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 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"
396 spin_lock_irqsave(&conf
->device_lock
, flags
);
397 if (!enough(conf
, rdev
->raid_disk
))
399 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
402 raid_end_bio_io(r10_bio
);
403 rdev_dec_pending(rdev
, conf
->mddev
);
406 * oops, read error - keep the refcount on the rdev
408 char b
[BDEVNAME_SIZE
];
409 printk_ratelimited(KERN_ERR
410 "md/raid10:%s: %s: rescheduling sector %llu\n",
412 bdevname(rdev
->bdev
, b
),
413 (unsigned long long)r10_bio
->sector
);
414 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
415 reschedule_retry(r10_bio
);
419 static void close_write(struct r10bio
*r10_bio
)
421 /* clear the bitmap if all writes complete successfully */
422 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
424 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
426 md_write_end(r10_bio
->mddev
);
429 static void one_write_done(struct r10bio
*r10_bio
)
431 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
432 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
433 reschedule_retry(r10_bio
);
435 close_write(r10_bio
);
436 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
437 reschedule_retry(r10_bio
);
439 raid_end_bio_io(r10_bio
);
444 static void raid10_end_write_request(struct bio
*bio
, int error
)
446 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
447 struct r10bio
*r10_bio
= bio
->bi_private
;
450 struct r10conf
*conf
= r10_bio
->mddev
->private;
452 struct md_rdev
*rdev
= NULL
;
454 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
457 rdev
= conf
->mirrors
[dev
].replacement
;
461 rdev
= conf
->mirrors
[dev
].rdev
;
464 * this branch is our 'one mirror IO has finished' event handler:
468 /* Never record new bad blocks to replacement,
471 md_error(rdev
->mddev
, rdev
);
473 set_bit(WriteErrorSeen
, &rdev
->flags
);
474 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
475 set_bit(MD_RECOVERY_NEEDED
,
476 &rdev
->mddev
->recovery
);
477 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
482 * Set R10BIO_Uptodate in our master bio, so that
483 * we will return a good error code for to the higher
484 * levels even if IO on some other mirrored buffer fails.
486 * The 'master' represents the composite IO operation to
487 * user-side. So if something waits for IO, then it will
488 * wait for the 'master' bio.
493 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
495 /* Maybe we can clear some bad blocks. */
496 if (is_badblock(rdev
,
497 r10_bio
->devs
[slot
].addr
,
499 &first_bad
, &bad_sectors
)) {
502 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
504 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
506 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
512 * Let's see if all mirrored write operations have finished
515 one_write_done(r10_bio
);
517 rdev_dec_pending(rdev
, conf
->mddev
);
521 * RAID10 layout manager
522 * As well as the chunksize and raid_disks count, there are two
523 * parameters: near_copies and far_copies.
524 * near_copies * far_copies must be <= raid_disks.
525 * Normally one of these will be 1.
526 * If both are 1, we get raid0.
527 * If near_copies == raid_disks, we get raid1.
529 * Chunks are laid out in raid0 style with near_copies copies of the
530 * first chunk, followed by near_copies copies of the next chunk and
532 * If far_copies > 1, then after 1/far_copies of the array has been assigned
533 * as described above, we start again with a device offset of near_copies.
534 * So we effectively have another copy of the whole array further down all
535 * the drives, but with blocks on different drives.
536 * With this layout, and block is never stored twice on the one device.
538 * raid10_find_phys finds the sector offset of a given virtual sector
539 * on each device that it is on.
541 * raid10_find_virt does the reverse mapping, from a device and a
542 * sector offset to a virtual address
545 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
553 int last_far_set_start
, last_far_set_size
;
555 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
556 last_far_set_start
*= geo
->far_set_size
;
558 last_far_set_size
= geo
->far_set_size
;
559 last_far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
561 /* now calculate first sector/dev */
562 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
563 sector
= r10bio
->sector
& geo
->chunk_mask
;
565 chunk
*= geo
->near_copies
;
567 dev
= sector_div(stripe
, geo
->raid_disks
);
569 stripe
*= geo
->far_copies
;
571 sector
+= stripe
<< geo
->chunk_shift
;
573 /* and calculate all the others */
574 for (n
= 0; n
< geo
->near_copies
; n
++) {
578 r10bio
->devs
[slot
].devnum
= d
;
579 r10bio
->devs
[slot
].addr
= s
;
582 for (f
= 1; f
< geo
->far_copies
; f
++) {
583 set
= d
/ geo
->far_set_size
;
584 d
+= geo
->near_copies
;
586 if ((geo
->raid_disks
% geo
->far_set_size
) &&
587 (d
> last_far_set_start
)) {
588 d
-= last_far_set_start
;
589 d
%= last_far_set_size
;
590 d
+= last_far_set_start
;
592 d
%= geo
->far_set_size
;
593 d
+= geo
->far_set_size
* set
;
596 r10bio
->devs
[slot
].devnum
= d
;
597 r10bio
->devs
[slot
].addr
= s
;
601 if (dev
>= geo
->raid_disks
) {
603 sector
+= (geo
->chunk_mask
+ 1);
608 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
610 struct geom
*geo
= &conf
->geo
;
612 if (conf
->reshape_progress
!= MaxSector
&&
613 ((r10bio
->sector
>= conf
->reshape_progress
) !=
614 conf
->mddev
->reshape_backwards
)) {
615 set_bit(R10BIO_Previous
, &r10bio
->state
);
618 clear_bit(R10BIO_Previous
, &r10bio
->state
);
620 __raid10_find_phys(geo
, r10bio
);
623 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
625 sector_t offset
, chunk
, vchunk
;
626 /* Never use conf->prev as this is only called during resync
627 * or recovery, so reshape isn't happening
629 struct geom
*geo
= &conf
->geo
;
630 int far_set_start
= (dev
/ geo
->far_set_size
) * geo
->far_set_size
;
631 int far_set_size
= geo
->far_set_size
;
632 int last_far_set_start
;
634 if (geo
->raid_disks
% geo
->far_set_size
) {
635 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
636 last_far_set_start
*= geo
->far_set_size
;
638 if (dev
>= last_far_set_start
) {
639 far_set_size
= geo
->far_set_size
;
640 far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
641 far_set_start
= last_far_set_start
;
645 offset
= sector
& geo
->chunk_mask
;
646 if (geo
->far_offset
) {
648 chunk
= sector
>> geo
->chunk_shift
;
649 fc
= sector_div(chunk
, geo
->far_copies
);
650 dev
-= fc
* geo
->near_copies
;
651 if (dev
< far_set_start
)
654 while (sector
>= geo
->stride
) {
655 sector
-= geo
->stride
;
656 if (dev
< (geo
->near_copies
+ far_set_start
))
657 dev
+= far_set_size
- geo
->near_copies
;
659 dev
-= geo
->near_copies
;
661 chunk
= sector
>> geo
->chunk_shift
;
663 vchunk
= chunk
* geo
->raid_disks
+ dev
;
664 sector_div(vchunk
, geo
->near_copies
);
665 return (vchunk
<< geo
->chunk_shift
) + offset
;
669 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
671 * @bvm: properties of new bio
672 * @biovec: the request that could be merged to it.
674 * Return amount of bytes we can accept at this offset
675 * This requires checking for end-of-chunk if near_copies != raid_disks,
676 * and for subordinate merge_bvec_fns if merge_check_needed.
678 static int raid10_mergeable_bvec(struct request_queue
*q
,
679 struct bvec_merge_data
*bvm
,
680 struct bio_vec
*biovec
)
682 struct mddev
*mddev
= q
->queuedata
;
683 struct r10conf
*conf
= mddev
->private;
684 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
686 unsigned int chunk_sectors
;
687 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
688 struct geom
*geo
= &conf
->geo
;
690 chunk_sectors
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
) + 1;
691 if (conf
->reshape_progress
!= MaxSector
&&
692 ((sector
>= conf
->reshape_progress
) !=
693 conf
->mddev
->reshape_backwards
))
696 if (geo
->near_copies
< geo
->raid_disks
) {
697 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1))
698 + bio_sectors
)) << 9;
700 /* bio_add cannot handle a negative return */
702 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
703 return biovec
->bv_len
;
705 max
= biovec
->bv_len
;
707 if (mddev
->merge_check_needed
) {
709 struct r10bio r10_bio
;
710 struct r10dev devs
[conf
->copies
];
712 struct r10bio
*r10_bio
= &on_stack
.r10_bio
;
714 if (conf
->reshape_progress
!= MaxSector
) {
715 /* Cannot give any guidance during reshape */
716 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
717 return biovec
->bv_len
;
720 r10_bio
->sector
= sector
;
721 raid10_find_phys(conf
, r10_bio
);
723 for (s
= 0; s
< conf
->copies
; s
++) {
724 int disk
= r10_bio
->devs
[s
].devnum
;
725 struct md_rdev
*rdev
= rcu_dereference(
726 conf
->mirrors
[disk
].rdev
);
727 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
728 struct request_queue
*q
=
729 bdev_get_queue(rdev
->bdev
);
730 if (q
->merge_bvec_fn
) {
731 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
733 bvm
->bi_bdev
= rdev
->bdev
;
734 max
= min(max
, q
->merge_bvec_fn(
738 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
739 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
740 struct request_queue
*q
=
741 bdev_get_queue(rdev
->bdev
);
742 if (q
->merge_bvec_fn
) {
743 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
745 bvm
->bi_bdev
= rdev
->bdev
;
746 max
= min(max
, q
->merge_bvec_fn(
757 * This routine returns the disk from which the requested read should
758 * be done. There is a per-array 'next expected sequential IO' sector
759 * number - if this matches on the next IO then we use the last disk.
760 * There is also a per-disk 'last know head position' sector that is
761 * maintained from IRQ contexts, both the normal and the resync IO
762 * completion handlers update this position correctly. If there is no
763 * perfect sequential match then we pick the disk whose head is closest.
765 * If there are 2 mirrors in the same 2 devices, performance degrades
766 * because position is mirror, not device based.
768 * The rdev for the device selected will have nr_pending incremented.
772 * FIXME: possibly should rethink readbalancing and do it differently
773 * depending on near_copies / far_copies geometry.
775 static struct md_rdev
*read_balance(struct r10conf
*conf
,
776 struct r10bio
*r10_bio
,
779 const sector_t this_sector
= r10_bio
->sector
;
781 int sectors
= r10_bio
->sectors
;
782 int best_good_sectors
;
783 sector_t new_distance
, best_dist
;
784 struct md_rdev
*best_rdev
, *rdev
= NULL
;
787 struct geom
*geo
= &conf
->geo
;
789 raid10_find_phys(conf
, r10_bio
);
792 sectors
= r10_bio
->sectors
;
795 best_dist
= MaxSector
;
796 best_good_sectors
= 0;
799 * Check if we can balance. We can balance on the whole
800 * device if no resync is going on (recovery is ok), or below
801 * the resync window. We take the first readable disk when
802 * above the resync window.
804 if (conf
->mddev
->recovery_cp
< MaxSector
805 && (this_sector
+ sectors
>= conf
->next_resync
))
808 for (slot
= 0; slot
< conf
->copies
; slot
++) {
813 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
815 disk
= r10_bio
->devs
[slot
].devnum
;
816 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
817 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
818 test_bit(Unmerged
, &rdev
->flags
) ||
819 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
820 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
822 test_bit(Faulty
, &rdev
->flags
) ||
823 test_bit(Unmerged
, &rdev
->flags
))
825 if (!test_bit(In_sync
, &rdev
->flags
) &&
826 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
829 dev_sector
= r10_bio
->devs
[slot
].addr
;
830 if (is_badblock(rdev
, dev_sector
, sectors
,
831 &first_bad
, &bad_sectors
)) {
832 if (best_dist
< MaxSector
)
833 /* Already have a better slot */
835 if (first_bad
<= dev_sector
) {
836 /* Cannot read here. If this is the
837 * 'primary' device, then we must not read
838 * beyond 'bad_sectors' from another device.
840 bad_sectors
-= (dev_sector
- first_bad
);
841 if (!do_balance
&& sectors
> bad_sectors
)
842 sectors
= bad_sectors
;
843 if (best_good_sectors
> sectors
)
844 best_good_sectors
= sectors
;
846 sector_t good_sectors
=
847 first_bad
- dev_sector
;
848 if (good_sectors
> best_good_sectors
) {
849 best_good_sectors
= good_sectors
;
854 /* Must read from here */
859 best_good_sectors
= sectors
;
864 /* This optimisation is debatable, and completely destroys
865 * sequential read speed for 'far copies' arrays. So only
866 * keep it for 'near' arrays, and review those later.
868 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
871 /* for far > 1 always use the lowest address */
872 if (geo
->far_copies
> 1)
873 new_distance
= r10_bio
->devs
[slot
].addr
;
875 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
876 conf
->mirrors
[disk
].head_position
);
877 if (new_distance
< best_dist
) {
878 best_dist
= new_distance
;
883 if (slot
>= conf
->copies
) {
889 atomic_inc(&rdev
->nr_pending
);
890 if (test_bit(Faulty
, &rdev
->flags
)) {
891 /* Cannot risk returning a device that failed
892 * before we inc'ed nr_pending
894 rdev_dec_pending(rdev
, conf
->mddev
);
897 r10_bio
->read_slot
= slot
;
901 *max_sectors
= best_good_sectors
;
906 int md_raid10_congested(struct mddev
*mddev
, int bits
)
908 struct r10conf
*conf
= mddev
->private;
911 if ((bits
& (1 << BDI_async_congested
)) &&
912 conf
->pending_count
>= max_queued_requests
)
917 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
920 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
921 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
922 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
924 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
930 EXPORT_SYMBOL_GPL(md_raid10_congested
);
932 static int raid10_congested(void *data
, int bits
)
934 struct mddev
*mddev
= data
;
936 return mddev_congested(mddev
, bits
) ||
937 md_raid10_congested(mddev
, bits
);
940 static void flush_pending_writes(struct r10conf
*conf
)
942 /* Any writes that have been queued but are awaiting
943 * bitmap updates get flushed here.
945 spin_lock_irq(&conf
->device_lock
);
947 if (conf
->pending_bio_list
.head
) {
949 bio
= bio_list_get(&conf
->pending_bio_list
);
950 conf
->pending_count
= 0;
951 spin_unlock_irq(&conf
->device_lock
);
952 /* flush any pending bitmap writes to disk
953 * before proceeding w/ I/O */
954 bitmap_unplug(conf
->mddev
->bitmap
);
955 wake_up(&conf
->wait_barrier
);
957 while (bio
) { /* submit pending writes */
958 struct bio
*next
= bio
->bi_next
;
960 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
961 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
965 generic_make_request(bio
);
969 spin_unlock_irq(&conf
->device_lock
);
973 * Sometimes we need to suspend IO while we do something else,
974 * either some resync/recovery, or reconfigure the array.
975 * To do this we raise a 'barrier'.
976 * The 'barrier' is a counter that can be raised multiple times
977 * to count how many activities are happening which preclude
979 * We can only raise the barrier if there is no pending IO.
980 * i.e. if nr_pending == 0.
981 * We choose only to raise the barrier if no-one is waiting for the
982 * barrier to go down. This means that as soon as an IO request
983 * is ready, no other operations which require a barrier will start
984 * until the IO request has had a chance.
986 * So: regular IO calls 'wait_barrier'. When that returns there
987 * is no backgroup IO happening, It must arrange to call
988 * allow_barrier when it has finished its IO.
989 * backgroup IO calls must call raise_barrier. Once that returns
990 * there is no normal IO happeing. It must arrange to call
991 * lower_barrier when the particular background IO completes.
994 static void raise_barrier(struct r10conf
*conf
, int force
)
996 BUG_ON(force
&& !conf
->barrier
);
997 spin_lock_irq(&conf
->resync_lock
);
999 /* Wait until no block IO is waiting (unless 'force') */
1000 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
1003 /* block any new IO from starting */
1006 /* Now wait for all pending IO to complete */
1007 wait_event_lock_irq(conf
->wait_barrier
,
1008 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
1011 spin_unlock_irq(&conf
->resync_lock
);
1014 static void lower_barrier(struct r10conf
*conf
)
1016 unsigned long flags
;
1017 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1019 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1020 wake_up(&conf
->wait_barrier
);
1023 static void wait_barrier(struct r10conf
*conf
)
1025 spin_lock_irq(&conf
->resync_lock
);
1026 if (conf
->barrier
) {
1028 /* Wait for the barrier to drop.
1029 * However if there are already pending
1030 * requests (preventing the barrier from
1031 * rising completely), and the
1032 * pre-process bio queue isn't empty,
1033 * then don't wait, as we need to empty
1034 * that queue to get the nr_pending
1037 wait_event_lock_irq(conf
->wait_barrier
,
1039 (conf
->nr_pending
&&
1040 current
->bio_list
&&
1041 !bio_list_empty(current
->bio_list
)),
1046 spin_unlock_irq(&conf
->resync_lock
);
1049 static void allow_barrier(struct r10conf
*conf
)
1051 unsigned long flags
;
1052 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1054 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1055 wake_up(&conf
->wait_barrier
);
1058 static void freeze_array(struct r10conf
*conf
)
1060 /* stop syncio and normal IO and wait for everything to
1062 * We increment barrier and nr_waiting, and then
1063 * wait until nr_pending match nr_queued+1
1064 * This is called in the context of one normal IO request
1065 * that has failed. Thus any sync request that might be pending
1066 * will be blocked by nr_pending, and we need to wait for
1067 * pending IO requests to complete or be queued for re-try.
1068 * Thus the number queued (nr_queued) plus this request (1)
1069 * must match the number of pending IOs (nr_pending) before
1072 spin_lock_irq(&conf
->resync_lock
);
1075 wait_event_lock_irq_cmd(conf
->wait_barrier
,
1076 conf
->nr_pending
== conf
->nr_queued
+1,
1078 flush_pending_writes(conf
));
1080 spin_unlock_irq(&conf
->resync_lock
);
1083 static void unfreeze_array(struct r10conf
*conf
)
1085 /* reverse the effect of the freeze */
1086 spin_lock_irq(&conf
->resync_lock
);
1089 wake_up(&conf
->wait_barrier
);
1090 spin_unlock_irq(&conf
->resync_lock
);
1093 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1094 struct md_rdev
*rdev
)
1096 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1097 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1098 return rdev
->data_offset
;
1100 return rdev
->new_data_offset
;
1103 struct raid10_plug_cb
{
1104 struct blk_plug_cb cb
;
1105 struct bio_list pending
;
1109 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1111 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1113 struct mddev
*mddev
= plug
->cb
.data
;
1114 struct r10conf
*conf
= mddev
->private;
1117 if (from_schedule
|| current
->bio_list
) {
1118 spin_lock_irq(&conf
->device_lock
);
1119 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1120 conf
->pending_count
+= plug
->pending_cnt
;
1121 spin_unlock_irq(&conf
->device_lock
);
1122 wake_up(&conf
->wait_barrier
);
1123 md_wakeup_thread(mddev
->thread
);
1128 /* we aren't scheduling, so we can do the write-out directly. */
1129 bio
= bio_list_get(&plug
->pending
);
1130 bitmap_unplug(mddev
->bitmap
);
1131 wake_up(&conf
->wait_barrier
);
1133 while (bio
) { /* submit pending writes */
1134 struct bio
*next
= bio
->bi_next
;
1135 bio
->bi_next
= NULL
;
1136 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
1137 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1138 /* Just ignore it */
1141 generic_make_request(bio
);
1147 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1149 struct r10conf
*conf
= mddev
->private;
1150 struct r10bio
*r10_bio
;
1151 struct bio
*read_bio
;
1153 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1154 int chunk_sects
= chunk_mask
+ 1;
1155 const int rw
= bio_data_dir(bio
);
1156 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1157 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1158 const unsigned long do_discard
= (bio
->bi_rw
1159 & (REQ_DISCARD
| REQ_SECURE
));
1160 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1161 unsigned long flags
;
1162 struct md_rdev
*blocked_rdev
;
1163 struct blk_plug_cb
*cb
;
1164 struct raid10_plug_cb
*plug
= NULL
;
1165 int sectors_handled
;
1169 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1170 md_flush_request(mddev
, bio
);
1174 /* If this request crosses a chunk boundary, we need to
1175 * split it. This will only happen for 1 PAGE (or less) requests.
1177 if (unlikely((bio
->bi_sector
& chunk_mask
) + (bio
->bi_size
>> 9)
1179 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1180 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1181 struct bio_pair
*bp
;
1182 /* Sanity check -- queue functions should prevent this happening */
1183 if ((bio
->bi_vcnt
!= 1 && bio
->bi_vcnt
!= 0) ||
1186 /* This is a one page bio that upper layers
1187 * refuse to split for us, so we need to split it.
1190 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
1192 /* Each of these 'make_request' calls will call 'wait_barrier'.
1193 * If the first succeeds but the second blocks due to the resync
1194 * thread raising the barrier, we will deadlock because the
1195 * IO to the underlying device will be queued in generic_make_request
1196 * and will never complete, so will never reduce nr_pending.
1197 * So increment nr_waiting here so no new raise_barriers will
1198 * succeed, and so the second wait_barrier cannot block.
1200 spin_lock_irq(&conf
->resync_lock
);
1202 spin_unlock_irq(&conf
->resync_lock
);
1204 make_request(mddev
, &bp
->bio1
);
1205 make_request(mddev
, &bp
->bio2
);
1207 spin_lock_irq(&conf
->resync_lock
);
1209 wake_up(&conf
->wait_barrier
);
1210 spin_unlock_irq(&conf
->resync_lock
);
1212 bio_pair_release(bp
);
1215 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1216 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1217 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
1223 md_write_start(mddev
, bio
);
1226 * Register the new request and wait if the reconstruction
1227 * thread has put up a bar for new requests.
1228 * Continue immediately if no resync is active currently.
1232 sectors
= bio
->bi_size
>> 9;
1233 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1234 bio
->bi_sector
< conf
->reshape_progress
&&
1235 bio
->bi_sector
+ sectors
> conf
->reshape_progress
) {
1236 /* IO spans the reshape position. Need to wait for
1239 allow_barrier(conf
);
1240 wait_event(conf
->wait_barrier
,
1241 conf
->reshape_progress
<= bio
->bi_sector
||
1242 conf
->reshape_progress
>= bio
->bi_sector
+ sectors
);
1245 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1246 bio_data_dir(bio
) == WRITE
&&
1247 (mddev
->reshape_backwards
1248 ? (bio
->bi_sector
< conf
->reshape_safe
&&
1249 bio
->bi_sector
+ sectors
> conf
->reshape_progress
)
1250 : (bio
->bi_sector
+ sectors
> conf
->reshape_safe
&&
1251 bio
->bi_sector
< conf
->reshape_progress
))) {
1252 /* Need to update reshape_position in metadata */
1253 mddev
->reshape_position
= conf
->reshape_progress
;
1254 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1255 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1256 md_wakeup_thread(mddev
->thread
);
1257 wait_event(mddev
->sb_wait
,
1258 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1260 conf
->reshape_safe
= mddev
->reshape_position
;
1263 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1265 r10_bio
->master_bio
= bio
;
1266 r10_bio
->sectors
= sectors
;
1268 r10_bio
->mddev
= mddev
;
1269 r10_bio
->sector
= bio
->bi_sector
;
1272 /* We might need to issue multiple reads to different
1273 * devices if there are bad blocks around, so we keep
1274 * track of the number of reads in bio->bi_phys_segments.
1275 * If this is 0, there is only one r10_bio and no locking
1276 * will be needed when the request completes. If it is
1277 * non-zero, then it is the number of not-completed requests.
1279 bio
->bi_phys_segments
= 0;
1280 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1284 * read balancing logic:
1286 struct md_rdev
*rdev
;
1290 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1292 raid_end_bio_io(r10_bio
);
1295 slot
= r10_bio
->read_slot
;
1297 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1298 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1301 r10_bio
->devs
[slot
].bio
= read_bio
;
1302 r10_bio
->devs
[slot
].rdev
= rdev
;
1304 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1305 choose_data_offset(r10_bio
, rdev
);
1306 read_bio
->bi_bdev
= rdev
->bdev
;
1307 read_bio
->bi_end_io
= raid10_end_read_request
;
1308 read_bio
->bi_rw
= READ
| do_sync
;
1309 read_bio
->bi_private
= r10_bio
;
1311 if (max_sectors
< r10_bio
->sectors
) {
1312 /* Could not read all from this device, so we will
1313 * need another r10_bio.
1315 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1317 r10_bio
->sectors
= max_sectors
;
1318 spin_lock_irq(&conf
->device_lock
);
1319 if (bio
->bi_phys_segments
== 0)
1320 bio
->bi_phys_segments
= 2;
1322 bio
->bi_phys_segments
++;
1323 spin_unlock(&conf
->device_lock
);
1324 /* Cannot call generic_make_request directly
1325 * as that will be queued in __generic_make_request
1326 * and subsequent mempool_alloc might block
1327 * waiting for it. so hand bio over to raid10d.
1329 reschedule_retry(r10_bio
);
1331 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1333 r10_bio
->master_bio
= bio
;
1334 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
1337 r10_bio
->mddev
= mddev
;
1338 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1341 generic_make_request(read_bio
);
1348 if (conf
->pending_count
>= max_queued_requests
) {
1349 md_wakeup_thread(mddev
->thread
);
1350 wait_event(conf
->wait_barrier
,
1351 conf
->pending_count
< max_queued_requests
);
1353 /* first select target devices under rcu_lock and
1354 * inc refcount on their rdev. Record them by setting
1356 * If there are known/acknowledged bad blocks on any device
1357 * on which we have seen a write error, we want to avoid
1358 * writing to those blocks. This potentially requires several
1359 * writes to write around the bad blocks. Each set of writes
1360 * gets its own r10_bio with a set of bios attached. The number
1361 * of r10_bios is recored in bio->bi_phys_segments just as with
1365 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1366 raid10_find_phys(conf
, r10_bio
);
1368 blocked_rdev
= NULL
;
1370 max_sectors
= r10_bio
->sectors
;
1372 for (i
= 0; i
< conf
->copies
; i
++) {
1373 int d
= r10_bio
->devs
[i
].devnum
;
1374 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1375 struct md_rdev
*rrdev
= rcu_dereference(
1376 conf
->mirrors
[d
].replacement
);
1379 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1380 atomic_inc(&rdev
->nr_pending
);
1381 blocked_rdev
= rdev
;
1384 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1385 atomic_inc(&rrdev
->nr_pending
);
1386 blocked_rdev
= rrdev
;
1389 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)
1390 || test_bit(Unmerged
, &rdev
->flags
)))
1392 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1393 || test_bit(Unmerged
, &rrdev
->flags
)))
1396 r10_bio
->devs
[i
].bio
= NULL
;
1397 r10_bio
->devs
[i
].repl_bio
= NULL
;
1399 if (!rdev
&& !rrdev
) {
1400 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1403 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1405 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1409 is_bad
= is_badblock(rdev
, dev_sector
,
1411 &first_bad
, &bad_sectors
);
1413 /* Mustn't write here until the bad block
1416 atomic_inc(&rdev
->nr_pending
);
1417 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1418 blocked_rdev
= rdev
;
1421 if (is_bad
&& first_bad
<= dev_sector
) {
1422 /* Cannot write here at all */
1423 bad_sectors
-= (dev_sector
- first_bad
);
1424 if (bad_sectors
< max_sectors
)
1425 /* Mustn't write more than bad_sectors
1426 * to other devices yet
1428 max_sectors
= bad_sectors
;
1429 /* We don't set R10BIO_Degraded as that
1430 * only applies if the disk is missing,
1431 * so it might be re-added, and we want to
1432 * know to recover this chunk.
1433 * In this case the device is here, and the
1434 * fact that this chunk is not in-sync is
1435 * recorded in the bad block log.
1440 int good_sectors
= first_bad
- dev_sector
;
1441 if (good_sectors
< max_sectors
)
1442 max_sectors
= good_sectors
;
1446 r10_bio
->devs
[i
].bio
= bio
;
1447 atomic_inc(&rdev
->nr_pending
);
1450 r10_bio
->devs
[i
].repl_bio
= bio
;
1451 atomic_inc(&rrdev
->nr_pending
);
1456 if (unlikely(blocked_rdev
)) {
1457 /* Have to wait for this device to get unblocked, then retry */
1461 for (j
= 0; j
< i
; j
++) {
1462 if (r10_bio
->devs
[j
].bio
) {
1463 d
= r10_bio
->devs
[j
].devnum
;
1464 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1466 if (r10_bio
->devs
[j
].repl_bio
) {
1467 struct md_rdev
*rdev
;
1468 d
= r10_bio
->devs
[j
].devnum
;
1469 rdev
= conf
->mirrors
[d
].replacement
;
1471 /* Race with remove_disk */
1473 rdev
= conf
->mirrors
[d
].rdev
;
1475 rdev_dec_pending(rdev
, mddev
);
1478 allow_barrier(conf
);
1479 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1484 if (max_sectors
< r10_bio
->sectors
) {
1485 /* We are splitting this into multiple parts, so
1486 * we need to prepare for allocating another r10_bio.
1488 r10_bio
->sectors
= max_sectors
;
1489 spin_lock_irq(&conf
->device_lock
);
1490 if (bio
->bi_phys_segments
== 0)
1491 bio
->bi_phys_segments
= 2;
1493 bio
->bi_phys_segments
++;
1494 spin_unlock_irq(&conf
->device_lock
);
1496 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1498 atomic_set(&r10_bio
->remaining
, 1);
1499 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1501 for (i
= 0; i
< conf
->copies
; i
++) {
1503 int d
= r10_bio
->devs
[i
].devnum
;
1504 if (r10_bio
->devs
[i
].bio
) {
1505 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1506 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1507 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1509 r10_bio
->devs
[i
].bio
= mbio
;
1511 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1512 choose_data_offset(r10_bio
,
1514 mbio
->bi_bdev
= rdev
->bdev
;
1515 mbio
->bi_end_io
= raid10_end_write_request
;
1517 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1518 mbio
->bi_private
= r10_bio
;
1520 atomic_inc(&r10_bio
->remaining
);
1522 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1525 plug
= container_of(cb
, struct raid10_plug_cb
,
1529 spin_lock_irqsave(&conf
->device_lock
, flags
);
1531 bio_list_add(&plug
->pending
, mbio
);
1532 plug
->pending_cnt
++;
1534 bio_list_add(&conf
->pending_bio_list
, mbio
);
1535 conf
->pending_count
++;
1537 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1539 md_wakeup_thread(mddev
->thread
);
1542 if (r10_bio
->devs
[i
].repl_bio
) {
1543 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1545 /* Replacement just got moved to main 'rdev' */
1547 rdev
= conf
->mirrors
[d
].rdev
;
1549 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1550 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1552 r10_bio
->devs
[i
].repl_bio
= mbio
;
1554 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1557 mbio
->bi_bdev
= rdev
->bdev
;
1558 mbio
->bi_end_io
= raid10_end_write_request
;
1560 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1561 mbio
->bi_private
= r10_bio
;
1563 atomic_inc(&r10_bio
->remaining
);
1564 spin_lock_irqsave(&conf
->device_lock
, flags
);
1565 bio_list_add(&conf
->pending_bio_list
, mbio
);
1566 conf
->pending_count
++;
1567 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1568 if (!mddev_check_plugged(mddev
))
1569 md_wakeup_thread(mddev
->thread
);
1573 /* Don't remove the bias on 'remaining' (one_write_done) until
1574 * after checking if we need to go around again.
1577 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1578 one_write_done(r10_bio
);
1579 /* We need another r10_bio. It has already been counted
1580 * in bio->bi_phys_segments.
1582 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1584 r10_bio
->master_bio
= bio
;
1585 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1587 r10_bio
->mddev
= mddev
;
1588 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1592 one_write_done(r10_bio
);
1594 /* In case raid10d snuck in to freeze_array */
1595 wake_up(&conf
->wait_barrier
);
1598 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1600 struct r10conf
*conf
= mddev
->private;
1603 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1604 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1605 if (conf
->geo
.near_copies
> 1)
1606 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1607 if (conf
->geo
.far_copies
> 1) {
1608 if (conf
->geo
.far_offset
)
1609 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1611 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1613 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1614 conf
->geo
.raid_disks
- mddev
->degraded
);
1615 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1616 seq_printf(seq
, "%s",
1617 conf
->mirrors
[i
].rdev
&&
1618 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1619 seq_printf(seq
, "]");
1622 /* check if there are enough drives for
1623 * every block to appear on atleast one.
1624 * Don't consider the device numbered 'ignore'
1625 * as we might be about to remove it.
1627 static int _enough(struct r10conf
*conf
, struct geom
*geo
, int ignore
)
1632 int n
= conf
->copies
;
1636 if (conf
->mirrors
[this].rdev
&&
1639 this = (this+1) % geo
->raid_disks
;
1643 first
= (first
+ geo
->near_copies
) % geo
->raid_disks
;
1644 } while (first
!= 0);
1648 static int enough(struct r10conf
*conf
, int ignore
)
1650 return _enough(conf
, &conf
->geo
, ignore
) &&
1651 _enough(conf
, &conf
->prev
, ignore
);
1654 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1656 char b
[BDEVNAME_SIZE
];
1657 struct r10conf
*conf
= mddev
->private;
1660 * If it is not operational, then we have already marked it as dead
1661 * else if it is the last working disks, ignore the error, let the
1662 * next level up know.
1663 * else mark the drive as failed
1665 if (test_bit(In_sync
, &rdev
->flags
)
1666 && !enough(conf
, rdev
->raid_disk
))
1668 * Don't fail the drive, just return an IO error.
1671 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1672 unsigned long flags
;
1673 spin_lock_irqsave(&conf
->device_lock
, flags
);
1675 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1677 * if recovery is running, make sure it aborts.
1679 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1681 set_bit(Blocked
, &rdev
->flags
);
1682 set_bit(Faulty
, &rdev
->flags
);
1683 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1685 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1686 "md/raid10:%s: Operation continuing on %d devices.\n",
1687 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1688 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1691 static void print_conf(struct r10conf
*conf
)
1694 struct raid10_info
*tmp
;
1696 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1698 printk(KERN_DEBUG
"(!conf)\n");
1701 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1702 conf
->geo
.raid_disks
);
1704 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1705 char b
[BDEVNAME_SIZE
];
1706 tmp
= conf
->mirrors
+ i
;
1708 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1709 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1710 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1711 bdevname(tmp
->rdev
->bdev
,b
));
1715 static void close_sync(struct r10conf
*conf
)
1718 allow_barrier(conf
);
1720 mempool_destroy(conf
->r10buf_pool
);
1721 conf
->r10buf_pool
= NULL
;
1724 static int raid10_spare_active(struct mddev
*mddev
)
1727 struct r10conf
*conf
= mddev
->private;
1728 struct raid10_info
*tmp
;
1730 unsigned long flags
;
1733 * Find all non-in_sync disks within the RAID10 configuration
1734 * and mark them in_sync
1736 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1737 tmp
= conf
->mirrors
+ i
;
1738 if (tmp
->replacement
1739 && tmp
->replacement
->recovery_offset
== MaxSector
1740 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1741 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1742 /* Replacement has just become active */
1744 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1747 /* Replaced device not technically faulty,
1748 * but we need to be sure it gets removed
1749 * and never re-added.
1751 set_bit(Faulty
, &tmp
->rdev
->flags
);
1752 sysfs_notify_dirent_safe(
1753 tmp
->rdev
->sysfs_state
);
1755 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1756 } else if (tmp
->rdev
1757 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1758 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1760 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1763 spin_lock_irqsave(&conf
->device_lock
, flags
);
1764 mddev
->degraded
-= count
;
1765 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1772 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1774 struct r10conf
*conf
= mddev
->private;
1778 int last
= conf
->geo
.raid_disks
- 1;
1779 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1781 if (mddev
->recovery_cp
< MaxSector
)
1782 /* only hot-add to in-sync arrays, as recovery is
1783 * very different from resync
1786 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, &conf
->prev
, -1))
1789 if (rdev
->raid_disk
>= 0)
1790 first
= last
= rdev
->raid_disk
;
1792 if (q
->merge_bvec_fn
) {
1793 set_bit(Unmerged
, &rdev
->flags
);
1794 mddev
->merge_check_needed
= 1;
1797 if (rdev
->saved_raid_disk
>= first
&&
1798 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1799 mirror
= rdev
->saved_raid_disk
;
1802 for ( ; mirror
<= last
; mirror
++) {
1803 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1804 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1807 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1808 p
->replacement
!= NULL
)
1810 clear_bit(In_sync
, &rdev
->flags
);
1811 set_bit(Replacement
, &rdev
->flags
);
1812 rdev
->raid_disk
= mirror
;
1814 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1815 rdev
->data_offset
<< 9);
1817 rcu_assign_pointer(p
->replacement
, rdev
);
1821 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1822 rdev
->data_offset
<< 9);
1824 p
->head_position
= 0;
1825 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1826 rdev
->raid_disk
= mirror
;
1828 if (rdev
->saved_raid_disk
!= mirror
)
1830 rcu_assign_pointer(p
->rdev
, rdev
);
1833 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1834 /* Some requests might not have seen this new
1835 * merge_bvec_fn. We must wait for them to complete
1836 * before merging the device fully.
1837 * First we make sure any code which has tested
1838 * our function has submitted the request, then
1839 * we wait for all outstanding requests to complete.
1841 synchronize_sched();
1842 raise_barrier(conf
, 0);
1843 lower_barrier(conf
);
1844 clear_bit(Unmerged
, &rdev
->flags
);
1846 md_integrity_add_rdev(rdev
, mddev
);
1847 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1848 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1854 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1856 struct r10conf
*conf
= mddev
->private;
1858 int number
= rdev
->raid_disk
;
1859 struct md_rdev
**rdevp
;
1860 struct raid10_info
*p
= conf
->mirrors
+ number
;
1863 if (rdev
== p
->rdev
)
1865 else if (rdev
== p
->replacement
)
1866 rdevp
= &p
->replacement
;
1870 if (test_bit(In_sync
, &rdev
->flags
) ||
1871 atomic_read(&rdev
->nr_pending
)) {
1875 /* Only remove faulty devices if recovery
1878 if (!test_bit(Faulty
, &rdev
->flags
) &&
1879 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1880 (!p
->replacement
|| p
->replacement
== rdev
) &&
1881 number
< conf
->geo
.raid_disks
&&
1888 if (atomic_read(&rdev
->nr_pending
)) {
1889 /* lost the race, try later */
1893 } else if (p
->replacement
) {
1894 /* We must have just cleared 'rdev' */
1895 p
->rdev
= p
->replacement
;
1896 clear_bit(Replacement
, &p
->replacement
->flags
);
1897 smp_mb(); /* Make sure other CPUs may see both as identical
1898 * but will never see neither -- if they are careful.
1900 p
->replacement
= NULL
;
1901 clear_bit(WantReplacement
, &rdev
->flags
);
1903 /* We might have just remove the Replacement as faulty
1904 * Clear the flag just in case
1906 clear_bit(WantReplacement
, &rdev
->flags
);
1908 err
= md_integrity_register(mddev
);
1917 static void end_sync_read(struct bio
*bio
, int error
)
1919 struct r10bio
*r10_bio
= bio
->bi_private
;
1920 struct r10conf
*conf
= r10_bio
->mddev
->private;
1923 if (bio
== r10_bio
->master_bio
) {
1924 /* this is a reshape read */
1925 d
= r10_bio
->read_slot
; /* really the read dev */
1927 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1929 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1930 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1932 /* The write handler will notice the lack of
1933 * R10BIO_Uptodate and record any errors etc
1935 atomic_add(r10_bio
->sectors
,
1936 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1938 /* for reconstruct, we always reschedule after a read.
1939 * for resync, only after all reads
1941 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1942 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1943 atomic_dec_and_test(&r10_bio
->remaining
)) {
1944 /* we have read all the blocks,
1945 * do the comparison in process context in raid10d
1947 reschedule_retry(r10_bio
);
1951 static void end_sync_request(struct r10bio
*r10_bio
)
1953 struct mddev
*mddev
= r10_bio
->mddev
;
1955 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1956 if (r10_bio
->master_bio
== NULL
) {
1957 /* the primary of several recovery bios */
1958 sector_t s
= r10_bio
->sectors
;
1959 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1960 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1961 reschedule_retry(r10_bio
);
1964 md_done_sync(mddev
, s
, 1);
1967 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1968 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1969 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1970 reschedule_retry(r10_bio
);
1978 static void end_sync_write(struct bio
*bio
, int error
)
1980 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1981 struct r10bio
*r10_bio
= bio
->bi_private
;
1982 struct mddev
*mddev
= r10_bio
->mddev
;
1983 struct r10conf
*conf
= mddev
->private;
1989 struct md_rdev
*rdev
= NULL
;
1991 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1993 rdev
= conf
->mirrors
[d
].replacement
;
1995 rdev
= conf
->mirrors
[d
].rdev
;
1999 md_error(mddev
, rdev
);
2001 set_bit(WriteErrorSeen
, &rdev
->flags
);
2002 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2003 set_bit(MD_RECOVERY_NEEDED
,
2004 &rdev
->mddev
->recovery
);
2005 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
2007 } else if (is_badblock(rdev
,
2008 r10_bio
->devs
[slot
].addr
,
2010 &first_bad
, &bad_sectors
))
2011 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
2013 rdev_dec_pending(rdev
, mddev
);
2015 end_sync_request(r10_bio
);
2019 * Note: sync and recover and handled very differently for raid10
2020 * This code is for resync.
2021 * For resync, we read through virtual addresses and read all blocks.
2022 * If there is any error, we schedule a write. The lowest numbered
2023 * drive is authoritative.
2024 * However requests come for physical address, so we need to map.
2025 * For every physical address there are raid_disks/copies virtual addresses,
2026 * which is always are least one, but is not necessarly an integer.
2027 * This means that a physical address can span multiple chunks, so we may
2028 * have to submit multiple io requests for a single sync request.
2031 * We check if all blocks are in-sync and only write to blocks that
2034 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2036 struct r10conf
*conf
= mddev
->private;
2038 struct bio
*tbio
, *fbio
;
2041 atomic_set(&r10_bio
->remaining
, 1);
2043 /* find the first device with a block */
2044 for (i
=0; i
<conf
->copies
; i
++)
2045 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
2048 if (i
== conf
->copies
)
2052 fbio
= r10_bio
->devs
[i
].bio
;
2054 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2055 /* now find blocks with errors */
2056 for (i
=0 ; i
< conf
->copies
; i
++) {
2059 tbio
= r10_bio
->devs
[i
].bio
;
2061 if (tbio
->bi_end_io
!= end_sync_read
)
2065 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2066 /* We know that the bi_io_vec layout is the same for
2067 * both 'first' and 'i', so we just compare them.
2068 * All vec entries are PAGE_SIZE;
2070 for (j
= 0; j
< vcnt
; j
++)
2071 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2072 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2073 fbio
->bi_io_vec
[j
].bv_len
))
2077 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2078 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2079 /* Don't fix anything. */
2082 /* Ok, we need to write this bio, either to correct an
2083 * inconsistency or to correct an unreadable block.
2084 * First we need to fixup bv_offset, bv_len and
2085 * bi_vecs, as the read request might have corrupted these
2087 tbio
->bi_vcnt
= vcnt
;
2088 tbio
->bi_size
= r10_bio
->sectors
<< 9;
2090 tbio
->bi_phys_segments
= 0;
2091 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2092 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
2093 tbio
->bi_next
= NULL
;
2094 tbio
->bi_rw
= WRITE
;
2095 tbio
->bi_private
= r10_bio
;
2096 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
2098 for (j
=0; j
< vcnt
; j
++) {
2099 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2100 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2102 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2103 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2106 tbio
->bi_end_io
= end_sync_write
;
2108 d
= r10_bio
->devs
[i
].devnum
;
2109 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2110 atomic_inc(&r10_bio
->remaining
);
2111 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
2113 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2114 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2115 generic_make_request(tbio
);
2118 /* Now write out to any replacement devices
2121 for (i
= 0; i
< conf
->copies
; i
++) {
2124 tbio
= r10_bio
->devs
[i
].repl_bio
;
2125 if (!tbio
|| !tbio
->bi_end_io
)
2127 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2128 && r10_bio
->devs
[i
].bio
!= fbio
)
2129 for (j
= 0; j
< vcnt
; j
++)
2130 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2131 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2133 d
= r10_bio
->devs
[i
].devnum
;
2134 atomic_inc(&r10_bio
->remaining
);
2135 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2136 tbio
->bi_size
>> 9);
2137 generic_make_request(tbio
);
2141 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2142 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2148 * Now for the recovery code.
2149 * Recovery happens across physical sectors.
2150 * We recover all non-is_sync drives by finding the virtual address of
2151 * each, and then choose a working drive that also has that virt address.
2152 * There is a separate r10_bio for each non-in_sync drive.
2153 * Only the first two slots are in use. The first for reading,
2154 * The second for writing.
2157 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2159 /* We got a read error during recovery.
2160 * We repeat the read in smaller page-sized sections.
2161 * If a read succeeds, write it to the new device or record
2162 * a bad block if we cannot.
2163 * If a read fails, record a bad block on both old and
2166 struct mddev
*mddev
= r10_bio
->mddev
;
2167 struct r10conf
*conf
= mddev
->private;
2168 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2170 int sectors
= r10_bio
->sectors
;
2172 int dr
= r10_bio
->devs
[0].devnum
;
2173 int dw
= r10_bio
->devs
[1].devnum
;
2177 struct md_rdev
*rdev
;
2181 if (s
> (PAGE_SIZE
>>9))
2184 rdev
= conf
->mirrors
[dr
].rdev
;
2185 addr
= r10_bio
->devs
[0].addr
+ sect
,
2186 ok
= sync_page_io(rdev
,
2189 bio
->bi_io_vec
[idx
].bv_page
,
2192 rdev
= conf
->mirrors
[dw
].rdev
;
2193 addr
= r10_bio
->devs
[1].addr
+ sect
;
2194 ok
= sync_page_io(rdev
,
2197 bio
->bi_io_vec
[idx
].bv_page
,
2200 set_bit(WriteErrorSeen
, &rdev
->flags
);
2201 if (!test_and_set_bit(WantReplacement
,
2203 set_bit(MD_RECOVERY_NEEDED
,
2204 &rdev
->mddev
->recovery
);
2208 /* We don't worry if we cannot set a bad block -
2209 * it really is bad so there is no loss in not
2212 rdev_set_badblocks(rdev
, addr
, s
, 0);
2214 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2215 /* need bad block on destination too */
2216 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2217 addr
= r10_bio
->devs
[1].addr
+ sect
;
2218 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2220 /* just abort the recovery */
2222 "md/raid10:%s: recovery aborted"
2223 " due to read error\n",
2226 conf
->mirrors
[dw
].recovery_disabled
2227 = mddev
->recovery_disabled
;
2228 set_bit(MD_RECOVERY_INTR
,
2241 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2243 struct r10conf
*conf
= mddev
->private;
2245 struct bio
*wbio
, *wbio2
;
2247 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2248 fix_recovery_read_error(r10_bio
);
2249 end_sync_request(r10_bio
);
2254 * share the pages with the first bio
2255 * and submit the write request
2257 d
= r10_bio
->devs
[1].devnum
;
2258 wbio
= r10_bio
->devs
[1].bio
;
2259 wbio2
= r10_bio
->devs
[1].repl_bio
;
2260 if (wbio
->bi_end_io
) {
2261 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2262 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
2263 generic_make_request(wbio
);
2265 if (wbio2
&& wbio2
->bi_end_io
) {
2266 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2267 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2268 wbio2
->bi_size
>> 9);
2269 generic_make_request(wbio2
);
2275 * Used by fix_read_error() to decay the per rdev read_errors.
2276 * We halve the read error count for every hour that has elapsed
2277 * since the last recorded read error.
2280 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2282 struct timespec cur_time_mon
;
2283 unsigned long hours_since_last
;
2284 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2286 ktime_get_ts(&cur_time_mon
);
2288 if (rdev
->last_read_error
.tv_sec
== 0 &&
2289 rdev
->last_read_error
.tv_nsec
== 0) {
2290 /* first time we've seen a read error */
2291 rdev
->last_read_error
= cur_time_mon
;
2295 hours_since_last
= (cur_time_mon
.tv_sec
-
2296 rdev
->last_read_error
.tv_sec
) / 3600;
2298 rdev
->last_read_error
= cur_time_mon
;
2301 * if hours_since_last is > the number of bits in read_errors
2302 * just set read errors to 0. We do this to avoid
2303 * overflowing the shift of read_errors by hours_since_last.
2305 if (hours_since_last
>= 8 * sizeof(read_errors
))
2306 atomic_set(&rdev
->read_errors
, 0);
2308 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2311 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2312 int sectors
, struct page
*page
, int rw
)
2317 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2318 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2320 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2324 set_bit(WriteErrorSeen
, &rdev
->flags
);
2325 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2326 set_bit(MD_RECOVERY_NEEDED
,
2327 &rdev
->mddev
->recovery
);
2329 /* need to record an error - either for the block or the device */
2330 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2331 md_error(rdev
->mddev
, rdev
);
2336 * This is a kernel thread which:
2338 * 1. Retries failed read operations on working mirrors.
2339 * 2. Updates the raid superblock when problems encounter.
2340 * 3. Performs writes following reads for array synchronising.
2343 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2345 int sect
= 0; /* Offset from r10_bio->sector */
2346 int sectors
= r10_bio
->sectors
;
2347 struct md_rdev
*rdev
;
2348 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2349 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2351 /* still own a reference to this rdev, so it cannot
2352 * have been cleared recently.
2354 rdev
= conf
->mirrors
[d
].rdev
;
2356 if (test_bit(Faulty
, &rdev
->flags
))
2357 /* drive has already been failed, just ignore any
2358 more fix_read_error() attempts */
2361 check_decay_read_errors(mddev
, rdev
);
2362 atomic_inc(&rdev
->read_errors
);
2363 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2364 char b
[BDEVNAME_SIZE
];
2365 bdevname(rdev
->bdev
, b
);
2368 "md/raid10:%s: %s: Raid device exceeded "
2369 "read_error threshold [cur %d:max %d]\n",
2371 atomic_read(&rdev
->read_errors
), max_read_errors
);
2373 "md/raid10:%s: %s: Failing raid device\n",
2375 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2376 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2382 int sl
= r10_bio
->read_slot
;
2386 if (s
> (PAGE_SIZE
>>9))
2394 d
= r10_bio
->devs
[sl
].devnum
;
2395 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2397 !test_bit(Unmerged
, &rdev
->flags
) &&
2398 test_bit(In_sync
, &rdev
->flags
) &&
2399 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2400 &first_bad
, &bad_sectors
) == 0) {
2401 atomic_inc(&rdev
->nr_pending
);
2403 success
= sync_page_io(rdev
,
2404 r10_bio
->devs
[sl
].addr
+
2407 conf
->tmppage
, READ
, false);
2408 rdev_dec_pending(rdev
, mddev
);
2414 if (sl
== conf
->copies
)
2416 } while (!success
&& sl
!= r10_bio
->read_slot
);
2420 /* Cannot read from anywhere, just mark the block
2421 * as bad on the first device to discourage future
2424 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2425 rdev
= conf
->mirrors
[dn
].rdev
;
2427 if (!rdev_set_badblocks(
2429 r10_bio
->devs
[r10_bio
->read_slot
].addr
2432 md_error(mddev
, rdev
);
2433 r10_bio
->devs
[r10_bio
->read_slot
].bio
2440 /* write it back and re-read */
2442 while (sl
!= r10_bio
->read_slot
) {
2443 char b
[BDEVNAME_SIZE
];
2448 d
= r10_bio
->devs
[sl
].devnum
;
2449 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2451 test_bit(Unmerged
, &rdev
->flags
) ||
2452 !test_bit(In_sync
, &rdev
->flags
))
2455 atomic_inc(&rdev
->nr_pending
);
2457 if (r10_sync_page_io(rdev
,
2458 r10_bio
->devs
[sl
].addr
+
2460 s
, conf
->tmppage
, WRITE
)
2462 /* Well, this device is dead */
2464 "md/raid10:%s: read correction "
2466 " (%d sectors at %llu on %s)\n",
2468 (unsigned long long)(
2470 choose_data_offset(r10_bio
,
2472 bdevname(rdev
->bdev
, b
));
2473 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2476 bdevname(rdev
->bdev
, b
));
2478 rdev_dec_pending(rdev
, mddev
);
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(In_sync
, &rdev
->flags
))
2494 atomic_inc(&rdev
->nr_pending
);
2496 switch (r10_sync_page_io(rdev
,
2497 r10_bio
->devs
[sl
].addr
+
2502 /* Well, this device is dead */
2504 "md/raid10:%s: unable to read back "
2506 " (%d sectors at %llu on %s)\n",
2508 (unsigned long long)(
2510 choose_data_offset(r10_bio
, rdev
)),
2511 bdevname(rdev
->bdev
, b
));
2512 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2515 bdevname(rdev
->bdev
, b
));
2519 "md/raid10:%s: read error corrected"
2520 " (%d sectors at %llu on %s)\n",
2522 (unsigned long long)(
2524 choose_data_offset(r10_bio
, rdev
)),
2525 bdevname(rdev
->bdev
, b
));
2526 atomic_add(s
, &rdev
->corrected_errors
);
2529 rdev_dec_pending(rdev
, mddev
);
2539 static void bi_complete(struct bio
*bio
, int error
)
2541 complete((struct completion
*)bio
->bi_private
);
2544 static int submit_bio_wait(int rw
, struct bio
*bio
)
2546 struct completion event
;
2549 init_completion(&event
);
2550 bio
->bi_private
= &event
;
2551 bio
->bi_end_io
= bi_complete
;
2552 submit_bio(rw
, bio
);
2553 wait_for_completion(&event
);
2555 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2558 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2560 struct bio
*bio
= r10_bio
->master_bio
;
2561 struct mddev
*mddev
= r10_bio
->mddev
;
2562 struct r10conf
*conf
= mddev
->private;
2563 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2564 /* bio has the data to be written to slot 'i' where
2565 * we just recently had a write error.
2566 * We repeatedly clone the bio and trim down to one block,
2567 * then try the write. Where the write fails we record
2569 * It is conceivable that the bio doesn't exactly align with
2570 * blocks. We must handle this.
2572 * We currently own a reference to the rdev.
2578 int sect_to_write
= r10_bio
->sectors
;
2581 if (rdev
->badblocks
.shift
< 0)
2584 block_sectors
= 1 << rdev
->badblocks
.shift
;
2585 sector
= r10_bio
->sector
;
2586 sectors
= ((r10_bio
->sector
+ block_sectors
)
2587 & ~(sector_t
)(block_sectors
- 1))
2590 while (sect_to_write
) {
2592 if (sectors
> sect_to_write
)
2593 sectors
= sect_to_write
;
2594 /* Write at 'sector' for 'sectors' */
2595 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2596 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2597 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2598 choose_data_offset(r10_bio
, rdev
) +
2599 (sector
- r10_bio
->sector
));
2600 wbio
->bi_bdev
= rdev
->bdev
;
2601 if (submit_bio_wait(WRITE
, wbio
) == 0)
2603 ok
= rdev_set_badblocks(rdev
, sector
,
2608 sect_to_write
-= sectors
;
2610 sectors
= block_sectors
;
2615 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2617 int slot
= r10_bio
->read_slot
;
2619 struct r10conf
*conf
= mddev
->private;
2620 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2621 char b
[BDEVNAME_SIZE
];
2622 unsigned long do_sync
;
2625 /* we got a read error. Maybe the drive is bad. Maybe just
2626 * the block and we can fix it.
2627 * We freeze all other IO, and try reading the block from
2628 * other devices. When we find one, we re-write
2629 * and check it that fixes the read error.
2630 * This is all done synchronously while the array is
2633 bio
= r10_bio
->devs
[slot
].bio
;
2634 bdevname(bio
->bi_bdev
, b
);
2636 r10_bio
->devs
[slot
].bio
= NULL
;
2638 if (mddev
->ro
== 0) {
2640 fix_read_error(conf
, mddev
, r10_bio
);
2641 unfreeze_array(conf
);
2643 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2645 rdev_dec_pending(rdev
, mddev
);
2648 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2650 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2651 " read error for block %llu\n",
2653 (unsigned long long)r10_bio
->sector
);
2654 raid_end_bio_io(r10_bio
);
2658 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2659 slot
= r10_bio
->read_slot
;
2662 "md/raid10:%s: %s: redirecting "
2663 "sector %llu to another mirror\n",
2665 bdevname(rdev
->bdev
, b
),
2666 (unsigned long long)r10_bio
->sector
);
2667 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2670 r10_bio
->sector
- bio
->bi_sector
,
2672 r10_bio
->devs
[slot
].bio
= bio
;
2673 r10_bio
->devs
[slot
].rdev
= rdev
;
2674 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2675 + choose_data_offset(r10_bio
, rdev
);
2676 bio
->bi_bdev
= rdev
->bdev
;
2677 bio
->bi_rw
= READ
| do_sync
;
2678 bio
->bi_private
= r10_bio
;
2679 bio
->bi_end_io
= raid10_end_read_request
;
2680 if (max_sectors
< r10_bio
->sectors
) {
2681 /* Drat - have to split this up more */
2682 struct bio
*mbio
= r10_bio
->master_bio
;
2683 int sectors_handled
=
2684 r10_bio
->sector
+ max_sectors
2686 r10_bio
->sectors
= max_sectors
;
2687 spin_lock_irq(&conf
->device_lock
);
2688 if (mbio
->bi_phys_segments
== 0)
2689 mbio
->bi_phys_segments
= 2;
2691 mbio
->bi_phys_segments
++;
2692 spin_unlock_irq(&conf
->device_lock
);
2693 generic_make_request(bio
);
2695 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2697 r10_bio
->master_bio
= mbio
;
2698 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2701 set_bit(R10BIO_ReadError
,
2703 r10_bio
->mddev
= mddev
;
2704 r10_bio
->sector
= mbio
->bi_sector
2709 generic_make_request(bio
);
2712 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2714 /* Some sort of write request has finished and it
2715 * succeeded in writing where we thought there was a
2716 * bad block. So forget the bad block.
2717 * Or possibly if failed and we need to record
2721 struct md_rdev
*rdev
;
2723 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2724 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2725 for (m
= 0; m
< conf
->copies
; m
++) {
2726 int dev
= r10_bio
->devs
[m
].devnum
;
2727 rdev
= conf
->mirrors
[dev
].rdev
;
2728 if (r10_bio
->devs
[m
].bio
== NULL
)
2730 if (test_bit(BIO_UPTODATE
,
2731 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2732 rdev_clear_badblocks(
2734 r10_bio
->devs
[m
].addr
,
2735 r10_bio
->sectors
, 0);
2737 if (!rdev_set_badblocks(
2739 r10_bio
->devs
[m
].addr
,
2740 r10_bio
->sectors
, 0))
2741 md_error(conf
->mddev
, rdev
);
2743 rdev
= conf
->mirrors
[dev
].replacement
;
2744 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2746 if (test_bit(BIO_UPTODATE
,
2747 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2748 rdev_clear_badblocks(
2750 r10_bio
->devs
[m
].addr
,
2751 r10_bio
->sectors
, 0);
2753 if (!rdev_set_badblocks(
2755 r10_bio
->devs
[m
].addr
,
2756 r10_bio
->sectors
, 0))
2757 md_error(conf
->mddev
, rdev
);
2762 for (m
= 0; m
< conf
->copies
; m
++) {
2763 int dev
= r10_bio
->devs
[m
].devnum
;
2764 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2765 rdev
= conf
->mirrors
[dev
].rdev
;
2766 if (bio
== IO_MADE_GOOD
) {
2767 rdev_clear_badblocks(
2769 r10_bio
->devs
[m
].addr
,
2770 r10_bio
->sectors
, 0);
2771 rdev_dec_pending(rdev
, conf
->mddev
);
2772 } else if (bio
!= NULL
&&
2773 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2774 if (!narrow_write_error(r10_bio
, m
)) {
2775 md_error(conf
->mddev
, rdev
);
2776 set_bit(R10BIO_Degraded
,
2779 rdev_dec_pending(rdev
, conf
->mddev
);
2781 bio
= r10_bio
->devs
[m
].repl_bio
;
2782 rdev
= conf
->mirrors
[dev
].replacement
;
2783 if (rdev
&& bio
== IO_MADE_GOOD
) {
2784 rdev_clear_badblocks(
2786 r10_bio
->devs
[m
].addr
,
2787 r10_bio
->sectors
, 0);
2788 rdev_dec_pending(rdev
, conf
->mddev
);
2791 if (test_bit(R10BIO_WriteError
,
2793 close_write(r10_bio
);
2794 raid_end_bio_io(r10_bio
);
2798 static void raid10d(struct md_thread
*thread
)
2800 struct mddev
*mddev
= thread
->mddev
;
2801 struct r10bio
*r10_bio
;
2802 unsigned long flags
;
2803 struct r10conf
*conf
= mddev
->private;
2804 struct list_head
*head
= &conf
->retry_list
;
2805 struct blk_plug plug
;
2807 md_check_recovery(mddev
);
2809 blk_start_plug(&plug
);
2812 flush_pending_writes(conf
);
2814 spin_lock_irqsave(&conf
->device_lock
, flags
);
2815 if (list_empty(head
)) {
2816 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2819 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2820 list_del(head
->prev
);
2822 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2824 mddev
= r10_bio
->mddev
;
2825 conf
= mddev
->private;
2826 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2827 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2828 handle_write_completed(conf
, r10_bio
);
2829 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2830 reshape_request_write(mddev
, r10_bio
);
2831 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2832 sync_request_write(mddev
, r10_bio
);
2833 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2834 recovery_request_write(mddev
, r10_bio
);
2835 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2836 handle_read_error(mddev
, r10_bio
);
2838 /* just a partial read to be scheduled from a
2841 int slot
= r10_bio
->read_slot
;
2842 generic_make_request(r10_bio
->devs
[slot
].bio
);
2846 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2847 md_check_recovery(mddev
);
2849 blk_finish_plug(&plug
);
2853 static int init_resync(struct r10conf
*conf
)
2858 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2859 BUG_ON(conf
->r10buf_pool
);
2860 conf
->have_replacement
= 0;
2861 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2862 if (conf
->mirrors
[i
].replacement
)
2863 conf
->have_replacement
= 1;
2864 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2865 if (!conf
->r10buf_pool
)
2867 conf
->next_resync
= 0;
2872 * perform a "sync" on one "block"
2874 * We need to make sure that no normal I/O request - particularly write
2875 * requests - conflict with active sync requests.
2877 * This is achieved by tracking pending requests and a 'barrier' concept
2878 * that can be installed to exclude normal IO requests.
2880 * Resync and recovery are handled very differently.
2881 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2883 * For resync, we iterate over virtual addresses, read all copies,
2884 * and update if there are differences. If only one copy is live,
2886 * For recovery, we iterate over physical addresses, read a good
2887 * value for each non-in_sync drive, and over-write.
2889 * So, for recovery we may have several outstanding complex requests for a
2890 * given address, one for each out-of-sync device. We model this by allocating
2891 * a number of r10_bio structures, one for each out-of-sync device.
2892 * As we setup these structures, we collect all bio's together into a list
2893 * which we then process collectively to add pages, and then process again
2894 * to pass to generic_make_request.
2896 * The r10_bio structures are linked using a borrowed master_bio pointer.
2897 * This link is counted in ->remaining. When the r10_bio that points to NULL
2898 * has its remaining count decremented to 0, the whole complex operation
2903 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2904 int *skipped
, int go_faster
)
2906 struct r10conf
*conf
= mddev
->private;
2907 struct r10bio
*r10_bio
;
2908 struct bio
*biolist
= NULL
, *bio
;
2909 sector_t max_sector
, nr_sectors
;
2912 sector_t sync_blocks
;
2913 sector_t sectors_skipped
= 0;
2914 int chunks_skipped
= 0;
2915 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2917 if (!conf
->r10buf_pool
)
2918 if (init_resync(conf
))
2922 * Allow skipping a full rebuild for incremental assembly
2923 * of a clean array, like RAID1 does.
2925 if (mddev
->bitmap
== NULL
&&
2926 mddev
->recovery_cp
== MaxSector
&&
2927 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2928 conf
->fullsync
== 0) {
2930 max_sector
= mddev
->dev_sectors
;
2931 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2932 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2933 max_sector
= mddev
->resync_max_sectors
;
2934 return max_sector
- sector_nr
;
2938 max_sector
= mddev
->dev_sectors
;
2939 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2940 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2941 max_sector
= mddev
->resync_max_sectors
;
2942 if (sector_nr
>= max_sector
) {
2943 /* If we aborted, we need to abort the
2944 * sync on the 'current' bitmap chucks (there can
2945 * be several when recovering multiple devices).
2946 * as we may have started syncing it but not finished.
2947 * We can find the current address in
2948 * mddev->curr_resync, but for recovery,
2949 * we need to convert that to several
2950 * virtual addresses.
2952 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2957 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2958 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2959 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2961 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2963 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2964 bitmap_end_sync(mddev
->bitmap
, sect
,
2968 /* completed sync */
2969 if ((!mddev
->bitmap
|| conf
->fullsync
)
2970 && conf
->have_replacement
2971 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2972 /* Completed a full sync so the replacements
2973 * are now fully recovered.
2975 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2976 if (conf
->mirrors
[i
].replacement
)
2977 conf
->mirrors
[i
].replacement
2983 bitmap_close_sync(mddev
->bitmap
);
2986 return sectors_skipped
;
2989 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2990 return reshape_request(mddev
, sector_nr
, skipped
);
2992 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2993 /* if there has been nothing to do on any drive,
2994 * then there is nothing to do at all..
2997 return (max_sector
- sector_nr
) + sectors_skipped
;
3000 if (max_sector
> mddev
->resync_max
)
3001 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
3003 /* make sure whole request will fit in a chunk - if chunks
3006 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
3007 max_sector
> (sector_nr
| chunk_mask
))
3008 max_sector
= (sector_nr
| chunk_mask
) + 1;
3010 * If there is non-resync activity waiting for us then
3011 * put in a delay to throttle resync.
3013 if (!go_faster
&& conf
->nr_waiting
)
3014 msleep_interruptible(1000);
3016 /* Again, very different code for resync and recovery.
3017 * Both must result in an r10bio with a list of bios that
3018 * have bi_end_io, bi_sector, bi_bdev set,
3019 * and bi_private set to the r10bio.
3020 * For recovery, we may actually create several r10bios
3021 * with 2 bios in each, that correspond to the bios in the main one.
3022 * In this case, the subordinate r10bios link back through a
3023 * borrowed master_bio pointer, and the counter in the master
3024 * includes a ref from each subordinate.
3026 /* First, we decide what to do and set ->bi_end_io
3027 * To end_sync_read if we want to read, and
3028 * end_sync_write if we will want to write.
3031 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
3032 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3033 /* recovery... the complicated one */
3037 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
3043 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
3045 if ((mirror
->rdev
== NULL
||
3046 test_bit(In_sync
, &mirror
->rdev
->flags
))
3048 (mirror
->replacement
== NULL
||
3050 &mirror
->replacement
->flags
)))
3054 /* want to reconstruct this device */
3056 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3057 if (sect
>= mddev
->resync_max_sectors
) {
3058 /* last stripe is not complete - don't
3059 * try to recover this sector.
3063 /* Unless we are doing a full sync, or a replacement
3064 * we only need to recover the block if it is set in
3067 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3069 if (sync_blocks
< max_sync
)
3070 max_sync
= sync_blocks
;
3072 mirror
->replacement
== NULL
&&
3074 /* yep, skip the sync_blocks here, but don't assume
3075 * that there will never be anything to do here
3077 chunks_skipped
= -1;
3081 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3082 raise_barrier(conf
, rb2
!= NULL
);
3083 atomic_set(&r10_bio
->remaining
, 0);
3085 r10_bio
->master_bio
= (struct bio
*)rb2
;
3087 atomic_inc(&rb2
->remaining
);
3088 r10_bio
->mddev
= mddev
;
3089 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3090 r10_bio
->sector
= sect
;
3092 raid10_find_phys(conf
, r10_bio
);
3094 /* Need to check if the array will still be
3097 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3098 if (conf
->mirrors
[j
].rdev
== NULL
||
3099 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3104 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3105 &sync_blocks
, still_degraded
);
3108 for (j
=0; j
<conf
->copies
;j
++) {
3110 int d
= r10_bio
->devs
[j
].devnum
;
3111 sector_t from_addr
, to_addr
;
3112 struct md_rdev
*rdev
;
3113 sector_t sector
, first_bad
;
3115 if (!conf
->mirrors
[d
].rdev
||
3116 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3118 /* This is where we read from */
3120 rdev
= conf
->mirrors
[d
].rdev
;
3121 sector
= r10_bio
->devs
[j
].addr
;
3123 if (is_badblock(rdev
, sector
, max_sync
,
3124 &first_bad
, &bad_sectors
)) {
3125 if (first_bad
> sector
)
3126 max_sync
= first_bad
- sector
;
3128 bad_sectors
-= (sector
3130 if (max_sync
> bad_sectors
)
3131 max_sync
= bad_sectors
;
3135 bio
= r10_bio
->devs
[0].bio
;
3136 bio
->bi_next
= biolist
;
3138 bio
->bi_private
= r10_bio
;
3139 bio
->bi_end_io
= end_sync_read
;
3141 from_addr
= r10_bio
->devs
[j
].addr
;
3142 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
3143 bio
->bi_bdev
= rdev
->bdev
;
3144 atomic_inc(&rdev
->nr_pending
);
3145 /* and we write to 'i' (if not in_sync) */
3147 for (k
=0; k
<conf
->copies
; k
++)
3148 if (r10_bio
->devs
[k
].devnum
== i
)
3150 BUG_ON(k
== conf
->copies
);
3151 to_addr
= r10_bio
->devs
[k
].addr
;
3152 r10_bio
->devs
[0].devnum
= d
;
3153 r10_bio
->devs
[0].addr
= from_addr
;
3154 r10_bio
->devs
[1].devnum
= i
;
3155 r10_bio
->devs
[1].addr
= to_addr
;
3157 rdev
= mirror
->rdev
;
3158 if (!test_bit(In_sync
, &rdev
->flags
)) {
3159 bio
= r10_bio
->devs
[1].bio
;
3160 bio
->bi_next
= biolist
;
3162 bio
->bi_private
= r10_bio
;
3163 bio
->bi_end_io
= end_sync_write
;
3165 bio
->bi_sector
= to_addr
3166 + rdev
->data_offset
;
3167 bio
->bi_bdev
= rdev
->bdev
;
3168 atomic_inc(&r10_bio
->remaining
);
3170 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3172 /* and maybe write to replacement */
3173 bio
= r10_bio
->devs
[1].repl_bio
;
3175 bio
->bi_end_io
= NULL
;
3176 rdev
= mirror
->replacement
;
3177 /* Note: if rdev != NULL, then bio
3178 * cannot be NULL as r10buf_pool_alloc will
3179 * have allocated it.
3180 * So the second test here is pointless.
3181 * But it keeps semantic-checkers happy, and
3182 * this comment keeps human reviewers
3185 if (rdev
== NULL
|| bio
== NULL
||
3186 test_bit(Faulty
, &rdev
->flags
))
3188 bio
->bi_next
= biolist
;
3190 bio
->bi_private
= r10_bio
;
3191 bio
->bi_end_io
= end_sync_write
;
3193 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3194 bio
->bi_bdev
= rdev
->bdev
;
3195 atomic_inc(&r10_bio
->remaining
);
3198 if (j
== conf
->copies
) {
3199 /* Cannot recover, so abort the recovery or
3200 * record a bad block */
3203 atomic_dec(&rb2
->remaining
);
3206 /* problem is that there are bad blocks
3207 * on other device(s)
3210 for (k
= 0; k
< conf
->copies
; k
++)
3211 if (r10_bio
->devs
[k
].devnum
== i
)
3213 if (!test_bit(In_sync
,
3214 &mirror
->rdev
->flags
)
3215 && !rdev_set_badblocks(
3217 r10_bio
->devs
[k
].addr
,
3220 if (mirror
->replacement
&&
3221 !rdev_set_badblocks(
3222 mirror
->replacement
,
3223 r10_bio
->devs
[k
].addr
,
3228 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3230 printk(KERN_INFO
"md/raid10:%s: insufficient "
3231 "working devices for recovery.\n",
3233 mirror
->recovery_disabled
3234 = mddev
->recovery_disabled
;
3239 if (biolist
== NULL
) {
3241 struct r10bio
*rb2
= r10_bio
;
3242 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3243 rb2
->master_bio
= NULL
;
3249 /* resync. Schedule a read for every block at this virt offset */
3252 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3254 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3255 &sync_blocks
, mddev
->degraded
) &&
3256 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3257 &mddev
->recovery
)) {
3258 /* We can skip this block */
3260 return sync_blocks
+ sectors_skipped
;
3262 if (sync_blocks
< max_sync
)
3263 max_sync
= sync_blocks
;
3264 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3266 r10_bio
->mddev
= mddev
;
3267 atomic_set(&r10_bio
->remaining
, 0);
3268 raise_barrier(conf
, 0);
3269 conf
->next_resync
= sector_nr
;
3271 r10_bio
->master_bio
= NULL
;
3272 r10_bio
->sector
= sector_nr
;
3273 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3274 raid10_find_phys(conf
, r10_bio
);
3275 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3277 for (i
= 0; i
< conf
->copies
; i
++) {
3278 int d
= r10_bio
->devs
[i
].devnum
;
3279 sector_t first_bad
, sector
;
3282 if (r10_bio
->devs
[i
].repl_bio
)
3283 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3285 bio
= r10_bio
->devs
[i
].bio
;
3286 bio
->bi_end_io
= NULL
;
3287 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3288 if (conf
->mirrors
[d
].rdev
== NULL
||
3289 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3291 sector
= r10_bio
->devs
[i
].addr
;
3292 if (is_badblock(conf
->mirrors
[d
].rdev
,
3294 &first_bad
, &bad_sectors
)) {
3295 if (first_bad
> sector
)
3296 max_sync
= first_bad
- sector
;
3298 bad_sectors
-= (sector
- first_bad
);
3299 if (max_sync
> bad_sectors
)
3300 max_sync
= bad_sectors
;
3304 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3305 atomic_inc(&r10_bio
->remaining
);
3306 bio
->bi_next
= biolist
;
3308 bio
->bi_private
= r10_bio
;
3309 bio
->bi_end_io
= end_sync_read
;
3311 bio
->bi_sector
= sector
+
3312 conf
->mirrors
[d
].rdev
->data_offset
;
3313 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3316 if (conf
->mirrors
[d
].replacement
== NULL
||
3318 &conf
->mirrors
[d
].replacement
->flags
))
3321 /* Need to set up for writing to the replacement */
3322 bio
= r10_bio
->devs
[i
].repl_bio
;
3323 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3325 sector
= r10_bio
->devs
[i
].addr
;
3326 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3327 bio
->bi_next
= biolist
;
3329 bio
->bi_private
= r10_bio
;
3330 bio
->bi_end_io
= end_sync_write
;
3332 bio
->bi_sector
= sector
+
3333 conf
->mirrors
[d
].replacement
->data_offset
;
3334 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3339 for (i
=0; i
<conf
->copies
; i
++) {
3340 int d
= r10_bio
->devs
[i
].devnum
;
3341 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3342 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3344 if (r10_bio
->devs
[i
].repl_bio
&&
3345 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3347 conf
->mirrors
[d
].replacement
,
3356 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3358 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
3360 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
3363 bio
->bi_phys_segments
= 0;
3368 if (sector_nr
+ max_sync
< max_sector
)
3369 max_sector
= sector_nr
+ max_sync
;
3372 int len
= PAGE_SIZE
;
3373 if (sector_nr
+ (len
>>9) > max_sector
)
3374 len
= (max_sector
- sector_nr
) << 9;
3377 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3379 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3380 if (bio_add_page(bio
, page
, len
, 0))
3384 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3385 for (bio2
= biolist
;
3386 bio2
&& bio2
!= bio
;
3387 bio2
= bio2
->bi_next
) {
3388 /* remove last page from this bio */
3390 bio2
->bi_size
-= len
;
3391 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3395 nr_sectors
+= len
>>9;
3396 sector_nr
+= len
>>9;
3397 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3399 r10_bio
->sectors
= nr_sectors
;
3403 biolist
= biolist
->bi_next
;
3405 bio
->bi_next
= NULL
;
3406 r10_bio
= bio
->bi_private
;
3407 r10_bio
->sectors
= nr_sectors
;
3409 if (bio
->bi_end_io
== end_sync_read
) {
3410 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3411 generic_make_request(bio
);
3415 if (sectors_skipped
)
3416 /* pretend they weren't skipped, it makes
3417 * no important difference in this case
3419 md_done_sync(mddev
, sectors_skipped
, 1);
3421 return sectors_skipped
+ nr_sectors
;
3423 /* There is nowhere to write, so all non-sync
3424 * drives must be failed or in resync, all drives
3425 * have a bad block, so try the next chunk...
3427 if (sector_nr
+ max_sync
< max_sector
)
3428 max_sector
= sector_nr
+ max_sync
;
3430 sectors_skipped
+= (max_sector
- sector_nr
);
3432 sector_nr
= max_sector
;
3437 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3440 struct r10conf
*conf
= mddev
->private;
3443 raid_disks
= min(conf
->geo
.raid_disks
,
3444 conf
->prev
.raid_disks
);
3446 sectors
= conf
->dev_sectors
;
3448 size
= sectors
>> conf
->geo
.chunk_shift
;
3449 sector_div(size
, conf
->geo
.far_copies
);
3450 size
= size
* raid_disks
;
3451 sector_div(size
, conf
->geo
.near_copies
);
3453 return size
<< conf
->geo
.chunk_shift
;
3456 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3458 /* Calculate the number of sectors-per-device that will
3459 * actually be used, and set conf->dev_sectors and
3463 size
= size
>> conf
->geo
.chunk_shift
;
3464 sector_div(size
, conf
->geo
.far_copies
);
3465 size
= size
* conf
->geo
.raid_disks
;
3466 sector_div(size
, conf
->geo
.near_copies
);
3467 /* 'size' is now the number of chunks in the array */
3468 /* calculate "used chunks per device" */
3469 size
= size
* conf
->copies
;
3471 /* We need to round up when dividing by raid_disks to
3472 * get the stride size.
3474 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3476 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3478 if (conf
->geo
.far_offset
)
3479 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3481 sector_div(size
, conf
->geo
.far_copies
);
3482 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3486 enum geo_type
{geo_new
, geo_old
, geo_start
};
3487 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3490 int layout
, chunk
, disks
;
3493 layout
= mddev
->layout
;
3494 chunk
= mddev
->chunk_sectors
;
3495 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3498 layout
= mddev
->new_layout
;
3499 chunk
= mddev
->new_chunk_sectors
;
3500 disks
= mddev
->raid_disks
;
3502 default: /* avoid 'may be unused' warnings */
3503 case geo_start
: /* new when starting reshape - raid_disks not
3505 layout
= mddev
->new_layout
;
3506 chunk
= mddev
->new_chunk_sectors
;
3507 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3512 if (chunk
< (PAGE_SIZE
>> 9) ||
3513 !is_power_of_2(chunk
))
3516 fc
= (layout
>> 8) & 255;
3517 fo
= layout
& (1<<16);
3518 geo
->raid_disks
= disks
;
3519 geo
->near_copies
= nc
;
3520 geo
->far_copies
= fc
;
3521 geo
->far_offset
= fo
;
3522 geo
->far_set_size
= (layout
& (1<<17)) ? disks
/ fc
: disks
;
3523 geo
->chunk_mask
= chunk
- 1;
3524 geo
->chunk_shift
= ffz(~chunk
);
3528 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3530 struct r10conf
*conf
= NULL
;
3535 copies
= setup_geo(&geo
, mddev
, geo_new
);
3538 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3539 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3540 mdname(mddev
), PAGE_SIZE
);
3544 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3545 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3546 mdname(mddev
), mddev
->new_layout
);
3551 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3555 /* FIXME calc properly */
3556 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3557 max(0,mddev
->delta_disks
)),
3562 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3567 conf
->copies
= copies
;
3568 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3569 r10bio_pool_free
, conf
);
3570 if (!conf
->r10bio_pool
)
3573 calc_sectors(conf
, mddev
->dev_sectors
);
3574 if (mddev
->reshape_position
== MaxSector
) {
3575 conf
->prev
= conf
->geo
;
3576 conf
->reshape_progress
= MaxSector
;
3578 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3582 conf
->reshape_progress
= mddev
->reshape_position
;
3583 if (conf
->prev
.far_offset
)
3584 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3586 /* far_copies must be 1 */
3587 conf
->prev
.stride
= conf
->dev_sectors
;
3589 spin_lock_init(&conf
->device_lock
);
3590 INIT_LIST_HEAD(&conf
->retry_list
);
3592 spin_lock_init(&conf
->resync_lock
);
3593 init_waitqueue_head(&conf
->wait_barrier
);
3595 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3599 conf
->mddev
= mddev
;
3604 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3607 if (conf
->r10bio_pool
)
3608 mempool_destroy(conf
->r10bio_pool
);
3609 kfree(conf
->mirrors
);
3610 safe_put_page(conf
->tmppage
);
3613 return ERR_PTR(err
);
3616 static int run(struct mddev
*mddev
)
3618 struct r10conf
*conf
;
3619 int i
, disk_idx
, chunk_size
;
3620 struct raid10_info
*disk
;
3621 struct md_rdev
*rdev
;
3623 sector_t min_offset_diff
= 0;
3625 bool discard_supported
= false;
3627 if (mddev
->private == NULL
) {
3628 conf
= setup_conf(mddev
);
3630 return PTR_ERR(conf
);
3631 mddev
->private = conf
;
3633 conf
= mddev
->private;
3637 mddev
->thread
= conf
->thread
;
3638 conf
->thread
= NULL
;
3640 chunk_size
= mddev
->chunk_sectors
<< 9;
3642 blk_queue_max_discard_sectors(mddev
->queue
,
3643 mddev
->chunk_sectors
);
3644 blk_queue_max_write_same_sectors(mddev
->queue
,
3645 mddev
->chunk_sectors
);
3646 blk_queue_io_min(mddev
->queue
, chunk_size
);
3647 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3648 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3650 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3651 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3654 rdev_for_each(rdev
, mddev
) {
3656 struct request_queue
*q
;
3658 disk_idx
= rdev
->raid_disk
;
3661 if (disk_idx
>= conf
->geo
.raid_disks
&&
3662 disk_idx
>= conf
->prev
.raid_disks
)
3664 disk
= conf
->mirrors
+ disk_idx
;
3666 if (test_bit(Replacement
, &rdev
->flags
)) {
3667 if (disk
->replacement
)
3669 disk
->replacement
= rdev
;
3675 q
= bdev_get_queue(rdev
->bdev
);
3676 if (q
->merge_bvec_fn
)
3677 mddev
->merge_check_needed
= 1;
3678 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3679 if (!mddev
->reshape_backwards
)
3683 if (first
|| diff
< min_offset_diff
)
3684 min_offset_diff
= diff
;
3687 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3688 rdev
->data_offset
<< 9);
3690 disk
->head_position
= 0;
3692 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3693 discard_supported
= true;
3697 if (discard_supported
)
3698 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3701 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3704 /* need to check that every block has at least one working mirror */
3705 if (!enough(conf
, -1)) {
3706 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3711 if (conf
->reshape_progress
!= MaxSector
) {
3712 /* must ensure that shape change is supported */
3713 if (conf
->geo
.far_copies
!= 1 &&
3714 conf
->geo
.far_offset
== 0)
3716 if (conf
->prev
.far_copies
!= 1 &&
3717 conf
->geo
.far_offset
== 0)
3721 mddev
->degraded
= 0;
3723 i
< conf
->geo
.raid_disks
3724 || i
< conf
->prev
.raid_disks
;
3727 disk
= conf
->mirrors
+ i
;
3729 if (!disk
->rdev
&& disk
->replacement
) {
3730 /* The replacement is all we have - use it */
3731 disk
->rdev
= disk
->replacement
;
3732 disk
->replacement
= NULL
;
3733 clear_bit(Replacement
, &disk
->rdev
->flags
);
3737 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3738 disk
->head_position
= 0;
3743 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3746 if (mddev
->recovery_cp
!= MaxSector
)
3747 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3748 " -- starting background reconstruction\n",
3751 "md/raid10:%s: active with %d out of %d devices\n",
3752 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3753 conf
->geo
.raid_disks
);
3755 * Ok, everything is just fine now
3757 mddev
->dev_sectors
= conf
->dev_sectors
;
3758 size
= raid10_size(mddev
, 0, 0);
3759 md_set_array_sectors(mddev
, size
);
3760 mddev
->resync_max_sectors
= size
;
3763 int stripe
= conf
->geo
.raid_disks
*
3764 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3765 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3766 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3768 /* Calculate max read-ahead size.
3769 * We need to readahead at least twice a whole stripe....
3772 stripe
/= conf
->geo
.near_copies
;
3773 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3774 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3775 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3779 if (md_integrity_register(mddev
))
3782 if (conf
->reshape_progress
!= MaxSector
) {
3783 unsigned long before_length
, after_length
;
3785 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3786 conf
->prev
.far_copies
);
3787 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3788 conf
->geo
.far_copies
);
3790 if (max(before_length
, after_length
) > min_offset_diff
) {
3791 /* This cannot work */
3792 printk("md/raid10: offset difference not enough to continue reshape\n");
3795 conf
->offset_diff
= min_offset_diff
;
3797 conf
->reshape_safe
= conf
->reshape_progress
;
3798 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3799 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3800 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3801 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3802 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3809 md_unregister_thread(&mddev
->thread
);
3810 if (conf
->r10bio_pool
)
3811 mempool_destroy(conf
->r10bio_pool
);
3812 safe_put_page(conf
->tmppage
);
3813 kfree(conf
->mirrors
);
3815 mddev
->private = NULL
;
3820 static int stop(struct mddev
*mddev
)
3822 struct r10conf
*conf
= mddev
->private;
3824 raise_barrier(conf
, 0);
3825 lower_barrier(conf
);
3827 md_unregister_thread(&mddev
->thread
);
3829 /* the unplug fn references 'conf'*/
3830 blk_sync_queue(mddev
->queue
);
3832 if (conf
->r10bio_pool
)
3833 mempool_destroy(conf
->r10bio_pool
);
3834 safe_put_page(conf
->tmppage
);
3835 kfree(conf
->mirrors
);
3837 mddev
->private = NULL
;
3841 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3843 struct r10conf
*conf
= mddev
->private;
3847 raise_barrier(conf
, 0);
3850 lower_barrier(conf
);
3855 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3857 /* Resize of 'far' arrays is not supported.
3858 * For 'near' and 'offset' arrays we can set the
3859 * number of sectors used to be an appropriate multiple
3860 * of the chunk size.
3861 * For 'offset', this is far_copies*chunksize.
3862 * For 'near' the multiplier is the LCM of
3863 * near_copies and raid_disks.
3864 * So if far_copies > 1 && !far_offset, fail.
3865 * Else find LCM(raid_disks, near_copy)*far_copies and
3866 * multiply by chunk_size. Then round to this number.
3867 * This is mostly done by raid10_size()
3869 struct r10conf
*conf
= mddev
->private;
3870 sector_t oldsize
, size
;
3872 if (mddev
->reshape_position
!= MaxSector
)
3875 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3878 oldsize
= raid10_size(mddev
, 0, 0);
3879 size
= raid10_size(mddev
, sectors
, 0);
3880 if (mddev
->external_size
&&
3881 mddev
->array_sectors
> size
)
3883 if (mddev
->bitmap
) {
3884 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3888 md_set_array_sectors(mddev
, size
);
3889 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3890 revalidate_disk(mddev
->gendisk
);
3891 if (sectors
> mddev
->dev_sectors
&&
3892 mddev
->recovery_cp
> oldsize
) {
3893 mddev
->recovery_cp
= oldsize
;
3894 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3896 calc_sectors(conf
, sectors
);
3897 mddev
->dev_sectors
= conf
->dev_sectors
;
3898 mddev
->resync_max_sectors
= size
;
3902 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3904 struct md_rdev
*rdev
;
3905 struct r10conf
*conf
;
3907 if (mddev
->degraded
> 0) {
3908 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3910 return ERR_PTR(-EINVAL
);
3913 /* Set new parameters */
3914 mddev
->new_level
= 10;
3915 /* new layout: far_copies = 1, near_copies = 2 */
3916 mddev
->new_layout
= (1<<8) + 2;
3917 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3918 mddev
->delta_disks
= mddev
->raid_disks
;
3919 mddev
->raid_disks
*= 2;
3920 /* make sure it will be not marked as dirty */
3921 mddev
->recovery_cp
= MaxSector
;
3923 conf
= setup_conf(mddev
);
3924 if (!IS_ERR(conf
)) {
3925 rdev_for_each(rdev
, mddev
)
3926 if (rdev
->raid_disk
>= 0)
3927 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3934 static void *raid10_takeover(struct mddev
*mddev
)
3936 struct r0conf
*raid0_conf
;
3938 /* raid10 can take over:
3939 * raid0 - providing it has only two drives
3941 if (mddev
->level
== 0) {
3942 /* for raid0 takeover only one zone is supported */
3943 raid0_conf
= mddev
->private;
3944 if (raid0_conf
->nr_strip_zones
> 1) {
3945 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3946 " with more than one zone.\n",
3948 return ERR_PTR(-EINVAL
);
3950 return raid10_takeover_raid0(mddev
);
3952 return ERR_PTR(-EINVAL
);
3955 static int raid10_check_reshape(struct mddev
*mddev
)
3957 /* Called when there is a request to change
3958 * - layout (to ->new_layout)
3959 * - chunk size (to ->new_chunk_sectors)
3960 * - raid_disks (by delta_disks)
3961 * or when trying to restart a reshape that was ongoing.
3963 * We need to validate the request and possibly allocate
3964 * space if that might be an issue later.
3966 * Currently we reject any reshape of a 'far' mode array,
3967 * allow chunk size to change if new is generally acceptable,
3968 * allow raid_disks to increase, and allow
3969 * a switch between 'near' mode and 'offset' mode.
3971 struct r10conf
*conf
= mddev
->private;
3974 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3977 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3978 /* mustn't change number of copies */
3980 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3981 /* Cannot switch to 'far' mode */
3984 if (mddev
->array_sectors
& geo
.chunk_mask
)
3985 /* not factor of array size */
3988 if (!enough(conf
, -1))
3991 kfree(conf
->mirrors_new
);
3992 conf
->mirrors_new
= NULL
;
3993 if (mddev
->delta_disks
> 0) {
3994 /* allocate new 'mirrors' list */
3995 conf
->mirrors_new
= kzalloc(
3996 sizeof(struct raid10_info
)
3997 *(mddev
->raid_disks
+
3998 mddev
->delta_disks
),
4000 if (!conf
->mirrors_new
)
4007 * Need to check if array has failed when deciding whether to:
4009 * - remove non-faulty devices
4012 * This determination is simple when no reshape is happening.
4013 * However if there is a reshape, we need to carefully check
4014 * both the before and after sections.
4015 * This is because some failed devices may only affect one
4016 * of the two sections, and some non-in_sync devices may
4017 * be insync in the section most affected by failed devices.
4019 static int calc_degraded(struct r10conf
*conf
)
4021 int degraded
, degraded2
;
4026 /* 'prev' section first */
4027 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4028 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4029 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4031 else if (!test_bit(In_sync
, &rdev
->flags
))
4032 /* When we can reduce the number of devices in
4033 * an array, this might not contribute to
4034 * 'degraded'. It does now.
4039 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4043 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4044 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4045 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4047 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4048 /* If reshape is increasing the number of devices,
4049 * this section has already been recovered, so
4050 * it doesn't contribute to degraded.
4053 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4058 if (degraded2
> degraded
)
4063 static int raid10_start_reshape(struct mddev
*mddev
)
4065 /* A 'reshape' has been requested. This commits
4066 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4067 * This also checks if there are enough spares and adds them
4069 * We currently require enough spares to make the final
4070 * array non-degraded. We also require that the difference
4071 * between old and new data_offset - on each device - is
4072 * enough that we never risk over-writing.
4075 unsigned long before_length
, after_length
;
4076 sector_t min_offset_diff
= 0;
4079 struct r10conf
*conf
= mddev
->private;
4080 struct md_rdev
*rdev
;
4084 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4087 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4090 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4091 conf
->prev
.far_copies
);
4092 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4093 conf
->geo
.far_copies
);
4095 rdev_for_each(rdev
, mddev
) {
4096 if (!test_bit(In_sync
, &rdev
->flags
)
4097 && !test_bit(Faulty
, &rdev
->flags
))
4099 if (rdev
->raid_disk
>= 0) {
4100 long long diff
= (rdev
->new_data_offset
4101 - rdev
->data_offset
);
4102 if (!mddev
->reshape_backwards
)
4106 if (first
|| diff
< min_offset_diff
)
4107 min_offset_diff
= diff
;
4111 if (max(before_length
, after_length
) > min_offset_diff
)
4114 if (spares
< mddev
->delta_disks
)
4117 conf
->offset_diff
= min_offset_diff
;
4118 spin_lock_irq(&conf
->device_lock
);
4119 if (conf
->mirrors_new
) {
4120 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4121 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4123 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
4124 conf
->mirrors_old
= conf
->mirrors
;
4125 conf
->mirrors
= conf
->mirrors_new
;
4126 conf
->mirrors_new
= NULL
;
4128 setup_geo(&conf
->geo
, mddev
, geo_start
);
4130 if (mddev
->reshape_backwards
) {
4131 sector_t size
= raid10_size(mddev
, 0, 0);
4132 if (size
< mddev
->array_sectors
) {
4133 spin_unlock_irq(&conf
->device_lock
);
4134 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4138 mddev
->resync_max_sectors
= size
;
4139 conf
->reshape_progress
= size
;
4141 conf
->reshape_progress
= 0;
4142 spin_unlock_irq(&conf
->device_lock
);
4144 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4145 ret
= bitmap_resize(mddev
->bitmap
,
4146 raid10_size(mddev
, 0,
4147 conf
->geo
.raid_disks
),
4152 if (mddev
->delta_disks
> 0) {
4153 rdev_for_each(rdev
, mddev
)
4154 if (rdev
->raid_disk
< 0 &&
4155 !test_bit(Faulty
, &rdev
->flags
)) {
4156 if (raid10_add_disk(mddev
, rdev
) == 0) {
4157 if (rdev
->raid_disk
>=
4158 conf
->prev
.raid_disks
)
4159 set_bit(In_sync
, &rdev
->flags
);
4161 rdev
->recovery_offset
= 0;
4163 if (sysfs_link_rdev(mddev
, rdev
))
4164 /* Failure here is OK */;
4166 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4167 && !test_bit(Faulty
, &rdev
->flags
)) {
4168 /* This is a spare that was manually added */
4169 set_bit(In_sync
, &rdev
->flags
);
4172 /* When a reshape changes the number of devices,
4173 * ->degraded is measured against the larger of the
4174 * pre and post numbers.
4176 spin_lock_irq(&conf
->device_lock
);
4177 mddev
->degraded
= calc_degraded(conf
);
4178 spin_unlock_irq(&conf
->device_lock
);
4179 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4180 mddev
->reshape_position
= conf
->reshape_progress
;
4181 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4183 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4184 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4185 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4186 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4188 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4190 if (!mddev
->sync_thread
) {
4194 conf
->reshape_checkpoint
= jiffies
;
4195 md_wakeup_thread(mddev
->sync_thread
);
4196 md_new_event(mddev
);
4200 mddev
->recovery
= 0;
4201 spin_lock_irq(&conf
->device_lock
);
4202 conf
->geo
= conf
->prev
;
4203 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4204 rdev_for_each(rdev
, mddev
)
4205 rdev
->new_data_offset
= rdev
->data_offset
;
4207 conf
->reshape_progress
= MaxSector
;
4208 mddev
->reshape_position
= MaxSector
;
4209 spin_unlock_irq(&conf
->device_lock
);
4213 /* Calculate the last device-address that could contain
4214 * any block from the chunk that includes the array-address 's'
4215 * and report the next address.
4216 * i.e. the address returned will be chunk-aligned and after
4217 * any data that is in the chunk containing 's'.
4219 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4221 s
= (s
| geo
->chunk_mask
) + 1;
4222 s
>>= geo
->chunk_shift
;
4223 s
*= geo
->near_copies
;
4224 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4225 s
*= geo
->far_copies
;
4226 s
<<= geo
->chunk_shift
;
4230 /* Calculate the first device-address that could contain
4231 * any block from the chunk that includes the array-address 's'.
4232 * This too will be the start of a chunk
4234 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4236 s
>>= geo
->chunk_shift
;
4237 s
*= geo
->near_copies
;
4238 sector_div(s
, geo
->raid_disks
);
4239 s
*= geo
->far_copies
;
4240 s
<<= geo
->chunk_shift
;
4244 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4247 /* We simply copy at most one chunk (smallest of old and new)
4248 * at a time, possibly less if that exceeds RESYNC_PAGES,
4249 * or we hit a bad block or something.
4250 * This might mean we pause for normal IO in the middle of
4251 * a chunk, but that is not a problem was mddev->reshape_position
4252 * can record any location.
4254 * If we will want to write to a location that isn't
4255 * yet recorded as 'safe' (i.e. in metadata on disk) then
4256 * we need to flush all reshape requests and update the metadata.
4258 * When reshaping forwards (e.g. to more devices), we interpret
4259 * 'safe' as the earliest block which might not have been copied
4260 * down yet. We divide this by previous stripe size and multiply
4261 * by previous stripe length to get lowest device offset that we
4262 * cannot write to yet.
4263 * We interpret 'sector_nr' as an address that we want to write to.
4264 * From this we use last_device_address() to find where we might
4265 * write to, and first_device_address on the 'safe' position.
4266 * If this 'next' write position is after the 'safe' position,
4267 * we must update the metadata to increase the 'safe' position.
4269 * When reshaping backwards, we round in the opposite direction
4270 * and perform the reverse test: next write position must not be
4271 * less than current safe position.
4273 * In all this the minimum difference in data offsets
4274 * (conf->offset_diff - always positive) allows a bit of slack,
4275 * so next can be after 'safe', but not by more than offset_disk
4277 * We need to prepare all the bios here before we start any IO
4278 * to ensure the size we choose is acceptable to all devices.
4279 * The means one for each copy for write-out and an extra one for
4281 * We store the read-in bio in ->master_bio and the others in
4282 * ->devs[x].bio and ->devs[x].repl_bio.
4284 struct r10conf
*conf
= mddev
->private;
4285 struct r10bio
*r10_bio
;
4286 sector_t next
, safe
, last
;
4290 struct md_rdev
*rdev
;
4293 struct bio
*bio
, *read_bio
;
4294 int sectors_done
= 0;
4296 if (sector_nr
== 0) {
4297 /* If restarting in the middle, skip the initial sectors */
4298 if (mddev
->reshape_backwards
&&
4299 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4300 sector_nr
= (raid10_size(mddev
, 0, 0)
4301 - conf
->reshape_progress
);
4302 } else if (!mddev
->reshape_backwards
&&
4303 conf
->reshape_progress
> 0)
4304 sector_nr
= conf
->reshape_progress
;
4306 mddev
->curr_resync_completed
= sector_nr
;
4307 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4313 /* We don't use sector_nr to track where we are up to
4314 * as that doesn't work well for ->reshape_backwards.
4315 * So just use ->reshape_progress.
4317 if (mddev
->reshape_backwards
) {
4318 /* 'next' is the earliest device address that we might
4319 * write to for this chunk in the new layout
4321 next
= first_dev_address(conf
->reshape_progress
- 1,
4324 /* 'safe' is the last device address that we might read from
4325 * in the old layout after a restart
4327 safe
= last_dev_address(conf
->reshape_safe
- 1,
4330 if (next
+ conf
->offset_diff
< safe
)
4333 last
= conf
->reshape_progress
- 1;
4334 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4335 & conf
->prev
.chunk_mask
);
4336 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4337 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4339 /* 'next' is after the last device address that we
4340 * might write to for this chunk in the new layout
4342 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4344 /* 'safe' is the earliest device address that we might
4345 * read from in the old layout after a restart
4347 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4349 /* Need to update metadata if 'next' might be beyond 'safe'
4350 * as that would possibly corrupt data
4352 if (next
> safe
+ conf
->offset_diff
)
4355 sector_nr
= conf
->reshape_progress
;
4356 last
= sector_nr
| (conf
->geo
.chunk_mask
4357 & conf
->prev
.chunk_mask
);
4359 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4360 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4364 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4365 /* Need to update reshape_position in metadata */
4367 mddev
->reshape_position
= conf
->reshape_progress
;
4368 if (mddev
->reshape_backwards
)
4369 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4370 - conf
->reshape_progress
;
4372 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4373 conf
->reshape_checkpoint
= jiffies
;
4374 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4375 md_wakeup_thread(mddev
->thread
);
4376 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4377 kthread_should_stop());
4378 conf
->reshape_safe
= mddev
->reshape_position
;
4379 allow_barrier(conf
);
4383 /* Now schedule reads for blocks from sector_nr to last */
4384 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4385 raise_barrier(conf
, sectors_done
!= 0);
4386 atomic_set(&r10_bio
->remaining
, 0);
4387 r10_bio
->mddev
= mddev
;
4388 r10_bio
->sector
= sector_nr
;
4389 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4390 r10_bio
->sectors
= last
- sector_nr
+ 1;
4391 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4392 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4395 /* Cannot read from here, so need to record bad blocks
4396 * on all the target devices.
4399 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4400 return sectors_done
;
4403 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4405 read_bio
->bi_bdev
= rdev
->bdev
;
4406 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4407 + rdev
->data_offset
);
4408 read_bio
->bi_private
= r10_bio
;
4409 read_bio
->bi_end_io
= end_sync_read
;
4410 read_bio
->bi_rw
= READ
;
4411 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4412 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4413 read_bio
->bi_vcnt
= 0;
4414 read_bio
->bi_idx
= 0;
4415 read_bio
->bi_size
= 0;
4416 r10_bio
->master_bio
= read_bio
;
4417 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4419 /* Now find the locations in the new layout */
4420 __raid10_find_phys(&conf
->geo
, r10_bio
);
4423 read_bio
->bi_next
= NULL
;
4425 for (s
= 0; s
< conf
->copies
*2; s
++) {
4427 int d
= r10_bio
->devs
[s
/2].devnum
;
4428 struct md_rdev
*rdev2
;
4430 rdev2
= conf
->mirrors
[d
].replacement
;
4431 b
= r10_bio
->devs
[s
/2].repl_bio
;
4433 rdev2
= conf
->mirrors
[d
].rdev
;
4434 b
= r10_bio
->devs
[s
/2].bio
;
4436 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4438 b
->bi_bdev
= rdev2
->bdev
;
4439 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4440 b
->bi_private
= r10_bio
;
4441 b
->bi_end_io
= end_reshape_write
;
4443 b
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4444 b
->bi_flags
|= 1 << BIO_UPTODATE
;
4452 /* Now add as many pages as possible to all of these bios. */
4455 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4456 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4457 int len
= (max_sectors
- s
) << 9;
4458 if (len
> PAGE_SIZE
)
4460 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4462 if (bio_add_page(bio
, page
, len
, 0))
4465 /* Didn't fit, must stop */
4467 bio2
&& bio2
!= bio
;
4468 bio2
= bio2
->bi_next
) {
4469 /* Remove last page from this bio */
4471 bio2
->bi_size
-= len
;
4472 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4476 sector_nr
+= len
>> 9;
4477 nr_sectors
+= len
>> 9;
4480 r10_bio
->sectors
= nr_sectors
;
4482 /* Now submit the read */
4483 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4484 atomic_inc(&r10_bio
->remaining
);
4485 read_bio
->bi_next
= NULL
;
4486 generic_make_request(read_bio
);
4487 sector_nr
+= nr_sectors
;
4488 sectors_done
+= nr_sectors
;
4489 if (sector_nr
<= last
)
4492 /* Now that we have done the whole section we can
4493 * update reshape_progress
4495 if (mddev
->reshape_backwards
)
4496 conf
->reshape_progress
-= sectors_done
;
4498 conf
->reshape_progress
+= sectors_done
;
4500 return sectors_done
;
4503 static void end_reshape_request(struct r10bio
*r10_bio
);
4504 static int handle_reshape_read_error(struct mddev
*mddev
,
4505 struct r10bio
*r10_bio
);
4506 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4508 /* Reshape read completed. Hopefully we have a block
4510 * If we got a read error then we do sync 1-page reads from
4511 * elsewhere until we find the data - or give up.
4513 struct r10conf
*conf
= mddev
->private;
4516 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4517 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4518 /* Reshape has been aborted */
4519 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4523 /* We definitely have the data in the pages, schedule the
4526 atomic_set(&r10_bio
->remaining
, 1);
4527 for (s
= 0; s
< conf
->copies
*2; s
++) {
4529 int d
= r10_bio
->devs
[s
/2].devnum
;
4530 struct md_rdev
*rdev
;
4532 rdev
= conf
->mirrors
[d
].replacement
;
4533 b
= r10_bio
->devs
[s
/2].repl_bio
;
4535 rdev
= conf
->mirrors
[d
].rdev
;
4536 b
= r10_bio
->devs
[s
/2].bio
;
4538 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4540 atomic_inc(&rdev
->nr_pending
);
4541 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4542 atomic_inc(&r10_bio
->remaining
);
4544 generic_make_request(b
);
4546 end_reshape_request(r10_bio
);
4549 static void end_reshape(struct r10conf
*conf
)
4551 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4554 spin_lock_irq(&conf
->device_lock
);
4555 conf
->prev
= conf
->geo
;
4556 md_finish_reshape(conf
->mddev
);
4558 conf
->reshape_progress
= MaxSector
;
4559 spin_unlock_irq(&conf
->device_lock
);
4561 /* read-ahead size must cover two whole stripes, which is
4562 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4564 if (conf
->mddev
->queue
) {
4565 int stripe
= conf
->geo
.raid_disks
*
4566 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4567 stripe
/= conf
->geo
.near_copies
;
4568 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4569 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4575 static int handle_reshape_read_error(struct mddev
*mddev
,
4576 struct r10bio
*r10_bio
)
4578 /* Use sync reads to get the blocks from somewhere else */
4579 int sectors
= r10_bio
->sectors
;
4580 struct r10conf
*conf
= mddev
->private;
4582 struct r10bio r10_bio
;
4583 struct r10dev devs
[conf
->copies
];
4585 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4588 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4590 r10b
->sector
= r10_bio
->sector
;
4591 __raid10_find_phys(&conf
->prev
, r10b
);
4596 int first_slot
= slot
;
4598 if (s
> (PAGE_SIZE
>> 9))
4602 int d
= r10b
->devs
[slot
].devnum
;
4603 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4606 test_bit(Faulty
, &rdev
->flags
) ||
4607 !test_bit(In_sync
, &rdev
->flags
))
4610 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4611 success
= sync_page_io(rdev
,
4620 if (slot
>= conf
->copies
)
4622 if (slot
== first_slot
)
4626 /* couldn't read this block, must give up */
4627 set_bit(MD_RECOVERY_INTR
,
4637 static void end_reshape_write(struct bio
*bio
, int error
)
4639 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4640 struct r10bio
*r10_bio
= bio
->bi_private
;
4641 struct mddev
*mddev
= r10_bio
->mddev
;
4642 struct r10conf
*conf
= mddev
->private;
4646 struct md_rdev
*rdev
= NULL
;
4648 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4650 rdev
= conf
->mirrors
[d
].replacement
;
4653 rdev
= conf
->mirrors
[d
].rdev
;
4657 /* FIXME should record badblock */
4658 md_error(mddev
, rdev
);
4661 rdev_dec_pending(rdev
, mddev
);
4662 end_reshape_request(r10_bio
);
4665 static void end_reshape_request(struct r10bio
*r10_bio
)
4667 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4669 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4670 bio_put(r10_bio
->master_bio
);
4674 static void raid10_finish_reshape(struct mddev
*mddev
)
4676 struct r10conf
*conf
= mddev
->private;
4678 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4681 if (mddev
->delta_disks
> 0) {
4682 sector_t size
= raid10_size(mddev
, 0, 0);
4683 md_set_array_sectors(mddev
, size
);
4684 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4685 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4686 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4688 mddev
->resync_max_sectors
= size
;
4689 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4690 revalidate_disk(mddev
->gendisk
);
4693 for (d
= conf
->geo
.raid_disks
;
4694 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4696 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4698 clear_bit(In_sync
, &rdev
->flags
);
4699 rdev
= conf
->mirrors
[d
].replacement
;
4701 clear_bit(In_sync
, &rdev
->flags
);
4704 mddev
->layout
= mddev
->new_layout
;
4705 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4706 mddev
->reshape_position
= MaxSector
;
4707 mddev
->delta_disks
= 0;
4708 mddev
->reshape_backwards
= 0;
4711 static struct md_personality raid10_personality
=
4715 .owner
= THIS_MODULE
,
4716 .make_request
= make_request
,
4720 .error_handler
= error
,
4721 .hot_add_disk
= raid10_add_disk
,
4722 .hot_remove_disk
= raid10_remove_disk
,
4723 .spare_active
= raid10_spare_active
,
4724 .sync_request
= sync_request
,
4725 .quiesce
= raid10_quiesce
,
4726 .size
= raid10_size
,
4727 .resize
= raid10_resize
,
4728 .takeover
= raid10_takeover
,
4729 .check_reshape
= raid10_check_reshape
,
4730 .start_reshape
= raid10_start_reshape
,
4731 .finish_reshape
= raid10_finish_reshape
,
4734 static int __init
raid_init(void)
4736 return register_md_personality(&raid10_personality
);
4739 static void raid_exit(void)
4741 unregister_md_personality(&raid10_personality
);
4744 module_init(raid_init
);
4745 module_exit(raid_exit
);
4746 MODULE_LICENSE("GPL");
4747 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4748 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4749 MODULE_ALIAS("md-raid10");
4750 MODULE_ALIAS("md-level-10");
4752 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);