2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
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.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 * use_far_sets (stored in bit 17 of layout )
43 * The data to be stored is divided into chunks using chunksize. Each device
44 * is divided into far_copies sections. In each section, chunks are laid out
45 * in a style similar to raid0, but near_copies copies of each chunk is stored
46 * (each on a different drive). The starting device for each section is offset
47 * near_copies from the starting device of the previous section. Thus there
48 * are (near_copies * far_copies) of each chunk, and each is on a different
49 * drive. near_copies and far_copies must be at least one, and their product
50 * is at most raid_disks.
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of being very far
54 * apart on disk, there are adjacent stripes.
56 * The far and offset algorithms are handled slightly differently if
57 * 'use_far_sets' is true. In this case, the array's devices are grouped into
58 * sets that are (near_copies * far_copies) in size. The far copied stripes
59 * are still shifted by 'near_copies' devices, but this shifting stays confined
60 * to the set rather than the entire array. This is done to improve the number
61 * of device combinations that can fail without causing the array to fail.
62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68 * [A B] [C D] [A B] [C D E]
69 * |...| |...| |...| | ... |
70 * [B A] [D C] [B A] [E C D]
74 * Number of guaranteed r10bios in case of extreme VM load:
76 #define NR_RAID10_BIOS 256
78 /* when we get a read error on a read-only array, we redirect to another
79 * device without failing the first device, or trying to over-write to
80 * correct the read error. To keep track of bad blocks on a per-bio
81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83 #define IO_BLOCKED ((struct bio *)1)
84 /* When we successfully write to a known bad-block, we need to remove the
85 * bad-block marking which must be done from process context. So we record
86 * the success by setting devs[n].bio to IO_MADE_GOOD
88 #define IO_MADE_GOOD ((struct bio *)2)
90 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92 /* When there are this many requests queued to be written by
93 * the raid10 thread, we become 'congested' to provide back-pressure
96 static int max_queued_requests
= 1024;
98 static void allow_barrier(struct r10conf
*conf
);
99 static void lower_barrier(struct r10conf
*conf
);
100 static int _enough(struct r10conf
*conf
, int previous
, int ignore
);
101 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
103 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
);
104 static void end_reshape_write(struct bio
*bio
, int error
);
105 static void end_reshape(struct r10conf
*conf
);
107 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
109 struct r10conf
*conf
= data
;
110 int size
= offsetof(struct r10bio
, devs
[conf
->copies
]);
112 /* allocate a r10bio with room for raid_disks entries in the
114 return kzalloc(size
, gfp_flags
);
117 static void r10bio_pool_free(void *r10_bio
, void *data
)
122 /* Maximum size of each resync request */
123 #define RESYNC_BLOCK_SIZE (64*1024)
124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
125 /* amount of memory to reserve for resync requests */
126 #define RESYNC_WINDOW (1024*1024)
127 /* maximum number of concurrent requests, memory permitting */
128 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
131 * When performing a resync, we need to read and compare, so
132 * we need as many pages are there are copies.
133 * When performing a recovery, we need 2 bios, one for read,
134 * one for write (we recover only one drive per r10buf)
137 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
139 struct r10conf
*conf
= data
;
141 struct r10bio
*r10_bio
;
146 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
150 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
) ||
151 test_bit(MD_RECOVERY_RESHAPE
, &conf
->mddev
->recovery
))
152 nalloc
= conf
->copies
; /* resync */
154 nalloc
= 2; /* recovery */
159 for (j
= nalloc
; j
-- ; ) {
160 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
163 r10_bio
->devs
[j
].bio
= bio
;
164 if (!conf
->have_replacement
)
166 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
169 r10_bio
->devs
[j
].repl_bio
= bio
;
172 * Allocate RESYNC_PAGES data pages and attach them
175 for (j
= 0 ; j
< nalloc
; j
++) {
176 struct bio
*rbio
= r10_bio
->devs
[j
].repl_bio
;
177 bio
= r10_bio
->devs
[j
].bio
;
178 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
179 if (j
> 0 && !test_bit(MD_RECOVERY_SYNC
,
180 &conf
->mddev
->recovery
)) {
181 /* we can share bv_page's during recovery
183 struct bio
*rbio
= r10_bio
->devs
[0].bio
;
184 page
= rbio
->bi_io_vec
[i
].bv_page
;
187 page
= alloc_page(gfp_flags
);
191 bio
->bi_io_vec
[i
].bv_page
= page
;
193 rbio
->bi_io_vec
[i
].bv_page
= page
;
201 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
203 for (i
= 0; i
< RESYNC_PAGES
; i
++)
204 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
207 for ( ; j
< nalloc
; j
++) {
208 if (r10_bio
->devs
[j
].bio
)
209 bio_put(r10_bio
->devs
[j
].bio
);
210 if (r10_bio
->devs
[j
].repl_bio
)
211 bio_put(r10_bio
->devs
[j
].repl_bio
);
213 r10bio_pool_free(r10_bio
, conf
);
217 static void r10buf_pool_free(void *__r10_bio
, void *data
)
220 struct r10conf
*conf
= data
;
221 struct r10bio
*r10bio
= __r10_bio
;
224 for (j
=0; j
< conf
->copies
; j
++) {
225 struct bio
*bio
= r10bio
->devs
[j
].bio
;
227 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
228 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
229 bio
->bi_io_vec
[i
].bv_page
= NULL
;
233 bio
= r10bio
->devs
[j
].repl_bio
;
237 r10bio_pool_free(r10bio
, conf
);
240 static void put_all_bios(struct r10conf
*conf
, struct r10bio
*r10_bio
)
244 for (i
= 0; i
< conf
->copies
; i
++) {
245 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
246 if (!BIO_SPECIAL(*bio
))
249 bio
= &r10_bio
->devs
[i
].repl_bio
;
250 if (r10_bio
->read_slot
< 0 && !BIO_SPECIAL(*bio
))
256 static void free_r10bio(struct r10bio
*r10_bio
)
258 struct r10conf
*conf
= r10_bio
->mddev
->private;
260 put_all_bios(conf
, r10_bio
);
261 mempool_free(r10_bio
, conf
->r10bio_pool
);
264 static void put_buf(struct r10bio
*r10_bio
)
266 struct r10conf
*conf
= r10_bio
->mddev
->private;
268 mempool_free(r10_bio
, conf
->r10buf_pool
);
273 static void reschedule_retry(struct r10bio
*r10_bio
)
276 struct mddev
*mddev
= r10_bio
->mddev
;
277 struct r10conf
*conf
= mddev
->private;
279 spin_lock_irqsave(&conf
->device_lock
, flags
);
280 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
282 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
284 /* wake up frozen array... */
285 wake_up(&conf
->wait_barrier
);
287 md_wakeup_thread(mddev
->thread
);
291 * raid_end_bio_io() is called when we have finished servicing a mirrored
292 * operation and are ready to return a success/failure code to the buffer
295 static void raid_end_bio_io(struct r10bio
*r10_bio
)
297 struct bio
*bio
= r10_bio
->master_bio
;
299 struct r10conf
*conf
= r10_bio
->mddev
->private;
301 if (bio
->bi_phys_segments
) {
303 spin_lock_irqsave(&conf
->device_lock
, flags
);
304 bio
->bi_phys_segments
--;
305 done
= (bio
->bi_phys_segments
== 0);
306 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
309 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
310 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
314 * Wake up any possible resync thread that waits for the device
319 free_r10bio(r10_bio
);
323 * Update disk head position estimator based on IRQ completion info.
325 static inline void update_head_pos(int slot
, struct r10bio
*r10_bio
)
327 struct r10conf
*conf
= r10_bio
->mddev
->private;
329 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
330 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
334 * Find the disk number which triggered given bio
336 static int find_bio_disk(struct r10conf
*conf
, struct r10bio
*r10_bio
,
337 struct bio
*bio
, int *slotp
, int *replp
)
342 for (slot
= 0; slot
< conf
->copies
; slot
++) {
343 if (r10_bio
->devs
[slot
].bio
== bio
)
345 if (r10_bio
->devs
[slot
].repl_bio
== bio
) {
351 BUG_ON(slot
== conf
->copies
);
352 update_head_pos(slot
, r10_bio
);
358 return r10_bio
->devs
[slot
].devnum
;
361 static void raid10_end_read_request(struct bio
*bio
, int error
)
363 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
364 struct r10bio
*r10_bio
= bio
->bi_private
;
366 struct md_rdev
*rdev
;
367 struct r10conf
*conf
= r10_bio
->mddev
->private;
369 slot
= r10_bio
->read_slot
;
370 dev
= r10_bio
->devs
[slot
].devnum
;
371 rdev
= r10_bio
->devs
[slot
].rdev
;
373 * this branch is our 'one mirror IO has finished' event handler:
375 update_head_pos(slot
, r10_bio
);
379 * Set R10BIO_Uptodate in our master bio, so that
380 * we will return a good error code to the higher
381 * levels even if IO on some other mirrored buffer fails.
383 * The 'master' represents the composite IO operation to
384 * user-side. So if something waits for IO, then it will
385 * wait for the 'master' bio.
387 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
389 /* If all other devices that store this block have
390 * failed, we want to return the error upwards rather
391 * than fail the last device. Here we redefine
392 * "uptodate" to mean "Don't want to retry"
394 if (!_enough(conf
, test_bit(R10BIO_Previous
, &r10_bio
->state
),
399 raid_end_bio_io(r10_bio
);
400 rdev_dec_pending(rdev
, conf
->mddev
);
403 * oops, read error - keep the refcount on the rdev
405 char b
[BDEVNAME_SIZE
];
406 printk_ratelimited(KERN_ERR
407 "md/raid10:%s: %s: rescheduling sector %llu\n",
409 bdevname(rdev
->bdev
, b
),
410 (unsigned long long)r10_bio
->sector
);
411 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
412 reschedule_retry(r10_bio
);
416 static void close_write(struct r10bio
*r10_bio
)
418 /* clear the bitmap if all writes complete successfully */
419 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
421 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
423 md_write_end(r10_bio
->mddev
);
426 static void one_write_done(struct r10bio
*r10_bio
)
428 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
429 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
430 reschedule_retry(r10_bio
);
432 close_write(r10_bio
);
433 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
434 reschedule_retry(r10_bio
);
436 raid_end_bio_io(r10_bio
);
441 static void raid10_end_write_request(struct bio
*bio
, int error
)
443 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
444 struct r10bio
*r10_bio
= bio
->bi_private
;
447 struct r10conf
*conf
= r10_bio
->mddev
->private;
449 struct md_rdev
*rdev
= NULL
;
451 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
454 rdev
= conf
->mirrors
[dev
].replacement
;
458 rdev
= conf
->mirrors
[dev
].rdev
;
461 * this branch is our 'one mirror IO has finished' event handler:
465 /* Never record new bad blocks to replacement,
468 md_error(rdev
->mddev
, rdev
);
470 set_bit(WriteErrorSeen
, &rdev
->flags
);
471 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
472 set_bit(MD_RECOVERY_NEEDED
,
473 &rdev
->mddev
->recovery
);
474 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
479 * Set R10BIO_Uptodate in our master bio, so that
480 * we will return a good error code for to the higher
481 * levels even if IO on some other mirrored buffer fails.
483 * The 'master' represents the composite IO operation to
484 * user-side. So if something waits for IO, then it will
485 * wait for the 'master' bio.
491 * Do not set R10BIO_Uptodate if the current device is
492 * rebuilding or Faulty. This is because we cannot use
493 * such device for properly reading the data back (we could
494 * potentially use it, if the current write would have felt
495 * before rdev->recovery_offset, but for simplicity we don't
498 if (test_bit(In_sync
, &rdev
->flags
) &&
499 !test_bit(Faulty
, &rdev
->flags
))
500 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
502 /* Maybe we can clear some bad blocks. */
503 if (is_badblock(rdev
,
504 r10_bio
->devs
[slot
].addr
,
506 &first_bad
, &bad_sectors
)) {
509 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
511 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
513 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
519 * Let's see if all mirrored write operations have finished
522 one_write_done(r10_bio
);
524 rdev_dec_pending(rdev
, conf
->mddev
);
528 * RAID10 layout manager
529 * As well as the chunksize and raid_disks count, there are two
530 * parameters: near_copies and far_copies.
531 * near_copies * far_copies must be <= raid_disks.
532 * Normally one of these will be 1.
533 * If both are 1, we get raid0.
534 * If near_copies == raid_disks, we get raid1.
536 * Chunks are laid out in raid0 style with near_copies copies of the
537 * first chunk, followed by near_copies copies of the next chunk and
539 * If far_copies > 1, then after 1/far_copies of the array has been assigned
540 * as described above, we start again with a device offset of near_copies.
541 * So we effectively have another copy of the whole array further down all
542 * the drives, but with blocks on different drives.
543 * With this layout, and block is never stored twice on the one device.
545 * raid10_find_phys finds the sector offset of a given virtual sector
546 * on each device that it is on.
548 * raid10_find_virt does the reverse mapping, from a device and a
549 * sector offset to a virtual address
552 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
560 int last_far_set_start
, last_far_set_size
;
562 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
563 last_far_set_start
*= geo
->far_set_size
;
565 last_far_set_size
= geo
->far_set_size
;
566 last_far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
568 /* now calculate first sector/dev */
569 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
570 sector
= r10bio
->sector
& geo
->chunk_mask
;
572 chunk
*= geo
->near_copies
;
574 dev
= sector_div(stripe
, geo
->raid_disks
);
576 stripe
*= geo
->far_copies
;
578 sector
+= stripe
<< geo
->chunk_shift
;
580 /* and calculate all the others */
581 for (n
= 0; n
< geo
->near_copies
; n
++) {
585 r10bio
->devs
[slot
].devnum
= d
;
586 r10bio
->devs
[slot
].addr
= s
;
589 for (f
= 1; f
< geo
->far_copies
; f
++) {
590 set
= d
/ geo
->far_set_size
;
591 d
+= geo
->near_copies
;
593 if ((geo
->raid_disks
% geo
->far_set_size
) &&
594 (d
> last_far_set_start
)) {
595 d
-= last_far_set_start
;
596 d
%= last_far_set_size
;
597 d
+= last_far_set_start
;
599 d
%= geo
->far_set_size
;
600 d
+= geo
->far_set_size
* set
;
603 r10bio
->devs
[slot
].devnum
= d
;
604 r10bio
->devs
[slot
].addr
= s
;
608 if (dev
>= geo
->raid_disks
) {
610 sector
+= (geo
->chunk_mask
+ 1);
615 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
617 struct geom
*geo
= &conf
->geo
;
619 if (conf
->reshape_progress
!= MaxSector
&&
620 ((r10bio
->sector
>= conf
->reshape_progress
) !=
621 conf
->mddev
->reshape_backwards
)) {
622 set_bit(R10BIO_Previous
, &r10bio
->state
);
625 clear_bit(R10BIO_Previous
, &r10bio
->state
);
627 __raid10_find_phys(geo
, r10bio
);
630 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
632 sector_t offset
, chunk
, vchunk
;
633 /* Never use conf->prev as this is only called during resync
634 * or recovery, so reshape isn't happening
636 struct geom
*geo
= &conf
->geo
;
637 int far_set_start
= (dev
/ geo
->far_set_size
) * geo
->far_set_size
;
638 int far_set_size
= geo
->far_set_size
;
639 int last_far_set_start
;
641 if (geo
->raid_disks
% geo
->far_set_size
) {
642 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
643 last_far_set_start
*= geo
->far_set_size
;
645 if (dev
>= last_far_set_start
) {
646 far_set_size
= geo
->far_set_size
;
647 far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
648 far_set_start
= last_far_set_start
;
652 offset
= sector
& geo
->chunk_mask
;
653 if (geo
->far_offset
) {
655 chunk
= sector
>> geo
->chunk_shift
;
656 fc
= sector_div(chunk
, geo
->far_copies
);
657 dev
-= fc
* geo
->near_copies
;
658 if (dev
< far_set_start
)
661 while (sector
>= geo
->stride
) {
662 sector
-= geo
->stride
;
663 if (dev
< (geo
->near_copies
+ far_set_start
))
664 dev
+= far_set_size
- geo
->near_copies
;
666 dev
-= geo
->near_copies
;
668 chunk
= sector
>> geo
->chunk_shift
;
670 vchunk
= chunk
* geo
->raid_disks
+ dev
;
671 sector_div(vchunk
, geo
->near_copies
);
672 return (vchunk
<< geo
->chunk_shift
) + offset
;
676 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
678 * @bvm: properties of new bio
679 * @biovec: the request that could be merged to it.
681 * Return amount of bytes we can accept at this offset
682 * This requires checking for end-of-chunk if near_copies != raid_disks,
683 * and for subordinate merge_bvec_fns if merge_check_needed.
685 static int raid10_mergeable_bvec(struct request_queue
*q
,
686 struct bvec_merge_data
*bvm
,
687 struct bio_vec
*biovec
)
689 struct mddev
*mddev
= q
->queuedata
;
690 struct r10conf
*conf
= mddev
->private;
691 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
693 unsigned int chunk_sectors
;
694 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
695 struct geom
*geo
= &conf
->geo
;
697 chunk_sectors
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
) + 1;
698 if (conf
->reshape_progress
!= MaxSector
&&
699 ((sector
>= conf
->reshape_progress
) !=
700 conf
->mddev
->reshape_backwards
))
703 if (geo
->near_copies
< geo
->raid_disks
) {
704 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1))
705 + bio_sectors
)) << 9;
707 /* bio_add cannot handle a negative return */
709 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
710 return biovec
->bv_len
;
712 max
= biovec
->bv_len
;
714 if (mddev
->merge_check_needed
) {
716 struct r10bio r10_bio
;
717 struct r10dev devs
[conf
->copies
];
719 struct r10bio
*r10_bio
= &on_stack
.r10_bio
;
721 if (conf
->reshape_progress
!= MaxSector
) {
722 /* Cannot give any guidance during reshape */
723 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
724 return biovec
->bv_len
;
727 r10_bio
->sector
= sector
;
728 raid10_find_phys(conf
, r10_bio
);
730 for (s
= 0; s
< conf
->copies
; s
++) {
731 int disk
= r10_bio
->devs
[s
].devnum
;
732 struct md_rdev
*rdev
= rcu_dereference(
733 conf
->mirrors
[disk
].rdev
);
734 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
735 struct request_queue
*q
=
736 bdev_get_queue(rdev
->bdev
);
737 if (q
->merge_bvec_fn
) {
738 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
740 bvm
->bi_bdev
= rdev
->bdev
;
741 max
= min(max
, q
->merge_bvec_fn(
745 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
746 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
747 struct request_queue
*q
=
748 bdev_get_queue(rdev
->bdev
);
749 if (q
->merge_bvec_fn
) {
750 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
752 bvm
->bi_bdev
= rdev
->bdev
;
753 max
= min(max
, q
->merge_bvec_fn(
764 * This routine returns the disk from which the requested read should
765 * be done. There is a per-array 'next expected sequential IO' sector
766 * number - if this matches on the next IO then we use the last disk.
767 * There is also a per-disk 'last know head position' sector that is
768 * maintained from IRQ contexts, both the normal and the resync IO
769 * completion handlers update this position correctly. If there is no
770 * perfect sequential match then we pick the disk whose head is closest.
772 * If there are 2 mirrors in the same 2 devices, performance degrades
773 * because position is mirror, not device based.
775 * The rdev for the device selected will have nr_pending incremented.
779 * FIXME: possibly should rethink readbalancing and do it differently
780 * depending on near_copies / far_copies geometry.
782 static struct md_rdev
*read_balance(struct r10conf
*conf
,
783 struct r10bio
*r10_bio
,
786 const sector_t this_sector
= r10_bio
->sector
;
788 int sectors
= r10_bio
->sectors
;
789 int best_good_sectors
;
790 sector_t new_distance
, best_dist
;
791 struct md_rdev
*best_rdev
, *rdev
= NULL
;
794 struct geom
*geo
= &conf
->geo
;
796 raid10_find_phys(conf
, r10_bio
);
799 sectors
= r10_bio
->sectors
;
802 best_dist
= MaxSector
;
803 best_good_sectors
= 0;
806 * Check if we can balance. We can balance on the whole
807 * device if no resync is going on (recovery is ok), or below
808 * the resync window. We take the first readable disk when
809 * above the resync window.
811 if (conf
->mddev
->recovery_cp
< MaxSector
812 && (this_sector
+ sectors
>= conf
->next_resync
))
815 for (slot
= 0; slot
< conf
->copies
; slot
++) {
820 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
822 disk
= r10_bio
->devs
[slot
].devnum
;
823 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
824 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
825 test_bit(Unmerged
, &rdev
->flags
) ||
826 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
827 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
829 test_bit(Faulty
, &rdev
->flags
) ||
830 test_bit(Unmerged
, &rdev
->flags
))
832 if (!test_bit(In_sync
, &rdev
->flags
) &&
833 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
836 dev_sector
= r10_bio
->devs
[slot
].addr
;
837 if (is_badblock(rdev
, dev_sector
, sectors
,
838 &first_bad
, &bad_sectors
)) {
839 if (best_dist
< MaxSector
)
840 /* Already have a better slot */
842 if (first_bad
<= dev_sector
) {
843 /* Cannot read here. If this is the
844 * 'primary' device, then we must not read
845 * beyond 'bad_sectors' from another device.
847 bad_sectors
-= (dev_sector
- first_bad
);
848 if (!do_balance
&& sectors
> bad_sectors
)
849 sectors
= bad_sectors
;
850 if (best_good_sectors
> sectors
)
851 best_good_sectors
= sectors
;
853 sector_t good_sectors
=
854 first_bad
- dev_sector
;
855 if (good_sectors
> best_good_sectors
) {
856 best_good_sectors
= good_sectors
;
861 /* Must read from here */
866 best_good_sectors
= sectors
;
871 /* This optimisation is debatable, and completely destroys
872 * sequential read speed for 'far copies' arrays. So only
873 * keep it for 'near' arrays, and review those later.
875 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
878 /* for far > 1 always use the lowest address */
879 if (geo
->far_copies
> 1)
880 new_distance
= r10_bio
->devs
[slot
].addr
;
882 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
883 conf
->mirrors
[disk
].head_position
);
884 if (new_distance
< best_dist
) {
885 best_dist
= new_distance
;
890 if (slot
>= conf
->copies
) {
896 atomic_inc(&rdev
->nr_pending
);
897 if (test_bit(Faulty
, &rdev
->flags
)) {
898 /* Cannot risk returning a device that failed
899 * before we inc'ed nr_pending
901 rdev_dec_pending(rdev
, conf
->mddev
);
904 r10_bio
->read_slot
= slot
;
908 *max_sectors
= best_good_sectors
;
913 int md_raid10_congested(struct mddev
*mddev
, int bits
)
915 struct r10conf
*conf
= mddev
->private;
918 if ((bits
& (1 << BDI_async_congested
)) &&
919 conf
->pending_count
>= max_queued_requests
)
924 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
927 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
928 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
929 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
931 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
937 EXPORT_SYMBOL_GPL(md_raid10_congested
);
939 static int raid10_congested(void *data
, int bits
)
941 struct mddev
*mddev
= data
;
943 return mddev_congested(mddev
, bits
) ||
944 md_raid10_congested(mddev
, bits
);
947 static void flush_pending_writes(struct r10conf
*conf
)
949 /* Any writes that have been queued but are awaiting
950 * bitmap updates get flushed here.
952 spin_lock_irq(&conf
->device_lock
);
954 if (conf
->pending_bio_list
.head
) {
956 bio
= bio_list_get(&conf
->pending_bio_list
);
957 conf
->pending_count
= 0;
958 spin_unlock_irq(&conf
->device_lock
);
959 /* flush any pending bitmap writes to disk
960 * before proceeding w/ I/O */
961 bitmap_unplug(conf
->mddev
->bitmap
);
962 wake_up(&conf
->wait_barrier
);
964 while (bio
) { /* submit pending writes */
965 struct bio
*next
= bio
->bi_next
;
967 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
968 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
972 generic_make_request(bio
);
976 spin_unlock_irq(&conf
->device_lock
);
980 * Sometimes we need to suspend IO while we do something else,
981 * either some resync/recovery, or reconfigure the array.
982 * To do this we raise a 'barrier'.
983 * The 'barrier' is a counter that can be raised multiple times
984 * to count how many activities are happening which preclude
986 * We can only raise the barrier if there is no pending IO.
987 * i.e. if nr_pending == 0.
988 * We choose only to raise the barrier if no-one is waiting for the
989 * barrier to go down. This means that as soon as an IO request
990 * is ready, no other operations which require a barrier will start
991 * until the IO request has had a chance.
993 * So: regular IO calls 'wait_barrier'. When that returns there
994 * is no backgroup IO happening, It must arrange to call
995 * allow_barrier when it has finished its IO.
996 * backgroup IO calls must call raise_barrier. Once that returns
997 * there is no normal IO happeing. It must arrange to call
998 * lower_barrier when the particular background IO completes.
1001 static void raise_barrier(struct r10conf
*conf
, int force
)
1003 BUG_ON(force
&& !conf
->barrier
);
1004 spin_lock_irq(&conf
->resync_lock
);
1006 /* Wait until no block IO is waiting (unless 'force') */
1007 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
1010 /* block any new IO from starting */
1013 /* Now wait for all pending IO to complete */
1014 wait_event_lock_irq(conf
->wait_barrier
,
1015 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
1018 spin_unlock_irq(&conf
->resync_lock
);
1021 static void lower_barrier(struct r10conf
*conf
)
1023 unsigned long flags
;
1024 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1026 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1027 wake_up(&conf
->wait_barrier
);
1030 static void wait_barrier(struct r10conf
*conf
)
1032 spin_lock_irq(&conf
->resync_lock
);
1033 if (conf
->barrier
) {
1035 /* Wait for the barrier to drop.
1036 * However if there are already pending
1037 * requests (preventing the barrier from
1038 * rising completely), and the
1039 * pre-process bio queue isn't empty,
1040 * then don't wait, as we need to empty
1041 * that queue to get the nr_pending
1044 wait_event_lock_irq(conf
->wait_barrier
,
1046 (conf
->nr_pending
&&
1047 current
->bio_list
&&
1048 !bio_list_empty(current
->bio_list
)),
1053 spin_unlock_irq(&conf
->resync_lock
);
1056 static void allow_barrier(struct r10conf
*conf
)
1058 unsigned long flags
;
1059 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1061 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1062 wake_up(&conf
->wait_barrier
);
1065 static void freeze_array(struct r10conf
*conf
, int extra
)
1067 /* stop syncio and normal IO and wait for everything to
1069 * We increment barrier and nr_waiting, and then
1070 * wait until nr_pending match nr_queued+extra
1071 * This is called in the context of one normal IO request
1072 * that has failed. Thus any sync request that might be pending
1073 * will be blocked by nr_pending, and we need to wait for
1074 * pending IO requests to complete or be queued for re-try.
1075 * Thus the number queued (nr_queued) plus this request (extra)
1076 * must match the number of pending IOs (nr_pending) before
1079 spin_lock_irq(&conf
->resync_lock
);
1082 wait_event_lock_irq_cmd(conf
->wait_barrier
,
1083 conf
->nr_pending
== conf
->nr_queued
+extra
,
1085 flush_pending_writes(conf
));
1087 spin_unlock_irq(&conf
->resync_lock
);
1090 static void unfreeze_array(struct r10conf
*conf
)
1092 /* reverse the effect of the freeze */
1093 spin_lock_irq(&conf
->resync_lock
);
1096 wake_up(&conf
->wait_barrier
);
1097 spin_unlock_irq(&conf
->resync_lock
);
1100 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1101 struct md_rdev
*rdev
)
1103 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1104 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1105 return rdev
->data_offset
;
1107 return rdev
->new_data_offset
;
1110 struct raid10_plug_cb
{
1111 struct blk_plug_cb cb
;
1112 struct bio_list pending
;
1116 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1118 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1120 struct mddev
*mddev
= plug
->cb
.data
;
1121 struct r10conf
*conf
= mddev
->private;
1124 if (from_schedule
|| current
->bio_list
) {
1125 spin_lock_irq(&conf
->device_lock
);
1126 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1127 conf
->pending_count
+= plug
->pending_cnt
;
1128 spin_unlock_irq(&conf
->device_lock
);
1129 wake_up(&conf
->wait_barrier
);
1130 md_wakeup_thread(mddev
->thread
);
1135 /* we aren't scheduling, so we can do the write-out directly. */
1136 bio
= bio_list_get(&plug
->pending
);
1137 bitmap_unplug(mddev
->bitmap
);
1138 wake_up(&conf
->wait_barrier
);
1140 while (bio
) { /* submit pending writes */
1141 struct bio
*next
= bio
->bi_next
;
1142 bio
->bi_next
= NULL
;
1143 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
1144 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1145 /* Just ignore it */
1148 generic_make_request(bio
);
1154 static void __make_request(struct mddev
*mddev
, struct bio
*bio
)
1156 struct r10conf
*conf
= mddev
->private;
1157 struct r10bio
*r10_bio
;
1158 struct bio
*read_bio
;
1160 const int rw
= bio_data_dir(bio
);
1161 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1162 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1163 const unsigned long do_discard
= (bio
->bi_rw
1164 & (REQ_DISCARD
| REQ_SECURE
));
1165 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1166 unsigned long flags
;
1167 struct md_rdev
*blocked_rdev
;
1168 struct blk_plug_cb
*cb
;
1169 struct raid10_plug_cb
*plug
= NULL
;
1170 int sectors_handled
;
1175 * Register the new request and wait if the reconstruction
1176 * thread has put up a bar for new requests.
1177 * Continue immediately if no resync is active currently.
1181 sectors
= bio_sectors(bio
);
1182 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1183 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1184 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1185 /* IO spans the reshape position. Need to wait for
1188 allow_barrier(conf
);
1189 wait_event(conf
->wait_barrier
,
1190 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1191 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1195 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1196 bio_data_dir(bio
) == WRITE
&&
1197 (mddev
->reshape_backwards
1198 ? (bio
->bi_iter
.bi_sector
< conf
->reshape_safe
&&
1199 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
)
1200 : (bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_safe
&&
1201 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
))) {
1202 /* Need to update reshape_position in metadata */
1203 mddev
->reshape_position
= conf
->reshape_progress
;
1204 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1205 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1206 md_wakeup_thread(mddev
->thread
);
1207 wait_event(mddev
->sb_wait
,
1208 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1210 conf
->reshape_safe
= mddev
->reshape_position
;
1213 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1215 r10_bio
->master_bio
= bio
;
1216 r10_bio
->sectors
= sectors
;
1218 r10_bio
->mddev
= mddev
;
1219 r10_bio
->sector
= bio
->bi_iter
.bi_sector
;
1222 /* We might need to issue multiple reads to different
1223 * devices if there are bad blocks around, so we keep
1224 * track of the number of reads in bio->bi_phys_segments.
1225 * If this is 0, there is only one r10_bio and no locking
1226 * will be needed when the request completes. If it is
1227 * non-zero, then it is the number of not-completed requests.
1229 bio
->bi_phys_segments
= 0;
1230 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1234 * read balancing logic:
1236 struct md_rdev
*rdev
;
1240 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1242 raid_end_bio_io(r10_bio
);
1245 slot
= r10_bio
->read_slot
;
1247 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1248 bio_trim(read_bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1251 r10_bio
->devs
[slot
].bio
= read_bio
;
1252 r10_bio
->devs
[slot
].rdev
= rdev
;
1254 read_bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
+
1255 choose_data_offset(r10_bio
, rdev
);
1256 read_bio
->bi_bdev
= rdev
->bdev
;
1257 read_bio
->bi_end_io
= raid10_end_read_request
;
1258 read_bio
->bi_rw
= READ
| do_sync
;
1259 read_bio
->bi_private
= r10_bio
;
1261 if (max_sectors
< r10_bio
->sectors
) {
1262 /* Could not read all from this device, so we will
1263 * need another r10_bio.
1265 sectors_handled
= (r10_bio
->sector
+ max_sectors
1266 - bio
->bi_iter
.bi_sector
);
1267 r10_bio
->sectors
= max_sectors
;
1268 spin_lock_irq(&conf
->device_lock
);
1269 if (bio
->bi_phys_segments
== 0)
1270 bio
->bi_phys_segments
= 2;
1272 bio
->bi_phys_segments
++;
1273 spin_unlock_irq(&conf
->device_lock
);
1274 /* Cannot call generic_make_request directly
1275 * as that will be queued in __generic_make_request
1276 * and subsequent mempool_alloc might block
1277 * waiting for it. so hand bio over to raid10d.
1279 reschedule_retry(r10_bio
);
1281 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1283 r10_bio
->master_bio
= bio
;
1284 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1286 r10_bio
->mddev
= mddev
;
1287 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+
1291 generic_make_request(read_bio
);
1298 if (conf
->pending_count
>= max_queued_requests
) {
1299 md_wakeup_thread(mddev
->thread
);
1300 wait_event(conf
->wait_barrier
,
1301 conf
->pending_count
< max_queued_requests
);
1303 /* first select target devices under rcu_lock and
1304 * inc refcount on their rdev. Record them by setting
1306 * If there are known/acknowledged bad blocks on any device
1307 * on which we have seen a write error, we want to avoid
1308 * writing to those blocks. This potentially requires several
1309 * writes to write around the bad blocks. Each set of writes
1310 * gets its own r10_bio with a set of bios attached. The number
1311 * of r10_bios is recored in bio->bi_phys_segments just as with
1315 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1316 raid10_find_phys(conf
, r10_bio
);
1318 blocked_rdev
= NULL
;
1320 max_sectors
= r10_bio
->sectors
;
1322 for (i
= 0; i
< conf
->copies
; i
++) {
1323 int d
= r10_bio
->devs
[i
].devnum
;
1324 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1325 struct md_rdev
*rrdev
= rcu_dereference(
1326 conf
->mirrors
[d
].replacement
);
1329 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1330 atomic_inc(&rdev
->nr_pending
);
1331 blocked_rdev
= rdev
;
1334 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1335 atomic_inc(&rrdev
->nr_pending
);
1336 blocked_rdev
= rrdev
;
1339 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)
1340 || test_bit(Unmerged
, &rdev
->flags
)))
1342 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1343 || test_bit(Unmerged
, &rrdev
->flags
)))
1346 r10_bio
->devs
[i
].bio
= NULL
;
1347 r10_bio
->devs
[i
].repl_bio
= NULL
;
1349 if (!rdev
&& !rrdev
) {
1350 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1353 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1355 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1359 is_bad
= is_badblock(rdev
, dev_sector
,
1361 &first_bad
, &bad_sectors
);
1363 /* Mustn't write here until the bad block
1366 atomic_inc(&rdev
->nr_pending
);
1367 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1368 blocked_rdev
= rdev
;
1371 if (is_bad
&& first_bad
<= dev_sector
) {
1372 /* Cannot write here at all */
1373 bad_sectors
-= (dev_sector
- first_bad
);
1374 if (bad_sectors
< max_sectors
)
1375 /* Mustn't write more than bad_sectors
1376 * to other devices yet
1378 max_sectors
= bad_sectors
;
1379 /* We don't set R10BIO_Degraded as that
1380 * only applies if the disk is missing,
1381 * so it might be re-added, and we want to
1382 * know to recover this chunk.
1383 * In this case the device is here, and the
1384 * fact that this chunk is not in-sync is
1385 * recorded in the bad block log.
1390 int good_sectors
= first_bad
- dev_sector
;
1391 if (good_sectors
< max_sectors
)
1392 max_sectors
= good_sectors
;
1396 r10_bio
->devs
[i
].bio
= bio
;
1397 atomic_inc(&rdev
->nr_pending
);
1400 r10_bio
->devs
[i
].repl_bio
= bio
;
1401 atomic_inc(&rrdev
->nr_pending
);
1406 if (unlikely(blocked_rdev
)) {
1407 /* Have to wait for this device to get unblocked, then retry */
1411 for (j
= 0; j
< i
; j
++) {
1412 if (r10_bio
->devs
[j
].bio
) {
1413 d
= r10_bio
->devs
[j
].devnum
;
1414 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1416 if (r10_bio
->devs
[j
].repl_bio
) {
1417 struct md_rdev
*rdev
;
1418 d
= r10_bio
->devs
[j
].devnum
;
1419 rdev
= conf
->mirrors
[d
].replacement
;
1421 /* Race with remove_disk */
1423 rdev
= conf
->mirrors
[d
].rdev
;
1425 rdev_dec_pending(rdev
, mddev
);
1428 allow_barrier(conf
);
1429 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1434 if (max_sectors
< r10_bio
->sectors
) {
1435 /* We are splitting this into multiple parts, so
1436 * we need to prepare for allocating another r10_bio.
1438 r10_bio
->sectors
= max_sectors
;
1439 spin_lock_irq(&conf
->device_lock
);
1440 if (bio
->bi_phys_segments
== 0)
1441 bio
->bi_phys_segments
= 2;
1443 bio
->bi_phys_segments
++;
1444 spin_unlock_irq(&conf
->device_lock
);
1446 sectors_handled
= r10_bio
->sector
+ max_sectors
-
1447 bio
->bi_iter
.bi_sector
;
1449 atomic_set(&r10_bio
->remaining
, 1);
1450 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1452 for (i
= 0; i
< conf
->copies
; i
++) {
1454 int d
= r10_bio
->devs
[i
].devnum
;
1455 if (r10_bio
->devs
[i
].bio
) {
1456 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1457 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1458 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1460 r10_bio
->devs
[i
].bio
= mbio
;
1462 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1463 choose_data_offset(r10_bio
,
1465 mbio
->bi_bdev
= rdev
->bdev
;
1466 mbio
->bi_end_io
= raid10_end_write_request
;
1468 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1469 mbio
->bi_private
= r10_bio
;
1471 atomic_inc(&r10_bio
->remaining
);
1473 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1476 plug
= container_of(cb
, struct raid10_plug_cb
,
1480 spin_lock_irqsave(&conf
->device_lock
, flags
);
1482 bio_list_add(&plug
->pending
, mbio
);
1483 plug
->pending_cnt
++;
1485 bio_list_add(&conf
->pending_bio_list
, mbio
);
1486 conf
->pending_count
++;
1488 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1490 md_wakeup_thread(mddev
->thread
);
1493 if (r10_bio
->devs
[i
].repl_bio
) {
1494 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1496 /* Replacement just got moved to main 'rdev' */
1498 rdev
= conf
->mirrors
[d
].rdev
;
1500 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1501 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1503 r10_bio
->devs
[i
].repl_bio
= mbio
;
1505 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1508 mbio
->bi_bdev
= rdev
->bdev
;
1509 mbio
->bi_end_io
= raid10_end_write_request
;
1511 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1512 mbio
->bi_private
= r10_bio
;
1514 atomic_inc(&r10_bio
->remaining
);
1515 spin_lock_irqsave(&conf
->device_lock
, flags
);
1516 bio_list_add(&conf
->pending_bio_list
, mbio
);
1517 conf
->pending_count
++;
1518 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1519 if (!mddev_check_plugged(mddev
))
1520 md_wakeup_thread(mddev
->thread
);
1524 /* Don't remove the bias on 'remaining' (one_write_done) until
1525 * after checking if we need to go around again.
1528 if (sectors_handled
< bio_sectors(bio
)) {
1529 one_write_done(r10_bio
);
1530 /* We need another r10_bio. It has already been counted
1531 * in bio->bi_phys_segments.
1533 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1535 r10_bio
->master_bio
= bio
;
1536 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1538 r10_bio
->mddev
= mddev
;
1539 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1543 one_write_done(r10_bio
);
1546 static void make_request(struct mddev
*mddev
, struct bio
*bio
)
1548 struct r10conf
*conf
= mddev
->private;
1549 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1550 int chunk_sects
= chunk_mask
+ 1;
1554 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1555 md_flush_request(mddev
, bio
);
1559 md_write_start(mddev
, bio
);
1564 * If this request crosses a chunk boundary, we need to split
1567 if (unlikely((bio
->bi_iter
.bi_sector
& chunk_mask
) +
1568 bio_sectors(bio
) > chunk_sects
1569 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1570 || conf
->prev
.near_copies
<
1571 conf
->prev
.raid_disks
))) {
1572 split
= bio_split(bio
, chunk_sects
-
1573 (bio
->bi_iter
.bi_sector
&
1575 GFP_NOIO
, fs_bio_set
);
1576 bio_chain(split
, bio
);
1581 __make_request(mddev
, split
);
1582 } while (split
!= bio
);
1584 /* In case raid10d snuck in to freeze_array */
1585 wake_up(&conf
->wait_barrier
);
1588 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1590 struct r10conf
*conf
= mddev
->private;
1593 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1594 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1595 if (conf
->geo
.near_copies
> 1)
1596 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1597 if (conf
->geo
.far_copies
> 1) {
1598 if (conf
->geo
.far_offset
)
1599 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1601 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1603 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1604 conf
->geo
.raid_disks
- mddev
->degraded
);
1605 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1606 seq_printf(seq
, "%s",
1607 conf
->mirrors
[i
].rdev
&&
1608 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1609 seq_printf(seq
, "]");
1612 /* check if there are enough drives for
1613 * every block to appear on atleast one.
1614 * Don't consider the device numbered 'ignore'
1615 * as we might be about to remove it.
1617 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1623 disks
= conf
->prev
.raid_disks
;
1624 ncopies
= conf
->prev
.near_copies
;
1626 disks
= conf
->geo
.raid_disks
;
1627 ncopies
= conf
->geo
.near_copies
;
1632 int n
= conf
->copies
;
1636 struct md_rdev
*rdev
;
1637 if (this != ignore
&&
1638 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1639 test_bit(In_sync
, &rdev
->flags
))
1641 this = (this+1) % disks
;
1645 first
= (first
+ ncopies
) % disks
;
1646 } while (first
!= 0);
1653 static int enough(struct r10conf
*conf
, int ignore
)
1655 /* when calling 'enough', both 'prev' and 'geo' must
1657 * This is ensured if ->reconfig_mutex or ->device_lock
1660 return _enough(conf
, 0, ignore
) &&
1661 _enough(conf
, 1, ignore
);
1664 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1666 char b
[BDEVNAME_SIZE
];
1667 struct r10conf
*conf
= mddev
->private;
1668 unsigned long flags
;
1671 * If it is not operational, then we have already marked it as dead
1672 * else if it is the last working disks, ignore the error, let the
1673 * next level up know.
1674 * else mark the drive as failed
1676 spin_lock_irqsave(&conf
->device_lock
, flags
);
1677 if (test_bit(In_sync
, &rdev
->flags
)
1678 && !enough(conf
, rdev
->raid_disk
)) {
1680 * Don't fail the drive, just return an IO error.
1682 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1685 if (test_and_clear_bit(In_sync
, &rdev
->flags
))
1688 * If recovery is running, make sure it aborts.
1690 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1691 set_bit(Blocked
, &rdev
->flags
);
1692 set_bit(Faulty
, &rdev
->flags
);
1693 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1694 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1696 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1697 "md/raid10:%s: Operation continuing on %d devices.\n",
1698 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1699 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1702 static void print_conf(struct r10conf
*conf
)
1705 struct raid10_info
*tmp
;
1707 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1709 printk(KERN_DEBUG
"(!conf)\n");
1712 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1713 conf
->geo
.raid_disks
);
1715 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1716 char b
[BDEVNAME_SIZE
];
1717 tmp
= conf
->mirrors
+ i
;
1719 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1720 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1721 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1722 bdevname(tmp
->rdev
->bdev
,b
));
1726 static void close_sync(struct r10conf
*conf
)
1729 allow_barrier(conf
);
1731 mempool_destroy(conf
->r10buf_pool
);
1732 conf
->r10buf_pool
= NULL
;
1735 static int raid10_spare_active(struct mddev
*mddev
)
1738 struct r10conf
*conf
= mddev
->private;
1739 struct raid10_info
*tmp
;
1741 unsigned long flags
;
1744 * Find all non-in_sync disks within the RAID10 configuration
1745 * and mark them in_sync
1747 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1748 tmp
= conf
->mirrors
+ i
;
1749 if (tmp
->replacement
1750 && tmp
->replacement
->recovery_offset
== MaxSector
1751 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1752 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1753 /* Replacement has just become active */
1755 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1758 /* Replaced device not technically faulty,
1759 * but we need to be sure it gets removed
1760 * and never re-added.
1762 set_bit(Faulty
, &tmp
->rdev
->flags
);
1763 sysfs_notify_dirent_safe(
1764 tmp
->rdev
->sysfs_state
);
1766 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1767 } else if (tmp
->rdev
1768 && tmp
->rdev
->recovery_offset
== MaxSector
1769 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1770 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1772 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1775 spin_lock_irqsave(&conf
->device_lock
, flags
);
1776 mddev
->degraded
-= count
;
1777 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1783 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1785 struct r10conf
*conf
= mddev
->private;
1789 int last
= conf
->geo
.raid_disks
- 1;
1790 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1792 if (mddev
->recovery_cp
< MaxSector
)
1793 /* only hot-add to in-sync arrays, as recovery is
1794 * very different from resync
1797 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1800 if (rdev
->raid_disk
>= 0)
1801 first
= last
= rdev
->raid_disk
;
1803 if (q
->merge_bvec_fn
) {
1804 set_bit(Unmerged
, &rdev
->flags
);
1805 mddev
->merge_check_needed
= 1;
1808 if (rdev
->saved_raid_disk
>= first
&&
1809 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1810 mirror
= rdev
->saved_raid_disk
;
1813 for ( ; mirror
<= last
; mirror
++) {
1814 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1815 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1818 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1819 p
->replacement
!= NULL
)
1821 clear_bit(In_sync
, &rdev
->flags
);
1822 set_bit(Replacement
, &rdev
->flags
);
1823 rdev
->raid_disk
= mirror
;
1826 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1827 rdev
->data_offset
<< 9);
1829 rcu_assign_pointer(p
->replacement
, rdev
);
1834 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1835 rdev
->data_offset
<< 9);
1837 p
->head_position
= 0;
1838 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1839 rdev
->raid_disk
= mirror
;
1841 if (rdev
->saved_raid_disk
!= mirror
)
1843 rcu_assign_pointer(p
->rdev
, rdev
);
1846 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1847 /* Some requests might not have seen this new
1848 * merge_bvec_fn. We must wait for them to complete
1849 * before merging the device fully.
1850 * First we make sure any code which has tested
1851 * our function has submitted the request, then
1852 * we wait for all outstanding requests to complete.
1854 synchronize_sched();
1855 freeze_array(conf
, 0);
1856 unfreeze_array(conf
);
1857 clear_bit(Unmerged
, &rdev
->flags
);
1859 md_integrity_add_rdev(rdev
, mddev
);
1860 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1861 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1867 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1869 struct r10conf
*conf
= mddev
->private;
1871 int number
= rdev
->raid_disk
;
1872 struct md_rdev
**rdevp
;
1873 struct raid10_info
*p
= conf
->mirrors
+ number
;
1876 if (rdev
== p
->rdev
)
1878 else if (rdev
== p
->replacement
)
1879 rdevp
= &p
->replacement
;
1883 if (test_bit(In_sync
, &rdev
->flags
) ||
1884 atomic_read(&rdev
->nr_pending
)) {
1888 /* Only remove faulty devices if recovery
1891 if (!test_bit(Faulty
, &rdev
->flags
) &&
1892 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1893 (!p
->replacement
|| p
->replacement
== rdev
) &&
1894 number
< conf
->geo
.raid_disks
&&
1901 if (atomic_read(&rdev
->nr_pending
)) {
1902 /* lost the race, try later */
1906 } else if (p
->replacement
) {
1907 /* We must have just cleared 'rdev' */
1908 p
->rdev
= p
->replacement
;
1909 clear_bit(Replacement
, &p
->replacement
->flags
);
1910 smp_mb(); /* Make sure other CPUs may see both as identical
1911 * but will never see neither -- if they are careful.
1913 p
->replacement
= NULL
;
1914 clear_bit(WantReplacement
, &rdev
->flags
);
1916 /* We might have just remove the Replacement as faulty
1917 * Clear the flag just in case
1919 clear_bit(WantReplacement
, &rdev
->flags
);
1921 err
= md_integrity_register(mddev
);
1929 static void end_sync_read(struct bio
*bio
, int error
)
1931 struct r10bio
*r10_bio
= bio
->bi_private
;
1932 struct r10conf
*conf
= r10_bio
->mddev
->private;
1935 if (bio
== r10_bio
->master_bio
) {
1936 /* this is a reshape read */
1937 d
= r10_bio
->read_slot
; /* really the read dev */
1939 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1941 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1942 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1944 /* The write handler will notice the lack of
1945 * R10BIO_Uptodate and record any errors etc
1947 atomic_add(r10_bio
->sectors
,
1948 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1950 /* for reconstruct, we always reschedule after a read.
1951 * for resync, only after all reads
1953 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1954 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1955 atomic_dec_and_test(&r10_bio
->remaining
)) {
1956 /* we have read all the blocks,
1957 * do the comparison in process context in raid10d
1959 reschedule_retry(r10_bio
);
1963 static void end_sync_request(struct r10bio
*r10_bio
)
1965 struct mddev
*mddev
= r10_bio
->mddev
;
1967 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1968 if (r10_bio
->master_bio
== NULL
) {
1969 /* the primary of several recovery bios */
1970 sector_t s
= r10_bio
->sectors
;
1971 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1972 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1973 reschedule_retry(r10_bio
);
1976 md_done_sync(mddev
, s
, 1);
1979 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1980 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1981 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1982 reschedule_retry(r10_bio
);
1990 static void end_sync_write(struct bio
*bio
, int error
)
1992 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1993 struct r10bio
*r10_bio
= bio
->bi_private
;
1994 struct mddev
*mddev
= r10_bio
->mddev
;
1995 struct r10conf
*conf
= mddev
->private;
2001 struct md_rdev
*rdev
= NULL
;
2003 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
2005 rdev
= conf
->mirrors
[d
].replacement
;
2007 rdev
= conf
->mirrors
[d
].rdev
;
2011 md_error(mddev
, rdev
);
2013 set_bit(WriteErrorSeen
, &rdev
->flags
);
2014 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2015 set_bit(MD_RECOVERY_NEEDED
,
2016 &rdev
->mddev
->recovery
);
2017 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
2019 } else if (is_badblock(rdev
,
2020 r10_bio
->devs
[slot
].addr
,
2022 &first_bad
, &bad_sectors
))
2023 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
2025 rdev_dec_pending(rdev
, mddev
);
2027 end_sync_request(r10_bio
);
2031 * Note: sync and recover and handled very differently for raid10
2032 * This code is for resync.
2033 * For resync, we read through virtual addresses and read all blocks.
2034 * If there is any error, we schedule a write. The lowest numbered
2035 * drive is authoritative.
2036 * However requests come for physical address, so we need to map.
2037 * For every physical address there are raid_disks/copies virtual addresses,
2038 * which is always are least one, but is not necessarly an integer.
2039 * This means that a physical address can span multiple chunks, so we may
2040 * have to submit multiple io requests for a single sync request.
2043 * We check if all blocks are in-sync and only write to blocks that
2046 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2048 struct r10conf
*conf
= mddev
->private;
2050 struct bio
*tbio
, *fbio
;
2053 atomic_set(&r10_bio
->remaining
, 1);
2055 /* find the first device with a block */
2056 for (i
=0; i
<conf
->copies
; i
++)
2057 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
2060 if (i
== conf
->copies
)
2064 fbio
= r10_bio
->devs
[i
].bio
;
2066 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2067 /* now find blocks with errors */
2068 for (i
=0 ; i
< conf
->copies
; i
++) {
2071 tbio
= r10_bio
->devs
[i
].bio
;
2073 if (tbio
->bi_end_io
!= end_sync_read
)
2077 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2078 /* We know that the bi_io_vec layout is the same for
2079 * both 'first' and 'i', so we just compare them.
2080 * All vec entries are PAGE_SIZE;
2082 int sectors
= r10_bio
->sectors
;
2083 for (j
= 0; j
< vcnt
; j
++) {
2084 int len
= PAGE_SIZE
;
2085 if (sectors
< (len
/ 512))
2086 len
= sectors
* 512;
2087 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2088 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2095 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2096 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2097 /* Don't fix anything. */
2100 /* Ok, we need to write this bio, either to correct an
2101 * inconsistency or to correct an unreadable block.
2102 * First we need to fixup bv_offset, bv_len and
2103 * bi_vecs, as the read request might have corrupted these
2107 tbio
->bi_vcnt
= vcnt
;
2108 tbio
->bi_iter
.bi_size
= r10_bio
->sectors
<< 9;
2109 tbio
->bi_rw
= WRITE
;
2110 tbio
->bi_private
= r10_bio
;
2111 tbio
->bi_iter
.bi_sector
= r10_bio
->devs
[i
].addr
;
2113 for (j
=0; j
< vcnt
; j
++) {
2114 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2115 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2117 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2118 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2121 tbio
->bi_end_io
= end_sync_write
;
2123 d
= r10_bio
->devs
[i
].devnum
;
2124 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2125 atomic_inc(&r10_bio
->remaining
);
2126 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2128 tbio
->bi_iter
.bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2129 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2130 generic_make_request(tbio
);
2133 /* Now write out to any replacement devices
2136 for (i
= 0; i
< conf
->copies
; i
++) {
2139 tbio
= r10_bio
->devs
[i
].repl_bio
;
2140 if (!tbio
|| !tbio
->bi_end_io
)
2142 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2143 && r10_bio
->devs
[i
].bio
!= fbio
)
2144 for (j
= 0; j
< vcnt
; j
++)
2145 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2146 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2148 d
= r10_bio
->devs
[i
].devnum
;
2149 atomic_inc(&r10_bio
->remaining
);
2150 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2152 generic_make_request(tbio
);
2156 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2157 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2163 * Now for the recovery code.
2164 * Recovery happens across physical sectors.
2165 * We recover all non-is_sync drives by finding the virtual address of
2166 * each, and then choose a working drive that also has that virt address.
2167 * There is a separate r10_bio for each non-in_sync drive.
2168 * Only the first two slots are in use. The first for reading,
2169 * The second for writing.
2172 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2174 /* We got a read error during recovery.
2175 * We repeat the read in smaller page-sized sections.
2176 * If a read succeeds, write it to the new device or record
2177 * a bad block if we cannot.
2178 * If a read fails, record a bad block on both old and
2181 struct mddev
*mddev
= r10_bio
->mddev
;
2182 struct r10conf
*conf
= mddev
->private;
2183 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2185 int sectors
= r10_bio
->sectors
;
2187 int dr
= r10_bio
->devs
[0].devnum
;
2188 int dw
= r10_bio
->devs
[1].devnum
;
2192 struct md_rdev
*rdev
;
2196 if (s
> (PAGE_SIZE
>>9))
2199 rdev
= conf
->mirrors
[dr
].rdev
;
2200 addr
= r10_bio
->devs
[0].addr
+ sect
,
2201 ok
= sync_page_io(rdev
,
2204 bio
->bi_io_vec
[idx
].bv_page
,
2207 rdev
= conf
->mirrors
[dw
].rdev
;
2208 addr
= r10_bio
->devs
[1].addr
+ sect
;
2209 ok
= sync_page_io(rdev
,
2212 bio
->bi_io_vec
[idx
].bv_page
,
2215 set_bit(WriteErrorSeen
, &rdev
->flags
);
2216 if (!test_and_set_bit(WantReplacement
,
2218 set_bit(MD_RECOVERY_NEEDED
,
2219 &rdev
->mddev
->recovery
);
2223 /* We don't worry if we cannot set a bad block -
2224 * it really is bad so there is no loss in not
2227 rdev_set_badblocks(rdev
, addr
, s
, 0);
2229 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2230 /* need bad block on destination too */
2231 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2232 addr
= r10_bio
->devs
[1].addr
+ sect
;
2233 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2235 /* just abort the recovery */
2237 "md/raid10:%s: recovery aborted"
2238 " due to read error\n",
2241 conf
->mirrors
[dw
].recovery_disabled
2242 = mddev
->recovery_disabled
;
2243 set_bit(MD_RECOVERY_INTR
,
2256 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2258 struct r10conf
*conf
= mddev
->private;
2260 struct bio
*wbio
, *wbio2
;
2262 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2263 fix_recovery_read_error(r10_bio
);
2264 end_sync_request(r10_bio
);
2269 * share the pages with the first bio
2270 * and submit the write request
2272 d
= r10_bio
->devs
[1].devnum
;
2273 wbio
= r10_bio
->devs
[1].bio
;
2274 wbio2
= r10_bio
->devs
[1].repl_bio
;
2275 /* Need to test wbio2->bi_end_io before we call
2276 * generic_make_request as if the former is NULL,
2277 * the latter is free to free wbio2.
2279 if (wbio2
&& !wbio2
->bi_end_io
)
2281 if (wbio
->bi_end_io
) {
2282 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2283 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2284 generic_make_request(wbio
);
2287 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2288 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2289 bio_sectors(wbio2
));
2290 generic_make_request(wbio2
);
2295 * Used by fix_read_error() to decay the per rdev read_errors.
2296 * We halve the read error count for every hour that has elapsed
2297 * since the last recorded read error.
2300 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2302 struct timespec cur_time_mon
;
2303 unsigned long hours_since_last
;
2304 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2306 ktime_get_ts(&cur_time_mon
);
2308 if (rdev
->last_read_error
.tv_sec
== 0 &&
2309 rdev
->last_read_error
.tv_nsec
== 0) {
2310 /* first time we've seen a read error */
2311 rdev
->last_read_error
= cur_time_mon
;
2315 hours_since_last
= (cur_time_mon
.tv_sec
-
2316 rdev
->last_read_error
.tv_sec
) / 3600;
2318 rdev
->last_read_error
= cur_time_mon
;
2321 * if hours_since_last is > the number of bits in read_errors
2322 * just set read errors to 0. We do this to avoid
2323 * overflowing the shift of read_errors by hours_since_last.
2325 if (hours_since_last
>= 8 * sizeof(read_errors
))
2326 atomic_set(&rdev
->read_errors
, 0);
2328 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2331 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2332 int sectors
, struct page
*page
, int rw
)
2337 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2338 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2340 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2344 set_bit(WriteErrorSeen
, &rdev
->flags
);
2345 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2346 set_bit(MD_RECOVERY_NEEDED
,
2347 &rdev
->mddev
->recovery
);
2349 /* need to record an error - either for the block or the device */
2350 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2351 md_error(rdev
->mddev
, rdev
);
2356 * This is a kernel thread which:
2358 * 1. Retries failed read operations on working mirrors.
2359 * 2. Updates the raid superblock when problems encounter.
2360 * 3. Performs writes following reads for array synchronising.
2363 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2365 int sect
= 0; /* Offset from r10_bio->sector */
2366 int sectors
= r10_bio
->sectors
;
2367 struct md_rdev
*rdev
;
2368 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2369 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2371 /* still own a reference to this rdev, so it cannot
2372 * have been cleared recently.
2374 rdev
= conf
->mirrors
[d
].rdev
;
2376 if (test_bit(Faulty
, &rdev
->flags
))
2377 /* drive has already been failed, just ignore any
2378 more fix_read_error() attempts */
2381 check_decay_read_errors(mddev
, rdev
);
2382 atomic_inc(&rdev
->read_errors
);
2383 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2384 char b
[BDEVNAME_SIZE
];
2385 bdevname(rdev
->bdev
, b
);
2388 "md/raid10:%s: %s: Raid device exceeded "
2389 "read_error threshold [cur %d:max %d]\n",
2391 atomic_read(&rdev
->read_errors
), max_read_errors
);
2393 "md/raid10:%s: %s: Failing raid device\n",
2395 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2396 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2402 int sl
= r10_bio
->read_slot
;
2406 if (s
> (PAGE_SIZE
>>9))
2414 d
= r10_bio
->devs
[sl
].devnum
;
2415 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2417 !test_bit(Unmerged
, &rdev
->flags
) &&
2418 test_bit(In_sync
, &rdev
->flags
) &&
2419 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2420 &first_bad
, &bad_sectors
) == 0) {
2421 atomic_inc(&rdev
->nr_pending
);
2423 success
= sync_page_io(rdev
,
2424 r10_bio
->devs
[sl
].addr
+
2427 conf
->tmppage
, READ
, false);
2428 rdev_dec_pending(rdev
, mddev
);
2434 if (sl
== conf
->copies
)
2436 } while (!success
&& sl
!= r10_bio
->read_slot
);
2440 /* Cannot read from anywhere, just mark the block
2441 * as bad on the first device to discourage future
2444 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2445 rdev
= conf
->mirrors
[dn
].rdev
;
2447 if (!rdev_set_badblocks(
2449 r10_bio
->devs
[r10_bio
->read_slot
].addr
2452 md_error(mddev
, rdev
);
2453 r10_bio
->devs
[r10_bio
->read_slot
].bio
2460 /* write it back and re-read */
2462 while (sl
!= r10_bio
->read_slot
) {
2463 char b
[BDEVNAME_SIZE
];
2468 d
= r10_bio
->devs
[sl
].devnum
;
2469 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2471 test_bit(Unmerged
, &rdev
->flags
) ||
2472 !test_bit(In_sync
, &rdev
->flags
))
2475 atomic_inc(&rdev
->nr_pending
);
2477 if (r10_sync_page_io(rdev
,
2478 r10_bio
->devs
[sl
].addr
+
2480 s
, conf
->tmppage
, WRITE
)
2482 /* Well, this device is dead */
2484 "md/raid10:%s: read correction "
2486 " (%d sectors at %llu on %s)\n",
2488 (unsigned long long)(
2490 choose_data_offset(r10_bio
,
2492 bdevname(rdev
->bdev
, b
));
2493 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2496 bdevname(rdev
->bdev
, b
));
2498 rdev_dec_pending(rdev
, mddev
);
2502 while (sl
!= r10_bio
->read_slot
) {
2503 char b
[BDEVNAME_SIZE
];
2508 d
= r10_bio
->devs
[sl
].devnum
;
2509 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2511 !test_bit(In_sync
, &rdev
->flags
))
2514 atomic_inc(&rdev
->nr_pending
);
2516 switch (r10_sync_page_io(rdev
,
2517 r10_bio
->devs
[sl
].addr
+
2522 /* Well, this device is dead */
2524 "md/raid10:%s: unable to read back "
2526 " (%d sectors at %llu on %s)\n",
2528 (unsigned long long)(
2530 choose_data_offset(r10_bio
, rdev
)),
2531 bdevname(rdev
->bdev
, b
));
2532 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2535 bdevname(rdev
->bdev
, b
));
2539 "md/raid10:%s: read error corrected"
2540 " (%d sectors at %llu on %s)\n",
2542 (unsigned long long)(
2544 choose_data_offset(r10_bio
, rdev
)),
2545 bdevname(rdev
->bdev
, b
));
2546 atomic_add(s
, &rdev
->corrected_errors
);
2549 rdev_dec_pending(rdev
, mddev
);
2559 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2561 struct bio
*bio
= r10_bio
->master_bio
;
2562 struct mddev
*mddev
= r10_bio
->mddev
;
2563 struct r10conf
*conf
= mddev
->private;
2564 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2565 /* bio has the data to be written to slot 'i' where
2566 * we just recently had a write error.
2567 * We repeatedly clone the bio and trim down to one block,
2568 * then try the write. Where the write fails we record
2570 * It is conceivable that the bio doesn't exactly align with
2571 * blocks. We must handle this.
2573 * We currently own a reference to the rdev.
2579 int sect_to_write
= r10_bio
->sectors
;
2582 if (rdev
->badblocks
.shift
< 0)
2585 block_sectors
= 1 << rdev
->badblocks
.shift
;
2586 sector
= r10_bio
->sector
;
2587 sectors
= ((r10_bio
->sector
+ block_sectors
)
2588 & ~(sector_t
)(block_sectors
- 1))
2591 while (sect_to_write
) {
2593 if (sectors
> sect_to_write
)
2594 sectors
= sect_to_write
;
2595 /* Write at 'sector' for 'sectors' */
2596 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2597 bio_trim(wbio
, sector
- bio
->bi_iter
.bi_sector
, sectors
);
2598 wbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
2599 choose_data_offset(r10_bio
, rdev
) +
2600 (sector
- r10_bio
->sector
));
2601 wbio
->bi_bdev
= rdev
->bdev
;
2602 if (submit_bio_wait(WRITE
, wbio
) == 0)
2604 ok
= rdev_set_badblocks(rdev
, sector
,
2609 sect_to_write
-= sectors
;
2611 sectors
= block_sectors
;
2616 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2618 int slot
= r10_bio
->read_slot
;
2620 struct r10conf
*conf
= mddev
->private;
2621 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2622 char b
[BDEVNAME_SIZE
];
2623 unsigned long do_sync
;
2626 /* we got a read error. Maybe the drive is bad. Maybe just
2627 * the block and we can fix it.
2628 * We freeze all other IO, and try reading the block from
2629 * other devices. When we find one, we re-write
2630 * and check it that fixes the read error.
2631 * This is all done synchronously while the array is
2634 bio
= r10_bio
->devs
[slot
].bio
;
2635 bdevname(bio
->bi_bdev
, b
);
2637 r10_bio
->devs
[slot
].bio
= NULL
;
2639 if (mddev
->ro
== 0) {
2640 freeze_array(conf
, 1);
2641 fix_read_error(conf
, mddev
, r10_bio
);
2642 unfreeze_array(conf
);
2644 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2646 rdev_dec_pending(rdev
, mddev
);
2649 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2651 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2652 " read error for block %llu\n",
2654 (unsigned long long)r10_bio
->sector
);
2655 raid_end_bio_io(r10_bio
);
2659 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2660 slot
= r10_bio
->read_slot
;
2663 "md/raid10:%s: %s: redirecting "
2664 "sector %llu to another mirror\n",
2666 bdevname(rdev
->bdev
, b
),
2667 (unsigned long long)r10_bio
->sector
);
2668 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2670 bio_trim(bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
2671 r10_bio
->devs
[slot
].bio
= bio
;
2672 r10_bio
->devs
[slot
].rdev
= rdev
;
2673 bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
2674 + choose_data_offset(r10_bio
, rdev
);
2675 bio
->bi_bdev
= rdev
->bdev
;
2676 bio
->bi_rw
= READ
| do_sync
;
2677 bio
->bi_private
= r10_bio
;
2678 bio
->bi_end_io
= raid10_end_read_request
;
2679 if (max_sectors
< r10_bio
->sectors
) {
2680 /* Drat - have to split this up more */
2681 struct bio
*mbio
= r10_bio
->master_bio
;
2682 int sectors_handled
=
2683 r10_bio
->sector
+ max_sectors
2684 - mbio
->bi_iter
.bi_sector
;
2685 r10_bio
->sectors
= max_sectors
;
2686 spin_lock_irq(&conf
->device_lock
);
2687 if (mbio
->bi_phys_segments
== 0)
2688 mbio
->bi_phys_segments
= 2;
2690 mbio
->bi_phys_segments
++;
2691 spin_unlock_irq(&conf
->device_lock
);
2692 generic_make_request(bio
);
2694 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2696 r10_bio
->master_bio
= mbio
;
2697 r10_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2699 set_bit(R10BIO_ReadError
,
2701 r10_bio
->mddev
= mddev
;
2702 r10_bio
->sector
= mbio
->bi_iter
.bi_sector
2707 generic_make_request(bio
);
2710 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2712 /* Some sort of write request has finished and it
2713 * succeeded in writing where we thought there was a
2714 * bad block. So forget the bad block.
2715 * Or possibly if failed and we need to record
2719 struct md_rdev
*rdev
;
2721 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2722 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2723 for (m
= 0; m
< conf
->copies
; m
++) {
2724 int dev
= r10_bio
->devs
[m
].devnum
;
2725 rdev
= conf
->mirrors
[dev
].rdev
;
2726 if (r10_bio
->devs
[m
].bio
== NULL
)
2728 if (test_bit(BIO_UPTODATE
,
2729 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2730 rdev_clear_badblocks(
2732 r10_bio
->devs
[m
].addr
,
2733 r10_bio
->sectors
, 0);
2735 if (!rdev_set_badblocks(
2737 r10_bio
->devs
[m
].addr
,
2738 r10_bio
->sectors
, 0))
2739 md_error(conf
->mddev
, rdev
);
2741 rdev
= conf
->mirrors
[dev
].replacement
;
2742 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2744 if (test_bit(BIO_UPTODATE
,
2745 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2746 rdev_clear_badblocks(
2748 r10_bio
->devs
[m
].addr
,
2749 r10_bio
->sectors
, 0);
2751 if (!rdev_set_badblocks(
2753 r10_bio
->devs
[m
].addr
,
2754 r10_bio
->sectors
, 0))
2755 md_error(conf
->mddev
, rdev
);
2760 for (m
= 0; m
< conf
->copies
; m
++) {
2761 int dev
= r10_bio
->devs
[m
].devnum
;
2762 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2763 rdev
= conf
->mirrors
[dev
].rdev
;
2764 if (bio
== IO_MADE_GOOD
) {
2765 rdev_clear_badblocks(
2767 r10_bio
->devs
[m
].addr
,
2768 r10_bio
->sectors
, 0);
2769 rdev_dec_pending(rdev
, conf
->mddev
);
2770 } else if (bio
!= NULL
&&
2771 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2772 if (!narrow_write_error(r10_bio
, m
)) {
2773 md_error(conf
->mddev
, rdev
);
2774 set_bit(R10BIO_Degraded
,
2777 rdev_dec_pending(rdev
, conf
->mddev
);
2779 bio
= r10_bio
->devs
[m
].repl_bio
;
2780 rdev
= conf
->mirrors
[dev
].replacement
;
2781 if (rdev
&& bio
== IO_MADE_GOOD
) {
2782 rdev_clear_badblocks(
2784 r10_bio
->devs
[m
].addr
,
2785 r10_bio
->sectors
, 0);
2786 rdev_dec_pending(rdev
, conf
->mddev
);
2789 if (test_bit(R10BIO_WriteError
,
2791 close_write(r10_bio
);
2792 raid_end_bio_io(r10_bio
);
2796 static void raid10d(struct md_thread
*thread
)
2798 struct mddev
*mddev
= thread
->mddev
;
2799 struct r10bio
*r10_bio
;
2800 unsigned long flags
;
2801 struct r10conf
*conf
= mddev
->private;
2802 struct list_head
*head
= &conf
->retry_list
;
2803 struct blk_plug plug
;
2805 md_check_recovery(mddev
);
2807 blk_start_plug(&plug
);
2810 flush_pending_writes(conf
);
2812 spin_lock_irqsave(&conf
->device_lock
, flags
);
2813 if (list_empty(head
)) {
2814 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2817 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2818 list_del(head
->prev
);
2820 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2822 mddev
= r10_bio
->mddev
;
2823 conf
= mddev
->private;
2824 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2825 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2826 handle_write_completed(conf
, r10_bio
);
2827 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2828 reshape_request_write(mddev
, r10_bio
);
2829 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2830 sync_request_write(mddev
, r10_bio
);
2831 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2832 recovery_request_write(mddev
, r10_bio
);
2833 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2834 handle_read_error(mddev
, r10_bio
);
2836 /* just a partial read to be scheduled from a
2839 int slot
= r10_bio
->read_slot
;
2840 generic_make_request(r10_bio
->devs
[slot
].bio
);
2844 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2845 md_check_recovery(mddev
);
2847 blk_finish_plug(&plug
);
2850 static int init_resync(struct r10conf
*conf
)
2855 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2856 BUG_ON(conf
->r10buf_pool
);
2857 conf
->have_replacement
= 0;
2858 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2859 if (conf
->mirrors
[i
].replacement
)
2860 conf
->have_replacement
= 1;
2861 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2862 if (!conf
->r10buf_pool
)
2864 conf
->next_resync
= 0;
2869 * perform a "sync" on one "block"
2871 * We need to make sure that no normal I/O request - particularly write
2872 * requests - conflict with active sync requests.
2874 * This is achieved by tracking pending requests and a 'barrier' concept
2875 * that can be installed to exclude normal IO requests.
2877 * Resync and recovery are handled very differently.
2878 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2880 * For resync, we iterate over virtual addresses, read all copies,
2881 * and update if there are differences. If only one copy is live,
2883 * For recovery, we iterate over physical addresses, read a good
2884 * value for each non-in_sync drive, and over-write.
2886 * So, for recovery we may have several outstanding complex requests for a
2887 * given address, one for each out-of-sync device. We model this by allocating
2888 * a number of r10_bio structures, one for each out-of-sync device.
2889 * As we setup these structures, we collect all bio's together into a list
2890 * which we then process collectively to add pages, and then process again
2891 * to pass to generic_make_request.
2893 * The r10_bio structures are linked using a borrowed master_bio pointer.
2894 * This link is counted in ->remaining. When the r10_bio that points to NULL
2895 * has its remaining count decremented to 0, the whole complex operation
2900 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2901 int *skipped
, int go_faster
)
2903 struct r10conf
*conf
= mddev
->private;
2904 struct r10bio
*r10_bio
;
2905 struct bio
*biolist
= NULL
, *bio
;
2906 sector_t max_sector
, nr_sectors
;
2909 sector_t sync_blocks
;
2910 sector_t sectors_skipped
= 0;
2911 int chunks_skipped
= 0;
2912 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2914 if (!conf
->r10buf_pool
)
2915 if (init_resync(conf
))
2919 * Allow skipping a full rebuild for incremental assembly
2920 * of a clean array, like RAID1 does.
2922 if (mddev
->bitmap
== NULL
&&
2923 mddev
->recovery_cp
== MaxSector
&&
2924 mddev
->reshape_position
== MaxSector
&&
2925 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2926 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2927 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2928 conf
->fullsync
== 0) {
2930 return mddev
->dev_sectors
- sector_nr
;
2934 max_sector
= mddev
->dev_sectors
;
2935 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2936 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2937 max_sector
= mddev
->resync_max_sectors
;
2938 if (sector_nr
>= max_sector
) {
2939 /* If we aborted, we need to abort the
2940 * sync on the 'current' bitmap chucks (there can
2941 * be several when recovering multiple devices).
2942 * as we may have started syncing it but not finished.
2943 * We can find the current address in
2944 * mddev->curr_resync, but for recovery,
2945 * we need to convert that to several
2946 * virtual addresses.
2948 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2954 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2955 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2956 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2958 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2960 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2961 bitmap_end_sync(mddev
->bitmap
, sect
,
2965 /* completed sync */
2966 if ((!mddev
->bitmap
|| conf
->fullsync
)
2967 && conf
->have_replacement
2968 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2969 /* Completed a full sync so the replacements
2970 * are now fully recovered.
2972 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2973 if (conf
->mirrors
[i
].replacement
)
2974 conf
->mirrors
[i
].replacement
2980 bitmap_close_sync(mddev
->bitmap
);
2983 return sectors_skipped
;
2986 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2987 return reshape_request(mddev
, sector_nr
, skipped
);
2989 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2990 /* if there has been nothing to do on any drive,
2991 * then there is nothing to do at all..
2994 return (max_sector
- sector_nr
) + sectors_skipped
;
2997 if (max_sector
> mddev
->resync_max
)
2998 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
3000 /* make sure whole request will fit in a chunk - if chunks
3003 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
3004 max_sector
> (sector_nr
| chunk_mask
))
3005 max_sector
= (sector_nr
| chunk_mask
) + 1;
3007 * If there is non-resync activity waiting for us then
3008 * put in a delay to throttle resync.
3010 if (!go_faster
&& conf
->nr_waiting
)
3011 msleep_interruptible(1000);
3013 /* Again, very different code for resync and recovery.
3014 * Both must result in an r10bio with a list of bios that
3015 * have bi_end_io, bi_sector, bi_bdev set,
3016 * and bi_private set to the r10bio.
3017 * For recovery, we may actually create several r10bios
3018 * with 2 bios in each, that correspond to the bios in the main one.
3019 * In this case, the subordinate r10bios link back through a
3020 * borrowed master_bio pointer, and the counter in the master
3021 * includes a ref from each subordinate.
3023 /* First, we decide what to do and set ->bi_end_io
3024 * To end_sync_read if we want to read, and
3025 * end_sync_write if we will want to write.
3028 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
3029 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3030 /* recovery... the complicated one */
3034 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
3040 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
3042 if ((mirror
->rdev
== NULL
||
3043 test_bit(In_sync
, &mirror
->rdev
->flags
))
3045 (mirror
->replacement
== NULL
||
3047 &mirror
->replacement
->flags
)))
3051 /* want to reconstruct this device */
3053 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3054 if (sect
>= mddev
->resync_max_sectors
) {
3055 /* last stripe is not complete - don't
3056 * try to recover this sector.
3060 /* Unless we are doing a full sync, or a replacement
3061 * we only need to recover the block if it is set in
3064 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3066 if (sync_blocks
< max_sync
)
3067 max_sync
= sync_blocks
;
3069 mirror
->replacement
== NULL
&&
3071 /* yep, skip the sync_blocks here, but don't assume
3072 * that there will never be anything to do here
3074 chunks_skipped
= -1;
3078 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3080 raise_barrier(conf
, rb2
!= NULL
);
3081 atomic_set(&r10_bio
->remaining
, 0);
3083 r10_bio
->master_bio
= (struct bio
*)rb2
;
3085 atomic_inc(&rb2
->remaining
);
3086 r10_bio
->mddev
= mddev
;
3087 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3088 r10_bio
->sector
= sect
;
3090 raid10_find_phys(conf
, r10_bio
);
3092 /* Need to check if the array will still be
3095 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3096 if (conf
->mirrors
[j
].rdev
== NULL
||
3097 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3102 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3103 &sync_blocks
, still_degraded
);
3106 for (j
=0; j
<conf
->copies
;j
++) {
3108 int d
= r10_bio
->devs
[j
].devnum
;
3109 sector_t from_addr
, to_addr
;
3110 struct md_rdev
*rdev
;
3111 sector_t sector
, first_bad
;
3113 if (!conf
->mirrors
[d
].rdev
||
3114 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3116 /* This is where we read from */
3118 rdev
= conf
->mirrors
[d
].rdev
;
3119 sector
= r10_bio
->devs
[j
].addr
;
3121 if (is_badblock(rdev
, sector
, max_sync
,
3122 &first_bad
, &bad_sectors
)) {
3123 if (first_bad
> sector
)
3124 max_sync
= first_bad
- sector
;
3126 bad_sectors
-= (sector
3128 if (max_sync
> bad_sectors
)
3129 max_sync
= bad_sectors
;
3133 bio
= r10_bio
->devs
[0].bio
;
3135 bio
->bi_next
= biolist
;
3137 bio
->bi_private
= r10_bio
;
3138 bio
->bi_end_io
= end_sync_read
;
3140 from_addr
= r10_bio
->devs
[j
].addr
;
3141 bio
->bi_iter
.bi_sector
= from_addr
+
3143 bio
->bi_bdev
= rdev
->bdev
;
3144 atomic_inc(&rdev
->nr_pending
);
3145 /* and we write to 'i' (if not in_sync) */
3147 for (k
=0; k
<conf
->copies
; k
++)
3148 if (r10_bio
->devs
[k
].devnum
== i
)
3150 BUG_ON(k
== conf
->copies
);
3151 to_addr
= r10_bio
->devs
[k
].addr
;
3152 r10_bio
->devs
[0].devnum
= d
;
3153 r10_bio
->devs
[0].addr
= from_addr
;
3154 r10_bio
->devs
[1].devnum
= i
;
3155 r10_bio
->devs
[1].addr
= to_addr
;
3157 rdev
= mirror
->rdev
;
3158 if (!test_bit(In_sync
, &rdev
->flags
)) {
3159 bio
= r10_bio
->devs
[1].bio
;
3161 bio
->bi_next
= biolist
;
3163 bio
->bi_private
= r10_bio
;
3164 bio
->bi_end_io
= end_sync_write
;
3166 bio
->bi_iter
.bi_sector
= to_addr
3167 + rdev
->data_offset
;
3168 bio
->bi_bdev
= rdev
->bdev
;
3169 atomic_inc(&r10_bio
->remaining
);
3171 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3173 /* and maybe write to replacement */
3174 bio
= r10_bio
->devs
[1].repl_bio
;
3176 bio
->bi_end_io
= NULL
;
3177 rdev
= mirror
->replacement
;
3178 /* Note: if rdev != NULL, then bio
3179 * cannot be NULL as r10buf_pool_alloc will
3180 * have allocated it.
3181 * So the second test here is pointless.
3182 * But it keeps semantic-checkers happy, and
3183 * this comment keeps human reviewers
3186 if (rdev
== NULL
|| bio
== NULL
||
3187 test_bit(Faulty
, &rdev
->flags
))
3190 bio
->bi_next
= biolist
;
3192 bio
->bi_private
= r10_bio
;
3193 bio
->bi_end_io
= end_sync_write
;
3195 bio
->bi_iter
.bi_sector
= to_addr
+
3197 bio
->bi_bdev
= rdev
->bdev
;
3198 atomic_inc(&r10_bio
->remaining
);
3201 if (j
== conf
->copies
) {
3202 /* Cannot recover, so abort the recovery or
3203 * record a bad block */
3205 /* problem is that there are bad blocks
3206 * on other device(s)
3209 for (k
= 0; k
< conf
->copies
; k
++)
3210 if (r10_bio
->devs
[k
].devnum
== i
)
3212 if (!test_bit(In_sync
,
3213 &mirror
->rdev
->flags
)
3214 && !rdev_set_badblocks(
3216 r10_bio
->devs
[k
].addr
,
3219 if (mirror
->replacement
&&
3220 !rdev_set_badblocks(
3221 mirror
->replacement
,
3222 r10_bio
->devs
[k
].addr
,
3227 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3229 printk(KERN_INFO
"md/raid10:%s: insufficient "
3230 "working devices for recovery.\n",
3232 mirror
->recovery_disabled
3233 = mddev
->recovery_disabled
;
3237 atomic_dec(&rb2
->remaining
);
3242 if (biolist
== NULL
) {
3244 struct r10bio
*rb2
= r10_bio
;
3245 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3246 rb2
->master_bio
= NULL
;
3252 /* resync. Schedule a read for every block at this virt offset */
3255 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3257 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3258 &sync_blocks
, mddev
->degraded
) &&
3259 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3260 &mddev
->recovery
)) {
3261 /* We can skip this block */
3263 return sync_blocks
+ sectors_skipped
;
3265 if (sync_blocks
< max_sync
)
3266 max_sync
= sync_blocks
;
3267 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3270 r10_bio
->mddev
= mddev
;
3271 atomic_set(&r10_bio
->remaining
, 0);
3272 raise_barrier(conf
, 0);
3273 conf
->next_resync
= sector_nr
;
3275 r10_bio
->master_bio
= NULL
;
3276 r10_bio
->sector
= sector_nr
;
3277 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3278 raid10_find_phys(conf
, r10_bio
);
3279 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3281 for (i
= 0; i
< conf
->copies
; i
++) {
3282 int d
= r10_bio
->devs
[i
].devnum
;
3283 sector_t first_bad
, sector
;
3286 if (r10_bio
->devs
[i
].repl_bio
)
3287 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3289 bio
= r10_bio
->devs
[i
].bio
;
3291 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3292 if (conf
->mirrors
[d
].rdev
== NULL
||
3293 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3295 sector
= r10_bio
->devs
[i
].addr
;
3296 if (is_badblock(conf
->mirrors
[d
].rdev
,
3298 &first_bad
, &bad_sectors
)) {
3299 if (first_bad
> sector
)
3300 max_sync
= first_bad
- sector
;
3302 bad_sectors
-= (sector
- first_bad
);
3303 if (max_sync
> bad_sectors
)
3304 max_sync
= bad_sectors
;
3308 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3309 atomic_inc(&r10_bio
->remaining
);
3310 bio
->bi_next
= biolist
;
3312 bio
->bi_private
= r10_bio
;
3313 bio
->bi_end_io
= end_sync_read
;
3315 bio
->bi_iter
.bi_sector
= sector
+
3316 conf
->mirrors
[d
].rdev
->data_offset
;
3317 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3320 if (conf
->mirrors
[d
].replacement
== NULL
||
3322 &conf
->mirrors
[d
].replacement
->flags
))
3325 /* Need to set up for writing to the replacement */
3326 bio
= r10_bio
->devs
[i
].repl_bio
;
3328 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3330 sector
= r10_bio
->devs
[i
].addr
;
3331 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3332 bio
->bi_next
= biolist
;
3334 bio
->bi_private
= r10_bio
;
3335 bio
->bi_end_io
= end_sync_write
;
3337 bio
->bi_iter
.bi_sector
= sector
+
3338 conf
->mirrors
[d
].replacement
->data_offset
;
3339 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3344 for (i
=0; i
<conf
->copies
; i
++) {
3345 int d
= r10_bio
->devs
[i
].devnum
;
3346 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3347 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3349 if (r10_bio
->devs
[i
].repl_bio
&&
3350 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3352 conf
->mirrors
[d
].replacement
,
3362 if (sector_nr
+ max_sync
< max_sector
)
3363 max_sector
= sector_nr
+ max_sync
;
3366 int len
= PAGE_SIZE
;
3367 if (sector_nr
+ (len
>>9) > max_sector
)
3368 len
= (max_sector
- sector_nr
) << 9;
3371 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3373 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3374 if (bio_add_page(bio
, page
, len
, 0))
3378 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3379 for (bio2
= biolist
;
3380 bio2
&& bio2
!= bio
;
3381 bio2
= bio2
->bi_next
) {
3382 /* remove last page from this bio */
3384 bio2
->bi_iter
.bi_size
-= len
;
3385 __clear_bit(BIO_SEG_VALID
, &bio2
->bi_flags
);
3389 nr_sectors
+= len
>>9;
3390 sector_nr
+= len
>>9;
3391 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3393 r10_bio
->sectors
= nr_sectors
;
3397 biolist
= biolist
->bi_next
;
3399 bio
->bi_next
= NULL
;
3400 r10_bio
= bio
->bi_private
;
3401 r10_bio
->sectors
= nr_sectors
;
3403 if (bio
->bi_end_io
== end_sync_read
) {
3404 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3405 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3406 generic_make_request(bio
);
3410 if (sectors_skipped
)
3411 /* pretend they weren't skipped, it makes
3412 * no important difference in this case
3414 md_done_sync(mddev
, sectors_skipped
, 1);
3416 return sectors_skipped
+ nr_sectors
;
3418 /* There is nowhere to write, so all non-sync
3419 * drives must be failed or in resync, all drives
3420 * have a bad block, so try the next chunk...
3422 if (sector_nr
+ max_sync
< max_sector
)
3423 max_sector
= sector_nr
+ max_sync
;
3425 sectors_skipped
+= (max_sector
- sector_nr
);
3427 sector_nr
= max_sector
;
3432 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3435 struct r10conf
*conf
= mddev
->private;
3438 raid_disks
= min(conf
->geo
.raid_disks
,
3439 conf
->prev
.raid_disks
);
3441 sectors
= conf
->dev_sectors
;
3443 size
= sectors
>> conf
->geo
.chunk_shift
;
3444 sector_div(size
, conf
->geo
.far_copies
);
3445 size
= size
* raid_disks
;
3446 sector_div(size
, conf
->geo
.near_copies
);
3448 return size
<< conf
->geo
.chunk_shift
;
3451 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3453 /* Calculate the number of sectors-per-device that will
3454 * actually be used, and set conf->dev_sectors and
3458 size
= size
>> conf
->geo
.chunk_shift
;
3459 sector_div(size
, conf
->geo
.far_copies
);
3460 size
= size
* conf
->geo
.raid_disks
;
3461 sector_div(size
, conf
->geo
.near_copies
);
3462 /* 'size' is now the number of chunks in the array */
3463 /* calculate "used chunks per device" */
3464 size
= size
* conf
->copies
;
3466 /* We need to round up when dividing by raid_disks to
3467 * get the stride size.
3469 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3471 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3473 if (conf
->geo
.far_offset
)
3474 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3476 sector_div(size
, conf
->geo
.far_copies
);
3477 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3481 enum geo_type
{geo_new
, geo_old
, geo_start
};
3482 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3485 int layout
, chunk
, disks
;
3488 layout
= mddev
->layout
;
3489 chunk
= mddev
->chunk_sectors
;
3490 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3493 layout
= mddev
->new_layout
;
3494 chunk
= mddev
->new_chunk_sectors
;
3495 disks
= mddev
->raid_disks
;
3497 default: /* avoid 'may be unused' warnings */
3498 case geo_start
: /* new when starting reshape - raid_disks not
3500 layout
= mddev
->new_layout
;
3501 chunk
= mddev
->new_chunk_sectors
;
3502 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3507 if (chunk
< (PAGE_SIZE
>> 9) ||
3508 !is_power_of_2(chunk
))
3511 fc
= (layout
>> 8) & 255;
3512 fo
= layout
& (1<<16);
3513 geo
->raid_disks
= disks
;
3514 geo
->near_copies
= nc
;
3515 geo
->far_copies
= fc
;
3516 geo
->far_offset
= fo
;
3517 geo
->far_set_size
= (layout
& (1<<17)) ? disks
/ fc
: disks
;
3518 geo
->chunk_mask
= chunk
- 1;
3519 geo
->chunk_shift
= ffz(~chunk
);
3523 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3525 struct r10conf
*conf
= NULL
;
3530 copies
= setup_geo(&geo
, mddev
, geo_new
);
3533 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3534 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3535 mdname(mddev
), PAGE_SIZE
);
3539 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3540 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3541 mdname(mddev
), mddev
->new_layout
);
3546 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3550 /* FIXME calc properly */
3551 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3552 max(0,-mddev
->delta_disks
)),
3557 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3562 conf
->copies
= copies
;
3563 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3564 r10bio_pool_free
, conf
);
3565 if (!conf
->r10bio_pool
)
3568 calc_sectors(conf
, mddev
->dev_sectors
);
3569 if (mddev
->reshape_position
== MaxSector
) {
3570 conf
->prev
= conf
->geo
;
3571 conf
->reshape_progress
= MaxSector
;
3573 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3577 conf
->reshape_progress
= mddev
->reshape_position
;
3578 if (conf
->prev
.far_offset
)
3579 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3581 /* far_copies must be 1 */
3582 conf
->prev
.stride
= conf
->dev_sectors
;
3584 spin_lock_init(&conf
->device_lock
);
3585 INIT_LIST_HEAD(&conf
->retry_list
);
3587 spin_lock_init(&conf
->resync_lock
);
3588 init_waitqueue_head(&conf
->wait_barrier
);
3590 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3594 conf
->mddev
= mddev
;
3599 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3602 if (conf
->r10bio_pool
)
3603 mempool_destroy(conf
->r10bio_pool
);
3604 kfree(conf
->mirrors
);
3605 safe_put_page(conf
->tmppage
);
3608 return ERR_PTR(err
);
3611 static int run(struct mddev
*mddev
)
3613 struct r10conf
*conf
;
3614 int i
, disk_idx
, chunk_size
;
3615 struct raid10_info
*disk
;
3616 struct md_rdev
*rdev
;
3618 sector_t min_offset_diff
= 0;
3620 bool discard_supported
= false;
3622 if (mddev
->private == NULL
) {
3623 conf
= setup_conf(mddev
);
3625 return PTR_ERR(conf
);
3626 mddev
->private = conf
;
3628 conf
= mddev
->private;
3632 mddev
->thread
= conf
->thread
;
3633 conf
->thread
= NULL
;
3635 chunk_size
= mddev
->chunk_sectors
<< 9;
3637 blk_queue_max_discard_sectors(mddev
->queue
,
3638 mddev
->chunk_sectors
);
3639 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3640 blk_queue_io_min(mddev
->queue
, chunk_size
);
3641 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3642 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3644 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3645 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3648 rdev_for_each(rdev
, mddev
) {
3650 struct request_queue
*q
;
3652 disk_idx
= rdev
->raid_disk
;
3655 if (disk_idx
>= conf
->geo
.raid_disks
&&
3656 disk_idx
>= conf
->prev
.raid_disks
)
3658 disk
= conf
->mirrors
+ disk_idx
;
3660 if (test_bit(Replacement
, &rdev
->flags
)) {
3661 if (disk
->replacement
)
3663 disk
->replacement
= rdev
;
3669 q
= bdev_get_queue(rdev
->bdev
);
3670 if (q
->merge_bvec_fn
)
3671 mddev
->merge_check_needed
= 1;
3672 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3673 if (!mddev
->reshape_backwards
)
3677 if (first
|| diff
< min_offset_diff
)
3678 min_offset_diff
= diff
;
3681 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3682 rdev
->data_offset
<< 9);
3684 disk
->head_position
= 0;
3686 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3687 discard_supported
= true;
3691 if (discard_supported
)
3692 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3695 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3698 /* need to check that every block has at least one working mirror */
3699 if (!enough(conf
, -1)) {
3700 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3705 if (conf
->reshape_progress
!= MaxSector
) {
3706 /* must ensure that shape change is supported */
3707 if (conf
->geo
.far_copies
!= 1 &&
3708 conf
->geo
.far_offset
== 0)
3710 if (conf
->prev
.far_copies
!= 1 &&
3711 conf
->prev
.far_offset
== 0)
3715 mddev
->degraded
= 0;
3717 i
< conf
->geo
.raid_disks
3718 || i
< conf
->prev
.raid_disks
;
3721 disk
= conf
->mirrors
+ i
;
3723 if (!disk
->rdev
&& disk
->replacement
) {
3724 /* The replacement is all we have - use it */
3725 disk
->rdev
= disk
->replacement
;
3726 disk
->replacement
= NULL
;
3727 clear_bit(Replacement
, &disk
->rdev
->flags
);
3731 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3732 disk
->head_position
= 0;
3735 disk
->rdev
->saved_raid_disk
< 0)
3738 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3741 if (mddev
->recovery_cp
!= MaxSector
)
3742 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3743 " -- starting background reconstruction\n",
3746 "md/raid10:%s: active with %d out of %d devices\n",
3747 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3748 conf
->geo
.raid_disks
);
3750 * Ok, everything is just fine now
3752 mddev
->dev_sectors
= conf
->dev_sectors
;
3753 size
= raid10_size(mddev
, 0, 0);
3754 md_set_array_sectors(mddev
, size
);
3755 mddev
->resync_max_sectors
= size
;
3758 int stripe
= conf
->geo
.raid_disks
*
3759 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3760 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3761 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3763 /* Calculate max read-ahead size.
3764 * We need to readahead at least twice a whole stripe....
3767 stripe
/= conf
->geo
.near_copies
;
3768 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3769 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3770 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3773 if (md_integrity_register(mddev
))
3776 if (conf
->reshape_progress
!= MaxSector
) {
3777 unsigned long before_length
, after_length
;
3779 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3780 conf
->prev
.far_copies
);
3781 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3782 conf
->geo
.far_copies
);
3784 if (max(before_length
, after_length
) > min_offset_diff
) {
3785 /* This cannot work */
3786 printk("md/raid10: offset difference not enough to continue reshape\n");
3789 conf
->offset_diff
= min_offset_diff
;
3791 conf
->reshape_safe
= conf
->reshape_progress
;
3792 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3793 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3794 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3795 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3796 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3803 md_unregister_thread(&mddev
->thread
);
3804 if (conf
->r10bio_pool
)
3805 mempool_destroy(conf
->r10bio_pool
);
3806 safe_put_page(conf
->tmppage
);
3807 kfree(conf
->mirrors
);
3809 mddev
->private = NULL
;
3814 static int stop(struct mddev
*mddev
)
3816 struct r10conf
*conf
= mddev
->private;
3818 raise_barrier(conf
, 0);
3819 lower_barrier(conf
);
3821 md_unregister_thread(&mddev
->thread
);
3823 /* the unplug fn references 'conf'*/
3824 blk_sync_queue(mddev
->queue
);
3826 if (conf
->r10bio_pool
)
3827 mempool_destroy(conf
->r10bio_pool
);
3828 safe_put_page(conf
->tmppage
);
3829 kfree(conf
->mirrors
);
3830 kfree(conf
->mirrors_old
);
3831 kfree(conf
->mirrors_new
);
3833 mddev
->private = NULL
;
3837 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3839 struct r10conf
*conf
= mddev
->private;
3843 raise_barrier(conf
, 0);
3846 lower_barrier(conf
);
3851 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3853 /* Resize of 'far' arrays is not supported.
3854 * For 'near' and 'offset' arrays we can set the
3855 * number of sectors used to be an appropriate multiple
3856 * of the chunk size.
3857 * For 'offset', this is far_copies*chunksize.
3858 * For 'near' the multiplier is the LCM of
3859 * near_copies and raid_disks.
3860 * So if far_copies > 1 && !far_offset, fail.
3861 * Else find LCM(raid_disks, near_copy)*far_copies and
3862 * multiply by chunk_size. Then round to this number.
3863 * This is mostly done by raid10_size()
3865 struct r10conf
*conf
= mddev
->private;
3866 sector_t oldsize
, size
;
3868 if (mddev
->reshape_position
!= MaxSector
)
3871 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3874 oldsize
= raid10_size(mddev
, 0, 0);
3875 size
= raid10_size(mddev
, sectors
, 0);
3876 if (mddev
->external_size
&&
3877 mddev
->array_sectors
> size
)
3879 if (mddev
->bitmap
) {
3880 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3884 md_set_array_sectors(mddev
, size
);
3885 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3886 revalidate_disk(mddev
->gendisk
);
3887 if (sectors
> mddev
->dev_sectors
&&
3888 mddev
->recovery_cp
> oldsize
) {
3889 mddev
->recovery_cp
= oldsize
;
3890 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3892 calc_sectors(conf
, sectors
);
3893 mddev
->dev_sectors
= conf
->dev_sectors
;
3894 mddev
->resync_max_sectors
= size
;
3898 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3900 struct md_rdev
*rdev
;
3901 struct r10conf
*conf
;
3903 if (mddev
->degraded
> 0) {
3904 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3906 return ERR_PTR(-EINVAL
);
3909 /* Set new parameters */
3910 mddev
->new_level
= 10;
3911 /* new layout: far_copies = 1, near_copies = 2 */
3912 mddev
->new_layout
= (1<<8) + 2;
3913 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3914 mddev
->delta_disks
= mddev
->raid_disks
;
3915 mddev
->raid_disks
*= 2;
3916 /* make sure it will be not marked as dirty */
3917 mddev
->recovery_cp
= MaxSector
;
3919 conf
= setup_conf(mddev
);
3920 if (!IS_ERR(conf
)) {
3921 rdev_for_each(rdev
, mddev
)
3922 if (rdev
->raid_disk
>= 0)
3923 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3930 static void *raid10_takeover(struct mddev
*mddev
)
3932 struct r0conf
*raid0_conf
;
3934 /* raid10 can take over:
3935 * raid0 - providing it has only two drives
3937 if (mddev
->level
== 0) {
3938 /* for raid0 takeover only one zone is supported */
3939 raid0_conf
= mddev
->private;
3940 if (raid0_conf
->nr_strip_zones
> 1) {
3941 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3942 " with more than one zone.\n",
3944 return ERR_PTR(-EINVAL
);
3946 return raid10_takeover_raid0(mddev
);
3948 return ERR_PTR(-EINVAL
);
3951 static int raid10_check_reshape(struct mddev
*mddev
)
3953 /* Called when there is a request to change
3954 * - layout (to ->new_layout)
3955 * - chunk size (to ->new_chunk_sectors)
3956 * - raid_disks (by delta_disks)
3957 * or when trying to restart a reshape that was ongoing.
3959 * We need to validate the request and possibly allocate
3960 * space if that might be an issue later.
3962 * Currently we reject any reshape of a 'far' mode array,
3963 * allow chunk size to change if new is generally acceptable,
3964 * allow raid_disks to increase, and allow
3965 * a switch between 'near' mode and 'offset' mode.
3967 struct r10conf
*conf
= mddev
->private;
3970 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3973 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3974 /* mustn't change number of copies */
3976 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3977 /* Cannot switch to 'far' mode */
3980 if (mddev
->array_sectors
& geo
.chunk_mask
)
3981 /* not factor of array size */
3984 if (!enough(conf
, -1))
3987 kfree(conf
->mirrors_new
);
3988 conf
->mirrors_new
= NULL
;
3989 if (mddev
->delta_disks
> 0) {
3990 /* allocate new 'mirrors' list */
3991 conf
->mirrors_new
= kzalloc(
3992 sizeof(struct raid10_info
)
3993 *(mddev
->raid_disks
+
3994 mddev
->delta_disks
),
3996 if (!conf
->mirrors_new
)
4003 * Need to check if array has failed when deciding whether to:
4005 * - remove non-faulty devices
4008 * This determination is simple when no reshape is happening.
4009 * However if there is a reshape, we need to carefully check
4010 * both the before and after sections.
4011 * This is because some failed devices may only affect one
4012 * of the two sections, and some non-in_sync devices may
4013 * be insync in the section most affected by failed devices.
4015 static int calc_degraded(struct r10conf
*conf
)
4017 int degraded
, degraded2
;
4022 /* 'prev' section first */
4023 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4024 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4025 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4027 else if (!test_bit(In_sync
, &rdev
->flags
))
4028 /* When we can reduce the number of devices in
4029 * an array, this might not contribute to
4030 * 'degraded'. It does now.
4035 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4039 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4040 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4041 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4043 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4044 /* If reshape is increasing the number of devices,
4045 * this section has already been recovered, so
4046 * it doesn't contribute to degraded.
4049 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4054 if (degraded2
> degraded
)
4059 static int raid10_start_reshape(struct mddev
*mddev
)
4061 /* A 'reshape' has been requested. This commits
4062 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4063 * This also checks if there are enough spares and adds them
4065 * We currently require enough spares to make the final
4066 * array non-degraded. We also require that the difference
4067 * between old and new data_offset - on each device - is
4068 * enough that we never risk over-writing.
4071 unsigned long before_length
, after_length
;
4072 sector_t min_offset_diff
= 0;
4075 struct r10conf
*conf
= mddev
->private;
4076 struct md_rdev
*rdev
;
4080 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4083 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4086 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4087 conf
->prev
.far_copies
);
4088 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4089 conf
->geo
.far_copies
);
4091 rdev_for_each(rdev
, mddev
) {
4092 if (!test_bit(In_sync
, &rdev
->flags
)
4093 && !test_bit(Faulty
, &rdev
->flags
))
4095 if (rdev
->raid_disk
>= 0) {
4096 long long diff
= (rdev
->new_data_offset
4097 - rdev
->data_offset
);
4098 if (!mddev
->reshape_backwards
)
4102 if (first
|| diff
< min_offset_diff
)
4103 min_offset_diff
= diff
;
4107 if (max(before_length
, after_length
) > min_offset_diff
)
4110 if (spares
< mddev
->delta_disks
)
4113 conf
->offset_diff
= min_offset_diff
;
4114 spin_lock_irq(&conf
->device_lock
);
4115 if (conf
->mirrors_new
) {
4116 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4117 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4119 kfree(conf
->mirrors_old
);
4120 conf
->mirrors_old
= conf
->mirrors
;
4121 conf
->mirrors
= conf
->mirrors_new
;
4122 conf
->mirrors_new
= NULL
;
4124 setup_geo(&conf
->geo
, mddev
, geo_start
);
4126 if (mddev
->reshape_backwards
) {
4127 sector_t size
= raid10_size(mddev
, 0, 0);
4128 if (size
< mddev
->array_sectors
) {
4129 spin_unlock_irq(&conf
->device_lock
);
4130 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4134 mddev
->resync_max_sectors
= size
;
4135 conf
->reshape_progress
= size
;
4137 conf
->reshape_progress
= 0;
4138 spin_unlock_irq(&conf
->device_lock
);
4140 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4141 ret
= bitmap_resize(mddev
->bitmap
,
4142 raid10_size(mddev
, 0,
4143 conf
->geo
.raid_disks
),
4148 if (mddev
->delta_disks
> 0) {
4149 rdev_for_each(rdev
, mddev
)
4150 if (rdev
->raid_disk
< 0 &&
4151 !test_bit(Faulty
, &rdev
->flags
)) {
4152 if (raid10_add_disk(mddev
, rdev
) == 0) {
4153 if (rdev
->raid_disk
>=
4154 conf
->prev
.raid_disks
)
4155 set_bit(In_sync
, &rdev
->flags
);
4157 rdev
->recovery_offset
= 0;
4159 if (sysfs_link_rdev(mddev
, rdev
))
4160 /* Failure here is OK */;
4162 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4163 && !test_bit(Faulty
, &rdev
->flags
)) {
4164 /* This is a spare that was manually added */
4165 set_bit(In_sync
, &rdev
->flags
);
4168 /* When a reshape changes the number of devices,
4169 * ->degraded is measured against the larger of the
4170 * pre and post numbers.
4172 spin_lock_irq(&conf
->device_lock
);
4173 mddev
->degraded
= calc_degraded(conf
);
4174 spin_unlock_irq(&conf
->device_lock
);
4175 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4176 mddev
->reshape_position
= conf
->reshape_progress
;
4177 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4179 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4180 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4181 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4182 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4184 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4186 if (!mddev
->sync_thread
) {
4190 conf
->reshape_checkpoint
= jiffies
;
4191 md_wakeup_thread(mddev
->sync_thread
);
4192 md_new_event(mddev
);
4196 mddev
->recovery
= 0;
4197 spin_lock_irq(&conf
->device_lock
);
4198 conf
->geo
= conf
->prev
;
4199 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4200 rdev_for_each(rdev
, mddev
)
4201 rdev
->new_data_offset
= rdev
->data_offset
;
4203 conf
->reshape_progress
= MaxSector
;
4204 mddev
->reshape_position
= MaxSector
;
4205 spin_unlock_irq(&conf
->device_lock
);
4209 /* Calculate the last device-address that could contain
4210 * any block from the chunk that includes the array-address 's'
4211 * and report the next address.
4212 * i.e. the address returned will be chunk-aligned and after
4213 * any data that is in the chunk containing 's'.
4215 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4217 s
= (s
| geo
->chunk_mask
) + 1;
4218 s
>>= geo
->chunk_shift
;
4219 s
*= geo
->near_copies
;
4220 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4221 s
*= geo
->far_copies
;
4222 s
<<= geo
->chunk_shift
;
4226 /* Calculate the first device-address that could contain
4227 * any block from the chunk that includes the array-address 's'.
4228 * This too will be the start of a chunk
4230 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4232 s
>>= geo
->chunk_shift
;
4233 s
*= geo
->near_copies
;
4234 sector_div(s
, geo
->raid_disks
);
4235 s
*= geo
->far_copies
;
4236 s
<<= geo
->chunk_shift
;
4240 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4243 /* We simply copy at most one chunk (smallest of old and new)
4244 * at a time, possibly less if that exceeds RESYNC_PAGES,
4245 * or we hit a bad block or something.
4246 * This might mean we pause for normal IO in the middle of
4247 * a chunk, but that is not a problem was mddev->reshape_position
4248 * can record any location.
4250 * If we will want to write to a location that isn't
4251 * yet recorded as 'safe' (i.e. in metadata on disk) then
4252 * we need to flush all reshape requests and update the metadata.
4254 * When reshaping forwards (e.g. to more devices), we interpret
4255 * 'safe' as the earliest block which might not have been copied
4256 * down yet. We divide this by previous stripe size and multiply
4257 * by previous stripe length to get lowest device offset that we
4258 * cannot write to yet.
4259 * We interpret 'sector_nr' as an address that we want to write to.
4260 * From this we use last_device_address() to find where we might
4261 * write to, and first_device_address on the 'safe' position.
4262 * If this 'next' write position is after the 'safe' position,
4263 * we must update the metadata to increase the 'safe' position.
4265 * When reshaping backwards, we round in the opposite direction
4266 * and perform the reverse test: next write position must not be
4267 * less than current safe position.
4269 * In all this the minimum difference in data offsets
4270 * (conf->offset_diff - always positive) allows a bit of slack,
4271 * so next can be after 'safe', but not by more than offset_disk
4273 * We need to prepare all the bios here before we start any IO
4274 * to ensure the size we choose is acceptable to all devices.
4275 * The means one for each copy for write-out and an extra one for
4277 * We store the read-in bio in ->master_bio and the others in
4278 * ->devs[x].bio and ->devs[x].repl_bio.
4280 struct r10conf
*conf
= mddev
->private;
4281 struct r10bio
*r10_bio
;
4282 sector_t next
, safe
, last
;
4286 struct md_rdev
*rdev
;
4289 struct bio
*bio
, *read_bio
;
4290 int sectors_done
= 0;
4292 if (sector_nr
== 0) {
4293 /* If restarting in the middle, skip the initial sectors */
4294 if (mddev
->reshape_backwards
&&
4295 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4296 sector_nr
= (raid10_size(mddev
, 0, 0)
4297 - conf
->reshape_progress
);
4298 } else if (!mddev
->reshape_backwards
&&
4299 conf
->reshape_progress
> 0)
4300 sector_nr
= conf
->reshape_progress
;
4302 mddev
->curr_resync_completed
= sector_nr
;
4303 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4309 /* We don't use sector_nr to track where we are up to
4310 * as that doesn't work well for ->reshape_backwards.
4311 * So just use ->reshape_progress.
4313 if (mddev
->reshape_backwards
) {
4314 /* 'next' is the earliest device address that we might
4315 * write to for this chunk in the new layout
4317 next
= first_dev_address(conf
->reshape_progress
- 1,
4320 /* 'safe' is the last device address that we might read from
4321 * in the old layout after a restart
4323 safe
= last_dev_address(conf
->reshape_safe
- 1,
4326 if (next
+ conf
->offset_diff
< safe
)
4329 last
= conf
->reshape_progress
- 1;
4330 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4331 & conf
->prev
.chunk_mask
);
4332 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4333 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4335 /* 'next' is after the last device address that we
4336 * might write to for this chunk in the new layout
4338 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4340 /* 'safe' is the earliest device address that we might
4341 * read from in the old layout after a restart
4343 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4345 /* Need to update metadata if 'next' might be beyond 'safe'
4346 * as that would possibly corrupt data
4348 if (next
> safe
+ conf
->offset_diff
)
4351 sector_nr
= conf
->reshape_progress
;
4352 last
= sector_nr
| (conf
->geo
.chunk_mask
4353 & conf
->prev
.chunk_mask
);
4355 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4356 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4360 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4361 /* Need to update reshape_position in metadata */
4363 mddev
->reshape_position
= conf
->reshape_progress
;
4364 if (mddev
->reshape_backwards
)
4365 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4366 - conf
->reshape_progress
;
4368 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4369 conf
->reshape_checkpoint
= jiffies
;
4370 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4371 md_wakeup_thread(mddev
->thread
);
4372 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4373 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4374 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4375 allow_barrier(conf
);
4376 return sectors_done
;
4378 conf
->reshape_safe
= mddev
->reshape_position
;
4379 allow_barrier(conf
);
4383 /* Now schedule reads for blocks from sector_nr to last */
4384 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4386 raise_barrier(conf
, sectors_done
!= 0);
4387 atomic_set(&r10_bio
->remaining
, 0);
4388 r10_bio
->mddev
= mddev
;
4389 r10_bio
->sector
= sector_nr
;
4390 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4391 r10_bio
->sectors
= last
- sector_nr
+ 1;
4392 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4393 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4396 /* Cannot read from here, so need to record bad blocks
4397 * on all the target devices.
4400 mempool_free(r10_bio
, conf
->r10buf_pool
);
4401 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4402 return sectors_done
;
4405 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4407 read_bio
->bi_bdev
= rdev
->bdev
;
4408 read_bio
->bi_iter
.bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4409 + rdev
->data_offset
);
4410 read_bio
->bi_private
= r10_bio
;
4411 read_bio
->bi_end_io
= end_sync_read
;
4412 read_bio
->bi_rw
= READ
;
4413 read_bio
->bi_flags
&= (~0UL << BIO_RESET_BITS
);
4414 __set_bit(BIO_UPTODATE
, &read_bio
->bi_flags
);
4415 read_bio
->bi_vcnt
= 0;
4416 read_bio
->bi_iter
.bi_size
= 0;
4417 r10_bio
->master_bio
= read_bio
;
4418 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4420 /* Now find the locations in the new layout */
4421 __raid10_find_phys(&conf
->geo
, r10_bio
);
4424 read_bio
->bi_next
= NULL
;
4426 for (s
= 0; s
< conf
->copies
*2; s
++) {
4428 int d
= r10_bio
->devs
[s
/2].devnum
;
4429 struct md_rdev
*rdev2
;
4431 rdev2
= conf
->mirrors
[d
].replacement
;
4432 b
= r10_bio
->devs
[s
/2].repl_bio
;
4434 rdev2
= conf
->mirrors
[d
].rdev
;
4435 b
= r10_bio
->devs
[s
/2].bio
;
4437 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4441 b
->bi_bdev
= rdev2
->bdev
;
4442 b
->bi_iter
.bi_sector
= r10_bio
->devs
[s
/2].addr
+
4443 rdev2
->new_data_offset
;
4444 b
->bi_private
= r10_bio
;
4445 b
->bi_end_io
= end_reshape_write
;
4451 /* Now add as many pages as possible to all of these bios. */
4454 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4455 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4456 int len
= (max_sectors
- s
) << 9;
4457 if (len
> PAGE_SIZE
)
4459 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4461 if (bio_add_page(bio
, page
, len
, 0))
4464 /* Didn't fit, must stop */
4466 bio2
&& bio2
!= bio
;
4467 bio2
= bio2
->bi_next
) {
4468 /* Remove last page from this bio */
4470 bio2
->bi_iter
.bi_size
-= len
;
4471 __clear_bit(BIO_SEG_VALID
, &bio2
->bi_flags
);
4475 sector_nr
+= len
>> 9;
4476 nr_sectors
+= len
>> 9;
4479 r10_bio
->sectors
= nr_sectors
;
4481 /* Now submit the read */
4482 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4483 atomic_inc(&r10_bio
->remaining
);
4484 read_bio
->bi_next
= NULL
;
4485 generic_make_request(read_bio
);
4486 sector_nr
+= nr_sectors
;
4487 sectors_done
+= nr_sectors
;
4488 if (sector_nr
<= last
)
4491 /* Now that we have done the whole section we can
4492 * update reshape_progress
4494 if (mddev
->reshape_backwards
)
4495 conf
->reshape_progress
-= sectors_done
;
4497 conf
->reshape_progress
+= sectors_done
;
4499 return sectors_done
;
4502 static void end_reshape_request(struct r10bio
*r10_bio
);
4503 static int handle_reshape_read_error(struct mddev
*mddev
,
4504 struct r10bio
*r10_bio
);
4505 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4507 /* Reshape read completed. Hopefully we have a block
4509 * If we got a read error then we do sync 1-page reads from
4510 * elsewhere until we find the data - or give up.
4512 struct r10conf
*conf
= mddev
->private;
4515 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4516 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4517 /* Reshape has been aborted */
4518 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4522 /* We definitely have the data in the pages, schedule the
4525 atomic_set(&r10_bio
->remaining
, 1);
4526 for (s
= 0; s
< conf
->copies
*2; s
++) {
4528 int d
= r10_bio
->devs
[s
/2].devnum
;
4529 struct md_rdev
*rdev
;
4531 rdev
= conf
->mirrors
[d
].replacement
;
4532 b
= r10_bio
->devs
[s
/2].repl_bio
;
4534 rdev
= conf
->mirrors
[d
].rdev
;
4535 b
= r10_bio
->devs
[s
/2].bio
;
4537 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4539 atomic_inc(&rdev
->nr_pending
);
4540 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4541 atomic_inc(&r10_bio
->remaining
);
4543 generic_make_request(b
);
4545 end_reshape_request(r10_bio
);
4548 static void end_reshape(struct r10conf
*conf
)
4550 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4553 spin_lock_irq(&conf
->device_lock
);
4554 conf
->prev
= conf
->geo
;
4555 md_finish_reshape(conf
->mddev
);
4557 conf
->reshape_progress
= MaxSector
;
4558 spin_unlock_irq(&conf
->device_lock
);
4560 /* read-ahead size must cover two whole stripes, which is
4561 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4563 if (conf
->mddev
->queue
) {
4564 int stripe
= conf
->geo
.raid_disks
*
4565 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4566 stripe
/= conf
->geo
.near_copies
;
4567 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4568 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4573 static int handle_reshape_read_error(struct mddev
*mddev
,
4574 struct r10bio
*r10_bio
)
4576 /* Use sync reads to get the blocks from somewhere else */
4577 int sectors
= r10_bio
->sectors
;
4578 struct r10conf
*conf
= mddev
->private;
4580 struct r10bio r10_bio
;
4581 struct r10dev devs
[conf
->copies
];
4583 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4586 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4588 r10b
->sector
= r10_bio
->sector
;
4589 __raid10_find_phys(&conf
->prev
, r10b
);
4594 int first_slot
= slot
;
4596 if (s
> (PAGE_SIZE
>> 9))
4600 int d
= r10b
->devs
[slot
].devnum
;
4601 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4604 test_bit(Faulty
, &rdev
->flags
) ||
4605 !test_bit(In_sync
, &rdev
->flags
))
4608 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4609 success
= sync_page_io(rdev
,
4618 if (slot
>= conf
->copies
)
4620 if (slot
== first_slot
)
4624 /* couldn't read this block, must give up */
4625 set_bit(MD_RECOVERY_INTR
,
4635 static void end_reshape_write(struct bio
*bio
, int error
)
4637 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4638 struct r10bio
*r10_bio
= bio
->bi_private
;
4639 struct mddev
*mddev
= r10_bio
->mddev
;
4640 struct r10conf
*conf
= mddev
->private;
4644 struct md_rdev
*rdev
= NULL
;
4646 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4648 rdev
= conf
->mirrors
[d
].replacement
;
4651 rdev
= conf
->mirrors
[d
].rdev
;
4655 /* FIXME should record badblock */
4656 md_error(mddev
, rdev
);
4659 rdev_dec_pending(rdev
, mddev
);
4660 end_reshape_request(r10_bio
);
4663 static void end_reshape_request(struct r10bio
*r10_bio
)
4665 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4667 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4668 bio_put(r10_bio
->master_bio
);
4672 static void raid10_finish_reshape(struct mddev
*mddev
)
4674 struct r10conf
*conf
= mddev
->private;
4676 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4679 if (mddev
->delta_disks
> 0) {
4680 sector_t size
= raid10_size(mddev
, 0, 0);
4681 md_set_array_sectors(mddev
, size
);
4682 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4683 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4684 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4686 mddev
->resync_max_sectors
= size
;
4687 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4688 revalidate_disk(mddev
->gendisk
);
4691 for (d
= conf
->geo
.raid_disks
;
4692 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4694 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4696 clear_bit(In_sync
, &rdev
->flags
);
4697 rdev
= conf
->mirrors
[d
].replacement
;
4699 clear_bit(In_sync
, &rdev
->flags
);
4702 mddev
->layout
= mddev
->new_layout
;
4703 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4704 mddev
->reshape_position
= MaxSector
;
4705 mddev
->delta_disks
= 0;
4706 mddev
->reshape_backwards
= 0;
4709 static struct md_personality raid10_personality
=
4713 .owner
= THIS_MODULE
,
4714 .make_request
= make_request
,
4718 .error_handler
= error
,
4719 .hot_add_disk
= raid10_add_disk
,
4720 .hot_remove_disk
= raid10_remove_disk
,
4721 .spare_active
= raid10_spare_active
,
4722 .sync_request
= sync_request
,
4723 .quiesce
= raid10_quiesce
,
4724 .size
= raid10_size
,
4725 .resize
= raid10_resize
,
4726 .takeover
= raid10_takeover
,
4727 .check_reshape
= raid10_check_reshape
,
4728 .start_reshape
= raid10_start_reshape
,
4729 .finish_reshape
= raid10_finish_reshape
,
4732 static int __init
raid_init(void)
4734 return register_md_personality(&raid10_personality
);
4737 static void raid_exit(void)
4739 unregister_md_personality(&raid10_personality
);
4742 module_init(raid_init
);
4743 module_exit(raid_exit
);
4744 MODULE_LICENSE("GPL");
4745 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4746 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4747 MODULE_ALIAS("md-raid10");
4748 MODULE_ALIAS("md-level-10");
4750 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);