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>
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)
114 * 'strct resync_pages' stores actual pages used for doing the resync
115 * IO, and it is per-bio, so make .bi_private points to it.
117 static inline struct resync_pages
*get_resync_pages(struct bio
*bio
)
119 return bio
->bi_private
;
123 * for resync bio, r10bio pointer can be retrieved from the per-bio
124 * 'struct resync_pages'.
126 static inline struct r10bio
*get_resync_r10bio(struct bio
*bio
)
128 return get_resync_pages(bio
)->raid_bio
;
131 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
133 struct r10conf
*conf
= data
;
134 int size
= offsetof(struct r10bio
, devs
[conf
->copies
]);
136 /* allocate a r10bio with room for raid_disks entries in the
138 return kzalloc(size
, gfp_flags
);
141 static void r10bio_pool_free(void *r10_bio
, void *data
)
146 /* amount of memory to reserve for resync requests */
147 #define RESYNC_WINDOW (1024*1024)
148 /* maximum number of concurrent requests, memory permitting */
149 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
152 * When performing a resync, we need to read and compare, so
153 * we need as many pages are there are copies.
154 * When performing a recovery, we need 2 bios, one for read,
155 * one for write (we recover only one drive per r10buf)
158 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
160 struct r10conf
*conf
= data
;
161 struct r10bio
*r10_bio
;
164 int nalloc
, nalloc_rp
;
165 struct resync_pages
*rps
;
167 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
171 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
) ||
172 test_bit(MD_RECOVERY_RESHAPE
, &conf
->mddev
->recovery
))
173 nalloc
= conf
->copies
; /* resync */
175 nalloc
= 2; /* recovery */
177 /* allocate once for all bios */
178 if (!conf
->have_replacement
)
181 nalloc_rp
= nalloc
* 2;
182 rps
= kmalloc(sizeof(struct resync_pages
) * nalloc_rp
, gfp_flags
);
184 goto out_free_r10bio
;
189 for (j
= nalloc
; j
-- ; ) {
190 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
193 r10_bio
->devs
[j
].bio
= bio
;
194 if (!conf
->have_replacement
)
196 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
199 r10_bio
->devs
[j
].repl_bio
= bio
;
202 * Allocate RESYNC_PAGES data pages and attach them
205 for (j
= 0; j
< nalloc
; j
++) {
206 struct bio
*rbio
= r10_bio
->devs
[j
].repl_bio
;
207 struct resync_pages
*rp
, *rp_repl
;
211 rp_repl
= &rps
[nalloc
+ j
];
213 bio
= r10_bio
->devs
[j
].bio
;
215 if (!j
|| test_bit(MD_RECOVERY_SYNC
,
216 &conf
->mddev
->recovery
)) {
217 if (resync_alloc_pages(rp
, gfp_flags
))
220 memcpy(rp
, &rps
[0], sizeof(*rp
));
221 resync_get_all_pages(rp
);
225 rp
->raid_bio
= r10_bio
;
226 bio
->bi_private
= rp
;
228 memcpy(rp_repl
, rp
, sizeof(*rp
));
229 rbio
->bi_private
= rp_repl
;
237 resync_free_pages(&rps
[j
* 2]);
241 for ( ; j
< nalloc
; j
++) {
242 if (r10_bio
->devs
[j
].bio
)
243 bio_put(r10_bio
->devs
[j
].bio
);
244 if (r10_bio
->devs
[j
].repl_bio
)
245 bio_put(r10_bio
->devs
[j
].repl_bio
);
249 r10bio_pool_free(r10_bio
, conf
);
253 static void r10buf_pool_free(void *__r10_bio
, void *data
)
255 struct r10conf
*conf
= data
;
256 struct r10bio
*r10bio
= __r10_bio
;
258 struct resync_pages
*rp
= NULL
;
260 for (j
= conf
->copies
; j
--; ) {
261 struct bio
*bio
= r10bio
->devs
[j
].bio
;
263 rp
= get_resync_pages(bio
);
264 resync_free_pages(rp
);
267 bio
= r10bio
->devs
[j
].repl_bio
;
272 /* resync pages array stored in the 1st bio's .bi_private */
275 r10bio_pool_free(r10bio
, conf
);
278 static void put_all_bios(struct r10conf
*conf
, struct r10bio
*r10_bio
)
282 for (i
= 0; i
< conf
->copies
; i
++) {
283 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
284 if (!BIO_SPECIAL(*bio
))
287 bio
= &r10_bio
->devs
[i
].repl_bio
;
288 if (r10_bio
->read_slot
< 0 && !BIO_SPECIAL(*bio
))
294 static void free_r10bio(struct r10bio
*r10_bio
)
296 struct r10conf
*conf
= r10_bio
->mddev
->private;
298 put_all_bios(conf
, r10_bio
);
299 mempool_free(r10_bio
, conf
->r10bio_pool
);
302 static void put_buf(struct r10bio
*r10_bio
)
304 struct r10conf
*conf
= r10_bio
->mddev
->private;
306 mempool_free(r10_bio
, conf
->r10buf_pool
);
311 static void reschedule_retry(struct r10bio
*r10_bio
)
314 struct mddev
*mddev
= r10_bio
->mddev
;
315 struct r10conf
*conf
= mddev
->private;
317 spin_lock_irqsave(&conf
->device_lock
, flags
);
318 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
320 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
322 /* wake up frozen array... */
323 wake_up(&conf
->wait_barrier
);
325 md_wakeup_thread(mddev
->thread
);
329 * raid_end_bio_io() is called when we have finished servicing a mirrored
330 * operation and are ready to return a success/failure code to the buffer
333 static void raid_end_bio_io(struct r10bio
*r10_bio
)
335 struct bio
*bio
= r10_bio
->master_bio
;
336 struct r10conf
*conf
= r10_bio
->mddev
->private;
338 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
339 bio
->bi_status
= BLK_STS_IOERR
;
343 * Wake up any possible resync thread that waits for the device
348 free_r10bio(r10_bio
);
352 * Update disk head position estimator based on IRQ completion info.
354 static inline void update_head_pos(int slot
, struct r10bio
*r10_bio
)
356 struct r10conf
*conf
= r10_bio
->mddev
->private;
358 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
359 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
363 * Find the disk number which triggered given bio
365 static int find_bio_disk(struct r10conf
*conf
, struct r10bio
*r10_bio
,
366 struct bio
*bio
, int *slotp
, int *replp
)
371 for (slot
= 0; slot
< conf
->copies
; slot
++) {
372 if (r10_bio
->devs
[slot
].bio
== bio
)
374 if (r10_bio
->devs
[slot
].repl_bio
== bio
) {
380 BUG_ON(slot
== conf
->copies
);
381 update_head_pos(slot
, r10_bio
);
387 return r10_bio
->devs
[slot
].devnum
;
390 static void raid10_end_read_request(struct bio
*bio
)
392 int uptodate
= !bio
->bi_status
;
393 struct r10bio
*r10_bio
= bio
->bi_private
;
395 struct md_rdev
*rdev
;
396 struct r10conf
*conf
= r10_bio
->mddev
->private;
398 slot
= r10_bio
->read_slot
;
399 dev
= r10_bio
->devs
[slot
].devnum
;
400 rdev
= r10_bio
->devs
[slot
].rdev
;
402 * this branch is our 'one mirror IO has finished' event handler:
404 update_head_pos(slot
, r10_bio
);
408 * Set R10BIO_Uptodate in our master bio, so that
409 * we will return a good error code to the higher
410 * levels even if IO on some other mirrored buffer fails.
412 * The 'master' represents the composite IO operation to
413 * user-side. So if something waits for IO, then it will
414 * wait for the 'master' bio.
416 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
418 /* If all other devices that store this block have
419 * failed, we want to return the error upwards rather
420 * than fail the last device. Here we redefine
421 * "uptodate" to mean "Don't want to retry"
423 if (!_enough(conf
, test_bit(R10BIO_Previous
, &r10_bio
->state
),
428 raid_end_bio_io(r10_bio
);
429 rdev_dec_pending(rdev
, conf
->mddev
);
432 * oops, read error - keep the refcount on the rdev
434 char b
[BDEVNAME_SIZE
];
435 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
437 bdevname(rdev
->bdev
, b
),
438 (unsigned long long)r10_bio
->sector
);
439 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
440 reschedule_retry(r10_bio
);
444 static void close_write(struct r10bio
*r10_bio
)
446 /* clear the bitmap if all writes complete successfully */
447 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
449 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
451 md_write_end(r10_bio
->mddev
);
454 static void one_write_done(struct r10bio
*r10_bio
)
456 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
457 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
458 reschedule_retry(r10_bio
);
460 close_write(r10_bio
);
461 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
462 reschedule_retry(r10_bio
);
464 raid_end_bio_io(r10_bio
);
469 static void raid10_end_write_request(struct bio
*bio
)
471 struct r10bio
*r10_bio
= bio
->bi_private
;
474 struct r10conf
*conf
= r10_bio
->mddev
->private;
476 struct md_rdev
*rdev
= NULL
;
477 struct bio
*to_put
= NULL
;
480 discard_error
= bio
->bi_status
&& bio_op(bio
) == REQ_OP_DISCARD
;
482 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
485 rdev
= conf
->mirrors
[dev
].replacement
;
489 rdev
= conf
->mirrors
[dev
].rdev
;
492 * this branch is our 'one mirror IO has finished' event handler:
494 if (bio
->bi_status
&& !discard_error
) {
496 /* Never record new bad blocks to replacement,
499 md_error(rdev
->mddev
, rdev
);
501 set_bit(WriteErrorSeen
, &rdev
->flags
);
502 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
503 set_bit(MD_RECOVERY_NEEDED
,
504 &rdev
->mddev
->recovery
);
507 if (test_bit(FailFast
, &rdev
->flags
) &&
508 (bio
->bi_opf
& MD_FAILFAST
)) {
509 md_error(rdev
->mddev
, rdev
);
510 if (!test_bit(Faulty
, &rdev
->flags
))
511 /* This is the only remaining device,
512 * We need to retry the write without
515 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
517 r10_bio
->devs
[slot
].bio
= NULL
;
522 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
526 * Set R10BIO_Uptodate in our master bio, so that
527 * we will return a good error code for to the higher
528 * levels even if IO on some other mirrored buffer fails.
530 * The 'master' represents the composite IO operation to
531 * user-side. So if something waits for IO, then it will
532 * wait for the 'master' bio.
538 * Do not set R10BIO_Uptodate if the current device is
539 * rebuilding or Faulty. This is because we cannot use
540 * such device for properly reading the data back (we could
541 * potentially use it, if the current write would have felt
542 * before rdev->recovery_offset, but for simplicity we don't
545 if (test_bit(In_sync
, &rdev
->flags
) &&
546 !test_bit(Faulty
, &rdev
->flags
))
547 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
549 /* Maybe we can clear some bad blocks. */
550 if (is_badblock(rdev
,
551 r10_bio
->devs
[slot
].addr
,
553 &first_bad
, &bad_sectors
) && !discard_error
) {
556 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
558 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
560 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
566 * Let's see if all mirrored write operations have finished
569 one_write_done(r10_bio
);
571 rdev_dec_pending(rdev
, conf
->mddev
);
577 * RAID10 layout manager
578 * As well as the chunksize and raid_disks count, there are two
579 * parameters: near_copies and far_copies.
580 * near_copies * far_copies must be <= raid_disks.
581 * Normally one of these will be 1.
582 * If both are 1, we get raid0.
583 * If near_copies == raid_disks, we get raid1.
585 * Chunks are laid out in raid0 style with near_copies copies of the
586 * first chunk, followed by near_copies copies of the next chunk and
588 * If far_copies > 1, then after 1/far_copies of the array has been assigned
589 * as described above, we start again with a device offset of near_copies.
590 * So we effectively have another copy of the whole array further down all
591 * the drives, but with blocks on different drives.
592 * With this layout, and block is never stored twice on the one device.
594 * raid10_find_phys finds the sector offset of a given virtual sector
595 * on each device that it is on.
597 * raid10_find_virt does the reverse mapping, from a device and a
598 * sector offset to a virtual address
601 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
609 int last_far_set_start
, last_far_set_size
;
611 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
612 last_far_set_start
*= geo
->far_set_size
;
614 last_far_set_size
= geo
->far_set_size
;
615 last_far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
617 /* now calculate first sector/dev */
618 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
619 sector
= r10bio
->sector
& geo
->chunk_mask
;
621 chunk
*= geo
->near_copies
;
623 dev
= sector_div(stripe
, geo
->raid_disks
);
625 stripe
*= geo
->far_copies
;
627 sector
+= stripe
<< geo
->chunk_shift
;
629 /* and calculate all the others */
630 for (n
= 0; n
< geo
->near_copies
; n
++) {
634 r10bio
->devs
[slot
].devnum
= d
;
635 r10bio
->devs
[slot
].addr
= s
;
638 for (f
= 1; f
< geo
->far_copies
; f
++) {
639 set
= d
/ geo
->far_set_size
;
640 d
+= geo
->near_copies
;
642 if ((geo
->raid_disks
% geo
->far_set_size
) &&
643 (d
> last_far_set_start
)) {
644 d
-= last_far_set_start
;
645 d
%= last_far_set_size
;
646 d
+= last_far_set_start
;
648 d
%= geo
->far_set_size
;
649 d
+= geo
->far_set_size
* set
;
652 r10bio
->devs
[slot
].devnum
= d
;
653 r10bio
->devs
[slot
].addr
= s
;
657 if (dev
>= geo
->raid_disks
) {
659 sector
+= (geo
->chunk_mask
+ 1);
664 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
666 struct geom
*geo
= &conf
->geo
;
668 if (conf
->reshape_progress
!= MaxSector
&&
669 ((r10bio
->sector
>= conf
->reshape_progress
) !=
670 conf
->mddev
->reshape_backwards
)) {
671 set_bit(R10BIO_Previous
, &r10bio
->state
);
674 clear_bit(R10BIO_Previous
, &r10bio
->state
);
676 __raid10_find_phys(geo
, r10bio
);
679 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
681 sector_t offset
, chunk
, vchunk
;
682 /* Never use conf->prev as this is only called during resync
683 * or recovery, so reshape isn't happening
685 struct geom
*geo
= &conf
->geo
;
686 int far_set_start
= (dev
/ geo
->far_set_size
) * geo
->far_set_size
;
687 int far_set_size
= geo
->far_set_size
;
688 int last_far_set_start
;
690 if (geo
->raid_disks
% geo
->far_set_size
) {
691 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
692 last_far_set_start
*= geo
->far_set_size
;
694 if (dev
>= last_far_set_start
) {
695 far_set_size
= geo
->far_set_size
;
696 far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
697 far_set_start
= last_far_set_start
;
701 offset
= sector
& geo
->chunk_mask
;
702 if (geo
->far_offset
) {
704 chunk
= sector
>> geo
->chunk_shift
;
705 fc
= sector_div(chunk
, geo
->far_copies
);
706 dev
-= fc
* geo
->near_copies
;
707 if (dev
< far_set_start
)
710 while (sector
>= geo
->stride
) {
711 sector
-= geo
->stride
;
712 if (dev
< (geo
->near_copies
+ far_set_start
))
713 dev
+= far_set_size
- geo
->near_copies
;
715 dev
-= geo
->near_copies
;
717 chunk
= sector
>> geo
->chunk_shift
;
719 vchunk
= chunk
* geo
->raid_disks
+ dev
;
720 sector_div(vchunk
, geo
->near_copies
);
721 return (vchunk
<< geo
->chunk_shift
) + offset
;
725 * This routine returns the disk from which the requested read should
726 * be done. There is a per-array 'next expected sequential IO' sector
727 * number - if this matches on the next IO then we use the last disk.
728 * There is also a per-disk 'last know head position' sector that is
729 * maintained from IRQ contexts, both the normal and the resync IO
730 * completion handlers update this position correctly. If there is no
731 * perfect sequential match then we pick the disk whose head is closest.
733 * If there are 2 mirrors in the same 2 devices, performance degrades
734 * because position is mirror, not device based.
736 * The rdev for the device selected will have nr_pending incremented.
740 * FIXME: possibly should rethink readbalancing and do it differently
741 * depending on near_copies / far_copies geometry.
743 static struct md_rdev
*read_balance(struct r10conf
*conf
,
744 struct r10bio
*r10_bio
,
747 const sector_t this_sector
= r10_bio
->sector
;
749 int sectors
= r10_bio
->sectors
;
750 int best_good_sectors
;
751 sector_t new_distance
, best_dist
;
752 struct md_rdev
*best_rdev
, *rdev
= NULL
;
755 struct geom
*geo
= &conf
->geo
;
757 raid10_find_phys(conf
, r10_bio
);
759 sectors
= r10_bio
->sectors
;
762 best_dist
= MaxSector
;
763 best_good_sectors
= 0;
765 clear_bit(R10BIO_FailFast
, &r10_bio
->state
);
767 * Check if we can balance. We can balance on the whole
768 * device if no resync is going on (recovery is ok), or below
769 * the resync window. We take the first readable disk when
770 * above the resync window.
772 if (conf
->mddev
->recovery_cp
< MaxSector
773 && (this_sector
+ sectors
>= conf
->next_resync
))
776 for (slot
= 0; slot
< conf
->copies
; slot
++) {
781 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
783 disk
= r10_bio
->devs
[slot
].devnum
;
784 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
785 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
786 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
787 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
789 test_bit(Faulty
, &rdev
->flags
))
791 if (!test_bit(In_sync
, &rdev
->flags
) &&
792 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
795 dev_sector
= r10_bio
->devs
[slot
].addr
;
796 if (is_badblock(rdev
, dev_sector
, sectors
,
797 &first_bad
, &bad_sectors
)) {
798 if (best_dist
< MaxSector
)
799 /* Already have a better slot */
801 if (first_bad
<= dev_sector
) {
802 /* Cannot read here. If this is the
803 * 'primary' device, then we must not read
804 * beyond 'bad_sectors' from another device.
806 bad_sectors
-= (dev_sector
- first_bad
);
807 if (!do_balance
&& sectors
> bad_sectors
)
808 sectors
= bad_sectors
;
809 if (best_good_sectors
> sectors
)
810 best_good_sectors
= sectors
;
812 sector_t good_sectors
=
813 first_bad
- dev_sector
;
814 if (good_sectors
> best_good_sectors
) {
815 best_good_sectors
= good_sectors
;
820 /* Must read from here */
825 best_good_sectors
= sectors
;
831 /* At least 2 disks to choose from so failfast is OK */
832 set_bit(R10BIO_FailFast
, &r10_bio
->state
);
833 /* This optimisation is debatable, and completely destroys
834 * sequential read speed for 'far copies' arrays. So only
835 * keep it for 'near' arrays, and review those later.
837 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
840 /* for far > 1 always use the lowest address */
841 else if (geo
->far_copies
> 1)
842 new_distance
= r10_bio
->devs
[slot
].addr
;
844 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
845 conf
->mirrors
[disk
].head_position
);
846 if (new_distance
< best_dist
) {
847 best_dist
= new_distance
;
852 if (slot
>= conf
->copies
) {
858 atomic_inc(&rdev
->nr_pending
);
859 r10_bio
->read_slot
= slot
;
863 *max_sectors
= best_good_sectors
;
868 static int raid10_congested(struct mddev
*mddev
, int bits
)
870 struct r10conf
*conf
= mddev
->private;
873 if ((bits
& (1 << WB_async_congested
)) &&
874 conf
->pending_count
>= max_queued_requests
)
879 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
882 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
883 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
884 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
886 ret
|= bdi_congested(q
->backing_dev_info
, bits
);
893 static void flush_pending_writes(struct r10conf
*conf
)
895 /* Any writes that have been queued but are awaiting
896 * bitmap updates get flushed here.
898 spin_lock_irq(&conf
->device_lock
);
900 if (conf
->pending_bio_list
.head
) {
902 bio
= bio_list_get(&conf
->pending_bio_list
);
903 conf
->pending_count
= 0;
904 spin_unlock_irq(&conf
->device_lock
);
905 /* flush any pending bitmap writes to disk
906 * before proceeding w/ I/O */
907 bitmap_unplug(conf
->mddev
->bitmap
);
908 wake_up(&conf
->wait_barrier
);
910 while (bio
) { /* submit pending writes */
911 struct bio
*next
= bio
->bi_next
;
912 struct md_rdev
*rdev
= (void*)bio
->bi_bdev
;
914 bio
->bi_bdev
= rdev
->bdev
;
915 if (test_bit(Faulty
, &rdev
->flags
)) {
916 bio
->bi_status
= BLK_STS_IOERR
;
918 } else if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
919 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
923 generic_make_request(bio
);
927 spin_unlock_irq(&conf
->device_lock
);
931 * Sometimes we need to suspend IO while we do something else,
932 * either some resync/recovery, or reconfigure the array.
933 * To do this we raise a 'barrier'.
934 * The 'barrier' is a counter that can be raised multiple times
935 * to count how many activities are happening which preclude
937 * We can only raise the barrier if there is no pending IO.
938 * i.e. if nr_pending == 0.
939 * We choose only to raise the barrier if no-one is waiting for the
940 * barrier to go down. This means that as soon as an IO request
941 * is ready, no other operations which require a barrier will start
942 * until the IO request has had a chance.
944 * So: regular IO calls 'wait_barrier'. When that returns there
945 * is no backgroup IO happening, It must arrange to call
946 * allow_barrier when it has finished its IO.
947 * backgroup IO calls must call raise_barrier. Once that returns
948 * there is no normal IO happeing. It must arrange to call
949 * lower_barrier when the particular background IO completes.
952 static void raise_barrier(struct r10conf
*conf
, int force
)
954 BUG_ON(force
&& !conf
->barrier
);
955 spin_lock_irq(&conf
->resync_lock
);
957 /* Wait until no block IO is waiting (unless 'force') */
958 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
961 /* block any new IO from starting */
964 /* Now wait for all pending IO to complete */
965 wait_event_lock_irq(conf
->wait_barrier
,
966 !atomic_read(&conf
->nr_pending
) && conf
->barrier
< RESYNC_DEPTH
,
969 spin_unlock_irq(&conf
->resync_lock
);
972 static void lower_barrier(struct r10conf
*conf
)
975 spin_lock_irqsave(&conf
->resync_lock
, flags
);
977 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
978 wake_up(&conf
->wait_barrier
);
981 static void wait_barrier(struct r10conf
*conf
)
983 spin_lock_irq(&conf
->resync_lock
);
986 /* Wait for the barrier to drop.
987 * However if there are already pending
988 * requests (preventing the barrier from
989 * rising completely), and the
990 * pre-process bio queue isn't empty,
991 * then don't wait, as we need to empty
992 * that queue to get the nr_pending
995 raid10_log(conf
->mddev
, "wait barrier");
996 wait_event_lock_irq(conf
->wait_barrier
,
998 (atomic_read(&conf
->nr_pending
) &&
1000 (!bio_list_empty(¤t
->bio_list
[0]) ||
1001 !bio_list_empty(¤t
->bio_list
[1]))),
1004 if (!conf
->nr_waiting
)
1005 wake_up(&conf
->wait_barrier
);
1007 atomic_inc(&conf
->nr_pending
);
1008 spin_unlock_irq(&conf
->resync_lock
);
1011 static void allow_barrier(struct r10conf
*conf
)
1013 if ((atomic_dec_and_test(&conf
->nr_pending
)) ||
1014 (conf
->array_freeze_pending
))
1015 wake_up(&conf
->wait_barrier
);
1018 static void freeze_array(struct r10conf
*conf
, int extra
)
1020 /* stop syncio and normal IO and wait for everything to
1022 * We increment barrier and nr_waiting, and then
1023 * wait until nr_pending match nr_queued+extra
1024 * This is called in the context of one normal IO request
1025 * that has failed. Thus any sync request that might be pending
1026 * will be blocked by nr_pending, and we need to wait for
1027 * pending IO requests to complete or be queued for re-try.
1028 * Thus the number queued (nr_queued) plus this request (extra)
1029 * must match the number of pending IOs (nr_pending) before
1032 spin_lock_irq(&conf
->resync_lock
);
1033 conf
->array_freeze_pending
++;
1036 wait_event_lock_irq_cmd(conf
->wait_barrier
,
1037 atomic_read(&conf
->nr_pending
) == conf
->nr_queued
+extra
,
1039 flush_pending_writes(conf
));
1041 conf
->array_freeze_pending
--;
1042 spin_unlock_irq(&conf
->resync_lock
);
1045 static void unfreeze_array(struct r10conf
*conf
)
1047 /* reverse the effect of the freeze */
1048 spin_lock_irq(&conf
->resync_lock
);
1051 wake_up(&conf
->wait_barrier
);
1052 spin_unlock_irq(&conf
->resync_lock
);
1055 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1056 struct md_rdev
*rdev
)
1058 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1059 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1060 return rdev
->data_offset
;
1062 return rdev
->new_data_offset
;
1065 struct raid10_plug_cb
{
1066 struct blk_plug_cb cb
;
1067 struct bio_list pending
;
1071 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1073 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1075 struct mddev
*mddev
= plug
->cb
.data
;
1076 struct r10conf
*conf
= mddev
->private;
1079 if (from_schedule
|| current
->bio_list
) {
1080 spin_lock_irq(&conf
->device_lock
);
1081 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1082 conf
->pending_count
+= plug
->pending_cnt
;
1083 spin_unlock_irq(&conf
->device_lock
);
1084 wake_up(&conf
->wait_barrier
);
1085 md_wakeup_thread(mddev
->thread
);
1090 /* we aren't scheduling, so we can do the write-out directly. */
1091 bio
= bio_list_get(&plug
->pending
);
1092 bitmap_unplug(mddev
->bitmap
);
1093 wake_up(&conf
->wait_barrier
);
1095 while (bio
) { /* submit pending writes */
1096 struct bio
*next
= bio
->bi_next
;
1097 struct md_rdev
*rdev
= (void*)bio
->bi_bdev
;
1098 bio
->bi_next
= NULL
;
1099 bio
->bi_bdev
= rdev
->bdev
;
1100 if (test_bit(Faulty
, &rdev
->flags
)) {
1101 bio
->bi_status
= BLK_STS_IOERR
;
1103 } else if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
1104 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1105 /* Just ignore it */
1108 generic_make_request(bio
);
1114 static void raid10_read_request(struct mddev
*mddev
, struct bio
*bio
,
1115 struct r10bio
*r10_bio
)
1117 struct r10conf
*conf
= mddev
->private;
1118 struct bio
*read_bio
;
1119 const int op
= bio_op(bio
);
1120 const unsigned long do_sync
= (bio
->bi_opf
& REQ_SYNC
);
1123 struct md_rdev
*rdev
;
1124 char b
[BDEVNAME_SIZE
];
1125 int slot
= r10_bio
->read_slot
;
1126 struct md_rdev
*err_rdev
= NULL
;
1127 gfp_t gfp
= GFP_NOIO
;
1129 if (r10_bio
->devs
[slot
].rdev
) {
1131 * This is an error retry, but we cannot
1132 * safely dereference the rdev in the r10_bio,
1133 * we must use the one in conf.
1134 * If it has already been disconnected (unlikely)
1135 * we lose the device name in error messages.
1139 * As we are blocking raid10, it is a little safer to
1142 gfp
= GFP_NOIO
| __GFP_HIGH
;
1145 disk
= r10_bio
->devs
[slot
].devnum
;
1146 err_rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
1148 bdevname(err_rdev
->bdev
, b
);
1151 /* This never gets dereferenced */
1152 err_rdev
= r10_bio
->devs
[slot
].rdev
;
1157 * Register the new request and wait if the reconstruction
1158 * thread has put up a bar for new requests.
1159 * Continue immediately if no resync is active currently.
1163 sectors
= r10_bio
->sectors
;
1164 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1165 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1166 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1168 * IO spans the reshape position. Need to wait for reshape to
1171 raid10_log(conf
->mddev
, "wait reshape");
1172 allow_barrier(conf
);
1173 wait_event(conf
->wait_barrier
,
1174 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1175 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1180 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1183 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1185 (unsigned long long)r10_bio
->sector
);
1187 raid_end_bio_io(r10_bio
);
1191 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1193 bdevname(rdev
->bdev
, b
),
1194 (unsigned long long)r10_bio
->sector
);
1195 if (max_sectors
< bio_sectors(bio
)) {
1196 struct bio
*split
= bio_split(bio
, max_sectors
,
1197 gfp
, conf
->bio_split
);
1198 bio_chain(split
, bio
);
1199 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 read_bio
->bi_bdev
= 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(bdev_get_queue(read_bio
->bi_bdev
),
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 mbio
->bi_bdev
= 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(bdev_get_queue(mbio
->bi_bdev
),
1273 mbio
, disk_devt(conf
->mddev
->gendisk
),
1275 /* flush_pending_writes() needs access to the rdev so...*/
1276 mbio
->bi_bdev
= (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
;
1307 * Register the new request and wait if the reconstruction
1308 * thread has put up a bar for new requests.
1309 * Continue immediately if no resync is active currently.
1313 sectors
= r10_bio
->sectors
;
1314 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1315 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1316 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1318 * IO spans the reshape position. Need to wait for reshape to
1321 raid10_log(conf
->mddev
, "wait reshape");
1322 allow_barrier(conf
);
1323 wait_event(conf
->wait_barrier
,
1324 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1325 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1330 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1331 (mddev
->reshape_backwards
1332 ? (bio
->bi_iter
.bi_sector
< conf
->reshape_safe
&&
1333 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
)
1334 : (bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_safe
&&
1335 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
))) {
1336 /* Need to update reshape_position in metadata */
1337 mddev
->reshape_position
= conf
->reshape_progress
;
1338 set_mask_bits(&mddev
->sb_flags
, 0,
1339 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
1340 md_wakeup_thread(mddev
->thread
);
1341 raid10_log(conf
->mddev
, "wait reshape metadata");
1342 wait_event(mddev
->sb_wait
,
1343 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
));
1345 conf
->reshape_safe
= mddev
->reshape_position
;
1348 if (conf
->pending_count
>= max_queued_requests
) {
1349 md_wakeup_thread(mddev
->thread
);
1350 raid10_log(mddev
, "wait queued");
1351 wait_event(conf
->wait_barrier
,
1352 conf
->pending_count
< max_queued_requests
);
1354 /* first select target devices under rcu_lock and
1355 * inc refcount on their rdev. Record them by setting
1357 * If there are known/acknowledged bad blocks on any device
1358 * on which we have seen a write error, we want to avoid
1359 * writing to those blocks. This potentially requires several
1360 * writes to write around the bad blocks. Each set of writes
1361 * gets its own r10_bio with a set of bios attached.
1364 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1365 raid10_find_phys(conf
, r10_bio
);
1367 blocked_rdev
= NULL
;
1369 max_sectors
= r10_bio
->sectors
;
1371 for (i
= 0; i
< conf
->copies
; i
++) {
1372 int d
= r10_bio
->devs
[i
].devnum
;
1373 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1374 struct md_rdev
*rrdev
= rcu_dereference(
1375 conf
->mirrors
[d
].replacement
);
1378 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1379 atomic_inc(&rdev
->nr_pending
);
1380 blocked_rdev
= rdev
;
1383 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1384 atomic_inc(&rrdev
->nr_pending
);
1385 blocked_rdev
= rrdev
;
1388 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)))
1390 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)))
1393 r10_bio
->devs
[i
].bio
= NULL
;
1394 r10_bio
->devs
[i
].repl_bio
= NULL
;
1396 if (!rdev
&& !rrdev
) {
1397 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1400 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1402 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1406 is_bad
= is_badblock(rdev
, dev_sector
, max_sectors
,
1407 &first_bad
, &bad_sectors
);
1409 /* Mustn't write here until the bad block
1412 atomic_inc(&rdev
->nr_pending
);
1413 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1414 blocked_rdev
= rdev
;
1417 if (is_bad
&& first_bad
<= dev_sector
) {
1418 /* Cannot write here at all */
1419 bad_sectors
-= (dev_sector
- first_bad
);
1420 if (bad_sectors
< max_sectors
)
1421 /* Mustn't write more than bad_sectors
1422 * to other devices yet
1424 max_sectors
= bad_sectors
;
1425 /* We don't set R10BIO_Degraded as that
1426 * only applies if the disk is missing,
1427 * so it might be re-added, and we want to
1428 * know to recover this chunk.
1429 * In this case the device is here, and the
1430 * fact that this chunk is not in-sync is
1431 * recorded in the bad block log.
1436 int good_sectors
= first_bad
- dev_sector
;
1437 if (good_sectors
< max_sectors
)
1438 max_sectors
= good_sectors
;
1442 r10_bio
->devs
[i
].bio
= bio
;
1443 atomic_inc(&rdev
->nr_pending
);
1446 r10_bio
->devs
[i
].repl_bio
= bio
;
1447 atomic_inc(&rrdev
->nr_pending
);
1452 if (unlikely(blocked_rdev
)) {
1453 /* Have to wait for this device to get unblocked, then retry */
1457 for (j
= 0; j
< i
; j
++) {
1458 if (r10_bio
->devs
[j
].bio
) {
1459 d
= r10_bio
->devs
[j
].devnum
;
1460 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1462 if (r10_bio
->devs
[j
].repl_bio
) {
1463 struct md_rdev
*rdev
;
1464 d
= r10_bio
->devs
[j
].devnum
;
1465 rdev
= conf
->mirrors
[d
].replacement
;
1467 /* Race with remove_disk */
1469 rdev
= conf
->mirrors
[d
].rdev
;
1471 rdev_dec_pending(rdev
, mddev
);
1474 allow_barrier(conf
);
1475 raid10_log(conf
->mddev
, "wait rdev %d blocked", blocked_rdev
->raid_disk
);
1476 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1481 if (max_sectors
< r10_bio
->sectors
)
1482 r10_bio
->sectors
= max_sectors
;
1484 if (r10_bio
->sectors
< bio_sectors(bio
)) {
1485 struct bio
*split
= bio_split(bio
, r10_bio
->sectors
,
1486 GFP_NOIO
, conf
->bio_split
);
1487 bio_chain(split
, bio
);
1488 generic_make_request(bio
);
1490 r10_bio
->master_bio
= bio
;
1493 atomic_set(&r10_bio
->remaining
, 1);
1494 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1496 for (i
= 0; i
< conf
->copies
; i
++) {
1497 if (r10_bio
->devs
[i
].bio
)
1498 raid10_write_one_disk(mddev
, r10_bio
, bio
, false, i
);
1499 if (r10_bio
->devs
[i
].repl_bio
)
1500 raid10_write_one_disk(mddev
, r10_bio
, bio
, true, i
);
1502 one_write_done(r10_bio
);
1505 static void __make_request(struct mddev
*mddev
, struct bio
*bio
, int sectors
)
1507 struct r10conf
*conf
= mddev
->private;
1508 struct r10bio
*r10_bio
;
1510 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1512 r10_bio
->master_bio
= bio
;
1513 r10_bio
->sectors
= sectors
;
1515 r10_bio
->mddev
= mddev
;
1516 r10_bio
->sector
= bio
->bi_iter
.bi_sector
;
1518 memset(r10_bio
->devs
, 0, sizeof(r10_bio
->devs
[0]) * conf
->copies
);
1520 if (bio_data_dir(bio
) == READ
)
1521 raid10_read_request(mddev
, bio
, r10_bio
);
1523 raid10_write_request(mddev
, bio
, r10_bio
);
1526 static bool raid10_make_request(struct mddev
*mddev
, struct bio
*bio
)
1528 struct r10conf
*conf
= mddev
->private;
1529 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1530 int chunk_sects
= chunk_mask
+ 1;
1531 int sectors
= bio_sectors(bio
);
1533 if (unlikely(bio
->bi_opf
& REQ_PREFLUSH
)) {
1534 md_flush_request(mddev
, bio
);
1538 if (!md_write_start(mddev
, bio
))
1542 * If this request crosses a chunk boundary, we need to split
1545 if (unlikely((bio
->bi_iter
.bi_sector
& chunk_mask
) +
1546 sectors
> chunk_sects
1547 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1548 || conf
->prev
.near_copies
<
1549 conf
->prev
.raid_disks
)))
1550 sectors
= chunk_sects
-
1551 (bio
->bi_iter
.bi_sector
&
1553 __make_request(mddev
, bio
, sectors
);
1555 /* In case raid10d snuck in to freeze_array */
1556 wake_up(&conf
->wait_barrier
);
1560 static void raid10_status(struct seq_file
*seq
, struct mddev
*mddev
)
1562 struct r10conf
*conf
= mddev
->private;
1565 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1566 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1567 if (conf
->geo
.near_copies
> 1)
1568 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1569 if (conf
->geo
.far_copies
> 1) {
1570 if (conf
->geo
.far_offset
)
1571 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1573 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1574 if (conf
->geo
.far_set_size
!= conf
->geo
.raid_disks
)
1575 seq_printf(seq
, " %d devices per set", conf
->geo
.far_set_size
);
1577 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1578 conf
->geo
.raid_disks
- mddev
->degraded
);
1580 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1581 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1582 seq_printf(seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1585 seq_printf(seq
, "]");
1588 /* check if there are enough drives for
1589 * every block to appear on atleast one.
1590 * Don't consider the device numbered 'ignore'
1591 * as we might be about to remove it.
1593 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1599 disks
= conf
->prev
.raid_disks
;
1600 ncopies
= conf
->prev
.near_copies
;
1602 disks
= conf
->geo
.raid_disks
;
1603 ncopies
= conf
->geo
.near_copies
;
1608 int n
= conf
->copies
;
1612 struct md_rdev
*rdev
;
1613 if (this != ignore
&&
1614 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1615 test_bit(In_sync
, &rdev
->flags
))
1617 this = (this+1) % disks
;
1621 first
= (first
+ ncopies
) % disks
;
1622 } while (first
!= 0);
1629 static int enough(struct r10conf
*conf
, int ignore
)
1631 /* when calling 'enough', both 'prev' and 'geo' must
1633 * This is ensured if ->reconfig_mutex or ->device_lock
1636 return _enough(conf
, 0, ignore
) &&
1637 _enough(conf
, 1, ignore
);
1640 static void raid10_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1642 char b
[BDEVNAME_SIZE
];
1643 struct r10conf
*conf
= mddev
->private;
1644 unsigned long flags
;
1647 * If it is not operational, then we have already marked it as dead
1648 * else if it is the last working disks, ignore the error, let the
1649 * next level up know.
1650 * else mark the drive as failed
1652 spin_lock_irqsave(&conf
->device_lock
, flags
);
1653 if (test_bit(In_sync
, &rdev
->flags
)
1654 && !enough(conf
, rdev
->raid_disk
)) {
1656 * Don't fail the drive, just return an IO error.
1658 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1661 if (test_and_clear_bit(In_sync
, &rdev
->flags
))
1664 * If recovery is running, make sure it aborts.
1666 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1667 set_bit(Blocked
, &rdev
->flags
);
1668 set_bit(Faulty
, &rdev
->flags
);
1669 set_mask_bits(&mddev
->sb_flags
, 0,
1670 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
1671 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1672 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1673 "md/raid10:%s: Operation continuing on %d devices.\n",
1674 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1675 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1678 static void print_conf(struct r10conf
*conf
)
1681 struct md_rdev
*rdev
;
1683 pr_debug("RAID10 conf printout:\n");
1685 pr_debug("(!conf)\n");
1688 pr_debug(" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1689 conf
->geo
.raid_disks
);
1691 /* This is only called with ->reconfix_mutex held, so
1692 * rcu protection of rdev is not needed */
1693 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1694 char b
[BDEVNAME_SIZE
];
1695 rdev
= conf
->mirrors
[i
].rdev
;
1697 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1698 i
, !test_bit(In_sync
, &rdev
->flags
),
1699 !test_bit(Faulty
, &rdev
->flags
),
1700 bdevname(rdev
->bdev
,b
));
1704 static void close_sync(struct r10conf
*conf
)
1707 allow_barrier(conf
);
1709 mempool_destroy(conf
->r10buf_pool
);
1710 conf
->r10buf_pool
= NULL
;
1713 static int raid10_spare_active(struct mddev
*mddev
)
1716 struct r10conf
*conf
= mddev
->private;
1717 struct raid10_info
*tmp
;
1719 unsigned long flags
;
1722 * Find all non-in_sync disks within the RAID10 configuration
1723 * and mark them in_sync
1725 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1726 tmp
= conf
->mirrors
+ i
;
1727 if (tmp
->replacement
1728 && tmp
->replacement
->recovery_offset
== MaxSector
1729 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1730 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1731 /* Replacement has just become active */
1733 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1736 /* Replaced device not technically faulty,
1737 * but we need to be sure it gets removed
1738 * and never re-added.
1740 set_bit(Faulty
, &tmp
->rdev
->flags
);
1741 sysfs_notify_dirent_safe(
1742 tmp
->rdev
->sysfs_state
);
1744 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1745 } else if (tmp
->rdev
1746 && tmp
->rdev
->recovery_offset
== MaxSector
1747 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1748 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1750 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1753 spin_lock_irqsave(&conf
->device_lock
, flags
);
1754 mddev
->degraded
-= count
;
1755 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1761 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1763 struct r10conf
*conf
= mddev
->private;
1767 int last
= conf
->geo
.raid_disks
- 1;
1769 if (mddev
->recovery_cp
< MaxSector
)
1770 /* only hot-add to in-sync arrays, as recovery is
1771 * very different from resync
1774 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1777 if (md_integrity_add_rdev(rdev
, mddev
))
1780 if (rdev
->raid_disk
>= 0)
1781 first
= last
= rdev
->raid_disk
;
1783 if (rdev
->saved_raid_disk
>= first
&&
1784 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1785 mirror
= rdev
->saved_raid_disk
;
1788 for ( ; mirror
<= last
; mirror
++) {
1789 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1790 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1793 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1794 p
->replacement
!= NULL
)
1796 clear_bit(In_sync
, &rdev
->flags
);
1797 set_bit(Replacement
, &rdev
->flags
);
1798 rdev
->raid_disk
= mirror
;
1801 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1802 rdev
->data_offset
<< 9);
1804 rcu_assign_pointer(p
->replacement
, rdev
);
1809 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1810 rdev
->data_offset
<< 9);
1812 p
->head_position
= 0;
1813 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1814 rdev
->raid_disk
= mirror
;
1816 if (rdev
->saved_raid_disk
!= mirror
)
1818 rcu_assign_pointer(p
->rdev
, rdev
);
1821 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1822 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1828 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1830 struct r10conf
*conf
= mddev
->private;
1832 int number
= rdev
->raid_disk
;
1833 struct md_rdev
**rdevp
;
1834 struct raid10_info
*p
= conf
->mirrors
+ number
;
1837 if (rdev
== p
->rdev
)
1839 else if (rdev
== p
->replacement
)
1840 rdevp
= &p
->replacement
;
1844 if (test_bit(In_sync
, &rdev
->flags
) ||
1845 atomic_read(&rdev
->nr_pending
)) {
1849 /* Only remove non-faulty devices if recovery
1852 if (!test_bit(Faulty
, &rdev
->flags
) &&
1853 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1854 (!p
->replacement
|| p
->replacement
== rdev
) &&
1855 number
< conf
->geo
.raid_disks
&&
1861 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
1863 if (atomic_read(&rdev
->nr_pending
)) {
1864 /* lost the race, try later */
1870 if (p
->replacement
) {
1871 /* We must have just cleared 'rdev' */
1872 p
->rdev
= p
->replacement
;
1873 clear_bit(Replacement
, &p
->replacement
->flags
);
1874 smp_mb(); /* Make sure other CPUs may see both as identical
1875 * but will never see neither -- if they are careful.
1877 p
->replacement
= NULL
;
1880 clear_bit(WantReplacement
, &rdev
->flags
);
1881 err
= md_integrity_register(mddev
);
1889 static void __end_sync_read(struct r10bio
*r10_bio
, struct bio
*bio
, int d
)
1891 struct r10conf
*conf
= r10_bio
->mddev
->private;
1893 if (!bio
->bi_status
)
1894 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1896 /* The write handler will notice the lack of
1897 * R10BIO_Uptodate and record any errors etc
1899 atomic_add(r10_bio
->sectors
,
1900 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1902 /* for reconstruct, we always reschedule after a read.
1903 * for resync, only after all reads
1905 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1906 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1907 atomic_dec_and_test(&r10_bio
->remaining
)) {
1908 /* we have read all the blocks,
1909 * do the comparison in process context in raid10d
1911 reschedule_retry(r10_bio
);
1915 static void end_sync_read(struct bio
*bio
)
1917 struct r10bio
*r10_bio
= get_resync_r10bio(bio
);
1918 struct r10conf
*conf
= r10_bio
->mddev
->private;
1919 int d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1921 __end_sync_read(r10_bio
, bio
, d
);
1924 static void end_reshape_read(struct bio
*bio
)
1926 /* reshape read bio isn't allocated from r10buf_pool */
1927 struct r10bio
*r10_bio
= bio
->bi_private
;
1929 __end_sync_read(r10_bio
, bio
, r10_bio
->read_slot
);
1932 static void end_sync_request(struct r10bio
*r10_bio
)
1934 struct mddev
*mddev
= r10_bio
->mddev
;
1936 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1937 if (r10_bio
->master_bio
== NULL
) {
1938 /* the primary of several recovery bios */
1939 sector_t s
= r10_bio
->sectors
;
1940 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1941 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1942 reschedule_retry(r10_bio
);
1945 md_done_sync(mddev
, s
, 1);
1948 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1949 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1950 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1951 reschedule_retry(r10_bio
);
1959 static void end_sync_write(struct bio
*bio
)
1961 struct r10bio
*r10_bio
= get_resync_r10bio(bio
);
1962 struct mddev
*mddev
= r10_bio
->mddev
;
1963 struct r10conf
*conf
= mddev
->private;
1969 struct md_rdev
*rdev
= NULL
;
1971 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1973 rdev
= conf
->mirrors
[d
].replacement
;
1975 rdev
= conf
->mirrors
[d
].rdev
;
1977 if (bio
->bi_status
) {
1979 md_error(mddev
, rdev
);
1981 set_bit(WriteErrorSeen
, &rdev
->flags
);
1982 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1983 set_bit(MD_RECOVERY_NEEDED
,
1984 &rdev
->mddev
->recovery
);
1985 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1987 } else if (is_badblock(rdev
,
1988 r10_bio
->devs
[slot
].addr
,
1990 &first_bad
, &bad_sectors
))
1991 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1993 rdev_dec_pending(rdev
, mddev
);
1995 end_sync_request(r10_bio
);
1999 * Note: sync and recover and handled very differently for raid10
2000 * This code is for resync.
2001 * For resync, we read through virtual addresses and read all blocks.
2002 * If there is any error, we schedule a write. The lowest numbered
2003 * drive is authoritative.
2004 * However requests come for physical address, so we need to map.
2005 * For every physical address there are raid_disks/copies virtual addresses,
2006 * which is always are least one, but is not necessarly an integer.
2007 * This means that a physical address can span multiple chunks, so we may
2008 * have to submit multiple io requests for a single sync request.
2011 * We check if all blocks are in-sync and only write to blocks that
2014 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2016 struct r10conf
*conf
= mddev
->private;
2018 struct bio
*tbio
, *fbio
;
2020 struct page
**tpages
, **fpages
;
2022 atomic_set(&r10_bio
->remaining
, 1);
2024 /* find the first device with a block */
2025 for (i
=0; i
<conf
->copies
; i
++)
2026 if (!r10_bio
->devs
[i
].bio
->bi_status
)
2029 if (i
== conf
->copies
)
2033 fbio
= r10_bio
->devs
[i
].bio
;
2034 fbio
->bi_iter
.bi_size
= r10_bio
->sectors
<< 9;
2035 fbio
->bi_iter
.bi_idx
= 0;
2036 fpages
= get_resync_pages(fbio
)->pages
;
2038 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2039 /* now find blocks with errors */
2040 for (i
=0 ; i
< conf
->copies
; i
++) {
2042 struct md_rdev
*rdev
;
2043 struct resync_pages
*rp
;
2045 tbio
= r10_bio
->devs
[i
].bio
;
2047 if (tbio
->bi_end_io
!= end_sync_read
)
2052 tpages
= get_resync_pages(tbio
)->pages
;
2053 d
= r10_bio
->devs
[i
].devnum
;
2054 rdev
= conf
->mirrors
[d
].rdev
;
2055 if (!r10_bio
->devs
[i
].bio
->bi_status
) {
2056 /* We know that the bi_io_vec layout is the same for
2057 * both 'first' and 'i', so we just compare them.
2058 * All vec entries are PAGE_SIZE;
2060 int sectors
= r10_bio
->sectors
;
2061 for (j
= 0; j
< vcnt
; j
++) {
2062 int len
= PAGE_SIZE
;
2063 if (sectors
< (len
/ 512))
2064 len
= sectors
* 512;
2065 if (memcmp(page_address(fpages
[j
]),
2066 page_address(tpages
[j
]),
2073 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2074 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2075 /* Don't fix anything. */
2077 } else if (test_bit(FailFast
, &rdev
->flags
)) {
2078 /* Just give up on this device */
2079 md_error(rdev
->mddev
, rdev
);
2082 /* Ok, we need to write this bio, either to correct an
2083 * inconsistency or to correct an unreadable block.
2084 * First we need to fixup bv_offset, bv_len and
2085 * bi_vecs, as the read request might have corrupted these
2087 rp
= get_resync_pages(tbio
);
2090 tbio
->bi_vcnt
= vcnt
;
2091 tbio
->bi_iter
.bi_size
= fbio
->bi_iter
.bi_size
;
2092 rp
->raid_bio
= r10_bio
;
2093 tbio
->bi_private
= rp
;
2094 tbio
->bi_iter
.bi_sector
= r10_bio
->devs
[i
].addr
;
2095 tbio
->bi_end_io
= end_sync_write
;
2096 bio_set_op_attrs(tbio
, REQ_OP_WRITE
, 0);
2098 bio_copy_data(tbio
, fbio
);
2100 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2101 atomic_inc(&r10_bio
->remaining
);
2102 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2104 if (test_bit(FailFast
, &conf
->mirrors
[d
].rdev
->flags
))
2105 tbio
->bi_opf
|= MD_FAILFAST
;
2106 tbio
->bi_iter
.bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2107 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2108 generic_make_request(tbio
);
2111 /* Now write out to any replacement devices
2114 for (i
= 0; i
< conf
->copies
; i
++) {
2117 tbio
= r10_bio
->devs
[i
].repl_bio
;
2118 if (!tbio
|| !tbio
->bi_end_io
)
2120 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2121 && r10_bio
->devs
[i
].bio
!= fbio
)
2122 bio_copy_data(tbio
, fbio
);
2123 d
= r10_bio
->devs
[i
].devnum
;
2124 atomic_inc(&r10_bio
->remaining
);
2125 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2127 generic_make_request(tbio
);
2131 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2132 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2138 * Now for the recovery code.
2139 * Recovery happens across physical sectors.
2140 * We recover all non-is_sync drives by finding the virtual address of
2141 * each, and then choose a working drive that also has that virt address.
2142 * There is a separate r10_bio for each non-in_sync drive.
2143 * Only the first two slots are in use. The first for reading,
2144 * The second for writing.
2147 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2149 /* We got a read error during recovery.
2150 * We repeat the read in smaller page-sized sections.
2151 * If a read succeeds, write it to the new device or record
2152 * a bad block if we cannot.
2153 * If a read fails, record a bad block on both old and
2156 struct mddev
*mddev
= r10_bio
->mddev
;
2157 struct r10conf
*conf
= mddev
->private;
2158 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2160 int sectors
= r10_bio
->sectors
;
2162 int dr
= r10_bio
->devs
[0].devnum
;
2163 int dw
= r10_bio
->devs
[1].devnum
;
2164 struct page
**pages
= get_resync_pages(bio
)->pages
;
2168 struct md_rdev
*rdev
;
2172 if (s
> (PAGE_SIZE
>>9))
2175 rdev
= conf
->mirrors
[dr
].rdev
;
2176 addr
= r10_bio
->devs
[0].addr
+ sect
,
2177 ok
= sync_page_io(rdev
,
2181 REQ_OP_READ
, 0, false);
2183 rdev
= conf
->mirrors
[dw
].rdev
;
2184 addr
= r10_bio
->devs
[1].addr
+ sect
;
2185 ok
= sync_page_io(rdev
,
2189 REQ_OP_WRITE
, 0, false);
2191 set_bit(WriteErrorSeen
, &rdev
->flags
);
2192 if (!test_and_set_bit(WantReplacement
,
2194 set_bit(MD_RECOVERY_NEEDED
,
2195 &rdev
->mddev
->recovery
);
2199 /* We don't worry if we cannot set a bad block -
2200 * it really is bad so there is no loss in not
2203 rdev_set_badblocks(rdev
, addr
, s
, 0);
2205 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2206 /* need bad block on destination too */
2207 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2208 addr
= r10_bio
->devs
[1].addr
+ sect
;
2209 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2211 /* just abort the recovery */
2212 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2215 conf
->mirrors
[dw
].recovery_disabled
2216 = mddev
->recovery_disabled
;
2217 set_bit(MD_RECOVERY_INTR
,
2230 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2232 struct r10conf
*conf
= mddev
->private;
2234 struct bio
*wbio
, *wbio2
;
2236 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2237 fix_recovery_read_error(r10_bio
);
2238 end_sync_request(r10_bio
);
2243 * share the pages with the first bio
2244 * and submit the write request
2246 d
= r10_bio
->devs
[1].devnum
;
2247 wbio
= r10_bio
->devs
[1].bio
;
2248 wbio2
= r10_bio
->devs
[1].repl_bio
;
2249 /* Need to test wbio2->bi_end_io before we call
2250 * generic_make_request as if the former is NULL,
2251 * the latter is free to free wbio2.
2253 if (wbio2
&& !wbio2
->bi_end_io
)
2255 if (wbio
->bi_end_io
) {
2256 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2257 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2258 generic_make_request(wbio
);
2261 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2262 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2263 bio_sectors(wbio2
));
2264 generic_make_request(wbio2
);
2269 * Used by fix_read_error() to decay the per rdev read_errors.
2270 * We halve the read error count for every hour that has elapsed
2271 * since the last recorded read error.
2274 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2277 unsigned long hours_since_last
;
2278 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2280 cur_time_mon
= ktime_get_seconds();
2282 if (rdev
->last_read_error
== 0) {
2283 /* first time we've seen a read error */
2284 rdev
->last_read_error
= cur_time_mon
;
2288 hours_since_last
= (long)(cur_time_mon
-
2289 rdev
->last_read_error
) / 3600;
2291 rdev
->last_read_error
= cur_time_mon
;
2294 * if hours_since_last is > the number of bits in read_errors
2295 * just set read errors to 0. We do this to avoid
2296 * overflowing the shift of read_errors by hours_since_last.
2298 if (hours_since_last
>= 8 * sizeof(read_errors
))
2299 atomic_set(&rdev
->read_errors
, 0);
2301 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2304 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2305 int sectors
, struct page
*page
, int rw
)
2310 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2311 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2313 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, 0, false))
2317 set_bit(WriteErrorSeen
, &rdev
->flags
);
2318 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2319 set_bit(MD_RECOVERY_NEEDED
,
2320 &rdev
->mddev
->recovery
);
2322 /* need to record an error - either for the block or the device */
2323 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2324 md_error(rdev
->mddev
, rdev
);
2329 * This is a kernel thread which:
2331 * 1. Retries failed read operations on working mirrors.
2332 * 2. Updates the raid superblock when problems encounter.
2333 * 3. Performs writes following reads for array synchronising.
2336 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2338 int sect
= 0; /* Offset from r10_bio->sector */
2339 int sectors
= r10_bio
->sectors
;
2340 struct md_rdev
*rdev
;
2341 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2342 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2344 /* still own a reference to this rdev, so it cannot
2345 * have been cleared recently.
2347 rdev
= conf
->mirrors
[d
].rdev
;
2349 if (test_bit(Faulty
, &rdev
->flags
))
2350 /* drive has already been failed, just ignore any
2351 more fix_read_error() attempts */
2354 check_decay_read_errors(mddev
, rdev
);
2355 atomic_inc(&rdev
->read_errors
);
2356 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2357 char b
[BDEVNAME_SIZE
];
2358 bdevname(rdev
->bdev
, b
);
2360 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2362 atomic_read(&rdev
->read_errors
), max_read_errors
);
2363 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2365 md_error(mddev
, rdev
);
2366 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2372 int sl
= r10_bio
->read_slot
;
2376 if (s
> (PAGE_SIZE
>>9))
2384 d
= r10_bio
->devs
[sl
].devnum
;
2385 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2387 test_bit(In_sync
, &rdev
->flags
) &&
2388 !test_bit(Faulty
, &rdev
->flags
) &&
2389 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2390 &first_bad
, &bad_sectors
) == 0) {
2391 atomic_inc(&rdev
->nr_pending
);
2393 success
= sync_page_io(rdev
,
2394 r10_bio
->devs
[sl
].addr
+
2398 REQ_OP_READ
, 0, false);
2399 rdev_dec_pending(rdev
, mddev
);
2405 if (sl
== conf
->copies
)
2407 } while (!success
&& sl
!= r10_bio
->read_slot
);
2411 /* Cannot read from anywhere, just mark the block
2412 * as bad on the first device to discourage future
2415 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2416 rdev
= conf
->mirrors
[dn
].rdev
;
2418 if (!rdev_set_badblocks(
2420 r10_bio
->devs
[r10_bio
->read_slot
].addr
2423 md_error(mddev
, rdev
);
2424 r10_bio
->devs
[r10_bio
->read_slot
].bio
2431 /* write it back and re-read */
2433 while (sl
!= r10_bio
->read_slot
) {
2434 char b
[BDEVNAME_SIZE
];
2439 d
= r10_bio
->devs
[sl
].devnum
;
2440 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2442 test_bit(Faulty
, &rdev
->flags
) ||
2443 !test_bit(In_sync
, &rdev
->flags
))
2446 atomic_inc(&rdev
->nr_pending
);
2448 if (r10_sync_page_io(rdev
,
2449 r10_bio
->devs
[sl
].addr
+
2451 s
, conf
->tmppage
, WRITE
)
2453 /* Well, this device is dead */
2454 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2456 (unsigned long long)(
2458 choose_data_offset(r10_bio
,
2460 bdevname(rdev
->bdev
, b
));
2461 pr_notice("md/raid10:%s: %s: failing drive\n",
2463 bdevname(rdev
->bdev
, b
));
2465 rdev_dec_pending(rdev
, mddev
);
2469 while (sl
!= r10_bio
->read_slot
) {
2470 char b
[BDEVNAME_SIZE
];
2475 d
= r10_bio
->devs
[sl
].devnum
;
2476 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2478 test_bit(Faulty
, &rdev
->flags
) ||
2479 !test_bit(In_sync
, &rdev
->flags
))
2482 atomic_inc(&rdev
->nr_pending
);
2484 switch (r10_sync_page_io(rdev
,
2485 r10_bio
->devs
[sl
].addr
+
2490 /* Well, this device is dead */
2491 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2493 (unsigned long long)(
2495 choose_data_offset(r10_bio
, rdev
)),
2496 bdevname(rdev
->bdev
, b
));
2497 pr_notice("md/raid10:%s: %s: failing drive\n",
2499 bdevname(rdev
->bdev
, b
));
2502 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2504 (unsigned long long)(
2506 choose_data_offset(r10_bio
, rdev
)),
2507 bdevname(rdev
->bdev
, b
));
2508 atomic_add(s
, &rdev
->corrected_errors
);
2511 rdev_dec_pending(rdev
, mddev
);
2521 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2523 struct bio
*bio
= r10_bio
->master_bio
;
2524 struct mddev
*mddev
= r10_bio
->mddev
;
2525 struct r10conf
*conf
= mddev
->private;
2526 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2527 /* bio has the data to be written to slot 'i' where
2528 * we just recently had a write error.
2529 * We repeatedly clone the bio and trim down to one block,
2530 * then try the write. Where the write fails we record
2532 * It is conceivable that the bio doesn't exactly align with
2533 * blocks. We must handle this.
2535 * We currently own a reference to the rdev.
2541 int sect_to_write
= r10_bio
->sectors
;
2544 if (rdev
->badblocks
.shift
< 0)
2547 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2548 bdev_logical_block_size(rdev
->bdev
) >> 9);
2549 sector
= r10_bio
->sector
;
2550 sectors
= ((r10_bio
->sector
+ block_sectors
)
2551 & ~(sector_t
)(block_sectors
- 1))
2554 while (sect_to_write
) {
2557 if (sectors
> sect_to_write
)
2558 sectors
= sect_to_write
;
2559 /* Write at 'sector' for 'sectors' */
2560 wbio
= bio_clone_fast(bio
, GFP_NOIO
, mddev
->bio_set
);
2561 bio_trim(wbio
, sector
- bio
->bi_iter
.bi_sector
, sectors
);
2562 wsector
= r10_bio
->devs
[i
].addr
+ (sector
- r10_bio
->sector
);
2563 wbio
->bi_iter
.bi_sector
= wsector
+
2564 choose_data_offset(r10_bio
, rdev
);
2565 wbio
->bi_bdev
= rdev
->bdev
;
2566 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2568 if (submit_bio_wait(wbio
) < 0)
2570 ok
= rdev_set_badblocks(rdev
, wsector
,
2575 sect_to_write
-= sectors
;
2577 sectors
= block_sectors
;
2582 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2584 int slot
= r10_bio
->read_slot
;
2586 struct r10conf
*conf
= mddev
->private;
2587 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2589 sector_t bio_last_sector
;
2591 /* we got a read error. Maybe the drive is bad. Maybe just
2592 * the block and we can fix it.
2593 * We freeze all other IO, and try reading the block from
2594 * other devices. When we find one, we re-write
2595 * and check it that fixes the read error.
2596 * This is all done synchronously while the array is
2599 bio
= r10_bio
->devs
[slot
].bio
;
2600 bio_dev
= bio
->bi_bdev
->bd_dev
;
2601 bio_last_sector
= r10_bio
->devs
[slot
].addr
+ rdev
->data_offset
+ r10_bio
->sectors
;
2603 r10_bio
->devs
[slot
].bio
= NULL
;
2606 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2607 else if (!test_bit(FailFast
, &rdev
->flags
)) {
2608 freeze_array(conf
, 1);
2609 fix_read_error(conf
, mddev
, r10_bio
);
2610 unfreeze_array(conf
);
2612 md_error(mddev
, rdev
);
2614 rdev_dec_pending(rdev
, mddev
);
2615 allow_barrier(conf
);
2617 raid10_read_request(mddev
, r10_bio
->master_bio
, r10_bio
);
2620 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2622 /* Some sort of write request has finished and it
2623 * succeeded in writing where we thought there was a
2624 * bad block. So forget the bad block.
2625 * Or possibly if failed and we need to record
2629 struct md_rdev
*rdev
;
2631 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2632 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2633 for (m
= 0; m
< conf
->copies
; m
++) {
2634 int dev
= r10_bio
->devs
[m
].devnum
;
2635 rdev
= conf
->mirrors
[dev
].rdev
;
2636 if (r10_bio
->devs
[m
].bio
== NULL
)
2638 if (!r10_bio
->devs
[m
].bio
->bi_status
) {
2639 rdev_clear_badblocks(
2641 r10_bio
->devs
[m
].addr
,
2642 r10_bio
->sectors
, 0);
2644 if (!rdev_set_badblocks(
2646 r10_bio
->devs
[m
].addr
,
2647 r10_bio
->sectors
, 0))
2648 md_error(conf
->mddev
, rdev
);
2650 rdev
= conf
->mirrors
[dev
].replacement
;
2651 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2654 if (!r10_bio
->devs
[m
].repl_bio
->bi_status
) {
2655 rdev_clear_badblocks(
2657 r10_bio
->devs
[m
].addr
,
2658 r10_bio
->sectors
, 0);
2660 if (!rdev_set_badblocks(
2662 r10_bio
->devs
[m
].addr
,
2663 r10_bio
->sectors
, 0))
2664 md_error(conf
->mddev
, rdev
);
2670 for (m
= 0; m
< conf
->copies
; m
++) {
2671 int dev
= r10_bio
->devs
[m
].devnum
;
2672 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2673 rdev
= conf
->mirrors
[dev
].rdev
;
2674 if (bio
== IO_MADE_GOOD
) {
2675 rdev_clear_badblocks(
2677 r10_bio
->devs
[m
].addr
,
2678 r10_bio
->sectors
, 0);
2679 rdev_dec_pending(rdev
, conf
->mddev
);
2680 } else if (bio
!= NULL
&& bio
->bi_status
) {
2682 if (!narrow_write_error(r10_bio
, m
)) {
2683 md_error(conf
->mddev
, rdev
);
2684 set_bit(R10BIO_Degraded
,
2687 rdev_dec_pending(rdev
, conf
->mddev
);
2689 bio
= r10_bio
->devs
[m
].repl_bio
;
2690 rdev
= conf
->mirrors
[dev
].replacement
;
2691 if (rdev
&& bio
== IO_MADE_GOOD
) {
2692 rdev_clear_badblocks(
2694 r10_bio
->devs
[m
].addr
,
2695 r10_bio
->sectors
, 0);
2696 rdev_dec_pending(rdev
, conf
->mddev
);
2700 spin_lock_irq(&conf
->device_lock
);
2701 list_add(&r10_bio
->retry_list
, &conf
->bio_end_io_list
);
2703 spin_unlock_irq(&conf
->device_lock
);
2705 * In case freeze_array() is waiting for condition
2706 * nr_pending == nr_queued + extra to be true.
2708 wake_up(&conf
->wait_barrier
);
2709 md_wakeup_thread(conf
->mddev
->thread
);
2711 if (test_bit(R10BIO_WriteError
,
2713 close_write(r10_bio
);
2714 raid_end_bio_io(r10_bio
);
2719 static void raid10d(struct md_thread
*thread
)
2721 struct mddev
*mddev
= thread
->mddev
;
2722 struct r10bio
*r10_bio
;
2723 unsigned long flags
;
2724 struct r10conf
*conf
= mddev
->private;
2725 struct list_head
*head
= &conf
->retry_list
;
2726 struct blk_plug plug
;
2728 md_check_recovery(mddev
);
2730 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2731 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
2733 spin_lock_irqsave(&conf
->device_lock
, flags
);
2734 if (!test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
2735 while (!list_empty(&conf
->bio_end_io_list
)) {
2736 list_move(conf
->bio_end_io_list
.prev
, &tmp
);
2740 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2741 while (!list_empty(&tmp
)) {
2742 r10_bio
= list_first_entry(&tmp
, struct r10bio
,
2744 list_del(&r10_bio
->retry_list
);
2745 if (mddev
->degraded
)
2746 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
2748 if (test_bit(R10BIO_WriteError
,
2750 close_write(r10_bio
);
2751 raid_end_bio_io(r10_bio
);
2755 blk_start_plug(&plug
);
2758 flush_pending_writes(conf
);
2760 spin_lock_irqsave(&conf
->device_lock
, flags
);
2761 if (list_empty(head
)) {
2762 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2765 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2766 list_del(head
->prev
);
2768 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2770 mddev
= r10_bio
->mddev
;
2771 conf
= mddev
->private;
2772 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2773 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2774 handle_write_completed(conf
, r10_bio
);
2775 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2776 reshape_request_write(mddev
, r10_bio
);
2777 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2778 sync_request_write(mddev
, r10_bio
);
2779 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2780 recovery_request_write(mddev
, r10_bio
);
2781 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2782 handle_read_error(mddev
, r10_bio
);
2787 if (mddev
->sb_flags
& ~(1<<MD_SB_CHANGE_PENDING
))
2788 md_check_recovery(mddev
);
2790 blk_finish_plug(&plug
);
2793 static int init_resync(struct r10conf
*conf
)
2798 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2799 BUG_ON(conf
->r10buf_pool
);
2800 conf
->have_replacement
= 0;
2801 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2802 if (conf
->mirrors
[i
].replacement
)
2803 conf
->have_replacement
= 1;
2804 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2805 if (!conf
->r10buf_pool
)
2807 conf
->next_resync
= 0;
2812 * perform a "sync" on one "block"
2814 * We need to make sure that no normal I/O request - particularly write
2815 * requests - conflict with active sync requests.
2817 * This is achieved by tracking pending requests and a 'barrier' concept
2818 * that can be installed to exclude normal IO requests.
2820 * Resync and recovery are handled very differently.
2821 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2823 * For resync, we iterate over virtual addresses, read all copies,
2824 * and update if there are differences. If only one copy is live,
2826 * For recovery, we iterate over physical addresses, read a good
2827 * value for each non-in_sync drive, and over-write.
2829 * So, for recovery we may have several outstanding complex requests for a
2830 * given address, one for each out-of-sync device. We model this by allocating
2831 * a number of r10_bio structures, one for each out-of-sync device.
2832 * As we setup these structures, we collect all bio's together into a list
2833 * which we then process collectively to add pages, and then process again
2834 * to pass to generic_make_request.
2836 * The r10_bio structures are linked using a borrowed master_bio pointer.
2837 * This link is counted in ->remaining. When the r10_bio that points to NULL
2838 * has its remaining count decremented to 0, the whole complex operation
2843 static sector_t
raid10_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2846 struct r10conf
*conf
= mddev
->private;
2847 struct r10bio
*r10_bio
;
2848 struct bio
*biolist
= NULL
, *bio
;
2849 sector_t max_sector
, nr_sectors
;
2852 sector_t sync_blocks
;
2853 sector_t sectors_skipped
= 0;
2854 int chunks_skipped
= 0;
2855 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2857 if (!conf
->r10buf_pool
)
2858 if (init_resync(conf
))
2862 * Allow skipping a full rebuild for incremental assembly
2863 * of a clean array, like RAID1 does.
2865 if (mddev
->bitmap
== NULL
&&
2866 mddev
->recovery_cp
== MaxSector
&&
2867 mddev
->reshape_position
== MaxSector
&&
2868 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2869 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2870 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2871 conf
->fullsync
== 0) {
2873 return mddev
->dev_sectors
- sector_nr
;
2877 max_sector
= mddev
->dev_sectors
;
2878 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2879 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2880 max_sector
= mddev
->resync_max_sectors
;
2881 if (sector_nr
>= max_sector
) {
2882 /* If we aborted, we need to abort the
2883 * sync on the 'current' bitmap chucks (there can
2884 * be several when recovering multiple devices).
2885 * as we may have started syncing it but not finished.
2886 * We can find the current address in
2887 * mddev->curr_resync, but for recovery,
2888 * we need to convert that to several
2889 * virtual addresses.
2891 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2897 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2898 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2899 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2901 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2903 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2904 bitmap_end_sync(mddev
->bitmap
, sect
,
2908 /* completed sync */
2909 if ((!mddev
->bitmap
|| conf
->fullsync
)
2910 && conf
->have_replacement
2911 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2912 /* Completed a full sync so the replacements
2913 * are now fully recovered.
2916 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2917 struct md_rdev
*rdev
=
2918 rcu_dereference(conf
->mirrors
[i
].replacement
);
2920 rdev
->recovery_offset
= MaxSector
;
2926 bitmap_close_sync(mddev
->bitmap
);
2929 return sectors_skipped
;
2932 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2933 return reshape_request(mddev
, sector_nr
, skipped
);
2935 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2936 /* if there has been nothing to do on any drive,
2937 * then there is nothing to do at all..
2940 return (max_sector
- sector_nr
) + sectors_skipped
;
2943 if (max_sector
> mddev
->resync_max
)
2944 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2946 /* make sure whole request will fit in a chunk - if chunks
2949 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2950 max_sector
> (sector_nr
| chunk_mask
))
2951 max_sector
= (sector_nr
| chunk_mask
) + 1;
2954 * If there is non-resync activity waiting for a turn, then let it
2955 * though before starting on this new sync request.
2957 if (conf
->nr_waiting
)
2958 schedule_timeout_uninterruptible(1);
2960 /* Again, very different code for resync and recovery.
2961 * Both must result in an r10bio with a list of bios that
2962 * have bi_end_io, bi_sector, bi_bdev set,
2963 * and bi_private set to the r10bio.
2964 * For recovery, we may actually create several r10bios
2965 * with 2 bios in each, that correspond to the bios in the main one.
2966 * In this case, the subordinate r10bios link back through a
2967 * borrowed master_bio pointer, and the counter in the master
2968 * includes a ref from each subordinate.
2970 /* First, we decide what to do and set ->bi_end_io
2971 * To end_sync_read if we want to read, and
2972 * end_sync_write if we will want to write.
2975 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2976 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2977 /* recovery... the complicated one */
2981 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2987 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2988 struct md_rdev
*mrdev
, *mreplace
;
2991 mrdev
= rcu_dereference(mirror
->rdev
);
2992 mreplace
= rcu_dereference(mirror
->replacement
);
2994 if ((mrdev
== NULL
||
2995 test_bit(Faulty
, &mrdev
->flags
) ||
2996 test_bit(In_sync
, &mrdev
->flags
)) &&
2997 (mreplace
== NULL
||
2998 test_bit(Faulty
, &mreplace
->flags
))) {
3004 /* want to reconstruct this device */
3006 sect
= raid10_find_virt(conf
, sector_nr
, i
);
3007 if (sect
>= mddev
->resync_max_sectors
) {
3008 /* last stripe is not complete - don't
3009 * try to recover this sector.
3014 if (mreplace
&& test_bit(Faulty
, &mreplace
->flags
))
3016 /* Unless we are doing a full sync, or a replacement
3017 * we only need to recover the block if it is set in
3020 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3022 if (sync_blocks
< max_sync
)
3023 max_sync
= sync_blocks
;
3027 /* yep, skip the sync_blocks here, but don't assume
3028 * that there will never be anything to do here
3030 chunks_skipped
= -1;
3034 atomic_inc(&mrdev
->nr_pending
);
3036 atomic_inc(&mreplace
->nr_pending
);
3039 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3041 raise_barrier(conf
, rb2
!= NULL
);
3042 atomic_set(&r10_bio
->remaining
, 0);
3044 r10_bio
->master_bio
= (struct bio
*)rb2
;
3046 atomic_inc(&rb2
->remaining
);
3047 r10_bio
->mddev
= mddev
;
3048 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3049 r10_bio
->sector
= sect
;
3051 raid10_find_phys(conf
, r10_bio
);
3053 /* Need to check if the array will still be
3057 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++) {
3058 struct md_rdev
*rdev
= rcu_dereference(
3059 conf
->mirrors
[j
].rdev
);
3060 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3066 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3067 &sync_blocks
, still_degraded
);
3070 for (j
=0; j
<conf
->copies
;j
++) {
3072 int d
= r10_bio
->devs
[j
].devnum
;
3073 sector_t from_addr
, to_addr
;
3074 struct md_rdev
*rdev
=
3075 rcu_dereference(conf
->mirrors
[d
].rdev
);
3076 sector_t sector
, first_bad
;
3079 !test_bit(In_sync
, &rdev
->flags
))
3081 /* This is where we read from */
3083 sector
= r10_bio
->devs
[j
].addr
;
3085 if (is_badblock(rdev
, sector
, max_sync
,
3086 &first_bad
, &bad_sectors
)) {
3087 if (first_bad
> sector
)
3088 max_sync
= first_bad
- sector
;
3090 bad_sectors
-= (sector
3092 if (max_sync
> bad_sectors
)
3093 max_sync
= bad_sectors
;
3097 bio
= r10_bio
->devs
[0].bio
;
3098 bio
->bi_next
= biolist
;
3100 bio
->bi_end_io
= end_sync_read
;
3101 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3102 if (test_bit(FailFast
, &rdev
->flags
))
3103 bio
->bi_opf
|= MD_FAILFAST
;
3104 from_addr
= r10_bio
->devs
[j
].addr
;
3105 bio
->bi_iter
.bi_sector
= from_addr
+
3107 bio
->bi_bdev
= rdev
->bdev
;
3108 atomic_inc(&rdev
->nr_pending
);
3109 /* and we write to 'i' (if not in_sync) */
3111 for (k
=0; k
<conf
->copies
; k
++)
3112 if (r10_bio
->devs
[k
].devnum
== i
)
3114 BUG_ON(k
== conf
->copies
);
3115 to_addr
= r10_bio
->devs
[k
].addr
;
3116 r10_bio
->devs
[0].devnum
= d
;
3117 r10_bio
->devs
[0].addr
= from_addr
;
3118 r10_bio
->devs
[1].devnum
= i
;
3119 r10_bio
->devs
[1].addr
= to_addr
;
3121 if (!test_bit(In_sync
, &mrdev
->flags
)) {
3122 bio
= r10_bio
->devs
[1].bio
;
3123 bio
->bi_next
= biolist
;
3125 bio
->bi_end_io
= end_sync_write
;
3126 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3127 bio
->bi_iter
.bi_sector
= to_addr
3128 + mrdev
->data_offset
;
3129 bio
->bi_bdev
= mrdev
->bdev
;
3130 atomic_inc(&r10_bio
->remaining
);
3132 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3134 /* and maybe write to replacement */
3135 bio
= r10_bio
->devs
[1].repl_bio
;
3137 bio
->bi_end_io
= NULL
;
3138 /* Note: if mreplace != NULL, then bio
3139 * cannot be NULL as r10buf_pool_alloc will
3140 * have allocated it.
3141 * So the second test here is pointless.
3142 * But it keeps semantic-checkers happy, and
3143 * this comment keeps human reviewers
3146 if (mreplace
== NULL
|| bio
== NULL
||
3147 test_bit(Faulty
, &mreplace
->flags
))
3149 bio
->bi_next
= biolist
;
3151 bio
->bi_end_io
= end_sync_write
;
3152 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3153 bio
->bi_iter
.bi_sector
= to_addr
+
3154 mreplace
->data_offset
;
3155 bio
->bi_bdev
= mreplace
->bdev
;
3156 atomic_inc(&r10_bio
->remaining
);
3160 if (j
== conf
->copies
) {
3161 /* Cannot recover, so abort the recovery or
3162 * record a bad block */
3164 /* problem is that there are bad blocks
3165 * on other device(s)
3168 for (k
= 0; k
< conf
->copies
; k
++)
3169 if (r10_bio
->devs
[k
].devnum
== i
)
3171 if (!test_bit(In_sync
,
3173 && !rdev_set_badblocks(
3175 r10_bio
->devs
[k
].addr
,
3179 !rdev_set_badblocks(
3181 r10_bio
->devs
[k
].addr
,
3186 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3188 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3190 mirror
->recovery_disabled
3191 = mddev
->recovery_disabled
;
3195 atomic_dec(&rb2
->remaining
);
3197 rdev_dec_pending(mrdev
, mddev
);
3199 rdev_dec_pending(mreplace
, mddev
);
3202 rdev_dec_pending(mrdev
, mddev
);
3204 rdev_dec_pending(mreplace
, mddev
);
3205 if (r10_bio
->devs
[0].bio
->bi_opf
& MD_FAILFAST
) {
3206 /* Only want this if there is elsewhere to
3207 * read from. 'j' is currently the first
3211 for (; j
< conf
->copies
; j
++) {
3212 int d
= r10_bio
->devs
[j
].devnum
;
3213 if (conf
->mirrors
[d
].rdev
&&
3215 &conf
->mirrors
[d
].rdev
->flags
))
3219 r10_bio
->devs
[0].bio
->bi_opf
3223 if (biolist
== NULL
) {
3225 struct r10bio
*rb2
= r10_bio
;
3226 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3227 rb2
->master_bio
= NULL
;
3233 /* resync. Schedule a read for every block at this virt offset */
3236 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, 0);
3238 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3239 &sync_blocks
, mddev
->degraded
) &&
3240 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3241 &mddev
->recovery
)) {
3242 /* We can skip this block */
3244 return sync_blocks
+ sectors_skipped
;
3246 if (sync_blocks
< max_sync
)
3247 max_sync
= sync_blocks
;
3248 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3251 r10_bio
->mddev
= mddev
;
3252 atomic_set(&r10_bio
->remaining
, 0);
3253 raise_barrier(conf
, 0);
3254 conf
->next_resync
= sector_nr
;
3256 r10_bio
->master_bio
= NULL
;
3257 r10_bio
->sector
= sector_nr
;
3258 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3259 raid10_find_phys(conf
, r10_bio
);
3260 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3262 for (i
= 0; i
< conf
->copies
; i
++) {
3263 int d
= r10_bio
->devs
[i
].devnum
;
3264 sector_t first_bad
, sector
;
3266 struct md_rdev
*rdev
;
3268 if (r10_bio
->devs
[i
].repl_bio
)
3269 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3271 bio
= r10_bio
->devs
[i
].bio
;
3272 bio
->bi_status
= BLK_STS_IOERR
;
3274 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
3275 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3279 sector
= r10_bio
->devs
[i
].addr
;
3280 if (is_badblock(rdev
, sector
, max_sync
,
3281 &first_bad
, &bad_sectors
)) {
3282 if (first_bad
> sector
)
3283 max_sync
= first_bad
- sector
;
3285 bad_sectors
-= (sector
- first_bad
);
3286 if (max_sync
> bad_sectors
)
3287 max_sync
= bad_sectors
;
3292 atomic_inc(&rdev
->nr_pending
);
3293 atomic_inc(&r10_bio
->remaining
);
3294 bio
->bi_next
= biolist
;
3296 bio
->bi_end_io
= end_sync_read
;
3297 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3298 if (test_bit(FailFast
, &rdev
->flags
))
3299 bio
->bi_opf
|= MD_FAILFAST
;
3300 bio
->bi_iter
.bi_sector
= sector
+ rdev
->data_offset
;
3301 bio
->bi_bdev
= rdev
->bdev
;
3304 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
3305 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3309 atomic_inc(&rdev
->nr_pending
);
3311 /* Need to set up for writing to the replacement */
3312 bio
= r10_bio
->devs
[i
].repl_bio
;
3313 bio
->bi_status
= BLK_STS_IOERR
;
3315 sector
= r10_bio
->devs
[i
].addr
;
3316 bio
->bi_next
= biolist
;
3318 bio
->bi_end_io
= end_sync_write
;
3319 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3320 if (test_bit(FailFast
, &rdev
->flags
))
3321 bio
->bi_opf
|= MD_FAILFAST
;
3322 bio
->bi_iter
.bi_sector
= sector
+ rdev
->data_offset
;
3323 bio
->bi_bdev
= rdev
->bdev
;
3329 for (i
=0; i
<conf
->copies
; i
++) {
3330 int d
= r10_bio
->devs
[i
].devnum
;
3331 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3332 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3334 if (r10_bio
->devs
[i
].repl_bio
&&
3335 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3337 conf
->mirrors
[d
].replacement
,
3347 if (sector_nr
+ max_sync
< max_sector
)
3348 max_sector
= sector_nr
+ max_sync
;
3351 int len
= PAGE_SIZE
;
3352 if (sector_nr
+ (len
>>9) > max_sector
)
3353 len
= (max_sector
- sector_nr
) << 9;
3356 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3357 struct resync_pages
*rp
= get_resync_pages(bio
);
3358 page
= resync_fetch_page(rp
, rp
->idx
++);
3360 * won't fail because the vec table is big enough
3361 * to hold all these pages
3363 bio_add_page(bio
, page
, len
, 0);
3365 nr_sectors
+= len
>>9;
3366 sector_nr
+= len
>>9;
3367 } while (get_resync_pages(biolist
)->idx
< RESYNC_PAGES
);
3368 r10_bio
->sectors
= nr_sectors
;
3372 biolist
= biolist
->bi_next
;
3374 bio
->bi_next
= NULL
;
3375 r10_bio
= get_resync_r10bio(bio
);
3376 r10_bio
->sectors
= nr_sectors
;
3378 if (bio
->bi_end_io
== end_sync_read
) {
3379 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3381 generic_make_request(bio
);
3385 if (sectors_skipped
)
3386 /* pretend they weren't skipped, it makes
3387 * no important difference in this case
3389 md_done_sync(mddev
, sectors_skipped
, 1);
3391 return sectors_skipped
+ nr_sectors
;
3393 /* There is nowhere to write, so all non-sync
3394 * drives must be failed or in resync, all drives
3395 * have a bad block, so try the next chunk...
3397 if (sector_nr
+ max_sync
< max_sector
)
3398 max_sector
= sector_nr
+ max_sync
;
3400 sectors_skipped
+= (max_sector
- sector_nr
);
3402 sector_nr
= max_sector
;
3407 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3410 struct r10conf
*conf
= mddev
->private;
3413 raid_disks
= min(conf
->geo
.raid_disks
,
3414 conf
->prev
.raid_disks
);
3416 sectors
= conf
->dev_sectors
;
3418 size
= sectors
>> conf
->geo
.chunk_shift
;
3419 sector_div(size
, conf
->geo
.far_copies
);
3420 size
= size
* raid_disks
;
3421 sector_div(size
, conf
->geo
.near_copies
);
3423 return size
<< conf
->geo
.chunk_shift
;
3426 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3428 /* Calculate the number of sectors-per-device that will
3429 * actually be used, and set conf->dev_sectors and
3433 size
= size
>> conf
->geo
.chunk_shift
;
3434 sector_div(size
, conf
->geo
.far_copies
);
3435 size
= size
* conf
->geo
.raid_disks
;
3436 sector_div(size
, conf
->geo
.near_copies
);
3437 /* 'size' is now the number of chunks in the array */
3438 /* calculate "used chunks per device" */
3439 size
= size
* conf
->copies
;
3441 /* We need to round up when dividing by raid_disks to
3442 * get the stride size.
3444 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3446 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3448 if (conf
->geo
.far_offset
)
3449 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3451 sector_div(size
, conf
->geo
.far_copies
);
3452 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3456 enum geo_type
{geo_new
, geo_old
, geo_start
};
3457 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3460 int layout
, chunk
, disks
;
3463 layout
= mddev
->layout
;
3464 chunk
= mddev
->chunk_sectors
;
3465 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3468 layout
= mddev
->new_layout
;
3469 chunk
= mddev
->new_chunk_sectors
;
3470 disks
= mddev
->raid_disks
;
3472 default: /* avoid 'may be unused' warnings */
3473 case geo_start
: /* new when starting reshape - raid_disks not
3475 layout
= mddev
->new_layout
;
3476 chunk
= mddev
->new_chunk_sectors
;
3477 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3482 if (chunk
< (PAGE_SIZE
>> 9) ||
3483 !is_power_of_2(chunk
))
3486 fc
= (layout
>> 8) & 255;
3487 fo
= layout
& (1<<16);
3488 geo
->raid_disks
= disks
;
3489 geo
->near_copies
= nc
;
3490 geo
->far_copies
= fc
;
3491 geo
->far_offset
= fo
;
3492 switch (layout
>> 17) {
3493 case 0: /* original layout. simple but not always optimal */
3494 geo
->far_set_size
= disks
;
3496 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3497 * actually using this, but leave code here just in case.*/
3498 geo
->far_set_size
= disks
/fc
;
3499 WARN(geo
->far_set_size
< fc
,
3500 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3502 case 2: /* "improved" layout fixed to match documentation */
3503 geo
->far_set_size
= fc
* nc
;
3505 default: /* Not a valid layout */
3508 geo
->chunk_mask
= chunk
- 1;
3509 geo
->chunk_shift
= ffz(~chunk
);
3513 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3515 struct r10conf
*conf
= NULL
;
3520 copies
= setup_geo(&geo
, mddev
, geo_new
);
3523 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3524 mdname(mddev
), PAGE_SIZE
);
3528 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3529 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3530 mdname(mddev
), mddev
->new_layout
);
3535 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3539 /* FIXME calc properly */
3540 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3541 max(0,-mddev
->delta_disks
)),
3546 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3551 conf
->copies
= copies
;
3552 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3553 r10bio_pool_free
, conf
);
3554 if (!conf
->r10bio_pool
)
3557 conf
->bio_split
= bioset_create(BIO_POOL_SIZE
, 0, 0);
3558 if (!conf
->bio_split
)
3561 calc_sectors(conf
, mddev
->dev_sectors
);
3562 if (mddev
->reshape_position
== MaxSector
) {
3563 conf
->prev
= conf
->geo
;
3564 conf
->reshape_progress
= MaxSector
;
3566 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3570 conf
->reshape_progress
= mddev
->reshape_position
;
3571 if (conf
->prev
.far_offset
)
3572 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3574 /* far_copies must be 1 */
3575 conf
->prev
.stride
= conf
->dev_sectors
;
3577 conf
->reshape_safe
= conf
->reshape_progress
;
3578 spin_lock_init(&conf
->device_lock
);
3579 INIT_LIST_HEAD(&conf
->retry_list
);
3580 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
3582 spin_lock_init(&conf
->resync_lock
);
3583 init_waitqueue_head(&conf
->wait_barrier
);
3584 atomic_set(&conf
->nr_pending
, 0);
3586 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3590 conf
->mddev
= mddev
;
3595 mempool_destroy(conf
->r10bio_pool
);
3596 kfree(conf
->mirrors
);
3597 safe_put_page(conf
->tmppage
);
3598 if (conf
->bio_split
)
3599 bioset_free(conf
->bio_split
);
3602 return ERR_PTR(err
);
3605 static int raid10_run(struct mddev
*mddev
)
3607 struct r10conf
*conf
;
3608 int i
, disk_idx
, chunk_size
;
3609 struct raid10_info
*disk
;
3610 struct md_rdev
*rdev
;
3612 sector_t min_offset_diff
= 0;
3614 bool discard_supported
= false;
3616 if (mddev_init_writes_pending(mddev
) < 0)
3619 if (mddev
->private == NULL
) {
3620 conf
= setup_conf(mddev
);
3622 return PTR_ERR(conf
);
3623 mddev
->private = conf
;
3625 conf
= mddev
->private;
3629 mddev
->thread
= conf
->thread
;
3630 conf
->thread
= NULL
;
3632 chunk_size
= mddev
->chunk_sectors
<< 9;
3634 blk_queue_max_discard_sectors(mddev
->queue
,
3635 mddev
->chunk_sectors
);
3636 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3637 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
3638 blk_queue_io_min(mddev
->queue
, chunk_size
);
3639 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3640 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3642 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3643 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3646 rdev_for_each(rdev
, mddev
) {
3649 disk_idx
= rdev
->raid_disk
;
3652 if (disk_idx
>= conf
->geo
.raid_disks
&&
3653 disk_idx
>= conf
->prev
.raid_disks
)
3655 disk
= conf
->mirrors
+ disk_idx
;
3657 if (test_bit(Replacement
, &rdev
->flags
)) {
3658 if (disk
->replacement
)
3660 disk
->replacement
= rdev
;
3666 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3667 if (!mddev
->reshape_backwards
)
3671 if (first
|| diff
< min_offset_diff
)
3672 min_offset_diff
= diff
;
3675 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3676 rdev
->data_offset
<< 9);
3678 disk
->head_position
= 0;
3680 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3681 discard_supported
= true;
3686 if (discard_supported
)
3687 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3690 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3693 /* need to check that every block has at least one working mirror */
3694 if (!enough(conf
, -1)) {
3695 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3700 if (conf
->reshape_progress
!= MaxSector
) {
3701 /* must ensure that shape change is supported */
3702 if (conf
->geo
.far_copies
!= 1 &&
3703 conf
->geo
.far_offset
== 0)
3705 if (conf
->prev
.far_copies
!= 1 &&
3706 conf
->prev
.far_offset
== 0)
3710 mddev
->degraded
= 0;
3712 i
< conf
->geo
.raid_disks
3713 || i
< conf
->prev
.raid_disks
;
3716 disk
= conf
->mirrors
+ i
;
3718 if (!disk
->rdev
&& disk
->replacement
) {
3719 /* The replacement is all we have - use it */
3720 disk
->rdev
= disk
->replacement
;
3721 disk
->replacement
= NULL
;
3722 clear_bit(Replacement
, &disk
->rdev
->flags
);
3726 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3727 disk
->head_position
= 0;
3730 disk
->rdev
->saved_raid_disk
< 0)
3733 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3736 if (mddev
->recovery_cp
!= MaxSector
)
3737 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3739 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3740 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3741 conf
->geo
.raid_disks
);
3743 * Ok, everything is just fine now
3745 mddev
->dev_sectors
= conf
->dev_sectors
;
3746 size
= raid10_size(mddev
, 0, 0);
3747 md_set_array_sectors(mddev
, size
);
3748 mddev
->resync_max_sectors
= size
;
3749 set_bit(MD_FAILFAST_SUPPORTED
, &mddev
->flags
);
3752 int stripe
= conf
->geo
.raid_disks
*
3753 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3755 /* Calculate max read-ahead size.
3756 * We need to readahead at least twice a whole stripe....
3759 stripe
/= conf
->geo
.near_copies
;
3760 if (mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
3761 mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
3764 if (md_integrity_register(mddev
))
3767 if (conf
->reshape_progress
!= MaxSector
) {
3768 unsigned long before_length
, after_length
;
3770 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3771 conf
->prev
.far_copies
);
3772 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3773 conf
->geo
.far_copies
);
3775 if (max(before_length
, after_length
) > min_offset_diff
) {
3776 /* This cannot work */
3777 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3780 conf
->offset_diff
= min_offset_diff
;
3782 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3783 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3784 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3785 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3786 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3793 md_unregister_thread(&mddev
->thread
);
3794 mempool_destroy(conf
->r10bio_pool
);
3795 safe_put_page(conf
->tmppage
);
3796 kfree(conf
->mirrors
);
3798 mddev
->private = NULL
;
3803 static void raid10_free(struct mddev
*mddev
, void *priv
)
3805 struct r10conf
*conf
= priv
;
3807 mempool_destroy(conf
->r10bio_pool
);
3808 safe_put_page(conf
->tmppage
);
3809 kfree(conf
->mirrors
);
3810 kfree(conf
->mirrors_old
);
3811 kfree(conf
->mirrors_new
);
3812 if (conf
->bio_split
)
3813 bioset_free(conf
->bio_split
);
3817 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3819 struct r10conf
*conf
= mddev
->private;
3823 raise_barrier(conf
, 0);
3826 lower_barrier(conf
);
3831 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3833 /* Resize of 'far' arrays is not supported.
3834 * For 'near' and 'offset' arrays we can set the
3835 * number of sectors used to be an appropriate multiple
3836 * of the chunk size.
3837 * For 'offset', this is far_copies*chunksize.
3838 * For 'near' the multiplier is the LCM of
3839 * near_copies and raid_disks.
3840 * So if far_copies > 1 && !far_offset, fail.
3841 * Else find LCM(raid_disks, near_copy)*far_copies and
3842 * multiply by chunk_size. Then round to this number.
3843 * This is mostly done by raid10_size()
3845 struct r10conf
*conf
= mddev
->private;
3846 sector_t oldsize
, size
;
3848 if (mddev
->reshape_position
!= MaxSector
)
3851 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3854 oldsize
= raid10_size(mddev
, 0, 0);
3855 size
= raid10_size(mddev
, sectors
, 0);
3856 if (mddev
->external_size
&&
3857 mddev
->array_sectors
> size
)
3859 if (mddev
->bitmap
) {
3860 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3864 md_set_array_sectors(mddev
, size
);
3865 if (sectors
> mddev
->dev_sectors
&&
3866 mddev
->recovery_cp
> oldsize
) {
3867 mddev
->recovery_cp
= oldsize
;
3868 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3870 calc_sectors(conf
, sectors
);
3871 mddev
->dev_sectors
= conf
->dev_sectors
;
3872 mddev
->resync_max_sectors
= size
;
3876 static void *raid10_takeover_raid0(struct mddev
*mddev
, sector_t size
, int devs
)
3878 struct md_rdev
*rdev
;
3879 struct r10conf
*conf
;
3881 if (mddev
->degraded
> 0) {
3882 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
3884 return ERR_PTR(-EINVAL
);
3886 sector_div(size
, devs
);
3888 /* Set new parameters */
3889 mddev
->new_level
= 10;
3890 /* new layout: far_copies = 1, near_copies = 2 */
3891 mddev
->new_layout
= (1<<8) + 2;
3892 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3893 mddev
->delta_disks
= mddev
->raid_disks
;
3894 mddev
->raid_disks
*= 2;
3895 /* make sure it will be not marked as dirty */
3896 mddev
->recovery_cp
= MaxSector
;
3897 mddev
->dev_sectors
= size
;
3899 conf
= setup_conf(mddev
);
3900 if (!IS_ERR(conf
)) {
3901 rdev_for_each(rdev
, mddev
)
3902 if (rdev
->raid_disk
>= 0) {
3903 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3904 rdev
->sectors
= size
;
3912 static void *raid10_takeover(struct mddev
*mddev
)
3914 struct r0conf
*raid0_conf
;
3916 /* raid10 can take over:
3917 * raid0 - providing it has only two drives
3919 if (mddev
->level
== 0) {
3920 /* for raid0 takeover only one zone is supported */
3921 raid0_conf
= mddev
->private;
3922 if (raid0_conf
->nr_strip_zones
> 1) {
3923 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
3925 return ERR_PTR(-EINVAL
);
3927 return raid10_takeover_raid0(mddev
,
3928 raid0_conf
->strip_zone
->zone_end
,
3929 raid0_conf
->strip_zone
->nb_dev
);
3931 return ERR_PTR(-EINVAL
);
3934 static int raid10_check_reshape(struct mddev
*mddev
)
3936 /* Called when there is a request to change
3937 * - layout (to ->new_layout)
3938 * - chunk size (to ->new_chunk_sectors)
3939 * - raid_disks (by delta_disks)
3940 * or when trying to restart a reshape that was ongoing.
3942 * We need to validate the request and possibly allocate
3943 * space if that might be an issue later.
3945 * Currently we reject any reshape of a 'far' mode array,
3946 * allow chunk size to change if new is generally acceptable,
3947 * allow raid_disks to increase, and allow
3948 * a switch between 'near' mode and 'offset' mode.
3950 struct r10conf
*conf
= mddev
->private;
3953 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3956 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3957 /* mustn't change number of copies */
3959 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3960 /* Cannot switch to 'far' mode */
3963 if (mddev
->array_sectors
& geo
.chunk_mask
)
3964 /* not factor of array size */
3967 if (!enough(conf
, -1))
3970 kfree(conf
->mirrors_new
);
3971 conf
->mirrors_new
= NULL
;
3972 if (mddev
->delta_disks
> 0) {
3973 /* allocate new 'mirrors' list */
3974 conf
->mirrors_new
= kzalloc(
3975 sizeof(struct raid10_info
)
3976 *(mddev
->raid_disks
+
3977 mddev
->delta_disks
),
3979 if (!conf
->mirrors_new
)
3986 * Need to check if array has failed when deciding whether to:
3988 * - remove non-faulty devices
3991 * This determination is simple when no reshape is happening.
3992 * However if there is a reshape, we need to carefully check
3993 * both the before and after sections.
3994 * This is because some failed devices may only affect one
3995 * of the two sections, and some non-in_sync devices may
3996 * be insync in the section most affected by failed devices.
3998 static int calc_degraded(struct r10conf
*conf
)
4000 int degraded
, degraded2
;
4005 /* 'prev' section first */
4006 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
4007 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4008 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4010 else if (!test_bit(In_sync
, &rdev
->flags
))
4011 /* When we can reduce the number of devices in
4012 * an array, this might not contribute to
4013 * 'degraded'. It does now.
4018 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
4022 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
4023 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
4024 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4026 else if (!test_bit(In_sync
, &rdev
->flags
)) {
4027 /* If reshape is increasing the number of devices,
4028 * this section has already been recovered, so
4029 * it doesn't contribute to degraded.
4032 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4037 if (degraded2
> degraded
)
4042 static int raid10_start_reshape(struct mddev
*mddev
)
4044 /* A 'reshape' has been requested. This commits
4045 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4046 * This also checks if there are enough spares and adds them
4048 * We currently require enough spares to make the final
4049 * array non-degraded. We also require that the difference
4050 * between old and new data_offset - on each device - is
4051 * enough that we never risk over-writing.
4054 unsigned long before_length
, after_length
;
4055 sector_t min_offset_diff
= 0;
4058 struct r10conf
*conf
= mddev
->private;
4059 struct md_rdev
*rdev
;
4063 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4066 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4069 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4070 conf
->prev
.far_copies
);
4071 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4072 conf
->geo
.far_copies
);
4074 rdev_for_each(rdev
, mddev
) {
4075 if (!test_bit(In_sync
, &rdev
->flags
)
4076 && !test_bit(Faulty
, &rdev
->flags
))
4078 if (rdev
->raid_disk
>= 0) {
4079 long long diff
= (rdev
->new_data_offset
4080 - rdev
->data_offset
);
4081 if (!mddev
->reshape_backwards
)
4085 if (first
|| diff
< min_offset_diff
)
4086 min_offset_diff
= diff
;
4091 if (max(before_length
, after_length
) > min_offset_diff
)
4094 if (spares
< mddev
->delta_disks
)
4097 conf
->offset_diff
= min_offset_diff
;
4098 spin_lock_irq(&conf
->device_lock
);
4099 if (conf
->mirrors_new
) {
4100 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4101 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4103 kfree(conf
->mirrors_old
);
4104 conf
->mirrors_old
= conf
->mirrors
;
4105 conf
->mirrors
= conf
->mirrors_new
;
4106 conf
->mirrors_new
= NULL
;
4108 setup_geo(&conf
->geo
, mddev
, geo_start
);
4110 if (mddev
->reshape_backwards
) {
4111 sector_t size
= raid10_size(mddev
, 0, 0);
4112 if (size
< mddev
->array_sectors
) {
4113 spin_unlock_irq(&conf
->device_lock
);
4114 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4118 mddev
->resync_max_sectors
= size
;
4119 conf
->reshape_progress
= size
;
4121 conf
->reshape_progress
= 0;
4122 conf
->reshape_safe
= conf
->reshape_progress
;
4123 spin_unlock_irq(&conf
->device_lock
);
4125 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4126 ret
= bitmap_resize(mddev
->bitmap
,
4127 raid10_size(mddev
, 0,
4128 conf
->geo
.raid_disks
),
4133 if (mddev
->delta_disks
> 0) {
4134 rdev_for_each(rdev
, mddev
)
4135 if (rdev
->raid_disk
< 0 &&
4136 !test_bit(Faulty
, &rdev
->flags
)) {
4137 if (raid10_add_disk(mddev
, rdev
) == 0) {
4138 if (rdev
->raid_disk
>=
4139 conf
->prev
.raid_disks
)
4140 set_bit(In_sync
, &rdev
->flags
);
4142 rdev
->recovery_offset
= 0;
4144 if (sysfs_link_rdev(mddev
, rdev
))
4145 /* Failure here is OK */;
4147 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4148 && !test_bit(Faulty
, &rdev
->flags
)) {
4149 /* This is a spare that was manually added */
4150 set_bit(In_sync
, &rdev
->flags
);
4153 /* When a reshape changes the number of devices,
4154 * ->degraded is measured against the larger of the
4155 * pre and post numbers.
4157 spin_lock_irq(&conf
->device_lock
);
4158 mddev
->degraded
= calc_degraded(conf
);
4159 spin_unlock_irq(&conf
->device_lock
);
4160 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4161 mddev
->reshape_position
= conf
->reshape_progress
;
4162 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
4164 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4165 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4166 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
4167 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4168 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4170 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4172 if (!mddev
->sync_thread
) {
4176 conf
->reshape_checkpoint
= jiffies
;
4177 md_wakeup_thread(mddev
->sync_thread
);
4178 md_new_event(mddev
);
4182 mddev
->recovery
= 0;
4183 spin_lock_irq(&conf
->device_lock
);
4184 conf
->geo
= conf
->prev
;
4185 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4186 rdev_for_each(rdev
, mddev
)
4187 rdev
->new_data_offset
= rdev
->data_offset
;
4189 conf
->reshape_progress
= MaxSector
;
4190 conf
->reshape_safe
= MaxSector
;
4191 mddev
->reshape_position
= MaxSector
;
4192 spin_unlock_irq(&conf
->device_lock
);
4196 /* Calculate the last device-address that could contain
4197 * any block from the chunk that includes the array-address 's'
4198 * and report the next address.
4199 * i.e. the address returned will be chunk-aligned and after
4200 * any data that is in the chunk containing 's'.
4202 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4204 s
= (s
| geo
->chunk_mask
) + 1;
4205 s
>>= geo
->chunk_shift
;
4206 s
*= geo
->near_copies
;
4207 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4208 s
*= geo
->far_copies
;
4209 s
<<= geo
->chunk_shift
;
4213 /* Calculate the first device-address that could contain
4214 * any block from the chunk that includes the array-address 's'.
4215 * This too will be the start of a chunk
4217 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4219 s
>>= geo
->chunk_shift
;
4220 s
*= geo
->near_copies
;
4221 sector_div(s
, geo
->raid_disks
);
4222 s
*= geo
->far_copies
;
4223 s
<<= geo
->chunk_shift
;
4227 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4230 /* We simply copy at most one chunk (smallest of old and new)
4231 * at a time, possibly less if that exceeds RESYNC_PAGES,
4232 * or we hit a bad block or something.
4233 * This might mean we pause for normal IO in the middle of
4234 * a chunk, but that is not a problem as mddev->reshape_position
4235 * can record any location.
4237 * If we will want to write to a location that isn't
4238 * yet recorded as 'safe' (i.e. in metadata on disk) then
4239 * we need to flush all reshape requests and update the metadata.
4241 * When reshaping forwards (e.g. to more devices), we interpret
4242 * 'safe' as the earliest block which might not have been copied
4243 * down yet. We divide this by previous stripe size and multiply
4244 * by previous stripe length to get lowest device offset that we
4245 * cannot write to yet.
4246 * We interpret 'sector_nr' as an address that we want to write to.
4247 * From this we use last_device_address() to find where we might
4248 * write to, and first_device_address on the 'safe' position.
4249 * If this 'next' write position is after the 'safe' position,
4250 * we must update the metadata to increase the 'safe' position.
4252 * When reshaping backwards, we round in the opposite direction
4253 * and perform the reverse test: next write position must not be
4254 * less than current safe position.
4256 * In all this the minimum difference in data offsets
4257 * (conf->offset_diff - always positive) allows a bit of slack,
4258 * so next can be after 'safe', but not by more than offset_diff
4260 * We need to prepare all the bios here before we start any IO
4261 * to ensure the size we choose is acceptable to all devices.
4262 * The means one for each copy for write-out and an extra one for
4264 * We store the read-in bio in ->master_bio and the others in
4265 * ->devs[x].bio and ->devs[x].repl_bio.
4267 struct r10conf
*conf
= mddev
->private;
4268 struct r10bio
*r10_bio
;
4269 sector_t next
, safe
, last
;
4273 struct md_rdev
*rdev
;
4276 struct bio
*bio
, *read_bio
;
4277 int sectors_done
= 0;
4278 struct page
**pages
;
4280 if (sector_nr
== 0) {
4281 /* If restarting in the middle, skip the initial sectors */
4282 if (mddev
->reshape_backwards
&&
4283 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4284 sector_nr
= (raid10_size(mddev
, 0, 0)
4285 - conf
->reshape_progress
);
4286 } else if (!mddev
->reshape_backwards
&&
4287 conf
->reshape_progress
> 0)
4288 sector_nr
= conf
->reshape_progress
;
4290 mddev
->curr_resync_completed
= sector_nr
;
4291 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4297 /* We don't use sector_nr to track where we are up to
4298 * as that doesn't work well for ->reshape_backwards.
4299 * So just use ->reshape_progress.
4301 if (mddev
->reshape_backwards
) {
4302 /* 'next' is the earliest device address that we might
4303 * write to for this chunk in the new layout
4305 next
= first_dev_address(conf
->reshape_progress
- 1,
4308 /* 'safe' is the last device address that we might read from
4309 * in the old layout after a restart
4311 safe
= last_dev_address(conf
->reshape_safe
- 1,
4314 if (next
+ conf
->offset_diff
< safe
)
4317 last
= conf
->reshape_progress
- 1;
4318 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4319 & conf
->prev
.chunk_mask
);
4320 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4321 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4323 /* 'next' is after the last device address that we
4324 * might write to for this chunk in the new layout
4326 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4328 /* 'safe' is the earliest device address that we might
4329 * read from in the old layout after a restart
4331 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4333 /* Need to update metadata if 'next' might be beyond 'safe'
4334 * as that would possibly corrupt data
4336 if (next
> safe
+ conf
->offset_diff
)
4339 sector_nr
= conf
->reshape_progress
;
4340 last
= sector_nr
| (conf
->geo
.chunk_mask
4341 & conf
->prev
.chunk_mask
);
4343 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4344 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4348 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4349 /* Need to update reshape_position in metadata */
4351 mddev
->reshape_position
= conf
->reshape_progress
;
4352 if (mddev
->reshape_backwards
)
4353 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4354 - conf
->reshape_progress
;
4356 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4357 conf
->reshape_checkpoint
= jiffies
;
4358 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
4359 md_wakeup_thread(mddev
->thread
);
4360 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
4361 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4362 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4363 allow_barrier(conf
);
4364 return sectors_done
;
4366 conf
->reshape_safe
= mddev
->reshape_position
;
4367 allow_barrier(conf
);
4371 /* Now schedule reads for blocks from sector_nr to last */
4372 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4374 raise_barrier(conf
, sectors_done
!= 0);
4375 atomic_set(&r10_bio
->remaining
, 0);
4376 r10_bio
->mddev
= mddev
;
4377 r10_bio
->sector
= sector_nr
;
4378 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4379 r10_bio
->sectors
= last
- sector_nr
+ 1;
4380 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4381 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4384 /* Cannot read from here, so need to record bad blocks
4385 * on all the target devices.
4388 mempool_free(r10_bio
, conf
->r10buf_pool
);
4389 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4390 return sectors_done
;
4393 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4395 read_bio
->bi_bdev
= rdev
->bdev
;
4396 read_bio
->bi_iter
.bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4397 + rdev
->data_offset
);
4398 read_bio
->bi_private
= r10_bio
;
4399 read_bio
->bi_end_io
= end_reshape_read
;
4400 bio_set_op_attrs(read_bio
, REQ_OP_READ
, 0);
4401 read_bio
->bi_flags
&= (~0UL << BIO_RESET_BITS
);
4402 read_bio
->bi_status
= 0;
4403 read_bio
->bi_vcnt
= 0;
4404 read_bio
->bi_iter
.bi_size
= 0;
4405 r10_bio
->master_bio
= read_bio
;
4406 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4408 /* Now find the locations in the new layout */
4409 __raid10_find_phys(&conf
->geo
, r10_bio
);
4412 read_bio
->bi_next
= NULL
;
4415 for (s
= 0; s
< conf
->copies
*2; s
++) {
4417 int d
= r10_bio
->devs
[s
/2].devnum
;
4418 struct md_rdev
*rdev2
;
4420 rdev2
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4421 b
= r10_bio
->devs
[s
/2].repl_bio
;
4423 rdev2
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4424 b
= r10_bio
->devs
[s
/2].bio
;
4426 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4429 b
->bi_bdev
= rdev2
->bdev
;
4430 b
->bi_iter
.bi_sector
= r10_bio
->devs
[s
/2].addr
+
4431 rdev2
->new_data_offset
;
4432 b
->bi_end_io
= end_reshape_write
;
4433 bio_set_op_attrs(b
, REQ_OP_WRITE
, 0);
4438 /* Now add as many pages as possible to all of these bios. */
4441 pages
= get_resync_pages(r10_bio
->devs
[0].bio
)->pages
;
4442 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4443 struct page
*page
= pages
[s
/ (PAGE_SIZE
>> 9)];
4444 int len
= (max_sectors
- s
) << 9;
4445 if (len
> PAGE_SIZE
)
4447 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4449 * won't fail because the vec table is big enough
4450 * to hold all these pages
4452 bio_add_page(bio
, page
, len
, 0);
4454 sector_nr
+= len
>> 9;
4455 nr_sectors
+= len
>> 9;
4458 r10_bio
->sectors
= nr_sectors
;
4460 /* Now submit the read */
4461 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4462 atomic_inc(&r10_bio
->remaining
);
4463 read_bio
->bi_next
= NULL
;
4464 generic_make_request(read_bio
);
4465 sector_nr
+= nr_sectors
;
4466 sectors_done
+= nr_sectors
;
4467 if (sector_nr
<= last
)
4470 /* Now that we have done the whole section we can
4471 * update reshape_progress
4473 if (mddev
->reshape_backwards
)
4474 conf
->reshape_progress
-= sectors_done
;
4476 conf
->reshape_progress
+= sectors_done
;
4478 return sectors_done
;
4481 static void end_reshape_request(struct r10bio
*r10_bio
);
4482 static int handle_reshape_read_error(struct mddev
*mddev
,
4483 struct r10bio
*r10_bio
);
4484 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4486 /* Reshape read completed. Hopefully we have a block
4488 * If we got a read error then we do sync 1-page reads from
4489 * elsewhere until we find the data - or give up.
4491 struct r10conf
*conf
= mddev
->private;
4494 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4495 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4496 /* Reshape has been aborted */
4497 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4501 /* We definitely have the data in the pages, schedule the
4504 atomic_set(&r10_bio
->remaining
, 1);
4505 for (s
= 0; s
< conf
->copies
*2; s
++) {
4507 int d
= r10_bio
->devs
[s
/2].devnum
;
4508 struct md_rdev
*rdev
;
4511 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4512 b
= r10_bio
->devs
[s
/2].repl_bio
;
4514 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4515 b
= r10_bio
->devs
[s
/2].bio
;
4517 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
4521 atomic_inc(&rdev
->nr_pending
);
4523 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4524 atomic_inc(&r10_bio
->remaining
);
4526 generic_make_request(b
);
4528 end_reshape_request(r10_bio
);
4531 static void end_reshape(struct r10conf
*conf
)
4533 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4536 spin_lock_irq(&conf
->device_lock
);
4537 conf
->prev
= conf
->geo
;
4538 md_finish_reshape(conf
->mddev
);
4540 conf
->reshape_progress
= MaxSector
;
4541 conf
->reshape_safe
= MaxSector
;
4542 spin_unlock_irq(&conf
->device_lock
);
4544 /* read-ahead size must cover two whole stripes, which is
4545 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4547 if (conf
->mddev
->queue
) {
4548 int stripe
= conf
->geo
.raid_disks
*
4549 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4550 stripe
/= conf
->geo
.near_copies
;
4551 if (conf
->mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
4552 conf
->mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
4557 static int handle_reshape_read_error(struct mddev
*mddev
,
4558 struct r10bio
*r10_bio
)
4560 /* Use sync reads to get the blocks from somewhere else */
4561 int sectors
= r10_bio
->sectors
;
4562 struct r10conf
*conf
= mddev
->private;
4564 struct r10bio r10_bio
;
4565 struct r10dev devs
[conf
->copies
];
4567 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4570 struct page
**pages
;
4572 /* reshape IOs share pages from .devs[0].bio */
4573 pages
= get_resync_pages(r10_bio
->devs
[0].bio
)->pages
;
4575 r10b
->sector
= r10_bio
->sector
;
4576 __raid10_find_phys(&conf
->prev
, r10b
);
4581 int first_slot
= slot
;
4583 if (s
> (PAGE_SIZE
>> 9))
4588 int d
= r10b
->devs
[slot
].devnum
;
4589 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4592 test_bit(Faulty
, &rdev
->flags
) ||
4593 !test_bit(In_sync
, &rdev
->flags
))
4596 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4597 atomic_inc(&rdev
->nr_pending
);
4599 success
= sync_page_io(rdev
,
4603 REQ_OP_READ
, 0, false);
4604 rdev_dec_pending(rdev
, mddev
);
4610 if (slot
>= conf
->copies
)
4612 if (slot
== first_slot
)
4617 /* couldn't read this block, must give up */
4618 set_bit(MD_RECOVERY_INTR
,
4628 static void end_reshape_write(struct bio
*bio
)
4630 struct r10bio
*r10_bio
= get_resync_r10bio(bio
);
4631 struct mddev
*mddev
= r10_bio
->mddev
;
4632 struct r10conf
*conf
= mddev
->private;
4636 struct md_rdev
*rdev
= NULL
;
4638 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4640 rdev
= conf
->mirrors
[d
].replacement
;
4643 rdev
= conf
->mirrors
[d
].rdev
;
4646 if (bio
->bi_status
) {
4647 /* FIXME should record badblock */
4648 md_error(mddev
, rdev
);
4651 rdev_dec_pending(rdev
, mddev
);
4652 end_reshape_request(r10_bio
);
4655 static void end_reshape_request(struct r10bio
*r10_bio
)
4657 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4659 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4660 bio_put(r10_bio
->master_bio
);
4664 static void raid10_finish_reshape(struct mddev
*mddev
)
4666 struct r10conf
*conf
= mddev
->private;
4668 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4671 if (mddev
->delta_disks
> 0) {
4672 sector_t size
= raid10_size(mddev
, 0, 0);
4673 md_set_array_sectors(mddev
, size
);
4674 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4675 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4676 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4678 mddev
->resync_max_sectors
= size
;
4680 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4681 revalidate_disk(mddev
->gendisk
);
4686 for (d
= conf
->geo
.raid_disks
;
4687 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4689 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4691 clear_bit(In_sync
, &rdev
->flags
);
4692 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4694 clear_bit(In_sync
, &rdev
->flags
);
4698 mddev
->layout
= mddev
->new_layout
;
4699 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4700 mddev
->reshape_position
= MaxSector
;
4701 mddev
->delta_disks
= 0;
4702 mddev
->reshape_backwards
= 0;
4705 static struct md_personality raid10_personality
=
4709 .owner
= THIS_MODULE
,
4710 .make_request
= raid10_make_request
,
4712 .free
= raid10_free
,
4713 .status
= raid10_status
,
4714 .error_handler
= raid10_error
,
4715 .hot_add_disk
= raid10_add_disk
,
4716 .hot_remove_disk
= raid10_remove_disk
,
4717 .spare_active
= raid10_spare_active
,
4718 .sync_request
= raid10_sync_request
,
4719 .quiesce
= raid10_quiesce
,
4720 .size
= raid10_size
,
4721 .resize
= raid10_resize
,
4722 .takeover
= raid10_takeover
,
4723 .check_reshape
= raid10_check_reshape
,
4724 .start_reshape
= raid10_start_reshape
,
4725 .finish_reshape
= raid10_finish_reshape
,
4726 .congested
= raid10_congested
,
4729 static int __init
raid_init(void)
4731 return register_md_personality(&raid10_personality
);
4734 static void raid_exit(void)
4736 unregister_md_personality(&raid10_personality
);
4739 module_init(raid_init
);
4740 module_exit(raid_exit
);
4741 MODULE_LICENSE("GPL");
4742 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4743 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4744 MODULE_ALIAS("md-raid10");
4745 MODULE_ALIAS("md-level-10");
4747 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);