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 const int rw
= bio_data_dir(bio
);
1162 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1163 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1164 const unsigned long do_discard
= (bio
->bi_rw
1165 & (REQ_DISCARD
| REQ_SECURE
));
1166 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1167 unsigned long flags
;
1168 struct md_rdev
*blocked_rdev
;
1169 struct blk_plug_cb
*cb
;
1170 struct raid10_plug_cb
*plug
= NULL
;
1171 int sectors_handled
;
1176 * Register the new request and wait if the reconstruction
1177 * thread has put up a bar for new requests.
1178 * Continue immediately if no resync is active currently.
1182 sectors
= bio_sectors(bio
);
1183 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1184 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1185 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1186 /* IO spans the reshape position. Need to wait for
1189 allow_barrier(conf
);
1190 wait_event(conf
->wait_barrier
,
1191 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1192 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1196 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1197 bio_data_dir(bio
) == WRITE
&&
1198 (mddev
->reshape_backwards
1199 ? (bio
->bi_iter
.bi_sector
< conf
->reshape_safe
&&
1200 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
)
1201 : (bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_safe
&&
1202 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
))) {
1203 /* Need to update reshape_position in metadata */
1204 mddev
->reshape_position
= conf
->reshape_progress
;
1205 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1206 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1207 md_wakeup_thread(mddev
->thread
);
1208 wait_event(mddev
->sb_wait
,
1209 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1211 conf
->reshape_safe
= mddev
->reshape_position
;
1214 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1216 r10_bio
->master_bio
= bio
;
1217 r10_bio
->sectors
= sectors
;
1219 r10_bio
->mddev
= mddev
;
1220 r10_bio
->sector
= bio
->bi_iter
.bi_sector
;
1223 /* We might need to issue multiple reads to different
1224 * devices if there are bad blocks around, so we keep
1225 * track of the number of reads in bio->bi_phys_segments.
1226 * If this is 0, there is only one r10_bio and no locking
1227 * will be needed when the request completes. If it is
1228 * non-zero, then it is the number of not-completed requests.
1230 bio
->bi_phys_segments
= 0;
1231 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1235 * read balancing logic:
1237 struct md_rdev
*rdev
;
1241 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1243 raid_end_bio_io(r10_bio
);
1246 slot
= r10_bio
->read_slot
;
1248 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1249 bio_trim(read_bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1252 r10_bio
->devs
[slot
].bio
= read_bio
;
1253 r10_bio
->devs
[slot
].rdev
= rdev
;
1255 read_bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
+
1256 choose_data_offset(r10_bio
, rdev
);
1257 read_bio
->bi_bdev
= rdev
->bdev
;
1258 read_bio
->bi_end_io
= raid10_end_read_request
;
1259 read_bio
->bi_rw
= READ
| do_sync
;
1260 read_bio
->bi_private
= r10_bio
;
1262 if (max_sectors
< r10_bio
->sectors
) {
1263 /* Could not read all from this device, so we will
1264 * need another r10_bio.
1266 sectors_handled
= (r10_bio
->sector
+ max_sectors
1267 - bio
->bi_iter
.bi_sector
);
1268 r10_bio
->sectors
= max_sectors
;
1269 spin_lock_irq(&conf
->device_lock
);
1270 if (bio
->bi_phys_segments
== 0)
1271 bio
->bi_phys_segments
= 2;
1273 bio
->bi_phys_segments
++;
1274 spin_unlock_irq(&conf
->device_lock
);
1275 /* Cannot call generic_make_request directly
1276 * as that will be queued in __generic_make_request
1277 * and subsequent mempool_alloc might block
1278 * waiting for it. so hand bio over to raid10d.
1280 reschedule_retry(r10_bio
);
1282 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1284 r10_bio
->master_bio
= bio
;
1285 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1287 r10_bio
->mddev
= mddev
;
1288 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+
1292 generic_make_request(read_bio
);
1299 if (conf
->pending_count
>= max_queued_requests
) {
1300 md_wakeup_thread(mddev
->thread
);
1301 wait_event(conf
->wait_barrier
,
1302 conf
->pending_count
< max_queued_requests
);
1304 /* first select target devices under rcu_lock and
1305 * inc refcount on their rdev. Record them by setting
1307 * If there are known/acknowledged bad blocks on any device
1308 * on which we have seen a write error, we want to avoid
1309 * writing to those blocks. This potentially requires several
1310 * writes to write around the bad blocks. Each set of writes
1311 * gets its own r10_bio with a set of bios attached. The number
1312 * of r10_bios is recored in bio->bi_phys_segments just as with
1316 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1317 raid10_find_phys(conf
, r10_bio
);
1319 blocked_rdev
= NULL
;
1321 max_sectors
= r10_bio
->sectors
;
1323 for (i
= 0; i
< conf
->copies
; i
++) {
1324 int d
= r10_bio
->devs
[i
].devnum
;
1325 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1326 struct md_rdev
*rrdev
= rcu_dereference(
1327 conf
->mirrors
[d
].replacement
);
1330 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1331 atomic_inc(&rdev
->nr_pending
);
1332 blocked_rdev
= rdev
;
1335 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1336 atomic_inc(&rrdev
->nr_pending
);
1337 blocked_rdev
= rrdev
;
1340 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)
1341 || test_bit(Unmerged
, &rdev
->flags
)))
1343 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1344 || test_bit(Unmerged
, &rrdev
->flags
)))
1347 r10_bio
->devs
[i
].bio
= NULL
;
1348 r10_bio
->devs
[i
].repl_bio
= NULL
;
1350 if (!rdev
&& !rrdev
) {
1351 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1354 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1356 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1360 is_bad
= is_badblock(rdev
, dev_sector
,
1362 &first_bad
, &bad_sectors
);
1364 /* Mustn't write here until the bad block
1367 atomic_inc(&rdev
->nr_pending
);
1368 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1369 blocked_rdev
= rdev
;
1372 if (is_bad
&& first_bad
<= dev_sector
) {
1373 /* Cannot write here at all */
1374 bad_sectors
-= (dev_sector
- first_bad
);
1375 if (bad_sectors
< max_sectors
)
1376 /* Mustn't write more than bad_sectors
1377 * to other devices yet
1379 max_sectors
= bad_sectors
;
1380 /* We don't set R10BIO_Degraded as that
1381 * only applies if the disk is missing,
1382 * so it might be re-added, and we want to
1383 * know to recover this chunk.
1384 * In this case the device is here, and the
1385 * fact that this chunk is not in-sync is
1386 * recorded in the bad block log.
1391 int good_sectors
= first_bad
- dev_sector
;
1392 if (good_sectors
< max_sectors
)
1393 max_sectors
= good_sectors
;
1397 r10_bio
->devs
[i
].bio
= bio
;
1398 atomic_inc(&rdev
->nr_pending
);
1401 r10_bio
->devs
[i
].repl_bio
= bio
;
1402 atomic_inc(&rrdev
->nr_pending
);
1407 if (unlikely(blocked_rdev
)) {
1408 /* Have to wait for this device to get unblocked, then retry */
1412 for (j
= 0; j
< i
; j
++) {
1413 if (r10_bio
->devs
[j
].bio
) {
1414 d
= r10_bio
->devs
[j
].devnum
;
1415 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1417 if (r10_bio
->devs
[j
].repl_bio
) {
1418 struct md_rdev
*rdev
;
1419 d
= r10_bio
->devs
[j
].devnum
;
1420 rdev
= conf
->mirrors
[d
].replacement
;
1422 /* Race with remove_disk */
1424 rdev
= conf
->mirrors
[d
].rdev
;
1426 rdev_dec_pending(rdev
, mddev
);
1429 allow_barrier(conf
);
1430 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1435 if (max_sectors
< r10_bio
->sectors
) {
1436 /* We are splitting this into multiple parts, so
1437 * we need to prepare for allocating another r10_bio.
1439 r10_bio
->sectors
= max_sectors
;
1440 spin_lock_irq(&conf
->device_lock
);
1441 if (bio
->bi_phys_segments
== 0)
1442 bio
->bi_phys_segments
= 2;
1444 bio
->bi_phys_segments
++;
1445 spin_unlock_irq(&conf
->device_lock
);
1447 sectors_handled
= r10_bio
->sector
+ max_sectors
-
1448 bio
->bi_iter
.bi_sector
;
1450 atomic_set(&r10_bio
->remaining
, 1);
1451 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1453 for (i
= 0; i
< conf
->copies
; i
++) {
1455 int d
= r10_bio
->devs
[i
].devnum
;
1456 if (r10_bio
->devs
[i
].bio
) {
1457 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1458 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1459 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1461 r10_bio
->devs
[i
].bio
= mbio
;
1463 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1464 choose_data_offset(r10_bio
,
1466 mbio
->bi_bdev
= rdev
->bdev
;
1467 mbio
->bi_end_io
= raid10_end_write_request
;
1469 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1470 mbio
->bi_private
= r10_bio
;
1472 atomic_inc(&r10_bio
->remaining
);
1474 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1477 plug
= container_of(cb
, struct raid10_plug_cb
,
1481 spin_lock_irqsave(&conf
->device_lock
, flags
);
1483 bio_list_add(&plug
->pending
, mbio
);
1484 plug
->pending_cnt
++;
1486 bio_list_add(&conf
->pending_bio_list
, mbio
);
1487 conf
->pending_count
++;
1489 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1491 md_wakeup_thread(mddev
->thread
);
1494 if (r10_bio
->devs
[i
].repl_bio
) {
1495 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1497 /* Replacement just got moved to main 'rdev' */
1499 rdev
= conf
->mirrors
[d
].rdev
;
1501 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1502 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1504 r10_bio
->devs
[i
].repl_bio
= mbio
;
1506 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1509 mbio
->bi_bdev
= rdev
->bdev
;
1510 mbio
->bi_end_io
= raid10_end_write_request
;
1512 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1513 mbio
->bi_private
= r10_bio
;
1515 atomic_inc(&r10_bio
->remaining
);
1516 spin_lock_irqsave(&conf
->device_lock
, flags
);
1517 bio_list_add(&conf
->pending_bio_list
, mbio
);
1518 conf
->pending_count
++;
1519 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1520 if (!mddev_check_plugged(mddev
))
1521 md_wakeup_thread(mddev
->thread
);
1525 /* Don't remove the bias on 'remaining' (one_write_done) until
1526 * after checking if we need to go around again.
1529 if (sectors_handled
< bio_sectors(bio
)) {
1530 one_write_done(r10_bio
);
1531 /* We need another r10_bio. It has already been counted
1532 * in bio->bi_phys_segments.
1534 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1536 r10_bio
->master_bio
= bio
;
1537 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1539 r10_bio
->mddev
= mddev
;
1540 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1544 one_write_done(r10_bio
);
1547 static void make_request(struct mddev
*mddev
, struct bio
*bio
)
1549 struct r10conf
*conf
= mddev
->private;
1550 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1551 int chunk_sects
= chunk_mask
+ 1;
1555 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1556 md_flush_request(mddev
, bio
);
1560 md_write_start(mddev
, bio
);
1566 * If this request crosses a chunk boundary, we need to split
1569 if (unlikely((bio
->bi_iter
.bi_sector
& chunk_mask
) +
1570 bio_sectors(bio
) > chunk_sects
1571 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1572 || conf
->prev
.near_copies
<
1573 conf
->prev
.raid_disks
))) {
1574 split
= bio_split(bio
, chunk_sects
-
1575 (bio
->bi_iter
.bi_sector
&
1577 GFP_NOIO
, fs_bio_set
);
1578 bio_chain(split
, bio
);
1583 __make_request(mddev
, split
);
1584 } while (split
!= bio
);
1586 /* In case raid10d snuck in to freeze_array */
1587 wake_up(&conf
->wait_barrier
);
1590 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1592 struct r10conf
*conf
= mddev
->private;
1595 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1596 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1597 if (conf
->geo
.near_copies
> 1)
1598 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1599 if (conf
->geo
.far_copies
> 1) {
1600 if (conf
->geo
.far_offset
)
1601 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1603 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1605 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1606 conf
->geo
.raid_disks
- mddev
->degraded
);
1607 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1608 seq_printf(seq
, "%s",
1609 conf
->mirrors
[i
].rdev
&&
1610 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1611 seq_printf(seq
, "]");
1614 /* check if there are enough drives for
1615 * every block to appear on atleast one.
1616 * Don't consider the device numbered 'ignore'
1617 * as we might be about to remove it.
1619 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1625 disks
= conf
->prev
.raid_disks
;
1626 ncopies
= conf
->prev
.near_copies
;
1628 disks
= conf
->geo
.raid_disks
;
1629 ncopies
= conf
->geo
.near_copies
;
1634 int n
= conf
->copies
;
1638 struct md_rdev
*rdev
;
1639 if (this != ignore
&&
1640 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1641 test_bit(In_sync
, &rdev
->flags
))
1643 this = (this+1) % disks
;
1647 first
= (first
+ ncopies
) % disks
;
1648 } while (first
!= 0);
1655 static int enough(struct r10conf
*conf
, int ignore
)
1657 /* when calling 'enough', both 'prev' and 'geo' must
1659 * This is ensured if ->reconfig_mutex or ->device_lock
1662 return _enough(conf
, 0, ignore
) &&
1663 _enough(conf
, 1, ignore
);
1666 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1668 char b
[BDEVNAME_SIZE
];
1669 struct r10conf
*conf
= mddev
->private;
1670 unsigned long flags
;
1673 * If it is not operational, then we have already marked it as dead
1674 * else if it is the last working disks, ignore the error, let the
1675 * next level up know.
1676 * else mark the drive as failed
1678 spin_lock_irqsave(&conf
->device_lock
, flags
);
1679 if (test_bit(In_sync
, &rdev
->flags
)
1680 && !enough(conf
, rdev
->raid_disk
)) {
1682 * Don't fail the drive, just return an IO error.
1684 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1687 if (test_and_clear_bit(In_sync
, &rdev
->flags
))
1690 * If recovery is running, make sure it aborts.
1692 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1693 set_bit(Blocked
, &rdev
->flags
);
1694 set_bit(Faulty
, &rdev
->flags
);
1695 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1696 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1698 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1699 "md/raid10:%s: Operation continuing on %d devices.\n",
1700 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1701 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1704 static void print_conf(struct r10conf
*conf
)
1707 struct raid10_info
*tmp
;
1709 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1711 printk(KERN_DEBUG
"(!conf)\n");
1714 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1715 conf
->geo
.raid_disks
);
1717 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1718 char b
[BDEVNAME_SIZE
];
1719 tmp
= conf
->mirrors
+ i
;
1721 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1722 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1723 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1724 bdevname(tmp
->rdev
->bdev
,b
));
1728 static void close_sync(struct r10conf
*conf
)
1731 allow_barrier(conf
);
1733 mempool_destroy(conf
->r10buf_pool
);
1734 conf
->r10buf_pool
= NULL
;
1737 static int raid10_spare_active(struct mddev
*mddev
)
1740 struct r10conf
*conf
= mddev
->private;
1741 struct raid10_info
*tmp
;
1743 unsigned long flags
;
1746 * Find all non-in_sync disks within the RAID10 configuration
1747 * and mark them in_sync
1749 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1750 tmp
= conf
->mirrors
+ i
;
1751 if (tmp
->replacement
1752 && tmp
->replacement
->recovery_offset
== MaxSector
1753 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1754 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1755 /* Replacement has just become active */
1757 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1760 /* Replaced device not technically faulty,
1761 * but we need to be sure it gets removed
1762 * and never re-added.
1764 set_bit(Faulty
, &tmp
->rdev
->flags
);
1765 sysfs_notify_dirent_safe(
1766 tmp
->rdev
->sysfs_state
);
1768 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1769 } else if (tmp
->rdev
1770 && tmp
->rdev
->recovery_offset
== MaxSector
1771 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1772 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1774 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1777 spin_lock_irqsave(&conf
->device_lock
, flags
);
1778 mddev
->degraded
-= count
;
1779 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1786 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1788 struct r10conf
*conf
= mddev
->private;
1792 int last
= conf
->geo
.raid_disks
- 1;
1793 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1795 if (mddev
->recovery_cp
< MaxSector
)
1796 /* only hot-add to in-sync arrays, as recovery is
1797 * very different from resync
1800 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1803 if (rdev
->raid_disk
>= 0)
1804 first
= last
= rdev
->raid_disk
;
1806 if (q
->merge_bvec_fn
) {
1807 set_bit(Unmerged
, &rdev
->flags
);
1808 mddev
->merge_check_needed
= 1;
1811 if (rdev
->saved_raid_disk
>= first
&&
1812 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1813 mirror
= rdev
->saved_raid_disk
;
1816 for ( ; mirror
<= last
; mirror
++) {
1817 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1818 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1821 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1822 p
->replacement
!= NULL
)
1824 clear_bit(In_sync
, &rdev
->flags
);
1825 set_bit(Replacement
, &rdev
->flags
);
1826 rdev
->raid_disk
= mirror
;
1829 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1830 rdev
->data_offset
<< 9);
1832 rcu_assign_pointer(p
->replacement
, rdev
);
1837 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1838 rdev
->data_offset
<< 9);
1840 p
->head_position
= 0;
1841 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1842 rdev
->raid_disk
= mirror
;
1844 if (rdev
->saved_raid_disk
!= mirror
)
1846 rcu_assign_pointer(p
->rdev
, rdev
);
1849 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1850 /* Some requests might not have seen this new
1851 * merge_bvec_fn. We must wait for them to complete
1852 * before merging the device fully.
1853 * First we make sure any code which has tested
1854 * our function has submitted the request, then
1855 * we wait for all outstanding requests to complete.
1857 synchronize_sched();
1858 freeze_array(conf
, 0);
1859 unfreeze_array(conf
);
1860 clear_bit(Unmerged
, &rdev
->flags
);
1862 md_integrity_add_rdev(rdev
, mddev
);
1863 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1864 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1870 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1872 struct r10conf
*conf
= mddev
->private;
1874 int number
= rdev
->raid_disk
;
1875 struct md_rdev
**rdevp
;
1876 struct raid10_info
*p
= conf
->mirrors
+ number
;
1879 if (rdev
== p
->rdev
)
1881 else if (rdev
== p
->replacement
)
1882 rdevp
= &p
->replacement
;
1886 if (test_bit(In_sync
, &rdev
->flags
) ||
1887 atomic_read(&rdev
->nr_pending
)) {
1891 /* Only remove faulty devices if recovery
1894 if (!test_bit(Faulty
, &rdev
->flags
) &&
1895 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1896 (!p
->replacement
|| p
->replacement
== rdev
) &&
1897 number
< conf
->geo
.raid_disks
&&
1904 if (atomic_read(&rdev
->nr_pending
)) {
1905 /* lost the race, try later */
1909 } else if (p
->replacement
) {
1910 /* We must have just cleared 'rdev' */
1911 p
->rdev
= p
->replacement
;
1912 clear_bit(Replacement
, &p
->replacement
->flags
);
1913 smp_mb(); /* Make sure other CPUs may see both as identical
1914 * but will never see neither -- if they are careful.
1916 p
->replacement
= NULL
;
1917 clear_bit(WantReplacement
, &rdev
->flags
);
1919 /* We might have just remove the Replacement as faulty
1920 * Clear the flag just in case
1922 clear_bit(WantReplacement
, &rdev
->flags
);
1924 err
= md_integrity_register(mddev
);
1933 static void end_sync_read(struct bio
*bio
, int error
)
1935 struct r10bio
*r10_bio
= bio
->bi_private
;
1936 struct r10conf
*conf
= r10_bio
->mddev
->private;
1939 if (bio
== r10_bio
->master_bio
) {
1940 /* this is a reshape read */
1941 d
= r10_bio
->read_slot
; /* really the read dev */
1943 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1945 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1946 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1948 /* The write handler will notice the lack of
1949 * R10BIO_Uptodate and record any errors etc
1951 atomic_add(r10_bio
->sectors
,
1952 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1954 /* for reconstruct, we always reschedule after a read.
1955 * for resync, only after all reads
1957 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1958 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1959 atomic_dec_and_test(&r10_bio
->remaining
)) {
1960 /* we have read all the blocks,
1961 * do the comparison in process context in raid10d
1963 reschedule_retry(r10_bio
);
1967 static void end_sync_request(struct r10bio
*r10_bio
)
1969 struct mddev
*mddev
= r10_bio
->mddev
;
1971 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1972 if (r10_bio
->master_bio
== NULL
) {
1973 /* the primary of several recovery bios */
1974 sector_t s
= r10_bio
->sectors
;
1975 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1976 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1977 reschedule_retry(r10_bio
);
1980 md_done_sync(mddev
, s
, 1);
1983 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1984 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1985 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1986 reschedule_retry(r10_bio
);
1994 static void end_sync_write(struct bio
*bio
, int error
)
1996 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1997 struct r10bio
*r10_bio
= bio
->bi_private
;
1998 struct mddev
*mddev
= r10_bio
->mddev
;
1999 struct r10conf
*conf
= mddev
->private;
2005 struct md_rdev
*rdev
= NULL
;
2007 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
2009 rdev
= conf
->mirrors
[d
].replacement
;
2011 rdev
= conf
->mirrors
[d
].rdev
;
2015 md_error(mddev
, rdev
);
2017 set_bit(WriteErrorSeen
, &rdev
->flags
);
2018 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2019 set_bit(MD_RECOVERY_NEEDED
,
2020 &rdev
->mddev
->recovery
);
2021 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
2023 } else if (is_badblock(rdev
,
2024 r10_bio
->devs
[slot
].addr
,
2026 &first_bad
, &bad_sectors
))
2027 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
2029 rdev_dec_pending(rdev
, mddev
);
2031 end_sync_request(r10_bio
);
2035 * Note: sync and recover and handled very differently for raid10
2036 * This code is for resync.
2037 * For resync, we read through virtual addresses and read all blocks.
2038 * If there is any error, we schedule a write. The lowest numbered
2039 * drive is authoritative.
2040 * However requests come for physical address, so we need to map.
2041 * For every physical address there are raid_disks/copies virtual addresses,
2042 * which is always are least one, but is not necessarly an integer.
2043 * This means that a physical address can span multiple chunks, so we may
2044 * have to submit multiple io requests for a single sync request.
2047 * We check if all blocks are in-sync and only write to blocks that
2050 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2052 struct r10conf
*conf
= mddev
->private;
2054 struct bio
*tbio
, *fbio
;
2057 atomic_set(&r10_bio
->remaining
, 1);
2059 /* find the first device with a block */
2060 for (i
=0; i
<conf
->copies
; i
++)
2061 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
2064 if (i
== conf
->copies
)
2068 fbio
= r10_bio
->devs
[i
].bio
;
2070 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2071 /* now find blocks with errors */
2072 for (i
=0 ; i
< conf
->copies
; i
++) {
2075 tbio
= r10_bio
->devs
[i
].bio
;
2077 if (tbio
->bi_end_io
!= end_sync_read
)
2081 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2082 /* We know that the bi_io_vec layout is the same for
2083 * both 'first' and 'i', so we just compare them.
2084 * All vec entries are PAGE_SIZE;
2086 int sectors
= r10_bio
->sectors
;
2087 for (j
= 0; j
< vcnt
; j
++) {
2088 int len
= PAGE_SIZE
;
2089 if (sectors
< (len
/ 512))
2090 len
= sectors
* 512;
2091 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2092 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2099 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2100 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2101 /* Don't fix anything. */
2104 /* Ok, we need to write this bio, either to correct an
2105 * inconsistency or to correct an unreadable block.
2106 * First we need to fixup bv_offset, bv_len and
2107 * bi_vecs, as the read request might have corrupted these
2111 tbio
->bi_vcnt
= vcnt
;
2112 tbio
->bi_iter
.bi_size
= r10_bio
->sectors
<< 9;
2113 tbio
->bi_rw
= WRITE
;
2114 tbio
->bi_private
= r10_bio
;
2115 tbio
->bi_iter
.bi_sector
= r10_bio
->devs
[i
].addr
;
2117 for (j
=0; j
< vcnt
; j
++) {
2118 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2119 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2121 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2122 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2125 tbio
->bi_end_io
= end_sync_write
;
2127 d
= r10_bio
->devs
[i
].devnum
;
2128 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2129 atomic_inc(&r10_bio
->remaining
);
2130 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2132 tbio
->bi_iter
.bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2133 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2134 generic_make_request(tbio
);
2137 /* Now write out to any replacement devices
2140 for (i
= 0; i
< conf
->copies
; i
++) {
2143 tbio
= r10_bio
->devs
[i
].repl_bio
;
2144 if (!tbio
|| !tbio
->bi_end_io
)
2146 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2147 && r10_bio
->devs
[i
].bio
!= fbio
)
2148 for (j
= 0; j
< vcnt
; j
++)
2149 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2150 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2152 d
= r10_bio
->devs
[i
].devnum
;
2153 atomic_inc(&r10_bio
->remaining
);
2154 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2156 generic_make_request(tbio
);
2160 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2161 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2167 * Now for the recovery code.
2168 * Recovery happens across physical sectors.
2169 * We recover all non-is_sync drives by finding the virtual address of
2170 * each, and then choose a working drive that also has that virt address.
2171 * There is a separate r10_bio for each non-in_sync drive.
2172 * Only the first two slots are in use. The first for reading,
2173 * The second for writing.
2176 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2178 /* We got a read error during recovery.
2179 * We repeat the read in smaller page-sized sections.
2180 * If a read succeeds, write it to the new device or record
2181 * a bad block if we cannot.
2182 * If a read fails, record a bad block on both old and
2185 struct mddev
*mddev
= r10_bio
->mddev
;
2186 struct r10conf
*conf
= mddev
->private;
2187 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2189 int sectors
= r10_bio
->sectors
;
2191 int dr
= r10_bio
->devs
[0].devnum
;
2192 int dw
= r10_bio
->devs
[1].devnum
;
2196 struct md_rdev
*rdev
;
2200 if (s
> (PAGE_SIZE
>>9))
2203 rdev
= conf
->mirrors
[dr
].rdev
;
2204 addr
= r10_bio
->devs
[0].addr
+ sect
,
2205 ok
= sync_page_io(rdev
,
2208 bio
->bi_io_vec
[idx
].bv_page
,
2211 rdev
= conf
->mirrors
[dw
].rdev
;
2212 addr
= r10_bio
->devs
[1].addr
+ sect
;
2213 ok
= sync_page_io(rdev
,
2216 bio
->bi_io_vec
[idx
].bv_page
,
2219 set_bit(WriteErrorSeen
, &rdev
->flags
);
2220 if (!test_and_set_bit(WantReplacement
,
2222 set_bit(MD_RECOVERY_NEEDED
,
2223 &rdev
->mddev
->recovery
);
2227 /* We don't worry if we cannot set a bad block -
2228 * it really is bad so there is no loss in not
2231 rdev_set_badblocks(rdev
, addr
, s
, 0);
2233 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2234 /* need bad block on destination too */
2235 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2236 addr
= r10_bio
->devs
[1].addr
+ sect
;
2237 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2239 /* just abort the recovery */
2241 "md/raid10:%s: recovery aborted"
2242 " due to read error\n",
2245 conf
->mirrors
[dw
].recovery_disabled
2246 = mddev
->recovery_disabled
;
2247 set_bit(MD_RECOVERY_INTR
,
2260 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2262 struct r10conf
*conf
= mddev
->private;
2264 struct bio
*wbio
, *wbio2
;
2266 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2267 fix_recovery_read_error(r10_bio
);
2268 end_sync_request(r10_bio
);
2273 * share the pages with the first bio
2274 * and submit the write request
2276 d
= r10_bio
->devs
[1].devnum
;
2277 wbio
= r10_bio
->devs
[1].bio
;
2278 wbio2
= r10_bio
->devs
[1].repl_bio
;
2279 /* Need to test wbio2->bi_end_io before we call
2280 * generic_make_request as if the former is NULL,
2281 * the latter is free to free wbio2.
2283 if (wbio2
&& !wbio2
->bi_end_io
)
2285 if (wbio
->bi_end_io
) {
2286 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2287 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2288 generic_make_request(wbio
);
2291 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2292 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2293 bio_sectors(wbio2
));
2294 generic_make_request(wbio2
);
2300 * Used by fix_read_error() to decay the per rdev read_errors.
2301 * We halve the read error count for every hour that has elapsed
2302 * since the last recorded read error.
2305 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2307 struct timespec cur_time_mon
;
2308 unsigned long hours_since_last
;
2309 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2311 ktime_get_ts(&cur_time_mon
);
2313 if (rdev
->last_read_error
.tv_sec
== 0 &&
2314 rdev
->last_read_error
.tv_nsec
== 0) {
2315 /* first time we've seen a read error */
2316 rdev
->last_read_error
= cur_time_mon
;
2320 hours_since_last
= (cur_time_mon
.tv_sec
-
2321 rdev
->last_read_error
.tv_sec
) / 3600;
2323 rdev
->last_read_error
= cur_time_mon
;
2326 * if hours_since_last is > the number of bits in read_errors
2327 * just set read errors to 0. We do this to avoid
2328 * overflowing the shift of read_errors by hours_since_last.
2330 if (hours_since_last
>= 8 * sizeof(read_errors
))
2331 atomic_set(&rdev
->read_errors
, 0);
2333 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2336 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2337 int sectors
, struct page
*page
, int rw
)
2342 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2343 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2345 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2349 set_bit(WriteErrorSeen
, &rdev
->flags
);
2350 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2351 set_bit(MD_RECOVERY_NEEDED
,
2352 &rdev
->mddev
->recovery
);
2354 /* need to record an error - either for the block or the device */
2355 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2356 md_error(rdev
->mddev
, rdev
);
2361 * This is a kernel thread which:
2363 * 1. Retries failed read operations on working mirrors.
2364 * 2. Updates the raid superblock when problems encounter.
2365 * 3. Performs writes following reads for array synchronising.
2368 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2370 int sect
= 0; /* Offset from r10_bio->sector */
2371 int sectors
= r10_bio
->sectors
;
2372 struct md_rdev
*rdev
;
2373 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2374 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2376 /* still own a reference to this rdev, so it cannot
2377 * have been cleared recently.
2379 rdev
= conf
->mirrors
[d
].rdev
;
2381 if (test_bit(Faulty
, &rdev
->flags
))
2382 /* drive has already been failed, just ignore any
2383 more fix_read_error() attempts */
2386 check_decay_read_errors(mddev
, rdev
);
2387 atomic_inc(&rdev
->read_errors
);
2388 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2389 char b
[BDEVNAME_SIZE
];
2390 bdevname(rdev
->bdev
, b
);
2393 "md/raid10:%s: %s: Raid device exceeded "
2394 "read_error threshold [cur %d:max %d]\n",
2396 atomic_read(&rdev
->read_errors
), max_read_errors
);
2398 "md/raid10:%s: %s: Failing raid device\n",
2400 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2401 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2407 int sl
= r10_bio
->read_slot
;
2411 if (s
> (PAGE_SIZE
>>9))
2419 d
= r10_bio
->devs
[sl
].devnum
;
2420 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2422 !test_bit(Unmerged
, &rdev
->flags
) &&
2423 test_bit(In_sync
, &rdev
->flags
) &&
2424 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2425 &first_bad
, &bad_sectors
) == 0) {
2426 atomic_inc(&rdev
->nr_pending
);
2428 success
= sync_page_io(rdev
,
2429 r10_bio
->devs
[sl
].addr
+
2432 conf
->tmppage
, READ
, false);
2433 rdev_dec_pending(rdev
, mddev
);
2439 if (sl
== conf
->copies
)
2441 } while (!success
&& sl
!= r10_bio
->read_slot
);
2445 /* Cannot read from anywhere, just mark the block
2446 * as bad on the first device to discourage future
2449 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2450 rdev
= conf
->mirrors
[dn
].rdev
;
2452 if (!rdev_set_badblocks(
2454 r10_bio
->devs
[r10_bio
->read_slot
].addr
2457 md_error(mddev
, rdev
);
2458 r10_bio
->devs
[r10_bio
->read_slot
].bio
2465 /* write it back and re-read */
2467 while (sl
!= r10_bio
->read_slot
) {
2468 char b
[BDEVNAME_SIZE
];
2473 d
= r10_bio
->devs
[sl
].devnum
;
2474 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2476 test_bit(Unmerged
, &rdev
->flags
) ||
2477 !test_bit(In_sync
, &rdev
->flags
))
2480 atomic_inc(&rdev
->nr_pending
);
2482 if (r10_sync_page_io(rdev
,
2483 r10_bio
->devs
[sl
].addr
+
2485 s
, conf
->tmppage
, WRITE
)
2487 /* Well, this device is dead */
2489 "md/raid10:%s: read correction "
2491 " (%d sectors at %llu on %s)\n",
2493 (unsigned long long)(
2495 choose_data_offset(r10_bio
,
2497 bdevname(rdev
->bdev
, b
));
2498 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2501 bdevname(rdev
->bdev
, b
));
2503 rdev_dec_pending(rdev
, mddev
);
2507 while (sl
!= r10_bio
->read_slot
) {
2508 char b
[BDEVNAME_SIZE
];
2513 d
= r10_bio
->devs
[sl
].devnum
;
2514 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2516 !test_bit(In_sync
, &rdev
->flags
))
2519 atomic_inc(&rdev
->nr_pending
);
2521 switch (r10_sync_page_io(rdev
,
2522 r10_bio
->devs
[sl
].addr
+
2527 /* Well, this device is dead */
2529 "md/raid10:%s: unable to read back "
2531 " (%d sectors at %llu on %s)\n",
2533 (unsigned long long)(
2535 choose_data_offset(r10_bio
, rdev
)),
2536 bdevname(rdev
->bdev
, b
));
2537 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2540 bdevname(rdev
->bdev
, b
));
2544 "md/raid10:%s: read error corrected"
2545 " (%d sectors at %llu on %s)\n",
2547 (unsigned long long)(
2549 choose_data_offset(r10_bio
, rdev
)),
2550 bdevname(rdev
->bdev
, b
));
2551 atomic_add(s
, &rdev
->corrected_errors
);
2554 rdev_dec_pending(rdev
, mddev
);
2564 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2566 struct bio
*bio
= r10_bio
->master_bio
;
2567 struct mddev
*mddev
= r10_bio
->mddev
;
2568 struct r10conf
*conf
= mddev
->private;
2569 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2570 /* bio has the data to be written to slot 'i' where
2571 * we just recently had a write error.
2572 * We repeatedly clone the bio and trim down to one block,
2573 * then try the write. Where the write fails we record
2575 * It is conceivable that the bio doesn't exactly align with
2576 * blocks. We must handle this.
2578 * We currently own a reference to the rdev.
2584 int sect_to_write
= r10_bio
->sectors
;
2587 if (rdev
->badblocks
.shift
< 0)
2590 block_sectors
= 1 << rdev
->badblocks
.shift
;
2591 sector
= r10_bio
->sector
;
2592 sectors
= ((r10_bio
->sector
+ block_sectors
)
2593 & ~(sector_t
)(block_sectors
- 1))
2596 while (sect_to_write
) {
2598 if (sectors
> sect_to_write
)
2599 sectors
= sect_to_write
;
2600 /* Write at 'sector' for 'sectors' */
2601 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2602 bio_trim(wbio
, sector
- bio
->bi_iter
.bi_sector
, sectors
);
2603 wbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
2604 choose_data_offset(r10_bio
, rdev
) +
2605 (sector
- r10_bio
->sector
));
2606 wbio
->bi_bdev
= rdev
->bdev
;
2607 if (submit_bio_wait(WRITE
, wbio
) == 0)
2609 ok
= rdev_set_badblocks(rdev
, sector
,
2614 sect_to_write
-= sectors
;
2616 sectors
= block_sectors
;
2621 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2623 int slot
= r10_bio
->read_slot
;
2625 struct r10conf
*conf
= mddev
->private;
2626 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2627 char b
[BDEVNAME_SIZE
];
2628 unsigned long do_sync
;
2631 /* we got a read error. Maybe the drive is bad. Maybe just
2632 * the block and we can fix it.
2633 * We freeze all other IO, and try reading the block from
2634 * other devices. When we find one, we re-write
2635 * and check it that fixes the read error.
2636 * This is all done synchronously while the array is
2639 bio
= r10_bio
->devs
[slot
].bio
;
2640 bdevname(bio
->bi_bdev
, b
);
2642 r10_bio
->devs
[slot
].bio
= NULL
;
2644 if (mddev
->ro
== 0) {
2645 freeze_array(conf
, 1);
2646 fix_read_error(conf
, mddev
, r10_bio
);
2647 unfreeze_array(conf
);
2649 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2651 rdev_dec_pending(rdev
, mddev
);
2654 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2656 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2657 " read error for block %llu\n",
2659 (unsigned long long)r10_bio
->sector
);
2660 raid_end_bio_io(r10_bio
);
2664 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2665 slot
= r10_bio
->read_slot
;
2668 "md/raid10:%s: %s: redirecting "
2669 "sector %llu to another mirror\n",
2671 bdevname(rdev
->bdev
, b
),
2672 (unsigned long long)r10_bio
->sector
);
2673 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2675 bio_trim(bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
2676 r10_bio
->devs
[slot
].bio
= bio
;
2677 r10_bio
->devs
[slot
].rdev
= rdev
;
2678 bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
2679 + choose_data_offset(r10_bio
, rdev
);
2680 bio
->bi_bdev
= rdev
->bdev
;
2681 bio
->bi_rw
= READ
| do_sync
;
2682 bio
->bi_private
= r10_bio
;
2683 bio
->bi_end_io
= raid10_end_read_request
;
2684 if (max_sectors
< r10_bio
->sectors
) {
2685 /* Drat - have to split this up more */
2686 struct bio
*mbio
= r10_bio
->master_bio
;
2687 int sectors_handled
=
2688 r10_bio
->sector
+ max_sectors
2689 - mbio
->bi_iter
.bi_sector
;
2690 r10_bio
->sectors
= max_sectors
;
2691 spin_lock_irq(&conf
->device_lock
);
2692 if (mbio
->bi_phys_segments
== 0)
2693 mbio
->bi_phys_segments
= 2;
2695 mbio
->bi_phys_segments
++;
2696 spin_unlock_irq(&conf
->device_lock
);
2697 generic_make_request(bio
);
2699 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2701 r10_bio
->master_bio
= mbio
;
2702 r10_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2704 set_bit(R10BIO_ReadError
,
2706 r10_bio
->mddev
= mddev
;
2707 r10_bio
->sector
= mbio
->bi_iter
.bi_sector
2712 generic_make_request(bio
);
2715 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2717 /* Some sort of write request has finished and it
2718 * succeeded in writing where we thought there was a
2719 * bad block. So forget the bad block.
2720 * Or possibly if failed and we need to record
2724 struct md_rdev
*rdev
;
2726 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2727 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2728 for (m
= 0; m
< conf
->copies
; m
++) {
2729 int dev
= r10_bio
->devs
[m
].devnum
;
2730 rdev
= conf
->mirrors
[dev
].rdev
;
2731 if (r10_bio
->devs
[m
].bio
== NULL
)
2733 if (test_bit(BIO_UPTODATE
,
2734 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2735 rdev_clear_badblocks(
2737 r10_bio
->devs
[m
].addr
,
2738 r10_bio
->sectors
, 0);
2740 if (!rdev_set_badblocks(
2742 r10_bio
->devs
[m
].addr
,
2743 r10_bio
->sectors
, 0))
2744 md_error(conf
->mddev
, rdev
);
2746 rdev
= conf
->mirrors
[dev
].replacement
;
2747 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2749 if (test_bit(BIO_UPTODATE
,
2750 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2751 rdev_clear_badblocks(
2753 r10_bio
->devs
[m
].addr
,
2754 r10_bio
->sectors
, 0);
2756 if (!rdev_set_badblocks(
2758 r10_bio
->devs
[m
].addr
,
2759 r10_bio
->sectors
, 0))
2760 md_error(conf
->mddev
, rdev
);
2765 for (m
= 0; m
< conf
->copies
; m
++) {
2766 int dev
= r10_bio
->devs
[m
].devnum
;
2767 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2768 rdev
= conf
->mirrors
[dev
].rdev
;
2769 if (bio
== IO_MADE_GOOD
) {
2770 rdev_clear_badblocks(
2772 r10_bio
->devs
[m
].addr
,
2773 r10_bio
->sectors
, 0);
2774 rdev_dec_pending(rdev
, conf
->mddev
);
2775 } else if (bio
!= NULL
&&
2776 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2777 if (!narrow_write_error(r10_bio
, m
)) {
2778 md_error(conf
->mddev
, rdev
);
2779 set_bit(R10BIO_Degraded
,
2782 rdev_dec_pending(rdev
, conf
->mddev
);
2784 bio
= r10_bio
->devs
[m
].repl_bio
;
2785 rdev
= conf
->mirrors
[dev
].replacement
;
2786 if (rdev
&& bio
== IO_MADE_GOOD
) {
2787 rdev_clear_badblocks(
2789 r10_bio
->devs
[m
].addr
,
2790 r10_bio
->sectors
, 0);
2791 rdev_dec_pending(rdev
, conf
->mddev
);
2794 if (test_bit(R10BIO_WriteError
,
2796 close_write(r10_bio
);
2797 raid_end_bio_io(r10_bio
);
2801 static void raid10d(struct md_thread
*thread
)
2803 struct mddev
*mddev
= thread
->mddev
;
2804 struct r10bio
*r10_bio
;
2805 unsigned long flags
;
2806 struct r10conf
*conf
= mddev
->private;
2807 struct list_head
*head
= &conf
->retry_list
;
2808 struct blk_plug plug
;
2810 md_check_recovery(mddev
);
2812 blk_start_plug(&plug
);
2815 flush_pending_writes(conf
);
2817 spin_lock_irqsave(&conf
->device_lock
, flags
);
2818 if (list_empty(head
)) {
2819 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2822 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2823 list_del(head
->prev
);
2825 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2827 mddev
= r10_bio
->mddev
;
2828 conf
= mddev
->private;
2829 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2830 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2831 handle_write_completed(conf
, r10_bio
);
2832 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2833 reshape_request_write(mddev
, r10_bio
);
2834 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2835 sync_request_write(mddev
, r10_bio
);
2836 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2837 recovery_request_write(mddev
, r10_bio
);
2838 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2839 handle_read_error(mddev
, r10_bio
);
2841 /* just a partial read to be scheduled from a
2844 int slot
= r10_bio
->read_slot
;
2845 generic_make_request(r10_bio
->devs
[slot
].bio
);
2849 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2850 md_check_recovery(mddev
);
2852 blk_finish_plug(&plug
);
2856 static int init_resync(struct r10conf
*conf
)
2861 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2862 BUG_ON(conf
->r10buf_pool
);
2863 conf
->have_replacement
= 0;
2864 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2865 if (conf
->mirrors
[i
].replacement
)
2866 conf
->have_replacement
= 1;
2867 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2868 if (!conf
->r10buf_pool
)
2870 conf
->next_resync
= 0;
2875 * perform a "sync" on one "block"
2877 * We need to make sure that no normal I/O request - particularly write
2878 * requests - conflict with active sync requests.
2880 * This is achieved by tracking pending requests and a 'barrier' concept
2881 * that can be installed to exclude normal IO requests.
2883 * Resync and recovery are handled very differently.
2884 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2886 * For resync, we iterate over virtual addresses, read all copies,
2887 * and update if there are differences. If only one copy is live,
2889 * For recovery, we iterate over physical addresses, read a good
2890 * value for each non-in_sync drive, and over-write.
2892 * So, for recovery we may have several outstanding complex requests for a
2893 * given address, one for each out-of-sync device. We model this by allocating
2894 * a number of r10_bio structures, one for each out-of-sync device.
2895 * As we setup these structures, we collect all bio's together into a list
2896 * which we then process collectively to add pages, and then process again
2897 * to pass to generic_make_request.
2899 * The r10_bio structures are linked using a borrowed master_bio pointer.
2900 * This link is counted in ->remaining. When the r10_bio that points to NULL
2901 * has its remaining count decremented to 0, the whole complex operation
2906 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2907 int *skipped
, int go_faster
)
2909 struct r10conf
*conf
= mddev
->private;
2910 struct r10bio
*r10_bio
;
2911 struct bio
*biolist
= NULL
, *bio
;
2912 sector_t max_sector
, nr_sectors
;
2915 sector_t sync_blocks
;
2916 sector_t sectors_skipped
= 0;
2917 int chunks_skipped
= 0;
2918 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2920 if (!conf
->r10buf_pool
)
2921 if (init_resync(conf
))
2925 * Allow skipping a full rebuild for incremental assembly
2926 * of a clean array, like RAID1 does.
2928 if (mddev
->bitmap
== NULL
&&
2929 mddev
->recovery_cp
== MaxSector
&&
2930 mddev
->reshape_position
== MaxSector
&&
2931 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2932 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2933 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2934 conf
->fullsync
== 0) {
2936 return mddev
->dev_sectors
- sector_nr
;
2940 max_sector
= mddev
->dev_sectors
;
2941 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2942 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2943 max_sector
= mddev
->resync_max_sectors
;
2944 if (sector_nr
>= max_sector
) {
2945 /* If we aborted, we need to abort the
2946 * sync on the 'current' bitmap chucks (there can
2947 * be several when recovering multiple devices).
2948 * as we may have started syncing it but not finished.
2949 * We can find the current address in
2950 * mddev->curr_resync, but for recovery,
2951 * we need to convert that to several
2952 * virtual addresses.
2954 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2960 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2961 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2962 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2964 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2966 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2967 bitmap_end_sync(mddev
->bitmap
, sect
,
2971 /* completed sync */
2972 if ((!mddev
->bitmap
|| conf
->fullsync
)
2973 && conf
->have_replacement
2974 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2975 /* Completed a full sync so the replacements
2976 * are now fully recovered.
2978 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2979 if (conf
->mirrors
[i
].replacement
)
2980 conf
->mirrors
[i
].replacement
2986 bitmap_close_sync(mddev
->bitmap
);
2989 return sectors_skipped
;
2992 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2993 return reshape_request(mddev
, sector_nr
, skipped
);
2995 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2996 /* if there has been nothing to do on any drive,
2997 * then there is nothing to do at all..
3000 return (max_sector
- sector_nr
) + sectors_skipped
;
3003 if (max_sector
> mddev
->resync_max
)
3004 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
3006 /* make sure whole request will fit in a chunk - if chunks
3009 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
3010 max_sector
> (sector_nr
| chunk_mask
))
3011 max_sector
= (sector_nr
| chunk_mask
) + 1;
3013 * If there is non-resync activity waiting for us then
3014 * put in a delay to throttle resync.
3016 if (!go_faster
&& conf
->nr_waiting
)
3017 msleep_interruptible(1000);
3019 /* Again, very different code for resync and recovery.
3020 * Both must result in an r10bio with a list of bios that
3021 * have bi_end_io, bi_sector, bi_bdev set,
3022 * and bi_private set to the r10bio.
3023 * For recovery, we may actually create several r10bios
3024 * with 2 bios in each, that correspond to the bios in the main one.
3025 * In this case, the subordinate r10bios link back through a
3026 * borrowed master_bio pointer, and the counter in the master
3027 * includes a ref from each subordinate.
3029 /* First, we decide what to do and set ->bi_end_io
3030 * To end_sync_read if we want to read, and
3031 * end_sync_write if we will want to write.
3034 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
3035 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3036 /* recovery... the complicated one */
3040 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
3046 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
3048 if ((mirror
->rdev
== NULL
||
3049 test_bit(In_sync
, &mirror
->rdev
->flags
))
3051 (mirror
->replacement
== NULL
||
3053 &mirror
->replacement
->flags
)))
3057 /* want to reconstruct this device */
3059 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3060 if (sect
>= mddev
->resync_max_sectors
) {
3061 /* last stripe is not complete - don't
3062 * try to recover this sector.
3066 /* Unless we are doing a full sync, or a replacement
3067 * we only need to recover the block if it is set in
3070 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3072 if (sync_blocks
< max_sync
)
3073 max_sync
= sync_blocks
;
3075 mirror
->replacement
== NULL
&&
3077 /* yep, skip the sync_blocks here, but don't assume
3078 * that there will never be anything to do here
3080 chunks_skipped
= -1;
3084 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3086 raise_barrier(conf
, rb2
!= NULL
);
3087 atomic_set(&r10_bio
->remaining
, 0);
3089 r10_bio
->master_bio
= (struct bio
*)rb2
;
3091 atomic_inc(&rb2
->remaining
);
3092 r10_bio
->mddev
= mddev
;
3093 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3094 r10_bio
->sector
= sect
;
3096 raid10_find_phys(conf
, r10_bio
);
3098 /* Need to check if the array will still be
3101 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3102 if (conf
->mirrors
[j
].rdev
== NULL
||
3103 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3108 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3109 &sync_blocks
, still_degraded
);
3112 for (j
=0; j
<conf
->copies
;j
++) {
3114 int d
= r10_bio
->devs
[j
].devnum
;
3115 sector_t from_addr
, to_addr
;
3116 struct md_rdev
*rdev
;
3117 sector_t sector
, first_bad
;
3119 if (!conf
->mirrors
[d
].rdev
||
3120 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3122 /* This is where we read from */
3124 rdev
= conf
->mirrors
[d
].rdev
;
3125 sector
= r10_bio
->devs
[j
].addr
;
3127 if (is_badblock(rdev
, sector
, max_sync
,
3128 &first_bad
, &bad_sectors
)) {
3129 if (first_bad
> sector
)
3130 max_sync
= first_bad
- sector
;
3132 bad_sectors
-= (sector
3134 if (max_sync
> bad_sectors
)
3135 max_sync
= bad_sectors
;
3139 bio
= r10_bio
->devs
[0].bio
;
3141 bio
->bi_next
= biolist
;
3143 bio
->bi_private
= r10_bio
;
3144 bio
->bi_end_io
= end_sync_read
;
3146 from_addr
= r10_bio
->devs
[j
].addr
;
3147 bio
->bi_iter
.bi_sector
= from_addr
+
3149 bio
->bi_bdev
= rdev
->bdev
;
3150 atomic_inc(&rdev
->nr_pending
);
3151 /* and we write to 'i' (if not in_sync) */
3153 for (k
=0; k
<conf
->copies
; k
++)
3154 if (r10_bio
->devs
[k
].devnum
== i
)
3156 BUG_ON(k
== conf
->copies
);
3157 to_addr
= r10_bio
->devs
[k
].addr
;
3158 r10_bio
->devs
[0].devnum
= d
;
3159 r10_bio
->devs
[0].addr
= from_addr
;
3160 r10_bio
->devs
[1].devnum
= i
;
3161 r10_bio
->devs
[1].addr
= to_addr
;
3163 rdev
= mirror
->rdev
;
3164 if (!test_bit(In_sync
, &rdev
->flags
)) {
3165 bio
= r10_bio
->devs
[1].bio
;
3167 bio
->bi_next
= biolist
;
3169 bio
->bi_private
= r10_bio
;
3170 bio
->bi_end_io
= end_sync_write
;
3172 bio
->bi_iter
.bi_sector
= to_addr
3173 + rdev
->data_offset
;
3174 bio
->bi_bdev
= rdev
->bdev
;
3175 atomic_inc(&r10_bio
->remaining
);
3177 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3179 /* and maybe write to replacement */
3180 bio
= r10_bio
->devs
[1].repl_bio
;
3182 bio
->bi_end_io
= NULL
;
3183 rdev
= mirror
->replacement
;
3184 /* Note: if rdev != NULL, then bio
3185 * cannot be NULL as r10buf_pool_alloc will
3186 * have allocated it.
3187 * So the second test here is pointless.
3188 * But it keeps semantic-checkers happy, and
3189 * this comment keeps human reviewers
3192 if (rdev
== NULL
|| bio
== NULL
||
3193 test_bit(Faulty
, &rdev
->flags
))
3196 bio
->bi_next
= biolist
;
3198 bio
->bi_private
= r10_bio
;
3199 bio
->bi_end_io
= end_sync_write
;
3201 bio
->bi_iter
.bi_sector
= to_addr
+
3203 bio
->bi_bdev
= rdev
->bdev
;
3204 atomic_inc(&r10_bio
->remaining
);
3207 if (j
== conf
->copies
) {
3208 /* Cannot recover, so abort the recovery or
3209 * record a bad block */
3211 /* problem is that there are bad blocks
3212 * on other device(s)
3215 for (k
= 0; k
< conf
->copies
; k
++)
3216 if (r10_bio
->devs
[k
].devnum
== i
)
3218 if (!test_bit(In_sync
,
3219 &mirror
->rdev
->flags
)
3220 && !rdev_set_badblocks(
3222 r10_bio
->devs
[k
].addr
,
3225 if (mirror
->replacement
&&
3226 !rdev_set_badblocks(
3227 mirror
->replacement
,
3228 r10_bio
->devs
[k
].addr
,
3233 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3235 printk(KERN_INFO
"md/raid10:%s: insufficient "
3236 "working devices for recovery.\n",
3238 mirror
->recovery_disabled
3239 = mddev
->recovery_disabled
;
3243 atomic_dec(&rb2
->remaining
);
3248 if (biolist
== NULL
) {
3250 struct r10bio
*rb2
= r10_bio
;
3251 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3252 rb2
->master_bio
= NULL
;
3258 /* resync. Schedule a read for every block at this virt offset */
3261 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3263 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3264 &sync_blocks
, mddev
->degraded
) &&
3265 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3266 &mddev
->recovery
)) {
3267 /* We can skip this block */
3269 return sync_blocks
+ sectors_skipped
;
3271 if (sync_blocks
< max_sync
)
3272 max_sync
= sync_blocks
;
3273 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3276 r10_bio
->mddev
= mddev
;
3277 atomic_set(&r10_bio
->remaining
, 0);
3278 raise_barrier(conf
, 0);
3279 conf
->next_resync
= sector_nr
;
3281 r10_bio
->master_bio
= NULL
;
3282 r10_bio
->sector
= sector_nr
;
3283 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3284 raid10_find_phys(conf
, r10_bio
);
3285 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3287 for (i
= 0; i
< conf
->copies
; i
++) {
3288 int d
= r10_bio
->devs
[i
].devnum
;
3289 sector_t first_bad
, sector
;
3292 if (r10_bio
->devs
[i
].repl_bio
)
3293 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3295 bio
= r10_bio
->devs
[i
].bio
;
3297 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3298 if (conf
->mirrors
[d
].rdev
== NULL
||
3299 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3301 sector
= r10_bio
->devs
[i
].addr
;
3302 if (is_badblock(conf
->mirrors
[d
].rdev
,
3304 &first_bad
, &bad_sectors
)) {
3305 if (first_bad
> sector
)
3306 max_sync
= first_bad
- sector
;
3308 bad_sectors
-= (sector
- first_bad
);
3309 if (max_sync
> bad_sectors
)
3310 max_sync
= bad_sectors
;
3314 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3315 atomic_inc(&r10_bio
->remaining
);
3316 bio
->bi_next
= biolist
;
3318 bio
->bi_private
= r10_bio
;
3319 bio
->bi_end_io
= end_sync_read
;
3321 bio
->bi_iter
.bi_sector
= sector
+
3322 conf
->mirrors
[d
].rdev
->data_offset
;
3323 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3326 if (conf
->mirrors
[d
].replacement
== NULL
||
3328 &conf
->mirrors
[d
].replacement
->flags
))
3331 /* Need to set up for writing to the replacement */
3332 bio
= r10_bio
->devs
[i
].repl_bio
;
3334 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3336 sector
= r10_bio
->devs
[i
].addr
;
3337 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3338 bio
->bi_next
= biolist
;
3340 bio
->bi_private
= r10_bio
;
3341 bio
->bi_end_io
= end_sync_write
;
3343 bio
->bi_iter
.bi_sector
= sector
+
3344 conf
->mirrors
[d
].replacement
->data_offset
;
3345 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3350 for (i
=0; i
<conf
->copies
; i
++) {
3351 int d
= r10_bio
->devs
[i
].devnum
;
3352 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3353 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3355 if (r10_bio
->devs
[i
].repl_bio
&&
3356 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3358 conf
->mirrors
[d
].replacement
,
3368 if (sector_nr
+ max_sync
< max_sector
)
3369 max_sector
= sector_nr
+ max_sync
;
3372 int len
= PAGE_SIZE
;
3373 if (sector_nr
+ (len
>>9) > max_sector
)
3374 len
= (max_sector
- sector_nr
) << 9;
3377 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3379 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3380 if (bio_add_page(bio
, page
, len
, 0))
3384 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3385 for (bio2
= biolist
;
3386 bio2
&& bio2
!= bio
;
3387 bio2
= bio2
->bi_next
) {
3388 /* remove last page from this bio */
3390 bio2
->bi_iter
.bi_size
-= len
;
3391 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3395 nr_sectors
+= len
>>9;
3396 sector_nr
+= len
>>9;
3397 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3399 r10_bio
->sectors
= nr_sectors
;
3403 biolist
= biolist
->bi_next
;
3405 bio
->bi_next
= NULL
;
3406 r10_bio
= bio
->bi_private
;
3407 r10_bio
->sectors
= nr_sectors
;
3409 if (bio
->bi_end_io
== end_sync_read
) {
3410 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3411 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3412 generic_make_request(bio
);
3416 if (sectors_skipped
)
3417 /* pretend they weren't skipped, it makes
3418 * no important difference in this case
3420 md_done_sync(mddev
, sectors_skipped
, 1);
3422 return sectors_skipped
+ nr_sectors
;
3424 /* There is nowhere to write, so all non-sync
3425 * drives must be failed or in resync, all drives
3426 * have a bad block, so try the next chunk...
3428 if (sector_nr
+ max_sync
< max_sector
)
3429 max_sector
= sector_nr
+ max_sync
;
3431 sectors_skipped
+= (max_sector
- sector_nr
);
3433 sector_nr
= max_sector
;
3438 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3441 struct r10conf
*conf
= mddev
->private;
3444 raid_disks
= min(conf
->geo
.raid_disks
,
3445 conf
->prev
.raid_disks
);
3447 sectors
= conf
->dev_sectors
;
3449 size
= sectors
>> conf
->geo
.chunk_shift
;
3450 sector_div(size
, conf
->geo
.far_copies
);
3451 size
= size
* raid_disks
;
3452 sector_div(size
, conf
->geo
.near_copies
);
3454 return size
<< conf
->geo
.chunk_shift
;
3457 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3459 /* Calculate the number of sectors-per-device that will
3460 * actually be used, and set conf->dev_sectors and
3464 size
= size
>> conf
->geo
.chunk_shift
;
3465 sector_div(size
, conf
->geo
.far_copies
);
3466 size
= size
* conf
->geo
.raid_disks
;
3467 sector_div(size
, conf
->geo
.near_copies
);
3468 /* 'size' is now the number of chunks in the array */
3469 /* calculate "used chunks per device" */
3470 size
= size
* conf
->copies
;
3472 /* We need to round up when dividing by raid_disks to
3473 * get the stride size.
3475 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3477 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3479 if (conf
->geo
.far_offset
)
3480 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3482 sector_div(size
, conf
->geo
.far_copies
);
3483 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3487 enum geo_type
{geo_new
, geo_old
, geo_start
};
3488 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3491 int layout
, chunk
, disks
;
3494 layout
= mddev
->layout
;
3495 chunk
= mddev
->chunk_sectors
;
3496 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3499 layout
= mddev
->new_layout
;
3500 chunk
= mddev
->new_chunk_sectors
;
3501 disks
= mddev
->raid_disks
;
3503 default: /* avoid 'may be unused' warnings */
3504 case geo_start
: /* new when starting reshape - raid_disks not
3506 layout
= mddev
->new_layout
;
3507 chunk
= mddev
->new_chunk_sectors
;
3508 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3513 if (chunk
< (PAGE_SIZE
>> 9) ||
3514 !is_power_of_2(chunk
))
3517 fc
= (layout
>> 8) & 255;
3518 fo
= layout
& (1<<16);
3519 geo
->raid_disks
= disks
;
3520 geo
->near_copies
= nc
;
3521 geo
->far_copies
= fc
;
3522 geo
->far_offset
= fo
;
3523 geo
->far_set_size
= (layout
& (1<<17)) ? disks
/ fc
: disks
;
3524 geo
->chunk_mask
= chunk
- 1;
3525 geo
->chunk_shift
= ffz(~chunk
);
3529 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3531 struct r10conf
*conf
= NULL
;
3536 copies
= setup_geo(&geo
, mddev
, geo_new
);
3539 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3540 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3541 mdname(mddev
), PAGE_SIZE
);
3545 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3546 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3547 mdname(mddev
), mddev
->new_layout
);
3552 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3556 /* FIXME calc properly */
3557 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3558 max(0,-mddev
->delta_disks
)),
3563 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3568 conf
->copies
= copies
;
3569 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3570 r10bio_pool_free
, conf
);
3571 if (!conf
->r10bio_pool
)
3574 calc_sectors(conf
, mddev
->dev_sectors
);
3575 if (mddev
->reshape_position
== MaxSector
) {
3576 conf
->prev
= conf
->geo
;
3577 conf
->reshape_progress
= MaxSector
;
3579 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3583 conf
->reshape_progress
= mddev
->reshape_position
;
3584 if (conf
->prev
.far_offset
)
3585 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3587 /* far_copies must be 1 */
3588 conf
->prev
.stride
= conf
->dev_sectors
;
3590 spin_lock_init(&conf
->device_lock
);
3591 INIT_LIST_HEAD(&conf
->retry_list
);
3593 spin_lock_init(&conf
->resync_lock
);
3594 init_waitqueue_head(&conf
->wait_barrier
);
3596 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3600 conf
->mddev
= mddev
;
3605 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3608 if (conf
->r10bio_pool
)
3609 mempool_destroy(conf
->r10bio_pool
);
3610 kfree(conf
->mirrors
);
3611 safe_put_page(conf
->tmppage
);
3614 return ERR_PTR(err
);
3617 static int run(struct mddev
*mddev
)
3619 struct r10conf
*conf
;
3620 int i
, disk_idx
, chunk_size
;
3621 struct raid10_info
*disk
;
3622 struct md_rdev
*rdev
;
3624 sector_t min_offset_diff
= 0;
3626 bool discard_supported
= false;
3628 if (mddev
->private == NULL
) {
3629 conf
= setup_conf(mddev
);
3631 return PTR_ERR(conf
);
3632 mddev
->private = conf
;
3634 conf
= mddev
->private;
3638 mddev
->thread
= conf
->thread
;
3639 conf
->thread
= NULL
;
3641 chunk_size
= mddev
->chunk_sectors
<< 9;
3643 blk_queue_max_discard_sectors(mddev
->queue
,
3644 mddev
->chunk_sectors
);
3645 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3646 blk_queue_io_min(mddev
->queue
, chunk_size
);
3647 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3648 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3650 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3651 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3654 rdev_for_each(rdev
, mddev
) {
3656 struct request_queue
*q
;
3658 disk_idx
= rdev
->raid_disk
;
3661 if (disk_idx
>= conf
->geo
.raid_disks
&&
3662 disk_idx
>= conf
->prev
.raid_disks
)
3664 disk
= conf
->mirrors
+ disk_idx
;
3666 if (test_bit(Replacement
, &rdev
->flags
)) {
3667 if (disk
->replacement
)
3669 disk
->replacement
= rdev
;
3675 q
= bdev_get_queue(rdev
->bdev
);
3676 if (q
->merge_bvec_fn
)
3677 mddev
->merge_check_needed
= 1;
3678 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3679 if (!mddev
->reshape_backwards
)
3683 if (first
|| diff
< min_offset_diff
)
3684 min_offset_diff
= diff
;
3687 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3688 rdev
->data_offset
<< 9);
3690 disk
->head_position
= 0;
3692 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3693 discard_supported
= true;
3697 if (discard_supported
)
3698 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3701 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3704 /* need to check that every block has at least one working mirror */
3705 if (!enough(conf
, -1)) {
3706 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3711 if (conf
->reshape_progress
!= MaxSector
) {
3712 /* must ensure that shape change is supported */
3713 if (conf
->geo
.far_copies
!= 1 &&
3714 conf
->geo
.far_offset
== 0)
3716 if (conf
->prev
.far_copies
!= 1 &&
3717 conf
->prev
.far_offset
== 0)
3721 mddev
->degraded
= 0;
3723 i
< conf
->geo
.raid_disks
3724 || i
< conf
->prev
.raid_disks
;
3727 disk
= conf
->mirrors
+ i
;
3729 if (!disk
->rdev
&& disk
->replacement
) {
3730 /* The replacement is all we have - use it */
3731 disk
->rdev
= disk
->replacement
;
3732 disk
->replacement
= NULL
;
3733 clear_bit(Replacement
, &disk
->rdev
->flags
);
3737 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3738 disk
->head_position
= 0;
3741 disk
->rdev
->saved_raid_disk
< 0)
3744 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3747 if (mddev
->recovery_cp
!= MaxSector
)
3748 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3749 " -- starting background reconstruction\n",
3752 "md/raid10:%s: active with %d out of %d devices\n",
3753 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3754 conf
->geo
.raid_disks
);
3756 * Ok, everything is just fine now
3758 mddev
->dev_sectors
= conf
->dev_sectors
;
3759 size
= raid10_size(mddev
, 0, 0);
3760 md_set_array_sectors(mddev
, size
);
3761 mddev
->resync_max_sectors
= size
;
3764 int stripe
= conf
->geo
.raid_disks
*
3765 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3766 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3767 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3769 /* Calculate max read-ahead size.
3770 * We need to readahead at least twice a whole stripe....
3773 stripe
/= conf
->geo
.near_copies
;
3774 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3775 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3776 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3780 if (md_integrity_register(mddev
))
3783 if (conf
->reshape_progress
!= MaxSector
) {
3784 unsigned long before_length
, after_length
;
3786 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3787 conf
->prev
.far_copies
);
3788 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3789 conf
->geo
.far_copies
);
3791 if (max(before_length
, after_length
) > min_offset_diff
) {
3792 /* This cannot work */
3793 printk("md/raid10: offset difference not enough to continue reshape\n");
3796 conf
->offset_diff
= min_offset_diff
;
3798 conf
->reshape_safe
= conf
->reshape_progress
;
3799 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3800 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3801 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3802 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3803 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3810 md_unregister_thread(&mddev
->thread
);
3811 if (conf
->r10bio_pool
)
3812 mempool_destroy(conf
->r10bio_pool
);
3813 safe_put_page(conf
->tmppage
);
3814 kfree(conf
->mirrors
);
3816 mddev
->private = NULL
;
3821 static int stop(struct mddev
*mddev
)
3823 struct r10conf
*conf
= mddev
->private;
3825 raise_barrier(conf
, 0);
3826 lower_barrier(conf
);
3828 md_unregister_thread(&mddev
->thread
);
3830 /* the unplug fn references 'conf'*/
3831 blk_sync_queue(mddev
->queue
);
3833 if (conf
->r10bio_pool
)
3834 mempool_destroy(conf
->r10bio_pool
);
3835 safe_put_page(conf
->tmppage
);
3836 kfree(conf
->mirrors
);
3838 mddev
->private = NULL
;
3842 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3844 struct r10conf
*conf
= mddev
->private;
3848 raise_barrier(conf
, 0);
3851 lower_barrier(conf
);
3856 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3858 /* Resize of 'far' arrays is not supported.
3859 * For 'near' and 'offset' arrays we can set the
3860 * number of sectors used to be an appropriate multiple
3861 * of the chunk size.
3862 * For 'offset', this is far_copies*chunksize.
3863 * For 'near' the multiplier is the LCM of
3864 * near_copies and raid_disks.
3865 * So if far_copies > 1 && !far_offset, fail.
3866 * Else find LCM(raid_disks, near_copy)*far_copies and
3867 * multiply by chunk_size. Then round to this number.
3868 * This is mostly done by raid10_size()
3870 struct r10conf
*conf
= mddev
->private;
3871 sector_t oldsize
, size
;
3873 if (mddev
->reshape_position
!= MaxSector
)
3876 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3879 oldsize
= raid10_size(mddev
, 0, 0);
3880 size
= raid10_size(mddev
, sectors
, 0);
3881 if (mddev
->external_size
&&
3882 mddev
->array_sectors
> size
)
3884 if (mddev
->bitmap
) {
3885 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3889 md_set_array_sectors(mddev
, size
);
3890 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3891 revalidate_disk(mddev
->gendisk
);
3892 if (sectors
> mddev
->dev_sectors
&&
3893 mddev
->recovery_cp
> oldsize
) {
3894 mddev
->recovery_cp
= oldsize
;
3895 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3897 calc_sectors(conf
, sectors
);
3898 mddev
->dev_sectors
= conf
->dev_sectors
;
3899 mddev
->resync_max_sectors
= size
;
3903 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3905 struct md_rdev
*rdev
;
3906 struct r10conf
*conf
;
3908 if (mddev
->degraded
> 0) {
3909 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3911 return ERR_PTR(-EINVAL
);
3914 /* Set new parameters */
3915 mddev
->new_level
= 10;
3916 /* new layout: far_copies = 1, near_copies = 2 */
3917 mddev
->new_layout
= (1<<8) + 2;
3918 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3919 mddev
->delta_disks
= mddev
->raid_disks
;
3920 mddev
->raid_disks
*= 2;
3921 /* make sure it will be not marked as dirty */
3922 mddev
->recovery_cp
= MaxSector
;
3924 conf
= setup_conf(mddev
);
3925 if (!IS_ERR(conf
)) {
3926 rdev_for_each(rdev
, mddev
)
3927 if (rdev
->raid_disk
>= 0)
3928 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3935 static void *raid10_takeover(struct mddev
*mddev
)
3937 struct r0conf
*raid0_conf
;
3939 /* raid10 can take over:
3940 * raid0 - providing it has only two drives
3942 if (mddev
->level
== 0) {
3943 /* for raid0 takeover only one zone is supported */
3944 raid0_conf
= mddev
->private;
3945 if (raid0_conf
->nr_strip_zones
> 1) {
3946 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3947 " with more than one zone.\n",
3949 return ERR_PTR(-EINVAL
);
3951 return raid10_takeover_raid0(mddev
);
3953 return ERR_PTR(-EINVAL
);
3956 static int raid10_check_reshape(struct mddev
*mddev
)
3958 /* Called when there is a request to change
3959 * - layout (to ->new_layout)
3960 * - chunk size (to ->new_chunk_sectors)
3961 * - raid_disks (by delta_disks)
3962 * or when trying to restart a reshape that was ongoing.
3964 * We need to validate the request and possibly allocate
3965 * space if that might be an issue later.
3967 * Currently we reject any reshape of a 'far' mode array,
3968 * allow chunk size to change if new is generally acceptable,
3969 * allow raid_disks to increase, and allow
3970 * a switch between 'near' mode and 'offset' mode.
3972 struct r10conf
*conf
= mddev
->private;
3975 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3978 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3979 /* mustn't change number of copies */
3981 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3982 /* Cannot switch to 'far' mode */
3985 if (mddev
->array_sectors
& geo
.chunk_mask
)
3986 /* not factor of array size */
3989 if (!enough(conf
, -1))
3992 kfree(conf
->mirrors_new
);
3993 conf
->mirrors_new
= NULL
;
3994 if (mddev
->delta_disks
> 0) {
3995 /* allocate new 'mirrors' list */
3996 conf
->mirrors_new
= kzalloc(
3997 sizeof(struct raid10_info
)
3998 *(mddev
->raid_disks
+
3999 mddev
->delta_disks
),
4001 if (!conf
->mirrors_new
)
4008 * Need to check if array has failed when deciding whether to:
4010 * - remove non-faulty devices
4013 * This determination is simple when no reshape is happening.
4014 * However if there is a reshape, we need to carefully check
4015 * both the before and after sections.
4016 * This is because some failed devices may only affect one
4017 * of the two sections, and some non-in_sync devices may
4018 * be insync in the section most affected by failed devices.
4020 static int calc_degraded(struct r10conf
*conf
)
4022 int degraded
, degraded2
;
4027 /* 'prev' section first */
4028 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4029 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4030 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4032 else if (!test_bit(In_sync
, &rdev
->flags
))
4033 /* When we can reduce the number of devices in
4034 * an array, this might not contribute to
4035 * 'degraded'. It does now.
4040 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4044 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4045 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4046 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4048 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4049 /* If reshape is increasing the number of devices,
4050 * this section has already been recovered, so
4051 * it doesn't contribute to degraded.
4054 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4059 if (degraded2
> degraded
)
4064 static int raid10_start_reshape(struct mddev
*mddev
)
4066 /* A 'reshape' has been requested. This commits
4067 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4068 * This also checks if there are enough spares and adds them
4070 * We currently require enough spares to make the final
4071 * array non-degraded. We also require that the difference
4072 * between old and new data_offset - on each device - is
4073 * enough that we never risk over-writing.
4076 unsigned long before_length
, after_length
;
4077 sector_t min_offset_diff
= 0;
4080 struct r10conf
*conf
= mddev
->private;
4081 struct md_rdev
*rdev
;
4085 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4088 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4091 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4092 conf
->prev
.far_copies
);
4093 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4094 conf
->geo
.far_copies
);
4096 rdev_for_each(rdev
, mddev
) {
4097 if (!test_bit(In_sync
, &rdev
->flags
)
4098 && !test_bit(Faulty
, &rdev
->flags
))
4100 if (rdev
->raid_disk
>= 0) {
4101 long long diff
= (rdev
->new_data_offset
4102 - rdev
->data_offset
);
4103 if (!mddev
->reshape_backwards
)
4107 if (first
|| diff
< min_offset_diff
)
4108 min_offset_diff
= diff
;
4112 if (max(before_length
, after_length
) > min_offset_diff
)
4115 if (spares
< mddev
->delta_disks
)
4118 conf
->offset_diff
= min_offset_diff
;
4119 spin_lock_irq(&conf
->device_lock
);
4120 if (conf
->mirrors_new
) {
4121 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4122 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4124 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
4125 conf
->mirrors_old
= conf
->mirrors
;
4126 conf
->mirrors
= conf
->mirrors_new
;
4127 conf
->mirrors_new
= NULL
;
4129 setup_geo(&conf
->geo
, mddev
, geo_start
);
4131 if (mddev
->reshape_backwards
) {
4132 sector_t size
= raid10_size(mddev
, 0, 0);
4133 if (size
< mddev
->array_sectors
) {
4134 spin_unlock_irq(&conf
->device_lock
);
4135 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4139 mddev
->resync_max_sectors
= size
;
4140 conf
->reshape_progress
= size
;
4142 conf
->reshape_progress
= 0;
4143 spin_unlock_irq(&conf
->device_lock
);
4145 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4146 ret
= bitmap_resize(mddev
->bitmap
,
4147 raid10_size(mddev
, 0,
4148 conf
->geo
.raid_disks
),
4153 if (mddev
->delta_disks
> 0) {
4154 rdev_for_each(rdev
, mddev
)
4155 if (rdev
->raid_disk
< 0 &&
4156 !test_bit(Faulty
, &rdev
->flags
)) {
4157 if (raid10_add_disk(mddev
, rdev
) == 0) {
4158 if (rdev
->raid_disk
>=
4159 conf
->prev
.raid_disks
)
4160 set_bit(In_sync
, &rdev
->flags
);
4162 rdev
->recovery_offset
= 0;
4164 if (sysfs_link_rdev(mddev
, rdev
))
4165 /* Failure here is OK */;
4167 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4168 && !test_bit(Faulty
, &rdev
->flags
)) {
4169 /* This is a spare that was manually added */
4170 set_bit(In_sync
, &rdev
->flags
);
4173 /* When a reshape changes the number of devices,
4174 * ->degraded is measured against the larger of the
4175 * pre and post numbers.
4177 spin_lock_irq(&conf
->device_lock
);
4178 mddev
->degraded
= calc_degraded(conf
);
4179 spin_unlock_irq(&conf
->device_lock
);
4180 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4181 mddev
->reshape_position
= conf
->reshape_progress
;
4182 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4184 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4185 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4186 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4187 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4189 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4191 if (!mddev
->sync_thread
) {
4195 conf
->reshape_checkpoint
= jiffies
;
4196 md_wakeup_thread(mddev
->sync_thread
);
4197 md_new_event(mddev
);
4201 mddev
->recovery
= 0;
4202 spin_lock_irq(&conf
->device_lock
);
4203 conf
->geo
= conf
->prev
;
4204 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4205 rdev_for_each(rdev
, mddev
)
4206 rdev
->new_data_offset
= rdev
->data_offset
;
4208 conf
->reshape_progress
= MaxSector
;
4209 mddev
->reshape_position
= MaxSector
;
4210 spin_unlock_irq(&conf
->device_lock
);
4214 /* Calculate the last device-address that could contain
4215 * any block from the chunk that includes the array-address 's'
4216 * and report the next address.
4217 * i.e. the address returned will be chunk-aligned and after
4218 * any data that is in the chunk containing 's'.
4220 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4222 s
= (s
| geo
->chunk_mask
) + 1;
4223 s
>>= geo
->chunk_shift
;
4224 s
*= geo
->near_copies
;
4225 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4226 s
*= geo
->far_copies
;
4227 s
<<= geo
->chunk_shift
;
4231 /* Calculate the first device-address that could contain
4232 * any block from the chunk that includes the array-address 's'.
4233 * This too will be the start of a chunk
4235 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4237 s
>>= geo
->chunk_shift
;
4238 s
*= geo
->near_copies
;
4239 sector_div(s
, geo
->raid_disks
);
4240 s
*= geo
->far_copies
;
4241 s
<<= geo
->chunk_shift
;
4245 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4248 /* We simply copy at most one chunk (smallest of old and new)
4249 * at a time, possibly less if that exceeds RESYNC_PAGES,
4250 * or we hit a bad block or something.
4251 * This might mean we pause for normal IO in the middle of
4252 * a chunk, but that is not a problem was mddev->reshape_position
4253 * can record any location.
4255 * If we will want to write to a location that isn't
4256 * yet recorded as 'safe' (i.e. in metadata on disk) then
4257 * we need to flush all reshape requests and update the metadata.
4259 * When reshaping forwards (e.g. to more devices), we interpret
4260 * 'safe' as the earliest block which might not have been copied
4261 * down yet. We divide this by previous stripe size and multiply
4262 * by previous stripe length to get lowest device offset that we
4263 * cannot write to yet.
4264 * We interpret 'sector_nr' as an address that we want to write to.
4265 * From this we use last_device_address() to find where we might
4266 * write to, and first_device_address on the 'safe' position.
4267 * If this 'next' write position is after the 'safe' position,
4268 * we must update the metadata to increase the 'safe' position.
4270 * When reshaping backwards, we round in the opposite direction
4271 * and perform the reverse test: next write position must not be
4272 * less than current safe position.
4274 * In all this the minimum difference in data offsets
4275 * (conf->offset_diff - always positive) allows a bit of slack,
4276 * so next can be after 'safe', but not by more than offset_disk
4278 * We need to prepare all the bios here before we start any IO
4279 * to ensure the size we choose is acceptable to all devices.
4280 * The means one for each copy for write-out and an extra one for
4282 * We store the read-in bio in ->master_bio and the others in
4283 * ->devs[x].bio and ->devs[x].repl_bio.
4285 struct r10conf
*conf
= mddev
->private;
4286 struct r10bio
*r10_bio
;
4287 sector_t next
, safe
, last
;
4291 struct md_rdev
*rdev
;
4294 struct bio
*bio
, *read_bio
;
4295 int sectors_done
= 0;
4297 if (sector_nr
== 0) {
4298 /* If restarting in the middle, skip the initial sectors */
4299 if (mddev
->reshape_backwards
&&
4300 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4301 sector_nr
= (raid10_size(mddev
, 0, 0)
4302 - conf
->reshape_progress
);
4303 } else if (!mddev
->reshape_backwards
&&
4304 conf
->reshape_progress
> 0)
4305 sector_nr
= conf
->reshape_progress
;
4307 mddev
->curr_resync_completed
= sector_nr
;
4308 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4314 /* We don't use sector_nr to track where we are up to
4315 * as that doesn't work well for ->reshape_backwards.
4316 * So just use ->reshape_progress.
4318 if (mddev
->reshape_backwards
) {
4319 /* 'next' is the earliest device address that we might
4320 * write to for this chunk in the new layout
4322 next
= first_dev_address(conf
->reshape_progress
- 1,
4325 /* 'safe' is the last device address that we might read from
4326 * in the old layout after a restart
4328 safe
= last_dev_address(conf
->reshape_safe
- 1,
4331 if (next
+ conf
->offset_diff
< safe
)
4334 last
= conf
->reshape_progress
- 1;
4335 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4336 & conf
->prev
.chunk_mask
);
4337 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4338 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4340 /* 'next' is after the last device address that we
4341 * might write to for this chunk in the new layout
4343 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4345 /* 'safe' is the earliest device address that we might
4346 * read from in the old layout after a restart
4348 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4350 /* Need to update metadata if 'next' might be beyond 'safe'
4351 * as that would possibly corrupt data
4353 if (next
> safe
+ conf
->offset_diff
)
4356 sector_nr
= conf
->reshape_progress
;
4357 last
= sector_nr
| (conf
->geo
.chunk_mask
4358 & conf
->prev
.chunk_mask
);
4360 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4361 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4365 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4366 /* Need to update reshape_position in metadata */
4368 mddev
->reshape_position
= conf
->reshape_progress
;
4369 if (mddev
->reshape_backwards
)
4370 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4371 - conf
->reshape_progress
;
4373 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4374 conf
->reshape_checkpoint
= jiffies
;
4375 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4376 md_wakeup_thread(mddev
->thread
);
4377 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4378 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4379 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4380 allow_barrier(conf
);
4381 return sectors_done
;
4383 conf
->reshape_safe
= mddev
->reshape_position
;
4384 allow_barrier(conf
);
4388 /* Now schedule reads for blocks from sector_nr to last */
4389 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4391 raise_barrier(conf
, sectors_done
!= 0);
4392 atomic_set(&r10_bio
->remaining
, 0);
4393 r10_bio
->mddev
= mddev
;
4394 r10_bio
->sector
= sector_nr
;
4395 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4396 r10_bio
->sectors
= last
- sector_nr
+ 1;
4397 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4398 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4401 /* Cannot read from here, so need to record bad blocks
4402 * on all the target devices.
4405 mempool_free(r10_bio
, conf
->r10buf_pool
);
4406 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4407 return sectors_done
;
4410 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4412 read_bio
->bi_bdev
= rdev
->bdev
;
4413 read_bio
->bi_iter
.bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4414 + rdev
->data_offset
);
4415 read_bio
->bi_private
= r10_bio
;
4416 read_bio
->bi_end_io
= end_sync_read
;
4417 read_bio
->bi_rw
= READ
;
4418 read_bio
->bi_flags
&= (~0UL << BIO_RESET_BITS
);
4419 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4420 read_bio
->bi_vcnt
= 0;
4421 read_bio
->bi_iter
.bi_size
= 0;
4422 r10_bio
->master_bio
= read_bio
;
4423 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4425 /* Now find the locations in the new layout */
4426 __raid10_find_phys(&conf
->geo
, r10_bio
);
4429 read_bio
->bi_next
= NULL
;
4431 for (s
= 0; s
< conf
->copies
*2; s
++) {
4433 int d
= r10_bio
->devs
[s
/2].devnum
;
4434 struct md_rdev
*rdev2
;
4436 rdev2
= conf
->mirrors
[d
].replacement
;
4437 b
= r10_bio
->devs
[s
/2].repl_bio
;
4439 rdev2
= conf
->mirrors
[d
].rdev
;
4440 b
= r10_bio
->devs
[s
/2].bio
;
4442 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4446 b
->bi_bdev
= rdev2
->bdev
;
4447 b
->bi_iter
.bi_sector
= r10_bio
->devs
[s
/2].addr
+
4448 rdev2
->new_data_offset
;
4449 b
->bi_private
= r10_bio
;
4450 b
->bi_end_io
= end_reshape_write
;
4456 /* Now add as many pages as possible to all of these bios. */
4459 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4460 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4461 int len
= (max_sectors
- s
) << 9;
4462 if (len
> PAGE_SIZE
)
4464 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4466 if (bio_add_page(bio
, page
, len
, 0))
4469 /* Didn't fit, must stop */
4471 bio2
&& bio2
!= bio
;
4472 bio2
= bio2
->bi_next
) {
4473 /* Remove last page from this bio */
4475 bio2
->bi_iter
.bi_size
-= len
;
4476 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4480 sector_nr
+= len
>> 9;
4481 nr_sectors
+= len
>> 9;
4484 r10_bio
->sectors
= nr_sectors
;
4486 /* Now submit the read */
4487 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4488 atomic_inc(&r10_bio
->remaining
);
4489 read_bio
->bi_next
= NULL
;
4490 generic_make_request(read_bio
);
4491 sector_nr
+= nr_sectors
;
4492 sectors_done
+= nr_sectors
;
4493 if (sector_nr
<= last
)
4496 /* Now that we have done the whole section we can
4497 * update reshape_progress
4499 if (mddev
->reshape_backwards
)
4500 conf
->reshape_progress
-= sectors_done
;
4502 conf
->reshape_progress
+= sectors_done
;
4504 return sectors_done
;
4507 static void end_reshape_request(struct r10bio
*r10_bio
);
4508 static int handle_reshape_read_error(struct mddev
*mddev
,
4509 struct r10bio
*r10_bio
);
4510 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4512 /* Reshape read completed. Hopefully we have a block
4514 * If we got a read error then we do sync 1-page reads from
4515 * elsewhere until we find the data - or give up.
4517 struct r10conf
*conf
= mddev
->private;
4520 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4521 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4522 /* Reshape has been aborted */
4523 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4527 /* We definitely have the data in the pages, schedule the
4530 atomic_set(&r10_bio
->remaining
, 1);
4531 for (s
= 0; s
< conf
->copies
*2; s
++) {
4533 int d
= r10_bio
->devs
[s
/2].devnum
;
4534 struct md_rdev
*rdev
;
4536 rdev
= conf
->mirrors
[d
].replacement
;
4537 b
= r10_bio
->devs
[s
/2].repl_bio
;
4539 rdev
= conf
->mirrors
[d
].rdev
;
4540 b
= r10_bio
->devs
[s
/2].bio
;
4542 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4544 atomic_inc(&rdev
->nr_pending
);
4545 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4546 atomic_inc(&r10_bio
->remaining
);
4548 generic_make_request(b
);
4550 end_reshape_request(r10_bio
);
4553 static void end_reshape(struct r10conf
*conf
)
4555 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4558 spin_lock_irq(&conf
->device_lock
);
4559 conf
->prev
= conf
->geo
;
4560 md_finish_reshape(conf
->mddev
);
4562 conf
->reshape_progress
= MaxSector
;
4563 spin_unlock_irq(&conf
->device_lock
);
4565 /* read-ahead size must cover two whole stripes, which is
4566 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4568 if (conf
->mddev
->queue
) {
4569 int stripe
= conf
->geo
.raid_disks
*
4570 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4571 stripe
/= conf
->geo
.near_copies
;
4572 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4573 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4579 static int handle_reshape_read_error(struct mddev
*mddev
,
4580 struct r10bio
*r10_bio
)
4582 /* Use sync reads to get the blocks from somewhere else */
4583 int sectors
= r10_bio
->sectors
;
4584 struct r10conf
*conf
= mddev
->private;
4586 struct r10bio r10_bio
;
4587 struct r10dev devs
[conf
->copies
];
4589 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4592 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4594 r10b
->sector
= r10_bio
->sector
;
4595 __raid10_find_phys(&conf
->prev
, r10b
);
4600 int first_slot
= slot
;
4602 if (s
> (PAGE_SIZE
>> 9))
4606 int d
= r10b
->devs
[slot
].devnum
;
4607 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4610 test_bit(Faulty
, &rdev
->flags
) ||
4611 !test_bit(In_sync
, &rdev
->flags
))
4614 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4615 success
= sync_page_io(rdev
,
4624 if (slot
>= conf
->copies
)
4626 if (slot
== first_slot
)
4630 /* couldn't read this block, must give up */
4631 set_bit(MD_RECOVERY_INTR
,
4641 static void end_reshape_write(struct bio
*bio
, int error
)
4643 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4644 struct r10bio
*r10_bio
= bio
->bi_private
;
4645 struct mddev
*mddev
= r10_bio
->mddev
;
4646 struct r10conf
*conf
= mddev
->private;
4650 struct md_rdev
*rdev
= NULL
;
4652 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4654 rdev
= conf
->mirrors
[d
].replacement
;
4657 rdev
= conf
->mirrors
[d
].rdev
;
4661 /* FIXME should record badblock */
4662 md_error(mddev
, rdev
);
4665 rdev_dec_pending(rdev
, mddev
);
4666 end_reshape_request(r10_bio
);
4669 static void end_reshape_request(struct r10bio
*r10_bio
)
4671 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4673 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4674 bio_put(r10_bio
->master_bio
);
4678 static void raid10_finish_reshape(struct mddev
*mddev
)
4680 struct r10conf
*conf
= mddev
->private;
4682 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4685 if (mddev
->delta_disks
> 0) {
4686 sector_t size
= raid10_size(mddev
, 0, 0);
4687 md_set_array_sectors(mddev
, size
);
4688 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4689 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4690 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4692 mddev
->resync_max_sectors
= size
;
4693 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4694 revalidate_disk(mddev
->gendisk
);
4697 for (d
= conf
->geo
.raid_disks
;
4698 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4700 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4702 clear_bit(In_sync
, &rdev
->flags
);
4703 rdev
= conf
->mirrors
[d
].replacement
;
4705 clear_bit(In_sync
, &rdev
->flags
);
4708 mddev
->layout
= mddev
->new_layout
;
4709 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4710 mddev
->reshape_position
= MaxSector
;
4711 mddev
->delta_disks
= 0;
4712 mddev
->reshape_backwards
= 0;
4715 static struct md_personality raid10_personality
=
4719 .owner
= THIS_MODULE
,
4720 .make_request
= make_request
,
4724 .error_handler
= error
,
4725 .hot_add_disk
= raid10_add_disk
,
4726 .hot_remove_disk
= raid10_remove_disk
,
4727 .spare_active
= raid10_spare_active
,
4728 .sync_request
= sync_request
,
4729 .quiesce
= raid10_quiesce
,
4730 .size
= raid10_size
,
4731 .resize
= raid10_resize
,
4732 .takeover
= raid10_takeover
,
4733 .check_reshape
= raid10_check_reshape
,
4734 .start_reshape
= raid10_start_reshape
,
4735 .finish_reshape
= raid10_finish_reshape
,
4738 static int __init
raid_init(void)
4740 return register_md_personality(&raid10_personality
);
4743 static void raid_exit(void)
4745 unregister_md_personality(&raid10_personality
);
4748 module_init(raid_init
);
4749 module_exit(raid_exit
);
4750 MODULE_LICENSE("GPL");
4751 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4752 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4753 MODULE_ALIAS("md-raid10");
4754 MODULE_ALIAS("md-level-10");
4756 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);