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 <trace/events/block.h>
64 #define cpu_to_group(cpu) cpu_to_node(cpu)
65 #define ANY_GROUP NUMA_NO_NODE
67 static struct workqueue_struct
*raid5_wq
;
72 #define NR_STRIPES 256
73 #define STRIPE_SIZE PAGE_SIZE
74 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
75 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
76 #define IO_THRESHOLD 1
77 #define BYPASS_THRESHOLD 1
78 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
79 #define HASH_MASK (NR_HASH - 1)
80 #define MAX_STRIPE_BATCH 8
82 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
84 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
85 return &conf
->stripe_hashtbl
[hash
];
88 static inline int stripe_hash_locks_hash(sector_t sect
)
90 return (sect
>> STRIPE_SHIFT
) & STRIPE_HASH_LOCKS_MASK
;
93 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
95 spin_lock_irq(conf
->hash_locks
+ hash
);
96 spin_lock(&conf
->device_lock
);
99 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
101 spin_unlock(&conf
->device_lock
);
102 spin_unlock_irq(conf
->hash_locks
+ hash
);
105 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
109 spin_lock(conf
->hash_locks
);
110 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
111 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
112 spin_lock(&conf
->device_lock
);
115 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
118 spin_unlock(&conf
->device_lock
);
119 for (i
= NR_STRIPE_HASH_LOCKS
; i
; i
--)
120 spin_unlock(conf
->hash_locks
+ i
- 1);
124 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
125 * order without overlap. There may be several bio's per stripe+device, and
126 * a bio could span several devices.
127 * When walking this list for a particular stripe+device, we must never proceed
128 * beyond a bio that extends past this device, as the next bio might no longer
130 * This function is used to determine the 'next' bio in the list, given the sector
131 * of the current stripe+device
133 static inline struct bio
*r5_next_bio(struct bio
*bio
, sector_t sector
)
135 int sectors
= bio_sectors(bio
);
136 if (bio
->bi_iter
.bi_sector
+ sectors
< sector
+ STRIPE_SECTORS
)
143 * We maintain a biased count of active stripes in the bottom 16 bits of
144 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
146 static inline int raid5_bi_processed_stripes(struct bio
*bio
)
148 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
149 return (atomic_read(segments
) >> 16) & 0xffff;
152 static inline int raid5_dec_bi_active_stripes(struct bio
*bio
)
154 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
155 return atomic_sub_return(1, segments
) & 0xffff;
158 static inline void raid5_inc_bi_active_stripes(struct bio
*bio
)
160 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
161 atomic_inc(segments
);
164 static inline void raid5_set_bi_processed_stripes(struct bio
*bio
,
167 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
171 old
= atomic_read(segments
);
172 new = (old
& 0xffff) | (cnt
<< 16);
173 } while (atomic_cmpxchg(segments
, old
, new) != old
);
176 static inline void raid5_set_bi_stripes(struct bio
*bio
, unsigned int cnt
)
178 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
179 atomic_set(segments
, cnt
);
182 /* Find first data disk in a raid6 stripe */
183 static inline int raid6_d0(struct stripe_head
*sh
)
186 /* ddf always start from first device */
188 /* md starts just after Q block */
189 if (sh
->qd_idx
== sh
->disks
- 1)
192 return sh
->qd_idx
+ 1;
194 static inline int raid6_next_disk(int disk
, int raid_disks
)
197 return (disk
< raid_disks
) ? disk
: 0;
200 /* When walking through the disks in a raid5, starting at raid6_d0,
201 * We need to map each disk to a 'slot', where the data disks are slot
202 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
203 * is raid_disks-1. This help does that mapping.
205 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
206 int *count
, int syndrome_disks
)
212 if (idx
== sh
->pd_idx
)
213 return syndrome_disks
;
214 if (idx
== sh
->qd_idx
)
215 return syndrome_disks
+ 1;
221 static void return_io(struct bio
*return_bi
)
223 struct bio
*bi
= return_bi
;
226 return_bi
= bi
->bi_next
;
228 bi
->bi_iter
.bi_size
= 0;
229 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
236 static void print_raid5_conf (struct r5conf
*conf
);
238 static int stripe_operations_active(struct stripe_head
*sh
)
240 return sh
->check_state
|| sh
->reconstruct_state
||
241 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
242 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
245 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
247 struct r5conf
*conf
= sh
->raid_conf
;
248 struct r5worker_group
*group
;
250 int i
, cpu
= sh
->cpu
;
252 if (!cpu_online(cpu
)) {
253 cpu
= cpumask_any(cpu_online_mask
);
257 if (list_empty(&sh
->lru
)) {
258 struct r5worker_group
*group
;
259 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
260 list_add_tail(&sh
->lru
, &group
->handle_list
);
261 group
->stripes_cnt
++;
265 if (conf
->worker_cnt_per_group
== 0) {
266 md_wakeup_thread(conf
->mddev
->thread
);
270 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
272 group
->workers
[0].working
= true;
273 /* at least one worker should run to avoid race */
274 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
276 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
277 /* wakeup more workers */
278 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
279 if (group
->workers
[i
].working
== false) {
280 group
->workers
[i
].working
= true;
281 queue_work_on(sh
->cpu
, raid5_wq
,
282 &group
->workers
[i
].work
);
288 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
289 struct list_head
*temp_inactive_list
)
291 BUG_ON(!list_empty(&sh
->lru
));
292 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
293 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
294 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
295 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
296 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
297 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
298 sh
->bm_seq
- conf
->seq_write
> 0)
299 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
301 clear_bit(STRIPE_DELAYED
, &sh
->state
);
302 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
303 if (conf
->worker_cnt_per_group
== 0) {
304 list_add_tail(&sh
->lru
, &conf
->handle_list
);
306 raid5_wakeup_stripe_thread(sh
);
310 md_wakeup_thread(conf
->mddev
->thread
);
312 BUG_ON(stripe_operations_active(sh
));
313 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
314 if (atomic_dec_return(&conf
->preread_active_stripes
)
316 md_wakeup_thread(conf
->mddev
->thread
);
317 atomic_dec(&conf
->active_stripes
);
318 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
))
319 list_add_tail(&sh
->lru
, temp_inactive_list
);
323 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
324 struct list_head
*temp_inactive_list
)
326 if (atomic_dec_and_test(&sh
->count
))
327 do_release_stripe(conf
, sh
, temp_inactive_list
);
331 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
333 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
334 * given time. Adding stripes only takes device lock, while deleting stripes
335 * only takes hash lock.
337 static void release_inactive_stripe_list(struct r5conf
*conf
,
338 struct list_head
*temp_inactive_list
,
342 bool do_wakeup
= false;
345 if (hash
== NR_STRIPE_HASH_LOCKS
) {
346 size
= NR_STRIPE_HASH_LOCKS
;
347 hash
= NR_STRIPE_HASH_LOCKS
- 1;
351 struct list_head
*list
= &temp_inactive_list
[size
- 1];
354 * We don't hold any lock here yet, get_active_stripe() might
355 * remove stripes from the list
357 if (!list_empty_careful(list
)) {
358 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
359 if (list_empty(conf
->inactive_list
+ hash
) &&
361 atomic_dec(&conf
->empty_inactive_list_nr
);
362 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
364 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
371 wake_up(&conf
->wait_for_stripe
);
372 if (conf
->retry_read_aligned
)
373 md_wakeup_thread(conf
->mddev
->thread
);
377 /* should hold conf->device_lock already */
378 static int release_stripe_list(struct r5conf
*conf
,
379 struct list_head
*temp_inactive_list
)
381 struct stripe_head
*sh
;
383 struct llist_node
*head
;
385 head
= llist_del_all(&conf
->released_stripes
);
386 head
= llist_reverse_order(head
);
390 sh
= llist_entry(head
, struct stripe_head
, release_list
);
391 head
= llist_next(head
);
392 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
394 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
396 * Don't worry the bit is set here, because if the bit is set
397 * again, the count is always > 1. This is true for
398 * STRIPE_ON_UNPLUG_LIST bit too.
400 hash
= sh
->hash_lock_index
;
401 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
408 static void release_stripe(struct stripe_head
*sh
)
410 struct r5conf
*conf
= sh
->raid_conf
;
412 struct list_head list
;
416 if (unlikely(!conf
->mddev
->thread
) ||
417 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
419 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
421 md_wakeup_thread(conf
->mddev
->thread
);
424 local_irq_save(flags
);
425 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
426 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
427 INIT_LIST_HEAD(&list
);
428 hash
= sh
->hash_lock_index
;
429 do_release_stripe(conf
, sh
, &list
);
430 spin_unlock(&conf
->device_lock
);
431 release_inactive_stripe_list(conf
, &list
, hash
);
433 local_irq_restore(flags
);
436 static inline void remove_hash(struct stripe_head
*sh
)
438 pr_debug("remove_hash(), stripe %llu\n",
439 (unsigned long long)sh
->sector
);
441 hlist_del_init(&sh
->hash
);
444 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
446 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
448 pr_debug("insert_hash(), stripe %llu\n",
449 (unsigned long long)sh
->sector
);
451 hlist_add_head(&sh
->hash
, hp
);
455 /* find an idle stripe, make sure it is unhashed, and return it. */
456 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
458 struct stripe_head
*sh
= NULL
;
459 struct list_head
*first
;
461 if (list_empty(conf
->inactive_list
+ hash
))
463 first
= (conf
->inactive_list
+ hash
)->next
;
464 sh
= list_entry(first
, struct stripe_head
, lru
);
465 list_del_init(first
);
467 atomic_inc(&conf
->active_stripes
);
468 BUG_ON(hash
!= sh
->hash_lock_index
);
469 if (list_empty(conf
->inactive_list
+ hash
))
470 atomic_inc(&conf
->empty_inactive_list_nr
);
475 static void shrink_buffers(struct stripe_head
*sh
)
479 int num
= sh
->raid_conf
->pool_size
;
481 for (i
= 0; i
< num
; i
++) {
485 sh
->dev
[i
].page
= NULL
;
490 static int grow_buffers(struct stripe_head
*sh
)
493 int num
= sh
->raid_conf
->pool_size
;
495 for (i
= 0; i
< num
; i
++) {
498 if (!(page
= alloc_page(GFP_KERNEL
))) {
501 sh
->dev
[i
].page
= page
;
506 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
507 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
508 struct stripe_head
*sh
);
510 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
512 struct r5conf
*conf
= sh
->raid_conf
;
515 BUG_ON(atomic_read(&sh
->count
) != 0);
516 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
517 BUG_ON(stripe_operations_active(sh
));
519 pr_debug("init_stripe called, stripe %llu\n",
520 (unsigned long long)sh
->sector
);
524 seq
= read_seqcount_begin(&conf
->gen_lock
);
525 sh
->generation
= conf
->generation
- previous
;
526 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
528 stripe_set_idx(sector
, conf
, previous
, sh
);
532 for (i
= sh
->disks
; i
--; ) {
533 struct r5dev
*dev
= &sh
->dev
[i
];
535 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
536 test_bit(R5_LOCKED
, &dev
->flags
)) {
537 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
538 (unsigned long long)sh
->sector
, i
, dev
->toread
,
539 dev
->read
, dev
->towrite
, dev
->written
,
540 test_bit(R5_LOCKED
, &dev
->flags
));
544 raid5_build_block(sh
, i
, previous
);
546 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
548 insert_hash(conf
, sh
);
549 sh
->cpu
= smp_processor_id();
552 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
555 struct stripe_head
*sh
;
557 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
558 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
559 if (sh
->sector
== sector
&& sh
->generation
== generation
)
561 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
566 * Need to check if array has failed when deciding whether to:
568 * - remove non-faulty devices
571 * This determination is simple when no reshape is happening.
572 * However if there is a reshape, we need to carefully check
573 * both the before and after sections.
574 * This is because some failed devices may only affect one
575 * of the two sections, and some non-in_sync devices may
576 * be insync in the section most affected by failed devices.
578 static int calc_degraded(struct r5conf
*conf
)
580 int degraded
, degraded2
;
585 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
586 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
587 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
588 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
589 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
591 else if (test_bit(In_sync
, &rdev
->flags
))
594 /* not in-sync or faulty.
595 * If the reshape increases the number of devices,
596 * this is being recovered by the reshape, so
597 * this 'previous' section is not in_sync.
598 * If the number of devices is being reduced however,
599 * the device can only be part of the array if
600 * we are reverting a reshape, so this section will
603 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
607 if (conf
->raid_disks
== conf
->previous_raid_disks
)
611 for (i
= 0; i
< conf
->raid_disks
; i
++) {
612 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
613 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
614 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
615 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
617 else if (test_bit(In_sync
, &rdev
->flags
))
620 /* not in-sync or faulty.
621 * If reshape increases the number of devices, this
622 * section has already been recovered, else it
623 * almost certainly hasn't.
625 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
629 if (degraded2
> degraded
)
634 static int has_failed(struct r5conf
*conf
)
638 if (conf
->mddev
->reshape_position
== MaxSector
)
639 return conf
->mddev
->degraded
> conf
->max_degraded
;
641 degraded
= calc_degraded(conf
);
642 if (degraded
> conf
->max_degraded
)
647 static struct stripe_head
*
648 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
649 int previous
, int noblock
, int noquiesce
)
651 struct stripe_head
*sh
;
652 int hash
= stripe_hash_locks_hash(sector
);
654 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
656 spin_lock_irq(conf
->hash_locks
+ hash
);
659 wait_event_lock_irq(conf
->wait_for_stripe
,
660 conf
->quiesce
== 0 || noquiesce
,
661 *(conf
->hash_locks
+ hash
));
662 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
664 if (!conf
->inactive_blocked
)
665 sh
= get_free_stripe(conf
, hash
);
666 if (noblock
&& sh
== NULL
)
669 conf
->inactive_blocked
= 1;
671 conf
->wait_for_stripe
,
672 !list_empty(conf
->inactive_list
+ hash
) &&
673 (atomic_read(&conf
->active_stripes
)
674 < (conf
->max_nr_stripes
* 3 / 4)
675 || !conf
->inactive_blocked
),
676 *(conf
->hash_locks
+ hash
));
677 conf
->inactive_blocked
= 0;
679 init_stripe(sh
, sector
, previous
);
681 if (atomic_read(&sh
->count
)) {
682 BUG_ON(!list_empty(&sh
->lru
)
683 && !test_bit(STRIPE_EXPANDING
, &sh
->state
)
684 && !test_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
)
685 && !test_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
));
687 spin_lock(&conf
->device_lock
);
688 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
689 atomic_inc(&conf
->active_stripes
);
690 if (list_empty(&sh
->lru
) &&
691 !test_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
) &&
692 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
694 list_del_init(&sh
->lru
);
696 sh
->group
->stripes_cnt
--;
699 spin_unlock(&conf
->device_lock
);
702 } while (sh
== NULL
);
705 atomic_inc(&sh
->count
);
707 spin_unlock_irq(conf
->hash_locks
+ hash
);
711 /* Determine if 'data_offset' or 'new_data_offset' should be used
712 * in this stripe_head.
714 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
716 sector_t progress
= conf
->reshape_progress
;
717 /* Need a memory barrier to make sure we see the value
718 * of conf->generation, or ->data_offset that was set before
719 * reshape_progress was updated.
722 if (progress
== MaxSector
)
724 if (sh
->generation
== conf
->generation
- 1)
726 /* We are in a reshape, and this is a new-generation stripe,
727 * so use new_data_offset.
733 raid5_end_read_request(struct bio
*bi
, int error
);
735 raid5_end_write_request(struct bio
*bi
, int error
);
737 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
739 struct r5conf
*conf
= sh
->raid_conf
;
740 int i
, disks
= sh
->disks
;
744 for (i
= disks
; i
--; ) {
746 int replace_only
= 0;
747 struct bio
*bi
, *rbi
;
748 struct md_rdev
*rdev
, *rrdev
= NULL
;
749 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
750 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
754 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
756 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
758 else if (test_and_clear_bit(R5_WantReplace
,
759 &sh
->dev
[i
].flags
)) {
764 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
767 bi
= &sh
->dev
[i
].req
;
768 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
771 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
772 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
773 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
782 /* We raced and saw duplicates */
785 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
) && rrdev
)
790 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
793 atomic_inc(&rdev
->nr_pending
);
794 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
797 atomic_inc(&rrdev
->nr_pending
);
800 /* We have already checked bad blocks for reads. Now
801 * need to check for writes. We never accept write errors
802 * on the replacement, so we don't to check rrdev.
804 while ((rw
& WRITE
) && rdev
&&
805 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
808 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
809 &first_bad
, &bad_sectors
);
814 set_bit(BlockedBadBlocks
, &rdev
->flags
);
815 if (!conf
->mddev
->external
&&
816 conf
->mddev
->flags
) {
817 /* It is very unlikely, but we might
818 * still need to write out the
819 * bad block log - better give it
821 md_check_recovery(conf
->mddev
);
824 * Because md_wait_for_blocked_rdev
825 * will dec nr_pending, we must
826 * increment it first.
828 atomic_inc(&rdev
->nr_pending
);
829 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
831 /* Acknowledged bad block - skip the write */
832 rdev_dec_pending(rdev
, conf
->mddev
);
838 if (s
->syncing
|| s
->expanding
|| s
->expanded
840 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
842 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
845 bi
->bi_bdev
= rdev
->bdev
;
847 bi
->bi_end_io
= (rw
& WRITE
)
848 ? raid5_end_write_request
849 : raid5_end_read_request
;
852 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
853 __func__
, (unsigned long long)sh
->sector
,
855 atomic_inc(&sh
->count
);
856 if (use_new_offset(conf
, sh
))
857 bi
->bi_iter
.bi_sector
= (sh
->sector
858 + rdev
->new_data_offset
);
860 bi
->bi_iter
.bi_sector
= (sh
->sector
861 + rdev
->data_offset
);
862 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
863 bi
->bi_rw
|= REQ_NOMERGE
;
866 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
867 bi
->bi_io_vec
[0].bv_offset
= 0;
868 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
870 * If this is discard request, set bi_vcnt 0. We don't
871 * want to confuse SCSI because SCSI will replace payload
873 if (rw
& REQ_DISCARD
)
876 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
878 if (conf
->mddev
->gendisk
)
879 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
880 bi
, disk_devt(conf
->mddev
->gendisk
),
882 generic_make_request(bi
);
885 if (s
->syncing
|| s
->expanding
|| s
->expanded
887 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
889 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
892 rbi
->bi_bdev
= rrdev
->bdev
;
894 BUG_ON(!(rw
& WRITE
));
895 rbi
->bi_end_io
= raid5_end_write_request
;
896 rbi
->bi_private
= sh
;
898 pr_debug("%s: for %llu schedule op %ld on "
899 "replacement disc %d\n",
900 __func__
, (unsigned long long)sh
->sector
,
902 atomic_inc(&sh
->count
);
903 if (use_new_offset(conf
, sh
))
904 rbi
->bi_iter
.bi_sector
= (sh
->sector
905 + rrdev
->new_data_offset
);
907 rbi
->bi_iter
.bi_sector
= (sh
->sector
908 + rrdev
->data_offset
);
910 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
911 rbi
->bi_io_vec
[0].bv_offset
= 0;
912 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
914 * If this is discard request, set bi_vcnt 0. We don't
915 * want to confuse SCSI because SCSI will replace payload
917 if (rw
& REQ_DISCARD
)
919 if (conf
->mddev
->gendisk
)
920 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
921 rbi
, disk_devt(conf
->mddev
->gendisk
),
923 generic_make_request(rbi
);
925 if (!rdev
&& !rrdev
) {
927 set_bit(STRIPE_DEGRADED
, &sh
->state
);
928 pr_debug("skip op %ld on disc %d for sector %llu\n",
929 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
930 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
931 set_bit(STRIPE_HANDLE
, &sh
->state
);
936 static struct dma_async_tx_descriptor
*
937 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
938 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
941 struct page
*bio_page
;
944 struct async_submit_ctl submit
;
945 enum async_tx_flags flags
= 0;
947 if (bio
->bi_iter
.bi_sector
>= sector
)
948 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
950 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
953 flags
|= ASYNC_TX_FENCE
;
954 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
956 bio_for_each_segment(bvl
, bio
, i
) {
957 int len
= bvl
->bv_len
;
961 if (page_offset
< 0) {
962 b_offset
= -page_offset
;
963 page_offset
+= b_offset
;
967 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
968 clen
= STRIPE_SIZE
- page_offset
;
973 b_offset
+= bvl
->bv_offset
;
974 bio_page
= bvl
->bv_page
;
976 tx
= async_memcpy(page
, bio_page
, page_offset
,
977 b_offset
, clen
, &submit
);
979 tx
= async_memcpy(bio_page
, page
, b_offset
,
980 page_offset
, clen
, &submit
);
982 /* chain the operations */
983 submit
.depend_tx
= tx
;
985 if (clen
< len
) /* hit end of page */
993 static void ops_complete_biofill(void *stripe_head_ref
)
995 struct stripe_head
*sh
= stripe_head_ref
;
996 struct bio
*return_bi
= NULL
;
999 pr_debug("%s: stripe %llu\n", __func__
,
1000 (unsigned long long)sh
->sector
);
1002 /* clear completed biofills */
1003 for (i
= sh
->disks
; i
--; ) {
1004 struct r5dev
*dev
= &sh
->dev
[i
];
1006 /* acknowledge completion of a biofill operation */
1007 /* and check if we need to reply to a read request,
1008 * new R5_Wantfill requests are held off until
1009 * !STRIPE_BIOFILL_RUN
1011 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1012 struct bio
*rbi
, *rbi2
;
1017 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1018 dev
->sector
+ STRIPE_SECTORS
) {
1019 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1020 if (!raid5_dec_bi_active_stripes(rbi
)) {
1021 rbi
->bi_next
= return_bi
;
1028 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1030 return_io(return_bi
);
1032 set_bit(STRIPE_HANDLE
, &sh
->state
);
1036 static void ops_run_biofill(struct stripe_head
*sh
)
1038 struct dma_async_tx_descriptor
*tx
= NULL
;
1039 struct async_submit_ctl submit
;
1042 pr_debug("%s: stripe %llu\n", __func__
,
1043 (unsigned long long)sh
->sector
);
1045 for (i
= sh
->disks
; i
--; ) {
1046 struct r5dev
*dev
= &sh
->dev
[i
];
1047 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1049 spin_lock_irq(&sh
->stripe_lock
);
1050 dev
->read
= rbi
= dev
->toread
;
1052 spin_unlock_irq(&sh
->stripe_lock
);
1053 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1054 dev
->sector
+ STRIPE_SECTORS
) {
1055 tx
= async_copy_data(0, rbi
, dev
->page
,
1057 rbi
= r5_next_bio(rbi
, dev
->sector
);
1062 atomic_inc(&sh
->count
);
1063 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1064 async_trigger_callback(&submit
);
1067 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1074 tgt
= &sh
->dev
[target
];
1075 set_bit(R5_UPTODATE
, &tgt
->flags
);
1076 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1077 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1080 static void ops_complete_compute(void *stripe_head_ref
)
1082 struct stripe_head
*sh
= stripe_head_ref
;
1084 pr_debug("%s: stripe %llu\n", __func__
,
1085 (unsigned long long)sh
->sector
);
1087 /* mark the computed target(s) as uptodate */
1088 mark_target_uptodate(sh
, sh
->ops
.target
);
1089 mark_target_uptodate(sh
, sh
->ops
.target2
);
1091 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1092 if (sh
->check_state
== check_state_compute_run
)
1093 sh
->check_state
= check_state_compute_result
;
1094 set_bit(STRIPE_HANDLE
, &sh
->state
);
1098 /* return a pointer to the address conversion region of the scribble buffer */
1099 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1100 struct raid5_percpu
*percpu
)
1102 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1105 static struct dma_async_tx_descriptor
*
1106 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1108 int disks
= sh
->disks
;
1109 struct page
**xor_srcs
= percpu
->scribble
;
1110 int target
= sh
->ops
.target
;
1111 struct r5dev
*tgt
= &sh
->dev
[target
];
1112 struct page
*xor_dest
= tgt
->page
;
1114 struct dma_async_tx_descriptor
*tx
;
1115 struct async_submit_ctl submit
;
1118 pr_debug("%s: stripe %llu block: %d\n",
1119 __func__
, (unsigned long long)sh
->sector
, target
);
1120 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1122 for (i
= disks
; i
--; )
1124 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1126 atomic_inc(&sh
->count
);
1128 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1129 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
1130 if (unlikely(count
== 1))
1131 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1133 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1138 /* set_syndrome_sources - populate source buffers for gen_syndrome
1139 * @srcs - (struct page *) array of size sh->disks
1140 * @sh - stripe_head to parse
1142 * Populates srcs in proper layout order for the stripe and returns the
1143 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1144 * destination buffer is recorded in srcs[count] and the Q destination
1145 * is recorded in srcs[count+1]].
1147 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
1149 int disks
= sh
->disks
;
1150 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1151 int d0_idx
= raid6_d0(sh
);
1155 for (i
= 0; i
< disks
; i
++)
1161 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1163 srcs
[slot
] = sh
->dev
[i
].page
;
1164 i
= raid6_next_disk(i
, disks
);
1165 } while (i
!= d0_idx
);
1167 return syndrome_disks
;
1170 static struct dma_async_tx_descriptor
*
1171 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1173 int disks
= sh
->disks
;
1174 struct page
**blocks
= percpu
->scribble
;
1176 int qd_idx
= sh
->qd_idx
;
1177 struct dma_async_tx_descriptor
*tx
;
1178 struct async_submit_ctl submit
;
1184 if (sh
->ops
.target
< 0)
1185 target
= sh
->ops
.target2
;
1186 else if (sh
->ops
.target2
< 0)
1187 target
= sh
->ops
.target
;
1189 /* we should only have one valid target */
1192 pr_debug("%s: stripe %llu block: %d\n",
1193 __func__
, (unsigned long long)sh
->sector
, target
);
1195 tgt
= &sh
->dev
[target
];
1196 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1199 atomic_inc(&sh
->count
);
1201 if (target
== qd_idx
) {
1202 count
= set_syndrome_sources(blocks
, sh
);
1203 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1204 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1205 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1206 ops_complete_compute
, sh
,
1207 to_addr_conv(sh
, percpu
));
1208 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1210 /* Compute any data- or p-drive using XOR */
1212 for (i
= disks
; i
-- ; ) {
1213 if (i
== target
|| i
== qd_idx
)
1215 blocks
[count
++] = sh
->dev
[i
].page
;
1218 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1219 NULL
, ops_complete_compute
, sh
,
1220 to_addr_conv(sh
, percpu
));
1221 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1227 static struct dma_async_tx_descriptor
*
1228 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1230 int i
, count
, disks
= sh
->disks
;
1231 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1232 int d0_idx
= raid6_d0(sh
);
1233 int faila
= -1, failb
= -1;
1234 int target
= sh
->ops
.target
;
1235 int target2
= sh
->ops
.target2
;
1236 struct r5dev
*tgt
= &sh
->dev
[target
];
1237 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1238 struct dma_async_tx_descriptor
*tx
;
1239 struct page
**blocks
= percpu
->scribble
;
1240 struct async_submit_ctl submit
;
1242 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1243 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1244 BUG_ON(target
< 0 || target2
< 0);
1245 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1246 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1248 /* we need to open-code set_syndrome_sources to handle the
1249 * slot number conversion for 'faila' and 'failb'
1251 for (i
= 0; i
< disks
; i
++)
1256 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1258 blocks
[slot
] = sh
->dev
[i
].page
;
1264 i
= raid6_next_disk(i
, disks
);
1265 } while (i
!= d0_idx
);
1267 BUG_ON(faila
== failb
);
1270 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1271 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1273 atomic_inc(&sh
->count
);
1275 if (failb
== syndrome_disks
+1) {
1276 /* Q disk is one of the missing disks */
1277 if (faila
== syndrome_disks
) {
1278 /* Missing P+Q, just recompute */
1279 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1280 ops_complete_compute
, sh
,
1281 to_addr_conv(sh
, percpu
));
1282 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1283 STRIPE_SIZE
, &submit
);
1287 int qd_idx
= sh
->qd_idx
;
1289 /* Missing D+Q: recompute D from P, then recompute Q */
1290 if (target
== qd_idx
)
1291 data_target
= target2
;
1293 data_target
= target
;
1296 for (i
= disks
; i
-- ; ) {
1297 if (i
== data_target
|| i
== qd_idx
)
1299 blocks
[count
++] = sh
->dev
[i
].page
;
1301 dest
= sh
->dev
[data_target
].page
;
1302 init_async_submit(&submit
,
1303 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1305 to_addr_conv(sh
, percpu
));
1306 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1309 count
= set_syndrome_sources(blocks
, sh
);
1310 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1311 ops_complete_compute
, sh
,
1312 to_addr_conv(sh
, percpu
));
1313 return async_gen_syndrome(blocks
, 0, count
+2,
1314 STRIPE_SIZE
, &submit
);
1317 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1318 ops_complete_compute
, sh
,
1319 to_addr_conv(sh
, percpu
));
1320 if (failb
== syndrome_disks
) {
1321 /* We're missing D+P. */
1322 return async_raid6_datap_recov(syndrome_disks
+2,
1326 /* We're missing D+D. */
1327 return async_raid6_2data_recov(syndrome_disks
+2,
1328 STRIPE_SIZE
, faila
, failb
,
1335 static void ops_complete_prexor(void *stripe_head_ref
)
1337 struct stripe_head
*sh
= stripe_head_ref
;
1339 pr_debug("%s: stripe %llu\n", __func__
,
1340 (unsigned long long)sh
->sector
);
1343 static struct dma_async_tx_descriptor
*
1344 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1345 struct dma_async_tx_descriptor
*tx
)
1347 int disks
= sh
->disks
;
1348 struct page
**xor_srcs
= percpu
->scribble
;
1349 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1350 struct async_submit_ctl submit
;
1352 /* existing parity data subtracted */
1353 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1355 pr_debug("%s: stripe %llu\n", __func__
,
1356 (unsigned long long)sh
->sector
);
1358 for (i
= disks
; i
--; ) {
1359 struct r5dev
*dev
= &sh
->dev
[i
];
1360 /* Only process blocks that are known to be uptodate */
1361 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1362 xor_srcs
[count
++] = dev
->page
;
1365 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1366 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1367 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1372 static struct dma_async_tx_descriptor
*
1373 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1375 int disks
= sh
->disks
;
1378 pr_debug("%s: stripe %llu\n", __func__
,
1379 (unsigned long long)sh
->sector
);
1381 for (i
= disks
; i
--; ) {
1382 struct r5dev
*dev
= &sh
->dev
[i
];
1385 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1388 spin_lock_irq(&sh
->stripe_lock
);
1389 chosen
= dev
->towrite
;
1390 dev
->towrite
= NULL
;
1391 BUG_ON(dev
->written
);
1392 wbi
= dev
->written
= chosen
;
1393 spin_unlock_irq(&sh
->stripe_lock
);
1395 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1396 dev
->sector
+ STRIPE_SECTORS
) {
1397 if (wbi
->bi_rw
& REQ_FUA
)
1398 set_bit(R5_WantFUA
, &dev
->flags
);
1399 if (wbi
->bi_rw
& REQ_SYNC
)
1400 set_bit(R5_SyncIO
, &dev
->flags
);
1401 if (wbi
->bi_rw
& REQ_DISCARD
)
1402 set_bit(R5_Discard
, &dev
->flags
);
1404 tx
= async_copy_data(1, wbi
, dev
->page
,
1406 wbi
= r5_next_bio(wbi
, dev
->sector
);
1414 static void ops_complete_reconstruct(void *stripe_head_ref
)
1416 struct stripe_head
*sh
= stripe_head_ref
;
1417 int disks
= sh
->disks
;
1418 int pd_idx
= sh
->pd_idx
;
1419 int qd_idx
= sh
->qd_idx
;
1421 bool fua
= false, sync
= false, discard
= false;
1423 pr_debug("%s: stripe %llu\n", __func__
,
1424 (unsigned long long)sh
->sector
);
1426 for (i
= disks
; i
--; ) {
1427 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1428 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1429 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1432 for (i
= disks
; i
--; ) {
1433 struct r5dev
*dev
= &sh
->dev
[i
];
1435 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1437 set_bit(R5_UPTODATE
, &dev
->flags
);
1439 set_bit(R5_WantFUA
, &dev
->flags
);
1441 set_bit(R5_SyncIO
, &dev
->flags
);
1445 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1446 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1447 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1448 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1450 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1451 sh
->reconstruct_state
= reconstruct_state_result
;
1454 set_bit(STRIPE_HANDLE
, &sh
->state
);
1459 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1460 struct dma_async_tx_descriptor
*tx
)
1462 int disks
= sh
->disks
;
1463 struct page
**xor_srcs
= percpu
->scribble
;
1464 struct async_submit_ctl submit
;
1465 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1466 struct page
*xor_dest
;
1468 unsigned long flags
;
1470 pr_debug("%s: stripe %llu\n", __func__
,
1471 (unsigned long long)sh
->sector
);
1473 for (i
= 0; i
< sh
->disks
; i
++) {
1476 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1479 if (i
>= sh
->disks
) {
1480 atomic_inc(&sh
->count
);
1481 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1482 ops_complete_reconstruct(sh
);
1485 /* check if prexor is active which means only process blocks
1486 * that are part of a read-modify-write (written)
1488 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1490 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1491 for (i
= disks
; i
--; ) {
1492 struct r5dev
*dev
= &sh
->dev
[i
];
1494 xor_srcs
[count
++] = dev
->page
;
1497 xor_dest
= sh
->dev
[pd_idx
].page
;
1498 for (i
= disks
; i
--; ) {
1499 struct r5dev
*dev
= &sh
->dev
[i
];
1501 xor_srcs
[count
++] = dev
->page
;
1505 /* 1/ if we prexor'd then the dest is reused as a source
1506 * 2/ if we did not prexor then we are redoing the parity
1507 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1508 * for the synchronous xor case
1510 flags
= ASYNC_TX_ACK
|
1511 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1513 atomic_inc(&sh
->count
);
1515 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1516 to_addr_conv(sh
, percpu
));
1517 if (unlikely(count
== 1))
1518 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1520 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1524 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1525 struct dma_async_tx_descriptor
*tx
)
1527 struct async_submit_ctl submit
;
1528 struct page
**blocks
= percpu
->scribble
;
1531 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1533 for (i
= 0; i
< sh
->disks
; i
++) {
1534 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1536 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1539 if (i
>= sh
->disks
) {
1540 atomic_inc(&sh
->count
);
1541 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1542 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1543 ops_complete_reconstruct(sh
);
1547 count
= set_syndrome_sources(blocks
, sh
);
1549 atomic_inc(&sh
->count
);
1551 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1552 sh
, to_addr_conv(sh
, percpu
));
1553 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1556 static void ops_complete_check(void *stripe_head_ref
)
1558 struct stripe_head
*sh
= stripe_head_ref
;
1560 pr_debug("%s: stripe %llu\n", __func__
,
1561 (unsigned long long)sh
->sector
);
1563 sh
->check_state
= check_state_check_result
;
1564 set_bit(STRIPE_HANDLE
, &sh
->state
);
1568 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1570 int disks
= sh
->disks
;
1571 int pd_idx
= sh
->pd_idx
;
1572 int qd_idx
= sh
->qd_idx
;
1573 struct page
*xor_dest
;
1574 struct page
**xor_srcs
= percpu
->scribble
;
1575 struct dma_async_tx_descriptor
*tx
;
1576 struct async_submit_ctl submit
;
1580 pr_debug("%s: stripe %llu\n", __func__
,
1581 (unsigned long long)sh
->sector
);
1584 xor_dest
= sh
->dev
[pd_idx
].page
;
1585 xor_srcs
[count
++] = xor_dest
;
1586 for (i
= disks
; i
--; ) {
1587 if (i
== pd_idx
|| i
== qd_idx
)
1589 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1592 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1593 to_addr_conv(sh
, percpu
));
1594 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1595 &sh
->ops
.zero_sum_result
, &submit
);
1597 atomic_inc(&sh
->count
);
1598 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1599 tx
= async_trigger_callback(&submit
);
1602 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1604 struct page
**srcs
= percpu
->scribble
;
1605 struct async_submit_ctl submit
;
1608 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1609 (unsigned long long)sh
->sector
, checkp
);
1611 count
= set_syndrome_sources(srcs
, sh
);
1615 atomic_inc(&sh
->count
);
1616 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1617 sh
, to_addr_conv(sh
, percpu
));
1618 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1619 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1622 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1624 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1625 struct dma_async_tx_descriptor
*tx
= NULL
;
1626 struct r5conf
*conf
= sh
->raid_conf
;
1627 int level
= conf
->level
;
1628 struct raid5_percpu
*percpu
;
1632 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1633 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1634 ops_run_biofill(sh
);
1638 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1640 tx
= ops_run_compute5(sh
, percpu
);
1642 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1643 tx
= ops_run_compute6_1(sh
, percpu
);
1645 tx
= ops_run_compute6_2(sh
, percpu
);
1647 /* terminate the chain if reconstruct is not set to be run */
1648 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1652 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1653 tx
= ops_run_prexor(sh
, percpu
, tx
);
1655 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1656 tx
= ops_run_biodrain(sh
, tx
);
1660 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1662 ops_run_reconstruct5(sh
, percpu
, tx
);
1664 ops_run_reconstruct6(sh
, percpu
, tx
);
1667 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1668 if (sh
->check_state
== check_state_run
)
1669 ops_run_check_p(sh
, percpu
);
1670 else if (sh
->check_state
== check_state_run_q
)
1671 ops_run_check_pq(sh
, percpu
, 0);
1672 else if (sh
->check_state
== check_state_run_pq
)
1673 ops_run_check_pq(sh
, percpu
, 1);
1679 for (i
= disks
; i
--; ) {
1680 struct r5dev
*dev
= &sh
->dev
[i
];
1681 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1682 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1687 static int grow_one_stripe(struct r5conf
*conf
, int hash
)
1689 struct stripe_head
*sh
;
1690 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1694 sh
->raid_conf
= conf
;
1696 spin_lock_init(&sh
->stripe_lock
);
1698 if (grow_buffers(sh
)) {
1700 kmem_cache_free(conf
->slab_cache
, sh
);
1703 sh
->hash_lock_index
= hash
;
1704 /* we just created an active stripe so... */
1705 atomic_set(&sh
->count
, 1);
1706 atomic_inc(&conf
->active_stripes
);
1707 INIT_LIST_HEAD(&sh
->lru
);
1712 static int grow_stripes(struct r5conf
*conf
, int num
)
1714 struct kmem_cache
*sc
;
1715 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1718 if (conf
->mddev
->gendisk
)
1719 sprintf(conf
->cache_name
[0],
1720 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1722 sprintf(conf
->cache_name
[0],
1723 "raid%d-%p", conf
->level
, conf
->mddev
);
1724 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1726 conf
->active_name
= 0;
1727 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1728 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1732 conf
->slab_cache
= sc
;
1733 conf
->pool_size
= devs
;
1734 hash
= conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
1736 if (!grow_one_stripe(conf
, hash
))
1738 conf
->max_nr_stripes
++;
1739 hash
= (hash
+ 1) % NR_STRIPE_HASH_LOCKS
;
1745 * scribble_len - return the required size of the scribble region
1746 * @num - total number of disks in the array
1748 * The size must be enough to contain:
1749 * 1/ a struct page pointer for each device in the array +2
1750 * 2/ room to convert each entry in (1) to its corresponding dma
1751 * (dma_map_page()) or page (page_address()) address.
1753 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1754 * calculate over all devices (not just the data blocks), using zeros in place
1755 * of the P and Q blocks.
1757 static size_t scribble_len(int num
)
1761 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1766 static int resize_stripes(struct r5conf
*conf
, int newsize
)
1768 /* Make all the stripes able to hold 'newsize' devices.
1769 * New slots in each stripe get 'page' set to a new page.
1771 * This happens in stages:
1772 * 1/ create a new kmem_cache and allocate the required number of
1774 * 2/ gather all the old stripe_heads and transfer the pages across
1775 * to the new stripe_heads. This will have the side effect of
1776 * freezing the array as once all stripe_heads have been collected,
1777 * no IO will be possible. Old stripe heads are freed once their
1778 * pages have been transferred over, and the old kmem_cache is
1779 * freed when all stripes are done.
1780 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1781 * we simple return a failre status - no need to clean anything up.
1782 * 4/ allocate new pages for the new slots in the new stripe_heads.
1783 * If this fails, we don't bother trying the shrink the
1784 * stripe_heads down again, we just leave them as they are.
1785 * As each stripe_head is processed the new one is released into
1788 * Once step2 is started, we cannot afford to wait for a write,
1789 * so we use GFP_NOIO allocations.
1791 struct stripe_head
*osh
, *nsh
;
1792 LIST_HEAD(newstripes
);
1793 struct disk_info
*ndisks
;
1796 struct kmem_cache
*sc
;
1800 if (newsize
<= conf
->pool_size
)
1801 return 0; /* never bother to shrink */
1803 err
= md_allow_write(conf
->mddev
);
1808 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1809 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1814 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1815 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1819 nsh
->raid_conf
= conf
;
1820 spin_lock_init(&nsh
->stripe_lock
);
1822 list_add(&nsh
->lru
, &newstripes
);
1825 /* didn't get enough, give up */
1826 while (!list_empty(&newstripes
)) {
1827 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1828 list_del(&nsh
->lru
);
1829 kmem_cache_free(sc
, nsh
);
1831 kmem_cache_destroy(sc
);
1834 /* Step 2 - Must use GFP_NOIO now.
1835 * OK, we have enough stripes, start collecting inactive
1836 * stripes and copying them over
1840 list_for_each_entry(nsh
, &newstripes
, lru
) {
1841 lock_device_hash_lock(conf
, hash
);
1842 wait_event_cmd(conf
->wait_for_stripe
,
1843 !list_empty(conf
->inactive_list
+ hash
),
1844 unlock_device_hash_lock(conf
, hash
),
1845 lock_device_hash_lock(conf
, hash
));
1846 osh
= get_free_stripe(conf
, hash
);
1847 unlock_device_hash_lock(conf
, hash
);
1848 atomic_set(&nsh
->count
, 1);
1849 for(i
=0; i
<conf
->pool_size
; i
++)
1850 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1851 for( ; i
<newsize
; i
++)
1852 nsh
->dev
[i
].page
= NULL
;
1853 nsh
->hash_lock_index
= hash
;
1854 kmem_cache_free(conf
->slab_cache
, osh
);
1856 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
1857 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
1862 kmem_cache_destroy(conf
->slab_cache
);
1865 * At this point, we are holding all the stripes so the array
1866 * is completely stalled, so now is a good time to resize
1867 * conf->disks and the scribble region
1869 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1871 for (i
=0; i
<conf
->raid_disks
; i
++)
1872 ndisks
[i
] = conf
->disks
[i
];
1874 conf
->disks
= ndisks
;
1879 conf
->scribble_len
= scribble_len(newsize
);
1880 for_each_present_cpu(cpu
) {
1881 struct raid5_percpu
*percpu
;
1884 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1885 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1888 kfree(percpu
->scribble
);
1889 percpu
->scribble
= scribble
;
1897 /* Step 4, return new stripes to service */
1898 while(!list_empty(&newstripes
)) {
1899 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1900 list_del_init(&nsh
->lru
);
1902 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1903 if (nsh
->dev
[i
].page
== NULL
) {
1904 struct page
*p
= alloc_page(GFP_NOIO
);
1905 nsh
->dev
[i
].page
= p
;
1909 release_stripe(nsh
);
1911 /* critical section pass, GFP_NOIO no longer needed */
1913 conf
->slab_cache
= sc
;
1914 conf
->active_name
= 1-conf
->active_name
;
1915 conf
->pool_size
= newsize
;
1919 static int drop_one_stripe(struct r5conf
*conf
, int hash
)
1921 struct stripe_head
*sh
;
1923 spin_lock_irq(conf
->hash_locks
+ hash
);
1924 sh
= get_free_stripe(conf
, hash
);
1925 spin_unlock_irq(conf
->hash_locks
+ hash
);
1928 BUG_ON(atomic_read(&sh
->count
));
1930 kmem_cache_free(conf
->slab_cache
, sh
);
1931 atomic_dec(&conf
->active_stripes
);
1935 static void shrink_stripes(struct r5conf
*conf
)
1938 for (hash
= 0; hash
< NR_STRIPE_HASH_LOCKS
; hash
++)
1939 while (drop_one_stripe(conf
, hash
))
1942 if (conf
->slab_cache
)
1943 kmem_cache_destroy(conf
->slab_cache
);
1944 conf
->slab_cache
= NULL
;
1947 static void raid5_end_read_request(struct bio
* bi
, int error
)
1949 struct stripe_head
*sh
= bi
->bi_private
;
1950 struct r5conf
*conf
= sh
->raid_conf
;
1951 int disks
= sh
->disks
, i
;
1952 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1953 char b
[BDEVNAME_SIZE
];
1954 struct md_rdev
*rdev
= NULL
;
1957 for (i
=0 ; i
<disks
; i
++)
1958 if (bi
== &sh
->dev
[i
].req
)
1961 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1962 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1968 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1969 /* If replacement finished while this request was outstanding,
1970 * 'replacement' might be NULL already.
1971 * In that case it moved down to 'rdev'.
1972 * rdev is not removed until all requests are finished.
1974 rdev
= conf
->disks
[i
].replacement
;
1976 rdev
= conf
->disks
[i
].rdev
;
1978 if (use_new_offset(conf
, sh
))
1979 s
= sh
->sector
+ rdev
->new_data_offset
;
1981 s
= sh
->sector
+ rdev
->data_offset
;
1983 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1984 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1985 /* Note that this cannot happen on a
1986 * replacement device. We just fail those on
1991 "md/raid:%s: read error corrected"
1992 " (%lu sectors at %llu on %s)\n",
1993 mdname(conf
->mddev
), STRIPE_SECTORS
,
1994 (unsigned long long)s
,
1995 bdevname(rdev
->bdev
, b
));
1996 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1997 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1998 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1999 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2000 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2002 if (atomic_read(&rdev
->read_errors
))
2003 atomic_set(&rdev
->read_errors
, 0);
2005 const char *bdn
= bdevname(rdev
->bdev
, b
);
2009 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2010 atomic_inc(&rdev
->read_errors
);
2011 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2014 "md/raid:%s: read error on replacement device "
2015 "(sector %llu on %s).\n",
2016 mdname(conf
->mddev
),
2017 (unsigned long long)s
,
2019 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2023 "md/raid:%s: read error not correctable "
2024 "(sector %llu on %s).\n",
2025 mdname(conf
->mddev
),
2026 (unsigned long long)s
,
2028 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2033 "md/raid:%s: read error NOT corrected!! "
2034 "(sector %llu on %s).\n",
2035 mdname(conf
->mddev
),
2036 (unsigned long long)s
,
2038 } else if (atomic_read(&rdev
->read_errors
)
2039 > conf
->max_nr_stripes
)
2041 "md/raid:%s: Too many read errors, failing device %s.\n",
2042 mdname(conf
->mddev
), bdn
);
2045 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2046 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2049 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2050 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2051 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2053 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2055 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2056 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2058 && test_bit(In_sync
, &rdev
->flags
)
2059 && rdev_set_badblocks(
2060 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2061 md_error(conf
->mddev
, rdev
);
2064 rdev_dec_pending(rdev
, conf
->mddev
);
2065 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2066 set_bit(STRIPE_HANDLE
, &sh
->state
);
2070 static void raid5_end_write_request(struct bio
*bi
, int error
)
2072 struct stripe_head
*sh
= bi
->bi_private
;
2073 struct r5conf
*conf
= sh
->raid_conf
;
2074 int disks
= sh
->disks
, i
;
2075 struct md_rdev
*uninitialized_var(rdev
);
2076 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2079 int replacement
= 0;
2081 for (i
= 0 ; i
< disks
; i
++) {
2082 if (bi
== &sh
->dev
[i
].req
) {
2083 rdev
= conf
->disks
[i
].rdev
;
2086 if (bi
== &sh
->dev
[i
].rreq
) {
2087 rdev
= conf
->disks
[i
].replacement
;
2091 /* rdev was removed and 'replacement'
2092 * replaced it. rdev is not removed
2093 * until all requests are finished.
2095 rdev
= conf
->disks
[i
].rdev
;
2099 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2100 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2109 md_error(conf
->mddev
, rdev
);
2110 else if (is_badblock(rdev
, sh
->sector
,
2112 &first_bad
, &bad_sectors
))
2113 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2116 set_bit(WriteErrorSeen
, &rdev
->flags
);
2117 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2118 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2119 set_bit(MD_RECOVERY_NEEDED
,
2120 &rdev
->mddev
->recovery
);
2121 } else if (is_badblock(rdev
, sh
->sector
,
2123 &first_bad
, &bad_sectors
)) {
2124 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2125 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2126 /* That was a successful write so make
2127 * sure it looks like we already did
2130 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2133 rdev_dec_pending(rdev
, conf
->mddev
);
2135 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2136 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2137 set_bit(STRIPE_HANDLE
, &sh
->state
);
2141 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
2143 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
2145 struct r5dev
*dev
= &sh
->dev
[i
];
2147 bio_init(&dev
->req
);
2148 dev
->req
.bi_io_vec
= &dev
->vec
;
2150 dev
->req
.bi_max_vecs
++;
2151 dev
->req
.bi_private
= sh
;
2152 dev
->vec
.bv_page
= dev
->page
;
2154 bio_init(&dev
->rreq
);
2155 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
2156 dev
->rreq
.bi_vcnt
++;
2157 dev
->rreq
.bi_max_vecs
++;
2158 dev
->rreq
.bi_private
= sh
;
2159 dev
->rvec
.bv_page
= dev
->page
;
2162 dev
->sector
= compute_blocknr(sh
, i
, previous
);
2165 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2167 char b
[BDEVNAME_SIZE
];
2168 struct r5conf
*conf
= mddev
->private;
2169 unsigned long flags
;
2170 pr_debug("raid456: error called\n");
2172 spin_lock_irqsave(&conf
->device_lock
, flags
);
2173 clear_bit(In_sync
, &rdev
->flags
);
2174 mddev
->degraded
= calc_degraded(conf
);
2175 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2176 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2178 set_bit(Blocked
, &rdev
->flags
);
2179 set_bit(Faulty
, &rdev
->flags
);
2180 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2182 "md/raid:%s: Disk failure on %s, disabling device.\n"
2183 "md/raid:%s: Operation continuing on %d devices.\n",
2185 bdevname(rdev
->bdev
, b
),
2187 conf
->raid_disks
- mddev
->degraded
);
2191 * Input: a 'big' sector number,
2192 * Output: index of the data and parity disk, and the sector # in them.
2194 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2195 int previous
, int *dd_idx
,
2196 struct stripe_head
*sh
)
2198 sector_t stripe
, stripe2
;
2199 sector_t chunk_number
;
2200 unsigned int chunk_offset
;
2203 sector_t new_sector
;
2204 int algorithm
= previous
? conf
->prev_algo
2206 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2207 : conf
->chunk_sectors
;
2208 int raid_disks
= previous
? conf
->previous_raid_disks
2210 int data_disks
= raid_disks
- conf
->max_degraded
;
2212 /* First compute the information on this sector */
2215 * Compute the chunk number and the sector offset inside the chunk
2217 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2218 chunk_number
= r_sector
;
2221 * Compute the stripe number
2223 stripe
= chunk_number
;
2224 *dd_idx
= sector_div(stripe
, data_disks
);
2227 * Select the parity disk based on the user selected algorithm.
2229 pd_idx
= qd_idx
= -1;
2230 switch(conf
->level
) {
2232 pd_idx
= data_disks
;
2235 switch (algorithm
) {
2236 case ALGORITHM_LEFT_ASYMMETRIC
:
2237 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2238 if (*dd_idx
>= pd_idx
)
2241 case ALGORITHM_RIGHT_ASYMMETRIC
:
2242 pd_idx
= sector_div(stripe2
, raid_disks
);
2243 if (*dd_idx
>= pd_idx
)
2246 case ALGORITHM_LEFT_SYMMETRIC
:
2247 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2248 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2250 case ALGORITHM_RIGHT_SYMMETRIC
:
2251 pd_idx
= sector_div(stripe2
, raid_disks
);
2252 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2254 case ALGORITHM_PARITY_0
:
2258 case ALGORITHM_PARITY_N
:
2259 pd_idx
= data_disks
;
2267 switch (algorithm
) {
2268 case ALGORITHM_LEFT_ASYMMETRIC
:
2269 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2270 qd_idx
= pd_idx
+ 1;
2271 if (pd_idx
== raid_disks
-1) {
2272 (*dd_idx
)++; /* Q D D D P */
2274 } else if (*dd_idx
>= pd_idx
)
2275 (*dd_idx
) += 2; /* D D P Q D */
2277 case ALGORITHM_RIGHT_ASYMMETRIC
:
2278 pd_idx
= sector_div(stripe2
, raid_disks
);
2279 qd_idx
= pd_idx
+ 1;
2280 if (pd_idx
== raid_disks
-1) {
2281 (*dd_idx
)++; /* Q D D D P */
2283 } else if (*dd_idx
>= pd_idx
)
2284 (*dd_idx
) += 2; /* D D P Q D */
2286 case ALGORITHM_LEFT_SYMMETRIC
:
2287 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2288 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2289 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2291 case ALGORITHM_RIGHT_SYMMETRIC
:
2292 pd_idx
= sector_div(stripe2
, raid_disks
);
2293 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2294 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2297 case ALGORITHM_PARITY_0
:
2302 case ALGORITHM_PARITY_N
:
2303 pd_idx
= data_disks
;
2304 qd_idx
= data_disks
+ 1;
2307 case ALGORITHM_ROTATING_ZERO_RESTART
:
2308 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2309 * of blocks for computing Q is different.
2311 pd_idx
= sector_div(stripe2
, raid_disks
);
2312 qd_idx
= pd_idx
+ 1;
2313 if (pd_idx
== raid_disks
-1) {
2314 (*dd_idx
)++; /* Q D D D P */
2316 } else if (*dd_idx
>= pd_idx
)
2317 (*dd_idx
) += 2; /* D D P Q D */
2321 case ALGORITHM_ROTATING_N_RESTART
:
2322 /* Same a left_asymmetric, by first stripe is
2323 * D D D P Q rather than
2327 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2328 qd_idx
= pd_idx
+ 1;
2329 if (pd_idx
== raid_disks
-1) {
2330 (*dd_idx
)++; /* Q D D D P */
2332 } else if (*dd_idx
>= pd_idx
)
2333 (*dd_idx
) += 2; /* D D P Q D */
2337 case ALGORITHM_ROTATING_N_CONTINUE
:
2338 /* Same as left_symmetric but Q is before P */
2339 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2340 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2341 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2345 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2346 /* RAID5 left_asymmetric, with Q on last device */
2347 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2348 if (*dd_idx
>= pd_idx
)
2350 qd_idx
= raid_disks
- 1;
2353 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2354 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2355 if (*dd_idx
>= pd_idx
)
2357 qd_idx
= raid_disks
- 1;
2360 case ALGORITHM_LEFT_SYMMETRIC_6
:
2361 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2362 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2363 qd_idx
= raid_disks
- 1;
2366 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2367 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2368 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2369 qd_idx
= raid_disks
- 1;
2372 case ALGORITHM_PARITY_0_6
:
2375 qd_idx
= raid_disks
- 1;
2385 sh
->pd_idx
= pd_idx
;
2386 sh
->qd_idx
= qd_idx
;
2387 sh
->ddf_layout
= ddf_layout
;
2390 * Finally, compute the new sector number
2392 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2397 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2399 struct r5conf
*conf
= sh
->raid_conf
;
2400 int raid_disks
= sh
->disks
;
2401 int data_disks
= raid_disks
- conf
->max_degraded
;
2402 sector_t new_sector
= sh
->sector
, check
;
2403 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2404 : conf
->chunk_sectors
;
2405 int algorithm
= previous
? conf
->prev_algo
2409 sector_t chunk_number
;
2410 int dummy1
, dd_idx
= i
;
2412 struct stripe_head sh2
;
2415 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2416 stripe
= new_sector
;
2418 if (i
== sh
->pd_idx
)
2420 switch(conf
->level
) {
2423 switch (algorithm
) {
2424 case ALGORITHM_LEFT_ASYMMETRIC
:
2425 case ALGORITHM_RIGHT_ASYMMETRIC
:
2429 case ALGORITHM_LEFT_SYMMETRIC
:
2430 case ALGORITHM_RIGHT_SYMMETRIC
:
2433 i
-= (sh
->pd_idx
+ 1);
2435 case ALGORITHM_PARITY_0
:
2438 case ALGORITHM_PARITY_N
:
2445 if (i
== sh
->qd_idx
)
2446 return 0; /* It is the Q disk */
2447 switch (algorithm
) {
2448 case ALGORITHM_LEFT_ASYMMETRIC
:
2449 case ALGORITHM_RIGHT_ASYMMETRIC
:
2450 case ALGORITHM_ROTATING_ZERO_RESTART
:
2451 case ALGORITHM_ROTATING_N_RESTART
:
2452 if (sh
->pd_idx
== raid_disks
-1)
2453 i
--; /* Q D D D P */
2454 else if (i
> sh
->pd_idx
)
2455 i
-= 2; /* D D P Q D */
2457 case ALGORITHM_LEFT_SYMMETRIC
:
2458 case ALGORITHM_RIGHT_SYMMETRIC
:
2459 if (sh
->pd_idx
== raid_disks
-1)
2460 i
--; /* Q D D D P */
2465 i
-= (sh
->pd_idx
+ 2);
2468 case ALGORITHM_PARITY_0
:
2471 case ALGORITHM_PARITY_N
:
2473 case ALGORITHM_ROTATING_N_CONTINUE
:
2474 /* Like left_symmetric, but P is before Q */
2475 if (sh
->pd_idx
== 0)
2476 i
--; /* P D D D Q */
2481 i
-= (sh
->pd_idx
+ 1);
2484 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2485 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2489 case ALGORITHM_LEFT_SYMMETRIC_6
:
2490 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2492 i
+= data_disks
+ 1;
2493 i
-= (sh
->pd_idx
+ 1);
2495 case ALGORITHM_PARITY_0_6
:
2504 chunk_number
= stripe
* data_disks
+ i
;
2505 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2507 check
= raid5_compute_sector(conf
, r_sector
,
2508 previous
, &dummy1
, &sh2
);
2509 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2510 || sh2
.qd_idx
!= sh
->qd_idx
) {
2511 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2512 mdname(conf
->mddev
));
2520 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2521 int rcw
, int expand
)
2523 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2524 struct r5conf
*conf
= sh
->raid_conf
;
2525 int level
= conf
->level
;
2529 for (i
= disks
; i
--; ) {
2530 struct r5dev
*dev
= &sh
->dev
[i
];
2533 set_bit(R5_LOCKED
, &dev
->flags
);
2534 set_bit(R5_Wantdrain
, &dev
->flags
);
2536 clear_bit(R5_UPTODATE
, &dev
->flags
);
2540 /* if we are not expanding this is a proper write request, and
2541 * there will be bios with new data to be drained into the
2546 /* False alarm, nothing to do */
2548 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2549 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2551 sh
->reconstruct_state
= reconstruct_state_run
;
2553 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2555 if (s
->locked
+ conf
->max_degraded
== disks
)
2556 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2557 atomic_inc(&conf
->pending_full_writes
);
2560 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2561 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2563 for (i
= disks
; i
--; ) {
2564 struct r5dev
*dev
= &sh
->dev
[i
];
2569 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2570 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2571 set_bit(R5_Wantdrain
, &dev
->flags
);
2572 set_bit(R5_LOCKED
, &dev
->flags
);
2573 clear_bit(R5_UPTODATE
, &dev
->flags
);
2578 /* False alarm - nothing to do */
2580 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2581 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2582 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2583 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2586 /* keep the parity disk(s) locked while asynchronous operations
2589 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2590 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2594 int qd_idx
= sh
->qd_idx
;
2595 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2597 set_bit(R5_LOCKED
, &dev
->flags
);
2598 clear_bit(R5_UPTODATE
, &dev
->flags
);
2602 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2603 __func__
, (unsigned long long)sh
->sector
,
2604 s
->locked
, s
->ops_request
);
2608 * Each stripe/dev can have one or more bion attached.
2609 * toread/towrite point to the first in a chain.
2610 * The bi_next chain must be in order.
2612 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2615 struct r5conf
*conf
= sh
->raid_conf
;
2618 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2619 (unsigned long long)bi
->bi_iter
.bi_sector
,
2620 (unsigned long long)sh
->sector
);
2623 * If several bio share a stripe. The bio bi_phys_segments acts as a
2624 * reference count to avoid race. The reference count should already be
2625 * increased before this function is called (for example, in
2626 * make_request()), so other bio sharing this stripe will not free the
2627 * stripe. If a stripe is owned by one stripe, the stripe lock will
2630 spin_lock_irq(&sh
->stripe_lock
);
2632 bip
= &sh
->dev
[dd_idx
].towrite
;
2636 bip
= &sh
->dev
[dd_idx
].toread
;
2637 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
2638 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
2640 bip
= & (*bip
)->bi_next
;
2642 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
2645 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2649 raid5_inc_bi_active_stripes(bi
);
2652 /* check if page is covered */
2653 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2654 for (bi
=sh
->dev
[dd_idx
].towrite
;
2655 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2656 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
2657 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2658 if (bio_end_sector(bi
) >= sector
)
2659 sector
= bio_end_sector(bi
);
2661 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2662 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2665 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2666 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
2667 (unsigned long long)sh
->sector
, dd_idx
);
2668 spin_unlock_irq(&sh
->stripe_lock
);
2670 if (conf
->mddev
->bitmap
&& firstwrite
) {
2671 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2673 sh
->bm_seq
= conf
->seq_flush
+1;
2674 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2679 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2680 spin_unlock_irq(&sh
->stripe_lock
);
2684 static void end_reshape(struct r5conf
*conf
);
2686 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2687 struct stripe_head
*sh
)
2689 int sectors_per_chunk
=
2690 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2692 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2693 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2695 raid5_compute_sector(conf
,
2696 stripe
* (disks
- conf
->max_degraded
)
2697 *sectors_per_chunk
+ chunk_offset
,
2703 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2704 struct stripe_head_state
*s
, int disks
,
2705 struct bio
**return_bi
)
2708 for (i
= disks
; i
--; ) {
2712 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2713 struct md_rdev
*rdev
;
2715 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2716 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2717 atomic_inc(&rdev
->nr_pending
);
2722 if (!rdev_set_badblocks(
2726 md_error(conf
->mddev
, rdev
);
2727 rdev_dec_pending(rdev
, conf
->mddev
);
2730 spin_lock_irq(&sh
->stripe_lock
);
2731 /* fail all writes first */
2732 bi
= sh
->dev
[i
].towrite
;
2733 sh
->dev
[i
].towrite
= NULL
;
2734 spin_unlock_irq(&sh
->stripe_lock
);
2738 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2739 wake_up(&conf
->wait_for_overlap
);
2741 while (bi
&& bi
->bi_iter
.bi_sector
<
2742 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2743 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2744 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2745 if (!raid5_dec_bi_active_stripes(bi
)) {
2746 md_write_end(conf
->mddev
);
2747 bi
->bi_next
= *return_bi
;
2753 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2754 STRIPE_SECTORS
, 0, 0);
2756 /* and fail all 'written' */
2757 bi
= sh
->dev
[i
].written
;
2758 sh
->dev
[i
].written
= NULL
;
2759 if (bi
) bitmap_end
= 1;
2760 while (bi
&& bi
->bi_iter
.bi_sector
<
2761 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2762 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2763 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2764 if (!raid5_dec_bi_active_stripes(bi
)) {
2765 md_write_end(conf
->mddev
);
2766 bi
->bi_next
= *return_bi
;
2772 /* fail any reads if this device is non-operational and
2773 * the data has not reached the cache yet.
2775 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2776 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2777 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2778 spin_lock_irq(&sh
->stripe_lock
);
2779 bi
= sh
->dev
[i
].toread
;
2780 sh
->dev
[i
].toread
= NULL
;
2781 spin_unlock_irq(&sh
->stripe_lock
);
2782 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2783 wake_up(&conf
->wait_for_overlap
);
2784 while (bi
&& bi
->bi_iter
.bi_sector
<
2785 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2786 struct bio
*nextbi
=
2787 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2788 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2789 if (!raid5_dec_bi_active_stripes(bi
)) {
2790 bi
->bi_next
= *return_bi
;
2797 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2798 STRIPE_SECTORS
, 0, 0);
2799 /* If we were in the middle of a write the parity block might
2800 * still be locked - so just clear all R5_LOCKED flags
2802 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2805 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2806 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2807 md_wakeup_thread(conf
->mddev
->thread
);
2811 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2812 struct stripe_head_state
*s
)
2817 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2818 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
2819 wake_up(&conf
->wait_for_overlap
);
2822 /* There is nothing more to do for sync/check/repair.
2823 * Don't even need to abort as that is handled elsewhere
2824 * if needed, and not always wanted e.g. if there is a known
2826 * For recover/replace we need to record a bad block on all
2827 * non-sync devices, or abort the recovery
2829 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
2830 /* During recovery devices cannot be removed, so
2831 * locking and refcounting of rdevs is not needed
2833 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2834 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2836 && !test_bit(Faulty
, &rdev
->flags
)
2837 && !test_bit(In_sync
, &rdev
->flags
)
2838 && !rdev_set_badblocks(rdev
, sh
->sector
,
2841 rdev
= conf
->disks
[i
].replacement
;
2843 && !test_bit(Faulty
, &rdev
->flags
)
2844 && !test_bit(In_sync
, &rdev
->flags
)
2845 && !rdev_set_badblocks(rdev
, sh
->sector
,
2850 conf
->recovery_disabled
=
2851 conf
->mddev
->recovery_disabled
;
2853 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
2856 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2858 struct md_rdev
*rdev
;
2860 /* Doing recovery so rcu locking not required */
2861 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2863 && !test_bit(Faulty
, &rdev
->flags
)
2864 && !test_bit(In_sync
, &rdev
->flags
)
2865 && (rdev
->recovery_offset
<= sh
->sector
2866 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2872 /* fetch_block - checks the given member device to see if its data needs
2873 * to be read or computed to satisfy a request.
2875 * Returns 1 when no more member devices need to be checked, otherwise returns
2876 * 0 to tell the loop in handle_stripe_fill to continue
2878 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2879 int disk_idx
, int disks
)
2881 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2882 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2883 &sh
->dev
[s
->failed_num
[1]] };
2885 /* is the data in this block needed, and can we get it? */
2886 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2887 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2889 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2890 s
->syncing
|| s
->expanding
||
2891 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2892 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2893 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2894 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2895 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2896 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2897 /* we would like to get this block, possibly by computing it,
2898 * otherwise read it if the backing disk is insync
2900 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2901 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2902 if ((s
->uptodate
== disks
- 1) &&
2903 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2904 disk_idx
== s
->failed_num
[1]))) {
2905 /* have disk failed, and we're requested to fetch it;
2908 pr_debug("Computing stripe %llu block %d\n",
2909 (unsigned long long)sh
->sector
, disk_idx
);
2910 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2911 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2912 set_bit(R5_Wantcompute
, &dev
->flags
);
2913 sh
->ops
.target
= disk_idx
;
2914 sh
->ops
.target2
= -1; /* no 2nd target */
2916 /* Careful: from this point on 'uptodate' is in the eye
2917 * of raid_run_ops which services 'compute' operations
2918 * before writes. R5_Wantcompute flags a block that will
2919 * be R5_UPTODATE by the time it is needed for a
2920 * subsequent operation.
2924 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2925 /* Computing 2-failure is *very* expensive; only
2926 * do it if failed >= 2
2929 for (other
= disks
; other
--; ) {
2930 if (other
== disk_idx
)
2932 if (!test_bit(R5_UPTODATE
,
2933 &sh
->dev
[other
].flags
))
2937 pr_debug("Computing stripe %llu blocks %d,%d\n",
2938 (unsigned long long)sh
->sector
,
2940 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2941 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2942 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2943 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2944 sh
->ops
.target
= disk_idx
;
2945 sh
->ops
.target2
= other
;
2949 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2950 set_bit(R5_LOCKED
, &dev
->flags
);
2951 set_bit(R5_Wantread
, &dev
->flags
);
2953 pr_debug("Reading block %d (sync=%d)\n",
2954 disk_idx
, s
->syncing
);
2962 * handle_stripe_fill - read or compute data to satisfy pending requests.
2964 static void handle_stripe_fill(struct stripe_head
*sh
,
2965 struct stripe_head_state
*s
,
2970 /* look for blocks to read/compute, skip this if a compute
2971 * is already in flight, or if the stripe contents are in the
2972 * midst of changing due to a write
2974 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2975 !sh
->reconstruct_state
)
2976 for (i
= disks
; i
--; )
2977 if (fetch_block(sh
, s
, i
, disks
))
2979 set_bit(STRIPE_HANDLE
, &sh
->state
);
2983 /* handle_stripe_clean_event
2984 * any written block on an uptodate or failed drive can be returned.
2985 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2986 * never LOCKED, so we don't need to test 'failed' directly.
2988 static void handle_stripe_clean_event(struct r5conf
*conf
,
2989 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2993 int discard_pending
= 0;
2995 for (i
= disks
; i
--; )
2996 if (sh
->dev
[i
].written
) {
2998 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2999 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3000 test_bit(R5_Discard
, &dev
->flags
))) {
3001 /* We can return any write requests */
3002 struct bio
*wbi
, *wbi2
;
3003 pr_debug("Return write for disc %d\n", i
);
3004 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3005 clear_bit(R5_UPTODATE
, &dev
->flags
);
3007 dev
->written
= NULL
;
3008 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3009 dev
->sector
+ STRIPE_SECTORS
) {
3010 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3011 if (!raid5_dec_bi_active_stripes(wbi
)) {
3012 md_write_end(conf
->mddev
);
3013 wbi
->bi_next
= *return_bi
;
3018 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3020 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3022 } else if (test_bit(R5_Discard
, &dev
->flags
))
3023 discard_pending
= 1;
3025 if (!discard_pending
&&
3026 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3027 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3028 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3029 if (sh
->qd_idx
>= 0) {
3030 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3031 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3033 /* now that discard is done we can proceed with any sync */
3034 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3036 * SCSI discard will change some bio fields and the stripe has
3037 * no updated data, so remove it from hash list and the stripe
3038 * will be reinitialized
3040 spin_lock_irq(&conf
->device_lock
);
3042 spin_unlock_irq(&conf
->device_lock
);
3043 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3044 set_bit(STRIPE_HANDLE
, &sh
->state
);
3048 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3049 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3050 md_wakeup_thread(conf
->mddev
->thread
);
3053 static void handle_stripe_dirtying(struct r5conf
*conf
,
3054 struct stripe_head
*sh
,
3055 struct stripe_head_state
*s
,
3058 int rmw
= 0, rcw
= 0, i
;
3059 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3061 /* RAID6 requires 'rcw' in current implementation.
3062 * Otherwise, check whether resync is now happening or should start.
3063 * If yes, then the array is dirty (after unclean shutdown or
3064 * initial creation), so parity in some stripes might be inconsistent.
3065 * In this case, we need to always do reconstruct-write, to ensure
3066 * that in case of drive failure or read-error correction, we
3067 * generate correct data from the parity.
3069 if (conf
->max_degraded
== 2 ||
3070 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
)) {
3071 /* Calculate the real rcw later - for now make it
3072 * look like rcw is cheaper
3075 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3076 conf
->max_degraded
, (unsigned long long)recovery_cp
,
3077 (unsigned long long)sh
->sector
);
3078 } else for (i
= disks
; i
--; ) {
3079 /* would I have to read this buffer for read_modify_write */
3080 struct r5dev
*dev
= &sh
->dev
[i
];
3081 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
3082 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3083 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3084 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3085 if (test_bit(R5_Insync
, &dev
->flags
))
3088 rmw
+= 2*disks
; /* cannot read it */
3090 /* Would I have to read this buffer for reconstruct_write */
3091 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
3092 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3093 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3094 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3095 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
3100 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3101 (unsigned long long)sh
->sector
, rmw
, rcw
);
3102 set_bit(STRIPE_HANDLE
, &sh
->state
);
3103 if (rmw
< rcw
&& rmw
> 0) {
3104 /* prefer read-modify-write, but need to get some data */
3105 if (conf
->mddev
->queue
)
3106 blk_add_trace_msg(conf
->mddev
->queue
,
3107 "raid5 rmw %llu %d",
3108 (unsigned long long)sh
->sector
, rmw
);
3109 for (i
= disks
; i
--; ) {
3110 struct r5dev
*dev
= &sh
->dev
[i
];
3111 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
3112 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3113 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3114 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3115 test_bit(R5_Insync
, &dev
->flags
)) {
3117 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
3118 pr_debug("Read_old block "
3119 "%d for r-m-w\n", i
);
3120 set_bit(R5_LOCKED
, &dev
->flags
);
3121 set_bit(R5_Wantread
, &dev
->flags
);
3124 set_bit(STRIPE_DELAYED
, &sh
->state
);
3125 set_bit(STRIPE_HANDLE
, &sh
->state
);
3130 if (rcw
<= rmw
&& rcw
> 0) {
3131 /* want reconstruct write, but need to get some data */
3134 for (i
= disks
; i
--; ) {
3135 struct r5dev
*dev
= &sh
->dev
[i
];
3136 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3137 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3138 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3139 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3140 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3142 if (!test_bit(R5_Insync
, &dev
->flags
))
3143 continue; /* it's a failed drive */
3145 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
3146 pr_debug("Read_old block "
3147 "%d for Reconstruct\n", i
);
3148 set_bit(R5_LOCKED
, &dev
->flags
);
3149 set_bit(R5_Wantread
, &dev
->flags
);
3153 set_bit(STRIPE_DELAYED
, &sh
->state
);
3154 set_bit(STRIPE_HANDLE
, &sh
->state
);
3158 if (rcw
&& conf
->mddev
->queue
)
3159 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3160 (unsigned long long)sh
->sector
,
3161 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3163 /* now if nothing is locked, and if we have enough data,
3164 * we can start a write request
3166 /* since handle_stripe can be called at any time we need to handle the
3167 * case where a compute block operation has been submitted and then a
3168 * subsequent call wants to start a write request. raid_run_ops only
3169 * handles the case where compute block and reconstruct are requested
3170 * simultaneously. If this is not the case then new writes need to be
3171 * held off until the compute completes.
3173 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
3174 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
3175 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
3176 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
3179 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
3180 struct stripe_head_state
*s
, int disks
)
3182 struct r5dev
*dev
= NULL
;
3184 set_bit(STRIPE_HANDLE
, &sh
->state
);
3186 switch (sh
->check_state
) {
3187 case check_state_idle
:
3188 /* start a new check operation if there are no failures */
3189 if (s
->failed
== 0) {
3190 BUG_ON(s
->uptodate
!= disks
);
3191 sh
->check_state
= check_state_run
;
3192 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3193 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3197 dev
= &sh
->dev
[s
->failed_num
[0]];
3199 case check_state_compute_result
:
3200 sh
->check_state
= check_state_idle
;
3202 dev
= &sh
->dev
[sh
->pd_idx
];
3204 /* check that a write has not made the stripe insync */
3205 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3208 /* either failed parity check, or recovery is happening */
3209 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3210 BUG_ON(s
->uptodate
!= disks
);
3212 set_bit(R5_LOCKED
, &dev
->flags
);
3214 set_bit(R5_Wantwrite
, &dev
->flags
);
3216 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3217 set_bit(STRIPE_INSYNC
, &sh
->state
);
3219 case check_state_run
:
3220 break; /* we will be called again upon completion */
3221 case check_state_check_result
:
3222 sh
->check_state
= check_state_idle
;
3224 /* if a failure occurred during the check operation, leave
3225 * STRIPE_INSYNC not set and let the stripe be handled again
3230 /* handle a successful check operation, if parity is correct
3231 * we are done. Otherwise update the mismatch count and repair
3232 * parity if !MD_RECOVERY_CHECK
3234 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3235 /* parity is correct (on disc,
3236 * not in buffer any more)
3238 set_bit(STRIPE_INSYNC
, &sh
->state
);
3240 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3241 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3242 /* don't try to repair!! */
3243 set_bit(STRIPE_INSYNC
, &sh
->state
);
3245 sh
->check_state
= check_state_compute_run
;
3246 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3247 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3248 set_bit(R5_Wantcompute
,
3249 &sh
->dev
[sh
->pd_idx
].flags
);
3250 sh
->ops
.target
= sh
->pd_idx
;
3251 sh
->ops
.target2
= -1;
3256 case check_state_compute_run
:
3259 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3260 __func__
, sh
->check_state
,
3261 (unsigned long long) sh
->sector
);
3267 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3268 struct stripe_head_state
*s
,
3271 int pd_idx
= sh
->pd_idx
;
3272 int qd_idx
= sh
->qd_idx
;
3275 set_bit(STRIPE_HANDLE
, &sh
->state
);
3277 BUG_ON(s
->failed
> 2);
3279 /* Want to check and possibly repair P and Q.
3280 * However there could be one 'failed' device, in which
3281 * case we can only check one of them, possibly using the
3282 * other to generate missing data
3285 switch (sh
->check_state
) {
3286 case check_state_idle
:
3287 /* start a new check operation if there are < 2 failures */
3288 if (s
->failed
== s
->q_failed
) {
3289 /* The only possible failed device holds Q, so it
3290 * makes sense to check P (If anything else were failed,
3291 * we would have used P to recreate it).
3293 sh
->check_state
= check_state_run
;
3295 if (!s
->q_failed
&& s
->failed
< 2) {
3296 /* Q is not failed, and we didn't use it to generate
3297 * anything, so it makes sense to check it
3299 if (sh
->check_state
== check_state_run
)
3300 sh
->check_state
= check_state_run_pq
;
3302 sh
->check_state
= check_state_run_q
;
3305 /* discard potentially stale zero_sum_result */
3306 sh
->ops
.zero_sum_result
= 0;
3308 if (sh
->check_state
== check_state_run
) {
3309 /* async_xor_zero_sum destroys the contents of P */
3310 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3313 if (sh
->check_state
>= check_state_run
&&
3314 sh
->check_state
<= check_state_run_pq
) {
3315 /* async_syndrome_zero_sum preserves P and Q, so
3316 * no need to mark them !uptodate here
3318 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3322 /* we have 2-disk failure */
3323 BUG_ON(s
->failed
!= 2);
3325 case check_state_compute_result
:
3326 sh
->check_state
= check_state_idle
;
3328 /* check that a write has not made the stripe insync */
3329 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3332 /* now write out any block on a failed drive,
3333 * or P or Q if they were recomputed
3335 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3336 if (s
->failed
== 2) {
3337 dev
= &sh
->dev
[s
->failed_num
[1]];
3339 set_bit(R5_LOCKED
, &dev
->flags
);
3340 set_bit(R5_Wantwrite
, &dev
->flags
);
3342 if (s
->failed
>= 1) {
3343 dev
= &sh
->dev
[s
->failed_num
[0]];
3345 set_bit(R5_LOCKED
, &dev
->flags
);
3346 set_bit(R5_Wantwrite
, &dev
->flags
);
3348 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3349 dev
= &sh
->dev
[pd_idx
];
3351 set_bit(R5_LOCKED
, &dev
->flags
);
3352 set_bit(R5_Wantwrite
, &dev
->flags
);
3354 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3355 dev
= &sh
->dev
[qd_idx
];
3357 set_bit(R5_LOCKED
, &dev
->flags
);
3358 set_bit(R5_Wantwrite
, &dev
->flags
);
3360 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3362 set_bit(STRIPE_INSYNC
, &sh
->state
);
3364 case check_state_run
:
3365 case check_state_run_q
:
3366 case check_state_run_pq
:
3367 break; /* we will be called again upon completion */
3368 case check_state_check_result
:
3369 sh
->check_state
= check_state_idle
;
3371 /* handle a successful check operation, if parity is correct
3372 * we are done. Otherwise update the mismatch count and repair
3373 * parity if !MD_RECOVERY_CHECK
3375 if (sh
->ops
.zero_sum_result
== 0) {
3376 /* both parities are correct */
3378 set_bit(STRIPE_INSYNC
, &sh
->state
);
3380 /* in contrast to the raid5 case we can validate
3381 * parity, but still have a failure to write
3384 sh
->check_state
= check_state_compute_result
;
3385 /* Returning at this point means that we may go
3386 * off and bring p and/or q uptodate again so
3387 * we make sure to check zero_sum_result again
3388 * to verify if p or q need writeback
3392 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3393 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3394 /* don't try to repair!! */
3395 set_bit(STRIPE_INSYNC
, &sh
->state
);
3397 int *target
= &sh
->ops
.target
;
3399 sh
->ops
.target
= -1;
3400 sh
->ops
.target2
= -1;
3401 sh
->check_state
= check_state_compute_run
;
3402 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3403 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3404 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3405 set_bit(R5_Wantcompute
,
3406 &sh
->dev
[pd_idx
].flags
);
3408 target
= &sh
->ops
.target2
;
3411 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3412 set_bit(R5_Wantcompute
,
3413 &sh
->dev
[qd_idx
].flags
);
3420 case check_state_compute_run
:
3423 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3424 __func__
, sh
->check_state
,
3425 (unsigned long long) sh
->sector
);
3430 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3434 /* We have read all the blocks in this stripe and now we need to
3435 * copy some of them into a target stripe for expand.
3437 struct dma_async_tx_descriptor
*tx
= NULL
;
3438 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3439 for (i
= 0; i
< sh
->disks
; i
++)
3440 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3442 struct stripe_head
*sh2
;
3443 struct async_submit_ctl submit
;
3445 sector_t bn
= compute_blocknr(sh
, i
, 1);
3446 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3448 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3450 /* so far only the early blocks of this stripe
3451 * have been requested. When later blocks
3452 * get requested, we will try again
3455 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3456 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3457 /* must have already done this block */
3458 release_stripe(sh2
);
3462 /* place all the copies on one channel */
3463 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3464 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3465 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3468 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3469 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3470 for (j
= 0; j
< conf
->raid_disks
; j
++)
3471 if (j
!= sh2
->pd_idx
&&
3473 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3475 if (j
== conf
->raid_disks
) {
3476 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3477 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3479 release_stripe(sh2
);
3482 /* done submitting copies, wait for them to complete */
3483 async_tx_quiesce(&tx
);
3487 * handle_stripe - do things to a stripe.
3489 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3490 * state of various bits to see what needs to be done.
3492 * return some read requests which now have data
3493 * return some write requests which are safely on storage
3494 * schedule a read on some buffers
3495 * schedule a write of some buffers
3496 * return confirmation of parity correctness
3500 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3502 struct r5conf
*conf
= sh
->raid_conf
;
3503 int disks
= sh
->disks
;
3506 int do_recovery
= 0;
3508 memset(s
, 0, sizeof(*s
));
3510 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3511 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3512 s
->failed_num
[0] = -1;
3513 s
->failed_num
[1] = -1;
3515 /* Now to look around and see what can be done */
3517 for (i
=disks
; i
--; ) {
3518 struct md_rdev
*rdev
;
3525 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3527 dev
->toread
, dev
->towrite
, dev
->written
);
3528 /* maybe we can reply to a read
3530 * new wantfill requests are only permitted while
3531 * ops_complete_biofill is guaranteed to be inactive
3533 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3534 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3535 set_bit(R5_Wantfill
, &dev
->flags
);
3537 /* now count some things */
3538 if (test_bit(R5_LOCKED
, &dev
->flags
))
3540 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3542 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3544 BUG_ON(s
->compute
> 2);
3547 if (test_bit(R5_Wantfill
, &dev
->flags
))
3549 else if (dev
->toread
)
3553 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3558 /* Prefer to use the replacement for reads, but only
3559 * if it is recovered enough and has no bad blocks.
3561 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3562 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3563 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3564 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3565 &first_bad
, &bad_sectors
))
3566 set_bit(R5_ReadRepl
, &dev
->flags
);
3569 set_bit(R5_NeedReplace
, &dev
->flags
);
3570 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3571 clear_bit(R5_ReadRepl
, &dev
->flags
);
3573 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3576 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3577 &first_bad
, &bad_sectors
);
3578 if (s
->blocked_rdev
== NULL
3579 && (test_bit(Blocked
, &rdev
->flags
)
3582 set_bit(BlockedBadBlocks
,
3584 s
->blocked_rdev
= rdev
;
3585 atomic_inc(&rdev
->nr_pending
);
3588 clear_bit(R5_Insync
, &dev
->flags
);
3592 /* also not in-sync */
3593 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
3594 test_bit(R5_UPTODATE
, &dev
->flags
)) {
3595 /* treat as in-sync, but with a read error
3596 * which we can now try to correct
3598 set_bit(R5_Insync
, &dev
->flags
);
3599 set_bit(R5_ReadError
, &dev
->flags
);
3601 } else if (test_bit(In_sync
, &rdev
->flags
))
3602 set_bit(R5_Insync
, &dev
->flags
);
3603 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3604 /* in sync if before recovery_offset */
3605 set_bit(R5_Insync
, &dev
->flags
);
3606 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3607 test_bit(R5_Expanded
, &dev
->flags
))
3608 /* If we've reshaped into here, we assume it is Insync.
3609 * We will shortly update recovery_offset to make
3612 set_bit(R5_Insync
, &dev
->flags
);
3614 if (rdev
&& test_bit(R5_WriteError
, &dev
->flags
)) {
3615 /* This flag does not apply to '.replacement'
3616 * only to .rdev, so make sure to check that*/
3617 struct md_rdev
*rdev2
= rcu_dereference(
3618 conf
->disks
[i
].rdev
);
3620 clear_bit(R5_Insync
, &dev
->flags
);
3621 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3622 s
->handle_bad_blocks
= 1;
3623 atomic_inc(&rdev2
->nr_pending
);
3625 clear_bit(R5_WriteError
, &dev
->flags
);
3627 if (rdev
&& test_bit(R5_MadeGood
, &dev
->flags
)) {
3628 /* This flag does not apply to '.replacement'
3629 * only to .rdev, so make sure to check that*/
3630 struct md_rdev
*rdev2
= rcu_dereference(
3631 conf
->disks
[i
].rdev
);
3632 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3633 s
->handle_bad_blocks
= 1;
3634 atomic_inc(&rdev2
->nr_pending
);
3636 clear_bit(R5_MadeGood
, &dev
->flags
);
3638 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3639 struct md_rdev
*rdev2
= rcu_dereference(
3640 conf
->disks
[i
].replacement
);
3641 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3642 s
->handle_bad_blocks
= 1;
3643 atomic_inc(&rdev2
->nr_pending
);
3645 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3647 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3648 /* The ReadError flag will just be confusing now */
3649 clear_bit(R5_ReadError
, &dev
->flags
);
3650 clear_bit(R5_ReWrite
, &dev
->flags
);
3652 if (test_bit(R5_ReadError
, &dev
->flags
))
3653 clear_bit(R5_Insync
, &dev
->flags
);
3654 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3656 s
->failed_num
[s
->failed
] = i
;
3658 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3662 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3663 /* If there is a failed device being replaced,
3664 * we must be recovering.
3665 * else if we are after recovery_cp, we must be syncing
3666 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3667 * else we can only be replacing
3668 * sync and recovery both need to read all devices, and so
3669 * use the same flag.
3672 sh
->sector
>= conf
->mddev
->recovery_cp
||
3673 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
3681 static void handle_stripe(struct stripe_head
*sh
)
3683 struct stripe_head_state s
;
3684 struct r5conf
*conf
= sh
->raid_conf
;
3687 int disks
= sh
->disks
;
3688 struct r5dev
*pdev
, *qdev
;
3690 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3691 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3692 /* already being handled, ensure it gets handled
3693 * again when current action finishes */
3694 set_bit(STRIPE_HANDLE
, &sh
->state
);
3698 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3699 spin_lock(&sh
->stripe_lock
);
3700 /* Cannot process 'sync' concurrently with 'discard' */
3701 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
3702 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3703 set_bit(STRIPE_SYNCING
, &sh
->state
);
3704 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3705 clear_bit(STRIPE_REPLACED
, &sh
->state
);
3707 spin_unlock(&sh
->stripe_lock
);
3709 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3711 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3712 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3713 (unsigned long long)sh
->sector
, sh
->state
,
3714 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3715 sh
->check_state
, sh
->reconstruct_state
);
3717 analyse_stripe(sh
, &s
);
3719 if (s
.handle_bad_blocks
) {
3720 set_bit(STRIPE_HANDLE
, &sh
->state
);
3724 if (unlikely(s
.blocked_rdev
)) {
3725 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3726 s
.replacing
|| s
.to_write
|| s
.written
) {
3727 set_bit(STRIPE_HANDLE
, &sh
->state
);
3730 /* There is nothing for the blocked_rdev to block */
3731 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3732 s
.blocked_rdev
= NULL
;
3735 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3736 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3737 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3740 pr_debug("locked=%d uptodate=%d to_read=%d"
3741 " to_write=%d failed=%d failed_num=%d,%d\n",
3742 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3743 s
.failed_num
[0], s
.failed_num
[1]);
3744 /* check if the array has lost more than max_degraded devices and,
3745 * if so, some requests might need to be failed.
3747 if (s
.failed
> conf
->max_degraded
) {
3748 sh
->check_state
= 0;
3749 sh
->reconstruct_state
= 0;
3750 if (s
.to_read
+s
.to_write
+s
.written
)
3751 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3752 if (s
.syncing
+ s
.replacing
)
3753 handle_failed_sync(conf
, sh
, &s
);
3756 /* Now we check to see if any write operations have recently
3760 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3762 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3763 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3764 sh
->reconstruct_state
= reconstruct_state_idle
;
3766 /* All the 'written' buffers and the parity block are ready to
3767 * be written back to disk
3769 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
3770 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
3771 BUG_ON(sh
->qd_idx
>= 0 &&
3772 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
3773 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
3774 for (i
= disks
; i
--; ) {
3775 struct r5dev
*dev
= &sh
->dev
[i
];
3776 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3777 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3779 pr_debug("Writing block %d\n", i
);
3780 set_bit(R5_Wantwrite
, &dev
->flags
);
3783 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3784 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3786 set_bit(STRIPE_INSYNC
, &sh
->state
);
3789 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3790 s
.dec_preread_active
= 1;
3794 * might be able to return some write requests if the parity blocks
3795 * are safe, or on a failed drive
3797 pdev
= &sh
->dev
[sh
->pd_idx
];
3798 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3799 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3800 qdev
= &sh
->dev
[sh
->qd_idx
];
3801 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3802 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3806 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3807 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3808 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
3809 test_bit(R5_Discard
, &pdev
->flags
))))) &&
3810 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3811 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3812 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
3813 test_bit(R5_Discard
, &qdev
->flags
))))))
3814 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3816 /* Now we might consider reading some blocks, either to check/generate
3817 * parity, or to satisfy requests
3818 * or to load a block that is being partially written.
3820 if (s
.to_read
|| s
.non_overwrite
3821 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3822 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3825 handle_stripe_fill(sh
, &s
, disks
);
3827 /* Now to consider new write requests and what else, if anything
3828 * should be read. We do not handle new writes when:
3829 * 1/ A 'write' operation (copy+xor) is already in flight.
3830 * 2/ A 'check' operation is in flight, as it may clobber the parity
3833 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3834 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3836 /* maybe we need to check and possibly fix the parity for this stripe
3837 * Any reads will already have been scheduled, so we just see if enough
3838 * data is available. The parity check is held off while parity
3839 * dependent operations are in flight.
3841 if (sh
->check_state
||
3842 (s
.syncing
&& s
.locked
== 0 &&
3843 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3844 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3845 if (conf
->level
== 6)
3846 handle_parity_checks6(conf
, sh
, &s
, disks
);
3848 handle_parity_checks5(conf
, sh
, &s
, disks
);
3851 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
3852 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
3853 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
3854 /* Write out to replacement devices where possible */
3855 for (i
= 0; i
< conf
->raid_disks
; i
++)
3856 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3857 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3858 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3859 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3863 set_bit(STRIPE_INSYNC
, &sh
->state
);
3864 set_bit(STRIPE_REPLACED
, &sh
->state
);
3866 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3867 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3868 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3869 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3870 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3871 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3872 wake_up(&conf
->wait_for_overlap
);
3875 /* If the failed drives are just a ReadError, then we might need
3876 * to progress the repair/check process
3878 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3879 for (i
= 0; i
< s
.failed
; i
++) {
3880 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3881 if (test_bit(R5_ReadError
, &dev
->flags
)
3882 && !test_bit(R5_LOCKED
, &dev
->flags
)
3883 && test_bit(R5_UPTODATE
, &dev
->flags
)
3885 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3886 set_bit(R5_Wantwrite
, &dev
->flags
);
3887 set_bit(R5_ReWrite
, &dev
->flags
);
3888 set_bit(R5_LOCKED
, &dev
->flags
);
3891 /* let's read it back */
3892 set_bit(R5_Wantread
, &dev
->flags
);
3893 set_bit(R5_LOCKED
, &dev
->flags
);
3900 /* Finish reconstruct operations initiated by the expansion process */
3901 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3902 struct stripe_head
*sh_src
3903 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3904 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3905 /* sh cannot be written until sh_src has been read.
3906 * so arrange for sh to be delayed a little
3908 set_bit(STRIPE_DELAYED
, &sh
->state
);
3909 set_bit(STRIPE_HANDLE
, &sh
->state
);
3910 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3912 atomic_inc(&conf
->preread_active_stripes
);
3913 release_stripe(sh_src
);
3917 release_stripe(sh_src
);
3919 sh
->reconstruct_state
= reconstruct_state_idle
;
3920 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3921 for (i
= conf
->raid_disks
; i
--; ) {
3922 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3923 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3928 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3929 !sh
->reconstruct_state
) {
3930 /* Need to write out all blocks after computing parity */
3931 sh
->disks
= conf
->raid_disks
;
3932 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3933 schedule_reconstruction(sh
, &s
, 1, 1);
3934 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3935 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3936 atomic_dec(&conf
->reshape_stripes
);
3937 wake_up(&conf
->wait_for_overlap
);
3938 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3941 if (s
.expanding
&& s
.locked
== 0 &&
3942 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3943 handle_stripe_expansion(conf
, sh
);
3946 /* wait for this device to become unblocked */
3947 if (unlikely(s
.blocked_rdev
)) {
3948 if (conf
->mddev
->external
)
3949 md_wait_for_blocked_rdev(s
.blocked_rdev
,
3952 /* Internal metadata will immediately
3953 * be written by raid5d, so we don't
3954 * need to wait here.
3956 rdev_dec_pending(s
.blocked_rdev
,
3960 if (s
.handle_bad_blocks
)
3961 for (i
= disks
; i
--; ) {
3962 struct md_rdev
*rdev
;
3963 struct r5dev
*dev
= &sh
->dev
[i
];
3964 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3965 /* We own a safe reference to the rdev */
3966 rdev
= conf
->disks
[i
].rdev
;
3967 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3969 md_error(conf
->mddev
, rdev
);
3970 rdev_dec_pending(rdev
, conf
->mddev
);
3972 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3973 rdev
= conf
->disks
[i
].rdev
;
3974 rdev_clear_badblocks(rdev
, sh
->sector
,
3976 rdev_dec_pending(rdev
, conf
->mddev
);
3978 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3979 rdev
= conf
->disks
[i
].replacement
;
3981 /* rdev have been moved down */
3982 rdev
= conf
->disks
[i
].rdev
;
3983 rdev_clear_badblocks(rdev
, sh
->sector
,
3985 rdev_dec_pending(rdev
, conf
->mddev
);
3990 raid_run_ops(sh
, s
.ops_request
);
3994 if (s
.dec_preread_active
) {
3995 /* We delay this until after ops_run_io so that if make_request
3996 * is waiting on a flush, it won't continue until the writes
3997 * have actually been submitted.
3999 atomic_dec(&conf
->preread_active_stripes
);
4000 if (atomic_read(&conf
->preread_active_stripes
) <
4002 md_wakeup_thread(conf
->mddev
->thread
);
4005 return_io(s
.return_bi
);
4007 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4010 static void raid5_activate_delayed(struct r5conf
*conf
)
4012 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
4013 while (!list_empty(&conf
->delayed_list
)) {
4014 struct list_head
*l
= conf
->delayed_list
.next
;
4015 struct stripe_head
*sh
;
4016 sh
= list_entry(l
, struct stripe_head
, lru
);
4018 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4019 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4020 atomic_inc(&conf
->preread_active_stripes
);
4021 list_add_tail(&sh
->lru
, &conf
->hold_list
);
4022 raid5_wakeup_stripe_thread(sh
);
4027 static void activate_bit_delay(struct r5conf
*conf
,
4028 struct list_head
*temp_inactive_list
)
4030 /* device_lock is held */
4031 struct list_head head
;
4032 list_add(&head
, &conf
->bitmap_list
);
4033 list_del_init(&conf
->bitmap_list
);
4034 while (!list_empty(&head
)) {
4035 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
4037 list_del_init(&sh
->lru
);
4038 atomic_inc(&sh
->count
);
4039 hash
= sh
->hash_lock_index
;
4040 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
4044 int md_raid5_congested(struct mddev
*mddev
, int bits
)
4046 struct r5conf
*conf
= mddev
->private;
4048 /* No difference between reads and writes. Just check
4049 * how busy the stripe_cache is
4052 if (conf
->inactive_blocked
)
4056 if (atomic_read(&conf
->empty_inactive_list_nr
))
4061 EXPORT_SYMBOL_GPL(md_raid5_congested
);
4063 static int raid5_congested(void *data
, int bits
)
4065 struct mddev
*mddev
= data
;
4067 return mddev_congested(mddev
, bits
) ||
4068 md_raid5_congested(mddev
, bits
);
4071 /* We want read requests to align with chunks where possible,
4072 * but write requests don't need to.
4074 static int raid5_mergeable_bvec(struct request_queue
*q
,
4075 struct bvec_merge_data
*bvm
,
4076 struct bio_vec
*biovec
)
4078 struct mddev
*mddev
= q
->queuedata
;
4079 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
4081 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4082 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
4084 if ((bvm
->bi_rw
& 1) == WRITE
)
4085 return biovec
->bv_len
; /* always allow writes to be mergeable */
4087 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4088 chunk_sectors
= mddev
->new_chunk_sectors
;
4089 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
4090 if (max
< 0) max
= 0;
4091 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
4092 return biovec
->bv_len
;
4098 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
4100 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
4101 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
4102 unsigned int bio_sectors
= bio_sectors(bio
);
4104 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
4105 chunk_sectors
= mddev
->new_chunk_sectors
;
4106 return chunk_sectors
>=
4107 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
4111 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4112 * later sampled by raid5d.
4114 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
4116 unsigned long flags
;
4118 spin_lock_irqsave(&conf
->device_lock
, flags
);
4120 bi
->bi_next
= conf
->retry_read_aligned_list
;
4121 conf
->retry_read_aligned_list
= bi
;
4123 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4124 md_wakeup_thread(conf
->mddev
->thread
);
4128 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
4132 bi
= conf
->retry_read_aligned
;
4134 conf
->retry_read_aligned
= NULL
;
4137 bi
= conf
->retry_read_aligned_list
;
4139 conf
->retry_read_aligned_list
= bi
->bi_next
;
4142 * this sets the active strip count to 1 and the processed
4143 * strip count to zero (upper 8 bits)
4145 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
4153 * The "raid5_align_endio" should check if the read succeeded and if it
4154 * did, call bio_endio on the original bio (having bio_put the new bio
4156 * If the read failed..
4158 static void raid5_align_endio(struct bio
*bi
, int error
)
4160 struct bio
* raid_bi
= bi
->bi_private
;
4161 struct mddev
*mddev
;
4162 struct r5conf
*conf
;
4163 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4164 struct md_rdev
*rdev
;
4168 rdev
= (void*)raid_bi
->bi_next
;
4169 raid_bi
->bi_next
= NULL
;
4170 mddev
= rdev
->mddev
;
4171 conf
= mddev
->private;
4173 rdev_dec_pending(rdev
, conf
->mddev
);
4175 if (!error
&& uptodate
) {
4176 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
4178 bio_endio(raid_bi
, 0);
4179 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4180 wake_up(&conf
->wait_for_stripe
);
4185 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4187 add_bio_to_retry(raid_bi
, conf
);
4190 static int bio_fits_rdev(struct bio
*bi
)
4192 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
4194 if (bio_sectors(bi
) > queue_max_sectors(q
))
4196 blk_recount_segments(q
, bi
);
4197 if (bi
->bi_phys_segments
> queue_max_segments(q
))
4200 if (q
->merge_bvec_fn
)
4201 /* it's too hard to apply the merge_bvec_fn at this stage,
4210 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
4212 struct r5conf
*conf
= mddev
->private;
4214 struct bio
* align_bi
;
4215 struct md_rdev
*rdev
;
4216 sector_t end_sector
;
4218 if (!in_chunk_boundary(mddev
, raid_bio
)) {
4219 pr_debug("chunk_aligned_read : non aligned\n");
4223 * use bio_clone_mddev to make a copy of the bio
4225 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
4229 * set bi_end_io to a new function, and set bi_private to the
4232 align_bi
->bi_end_io
= raid5_align_endio
;
4233 align_bi
->bi_private
= raid_bio
;
4237 align_bi
->bi_iter
.bi_sector
=
4238 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
4241 end_sector
= bio_end_sector(align_bi
);
4243 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
4244 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
4245 rdev
->recovery_offset
< end_sector
) {
4246 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
4248 (test_bit(Faulty
, &rdev
->flags
) ||
4249 !(test_bit(In_sync
, &rdev
->flags
) ||
4250 rdev
->recovery_offset
>= end_sector
)))
4257 atomic_inc(&rdev
->nr_pending
);
4259 raid_bio
->bi_next
= (void*)rdev
;
4260 align_bi
->bi_bdev
= rdev
->bdev
;
4261 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
4263 if (!bio_fits_rdev(align_bi
) ||
4264 is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
4265 bio_sectors(align_bi
),
4266 &first_bad
, &bad_sectors
)) {
4267 /* too big in some way, or has a known bad block */
4269 rdev_dec_pending(rdev
, mddev
);
4273 /* No reshape active, so we can trust rdev->data_offset */
4274 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
4276 spin_lock_irq(&conf
->device_lock
);
4277 wait_event_lock_irq(conf
->wait_for_stripe
,
4280 atomic_inc(&conf
->active_aligned_reads
);
4281 spin_unlock_irq(&conf
->device_lock
);
4284 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4285 align_bi
, disk_devt(mddev
->gendisk
),
4286 raid_bio
->bi_iter
.bi_sector
);
4287 generic_make_request(align_bi
);
4296 /* __get_priority_stripe - get the next stripe to process
4298 * Full stripe writes are allowed to pass preread active stripes up until
4299 * the bypass_threshold is exceeded. In general the bypass_count
4300 * increments when the handle_list is handled before the hold_list; however, it
4301 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4302 * stripe with in flight i/o. The bypass_count will be reset when the
4303 * head of the hold_list has changed, i.e. the head was promoted to the
4306 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
4308 struct stripe_head
*sh
= NULL
, *tmp
;
4309 struct list_head
*handle_list
= NULL
;
4310 struct r5worker_group
*wg
= NULL
;
4312 if (conf
->worker_cnt_per_group
== 0) {
4313 handle_list
= &conf
->handle_list
;
4314 } else if (group
!= ANY_GROUP
) {
4315 handle_list
= &conf
->worker_groups
[group
].handle_list
;
4316 wg
= &conf
->worker_groups
[group
];
4319 for (i
= 0; i
< conf
->group_cnt
; i
++) {
4320 handle_list
= &conf
->worker_groups
[i
].handle_list
;
4321 wg
= &conf
->worker_groups
[i
];
4322 if (!list_empty(handle_list
))
4327 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4329 list_empty(handle_list
) ? "empty" : "busy",
4330 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4331 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4333 if (!list_empty(handle_list
)) {
4334 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
4336 if (list_empty(&conf
->hold_list
))
4337 conf
->bypass_count
= 0;
4338 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4339 if (conf
->hold_list
.next
== conf
->last_hold
)
4340 conf
->bypass_count
++;
4342 conf
->last_hold
= conf
->hold_list
.next
;
4343 conf
->bypass_count
-= conf
->bypass_threshold
;
4344 if (conf
->bypass_count
< 0)
4345 conf
->bypass_count
= 0;
4348 } else if (!list_empty(&conf
->hold_list
) &&
4349 ((conf
->bypass_threshold
&&
4350 conf
->bypass_count
> conf
->bypass_threshold
) ||
4351 atomic_read(&conf
->pending_full_writes
) == 0)) {
4353 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
4354 if (conf
->worker_cnt_per_group
== 0 ||
4355 group
== ANY_GROUP
||
4356 !cpu_online(tmp
->cpu
) ||
4357 cpu_to_group(tmp
->cpu
) == group
) {
4364 conf
->bypass_count
-= conf
->bypass_threshold
;
4365 if (conf
->bypass_count
< 0)
4366 conf
->bypass_count
= 0;
4378 list_del_init(&sh
->lru
);
4379 atomic_inc(&sh
->count
);
4380 BUG_ON(atomic_read(&sh
->count
) != 1);
4384 struct raid5_plug_cb
{
4385 struct blk_plug_cb cb
;
4386 struct list_head list
;
4387 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
4390 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4392 struct raid5_plug_cb
*cb
= container_of(
4393 blk_cb
, struct raid5_plug_cb
, cb
);
4394 struct stripe_head
*sh
;
4395 struct mddev
*mddev
= cb
->cb
.data
;
4396 struct r5conf
*conf
= mddev
->private;
4400 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4401 spin_lock_irq(&conf
->device_lock
);
4402 while (!list_empty(&cb
->list
)) {
4403 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4404 list_del_init(&sh
->lru
);
4406 * avoid race release_stripe_plug() sees
4407 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4408 * is still in our list
4410 smp_mb__before_clear_bit();
4411 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4413 * STRIPE_ON_RELEASE_LIST could be set here. In that
4414 * case, the count is always > 1 here
4416 hash
= sh
->hash_lock_index
;
4417 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
4420 spin_unlock_irq(&conf
->device_lock
);
4422 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
4423 NR_STRIPE_HASH_LOCKS
);
4425 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
4429 static void release_stripe_plug(struct mddev
*mddev
,
4430 struct stripe_head
*sh
)
4432 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4433 raid5_unplug
, mddev
,
4434 sizeof(struct raid5_plug_cb
));
4435 struct raid5_plug_cb
*cb
;
4442 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4444 if (cb
->list
.next
== NULL
) {
4446 INIT_LIST_HEAD(&cb
->list
);
4447 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
4448 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
4451 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
4452 list_add_tail(&sh
->lru
, &cb
->list
);
4457 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
4459 struct r5conf
*conf
= mddev
->private;
4460 sector_t logical_sector
, last_sector
;
4461 struct stripe_head
*sh
;
4465 if (mddev
->reshape_position
!= MaxSector
)
4466 /* Skip discard while reshape is happening */
4469 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4470 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
4473 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4475 stripe_sectors
= conf
->chunk_sectors
*
4476 (conf
->raid_disks
- conf
->max_degraded
);
4477 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
4479 sector_div(last_sector
, stripe_sectors
);
4481 logical_sector
*= conf
->chunk_sectors
;
4482 last_sector
*= conf
->chunk_sectors
;
4484 for (; logical_sector
< last_sector
;
4485 logical_sector
+= STRIPE_SECTORS
) {
4489 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
4490 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
4491 TASK_UNINTERRUPTIBLE
);
4492 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4493 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4498 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4499 spin_lock_irq(&sh
->stripe_lock
);
4500 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4501 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4503 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
4504 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
4505 spin_unlock_irq(&sh
->stripe_lock
);
4511 set_bit(STRIPE_DISCARD
, &sh
->state
);
4512 finish_wait(&conf
->wait_for_overlap
, &w
);
4513 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4514 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4516 sh
->dev
[d
].towrite
= bi
;
4517 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
4518 raid5_inc_bi_active_stripes(bi
);
4520 spin_unlock_irq(&sh
->stripe_lock
);
4521 if (conf
->mddev
->bitmap
) {
4523 d
< conf
->raid_disks
- conf
->max_degraded
;
4525 bitmap_startwrite(mddev
->bitmap
,
4529 sh
->bm_seq
= conf
->seq_flush
+ 1;
4530 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
4533 set_bit(STRIPE_HANDLE
, &sh
->state
);
4534 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4535 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4536 atomic_inc(&conf
->preread_active_stripes
);
4537 release_stripe_plug(mddev
, sh
);
4540 remaining
= raid5_dec_bi_active_stripes(bi
);
4541 if (remaining
== 0) {
4542 md_write_end(mddev
);
4547 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
4549 struct r5conf
*conf
= mddev
->private;
4551 sector_t new_sector
;
4552 sector_t logical_sector
, last_sector
;
4553 struct stripe_head
*sh
;
4554 const int rw
= bio_data_dir(bi
);
4557 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
4558 md_flush_request(mddev
, bi
);
4562 md_write_start(mddev
, bi
);
4565 mddev
->reshape_position
== MaxSector
&&
4566 chunk_aligned_read(mddev
,bi
))
4569 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
4570 make_discard_request(mddev
, bi
);
4574 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4575 last_sector
= bio_end_sector(bi
);
4577 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4579 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4585 seq
= read_seqcount_begin(&conf
->gen_lock
);
4587 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4588 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4589 /* spinlock is needed as reshape_progress may be
4590 * 64bit on a 32bit platform, and so it might be
4591 * possible to see a half-updated value
4592 * Of course reshape_progress could change after
4593 * the lock is dropped, so once we get a reference
4594 * to the stripe that we think it is, we will have
4597 spin_lock_irq(&conf
->device_lock
);
4598 if (mddev
->reshape_backwards
4599 ? logical_sector
< conf
->reshape_progress
4600 : logical_sector
>= conf
->reshape_progress
) {
4603 if (mddev
->reshape_backwards
4604 ? logical_sector
< conf
->reshape_safe
4605 : logical_sector
>= conf
->reshape_safe
) {
4606 spin_unlock_irq(&conf
->device_lock
);
4611 spin_unlock_irq(&conf
->device_lock
);
4614 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4617 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4618 (unsigned long long)new_sector
,
4619 (unsigned long long)logical_sector
);
4621 sh
= get_active_stripe(conf
, new_sector
, previous
,
4622 (bi
->bi_rw
&RWA_MASK
), 0);
4624 if (unlikely(previous
)) {
4625 /* expansion might have moved on while waiting for a
4626 * stripe, so we must do the range check again.
4627 * Expansion could still move past after this
4628 * test, but as we are holding a reference to
4629 * 'sh', we know that if that happens,
4630 * STRIPE_EXPANDING will get set and the expansion
4631 * won't proceed until we finish with the stripe.
4634 spin_lock_irq(&conf
->device_lock
);
4635 if (mddev
->reshape_backwards
4636 ? logical_sector
>= conf
->reshape_progress
4637 : logical_sector
< conf
->reshape_progress
)
4638 /* mismatch, need to try again */
4640 spin_unlock_irq(&conf
->device_lock
);
4647 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
4648 /* Might have got the wrong stripe_head
4656 logical_sector
>= mddev
->suspend_lo
&&
4657 logical_sector
< mddev
->suspend_hi
) {
4659 /* As the suspend_* range is controlled by
4660 * userspace, we want an interruptible
4663 flush_signals(current
);
4664 prepare_to_wait(&conf
->wait_for_overlap
,
4665 &w
, TASK_INTERRUPTIBLE
);
4666 if (logical_sector
>= mddev
->suspend_lo
&&
4667 logical_sector
< mddev
->suspend_hi
)
4672 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4673 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4674 /* Stripe is busy expanding or
4675 * add failed due to overlap. Flush everything
4678 md_wakeup_thread(mddev
->thread
);
4683 finish_wait(&conf
->wait_for_overlap
, &w
);
4684 set_bit(STRIPE_HANDLE
, &sh
->state
);
4685 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4686 if ((bi
->bi_rw
& REQ_SYNC
) &&
4687 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4688 atomic_inc(&conf
->preread_active_stripes
);
4689 release_stripe_plug(mddev
, sh
);
4691 /* cannot get stripe for read-ahead, just give-up */
4692 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4693 finish_wait(&conf
->wait_for_overlap
, &w
);
4698 remaining
= raid5_dec_bi_active_stripes(bi
);
4699 if (remaining
== 0) {
4702 md_write_end(mddev
);
4704 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
4710 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4712 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4714 /* reshaping is quite different to recovery/resync so it is
4715 * handled quite separately ... here.
4717 * On each call to sync_request, we gather one chunk worth of
4718 * destination stripes and flag them as expanding.
4719 * Then we find all the source stripes and request reads.
4720 * As the reads complete, handle_stripe will copy the data
4721 * into the destination stripe and release that stripe.
4723 struct r5conf
*conf
= mddev
->private;
4724 struct stripe_head
*sh
;
4725 sector_t first_sector
, last_sector
;
4726 int raid_disks
= conf
->previous_raid_disks
;
4727 int data_disks
= raid_disks
- conf
->max_degraded
;
4728 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4731 sector_t writepos
, readpos
, safepos
;
4732 sector_t stripe_addr
;
4733 int reshape_sectors
;
4734 struct list_head stripes
;
4736 if (sector_nr
== 0) {
4737 /* If restarting in the middle, skip the initial sectors */
4738 if (mddev
->reshape_backwards
&&
4739 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4740 sector_nr
= raid5_size(mddev
, 0, 0)
4741 - conf
->reshape_progress
;
4742 } else if (!mddev
->reshape_backwards
&&
4743 conf
->reshape_progress
> 0)
4744 sector_nr
= conf
->reshape_progress
;
4745 sector_div(sector_nr
, new_data_disks
);
4747 mddev
->curr_resync_completed
= sector_nr
;
4748 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4754 /* We need to process a full chunk at a time.
4755 * If old and new chunk sizes differ, we need to process the
4758 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4759 reshape_sectors
= mddev
->new_chunk_sectors
;
4761 reshape_sectors
= mddev
->chunk_sectors
;
4763 /* We update the metadata at least every 10 seconds, or when
4764 * the data about to be copied would over-write the source of
4765 * the data at the front of the range. i.e. one new_stripe
4766 * along from reshape_progress new_maps to after where
4767 * reshape_safe old_maps to
4769 writepos
= conf
->reshape_progress
;
4770 sector_div(writepos
, new_data_disks
);
4771 readpos
= conf
->reshape_progress
;
4772 sector_div(readpos
, data_disks
);
4773 safepos
= conf
->reshape_safe
;
4774 sector_div(safepos
, data_disks
);
4775 if (mddev
->reshape_backwards
) {
4776 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4777 readpos
+= reshape_sectors
;
4778 safepos
+= reshape_sectors
;
4780 writepos
+= reshape_sectors
;
4781 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4782 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4785 /* Having calculated the 'writepos' possibly use it
4786 * to set 'stripe_addr' which is where we will write to.
4788 if (mddev
->reshape_backwards
) {
4789 BUG_ON(conf
->reshape_progress
== 0);
4790 stripe_addr
= writepos
;
4791 BUG_ON((mddev
->dev_sectors
&
4792 ~((sector_t
)reshape_sectors
- 1))
4793 - reshape_sectors
- stripe_addr
4796 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4797 stripe_addr
= sector_nr
;
4800 /* 'writepos' is the most advanced device address we might write.
4801 * 'readpos' is the least advanced device address we might read.
4802 * 'safepos' is the least address recorded in the metadata as having
4804 * If there is a min_offset_diff, these are adjusted either by
4805 * increasing the safepos/readpos if diff is negative, or
4806 * increasing writepos if diff is positive.
4807 * If 'readpos' is then behind 'writepos', there is no way that we can
4808 * ensure safety in the face of a crash - that must be done by userspace
4809 * making a backup of the data. So in that case there is no particular
4810 * rush to update metadata.
4811 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4812 * update the metadata to advance 'safepos' to match 'readpos' so that
4813 * we can be safe in the event of a crash.
4814 * So we insist on updating metadata if safepos is behind writepos and
4815 * readpos is beyond writepos.
4816 * In any case, update the metadata every 10 seconds.
4817 * Maybe that number should be configurable, but I'm not sure it is
4818 * worth it.... maybe it could be a multiple of safemode_delay???
4820 if (conf
->min_offset_diff
< 0) {
4821 safepos
+= -conf
->min_offset_diff
;
4822 readpos
+= -conf
->min_offset_diff
;
4824 writepos
+= conf
->min_offset_diff
;
4826 if ((mddev
->reshape_backwards
4827 ? (safepos
> writepos
&& readpos
< writepos
)
4828 : (safepos
< writepos
&& readpos
> writepos
)) ||
4829 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4830 /* Cannot proceed until we've updated the superblock... */
4831 wait_event(conf
->wait_for_overlap
,
4832 atomic_read(&conf
->reshape_stripes
)==0
4833 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4834 if (atomic_read(&conf
->reshape_stripes
) != 0)
4836 mddev
->reshape_position
= conf
->reshape_progress
;
4837 mddev
->curr_resync_completed
= sector_nr
;
4838 conf
->reshape_checkpoint
= jiffies
;
4839 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4840 md_wakeup_thread(mddev
->thread
);
4841 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4842 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4843 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4845 spin_lock_irq(&conf
->device_lock
);
4846 conf
->reshape_safe
= mddev
->reshape_position
;
4847 spin_unlock_irq(&conf
->device_lock
);
4848 wake_up(&conf
->wait_for_overlap
);
4849 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4852 INIT_LIST_HEAD(&stripes
);
4853 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4855 int skipped_disk
= 0;
4856 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4857 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4858 atomic_inc(&conf
->reshape_stripes
);
4859 /* If any of this stripe is beyond the end of the old
4860 * array, then we need to zero those blocks
4862 for (j
=sh
->disks
; j
--;) {
4864 if (j
== sh
->pd_idx
)
4866 if (conf
->level
== 6 &&
4869 s
= compute_blocknr(sh
, j
, 0);
4870 if (s
< raid5_size(mddev
, 0, 0)) {
4874 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4875 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4876 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4878 if (!skipped_disk
) {
4879 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4880 set_bit(STRIPE_HANDLE
, &sh
->state
);
4882 list_add(&sh
->lru
, &stripes
);
4884 spin_lock_irq(&conf
->device_lock
);
4885 if (mddev
->reshape_backwards
)
4886 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4888 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4889 spin_unlock_irq(&conf
->device_lock
);
4890 /* Ok, those stripe are ready. We can start scheduling
4891 * reads on the source stripes.
4892 * The source stripes are determined by mapping the first and last
4893 * block on the destination stripes.
4896 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4899 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4900 * new_data_disks
- 1),
4902 if (last_sector
>= mddev
->dev_sectors
)
4903 last_sector
= mddev
->dev_sectors
- 1;
4904 while (first_sector
<= last_sector
) {
4905 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4906 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4907 set_bit(STRIPE_HANDLE
, &sh
->state
);
4909 first_sector
+= STRIPE_SECTORS
;
4911 /* Now that the sources are clearly marked, we can release
4912 * the destination stripes
4914 while (!list_empty(&stripes
)) {
4915 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4916 list_del_init(&sh
->lru
);
4919 /* If this takes us to the resync_max point where we have to pause,
4920 * then we need to write out the superblock.
4922 sector_nr
+= reshape_sectors
;
4923 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4924 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4925 /* Cannot proceed until we've updated the superblock... */
4926 wait_event(conf
->wait_for_overlap
,
4927 atomic_read(&conf
->reshape_stripes
) == 0
4928 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4929 if (atomic_read(&conf
->reshape_stripes
) != 0)
4931 mddev
->reshape_position
= conf
->reshape_progress
;
4932 mddev
->curr_resync_completed
= sector_nr
;
4933 conf
->reshape_checkpoint
= jiffies
;
4934 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4935 md_wakeup_thread(mddev
->thread
);
4936 wait_event(mddev
->sb_wait
,
4937 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4938 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4939 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4941 spin_lock_irq(&conf
->device_lock
);
4942 conf
->reshape_safe
= mddev
->reshape_position
;
4943 spin_unlock_irq(&conf
->device_lock
);
4944 wake_up(&conf
->wait_for_overlap
);
4945 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4948 return reshape_sectors
;
4951 /* FIXME go_faster isn't used */
4952 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4954 struct r5conf
*conf
= mddev
->private;
4955 struct stripe_head
*sh
;
4956 sector_t max_sector
= mddev
->dev_sectors
;
4957 sector_t sync_blocks
;
4958 int still_degraded
= 0;
4961 if (sector_nr
>= max_sector
) {
4962 /* just being told to finish up .. nothing much to do */
4964 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4969 if (mddev
->curr_resync
< max_sector
) /* aborted */
4970 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4972 else /* completed sync */
4974 bitmap_close_sync(mddev
->bitmap
);
4979 /* Allow raid5_quiesce to complete */
4980 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4982 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4983 return reshape_request(mddev
, sector_nr
, skipped
);
4985 /* No need to check resync_max as we never do more than one
4986 * stripe, and as resync_max will always be on a chunk boundary,
4987 * if the check in md_do_sync didn't fire, there is no chance
4988 * of overstepping resync_max here
4991 /* if there is too many failed drives and we are trying
4992 * to resync, then assert that we are finished, because there is
4993 * nothing we can do.
4995 if (mddev
->degraded
>= conf
->max_degraded
&&
4996 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4997 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
5001 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
5003 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
5004 sync_blocks
>= STRIPE_SECTORS
) {
5005 /* we can skip this block, and probably more */
5006 sync_blocks
/= STRIPE_SECTORS
;
5008 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
5011 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
5013 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
5015 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
5016 /* make sure we don't swamp the stripe cache if someone else
5017 * is trying to get access
5019 schedule_timeout_uninterruptible(1);
5021 /* Need to check if array will still be degraded after recovery/resync
5022 * We don't need to check the 'failed' flag as when that gets set,
5025 for (i
= 0; i
< conf
->raid_disks
; i
++)
5026 if (conf
->disks
[i
].rdev
== NULL
)
5029 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
5031 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
5036 return STRIPE_SECTORS
;
5039 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
5041 /* We may not be able to submit a whole bio at once as there
5042 * may not be enough stripe_heads available.
5043 * We cannot pre-allocate enough stripe_heads as we may need
5044 * more than exist in the cache (if we allow ever large chunks).
5045 * So we do one stripe head at a time and record in
5046 * ->bi_hw_segments how many have been done.
5048 * We *know* that this entire raid_bio is in one chunk, so
5049 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5051 struct stripe_head
*sh
;
5053 sector_t sector
, logical_sector
, last_sector
;
5058 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
5059 ~((sector_t
)STRIPE_SECTORS
-1);
5060 sector
= raid5_compute_sector(conf
, logical_sector
,
5062 last_sector
= bio_end_sector(raid_bio
);
5064 for (; logical_sector
< last_sector
;
5065 logical_sector
+= STRIPE_SECTORS
,
5066 sector
+= STRIPE_SECTORS
,
5069 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
5070 /* already done this stripe */
5073 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
5076 /* failed to get a stripe - must wait */
5077 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5078 conf
->retry_read_aligned
= raid_bio
;
5082 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
5084 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
5085 conf
->retry_read_aligned
= raid_bio
;
5089 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
5094 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
5095 if (remaining
== 0) {
5096 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
5098 bio_endio(raid_bio
, 0);
5100 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5101 wake_up(&conf
->wait_for_stripe
);
5105 static int handle_active_stripes(struct r5conf
*conf
, int group
,
5106 struct r5worker
*worker
,
5107 struct list_head
*temp_inactive_list
)
5109 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
5110 int i
, batch_size
= 0, hash
;
5111 bool release_inactive
= false;
5113 while (batch_size
< MAX_STRIPE_BATCH
&&
5114 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
5115 batch
[batch_size
++] = sh
;
5117 if (batch_size
== 0) {
5118 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5119 if (!list_empty(temp_inactive_list
+ i
))
5121 if (i
== NR_STRIPE_HASH_LOCKS
)
5123 release_inactive
= true;
5125 spin_unlock_irq(&conf
->device_lock
);
5127 release_inactive_stripe_list(conf
, temp_inactive_list
,
5128 NR_STRIPE_HASH_LOCKS
);
5130 if (release_inactive
) {
5131 spin_lock_irq(&conf
->device_lock
);
5135 for (i
= 0; i
< batch_size
; i
++)
5136 handle_stripe(batch
[i
]);
5140 spin_lock_irq(&conf
->device_lock
);
5141 for (i
= 0; i
< batch_size
; i
++) {
5142 hash
= batch
[i
]->hash_lock_index
;
5143 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
5148 static void raid5_do_work(struct work_struct
*work
)
5150 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
5151 struct r5worker_group
*group
= worker
->group
;
5152 struct r5conf
*conf
= group
->conf
;
5153 int group_id
= group
- conf
->worker_groups
;
5155 struct blk_plug plug
;
5157 pr_debug("+++ raid5worker active\n");
5159 blk_start_plug(&plug
);
5161 spin_lock_irq(&conf
->device_lock
);
5163 int batch_size
, released
;
5165 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
5167 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
5168 worker
->temp_inactive_list
);
5169 worker
->working
= false;
5170 if (!batch_size
&& !released
)
5172 handled
+= batch_size
;
5174 pr_debug("%d stripes handled\n", handled
);
5176 spin_unlock_irq(&conf
->device_lock
);
5177 blk_finish_plug(&plug
);
5179 pr_debug("--- raid5worker inactive\n");
5183 * This is our raid5 kernel thread.
5185 * We scan the hash table for stripes which can be handled now.
5186 * During the scan, completed stripes are saved for us by the interrupt
5187 * handler, so that they will not have to wait for our next wakeup.
5189 static void raid5d(struct md_thread
*thread
)
5191 struct mddev
*mddev
= thread
->mddev
;
5192 struct r5conf
*conf
= mddev
->private;
5194 struct blk_plug plug
;
5196 pr_debug("+++ raid5d active\n");
5198 md_check_recovery(mddev
);
5200 blk_start_plug(&plug
);
5202 spin_lock_irq(&conf
->device_lock
);
5205 int batch_size
, released
;
5207 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
5210 !list_empty(&conf
->bitmap_list
)) {
5211 /* Now is a good time to flush some bitmap updates */
5213 spin_unlock_irq(&conf
->device_lock
);
5214 bitmap_unplug(mddev
->bitmap
);
5215 spin_lock_irq(&conf
->device_lock
);
5216 conf
->seq_write
= conf
->seq_flush
;
5217 activate_bit_delay(conf
, conf
->temp_inactive_list
);
5219 raid5_activate_delayed(conf
);
5221 while ((bio
= remove_bio_from_retry(conf
))) {
5223 spin_unlock_irq(&conf
->device_lock
);
5224 ok
= retry_aligned_read(conf
, bio
);
5225 spin_lock_irq(&conf
->device_lock
);
5231 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
5232 conf
->temp_inactive_list
);
5233 if (!batch_size
&& !released
)
5235 handled
+= batch_size
;
5237 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
5238 spin_unlock_irq(&conf
->device_lock
);
5239 md_check_recovery(mddev
);
5240 spin_lock_irq(&conf
->device_lock
);
5243 pr_debug("%d stripes handled\n", handled
);
5245 spin_unlock_irq(&conf
->device_lock
);
5247 async_tx_issue_pending_all();
5248 blk_finish_plug(&plug
);
5250 pr_debug("--- raid5d inactive\n");
5254 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
5256 struct r5conf
*conf
= mddev
->private;
5258 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
5264 raid5_set_cache_size(struct mddev
*mddev
, int size
)
5266 struct r5conf
*conf
= mddev
->private;
5270 if (size
<= 16 || size
> 32768)
5272 hash
= (conf
->max_nr_stripes
- 1) % NR_STRIPE_HASH_LOCKS
;
5273 while (size
< conf
->max_nr_stripes
) {
5274 if (drop_one_stripe(conf
, hash
))
5275 conf
->max_nr_stripes
--;
5280 hash
= NR_STRIPE_HASH_LOCKS
- 1;
5282 err
= md_allow_write(mddev
);
5285 hash
= conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
5286 while (size
> conf
->max_nr_stripes
) {
5287 if (grow_one_stripe(conf
, hash
))
5288 conf
->max_nr_stripes
++;
5290 hash
= (hash
+ 1) % NR_STRIPE_HASH_LOCKS
;
5294 EXPORT_SYMBOL(raid5_set_cache_size
);
5297 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
5299 struct r5conf
*conf
= mddev
->private;
5303 if (len
>= PAGE_SIZE
)
5308 if (kstrtoul(page
, 10, &new))
5310 err
= raid5_set_cache_size(mddev
, new);
5316 static struct md_sysfs_entry
5317 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
5318 raid5_show_stripe_cache_size
,
5319 raid5_store_stripe_cache_size
);
5322 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
5324 struct r5conf
*conf
= mddev
->private;
5326 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
5332 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
5334 struct r5conf
*conf
= mddev
->private;
5336 if (len
>= PAGE_SIZE
)
5341 if (kstrtoul(page
, 10, &new))
5343 if (new > conf
->max_nr_stripes
)
5345 conf
->bypass_threshold
= new;
5349 static struct md_sysfs_entry
5350 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
5352 raid5_show_preread_threshold
,
5353 raid5_store_preread_threshold
);
5356 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
5358 struct r5conf
*conf
= mddev
->private;
5360 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
5365 static struct md_sysfs_entry
5366 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
5369 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
5371 struct r5conf
*conf
= mddev
->private;
5373 return sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
5378 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
5380 int *worker_cnt_per_group
,
5381 struct r5worker_group
**worker_groups
);
5383 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
5385 struct r5conf
*conf
= mddev
->private;
5388 struct r5worker_group
*new_groups
, *old_groups
;
5389 int group_cnt
, worker_cnt_per_group
;
5391 if (len
>= PAGE_SIZE
)
5396 if (kstrtoul(page
, 10, &new))
5399 if (new == conf
->worker_cnt_per_group
)
5402 mddev_suspend(mddev
);
5404 old_groups
= conf
->worker_groups
;
5406 flush_workqueue(raid5_wq
);
5408 err
= alloc_thread_groups(conf
, new,
5409 &group_cnt
, &worker_cnt_per_group
,
5412 spin_lock_irq(&conf
->device_lock
);
5413 conf
->group_cnt
= group_cnt
;
5414 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
5415 conf
->worker_groups
= new_groups
;
5416 spin_unlock_irq(&conf
->device_lock
);
5419 kfree(old_groups
[0].workers
);
5423 mddev_resume(mddev
);
5430 static struct md_sysfs_entry
5431 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
5432 raid5_show_group_thread_cnt
,
5433 raid5_store_group_thread_cnt
);
5435 static struct attribute
*raid5_attrs
[] = {
5436 &raid5_stripecache_size
.attr
,
5437 &raid5_stripecache_active
.attr
,
5438 &raid5_preread_bypass_threshold
.attr
,
5439 &raid5_group_thread_cnt
.attr
,
5442 static struct attribute_group raid5_attrs_group
= {
5444 .attrs
= raid5_attrs
,
5447 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
5449 int *worker_cnt_per_group
,
5450 struct r5worker_group
**worker_groups
)
5454 struct r5worker
*workers
;
5456 *worker_cnt_per_group
= cnt
;
5459 *worker_groups
= NULL
;
5462 *group_cnt
= num_possible_nodes();
5463 size
= sizeof(struct r5worker
) * cnt
;
5464 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
5465 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
5466 *group_cnt
, GFP_NOIO
);
5467 if (!*worker_groups
|| !workers
) {
5469 kfree(*worker_groups
);
5473 for (i
= 0; i
< *group_cnt
; i
++) {
5474 struct r5worker_group
*group
;
5476 group
= worker_groups
[i
];
5477 INIT_LIST_HEAD(&group
->handle_list
);
5479 group
->workers
= workers
+ i
* cnt
;
5481 for (j
= 0; j
< cnt
; j
++) {
5482 struct r5worker
*worker
= group
->workers
+ j
;
5483 worker
->group
= group
;
5484 INIT_WORK(&worker
->work
, raid5_do_work
);
5486 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
5487 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
5494 static void free_thread_groups(struct r5conf
*conf
)
5496 if (conf
->worker_groups
)
5497 kfree(conf
->worker_groups
[0].workers
);
5498 kfree(conf
->worker_groups
);
5499 conf
->worker_groups
= NULL
;
5503 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
5505 struct r5conf
*conf
= mddev
->private;
5508 sectors
= mddev
->dev_sectors
;
5510 /* size is defined by the smallest of previous and new size */
5511 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
5513 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5514 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
5515 return sectors
* (raid_disks
- conf
->max_degraded
);
5518 static void raid5_free_percpu(struct r5conf
*conf
)
5520 struct raid5_percpu
*percpu
;
5527 for_each_possible_cpu(cpu
) {
5528 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
5529 safe_put_page(percpu
->spare_page
);
5530 kfree(percpu
->scribble
);
5532 #ifdef CONFIG_HOTPLUG_CPU
5533 unregister_cpu_notifier(&conf
->cpu_notify
);
5537 free_percpu(conf
->percpu
);
5540 static void free_conf(struct r5conf
*conf
)
5542 free_thread_groups(conf
);
5543 shrink_stripes(conf
);
5544 raid5_free_percpu(conf
);
5546 kfree(conf
->stripe_hashtbl
);
5550 #ifdef CONFIG_HOTPLUG_CPU
5551 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
5554 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
5555 long cpu
= (long)hcpu
;
5556 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
5559 case CPU_UP_PREPARE
:
5560 case CPU_UP_PREPARE_FROZEN
:
5561 if (conf
->level
== 6 && !percpu
->spare_page
)
5562 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
5563 if (!percpu
->scribble
)
5564 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5566 if (!percpu
->scribble
||
5567 (conf
->level
== 6 && !percpu
->spare_page
)) {
5568 safe_put_page(percpu
->spare_page
);
5569 kfree(percpu
->scribble
);
5570 pr_err("%s: failed memory allocation for cpu%ld\n",
5572 return notifier_from_errno(-ENOMEM
);
5576 case CPU_DEAD_FROZEN
:
5577 safe_put_page(percpu
->spare_page
);
5578 kfree(percpu
->scribble
);
5579 percpu
->spare_page
= NULL
;
5580 percpu
->scribble
= NULL
;
5589 static int raid5_alloc_percpu(struct r5conf
*conf
)
5592 struct page
*spare_page
;
5593 struct raid5_percpu __percpu
*allcpus
;
5597 allcpus
= alloc_percpu(struct raid5_percpu
);
5600 conf
->percpu
= allcpus
;
5604 for_each_present_cpu(cpu
) {
5605 if (conf
->level
== 6) {
5606 spare_page
= alloc_page(GFP_KERNEL
);
5611 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
5613 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5618 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
5620 #ifdef CONFIG_HOTPLUG_CPU
5621 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
5622 conf
->cpu_notify
.priority
= 0;
5624 err
= register_cpu_notifier(&conf
->cpu_notify
);
5631 static struct r5conf
*setup_conf(struct mddev
*mddev
)
5633 struct r5conf
*conf
;
5634 int raid_disk
, memory
, max_disks
;
5635 struct md_rdev
*rdev
;
5636 struct disk_info
*disk
;
5639 int group_cnt
, worker_cnt_per_group
;
5640 struct r5worker_group
*new_group
;
5642 if (mddev
->new_level
!= 5
5643 && mddev
->new_level
!= 4
5644 && mddev
->new_level
!= 6) {
5645 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5646 mdname(mddev
), mddev
->new_level
);
5647 return ERR_PTR(-EIO
);
5649 if ((mddev
->new_level
== 5
5650 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
5651 (mddev
->new_level
== 6
5652 && !algorithm_valid_raid6(mddev
->new_layout
))) {
5653 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
5654 mdname(mddev
), mddev
->new_layout
);
5655 return ERR_PTR(-EIO
);
5657 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
5658 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5659 mdname(mddev
), mddev
->raid_disks
);
5660 return ERR_PTR(-EINVAL
);
5663 if (!mddev
->new_chunk_sectors
||
5664 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
5665 !is_power_of_2(mddev
->new_chunk_sectors
)) {
5666 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
5667 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
5668 return ERR_PTR(-EINVAL
);
5671 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
5674 /* Don't enable multi-threading by default*/
5675 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
5677 conf
->group_cnt
= group_cnt
;
5678 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
5679 conf
->worker_groups
= new_group
;
5682 spin_lock_init(&conf
->device_lock
);
5683 seqcount_init(&conf
->gen_lock
);
5684 init_waitqueue_head(&conf
->wait_for_stripe
);
5685 init_waitqueue_head(&conf
->wait_for_overlap
);
5686 INIT_LIST_HEAD(&conf
->handle_list
);
5687 INIT_LIST_HEAD(&conf
->hold_list
);
5688 INIT_LIST_HEAD(&conf
->delayed_list
);
5689 INIT_LIST_HEAD(&conf
->bitmap_list
);
5690 init_llist_head(&conf
->released_stripes
);
5691 atomic_set(&conf
->active_stripes
, 0);
5692 atomic_set(&conf
->preread_active_stripes
, 0);
5693 atomic_set(&conf
->active_aligned_reads
, 0);
5694 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
5695 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
5697 conf
->raid_disks
= mddev
->raid_disks
;
5698 if (mddev
->reshape_position
== MaxSector
)
5699 conf
->previous_raid_disks
= mddev
->raid_disks
;
5701 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5702 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
5703 conf
->scribble_len
= scribble_len(max_disks
);
5705 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
5710 conf
->mddev
= mddev
;
5712 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
5715 /* We init hash_locks[0] separately to that it can be used
5716 * as the reference lock in the spin_lock_nest_lock() call
5717 * in lock_all_device_hash_locks_irq in order to convince
5718 * lockdep that we know what we are doing.
5720 spin_lock_init(conf
->hash_locks
);
5721 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5722 spin_lock_init(conf
->hash_locks
+ i
);
5724 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5725 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
5727 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5728 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
5730 conf
->level
= mddev
->new_level
;
5731 if (raid5_alloc_percpu(conf
) != 0)
5734 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
5736 rdev_for_each(rdev
, mddev
) {
5737 raid_disk
= rdev
->raid_disk
;
5738 if (raid_disk
>= max_disks
5741 disk
= conf
->disks
+ raid_disk
;
5743 if (test_bit(Replacement
, &rdev
->flags
)) {
5744 if (disk
->replacement
)
5746 disk
->replacement
= rdev
;
5753 if (test_bit(In_sync
, &rdev
->flags
)) {
5754 char b
[BDEVNAME_SIZE
];
5755 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
5757 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
5758 } else if (rdev
->saved_raid_disk
!= raid_disk
)
5759 /* Cannot rely on bitmap to complete recovery */
5763 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5764 conf
->level
= mddev
->new_level
;
5765 if (conf
->level
== 6)
5766 conf
->max_degraded
= 2;
5768 conf
->max_degraded
= 1;
5769 conf
->algorithm
= mddev
->new_layout
;
5770 conf
->reshape_progress
= mddev
->reshape_position
;
5771 if (conf
->reshape_progress
!= MaxSector
) {
5772 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
5773 conf
->prev_algo
= mddev
->layout
;
5776 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
5777 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
5778 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
5779 if (grow_stripes(conf
, NR_STRIPES
)) {
5781 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5782 mdname(mddev
), memory
);
5785 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
5786 mdname(mddev
), memory
);
5788 sprintf(pers_name
, "raid%d", mddev
->new_level
);
5789 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
5790 if (!conf
->thread
) {
5792 "md/raid:%s: couldn't allocate thread.\n",
5802 return ERR_PTR(-EIO
);
5804 return ERR_PTR(-ENOMEM
);
5808 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
5811 case ALGORITHM_PARITY_0
:
5812 if (raid_disk
< max_degraded
)
5815 case ALGORITHM_PARITY_N
:
5816 if (raid_disk
>= raid_disks
- max_degraded
)
5819 case ALGORITHM_PARITY_0_6
:
5820 if (raid_disk
== 0 ||
5821 raid_disk
== raid_disks
- 1)
5824 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5825 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5826 case ALGORITHM_LEFT_SYMMETRIC_6
:
5827 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5828 if (raid_disk
== raid_disks
- 1)
5834 static int run(struct mddev
*mddev
)
5836 struct r5conf
*conf
;
5837 int working_disks
= 0;
5838 int dirty_parity_disks
= 0;
5839 struct md_rdev
*rdev
;
5840 sector_t reshape_offset
= 0;
5842 long long min_offset_diff
= 0;
5845 if (mddev
->recovery_cp
!= MaxSector
)
5846 printk(KERN_NOTICE
"md/raid:%s: not clean"
5847 " -- starting background reconstruction\n",
5850 rdev_for_each(rdev
, mddev
) {
5852 if (rdev
->raid_disk
< 0)
5854 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
5856 min_offset_diff
= diff
;
5858 } else if (mddev
->reshape_backwards
&&
5859 diff
< min_offset_diff
)
5860 min_offset_diff
= diff
;
5861 else if (!mddev
->reshape_backwards
&&
5862 diff
> min_offset_diff
)
5863 min_offset_diff
= diff
;
5866 if (mddev
->reshape_position
!= MaxSector
) {
5867 /* Check that we can continue the reshape.
5868 * Difficulties arise if the stripe we would write to
5869 * next is at or after the stripe we would read from next.
5870 * For a reshape that changes the number of devices, this
5871 * is only possible for a very short time, and mdadm makes
5872 * sure that time appears to have past before assembling
5873 * the array. So we fail if that time hasn't passed.
5874 * For a reshape that keeps the number of devices the same
5875 * mdadm must be monitoring the reshape can keeping the
5876 * critical areas read-only and backed up. It will start
5877 * the array in read-only mode, so we check for that.
5879 sector_t here_new
, here_old
;
5881 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5883 if (mddev
->new_level
!= mddev
->level
) {
5884 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5885 "required - aborting.\n",
5889 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5890 /* reshape_position must be on a new-stripe boundary, and one
5891 * further up in new geometry must map after here in old
5894 here_new
= mddev
->reshape_position
;
5895 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5896 (mddev
->raid_disks
- max_degraded
))) {
5897 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5898 "on a stripe boundary\n", mdname(mddev
));
5901 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5902 /* here_new is the stripe we will write to */
5903 here_old
= mddev
->reshape_position
;
5904 sector_div(here_old
, mddev
->chunk_sectors
*
5905 (old_disks
-max_degraded
));
5906 /* here_old is the first stripe that we might need to read
5908 if (mddev
->delta_disks
== 0) {
5909 if ((here_new
* mddev
->new_chunk_sectors
!=
5910 here_old
* mddev
->chunk_sectors
)) {
5911 printk(KERN_ERR
"md/raid:%s: reshape position is"
5912 " confused - aborting\n", mdname(mddev
));
5915 /* We cannot be sure it is safe to start an in-place
5916 * reshape. It is only safe if user-space is monitoring
5917 * and taking constant backups.
5918 * mdadm always starts a situation like this in
5919 * readonly mode so it can take control before
5920 * allowing any writes. So just check for that.
5922 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
5923 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
5924 /* not really in-place - so OK */;
5925 else if (mddev
->ro
== 0) {
5926 printk(KERN_ERR
"md/raid:%s: in-place reshape "
5927 "must be started in read-only mode "
5932 } else if (mddev
->reshape_backwards
5933 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
5934 here_old
* mddev
->chunk_sectors
)
5935 : (here_new
* mddev
->new_chunk_sectors
>=
5936 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
5937 /* Reading from the same stripe as writing to - bad */
5938 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5939 "auto-recovery - aborting.\n",
5943 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5945 /* OK, we should be able to continue; */
5947 BUG_ON(mddev
->level
!= mddev
->new_level
);
5948 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5949 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5950 BUG_ON(mddev
->delta_disks
!= 0);
5953 if (mddev
->private == NULL
)
5954 conf
= setup_conf(mddev
);
5956 conf
= mddev
->private;
5959 return PTR_ERR(conf
);
5961 conf
->min_offset_diff
= min_offset_diff
;
5962 mddev
->thread
= conf
->thread
;
5963 conf
->thread
= NULL
;
5964 mddev
->private = conf
;
5966 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5968 rdev
= conf
->disks
[i
].rdev
;
5969 if (!rdev
&& conf
->disks
[i
].replacement
) {
5970 /* The replacement is all we have yet */
5971 rdev
= conf
->disks
[i
].replacement
;
5972 conf
->disks
[i
].replacement
= NULL
;
5973 clear_bit(Replacement
, &rdev
->flags
);
5974 conf
->disks
[i
].rdev
= rdev
;
5978 if (conf
->disks
[i
].replacement
&&
5979 conf
->reshape_progress
!= MaxSector
) {
5980 /* replacements and reshape simply do not mix. */
5981 printk(KERN_ERR
"md: cannot handle concurrent "
5982 "replacement and reshape.\n");
5985 if (test_bit(In_sync
, &rdev
->flags
)) {
5989 /* This disc is not fully in-sync. However if it
5990 * just stored parity (beyond the recovery_offset),
5991 * when we don't need to be concerned about the
5992 * array being dirty.
5993 * When reshape goes 'backwards', we never have
5994 * partially completed devices, so we only need
5995 * to worry about reshape going forwards.
5997 /* Hack because v0.91 doesn't store recovery_offset properly. */
5998 if (mddev
->major_version
== 0 &&
5999 mddev
->minor_version
> 90)
6000 rdev
->recovery_offset
= reshape_offset
;
6002 if (rdev
->recovery_offset
< reshape_offset
) {
6003 /* We need to check old and new layout */
6004 if (!only_parity(rdev
->raid_disk
,
6007 conf
->max_degraded
))
6010 if (!only_parity(rdev
->raid_disk
,
6012 conf
->previous_raid_disks
,
6013 conf
->max_degraded
))
6015 dirty_parity_disks
++;
6019 * 0 for a fully functional array, 1 or 2 for a degraded array.
6021 mddev
->degraded
= calc_degraded(conf
);
6023 if (has_failed(conf
)) {
6024 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
6025 " (%d/%d failed)\n",
6026 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
6030 /* device size must be a multiple of chunk size */
6031 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
6032 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
6034 if (mddev
->degraded
> dirty_parity_disks
&&
6035 mddev
->recovery_cp
!= MaxSector
) {
6036 if (mddev
->ok_start_degraded
)
6038 "md/raid:%s: starting dirty degraded array"
6039 " - data corruption possible.\n",
6043 "md/raid:%s: cannot start dirty degraded array.\n",
6049 if (mddev
->degraded
== 0)
6050 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
6051 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
6052 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
6055 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
6056 " out of %d devices, algorithm %d\n",
6057 mdname(mddev
), conf
->level
,
6058 mddev
->raid_disks
- mddev
->degraded
,
6059 mddev
->raid_disks
, mddev
->new_layout
);
6061 print_raid5_conf(conf
);
6063 if (conf
->reshape_progress
!= MaxSector
) {
6064 conf
->reshape_safe
= conf
->reshape_progress
;
6065 atomic_set(&conf
->reshape_stripes
, 0);
6066 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6067 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6068 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6069 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6070 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6075 /* Ok, everything is just fine now */
6076 if (mddev
->to_remove
== &raid5_attrs_group
)
6077 mddev
->to_remove
= NULL
;
6078 else if (mddev
->kobj
.sd
&&
6079 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
6081 "raid5: failed to create sysfs attributes for %s\n",
6083 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6087 bool discard_supported
= true;
6088 /* read-ahead size must cover two whole stripes, which
6089 * is 2 * (datadisks) * chunksize where 'n' is the
6090 * number of raid devices
6092 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
6093 int stripe
= data_disks
*
6094 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
6095 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6096 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6098 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
6100 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
6101 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
6103 chunk_size
= mddev
->chunk_sectors
<< 9;
6104 blk_queue_io_min(mddev
->queue
, chunk_size
);
6105 blk_queue_io_opt(mddev
->queue
, chunk_size
*
6106 (conf
->raid_disks
- conf
->max_degraded
));
6108 * We can only discard a whole stripe. It doesn't make sense to
6109 * discard data disk but write parity disk
6111 stripe
= stripe
* PAGE_SIZE
;
6112 /* Round up to power of 2, as discard handling
6113 * currently assumes that */
6114 while ((stripe
-1) & stripe
)
6115 stripe
= (stripe
| (stripe
-1)) + 1;
6116 mddev
->queue
->limits
.discard_alignment
= stripe
;
6117 mddev
->queue
->limits
.discard_granularity
= stripe
;
6119 * unaligned part of discard request will be ignored, so can't
6120 * guarantee discard_zerors_data
6122 mddev
->queue
->limits
.discard_zeroes_data
= 0;
6124 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
6126 rdev_for_each(rdev
, mddev
) {
6127 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6128 rdev
->data_offset
<< 9);
6129 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
6130 rdev
->new_data_offset
<< 9);
6132 * discard_zeroes_data is required, otherwise data
6133 * could be lost. Consider a scenario: discard a stripe
6134 * (the stripe could be inconsistent if
6135 * discard_zeroes_data is 0); write one disk of the
6136 * stripe (the stripe could be inconsistent again
6137 * depending on which disks are used to calculate
6138 * parity); the disk is broken; The stripe data of this
6141 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
6142 !bdev_get_queue(rdev
->bdev
)->
6143 limits
.discard_zeroes_data
)
6144 discard_supported
= false;
6147 if (discard_supported
&&
6148 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
6149 mddev
->queue
->limits
.discard_granularity
>= stripe
)
6150 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
6153 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
6159 md_unregister_thread(&mddev
->thread
);
6160 print_raid5_conf(conf
);
6162 mddev
->private = NULL
;
6163 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
6167 static int stop(struct mddev
*mddev
)
6169 struct r5conf
*conf
= mddev
->private;
6171 md_unregister_thread(&mddev
->thread
);
6173 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
6175 mddev
->private = NULL
;
6176 mddev
->to_remove
= &raid5_attrs_group
;
6180 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
6182 struct r5conf
*conf
= mddev
->private;
6185 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
6186 mddev
->chunk_sectors
/ 2, mddev
->layout
);
6187 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
6188 for (i
= 0; i
< conf
->raid_disks
; i
++)
6189 seq_printf (seq
, "%s",
6190 conf
->disks
[i
].rdev
&&
6191 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
6192 seq_printf (seq
, "]");
6195 static void print_raid5_conf (struct r5conf
*conf
)
6198 struct disk_info
*tmp
;
6200 printk(KERN_DEBUG
"RAID conf printout:\n");
6202 printk("(conf==NULL)\n");
6205 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
6207 conf
->raid_disks
- conf
->mddev
->degraded
);
6209 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6210 char b
[BDEVNAME_SIZE
];
6211 tmp
= conf
->disks
+ i
;
6213 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
6214 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
6215 bdevname(tmp
->rdev
->bdev
, b
));
6219 static int raid5_spare_active(struct mddev
*mddev
)
6222 struct r5conf
*conf
= mddev
->private;
6223 struct disk_info
*tmp
;
6225 unsigned long flags
;
6227 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6228 tmp
= conf
->disks
+ i
;
6229 if (tmp
->replacement
6230 && tmp
->replacement
->recovery_offset
== MaxSector
6231 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
6232 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
6233 /* Replacement has just become active. */
6235 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
6238 /* Replaced device not technically faulty,
6239 * but we need to be sure it gets removed
6240 * and never re-added.
6242 set_bit(Faulty
, &tmp
->rdev
->flags
);
6243 sysfs_notify_dirent_safe(
6244 tmp
->rdev
->sysfs_state
);
6246 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
6247 } else if (tmp
->rdev
6248 && tmp
->rdev
->recovery_offset
== MaxSector
6249 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
6250 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
6252 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
6255 spin_lock_irqsave(&conf
->device_lock
, flags
);
6256 mddev
->degraded
= calc_degraded(conf
);
6257 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
6258 print_raid5_conf(conf
);
6262 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
6264 struct r5conf
*conf
= mddev
->private;
6266 int number
= rdev
->raid_disk
;
6267 struct md_rdev
**rdevp
;
6268 struct disk_info
*p
= conf
->disks
+ number
;
6270 print_raid5_conf(conf
);
6271 if (rdev
== p
->rdev
)
6273 else if (rdev
== p
->replacement
)
6274 rdevp
= &p
->replacement
;
6278 if (number
>= conf
->raid_disks
&&
6279 conf
->reshape_progress
== MaxSector
)
6280 clear_bit(In_sync
, &rdev
->flags
);
6282 if (test_bit(In_sync
, &rdev
->flags
) ||
6283 atomic_read(&rdev
->nr_pending
)) {
6287 /* Only remove non-faulty devices if recovery
6290 if (!test_bit(Faulty
, &rdev
->flags
) &&
6291 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
6292 !has_failed(conf
) &&
6293 (!p
->replacement
|| p
->replacement
== rdev
) &&
6294 number
< conf
->raid_disks
) {
6300 if (atomic_read(&rdev
->nr_pending
)) {
6301 /* lost the race, try later */
6304 } else if (p
->replacement
) {
6305 /* We must have just cleared 'rdev' */
6306 p
->rdev
= p
->replacement
;
6307 clear_bit(Replacement
, &p
->replacement
->flags
);
6308 smp_mb(); /* Make sure other CPUs may see both as identical
6309 * but will never see neither - if they are careful
6311 p
->replacement
= NULL
;
6312 clear_bit(WantReplacement
, &rdev
->flags
);
6314 /* We might have just removed the Replacement as faulty-
6315 * clear the bit just in case
6317 clear_bit(WantReplacement
, &rdev
->flags
);
6320 print_raid5_conf(conf
);
6324 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
6326 struct r5conf
*conf
= mddev
->private;
6329 struct disk_info
*p
;
6331 int last
= conf
->raid_disks
- 1;
6333 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
6336 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
6337 /* no point adding a device */
6340 if (rdev
->raid_disk
>= 0)
6341 first
= last
= rdev
->raid_disk
;
6344 * find the disk ... but prefer rdev->saved_raid_disk
6347 if (rdev
->saved_raid_disk
>= 0 &&
6348 rdev
->saved_raid_disk
>= first
&&
6349 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
6350 first
= rdev
->saved_raid_disk
;
6352 for (disk
= first
; disk
<= last
; disk
++) {
6353 p
= conf
->disks
+ disk
;
6354 if (p
->rdev
== NULL
) {
6355 clear_bit(In_sync
, &rdev
->flags
);
6356 rdev
->raid_disk
= disk
;
6358 if (rdev
->saved_raid_disk
!= disk
)
6360 rcu_assign_pointer(p
->rdev
, rdev
);
6364 for (disk
= first
; disk
<= last
; disk
++) {
6365 p
= conf
->disks
+ disk
;
6366 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
6367 p
->replacement
== NULL
) {
6368 clear_bit(In_sync
, &rdev
->flags
);
6369 set_bit(Replacement
, &rdev
->flags
);
6370 rdev
->raid_disk
= disk
;
6373 rcu_assign_pointer(p
->replacement
, rdev
);
6378 print_raid5_conf(conf
);
6382 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
6384 /* no resync is happening, and there is enough space
6385 * on all devices, so we can resize.
6386 * We need to make sure resync covers any new space.
6387 * If the array is shrinking we should possibly wait until
6388 * any io in the removed space completes, but it hardly seems
6392 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
6393 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
6394 if (mddev
->external_size
&&
6395 mddev
->array_sectors
> newsize
)
6397 if (mddev
->bitmap
) {
6398 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
6402 md_set_array_sectors(mddev
, newsize
);
6403 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
6404 revalidate_disk(mddev
->gendisk
);
6405 if (sectors
> mddev
->dev_sectors
&&
6406 mddev
->recovery_cp
> mddev
->dev_sectors
) {
6407 mddev
->recovery_cp
= mddev
->dev_sectors
;
6408 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
6410 mddev
->dev_sectors
= sectors
;
6411 mddev
->resync_max_sectors
= sectors
;
6415 static int check_stripe_cache(struct mddev
*mddev
)
6417 /* Can only proceed if there are plenty of stripe_heads.
6418 * We need a minimum of one full stripe,, and for sensible progress
6419 * it is best to have about 4 times that.
6420 * If we require 4 times, then the default 256 4K stripe_heads will
6421 * allow for chunk sizes up to 256K, which is probably OK.
6422 * If the chunk size is greater, user-space should request more
6423 * stripe_heads first.
6425 struct r5conf
*conf
= mddev
->private;
6426 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
6427 > conf
->max_nr_stripes
||
6428 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
6429 > conf
->max_nr_stripes
) {
6430 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
6432 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
6439 static int check_reshape(struct mddev
*mddev
)
6441 struct r5conf
*conf
= mddev
->private;
6443 if (mddev
->delta_disks
== 0 &&
6444 mddev
->new_layout
== mddev
->layout
&&
6445 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
6446 return 0; /* nothing to do */
6447 if (has_failed(conf
))
6449 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
6450 /* We might be able to shrink, but the devices must
6451 * be made bigger first.
6452 * For raid6, 4 is the minimum size.
6453 * Otherwise 2 is the minimum
6456 if (mddev
->level
== 6)
6458 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
6462 if (!check_stripe_cache(mddev
))
6465 return resize_stripes(conf
, (conf
->previous_raid_disks
6466 + mddev
->delta_disks
));
6469 static int raid5_start_reshape(struct mddev
*mddev
)
6471 struct r5conf
*conf
= mddev
->private;
6472 struct md_rdev
*rdev
;
6474 unsigned long flags
;
6476 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
6479 if (!check_stripe_cache(mddev
))
6482 if (has_failed(conf
))
6485 rdev_for_each(rdev
, mddev
) {
6486 if (!test_bit(In_sync
, &rdev
->flags
)
6487 && !test_bit(Faulty
, &rdev
->flags
))
6491 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
6492 /* Not enough devices even to make a degraded array
6497 /* Refuse to reduce size of the array. Any reductions in
6498 * array size must be through explicit setting of array_size
6501 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
6502 < mddev
->array_sectors
) {
6503 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
6504 "before number of disks\n", mdname(mddev
));
6508 atomic_set(&conf
->reshape_stripes
, 0);
6509 spin_lock_irq(&conf
->device_lock
);
6510 write_seqcount_begin(&conf
->gen_lock
);
6511 conf
->previous_raid_disks
= conf
->raid_disks
;
6512 conf
->raid_disks
+= mddev
->delta_disks
;
6513 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
6514 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6515 conf
->prev_algo
= conf
->algorithm
;
6516 conf
->algorithm
= mddev
->new_layout
;
6518 /* Code that selects data_offset needs to see the generation update
6519 * if reshape_progress has been set - so a memory barrier needed.
6522 if (mddev
->reshape_backwards
)
6523 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
6525 conf
->reshape_progress
= 0;
6526 conf
->reshape_safe
= conf
->reshape_progress
;
6527 write_seqcount_end(&conf
->gen_lock
);
6528 spin_unlock_irq(&conf
->device_lock
);
6530 /* Now make sure any requests that proceeded on the assumption
6531 * the reshape wasn't running - like Discard or Read - have
6534 mddev_suspend(mddev
);
6535 mddev_resume(mddev
);
6537 /* Add some new drives, as many as will fit.
6538 * We know there are enough to make the newly sized array work.
6539 * Don't add devices if we are reducing the number of
6540 * devices in the array. This is because it is not possible
6541 * to correctly record the "partially reconstructed" state of
6542 * such devices during the reshape and confusion could result.
6544 if (mddev
->delta_disks
>= 0) {
6545 rdev_for_each(rdev
, mddev
)
6546 if (rdev
->raid_disk
< 0 &&
6547 !test_bit(Faulty
, &rdev
->flags
)) {
6548 if (raid5_add_disk(mddev
, rdev
) == 0) {
6550 >= conf
->previous_raid_disks
)
6551 set_bit(In_sync
, &rdev
->flags
);
6553 rdev
->recovery_offset
= 0;
6555 if (sysfs_link_rdev(mddev
, rdev
))
6556 /* Failure here is OK */;
6558 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
6559 && !test_bit(Faulty
, &rdev
->flags
)) {
6560 /* This is a spare that was manually added */
6561 set_bit(In_sync
, &rdev
->flags
);
6564 /* When a reshape changes the number of devices,
6565 * ->degraded is measured against the larger of the
6566 * pre and post number of devices.
6568 spin_lock_irqsave(&conf
->device_lock
, flags
);
6569 mddev
->degraded
= calc_degraded(conf
);
6570 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
6572 mddev
->raid_disks
= conf
->raid_disks
;
6573 mddev
->reshape_position
= conf
->reshape_progress
;
6574 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6576 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6577 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6578 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6579 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6580 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6582 if (!mddev
->sync_thread
) {
6583 mddev
->recovery
= 0;
6584 spin_lock_irq(&conf
->device_lock
);
6585 write_seqcount_begin(&conf
->gen_lock
);
6586 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
6587 mddev
->new_chunk_sectors
=
6588 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
6589 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
6590 rdev_for_each(rdev
, mddev
)
6591 rdev
->new_data_offset
= rdev
->data_offset
;
6593 conf
->generation
--;
6594 conf
->reshape_progress
= MaxSector
;
6595 mddev
->reshape_position
= MaxSector
;
6596 write_seqcount_end(&conf
->gen_lock
);
6597 spin_unlock_irq(&conf
->device_lock
);
6600 conf
->reshape_checkpoint
= jiffies
;
6601 md_wakeup_thread(mddev
->sync_thread
);
6602 md_new_event(mddev
);
6606 /* This is called from the reshape thread and should make any
6607 * changes needed in 'conf'
6609 static void end_reshape(struct r5conf
*conf
)
6612 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
6613 struct md_rdev
*rdev
;
6615 spin_lock_irq(&conf
->device_lock
);
6616 conf
->previous_raid_disks
= conf
->raid_disks
;
6617 rdev_for_each(rdev
, conf
->mddev
)
6618 rdev
->data_offset
= rdev
->new_data_offset
;
6620 conf
->reshape_progress
= MaxSector
;
6621 spin_unlock_irq(&conf
->device_lock
);
6622 wake_up(&conf
->wait_for_overlap
);
6624 /* read-ahead size must cover two whole stripes, which is
6625 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6627 if (conf
->mddev
->queue
) {
6628 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
6629 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
6631 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6632 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6637 /* This is called from the raid5d thread with mddev_lock held.
6638 * It makes config changes to the device.
6640 static void raid5_finish_reshape(struct mddev
*mddev
)
6642 struct r5conf
*conf
= mddev
->private;
6644 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
6646 if (mddev
->delta_disks
> 0) {
6647 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6648 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
6649 revalidate_disk(mddev
->gendisk
);
6652 spin_lock_irq(&conf
->device_lock
);
6653 mddev
->degraded
= calc_degraded(conf
);
6654 spin_unlock_irq(&conf
->device_lock
);
6655 for (d
= conf
->raid_disks
;
6656 d
< conf
->raid_disks
- mddev
->delta_disks
;
6658 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
6660 clear_bit(In_sync
, &rdev
->flags
);
6661 rdev
= conf
->disks
[d
].replacement
;
6663 clear_bit(In_sync
, &rdev
->flags
);
6666 mddev
->layout
= conf
->algorithm
;
6667 mddev
->chunk_sectors
= conf
->chunk_sectors
;
6668 mddev
->reshape_position
= MaxSector
;
6669 mddev
->delta_disks
= 0;
6670 mddev
->reshape_backwards
= 0;
6674 static void raid5_quiesce(struct mddev
*mddev
, int state
)
6676 struct r5conf
*conf
= mddev
->private;
6679 case 2: /* resume for a suspend */
6680 wake_up(&conf
->wait_for_overlap
);
6683 case 1: /* stop all writes */
6684 lock_all_device_hash_locks_irq(conf
);
6685 /* '2' tells resync/reshape to pause so that all
6686 * active stripes can drain
6689 wait_event_cmd(conf
->wait_for_stripe
,
6690 atomic_read(&conf
->active_stripes
) == 0 &&
6691 atomic_read(&conf
->active_aligned_reads
) == 0,
6692 unlock_all_device_hash_locks_irq(conf
),
6693 lock_all_device_hash_locks_irq(conf
));
6695 unlock_all_device_hash_locks_irq(conf
);
6696 /* allow reshape to continue */
6697 wake_up(&conf
->wait_for_overlap
);
6700 case 0: /* re-enable writes */
6701 lock_all_device_hash_locks_irq(conf
);
6703 wake_up(&conf
->wait_for_stripe
);
6704 wake_up(&conf
->wait_for_overlap
);
6705 unlock_all_device_hash_locks_irq(conf
);
6711 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
6713 struct r0conf
*raid0_conf
= mddev
->private;
6716 /* for raid0 takeover only one zone is supported */
6717 if (raid0_conf
->nr_strip_zones
> 1) {
6718 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6720 return ERR_PTR(-EINVAL
);
6723 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
6724 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
6725 mddev
->dev_sectors
= sectors
;
6726 mddev
->new_level
= level
;
6727 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6728 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
6729 mddev
->raid_disks
+= 1;
6730 mddev
->delta_disks
= 1;
6731 /* make sure it will be not marked as dirty */
6732 mddev
->recovery_cp
= MaxSector
;
6734 return setup_conf(mddev
);
6738 static void *raid5_takeover_raid1(struct mddev
*mddev
)
6742 if (mddev
->raid_disks
!= 2 ||
6743 mddev
->degraded
> 1)
6744 return ERR_PTR(-EINVAL
);
6746 /* Should check if there are write-behind devices? */
6748 chunksect
= 64*2; /* 64K by default */
6750 /* The array must be an exact multiple of chunksize */
6751 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
6754 if ((chunksect
<<9) < STRIPE_SIZE
)
6755 /* array size does not allow a suitable chunk size */
6756 return ERR_PTR(-EINVAL
);
6758 mddev
->new_level
= 5;
6759 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6760 mddev
->new_chunk_sectors
= chunksect
;
6762 return setup_conf(mddev
);
6765 static void *raid5_takeover_raid6(struct mddev
*mddev
)
6769 switch (mddev
->layout
) {
6770 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6771 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
6773 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6774 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
6776 case ALGORITHM_LEFT_SYMMETRIC_6
:
6777 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6779 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6780 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
6782 case ALGORITHM_PARITY_0_6
:
6783 new_layout
= ALGORITHM_PARITY_0
;
6785 case ALGORITHM_PARITY_N
:
6786 new_layout
= ALGORITHM_PARITY_N
;
6789 return ERR_PTR(-EINVAL
);
6791 mddev
->new_level
= 5;
6792 mddev
->new_layout
= new_layout
;
6793 mddev
->delta_disks
= -1;
6794 mddev
->raid_disks
-= 1;
6795 return setup_conf(mddev
);
6799 static int raid5_check_reshape(struct mddev
*mddev
)
6801 /* For a 2-drive array, the layout and chunk size can be changed
6802 * immediately as not restriping is needed.
6803 * For larger arrays we record the new value - after validation
6804 * to be used by a reshape pass.
6806 struct r5conf
*conf
= mddev
->private;
6807 int new_chunk
= mddev
->new_chunk_sectors
;
6809 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
6811 if (new_chunk
> 0) {
6812 if (!is_power_of_2(new_chunk
))
6814 if (new_chunk
< (PAGE_SIZE
>>9))
6816 if (mddev
->array_sectors
& (new_chunk
-1))
6817 /* not factor of array size */
6821 /* They look valid */
6823 if (mddev
->raid_disks
== 2) {
6824 /* can make the change immediately */
6825 if (mddev
->new_layout
>= 0) {
6826 conf
->algorithm
= mddev
->new_layout
;
6827 mddev
->layout
= mddev
->new_layout
;
6829 if (new_chunk
> 0) {
6830 conf
->chunk_sectors
= new_chunk
;
6831 mddev
->chunk_sectors
= new_chunk
;
6833 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6834 md_wakeup_thread(mddev
->thread
);
6836 return check_reshape(mddev
);
6839 static int raid6_check_reshape(struct mddev
*mddev
)
6841 int new_chunk
= mddev
->new_chunk_sectors
;
6843 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
6845 if (new_chunk
> 0) {
6846 if (!is_power_of_2(new_chunk
))
6848 if (new_chunk
< (PAGE_SIZE
>> 9))
6850 if (mddev
->array_sectors
& (new_chunk
-1))
6851 /* not factor of array size */
6855 /* They look valid */
6856 return check_reshape(mddev
);
6859 static void *raid5_takeover(struct mddev
*mddev
)
6861 /* raid5 can take over:
6862 * raid0 - if there is only one strip zone - make it a raid4 layout
6863 * raid1 - if there are two drives. We need to know the chunk size
6864 * raid4 - trivial - just use a raid4 layout.
6865 * raid6 - Providing it is a *_6 layout
6867 if (mddev
->level
== 0)
6868 return raid45_takeover_raid0(mddev
, 5);
6869 if (mddev
->level
== 1)
6870 return raid5_takeover_raid1(mddev
);
6871 if (mddev
->level
== 4) {
6872 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6873 mddev
->new_level
= 5;
6874 return setup_conf(mddev
);
6876 if (mddev
->level
== 6)
6877 return raid5_takeover_raid6(mddev
);
6879 return ERR_PTR(-EINVAL
);
6882 static void *raid4_takeover(struct mddev
*mddev
)
6884 /* raid4 can take over:
6885 * raid0 - if there is only one strip zone
6886 * raid5 - if layout is right
6888 if (mddev
->level
== 0)
6889 return raid45_takeover_raid0(mddev
, 4);
6890 if (mddev
->level
== 5 &&
6891 mddev
->layout
== ALGORITHM_PARITY_N
) {
6892 mddev
->new_layout
= 0;
6893 mddev
->new_level
= 4;
6894 return setup_conf(mddev
);
6896 return ERR_PTR(-EINVAL
);
6899 static struct md_personality raid5_personality
;
6901 static void *raid6_takeover(struct mddev
*mddev
)
6903 /* Currently can only take over a raid5. We map the
6904 * personality to an equivalent raid6 personality
6905 * with the Q block at the end.
6909 if (mddev
->pers
!= &raid5_personality
)
6910 return ERR_PTR(-EINVAL
);
6911 if (mddev
->degraded
> 1)
6912 return ERR_PTR(-EINVAL
);
6913 if (mddev
->raid_disks
> 253)
6914 return ERR_PTR(-EINVAL
);
6915 if (mddev
->raid_disks
< 3)
6916 return ERR_PTR(-EINVAL
);
6918 switch (mddev
->layout
) {
6919 case ALGORITHM_LEFT_ASYMMETRIC
:
6920 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
6922 case ALGORITHM_RIGHT_ASYMMETRIC
:
6923 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
6925 case ALGORITHM_LEFT_SYMMETRIC
:
6926 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
6928 case ALGORITHM_RIGHT_SYMMETRIC
:
6929 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
6931 case ALGORITHM_PARITY_0
:
6932 new_layout
= ALGORITHM_PARITY_0_6
;
6934 case ALGORITHM_PARITY_N
:
6935 new_layout
= ALGORITHM_PARITY_N
;
6938 return ERR_PTR(-EINVAL
);
6940 mddev
->new_level
= 6;
6941 mddev
->new_layout
= new_layout
;
6942 mddev
->delta_disks
= 1;
6943 mddev
->raid_disks
+= 1;
6944 return setup_conf(mddev
);
6948 static struct md_personality raid6_personality
=
6952 .owner
= THIS_MODULE
,
6953 .make_request
= make_request
,
6957 .error_handler
= error
,
6958 .hot_add_disk
= raid5_add_disk
,
6959 .hot_remove_disk
= raid5_remove_disk
,
6960 .spare_active
= raid5_spare_active
,
6961 .sync_request
= sync_request
,
6962 .resize
= raid5_resize
,
6964 .check_reshape
= raid6_check_reshape
,
6965 .start_reshape
= raid5_start_reshape
,
6966 .finish_reshape
= raid5_finish_reshape
,
6967 .quiesce
= raid5_quiesce
,
6968 .takeover
= raid6_takeover
,
6970 static struct md_personality raid5_personality
=
6974 .owner
= THIS_MODULE
,
6975 .make_request
= make_request
,
6979 .error_handler
= error
,
6980 .hot_add_disk
= raid5_add_disk
,
6981 .hot_remove_disk
= raid5_remove_disk
,
6982 .spare_active
= raid5_spare_active
,
6983 .sync_request
= sync_request
,
6984 .resize
= raid5_resize
,
6986 .check_reshape
= raid5_check_reshape
,
6987 .start_reshape
= raid5_start_reshape
,
6988 .finish_reshape
= raid5_finish_reshape
,
6989 .quiesce
= raid5_quiesce
,
6990 .takeover
= raid5_takeover
,
6993 static struct md_personality raid4_personality
=
6997 .owner
= THIS_MODULE
,
6998 .make_request
= make_request
,
7002 .error_handler
= error
,
7003 .hot_add_disk
= raid5_add_disk
,
7004 .hot_remove_disk
= raid5_remove_disk
,
7005 .spare_active
= raid5_spare_active
,
7006 .sync_request
= sync_request
,
7007 .resize
= raid5_resize
,
7009 .check_reshape
= raid5_check_reshape
,
7010 .start_reshape
= raid5_start_reshape
,
7011 .finish_reshape
= raid5_finish_reshape
,
7012 .quiesce
= raid5_quiesce
,
7013 .takeover
= raid4_takeover
,
7016 static int __init
raid5_init(void)
7018 raid5_wq
= alloc_workqueue("raid5wq",
7019 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
7022 register_md_personality(&raid6_personality
);
7023 register_md_personality(&raid5_personality
);
7024 register_md_personality(&raid4_personality
);
7028 static void raid5_exit(void)
7030 unregister_md_personality(&raid6_personality
);
7031 unregister_md_personality(&raid5_personality
);
7032 unregister_md_personality(&raid4_personality
);
7033 destroy_workqueue(raid5_wq
);
7036 module_init(raid5_init
);
7037 module_exit(raid5_exit
);
7038 MODULE_LICENSE("GPL");
7039 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7040 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7041 MODULE_ALIAS("md-raid5");
7042 MODULE_ALIAS("md-raid4");
7043 MODULE_ALIAS("md-level-5");
7044 MODULE_ALIAS("md-level-4");
7045 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7046 MODULE_ALIAS("md-raid6");
7047 MODULE_ALIAS("md-level-6");
7049 /* This used to be two separate modules, they were: */
7050 MODULE_ALIAS("raid5");
7051 MODULE_ALIAS("raid6");