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>
28 #include <trace/events/block.h>
32 #include "md-bitmap.h"
35 * RAID10 provides a combination of RAID0 and RAID1 functionality.
36 * The layout of data is defined by
39 * near_copies (stored in low byte of layout)
40 * far_copies (stored in second byte of layout)
41 * far_offset (stored in bit 16 of layout )
42 * use_far_sets (stored in bit 17 of layout )
43 * use_far_sets_bugfixed (stored in bit 18 of layout )
45 * The data to be stored is divided into chunks using chunksize. Each device
46 * is divided into far_copies sections. In each section, chunks are laid out
47 * in a style similar to raid0, but near_copies copies of each chunk is stored
48 * (each on a different drive). The starting device for each section is offset
49 * near_copies from the starting device of the previous section. Thus there
50 * are (near_copies * far_copies) of each chunk, and each is on a different
51 * drive. near_copies and far_copies must be at least one, and their product
52 * is at most raid_disks.
54 * If far_offset is true, then the far_copies are handled a bit differently.
55 * The copies are still in different stripes, but instead of being very far
56 * apart on disk, there are adjacent stripes.
58 * The far and offset algorithms are handled slightly differently if
59 * 'use_far_sets' is true. In this case, the array's devices are grouped into
60 * sets that are (near_copies * far_copies) in size. The far copied stripes
61 * are still shifted by 'near_copies' devices, but this shifting stays confined
62 * to the set rather than the entire array. This is done to improve the number
63 * of device combinations that can fail without causing the array to fail.
64 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
69 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70 * [A B] [C D] [A B] [C D E]
71 * |...| |...| |...| | ... |
72 * [B A] [D C] [B A] [E C D]
76 * Number of guaranteed r10bios in case of extreme VM load:
78 #define NR_RAID10_BIOS 256
80 /* when we get a read error on a read-only array, we redirect to another
81 * device without failing the first device, or trying to over-write to
82 * correct the read error. To keep track of bad blocks on a per-bio
83 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
85 #define IO_BLOCKED ((struct bio *)1)
86 /* When we successfully write to a known bad-block, we need to remove the
87 * bad-block marking which must be done from process context. So we record
88 * the success by setting devs[n].bio to IO_MADE_GOOD
90 #define IO_MADE_GOOD ((struct bio *)2)
92 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
94 /* When there are this many requests queued to be written by
95 * the raid10 thread, we become 'congested' to provide back-pressure
98 static int max_queued_requests
= 1024;
100 static void allow_barrier(struct r10conf
*conf
);
101 static void lower_barrier(struct r10conf
*conf
);
102 static int _enough(struct r10conf
*conf
, int previous
, int ignore
);
103 static int enough(struct r10conf
*conf
, int ignore
);
104 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
106 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
);
107 static void end_reshape_write(struct bio
*bio
);
108 static void end_reshape(struct r10conf
*conf
);
110 #define raid10_log(md, fmt, args...) \
111 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
113 #include "raid1-10.c"
116 * for resync bio, r10bio pointer can be retrieved from the per-bio
117 * 'struct resync_pages'.
119 static inline struct r10bio
*get_resync_r10bio(struct bio
*bio
)
121 return get_resync_pages(bio
)->raid_bio
;
124 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
126 struct r10conf
*conf
= data
;
127 int size
= offsetof(struct r10bio
, devs
[conf
->copies
]);
129 /* allocate a r10bio with room for raid_disks entries in the
131 return kzalloc(size
, gfp_flags
);
134 static void r10bio_pool_free(void *r10_bio
, void *data
)
139 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
140 /* amount of memory to reserve for resync requests */
141 #define RESYNC_WINDOW (1024*1024)
142 /* maximum number of concurrent requests, memory permitting */
143 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
144 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
145 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
148 * When performing a resync, we need to read and compare, so
149 * we need as many pages are there are copies.
150 * When performing a recovery, we need 2 bios, one for read,
151 * one for write (we recover only one drive per r10buf)
154 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
156 struct r10conf
*conf
= data
;
157 struct r10bio
*r10_bio
;
160 int nalloc
, nalloc_rp
;
161 struct resync_pages
*rps
;
163 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
167 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
) ||
168 test_bit(MD_RECOVERY_RESHAPE
, &conf
->mddev
->recovery
))
169 nalloc
= conf
->copies
; /* resync */
171 nalloc
= 2; /* recovery */
173 /* allocate once for all bios */
174 if (!conf
->have_replacement
)
177 nalloc_rp
= nalloc
* 2;
178 rps
= kmalloc(sizeof(struct resync_pages
) * nalloc_rp
, gfp_flags
);
180 goto out_free_r10bio
;
185 for (j
= nalloc
; j
-- ; ) {
186 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
189 r10_bio
->devs
[j
].bio
= bio
;
190 if (!conf
->have_replacement
)
192 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
195 r10_bio
->devs
[j
].repl_bio
= bio
;
198 * Allocate RESYNC_PAGES data pages and attach them
201 for (j
= 0; j
< nalloc
; j
++) {
202 struct bio
*rbio
= r10_bio
->devs
[j
].repl_bio
;
203 struct resync_pages
*rp
, *rp_repl
;
207 rp_repl
= &rps
[nalloc
+ j
];
209 bio
= r10_bio
->devs
[j
].bio
;
211 if (!j
|| test_bit(MD_RECOVERY_SYNC
,
212 &conf
->mddev
->recovery
)) {
213 if (resync_alloc_pages(rp
, gfp_flags
))
216 memcpy(rp
, &rps
[0], sizeof(*rp
));
217 resync_get_all_pages(rp
);
220 rp
->raid_bio
= r10_bio
;
221 bio
->bi_private
= rp
;
223 memcpy(rp_repl
, rp
, sizeof(*rp
));
224 rbio
->bi_private
= rp_repl
;
232 resync_free_pages(&rps
[j
* 2]);
236 for ( ; j
< nalloc
; j
++) {
237 if (r10_bio
->devs
[j
].bio
)
238 bio_put(r10_bio
->devs
[j
].bio
);
239 if (r10_bio
->devs
[j
].repl_bio
)
240 bio_put(r10_bio
->devs
[j
].repl_bio
);
244 r10bio_pool_free(r10_bio
, conf
);
248 static void r10buf_pool_free(void *__r10_bio
, void *data
)
250 struct r10conf
*conf
= data
;
251 struct r10bio
*r10bio
= __r10_bio
;
253 struct resync_pages
*rp
= NULL
;
255 for (j
= conf
->copies
; j
--; ) {
256 struct bio
*bio
= r10bio
->devs
[j
].bio
;
258 rp
= get_resync_pages(bio
);
259 resync_free_pages(rp
);
262 bio
= r10bio
->devs
[j
].repl_bio
;
267 /* resync pages array stored in the 1st bio's .bi_private */
270 r10bio_pool_free(r10bio
, conf
);
273 static void put_all_bios(struct r10conf
*conf
, struct r10bio
*r10_bio
)
277 for (i
= 0; i
< conf
->copies
; i
++) {
278 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
279 if (!BIO_SPECIAL(*bio
))
282 bio
= &r10_bio
->devs
[i
].repl_bio
;
283 if (r10_bio
->read_slot
< 0 && !BIO_SPECIAL(*bio
))
289 static void free_r10bio(struct r10bio
*r10_bio
)
291 struct r10conf
*conf
= r10_bio
->mddev
->private;
293 put_all_bios(conf
, r10_bio
);
294 mempool_free(r10_bio
, conf
->r10bio_pool
);
297 static void put_buf(struct r10bio
*r10_bio
)
299 struct r10conf
*conf
= r10_bio
->mddev
->private;
301 mempool_free(r10_bio
, conf
->r10buf_pool
);
306 static void reschedule_retry(struct r10bio
*r10_bio
)
309 struct mddev
*mddev
= r10_bio
->mddev
;
310 struct r10conf
*conf
= mddev
->private;
312 spin_lock_irqsave(&conf
->device_lock
, flags
);
313 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
315 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
317 /* wake up frozen array... */
318 wake_up(&conf
->wait_barrier
);
320 md_wakeup_thread(mddev
->thread
);
324 * raid_end_bio_io() is called when we have finished servicing a mirrored
325 * operation and are ready to return a success/failure code to the buffer
328 static void raid_end_bio_io(struct r10bio
*r10_bio
)
330 struct bio
*bio
= r10_bio
->master_bio
;
331 struct r10conf
*conf
= r10_bio
->mddev
->private;
333 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
334 bio
->bi_status
= BLK_STS_IOERR
;
338 * Wake up any possible resync thread that waits for the device
343 free_r10bio(r10_bio
);
347 * Update disk head position estimator based on IRQ completion info.
349 static inline void update_head_pos(int slot
, struct r10bio
*r10_bio
)
351 struct r10conf
*conf
= r10_bio
->mddev
->private;
353 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
354 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
358 * Find the disk number which triggered given bio
360 static int find_bio_disk(struct r10conf
*conf
, struct r10bio
*r10_bio
,
361 struct bio
*bio
, int *slotp
, int *replp
)
366 for (slot
= 0; slot
< conf
->copies
; slot
++) {
367 if (r10_bio
->devs
[slot
].bio
== bio
)
369 if (r10_bio
->devs
[slot
].repl_bio
== bio
) {
375 BUG_ON(slot
== conf
->copies
);
376 update_head_pos(slot
, r10_bio
);
382 return r10_bio
->devs
[slot
].devnum
;
385 static void raid10_end_read_request(struct bio
*bio
)
387 int uptodate
= !bio
->bi_status
;
388 struct r10bio
*r10_bio
= bio
->bi_private
;
390 struct md_rdev
*rdev
;
391 struct r10conf
*conf
= r10_bio
->mddev
->private;
393 slot
= r10_bio
->read_slot
;
394 rdev
= r10_bio
->devs
[slot
].rdev
;
396 * this branch is our 'one mirror IO has finished' event handler:
398 update_head_pos(slot
, r10_bio
);
402 * Set R10BIO_Uptodate in our master bio, so that
403 * we will return a good error code to the higher
404 * levels even if IO on some other mirrored buffer fails.
406 * The 'master' represents the composite IO operation to
407 * user-side. So if something waits for IO, then it will
408 * wait for the 'master' bio.
410 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
412 /* If all other devices that store this block have
413 * failed, we want to return the error upwards rather
414 * than fail the last device. Here we redefine
415 * "uptodate" to mean "Don't want to retry"
417 if (!_enough(conf
, test_bit(R10BIO_Previous
, &r10_bio
->state
),
422 raid_end_bio_io(r10_bio
);
423 rdev_dec_pending(rdev
, conf
->mddev
);
426 * oops, read error - keep the refcount on the rdev
428 char b
[BDEVNAME_SIZE
];
429 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
431 bdevname(rdev
->bdev
, b
),
432 (unsigned long long)r10_bio
->sector
);
433 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
434 reschedule_retry(r10_bio
);
438 static void close_write(struct r10bio
*r10_bio
)
440 /* clear the bitmap if all writes complete successfully */
441 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
443 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
445 md_write_end(r10_bio
->mddev
);
448 static void one_write_done(struct r10bio
*r10_bio
)
450 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
451 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
452 reschedule_retry(r10_bio
);
454 close_write(r10_bio
);
455 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
456 reschedule_retry(r10_bio
);
458 raid_end_bio_io(r10_bio
);
463 static void raid10_end_write_request(struct bio
*bio
)
465 struct r10bio
*r10_bio
= bio
->bi_private
;
468 struct r10conf
*conf
= r10_bio
->mddev
->private;
470 struct md_rdev
*rdev
= NULL
;
471 struct bio
*to_put
= NULL
;
474 discard_error
= bio
->bi_status
&& bio_op(bio
) == REQ_OP_DISCARD
;
476 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
479 rdev
= conf
->mirrors
[dev
].replacement
;
483 rdev
= conf
->mirrors
[dev
].rdev
;
486 * this branch is our 'one mirror IO has finished' event handler:
488 if (bio
->bi_status
&& !discard_error
) {
490 /* Never record new bad blocks to replacement,
493 md_error(rdev
->mddev
, rdev
);
495 set_bit(WriteErrorSeen
, &rdev
->flags
);
496 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
497 set_bit(MD_RECOVERY_NEEDED
,
498 &rdev
->mddev
->recovery
);
501 if (test_bit(FailFast
, &rdev
->flags
) &&
502 (bio
->bi_opf
& MD_FAILFAST
)) {
503 md_error(rdev
->mddev
, rdev
);
504 if (!test_bit(Faulty
, &rdev
->flags
))
505 /* This is the only remaining device,
506 * We need to retry the write without
509 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
511 r10_bio
->devs
[slot
].bio
= NULL
;
516 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
520 * Set R10BIO_Uptodate in our master bio, so that
521 * we will return a good error code for to the higher
522 * levels even if IO on some other mirrored buffer fails.
524 * The 'master' represents the composite IO operation to
525 * user-side. So if something waits for IO, then it will
526 * wait for the 'master' bio.
532 * Do not set R10BIO_Uptodate if the current device is
533 * rebuilding or Faulty. This is because we cannot use
534 * such device for properly reading the data back (we could
535 * potentially use it, if the current write would have felt
536 * before rdev->recovery_offset, but for simplicity we don't
539 if (test_bit(In_sync
, &rdev
->flags
) &&
540 !test_bit(Faulty
, &rdev
->flags
))
541 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
543 /* Maybe we can clear some bad blocks. */
544 if (is_badblock(rdev
,
545 r10_bio
->devs
[slot
].addr
,
547 &first_bad
, &bad_sectors
) && !discard_error
) {
550 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
552 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
554 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
560 * Let's see if all mirrored write operations have finished
563 one_write_done(r10_bio
);
565 rdev_dec_pending(rdev
, conf
->mddev
);
571 * RAID10 layout manager
572 * As well as the chunksize and raid_disks count, there are two
573 * parameters: near_copies and far_copies.
574 * near_copies * far_copies must be <= raid_disks.
575 * Normally one of these will be 1.
576 * If both are 1, we get raid0.
577 * If near_copies == raid_disks, we get raid1.
579 * Chunks are laid out in raid0 style with near_copies copies of the
580 * first chunk, followed by near_copies copies of the next chunk and
582 * If far_copies > 1, then after 1/far_copies of the array has been assigned
583 * as described above, we start again with a device offset of near_copies.
584 * So we effectively have another copy of the whole array further down all
585 * the drives, but with blocks on different drives.
586 * With this layout, and block is never stored twice on the one device.
588 * raid10_find_phys finds the sector offset of a given virtual sector
589 * on each device that it is on.
591 * raid10_find_virt does the reverse mapping, from a device and a
592 * sector offset to a virtual address
595 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
603 int last_far_set_start
, last_far_set_size
;
605 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
606 last_far_set_start
*= geo
->far_set_size
;
608 last_far_set_size
= geo
->far_set_size
;
609 last_far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
611 /* now calculate first sector/dev */
612 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
613 sector
= r10bio
->sector
& geo
->chunk_mask
;
615 chunk
*= geo
->near_copies
;
617 dev
= sector_div(stripe
, geo
->raid_disks
);
619 stripe
*= geo
->far_copies
;
621 sector
+= stripe
<< geo
->chunk_shift
;
623 /* and calculate all the others */
624 for (n
= 0; n
< geo
->near_copies
; n
++) {
628 r10bio
->devs
[slot
].devnum
= d
;
629 r10bio
->devs
[slot
].addr
= s
;
632 for (f
= 1; f
< geo
->far_copies
; f
++) {
633 set
= d
/ geo
->far_set_size
;
634 d
+= geo
->near_copies
;
636 if ((geo
->raid_disks
% geo
->far_set_size
) &&
637 (d
> last_far_set_start
)) {
638 d
-= last_far_set_start
;
639 d
%= last_far_set_size
;
640 d
+= last_far_set_start
;
642 d
%= geo
->far_set_size
;
643 d
+= geo
->far_set_size
* set
;
646 r10bio
->devs
[slot
].devnum
= d
;
647 r10bio
->devs
[slot
].addr
= s
;
651 if (dev
>= geo
->raid_disks
) {
653 sector
+= (geo
->chunk_mask
+ 1);
658 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
660 struct geom
*geo
= &conf
->geo
;
662 if (conf
->reshape_progress
!= MaxSector
&&
663 ((r10bio
->sector
>= conf
->reshape_progress
) !=
664 conf
->mddev
->reshape_backwards
)) {
665 set_bit(R10BIO_Previous
, &r10bio
->state
);
668 clear_bit(R10BIO_Previous
, &r10bio
->state
);
670 __raid10_find_phys(geo
, r10bio
);
673 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
675 sector_t offset
, chunk
, vchunk
;
676 /* Never use conf->prev as this is only called during resync
677 * or recovery, so reshape isn't happening
679 struct geom
*geo
= &conf
->geo
;
680 int far_set_start
= (dev
/ geo
->far_set_size
) * geo
->far_set_size
;
681 int far_set_size
= geo
->far_set_size
;
682 int last_far_set_start
;
684 if (geo
->raid_disks
% geo
->far_set_size
) {
685 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
686 last_far_set_start
*= geo
->far_set_size
;
688 if (dev
>= last_far_set_start
) {
689 far_set_size
= geo
->far_set_size
;
690 far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
691 far_set_start
= last_far_set_start
;
695 offset
= sector
& geo
->chunk_mask
;
696 if (geo
->far_offset
) {
698 chunk
= sector
>> geo
->chunk_shift
;
699 fc
= sector_div(chunk
, geo
->far_copies
);
700 dev
-= fc
* geo
->near_copies
;
701 if (dev
< far_set_start
)
704 while (sector
>= geo
->stride
) {
705 sector
-= geo
->stride
;
706 if (dev
< (geo
->near_copies
+ far_set_start
))
707 dev
+= far_set_size
- geo
->near_copies
;
709 dev
-= geo
->near_copies
;
711 chunk
= sector
>> geo
->chunk_shift
;
713 vchunk
= chunk
* geo
->raid_disks
+ dev
;
714 sector_div(vchunk
, geo
->near_copies
);
715 return (vchunk
<< geo
->chunk_shift
) + offset
;
719 * This routine returns the disk from which the requested read should
720 * be done. There is a per-array 'next expected sequential IO' sector
721 * number - if this matches on the next IO then we use the last disk.
722 * There is also a per-disk 'last know head position' sector that is
723 * maintained from IRQ contexts, both the normal and the resync IO
724 * completion handlers update this position correctly. If there is no
725 * perfect sequential match then we pick the disk whose head is closest.
727 * If there are 2 mirrors in the same 2 devices, performance degrades
728 * because position is mirror, not device based.
730 * The rdev for the device selected will have nr_pending incremented.
734 * FIXME: possibly should rethink readbalancing and do it differently
735 * depending on near_copies / far_copies geometry.
737 static struct md_rdev
*read_balance(struct r10conf
*conf
,
738 struct r10bio
*r10_bio
,
741 const sector_t this_sector
= r10_bio
->sector
;
743 int sectors
= r10_bio
->sectors
;
744 int best_good_sectors
;
745 sector_t new_distance
, best_dist
;
746 struct md_rdev
*best_rdev
, *rdev
= NULL
;
749 struct geom
*geo
= &conf
->geo
;
751 raid10_find_phys(conf
, r10_bio
);
755 best_dist
= MaxSector
;
756 best_good_sectors
= 0;
758 clear_bit(R10BIO_FailFast
, &r10_bio
->state
);
760 * Check if we can balance. We can balance on the whole
761 * device if no resync is going on (recovery is ok), or below
762 * the resync window. We take the first readable disk when
763 * above the resync window.
765 if ((conf
->mddev
->recovery_cp
< MaxSector
766 && (this_sector
+ sectors
>= conf
->next_resync
)) ||
767 (mddev_is_clustered(conf
->mddev
) &&
768 md_cluster_ops
->area_resyncing(conf
->mddev
, READ
, this_sector
,
769 this_sector
+ sectors
)))
772 for (slot
= 0; slot
< conf
->copies
; slot
++) {
777 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
779 disk
= r10_bio
->devs
[slot
].devnum
;
780 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
781 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
782 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
783 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
785 test_bit(Faulty
, &rdev
->flags
))
787 if (!test_bit(In_sync
, &rdev
->flags
) &&
788 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
791 dev_sector
= r10_bio
->devs
[slot
].addr
;
792 if (is_badblock(rdev
, dev_sector
, sectors
,
793 &first_bad
, &bad_sectors
)) {
794 if (best_dist
< MaxSector
)
795 /* Already have a better slot */
797 if (first_bad
<= dev_sector
) {
798 /* Cannot read here. If this is the
799 * 'primary' device, then we must not read
800 * beyond 'bad_sectors' from another device.
802 bad_sectors
-= (dev_sector
- first_bad
);
803 if (!do_balance
&& sectors
> bad_sectors
)
804 sectors
= bad_sectors
;
805 if (best_good_sectors
> sectors
)
806 best_good_sectors
= sectors
;
808 sector_t good_sectors
=
809 first_bad
- dev_sector
;
810 if (good_sectors
> best_good_sectors
) {
811 best_good_sectors
= good_sectors
;
816 /* Must read from here */
821 best_good_sectors
= sectors
;
827 /* At least 2 disks to choose from so failfast is OK */
828 set_bit(R10BIO_FailFast
, &r10_bio
->state
);
829 /* This optimisation is debatable, and completely destroys
830 * sequential read speed for 'far copies' arrays. So only
831 * keep it for 'near' arrays, and review those later.
833 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
836 /* for far > 1 always use the lowest address */
837 else if (geo
->far_copies
> 1)
838 new_distance
= r10_bio
->devs
[slot
].addr
;
840 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
841 conf
->mirrors
[disk
].head_position
);
842 if (new_distance
< best_dist
) {
843 best_dist
= new_distance
;
848 if (slot
>= conf
->copies
) {
854 atomic_inc(&rdev
->nr_pending
);
855 r10_bio
->read_slot
= slot
;
859 *max_sectors
= best_good_sectors
;
864 static int raid10_congested(struct mddev
*mddev
, int bits
)
866 struct r10conf
*conf
= mddev
->private;
869 if ((bits
& (1 << WB_async_congested
)) &&
870 conf
->pending_count
>= max_queued_requests
)
875 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
878 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
879 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
880 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
882 ret
|= bdi_congested(q
->backing_dev_info
, bits
);
889 static void flush_pending_writes(struct r10conf
*conf
)
891 /* Any writes that have been queued but are awaiting
892 * bitmap updates get flushed here.
894 spin_lock_irq(&conf
->device_lock
);
896 if (conf
->pending_bio_list
.head
) {
897 struct blk_plug plug
;
900 bio
= bio_list_get(&conf
->pending_bio_list
);
901 conf
->pending_count
= 0;
902 spin_unlock_irq(&conf
->device_lock
);
903 blk_start_plug(&plug
);
904 /* flush any pending bitmap writes to disk
905 * before proceeding w/ I/O */
906 bitmap_unplug(conf
->mddev
->bitmap
);
907 wake_up(&conf
->wait_barrier
);
909 while (bio
) { /* submit pending writes */
910 struct bio
*next
= bio
->bi_next
;
911 struct md_rdev
*rdev
= (void*)bio
->bi_disk
;
913 bio_set_dev(bio
, rdev
->bdev
);
914 if (test_bit(Faulty
, &rdev
->flags
)) {
916 } else if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
917 !blk_queue_discard(bio
->bi_disk
->queue
)))
921 generic_make_request(bio
);
924 blk_finish_plug(&plug
);
926 spin_unlock_irq(&conf
->device_lock
);
930 * Sometimes we need to suspend IO while we do something else,
931 * either some resync/recovery, or reconfigure the array.
932 * To do this we raise a 'barrier'.
933 * The 'barrier' is a counter that can be raised multiple times
934 * to count how many activities are happening which preclude
936 * We can only raise the barrier if there is no pending IO.
937 * i.e. if nr_pending == 0.
938 * We choose only to raise the barrier if no-one is waiting for the
939 * barrier to go down. This means that as soon as an IO request
940 * is ready, no other operations which require a barrier will start
941 * until the IO request has had a chance.
943 * So: regular IO calls 'wait_barrier'. When that returns there
944 * is no backgroup IO happening, It must arrange to call
945 * allow_barrier when it has finished its IO.
946 * backgroup IO calls must call raise_barrier. Once that returns
947 * there is no normal IO happeing. It must arrange to call
948 * lower_barrier when the particular background IO completes.
951 static void raise_barrier(struct r10conf
*conf
, int force
)
953 BUG_ON(force
&& !conf
->barrier
);
954 spin_lock_irq(&conf
->resync_lock
);
956 /* Wait until no block IO is waiting (unless 'force') */
957 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
960 /* block any new IO from starting */
963 /* Now wait for all pending IO to complete */
964 wait_event_lock_irq(conf
->wait_barrier
,
965 !atomic_read(&conf
->nr_pending
) && conf
->barrier
< RESYNC_DEPTH
,
968 spin_unlock_irq(&conf
->resync_lock
);
971 static void lower_barrier(struct r10conf
*conf
)
974 spin_lock_irqsave(&conf
->resync_lock
, flags
);
976 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
977 wake_up(&conf
->wait_barrier
);
980 static void wait_barrier(struct r10conf
*conf
)
982 spin_lock_irq(&conf
->resync_lock
);
985 /* Wait for the barrier to drop.
986 * However if there are already pending
987 * requests (preventing the barrier from
988 * rising completely), and the
989 * pre-process bio queue isn't empty,
990 * then don't wait, as we need to empty
991 * that queue to get the nr_pending
994 raid10_log(conf
->mddev
, "wait barrier");
995 wait_event_lock_irq(conf
->wait_barrier
,
997 (atomic_read(&conf
->nr_pending
) &&
999 (!bio_list_empty(¤t
->bio_list
[0]) ||
1000 !bio_list_empty(¤t
->bio_list
[1]))),
1003 if (!conf
->nr_waiting
)
1004 wake_up(&conf
->wait_barrier
);
1006 atomic_inc(&conf
->nr_pending
);
1007 spin_unlock_irq(&conf
->resync_lock
);
1010 static void allow_barrier(struct r10conf
*conf
)
1012 if ((atomic_dec_and_test(&conf
->nr_pending
)) ||
1013 (conf
->array_freeze_pending
))
1014 wake_up(&conf
->wait_barrier
);
1017 static void freeze_array(struct r10conf
*conf
, int extra
)
1019 /* stop syncio and normal IO and wait for everything to
1021 * We increment barrier and nr_waiting, and then
1022 * wait until nr_pending match nr_queued+extra
1023 * This is called in the context of one normal IO request
1024 * that has failed. Thus any sync request that might be pending
1025 * will be blocked by nr_pending, and we need to wait for
1026 * pending IO requests to complete or be queued for re-try.
1027 * Thus the number queued (nr_queued) plus this request (extra)
1028 * must match the number of pending IOs (nr_pending) before
1031 spin_lock_irq(&conf
->resync_lock
);
1032 conf
->array_freeze_pending
++;
1035 wait_event_lock_irq_cmd(conf
->wait_barrier
,
1036 atomic_read(&conf
->nr_pending
) == conf
->nr_queued
+extra
,
1038 flush_pending_writes(conf
));
1040 conf
->array_freeze_pending
--;
1041 spin_unlock_irq(&conf
->resync_lock
);
1044 static void unfreeze_array(struct r10conf
*conf
)
1046 /* reverse the effect of the freeze */
1047 spin_lock_irq(&conf
->resync_lock
);
1050 wake_up(&conf
->wait_barrier
);
1051 spin_unlock_irq(&conf
->resync_lock
);
1054 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1055 struct md_rdev
*rdev
)
1057 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1058 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1059 return rdev
->data_offset
;
1061 return rdev
->new_data_offset
;
1064 struct raid10_plug_cb
{
1065 struct blk_plug_cb cb
;
1066 struct bio_list pending
;
1070 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1072 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1074 struct mddev
*mddev
= plug
->cb
.data
;
1075 struct r10conf
*conf
= mddev
->private;
1078 if (from_schedule
|| current
->bio_list
) {
1079 spin_lock_irq(&conf
->device_lock
);
1080 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1081 conf
->pending_count
+= plug
->pending_cnt
;
1082 spin_unlock_irq(&conf
->device_lock
);
1083 wake_up(&conf
->wait_barrier
);
1084 md_wakeup_thread(mddev
->thread
);
1089 /* we aren't scheduling, so we can do the write-out directly. */
1090 bio
= bio_list_get(&plug
->pending
);
1091 bitmap_unplug(mddev
->bitmap
);
1092 wake_up(&conf
->wait_barrier
);
1094 while (bio
) { /* submit pending writes */
1095 struct bio
*next
= bio
->bi_next
;
1096 struct md_rdev
*rdev
= (void*)bio
->bi_disk
;
1097 bio
->bi_next
= NULL
;
1098 bio_set_dev(bio
, rdev
->bdev
);
1099 if (test_bit(Faulty
, &rdev
->flags
)) {
1101 } else if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
1102 !blk_queue_discard(bio
->bi_disk
->queue
)))
1103 /* Just ignore it */
1106 generic_make_request(bio
);
1112 static void raid10_read_request(struct mddev
*mddev
, struct bio
*bio
,
1113 struct r10bio
*r10_bio
)
1115 struct r10conf
*conf
= mddev
->private;
1116 struct bio
*read_bio
;
1117 const int op
= bio_op(bio
);
1118 const unsigned long do_sync
= (bio
->bi_opf
& REQ_SYNC
);
1121 struct md_rdev
*rdev
;
1122 char b
[BDEVNAME_SIZE
];
1123 int slot
= r10_bio
->read_slot
;
1124 struct md_rdev
*err_rdev
= NULL
;
1125 gfp_t gfp
= GFP_NOIO
;
1127 if (r10_bio
->devs
[slot
].rdev
) {
1129 * This is an error retry, but we cannot
1130 * safely dereference the rdev in the r10_bio,
1131 * we must use the one in conf.
1132 * If it has already been disconnected (unlikely)
1133 * we lose the device name in error messages.
1137 * As we are blocking raid10, it is a little safer to
1140 gfp
= GFP_NOIO
| __GFP_HIGH
;
1143 disk
= r10_bio
->devs
[slot
].devnum
;
1144 err_rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
1146 bdevname(err_rdev
->bdev
, b
);
1149 /* This never gets dereferenced */
1150 err_rdev
= r10_bio
->devs
[slot
].rdev
;
1155 * Register the new request and wait if the reconstruction
1156 * thread has put up a bar for new requests.
1157 * Continue immediately if no resync is active currently.
1161 sectors
= r10_bio
->sectors
;
1162 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1163 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1164 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1166 * IO spans the reshape position. Need to wait for reshape to
1169 raid10_log(conf
->mddev
, "wait reshape");
1170 allow_barrier(conf
);
1171 wait_event(conf
->wait_barrier
,
1172 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1173 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1178 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1181 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1183 (unsigned long long)r10_bio
->sector
);
1185 raid_end_bio_io(r10_bio
);
1189 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1191 bdevname(rdev
->bdev
, b
),
1192 (unsigned long long)r10_bio
->sector
);
1193 if (max_sectors
< bio_sectors(bio
)) {
1194 struct bio
*split
= bio_split(bio
, max_sectors
,
1195 gfp
, conf
->bio_split
);
1196 bio_chain(split
, bio
);
1197 generic_make_request(bio
);
1199 r10_bio
->master_bio
= bio
;
1200 r10_bio
->sectors
= max_sectors
;
1202 slot
= r10_bio
->read_slot
;
1204 read_bio
= bio_clone_fast(bio
, gfp
, mddev
->bio_set
);
1206 r10_bio
->devs
[slot
].bio
= read_bio
;
1207 r10_bio
->devs
[slot
].rdev
= rdev
;
1209 read_bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
+
1210 choose_data_offset(r10_bio
, rdev
);
1211 bio_set_dev(read_bio
, rdev
->bdev
);
1212 read_bio
->bi_end_io
= raid10_end_read_request
;
1213 bio_set_op_attrs(read_bio
, op
, do_sync
);
1214 if (test_bit(FailFast
, &rdev
->flags
) &&
1215 test_bit(R10BIO_FailFast
, &r10_bio
->state
))
1216 read_bio
->bi_opf
|= MD_FAILFAST
;
1217 read_bio
->bi_private
= r10_bio
;
1220 trace_block_bio_remap(read_bio
->bi_disk
->queue
,
1221 read_bio
, disk_devt(mddev
->gendisk
),
1223 generic_make_request(read_bio
);
1227 static void raid10_write_one_disk(struct mddev
*mddev
, struct r10bio
*r10_bio
,
1228 struct bio
*bio
, bool replacement
,
1231 const int op
= bio_op(bio
);
1232 const unsigned long do_sync
= (bio
->bi_opf
& REQ_SYNC
);
1233 const unsigned long do_fua
= (bio
->bi_opf
& REQ_FUA
);
1234 unsigned long flags
;
1235 struct blk_plug_cb
*cb
;
1236 struct raid10_plug_cb
*plug
= NULL
;
1237 struct r10conf
*conf
= mddev
->private;
1238 struct md_rdev
*rdev
;
1239 int devnum
= r10_bio
->devs
[n_copy
].devnum
;
1243 rdev
= conf
->mirrors
[devnum
].replacement
;
1245 /* Replacement just got moved to main 'rdev' */
1247 rdev
= conf
->mirrors
[devnum
].rdev
;
1250 rdev
= conf
->mirrors
[devnum
].rdev
;
1252 mbio
= bio_clone_fast(bio
, GFP_NOIO
, mddev
->bio_set
);
1254 r10_bio
->devs
[n_copy
].repl_bio
= mbio
;
1256 r10_bio
->devs
[n_copy
].bio
= mbio
;
1258 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[n_copy
].addr
+
1259 choose_data_offset(r10_bio
, rdev
));
1260 bio_set_dev(mbio
, rdev
->bdev
);
1261 mbio
->bi_end_io
= raid10_end_write_request
;
1262 bio_set_op_attrs(mbio
, op
, do_sync
| do_fua
);
1263 if (!replacement
&& test_bit(FailFast
,
1264 &conf
->mirrors
[devnum
].rdev
->flags
)
1265 && enough(conf
, devnum
))
1266 mbio
->bi_opf
|= MD_FAILFAST
;
1267 mbio
->bi_private
= r10_bio
;
1269 if (conf
->mddev
->gendisk
)
1270 trace_block_bio_remap(mbio
->bi_disk
->queue
,
1271 mbio
, disk_devt(conf
->mddev
->gendisk
),
1273 /* flush_pending_writes() needs access to the rdev so...*/
1274 mbio
->bi_disk
= (void *)rdev
;
1276 atomic_inc(&r10_bio
->remaining
);
1278 cb
= blk_check_plugged(raid10_unplug
, mddev
, sizeof(*plug
));
1280 plug
= container_of(cb
, struct raid10_plug_cb
, cb
);
1284 bio_list_add(&plug
->pending
, mbio
);
1285 plug
->pending_cnt
++;
1287 spin_lock_irqsave(&conf
->device_lock
, flags
);
1288 bio_list_add(&conf
->pending_bio_list
, mbio
);
1289 conf
->pending_count
++;
1290 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1291 md_wakeup_thread(mddev
->thread
);
1295 static void raid10_write_request(struct mddev
*mddev
, struct bio
*bio
,
1296 struct r10bio
*r10_bio
)
1298 struct r10conf
*conf
= mddev
->private;
1300 struct md_rdev
*blocked_rdev
;
1304 if ((mddev_is_clustered(mddev
) &&
1305 md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1306 bio
->bi_iter
.bi_sector
,
1307 bio_end_sector(bio
)))) {
1310 prepare_to_wait(&conf
->wait_barrier
,
1312 if (!md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1313 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
)))
1317 finish_wait(&conf
->wait_barrier
, &w
);
1321 * Register the new request and wait if the reconstruction
1322 * thread has put up a bar for new requests.
1323 * Continue immediately if no resync is active currently.
1327 sectors
= r10_bio
->sectors
;
1328 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1329 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1330 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1332 * IO spans the reshape position. Need to wait for reshape to
1335 raid10_log(conf
->mddev
, "wait reshape");
1336 allow_barrier(conf
);
1337 wait_event(conf
->wait_barrier
,
1338 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1339 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1344 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1345 (mddev
->reshape_backwards
1346 ? (bio
->bi_iter
.bi_sector
< conf
->reshape_safe
&&
1347 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
)
1348 : (bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_safe
&&
1349 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
))) {
1350 /* Need to update reshape_position in metadata */
1351 mddev
->reshape_position
= conf
->reshape_progress
;
1352 set_mask_bits(&mddev
->sb_flags
, 0,
1353 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
1354 md_wakeup_thread(mddev
->thread
);
1355 raid10_log(conf
->mddev
, "wait reshape metadata");
1356 wait_event(mddev
->sb_wait
,
1357 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
));
1359 conf
->reshape_safe
= mddev
->reshape_position
;
1362 if (conf
->pending_count
>= max_queued_requests
) {
1363 md_wakeup_thread(mddev
->thread
);
1364 raid10_log(mddev
, "wait queued");
1365 wait_event(conf
->wait_barrier
,
1366 conf
->pending_count
< max_queued_requests
);
1368 /* first select target devices under rcu_lock and
1369 * inc refcount on their rdev. Record them by setting
1371 * If there are known/acknowledged bad blocks on any device
1372 * on which we have seen a write error, we want to avoid
1373 * writing to those blocks. This potentially requires several
1374 * writes to write around the bad blocks. Each set of writes
1375 * gets its own r10_bio with a set of bios attached.
1378 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1379 raid10_find_phys(conf
, r10_bio
);
1381 blocked_rdev
= NULL
;
1383 max_sectors
= r10_bio
->sectors
;
1385 for (i
= 0; i
< conf
->copies
; i
++) {
1386 int d
= r10_bio
->devs
[i
].devnum
;
1387 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1388 struct md_rdev
*rrdev
= rcu_dereference(
1389 conf
->mirrors
[d
].replacement
);
1392 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1393 atomic_inc(&rdev
->nr_pending
);
1394 blocked_rdev
= rdev
;
1397 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1398 atomic_inc(&rrdev
->nr_pending
);
1399 blocked_rdev
= rrdev
;
1402 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)))
1404 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)))
1407 r10_bio
->devs
[i
].bio
= NULL
;
1408 r10_bio
->devs
[i
].repl_bio
= NULL
;
1410 if (!rdev
&& !rrdev
) {
1411 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1414 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1416 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1420 is_bad
= is_badblock(rdev
, dev_sector
, max_sectors
,
1421 &first_bad
, &bad_sectors
);
1423 /* Mustn't write here until the bad block
1426 atomic_inc(&rdev
->nr_pending
);
1427 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1428 blocked_rdev
= rdev
;
1431 if (is_bad
&& first_bad
<= dev_sector
) {
1432 /* Cannot write here at all */
1433 bad_sectors
-= (dev_sector
- first_bad
);
1434 if (bad_sectors
< max_sectors
)
1435 /* Mustn't write more than bad_sectors
1436 * to other devices yet
1438 max_sectors
= bad_sectors
;
1439 /* We don't set R10BIO_Degraded as that
1440 * only applies if the disk is missing,
1441 * so it might be re-added, and we want to
1442 * know to recover this chunk.
1443 * In this case the device is here, and the
1444 * fact that this chunk is not in-sync is
1445 * recorded in the bad block log.
1450 int good_sectors
= first_bad
- dev_sector
;
1451 if (good_sectors
< max_sectors
)
1452 max_sectors
= good_sectors
;
1456 r10_bio
->devs
[i
].bio
= bio
;
1457 atomic_inc(&rdev
->nr_pending
);
1460 r10_bio
->devs
[i
].repl_bio
= bio
;
1461 atomic_inc(&rrdev
->nr_pending
);
1466 if (unlikely(blocked_rdev
)) {
1467 /* Have to wait for this device to get unblocked, then retry */
1471 for (j
= 0; j
< i
; j
++) {
1472 if (r10_bio
->devs
[j
].bio
) {
1473 d
= r10_bio
->devs
[j
].devnum
;
1474 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1476 if (r10_bio
->devs
[j
].repl_bio
) {
1477 struct md_rdev
*rdev
;
1478 d
= r10_bio
->devs
[j
].devnum
;
1479 rdev
= conf
->mirrors
[d
].replacement
;
1481 /* Race with remove_disk */
1483 rdev
= conf
->mirrors
[d
].rdev
;
1485 rdev_dec_pending(rdev
, mddev
);
1488 allow_barrier(conf
);
1489 raid10_log(conf
->mddev
, "wait rdev %d blocked", blocked_rdev
->raid_disk
);
1490 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1495 if (max_sectors
< r10_bio
->sectors
)
1496 r10_bio
->sectors
= max_sectors
;
1498 if (r10_bio
->sectors
< bio_sectors(bio
)) {
1499 struct bio
*split
= bio_split(bio
, r10_bio
->sectors
,
1500 GFP_NOIO
, conf
->bio_split
);
1501 bio_chain(split
, bio
);
1502 generic_make_request(bio
);
1504 r10_bio
->master_bio
= bio
;
1507 atomic_set(&r10_bio
->remaining
, 1);
1508 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1510 for (i
= 0; i
< conf
->copies
; i
++) {
1511 if (r10_bio
->devs
[i
].bio
)
1512 raid10_write_one_disk(mddev
, r10_bio
, bio
, false, i
);
1513 if (r10_bio
->devs
[i
].repl_bio
)
1514 raid10_write_one_disk(mddev
, r10_bio
, bio
, true, i
);
1516 one_write_done(r10_bio
);
1519 static void __make_request(struct mddev
*mddev
, struct bio
*bio
, int sectors
)
1521 struct r10conf
*conf
= mddev
->private;
1522 struct r10bio
*r10_bio
;
1524 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1526 r10_bio
->master_bio
= bio
;
1527 r10_bio
->sectors
= sectors
;
1529 r10_bio
->mddev
= mddev
;
1530 r10_bio
->sector
= bio
->bi_iter
.bi_sector
;
1532 memset(r10_bio
->devs
, 0, sizeof(r10_bio
->devs
[0]) * conf
->copies
);
1534 if (bio_data_dir(bio
) == READ
)
1535 raid10_read_request(mddev
, bio
, r10_bio
);
1537 raid10_write_request(mddev
, bio
, r10_bio
);
1540 static bool raid10_make_request(struct mddev
*mddev
, struct bio
*bio
)
1542 struct r10conf
*conf
= mddev
->private;
1543 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1544 int chunk_sects
= chunk_mask
+ 1;
1545 int sectors
= bio_sectors(bio
);
1547 if (unlikely(bio
->bi_opf
& REQ_PREFLUSH
)) {
1548 md_flush_request(mddev
, bio
);
1552 if (!md_write_start(mddev
, bio
))
1556 * If this request crosses a chunk boundary, we need to split
1559 if (unlikely((bio
->bi_iter
.bi_sector
& chunk_mask
) +
1560 sectors
> chunk_sects
1561 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1562 || conf
->prev
.near_copies
<
1563 conf
->prev
.raid_disks
)))
1564 sectors
= chunk_sects
-
1565 (bio
->bi_iter
.bi_sector
&
1567 __make_request(mddev
, bio
, sectors
);
1569 /* In case raid10d snuck in to freeze_array */
1570 wake_up(&conf
->wait_barrier
);
1574 static void raid10_status(struct seq_file
*seq
, struct mddev
*mddev
)
1576 struct r10conf
*conf
= mddev
->private;
1579 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1580 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1581 if (conf
->geo
.near_copies
> 1)
1582 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1583 if (conf
->geo
.far_copies
> 1) {
1584 if (conf
->geo
.far_offset
)
1585 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1587 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1588 if (conf
->geo
.far_set_size
!= conf
->geo
.raid_disks
)
1589 seq_printf(seq
, " %d devices per set", conf
->geo
.far_set_size
);
1591 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1592 conf
->geo
.raid_disks
- mddev
->degraded
);
1594 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1595 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1596 seq_printf(seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1599 seq_printf(seq
, "]");
1602 /* check if there are enough drives for
1603 * every block to appear on atleast one.
1604 * Don't consider the device numbered 'ignore'
1605 * as we might be about to remove it.
1607 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1613 disks
= conf
->prev
.raid_disks
;
1614 ncopies
= conf
->prev
.near_copies
;
1616 disks
= conf
->geo
.raid_disks
;
1617 ncopies
= conf
->geo
.near_copies
;
1622 int n
= conf
->copies
;
1626 struct md_rdev
*rdev
;
1627 if (this != ignore
&&
1628 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1629 test_bit(In_sync
, &rdev
->flags
))
1631 this = (this+1) % disks
;
1635 first
= (first
+ ncopies
) % disks
;
1636 } while (first
!= 0);
1643 static int enough(struct r10conf
*conf
, int ignore
)
1645 /* when calling 'enough', both 'prev' and 'geo' must
1647 * This is ensured if ->reconfig_mutex or ->device_lock
1650 return _enough(conf
, 0, ignore
) &&
1651 _enough(conf
, 1, ignore
);
1654 static void raid10_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1656 char b
[BDEVNAME_SIZE
];
1657 struct r10conf
*conf
= mddev
->private;
1658 unsigned long flags
;
1661 * If it is not operational, then we have already marked it as dead
1662 * else if it is the last working disks, ignore the error, let the
1663 * next level up know.
1664 * else mark the drive as failed
1666 spin_lock_irqsave(&conf
->device_lock
, flags
);
1667 if (test_bit(In_sync
, &rdev
->flags
)
1668 && !enough(conf
, rdev
->raid_disk
)) {
1670 * Don't fail the drive, just return an IO error.
1672 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1675 if (test_and_clear_bit(In_sync
, &rdev
->flags
))
1678 * If recovery is running, make sure it aborts.
1680 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1681 set_bit(Blocked
, &rdev
->flags
);
1682 set_bit(Faulty
, &rdev
->flags
);
1683 set_mask_bits(&mddev
->sb_flags
, 0,
1684 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
1685 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1686 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1687 "md/raid10:%s: Operation continuing on %d devices.\n",
1688 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1689 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1692 static void print_conf(struct r10conf
*conf
)
1695 struct md_rdev
*rdev
;
1697 pr_debug("RAID10 conf printout:\n");
1699 pr_debug("(!conf)\n");
1702 pr_debug(" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1703 conf
->geo
.raid_disks
);
1705 /* This is only called with ->reconfix_mutex held, so
1706 * rcu protection of rdev is not needed */
1707 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1708 char b
[BDEVNAME_SIZE
];
1709 rdev
= conf
->mirrors
[i
].rdev
;
1711 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1712 i
, !test_bit(In_sync
, &rdev
->flags
),
1713 !test_bit(Faulty
, &rdev
->flags
),
1714 bdevname(rdev
->bdev
,b
));
1718 static void close_sync(struct r10conf
*conf
)
1721 allow_barrier(conf
);
1723 mempool_destroy(conf
->r10buf_pool
);
1724 conf
->r10buf_pool
= NULL
;
1727 static int raid10_spare_active(struct mddev
*mddev
)
1730 struct r10conf
*conf
= mddev
->private;
1731 struct raid10_info
*tmp
;
1733 unsigned long flags
;
1736 * Find all non-in_sync disks within the RAID10 configuration
1737 * and mark them in_sync
1739 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1740 tmp
= conf
->mirrors
+ i
;
1741 if (tmp
->replacement
1742 && tmp
->replacement
->recovery_offset
== MaxSector
1743 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1744 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1745 /* Replacement has just become active */
1747 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1750 /* Replaced device not technically faulty,
1751 * but we need to be sure it gets removed
1752 * and never re-added.
1754 set_bit(Faulty
, &tmp
->rdev
->flags
);
1755 sysfs_notify_dirent_safe(
1756 tmp
->rdev
->sysfs_state
);
1758 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1759 } else if (tmp
->rdev
1760 && tmp
->rdev
->recovery_offset
== MaxSector
1761 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1762 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1764 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1767 spin_lock_irqsave(&conf
->device_lock
, flags
);
1768 mddev
->degraded
-= count
;
1769 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1775 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1777 struct r10conf
*conf
= mddev
->private;
1781 int last
= conf
->geo
.raid_disks
- 1;
1783 if (mddev
->recovery_cp
< MaxSector
)
1784 /* only hot-add to in-sync arrays, as recovery is
1785 * very different from resync
1788 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1791 if (md_integrity_add_rdev(rdev
, mddev
))
1794 if (rdev
->raid_disk
>= 0)
1795 first
= last
= rdev
->raid_disk
;
1797 if (rdev
->saved_raid_disk
>= first
&&
1798 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1799 mirror
= rdev
->saved_raid_disk
;
1802 for ( ; mirror
<= last
; mirror
++) {
1803 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1804 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1807 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1808 p
->replacement
!= NULL
)
1810 clear_bit(In_sync
, &rdev
->flags
);
1811 set_bit(Replacement
, &rdev
->flags
);
1812 rdev
->raid_disk
= mirror
;
1815 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1816 rdev
->data_offset
<< 9);
1818 rcu_assign_pointer(p
->replacement
, rdev
);
1823 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1824 rdev
->data_offset
<< 9);
1826 p
->head_position
= 0;
1827 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1828 rdev
->raid_disk
= mirror
;
1830 if (rdev
->saved_raid_disk
!= mirror
)
1832 rcu_assign_pointer(p
->rdev
, rdev
);
1835 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1836 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1842 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1844 struct r10conf
*conf
= mddev
->private;
1846 int number
= rdev
->raid_disk
;
1847 struct md_rdev
**rdevp
;
1848 struct raid10_info
*p
= conf
->mirrors
+ number
;
1851 if (rdev
== p
->rdev
)
1853 else if (rdev
== p
->replacement
)
1854 rdevp
= &p
->replacement
;
1858 if (test_bit(In_sync
, &rdev
->flags
) ||
1859 atomic_read(&rdev
->nr_pending
)) {
1863 /* Only remove non-faulty devices if recovery
1866 if (!test_bit(Faulty
, &rdev
->flags
) &&
1867 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1868 (!p
->replacement
|| p
->replacement
== rdev
) &&
1869 number
< conf
->geo
.raid_disks
&&
1875 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
1877 if (atomic_read(&rdev
->nr_pending
)) {
1878 /* lost the race, try later */
1884 if (p
->replacement
) {
1885 /* We must have just cleared 'rdev' */
1886 p
->rdev
= p
->replacement
;
1887 clear_bit(Replacement
, &p
->replacement
->flags
);
1888 smp_mb(); /* Make sure other CPUs may see both as identical
1889 * but will never see neither -- if they are careful.
1891 p
->replacement
= NULL
;
1894 clear_bit(WantReplacement
, &rdev
->flags
);
1895 err
= md_integrity_register(mddev
);
1903 static void __end_sync_read(struct r10bio
*r10_bio
, struct bio
*bio
, int d
)
1905 struct r10conf
*conf
= r10_bio
->mddev
->private;
1907 if (!bio
->bi_status
)
1908 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1910 /* The write handler will notice the lack of
1911 * R10BIO_Uptodate and record any errors etc
1913 atomic_add(r10_bio
->sectors
,
1914 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1916 /* for reconstruct, we always reschedule after a read.
1917 * for resync, only after all reads
1919 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1920 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1921 atomic_dec_and_test(&r10_bio
->remaining
)) {
1922 /* we have read all the blocks,
1923 * do the comparison in process context in raid10d
1925 reschedule_retry(r10_bio
);
1929 static void end_sync_read(struct bio
*bio
)
1931 struct r10bio
*r10_bio
= get_resync_r10bio(bio
);
1932 struct r10conf
*conf
= r10_bio
->mddev
->private;
1933 int d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1935 __end_sync_read(r10_bio
, bio
, d
);
1938 static void end_reshape_read(struct bio
*bio
)
1940 /* reshape read bio isn't allocated from r10buf_pool */
1941 struct r10bio
*r10_bio
= bio
->bi_private
;
1943 __end_sync_read(r10_bio
, bio
, r10_bio
->read_slot
);
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
)
1975 struct r10bio
*r10_bio
= get_resync_r10bio(bio
);
1976 struct mddev
*mddev
= r10_bio
->mddev
;
1977 struct r10conf
*conf
= mddev
->private;
1983 struct md_rdev
*rdev
= NULL
;
1985 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1987 rdev
= conf
->mirrors
[d
].replacement
;
1989 rdev
= conf
->mirrors
[d
].rdev
;
1991 if (bio
->bi_status
) {
1993 md_error(mddev
, rdev
);
1995 set_bit(WriteErrorSeen
, &rdev
->flags
);
1996 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1997 set_bit(MD_RECOVERY_NEEDED
,
1998 &rdev
->mddev
->recovery
);
1999 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
2001 } else if (is_badblock(rdev
,
2002 r10_bio
->devs
[slot
].addr
,
2004 &first_bad
, &bad_sectors
))
2005 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
2007 rdev_dec_pending(rdev
, mddev
);
2009 end_sync_request(r10_bio
);
2013 * Note: sync and recover and handled very differently for raid10
2014 * This code is for resync.
2015 * For resync, we read through virtual addresses and read all blocks.
2016 * If there is any error, we schedule a write. The lowest numbered
2017 * drive is authoritative.
2018 * However requests come for physical address, so we need to map.
2019 * For every physical address there are raid_disks/copies virtual addresses,
2020 * which is always are least one, but is not necessarly an integer.
2021 * This means that a physical address can span multiple chunks, so we may
2022 * have to submit multiple io requests for a single sync request.
2025 * We check if all blocks are in-sync and only write to blocks that
2028 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2030 struct r10conf
*conf
= mddev
->private;
2032 struct bio
*tbio
, *fbio
;
2034 struct page
**tpages
, **fpages
;
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 (!r10_bio
->devs
[i
].bio
->bi_status
)
2043 if (i
== conf
->copies
)
2047 fbio
= r10_bio
->devs
[i
].bio
;
2048 fbio
->bi_iter
.bi_size
= r10_bio
->sectors
<< 9;
2049 fbio
->bi_iter
.bi_idx
= 0;
2050 fpages
= get_resync_pages(fbio
)->pages
;
2052 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2053 /* now find blocks with errors */
2054 for (i
=0 ; i
< conf
->copies
; i
++) {
2056 struct md_rdev
*rdev
;
2057 struct resync_pages
*rp
;
2059 tbio
= r10_bio
->devs
[i
].bio
;
2061 if (tbio
->bi_end_io
!= end_sync_read
)
2066 tpages
= get_resync_pages(tbio
)->pages
;
2067 d
= r10_bio
->devs
[i
].devnum
;
2068 rdev
= conf
->mirrors
[d
].rdev
;
2069 if (!r10_bio
->devs
[i
].bio
->bi_status
) {
2070 /* We know that the bi_io_vec layout is the same for
2071 * both 'first' and 'i', so we just compare them.
2072 * All vec entries are PAGE_SIZE;
2074 int sectors
= r10_bio
->sectors
;
2075 for (j
= 0; j
< vcnt
; j
++) {
2076 int len
= PAGE_SIZE
;
2077 if (sectors
< (len
/ 512))
2078 len
= sectors
* 512;
2079 if (memcmp(page_address(fpages
[j
]),
2080 page_address(tpages
[j
]),
2087 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2088 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2089 /* Don't fix anything. */
2091 } else if (test_bit(FailFast
, &rdev
->flags
)) {
2092 /* Just give up on this device */
2093 md_error(rdev
->mddev
, rdev
);
2096 /* Ok, we need to write this bio, either to correct an
2097 * inconsistency or to correct an unreadable block.
2098 * First we need to fixup bv_offset, bv_len and
2099 * bi_vecs, as the read request might have corrupted these
2101 rp
= get_resync_pages(tbio
);
2104 md_bio_reset_resync_pages(tbio
, rp
, fbio
->bi_iter
.bi_size
);
2106 rp
->raid_bio
= r10_bio
;
2107 tbio
->bi_private
= rp
;
2108 tbio
->bi_iter
.bi_sector
= r10_bio
->devs
[i
].addr
;
2109 tbio
->bi_end_io
= end_sync_write
;
2110 bio_set_op_attrs(tbio
, REQ_OP_WRITE
, 0);
2112 bio_copy_data(tbio
, fbio
);
2114 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2115 atomic_inc(&r10_bio
->remaining
);
2116 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2118 if (test_bit(FailFast
, &conf
->mirrors
[d
].rdev
->flags
))
2119 tbio
->bi_opf
|= MD_FAILFAST
;
2120 tbio
->bi_iter
.bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2121 bio_set_dev(tbio
, conf
->mirrors
[d
].rdev
->bdev
);
2122 generic_make_request(tbio
);
2125 /* Now write out to any replacement devices
2128 for (i
= 0; i
< conf
->copies
; i
++) {
2131 tbio
= r10_bio
->devs
[i
].repl_bio
;
2132 if (!tbio
|| !tbio
->bi_end_io
)
2134 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2135 && r10_bio
->devs
[i
].bio
!= fbio
)
2136 bio_copy_data(tbio
, fbio
);
2137 d
= r10_bio
->devs
[i
].devnum
;
2138 atomic_inc(&r10_bio
->remaining
);
2139 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2141 generic_make_request(tbio
);
2145 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2146 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2152 * Now for the recovery code.
2153 * Recovery happens across physical sectors.
2154 * We recover all non-is_sync drives by finding the virtual address of
2155 * each, and then choose a working drive that also has that virt address.
2156 * There is a separate r10_bio for each non-in_sync drive.
2157 * Only the first two slots are in use. The first for reading,
2158 * The second for writing.
2161 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2163 /* We got a read error during recovery.
2164 * We repeat the read in smaller page-sized sections.
2165 * If a read succeeds, write it to the new device or record
2166 * a bad block if we cannot.
2167 * If a read fails, record a bad block on both old and
2170 struct mddev
*mddev
= r10_bio
->mddev
;
2171 struct r10conf
*conf
= mddev
->private;
2172 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2174 int sectors
= r10_bio
->sectors
;
2176 int dr
= r10_bio
->devs
[0].devnum
;
2177 int dw
= r10_bio
->devs
[1].devnum
;
2178 struct page
**pages
= get_resync_pages(bio
)->pages
;
2182 struct md_rdev
*rdev
;
2186 if (s
> (PAGE_SIZE
>>9))
2189 rdev
= conf
->mirrors
[dr
].rdev
;
2190 addr
= r10_bio
->devs
[0].addr
+ sect
,
2191 ok
= sync_page_io(rdev
,
2195 REQ_OP_READ
, 0, false);
2197 rdev
= conf
->mirrors
[dw
].rdev
;
2198 addr
= r10_bio
->devs
[1].addr
+ sect
;
2199 ok
= sync_page_io(rdev
,
2203 REQ_OP_WRITE
, 0, false);
2205 set_bit(WriteErrorSeen
, &rdev
->flags
);
2206 if (!test_and_set_bit(WantReplacement
,
2208 set_bit(MD_RECOVERY_NEEDED
,
2209 &rdev
->mddev
->recovery
);
2213 /* We don't worry if we cannot set a bad block -
2214 * it really is bad so there is no loss in not
2217 rdev_set_badblocks(rdev
, addr
, s
, 0);
2219 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2220 /* need bad block on destination too */
2221 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2222 addr
= r10_bio
->devs
[1].addr
+ sect
;
2223 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2225 /* just abort the recovery */
2226 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2229 conf
->mirrors
[dw
].recovery_disabled
2230 = mddev
->recovery_disabled
;
2231 set_bit(MD_RECOVERY_INTR
,
2244 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2246 struct r10conf
*conf
= mddev
->private;
2248 struct bio
*wbio
, *wbio2
;
2250 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2251 fix_recovery_read_error(r10_bio
);
2252 end_sync_request(r10_bio
);
2257 * share the pages with the first bio
2258 * and submit the write request
2260 d
= r10_bio
->devs
[1].devnum
;
2261 wbio
= r10_bio
->devs
[1].bio
;
2262 wbio2
= r10_bio
->devs
[1].repl_bio
;
2263 /* Need to test wbio2->bi_end_io before we call
2264 * generic_make_request as if the former is NULL,
2265 * the latter is free to free wbio2.
2267 if (wbio2
&& !wbio2
->bi_end_io
)
2269 if (wbio
->bi_end_io
) {
2270 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2271 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2272 generic_make_request(wbio
);
2275 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2276 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2277 bio_sectors(wbio2
));
2278 generic_make_request(wbio2
);
2283 * Used by fix_read_error() to decay the per rdev read_errors.
2284 * We halve the read error count for every hour that has elapsed
2285 * since the last recorded read error.
2288 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2291 unsigned long hours_since_last
;
2292 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2294 cur_time_mon
= ktime_get_seconds();
2296 if (rdev
->last_read_error
== 0) {
2297 /* first time we've seen a read error */
2298 rdev
->last_read_error
= cur_time_mon
;
2302 hours_since_last
= (long)(cur_time_mon
-
2303 rdev
->last_read_error
) / 3600;
2305 rdev
->last_read_error
= cur_time_mon
;
2308 * if hours_since_last is > the number of bits in read_errors
2309 * just set read errors to 0. We do this to avoid
2310 * overflowing the shift of read_errors by hours_since_last.
2312 if (hours_since_last
>= 8 * sizeof(read_errors
))
2313 atomic_set(&rdev
->read_errors
, 0);
2315 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2318 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2319 int sectors
, struct page
*page
, int rw
)
2324 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2325 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2327 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, 0, false))
2331 set_bit(WriteErrorSeen
, &rdev
->flags
);
2332 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2333 set_bit(MD_RECOVERY_NEEDED
,
2334 &rdev
->mddev
->recovery
);
2336 /* need to record an error - either for the block or the device */
2337 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2338 md_error(rdev
->mddev
, rdev
);
2343 * This is a kernel thread which:
2345 * 1. Retries failed read operations on working mirrors.
2346 * 2. Updates the raid superblock when problems encounter.
2347 * 3. Performs writes following reads for array synchronising.
2350 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2352 int sect
= 0; /* Offset from r10_bio->sector */
2353 int sectors
= r10_bio
->sectors
;
2354 struct md_rdev
*rdev
;
2355 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2356 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2358 /* still own a reference to this rdev, so it cannot
2359 * have been cleared recently.
2361 rdev
= conf
->mirrors
[d
].rdev
;
2363 if (test_bit(Faulty
, &rdev
->flags
))
2364 /* drive has already been failed, just ignore any
2365 more fix_read_error() attempts */
2368 check_decay_read_errors(mddev
, rdev
);
2369 atomic_inc(&rdev
->read_errors
);
2370 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2371 char b
[BDEVNAME_SIZE
];
2372 bdevname(rdev
->bdev
, b
);
2374 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2376 atomic_read(&rdev
->read_errors
), max_read_errors
);
2377 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2379 md_error(mddev
, rdev
);
2380 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2386 int sl
= r10_bio
->read_slot
;
2390 if (s
> (PAGE_SIZE
>>9))
2398 d
= r10_bio
->devs
[sl
].devnum
;
2399 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2401 test_bit(In_sync
, &rdev
->flags
) &&
2402 !test_bit(Faulty
, &rdev
->flags
) &&
2403 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2404 &first_bad
, &bad_sectors
) == 0) {
2405 atomic_inc(&rdev
->nr_pending
);
2407 success
= sync_page_io(rdev
,
2408 r10_bio
->devs
[sl
].addr
+
2412 REQ_OP_READ
, 0, false);
2413 rdev_dec_pending(rdev
, mddev
);
2419 if (sl
== conf
->copies
)
2421 } while (!success
&& sl
!= r10_bio
->read_slot
);
2425 /* Cannot read from anywhere, just mark the block
2426 * as bad on the first device to discourage future
2429 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2430 rdev
= conf
->mirrors
[dn
].rdev
;
2432 if (!rdev_set_badblocks(
2434 r10_bio
->devs
[r10_bio
->read_slot
].addr
2437 md_error(mddev
, rdev
);
2438 r10_bio
->devs
[r10_bio
->read_slot
].bio
2445 /* write it back and re-read */
2447 while (sl
!= r10_bio
->read_slot
) {
2448 char b
[BDEVNAME_SIZE
];
2453 d
= r10_bio
->devs
[sl
].devnum
;
2454 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2456 test_bit(Faulty
, &rdev
->flags
) ||
2457 !test_bit(In_sync
, &rdev
->flags
))
2460 atomic_inc(&rdev
->nr_pending
);
2462 if (r10_sync_page_io(rdev
,
2463 r10_bio
->devs
[sl
].addr
+
2465 s
, conf
->tmppage
, WRITE
)
2467 /* Well, this device is dead */
2468 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2470 (unsigned long long)(
2472 choose_data_offset(r10_bio
,
2474 bdevname(rdev
->bdev
, b
));
2475 pr_notice("md/raid10:%s: %s: failing drive\n",
2477 bdevname(rdev
->bdev
, b
));
2479 rdev_dec_pending(rdev
, mddev
);
2483 while (sl
!= r10_bio
->read_slot
) {
2484 char b
[BDEVNAME_SIZE
];
2489 d
= r10_bio
->devs
[sl
].devnum
;
2490 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2492 test_bit(Faulty
, &rdev
->flags
) ||
2493 !test_bit(In_sync
, &rdev
->flags
))
2496 atomic_inc(&rdev
->nr_pending
);
2498 switch (r10_sync_page_io(rdev
,
2499 r10_bio
->devs
[sl
].addr
+
2504 /* Well, this device is dead */
2505 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2507 (unsigned long long)(
2509 choose_data_offset(r10_bio
, rdev
)),
2510 bdevname(rdev
->bdev
, b
));
2511 pr_notice("md/raid10:%s: %s: failing drive\n",
2513 bdevname(rdev
->bdev
, b
));
2516 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2518 (unsigned long long)(
2520 choose_data_offset(r10_bio
, rdev
)),
2521 bdevname(rdev
->bdev
, b
));
2522 atomic_add(s
, &rdev
->corrected_errors
);
2525 rdev_dec_pending(rdev
, mddev
);
2535 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2537 struct bio
*bio
= r10_bio
->master_bio
;
2538 struct mddev
*mddev
= r10_bio
->mddev
;
2539 struct r10conf
*conf
= mddev
->private;
2540 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2541 /* bio has the data to be written to slot 'i' where
2542 * we just recently had a write error.
2543 * We repeatedly clone the bio and trim down to one block,
2544 * then try the write. Where the write fails we record
2546 * It is conceivable that the bio doesn't exactly align with
2547 * blocks. We must handle this.
2549 * We currently own a reference to the rdev.
2555 int sect_to_write
= r10_bio
->sectors
;
2558 if (rdev
->badblocks
.shift
< 0)
2561 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2562 bdev_logical_block_size(rdev
->bdev
) >> 9);
2563 sector
= r10_bio
->sector
;
2564 sectors
= ((r10_bio
->sector
+ block_sectors
)
2565 & ~(sector_t
)(block_sectors
- 1))
2568 while (sect_to_write
) {
2571 if (sectors
> sect_to_write
)
2572 sectors
= sect_to_write
;
2573 /* Write at 'sector' for 'sectors' */
2574 wbio
= bio_clone_fast(bio
, GFP_NOIO
, mddev
->bio_set
);
2575 bio_trim(wbio
, sector
- bio
->bi_iter
.bi_sector
, sectors
);
2576 wsector
= r10_bio
->devs
[i
].addr
+ (sector
- r10_bio
->sector
);
2577 wbio
->bi_iter
.bi_sector
= wsector
+
2578 choose_data_offset(r10_bio
, rdev
);
2579 bio_set_dev(wbio
, rdev
->bdev
);
2580 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2582 if (submit_bio_wait(wbio
) < 0)
2584 ok
= rdev_set_badblocks(rdev
, wsector
,
2589 sect_to_write
-= sectors
;
2591 sectors
= block_sectors
;
2596 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2598 int slot
= r10_bio
->read_slot
;
2600 struct r10conf
*conf
= mddev
->private;
2601 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2603 /* we got a read error. Maybe the drive is bad. Maybe just
2604 * the block and we can fix it.
2605 * We freeze all other IO, and try reading the block from
2606 * other devices. When we find one, we re-write
2607 * and check it that fixes the read error.
2608 * This is all done synchronously while the array is
2611 bio
= r10_bio
->devs
[slot
].bio
;
2613 r10_bio
->devs
[slot
].bio
= NULL
;
2616 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2617 else if (!test_bit(FailFast
, &rdev
->flags
)) {
2618 freeze_array(conf
, 1);
2619 fix_read_error(conf
, mddev
, r10_bio
);
2620 unfreeze_array(conf
);
2622 md_error(mddev
, rdev
);
2624 rdev_dec_pending(rdev
, mddev
);
2625 allow_barrier(conf
);
2627 raid10_read_request(mddev
, r10_bio
->master_bio
, r10_bio
);
2630 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2632 /* Some sort of write request has finished and it
2633 * succeeded in writing where we thought there was a
2634 * bad block. So forget the bad block.
2635 * Or possibly if failed and we need to record
2639 struct md_rdev
*rdev
;
2641 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2642 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2643 for (m
= 0; m
< conf
->copies
; m
++) {
2644 int dev
= r10_bio
->devs
[m
].devnum
;
2645 rdev
= conf
->mirrors
[dev
].rdev
;
2646 if (r10_bio
->devs
[m
].bio
== NULL
)
2648 if (!r10_bio
->devs
[m
].bio
->bi_status
) {
2649 rdev_clear_badblocks(
2651 r10_bio
->devs
[m
].addr
,
2652 r10_bio
->sectors
, 0);
2654 if (!rdev_set_badblocks(
2656 r10_bio
->devs
[m
].addr
,
2657 r10_bio
->sectors
, 0))
2658 md_error(conf
->mddev
, rdev
);
2660 rdev
= conf
->mirrors
[dev
].replacement
;
2661 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2664 if (!r10_bio
->devs
[m
].repl_bio
->bi_status
) {
2665 rdev_clear_badblocks(
2667 r10_bio
->devs
[m
].addr
,
2668 r10_bio
->sectors
, 0);
2670 if (!rdev_set_badblocks(
2672 r10_bio
->devs
[m
].addr
,
2673 r10_bio
->sectors
, 0))
2674 md_error(conf
->mddev
, rdev
);
2680 for (m
= 0; m
< conf
->copies
; m
++) {
2681 int dev
= r10_bio
->devs
[m
].devnum
;
2682 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2683 rdev
= conf
->mirrors
[dev
].rdev
;
2684 if (bio
== IO_MADE_GOOD
) {
2685 rdev_clear_badblocks(
2687 r10_bio
->devs
[m
].addr
,
2688 r10_bio
->sectors
, 0);
2689 rdev_dec_pending(rdev
, conf
->mddev
);
2690 } else if (bio
!= NULL
&& bio
->bi_status
) {
2692 if (!narrow_write_error(r10_bio
, m
)) {
2693 md_error(conf
->mddev
, rdev
);
2694 set_bit(R10BIO_Degraded
,
2697 rdev_dec_pending(rdev
, conf
->mddev
);
2699 bio
= r10_bio
->devs
[m
].repl_bio
;
2700 rdev
= conf
->mirrors
[dev
].replacement
;
2701 if (rdev
&& bio
== IO_MADE_GOOD
) {
2702 rdev_clear_badblocks(
2704 r10_bio
->devs
[m
].addr
,
2705 r10_bio
->sectors
, 0);
2706 rdev_dec_pending(rdev
, conf
->mddev
);
2710 spin_lock_irq(&conf
->device_lock
);
2711 list_add(&r10_bio
->retry_list
, &conf
->bio_end_io_list
);
2713 spin_unlock_irq(&conf
->device_lock
);
2715 * In case freeze_array() is waiting for condition
2716 * nr_pending == nr_queued + extra to be true.
2718 wake_up(&conf
->wait_barrier
);
2719 md_wakeup_thread(conf
->mddev
->thread
);
2721 if (test_bit(R10BIO_WriteError
,
2723 close_write(r10_bio
);
2724 raid_end_bio_io(r10_bio
);
2729 static void raid10d(struct md_thread
*thread
)
2731 struct mddev
*mddev
= thread
->mddev
;
2732 struct r10bio
*r10_bio
;
2733 unsigned long flags
;
2734 struct r10conf
*conf
= mddev
->private;
2735 struct list_head
*head
= &conf
->retry_list
;
2736 struct blk_plug plug
;
2738 md_check_recovery(mddev
);
2740 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2741 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
2743 spin_lock_irqsave(&conf
->device_lock
, flags
);
2744 if (!test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
2745 while (!list_empty(&conf
->bio_end_io_list
)) {
2746 list_move(conf
->bio_end_io_list
.prev
, &tmp
);
2750 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2751 while (!list_empty(&tmp
)) {
2752 r10_bio
= list_first_entry(&tmp
, struct r10bio
,
2754 list_del(&r10_bio
->retry_list
);
2755 if (mddev
->degraded
)
2756 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
2758 if (test_bit(R10BIO_WriteError
,
2760 close_write(r10_bio
);
2761 raid_end_bio_io(r10_bio
);
2765 blk_start_plug(&plug
);
2768 flush_pending_writes(conf
);
2770 spin_lock_irqsave(&conf
->device_lock
, flags
);
2771 if (list_empty(head
)) {
2772 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2775 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2776 list_del(head
->prev
);
2778 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2780 mddev
= r10_bio
->mddev
;
2781 conf
= mddev
->private;
2782 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2783 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2784 handle_write_completed(conf
, r10_bio
);
2785 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2786 reshape_request_write(mddev
, r10_bio
);
2787 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2788 sync_request_write(mddev
, r10_bio
);
2789 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2790 recovery_request_write(mddev
, r10_bio
);
2791 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2792 handle_read_error(mddev
, r10_bio
);
2797 if (mddev
->sb_flags
& ~(1<<MD_SB_CHANGE_PENDING
))
2798 md_check_recovery(mddev
);
2800 blk_finish_plug(&plug
);
2803 static int init_resync(struct r10conf
*conf
)
2808 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2809 BUG_ON(conf
->r10buf_pool
);
2810 conf
->have_replacement
= 0;
2811 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2812 if (conf
->mirrors
[i
].replacement
)
2813 conf
->have_replacement
= 1;
2814 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2815 if (!conf
->r10buf_pool
)
2817 conf
->next_resync
= 0;
2821 static struct r10bio
*raid10_alloc_init_r10buf(struct r10conf
*conf
)
2823 struct r10bio
*r10bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2824 struct rsync_pages
*rp
;
2829 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
) ||
2830 test_bit(MD_RECOVERY_RESHAPE
, &conf
->mddev
->recovery
))
2831 nalloc
= conf
->copies
; /* resync */
2833 nalloc
= 2; /* recovery */
2835 for (i
= 0; i
< nalloc
; i
++) {
2836 bio
= r10bio
->devs
[i
].bio
;
2837 rp
= bio
->bi_private
;
2839 bio
->bi_private
= rp
;
2840 bio
= r10bio
->devs
[i
].repl_bio
;
2842 rp
= bio
->bi_private
;
2844 bio
->bi_private
= rp
;
2851 * Set cluster_sync_high since we need other nodes to add the
2852 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2854 static void raid10_set_cluster_sync_high(struct r10conf
*conf
)
2856 sector_t window_size
;
2857 int extra_chunk
, chunks
;
2860 * First, here we define "stripe" as a unit which across
2861 * all member devices one time, so we get chunks by use
2862 * raid_disks / near_copies. Otherwise, if near_copies is
2863 * close to raid_disks, then resync window could increases
2864 * linearly with the increase of raid_disks, which means
2865 * we will suspend a really large IO window while it is not
2866 * necessary. If raid_disks is not divisible by near_copies,
2867 * an extra chunk is needed to ensure the whole "stripe" is
2871 chunks
= conf
->geo
.raid_disks
/ conf
->geo
.near_copies
;
2872 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
== 0)
2876 window_size
= (chunks
+ extra_chunk
) * conf
->mddev
->chunk_sectors
;
2879 * At least use a 32M window to align with raid1's resync window
2881 window_size
= (CLUSTER_RESYNC_WINDOW_SECTORS
> window_size
) ?
2882 CLUSTER_RESYNC_WINDOW_SECTORS
: window_size
;
2884 conf
->cluster_sync_high
= conf
->cluster_sync_low
+ window_size
;
2888 * perform a "sync" on one "block"
2890 * We need to make sure that no normal I/O request - particularly write
2891 * requests - conflict with active sync requests.
2893 * This is achieved by tracking pending requests and a 'barrier' concept
2894 * that can be installed to exclude normal IO requests.
2896 * Resync and recovery are handled very differently.
2897 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2899 * For resync, we iterate over virtual addresses, read all copies,
2900 * and update if there are differences. If only one copy is live,
2902 * For recovery, we iterate over physical addresses, read a good
2903 * value for each non-in_sync drive, and over-write.
2905 * So, for recovery we may have several outstanding complex requests for a
2906 * given address, one for each out-of-sync device. We model this by allocating
2907 * a number of r10_bio structures, one for each out-of-sync device.
2908 * As we setup these structures, we collect all bio's together into a list
2909 * which we then process collectively to add pages, and then process again
2910 * to pass to generic_make_request.
2912 * The r10_bio structures are linked using a borrowed master_bio pointer.
2913 * This link is counted in ->remaining. When the r10_bio that points to NULL
2914 * has its remaining count decremented to 0, the whole complex operation
2919 static sector_t
raid10_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2922 struct r10conf
*conf
= mddev
->private;
2923 struct r10bio
*r10_bio
;
2924 struct bio
*biolist
= NULL
, *bio
;
2925 sector_t max_sector
, nr_sectors
;
2928 sector_t sync_blocks
;
2929 sector_t sectors_skipped
= 0;
2930 int chunks_skipped
= 0;
2931 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2934 if (!conf
->r10buf_pool
)
2935 if (init_resync(conf
))
2939 * Allow skipping a full rebuild for incremental assembly
2940 * of a clean array, like RAID1 does.
2942 if (mddev
->bitmap
== NULL
&&
2943 mddev
->recovery_cp
== MaxSector
&&
2944 mddev
->reshape_position
== MaxSector
&&
2945 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2946 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2947 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2948 conf
->fullsync
== 0) {
2950 return mddev
->dev_sectors
- sector_nr
;
2954 max_sector
= mddev
->dev_sectors
;
2955 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2956 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2957 max_sector
= mddev
->resync_max_sectors
;
2958 if (sector_nr
>= max_sector
) {
2959 conf
->cluster_sync_low
= 0;
2960 conf
->cluster_sync_high
= 0;
2962 /* If we aborted, we need to abort the
2963 * sync on the 'current' bitmap chucks (there can
2964 * be several when recovering multiple devices).
2965 * as we may have started syncing it but not finished.
2966 * We can find the current address in
2967 * mddev->curr_resync, but for recovery,
2968 * we need to convert that to several
2969 * virtual addresses.
2971 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2977 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2978 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2979 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2981 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2983 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2984 bitmap_end_sync(mddev
->bitmap
, sect
,
2988 /* completed sync */
2989 if ((!mddev
->bitmap
|| conf
->fullsync
)
2990 && conf
->have_replacement
2991 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2992 /* Completed a full sync so the replacements
2993 * are now fully recovered.
2996 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2997 struct md_rdev
*rdev
=
2998 rcu_dereference(conf
->mirrors
[i
].replacement
);
3000 rdev
->recovery_offset
= MaxSector
;
3006 bitmap_close_sync(mddev
->bitmap
);
3009 return sectors_skipped
;
3012 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
3013 return reshape_request(mddev
, sector_nr
, skipped
);
3015 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
3016 /* if there has been nothing to do on any drive,
3017 * then there is nothing to do at all..
3020 return (max_sector
- sector_nr
) + sectors_skipped
;
3023 if (max_sector
> mddev
->resync_max
)
3024 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
3026 /* make sure whole request will fit in a chunk - if chunks
3029 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
3030 max_sector
> (sector_nr
| chunk_mask
))
3031 max_sector
= (sector_nr
| chunk_mask
) + 1;
3034 * If there is non-resync activity waiting for a turn, then let it
3035 * though before starting on this new sync request.
3037 if (conf
->nr_waiting
)
3038 schedule_timeout_uninterruptible(1);
3040 /* Again, very different code for resync and recovery.
3041 * Both must result in an r10bio with a list of bios that
3042 * have bi_end_io, bi_sector, bi_disk set,
3043 * and bi_private set to the r10bio.
3044 * For recovery, we may actually create several r10bios
3045 * with 2 bios in each, that correspond to the bios in the main one.
3046 * In this case, the subordinate r10bios link back through a
3047 * borrowed master_bio pointer, and the counter in the master
3048 * includes a ref from each subordinate.
3050 /* First, we decide what to do and set ->bi_end_io
3051 * To end_sync_read if we want to read, and
3052 * end_sync_write if we will want to write.
3055 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
3056 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3057 /* recovery... the complicated one */
3061 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
3067 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
3068 struct md_rdev
*mrdev
, *mreplace
;
3071 mrdev
= rcu_dereference(mirror
->rdev
);
3072 mreplace
= rcu_dereference(mirror
->replacement
);
3074 if ((mrdev
== NULL
||
3075 test_bit(Faulty
, &mrdev
->flags
) ||
3076 test_bit(In_sync
, &mrdev
->flags
)) &&
3077 (mreplace
== NULL
||
3078 test_bit(Faulty
, &mreplace
->flags
))) {
3084 /* want to reconstruct this device */
3086 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3087 if (sect
>= mddev
->resync_max_sectors
) {
3088 /* last stripe is not complete - don't
3089 * try to recover this sector.
3094 if (mreplace
&& test_bit(Faulty
, &mreplace
->flags
))
3096 /* Unless we are doing a full sync, or a replacement
3097 * we only need to recover the block if it is set in
3100 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3102 if (sync_blocks
< max_sync
)
3103 max_sync
= sync_blocks
;
3107 /* yep, skip the sync_blocks here, but don't assume
3108 * that there will never be anything to do here
3110 chunks_skipped
= -1;
3114 atomic_inc(&mrdev
->nr_pending
);
3116 atomic_inc(&mreplace
->nr_pending
);
3119 r10_bio
= raid10_alloc_init_r10buf(conf
);
3121 raise_barrier(conf
, rb2
!= NULL
);
3122 atomic_set(&r10_bio
->remaining
, 0);
3124 r10_bio
->master_bio
= (struct bio
*)rb2
;
3126 atomic_inc(&rb2
->remaining
);
3127 r10_bio
->mddev
= mddev
;
3128 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3129 r10_bio
->sector
= sect
;
3131 raid10_find_phys(conf
, r10_bio
);
3133 /* Need to check if the array will still be
3137 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++) {
3138 struct md_rdev
*rdev
= rcu_dereference(
3139 conf
->mirrors
[j
].rdev
);
3140 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3146 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3147 &sync_blocks
, still_degraded
);
3150 for (j
=0; j
<conf
->copies
;j
++) {
3152 int d
= r10_bio
->devs
[j
].devnum
;
3153 sector_t from_addr
, to_addr
;
3154 struct md_rdev
*rdev
=
3155 rcu_dereference(conf
->mirrors
[d
].rdev
);
3156 sector_t sector
, first_bad
;
3159 !test_bit(In_sync
, &rdev
->flags
))
3161 /* This is where we read from */
3163 sector
= r10_bio
->devs
[j
].addr
;
3165 if (is_badblock(rdev
, sector
, max_sync
,
3166 &first_bad
, &bad_sectors
)) {
3167 if (first_bad
> sector
)
3168 max_sync
= first_bad
- sector
;
3170 bad_sectors
-= (sector
3172 if (max_sync
> bad_sectors
)
3173 max_sync
= bad_sectors
;
3177 bio
= r10_bio
->devs
[0].bio
;
3178 bio
->bi_next
= biolist
;
3180 bio
->bi_end_io
= end_sync_read
;
3181 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3182 if (test_bit(FailFast
, &rdev
->flags
))
3183 bio
->bi_opf
|= MD_FAILFAST
;
3184 from_addr
= r10_bio
->devs
[j
].addr
;
3185 bio
->bi_iter
.bi_sector
= from_addr
+
3187 bio_set_dev(bio
, rdev
->bdev
);
3188 atomic_inc(&rdev
->nr_pending
);
3189 /* and we write to 'i' (if not in_sync) */
3191 for (k
=0; k
<conf
->copies
; k
++)
3192 if (r10_bio
->devs
[k
].devnum
== i
)
3194 BUG_ON(k
== conf
->copies
);
3195 to_addr
= r10_bio
->devs
[k
].addr
;
3196 r10_bio
->devs
[0].devnum
= d
;
3197 r10_bio
->devs
[0].addr
= from_addr
;
3198 r10_bio
->devs
[1].devnum
= i
;
3199 r10_bio
->devs
[1].addr
= to_addr
;
3201 if (!test_bit(In_sync
, &mrdev
->flags
)) {
3202 bio
= r10_bio
->devs
[1].bio
;
3203 bio
->bi_next
= biolist
;
3205 bio
->bi_end_io
= end_sync_write
;
3206 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3207 bio
->bi_iter
.bi_sector
= to_addr
3208 + mrdev
->data_offset
;
3209 bio_set_dev(bio
, mrdev
->bdev
);
3210 atomic_inc(&r10_bio
->remaining
);
3212 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3214 /* and maybe write to replacement */
3215 bio
= r10_bio
->devs
[1].repl_bio
;
3217 bio
->bi_end_io
= NULL
;
3218 /* Note: if mreplace != NULL, then bio
3219 * cannot be NULL as r10buf_pool_alloc will
3220 * have allocated it.
3221 * So the second test here is pointless.
3222 * But it keeps semantic-checkers happy, and
3223 * this comment keeps human reviewers
3226 if (mreplace
== NULL
|| bio
== NULL
||
3227 test_bit(Faulty
, &mreplace
->flags
))
3229 bio
->bi_next
= biolist
;
3231 bio
->bi_end_io
= end_sync_write
;
3232 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3233 bio
->bi_iter
.bi_sector
= to_addr
+
3234 mreplace
->data_offset
;
3235 bio_set_dev(bio
, mreplace
->bdev
);
3236 atomic_inc(&r10_bio
->remaining
);
3240 if (j
== conf
->copies
) {
3241 /* Cannot recover, so abort the recovery or
3242 * record a bad block */
3244 /* problem is that there are bad blocks
3245 * on other device(s)
3248 for (k
= 0; k
< conf
->copies
; k
++)
3249 if (r10_bio
->devs
[k
].devnum
== i
)
3251 if (!test_bit(In_sync
,
3253 && !rdev_set_badblocks(
3255 r10_bio
->devs
[k
].addr
,
3259 !rdev_set_badblocks(
3261 r10_bio
->devs
[k
].addr
,
3266 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3268 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3270 mirror
->recovery_disabled
3271 = mddev
->recovery_disabled
;
3275 atomic_dec(&rb2
->remaining
);
3277 rdev_dec_pending(mrdev
, mddev
);
3279 rdev_dec_pending(mreplace
, mddev
);
3282 rdev_dec_pending(mrdev
, mddev
);
3284 rdev_dec_pending(mreplace
, mddev
);
3285 if (r10_bio
->devs
[0].bio
->bi_opf
& MD_FAILFAST
) {
3286 /* Only want this if there is elsewhere to
3287 * read from. 'j' is currently the first
3291 for (; j
< conf
->copies
; j
++) {
3292 int d
= r10_bio
->devs
[j
].devnum
;
3293 if (conf
->mirrors
[d
].rdev
&&
3295 &conf
->mirrors
[d
].rdev
->flags
))
3299 r10_bio
->devs
[0].bio
->bi_opf
3303 if (biolist
== NULL
) {
3305 struct r10bio
*rb2
= r10_bio
;
3306 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3307 rb2
->master_bio
= NULL
;
3313 /* resync. Schedule a read for every block at this virt offset */
3317 * Since curr_resync_completed could probably not update in
3318 * time, and we will set cluster_sync_low based on it.
3319 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3320 * safety reason, which ensures curr_resync_completed is
3321 * updated in bitmap_cond_end_sync.
3323 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
,
3324 mddev_is_clustered(mddev
) &&
3325 (sector_nr
+ 2 * RESYNC_SECTORS
>
3326 conf
->cluster_sync_high
));
3328 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3329 &sync_blocks
, mddev
->degraded
) &&
3330 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3331 &mddev
->recovery
)) {
3332 /* We can skip this block */
3334 return sync_blocks
+ sectors_skipped
;
3336 if (sync_blocks
< max_sync
)
3337 max_sync
= sync_blocks
;
3338 r10_bio
= raid10_alloc_init_r10buf(conf
);
3341 r10_bio
->mddev
= mddev
;
3342 atomic_set(&r10_bio
->remaining
, 0);
3343 raise_barrier(conf
, 0);
3344 conf
->next_resync
= sector_nr
;
3346 r10_bio
->master_bio
= NULL
;
3347 r10_bio
->sector
= sector_nr
;
3348 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3349 raid10_find_phys(conf
, r10_bio
);
3350 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3352 for (i
= 0; i
< conf
->copies
; i
++) {
3353 int d
= r10_bio
->devs
[i
].devnum
;
3354 sector_t first_bad
, sector
;
3356 struct md_rdev
*rdev
;
3358 if (r10_bio
->devs
[i
].repl_bio
)
3359 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3361 bio
= r10_bio
->devs
[i
].bio
;
3362 bio
->bi_status
= BLK_STS_IOERR
;
3364 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
3365 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3369 sector
= r10_bio
->devs
[i
].addr
;
3370 if (is_badblock(rdev
, sector
, max_sync
,
3371 &first_bad
, &bad_sectors
)) {
3372 if (first_bad
> sector
)
3373 max_sync
= first_bad
- sector
;
3375 bad_sectors
-= (sector
- first_bad
);
3376 if (max_sync
> bad_sectors
)
3377 max_sync
= bad_sectors
;
3382 atomic_inc(&rdev
->nr_pending
);
3383 atomic_inc(&r10_bio
->remaining
);
3384 bio
->bi_next
= biolist
;
3386 bio
->bi_end_io
= end_sync_read
;
3387 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3388 if (test_bit(FailFast
, &rdev
->flags
))
3389 bio
->bi_opf
|= MD_FAILFAST
;
3390 bio
->bi_iter
.bi_sector
= sector
+ rdev
->data_offset
;
3391 bio_set_dev(bio
, rdev
->bdev
);
3394 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
3395 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3399 atomic_inc(&rdev
->nr_pending
);
3401 /* Need to set up for writing to the replacement */
3402 bio
= r10_bio
->devs
[i
].repl_bio
;
3403 bio
->bi_status
= BLK_STS_IOERR
;
3405 sector
= r10_bio
->devs
[i
].addr
;
3406 bio
->bi_next
= biolist
;
3408 bio
->bi_end_io
= end_sync_write
;
3409 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3410 if (test_bit(FailFast
, &rdev
->flags
))
3411 bio
->bi_opf
|= MD_FAILFAST
;
3412 bio
->bi_iter
.bi_sector
= sector
+ rdev
->data_offset
;
3413 bio_set_dev(bio
, rdev
->bdev
);
3419 for (i
=0; i
<conf
->copies
; i
++) {
3420 int d
= r10_bio
->devs
[i
].devnum
;
3421 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3422 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3424 if (r10_bio
->devs
[i
].repl_bio
&&
3425 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3427 conf
->mirrors
[d
].replacement
,
3437 if (sector_nr
+ max_sync
< max_sector
)
3438 max_sector
= sector_nr
+ max_sync
;
3441 int len
= PAGE_SIZE
;
3442 if (sector_nr
+ (len
>>9) > max_sector
)
3443 len
= (max_sector
- sector_nr
) << 9;
3446 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3447 struct resync_pages
*rp
= get_resync_pages(bio
);
3448 page
= resync_fetch_page(rp
, page_idx
);
3450 * won't fail because the vec table is big enough
3451 * to hold all these pages
3453 bio_add_page(bio
, page
, len
, 0);
3455 nr_sectors
+= len
>>9;
3456 sector_nr
+= len
>>9;
3457 } while (++page_idx
< RESYNC_PAGES
);
3458 r10_bio
->sectors
= nr_sectors
;
3460 if (mddev_is_clustered(mddev
) &&
3461 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3462 /* It is resync not recovery */
3463 if (conf
->cluster_sync_high
< sector_nr
+ nr_sectors
) {
3464 conf
->cluster_sync_low
= mddev
->curr_resync_completed
;
3465 raid10_set_cluster_sync_high(conf
);
3466 /* Send resync message */
3467 md_cluster_ops
->resync_info_update(mddev
,
3468 conf
->cluster_sync_low
,
3469 conf
->cluster_sync_high
);
3471 } else if (mddev_is_clustered(mddev
)) {
3472 /* This is recovery not resync */
3473 sector_t sect_va1
, sect_va2
;
3474 bool broadcast_msg
= false;
3476 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3478 * sector_nr is a device address for recovery, so we
3479 * need translate it to array address before compare
3480 * with cluster_sync_high.
3482 sect_va1
= raid10_find_virt(conf
, sector_nr
, i
);
3484 if (conf
->cluster_sync_high
< sect_va1
+ nr_sectors
) {
3485 broadcast_msg
= true;
3487 * curr_resync_completed is similar as
3488 * sector_nr, so make the translation too.
3490 sect_va2
= raid10_find_virt(conf
,
3491 mddev
->curr_resync_completed
, i
);
3493 if (conf
->cluster_sync_low
== 0 ||
3494 conf
->cluster_sync_low
> sect_va2
)
3495 conf
->cluster_sync_low
= sect_va2
;
3498 if (broadcast_msg
) {
3499 raid10_set_cluster_sync_high(conf
);
3500 md_cluster_ops
->resync_info_update(mddev
,
3501 conf
->cluster_sync_low
,
3502 conf
->cluster_sync_high
);
3508 biolist
= biolist
->bi_next
;
3510 bio
->bi_next
= NULL
;
3511 r10_bio
= get_resync_r10bio(bio
);
3512 r10_bio
->sectors
= nr_sectors
;
3514 if (bio
->bi_end_io
== end_sync_read
) {
3515 md_sync_acct_bio(bio
, nr_sectors
);
3517 generic_make_request(bio
);
3521 if (sectors_skipped
)
3522 /* pretend they weren't skipped, it makes
3523 * no important difference in this case
3525 md_done_sync(mddev
, sectors_skipped
, 1);
3527 return sectors_skipped
+ nr_sectors
;
3529 /* There is nowhere to write, so all non-sync
3530 * drives must be failed or in resync, all drives
3531 * have a bad block, so try the next chunk...
3533 if (sector_nr
+ max_sync
< max_sector
)
3534 max_sector
= sector_nr
+ max_sync
;
3536 sectors_skipped
+= (max_sector
- sector_nr
);
3538 sector_nr
= max_sector
;
3543 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3546 struct r10conf
*conf
= mddev
->private;
3549 raid_disks
= min(conf
->geo
.raid_disks
,
3550 conf
->prev
.raid_disks
);
3552 sectors
= conf
->dev_sectors
;
3554 size
= sectors
>> conf
->geo
.chunk_shift
;
3555 sector_div(size
, conf
->geo
.far_copies
);
3556 size
= size
* raid_disks
;
3557 sector_div(size
, conf
->geo
.near_copies
);
3559 return size
<< conf
->geo
.chunk_shift
;
3562 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3564 /* Calculate the number of sectors-per-device that will
3565 * actually be used, and set conf->dev_sectors and
3569 size
= size
>> conf
->geo
.chunk_shift
;
3570 sector_div(size
, conf
->geo
.far_copies
);
3571 size
= size
* conf
->geo
.raid_disks
;
3572 sector_div(size
, conf
->geo
.near_copies
);
3573 /* 'size' is now the number of chunks in the array */
3574 /* calculate "used chunks per device" */
3575 size
= size
* conf
->copies
;
3577 /* We need to round up when dividing by raid_disks to
3578 * get the stride size.
3580 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3582 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3584 if (conf
->geo
.far_offset
)
3585 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3587 sector_div(size
, conf
->geo
.far_copies
);
3588 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3592 enum geo_type
{geo_new
, geo_old
, geo_start
};
3593 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3596 int layout
, chunk
, disks
;
3599 layout
= mddev
->layout
;
3600 chunk
= mddev
->chunk_sectors
;
3601 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3604 layout
= mddev
->new_layout
;
3605 chunk
= mddev
->new_chunk_sectors
;
3606 disks
= mddev
->raid_disks
;
3608 default: /* avoid 'may be unused' warnings */
3609 case geo_start
: /* new when starting reshape - raid_disks not
3611 layout
= mddev
->new_layout
;
3612 chunk
= mddev
->new_chunk_sectors
;
3613 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3618 if (chunk
< (PAGE_SIZE
>> 9) ||
3619 !is_power_of_2(chunk
))
3622 fc
= (layout
>> 8) & 255;
3623 fo
= layout
& (1<<16);
3624 geo
->raid_disks
= disks
;
3625 geo
->near_copies
= nc
;
3626 geo
->far_copies
= fc
;
3627 geo
->far_offset
= fo
;
3628 switch (layout
>> 17) {
3629 case 0: /* original layout. simple but not always optimal */
3630 geo
->far_set_size
= disks
;
3632 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3633 * actually using this, but leave code here just in case.*/
3634 geo
->far_set_size
= disks
/fc
;
3635 WARN(geo
->far_set_size
< fc
,
3636 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3638 case 2: /* "improved" layout fixed to match documentation */
3639 geo
->far_set_size
= fc
* nc
;
3641 default: /* Not a valid layout */
3644 geo
->chunk_mask
= chunk
- 1;
3645 geo
->chunk_shift
= ffz(~chunk
);
3649 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3651 struct r10conf
*conf
= NULL
;
3656 copies
= setup_geo(&geo
, mddev
, geo_new
);
3659 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3660 mdname(mddev
), PAGE_SIZE
);
3664 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3665 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3666 mdname(mddev
), mddev
->new_layout
);
3671 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3675 /* FIXME calc properly */
3676 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3677 max(0,-mddev
->delta_disks
)),
3682 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3687 conf
->copies
= copies
;
3688 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3689 r10bio_pool_free
, conf
);
3690 if (!conf
->r10bio_pool
)
3693 conf
->bio_split
= bioset_create(BIO_POOL_SIZE
, 0, 0);
3694 if (!conf
->bio_split
)
3697 calc_sectors(conf
, mddev
->dev_sectors
);
3698 if (mddev
->reshape_position
== MaxSector
) {
3699 conf
->prev
= conf
->geo
;
3700 conf
->reshape_progress
= MaxSector
;
3702 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3706 conf
->reshape_progress
= mddev
->reshape_position
;
3707 if (conf
->prev
.far_offset
)
3708 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3710 /* far_copies must be 1 */
3711 conf
->prev
.stride
= conf
->dev_sectors
;
3713 conf
->reshape_safe
= conf
->reshape_progress
;
3714 spin_lock_init(&conf
->device_lock
);
3715 INIT_LIST_HEAD(&conf
->retry_list
);
3716 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
3718 spin_lock_init(&conf
->resync_lock
);
3719 init_waitqueue_head(&conf
->wait_barrier
);
3720 atomic_set(&conf
->nr_pending
, 0);
3722 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3726 conf
->mddev
= mddev
;
3731 mempool_destroy(conf
->r10bio_pool
);
3732 kfree(conf
->mirrors
);
3733 safe_put_page(conf
->tmppage
);
3734 if (conf
->bio_split
)
3735 bioset_free(conf
->bio_split
);
3738 return ERR_PTR(err
);
3741 static int raid10_run(struct mddev
*mddev
)
3743 struct r10conf
*conf
;
3744 int i
, disk_idx
, chunk_size
;
3745 struct raid10_info
*disk
;
3746 struct md_rdev
*rdev
;
3748 sector_t min_offset_diff
= 0;
3750 bool discard_supported
= false;
3752 if (mddev_init_writes_pending(mddev
) < 0)
3755 if (mddev
->private == NULL
) {
3756 conf
= setup_conf(mddev
);
3758 return PTR_ERR(conf
);
3759 mddev
->private = conf
;
3761 conf
= mddev
->private;
3765 if (mddev_is_clustered(conf
->mddev
)) {
3768 fc
= (mddev
->layout
>> 8) & 255;
3769 fo
= mddev
->layout
& (1<<16);
3770 if (fc
> 1 || fo
> 0) {
3771 pr_err("only near layout is supported by clustered"
3777 mddev
->thread
= conf
->thread
;
3778 conf
->thread
= NULL
;
3780 chunk_size
= mddev
->chunk_sectors
<< 9;
3782 blk_queue_max_discard_sectors(mddev
->queue
,
3783 mddev
->chunk_sectors
);
3784 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3785 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
3786 blk_queue_io_min(mddev
->queue
, chunk_size
);
3787 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3788 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3790 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3791 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3794 rdev_for_each(rdev
, mddev
) {
3797 disk_idx
= rdev
->raid_disk
;
3800 if (disk_idx
>= conf
->geo
.raid_disks
&&
3801 disk_idx
>= conf
->prev
.raid_disks
)
3803 disk
= conf
->mirrors
+ disk_idx
;
3805 if (test_bit(Replacement
, &rdev
->flags
)) {
3806 if (disk
->replacement
)
3808 disk
->replacement
= rdev
;
3814 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3815 if (!mddev
->reshape_backwards
)
3819 if (first
|| diff
< min_offset_diff
)
3820 min_offset_diff
= diff
;
3823 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3824 rdev
->data_offset
<< 9);
3826 disk
->head_position
= 0;
3828 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3829 discard_supported
= true;
3834 if (discard_supported
)
3835 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3838 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3841 /* need to check that every block has at least one working mirror */
3842 if (!enough(conf
, -1)) {
3843 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3848 if (conf
->reshape_progress
!= MaxSector
) {
3849 /* must ensure that shape change is supported */
3850 if (conf
->geo
.far_copies
!= 1 &&
3851 conf
->geo
.far_offset
== 0)
3853 if (conf
->prev
.far_copies
!= 1 &&
3854 conf
->prev
.far_offset
== 0)
3858 mddev
->degraded
= 0;
3860 i
< conf
->geo
.raid_disks
3861 || i
< conf
->prev
.raid_disks
;
3864 disk
= conf
->mirrors
+ i
;
3866 if (!disk
->rdev
&& disk
->replacement
) {
3867 /* The replacement is all we have - use it */
3868 disk
->rdev
= disk
->replacement
;
3869 disk
->replacement
= NULL
;
3870 clear_bit(Replacement
, &disk
->rdev
->flags
);
3874 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3875 disk
->head_position
= 0;
3878 disk
->rdev
->saved_raid_disk
< 0)
3881 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3884 if (mddev
->recovery_cp
!= MaxSector
)
3885 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3887 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3888 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3889 conf
->geo
.raid_disks
);
3891 * Ok, everything is just fine now
3893 mddev
->dev_sectors
= conf
->dev_sectors
;
3894 size
= raid10_size(mddev
, 0, 0);
3895 md_set_array_sectors(mddev
, size
);
3896 mddev
->resync_max_sectors
= size
;
3897 set_bit(MD_FAILFAST_SUPPORTED
, &mddev
->flags
);
3900 int stripe
= conf
->geo
.raid_disks
*
3901 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3903 /* Calculate max read-ahead size.
3904 * We need to readahead at least twice a whole stripe....
3907 stripe
/= conf
->geo
.near_copies
;
3908 if (mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
3909 mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
3912 if (md_integrity_register(mddev
))
3915 if (conf
->reshape_progress
!= MaxSector
) {
3916 unsigned long before_length
, after_length
;
3918 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3919 conf
->prev
.far_copies
);
3920 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3921 conf
->geo
.far_copies
);
3923 if (max(before_length
, after_length
) > min_offset_diff
) {
3924 /* This cannot work */
3925 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3928 conf
->offset_diff
= min_offset_diff
;
3930 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3931 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3932 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3933 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3934 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3941 md_unregister_thread(&mddev
->thread
);
3942 mempool_destroy(conf
->r10bio_pool
);
3943 safe_put_page(conf
->tmppage
);
3944 kfree(conf
->mirrors
);
3946 mddev
->private = NULL
;
3951 static void raid10_free(struct mddev
*mddev
, void *priv
)
3953 struct r10conf
*conf
= priv
;
3955 mempool_destroy(conf
->r10bio_pool
);
3956 safe_put_page(conf
->tmppage
);
3957 kfree(conf
->mirrors
);
3958 kfree(conf
->mirrors_old
);
3959 kfree(conf
->mirrors_new
);
3960 if (conf
->bio_split
)
3961 bioset_free(conf
->bio_split
);
3965 static void raid10_quiesce(struct mddev
*mddev
, int quiesce
)
3967 struct r10conf
*conf
= mddev
->private;
3970 raise_barrier(conf
, 0);
3972 lower_barrier(conf
);
3975 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3977 /* Resize of 'far' arrays is not supported.
3978 * For 'near' and 'offset' arrays we can set the
3979 * number of sectors used to be an appropriate multiple
3980 * of the chunk size.
3981 * For 'offset', this is far_copies*chunksize.
3982 * For 'near' the multiplier is the LCM of
3983 * near_copies and raid_disks.
3984 * So if far_copies > 1 && !far_offset, fail.
3985 * Else find LCM(raid_disks, near_copy)*far_copies and
3986 * multiply by chunk_size. Then round to this number.
3987 * This is mostly done by raid10_size()
3989 struct r10conf
*conf
= mddev
->private;
3990 sector_t oldsize
, size
;
3992 if (mddev
->reshape_position
!= MaxSector
)
3995 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3998 oldsize
= raid10_size(mddev
, 0, 0);
3999 size
= raid10_size(mddev
, sectors
, 0);
4000 if (mddev
->external_size
&&
4001 mddev
->array_sectors
> size
)
4003 if (mddev
->bitmap
) {
4004 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
4008 md_set_array_sectors(mddev
, size
);
4009 if (sectors
> mddev
->dev_sectors
&&
4010 mddev
->recovery_cp
> oldsize
) {
4011 mddev
->recovery_cp
= oldsize
;
4012 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4014 calc_sectors(conf
, sectors
);
4015 mddev
->dev_sectors
= conf
->dev_sectors
;
4016 mddev
->resync_max_sectors
= size
;
4020 static void *raid10_takeover_raid0(struct mddev
*mddev
, sector_t size
, int devs
)
4022 struct md_rdev
*rdev
;
4023 struct r10conf
*conf
;
4025 if (mddev
->degraded
> 0) {
4026 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4028 return ERR_PTR(-EINVAL
);
4030 sector_div(size
, devs
);
4032 /* Set new parameters */
4033 mddev
->new_level
= 10;
4034 /* new layout: far_copies = 1, near_copies = 2 */
4035 mddev
->new_layout
= (1<<8) + 2;
4036 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
4037 mddev
->delta_disks
= mddev
->raid_disks
;
4038 mddev
->raid_disks
*= 2;
4039 /* make sure it will be not marked as dirty */
4040 mddev
->recovery_cp
= MaxSector
;
4041 mddev
->dev_sectors
= size
;
4043 conf
= setup_conf(mddev
);
4044 if (!IS_ERR(conf
)) {
4045 rdev_for_each(rdev
, mddev
)
4046 if (rdev
->raid_disk
>= 0) {
4047 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
4048 rdev
->sectors
= size
;
4056 static void *raid10_takeover(struct mddev
*mddev
)
4058 struct r0conf
*raid0_conf
;
4060 /* raid10 can take over:
4061 * raid0 - providing it has only two drives
4063 if (mddev
->level
== 0) {
4064 /* for raid0 takeover only one zone is supported */
4065 raid0_conf
= mddev
->private;
4066 if (raid0_conf
->nr_strip_zones
> 1) {
4067 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4069 return ERR_PTR(-EINVAL
);
4071 return raid10_takeover_raid0(mddev
,
4072 raid0_conf
->strip_zone
->zone_end
,
4073 raid0_conf
->strip_zone
->nb_dev
);
4075 return ERR_PTR(-EINVAL
);
4078 static int raid10_check_reshape(struct mddev
*mddev
)
4080 /* Called when there is a request to change
4081 * - layout (to ->new_layout)
4082 * - chunk size (to ->new_chunk_sectors)
4083 * - raid_disks (by delta_disks)
4084 * or when trying to restart a reshape that was ongoing.
4086 * We need to validate the request and possibly allocate
4087 * space if that might be an issue later.
4089 * Currently we reject any reshape of a 'far' mode array,
4090 * allow chunk size to change if new is generally acceptable,
4091 * allow raid_disks to increase, and allow
4092 * a switch between 'near' mode and 'offset' mode.
4094 struct r10conf
*conf
= mddev
->private;
4097 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
4100 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
4101 /* mustn't change number of copies */
4103 if (geo
.far_copies
> 1 && !geo
.far_offset
)
4104 /* Cannot switch to 'far' mode */
4107 if (mddev
->array_sectors
& geo
.chunk_mask
)
4108 /* not factor of array size */
4111 if (!enough(conf
, -1))
4114 kfree(conf
->mirrors_new
);
4115 conf
->mirrors_new
= NULL
;
4116 if (mddev
->delta_disks
> 0) {
4117 /* allocate new 'mirrors' list */
4118 conf
->mirrors_new
= kzalloc(
4119 sizeof(struct raid10_info
)
4120 *(mddev
->raid_disks
+
4121 mddev
->delta_disks
),
4123 if (!conf
->mirrors_new
)
4130 * Need to check if array has failed when deciding whether to:
4132 * - remove non-faulty devices
4135 * This determination is simple when no reshape is happening.
4136 * However if there is a reshape, we need to carefully check
4137 * both the before and after sections.
4138 * This is because some failed devices may only affect one
4139 * of the two sections, and some non-in_sync devices may
4140 * be insync in the section most affected by failed devices.
4142 static int calc_degraded(struct r10conf
*conf
)
4144 int degraded
, degraded2
;
4149 /* 'prev' section first */
4150 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4151 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4152 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4154 else if (!test_bit(In_sync
, &rdev
->flags
))
4155 /* When we can reduce the number of devices in
4156 * an array, this might not contribute to
4157 * 'degraded'. It does now.
4162 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4166 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4167 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4168 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4170 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4171 /* If reshape is increasing the number of devices,
4172 * this section has already been recovered, so
4173 * it doesn't contribute to degraded.
4176 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4181 if (degraded2
> degraded
)
4186 static int raid10_start_reshape(struct mddev
*mddev
)
4188 /* A 'reshape' has been requested. This commits
4189 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4190 * This also checks if there are enough spares and adds them
4192 * We currently require enough spares to make the final
4193 * array non-degraded. We also require that the difference
4194 * between old and new data_offset - on each device - is
4195 * enough that we never risk over-writing.
4198 unsigned long before_length
, after_length
;
4199 sector_t min_offset_diff
= 0;
4202 struct r10conf
*conf
= mddev
->private;
4203 struct md_rdev
*rdev
;
4207 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4210 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4213 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4214 conf
->prev
.far_copies
);
4215 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4216 conf
->geo
.far_copies
);
4218 rdev_for_each(rdev
, mddev
) {
4219 if (!test_bit(In_sync
, &rdev
->flags
)
4220 && !test_bit(Faulty
, &rdev
->flags
))
4222 if (rdev
->raid_disk
>= 0) {
4223 long long diff
= (rdev
->new_data_offset
4224 - rdev
->data_offset
);
4225 if (!mddev
->reshape_backwards
)
4229 if (first
|| diff
< min_offset_diff
)
4230 min_offset_diff
= diff
;
4235 if (max(before_length
, after_length
) > min_offset_diff
)
4238 if (spares
< mddev
->delta_disks
)
4241 conf
->offset_diff
= min_offset_diff
;
4242 spin_lock_irq(&conf
->device_lock
);
4243 if (conf
->mirrors_new
) {
4244 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4245 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4247 kfree(conf
->mirrors_old
);
4248 conf
->mirrors_old
= conf
->mirrors
;
4249 conf
->mirrors
= conf
->mirrors_new
;
4250 conf
->mirrors_new
= NULL
;
4252 setup_geo(&conf
->geo
, mddev
, geo_start
);
4254 if (mddev
->reshape_backwards
) {
4255 sector_t size
= raid10_size(mddev
, 0, 0);
4256 if (size
< mddev
->array_sectors
) {
4257 spin_unlock_irq(&conf
->device_lock
);
4258 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4262 mddev
->resync_max_sectors
= size
;
4263 conf
->reshape_progress
= size
;
4265 conf
->reshape_progress
= 0;
4266 conf
->reshape_safe
= conf
->reshape_progress
;
4267 spin_unlock_irq(&conf
->device_lock
);
4269 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4270 ret
= bitmap_resize(mddev
->bitmap
,
4271 raid10_size(mddev
, 0,
4272 conf
->geo
.raid_disks
),
4277 if (mddev
->delta_disks
> 0) {
4278 rdev_for_each(rdev
, mddev
)
4279 if (rdev
->raid_disk
< 0 &&
4280 !test_bit(Faulty
, &rdev
->flags
)) {
4281 if (raid10_add_disk(mddev
, rdev
) == 0) {
4282 if (rdev
->raid_disk
>=
4283 conf
->prev
.raid_disks
)
4284 set_bit(In_sync
, &rdev
->flags
);
4286 rdev
->recovery_offset
= 0;
4288 if (sysfs_link_rdev(mddev
, rdev
))
4289 /* Failure here is OK */;
4291 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4292 && !test_bit(Faulty
, &rdev
->flags
)) {
4293 /* This is a spare that was manually added */
4294 set_bit(In_sync
, &rdev
->flags
);
4297 /* When a reshape changes the number of devices,
4298 * ->degraded is measured against the larger of the
4299 * pre and post numbers.
4301 spin_lock_irq(&conf
->device_lock
);
4302 mddev
->degraded
= calc_degraded(conf
);
4303 spin_unlock_irq(&conf
->device_lock
);
4304 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4305 mddev
->reshape_position
= conf
->reshape_progress
;
4306 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
4308 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4309 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4310 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
4311 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4312 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4314 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4316 if (!mddev
->sync_thread
) {
4320 conf
->reshape_checkpoint
= jiffies
;
4321 md_wakeup_thread(mddev
->sync_thread
);
4322 md_new_event(mddev
);
4326 mddev
->recovery
= 0;
4327 spin_lock_irq(&conf
->device_lock
);
4328 conf
->geo
= conf
->prev
;
4329 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4330 rdev_for_each(rdev
, mddev
)
4331 rdev
->new_data_offset
= rdev
->data_offset
;
4333 conf
->reshape_progress
= MaxSector
;
4334 conf
->reshape_safe
= MaxSector
;
4335 mddev
->reshape_position
= MaxSector
;
4336 spin_unlock_irq(&conf
->device_lock
);
4340 /* Calculate the last device-address that could contain
4341 * any block from the chunk that includes the array-address 's'
4342 * and report the next address.
4343 * i.e. the address returned will be chunk-aligned and after
4344 * any data that is in the chunk containing 's'.
4346 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4348 s
= (s
| geo
->chunk_mask
) + 1;
4349 s
>>= geo
->chunk_shift
;
4350 s
*= geo
->near_copies
;
4351 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4352 s
*= geo
->far_copies
;
4353 s
<<= geo
->chunk_shift
;
4357 /* Calculate the first device-address that could contain
4358 * any block from the chunk that includes the array-address 's'.
4359 * This too will be the start of a chunk
4361 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4363 s
>>= geo
->chunk_shift
;
4364 s
*= geo
->near_copies
;
4365 sector_div(s
, geo
->raid_disks
);
4366 s
*= geo
->far_copies
;
4367 s
<<= geo
->chunk_shift
;
4371 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4374 /* We simply copy at most one chunk (smallest of old and new)
4375 * at a time, possibly less if that exceeds RESYNC_PAGES,
4376 * or we hit a bad block or something.
4377 * This might mean we pause for normal IO in the middle of
4378 * a chunk, but that is not a problem as mddev->reshape_position
4379 * can record any location.
4381 * If we will want to write to a location that isn't
4382 * yet recorded as 'safe' (i.e. in metadata on disk) then
4383 * we need to flush all reshape requests and update the metadata.
4385 * When reshaping forwards (e.g. to more devices), we interpret
4386 * 'safe' as the earliest block which might not have been copied
4387 * down yet. We divide this by previous stripe size and multiply
4388 * by previous stripe length to get lowest device offset that we
4389 * cannot write to yet.
4390 * We interpret 'sector_nr' as an address that we want to write to.
4391 * From this we use last_device_address() to find where we might
4392 * write to, and first_device_address on the 'safe' position.
4393 * If this 'next' write position is after the 'safe' position,
4394 * we must update the metadata to increase the 'safe' position.
4396 * When reshaping backwards, we round in the opposite direction
4397 * and perform the reverse test: next write position must not be
4398 * less than current safe position.
4400 * In all this the minimum difference in data offsets
4401 * (conf->offset_diff - always positive) allows a bit of slack,
4402 * so next can be after 'safe', but not by more than offset_diff
4404 * We need to prepare all the bios here before we start any IO
4405 * to ensure the size we choose is acceptable to all devices.
4406 * The means one for each copy for write-out and an extra one for
4408 * We store the read-in bio in ->master_bio and the others in
4409 * ->devs[x].bio and ->devs[x].repl_bio.
4411 struct r10conf
*conf
= mddev
->private;
4412 struct r10bio
*r10_bio
;
4413 sector_t next
, safe
, last
;
4417 struct md_rdev
*rdev
;
4420 struct bio
*bio
, *read_bio
;
4421 int sectors_done
= 0;
4422 struct page
**pages
;
4424 if (sector_nr
== 0) {
4425 /* If restarting in the middle, skip the initial sectors */
4426 if (mddev
->reshape_backwards
&&
4427 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4428 sector_nr
= (raid10_size(mddev
, 0, 0)
4429 - conf
->reshape_progress
);
4430 } else if (!mddev
->reshape_backwards
&&
4431 conf
->reshape_progress
> 0)
4432 sector_nr
= conf
->reshape_progress
;
4434 mddev
->curr_resync_completed
= sector_nr
;
4435 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4441 /* We don't use sector_nr to track where we are up to
4442 * as that doesn't work well for ->reshape_backwards.
4443 * So just use ->reshape_progress.
4445 if (mddev
->reshape_backwards
) {
4446 /* 'next' is the earliest device address that we might
4447 * write to for this chunk in the new layout
4449 next
= first_dev_address(conf
->reshape_progress
- 1,
4452 /* 'safe' is the last device address that we might read from
4453 * in the old layout after a restart
4455 safe
= last_dev_address(conf
->reshape_safe
- 1,
4458 if (next
+ conf
->offset_diff
< safe
)
4461 last
= conf
->reshape_progress
- 1;
4462 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4463 & conf
->prev
.chunk_mask
);
4464 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4465 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4467 /* 'next' is after the last device address that we
4468 * might write to for this chunk in the new layout
4470 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4472 /* 'safe' is the earliest device address that we might
4473 * read from in the old layout after a restart
4475 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4477 /* Need to update metadata if 'next' might be beyond 'safe'
4478 * as that would possibly corrupt data
4480 if (next
> safe
+ conf
->offset_diff
)
4483 sector_nr
= conf
->reshape_progress
;
4484 last
= sector_nr
| (conf
->geo
.chunk_mask
4485 & conf
->prev
.chunk_mask
);
4487 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4488 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4492 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4493 /* Need to update reshape_position in metadata */
4495 mddev
->reshape_position
= conf
->reshape_progress
;
4496 if (mddev
->reshape_backwards
)
4497 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4498 - conf
->reshape_progress
;
4500 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4501 conf
->reshape_checkpoint
= jiffies
;
4502 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
4503 md_wakeup_thread(mddev
->thread
);
4504 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
4505 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4506 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4507 allow_barrier(conf
);
4508 return sectors_done
;
4510 conf
->reshape_safe
= mddev
->reshape_position
;
4511 allow_barrier(conf
);
4515 /* Now schedule reads for blocks from sector_nr to last */
4516 r10_bio
= raid10_alloc_init_r10buf(conf
);
4518 raise_barrier(conf
, sectors_done
!= 0);
4519 atomic_set(&r10_bio
->remaining
, 0);
4520 r10_bio
->mddev
= mddev
;
4521 r10_bio
->sector
= sector_nr
;
4522 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4523 r10_bio
->sectors
= last
- sector_nr
+ 1;
4524 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4525 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4528 /* Cannot read from here, so need to record bad blocks
4529 * on all the target devices.
4532 mempool_free(r10_bio
, conf
->r10buf_pool
);
4533 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4534 return sectors_done
;
4537 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4539 bio_set_dev(read_bio
, rdev
->bdev
);
4540 read_bio
->bi_iter
.bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4541 + rdev
->data_offset
);
4542 read_bio
->bi_private
= r10_bio
;
4543 read_bio
->bi_end_io
= end_reshape_read
;
4544 bio_set_op_attrs(read_bio
, REQ_OP_READ
, 0);
4545 read_bio
->bi_flags
&= (~0UL << BIO_RESET_BITS
);
4546 read_bio
->bi_status
= 0;
4547 read_bio
->bi_vcnt
= 0;
4548 read_bio
->bi_iter
.bi_size
= 0;
4549 r10_bio
->master_bio
= read_bio
;
4550 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4552 /* Now find the locations in the new layout */
4553 __raid10_find_phys(&conf
->geo
, r10_bio
);
4556 read_bio
->bi_next
= NULL
;
4559 for (s
= 0; s
< conf
->copies
*2; s
++) {
4561 int d
= r10_bio
->devs
[s
/2].devnum
;
4562 struct md_rdev
*rdev2
;
4564 rdev2
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4565 b
= r10_bio
->devs
[s
/2].repl_bio
;
4567 rdev2
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4568 b
= r10_bio
->devs
[s
/2].bio
;
4570 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4573 bio_set_dev(b
, rdev2
->bdev
);
4574 b
->bi_iter
.bi_sector
= r10_bio
->devs
[s
/2].addr
+
4575 rdev2
->new_data_offset
;
4576 b
->bi_end_io
= end_reshape_write
;
4577 bio_set_op_attrs(b
, REQ_OP_WRITE
, 0);
4582 /* Now add as many pages as possible to all of these bios. */
4585 pages
= get_resync_pages(r10_bio
->devs
[0].bio
)->pages
;
4586 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4587 struct page
*page
= pages
[s
/ (PAGE_SIZE
>> 9)];
4588 int len
= (max_sectors
- s
) << 9;
4589 if (len
> PAGE_SIZE
)
4591 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4593 * won't fail because the vec table is big enough
4594 * to hold all these pages
4596 bio_add_page(bio
, page
, len
, 0);
4598 sector_nr
+= len
>> 9;
4599 nr_sectors
+= len
>> 9;
4602 r10_bio
->sectors
= nr_sectors
;
4604 /* Now submit the read */
4605 md_sync_acct_bio(read_bio
, r10_bio
->sectors
);
4606 atomic_inc(&r10_bio
->remaining
);
4607 read_bio
->bi_next
= NULL
;
4608 generic_make_request(read_bio
);
4609 sector_nr
+= nr_sectors
;
4610 sectors_done
+= nr_sectors
;
4611 if (sector_nr
<= last
)
4614 /* Now that we have done the whole section we can
4615 * update reshape_progress
4617 if (mddev
->reshape_backwards
)
4618 conf
->reshape_progress
-= sectors_done
;
4620 conf
->reshape_progress
+= sectors_done
;
4622 return sectors_done
;
4625 static void end_reshape_request(struct r10bio
*r10_bio
);
4626 static int handle_reshape_read_error(struct mddev
*mddev
,
4627 struct r10bio
*r10_bio
);
4628 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4630 /* Reshape read completed. Hopefully we have a block
4632 * If we got a read error then we do sync 1-page reads from
4633 * elsewhere until we find the data - or give up.
4635 struct r10conf
*conf
= mddev
->private;
4638 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4639 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4640 /* Reshape has been aborted */
4641 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4645 /* We definitely have the data in the pages, schedule the
4648 atomic_set(&r10_bio
->remaining
, 1);
4649 for (s
= 0; s
< conf
->copies
*2; s
++) {
4651 int d
= r10_bio
->devs
[s
/2].devnum
;
4652 struct md_rdev
*rdev
;
4655 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4656 b
= r10_bio
->devs
[s
/2].repl_bio
;
4658 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4659 b
= r10_bio
->devs
[s
/2].bio
;
4661 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
4665 atomic_inc(&rdev
->nr_pending
);
4667 md_sync_acct_bio(b
, r10_bio
->sectors
);
4668 atomic_inc(&r10_bio
->remaining
);
4670 generic_make_request(b
);
4672 end_reshape_request(r10_bio
);
4675 static void end_reshape(struct r10conf
*conf
)
4677 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4680 spin_lock_irq(&conf
->device_lock
);
4681 conf
->prev
= conf
->geo
;
4682 md_finish_reshape(conf
->mddev
);
4684 conf
->reshape_progress
= MaxSector
;
4685 conf
->reshape_safe
= MaxSector
;
4686 spin_unlock_irq(&conf
->device_lock
);
4688 /* read-ahead size must cover two whole stripes, which is
4689 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4691 if (conf
->mddev
->queue
) {
4692 int stripe
= conf
->geo
.raid_disks
*
4693 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4694 stripe
/= conf
->geo
.near_copies
;
4695 if (conf
->mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
4696 conf
->mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
4701 static int handle_reshape_read_error(struct mddev
*mddev
,
4702 struct r10bio
*r10_bio
)
4704 /* Use sync reads to get the blocks from somewhere else */
4705 int sectors
= r10_bio
->sectors
;
4706 struct r10conf
*conf
= mddev
->private;
4707 struct r10bio
*r10b
;
4710 struct page
**pages
;
4712 r10b
= kmalloc(sizeof(*r10b
) +
4713 sizeof(struct r10dev
) * conf
->copies
, GFP_NOIO
);
4715 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4719 /* reshape IOs share pages from .devs[0].bio */
4720 pages
= get_resync_pages(r10_bio
->devs
[0].bio
)->pages
;
4722 r10b
->sector
= r10_bio
->sector
;
4723 __raid10_find_phys(&conf
->prev
, r10b
);
4728 int first_slot
= slot
;
4730 if (s
> (PAGE_SIZE
>> 9))
4735 int d
= r10b
->devs
[slot
].devnum
;
4736 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4739 test_bit(Faulty
, &rdev
->flags
) ||
4740 !test_bit(In_sync
, &rdev
->flags
))
4743 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4744 atomic_inc(&rdev
->nr_pending
);
4746 success
= sync_page_io(rdev
,
4750 REQ_OP_READ
, 0, false);
4751 rdev_dec_pending(rdev
, mddev
);
4757 if (slot
>= conf
->copies
)
4759 if (slot
== first_slot
)
4764 /* couldn't read this block, must give up */
4765 set_bit(MD_RECOVERY_INTR
,
4777 static void end_reshape_write(struct bio
*bio
)
4779 struct r10bio
*r10_bio
= get_resync_r10bio(bio
);
4780 struct mddev
*mddev
= r10_bio
->mddev
;
4781 struct r10conf
*conf
= mddev
->private;
4785 struct md_rdev
*rdev
= NULL
;
4787 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4789 rdev
= conf
->mirrors
[d
].replacement
;
4792 rdev
= conf
->mirrors
[d
].rdev
;
4795 if (bio
->bi_status
) {
4796 /* FIXME should record badblock */
4797 md_error(mddev
, rdev
);
4800 rdev_dec_pending(rdev
, mddev
);
4801 end_reshape_request(r10_bio
);
4804 static void end_reshape_request(struct r10bio
*r10_bio
)
4806 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4808 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4809 bio_put(r10_bio
->master_bio
);
4813 static void raid10_finish_reshape(struct mddev
*mddev
)
4815 struct r10conf
*conf
= mddev
->private;
4817 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4820 if (mddev
->delta_disks
> 0) {
4821 sector_t size
= raid10_size(mddev
, 0, 0);
4822 md_set_array_sectors(mddev
, size
);
4823 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4824 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4825 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4827 mddev
->resync_max_sectors
= size
;
4829 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4830 revalidate_disk(mddev
->gendisk
);
4835 for (d
= conf
->geo
.raid_disks
;
4836 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4838 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4840 clear_bit(In_sync
, &rdev
->flags
);
4841 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4843 clear_bit(In_sync
, &rdev
->flags
);
4847 mddev
->layout
= mddev
->new_layout
;
4848 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4849 mddev
->reshape_position
= MaxSector
;
4850 mddev
->delta_disks
= 0;
4851 mddev
->reshape_backwards
= 0;
4854 static struct md_personality raid10_personality
=
4858 .owner
= THIS_MODULE
,
4859 .make_request
= raid10_make_request
,
4861 .free
= raid10_free
,
4862 .status
= raid10_status
,
4863 .error_handler
= raid10_error
,
4864 .hot_add_disk
= raid10_add_disk
,
4865 .hot_remove_disk
= raid10_remove_disk
,
4866 .spare_active
= raid10_spare_active
,
4867 .sync_request
= raid10_sync_request
,
4868 .quiesce
= raid10_quiesce
,
4869 .size
= raid10_size
,
4870 .resize
= raid10_resize
,
4871 .takeover
= raid10_takeover
,
4872 .check_reshape
= raid10_check_reshape
,
4873 .start_reshape
= raid10_start_reshape
,
4874 .finish_reshape
= raid10_finish_reshape
,
4875 .congested
= raid10_congested
,
4878 static int __init
raid_init(void)
4880 return register_md_personality(&raid10_personality
);
4883 static void raid_exit(void)
4885 unregister_md_personality(&raid10_personality
);
4888 module_init(raid_init
);
4889 module_exit(raid_exit
);
4890 MODULE_LICENSE("GPL");
4891 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4892 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4893 MODULE_ALIAS("md-raid10");
4894 MODULE_ALIAS("md-level-10");
4896 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);