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 <trace/events/block.h>
67 #define NR_STRIPES 256
68 #define STRIPE_SIZE PAGE_SIZE
69 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
70 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
71 #define IO_THRESHOLD 1
72 #define BYPASS_THRESHOLD 1
73 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
74 #define HASH_MASK (NR_HASH - 1)
76 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
78 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
79 return &conf
->stripe_hashtbl
[hash
];
82 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
83 * order without overlap. There may be several bio's per stripe+device, and
84 * a bio could span several devices.
85 * When walking this list for a particular stripe+device, we must never proceed
86 * beyond a bio that extends past this device, as the next bio might no longer
88 * This function is used to determine the 'next' bio in the list, given the sector
89 * of the current stripe+device
91 static inline struct bio
*r5_next_bio(struct bio
*bio
, sector_t sector
)
93 int sectors
= bio
->bi_size
>> 9;
94 if (bio
->bi_sector
+ sectors
< sector
+ STRIPE_SECTORS
)
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104 static inline int raid5_bi_processed_stripes(struct bio
*bio
)
106 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
107 return (atomic_read(segments
) >> 16) & 0xffff;
110 static inline int raid5_dec_bi_active_stripes(struct bio
*bio
)
112 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
113 return atomic_sub_return(1, segments
) & 0xffff;
116 static inline void raid5_inc_bi_active_stripes(struct bio
*bio
)
118 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
119 atomic_inc(segments
);
122 static inline void raid5_set_bi_processed_stripes(struct bio
*bio
,
125 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
129 old
= atomic_read(segments
);
130 new = (old
& 0xffff) | (cnt
<< 16);
131 } while (atomic_cmpxchg(segments
, old
, new) != old
);
134 static inline void raid5_set_bi_stripes(struct bio
*bio
, unsigned int cnt
)
136 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
137 atomic_set(segments
, cnt
);
140 /* Find first data disk in a raid6 stripe */
141 static inline int raid6_d0(struct stripe_head
*sh
)
144 /* ddf always start from first device */
146 /* md starts just after Q block */
147 if (sh
->qd_idx
== sh
->disks
- 1)
150 return sh
->qd_idx
+ 1;
152 static inline int raid6_next_disk(int disk
, int raid_disks
)
155 return (disk
< raid_disks
) ? disk
: 0;
158 /* When walking through the disks in a raid5, starting at raid6_d0,
159 * We need to map each disk to a 'slot', where the data disks are slot
160 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
161 * is raid_disks-1. This help does that mapping.
163 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
164 int *count
, int syndrome_disks
)
170 if (idx
== sh
->pd_idx
)
171 return syndrome_disks
;
172 if (idx
== sh
->qd_idx
)
173 return syndrome_disks
+ 1;
179 static void return_io(struct bio
*return_bi
)
181 struct bio
*bi
= return_bi
;
184 return_bi
= bi
->bi_next
;
192 static void print_raid5_conf (struct r5conf
*conf
);
194 static int stripe_operations_active(struct stripe_head
*sh
)
196 return sh
->check_state
|| sh
->reconstruct_state
||
197 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
198 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
201 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
)
203 BUG_ON(!list_empty(&sh
->lru
));
204 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
205 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
206 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
207 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
208 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
209 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
210 sh
->bm_seq
- conf
->seq_write
> 0)
211 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
213 clear_bit(STRIPE_DELAYED
, &sh
->state
);
214 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
215 list_add_tail(&sh
->lru
, &conf
->handle_list
);
217 md_wakeup_thread(conf
->mddev
->thread
);
219 BUG_ON(stripe_operations_active(sh
));
220 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
221 if (atomic_dec_return(&conf
->preread_active_stripes
)
223 md_wakeup_thread(conf
->mddev
->thread
);
224 atomic_dec(&conf
->active_stripes
);
225 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
226 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
227 wake_up(&conf
->wait_for_stripe
);
228 if (conf
->retry_read_aligned
)
229 md_wakeup_thread(conf
->mddev
->thread
);
234 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
)
236 if (atomic_dec_and_test(&sh
->count
))
237 do_release_stripe(conf
, sh
);
240 static void release_stripe(struct stripe_head
*sh
)
242 struct r5conf
*conf
= sh
->raid_conf
;
245 local_irq_save(flags
);
246 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
247 do_release_stripe(conf
, sh
);
248 spin_unlock(&conf
->device_lock
);
250 local_irq_restore(flags
);
253 static inline void remove_hash(struct stripe_head
*sh
)
255 pr_debug("remove_hash(), stripe %llu\n",
256 (unsigned long long)sh
->sector
);
258 hlist_del_init(&sh
->hash
);
261 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
263 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
265 pr_debug("insert_hash(), stripe %llu\n",
266 (unsigned long long)sh
->sector
);
268 hlist_add_head(&sh
->hash
, hp
);
272 /* find an idle stripe, make sure it is unhashed, and return it. */
273 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
)
275 struct stripe_head
*sh
= NULL
;
276 struct list_head
*first
;
278 if (list_empty(&conf
->inactive_list
))
280 first
= conf
->inactive_list
.next
;
281 sh
= list_entry(first
, struct stripe_head
, lru
);
282 list_del_init(first
);
284 atomic_inc(&conf
->active_stripes
);
289 static void shrink_buffers(struct stripe_head
*sh
)
293 int num
= sh
->raid_conf
->pool_size
;
295 for (i
= 0; i
< num
; i
++) {
299 sh
->dev
[i
].page
= NULL
;
304 static int grow_buffers(struct stripe_head
*sh
)
307 int num
= sh
->raid_conf
->pool_size
;
309 for (i
= 0; i
< num
; i
++) {
312 if (!(page
= alloc_page(GFP_KERNEL
))) {
315 sh
->dev
[i
].page
= page
;
320 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
321 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
322 struct stripe_head
*sh
);
324 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
326 struct r5conf
*conf
= sh
->raid_conf
;
329 BUG_ON(atomic_read(&sh
->count
) != 0);
330 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
331 BUG_ON(stripe_operations_active(sh
));
333 pr_debug("init_stripe called, stripe %llu\n",
334 (unsigned long long)sh
->sector
);
338 sh
->generation
= conf
->generation
- previous
;
339 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
341 stripe_set_idx(sector
, conf
, previous
, sh
);
345 for (i
= sh
->disks
; i
--; ) {
346 struct r5dev
*dev
= &sh
->dev
[i
];
348 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
349 test_bit(R5_LOCKED
, &dev
->flags
)) {
350 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
351 (unsigned long long)sh
->sector
, i
, dev
->toread
,
352 dev
->read
, dev
->towrite
, dev
->written
,
353 test_bit(R5_LOCKED
, &dev
->flags
));
357 raid5_build_block(sh
, i
, previous
);
359 insert_hash(conf
, sh
);
362 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
365 struct stripe_head
*sh
;
367 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
368 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
369 if (sh
->sector
== sector
&& sh
->generation
== generation
)
371 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
376 * Need to check if array has failed when deciding whether to:
378 * - remove non-faulty devices
381 * This determination is simple when no reshape is happening.
382 * However if there is a reshape, we need to carefully check
383 * both the before and after sections.
384 * This is because some failed devices may only affect one
385 * of the two sections, and some non-in_sync devices may
386 * be insync in the section most affected by failed devices.
388 static int calc_degraded(struct r5conf
*conf
)
390 int degraded
, degraded2
;
395 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
396 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
397 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
398 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
399 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
401 else if (test_bit(In_sync
, &rdev
->flags
))
404 /* not in-sync or faulty.
405 * If the reshape increases the number of devices,
406 * this is being recovered by the reshape, so
407 * this 'previous' section is not in_sync.
408 * If the number of devices is being reduced however,
409 * the device can only be part of the array if
410 * we are reverting a reshape, so this section will
413 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
417 if (conf
->raid_disks
== conf
->previous_raid_disks
)
421 for (i
= 0; i
< conf
->raid_disks
; i
++) {
422 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
423 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
424 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
425 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
427 else if (test_bit(In_sync
, &rdev
->flags
))
430 /* not in-sync or faulty.
431 * If reshape increases the number of devices, this
432 * section has already been recovered, else it
433 * almost certainly hasn't.
435 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
439 if (degraded2
> degraded
)
444 static int has_failed(struct r5conf
*conf
)
448 if (conf
->mddev
->reshape_position
== MaxSector
)
449 return conf
->mddev
->degraded
> conf
->max_degraded
;
451 degraded
= calc_degraded(conf
);
452 if (degraded
> conf
->max_degraded
)
457 static struct stripe_head
*
458 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
459 int previous
, int noblock
, int noquiesce
)
461 struct stripe_head
*sh
;
463 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
465 spin_lock_irq(&conf
->device_lock
);
468 wait_event_lock_irq(conf
->wait_for_stripe
,
469 conf
->quiesce
== 0 || noquiesce
,
471 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
473 if (!conf
->inactive_blocked
)
474 sh
= get_free_stripe(conf
);
475 if (noblock
&& sh
== NULL
)
478 conf
->inactive_blocked
= 1;
479 wait_event_lock_irq(conf
->wait_for_stripe
,
480 !list_empty(&conf
->inactive_list
) &&
481 (atomic_read(&conf
->active_stripes
)
482 < (conf
->max_nr_stripes
*3/4)
483 || !conf
->inactive_blocked
),
485 conf
->inactive_blocked
= 0;
487 init_stripe(sh
, sector
, previous
);
489 if (atomic_read(&sh
->count
)) {
490 BUG_ON(!list_empty(&sh
->lru
)
491 && !test_bit(STRIPE_EXPANDING
, &sh
->state
)
492 && !test_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
));
494 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
495 atomic_inc(&conf
->active_stripes
);
496 if (list_empty(&sh
->lru
) &&
497 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
499 list_del_init(&sh
->lru
);
502 } while (sh
== NULL
);
505 atomic_inc(&sh
->count
);
507 spin_unlock_irq(&conf
->device_lock
);
511 /* Determine if 'data_offset' or 'new_data_offset' should be used
512 * in this stripe_head.
514 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
516 sector_t progress
= conf
->reshape_progress
;
517 /* Need a memory barrier to make sure we see the value
518 * of conf->generation, or ->data_offset that was set before
519 * reshape_progress was updated.
522 if (progress
== MaxSector
)
524 if (sh
->generation
== conf
->generation
- 1)
526 /* We are in a reshape, and this is a new-generation stripe,
527 * so use new_data_offset.
533 raid5_end_read_request(struct bio
*bi
, int error
);
535 raid5_end_write_request(struct bio
*bi
, int error
);
537 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
539 struct r5conf
*conf
= sh
->raid_conf
;
540 int i
, disks
= sh
->disks
;
544 for (i
= disks
; i
--; ) {
546 int replace_only
= 0;
547 struct bio
*bi
, *rbi
;
548 struct md_rdev
*rdev
, *rrdev
= NULL
;
549 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
550 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
554 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
556 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
558 else if (test_and_clear_bit(R5_WantReplace
,
559 &sh
->dev
[i
].flags
)) {
564 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
567 bi
= &sh
->dev
[i
].req
;
568 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
573 bi
->bi_end_io
= raid5_end_write_request
;
574 rbi
->bi_end_io
= raid5_end_write_request
;
576 bi
->bi_end_io
= raid5_end_read_request
;
579 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
580 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
581 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
590 /* We raced and saw duplicates */
593 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
) && rrdev
)
598 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
601 atomic_inc(&rdev
->nr_pending
);
602 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
605 atomic_inc(&rrdev
->nr_pending
);
608 /* We have already checked bad blocks for reads. Now
609 * need to check for writes. We never accept write errors
610 * on the replacement, so we don't to check rrdev.
612 while ((rw
& WRITE
) && rdev
&&
613 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
616 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
617 &first_bad
, &bad_sectors
);
622 set_bit(BlockedBadBlocks
, &rdev
->flags
);
623 if (!conf
->mddev
->external
&&
624 conf
->mddev
->flags
) {
625 /* It is very unlikely, but we might
626 * still need to write out the
627 * bad block log - better give it
629 md_check_recovery(conf
->mddev
);
632 * Because md_wait_for_blocked_rdev
633 * will dec nr_pending, we must
634 * increment it first.
636 atomic_inc(&rdev
->nr_pending
);
637 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
639 /* Acknowledged bad block - skip the write */
640 rdev_dec_pending(rdev
, conf
->mddev
);
646 if (s
->syncing
|| s
->expanding
|| s
->expanded
648 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
650 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
652 bi
->bi_bdev
= rdev
->bdev
;
653 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
654 __func__
, (unsigned long long)sh
->sector
,
656 atomic_inc(&sh
->count
);
657 if (use_new_offset(conf
, sh
))
658 bi
->bi_sector
= (sh
->sector
659 + rdev
->new_data_offset
);
661 bi
->bi_sector
= (sh
->sector
662 + rdev
->data_offset
);
663 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
664 bi
->bi_rw
|= REQ_FLUSH
;
666 bi
->bi_flags
= 1 << BIO_UPTODATE
;
668 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
669 bi
->bi_io_vec
[0].bv_offset
= 0;
670 bi
->bi_size
= STRIPE_SIZE
;
673 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
675 if (conf
->mddev
->gendisk
)
676 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
677 bi
, disk_devt(conf
->mddev
->gendisk
),
679 generic_make_request(bi
);
682 if (s
->syncing
|| s
->expanding
|| s
->expanded
684 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
686 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
688 rbi
->bi_bdev
= rrdev
->bdev
;
689 pr_debug("%s: for %llu schedule op %ld on "
690 "replacement disc %d\n",
691 __func__
, (unsigned long long)sh
->sector
,
693 atomic_inc(&sh
->count
);
694 if (use_new_offset(conf
, sh
))
695 rbi
->bi_sector
= (sh
->sector
696 + rrdev
->new_data_offset
);
698 rbi
->bi_sector
= (sh
->sector
699 + rrdev
->data_offset
);
700 rbi
->bi_flags
= 1 << BIO_UPTODATE
;
702 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
703 rbi
->bi_io_vec
[0].bv_offset
= 0;
704 rbi
->bi_size
= STRIPE_SIZE
;
706 if (conf
->mddev
->gendisk
)
707 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
708 rbi
, disk_devt(conf
->mddev
->gendisk
),
710 generic_make_request(rbi
);
712 if (!rdev
&& !rrdev
) {
714 set_bit(STRIPE_DEGRADED
, &sh
->state
);
715 pr_debug("skip op %ld on disc %d for sector %llu\n",
716 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
717 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
718 set_bit(STRIPE_HANDLE
, &sh
->state
);
723 static struct dma_async_tx_descriptor
*
724 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
725 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
728 struct page
*bio_page
;
731 struct async_submit_ctl submit
;
732 enum async_tx_flags flags
= 0;
734 if (bio
->bi_sector
>= sector
)
735 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
737 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
740 flags
|= ASYNC_TX_FENCE
;
741 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
743 bio_for_each_segment(bvl
, bio
, i
) {
744 int len
= bvl
->bv_len
;
748 if (page_offset
< 0) {
749 b_offset
= -page_offset
;
750 page_offset
+= b_offset
;
754 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
755 clen
= STRIPE_SIZE
- page_offset
;
760 b_offset
+= bvl
->bv_offset
;
761 bio_page
= bvl
->bv_page
;
763 tx
= async_memcpy(page
, bio_page
, page_offset
,
764 b_offset
, clen
, &submit
);
766 tx
= async_memcpy(bio_page
, page
, b_offset
,
767 page_offset
, clen
, &submit
);
769 /* chain the operations */
770 submit
.depend_tx
= tx
;
772 if (clen
< len
) /* hit end of page */
780 static void ops_complete_biofill(void *stripe_head_ref
)
782 struct stripe_head
*sh
= stripe_head_ref
;
783 struct bio
*return_bi
= NULL
;
786 pr_debug("%s: stripe %llu\n", __func__
,
787 (unsigned long long)sh
->sector
);
789 /* clear completed biofills */
790 for (i
= sh
->disks
; i
--; ) {
791 struct r5dev
*dev
= &sh
->dev
[i
];
793 /* acknowledge completion of a biofill operation */
794 /* and check if we need to reply to a read request,
795 * new R5_Wantfill requests are held off until
796 * !STRIPE_BIOFILL_RUN
798 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
799 struct bio
*rbi
, *rbi2
;
804 while (rbi
&& rbi
->bi_sector
<
805 dev
->sector
+ STRIPE_SECTORS
) {
806 rbi2
= r5_next_bio(rbi
, dev
->sector
);
807 if (!raid5_dec_bi_active_stripes(rbi
)) {
808 rbi
->bi_next
= return_bi
;
815 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
817 return_io(return_bi
);
819 set_bit(STRIPE_HANDLE
, &sh
->state
);
823 static void ops_run_biofill(struct stripe_head
*sh
)
825 struct dma_async_tx_descriptor
*tx
= NULL
;
826 struct async_submit_ctl submit
;
829 pr_debug("%s: stripe %llu\n", __func__
,
830 (unsigned long long)sh
->sector
);
832 for (i
= sh
->disks
; i
--; ) {
833 struct r5dev
*dev
= &sh
->dev
[i
];
834 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
836 spin_lock_irq(&sh
->stripe_lock
);
837 dev
->read
= rbi
= dev
->toread
;
839 spin_unlock_irq(&sh
->stripe_lock
);
840 while (rbi
&& rbi
->bi_sector
<
841 dev
->sector
+ STRIPE_SECTORS
) {
842 tx
= async_copy_data(0, rbi
, dev
->page
,
844 rbi
= r5_next_bio(rbi
, dev
->sector
);
849 atomic_inc(&sh
->count
);
850 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
851 async_trigger_callback(&submit
);
854 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
861 tgt
= &sh
->dev
[target
];
862 set_bit(R5_UPTODATE
, &tgt
->flags
);
863 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
864 clear_bit(R5_Wantcompute
, &tgt
->flags
);
867 static void ops_complete_compute(void *stripe_head_ref
)
869 struct stripe_head
*sh
= stripe_head_ref
;
871 pr_debug("%s: stripe %llu\n", __func__
,
872 (unsigned long long)sh
->sector
);
874 /* mark the computed target(s) as uptodate */
875 mark_target_uptodate(sh
, sh
->ops
.target
);
876 mark_target_uptodate(sh
, sh
->ops
.target2
);
878 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
879 if (sh
->check_state
== check_state_compute_run
)
880 sh
->check_state
= check_state_compute_result
;
881 set_bit(STRIPE_HANDLE
, &sh
->state
);
885 /* return a pointer to the address conversion region of the scribble buffer */
886 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
887 struct raid5_percpu
*percpu
)
889 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
892 static struct dma_async_tx_descriptor
*
893 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
895 int disks
= sh
->disks
;
896 struct page
**xor_srcs
= percpu
->scribble
;
897 int target
= sh
->ops
.target
;
898 struct r5dev
*tgt
= &sh
->dev
[target
];
899 struct page
*xor_dest
= tgt
->page
;
901 struct dma_async_tx_descriptor
*tx
;
902 struct async_submit_ctl submit
;
905 pr_debug("%s: stripe %llu block: %d\n",
906 __func__
, (unsigned long long)sh
->sector
, target
);
907 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
909 for (i
= disks
; i
--; )
911 xor_srcs
[count
++] = sh
->dev
[i
].page
;
913 atomic_inc(&sh
->count
);
915 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
916 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
917 if (unlikely(count
== 1))
918 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
920 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
925 /* set_syndrome_sources - populate source buffers for gen_syndrome
926 * @srcs - (struct page *) array of size sh->disks
927 * @sh - stripe_head to parse
929 * Populates srcs in proper layout order for the stripe and returns the
930 * 'count' of sources to be used in a call to async_gen_syndrome. The P
931 * destination buffer is recorded in srcs[count] and the Q destination
932 * is recorded in srcs[count+1]].
934 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
936 int disks
= sh
->disks
;
937 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
938 int d0_idx
= raid6_d0(sh
);
942 for (i
= 0; i
< disks
; i
++)
948 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
950 srcs
[slot
] = sh
->dev
[i
].page
;
951 i
= raid6_next_disk(i
, disks
);
952 } while (i
!= d0_idx
);
954 return syndrome_disks
;
957 static struct dma_async_tx_descriptor
*
958 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
960 int disks
= sh
->disks
;
961 struct page
**blocks
= percpu
->scribble
;
963 int qd_idx
= sh
->qd_idx
;
964 struct dma_async_tx_descriptor
*tx
;
965 struct async_submit_ctl submit
;
971 if (sh
->ops
.target
< 0)
972 target
= sh
->ops
.target2
;
973 else if (sh
->ops
.target2
< 0)
974 target
= sh
->ops
.target
;
976 /* we should only have one valid target */
979 pr_debug("%s: stripe %llu block: %d\n",
980 __func__
, (unsigned long long)sh
->sector
, target
);
982 tgt
= &sh
->dev
[target
];
983 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
986 atomic_inc(&sh
->count
);
988 if (target
== qd_idx
) {
989 count
= set_syndrome_sources(blocks
, sh
);
990 blocks
[count
] = NULL
; /* regenerating p is not necessary */
991 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
992 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
993 ops_complete_compute
, sh
,
994 to_addr_conv(sh
, percpu
));
995 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
997 /* Compute any data- or p-drive using XOR */
999 for (i
= disks
; i
-- ; ) {
1000 if (i
== target
|| i
== qd_idx
)
1002 blocks
[count
++] = sh
->dev
[i
].page
;
1005 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1006 NULL
, ops_complete_compute
, sh
,
1007 to_addr_conv(sh
, percpu
));
1008 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1014 static struct dma_async_tx_descriptor
*
1015 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1017 int i
, count
, disks
= sh
->disks
;
1018 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1019 int d0_idx
= raid6_d0(sh
);
1020 int faila
= -1, failb
= -1;
1021 int target
= sh
->ops
.target
;
1022 int target2
= sh
->ops
.target2
;
1023 struct r5dev
*tgt
= &sh
->dev
[target
];
1024 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1025 struct dma_async_tx_descriptor
*tx
;
1026 struct page
**blocks
= percpu
->scribble
;
1027 struct async_submit_ctl submit
;
1029 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1030 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1031 BUG_ON(target
< 0 || target2
< 0);
1032 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1033 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1035 /* we need to open-code set_syndrome_sources to handle the
1036 * slot number conversion for 'faila' and 'failb'
1038 for (i
= 0; i
< disks
; i
++)
1043 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1045 blocks
[slot
] = sh
->dev
[i
].page
;
1051 i
= raid6_next_disk(i
, disks
);
1052 } while (i
!= d0_idx
);
1054 BUG_ON(faila
== failb
);
1057 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1058 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1060 atomic_inc(&sh
->count
);
1062 if (failb
== syndrome_disks
+1) {
1063 /* Q disk is one of the missing disks */
1064 if (faila
== syndrome_disks
) {
1065 /* Missing P+Q, just recompute */
1066 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1067 ops_complete_compute
, sh
,
1068 to_addr_conv(sh
, percpu
));
1069 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1070 STRIPE_SIZE
, &submit
);
1074 int qd_idx
= sh
->qd_idx
;
1076 /* Missing D+Q: recompute D from P, then recompute Q */
1077 if (target
== qd_idx
)
1078 data_target
= target2
;
1080 data_target
= target
;
1083 for (i
= disks
; i
-- ; ) {
1084 if (i
== data_target
|| i
== qd_idx
)
1086 blocks
[count
++] = sh
->dev
[i
].page
;
1088 dest
= sh
->dev
[data_target
].page
;
1089 init_async_submit(&submit
,
1090 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1092 to_addr_conv(sh
, percpu
));
1093 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1096 count
= set_syndrome_sources(blocks
, sh
);
1097 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1098 ops_complete_compute
, sh
,
1099 to_addr_conv(sh
, percpu
));
1100 return async_gen_syndrome(blocks
, 0, count
+2,
1101 STRIPE_SIZE
, &submit
);
1104 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1105 ops_complete_compute
, sh
,
1106 to_addr_conv(sh
, percpu
));
1107 if (failb
== syndrome_disks
) {
1108 /* We're missing D+P. */
1109 return async_raid6_datap_recov(syndrome_disks
+2,
1113 /* We're missing D+D. */
1114 return async_raid6_2data_recov(syndrome_disks
+2,
1115 STRIPE_SIZE
, faila
, failb
,
1122 static void ops_complete_prexor(void *stripe_head_ref
)
1124 struct stripe_head
*sh
= stripe_head_ref
;
1126 pr_debug("%s: stripe %llu\n", __func__
,
1127 (unsigned long long)sh
->sector
);
1130 static struct dma_async_tx_descriptor
*
1131 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1132 struct dma_async_tx_descriptor
*tx
)
1134 int disks
= sh
->disks
;
1135 struct page
**xor_srcs
= percpu
->scribble
;
1136 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1137 struct async_submit_ctl submit
;
1139 /* existing parity data subtracted */
1140 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1142 pr_debug("%s: stripe %llu\n", __func__
,
1143 (unsigned long long)sh
->sector
);
1145 for (i
= disks
; i
--; ) {
1146 struct r5dev
*dev
= &sh
->dev
[i
];
1147 /* Only process blocks that are known to be uptodate */
1148 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1149 xor_srcs
[count
++] = dev
->page
;
1152 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1153 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1154 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1159 static struct dma_async_tx_descriptor
*
1160 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1162 int disks
= sh
->disks
;
1165 pr_debug("%s: stripe %llu\n", __func__
,
1166 (unsigned long long)sh
->sector
);
1168 for (i
= disks
; i
--; ) {
1169 struct r5dev
*dev
= &sh
->dev
[i
];
1172 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1175 spin_lock_irq(&sh
->stripe_lock
);
1176 chosen
= dev
->towrite
;
1177 dev
->towrite
= NULL
;
1178 BUG_ON(dev
->written
);
1179 wbi
= dev
->written
= chosen
;
1180 spin_unlock_irq(&sh
->stripe_lock
);
1182 while (wbi
&& wbi
->bi_sector
<
1183 dev
->sector
+ STRIPE_SECTORS
) {
1184 if (wbi
->bi_rw
& REQ_FUA
)
1185 set_bit(R5_WantFUA
, &dev
->flags
);
1186 if (wbi
->bi_rw
& REQ_SYNC
)
1187 set_bit(R5_SyncIO
, &dev
->flags
);
1188 if (wbi
->bi_rw
& REQ_DISCARD
)
1189 set_bit(R5_Discard
, &dev
->flags
);
1191 tx
= async_copy_data(1, wbi
, dev
->page
,
1193 wbi
= r5_next_bio(wbi
, dev
->sector
);
1201 static void ops_complete_reconstruct(void *stripe_head_ref
)
1203 struct stripe_head
*sh
= stripe_head_ref
;
1204 int disks
= sh
->disks
;
1205 int pd_idx
= sh
->pd_idx
;
1206 int qd_idx
= sh
->qd_idx
;
1208 bool fua
= false, sync
= false, discard
= false;
1210 pr_debug("%s: stripe %llu\n", __func__
,
1211 (unsigned long long)sh
->sector
);
1213 for (i
= disks
; i
--; ) {
1214 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1215 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1216 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1219 for (i
= disks
; i
--; ) {
1220 struct r5dev
*dev
= &sh
->dev
[i
];
1222 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1224 set_bit(R5_UPTODATE
, &dev
->flags
);
1226 set_bit(R5_WantFUA
, &dev
->flags
);
1228 set_bit(R5_SyncIO
, &dev
->flags
);
1232 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1233 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1234 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1235 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1237 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1238 sh
->reconstruct_state
= reconstruct_state_result
;
1241 set_bit(STRIPE_HANDLE
, &sh
->state
);
1246 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1247 struct dma_async_tx_descriptor
*tx
)
1249 int disks
= sh
->disks
;
1250 struct page
**xor_srcs
= percpu
->scribble
;
1251 struct async_submit_ctl submit
;
1252 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1253 struct page
*xor_dest
;
1255 unsigned long flags
;
1257 pr_debug("%s: stripe %llu\n", __func__
,
1258 (unsigned long long)sh
->sector
);
1260 for (i
= 0; i
< sh
->disks
; i
++) {
1263 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1266 if (i
>= sh
->disks
) {
1267 atomic_inc(&sh
->count
);
1268 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1269 ops_complete_reconstruct(sh
);
1272 /* check if prexor is active which means only process blocks
1273 * that are part of a read-modify-write (written)
1275 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1277 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1278 for (i
= disks
; i
--; ) {
1279 struct r5dev
*dev
= &sh
->dev
[i
];
1281 xor_srcs
[count
++] = dev
->page
;
1284 xor_dest
= sh
->dev
[pd_idx
].page
;
1285 for (i
= disks
; i
--; ) {
1286 struct r5dev
*dev
= &sh
->dev
[i
];
1288 xor_srcs
[count
++] = dev
->page
;
1292 /* 1/ if we prexor'd then the dest is reused as a source
1293 * 2/ if we did not prexor then we are redoing the parity
1294 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1295 * for the synchronous xor case
1297 flags
= ASYNC_TX_ACK
|
1298 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1300 atomic_inc(&sh
->count
);
1302 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1303 to_addr_conv(sh
, percpu
));
1304 if (unlikely(count
== 1))
1305 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1307 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1311 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1312 struct dma_async_tx_descriptor
*tx
)
1314 struct async_submit_ctl submit
;
1315 struct page
**blocks
= percpu
->scribble
;
1318 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1320 for (i
= 0; i
< sh
->disks
; i
++) {
1321 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1323 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1326 if (i
>= sh
->disks
) {
1327 atomic_inc(&sh
->count
);
1328 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1329 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1330 ops_complete_reconstruct(sh
);
1334 count
= set_syndrome_sources(blocks
, sh
);
1336 atomic_inc(&sh
->count
);
1338 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1339 sh
, to_addr_conv(sh
, percpu
));
1340 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1343 static void ops_complete_check(void *stripe_head_ref
)
1345 struct stripe_head
*sh
= stripe_head_ref
;
1347 pr_debug("%s: stripe %llu\n", __func__
,
1348 (unsigned long long)sh
->sector
);
1350 sh
->check_state
= check_state_check_result
;
1351 set_bit(STRIPE_HANDLE
, &sh
->state
);
1355 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1357 int disks
= sh
->disks
;
1358 int pd_idx
= sh
->pd_idx
;
1359 int qd_idx
= sh
->qd_idx
;
1360 struct page
*xor_dest
;
1361 struct page
**xor_srcs
= percpu
->scribble
;
1362 struct dma_async_tx_descriptor
*tx
;
1363 struct async_submit_ctl submit
;
1367 pr_debug("%s: stripe %llu\n", __func__
,
1368 (unsigned long long)sh
->sector
);
1371 xor_dest
= sh
->dev
[pd_idx
].page
;
1372 xor_srcs
[count
++] = xor_dest
;
1373 for (i
= disks
; i
--; ) {
1374 if (i
== pd_idx
|| i
== qd_idx
)
1376 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1379 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1380 to_addr_conv(sh
, percpu
));
1381 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1382 &sh
->ops
.zero_sum_result
, &submit
);
1384 atomic_inc(&sh
->count
);
1385 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1386 tx
= async_trigger_callback(&submit
);
1389 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1391 struct page
**srcs
= percpu
->scribble
;
1392 struct async_submit_ctl submit
;
1395 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1396 (unsigned long long)sh
->sector
, checkp
);
1398 count
= set_syndrome_sources(srcs
, sh
);
1402 atomic_inc(&sh
->count
);
1403 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1404 sh
, to_addr_conv(sh
, percpu
));
1405 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1406 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1409 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1411 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1412 struct dma_async_tx_descriptor
*tx
= NULL
;
1413 struct r5conf
*conf
= sh
->raid_conf
;
1414 int level
= conf
->level
;
1415 struct raid5_percpu
*percpu
;
1419 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1420 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1421 ops_run_biofill(sh
);
1425 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1427 tx
= ops_run_compute5(sh
, percpu
);
1429 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1430 tx
= ops_run_compute6_1(sh
, percpu
);
1432 tx
= ops_run_compute6_2(sh
, percpu
);
1434 /* terminate the chain if reconstruct is not set to be run */
1435 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1439 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1440 tx
= ops_run_prexor(sh
, percpu
, tx
);
1442 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1443 tx
= ops_run_biodrain(sh
, tx
);
1447 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1449 ops_run_reconstruct5(sh
, percpu
, tx
);
1451 ops_run_reconstruct6(sh
, percpu
, tx
);
1454 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1455 if (sh
->check_state
== check_state_run
)
1456 ops_run_check_p(sh
, percpu
);
1457 else if (sh
->check_state
== check_state_run_q
)
1458 ops_run_check_pq(sh
, percpu
, 0);
1459 else if (sh
->check_state
== check_state_run_pq
)
1460 ops_run_check_pq(sh
, percpu
, 1);
1466 for (i
= disks
; i
--; ) {
1467 struct r5dev
*dev
= &sh
->dev
[i
];
1468 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1469 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1474 static int grow_one_stripe(struct r5conf
*conf
)
1476 struct stripe_head
*sh
;
1477 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1481 sh
->raid_conf
= conf
;
1483 spin_lock_init(&sh
->stripe_lock
);
1485 if (grow_buffers(sh
)) {
1487 kmem_cache_free(conf
->slab_cache
, sh
);
1490 /* we just created an active stripe so... */
1491 atomic_set(&sh
->count
, 1);
1492 atomic_inc(&conf
->active_stripes
);
1493 INIT_LIST_HEAD(&sh
->lru
);
1498 static int grow_stripes(struct r5conf
*conf
, int num
)
1500 struct kmem_cache
*sc
;
1501 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1503 if (conf
->mddev
->gendisk
)
1504 sprintf(conf
->cache_name
[0],
1505 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1507 sprintf(conf
->cache_name
[0],
1508 "raid%d-%p", conf
->level
, conf
->mddev
);
1509 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1511 conf
->active_name
= 0;
1512 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1513 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1517 conf
->slab_cache
= sc
;
1518 conf
->pool_size
= devs
;
1520 if (!grow_one_stripe(conf
))
1526 * scribble_len - return the required size of the scribble region
1527 * @num - total number of disks in the array
1529 * The size must be enough to contain:
1530 * 1/ a struct page pointer for each device in the array +2
1531 * 2/ room to convert each entry in (1) to its corresponding dma
1532 * (dma_map_page()) or page (page_address()) address.
1534 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1535 * calculate over all devices (not just the data blocks), using zeros in place
1536 * of the P and Q blocks.
1538 static size_t scribble_len(int num
)
1542 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1547 static int resize_stripes(struct r5conf
*conf
, int newsize
)
1549 /* Make all the stripes able to hold 'newsize' devices.
1550 * New slots in each stripe get 'page' set to a new page.
1552 * This happens in stages:
1553 * 1/ create a new kmem_cache and allocate the required number of
1555 * 2/ gather all the old stripe_heads and transfer the pages across
1556 * to the new stripe_heads. This will have the side effect of
1557 * freezing the array as once all stripe_heads have been collected,
1558 * no IO will be possible. Old stripe heads are freed once their
1559 * pages have been transferred over, and the old kmem_cache is
1560 * freed when all stripes are done.
1561 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1562 * we simple return a failre status - no need to clean anything up.
1563 * 4/ allocate new pages for the new slots in the new stripe_heads.
1564 * If this fails, we don't bother trying the shrink the
1565 * stripe_heads down again, we just leave them as they are.
1566 * As each stripe_head is processed the new one is released into
1569 * Once step2 is started, we cannot afford to wait for a write,
1570 * so we use GFP_NOIO allocations.
1572 struct stripe_head
*osh
, *nsh
;
1573 LIST_HEAD(newstripes
);
1574 struct disk_info
*ndisks
;
1577 struct kmem_cache
*sc
;
1580 if (newsize
<= conf
->pool_size
)
1581 return 0; /* never bother to shrink */
1583 err
= md_allow_write(conf
->mddev
);
1588 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1589 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1594 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1595 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1599 nsh
->raid_conf
= conf
;
1600 spin_lock_init(&nsh
->stripe_lock
);
1602 list_add(&nsh
->lru
, &newstripes
);
1605 /* didn't get enough, give up */
1606 while (!list_empty(&newstripes
)) {
1607 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1608 list_del(&nsh
->lru
);
1609 kmem_cache_free(sc
, nsh
);
1611 kmem_cache_destroy(sc
);
1614 /* Step 2 - Must use GFP_NOIO now.
1615 * OK, we have enough stripes, start collecting inactive
1616 * stripes and copying them over
1618 list_for_each_entry(nsh
, &newstripes
, lru
) {
1619 spin_lock_irq(&conf
->device_lock
);
1620 wait_event_lock_irq(conf
->wait_for_stripe
,
1621 !list_empty(&conf
->inactive_list
),
1623 osh
= get_free_stripe(conf
);
1624 spin_unlock_irq(&conf
->device_lock
);
1625 atomic_set(&nsh
->count
, 1);
1626 for(i
=0; i
<conf
->pool_size
; i
++)
1627 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1628 for( ; i
<newsize
; i
++)
1629 nsh
->dev
[i
].page
= NULL
;
1630 kmem_cache_free(conf
->slab_cache
, osh
);
1632 kmem_cache_destroy(conf
->slab_cache
);
1635 * At this point, we are holding all the stripes so the array
1636 * is completely stalled, so now is a good time to resize
1637 * conf->disks and the scribble region
1639 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1641 for (i
=0; i
<conf
->raid_disks
; i
++)
1642 ndisks
[i
] = conf
->disks
[i
];
1644 conf
->disks
= ndisks
;
1649 conf
->scribble_len
= scribble_len(newsize
);
1650 for_each_present_cpu(cpu
) {
1651 struct raid5_percpu
*percpu
;
1654 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1655 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1658 kfree(percpu
->scribble
);
1659 percpu
->scribble
= scribble
;
1667 /* Step 4, return new stripes to service */
1668 while(!list_empty(&newstripes
)) {
1669 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1670 list_del_init(&nsh
->lru
);
1672 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1673 if (nsh
->dev
[i
].page
== NULL
) {
1674 struct page
*p
= alloc_page(GFP_NOIO
);
1675 nsh
->dev
[i
].page
= p
;
1679 release_stripe(nsh
);
1681 /* critical section pass, GFP_NOIO no longer needed */
1683 conf
->slab_cache
= sc
;
1684 conf
->active_name
= 1-conf
->active_name
;
1685 conf
->pool_size
= newsize
;
1689 static int drop_one_stripe(struct r5conf
*conf
)
1691 struct stripe_head
*sh
;
1693 spin_lock_irq(&conf
->device_lock
);
1694 sh
= get_free_stripe(conf
);
1695 spin_unlock_irq(&conf
->device_lock
);
1698 BUG_ON(atomic_read(&sh
->count
));
1700 kmem_cache_free(conf
->slab_cache
, sh
);
1701 atomic_dec(&conf
->active_stripes
);
1705 static void shrink_stripes(struct r5conf
*conf
)
1707 while (drop_one_stripe(conf
))
1710 if (conf
->slab_cache
)
1711 kmem_cache_destroy(conf
->slab_cache
);
1712 conf
->slab_cache
= NULL
;
1715 static void raid5_end_read_request(struct bio
* bi
, int error
)
1717 struct stripe_head
*sh
= bi
->bi_private
;
1718 struct r5conf
*conf
= sh
->raid_conf
;
1719 int disks
= sh
->disks
, i
;
1720 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1721 char b
[BDEVNAME_SIZE
];
1722 struct md_rdev
*rdev
= NULL
;
1725 for (i
=0 ; i
<disks
; i
++)
1726 if (bi
== &sh
->dev
[i
].req
)
1729 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1730 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1736 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1737 /* If replacement finished while this request was outstanding,
1738 * 'replacement' might be NULL already.
1739 * In that case it moved down to 'rdev'.
1740 * rdev is not removed until all requests are finished.
1742 rdev
= conf
->disks
[i
].replacement
;
1744 rdev
= conf
->disks
[i
].rdev
;
1746 if (use_new_offset(conf
, sh
))
1747 s
= sh
->sector
+ rdev
->new_data_offset
;
1749 s
= sh
->sector
+ rdev
->data_offset
;
1751 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1752 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1753 /* Note that this cannot happen on a
1754 * replacement device. We just fail those on
1759 "md/raid:%s: read error corrected"
1760 " (%lu sectors at %llu on %s)\n",
1761 mdname(conf
->mddev
), STRIPE_SECTORS
,
1762 (unsigned long long)s
,
1763 bdevname(rdev
->bdev
, b
));
1764 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1765 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1766 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1767 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
1768 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1770 if (atomic_read(&rdev
->read_errors
))
1771 atomic_set(&rdev
->read_errors
, 0);
1773 const char *bdn
= bdevname(rdev
->bdev
, b
);
1777 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1778 atomic_inc(&rdev
->read_errors
);
1779 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1782 "md/raid:%s: read error on replacement device "
1783 "(sector %llu on %s).\n",
1784 mdname(conf
->mddev
),
1785 (unsigned long long)s
,
1787 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
1791 "md/raid:%s: read error not correctable "
1792 "(sector %llu on %s).\n",
1793 mdname(conf
->mddev
),
1794 (unsigned long long)s
,
1796 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
1801 "md/raid:%s: read error NOT corrected!! "
1802 "(sector %llu on %s).\n",
1803 mdname(conf
->mddev
),
1804 (unsigned long long)s
,
1806 } else if (atomic_read(&rdev
->read_errors
)
1807 > conf
->max_nr_stripes
)
1809 "md/raid:%s: Too many read errors, failing device %s.\n",
1810 mdname(conf
->mddev
), bdn
);
1814 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
1815 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1816 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1818 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1820 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1821 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1823 && test_bit(In_sync
, &rdev
->flags
)
1824 && rdev_set_badblocks(
1825 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
1826 md_error(conf
->mddev
, rdev
);
1829 rdev_dec_pending(rdev
, conf
->mddev
);
1830 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1831 set_bit(STRIPE_HANDLE
, &sh
->state
);
1835 static void raid5_end_write_request(struct bio
*bi
, int error
)
1837 struct stripe_head
*sh
= bi
->bi_private
;
1838 struct r5conf
*conf
= sh
->raid_conf
;
1839 int disks
= sh
->disks
, i
;
1840 struct md_rdev
*uninitialized_var(rdev
);
1841 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1844 int replacement
= 0;
1846 for (i
= 0 ; i
< disks
; i
++) {
1847 if (bi
== &sh
->dev
[i
].req
) {
1848 rdev
= conf
->disks
[i
].rdev
;
1851 if (bi
== &sh
->dev
[i
].rreq
) {
1852 rdev
= conf
->disks
[i
].replacement
;
1856 /* rdev was removed and 'replacement'
1857 * replaced it. rdev is not removed
1858 * until all requests are finished.
1860 rdev
= conf
->disks
[i
].rdev
;
1864 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1865 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1874 md_error(conf
->mddev
, rdev
);
1875 else if (is_badblock(rdev
, sh
->sector
,
1877 &first_bad
, &bad_sectors
))
1878 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
1881 set_bit(WriteErrorSeen
, &rdev
->flags
);
1882 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
1883 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1884 set_bit(MD_RECOVERY_NEEDED
,
1885 &rdev
->mddev
->recovery
);
1886 } else if (is_badblock(rdev
, sh
->sector
,
1888 &first_bad
, &bad_sectors
))
1889 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
1891 rdev_dec_pending(rdev
, conf
->mddev
);
1893 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
1894 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1895 set_bit(STRIPE_HANDLE
, &sh
->state
);
1899 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1901 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1903 struct r5dev
*dev
= &sh
->dev
[i
];
1905 bio_init(&dev
->req
);
1906 dev
->req
.bi_io_vec
= &dev
->vec
;
1908 dev
->req
.bi_max_vecs
++;
1909 dev
->req
.bi_private
= sh
;
1910 dev
->vec
.bv_page
= dev
->page
;
1912 bio_init(&dev
->rreq
);
1913 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
1914 dev
->rreq
.bi_vcnt
++;
1915 dev
->rreq
.bi_max_vecs
++;
1916 dev
->rreq
.bi_private
= sh
;
1917 dev
->rvec
.bv_page
= dev
->page
;
1920 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1923 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1925 char b
[BDEVNAME_SIZE
];
1926 struct r5conf
*conf
= mddev
->private;
1927 unsigned long flags
;
1928 pr_debug("raid456: error called\n");
1930 spin_lock_irqsave(&conf
->device_lock
, flags
);
1931 clear_bit(In_sync
, &rdev
->flags
);
1932 mddev
->degraded
= calc_degraded(conf
);
1933 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1934 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1936 set_bit(Blocked
, &rdev
->flags
);
1937 set_bit(Faulty
, &rdev
->flags
);
1938 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1940 "md/raid:%s: Disk failure on %s, disabling device.\n"
1941 "md/raid:%s: Operation continuing on %d devices.\n",
1943 bdevname(rdev
->bdev
, b
),
1945 conf
->raid_disks
- mddev
->degraded
);
1949 * Input: a 'big' sector number,
1950 * Output: index of the data and parity disk, and the sector # in them.
1952 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
1953 int previous
, int *dd_idx
,
1954 struct stripe_head
*sh
)
1956 sector_t stripe
, stripe2
;
1957 sector_t chunk_number
;
1958 unsigned int chunk_offset
;
1961 sector_t new_sector
;
1962 int algorithm
= previous
? conf
->prev_algo
1964 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1965 : conf
->chunk_sectors
;
1966 int raid_disks
= previous
? conf
->previous_raid_disks
1968 int data_disks
= raid_disks
- conf
->max_degraded
;
1970 /* First compute the information on this sector */
1973 * Compute the chunk number and the sector offset inside the chunk
1975 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1976 chunk_number
= r_sector
;
1979 * Compute the stripe number
1981 stripe
= chunk_number
;
1982 *dd_idx
= sector_div(stripe
, data_disks
);
1985 * Select the parity disk based on the user selected algorithm.
1987 pd_idx
= qd_idx
= -1;
1988 switch(conf
->level
) {
1990 pd_idx
= data_disks
;
1993 switch (algorithm
) {
1994 case ALGORITHM_LEFT_ASYMMETRIC
:
1995 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1996 if (*dd_idx
>= pd_idx
)
1999 case ALGORITHM_RIGHT_ASYMMETRIC
:
2000 pd_idx
= sector_div(stripe2
, raid_disks
);
2001 if (*dd_idx
>= pd_idx
)
2004 case ALGORITHM_LEFT_SYMMETRIC
:
2005 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2006 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2008 case ALGORITHM_RIGHT_SYMMETRIC
:
2009 pd_idx
= sector_div(stripe2
, raid_disks
);
2010 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2012 case ALGORITHM_PARITY_0
:
2016 case ALGORITHM_PARITY_N
:
2017 pd_idx
= data_disks
;
2025 switch (algorithm
) {
2026 case ALGORITHM_LEFT_ASYMMETRIC
:
2027 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2028 qd_idx
= pd_idx
+ 1;
2029 if (pd_idx
== raid_disks
-1) {
2030 (*dd_idx
)++; /* Q D D D P */
2032 } else if (*dd_idx
>= pd_idx
)
2033 (*dd_idx
) += 2; /* D D P Q D */
2035 case ALGORITHM_RIGHT_ASYMMETRIC
:
2036 pd_idx
= sector_div(stripe2
, raid_disks
);
2037 qd_idx
= pd_idx
+ 1;
2038 if (pd_idx
== raid_disks
-1) {
2039 (*dd_idx
)++; /* Q D D D P */
2041 } else if (*dd_idx
>= pd_idx
)
2042 (*dd_idx
) += 2; /* D D P Q D */
2044 case ALGORITHM_LEFT_SYMMETRIC
:
2045 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2046 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2047 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2049 case ALGORITHM_RIGHT_SYMMETRIC
:
2050 pd_idx
= sector_div(stripe2
, raid_disks
);
2051 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2052 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2055 case ALGORITHM_PARITY_0
:
2060 case ALGORITHM_PARITY_N
:
2061 pd_idx
= data_disks
;
2062 qd_idx
= data_disks
+ 1;
2065 case ALGORITHM_ROTATING_ZERO_RESTART
:
2066 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2067 * of blocks for computing Q is different.
2069 pd_idx
= sector_div(stripe2
, raid_disks
);
2070 qd_idx
= pd_idx
+ 1;
2071 if (pd_idx
== raid_disks
-1) {
2072 (*dd_idx
)++; /* Q D D D P */
2074 } else if (*dd_idx
>= pd_idx
)
2075 (*dd_idx
) += 2; /* D D P Q D */
2079 case ALGORITHM_ROTATING_N_RESTART
:
2080 /* Same a left_asymmetric, by first stripe is
2081 * D D D P Q rather than
2085 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2086 qd_idx
= pd_idx
+ 1;
2087 if (pd_idx
== raid_disks
-1) {
2088 (*dd_idx
)++; /* Q D D D P */
2090 } else if (*dd_idx
>= pd_idx
)
2091 (*dd_idx
) += 2; /* D D P Q D */
2095 case ALGORITHM_ROTATING_N_CONTINUE
:
2096 /* Same as left_symmetric but Q is before P */
2097 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2098 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2099 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2103 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2104 /* RAID5 left_asymmetric, with Q on last device */
2105 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2106 if (*dd_idx
>= pd_idx
)
2108 qd_idx
= raid_disks
- 1;
2111 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2112 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2113 if (*dd_idx
>= pd_idx
)
2115 qd_idx
= raid_disks
- 1;
2118 case ALGORITHM_LEFT_SYMMETRIC_6
:
2119 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2120 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2121 qd_idx
= raid_disks
- 1;
2124 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2125 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2126 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2127 qd_idx
= raid_disks
- 1;
2130 case ALGORITHM_PARITY_0_6
:
2133 qd_idx
= raid_disks
- 1;
2143 sh
->pd_idx
= pd_idx
;
2144 sh
->qd_idx
= qd_idx
;
2145 sh
->ddf_layout
= ddf_layout
;
2148 * Finally, compute the new sector number
2150 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2155 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2157 struct r5conf
*conf
= sh
->raid_conf
;
2158 int raid_disks
= sh
->disks
;
2159 int data_disks
= raid_disks
- conf
->max_degraded
;
2160 sector_t new_sector
= sh
->sector
, check
;
2161 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2162 : conf
->chunk_sectors
;
2163 int algorithm
= previous
? conf
->prev_algo
2167 sector_t chunk_number
;
2168 int dummy1
, dd_idx
= i
;
2170 struct stripe_head sh2
;
2173 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2174 stripe
= new_sector
;
2176 if (i
== sh
->pd_idx
)
2178 switch(conf
->level
) {
2181 switch (algorithm
) {
2182 case ALGORITHM_LEFT_ASYMMETRIC
:
2183 case ALGORITHM_RIGHT_ASYMMETRIC
:
2187 case ALGORITHM_LEFT_SYMMETRIC
:
2188 case ALGORITHM_RIGHT_SYMMETRIC
:
2191 i
-= (sh
->pd_idx
+ 1);
2193 case ALGORITHM_PARITY_0
:
2196 case ALGORITHM_PARITY_N
:
2203 if (i
== sh
->qd_idx
)
2204 return 0; /* It is the Q disk */
2205 switch (algorithm
) {
2206 case ALGORITHM_LEFT_ASYMMETRIC
:
2207 case ALGORITHM_RIGHT_ASYMMETRIC
:
2208 case ALGORITHM_ROTATING_ZERO_RESTART
:
2209 case ALGORITHM_ROTATING_N_RESTART
:
2210 if (sh
->pd_idx
== raid_disks
-1)
2211 i
--; /* Q D D D P */
2212 else if (i
> sh
->pd_idx
)
2213 i
-= 2; /* D D P Q D */
2215 case ALGORITHM_LEFT_SYMMETRIC
:
2216 case ALGORITHM_RIGHT_SYMMETRIC
:
2217 if (sh
->pd_idx
== raid_disks
-1)
2218 i
--; /* Q D D D P */
2223 i
-= (sh
->pd_idx
+ 2);
2226 case ALGORITHM_PARITY_0
:
2229 case ALGORITHM_PARITY_N
:
2231 case ALGORITHM_ROTATING_N_CONTINUE
:
2232 /* Like left_symmetric, but P is before Q */
2233 if (sh
->pd_idx
== 0)
2234 i
--; /* P D D D Q */
2239 i
-= (sh
->pd_idx
+ 1);
2242 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2243 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2247 case ALGORITHM_LEFT_SYMMETRIC_6
:
2248 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2250 i
+= data_disks
+ 1;
2251 i
-= (sh
->pd_idx
+ 1);
2253 case ALGORITHM_PARITY_0_6
:
2262 chunk_number
= stripe
* data_disks
+ i
;
2263 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2265 check
= raid5_compute_sector(conf
, r_sector
,
2266 previous
, &dummy1
, &sh2
);
2267 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2268 || sh2
.qd_idx
!= sh
->qd_idx
) {
2269 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2270 mdname(conf
->mddev
));
2278 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2279 int rcw
, int expand
)
2281 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2282 struct r5conf
*conf
= sh
->raid_conf
;
2283 int level
= conf
->level
;
2287 for (i
= disks
; i
--; ) {
2288 struct r5dev
*dev
= &sh
->dev
[i
];
2291 set_bit(R5_LOCKED
, &dev
->flags
);
2292 set_bit(R5_Wantdrain
, &dev
->flags
);
2294 clear_bit(R5_UPTODATE
, &dev
->flags
);
2298 /* if we are not expanding this is a proper write request, and
2299 * there will be bios with new data to be drained into the
2304 /* False alarm, nothing to do */
2306 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2307 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2309 sh
->reconstruct_state
= reconstruct_state_run
;
2311 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2313 if (s
->locked
+ conf
->max_degraded
== disks
)
2314 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2315 atomic_inc(&conf
->pending_full_writes
);
2318 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2319 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2321 for (i
= disks
; i
--; ) {
2322 struct r5dev
*dev
= &sh
->dev
[i
];
2327 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2328 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2329 set_bit(R5_Wantdrain
, &dev
->flags
);
2330 set_bit(R5_LOCKED
, &dev
->flags
);
2331 clear_bit(R5_UPTODATE
, &dev
->flags
);
2336 /* False alarm - nothing to do */
2338 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2339 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2340 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2341 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2344 /* keep the parity disk(s) locked while asynchronous operations
2347 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2348 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2352 int qd_idx
= sh
->qd_idx
;
2353 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2355 set_bit(R5_LOCKED
, &dev
->flags
);
2356 clear_bit(R5_UPTODATE
, &dev
->flags
);
2360 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2361 __func__
, (unsigned long long)sh
->sector
,
2362 s
->locked
, s
->ops_request
);
2366 * Each stripe/dev can have one or more bion attached.
2367 * toread/towrite point to the first in a chain.
2368 * The bi_next chain must be in order.
2370 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2373 struct r5conf
*conf
= sh
->raid_conf
;
2376 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2377 (unsigned long long)bi
->bi_sector
,
2378 (unsigned long long)sh
->sector
);
2381 * If several bio share a stripe. The bio bi_phys_segments acts as a
2382 * reference count to avoid race. The reference count should already be
2383 * increased before this function is called (for example, in
2384 * make_request()), so other bio sharing this stripe will not free the
2385 * stripe. If a stripe is owned by one stripe, the stripe lock will
2388 spin_lock_irq(&sh
->stripe_lock
);
2390 bip
= &sh
->dev
[dd_idx
].towrite
;
2394 bip
= &sh
->dev
[dd_idx
].toread
;
2395 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2396 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2398 bip
= & (*bip
)->bi_next
;
2400 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2403 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2407 raid5_inc_bi_active_stripes(bi
);
2410 /* check if page is covered */
2411 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2412 for (bi
=sh
->dev
[dd_idx
].towrite
;
2413 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2414 bi
&& bi
->bi_sector
<= sector
;
2415 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2416 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2417 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2419 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2420 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2423 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2424 (unsigned long long)(*bip
)->bi_sector
,
2425 (unsigned long long)sh
->sector
, dd_idx
);
2426 spin_unlock_irq(&sh
->stripe_lock
);
2428 if (conf
->mddev
->bitmap
&& firstwrite
) {
2429 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2431 sh
->bm_seq
= conf
->seq_flush
+1;
2432 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2437 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2438 spin_unlock_irq(&sh
->stripe_lock
);
2442 static void end_reshape(struct r5conf
*conf
);
2444 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2445 struct stripe_head
*sh
)
2447 int sectors_per_chunk
=
2448 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2450 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2451 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2453 raid5_compute_sector(conf
,
2454 stripe
* (disks
- conf
->max_degraded
)
2455 *sectors_per_chunk
+ chunk_offset
,
2461 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2462 struct stripe_head_state
*s
, int disks
,
2463 struct bio
**return_bi
)
2466 for (i
= disks
; i
--; ) {
2470 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2471 struct md_rdev
*rdev
;
2473 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2474 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2475 atomic_inc(&rdev
->nr_pending
);
2480 if (!rdev_set_badblocks(
2484 md_error(conf
->mddev
, rdev
);
2485 rdev_dec_pending(rdev
, conf
->mddev
);
2488 spin_lock_irq(&sh
->stripe_lock
);
2489 /* fail all writes first */
2490 bi
= sh
->dev
[i
].towrite
;
2491 sh
->dev
[i
].towrite
= NULL
;
2492 spin_unlock_irq(&sh
->stripe_lock
);
2496 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2497 wake_up(&conf
->wait_for_overlap
);
2499 while (bi
&& bi
->bi_sector
<
2500 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2501 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2502 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2503 if (!raid5_dec_bi_active_stripes(bi
)) {
2504 md_write_end(conf
->mddev
);
2505 bi
->bi_next
= *return_bi
;
2511 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2512 STRIPE_SECTORS
, 0, 0);
2514 /* and fail all 'written' */
2515 bi
= sh
->dev
[i
].written
;
2516 sh
->dev
[i
].written
= NULL
;
2517 if (bi
) bitmap_end
= 1;
2518 while (bi
&& bi
->bi_sector
<
2519 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2520 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2521 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2522 if (!raid5_dec_bi_active_stripes(bi
)) {
2523 md_write_end(conf
->mddev
);
2524 bi
->bi_next
= *return_bi
;
2530 /* fail any reads if this device is non-operational and
2531 * the data has not reached the cache yet.
2533 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2534 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2535 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2536 spin_lock_irq(&sh
->stripe_lock
);
2537 bi
= sh
->dev
[i
].toread
;
2538 sh
->dev
[i
].toread
= NULL
;
2539 spin_unlock_irq(&sh
->stripe_lock
);
2540 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2541 wake_up(&conf
->wait_for_overlap
);
2542 while (bi
&& bi
->bi_sector
<
2543 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2544 struct bio
*nextbi
=
2545 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2546 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2547 if (!raid5_dec_bi_active_stripes(bi
)) {
2548 bi
->bi_next
= *return_bi
;
2555 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2556 STRIPE_SECTORS
, 0, 0);
2557 /* If we were in the middle of a write the parity block might
2558 * still be locked - so just clear all R5_LOCKED flags
2560 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2563 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2564 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2565 md_wakeup_thread(conf
->mddev
->thread
);
2569 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2570 struct stripe_head_state
*s
)
2575 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2576 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
2577 wake_up(&conf
->wait_for_overlap
);
2580 /* There is nothing more to do for sync/check/repair.
2581 * Don't even need to abort as that is handled elsewhere
2582 * if needed, and not always wanted e.g. if there is a known
2584 * For recover/replace we need to record a bad block on all
2585 * non-sync devices, or abort the recovery
2587 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
2588 /* During recovery devices cannot be removed, so
2589 * locking and refcounting of rdevs is not needed
2591 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2592 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2594 && !test_bit(Faulty
, &rdev
->flags
)
2595 && !test_bit(In_sync
, &rdev
->flags
)
2596 && !rdev_set_badblocks(rdev
, sh
->sector
,
2599 rdev
= conf
->disks
[i
].replacement
;
2601 && !test_bit(Faulty
, &rdev
->flags
)
2602 && !test_bit(In_sync
, &rdev
->flags
)
2603 && !rdev_set_badblocks(rdev
, sh
->sector
,
2608 conf
->recovery_disabled
=
2609 conf
->mddev
->recovery_disabled
;
2611 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
2614 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2616 struct md_rdev
*rdev
;
2618 /* Doing recovery so rcu locking not required */
2619 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2621 && !test_bit(Faulty
, &rdev
->flags
)
2622 && !test_bit(In_sync
, &rdev
->flags
)
2623 && (rdev
->recovery_offset
<= sh
->sector
2624 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2630 /* fetch_block - checks the given member device to see if its data needs
2631 * to be read or computed to satisfy a request.
2633 * Returns 1 when no more member devices need to be checked, otherwise returns
2634 * 0 to tell the loop in handle_stripe_fill to continue
2636 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2637 int disk_idx
, int disks
)
2639 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2640 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2641 &sh
->dev
[s
->failed_num
[1]] };
2643 /* is the data in this block needed, and can we get it? */
2644 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2645 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2647 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2648 s
->syncing
|| s
->expanding
||
2649 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2650 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2651 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2652 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2653 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2654 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2655 /* we would like to get this block, possibly by computing it,
2656 * otherwise read it if the backing disk is insync
2658 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2659 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2660 if ((s
->uptodate
== disks
- 1) &&
2661 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2662 disk_idx
== s
->failed_num
[1]))) {
2663 /* have disk failed, and we're requested to fetch it;
2666 pr_debug("Computing stripe %llu block %d\n",
2667 (unsigned long long)sh
->sector
, disk_idx
);
2668 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2669 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2670 set_bit(R5_Wantcompute
, &dev
->flags
);
2671 sh
->ops
.target
= disk_idx
;
2672 sh
->ops
.target2
= -1; /* no 2nd target */
2674 /* Careful: from this point on 'uptodate' is in the eye
2675 * of raid_run_ops which services 'compute' operations
2676 * before writes. R5_Wantcompute flags a block that will
2677 * be R5_UPTODATE by the time it is needed for a
2678 * subsequent operation.
2682 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2683 /* Computing 2-failure is *very* expensive; only
2684 * do it if failed >= 2
2687 for (other
= disks
; other
--; ) {
2688 if (other
== disk_idx
)
2690 if (!test_bit(R5_UPTODATE
,
2691 &sh
->dev
[other
].flags
))
2695 pr_debug("Computing stripe %llu blocks %d,%d\n",
2696 (unsigned long long)sh
->sector
,
2698 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2699 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2700 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2701 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2702 sh
->ops
.target
= disk_idx
;
2703 sh
->ops
.target2
= other
;
2707 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2708 set_bit(R5_LOCKED
, &dev
->flags
);
2709 set_bit(R5_Wantread
, &dev
->flags
);
2711 pr_debug("Reading block %d (sync=%d)\n",
2712 disk_idx
, s
->syncing
);
2720 * handle_stripe_fill - read or compute data to satisfy pending requests.
2722 static void handle_stripe_fill(struct stripe_head
*sh
,
2723 struct stripe_head_state
*s
,
2728 /* look for blocks to read/compute, skip this if a compute
2729 * is already in flight, or if the stripe contents are in the
2730 * midst of changing due to a write
2732 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2733 !sh
->reconstruct_state
)
2734 for (i
= disks
; i
--; )
2735 if (fetch_block(sh
, s
, i
, disks
))
2737 set_bit(STRIPE_HANDLE
, &sh
->state
);
2741 /* handle_stripe_clean_event
2742 * any written block on an uptodate or failed drive can be returned.
2743 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2744 * never LOCKED, so we don't need to test 'failed' directly.
2746 static void handle_stripe_clean_event(struct r5conf
*conf
,
2747 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2751 int discard_pending
= 0;
2753 for (i
= disks
; i
--; )
2754 if (sh
->dev
[i
].written
) {
2756 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2757 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2758 test_bit(R5_Discard
, &dev
->flags
))) {
2759 /* We can return any write requests */
2760 struct bio
*wbi
, *wbi2
;
2761 pr_debug("Return write for disc %d\n", i
);
2762 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
2763 clear_bit(R5_UPTODATE
, &dev
->flags
);
2765 dev
->written
= NULL
;
2766 while (wbi
&& wbi
->bi_sector
<
2767 dev
->sector
+ STRIPE_SECTORS
) {
2768 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2769 if (!raid5_dec_bi_active_stripes(wbi
)) {
2770 md_write_end(conf
->mddev
);
2771 wbi
->bi_next
= *return_bi
;
2776 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2778 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2780 } else if (test_bit(R5_Discard
, &dev
->flags
))
2781 discard_pending
= 1;
2783 if (!discard_pending
&&
2784 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
2785 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
2786 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2787 if (sh
->qd_idx
>= 0) {
2788 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
2789 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
2791 /* now that discard is done we can proceed with any sync */
2792 clear_bit(STRIPE_DISCARD
, &sh
->state
);
2793 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
2794 set_bit(STRIPE_HANDLE
, &sh
->state
);
2798 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2799 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2800 md_wakeup_thread(conf
->mddev
->thread
);
2803 static void handle_stripe_dirtying(struct r5conf
*conf
,
2804 struct stripe_head
*sh
,
2805 struct stripe_head_state
*s
,
2808 int rmw
= 0, rcw
= 0, i
;
2809 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
2811 /* RAID6 requires 'rcw' in current implementation.
2812 * Otherwise, check whether resync is now happening or should start.
2813 * If yes, then the array is dirty (after unclean shutdown or
2814 * initial creation), so parity in some stripes might be inconsistent.
2815 * In this case, we need to always do reconstruct-write, to ensure
2816 * that in case of drive failure or read-error correction, we
2817 * generate correct data from the parity.
2819 if (conf
->max_degraded
== 2 ||
2820 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
)) {
2821 /* Calculate the real rcw later - for now make it
2822 * look like rcw is cheaper
2825 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2826 conf
->max_degraded
, (unsigned long long)recovery_cp
,
2827 (unsigned long long)sh
->sector
);
2828 } else for (i
= disks
; i
--; ) {
2829 /* would I have to read this buffer for read_modify_write */
2830 struct r5dev
*dev
= &sh
->dev
[i
];
2831 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2832 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2833 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2834 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2835 if (test_bit(R5_Insync
, &dev
->flags
))
2838 rmw
+= 2*disks
; /* cannot read it */
2840 /* Would I have to read this buffer for reconstruct_write */
2841 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2842 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2843 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2844 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2845 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2850 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2851 (unsigned long long)sh
->sector
, rmw
, rcw
);
2852 set_bit(STRIPE_HANDLE
, &sh
->state
);
2853 if (rmw
< rcw
&& rmw
> 0) {
2854 /* prefer read-modify-write, but need to get some data */
2855 if (conf
->mddev
->queue
)
2856 blk_add_trace_msg(conf
->mddev
->queue
,
2857 "raid5 rmw %llu %d",
2858 (unsigned long long)sh
->sector
, rmw
);
2859 for (i
= disks
; i
--; ) {
2860 struct r5dev
*dev
= &sh
->dev
[i
];
2861 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2862 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2863 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2864 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2865 test_bit(R5_Insync
, &dev
->flags
)) {
2867 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2868 pr_debug("Read_old block "
2869 "%d for r-m-w\n", i
);
2870 set_bit(R5_LOCKED
, &dev
->flags
);
2871 set_bit(R5_Wantread
, &dev
->flags
);
2874 set_bit(STRIPE_DELAYED
, &sh
->state
);
2875 set_bit(STRIPE_HANDLE
, &sh
->state
);
2880 if (rcw
<= rmw
&& rcw
> 0) {
2881 /* want reconstruct write, but need to get some data */
2884 for (i
= disks
; i
--; ) {
2885 struct r5dev
*dev
= &sh
->dev
[i
];
2886 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2887 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2888 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2889 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2890 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2892 if (!test_bit(R5_Insync
, &dev
->flags
))
2893 continue; /* it's a failed drive */
2895 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2896 pr_debug("Read_old block "
2897 "%d for Reconstruct\n", i
);
2898 set_bit(R5_LOCKED
, &dev
->flags
);
2899 set_bit(R5_Wantread
, &dev
->flags
);
2903 set_bit(STRIPE_DELAYED
, &sh
->state
);
2904 set_bit(STRIPE_HANDLE
, &sh
->state
);
2908 if (rcw
&& conf
->mddev
->queue
)
2909 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
2910 (unsigned long long)sh
->sector
,
2911 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
2913 /* now if nothing is locked, and if we have enough data,
2914 * we can start a write request
2916 /* since handle_stripe can be called at any time we need to handle the
2917 * case where a compute block operation has been submitted and then a
2918 * subsequent call wants to start a write request. raid_run_ops only
2919 * handles the case where compute block and reconstruct are requested
2920 * simultaneously. If this is not the case then new writes need to be
2921 * held off until the compute completes.
2923 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2924 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2925 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2926 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2929 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
2930 struct stripe_head_state
*s
, int disks
)
2932 struct r5dev
*dev
= NULL
;
2934 set_bit(STRIPE_HANDLE
, &sh
->state
);
2936 switch (sh
->check_state
) {
2937 case check_state_idle
:
2938 /* start a new check operation if there are no failures */
2939 if (s
->failed
== 0) {
2940 BUG_ON(s
->uptodate
!= disks
);
2941 sh
->check_state
= check_state_run
;
2942 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2943 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2947 dev
= &sh
->dev
[s
->failed_num
[0]];
2949 case check_state_compute_result
:
2950 sh
->check_state
= check_state_idle
;
2952 dev
= &sh
->dev
[sh
->pd_idx
];
2954 /* check that a write has not made the stripe insync */
2955 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2958 /* either failed parity check, or recovery is happening */
2959 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2960 BUG_ON(s
->uptodate
!= disks
);
2962 set_bit(R5_LOCKED
, &dev
->flags
);
2964 set_bit(R5_Wantwrite
, &dev
->flags
);
2966 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2967 set_bit(STRIPE_INSYNC
, &sh
->state
);
2969 case check_state_run
:
2970 break; /* we will be called again upon completion */
2971 case check_state_check_result
:
2972 sh
->check_state
= check_state_idle
;
2974 /* if a failure occurred during the check operation, leave
2975 * STRIPE_INSYNC not set and let the stripe be handled again
2980 /* handle a successful check operation, if parity is correct
2981 * we are done. Otherwise update the mismatch count and repair
2982 * parity if !MD_RECOVERY_CHECK
2984 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2985 /* parity is correct (on disc,
2986 * not in buffer any more)
2988 set_bit(STRIPE_INSYNC
, &sh
->state
);
2990 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
2991 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2992 /* don't try to repair!! */
2993 set_bit(STRIPE_INSYNC
, &sh
->state
);
2995 sh
->check_state
= check_state_compute_run
;
2996 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2997 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2998 set_bit(R5_Wantcompute
,
2999 &sh
->dev
[sh
->pd_idx
].flags
);
3000 sh
->ops
.target
= sh
->pd_idx
;
3001 sh
->ops
.target2
= -1;
3006 case check_state_compute_run
:
3009 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3010 __func__
, sh
->check_state
,
3011 (unsigned long long) sh
->sector
);
3017 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3018 struct stripe_head_state
*s
,
3021 int pd_idx
= sh
->pd_idx
;
3022 int qd_idx
= sh
->qd_idx
;
3025 set_bit(STRIPE_HANDLE
, &sh
->state
);
3027 BUG_ON(s
->failed
> 2);
3029 /* Want to check and possibly repair P and Q.
3030 * However there could be one 'failed' device, in which
3031 * case we can only check one of them, possibly using the
3032 * other to generate missing data
3035 switch (sh
->check_state
) {
3036 case check_state_idle
:
3037 /* start a new check operation if there are < 2 failures */
3038 if (s
->failed
== s
->q_failed
) {
3039 /* The only possible failed device holds Q, so it
3040 * makes sense to check P (If anything else were failed,
3041 * we would have used P to recreate it).
3043 sh
->check_state
= check_state_run
;
3045 if (!s
->q_failed
&& s
->failed
< 2) {
3046 /* Q is not failed, and we didn't use it to generate
3047 * anything, so it makes sense to check it
3049 if (sh
->check_state
== check_state_run
)
3050 sh
->check_state
= check_state_run_pq
;
3052 sh
->check_state
= check_state_run_q
;
3055 /* discard potentially stale zero_sum_result */
3056 sh
->ops
.zero_sum_result
= 0;
3058 if (sh
->check_state
== check_state_run
) {
3059 /* async_xor_zero_sum destroys the contents of P */
3060 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3063 if (sh
->check_state
>= check_state_run
&&
3064 sh
->check_state
<= check_state_run_pq
) {
3065 /* async_syndrome_zero_sum preserves P and Q, so
3066 * no need to mark them !uptodate here
3068 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3072 /* we have 2-disk failure */
3073 BUG_ON(s
->failed
!= 2);
3075 case check_state_compute_result
:
3076 sh
->check_state
= check_state_idle
;
3078 /* check that a write has not made the stripe insync */
3079 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3082 /* now write out any block on a failed drive,
3083 * or P or Q if they were recomputed
3085 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3086 if (s
->failed
== 2) {
3087 dev
= &sh
->dev
[s
->failed_num
[1]];
3089 set_bit(R5_LOCKED
, &dev
->flags
);
3090 set_bit(R5_Wantwrite
, &dev
->flags
);
3092 if (s
->failed
>= 1) {
3093 dev
= &sh
->dev
[s
->failed_num
[0]];
3095 set_bit(R5_LOCKED
, &dev
->flags
);
3096 set_bit(R5_Wantwrite
, &dev
->flags
);
3098 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3099 dev
= &sh
->dev
[pd_idx
];
3101 set_bit(R5_LOCKED
, &dev
->flags
);
3102 set_bit(R5_Wantwrite
, &dev
->flags
);
3104 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3105 dev
= &sh
->dev
[qd_idx
];
3107 set_bit(R5_LOCKED
, &dev
->flags
);
3108 set_bit(R5_Wantwrite
, &dev
->flags
);
3110 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3112 set_bit(STRIPE_INSYNC
, &sh
->state
);
3114 case check_state_run
:
3115 case check_state_run_q
:
3116 case check_state_run_pq
:
3117 break; /* we will be called again upon completion */
3118 case check_state_check_result
:
3119 sh
->check_state
= check_state_idle
;
3121 /* handle a successful check operation, if parity is correct
3122 * we are done. Otherwise update the mismatch count and repair
3123 * parity if !MD_RECOVERY_CHECK
3125 if (sh
->ops
.zero_sum_result
== 0) {
3126 /* both parities are correct */
3128 set_bit(STRIPE_INSYNC
, &sh
->state
);
3130 /* in contrast to the raid5 case we can validate
3131 * parity, but still have a failure to write
3134 sh
->check_state
= check_state_compute_result
;
3135 /* Returning at this point means that we may go
3136 * off and bring p and/or q uptodate again so
3137 * we make sure to check zero_sum_result again
3138 * to verify if p or q need writeback
3142 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3143 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3144 /* don't try to repair!! */
3145 set_bit(STRIPE_INSYNC
, &sh
->state
);
3147 int *target
= &sh
->ops
.target
;
3149 sh
->ops
.target
= -1;
3150 sh
->ops
.target2
= -1;
3151 sh
->check_state
= check_state_compute_run
;
3152 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3153 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3154 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3155 set_bit(R5_Wantcompute
,
3156 &sh
->dev
[pd_idx
].flags
);
3158 target
= &sh
->ops
.target2
;
3161 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3162 set_bit(R5_Wantcompute
,
3163 &sh
->dev
[qd_idx
].flags
);
3170 case check_state_compute_run
:
3173 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3174 __func__
, sh
->check_state
,
3175 (unsigned long long) sh
->sector
);
3180 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3184 /* We have read all the blocks in this stripe and now we need to
3185 * copy some of them into a target stripe for expand.
3187 struct dma_async_tx_descriptor
*tx
= NULL
;
3188 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3189 for (i
= 0; i
< sh
->disks
; i
++)
3190 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3192 struct stripe_head
*sh2
;
3193 struct async_submit_ctl submit
;
3195 sector_t bn
= compute_blocknr(sh
, i
, 1);
3196 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3198 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3200 /* so far only the early blocks of this stripe
3201 * have been requested. When later blocks
3202 * get requested, we will try again
3205 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3206 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3207 /* must have already done this block */
3208 release_stripe(sh2
);
3212 /* place all the copies on one channel */
3213 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3214 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3215 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3218 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3219 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3220 for (j
= 0; j
< conf
->raid_disks
; j
++)
3221 if (j
!= sh2
->pd_idx
&&
3223 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3225 if (j
== conf
->raid_disks
) {
3226 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3227 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3229 release_stripe(sh2
);
3232 /* done submitting copies, wait for them to complete */
3233 async_tx_quiesce(&tx
);
3237 * handle_stripe - do things to a stripe.
3239 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3240 * state of various bits to see what needs to be done.
3242 * return some read requests which now have data
3243 * return some write requests which are safely on storage
3244 * schedule a read on some buffers
3245 * schedule a write of some buffers
3246 * return confirmation of parity correctness
3250 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3252 struct r5conf
*conf
= sh
->raid_conf
;
3253 int disks
= sh
->disks
;
3256 int do_recovery
= 0;
3258 memset(s
, 0, sizeof(*s
));
3260 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3261 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3262 s
->failed_num
[0] = -1;
3263 s
->failed_num
[1] = -1;
3265 /* Now to look around and see what can be done */
3267 for (i
=disks
; i
--; ) {
3268 struct md_rdev
*rdev
;
3275 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3277 dev
->toread
, dev
->towrite
, dev
->written
);
3278 /* maybe we can reply to a read
3280 * new wantfill requests are only permitted while
3281 * ops_complete_biofill is guaranteed to be inactive
3283 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3284 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3285 set_bit(R5_Wantfill
, &dev
->flags
);
3287 /* now count some things */
3288 if (test_bit(R5_LOCKED
, &dev
->flags
))
3290 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3292 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3294 BUG_ON(s
->compute
> 2);
3297 if (test_bit(R5_Wantfill
, &dev
->flags
))
3299 else if (dev
->toread
)
3303 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3308 /* Prefer to use the replacement for reads, but only
3309 * if it is recovered enough and has no bad blocks.
3311 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3312 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3313 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3314 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3315 &first_bad
, &bad_sectors
))
3316 set_bit(R5_ReadRepl
, &dev
->flags
);
3319 set_bit(R5_NeedReplace
, &dev
->flags
);
3320 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3321 clear_bit(R5_ReadRepl
, &dev
->flags
);
3323 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3326 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3327 &first_bad
, &bad_sectors
);
3328 if (s
->blocked_rdev
== NULL
3329 && (test_bit(Blocked
, &rdev
->flags
)
3332 set_bit(BlockedBadBlocks
,
3334 s
->blocked_rdev
= rdev
;
3335 atomic_inc(&rdev
->nr_pending
);
3338 clear_bit(R5_Insync
, &dev
->flags
);
3342 /* also not in-sync */
3343 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
3344 test_bit(R5_UPTODATE
, &dev
->flags
)) {
3345 /* treat as in-sync, but with a read error
3346 * which we can now try to correct
3348 set_bit(R5_Insync
, &dev
->flags
);
3349 set_bit(R5_ReadError
, &dev
->flags
);
3351 } else if (test_bit(In_sync
, &rdev
->flags
))
3352 set_bit(R5_Insync
, &dev
->flags
);
3353 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3354 /* in sync if before recovery_offset */
3355 set_bit(R5_Insync
, &dev
->flags
);
3356 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3357 test_bit(R5_Expanded
, &dev
->flags
))
3358 /* If we've reshaped into here, we assume it is Insync.
3359 * We will shortly update recovery_offset to make
3362 set_bit(R5_Insync
, &dev
->flags
);
3364 if (rdev
&& test_bit(R5_WriteError
, &dev
->flags
)) {
3365 /* This flag does not apply to '.replacement'
3366 * only to .rdev, so make sure to check that*/
3367 struct md_rdev
*rdev2
= rcu_dereference(
3368 conf
->disks
[i
].rdev
);
3370 clear_bit(R5_Insync
, &dev
->flags
);
3371 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3372 s
->handle_bad_blocks
= 1;
3373 atomic_inc(&rdev2
->nr_pending
);
3375 clear_bit(R5_WriteError
, &dev
->flags
);
3377 if (rdev
&& test_bit(R5_MadeGood
, &dev
->flags
)) {
3378 /* This flag does not apply to '.replacement'
3379 * only to .rdev, so make sure to check that*/
3380 struct md_rdev
*rdev2
= rcu_dereference(
3381 conf
->disks
[i
].rdev
);
3382 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3383 s
->handle_bad_blocks
= 1;
3384 atomic_inc(&rdev2
->nr_pending
);
3386 clear_bit(R5_MadeGood
, &dev
->flags
);
3388 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3389 struct md_rdev
*rdev2
= rcu_dereference(
3390 conf
->disks
[i
].replacement
);
3391 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3392 s
->handle_bad_blocks
= 1;
3393 atomic_inc(&rdev2
->nr_pending
);
3395 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3397 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3398 /* The ReadError flag will just be confusing now */
3399 clear_bit(R5_ReadError
, &dev
->flags
);
3400 clear_bit(R5_ReWrite
, &dev
->flags
);
3402 if (test_bit(R5_ReadError
, &dev
->flags
))
3403 clear_bit(R5_Insync
, &dev
->flags
);
3404 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3406 s
->failed_num
[s
->failed
] = i
;
3408 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3412 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3413 /* If there is a failed device being replaced,
3414 * we must be recovering.
3415 * else if we are after recovery_cp, we must be syncing
3416 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3417 * else we can only be replacing
3418 * sync and recovery both need to read all devices, and so
3419 * use the same flag.
3422 sh
->sector
>= conf
->mddev
->recovery_cp
||
3423 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
3431 static void handle_stripe(struct stripe_head
*sh
)
3433 struct stripe_head_state s
;
3434 struct r5conf
*conf
= sh
->raid_conf
;
3437 int disks
= sh
->disks
;
3438 struct r5dev
*pdev
, *qdev
;
3440 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3441 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3442 /* already being handled, ensure it gets handled
3443 * again when current action finishes */
3444 set_bit(STRIPE_HANDLE
, &sh
->state
);
3448 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3449 spin_lock(&sh
->stripe_lock
);
3450 /* Cannot process 'sync' concurrently with 'discard' */
3451 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
3452 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3453 set_bit(STRIPE_SYNCING
, &sh
->state
);
3454 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3456 spin_unlock(&sh
->stripe_lock
);
3458 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3460 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3461 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3462 (unsigned long long)sh
->sector
, sh
->state
,
3463 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3464 sh
->check_state
, sh
->reconstruct_state
);
3466 analyse_stripe(sh
, &s
);
3468 if (s
.handle_bad_blocks
) {
3469 set_bit(STRIPE_HANDLE
, &sh
->state
);
3473 if (unlikely(s
.blocked_rdev
)) {
3474 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3475 s
.replacing
|| s
.to_write
|| s
.written
) {
3476 set_bit(STRIPE_HANDLE
, &sh
->state
);
3479 /* There is nothing for the blocked_rdev to block */
3480 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3481 s
.blocked_rdev
= NULL
;
3484 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3485 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3486 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3489 pr_debug("locked=%d uptodate=%d to_read=%d"
3490 " to_write=%d failed=%d failed_num=%d,%d\n",
3491 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3492 s
.failed_num
[0], s
.failed_num
[1]);
3493 /* check if the array has lost more than max_degraded devices and,
3494 * if so, some requests might need to be failed.
3496 if (s
.failed
> conf
->max_degraded
) {
3497 sh
->check_state
= 0;
3498 sh
->reconstruct_state
= 0;
3499 if (s
.to_read
+s
.to_write
+s
.written
)
3500 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3501 if (s
.syncing
+ s
.replacing
)
3502 handle_failed_sync(conf
, sh
, &s
);
3505 /* Now we check to see if any write operations have recently
3509 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3511 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3512 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3513 sh
->reconstruct_state
= reconstruct_state_idle
;
3515 /* All the 'written' buffers and the parity block are ready to
3516 * be written back to disk
3518 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
3519 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
3520 BUG_ON(sh
->qd_idx
>= 0 &&
3521 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
3522 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
3523 for (i
= disks
; i
--; ) {
3524 struct r5dev
*dev
= &sh
->dev
[i
];
3525 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3526 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3528 pr_debug("Writing block %d\n", i
);
3529 set_bit(R5_Wantwrite
, &dev
->flags
);
3532 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3533 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3535 set_bit(STRIPE_INSYNC
, &sh
->state
);
3538 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3539 s
.dec_preread_active
= 1;
3543 * might be able to return some write requests if the parity blocks
3544 * are safe, or on a failed drive
3546 pdev
= &sh
->dev
[sh
->pd_idx
];
3547 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3548 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3549 qdev
= &sh
->dev
[sh
->qd_idx
];
3550 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3551 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3555 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3556 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3557 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
3558 test_bit(R5_Discard
, &pdev
->flags
))))) &&
3559 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3560 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3561 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
3562 test_bit(R5_Discard
, &qdev
->flags
))))))
3563 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3565 /* Now we might consider reading some blocks, either to check/generate
3566 * parity, or to satisfy requests
3567 * or to load a block that is being partially written.
3569 if (s
.to_read
|| s
.non_overwrite
3570 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3571 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3574 handle_stripe_fill(sh
, &s
, disks
);
3576 /* Now to consider new write requests and what else, if anything
3577 * should be read. We do not handle new writes when:
3578 * 1/ A 'write' operation (copy+xor) is already in flight.
3579 * 2/ A 'check' operation is in flight, as it may clobber the parity
3582 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3583 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3585 /* maybe we need to check and possibly fix the parity for this stripe
3586 * Any reads will already have been scheduled, so we just see if enough
3587 * data is available. The parity check is held off while parity
3588 * dependent operations are in flight.
3590 if (sh
->check_state
||
3591 (s
.syncing
&& s
.locked
== 0 &&
3592 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3593 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3594 if (conf
->level
== 6)
3595 handle_parity_checks6(conf
, sh
, &s
, disks
);
3597 handle_parity_checks5(conf
, sh
, &s
, disks
);
3600 if (s
.replacing
&& s
.locked
== 0
3601 && !test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3602 /* Write out to replacement devices where possible */
3603 for (i
= 0; i
< conf
->raid_disks
; i
++)
3604 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
) &&
3605 test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3606 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3607 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3610 set_bit(STRIPE_INSYNC
, &sh
->state
);
3612 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3613 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3614 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3615 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3616 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3617 wake_up(&conf
->wait_for_overlap
);
3620 /* If the failed drives are just a ReadError, then we might need
3621 * to progress the repair/check process
3623 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3624 for (i
= 0; i
< s
.failed
; i
++) {
3625 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3626 if (test_bit(R5_ReadError
, &dev
->flags
)
3627 && !test_bit(R5_LOCKED
, &dev
->flags
)
3628 && test_bit(R5_UPTODATE
, &dev
->flags
)
3630 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3631 set_bit(R5_Wantwrite
, &dev
->flags
);
3632 set_bit(R5_ReWrite
, &dev
->flags
);
3633 set_bit(R5_LOCKED
, &dev
->flags
);
3636 /* let's read it back */
3637 set_bit(R5_Wantread
, &dev
->flags
);
3638 set_bit(R5_LOCKED
, &dev
->flags
);
3645 /* Finish reconstruct operations initiated by the expansion process */
3646 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3647 struct stripe_head
*sh_src
3648 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3649 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3650 /* sh cannot be written until sh_src has been read.
3651 * so arrange for sh to be delayed a little
3653 set_bit(STRIPE_DELAYED
, &sh
->state
);
3654 set_bit(STRIPE_HANDLE
, &sh
->state
);
3655 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3657 atomic_inc(&conf
->preread_active_stripes
);
3658 release_stripe(sh_src
);
3662 release_stripe(sh_src
);
3664 sh
->reconstruct_state
= reconstruct_state_idle
;
3665 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3666 for (i
= conf
->raid_disks
; i
--; ) {
3667 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3668 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3673 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3674 !sh
->reconstruct_state
) {
3675 /* Need to write out all blocks after computing parity */
3676 sh
->disks
= conf
->raid_disks
;
3677 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3678 schedule_reconstruction(sh
, &s
, 1, 1);
3679 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3680 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3681 atomic_dec(&conf
->reshape_stripes
);
3682 wake_up(&conf
->wait_for_overlap
);
3683 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3686 if (s
.expanding
&& s
.locked
== 0 &&
3687 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3688 handle_stripe_expansion(conf
, sh
);
3691 /* wait for this device to become unblocked */
3692 if (unlikely(s
.blocked_rdev
)) {
3693 if (conf
->mddev
->external
)
3694 md_wait_for_blocked_rdev(s
.blocked_rdev
,
3697 /* Internal metadata will immediately
3698 * be written by raid5d, so we don't
3699 * need to wait here.
3701 rdev_dec_pending(s
.blocked_rdev
,
3705 if (s
.handle_bad_blocks
)
3706 for (i
= disks
; i
--; ) {
3707 struct md_rdev
*rdev
;
3708 struct r5dev
*dev
= &sh
->dev
[i
];
3709 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3710 /* We own a safe reference to the rdev */
3711 rdev
= conf
->disks
[i
].rdev
;
3712 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3714 md_error(conf
->mddev
, rdev
);
3715 rdev_dec_pending(rdev
, conf
->mddev
);
3717 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3718 rdev
= conf
->disks
[i
].rdev
;
3719 rdev_clear_badblocks(rdev
, sh
->sector
,
3721 rdev_dec_pending(rdev
, conf
->mddev
);
3723 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3724 rdev
= conf
->disks
[i
].replacement
;
3726 /* rdev have been moved down */
3727 rdev
= conf
->disks
[i
].rdev
;
3728 rdev_clear_badblocks(rdev
, sh
->sector
,
3730 rdev_dec_pending(rdev
, conf
->mddev
);
3735 raid_run_ops(sh
, s
.ops_request
);
3739 if (s
.dec_preread_active
) {
3740 /* We delay this until after ops_run_io so that if make_request
3741 * is waiting on a flush, it won't continue until the writes
3742 * have actually been submitted.
3744 atomic_dec(&conf
->preread_active_stripes
);
3745 if (atomic_read(&conf
->preread_active_stripes
) <
3747 md_wakeup_thread(conf
->mddev
->thread
);
3750 return_io(s
.return_bi
);
3752 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3755 static void raid5_activate_delayed(struct r5conf
*conf
)
3757 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3758 while (!list_empty(&conf
->delayed_list
)) {
3759 struct list_head
*l
= conf
->delayed_list
.next
;
3760 struct stripe_head
*sh
;
3761 sh
= list_entry(l
, struct stripe_head
, lru
);
3763 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3764 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3765 atomic_inc(&conf
->preread_active_stripes
);
3766 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3771 static void activate_bit_delay(struct r5conf
*conf
)
3773 /* device_lock is held */
3774 struct list_head head
;
3775 list_add(&head
, &conf
->bitmap_list
);
3776 list_del_init(&conf
->bitmap_list
);
3777 while (!list_empty(&head
)) {
3778 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3779 list_del_init(&sh
->lru
);
3780 atomic_inc(&sh
->count
);
3781 __release_stripe(conf
, sh
);
3785 int md_raid5_congested(struct mddev
*mddev
, int bits
)
3787 struct r5conf
*conf
= mddev
->private;
3789 /* No difference between reads and writes. Just check
3790 * how busy the stripe_cache is
3793 if (conf
->inactive_blocked
)
3797 if (list_empty_careful(&conf
->inactive_list
))
3802 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3804 static int raid5_congested(void *data
, int bits
)
3806 struct mddev
*mddev
= data
;
3808 return mddev_congested(mddev
, bits
) ||
3809 md_raid5_congested(mddev
, bits
);
3812 /* We want read requests to align with chunks where possible,
3813 * but write requests don't need to.
3815 static int raid5_mergeable_bvec(struct request_queue
*q
,
3816 struct bvec_merge_data
*bvm
,
3817 struct bio_vec
*biovec
)
3819 struct mddev
*mddev
= q
->queuedata
;
3820 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3822 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3823 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3825 if ((bvm
->bi_rw
& 1) == WRITE
)
3826 return biovec
->bv_len
; /* always allow writes to be mergeable */
3828 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3829 chunk_sectors
= mddev
->new_chunk_sectors
;
3830 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3831 if (max
< 0) max
= 0;
3832 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3833 return biovec
->bv_len
;
3839 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
3841 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3842 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3843 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3845 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3846 chunk_sectors
= mddev
->new_chunk_sectors
;
3847 return chunk_sectors
>=
3848 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3852 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3853 * later sampled by raid5d.
3855 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
3857 unsigned long flags
;
3859 spin_lock_irqsave(&conf
->device_lock
, flags
);
3861 bi
->bi_next
= conf
->retry_read_aligned_list
;
3862 conf
->retry_read_aligned_list
= bi
;
3864 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3865 md_wakeup_thread(conf
->mddev
->thread
);
3869 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
3873 bi
= conf
->retry_read_aligned
;
3875 conf
->retry_read_aligned
= NULL
;
3878 bi
= conf
->retry_read_aligned_list
;
3880 conf
->retry_read_aligned_list
= bi
->bi_next
;
3883 * this sets the active strip count to 1 and the processed
3884 * strip count to zero (upper 8 bits)
3886 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
3894 * The "raid5_align_endio" should check if the read succeeded and if it
3895 * did, call bio_endio on the original bio (having bio_put the new bio
3897 * If the read failed..
3899 static void raid5_align_endio(struct bio
*bi
, int error
)
3901 struct bio
* raid_bi
= bi
->bi_private
;
3902 struct mddev
*mddev
;
3903 struct r5conf
*conf
;
3904 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3905 struct md_rdev
*rdev
;
3909 rdev
= (void*)raid_bi
->bi_next
;
3910 raid_bi
->bi_next
= NULL
;
3911 mddev
= rdev
->mddev
;
3912 conf
= mddev
->private;
3914 rdev_dec_pending(rdev
, conf
->mddev
);
3916 if (!error
&& uptodate
) {
3917 bio_endio(raid_bi
, 0);
3918 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3919 wake_up(&conf
->wait_for_stripe
);
3924 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3926 add_bio_to_retry(raid_bi
, conf
);
3929 static int bio_fits_rdev(struct bio
*bi
)
3931 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3933 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3935 blk_recount_segments(q
, bi
);
3936 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3939 if (q
->merge_bvec_fn
)
3940 /* it's too hard to apply the merge_bvec_fn at this stage,
3949 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
3951 struct r5conf
*conf
= mddev
->private;
3953 struct bio
* align_bi
;
3954 struct md_rdev
*rdev
;
3955 sector_t end_sector
;
3957 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3958 pr_debug("chunk_aligned_read : non aligned\n");
3962 * use bio_clone_mddev to make a copy of the bio
3964 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3968 * set bi_end_io to a new function, and set bi_private to the
3971 align_bi
->bi_end_io
= raid5_align_endio
;
3972 align_bi
->bi_private
= raid_bio
;
3976 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3980 end_sector
= align_bi
->bi_sector
+ (align_bi
->bi_size
>> 9);
3982 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
3983 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
3984 rdev
->recovery_offset
< end_sector
) {
3985 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3987 (test_bit(Faulty
, &rdev
->flags
) ||
3988 !(test_bit(In_sync
, &rdev
->flags
) ||
3989 rdev
->recovery_offset
>= end_sector
)))
3996 atomic_inc(&rdev
->nr_pending
);
3998 raid_bio
->bi_next
= (void*)rdev
;
3999 align_bi
->bi_bdev
= rdev
->bdev
;
4000 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
4002 if (!bio_fits_rdev(align_bi
) ||
4003 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
4004 &first_bad
, &bad_sectors
)) {
4005 /* too big in some way, or has a known bad block */
4007 rdev_dec_pending(rdev
, mddev
);
4011 /* No reshape active, so we can trust rdev->data_offset */
4012 align_bi
->bi_sector
+= rdev
->data_offset
;
4014 spin_lock_irq(&conf
->device_lock
);
4015 wait_event_lock_irq(conf
->wait_for_stripe
,
4018 atomic_inc(&conf
->active_aligned_reads
);
4019 spin_unlock_irq(&conf
->device_lock
);
4022 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4023 align_bi
, disk_devt(mddev
->gendisk
),
4024 raid_bio
->bi_sector
);
4025 generic_make_request(align_bi
);
4034 /* __get_priority_stripe - get the next stripe to process
4036 * Full stripe writes are allowed to pass preread active stripes up until
4037 * the bypass_threshold is exceeded. In general the bypass_count
4038 * increments when the handle_list is handled before the hold_list; however, it
4039 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4040 * stripe with in flight i/o. The bypass_count will be reset when the
4041 * head of the hold_list has changed, i.e. the head was promoted to the
4044 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
)
4046 struct stripe_head
*sh
;
4048 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4050 list_empty(&conf
->handle_list
) ? "empty" : "busy",
4051 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4052 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4054 if (!list_empty(&conf
->handle_list
)) {
4055 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
4057 if (list_empty(&conf
->hold_list
))
4058 conf
->bypass_count
= 0;
4059 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4060 if (conf
->hold_list
.next
== conf
->last_hold
)
4061 conf
->bypass_count
++;
4063 conf
->last_hold
= conf
->hold_list
.next
;
4064 conf
->bypass_count
-= conf
->bypass_threshold
;
4065 if (conf
->bypass_count
< 0)
4066 conf
->bypass_count
= 0;
4069 } else if (!list_empty(&conf
->hold_list
) &&
4070 ((conf
->bypass_threshold
&&
4071 conf
->bypass_count
> conf
->bypass_threshold
) ||
4072 atomic_read(&conf
->pending_full_writes
) == 0)) {
4073 sh
= list_entry(conf
->hold_list
.next
,
4075 conf
->bypass_count
-= conf
->bypass_threshold
;
4076 if (conf
->bypass_count
< 0)
4077 conf
->bypass_count
= 0;
4081 list_del_init(&sh
->lru
);
4082 atomic_inc(&sh
->count
);
4083 BUG_ON(atomic_read(&sh
->count
) != 1);
4087 struct raid5_plug_cb
{
4088 struct blk_plug_cb cb
;
4089 struct list_head list
;
4092 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4094 struct raid5_plug_cb
*cb
= container_of(
4095 blk_cb
, struct raid5_plug_cb
, cb
);
4096 struct stripe_head
*sh
;
4097 struct mddev
*mddev
= cb
->cb
.data
;
4098 struct r5conf
*conf
= mddev
->private;
4101 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4102 spin_lock_irq(&conf
->device_lock
);
4103 while (!list_empty(&cb
->list
)) {
4104 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4105 list_del_init(&sh
->lru
);
4107 * avoid race release_stripe_plug() sees
4108 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4109 * is still in our list
4111 smp_mb__before_clear_bit();
4112 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4113 __release_stripe(conf
, sh
);
4116 spin_unlock_irq(&conf
->device_lock
);
4119 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
4123 static void release_stripe_plug(struct mddev
*mddev
,
4124 struct stripe_head
*sh
)
4126 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4127 raid5_unplug
, mddev
,
4128 sizeof(struct raid5_plug_cb
));
4129 struct raid5_plug_cb
*cb
;
4136 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4138 if (cb
->list
.next
== NULL
)
4139 INIT_LIST_HEAD(&cb
->list
);
4141 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
4142 list_add_tail(&sh
->lru
, &cb
->list
);
4147 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
4149 struct r5conf
*conf
= mddev
->private;
4150 sector_t logical_sector
, last_sector
;
4151 struct stripe_head
*sh
;
4155 if (mddev
->reshape_position
!= MaxSector
)
4156 /* Skip discard while reshape is happening */
4159 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4160 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4163 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4165 stripe_sectors
= conf
->chunk_sectors
*
4166 (conf
->raid_disks
- conf
->max_degraded
);
4167 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
4169 sector_div(last_sector
, stripe_sectors
);
4171 logical_sector
*= conf
->chunk_sectors
;
4172 last_sector
*= conf
->chunk_sectors
;
4174 for (; logical_sector
< last_sector
;
4175 logical_sector
+= STRIPE_SECTORS
) {
4179 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
4180 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
4181 TASK_UNINTERRUPTIBLE
);
4182 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4183 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4188 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4189 spin_lock_irq(&sh
->stripe_lock
);
4190 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4191 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4193 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
4194 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
4195 spin_unlock_irq(&sh
->stripe_lock
);
4201 set_bit(STRIPE_DISCARD
, &sh
->state
);
4202 finish_wait(&conf
->wait_for_overlap
, &w
);
4203 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4204 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4206 sh
->dev
[d
].towrite
= bi
;
4207 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
4208 raid5_inc_bi_active_stripes(bi
);
4210 spin_unlock_irq(&sh
->stripe_lock
);
4211 if (conf
->mddev
->bitmap
) {
4213 d
< conf
->raid_disks
- conf
->max_degraded
;
4215 bitmap_startwrite(mddev
->bitmap
,
4219 sh
->bm_seq
= conf
->seq_flush
+ 1;
4220 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
4223 set_bit(STRIPE_HANDLE
, &sh
->state
);
4224 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4225 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4226 atomic_inc(&conf
->preread_active_stripes
);
4227 release_stripe_plug(mddev
, sh
);
4230 remaining
= raid5_dec_bi_active_stripes(bi
);
4231 if (remaining
== 0) {
4232 md_write_end(mddev
);
4237 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
4239 struct r5conf
*conf
= mddev
->private;
4241 sector_t new_sector
;
4242 sector_t logical_sector
, last_sector
;
4243 struct stripe_head
*sh
;
4244 const int rw
= bio_data_dir(bi
);
4247 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
4248 md_flush_request(mddev
, bi
);
4252 md_write_start(mddev
, bi
);
4255 mddev
->reshape_position
== MaxSector
&&
4256 chunk_aligned_read(mddev
,bi
))
4259 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
4260 make_discard_request(mddev
, bi
);
4264 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4265 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4267 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4269 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4275 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4276 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4277 /* spinlock is needed as reshape_progress may be
4278 * 64bit on a 32bit platform, and so it might be
4279 * possible to see a half-updated value
4280 * Of course reshape_progress could change after
4281 * the lock is dropped, so once we get a reference
4282 * to the stripe that we think it is, we will have
4285 spin_lock_irq(&conf
->device_lock
);
4286 if (mddev
->reshape_backwards
4287 ? logical_sector
< conf
->reshape_progress
4288 : logical_sector
>= conf
->reshape_progress
) {
4291 if (mddev
->reshape_backwards
4292 ? logical_sector
< conf
->reshape_safe
4293 : logical_sector
>= conf
->reshape_safe
) {
4294 spin_unlock_irq(&conf
->device_lock
);
4299 spin_unlock_irq(&conf
->device_lock
);
4302 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4305 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4306 (unsigned long long)new_sector
,
4307 (unsigned long long)logical_sector
);
4309 sh
= get_active_stripe(conf
, new_sector
, previous
,
4310 (bi
->bi_rw
&RWA_MASK
), 0);
4312 if (unlikely(previous
)) {
4313 /* expansion might have moved on while waiting for a
4314 * stripe, so we must do the range check again.
4315 * Expansion could still move past after this
4316 * test, but as we are holding a reference to
4317 * 'sh', we know that if that happens,
4318 * STRIPE_EXPANDING will get set and the expansion
4319 * won't proceed until we finish with the stripe.
4322 spin_lock_irq(&conf
->device_lock
);
4323 if (mddev
->reshape_backwards
4324 ? logical_sector
>= conf
->reshape_progress
4325 : logical_sector
< conf
->reshape_progress
)
4326 /* mismatch, need to try again */
4328 spin_unlock_irq(&conf
->device_lock
);
4337 logical_sector
>= mddev
->suspend_lo
&&
4338 logical_sector
< mddev
->suspend_hi
) {
4340 /* As the suspend_* range is controlled by
4341 * userspace, we want an interruptible
4344 flush_signals(current
);
4345 prepare_to_wait(&conf
->wait_for_overlap
,
4346 &w
, TASK_INTERRUPTIBLE
);
4347 if (logical_sector
>= mddev
->suspend_lo
&&
4348 logical_sector
< mddev
->suspend_hi
)
4353 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4354 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4355 /* Stripe is busy expanding or
4356 * add failed due to overlap. Flush everything
4359 md_wakeup_thread(mddev
->thread
);
4364 finish_wait(&conf
->wait_for_overlap
, &w
);
4365 set_bit(STRIPE_HANDLE
, &sh
->state
);
4366 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4367 if ((bi
->bi_rw
& REQ_SYNC
) &&
4368 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4369 atomic_inc(&conf
->preread_active_stripes
);
4370 release_stripe_plug(mddev
, sh
);
4372 /* cannot get stripe for read-ahead, just give-up */
4373 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4374 finish_wait(&conf
->wait_for_overlap
, &w
);
4379 remaining
= raid5_dec_bi_active_stripes(bi
);
4380 if (remaining
== 0) {
4383 md_write_end(mddev
);
4389 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4391 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4393 /* reshaping is quite different to recovery/resync so it is
4394 * handled quite separately ... here.
4396 * On each call to sync_request, we gather one chunk worth of
4397 * destination stripes and flag them as expanding.
4398 * Then we find all the source stripes and request reads.
4399 * As the reads complete, handle_stripe will copy the data
4400 * into the destination stripe and release that stripe.
4402 struct r5conf
*conf
= mddev
->private;
4403 struct stripe_head
*sh
;
4404 sector_t first_sector
, last_sector
;
4405 int raid_disks
= conf
->previous_raid_disks
;
4406 int data_disks
= raid_disks
- conf
->max_degraded
;
4407 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4410 sector_t writepos
, readpos
, safepos
;
4411 sector_t stripe_addr
;
4412 int reshape_sectors
;
4413 struct list_head stripes
;
4415 if (sector_nr
== 0) {
4416 /* If restarting in the middle, skip the initial sectors */
4417 if (mddev
->reshape_backwards
&&
4418 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4419 sector_nr
= raid5_size(mddev
, 0, 0)
4420 - conf
->reshape_progress
;
4421 } else if (!mddev
->reshape_backwards
&&
4422 conf
->reshape_progress
> 0)
4423 sector_nr
= conf
->reshape_progress
;
4424 sector_div(sector_nr
, new_data_disks
);
4426 mddev
->curr_resync_completed
= sector_nr
;
4427 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4433 /* We need to process a full chunk at a time.
4434 * If old and new chunk sizes differ, we need to process the
4437 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4438 reshape_sectors
= mddev
->new_chunk_sectors
;
4440 reshape_sectors
= mddev
->chunk_sectors
;
4442 /* We update the metadata at least every 10 seconds, or when
4443 * the data about to be copied would over-write the source of
4444 * the data at the front of the range. i.e. one new_stripe
4445 * along from reshape_progress new_maps to after where
4446 * reshape_safe old_maps to
4448 writepos
= conf
->reshape_progress
;
4449 sector_div(writepos
, new_data_disks
);
4450 readpos
= conf
->reshape_progress
;
4451 sector_div(readpos
, data_disks
);
4452 safepos
= conf
->reshape_safe
;
4453 sector_div(safepos
, data_disks
);
4454 if (mddev
->reshape_backwards
) {
4455 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4456 readpos
+= reshape_sectors
;
4457 safepos
+= reshape_sectors
;
4459 writepos
+= reshape_sectors
;
4460 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4461 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4464 /* Having calculated the 'writepos' possibly use it
4465 * to set 'stripe_addr' which is where we will write to.
4467 if (mddev
->reshape_backwards
) {
4468 BUG_ON(conf
->reshape_progress
== 0);
4469 stripe_addr
= writepos
;
4470 BUG_ON((mddev
->dev_sectors
&
4471 ~((sector_t
)reshape_sectors
- 1))
4472 - reshape_sectors
- stripe_addr
4475 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4476 stripe_addr
= sector_nr
;
4479 /* 'writepos' is the most advanced device address we might write.
4480 * 'readpos' is the least advanced device address we might read.
4481 * 'safepos' is the least address recorded in the metadata as having
4483 * If there is a min_offset_diff, these are adjusted either by
4484 * increasing the safepos/readpos if diff is negative, or
4485 * increasing writepos if diff is positive.
4486 * If 'readpos' is then behind 'writepos', there is no way that we can
4487 * ensure safety in the face of a crash - that must be done by userspace
4488 * making a backup of the data. So in that case there is no particular
4489 * rush to update metadata.
4490 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4491 * update the metadata to advance 'safepos' to match 'readpos' so that
4492 * we can be safe in the event of a crash.
4493 * So we insist on updating metadata if safepos is behind writepos and
4494 * readpos is beyond writepos.
4495 * In any case, update the metadata every 10 seconds.
4496 * Maybe that number should be configurable, but I'm not sure it is
4497 * worth it.... maybe it could be a multiple of safemode_delay???
4499 if (conf
->min_offset_diff
< 0) {
4500 safepos
+= -conf
->min_offset_diff
;
4501 readpos
+= -conf
->min_offset_diff
;
4503 writepos
+= conf
->min_offset_diff
;
4505 if ((mddev
->reshape_backwards
4506 ? (safepos
> writepos
&& readpos
< writepos
)
4507 : (safepos
< writepos
&& readpos
> writepos
)) ||
4508 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4509 /* Cannot proceed until we've updated the superblock... */
4510 wait_event(conf
->wait_for_overlap
,
4511 atomic_read(&conf
->reshape_stripes
)==0);
4512 mddev
->reshape_position
= conf
->reshape_progress
;
4513 mddev
->curr_resync_completed
= sector_nr
;
4514 conf
->reshape_checkpoint
= jiffies
;
4515 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4516 md_wakeup_thread(mddev
->thread
);
4517 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4518 kthread_should_stop());
4519 spin_lock_irq(&conf
->device_lock
);
4520 conf
->reshape_safe
= mddev
->reshape_position
;
4521 spin_unlock_irq(&conf
->device_lock
);
4522 wake_up(&conf
->wait_for_overlap
);
4523 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4526 INIT_LIST_HEAD(&stripes
);
4527 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4529 int skipped_disk
= 0;
4530 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4531 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4532 atomic_inc(&conf
->reshape_stripes
);
4533 /* If any of this stripe is beyond the end of the old
4534 * array, then we need to zero those blocks
4536 for (j
=sh
->disks
; j
--;) {
4538 if (j
== sh
->pd_idx
)
4540 if (conf
->level
== 6 &&
4543 s
= compute_blocknr(sh
, j
, 0);
4544 if (s
< raid5_size(mddev
, 0, 0)) {
4548 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4549 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4550 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4552 if (!skipped_disk
) {
4553 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4554 set_bit(STRIPE_HANDLE
, &sh
->state
);
4556 list_add(&sh
->lru
, &stripes
);
4558 spin_lock_irq(&conf
->device_lock
);
4559 if (mddev
->reshape_backwards
)
4560 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4562 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4563 spin_unlock_irq(&conf
->device_lock
);
4564 /* Ok, those stripe are ready. We can start scheduling
4565 * reads on the source stripes.
4566 * The source stripes are determined by mapping the first and last
4567 * block on the destination stripes.
4570 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4573 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4574 * new_data_disks
- 1),
4576 if (last_sector
>= mddev
->dev_sectors
)
4577 last_sector
= mddev
->dev_sectors
- 1;
4578 while (first_sector
<= last_sector
) {
4579 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4580 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4581 set_bit(STRIPE_HANDLE
, &sh
->state
);
4583 first_sector
+= STRIPE_SECTORS
;
4585 /* Now that the sources are clearly marked, we can release
4586 * the destination stripes
4588 while (!list_empty(&stripes
)) {
4589 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4590 list_del_init(&sh
->lru
);
4593 /* If this takes us to the resync_max point where we have to pause,
4594 * then we need to write out the superblock.
4596 sector_nr
+= reshape_sectors
;
4597 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4598 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4599 /* Cannot proceed until we've updated the superblock... */
4600 wait_event(conf
->wait_for_overlap
,
4601 atomic_read(&conf
->reshape_stripes
) == 0);
4602 mddev
->reshape_position
= conf
->reshape_progress
;
4603 mddev
->curr_resync_completed
= sector_nr
;
4604 conf
->reshape_checkpoint
= jiffies
;
4605 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4606 md_wakeup_thread(mddev
->thread
);
4607 wait_event(mddev
->sb_wait
,
4608 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4609 || kthread_should_stop());
4610 spin_lock_irq(&conf
->device_lock
);
4611 conf
->reshape_safe
= mddev
->reshape_position
;
4612 spin_unlock_irq(&conf
->device_lock
);
4613 wake_up(&conf
->wait_for_overlap
);
4614 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4616 return reshape_sectors
;
4619 /* FIXME go_faster isn't used */
4620 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4622 struct r5conf
*conf
= mddev
->private;
4623 struct stripe_head
*sh
;
4624 sector_t max_sector
= mddev
->dev_sectors
;
4625 sector_t sync_blocks
;
4626 int still_degraded
= 0;
4629 if (sector_nr
>= max_sector
) {
4630 /* just being told to finish up .. nothing much to do */
4632 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4637 if (mddev
->curr_resync
< max_sector
) /* aborted */
4638 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4640 else /* completed sync */
4642 bitmap_close_sync(mddev
->bitmap
);
4647 /* Allow raid5_quiesce to complete */
4648 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4650 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4651 return reshape_request(mddev
, sector_nr
, skipped
);
4653 /* No need to check resync_max as we never do more than one
4654 * stripe, and as resync_max will always be on a chunk boundary,
4655 * if the check in md_do_sync didn't fire, there is no chance
4656 * of overstepping resync_max here
4659 /* if there is too many failed drives and we are trying
4660 * to resync, then assert that we are finished, because there is
4661 * nothing we can do.
4663 if (mddev
->degraded
>= conf
->max_degraded
&&
4664 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4665 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4669 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4670 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4671 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4672 /* we can skip this block, and probably more */
4673 sync_blocks
/= STRIPE_SECTORS
;
4675 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4678 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4680 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4682 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4683 /* make sure we don't swamp the stripe cache if someone else
4684 * is trying to get access
4686 schedule_timeout_uninterruptible(1);
4688 /* Need to check if array will still be degraded after recovery/resync
4689 * We don't need to check the 'failed' flag as when that gets set,
4692 for (i
= 0; i
< conf
->raid_disks
; i
++)
4693 if (conf
->disks
[i
].rdev
== NULL
)
4696 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4698 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4703 return STRIPE_SECTORS
;
4706 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
4708 /* We may not be able to submit a whole bio at once as there
4709 * may not be enough stripe_heads available.
4710 * We cannot pre-allocate enough stripe_heads as we may need
4711 * more than exist in the cache (if we allow ever large chunks).
4712 * So we do one stripe head at a time and record in
4713 * ->bi_hw_segments how many have been done.
4715 * We *know* that this entire raid_bio is in one chunk, so
4716 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4718 struct stripe_head
*sh
;
4720 sector_t sector
, logical_sector
, last_sector
;
4725 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4726 sector
= raid5_compute_sector(conf
, logical_sector
,
4728 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4730 for (; logical_sector
< last_sector
;
4731 logical_sector
+= STRIPE_SECTORS
,
4732 sector
+= STRIPE_SECTORS
,
4735 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
4736 /* already done this stripe */
4739 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4742 /* failed to get a stripe - must wait */
4743 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4744 conf
->retry_read_aligned
= raid_bio
;
4748 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4750 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4751 conf
->retry_read_aligned
= raid_bio
;
4755 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
4760 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
4762 bio_endio(raid_bio
, 0);
4763 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4764 wake_up(&conf
->wait_for_stripe
);
4768 #define MAX_STRIPE_BATCH 8
4769 static int handle_active_stripes(struct r5conf
*conf
)
4771 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
4772 int i
, batch_size
= 0;
4774 while (batch_size
< MAX_STRIPE_BATCH
&&
4775 (sh
= __get_priority_stripe(conf
)) != NULL
)
4776 batch
[batch_size
++] = sh
;
4778 if (batch_size
== 0)
4780 spin_unlock_irq(&conf
->device_lock
);
4782 for (i
= 0; i
< batch_size
; i
++)
4783 handle_stripe(batch
[i
]);
4787 spin_lock_irq(&conf
->device_lock
);
4788 for (i
= 0; i
< batch_size
; i
++)
4789 __release_stripe(conf
, batch
[i
]);
4794 * This is our raid5 kernel thread.
4796 * We scan the hash table for stripes which can be handled now.
4797 * During the scan, completed stripes are saved for us by the interrupt
4798 * handler, so that they will not have to wait for our next wakeup.
4800 static void raid5d(struct md_thread
*thread
)
4802 struct mddev
*mddev
= thread
->mddev
;
4803 struct r5conf
*conf
= mddev
->private;
4805 struct blk_plug plug
;
4807 pr_debug("+++ raid5d active\n");
4809 md_check_recovery(mddev
);
4811 blk_start_plug(&plug
);
4813 spin_lock_irq(&conf
->device_lock
);
4819 !list_empty(&conf
->bitmap_list
)) {
4820 /* Now is a good time to flush some bitmap updates */
4822 spin_unlock_irq(&conf
->device_lock
);
4823 bitmap_unplug(mddev
->bitmap
);
4824 spin_lock_irq(&conf
->device_lock
);
4825 conf
->seq_write
= conf
->seq_flush
;
4826 activate_bit_delay(conf
);
4828 raid5_activate_delayed(conf
);
4830 while ((bio
= remove_bio_from_retry(conf
))) {
4832 spin_unlock_irq(&conf
->device_lock
);
4833 ok
= retry_aligned_read(conf
, bio
);
4834 spin_lock_irq(&conf
->device_lock
);
4840 batch_size
= handle_active_stripes(conf
);
4843 handled
+= batch_size
;
4845 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
4846 spin_unlock_irq(&conf
->device_lock
);
4847 md_check_recovery(mddev
);
4848 spin_lock_irq(&conf
->device_lock
);
4851 pr_debug("%d stripes handled\n", handled
);
4853 spin_unlock_irq(&conf
->device_lock
);
4855 async_tx_issue_pending_all();
4856 blk_finish_plug(&plug
);
4858 pr_debug("--- raid5d inactive\n");
4862 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
4864 struct r5conf
*conf
= mddev
->private;
4866 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4872 raid5_set_cache_size(struct mddev
*mddev
, int size
)
4874 struct r5conf
*conf
= mddev
->private;
4877 if (size
<= 16 || size
> 32768)
4879 while (size
< conf
->max_nr_stripes
) {
4880 if (drop_one_stripe(conf
))
4881 conf
->max_nr_stripes
--;
4885 err
= md_allow_write(mddev
);
4888 while (size
> conf
->max_nr_stripes
) {
4889 if (grow_one_stripe(conf
))
4890 conf
->max_nr_stripes
++;
4895 EXPORT_SYMBOL(raid5_set_cache_size
);
4898 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
4900 struct r5conf
*conf
= mddev
->private;
4904 if (len
>= PAGE_SIZE
)
4909 if (strict_strtoul(page
, 10, &new))
4911 err
= raid5_set_cache_size(mddev
, new);
4917 static struct md_sysfs_entry
4918 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4919 raid5_show_stripe_cache_size
,
4920 raid5_store_stripe_cache_size
);
4923 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
4925 struct r5conf
*conf
= mddev
->private;
4927 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4933 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
4935 struct r5conf
*conf
= mddev
->private;
4937 if (len
>= PAGE_SIZE
)
4942 if (strict_strtoul(page
, 10, &new))
4944 if (new > conf
->max_nr_stripes
)
4946 conf
->bypass_threshold
= new;
4950 static struct md_sysfs_entry
4951 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4953 raid5_show_preread_threshold
,
4954 raid5_store_preread_threshold
);
4957 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
4959 struct r5conf
*conf
= mddev
->private;
4961 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4966 static struct md_sysfs_entry
4967 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4969 static struct attribute
*raid5_attrs
[] = {
4970 &raid5_stripecache_size
.attr
,
4971 &raid5_stripecache_active
.attr
,
4972 &raid5_preread_bypass_threshold
.attr
,
4975 static struct attribute_group raid5_attrs_group
= {
4977 .attrs
= raid5_attrs
,
4981 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
4983 struct r5conf
*conf
= mddev
->private;
4986 sectors
= mddev
->dev_sectors
;
4988 /* size is defined by the smallest of previous and new size */
4989 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4991 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4992 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4993 return sectors
* (raid_disks
- conf
->max_degraded
);
4996 static void raid5_free_percpu(struct r5conf
*conf
)
4998 struct raid5_percpu
*percpu
;
5005 for_each_possible_cpu(cpu
) {
5006 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
5007 safe_put_page(percpu
->spare_page
);
5008 kfree(percpu
->scribble
);
5010 #ifdef CONFIG_HOTPLUG_CPU
5011 unregister_cpu_notifier(&conf
->cpu_notify
);
5015 free_percpu(conf
->percpu
);
5018 static void free_conf(struct r5conf
*conf
)
5020 shrink_stripes(conf
);
5021 raid5_free_percpu(conf
);
5023 kfree(conf
->stripe_hashtbl
);
5027 #ifdef CONFIG_HOTPLUG_CPU
5028 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
5031 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
5032 long cpu
= (long)hcpu
;
5033 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
5036 case CPU_UP_PREPARE
:
5037 case CPU_UP_PREPARE_FROZEN
:
5038 if (conf
->level
== 6 && !percpu
->spare_page
)
5039 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
5040 if (!percpu
->scribble
)
5041 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5043 if (!percpu
->scribble
||
5044 (conf
->level
== 6 && !percpu
->spare_page
)) {
5045 safe_put_page(percpu
->spare_page
);
5046 kfree(percpu
->scribble
);
5047 pr_err("%s: failed memory allocation for cpu%ld\n",
5049 return notifier_from_errno(-ENOMEM
);
5053 case CPU_DEAD_FROZEN
:
5054 safe_put_page(percpu
->spare_page
);
5055 kfree(percpu
->scribble
);
5056 percpu
->spare_page
= NULL
;
5057 percpu
->scribble
= NULL
;
5066 static int raid5_alloc_percpu(struct r5conf
*conf
)
5069 struct page
*spare_page
;
5070 struct raid5_percpu __percpu
*allcpus
;
5074 allcpus
= alloc_percpu(struct raid5_percpu
);
5077 conf
->percpu
= allcpus
;
5081 for_each_present_cpu(cpu
) {
5082 if (conf
->level
== 6) {
5083 spare_page
= alloc_page(GFP_KERNEL
);
5088 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
5090 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5095 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
5097 #ifdef CONFIG_HOTPLUG_CPU
5098 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
5099 conf
->cpu_notify
.priority
= 0;
5101 err
= register_cpu_notifier(&conf
->cpu_notify
);
5108 static struct r5conf
*setup_conf(struct mddev
*mddev
)
5110 struct r5conf
*conf
;
5111 int raid_disk
, memory
, max_disks
;
5112 struct md_rdev
*rdev
;
5113 struct disk_info
*disk
;
5116 if (mddev
->new_level
!= 5
5117 && mddev
->new_level
!= 4
5118 && mddev
->new_level
!= 6) {
5119 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5120 mdname(mddev
), mddev
->new_level
);
5121 return ERR_PTR(-EIO
);
5123 if ((mddev
->new_level
== 5
5124 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
5125 (mddev
->new_level
== 6
5126 && !algorithm_valid_raid6(mddev
->new_layout
))) {
5127 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
5128 mdname(mddev
), mddev
->new_layout
);
5129 return ERR_PTR(-EIO
);
5131 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
5132 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5133 mdname(mddev
), mddev
->raid_disks
);
5134 return ERR_PTR(-EINVAL
);
5137 if (!mddev
->new_chunk_sectors
||
5138 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
5139 !is_power_of_2(mddev
->new_chunk_sectors
)) {
5140 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
5141 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
5142 return ERR_PTR(-EINVAL
);
5145 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
5148 spin_lock_init(&conf
->device_lock
);
5149 init_waitqueue_head(&conf
->wait_for_stripe
);
5150 init_waitqueue_head(&conf
->wait_for_overlap
);
5151 INIT_LIST_HEAD(&conf
->handle_list
);
5152 INIT_LIST_HEAD(&conf
->hold_list
);
5153 INIT_LIST_HEAD(&conf
->delayed_list
);
5154 INIT_LIST_HEAD(&conf
->bitmap_list
);
5155 INIT_LIST_HEAD(&conf
->inactive_list
);
5156 atomic_set(&conf
->active_stripes
, 0);
5157 atomic_set(&conf
->preread_active_stripes
, 0);
5158 atomic_set(&conf
->active_aligned_reads
, 0);
5159 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
5160 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
5162 conf
->raid_disks
= mddev
->raid_disks
;
5163 if (mddev
->reshape_position
== MaxSector
)
5164 conf
->previous_raid_disks
= mddev
->raid_disks
;
5166 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5167 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
5168 conf
->scribble_len
= scribble_len(max_disks
);
5170 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
5175 conf
->mddev
= mddev
;
5177 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
5180 conf
->level
= mddev
->new_level
;
5181 if (raid5_alloc_percpu(conf
) != 0)
5184 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
5186 rdev_for_each(rdev
, mddev
) {
5187 raid_disk
= rdev
->raid_disk
;
5188 if (raid_disk
>= max_disks
5191 disk
= conf
->disks
+ raid_disk
;
5193 if (test_bit(Replacement
, &rdev
->flags
)) {
5194 if (disk
->replacement
)
5196 disk
->replacement
= rdev
;
5203 if (test_bit(In_sync
, &rdev
->flags
)) {
5204 char b
[BDEVNAME_SIZE
];
5205 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
5207 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
5208 } else if (rdev
->saved_raid_disk
!= raid_disk
)
5209 /* Cannot rely on bitmap to complete recovery */
5213 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5214 conf
->level
= mddev
->new_level
;
5215 if (conf
->level
== 6)
5216 conf
->max_degraded
= 2;
5218 conf
->max_degraded
= 1;
5219 conf
->algorithm
= mddev
->new_layout
;
5220 conf
->max_nr_stripes
= NR_STRIPES
;
5221 conf
->reshape_progress
= mddev
->reshape_position
;
5222 if (conf
->reshape_progress
!= MaxSector
) {
5223 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
5224 conf
->prev_algo
= mddev
->layout
;
5227 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
5228 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
5229 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
5231 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5232 mdname(mddev
), memory
);
5235 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
5236 mdname(mddev
), memory
);
5238 sprintf(pers_name
, "raid%d", mddev
->new_level
);
5239 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
5240 if (!conf
->thread
) {
5242 "md/raid:%s: couldn't allocate thread.\n",
5252 return ERR_PTR(-EIO
);
5254 return ERR_PTR(-ENOMEM
);
5258 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
5261 case ALGORITHM_PARITY_0
:
5262 if (raid_disk
< max_degraded
)
5265 case ALGORITHM_PARITY_N
:
5266 if (raid_disk
>= raid_disks
- max_degraded
)
5269 case ALGORITHM_PARITY_0_6
:
5270 if (raid_disk
== 0 ||
5271 raid_disk
== raid_disks
- 1)
5274 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5275 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5276 case ALGORITHM_LEFT_SYMMETRIC_6
:
5277 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5278 if (raid_disk
== raid_disks
- 1)
5284 static int run(struct mddev
*mddev
)
5286 struct r5conf
*conf
;
5287 int working_disks
= 0;
5288 int dirty_parity_disks
= 0;
5289 struct md_rdev
*rdev
;
5290 sector_t reshape_offset
= 0;
5292 long long min_offset_diff
= 0;
5295 if (mddev
->recovery_cp
!= MaxSector
)
5296 printk(KERN_NOTICE
"md/raid:%s: not clean"
5297 " -- starting background reconstruction\n",
5300 rdev_for_each(rdev
, mddev
) {
5302 if (rdev
->raid_disk
< 0)
5304 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
5306 min_offset_diff
= diff
;
5308 } else if (mddev
->reshape_backwards
&&
5309 diff
< min_offset_diff
)
5310 min_offset_diff
= diff
;
5311 else if (!mddev
->reshape_backwards
&&
5312 diff
> min_offset_diff
)
5313 min_offset_diff
= diff
;
5316 if (mddev
->reshape_position
!= MaxSector
) {
5317 /* Check that we can continue the reshape.
5318 * Difficulties arise if the stripe we would write to
5319 * next is at or after the stripe we would read from next.
5320 * For a reshape that changes the number of devices, this
5321 * is only possible for a very short time, and mdadm makes
5322 * sure that time appears to have past before assembling
5323 * the array. So we fail if that time hasn't passed.
5324 * For a reshape that keeps the number of devices the same
5325 * mdadm must be monitoring the reshape can keeping the
5326 * critical areas read-only and backed up. It will start
5327 * the array in read-only mode, so we check for that.
5329 sector_t here_new
, here_old
;
5331 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5333 if (mddev
->new_level
!= mddev
->level
) {
5334 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5335 "required - aborting.\n",
5339 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5340 /* reshape_position must be on a new-stripe boundary, and one
5341 * further up in new geometry must map after here in old
5344 here_new
= mddev
->reshape_position
;
5345 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5346 (mddev
->raid_disks
- max_degraded
))) {
5347 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5348 "on a stripe boundary\n", mdname(mddev
));
5351 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5352 /* here_new is the stripe we will write to */
5353 here_old
= mddev
->reshape_position
;
5354 sector_div(here_old
, mddev
->chunk_sectors
*
5355 (old_disks
-max_degraded
));
5356 /* here_old is the first stripe that we might need to read
5358 if (mddev
->delta_disks
== 0) {
5359 if ((here_new
* mddev
->new_chunk_sectors
!=
5360 here_old
* mddev
->chunk_sectors
)) {
5361 printk(KERN_ERR
"md/raid:%s: reshape position is"
5362 " confused - aborting\n", mdname(mddev
));
5365 /* We cannot be sure it is safe to start an in-place
5366 * reshape. It is only safe if user-space is monitoring
5367 * and taking constant backups.
5368 * mdadm always starts a situation like this in
5369 * readonly mode so it can take control before
5370 * allowing any writes. So just check for that.
5372 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
5373 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
5374 /* not really in-place - so OK */;
5375 else if (mddev
->ro
== 0) {
5376 printk(KERN_ERR
"md/raid:%s: in-place reshape "
5377 "must be started in read-only mode "
5382 } else if (mddev
->reshape_backwards
5383 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
5384 here_old
* mddev
->chunk_sectors
)
5385 : (here_new
* mddev
->new_chunk_sectors
>=
5386 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
5387 /* Reading from the same stripe as writing to - bad */
5388 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5389 "auto-recovery - aborting.\n",
5393 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5395 /* OK, we should be able to continue; */
5397 BUG_ON(mddev
->level
!= mddev
->new_level
);
5398 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5399 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5400 BUG_ON(mddev
->delta_disks
!= 0);
5403 if (mddev
->private == NULL
)
5404 conf
= setup_conf(mddev
);
5406 conf
= mddev
->private;
5409 return PTR_ERR(conf
);
5411 conf
->min_offset_diff
= min_offset_diff
;
5412 mddev
->thread
= conf
->thread
;
5413 conf
->thread
= NULL
;
5414 mddev
->private = conf
;
5416 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5418 rdev
= conf
->disks
[i
].rdev
;
5419 if (!rdev
&& conf
->disks
[i
].replacement
) {
5420 /* The replacement is all we have yet */
5421 rdev
= conf
->disks
[i
].replacement
;
5422 conf
->disks
[i
].replacement
= NULL
;
5423 clear_bit(Replacement
, &rdev
->flags
);
5424 conf
->disks
[i
].rdev
= rdev
;
5428 if (conf
->disks
[i
].replacement
&&
5429 conf
->reshape_progress
!= MaxSector
) {
5430 /* replacements and reshape simply do not mix. */
5431 printk(KERN_ERR
"md: cannot handle concurrent "
5432 "replacement and reshape.\n");
5435 if (test_bit(In_sync
, &rdev
->flags
)) {
5439 /* This disc is not fully in-sync. However if it
5440 * just stored parity (beyond the recovery_offset),
5441 * when we don't need to be concerned about the
5442 * array being dirty.
5443 * When reshape goes 'backwards', we never have
5444 * partially completed devices, so we only need
5445 * to worry about reshape going forwards.
5447 /* Hack because v0.91 doesn't store recovery_offset properly. */
5448 if (mddev
->major_version
== 0 &&
5449 mddev
->minor_version
> 90)
5450 rdev
->recovery_offset
= reshape_offset
;
5452 if (rdev
->recovery_offset
< reshape_offset
) {
5453 /* We need to check old and new layout */
5454 if (!only_parity(rdev
->raid_disk
,
5457 conf
->max_degraded
))
5460 if (!only_parity(rdev
->raid_disk
,
5462 conf
->previous_raid_disks
,
5463 conf
->max_degraded
))
5465 dirty_parity_disks
++;
5469 * 0 for a fully functional array, 1 or 2 for a degraded array.
5471 mddev
->degraded
= calc_degraded(conf
);
5473 if (has_failed(conf
)) {
5474 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5475 " (%d/%d failed)\n",
5476 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5480 /* device size must be a multiple of chunk size */
5481 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5482 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5484 if (mddev
->degraded
> dirty_parity_disks
&&
5485 mddev
->recovery_cp
!= MaxSector
) {
5486 if (mddev
->ok_start_degraded
)
5488 "md/raid:%s: starting dirty degraded array"
5489 " - data corruption possible.\n",
5493 "md/raid:%s: cannot start dirty degraded array.\n",
5499 if (mddev
->degraded
== 0)
5500 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5501 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5502 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5505 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5506 " out of %d devices, algorithm %d\n",
5507 mdname(mddev
), conf
->level
,
5508 mddev
->raid_disks
- mddev
->degraded
,
5509 mddev
->raid_disks
, mddev
->new_layout
);
5511 print_raid5_conf(conf
);
5513 if (conf
->reshape_progress
!= MaxSector
) {
5514 conf
->reshape_safe
= conf
->reshape_progress
;
5515 atomic_set(&conf
->reshape_stripes
, 0);
5516 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5517 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5518 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5519 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5520 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5525 /* Ok, everything is just fine now */
5526 if (mddev
->to_remove
== &raid5_attrs_group
)
5527 mddev
->to_remove
= NULL
;
5528 else if (mddev
->kobj
.sd
&&
5529 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5531 "raid5: failed to create sysfs attributes for %s\n",
5533 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5537 bool discard_supported
= true;
5538 /* read-ahead size must cover two whole stripes, which
5539 * is 2 * (datadisks) * chunksize where 'n' is the
5540 * number of raid devices
5542 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5543 int stripe
= data_disks
*
5544 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5545 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5546 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5548 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5550 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5551 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5553 chunk_size
= mddev
->chunk_sectors
<< 9;
5554 blk_queue_io_min(mddev
->queue
, chunk_size
);
5555 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5556 (conf
->raid_disks
- conf
->max_degraded
));
5558 * We can only discard a whole stripe. It doesn't make sense to
5559 * discard data disk but write parity disk
5561 stripe
= stripe
* PAGE_SIZE
;
5562 /* Round up to power of 2, as discard handling
5563 * currently assumes that */
5564 while ((stripe
-1) & stripe
)
5565 stripe
= (stripe
| (stripe
-1)) + 1;
5566 mddev
->queue
->limits
.discard_alignment
= stripe
;
5567 mddev
->queue
->limits
.discard_granularity
= stripe
;
5569 * unaligned part of discard request will be ignored, so can't
5570 * guarantee discard_zerors_data
5572 mddev
->queue
->limits
.discard_zeroes_data
= 0;
5574 rdev_for_each(rdev
, mddev
) {
5575 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5576 rdev
->data_offset
<< 9);
5577 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5578 rdev
->new_data_offset
<< 9);
5580 * discard_zeroes_data is required, otherwise data
5581 * could be lost. Consider a scenario: discard a stripe
5582 * (the stripe could be inconsistent if
5583 * discard_zeroes_data is 0); write one disk of the
5584 * stripe (the stripe could be inconsistent again
5585 * depending on which disks are used to calculate
5586 * parity); the disk is broken; The stripe data of this
5589 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
5590 !bdev_get_queue(rdev
->bdev
)->
5591 limits
.discard_zeroes_data
)
5592 discard_supported
= false;
5595 if (discard_supported
&&
5596 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
5597 mddev
->queue
->limits
.discard_granularity
>= stripe
)
5598 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
5601 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
5607 md_unregister_thread(&mddev
->thread
);
5608 print_raid5_conf(conf
);
5610 mddev
->private = NULL
;
5611 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5615 static int stop(struct mddev
*mddev
)
5617 struct r5conf
*conf
= mddev
->private;
5619 md_unregister_thread(&mddev
->thread
);
5621 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5623 mddev
->private = NULL
;
5624 mddev
->to_remove
= &raid5_attrs_group
;
5628 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
5630 struct r5conf
*conf
= mddev
->private;
5633 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5634 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5635 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5636 for (i
= 0; i
< conf
->raid_disks
; i
++)
5637 seq_printf (seq
, "%s",
5638 conf
->disks
[i
].rdev
&&
5639 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5640 seq_printf (seq
, "]");
5643 static void print_raid5_conf (struct r5conf
*conf
)
5646 struct disk_info
*tmp
;
5648 printk(KERN_DEBUG
"RAID conf printout:\n");
5650 printk("(conf==NULL)\n");
5653 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5655 conf
->raid_disks
- conf
->mddev
->degraded
);
5657 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5658 char b
[BDEVNAME_SIZE
];
5659 tmp
= conf
->disks
+ i
;
5661 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5662 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5663 bdevname(tmp
->rdev
->bdev
, b
));
5667 static int raid5_spare_active(struct mddev
*mddev
)
5670 struct r5conf
*conf
= mddev
->private;
5671 struct disk_info
*tmp
;
5673 unsigned long flags
;
5675 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5676 tmp
= conf
->disks
+ i
;
5677 if (tmp
->replacement
5678 && tmp
->replacement
->recovery_offset
== MaxSector
5679 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
5680 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
5681 /* Replacement has just become active. */
5683 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
5686 /* Replaced device not technically faulty,
5687 * but we need to be sure it gets removed
5688 * and never re-added.
5690 set_bit(Faulty
, &tmp
->rdev
->flags
);
5691 sysfs_notify_dirent_safe(
5692 tmp
->rdev
->sysfs_state
);
5694 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
5695 } else if (tmp
->rdev
5696 && tmp
->rdev
->recovery_offset
== MaxSector
5697 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5698 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5700 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5703 spin_lock_irqsave(&conf
->device_lock
, flags
);
5704 mddev
->degraded
= calc_degraded(conf
);
5705 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5706 print_raid5_conf(conf
);
5710 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5712 struct r5conf
*conf
= mddev
->private;
5714 int number
= rdev
->raid_disk
;
5715 struct md_rdev
**rdevp
;
5716 struct disk_info
*p
= conf
->disks
+ number
;
5718 print_raid5_conf(conf
);
5719 if (rdev
== p
->rdev
)
5721 else if (rdev
== p
->replacement
)
5722 rdevp
= &p
->replacement
;
5726 if (number
>= conf
->raid_disks
&&
5727 conf
->reshape_progress
== MaxSector
)
5728 clear_bit(In_sync
, &rdev
->flags
);
5730 if (test_bit(In_sync
, &rdev
->flags
) ||
5731 atomic_read(&rdev
->nr_pending
)) {
5735 /* Only remove non-faulty devices if recovery
5738 if (!test_bit(Faulty
, &rdev
->flags
) &&
5739 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5740 !has_failed(conf
) &&
5741 (!p
->replacement
|| p
->replacement
== rdev
) &&
5742 number
< conf
->raid_disks
) {
5748 if (atomic_read(&rdev
->nr_pending
)) {
5749 /* lost the race, try later */
5752 } else if (p
->replacement
) {
5753 /* We must have just cleared 'rdev' */
5754 p
->rdev
= p
->replacement
;
5755 clear_bit(Replacement
, &p
->replacement
->flags
);
5756 smp_mb(); /* Make sure other CPUs may see both as identical
5757 * but will never see neither - if they are careful
5759 p
->replacement
= NULL
;
5760 clear_bit(WantReplacement
, &rdev
->flags
);
5762 /* We might have just removed the Replacement as faulty-
5763 * clear the bit just in case
5765 clear_bit(WantReplacement
, &rdev
->flags
);
5768 print_raid5_conf(conf
);
5772 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5774 struct r5conf
*conf
= mddev
->private;
5777 struct disk_info
*p
;
5779 int last
= conf
->raid_disks
- 1;
5781 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5784 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
5785 /* no point adding a device */
5788 if (rdev
->raid_disk
>= 0)
5789 first
= last
= rdev
->raid_disk
;
5792 * find the disk ... but prefer rdev->saved_raid_disk
5795 if (rdev
->saved_raid_disk
>= 0 &&
5796 rdev
->saved_raid_disk
>= first
&&
5797 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5798 first
= rdev
->saved_raid_disk
;
5800 for (disk
= first
; disk
<= last
; disk
++) {
5801 p
= conf
->disks
+ disk
;
5802 if (p
->rdev
== NULL
) {
5803 clear_bit(In_sync
, &rdev
->flags
);
5804 rdev
->raid_disk
= disk
;
5806 if (rdev
->saved_raid_disk
!= disk
)
5808 rcu_assign_pointer(p
->rdev
, rdev
);
5812 for (disk
= first
; disk
<= last
; disk
++) {
5813 p
= conf
->disks
+ disk
;
5814 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
5815 p
->replacement
== NULL
) {
5816 clear_bit(In_sync
, &rdev
->flags
);
5817 set_bit(Replacement
, &rdev
->flags
);
5818 rdev
->raid_disk
= disk
;
5821 rcu_assign_pointer(p
->replacement
, rdev
);
5826 print_raid5_conf(conf
);
5830 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
5832 /* no resync is happening, and there is enough space
5833 * on all devices, so we can resize.
5834 * We need to make sure resync covers any new space.
5835 * If the array is shrinking we should possibly wait until
5836 * any io in the removed space completes, but it hardly seems
5840 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5841 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
5842 if (mddev
->external_size
&&
5843 mddev
->array_sectors
> newsize
)
5845 if (mddev
->bitmap
) {
5846 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
5850 md_set_array_sectors(mddev
, newsize
);
5851 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5852 revalidate_disk(mddev
->gendisk
);
5853 if (sectors
> mddev
->dev_sectors
&&
5854 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5855 mddev
->recovery_cp
= mddev
->dev_sectors
;
5856 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5858 mddev
->dev_sectors
= sectors
;
5859 mddev
->resync_max_sectors
= sectors
;
5863 static int check_stripe_cache(struct mddev
*mddev
)
5865 /* Can only proceed if there are plenty of stripe_heads.
5866 * We need a minimum of one full stripe,, and for sensible progress
5867 * it is best to have about 4 times that.
5868 * If we require 4 times, then the default 256 4K stripe_heads will
5869 * allow for chunk sizes up to 256K, which is probably OK.
5870 * If the chunk size is greater, user-space should request more
5871 * stripe_heads first.
5873 struct r5conf
*conf
= mddev
->private;
5874 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5875 > conf
->max_nr_stripes
||
5876 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5877 > conf
->max_nr_stripes
) {
5878 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5880 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5887 static int check_reshape(struct mddev
*mddev
)
5889 struct r5conf
*conf
= mddev
->private;
5891 if (mddev
->delta_disks
== 0 &&
5892 mddev
->new_layout
== mddev
->layout
&&
5893 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5894 return 0; /* nothing to do */
5895 if (has_failed(conf
))
5897 if (mddev
->delta_disks
< 0) {
5898 /* We might be able to shrink, but the devices must
5899 * be made bigger first.
5900 * For raid6, 4 is the minimum size.
5901 * Otherwise 2 is the minimum
5904 if (mddev
->level
== 6)
5906 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5910 if (!check_stripe_cache(mddev
))
5913 return resize_stripes(conf
, (conf
->previous_raid_disks
5914 + mddev
->delta_disks
));
5917 static int raid5_start_reshape(struct mddev
*mddev
)
5919 struct r5conf
*conf
= mddev
->private;
5920 struct md_rdev
*rdev
;
5922 unsigned long flags
;
5924 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5927 if (!check_stripe_cache(mddev
))
5930 if (has_failed(conf
))
5933 rdev_for_each(rdev
, mddev
) {
5934 if (!test_bit(In_sync
, &rdev
->flags
)
5935 && !test_bit(Faulty
, &rdev
->flags
))
5939 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5940 /* Not enough devices even to make a degraded array
5945 /* Refuse to reduce size of the array. Any reductions in
5946 * array size must be through explicit setting of array_size
5949 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5950 < mddev
->array_sectors
) {
5951 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5952 "before number of disks\n", mdname(mddev
));
5956 atomic_set(&conf
->reshape_stripes
, 0);
5957 spin_lock_irq(&conf
->device_lock
);
5958 conf
->previous_raid_disks
= conf
->raid_disks
;
5959 conf
->raid_disks
+= mddev
->delta_disks
;
5960 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5961 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5962 conf
->prev_algo
= conf
->algorithm
;
5963 conf
->algorithm
= mddev
->new_layout
;
5965 /* Code that selects data_offset needs to see the generation update
5966 * if reshape_progress has been set - so a memory barrier needed.
5969 if (mddev
->reshape_backwards
)
5970 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5972 conf
->reshape_progress
= 0;
5973 conf
->reshape_safe
= conf
->reshape_progress
;
5974 spin_unlock_irq(&conf
->device_lock
);
5976 /* Add some new drives, as many as will fit.
5977 * We know there are enough to make the newly sized array work.
5978 * Don't add devices if we are reducing the number of
5979 * devices in the array. This is because it is not possible
5980 * to correctly record the "partially reconstructed" state of
5981 * such devices during the reshape and confusion could result.
5983 if (mddev
->delta_disks
>= 0) {
5984 rdev_for_each(rdev
, mddev
)
5985 if (rdev
->raid_disk
< 0 &&
5986 !test_bit(Faulty
, &rdev
->flags
)) {
5987 if (raid5_add_disk(mddev
, rdev
) == 0) {
5989 >= conf
->previous_raid_disks
)
5990 set_bit(In_sync
, &rdev
->flags
);
5992 rdev
->recovery_offset
= 0;
5994 if (sysfs_link_rdev(mddev
, rdev
))
5995 /* Failure here is OK */;
5997 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5998 && !test_bit(Faulty
, &rdev
->flags
)) {
5999 /* This is a spare that was manually added */
6000 set_bit(In_sync
, &rdev
->flags
);
6003 /* When a reshape changes the number of devices,
6004 * ->degraded is measured against the larger of the
6005 * pre and post number of devices.
6007 spin_lock_irqsave(&conf
->device_lock
, flags
);
6008 mddev
->degraded
= calc_degraded(conf
);
6009 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
6011 mddev
->raid_disks
= conf
->raid_disks
;
6012 mddev
->reshape_position
= conf
->reshape_progress
;
6013 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6015 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6016 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6017 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6018 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6019 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6021 if (!mddev
->sync_thread
) {
6022 mddev
->recovery
= 0;
6023 spin_lock_irq(&conf
->device_lock
);
6024 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
6025 rdev_for_each(rdev
, mddev
)
6026 rdev
->new_data_offset
= rdev
->data_offset
;
6028 conf
->reshape_progress
= MaxSector
;
6029 mddev
->reshape_position
= MaxSector
;
6030 spin_unlock_irq(&conf
->device_lock
);
6033 conf
->reshape_checkpoint
= jiffies
;
6034 md_wakeup_thread(mddev
->sync_thread
);
6035 md_new_event(mddev
);
6039 /* This is called from the reshape thread and should make any
6040 * changes needed in 'conf'
6042 static void end_reshape(struct r5conf
*conf
)
6045 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
6046 struct md_rdev
*rdev
;
6048 spin_lock_irq(&conf
->device_lock
);
6049 conf
->previous_raid_disks
= conf
->raid_disks
;
6050 rdev_for_each(rdev
, conf
->mddev
)
6051 rdev
->data_offset
= rdev
->new_data_offset
;
6053 conf
->reshape_progress
= MaxSector
;
6054 spin_unlock_irq(&conf
->device_lock
);
6055 wake_up(&conf
->wait_for_overlap
);
6057 /* read-ahead size must cover two whole stripes, which is
6058 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6060 if (conf
->mddev
->queue
) {
6061 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
6062 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
6064 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6065 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6070 /* This is called from the raid5d thread with mddev_lock held.
6071 * It makes config changes to the device.
6073 static void raid5_finish_reshape(struct mddev
*mddev
)
6075 struct r5conf
*conf
= mddev
->private;
6077 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
6079 if (mddev
->delta_disks
> 0) {
6080 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6081 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
6082 revalidate_disk(mddev
->gendisk
);
6085 spin_lock_irq(&conf
->device_lock
);
6086 mddev
->degraded
= calc_degraded(conf
);
6087 spin_unlock_irq(&conf
->device_lock
);
6088 for (d
= conf
->raid_disks
;
6089 d
< conf
->raid_disks
- mddev
->delta_disks
;
6091 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
6093 clear_bit(In_sync
, &rdev
->flags
);
6094 rdev
= conf
->disks
[d
].replacement
;
6096 clear_bit(In_sync
, &rdev
->flags
);
6099 mddev
->layout
= conf
->algorithm
;
6100 mddev
->chunk_sectors
= conf
->chunk_sectors
;
6101 mddev
->reshape_position
= MaxSector
;
6102 mddev
->delta_disks
= 0;
6103 mddev
->reshape_backwards
= 0;
6107 static void raid5_quiesce(struct mddev
*mddev
, int state
)
6109 struct r5conf
*conf
= mddev
->private;
6112 case 2: /* resume for a suspend */
6113 wake_up(&conf
->wait_for_overlap
);
6116 case 1: /* stop all writes */
6117 spin_lock_irq(&conf
->device_lock
);
6118 /* '2' tells resync/reshape to pause so that all
6119 * active stripes can drain
6122 wait_event_lock_irq(conf
->wait_for_stripe
,
6123 atomic_read(&conf
->active_stripes
) == 0 &&
6124 atomic_read(&conf
->active_aligned_reads
) == 0,
6127 spin_unlock_irq(&conf
->device_lock
);
6128 /* allow reshape to continue */
6129 wake_up(&conf
->wait_for_overlap
);
6132 case 0: /* re-enable writes */
6133 spin_lock_irq(&conf
->device_lock
);
6135 wake_up(&conf
->wait_for_stripe
);
6136 wake_up(&conf
->wait_for_overlap
);
6137 spin_unlock_irq(&conf
->device_lock
);
6143 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
6145 struct r0conf
*raid0_conf
= mddev
->private;
6148 /* for raid0 takeover only one zone is supported */
6149 if (raid0_conf
->nr_strip_zones
> 1) {
6150 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6152 return ERR_PTR(-EINVAL
);
6155 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
6156 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
6157 mddev
->dev_sectors
= sectors
;
6158 mddev
->new_level
= level
;
6159 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6160 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
6161 mddev
->raid_disks
+= 1;
6162 mddev
->delta_disks
= 1;
6163 /* make sure it will be not marked as dirty */
6164 mddev
->recovery_cp
= MaxSector
;
6166 return setup_conf(mddev
);
6170 static void *raid5_takeover_raid1(struct mddev
*mddev
)
6174 if (mddev
->raid_disks
!= 2 ||
6175 mddev
->degraded
> 1)
6176 return ERR_PTR(-EINVAL
);
6178 /* Should check if there are write-behind devices? */
6180 chunksect
= 64*2; /* 64K by default */
6182 /* The array must be an exact multiple of chunksize */
6183 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
6186 if ((chunksect
<<9) < STRIPE_SIZE
)
6187 /* array size does not allow a suitable chunk size */
6188 return ERR_PTR(-EINVAL
);
6190 mddev
->new_level
= 5;
6191 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6192 mddev
->new_chunk_sectors
= chunksect
;
6194 return setup_conf(mddev
);
6197 static void *raid5_takeover_raid6(struct mddev
*mddev
)
6201 switch (mddev
->layout
) {
6202 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6203 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
6205 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6206 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
6208 case ALGORITHM_LEFT_SYMMETRIC_6
:
6209 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6211 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6212 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
6214 case ALGORITHM_PARITY_0_6
:
6215 new_layout
= ALGORITHM_PARITY_0
;
6217 case ALGORITHM_PARITY_N
:
6218 new_layout
= ALGORITHM_PARITY_N
;
6221 return ERR_PTR(-EINVAL
);
6223 mddev
->new_level
= 5;
6224 mddev
->new_layout
= new_layout
;
6225 mddev
->delta_disks
= -1;
6226 mddev
->raid_disks
-= 1;
6227 return setup_conf(mddev
);
6231 static int raid5_check_reshape(struct mddev
*mddev
)
6233 /* For a 2-drive array, the layout and chunk size can be changed
6234 * immediately as not restriping is needed.
6235 * For larger arrays we record the new value - after validation
6236 * to be used by a reshape pass.
6238 struct r5conf
*conf
= mddev
->private;
6239 int new_chunk
= mddev
->new_chunk_sectors
;
6241 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
6243 if (new_chunk
> 0) {
6244 if (!is_power_of_2(new_chunk
))
6246 if (new_chunk
< (PAGE_SIZE
>>9))
6248 if (mddev
->array_sectors
& (new_chunk
-1))
6249 /* not factor of array size */
6253 /* They look valid */
6255 if (mddev
->raid_disks
== 2) {
6256 /* can make the change immediately */
6257 if (mddev
->new_layout
>= 0) {
6258 conf
->algorithm
= mddev
->new_layout
;
6259 mddev
->layout
= mddev
->new_layout
;
6261 if (new_chunk
> 0) {
6262 conf
->chunk_sectors
= new_chunk
;
6263 mddev
->chunk_sectors
= new_chunk
;
6265 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6266 md_wakeup_thread(mddev
->thread
);
6268 return check_reshape(mddev
);
6271 static int raid6_check_reshape(struct mddev
*mddev
)
6273 int new_chunk
= mddev
->new_chunk_sectors
;
6275 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
6277 if (new_chunk
> 0) {
6278 if (!is_power_of_2(new_chunk
))
6280 if (new_chunk
< (PAGE_SIZE
>> 9))
6282 if (mddev
->array_sectors
& (new_chunk
-1))
6283 /* not factor of array size */
6287 /* They look valid */
6288 return check_reshape(mddev
);
6291 static void *raid5_takeover(struct mddev
*mddev
)
6293 /* raid5 can take over:
6294 * raid0 - if there is only one strip zone - make it a raid4 layout
6295 * raid1 - if there are two drives. We need to know the chunk size
6296 * raid4 - trivial - just use a raid4 layout.
6297 * raid6 - Providing it is a *_6 layout
6299 if (mddev
->level
== 0)
6300 return raid45_takeover_raid0(mddev
, 5);
6301 if (mddev
->level
== 1)
6302 return raid5_takeover_raid1(mddev
);
6303 if (mddev
->level
== 4) {
6304 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6305 mddev
->new_level
= 5;
6306 return setup_conf(mddev
);
6308 if (mddev
->level
== 6)
6309 return raid5_takeover_raid6(mddev
);
6311 return ERR_PTR(-EINVAL
);
6314 static void *raid4_takeover(struct mddev
*mddev
)
6316 /* raid4 can take over:
6317 * raid0 - if there is only one strip zone
6318 * raid5 - if layout is right
6320 if (mddev
->level
== 0)
6321 return raid45_takeover_raid0(mddev
, 4);
6322 if (mddev
->level
== 5 &&
6323 mddev
->layout
== ALGORITHM_PARITY_N
) {
6324 mddev
->new_layout
= 0;
6325 mddev
->new_level
= 4;
6326 return setup_conf(mddev
);
6328 return ERR_PTR(-EINVAL
);
6331 static struct md_personality raid5_personality
;
6333 static void *raid6_takeover(struct mddev
*mddev
)
6335 /* Currently can only take over a raid5. We map the
6336 * personality to an equivalent raid6 personality
6337 * with the Q block at the end.
6341 if (mddev
->pers
!= &raid5_personality
)
6342 return ERR_PTR(-EINVAL
);
6343 if (mddev
->degraded
> 1)
6344 return ERR_PTR(-EINVAL
);
6345 if (mddev
->raid_disks
> 253)
6346 return ERR_PTR(-EINVAL
);
6347 if (mddev
->raid_disks
< 3)
6348 return ERR_PTR(-EINVAL
);
6350 switch (mddev
->layout
) {
6351 case ALGORITHM_LEFT_ASYMMETRIC
:
6352 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
6354 case ALGORITHM_RIGHT_ASYMMETRIC
:
6355 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
6357 case ALGORITHM_LEFT_SYMMETRIC
:
6358 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
6360 case ALGORITHM_RIGHT_SYMMETRIC
:
6361 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
6363 case ALGORITHM_PARITY_0
:
6364 new_layout
= ALGORITHM_PARITY_0_6
;
6366 case ALGORITHM_PARITY_N
:
6367 new_layout
= ALGORITHM_PARITY_N
;
6370 return ERR_PTR(-EINVAL
);
6372 mddev
->new_level
= 6;
6373 mddev
->new_layout
= new_layout
;
6374 mddev
->delta_disks
= 1;
6375 mddev
->raid_disks
+= 1;
6376 return setup_conf(mddev
);
6380 static struct md_personality raid6_personality
=
6384 .owner
= THIS_MODULE
,
6385 .make_request
= make_request
,
6389 .error_handler
= error
,
6390 .hot_add_disk
= raid5_add_disk
,
6391 .hot_remove_disk
= raid5_remove_disk
,
6392 .spare_active
= raid5_spare_active
,
6393 .sync_request
= sync_request
,
6394 .resize
= raid5_resize
,
6396 .check_reshape
= raid6_check_reshape
,
6397 .start_reshape
= raid5_start_reshape
,
6398 .finish_reshape
= raid5_finish_reshape
,
6399 .quiesce
= raid5_quiesce
,
6400 .takeover
= raid6_takeover
,
6402 static struct md_personality raid5_personality
=
6406 .owner
= THIS_MODULE
,
6407 .make_request
= make_request
,
6411 .error_handler
= error
,
6412 .hot_add_disk
= raid5_add_disk
,
6413 .hot_remove_disk
= raid5_remove_disk
,
6414 .spare_active
= raid5_spare_active
,
6415 .sync_request
= sync_request
,
6416 .resize
= raid5_resize
,
6418 .check_reshape
= raid5_check_reshape
,
6419 .start_reshape
= raid5_start_reshape
,
6420 .finish_reshape
= raid5_finish_reshape
,
6421 .quiesce
= raid5_quiesce
,
6422 .takeover
= raid5_takeover
,
6425 static struct md_personality raid4_personality
=
6429 .owner
= THIS_MODULE
,
6430 .make_request
= make_request
,
6434 .error_handler
= error
,
6435 .hot_add_disk
= raid5_add_disk
,
6436 .hot_remove_disk
= raid5_remove_disk
,
6437 .spare_active
= raid5_spare_active
,
6438 .sync_request
= sync_request
,
6439 .resize
= raid5_resize
,
6441 .check_reshape
= raid5_check_reshape
,
6442 .start_reshape
= raid5_start_reshape
,
6443 .finish_reshape
= raid5_finish_reshape
,
6444 .quiesce
= raid5_quiesce
,
6445 .takeover
= raid4_takeover
,
6448 static int __init
raid5_init(void)
6450 register_md_personality(&raid6_personality
);
6451 register_md_personality(&raid5_personality
);
6452 register_md_personality(&raid4_personality
);
6456 static void raid5_exit(void)
6458 unregister_md_personality(&raid6_personality
);
6459 unregister_md_personality(&raid5_personality
);
6460 unregister_md_personality(&raid4_personality
);
6463 module_init(raid5_init
);
6464 module_exit(raid5_exit
);
6465 MODULE_LICENSE("GPL");
6466 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6467 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6468 MODULE_ALIAS("md-raid5");
6469 MODULE_ALIAS("md-raid4");
6470 MODULE_ALIAS("md-level-5");
6471 MODULE_ALIAS("md-level-4");
6472 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6473 MODULE_ALIAS("md-raid6");
6474 MODULE_ALIAS("md-level-6");
6476 /* This used to be two separate modules, they were: */
6477 MODULE_ALIAS("raid5");
6478 MODULE_ALIAS("raid6");