bit(STRIPE_BITMAP_PENDING
, &sh
->state
) &&
760 is_full_stripe_write(sh
);
763 /* we only do back search */
764 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
766 struct stripe_head
*head
;
767 sector_t head_sector
, tmp_sec
;
771 if (!stripe_can_batch(sh
))
773 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
774 tmp_sec
= sh
->sector
;
775 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
777 head_sector
= sh
->sector
- STRIPE_SECTORS
;
779 hash
= stripe_hash_locks_hash(head_sector
);
780 spin_lock_irq(conf
->hash_locks
+ hash
);
781 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
782 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
783 spin_lock(&conf
->device_lock
);
784 if (!atomic_read(&head
->count
)) {
785 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
786 atomic_inc(&conf
->active_stripes
);
787 BUG_ON(list_empty(&head
->lru
) &&
788 !test_bit(STRIPE_EXPANDING
, &head
->state
));
789 list_del_init(&head
->lru
);
791 head
->group
->stripes_cnt
--;
795 atomic_inc(&head
->count
);
796 spin_unlock(&conf
->device_lock
);
798 spin_unlock_irq(conf
->hash_locks
+ hash
);
802 if (!stripe_can_batch(head
))
805 lock_two_stripes(head
, sh
);
806 /* clear_batch_ready clear the flag */
807 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
814 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
816 if (head
->dev
[dd_idx
].towrite
->bi_rw
!= sh
->dev
[dd_idx
].towrite
->bi_rw
)
819 if (head
->batch_head
) {
820 spin_lock(&head
->batch_head
->batch_lock
);
821 /* This batch list is already running */
822 if (!stripe_can_batch(head
)) {
823 spin_unlock(&head
->batch_head
->batch_lock
);
828 * at this point, head's BATCH_READY could be cleared, but we
829 * can still add the stripe to batch list
831 list_add(&sh
->batch_list
, &head
->batch_list
);
832 spin_unlock(&head
->batch_head
->batch_lock
);
834 sh
->batch_head
= head
->batch_head
;
836 head
->batch_head
= head
;
837 sh
->batch_head
= head
->batch_head
;
838 spin_lock(&head
->batch_lock
);
839 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
840 spin_unlock(&head
->batch_lock
);
843 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
844 if (atomic_dec_return(&conf
->preread_active_stripes
)
846 md_wakeup_thread(conf
->mddev
->thread
);
848 if (test_and_clear_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
849 int seq
= sh
->bm_seq
;
850 if (test_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
) &&
851 sh
->batch_head
->bm_seq
> seq
)
852 seq
= sh
->batch_head
->bm_seq
;
853 set_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
);
854 sh
->batch_head
->bm_seq
= seq
;
857 atomic_inc(&sh
->count
);
859 unlock_two_stripes(head
, sh
);
861 release_stripe(head
);
864 /* Determine if 'data_offset' or 'new_data_offset' should be used
865 * in this stripe_head.
867 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
869 sector_t progress
= conf
->reshape_progress
;
870 /* Need a memory barrier to make sure we see the value
871 * of conf->generation, or ->data_offset that was set before
872 * reshape_progress was updated.
875 if (progress
== MaxSector
)
877 if (sh
->generation
== conf
->generation
- 1)
879 /* We are in a reshape, and this is a new-generation stripe,
880 * so use new_data_offset.
886 raid5_end_read_request(struct bio
*bi
);
888 raid5_end_write_request(struct bio
*bi
);
890 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
892 struct r5conf
*conf
= sh
->raid_conf
;
893 int i
, disks
= sh
->disks
;
894 struct stripe_head
*head_sh
= sh
;
898 for (i
= disks
; i
--; ) {
900 int replace_only
= 0;
901 struct bio
*bi
, *rbi
;
902 struct md_rdev
*rdev
, *rrdev
= NULL
;
905 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
906 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
910 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
912 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
914 else if (test_and_clear_bit(R5_WantReplace
,
915 &sh
->dev
[i
].flags
)) {
920 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
924 bi
= &sh
->dev
[i
].req
;
925 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
928 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
929 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
930 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
939 /* We raced and saw duplicates */
942 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
947 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
950 atomic_inc(&rdev
->nr_pending
);
951 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
954 atomic_inc(&rrdev
->nr_pending
);
957 /* We have already checked bad blocks for reads. Now
958 * need to check for writes. We never accept write errors
959 * on the replacement, so we don't to check rrdev.
961 while ((rw
& WRITE
) && rdev
&&
962 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
965 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
966 &first_bad
, &bad_sectors
);
971 set_bit(BlockedBadBlocks
, &rdev
->flags
);
972 if (!conf
->mddev
->external
&&
973 conf
->mddev
->flags
) {
974 /* It is very unlikely, but we might
975 * still need to write out the
976 * bad block log - better give it
978 md_check_recovery(conf
->mddev
);
981 * Because md_wait_for_blocked_rdev
982 * will dec nr_pending, we must
983 * increment it first.
985 atomic_inc(&rdev
->nr_pending
);
986 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
988 /* Acknowledged bad block - skip the write */
989 rdev_dec_pending(rdev
, conf
->mddev
);
995 if (s
->syncing
|| s
->expanding
|| s
->expanded
997 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
999 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1002 bi
->bi_bdev
= rdev
->bdev
;
1004 bi
->bi_end_io
= (rw
& WRITE
)
1005 ? raid5_end_write_request
1006 : raid5_end_read_request
;
1007 bi
->bi_private
= sh
;
1009 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1010 __func__
, (unsigned long long)sh
->sector
,
1012 atomic_inc(&sh
->count
);
1014 atomic_inc(&head_sh
->count
);
1015 if (use_new_offset(conf
, sh
))
1016 bi
->bi_iter
.bi_sector
= (sh
->sector
1017 + rdev
->new_data_offset
);
1019 bi
->bi_iter
.bi_sector
= (sh
->sector
1020 + rdev
->data_offset
);
1021 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1022 bi
->bi_rw
|= REQ_NOMERGE
;
1024 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1025 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1026 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1028 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1029 bi
->bi_io_vec
[0].bv_offset
= 0;
1030 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1032 * If this is discard request, set bi_vcnt 0. We don't
1033 * want to confuse SCSI because SCSI will replace payload
1035 if (rw
& REQ_DISCARD
)
1038 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1040 if (conf
->mddev
->gendisk
)
1041 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
1042 bi
, disk_devt(conf
->mddev
->gendisk
),
1044 generic_make_request(bi
);
1047 if (s
->syncing
|| s
->expanding
|| s
->expanded
1049 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1051 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1054 rbi
->bi_bdev
= rrdev
->bdev
;
1056 BUG_ON(!(rw
& WRITE
));
1057 rbi
->bi_end_io
= raid5_end_write_request
;
1058 rbi
->bi_private
= sh
;
1060 pr_debug("%s: for %llu schedule op %ld on "
1061 "replacement disc %d\n",
1062 __func__
, (unsigned long long)sh
->sector
,
1064 atomic_inc(&sh
->count
);
1066 atomic_inc(&head_sh
->count
);
1067 if (use_new_offset(conf
, sh
))
1068 rbi
->bi_iter
.bi_sector
= (sh
->sector
1069 + rrdev
->new_data_offset
);
1071 rbi
->bi_iter
.bi_sector
= (sh
->sector
1072 + rrdev
->data_offset
);
1073 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1074 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1075 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1077 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1078 rbi
->bi_io_vec
[0].bv_offset
= 0;
1079 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1081 * If this is discard request, set bi_vcnt 0. We don't
1082 * want to confuse SCSI because SCSI will replace payload
1084 if (rw
& REQ_DISCARD
)
1086 if (conf
->mddev
->gendisk
)
1087 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
1088 rbi
, disk_devt(conf
->mddev
->gendisk
),
1090 generic_make_request(rbi
);
1092 if (!rdev
&& !rrdev
) {
1094 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1095 pr_debug("skip op %ld on disc %d for sector %llu\n",
1096 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
1097 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1098 set_bit(STRIPE_HANDLE
, &sh
->state
);
1101 if (!head_sh
->batch_head
)
1103 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1110 static struct dma_async_tx_descriptor
*
1111 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1112 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1113 struct stripe_head
*sh
)
1116 struct bvec_iter iter
;
1117 struct page
*bio_page
;
1119 struct async_submit_ctl submit
;
1120 enum async_tx_flags flags
= 0;
1122 if (bio
->bi_iter
.bi_sector
>= sector
)
1123 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1125 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1128 flags
|= ASYNC_TX_FENCE
;
1129 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1131 bio_for_each_segment(bvl
, bio
, iter
) {
1132 int len
= bvl
.bv_len
;
1136 if (page_offset
< 0) {
1137 b_offset
= -page_offset
;
1138 page_offset
+= b_offset
;
1142 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1143 clen
= STRIPE_SIZE
- page_offset
;
1148 b_offset
+= bvl
.bv_offset
;
1149 bio_page
= bvl
.bv_page
;
1151 if (sh
->raid_conf
->skip_copy
&&
1152 b_offset
== 0 && page_offset
== 0 &&
1153 clen
== STRIPE_SIZE
)
1156 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1157 b_offset
, clen
, &submit
);
1159 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1160 page_offset
, clen
, &submit
);
1162 /* chain the operations */
1163 submit
.depend_tx
= tx
;
1165 if (clen
< len
) /* hit end of page */
1173 static void ops_complete_biofill(void *stripe_head_ref
)
1175 struct stripe_head
*sh
= stripe_head_ref
;
1176 struct bio_list return_bi
= BIO_EMPTY_LIST
;
1179 pr_debug("%s: stripe %llu\n", __func__
,
1180 (unsigned long long)sh
->sector
);
1182 /* clear completed biofills */
1183 for (i
= sh
->disks
; i
--; ) {
1184 struct r5dev
*dev
= &sh
->dev
[i
];
1186 /* acknowledge completion of a biofill operation */
1187 /* and check if we need to reply to a read request,
1188 * new R5_Wantfill requests are held off until
1189 * !STRIPE_BIOFILL_RUN
1191 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1192 struct bio
*rbi
, *rbi2
;
1197 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1198 dev
->sector
+ STRIPE_SECTORS
) {
1199 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1200 if (!raid5_dec_bi_active_stripes(rbi
))
1201 bio_list_add(&return_bi
, rbi
);
1206 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1208 return_io(&return_bi
);
1210 set_bit(STRIPE_HANDLE
, &sh
->state
);
1214 static void ops_run_biofill(struct stripe_head
*sh
)
1216 struct dma_async_tx_descriptor
*tx
= NULL
;
1217 struct async_submit_ctl submit
;
1220 BUG_ON(sh
->batch_head
);
1221 pr_debug("%s: stripe %llu\n", __func__
,
1222 (unsigned long long)sh
->sector
);
1224 for (i
= sh
->disks
; i
--; ) {
1225 struct r5dev
*dev
= &sh
->dev
[i
];
1226 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1228 spin_lock_irq(&sh
->stripe_lock
);
1229 dev
->read
= rbi
= dev
->toread
;
1231 spin_unlock_irq(&sh
->stripe_lock
);
1232 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1233 dev
->sector
+ STRIPE_SECTORS
) {
1234 tx
= async_copy_data(0, rbi
, &dev
->page
,
1235 dev
->sector
, tx
, sh
);
1236 rbi
= r5_next_bio(rbi
, dev
->sector
);
1241 atomic_inc(&sh
->count
);
1242 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1243 async_trigger_callback(&submit
);
1246 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1253 tgt
= &sh
->dev
[target
];
1254 set_bit(R5_UPTODATE
, &tgt
->flags
);
1255 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1256 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1259 static void ops_complete_compute(void *stripe_head_ref
)
1261 struct stripe_head
*sh
= stripe_head_ref
;
1263 pr_debug("%s: stripe %llu\n", __func__
,
1264 (unsigned long long)sh
->sector
);
1266 /* mark the computed target(s) as uptodate */
1267 mark_target_uptodate(sh
, sh
->ops
.target
);
1268 mark_target_uptodate(sh
, sh
->ops
.target2
);
1270 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1271 if (sh
->check_state
== check_state_compute_run
)
1272 sh
->check_state
= check_state_compute_result
;
1273 set_bit(STRIPE_HANDLE
, &sh
->state
);
1277 /* return a pointer to the address conversion region of the scribble buffer */
1278 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1279 struct raid5_percpu
*percpu
, int i
)
1283 addr
= flex_array_get(percpu
->scribble
, i
);
1284 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1287 /* return a pointer to the address conversion region of the scribble buffer */
1288 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1292 addr
= flex_array_get(percpu
->scribble
, i
);
1296 static struct dma_async_tx_descriptor
*
1297 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1299 int disks
= sh
->disks
;
1300 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1301 int target
= sh
->ops
.target
;
1302 struct r5dev
*tgt
= &sh
->dev
[target
];
1303 struct page
*xor_dest
= tgt
->page
;
1305 struct dma_async_tx_descriptor
*tx
;
1306 struct async_submit_ctl submit
;
1309 BUG_ON(sh
->batch_head
);
1311 pr_debug("%s: stripe %llu block: %d\n",
1312 __func__
, (unsigned long long)sh
->sector
, target
);
1313 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1315 for (i
= disks
; i
--; )
1317 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1319 atomic_inc(&sh
->count
);
1321 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1322 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1323 if (unlikely(count
== 1))
1324 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1326 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1331 /* set_syndrome_sources - populate source buffers for gen_syndrome
1332 * @srcs - (struct page *) array of size sh->disks
1333 * @sh - stripe_head to parse
1335 * Populates srcs in proper layout order for the stripe and returns the
1336 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1337 * destination buffer is recorded in srcs[count] and the Q destination
1338 * is recorded in srcs[count+1]].
1340 static int set_syndrome_sources(struct page
**srcs
,
1341 struct stripe_head
*sh
,
1344 int disks
= sh
->disks
;
1345 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1346 int d0_idx
= raid6_d0(sh
);
1350 for (i
= 0; i
< disks
; i
++)
1356 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1357 struct r5dev
*dev
= &sh
->dev
[i
];
1359 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1360 (srctype
== SYNDROME_SRC_ALL
) ||
1361 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1362 test_bit(R5_Wantdrain
, &dev
->flags
)) ||
1363 (srctype
== SYNDROME_SRC_WRITTEN
&&
1365 srcs
[slot
] = sh
->dev
[i
].page
;
1366 i
= raid6_next_disk(i
, disks
);
1367 } while (i
!= d0_idx
);
1369 return syndrome_disks
;
1372 static struct dma_async_tx_descriptor
*
1373 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1375 int disks
= sh
->disks
;
1376 struct page
**blocks
= to_addr_page(percpu
, 0);
1378 int qd_idx
= sh
->qd_idx
;
1379 struct dma_async_tx_descriptor
*tx
;
1380 struct async_submit_ctl submit
;
1386 BUG_ON(sh
->batch_head
);
1387 if (sh
->ops
.target
< 0)
1388 target
= sh
->ops
.target2
;
1389 else if (sh
->ops
.target2
< 0)
1390 target
= sh
->ops
.target
;
1392 /* we should only have one valid target */
1395 pr_debug("%s: stripe %llu block: %d\n",
1396 __func__
, (unsigned long long)sh
->sector
, target
);
1398 tgt
= &sh
->dev
[target
];
1399 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1402 atomic_inc(&sh
->count
);
1404 if (target
== qd_idx
) {
1405 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1406 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1407 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1408 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1409 ops_complete_compute
, sh
,
1410 to_addr_conv(sh
, percpu
, 0));
1411 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1413 /* Compute any data- or p-drive using XOR */
1415 for (i
= disks
; i
-- ; ) {
1416 if (i
== target
|| i
== qd_idx
)
1418 blocks
[count
++] = sh
->dev
[i
].page
;
1421 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1422 NULL
, ops_complete_compute
, sh
,
1423 to_addr_conv(sh
, percpu
, 0));
1424 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1430 static struct dma_async_tx_descriptor
*
1431 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1433 int i
, count
, disks
= sh
->disks
;
1434 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1435 int d0_idx
= raid6_d0(sh
);
1436 int faila
= -1, failb
= -1;
1437 int target
= sh
->ops
.target
;
1438 int target2
= sh
->ops
.target2
;
1439 struct r5dev
*tgt
= &sh
->dev
[target
];
1440 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1441 struct dma_async_tx_descriptor
*tx
;
1442 struct page
**blocks
= to_addr_page(percpu
, 0);
1443 struct async_submit_ctl submit
;
1445 BUG_ON(sh
->batch_head
);
1446 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1447 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1448 BUG_ON(target
< 0 || target2
< 0);
1449 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1450 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1452 /* we need to open-code set_syndrome_sources to handle the
1453 * slot number conversion for 'faila' and 'failb'
1455 for (i
= 0; i
< disks
; i
++)
1460 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1462 blocks
[slot
] = sh
->dev
[i
].page
;
1468 i
= raid6_next_disk(i
, disks
);
1469 } while (i
!= d0_idx
);
1471 BUG_ON(faila
== failb
);
1474 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1475 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1477 atomic_inc(&sh
->count
);
1479 if (failb
== syndrome_disks
+1) {
1480 /* Q disk is one of the missing disks */
1481 if (faila
== syndrome_disks
) {
1482 /* Missing P+Q, just recompute */
1483 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1484 ops_complete_compute
, sh
,
1485 to_addr_conv(sh
, percpu
, 0));
1486 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1487 STRIPE_SIZE
, &submit
);
1491 int qd_idx
= sh
->qd_idx
;
1493 /* Missing D+Q: recompute D from P, then recompute Q */
1494 if (target
== qd_idx
)
1495 data_target
= target2
;
1497 data_target
= target
;
1500 for (i
= disks
; i
-- ; ) {
1501 if (i
== data_target
|| i
== qd_idx
)
1503 blocks
[count
++] = sh
->dev
[i
].page
;
1505 dest
= sh
->dev
[data_target
].page
;
1506 init_async_submit(&submit
,
1507 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1509 to_addr_conv(sh
, percpu
, 0));
1510 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1513 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1514 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1515 ops_complete_compute
, sh
,
1516 to_addr_conv(sh
, percpu
, 0));
1517 return async_gen_syndrome(blocks
, 0, count
+2,
1518 STRIPE_SIZE
, &submit
);
1521 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1522 ops_complete_compute
, sh
,
1523 to_addr_conv(sh
, percpu
, 0));
1524 if (failb
== syndrome_disks
) {
1525 /* We're missing D+P. */
1526 return async_raid6_datap_recov(syndrome_disks
+2,
1530 /* We're missing D+D. */
1531 return async_raid6_2data_recov(syndrome_disks
+2,
1532 STRIPE_SIZE
, faila
, failb
,
1538 static void ops_complete_prexor(void *stripe_head_ref
)
1540 struct stripe_head
*sh
= stripe_head_ref
;
1542 pr_debug("%s: stripe %llu\n", __func__
,
1543 (unsigned long long)sh
->sector
);
1546 static struct dma_async_tx_descriptor
*
1547 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1548 struct dma_async_tx_descriptor
*tx
)
1550 int disks
= sh
->disks
;
1551 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1552 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1553 struct async_submit_ctl submit
;
1555 /* existing parity data subtracted */
1556 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1558 BUG_ON(sh
->batch_head
);
1559 pr_debug("%s: stripe %llu\n", __func__
,
1560 (unsigned long long)sh
->sector
);
1562 for (i
= disks
; i
--; ) {
1563 struct r5dev
*dev
= &sh
->dev
[i
];
1564 /* Only process blocks that are known to be uptodate */
1565 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1566 xor_srcs
[count
++] = dev
->page
;
1569 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1570 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1571 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1576 static struct dma_async_tx_descriptor
*
1577 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1578 struct dma_async_tx_descriptor
*tx
)
1580 struct page
**blocks
= to_addr_page(percpu
, 0);
1582 struct async_submit_ctl submit
;
1584 pr_debug("%s: stripe %llu\n", __func__
,
1585 (unsigned long long)sh
->sector
);
1587 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1589 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1590 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1591 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1596 static struct dma_async_tx_descriptor
*
1597 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1599 int disks
= sh
->disks
;
1601 struct stripe_head
*head_sh
= sh
;
1603 pr_debug("%s: stripe %llu\n", __func__
,
1604 (unsigned long long)sh
->sector
);
1606 for (i
= disks
; i
--; ) {
1611 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1616 spin_lock_irq(&sh
->stripe_lock
);
1617 chosen
= dev
->towrite
;
1618 dev
->towrite
= NULL
;
1619 sh
->overwrite_disks
= 0;
1620 BUG_ON(dev
->written
);
1621 wbi
= dev
->written
= chosen
;
1622 spin_unlock_irq(&sh
->stripe_lock
);
1623 WARN_ON(dev
->page
!= dev
->orig_page
);
1625 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1626 dev
->sector
+ STRIPE_SECTORS
) {
1627 if (wbi
->bi_rw
& REQ_FUA
)
1628 set_bit(R5_WantFUA
, &dev
->flags
);
1629 if (wbi
->bi_rw
& REQ_SYNC
)
1630 set_bit(R5_SyncIO
, &dev
->flags
);
1631 if (wbi
->bi_rw
& REQ_DISCARD
)
1632 set_bit(R5_Discard
, &dev
->flags
);
1634 tx
= async_copy_data(1, wbi
, &dev
->page
,
1635 dev
->sector
, tx
, sh
);
1636 if (dev
->page
!= dev
->orig_page
) {
1637 set_bit(R5_SkipCopy
, &dev
->flags
);
1638 clear_bit(R5_UPTODATE
, &dev
->flags
);
1639 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1642 wbi
= r5_next_bio(wbi
, dev
->sector
);
1645 if (head_sh
->batch_head
) {
1646 sh
= list_first_entry(&sh
->batch_list
,
1659 static void ops_complete_reconstruct(void *stripe_head_ref
)
1661 struct stripe_head
*sh
= stripe_head_ref
;
1662 int disks
= sh
->disks
;
1663 int pd_idx
= sh
->pd_idx
;
1664 int qd_idx
= sh
->qd_idx
;
1666 bool fua
= false, sync
= false, discard
= false;
1668 pr_debug("%s: stripe %llu\n", __func__
,
1669 (unsigned long long)sh
->sector
);
1671 for (i
= disks
; i
--; ) {
1672 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1673 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1674 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1677 for (i
= disks
; i
--; ) {
1678 struct r5dev
*dev
= &sh
->dev
[i
];
1680 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1681 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1682 set_bit(R5_UPTODATE
, &dev
->flags
);
1684 set_bit(R5_WantFUA
, &dev
->flags
);
1686 set_bit(R5_SyncIO
, &dev
->flags
);
1690 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1691 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1692 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1693 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1695 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1696 sh
->reconstruct_state
= reconstruct_state_result
;
1699 set_bit(STRIPE_HANDLE
, &sh
->state
);
1704 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1705 struct dma_async_tx_descriptor
*tx
)
1707 int disks
= sh
->disks
;
1708 struct page
**xor_srcs
;
1709 struct async_submit_ctl submit
;
1710 int count
, pd_idx
= sh
->pd_idx
, i
;
1711 struct page
*xor_dest
;
1713 unsigned long flags
;
1715 struct stripe_head
*head_sh
= sh
;
1718 pr_debug("%s: stripe %llu\n", __func__
,
1719 (unsigned long long)sh
->sector
);
1721 for (i
= 0; i
< sh
->disks
; i
++) {
1724 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1727 if (i
>= sh
->disks
) {
1728 atomic_inc(&sh
->count
);
1729 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1730 ops_complete_reconstruct(sh
);
1735 xor_srcs
= to_addr_page(percpu
, j
);
1736 /* check if prexor is active which means only process blocks
1737 * that are part of a read-modify-write (written)
1739 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1741 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1742 for (i
= disks
; i
--; ) {
1743 struct r5dev
*dev
= &sh
->dev
[i
];
1744 if (head_sh
->dev
[i
].written
)
1745 xor_srcs
[count
++] = dev
->page
;
1748 xor_dest
= sh
->dev
[pd_idx
].page
;
1749 for (i
= disks
; i
--; ) {
1750 struct r5dev
*dev
= &sh
->dev
[i
];
1752 xor_srcs
[count
++] = dev
->page
;
1756 /* 1/ if we prexor'd then the dest is reused as a source
1757 * 2/ if we did not prexor then we are redoing the parity
1758 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1759 * for the synchronous xor case
1761 last_stripe
= !head_sh
->batch_head
||
1762 list_first_entry(&sh
->batch_list
,
1763 struct stripe_head
, batch_list
) == head_sh
;
1765 flags
= ASYNC_TX_ACK
|
1766 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1768 atomic_inc(&head_sh
->count
);
1769 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1770 to_addr_conv(sh
, percpu
, j
));
1772 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1773 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1774 to_addr_conv(sh
, percpu
, j
));
1777 if (unlikely(count
== 1))
1778 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1780 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1783 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1790 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1791 struct dma_async_tx_descriptor
*tx
)
1793 struct async_submit_ctl submit
;
1794 struct page
**blocks
;
1795 int count
, i
, j
= 0;
1796 struct stripe_head
*head_sh
= sh
;
1799 unsigned long txflags
;
1801 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1803 for (i
= 0; i
< sh
->disks
; i
++) {
1804 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1806 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1809 if (i
>= sh
->disks
) {
1810 atomic_inc(&sh
->count
);
1811 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1812 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1813 ops_complete_reconstruct(sh
);
1818 blocks
= to_addr_page(percpu
, j
);
1820 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1821 synflags
= SYNDROME_SRC_WRITTEN
;
1822 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1824 synflags
= SYNDROME_SRC_ALL
;
1825 txflags
= ASYNC_TX_ACK
;
1828 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1829 last_stripe
= !head_sh
->batch_head
||
1830 list_first_entry(&sh
->batch_list
,
1831 struct stripe_head
, batch_list
) == head_sh
;
1834 atomic_inc(&head_sh
->count
);
1835 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1836 head_sh
, to_addr_conv(sh
, percpu
, j
));
1838 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1839 to_addr_conv(sh
, percpu
, j
));
1840 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1843 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1849 static void ops_complete_check(void *stripe_head_ref
)
1851 struct stripe_head
*sh
= stripe_head_ref
;
1853 pr_debug("%s: stripe %llu\n", __func__
,
1854 (unsigned long long)sh
->sector
);
1856 sh
->check_state
= check_state_check_result
;
1857 set_bit(STRIPE_HANDLE
, &sh
->state
);
1861 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1863 int disks
= sh
->disks
;
1864 int pd_idx
= sh
->pd_idx
;
1865 int qd_idx
= sh
->qd_idx
;
1866 struct page
*xor_dest
;
1867 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1868 struct dma_async_tx_descriptor
*tx
;
1869 struct async_submit_ctl submit
;
1873 pr_debug("%s: stripe %llu\n", __func__
,
1874 (unsigned long long)sh
->sector
);
1876 BUG_ON(sh
->batch_head
);
1878 xor_dest
= sh
->dev
[pd_idx
].page
;
1879 xor_srcs
[count
++] = xor_dest
;
1880 for (i
= disks
; i
--; ) {
1881 if (i
== pd_idx
|| i
== qd_idx
)
1883 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1886 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1887 to_addr_conv(sh
, percpu
, 0));
1888 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1889 &sh
->ops
.zero_sum_result
, &submit
);
1891 atomic_inc(&sh
->count
);
1892 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1893 tx
= async_trigger_callback(&submit
);
1896 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1898 struct page
**srcs
= to_addr_page(percpu
, 0);
1899 struct async_submit_ctl submit
;
1902 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1903 (unsigned long long)sh
->sector
, checkp
);
1905 BUG_ON(sh
->batch_head
);
1906 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
1910 atomic_inc(&sh
->count
);
1911 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1912 sh
, to_addr_conv(sh
, percpu
, 0));
1913 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1914 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1917 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1919 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1920 struct dma_async_tx_descriptor
*tx
= NULL
;
1921 struct r5conf
*conf
= sh
->raid_conf
;
1922 int level
= conf
->level
;
1923 struct raid5_percpu
*percpu
;
1927 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1928 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1929 ops_run_biofill(sh
);
1933 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1935 tx
= ops_run_compute5(sh
, percpu
);
1937 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1938 tx
= ops_run_compute6_1(sh
, percpu
);
1940 tx
= ops_run_compute6_2(sh
, percpu
);
1942 /* terminate the chain if reconstruct is not set to be run */
1943 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1947 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
1949 tx
= ops_run_prexor5(sh
, percpu
, tx
);
1951 tx
= ops_run_prexor6(sh
, percpu
, tx
);
1954 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1955 tx
= ops_run_biodrain(sh
, tx
);
1959 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1961 ops_run_reconstruct5(sh
, percpu
, tx
);
1963 ops_run_reconstruct6(sh
, percpu
, tx
);
1966 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1967 if (sh
->check_state
== check_state_run
)
1968 ops_run_check_p(sh
, percpu
);
1969 else if (sh
->check_state
== check_state_run_q
)
1970 ops_run_check_pq(sh
, percpu
, 0);
1971 else if (sh
->check_state
== check_state_run_pq
)
1972 ops_run_check_pq(sh
, percpu
, 1);
1977 if (overlap_clear
&& !sh
->batch_head
)
1978 for (i
= disks
; i
--; ) {
1979 struct r5dev
*dev
= &sh
->dev
[i
];
1980 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1981 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1986 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
)
1988 struct stripe_head
*sh
;
1990 sh
= kmem_cache_zalloc(sc
, gfp
);
1992 spin_lock_init(&sh
->stripe_lock
);
1993 spin_lock_init(&sh
->batch_lock
);
1994 INIT_LIST_HEAD(&sh
->batch_list
);
1995 INIT_LIST_HEAD(&sh
->lru
);
1996 atomic_set(&sh
->count
, 1);
2000 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2002 struct stripe_head
*sh
;
2004 sh
= alloc_stripe(conf
->slab_cache
, gfp
);
2008 sh
->raid_conf
= conf
;
2010 if (grow_buffers(sh
, gfp
)) {
2012 kmem_cache_free(conf
->slab_cache
, sh
);
2015 sh
->hash_lock_index
=
2016 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2017 /* we just created an active stripe so... */
2018 atomic_inc(&conf
->active_stripes
);
2021 conf
->max_nr_stripes
++;
2025 static int grow_stripes(struct r5conf
*conf
, int num
)
2027 struct kmem_cache
*sc
;
2028 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2030 if (conf
->mddev
->gendisk
)
2031 sprintf(conf
->cache_name
[0],
2032 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2034 sprintf(conf
->cache_name
[0],
2035 "raid%d-%p", conf
->level
, conf
->mddev
);
2036 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2038 conf
->active_name
= 0;
2039 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2040 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2044 conf
->slab_cache
= sc
;
2045 conf
->pool_size
= devs
;
2047 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2054 * scribble_len - return the required size of the scribble region
2055 * @num - total number of disks in the array
2057 * The size must be enough to contain:
2058 * 1/ a struct page pointer for each device in the array +2
2059 * 2/ room to convert each entry in (1) to its corresponding dma
2060 * (dma_map_page()) or page (page_address()) address.
2062 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2063 * calculate over all devices (not just the data blocks), using zeros in place
2064 * of the P and Q blocks.
2066 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2068 struct flex_array
*ret
;
2071 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2072 ret
= flex_array_alloc(len
, cnt
, flags
);
2075 /* always prealloc all elements, so no locking is required */
2076 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2077 flex_array_free(ret
);
2083 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2088 mddev_suspend(conf
->mddev
);
2090 for_each_present_cpu(cpu
) {
2091 struct raid5_percpu
*percpu
;
2092 struct flex_array
*scribble
;
2094 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2095 scribble
= scribble_alloc(new_disks
,
2096 new_sectors
/ STRIPE_SECTORS
,
2100 flex_array_free(percpu
->scribble
);
2101 percpu
->scribble
= scribble
;
2108 mddev_resume(conf
->mddev
);
2112 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2114 /* Make all the stripes able to hold 'newsize' devices.
2115 * New slots in each stripe get 'page' set to a new page.
2117 * This happens in stages:
2118 * 1/ create a new kmem_cache and allocate the required number of
2120 * 2/ gather all the old stripe_heads and transfer the pages across
2121 * to the new stripe_heads. This will have the side effect of
2122 * freezing the array as once all stripe_heads have been collected,
2123 * no IO will be possible. Old stripe heads are freed once their
2124 * pages have been transferred over, and the old kmem_cache is
2125 * freed when all stripes are done.
2126 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2127 * we simple return a failre status - no need to clean anything up.
2128 * 4/ allocate new pages for the new slots in the new stripe_heads.
2129 * If this fails, we don't bother trying the shrink the
2130 * stripe_heads down again, we just leave them as they are.
2131 * As each stripe_head is processed the new one is released into
2134 * Once step2 is started, we cannot afford to wait for a write,
2135 * so we use GFP_NOIO allocations.
2137 struct stripe_head
*osh
, *nsh
;
2138 LIST_HEAD(newstripes
);
2139 struct disk_info
*ndisks
;
2141 struct kmem_cache
*sc
;
2145 if (newsize
<= conf
->pool_size
)
2146 return 0; /* never bother to shrink */
2148 err
= md_allow_write(conf
->mddev
);
2153 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2154 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2159 /* Need to ensure auto-resizing doesn't interfere */
2160 mutex_lock(&conf
->cache_size_mutex
);
2162 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2163 nsh
= alloc_stripe(sc
, GFP_KERNEL
);
2167 nsh
->raid_conf
= conf
;
2168 list_add(&nsh
->lru
, &newstripes
);
2171 /* didn't get enough, give up */
2172 while (!list_empty(&newstripes
)) {
2173 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2174 list_del(&nsh
->lru
);
2175 kmem_cache_free(sc
, nsh
);
2177 kmem_cache_destroy(sc
);
2178 mutex_unlock(&conf
->cache_size_mutex
);
2181 /* Step 2 - Must use GFP_NOIO now.
2182 * OK, we have enough stripes, start collecting inactive
2183 * stripes and copying them over
2187 list_for_each_entry(nsh
, &newstripes
, lru
) {
2188 lock_device_hash_lock(conf
, hash
);
2189 wait_event_exclusive_cmd(conf
->wait_for_stripe
[hash
],
2190 !list_empty(conf
->inactive_list
+ hash
),
2191 unlock_device_hash_lock(conf
, hash
),
2192 lock_device_hash_lock(conf
, hash
));
2193 osh
= get_free_stripe(conf
, hash
);
2194 unlock_device_hash_lock(conf
, hash
);
2196 for(i
=0; i
<conf
->pool_size
; i
++) {
2197 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2198 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2200 nsh
->hash_lock_index
= hash
;
2201 kmem_cache_free(conf
->slab_cache
, osh
);
2203 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2204 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2209 kmem_cache_destroy(conf
->slab_cache
);
2212 * At this point, we are holding all the stripes so the array
2213 * is completely stalled, so now is a good time to resize
2214 * conf->disks and the scribble region
2216 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2218 for (i
=0; i
<conf
->raid_disks
; i
++)
2219 ndisks
[i
] = conf
->disks
[i
];
2221 conf
->disks
= ndisks
;
2225 mutex_unlock(&conf
->cache_size_mutex
);
2226 /* Step 4, return new stripes to service */
2227 while(!list_empty(&newstripes
)) {
2228 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2229 list_del_init(&nsh
->lru
);
2231 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2232 if (nsh
->dev
[i
].page
== NULL
) {
2233 struct page
*p
= alloc_page(GFP_NOIO
);
2234 nsh
->dev
[i
].page
= p
;
2235 nsh
->dev
[i
].orig_page
= p
;
2239 release_stripe(nsh
);
2241 /* critical section pass, GFP_NOIO no longer needed */
2243 conf
->slab_cache
= sc
;
2244 conf
->active_name
= 1-conf
->active_name
;
2246 conf
->pool_size
= newsize
;
2250 static int drop_one_stripe(struct r5conf
*conf
)
2252 struct stripe_head
*sh
;
2253 int hash
= (conf
->max_nr_stripes
- 1) & STRIPE_HASH_LOCKS_MASK
;
2255 spin_lock_irq(conf
->hash_locks
+ hash
);
2256 sh
= get_free_stripe(conf
, hash
);
2257 spin_unlock_irq(conf
->hash_locks
+ hash
);
2260 BUG_ON(atomic_read(&sh
->count
));
2262 kmem_cache_free(conf
->slab_cache
, sh
);
2263 atomic_dec(&conf
->active_stripes
);
2264 conf
->max_nr_stripes
--;
2268 static void shrink_stripes(struct r5conf
*conf
)
2270 while (conf
->max_nr_stripes
&&
2271 drop_one_stripe(conf
))
2274 if (conf
->slab_cache
)
2275 kmem_cache_destroy(conf
->slab_cache
);
2276 conf
->slab_cache
= NULL
;
2279 static void raid5_end_read_request(struct bio
* bi
)
2281 struct stripe_head
*sh
= bi
->bi_private
;
2282 struct r5conf
*conf
= sh
->raid_conf
;
2283 int disks
= sh
->disks
, i
;
2284 char b
[BDEVNAME_SIZE
];
2285 struct md_rdev
*rdev
= NULL
;
2288 for (i
=0 ; i
<disks
; i
++)
2289 if (bi
== &sh
->dev
[i
].req
)
2292 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2293 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2299 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2300 /* If replacement finished while this request was outstanding,
2301 * 'replacement' might be NULL already.
2302 * In that case it moved down to 'rdev'.
2303 * rdev is not removed until all requests are finished.
2305 rdev
= conf
->disks
[i
].replacement
;
2307 rdev
= conf
->disks
[i
].rdev
;
2309 if (use_new_offset(conf
, sh
))
2310 s
= sh
->sector
+ rdev
->new_data_offset
;
2312 s
= sh
->sector
+ rdev
->data_offset
;
2313 if (!bi
->bi_error
) {
2314 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2315 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2316 /* Note that this cannot happen on a
2317 * replacement device. We just fail those on
2322 "md/raid:%s: read error corrected"
2323 " (%lu sectors at %llu on %s)\n",
2324 mdname(conf
->mddev
), STRIPE_SECTORS
,
2325 (unsigned long long)s
,
2326 bdevname(rdev
->bdev
, b
));
2327 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2328 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2329 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2330 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2331 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2333 if (atomic_read(&rdev
->read_errors
))
2334 atomic_set(&rdev
->read_errors
, 0);
2336 const char *bdn
= bdevname(rdev
->bdev
, b
);
2340 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2341 atomic_inc(&rdev
->read_errors
);
2342 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2345 "md/raid:%s: read error on replacement device "
2346 "(sector %llu on %s).\n",
2347 mdname(conf
->mddev
),
2348 (unsigned long long)s
,
2350 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2354 "md/raid:%s: read error not correctable "
2355 "(sector %llu on %s).\n",
2356 mdname(conf
->mddev
),
2357 (unsigned long long)s
,
2359 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2364 "md/raid:%s: read error NOT corrected!! "
2365 "(sector %llu on %s).\n",
2366 mdname(conf
->mddev
),
2367 (unsigned long long)s
,
2369 } else if (atomic_read(&rdev
->read_errors
)
2370 > conf
->max_nr_stripes
)
2372 "md/raid:%s: Too many read errors, failing device %s.\n",
2373 mdname(conf
->mddev
), bdn
);
2376 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2377 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2380 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2381 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2382 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2384 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2386 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2387 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2389 && test_bit(In_sync
, &rdev
->flags
)
2390 && rdev_set_badblocks(
2391 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2392 md_error(conf
->mddev
, rdev
);
2395 rdev_dec_pending(rdev
, conf
->mddev
);
2396 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2397 set_bit(STRIPE_HANDLE
, &sh
->state
);
2401 static void raid5_end_write_request(struct bio
*bi
)
2403 struct stripe_head
*sh
= bi
->bi_private
;
2404 struct r5conf
*conf
= sh
->raid_conf
;
2405 int disks
= sh
->disks
, i
;
2406 struct md_rdev
*uninitialized_var(rdev
);
2409 int replacement
= 0;
2411 for (i
= 0 ; i
< disks
; i
++) {
2412 if (bi
== &sh
->dev
[i
].req
) {
2413 rdev
= conf
->disks
[i
].rdev
;
2416 if (bi
== &sh
->dev
[i
].rreq
) {
2417 rdev
= conf
->disks
[i
].replacement
;
2421 /* rdev was removed and 'replacement'
2422 * replaced it. rdev is not removed
2423 * until all requests are finished.
2425 rdev
= conf
->disks
[i
].rdev
;
2429 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2430 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2439 md_error(conf
->mddev
, rdev
);
2440 else if (is_badblock(rdev
, sh
->sector
,
2442 &first_bad
, &bad_sectors
))
2443 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2446 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2447 set_bit(WriteErrorSeen
, &rdev
->flags
);
2448 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2449 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2450 set_bit(MD_RECOVERY_NEEDED
,
2451 &rdev
->mddev
->recovery
);
2452 } else if (is_badblock(rdev
, sh
->sector
,
2454 &first_bad
, &bad_sectors
)) {
2455 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2456 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2457 /* That was a successful write so make
2458 * sure it looks like we already did
2461 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2464 rdev_dec_pending(rdev
, conf
->mddev
);
2466 if (sh
->batch_head
&& bi
->bi_error
&& !replacement
)
2467 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2469 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2470 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2471 set_bit(STRIPE_HANDLE
, &sh
->state
);
2474 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2475 release_stripe(sh
->batch_head
);
2478 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
2480 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
2482 struct r5dev
*dev
= &sh
->dev
[i
];
2484 bio_init(&dev
->req
);
2485 dev
->req
.bi_io_vec
= &dev
->vec
;
2486 dev
->req
.bi_max_vecs
= 1;
2487 dev
->req
.bi_private
= sh
;
2489 bio_init(&dev
->rreq
);
2490 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
2491 dev
->rreq
.bi_max_vecs
= 1;
2492 dev
->rreq
.bi_private
= sh
;
2495 dev
->sector
= compute_blocknr(sh
, i
, previous
);
2498 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2500 char b
[BDEVNAME_SIZE
];
2501 struct r5conf
*conf
= mddev
->private;
2502 unsigned long flags
;
2503 pr_debug("raid456: error called\n");
2505 spin_lock_irqsave(&conf
->device_lock
, flags
);
2506 clear_bit(In_sync
, &rdev
->flags
);
2507 mddev
->degraded
= calc_degraded(conf
);
2508 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2509 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2511 set_bit(Blocked
, &rdev
->flags
);
2512 set_bit(Faulty
, &rdev
->flags
);
2513 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2514 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
2516 "md/raid:%s: Disk failure on %s, disabling device.\n"
2517 "md/raid:%s: Operation continuing on %d devices.\n",
2519 bdevname(rdev
->bdev
, b
),
2521 conf
->raid_disks
- mddev
->degraded
);
2525 * Input: a 'big' sector number,
2526 * Output: index of the data and parity disk, and the sector # in them.
2528 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2529 int previous
, int *dd_idx
,
2530 struct stripe_head
*sh
)
2532 sector_t stripe
, stripe2
;
2533 sector_t chunk_number
;
2534 unsigned int chunk_offset
;
2537 sector_t new_sector
;
2538 int algorithm
= previous
? conf
->prev_algo
2540 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2541 : conf
->chunk_sectors
;
2542 int raid_disks
= previous
? conf
->previous_raid_disks
2544 int data_disks
= raid_disks
- conf
->max_degraded
;
2546 /* First compute the information on this sector */
2549 * Compute the chunk number and the sector offset inside the chunk
2551 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2552 chunk_number
= r_sector
;
2555 * Compute the stripe number
2557 stripe
= chunk_number
;
2558 *dd_idx
= sector_div(stripe
, data_disks
);
2561 * Select the parity disk based on the user selected algorithm.
2563 pd_idx
= qd_idx
= -1;
2564 switch(conf
->level
) {
2566 pd_idx
= data_disks
;
2569 switch (algorithm
) {
2570 case ALGORITHM_LEFT_ASYMMETRIC
:
2571 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2572 if (*dd_idx
>= pd_idx
)
2575 case ALGORITHM_RIGHT_ASYMMETRIC
:
2576 pd_idx
= sector_div(stripe2
, raid_disks
);
2577 if (*dd_idx
>= pd_idx
)
2580 case ALGORITHM_LEFT_SYMMETRIC
:
2581 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2582 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2584 case ALGORITHM_RIGHT_SYMMETRIC
:
2585 pd_idx
= sector_div(stripe2
, raid_disks
);
2586 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2588 case ALGORITHM_PARITY_0
:
2592 case ALGORITHM_PARITY_N
:
2593 pd_idx
= data_disks
;
2601 switch (algorithm
) {
2602 case ALGORITHM_LEFT_ASYMMETRIC
:
2603 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2604 qd_idx
= pd_idx
+ 1;
2605 if (pd_idx
== raid_disks
-1) {
2606 (*dd_idx
)++; /* Q D D D P */
2608 } else if (*dd_idx
>= pd_idx
)
2609 (*dd_idx
) += 2; /* D D P Q D */
2611 case ALGORITHM_RIGHT_ASYMMETRIC
:
2612 pd_idx
= sector_div(stripe2
, raid_disks
);
2613 qd_idx
= pd_idx
+ 1;
2614 if (pd_idx
== raid_disks
-1) {
2615 (*dd_idx
)++; /* Q D D D P */
2617 } else if (*dd_idx
>= pd_idx
)
2618 (*dd_idx
) += 2; /* D D P Q D */
2620 case ALGORITHM_LEFT_SYMMETRIC
:
2621 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2622 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2623 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2625 case ALGORITHM_RIGHT_SYMMETRIC
:
2626 pd_idx
= sector_div(stripe2
, raid_disks
);
2627 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2628 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2631 case ALGORITHM_PARITY_0
:
2636 case ALGORITHM_PARITY_N
:
2637 pd_idx
= data_disks
;
2638 qd_idx
= data_disks
+ 1;
2641 case ALGORITHM_ROTATING_ZERO_RESTART
:
2642 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2643 * of blocks for computing Q is different.
2645 pd_idx
= sector_div(stripe2
, raid_disks
);
2646 qd_idx
= pd_idx
+ 1;
2647 if (pd_idx
== raid_disks
-1) {
2648 (*dd_idx
)++; /* Q D D D P */
2650 } else if (*dd_idx
>= pd_idx
)
2651 (*dd_idx
) += 2; /* D D P Q D */
2655 case ALGORITHM_ROTATING_N_RESTART
:
2656 /* Same a left_asymmetric, by first stripe is
2657 * D D D P Q rather than
2661 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2662 qd_idx
= pd_idx
+ 1;
2663 if (pd_idx
== raid_disks
-1) {
2664 (*dd_idx
)++; /* Q D D D P */
2666 } else if (*dd_idx
>= pd_idx
)
2667 (*dd_idx
) += 2; /* D D P Q D */
2671 case ALGORITHM_ROTATING_N_CONTINUE
:
2672 /* Same as left_symmetric but Q is before P */
2673 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2674 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2675 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2679 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2680 /* RAID5 left_asymmetric, with Q on last device */
2681 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2682 if (*dd_idx
>= pd_idx
)
2684 qd_idx
= raid_disks
- 1;
2687 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2688 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2689 if (*dd_idx
>= pd_idx
)
2691 qd_idx
= raid_disks
- 1;
2694 case ALGORITHM_LEFT_SYMMETRIC_6
:
2695 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2696 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2697 qd_idx
= raid_disks
- 1;
2700 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2701 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2702 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2703 qd_idx
= raid_disks
- 1;
2706 case ALGORITHM_PARITY_0_6
:
2709 qd_idx
= raid_disks
- 1;
2719 sh
->pd_idx
= pd_idx
;
2720 sh
->qd_idx
= qd_idx
;
2721 sh
->ddf_layout
= ddf_layout
;
2724 * Finally, compute the new sector number
2726 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2730 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2732 struct r5conf
*conf
= sh
->raid_conf
;
2733 int raid_disks
= sh
->disks
;
2734 int data_disks
= raid_disks
- conf
->max_degraded
;
2735 sector_t new_sector
= sh
->sector
, check
;
2736 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2737 : conf
->chunk_sectors
;
2738 int algorithm
= previous
? conf
->prev_algo
2742 sector_t chunk_number
;
2743 int dummy1
, dd_idx
= i
;
2745 struct stripe_head sh2
;
2747 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2748 stripe
= new_sector
;
2750 if (i
== sh
->pd_idx
)
2752 switch(conf
->level
) {
2755 switch (algorithm
) {
2756 case ALGORITHM_LEFT_ASYMMETRIC
:
2757 case ALGORITHM_RIGHT_ASYMMETRIC
:
2761 case ALGORITHM_LEFT_SYMMETRIC
:
2762 case ALGORITHM_RIGHT_SYMMETRIC
:
2765 i
-= (sh
->pd_idx
+ 1);
2767 case ALGORITHM_PARITY_0
:
2770 case ALGORITHM_PARITY_N
:
2777 if (i
== sh
->qd_idx
)
2778 return 0; /* It is the Q disk */
2779 switch (algorithm
) {
2780 case ALGORITHM_LEFT_ASYMMETRIC
:
2781 case ALGORITHM_RIGHT_ASYMMETRIC
:
2782 case ALGORITHM_ROTATING_ZERO_RESTART
:
2783 case ALGORITHM_ROTATING_N_RESTART
:
2784 if (sh
->pd_idx
== raid_disks
-1)
2785 i
--; /* Q D D D P */
2786 else if (i
> sh
->pd_idx
)
2787 i
-= 2; /* D D P Q D */
2789 case ALGORITHM_LEFT_SYMMETRIC
:
2790 case ALGORITHM_RIGHT_SYMMETRIC
:
2791 if (sh
->pd_idx
== raid_disks
-1)
2792 i
--; /* Q D D D P */
2797 i
-= (sh
->pd_idx
+ 2);
2800 case ALGORITHM_PARITY_0
:
2803 case ALGORITHM_PARITY_N
:
2805 case ALGORITHM_ROTATING_N_CONTINUE
:
2806 /* Like left_symmetric, but P is before Q */
2807 if (sh
->pd_idx
== 0)
2808 i
--; /* P D D D Q */
2813 i
-= (sh
->pd_idx
+ 1);
2816 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2817 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2821 case ALGORITHM_LEFT_SYMMETRIC_6
:
2822 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2824 i
+= data_disks
+ 1;
2825 i
-= (sh
->pd_idx
+ 1);
2827 case ALGORITHM_PARITY_0_6
:
2836 chunk_number
= stripe
* data_disks
+ i
;
2837 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2839 check
= raid5_compute_sector(conf
, r_sector
,
2840 previous
, &dummy1
, &sh2
);
2841 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2842 || sh2
.qd_idx
!= sh
->qd_idx
) {
2843 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2844 mdname(conf
->mddev
));
2851 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2852 int rcw
, int expand
)
2854 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
2855 struct r5conf
*conf
= sh
->raid_conf
;
2856 int level
= conf
->level
;
2860 for (i
= disks
; i
--; ) {
2861 struct r5dev
*dev
= &sh
->dev
[i
];
2864 set_bit(R5_LOCKED
, &dev
->flags
);
2865 set_bit(R5_Wantdrain
, &dev
->flags
);
2867 clear_bit(R5_UPTODATE
, &dev
->flags
);
2871 /* if we are not expanding this is a proper write request, and
2872 * there will be bios with new data to be drained into the
2877 /* False alarm, nothing to do */
2879 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2880 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2882 sh
->reconstruct_state
= reconstruct_state_run
;
2884 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2886 if (s
->locked
+ conf
->max_degraded
== disks
)
2887 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2888 atomic_inc(&conf
->pending_full_writes
);
2890 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2891 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2892 BUG_ON(level
== 6 &&
2893 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
2894 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
2896 for (i
= disks
; i
--; ) {
2897 struct r5dev
*dev
= &sh
->dev
[i
];
2898 if (i
== pd_idx
|| i
== qd_idx
)
2902 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2903 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2904 set_bit(R5_Wantdrain
, &dev
->flags
);
2905 set_bit(R5_LOCKED
, &dev
->flags
);
2906 clear_bit(R5_UPTODATE
, &dev
->flags
);
2911 /* False alarm - nothing to do */
2913 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2914 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2915 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2916 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2919 /* keep the parity disk(s) locked while asynchronous operations
2922 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2923 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2927 int qd_idx
= sh
->qd_idx
;
2928 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2930 set_bit(R5_LOCKED
, &dev
->flags
);
2931 clear_bit(R5_UPTODATE
, &dev
->flags
);
2935 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2936 __func__
, (unsigned long long)sh
->sector
,
2937 s
->locked
, s
->ops_request
);
2941 * Each stripe/dev can have one or more bion attached.
2942 * toread/towrite point to the first in a chain.
2943 * The bi_next chain must be in order.
2945 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
2946 int forwrite
, int previous
)
2949 struct r5conf
*conf
= sh
->raid_conf
;
2952 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2953 (unsigned long long)bi
->bi_iter
.bi_sector
,
2954 (unsigned long long)sh
->sector
);
2957 * If several bio share a stripe. The bio bi_phys_segments acts as a
2958 * reference count to avoid race. The reference count should already be
2959 * increased before this function is called (for example, in
2960 * make_request()), so other bio sharing this stripe will not free the
2961 * stripe. If a stripe is owned by one stripe, the stripe lock will
2964 spin_lock_irq(&sh
->stripe_lock
);
2965 /* Don't allow new IO added to stripes in batch list */
2969 bip
= &sh
->dev
[dd_idx
].towrite
;
2973 bip
= &sh
->dev
[dd_idx
].toread
;
2974 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
2975 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
2977 bip
= & (*bip
)->bi_next
;
2979 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
2982 if (!forwrite
|| previous
)
2983 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
2985 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2989 raid5_inc_bi_active_stripes(bi
);
2992 /* check if page is covered */
2993 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2994 for (bi
=sh
->dev
[dd_idx
].towrite
;
2995 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2996 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
2997 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2998 if (bio_end_sector(bi
) >= sector
)
2999 sector
= bio_end_sector(bi
);
3001 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
3002 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
3003 sh
->overwrite_disks
++;
3006 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3007 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
3008 (unsigned long long)sh
->sector
, dd_idx
);
3010 if (conf
->mddev
->bitmap
&& firstwrite
) {
3011 /* Cannot hold spinlock over bitmap_startwrite,
3012 * but must ensure this isn't added to a batch until
3013 * we have added to the bitmap and set bm_seq.
3014 * So set STRIPE_BITMAP_PENDING to prevent
3016 * If multiple add_stripe_bio() calls race here they
3017 * much all set STRIPE_BITMAP_PENDING. So only the first one
3018 * to complete "bitmap_startwrite" gets to set
3019 * STRIPE_BIT_DELAY. This is important as once a stripe
3020 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3023 set_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3024 spin_unlock_irq(&sh
->stripe_lock
);
3025 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
3027 spin_lock_irq(&sh
->stripe_lock
);
3028 clear_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3029 if (!sh
->batch_head
) {
3030 sh
->bm_seq
= conf
->seq_flush
+1;
3031 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
3034 spin_unlock_irq(&sh
->stripe_lock
);
3036 if (stripe_can_batch(sh
))
3037 stripe_add_to_batch_list(conf
, sh
);
3041 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
3042 spin_unlock_irq(&sh
->stripe_lock
);
3046 static void end_reshape(struct r5conf
*conf
);
3048 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3049 struct stripe_head
*sh
)
3051 int sectors_per_chunk
=
3052 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3054 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3055 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3057 raid5_compute_sector(conf
,
3058 stripe
* (disks
- conf
->max_degraded
)
3059 *sectors_per_chunk
+ chunk_offset
,
3065 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3066 struct stripe_head_state
*s
, int disks
,
3067 struct bio_list
*return_bi
)
3070 BUG_ON(sh
->batch_head
);
3071 for (i
= disks
; i
--; ) {
3075 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3076 struct md_rdev
*rdev
;
3078 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3079 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
3080 atomic_inc(&rdev
->nr_pending
);
3085 if (!rdev_set_badblocks(
3089 md_error(conf
->mddev
, rdev
);
3090 rdev_dec_pending(rdev
, conf
->mddev
);
3093 spin_lock_irq(&sh
->stripe_lock
);
3094 /* fail all writes first */
3095 bi
= sh
->dev
[i
].towrite
;
3096 sh
->dev
[i
].towrite
= NULL
;
3097 sh
->overwrite_disks
= 0;
3098 spin_unlock_irq(&sh
->stripe_lock
);
3102 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3103 wake_up(&conf
->wait_for_overlap
);
3105 while (bi
&& bi
->bi_iter
.bi_sector
<
3106 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3107 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3109 bi
->bi_error
= -EIO
;
3110 if (!raid5_dec_bi_active_stripes(bi
)) {
3111 md_write_end(conf
->mddev
);
3112 bio_list_add(return_bi
, bi
);
3117 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3118 STRIPE_SECTORS
, 0, 0);
3120 /* and fail all 'written' */
3121 bi
= sh
->dev
[i
].written
;
3122 sh
->dev
[i
].written
= NULL
;
3123 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3124 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3125 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3128 if (bi
) bitmap_end
= 1;
3129 while (bi
&& bi
->bi_iter
.bi_sector
<
3130 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3131 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3133 bi
->bi_error
= -EIO
;
3134 if (!raid5_dec_bi_active_stripes(bi
)) {
3135 md_write_end(conf
->mddev
);
3136 bio_list_add(return_bi
, bi
);
3141 /* fail any reads if this device is non-operational and
3142 * the data has not reached the cache yet.
3144 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3145 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3146 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3147 spin_lock_irq(&sh
->stripe_lock
);
3148 bi
= sh
->dev
[i
].toread
;
3149 sh
->dev
[i
].toread
= NULL
;
3150 spin_unlock_irq(&sh
->stripe_lock
);
3151 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3152 wake_up(&conf
->wait_for_overlap
);
3153 while (bi
&& bi
->bi_iter
.bi_sector
<
3154 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3155 struct bio
*nextbi
=
3156 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3158 bi
->bi_error
= -EIO
;
3159 if (!raid5_dec_bi_active_stripes(bi
))
3160 bio_list_add(return_bi
, bi
);
3165 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3166 STRIPE_SECTORS
, 0, 0);
3167 /* If we were in the middle of a write the parity block might
3168 * still be locked - so just clear all R5_LOCKED flags
3170 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3173 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3174 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3175 md_wakeup_thread(conf
->mddev
->thread
);
3179 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3180 struct stripe_head_state
*s
)
3185 BUG_ON(sh
->batch_head
);
3186 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3187 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3188 wake_up(&conf
->wait_for_overlap
);
3191 /* There is nothing more to do for sync/check/repair.
3192 * Don't even need to abort as that is handled elsewhere
3193 * if needed, and not always wanted e.g. if there is a known
3195 * For recover/replace we need to record a bad block on all
3196 * non-sync devices, or abort the recovery
3198 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3199 /* During recovery devices cannot be removed, so
3200 * locking and refcounting of rdevs is not needed
3202 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3203 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
3205 && !test_bit(Faulty
, &rdev
->flags
)
3206 && !test_bit(In_sync
, &rdev
->flags
)
3207 && !rdev_set_badblocks(rdev
, sh
->sector
,
3210 rdev
= conf
->disks
[i
].replacement
;
3212 && !test_bit(Faulty
, &rdev
->flags
)
3213 && !test_bit(In_sync
, &rdev
->flags
)
3214 && !rdev_set_badblocks(rdev
, sh
->sector
,
3219 conf
->recovery_disabled
=
3220 conf
->mddev
->recovery_disabled
;
3222 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3225 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3227 struct md_rdev
*rdev
;
3229 /* Doing recovery so rcu locking not required */
3230 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
3232 && !test_bit(Faulty
, &rdev
->flags
)
3233 && !test_bit(In_sync
, &rdev
->flags
)
3234 && (rdev
->recovery_offset
<= sh
->sector
3235 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3241 /* fetch_block - checks the given member device to see if its data needs
3242 * to be read or computed to satisfy a request.
3244 * Returns 1 when no more member devices need to be checked, otherwise returns
3245 * 0 to tell the loop in handle_stripe_fill to continue
3248 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3249 int disk_idx
, int disks
)
3251 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3252 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3253 &sh
->dev
[s
->failed_num
[1]] };
3257 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3258 test_bit(R5_UPTODATE
, &dev
->flags
))
3259 /* No point reading this as we already have it or have
3260 * decided to get it.
3265 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3266 /* We need this block to directly satisfy a request */
3269 if (s
->syncing
|| s
->expanding
||
3270 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3271 /* When syncing, or expanding we read everything.
3272 * When replacing, we need the replaced block.
3276 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3277 (s
->failed
>= 2 && fdev
[1]->toread
))
3278 /* If we want to read from a failed device, then
3279 * we need to actually read every other device.
3283 /* Sometimes neither read-modify-write nor reconstruct-write
3284 * cycles can work. In those cases we read every block we
3285 * can. Then the parity-update is certain to have enough to
3287 * This can only be a problem when we need to write something,
3288 * and some device has failed. If either of those tests
3289 * fail we need look no further.
3291 if (!s
->failed
|| !s
->to_write
)
3294 if (test_bit(R5_Insync
, &dev
->flags
) &&
3295 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3296 /* Pre-reads at not permitted until after short delay
3297 * to gather multiple requests. However if this
3298 * device is no Insync, the block could only be be computed
3299 * and there is no need to delay that.
3303 for (i
= 0; i
< s
->failed
; i
++) {
3304 if (fdev
[i
]->towrite
&&
3305 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3306 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3307 /* If we have a partial write to a failed
3308 * device, then we will need to reconstruct
3309 * the content of that device, so all other
3310 * devices must be read.
3315 /* If we are forced to do a reconstruct-write, either because
3316 * the current RAID6 implementation only supports that, or
3317 * or because parity cannot be trusted and we are currently
3318 * recovering it, there is extra need to be careful.
3319 * If one of the devices that we would need to read, because
3320 * it is not being overwritten (and maybe not written at all)
3321 * is missing/faulty, then we need to read everything we can.
3323 if (sh
->raid_conf
->level
!= 6 &&
3324 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3325 /* reconstruct-write isn't being forced */
3327 for (i
= 0; i
< s
->failed
; i
++) {
3328 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3329 s
->failed_num
[i
] != sh
->qd_idx
&&
3330 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3331 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3338 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3339 int disk_idx
, int disks
)
3341 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3343 /* is the data in this block needed, and can we get it? */
3344 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3345 /* we would like to get this block, possibly by computing it,
3346 * otherwise read it if the backing disk is insync
3348 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3349 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3350 BUG_ON(sh
->batch_head
);
3351 if ((s
->uptodate
== disks
- 1) &&
3352 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3353 disk_idx
== s
->failed_num
[1]))) {
3354 /* have disk failed, and we're requested to fetch it;
3357 pr_debug("Computing stripe %llu block %d\n",
3358 (unsigned long long)sh
->sector
, disk_idx
);
3359 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3360 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3361 set_bit(R5_Wantcompute
, &dev
->flags
);
3362 sh
->ops
.target
= disk_idx
;
3363 sh
->ops
.target2
= -1; /* no 2nd target */
3365 /* Careful: from this point on 'uptodate' is in the eye
3366 * of raid_run_ops which services 'compute' operations
3367 * before writes. R5_Wantcompute flags a block that will
3368 * be R5_UPTODATE by the time it is needed for a
3369 * subsequent operation.
3373 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3374 /* Computing 2-failure is *very* expensive; only
3375 * do it if failed >= 2
3378 for (other
= disks
; other
--; ) {
3379 if (other
== disk_idx
)
3381 if (!test_bit(R5_UPTODATE
,
3382 &sh
->dev
[other
].flags
))
3386 pr_debug("Computing stripe %llu blocks %d,%d\n",
3387 (unsigned long long)sh
->sector
,
3389 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3390 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3391 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3392 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3393 sh
->ops
.target
= disk_idx
;
3394 sh
->ops
.target2
= other
;
3398 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3399 set_bit(R5_LOCKED
, &dev
->flags
);
3400 set_bit(R5_Wantread
, &dev
->flags
);
3402 pr_debug("Reading block %d (sync=%d)\n",
3403 disk_idx
, s
->syncing
);
3411 * handle_stripe_fill - read or compute data to satisfy pending requests.
3413 static void handle_stripe_fill(struct stripe_head
*sh
,
3414 struct stripe_head_state
*s
,
3419 /* look for blocks to read/compute, skip this if a compute
3420 * is already in flight, or if the stripe contents are in the
3421 * midst of changing due to a write
3423 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3424 !sh
->reconstruct_state
)
3425 for (i
= disks
; i
--; )
3426 if (fetch_block(sh
, s
, i
, disks
))
3428 set_bit(STRIPE_HANDLE
, &sh
->state
);
3431 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
3432 unsigned long handle_flags
);
3433 /* handle_stripe_clean_event
3434 * any written block on an uptodate or failed drive can be returned.
3435 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3436 * never LOCKED, so we don't need to test 'failed' directly.
3438 static void handle_stripe_clean_event(struct r5conf
*conf
,
3439 struct stripe_head
*sh
, int disks
, struct bio_list
*return_bi
)
3443 int discard_pending
= 0;
3444 struct stripe_head
*head_sh
= sh
;
3445 bool do_endio
= false;
3447 for (i
= disks
; i
--; )
3448 if (sh
->dev
[i
].written
) {
3450 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3451 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3452 test_bit(R5_Discard
, &dev
->flags
) ||
3453 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3454 /* We can return any write requests */
3455 struct bio
*wbi
, *wbi2
;
3456 pr_debug("Return write for disc %d\n", i
);
3457 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3458 clear_bit(R5_UPTODATE
, &dev
->flags
);
3459 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3460 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3465 dev
->page
= dev
->orig_page
;
3467 dev
->written
= NULL
;
3468 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3469 dev
->sector
+ STRIPE_SECTORS
) {
3470 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3471 if (!raid5_dec_bi_active_stripes(wbi
)) {
3472 md_write_end(conf
->mddev
);
3473 bio_list_add(return_bi
, wbi
);
3477 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3479 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3481 if (head_sh
->batch_head
) {
3482 sh
= list_first_entry(&sh
->batch_list
,
3485 if (sh
!= head_sh
) {
3492 } else if (test_bit(R5_Discard
, &dev
->flags
))
3493 discard_pending
= 1;
3494 WARN_ON(test_bit(R5_SkipCopy
, &dev
->flags
));
3495 WARN_ON(dev
->page
!= dev
->orig_page
);
3497 if (!discard_pending
&&
3498 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3499 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3500 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3501 if (sh
->qd_idx
>= 0) {
3502 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3503 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3505 /* now that discard is done we can proceed with any sync */
3506 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3508 * SCSI discard will change some bio fields and the stripe has
3509 * no updated data, so remove it from hash list and the stripe
3510 * will be reinitialized
3512 spin_lock_irq(&conf
->device_lock
);
3515 if (head_sh
->batch_head
) {
3516 sh
= list_first_entry(&sh
->batch_list
,
3517 struct stripe_head
, batch_list
);
3521 spin_unlock_irq(&conf
->device_lock
);
3524 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3525 set_bit(STRIPE_HANDLE
, &sh
->state
);
3529 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3530 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3531 md_wakeup_thread(conf
->mddev
->thread
);
3533 if (head_sh
->batch_head
&& do_endio
)
3534 break_stripe_batch_list(head_sh
, STRIPE_EXPAND_SYNC_FLAGS
);
3537 static void handle_stripe_dirtying(struct r5conf
*conf
,
3538 struct stripe_head
*sh
,
3539 struct stripe_head_state
*s
,
3542 int rmw
= 0, rcw
= 0, i
;
3543 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3545 /* Check whether resync is now happening or should start.
3546 * If yes, then the array is dirty (after unclean shutdown or
3547 * initial creation), so parity in some stripes might be inconsistent.
3548 * In this case, we need to always do reconstruct-write, to ensure
3549 * that in case of drive failure or read-error correction, we
3550 * generate correct data from the parity.
3552 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3553 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3555 /* Calculate the real rcw later - for now make it
3556 * look like rcw is cheaper
3559 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3560 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3561 (unsigned long long)sh
->sector
);
3562 } else for (i
= disks
; i
--; ) {
3563 /* would I have to read this buffer for read_modify_write */
3564 struct r5dev
*dev
= &sh
->dev
[i
];
3565 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3566 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3567 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3568 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3569 if (test_bit(R5_Insync
, &dev
->flags
))
3572 rmw
+= 2*disks
; /* cannot read it */
3574 /* Would I have to read this buffer for reconstruct_write */
3575 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3576 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3577 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3578 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3579 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3580 if (test_bit(R5_Insync
, &dev
->flags
))
3586 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3587 (unsigned long long)sh
->sector
, rmw
, rcw
);
3588 set_bit(STRIPE_HANDLE
, &sh
->state
);
3589 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_ENABLE_RMW
)) && rmw
> 0) {
3590 /* prefer read-modify-write, but need to get some data */
3591 if (conf
->mddev
->queue
)
3592 blk_add_trace_msg(conf
->mddev
->queue
,
3593 "raid5 rmw %llu %d",
3594 (unsigned long long)sh
->sector
, rmw
);
3595 for (i
= disks
; i
--; ) {
3596 struct r5dev
*dev
= &sh
->dev
[i
];
3597 if ((dev
->towrite
|| i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3598 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3599 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3600 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3601 test_bit(R5_Insync
, &dev
->flags
)) {
3602 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3604 pr_debug("Read_old block %d for r-m-w\n",
3606 set_bit(R5_LOCKED
, &dev
->flags
);
3607 set_bit(R5_Wantread
, &dev
->flags
);
3610 set_bit(STRIPE_DELAYED
, &sh
->state
);
3611 set_bit(STRIPE_HANDLE
, &sh
->state
);
3616 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_ENABLE_RMW
)) && rcw
> 0) {
3617 /* want reconstruct write, but need to get some data */
3620 for (i
= disks
; i
--; ) {
3621 struct r5dev
*dev
= &sh
->dev
[i
];
3622 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3623 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3624 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3625 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3626 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3628 if (test_bit(R5_Insync
, &dev
->flags
) &&
3629 test_bit(STRIPE_PREREAD_ACTIVE
,
3631 pr_debug("Read_old block "
3632 "%d for Reconstruct\n", i
);
3633 set_bit(R5_LOCKED
, &dev
->flags
);
3634 set_bit(R5_Wantread
, &dev
->flags
);
3638 set_bit(STRIPE_DELAYED
, &sh
->state
);
3639 set_bit(STRIPE_HANDLE
, &sh
->state
);
3643 if (rcw
&& conf
->mddev
->queue
)
3644 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3645 (unsigned long long)sh
->sector
,
3646 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3649 if (rcw
> disks
&& rmw
> disks
&&
3650 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3651 set_bit(STRIPE_DELAYED
, &sh
->state
);
3653 /* now if nothing is locked, and if we have enough data,
3654 * we can start a write request
3656 /* since handle_stripe can be called at any time we need to handle the
3657 * case where a compute block operation has been submitted and then a
3658 * subsequent call wants to start a write request. raid_run_ops only
3659 * handles the case where compute block and reconstruct are requested
3660 * simultaneously. If this is not the case then new writes need to be
3661 * held off until the compute completes.
3663 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
3664 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
3665 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
3666 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
3669 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
3670 struct stripe_head_state
*s
, int disks
)
3672 struct r5dev
*dev
= NULL
;
3674 BUG_ON(sh
->batch_head
);
3675 set_bit(STRIPE_HANDLE
, &sh
->state
);
3677 switch (sh
->check_state
) {
3678 case check_state_idle
:
3679 /* start a new check operation if there are no failures */
3680 if (s
->failed
== 0) {
3681 BUG_ON(s
->uptodate
!= disks
);
3682 sh
->check_state
= check_state_run
;
3683 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3684 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3688 dev
= &sh
->dev
[s
->failed_num
[0]];
3690 case check_state_compute_result
:
3691 sh
->check_state
= check_state_idle
;
3693 dev
= &sh
->dev
[sh
->pd_idx
];
3695 /* check that a write has not made the stripe insync */
3696 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3699 /* either failed parity check, or recovery is happening */
3700 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3701 BUG_ON(s
->uptodate
!= disks
);
3703 set_bit(R5_LOCKED
, &dev
->flags
);
3705 set_bit(R5_Wantwrite
, &dev
->flags
);
3707 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3708 set_bit(STRIPE_INSYNC
, &sh
->state
);
3710 case check_state_run
:
3711 break; /* we will be called again upon completion */
3712 case check_state_check_result
:
3713 sh
->check_state
= check_state_idle
;
3715 /* if a failure occurred during the check operation, leave
3716 * STRIPE_INSYNC not set and let the stripe be handled again
3721 /* handle a successful check operation, if parity is correct
3722 * we are done. Otherwise update the mismatch count and repair
3723 * parity if !MD_RECOVERY_CHECK
3725 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3726 /* parity is correct (on disc,
3727 * not in buffer any more)
3729 set_bit(STRIPE_INSYNC
, &sh
->state
);
3731 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3732 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3733 /* don't try to repair!! */
3734 set_bit(STRIPE_INSYNC
, &sh
->state
);
3736 sh
->check_state
= check_state_compute_run
;
3737 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3738 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3739 set_bit(R5_Wantcompute
,
3740 &sh
->dev
[sh
->pd_idx
].flags
);
3741 sh
->ops
.target
= sh
->pd_idx
;
3742 sh
->ops
.target2
= -1;
3747 case check_state_compute_run
:
3750 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3751 __func__
, sh
->check_state
,
3752 (unsigned long long) sh
->sector
);
3757 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3758 struct stripe_head_state
*s
,
3761 int pd_idx
= sh
->pd_idx
;
3762 int qd_idx
= sh
->qd_idx
;
3765 BUG_ON(sh
->batch_head
);
3766 set_bit(STRIPE_HANDLE
, &sh
->state
);
3768 BUG_ON(s
->failed
> 2);
3770 /* Want to check and possibly repair P and Q.
3771 * However there could be one 'failed' device, in which
3772 * case we can only check one of them, possibly using the
3773 * other to generate missing data
3776 switch (sh
->check_state
) {
3777 case check_state_idle
:
3778 /* start a new check operation if there are < 2 failures */
3779 if (s
->failed
== s
->q_failed
) {
3780 /* The only possible failed device holds Q, so it
3781 * makes sense to check P (If anything else were failed,
3782 * we would have used P to recreate it).
3784 sh
->check_state
= check_state_run
;
3786 if (!s
->q_failed
&& s
->failed
< 2) {
3787 /* Q is not failed, and we didn't use it to generate
3788 * anything, so it makes sense to check it
3790 if (sh
->check_state
== check_state_run
)
3791 sh
->check_state
= check_state_run_pq
;
3793 sh
->check_state
= check_state_run_q
;
3796 /* discard potentially stale zero_sum_result */
3797 sh
->ops
.zero_sum_result
= 0;
3799 if (sh
->check_state
== check_state_run
) {
3800 /* async_xor_zero_sum destroys the contents of P */
3801 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3804 if (sh
->check_state
>= check_state_run
&&
3805 sh
->check_state
<= check_state_run_pq
) {
3806 /* async_syndrome_zero_sum preserves P and Q, so
3807 * no need to mark them !uptodate here
3809 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3813 /* we have 2-disk failure */
3814 BUG_ON(s
->failed
!= 2);
3816 case check_state_compute_result
:
3817 sh
->check_state
= check_state_idle
;
3819 /* check that a write has not made the stripe insync */
3820 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3823 /* now write out any block on a failed drive,
3824 * or P or Q if they were recomputed
3826 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3827 if (s
->failed
== 2) {
3828 dev
= &sh
->dev
[s
->failed_num
[1]];
3830 set_bit(R5_LOCKED
, &dev
->flags
);
3831 set_bit(R5_Wantwrite
, &dev
->flags
);
3833 if (s
->failed
>= 1) {
3834 dev
= &sh
->dev
[s
->failed_num
[0]];
3836 set_bit(R5_LOCKED
, &dev
->flags
);
3837 set_bit(R5_Wantwrite
, &dev
->flags
);
3839 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3840 dev
= &sh
->dev
[pd_idx
];
3842 set_bit(R5_LOCKED
, &dev
->flags
);
3843 set_bit(R5_Wantwrite
, &dev
->flags
);
3845 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3846 dev
= &sh
->dev
[qd_idx
];
3848 set_bit(R5_LOCKED
, &dev
->flags
);
3849 set_bit(R5_Wantwrite
, &dev
->flags
);
3851 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3853 set_bit(STRIPE_INSYNC
, &sh
->state
);
3855 case check_state_run
:
3856 case check_state_run_q
:
3857 case check_state_run_pq
:
3858 break; /* we will be called again upon completion */
3859 case check_state_check_result
:
3860 sh
->check_state
= check_state_idle
;
3862 /* handle a successful check operation, if parity is correct
3863 * we are done. Otherwise update the mismatch count and repair
3864 * parity if !MD_RECOVERY_CHECK
3866 if (sh
->ops
.zero_sum_result
== 0) {
3867 /* both parities are correct */
3869 set_bit(STRIPE_INSYNC
, &sh
->state
);
3871 /* in contrast to the raid5 case we can validate
3872 * parity, but still have a failure to write
3875 sh
->check_state
= check_state_compute_result
;
3876 /* Returning at this point means that we may go
3877 * off and bring p and/or q uptodate again so
3878 * we make sure to check zero_sum_result again
3879 * to verify if p or q need writeback
3883 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3884 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3885 /* don't try to repair!! */
3886 set_bit(STRIPE_INSYNC
, &sh
->state
);
3888 int *target
= &sh
->ops
.target
;
3890 sh
->ops
.target
= -1;
3891 sh
->ops
.target2
= -1;
3892 sh
->check_state
= check_state_compute_run
;
3893 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3894 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3895 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3896 set_bit(R5_Wantcompute
,
3897 &sh
->dev
[pd_idx
].flags
);
3899 target
= &sh
->ops
.target2
;
3902 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3903 set_bit(R5_Wantcompute
,
3904 &sh
->dev
[qd_idx
].flags
);
3911 case check_state_compute_run
:
3914 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3915 __func__
, sh
->check_state
,
3916 (unsigned long long) sh
->sector
);
3921 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3925 /* We have read all the blocks in this stripe and now we need to
3926 * copy some of them into a target stripe for expand.
3928 struct dma_async_tx_descriptor
*tx
= NULL
;
3929 BUG_ON(sh
->batch_head
);
3930 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3931 for (i
= 0; i
< sh
->disks
; i
++)
3932 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3934 struct stripe_head
*sh2
;
3935 struct async_submit_ctl submit
;
3937 sector_t bn
= compute_blocknr(sh
, i
, 1);
3938 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3940 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3942 /* so far only the early blocks of this stripe
3943 * have been requested. When later blocks
3944 * get requested, we will try again
3947 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3948 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3949 /* must have already done this block */
3950 release_stripe(sh2
);
3954 /* place all the copies on one channel */
3955 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3956 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3957 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3960 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3961 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3962 for (j
= 0; j
< conf
->raid_disks
; j
++)
3963 if (j
!= sh2
->pd_idx
&&
3965 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3967 if (j
== conf
->raid_disks
) {
3968 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3969 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3971 release_stripe(sh2
);
3974 /* done submitting copies, wait for them to complete */
3975 async_tx_quiesce(&tx
);
3979 * handle_stripe - do things to a stripe.
3981 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3982 * state of various bits to see what needs to be done.
3984 * return some read requests which now have data
3985 * return some write requests which are safely on storage
3986 * schedule a read on some buffers
3987 * schedule a write of some buffers
3988 * return confirmation of parity correctness
3992 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3994 struct r5conf
*conf
= sh
->raid_conf
;
3995 int disks
= sh
->disks
;
3998 int do_recovery
= 0;
4000 memset(s
, 0, sizeof(*s
));
4002 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
4003 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
4004 s
->failed_num
[0] = -1;
4005 s
->failed_num
[1] = -1;
4007 /* Now to look around and see what can be done */
4009 for (i
=disks
; i
--; ) {
4010 struct md_rdev
*rdev
;
4017 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4019 dev
->toread
, dev
->towrite
, dev
->written
);
4020 /* maybe we can reply to a read
4022 * new wantfill requests are only permitted while
4023 * ops_complete_biofill is guaranteed to be inactive
4025 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4026 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4027 set_bit(R5_Wantfill
, &dev
->flags
);
4029 /* now count some things */
4030 if (test_bit(R5_LOCKED
, &dev
->flags
))
4032 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4034 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4036 BUG_ON(s
->compute
> 2);
4039 if (test_bit(R5_Wantfill
, &dev
->flags
))
4041 else if (dev
->toread
)
4045 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4050 /* Prefer to use the replacement for reads, but only
4051 * if it is recovered enough and has no bad blocks.
4053 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4054 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4055 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4056 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4057 &first_bad
, &bad_sectors
))
4058 set_bit(R5_ReadRepl
, &dev
->flags
);
4060 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4061 set_bit(R5_NeedReplace
, &dev
->flags
);
4063 clear_bit(R5_NeedReplace
, &dev
->flags
);
4064 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4065 clear_bit(R5_ReadRepl
, &dev
->flags
);
4067 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4070 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4071 &first_bad
, &bad_sectors
);
4072 if (s
->blocked_rdev
== NULL
4073 && (test_bit(Blocked
, &rdev
->flags
)
4076 set_bit(BlockedBadBlocks
,
4078 s
->blocked_rdev
= rdev
;
4079 atomic_inc(&rdev
->nr_pending
);
4082 clear_bit(R5_Insync
, &dev
->flags
);
4086 /* also not in-sync */
4087 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4088 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4089 /* treat as in-sync, but with a read error
4090 * which we can now try to correct
4092 set_bit(R5_Insync
, &dev
->flags
);
4093 set_bit(R5_ReadError
, &dev
->flags
);
4095 } else if (test_bit(In_sync
, &rdev
->flags
))
4096 set_bit(R5_Insync
, &dev
->flags
);
4097 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4098 /* in sync if before recovery_offset */
4099 set_bit(R5_Insync
, &dev
->flags
);
4100 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4101 test_bit(R5_Expanded
, &dev
->flags
))
4102 /* If we've reshaped into here, we assume it is Insync.
4103 * We will shortly update recovery_offset to make
4106 set_bit(R5_Insync
, &dev
->flags
);
4108 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4109 /* This flag does not apply to '.replacement'
4110 * only to .rdev, so make sure to check that*/
4111 struct md_rdev
*rdev2
= rcu_dereference(
4112 conf
->disks
[i
].rdev
);
4114 clear_bit(R5_Insync
, &dev
->flags
);
4115 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4116 s
->handle_bad_blocks
= 1;
4117 atomic_inc(&rdev2
->nr_pending
);
4119 clear_bit(R5_WriteError
, &dev
->flags
);
4121 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4122 /* This flag does not apply to '.replacement'
4123 * only to .rdev, so make sure to check that*/
4124 struct md_rdev
*rdev2
= rcu_dereference(
4125 conf
->disks
[i
].rdev
);
4126 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4127 s
->handle_bad_blocks
= 1;
4128 atomic_inc(&rdev2
->nr_pending
);
4130 clear_bit(R5_MadeGood
, &dev
->flags
);
4132 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4133 struct md_rdev
*rdev2
= rcu_dereference(
4134 conf
->disks
[i
].replacement
);
4135 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4136 s
->handle_bad_blocks
= 1;
4137 atomic_inc(&rdev2
->nr_pending
);
4139 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4141 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4142 /* The ReadError flag will just be confusing now */
4143 clear_bit(R5_ReadError
, &dev
->flags
);
4144 clear_bit(R5_ReWrite
, &dev
->flags
);
4146 if (test_bit(R5_ReadError
, &dev
->flags
))
4147 clear_bit(R5_Insync
, &dev
->flags
);
4148 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4150 s
->failed_num
[s
->failed
] = i
;
4152 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4156 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4157 /* If there is a failed device being replaced,
4158 * we must be recovering.
4159 * else if we are after recovery_cp, we must be syncing
4160 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4161 * else we can only be replacing
4162 * sync and recovery both need to read all devices, and so
4163 * use the same flag.
4166 sh
->sector
>= conf
->mddev
->recovery_cp
||
4167 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4175 static int clear_batch_ready(struct stripe_head
*sh
)
4177 /* Return '1' if this is a member of batch, or
4178 * '0' if it is a lone stripe or a head which can now be
4181 struct stripe_head
*tmp
;
4182 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4183 return (sh
->batch_head
&& sh
->batch_head
!= sh
);
4184 spin_lock(&sh
->stripe_lock
);
4185 if (!sh
->batch_head
) {
4186 spin_unlock(&sh
->stripe_lock
);
4191 * this stripe could be added to a batch list before we check
4192 * BATCH_READY, skips it
4194 if (sh
->batch_head
!= sh
) {
4195 spin_unlock(&sh
->stripe_lock
);
4198 spin_lock(&sh
->batch_lock
);
4199 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4200 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4201 spin_unlock(&sh
->batch_lock
);
4202 spin_unlock(&sh
->stripe_lock
);
4205 * BATCH_READY is cleared, no new stripes can be added.
4206 * batch_list can be accessed without lock
4211 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4212 unsigned long handle_flags
)
4214 struct stripe_head
*sh
, *next
;
4218 list_for_each_entry_safe(sh
, next
, &head_sh
->batch_list
, batch_list
) {
4220 list_del_init(&sh
->batch_list
);
4222 WARN_ON_ONCE(sh
->state
& ((1 << STRIPE_ACTIVE
) |
4223 (1 << STRIPE_SYNCING
) |
4224 (1 << STRIPE_REPLACED
) |
4225 (1 << STRIPE_PREREAD_ACTIVE
) |
4226 (1 << STRIPE_DELAYED
) |
4227 (1 << STRIPE_BIT_DELAY
) |
4228 (1 << STRIPE_FULL_WRITE
) |
4229 (1 << STRIPE_BIOFILL_RUN
) |
4230 (1 << STRIPE_COMPUTE_RUN
) |
4231 (1 << STRIPE_OPS_REQ_PENDING
) |
4232 (1 << STRIPE_DISCARD
) |
4233 (1 << STRIPE_BATCH_READY
) |
4234 (1 << STRIPE_BATCH_ERR
) |
4235 (1 << STRIPE_BITMAP_PENDING
)));
4236 WARN_ON_ONCE(head_sh
->state
& ((1 << STRIPE_DISCARD
) |
4237 (1 << STRIPE_REPLACED
)));
4239 set_mask_bits(&sh
->state
, ~(STRIPE_EXPAND_SYNC_FLAGS
|
4240 (1 << STRIPE_DEGRADED
)),
4241 head_sh
->state
& (1 << STRIPE_INSYNC
));
4243 sh
->check_state
= head_sh
->check_state
;
4244 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4245 for (i
= 0; i
< sh
->disks
; i
++) {
4246 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
4248 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4249 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4251 spin_lock_irq(&sh
->stripe_lock
);
4252 sh
->batch_head
= NULL
;
4253 spin_unlock_irq(&sh
->stripe_lock
);
4254 if (handle_flags
== 0 ||
4255 sh
->state
& handle_flags
)
4256 set_bit(STRIPE_HANDLE
, &sh
->state
);
4259 spin_lock_irq(&head_sh
->stripe_lock
);
4260 head_sh
->batch_head
= NULL
;
4261 spin_unlock_irq(&head_sh
->stripe_lock
);
4262 for (i
= 0; i
< head_sh
->disks
; i
++)
4263 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
4265 if (head_sh
->state
& handle_flags
)
4266 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
4269 wake_up(&head_sh
->raid_conf
->wait_for_overlap
);
4272 static void handle_stripe(struct stripe_head
*sh
)
4274 struct stripe_head_state s
;
4275 struct r5conf
*conf
= sh
->raid_conf
;
4278 int disks
= sh
->disks
;
4279 struct r5dev
*pdev
, *qdev
;
4281 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4282 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4283 /* already being handled, ensure it gets handled
4284 * again when current action finishes */
4285 set_bit(STRIPE_HANDLE
, &sh
->state
);
4289 if (clear_batch_ready(sh
) ) {
4290 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4294 if (test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4295 break_stripe_batch_list(sh
, 0);
4297 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4298 spin_lock(&sh
->stripe_lock
);
4299 /* Cannot process 'sync' concurrently with 'discard' */
4300 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4301 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4302 set_bit(STRIPE_SYNCING
, &sh
->state
);
4303 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4304 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4306 spin_unlock(&sh
->stripe_lock
);
4308 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4310 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4311 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4312 (unsigned long long)sh
->sector
, sh
->state
,
4313 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4314 sh
->check_state
, sh
->reconstruct_state
);
4316 analyse_stripe(sh
, &s
);
4318 if (s
.handle_bad_blocks
) {
4319 set_bit(STRIPE_HANDLE
, &sh
->state
);
4323 if (unlikely(s
.blocked_rdev
)) {
4324 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4325 s
.replacing
|| s
.to_write
|| s
.written
) {
4326 set_bit(STRIPE_HANDLE
, &sh
->state
);
4329 /* There is nothing for the blocked_rdev to block */
4330 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4331 s
.blocked_rdev
= NULL
;
4334 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4335 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4336 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4339 pr_debug("locked=%d uptodate=%d to_read=%d"
4340 " to_write=%d failed=%d failed_num=%d,%d\n",
4341 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4342 s
.failed_num
[0], s
.failed_num
[1]);
4343 /* check if the array has lost more than max_degraded devices and,
4344 * if so, some requests might need to be failed.
4346 if (s
.failed
> conf
->max_degraded
) {
4347 sh
->check_state
= 0;
4348 sh
->reconstruct_state
= 0;
4349 break_stripe_batch_list(sh
, 0);
4350 if (s
.to_read
+s
.to_write
+s
.written
)
4351 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
4352 if (s
.syncing
+ s
.replacing
)
4353 handle_failed_sync(conf
, sh
, &s
);
4356 /* Now we check to see if any write operations have recently
4360 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4362 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4363 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4364 sh
->reconstruct_state
= reconstruct_state_idle
;
4366 /* All the 'written' buffers and the parity block are ready to
4367 * be written back to disk
4369 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4370 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4371 BUG_ON(sh
->qd_idx
>= 0 &&
4372 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4373 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4374 for (i
= disks
; i
--; ) {
4375 struct r5dev
*dev
= &sh
->dev
[i
];
4376 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4377 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4379 pr_debug("Writing block %d\n", i
);
4380 set_bit(R5_Wantwrite
, &dev
->flags
);
4385 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4386 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4388 set_bit(STRIPE_INSYNC
, &sh
->state
);
4391 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4392 s
.dec_preread_active
= 1;
4396 * might be able to return some write requests if the parity blocks
4397 * are safe, or on a failed drive
4399 pdev
= &sh
->dev
[sh
->pd_idx
];
4400 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4401 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4402 qdev
= &sh
->dev
[sh
->qd_idx
];
4403 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4404 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4408 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4409 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4410 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4411 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4412 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4413 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4414 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4415 test_bit(R5_Discard
, &qdev
->flags
))))))
4416 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
4418 /* Now we might consider reading some blocks, either to check/generate
4419 * parity, or to satisfy requests
4420 * or to load a block that is being partially written.
4422 if (s
.to_read
|| s
.non_overwrite
4423 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4424 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4427 handle_stripe_fill(sh
, &s
, disks
);
4429 /* Now to consider new write requests and what else, if anything
4430 * should be read. We do not handle new writes when:
4431 * 1/ A 'write' operation (copy+xor) is already in flight.
4432 * 2/ A 'check' operation is in flight, as it may clobber the parity
4435 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
4436 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4438 /* maybe we need to check and possibly fix the parity for this stripe
4439 * Any reads will already have been scheduled, so we just see if enough
4440 * data is available. The parity check is held off while parity
4441 * dependent operations are in flight.
4443 if (sh
->check_state
||
4444 (s
.syncing
&& s
.locked
== 0 &&
4445 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4446 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4447 if (conf
->level
== 6)
4448 handle_parity_checks6(conf
, sh
, &s
, disks
);
4450 handle_parity_checks5(conf
, sh
, &s
, disks
);
4453 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4454 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4455 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4456 /* Write out to replacement devices where possible */
4457 for (i
= 0; i
< conf
->raid_disks
; i
++)
4458 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4459 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4460 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4461 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4465 set_bit(STRIPE_INSYNC
, &sh
->state
);
4466 set_bit(STRIPE_REPLACED
, &sh
->state
);
4468 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4469 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4470 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4471 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4472 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4473 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4474 wake_up(&conf
->wait_for_overlap
);
4477 /* If the failed drives are just a ReadError, then we might need
4478 * to progress the repair/check process
4480 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4481 for (i
= 0; i
< s
.failed
; i
++) {
4482 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4483 if (test_bit(R5_ReadError
, &dev
->flags
)
4484 && !test_bit(R5_LOCKED
, &dev
->flags
)
4485 && test_bit(R5_UPTODATE
, &dev
->flags
)
4487 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4488 set_bit(R5_Wantwrite
, &dev
->flags
);
4489 set_bit(R5_ReWrite
, &dev
->flags
);
4490 set_bit(R5_LOCKED
, &dev
->flags
);
4493 /* let's read it back */
4494 set_bit(R5_Wantread
, &dev
->flags
);
4495 set_bit(R5_LOCKED
, &dev
->flags
);
4501 /* Finish reconstruct operations initiated by the expansion process */
4502 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4503 struct stripe_head
*sh_src
4504 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4505 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4506 /* sh cannot be written until sh_src has been read.
4507 * so arrange for sh to be delayed a little
4509 set_bit(STRIPE_DELAYED
, &sh
->state
);
4510 set_bit(STRIPE_HANDLE
, &sh
->state
);
4511 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4513 atomic_inc(&conf
->preread_active_stripes
);
4514 release_stripe(sh_src
);
4518 release_stripe(sh_src
);
4520 sh
->reconstruct_state
= reconstruct_state_idle
;
4521 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4522 for (i
= conf
->raid_disks
; i
--; ) {
4523 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4524 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4529 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4530 !sh
->reconstruct_state
) {
4531 /* Need to write out all blocks after computing parity */
4532 sh
->disks
= conf
->raid_disks
;
4533 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4534 schedule_reconstruction(sh
, &s
, 1, 1);
4535 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4536 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4537 atomic_dec(&conf
->reshape_stripes
);
4538 wake_up(&conf
->wait_for_overlap
);
4539 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4542 if (s
.expanding
&& s
.locked
== 0 &&
4543 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4544 handle_stripe_expansion(conf
, sh
);
4547 /* wait for this device to become unblocked */
4548 if (unlikely(s
.blocked_rdev
)) {
4549 if (conf
->mddev
->external
)
4550 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4553 /* Internal metadata will immediately
4554 * be written by raid5d, so we don't
4555 * need to wait here.
4557 rdev_dec_pending(s
.blocked_rdev
,
4561 if (s
.handle_bad_blocks
)
4562 for (i
= disks
; i
--; ) {
4563 struct md_rdev
*rdev
;
4564 struct r5dev
*dev
= &sh
->dev
[i
];
4565 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4566 /* We own a safe reference to the rdev */
4567 rdev
= conf
->disks
[i
].rdev
;
4568 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4570 md_error(conf
->mddev
, rdev
);
4571 rdev_dec_pending(rdev
, conf
->mddev
);
4573 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4574 rdev
= conf
->disks
[i
].rdev
;
4575 rdev_clear_badblocks(rdev
, sh
->sector
,
4577 rdev_dec_pending(rdev
, conf
->mddev
);
4579 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4580 rdev
= conf
->disks
[i
].replacement
;
4582 /* rdev have been moved down */
4583 rdev
= conf
->disks
[i
].rdev
;
4584 rdev_clear_badblocks(rdev
, sh
->sector
,
4586 rdev_dec_pending(rdev
, conf
->mddev
);
4591 raid_run_ops(sh
, s
.ops_request
);
4595 if (s
.dec_preread_active
) {
4596 /* We delay this until after ops_run_io so that if make_request
4597 * is waiting on a flush, it won't continue until the writes
4598 * have actually been submitted.
4600 atomic_dec(&conf
->preread_active_stripes
);
4601 if (atomic_read(&conf
->preread_active_stripes
) <
4603 md_wakeup_thread(conf
->mddev
->thread
);
4606 if (!bio_list_empty(&s
.return_bi
)) {
4607 if (test_bit(MD_CHANGE_PENDING
, &conf
->mddev
->flags
)) {
4608 spin_lock_irq(&conf
->device_lock
);
4609 bio_list_merge(&conf
->return_bi
, &s
.return_bi
);
4610 spin_unlock_irq(&conf
->device_lock
);
4611 md_wakeup_thread(conf
->mddev
->thread
);
4613 return_io(&s
.return_bi
);
4616 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4619 static void raid5_activate_delayed(struct r5conf
*conf
)
4621 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
4622 while (!list_empty(&conf
->delayed_list
)) {
4623 struct list_head
*l
= conf
->delayed_list
.next
;
4624 struct stripe_head
*sh
;
4625 sh
= list_entry(l
, struct stripe_head
, lru
);
4627 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4628 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4629 atomic_inc(&conf
->preread_active_stripes
);
4630 list_add_tail(&sh
->lru
, &conf
->hold_list
);
4631 raid5_wakeup_stripe_thread(sh
);
4636 static void activate_bit_delay(struct r5conf
*conf
,
4637 struct list_head
*temp_inactive_list
)
4639 /* device_lock is held */
4640 struct list_head head
;
4641 list_add(&head
, &conf
->bitmap_list
);
4642 list_del_init(&conf
->bitmap_list
);
4643 while (!list_empty(&head
)) {
4644 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
4646 list_del_init(&sh
->lru
);
4647 atomic_inc(&sh
->count
);
4648 hash
= sh
->hash_lock_index
;
4649 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
4653 static int raid5_congested(struct mddev
*mddev
, int bits
)
4655 struct r5conf
*conf
= mddev
->private;
4657 /* No difference between reads and writes. Just check
4658 * how busy the stripe_cache is
4661 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
4665 if (atomic_read(&conf
->empty_inactive_list_nr
))
4671 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
4673 struct r5conf
*conf
= mddev
->private;
4674 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
4675 unsigned int chunk_sectors
;
4676 unsigned int bio_sectors
= bio_sectors(bio
);
4678 chunk_sectors
= min(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
4679 return chunk_sectors
>=
4680 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
4684 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4685 * later sampled by raid5d.
4687 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
4689 unsigned long flags
;
4691 spin_lock_irqsave(&conf
->device_lock
, flags
);
4693 bi
->bi_next
= conf
->retry_read_aligned_list
;
4694 conf
->retry_read_aligned_list
= bi
;
4696 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4697 md_wakeup_thread(conf
->mddev
->thread
);
4700 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
4704 bi
= conf
->retry_read_aligned
;
4706 conf
->retry_read_aligned
= NULL
;
4709 bi
= conf
->retry_read_aligned_list
;
4711 conf
->retry_read_aligned_list
= bi
->bi_next
;
4714 * this sets the active strip count to 1 and the processed
4715 * strip count to zero (upper 8 bits)
4717 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
4724 * The "raid5_align_endio" should check if the read succeeded and if it
4725 * did, call bio_endio on the original bio (having bio_put the new bio
4727 * If the read failed..
4729 static void raid5_align_endio(struct bio
*bi
)
4731 struct bio
* raid_bi
= bi
->bi_private
;
4732 struct mddev
*mddev
;
4733 struct r5conf
*conf
;
4734 struct md_rdev
*rdev
;
4735 int error
= bi
->bi_error
;
4739 rdev
= (void*)raid_bi
->bi_next
;
4740 raid_bi
->bi_next
= NULL
;
4741 mddev
= rdev
->mddev
;
4742 conf
= mddev
->private;
4744 rdev_dec_pending(rdev
, conf
->mddev
);
4747 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
4750 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4751 wake_up(&conf
->wait_for_quiescent
);
4755 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4757 add_bio_to_retry(raid_bi
, conf
);
4760 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
4762 struct r5conf
*conf
= mddev
->private;
4764 struct bio
* align_bi
;
4765 struct md_rdev
*rdev
;
4766 sector_t end_sector
;
4768 if (!in_chunk_boundary(mddev
, raid_bio
)) {
4769 pr_debug("%s: non aligned\n", __func__
);
4773 * use bio_clone_mddev to make a copy of the bio
4775 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
4779 * set bi_end_io to a new function, and set bi_private to the
4782 align_bi
->bi_end_io
= raid5_align_endio
;
4783 align_bi
->bi_private
= raid_bio
;
4787 align_bi
->bi_iter
.bi_sector
=
4788 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
4791 end_sector
= bio_end_sector(align_bi
);
4793 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
4794 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
4795 rdev
->recovery_offset
< end_sector
) {
4796 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
4798 (test_bit(Faulty
, &rdev
->flags
) ||
4799 !(test_bit(In_sync
, &rdev
->flags
) ||
4800 rdev
->recovery_offset
>= end_sector
)))
4807 atomic_inc(&rdev
->nr_pending
);
4809 raid_bio
->bi_next
= (void*)rdev
;
4810 align_bi
->bi_bdev
= rdev
->bdev
;
4811 bio_clear_flag(align_bi
, BIO_SEG_VALID
);
4813 if (is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
4814 bio_sectors(align_bi
),
4815 &first_bad
, &bad_sectors
)) {
4817 rdev_dec_pending(rdev
, mddev
);
4821 /* No reshape active, so we can trust rdev->data_offset */
4822 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
4824 spin_lock_irq(&conf
->device_lock
);
4825 wait_event_lock_irq(conf
->wait_for_quiescent
,
4828 atomic_inc(&conf
->active_aligned_reads
);
4829 spin_unlock_irq(&conf
->device_lock
);
4832 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4833 align_bi
, disk_devt(mddev
->gendisk
),
4834 raid_bio
->bi_iter
.bi_sector
);
4835 generic_make_request(align_bi
);
4844 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
4849 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
4850 unsigned chunk_sects
= mddev
->chunk_sectors
;
4851 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
4853 if (sectors
< bio_sectors(raid_bio
)) {
4854 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, fs_bio_set
);
4855 bio_chain(split
, raid_bio
);
4859 if (!raid5_read_one_chunk(mddev
, split
)) {
4860 if (split
!= raid_bio
)
4861 generic_make_request(raid_bio
);
4864 } while (split
!= raid_bio
);
4869 /* __get_priority_stripe - get the next stripe to process
4871 * Full stripe writes are allowed to pass preread active stripes up until
4872 * the bypass_threshold is exceeded. In general the bypass_count
4873 * increments when the handle_list is handled before the hold_list; however, it
4874 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4875 * stripe with in flight i/o. The bypass_count will be reset when the
4876 * head of the hold_list has changed, i.e. the head was promoted to the
4879 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
4881 struct stripe_head
*sh
= NULL
, *tmp
;
4882 struct list_head
*handle_list
= NULL
;
4883 struct r5worker_group
*wg
= NULL
;
4885 if (conf
->worker_cnt_per_group
== 0) {
4886 handle_list
= &conf
->handle_list
;
4887 } else if (group
!= ANY_GROUP
) {
4888 handle_list
= &conf
->worker_groups
[group
].handle_list
;
4889 wg
= &conf
->worker_groups
[group
];
4892 for (i
= 0; i
< conf
->group_cnt
; i
++) {
4893 handle_list
= &conf
->worker_groups
[i
].handle_list
;
4894 wg
= &conf
->worker_groups
[i
];
4895 if (!list_empty(handle_list
))
4900 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4902 list_empty(handle_list
) ? "empty" : "busy",
4903 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4904 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4906 if (!list_empty(handle_list
)) {
4907 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
4909 if (list_empty(&conf
->hold_list
))
4910 conf
->bypass_count
= 0;
4911 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4912 if (conf
->hold_list
.next
== conf
->last_hold
)
4913 conf
->bypass_count
++;
4915 conf
->last_hold
= conf
->hold_list
.next
;
4916 conf
->bypass_count
-= conf
->bypass_threshold
;
4917 if (conf
->bypass_count
< 0)
4918 conf
->bypass_count
= 0;
4921 } else if (!list_empty(&conf
->hold_list
) &&
4922 ((conf
->bypass_threshold
&&
4923 conf
->bypass_count
> conf
->bypass_threshold
) ||
4924 atomic_read(&conf
->pending_full_writes
) == 0)) {
4926 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
4927 if (conf
->worker_cnt_per_group
== 0 ||
4928 group
== ANY_GROUP
||
4929 !cpu_online(tmp
->cpu
) ||
4930 cpu_to_group(tmp
->cpu
) == group
) {
4937 conf
->bypass_count
-= conf
->bypass_threshold
;
4938 if (conf
->bypass_count
< 0)
4939 conf
->bypass_count
= 0;
4951 list_del_init(&sh
->lru
);
4952 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
4956 struct raid5_plug_cb
{
4957 struct blk_plug_cb cb
;
4958 struct list_head list
;
4959 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
4962 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4964 struct raid5_plug_cb
*cb
= container_of(
4965 blk_cb
, struct raid5_plug_cb
, cb
);
4966 struct stripe_head
*sh
;
4967 struct mddev
*mddev
= cb
->cb
.data
;
4968 struct r5conf
*conf
= mddev
->private;
4972 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4973 spin_lock_irq(&conf
->device_lock
);
4974 while (!list_empty(&cb
->list
)) {
4975 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4976 list_del_init(&sh
->lru
);
4978 * avoid race release_stripe_plug() sees
4979 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4980 * is still in our list
4982 smp_mb__before_atomic();
4983 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4985 * STRIPE_ON_RELEASE_LIST could be set here. In that
4986 * case, the count is always > 1 here
4988 hash
= sh
->hash_lock_index
;
4989 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
4992 spin_unlock_irq(&conf
->device_lock
);
4994 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
4995 NR_STRIPE_HASH_LOCKS
);
4997 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5001 static void release_stripe_plug(struct mddev
*mddev
,
5002 struct stripe_head
*sh
)
5004 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5005 raid5_unplug
, mddev
,
5006 sizeof(struct raid5_plug_cb
));
5007 struct raid5_plug_cb
*cb
;
5014 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5016 if (cb
->list
.next
== NULL
) {
5018 INIT_LIST_HEAD(&cb
->list
);
5019 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5020 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5023 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5024 list_add_tail(&sh
->lru
, &cb
->list
);
5029 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5031 struct r5conf
*conf
= mddev
->private;
5032 sector_t logical_sector
, last_sector
;
5033 struct stripe_head
*sh
;
5037 if (mddev
->reshape_position
!= MaxSector
)
5038 /* Skip discard while reshape is happening */
5041 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5042 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5045 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5047 stripe_sectors
= conf
->chunk_sectors
*
5048 (conf
->raid_disks
- conf
->max_degraded
);
5049 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5051 sector_div(last_sector
, stripe_sectors
);
5053 logical_sector
*= conf
->chunk_sectors
;
5054 last_sector
*= conf
->chunk_sectors
;
5056 for (; logical_sector
< last_sector
;
5057 logical_sector
+= STRIPE_SECTORS
) {
5061 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5062 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5063 TASK_UNINTERRUPTIBLE
);
5064 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5065 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5070 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5071 spin_lock_irq(&sh
->stripe_lock
);
5072 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5073 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5075 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5076 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5077 spin_unlock_irq(&sh
->stripe_lock
);
5083 set_bit(STRIPE_DISCARD
, &sh
->state
);
5084 finish_wait(&conf
->wait_for_overlap
, &w
);
5085 sh
->overwrite_disks
= 0;
5086 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5087 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5089 sh
->dev
[d
].towrite
= bi
;
5090 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5091 raid5_inc_bi_active_stripes(bi
);
5092 sh
->overwrite_disks
++;
5094 spin_unlock_irq(&sh
->stripe_lock
);
5095 if (conf
->mddev
->bitmap
) {
5097 d
< conf
->raid_disks
- conf
->max_degraded
;
5099 bitmap_startwrite(mddev
->bitmap
,
5103 sh
->bm_seq
= conf
->seq_flush
+ 1;
5104 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5107 set_bit(STRIPE_HANDLE
, &sh
->state
);
5108 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5109 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5110 atomic_inc(&conf
->preread_active_stripes
);
5111 release_stripe_plug(mddev
, sh
);
5114 remaining
= raid5_dec_bi_active_stripes(bi
);
5115 if (remaining
== 0) {
5116 md_write_end(mddev
);
5121 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
5123 struct r5conf
*conf
= mddev
->private;
5125 sector_t new_sector
;
5126 sector_t logical_sector
, last_sector
;
5127 struct stripe_head
*sh
;
5128 const int rw
= bio_data_dir(bi
);
5133 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
5134 md_flush_request(mddev
, bi
);
5138 md_write_start(mddev
, bi
);
5141 * If array is degraded, better not do chunk aligned read because
5142 * later we might have to read it again in order to reconstruct
5143 * data on failed drives.
5145 if (rw
== READ
&& mddev
->degraded
== 0 &&
5146 mddev
->reshape_position
== MaxSector
) {
5147 bi
= chunk_aligned_read(mddev
, bi
);
5152 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
5153 make_discard_request(mddev
, bi
);
5157 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5158 last_sector
= bio_end_sector(bi
);
5160 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
5162 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5163 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5169 seq
= read_seqcount_begin(&conf
->gen_lock
);
5172 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5173 TASK_UNINTERRUPTIBLE
);
5174 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5175 /* spinlock is needed as reshape_progress may be
5176 * 64bit on a 32bit platform, and so it might be
5177 * possible to see a half-updated value
5178 * Of course reshape_progress could change after
5179 * the lock is dropped, so once we get a reference
5180 * to the stripe that we think it is, we will have
5183 spin_lock_irq(&conf
->device_lock
);
5184 if (mddev
->reshape_backwards
5185 ? logical_sector
< conf
->reshape_progress
5186 : logical_sector
>= conf
->reshape_progress
) {
5189 if (mddev
->reshape_backwards
5190 ? logical_sector
< conf
->reshape_safe
5191 : logical_sector
>= conf
->reshape_safe
) {
5192 spin_unlock_irq(&conf
->device_lock
);
5198 spin_unlock_irq(&conf
->device_lock
);
5201 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5204 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5205 (unsigned long long)new_sector
,
5206 (unsigned long long)logical_sector
);
5208 sh
= get_active_stripe(conf
, new_sector
, previous
,
5209 (bi
->bi_rw
&RWA_MASK
), 0);
5211 if (unlikely(previous
)) {
5212 /* expansion might have moved on while waiting for a
5213 * stripe, so we must do the range check again.
5214 * Expansion could still move past after this
5215 * test, but as we are holding a reference to
5216 * 'sh', we know that if that happens,
5217 * STRIPE_EXPANDING will get set and the expansion
5218 * won't proceed until we finish with the stripe.
5221 spin_lock_irq(&conf
->device_lock
);
5222 if (mddev
->reshape_backwards
5223 ? logical_sector
>= conf
->reshape_progress
5224 : logical_sector
< conf
->reshape_progress
)
5225 /* mismatch, need to try again */
5227 spin_unlock_irq(&conf
->device_lock
);
5235 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5236 /* Might have got the wrong stripe_head
5244 logical_sector
>= mddev
->suspend_lo
&&
5245 logical_sector
< mddev
->suspend_hi
) {
5247 /* As the suspend_* range is controlled by
5248 * userspace, we want an interruptible
5251 flush_signals(current
);
5252 prepare_to_wait(&conf
->wait_for_overlap
,
5253 &w
, TASK_INTERRUPTIBLE
);
5254 if (logical_sector
>= mddev
->suspend_lo
&&
5255 logical_sector
< mddev
->suspend_hi
) {
5262 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5263 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5264 /* Stripe is busy expanding or
5265 * add failed due to overlap. Flush everything
5268 md_wakeup_thread(mddev
->thread
);
5274 set_bit(STRIPE_HANDLE
, &sh
->state
);
5275 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5276 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5277 (bi
->bi_rw
& REQ_SYNC
) &&
5278 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5279 atomic_inc(&conf
->preread_active_stripes
);
5280 release_stripe_plug(mddev
, sh
);
5282 /* cannot get stripe for read-ahead, just give-up */
5283 bi
->bi_error
= -EIO
;
5287 finish_wait(&conf
->wait_for_overlap
, &w
);
5289 remaining
= raid5_dec_bi_active_stripes(bi
);
5290 if (remaining
== 0) {
5293 md_write_end(mddev
);
5295 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
5301 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5303 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5305 /* reshaping is quite different to recovery/resync so it is
5306 * handled quite separately ... here.
5308 * On each call to sync_request, we gather one chunk worth of
5309 * destination stripes and flag them as expanding.
5310 * Then we find all the source stripes and request reads.
5311 * As the reads complete, handle_stripe will copy the data
5312 * into the destination stripe and release that stripe.
5314 struct r5conf
*conf
= mddev
->private;
5315 struct stripe_head
*sh
;
5316 sector_t first_sector
, last_sector
;
5317 int raid_disks
= conf
->previous_raid_disks
;
5318 int data_disks
= raid_disks
- conf
->max_degraded
;
5319 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5322 sector_t writepos
, readpos
, safepos
;
5323 sector_t stripe_addr
;
5324 int reshape_sectors
;
5325 struct list_head stripes
;
5328 if (sector_nr
== 0) {
5329 /* If restarting in the middle, skip the initial sectors */
5330 if (mddev
->reshape_backwards
&&
5331 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5332 sector_nr
= raid5_size(mddev
, 0, 0)
5333 - conf
->reshape_progress
;
5334 } else if (mddev
->reshape_backwards
&&
5335 conf
->reshape_progress
== MaxSector
) {
5336 /* shouldn't happen, but just in case, finish up.*/
5337 sector_nr
= MaxSector
;
5338 } else if (!mddev
->reshape_backwards
&&
5339 conf
->reshape_progress
> 0)
5340 sector_nr
= conf
->reshape_progress
;
5341 sector_div(sector_nr
, new_data_disks
);
5343 mddev
->curr_resync_completed
= sector_nr
;
5344 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5351 /* We need to process a full chunk at a time.
5352 * If old and new chunk sizes differ, we need to process the
5356 reshape_sectors
= max(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5358 /* We update the metadata at least every 10 seconds, or when
5359 * the data about to be copied would over-write the source of
5360 * the data at the front of the range. i.e. one new_stripe
5361 * along from reshape_progress new_maps to after where
5362 * reshape_safe old_maps to
5364 writepos
= conf
->reshape_progress
;
5365 sector_div(writepos
, new_data_disks
);
5366 readpos
= conf
->reshape_progress
;
5367 sector_div(readpos
, data_disks
);
5368 safepos
= conf
->reshape_safe
;
5369 sector_div(safepos
, data_disks
);
5370 if (mddev
->reshape_backwards
) {
5371 BUG_ON(writepos
< reshape_sectors
);
5372 writepos
-= reshape_sectors
;
5373 readpos
+= reshape_sectors
;
5374 safepos
+= reshape_sectors
;
5376 writepos
+= reshape_sectors
;
5377 /* readpos and safepos are worst-case calculations.
5378 * A negative number is overly pessimistic, and causes
5379 * obvious problems for unsigned storage. So clip to 0.
5381 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5382 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5385 /* Having calculated the 'writepos' possibly use it
5386 * to set 'stripe_addr' which is where we will write to.
5388 if (mddev
->reshape_backwards
) {
5389 BUG_ON(conf
->reshape_progress
== 0);
5390 stripe_addr
= writepos
;
5391 BUG_ON((mddev
->dev_sectors
&
5392 ~((sector_t
)reshape_sectors
- 1))
5393 - reshape_sectors
- stripe_addr
5396 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5397 stripe_addr
= sector_nr
;
5400 /* 'writepos' is the most advanced device address we might write.
5401 * 'readpos' is the least advanced device address we might read.
5402 * 'safepos' is the least address recorded in the metadata as having
5404 * If there is a min_offset_diff, these are adjusted either by
5405 * increasing the safepos/readpos if diff is negative, or
5406 * increasing writepos if diff is positive.
5407 * If 'readpos' is then behind 'writepos', there is no way that we can
5408 * ensure safety in the face of a crash - that must be done by userspace
5409 * making a backup of the data. So in that case there is no particular
5410 * rush to update metadata.
5411 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5412 * update the metadata to advance 'safepos' to match 'readpos' so that
5413 * we can be safe in the event of a crash.
5414 * So we insist on updating metadata if safepos is behind writepos and
5415 * readpos is beyond writepos.
5416 * In any case, update the metadata every 10 seconds.
5417 * Maybe that number should be configurable, but I'm not sure it is
5418 * worth it.... maybe it could be a multiple of safemode_delay???
5420 if (conf
->min_offset_diff
< 0) {
5421 safepos
+= -conf
->min_offset_diff
;
5422 readpos
+= -conf
->min_offset_diff
;
5424 writepos
+= conf
->min_offset_diff
;
5426 if ((mddev
->reshape_backwards
5427 ? (safepos
> writepos
&& readpos
< writepos
)
5428 : (safepos
< writepos
&& readpos
> writepos
)) ||
5429 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5430 /* Cannot proceed until we've updated the superblock... */
5431 wait_event(conf
->wait_for_overlap
,
5432 atomic_read(&conf
->reshape_stripes
)==0
5433 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5434 if (atomic_read(&conf
->reshape_stripes
) != 0)
5436 mddev
->reshape_position
= conf
->reshape_progress
;
5437 mddev
->curr_resync_completed
= sector_nr
;
5438 conf
->reshape_checkpoint
= jiffies
;
5439 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5440 md_wakeup_thread(mddev
->thread
);
5441 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
5442 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5443 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5445 spin_lock_irq(&conf
->device_lock
);
5446 conf
->reshape_safe
= mddev
->reshape_position
;
5447 spin_unlock_irq(&conf
->device_lock
);
5448 wake_up(&conf
->wait_for_overlap
);
5449 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5452 INIT_LIST_HEAD(&stripes
);
5453 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5455 int skipped_disk
= 0;
5456 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5457 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5458 atomic_inc(&conf
->reshape_stripes
);
5459 /* If any of this stripe is beyond the end of the old
5460 * array, then we need to zero those blocks
5462 for (j
=sh
->disks
; j
--;) {
5464 if (j
== sh
->pd_idx
)
5466 if (conf
->level
== 6 &&
5469 s
= compute_blocknr(sh
, j
, 0);
5470 if (s
< raid5_size(mddev
, 0, 0)) {
5474 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5475 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5476 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5478 if (!skipped_disk
) {
5479 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5480 set_bit(STRIPE_HANDLE
, &sh
->state
);
5482 list_add(&sh
->lru
, &stripes
);
5484 spin_lock_irq(&conf
->device_lock
);
5485 if (mddev
->reshape_backwards
)
5486 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5488 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5489 spin_unlock_irq(&conf
->device_lock
);
5490 /* Ok, those stripe are ready. We can start scheduling
5491 * reads on the source stripes.
5492 * The source stripes are determined by mapping the first and last
5493 * block on the destination stripes.
5496 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5499 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5500 * new_data_disks
- 1),
5502 if (last_sector
>= mddev
->dev_sectors
)
5503 last_sector
= mddev
->dev_sectors
- 1;
5504 while (first_sector
<= last_sector
) {
5505 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
5506 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5507 set_bit(STRIPE_HANDLE
, &sh
->state
);
5509 first_sector
+= STRIPE_SECTORS
;
5511 /* Now that the sources are clearly marked, we can release
5512 * the destination stripes
5514 while (!list_empty(&stripes
)) {
5515 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5516 list_del_init(&sh
->lru
);
5519 /* If this takes us to the resync_max point where we have to pause,
5520 * then we need to write out the superblock.
5522 sector_nr
+= reshape_sectors
;
5523 retn
= reshape_sectors
;
5525 if (mddev
->curr_resync_completed
> mddev
->resync_max
||
5526 (sector_nr
- mddev
->curr_resync_completed
) * 2
5527 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5528 /* Cannot proceed until we've updated the superblock... */
5529 wait_event(conf
->wait_for_overlap
,
5530 atomic_read(&conf
->reshape_stripes
) == 0
5531 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5532 if (atomic_read(&conf
->reshape_stripes
) != 0)
5534 mddev
->reshape_position
= conf
->reshape_progress
;
5535 mddev
->curr_resync_completed
= sector_nr
;
5536 conf
->reshape_checkpoint
= jiffies
;
5537 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5538 md_wakeup_thread(mddev
->thread
);
5539 wait_event(mddev
->sb_wait
,
5540 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
5541 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5542 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5544 spin_lock_irq(&conf
->device_lock
);
5545 conf
->reshape_safe
= mddev
->reshape_position
;
5546 spin_unlock_irq(&conf
->device_lock
);
5547 wake_up(&conf
->wait_for_overlap
);
5548 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5554 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5556 struct r5conf
*conf
= mddev
->private;
5557 struct stripe_head
*sh
;
5558 sector_t max_sector
= mddev
->dev_sectors
;
5559 sector_t sync_blocks
;
5560 int still_degraded
= 0;
5563 if (sector_nr
>= max_sector
) {
5564 /* just being told to finish up .. nothing much to do */
5566 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
5571 if (mddev
->curr_resync
< max_sector
) /* aborted */
5572 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
5574 else /* completed sync */
5576 bitmap_close_sync(mddev
->bitmap
);
5581 /* Allow raid5_quiesce to complete */
5582 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
5584 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
5585 return reshape_request(mddev
, sector_nr
, skipped
);
5587 /* No need to check resync_max as we never do more than one
5588 * stripe, and as resync_max will always be on a chunk boundary,
5589 * if the check in md_do_sync didn't fire, there is no chance
5590 * of overstepping resync_max here
5593 /* if there is too many failed drives and we are trying
5594 * to resync, then assert that we are finished, because there is
5595 * nothing we can do.
5597 if (mddev
->degraded
>= conf
->max_degraded
&&
5598 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
5599 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
5603 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
5605 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
5606 sync_blocks
>= STRIPE_SECTORS
) {
5607 /* we can skip this block, and probably more */
5608 sync_blocks
/= STRIPE_SECTORS
;
5610 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
5613 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
5615 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
5617 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
5618 /* make sure we don't swamp the stripe cache if someone else
5619 * is trying to get access
5621 schedule_timeout_uninterruptible(1);
5623 /* Need to check if array will still be degraded after recovery/resync
5624 * Note in case of > 1 drive failures it's possible we're rebuilding
5625 * one drive while leaving another faulty drive in array.
5628 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5629 struct md_rdev
*rdev
= ACCESS_ONCE(conf
->disks
[i
].rdev
);
5631 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
5636 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
5638 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
5639 set_bit(STRIPE_HANDLE
, &sh
->state
);
5643 return STRIPE_SECTORS
;
5646 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
5648 /* We may not be able to submit a whole bio at once as there
5649 * may not be enough stripe_heads available.
5650 * We cannot pre-allocate enough stripe_heads as we may need
5651 * more than exist in the cache (if we allow ever large chunks).
5652 * So we do one stripe head at a time and record in
5653 * ->bi_hw_segments how many have been done.
5655 * We *know* that this entire raid_bio is in one chunk, so
5656 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5658 struct stripe_head
*sh
;
5660 sector_t sector
, logical_sector
, last_sector
;
5665 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
5666 ~((sector_t
)STRIPE_SECTORS
-1);
5667 sector
= raid5_compute_sector(conf
, logical_sector
,
5669 last_sector
= bio_end_sector(raid_bio
);
5671 for (; logical_sector
< last_sector
;
5672 logical_sector
+= STRIPE_SECTORS
,
5673 sector
+= STRIPE_SECTORS
,
5676 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
5677 /* already done this stripe */
5680 sh
= get_active_stripe(conf
, sector
, 0, 1, 1);
5683 /* failed to get a stripe - must wait */
5684 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5685 conf
->retry_read_aligned
= raid_bio
;
5689 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
5691 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5692 conf
->retry_read_aligned
= raid_bio
;
5696 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
5701 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
5702 if (remaining
== 0) {
5703 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
5705 bio_endio(raid_bio
);
5707 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5708 wake_up(&conf
->wait_for_quiescent
);
5712 static int handle_active_stripes(struct r5conf
*conf
, int group
,
5713 struct r5worker
*worker
,
5714 struct list_head
*temp_inactive_list
)
5716 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
5717 int i
, batch_size
= 0, hash
;
5718 bool release_inactive
= false;
5720 while (batch_size
< MAX_STRIPE_BATCH
&&
5721 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
5722 batch
[batch_size
++] = sh
;
5724 if (batch_size
== 0) {
5725 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5726 if (!list_empty(temp_inactive_list
+ i
))
5728 if (i
== NR_STRIPE_HASH_LOCKS
)
5730 release_inactive
= true;
5732 spin_unlock_irq(&conf
->device_lock
);
5734 release_inactive_stripe_list(conf
, temp_inactive_list
,
5735 NR_STRIPE_HASH_LOCKS
);
5737 if (release_inactive
) {
5738 spin_lock_irq(&conf
->device_lock
);
5742 for (i
= 0; i
< batch_size
; i
++)
5743 handle_stripe(batch
[i
]);
5747 spin_lock_irq(&conf
->device_lock
);
5748 for (i
= 0; i
< batch_size
; i
++) {
5749 hash
= batch
[i
]->hash_lock_index
;
5750 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
5755 static void raid5_do_work(struct work_struct
*work
)
5757 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
5758 struct r5worker_group
*group
= worker
->group
;
5759 struct r5conf
*conf
= group
->conf
;
5760 int group_id
= group
- conf
->worker_groups
;
5762 struct blk_plug plug
;
5764 pr_debug("+++ raid5worker active\n");
5766 blk_start_plug(&plug
);
5768 spin_lock_irq(&conf
->device_lock
);
5770 int batch_size
, released
;
5772 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
5774 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
5775 worker
->temp_inactive_list
);
5776 worker
->working
= false;
5777 if (!batch_size
&& !released
)
5779 handled
+= batch_size
;
5781 pr_debug("%d stripes handled\n", handled
);
5783 spin_unlock_irq(&conf
->device_lock
);
5784 blk_finish_plug(&plug
);
5786 pr_debug("--- raid5worker inactive\n");
5790 * This is our raid5 kernel thread.
5792 * We scan the hash table for stripes which can be handled now.
5793 * During the scan, completed stripes are saved for us by the interrupt
5794 * handler, so that they will not have to wait for our next wakeup.
5796 static void raid5d(struct md_thread
*thread
)
5798 struct mddev
*mddev
= thread
->mddev
;
5799 struct r5conf
*conf
= mddev
->private;
5801 struct blk_plug plug
;
5803 pr_debug("+++ raid5d active\n");
5805 md_check_recovery(mddev
);
5807 if (!bio_list_empty(&conf
->return_bi
) &&
5808 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
5809 struct bio_list tmp
= BIO_EMPTY_LIST
;
5810 spin_lock_irq(&conf
->device_lock
);
5811 if (!test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
5812 bio_list_merge(&tmp
, &conf
->return_bi
);
5813 bio_list_init(&conf
->return_bi
);
5815 spin_unlock_irq(&conf
->device_lock
);
5819 blk_start_plug(&plug
);
5821 spin_lock_irq(&conf
->device_lock
);
5824 int batch_size
, released
;
5826 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
5828 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5831 !list_empty(&conf
->bitmap_list
)) {
5832 /* Now is a good time to flush some bitmap updates */
5834 spin_unlock_irq(&conf
->device_lock
);
5835 bitmap_unplug(mddev
->bitmap
);
5836 spin_lock_irq(&conf
->device_lock
);
5837 conf
->seq_write
= conf
->seq_flush
;
5838 activate_bit_delay(conf
, conf
->temp_inactive_list
);
5840 raid5_activate_delayed(conf
);
5842 while ((bio
= remove_bio_from_retry(conf
))) {
5844 spin_unlock_irq(&conf
->device_lock
);
5845 ok
= retry_aligned_read(conf
, bio
);
5846 spin_lock_irq(&conf
->device_lock
);
5852 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
5853 conf
->temp_inactive_list
);
5854 if (!batch_size
&& !released
)
5856 handled
+= batch_size
;
5858 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
5859 spin_unlock_irq(&conf
->device_lock
);
5860 md_check_recovery(mddev
);
5861 spin_lock_irq(&conf
->device_lock
);
5864 pr_debug("%d stripes handled\n", handled
);
5866 spin_unlock_irq(&conf
->device_lock
);
5867 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
) &&
5868 mutex_trylock(&conf
->cache_size_mutex
)) {
5869 grow_one_stripe(conf
, __GFP_NOWARN
);
5870 /* Set flag even if allocation failed. This helps
5871 * slow down allocation requests when mem is short
5873 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
5874 mutex_unlock(&conf
->cache_size_mutex
);
5877 async_tx_issue_pending_all();
5878 blk_finish_plug(&plug
);
5880 pr_debug("--- raid5d inactive\n");
5884 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
5886 struct r5conf
*conf
;
5888 spin_lock(&mddev
->lock
);
5889 conf
= mddev
->private;
5891 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
5892 spin_unlock(&mddev
->lock
);
5897 raid5_set_cache_size(struct mddev
*mddev
, int size
)
5899 struct r5conf
*conf
= mddev
->private;
5902 if (size
<= 16 || size
> 32768)
5905 conf
->min_nr_stripes
= size
;
5906 mutex_lock(&conf
->cache_size_mutex
);
5907 while (size
< conf
->max_nr_stripes
&&
5908 drop_one_stripe(conf
))
5910 mutex_unlock(&conf
->cache_size_mutex
);
5913 err
= md_allow_write(mddev
);
5917 mutex_lock(&conf
->cache_size_mutex
);
5918 while (size
> conf
->max_nr_stripes
)
5919 if (!grow_one_stripe(conf
, GFP_KERNEL
))
5921 mutex_unlock(&conf
->cache_size_mutex
);
5925 EXPORT_SYMBOL(raid5_set_cache_size
);
5928 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
5930 struct r5conf
*conf
;
5934 if (len
>= PAGE_SIZE
)
5936 if (kstrtoul(page
, 10, &new))
5938 err
= mddev_lock(mddev
);
5941 conf
= mddev
->private;
5945 err
= raid5_set_cache_size(mddev
, new);
5946 mddev_unlock(mddev
);
5951 static struct md_sysfs_entry
5952 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
5953 raid5_show_stripe_cache_size
,
5954 raid5_store_stripe_cache_size
);
5957 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
5959 struct r5conf
*conf
= mddev
->private;
5961 return sprintf(page
, "%d\n", conf
->rmw_level
);
5967 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
5969 struct r5conf
*conf
= mddev
->private;
5975 if (len
>= PAGE_SIZE
)
5978 if (kstrtoul(page
, 10, &new))
5981 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
5984 if (new != PARITY_DISABLE_RMW
&&
5985 new != PARITY_ENABLE_RMW
&&
5986 new != PARITY_PREFER_RMW
)
5989 conf
->rmw_level
= new;
5993 static struct md_sysfs_entry
5994 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
5995 raid5_show_rmw_level
,
5996 raid5_store_rmw_level
);
6000 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
6002 struct r5conf
*conf
;
6004 spin_lock(&mddev
->lock
);
6005 conf
= mddev
->private;
6007 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
6008 spin_unlock(&mddev
->lock
);
6013 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
6015 struct r5conf
*conf
;
6019 if (len
>= PAGE_SIZE
)
6021 if (kstrtoul(page
, 10, &new))
6024 err
= mddev_lock(mddev
);
6027 conf
= mddev
->private;
6030 else if (new > conf
->min_nr_stripes
)
6033 conf
->bypass_threshold
= new;
6034 mddev_unlock(mddev
);
6038 static struct md_sysfs_entry
6039 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
6041 raid5_show_preread_threshold
,
6042 raid5_store_preread_threshold
);
6045 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
6047 struct r5conf
*conf
;
6049 spin_lock(&mddev
->lock
);
6050 conf
= mddev
->private;
6052 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6053 spin_unlock(&mddev
->lock
);
6058 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6060 struct r5conf
*conf
;
6064 if (len
>= PAGE_SIZE
)
6066 if (kstrtoul(page
, 10, &new))
6070 err
= mddev_lock(mddev
);
6073 conf
= mddev
->private;
6076 else if (new != conf
->skip_copy
) {
6077 mddev_suspend(mddev
);
6078 conf
->skip_copy
= new;
6080 mddev
->queue
->backing_dev_info
.capabilities
|=
6081 BDI_CAP_STABLE_WRITES
;
6083 mddev
->queue
->backing_dev_info
.capabilities
&=
6084 ~BDI_CAP_STABLE_WRITES
;
6085 mddev_resume(mddev
);
6087 mddev_unlock(mddev
);
6091 static struct md_sysfs_entry
6092 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6093 raid5_show_skip_copy
,
6094 raid5_store_skip_copy
);
6097 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6099 struct r5conf
*conf
= mddev
->private;
6101 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6106 static struct md_sysfs_entry
6107 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6110 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6112 struct r5conf
*conf
;
6114 spin_lock(&mddev
->lock
);
6115 conf
= mddev
->private;
6117 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6118 spin_unlock(&mddev
->lock
);
6122 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6124 int *worker_cnt_per_group
,
6125 struct r5worker_group
**worker_groups
);
6127 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6129 struct r5conf
*conf
;
6132 struct r5worker_group
*new_groups
, *old_groups
;
6133 int group_cnt
, worker_cnt_per_group
;
6135 if (len
>= PAGE_SIZE
)
6137 if (kstrtoul(page
, 10, &new))
6140 err
= mddev_lock(mddev
);
6143 conf
= mddev
->private;
6146 else if (new != conf
->worker_cnt_per_group
) {
6147 mddev_suspend(mddev
);
6149 old_groups
= conf
->worker_groups
;
6151 flush_workqueue(raid5_wq
);
6153 err
= alloc_thread_groups(conf
, new,
6154 &group_cnt
, &worker_cnt_per_group
,
6157 spin_lock_irq(&conf
->device_lock
);
6158 conf
->group_cnt
= group_cnt
;
6159 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6160 conf
->worker_groups
= new_groups
;
6161 spin_unlock_irq(&conf
->device_lock
);
6164 kfree(old_groups
[0].workers
);
6167 mddev_resume(mddev
);
6169 mddev_unlock(mddev
);
6174 static struct md_sysfs_entry
6175 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6176 raid5_show_group_thread_cnt
,
6177 raid5_store_group_thread_cnt
);
6179 static struct attribute
*raid5_attrs
[] = {
6180 &raid5_stripecache_size
.attr
,
6181 &raid5_stripecache_active
.attr
,
6182 &raid5_preread_bypass_threshold
.attr
,
6183 &raid5_group_thread_cnt
.attr
,
6184 &raid5_skip_copy
.attr
,
6185 &raid5_rmw_level
.attr
,
6188 static struct attribute_group raid5_attrs_group
= {
6190 .attrs
= raid5_attrs
,
6193 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6195 int *worker_cnt_per_group
,
6196 struct r5worker_group
**worker_groups
)
6200 struct r5worker
*workers
;
6202 *worker_cnt_per_group
= cnt
;
6205 *worker_groups
= NULL
;
6208 *group_cnt
= num_possible_nodes();
6209 size
= sizeof(struct r5worker
) * cnt
;
6210 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6211 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6212 *group_cnt
, GFP_NOIO
);
6213 if (!*worker_groups
|| !workers
) {
6215 kfree(*worker_groups
);
6219 for (i
= 0; i
< *group_cnt
; i
++) {
6220 struct r5worker_group
*group
;
6222 group
= &(*worker_groups
)[i
];
6223 INIT_LIST_HEAD(&group
->handle_list
);
6225 group
->workers
= workers
+ i
* cnt
;
6227 for (j
= 0; j
< cnt
; j
++) {
6228 struct r5worker
*worker
= group
->workers
+ j
;
6229 worker
->group
= group
;
6230 INIT_WORK(&worker
->work
, raid5_do_work
);
6232 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6233 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6240 static void free_thread_groups(struct r5conf
*conf
)
6242 if (conf
->worker_groups
)
6243 kfree(conf
->worker_groups
[0].workers
);
6244 kfree(conf
->worker_groups
);
6245 conf
->worker_groups
= NULL
;
6249 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6251 struct r5conf
*conf
= mddev
->private;
6254 sectors
= mddev
->dev_sectors
;
6256 /* size is defined by the smallest of previous and new size */
6257 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6259 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
6260 sectors
&= ~((sector_t
)conf
->prev_chunk_sectors
- 1);
6261 return sectors
* (raid_disks
- conf
->max_degraded
);
6264 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6266 safe_put_page(percpu
->spare_page
);
6267 if (percpu
->scribble
)
6268 flex_array_free(percpu
->scribble
);
6269 percpu
->spare_page
= NULL
;
6270 percpu
->scribble
= NULL
;
6273 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6275 if (conf
->level
== 6 && !percpu
->spare_page
)
6276 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6277 if (!percpu
->scribble
)
6278 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6279 conf
->previous_raid_disks
),
6280 max(conf
->chunk_sectors
,
6281 conf
->prev_chunk_sectors
)
6285 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6286 free_scratch_buffer(conf
, percpu
);
6293 static void raid5_free_percpu(struct r5conf
*conf
)
6300 #ifdef CONFIG_HOTPLUG_CPU
6301 unregister_cpu_notifier(&conf
->cpu_notify
);
6305 for_each_possible_cpu(cpu
)
6306 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6309 free_percpu(conf
->percpu
);
6312 static void free_conf(struct r5conf
*conf
)
6314 if (conf
->shrinker
.seeks
)
6315 unregister_shrinker(&conf
->shrinker
);
6316 free_thread_groups(conf
);
6317 shrink_stripes(conf
);
6318 raid5_free_percpu(conf
);
6320 kfree(conf
->stripe_hashtbl
);
6324 #ifdef CONFIG_HOTPLUG_CPU
6325 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
6328 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
6329 long cpu
= (long)hcpu
;
6330 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6333 case CPU_UP_PREPARE
:
6334 case CPU_UP_PREPARE_FROZEN
:
6335 if (alloc_scratch_buffer(conf
, percpu
)) {
6336 pr_err("%s: failed memory allocation for cpu%ld\n",
6338 return notifier_from_errno(-ENOMEM
);
6342 case CPU_DEAD_FROZEN
:
6343 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6352 static int raid5_alloc_percpu(struct r5conf
*conf
)
6357 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6361 #ifdef CONFIG_HOTPLUG_CPU
6362 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
6363 conf
->cpu_notify
.priority
= 0;
6364 err
= register_cpu_notifier(&conf
->cpu_notify
);
6370 for_each_present_cpu(cpu
) {
6371 err
= alloc_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6373 pr_err("%s: failed memory allocation for cpu%ld\n",
6383 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6384 struct shrink_control
*sc
)
6386 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6387 unsigned long ret
= SHRINK_STOP
;
6389 if (mutex_trylock(&conf
->cache_size_mutex
)) {
6391 while (ret
< sc
->nr_to_scan
&&
6392 conf
->max_nr_stripes
> conf
->min_nr_stripes
) {
6393 if (drop_one_stripe(conf
) == 0) {
6399 mutex_unlock(&conf
->cache_size_mutex
);
6404 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6405 struct shrink_control
*sc
)
6407 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6409 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6410 /* unlikely, but not impossible */
6412 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6415 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6417 struct r5conf
*conf
;
6418 int raid_disk
, memory
, max_disks
;
6419 struct md_rdev
*rdev
;
6420 struct disk_info
*disk
;
6423 int group_cnt
, worker_cnt_per_group
;
6424 struct r5worker_group
*new_group
;
6426 if (mddev
->new_level
!= 5
6427 && mddev
->new_level
!= 4
6428 && mddev
->new_level
!= 6) {
6429 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6430 mdname(mddev
), mddev
->new_level
);
6431 return ERR_PTR(-EIO
);
6433 if ((mddev
->new_level
== 5
6434 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6435 (mddev
->new_level
== 6
6436 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6437 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
6438 mdname(mddev
), mddev
->new_layout
);
6439 return ERR_PTR(-EIO
);
6441 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6442 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6443 mdname(mddev
), mddev
->raid_disks
);
6444 return ERR_PTR(-EINVAL
);
6447 if (!mddev
->new_chunk_sectors
||
6448 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6449 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6450 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
6451 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6452 return ERR_PTR(-EINVAL
);
6455 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6458 /* Don't enable multi-threading by default*/
6459 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6461 conf
->group_cnt
= group_cnt
;
6462 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6463 conf
->worker_groups
= new_group
;
6466 spin_lock_init(&conf
->device_lock
);
6467 seqcount_init(&conf
->gen_lock
);
6468 mutex_init(&conf
->cache_size_mutex
);
6469 init_waitqueue_head(&conf
->wait_for_quiescent
);
6470 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++) {
6471 init_waitqueue_head(&conf
->wait_for_stripe
[i
]);
6473 init_waitqueue_head(&conf
->wait_for_overlap
);
6474 INIT_LIST_HEAD(&conf
->handle_list
);
6475 INIT_LIST_HEAD(&conf
->hold_list
);
6476 INIT_LIST_HEAD(&conf
->delayed_list
);
6477 INIT_LIST_HEAD(&conf
->bitmap_list
);
6478 bio_list_init(&conf
->return_bi
);
6479 init_llist_head(&conf
->released_stripes
);
6480 atomic_set(&conf
->active_stripes
, 0);
6481 atomic_set(&conf
->preread_active_stripes
, 0);
6482 atomic_set(&conf
->active_aligned_reads
, 0);
6483 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6484 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6486 conf
->raid_disks
= mddev
->raid_disks
;
6487 if (mddev
->reshape_position
== MaxSector
)
6488 conf
->previous_raid_disks
= mddev
->raid_disks
;
6490 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6491 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6493 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6498 conf
->mddev
= mddev
;
6500 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6503 /* We init hash_locks[0] separately to that it can be used
6504 * as the reference lock in the spin_lock_nest_lock() call
6505 * in lock_all_device_hash_locks_irq in order to convince
6506 * lockdep that we know what we are doing.
6508 spin_lock_init(conf
->hash_locks
);
6509 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6510 spin_lock_init(conf
->hash_locks
+ i
);
6512 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6513 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6515 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6516 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6518 conf
->level
= mddev
->new_level
;
6519 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6520 if (raid5_alloc_percpu(conf
) != 0)
6523 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6525 rdev_for_each(rdev
, mddev
) {
6526 raid_disk
= rdev
->raid_disk
;
6527 if (raid_disk
>= max_disks
6530 disk
= conf
->disks
+ raid_disk
;
6532 if (test_bit(Replacement
, &rdev
->flags
)) {
6533 if (disk
->replacement
)
6535 disk
->replacement
= rdev
;
6542 if (test_bit(In_sync
, &rdev
->flags
)) {
6543 char b
[BDEVNAME_SIZE
];
6544 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
6546 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
6547 } else if (rdev
->saved_raid_disk
!= raid_disk
)
6548 /* Cannot rely on bitmap to complete recovery */
6552 conf
->level
= mddev
->new_level
;
6553 if (conf
->level
== 6) {
6554 conf
->max_degraded
= 2;
6555 if (raid6_call
.xor_syndrome
)
6556 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6558 conf
->rmw_level
= PARITY_DISABLE_RMW
;
6560 conf
->max_degraded
= 1;
6561 conf
->rmw_level
= PARITY_ENABLE_RMW
;
6563 conf
->algorithm
= mddev
->new_layout
;
6564 conf
->reshape_progress
= mddev
->reshape_position
;
6565 if (conf
->reshape_progress
!= MaxSector
) {
6566 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
6567 conf
->prev_algo
= mddev
->layout
;
6569 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
6570 conf
->prev_algo
= conf
->algorithm
;
6573 conf
->min_nr_stripes
= NR_STRIPES
;
6574 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
6575 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
6576 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
6577 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
6579 "md/raid:%s: couldn't allocate %dkB for buffers\n",
6580 mdname(mddev
), memory
);
6583 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
6584 mdname(mddev
), memory
);
6586 * Losing a stripe head costs more than the time to refill it,
6587 * it reduces the queue depth and so can hurt throughput.
6588 * So set it rather large, scaled by number of devices.
6590 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
6591 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
6592 conf
->shrinker
.count_objects
= raid5_cache_count
;
6593 conf
->shrinker
.batch
= 128;
6594 conf
->shrinker
.flags
= 0;
6595 register_shrinker(&conf
->shrinker
);
6597 sprintf(pers_name
, "raid%d", mddev
->new_level
);
6598 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
6599 if (!conf
->thread
) {
6601 "md/raid:%s: couldn't allocate thread.\n",
6611 return ERR_PTR(-EIO
);
6613 return ERR_PTR(-ENOMEM
);
6616 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
6619 case ALGORITHM_PARITY_0
:
6620 if (raid_disk
< max_degraded
)
6623 case ALGORITHM_PARITY_N
:
6624 if (raid_disk
>= raid_disks
- max_degraded
)
6627 case ALGORITHM_PARITY_0_6
:
6628 if (raid_disk
== 0 ||
6629 raid_disk
== raid_disks
- 1)
6632 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6633 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6634 case ALGORITHM_LEFT_SYMMETRIC_6
:
6635 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6636 if (raid_disk
== raid_disks
- 1)
6642 static int run(struct mddev
*mddev
)
6644 struct r5conf
*conf
;
6645 int working_disks
= 0;
6646 int dirty_parity_disks
= 0;
6647 struct md_rdev
*rdev
;
6648 sector_t reshape_offset
= 0;
6650 long long min_offset_diff
= 0;
6653 if (mddev
->recovery_cp
!= MaxSector
)
6654 printk(KERN_NOTICE
"md/raid:%s: not clean"
6655 " -- starting background reconstruction\n",
6658 rdev_for_each(rdev
, mddev
) {
6660 if (rdev
->raid_disk
< 0)
6662 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
6664 min_offset_diff
= diff
;
6666 } else if (mddev
->reshape_backwards
&&
6667 diff
< min_offset_diff
)
6668 min_offset_diff
= diff
;
6669 else if (!mddev
->reshape_backwards
&&
6670 diff
> min_offset_diff
)
6671 min_offset_diff
= diff
;
6674 if (mddev
->reshape_position
!= MaxSector
) {
6675 /* Check that we can continue the reshape.
6676 * Difficulties arise if the stripe we would write to
6677 * next is at or after the stripe we would read from next.
6678 * For a reshape that changes the number of devices, this
6679 * is only possible for a very short time, and mdadm makes
6680 * sure that time appears to have past before assembling
6681 * the array. So we fail if that time hasn't passed.
6682 * For a reshape that keeps the number of devices the same
6683 * mdadm must be monitoring the reshape can keeping the
6684 * critical areas read-only and backed up. It will start
6685 * the array in read-only mode, so we check for that.
6687 sector_t here_new
, here_old
;
6689 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
6693 if (mddev
->new_level
!= mddev
->level
) {
6694 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
6695 "required - aborting.\n",
6699 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6700 /* reshape_position must be on a new-stripe boundary, and one
6701 * further up in new geometry must map after here in old
6703 * If the chunk sizes are different, then as we perform reshape
6704 * in units of the largest of the two, reshape_position needs
6705 * be a multiple of the largest chunk size times new data disks.
6707 here_new
= mddev
->reshape_position
;
6708 chunk_sectors
= max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
);
6709 new_data_disks
= mddev
->raid_disks
- max_degraded
;
6710 if (sector_div(here_new
, chunk_sectors
* new_data_disks
)) {
6711 printk(KERN_ERR
"md/raid:%s: reshape_position not "
6712 "on a stripe boundary\n", mdname(mddev
));
6715 reshape_offset
= here_new
* chunk_sectors
;
6716 /* here_new is the stripe we will write to */
6717 here_old
= mddev
->reshape_position
;
6718 sector_div(here_old
, chunk_sectors
* (old_disks
-max_degraded
));
6719 /* here_old is the first stripe that we might need to read
6721 if (mddev
->delta_disks
== 0) {
6722 /* We cannot be sure it is safe to start an in-place
6723 * reshape. It is only safe if user-space is monitoring
6724 * and taking constant backups.
6725 * mdadm always starts a situation like this in
6726 * readonly mode so it can take control before
6727 * allowing any writes. So just check for that.
6729 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
6730 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
6731 /* not really in-place - so OK */;
6732 else if (mddev
->ro
== 0) {
6733 printk(KERN_ERR
"md/raid:%s: in-place reshape "
6734 "must be started in read-only mode "
6739 } else if (mddev
->reshape_backwards
6740 ? (here_new
* chunk_sectors
+ min_offset_diff
<=
6741 here_old
* chunk_sectors
)
6742 : (here_new
* chunk_sectors
>=
6743 here_old
* chunk_sectors
+ (-min_offset_diff
))) {
6744 /* Reading from the same stripe as writing to - bad */
6745 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
6746 "auto-recovery - aborting.\n",
6750 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
6752 /* OK, we should be able to continue; */
6754 BUG_ON(mddev
->level
!= mddev
->new_level
);
6755 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
6756 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
6757 BUG_ON(mddev
->delta_disks
!= 0);
6760 if (mddev
->private == NULL
)
6761 conf
= setup_conf(mddev
);
6763 conf
= mddev
->private;
6766 return PTR_ERR(conf
);
6768 conf
->min_offset_diff
= min_offset_diff
;
6769 mddev
->thread
= conf
->thread
;
6770 conf
->thread
= NULL
;
6771 mddev
->private = conf
;
6773 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
6775 rdev
= conf
->disks
[i
].rdev
;
6776 if (!rdev
&& conf
->disks
[i
].replacement
) {
6777 /* The replacement is all we have yet */
6778 rdev
= conf
->disks
[i
].replacement
;
6779 conf
->disks
[i
].replacement
= NULL
;
6780 clear_bit(Replacement
, &rdev
->flags
);
6781 conf
->disks
[i
].rdev
= rdev
;
6785 if (conf
->disks
[i
].replacement
&&
6786 conf
->reshape_progress
!= MaxSector
) {
6787 /* replacements and reshape simply do not mix. */
6788 printk(KERN_ERR
"md: cannot handle concurrent "
6789 "replacement and reshape.\n");
6792 if (test_bit(In_sync
, &rdev
->flags
)) {
6796 /* This disc is not fully in-sync. However if it
6797 * just stored parity (beyond the recovery_offset),
6798 * when we don't need to be concerned about the
6799 * array being dirty.
6800 * When reshape goes 'backwards', we never have
6801 * partially completed devices, so we only need
6802 * to worry about reshape going forwards.
6804 /* Hack because v0.91 doesn't store recovery_offset properly. */
6805 if (mddev
->major_version
== 0 &&
6806 mddev
->minor_version
> 90)
6807 rdev
->recovery_offset
= reshape_offset
;
6809 if (rdev
->recovery_offset
< reshape_offset
) {
6810 /* We need to check old and new layout */
6811 if (!only_parity(rdev
->raid_disk
,
6814 conf
->max_degraded
))
6817 if (!only_parity(rdev
->raid_disk
,
6819 conf
->previous_raid_disks
,
6820 conf
->max_degraded
))
6822 dirty_parity_disks
++;
6826 * 0 for a fully functional array, 1 or 2 for a degraded array.
6828 mddev
->degraded
= calc_degraded(conf
);
6830 if (has_failed(conf
)) {
6831 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
6832 " (%d/%d failed)\n",
6833 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
6837 /* device size must be a multiple of chunk size */
6838 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
6839 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
6841 if (mddev
->degraded
> dirty_parity_disks
&&
6842 mddev
->recovery_cp
!= MaxSector
) {
6843 if (mddev
->ok_start_degraded
)
6845 "md/raid:%s: starting dirty degraded array"
6846 " - data corruption possible.\n",
6850 "md/raid:%s: cannot start dirty degraded array.\n",
6856 if (mddev
->degraded
== 0)
6857 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
6858 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
6859 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
6862 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
6863 " out of %d devices, algorithm %d\n",
6864 mdname(mddev
), conf
->level
,
6865 mddev
->raid_disks
- mddev
->degraded
,
6866 mddev
->raid_disks
, mddev
->new_layout
);
6868 print_raid5_conf(conf
);
6870 if (conf
->reshape_progress
!= MaxSector
) {
6871 conf
->reshape_safe
= conf
->reshape_progress
;
6872 atomic_set(&conf
->reshape_stripes
, 0);
6873 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6874 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6875 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6876 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6877 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6881 /* Ok, everything is just fine now */
6882 if (mddev
->to_remove
== &raid5_attrs_group
)
6883 mddev
->to_remove
= NULL
;
6884 else if (mddev
->kobj
.sd
&&
6885 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
6887 "raid5: failed to create sysfs attributes for %s\n",
6889 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6893 bool discard_supported
= true;
6894 /* read-ahead size must cover two whole stripes, which
6895 * is 2 * (datadisks) * chunksize where 'n' is the
6896 * number of raid devices
6898 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
6899 int stripe
= data_disks
*
6900 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
6901 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6902 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6904 chunk_size
= mddev
->chunk_sectors
<< 9;
6905 blk_queue_io_min(mddev
->queue
, chunk_size
);
6906 blk_queue_io_opt(mddev
->queue
, chunk_size
*
6907 (conf
->raid_disks
- conf
->max_degraded
));
6908 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
6910 * We can only discard a whole stripe. It doesn't make sense to
6911 * discard data disk but write parity disk
6913 stripe
= stripe
* PAGE_SIZE
;
6914 /* Round up to power of 2, as discard handling
6915 * currently assumes that */
6916 while ((stripe
-1) & stripe
)
6917 stripe
= (stripe
| (stripe
-1)) + 1;
6918 mddev
->queue
->limits
.discard_alignment
= stripe
;
6919 mddev
->queue
->limits
.discard_granularity
= stripe
;
6921 * unaligned part of discard request will be ignored, so can't
6922 * guarantee discard_zeroes_data
6924 mddev
->queue
->limits
.discard_zeroes_data
= 0;
6926 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
6928 rdev_for_each(rdev
, mddev
) {
6929 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6930 rdev
->data_offset
<< 9);
6931 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6932 rdev
->new_data_offset
<< 9);
6934 * discard_zeroes_data is required, otherwise data
6935 * could be lost. Consider a scenario: discard a stripe
6936 * (the stripe could be inconsistent if
6937 * discard_zeroes_data is 0); write one disk of the
6938 * stripe (the stripe could be inconsistent again
6939 * depending on which disks are used to calculate
6940 * parity); the disk is broken; The stripe data of this
6943 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
6944 !bdev_get_queue(rdev
->bdev
)->
6945 limits
.discard_zeroes_data
)
6946 discard_supported
= false;
6947 /* Unfortunately, discard_zeroes_data is not currently
6948 * a guarantee - just a hint. So we only allow DISCARD
6949 * if the sysadmin has confirmed that only safe devices
6950 * are in use by setting a module parameter.
6952 if (!devices_handle_discard_safely
) {
6953 if (discard_supported
) {
6954 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6955 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6957 discard_supported
= false;
6961 if (discard_supported
&&
6962 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
6963 mddev
->queue
->limits
.discard_granularity
>= stripe
)
6964 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
6967 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
6973 md_unregister_thread(&mddev
->thread
);
6974 print_raid5_conf(conf
);
6976 mddev
->private = NULL
;
6977 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
6981 static void raid5_free(struct mddev
*mddev
, void *priv
)
6983 struct r5conf
*conf
= priv
;
6986 mddev
->to_remove
= &raid5_attrs_group
;
6989 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
6991 struct r5conf
*conf
= mddev
->private;
6994 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
6995 conf
->chunk_sectors
/ 2, mddev
->layout
);
6996 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
6997 for (i
= 0; i
< conf
->raid_disks
; i
++)
6998 seq_printf (seq
, "%s",
6999 conf
->disks
[i
].rdev
&&
7000 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
7001 seq_printf (seq
, "]");
7004 static void print_raid5_conf (struct r5conf
*conf
)
7007 struct disk_info
*tmp
;
7009 printk(KERN_DEBUG
"RAID conf printout:\n");
7011 printk("(conf==NULL)\n");
7014 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
7016 conf
->raid_disks
- conf
->mddev
->degraded
);
7018 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7019 char b
[BDEVNAME_SIZE
];
7020 tmp
= conf
->disks
+ i
;
7022 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
7023 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
7024 bdevname(tmp
->rdev
->bdev
, b
));
7028 static int raid5_spare_active(struct mddev
*mddev
)
7031 struct r5conf
*conf
= mddev
->private;
7032 struct disk_info
*tmp
;
7034 unsigned long flags
;
7036 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7037 tmp
= conf
->disks
+ i
;
7038 if (tmp
->replacement
7039 && tmp
->replacement
->recovery_offset
== MaxSector
7040 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
7041 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
7042 /* Replacement has just become active. */
7044 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
7047 /* Replaced device not technically faulty,
7048 * but we need to be sure it gets removed
7049 * and never re-added.
7051 set_bit(Faulty
, &tmp
->rdev
->flags
);
7052 sysfs_notify_dirent_safe(
7053 tmp
->rdev
->sysfs_state
);
7055 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
7056 } else if (tmp
->rdev
7057 && tmp
->rdev
->recovery_offset
== MaxSector
7058 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
7059 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
7061 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
7064 spin_lock_irqsave(&conf
->device_lock
, flags
);
7065 mddev
->degraded
= calc_degraded(conf
);
7066 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7067 print_raid5_conf(conf
);
7071 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7073 struct r5conf
*conf
= mddev
->private;
7075 int number
= rdev
->raid_disk
;
7076 struct md_rdev
**rdevp
;
7077 struct disk_info
*p
= conf
->disks
+ number
;
7079 print_raid5_conf(conf
);
7080 if (rdev
== p
->rdev
)
7082 else if (rdev
== p
->replacement
)
7083 rdevp
= &p
->replacement
;
7087 if (number
>= conf
->raid_disks
&&
7088 conf
->reshape_progress
== MaxSector
)
7089 clear_bit(In_sync
, &rdev
->flags
);
7091 if (test_bit(In_sync
, &rdev
->flags
) ||
7092 atomic_read(&rdev
->nr_pending
)) {
7096 /* Only remove non-faulty devices if recovery
7099 if (!test_bit(Faulty
, &rdev
->flags
) &&
7100 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7101 !has_failed(conf
) &&
7102 (!p
->replacement
|| p
->replacement
== rdev
) &&
7103 number
< conf
->raid_disks
) {
7109 if (atomic_read(&rdev
->nr_pending
)) {
7110 /* lost the race, try later */
7113 } else if (p
->replacement
) {
7114 /* We must have just cleared 'rdev' */
7115 p
->rdev
= p
->replacement
;
7116 clear_bit(Replacement
, &p
->replacement
->flags
);
7117 smp_mb(); /* Make sure other CPUs may see both as identical
7118 * but will never see neither - if they are careful
7120 p
->replacement
= NULL
;
7121 clear_bit(WantReplacement
, &rdev
->flags
);
7123 /* We might have just removed the Replacement as faulty-
7124 * clear the bit just in case
7126 clear_bit(WantReplacement
, &rdev
->flags
);
7129 print_raid5_conf(conf
);
7133 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7135 struct r5conf
*conf
= mddev
->private;
7138 struct disk_info
*p
;
7140 int last
= conf
->raid_disks
- 1;
7142 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7145 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7146 /* no point adding a device */
7149 if (rdev
->raid_disk
>= 0)
7150 first
= last
= rdev
->raid_disk
;
7153 * find the disk ... but prefer rdev->saved_raid_disk
7156 if (rdev
->saved_raid_disk
>= 0 &&
7157 rdev
->saved_raid_disk
>= first
&&
7158 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7159 first
= rdev
->saved_raid_disk
;
7161 for (disk
= first
; disk
<= last
; disk
++) {
7162 p
= conf
->disks
+ disk
;
7163 if (p
->rdev
== NULL
) {
7164 clear_bit(In_sync
, &rdev
->flags
);
7165 rdev
->raid_disk
= disk
;
7167 if (rdev
->saved_raid_disk
!= disk
)
7169 rcu_assign_pointer(p
->rdev
, rdev
);
7173 for (disk
= first
; disk
<= last
; disk
++) {
7174 p
= conf
->disks
+ disk
;
7175 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7176 p
->replacement
== NULL
) {
7177 clear_bit(In_sync
, &rdev
->flags
);
7178 set_bit(Replacement
, &rdev
->flags
);
7179 rdev
->raid_disk
= disk
;
7182 rcu_assign_pointer(p
->replacement
, rdev
);
7187 print_raid5_conf(conf
);
7191 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7193 /* no resync is happening, and there is enough space
7194 * on all devices, so we can resize.
7195 * We need to make sure resync covers any new space.
7196 * If the array is shrinking we should possibly wait until
7197 * any io in the removed space completes, but it hardly seems
7201 struct r5conf
*conf
= mddev
->private;
7203 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
7204 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7205 if (mddev
->external_size
&&
7206 mddev
->array_sectors
> newsize
)
7208 if (mddev
->bitmap
) {
7209 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7213 md_set_array_sectors(mddev
, newsize
);
7214 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7215 revalidate_disk(mddev
->gendisk
);
7216 if (sectors
> mddev
->dev_sectors
&&
7217 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7218 mddev
->recovery_cp
= mddev
->dev_sectors
;
7219 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7221 mddev
->dev_sectors
= sectors
;
7222 mddev
->resync_max_sectors
= sectors
;
7226 static int check_stripe_cache(struct mddev
*mddev
)
7228 /* Can only proceed if there are plenty of stripe_heads.
7229 * We need a minimum of one full stripe,, and for sensible progress
7230 * it is best to have about 4 times that.
7231 * If we require 4 times, then the default 256 4K stripe_heads will
7232 * allow for chunk sizes up to 256K, which is probably OK.
7233 * If the chunk size is greater, user-space should request more
7234 * stripe_heads first.
7236 struct r5conf
*conf
= mddev
->private;
7237 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7238 > conf
->min_nr_stripes
||
7239 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7240 > conf
->min_nr_stripes
) {
7241 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7243 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7250 static int check_reshape(struct mddev
*mddev
)
7252 struct r5conf
*conf
= mddev
->private;
7254 if (mddev
->delta_disks
== 0 &&
7255 mddev
->new_layout
== mddev
->layout
&&
7256 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7257 return 0; /* nothing to do */
7258 if (has_failed(conf
))
7260 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7261 /* We might be able to shrink, but the devices must
7262 * be made bigger first.
7263 * For raid6, 4 is the minimum size.
7264 * Otherwise 2 is the minimum
7267 if (mddev
->level
== 6)
7269 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7273 if (!check_stripe_cache(mddev
))
7276 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
7277 mddev
->delta_disks
> 0)
7278 if (resize_chunks(conf
,
7279 conf
->previous_raid_disks
7280 + max(0, mddev
->delta_disks
),
7281 max(mddev
->new_chunk_sectors
,
7282 mddev
->chunk_sectors
)
7285 return resize_stripes(conf
, (conf
->previous_raid_disks
7286 + mddev
->delta_disks
));
7289 static int raid5_start_reshape(struct mddev
*mddev
)
7291 struct r5conf
*conf
= mddev
->private;
7292 struct md_rdev
*rdev
;
7294 unsigned long flags
;
7296 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7299 if (!check_stripe_cache(mddev
))
7302 if (has_failed(conf
))
7305 rdev_for_each(rdev
, mddev
) {
7306 if (!test_bit(In_sync
, &rdev
->flags
)
7307 && !test_bit(Faulty
, &rdev
->flags
))
7311 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7312 /* Not enough devices even to make a degraded array
7317 /* Refuse to reduce size of the array. Any reductions in
7318 * array size must be through explicit setting of array_size
7321 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7322 < mddev
->array_sectors
) {
7323 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
7324 "before number of disks\n", mdname(mddev
));
7328 atomic_set(&conf
->reshape_stripes
, 0);
7329 spin_lock_irq(&conf
->device_lock
);
7330 write_seqcount_begin(&conf
->gen_lock
);
7331 conf
->previous_raid_disks
= conf
->raid_disks
;
7332 conf
->raid_disks
+= mddev
->delta_disks
;
7333 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7334 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7335 conf
->prev_algo
= conf
->algorithm
;
7336 conf
->algorithm
= mddev
->new_layout
;
7338 /* Code that selects data_offset needs to see the generation update
7339 * if reshape_progress has been set - so a memory barrier needed.
7342 if (mddev
->reshape_backwards
)
7343 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7345 conf
->reshape_progress
= 0;
7346 conf
->reshape_safe
= conf
->reshape_progress
;
7347 write_seqcount_end(&conf
->gen_lock
);
7348 spin_unlock_irq(&conf
->device_lock
);
7350 /* Now make sure any requests that proceeded on the assumption
7351 * the reshape wasn't running - like Discard or Read - have
7354 mddev_suspend(mddev
);
7355 mddev_resume(mddev
);
7357 /* Add some new drives, as many as will fit.
7358 * We know there are enough to make the newly sized array work.
7359 * Don't add devices if we are reducing the number of
7360 * devices in the array. This is because it is not possible
7361 * to correctly record the "partially reconstructed" state of
7362 * such devices during the reshape and confusion could result.
7364 if (mddev
->delta_disks
>= 0) {
7365 rdev_for_each(rdev
, mddev
)
7366 if (rdev
->raid_disk
< 0 &&
7367 !test_bit(Faulty
, &rdev
->flags
)) {
7368 if (raid5_add_disk(mddev
, rdev
) == 0) {
7370 >= conf
->previous_raid_disks
)
7371 set_bit(In_sync
, &rdev
->flags
);
7373 rdev
->recovery_offset
= 0;
7375 if (sysfs_link_rdev(mddev
, rdev
))
7376 /* Failure here is OK */;
7378 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7379 && !test_bit(Faulty
, &rdev
->flags
)) {
7380 /* This is a spare that was manually added */
7381 set_bit(In_sync
, &rdev
->flags
);
7384 /* When a reshape changes the number of devices,
7385 * ->degraded is measured against the larger of the
7386 * pre and post number of devices.
7388 spin_lock_irqsave(&conf
->device_lock
, flags
);
7389 mddev
->degraded
= calc_degraded(conf
);
7390 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7392 mddev
->raid_disks
= conf
->raid_disks
;
7393 mddev
->reshape_position
= conf
->reshape_progress
;
7394 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7396 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7397 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7398 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
7399 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7400 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7401 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7403 if (!mddev
->sync_thread
) {
7404 mddev
->recovery
= 0;
7405 spin_lock_irq(&conf
->device_lock
);
7406 write_seqcount_begin(&conf
->gen_lock
);
7407 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7408 mddev
->new_chunk_sectors
=
7409 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7410 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7411 rdev_for_each(rdev
, mddev
)
7412 rdev
->new_data_offset
= rdev
->data_offset
;
7414 conf
->generation
--;
7415 conf
->reshape_progress
= MaxSector
;
7416 mddev
->reshape_position
= MaxSector
;
7417 write_seqcount_end(&conf
->gen_lock
);
7418 spin_unlock_irq(&conf
->device_lock
);
7421 conf
->reshape_checkpoint
= jiffies
;
7422 md_wakeup_thread(mddev
->sync_thread
);
7423 md_new_event(mddev
);
7427 /* This is called from the reshape thread and should make any
7428 * changes needed in 'conf'
7430 static void end_reshape(struct r5conf
*conf
)
7433 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7434 struct md_rdev
*rdev
;
7436 spin_lock_irq(&conf
->device_lock
);
7437 conf
->previous_raid_disks
= conf
->raid_disks
;
7438 rdev_for_each(rdev
, conf
->mddev
)
7439 rdev
->data_offset
= rdev
->new_data_offset
;
7441 conf
->reshape_progress
= MaxSector
;
7442 conf
->mddev
->reshape_position
= MaxSector
;
7443 spin_unlock_irq(&conf
->device_lock
);
7444 wake_up(&conf
->wait_for_overlap
);
7446 /* read-ahead size must cover two whole stripes, which is
7447 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7449 if (conf
->mddev
->queue
) {
7450 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7451 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7453 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
7454 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
7459 /* This is called from the raid5d thread with mddev_lock held.
7460 * It makes config changes to the device.
7462 static void raid5_finish_reshape(struct mddev
*mddev
)
7464 struct r5conf
*conf
= mddev
->private;
7466 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7468 if (mddev
->delta_disks
> 0) {
7469 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7470 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7471 revalidate_disk(mddev
->gendisk
);
7474 spin_lock_irq(&conf
->device_lock
);
7475 mddev
->degraded
= calc_degraded(conf
);
7476 spin_unlock_irq(&conf
->device_lock
);
7477 for (d
= conf
->raid_disks
;
7478 d
< conf
->raid_disks
- mddev
->delta_disks
;
7480 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
7482 clear_bit(In_sync
, &rdev
->flags
);
7483 rdev
= conf
->disks
[d
].replacement
;
7485 clear_bit(In_sync
, &rdev
->flags
);
7488 mddev
->layout
= conf
->algorithm
;
7489 mddev
->chunk_sectors
= conf
->chunk_sectors
;
7490 mddev
->reshape_position
= MaxSector
;
7491 mddev
->delta_disks
= 0;
7492 mddev
->reshape_backwards
= 0;
7496 static void raid5_quiesce(struct mddev
*mddev
, int state
)
7498 struct r5conf
*conf
= mddev
->private;
7501 case 2: /* resume for a suspend */
7502 wake_up(&conf
->wait_for_overlap
);
7505 case 1: /* stop all writes */
7506 lock_all_device_hash_locks_irq(conf
);
7507 /* '2' tells resync/reshape to pause so that all
7508 * active stripes can drain
7511 wait_event_cmd(conf
->wait_for_quiescent
,
7512 atomic_read(&conf
->active_stripes
) == 0 &&
7513 atomic_read(&conf
->active_aligned_reads
) == 0,
7514 unlock_all_device_hash_locks_irq(conf
),
7515 lock_all_device_hash_locks_irq(conf
));
7517 unlock_all_device_hash_locks_irq(conf
);
7518 /* allow reshape to continue */
7519 wake_up(&conf
->wait_for_overlap
);
7522 case 0: /* re-enable writes */
7523 lock_all_device_hash_locks_irq(conf
);
7525 wake_up(&conf
->wait_for_quiescent
);
7526 wake_up(&conf
->wait_for_overlap
);
7527 unlock_all_device_hash_locks_irq(conf
);
7532 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
7534 struct r0conf
*raid0_conf
= mddev
->private;
7537 /* for raid0 takeover only one zone is supported */
7538 if (raid0_conf
->nr_strip_zones
> 1) {
7539 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7541 return ERR_PTR(-EINVAL
);
7544 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
7545 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
7546 mddev
->dev_sectors
= sectors
;
7547 mddev
->new_level
= level
;
7548 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7549 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
7550 mddev
->raid_disks
+= 1;
7551 mddev
->delta_disks
= 1;
7552 /* make sure it will be not marked as dirty */
7553 mddev
->recovery_cp
= MaxSector
;
7555 return setup_conf(mddev
);
7558 static void *raid5_takeover_raid1(struct mddev
*mddev
)
7562 if (mddev
->raid_disks
!= 2 ||
7563 mddev
->degraded
> 1)
7564 return ERR_PTR(-EINVAL
);
7566 /* Should check if there are write-behind devices? */
7568 chunksect
= 64*2; /* 64K by default */
7570 /* The array must be an exact multiple of chunksize */
7571 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
7574 if ((chunksect
<<9) < STRIPE_SIZE
)
7575 /* array size does not allow a suitable chunk size */
7576 return ERR_PTR(-EINVAL
);
7578 mddev
->new_level
= 5;
7579 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7580 mddev
->new_chunk_sectors
= chunksect
;
7582 return setup_conf(mddev
);
7585 static void *raid5_takeover_raid6(struct mddev
*mddev
)
7589 switch (mddev
->layout
) {
7590 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7591 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
7593 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7594 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
7596 case ALGORITHM_LEFT_SYMMETRIC_6
:
7597 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
7599 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7600 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
7602 case ALGORITHM_PARITY_0_6
:
7603 new_layout
= ALGORITHM_PARITY_0
;
7605 case ALGORITHM_PARITY_N
:
7606 new_layout
= ALGORITHM_PARITY_N
;
7609 return ERR_PTR(-EINVAL
);
7611 mddev
->new_level
= 5;
7612 mddev
->new_layout
= new_layout
;
7613 mddev
->delta_disks
= -1;
7614 mddev
->raid_disks
-= 1;
7615 return setup_conf(mddev
);
7618 static int raid5_check_reshape(struct mddev
*mddev
)
7620 /* For a 2-drive array, the layout and chunk size can be changed
7621 * immediately as not restriping is needed.
7622 * For larger arrays we record the new value - after validation
7623 * to be used by a reshape pass.
7625 struct r5conf
*conf
= mddev
->private;
7626 int new_chunk
= mddev
->new_chunk_sectors
;
7628 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
7630 if (new_chunk
> 0) {
7631 if (!is_power_of_2(new_chunk
))
7633 if (new_chunk
< (PAGE_SIZE
>>9))
7635 if (mddev
->array_sectors
& (new_chunk
-1))
7636 /* not factor of array size */
7640 /* They look valid */
7642 if (mddev
->raid_disks
== 2) {
7643 /* can make the change immediately */
7644 if (mddev
->new_layout
>= 0) {
7645 conf
->algorithm
= mddev
->new_layout
;
7646 mddev
->layout
= mddev
->new_layout
;
7648 if (new_chunk
> 0) {
7649 conf
->chunk_sectors
= new_chunk
;
7650 mddev
->chunk_sectors
= new_chunk
;
7652 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
7653 md_wakeup_thread(mddev
->thread
);
7655 return check_reshape(mddev
);
7658 static int raid6_check_reshape(struct mddev
*mddev
)
7660 int new_chunk
= mddev
->new_chunk_sectors
;
7662 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
7664 if (new_chunk
> 0) {
7665 if (!is_power_of_2(new_chunk
))
7667 if (new_chunk
< (PAGE_SIZE
>> 9))
7669 if (mddev
->array_sectors
& (new_chunk
-1))
7670 /* not factor of array size */
7674 /* They look valid */
7675 return check_reshape(mddev
);
7678 static void *raid5_takeover(struct mddev
*mddev
)
7680 /* raid5 can take over:
7681 * raid0 - if there is only one strip zone - make it a raid4 layout
7682 * raid1 - if there are two drives. We need to know the chunk size
7683 * raid4 - trivial - just use a raid4 layout.
7684 * raid6 - Providing it is a *_6 layout
7686 if (mddev
->level
== 0)
7687 return raid45_takeover_raid0(mddev
, 5);
7688 if (mddev
->level
== 1)
7689 return raid5_takeover_raid1(mddev
);
7690 if (mddev
->level
== 4) {
7691 mddev
->new_layout
= ALGORITHM_PARITY_N
;
7692 mddev
->new_level
= 5;
7693 return setup_conf(mddev
);
7695 if (mddev
->level
== 6)
7696 return raid5_takeover_raid6(mddev
);
7698 return ERR_PTR(-EINVAL
);
7701 static void *raid4_takeover(struct mddev
*mddev
)
7703 /* raid4 can take over:
7704 * raid0 - if there is only one strip zone
7705 * raid5 - if layout is right
7707 if (mddev
->level
== 0)
7708 return raid45_takeover_raid0(mddev
, 4);
7709 if (mddev
->level
== 5 &&
7710 mddev
->layout
== ALGORITHM_PARITY_N
) {
7711 mddev
->new_layout
= 0;
7712 mddev
->new_level
= 4;
7713 return setup_conf(mddev
);
7715 return ERR_PTR(-EINVAL
);
7718 static struct md_personality raid5_personality
;
7720 static void *raid6_takeover(struct mddev
*mddev
)
7722 /* Currently can only take over a raid5. We map the
7723 * personality to an equivalent raid6 personality
7724 * with the Q block at the end.
7728 if (mddev
->pers
!= &raid5_personality
)
7729 return ERR_PTR(-EINVAL
);
7730 if (mddev
->degraded
> 1)
7731 return ERR_PTR(-EINVAL
);
7732 if (mddev
->raid_disks
> 253)
7733 return ERR_PTR(-EINVAL
);
7734 if (mddev
->raid_disks
< 3)
7735 return ERR_PTR(-EINVAL
);
7737 switch (mddev
->layout
) {
7738 case ALGORITHM_LEFT_ASYMMETRIC
:
7739 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
7741 case ALGORITHM_RIGHT_ASYMMETRIC
:
7742 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
7744 case ALGORITHM_LEFT_SYMMETRIC
:
7745 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
7747 case ALGORITHM_RIGHT_SYMMETRIC
:
7748 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
7750 case ALGORITHM_PARITY_0
:
7751 new_layout
= ALGORITHM_PARITY_0_6
;
7753 case ALGORITHM_PARITY_N
:
7754 new_layout
= ALGORITHM_PARITY_N
;
7757 return ERR_PTR(-EINVAL
);
7759 mddev
->new_level
= 6;
7760 mddev
->new_layout
= new_layout
;
7761 mddev
->delta_disks
= 1;
7762 mddev
->raid_disks
+= 1;
7763 return setup_conf(mddev
);
7766 static struct md_personality raid6_personality
=
7770 .owner
= THIS_MODULE
,
7771 .make_request
= make_request
,
7775 .error_handler
= error
,
7776 .hot_add_disk
= raid5_add_disk
,
7777 .hot_remove_disk
= raid5_remove_disk
,
7778 .spare_active
= raid5_spare_active
,
7779 .sync_request
= sync_request
,
7780 .resize
= raid5_resize
,
7782 .check_reshape
= raid6_check_reshape
,
7783 .start_reshape
= raid5_start_reshape
,
7784 .finish_reshape
= raid5_finish_reshape
,
7785 .quiesce
= raid5_quiesce
,
7786 .takeover
= raid6_takeover
,
7787 .congested
= raid5_congested
,
7789 static struct md_personality raid5_personality
=
7793 .owner
= THIS_MODULE
,
7794 .make_request
= make_request
,
7798 .error_handler
= error
,
7799 .hot_add_disk
= raid5_add_disk
,
7800 .hot_remove_disk
= raid5_remove_disk
,
7801 .spare_active
= raid5_spare_active
,
7802 .sync_request
= sync_request
,
7803 .resize
= raid5_resize
,
7805 .check_reshape
= raid5_check_reshape
,
7806 .start_reshape
= raid5_start_reshape
,
7807 .finish_reshape
= raid5_finish_reshape
,
7808 .quiesce
= raid5_quiesce
,
7809 .takeover
= raid5_takeover
,
7810 .congested
= raid5_congested
,
7813 static struct md_personality raid4_personality
=
7817 .owner
= THIS_MODULE
,
7818 .make_request
= make_request
,
7822 .error_handler
= error
,
7823 .hot_add_disk
= raid5_add_disk
,
7824 .hot_remove_disk
= raid5_remove_disk
,
7825 .spare_active
= raid5_spare_active
,
7826 .sync_request
= sync_request
,
7827 .resize
= raid5_resize
,
7829 .check_reshape
= raid5_check_reshape
,
7830 .start_reshape
= raid5_start_reshape
,
7831 .finish_reshape
= raid5_finish_reshape
,
7832 .quiesce
= raid5_quiesce
,
7833 .takeover
= raid4_takeover
,
7834 .congested
= raid5_congested
,
7837 static int __init
raid5_init(void)
7839 raid5_wq
= alloc_workqueue("raid5wq",
7840 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
7843 register_md_personality(&raid6_personality
);
7844 register_md_personality(&raid5_personality
);
7845 register_md_personality(&raid4_personality
);
7849 static void raid5_exit(void)
7851 unregister_md_personality(&raid6_personality
);
7852 unregister_md_personality(&raid5_personality
);
7853 unregister_md_personality(&raid4_personality
);
7854 destroy_workqueue(raid5_wq
);
7857 module_init(raid5_init
);
7858 module_exit(raid5_exit
);
7859 MODULE_LICENSE("GPL");
7860 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7861 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7862 MODULE_ALIAS("md-raid5");
7863 MODULE_ALIAS("md-raid4");
7864 MODULE_ALIAS("md-level-5");
7865 MODULE_ALIAS("md-level-4");
7866 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7867 MODULE_ALIAS("md-raid6");
7868 MODULE_ALIAS("md-level-6");
7870 /* This used to be two separate modules, they were: */
7871 MODULE_ALIAS("raid5");
7872 MODULE_ALIAS("raid6");