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
]);
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 allow_barrier(conf
);
1198 generic_make_request(bio
);
1201 r10_bio
->master_bio
= bio
;
1202 r10_bio
->sectors
= max_sectors
;
1204 slot
= r10_bio
->read_slot
;
1206 read_bio
= bio_clone_fast(bio
, gfp
, mddev
->bio_set
);
1208 r10_bio
->devs
[slot
].bio
= read_bio
;
1209 r10_bio
->devs
[slot
].rdev
= rdev
;
1211 read_bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
+
1212 choose_data_offset(r10_bio
, rdev
);
1213 bio_set_dev(read_bio
, rdev
->bdev
);
1214 read_bio
->bi_end_io
= raid10_end_read_request
;
1215 bio_set_op_attrs(read_bio
, op
, do_sync
);
1216 if (test_bit(FailFast
, &rdev
->flags
) &&
1217 test_bit(R10BIO_FailFast
, &r10_bio
->state
))
1218 read_bio
->bi_opf
|= MD_FAILFAST
;
1219 read_bio
->bi_private
= r10_bio
;
1222 trace_block_bio_remap(read_bio
->bi_disk
->queue
,
1223 read_bio
, disk_devt(mddev
->gendisk
),
1225 generic_make_request(read_bio
);
1229 static void raid10_write_one_disk(struct mddev
*mddev
, struct r10bio
*r10_bio
,
1230 struct bio
*bio
, bool replacement
,
1233 const int op
= bio_op(bio
);
1234 const unsigned long do_sync
= (bio
->bi_opf
& REQ_SYNC
);
1235 const unsigned long do_fua
= (bio
->bi_opf
& REQ_FUA
);
1236 unsigned long flags
;
1237 struct blk_plug_cb
*cb
;
1238 struct raid10_plug_cb
*plug
= NULL
;
1239 struct r10conf
*conf
= mddev
->private;
1240 struct md_rdev
*rdev
;
1241 int devnum
= r10_bio
->devs
[n_copy
].devnum
;
1245 rdev
= conf
->mirrors
[devnum
].replacement
;
1247 /* Replacement just got moved to main 'rdev' */
1249 rdev
= conf
->mirrors
[devnum
].rdev
;
1252 rdev
= conf
->mirrors
[devnum
].rdev
;
1254 mbio
= bio_clone_fast(bio
, GFP_NOIO
, mddev
->bio_set
);
1256 r10_bio
->devs
[n_copy
].repl_bio
= mbio
;
1258 r10_bio
->devs
[n_copy
].bio
= mbio
;
1260 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[n_copy
].addr
+
1261 choose_data_offset(r10_bio
, rdev
));
1262 bio_set_dev(mbio
, rdev
->bdev
);
1263 mbio
->bi_end_io
= raid10_end_write_request
;
1264 bio_set_op_attrs(mbio
, op
, do_sync
| do_fua
);
1265 if (!replacement
&& test_bit(FailFast
,
1266 &conf
->mirrors
[devnum
].rdev
->flags
)
1267 && enough(conf
, devnum
))
1268 mbio
->bi_opf
|= MD_FAILFAST
;
1269 mbio
->bi_private
= r10_bio
;
1271 if (conf
->mddev
->gendisk
)
1272 trace_block_bio_remap(mbio
->bi_disk
->queue
,
1273 mbio
, disk_devt(conf
->mddev
->gendisk
),
1275 /* flush_pending_writes() needs access to the rdev so...*/
1276 mbio
->bi_disk
= (void *)rdev
;
1278 atomic_inc(&r10_bio
->remaining
);
1280 cb
= blk_check_plugged(raid10_unplug
, mddev
, sizeof(*plug
));
1282 plug
= container_of(cb
, struct raid10_plug_cb
, cb
);
1286 bio_list_add(&plug
->pending
, mbio
);
1287 plug
->pending_cnt
++;
1289 spin_lock_irqsave(&conf
->device_lock
, flags
);
1290 bio_list_add(&conf
->pending_bio_list
, mbio
);
1291 conf
->pending_count
++;
1292 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1293 md_wakeup_thread(mddev
->thread
);
1297 static void raid10_write_request(struct mddev
*mddev
, struct bio
*bio
,
1298 struct r10bio
*r10_bio
)
1300 struct r10conf
*conf
= mddev
->private;
1302 struct md_rdev
*blocked_rdev
;
1306 if ((mddev_is_clustered(mddev
) &&
1307 md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1308 bio
->bi_iter
.bi_sector
,
1309 bio_end_sector(bio
)))) {
1312 prepare_to_wait(&conf
->wait_barrier
,
1314 if (!md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1315 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
)))
1319 finish_wait(&conf
->wait_barrier
, &w
);
1323 * Register the new request and wait if the reconstruction
1324 * thread has put up a bar for new requests.
1325 * Continue immediately if no resync is active currently.
1329 sectors
= r10_bio
->sectors
;
1330 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1331 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1332 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1334 * IO spans the reshape position. Need to wait for reshape to
1337 raid10_log(conf
->mddev
, "wait reshape");
1338 allow_barrier(conf
);
1339 wait_event(conf
->wait_barrier
,
1340 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1341 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1346 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1347 (mddev
->reshape_backwards
1348 ? (bio
->bi_iter
.bi_sector
< conf
->reshape_safe
&&
1349 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
)
1350 : (bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_safe
&&
1351 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
))) {
1352 /* Need to update reshape_position in metadata */
1353 mddev
->reshape_position
= conf
->reshape_progress
;
1354 set_mask_bits(&mddev
->sb_flags
, 0,
1355 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
1356 md_wakeup_thread(mddev
->thread
);
1357 raid10_log(conf
->mddev
, "wait reshape metadata");
1358 wait_event(mddev
->sb_wait
,
1359 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
));
1361 conf
->reshape_safe
= mddev
->reshape_position
;
1364 if (conf
->pending_count
>= max_queued_requests
) {
1365 md_wakeup_thread(mddev
->thread
);
1366 raid10_log(mddev
, "wait queued");
1367 wait_event(conf
->wait_barrier
,
1368 conf
->pending_count
< max_queued_requests
);
1370 /* first select target devices under rcu_lock and
1371 * inc refcount on their rdev. Record them by setting
1373 * If there are known/acknowledged bad blocks on any device
1374 * on which we have seen a write error, we want to avoid
1375 * writing to those blocks. This potentially requires several
1376 * writes to write around the bad blocks. Each set of writes
1377 * gets its own r10_bio with a set of bios attached.
1380 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1381 raid10_find_phys(conf
, r10_bio
);
1383 blocked_rdev
= NULL
;
1385 max_sectors
= r10_bio
->sectors
;
1387 for (i
= 0; i
< conf
->copies
; i
++) {
1388 int d
= r10_bio
->devs
[i
].devnum
;
1389 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1390 struct md_rdev
*rrdev
= rcu_dereference(
1391 conf
->mirrors
[d
].replacement
);
1394 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1395 atomic_inc(&rdev
->nr_pending
);
1396 blocked_rdev
= rdev
;
1399 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1400 atomic_inc(&rrdev
->nr_pending
);
1401 blocked_rdev
= rrdev
;
1404 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)))
1406 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)))
1409 r10_bio
->devs
[i
].bio
= NULL
;
1410 r10_bio
->devs
[i
].repl_bio
= NULL
;
1412 if (!rdev
&& !rrdev
) {
1413 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1416 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1418 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1422 is_bad
= is_badblock(rdev
, dev_sector
, max_sectors
,
1423 &first_bad
, &bad_sectors
);
1425 /* Mustn't write here until the bad block
1428 atomic_inc(&rdev
->nr_pending
);
1429 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1430 blocked_rdev
= rdev
;
1433 if (is_bad
&& first_bad
<= dev_sector
) {
1434 /* Cannot write here at all */
1435 bad_sectors
-= (dev_sector
- first_bad
);
1436 if (bad_sectors
< max_sectors
)
1437 /* Mustn't write more than bad_sectors
1438 * to other devices yet
1440 max_sectors
= bad_sectors
;
1441 /* We don't set R10BIO_Degraded as that
1442 * only applies if the disk is missing,
1443 * so it might be re-added, and we want to
1444 * know to recover this chunk.
1445 * In this case the device is here, and the
1446 * fact that this chunk is not in-sync is
1447 * recorded in the bad block log.
1452 int good_sectors
= first_bad
- dev_sector
;
1453 if (good_sectors
< max_sectors
)
1454 max_sectors
= good_sectors
;
1458 r10_bio
->devs
[i
].bio
= bio
;
1459 atomic_inc(&rdev
->nr_pending
);
1462 r10_bio
->devs
[i
].repl_bio
= bio
;
1463 atomic_inc(&rrdev
->nr_pending
);
1468 if (unlikely(blocked_rdev
)) {
1469 /* Have to wait for this device to get unblocked, then retry */
1473 for (j
= 0; j
< i
; j
++) {
1474 if (r10_bio
->devs
[j
].bio
) {
1475 d
= r10_bio
->devs
[j
].devnum
;
1476 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1478 if (r10_bio
->devs
[j
].repl_bio
) {
1479 struct md_rdev
*rdev
;
1480 d
= r10_bio
->devs
[j
].devnum
;
1481 rdev
= conf
->mirrors
[d
].replacement
;
1483 /* Race with remove_disk */
1485 rdev
= conf
->mirrors
[d
].rdev
;
1487 rdev_dec_pending(rdev
, mddev
);
1490 allow_barrier(conf
);
1491 raid10_log(conf
->mddev
, "wait rdev %d blocked", blocked_rdev
->raid_disk
);
1492 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1497 if (max_sectors
< r10_bio
->sectors
)
1498 r10_bio
->sectors
= max_sectors
;
1500 if (r10_bio
->sectors
< bio_sectors(bio
)) {
1501 struct bio
*split
= bio_split(bio
, r10_bio
->sectors
,
1502 GFP_NOIO
, conf
->bio_split
);
1503 bio_chain(split
, bio
);
1504 allow_barrier(conf
);
1505 generic_make_request(bio
);
1508 r10_bio
->master_bio
= bio
;
1511 atomic_set(&r10_bio
->remaining
, 1);
1512 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1514 for (i
= 0; i
< conf
->copies
; i
++) {
1515 if (r10_bio
->devs
[i
].bio
)
1516 raid10_write_one_disk(mddev
, r10_bio
, bio
, false, i
);
1517 if (r10_bio
->devs
[i
].repl_bio
)
1518 raid10_write_one_disk(mddev
, r10_bio
, bio
, true, i
);
1520 one_write_done(r10_bio
);
1523 static void __make_request(struct mddev
*mddev
, struct bio
*bio
, int sectors
)
1525 struct r10conf
*conf
= mddev
->private;
1526 struct r10bio
*r10_bio
;
1528 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1530 r10_bio
->master_bio
= bio
;
1531 r10_bio
->sectors
= sectors
;
1533 r10_bio
->mddev
= mddev
;
1534 r10_bio
->sector
= bio
->bi_iter
.bi_sector
;
1536 memset(r10_bio
->devs
, 0, sizeof(r10_bio
->devs
[0]) * conf
->copies
);
1538 if (bio_data_dir(bio
) == READ
)
1539 raid10_read_request(mddev
, bio
, r10_bio
);
1541 raid10_write_request(mddev
, bio
, r10_bio
);
1544 static bool raid10_make_request(struct mddev
*mddev
, struct bio
*bio
)
1546 struct r10conf
*conf
= mddev
->private;
1547 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1548 int chunk_sects
= chunk_mask
+ 1;
1549 int sectors
= bio_sectors(bio
);
1551 if (unlikely(bio
->bi_opf
& REQ_PREFLUSH
)
1552 && md_flush_request(mddev
, bio
))
1555 if (!md_write_start(mddev
, bio
))
1559 * If this request crosses a chunk boundary, we need to split
1562 if (unlikely((bio
->bi_iter
.bi_sector
& chunk_mask
) +
1563 sectors
> chunk_sects
1564 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1565 || conf
->prev
.near_copies
<
1566 conf
->prev
.raid_disks
)))
1567 sectors
= chunk_sects
-
1568 (bio
->bi_iter
.bi_sector
&
1570 __make_request(mddev
, bio
, sectors
);
1572 /* In case raid10d snuck in to freeze_array */
1573 wake_up(&conf
->wait_barrier
);
1577 static void raid10_status(struct seq_file
*seq
, struct mddev
*mddev
)
1579 struct r10conf
*conf
= mddev
->private;
1582 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1583 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1584 if (conf
->geo
.near_copies
> 1)
1585 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1586 if (conf
->geo
.far_copies
> 1) {
1587 if (conf
->geo
.far_offset
)
1588 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1590 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1591 if (conf
->geo
.far_set_size
!= conf
->geo
.raid_disks
)
1592 seq_printf(seq
, " %d devices per set", conf
->geo
.far_set_size
);
1594 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1595 conf
->geo
.raid_disks
- mddev
->degraded
);
1597 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1598 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1599 seq_printf(seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1602 seq_printf(seq
, "]");
1605 /* check if there are enough drives for
1606 * every block to appear on atleast one.
1607 * Don't consider the device numbered 'ignore'
1608 * as we might be about to remove it.
1610 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1616 disks
= conf
->prev
.raid_disks
;
1617 ncopies
= conf
->prev
.near_copies
;
1619 disks
= conf
->geo
.raid_disks
;
1620 ncopies
= conf
->geo
.near_copies
;
1625 int n
= conf
->copies
;
1629 struct md_rdev
*rdev
;
1630 if (this != ignore
&&
1631 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1632 test_bit(In_sync
, &rdev
->flags
))
1634 this = (this+1) % disks
;
1638 first
= (first
+ ncopies
) % disks
;
1639 } while (first
!= 0);
1646 static int enough(struct r10conf
*conf
, int ignore
)
1648 /* when calling 'enough', both 'prev' and 'geo' must
1650 * This is ensured if ->reconfig_mutex or ->device_lock
1653 return _enough(conf
, 0, ignore
) &&
1654 _enough(conf
, 1, ignore
);
1657 static void raid10_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1659 char b
[BDEVNAME_SIZE
];
1660 struct r10conf
*conf
= mddev
->private;
1661 unsigned long flags
;
1664 * If it is not operational, then we have already marked it as dead
1665 * else if it is the last working disks, ignore the error, let the
1666 * next level up know.
1667 * else mark the drive as failed
1669 spin_lock_irqsave(&conf
->device_lock
, flags
);
1670 if (test_bit(In_sync
, &rdev
->flags
)
1671 && !enough(conf
, rdev
->raid_disk
)) {
1673 * Don't fail the drive, just return an IO error.
1675 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1678 if (test_and_clear_bit(In_sync
, &rdev
->flags
))
1681 * If recovery is running, make sure it aborts.
1683 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1684 set_bit(Blocked
, &rdev
->flags
);
1685 set_bit(Faulty
, &rdev
->flags
);
1686 set_mask_bits(&mddev
->sb_flags
, 0,
1687 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
1688 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1689 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1690 "md/raid10:%s: Operation continuing on %d devices.\n",
1691 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1692 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1695 static void print_conf(struct r10conf
*conf
)
1698 struct md_rdev
*rdev
;
1700 pr_debug("RAID10 conf printout:\n");
1702 pr_debug("(!conf)\n");
1705 pr_debug(" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1706 conf
->geo
.raid_disks
);
1708 /* This is only called with ->reconfix_mutex held, so
1709 * rcu protection of rdev is not needed */
1710 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1711 char b
[BDEVNAME_SIZE
];
1712 rdev
= conf
->mirrors
[i
].rdev
;
1714 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1715 i
, !test_bit(In_sync
, &rdev
->flags
),
1716 !test_bit(Faulty
, &rdev
->flags
),
1717 bdevname(rdev
->bdev
,b
));
1721 static void close_sync(struct r10conf
*conf
)
1724 allow_barrier(conf
);
1726 mempool_destroy(conf
->r10buf_pool
);
1727 conf
->r10buf_pool
= NULL
;
1730 static int raid10_spare_active(struct mddev
*mddev
)
1733 struct r10conf
*conf
= mddev
->private;
1734 struct raid10_info
*tmp
;
1736 unsigned long flags
;
1739 * Find all non-in_sync disks within the RAID10 configuration
1740 * and mark them in_sync
1742 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1743 tmp
= conf
->mirrors
+ i
;
1744 if (tmp
->replacement
1745 && tmp
->replacement
->recovery_offset
== MaxSector
1746 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1747 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1748 /* Replacement has just become active */
1750 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1753 /* Replaced device not technically faulty,
1754 * but we need to be sure it gets removed
1755 * and never re-added.
1757 set_bit(Faulty
, &tmp
->rdev
->flags
);
1758 sysfs_notify_dirent_safe(
1759 tmp
->rdev
->sysfs_state
);
1761 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1762 } else if (tmp
->rdev
1763 && tmp
->rdev
->recovery_offset
== MaxSector
1764 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1765 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1767 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1770 spin_lock_irqsave(&conf
->device_lock
, flags
);
1771 mddev
->degraded
-= count
;
1772 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1778 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1780 struct r10conf
*conf
= mddev
->private;
1784 int last
= conf
->geo
.raid_disks
- 1;
1786 if (mddev
->recovery_cp
< MaxSector
)
1787 /* only hot-add to in-sync arrays, as recovery is
1788 * very different from resync
1791 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1794 if (md_integrity_add_rdev(rdev
, mddev
))
1797 if (rdev
->raid_disk
>= 0)
1798 first
= last
= rdev
->raid_disk
;
1800 if (rdev
->saved_raid_disk
>= first
&&
1801 rdev
->saved_raid_disk
< conf
->geo
.raid_disks
&&
1802 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1803 mirror
= rdev
->saved_raid_disk
;
1806 for ( ; mirror
<= last
; mirror
++) {
1807 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1808 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1811 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1812 p
->replacement
!= NULL
)
1814 clear_bit(In_sync
, &rdev
->flags
);
1815 set_bit(Replacement
, &rdev
->flags
);
1816 rdev
->raid_disk
= mirror
;
1819 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1820 rdev
->data_offset
<< 9);
1822 rcu_assign_pointer(p
->replacement
, rdev
);
1827 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1828 rdev
->data_offset
<< 9);
1830 p
->head_position
= 0;
1831 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1832 rdev
->raid_disk
= mirror
;
1834 if (rdev
->saved_raid_disk
!= mirror
)
1836 rcu_assign_pointer(p
->rdev
, rdev
);
1839 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1840 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1846 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1848 struct r10conf
*conf
= mddev
->private;
1850 int number
= rdev
->raid_disk
;
1851 struct md_rdev
**rdevp
;
1852 struct raid10_info
*p
= conf
->mirrors
+ number
;
1855 if (rdev
== p
->rdev
)
1857 else if (rdev
== p
->replacement
)
1858 rdevp
= &p
->replacement
;
1862 if (test_bit(In_sync
, &rdev
->flags
) ||
1863 atomic_read(&rdev
->nr_pending
)) {
1867 /* Only remove non-faulty devices if recovery
1870 if (!test_bit(Faulty
, &rdev
->flags
) &&
1871 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1872 (!p
->replacement
|| p
->replacement
== rdev
) &&
1873 number
< conf
->geo
.raid_disks
&&
1879 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
1881 if (atomic_read(&rdev
->nr_pending
)) {
1882 /* lost the race, try later */
1888 if (p
->replacement
) {
1889 /* We must have just cleared 'rdev' */
1890 p
->rdev
= p
->replacement
;
1891 clear_bit(Replacement
, &p
->replacement
->flags
);
1892 smp_mb(); /* Make sure other CPUs may see both as identical
1893 * but will never see neither -- if they are careful.
1895 p
->replacement
= NULL
;
1898 clear_bit(WantReplacement
, &rdev
->flags
);
1899 err
= md_integrity_register(mddev
);
1907 static void __end_sync_read(struct r10bio
*r10_bio
, struct bio
*bio
, int d
)
1909 struct r10conf
*conf
= r10_bio
->mddev
->private;
1911 if (!bio
->bi_status
)
1912 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1914 /* The write handler will notice the lack of
1915 * R10BIO_Uptodate and record any errors etc
1917 atomic_add(r10_bio
->sectors
,
1918 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1920 /* for reconstruct, we always reschedule after a read.
1921 * for resync, only after all reads
1923 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1924 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1925 atomic_dec_and_test(&r10_bio
->remaining
)) {
1926 /* we have read all the blocks,
1927 * do the comparison in process context in raid10d
1929 reschedule_retry(r10_bio
);
1933 static void end_sync_read(struct bio
*bio
)
1935 struct r10bio
*r10_bio
= get_resync_r10bio(bio
);
1936 struct r10conf
*conf
= r10_bio
->mddev
->private;
1937 int d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1939 __end_sync_read(r10_bio
, bio
, d
);
1942 static void end_reshape_read(struct bio
*bio
)
1944 /* reshape read bio isn't allocated from r10buf_pool */
1945 struct r10bio
*r10_bio
= bio
->bi_private
;
1947 __end_sync_read(r10_bio
, bio
, r10_bio
->read_slot
);
1950 static void end_sync_request(struct r10bio
*r10_bio
)
1952 struct mddev
*mddev
= r10_bio
->mddev
;
1954 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1955 if (r10_bio
->master_bio
== NULL
) {
1956 /* the primary of several recovery bios */
1957 sector_t s
= r10_bio
->sectors
;
1958 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1959 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1960 reschedule_retry(r10_bio
);
1963 md_done_sync(mddev
, s
, 1);
1966 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1967 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1968 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1969 reschedule_retry(r10_bio
);
1977 static void end_sync_write(struct bio
*bio
)
1979 struct r10bio
*r10_bio
= get_resync_r10bio(bio
);
1980 struct mddev
*mddev
= r10_bio
->mddev
;
1981 struct r10conf
*conf
= mddev
->private;
1987 struct md_rdev
*rdev
= NULL
;
1989 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1991 rdev
= conf
->mirrors
[d
].replacement
;
1993 rdev
= conf
->mirrors
[d
].rdev
;
1995 if (bio
->bi_status
) {
1997 md_error(mddev
, rdev
);
1999 set_bit(WriteErrorSeen
, &rdev
->flags
);
2000 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2001 set_bit(MD_RECOVERY_NEEDED
,
2002 &rdev
->mddev
->recovery
);
2003 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
2005 } else if (is_badblock(rdev
,
2006 r10_bio
->devs
[slot
].addr
,
2008 &first_bad
, &bad_sectors
))
2009 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
2011 rdev_dec_pending(rdev
, mddev
);
2013 end_sync_request(r10_bio
);
2017 * Note: sync and recover and handled very differently for raid10
2018 * This code is for resync.
2019 * For resync, we read through virtual addresses and read all blocks.
2020 * If there is any error, we schedule a write. The lowest numbered
2021 * drive is authoritative.
2022 * However requests come for physical address, so we need to map.
2023 * For every physical address there are raid_disks/copies virtual addresses,
2024 * which is always are least one, but is not necessarly an integer.
2025 * This means that a physical address can span multiple chunks, so we may
2026 * have to submit multiple io requests for a single sync request.
2029 * We check if all blocks are in-sync and only write to blocks that
2032 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2034 struct r10conf
*conf
= mddev
->private;
2036 struct bio
*tbio
, *fbio
;
2038 struct page
**tpages
, **fpages
;
2040 atomic_set(&r10_bio
->remaining
, 1);
2042 /* find the first device with a block */
2043 for (i
=0; i
<conf
->copies
; i
++)
2044 if (!r10_bio
->devs
[i
].bio
->bi_status
)
2047 if (i
== conf
->copies
)
2051 fbio
= r10_bio
->devs
[i
].bio
;
2052 fbio
->bi_iter
.bi_size
= r10_bio
->sectors
<< 9;
2053 fbio
->bi_iter
.bi_idx
= 0;
2054 fpages
= get_resync_pages(fbio
)->pages
;
2056 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2057 /* now find blocks with errors */
2058 for (i
=0 ; i
< conf
->copies
; i
++) {
2060 struct md_rdev
*rdev
;
2061 struct resync_pages
*rp
;
2063 tbio
= r10_bio
->devs
[i
].bio
;
2065 if (tbio
->bi_end_io
!= end_sync_read
)
2070 tpages
= get_resync_pages(tbio
)->pages
;
2071 d
= r10_bio
->devs
[i
].devnum
;
2072 rdev
= conf
->mirrors
[d
].rdev
;
2073 if (!r10_bio
->devs
[i
].bio
->bi_status
) {
2074 /* We know that the bi_io_vec layout is the same for
2075 * both 'first' and 'i', so we just compare them.
2076 * All vec entries are PAGE_SIZE;
2078 int sectors
= r10_bio
->sectors
;
2079 for (j
= 0; j
< vcnt
; j
++) {
2080 int len
= PAGE_SIZE
;
2081 if (sectors
< (len
/ 512))
2082 len
= sectors
* 512;
2083 if (memcmp(page_address(fpages
[j
]),
2084 page_address(tpages
[j
]),
2091 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2092 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2093 /* Don't fix anything. */
2095 } else if (test_bit(FailFast
, &rdev
->flags
)) {
2096 /* Just give up on this device */
2097 md_error(rdev
->mddev
, rdev
);
2100 /* Ok, we need to write this bio, either to correct an
2101 * inconsistency or to correct an unreadable block.
2102 * First we need to fixup bv_offset, bv_len and
2103 * bi_vecs, as the read request might have corrupted these
2105 rp
= get_resync_pages(tbio
);
2108 md_bio_reset_resync_pages(tbio
, rp
, fbio
->bi_iter
.bi_size
);
2110 rp
->raid_bio
= r10_bio
;
2111 tbio
->bi_private
= rp
;
2112 tbio
->bi_iter
.bi_sector
= r10_bio
->devs
[i
].addr
;
2113 tbio
->bi_end_io
= end_sync_write
;
2114 bio_set_op_attrs(tbio
, REQ_OP_WRITE
, 0);
2116 bio_copy_data(tbio
, fbio
);
2118 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2119 atomic_inc(&r10_bio
->remaining
);
2120 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2122 if (test_bit(FailFast
, &conf
->mirrors
[d
].rdev
->flags
))
2123 tbio
->bi_opf
|= MD_FAILFAST
;
2124 tbio
->bi_iter
.bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2125 bio_set_dev(tbio
, conf
->mirrors
[d
].rdev
->bdev
);
2126 generic_make_request(tbio
);
2129 /* Now write out to any replacement devices
2132 for (i
= 0; i
< conf
->copies
; i
++) {
2135 tbio
= r10_bio
->devs
[i
].repl_bio
;
2136 if (!tbio
|| !tbio
->bi_end_io
)
2138 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2139 && r10_bio
->devs
[i
].bio
!= fbio
)
2140 bio_copy_data(tbio
, fbio
);
2141 d
= r10_bio
->devs
[i
].devnum
;
2142 atomic_inc(&r10_bio
->remaining
);
2143 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2145 generic_make_request(tbio
);
2149 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2150 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2156 * Now for the recovery code.
2157 * Recovery happens across physical sectors.
2158 * We recover all non-is_sync drives by finding the virtual address of
2159 * each, and then choose a working drive that also has that virt address.
2160 * There is a separate r10_bio for each non-in_sync drive.
2161 * Only the first two slots are in use. The first for reading,
2162 * The second for writing.
2165 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2167 /* We got a read error during recovery.
2168 * We repeat the read in smaller page-sized sections.
2169 * If a read succeeds, write it to the new device or record
2170 * a bad block if we cannot.
2171 * If a read fails, record a bad block on both old and
2174 struct mddev
*mddev
= r10_bio
->mddev
;
2175 struct r10conf
*conf
= mddev
->private;
2176 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2178 int sectors
= r10_bio
->sectors
;
2180 int dr
= r10_bio
->devs
[0].devnum
;
2181 int dw
= r10_bio
->devs
[1].devnum
;
2182 struct page
**pages
= get_resync_pages(bio
)->pages
;
2186 struct md_rdev
*rdev
;
2190 if (s
> (PAGE_SIZE
>>9))
2193 rdev
= conf
->mirrors
[dr
].rdev
;
2194 addr
= r10_bio
->devs
[0].addr
+ sect
,
2195 ok
= sync_page_io(rdev
,
2199 REQ_OP_READ
, 0, false);
2201 rdev
= conf
->mirrors
[dw
].rdev
;
2202 addr
= r10_bio
->devs
[1].addr
+ sect
;
2203 ok
= sync_page_io(rdev
,
2207 REQ_OP_WRITE
, 0, false);
2209 set_bit(WriteErrorSeen
, &rdev
->flags
);
2210 if (!test_and_set_bit(WantReplacement
,
2212 set_bit(MD_RECOVERY_NEEDED
,
2213 &rdev
->mddev
->recovery
);
2217 /* We don't worry if we cannot set a bad block -
2218 * it really is bad so there is no loss in not
2221 rdev_set_badblocks(rdev
, addr
, s
, 0);
2223 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2224 /* need bad block on destination too */
2225 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2226 addr
= r10_bio
->devs
[1].addr
+ sect
;
2227 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2229 /* just abort the recovery */
2230 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2233 conf
->mirrors
[dw
].recovery_disabled
2234 = mddev
->recovery_disabled
;
2235 set_bit(MD_RECOVERY_INTR
,
2248 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2250 struct r10conf
*conf
= mddev
->private;
2252 struct bio
*wbio
, *wbio2
;
2254 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2255 fix_recovery_read_error(r10_bio
);
2256 end_sync_request(r10_bio
);
2261 * share the pages with the first bio
2262 * and submit the write request
2264 d
= r10_bio
->devs
[1].devnum
;
2265 wbio
= r10_bio
->devs
[1].bio
;
2266 wbio2
= r10_bio
->devs
[1].repl_bio
;
2267 /* Need to test wbio2->bi_end_io before we call
2268 * generic_make_request as if the former is NULL,
2269 * the latter is free to free wbio2.
2271 if (wbio2
&& !wbio2
->bi_end_io
)
2273 if (wbio
->bi_end_io
) {
2274 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2275 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2276 generic_make_request(wbio
);
2279 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2280 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2281 bio_sectors(wbio2
));
2282 generic_make_request(wbio2
);
2287 * Used by fix_read_error() to decay the per rdev read_errors.
2288 * We halve the read error count for every hour that has elapsed
2289 * since the last recorded read error.
2292 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2295 unsigned long hours_since_last
;
2296 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2298 cur_time_mon
= ktime_get_seconds();
2300 if (rdev
->last_read_error
== 0) {
2301 /* first time we've seen a read error */
2302 rdev
->last_read_error
= cur_time_mon
;
2306 hours_since_last
= (long)(cur_time_mon
-
2307 rdev
->last_read_error
) / 3600;
2309 rdev
->last_read_error
= cur_time_mon
;
2312 * if hours_since_last is > the number of bits in read_errors
2313 * just set read errors to 0. We do this to avoid
2314 * overflowing the shift of read_errors by hours_since_last.
2316 if (hours_since_last
>= 8 * sizeof(read_errors
))
2317 atomic_set(&rdev
->read_errors
, 0);
2319 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2322 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2323 int sectors
, struct page
*page
, int rw
)
2328 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2329 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2331 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, 0, false))
2335 set_bit(WriteErrorSeen
, &rdev
->flags
);
2336 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2337 set_bit(MD_RECOVERY_NEEDED
,
2338 &rdev
->mddev
->recovery
);
2340 /* need to record an error - either for the block or the device */
2341 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2342 md_error(rdev
->mddev
, rdev
);
2347 * This is a kernel thread which:
2349 * 1. Retries failed read operations on working mirrors.
2350 * 2. Updates the raid superblock when problems encounter.
2351 * 3. Performs writes following reads for array synchronising.
2354 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2356 int sect
= 0; /* Offset from r10_bio->sector */
2357 int sectors
= r10_bio
->sectors
;
2358 struct md_rdev
*rdev
;
2359 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2360 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2362 /* still own a reference to this rdev, so it cannot
2363 * have been cleared recently.
2365 rdev
= conf
->mirrors
[d
].rdev
;
2367 if (test_bit(Faulty
, &rdev
->flags
))
2368 /* drive has already been failed, just ignore any
2369 more fix_read_error() attempts */
2372 check_decay_read_errors(mddev
, rdev
);
2373 atomic_inc(&rdev
->read_errors
);
2374 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2375 char b
[BDEVNAME_SIZE
];
2376 bdevname(rdev
->bdev
, b
);
2378 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2380 atomic_read(&rdev
->read_errors
), max_read_errors
);
2381 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2383 md_error(mddev
, rdev
);
2384 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2390 int sl
= r10_bio
->read_slot
;
2394 if (s
> (PAGE_SIZE
>>9))
2402 d
= r10_bio
->devs
[sl
].devnum
;
2403 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2405 test_bit(In_sync
, &rdev
->flags
) &&
2406 !test_bit(Faulty
, &rdev
->flags
) &&
2407 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2408 &first_bad
, &bad_sectors
) == 0) {
2409 atomic_inc(&rdev
->nr_pending
);
2411 success
= sync_page_io(rdev
,
2412 r10_bio
->devs
[sl
].addr
+
2416 REQ_OP_READ
, 0, false);
2417 rdev_dec_pending(rdev
, mddev
);
2423 if (sl
== conf
->copies
)
2425 } while (!success
&& sl
!= r10_bio
->read_slot
);
2429 /* Cannot read from anywhere, just mark the block
2430 * as bad on the first device to discourage future
2433 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2434 rdev
= conf
->mirrors
[dn
].rdev
;
2436 if (!rdev_set_badblocks(
2438 r10_bio
->devs
[r10_bio
->read_slot
].addr
2441 md_error(mddev
, rdev
);
2442 r10_bio
->devs
[r10_bio
->read_slot
].bio
2449 /* write it back and re-read */
2451 while (sl
!= r10_bio
->read_slot
) {
2452 char b
[BDEVNAME_SIZE
];
2457 d
= r10_bio
->devs
[sl
].devnum
;
2458 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2460 test_bit(Faulty
, &rdev
->flags
) ||
2461 !test_bit(In_sync
, &rdev
->flags
))
2464 atomic_inc(&rdev
->nr_pending
);
2466 if (r10_sync_page_io(rdev
,
2467 r10_bio
->devs
[sl
].addr
+
2469 s
, conf
->tmppage
, WRITE
)
2471 /* Well, this device is dead */
2472 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2474 (unsigned long long)(
2476 choose_data_offset(r10_bio
,
2478 bdevname(rdev
->bdev
, b
));
2479 pr_notice("md/raid10:%s: %s: failing drive\n",
2481 bdevname(rdev
->bdev
, b
));
2483 rdev_dec_pending(rdev
, mddev
);
2487 while (sl
!= r10_bio
->read_slot
) {
2488 char b
[BDEVNAME_SIZE
];
2493 d
= r10_bio
->devs
[sl
].devnum
;
2494 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2496 test_bit(Faulty
, &rdev
->flags
) ||
2497 !test_bit(In_sync
, &rdev
->flags
))
2500 atomic_inc(&rdev
->nr_pending
);
2502 switch (r10_sync_page_io(rdev
,
2503 r10_bio
->devs
[sl
].addr
+
2508 /* Well, this device is dead */
2509 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2511 (unsigned long long)(
2513 choose_data_offset(r10_bio
, rdev
)),
2514 bdevname(rdev
->bdev
, b
));
2515 pr_notice("md/raid10:%s: %s: failing drive\n",
2517 bdevname(rdev
->bdev
, b
));
2520 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2522 (unsigned long long)(
2524 choose_data_offset(r10_bio
, rdev
)),
2525 bdevname(rdev
->bdev
, b
));
2526 atomic_add(s
, &rdev
->corrected_errors
);
2529 rdev_dec_pending(rdev
, mddev
);
2539 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2541 struct bio
*bio
= r10_bio
->master_bio
;
2542 struct mddev
*mddev
= r10_bio
->mddev
;
2543 struct r10conf
*conf
= mddev
->private;
2544 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2545 /* bio has the data to be written to slot 'i' where
2546 * we just recently had a write error.
2547 * We repeatedly clone the bio and trim down to one block,
2548 * then try the write. Where the write fails we record
2550 * It is conceivable that the bio doesn't exactly align with
2551 * blocks. We must handle this.
2553 * We currently own a reference to the rdev.
2559 int sect_to_write
= r10_bio
->sectors
;
2562 if (rdev
->badblocks
.shift
< 0)
2565 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2566 bdev_logical_block_size(rdev
->bdev
) >> 9);
2567 sector
= r10_bio
->sector
;
2568 sectors
= ((r10_bio
->sector
+ block_sectors
)
2569 & ~(sector_t
)(block_sectors
- 1))
2572 while (sect_to_write
) {
2575 if (sectors
> sect_to_write
)
2576 sectors
= sect_to_write
;
2577 /* Write at 'sector' for 'sectors' */
2578 wbio
= bio_clone_fast(bio
, GFP_NOIO
, mddev
->bio_set
);
2579 bio_trim(wbio
, sector
- bio
->bi_iter
.bi_sector
, sectors
);
2580 wsector
= r10_bio
->devs
[i
].addr
+ (sector
- r10_bio
->sector
);
2581 wbio
->bi_iter
.bi_sector
= wsector
+
2582 choose_data_offset(r10_bio
, rdev
);
2583 bio_set_dev(wbio
, rdev
->bdev
);
2584 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2586 if (submit_bio_wait(wbio
) < 0)
2588 ok
= rdev_set_badblocks(rdev
, wsector
,
2593 sect_to_write
-= sectors
;
2595 sectors
= block_sectors
;
2600 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2602 int slot
= r10_bio
->read_slot
;
2604 struct r10conf
*conf
= mddev
->private;
2605 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2607 /* we got a read error. Maybe the drive is bad. Maybe just
2608 * the block and we can fix it.
2609 * We freeze all other IO, and try reading the block from
2610 * other devices. When we find one, we re-write
2611 * and check it that fixes the read error.
2612 * This is all done synchronously while the array is
2615 bio
= r10_bio
->devs
[slot
].bio
;
2617 r10_bio
->devs
[slot
].bio
= NULL
;
2620 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2621 else if (!test_bit(FailFast
, &rdev
->flags
)) {
2622 freeze_array(conf
, 1);
2623 fix_read_error(conf
, mddev
, r10_bio
);
2624 unfreeze_array(conf
);
2626 md_error(mddev
, rdev
);
2628 rdev_dec_pending(rdev
, mddev
);
2629 allow_barrier(conf
);
2631 raid10_read_request(mddev
, r10_bio
->master_bio
, r10_bio
);
2634 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2636 /* Some sort of write request has finished and it
2637 * succeeded in writing where we thought there was a
2638 * bad block. So forget the bad block.
2639 * Or possibly if failed and we need to record
2643 struct md_rdev
*rdev
;
2645 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2646 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2647 for (m
= 0; m
< conf
->copies
; m
++) {
2648 int dev
= r10_bio
->devs
[m
].devnum
;
2649 rdev
= conf
->mirrors
[dev
].rdev
;
2650 if (r10_bio
->devs
[m
].bio
== NULL
||
2651 r10_bio
->devs
[m
].bio
->bi_end_io
== NULL
)
2653 if (!r10_bio
->devs
[m
].bio
->bi_status
) {
2654 rdev_clear_badblocks(
2656 r10_bio
->devs
[m
].addr
,
2657 r10_bio
->sectors
, 0);
2659 if (!rdev_set_badblocks(
2661 r10_bio
->devs
[m
].addr
,
2662 r10_bio
->sectors
, 0))
2663 md_error(conf
->mddev
, rdev
);
2665 rdev
= conf
->mirrors
[dev
].replacement
;
2666 if (r10_bio
->devs
[m
].repl_bio
== NULL
||
2667 r10_bio
->devs
[m
].repl_bio
->bi_end_io
== NULL
)
2670 if (!r10_bio
->devs
[m
].repl_bio
->bi_status
) {
2671 rdev_clear_badblocks(
2673 r10_bio
->devs
[m
].addr
,
2674 r10_bio
->sectors
, 0);
2676 if (!rdev_set_badblocks(
2678 r10_bio
->devs
[m
].addr
,
2679 r10_bio
->sectors
, 0))
2680 md_error(conf
->mddev
, rdev
);
2686 for (m
= 0; m
< conf
->copies
; m
++) {
2687 int dev
= r10_bio
->devs
[m
].devnum
;
2688 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2689 rdev
= conf
->mirrors
[dev
].rdev
;
2690 if (bio
== IO_MADE_GOOD
) {
2691 rdev_clear_badblocks(
2693 r10_bio
->devs
[m
].addr
,
2694 r10_bio
->sectors
, 0);
2695 rdev_dec_pending(rdev
, conf
->mddev
);
2696 } else if (bio
!= NULL
&& bio
->bi_status
) {
2698 if (!narrow_write_error(r10_bio
, m
)) {
2699 md_error(conf
->mddev
, rdev
);
2700 set_bit(R10BIO_Degraded
,
2703 rdev_dec_pending(rdev
, conf
->mddev
);
2705 bio
= r10_bio
->devs
[m
].repl_bio
;
2706 rdev
= conf
->mirrors
[dev
].replacement
;
2707 if (rdev
&& bio
== IO_MADE_GOOD
) {
2708 rdev_clear_badblocks(
2710 r10_bio
->devs
[m
].addr
,
2711 r10_bio
->sectors
, 0);
2712 rdev_dec_pending(rdev
, conf
->mddev
);
2716 spin_lock_irq(&conf
->device_lock
);
2717 list_add(&r10_bio
->retry_list
, &conf
->bio_end_io_list
);
2719 spin_unlock_irq(&conf
->device_lock
);
2721 * In case freeze_array() is waiting for condition
2722 * nr_pending == nr_queued + extra to be true.
2724 wake_up(&conf
->wait_barrier
);
2725 md_wakeup_thread(conf
->mddev
->thread
);
2727 if (test_bit(R10BIO_WriteError
,
2729 close_write(r10_bio
);
2730 raid_end_bio_io(r10_bio
);
2735 static void raid10d(struct md_thread
*thread
)
2737 struct mddev
*mddev
= thread
->mddev
;
2738 struct r10bio
*r10_bio
;
2739 unsigned long flags
;
2740 struct r10conf
*conf
= mddev
->private;
2741 struct list_head
*head
= &conf
->retry_list
;
2742 struct blk_plug plug
;
2744 md_check_recovery(mddev
);
2746 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2747 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
2749 spin_lock_irqsave(&conf
->device_lock
, flags
);
2750 if (!test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
2751 while (!list_empty(&conf
->bio_end_io_list
)) {
2752 list_move(conf
->bio_end_io_list
.prev
, &tmp
);
2756 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2757 while (!list_empty(&tmp
)) {
2758 r10_bio
= list_first_entry(&tmp
, struct r10bio
,
2760 list_del(&r10_bio
->retry_list
);
2761 if (mddev
->degraded
)
2762 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
2764 if (test_bit(R10BIO_WriteError
,
2766 close_write(r10_bio
);
2767 raid_end_bio_io(r10_bio
);
2771 blk_start_plug(&plug
);
2774 flush_pending_writes(conf
);
2776 spin_lock_irqsave(&conf
->device_lock
, flags
);
2777 if (list_empty(head
)) {
2778 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2781 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2782 list_del(head
->prev
);
2784 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2786 mddev
= r10_bio
->mddev
;
2787 conf
= mddev
->private;
2788 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2789 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2790 handle_write_completed(conf
, r10_bio
);
2791 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2792 reshape_request_write(mddev
, r10_bio
);
2793 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2794 sync_request_write(mddev
, r10_bio
);
2795 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2796 recovery_request_write(mddev
, r10_bio
);
2797 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2798 handle_read_error(mddev
, r10_bio
);
2803 if (mddev
->sb_flags
& ~(1<<MD_SB_CHANGE_PENDING
))
2804 md_check_recovery(mddev
);
2806 blk_finish_plug(&plug
);
2809 static int init_resync(struct r10conf
*conf
)
2814 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2815 BUG_ON(conf
->r10buf_pool
);
2816 conf
->have_replacement
= 0;
2817 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2818 if (conf
->mirrors
[i
].replacement
)
2819 conf
->have_replacement
= 1;
2820 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2821 if (!conf
->r10buf_pool
)
2823 conf
->next_resync
= 0;
2827 static struct r10bio
*raid10_alloc_init_r10buf(struct r10conf
*conf
)
2829 struct r10bio
*r10bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2830 struct rsync_pages
*rp
;
2835 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
) ||
2836 test_bit(MD_RECOVERY_RESHAPE
, &conf
->mddev
->recovery
))
2837 nalloc
= conf
->copies
; /* resync */
2839 nalloc
= 2; /* recovery */
2841 for (i
= 0; i
< nalloc
; i
++) {
2842 bio
= r10bio
->devs
[i
].bio
;
2843 rp
= bio
->bi_private
;
2845 bio
->bi_private
= rp
;
2846 bio
= r10bio
->devs
[i
].repl_bio
;
2848 rp
= bio
->bi_private
;
2850 bio
->bi_private
= rp
;
2857 * Set cluster_sync_high since we need other nodes to add the
2858 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2860 static void raid10_set_cluster_sync_high(struct r10conf
*conf
)
2862 sector_t window_size
;
2863 int extra_chunk
, chunks
;
2866 * First, here we define "stripe" as a unit which across
2867 * all member devices one time, so we get chunks by use
2868 * raid_disks / near_copies. Otherwise, if near_copies is
2869 * close to raid_disks, then resync window could increases
2870 * linearly with the increase of raid_disks, which means
2871 * we will suspend a really large IO window while it is not
2872 * necessary. If raid_disks is not divisible by near_copies,
2873 * an extra chunk is needed to ensure the whole "stripe" is
2877 chunks
= conf
->geo
.raid_disks
/ conf
->geo
.near_copies
;
2878 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
== 0)
2882 window_size
= (chunks
+ extra_chunk
) * conf
->mddev
->chunk_sectors
;
2885 * At least use a 32M window to align with raid1's resync window
2887 window_size
= (CLUSTER_RESYNC_WINDOW_SECTORS
> window_size
) ?
2888 CLUSTER_RESYNC_WINDOW_SECTORS
: window_size
;
2890 conf
->cluster_sync_high
= conf
->cluster_sync_low
+ window_size
;
2894 * perform a "sync" on one "block"
2896 * We need to make sure that no normal I/O request - particularly write
2897 * requests - conflict with active sync requests.
2899 * This is achieved by tracking pending requests and a 'barrier' concept
2900 * that can be installed to exclude normal IO requests.
2902 * Resync and recovery are handled very differently.
2903 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2905 * For resync, we iterate over virtual addresses, read all copies,
2906 * and update if there are differences. If only one copy is live,
2908 * For recovery, we iterate over physical addresses, read a good
2909 * value for each non-in_sync drive, and over-write.
2911 * So, for recovery we may have several outstanding complex requests for a
2912 * given address, one for each out-of-sync device. We model this by allocating
2913 * a number of r10_bio structures, one for each out-of-sync device.
2914 * As we setup these structures, we collect all bio's together into a list
2915 * which we then process collectively to add pages, and then process again
2916 * to pass to generic_make_request.
2918 * The r10_bio structures are linked using a borrowed master_bio pointer.
2919 * This link is counted in ->remaining. When the r10_bio that points to NULL
2920 * has its remaining count decremented to 0, the whole complex operation
2925 static sector_t
raid10_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2928 struct r10conf
*conf
= mddev
->private;
2929 struct r10bio
*r10_bio
;
2930 struct bio
*biolist
= NULL
, *bio
;
2931 sector_t max_sector
, nr_sectors
;
2934 sector_t sync_blocks
;
2935 sector_t sectors_skipped
= 0;
2936 int chunks_skipped
= 0;
2937 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2940 if (!conf
->r10buf_pool
)
2941 if (init_resync(conf
))
2945 * Allow skipping a full rebuild for incremental assembly
2946 * of a clean array, like RAID1 does.
2948 if (mddev
->bitmap
== NULL
&&
2949 mddev
->recovery_cp
== MaxSector
&&
2950 mddev
->reshape_position
== MaxSector
&&
2951 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2952 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2953 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2954 conf
->fullsync
== 0) {
2956 return mddev
->dev_sectors
- sector_nr
;
2960 max_sector
= mddev
->dev_sectors
;
2961 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2962 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2963 max_sector
= mddev
->resync_max_sectors
;
2964 if (sector_nr
>= max_sector
) {
2965 conf
->cluster_sync_low
= 0;
2966 conf
->cluster_sync_high
= 0;
2968 /* If we aborted, we need to abort the
2969 * sync on the 'current' bitmap chucks (there can
2970 * be several when recovering multiple devices).
2971 * as we may have started syncing it but not finished.
2972 * We can find the current address in
2973 * mddev->curr_resync, but for recovery,
2974 * we need to convert that to several
2975 * virtual addresses.
2977 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2983 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2984 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2985 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2987 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2989 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2990 bitmap_end_sync(mddev
->bitmap
, sect
,
2994 /* completed sync */
2995 if ((!mddev
->bitmap
|| conf
->fullsync
)
2996 && conf
->have_replacement
2997 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2998 /* Completed a full sync so the replacements
2999 * are now fully recovered.
3002 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3003 struct md_rdev
*rdev
=
3004 rcu_dereference(conf
->mirrors
[i
].replacement
);
3006 rdev
->recovery_offset
= MaxSector
;
3012 bitmap_close_sync(mddev
->bitmap
);
3015 return sectors_skipped
;
3018 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
3019 return reshape_request(mddev
, sector_nr
, skipped
);
3021 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
3022 /* if there has been nothing to do on any drive,
3023 * then there is nothing to do at all..
3026 return (max_sector
- sector_nr
) + sectors_skipped
;
3029 if (max_sector
> mddev
->resync_max
)
3030 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
3032 /* make sure whole request will fit in a chunk - if chunks
3035 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
3036 max_sector
> (sector_nr
| chunk_mask
))
3037 max_sector
= (sector_nr
| chunk_mask
) + 1;
3040 * If there is non-resync activity waiting for a turn, then let it
3041 * though before starting on this new sync request.
3043 if (conf
->nr_waiting
)
3044 schedule_timeout_uninterruptible(1);
3046 /* Again, very different code for resync and recovery.
3047 * Both must result in an r10bio with a list of bios that
3048 * have bi_end_io, bi_sector, bi_disk set,
3049 * and bi_private set to the r10bio.
3050 * For recovery, we may actually create several r10bios
3051 * with 2 bios in each, that correspond to the bios in the main one.
3052 * In this case, the subordinate r10bios link back through a
3053 * borrowed master_bio pointer, and the counter in the master
3054 * includes a ref from each subordinate.
3056 /* First, we decide what to do and set ->bi_end_io
3057 * To end_sync_read if we want to read, and
3058 * end_sync_write if we will want to write.
3061 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
3062 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3063 /* recovery... the complicated one */
3067 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
3073 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
3074 struct md_rdev
*mrdev
, *mreplace
;
3077 mrdev
= rcu_dereference(mirror
->rdev
);
3078 mreplace
= rcu_dereference(mirror
->replacement
);
3080 if ((mrdev
== NULL
||
3081 test_bit(Faulty
, &mrdev
->flags
) ||
3082 test_bit(In_sync
, &mrdev
->flags
)) &&
3083 (mreplace
== NULL
||
3084 test_bit(Faulty
, &mreplace
->flags
))) {
3090 /* want to reconstruct this device */
3092 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3093 if (sect
>= mddev
->resync_max_sectors
) {
3094 /* last stripe is not complete - don't
3095 * try to recover this sector.
3100 if (mreplace
&& test_bit(Faulty
, &mreplace
->flags
))
3102 /* Unless we are doing a full sync, or a replacement
3103 * we only need to recover the block if it is set in
3106 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3108 if (sync_blocks
< max_sync
)
3109 max_sync
= sync_blocks
;
3113 /* yep, skip the sync_blocks here, but don't assume
3114 * that there will never be anything to do here
3116 chunks_skipped
= -1;
3120 atomic_inc(&mrdev
->nr_pending
);
3122 atomic_inc(&mreplace
->nr_pending
);
3125 r10_bio
= raid10_alloc_init_r10buf(conf
);
3127 raise_barrier(conf
, rb2
!= NULL
);
3128 atomic_set(&r10_bio
->remaining
, 0);
3130 r10_bio
->master_bio
= (struct bio
*)rb2
;
3132 atomic_inc(&rb2
->remaining
);
3133 r10_bio
->mddev
= mddev
;
3134 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3135 r10_bio
->sector
= sect
;
3137 raid10_find_phys(conf
, r10_bio
);
3139 /* Need to check if the array will still be
3143 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++) {
3144 struct md_rdev
*rdev
= rcu_dereference(
3145 conf
->mirrors
[j
].rdev
);
3146 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3152 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3153 &sync_blocks
, still_degraded
);
3156 for (j
=0; j
<conf
->copies
;j
++) {
3158 int d
= r10_bio
->devs
[j
].devnum
;
3159 sector_t from_addr
, to_addr
;
3160 struct md_rdev
*rdev
=
3161 rcu_dereference(conf
->mirrors
[d
].rdev
);
3162 sector_t sector
, first_bad
;
3165 !test_bit(In_sync
, &rdev
->flags
))
3167 /* This is where we read from */
3169 sector
= r10_bio
->devs
[j
].addr
;
3171 if (is_badblock(rdev
, sector
, max_sync
,
3172 &first_bad
, &bad_sectors
)) {
3173 if (first_bad
> sector
)
3174 max_sync
= first_bad
- sector
;
3176 bad_sectors
-= (sector
3178 if (max_sync
> bad_sectors
)
3179 max_sync
= bad_sectors
;
3183 bio
= r10_bio
->devs
[0].bio
;
3184 bio
->bi_next
= biolist
;
3186 bio
->bi_end_io
= end_sync_read
;
3187 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3188 if (test_bit(FailFast
, &rdev
->flags
))
3189 bio
->bi_opf
|= MD_FAILFAST
;
3190 from_addr
= r10_bio
->devs
[j
].addr
;
3191 bio
->bi_iter
.bi_sector
= from_addr
+
3193 bio_set_dev(bio
, rdev
->bdev
);
3194 atomic_inc(&rdev
->nr_pending
);
3195 /* and we write to 'i' (if not in_sync) */
3197 for (k
=0; k
<conf
->copies
; k
++)
3198 if (r10_bio
->devs
[k
].devnum
== i
)
3200 BUG_ON(k
== conf
->copies
);
3201 to_addr
= r10_bio
->devs
[k
].addr
;
3202 r10_bio
->devs
[0].devnum
= d
;
3203 r10_bio
->devs
[0].addr
= from_addr
;
3204 r10_bio
->devs
[1].devnum
= i
;
3205 r10_bio
->devs
[1].addr
= to_addr
;
3207 if (!test_bit(In_sync
, &mrdev
->flags
)) {
3208 bio
= r10_bio
->devs
[1].bio
;
3209 bio
->bi_next
= biolist
;
3211 bio
->bi_end_io
= end_sync_write
;
3212 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3213 bio
->bi_iter
.bi_sector
= to_addr
3214 + mrdev
->data_offset
;
3215 bio_set_dev(bio
, mrdev
->bdev
);
3216 atomic_inc(&r10_bio
->remaining
);
3218 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3220 /* and maybe write to replacement */
3221 bio
= r10_bio
->devs
[1].repl_bio
;
3223 bio
->bi_end_io
= NULL
;
3224 /* Note: if mreplace != NULL, then bio
3225 * cannot be NULL as r10buf_pool_alloc will
3226 * have allocated it.
3227 * So the second test here is pointless.
3228 * But it keeps semantic-checkers happy, and
3229 * this comment keeps human reviewers
3232 if (mreplace
== NULL
|| bio
== NULL
||
3233 test_bit(Faulty
, &mreplace
->flags
))
3235 bio
->bi_next
= biolist
;
3237 bio
->bi_end_io
= end_sync_write
;
3238 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3239 bio
->bi_iter
.bi_sector
= to_addr
+
3240 mreplace
->data_offset
;
3241 bio_set_dev(bio
, mreplace
->bdev
);
3242 atomic_inc(&r10_bio
->remaining
);
3246 if (j
== conf
->copies
) {
3247 /* Cannot recover, so abort the recovery or
3248 * record a bad block */
3250 /* problem is that there are bad blocks
3251 * on other device(s)
3254 for (k
= 0; k
< conf
->copies
; k
++)
3255 if (r10_bio
->devs
[k
].devnum
== i
)
3257 if (!test_bit(In_sync
,
3259 && !rdev_set_badblocks(
3261 r10_bio
->devs
[k
].addr
,
3265 !rdev_set_badblocks(
3267 r10_bio
->devs
[k
].addr
,
3272 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3274 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3276 mirror
->recovery_disabled
3277 = mddev
->recovery_disabled
;
3281 atomic_dec(&rb2
->remaining
);
3283 rdev_dec_pending(mrdev
, mddev
);
3285 rdev_dec_pending(mreplace
, mddev
);
3288 rdev_dec_pending(mrdev
, mddev
);
3290 rdev_dec_pending(mreplace
, mddev
);
3291 if (r10_bio
->devs
[0].bio
->bi_opf
& MD_FAILFAST
) {
3292 /* Only want this if there is elsewhere to
3293 * read from. 'j' is currently the first
3297 for (; j
< conf
->copies
; j
++) {
3298 int d
= r10_bio
->devs
[j
].devnum
;
3299 if (conf
->mirrors
[d
].rdev
&&
3301 &conf
->mirrors
[d
].rdev
->flags
))
3305 r10_bio
->devs
[0].bio
->bi_opf
3309 if (biolist
== NULL
) {
3311 struct r10bio
*rb2
= r10_bio
;
3312 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3313 rb2
->master_bio
= NULL
;
3319 /* resync. Schedule a read for every block at this virt offset */
3323 * Since curr_resync_completed could probably not update in
3324 * time, and we will set cluster_sync_low based on it.
3325 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3326 * safety reason, which ensures curr_resync_completed is
3327 * updated in bitmap_cond_end_sync.
3329 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
,
3330 mddev_is_clustered(mddev
) &&
3331 (sector_nr
+ 2 * RESYNC_SECTORS
>
3332 conf
->cluster_sync_high
));
3334 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3335 &sync_blocks
, mddev
->degraded
) &&
3336 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3337 &mddev
->recovery
)) {
3338 /* We can skip this block */
3340 return sync_blocks
+ sectors_skipped
;
3342 if (sync_blocks
< max_sync
)
3343 max_sync
= sync_blocks
;
3344 r10_bio
= raid10_alloc_init_r10buf(conf
);
3347 r10_bio
->mddev
= mddev
;
3348 atomic_set(&r10_bio
->remaining
, 0);
3349 raise_barrier(conf
, 0);
3350 conf
->next_resync
= sector_nr
;
3352 r10_bio
->master_bio
= NULL
;
3353 r10_bio
->sector
= sector_nr
;
3354 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3355 raid10_find_phys(conf
, r10_bio
);
3356 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3358 for (i
= 0; i
< conf
->copies
; i
++) {
3359 int d
= r10_bio
->devs
[i
].devnum
;
3360 sector_t first_bad
, sector
;
3362 struct md_rdev
*rdev
;
3364 if (r10_bio
->devs
[i
].repl_bio
)
3365 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3367 bio
= r10_bio
->devs
[i
].bio
;
3368 bio
->bi_status
= BLK_STS_IOERR
;
3370 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
3371 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3375 sector
= r10_bio
->devs
[i
].addr
;
3376 if (is_badblock(rdev
, sector
, max_sync
,
3377 &first_bad
, &bad_sectors
)) {
3378 if (first_bad
> sector
)
3379 max_sync
= first_bad
- sector
;
3381 bad_sectors
-= (sector
- first_bad
);
3382 if (max_sync
> bad_sectors
)
3383 max_sync
= bad_sectors
;
3388 atomic_inc(&rdev
->nr_pending
);
3389 atomic_inc(&r10_bio
->remaining
);
3390 bio
->bi_next
= biolist
;
3392 bio
->bi_end_io
= end_sync_read
;
3393 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3394 if (test_bit(FailFast
, &rdev
->flags
))
3395 bio
->bi_opf
|= MD_FAILFAST
;
3396 bio
->bi_iter
.bi_sector
= sector
+ rdev
->data_offset
;
3397 bio_set_dev(bio
, rdev
->bdev
);
3400 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
3401 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3405 atomic_inc(&rdev
->nr_pending
);
3407 /* Need to set up for writing to the replacement */
3408 bio
= r10_bio
->devs
[i
].repl_bio
;
3409 bio
->bi_status
= BLK_STS_IOERR
;
3411 sector
= r10_bio
->devs
[i
].addr
;
3412 bio
->bi_next
= biolist
;
3414 bio
->bi_end_io
= end_sync_write
;
3415 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3416 if (test_bit(FailFast
, &rdev
->flags
))
3417 bio
->bi_opf
|= MD_FAILFAST
;
3418 bio
->bi_iter
.bi_sector
= sector
+ rdev
->data_offset
;
3419 bio_set_dev(bio
, rdev
->bdev
);
3425 for (i
=0; i
<conf
->copies
; i
++) {
3426 int d
= r10_bio
->devs
[i
].devnum
;
3427 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3428 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3430 if (r10_bio
->devs
[i
].repl_bio
&&
3431 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3433 conf
->mirrors
[d
].replacement
,
3443 if (sector_nr
+ max_sync
< max_sector
)
3444 max_sector
= sector_nr
+ max_sync
;
3447 int len
= PAGE_SIZE
;
3448 if (sector_nr
+ (len
>>9) > max_sector
)
3449 len
= (max_sector
- sector_nr
) << 9;
3452 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3453 struct resync_pages
*rp
= get_resync_pages(bio
);
3454 page
= resync_fetch_page(rp
, page_idx
);
3456 * won't fail because the vec table is big enough
3457 * to hold all these pages
3459 bio_add_page(bio
, page
, len
, 0);
3461 nr_sectors
+= len
>>9;
3462 sector_nr
+= len
>>9;
3463 } while (++page_idx
< RESYNC_PAGES
);
3464 r10_bio
->sectors
= nr_sectors
;
3466 if (mddev_is_clustered(mddev
) &&
3467 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3468 /* It is resync not recovery */
3469 if (conf
->cluster_sync_high
< sector_nr
+ nr_sectors
) {
3470 conf
->cluster_sync_low
= mddev
->curr_resync_completed
;
3471 raid10_set_cluster_sync_high(conf
);
3472 /* Send resync message */
3473 md_cluster_ops
->resync_info_update(mddev
,
3474 conf
->cluster_sync_low
,
3475 conf
->cluster_sync_high
);
3477 } else if (mddev_is_clustered(mddev
)) {
3478 /* This is recovery not resync */
3479 sector_t sect_va1
, sect_va2
;
3480 bool broadcast_msg
= false;
3482 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3484 * sector_nr is a device address for recovery, so we
3485 * need translate it to array address before compare
3486 * with cluster_sync_high.
3488 sect_va1
= raid10_find_virt(conf
, sector_nr
, i
);
3490 if (conf
->cluster_sync_high
< sect_va1
+ nr_sectors
) {
3491 broadcast_msg
= true;
3493 * curr_resync_completed is similar as
3494 * sector_nr, so make the translation too.
3496 sect_va2
= raid10_find_virt(conf
,
3497 mddev
->curr_resync_completed
, i
);
3499 if (conf
->cluster_sync_low
== 0 ||
3500 conf
->cluster_sync_low
> sect_va2
)
3501 conf
->cluster_sync_low
= sect_va2
;
3504 if (broadcast_msg
) {
3505 raid10_set_cluster_sync_high(conf
);
3506 md_cluster_ops
->resync_info_update(mddev
,
3507 conf
->cluster_sync_low
,
3508 conf
->cluster_sync_high
);
3514 biolist
= biolist
->bi_next
;
3516 bio
->bi_next
= NULL
;
3517 r10_bio
= get_resync_r10bio(bio
);
3518 r10_bio
->sectors
= nr_sectors
;
3520 if (bio
->bi_end_io
== end_sync_read
) {
3521 md_sync_acct_bio(bio
, nr_sectors
);
3523 generic_make_request(bio
);
3527 if (sectors_skipped
)
3528 /* pretend they weren't skipped, it makes
3529 * no important difference in this case
3531 md_done_sync(mddev
, sectors_skipped
, 1);
3533 return sectors_skipped
+ nr_sectors
;
3535 /* There is nowhere to write, so all non-sync
3536 * drives must be failed or in resync, all drives
3537 * have a bad block, so try the next chunk...
3539 if (sector_nr
+ max_sync
< max_sector
)
3540 max_sector
= sector_nr
+ max_sync
;
3542 sectors_skipped
+= (max_sector
- sector_nr
);
3544 sector_nr
= max_sector
;
3549 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3552 struct r10conf
*conf
= mddev
->private;
3555 raid_disks
= min(conf
->geo
.raid_disks
,
3556 conf
->prev
.raid_disks
);
3558 sectors
= conf
->dev_sectors
;
3560 size
= sectors
>> conf
->geo
.chunk_shift
;
3561 sector_div(size
, conf
->geo
.far_copies
);
3562 size
= size
* raid_disks
;
3563 sector_div(size
, conf
->geo
.near_copies
);
3565 return size
<< conf
->geo
.chunk_shift
;
3568 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3570 /* Calculate the number of sectors-per-device that will
3571 * actually be used, and set conf->dev_sectors and
3575 size
= size
>> conf
->geo
.chunk_shift
;
3576 sector_div(size
, conf
->geo
.far_copies
);
3577 size
= size
* conf
->geo
.raid_disks
;
3578 sector_div(size
, conf
->geo
.near_copies
);
3579 /* 'size' is now the number of chunks in the array */
3580 /* calculate "used chunks per device" */
3581 size
= size
* conf
->copies
;
3583 /* We need to round up when dividing by raid_disks to
3584 * get the stride size.
3586 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3588 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3590 if (conf
->geo
.far_offset
)
3591 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3593 sector_div(size
, conf
->geo
.far_copies
);
3594 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3598 enum geo_type
{geo_new
, geo_old
, geo_start
};
3599 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3602 int layout
, chunk
, disks
;
3605 layout
= mddev
->layout
;
3606 chunk
= mddev
->chunk_sectors
;
3607 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3610 layout
= mddev
->new_layout
;
3611 chunk
= mddev
->new_chunk_sectors
;
3612 disks
= mddev
->raid_disks
;
3614 default: /* avoid 'may be unused' warnings */
3615 case geo_start
: /* new when starting reshape - raid_disks not
3617 layout
= mddev
->new_layout
;
3618 chunk
= mddev
->new_chunk_sectors
;
3619 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3624 if (chunk
< (PAGE_SIZE
>> 9) ||
3625 !is_power_of_2(chunk
))
3628 fc
= (layout
>> 8) & 255;
3629 fo
= layout
& (1<<16);
3630 geo
->raid_disks
= disks
;
3631 geo
->near_copies
= nc
;
3632 geo
->far_copies
= fc
;
3633 geo
->far_offset
= fo
;
3634 switch (layout
>> 17) {
3635 case 0: /* original layout. simple but not always optimal */
3636 geo
->far_set_size
= disks
;
3638 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3639 * actually using this, but leave code here just in case.*/
3640 geo
->far_set_size
= disks
/fc
;
3641 WARN(geo
->far_set_size
< fc
,
3642 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3644 case 2: /* "improved" layout fixed to match documentation */
3645 geo
->far_set_size
= fc
* nc
;
3647 default: /* Not a valid layout */
3650 geo
->chunk_mask
= chunk
- 1;
3651 geo
->chunk_shift
= ffz(~chunk
);
3655 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3657 struct r10conf
*conf
= NULL
;
3662 copies
= setup_geo(&geo
, mddev
, geo_new
);
3665 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3666 mdname(mddev
), PAGE_SIZE
);
3670 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3671 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3672 mdname(mddev
), mddev
->new_layout
);
3677 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3681 /* FIXME calc properly */
3682 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3683 max(0,-mddev
->delta_disks
)),
3688 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3693 conf
->copies
= copies
;
3694 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3695 r10bio_pool_free
, conf
);
3696 if (!conf
->r10bio_pool
)
3699 conf
->bio_split
= bioset_create(BIO_POOL_SIZE
, 0, 0);
3700 if (!conf
->bio_split
)
3703 calc_sectors(conf
, mddev
->dev_sectors
);
3704 if (mddev
->reshape_position
== MaxSector
) {
3705 conf
->prev
= conf
->geo
;
3706 conf
->reshape_progress
= MaxSector
;
3708 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3712 conf
->reshape_progress
= mddev
->reshape_position
;
3713 if (conf
->prev
.far_offset
)
3714 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3716 /* far_copies must be 1 */
3717 conf
->prev
.stride
= conf
->dev_sectors
;
3719 conf
->reshape_safe
= conf
->reshape_progress
;
3720 spin_lock_init(&conf
->device_lock
);
3721 INIT_LIST_HEAD(&conf
->retry_list
);
3722 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
3724 spin_lock_init(&conf
->resync_lock
);
3725 init_waitqueue_head(&conf
->wait_barrier
);
3726 atomic_set(&conf
->nr_pending
, 0);
3728 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3732 conf
->mddev
= mddev
;
3737 mempool_destroy(conf
->r10bio_pool
);
3738 kfree(conf
->mirrors
);
3739 safe_put_page(conf
->tmppage
);
3740 if (conf
->bio_split
)
3741 bioset_free(conf
->bio_split
);
3744 return ERR_PTR(err
);
3747 static int raid10_run(struct mddev
*mddev
)
3749 struct r10conf
*conf
;
3750 int i
, disk_idx
, chunk_size
;
3751 struct raid10_info
*disk
;
3752 struct md_rdev
*rdev
;
3754 sector_t min_offset_diff
= 0;
3756 bool discard_supported
= false;
3758 if (mddev_init_writes_pending(mddev
) < 0)
3761 if (mddev
->private == NULL
) {
3762 conf
= setup_conf(mddev
);
3764 return PTR_ERR(conf
);
3765 mddev
->private = conf
;
3767 conf
= mddev
->private;
3771 if (mddev_is_clustered(conf
->mddev
)) {
3774 fc
= (mddev
->layout
>> 8) & 255;
3775 fo
= mddev
->layout
& (1<<16);
3776 if (fc
> 1 || fo
> 0) {
3777 pr_err("only near layout is supported by clustered"
3783 mddev
->thread
= conf
->thread
;
3784 conf
->thread
= NULL
;
3786 chunk_size
= mddev
->chunk_sectors
<< 9;
3788 blk_queue_max_discard_sectors(mddev
->queue
,
3789 mddev
->chunk_sectors
);
3790 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3791 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
3792 blk_queue_io_min(mddev
->queue
, chunk_size
);
3793 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3794 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3796 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3797 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3800 rdev_for_each(rdev
, mddev
) {
3803 disk_idx
= rdev
->raid_disk
;
3806 if (disk_idx
>= conf
->geo
.raid_disks
&&
3807 disk_idx
>= conf
->prev
.raid_disks
)
3809 disk
= conf
->mirrors
+ disk_idx
;
3811 if (test_bit(Replacement
, &rdev
->flags
)) {
3812 if (disk
->replacement
)
3814 disk
->replacement
= rdev
;
3820 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3821 if (!mddev
->reshape_backwards
)
3825 if (first
|| diff
< min_offset_diff
)
3826 min_offset_diff
= diff
;
3829 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3830 rdev
->data_offset
<< 9);
3832 disk
->head_position
= 0;
3834 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3835 discard_supported
= true;
3840 if (discard_supported
)
3841 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3844 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3847 /* need to check that every block has at least one working mirror */
3848 if (!enough(conf
, -1)) {
3849 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3854 if (conf
->reshape_progress
!= MaxSector
) {
3855 /* must ensure that shape change is supported */
3856 if (conf
->geo
.far_copies
!= 1 &&
3857 conf
->geo
.far_offset
== 0)
3859 if (conf
->prev
.far_copies
!= 1 &&
3860 conf
->prev
.far_offset
== 0)
3864 mddev
->degraded
= 0;
3866 i
< conf
->geo
.raid_disks
3867 || i
< conf
->prev
.raid_disks
;
3870 disk
= conf
->mirrors
+ i
;
3872 if (!disk
->rdev
&& disk
->replacement
) {
3873 /* The replacement is all we have - use it */
3874 disk
->rdev
= disk
->replacement
;
3875 disk
->replacement
= NULL
;
3876 clear_bit(Replacement
, &disk
->rdev
->flags
);
3880 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3881 disk
->head_position
= 0;
3884 disk
->rdev
->saved_raid_disk
< 0)
3888 if (disk
->replacement
&&
3889 !test_bit(In_sync
, &disk
->replacement
->flags
) &&
3890 disk
->replacement
->saved_raid_disk
< 0) {
3894 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3897 if (mddev
->recovery_cp
!= MaxSector
)
3898 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3900 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3901 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3902 conf
->geo
.raid_disks
);
3904 * Ok, everything is just fine now
3906 mddev
->dev_sectors
= conf
->dev_sectors
;
3907 size
= raid10_size(mddev
, 0, 0);
3908 md_set_array_sectors(mddev
, size
);
3909 mddev
->resync_max_sectors
= size
;
3910 set_bit(MD_FAILFAST_SUPPORTED
, &mddev
->flags
);
3913 int stripe
= conf
->geo
.raid_disks
*
3914 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3916 /* Calculate max read-ahead size.
3917 * We need to readahead at least twice a whole stripe....
3920 stripe
/= conf
->geo
.near_copies
;
3921 if (mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
3922 mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
3925 if (md_integrity_register(mddev
))
3928 if (conf
->reshape_progress
!= MaxSector
) {
3929 unsigned long before_length
, after_length
;
3931 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3932 conf
->prev
.far_copies
);
3933 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3934 conf
->geo
.far_copies
);
3936 if (max(before_length
, after_length
) > min_offset_diff
) {
3937 /* This cannot work */
3938 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3941 conf
->offset_diff
= min_offset_diff
;
3943 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3944 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3945 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3946 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3947 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3949 if (!mddev
->sync_thread
)
3956 md_unregister_thread(&mddev
->thread
);
3957 mempool_destroy(conf
->r10bio_pool
);
3958 safe_put_page(conf
->tmppage
);
3959 kfree(conf
->mirrors
);
3961 mddev
->private = NULL
;
3966 static void raid10_free(struct mddev
*mddev
, void *priv
)
3968 struct r10conf
*conf
= priv
;
3970 mempool_destroy(conf
->r10bio_pool
);
3971 safe_put_page(conf
->tmppage
);
3972 kfree(conf
->mirrors
);
3973 kfree(conf
->mirrors_old
);
3974 kfree(conf
->mirrors_new
);
3975 if (conf
->bio_split
)
3976 bioset_free(conf
->bio_split
);
3980 static void raid10_quiesce(struct mddev
*mddev
, int quiesce
)
3982 struct r10conf
*conf
= mddev
->private;
3985 raise_barrier(conf
, 0);
3987 lower_barrier(conf
);
3990 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3992 /* Resize of 'far' arrays is not supported.
3993 * For 'near' and 'offset' arrays we can set the
3994 * number of sectors used to be an appropriate multiple
3995 * of the chunk size.
3996 * For 'offset', this is far_copies*chunksize.
3997 * For 'near' the multiplier is the LCM of
3998 * near_copies and raid_disks.
3999 * So if far_copies > 1 && !far_offset, fail.
4000 * Else find LCM(raid_disks, near_copy)*far_copies and
4001 * multiply by chunk_size. Then round to this number.
4002 * This is mostly done by raid10_size()
4004 struct r10conf
*conf
= mddev
->private;
4005 sector_t oldsize
, size
;
4007 if (mddev
->reshape_position
!= MaxSector
)
4010 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
4013 oldsize
= raid10_size(mddev
, 0, 0);
4014 size
= raid10_size(mddev
, sectors
, 0);
4015 if (mddev
->external_size
&&
4016 mddev
->array_sectors
> size
)
4018 if (mddev
->bitmap
) {
4019 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
4023 md_set_array_sectors(mddev
, size
);
4024 if (sectors
> mddev
->dev_sectors
&&
4025 mddev
->recovery_cp
> oldsize
) {
4026 mddev
->recovery_cp
= oldsize
;
4027 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4029 calc_sectors(conf
, sectors
);
4030 mddev
->dev_sectors
= conf
->dev_sectors
;
4031 mddev
->resync_max_sectors
= size
;
4035 static void *raid10_takeover_raid0(struct mddev
*mddev
, sector_t size
, int devs
)
4037 struct md_rdev
*rdev
;
4038 struct r10conf
*conf
;
4040 if (mddev
->degraded
> 0) {
4041 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4043 return ERR_PTR(-EINVAL
);
4045 sector_div(size
, devs
);
4047 /* Set new parameters */
4048 mddev
->new_level
= 10;
4049 /* new layout: far_copies = 1, near_copies = 2 */
4050 mddev
->new_layout
= (1<<8) + 2;
4051 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
4052 mddev
->delta_disks
= mddev
->raid_disks
;
4053 mddev
->raid_disks
*= 2;
4054 /* make sure it will be not marked as dirty */
4055 mddev
->recovery_cp
= MaxSector
;
4056 mddev
->dev_sectors
= size
;
4058 conf
= setup_conf(mddev
);
4059 if (!IS_ERR(conf
)) {
4060 rdev_for_each(rdev
, mddev
)
4061 if (rdev
->raid_disk
>= 0) {
4062 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
4063 rdev
->sectors
= size
;
4071 static void *raid10_takeover(struct mddev
*mddev
)
4073 struct r0conf
*raid0_conf
;
4075 /* raid10 can take over:
4076 * raid0 - providing it has only two drives
4078 if (mddev
->level
== 0) {
4079 /* for raid0 takeover only one zone is supported */
4080 raid0_conf
= mddev
->private;
4081 if (raid0_conf
->nr_strip_zones
> 1) {
4082 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4084 return ERR_PTR(-EINVAL
);
4086 return raid10_takeover_raid0(mddev
,
4087 raid0_conf
->strip_zone
->zone_end
,
4088 raid0_conf
->strip_zone
->nb_dev
);
4090 return ERR_PTR(-EINVAL
);
4093 static int raid10_check_reshape(struct mddev
*mddev
)
4095 /* Called when there is a request to change
4096 * - layout (to ->new_layout)
4097 * - chunk size (to ->new_chunk_sectors)
4098 * - raid_disks (by delta_disks)
4099 * or when trying to restart a reshape that was ongoing.
4101 * We need to validate the request and possibly allocate
4102 * space if that might be an issue later.
4104 * Currently we reject any reshape of a 'far' mode array,
4105 * allow chunk size to change if new is generally acceptable,
4106 * allow raid_disks to increase, and allow
4107 * a switch between 'near' mode and 'offset' mode.
4109 struct r10conf
*conf
= mddev
->private;
4112 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
4115 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
4116 /* mustn't change number of copies */
4118 if (geo
.far_copies
> 1 && !geo
.far_offset
)
4119 /* Cannot switch to 'far' mode */
4122 if (mddev
->array_sectors
& geo
.chunk_mask
)
4123 /* not factor of array size */
4126 if (!enough(conf
, -1))
4129 kfree(conf
->mirrors_new
);
4130 conf
->mirrors_new
= NULL
;
4131 if (mddev
->delta_disks
> 0) {
4132 /* allocate new 'mirrors' list */
4133 conf
->mirrors_new
= kzalloc(
4134 sizeof(struct raid10_info
)
4135 *(mddev
->raid_disks
+
4136 mddev
->delta_disks
),
4138 if (!conf
->mirrors_new
)
4145 * Need to check if array has failed when deciding whether to:
4147 * - remove non-faulty devices
4150 * This determination is simple when no reshape is happening.
4151 * However if there is a reshape, we need to carefully check
4152 * both the before and after sections.
4153 * This is because some failed devices may only affect one
4154 * of the two sections, and some non-in_sync devices may
4155 * be insync in the section most affected by failed devices.
4157 static int calc_degraded(struct r10conf
*conf
)
4159 int degraded
, degraded2
;
4164 /* 'prev' section first */
4165 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4166 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4167 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4169 else if (!test_bit(In_sync
, &rdev
->flags
))
4170 /* When we can reduce the number of devices in
4171 * an array, this might not contribute to
4172 * 'degraded'. It does now.
4177 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4181 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4182 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4183 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4185 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4186 /* If reshape is increasing the number of devices,
4187 * this section has already been recovered, so
4188 * it doesn't contribute to degraded.
4191 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4196 if (degraded2
> degraded
)
4201 static int raid10_start_reshape(struct mddev
*mddev
)
4203 /* A 'reshape' has been requested. This commits
4204 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4205 * This also checks if there are enough spares and adds them
4207 * We currently require enough spares to make the final
4208 * array non-degraded. We also require that the difference
4209 * between old and new data_offset - on each device - is
4210 * enough that we never risk over-writing.
4213 unsigned long before_length
, after_length
;
4214 sector_t min_offset_diff
= 0;
4217 struct r10conf
*conf
= mddev
->private;
4218 struct md_rdev
*rdev
;
4222 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4225 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4228 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4229 conf
->prev
.far_copies
);
4230 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4231 conf
->geo
.far_copies
);
4233 rdev_for_each(rdev
, mddev
) {
4234 if (!test_bit(In_sync
, &rdev
->flags
)
4235 && !test_bit(Faulty
, &rdev
->flags
))
4237 if (rdev
->raid_disk
>= 0) {
4238 long long diff
= (rdev
->new_data_offset
4239 - rdev
->data_offset
);
4240 if (!mddev
->reshape_backwards
)
4244 if (first
|| diff
< min_offset_diff
)
4245 min_offset_diff
= diff
;
4250 if (max(before_length
, after_length
) > min_offset_diff
)
4253 if (spares
< mddev
->delta_disks
)
4256 conf
->offset_diff
= min_offset_diff
;
4257 spin_lock_irq(&conf
->device_lock
);
4258 if (conf
->mirrors_new
) {
4259 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4260 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4262 kfree(conf
->mirrors_old
);
4263 conf
->mirrors_old
= conf
->mirrors
;
4264 conf
->mirrors
= conf
->mirrors_new
;
4265 conf
->mirrors_new
= NULL
;
4267 setup_geo(&conf
->geo
, mddev
, geo_start
);
4269 if (mddev
->reshape_backwards
) {
4270 sector_t size
= raid10_size(mddev
, 0, 0);
4271 if (size
< mddev
->array_sectors
) {
4272 spin_unlock_irq(&conf
->device_lock
);
4273 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4277 mddev
->resync_max_sectors
= size
;
4278 conf
->reshape_progress
= size
;
4280 conf
->reshape_progress
= 0;
4281 conf
->reshape_safe
= conf
->reshape_progress
;
4282 spin_unlock_irq(&conf
->device_lock
);
4284 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4285 ret
= bitmap_resize(mddev
->bitmap
,
4286 raid10_size(mddev
, 0,
4287 conf
->geo
.raid_disks
),
4292 if (mddev
->delta_disks
> 0) {
4293 rdev_for_each(rdev
, mddev
)
4294 if (rdev
->raid_disk
< 0 &&
4295 !test_bit(Faulty
, &rdev
->flags
)) {
4296 if (raid10_add_disk(mddev
, rdev
) == 0) {
4297 if (rdev
->raid_disk
>=
4298 conf
->prev
.raid_disks
)
4299 set_bit(In_sync
, &rdev
->flags
);
4301 rdev
->recovery_offset
= 0;
4303 if (sysfs_link_rdev(mddev
, rdev
))
4304 /* Failure here is OK */;
4306 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4307 && !test_bit(Faulty
, &rdev
->flags
)) {
4308 /* This is a spare that was manually added */
4309 set_bit(In_sync
, &rdev
->flags
);
4312 /* When a reshape changes the number of devices,
4313 * ->degraded is measured against the larger of the
4314 * pre and post numbers.
4316 spin_lock_irq(&conf
->device_lock
);
4317 mddev
->degraded
= calc_degraded(conf
);
4318 spin_unlock_irq(&conf
->device_lock
);
4319 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4320 mddev
->reshape_position
= conf
->reshape_progress
;
4321 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
4323 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4324 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4325 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
4326 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4327 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4329 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4331 if (!mddev
->sync_thread
) {
4335 conf
->reshape_checkpoint
= jiffies
;
4336 md_wakeup_thread(mddev
->sync_thread
);
4337 md_new_event(mddev
);
4341 mddev
->recovery
= 0;
4342 spin_lock_irq(&conf
->device_lock
);
4343 conf
->geo
= conf
->prev
;
4344 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4345 rdev_for_each(rdev
, mddev
)
4346 rdev
->new_data_offset
= rdev
->data_offset
;
4348 conf
->reshape_progress
= MaxSector
;
4349 conf
->reshape_safe
= MaxSector
;
4350 mddev
->reshape_position
= MaxSector
;
4351 spin_unlock_irq(&conf
->device_lock
);
4355 /* Calculate the last device-address that could contain
4356 * any block from the chunk that includes the array-address 's'
4357 * and report the next address.
4358 * i.e. the address returned will be chunk-aligned and after
4359 * any data that is in the chunk containing 's'.
4361 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4363 s
= (s
| geo
->chunk_mask
) + 1;
4364 s
>>= geo
->chunk_shift
;
4365 s
*= geo
->near_copies
;
4366 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4367 s
*= geo
->far_copies
;
4368 s
<<= geo
->chunk_shift
;
4372 /* Calculate the first device-address that could contain
4373 * any block from the chunk that includes the array-address 's'.
4374 * This too will be the start of a chunk
4376 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4378 s
>>= geo
->chunk_shift
;
4379 s
*= geo
->near_copies
;
4380 sector_div(s
, geo
->raid_disks
);
4381 s
*= geo
->far_copies
;
4382 s
<<= geo
->chunk_shift
;
4386 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4389 /* We simply copy at most one chunk (smallest of old and new)
4390 * at a time, possibly less if that exceeds RESYNC_PAGES,
4391 * or we hit a bad block or something.
4392 * This might mean we pause for normal IO in the middle of
4393 * a chunk, but that is not a problem as mddev->reshape_position
4394 * can record any location.
4396 * If we will want to write to a location that isn't
4397 * yet recorded as 'safe' (i.e. in metadata on disk) then
4398 * we need to flush all reshape requests and update the metadata.
4400 * When reshaping forwards (e.g. to more devices), we interpret
4401 * 'safe' as the earliest block which might not have been copied
4402 * down yet. We divide this by previous stripe size and multiply
4403 * by previous stripe length to get lowest device offset that we
4404 * cannot write to yet.
4405 * We interpret 'sector_nr' as an address that we want to write to.
4406 * From this we use last_device_address() to find where we might
4407 * write to, and first_device_address on the 'safe' position.
4408 * If this 'next' write position is after the 'safe' position,
4409 * we must update the metadata to increase the 'safe' position.
4411 * When reshaping backwards, we round in the opposite direction
4412 * and perform the reverse test: next write position must not be
4413 * less than current safe position.
4415 * In all this the minimum difference in data offsets
4416 * (conf->offset_diff - always positive) allows a bit of slack,
4417 * so next can be after 'safe', but not by more than offset_diff
4419 * We need to prepare all the bios here before we start any IO
4420 * to ensure the size we choose is acceptable to all devices.
4421 * The means one for each copy for write-out and an extra one for
4423 * We store the read-in bio in ->master_bio and the others in
4424 * ->devs[x].bio and ->devs[x].repl_bio.
4426 struct r10conf
*conf
= mddev
->private;
4427 struct r10bio
*r10_bio
;
4428 sector_t next
, safe
, last
;
4432 struct md_rdev
*rdev
;
4435 struct bio
*bio
, *read_bio
;
4436 int sectors_done
= 0;
4437 struct page
**pages
;
4439 if (sector_nr
== 0) {
4440 /* If restarting in the middle, skip the initial sectors */
4441 if (mddev
->reshape_backwards
&&
4442 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4443 sector_nr
= (raid10_size(mddev
, 0, 0)
4444 - conf
->reshape_progress
);
4445 } else if (!mddev
->reshape_backwards
&&
4446 conf
->reshape_progress
> 0)
4447 sector_nr
= conf
->reshape_progress
;
4449 mddev
->curr_resync_completed
= sector_nr
;
4450 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4456 /* We don't use sector_nr to track where we are up to
4457 * as that doesn't work well for ->reshape_backwards.
4458 * So just use ->reshape_progress.
4460 if (mddev
->reshape_backwards
) {
4461 /* 'next' is the earliest device address that we might
4462 * write to for this chunk in the new layout
4464 next
= first_dev_address(conf
->reshape_progress
- 1,
4467 /* 'safe' is the last device address that we might read from
4468 * in the old layout after a restart
4470 safe
= last_dev_address(conf
->reshape_safe
- 1,
4473 if (next
+ conf
->offset_diff
< safe
)
4476 last
= conf
->reshape_progress
- 1;
4477 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4478 & conf
->prev
.chunk_mask
);
4479 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4480 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4482 /* 'next' is after the last device address that we
4483 * might write to for this chunk in the new layout
4485 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4487 /* 'safe' is the earliest device address that we might
4488 * read from in the old layout after a restart
4490 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4492 /* Need to update metadata if 'next' might be beyond 'safe'
4493 * as that would possibly corrupt data
4495 if (next
> safe
+ conf
->offset_diff
)
4498 sector_nr
= conf
->reshape_progress
;
4499 last
= sector_nr
| (conf
->geo
.chunk_mask
4500 & conf
->prev
.chunk_mask
);
4502 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4503 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4507 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4508 /* Need to update reshape_position in metadata */
4510 mddev
->reshape_position
= conf
->reshape_progress
;
4511 if (mddev
->reshape_backwards
)
4512 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4513 - conf
->reshape_progress
;
4515 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4516 conf
->reshape_checkpoint
= jiffies
;
4517 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
4518 md_wakeup_thread(mddev
->thread
);
4519 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
4520 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4521 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4522 allow_barrier(conf
);
4523 return sectors_done
;
4525 conf
->reshape_safe
= mddev
->reshape_position
;
4526 allow_barrier(conf
);
4529 raise_barrier(conf
, 0);
4531 /* Now schedule reads for blocks from sector_nr to last */
4532 r10_bio
= raid10_alloc_init_r10buf(conf
);
4534 raise_barrier(conf
, 1);
4535 atomic_set(&r10_bio
->remaining
, 0);
4536 r10_bio
->mddev
= mddev
;
4537 r10_bio
->sector
= sector_nr
;
4538 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4539 r10_bio
->sectors
= last
- sector_nr
+ 1;
4540 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4541 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4544 /* Cannot read from here, so need to record bad blocks
4545 * on all the target devices.
4548 mempool_free(r10_bio
, conf
->r10buf_pool
);
4549 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4550 return sectors_done
;
4553 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4555 bio_set_dev(read_bio
, rdev
->bdev
);
4556 read_bio
->bi_iter
.bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4557 + rdev
->data_offset
);
4558 read_bio
->bi_private
= r10_bio
;
4559 read_bio
->bi_end_io
= end_reshape_read
;
4560 bio_set_op_attrs(read_bio
, REQ_OP_READ
, 0);
4561 read_bio
->bi_flags
&= (~0UL << BIO_RESET_BITS
);
4562 read_bio
->bi_status
= 0;
4563 read_bio
->bi_vcnt
= 0;
4564 read_bio
->bi_iter
.bi_size
= 0;
4565 r10_bio
->master_bio
= read_bio
;
4566 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4568 /* Now find the locations in the new layout */
4569 __raid10_find_phys(&conf
->geo
, r10_bio
);
4572 read_bio
->bi_next
= NULL
;
4575 for (s
= 0; s
< conf
->copies
*2; s
++) {
4577 int d
= r10_bio
->devs
[s
/2].devnum
;
4578 struct md_rdev
*rdev2
;
4580 rdev2
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4581 b
= r10_bio
->devs
[s
/2].repl_bio
;
4583 rdev2
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4584 b
= r10_bio
->devs
[s
/2].bio
;
4586 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4589 bio_set_dev(b
, rdev2
->bdev
);
4590 b
->bi_iter
.bi_sector
= r10_bio
->devs
[s
/2].addr
+
4591 rdev2
->new_data_offset
;
4592 b
->bi_end_io
= end_reshape_write
;
4593 bio_set_op_attrs(b
, REQ_OP_WRITE
, 0);
4598 /* Now add as many pages as possible to all of these bios. */
4601 pages
= get_resync_pages(r10_bio
->devs
[0].bio
)->pages
;
4602 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4603 struct page
*page
= pages
[s
/ (PAGE_SIZE
>> 9)];
4604 int len
= (max_sectors
- s
) << 9;
4605 if (len
> PAGE_SIZE
)
4607 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4609 * won't fail because the vec table is big enough
4610 * to hold all these pages
4612 bio_add_page(bio
, page
, len
, 0);
4614 sector_nr
+= len
>> 9;
4615 nr_sectors
+= len
>> 9;
4618 r10_bio
->sectors
= nr_sectors
;
4620 /* Now submit the read */
4621 md_sync_acct_bio(read_bio
, r10_bio
->sectors
);
4622 atomic_inc(&r10_bio
->remaining
);
4623 read_bio
->bi_next
= NULL
;
4624 generic_make_request(read_bio
);
4625 sectors_done
+= nr_sectors
;
4626 if (sector_nr
<= last
)
4629 lower_barrier(conf
);
4631 /* Now that we have done the whole section we can
4632 * update reshape_progress
4634 if (mddev
->reshape_backwards
)
4635 conf
->reshape_progress
-= sectors_done
;
4637 conf
->reshape_progress
+= sectors_done
;
4639 return sectors_done
;
4642 static void end_reshape_request(struct r10bio
*r10_bio
);
4643 static int handle_reshape_read_error(struct mddev
*mddev
,
4644 struct r10bio
*r10_bio
);
4645 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4647 /* Reshape read completed. Hopefully we have a block
4649 * If we got a read error then we do sync 1-page reads from
4650 * elsewhere until we find the data - or give up.
4652 struct r10conf
*conf
= mddev
->private;
4655 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4656 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4657 /* Reshape has been aborted */
4658 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4662 /* We definitely have the data in the pages, schedule the
4665 atomic_set(&r10_bio
->remaining
, 1);
4666 for (s
= 0; s
< conf
->copies
*2; s
++) {
4668 int d
= r10_bio
->devs
[s
/2].devnum
;
4669 struct md_rdev
*rdev
;
4672 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4673 b
= r10_bio
->devs
[s
/2].repl_bio
;
4675 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4676 b
= r10_bio
->devs
[s
/2].bio
;
4678 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
4682 atomic_inc(&rdev
->nr_pending
);
4684 md_sync_acct_bio(b
, r10_bio
->sectors
);
4685 atomic_inc(&r10_bio
->remaining
);
4687 generic_make_request(b
);
4689 end_reshape_request(r10_bio
);
4692 static void end_reshape(struct r10conf
*conf
)
4694 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4697 spin_lock_irq(&conf
->device_lock
);
4698 conf
->prev
= conf
->geo
;
4699 md_finish_reshape(conf
->mddev
);
4701 conf
->reshape_progress
= MaxSector
;
4702 conf
->reshape_safe
= MaxSector
;
4703 spin_unlock_irq(&conf
->device_lock
);
4705 /* read-ahead size must cover two whole stripes, which is
4706 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4708 if (conf
->mddev
->queue
) {
4709 int stripe
= conf
->geo
.raid_disks
*
4710 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4711 stripe
/= conf
->geo
.near_copies
;
4712 if (conf
->mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
4713 conf
->mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
4718 static int handle_reshape_read_error(struct mddev
*mddev
,
4719 struct r10bio
*r10_bio
)
4721 /* Use sync reads to get the blocks from somewhere else */
4722 int sectors
= r10_bio
->sectors
;
4723 struct r10conf
*conf
= mddev
->private;
4724 struct r10bio
*r10b
;
4727 struct page
**pages
;
4729 r10b
= kmalloc(sizeof(*r10b
) +
4730 sizeof(struct r10dev
) * conf
->copies
, GFP_NOIO
);
4732 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4736 /* reshape IOs share pages from .devs[0].bio */
4737 pages
= get_resync_pages(r10_bio
->devs
[0].bio
)->pages
;
4739 r10b
->sector
= r10_bio
->sector
;
4740 __raid10_find_phys(&conf
->prev
, r10b
);
4745 int first_slot
= slot
;
4747 if (s
> (PAGE_SIZE
>> 9))
4752 int d
= r10b
->devs
[slot
].devnum
;
4753 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4756 test_bit(Faulty
, &rdev
->flags
) ||
4757 !test_bit(In_sync
, &rdev
->flags
))
4760 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4761 atomic_inc(&rdev
->nr_pending
);
4763 success
= sync_page_io(rdev
,
4767 REQ_OP_READ
, 0, false);
4768 rdev_dec_pending(rdev
, mddev
);
4774 if (slot
>= conf
->copies
)
4776 if (slot
== first_slot
)
4781 /* couldn't read this block, must give up */
4782 set_bit(MD_RECOVERY_INTR
,
4794 static void end_reshape_write(struct bio
*bio
)
4796 struct r10bio
*r10_bio
= get_resync_r10bio(bio
);
4797 struct mddev
*mddev
= r10_bio
->mddev
;
4798 struct r10conf
*conf
= mddev
->private;
4802 struct md_rdev
*rdev
= NULL
;
4804 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4806 rdev
= conf
->mirrors
[d
].replacement
;
4809 rdev
= conf
->mirrors
[d
].rdev
;
4812 if (bio
->bi_status
) {
4813 /* FIXME should record badblock */
4814 md_error(mddev
, rdev
);
4817 rdev_dec_pending(rdev
, mddev
);
4818 end_reshape_request(r10_bio
);
4821 static void end_reshape_request(struct r10bio
*r10_bio
)
4823 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4825 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4826 bio_put(r10_bio
->master_bio
);
4830 static void raid10_finish_reshape(struct mddev
*mddev
)
4832 struct r10conf
*conf
= mddev
->private;
4834 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4837 if (mddev
->delta_disks
> 0) {
4838 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4839 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4840 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4842 mddev
->resync_max_sectors
= mddev
->array_sectors
;
4846 for (d
= conf
->geo
.raid_disks
;
4847 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4849 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4851 clear_bit(In_sync
, &rdev
->flags
);
4852 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4854 clear_bit(In_sync
, &rdev
->flags
);
4858 mddev
->layout
= mddev
->new_layout
;
4859 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4860 mddev
->reshape_position
= MaxSector
;
4861 mddev
->delta_disks
= 0;
4862 mddev
->reshape_backwards
= 0;
4865 static struct md_personality raid10_personality
=
4869 .owner
= THIS_MODULE
,
4870 .make_request
= raid10_make_request
,
4872 .free
= raid10_free
,
4873 .status
= raid10_status
,
4874 .error_handler
= raid10_error
,
4875 .hot_add_disk
= raid10_add_disk
,
4876 .hot_remove_disk
= raid10_remove_disk
,
4877 .spare_active
= raid10_spare_active
,
4878 .sync_request
= raid10_sync_request
,
4879 .quiesce
= raid10_quiesce
,
4880 .size
= raid10_size
,
4881 .resize
= raid10_resize
,
4882 .takeover
= raid10_takeover
,
4883 .check_reshape
= raid10_check_reshape
,
4884 .start_reshape
= raid10_start_reshape
,
4885 .finish_reshape
= raid10_finish_reshape
,
4886 .congested
= raid10_congested
,
4889 static int __init
raid_init(void)
4891 return register_md_personality(&raid10_personality
);
4894 static void raid_exit(void)
4896 unregister_md_personality(&raid10_personality
);
4899 module_init(raid_init
);
4900 module_exit(raid_exit
);
4901 MODULE_LICENSE("GPL");
4902 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4903 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4904 MODULE_ALIAS("md-raid10");
4905 MODULE_ALIAS("md-level-10");
4907 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);