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;
370 slot
= r10_bio
->read_slot
;
371 dev
= r10_bio
->devs
[slot
].devnum
;
372 rdev
= r10_bio
->devs
[slot
].rdev
;
374 * this branch is our 'one mirror IO has finished' event handler:
376 update_head_pos(slot
, r10_bio
);
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
388 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
395 if (!_enough(conf
, test_bit(R10BIO_Previous
, &r10_bio
->state
),
400 raid_end_bio_io(r10_bio
);
401 rdev_dec_pending(rdev
, conf
->mddev
);
404 * oops, read error - keep the refcount on the rdev
406 char b
[BDEVNAME_SIZE
];
407 printk_ratelimited(KERN_ERR
408 "md/raid10:%s: %s: rescheduling sector %llu\n",
410 bdevname(rdev
->bdev
, b
),
411 (unsigned long long)r10_bio
->sector
);
412 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
413 reschedule_retry(r10_bio
);
417 static void close_write(struct r10bio
*r10_bio
)
419 /* clear the bitmap if all writes complete successfully */
420 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
422 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
424 md_write_end(r10_bio
->mddev
);
427 static void one_write_done(struct r10bio
*r10_bio
)
429 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
430 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
431 reschedule_retry(r10_bio
);
433 close_write(r10_bio
);
434 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
435 reschedule_retry(r10_bio
);
437 raid_end_bio_io(r10_bio
);
442 static void raid10_end_write_request(struct bio
*bio
, int error
)
444 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
445 struct r10bio
*r10_bio
= bio
->bi_private
;
448 struct r10conf
*conf
= r10_bio
->mddev
->private;
450 struct md_rdev
*rdev
= NULL
;
452 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
455 rdev
= conf
->mirrors
[dev
].replacement
;
459 rdev
= conf
->mirrors
[dev
].rdev
;
462 * this branch is our 'one mirror IO has finished' event handler:
466 /* Never record new bad blocks to replacement,
469 md_error(rdev
->mddev
, rdev
);
471 set_bit(WriteErrorSeen
, &rdev
->flags
);
472 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
473 set_bit(MD_RECOVERY_NEEDED
,
474 &rdev
->mddev
->recovery
);
475 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
480 * Set R10BIO_Uptodate in our master bio, so that
481 * we will return a good error code for to the higher
482 * levels even if IO on some other mirrored buffer fails.
484 * The 'master' represents the composite IO operation to
485 * user-side. So if something waits for IO, then it will
486 * wait for the 'master' bio.
492 * Do not set R10BIO_Uptodate if the current device is
493 * rebuilding or Faulty. This is because we cannot use
494 * such device for properly reading the data back (we could
495 * potentially use it, if the current write would have felt
496 * before rdev->recovery_offset, but for simplicity we don't
499 if (test_bit(In_sync
, &rdev
->flags
) &&
500 !test_bit(Faulty
, &rdev
->flags
))
501 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
503 /* Maybe we can clear some bad blocks. */
504 if (is_badblock(rdev
,
505 r10_bio
->devs
[slot
].addr
,
507 &first_bad
, &bad_sectors
)) {
510 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
512 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
514 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
520 * Let's see if all mirrored write operations have finished
523 one_write_done(r10_bio
);
525 rdev_dec_pending(rdev
, conf
->mddev
);
529 * RAID10 layout manager
530 * As well as the chunksize and raid_disks count, there are two
531 * parameters: near_copies and far_copies.
532 * near_copies * far_copies must be <= raid_disks.
533 * Normally one of these will be 1.
534 * If both are 1, we get raid0.
535 * If near_copies == raid_disks, we get raid1.
537 * Chunks are laid out in raid0 style with near_copies copies of the
538 * first chunk, followed by near_copies copies of the next chunk and
540 * If far_copies > 1, then after 1/far_copies of the array has been assigned
541 * as described above, we start again with a device offset of near_copies.
542 * So we effectively have another copy of the whole array further down all
543 * the drives, but with blocks on different drives.
544 * With this layout, and block is never stored twice on the one device.
546 * raid10_find_phys finds the sector offset of a given virtual sector
547 * on each device that it is on.
549 * raid10_find_virt does the reverse mapping, from a device and a
550 * sector offset to a virtual address
553 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
561 int last_far_set_start
, last_far_set_size
;
563 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
564 last_far_set_start
*= geo
->far_set_size
;
566 last_far_set_size
= geo
->far_set_size
;
567 last_far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
569 /* now calculate first sector/dev */
570 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
571 sector
= r10bio
->sector
& geo
->chunk_mask
;
573 chunk
*= geo
->near_copies
;
575 dev
= sector_div(stripe
, geo
->raid_disks
);
577 stripe
*= geo
->far_copies
;
579 sector
+= stripe
<< geo
->chunk_shift
;
581 /* and calculate all the others */
582 for (n
= 0; n
< geo
->near_copies
; n
++) {
586 r10bio
->devs
[slot
].devnum
= d
;
587 r10bio
->devs
[slot
].addr
= s
;
590 for (f
= 1; f
< geo
->far_copies
; f
++) {
591 set
= d
/ geo
->far_set_size
;
592 d
+= geo
->near_copies
;
594 if ((geo
->raid_disks
% geo
->far_set_size
) &&
595 (d
> last_far_set_start
)) {
596 d
-= last_far_set_start
;
597 d
%= last_far_set_size
;
598 d
+= last_far_set_start
;
600 d
%= geo
->far_set_size
;
601 d
+= geo
->far_set_size
* set
;
604 r10bio
->devs
[slot
].devnum
= d
;
605 r10bio
->devs
[slot
].addr
= s
;
609 if (dev
>= geo
->raid_disks
) {
611 sector
+= (geo
->chunk_mask
+ 1);
616 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
618 struct geom
*geo
= &conf
->geo
;
620 if (conf
->reshape_progress
!= MaxSector
&&
621 ((r10bio
->sector
>= conf
->reshape_progress
) !=
622 conf
->mddev
->reshape_backwards
)) {
623 set_bit(R10BIO_Previous
, &r10bio
->state
);
626 clear_bit(R10BIO_Previous
, &r10bio
->state
);
628 __raid10_find_phys(geo
, r10bio
);
631 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
633 sector_t offset
, chunk
, vchunk
;
634 /* Never use conf->prev as this is only called during resync
635 * or recovery, so reshape isn't happening
637 struct geom
*geo
= &conf
->geo
;
638 int far_set_start
= (dev
/ geo
->far_set_size
) * geo
->far_set_size
;
639 int far_set_size
= geo
->far_set_size
;
640 int last_far_set_start
;
642 if (geo
->raid_disks
% geo
->far_set_size
) {
643 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
644 last_far_set_start
*= geo
->far_set_size
;
646 if (dev
>= last_far_set_start
) {
647 far_set_size
= geo
->far_set_size
;
648 far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
649 far_set_start
= last_far_set_start
;
653 offset
= sector
& geo
->chunk_mask
;
654 if (geo
->far_offset
) {
656 chunk
= sector
>> geo
->chunk_shift
;
657 fc
= sector_div(chunk
, geo
->far_copies
);
658 dev
-= fc
* geo
->near_copies
;
659 if (dev
< far_set_start
)
662 while (sector
>= geo
->stride
) {
663 sector
-= geo
->stride
;
664 if (dev
< (geo
->near_copies
+ far_set_start
))
665 dev
+= far_set_size
- geo
->near_copies
;
667 dev
-= geo
->near_copies
;
669 chunk
= sector
>> geo
->chunk_shift
;
671 vchunk
= chunk
* geo
->raid_disks
+ dev
;
672 sector_div(vchunk
, geo
->near_copies
);
673 return (vchunk
<< geo
->chunk_shift
) + offset
;
677 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
679 * @bvm: properties of new bio
680 * @biovec: the request that could be merged to it.
682 * Return amount of bytes we can accept at this offset
683 * This requires checking for end-of-chunk if near_copies != raid_disks,
684 * and for subordinate merge_bvec_fns if merge_check_needed.
686 static int raid10_mergeable_bvec(struct request_queue
*q
,
687 struct bvec_merge_data
*bvm
,
688 struct bio_vec
*biovec
)
690 struct mddev
*mddev
= q
->queuedata
;
691 struct r10conf
*conf
= mddev
->private;
692 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
694 unsigned int chunk_sectors
;
695 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
696 struct geom
*geo
= &conf
->geo
;
698 chunk_sectors
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
) + 1;
699 if (conf
->reshape_progress
!= MaxSector
&&
700 ((sector
>= conf
->reshape_progress
) !=
701 conf
->mddev
->reshape_backwards
))
704 if (geo
->near_copies
< geo
->raid_disks
) {
705 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1))
706 + bio_sectors
)) << 9;
708 /* bio_add cannot handle a negative return */
710 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
711 return biovec
->bv_len
;
713 max
= biovec
->bv_len
;
715 if (mddev
->merge_check_needed
) {
717 struct r10bio r10_bio
;
718 struct r10dev devs
[conf
->copies
];
720 struct r10bio
*r10_bio
= &on_stack
.r10_bio
;
722 if (conf
->reshape_progress
!= MaxSector
) {
723 /* Cannot give any guidance during reshape */
724 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
725 return biovec
->bv_len
;
728 r10_bio
->sector
= sector
;
729 raid10_find_phys(conf
, r10_bio
);
731 for (s
= 0; s
< conf
->copies
; s
++) {
732 int disk
= r10_bio
->devs
[s
].devnum
;
733 struct md_rdev
*rdev
= rcu_dereference(
734 conf
->mirrors
[disk
].rdev
);
735 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
736 struct request_queue
*q
=
737 bdev_get_queue(rdev
->bdev
);
738 if (q
->merge_bvec_fn
) {
739 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
741 bvm
->bi_bdev
= rdev
->bdev
;
742 max
= min(max
, q
->merge_bvec_fn(
746 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
747 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
748 struct request_queue
*q
=
749 bdev_get_queue(rdev
->bdev
);
750 if (q
->merge_bvec_fn
) {
751 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
753 bvm
->bi_bdev
= rdev
->bdev
;
754 max
= min(max
, q
->merge_bvec_fn(
765 * This routine returns the disk from which the requested read should
766 * be done. There is a per-array 'next expected sequential IO' sector
767 * number - if this matches on the next IO then we use the last disk.
768 * There is also a per-disk 'last know head position' sector that is
769 * maintained from IRQ contexts, both the normal and the resync IO
770 * completion handlers update this position correctly. If there is no
771 * perfect sequential match then we pick the disk whose head is closest.
773 * If there are 2 mirrors in the same 2 devices, performance degrades
774 * because position is mirror, not device based.
776 * The rdev for the device selected will have nr_pending incremented.
780 * FIXME: possibly should rethink readbalancing and do it differently
781 * depending on near_copies / far_copies geometry.
783 static struct md_rdev
*read_balance(struct r10conf
*conf
,
784 struct r10bio
*r10_bio
,
787 const sector_t this_sector
= r10_bio
->sector
;
789 int sectors
= r10_bio
->sectors
;
790 int best_good_sectors
;
791 sector_t new_distance
, best_dist
;
792 struct md_rdev
*best_rdev
, *rdev
= NULL
;
795 struct geom
*geo
= &conf
->geo
;
797 raid10_find_phys(conf
, r10_bio
);
800 sectors
= r10_bio
->sectors
;
803 best_dist
= MaxSector
;
804 best_good_sectors
= 0;
807 * Check if we can balance. We can balance on the whole
808 * device if no resync is going on (recovery is ok), or below
809 * the resync window. We take the first readable disk when
810 * above the resync window.
812 if (conf
->mddev
->recovery_cp
< MaxSector
813 && (this_sector
+ sectors
>= conf
->next_resync
))
816 for (slot
= 0; slot
< conf
->copies
; slot
++) {
821 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
823 disk
= r10_bio
->devs
[slot
].devnum
;
824 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
825 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
826 test_bit(Unmerged
, &rdev
->flags
) ||
827 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
828 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
830 test_bit(Faulty
, &rdev
->flags
) ||
831 test_bit(Unmerged
, &rdev
->flags
))
833 if (!test_bit(In_sync
, &rdev
->flags
) &&
834 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
837 dev_sector
= r10_bio
->devs
[slot
].addr
;
838 if (is_badblock(rdev
, dev_sector
, sectors
,
839 &first_bad
, &bad_sectors
)) {
840 if (best_dist
< MaxSector
)
841 /* Already have a better slot */
843 if (first_bad
<= dev_sector
) {
844 /* Cannot read here. If this is the
845 * 'primary' device, then we must not read
846 * beyond 'bad_sectors' from another device.
848 bad_sectors
-= (dev_sector
- first_bad
);
849 if (!do_balance
&& sectors
> bad_sectors
)
850 sectors
= bad_sectors
;
851 if (best_good_sectors
> sectors
)
852 best_good_sectors
= sectors
;
854 sector_t good_sectors
=
855 first_bad
- dev_sector
;
856 if (good_sectors
> best_good_sectors
) {
857 best_good_sectors
= good_sectors
;
862 /* Must read from here */
867 best_good_sectors
= sectors
;
872 /* This optimisation is debatable, and completely destroys
873 * sequential read speed for 'far copies' arrays. So only
874 * keep it for 'near' arrays, and review those later.
876 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
879 /* for far > 1 always use the lowest address */
880 if (geo
->far_copies
> 1)
881 new_distance
= r10_bio
->devs
[slot
].addr
;
883 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
884 conf
->mirrors
[disk
].head_position
);
885 if (new_distance
< best_dist
) {
886 best_dist
= new_distance
;
891 if (slot
>= conf
->copies
) {
897 atomic_inc(&rdev
->nr_pending
);
898 if (test_bit(Faulty
, &rdev
->flags
)) {
899 /* Cannot risk returning a device that failed
900 * before we inc'ed nr_pending
902 rdev_dec_pending(rdev
, conf
->mddev
);
905 r10_bio
->read_slot
= slot
;
909 *max_sectors
= best_good_sectors
;
914 int md_raid10_congested(struct mddev
*mddev
, int bits
)
916 struct r10conf
*conf
= mddev
->private;
919 if ((bits
& (1 << BDI_async_congested
)) &&
920 conf
->pending_count
>= max_queued_requests
)
925 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
928 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
929 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
930 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
932 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
938 EXPORT_SYMBOL_GPL(md_raid10_congested
);
940 static int raid10_congested(void *data
, int bits
)
942 struct mddev
*mddev
= data
;
944 return mddev_congested(mddev
, bits
) ||
945 md_raid10_congested(mddev
, bits
);
948 static void flush_pending_writes(struct r10conf
*conf
)
950 /* Any writes that have been queued but are awaiting
951 * bitmap updates get flushed here.
953 spin_lock_irq(&conf
->device_lock
);
955 if (conf
->pending_bio_list
.head
) {
957 bio
= bio_list_get(&conf
->pending_bio_list
);
958 conf
->pending_count
= 0;
959 spin_unlock_irq(&conf
->device_lock
);
960 /* flush any pending bitmap writes to disk
961 * before proceeding w/ I/O */
962 bitmap_unplug(conf
->mddev
->bitmap
);
963 wake_up(&conf
->wait_barrier
);
965 while (bio
) { /* submit pending writes */
966 struct bio
*next
= bio
->bi_next
;
968 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
969 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
973 generic_make_request(bio
);
977 spin_unlock_irq(&conf
->device_lock
);
981 * Sometimes we need to suspend IO while we do something else,
982 * either some resync/recovery, or reconfigure the array.
983 * To do this we raise a 'barrier'.
984 * The 'barrier' is a counter that can be raised multiple times
985 * to count how many activities are happening which preclude
987 * We can only raise the barrier if there is no pending IO.
988 * i.e. if nr_pending == 0.
989 * We choose only to raise the barrier if no-one is waiting for the
990 * barrier to go down. This means that as soon as an IO request
991 * is ready, no other operations which require a barrier will start
992 * until the IO request has had a chance.
994 * So: regular IO calls 'wait_barrier'. When that returns there
995 * is no backgroup IO happening, It must arrange to call
996 * allow_barrier when it has finished its IO.
997 * backgroup IO calls must call raise_barrier. Once that returns
998 * there is no normal IO happeing. It must arrange to call
999 * lower_barrier when the particular background IO completes.
1002 static void raise_barrier(struct r10conf
*conf
, int force
)
1004 BUG_ON(force
&& !conf
->barrier
);
1005 spin_lock_irq(&conf
->resync_lock
);
1007 /* Wait until no block IO is waiting (unless 'force') */
1008 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
1011 /* block any new IO from starting */
1014 /* Now wait for all pending IO to complete */
1015 wait_event_lock_irq(conf
->wait_barrier
,
1016 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
1019 spin_unlock_irq(&conf
->resync_lock
);
1022 static void lower_barrier(struct r10conf
*conf
)
1024 unsigned long flags
;
1025 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1027 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1028 wake_up(&conf
->wait_barrier
);
1031 static void wait_barrier(struct r10conf
*conf
)
1033 spin_lock_irq(&conf
->resync_lock
);
1034 if (conf
->barrier
) {
1036 /* Wait for the barrier to drop.
1037 * However if there are already pending
1038 * requests (preventing the barrier from
1039 * rising completely), and the
1040 * pre-process bio queue isn't empty,
1041 * then don't wait, as we need to empty
1042 * that queue to get the nr_pending
1045 wait_event_lock_irq(conf
->wait_barrier
,
1047 (conf
->nr_pending
&&
1048 current
->bio_list
&&
1049 !bio_list_empty(current
->bio_list
)),
1054 spin_unlock_irq(&conf
->resync_lock
);
1057 static void allow_barrier(struct r10conf
*conf
)
1059 unsigned long flags
;
1060 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1062 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1063 wake_up(&conf
->wait_barrier
);
1066 static void freeze_array(struct r10conf
*conf
, int extra
)
1068 /* stop syncio and normal IO and wait for everything to
1070 * We increment barrier and nr_waiting, and then
1071 * wait until nr_pending match nr_queued+extra
1072 * This is called in the context of one normal IO request
1073 * that has failed. Thus any sync request that might be pending
1074 * will be blocked by nr_pending, and we need to wait for
1075 * pending IO requests to complete or be queued for re-try.
1076 * Thus the number queued (nr_queued) plus this request (extra)
1077 * must match the number of pending IOs (nr_pending) before
1080 spin_lock_irq(&conf
->resync_lock
);
1083 wait_event_lock_irq_cmd(conf
->wait_barrier
,
1084 conf
->nr_pending
== conf
->nr_queued
+extra
,
1086 flush_pending_writes(conf
));
1088 spin_unlock_irq(&conf
->resync_lock
);
1091 static void unfreeze_array(struct r10conf
*conf
)
1093 /* reverse the effect of the freeze */
1094 spin_lock_irq(&conf
->resync_lock
);
1097 wake_up(&conf
->wait_barrier
);
1098 spin_unlock_irq(&conf
->resync_lock
);
1101 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1102 struct md_rdev
*rdev
)
1104 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1105 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1106 return rdev
->data_offset
;
1108 return rdev
->new_data_offset
;
1111 struct raid10_plug_cb
{
1112 struct blk_plug_cb cb
;
1113 struct bio_list pending
;
1117 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1119 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1121 struct mddev
*mddev
= plug
->cb
.data
;
1122 struct r10conf
*conf
= mddev
->private;
1125 if (from_schedule
|| current
->bio_list
) {
1126 spin_lock_irq(&conf
->device_lock
);
1127 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1128 conf
->pending_count
+= plug
->pending_cnt
;
1129 spin_unlock_irq(&conf
->device_lock
);
1130 wake_up(&conf
->wait_barrier
);
1131 md_wakeup_thread(mddev
->thread
);
1136 /* we aren't scheduling, so we can do the write-out directly. */
1137 bio
= bio_list_get(&plug
->pending
);
1138 bitmap_unplug(mddev
->bitmap
);
1139 wake_up(&conf
->wait_barrier
);
1141 while (bio
) { /* submit pending writes */
1142 struct bio
*next
= bio
->bi_next
;
1143 bio
->bi_next
= NULL
;
1144 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
1145 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1146 /* Just ignore it */
1149 generic_make_request(bio
);
1155 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1157 struct r10conf
*conf
= mddev
->private;
1158 struct r10bio
*r10_bio
;
1159 struct bio
*read_bio
;
1161 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1162 int chunk_sects
= chunk_mask
+ 1;
1163 const int rw
= bio_data_dir(bio
);
1164 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1165 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1166 const unsigned long do_discard
= (bio
->bi_rw
1167 & (REQ_DISCARD
| REQ_SECURE
));
1168 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1169 unsigned long flags
;
1170 struct md_rdev
*blocked_rdev
;
1171 struct blk_plug_cb
*cb
;
1172 struct raid10_plug_cb
*plug
= NULL
;
1173 int sectors_handled
;
1177 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1178 md_flush_request(mddev
, bio
);
1182 /* If this request crosses a chunk boundary, we need to
1183 * split it. This will only happen for 1 PAGE (or less) requests.
1185 if (unlikely((bio
->bi_iter
.bi_sector
& chunk_mask
) + bio_sectors(bio
)
1187 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1188 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1189 struct bio_pair
*bp
;
1190 /* Sanity check -- queue functions should prevent this happening */
1191 if (bio_segments(bio
) > 1)
1193 /* This is a one page bio that upper layers
1194 * refuse to split for us, so we need to split it.
1196 bp
= bio_split(bio
, chunk_sects
-
1197 (bio
->bi_iter
.bi_sector
& (chunk_sects
- 1)));
1199 /* Each of these 'make_request' calls will call 'wait_barrier'.
1200 * If the first succeeds but the second blocks due to the resync
1201 * thread raising the barrier, we will deadlock because the
1202 * IO to the underlying device will be queued in generic_make_request
1203 * and will never complete, so will never reduce nr_pending.
1204 * So increment nr_waiting here so no new raise_barriers will
1205 * succeed, and so the second wait_barrier cannot block.
1207 spin_lock_irq(&conf
->resync_lock
);
1209 spin_unlock_irq(&conf
->resync_lock
);
1211 make_request(mddev
, &bp
->bio1
);
1212 make_request(mddev
, &bp
->bio2
);
1214 spin_lock_irq(&conf
->resync_lock
);
1216 wake_up(&conf
->wait_barrier
);
1217 spin_unlock_irq(&conf
->resync_lock
);
1219 bio_pair_release(bp
);
1222 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1223 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1224 (unsigned long long)bio
->bi_iter
.bi_sector
,
1225 bio_sectors(bio
) / 2);
1231 md_write_start(mddev
, bio
);
1234 * Register the new request and wait if the reconstruction
1235 * thread has put up a bar for new requests.
1236 * Continue immediately if no resync is active currently.
1240 sectors
= bio_sectors(bio
);
1241 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1242 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1243 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1244 /* IO spans the reshape position. Need to wait for
1247 allow_barrier(conf
);
1248 wait_event(conf
->wait_barrier
,
1249 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1250 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1254 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1255 bio_data_dir(bio
) == WRITE
&&
1256 (mddev
->reshape_backwards
1257 ? (bio
->bi_iter
.bi_sector
< conf
->reshape_safe
&&
1258 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
)
1259 : (bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_safe
&&
1260 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
))) {
1261 /* Need to update reshape_position in metadata */
1262 mddev
->reshape_position
= conf
->reshape_progress
;
1263 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1264 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1265 md_wakeup_thread(mddev
->thread
);
1266 wait_event(mddev
->sb_wait
,
1267 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1269 conf
->reshape_safe
= mddev
->reshape_position
;
1272 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1274 r10_bio
->master_bio
= bio
;
1275 r10_bio
->sectors
= sectors
;
1277 r10_bio
->mddev
= mddev
;
1278 r10_bio
->sector
= bio
->bi_iter
.bi_sector
;
1281 /* We might need to issue multiple reads to different
1282 * devices if there are bad blocks around, so we keep
1283 * track of the number of reads in bio->bi_phys_segments.
1284 * If this is 0, there is only one r10_bio and no locking
1285 * will be needed when the request completes. If it is
1286 * non-zero, then it is the number of not-completed requests.
1288 bio
->bi_phys_segments
= 0;
1289 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1293 * read balancing logic:
1295 struct md_rdev
*rdev
;
1299 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1301 raid_end_bio_io(r10_bio
);
1304 slot
= r10_bio
->read_slot
;
1306 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1307 bio_trim(read_bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1310 r10_bio
->devs
[slot
].bio
= read_bio
;
1311 r10_bio
->devs
[slot
].rdev
= rdev
;
1313 read_bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
+
1314 choose_data_offset(r10_bio
, rdev
);
1315 read_bio
->bi_bdev
= rdev
->bdev
;
1316 read_bio
->bi_end_io
= raid10_end_read_request
;
1317 read_bio
->bi_rw
= READ
| do_sync
;
1318 read_bio
->bi_private
= r10_bio
;
1320 if (max_sectors
< r10_bio
->sectors
) {
1321 /* Could not read all from this device, so we will
1322 * need another r10_bio.
1324 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1325 - bio
->bi_iter
.bi_sector
);
1326 r10_bio
->sectors
= max_sectors
;
1327 spin_lock_irq(&conf
->device_lock
);
1328 if (bio
->bi_phys_segments
== 0)
1329 bio
->bi_phys_segments
= 2;
1331 bio
->bi_phys_segments
++;
1332 spin_unlock(&conf
->device_lock
);
1333 /* Cannot call generic_make_request directly
1334 * as that will be queued in __generic_make_request
1335 * and subsequent mempool_alloc might block
1336 * waiting for it. so hand bio over to raid10d.
1338 reschedule_retry(r10_bio
);
1340 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1342 r10_bio
->master_bio
= bio
;
1343 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1345 r10_bio
->mddev
= mddev
;
1346 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+
1350 generic_make_request(read_bio
);
1357 if (conf
->pending_count
>= max_queued_requests
) {
1358 md_wakeup_thread(mddev
->thread
);
1359 wait_event(conf
->wait_barrier
,
1360 conf
->pending_count
< max_queued_requests
);
1362 /* first select target devices under rcu_lock and
1363 * inc refcount on their rdev. Record them by setting
1365 * If there are known/acknowledged bad blocks on any device
1366 * on which we have seen a write error, we want to avoid
1367 * writing to those blocks. This potentially requires several
1368 * writes to write around the bad blocks. Each set of writes
1369 * gets its own r10_bio with a set of bios attached. The number
1370 * of r10_bios is recored in bio->bi_phys_segments just as with
1374 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1375 raid10_find_phys(conf
, r10_bio
);
1377 blocked_rdev
= NULL
;
1379 max_sectors
= r10_bio
->sectors
;
1381 for (i
= 0; i
< conf
->copies
; i
++) {
1382 int d
= r10_bio
->devs
[i
].devnum
;
1383 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1384 struct md_rdev
*rrdev
= rcu_dereference(
1385 conf
->mirrors
[d
].replacement
);
1388 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1389 atomic_inc(&rdev
->nr_pending
);
1390 blocked_rdev
= rdev
;
1393 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1394 atomic_inc(&rrdev
->nr_pending
);
1395 blocked_rdev
= rrdev
;
1398 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)
1399 || test_bit(Unmerged
, &rdev
->flags
)))
1401 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1402 || test_bit(Unmerged
, &rrdev
->flags
)))
1405 r10_bio
->devs
[i
].bio
= NULL
;
1406 r10_bio
->devs
[i
].repl_bio
= NULL
;
1408 if (!rdev
&& !rrdev
) {
1409 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1412 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1414 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1418 is_bad
= is_badblock(rdev
, dev_sector
,
1420 &first_bad
, &bad_sectors
);
1422 /* Mustn't write here until the bad block
1425 atomic_inc(&rdev
->nr_pending
);
1426 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1427 blocked_rdev
= rdev
;
1430 if (is_bad
&& first_bad
<= dev_sector
) {
1431 /* Cannot write here at all */
1432 bad_sectors
-= (dev_sector
- first_bad
);
1433 if (bad_sectors
< max_sectors
)
1434 /* Mustn't write more than bad_sectors
1435 * to other devices yet
1437 max_sectors
= bad_sectors
;
1438 /* We don't set R10BIO_Degraded as that
1439 * only applies if the disk is missing,
1440 * so it might be re-added, and we want to
1441 * know to recover this chunk.
1442 * In this case the device is here, and the
1443 * fact that this chunk is not in-sync is
1444 * recorded in the bad block log.
1449 int good_sectors
= first_bad
- dev_sector
;
1450 if (good_sectors
< max_sectors
)
1451 max_sectors
= good_sectors
;
1455 r10_bio
->devs
[i
].bio
= bio
;
1456 atomic_inc(&rdev
->nr_pending
);
1459 r10_bio
->devs
[i
].repl_bio
= bio
;
1460 atomic_inc(&rrdev
->nr_pending
);
1465 if (unlikely(blocked_rdev
)) {
1466 /* Have to wait for this device to get unblocked, then retry */
1470 for (j
= 0; j
< i
; j
++) {
1471 if (r10_bio
->devs
[j
].bio
) {
1472 d
= r10_bio
->devs
[j
].devnum
;
1473 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1475 if (r10_bio
->devs
[j
].repl_bio
) {
1476 struct md_rdev
*rdev
;
1477 d
= r10_bio
->devs
[j
].devnum
;
1478 rdev
= conf
->mirrors
[d
].replacement
;
1480 /* Race with remove_disk */
1482 rdev
= conf
->mirrors
[d
].rdev
;
1484 rdev_dec_pending(rdev
, mddev
);
1487 allow_barrier(conf
);
1488 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1493 if (max_sectors
< r10_bio
->sectors
) {
1494 /* We are splitting this into multiple parts, so
1495 * we need to prepare for allocating another r10_bio.
1497 r10_bio
->sectors
= max_sectors
;
1498 spin_lock_irq(&conf
->device_lock
);
1499 if (bio
->bi_phys_segments
== 0)
1500 bio
->bi_phys_segments
= 2;
1502 bio
->bi_phys_segments
++;
1503 spin_unlock_irq(&conf
->device_lock
);
1505 sectors_handled
= r10_bio
->sector
+ max_sectors
-
1506 bio
->bi_iter
.bi_sector
;
1508 atomic_set(&r10_bio
->remaining
, 1);
1509 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1511 for (i
= 0; i
< conf
->copies
; i
++) {
1513 int d
= r10_bio
->devs
[i
].devnum
;
1514 if (r10_bio
->devs
[i
].bio
) {
1515 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1516 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1517 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1519 r10_bio
->devs
[i
].bio
= mbio
;
1521 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1522 choose_data_offset(r10_bio
,
1524 mbio
->bi_bdev
= rdev
->bdev
;
1525 mbio
->bi_end_io
= raid10_end_write_request
;
1527 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1528 mbio
->bi_private
= r10_bio
;
1530 atomic_inc(&r10_bio
->remaining
);
1532 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1535 plug
= container_of(cb
, struct raid10_plug_cb
,
1539 spin_lock_irqsave(&conf
->device_lock
, flags
);
1541 bio_list_add(&plug
->pending
, mbio
);
1542 plug
->pending_cnt
++;
1544 bio_list_add(&conf
->pending_bio_list
, mbio
);
1545 conf
->pending_count
++;
1547 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1549 md_wakeup_thread(mddev
->thread
);
1552 if (r10_bio
->devs
[i
].repl_bio
) {
1553 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1555 /* Replacement just got moved to main 'rdev' */
1557 rdev
= conf
->mirrors
[d
].rdev
;
1559 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1560 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1562 r10_bio
->devs
[i
].repl_bio
= mbio
;
1564 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1567 mbio
->bi_bdev
= rdev
->bdev
;
1568 mbio
->bi_end_io
= raid10_end_write_request
;
1570 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1571 mbio
->bi_private
= r10_bio
;
1573 atomic_inc(&r10_bio
->remaining
);
1574 spin_lock_irqsave(&conf
->device_lock
, flags
);
1575 bio_list_add(&conf
->pending_bio_list
, mbio
);
1576 conf
->pending_count
++;
1577 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1578 if (!mddev_check_plugged(mddev
))
1579 md_wakeup_thread(mddev
->thread
);
1583 /* Don't remove the bias on 'remaining' (one_write_done) until
1584 * after checking if we need to go around again.
1587 if (sectors_handled
< bio_sectors(bio
)) {
1588 one_write_done(r10_bio
);
1589 /* We need another r10_bio. It has already been counted
1590 * in bio->bi_phys_segments.
1592 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1594 r10_bio
->master_bio
= bio
;
1595 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1597 r10_bio
->mddev
= mddev
;
1598 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1602 one_write_done(r10_bio
);
1604 /* In case raid10d snuck in to freeze_array */
1605 wake_up(&conf
->wait_barrier
);
1608 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1610 struct r10conf
*conf
= mddev
->private;
1613 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1614 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1615 if (conf
->geo
.near_copies
> 1)
1616 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1617 if (conf
->geo
.far_copies
> 1) {
1618 if (conf
->geo
.far_offset
)
1619 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1621 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1623 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1624 conf
->geo
.raid_disks
- mddev
->degraded
);
1625 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1626 seq_printf(seq
, "%s",
1627 conf
->mirrors
[i
].rdev
&&
1628 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1629 seq_printf(seq
, "]");
1632 /* check if there are enough drives for
1633 * every block to appear on atleast one.
1634 * Don't consider the device numbered 'ignore'
1635 * as we might be about to remove it.
1637 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1643 disks
= conf
->prev
.raid_disks
;
1644 ncopies
= conf
->prev
.near_copies
;
1646 disks
= conf
->geo
.raid_disks
;
1647 ncopies
= conf
->geo
.near_copies
;
1652 int n
= conf
->copies
;
1656 struct md_rdev
*rdev
;
1657 if (this != ignore
&&
1658 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1659 test_bit(In_sync
, &rdev
->flags
))
1661 this = (this+1) % disks
;
1665 first
= (first
+ ncopies
) % disks
;
1666 } while (first
!= 0);
1673 static int enough(struct r10conf
*conf
, int ignore
)
1675 /* when calling 'enough', both 'prev' and 'geo' must
1677 * This is ensured if ->reconfig_mutex or ->device_lock
1680 return _enough(conf
, 0, ignore
) &&
1681 _enough(conf
, 1, ignore
);
1684 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1686 char b
[BDEVNAME_SIZE
];
1687 struct r10conf
*conf
= mddev
->private;
1688 unsigned long flags
;
1691 * If it is not operational, then we have already marked it as dead
1692 * else if it is the last working disks, ignore the error, let the
1693 * next level up know.
1694 * else mark the drive as failed
1696 spin_lock_irqsave(&conf
->device_lock
, flags
);
1697 if (test_bit(In_sync
, &rdev
->flags
)
1698 && !enough(conf
, rdev
->raid_disk
)) {
1700 * Don't fail the drive, just return an IO error.
1702 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1705 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1708 * if recovery is running, make sure it aborts.
1710 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1712 set_bit(Blocked
, &rdev
->flags
);
1713 set_bit(Faulty
, &rdev
->flags
);
1714 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1715 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1717 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1718 "md/raid10:%s: Operation continuing on %d devices.\n",
1719 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1720 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1723 static void print_conf(struct r10conf
*conf
)
1726 struct raid10_info
*tmp
;
1728 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1730 printk(KERN_DEBUG
"(!conf)\n");
1733 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1734 conf
->geo
.raid_disks
);
1736 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1737 char b
[BDEVNAME_SIZE
];
1738 tmp
= conf
->mirrors
+ i
;
1740 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1741 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1742 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1743 bdevname(tmp
->rdev
->bdev
,b
));
1747 static void close_sync(struct r10conf
*conf
)
1750 allow_barrier(conf
);
1752 mempool_destroy(conf
->r10buf_pool
);
1753 conf
->r10buf_pool
= NULL
;
1756 static int raid10_spare_active(struct mddev
*mddev
)
1759 struct r10conf
*conf
= mddev
->private;
1760 struct raid10_info
*tmp
;
1762 unsigned long flags
;
1765 * Find all non-in_sync disks within the RAID10 configuration
1766 * and mark them in_sync
1768 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1769 tmp
= conf
->mirrors
+ i
;
1770 if (tmp
->replacement
1771 && tmp
->replacement
->recovery_offset
== MaxSector
1772 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1773 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1774 /* Replacement has just become active */
1776 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1779 /* Replaced device not technically faulty,
1780 * but we need to be sure it gets removed
1781 * and never re-added.
1783 set_bit(Faulty
, &tmp
->rdev
->flags
);
1784 sysfs_notify_dirent_safe(
1785 tmp
->rdev
->sysfs_state
);
1787 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1788 } else if (tmp
->rdev
1789 && tmp
->rdev
->recovery_offset
== MaxSector
1790 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1791 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1793 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1796 spin_lock_irqsave(&conf
->device_lock
, flags
);
1797 mddev
->degraded
-= count
;
1798 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1805 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1807 struct r10conf
*conf
= mddev
->private;
1811 int last
= conf
->geo
.raid_disks
- 1;
1812 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1814 if (mddev
->recovery_cp
< MaxSector
)
1815 /* only hot-add to in-sync arrays, as recovery is
1816 * very different from resync
1819 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1822 if (rdev
->raid_disk
>= 0)
1823 first
= last
= rdev
->raid_disk
;
1825 if (q
->merge_bvec_fn
) {
1826 set_bit(Unmerged
, &rdev
->flags
);
1827 mddev
->merge_check_needed
= 1;
1830 if (rdev
->saved_raid_disk
>= first
&&
1831 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1832 mirror
= rdev
->saved_raid_disk
;
1835 for ( ; mirror
<= last
; mirror
++) {
1836 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1837 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1840 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1841 p
->replacement
!= NULL
)
1843 clear_bit(In_sync
, &rdev
->flags
);
1844 set_bit(Replacement
, &rdev
->flags
);
1845 rdev
->raid_disk
= mirror
;
1848 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1849 rdev
->data_offset
<< 9);
1851 rcu_assign_pointer(p
->replacement
, rdev
);
1856 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1857 rdev
->data_offset
<< 9);
1859 p
->head_position
= 0;
1860 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1861 rdev
->raid_disk
= mirror
;
1863 if (rdev
->saved_raid_disk
!= mirror
)
1865 rcu_assign_pointer(p
->rdev
, rdev
);
1868 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1869 /* Some requests might not have seen this new
1870 * merge_bvec_fn. We must wait for them to complete
1871 * before merging the device fully.
1872 * First we make sure any code which has tested
1873 * our function has submitted the request, then
1874 * we wait for all outstanding requests to complete.
1876 synchronize_sched();
1877 freeze_array(conf
, 0);
1878 unfreeze_array(conf
);
1879 clear_bit(Unmerged
, &rdev
->flags
);
1881 md_integrity_add_rdev(rdev
, mddev
);
1882 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1883 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1889 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1891 struct r10conf
*conf
= mddev
->private;
1893 int number
= rdev
->raid_disk
;
1894 struct md_rdev
**rdevp
;
1895 struct raid10_info
*p
= conf
->mirrors
+ number
;
1898 if (rdev
== p
->rdev
)
1900 else if (rdev
== p
->replacement
)
1901 rdevp
= &p
->replacement
;
1905 if (test_bit(In_sync
, &rdev
->flags
) ||
1906 atomic_read(&rdev
->nr_pending
)) {
1910 /* Only remove faulty devices if recovery
1913 if (!test_bit(Faulty
, &rdev
->flags
) &&
1914 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1915 (!p
->replacement
|| p
->replacement
== rdev
) &&
1916 number
< conf
->geo
.raid_disks
&&
1923 if (atomic_read(&rdev
->nr_pending
)) {
1924 /* lost the race, try later */
1928 } else if (p
->replacement
) {
1929 /* We must have just cleared 'rdev' */
1930 p
->rdev
= p
->replacement
;
1931 clear_bit(Replacement
, &p
->replacement
->flags
);
1932 smp_mb(); /* Make sure other CPUs may see both as identical
1933 * but will never see neither -- if they are careful.
1935 p
->replacement
= NULL
;
1936 clear_bit(WantReplacement
, &rdev
->flags
);
1938 /* We might have just remove the Replacement as faulty
1939 * Clear the flag just in case
1941 clear_bit(WantReplacement
, &rdev
->flags
);
1943 err
= md_integrity_register(mddev
);
1952 static void end_sync_read(struct bio
*bio
, int error
)
1954 struct r10bio
*r10_bio
= bio
->bi_private
;
1955 struct r10conf
*conf
= r10_bio
->mddev
->private;
1958 if (bio
== r10_bio
->master_bio
) {
1959 /* this is a reshape read */
1960 d
= r10_bio
->read_slot
; /* really the read dev */
1962 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1964 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1965 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1967 /* The write handler will notice the lack of
1968 * R10BIO_Uptodate and record any errors etc
1970 atomic_add(r10_bio
->sectors
,
1971 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1973 /* for reconstruct, we always reschedule after a read.
1974 * for resync, only after all reads
1976 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1977 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1978 atomic_dec_and_test(&r10_bio
->remaining
)) {
1979 /* we have read all the blocks,
1980 * do the comparison in process context in raid10d
1982 reschedule_retry(r10_bio
);
1986 static void end_sync_request(struct r10bio
*r10_bio
)
1988 struct mddev
*mddev
= r10_bio
->mddev
;
1990 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1991 if (r10_bio
->master_bio
== NULL
) {
1992 /* the primary of several recovery bios */
1993 sector_t s
= r10_bio
->sectors
;
1994 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1995 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1996 reschedule_retry(r10_bio
);
1999 md_done_sync(mddev
, s
, 1);
2002 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
2003 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2004 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2005 reschedule_retry(r10_bio
);
2013 static void end_sync_write(struct bio
*bio
, int error
)
2015 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2016 struct r10bio
*r10_bio
= bio
->bi_private
;
2017 struct mddev
*mddev
= r10_bio
->mddev
;
2018 struct r10conf
*conf
= mddev
->private;
2024 struct md_rdev
*rdev
= NULL
;
2026 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
2028 rdev
= conf
->mirrors
[d
].replacement
;
2030 rdev
= conf
->mirrors
[d
].rdev
;
2034 md_error(mddev
, rdev
);
2036 set_bit(WriteErrorSeen
, &rdev
->flags
);
2037 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2038 set_bit(MD_RECOVERY_NEEDED
,
2039 &rdev
->mddev
->recovery
);
2040 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
2042 } else if (is_badblock(rdev
,
2043 r10_bio
->devs
[slot
].addr
,
2045 &first_bad
, &bad_sectors
))
2046 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
2048 rdev_dec_pending(rdev
, mddev
);
2050 end_sync_request(r10_bio
);
2054 * Note: sync and recover and handled very differently for raid10
2055 * This code is for resync.
2056 * For resync, we read through virtual addresses and read all blocks.
2057 * If there is any error, we schedule a write. The lowest numbered
2058 * drive is authoritative.
2059 * However requests come for physical address, so we need to map.
2060 * For every physical address there are raid_disks/copies virtual addresses,
2061 * which is always are least one, but is not necessarly an integer.
2062 * This means that a physical address can span multiple chunks, so we may
2063 * have to submit multiple io requests for a single sync request.
2066 * We check if all blocks are in-sync and only write to blocks that
2069 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2071 struct r10conf
*conf
= mddev
->private;
2073 struct bio
*tbio
, *fbio
;
2076 atomic_set(&r10_bio
->remaining
, 1);
2078 /* find the first device with a block */
2079 for (i
=0; i
<conf
->copies
; i
++)
2080 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
2083 if (i
== conf
->copies
)
2087 fbio
= r10_bio
->devs
[i
].bio
;
2089 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2090 /* now find blocks with errors */
2091 for (i
=0 ; i
< conf
->copies
; i
++) {
2094 tbio
= r10_bio
->devs
[i
].bio
;
2096 if (tbio
->bi_end_io
!= end_sync_read
)
2100 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2101 /* We know that the bi_io_vec layout is the same for
2102 * both 'first' and 'i', so we just compare them.
2103 * All vec entries are PAGE_SIZE;
2105 int sectors
= r10_bio
->sectors
;
2106 for (j
= 0; j
< vcnt
; j
++) {
2107 int len
= PAGE_SIZE
;
2108 if (sectors
< (len
/ 512))
2109 len
= sectors
* 512;
2110 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2111 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2118 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2119 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2120 /* Don't fix anything. */
2123 /* Ok, we need to write this bio, either to correct an
2124 * inconsistency or to correct an unreadable block.
2125 * First we need to fixup bv_offset, bv_len and
2126 * bi_vecs, as the read request might have corrupted these
2130 tbio
->bi_vcnt
= vcnt
;
2131 tbio
->bi_iter
.bi_size
= r10_bio
->sectors
<< 9;
2132 tbio
->bi_rw
= WRITE
;
2133 tbio
->bi_private
= r10_bio
;
2134 tbio
->bi_iter
.bi_sector
= r10_bio
->devs
[i
].addr
;
2136 for (j
=0; j
< vcnt
; j
++) {
2137 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2138 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2140 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2141 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2144 tbio
->bi_end_io
= end_sync_write
;
2146 d
= r10_bio
->devs
[i
].devnum
;
2147 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2148 atomic_inc(&r10_bio
->remaining
);
2149 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2151 tbio
->bi_iter
.bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2152 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2153 generic_make_request(tbio
);
2156 /* Now write out to any replacement devices
2159 for (i
= 0; i
< conf
->copies
; i
++) {
2162 tbio
= r10_bio
->devs
[i
].repl_bio
;
2163 if (!tbio
|| !tbio
->bi_end_io
)
2165 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2166 && r10_bio
->devs
[i
].bio
!= fbio
)
2167 for (j
= 0; j
< vcnt
; j
++)
2168 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2169 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2171 d
= r10_bio
->devs
[i
].devnum
;
2172 atomic_inc(&r10_bio
->remaining
);
2173 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2175 generic_make_request(tbio
);
2179 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2180 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2186 * Now for the recovery code.
2187 * Recovery happens across physical sectors.
2188 * We recover all non-is_sync drives by finding the virtual address of
2189 * each, and then choose a working drive that also has that virt address.
2190 * There is a separate r10_bio for each non-in_sync drive.
2191 * Only the first two slots are in use. The first for reading,
2192 * The second for writing.
2195 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2197 /* We got a read error during recovery.
2198 * We repeat the read in smaller page-sized sections.
2199 * If a read succeeds, write it to the new device or record
2200 * a bad block if we cannot.
2201 * If a read fails, record a bad block on both old and
2204 struct mddev
*mddev
= r10_bio
->mddev
;
2205 struct r10conf
*conf
= mddev
->private;
2206 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2208 int sectors
= r10_bio
->sectors
;
2210 int dr
= r10_bio
->devs
[0].devnum
;
2211 int dw
= r10_bio
->devs
[1].devnum
;
2215 struct md_rdev
*rdev
;
2219 if (s
> (PAGE_SIZE
>>9))
2222 rdev
= conf
->mirrors
[dr
].rdev
;
2223 addr
= r10_bio
->devs
[0].addr
+ sect
,
2224 ok
= sync_page_io(rdev
,
2227 bio
->bi_io_vec
[idx
].bv_page
,
2230 rdev
= conf
->mirrors
[dw
].rdev
;
2231 addr
= r10_bio
->devs
[1].addr
+ sect
;
2232 ok
= sync_page_io(rdev
,
2235 bio
->bi_io_vec
[idx
].bv_page
,
2238 set_bit(WriteErrorSeen
, &rdev
->flags
);
2239 if (!test_and_set_bit(WantReplacement
,
2241 set_bit(MD_RECOVERY_NEEDED
,
2242 &rdev
->mddev
->recovery
);
2246 /* We don't worry if we cannot set a bad block -
2247 * it really is bad so there is no loss in not
2250 rdev_set_badblocks(rdev
, addr
, s
, 0);
2252 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2253 /* need bad block on destination too */
2254 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2255 addr
= r10_bio
->devs
[1].addr
+ sect
;
2256 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2258 /* just abort the recovery */
2260 "md/raid10:%s: recovery aborted"
2261 " due to read error\n",
2264 conf
->mirrors
[dw
].recovery_disabled
2265 = mddev
->recovery_disabled
;
2266 set_bit(MD_RECOVERY_INTR
,
2279 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2281 struct r10conf
*conf
= mddev
->private;
2283 struct bio
*wbio
, *wbio2
;
2285 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2286 fix_recovery_read_error(r10_bio
);
2287 end_sync_request(r10_bio
);
2292 * share the pages with the first bio
2293 * and submit the write request
2295 d
= r10_bio
->devs
[1].devnum
;
2296 wbio
= r10_bio
->devs
[1].bio
;
2297 wbio2
= r10_bio
->devs
[1].repl_bio
;
2298 /* Need to test wbio2->bi_end_io before we call
2299 * generic_make_request as if the former is NULL,
2300 * the latter is free to free wbio2.
2302 if (wbio2
&& !wbio2
->bi_end_io
)
2304 if (wbio
->bi_end_io
) {
2305 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2306 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2307 generic_make_request(wbio
);
2310 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2311 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2312 bio_sectors(wbio2
));
2313 generic_make_request(wbio2
);
2319 * Used by fix_read_error() to decay the per rdev read_errors.
2320 * We halve the read error count for every hour that has elapsed
2321 * since the last recorded read error.
2324 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2326 struct timespec cur_time_mon
;
2327 unsigned long hours_since_last
;
2328 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2330 ktime_get_ts(&cur_time_mon
);
2332 if (rdev
->last_read_error
.tv_sec
== 0 &&
2333 rdev
->last_read_error
.tv_nsec
== 0) {
2334 /* first time we've seen a read error */
2335 rdev
->last_read_error
= cur_time_mon
;
2339 hours_since_last
= (cur_time_mon
.tv_sec
-
2340 rdev
->last_read_error
.tv_sec
) / 3600;
2342 rdev
->last_read_error
= cur_time_mon
;
2345 * if hours_since_last is > the number of bits in read_errors
2346 * just set read errors to 0. We do this to avoid
2347 * overflowing the shift of read_errors by hours_since_last.
2349 if (hours_since_last
>= 8 * sizeof(read_errors
))
2350 atomic_set(&rdev
->read_errors
, 0);
2352 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2355 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2356 int sectors
, struct page
*page
, int rw
)
2361 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2362 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2364 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2368 set_bit(WriteErrorSeen
, &rdev
->flags
);
2369 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2370 set_bit(MD_RECOVERY_NEEDED
,
2371 &rdev
->mddev
->recovery
);
2373 /* need to record an error - either for the block or the device */
2374 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2375 md_error(rdev
->mddev
, rdev
);
2380 * This is a kernel thread which:
2382 * 1. Retries failed read operations on working mirrors.
2383 * 2. Updates the raid superblock when problems encounter.
2384 * 3. Performs writes following reads for array synchronising.
2387 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2389 int sect
= 0; /* Offset from r10_bio->sector */
2390 int sectors
= r10_bio
->sectors
;
2391 struct md_rdev
*rdev
;
2392 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2393 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2395 /* still own a reference to this rdev, so it cannot
2396 * have been cleared recently.
2398 rdev
= conf
->mirrors
[d
].rdev
;
2400 if (test_bit(Faulty
, &rdev
->flags
))
2401 /* drive has already been failed, just ignore any
2402 more fix_read_error() attempts */
2405 check_decay_read_errors(mddev
, rdev
);
2406 atomic_inc(&rdev
->read_errors
);
2407 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2408 char b
[BDEVNAME_SIZE
];
2409 bdevname(rdev
->bdev
, b
);
2412 "md/raid10:%s: %s: Raid device exceeded "
2413 "read_error threshold [cur %d:max %d]\n",
2415 atomic_read(&rdev
->read_errors
), max_read_errors
);
2417 "md/raid10:%s: %s: Failing raid device\n",
2419 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2420 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2426 int sl
= r10_bio
->read_slot
;
2430 if (s
> (PAGE_SIZE
>>9))
2438 d
= r10_bio
->devs
[sl
].devnum
;
2439 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2441 !test_bit(Unmerged
, &rdev
->flags
) &&
2442 test_bit(In_sync
, &rdev
->flags
) &&
2443 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2444 &first_bad
, &bad_sectors
) == 0) {
2445 atomic_inc(&rdev
->nr_pending
);
2447 success
= sync_page_io(rdev
,
2448 r10_bio
->devs
[sl
].addr
+
2451 conf
->tmppage
, READ
, false);
2452 rdev_dec_pending(rdev
, mddev
);
2458 if (sl
== conf
->copies
)
2460 } while (!success
&& sl
!= r10_bio
->read_slot
);
2464 /* Cannot read from anywhere, just mark the block
2465 * as bad on the first device to discourage future
2468 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2469 rdev
= conf
->mirrors
[dn
].rdev
;
2471 if (!rdev_set_badblocks(
2473 r10_bio
->devs
[r10_bio
->read_slot
].addr
2476 md_error(mddev
, rdev
);
2477 r10_bio
->devs
[r10_bio
->read_slot
].bio
2484 /* write it back and re-read */
2486 while (sl
!= r10_bio
->read_slot
) {
2487 char b
[BDEVNAME_SIZE
];
2492 d
= r10_bio
->devs
[sl
].devnum
;
2493 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2495 test_bit(Unmerged
, &rdev
->flags
) ||
2496 !test_bit(In_sync
, &rdev
->flags
))
2499 atomic_inc(&rdev
->nr_pending
);
2501 if (r10_sync_page_io(rdev
,
2502 r10_bio
->devs
[sl
].addr
+
2504 s
, conf
->tmppage
, WRITE
)
2506 /* Well, this device is dead */
2508 "md/raid10:%s: read correction "
2510 " (%d sectors at %llu on %s)\n",
2512 (unsigned long long)(
2514 choose_data_offset(r10_bio
,
2516 bdevname(rdev
->bdev
, b
));
2517 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2520 bdevname(rdev
->bdev
, b
));
2522 rdev_dec_pending(rdev
, mddev
);
2526 while (sl
!= r10_bio
->read_slot
) {
2527 char b
[BDEVNAME_SIZE
];
2532 d
= r10_bio
->devs
[sl
].devnum
;
2533 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2535 !test_bit(In_sync
, &rdev
->flags
))
2538 atomic_inc(&rdev
->nr_pending
);
2540 switch (r10_sync_page_io(rdev
,
2541 r10_bio
->devs
[sl
].addr
+
2546 /* Well, this device is dead */
2548 "md/raid10:%s: unable to read back "
2550 " (%d sectors at %llu on %s)\n",
2552 (unsigned long long)(
2554 choose_data_offset(r10_bio
, rdev
)),
2555 bdevname(rdev
->bdev
, b
));
2556 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2559 bdevname(rdev
->bdev
, b
));
2563 "md/raid10:%s: read error corrected"
2564 " (%d sectors at %llu on %s)\n",
2566 (unsigned long long)(
2568 choose_data_offset(r10_bio
, rdev
)),
2569 bdevname(rdev
->bdev
, b
));
2570 atomic_add(s
, &rdev
->corrected_errors
);
2573 rdev_dec_pending(rdev
, mddev
);
2583 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2585 struct bio
*bio
= r10_bio
->master_bio
;
2586 struct mddev
*mddev
= r10_bio
->mddev
;
2587 struct r10conf
*conf
= mddev
->private;
2588 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2589 /* bio has the data to be written to slot 'i' where
2590 * we just recently had a write error.
2591 * We repeatedly clone the bio and trim down to one block,
2592 * then try the write. Where the write fails we record
2594 * It is conceivable that the bio doesn't exactly align with
2595 * blocks. We must handle this.
2597 * We currently own a reference to the rdev.
2603 int sect_to_write
= r10_bio
->sectors
;
2606 if (rdev
->badblocks
.shift
< 0)
2609 block_sectors
= 1 << rdev
->badblocks
.shift
;
2610 sector
= r10_bio
->sector
;
2611 sectors
= ((r10_bio
->sector
+ block_sectors
)
2612 & ~(sector_t
)(block_sectors
- 1))
2615 while (sect_to_write
) {
2617 if (sectors
> sect_to_write
)
2618 sectors
= sect_to_write
;
2619 /* Write at 'sector' for 'sectors' */
2620 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2621 bio_trim(wbio
, sector
- bio
->bi_iter
.bi_sector
, sectors
);
2622 wbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
2623 choose_data_offset(r10_bio
, rdev
) +
2624 (sector
- r10_bio
->sector
));
2625 wbio
->bi_bdev
= rdev
->bdev
;
2626 if (submit_bio_wait(WRITE
, wbio
) == 0)
2628 ok
= rdev_set_badblocks(rdev
, sector
,
2633 sect_to_write
-= sectors
;
2635 sectors
= block_sectors
;
2640 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2642 int slot
= r10_bio
->read_slot
;
2644 struct r10conf
*conf
= mddev
->private;
2645 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2646 char b
[BDEVNAME_SIZE
];
2647 unsigned long do_sync
;
2650 /* we got a read error. Maybe the drive is bad. Maybe just
2651 * the block and we can fix it.
2652 * We freeze all other IO, and try reading the block from
2653 * other devices. When we find one, we re-write
2654 * and check it that fixes the read error.
2655 * This is all done synchronously while the array is
2658 bio
= r10_bio
->devs
[slot
].bio
;
2659 bdevname(bio
->bi_bdev
, b
);
2661 r10_bio
->devs
[slot
].bio
= NULL
;
2663 if (mddev
->ro
== 0) {
2664 freeze_array(conf
, 1);
2665 fix_read_error(conf
, mddev
, r10_bio
);
2666 unfreeze_array(conf
);
2668 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2670 rdev_dec_pending(rdev
, mddev
);
2673 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2675 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2676 " read error for block %llu\n",
2678 (unsigned long long)r10_bio
->sector
);
2679 raid_end_bio_io(r10_bio
);
2683 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2684 slot
= r10_bio
->read_slot
;
2687 "md/raid10:%s: %s: redirecting "
2688 "sector %llu to another mirror\n",
2690 bdevname(rdev
->bdev
, b
),
2691 (unsigned long long)r10_bio
->sector
);
2692 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2694 bio_trim(bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
2695 r10_bio
->devs
[slot
].bio
= bio
;
2696 r10_bio
->devs
[slot
].rdev
= rdev
;
2697 bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
2698 + choose_data_offset(r10_bio
, rdev
);
2699 bio
->bi_bdev
= rdev
->bdev
;
2700 bio
->bi_rw
= READ
| do_sync
;
2701 bio
->bi_private
= r10_bio
;
2702 bio
->bi_end_io
= raid10_end_read_request
;
2703 if (max_sectors
< r10_bio
->sectors
) {
2704 /* Drat - have to split this up more */
2705 struct bio
*mbio
= r10_bio
->master_bio
;
2706 int sectors_handled
=
2707 r10_bio
->sector
+ max_sectors
2708 - mbio
->bi_iter
.bi_sector
;
2709 r10_bio
->sectors
= max_sectors
;
2710 spin_lock_irq(&conf
->device_lock
);
2711 if (mbio
->bi_phys_segments
== 0)
2712 mbio
->bi_phys_segments
= 2;
2714 mbio
->bi_phys_segments
++;
2715 spin_unlock_irq(&conf
->device_lock
);
2716 generic_make_request(bio
);
2718 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2720 r10_bio
->master_bio
= mbio
;
2721 r10_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2723 set_bit(R10BIO_ReadError
,
2725 r10_bio
->mddev
= mddev
;
2726 r10_bio
->sector
= mbio
->bi_iter
.bi_sector
2731 generic_make_request(bio
);
2734 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2736 /* Some sort of write request has finished and it
2737 * succeeded in writing where we thought there was a
2738 * bad block. So forget the bad block.
2739 * Or possibly if failed and we need to record
2743 struct md_rdev
*rdev
;
2745 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2746 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2747 for (m
= 0; m
< conf
->copies
; m
++) {
2748 int dev
= r10_bio
->devs
[m
].devnum
;
2749 rdev
= conf
->mirrors
[dev
].rdev
;
2750 if (r10_bio
->devs
[m
].bio
== NULL
)
2752 if (test_bit(BIO_UPTODATE
,
2753 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2754 rdev_clear_badblocks(
2756 r10_bio
->devs
[m
].addr
,
2757 r10_bio
->sectors
, 0);
2759 if (!rdev_set_badblocks(
2761 r10_bio
->devs
[m
].addr
,
2762 r10_bio
->sectors
, 0))
2763 md_error(conf
->mddev
, rdev
);
2765 rdev
= conf
->mirrors
[dev
].replacement
;
2766 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2768 if (test_bit(BIO_UPTODATE
,
2769 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2770 rdev_clear_badblocks(
2772 r10_bio
->devs
[m
].addr
,
2773 r10_bio
->sectors
, 0);
2775 if (!rdev_set_badblocks(
2777 r10_bio
->devs
[m
].addr
,
2778 r10_bio
->sectors
, 0))
2779 md_error(conf
->mddev
, rdev
);
2784 for (m
= 0; m
< conf
->copies
; m
++) {
2785 int dev
= r10_bio
->devs
[m
].devnum
;
2786 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2787 rdev
= conf
->mirrors
[dev
].rdev
;
2788 if (bio
== IO_MADE_GOOD
) {
2789 rdev_clear_badblocks(
2791 r10_bio
->devs
[m
].addr
,
2792 r10_bio
->sectors
, 0);
2793 rdev_dec_pending(rdev
, conf
->mddev
);
2794 } else if (bio
!= NULL
&&
2795 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2796 if (!narrow_write_error(r10_bio
, m
)) {
2797 md_error(conf
->mddev
, rdev
);
2798 set_bit(R10BIO_Degraded
,
2801 rdev_dec_pending(rdev
, conf
->mddev
);
2803 bio
= r10_bio
->devs
[m
].repl_bio
;
2804 rdev
= conf
->mirrors
[dev
].replacement
;
2805 if (rdev
&& bio
== IO_MADE_GOOD
) {
2806 rdev_clear_badblocks(
2808 r10_bio
->devs
[m
].addr
,
2809 r10_bio
->sectors
, 0);
2810 rdev_dec_pending(rdev
, conf
->mddev
);
2813 if (test_bit(R10BIO_WriteError
,
2815 close_write(r10_bio
);
2816 raid_end_bio_io(r10_bio
);
2820 static void raid10d(struct md_thread
*thread
)
2822 struct mddev
*mddev
= thread
->mddev
;
2823 struct r10bio
*r10_bio
;
2824 unsigned long flags
;
2825 struct r10conf
*conf
= mddev
->private;
2826 struct list_head
*head
= &conf
->retry_list
;
2827 struct blk_plug plug
;
2829 md_check_recovery(mddev
);
2831 blk_start_plug(&plug
);
2834 flush_pending_writes(conf
);
2836 spin_lock_irqsave(&conf
->device_lock
, flags
);
2837 if (list_empty(head
)) {
2838 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2841 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2842 list_del(head
->prev
);
2844 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2846 mddev
= r10_bio
->mddev
;
2847 conf
= mddev
->private;
2848 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2849 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2850 handle_write_completed(conf
, r10_bio
);
2851 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2852 reshape_request_write(mddev
, r10_bio
);
2853 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2854 sync_request_write(mddev
, r10_bio
);
2855 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2856 recovery_request_write(mddev
, r10_bio
);
2857 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2858 handle_read_error(mddev
, r10_bio
);
2860 /* just a partial read to be scheduled from a
2863 int slot
= r10_bio
->read_slot
;
2864 generic_make_request(r10_bio
->devs
[slot
].bio
);
2868 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2869 md_check_recovery(mddev
);
2871 blk_finish_plug(&plug
);
2875 static int init_resync(struct r10conf
*conf
)
2880 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2881 BUG_ON(conf
->r10buf_pool
);
2882 conf
->have_replacement
= 0;
2883 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2884 if (conf
->mirrors
[i
].replacement
)
2885 conf
->have_replacement
= 1;
2886 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2887 if (!conf
->r10buf_pool
)
2889 conf
->next_resync
= 0;
2894 * perform a "sync" on one "block"
2896 * We need to make sure that no normal I/O request - particularly write
2897 * requests - conflict with active sync requests.
2899 * This is achieved by tracking pending requests and a 'barrier' concept
2900 * that can be installed to exclude normal IO requests.
2902 * Resync and recovery are handled very differently.
2903 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2905 * For resync, we iterate over virtual addresses, read all copies,
2906 * and update if there are differences. If only one copy is live,
2908 * For recovery, we iterate over physical addresses, read a good
2909 * value for each non-in_sync drive, and over-write.
2911 * So, for recovery we may have several outstanding complex requests for a
2912 * given address, one for each out-of-sync device. We model this by allocating
2913 * a number of r10_bio structures, one for each out-of-sync device.
2914 * As we setup these structures, we collect all bio's together into a list
2915 * which we then process collectively to add pages, and then process again
2916 * to pass to generic_make_request.
2918 * The r10_bio structures are linked using a borrowed master_bio pointer.
2919 * This link is counted in ->remaining. When the r10_bio that points to NULL
2920 * has its remaining count decremented to 0, the whole complex operation
2925 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2926 int *skipped
, int go_faster
)
2928 struct r10conf
*conf
= mddev
->private;
2929 struct r10bio
*r10_bio
;
2930 struct bio
*biolist
= NULL
, *bio
;
2931 sector_t max_sector
, nr_sectors
;
2934 sector_t sync_blocks
;
2935 sector_t sectors_skipped
= 0;
2936 int chunks_skipped
= 0;
2937 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2939 if (!conf
->r10buf_pool
)
2940 if (init_resync(conf
))
2944 * Allow skipping a full rebuild for incremental assembly
2945 * of a clean array, like RAID1 does.
2947 if (mddev
->bitmap
== NULL
&&
2948 mddev
->recovery_cp
== MaxSector
&&
2949 mddev
->reshape_position
== MaxSector
&&
2950 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2951 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2952 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2953 conf
->fullsync
== 0) {
2955 return mddev
->dev_sectors
- sector_nr
;
2959 max_sector
= mddev
->dev_sectors
;
2960 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2961 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2962 max_sector
= mddev
->resync_max_sectors
;
2963 if (sector_nr
>= max_sector
) {
2964 /* If we aborted, we need to abort the
2965 * sync on the 'current' bitmap chucks (there can
2966 * be several when recovering multiple devices).
2967 * as we may have started syncing it but not finished.
2968 * We can find the current address in
2969 * mddev->curr_resync, but for recovery,
2970 * we need to convert that to several
2971 * virtual addresses.
2973 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2978 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2979 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2980 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2982 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2984 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2985 bitmap_end_sync(mddev
->bitmap
, sect
,
2989 /* completed sync */
2990 if ((!mddev
->bitmap
|| conf
->fullsync
)
2991 && conf
->have_replacement
2992 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2993 /* Completed a full sync so the replacements
2994 * are now fully recovered.
2996 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2997 if (conf
->mirrors
[i
].replacement
)
2998 conf
->mirrors
[i
].replacement
3004 bitmap_close_sync(mddev
->bitmap
);
3007 return sectors_skipped
;
3010 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
3011 return reshape_request(mddev
, sector_nr
, skipped
);
3013 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
3014 /* if there has been nothing to do on any drive,
3015 * then there is nothing to do at all..
3018 return (max_sector
- sector_nr
) + sectors_skipped
;
3021 if (max_sector
> mddev
->resync_max
)
3022 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
3024 /* make sure whole request will fit in a chunk - if chunks
3027 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
3028 max_sector
> (sector_nr
| chunk_mask
))
3029 max_sector
= (sector_nr
| chunk_mask
) + 1;
3031 * If there is non-resync activity waiting for us then
3032 * put in a delay to throttle resync.
3034 if (!go_faster
&& conf
->nr_waiting
)
3035 msleep_interruptible(1000);
3037 /* Again, very different code for resync and recovery.
3038 * Both must result in an r10bio with a list of bios that
3039 * have bi_end_io, bi_sector, bi_bdev set,
3040 * and bi_private set to the r10bio.
3041 * For recovery, we may actually create several r10bios
3042 * with 2 bios in each, that correspond to the bios in the main one.
3043 * In this case, the subordinate r10bios link back through a
3044 * borrowed master_bio pointer, and the counter in the master
3045 * includes a ref from each subordinate.
3047 /* First, we decide what to do and set ->bi_end_io
3048 * To end_sync_read if we want to read, and
3049 * end_sync_write if we will want to write.
3052 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
3053 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3054 /* recovery... the complicated one */
3058 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
3064 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
3066 if ((mirror
->rdev
== NULL
||
3067 test_bit(In_sync
, &mirror
->rdev
->flags
))
3069 (mirror
->replacement
== NULL
||
3071 &mirror
->replacement
->flags
)))
3075 /* want to reconstruct this device */
3077 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3078 if (sect
>= mddev
->resync_max_sectors
) {
3079 /* last stripe is not complete - don't
3080 * try to recover this sector.
3084 /* Unless we are doing a full sync, or a replacement
3085 * we only need to recover the block if it is set in
3088 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3090 if (sync_blocks
< max_sync
)
3091 max_sync
= sync_blocks
;
3093 mirror
->replacement
== NULL
&&
3095 /* yep, skip the sync_blocks here, but don't assume
3096 * that there will never be anything to do here
3098 chunks_skipped
= -1;
3102 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3103 raise_barrier(conf
, rb2
!= NULL
);
3104 atomic_set(&r10_bio
->remaining
, 0);
3106 r10_bio
->master_bio
= (struct bio
*)rb2
;
3108 atomic_inc(&rb2
->remaining
);
3109 r10_bio
->mddev
= mddev
;
3110 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3111 r10_bio
->sector
= sect
;
3113 raid10_find_phys(conf
, r10_bio
);
3115 /* Need to check if the array will still be
3118 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3119 if (conf
->mirrors
[j
].rdev
== NULL
||
3120 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3125 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3126 &sync_blocks
, still_degraded
);
3129 for (j
=0; j
<conf
->copies
;j
++) {
3131 int d
= r10_bio
->devs
[j
].devnum
;
3132 sector_t from_addr
, to_addr
;
3133 struct md_rdev
*rdev
;
3134 sector_t sector
, first_bad
;
3136 if (!conf
->mirrors
[d
].rdev
||
3137 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3139 /* This is where we read from */
3141 rdev
= conf
->mirrors
[d
].rdev
;
3142 sector
= r10_bio
->devs
[j
].addr
;
3144 if (is_badblock(rdev
, sector
, max_sync
,
3145 &first_bad
, &bad_sectors
)) {
3146 if (first_bad
> sector
)
3147 max_sync
= first_bad
- sector
;
3149 bad_sectors
-= (sector
3151 if (max_sync
> bad_sectors
)
3152 max_sync
= bad_sectors
;
3156 bio
= r10_bio
->devs
[0].bio
;
3158 bio
->bi_next
= biolist
;
3160 bio
->bi_private
= r10_bio
;
3161 bio
->bi_end_io
= end_sync_read
;
3163 from_addr
= r10_bio
->devs
[j
].addr
;
3164 bio
->bi_iter
.bi_sector
= from_addr
+
3166 bio
->bi_bdev
= rdev
->bdev
;
3167 atomic_inc(&rdev
->nr_pending
);
3168 /* and we write to 'i' (if not in_sync) */
3170 for (k
=0; k
<conf
->copies
; k
++)
3171 if (r10_bio
->devs
[k
].devnum
== i
)
3173 BUG_ON(k
== conf
->copies
);
3174 to_addr
= r10_bio
->devs
[k
].addr
;
3175 r10_bio
->devs
[0].devnum
= d
;
3176 r10_bio
->devs
[0].addr
= from_addr
;
3177 r10_bio
->devs
[1].devnum
= i
;
3178 r10_bio
->devs
[1].addr
= to_addr
;
3180 rdev
= mirror
->rdev
;
3181 if (!test_bit(In_sync
, &rdev
->flags
)) {
3182 bio
= r10_bio
->devs
[1].bio
;
3184 bio
->bi_next
= biolist
;
3186 bio
->bi_private
= r10_bio
;
3187 bio
->bi_end_io
= end_sync_write
;
3189 bio
->bi_iter
.bi_sector
= to_addr
3190 + rdev
->data_offset
;
3191 bio
->bi_bdev
= rdev
->bdev
;
3192 atomic_inc(&r10_bio
->remaining
);
3194 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3196 /* and maybe write to replacement */
3197 bio
= r10_bio
->devs
[1].repl_bio
;
3199 bio
->bi_end_io
= NULL
;
3200 rdev
= mirror
->replacement
;
3201 /* Note: if rdev != NULL, then bio
3202 * cannot be NULL as r10buf_pool_alloc will
3203 * have allocated it.
3204 * So the second test here is pointless.
3205 * But it keeps semantic-checkers happy, and
3206 * this comment keeps human reviewers
3209 if (rdev
== NULL
|| bio
== NULL
||
3210 test_bit(Faulty
, &rdev
->flags
))
3213 bio
->bi_next
= biolist
;
3215 bio
->bi_private
= r10_bio
;
3216 bio
->bi_end_io
= end_sync_write
;
3218 bio
->bi_iter
.bi_sector
= to_addr
+
3220 bio
->bi_bdev
= rdev
->bdev
;
3221 atomic_inc(&r10_bio
->remaining
);
3224 if (j
== conf
->copies
) {
3225 /* Cannot recover, so abort the recovery or
3226 * record a bad block */
3229 atomic_dec(&rb2
->remaining
);
3232 /* problem is that there are bad blocks
3233 * on other device(s)
3236 for (k
= 0; k
< conf
->copies
; k
++)
3237 if (r10_bio
->devs
[k
].devnum
== i
)
3239 if (!test_bit(In_sync
,
3240 &mirror
->rdev
->flags
)
3241 && !rdev_set_badblocks(
3243 r10_bio
->devs
[k
].addr
,
3246 if (mirror
->replacement
&&
3247 !rdev_set_badblocks(
3248 mirror
->replacement
,
3249 r10_bio
->devs
[k
].addr
,
3254 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3256 printk(KERN_INFO
"md/raid10:%s: insufficient "
3257 "working devices for recovery.\n",
3259 mirror
->recovery_disabled
3260 = mddev
->recovery_disabled
;
3265 if (biolist
== NULL
) {
3267 struct r10bio
*rb2
= r10_bio
;
3268 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3269 rb2
->master_bio
= NULL
;
3275 /* resync. Schedule a read for every block at this virt offset */
3278 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3280 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3281 &sync_blocks
, mddev
->degraded
) &&
3282 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3283 &mddev
->recovery
)) {
3284 /* We can skip this block */
3286 return sync_blocks
+ sectors_skipped
;
3288 if (sync_blocks
< max_sync
)
3289 max_sync
= sync_blocks
;
3290 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3292 r10_bio
->mddev
= mddev
;
3293 atomic_set(&r10_bio
->remaining
, 0);
3294 raise_barrier(conf
, 0);
3295 conf
->next_resync
= sector_nr
;
3297 r10_bio
->master_bio
= NULL
;
3298 r10_bio
->sector
= sector_nr
;
3299 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3300 raid10_find_phys(conf
, r10_bio
);
3301 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3303 for (i
= 0; i
< conf
->copies
; i
++) {
3304 int d
= r10_bio
->devs
[i
].devnum
;
3305 sector_t first_bad
, sector
;
3308 if (r10_bio
->devs
[i
].repl_bio
)
3309 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3311 bio
= r10_bio
->devs
[i
].bio
;
3313 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3314 if (conf
->mirrors
[d
].rdev
== NULL
||
3315 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3317 sector
= r10_bio
->devs
[i
].addr
;
3318 if (is_badblock(conf
->mirrors
[d
].rdev
,
3320 &first_bad
, &bad_sectors
)) {
3321 if (first_bad
> sector
)
3322 max_sync
= first_bad
- sector
;
3324 bad_sectors
-= (sector
- first_bad
);
3325 if (max_sync
> bad_sectors
)
3326 max_sync
= bad_sectors
;
3330 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3331 atomic_inc(&r10_bio
->remaining
);
3332 bio
->bi_next
= biolist
;
3334 bio
->bi_private
= r10_bio
;
3335 bio
->bi_end_io
= end_sync_read
;
3337 bio
->bi_iter
.bi_sector
= sector
+
3338 conf
->mirrors
[d
].rdev
->data_offset
;
3339 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3342 if (conf
->mirrors
[d
].replacement
== NULL
||
3344 &conf
->mirrors
[d
].replacement
->flags
))
3347 /* Need to set up for writing to the replacement */
3348 bio
= r10_bio
->devs
[i
].repl_bio
;
3350 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3352 sector
= r10_bio
->devs
[i
].addr
;
3353 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3354 bio
->bi_next
= biolist
;
3356 bio
->bi_private
= r10_bio
;
3357 bio
->bi_end_io
= end_sync_write
;
3359 bio
->bi_iter
.bi_sector
= sector
+
3360 conf
->mirrors
[d
].replacement
->data_offset
;
3361 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3366 for (i
=0; i
<conf
->copies
; i
++) {
3367 int d
= r10_bio
->devs
[i
].devnum
;
3368 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3369 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3371 if (r10_bio
->devs
[i
].repl_bio
&&
3372 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3374 conf
->mirrors
[d
].replacement
,
3384 if (sector_nr
+ max_sync
< max_sector
)
3385 max_sector
= sector_nr
+ max_sync
;
3388 int len
= PAGE_SIZE
;
3389 if (sector_nr
+ (len
>>9) > max_sector
)
3390 len
= (max_sector
- sector_nr
) << 9;
3393 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3395 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3396 if (bio_add_page(bio
, page
, len
, 0))
3400 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3401 for (bio2
= biolist
;
3402 bio2
&& bio2
!= bio
;
3403 bio2
= bio2
->bi_next
) {
3404 /* remove last page from this bio */
3406 bio2
->bi_iter
.bi_size
-= len
;
3407 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3411 nr_sectors
+= len
>>9;
3412 sector_nr
+= len
>>9;
3413 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3415 r10_bio
->sectors
= nr_sectors
;
3419 biolist
= biolist
->bi_next
;
3421 bio
->bi_next
= NULL
;
3422 r10_bio
= bio
->bi_private
;
3423 r10_bio
->sectors
= nr_sectors
;
3425 if (bio
->bi_end_io
== end_sync_read
) {
3426 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3427 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3428 generic_make_request(bio
);
3432 if (sectors_skipped
)
3433 /* pretend they weren't skipped, it makes
3434 * no important difference in this case
3436 md_done_sync(mddev
, sectors_skipped
, 1);
3438 return sectors_skipped
+ nr_sectors
;
3440 /* There is nowhere to write, so all non-sync
3441 * drives must be failed or in resync, all drives
3442 * have a bad block, so try the next chunk...
3444 if (sector_nr
+ max_sync
< max_sector
)
3445 max_sector
= sector_nr
+ max_sync
;
3447 sectors_skipped
+= (max_sector
- sector_nr
);
3449 sector_nr
= max_sector
;
3454 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3457 struct r10conf
*conf
= mddev
->private;
3460 raid_disks
= min(conf
->geo
.raid_disks
,
3461 conf
->prev
.raid_disks
);
3463 sectors
= conf
->dev_sectors
;
3465 size
= sectors
>> conf
->geo
.chunk_shift
;
3466 sector_div(size
, conf
->geo
.far_copies
);
3467 size
= size
* raid_disks
;
3468 sector_div(size
, conf
->geo
.near_copies
);
3470 return size
<< conf
->geo
.chunk_shift
;
3473 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3475 /* Calculate the number of sectors-per-device that will
3476 * actually be used, and set conf->dev_sectors and
3480 size
= size
>> conf
->geo
.chunk_shift
;
3481 sector_div(size
, conf
->geo
.far_copies
);
3482 size
= size
* conf
->geo
.raid_disks
;
3483 sector_div(size
, conf
->geo
.near_copies
);
3484 /* 'size' is now the number of chunks in the array */
3485 /* calculate "used chunks per device" */
3486 size
= size
* conf
->copies
;
3488 /* We need to round up when dividing by raid_disks to
3489 * get the stride size.
3491 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3493 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3495 if (conf
->geo
.far_offset
)
3496 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3498 sector_div(size
, conf
->geo
.far_copies
);
3499 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3503 enum geo_type
{geo_new
, geo_old
, geo_start
};
3504 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3507 int layout
, chunk
, disks
;
3510 layout
= mddev
->layout
;
3511 chunk
= mddev
->chunk_sectors
;
3512 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3515 layout
= mddev
->new_layout
;
3516 chunk
= mddev
->new_chunk_sectors
;
3517 disks
= mddev
->raid_disks
;
3519 default: /* avoid 'may be unused' warnings */
3520 case geo_start
: /* new when starting reshape - raid_disks not
3522 layout
= mddev
->new_layout
;
3523 chunk
= mddev
->new_chunk_sectors
;
3524 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3529 if (chunk
< (PAGE_SIZE
>> 9) ||
3530 !is_power_of_2(chunk
))
3533 fc
= (layout
>> 8) & 255;
3534 fo
= layout
& (1<<16);
3535 geo
->raid_disks
= disks
;
3536 geo
->near_copies
= nc
;
3537 geo
->far_copies
= fc
;
3538 geo
->far_offset
= fo
;
3539 geo
->far_set_size
= (layout
& (1<<17)) ? disks
/ fc
: disks
;
3540 geo
->chunk_mask
= chunk
- 1;
3541 geo
->chunk_shift
= ffz(~chunk
);
3545 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3547 struct r10conf
*conf
= NULL
;
3552 copies
= setup_geo(&geo
, mddev
, geo_new
);
3555 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3556 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3557 mdname(mddev
), PAGE_SIZE
);
3561 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3562 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3563 mdname(mddev
), mddev
->new_layout
);
3568 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3572 /* FIXME calc properly */
3573 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3574 max(0,-mddev
->delta_disks
)),
3579 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3584 conf
->copies
= copies
;
3585 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3586 r10bio_pool_free
, conf
);
3587 if (!conf
->r10bio_pool
)
3590 calc_sectors(conf
, mddev
->dev_sectors
);
3591 if (mddev
->reshape_position
== MaxSector
) {
3592 conf
->prev
= conf
->geo
;
3593 conf
->reshape_progress
= MaxSector
;
3595 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3599 conf
->reshape_progress
= mddev
->reshape_position
;
3600 if (conf
->prev
.far_offset
)
3601 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3603 /* far_copies must be 1 */
3604 conf
->prev
.stride
= conf
->dev_sectors
;
3606 spin_lock_init(&conf
->device_lock
);
3607 INIT_LIST_HEAD(&conf
->retry_list
);
3609 spin_lock_init(&conf
->resync_lock
);
3610 init_waitqueue_head(&conf
->wait_barrier
);
3612 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3616 conf
->mddev
= mddev
;
3621 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3624 if (conf
->r10bio_pool
)
3625 mempool_destroy(conf
->r10bio_pool
);
3626 kfree(conf
->mirrors
);
3627 safe_put_page(conf
->tmppage
);
3630 return ERR_PTR(err
);
3633 static int run(struct mddev
*mddev
)
3635 struct r10conf
*conf
;
3636 int i
, disk_idx
, chunk_size
;
3637 struct raid10_info
*disk
;
3638 struct md_rdev
*rdev
;
3640 sector_t min_offset_diff
= 0;
3642 bool discard_supported
= false;
3644 if (mddev
->private == NULL
) {
3645 conf
= setup_conf(mddev
);
3647 return PTR_ERR(conf
);
3648 mddev
->private = conf
;
3650 conf
= mddev
->private;
3654 mddev
->thread
= conf
->thread
;
3655 conf
->thread
= NULL
;
3657 chunk_size
= mddev
->chunk_sectors
<< 9;
3659 blk_queue_max_discard_sectors(mddev
->queue
,
3660 mddev
->chunk_sectors
);
3661 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3662 blk_queue_io_min(mddev
->queue
, chunk_size
);
3663 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3664 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3666 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3667 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3670 rdev_for_each(rdev
, mddev
) {
3672 struct request_queue
*q
;
3674 disk_idx
= rdev
->raid_disk
;
3677 if (disk_idx
>= conf
->geo
.raid_disks
&&
3678 disk_idx
>= conf
->prev
.raid_disks
)
3680 disk
= conf
->mirrors
+ disk_idx
;
3682 if (test_bit(Replacement
, &rdev
->flags
)) {
3683 if (disk
->replacement
)
3685 disk
->replacement
= rdev
;
3691 q
= bdev_get_queue(rdev
->bdev
);
3692 if (q
->merge_bvec_fn
)
3693 mddev
->merge_check_needed
= 1;
3694 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3695 if (!mddev
->reshape_backwards
)
3699 if (first
|| diff
< min_offset_diff
)
3700 min_offset_diff
= diff
;
3703 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3704 rdev
->data_offset
<< 9);
3706 disk
->head_position
= 0;
3708 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3709 discard_supported
= true;
3713 if (discard_supported
)
3714 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3717 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3720 /* need to check that every block has at least one working mirror */
3721 if (!enough(conf
, -1)) {
3722 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3727 if (conf
->reshape_progress
!= MaxSector
) {
3728 /* must ensure that shape change is supported */
3729 if (conf
->geo
.far_copies
!= 1 &&
3730 conf
->geo
.far_offset
== 0)
3732 if (conf
->prev
.far_copies
!= 1 &&
3733 conf
->prev
.far_offset
== 0)
3737 mddev
->degraded
= 0;
3739 i
< conf
->geo
.raid_disks
3740 || i
< conf
->prev
.raid_disks
;
3743 disk
= conf
->mirrors
+ i
;
3745 if (!disk
->rdev
&& disk
->replacement
) {
3746 /* The replacement is all we have - use it */
3747 disk
->rdev
= disk
->replacement
;
3748 disk
->replacement
= NULL
;
3749 clear_bit(Replacement
, &disk
->rdev
->flags
);
3753 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3754 disk
->head_position
= 0;
3759 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3762 if (mddev
->recovery_cp
!= MaxSector
)
3763 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3764 " -- starting background reconstruction\n",
3767 "md/raid10:%s: active with %d out of %d devices\n",
3768 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3769 conf
->geo
.raid_disks
);
3771 * Ok, everything is just fine now
3773 mddev
->dev_sectors
= conf
->dev_sectors
;
3774 size
= raid10_size(mddev
, 0, 0);
3775 md_set_array_sectors(mddev
, size
);
3776 mddev
->resync_max_sectors
= size
;
3779 int stripe
= conf
->geo
.raid_disks
*
3780 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3781 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3782 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3784 /* Calculate max read-ahead size.
3785 * We need to readahead at least twice a whole stripe....
3788 stripe
/= conf
->geo
.near_copies
;
3789 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3790 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3791 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3795 if (md_integrity_register(mddev
))
3798 if (conf
->reshape_progress
!= MaxSector
) {
3799 unsigned long before_length
, after_length
;
3801 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3802 conf
->prev
.far_copies
);
3803 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3804 conf
->geo
.far_copies
);
3806 if (max(before_length
, after_length
) > min_offset_diff
) {
3807 /* This cannot work */
3808 printk("md/raid10: offset difference not enough to continue reshape\n");
3811 conf
->offset_diff
= min_offset_diff
;
3813 conf
->reshape_safe
= conf
->reshape_progress
;
3814 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3815 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3816 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3817 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3818 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3825 md_unregister_thread(&mddev
->thread
);
3826 if (conf
->r10bio_pool
)
3827 mempool_destroy(conf
->r10bio_pool
);
3828 safe_put_page(conf
->tmppage
);
3829 kfree(conf
->mirrors
);
3831 mddev
->private = NULL
;
3836 static int stop(struct mddev
*mddev
)
3838 struct r10conf
*conf
= mddev
->private;
3840 raise_barrier(conf
, 0);
3841 lower_barrier(conf
);
3843 md_unregister_thread(&mddev
->thread
);
3845 /* the unplug fn references 'conf'*/
3846 blk_sync_queue(mddev
->queue
);
3848 if (conf
->r10bio_pool
)
3849 mempool_destroy(conf
->r10bio_pool
);
3850 safe_put_page(conf
->tmppage
);
3851 kfree(conf
->mirrors
);
3853 mddev
->private = NULL
;
3857 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3859 struct r10conf
*conf
= mddev
->private;
3863 raise_barrier(conf
, 0);
3866 lower_barrier(conf
);
3871 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3873 /* Resize of 'far' arrays is not supported.
3874 * For 'near' and 'offset' arrays we can set the
3875 * number of sectors used to be an appropriate multiple
3876 * of the chunk size.
3877 * For 'offset', this is far_copies*chunksize.
3878 * For 'near' the multiplier is the LCM of
3879 * near_copies and raid_disks.
3880 * So if far_copies > 1 && !far_offset, fail.
3881 * Else find LCM(raid_disks, near_copy)*far_copies and
3882 * multiply by chunk_size. Then round to this number.
3883 * This is mostly done by raid10_size()
3885 struct r10conf
*conf
= mddev
->private;
3886 sector_t oldsize
, size
;
3888 if (mddev
->reshape_position
!= MaxSector
)
3891 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3894 oldsize
= raid10_size(mddev
, 0, 0);
3895 size
= raid10_size(mddev
, sectors
, 0);
3896 if (mddev
->external_size
&&
3897 mddev
->array_sectors
> size
)
3899 if (mddev
->bitmap
) {
3900 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3904 md_set_array_sectors(mddev
, size
);
3905 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3906 revalidate_disk(mddev
->gendisk
);
3907 if (sectors
> mddev
->dev_sectors
&&
3908 mddev
->recovery_cp
> oldsize
) {
3909 mddev
->recovery_cp
= oldsize
;
3910 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3912 calc_sectors(conf
, sectors
);
3913 mddev
->dev_sectors
= conf
->dev_sectors
;
3914 mddev
->resync_max_sectors
= size
;
3918 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3920 struct md_rdev
*rdev
;
3921 struct r10conf
*conf
;
3923 if (mddev
->degraded
> 0) {
3924 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3926 return ERR_PTR(-EINVAL
);
3929 /* Set new parameters */
3930 mddev
->new_level
= 10;
3931 /* new layout: far_copies = 1, near_copies = 2 */
3932 mddev
->new_layout
= (1<<8) + 2;
3933 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3934 mddev
->delta_disks
= mddev
->raid_disks
;
3935 mddev
->raid_disks
*= 2;
3936 /* make sure it will be not marked as dirty */
3937 mddev
->recovery_cp
= MaxSector
;
3939 conf
= setup_conf(mddev
);
3940 if (!IS_ERR(conf
)) {
3941 rdev_for_each(rdev
, mddev
)
3942 if (rdev
->raid_disk
>= 0)
3943 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3950 static void *raid10_takeover(struct mddev
*mddev
)
3952 struct r0conf
*raid0_conf
;
3954 /* raid10 can take over:
3955 * raid0 - providing it has only two drives
3957 if (mddev
->level
== 0) {
3958 /* for raid0 takeover only one zone is supported */
3959 raid0_conf
= mddev
->private;
3960 if (raid0_conf
->nr_strip_zones
> 1) {
3961 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3962 " with more than one zone.\n",
3964 return ERR_PTR(-EINVAL
);
3966 return raid10_takeover_raid0(mddev
);
3968 return ERR_PTR(-EINVAL
);
3971 static int raid10_check_reshape(struct mddev
*mddev
)
3973 /* Called when there is a request to change
3974 * - layout (to ->new_layout)
3975 * - chunk size (to ->new_chunk_sectors)
3976 * - raid_disks (by delta_disks)
3977 * or when trying to restart a reshape that was ongoing.
3979 * We need to validate the request and possibly allocate
3980 * space if that might be an issue later.
3982 * Currently we reject any reshape of a 'far' mode array,
3983 * allow chunk size to change if new is generally acceptable,
3984 * allow raid_disks to increase, and allow
3985 * a switch between 'near' mode and 'offset' mode.
3987 struct r10conf
*conf
= mddev
->private;
3990 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3993 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3994 /* mustn't change number of copies */
3996 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3997 /* Cannot switch to 'far' mode */
4000 if (mddev
->array_sectors
& geo
.chunk_mask
)
4001 /* not factor of array size */
4004 if (!enough(conf
, -1))
4007 kfree(conf
->mirrors_new
);
4008 conf
->mirrors_new
= NULL
;
4009 if (mddev
->delta_disks
> 0) {
4010 /* allocate new 'mirrors' list */
4011 conf
->mirrors_new
= kzalloc(
4012 sizeof(struct raid10_info
)
4013 *(mddev
->raid_disks
+
4014 mddev
->delta_disks
),
4016 if (!conf
->mirrors_new
)
4023 * Need to check if array has failed when deciding whether to:
4025 * - remove non-faulty devices
4028 * This determination is simple when no reshape is happening.
4029 * However if there is a reshape, we need to carefully check
4030 * both the before and after sections.
4031 * This is because some failed devices may only affect one
4032 * of the two sections, and some non-in_sync devices may
4033 * be insync in the section most affected by failed devices.
4035 static int calc_degraded(struct r10conf
*conf
)
4037 int degraded
, degraded2
;
4042 /* 'prev' section first */
4043 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4044 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4045 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4047 else if (!test_bit(In_sync
, &rdev
->flags
))
4048 /* When we can reduce the number of devices in
4049 * an array, this might not contribute to
4050 * 'degraded'. It does now.
4055 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4059 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4060 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4061 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4063 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4064 /* If reshape is increasing the number of devices,
4065 * this section has already been recovered, so
4066 * it doesn't contribute to degraded.
4069 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4074 if (degraded2
> degraded
)
4079 static int raid10_start_reshape(struct mddev
*mddev
)
4081 /* A 'reshape' has been requested. This commits
4082 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4083 * This also checks if there are enough spares and adds them
4085 * We currently require enough spares to make the final
4086 * array non-degraded. We also require that the difference
4087 * between old and new data_offset - on each device - is
4088 * enough that we never risk over-writing.
4091 unsigned long before_length
, after_length
;
4092 sector_t min_offset_diff
= 0;
4095 struct r10conf
*conf
= mddev
->private;
4096 struct md_rdev
*rdev
;
4100 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4103 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4106 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4107 conf
->prev
.far_copies
);
4108 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4109 conf
->geo
.far_copies
);
4111 rdev_for_each(rdev
, mddev
) {
4112 if (!test_bit(In_sync
, &rdev
->flags
)
4113 && !test_bit(Faulty
, &rdev
->flags
))
4115 if (rdev
->raid_disk
>= 0) {
4116 long long diff
= (rdev
->new_data_offset
4117 - rdev
->data_offset
);
4118 if (!mddev
->reshape_backwards
)
4122 if (first
|| diff
< min_offset_diff
)
4123 min_offset_diff
= diff
;
4127 if (max(before_length
, after_length
) > min_offset_diff
)
4130 if (spares
< mddev
->delta_disks
)
4133 conf
->offset_diff
= min_offset_diff
;
4134 spin_lock_irq(&conf
->device_lock
);
4135 if (conf
->mirrors_new
) {
4136 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4137 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4139 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
4140 conf
->mirrors_old
= conf
->mirrors
;
4141 conf
->mirrors
= conf
->mirrors_new
;
4142 conf
->mirrors_new
= NULL
;
4144 setup_geo(&conf
->geo
, mddev
, geo_start
);
4146 if (mddev
->reshape_backwards
) {
4147 sector_t size
= raid10_size(mddev
, 0, 0);
4148 if (size
< mddev
->array_sectors
) {
4149 spin_unlock_irq(&conf
->device_lock
);
4150 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4154 mddev
->resync_max_sectors
= size
;
4155 conf
->reshape_progress
= size
;
4157 conf
->reshape_progress
= 0;
4158 spin_unlock_irq(&conf
->device_lock
);
4160 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4161 ret
= bitmap_resize(mddev
->bitmap
,
4162 raid10_size(mddev
, 0,
4163 conf
->geo
.raid_disks
),
4168 if (mddev
->delta_disks
> 0) {
4169 rdev_for_each(rdev
, mddev
)
4170 if (rdev
->raid_disk
< 0 &&
4171 !test_bit(Faulty
, &rdev
->flags
)) {
4172 if (raid10_add_disk(mddev
, rdev
) == 0) {
4173 if (rdev
->raid_disk
>=
4174 conf
->prev
.raid_disks
)
4175 set_bit(In_sync
, &rdev
->flags
);
4177 rdev
->recovery_offset
= 0;
4179 if (sysfs_link_rdev(mddev
, rdev
))
4180 /* Failure here is OK */;
4182 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4183 && !test_bit(Faulty
, &rdev
->flags
)) {
4184 /* This is a spare that was manually added */
4185 set_bit(In_sync
, &rdev
->flags
);
4188 /* When a reshape changes the number of devices,
4189 * ->degraded is measured against the larger of the
4190 * pre and post numbers.
4192 spin_lock_irq(&conf
->device_lock
);
4193 mddev
->degraded
= calc_degraded(conf
);
4194 spin_unlock_irq(&conf
->device_lock
);
4195 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4196 mddev
->reshape_position
= conf
->reshape_progress
;
4197 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4199 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4200 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4201 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4202 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4204 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4206 if (!mddev
->sync_thread
) {
4210 conf
->reshape_checkpoint
= jiffies
;
4211 md_wakeup_thread(mddev
->sync_thread
);
4212 md_new_event(mddev
);
4216 mddev
->recovery
= 0;
4217 spin_lock_irq(&conf
->device_lock
);
4218 conf
->geo
= conf
->prev
;
4219 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4220 rdev_for_each(rdev
, mddev
)
4221 rdev
->new_data_offset
= rdev
->data_offset
;
4223 conf
->reshape_progress
= MaxSector
;
4224 mddev
->reshape_position
= MaxSector
;
4225 spin_unlock_irq(&conf
->device_lock
);
4229 /* Calculate the last device-address that could contain
4230 * any block from the chunk that includes the array-address 's'
4231 * and report the next address.
4232 * i.e. the address returned will be chunk-aligned and after
4233 * any data that is in the chunk containing 's'.
4235 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4237 s
= (s
| geo
->chunk_mask
) + 1;
4238 s
>>= geo
->chunk_shift
;
4239 s
*= geo
->near_copies
;
4240 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4241 s
*= geo
->far_copies
;
4242 s
<<= geo
->chunk_shift
;
4246 /* Calculate the first device-address that could contain
4247 * any block from the chunk that includes the array-address 's'.
4248 * This too will be the start of a chunk
4250 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4252 s
>>= geo
->chunk_shift
;
4253 s
*= geo
->near_copies
;
4254 sector_div(s
, geo
->raid_disks
);
4255 s
*= geo
->far_copies
;
4256 s
<<= geo
->chunk_shift
;
4260 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4263 /* We simply copy at most one chunk (smallest of old and new)
4264 * at a time, possibly less if that exceeds RESYNC_PAGES,
4265 * or we hit a bad block or something.
4266 * This might mean we pause for normal IO in the middle of
4267 * a chunk, but that is not a problem was mddev->reshape_position
4268 * can record any location.
4270 * If we will want to write to a location that isn't
4271 * yet recorded as 'safe' (i.e. in metadata on disk) then
4272 * we need to flush all reshape requests and update the metadata.
4274 * When reshaping forwards (e.g. to more devices), we interpret
4275 * 'safe' as the earliest block which might not have been copied
4276 * down yet. We divide this by previous stripe size and multiply
4277 * by previous stripe length to get lowest device offset that we
4278 * cannot write to yet.
4279 * We interpret 'sector_nr' as an address that we want to write to.
4280 * From this we use last_device_address() to find where we might
4281 * write to, and first_device_address on the 'safe' position.
4282 * If this 'next' write position is after the 'safe' position,
4283 * we must update the metadata to increase the 'safe' position.
4285 * When reshaping backwards, we round in the opposite direction
4286 * and perform the reverse test: next write position must not be
4287 * less than current safe position.
4289 * In all this the minimum difference in data offsets
4290 * (conf->offset_diff - always positive) allows a bit of slack,
4291 * so next can be after 'safe', but not by more than offset_disk
4293 * We need to prepare all the bios here before we start any IO
4294 * to ensure the size we choose is acceptable to all devices.
4295 * The means one for each copy for write-out and an extra one for
4297 * We store the read-in bio in ->master_bio and the others in
4298 * ->devs[x].bio and ->devs[x].repl_bio.
4300 struct r10conf
*conf
= mddev
->private;
4301 struct r10bio
*r10_bio
;
4302 sector_t next
, safe
, last
;
4306 struct md_rdev
*rdev
;
4309 struct bio
*bio
, *read_bio
;
4310 int sectors_done
= 0;
4312 if (sector_nr
== 0) {
4313 /* If restarting in the middle, skip the initial sectors */
4314 if (mddev
->reshape_backwards
&&
4315 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4316 sector_nr
= (raid10_size(mddev
, 0, 0)
4317 - conf
->reshape_progress
);
4318 } else if (!mddev
->reshape_backwards
&&
4319 conf
->reshape_progress
> 0)
4320 sector_nr
= conf
->reshape_progress
;
4322 mddev
->curr_resync_completed
= sector_nr
;
4323 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4329 /* We don't use sector_nr to track where we are up to
4330 * as that doesn't work well for ->reshape_backwards.
4331 * So just use ->reshape_progress.
4333 if (mddev
->reshape_backwards
) {
4334 /* 'next' is the earliest device address that we might
4335 * write to for this chunk in the new layout
4337 next
= first_dev_address(conf
->reshape_progress
- 1,
4340 /* 'safe' is the last device address that we might read from
4341 * in the old layout after a restart
4343 safe
= last_dev_address(conf
->reshape_safe
- 1,
4346 if (next
+ conf
->offset_diff
< safe
)
4349 last
= conf
->reshape_progress
- 1;
4350 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4351 & conf
->prev
.chunk_mask
);
4352 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4353 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4355 /* 'next' is after the last device address that we
4356 * might write to for this chunk in the new layout
4358 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4360 /* 'safe' is the earliest device address that we might
4361 * read from in the old layout after a restart
4363 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4365 /* Need to update metadata if 'next' might be beyond 'safe'
4366 * as that would possibly corrupt data
4368 if (next
> safe
+ conf
->offset_diff
)
4371 sector_nr
= conf
->reshape_progress
;
4372 last
= sector_nr
| (conf
->geo
.chunk_mask
4373 & conf
->prev
.chunk_mask
);
4375 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4376 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4380 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4381 /* Need to update reshape_position in metadata */
4383 mddev
->reshape_position
= conf
->reshape_progress
;
4384 if (mddev
->reshape_backwards
)
4385 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4386 - conf
->reshape_progress
;
4388 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4389 conf
->reshape_checkpoint
= jiffies
;
4390 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4391 md_wakeup_thread(mddev
->thread
);
4392 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4393 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4394 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4395 allow_barrier(conf
);
4396 return sectors_done
;
4398 conf
->reshape_safe
= mddev
->reshape_position
;
4399 allow_barrier(conf
);
4403 /* Now schedule reads for blocks from sector_nr to last */
4404 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4405 raise_barrier(conf
, sectors_done
!= 0);
4406 atomic_set(&r10_bio
->remaining
, 0);
4407 r10_bio
->mddev
= mddev
;
4408 r10_bio
->sector
= sector_nr
;
4409 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4410 r10_bio
->sectors
= last
- sector_nr
+ 1;
4411 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4412 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4415 /* Cannot read from here, so need to record bad blocks
4416 * on all the target devices.
4419 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4420 return sectors_done
;
4423 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4425 read_bio
->bi_bdev
= rdev
->bdev
;
4426 read_bio
->bi_iter
.bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4427 + rdev
->data_offset
);
4428 read_bio
->bi_private
= r10_bio
;
4429 read_bio
->bi_end_io
= end_sync_read
;
4430 read_bio
->bi_rw
= READ
;
4431 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4432 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4433 read_bio
->bi_vcnt
= 0;
4434 read_bio
->bi_iter
.bi_size
= 0;
4435 r10_bio
->master_bio
= read_bio
;
4436 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4438 /* Now find the locations in the new layout */
4439 __raid10_find_phys(&conf
->geo
, r10_bio
);
4442 read_bio
->bi_next
= NULL
;
4444 for (s
= 0; s
< conf
->copies
*2; s
++) {
4446 int d
= r10_bio
->devs
[s
/2].devnum
;
4447 struct md_rdev
*rdev2
;
4449 rdev2
= conf
->mirrors
[d
].replacement
;
4450 b
= r10_bio
->devs
[s
/2].repl_bio
;
4452 rdev2
= conf
->mirrors
[d
].rdev
;
4453 b
= r10_bio
->devs
[s
/2].bio
;
4455 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4459 b
->bi_bdev
= rdev2
->bdev
;
4460 b
->bi_iter
.bi_sector
= r10_bio
->devs
[s
/2].addr
+
4461 rdev2
->new_data_offset
;
4462 b
->bi_private
= r10_bio
;
4463 b
->bi_end_io
= end_reshape_write
;
4469 /* Now add as many pages as possible to all of these bios. */
4472 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4473 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4474 int len
= (max_sectors
- s
) << 9;
4475 if (len
> PAGE_SIZE
)
4477 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4479 if (bio_add_page(bio
, page
, len
, 0))
4482 /* Didn't fit, must stop */
4484 bio2
&& bio2
!= bio
;
4485 bio2
= bio2
->bi_next
) {
4486 /* Remove last page from this bio */
4488 bio2
->bi_iter
.bi_size
-= len
;
4489 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4493 sector_nr
+= len
>> 9;
4494 nr_sectors
+= len
>> 9;
4497 r10_bio
->sectors
= nr_sectors
;
4499 /* Now submit the read */
4500 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4501 atomic_inc(&r10_bio
->remaining
);
4502 read_bio
->bi_next
= NULL
;
4503 generic_make_request(read_bio
);
4504 sector_nr
+= nr_sectors
;
4505 sectors_done
+= nr_sectors
;
4506 if (sector_nr
<= last
)
4509 /* Now that we have done the whole section we can
4510 * update reshape_progress
4512 if (mddev
->reshape_backwards
)
4513 conf
->reshape_progress
-= sectors_done
;
4515 conf
->reshape_progress
+= sectors_done
;
4517 return sectors_done
;
4520 static void end_reshape_request(struct r10bio
*r10_bio
);
4521 static int handle_reshape_read_error(struct mddev
*mddev
,
4522 struct r10bio
*r10_bio
);
4523 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4525 /* Reshape read completed. Hopefully we have a block
4527 * If we got a read error then we do sync 1-page reads from
4528 * elsewhere until we find the data - or give up.
4530 struct r10conf
*conf
= mddev
->private;
4533 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4534 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4535 /* Reshape has been aborted */
4536 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4540 /* We definitely have the data in the pages, schedule the
4543 atomic_set(&r10_bio
->remaining
, 1);
4544 for (s
= 0; s
< conf
->copies
*2; s
++) {
4546 int d
= r10_bio
->devs
[s
/2].devnum
;
4547 struct md_rdev
*rdev
;
4549 rdev
= conf
->mirrors
[d
].replacement
;
4550 b
= r10_bio
->devs
[s
/2].repl_bio
;
4552 rdev
= conf
->mirrors
[d
].rdev
;
4553 b
= r10_bio
->devs
[s
/2].bio
;
4555 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4557 atomic_inc(&rdev
->nr_pending
);
4558 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4559 atomic_inc(&r10_bio
->remaining
);
4561 generic_make_request(b
);
4563 end_reshape_request(r10_bio
);
4566 static void end_reshape(struct r10conf
*conf
)
4568 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4571 spin_lock_irq(&conf
->device_lock
);
4572 conf
->prev
= conf
->geo
;
4573 md_finish_reshape(conf
->mddev
);
4575 conf
->reshape_progress
= MaxSector
;
4576 spin_unlock_irq(&conf
->device_lock
);
4578 /* read-ahead size must cover two whole stripes, which is
4579 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4581 if (conf
->mddev
->queue
) {
4582 int stripe
= conf
->geo
.raid_disks
*
4583 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4584 stripe
/= conf
->geo
.near_copies
;
4585 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4586 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4592 static int handle_reshape_read_error(struct mddev
*mddev
,
4593 struct r10bio
*r10_bio
)
4595 /* Use sync reads to get the blocks from somewhere else */
4596 int sectors
= r10_bio
->sectors
;
4597 struct r10conf
*conf
= mddev
->private;
4599 struct r10bio r10_bio
;
4600 struct r10dev devs
[conf
->copies
];
4602 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4605 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4607 r10b
->sector
= r10_bio
->sector
;
4608 __raid10_find_phys(&conf
->prev
, r10b
);
4613 int first_slot
= slot
;
4615 if (s
> (PAGE_SIZE
>> 9))
4619 int d
= r10b
->devs
[slot
].devnum
;
4620 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4623 test_bit(Faulty
, &rdev
->flags
) ||
4624 !test_bit(In_sync
, &rdev
->flags
))
4627 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4628 success
= sync_page_io(rdev
,
4637 if (slot
>= conf
->copies
)
4639 if (slot
== first_slot
)
4643 /* couldn't read this block, must give up */
4644 set_bit(MD_RECOVERY_INTR
,
4654 static void end_reshape_write(struct bio
*bio
, int error
)
4656 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4657 struct r10bio
*r10_bio
= bio
->bi_private
;
4658 struct mddev
*mddev
= r10_bio
->mddev
;
4659 struct r10conf
*conf
= mddev
->private;
4663 struct md_rdev
*rdev
= NULL
;
4665 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4667 rdev
= conf
->mirrors
[d
].replacement
;
4670 rdev
= conf
->mirrors
[d
].rdev
;
4674 /* FIXME should record badblock */
4675 md_error(mddev
, rdev
);
4678 rdev_dec_pending(rdev
, mddev
);
4679 end_reshape_request(r10_bio
);
4682 static void end_reshape_request(struct r10bio
*r10_bio
)
4684 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4686 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4687 bio_put(r10_bio
->master_bio
);
4691 static void raid10_finish_reshape(struct mddev
*mddev
)
4693 struct r10conf
*conf
= mddev
->private;
4695 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4698 if (mddev
->delta_disks
> 0) {
4699 sector_t size
= raid10_size(mddev
, 0, 0);
4700 md_set_array_sectors(mddev
, size
);
4701 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4702 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4703 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4705 mddev
->resync_max_sectors
= size
;
4706 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4707 revalidate_disk(mddev
->gendisk
);
4710 for (d
= conf
->geo
.raid_disks
;
4711 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4713 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4715 clear_bit(In_sync
, &rdev
->flags
);
4716 rdev
= conf
->mirrors
[d
].replacement
;
4718 clear_bit(In_sync
, &rdev
->flags
);
4721 mddev
->layout
= mddev
->new_layout
;
4722 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4723 mddev
->reshape_position
= MaxSector
;
4724 mddev
->delta_disks
= 0;
4725 mddev
->reshape_backwards
= 0;
4728 static struct md_personality raid10_personality
=
4732 .owner
= THIS_MODULE
,
4733 .make_request
= make_request
,
4737 .error_handler
= error
,
4738 .hot_add_disk
= raid10_add_disk
,
4739 .hot_remove_disk
= raid10_remove_disk
,
4740 .spare_active
= raid10_spare_active
,
4741 .sync_request
= sync_request
,
4742 .quiesce
= raid10_quiesce
,
4743 .size
= raid10_size
,
4744 .resize
= raid10_resize
,
4745 .takeover
= raid10_takeover
,
4746 .check_reshape
= raid10_check_reshape
,
4747 .start_reshape
= raid10_start_reshape
,
4748 .finish_reshape
= raid10_finish_reshape
,
4751 static int __init
raid_init(void)
4753 return register_md_personality(&raid10_personality
);
4756 static void raid_exit(void)
4758 unregister_md_personality(&raid10_personality
);
4761 module_init(raid_init
);
4762 module_exit(raid_exit
);
4763 MODULE_LICENSE("GPL");
4764 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4765 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4766 MODULE_ALIAS("md-raid10");
4767 MODULE_ALIAS("md-level-10");
4769 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);