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 generic_make_request(bio
);
1142 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1144 struct r10conf
*conf
= mddev
->private;
1145 struct r10bio
*r10_bio
;
1146 struct bio
*read_bio
;
1148 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1149 int chunk_sects
= chunk_mask
+ 1;
1150 const int rw
= bio_data_dir(bio
);
1151 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1152 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1153 const unsigned long do_discard
= (bio
->bi_rw
1154 & (REQ_DISCARD
| REQ_SECURE
));
1155 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1156 unsigned long flags
;
1157 struct md_rdev
*blocked_rdev
;
1158 struct blk_plug_cb
*cb
;
1159 struct raid10_plug_cb
*plug
= NULL
;
1160 int sectors_handled
;
1164 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1165 md_flush_request(mddev
, bio
);
1169 /* If this request crosses a chunk boundary, we need to
1170 * split it. This will only happen for 1 PAGE (or less) requests.
1172 if (unlikely((bio
->bi_sector
& chunk_mask
) + (bio
->bi_size
>> 9)
1174 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1175 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1176 struct bio_pair
*bp
;
1177 /* Sanity check -- queue functions should prevent this happening */
1178 if ((bio
->bi_vcnt
!= 1 && bio
->bi_vcnt
!= 0) ||
1181 /* This is a one page bio that upper layers
1182 * refuse to split for us, so we need to split it.
1185 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
1187 /* Each of these 'make_request' calls will call 'wait_barrier'.
1188 * If the first succeeds but the second blocks due to the resync
1189 * thread raising the barrier, we will deadlock because the
1190 * IO to the underlying device will be queued in generic_make_request
1191 * and will never complete, so will never reduce nr_pending.
1192 * So increment nr_waiting here so no new raise_barriers will
1193 * succeed, and so the second wait_barrier cannot block.
1195 spin_lock_irq(&conf
->resync_lock
);
1197 spin_unlock_irq(&conf
->resync_lock
);
1199 make_request(mddev
, &bp
->bio1
);
1200 make_request(mddev
, &bp
->bio2
);
1202 spin_lock_irq(&conf
->resync_lock
);
1204 wake_up(&conf
->wait_barrier
);
1205 spin_unlock_irq(&conf
->resync_lock
);
1207 bio_pair_release(bp
);
1210 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1211 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1212 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
1218 md_write_start(mddev
, bio
);
1221 * Register the new request and wait if the reconstruction
1222 * thread has put up a bar for new requests.
1223 * Continue immediately if no resync is active currently.
1227 sectors
= bio
->bi_size
>> 9;
1228 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1229 bio
->bi_sector
< conf
->reshape_progress
&&
1230 bio
->bi_sector
+ sectors
> conf
->reshape_progress
) {
1231 /* IO spans the reshape position. Need to wait for
1234 allow_barrier(conf
);
1235 wait_event(conf
->wait_barrier
,
1236 conf
->reshape_progress
<= bio
->bi_sector
||
1237 conf
->reshape_progress
>= bio
->bi_sector
+ sectors
);
1240 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1241 bio_data_dir(bio
) == WRITE
&&
1242 (mddev
->reshape_backwards
1243 ? (bio
->bi_sector
< conf
->reshape_safe
&&
1244 bio
->bi_sector
+ sectors
> conf
->reshape_progress
)
1245 : (bio
->bi_sector
+ sectors
> conf
->reshape_safe
&&
1246 bio
->bi_sector
< conf
->reshape_progress
))) {
1247 /* Need to update reshape_position in metadata */
1248 mddev
->reshape_position
= conf
->reshape_progress
;
1249 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1250 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1251 md_wakeup_thread(mddev
->thread
);
1252 wait_event(mddev
->sb_wait
,
1253 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1255 conf
->reshape_safe
= mddev
->reshape_position
;
1258 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1260 r10_bio
->master_bio
= bio
;
1261 r10_bio
->sectors
= sectors
;
1263 r10_bio
->mddev
= mddev
;
1264 r10_bio
->sector
= bio
->bi_sector
;
1267 /* We might need to issue multiple reads to different
1268 * devices if there are bad blocks around, so we keep
1269 * track of the number of reads in bio->bi_phys_segments.
1270 * If this is 0, there is only one r10_bio and no locking
1271 * will be needed when the request completes. If it is
1272 * non-zero, then it is the number of not-completed requests.
1274 bio
->bi_phys_segments
= 0;
1275 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1279 * read balancing logic:
1281 struct md_rdev
*rdev
;
1285 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1287 raid_end_bio_io(r10_bio
);
1290 slot
= r10_bio
->read_slot
;
1292 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1293 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1296 r10_bio
->devs
[slot
].bio
= read_bio
;
1297 r10_bio
->devs
[slot
].rdev
= rdev
;
1299 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1300 choose_data_offset(r10_bio
, rdev
);
1301 read_bio
->bi_bdev
= rdev
->bdev
;
1302 read_bio
->bi_end_io
= raid10_end_read_request
;
1303 read_bio
->bi_rw
= READ
| do_sync
;
1304 read_bio
->bi_private
= r10_bio
;
1306 if (max_sectors
< r10_bio
->sectors
) {
1307 /* Could not read all from this device, so we will
1308 * need another r10_bio.
1310 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1312 r10_bio
->sectors
= max_sectors
;
1313 spin_lock_irq(&conf
->device_lock
);
1314 if (bio
->bi_phys_segments
== 0)
1315 bio
->bi_phys_segments
= 2;
1317 bio
->bi_phys_segments
++;
1318 spin_unlock(&conf
->device_lock
);
1319 /* Cannot call generic_make_request directly
1320 * as that will be queued in __generic_make_request
1321 * and subsequent mempool_alloc might block
1322 * waiting for it. so hand bio over to raid10d.
1324 reschedule_retry(r10_bio
);
1326 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1328 r10_bio
->master_bio
= bio
;
1329 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
1332 r10_bio
->mddev
= mddev
;
1333 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1336 generic_make_request(read_bio
);
1343 if (conf
->pending_count
>= max_queued_requests
) {
1344 md_wakeup_thread(mddev
->thread
);
1345 wait_event(conf
->wait_barrier
,
1346 conf
->pending_count
< max_queued_requests
);
1348 /* first select target devices under rcu_lock and
1349 * inc refcount on their rdev. Record them by setting
1351 * If there are known/acknowledged bad blocks on any device
1352 * on which we have seen a write error, we want to avoid
1353 * writing to those blocks. This potentially requires several
1354 * writes to write around the bad blocks. Each set of writes
1355 * gets its own r10_bio with a set of bios attached. The number
1356 * of r10_bios is recored in bio->bi_phys_segments just as with
1360 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1361 raid10_find_phys(conf
, r10_bio
);
1363 blocked_rdev
= NULL
;
1365 max_sectors
= r10_bio
->sectors
;
1367 for (i
= 0; i
< conf
->copies
; i
++) {
1368 int d
= r10_bio
->devs
[i
].devnum
;
1369 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1370 struct md_rdev
*rrdev
= rcu_dereference(
1371 conf
->mirrors
[d
].replacement
);
1374 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1375 atomic_inc(&rdev
->nr_pending
);
1376 blocked_rdev
= rdev
;
1379 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1380 atomic_inc(&rrdev
->nr_pending
);
1381 blocked_rdev
= rrdev
;
1384 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)
1385 || test_bit(Unmerged
, &rdev
->flags
)))
1387 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1388 || test_bit(Unmerged
, &rrdev
->flags
)))
1391 r10_bio
->devs
[i
].bio
= NULL
;
1392 r10_bio
->devs
[i
].repl_bio
= NULL
;
1394 if (!rdev
&& !rrdev
) {
1395 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1398 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1400 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1404 is_bad
= is_badblock(rdev
, dev_sector
,
1406 &first_bad
, &bad_sectors
);
1408 /* Mustn't write here until the bad block
1411 atomic_inc(&rdev
->nr_pending
);
1412 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1413 blocked_rdev
= rdev
;
1416 if (is_bad
&& first_bad
<= dev_sector
) {
1417 /* Cannot write here at all */
1418 bad_sectors
-= (dev_sector
- first_bad
);
1419 if (bad_sectors
< max_sectors
)
1420 /* Mustn't write more than bad_sectors
1421 * to other devices yet
1423 max_sectors
= bad_sectors
;
1424 /* We don't set R10BIO_Degraded as that
1425 * only applies if the disk is missing,
1426 * so it might be re-added, and we want to
1427 * know to recover this chunk.
1428 * In this case the device is here, and the
1429 * fact that this chunk is not in-sync is
1430 * recorded in the bad block log.
1435 int good_sectors
= first_bad
- dev_sector
;
1436 if (good_sectors
< max_sectors
)
1437 max_sectors
= good_sectors
;
1441 r10_bio
->devs
[i
].bio
= bio
;
1442 atomic_inc(&rdev
->nr_pending
);
1445 r10_bio
->devs
[i
].repl_bio
= bio
;
1446 atomic_inc(&rrdev
->nr_pending
);
1451 if (unlikely(blocked_rdev
)) {
1452 /* Have to wait for this device to get unblocked, then retry */
1456 for (j
= 0; j
< i
; j
++) {
1457 if (r10_bio
->devs
[j
].bio
) {
1458 d
= r10_bio
->devs
[j
].devnum
;
1459 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1461 if (r10_bio
->devs
[j
].repl_bio
) {
1462 struct md_rdev
*rdev
;
1463 d
= r10_bio
->devs
[j
].devnum
;
1464 rdev
= conf
->mirrors
[d
].replacement
;
1466 /* Race with remove_disk */
1468 rdev
= conf
->mirrors
[d
].rdev
;
1470 rdev_dec_pending(rdev
, mddev
);
1473 allow_barrier(conf
);
1474 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1479 if (max_sectors
< r10_bio
->sectors
) {
1480 /* We are splitting this into multiple parts, so
1481 * we need to prepare for allocating another r10_bio.
1483 r10_bio
->sectors
= max_sectors
;
1484 spin_lock_irq(&conf
->device_lock
);
1485 if (bio
->bi_phys_segments
== 0)
1486 bio
->bi_phys_segments
= 2;
1488 bio
->bi_phys_segments
++;
1489 spin_unlock_irq(&conf
->device_lock
);
1491 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1493 atomic_set(&r10_bio
->remaining
, 1);
1494 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1496 for (i
= 0; i
< conf
->copies
; i
++) {
1498 int d
= r10_bio
->devs
[i
].devnum
;
1499 if (r10_bio
->devs
[i
].bio
) {
1500 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1501 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1502 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1504 r10_bio
->devs
[i
].bio
= mbio
;
1506 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1507 choose_data_offset(r10_bio
,
1509 mbio
->bi_bdev
= rdev
->bdev
;
1510 mbio
->bi_end_io
= raid10_end_write_request
;
1512 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1513 mbio
->bi_private
= r10_bio
;
1515 atomic_inc(&r10_bio
->remaining
);
1517 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1520 plug
= container_of(cb
, struct raid10_plug_cb
,
1524 spin_lock_irqsave(&conf
->device_lock
, flags
);
1526 bio_list_add(&plug
->pending
, mbio
);
1527 plug
->pending_cnt
++;
1529 bio_list_add(&conf
->pending_bio_list
, mbio
);
1530 conf
->pending_count
++;
1532 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1534 md_wakeup_thread(mddev
->thread
);
1537 if (r10_bio
->devs
[i
].repl_bio
) {
1538 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1540 /* Replacement just got moved to main 'rdev' */
1542 rdev
= conf
->mirrors
[d
].rdev
;
1544 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1545 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1547 r10_bio
->devs
[i
].repl_bio
= mbio
;
1549 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1552 mbio
->bi_bdev
= rdev
->bdev
;
1553 mbio
->bi_end_io
= raid10_end_write_request
;
1555 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1556 mbio
->bi_private
= r10_bio
;
1558 atomic_inc(&r10_bio
->remaining
);
1559 spin_lock_irqsave(&conf
->device_lock
, flags
);
1560 bio_list_add(&conf
->pending_bio_list
, mbio
);
1561 conf
->pending_count
++;
1562 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1563 if (!mddev_check_plugged(mddev
))
1564 md_wakeup_thread(mddev
->thread
);
1568 /* Don't remove the bias on 'remaining' (one_write_done) until
1569 * after checking if we need to go around again.
1572 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1573 one_write_done(r10_bio
);
1574 /* We need another r10_bio. It has already been counted
1575 * in bio->bi_phys_segments.
1577 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1579 r10_bio
->master_bio
= bio
;
1580 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1582 r10_bio
->mddev
= mddev
;
1583 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1587 one_write_done(r10_bio
);
1589 /* In case raid10d snuck in to freeze_array */
1590 wake_up(&conf
->wait_barrier
);
1593 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1595 struct r10conf
*conf
= mddev
->private;
1598 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1599 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1600 if (conf
->geo
.near_copies
> 1)
1601 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1602 if (conf
->geo
.far_copies
> 1) {
1603 if (conf
->geo
.far_offset
)
1604 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1606 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1608 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1609 conf
->geo
.raid_disks
- mddev
->degraded
);
1610 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1611 seq_printf(seq
, "%s",
1612 conf
->mirrors
[i
].rdev
&&
1613 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1614 seq_printf(seq
, "]");
1617 /* check if there are enough drives for
1618 * every block to appear on atleast one.
1619 * Don't consider the device numbered 'ignore'
1620 * as we might be about to remove it.
1622 static int _enough(struct r10conf
*conf
, struct geom
*geo
, int ignore
)
1627 int n
= conf
->copies
;
1631 if (conf
->mirrors
[this].rdev
&&
1634 this = (this+1) % geo
->raid_disks
;
1638 first
= (first
+ geo
->near_copies
) % geo
->raid_disks
;
1639 } while (first
!= 0);
1643 static int enough(struct r10conf
*conf
, int ignore
)
1645 return _enough(conf
, &conf
->geo
, ignore
) &&
1646 _enough(conf
, &conf
->prev
, ignore
);
1649 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1651 char b
[BDEVNAME_SIZE
];
1652 struct r10conf
*conf
= mddev
->private;
1655 * If it is not operational, then we have already marked it as dead
1656 * else if it is the last working disks, ignore the error, let the
1657 * next level up know.
1658 * else mark the drive as failed
1660 if (test_bit(In_sync
, &rdev
->flags
)
1661 && !enough(conf
, rdev
->raid_disk
))
1663 * Don't fail the drive, just return an IO error.
1666 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1667 unsigned long flags
;
1668 spin_lock_irqsave(&conf
->device_lock
, flags
);
1670 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1672 * if recovery is running, make sure it aborts.
1674 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1676 set_bit(Blocked
, &rdev
->flags
);
1677 set_bit(Faulty
, &rdev
->flags
);
1678 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1680 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1681 "md/raid10:%s: Operation continuing on %d devices.\n",
1682 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1683 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1686 static void print_conf(struct r10conf
*conf
)
1689 struct raid10_info
*tmp
;
1691 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1693 printk(KERN_DEBUG
"(!conf)\n");
1696 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1697 conf
->geo
.raid_disks
);
1699 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1700 char b
[BDEVNAME_SIZE
];
1701 tmp
= conf
->mirrors
+ i
;
1703 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1704 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1705 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1706 bdevname(tmp
->rdev
->bdev
,b
));
1710 static void close_sync(struct r10conf
*conf
)
1713 allow_barrier(conf
);
1715 mempool_destroy(conf
->r10buf_pool
);
1716 conf
->r10buf_pool
= NULL
;
1719 static int raid10_spare_active(struct mddev
*mddev
)
1722 struct r10conf
*conf
= mddev
->private;
1723 struct raid10_info
*tmp
;
1725 unsigned long flags
;
1728 * Find all non-in_sync disks within the RAID10 configuration
1729 * and mark them in_sync
1731 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1732 tmp
= conf
->mirrors
+ i
;
1733 if (tmp
->replacement
1734 && tmp
->replacement
->recovery_offset
== MaxSector
1735 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1736 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1737 /* Replacement has just become active */
1739 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1742 /* Replaced device not technically faulty,
1743 * but we need to be sure it gets removed
1744 * and never re-added.
1746 set_bit(Faulty
, &tmp
->rdev
->flags
);
1747 sysfs_notify_dirent_safe(
1748 tmp
->rdev
->sysfs_state
);
1750 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1751 } else if (tmp
->rdev
1752 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1753 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1755 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1758 spin_lock_irqsave(&conf
->device_lock
, flags
);
1759 mddev
->degraded
-= count
;
1760 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1767 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1769 struct r10conf
*conf
= mddev
->private;
1773 int last
= conf
->geo
.raid_disks
- 1;
1774 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1776 if (mddev
->recovery_cp
< MaxSector
)
1777 /* only hot-add to in-sync arrays, as recovery is
1778 * very different from resync
1781 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, &conf
->prev
, -1))
1784 if (rdev
->raid_disk
>= 0)
1785 first
= last
= rdev
->raid_disk
;
1787 if (q
->merge_bvec_fn
) {
1788 set_bit(Unmerged
, &rdev
->flags
);
1789 mddev
->merge_check_needed
= 1;
1792 if (rdev
->saved_raid_disk
>= first
&&
1793 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1794 mirror
= rdev
->saved_raid_disk
;
1797 for ( ; mirror
<= last
; mirror
++) {
1798 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1799 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1802 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1803 p
->replacement
!= NULL
)
1805 clear_bit(In_sync
, &rdev
->flags
);
1806 set_bit(Replacement
, &rdev
->flags
);
1807 rdev
->raid_disk
= mirror
;
1809 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1810 rdev
->data_offset
<< 9);
1812 rcu_assign_pointer(p
->replacement
, rdev
);
1816 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1817 rdev
->data_offset
<< 9);
1819 p
->head_position
= 0;
1820 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1821 rdev
->raid_disk
= mirror
;
1823 if (rdev
->saved_raid_disk
!= mirror
)
1825 rcu_assign_pointer(p
->rdev
, rdev
);
1828 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1829 /* Some requests might not have seen this new
1830 * merge_bvec_fn. We must wait for them to complete
1831 * before merging the device fully.
1832 * First we make sure any code which has tested
1833 * our function has submitted the request, then
1834 * we wait for all outstanding requests to complete.
1836 synchronize_sched();
1837 raise_barrier(conf
, 0);
1838 lower_barrier(conf
);
1839 clear_bit(Unmerged
, &rdev
->flags
);
1841 md_integrity_add_rdev(rdev
, mddev
);
1842 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1843 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1849 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1851 struct r10conf
*conf
= mddev
->private;
1853 int number
= rdev
->raid_disk
;
1854 struct md_rdev
**rdevp
;
1855 struct raid10_info
*p
= conf
->mirrors
+ number
;
1858 if (rdev
== p
->rdev
)
1860 else if (rdev
== p
->replacement
)
1861 rdevp
= &p
->replacement
;
1865 if (test_bit(In_sync
, &rdev
->flags
) ||
1866 atomic_read(&rdev
->nr_pending
)) {
1870 /* Only remove faulty devices if recovery
1873 if (!test_bit(Faulty
, &rdev
->flags
) &&
1874 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1875 (!p
->replacement
|| p
->replacement
== rdev
) &&
1876 number
< conf
->geo
.raid_disks
&&
1883 if (atomic_read(&rdev
->nr_pending
)) {
1884 /* lost the race, try later */
1888 } else if (p
->replacement
) {
1889 /* We must have just cleared 'rdev' */
1890 p
->rdev
= p
->replacement
;
1891 clear_bit(Replacement
, &p
->replacement
->flags
);
1892 smp_mb(); /* Make sure other CPUs may see both as identical
1893 * but will never see neither -- if they are careful.
1895 p
->replacement
= NULL
;
1896 clear_bit(WantReplacement
, &rdev
->flags
);
1898 /* We might have just remove the Replacement as faulty
1899 * Clear the flag just in case
1901 clear_bit(WantReplacement
, &rdev
->flags
);
1903 err
= md_integrity_register(mddev
);
1912 static void end_sync_read(struct bio
*bio
, int error
)
1914 struct r10bio
*r10_bio
= bio
->bi_private
;
1915 struct r10conf
*conf
= r10_bio
->mddev
->private;
1918 if (bio
== r10_bio
->master_bio
) {
1919 /* this is a reshape read */
1920 d
= r10_bio
->read_slot
; /* really the read dev */
1922 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1924 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1925 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1927 /* The write handler will notice the lack of
1928 * R10BIO_Uptodate and record any errors etc
1930 atomic_add(r10_bio
->sectors
,
1931 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1933 /* for reconstruct, we always reschedule after a read.
1934 * for resync, only after all reads
1936 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1937 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1938 atomic_dec_and_test(&r10_bio
->remaining
)) {
1939 /* we have read all the blocks,
1940 * do the comparison in process context in raid10d
1942 reschedule_retry(r10_bio
);
1946 static void end_sync_request(struct r10bio
*r10_bio
)
1948 struct mddev
*mddev
= r10_bio
->mddev
;
1950 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1951 if (r10_bio
->master_bio
== NULL
) {
1952 /* the primary of several recovery bios */
1953 sector_t s
= r10_bio
->sectors
;
1954 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1955 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1956 reschedule_retry(r10_bio
);
1959 md_done_sync(mddev
, s
, 1);
1962 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1963 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1964 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1965 reschedule_retry(r10_bio
);
1973 static void end_sync_write(struct bio
*bio
, int error
)
1975 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1976 struct r10bio
*r10_bio
= bio
->bi_private
;
1977 struct mddev
*mddev
= r10_bio
->mddev
;
1978 struct r10conf
*conf
= mddev
->private;
1984 struct md_rdev
*rdev
= NULL
;
1986 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1988 rdev
= conf
->mirrors
[d
].replacement
;
1990 rdev
= conf
->mirrors
[d
].rdev
;
1994 md_error(mddev
, rdev
);
1996 set_bit(WriteErrorSeen
, &rdev
->flags
);
1997 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1998 set_bit(MD_RECOVERY_NEEDED
,
1999 &rdev
->mddev
->recovery
);
2000 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
2002 } else if (is_badblock(rdev
,
2003 r10_bio
->devs
[slot
].addr
,
2005 &first_bad
, &bad_sectors
))
2006 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
2008 rdev_dec_pending(rdev
, mddev
);
2010 end_sync_request(r10_bio
);
2014 * Note: sync and recover and handled very differently for raid10
2015 * This code is for resync.
2016 * For resync, we read through virtual addresses and read all blocks.
2017 * If there is any error, we schedule a write. The lowest numbered
2018 * drive is authoritative.
2019 * However requests come for physical address, so we need to map.
2020 * For every physical address there are raid_disks/copies virtual addresses,
2021 * which is always are least one, but is not necessarly an integer.
2022 * This means that a physical address can span multiple chunks, so we may
2023 * have to submit multiple io requests for a single sync request.
2026 * We check if all blocks are in-sync and only write to blocks that
2029 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2031 struct r10conf
*conf
= mddev
->private;
2033 struct bio
*tbio
, *fbio
;
2036 atomic_set(&r10_bio
->remaining
, 1);
2038 /* find the first device with a block */
2039 for (i
=0; i
<conf
->copies
; i
++)
2040 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
2043 if (i
== conf
->copies
)
2047 fbio
= r10_bio
->devs
[i
].bio
;
2049 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2050 /* now find blocks with errors */
2051 for (i
=0 ; i
< conf
->copies
; i
++) {
2054 tbio
= r10_bio
->devs
[i
].bio
;
2056 if (tbio
->bi_end_io
!= end_sync_read
)
2060 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2061 /* We know that the bi_io_vec layout is the same for
2062 * both 'first' and 'i', so we just compare them.
2063 * All vec entries are PAGE_SIZE;
2065 for (j
= 0; j
< vcnt
; j
++)
2066 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2067 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2068 fbio
->bi_io_vec
[j
].bv_len
))
2072 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2073 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2074 /* Don't fix anything. */
2077 /* Ok, we need to write this bio, either to correct an
2078 * inconsistency or to correct an unreadable block.
2079 * First we need to fixup bv_offset, bv_len and
2080 * bi_vecs, as the read request might have corrupted these
2082 tbio
->bi_vcnt
= vcnt
;
2083 tbio
->bi_size
= r10_bio
->sectors
<< 9;
2085 tbio
->bi_phys_segments
= 0;
2086 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2087 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
2088 tbio
->bi_next
= NULL
;
2089 tbio
->bi_rw
= WRITE
;
2090 tbio
->bi_private
= r10_bio
;
2091 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
2093 for (j
=0; j
< vcnt
; j
++) {
2094 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2095 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2097 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2098 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2101 tbio
->bi_end_io
= end_sync_write
;
2103 d
= r10_bio
->devs
[i
].devnum
;
2104 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2105 atomic_inc(&r10_bio
->remaining
);
2106 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
2108 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2109 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2110 generic_make_request(tbio
);
2113 /* Now write out to any replacement devices
2116 for (i
= 0; i
< conf
->copies
; i
++) {
2119 tbio
= r10_bio
->devs
[i
].repl_bio
;
2120 if (!tbio
|| !tbio
->bi_end_io
)
2122 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2123 && r10_bio
->devs
[i
].bio
!= fbio
)
2124 for (j
= 0; j
< vcnt
; j
++)
2125 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2126 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2128 d
= r10_bio
->devs
[i
].devnum
;
2129 atomic_inc(&r10_bio
->remaining
);
2130 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2131 tbio
->bi_size
>> 9);
2132 generic_make_request(tbio
);
2136 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2137 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2143 * Now for the recovery code.
2144 * Recovery happens across physical sectors.
2145 * We recover all non-is_sync drives by finding the virtual address of
2146 * each, and then choose a working drive that also has that virt address.
2147 * There is a separate r10_bio for each non-in_sync drive.
2148 * Only the first two slots are in use. The first for reading,
2149 * The second for writing.
2152 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2154 /* We got a read error during recovery.
2155 * We repeat the read in smaller page-sized sections.
2156 * If a read succeeds, write it to the new device or record
2157 * a bad block if we cannot.
2158 * If a read fails, record a bad block on both old and
2161 struct mddev
*mddev
= r10_bio
->mddev
;
2162 struct r10conf
*conf
= mddev
->private;
2163 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2165 int sectors
= r10_bio
->sectors
;
2167 int dr
= r10_bio
->devs
[0].devnum
;
2168 int dw
= r10_bio
->devs
[1].devnum
;
2172 struct md_rdev
*rdev
;
2176 if (s
> (PAGE_SIZE
>>9))
2179 rdev
= conf
->mirrors
[dr
].rdev
;
2180 addr
= r10_bio
->devs
[0].addr
+ sect
,
2181 ok
= sync_page_io(rdev
,
2184 bio
->bi_io_vec
[idx
].bv_page
,
2187 rdev
= conf
->mirrors
[dw
].rdev
;
2188 addr
= r10_bio
->devs
[1].addr
+ sect
;
2189 ok
= sync_page_io(rdev
,
2192 bio
->bi_io_vec
[idx
].bv_page
,
2195 set_bit(WriteErrorSeen
, &rdev
->flags
);
2196 if (!test_and_set_bit(WantReplacement
,
2198 set_bit(MD_RECOVERY_NEEDED
,
2199 &rdev
->mddev
->recovery
);
2203 /* We don't worry if we cannot set a bad block -
2204 * it really is bad so there is no loss in not
2207 rdev_set_badblocks(rdev
, addr
, s
, 0);
2209 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2210 /* need bad block on destination too */
2211 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2212 addr
= r10_bio
->devs
[1].addr
+ sect
;
2213 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2215 /* just abort the recovery */
2217 "md/raid10:%s: recovery aborted"
2218 " due to read error\n",
2221 conf
->mirrors
[dw
].recovery_disabled
2222 = mddev
->recovery_disabled
;
2223 set_bit(MD_RECOVERY_INTR
,
2236 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2238 struct r10conf
*conf
= mddev
->private;
2240 struct bio
*wbio
, *wbio2
;
2242 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2243 fix_recovery_read_error(r10_bio
);
2244 end_sync_request(r10_bio
);
2249 * share the pages with the first bio
2250 * and submit the write request
2252 d
= r10_bio
->devs
[1].devnum
;
2253 wbio
= r10_bio
->devs
[1].bio
;
2254 wbio2
= r10_bio
->devs
[1].repl_bio
;
2255 if (wbio
->bi_end_io
) {
2256 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2257 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
2258 generic_make_request(wbio
);
2260 if (wbio2
&& wbio2
->bi_end_io
) {
2261 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2262 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2263 wbio2
->bi_size
>> 9);
2264 generic_make_request(wbio2
);
2270 * Used by fix_read_error() to decay the per rdev read_errors.
2271 * We halve the read error count for every hour that has elapsed
2272 * since the last recorded read error.
2275 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2277 struct timespec cur_time_mon
;
2278 unsigned long hours_since_last
;
2279 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2281 ktime_get_ts(&cur_time_mon
);
2283 if (rdev
->last_read_error
.tv_sec
== 0 &&
2284 rdev
->last_read_error
.tv_nsec
== 0) {
2285 /* first time we've seen a read error */
2286 rdev
->last_read_error
= cur_time_mon
;
2290 hours_since_last
= (cur_time_mon
.tv_sec
-
2291 rdev
->last_read_error
.tv_sec
) / 3600;
2293 rdev
->last_read_error
= cur_time_mon
;
2296 * if hours_since_last is > the number of bits in read_errors
2297 * just set read errors to 0. We do this to avoid
2298 * overflowing the shift of read_errors by hours_since_last.
2300 if (hours_since_last
>= 8 * sizeof(read_errors
))
2301 atomic_set(&rdev
->read_errors
, 0);
2303 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2306 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2307 int sectors
, struct page
*page
, int rw
)
2312 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2313 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2315 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2319 set_bit(WriteErrorSeen
, &rdev
->flags
);
2320 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2321 set_bit(MD_RECOVERY_NEEDED
,
2322 &rdev
->mddev
->recovery
);
2324 /* need to record an error - either for the block or the device */
2325 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2326 md_error(rdev
->mddev
, rdev
);
2331 * This is a kernel thread which:
2333 * 1. Retries failed read operations on working mirrors.
2334 * 2. Updates the raid superblock when problems encounter.
2335 * 3. Performs writes following reads for array synchronising.
2338 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2340 int sect
= 0; /* Offset from r10_bio->sector */
2341 int sectors
= r10_bio
->sectors
;
2342 struct md_rdev
*rdev
;
2343 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2344 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2346 /* still own a reference to this rdev, so it cannot
2347 * have been cleared recently.
2349 rdev
= conf
->mirrors
[d
].rdev
;
2351 if (test_bit(Faulty
, &rdev
->flags
))
2352 /* drive has already been failed, just ignore any
2353 more fix_read_error() attempts */
2356 check_decay_read_errors(mddev
, rdev
);
2357 atomic_inc(&rdev
->read_errors
);
2358 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2359 char b
[BDEVNAME_SIZE
];
2360 bdevname(rdev
->bdev
, b
);
2363 "md/raid10:%s: %s: Raid device exceeded "
2364 "read_error threshold [cur %d:max %d]\n",
2366 atomic_read(&rdev
->read_errors
), max_read_errors
);
2368 "md/raid10:%s: %s: Failing raid device\n",
2370 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2371 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2377 int sl
= r10_bio
->read_slot
;
2381 if (s
> (PAGE_SIZE
>>9))
2389 d
= r10_bio
->devs
[sl
].devnum
;
2390 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2392 !test_bit(Unmerged
, &rdev
->flags
) &&
2393 test_bit(In_sync
, &rdev
->flags
) &&
2394 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2395 &first_bad
, &bad_sectors
) == 0) {
2396 atomic_inc(&rdev
->nr_pending
);
2398 success
= sync_page_io(rdev
,
2399 r10_bio
->devs
[sl
].addr
+
2402 conf
->tmppage
, READ
, false);
2403 rdev_dec_pending(rdev
, mddev
);
2409 if (sl
== conf
->copies
)
2411 } while (!success
&& sl
!= r10_bio
->read_slot
);
2415 /* Cannot read from anywhere, just mark the block
2416 * as bad on the first device to discourage future
2419 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2420 rdev
= conf
->mirrors
[dn
].rdev
;
2422 if (!rdev_set_badblocks(
2424 r10_bio
->devs
[r10_bio
->read_slot
].addr
2427 md_error(mddev
, rdev
);
2428 r10_bio
->devs
[r10_bio
->read_slot
].bio
2435 /* write it back and re-read */
2437 while (sl
!= r10_bio
->read_slot
) {
2438 char b
[BDEVNAME_SIZE
];
2443 d
= r10_bio
->devs
[sl
].devnum
;
2444 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2446 test_bit(Unmerged
, &rdev
->flags
) ||
2447 !test_bit(In_sync
, &rdev
->flags
))
2450 atomic_inc(&rdev
->nr_pending
);
2452 if (r10_sync_page_io(rdev
,
2453 r10_bio
->devs
[sl
].addr
+
2455 s
, conf
->tmppage
, WRITE
)
2457 /* Well, this device is dead */
2459 "md/raid10:%s: read correction "
2461 " (%d sectors at %llu on %s)\n",
2463 (unsigned long long)(
2465 choose_data_offset(r10_bio
,
2467 bdevname(rdev
->bdev
, b
));
2468 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2471 bdevname(rdev
->bdev
, b
));
2473 rdev_dec_pending(rdev
, mddev
);
2477 while (sl
!= r10_bio
->read_slot
) {
2478 char b
[BDEVNAME_SIZE
];
2483 d
= r10_bio
->devs
[sl
].devnum
;
2484 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2486 !test_bit(In_sync
, &rdev
->flags
))
2489 atomic_inc(&rdev
->nr_pending
);
2491 switch (r10_sync_page_io(rdev
,
2492 r10_bio
->devs
[sl
].addr
+
2497 /* Well, this device is dead */
2499 "md/raid10:%s: unable to read back "
2501 " (%d sectors at %llu on %s)\n",
2503 (unsigned long long)(
2505 choose_data_offset(r10_bio
, rdev
)),
2506 bdevname(rdev
->bdev
, b
));
2507 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2510 bdevname(rdev
->bdev
, b
));
2514 "md/raid10:%s: read error corrected"
2515 " (%d sectors at %llu on %s)\n",
2517 (unsigned long long)(
2519 choose_data_offset(r10_bio
, rdev
)),
2520 bdevname(rdev
->bdev
, b
));
2521 atomic_add(s
, &rdev
->corrected_errors
);
2524 rdev_dec_pending(rdev
, mddev
);
2534 static void bi_complete(struct bio
*bio
, int error
)
2536 complete((struct completion
*)bio
->bi_private
);
2539 static int submit_bio_wait(int rw
, struct bio
*bio
)
2541 struct completion event
;
2544 init_completion(&event
);
2545 bio
->bi_private
= &event
;
2546 bio
->bi_end_io
= bi_complete
;
2547 submit_bio(rw
, bio
);
2548 wait_for_completion(&event
);
2550 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2553 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2555 struct bio
*bio
= r10_bio
->master_bio
;
2556 struct mddev
*mddev
= r10_bio
->mddev
;
2557 struct r10conf
*conf
= mddev
->private;
2558 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2559 /* bio has the data to be written to slot 'i' where
2560 * we just recently had a write error.
2561 * We repeatedly clone the bio and trim down to one block,
2562 * then try the write. Where the write fails we record
2564 * It is conceivable that the bio doesn't exactly align with
2565 * blocks. We must handle this.
2567 * We currently own a reference to the rdev.
2573 int sect_to_write
= r10_bio
->sectors
;
2576 if (rdev
->badblocks
.shift
< 0)
2579 block_sectors
= 1 << rdev
->badblocks
.shift
;
2580 sector
= r10_bio
->sector
;
2581 sectors
= ((r10_bio
->sector
+ block_sectors
)
2582 & ~(sector_t
)(block_sectors
- 1))
2585 while (sect_to_write
) {
2587 if (sectors
> sect_to_write
)
2588 sectors
= sect_to_write
;
2589 /* Write at 'sector' for 'sectors' */
2590 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2591 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2592 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2593 choose_data_offset(r10_bio
, rdev
) +
2594 (sector
- r10_bio
->sector
));
2595 wbio
->bi_bdev
= rdev
->bdev
;
2596 if (submit_bio_wait(WRITE
, wbio
) == 0)
2598 ok
= rdev_set_badblocks(rdev
, sector
,
2603 sect_to_write
-= sectors
;
2605 sectors
= block_sectors
;
2610 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2612 int slot
= r10_bio
->read_slot
;
2614 struct r10conf
*conf
= mddev
->private;
2615 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2616 char b
[BDEVNAME_SIZE
];
2617 unsigned long do_sync
;
2620 /* we got a read error. Maybe the drive is bad. Maybe just
2621 * the block and we can fix it.
2622 * We freeze all other IO, and try reading the block from
2623 * other devices. When we find one, we re-write
2624 * and check it that fixes the read error.
2625 * This is all done synchronously while the array is
2628 bio
= r10_bio
->devs
[slot
].bio
;
2629 bdevname(bio
->bi_bdev
, b
);
2631 r10_bio
->devs
[slot
].bio
= NULL
;
2633 if (mddev
->ro
== 0) {
2635 fix_read_error(conf
, mddev
, r10_bio
);
2636 unfreeze_array(conf
);
2638 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2640 rdev_dec_pending(rdev
, mddev
);
2643 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2645 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2646 " read error for block %llu\n",
2648 (unsigned long long)r10_bio
->sector
);
2649 raid_end_bio_io(r10_bio
);
2653 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2654 slot
= r10_bio
->read_slot
;
2657 "md/raid10:%s: %s: redirecting "
2658 "sector %llu to another mirror\n",
2660 bdevname(rdev
->bdev
, b
),
2661 (unsigned long long)r10_bio
->sector
);
2662 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2665 r10_bio
->sector
- bio
->bi_sector
,
2667 r10_bio
->devs
[slot
].bio
= bio
;
2668 r10_bio
->devs
[slot
].rdev
= rdev
;
2669 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2670 + choose_data_offset(r10_bio
, rdev
);
2671 bio
->bi_bdev
= rdev
->bdev
;
2672 bio
->bi_rw
= READ
| do_sync
;
2673 bio
->bi_private
= r10_bio
;
2674 bio
->bi_end_io
= raid10_end_read_request
;
2675 if (max_sectors
< r10_bio
->sectors
) {
2676 /* Drat - have to split this up more */
2677 struct bio
*mbio
= r10_bio
->master_bio
;
2678 int sectors_handled
=
2679 r10_bio
->sector
+ max_sectors
2681 r10_bio
->sectors
= max_sectors
;
2682 spin_lock_irq(&conf
->device_lock
);
2683 if (mbio
->bi_phys_segments
== 0)
2684 mbio
->bi_phys_segments
= 2;
2686 mbio
->bi_phys_segments
++;
2687 spin_unlock_irq(&conf
->device_lock
);
2688 generic_make_request(bio
);
2690 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2692 r10_bio
->master_bio
= mbio
;
2693 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2696 set_bit(R10BIO_ReadError
,
2698 r10_bio
->mddev
= mddev
;
2699 r10_bio
->sector
= mbio
->bi_sector
2704 generic_make_request(bio
);
2707 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2709 /* Some sort of write request has finished and it
2710 * succeeded in writing where we thought there was a
2711 * bad block. So forget the bad block.
2712 * Or possibly if failed and we need to record
2716 struct md_rdev
*rdev
;
2718 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2719 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2720 for (m
= 0; m
< conf
->copies
; m
++) {
2721 int dev
= r10_bio
->devs
[m
].devnum
;
2722 rdev
= conf
->mirrors
[dev
].rdev
;
2723 if (r10_bio
->devs
[m
].bio
== NULL
)
2725 if (test_bit(BIO_UPTODATE
,
2726 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2727 rdev_clear_badblocks(
2729 r10_bio
->devs
[m
].addr
,
2730 r10_bio
->sectors
, 0);
2732 if (!rdev_set_badblocks(
2734 r10_bio
->devs
[m
].addr
,
2735 r10_bio
->sectors
, 0))
2736 md_error(conf
->mddev
, rdev
);
2738 rdev
= conf
->mirrors
[dev
].replacement
;
2739 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2741 if (test_bit(BIO_UPTODATE
,
2742 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2743 rdev_clear_badblocks(
2745 r10_bio
->devs
[m
].addr
,
2746 r10_bio
->sectors
, 0);
2748 if (!rdev_set_badblocks(
2750 r10_bio
->devs
[m
].addr
,
2751 r10_bio
->sectors
, 0))
2752 md_error(conf
->mddev
, rdev
);
2757 for (m
= 0; m
< conf
->copies
; m
++) {
2758 int dev
= r10_bio
->devs
[m
].devnum
;
2759 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2760 rdev
= conf
->mirrors
[dev
].rdev
;
2761 if (bio
== IO_MADE_GOOD
) {
2762 rdev_clear_badblocks(
2764 r10_bio
->devs
[m
].addr
,
2765 r10_bio
->sectors
, 0);
2766 rdev_dec_pending(rdev
, conf
->mddev
);
2767 } else if (bio
!= NULL
&&
2768 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2769 if (!narrow_write_error(r10_bio
, m
)) {
2770 md_error(conf
->mddev
, rdev
);
2771 set_bit(R10BIO_Degraded
,
2774 rdev_dec_pending(rdev
, conf
->mddev
);
2776 bio
= r10_bio
->devs
[m
].repl_bio
;
2777 rdev
= conf
->mirrors
[dev
].replacement
;
2778 if (rdev
&& bio
== IO_MADE_GOOD
) {
2779 rdev_clear_badblocks(
2781 r10_bio
->devs
[m
].addr
,
2782 r10_bio
->sectors
, 0);
2783 rdev_dec_pending(rdev
, conf
->mddev
);
2786 if (test_bit(R10BIO_WriteError
,
2788 close_write(r10_bio
);
2789 raid_end_bio_io(r10_bio
);
2793 static void raid10d(struct md_thread
*thread
)
2795 struct mddev
*mddev
= thread
->mddev
;
2796 struct r10bio
*r10_bio
;
2797 unsigned long flags
;
2798 struct r10conf
*conf
= mddev
->private;
2799 struct list_head
*head
= &conf
->retry_list
;
2800 struct blk_plug plug
;
2802 md_check_recovery(mddev
);
2804 blk_start_plug(&plug
);
2807 flush_pending_writes(conf
);
2809 spin_lock_irqsave(&conf
->device_lock
, flags
);
2810 if (list_empty(head
)) {
2811 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2814 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2815 list_del(head
->prev
);
2817 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2819 mddev
= r10_bio
->mddev
;
2820 conf
= mddev
->private;
2821 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2822 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2823 handle_write_completed(conf
, r10_bio
);
2824 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2825 reshape_request_write(mddev
, r10_bio
);
2826 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2827 sync_request_write(mddev
, r10_bio
);
2828 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2829 recovery_request_write(mddev
, r10_bio
);
2830 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2831 handle_read_error(mddev
, r10_bio
);
2833 /* just a partial read to be scheduled from a
2836 int slot
= r10_bio
->read_slot
;
2837 generic_make_request(r10_bio
->devs
[slot
].bio
);
2841 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2842 md_check_recovery(mddev
);
2844 blk_finish_plug(&plug
);
2848 static int init_resync(struct r10conf
*conf
)
2853 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2854 BUG_ON(conf
->r10buf_pool
);
2855 conf
->have_replacement
= 0;
2856 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2857 if (conf
->mirrors
[i
].replacement
)
2858 conf
->have_replacement
= 1;
2859 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2860 if (!conf
->r10buf_pool
)
2862 conf
->next_resync
= 0;
2867 * perform a "sync" on one "block"
2869 * We need to make sure that no normal I/O request - particularly write
2870 * requests - conflict with active sync requests.
2872 * This is achieved by tracking pending requests and a 'barrier' concept
2873 * that can be installed to exclude normal IO requests.
2875 * Resync and recovery are handled very differently.
2876 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2878 * For resync, we iterate over virtual addresses, read all copies,
2879 * and update if there are differences. If only one copy is live,
2881 * For recovery, we iterate over physical addresses, read a good
2882 * value for each non-in_sync drive, and over-write.
2884 * So, for recovery we may have several outstanding complex requests for a
2885 * given address, one for each out-of-sync device. We model this by allocating
2886 * a number of r10_bio structures, one for each out-of-sync device.
2887 * As we setup these structures, we collect all bio's together into a list
2888 * which we then process collectively to add pages, and then process again
2889 * to pass to generic_make_request.
2891 * The r10_bio structures are linked using a borrowed master_bio pointer.
2892 * This link is counted in ->remaining. When the r10_bio that points to NULL
2893 * has its remaining count decremented to 0, the whole complex operation
2898 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2899 int *skipped
, int go_faster
)
2901 struct r10conf
*conf
= mddev
->private;
2902 struct r10bio
*r10_bio
;
2903 struct bio
*biolist
= NULL
, *bio
;
2904 sector_t max_sector
, nr_sectors
;
2907 sector_t sync_blocks
;
2908 sector_t sectors_skipped
= 0;
2909 int chunks_skipped
= 0;
2910 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2912 if (!conf
->r10buf_pool
)
2913 if (init_resync(conf
))
2917 max_sector
= mddev
->dev_sectors
;
2918 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2919 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2920 max_sector
= mddev
->resync_max_sectors
;
2921 if (sector_nr
>= max_sector
) {
2922 /* If we aborted, we need to abort the
2923 * sync on the 'current' bitmap chucks (there can
2924 * be several when recovering multiple devices).
2925 * as we may have started syncing it but not finished.
2926 * We can find the current address in
2927 * mddev->curr_resync, but for recovery,
2928 * we need to convert that to several
2929 * virtual addresses.
2931 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2936 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2937 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2938 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2940 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2942 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2943 bitmap_end_sync(mddev
->bitmap
, sect
,
2947 /* completed sync */
2948 if ((!mddev
->bitmap
|| conf
->fullsync
)
2949 && conf
->have_replacement
2950 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2951 /* Completed a full sync so the replacements
2952 * are now fully recovered.
2954 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2955 if (conf
->mirrors
[i
].replacement
)
2956 conf
->mirrors
[i
].replacement
2962 bitmap_close_sync(mddev
->bitmap
);
2965 return sectors_skipped
;
2968 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2969 return reshape_request(mddev
, sector_nr
, skipped
);
2971 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2972 /* if there has been nothing to do on any drive,
2973 * then there is nothing to do at all..
2976 return (max_sector
- sector_nr
) + sectors_skipped
;
2979 if (max_sector
> mddev
->resync_max
)
2980 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2982 /* make sure whole request will fit in a chunk - if chunks
2985 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2986 max_sector
> (sector_nr
| chunk_mask
))
2987 max_sector
= (sector_nr
| chunk_mask
) + 1;
2989 * If there is non-resync activity waiting for us then
2990 * put in a delay to throttle resync.
2992 if (!go_faster
&& conf
->nr_waiting
)
2993 msleep_interruptible(1000);
2995 /* Again, very different code for resync and recovery.
2996 * Both must result in an r10bio with a list of bios that
2997 * have bi_end_io, bi_sector, bi_bdev set,
2998 * and bi_private set to the r10bio.
2999 * For recovery, we may actually create several r10bios
3000 * with 2 bios in each, that correspond to the bios in the main one.
3001 * In this case, the subordinate r10bios link back through a
3002 * borrowed master_bio pointer, and the counter in the master
3003 * includes a ref from each subordinate.
3005 /* First, we decide what to do and set ->bi_end_io
3006 * To end_sync_read if we want to read, and
3007 * end_sync_write if we will want to write.
3010 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
3011 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3012 /* recovery... the complicated one */
3016 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
3022 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
3024 if ((mirror
->rdev
== NULL
||
3025 test_bit(In_sync
, &mirror
->rdev
->flags
))
3027 (mirror
->replacement
== NULL
||
3029 &mirror
->replacement
->flags
)))
3033 /* want to reconstruct this device */
3035 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3036 if (sect
>= mddev
->resync_max_sectors
) {
3037 /* last stripe is not complete - don't
3038 * try to recover this sector.
3042 /* Unless we are doing a full sync, or a replacement
3043 * we only need to recover the block if it is set in
3046 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3048 if (sync_blocks
< max_sync
)
3049 max_sync
= sync_blocks
;
3051 mirror
->replacement
== NULL
&&
3053 /* yep, skip the sync_blocks here, but don't assume
3054 * that there will never be anything to do here
3056 chunks_skipped
= -1;
3060 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3061 raise_barrier(conf
, rb2
!= NULL
);
3062 atomic_set(&r10_bio
->remaining
, 0);
3064 r10_bio
->master_bio
= (struct bio
*)rb2
;
3066 atomic_inc(&rb2
->remaining
);
3067 r10_bio
->mddev
= mddev
;
3068 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3069 r10_bio
->sector
= sect
;
3071 raid10_find_phys(conf
, r10_bio
);
3073 /* Need to check if the array will still be
3076 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3077 if (conf
->mirrors
[j
].rdev
== NULL
||
3078 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3083 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3084 &sync_blocks
, still_degraded
);
3087 for (j
=0; j
<conf
->copies
;j
++) {
3089 int d
= r10_bio
->devs
[j
].devnum
;
3090 sector_t from_addr
, to_addr
;
3091 struct md_rdev
*rdev
;
3092 sector_t sector
, first_bad
;
3094 if (!conf
->mirrors
[d
].rdev
||
3095 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3097 /* This is where we read from */
3099 rdev
= conf
->mirrors
[d
].rdev
;
3100 sector
= r10_bio
->devs
[j
].addr
;
3102 if (is_badblock(rdev
, sector
, max_sync
,
3103 &first_bad
, &bad_sectors
)) {
3104 if (first_bad
> sector
)
3105 max_sync
= first_bad
- sector
;
3107 bad_sectors
-= (sector
3109 if (max_sync
> bad_sectors
)
3110 max_sync
= bad_sectors
;
3114 bio
= r10_bio
->devs
[0].bio
;
3115 bio
->bi_next
= biolist
;
3117 bio
->bi_private
= r10_bio
;
3118 bio
->bi_end_io
= end_sync_read
;
3120 from_addr
= r10_bio
->devs
[j
].addr
;
3121 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
3122 bio
->bi_bdev
= rdev
->bdev
;
3123 atomic_inc(&rdev
->nr_pending
);
3124 /* and we write to 'i' (if not in_sync) */
3126 for (k
=0; k
<conf
->copies
; k
++)
3127 if (r10_bio
->devs
[k
].devnum
== i
)
3129 BUG_ON(k
== conf
->copies
);
3130 to_addr
= r10_bio
->devs
[k
].addr
;
3131 r10_bio
->devs
[0].devnum
= d
;
3132 r10_bio
->devs
[0].addr
= from_addr
;
3133 r10_bio
->devs
[1].devnum
= i
;
3134 r10_bio
->devs
[1].addr
= to_addr
;
3136 rdev
= mirror
->rdev
;
3137 if (!test_bit(In_sync
, &rdev
->flags
)) {
3138 bio
= r10_bio
->devs
[1].bio
;
3139 bio
->bi_next
= biolist
;
3141 bio
->bi_private
= r10_bio
;
3142 bio
->bi_end_io
= end_sync_write
;
3144 bio
->bi_sector
= to_addr
3145 + rdev
->data_offset
;
3146 bio
->bi_bdev
= rdev
->bdev
;
3147 atomic_inc(&r10_bio
->remaining
);
3149 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3151 /* and maybe write to replacement */
3152 bio
= r10_bio
->devs
[1].repl_bio
;
3154 bio
->bi_end_io
= NULL
;
3155 rdev
= mirror
->replacement
;
3156 /* Note: if rdev != NULL, then bio
3157 * cannot be NULL as r10buf_pool_alloc will
3158 * have allocated it.
3159 * So the second test here is pointless.
3160 * But it keeps semantic-checkers happy, and
3161 * this comment keeps human reviewers
3164 if (rdev
== NULL
|| bio
== NULL
||
3165 test_bit(Faulty
, &rdev
->flags
))
3167 bio
->bi_next
= biolist
;
3169 bio
->bi_private
= r10_bio
;
3170 bio
->bi_end_io
= end_sync_write
;
3172 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3173 bio
->bi_bdev
= rdev
->bdev
;
3174 atomic_inc(&r10_bio
->remaining
);
3177 if (j
== conf
->copies
) {
3178 /* Cannot recover, so abort the recovery or
3179 * record a bad block */
3182 atomic_dec(&rb2
->remaining
);
3185 /* problem is that there are bad blocks
3186 * on other device(s)
3189 for (k
= 0; k
< conf
->copies
; k
++)
3190 if (r10_bio
->devs
[k
].devnum
== i
)
3192 if (!test_bit(In_sync
,
3193 &mirror
->rdev
->flags
)
3194 && !rdev_set_badblocks(
3196 r10_bio
->devs
[k
].addr
,
3199 if (mirror
->replacement
&&
3200 !rdev_set_badblocks(
3201 mirror
->replacement
,
3202 r10_bio
->devs
[k
].addr
,
3207 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3209 printk(KERN_INFO
"md/raid10:%s: insufficient "
3210 "working devices for recovery.\n",
3212 mirror
->recovery_disabled
3213 = mddev
->recovery_disabled
;
3218 if (biolist
== NULL
) {
3220 struct r10bio
*rb2
= r10_bio
;
3221 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3222 rb2
->master_bio
= NULL
;
3228 /* resync. Schedule a read for every block at this virt offset */
3231 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3233 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3234 &sync_blocks
, mddev
->degraded
) &&
3235 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3236 &mddev
->recovery
)) {
3237 /* We can skip this block */
3239 return sync_blocks
+ sectors_skipped
;
3241 if (sync_blocks
< max_sync
)
3242 max_sync
= sync_blocks
;
3243 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3245 r10_bio
->mddev
= mddev
;
3246 atomic_set(&r10_bio
->remaining
, 0);
3247 raise_barrier(conf
, 0);
3248 conf
->next_resync
= sector_nr
;
3250 r10_bio
->master_bio
= NULL
;
3251 r10_bio
->sector
= sector_nr
;
3252 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3253 raid10_find_phys(conf
, r10_bio
);
3254 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3256 for (i
= 0; i
< conf
->copies
; i
++) {
3257 int d
= r10_bio
->devs
[i
].devnum
;
3258 sector_t first_bad
, sector
;
3261 if (r10_bio
->devs
[i
].repl_bio
)
3262 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3264 bio
= r10_bio
->devs
[i
].bio
;
3265 bio
->bi_end_io
= NULL
;
3266 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3267 if (conf
->mirrors
[d
].rdev
== NULL
||
3268 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3270 sector
= r10_bio
->devs
[i
].addr
;
3271 if (is_badblock(conf
->mirrors
[d
].rdev
,
3273 &first_bad
, &bad_sectors
)) {
3274 if (first_bad
> sector
)
3275 max_sync
= first_bad
- sector
;
3277 bad_sectors
-= (sector
- first_bad
);
3278 if (max_sync
> bad_sectors
)
3279 max_sync
= bad_sectors
;
3283 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3284 atomic_inc(&r10_bio
->remaining
);
3285 bio
->bi_next
= biolist
;
3287 bio
->bi_private
= r10_bio
;
3288 bio
->bi_end_io
= end_sync_read
;
3290 bio
->bi_sector
= sector
+
3291 conf
->mirrors
[d
].rdev
->data_offset
;
3292 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3295 if (conf
->mirrors
[d
].replacement
== NULL
||
3297 &conf
->mirrors
[d
].replacement
->flags
))
3300 /* Need to set up for writing to the replacement */
3301 bio
= r10_bio
->devs
[i
].repl_bio
;
3302 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3304 sector
= r10_bio
->devs
[i
].addr
;
3305 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3306 bio
->bi_next
= biolist
;
3308 bio
->bi_private
= r10_bio
;
3309 bio
->bi_end_io
= end_sync_write
;
3311 bio
->bi_sector
= sector
+
3312 conf
->mirrors
[d
].replacement
->data_offset
;
3313 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3318 for (i
=0; i
<conf
->copies
; i
++) {
3319 int d
= r10_bio
->devs
[i
].devnum
;
3320 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3321 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3323 if (r10_bio
->devs
[i
].repl_bio
&&
3324 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3326 conf
->mirrors
[d
].replacement
,
3335 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3337 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
3339 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
3342 bio
->bi_phys_segments
= 0;
3347 if (sector_nr
+ max_sync
< max_sector
)
3348 max_sector
= sector_nr
+ max_sync
;
3351 int len
= PAGE_SIZE
;
3352 if (sector_nr
+ (len
>>9) > max_sector
)
3353 len
= (max_sector
- sector_nr
) << 9;
3356 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3358 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3359 if (bio_add_page(bio
, page
, len
, 0))
3363 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3364 for (bio2
= biolist
;
3365 bio2
&& bio2
!= bio
;
3366 bio2
= bio2
->bi_next
) {
3367 /* remove last page from this bio */
3369 bio2
->bi_size
-= len
;
3370 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3374 nr_sectors
+= len
>>9;
3375 sector_nr
+= len
>>9;
3376 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3378 r10_bio
->sectors
= nr_sectors
;
3382 biolist
= biolist
->bi_next
;
3384 bio
->bi_next
= NULL
;
3385 r10_bio
= bio
->bi_private
;
3386 r10_bio
->sectors
= nr_sectors
;
3388 if (bio
->bi_end_io
== end_sync_read
) {
3389 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3390 generic_make_request(bio
);
3394 if (sectors_skipped
)
3395 /* pretend they weren't skipped, it makes
3396 * no important difference in this case
3398 md_done_sync(mddev
, sectors_skipped
, 1);
3400 return sectors_skipped
+ nr_sectors
;
3402 /* There is nowhere to write, so all non-sync
3403 * drives must be failed or in resync, all drives
3404 * have a bad block, so try the next chunk...
3406 if (sector_nr
+ max_sync
< max_sector
)
3407 max_sector
= sector_nr
+ max_sync
;
3409 sectors_skipped
+= (max_sector
- sector_nr
);
3411 sector_nr
= max_sector
;
3416 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3419 struct r10conf
*conf
= mddev
->private;
3422 raid_disks
= min(conf
->geo
.raid_disks
,
3423 conf
->prev
.raid_disks
);
3425 sectors
= conf
->dev_sectors
;
3427 size
= sectors
>> conf
->geo
.chunk_shift
;
3428 sector_div(size
, conf
->geo
.far_copies
);
3429 size
= size
* raid_disks
;
3430 sector_div(size
, conf
->geo
.near_copies
);
3432 return size
<< conf
->geo
.chunk_shift
;
3435 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3437 /* Calculate the number of sectors-per-device that will
3438 * actually be used, and set conf->dev_sectors and
3442 size
= size
>> conf
->geo
.chunk_shift
;
3443 sector_div(size
, conf
->geo
.far_copies
);
3444 size
= size
* conf
->geo
.raid_disks
;
3445 sector_div(size
, conf
->geo
.near_copies
);
3446 /* 'size' is now the number of chunks in the array */
3447 /* calculate "used chunks per device" */
3448 size
= size
* conf
->copies
;
3450 /* We need to round up when dividing by raid_disks to
3451 * get the stride size.
3453 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3455 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3457 if (conf
->geo
.far_offset
)
3458 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3460 sector_div(size
, conf
->geo
.far_copies
);
3461 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3465 enum geo_type
{geo_new
, geo_old
, geo_start
};
3466 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3469 int layout
, chunk
, disks
;
3472 layout
= mddev
->layout
;
3473 chunk
= mddev
->chunk_sectors
;
3474 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3477 layout
= mddev
->new_layout
;
3478 chunk
= mddev
->new_chunk_sectors
;
3479 disks
= mddev
->raid_disks
;
3481 default: /* avoid 'may be unused' warnings */
3482 case geo_start
: /* new when starting reshape - raid_disks not
3484 layout
= mddev
->new_layout
;
3485 chunk
= mddev
->new_chunk_sectors
;
3486 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3491 if (chunk
< (PAGE_SIZE
>> 9) ||
3492 !is_power_of_2(chunk
))
3495 fc
= (layout
>> 8) & 255;
3496 fo
= layout
& (1<<16);
3497 geo
->raid_disks
= disks
;
3498 geo
->near_copies
= nc
;
3499 geo
->far_copies
= fc
;
3500 geo
->far_offset
= fo
;
3501 geo
->far_set_size
= (layout
& (1<<17)) ? disks
/ fc
: disks
;
3502 geo
->chunk_mask
= chunk
- 1;
3503 geo
->chunk_shift
= ffz(~chunk
);
3507 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3509 struct r10conf
*conf
= NULL
;
3514 copies
= setup_geo(&geo
, mddev
, geo_new
);
3517 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3518 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3519 mdname(mddev
), PAGE_SIZE
);
3523 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3524 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3525 mdname(mddev
), mddev
->new_layout
);
3530 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3534 /* FIXME calc properly */
3535 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3536 max(0,mddev
->delta_disks
)),
3541 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3546 conf
->copies
= copies
;
3547 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3548 r10bio_pool_free
, conf
);
3549 if (!conf
->r10bio_pool
)
3552 calc_sectors(conf
, mddev
->dev_sectors
);
3553 if (mddev
->reshape_position
== MaxSector
) {
3554 conf
->prev
= conf
->geo
;
3555 conf
->reshape_progress
= MaxSector
;
3557 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3561 conf
->reshape_progress
= mddev
->reshape_position
;
3562 if (conf
->prev
.far_offset
)
3563 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3565 /* far_copies must be 1 */
3566 conf
->prev
.stride
= conf
->dev_sectors
;
3568 spin_lock_init(&conf
->device_lock
);
3569 INIT_LIST_HEAD(&conf
->retry_list
);
3571 spin_lock_init(&conf
->resync_lock
);
3572 init_waitqueue_head(&conf
->wait_barrier
);
3574 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3578 conf
->mddev
= mddev
;
3583 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3586 if (conf
->r10bio_pool
)
3587 mempool_destroy(conf
->r10bio_pool
);
3588 kfree(conf
->mirrors
);
3589 safe_put_page(conf
->tmppage
);
3592 return ERR_PTR(err
);
3595 static int run(struct mddev
*mddev
)
3597 struct r10conf
*conf
;
3598 int i
, disk_idx
, chunk_size
;
3599 struct raid10_info
*disk
;
3600 struct md_rdev
*rdev
;
3602 sector_t min_offset_diff
= 0;
3604 bool discard_supported
= false;
3606 if (mddev
->private == NULL
) {
3607 conf
= setup_conf(mddev
);
3609 return PTR_ERR(conf
);
3610 mddev
->private = conf
;
3612 conf
= mddev
->private;
3616 mddev
->thread
= conf
->thread
;
3617 conf
->thread
= NULL
;
3619 chunk_size
= mddev
->chunk_sectors
<< 9;
3621 blk_queue_max_discard_sectors(mddev
->queue
,
3622 mddev
->chunk_sectors
);
3623 blk_queue_max_write_same_sectors(mddev
->queue
,
3624 mddev
->chunk_sectors
);
3625 blk_queue_io_min(mddev
->queue
, chunk_size
);
3626 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3627 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3629 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3630 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3633 rdev_for_each(rdev
, mddev
) {
3635 struct request_queue
*q
;
3637 disk_idx
= rdev
->raid_disk
;
3640 if (disk_idx
>= conf
->geo
.raid_disks
&&
3641 disk_idx
>= conf
->prev
.raid_disks
)
3643 disk
= conf
->mirrors
+ disk_idx
;
3645 if (test_bit(Replacement
, &rdev
->flags
)) {
3646 if (disk
->replacement
)
3648 disk
->replacement
= rdev
;
3654 q
= bdev_get_queue(rdev
->bdev
);
3655 if (q
->merge_bvec_fn
)
3656 mddev
->merge_check_needed
= 1;
3657 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3658 if (!mddev
->reshape_backwards
)
3662 if (first
|| diff
< min_offset_diff
)
3663 min_offset_diff
= diff
;
3666 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3667 rdev
->data_offset
<< 9);
3669 disk
->head_position
= 0;
3671 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3672 discard_supported
= true;
3676 if (discard_supported
)
3677 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3680 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3683 /* need to check that every block has at least one working mirror */
3684 if (!enough(conf
, -1)) {
3685 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3690 if (conf
->reshape_progress
!= MaxSector
) {
3691 /* must ensure that shape change is supported */
3692 if (conf
->geo
.far_copies
!= 1 &&
3693 conf
->geo
.far_offset
== 0)
3695 if (conf
->prev
.far_copies
!= 1 &&
3696 conf
->geo
.far_offset
== 0)
3700 mddev
->degraded
= 0;
3702 i
< conf
->geo
.raid_disks
3703 || i
< conf
->prev
.raid_disks
;
3706 disk
= conf
->mirrors
+ i
;
3708 if (!disk
->rdev
&& disk
->replacement
) {
3709 /* The replacement is all we have - use it */
3710 disk
->rdev
= disk
->replacement
;
3711 disk
->replacement
= NULL
;
3712 clear_bit(Replacement
, &disk
->rdev
->flags
);
3716 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3717 disk
->head_position
= 0;
3722 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3725 if (mddev
->recovery_cp
!= MaxSector
)
3726 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3727 " -- starting background reconstruction\n",
3730 "md/raid10:%s: active with %d out of %d devices\n",
3731 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3732 conf
->geo
.raid_disks
);
3734 * Ok, everything is just fine now
3736 mddev
->dev_sectors
= conf
->dev_sectors
;
3737 size
= raid10_size(mddev
, 0, 0);
3738 md_set_array_sectors(mddev
, size
);
3739 mddev
->resync_max_sectors
= size
;
3742 int stripe
= conf
->geo
.raid_disks
*
3743 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3744 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3745 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3747 /* Calculate max read-ahead size.
3748 * We need to readahead at least twice a whole stripe....
3751 stripe
/= conf
->geo
.near_copies
;
3752 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3753 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3754 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3758 if (md_integrity_register(mddev
))
3761 if (conf
->reshape_progress
!= MaxSector
) {
3762 unsigned long before_length
, after_length
;
3764 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3765 conf
->prev
.far_copies
);
3766 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3767 conf
->geo
.far_copies
);
3769 if (max(before_length
, after_length
) > min_offset_diff
) {
3770 /* This cannot work */
3771 printk("md/raid10: offset difference not enough to continue reshape\n");
3774 conf
->offset_diff
= min_offset_diff
;
3776 conf
->reshape_safe
= conf
->reshape_progress
;
3777 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3778 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3779 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3780 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3781 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3788 md_unregister_thread(&mddev
->thread
);
3789 if (conf
->r10bio_pool
)
3790 mempool_destroy(conf
->r10bio_pool
);
3791 safe_put_page(conf
->tmppage
);
3792 kfree(conf
->mirrors
);
3794 mddev
->private = NULL
;
3799 static int stop(struct mddev
*mddev
)
3801 struct r10conf
*conf
= mddev
->private;
3803 raise_barrier(conf
, 0);
3804 lower_barrier(conf
);
3806 md_unregister_thread(&mddev
->thread
);
3808 /* the unplug fn references 'conf'*/
3809 blk_sync_queue(mddev
->queue
);
3811 if (conf
->r10bio_pool
)
3812 mempool_destroy(conf
->r10bio_pool
);
3813 kfree(conf
->mirrors
);
3815 mddev
->private = NULL
;
3819 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3821 struct r10conf
*conf
= mddev
->private;
3825 raise_barrier(conf
, 0);
3828 lower_barrier(conf
);
3833 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3835 /* Resize of 'far' arrays is not supported.
3836 * For 'near' and 'offset' arrays we can set the
3837 * number of sectors used to be an appropriate multiple
3838 * of the chunk size.
3839 * For 'offset', this is far_copies*chunksize.
3840 * For 'near' the multiplier is the LCM of
3841 * near_copies and raid_disks.
3842 * So if far_copies > 1 && !far_offset, fail.
3843 * Else find LCM(raid_disks, near_copy)*far_copies and
3844 * multiply by chunk_size. Then round to this number.
3845 * This is mostly done by raid10_size()
3847 struct r10conf
*conf
= mddev
->private;
3848 sector_t oldsize
, size
;
3850 if (mddev
->reshape_position
!= MaxSector
)
3853 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3856 oldsize
= raid10_size(mddev
, 0, 0);
3857 size
= raid10_size(mddev
, sectors
, 0);
3858 if (mddev
->external_size
&&
3859 mddev
->array_sectors
> size
)
3861 if (mddev
->bitmap
) {
3862 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3866 md_set_array_sectors(mddev
, size
);
3867 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3868 revalidate_disk(mddev
->gendisk
);
3869 if (sectors
> mddev
->dev_sectors
&&
3870 mddev
->recovery_cp
> oldsize
) {
3871 mddev
->recovery_cp
= oldsize
;
3872 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3874 calc_sectors(conf
, sectors
);
3875 mddev
->dev_sectors
= conf
->dev_sectors
;
3876 mddev
->resync_max_sectors
= size
;
3880 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3882 struct md_rdev
*rdev
;
3883 struct r10conf
*conf
;
3885 if (mddev
->degraded
> 0) {
3886 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3888 return ERR_PTR(-EINVAL
);
3891 /* Set new parameters */
3892 mddev
->new_level
= 10;
3893 /* new layout: far_copies = 1, near_copies = 2 */
3894 mddev
->new_layout
= (1<<8) + 2;
3895 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3896 mddev
->delta_disks
= mddev
->raid_disks
;
3897 mddev
->raid_disks
*= 2;
3898 /* make sure it will be not marked as dirty */
3899 mddev
->recovery_cp
= MaxSector
;
3901 conf
= setup_conf(mddev
);
3902 if (!IS_ERR(conf
)) {
3903 rdev_for_each(rdev
, mddev
)
3904 if (rdev
->raid_disk
>= 0)
3905 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3912 static void *raid10_takeover(struct mddev
*mddev
)
3914 struct r0conf
*raid0_conf
;
3916 /* raid10 can take over:
3917 * raid0 - providing it has only two drives
3919 if (mddev
->level
== 0) {
3920 /* for raid0 takeover only one zone is supported */
3921 raid0_conf
= mddev
->private;
3922 if (raid0_conf
->nr_strip_zones
> 1) {
3923 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3924 " with more than one zone.\n",
3926 return ERR_PTR(-EINVAL
);
3928 return raid10_takeover_raid0(mddev
);
3930 return ERR_PTR(-EINVAL
);
3933 static int raid10_check_reshape(struct mddev
*mddev
)
3935 /* Called when there is a request to change
3936 * - layout (to ->new_layout)
3937 * - chunk size (to ->new_chunk_sectors)
3938 * - raid_disks (by delta_disks)
3939 * or when trying to restart a reshape that was ongoing.
3941 * We need to validate the request and possibly allocate
3942 * space if that might be an issue later.
3944 * Currently we reject any reshape of a 'far' mode array,
3945 * allow chunk size to change if new is generally acceptable,
3946 * allow raid_disks to increase, and allow
3947 * a switch between 'near' mode and 'offset' mode.
3949 struct r10conf
*conf
= mddev
->private;
3952 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3955 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3956 /* mustn't change number of copies */
3958 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3959 /* Cannot switch to 'far' mode */
3962 if (mddev
->array_sectors
& geo
.chunk_mask
)
3963 /* not factor of array size */
3966 if (!enough(conf
, -1))
3969 kfree(conf
->mirrors_new
);
3970 conf
->mirrors_new
= NULL
;
3971 if (mddev
->delta_disks
> 0) {
3972 /* allocate new 'mirrors' list */
3973 conf
->mirrors_new
= kzalloc(
3974 sizeof(struct raid10_info
)
3975 *(mddev
->raid_disks
+
3976 mddev
->delta_disks
),
3978 if (!conf
->mirrors_new
)
3985 * Need to check if array has failed when deciding whether to:
3987 * - remove non-faulty devices
3990 * This determination is simple when no reshape is happening.
3991 * However if there is a reshape, we need to carefully check
3992 * both the before and after sections.
3993 * This is because some failed devices may only affect one
3994 * of the two sections, and some non-in_sync devices may
3995 * be insync in the section most affected by failed devices.
3997 static int calc_degraded(struct r10conf
*conf
)
3999 int degraded
, degraded2
;
4004 /* 'prev' section first */
4005 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4006 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4007 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4009 else if (!test_bit(In_sync
, &rdev
->flags
))
4010 /* When we can reduce the number of devices in
4011 * an array, this might not contribute to
4012 * 'degraded'. It does now.
4017 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4021 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4022 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4023 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4025 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4026 /* If reshape is increasing the number of devices,
4027 * this section has already been recovered, so
4028 * it doesn't contribute to degraded.
4031 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4036 if (degraded2
> degraded
)
4041 static int raid10_start_reshape(struct mddev
*mddev
)
4043 /* A 'reshape' has been requested. This commits
4044 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4045 * This also checks if there are enough spares and adds them
4047 * We currently require enough spares to make the final
4048 * array non-degraded. We also require that the difference
4049 * between old and new data_offset - on each device - is
4050 * enough that we never risk over-writing.
4053 unsigned long before_length
, after_length
;
4054 sector_t min_offset_diff
= 0;
4057 struct r10conf
*conf
= mddev
->private;
4058 struct md_rdev
*rdev
;
4062 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4065 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4068 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4069 conf
->prev
.far_copies
);
4070 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4071 conf
->geo
.far_copies
);
4073 rdev_for_each(rdev
, mddev
) {
4074 if (!test_bit(In_sync
, &rdev
->flags
)
4075 && !test_bit(Faulty
, &rdev
->flags
))
4077 if (rdev
->raid_disk
>= 0) {
4078 long long diff
= (rdev
->new_data_offset
4079 - rdev
->data_offset
);
4080 if (!mddev
->reshape_backwards
)
4084 if (first
|| diff
< min_offset_diff
)
4085 min_offset_diff
= diff
;
4089 if (max(before_length
, after_length
) > min_offset_diff
)
4092 if (spares
< mddev
->delta_disks
)
4095 conf
->offset_diff
= min_offset_diff
;
4096 spin_lock_irq(&conf
->device_lock
);
4097 if (conf
->mirrors_new
) {
4098 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4099 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4101 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
4102 conf
->mirrors_old
= conf
->mirrors
;
4103 conf
->mirrors
= conf
->mirrors_new
;
4104 conf
->mirrors_new
= NULL
;
4106 setup_geo(&conf
->geo
, mddev
, geo_start
);
4108 if (mddev
->reshape_backwards
) {
4109 sector_t size
= raid10_size(mddev
, 0, 0);
4110 if (size
< mddev
->array_sectors
) {
4111 spin_unlock_irq(&conf
->device_lock
);
4112 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4116 mddev
->resync_max_sectors
= size
;
4117 conf
->reshape_progress
= size
;
4119 conf
->reshape_progress
= 0;
4120 spin_unlock_irq(&conf
->device_lock
);
4122 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4123 ret
= bitmap_resize(mddev
->bitmap
,
4124 raid10_size(mddev
, 0,
4125 conf
->geo
.raid_disks
),
4130 if (mddev
->delta_disks
> 0) {
4131 rdev_for_each(rdev
, mddev
)
4132 if (rdev
->raid_disk
< 0 &&
4133 !test_bit(Faulty
, &rdev
->flags
)) {
4134 if (raid10_add_disk(mddev
, rdev
) == 0) {
4135 if (rdev
->raid_disk
>=
4136 conf
->prev
.raid_disks
)
4137 set_bit(In_sync
, &rdev
->flags
);
4139 rdev
->recovery_offset
= 0;
4141 if (sysfs_link_rdev(mddev
, rdev
))
4142 /* Failure here is OK */;
4144 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4145 && !test_bit(Faulty
, &rdev
->flags
)) {
4146 /* This is a spare that was manually added */
4147 set_bit(In_sync
, &rdev
->flags
);
4150 /* When a reshape changes the number of devices,
4151 * ->degraded is measured against the larger of the
4152 * pre and post numbers.
4154 spin_lock_irq(&conf
->device_lock
);
4155 mddev
->degraded
= calc_degraded(conf
);
4156 spin_unlock_irq(&conf
->device_lock
);
4157 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4158 mddev
->reshape_position
= conf
->reshape_progress
;
4159 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4161 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4162 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4163 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4164 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4166 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4168 if (!mddev
->sync_thread
) {
4172 conf
->reshape_checkpoint
= jiffies
;
4173 md_wakeup_thread(mddev
->sync_thread
);
4174 md_new_event(mddev
);
4178 mddev
->recovery
= 0;
4179 spin_lock_irq(&conf
->device_lock
);
4180 conf
->geo
= conf
->prev
;
4181 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4182 rdev_for_each(rdev
, mddev
)
4183 rdev
->new_data_offset
= rdev
->data_offset
;
4185 conf
->reshape_progress
= MaxSector
;
4186 mddev
->reshape_position
= MaxSector
;
4187 spin_unlock_irq(&conf
->device_lock
);
4191 /* Calculate the last device-address that could contain
4192 * any block from the chunk that includes the array-address 's'
4193 * and report the next address.
4194 * i.e. the address returned will be chunk-aligned and after
4195 * any data that is in the chunk containing 's'.
4197 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4199 s
= (s
| geo
->chunk_mask
) + 1;
4200 s
>>= geo
->chunk_shift
;
4201 s
*= geo
->near_copies
;
4202 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4203 s
*= geo
->far_copies
;
4204 s
<<= geo
->chunk_shift
;
4208 /* Calculate the first device-address that could contain
4209 * any block from the chunk that includes the array-address 's'.
4210 * This too will be the start of a chunk
4212 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4214 s
>>= geo
->chunk_shift
;
4215 s
*= geo
->near_copies
;
4216 sector_div(s
, geo
->raid_disks
);
4217 s
*= geo
->far_copies
;
4218 s
<<= geo
->chunk_shift
;
4222 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4225 /* We simply copy at most one chunk (smallest of old and new)
4226 * at a time, possibly less if that exceeds RESYNC_PAGES,
4227 * or we hit a bad block or something.
4228 * This might mean we pause for normal IO in the middle of
4229 * a chunk, but that is not a problem was mddev->reshape_position
4230 * can record any location.
4232 * If we will want to write to a location that isn't
4233 * yet recorded as 'safe' (i.e. in metadata on disk) then
4234 * we need to flush all reshape requests and update the metadata.
4236 * When reshaping forwards (e.g. to more devices), we interpret
4237 * 'safe' as the earliest block which might not have been copied
4238 * down yet. We divide this by previous stripe size and multiply
4239 * by previous stripe length to get lowest device offset that we
4240 * cannot write to yet.
4241 * We interpret 'sector_nr' as an address that we want to write to.
4242 * From this we use last_device_address() to find where we might
4243 * write to, and first_device_address on the 'safe' position.
4244 * If this 'next' write position is after the 'safe' position,
4245 * we must update the metadata to increase the 'safe' position.
4247 * When reshaping backwards, we round in the opposite direction
4248 * and perform the reverse test: next write position must not be
4249 * less than current safe position.
4251 * In all this the minimum difference in data offsets
4252 * (conf->offset_diff - always positive) allows a bit of slack,
4253 * so next can be after 'safe', but not by more than offset_disk
4255 * We need to prepare all the bios here before we start any IO
4256 * to ensure the size we choose is acceptable to all devices.
4257 * The means one for each copy for write-out and an extra one for
4259 * We store the read-in bio in ->master_bio and the others in
4260 * ->devs[x].bio and ->devs[x].repl_bio.
4262 struct r10conf
*conf
= mddev
->private;
4263 struct r10bio
*r10_bio
;
4264 sector_t next
, safe
, last
;
4268 struct md_rdev
*rdev
;
4271 struct bio
*bio
, *read_bio
;
4272 int sectors_done
= 0;
4274 if (sector_nr
== 0) {
4275 /* If restarting in the middle, skip the initial sectors */
4276 if (mddev
->reshape_backwards
&&
4277 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4278 sector_nr
= (raid10_size(mddev
, 0, 0)
4279 - conf
->reshape_progress
);
4280 } else if (!mddev
->reshape_backwards
&&
4281 conf
->reshape_progress
> 0)
4282 sector_nr
= conf
->reshape_progress
;
4284 mddev
->curr_resync_completed
= sector_nr
;
4285 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4291 /* We don't use sector_nr to track where we are up to
4292 * as that doesn't work well for ->reshape_backwards.
4293 * So just use ->reshape_progress.
4295 if (mddev
->reshape_backwards
) {
4296 /* 'next' is the earliest device address that we might
4297 * write to for this chunk in the new layout
4299 next
= first_dev_address(conf
->reshape_progress
- 1,
4302 /* 'safe' is the last device address that we might read from
4303 * in the old layout after a restart
4305 safe
= last_dev_address(conf
->reshape_safe
- 1,
4308 if (next
+ conf
->offset_diff
< safe
)
4311 last
= conf
->reshape_progress
- 1;
4312 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4313 & conf
->prev
.chunk_mask
);
4314 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4315 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4317 /* 'next' is after the last device address that we
4318 * might write to for this chunk in the new layout
4320 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4322 /* 'safe' is the earliest device address that we might
4323 * read from in the old layout after a restart
4325 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4327 /* Need to update metadata if 'next' might be beyond 'safe'
4328 * as that would possibly corrupt data
4330 if (next
> safe
+ conf
->offset_diff
)
4333 sector_nr
= conf
->reshape_progress
;
4334 last
= sector_nr
| (conf
->geo
.chunk_mask
4335 & conf
->prev
.chunk_mask
);
4337 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4338 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4342 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4343 /* Need to update reshape_position in metadata */
4345 mddev
->reshape_position
= conf
->reshape_progress
;
4346 if (mddev
->reshape_backwards
)
4347 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4348 - conf
->reshape_progress
;
4350 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4351 conf
->reshape_checkpoint
= jiffies
;
4352 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4353 md_wakeup_thread(mddev
->thread
);
4354 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4355 kthread_should_stop());
4356 conf
->reshape_safe
= mddev
->reshape_position
;
4357 allow_barrier(conf
);
4361 /* Now schedule reads for blocks from sector_nr to last */
4362 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4363 raise_barrier(conf
, sectors_done
!= 0);
4364 atomic_set(&r10_bio
->remaining
, 0);
4365 r10_bio
->mddev
= mddev
;
4366 r10_bio
->sector
= sector_nr
;
4367 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4368 r10_bio
->sectors
= last
- sector_nr
+ 1;
4369 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4370 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4373 /* Cannot read from here, so need to record bad blocks
4374 * on all the target devices.
4377 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4378 return sectors_done
;
4381 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4383 read_bio
->bi_bdev
= rdev
->bdev
;
4384 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4385 + rdev
->data_offset
);
4386 read_bio
->bi_private
= r10_bio
;
4387 read_bio
->bi_end_io
= end_sync_read
;
4388 read_bio
->bi_rw
= READ
;
4389 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4390 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4391 read_bio
->bi_vcnt
= 0;
4392 read_bio
->bi_idx
= 0;
4393 read_bio
->bi_size
= 0;
4394 r10_bio
->master_bio
= read_bio
;
4395 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4397 /* Now find the locations in the new layout */
4398 __raid10_find_phys(&conf
->geo
, r10_bio
);
4401 read_bio
->bi_next
= NULL
;
4403 for (s
= 0; s
< conf
->copies
*2; s
++) {
4405 int d
= r10_bio
->devs
[s
/2].devnum
;
4406 struct md_rdev
*rdev2
;
4408 rdev2
= conf
->mirrors
[d
].replacement
;
4409 b
= r10_bio
->devs
[s
/2].repl_bio
;
4411 rdev2
= conf
->mirrors
[d
].rdev
;
4412 b
= r10_bio
->devs
[s
/2].bio
;
4414 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4416 b
->bi_bdev
= rdev2
->bdev
;
4417 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4418 b
->bi_private
= r10_bio
;
4419 b
->bi_end_io
= end_reshape_write
;
4421 b
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4422 b
->bi_flags
|= 1 << BIO_UPTODATE
;
4430 /* Now add as many pages as possible to all of these bios. */
4433 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4434 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4435 int len
= (max_sectors
- s
) << 9;
4436 if (len
> PAGE_SIZE
)
4438 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4440 if (bio_add_page(bio
, page
, len
, 0))
4443 /* Didn't fit, must stop */
4445 bio2
&& bio2
!= bio
;
4446 bio2
= bio2
->bi_next
) {
4447 /* Remove last page from this bio */
4449 bio2
->bi_size
-= len
;
4450 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4454 sector_nr
+= len
>> 9;
4455 nr_sectors
+= len
>> 9;
4458 r10_bio
->sectors
= nr_sectors
;
4460 /* Now submit the read */
4461 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4462 atomic_inc(&r10_bio
->remaining
);
4463 read_bio
->bi_next
= NULL
;
4464 generic_make_request(read_bio
);
4465 sector_nr
+= nr_sectors
;
4466 sectors_done
+= nr_sectors
;
4467 if (sector_nr
<= last
)
4470 /* Now that we have done the whole section we can
4471 * update reshape_progress
4473 if (mddev
->reshape_backwards
)
4474 conf
->reshape_progress
-= sectors_done
;
4476 conf
->reshape_progress
+= sectors_done
;
4478 return sectors_done
;
4481 static void end_reshape_request(struct r10bio
*r10_bio
);
4482 static int handle_reshape_read_error(struct mddev
*mddev
,
4483 struct r10bio
*r10_bio
);
4484 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4486 /* Reshape read completed. Hopefully we have a block
4488 * If we got a read error then we do sync 1-page reads from
4489 * elsewhere until we find the data - or give up.
4491 struct r10conf
*conf
= mddev
->private;
4494 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4495 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4496 /* Reshape has been aborted */
4497 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4501 /* We definitely have the data in the pages, schedule the
4504 atomic_set(&r10_bio
->remaining
, 1);
4505 for (s
= 0; s
< conf
->copies
*2; s
++) {
4507 int d
= r10_bio
->devs
[s
/2].devnum
;
4508 struct md_rdev
*rdev
;
4510 rdev
= conf
->mirrors
[d
].replacement
;
4511 b
= r10_bio
->devs
[s
/2].repl_bio
;
4513 rdev
= conf
->mirrors
[d
].rdev
;
4514 b
= r10_bio
->devs
[s
/2].bio
;
4516 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4518 atomic_inc(&rdev
->nr_pending
);
4519 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4520 atomic_inc(&r10_bio
->remaining
);
4522 generic_make_request(b
);
4524 end_reshape_request(r10_bio
);
4527 static void end_reshape(struct r10conf
*conf
)
4529 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4532 spin_lock_irq(&conf
->device_lock
);
4533 conf
->prev
= conf
->geo
;
4534 md_finish_reshape(conf
->mddev
);
4536 conf
->reshape_progress
= MaxSector
;
4537 spin_unlock_irq(&conf
->device_lock
);
4539 /* read-ahead size must cover two whole stripes, which is
4540 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4542 if (conf
->mddev
->queue
) {
4543 int stripe
= conf
->geo
.raid_disks
*
4544 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4545 stripe
/= conf
->geo
.near_copies
;
4546 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4547 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4553 static int handle_reshape_read_error(struct mddev
*mddev
,
4554 struct r10bio
*r10_bio
)
4556 /* Use sync reads to get the blocks from somewhere else */
4557 int sectors
= r10_bio
->sectors
;
4558 struct r10conf
*conf
= mddev
->private;
4560 struct r10bio r10_bio
;
4561 struct r10dev devs
[conf
->copies
];
4563 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4566 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4568 r10b
->sector
= r10_bio
->sector
;
4569 __raid10_find_phys(&conf
->prev
, r10b
);
4574 int first_slot
= slot
;
4576 if (s
> (PAGE_SIZE
>> 9))
4580 int d
= r10b
->devs
[slot
].devnum
;
4581 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4584 test_bit(Faulty
, &rdev
->flags
) ||
4585 !test_bit(In_sync
, &rdev
->flags
))
4588 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4589 success
= sync_page_io(rdev
,
4598 if (slot
>= conf
->copies
)
4600 if (slot
== first_slot
)
4604 /* couldn't read this block, must give up */
4605 set_bit(MD_RECOVERY_INTR
,
4615 static void end_reshape_write(struct bio
*bio
, int error
)
4617 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4618 struct r10bio
*r10_bio
= bio
->bi_private
;
4619 struct mddev
*mddev
= r10_bio
->mddev
;
4620 struct r10conf
*conf
= mddev
->private;
4624 struct md_rdev
*rdev
= NULL
;
4626 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4628 rdev
= conf
->mirrors
[d
].replacement
;
4631 rdev
= conf
->mirrors
[d
].rdev
;
4635 /* FIXME should record badblock */
4636 md_error(mddev
, rdev
);
4639 rdev_dec_pending(rdev
, mddev
);
4640 end_reshape_request(r10_bio
);
4643 static void end_reshape_request(struct r10bio
*r10_bio
)
4645 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4647 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4648 bio_put(r10_bio
->master_bio
);
4652 static void raid10_finish_reshape(struct mddev
*mddev
)
4654 struct r10conf
*conf
= mddev
->private;
4656 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4659 if (mddev
->delta_disks
> 0) {
4660 sector_t size
= raid10_size(mddev
, 0, 0);
4661 md_set_array_sectors(mddev
, size
);
4662 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4663 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4664 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4666 mddev
->resync_max_sectors
= size
;
4667 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4668 revalidate_disk(mddev
->gendisk
);
4671 for (d
= conf
->geo
.raid_disks
;
4672 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4674 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4676 clear_bit(In_sync
, &rdev
->flags
);
4677 rdev
= conf
->mirrors
[d
].replacement
;
4679 clear_bit(In_sync
, &rdev
->flags
);
4682 mddev
->layout
= mddev
->new_layout
;
4683 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4684 mddev
->reshape_position
= MaxSector
;
4685 mddev
->delta_disks
= 0;
4686 mddev
->reshape_backwards
= 0;
4689 static struct md_personality raid10_personality
=
4693 .owner
= THIS_MODULE
,
4694 .make_request
= make_request
,
4698 .error_handler
= error
,
4699 .hot_add_disk
= raid10_add_disk
,
4700 .hot_remove_disk
= raid10_remove_disk
,
4701 .spare_active
= raid10_spare_active
,
4702 .sync_request
= sync_request
,
4703 .quiesce
= raid10_quiesce
,
4704 .size
= raid10_size
,
4705 .resize
= raid10_resize
,
4706 .takeover
= raid10_takeover
,
4707 .check_reshape
= raid10_check_reshape
,
4708 .start_reshape
= raid10_start_reshape
,
4709 .finish_reshape
= raid10_finish_reshape
,
4712 static int __init
raid_init(void)
4714 return register_md_personality(&raid10_personality
);
4717 static void raid_exit(void)
4719 unregister_md_personality(&raid10_personality
);
4722 module_init(raid_init
);
4723 module_exit(raid_exit
);
4724 MODULE_LICENSE("GPL");
4725 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4726 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4727 MODULE_ALIAS("md-raid10");
4728 MODULE_ALIAS("md-level-10");
4730 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);