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 bio_trim(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 bio_trim(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 bio_trim(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 && tmp
->rdev
->recovery_offset
== MaxSector
1786 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1787 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1789 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1792 spin_lock_irqsave(&conf
->device_lock
, flags
);
1793 mddev
->degraded
-= count
;
1794 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1801 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1803 struct r10conf
*conf
= mddev
->private;
1807 int last
= conf
->geo
.raid_disks
- 1;
1808 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1810 if (mddev
->recovery_cp
< MaxSector
)
1811 /* only hot-add to in-sync arrays, as recovery is
1812 * very different from resync
1815 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1818 if (rdev
->raid_disk
>= 0)
1819 first
= last
= rdev
->raid_disk
;
1821 if (q
->merge_bvec_fn
) {
1822 set_bit(Unmerged
, &rdev
->flags
);
1823 mddev
->merge_check_needed
= 1;
1826 if (rdev
->saved_raid_disk
>= first
&&
1827 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1828 mirror
= rdev
->saved_raid_disk
;
1831 for ( ; mirror
<= last
; mirror
++) {
1832 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1833 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1836 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1837 p
->replacement
!= NULL
)
1839 clear_bit(In_sync
, &rdev
->flags
);
1840 set_bit(Replacement
, &rdev
->flags
);
1841 rdev
->raid_disk
= mirror
;
1844 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1845 rdev
->data_offset
<< 9);
1847 rcu_assign_pointer(p
->replacement
, rdev
);
1852 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1853 rdev
->data_offset
<< 9);
1855 p
->head_position
= 0;
1856 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1857 rdev
->raid_disk
= mirror
;
1859 if (rdev
->saved_raid_disk
!= mirror
)
1861 rcu_assign_pointer(p
->rdev
, rdev
);
1864 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1865 /* Some requests might not have seen this new
1866 * merge_bvec_fn. We must wait for them to complete
1867 * before merging the device fully.
1868 * First we make sure any code which has tested
1869 * our function has submitted the request, then
1870 * we wait for all outstanding requests to complete.
1872 synchronize_sched();
1873 freeze_array(conf
, 0);
1874 unfreeze_array(conf
);
1875 clear_bit(Unmerged
, &rdev
->flags
);
1877 md_integrity_add_rdev(rdev
, mddev
);
1878 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1879 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1885 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1887 struct r10conf
*conf
= mddev
->private;
1889 int number
= rdev
->raid_disk
;
1890 struct md_rdev
**rdevp
;
1891 struct raid10_info
*p
= conf
->mirrors
+ number
;
1894 if (rdev
== p
->rdev
)
1896 else if (rdev
== p
->replacement
)
1897 rdevp
= &p
->replacement
;
1901 if (test_bit(In_sync
, &rdev
->flags
) ||
1902 atomic_read(&rdev
->nr_pending
)) {
1906 /* Only remove faulty devices if recovery
1909 if (!test_bit(Faulty
, &rdev
->flags
) &&
1910 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1911 (!p
->replacement
|| p
->replacement
== rdev
) &&
1912 number
< conf
->geo
.raid_disks
&&
1919 if (atomic_read(&rdev
->nr_pending
)) {
1920 /* lost the race, try later */
1924 } else if (p
->replacement
) {
1925 /* We must have just cleared 'rdev' */
1926 p
->rdev
= p
->replacement
;
1927 clear_bit(Replacement
, &p
->replacement
->flags
);
1928 smp_mb(); /* Make sure other CPUs may see both as identical
1929 * but will never see neither -- if they are careful.
1931 p
->replacement
= NULL
;
1932 clear_bit(WantReplacement
, &rdev
->flags
);
1934 /* We might have just remove the Replacement as faulty
1935 * Clear the flag just in case
1937 clear_bit(WantReplacement
, &rdev
->flags
);
1939 err
= md_integrity_register(mddev
);
1948 static void end_sync_read(struct bio
*bio
, int error
)
1950 struct r10bio
*r10_bio
= bio
->bi_private
;
1951 struct r10conf
*conf
= r10_bio
->mddev
->private;
1954 if (bio
== r10_bio
->master_bio
) {
1955 /* this is a reshape read */
1956 d
= r10_bio
->read_slot
; /* really the read dev */
1958 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1960 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1961 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1963 /* The write handler will notice the lack of
1964 * R10BIO_Uptodate and record any errors etc
1966 atomic_add(r10_bio
->sectors
,
1967 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1969 /* for reconstruct, we always reschedule after a read.
1970 * for resync, only after all reads
1972 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1973 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1974 atomic_dec_and_test(&r10_bio
->remaining
)) {
1975 /* we have read all the blocks,
1976 * do the comparison in process context in raid10d
1978 reschedule_retry(r10_bio
);
1982 static void end_sync_request(struct r10bio
*r10_bio
)
1984 struct mddev
*mddev
= r10_bio
->mddev
;
1986 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1987 if (r10_bio
->master_bio
== NULL
) {
1988 /* the primary of several recovery bios */
1989 sector_t s
= r10_bio
->sectors
;
1990 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1991 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1992 reschedule_retry(r10_bio
);
1995 md_done_sync(mddev
, s
, 1);
1998 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1999 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2000 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2001 reschedule_retry(r10_bio
);
2009 static void end_sync_write(struct bio
*bio
, int error
)
2011 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2012 struct r10bio
*r10_bio
= bio
->bi_private
;
2013 struct mddev
*mddev
= r10_bio
->mddev
;
2014 struct r10conf
*conf
= mddev
->private;
2020 struct md_rdev
*rdev
= NULL
;
2022 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
2024 rdev
= conf
->mirrors
[d
].replacement
;
2026 rdev
= conf
->mirrors
[d
].rdev
;
2030 md_error(mddev
, rdev
);
2032 set_bit(WriteErrorSeen
, &rdev
->flags
);
2033 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2034 set_bit(MD_RECOVERY_NEEDED
,
2035 &rdev
->mddev
->recovery
);
2036 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
2038 } else if (is_badblock(rdev
,
2039 r10_bio
->devs
[slot
].addr
,
2041 &first_bad
, &bad_sectors
))
2042 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
2044 rdev_dec_pending(rdev
, mddev
);
2046 end_sync_request(r10_bio
);
2050 * Note: sync and recover and handled very differently for raid10
2051 * This code is for resync.
2052 * For resync, we read through virtual addresses and read all blocks.
2053 * If there is any error, we schedule a write. The lowest numbered
2054 * drive is authoritative.
2055 * However requests come for physical address, so we need to map.
2056 * For every physical address there are raid_disks/copies virtual addresses,
2057 * which is always are least one, but is not necessarly an integer.
2058 * This means that a physical address can span multiple chunks, so we may
2059 * have to submit multiple io requests for a single sync request.
2062 * We check if all blocks are in-sync and only write to blocks that
2065 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2067 struct r10conf
*conf
= mddev
->private;
2069 struct bio
*tbio
, *fbio
;
2072 atomic_set(&r10_bio
->remaining
, 1);
2074 /* find the first device with a block */
2075 for (i
=0; i
<conf
->copies
; i
++)
2076 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
2079 if (i
== conf
->copies
)
2083 fbio
= r10_bio
->devs
[i
].bio
;
2085 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2086 /* now find blocks with errors */
2087 for (i
=0 ; i
< conf
->copies
; i
++) {
2090 tbio
= r10_bio
->devs
[i
].bio
;
2092 if (tbio
->bi_end_io
!= end_sync_read
)
2096 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2097 /* We know that the bi_io_vec layout is the same for
2098 * both 'first' and 'i', so we just compare them.
2099 * All vec entries are PAGE_SIZE;
2101 int sectors
= r10_bio
->sectors
;
2102 for (j
= 0; j
< vcnt
; j
++) {
2103 int len
= PAGE_SIZE
;
2104 if (sectors
< (len
/ 512))
2105 len
= sectors
* 512;
2106 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2107 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2114 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2115 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2116 /* Don't fix anything. */
2119 /* Ok, we need to write this bio, either to correct an
2120 * inconsistency or to correct an unreadable block.
2121 * First we need to fixup bv_offset, bv_len and
2122 * bi_vecs, as the read request might have corrupted these
2126 tbio
->bi_vcnt
= vcnt
;
2127 tbio
->bi_size
= r10_bio
->sectors
<< 9;
2128 tbio
->bi_rw
= WRITE
;
2129 tbio
->bi_private
= r10_bio
;
2130 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
2132 for (j
=0; j
< vcnt
; j
++) {
2133 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2134 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2136 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2137 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2140 tbio
->bi_end_io
= end_sync_write
;
2142 d
= r10_bio
->devs
[i
].devnum
;
2143 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2144 atomic_inc(&r10_bio
->remaining
);
2145 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2147 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2148 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2149 generic_make_request(tbio
);
2152 /* Now write out to any replacement devices
2155 for (i
= 0; i
< conf
->copies
; i
++) {
2158 tbio
= r10_bio
->devs
[i
].repl_bio
;
2159 if (!tbio
|| !tbio
->bi_end_io
)
2161 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2162 && r10_bio
->devs
[i
].bio
!= fbio
)
2163 for (j
= 0; j
< vcnt
; j
++)
2164 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2165 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2167 d
= r10_bio
->devs
[i
].devnum
;
2168 atomic_inc(&r10_bio
->remaining
);
2169 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2171 generic_make_request(tbio
);
2175 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2176 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2182 * Now for the recovery code.
2183 * Recovery happens across physical sectors.
2184 * We recover all non-is_sync drives by finding the virtual address of
2185 * each, and then choose a working drive that also has that virt address.
2186 * There is a separate r10_bio for each non-in_sync drive.
2187 * Only the first two slots are in use. The first for reading,
2188 * The second for writing.
2191 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2193 /* We got a read error during recovery.
2194 * We repeat the read in smaller page-sized sections.
2195 * If a read succeeds, write it to the new device or record
2196 * a bad block if we cannot.
2197 * If a read fails, record a bad block on both old and
2200 struct mddev
*mddev
= r10_bio
->mddev
;
2201 struct r10conf
*conf
= mddev
->private;
2202 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2204 int sectors
= r10_bio
->sectors
;
2206 int dr
= r10_bio
->devs
[0].devnum
;
2207 int dw
= r10_bio
->devs
[1].devnum
;
2211 struct md_rdev
*rdev
;
2215 if (s
> (PAGE_SIZE
>>9))
2218 rdev
= conf
->mirrors
[dr
].rdev
;
2219 addr
= r10_bio
->devs
[0].addr
+ sect
,
2220 ok
= sync_page_io(rdev
,
2223 bio
->bi_io_vec
[idx
].bv_page
,
2226 rdev
= conf
->mirrors
[dw
].rdev
;
2227 addr
= r10_bio
->devs
[1].addr
+ sect
;
2228 ok
= sync_page_io(rdev
,
2231 bio
->bi_io_vec
[idx
].bv_page
,
2234 set_bit(WriteErrorSeen
, &rdev
->flags
);
2235 if (!test_and_set_bit(WantReplacement
,
2237 set_bit(MD_RECOVERY_NEEDED
,
2238 &rdev
->mddev
->recovery
);
2242 /* We don't worry if we cannot set a bad block -
2243 * it really is bad so there is no loss in not
2246 rdev_set_badblocks(rdev
, addr
, s
, 0);
2248 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2249 /* need bad block on destination too */
2250 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2251 addr
= r10_bio
->devs
[1].addr
+ sect
;
2252 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2254 /* just abort the recovery */
2256 "md/raid10:%s: recovery aborted"
2257 " due to read error\n",
2260 conf
->mirrors
[dw
].recovery_disabled
2261 = mddev
->recovery_disabled
;
2262 set_bit(MD_RECOVERY_INTR
,
2275 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2277 struct r10conf
*conf
= mddev
->private;
2279 struct bio
*wbio
, *wbio2
;
2281 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2282 fix_recovery_read_error(r10_bio
);
2283 end_sync_request(r10_bio
);
2288 * share the pages with the first bio
2289 * and submit the write request
2291 d
= r10_bio
->devs
[1].devnum
;
2292 wbio
= r10_bio
->devs
[1].bio
;
2293 wbio2
= r10_bio
->devs
[1].repl_bio
;
2294 /* Need to test wbio2->bi_end_io before we call
2295 * generic_make_request as if the former is NULL,
2296 * the latter is free to free wbio2.
2298 if (wbio2
&& !wbio2
->bi_end_io
)
2300 if (wbio
->bi_end_io
) {
2301 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2302 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2303 generic_make_request(wbio
);
2306 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2307 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2308 bio_sectors(wbio2
));
2309 generic_make_request(wbio2
);
2315 * Used by fix_read_error() to decay the per rdev read_errors.
2316 * We halve the read error count for every hour that has elapsed
2317 * since the last recorded read error.
2320 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2322 struct timespec cur_time_mon
;
2323 unsigned long hours_since_last
;
2324 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2326 ktime_get_ts(&cur_time_mon
);
2328 if (rdev
->last_read_error
.tv_sec
== 0 &&
2329 rdev
->last_read_error
.tv_nsec
== 0) {
2330 /* first time we've seen a read error */
2331 rdev
->last_read_error
= cur_time_mon
;
2335 hours_since_last
= (cur_time_mon
.tv_sec
-
2336 rdev
->last_read_error
.tv_sec
) / 3600;
2338 rdev
->last_read_error
= cur_time_mon
;
2341 * if hours_since_last is > the number of bits in read_errors
2342 * just set read errors to 0. We do this to avoid
2343 * overflowing the shift of read_errors by hours_since_last.
2345 if (hours_since_last
>= 8 * sizeof(read_errors
))
2346 atomic_set(&rdev
->read_errors
, 0);
2348 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2351 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2352 int sectors
, struct page
*page
, int rw
)
2357 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2358 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2360 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2364 set_bit(WriteErrorSeen
, &rdev
->flags
);
2365 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2366 set_bit(MD_RECOVERY_NEEDED
,
2367 &rdev
->mddev
->recovery
);
2369 /* need to record an error - either for the block or the device */
2370 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2371 md_error(rdev
->mddev
, rdev
);
2376 * This is a kernel thread which:
2378 * 1. Retries failed read operations on working mirrors.
2379 * 2. Updates the raid superblock when problems encounter.
2380 * 3. Performs writes following reads for array synchronising.
2383 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2385 int sect
= 0; /* Offset from r10_bio->sector */
2386 int sectors
= r10_bio
->sectors
;
2387 struct md_rdev
*rdev
;
2388 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2389 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2391 /* still own a reference to this rdev, so it cannot
2392 * have been cleared recently.
2394 rdev
= conf
->mirrors
[d
].rdev
;
2396 if (test_bit(Faulty
, &rdev
->flags
))
2397 /* drive has already been failed, just ignore any
2398 more fix_read_error() attempts */
2401 check_decay_read_errors(mddev
, rdev
);
2402 atomic_inc(&rdev
->read_errors
);
2403 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2404 char b
[BDEVNAME_SIZE
];
2405 bdevname(rdev
->bdev
, b
);
2408 "md/raid10:%s: %s: Raid device exceeded "
2409 "read_error threshold [cur %d:max %d]\n",
2411 atomic_read(&rdev
->read_errors
), max_read_errors
);
2413 "md/raid10:%s: %s: Failing raid device\n",
2415 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2416 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2422 int sl
= r10_bio
->read_slot
;
2426 if (s
> (PAGE_SIZE
>>9))
2434 d
= r10_bio
->devs
[sl
].devnum
;
2435 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2437 !test_bit(Unmerged
, &rdev
->flags
) &&
2438 test_bit(In_sync
, &rdev
->flags
) &&
2439 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2440 &first_bad
, &bad_sectors
) == 0) {
2441 atomic_inc(&rdev
->nr_pending
);
2443 success
= sync_page_io(rdev
,
2444 r10_bio
->devs
[sl
].addr
+
2447 conf
->tmppage
, READ
, false);
2448 rdev_dec_pending(rdev
, mddev
);
2454 if (sl
== conf
->copies
)
2456 } while (!success
&& sl
!= r10_bio
->read_slot
);
2460 /* Cannot read from anywhere, just mark the block
2461 * as bad on the first device to discourage future
2464 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2465 rdev
= conf
->mirrors
[dn
].rdev
;
2467 if (!rdev_set_badblocks(
2469 r10_bio
->devs
[r10_bio
->read_slot
].addr
2472 md_error(mddev
, rdev
);
2473 r10_bio
->devs
[r10_bio
->read_slot
].bio
2480 /* write it back and re-read */
2482 while (sl
!= r10_bio
->read_slot
) {
2483 char b
[BDEVNAME_SIZE
];
2488 d
= r10_bio
->devs
[sl
].devnum
;
2489 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2491 test_bit(Unmerged
, &rdev
->flags
) ||
2492 !test_bit(In_sync
, &rdev
->flags
))
2495 atomic_inc(&rdev
->nr_pending
);
2497 if (r10_sync_page_io(rdev
,
2498 r10_bio
->devs
[sl
].addr
+
2500 s
, conf
->tmppage
, WRITE
)
2502 /* Well, this device is dead */
2504 "md/raid10:%s: read correction "
2506 " (%d sectors at %llu on %s)\n",
2508 (unsigned long long)(
2510 choose_data_offset(r10_bio
,
2512 bdevname(rdev
->bdev
, b
));
2513 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2516 bdevname(rdev
->bdev
, b
));
2518 rdev_dec_pending(rdev
, mddev
);
2522 while (sl
!= r10_bio
->read_slot
) {
2523 char b
[BDEVNAME_SIZE
];
2528 d
= r10_bio
->devs
[sl
].devnum
;
2529 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2531 !test_bit(In_sync
, &rdev
->flags
))
2534 atomic_inc(&rdev
->nr_pending
);
2536 switch (r10_sync_page_io(rdev
,
2537 r10_bio
->devs
[sl
].addr
+
2542 /* Well, this device is dead */
2544 "md/raid10:%s: unable to read back "
2546 " (%d sectors at %llu on %s)\n",
2548 (unsigned long long)(
2550 choose_data_offset(r10_bio
, rdev
)),
2551 bdevname(rdev
->bdev
, b
));
2552 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2555 bdevname(rdev
->bdev
, b
));
2559 "md/raid10:%s: read error corrected"
2560 " (%d sectors at %llu on %s)\n",
2562 (unsigned long long)(
2564 choose_data_offset(r10_bio
, rdev
)),
2565 bdevname(rdev
->bdev
, b
));
2566 atomic_add(s
, &rdev
->corrected_errors
);
2569 rdev_dec_pending(rdev
, mddev
);
2579 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2581 struct bio
*bio
= r10_bio
->master_bio
;
2582 struct mddev
*mddev
= r10_bio
->mddev
;
2583 struct r10conf
*conf
= mddev
->private;
2584 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2585 /* bio has the data to be written to slot 'i' where
2586 * we just recently had a write error.
2587 * We repeatedly clone the bio and trim down to one block,
2588 * then try the write. Where the write fails we record
2590 * It is conceivable that the bio doesn't exactly align with
2591 * blocks. We must handle this.
2593 * We currently own a reference to the rdev.
2599 int sect_to_write
= r10_bio
->sectors
;
2602 if (rdev
->badblocks
.shift
< 0)
2605 block_sectors
= 1 << rdev
->badblocks
.shift
;
2606 sector
= r10_bio
->sector
;
2607 sectors
= ((r10_bio
->sector
+ block_sectors
)
2608 & ~(sector_t
)(block_sectors
- 1))
2611 while (sect_to_write
) {
2613 if (sectors
> sect_to_write
)
2614 sectors
= sect_to_write
;
2615 /* Write at 'sector' for 'sectors' */
2616 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2617 bio_trim(wbio
, sector
- bio
->bi_sector
, sectors
);
2618 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2619 choose_data_offset(r10_bio
, rdev
) +
2620 (sector
- r10_bio
->sector
));
2621 wbio
->bi_bdev
= rdev
->bdev
;
2622 if (submit_bio_wait(WRITE
, wbio
) == 0)
2624 ok
= rdev_set_badblocks(rdev
, sector
,
2629 sect_to_write
-= sectors
;
2631 sectors
= block_sectors
;
2636 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2638 int slot
= r10_bio
->read_slot
;
2640 struct r10conf
*conf
= mddev
->private;
2641 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2642 char b
[BDEVNAME_SIZE
];
2643 unsigned long do_sync
;
2646 /* we got a read error. Maybe the drive is bad. Maybe just
2647 * the block and we can fix it.
2648 * We freeze all other IO, and try reading the block from
2649 * other devices. When we find one, we re-write
2650 * and check it that fixes the read error.
2651 * This is all done synchronously while the array is
2654 bio
= r10_bio
->devs
[slot
].bio
;
2655 bdevname(bio
->bi_bdev
, b
);
2657 r10_bio
->devs
[slot
].bio
= NULL
;
2659 if (mddev
->ro
== 0) {
2660 freeze_array(conf
, 1);
2661 fix_read_error(conf
, mddev
, r10_bio
);
2662 unfreeze_array(conf
);
2664 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2666 rdev_dec_pending(rdev
, mddev
);
2669 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2671 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2672 " read error for block %llu\n",
2674 (unsigned long long)r10_bio
->sector
);
2675 raid_end_bio_io(r10_bio
);
2679 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2680 slot
= r10_bio
->read_slot
;
2683 "md/raid10:%s: %s: redirecting "
2684 "sector %llu to another mirror\n",
2686 bdevname(rdev
->bdev
, b
),
2687 (unsigned long long)r10_bio
->sector
);
2688 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2690 bio_trim(bio
, r10_bio
->sector
- bio
->bi_sector
, max_sectors
);
2691 r10_bio
->devs
[slot
].bio
= bio
;
2692 r10_bio
->devs
[slot
].rdev
= rdev
;
2693 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2694 + choose_data_offset(r10_bio
, rdev
);
2695 bio
->bi_bdev
= rdev
->bdev
;
2696 bio
->bi_rw
= READ
| do_sync
;
2697 bio
->bi_private
= r10_bio
;
2698 bio
->bi_end_io
= raid10_end_read_request
;
2699 if (max_sectors
< r10_bio
->sectors
) {
2700 /* Drat - have to split this up more */
2701 struct bio
*mbio
= r10_bio
->master_bio
;
2702 int sectors_handled
=
2703 r10_bio
->sector
+ max_sectors
2705 r10_bio
->sectors
= max_sectors
;
2706 spin_lock_irq(&conf
->device_lock
);
2707 if (mbio
->bi_phys_segments
== 0)
2708 mbio
->bi_phys_segments
= 2;
2710 mbio
->bi_phys_segments
++;
2711 spin_unlock_irq(&conf
->device_lock
);
2712 generic_make_request(bio
);
2714 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2716 r10_bio
->master_bio
= mbio
;
2717 r10_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2719 set_bit(R10BIO_ReadError
,
2721 r10_bio
->mddev
= mddev
;
2722 r10_bio
->sector
= mbio
->bi_sector
2727 generic_make_request(bio
);
2730 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2732 /* Some sort of write request has finished and it
2733 * succeeded in writing where we thought there was a
2734 * bad block. So forget the bad block.
2735 * Or possibly if failed and we need to record
2739 struct md_rdev
*rdev
;
2741 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2742 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2743 for (m
= 0; m
< conf
->copies
; m
++) {
2744 int dev
= r10_bio
->devs
[m
].devnum
;
2745 rdev
= conf
->mirrors
[dev
].rdev
;
2746 if (r10_bio
->devs
[m
].bio
== NULL
)
2748 if (test_bit(BIO_UPTODATE
,
2749 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2750 rdev_clear_badblocks(
2752 r10_bio
->devs
[m
].addr
,
2753 r10_bio
->sectors
, 0);
2755 if (!rdev_set_badblocks(
2757 r10_bio
->devs
[m
].addr
,
2758 r10_bio
->sectors
, 0))
2759 md_error(conf
->mddev
, rdev
);
2761 rdev
= conf
->mirrors
[dev
].replacement
;
2762 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2764 if (test_bit(BIO_UPTODATE
,
2765 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2766 rdev_clear_badblocks(
2768 r10_bio
->devs
[m
].addr
,
2769 r10_bio
->sectors
, 0);
2771 if (!rdev_set_badblocks(
2773 r10_bio
->devs
[m
].addr
,
2774 r10_bio
->sectors
, 0))
2775 md_error(conf
->mddev
, rdev
);
2780 for (m
= 0; m
< conf
->copies
; m
++) {
2781 int dev
= r10_bio
->devs
[m
].devnum
;
2782 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2783 rdev
= conf
->mirrors
[dev
].rdev
;
2784 if (bio
== IO_MADE_GOOD
) {
2785 rdev_clear_badblocks(
2787 r10_bio
->devs
[m
].addr
,
2788 r10_bio
->sectors
, 0);
2789 rdev_dec_pending(rdev
, conf
->mddev
);
2790 } else if (bio
!= NULL
&&
2791 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2792 if (!narrow_write_error(r10_bio
, m
)) {
2793 md_error(conf
->mddev
, rdev
);
2794 set_bit(R10BIO_Degraded
,
2797 rdev_dec_pending(rdev
, conf
->mddev
);
2799 bio
= r10_bio
->devs
[m
].repl_bio
;
2800 rdev
= conf
->mirrors
[dev
].replacement
;
2801 if (rdev
&& bio
== IO_MADE_GOOD
) {
2802 rdev_clear_badblocks(
2804 r10_bio
->devs
[m
].addr
,
2805 r10_bio
->sectors
, 0);
2806 rdev_dec_pending(rdev
, conf
->mddev
);
2809 if (test_bit(R10BIO_WriteError
,
2811 close_write(r10_bio
);
2812 raid_end_bio_io(r10_bio
);
2816 static void raid10d(struct md_thread
*thread
)
2818 struct mddev
*mddev
= thread
->mddev
;
2819 struct r10bio
*r10_bio
;
2820 unsigned long flags
;
2821 struct r10conf
*conf
= mddev
->private;
2822 struct list_head
*head
= &conf
->retry_list
;
2823 struct blk_plug plug
;
2825 md_check_recovery(mddev
);
2827 blk_start_plug(&plug
);
2830 flush_pending_writes(conf
);
2832 spin_lock_irqsave(&conf
->device_lock
, flags
);
2833 if (list_empty(head
)) {
2834 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2837 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2838 list_del(head
->prev
);
2840 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2842 mddev
= r10_bio
->mddev
;
2843 conf
= mddev
->private;
2844 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2845 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2846 handle_write_completed(conf
, r10_bio
);
2847 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2848 reshape_request_write(mddev
, r10_bio
);
2849 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2850 sync_request_write(mddev
, r10_bio
);
2851 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2852 recovery_request_write(mddev
, r10_bio
);
2853 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2854 handle_read_error(mddev
, r10_bio
);
2856 /* just a partial read to be scheduled from a
2859 int slot
= r10_bio
->read_slot
;
2860 generic_make_request(r10_bio
->devs
[slot
].bio
);
2864 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2865 md_check_recovery(mddev
);
2867 blk_finish_plug(&plug
);
2871 static int init_resync(struct r10conf
*conf
)
2876 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2877 BUG_ON(conf
->r10buf_pool
);
2878 conf
->have_replacement
= 0;
2879 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2880 if (conf
->mirrors
[i
].replacement
)
2881 conf
->have_replacement
= 1;
2882 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2883 if (!conf
->r10buf_pool
)
2885 conf
->next_resync
= 0;
2890 * perform a "sync" on one "block"
2892 * We need to make sure that no normal I/O request - particularly write
2893 * requests - conflict with active sync requests.
2895 * This is achieved by tracking pending requests and a 'barrier' concept
2896 * that can be installed to exclude normal IO requests.
2898 * Resync and recovery are handled very differently.
2899 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2901 * For resync, we iterate over virtual addresses, read all copies,
2902 * and update if there are differences. If only one copy is live,
2904 * For recovery, we iterate over physical addresses, read a good
2905 * value for each non-in_sync drive, and over-write.
2907 * So, for recovery we may have several outstanding complex requests for a
2908 * given address, one for each out-of-sync device. We model this by allocating
2909 * a number of r10_bio structures, one for each out-of-sync device.
2910 * As we setup these structures, we collect all bio's together into a list
2911 * which we then process collectively to add pages, and then process again
2912 * to pass to generic_make_request.
2914 * The r10_bio structures are linked using a borrowed master_bio pointer.
2915 * This link is counted in ->remaining. When the r10_bio that points to NULL
2916 * has its remaining count decremented to 0, the whole complex operation
2921 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2922 int *skipped
, int go_faster
)
2924 struct r10conf
*conf
= mddev
->private;
2925 struct r10bio
*r10_bio
;
2926 struct bio
*biolist
= NULL
, *bio
;
2927 sector_t max_sector
, nr_sectors
;
2930 sector_t sync_blocks
;
2931 sector_t sectors_skipped
= 0;
2932 int chunks_skipped
= 0;
2933 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2935 if (!conf
->r10buf_pool
)
2936 if (init_resync(conf
))
2940 * Allow skipping a full rebuild for incremental assembly
2941 * of a clean array, like RAID1 does.
2943 if (mddev
->bitmap
== NULL
&&
2944 mddev
->recovery_cp
== MaxSector
&&
2945 mddev
->reshape_position
== MaxSector
&&
2946 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2947 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2948 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2949 conf
->fullsync
== 0) {
2951 return mddev
->dev_sectors
- sector_nr
;
2955 max_sector
= mddev
->dev_sectors
;
2956 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2957 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2958 max_sector
= mddev
->resync_max_sectors
;
2959 if (sector_nr
>= max_sector
) {
2960 /* If we aborted, we need to abort the
2961 * sync on the 'current' bitmap chucks (there can
2962 * be several when recovering multiple devices).
2963 * as we may have started syncing it but not finished.
2964 * We can find the current address in
2965 * mddev->curr_resync, but for recovery,
2966 * we need to convert that to several
2967 * virtual addresses.
2969 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2974 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2975 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2976 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2978 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2980 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2981 bitmap_end_sync(mddev
->bitmap
, sect
,
2985 /* completed sync */
2986 if ((!mddev
->bitmap
|| conf
->fullsync
)
2987 && conf
->have_replacement
2988 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2989 /* Completed a full sync so the replacements
2990 * are now fully recovered.
2992 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2993 if (conf
->mirrors
[i
].replacement
)
2994 conf
->mirrors
[i
].replacement
3000 bitmap_close_sync(mddev
->bitmap
);
3003 return sectors_skipped
;
3006 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
3007 return reshape_request(mddev
, sector_nr
, skipped
);
3009 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
3010 /* if there has been nothing to do on any drive,
3011 * then there is nothing to do at all..
3014 return (max_sector
- sector_nr
) + sectors_skipped
;
3017 if (max_sector
> mddev
->resync_max
)
3018 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
3020 /* make sure whole request will fit in a chunk - if chunks
3023 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
3024 max_sector
> (sector_nr
| chunk_mask
))
3025 max_sector
= (sector_nr
| chunk_mask
) + 1;
3027 * If there is non-resync activity waiting for us then
3028 * put in a delay to throttle resync.
3030 if (!go_faster
&& conf
->nr_waiting
)
3031 msleep_interruptible(1000);
3033 /* Again, very different code for resync and recovery.
3034 * Both must result in an r10bio with a list of bios that
3035 * have bi_end_io, bi_sector, bi_bdev set,
3036 * and bi_private set to the r10bio.
3037 * For recovery, we may actually create several r10bios
3038 * with 2 bios in each, that correspond to the bios in the main one.
3039 * In this case, the subordinate r10bios link back through a
3040 * borrowed master_bio pointer, and the counter in the master
3041 * includes a ref from each subordinate.
3043 /* First, we decide what to do and set ->bi_end_io
3044 * To end_sync_read if we want to read, and
3045 * end_sync_write if we will want to write.
3048 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
3049 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3050 /* recovery... the complicated one */
3054 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
3060 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
3062 if ((mirror
->rdev
== NULL
||
3063 test_bit(In_sync
, &mirror
->rdev
->flags
))
3065 (mirror
->replacement
== NULL
||
3067 &mirror
->replacement
->flags
)))
3071 /* want to reconstruct this device */
3073 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3074 if (sect
>= mddev
->resync_max_sectors
) {
3075 /* last stripe is not complete - don't
3076 * try to recover this sector.
3080 /* Unless we are doing a full sync, or a replacement
3081 * we only need to recover the block if it is set in
3084 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3086 if (sync_blocks
< max_sync
)
3087 max_sync
= sync_blocks
;
3089 mirror
->replacement
== NULL
&&
3091 /* yep, skip the sync_blocks here, but don't assume
3092 * that there will never be anything to do here
3094 chunks_skipped
= -1;
3098 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3099 raise_barrier(conf
, rb2
!= NULL
);
3100 atomic_set(&r10_bio
->remaining
, 0);
3102 r10_bio
->master_bio
= (struct bio
*)rb2
;
3104 atomic_inc(&rb2
->remaining
);
3105 r10_bio
->mddev
= mddev
;
3106 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3107 r10_bio
->sector
= sect
;
3109 raid10_find_phys(conf
, r10_bio
);
3111 /* Need to check if the array will still be
3114 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3115 if (conf
->mirrors
[j
].rdev
== NULL
||
3116 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3121 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3122 &sync_blocks
, still_degraded
);
3125 for (j
=0; j
<conf
->copies
;j
++) {
3127 int d
= r10_bio
->devs
[j
].devnum
;
3128 sector_t from_addr
, to_addr
;
3129 struct md_rdev
*rdev
;
3130 sector_t sector
, first_bad
;
3132 if (!conf
->mirrors
[d
].rdev
||
3133 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3135 /* This is where we read from */
3137 rdev
= conf
->mirrors
[d
].rdev
;
3138 sector
= r10_bio
->devs
[j
].addr
;
3140 if (is_badblock(rdev
, sector
, max_sync
,
3141 &first_bad
, &bad_sectors
)) {
3142 if (first_bad
> sector
)
3143 max_sync
= first_bad
- sector
;
3145 bad_sectors
-= (sector
3147 if (max_sync
> bad_sectors
)
3148 max_sync
= bad_sectors
;
3152 bio
= r10_bio
->devs
[0].bio
;
3154 bio
->bi_next
= biolist
;
3156 bio
->bi_private
= r10_bio
;
3157 bio
->bi_end_io
= end_sync_read
;
3159 from_addr
= r10_bio
->devs
[j
].addr
;
3160 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
3161 bio
->bi_bdev
= rdev
->bdev
;
3162 atomic_inc(&rdev
->nr_pending
);
3163 /* and we write to 'i' (if not in_sync) */
3165 for (k
=0; k
<conf
->copies
; k
++)
3166 if (r10_bio
->devs
[k
].devnum
== i
)
3168 BUG_ON(k
== conf
->copies
);
3169 to_addr
= r10_bio
->devs
[k
].addr
;
3170 r10_bio
->devs
[0].devnum
= d
;
3171 r10_bio
->devs
[0].addr
= from_addr
;
3172 r10_bio
->devs
[1].devnum
= i
;
3173 r10_bio
->devs
[1].addr
= to_addr
;
3175 rdev
= mirror
->rdev
;
3176 if (!test_bit(In_sync
, &rdev
->flags
)) {
3177 bio
= r10_bio
->devs
[1].bio
;
3179 bio
->bi_next
= biolist
;
3181 bio
->bi_private
= r10_bio
;
3182 bio
->bi_end_io
= end_sync_write
;
3184 bio
->bi_sector
= to_addr
3185 + rdev
->data_offset
;
3186 bio
->bi_bdev
= rdev
->bdev
;
3187 atomic_inc(&r10_bio
->remaining
);
3189 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3191 /* and maybe write to replacement */
3192 bio
= r10_bio
->devs
[1].repl_bio
;
3194 bio
->bi_end_io
= NULL
;
3195 rdev
= mirror
->replacement
;
3196 /* Note: if rdev != NULL, then bio
3197 * cannot be NULL as r10buf_pool_alloc will
3198 * have allocated it.
3199 * So the second test here is pointless.
3200 * But it keeps semantic-checkers happy, and
3201 * this comment keeps human reviewers
3204 if (rdev
== NULL
|| bio
== NULL
||
3205 test_bit(Faulty
, &rdev
->flags
))
3208 bio
->bi_next
= biolist
;
3210 bio
->bi_private
= r10_bio
;
3211 bio
->bi_end_io
= end_sync_write
;
3213 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3214 bio
->bi_bdev
= rdev
->bdev
;
3215 atomic_inc(&r10_bio
->remaining
);
3218 if (j
== conf
->copies
) {
3219 /* Cannot recover, so abort the recovery or
3220 * record a bad block */
3223 atomic_dec(&rb2
->remaining
);
3226 /* problem is that there are bad blocks
3227 * on other device(s)
3230 for (k
= 0; k
< conf
->copies
; k
++)
3231 if (r10_bio
->devs
[k
].devnum
== i
)
3233 if (!test_bit(In_sync
,
3234 &mirror
->rdev
->flags
)
3235 && !rdev_set_badblocks(
3237 r10_bio
->devs
[k
].addr
,
3240 if (mirror
->replacement
&&
3241 !rdev_set_badblocks(
3242 mirror
->replacement
,
3243 r10_bio
->devs
[k
].addr
,
3248 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3250 printk(KERN_INFO
"md/raid10:%s: insufficient "
3251 "working devices for recovery.\n",
3253 mirror
->recovery_disabled
3254 = mddev
->recovery_disabled
;
3259 if (biolist
== NULL
) {
3261 struct r10bio
*rb2
= r10_bio
;
3262 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3263 rb2
->master_bio
= NULL
;
3269 /* resync. Schedule a read for every block at this virt offset */
3272 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3274 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3275 &sync_blocks
, mddev
->degraded
) &&
3276 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3277 &mddev
->recovery
)) {
3278 /* We can skip this block */
3280 return sync_blocks
+ sectors_skipped
;
3282 if (sync_blocks
< max_sync
)
3283 max_sync
= sync_blocks
;
3284 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3286 r10_bio
->mddev
= mddev
;
3287 atomic_set(&r10_bio
->remaining
, 0);
3288 raise_barrier(conf
, 0);
3289 conf
->next_resync
= sector_nr
;
3291 r10_bio
->master_bio
= NULL
;
3292 r10_bio
->sector
= sector_nr
;
3293 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3294 raid10_find_phys(conf
, r10_bio
);
3295 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3297 for (i
= 0; i
< conf
->copies
; i
++) {
3298 int d
= r10_bio
->devs
[i
].devnum
;
3299 sector_t first_bad
, sector
;
3302 if (r10_bio
->devs
[i
].repl_bio
)
3303 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3305 bio
= r10_bio
->devs
[i
].bio
;
3307 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3308 if (conf
->mirrors
[d
].rdev
== NULL
||
3309 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3311 sector
= r10_bio
->devs
[i
].addr
;
3312 if (is_badblock(conf
->mirrors
[d
].rdev
,
3314 &first_bad
, &bad_sectors
)) {
3315 if (first_bad
> sector
)
3316 max_sync
= first_bad
- sector
;
3318 bad_sectors
-= (sector
- first_bad
);
3319 if (max_sync
> bad_sectors
)
3320 max_sync
= bad_sectors
;
3324 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3325 atomic_inc(&r10_bio
->remaining
);
3326 bio
->bi_next
= biolist
;
3328 bio
->bi_private
= r10_bio
;
3329 bio
->bi_end_io
= end_sync_read
;
3331 bio
->bi_sector
= sector
+
3332 conf
->mirrors
[d
].rdev
->data_offset
;
3333 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3336 if (conf
->mirrors
[d
].replacement
== NULL
||
3338 &conf
->mirrors
[d
].replacement
->flags
))
3341 /* Need to set up for writing to the replacement */
3342 bio
= r10_bio
->devs
[i
].repl_bio
;
3344 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3346 sector
= r10_bio
->devs
[i
].addr
;
3347 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3348 bio
->bi_next
= biolist
;
3350 bio
->bi_private
= r10_bio
;
3351 bio
->bi_end_io
= end_sync_write
;
3353 bio
->bi_sector
= sector
+
3354 conf
->mirrors
[d
].replacement
->data_offset
;
3355 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3360 for (i
=0; i
<conf
->copies
; i
++) {
3361 int d
= r10_bio
->devs
[i
].devnum
;
3362 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3363 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3365 if (r10_bio
->devs
[i
].repl_bio
&&
3366 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3368 conf
->mirrors
[d
].replacement
,
3378 if (sector_nr
+ max_sync
< max_sector
)
3379 max_sector
= sector_nr
+ max_sync
;
3382 int len
= PAGE_SIZE
;
3383 if (sector_nr
+ (len
>>9) > max_sector
)
3384 len
= (max_sector
- sector_nr
) << 9;
3387 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3389 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3390 if (bio_add_page(bio
, page
, len
, 0))
3394 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3395 for (bio2
= biolist
;
3396 bio2
&& bio2
!= bio
;
3397 bio2
= bio2
->bi_next
) {
3398 /* remove last page from this bio */
3400 bio2
->bi_size
-= len
;
3401 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3405 nr_sectors
+= len
>>9;
3406 sector_nr
+= len
>>9;
3407 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3409 r10_bio
->sectors
= nr_sectors
;
3413 biolist
= biolist
->bi_next
;
3415 bio
->bi_next
= NULL
;
3416 r10_bio
= bio
->bi_private
;
3417 r10_bio
->sectors
= nr_sectors
;
3419 if (bio
->bi_end_io
== end_sync_read
) {
3420 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3421 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3422 generic_make_request(bio
);
3426 if (sectors_skipped
)
3427 /* pretend they weren't skipped, it makes
3428 * no important difference in this case
3430 md_done_sync(mddev
, sectors_skipped
, 1);
3432 return sectors_skipped
+ nr_sectors
;
3434 /* There is nowhere to write, so all non-sync
3435 * drives must be failed or in resync, all drives
3436 * have a bad block, so try the next chunk...
3438 if (sector_nr
+ max_sync
< max_sector
)
3439 max_sector
= sector_nr
+ max_sync
;
3441 sectors_skipped
+= (max_sector
- sector_nr
);
3443 sector_nr
= max_sector
;
3448 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3451 struct r10conf
*conf
= mddev
->private;
3454 raid_disks
= min(conf
->geo
.raid_disks
,
3455 conf
->prev
.raid_disks
);
3457 sectors
= conf
->dev_sectors
;
3459 size
= sectors
>> conf
->geo
.chunk_shift
;
3460 sector_div(size
, conf
->geo
.far_copies
);
3461 size
= size
* raid_disks
;
3462 sector_div(size
, conf
->geo
.near_copies
);
3464 return size
<< conf
->geo
.chunk_shift
;
3467 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3469 /* Calculate the number of sectors-per-device that will
3470 * actually be used, and set conf->dev_sectors and
3474 size
= size
>> conf
->geo
.chunk_shift
;
3475 sector_div(size
, conf
->geo
.far_copies
);
3476 size
= size
* conf
->geo
.raid_disks
;
3477 sector_div(size
, conf
->geo
.near_copies
);
3478 /* 'size' is now the number of chunks in the array */
3479 /* calculate "used chunks per device" */
3480 size
= size
* conf
->copies
;
3482 /* We need to round up when dividing by raid_disks to
3483 * get the stride size.
3485 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3487 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3489 if (conf
->geo
.far_offset
)
3490 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3492 sector_div(size
, conf
->geo
.far_copies
);
3493 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3497 enum geo_type
{geo_new
, geo_old
, geo_start
};
3498 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3501 int layout
, chunk
, disks
;
3504 layout
= mddev
->layout
;
3505 chunk
= mddev
->chunk_sectors
;
3506 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3509 layout
= mddev
->new_layout
;
3510 chunk
= mddev
->new_chunk_sectors
;
3511 disks
= mddev
->raid_disks
;
3513 default: /* avoid 'may be unused' warnings */
3514 case geo_start
: /* new when starting reshape - raid_disks not
3516 layout
= mddev
->new_layout
;
3517 chunk
= mddev
->new_chunk_sectors
;
3518 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3523 if (chunk
< (PAGE_SIZE
>> 9) ||
3524 !is_power_of_2(chunk
))
3527 fc
= (layout
>> 8) & 255;
3528 fo
= layout
& (1<<16);
3529 geo
->raid_disks
= disks
;
3530 geo
->near_copies
= nc
;
3531 geo
->far_copies
= fc
;
3532 geo
->far_offset
= fo
;
3533 geo
->far_set_size
= (layout
& (1<<17)) ? disks
/ fc
: disks
;
3534 geo
->chunk_mask
= chunk
- 1;
3535 geo
->chunk_shift
= ffz(~chunk
);
3539 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3541 struct r10conf
*conf
= NULL
;
3546 copies
= setup_geo(&geo
, mddev
, geo_new
);
3549 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3550 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3551 mdname(mddev
), PAGE_SIZE
);
3555 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3556 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3557 mdname(mddev
), mddev
->new_layout
);
3562 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3566 /* FIXME calc properly */
3567 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3568 max(0,-mddev
->delta_disks
)),
3573 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3578 conf
->copies
= copies
;
3579 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3580 r10bio_pool_free
, conf
);
3581 if (!conf
->r10bio_pool
)
3584 calc_sectors(conf
, mddev
->dev_sectors
);
3585 if (mddev
->reshape_position
== MaxSector
) {
3586 conf
->prev
= conf
->geo
;
3587 conf
->reshape_progress
= MaxSector
;
3589 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3593 conf
->reshape_progress
= mddev
->reshape_position
;
3594 if (conf
->prev
.far_offset
)
3595 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3597 /* far_copies must be 1 */
3598 conf
->prev
.stride
= conf
->dev_sectors
;
3600 spin_lock_init(&conf
->device_lock
);
3601 INIT_LIST_HEAD(&conf
->retry_list
);
3603 spin_lock_init(&conf
->resync_lock
);
3604 init_waitqueue_head(&conf
->wait_barrier
);
3606 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3610 conf
->mddev
= mddev
;
3615 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3618 if (conf
->r10bio_pool
)
3619 mempool_destroy(conf
->r10bio_pool
);
3620 kfree(conf
->mirrors
);
3621 safe_put_page(conf
->tmppage
);
3624 return ERR_PTR(err
);
3627 static int run(struct mddev
*mddev
)
3629 struct r10conf
*conf
;
3630 int i
, disk_idx
, chunk_size
;
3631 struct raid10_info
*disk
;
3632 struct md_rdev
*rdev
;
3634 sector_t min_offset_diff
= 0;
3636 bool discard_supported
= false;
3638 if (mddev
->private == NULL
) {
3639 conf
= setup_conf(mddev
);
3641 return PTR_ERR(conf
);
3642 mddev
->private = conf
;
3644 conf
= mddev
->private;
3648 mddev
->thread
= conf
->thread
;
3649 conf
->thread
= NULL
;
3651 chunk_size
= mddev
->chunk_sectors
<< 9;
3653 blk_queue_max_discard_sectors(mddev
->queue
,
3654 mddev
->chunk_sectors
);
3655 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3656 blk_queue_io_min(mddev
->queue
, chunk_size
);
3657 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3658 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3660 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3661 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3664 rdev_for_each(rdev
, mddev
) {
3666 struct request_queue
*q
;
3668 disk_idx
= rdev
->raid_disk
;
3671 if (disk_idx
>= conf
->geo
.raid_disks
&&
3672 disk_idx
>= conf
->prev
.raid_disks
)
3674 disk
= conf
->mirrors
+ disk_idx
;
3676 if (test_bit(Replacement
, &rdev
->flags
)) {
3677 if (disk
->replacement
)
3679 disk
->replacement
= rdev
;
3685 q
= bdev_get_queue(rdev
->bdev
);
3686 if (q
->merge_bvec_fn
)
3687 mddev
->merge_check_needed
= 1;
3688 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3689 if (!mddev
->reshape_backwards
)
3693 if (first
|| diff
< min_offset_diff
)
3694 min_offset_diff
= diff
;
3697 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3698 rdev
->data_offset
<< 9);
3700 disk
->head_position
= 0;
3702 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3703 discard_supported
= true;
3707 if (discard_supported
)
3708 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3711 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3714 /* need to check that every block has at least one working mirror */
3715 if (!enough(conf
, -1)) {
3716 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3721 if (conf
->reshape_progress
!= MaxSector
) {
3722 /* must ensure that shape change is supported */
3723 if (conf
->geo
.far_copies
!= 1 &&
3724 conf
->geo
.far_offset
== 0)
3726 if (conf
->prev
.far_copies
!= 1 &&
3727 conf
->prev
.far_offset
== 0)
3731 mddev
->degraded
= 0;
3733 i
< conf
->geo
.raid_disks
3734 || i
< conf
->prev
.raid_disks
;
3737 disk
= conf
->mirrors
+ i
;
3739 if (!disk
->rdev
&& disk
->replacement
) {
3740 /* The replacement is all we have - use it */
3741 disk
->rdev
= disk
->replacement
;
3742 disk
->replacement
= NULL
;
3743 clear_bit(Replacement
, &disk
->rdev
->flags
);
3747 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3748 disk
->head_position
= 0;
3753 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3756 if (mddev
->recovery_cp
!= MaxSector
)
3757 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3758 " -- starting background reconstruction\n",
3761 "md/raid10:%s: active with %d out of %d devices\n",
3762 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3763 conf
->geo
.raid_disks
);
3765 * Ok, everything is just fine now
3767 mddev
->dev_sectors
= conf
->dev_sectors
;
3768 size
= raid10_size(mddev
, 0, 0);
3769 md_set_array_sectors(mddev
, size
);
3770 mddev
->resync_max_sectors
= size
;
3773 int stripe
= conf
->geo
.raid_disks
*
3774 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3775 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3776 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3778 /* Calculate max read-ahead size.
3779 * We need to readahead at least twice a whole stripe....
3782 stripe
/= conf
->geo
.near_copies
;
3783 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3784 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3785 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3789 if (md_integrity_register(mddev
))
3792 if (conf
->reshape_progress
!= MaxSector
) {
3793 unsigned long before_length
, after_length
;
3795 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3796 conf
->prev
.far_copies
);
3797 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3798 conf
->geo
.far_copies
);
3800 if (max(before_length
, after_length
) > min_offset_diff
) {
3801 /* This cannot work */
3802 printk("md/raid10: offset difference not enough to continue reshape\n");
3805 conf
->offset_diff
= min_offset_diff
;
3807 conf
->reshape_safe
= conf
->reshape_progress
;
3808 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3809 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3810 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3811 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3812 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3819 md_unregister_thread(&mddev
->thread
);
3820 if (conf
->r10bio_pool
)
3821 mempool_destroy(conf
->r10bio_pool
);
3822 safe_put_page(conf
->tmppage
);
3823 kfree(conf
->mirrors
);
3825 mddev
->private = NULL
;
3830 static int stop(struct mddev
*mddev
)
3832 struct r10conf
*conf
= mddev
->private;
3834 raise_barrier(conf
, 0);
3835 lower_barrier(conf
);
3837 md_unregister_thread(&mddev
->thread
);
3839 /* the unplug fn references 'conf'*/
3840 blk_sync_queue(mddev
->queue
);
3842 if (conf
->r10bio_pool
)
3843 mempool_destroy(conf
->r10bio_pool
);
3844 safe_put_page(conf
->tmppage
);
3845 kfree(conf
->mirrors
);
3847 mddev
->private = NULL
;
3851 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3853 struct r10conf
*conf
= mddev
->private;
3857 raise_barrier(conf
, 0);
3860 lower_barrier(conf
);
3865 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3867 /* Resize of 'far' arrays is not supported.
3868 * For 'near' and 'offset' arrays we can set the
3869 * number of sectors used to be an appropriate multiple
3870 * of the chunk size.
3871 * For 'offset', this is far_copies*chunksize.
3872 * For 'near' the multiplier is the LCM of
3873 * near_copies and raid_disks.
3874 * So if far_copies > 1 && !far_offset, fail.
3875 * Else find LCM(raid_disks, near_copy)*far_copies and
3876 * multiply by chunk_size. Then round to this number.
3877 * This is mostly done by raid10_size()
3879 struct r10conf
*conf
= mddev
->private;
3880 sector_t oldsize
, size
;
3882 if (mddev
->reshape_position
!= MaxSector
)
3885 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3888 oldsize
= raid10_size(mddev
, 0, 0);
3889 size
= raid10_size(mddev
, sectors
, 0);
3890 if (mddev
->external_size
&&
3891 mddev
->array_sectors
> size
)
3893 if (mddev
->bitmap
) {
3894 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3898 md_set_array_sectors(mddev
, size
);
3899 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3900 revalidate_disk(mddev
->gendisk
);
3901 if (sectors
> mddev
->dev_sectors
&&
3902 mddev
->recovery_cp
> oldsize
) {
3903 mddev
->recovery_cp
= oldsize
;
3904 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3906 calc_sectors(conf
, sectors
);
3907 mddev
->dev_sectors
= conf
->dev_sectors
;
3908 mddev
->resync_max_sectors
= size
;
3912 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3914 struct md_rdev
*rdev
;
3915 struct r10conf
*conf
;
3917 if (mddev
->degraded
> 0) {
3918 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3920 return ERR_PTR(-EINVAL
);
3923 /* Set new parameters */
3924 mddev
->new_level
= 10;
3925 /* new layout: far_copies = 1, near_copies = 2 */
3926 mddev
->new_layout
= (1<<8) + 2;
3927 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3928 mddev
->delta_disks
= mddev
->raid_disks
;
3929 mddev
->raid_disks
*= 2;
3930 /* make sure it will be not marked as dirty */
3931 mddev
->recovery_cp
= MaxSector
;
3933 conf
= setup_conf(mddev
);
3934 if (!IS_ERR(conf
)) {
3935 rdev_for_each(rdev
, mddev
)
3936 if (rdev
->raid_disk
>= 0)
3937 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3944 static void *raid10_takeover(struct mddev
*mddev
)
3946 struct r0conf
*raid0_conf
;
3948 /* raid10 can take over:
3949 * raid0 - providing it has only two drives
3951 if (mddev
->level
== 0) {
3952 /* for raid0 takeover only one zone is supported */
3953 raid0_conf
= mddev
->private;
3954 if (raid0_conf
->nr_strip_zones
> 1) {
3955 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3956 " with more than one zone.\n",
3958 return ERR_PTR(-EINVAL
);
3960 return raid10_takeover_raid0(mddev
);
3962 return ERR_PTR(-EINVAL
);
3965 static int raid10_check_reshape(struct mddev
*mddev
)
3967 /* Called when there is a request to change
3968 * - layout (to ->new_layout)
3969 * - chunk size (to ->new_chunk_sectors)
3970 * - raid_disks (by delta_disks)
3971 * or when trying to restart a reshape that was ongoing.
3973 * We need to validate the request and possibly allocate
3974 * space if that might be an issue later.
3976 * Currently we reject any reshape of a 'far' mode array,
3977 * allow chunk size to change if new is generally acceptable,
3978 * allow raid_disks to increase, and allow
3979 * a switch between 'near' mode and 'offset' mode.
3981 struct r10conf
*conf
= mddev
->private;
3984 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3987 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3988 /* mustn't change number of copies */
3990 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3991 /* Cannot switch to 'far' mode */
3994 if (mddev
->array_sectors
& geo
.chunk_mask
)
3995 /* not factor of array size */
3998 if (!enough(conf
, -1))
4001 kfree(conf
->mirrors_new
);
4002 conf
->mirrors_new
= NULL
;
4003 if (mddev
->delta_disks
> 0) {
4004 /* allocate new 'mirrors' list */
4005 conf
->mirrors_new
= kzalloc(
4006 sizeof(struct raid10_info
)
4007 *(mddev
->raid_disks
+
4008 mddev
->delta_disks
),
4010 if (!conf
->mirrors_new
)
4017 * Need to check if array has failed when deciding whether to:
4019 * - remove non-faulty devices
4022 * This determination is simple when no reshape is happening.
4023 * However if there is a reshape, we need to carefully check
4024 * both the before and after sections.
4025 * This is because some failed devices may only affect one
4026 * of the two sections, and some non-in_sync devices may
4027 * be insync in the section most affected by failed devices.
4029 static int calc_degraded(struct r10conf
*conf
)
4031 int degraded
, degraded2
;
4036 /* 'prev' section first */
4037 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4038 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4039 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4041 else if (!test_bit(In_sync
, &rdev
->flags
))
4042 /* When we can reduce the number of devices in
4043 * an array, this might not contribute to
4044 * 'degraded'. It does now.
4049 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4053 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4054 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4055 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4057 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4058 /* If reshape is increasing the number of devices,
4059 * this section has already been recovered, so
4060 * it doesn't contribute to degraded.
4063 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4068 if (degraded2
> degraded
)
4073 static int raid10_start_reshape(struct mddev
*mddev
)
4075 /* A 'reshape' has been requested. This commits
4076 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4077 * This also checks if there are enough spares and adds them
4079 * We currently require enough spares to make the final
4080 * array non-degraded. We also require that the difference
4081 * between old and new data_offset - on each device - is
4082 * enough that we never risk over-writing.
4085 unsigned long before_length
, after_length
;
4086 sector_t min_offset_diff
= 0;
4089 struct r10conf
*conf
= mddev
->private;
4090 struct md_rdev
*rdev
;
4094 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4097 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4100 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4101 conf
->prev
.far_copies
);
4102 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4103 conf
->geo
.far_copies
);
4105 rdev_for_each(rdev
, mddev
) {
4106 if (!test_bit(In_sync
, &rdev
->flags
)
4107 && !test_bit(Faulty
, &rdev
->flags
))
4109 if (rdev
->raid_disk
>= 0) {
4110 long long diff
= (rdev
->new_data_offset
4111 - rdev
->data_offset
);
4112 if (!mddev
->reshape_backwards
)
4116 if (first
|| diff
< min_offset_diff
)
4117 min_offset_diff
= diff
;
4121 if (max(before_length
, after_length
) > min_offset_diff
)
4124 if (spares
< mddev
->delta_disks
)
4127 conf
->offset_diff
= min_offset_diff
;
4128 spin_lock_irq(&conf
->device_lock
);
4129 if (conf
->mirrors_new
) {
4130 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4131 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4133 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
4134 conf
->mirrors_old
= conf
->mirrors
;
4135 conf
->mirrors
= conf
->mirrors_new
;
4136 conf
->mirrors_new
= NULL
;
4138 setup_geo(&conf
->geo
, mddev
, geo_start
);
4140 if (mddev
->reshape_backwards
) {
4141 sector_t size
= raid10_size(mddev
, 0, 0);
4142 if (size
< mddev
->array_sectors
) {
4143 spin_unlock_irq(&conf
->device_lock
);
4144 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4148 mddev
->resync_max_sectors
= size
;
4149 conf
->reshape_progress
= size
;
4151 conf
->reshape_progress
= 0;
4152 spin_unlock_irq(&conf
->device_lock
);
4154 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4155 ret
= bitmap_resize(mddev
->bitmap
,
4156 raid10_size(mddev
, 0,
4157 conf
->geo
.raid_disks
),
4162 if (mddev
->delta_disks
> 0) {
4163 rdev_for_each(rdev
, mddev
)
4164 if (rdev
->raid_disk
< 0 &&
4165 !test_bit(Faulty
, &rdev
->flags
)) {
4166 if (raid10_add_disk(mddev
, rdev
) == 0) {
4167 if (rdev
->raid_disk
>=
4168 conf
->prev
.raid_disks
)
4169 set_bit(In_sync
, &rdev
->flags
);
4171 rdev
->recovery_offset
= 0;
4173 if (sysfs_link_rdev(mddev
, rdev
))
4174 /* Failure here is OK */;
4176 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4177 && !test_bit(Faulty
, &rdev
->flags
)) {
4178 /* This is a spare that was manually added */
4179 set_bit(In_sync
, &rdev
->flags
);
4182 /* When a reshape changes the number of devices,
4183 * ->degraded is measured against the larger of the
4184 * pre and post numbers.
4186 spin_lock_irq(&conf
->device_lock
);
4187 mddev
->degraded
= calc_degraded(conf
);
4188 spin_unlock_irq(&conf
->device_lock
);
4189 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4190 mddev
->reshape_position
= conf
->reshape_progress
;
4191 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4193 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4194 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4195 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4196 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4198 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4200 if (!mddev
->sync_thread
) {
4204 conf
->reshape_checkpoint
= jiffies
;
4205 md_wakeup_thread(mddev
->sync_thread
);
4206 md_new_event(mddev
);
4210 mddev
->recovery
= 0;
4211 spin_lock_irq(&conf
->device_lock
);
4212 conf
->geo
= conf
->prev
;
4213 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4214 rdev_for_each(rdev
, mddev
)
4215 rdev
->new_data_offset
= rdev
->data_offset
;
4217 conf
->reshape_progress
= MaxSector
;
4218 mddev
->reshape_position
= MaxSector
;
4219 spin_unlock_irq(&conf
->device_lock
);
4223 /* Calculate the last device-address that could contain
4224 * any block from the chunk that includes the array-address 's'
4225 * and report the next address.
4226 * i.e. the address returned will be chunk-aligned and after
4227 * any data that is in the chunk containing 's'.
4229 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4231 s
= (s
| geo
->chunk_mask
) + 1;
4232 s
>>= geo
->chunk_shift
;
4233 s
*= geo
->near_copies
;
4234 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4235 s
*= geo
->far_copies
;
4236 s
<<= geo
->chunk_shift
;
4240 /* Calculate the first device-address that could contain
4241 * any block from the chunk that includes the array-address 's'.
4242 * This too will be the start of a chunk
4244 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4246 s
>>= geo
->chunk_shift
;
4247 s
*= geo
->near_copies
;
4248 sector_div(s
, geo
->raid_disks
);
4249 s
*= geo
->far_copies
;
4250 s
<<= geo
->chunk_shift
;
4254 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4257 /* We simply copy at most one chunk (smallest of old and new)
4258 * at a time, possibly less if that exceeds RESYNC_PAGES,
4259 * or we hit a bad block or something.
4260 * This might mean we pause for normal IO in the middle of
4261 * a chunk, but that is not a problem was mddev->reshape_position
4262 * can record any location.
4264 * If we will want to write to a location that isn't
4265 * yet recorded as 'safe' (i.e. in metadata on disk) then
4266 * we need to flush all reshape requests and update the metadata.
4268 * When reshaping forwards (e.g. to more devices), we interpret
4269 * 'safe' as the earliest block which might not have been copied
4270 * down yet. We divide this by previous stripe size and multiply
4271 * by previous stripe length to get lowest device offset that we
4272 * cannot write to yet.
4273 * We interpret 'sector_nr' as an address that we want to write to.
4274 * From this we use last_device_address() to find where we might
4275 * write to, and first_device_address on the 'safe' position.
4276 * If this 'next' write position is after the 'safe' position,
4277 * we must update the metadata to increase the 'safe' position.
4279 * When reshaping backwards, we round in the opposite direction
4280 * and perform the reverse test: next write position must not be
4281 * less than current safe position.
4283 * In all this the minimum difference in data offsets
4284 * (conf->offset_diff - always positive) allows a bit of slack,
4285 * so next can be after 'safe', but not by more than offset_disk
4287 * We need to prepare all the bios here before we start any IO
4288 * to ensure the size we choose is acceptable to all devices.
4289 * The means one for each copy for write-out and an extra one for
4291 * We store the read-in bio in ->master_bio and the others in
4292 * ->devs[x].bio and ->devs[x].repl_bio.
4294 struct r10conf
*conf
= mddev
->private;
4295 struct r10bio
*r10_bio
;
4296 sector_t next
, safe
, last
;
4300 struct md_rdev
*rdev
;
4303 struct bio
*bio
, *read_bio
;
4304 int sectors_done
= 0;
4306 if (sector_nr
== 0) {
4307 /* If restarting in the middle, skip the initial sectors */
4308 if (mddev
->reshape_backwards
&&
4309 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4310 sector_nr
= (raid10_size(mddev
, 0, 0)
4311 - conf
->reshape_progress
);
4312 } else if (!mddev
->reshape_backwards
&&
4313 conf
->reshape_progress
> 0)
4314 sector_nr
= conf
->reshape_progress
;
4316 mddev
->curr_resync_completed
= sector_nr
;
4317 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4323 /* We don't use sector_nr to track where we are up to
4324 * as that doesn't work well for ->reshape_backwards.
4325 * So just use ->reshape_progress.
4327 if (mddev
->reshape_backwards
) {
4328 /* 'next' is the earliest device address that we might
4329 * write to for this chunk in the new layout
4331 next
= first_dev_address(conf
->reshape_progress
- 1,
4334 /* 'safe' is the last device address that we might read from
4335 * in the old layout after a restart
4337 safe
= last_dev_address(conf
->reshape_safe
- 1,
4340 if (next
+ conf
->offset_diff
< safe
)
4343 last
= conf
->reshape_progress
- 1;
4344 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4345 & conf
->prev
.chunk_mask
);
4346 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4347 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4349 /* 'next' is after the last device address that we
4350 * might write to for this chunk in the new layout
4352 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4354 /* 'safe' is the earliest device address that we might
4355 * read from in the old layout after a restart
4357 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4359 /* Need to update metadata if 'next' might be beyond 'safe'
4360 * as that would possibly corrupt data
4362 if (next
> safe
+ conf
->offset_diff
)
4365 sector_nr
= conf
->reshape_progress
;
4366 last
= sector_nr
| (conf
->geo
.chunk_mask
4367 & conf
->prev
.chunk_mask
);
4369 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4370 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4374 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4375 /* Need to update reshape_position in metadata */
4377 mddev
->reshape_position
= conf
->reshape_progress
;
4378 if (mddev
->reshape_backwards
)
4379 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4380 - conf
->reshape_progress
;
4382 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4383 conf
->reshape_checkpoint
= jiffies
;
4384 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4385 md_wakeup_thread(mddev
->thread
);
4386 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4387 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4388 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4389 allow_barrier(conf
);
4390 return sectors_done
;
4392 conf
->reshape_safe
= mddev
->reshape_position
;
4393 allow_barrier(conf
);
4397 /* Now schedule reads for blocks from sector_nr to last */
4398 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4399 raise_barrier(conf
, sectors_done
!= 0);
4400 atomic_set(&r10_bio
->remaining
, 0);
4401 r10_bio
->mddev
= mddev
;
4402 r10_bio
->sector
= sector_nr
;
4403 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4404 r10_bio
->sectors
= last
- sector_nr
+ 1;
4405 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4406 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4409 /* Cannot read from here, so need to record bad blocks
4410 * on all the target devices.
4413 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4414 return sectors_done
;
4417 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4419 read_bio
->bi_bdev
= rdev
->bdev
;
4420 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4421 + rdev
->data_offset
);
4422 read_bio
->bi_private
= r10_bio
;
4423 read_bio
->bi_end_io
= end_sync_read
;
4424 read_bio
->bi_rw
= READ
;
4425 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4426 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4427 read_bio
->bi_vcnt
= 0;
4428 read_bio
->bi_size
= 0;
4429 r10_bio
->master_bio
= read_bio
;
4430 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4432 /* Now find the locations in the new layout */
4433 __raid10_find_phys(&conf
->geo
, r10_bio
);
4436 read_bio
->bi_next
= NULL
;
4438 for (s
= 0; s
< conf
->copies
*2; s
++) {
4440 int d
= r10_bio
->devs
[s
/2].devnum
;
4441 struct md_rdev
*rdev2
;
4443 rdev2
= conf
->mirrors
[d
].replacement
;
4444 b
= r10_bio
->devs
[s
/2].repl_bio
;
4446 rdev2
= conf
->mirrors
[d
].rdev
;
4447 b
= r10_bio
->devs
[s
/2].bio
;
4449 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4453 b
->bi_bdev
= rdev2
->bdev
;
4454 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4455 b
->bi_private
= r10_bio
;
4456 b
->bi_end_io
= end_reshape_write
;
4462 /* Now add as many pages as possible to all of these bios. */
4465 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4466 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4467 int len
= (max_sectors
- s
) << 9;
4468 if (len
> PAGE_SIZE
)
4470 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4472 if (bio_add_page(bio
, page
, len
, 0))
4475 /* Didn't fit, must stop */
4477 bio2
&& bio2
!= bio
;
4478 bio2
= bio2
->bi_next
) {
4479 /* Remove last page from this bio */
4481 bio2
->bi_size
-= len
;
4482 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4486 sector_nr
+= len
>> 9;
4487 nr_sectors
+= len
>> 9;
4490 r10_bio
->sectors
= nr_sectors
;
4492 /* Now submit the read */
4493 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4494 atomic_inc(&r10_bio
->remaining
);
4495 read_bio
->bi_next
= NULL
;
4496 generic_make_request(read_bio
);
4497 sector_nr
+= nr_sectors
;
4498 sectors_done
+= nr_sectors
;
4499 if (sector_nr
<= last
)
4502 /* Now that we have done the whole section we can
4503 * update reshape_progress
4505 if (mddev
->reshape_backwards
)
4506 conf
->reshape_progress
-= sectors_done
;
4508 conf
->reshape_progress
+= sectors_done
;
4510 return sectors_done
;
4513 static void end_reshape_request(struct r10bio
*r10_bio
);
4514 static int handle_reshape_read_error(struct mddev
*mddev
,
4515 struct r10bio
*r10_bio
);
4516 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4518 /* Reshape read completed. Hopefully we have a block
4520 * If we got a read error then we do sync 1-page reads from
4521 * elsewhere until we find the data - or give up.
4523 struct r10conf
*conf
= mddev
->private;
4526 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4527 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4528 /* Reshape has been aborted */
4529 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4533 /* We definitely have the data in the pages, schedule the
4536 atomic_set(&r10_bio
->remaining
, 1);
4537 for (s
= 0; s
< conf
->copies
*2; s
++) {
4539 int d
= r10_bio
->devs
[s
/2].devnum
;
4540 struct md_rdev
*rdev
;
4542 rdev
= conf
->mirrors
[d
].replacement
;
4543 b
= r10_bio
->devs
[s
/2].repl_bio
;
4545 rdev
= conf
->mirrors
[d
].rdev
;
4546 b
= r10_bio
->devs
[s
/2].bio
;
4548 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4550 atomic_inc(&rdev
->nr_pending
);
4551 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4552 atomic_inc(&r10_bio
->remaining
);
4554 generic_make_request(b
);
4556 end_reshape_request(r10_bio
);
4559 static void end_reshape(struct r10conf
*conf
)
4561 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4564 spin_lock_irq(&conf
->device_lock
);
4565 conf
->prev
= conf
->geo
;
4566 md_finish_reshape(conf
->mddev
);
4568 conf
->reshape_progress
= MaxSector
;
4569 spin_unlock_irq(&conf
->device_lock
);
4571 /* read-ahead size must cover two whole stripes, which is
4572 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4574 if (conf
->mddev
->queue
) {
4575 int stripe
= conf
->geo
.raid_disks
*
4576 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4577 stripe
/= conf
->geo
.near_copies
;
4578 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4579 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4585 static int handle_reshape_read_error(struct mddev
*mddev
,
4586 struct r10bio
*r10_bio
)
4588 /* Use sync reads to get the blocks from somewhere else */
4589 int sectors
= r10_bio
->sectors
;
4590 struct r10conf
*conf
= mddev
->private;
4592 struct r10bio r10_bio
;
4593 struct r10dev devs
[conf
->copies
];
4595 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4598 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4600 r10b
->sector
= r10_bio
->sector
;
4601 __raid10_find_phys(&conf
->prev
, r10b
);
4606 int first_slot
= slot
;
4608 if (s
> (PAGE_SIZE
>> 9))
4612 int d
= r10b
->devs
[slot
].devnum
;
4613 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4616 test_bit(Faulty
, &rdev
->flags
) ||
4617 !test_bit(In_sync
, &rdev
->flags
))
4620 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4621 success
= sync_page_io(rdev
,
4630 if (slot
>= conf
->copies
)
4632 if (slot
== first_slot
)
4636 /* couldn't read this block, must give up */
4637 set_bit(MD_RECOVERY_INTR
,
4647 static void end_reshape_write(struct bio
*bio
, int error
)
4649 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4650 struct r10bio
*r10_bio
= bio
->bi_private
;
4651 struct mddev
*mddev
= r10_bio
->mddev
;
4652 struct r10conf
*conf
= mddev
->private;
4656 struct md_rdev
*rdev
= NULL
;
4658 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4660 rdev
= conf
->mirrors
[d
].replacement
;
4663 rdev
= conf
->mirrors
[d
].rdev
;
4667 /* FIXME should record badblock */
4668 md_error(mddev
, rdev
);
4671 rdev_dec_pending(rdev
, mddev
);
4672 end_reshape_request(r10_bio
);
4675 static void end_reshape_request(struct r10bio
*r10_bio
)
4677 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4679 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4680 bio_put(r10_bio
->master_bio
);
4684 static void raid10_finish_reshape(struct mddev
*mddev
)
4686 struct r10conf
*conf
= mddev
->private;
4688 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4691 if (mddev
->delta_disks
> 0) {
4692 sector_t size
= raid10_size(mddev
, 0, 0);
4693 md_set_array_sectors(mddev
, size
);
4694 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4695 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4696 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4698 mddev
->resync_max_sectors
= size
;
4699 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4700 revalidate_disk(mddev
->gendisk
);
4703 for (d
= conf
->geo
.raid_disks
;
4704 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4706 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4708 clear_bit(In_sync
, &rdev
->flags
);
4709 rdev
= conf
->mirrors
[d
].replacement
;
4711 clear_bit(In_sync
, &rdev
->flags
);
4714 mddev
->layout
= mddev
->new_layout
;
4715 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4716 mddev
->reshape_position
= MaxSector
;
4717 mddev
->delta_disks
= 0;
4718 mddev
->reshape_backwards
= 0;
4721 static struct md_personality raid10_personality
=
4725 .owner
= THIS_MODULE
,
4726 .make_request
= make_request
,
4730 .error_handler
= error
,
4731 .hot_add_disk
= raid10_add_disk
,
4732 .hot_remove_disk
= raid10_remove_disk
,
4733 .spare_active
= raid10_spare_active
,
4734 .sync_request
= sync_request
,
4735 .quiesce
= raid10_quiesce
,
4736 .size
= raid10_size
,
4737 .resize
= raid10_resize
,
4738 .takeover
= raid10_takeover
,
4739 .check_reshape
= raid10_check_reshape
,
4740 .start_reshape
= raid10_start_reshape
,
4741 .finish_reshape
= raid10_finish_reshape
,
4744 static int __init
raid_init(void)
4746 return register_md_personality(&raid10_personality
);
4749 static void raid_exit(void)
4751 unregister_md_personality(&raid10_personality
);
4754 module_init(raid_init
);
4755 module_exit(raid_exit
);
4756 MODULE_LICENSE("GPL");
4757 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4758 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4759 MODULE_ALIAS("md-raid10");
4760 MODULE_ALIAS("md-level-10");
4762 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);