2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 * use_far_sets (stored in bit 17 of layout )
43 * The data to be stored is divided into chunks using chunksize. Each device
44 * is divided into far_copies sections. In each section, chunks are laid out
45 * in a style similar to raid0, but near_copies copies of each chunk is stored
46 * (each on a different drive). The starting device for each section is offset
47 * near_copies from the starting device of the previous section. Thus there
48 * are (near_copies * far_copies) of each chunk, and each is on a different
49 * drive. near_copies and far_copies must be at least one, and their product
50 * is at most raid_disks.
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of being very far
54 * apart on disk, there are adjacent stripes.
56 * The far and offset algorithms are handled slightly differently if
57 * 'use_far_sets' is true. In this case, the array's devices are grouped into
58 * sets that are (near_copies * far_copies) in size. The far copied stripes
59 * are still shifted by 'near_copies' devices, but this shifting stays confined
60 * to the set rather than the entire array. This is done to improve the number
61 * of device combinations that can fail without causing the array to fail.
62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68 * [A B] [C D] [A B] [C D E]
69 * |...| |...| |...| | ... |
70 * [B A] [D C] [B A] [E C D]
74 * Number of guaranteed r10bios in case of extreme VM load:
76 #define NR_RAID10_BIOS 256
78 /* when we get a read error on a read-only array, we redirect to another
79 * device without failing the first device, or trying to over-write to
80 * correct the read error. To keep track of bad blocks on a per-bio
81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83 #define IO_BLOCKED ((struct bio *)1)
84 /* When we successfully write to a known bad-block, we need to remove the
85 * bad-block marking which must be done from process context. So we record
86 * the success by setting devs[n].bio to IO_MADE_GOOD
88 #define IO_MADE_GOOD ((struct bio *)2)
90 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92 /* When there are this many requests queued to be written by
93 * the raid10 thread, we become 'congested' to provide back-pressure
96 static int max_queued_requests
= 1024;
98 static void allow_barrier(struct r10conf
*conf
);
99 static void lower_barrier(struct r10conf
*conf
);
100 static int _enough(struct r10conf
*conf
, int previous
, int ignore
);
101 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
103 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
);
104 static void end_reshape_write(struct bio
*bio
, int error
);
105 static void end_reshape(struct r10conf
*conf
);
107 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
109 struct r10conf
*conf
= data
;
110 int size
= offsetof(struct r10bio
, devs
[conf
->copies
]);
112 /* allocate a r10bio with room for raid_disks entries in the
114 return kzalloc(size
, gfp_flags
);
117 static void r10bio_pool_free(void *r10_bio
, void *data
)
122 /* Maximum size of each resync request */
123 #define RESYNC_BLOCK_SIZE (64*1024)
124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
125 /* amount of memory to reserve for resync requests */
126 #define RESYNC_WINDOW (1024*1024)
127 /* maximum number of concurrent requests, memory permitting */
128 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
131 * When performing a resync, we need to read and compare, so
132 * we need as many pages are there are copies.
133 * When performing a recovery, we need 2 bios, one for read,
134 * one for write (we recover only one drive per r10buf)
137 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
139 struct r10conf
*conf
= data
;
141 struct r10bio
*r10_bio
;
146 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
150 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
) ||
151 test_bit(MD_RECOVERY_RESHAPE
, &conf
->mddev
->recovery
))
152 nalloc
= conf
->copies
; /* resync */
154 nalloc
= 2; /* recovery */
159 for (j
= nalloc
; j
-- ; ) {
160 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
163 r10_bio
->devs
[j
].bio
= bio
;
164 if (!conf
->have_replacement
)
166 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
169 r10_bio
->devs
[j
].repl_bio
= bio
;
172 * Allocate RESYNC_PAGES data pages and attach them
175 for (j
= 0 ; j
< nalloc
; j
++) {
176 struct bio
*rbio
= r10_bio
->devs
[j
].repl_bio
;
177 bio
= r10_bio
->devs
[j
].bio
;
178 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
179 if (j
> 0 && !test_bit(MD_RECOVERY_SYNC
,
180 &conf
->mddev
->recovery
)) {
181 /* we can share bv_page's during recovery
183 struct bio
*rbio
= r10_bio
->devs
[0].bio
;
184 page
= rbio
->bi_io_vec
[i
].bv_page
;
187 page
= alloc_page(gfp_flags
);
191 bio
->bi_io_vec
[i
].bv_page
= page
;
193 rbio
->bi_io_vec
[i
].bv_page
= page
;
201 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
203 for (i
= 0; i
< RESYNC_PAGES
; i
++)
204 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
207 for ( ; j
< nalloc
; j
++) {
208 if (r10_bio
->devs
[j
].bio
)
209 bio_put(r10_bio
->devs
[j
].bio
);
210 if (r10_bio
->devs
[j
].repl_bio
)
211 bio_put(r10_bio
->devs
[j
].repl_bio
);
213 r10bio_pool_free(r10_bio
, conf
);
217 static void r10buf_pool_free(void *__r10_bio
, void *data
)
220 struct r10conf
*conf
= data
;
221 struct r10bio
*r10bio
= __r10_bio
;
224 for (j
=0; j
< conf
->copies
; j
++) {
225 struct bio
*bio
= r10bio
->devs
[j
].bio
;
227 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
228 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
229 bio
->bi_io_vec
[i
].bv_page
= NULL
;
233 bio
= r10bio
->devs
[j
].repl_bio
;
237 r10bio_pool_free(r10bio
, conf
);
240 static void put_all_bios(struct r10conf
*conf
, struct r10bio
*r10_bio
)
244 for (i
= 0; i
< conf
->copies
; i
++) {
245 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
246 if (!BIO_SPECIAL(*bio
))
249 bio
= &r10_bio
->devs
[i
].repl_bio
;
250 if (r10_bio
->read_slot
< 0 && !BIO_SPECIAL(*bio
))
256 static void free_r10bio(struct r10bio
*r10_bio
)
258 struct r10conf
*conf
= r10_bio
->mddev
->private;
260 put_all_bios(conf
, r10_bio
);
261 mempool_free(r10_bio
, conf
->r10bio_pool
);
264 static void put_buf(struct r10bio
*r10_bio
)
266 struct r10conf
*conf
= r10_bio
->mddev
->private;
268 mempool_free(r10_bio
, conf
->r10buf_pool
);
273 static void reschedule_retry(struct r10bio
*r10_bio
)
276 struct mddev
*mddev
= r10_bio
->mddev
;
277 struct r10conf
*conf
= mddev
->private;
279 spin_lock_irqsave(&conf
->device_lock
, flags
);
280 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
282 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
284 /* wake up frozen array... */
285 wake_up(&conf
->wait_barrier
);
287 md_wakeup_thread(mddev
->thread
);
291 * raid_end_bio_io() is called when we have finished servicing a mirrored
292 * operation and are ready to return a success/failure code to the buffer
295 static void raid_end_bio_io(struct r10bio
*r10_bio
)
297 struct bio
*bio
= r10_bio
->master_bio
;
299 struct r10conf
*conf
= r10_bio
->mddev
->private;
301 if (bio
->bi_phys_segments
) {
303 spin_lock_irqsave(&conf
->device_lock
, flags
);
304 bio
->bi_phys_segments
--;
305 done
= (bio
->bi_phys_segments
== 0);
306 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
309 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
310 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
314 * Wake up any possible resync thread that waits for the device
319 free_r10bio(r10_bio
);
323 * Update disk head position estimator based on IRQ completion info.
325 static inline void update_head_pos(int slot
, struct r10bio
*r10_bio
)
327 struct r10conf
*conf
= r10_bio
->mddev
->private;
329 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
330 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
334 * Find the disk number which triggered given bio
336 static int find_bio_disk(struct r10conf
*conf
, struct r10bio
*r10_bio
,
337 struct bio
*bio
, int *slotp
, int *replp
)
342 for (slot
= 0; slot
< conf
->copies
; slot
++) {
343 if (r10_bio
->devs
[slot
].bio
== bio
)
345 if (r10_bio
->devs
[slot
].repl_bio
== bio
) {
351 BUG_ON(slot
== conf
->copies
);
352 update_head_pos(slot
, r10_bio
);
358 return r10_bio
->devs
[slot
].devnum
;
361 static void raid10_end_read_request(struct bio
*bio
, int error
)
363 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
364 struct r10bio
*r10_bio
= bio
->bi_private
;
366 struct md_rdev
*rdev
;
367 struct r10conf
*conf
= r10_bio
->mddev
->private;
370 slot
= r10_bio
->read_slot
;
371 dev
= r10_bio
->devs
[slot
].devnum
;
372 rdev
= r10_bio
->devs
[slot
].rdev
;
374 * this branch is our 'one mirror IO has finished' event handler:
376 update_head_pos(slot
, r10_bio
);
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
388 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
395 if (!_enough(conf
, test_bit(R10BIO_Previous
, &r10_bio
->state
),
400 raid_end_bio_io(r10_bio
);
401 rdev_dec_pending(rdev
, conf
->mddev
);
404 * oops, read error - keep the refcount on the rdev
406 char b
[BDEVNAME_SIZE
];
407 printk_ratelimited(KERN_ERR
408 "md/raid10:%s: %s: rescheduling sector %llu\n",
410 bdevname(rdev
->bdev
, b
),
411 (unsigned long long)r10_bio
->sector
);
412 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
413 reschedule_retry(r10_bio
);
417 static void close_write(struct r10bio
*r10_bio
)
419 /* clear the bitmap if all writes complete successfully */
420 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
422 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
424 md_write_end(r10_bio
->mddev
);
427 static void one_write_done(struct r10bio
*r10_bio
)
429 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
430 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
431 reschedule_retry(r10_bio
);
433 close_write(r10_bio
);
434 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
435 reschedule_retry(r10_bio
);
437 raid_end_bio_io(r10_bio
);
442 static void raid10_end_write_request(struct bio
*bio
, int error
)
444 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
445 struct r10bio
*r10_bio
= bio
->bi_private
;
448 struct r10conf
*conf
= r10_bio
->mddev
->private;
450 struct md_rdev
*rdev
= NULL
;
452 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
455 rdev
= conf
->mirrors
[dev
].replacement
;
459 rdev
= conf
->mirrors
[dev
].rdev
;
462 * this branch is our 'one mirror IO has finished' event handler:
466 /* Never record new bad blocks to replacement,
469 md_error(rdev
->mddev
, rdev
);
471 set_bit(WriteErrorSeen
, &rdev
->flags
);
472 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
473 set_bit(MD_RECOVERY_NEEDED
,
474 &rdev
->mddev
->recovery
);
475 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
480 * Set R10BIO_Uptodate in our master bio, so that
481 * we will return a good error code for to the higher
482 * levels even if IO on some other mirrored buffer fails.
484 * The 'master' represents the composite IO operation to
485 * user-side. So if something waits for IO, then it will
486 * wait for the 'master' bio.
492 * Do not set R10BIO_Uptodate if the current device is
493 * rebuilding or Faulty. This is because we cannot use
494 * such device for properly reading the data back (we could
495 * potentially use it, if the current write would have felt
496 * before rdev->recovery_offset, but for simplicity we don't
499 if (test_bit(In_sync
, &rdev
->flags
) &&
500 !test_bit(Faulty
, &rdev
->flags
))
501 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
503 /* Maybe we can clear some bad blocks. */
504 if (is_badblock(rdev
,
505 r10_bio
->devs
[slot
].addr
,
507 &first_bad
, &bad_sectors
)) {
510 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
512 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
514 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
520 * Let's see if all mirrored write operations have finished
523 one_write_done(r10_bio
);
525 rdev_dec_pending(rdev
, conf
->mddev
);
529 * RAID10 layout manager
530 * As well as the chunksize and raid_disks count, there are two
531 * parameters: near_copies and far_copies.
532 * near_copies * far_copies must be <= raid_disks.
533 * Normally one of these will be 1.
534 * If both are 1, we get raid0.
535 * If near_copies == raid_disks, we get raid1.
537 * Chunks are laid out in raid0 style with near_copies copies of the
538 * first chunk, followed by near_copies copies of the next chunk and
540 * If far_copies > 1, then after 1/far_copies of the array has been assigned
541 * as described above, we start again with a device offset of near_copies.
542 * So we effectively have another copy of the whole array further down all
543 * the drives, but with blocks on different drives.
544 * With this layout, and block is never stored twice on the one device.
546 * raid10_find_phys finds the sector offset of a given virtual sector
547 * on each device that it is on.
549 * raid10_find_virt does the reverse mapping, from a device and a
550 * sector offset to a virtual address
553 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
561 int last_far_set_start
, last_far_set_size
;
563 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
564 last_far_set_start
*= geo
->far_set_size
;
566 last_far_set_size
= geo
->far_set_size
;
567 last_far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
569 /* now calculate first sector/dev */
570 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
571 sector
= r10bio
->sector
& geo
->chunk_mask
;
573 chunk
*= geo
->near_copies
;
575 dev
= sector_div(stripe
, geo
->raid_disks
);
577 stripe
*= geo
->far_copies
;
579 sector
+= stripe
<< geo
->chunk_shift
;
581 /* and calculate all the others */
582 for (n
= 0; n
< geo
->near_copies
; n
++) {
586 r10bio
->devs
[slot
].devnum
= d
;
587 r10bio
->devs
[slot
].addr
= s
;
590 for (f
= 1; f
< geo
->far_copies
; f
++) {
591 set
= d
/ geo
->far_set_size
;
592 d
+= geo
->near_copies
;
594 if ((geo
->raid_disks
% geo
->far_set_size
) &&
595 (d
> last_far_set_start
)) {
596 d
-= last_far_set_start
;
597 d
%= last_far_set_size
;
598 d
+= last_far_set_start
;
600 d
%= geo
->far_set_size
;
601 d
+= geo
->far_set_size
* set
;
604 r10bio
->devs
[slot
].devnum
= d
;
605 r10bio
->devs
[slot
].addr
= s
;
609 if (dev
>= geo
->raid_disks
) {
611 sector
+= (geo
->chunk_mask
+ 1);
616 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
618 struct geom
*geo
= &conf
->geo
;
620 if (conf
->reshape_progress
!= MaxSector
&&
621 ((r10bio
->sector
>= conf
->reshape_progress
) !=
622 conf
->mddev
->reshape_backwards
)) {
623 set_bit(R10BIO_Previous
, &r10bio
->state
);
626 clear_bit(R10BIO_Previous
, &r10bio
->state
);
628 __raid10_find_phys(geo
, r10bio
);
631 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
633 sector_t offset
, chunk
, vchunk
;
634 /* Never use conf->prev as this is only called during resync
635 * or recovery, so reshape isn't happening
637 struct geom
*geo
= &conf
->geo
;
638 int far_set_start
= (dev
/ geo
->far_set_size
) * geo
->far_set_size
;
639 int far_set_size
= geo
->far_set_size
;
640 int last_far_set_start
;
642 if (geo
->raid_disks
% geo
->far_set_size
) {
643 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
644 last_far_set_start
*= geo
->far_set_size
;
646 if (dev
>= last_far_set_start
) {
647 far_set_size
= geo
->far_set_size
;
648 far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
649 far_set_start
= last_far_set_start
;
653 offset
= sector
& geo
->chunk_mask
;
654 if (geo
->far_offset
) {
656 chunk
= sector
>> geo
->chunk_shift
;
657 fc
= sector_div(chunk
, geo
->far_copies
);
658 dev
-= fc
* geo
->near_copies
;
659 if (dev
< far_set_start
)
662 while (sector
>= geo
->stride
) {
663 sector
-= geo
->stride
;
664 if (dev
< (geo
->near_copies
+ far_set_start
))
665 dev
+= far_set_size
- geo
->near_copies
;
667 dev
-= geo
->near_copies
;
669 chunk
= sector
>> geo
->chunk_shift
;
671 vchunk
= chunk
* geo
->raid_disks
+ dev
;
672 sector_div(vchunk
, geo
->near_copies
);
673 return (vchunk
<< geo
->chunk_shift
) + offset
;
677 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
679 * @bvm: properties of new bio
680 * @biovec: the request that could be merged to it.
682 * Return amount of bytes we can accept at this offset
683 * This requires checking for end-of-chunk if near_copies != raid_disks,
684 * and for subordinate merge_bvec_fns if merge_check_needed.
686 static int raid10_mergeable_bvec(struct request_queue
*q
,
687 struct bvec_merge_data
*bvm
,
688 struct bio_vec
*biovec
)
690 struct mddev
*mddev
= q
->queuedata
;
691 struct r10conf
*conf
= mddev
->private;
692 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
694 unsigned int chunk_sectors
;
695 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
696 struct geom
*geo
= &conf
->geo
;
698 chunk_sectors
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
) + 1;
699 if (conf
->reshape_progress
!= MaxSector
&&
700 ((sector
>= conf
->reshape_progress
) !=
701 conf
->mddev
->reshape_backwards
))
704 if (geo
->near_copies
< geo
->raid_disks
) {
705 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1))
706 + bio_sectors
)) << 9;
708 /* bio_add cannot handle a negative return */
710 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
711 return biovec
->bv_len
;
713 max
= biovec
->bv_len
;
715 if (mddev
->merge_check_needed
) {
717 struct r10bio r10_bio
;
718 struct r10dev devs
[conf
->copies
];
720 struct r10bio
*r10_bio
= &on_stack
.r10_bio
;
722 if (conf
->reshape_progress
!= MaxSector
) {
723 /* Cannot give any guidance during reshape */
724 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
725 return biovec
->bv_len
;
728 r10_bio
->sector
= sector
;
729 raid10_find_phys(conf
, r10_bio
);
731 for (s
= 0; s
< conf
->copies
; s
++) {
732 int disk
= r10_bio
->devs
[s
].devnum
;
733 struct md_rdev
*rdev
= rcu_dereference(
734 conf
->mirrors
[disk
].rdev
);
735 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
736 struct request_queue
*q
=
737 bdev_get_queue(rdev
->bdev
);
738 if (q
->merge_bvec_fn
) {
739 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
741 bvm
->bi_bdev
= rdev
->bdev
;
742 max
= min(max
, q
->merge_bvec_fn(
746 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
747 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
748 struct request_queue
*q
=
749 bdev_get_queue(rdev
->bdev
);
750 if (q
->merge_bvec_fn
) {
751 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
753 bvm
->bi_bdev
= rdev
->bdev
;
754 max
= min(max
, q
->merge_bvec_fn(
765 * This routine returns the disk from which the requested read should
766 * be done. There is a per-array 'next expected sequential IO' sector
767 * number - if this matches on the next IO then we use the last disk.
768 * There is also a per-disk 'last know head position' sector that is
769 * maintained from IRQ contexts, both the normal and the resync IO
770 * completion handlers update this position correctly. If there is no
771 * perfect sequential match then we pick the disk whose head is closest.
773 * If there are 2 mirrors in the same 2 devices, performance degrades
774 * because position is mirror, not device based.
776 * The rdev for the device selected will have nr_pending incremented.
780 * FIXME: possibly should rethink readbalancing and do it differently
781 * depending on near_copies / far_copies geometry.
783 static struct md_rdev
*read_balance(struct r10conf
*conf
,
784 struct r10bio
*r10_bio
,
787 const sector_t this_sector
= r10_bio
->sector
;
789 int sectors
= r10_bio
->sectors
;
790 int best_good_sectors
;
791 sector_t new_distance
, best_dist
;
792 struct md_rdev
*best_rdev
, *rdev
= NULL
;
795 struct geom
*geo
= &conf
->geo
;
797 raid10_find_phys(conf
, r10_bio
);
800 sectors
= r10_bio
->sectors
;
803 best_dist
= MaxSector
;
804 best_good_sectors
= 0;
807 * Check if we can balance. We can balance on the whole
808 * device if no resync is going on (recovery is ok), or below
809 * the resync window. We take the first readable disk when
810 * above the resync window.
812 if (conf
->mddev
->recovery_cp
< MaxSector
813 && (this_sector
+ sectors
>= conf
->next_resync
))
816 for (slot
= 0; slot
< conf
->copies
; slot
++) {
821 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
823 disk
= r10_bio
->devs
[slot
].devnum
;
824 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
825 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
826 test_bit(Unmerged
, &rdev
->flags
) ||
827 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
828 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
830 test_bit(Faulty
, &rdev
->flags
) ||
831 test_bit(Unmerged
, &rdev
->flags
))
833 if (!test_bit(In_sync
, &rdev
->flags
) &&
834 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
837 dev_sector
= r10_bio
->devs
[slot
].addr
;
838 if (is_badblock(rdev
, dev_sector
, sectors
,
839 &first_bad
, &bad_sectors
)) {
840 if (best_dist
< MaxSector
)
841 /* Already have a better slot */
843 if (first_bad
<= dev_sector
) {
844 /* Cannot read here. If this is the
845 * 'primary' device, then we must not read
846 * beyond 'bad_sectors' from another device.
848 bad_sectors
-= (dev_sector
- first_bad
);
849 if (!do_balance
&& sectors
> bad_sectors
)
850 sectors
= bad_sectors
;
851 if (best_good_sectors
> sectors
)
852 best_good_sectors
= sectors
;
854 sector_t good_sectors
=
855 first_bad
- dev_sector
;
856 if (good_sectors
> best_good_sectors
) {
857 best_good_sectors
= good_sectors
;
862 /* Must read from here */
867 best_good_sectors
= sectors
;
872 /* This optimisation is debatable, and completely destroys
873 * sequential read speed for 'far copies' arrays. So only
874 * keep it for 'near' arrays, and review those later.
876 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
879 /* for far > 1 always use the lowest address */
880 if (geo
->far_copies
> 1)
881 new_distance
= r10_bio
->devs
[slot
].addr
;
883 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
884 conf
->mirrors
[disk
].head_position
);
885 if (new_distance
< best_dist
) {
886 best_dist
= new_distance
;
891 if (slot
>= conf
->copies
) {
897 atomic_inc(&rdev
->nr_pending
);
898 if (test_bit(Faulty
, &rdev
->flags
)) {
899 /* Cannot risk returning a device that failed
900 * before we inc'ed nr_pending
902 rdev_dec_pending(rdev
, conf
->mddev
);
905 r10_bio
->read_slot
= slot
;
909 *max_sectors
= best_good_sectors
;
914 int md_raid10_congested(struct mddev
*mddev
, int bits
)
916 struct r10conf
*conf
= mddev
->private;
919 if ((bits
& (1 << BDI_async_congested
)) &&
920 conf
->pending_count
>= max_queued_requests
)
925 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
928 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
929 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
930 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
932 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
938 EXPORT_SYMBOL_GPL(md_raid10_congested
);
940 static int raid10_congested(void *data
, int bits
)
942 struct mddev
*mddev
= data
;
944 return mddev_congested(mddev
, bits
) ||
945 md_raid10_congested(mddev
, bits
);
948 static void flush_pending_writes(struct r10conf
*conf
)
950 /* Any writes that have been queued but are awaiting
951 * bitmap updates get flushed here.
953 spin_lock_irq(&conf
->device_lock
);
955 if (conf
->pending_bio_list
.head
) {
957 bio
= bio_list_get(&conf
->pending_bio_list
);
958 conf
->pending_count
= 0;
959 spin_unlock_irq(&conf
->device_lock
);
960 /* flush any pending bitmap writes to disk
961 * before proceeding w/ I/O */
962 bitmap_unplug(conf
->mddev
->bitmap
);
963 wake_up(&conf
->wait_barrier
);
965 while (bio
) { /* submit pending writes */
966 struct bio
*next
= bio
->bi_next
;
968 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
969 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
973 generic_make_request(bio
);
977 spin_unlock_irq(&conf
->device_lock
);
981 * Sometimes we need to suspend IO while we do something else,
982 * either some resync/recovery, or reconfigure the array.
983 * To do this we raise a 'barrier'.
984 * The 'barrier' is a counter that can be raised multiple times
985 * to count how many activities are happening which preclude
987 * We can only raise the barrier if there is no pending IO.
988 * i.e. if nr_pending == 0.
989 * We choose only to raise the barrier if no-one is waiting for the
990 * barrier to go down. This means that as soon as an IO request
991 * is ready, no other operations which require a barrier will start
992 * until the IO request has had a chance.
994 * So: regular IO calls 'wait_barrier'. When that returns there
995 * is no backgroup IO happening, It must arrange to call
996 * allow_barrier when it has finished its IO.
997 * backgroup IO calls must call raise_barrier. Once that returns
998 * there is no normal IO happeing. It must arrange to call
999 * lower_barrier when the particular background IO completes.
1002 static void raise_barrier(struct r10conf
*conf
, int force
)
1004 BUG_ON(force
&& !conf
->barrier
);
1005 spin_lock_irq(&conf
->resync_lock
);
1007 /* Wait until no block IO is waiting (unless 'force') */
1008 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
1011 /* block any new IO from starting */
1014 /* Now wait for all pending IO to complete */
1015 wait_event_lock_irq(conf
->wait_barrier
,
1016 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
1019 spin_unlock_irq(&conf
->resync_lock
);
1022 static void lower_barrier(struct r10conf
*conf
)
1024 unsigned long flags
;
1025 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1027 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1028 wake_up(&conf
->wait_barrier
);
1031 static void wait_barrier(struct r10conf
*conf
)
1033 spin_lock_irq(&conf
->resync_lock
);
1034 if (conf
->barrier
) {
1036 /* Wait for the barrier to drop.
1037 * However if there are already pending
1038 * requests (preventing the barrier from
1039 * rising completely), and the
1040 * pre-process bio queue isn't empty,
1041 * then don't wait, as we need to empty
1042 * that queue to get the nr_pending
1045 wait_event_lock_irq(conf
->wait_barrier
,
1047 (conf
->nr_pending
&&
1048 current
->bio_list
&&
1049 !bio_list_empty(current
->bio_list
)),
1054 spin_unlock_irq(&conf
->resync_lock
);
1057 static void allow_barrier(struct r10conf
*conf
)
1059 unsigned long flags
;
1060 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1062 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1063 wake_up(&conf
->wait_barrier
);
1066 static void freeze_array(struct r10conf
*conf
, int extra
)
1068 /* stop syncio and normal IO and wait for everything to
1070 * We increment barrier and nr_waiting, and then
1071 * wait until nr_pending match nr_queued+extra
1072 * This is called in the context of one normal IO request
1073 * that has failed. Thus any sync request that might be pending
1074 * will be blocked by nr_pending, and we need to wait for
1075 * pending IO requests to complete or be queued for re-try.
1076 * Thus the number queued (nr_queued) plus this request (extra)
1077 * must match the number of pending IOs (nr_pending) before
1080 spin_lock_irq(&conf
->resync_lock
);
1083 wait_event_lock_irq_cmd(conf
->wait_barrier
,
1084 conf
->nr_pending
== conf
->nr_queued
+extra
,
1086 flush_pending_writes(conf
));
1088 spin_unlock_irq(&conf
->resync_lock
);
1091 static void unfreeze_array(struct r10conf
*conf
)
1093 /* reverse the effect of the freeze */
1094 spin_lock_irq(&conf
->resync_lock
);
1097 wake_up(&conf
->wait_barrier
);
1098 spin_unlock_irq(&conf
->resync_lock
);
1101 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1102 struct md_rdev
*rdev
)
1104 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1105 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1106 return rdev
->data_offset
;
1108 return rdev
->new_data_offset
;
1111 struct raid10_plug_cb
{
1112 struct blk_plug_cb cb
;
1113 struct bio_list pending
;
1117 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1119 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1121 struct mddev
*mddev
= plug
->cb
.data
;
1122 struct r10conf
*conf
= mddev
->private;
1125 if (from_schedule
|| current
->bio_list
) {
1126 spin_lock_irq(&conf
->device_lock
);
1127 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1128 conf
->pending_count
+= plug
->pending_cnt
;
1129 spin_unlock_irq(&conf
->device_lock
);
1130 wake_up(&conf
->wait_barrier
);
1131 md_wakeup_thread(mddev
->thread
);
1136 /* we aren't scheduling, so we can do the write-out directly. */
1137 bio
= bio_list_get(&plug
->pending
);
1138 bitmap_unplug(mddev
->bitmap
);
1139 wake_up(&conf
->wait_barrier
);
1141 while (bio
) { /* submit pending writes */
1142 struct bio
*next
= bio
->bi_next
;
1143 bio
->bi_next
= NULL
;
1144 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
1145 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1146 /* Just ignore it */
1149 generic_make_request(bio
);
1155 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1157 struct r10conf
*conf
= mddev
->private;
1158 struct r10bio
*r10_bio
;
1159 struct bio
*read_bio
;
1161 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1162 int chunk_sects
= chunk_mask
+ 1;
1163 const int rw
= bio_data_dir(bio
);
1164 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1165 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1166 const unsigned long do_discard
= (bio
->bi_rw
1167 & (REQ_DISCARD
| REQ_SECURE
));
1168 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1169 unsigned long flags
;
1170 struct md_rdev
*blocked_rdev
;
1171 struct blk_plug_cb
*cb
;
1172 struct raid10_plug_cb
*plug
= NULL
;
1173 int sectors_handled
;
1177 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1178 md_flush_request(mddev
, bio
);
1182 /* If this request crosses a chunk boundary, we need to
1183 * split it. This will only happen for 1 PAGE (or less) requests.
1185 if (unlikely((bio
->bi_sector
& chunk_mask
) + bio_sectors(bio
)
1187 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1188 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1189 struct bio_pair
*bp
;
1190 /* Sanity check -- queue functions should prevent this happening */
1191 if (bio_segments(bio
) > 1)
1193 /* This is a one page bio that upper layers
1194 * refuse to split for us, so we need to split it.
1197 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
1199 /* Each of these 'make_request' calls will call 'wait_barrier'.
1200 * If the first succeeds but the second blocks due to the resync
1201 * thread raising the barrier, we will deadlock because the
1202 * IO to the underlying device will be queued in generic_make_request
1203 * and will never complete, so will never reduce nr_pending.
1204 * So increment nr_waiting here so no new raise_barriers will
1205 * succeed, and so the second wait_barrier cannot block.
1207 spin_lock_irq(&conf
->resync_lock
);
1209 spin_unlock_irq(&conf
->resync_lock
);
1211 make_request(mddev
, &bp
->bio1
);
1212 make_request(mddev
, &bp
->bio2
);
1214 spin_lock_irq(&conf
->resync_lock
);
1216 wake_up(&conf
->wait_barrier
);
1217 spin_unlock_irq(&conf
->resync_lock
);
1219 bio_pair_release(bp
);
1222 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1223 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1224 (unsigned long long)bio
->bi_sector
, bio_sectors(bio
) / 2);
1230 md_write_start(mddev
, bio
);
1233 * Register the new request and wait if the reconstruction
1234 * thread has put up a bar for new requests.
1235 * Continue immediately if no resync is active currently.
1239 sectors
= bio_sectors(bio
);
1240 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1241 bio
->bi_sector
< conf
->reshape_progress
&&
1242 bio
->bi_sector
+ sectors
> conf
->reshape_progress
) {
1243 /* IO spans the reshape position. Need to wait for
1246 allow_barrier(conf
);
1247 wait_event(conf
->wait_barrier
,
1248 conf
->reshape_progress
<= bio
->bi_sector
||
1249 conf
->reshape_progress
>= bio
->bi_sector
+ sectors
);
1252 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1253 bio_data_dir(bio
) == WRITE
&&
1254 (mddev
->reshape_backwards
1255 ? (bio
->bi_sector
< conf
->reshape_safe
&&
1256 bio
->bi_sector
+ sectors
> conf
->reshape_progress
)
1257 : (bio
->bi_sector
+ sectors
> conf
->reshape_safe
&&
1258 bio
->bi_sector
< conf
->reshape_progress
))) {
1259 /* Need to update reshape_position in metadata */
1260 mddev
->reshape_position
= conf
->reshape_progress
;
1261 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1262 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1263 md_wakeup_thread(mddev
->thread
);
1264 wait_event(mddev
->sb_wait
,
1265 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1267 conf
->reshape_safe
= mddev
->reshape_position
;
1270 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1272 r10_bio
->master_bio
= bio
;
1273 r10_bio
->sectors
= sectors
;
1275 r10_bio
->mddev
= mddev
;
1276 r10_bio
->sector
= bio
->bi_sector
;
1279 /* We might need to issue multiple reads to different
1280 * devices if there are bad blocks around, so we keep
1281 * track of the number of reads in bio->bi_phys_segments.
1282 * If this is 0, there is only one r10_bio and no locking
1283 * will be needed when the request completes. If it is
1284 * non-zero, then it is the number of not-completed requests.
1286 bio
->bi_phys_segments
= 0;
1287 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1291 * read balancing logic:
1293 struct md_rdev
*rdev
;
1297 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1299 raid_end_bio_io(r10_bio
);
1302 slot
= r10_bio
->read_slot
;
1304 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1305 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1308 r10_bio
->devs
[slot
].bio
= read_bio
;
1309 r10_bio
->devs
[slot
].rdev
= rdev
;
1311 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1312 choose_data_offset(r10_bio
, rdev
);
1313 read_bio
->bi_bdev
= rdev
->bdev
;
1314 read_bio
->bi_end_io
= raid10_end_read_request
;
1315 read_bio
->bi_rw
= READ
| do_sync
;
1316 read_bio
->bi_private
= r10_bio
;
1318 if (max_sectors
< r10_bio
->sectors
) {
1319 /* Could not read all from this device, so we will
1320 * need another r10_bio.
1322 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1324 r10_bio
->sectors
= max_sectors
;
1325 spin_lock_irq(&conf
->device_lock
);
1326 if (bio
->bi_phys_segments
== 0)
1327 bio
->bi_phys_segments
= 2;
1329 bio
->bi_phys_segments
++;
1330 spin_unlock(&conf
->device_lock
);
1331 /* Cannot call generic_make_request directly
1332 * as that will be queued in __generic_make_request
1333 * and subsequent mempool_alloc might block
1334 * waiting for it. so hand bio over to raid10d.
1336 reschedule_retry(r10_bio
);
1338 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1340 r10_bio
->master_bio
= bio
;
1341 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1343 r10_bio
->mddev
= mddev
;
1344 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1347 generic_make_request(read_bio
);
1354 if (conf
->pending_count
>= max_queued_requests
) {
1355 md_wakeup_thread(mddev
->thread
);
1356 wait_event(conf
->wait_barrier
,
1357 conf
->pending_count
< max_queued_requests
);
1359 /* first select target devices under rcu_lock and
1360 * inc refcount on their rdev. Record them by setting
1362 * If there are known/acknowledged bad blocks on any device
1363 * on which we have seen a write error, we want to avoid
1364 * writing to those blocks. This potentially requires several
1365 * writes to write around the bad blocks. Each set of writes
1366 * gets its own r10_bio with a set of bios attached. The number
1367 * of r10_bios is recored in bio->bi_phys_segments just as with
1371 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1372 raid10_find_phys(conf
, r10_bio
);
1374 blocked_rdev
= NULL
;
1376 max_sectors
= r10_bio
->sectors
;
1378 for (i
= 0; i
< conf
->copies
; i
++) {
1379 int d
= r10_bio
->devs
[i
].devnum
;
1380 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1381 struct md_rdev
*rrdev
= rcu_dereference(
1382 conf
->mirrors
[d
].replacement
);
1385 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1386 atomic_inc(&rdev
->nr_pending
);
1387 blocked_rdev
= rdev
;
1390 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1391 atomic_inc(&rrdev
->nr_pending
);
1392 blocked_rdev
= rrdev
;
1395 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)
1396 || test_bit(Unmerged
, &rdev
->flags
)))
1398 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1399 || test_bit(Unmerged
, &rrdev
->flags
)))
1402 r10_bio
->devs
[i
].bio
= NULL
;
1403 r10_bio
->devs
[i
].repl_bio
= NULL
;
1405 if (!rdev
&& !rrdev
) {
1406 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1409 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1411 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1415 is_bad
= is_badblock(rdev
, dev_sector
,
1417 &first_bad
, &bad_sectors
);
1419 /* Mustn't write here until the bad block
1422 atomic_inc(&rdev
->nr_pending
);
1423 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1424 blocked_rdev
= rdev
;
1427 if (is_bad
&& first_bad
<= dev_sector
) {
1428 /* Cannot write here at all */
1429 bad_sectors
-= (dev_sector
- first_bad
);
1430 if (bad_sectors
< max_sectors
)
1431 /* Mustn't write more than bad_sectors
1432 * to other devices yet
1434 max_sectors
= bad_sectors
;
1435 /* We don't set R10BIO_Degraded as that
1436 * only applies if the disk is missing,
1437 * so it might be re-added, and we want to
1438 * know to recover this chunk.
1439 * In this case the device is here, and the
1440 * fact that this chunk is not in-sync is
1441 * recorded in the bad block log.
1446 int good_sectors
= first_bad
- dev_sector
;
1447 if (good_sectors
< max_sectors
)
1448 max_sectors
= good_sectors
;
1452 r10_bio
->devs
[i
].bio
= bio
;
1453 atomic_inc(&rdev
->nr_pending
);
1456 r10_bio
->devs
[i
].repl_bio
= bio
;
1457 atomic_inc(&rrdev
->nr_pending
);
1462 if (unlikely(blocked_rdev
)) {
1463 /* Have to wait for this device to get unblocked, then retry */
1467 for (j
= 0; j
< i
; j
++) {
1468 if (r10_bio
->devs
[j
].bio
) {
1469 d
= r10_bio
->devs
[j
].devnum
;
1470 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1472 if (r10_bio
->devs
[j
].repl_bio
) {
1473 struct md_rdev
*rdev
;
1474 d
= r10_bio
->devs
[j
].devnum
;
1475 rdev
= conf
->mirrors
[d
].replacement
;
1477 /* Race with remove_disk */
1479 rdev
= conf
->mirrors
[d
].rdev
;
1481 rdev_dec_pending(rdev
, mddev
);
1484 allow_barrier(conf
);
1485 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1490 if (max_sectors
< r10_bio
->sectors
) {
1491 /* We are splitting this into multiple parts, so
1492 * we need to prepare for allocating another r10_bio.
1494 r10_bio
->sectors
= max_sectors
;
1495 spin_lock_irq(&conf
->device_lock
);
1496 if (bio
->bi_phys_segments
== 0)
1497 bio
->bi_phys_segments
= 2;
1499 bio
->bi_phys_segments
++;
1500 spin_unlock_irq(&conf
->device_lock
);
1502 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1504 atomic_set(&r10_bio
->remaining
, 1);
1505 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1507 for (i
= 0; i
< conf
->copies
; i
++) {
1509 int d
= r10_bio
->devs
[i
].devnum
;
1510 if (r10_bio
->devs
[i
].bio
) {
1511 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1512 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1513 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1515 r10_bio
->devs
[i
].bio
= mbio
;
1517 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1518 choose_data_offset(r10_bio
,
1520 mbio
->bi_bdev
= rdev
->bdev
;
1521 mbio
->bi_end_io
= raid10_end_write_request
;
1523 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1524 mbio
->bi_private
= r10_bio
;
1526 atomic_inc(&r10_bio
->remaining
);
1528 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1531 plug
= container_of(cb
, struct raid10_plug_cb
,
1535 spin_lock_irqsave(&conf
->device_lock
, flags
);
1537 bio_list_add(&plug
->pending
, mbio
);
1538 plug
->pending_cnt
++;
1540 bio_list_add(&conf
->pending_bio_list
, mbio
);
1541 conf
->pending_count
++;
1543 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1545 md_wakeup_thread(mddev
->thread
);
1548 if (r10_bio
->devs
[i
].repl_bio
) {
1549 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1551 /* Replacement just got moved to main 'rdev' */
1553 rdev
= conf
->mirrors
[d
].rdev
;
1555 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1556 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1558 r10_bio
->devs
[i
].repl_bio
= mbio
;
1560 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1563 mbio
->bi_bdev
= rdev
->bdev
;
1564 mbio
->bi_end_io
= raid10_end_write_request
;
1566 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1567 mbio
->bi_private
= r10_bio
;
1569 atomic_inc(&r10_bio
->remaining
);
1570 spin_lock_irqsave(&conf
->device_lock
, flags
);
1571 bio_list_add(&conf
->pending_bio_list
, mbio
);
1572 conf
->pending_count
++;
1573 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1574 if (!mddev_check_plugged(mddev
))
1575 md_wakeup_thread(mddev
->thread
);
1579 /* Don't remove the bias on 'remaining' (one_write_done) until
1580 * after checking if we need to go around again.
1583 if (sectors_handled
< bio_sectors(bio
)) {
1584 one_write_done(r10_bio
);
1585 /* We need another r10_bio. It has already been counted
1586 * in bio->bi_phys_segments.
1588 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1590 r10_bio
->master_bio
= bio
;
1591 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1593 r10_bio
->mddev
= mddev
;
1594 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1598 one_write_done(r10_bio
);
1600 /* In case raid10d snuck in to freeze_array */
1601 wake_up(&conf
->wait_barrier
);
1604 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1606 struct r10conf
*conf
= mddev
->private;
1609 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1610 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1611 if (conf
->geo
.near_copies
> 1)
1612 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1613 if (conf
->geo
.far_copies
> 1) {
1614 if (conf
->geo
.far_offset
)
1615 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1617 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1619 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1620 conf
->geo
.raid_disks
- mddev
->degraded
);
1621 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1622 seq_printf(seq
, "%s",
1623 conf
->mirrors
[i
].rdev
&&
1624 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1625 seq_printf(seq
, "]");
1628 /* check if there are enough drives for
1629 * every block to appear on atleast one.
1630 * Don't consider the device numbered 'ignore'
1631 * as we might be about to remove it.
1633 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1639 disks
= conf
->prev
.raid_disks
;
1640 ncopies
= conf
->prev
.near_copies
;
1642 disks
= conf
->geo
.raid_disks
;
1643 ncopies
= conf
->geo
.near_copies
;
1648 int n
= conf
->copies
;
1652 struct md_rdev
*rdev
;
1653 if (this != ignore
&&
1654 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1655 test_bit(In_sync
, &rdev
->flags
))
1657 this = (this+1) % disks
;
1661 first
= (first
+ ncopies
) % disks
;
1662 } while (first
!= 0);
1669 static int enough(struct r10conf
*conf
, int ignore
)
1671 /* when calling 'enough', both 'prev' and 'geo' must
1673 * This is ensured if ->reconfig_mutex or ->device_lock
1676 return _enough(conf
, 0, ignore
) &&
1677 _enough(conf
, 1, ignore
);
1680 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1682 char b
[BDEVNAME_SIZE
];
1683 struct r10conf
*conf
= mddev
->private;
1684 unsigned long flags
;
1687 * If it is not operational, then we have already marked it as dead
1688 * else if it is the last working disks, ignore the error, let the
1689 * next level up know.
1690 * else mark the drive as failed
1692 spin_lock_irqsave(&conf
->device_lock
, flags
);
1693 if (test_bit(In_sync
, &rdev
->flags
)
1694 && !enough(conf
, rdev
->raid_disk
)) {
1696 * Don't fail the drive, just return an IO error.
1698 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1701 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1704 * if recovery is running, make sure it aborts.
1706 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1708 set_bit(Blocked
, &rdev
->flags
);
1709 set_bit(Faulty
, &rdev
->flags
);
1710 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1711 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1713 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1714 "md/raid10:%s: Operation continuing on %d devices.\n",
1715 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1716 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1719 static void print_conf(struct r10conf
*conf
)
1722 struct raid10_info
*tmp
;
1724 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1726 printk(KERN_DEBUG
"(!conf)\n");
1729 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1730 conf
->geo
.raid_disks
);
1732 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1733 char b
[BDEVNAME_SIZE
];
1734 tmp
= conf
->mirrors
+ i
;
1736 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1737 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1738 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1739 bdevname(tmp
->rdev
->bdev
,b
));
1743 static void close_sync(struct r10conf
*conf
)
1746 allow_barrier(conf
);
1748 mempool_destroy(conf
->r10buf_pool
);
1749 conf
->r10buf_pool
= NULL
;
1752 static int raid10_spare_active(struct mddev
*mddev
)
1755 struct r10conf
*conf
= mddev
->private;
1756 struct raid10_info
*tmp
;
1758 unsigned long flags
;
1761 * Find all non-in_sync disks within the RAID10 configuration
1762 * and mark them in_sync
1764 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1765 tmp
= conf
->mirrors
+ i
;
1766 if (tmp
->replacement
1767 && tmp
->replacement
->recovery_offset
== MaxSector
1768 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1769 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1770 /* Replacement has just become active */
1772 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1775 /* Replaced device not technically faulty,
1776 * but we need to be sure it gets removed
1777 * and never re-added.
1779 set_bit(Faulty
, &tmp
->rdev
->flags
);
1780 sysfs_notify_dirent_safe(
1781 tmp
->rdev
->sysfs_state
);
1783 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1784 } else if (tmp
->rdev
1785 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1786 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1788 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1791 spin_lock_irqsave(&conf
->device_lock
, flags
);
1792 mddev
->degraded
-= count
;
1793 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1800 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1802 struct r10conf
*conf
= mddev
->private;
1806 int last
= conf
->geo
.raid_disks
- 1;
1807 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1809 if (mddev
->recovery_cp
< MaxSector
)
1810 /* only hot-add to in-sync arrays, as recovery is
1811 * very different from resync
1814 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1817 if (rdev
->raid_disk
>= 0)
1818 first
= last
= rdev
->raid_disk
;
1820 if (q
->merge_bvec_fn
) {
1821 set_bit(Unmerged
, &rdev
->flags
);
1822 mddev
->merge_check_needed
= 1;
1825 if (rdev
->saved_raid_disk
>= first
&&
1826 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1827 mirror
= rdev
->saved_raid_disk
;
1830 for ( ; mirror
<= last
; mirror
++) {
1831 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1832 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1835 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1836 p
->replacement
!= NULL
)
1838 clear_bit(In_sync
, &rdev
->flags
);
1839 set_bit(Replacement
, &rdev
->flags
);
1840 rdev
->raid_disk
= mirror
;
1843 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1844 rdev
->data_offset
<< 9);
1846 rcu_assign_pointer(p
->replacement
, rdev
);
1851 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1852 rdev
->data_offset
<< 9);
1854 p
->head_position
= 0;
1855 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1856 rdev
->raid_disk
= mirror
;
1858 if (rdev
->saved_raid_disk
!= mirror
)
1860 rcu_assign_pointer(p
->rdev
, rdev
);
1863 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1864 /* Some requests might not have seen this new
1865 * merge_bvec_fn. We must wait for them to complete
1866 * before merging the device fully.
1867 * First we make sure any code which has tested
1868 * our function has submitted the request, then
1869 * we wait for all outstanding requests to complete.
1871 synchronize_sched();
1872 freeze_array(conf
, 0);
1873 unfreeze_array(conf
);
1874 clear_bit(Unmerged
, &rdev
->flags
);
1876 md_integrity_add_rdev(rdev
, mddev
);
1877 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1878 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1884 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1886 struct r10conf
*conf
= mddev
->private;
1888 int number
= rdev
->raid_disk
;
1889 struct md_rdev
**rdevp
;
1890 struct raid10_info
*p
= conf
->mirrors
+ number
;
1893 if (rdev
== p
->rdev
)
1895 else if (rdev
== p
->replacement
)
1896 rdevp
= &p
->replacement
;
1900 if (test_bit(In_sync
, &rdev
->flags
) ||
1901 atomic_read(&rdev
->nr_pending
)) {
1905 /* Only remove faulty devices if recovery
1908 if (!test_bit(Faulty
, &rdev
->flags
) &&
1909 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1910 (!p
->replacement
|| p
->replacement
== rdev
) &&
1911 number
< conf
->geo
.raid_disks
&&
1918 if (atomic_read(&rdev
->nr_pending
)) {
1919 /* lost the race, try later */
1923 } else if (p
->replacement
) {
1924 /* We must have just cleared 'rdev' */
1925 p
->rdev
= p
->replacement
;
1926 clear_bit(Replacement
, &p
->replacement
->flags
);
1927 smp_mb(); /* Make sure other CPUs may see both as identical
1928 * but will never see neither -- if they are careful.
1930 p
->replacement
= NULL
;
1931 clear_bit(WantReplacement
, &rdev
->flags
);
1933 /* We might have just remove the Replacement as faulty
1934 * Clear the flag just in case
1936 clear_bit(WantReplacement
, &rdev
->flags
);
1938 err
= md_integrity_register(mddev
);
1947 static void end_sync_read(struct bio
*bio
, int error
)
1949 struct r10bio
*r10_bio
= bio
->bi_private
;
1950 struct r10conf
*conf
= r10_bio
->mddev
->private;
1953 if (bio
== r10_bio
->master_bio
) {
1954 /* this is a reshape read */
1955 d
= r10_bio
->read_slot
; /* really the read dev */
1957 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1959 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1960 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1962 /* The write handler will notice the lack of
1963 * R10BIO_Uptodate and record any errors etc
1965 atomic_add(r10_bio
->sectors
,
1966 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1968 /* for reconstruct, we always reschedule after a read.
1969 * for resync, only after all reads
1971 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1972 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1973 atomic_dec_and_test(&r10_bio
->remaining
)) {
1974 /* we have read all the blocks,
1975 * do the comparison in process context in raid10d
1977 reschedule_retry(r10_bio
);
1981 static void end_sync_request(struct r10bio
*r10_bio
)
1983 struct mddev
*mddev
= r10_bio
->mddev
;
1985 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1986 if (r10_bio
->master_bio
== NULL
) {
1987 /* the primary of several recovery bios */
1988 sector_t s
= r10_bio
->sectors
;
1989 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1990 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1991 reschedule_retry(r10_bio
);
1994 md_done_sync(mddev
, s
, 1);
1997 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1998 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1999 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2000 reschedule_retry(r10_bio
);
2008 static void end_sync_write(struct bio
*bio
, int error
)
2010 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2011 struct r10bio
*r10_bio
= bio
->bi_private
;
2012 struct mddev
*mddev
= r10_bio
->mddev
;
2013 struct r10conf
*conf
= mddev
->private;
2019 struct md_rdev
*rdev
= NULL
;
2021 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
2023 rdev
= conf
->mirrors
[d
].replacement
;
2025 rdev
= conf
->mirrors
[d
].rdev
;
2029 md_error(mddev
, rdev
);
2031 set_bit(WriteErrorSeen
, &rdev
->flags
);
2032 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2033 set_bit(MD_RECOVERY_NEEDED
,
2034 &rdev
->mddev
->recovery
);
2035 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
2037 } else if (is_badblock(rdev
,
2038 r10_bio
->devs
[slot
].addr
,
2040 &first_bad
, &bad_sectors
))
2041 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
2043 rdev_dec_pending(rdev
, mddev
);
2045 end_sync_request(r10_bio
);
2049 * Note: sync and recover and handled very differently for raid10
2050 * This code is for resync.
2051 * For resync, we read through virtual addresses and read all blocks.
2052 * If there is any error, we schedule a write. The lowest numbered
2053 * drive is authoritative.
2054 * However requests come for physical address, so we need to map.
2055 * For every physical address there are raid_disks/copies virtual addresses,
2056 * which is always are least one, but is not necessarly an integer.
2057 * This means that a physical address can span multiple chunks, so we may
2058 * have to submit multiple io requests for a single sync request.
2061 * We check if all blocks are in-sync and only write to blocks that
2064 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2066 struct r10conf
*conf
= mddev
->private;
2068 struct bio
*tbio
, *fbio
;
2071 atomic_set(&r10_bio
->remaining
, 1);
2073 /* find the first device with a block */
2074 for (i
=0; i
<conf
->copies
; i
++)
2075 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
2078 if (i
== conf
->copies
)
2082 fbio
= r10_bio
->devs
[i
].bio
;
2084 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2085 /* now find blocks with errors */
2086 for (i
=0 ; i
< conf
->copies
; i
++) {
2089 tbio
= r10_bio
->devs
[i
].bio
;
2091 if (tbio
->bi_end_io
!= end_sync_read
)
2095 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2096 /* We know that the bi_io_vec layout is the same for
2097 * both 'first' and 'i', so we just compare them.
2098 * All vec entries are PAGE_SIZE;
2100 for (j
= 0; j
< vcnt
; j
++)
2101 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2102 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2103 fbio
->bi_io_vec
[j
].bv_len
))
2107 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2108 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2109 /* Don't fix anything. */
2112 /* Ok, we need to write this bio, either to correct an
2113 * inconsistency or to correct an unreadable block.
2114 * First we need to fixup bv_offset, bv_len and
2115 * bi_vecs, as the read request might have corrupted these
2119 tbio
->bi_vcnt
= vcnt
;
2120 tbio
->bi_size
= r10_bio
->sectors
<< 9;
2121 tbio
->bi_rw
= WRITE
;
2122 tbio
->bi_private
= r10_bio
;
2123 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
2125 for (j
=0; j
< vcnt
; j
++) {
2126 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2127 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2129 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2130 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2133 tbio
->bi_end_io
= end_sync_write
;
2135 d
= r10_bio
->devs
[i
].devnum
;
2136 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2137 atomic_inc(&r10_bio
->remaining
);
2138 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2140 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2141 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2142 generic_make_request(tbio
);
2145 /* Now write out to any replacement devices
2148 for (i
= 0; i
< conf
->copies
; i
++) {
2151 tbio
= r10_bio
->devs
[i
].repl_bio
;
2152 if (!tbio
|| !tbio
->bi_end_io
)
2154 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2155 && r10_bio
->devs
[i
].bio
!= fbio
)
2156 for (j
= 0; j
< vcnt
; j
++)
2157 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2158 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2160 d
= r10_bio
->devs
[i
].devnum
;
2161 atomic_inc(&r10_bio
->remaining
);
2162 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2164 generic_make_request(tbio
);
2168 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2169 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2175 * Now for the recovery code.
2176 * Recovery happens across physical sectors.
2177 * We recover all non-is_sync drives by finding the virtual address of
2178 * each, and then choose a working drive that also has that virt address.
2179 * There is a separate r10_bio for each non-in_sync drive.
2180 * Only the first two slots are in use. The first for reading,
2181 * The second for writing.
2184 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2186 /* We got a read error during recovery.
2187 * We repeat the read in smaller page-sized sections.
2188 * If a read succeeds, write it to the new device or record
2189 * a bad block if we cannot.
2190 * If a read fails, record a bad block on both old and
2193 struct mddev
*mddev
= r10_bio
->mddev
;
2194 struct r10conf
*conf
= mddev
->private;
2195 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2197 int sectors
= r10_bio
->sectors
;
2199 int dr
= r10_bio
->devs
[0].devnum
;
2200 int dw
= r10_bio
->devs
[1].devnum
;
2204 struct md_rdev
*rdev
;
2208 if (s
> (PAGE_SIZE
>>9))
2211 rdev
= conf
->mirrors
[dr
].rdev
;
2212 addr
= r10_bio
->devs
[0].addr
+ sect
,
2213 ok
= sync_page_io(rdev
,
2216 bio
->bi_io_vec
[idx
].bv_page
,
2219 rdev
= conf
->mirrors
[dw
].rdev
;
2220 addr
= r10_bio
->devs
[1].addr
+ sect
;
2221 ok
= sync_page_io(rdev
,
2224 bio
->bi_io_vec
[idx
].bv_page
,
2227 set_bit(WriteErrorSeen
, &rdev
->flags
);
2228 if (!test_and_set_bit(WantReplacement
,
2230 set_bit(MD_RECOVERY_NEEDED
,
2231 &rdev
->mddev
->recovery
);
2235 /* We don't worry if we cannot set a bad block -
2236 * it really is bad so there is no loss in not
2239 rdev_set_badblocks(rdev
, addr
, s
, 0);
2241 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2242 /* need bad block on destination too */
2243 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2244 addr
= r10_bio
->devs
[1].addr
+ sect
;
2245 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2247 /* just abort the recovery */
2249 "md/raid10:%s: recovery aborted"
2250 " due to read error\n",
2253 conf
->mirrors
[dw
].recovery_disabled
2254 = mddev
->recovery_disabled
;
2255 set_bit(MD_RECOVERY_INTR
,
2268 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2270 struct r10conf
*conf
= mddev
->private;
2272 struct bio
*wbio
, *wbio2
;
2274 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2275 fix_recovery_read_error(r10_bio
);
2276 end_sync_request(r10_bio
);
2281 * share the pages with the first bio
2282 * and submit the write request
2284 d
= r10_bio
->devs
[1].devnum
;
2285 wbio
= r10_bio
->devs
[1].bio
;
2286 wbio2
= r10_bio
->devs
[1].repl_bio
;
2287 if (wbio
->bi_end_io
) {
2288 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2289 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2290 generic_make_request(wbio
);
2292 if (wbio2
&& wbio2
->bi_end_io
) {
2293 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2294 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2295 bio_sectors(wbio2
));
2296 generic_make_request(wbio2
);
2302 * Used by fix_read_error() to decay the per rdev read_errors.
2303 * We halve the read error count for every hour that has elapsed
2304 * since the last recorded read error.
2307 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2309 struct timespec cur_time_mon
;
2310 unsigned long hours_since_last
;
2311 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2313 ktime_get_ts(&cur_time_mon
);
2315 if (rdev
->last_read_error
.tv_sec
== 0 &&
2316 rdev
->last_read_error
.tv_nsec
== 0) {
2317 /* first time we've seen a read error */
2318 rdev
->last_read_error
= cur_time_mon
;
2322 hours_since_last
= (cur_time_mon
.tv_sec
-
2323 rdev
->last_read_error
.tv_sec
) / 3600;
2325 rdev
->last_read_error
= cur_time_mon
;
2328 * if hours_since_last is > the number of bits in read_errors
2329 * just set read errors to 0. We do this to avoid
2330 * overflowing the shift of read_errors by hours_since_last.
2332 if (hours_since_last
>= 8 * sizeof(read_errors
))
2333 atomic_set(&rdev
->read_errors
, 0);
2335 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2338 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2339 int sectors
, struct page
*page
, int rw
)
2344 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2345 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2347 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2351 set_bit(WriteErrorSeen
, &rdev
->flags
);
2352 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2353 set_bit(MD_RECOVERY_NEEDED
,
2354 &rdev
->mddev
->recovery
);
2356 /* need to record an error - either for the block or the device */
2357 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2358 md_error(rdev
->mddev
, rdev
);
2363 * This is a kernel thread which:
2365 * 1. Retries failed read operations on working mirrors.
2366 * 2. Updates the raid superblock when problems encounter.
2367 * 3. Performs writes following reads for array synchronising.
2370 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2372 int sect
= 0; /* Offset from r10_bio->sector */
2373 int sectors
= r10_bio
->sectors
;
2374 struct md_rdev
*rdev
;
2375 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2376 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2378 /* still own a reference to this rdev, so it cannot
2379 * have been cleared recently.
2381 rdev
= conf
->mirrors
[d
].rdev
;
2383 if (test_bit(Faulty
, &rdev
->flags
))
2384 /* drive has already been failed, just ignore any
2385 more fix_read_error() attempts */
2388 check_decay_read_errors(mddev
, rdev
);
2389 atomic_inc(&rdev
->read_errors
);
2390 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2391 char b
[BDEVNAME_SIZE
];
2392 bdevname(rdev
->bdev
, b
);
2395 "md/raid10:%s: %s: Raid device exceeded "
2396 "read_error threshold [cur %d:max %d]\n",
2398 atomic_read(&rdev
->read_errors
), max_read_errors
);
2400 "md/raid10:%s: %s: Failing raid device\n",
2402 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2403 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2409 int sl
= r10_bio
->read_slot
;
2413 if (s
> (PAGE_SIZE
>>9))
2421 d
= r10_bio
->devs
[sl
].devnum
;
2422 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2424 !test_bit(Unmerged
, &rdev
->flags
) &&
2425 test_bit(In_sync
, &rdev
->flags
) &&
2426 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2427 &first_bad
, &bad_sectors
) == 0) {
2428 atomic_inc(&rdev
->nr_pending
);
2430 success
= sync_page_io(rdev
,
2431 r10_bio
->devs
[sl
].addr
+
2434 conf
->tmppage
, READ
, false);
2435 rdev_dec_pending(rdev
, mddev
);
2441 if (sl
== conf
->copies
)
2443 } while (!success
&& sl
!= r10_bio
->read_slot
);
2447 /* Cannot read from anywhere, just mark the block
2448 * as bad on the first device to discourage future
2451 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2452 rdev
= conf
->mirrors
[dn
].rdev
;
2454 if (!rdev_set_badblocks(
2456 r10_bio
->devs
[r10_bio
->read_slot
].addr
2459 md_error(mddev
, rdev
);
2460 r10_bio
->devs
[r10_bio
->read_slot
].bio
2467 /* write it back and re-read */
2469 while (sl
!= r10_bio
->read_slot
) {
2470 char b
[BDEVNAME_SIZE
];
2475 d
= r10_bio
->devs
[sl
].devnum
;
2476 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2478 test_bit(Unmerged
, &rdev
->flags
) ||
2479 !test_bit(In_sync
, &rdev
->flags
))
2482 atomic_inc(&rdev
->nr_pending
);
2484 if (r10_sync_page_io(rdev
,
2485 r10_bio
->devs
[sl
].addr
+
2487 s
, conf
->tmppage
, WRITE
)
2489 /* Well, this device is dead */
2491 "md/raid10:%s: read correction "
2493 " (%d sectors at %llu on %s)\n",
2495 (unsigned long long)(
2497 choose_data_offset(r10_bio
,
2499 bdevname(rdev
->bdev
, b
));
2500 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2503 bdevname(rdev
->bdev
, b
));
2505 rdev_dec_pending(rdev
, mddev
);
2509 while (sl
!= r10_bio
->read_slot
) {
2510 char b
[BDEVNAME_SIZE
];
2515 d
= r10_bio
->devs
[sl
].devnum
;
2516 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2518 !test_bit(In_sync
, &rdev
->flags
))
2521 atomic_inc(&rdev
->nr_pending
);
2523 switch (r10_sync_page_io(rdev
,
2524 r10_bio
->devs
[sl
].addr
+
2529 /* Well, this device is dead */
2531 "md/raid10:%s: unable to read back "
2533 " (%d sectors at %llu on %s)\n",
2535 (unsigned long long)(
2537 choose_data_offset(r10_bio
, rdev
)),
2538 bdevname(rdev
->bdev
, b
));
2539 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2542 bdevname(rdev
->bdev
, b
));
2546 "md/raid10:%s: read error corrected"
2547 " (%d sectors at %llu on %s)\n",
2549 (unsigned long long)(
2551 choose_data_offset(r10_bio
, rdev
)),
2552 bdevname(rdev
->bdev
, b
));
2553 atomic_add(s
, &rdev
->corrected_errors
);
2556 rdev_dec_pending(rdev
, mddev
);
2566 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2568 struct bio
*bio
= r10_bio
->master_bio
;
2569 struct mddev
*mddev
= r10_bio
->mddev
;
2570 struct r10conf
*conf
= mddev
->private;
2571 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2572 /* bio has the data to be written to slot 'i' where
2573 * we just recently had a write error.
2574 * We repeatedly clone the bio and trim down to one block,
2575 * then try the write. Where the write fails we record
2577 * It is conceivable that the bio doesn't exactly align with
2578 * blocks. We must handle this.
2580 * We currently own a reference to the rdev.
2586 int sect_to_write
= r10_bio
->sectors
;
2589 if (rdev
->badblocks
.shift
< 0)
2592 block_sectors
= 1 << rdev
->badblocks
.shift
;
2593 sector
= r10_bio
->sector
;
2594 sectors
= ((r10_bio
->sector
+ block_sectors
)
2595 & ~(sector_t
)(block_sectors
- 1))
2598 while (sect_to_write
) {
2600 if (sectors
> sect_to_write
)
2601 sectors
= sect_to_write
;
2602 /* Write at 'sector' for 'sectors' */
2603 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2604 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2605 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2606 choose_data_offset(r10_bio
, rdev
) +
2607 (sector
- r10_bio
->sector
));
2608 wbio
->bi_bdev
= rdev
->bdev
;
2609 if (submit_bio_wait(WRITE
, wbio
) == 0)
2611 ok
= rdev_set_badblocks(rdev
, sector
,
2616 sect_to_write
-= sectors
;
2618 sectors
= block_sectors
;
2623 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2625 int slot
= r10_bio
->read_slot
;
2627 struct r10conf
*conf
= mddev
->private;
2628 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2629 char b
[BDEVNAME_SIZE
];
2630 unsigned long do_sync
;
2633 /* we got a read error. Maybe the drive is bad. Maybe just
2634 * the block and we can fix it.
2635 * We freeze all other IO, and try reading the block from
2636 * other devices. When we find one, we re-write
2637 * and check it that fixes the read error.
2638 * This is all done synchronously while the array is
2641 bio
= r10_bio
->devs
[slot
].bio
;
2642 bdevname(bio
->bi_bdev
, b
);
2644 r10_bio
->devs
[slot
].bio
= NULL
;
2646 if (mddev
->ro
== 0) {
2647 freeze_array(conf
, 1);
2648 fix_read_error(conf
, mddev
, r10_bio
);
2649 unfreeze_array(conf
);
2651 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2653 rdev_dec_pending(rdev
, mddev
);
2656 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2658 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2659 " read error for block %llu\n",
2661 (unsigned long long)r10_bio
->sector
);
2662 raid_end_bio_io(r10_bio
);
2666 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2667 slot
= r10_bio
->read_slot
;
2670 "md/raid10:%s: %s: redirecting "
2671 "sector %llu to another mirror\n",
2673 bdevname(rdev
->bdev
, b
),
2674 (unsigned long long)r10_bio
->sector
);
2675 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2678 r10_bio
->sector
- bio
->bi_sector
,
2680 r10_bio
->devs
[slot
].bio
= bio
;
2681 r10_bio
->devs
[slot
].rdev
= rdev
;
2682 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2683 + choose_data_offset(r10_bio
, rdev
);
2684 bio
->bi_bdev
= rdev
->bdev
;
2685 bio
->bi_rw
= READ
| do_sync
;
2686 bio
->bi_private
= r10_bio
;
2687 bio
->bi_end_io
= raid10_end_read_request
;
2688 if (max_sectors
< r10_bio
->sectors
) {
2689 /* Drat - have to split this up more */
2690 struct bio
*mbio
= r10_bio
->master_bio
;
2691 int sectors_handled
=
2692 r10_bio
->sector
+ max_sectors
2694 r10_bio
->sectors
= max_sectors
;
2695 spin_lock_irq(&conf
->device_lock
);
2696 if (mbio
->bi_phys_segments
== 0)
2697 mbio
->bi_phys_segments
= 2;
2699 mbio
->bi_phys_segments
++;
2700 spin_unlock_irq(&conf
->device_lock
);
2701 generic_make_request(bio
);
2703 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2705 r10_bio
->master_bio
= mbio
;
2706 r10_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2708 set_bit(R10BIO_ReadError
,
2710 r10_bio
->mddev
= mddev
;
2711 r10_bio
->sector
= mbio
->bi_sector
2716 generic_make_request(bio
);
2719 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2721 /* Some sort of write request has finished and it
2722 * succeeded in writing where we thought there was a
2723 * bad block. So forget the bad block.
2724 * Or possibly if failed and we need to record
2728 struct md_rdev
*rdev
;
2730 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2731 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2732 for (m
= 0; m
< conf
->copies
; m
++) {
2733 int dev
= r10_bio
->devs
[m
].devnum
;
2734 rdev
= conf
->mirrors
[dev
].rdev
;
2735 if (r10_bio
->devs
[m
].bio
== NULL
)
2737 if (test_bit(BIO_UPTODATE
,
2738 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2739 rdev_clear_badblocks(
2741 r10_bio
->devs
[m
].addr
,
2742 r10_bio
->sectors
, 0);
2744 if (!rdev_set_badblocks(
2746 r10_bio
->devs
[m
].addr
,
2747 r10_bio
->sectors
, 0))
2748 md_error(conf
->mddev
, rdev
);
2750 rdev
= conf
->mirrors
[dev
].replacement
;
2751 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2753 if (test_bit(BIO_UPTODATE
,
2754 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2755 rdev_clear_badblocks(
2757 r10_bio
->devs
[m
].addr
,
2758 r10_bio
->sectors
, 0);
2760 if (!rdev_set_badblocks(
2762 r10_bio
->devs
[m
].addr
,
2763 r10_bio
->sectors
, 0))
2764 md_error(conf
->mddev
, rdev
);
2769 for (m
= 0; m
< conf
->copies
; m
++) {
2770 int dev
= r10_bio
->devs
[m
].devnum
;
2771 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2772 rdev
= conf
->mirrors
[dev
].rdev
;
2773 if (bio
== IO_MADE_GOOD
) {
2774 rdev_clear_badblocks(
2776 r10_bio
->devs
[m
].addr
,
2777 r10_bio
->sectors
, 0);
2778 rdev_dec_pending(rdev
, conf
->mddev
);
2779 } else if (bio
!= NULL
&&
2780 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2781 if (!narrow_write_error(r10_bio
, m
)) {
2782 md_error(conf
->mddev
, rdev
);
2783 set_bit(R10BIO_Degraded
,
2786 rdev_dec_pending(rdev
, conf
->mddev
);
2788 bio
= r10_bio
->devs
[m
].repl_bio
;
2789 rdev
= conf
->mirrors
[dev
].replacement
;
2790 if (rdev
&& bio
== IO_MADE_GOOD
) {
2791 rdev_clear_badblocks(
2793 r10_bio
->devs
[m
].addr
,
2794 r10_bio
->sectors
, 0);
2795 rdev_dec_pending(rdev
, conf
->mddev
);
2798 if (test_bit(R10BIO_WriteError
,
2800 close_write(r10_bio
);
2801 raid_end_bio_io(r10_bio
);
2805 static void raid10d(struct md_thread
*thread
)
2807 struct mddev
*mddev
= thread
->mddev
;
2808 struct r10bio
*r10_bio
;
2809 unsigned long flags
;
2810 struct r10conf
*conf
= mddev
->private;
2811 struct list_head
*head
= &conf
->retry_list
;
2812 struct blk_plug plug
;
2814 md_check_recovery(mddev
);
2816 blk_start_plug(&plug
);
2819 flush_pending_writes(conf
);
2821 spin_lock_irqsave(&conf
->device_lock
, flags
);
2822 if (list_empty(head
)) {
2823 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2826 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2827 list_del(head
->prev
);
2829 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2831 mddev
= r10_bio
->mddev
;
2832 conf
= mddev
->private;
2833 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2834 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2835 handle_write_completed(conf
, r10_bio
);
2836 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2837 reshape_request_write(mddev
, r10_bio
);
2838 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2839 sync_request_write(mddev
, r10_bio
);
2840 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2841 recovery_request_write(mddev
, r10_bio
);
2842 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2843 handle_read_error(mddev
, r10_bio
);
2845 /* just a partial read to be scheduled from a
2848 int slot
= r10_bio
->read_slot
;
2849 generic_make_request(r10_bio
->devs
[slot
].bio
);
2853 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2854 md_check_recovery(mddev
);
2856 blk_finish_plug(&plug
);
2860 static int init_resync(struct r10conf
*conf
)
2865 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2866 BUG_ON(conf
->r10buf_pool
);
2867 conf
->have_replacement
= 0;
2868 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2869 if (conf
->mirrors
[i
].replacement
)
2870 conf
->have_replacement
= 1;
2871 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2872 if (!conf
->r10buf_pool
)
2874 conf
->next_resync
= 0;
2879 * perform a "sync" on one "block"
2881 * We need to make sure that no normal I/O request - particularly write
2882 * requests - conflict with active sync requests.
2884 * This is achieved by tracking pending requests and a 'barrier' concept
2885 * that can be installed to exclude normal IO requests.
2887 * Resync and recovery are handled very differently.
2888 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2890 * For resync, we iterate over virtual addresses, read all copies,
2891 * and update if there are differences. If only one copy is live,
2893 * For recovery, we iterate over physical addresses, read a good
2894 * value for each non-in_sync drive, and over-write.
2896 * So, for recovery we may have several outstanding complex requests for a
2897 * given address, one for each out-of-sync device. We model this by allocating
2898 * a number of r10_bio structures, one for each out-of-sync device.
2899 * As we setup these structures, we collect all bio's together into a list
2900 * which we then process collectively to add pages, and then process again
2901 * to pass to generic_make_request.
2903 * The r10_bio structures are linked using a borrowed master_bio pointer.
2904 * This link is counted in ->remaining. When the r10_bio that points to NULL
2905 * has its remaining count decremented to 0, the whole complex operation
2910 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2911 int *skipped
, int go_faster
)
2913 struct r10conf
*conf
= mddev
->private;
2914 struct r10bio
*r10_bio
;
2915 struct bio
*biolist
= NULL
, *bio
;
2916 sector_t max_sector
, nr_sectors
;
2919 sector_t sync_blocks
;
2920 sector_t sectors_skipped
= 0;
2921 int chunks_skipped
= 0;
2922 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2924 if (!conf
->r10buf_pool
)
2925 if (init_resync(conf
))
2929 * Allow skipping a full rebuild for incremental assembly
2930 * of a clean array, like RAID1 does.
2932 if (mddev
->bitmap
== NULL
&&
2933 mddev
->recovery_cp
== MaxSector
&&
2934 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2935 conf
->fullsync
== 0) {
2937 max_sector
= mddev
->dev_sectors
;
2938 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2939 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2940 max_sector
= mddev
->resync_max_sectors
;
2941 return max_sector
- sector_nr
;
2945 max_sector
= mddev
->dev_sectors
;
2946 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2947 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2948 max_sector
= mddev
->resync_max_sectors
;
2949 if (sector_nr
>= max_sector
) {
2950 /* If we aborted, we need to abort the
2951 * sync on the 'current' bitmap chucks (there can
2952 * be several when recovering multiple devices).
2953 * as we may have started syncing it but not finished.
2954 * We can find the current address in
2955 * mddev->curr_resync, but for recovery,
2956 * we need to convert that to several
2957 * virtual addresses.
2959 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2964 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2965 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2966 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2968 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2970 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2971 bitmap_end_sync(mddev
->bitmap
, sect
,
2975 /* completed sync */
2976 if ((!mddev
->bitmap
|| conf
->fullsync
)
2977 && conf
->have_replacement
2978 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2979 /* Completed a full sync so the replacements
2980 * are now fully recovered.
2982 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2983 if (conf
->mirrors
[i
].replacement
)
2984 conf
->mirrors
[i
].replacement
2990 bitmap_close_sync(mddev
->bitmap
);
2993 return sectors_skipped
;
2996 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2997 return reshape_request(mddev
, sector_nr
, skipped
);
2999 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
3000 /* if there has been nothing to do on any drive,
3001 * then there is nothing to do at all..
3004 return (max_sector
- sector_nr
) + sectors_skipped
;
3007 if (max_sector
> mddev
->resync_max
)
3008 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
3010 /* make sure whole request will fit in a chunk - if chunks
3013 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
3014 max_sector
> (sector_nr
| chunk_mask
))
3015 max_sector
= (sector_nr
| chunk_mask
) + 1;
3017 * If there is non-resync activity waiting for us then
3018 * put in a delay to throttle resync.
3020 if (!go_faster
&& conf
->nr_waiting
)
3021 msleep_interruptible(1000);
3023 /* Again, very different code for resync and recovery.
3024 * Both must result in an r10bio with a list of bios that
3025 * have bi_end_io, bi_sector, bi_bdev set,
3026 * and bi_private set to the r10bio.
3027 * For recovery, we may actually create several r10bios
3028 * with 2 bios in each, that correspond to the bios in the main one.
3029 * In this case, the subordinate r10bios link back through a
3030 * borrowed master_bio pointer, and the counter in the master
3031 * includes a ref from each subordinate.
3033 /* First, we decide what to do and set ->bi_end_io
3034 * To end_sync_read if we want to read, and
3035 * end_sync_write if we will want to write.
3038 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
3039 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3040 /* recovery... the complicated one */
3044 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
3050 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
3052 if ((mirror
->rdev
== NULL
||
3053 test_bit(In_sync
, &mirror
->rdev
->flags
))
3055 (mirror
->replacement
== NULL
||
3057 &mirror
->replacement
->flags
)))
3061 /* want to reconstruct this device */
3063 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3064 if (sect
>= mddev
->resync_max_sectors
) {
3065 /* last stripe is not complete - don't
3066 * try to recover this sector.
3070 /* Unless we are doing a full sync, or a replacement
3071 * we only need to recover the block if it is set in
3074 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3076 if (sync_blocks
< max_sync
)
3077 max_sync
= sync_blocks
;
3079 mirror
->replacement
== NULL
&&
3081 /* yep, skip the sync_blocks here, but don't assume
3082 * that there will never be anything to do here
3084 chunks_skipped
= -1;
3088 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3089 raise_barrier(conf
, rb2
!= NULL
);
3090 atomic_set(&r10_bio
->remaining
, 0);
3092 r10_bio
->master_bio
= (struct bio
*)rb2
;
3094 atomic_inc(&rb2
->remaining
);
3095 r10_bio
->mddev
= mddev
;
3096 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3097 r10_bio
->sector
= sect
;
3099 raid10_find_phys(conf
, r10_bio
);
3101 /* Need to check if the array will still be
3104 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3105 if (conf
->mirrors
[j
].rdev
== NULL
||
3106 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3111 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3112 &sync_blocks
, still_degraded
);
3115 for (j
=0; j
<conf
->copies
;j
++) {
3117 int d
= r10_bio
->devs
[j
].devnum
;
3118 sector_t from_addr
, to_addr
;
3119 struct md_rdev
*rdev
;
3120 sector_t sector
, first_bad
;
3122 if (!conf
->mirrors
[d
].rdev
||
3123 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3125 /* This is where we read from */
3127 rdev
= conf
->mirrors
[d
].rdev
;
3128 sector
= r10_bio
->devs
[j
].addr
;
3130 if (is_badblock(rdev
, sector
, max_sync
,
3131 &first_bad
, &bad_sectors
)) {
3132 if (first_bad
> sector
)
3133 max_sync
= first_bad
- sector
;
3135 bad_sectors
-= (sector
3137 if (max_sync
> bad_sectors
)
3138 max_sync
= bad_sectors
;
3142 bio
= r10_bio
->devs
[0].bio
;
3144 bio
->bi_next
= biolist
;
3146 bio
->bi_private
= r10_bio
;
3147 bio
->bi_end_io
= end_sync_read
;
3149 from_addr
= r10_bio
->devs
[j
].addr
;
3150 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
3151 bio
->bi_bdev
= rdev
->bdev
;
3152 atomic_inc(&rdev
->nr_pending
);
3153 /* and we write to 'i' (if not in_sync) */
3155 for (k
=0; k
<conf
->copies
; k
++)
3156 if (r10_bio
->devs
[k
].devnum
== i
)
3158 BUG_ON(k
== conf
->copies
);
3159 to_addr
= r10_bio
->devs
[k
].addr
;
3160 r10_bio
->devs
[0].devnum
= d
;
3161 r10_bio
->devs
[0].addr
= from_addr
;
3162 r10_bio
->devs
[1].devnum
= i
;
3163 r10_bio
->devs
[1].addr
= to_addr
;
3165 rdev
= mirror
->rdev
;
3166 if (!test_bit(In_sync
, &rdev
->flags
)) {
3167 bio
= r10_bio
->devs
[1].bio
;
3169 bio
->bi_next
= biolist
;
3171 bio
->bi_private
= r10_bio
;
3172 bio
->bi_end_io
= end_sync_write
;
3174 bio
->bi_sector
= to_addr
3175 + rdev
->data_offset
;
3176 bio
->bi_bdev
= rdev
->bdev
;
3177 atomic_inc(&r10_bio
->remaining
);
3179 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3181 /* and maybe write to replacement */
3182 bio
= r10_bio
->devs
[1].repl_bio
;
3184 bio
->bi_end_io
= NULL
;
3185 rdev
= mirror
->replacement
;
3186 /* Note: if rdev != NULL, then bio
3187 * cannot be NULL as r10buf_pool_alloc will
3188 * have allocated it.
3189 * So the second test here is pointless.
3190 * But it keeps semantic-checkers happy, and
3191 * this comment keeps human reviewers
3194 if (rdev
== NULL
|| bio
== NULL
||
3195 test_bit(Faulty
, &rdev
->flags
))
3198 bio
->bi_next
= biolist
;
3200 bio
->bi_private
= r10_bio
;
3201 bio
->bi_end_io
= end_sync_write
;
3203 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3204 bio
->bi_bdev
= rdev
->bdev
;
3205 atomic_inc(&r10_bio
->remaining
);
3208 if (j
== conf
->copies
) {
3209 /* Cannot recover, so abort the recovery or
3210 * record a bad block */
3213 atomic_dec(&rb2
->remaining
);
3216 /* problem is that there are bad blocks
3217 * on other device(s)
3220 for (k
= 0; k
< conf
->copies
; k
++)
3221 if (r10_bio
->devs
[k
].devnum
== i
)
3223 if (!test_bit(In_sync
,
3224 &mirror
->rdev
->flags
)
3225 && !rdev_set_badblocks(
3227 r10_bio
->devs
[k
].addr
,
3230 if (mirror
->replacement
&&
3231 !rdev_set_badblocks(
3232 mirror
->replacement
,
3233 r10_bio
->devs
[k
].addr
,
3238 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3240 printk(KERN_INFO
"md/raid10:%s: insufficient "
3241 "working devices for recovery.\n",
3243 mirror
->recovery_disabled
3244 = mddev
->recovery_disabled
;
3249 if (biolist
== NULL
) {
3251 struct r10bio
*rb2
= r10_bio
;
3252 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3253 rb2
->master_bio
= NULL
;
3259 /* resync. Schedule a read for every block at this virt offset */
3262 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3264 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3265 &sync_blocks
, mddev
->degraded
) &&
3266 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3267 &mddev
->recovery
)) {
3268 /* We can skip this block */
3270 return sync_blocks
+ sectors_skipped
;
3272 if (sync_blocks
< max_sync
)
3273 max_sync
= sync_blocks
;
3274 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_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_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_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 generic_make_request(bio
);
3415 if (sectors_skipped
)
3416 /* pretend they weren't skipped, it makes
3417 * no important difference in this case
3419 md_done_sync(mddev
, sectors_skipped
, 1);
3421 return sectors_skipped
+ nr_sectors
;
3423 /* There is nowhere to write, so all non-sync
3424 * drives must be failed or in resync, all drives
3425 * have a bad block, so try the next chunk...
3427 if (sector_nr
+ max_sync
< max_sector
)
3428 max_sector
= sector_nr
+ max_sync
;
3430 sectors_skipped
+= (max_sector
- sector_nr
);
3432 sector_nr
= max_sector
;
3437 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3440 struct r10conf
*conf
= mddev
->private;
3443 raid_disks
= min(conf
->geo
.raid_disks
,
3444 conf
->prev
.raid_disks
);
3446 sectors
= conf
->dev_sectors
;
3448 size
= sectors
>> conf
->geo
.chunk_shift
;
3449 sector_div(size
, conf
->geo
.far_copies
);
3450 size
= size
* raid_disks
;
3451 sector_div(size
, conf
->geo
.near_copies
);
3453 return size
<< conf
->geo
.chunk_shift
;
3456 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3458 /* Calculate the number of sectors-per-device that will
3459 * actually be used, and set conf->dev_sectors and
3463 size
= size
>> conf
->geo
.chunk_shift
;
3464 sector_div(size
, conf
->geo
.far_copies
);
3465 size
= size
* conf
->geo
.raid_disks
;
3466 sector_div(size
, conf
->geo
.near_copies
);
3467 /* 'size' is now the number of chunks in the array */
3468 /* calculate "used chunks per device" */
3469 size
= size
* conf
->copies
;
3471 /* We need to round up when dividing by raid_disks to
3472 * get the stride size.
3474 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3476 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3478 if (conf
->geo
.far_offset
)
3479 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3481 sector_div(size
, conf
->geo
.far_copies
);
3482 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3486 enum geo_type
{geo_new
, geo_old
, geo_start
};
3487 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3490 int layout
, chunk
, disks
;
3493 layout
= mddev
->layout
;
3494 chunk
= mddev
->chunk_sectors
;
3495 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3498 layout
= mddev
->new_layout
;
3499 chunk
= mddev
->new_chunk_sectors
;
3500 disks
= mddev
->raid_disks
;
3502 default: /* avoid 'may be unused' warnings */
3503 case geo_start
: /* new when starting reshape - raid_disks not
3505 layout
= mddev
->new_layout
;
3506 chunk
= mddev
->new_chunk_sectors
;
3507 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3512 if (chunk
< (PAGE_SIZE
>> 9) ||
3513 !is_power_of_2(chunk
))
3516 fc
= (layout
>> 8) & 255;
3517 fo
= layout
& (1<<16);
3518 geo
->raid_disks
= disks
;
3519 geo
->near_copies
= nc
;
3520 geo
->far_copies
= fc
;
3521 geo
->far_offset
= fo
;
3522 geo
->far_set_size
= (layout
& (1<<17)) ? disks
/ fc
: disks
;
3523 geo
->chunk_mask
= chunk
- 1;
3524 geo
->chunk_shift
= ffz(~chunk
);
3528 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3530 struct r10conf
*conf
= NULL
;
3535 copies
= setup_geo(&geo
, mddev
, geo_new
);
3538 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3539 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3540 mdname(mddev
), PAGE_SIZE
);
3544 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3545 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3546 mdname(mddev
), mddev
->new_layout
);
3551 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3555 /* FIXME calc properly */
3556 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3557 max(0,mddev
->delta_disks
)),
3562 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3567 conf
->copies
= copies
;
3568 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3569 r10bio_pool_free
, conf
);
3570 if (!conf
->r10bio_pool
)
3573 calc_sectors(conf
, mddev
->dev_sectors
);
3574 if (mddev
->reshape_position
== MaxSector
) {
3575 conf
->prev
= conf
->geo
;
3576 conf
->reshape_progress
= MaxSector
;
3578 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3582 conf
->reshape_progress
= mddev
->reshape_position
;
3583 if (conf
->prev
.far_offset
)
3584 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3586 /* far_copies must be 1 */
3587 conf
->prev
.stride
= conf
->dev_sectors
;
3589 spin_lock_init(&conf
->device_lock
);
3590 INIT_LIST_HEAD(&conf
->retry_list
);
3592 spin_lock_init(&conf
->resync_lock
);
3593 init_waitqueue_head(&conf
->wait_barrier
);
3595 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3599 conf
->mddev
= mddev
;
3604 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3607 if (conf
->r10bio_pool
)
3608 mempool_destroy(conf
->r10bio_pool
);
3609 kfree(conf
->mirrors
);
3610 safe_put_page(conf
->tmppage
);
3613 return ERR_PTR(err
);
3616 static int run(struct mddev
*mddev
)
3618 struct r10conf
*conf
;
3619 int i
, disk_idx
, chunk_size
;
3620 struct raid10_info
*disk
;
3621 struct md_rdev
*rdev
;
3623 sector_t min_offset_diff
= 0;
3625 bool discard_supported
= false;
3627 if (mddev
->private == NULL
) {
3628 conf
= setup_conf(mddev
);
3630 return PTR_ERR(conf
);
3631 mddev
->private = conf
;
3633 conf
= mddev
->private;
3637 mddev
->thread
= conf
->thread
;
3638 conf
->thread
= NULL
;
3640 chunk_size
= mddev
->chunk_sectors
<< 9;
3642 blk_queue_max_discard_sectors(mddev
->queue
,
3643 mddev
->chunk_sectors
);
3644 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3645 blk_queue_io_min(mddev
->queue
, chunk_size
);
3646 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3647 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3649 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3650 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3653 rdev_for_each(rdev
, mddev
) {
3655 struct request_queue
*q
;
3657 disk_idx
= rdev
->raid_disk
;
3660 if (disk_idx
>= conf
->geo
.raid_disks
&&
3661 disk_idx
>= conf
->prev
.raid_disks
)
3663 disk
= conf
->mirrors
+ disk_idx
;
3665 if (test_bit(Replacement
, &rdev
->flags
)) {
3666 if (disk
->replacement
)
3668 disk
->replacement
= rdev
;
3674 q
= bdev_get_queue(rdev
->bdev
);
3675 if (q
->merge_bvec_fn
)
3676 mddev
->merge_check_needed
= 1;
3677 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3678 if (!mddev
->reshape_backwards
)
3682 if (first
|| diff
< min_offset_diff
)
3683 min_offset_diff
= diff
;
3686 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3687 rdev
->data_offset
<< 9);
3689 disk
->head_position
= 0;
3691 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3692 discard_supported
= true;
3696 if (discard_supported
)
3697 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3700 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3703 /* need to check that every block has at least one working mirror */
3704 if (!enough(conf
, -1)) {
3705 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3710 if (conf
->reshape_progress
!= MaxSector
) {
3711 /* must ensure that shape change is supported */
3712 if (conf
->geo
.far_copies
!= 1 &&
3713 conf
->geo
.far_offset
== 0)
3715 if (conf
->prev
.far_copies
!= 1 &&
3716 conf
->geo
.far_offset
== 0)
3720 mddev
->degraded
= 0;
3722 i
< conf
->geo
.raid_disks
3723 || i
< conf
->prev
.raid_disks
;
3726 disk
= conf
->mirrors
+ i
;
3728 if (!disk
->rdev
&& disk
->replacement
) {
3729 /* The replacement is all we have - use it */
3730 disk
->rdev
= disk
->replacement
;
3731 disk
->replacement
= NULL
;
3732 clear_bit(Replacement
, &disk
->rdev
->flags
);
3736 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3737 disk
->head_position
= 0;
3742 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3745 if (mddev
->recovery_cp
!= MaxSector
)
3746 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3747 " -- starting background reconstruction\n",
3750 "md/raid10:%s: active with %d out of %d devices\n",
3751 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3752 conf
->geo
.raid_disks
);
3754 * Ok, everything is just fine now
3756 mddev
->dev_sectors
= conf
->dev_sectors
;
3757 size
= raid10_size(mddev
, 0, 0);
3758 md_set_array_sectors(mddev
, size
);
3759 mddev
->resync_max_sectors
= size
;
3762 int stripe
= conf
->geo
.raid_disks
*
3763 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3764 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3765 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3767 /* Calculate max read-ahead size.
3768 * We need to readahead at least twice a whole stripe....
3771 stripe
/= conf
->geo
.near_copies
;
3772 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3773 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3774 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3778 if (md_integrity_register(mddev
))
3781 if (conf
->reshape_progress
!= MaxSector
) {
3782 unsigned long before_length
, after_length
;
3784 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3785 conf
->prev
.far_copies
);
3786 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3787 conf
->geo
.far_copies
);
3789 if (max(before_length
, after_length
) > min_offset_diff
) {
3790 /* This cannot work */
3791 printk("md/raid10: offset difference not enough to continue reshape\n");
3794 conf
->offset_diff
= min_offset_diff
;
3796 conf
->reshape_safe
= conf
->reshape_progress
;
3797 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3798 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3799 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3800 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3801 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3808 md_unregister_thread(&mddev
->thread
);
3809 if (conf
->r10bio_pool
)
3810 mempool_destroy(conf
->r10bio_pool
);
3811 safe_put_page(conf
->tmppage
);
3812 kfree(conf
->mirrors
);
3814 mddev
->private = NULL
;
3819 static int stop(struct mddev
*mddev
)
3821 struct r10conf
*conf
= mddev
->private;
3823 raise_barrier(conf
, 0);
3824 lower_barrier(conf
);
3826 md_unregister_thread(&mddev
->thread
);
3828 /* the unplug fn references 'conf'*/
3829 blk_sync_queue(mddev
->queue
);
3831 if (conf
->r10bio_pool
)
3832 mempool_destroy(conf
->r10bio_pool
);
3833 safe_put_page(conf
->tmppage
);
3834 kfree(conf
->mirrors
);
3836 mddev
->private = NULL
;
3840 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3842 struct r10conf
*conf
= mddev
->private;
3846 raise_barrier(conf
, 0);
3849 lower_barrier(conf
);
3854 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3856 /* Resize of 'far' arrays is not supported.
3857 * For 'near' and 'offset' arrays we can set the
3858 * number of sectors used to be an appropriate multiple
3859 * of the chunk size.
3860 * For 'offset', this is far_copies*chunksize.
3861 * For 'near' the multiplier is the LCM of
3862 * near_copies and raid_disks.
3863 * So if far_copies > 1 && !far_offset, fail.
3864 * Else find LCM(raid_disks, near_copy)*far_copies and
3865 * multiply by chunk_size. Then round to this number.
3866 * This is mostly done by raid10_size()
3868 struct r10conf
*conf
= mddev
->private;
3869 sector_t oldsize
, size
;
3871 if (mddev
->reshape_position
!= MaxSector
)
3874 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3877 oldsize
= raid10_size(mddev
, 0, 0);
3878 size
= raid10_size(mddev
, sectors
, 0);
3879 if (mddev
->external_size
&&
3880 mddev
->array_sectors
> size
)
3882 if (mddev
->bitmap
) {
3883 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3887 md_set_array_sectors(mddev
, size
);
3888 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3889 revalidate_disk(mddev
->gendisk
);
3890 if (sectors
> mddev
->dev_sectors
&&
3891 mddev
->recovery_cp
> oldsize
) {
3892 mddev
->recovery_cp
= oldsize
;
3893 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3895 calc_sectors(conf
, sectors
);
3896 mddev
->dev_sectors
= conf
->dev_sectors
;
3897 mddev
->resync_max_sectors
= size
;
3901 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3903 struct md_rdev
*rdev
;
3904 struct r10conf
*conf
;
3906 if (mddev
->degraded
> 0) {
3907 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3909 return ERR_PTR(-EINVAL
);
3912 /* Set new parameters */
3913 mddev
->new_level
= 10;
3914 /* new layout: far_copies = 1, near_copies = 2 */
3915 mddev
->new_layout
= (1<<8) + 2;
3916 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3917 mddev
->delta_disks
= mddev
->raid_disks
;
3918 mddev
->raid_disks
*= 2;
3919 /* make sure it will be not marked as dirty */
3920 mddev
->recovery_cp
= MaxSector
;
3922 conf
= setup_conf(mddev
);
3923 if (!IS_ERR(conf
)) {
3924 rdev_for_each(rdev
, mddev
)
3925 if (rdev
->raid_disk
>= 0)
3926 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3933 static void *raid10_takeover(struct mddev
*mddev
)
3935 struct r0conf
*raid0_conf
;
3937 /* raid10 can take over:
3938 * raid0 - providing it has only two drives
3940 if (mddev
->level
== 0) {
3941 /* for raid0 takeover only one zone is supported */
3942 raid0_conf
= mddev
->private;
3943 if (raid0_conf
->nr_strip_zones
> 1) {
3944 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3945 " with more than one zone.\n",
3947 return ERR_PTR(-EINVAL
);
3949 return raid10_takeover_raid0(mddev
);
3951 return ERR_PTR(-EINVAL
);
3954 static int raid10_check_reshape(struct mddev
*mddev
)
3956 /* Called when there is a request to change
3957 * - layout (to ->new_layout)
3958 * - chunk size (to ->new_chunk_sectors)
3959 * - raid_disks (by delta_disks)
3960 * or when trying to restart a reshape that was ongoing.
3962 * We need to validate the request and possibly allocate
3963 * space if that might be an issue later.
3965 * Currently we reject any reshape of a 'far' mode array,
3966 * allow chunk size to change if new is generally acceptable,
3967 * allow raid_disks to increase, and allow
3968 * a switch between 'near' mode and 'offset' mode.
3970 struct r10conf
*conf
= mddev
->private;
3973 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3976 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3977 /* mustn't change number of copies */
3979 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3980 /* Cannot switch to 'far' mode */
3983 if (mddev
->array_sectors
& geo
.chunk_mask
)
3984 /* not factor of array size */
3987 if (!enough(conf
, -1))
3990 kfree(conf
->mirrors_new
);
3991 conf
->mirrors_new
= NULL
;
3992 if (mddev
->delta_disks
> 0) {
3993 /* allocate new 'mirrors' list */
3994 conf
->mirrors_new
= kzalloc(
3995 sizeof(struct raid10_info
)
3996 *(mddev
->raid_disks
+
3997 mddev
->delta_disks
),
3999 if (!conf
->mirrors_new
)
4006 * Need to check if array has failed when deciding whether to:
4008 * - remove non-faulty devices
4011 * This determination is simple when no reshape is happening.
4012 * However if there is a reshape, we need to carefully check
4013 * both the before and after sections.
4014 * This is because some failed devices may only affect one
4015 * of the two sections, and some non-in_sync devices may
4016 * be insync in the section most affected by failed devices.
4018 static int calc_degraded(struct r10conf
*conf
)
4020 int degraded
, degraded2
;
4025 /* 'prev' section first */
4026 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4027 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4028 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4030 else if (!test_bit(In_sync
, &rdev
->flags
))
4031 /* When we can reduce the number of devices in
4032 * an array, this might not contribute to
4033 * 'degraded'. It does now.
4038 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4042 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4043 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4044 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4046 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4047 /* If reshape is increasing the number of devices,
4048 * this section has already been recovered, so
4049 * it doesn't contribute to degraded.
4052 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4057 if (degraded2
> degraded
)
4062 static int raid10_start_reshape(struct mddev
*mddev
)
4064 /* A 'reshape' has been requested. This commits
4065 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4066 * This also checks if there are enough spares and adds them
4068 * We currently require enough spares to make the final
4069 * array non-degraded. We also require that the difference
4070 * between old and new data_offset - on each device - is
4071 * enough that we never risk over-writing.
4074 unsigned long before_length
, after_length
;
4075 sector_t min_offset_diff
= 0;
4078 struct r10conf
*conf
= mddev
->private;
4079 struct md_rdev
*rdev
;
4083 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4086 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4089 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4090 conf
->prev
.far_copies
);
4091 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4092 conf
->geo
.far_copies
);
4094 rdev_for_each(rdev
, mddev
) {
4095 if (!test_bit(In_sync
, &rdev
->flags
)
4096 && !test_bit(Faulty
, &rdev
->flags
))
4098 if (rdev
->raid_disk
>= 0) {
4099 long long diff
= (rdev
->new_data_offset
4100 - rdev
->data_offset
);
4101 if (!mddev
->reshape_backwards
)
4105 if (first
|| diff
< min_offset_diff
)
4106 min_offset_diff
= diff
;
4110 if (max(before_length
, after_length
) > min_offset_diff
)
4113 if (spares
< mddev
->delta_disks
)
4116 conf
->offset_diff
= min_offset_diff
;
4117 spin_lock_irq(&conf
->device_lock
);
4118 if (conf
->mirrors_new
) {
4119 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4120 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4122 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
4123 conf
->mirrors_old
= conf
->mirrors
;
4124 conf
->mirrors
= conf
->mirrors_new
;
4125 conf
->mirrors_new
= NULL
;
4127 setup_geo(&conf
->geo
, mddev
, geo_start
);
4129 if (mddev
->reshape_backwards
) {
4130 sector_t size
= raid10_size(mddev
, 0, 0);
4131 if (size
< mddev
->array_sectors
) {
4132 spin_unlock_irq(&conf
->device_lock
);
4133 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4137 mddev
->resync_max_sectors
= size
;
4138 conf
->reshape_progress
= size
;
4140 conf
->reshape_progress
= 0;
4141 spin_unlock_irq(&conf
->device_lock
);
4143 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4144 ret
= bitmap_resize(mddev
->bitmap
,
4145 raid10_size(mddev
, 0,
4146 conf
->geo
.raid_disks
),
4151 if (mddev
->delta_disks
> 0) {
4152 rdev_for_each(rdev
, mddev
)
4153 if (rdev
->raid_disk
< 0 &&
4154 !test_bit(Faulty
, &rdev
->flags
)) {
4155 if (raid10_add_disk(mddev
, rdev
) == 0) {
4156 if (rdev
->raid_disk
>=
4157 conf
->prev
.raid_disks
)
4158 set_bit(In_sync
, &rdev
->flags
);
4160 rdev
->recovery_offset
= 0;
4162 if (sysfs_link_rdev(mddev
, rdev
))
4163 /* Failure here is OK */;
4165 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4166 && !test_bit(Faulty
, &rdev
->flags
)) {
4167 /* This is a spare that was manually added */
4168 set_bit(In_sync
, &rdev
->flags
);
4171 /* When a reshape changes the number of devices,
4172 * ->degraded is measured against the larger of the
4173 * pre and post numbers.
4175 spin_lock_irq(&conf
->device_lock
);
4176 mddev
->degraded
= calc_degraded(conf
);
4177 spin_unlock_irq(&conf
->device_lock
);
4178 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4179 mddev
->reshape_position
= conf
->reshape_progress
;
4180 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4182 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4183 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4184 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4185 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4187 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4189 if (!mddev
->sync_thread
) {
4193 conf
->reshape_checkpoint
= jiffies
;
4194 md_wakeup_thread(mddev
->sync_thread
);
4195 md_new_event(mddev
);
4199 mddev
->recovery
= 0;
4200 spin_lock_irq(&conf
->device_lock
);
4201 conf
->geo
= conf
->prev
;
4202 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4203 rdev_for_each(rdev
, mddev
)
4204 rdev
->new_data_offset
= rdev
->data_offset
;
4206 conf
->reshape_progress
= MaxSector
;
4207 mddev
->reshape_position
= MaxSector
;
4208 spin_unlock_irq(&conf
->device_lock
);
4212 /* Calculate the last device-address that could contain
4213 * any block from the chunk that includes the array-address 's'
4214 * and report the next address.
4215 * i.e. the address returned will be chunk-aligned and after
4216 * any data that is in the chunk containing 's'.
4218 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4220 s
= (s
| geo
->chunk_mask
) + 1;
4221 s
>>= geo
->chunk_shift
;
4222 s
*= geo
->near_copies
;
4223 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4224 s
*= geo
->far_copies
;
4225 s
<<= geo
->chunk_shift
;
4229 /* Calculate the first device-address that could contain
4230 * any block from the chunk that includes the array-address 's'.
4231 * This too will be the start of a chunk
4233 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4235 s
>>= geo
->chunk_shift
;
4236 s
*= geo
->near_copies
;
4237 sector_div(s
, geo
->raid_disks
);
4238 s
*= geo
->far_copies
;
4239 s
<<= geo
->chunk_shift
;
4243 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4246 /* We simply copy at most one chunk (smallest of old and new)
4247 * at a time, possibly less if that exceeds RESYNC_PAGES,
4248 * or we hit a bad block or something.
4249 * This might mean we pause for normal IO in the middle of
4250 * a chunk, but that is not a problem was mddev->reshape_position
4251 * can record any location.
4253 * If we will want to write to a location that isn't
4254 * yet recorded as 'safe' (i.e. in metadata on disk) then
4255 * we need to flush all reshape requests and update the metadata.
4257 * When reshaping forwards (e.g. to more devices), we interpret
4258 * 'safe' as the earliest block which might not have been copied
4259 * down yet. We divide this by previous stripe size and multiply
4260 * by previous stripe length to get lowest device offset that we
4261 * cannot write to yet.
4262 * We interpret 'sector_nr' as an address that we want to write to.
4263 * From this we use last_device_address() to find where we might
4264 * write to, and first_device_address on the 'safe' position.
4265 * If this 'next' write position is after the 'safe' position,
4266 * we must update the metadata to increase the 'safe' position.
4268 * When reshaping backwards, we round in the opposite direction
4269 * and perform the reverse test: next write position must not be
4270 * less than current safe position.
4272 * In all this the minimum difference in data offsets
4273 * (conf->offset_diff - always positive) allows a bit of slack,
4274 * so next can be after 'safe', but not by more than offset_disk
4276 * We need to prepare all the bios here before we start any IO
4277 * to ensure the size we choose is acceptable to all devices.
4278 * The means one for each copy for write-out and an extra one for
4280 * We store the read-in bio in ->master_bio and the others in
4281 * ->devs[x].bio and ->devs[x].repl_bio.
4283 struct r10conf
*conf
= mddev
->private;
4284 struct r10bio
*r10_bio
;
4285 sector_t next
, safe
, last
;
4289 struct md_rdev
*rdev
;
4292 struct bio
*bio
, *read_bio
;
4293 int sectors_done
= 0;
4295 if (sector_nr
== 0) {
4296 /* If restarting in the middle, skip the initial sectors */
4297 if (mddev
->reshape_backwards
&&
4298 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4299 sector_nr
= (raid10_size(mddev
, 0, 0)
4300 - conf
->reshape_progress
);
4301 } else if (!mddev
->reshape_backwards
&&
4302 conf
->reshape_progress
> 0)
4303 sector_nr
= conf
->reshape_progress
;
4305 mddev
->curr_resync_completed
= sector_nr
;
4306 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4312 /* We don't use sector_nr to track where we are up to
4313 * as that doesn't work well for ->reshape_backwards.
4314 * So just use ->reshape_progress.
4316 if (mddev
->reshape_backwards
) {
4317 /* 'next' is the earliest device address that we might
4318 * write to for this chunk in the new layout
4320 next
= first_dev_address(conf
->reshape_progress
- 1,
4323 /* 'safe' is the last device address that we might read from
4324 * in the old layout after a restart
4326 safe
= last_dev_address(conf
->reshape_safe
- 1,
4329 if (next
+ conf
->offset_diff
< safe
)
4332 last
= conf
->reshape_progress
- 1;
4333 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4334 & conf
->prev
.chunk_mask
);
4335 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4336 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4338 /* 'next' is after the last device address that we
4339 * might write to for this chunk in the new layout
4341 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4343 /* 'safe' is the earliest device address that we might
4344 * read from in the old layout after a restart
4346 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4348 /* Need to update metadata if 'next' might be beyond 'safe'
4349 * as that would possibly corrupt data
4351 if (next
> safe
+ conf
->offset_diff
)
4354 sector_nr
= conf
->reshape_progress
;
4355 last
= sector_nr
| (conf
->geo
.chunk_mask
4356 & conf
->prev
.chunk_mask
);
4358 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4359 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4363 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4364 /* Need to update reshape_position in metadata */
4366 mddev
->reshape_position
= conf
->reshape_progress
;
4367 if (mddev
->reshape_backwards
)
4368 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4369 - conf
->reshape_progress
;
4371 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4372 conf
->reshape_checkpoint
= jiffies
;
4373 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4374 md_wakeup_thread(mddev
->thread
);
4375 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4376 kthread_should_stop());
4377 conf
->reshape_safe
= mddev
->reshape_position
;
4378 allow_barrier(conf
);
4382 /* Now schedule reads for blocks from sector_nr to last */
4383 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4384 raise_barrier(conf
, sectors_done
!= 0);
4385 atomic_set(&r10_bio
->remaining
, 0);
4386 r10_bio
->mddev
= mddev
;
4387 r10_bio
->sector
= sector_nr
;
4388 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4389 r10_bio
->sectors
= last
- sector_nr
+ 1;
4390 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4391 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4394 /* Cannot read from here, so need to record bad blocks
4395 * on all the target devices.
4398 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4399 return sectors_done
;
4402 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4404 read_bio
->bi_bdev
= rdev
->bdev
;
4405 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4406 + rdev
->data_offset
);
4407 read_bio
->bi_private
= r10_bio
;
4408 read_bio
->bi_end_io
= end_sync_read
;
4409 read_bio
->bi_rw
= READ
;
4410 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4411 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4412 read_bio
->bi_vcnt
= 0;
4413 read_bio
->bi_size
= 0;
4414 r10_bio
->master_bio
= read_bio
;
4415 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4417 /* Now find the locations in the new layout */
4418 __raid10_find_phys(&conf
->geo
, r10_bio
);
4421 read_bio
->bi_next
= NULL
;
4423 for (s
= 0; s
< conf
->copies
*2; s
++) {
4425 int d
= r10_bio
->devs
[s
/2].devnum
;
4426 struct md_rdev
*rdev2
;
4428 rdev2
= conf
->mirrors
[d
].replacement
;
4429 b
= r10_bio
->devs
[s
/2].repl_bio
;
4431 rdev2
= conf
->mirrors
[d
].rdev
;
4432 b
= r10_bio
->devs
[s
/2].bio
;
4434 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4438 b
->bi_bdev
= rdev2
->bdev
;
4439 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4440 b
->bi_private
= r10_bio
;
4441 b
->bi_end_io
= end_reshape_write
;
4447 /* Now add as many pages as possible to all of these bios. */
4450 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4451 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4452 int len
= (max_sectors
- s
) << 9;
4453 if (len
> PAGE_SIZE
)
4455 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4457 if (bio_add_page(bio
, page
, len
, 0))
4460 /* Didn't fit, must stop */
4462 bio2
&& bio2
!= bio
;
4463 bio2
= bio2
->bi_next
) {
4464 /* Remove last page from this bio */
4466 bio2
->bi_size
-= len
;
4467 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4471 sector_nr
+= len
>> 9;
4472 nr_sectors
+= len
>> 9;
4475 r10_bio
->sectors
= nr_sectors
;
4477 /* Now submit the read */
4478 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4479 atomic_inc(&r10_bio
->remaining
);
4480 read_bio
->bi_next
= NULL
;
4481 generic_make_request(read_bio
);
4482 sector_nr
+= nr_sectors
;
4483 sectors_done
+= nr_sectors
;
4484 if (sector_nr
<= last
)
4487 /* Now that we have done the whole section we can
4488 * update reshape_progress
4490 if (mddev
->reshape_backwards
)
4491 conf
->reshape_progress
-= sectors_done
;
4493 conf
->reshape_progress
+= sectors_done
;
4495 return sectors_done
;
4498 static void end_reshape_request(struct r10bio
*r10_bio
);
4499 static int handle_reshape_read_error(struct mddev
*mddev
,
4500 struct r10bio
*r10_bio
);
4501 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4503 /* Reshape read completed. Hopefully we have a block
4505 * If we got a read error then we do sync 1-page reads from
4506 * elsewhere until we find the data - or give up.
4508 struct r10conf
*conf
= mddev
->private;
4511 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4512 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4513 /* Reshape has been aborted */
4514 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4518 /* We definitely have the data in the pages, schedule the
4521 atomic_set(&r10_bio
->remaining
, 1);
4522 for (s
= 0; s
< conf
->copies
*2; s
++) {
4524 int d
= r10_bio
->devs
[s
/2].devnum
;
4525 struct md_rdev
*rdev
;
4527 rdev
= conf
->mirrors
[d
].replacement
;
4528 b
= r10_bio
->devs
[s
/2].repl_bio
;
4530 rdev
= conf
->mirrors
[d
].rdev
;
4531 b
= r10_bio
->devs
[s
/2].bio
;
4533 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4535 atomic_inc(&rdev
->nr_pending
);
4536 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4537 atomic_inc(&r10_bio
->remaining
);
4539 generic_make_request(b
);
4541 end_reshape_request(r10_bio
);
4544 static void end_reshape(struct r10conf
*conf
)
4546 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4549 spin_lock_irq(&conf
->device_lock
);
4550 conf
->prev
= conf
->geo
;
4551 md_finish_reshape(conf
->mddev
);
4553 conf
->reshape_progress
= MaxSector
;
4554 spin_unlock_irq(&conf
->device_lock
);
4556 /* read-ahead size must cover two whole stripes, which is
4557 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4559 if (conf
->mddev
->queue
) {
4560 int stripe
= conf
->geo
.raid_disks
*
4561 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4562 stripe
/= conf
->geo
.near_copies
;
4563 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4564 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4570 static int handle_reshape_read_error(struct mddev
*mddev
,
4571 struct r10bio
*r10_bio
)
4573 /* Use sync reads to get the blocks from somewhere else */
4574 int sectors
= r10_bio
->sectors
;
4575 struct r10conf
*conf
= mddev
->private;
4577 struct r10bio r10_bio
;
4578 struct r10dev devs
[conf
->copies
];
4580 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4583 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4585 r10b
->sector
= r10_bio
->sector
;
4586 __raid10_find_phys(&conf
->prev
, r10b
);
4591 int first_slot
= slot
;
4593 if (s
> (PAGE_SIZE
>> 9))
4597 int d
= r10b
->devs
[slot
].devnum
;
4598 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4601 test_bit(Faulty
, &rdev
->flags
) ||
4602 !test_bit(In_sync
, &rdev
->flags
))
4605 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4606 success
= sync_page_io(rdev
,
4615 if (slot
>= conf
->copies
)
4617 if (slot
== first_slot
)
4621 /* couldn't read this block, must give up */
4622 set_bit(MD_RECOVERY_INTR
,
4632 static void end_reshape_write(struct bio
*bio
, int error
)
4634 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4635 struct r10bio
*r10_bio
= bio
->bi_private
;
4636 struct mddev
*mddev
= r10_bio
->mddev
;
4637 struct r10conf
*conf
= mddev
->private;
4641 struct md_rdev
*rdev
= NULL
;
4643 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4645 rdev
= conf
->mirrors
[d
].replacement
;
4648 rdev
= conf
->mirrors
[d
].rdev
;
4652 /* FIXME should record badblock */
4653 md_error(mddev
, rdev
);
4656 rdev_dec_pending(rdev
, mddev
);
4657 end_reshape_request(r10_bio
);
4660 static void end_reshape_request(struct r10bio
*r10_bio
)
4662 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4664 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4665 bio_put(r10_bio
->master_bio
);
4669 static void raid10_finish_reshape(struct mddev
*mddev
)
4671 struct r10conf
*conf
= mddev
->private;
4673 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4676 if (mddev
->delta_disks
> 0) {
4677 sector_t size
= raid10_size(mddev
, 0, 0);
4678 md_set_array_sectors(mddev
, size
);
4679 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4680 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4681 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4683 mddev
->resync_max_sectors
= size
;
4684 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4685 revalidate_disk(mddev
->gendisk
);
4688 for (d
= conf
->geo
.raid_disks
;
4689 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4691 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4693 clear_bit(In_sync
, &rdev
->flags
);
4694 rdev
= conf
->mirrors
[d
].replacement
;
4696 clear_bit(In_sync
, &rdev
->flags
);
4699 mddev
->layout
= mddev
->new_layout
;
4700 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4701 mddev
->reshape_position
= MaxSector
;
4702 mddev
->delta_disks
= 0;
4703 mddev
->reshape_backwards
= 0;
4706 static struct md_personality raid10_personality
=
4710 .owner
= THIS_MODULE
,
4711 .make_request
= make_request
,
4715 .error_handler
= error
,
4716 .hot_add_disk
= raid10_add_disk
,
4717 .hot_remove_disk
= raid10_remove_disk
,
4718 .spare_active
= raid10_spare_active
,
4719 .sync_request
= sync_request
,
4720 .quiesce
= raid10_quiesce
,
4721 .size
= raid10_size
,
4722 .resize
= raid10_resize
,
4723 .takeover
= raid10_takeover
,
4724 .check_reshape
= raid10_check_reshape
,
4725 .start_reshape
= raid10_start_reshape
,
4726 .finish_reshape
= raid10_finish_reshape
,
4729 static int __init
raid_init(void)
4731 return register_md_personality(&raid10_personality
);
4734 static void raid_exit(void)
4736 unregister_md_personality(&raid10_personality
);
4739 module_init(raid_init
);
4740 module_exit(raid_exit
);
4741 MODULE_LICENSE("GPL");
4742 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4743 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4744 MODULE_ALIAS("md-raid10");
4745 MODULE_ALIAS("md-level-10");
4747 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);