2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 * use_far_sets (stored in bit 17 of layout )
42 * use_far_sets_bugfixed (stored in bit 18 of layout )
44 * The data to be stored is divided into chunks using chunksize. Each device
45 * is divided into far_copies sections. In each section, chunks are laid out
46 * in a style similar to raid0, but near_copies copies of each chunk is stored
47 * (each on a different drive). The starting device for each section is offset
48 * near_copies from the starting device of the previous section. Thus there
49 * are (near_copies * far_copies) of each chunk, and each is on a different
50 * drive. near_copies and far_copies must be at least one, and their product
51 * is at most raid_disks.
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of being very far
55 * apart on disk, there are adjacent stripes.
57 * The far and offset algorithms are handled slightly differently if
58 * 'use_far_sets' is true. In this case, the array's devices are grouped into
59 * sets that are (near_copies * far_copies) in size. The far copied stripes
60 * are still shifted by 'near_copies' devices, but this shifting stays confined
61 * to the set rather than the entire array. This is done to improve the number
62 * of device combinations that can fail without causing the array to fail.
63 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
68 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
69 * [A B] [C D] [A B] [C D E]
70 * |...| |...| |...| | ... |
71 * [B A] [D C] [B A] [E C D]
75 * Number of guaranteed r10bios in case of extreme VM load:
77 #define NR_RAID10_BIOS 256
79 /* when we get a read error on a read-only array, we redirect to another
80 * device without failing the first device, or trying to over-write to
81 * correct the read error. To keep track of bad blocks on a per-bio
82 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
84 #define IO_BLOCKED ((struct bio *)1)
85 /* When we successfully write to a known bad-block, we need to remove the
86 * bad-block marking which must be done from process context. So we record
87 * the success by setting devs[n].bio to IO_MADE_GOOD
89 #define IO_MADE_GOOD ((struct bio *)2)
91 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
93 /* When there are this many requests queued to be written by
94 * the raid10 thread, we become 'congested' to provide back-pressure
97 static int max_queued_requests
= 1024;
99 static void allow_barrier(struct r10conf
*conf
);
100 static void lower_barrier(struct r10conf
*conf
);
101 static int _enough(struct r10conf
*conf
, int previous
, int ignore
);
102 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
104 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
);
105 static void end_reshape_write(struct bio
*bio
);
106 static void end_reshape(struct r10conf
*conf
);
108 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
110 struct r10conf
*conf
= data
;
111 int size
= offsetof(struct r10bio
, devs
[conf
->copies
]);
113 /* allocate a r10bio with room for raid_disks entries in the
115 return kzalloc(size
, gfp_flags
);
118 static void r10bio_pool_free(void *r10_bio
, void *data
)
123 /* Maximum size of each resync request */
124 #define RESYNC_BLOCK_SIZE (64*1024)
125 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
126 /* amount of memory to reserve for resync requests */
127 #define RESYNC_WINDOW (1024*1024)
128 /* maximum number of concurrent requests, memory permitting */
129 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
132 * When performing a resync, we need to read and compare, so
133 * we need as many pages are there are copies.
134 * When performing a recovery, we need 2 bios, one for read,
135 * one for write (we recover only one drive per r10buf)
138 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
140 struct r10conf
*conf
= data
;
142 struct r10bio
*r10_bio
;
147 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
151 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
) ||
152 test_bit(MD_RECOVERY_RESHAPE
, &conf
->mddev
->recovery
))
153 nalloc
= conf
->copies
; /* resync */
155 nalloc
= 2; /* recovery */
160 for (j
= nalloc
; j
-- ; ) {
161 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
164 r10_bio
->devs
[j
].bio
= bio
;
165 if (!conf
->have_replacement
)
167 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
170 r10_bio
->devs
[j
].repl_bio
= bio
;
173 * Allocate RESYNC_PAGES data pages and attach them
176 for (j
= 0 ; j
< nalloc
; j
++) {
177 struct bio
*rbio
= r10_bio
->devs
[j
].repl_bio
;
178 bio
= r10_bio
->devs
[j
].bio
;
179 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
180 if (j
> 0 && !test_bit(MD_RECOVERY_SYNC
,
181 &conf
->mddev
->recovery
)) {
182 /* we can share bv_page's during recovery
184 struct bio
*rbio
= r10_bio
->devs
[0].bio
;
185 page
= rbio
->bi_io_vec
[i
].bv_page
;
188 page
= alloc_page(gfp_flags
);
192 bio
->bi_io_vec
[i
].bv_page
= page
;
194 rbio
->bi_io_vec
[i
].bv_page
= page
;
202 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
204 for (i
= 0; i
< RESYNC_PAGES
; i
++)
205 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
208 for ( ; j
< nalloc
; j
++) {
209 if (r10_bio
->devs
[j
].bio
)
210 bio_put(r10_bio
->devs
[j
].bio
);
211 if (r10_bio
->devs
[j
].repl_bio
)
212 bio_put(r10_bio
->devs
[j
].repl_bio
);
214 r10bio_pool_free(r10_bio
, conf
);
218 static void r10buf_pool_free(void *__r10_bio
, void *data
)
221 struct r10conf
*conf
= data
;
222 struct r10bio
*r10bio
= __r10_bio
;
225 for (j
=0; j
< conf
->copies
; j
++) {
226 struct bio
*bio
= r10bio
->devs
[j
].bio
;
228 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
229 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
230 bio
->bi_io_vec
[i
].bv_page
= NULL
;
234 bio
= r10bio
->devs
[j
].repl_bio
;
238 r10bio_pool_free(r10bio
, conf
);
241 static void put_all_bios(struct r10conf
*conf
, struct r10bio
*r10_bio
)
245 for (i
= 0; i
< conf
->copies
; i
++) {
246 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
247 if (!BIO_SPECIAL(*bio
))
250 bio
= &r10_bio
->devs
[i
].repl_bio
;
251 if (r10_bio
->read_slot
< 0 && !BIO_SPECIAL(*bio
))
257 static void free_r10bio(struct r10bio
*r10_bio
)
259 struct r10conf
*conf
= r10_bio
->mddev
->private;
261 put_all_bios(conf
, r10_bio
);
262 mempool_free(r10_bio
, conf
->r10bio_pool
);
265 static void put_buf(struct r10bio
*r10_bio
)
267 struct r10conf
*conf
= r10_bio
->mddev
->private;
269 mempool_free(r10_bio
, conf
->r10buf_pool
);
274 static void reschedule_retry(struct r10bio
*r10_bio
)
277 struct mddev
*mddev
= r10_bio
->mddev
;
278 struct r10conf
*conf
= mddev
->private;
280 spin_lock_irqsave(&conf
->device_lock
, flags
);
281 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
283 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
285 /* wake up frozen array... */
286 wake_up(&conf
->wait_barrier
);
288 md_wakeup_thread(mddev
->thread
);
292 * raid_end_bio_io() is called when we have finished servicing a mirrored
293 * operation and are ready to return a success/failure code to the buffer
296 static void raid_end_bio_io(struct r10bio
*r10_bio
)
298 struct bio
*bio
= r10_bio
->master_bio
;
300 struct r10conf
*conf
= r10_bio
->mddev
->private;
302 if (bio
->bi_phys_segments
) {
304 spin_lock_irqsave(&conf
->device_lock
, flags
);
305 bio
->bi_phys_segments
--;
306 done
= (bio
->bi_phys_segments
== 0);
307 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
310 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
311 bio
->bi_error
= -EIO
;
315 * Wake up any possible resync thread that waits for the device
320 free_r10bio(r10_bio
);
324 * Update disk head position estimator based on IRQ completion info.
326 static inline void update_head_pos(int slot
, struct r10bio
*r10_bio
)
328 struct r10conf
*conf
= r10_bio
->mddev
->private;
330 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
331 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
335 * Find the disk number which triggered given bio
337 static int find_bio_disk(struct r10conf
*conf
, struct r10bio
*r10_bio
,
338 struct bio
*bio
, int *slotp
, int *replp
)
343 for (slot
= 0; slot
< conf
->copies
; slot
++) {
344 if (r10_bio
->devs
[slot
].bio
== bio
)
346 if (r10_bio
->devs
[slot
].repl_bio
== bio
) {
352 BUG_ON(slot
== conf
->copies
);
353 update_head_pos(slot
, r10_bio
);
359 return r10_bio
->devs
[slot
].devnum
;
362 static void raid10_end_read_request(struct bio
*bio
)
364 int uptodate
= !bio
->bi_error
;
365 struct r10bio
*r10_bio
= bio
->bi_private
;
367 struct md_rdev
*rdev
;
368 struct r10conf
*conf
= r10_bio
->mddev
->private;
370 slot
= r10_bio
->read_slot
;
371 dev
= r10_bio
->devs
[slot
].devnum
;
372 rdev
= r10_bio
->devs
[slot
].rdev
;
374 * this branch is our 'one mirror IO has finished' event handler:
376 update_head_pos(slot
, r10_bio
);
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
388 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
395 if (!_enough(conf
, test_bit(R10BIO_Previous
, &r10_bio
->state
),
400 raid_end_bio_io(r10_bio
);
401 rdev_dec_pending(rdev
, conf
->mddev
);
404 * oops, read error - keep the refcount on the rdev
406 char b
[BDEVNAME_SIZE
];
407 printk_ratelimited(KERN_ERR
408 "md/raid10:%s: %s: rescheduling sector %llu\n",
410 bdevname(rdev
->bdev
, b
),
411 (unsigned long long)r10_bio
->sector
);
412 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
413 reschedule_retry(r10_bio
);
417 static void close_write(struct r10bio
*r10_bio
)
419 /* clear the bitmap if all writes complete successfully */
420 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
422 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
424 md_write_end(r10_bio
->mddev
);
427 static void one_write_done(struct r10bio
*r10_bio
)
429 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
430 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
431 reschedule_retry(r10_bio
);
433 close_write(r10_bio
);
434 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
435 reschedule_retry(r10_bio
);
437 raid_end_bio_io(r10_bio
);
442 static void raid10_end_write_request(struct bio
*bio
)
444 struct r10bio
*r10_bio
= bio
->bi_private
;
447 struct r10conf
*conf
= r10_bio
->mddev
->private;
449 struct md_rdev
*rdev
= NULL
;
451 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
454 rdev
= conf
->mirrors
[dev
].replacement
;
458 rdev
= conf
->mirrors
[dev
].rdev
;
461 * this branch is our 'one mirror IO has finished' event handler:
465 /* Never record new bad blocks to replacement,
468 md_error(rdev
->mddev
, rdev
);
470 set_bit(WriteErrorSeen
, &rdev
->flags
);
471 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
472 set_bit(MD_RECOVERY_NEEDED
,
473 &rdev
->mddev
->recovery
);
474 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
479 * Set R10BIO_Uptodate in our master bio, so that
480 * we will return a good error code for to the higher
481 * levels even if IO on some other mirrored buffer fails.
483 * The 'master' represents the composite IO operation to
484 * user-side. So if something waits for IO, then it will
485 * wait for the 'master' bio.
491 * Do not set R10BIO_Uptodate if the current device is
492 * rebuilding or Faulty. This is because we cannot use
493 * such device for properly reading the data back (we could
494 * potentially use it, if the current write would have felt
495 * before rdev->recovery_offset, but for simplicity we don't
498 if (test_bit(In_sync
, &rdev
->flags
) &&
499 !test_bit(Faulty
, &rdev
->flags
))
500 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
502 /* Maybe we can clear some bad blocks. */
503 if (is_badblock(rdev
,
504 r10_bio
->devs
[slot
].addr
,
506 &first_bad
, &bad_sectors
)) {
509 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
511 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
513 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
519 * Let's see if all mirrored write operations have finished
522 one_write_done(r10_bio
);
524 rdev_dec_pending(rdev
, conf
->mddev
);
528 * RAID10 layout manager
529 * As well as the chunksize and raid_disks count, there are two
530 * parameters: near_copies and far_copies.
531 * near_copies * far_copies must be <= raid_disks.
532 * Normally one of these will be 1.
533 * If both are 1, we get raid0.
534 * If near_copies == raid_disks, we get raid1.
536 * Chunks are laid out in raid0 style with near_copies copies of the
537 * first chunk, followed by near_copies copies of the next chunk and
539 * If far_copies > 1, then after 1/far_copies of the array has been assigned
540 * as described above, we start again with a device offset of near_copies.
541 * So we effectively have another copy of the whole array further down all
542 * the drives, but with blocks on different drives.
543 * With this layout, and block is never stored twice on the one device.
545 * raid10_find_phys finds the sector offset of a given virtual sector
546 * on each device that it is on.
548 * raid10_find_virt does the reverse mapping, from a device and a
549 * sector offset to a virtual address
552 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
560 int last_far_set_start
, last_far_set_size
;
562 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
563 last_far_set_start
*= geo
->far_set_size
;
565 last_far_set_size
= geo
->far_set_size
;
566 last_far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
568 /* now calculate first sector/dev */
569 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
570 sector
= r10bio
->sector
& geo
->chunk_mask
;
572 chunk
*= geo
->near_copies
;
574 dev
= sector_div(stripe
, geo
->raid_disks
);
576 stripe
*= geo
->far_copies
;
578 sector
+= stripe
<< geo
->chunk_shift
;
580 /* and calculate all the others */
581 for (n
= 0; n
< geo
->near_copies
; n
++) {
585 r10bio
->devs
[slot
].devnum
= d
;
586 r10bio
->devs
[slot
].addr
= s
;
589 for (f
= 1; f
< geo
->far_copies
; f
++) {
590 set
= d
/ geo
->far_set_size
;
591 d
+= geo
->near_copies
;
593 if ((geo
->raid_disks
% geo
->far_set_size
) &&
594 (d
> last_far_set_start
)) {
595 d
-= last_far_set_start
;
596 d
%= last_far_set_size
;
597 d
+= last_far_set_start
;
599 d
%= geo
->far_set_size
;
600 d
+= geo
->far_set_size
* set
;
603 r10bio
->devs
[slot
].devnum
= d
;
604 r10bio
->devs
[slot
].addr
= s
;
608 if (dev
>= geo
->raid_disks
) {
610 sector
+= (geo
->chunk_mask
+ 1);
615 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
617 struct geom
*geo
= &conf
->geo
;
619 if (conf
->reshape_progress
!= MaxSector
&&
620 ((r10bio
->sector
>= conf
->reshape_progress
) !=
621 conf
->mddev
->reshape_backwards
)) {
622 set_bit(R10BIO_Previous
, &r10bio
->state
);
625 clear_bit(R10BIO_Previous
, &r10bio
->state
);
627 __raid10_find_phys(geo
, r10bio
);
630 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
632 sector_t offset
, chunk
, vchunk
;
633 /* Never use conf->prev as this is only called during resync
634 * or recovery, so reshape isn't happening
636 struct geom
*geo
= &conf
->geo
;
637 int far_set_start
= (dev
/ geo
->far_set_size
) * geo
->far_set_size
;
638 int far_set_size
= geo
->far_set_size
;
639 int last_far_set_start
;
641 if (geo
->raid_disks
% geo
->far_set_size
) {
642 last_far_set_start
= (geo
->raid_disks
/ geo
->far_set_size
) - 1;
643 last_far_set_start
*= geo
->far_set_size
;
645 if (dev
>= last_far_set_start
) {
646 far_set_size
= geo
->far_set_size
;
647 far_set_size
+= (geo
->raid_disks
% geo
->far_set_size
);
648 far_set_start
= last_far_set_start
;
652 offset
= sector
& geo
->chunk_mask
;
653 if (geo
->far_offset
) {
655 chunk
= sector
>> geo
->chunk_shift
;
656 fc
= sector_div(chunk
, geo
->far_copies
);
657 dev
-= fc
* geo
->near_copies
;
658 if (dev
< far_set_start
)
661 while (sector
>= geo
->stride
) {
662 sector
-= geo
->stride
;
663 if (dev
< (geo
->near_copies
+ far_set_start
))
664 dev
+= far_set_size
- geo
->near_copies
;
666 dev
-= geo
->near_copies
;
668 chunk
= sector
>> geo
->chunk_shift
;
670 vchunk
= chunk
* geo
->raid_disks
+ dev
;
671 sector_div(vchunk
, geo
->near_copies
);
672 return (vchunk
<< geo
->chunk_shift
) + offset
;
676 * This routine returns the disk from which the requested read should
677 * be done. There is a per-array 'next expected sequential IO' sector
678 * number - if this matches on the next IO then we use the last disk.
679 * There is also a per-disk 'last know head position' sector that is
680 * maintained from IRQ contexts, both the normal and the resync IO
681 * completion handlers update this position correctly. If there is no
682 * perfect sequential match then we pick the disk whose head is closest.
684 * If there are 2 mirrors in the same 2 devices, performance degrades
685 * because position is mirror, not device based.
687 * The rdev for the device selected will have nr_pending incremented.
691 * FIXME: possibly should rethink readbalancing and do it differently
692 * depending on near_copies / far_copies geometry.
694 static struct md_rdev
*read_balance(struct r10conf
*conf
,
695 struct r10bio
*r10_bio
,
698 const sector_t this_sector
= r10_bio
->sector
;
700 int sectors
= r10_bio
->sectors
;
701 int best_good_sectors
;
702 sector_t new_distance
, best_dist
;
703 struct md_rdev
*best_rdev
, *rdev
= NULL
;
706 struct geom
*geo
= &conf
->geo
;
708 raid10_find_phys(conf
, r10_bio
);
710 sectors
= r10_bio
->sectors
;
713 best_dist
= MaxSector
;
714 best_good_sectors
= 0;
717 * Check if we can balance. We can balance on the whole
718 * device if no resync is going on (recovery is ok), or below
719 * the resync window. We take the first readable disk when
720 * above the resync window.
722 if (conf
->mddev
->recovery_cp
< MaxSector
723 && (this_sector
+ sectors
>= conf
->next_resync
))
726 for (slot
= 0; slot
< conf
->copies
; slot
++) {
731 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
733 disk
= r10_bio
->devs
[slot
].devnum
;
734 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
735 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
736 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
737 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
739 test_bit(Faulty
, &rdev
->flags
))
741 if (!test_bit(In_sync
, &rdev
->flags
) &&
742 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
745 dev_sector
= r10_bio
->devs
[slot
].addr
;
746 if (is_badblock(rdev
, dev_sector
, sectors
,
747 &first_bad
, &bad_sectors
)) {
748 if (best_dist
< MaxSector
)
749 /* Already have a better slot */
751 if (first_bad
<= dev_sector
) {
752 /* Cannot read here. If this is the
753 * 'primary' device, then we must not read
754 * beyond 'bad_sectors' from another device.
756 bad_sectors
-= (dev_sector
- first_bad
);
757 if (!do_balance
&& sectors
> bad_sectors
)
758 sectors
= bad_sectors
;
759 if (best_good_sectors
> sectors
)
760 best_good_sectors
= sectors
;
762 sector_t good_sectors
=
763 first_bad
- dev_sector
;
764 if (good_sectors
> best_good_sectors
) {
765 best_good_sectors
= good_sectors
;
770 /* Must read from here */
775 best_good_sectors
= sectors
;
780 /* This optimisation is debatable, and completely destroys
781 * sequential read speed for 'far copies' arrays. So only
782 * keep it for 'near' arrays, and review those later.
784 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
787 /* for far > 1 always use the lowest address */
788 if (geo
->far_copies
> 1)
789 new_distance
= r10_bio
->devs
[slot
].addr
;
791 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
792 conf
->mirrors
[disk
].head_position
);
793 if (new_distance
< best_dist
) {
794 best_dist
= new_distance
;
799 if (slot
>= conf
->copies
) {
805 atomic_inc(&rdev
->nr_pending
);
806 r10_bio
->read_slot
= slot
;
810 *max_sectors
= best_good_sectors
;
815 static int raid10_congested(struct mddev
*mddev
, int bits
)
817 struct r10conf
*conf
= mddev
->private;
820 if ((bits
& (1 << WB_async_congested
)) &&
821 conf
->pending_count
>= max_queued_requests
)
826 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
829 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
830 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
831 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
833 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
840 static void flush_pending_writes(struct r10conf
*conf
)
842 /* Any writes that have been queued but are awaiting
843 * bitmap updates get flushed here.
845 spin_lock_irq(&conf
->device_lock
);
847 if (conf
->pending_bio_list
.head
) {
849 bio
= bio_list_get(&conf
->pending_bio_list
);
850 conf
->pending_count
= 0;
851 spin_unlock_irq(&conf
->device_lock
);
852 /* flush any pending bitmap writes to disk
853 * before proceeding w/ I/O */
854 bitmap_unplug(conf
->mddev
->bitmap
);
855 wake_up(&conf
->wait_barrier
);
857 while (bio
) { /* submit pending writes */
858 struct bio
*next
= bio
->bi_next
;
860 if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
861 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
865 generic_make_request(bio
);
869 spin_unlock_irq(&conf
->device_lock
);
873 * Sometimes we need to suspend IO while we do something else,
874 * either some resync/recovery, or reconfigure the array.
875 * To do this we raise a 'barrier'.
876 * The 'barrier' is a counter that can be raised multiple times
877 * to count how many activities are happening which preclude
879 * We can only raise the barrier if there is no pending IO.
880 * i.e. if nr_pending == 0.
881 * We choose only to raise the barrier if no-one is waiting for the
882 * barrier to go down. This means that as soon as an IO request
883 * is ready, no other operations which require a barrier will start
884 * until the IO request has had a chance.
886 * So: regular IO calls 'wait_barrier'. When that returns there
887 * is no backgroup IO happening, It must arrange to call
888 * allow_barrier when it has finished its IO.
889 * backgroup IO calls must call raise_barrier. Once that returns
890 * there is no normal IO happeing. It must arrange to call
891 * lower_barrier when the particular background IO completes.
894 static void raise_barrier(struct r10conf
*conf
, int force
)
896 BUG_ON(force
&& !conf
->barrier
);
897 spin_lock_irq(&conf
->resync_lock
);
899 /* Wait until no block IO is waiting (unless 'force') */
900 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
903 /* block any new IO from starting */
906 /* Now wait for all pending IO to complete */
907 wait_event_lock_irq(conf
->wait_barrier
,
908 !atomic_read(&conf
->nr_pending
) && conf
->barrier
< RESYNC_DEPTH
,
911 spin_unlock_irq(&conf
->resync_lock
);
914 static void lower_barrier(struct r10conf
*conf
)
917 spin_lock_irqsave(&conf
->resync_lock
, flags
);
919 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
920 wake_up(&conf
->wait_barrier
);
923 static void wait_barrier(struct r10conf
*conf
)
925 spin_lock_irq(&conf
->resync_lock
);
928 /* Wait for the barrier to drop.
929 * However if there are already pending
930 * requests (preventing the barrier from
931 * rising completely), and the
932 * pre-process bio queue isn't empty,
933 * then don't wait, as we need to empty
934 * that queue to get the nr_pending
937 wait_event_lock_irq(conf
->wait_barrier
,
939 (atomic_read(&conf
->nr_pending
) &&
941 !bio_list_empty(current
->bio_list
)),
944 if (!conf
->nr_waiting
)
945 wake_up(&conf
->wait_barrier
);
947 atomic_inc(&conf
->nr_pending
);
948 spin_unlock_irq(&conf
->resync_lock
);
951 static void allow_barrier(struct r10conf
*conf
)
953 if ((atomic_dec_and_test(&conf
->nr_pending
)) ||
954 (conf
->array_freeze_pending
))
955 wake_up(&conf
->wait_barrier
);
958 static void freeze_array(struct r10conf
*conf
, int extra
)
960 /* stop syncio and normal IO and wait for everything to
962 * We increment barrier and nr_waiting, and then
963 * wait until nr_pending match nr_queued+extra
964 * This is called in the context of one normal IO request
965 * that has failed. Thus any sync request that might be pending
966 * will be blocked by nr_pending, and we need to wait for
967 * pending IO requests to complete or be queued for re-try.
968 * Thus the number queued (nr_queued) plus this request (extra)
969 * must match the number of pending IOs (nr_pending) before
972 spin_lock_irq(&conf
->resync_lock
);
973 conf
->array_freeze_pending
++;
976 wait_event_lock_irq_cmd(conf
->wait_barrier
,
977 atomic_read(&conf
->nr_pending
) == conf
->nr_queued
+extra
,
979 flush_pending_writes(conf
));
981 conf
->array_freeze_pending
--;
982 spin_unlock_irq(&conf
->resync_lock
);
985 static void unfreeze_array(struct r10conf
*conf
)
987 /* reverse the effect of the freeze */
988 spin_lock_irq(&conf
->resync_lock
);
991 wake_up(&conf
->wait_barrier
);
992 spin_unlock_irq(&conf
->resync_lock
);
995 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
996 struct md_rdev
*rdev
)
998 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
999 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1000 return rdev
->data_offset
;
1002 return rdev
->new_data_offset
;
1005 struct raid10_plug_cb
{
1006 struct blk_plug_cb cb
;
1007 struct bio_list pending
;
1011 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1013 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1015 struct mddev
*mddev
= plug
->cb
.data
;
1016 struct r10conf
*conf
= mddev
->private;
1019 if (from_schedule
|| current
->bio_list
) {
1020 spin_lock_irq(&conf
->device_lock
);
1021 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1022 conf
->pending_count
+= plug
->pending_cnt
;
1023 spin_unlock_irq(&conf
->device_lock
);
1024 wake_up(&conf
->wait_barrier
);
1025 md_wakeup_thread(mddev
->thread
);
1030 /* we aren't scheduling, so we can do the write-out directly. */
1031 bio
= bio_list_get(&plug
->pending
);
1032 bitmap_unplug(mddev
->bitmap
);
1033 wake_up(&conf
->wait_barrier
);
1035 while (bio
) { /* submit pending writes */
1036 struct bio
*next
= bio
->bi_next
;
1037 bio
->bi_next
= NULL
;
1038 if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
1039 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1040 /* Just ignore it */
1043 generic_make_request(bio
);
1049 static void __make_request(struct mddev
*mddev
, struct bio
*bio
)
1051 struct r10conf
*conf
= mddev
->private;
1052 struct r10bio
*r10_bio
;
1053 struct bio
*read_bio
;
1055 const int op
= bio_op(bio
);
1056 const int rw
= bio_data_dir(bio
);
1057 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1058 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1059 unsigned long flags
;
1060 struct md_rdev
*blocked_rdev
;
1061 struct blk_plug_cb
*cb
;
1062 struct raid10_plug_cb
*plug
= NULL
;
1063 int sectors_handled
;
1068 * Register the new request and wait if the reconstruction
1069 * thread has put up a bar for new requests.
1070 * Continue immediately if no resync is active currently.
1074 sectors
= bio_sectors(bio
);
1075 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1076 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1077 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1078 /* IO spans the reshape position. Need to wait for
1081 allow_barrier(conf
);
1082 wait_event(conf
->wait_barrier
,
1083 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1084 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1088 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1089 bio_data_dir(bio
) == WRITE
&&
1090 (mddev
->reshape_backwards
1091 ? (bio
->bi_iter
.bi_sector
< conf
->reshape_safe
&&
1092 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
)
1093 : (bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_safe
&&
1094 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
))) {
1095 /* Need to update reshape_position in metadata */
1096 mddev
->reshape_position
= conf
->reshape_progress
;
1097 set_mask_bits(&mddev
->flags
, 0,
1098 BIT(MD_CHANGE_DEVS
) | BIT(MD_CHANGE_PENDING
));
1099 md_wakeup_thread(mddev
->thread
);
1100 wait_event(mddev
->sb_wait
,
1101 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1103 conf
->reshape_safe
= mddev
->reshape_position
;
1106 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1108 r10_bio
->master_bio
= bio
;
1109 r10_bio
->sectors
= sectors
;
1111 r10_bio
->mddev
= mddev
;
1112 r10_bio
->sector
= bio
->bi_iter
.bi_sector
;
1115 /* We might need to issue multiple reads to different
1116 * devices if there are bad blocks around, so we keep
1117 * track of the number of reads in bio->bi_phys_segments.
1118 * If this is 0, there is only one r10_bio and no locking
1119 * will be needed when the request completes. If it is
1120 * non-zero, then it is the number of not-completed requests.
1122 bio
->bi_phys_segments
= 0;
1123 bio_clear_flag(bio
, BIO_SEG_VALID
);
1127 * read balancing logic:
1129 struct md_rdev
*rdev
;
1133 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1135 raid_end_bio_io(r10_bio
);
1138 slot
= r10_bio
->read_slot
;
1140 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1141 bio_trim(read_bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1144 r10_bio
->devs
[slot
].bio
= read_bio
;
1145 r10_bio
->devs
[slot
].rdev
= rdev
;
1147 read_bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
+
1148 choose_data_offset(r10_bio
, rdev
);
1149 read_bio
->bi_bdev
= rdev
->bdev
;
1150 read_bio
->bi_end_io
= raid10_end_read_request
;
1151 bio_set_op_attrs(read_bio
, op
, do_sync
);
1152 read_bio
->bi_private
= r10_bio
;
1154 if (max_sectors
< r10_bio
->sectors
) {
1155 /* Could not read all from this device, so we will
1156 * need another r10_bio.
1158 sectors_handled
= (r10_bio
->sector
+ max_sectors
1159 - bio
->bi_iter
.bi_sector
);
1160 r10_bio
->sectors
= max_sectors
;
1161 spin_lock_irq(&conf
->device_lock
);
1162 if (bio
->bi_phys_segments
== 0)
1163 bio
->bi_phys_segments
= 2;
1165 bio
->bi_phys_segments
++;
1166 spin_unlock_irq(&conf
->device_lock
);
1167 /* Cannot call generic_make_request directly
1168 * as that will be queued in __generic_make_request
1169 * and subsequent mempool_alloc might block
1170 * waiting for it. so hand bio over to raid10d.
1172 reschedule_retry(r10_bio
);
1174 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1176 r10_bio
->master_bio
= bio
;
1177 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1179 r10_bio
->mddev
= mddev
;
1180 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+
1184 generic_make_request(read_bio
);
1191 if (conf
->pending_count
>= max_queued_requests
) {
1192 md_wakeup_thread(mddev
->thread
);
1193 wait_event(conf
->wait_barrier
,
1194 conf
->pending_count
< max_queued_requests
);
1196 /* first select target devices under rcu_lock and
1197 * inc refcount on their rdev. Record them by setting
1199 * If there are known/acknowledged bad blocks on any device
1200 * on which we have seen a write error, we want to avoid
1201 * writing to those blocks. This potentially requires several
1202 * writes to write around the bad blocks. Each set of writes
1203 * gets its own r10_bio with a set of bios attached. The number
1204 * of r10_bios is recored in bio->bi_phys_segments just as with
1208 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1209 raid10_find_phys(conf
, r10_bio
);
1211 blocked_rdev
= NULL
;
1213 max_sectors
= r10_bio
->sectors
;
1215 for (i
= 0; i
< conf
->copies
; i
++) {
1216 int d
= r10_bio
->devs
[i
].devnum
;
1217 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1218 struct md_rdev
*rrdev
= rcu_dereference(
1219 conf
->mirrors
[d
].replacement
);
1222 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1223 atomic_inc(&rdev
->nr_pending
);
1224 blocked_rdev
= rdev
;
1227 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1228 atomic_inc(&rrdev
->nr_pending
);
1229 blocked_rdev
= rrdev
;
1232 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)))
1234 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)))
1237 r10_bio
->devs
[i
].bio
= NULL
;
1238 r10_bio
->devs
[i
].repl_bio
= NULL
;
1240 if (!rdev
&& !rrdev
) {
1241 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1244 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1246 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1250 is_bad
= is_badblock(rdev
, dev_sector
,
1252 &first_bad
, &bad_sectors
);
1254 /* Mustn't write here until the bad block
1257 atomic_inc(&rdev
->nr_pending
);
1258 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1259 blocked_rdev
= rdev
;
1262 if (is_bad
&& first_bad
<= dev_sector
) {
1263 /* Cannot write here at all */
1264 bad_sectors
-= (dev_sector
- first_bad
);
1265 if (bad_sectors
< max_sectors
)
1266 /* Mustn't write more than bad_sectors
1267 * to other devices yet
1269 max_sectors
= bad_sectors
;
1270 /* We don't set R10BIO_Degraded as that
1271 * only applies if the disk is missing,
1272 * so it might be re-added, and we want to
1273 * know to recover this chunk.
1274 * In this case the device is here, and the
1275 * fact that this chunk is not in-sync is
1276 * recorded in the bad block log.
1281 int good_sectors
= first_bad
- dev_sector
;
1282 if (good_sectors
< max_sectors
)
1283 max_sectors
= good_sectors
;
1287 r10_bio
->devs
[i
].bio
= bio
;
1288 atomic_inc(&rdev
->nr_pending
);
1291 r10_bio
->devs
[i
].repl_bio
= bio
;
1292 atomic_inc(&rrdev
->nr_pending
);
1297 if (unlikely(blocked_rdev
)) {
1298 /* Have to wait for this device to get unblocked, then retry */
1302 for (j
= 0; j
< i
; j
++) {
1303 if (r10_bio
->devs
[j
].bio
) {
1304 d
= r10_bio
->devs
[j
].devnum
;
1305 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1307 if (r10_bio
->devs
[j
].repl_bio
) {
1308 struct md_rdev
*rdev
;
1309 d
= r10_bio
->devs
[j
].devnum
;
1310 rdev
= conf
->mirrors
[d
].replacement
;
1312 /* Race with remove_disk */
1314 rdev
= conf
->mirrors
[d
].rdev
;
1316 rdev_dec_pending(rdev
, mddev
);
1319 allow_barrier(conf
);
1320 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1325 if (max_sectors
< r10_bio
->sectors
) {
1326 /* We are splitting this into multiple parts, so
1327 * we need to prepare for allocating another r10_bio.
1329 r10_bio
->sectors
= max_sectors
;
1330 spin_lock_irq(&conf
->device_lock
);
1331 if (bio
->bi_phys_segments
== 0)
1332 bio
->bi_phys_segments
= 2;
1334 bio
->bi_phys_segments
++;
1335 spin_unlock_irq(&conf
->device_lock
);
1337 sectors_handled
= r10_bio
->sector
+ max_sectors
-
1338 bio
->bi_iter
.bi_sector
;
1340 atomic_set(&r10_bio
->remaining
, 1);
1341 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1343 for (i
= 0; i
< conf
->copies
; i
++) {
1345 int d
= r10_bio
->devs
[i
].devnum
;
1346 if (r10_bio
->devs
[i
].bio
) {
1347 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1348 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1349 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1351 r10_bio
->devs
[i
].bio
= mbio
;
1353 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1354 choose_data_offset(r10_bio
,
1356 mbio
->bi_bdev
= rdev
->bdev
;
1357 mbio
->bi_end_io
= raid10_end_write_request
;
1358 bio_set_op_attrs(mbio
, op
, do_sync
| do_fua
);
1359 mbio
->bi_private
= r10_bio
;
1361 atomic_inc(&r10_bio
->remaining
);
1363 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1366 plug
= container_of(cb
, struct raid10_plug_cb
,
1370 spin_lock_irqsave(&conf
->device_lock
, flags
);
1372 bio_list_add(&plug
->pending
, mbio
);
1373 plug
->pending_cnt
++;
1375 bio_list_add(&conf
->pending_bio_list
, mbio
);
1376 conf
->pending_count
++;
1378 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1380 md_wakeup_thread(mddev
->thread
);
1383 if (r10_bio
->devs
[i
].repl_bio
) {
1384 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1386 /* Replacement just got moved to main 'rdev' */
1388 rdev
= conf
->mirrors
[d
].rdev
;
1390 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1391 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1393 r10_bio
->devs
[i
].repl_bio
= mbio
;
1395 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1398 mbio
->bi_bdev
= rdev
->bdev
;
1399 mbio
->bi_end_io
= raid10_end_write_request
;
1400 bio_set_op_attrs(mbio
, op
, do_sync
| do_fua
);
1401 mbio
->bi_private
= r10_bio
;
1403 atomic_inc(&r10_bio
->remaining
);
1404 spin_lock_irqsave(&conf
->device_lock
, flags
);
1405 bio_list_add(&conf
->pending_bio_list
, mbio
);
1406 conf
->pending_count
++;
1407 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1408 if (!mddev_check_plugged(mddev
))
1409 md_wakeup_thread(mddev
->thread
);
1413 /* Don't remove the bias on 'remaining' (one_write_done) until
1414 * after checking if we need to go around again.
1417 if (sectors_handled
< bio_sectors(bio
)) {
1418 one_write_done(r10_bio
);
1419 /* We need another r10_bio. It has already been counted
1420 * in bio->bi_phys_segments.
1422 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1424 r10_bio
->master_bio
= bio
;
1425 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1427 r10_bio
->mddev
= mddev
;
1428 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1432 one_write_done(r10_bio
);
1435 static void raid10_make_request(struct mddev
*mddev
, struct bio
*bio
)
1437 struct r10conf
*conf
= mddev
->private;
1438 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1439 int chunk_sects
= chunk_mask
+ 1;
1443 if (unlikely(bio
->bi_rw
& REQ_PREFLUSH
)) {
1444 md_flush_request(mddev
, bio
);
1448 md_write_start(mddev
, bio
);
1453 * If this request crosses a chunk boundary, we need to split
1456 if (unlikely((bio
->bi_iter
.bi_sector
& chunk_mask
) +
1457 bio_sectors(bio
) > chunk_sects
1458 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1459 || conf
->prev
.near_copies
<
1460 conf
->prev
.raid_disks
))) {
1461 split
= bio_split(bio
, chunk_sects
-
1462 (bio
->bi_iter
.bi_sector
&
1464 GFP_NOIO
, fs_bio_set
);
1465 bio_chain(split
, bio
);
1470 __make_request(mddev
, split
);
1471 } while (split
!= bio
);
1473 /* In case raid10d snuck in to freeze_array */
1474 wake_up(&conf
->wait_barrier
);
1477 static void raid10_status(struct seq_file
*seq
, struct mddev
*mddev
)
1479 struct r10conf
*conf
= mddev
->private;
1482 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1483 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1484 if (conf
->geo
.near_copies
> 1)
1485 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1486 if (conf
->geo
.far_copies
> 1) {
1487 if (conf
->geo
.far_offset
)
1488 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1490 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1491 if (conf
->geo
.far_set_size
!= conf
->geo
.raid_disks
)
1492 seq_printf(seq
, " %d devices per set", conf
->geo
.far_set_size
);
1494 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1495 conf
->geo
.raid_disks
- mddev
->degraded
);
1497 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1498 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1499 seq_printf(seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1502 seq_printf(seq
, "]");
1505 /* check if there are enough drives for
1506 * every block to appear on atleast one.
1507 * Don't consider the device numbered 'ignore'
1508 * as we might be about to remove it.
1510 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1516 disks
= conf
->prev
.raid_disks
;
1517 ncopies
= conf
->prev
.near_copies
;
1519 disks
= conf
->geo
.raid_disks
;
1520 ncopies
= conf
->geo
.near_copies
;
1525 int n
= conf
->copies
;
1529 struct md_rdev
*rdev
;
1530 if (this != ignore
&&
1531 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1532 test_bit(In_sync
, &rdev
->flags
))
1534 this = (this+1) % disks
;
1538 first
= (first
+ ncopies
) % disks
;
1539 } while (first
!= 0);
1546 static int enough(struct r10conf
*conf
, int ignore
)
1548 /* when calling 'enough', both 'prev' and 'geo' must
1550 * This is ensured if ->reconfig_mutex or ->device_lock
1553 return _enough(conf
, 0, ignore
) &&
1554 _enough(conf
, 1, ignore
);
1557 static void raid10_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1559 char b
[BDEVNAME_SIZE
];
1560 struct r10conf
*conf
= mddev
->private;
1561 unsigned long flags
;
1564 * If it is not operational, then we have already marked it as dead
1565 * else if it is the last working disks, ignore the error, let the
1566 * next level up know.
1567 * else mark the drive as failed
1569 spin_lock_irqsave(&conf
->device_lock
, flags
);
1570 if (test_bit(In_sync
, &rdev
->flags
)
1571 && !enough(conf
, rdev
->raid_disk
)) {
1573 * Don't fail the drive, just return an IO error.
1575 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1578 if (test_and_clear_bit(In_sync
, &rdev
->flags
))
1581 * If recovery is running, make sure it aborts.
1583 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1584 set_bit(Blocked
, &rdev
->flags
);
1585 set_bit(Faulty
, &rdev
->flags
);
1586 set_mask_bits(&mddev
->flags
, 0,
1587 BIT(MD_CHANGE_DEVS
) | BIT(MD_CHANGE_PENDING
));
1588 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1590 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1591 "md/raid10:%s: Operation continuing on %d devices.\n",
1592 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1593 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1596 static void print_conf(struct r10conf
*conf
)
1599 struct md_rdev
*rdev
;
1601 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1603 printk(KERN_DEBUG
"(!conf)\n");
1606 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1607 conf
->geo
.raid_disks
);
1609 /* This is only called with ->reconfix_mutex held, so
1610 * rcu protection of rdev is not needed */
1611 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1612 char b
[BDEVNAME_SIZE
];
1613 rdev
= conf
->mirrors
[i
].rdev
;
1615 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1616 i
, !test_bit(In_sync
, &rdev
->flags
),
1617 !test_bit(Faulty
, &rdev
->flags
),
1618 bdevname(rdev
->bdev
,b
));
1622 static void close_sync(struct r10conf
*conf
)
1625 allow_barrier(conf
);
1627 mempool_destroy(conf
->r10buf_pool
);
1628 conf
->r10buf_pool
= NULL
;
1631 static int raid10_spare_active(struct mddev
*mddev
)
1634 struct r10conf
*conf
= mddev
->private;
1635 struct raid10_info
*tmp
;
1637 unsigned long flags
;
1640 * Find all non-in_sync disks within the RAID10 configuration
1641 * and mark them in_sync
1643 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1644 tmp
= conf
->mirrors
+ i
;
1645 if (tmp
->replacement
1646 && tmp
->replacement
->recovery_offset
== MaxSector
1647 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1648 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1649 /* Replacement has just become active */
1651 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1654 /* Replaced device not technically faulty,
1655 * but we need to be sure it gets removed
1656 * and never re-added.
1658 set_bit(Faulty
, &tmp
->rdev
->flags
);
1659 sysfs_notify_dirent_safe(
1660 tmp
->rdev
->sysfs_state
);
1662 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1663 } else if (tmp
->rdev
1664 && tmp
->rdev
->recovery_offset
== MaxSector
1665 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1666 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1668 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1671 spin_lock_irqsave(&conf
->device_lock
, flags
);
1672 mddev
->degraded
-= count
;
1673 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1679 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1681 struct r10conf
*conf
= mddev
->private;
1685 int last
= conf
->geo
.raid_disks
- 1;
1687 if (mddev
->recovery_cp
< MaxSector
)
1688 /* only hot-add to in-sync arrays, as recovery is
1689 * very different from resync
1692 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1695 if (md_integrity_add_rdev(rdev
, mddev
))
1698 if (rdev
->raid_disk
>= 0)
1699 first
= last
= rdev
->raid_disk
;
1701 if (rdev
->saved_raid_disk
>= first
&&
1702 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1703 mirror
= rdev
->saved_raid_disk
;
1706 for ( ; mirror
<= last
; mirror
++) {
1707 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1708 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1711 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1712 p
->replacement
!= NULL
)
1714 clear_bit(In_sync
, &rdev
->flags
);
1715 set_bit(Replacement
, &rdev
->flags
);
1716 rdev
->raid_disk
= mirror
;
1719 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1720 rdev
->data_offset
<< 9);
1722 rcu_assign_pointer(p
->replacement
, rdev
);
1727 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1728 rdev
->data_offset
<< 9);
1730 p
->head_position
= 0;
1731 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1732 rdev
->raid_disk
= mirror
;
1734 if (rdev
->saved_raid_disk
!= mirror
)
1736 rcu_assign_pointer(p
->rdev
, rdev
);
1739 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1740 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1746 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1748 struct r10conf
*conf
= mddev
->private;
1750 int number
= rdev
->raid_disk
;
1751 struct md_rdev
**rdevp
;
1752 struct raid10_info
*p
= conf
->mirrors
+ number
;
1755 if (rdev
== p
->rdev
)
1757 else if (rdev
== p
->replacement
)
1758 rdevp
= &p
->replacement
;
1762 if (test_bit(In_sync
, &rdev
->flags
) ||
1763 atomic_read(&rdev
->nr_pending
)) {
1767 /* Only remove non-faulty devices if recovery
1770 if (!test_bit(Faulty
, &rdev
->flags
) &&
1771 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1772 (!p
->replacement
|| p
->replacement
== rdev
) &&
1773 number
< conf
->geo
.raid_disks
&&
1779 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
1781 if (atomic_read(&rdev
->nr_pending
)) {
1782 /* lost the race, try later */
1788 if (p
->replacement
) {
1789 /* We must have just cleared 'rdev' */
1790 p
->rdev
= p
->replacement
;
1791 clear_bit(Replacement
, &p
->replacement
->flags
);
1792 smp_mb(); /* Make sure other CPUs may see both as identical
1793 * but will never see neither -- if they are careful.
1795 p
->replacement
= NULL
;
1796 clear_bit(WantReplacement
, &rdev
->flags
);
1798 /* We might have just remove the Replacement as faulty
1799 * Clear the flag just in case
1801 clear_bit(WantReplacement
, &rdev
->flags
);
1803 err
= md_integrity_register(mddev
);
1811 static void end_sync_read(struct bio
*bio
)
1813 struct r10bio
*r10_bio
= bio
->bi_private
;
1814 struct r10conf
*conf
= r10_bio
->mddev
->private;
1817 if (bio
== r10_bio
->master_bio
) {
1818 /* this is a reshape read */
1819 d
= r10_bio
->read_slot
; /* really the read dev */
1821 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1824 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1826 /* The write handler will notice the lack of
1827 * R10BIO_Uptodate and record any errors etc
1829 atomic_add(r10_bio
->sectors
,
1830 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1832 /* for reconstruct, we always reschedule after a read.
1833 * for resync, only after all reads
1835 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1836 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1837 atomic_dec_and_test(&r10_bio
->remaining
)) {
1838 /* we have read all the blocks,
1839 * do the comparison in process context in raid10d
1841 reschedule_retry(r10_bio
);
1845 static void end_sync_request(struct r10bio
*r10_bio
)
1847 struct mddev
*mddev
= r10_bio
->mddev
;
1849 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1850 if (r10_bio
->master_bio
== NULL
) {
1851 /* the primary of several recovery bios */
1852 sector_t s
= r10_bio
->sectors
;
1853 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1854 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1855 reschedule_retry(r10_bio
);
1858 md_done_sync(mddev
, s
, 1);
1861 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1862 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1863 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1864 reschedule_retry(r10_bio
);
1872 static void end_sync_write(struct bio
*bio
)
1874 struct r10bio
*r10_bio
= bio
->bi_private
;
1875 struct mddev
*mddev
= r10_bio
->mddev
;
1876 struct r10conf
*conf
= mddev
->private;
1882 struct md_rdev
*rdev
= NULL
;
1884 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1886 rdev
= conf
->mirrors
[d
].replacement
;
1888 rdev
= conf
->mirrors
[d
].rdev
;
1890 if (bio
->bi_error
) {
1892 md_error(mddev
, rdev
);
1894 set_bit(WriteErrorSeen
, &rdev
->flags
);
1895 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1896 set_bit(MD_RECOVERY_NEEDED
,
1897 &rdev
->mddev
->recovery
);
1898 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1900 } else if (is_badblock(rdev
,
1901 r10_bio
->devs
[slot
].addr
,
1903 &first_bad
, &bad_sectors
))
1904 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1906 rdev_dec_pending(rdev
, mddev
);
1908 end_sync_request(r10_bio
);
1912 * Note: sync and recover and handled very differently for raid10
1913 * This code is for resync.
1914 * For resync, we read through virtual addresses and read all blocks.
1915 * If there is any error, we schedule a write. The lowest numbered
1916 * drive is authoritative.
1917 * However requests come for physical address, so we need to map.
1918 * For every physical address there are raid_disks/copies virtual addresses,
1919 * which is always are least one, but is not necessarly an integer.
1920 * This means that a physical address can span multiple chunks, so we may
1921 * have to submit multiple io requests for a single sync request.
1924 * We check if all blocks are in-sync and only write to blocks that
1927 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1929 struct r10conf
*conf
= mddev
->private;
1931 struct bio
*tbio
, *fbio
;
1934 atomic_set(&r10_bio
->remaining
, 1);
1936 /* find the first device with a block */
1937 for (i
=0; i
<conf
->copies
; i
++)
1938 if (!r10_bio
->devs
[i
].bio
->bi_error
)
1941 if (i
== conf
->copies
)
1945 fbio
= r10_bio
->devs
[i
].bio
;
1946 fbio
->bi_iter
.bi_size
= r10_bio
->sectors
<< 9;
1947 fbio
->bi_iter
.bi_idx
= 0;
1949 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
1950 /* now find blocks with errors */
1951 for (i
=0 ; i
< conf
->copies
; i
++) {
1954 tbio
= r10_bio
->devs
[i
].bio
;
1956 if (tbio
->bi_end_io
!= end_sync_read
)
1960 if (!r10_bio
->devs
[i
].bio
->bi_error
) {
1961 /* We know that the bi_io_vec layout is the same for
1962 * both 'first' and 'i', so we just compare them.
1963 * All vec entries are PAGE_SIZE;
1965 int sectors
= r10_bio
->sectors
;
1966 for (j
= 0; j
< vcnt
; j
++) {
1967 int len
= PAGE_SIZE
;
1968 if (sectors
< (len
/ 512))
1969 len
= sectors
* 512;
1970 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1971 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1978 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
1979 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1980 /* Don't fix anything. */
1983 /* Ok, we need to write this bio, either to correct an
1984 * inconsistency or to correct an unreadable block.
1985 * First we need to fixup bv_offset, bv_len and
1986 * bi_vecs, as the read request might have corrupted these
1990 tbio
->bi_vcnt
= vcnt
;
1991 tbio
->bi_iter
.bi_size
= fbio
->bi_iter
.bi_size
;
1992 tbio
->bi_private
= r10_bio
;
1993 tbio
->bi_iter
.bi_sector
= r10_bio
->devs
[i
].addr
;
1994 tbio
->bi_end_io
= end_sync_write
;
1995 bio_set_op_attrs(tbio
, REQ_OP_WRITE
, 0);
1997 bio_copy_data(tbio
, fbio
);
1999 d
= r10_bio
->devs
[i
].devnum
;
2000 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2001 atomic_inc(&r10_bio
->remaining
);
2002 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2004 tbio
->bi_iter
.bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2005 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2006 generic_make_request(tbio
);
2009 /* Now write out to any replacement devices
2012 for (i
= 0; i
< conf
->copies
; i
++) {
2015 tbio
= r10_bio
->devs
[i
].repl_bio
;
2016 if (!tbio
|| !tbio
->bi_end_io
)
2018 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2019 && r10_bio
->devs
[i
].bio
!= fbio
)
2020 bio_copy_data(tbio
, fbio
);
2021 d
= r10_bio
->devs
[i
].devnum
;
2022 atomic_inc(&r10_bio
->remaining
);
2023 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2025 generic_make_request(tbio
);
2029 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2030 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2036 * Now for the recovery code.
2037 * Recovery happens across physical sectors.
2038 * We recover all non-is_sync drives by finding the virtual address of
2039 * each, and then choose a working drive that also has that virt address.
2040 * There is a separate r10_bio for each non-in_sync drive.
2041 * Only the first two slots are in use. The first for reading,
2042 * The second for writing.
2045 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2047 /* We got a read error during recovery.
2048 * We repeat the read in smaller page-sized sections.
2049 * If a read succeeds, write it to the new device or record
2050 * a bad block if we cannot.
2051 * If a read fails, record a bad block on both old and
2054 struct mddev
*mddev
= r10_bio
->mddev
;
2055 struct r10conf
*conf
= mddev
->private;
2056 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2058 int sectors
= r10_bio
->sectors
;
2060 int dr
= r10_bio
->devs
[0].devnum
;
2061 int dw
= r10_bio
->devs
[1].devnum
;
2065 struct md_rdev
*rdev
;
2069 if (s
> (PAGE_SIZE
>>9))
2072 rdev
= conf
->mirrors
[dr
].rdev
;
2073 addr
= r10_bio
->devs
[0].addr
+ sect
,
2074 ok
= sync_page_io(rdev
,
2077 bio
->bi_io_vec
[idx
].bv_page
,
2078 REQ_OP_READ
, 0, false);
2080 rdev
= conf
->mirrors
[dw
].rdev
;
2081 addr
= r10_bio
->devs
[1].addr
+ sect
;
2082 ok
= sync_page_io(rdev
,
2085 bio
->bi_io_vec
[idx
].bv_page
,
2086 REQ_OP_WRITE
, 0, false);
2088 set_bit(WriteErrorSeen
, &rdev
->flags
);
2089 if (!test_and_set_bit(WantReplacement
,
2091 set_bit(MD_RECOVERY_NEEDED
,
2092 &rdev
->mddev
->recovery
);
2096 /* We don't worry if we cannot set a bad block -
2097 * it really is bad so there is no loss in not
2100 rdev_set_badblocks(rdev
, addr
, s
, 0);
2102 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2103 /* need bad block on destination too */
2104 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2105 addr
= r10_bio
->devs
[1].addr
+ sect
;
2106 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2108 /* just abort the recovery */
2110 "md/raid10:%s: recovery aborted"
2111 " due to read error\n",
2114 conf
->mirrors
[dw
].recovery_disabled
2115 = mddev
->recovery_disabled
;
2116 set_bit(MD_RECOVERY_INTR
,
2129 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2131 struct r10conf
*conf
= mddev
->private;
2133 struct bio
*wbio
, *wbio2
;
2135 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2136 fix_recovery_read_error(r10_bio
);
2137 end_sync_request(r10_bio
);
2142 * share the pages with the first bio
2143 * and submit the write request
2145 d
= r10_bio
->devs
[1].devnum
;
2146 wbio
= r10_bio
->devs
[1].bio
;
2147 wbio2
= r10_bio
->devs
[1].repl_bio
;
2148 /* Need to test wbio2->bi_end_io before we call
2149 * generic_make_request as if the former is NULL,
2150 * the latter is free to free wbio2.
2152 if (wbio2
&& !wbio2
->bi_end_io
)
2154 if (wbio
->bi_end_io
) {
2155 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2156 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2157 generic_make_request(wbio
);
2160 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2161 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2162 bio_sectors(wbio2
));
2163 generic_make_request(wbio2
);
2168 * Used by fix_read_error() to decay the per rdev read_errors.
2169 * We halve the read error count for every hour that has elapsed
2170 * since the last recorded read error.
2173 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2176 unsigned long hours_since_last
;
2177 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2179 cur_time_mon
= ktime_get_seconds();
2181 if (rdev
->last_read_error
== 0) {
2182 /* first time we've seen a read error */
2183 rdev
->last_read_error
= cur_time_mon
;
2187 hours_since_last
= (long)(cur_time_mon
-
2188 rdev
->last_read_error
) / 3600;
2190 rdev
->last_read_error
= cur_time_mon
;
2193 * if hours_since_last is > the number of bits in read_errors
2194 * just set read errors to 0. We do this to avoid
2195 * overflowing the shift of read_errors by hours_since_last.
2197 if (hours_since_last
>= 8 * sizeof(read_errors
))
2198 atomic_set(&rdev
->read_errors
, 0);
2200 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2203 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2204 int sectors
, struct page
*page
, int rw
)
2209 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2210 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2212 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, 0, false))
2216 set_bit(WriteErrorSeen
, &rdev
->flags
);
2217 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2218 set_bit(MD_RECOVERY_NEEDED
,
2219 &rdev
->mddev
->recovery
);
2221 /* need to record an error - either for the block or the device */
2222 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2223 md_error(rdev
->mddev
, rdev
);
2228 * This is a kernel thread which:
2230 * 1. Retries failed read operations on working mirrors.
2231 * 2. Updates the raid superblock when problems encounter.
2232 * 3. Performs writes following reads for array synchronising.
2235 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2237 int sect
= 0; /* Offset from r10_bio->sector */
2238 int sectors
= r10_bio
->sectors
;
2239 struct md_rdev
*rdev
;
2240 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2241 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2243 /* still own a reference to this rdev, so it cannot
2244 * have been cleared recently.
2246 rdev
= conf
->mirrors
[d
].rdev
;
2248 if (test_bit(Faulty
, &rdev
->flags
))
2249 /* drive has already been failed, just ignore any
2250 more fix_read_error() attempts */
2253 check_decay_read_errors(mddev
, rdev
);
2254 atomic_inc(&rdev
->read_errors
);
2255 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2256 char b
[BDEVNAME_SIZE
];
2257 bdevname(rdev
->bdev
, b
);
2260 "md/raid10:%s: %s: Raid device exceeded "
2261 "read_error threshold [cur %d:max %d]\n",
2263 atomic_read(&rdev
->read_errors
), max_read_errors
);
2265 "md/raid10:%s: %s: Failing raid device\n",
2267 md_error(mddev
, rdev
);
2268 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2274 int sl
= r10_bio
->read_slot
;
2278 if (s
> (PAGE_SIZE
>>9))
2286 d
= r10_bio
->devs
[sl
].devnum
;
2287 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2289 test_bit(In_sync
, &rdev
->flags
) &&
2290 !test_bit(Faulty
, &rdev
->flags
) &&
2291 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2292 &first_bad
, &bad_sectors
) == 0) {
2293 atomic_inc(&rdev
->nr_pending
);
2295 success
= sync_page_io(rdev
,
2296 r10_bio
->devs
[sl
].addr
+
2300 REQ_OP_READ
, 0, false);
2301 rdev_dec_pending(rdev
, mddev
);
2307 if (sl
== conf
->copies
)
2309 } while (!success
&& sl
!= r10_bio
->read_slot
);
2313 /* Cannot read from anywhere, just mark the block
2314 * as bad on the first device to discourage future
2317 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2318 rdev
= conf
->mirrors
[dn
].rdev
;
2320 if (!rdev_set_badblocks(
2322 r10_bio
->devs
[r10_bio
->read_slot
].addr
2325 md_error(mddev
, rdev
);
2326 r10_bio
->devs
[r10_bio
->read_slot
].bio
2333 /* write it back and re-read */
2335 while (sl
!= r10_bio
->read_slot
) {
2336 char b
[BDEVNAME_SIZE
];
2341 d
= r10_bio
->devs
[sl
].devnum
;
2342 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2344 test_bit(Faulty
, &rdev
->flags
) ||
2345 !test_bit(In_sync
, &rdev
->flags
))
2348 atomic_inc(&rdev
->nr_pending
);
2350 if (r10_sync_page_io(rdev
,
2351 r10_bio
->devs
[sl
].addr
+
2353 s
, conf
->tmppage
, WRITE
)
2355 /* Well, this device is dead */
2357 "md/raid10:%s: read correction "
2359 " (%d sectors at %llu on %s)\n",
2361 (unsigned long long)(
2363 choose_data_offset(r10_bio
,
2365 bdevname(rdev
->bdev
, b
));
2366 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2369 bdevname(rdev
->bdev
, b
));
2371 rdev_dec_pending(rdev
, mddev
);
2375 while (sl
!= r10_bio
->read_slot
) {
2376 char b
[BDEVNAME_SIZE
];
2381 d
= r10_bio
->devs
[sl
].devnum
;
2382 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2384 test_bit(Faulty
, &rdev
->flags
) ||
2385 !test_bit(In_sync
, &rdev
->flags
))
2388 atomic_inc(&rdev
->nr_pending
);
2390 switch (r10_sync_page_io(rdev
,
2391 r10_bio
->devs
[sl
].addr
+
2396 /* Well, this device is dead */
2398 "md/raid10:%s: unable to read back "
2400 " (%d sectors at %llu on %s)\n",
2402 (unsigned long long)(
2404 choose_data_offset(r10_bio
, rdev
)),
2405 bdevname(rdev
->bdev
, b
));
2406 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2409 bdevname(rdev
->bdev
, b
));
2413 "md/raid10:%s: read error corrected"
2414 " (%d sectors at %llu on %s)\n",
2416 (unsigned long long)(
2418 choose_data_offset(r10_bio
, rdev
)),
2419 bdevname(rdev
->bdev
, b
));
2420 atomic_add(s
, &rdev
->corrected_errors
);
2423 rdev_dec_pending(rdev
, mddev
);
2433 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2435 struct bio
*bio
= r10_bio
->master_bio
;
2436 struct mddev
*mddev
= r10_bio
->mddev
;
2437 struct r10conf
*conf
= mddev
->private;
2438 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2439 /* bio has the data to be written to slot 'i' where
2440 * we just recently had a write error.
2441 * We repeatedly clone the bio and trim down to one block,
2442 * then try the write. Where the write fails we record
2444 * It is conceivable that the bio doesn't exactly align with
2445 * blocks. We must handle this.
2447 * We currently own a reference to the rdev.
2453 int sect_to_write
= r10_bio
->sectors
;
2456 if (rdev
->badblocks
.shift
< 0)
2459 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2460 bdev_logical_block_size(rdev
->bdev
) >> 9);
2461 sector
= r10_bio
->sector
;
2462 sectors
= ((r10_bio
->sector
+ block_sectors
)
2463 & ~(sector_t
)(block_sectors
- 1))
2466 while (sect_to_write
) {
2468 if (sectors
> sect_to_write
)
2469 sectors
= sect_to_write
;
2470 /* Write at 'sector' for 'sectors' */
2471 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2472 bio_trim(wbio
, sector
- bio
->bi_iter
.bi_sector
, sectors
);
2473 wbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
2474 choose_data_offset(r10_bio
, rdev
) +
2475 (sector
- r10_bio
->sector
));
2476 wbio
->bi_bdev
= rdev
->bdev
;
2477 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2479 if (submit_bio_wait(wbio
) < 0)
2481 ok
= rdev_set_badblocks(rdev
, sector
,
2486 sect_to_write
-= sectors
;
2488 sectors
= block_sectors
;
2493 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2495 int slot
= r10_bio
->read_slot
;
2497 struct r10conf
*conf
= mddev
->private;
2498 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2499 char b
[BDEVNAME_SIZE
];
2500 unsigned long do_sync
;
2503 /* we got a read error. Maybe the drive is bad. Maybe just
2504 * the block and we can fix it.
2505 * We freeze all other IO, and try reading the block from
2506 * other devices. When we find one, we re-write
2507 * and check it that fixes the read error.
2508 * This is all done synchronously while the array is
2511 bio
= r10_bio
->devs
[slot
].bio
;
2512 bdevname(bio
->bi_bdev
, b
);
2514 r10_bio
->devs
[slot
].bio
= NULL
;
2516 if (mddev
->ro
== 0) {
2517 freeze_array(conf
, 1);
2518 fix_read_error(conf
, mddev
, r10_bio
);
2519 unfreeze_array(conf
);
2521 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2523 rdev_dec_pending(rdev
, mddev
);
2526 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2528 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2529 " read error for block %llu\n",
2531 (unsigned long long)r10_bio
->sector
);
2532 raid_end_bio_io(r10_bio
);
2536 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2537 slot
= r10_bio
->read_slot
;
2540 "md/raid10:%s: %s: redirecting "
2541 "sector %llu to another mirror\n",
2543 bdevname(rdev
->bdev
, b
),
2544 (unsigned long long)r10_bio
->sector
);
2545 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2547 bio_trim(bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
2548 r10_bio
->devs
[slot
].bio
= bio
;
2549 r10_bio
->devs
[slot
].rdev
= rdev
;
2550 bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
2551 + choose_data_offset(r10_bio
, rdev
);
2552 bio
->bi_bdev
= rdev
->bdev
;
2553 bio_set_op_attrs(bio
, REQ_OP_READ
, do_sync
);
2554 bio
->bi_private
= r10_bio
;
2555 bio
->bi_end_io
= raid10_end_read_request
;
2556 if (max_sectors
< r10_bio
->sectors
) {
2557 /* Drat - have to split this up more */
2558 struct bio
*mbio
= r10_bio
->master_bio
;
2559 int sectors_handled
=
2560 r10_bio
->sector
+ max_sectors
2561 - mbio
->bi_iter
.bi_sector
;
2562 r10_bio
->sectors
= max_sectors
;
2563 spin_lock_irq(&conf
->device_lock
);
2564 if (mbio
->bi_phys_segments
== 0)
2565 mbio
->bi_phys_segments
= 2;
2567 mbio
->bi_phys_segments
++;
2568 spin_unlock_irq(&conf
->device_lock
);
2569 generic_make_request(bio
);
2571 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2573 r10_bio
->master_bio
= mbio
;
2574 r10_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2576 set_bit(R10BIO_ReadError
,
2578 r10_bio
->mddev
= mddev
;
2579 r10_bio
->sector
= mbio
->bi_iter
.bi_sector
2584 generic_make_request(bio
);
2587 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2589 /* Some sort of write request has finished and it
2590 * succeeded in writing where we thought there was a
2591 * bad block. So forget the bad block.
2592 * Or possibly if failed and we need to record
2596 struct md_rdev
*rdev
;
2598 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2599 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2600 for (m
= 0; m
< conf
->copies
; m
++) {
2601 int dev
= r10_bio
->devs
[m
].devnum
;
2602 rdev
= conf
->mirrors
[dev
].rdev
;
2603 if (r10_bio
->devs
[m
].bio
== NULL
)
2605 if (!r10_bio
->devs
[m
].bio
->bi_error
) {
2606 rdev_clear_badblocks(
2608 r10_bio
->devs
[m
].addr
,
2609 r10_bio
->sectors
, 0);
2611 if (!rdev_set_badblocks(
2613 r10_bio
->devs
[m
].addr
,
2614 r10_bio
->sectors
, 0))
2615 md_error(conf
->mddev
, rdev
);
2617 rdev
= conf
->mirrors
[dev
].replacement
;
2618 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2621 if (!r10_bio
->devs
[m
].repl_bio
->bi_error
) {
2622 rdev_clear_badblocks(
2624 r10_bio
->devs
[m
].addr
,
2625 r10_bio
->sectors
, 0);
2627 if (!rdev_set_badblocks(
2629 r10_bio
->devs
[m
].addr
,
2630 r10_bio
->sectors
, 0))
2631 md_error(conf
->mddev
, rdev
);
2637 for (m
= 0; m
< conf
->copies
; m
++) {
2638 int dev
= r10_bio
->devs
[m
].devnum
;
2639 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2640 rdev
= conf
->mirrors
[dev
].rdev
;
2641 if (bio
== IO_MADE_GOOD
) {
2642 rdev_clear_badblocks(
2644 r10_bio
->devs
[m
].addr
,
2645 r10_bio
->sectors
, 0);
2646 rdev_dec_pending(rdev
, conf
->mddev
);
2647 } else if (bio
!= NULL
&& bio
->bi_error
) {
2649 if (!narrow_write_error(r10_bio
, m
)) {
2650 md_error(conf
->mddev
, rdev
);
2651 set_bit(R10BIO_Degraded
,
2654 rdev_dec_pending(rdev
, conf
->mddev
);
2656 bio
= r10_bio
->devs
[m
].repl_bio
;
2657 rdev
= conf
->mirrors
[dev
].replacement
;
2658 if (rdev
&& bio
== IO_MADE_GOOD
) {
2659 rdev_clear_badblocks(
2661 r10_bio
->devs
[m
].addr
,
2662 r10_bio
->sectors
, 0);
2663 rdev_dec_pending(rdev
, conf
->mddev
);
2667 spin_lock_irq(&conf
->device_lock
);
2668 list_add(&r10_bio
->retry_list
, &conf
->bio_end_io_list
);
2670 spin_unlock_irq(&conf
->device_lock
);
2671 md_wakeup_thread(conf
->mddev
->thread
);
2673 if (test_bit(R10BIO_WriteError
,
2675 close_write(r10_bio
);
2676 raid_end_bio_io(r10_bio
);
2681 static void raid10d(struct md_thread
*thread
)
2683 struct mddev
*mddev
= thread
->mddev
;
2684 struct r10bio
*r10_bio
;
2685 unsigned long flags
;
2686 struct r10conf
*conf
= mddev
->private;
2687 struct list_head
*head
= &conf
->retry_list
;
2688 struct blk_plug plug
;
2690 md_check_recovery(mddev
);
2692 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2693 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2695 spin_lock_irqsave(&conf
->device_lock
, flags
);
2696 if (!test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2697 while (!list_empty(&conf
->bio_end_io_list
)) {
2698 list_move(conf
->bio_end_io_list
.prev
, &tmp
);
2702 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2703 while (!list_empty(&tmp
)) {
2704 r10_bio
= list_first_entry(&tmp
, struct r10bio
,
2706 list_del(&r10_bio
->retry_list
);
2707 if (mddev
->degraded
)
2708 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
2710 if (test_bit(R10BIO_WriteError
,
2712 close_write(r10_bio
);
2713 raid_end_bio_io(r10_bio
);
2717 blk_start_plug(&plug
);
2720 flush_pending_writes(conf
);
2722 spin_lock_irqsave(&conf
->device_lock
, flags
);
2723 if (list_empty(head
)) {
2724 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2727 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2728 list_del(head
->prev
);
2730 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2732 mddev
= r10_bio
->mddev
;
2733 conf
= mddev
->private;
2734 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2735 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2736 handle_write_completed(conf
, r10_bio
);
2737 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2738 reshape_request_write(mddev
, r10_bio
);
2739 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2740 sync_request_write(mddev
, r10_bio
);
2741 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2742 recovery_request_write(mddev
, r10_bio
);
2743 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2744 handle_read_error(mddev
, r10_bio
);
2746 /* just a partial read to be scheduled from a
2749 int slot
= r10_bio
->read_slot
;
2750 generic_make_request(r10_bio
->devs
[slot
].bio
);
2754 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2755 md_check_recovery(mddev
);
2757 blk_finish_plug(&plug
);
2760 static int init_resync(struct r10conf
*conf
)
2765 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2766 BUG_ON(conf
->r10buf_pool
);
2767 conf
->have_replacement
= 0;
2768 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2769 if (conf
->mirrors
[i
].replacement
)
2770 conf
->have_replacement
= 1;
2771 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2772 if (!conf
->r10buf_pool
)
2774 conf
->next_resync
= 0;
2779 * perform a "sync" on one "block"
2781 * We need to make sure that no normal I/O request - particularly write
2782 * requests - conflict with active sync requests.
2784 * This is achieved by tracking pending requests and a 'barrier' concept
2785 * that can be installed to exclude normal IO requests.
2787 * Resync and recovery are handled very differently.
2788 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2790 * For resync, we iterate over virtual addresses, read all copies,
2791 * and update if there are differences. If only one copy is live,
2793 * For recovery, we iterate over physical addresses, read a good
2794 * value for each non-in_sync drive, and over-write.
2796 * So, for recovery we may have several outstanding complex requests for a
2797 * given address, one for each out-of-sync device. We model this by allocating
2798 * a number of r10_bio structures, one for each out-of-sync device.
2799 * As we setup these structures, we collect all bio's together into a list
2800 * which we then process collectively to add pages, and then process again
2801 * to pass to generic_make_request.
2803 * The r10_bio structures are linked using a borrowed master_bio pointer.
2804 * This link is counted in ->remaining. When the r10_bio that points to NULL
2805 * has its remaining count decremented to 0, the whole complex operation
2810 static sector_t
raid10_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2813 struct r10conf
*conf
= mddev
->private;
2814 struct r10bio
*r10_bio
;
2815 struct bio
*biolist
= NULL
, *bio
;
2816 sector_t max_sector
, nr_sectors
;
2819 sector_t sync_blocks
;
2820 sector_t sectors_skipped
= 0;
2821 int chunks_skipped
= 0;
2822 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2824 if (!conf
->r10buf_pool
)
2825 if (init_resync(conf
))
2829 * Allow skipping a full rebuild for incremental assembly
2830 * of a clean array, like RAID1 does.
2832 if (mddev
->bitmap
== NULL
&&
2833 mddev
->recovery_cp
== MaxSector
&&
2834 mddev
->reshape_position
== MaxSector
&&
2835 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2836 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2837 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2838 conf
->fullsync
== 0) {
2840 return mddev
->dev_sectors
- sector_nr
;
2844 max_sector
= mddev
->dev_sectors
;
2845 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2846 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2847 max_sector
= mddev
->resync_max_sectors
;
2848 if (sector_nr
>= max_sector
) {
2849 /* If we aborted, we need to abort the
2850 * sync on the 'current' bitmap chucks (there can
2851 * be several when recovering multiple devices).
2852 * as we may have started syncing it but not finished.
2853 * We can find the current address in
2854 * mddev->curr_resync, but for recovery,
2855 * we need to convert that to several
2856 * virtual addresses.
2858 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2864 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2865 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2866 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2868 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2870 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2871 bitmap_end_sync(mddev
->bitmap
, sect
,
2875 /* completed sync */
2876 if ((!mddev
->bitmap
|| conf
->fullsync
)
2877 && conf
->have_replacement
2878 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2879 /* Completed a full sync so the replacements
2880 * are now fully recovered.
2883 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2884 struct md_rdev
*rdev
=
2885 rcu_dereference(conf
->mirrors
[i
].replacement
);
2887 rdev
->recovery_offset
= MaxSector
;
2893 bitmap_close_sync(mddev
->bitmap
);
2896 return sectors_skipped
;
2899 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2900 return reshape_request(mddev
, sector_nr
, skipped
);
2902 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2903 /* if there has been nothing to do on any drive,
2904 * then there is nothing to do at all..
2907 return (max_sector
- sector_nr
) + sectors_skipped
;
2910 if (max_sector
> mddev
->resync_max
)
2911 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2913 /* make sure whole request will fit in a chunk - if chunks
2916 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2917 max_sector
> (sector_nr
| chunk_mask
))
2918 max_sector
= (sector_nr
| chunk_mask
) + 1;
2921 * If there is non-resync activity waiting for a turn, then let it
2922 * though before starting on this new sync request.
2924 if (conf
->nr_waiting
)
2925 schedule_timeout_uninterruptible(1);
2927 /* Again, very different code for resync and recovery.
2928 * Both must result in an r10bio with a list of bios that
2929 * have bi_end_io, bi_sector, bi_bdev set,
2930 * and bi_private set to the r10bio.
2931 * For recovery, we may actually create several r10bios
2932 * with 2 bios in each, that correspond to the bios in the main one.
2933 * In this case, the subordinate r10bios link back through a
2934 * borrowed master_bio pointer, and the counter in the master
2935 * includes a ref from each subordinate.
2937 /* First, we decide what to do and set ->bi_end_io
2938 * To end_sync_read if we want to read, and
2939 * end_sync_write if we will want to write.
2942 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2943 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2944 /* recovery... the complicated one */
2948 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2954 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2955 struct md_rdev
*mrdev
, *mreplace
;
2958 mrdev
= rcu_dereference(mirror
->rdev
);
2959 mreplace
= rcu_dereference(mirror
->replacement
);
2961 if ((mrdev
== NULL
||
2962 test_bit(Faulty
, &mrdev
->flags
) ||
2963 test_bit(In_sync
, &mrdev
->flags
)) &&
2964 (mreplace
== NULL
||
2965 test_bit(Faulty
, &mreplace
->flags
))) {
2971 /* want to reconstruct this device */
2973 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2974 if (sect
>= mddev
->resync_max_sectors
) {
2975 /* last stripe is not complete - don't
2976 * try to recover this sector.
2981 if (mreplace
&& test_bit(Faulty
, &mreplace
->flags
))
2983 /* Unless we are doing a full sync, or a replacement
2984 * we only need to recover the block if it is set in
2987 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2989 if (sync_blocks
< max_sync
)
2990 max_sync
= sync_blocks
;
2994 /* yep, skip the sync_blocks here, but don't assume
2995 * that there will never be anything to do here
2997 chunks_skipped
= -1;
3001 atomic_inc(&mrdev
->nr_pending
);
3003 atomic_inc(&mreplace
->nr_pending
);
3006 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3008 raise_barrier(conf
, rb2
!= NULL
);
3009 atomic_set(&r10_bio
->remaining
, 0);
3011 r10_bio
->master_bio
= (struct bio
*)rb2
;
3013 atomic_inc(&rb2
->remaining
);
3014 r10_bio
->mddev
= mddev
;
3015 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3016 r10_bio
->sector
= sect
;
3018 raid10_find_phys(conf
, r10_bio
);
3020 /* Need to check if the array will still be
3024 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++) {
3025 struct md_rdev
*rdev
= rcu_dereference(
3026 conf
->mirrors
[j
].rdev
);
3027 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3033 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3034 &sync_blocks
, still_degraded
);
3037 for (j
=0; j
<conf
->copies
;j
++) {
3039 int d
= r10_bio
->devs
[j
].devnum
;
3040 sector_t from_addr
, to_addr
;
3041 struct md_rdev
*rdev
=
3042 rcu_dereference(conf
->mirrors
[d
].rdev
);
3043 sector_t sector
, first_bad
;
3046 !test_bit(In_sync
, &rdev
->flags
))
3048 /* This is where we read from */
3050 sector
= r10_bio
->devs
[j
].addr
;
3052 if (is_badblock(rdev
, sector
, max_sync
,
3053 &first_bad
, &bad_sectors
)) {
3054 if (first_bad
> sector
)
3055 max_sync
= first_bad
- sector
;
3057 bad_sectors
-= (sector
3059 if (max_sync
> bad_sectors
)
3060 max_sync
= bad_sectors
;
3064 bio
= r10_bio
->devs
[0].bio
;
3066 bio
->bi_next
= biolist
;
3068 bio
->bi_private
= r10_bio
;
3069 bio
->bi_end_io
= end_sync_read
;
3070 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3071 from_addr
= r10_bio
->devs
[j
].addr
;
3072 bio
->bi_iter
.bi_sector
= from_addr
+
3074 bio
->bi_bdev
= rdev
->bdev
;
3075 atomic_inc(&rdev
->nr_pending
);
3076 /* and we write to 'i' (if not in_sync) */
3078 for (k
=0; k
<conf
->copies
; k
++)
3079 if (r10_bio
->devs
[k
].devnum
== i
)
3081 BUG_ON(k
== conf
->copies
);
3082 to_addr
= r10_bio
->devs
[k
].addr
;
3083 r10_bio
->devs
[0].devnum
= d
;
3084 r10_bio
->devs
[0].addr
= from_addr
;
3085 r10_bio
->devs
[1].devnum
= i
;
3086 r10_bio
->devs
[1].addr
= to_addr
;
3088 if (!test_bit(In_sync
, &mrdev
->flags
)) {
3089 bio
= r10_bio
->devs
[1].bio
;
3091 bio
->bi_next
= biolist
;
3093 bio
->bi_private
= r10_bio
;
3094 bio
->bi_end_io
= end_sync_write
;
3095 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3096 bio
->bi_iter
.bi_sector
= to_addr
3097 + mrdev
->data_offset
;
3098 bio
->bi_bdev
= mrdev
->bdev
;
3099 atomic_inc(&r10_bio
->remaining
);
3101 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3103 /* and maybe write to replacement */
3104 bio
= r10_bio
->devs
[1].repl_bio
;
3106 bio
->bi_end_io
= NULL
;
3107 /* Note: if mreplace != NULL, then bio
3108 * cannot be NULL as r10buf_pool_alloc will
3109 * have allocated it.
3110 * So the second test here is pointless.
3111 * But it keeps semantic-checkers happy, and
3112 * this comment keeps human reviewers
3115 if (mreplace
== NULL
|| bio
== NULL
||
3116 test_bit(Faulty
, &mreplace
->flags
))
3119 bio
->bi_next
= biolist
;
3121 bio
->bi_private
= r10_bio
;
3122 bio
->bi_end_io
= end_sync_write
;
3123 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3124 bio
->bi_iter
.bi_sector
= to_addr
+
3125 mreplace
->data_offset
;
3126 bio
->bi_bdev
= mreplace
->bdev
;
3127 atomic_inc(&r10_bio
->remaining
);
3131 if (j
== conf
->copies
) {
3132 /* Cannot recover, so abort the recovery or
3133 * record a bad block */
3135 /* problem is that there are bad blocks
3136 * on other device(s)
3139 for (k
= 0; k
< conf
->copies
; k
++)
3140 if (r10_bio
->devs
[k
].devnum
== i
)
3142 if (!test_bit(In_sync
,
3144 && !rdev_set_badblocks(
3146 r10_bio
->devs
[k
].addr
,
3150 !rdev_set_badblocks(
3152 r10_bio
->devs
[k
].addr
,
3157 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3159 printk(KERN_INFO
"md/raid10:%s: insufficient "
3160 "working devices for recovery.\n",
3162 mirror
->recovery_disabled
3163 = mddev
->recovery_disabled
;
3167 atomic_dec(&rb2
->remaining
);
3169 rdev_dec_pending(mrdev
, mddev
);
3171 rdev_dec_pending(mreplace
, mddev
);
3174 rdev_dec_pending(mrdev
, mddev
);
3176 rdev_dec_pending(mreplace
, mddev
);
3178 if (biolist
== NULL
) {
3180 struct r10bio
*rb2
= r10_bio
;
3181 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3182 rb2
->master_bio
= NULL
;
3188 /* resync. Schedule a read for every block at this virt offset */
3191 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, 0);
3193 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3194 &sync_blocks
, mddev
->degraded
) &&
3195 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3196 &mddev
->recovery
)) {
3197 /* We can skip this block */
3199 return sync_blocks
+ sectors_skipped
;
3201 if (sync_blocks
< max_sync
)
3202 max_sync
= sync_blocks
;
3203 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3206 r10_bio
->mddev
= mddev
;
3207 atomic_set(&r10_bio
->remaining
, 0);
3208 raise_barrier(conf
, 0);
3209 conf
->next_resync
= sector_nr
;
3211 r10_bio
->master_bio
= NULL
;
3212 r10_bio
->sector
= sector_nr
;
3213 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3214 raid10_find_phys(conf
, r10_bio
);
3215 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3217 for (i
= 0; i
< conf
->copies
; i
++) {
3218 int d
= r10_bio
->devs
[i
].devnum
;
3219 sector_t first_bad
, sector
;
3221 struct md_rdev
*rdev
;
3223 if (r10_bio
->devs
[i
].repl_bio
)
3224 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3226 bio
= r10_bio
->devs
[i
].bio
;
3228 bio
->bi_error
= -EIO
;
3230 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
3231 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3235 sector
= r10_bio
->devs
[i
].addr
;
3236 if (is_badblock(rdev
, sector
, max_sync
,
3237 &first_bad
, &bad_sectors
)) {
3238 if (first_bad
> sector
)
3239 max_sync
= first_bad
- sector
;
3241 bad_sectors
-= (sector
- first_bad
);
3242 if (max_sync
> bad_sectors
)
3243 max_sync
= bad_sectors
;
3248 atomic_inc(&rdev
->nr_pending
);
3249 atomic_inc(&r10_bio
->remaining
);
3250 bio
->bi_next
= biolist
;
3252 bio
->bi_private
= r10_bio
;
3253 bio
->bi_end_io
= end_sync_read
;
3254 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3255 bio
->bi_iter
.bi_sector
= sector
+ rdev
->data_offset
;
3256 bio
->bi_bdev
= rdev
->bdev
;
3259 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
3260 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
)) {
3264 atomic_inc(&rdev
->nr_pending
);
3267 /* Need to set up for writing to the replacement */
3268 bio
= r10_bio
->devs
[i
].repl_bio
;
3270 bio
->bi_error
= -EIO
;
3272 sector
= r10_bio
->devs
[i
].addr
;
3273 bio
->bi_next
= biolist
;
3275 bio
->bi_private
= r10_bio
;
3276 bio
->bi_end_io
= end_sync_write
;
3277 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3278 bio
->bi_iter
.bi_sector
= sector
+ rdev
->data_offset
;
3279 bio
->bi_bdev
= rdev
->bdev
;
3284 for (i
=0; i
<conf
->copies
; i
++) {
3285 int d
= r10_bio
->devs
[i
].devnum
;
3286 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3287 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3289 if (r10_bio
->devs
[i
].repl_bio
&&
3290 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3292 conf
->mirrors
[d
].replacement
,
3302 if (sector_nr
+ max_sync
< max_sector
)
3303 max_sector
= sector_nr
+ max_sync
;
3306 int len
= PAGE_SIZE
;
3307 if (sector_nr
+ (len
>>9) > max_sector
)
3308 len
= (max_sector
- sector_nr
) << 9;
3311 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3313 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3314 if (bio_add_page(bio
, page
, len
, 0))
3318 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3319 for (bio2
= biolist
;
3320 bio2
&& bio2
!= bio
;
3321 bio2
= bio2
->bi_next
) {
3322 /* remove last page from this bio */
3324 bio2
->bi_iter
.bi_size
-= len
;
3325 bio_clear_flag(bio2
, BIO_SEG_VALID
);
3329 nr_sectors
+= len
>>9;
3330 sector_nr
+= len
>>9;
3331 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3333 r10_bio
->sectors
= nr_sectors
;
3337 biolist
= biolist
->bi_next
;
3339 bio
->bi_next
= NULL
;
3340 r10_bio
= bio
->bi_private
;
3341 r10_bio
->sectors
= nr_sectors
;
3343 if (bio
->bi_end_io
== end_sync_read
) {
3344 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3346 generic_make_request(bio
);
3350 if (sectors_skipped
)
3351 /* pretend they weren't skipped, it makes
3352 * no important difference in this case
3354 md_done_sync(mddev
, sectors_skipped
, 1);
3356 return sectors_skipped
+ nr_sectors
;
3358 /* There is nowhere to write, so all non-sync
3359 * drives must be failed or in resync, all drives
3360 * have a bad block, so try the next chunk...
3362 if (sector_nr
+ max_sync
< max_sector
)
3363 max_sector
= sector_nr
+ max_sync
;
3365 sectors_skipped
+= (max_sector
- sector_nr
);
3367 sector_nr
= max_sector
;
3372 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3375 struct r10conf
*conf
= mddev
->private;
3378 raid_disks
= min(conf
->geo
.raid_disks
,
3379 conf
->prev
.raid_disks
);
3381 sectors
= conf
->dev_sectors
;
3383 size
= sectors
>> conf
->geo
.chunk_shift
;
3384 sector_div(size
, conf
->geo
.far_copies
);
3385 size
= size
* raid_disks
;
3386 sector_div(size
, conf
->geo
.near_copies
);
3388 return size
<< conf
->geo
.chunk_shift
;
3391 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3393 /* Calculate the number of sectors-per-device that will
3394 * actually be used, and set conf->dev_sectors and
3398 size
= size
>> conf
->geo
.chunk_shift
;
3399 sector_div(size
, conf
->geo
.far_copies
);
3400 size
= size
* conf
->geo
.raid_disks
;
3401 sector_div(size
, conf
->geo
.near_copies
);
3402 /* 'size' is now the number of chunks in the array */
3403 /* calculate "used chunks per device" */
3404 size
= size
* conf
->copies
;
3406 /* We need to round up when dividing by raid_disks to
3407 * get the stride size.
3409 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3411 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3413 if (conf
->geo
.far_offset
)
3414 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3416 sector_div(size
, conf
->geo
.far_copies
);
3417 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3421 enum geo_type
{geo_new
, geo_old
, geo_start
};
3422 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3425 int layout
, chunk
, disks
;
3428 layout
= mddev
->layout
;
3429 chunk
= mddev
->chunk_sectors
;
3430 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3433 layout
= mddev
->new_layout
;
3434 chunk
= mddev
->new_chunk_sectors
;
3435 disks
= mddev
->raid_disks
;
3437 default: /* avoid 'may be unused' warnings */
3438 case geo_start
: /* new when starting reshape - raid_disks not
3440 layout
= mddev
->new_layout
;
3441 chunk
= mddev
->new_chunk_sectors
;
3442 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3447 if (chunk
< (PAGE_SIZE
>> 9) ||
3448 !is_power_of_2(chunk
))
3451 fc
= (layout
>> 8) & 255;
3452 fo
= layout
& (1<<16);
3453 geo
->raid_disks
= disks
;
3454 geo
->near_copies
= nc
;
3455 geo
->far_copies
= fc
;
3456 geo
->far_offset
= fo
;
3457 switch (layout
>> 17) {
3458 case 0: /* original layout. simple but not always optimal */
3459 geo
->far_set_size
= disks
;
3461 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3462 * actually using this, but leave code here just in case.*/
3463 geo
->far_set_size
= disks
/fc
;
3464 WARN(geo
->far_set_size
< fc
,
3465 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3467 case 2: /* "improved" layout fixed to match documentation */
3468 geo
->far_set_size
= fc
* nc
;
3470 default: /* Not a valid layout */
3473 geo
->chunk_mask
= chunk
- 1;
3474 geo
->chunk_shift
= ffz(~chunk
);
3478 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3480 struct r10conf
*conf
= NULL
;
3485 copies
= setup_geo(&geo
, mddev
, geo_new
);
3488 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3489 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3490 mdname(mddev
), PAGE_SIZE
);
3494 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3495 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3496 mdname(mddev
), mddev
->new_layout
);
3501 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3505 /* FIXME calc properly */
3506 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3507 max(0,-mddev
->delta_disks
)),
3512 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3517 conf
->copies
= copies
;
3518 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3519 r10bio_pool_free
, conf
);
3520 if (!conf
->r10bio_pool
)
3523 calc_sectors(conf
, mddev
->dev_sectors
);
3524 if (mddev
->reshape_position
== MaxSector
) {
3525 conf
->prev
= conf
->geo
;
3526 conf
->reshape_progress
= MaxSector
;
3528 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3532 conf
->reshape_progress
= mddev
->reshape_position
;
3533 if (conf
->prev
.far_offset
)
3534 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3536 /* far_copies must be 1 */
3537 conf
->prev
.stride
= conf
->dev_sectors
;
3539 conf
->reshape_safe
= conf
->reshape_progress
;
3540 spin_lock_init(&conf
->device_lock
);
3541 INIT_LIST_HEAD(&conf
->retry_list
);
3542 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
3544 spin_lock_init(&conf
->resync_lock
);
3545 init_waitqueue_head(&conf
->wait_barrier
);
3546 atomic_set(&conf
->nr_pending
, 0);
3548 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3552 conf
->mddev
= mddev
;
3557 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3560 mempool_destroy(conf
->r10bio_pool
);
3561 kfree(conf
->mirrors
);
3562 safe_put_page(conf
->tmppage
);
3565 return ERR_PTR(err
);
3568 static int raid10_run(struct mddev
*mddev
)
3570 struct r10conf
*conf
;
3571 int i
, disk_idx
, chunk_size
;
3572 struct raid10_info
*disk
;
3573 struct md_rdev
*rdev
;
3575 sector_t min_offset_diff
= 0;
3577 bool discard_supported
= false;
3579 if (mddev
->private == NULL
) {
3580 conf
= setup_conf(mddev
);
3582 return PTR_ERR(conf
);
3583 mddev
->private = conf
;
3585 conf
= mddev
->private;
3589 mddev
->thread
= conf
->thread
;
3590 conf
->thread
= NULL
;
3592 chunk_size
= mddev
->chunk_sectors
<< 9;
3594 blk_queue_max_discard_sectors(mddev
->queue
,
3595 mddev
->chunk_sectors
);
3596 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3597 blk_queue_io_min(mddev
->queue
, chunk_size
);
3598 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3599 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3601 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3602 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3605 rdev_for_each(rdev
, mddev
) {
3607 struct request_queue
*q
;
3609 disk_idx
= rdev
->raid_disk
;
3612 if (disk_idx
>= conf
->geo
.raid_disks
&&
3613 disk_idx
>= conf
->prev
.raid_disks
)
3615 disk
= conf
->mirrors
+ disk_idx
;
3617 if (test_bit(Replacement
, &rdev
->flags
)) {
3618 if (disk
->replacement
)
3620 disk
->replacement
= rdev
;
3626 q
= bdev_get_queue(rdev
->bdev
);
3627 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3628 if (!mddev
->reshape_backwards
)
3632 if (first
|| diff
< min_offset_diff
)
3633 min_offset_diff
= diff
;
3636 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3637 rdev
->data_offset
<< 9);
3639 disk
->head_position
= 0;
3641 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3642 discard_supported
= true;
3646 if (discard_supported
)
3647 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3650 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3653 /* need to check that every block has at least one working mirror */
3654 if (!enough(conf
, -1)) {
3655 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3660 if (conf
->reshape_progress
!= MaxSector
) {
3661 /* must ensure that shape change is supported */
3662 if (conf
->geo
.far_copies
!= 1 &&
3663 conf
->geo
.far_offset
== 0)
3665 if (conf
->prev
.far_copies
!= 1 &&
3666 conf
->prev
.far_offset
== 0)
3670 mddev
->degraded
= 0;
3672 i
< conf
->geo
.raid_disks
3673 || i
< conf
->prev
.raid_disks
;
3676 disk
= conf
->mirrors
+ i
;
3678 if (!disk
->rdev
&& disk
->replacement
) {
3679 /* The replacement is all we have - use it */
3680 disk
->rdev
= disk
->replacement
;
3681 disk
->replacement
= NULL
;
3682 clear_bit(Replacement
, &disk
->rdev
->flags
);
3686 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3687 disk
->head_position
= 0;
3690 disk
->rdev
->saved_raid_disk
< 0)
3693 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3696 if (mddev
->recovery_cp
!= MaxSector
)
3697 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3698 " -- starting background reconstruction\n",
3701 "md/raid10:%s: active with %d out of %d devices\n",
3702 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3703 conf
->geo
.raid_disks
);
3705 * Ok, everything is just fine now
3707 mddev
->dev_sectors
= conf
->dev_sectors
;
3708 size
= raid10_size(mddev
, 0, 0);
3709 md_set_array_sectors(mddev
, size
);
3710 mddev
->resync_max_sectors
= size
;
3713 int stripe
= conf
->geo
.raid_disks
*
3714 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3716 /* Calculate max read-ahead size.
3717 * We need to readahead at least twice a whole stripe....
3720 stripe
/= conf
->geo
.near_copies
;
3721 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3722 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3725 if (md_integrity_register(mddev
))
3728 if (conf
->reshape_progress
!= MaxSector
) {
3729 unsigned long before_length
, after_length
;
3731 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3732 conf
->prev
.far_copies
);
3733 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3734 conf
->geo
.far_copies
);
3736 if (max(before_length
, after_length
) > min_offset_diff
) {
3737 /* This cannot work */
3738 printk("md/raid10: offset difference not enough to continue reshape\n");
3741 conf
->offset_diff
= min_offset_diff
;
3743 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3744 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3745 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3746 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3747 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3754 md_unregister_thread(&mddev
->thread
);
3755 mempool_destroy(conf
->r10bio_pool
);
3756 safe_put_page(conf
->tmppage
);
3757 kfree(conf
->mirrors
);
3759 mddev
->private = NULL
;
3764 static void raid10_free(struct mddev
*mddev
, void *priv
)
3766 struct r10conf
*conf
= priv
;
3768 mempool_destroy(conf
->r10bio_pool
);
3769 safe_put_page(conf
->tmppage
);
3770 kfree(conf
->mirrors
);
3771 kfree(conf
->mirrors_old
);
3772 kfree(conf
->mirrors_new
);
3776 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3778 struct r10conf
*conf
= mddev
->private;
3782 raise_barrier(conf
, 0);
3785 lower_barrier(conf
);
3790 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3792 /* Resize of 'far' arrays is not supported.
3793 * For 'near' and 'offset' arrays we can set the
3794 * number of sectors used to be an appropriate multiple
3795 * of the chunk size.
3796 * For 'offset', this is far_copies*chunksize.
3797 * For 'near' the multiplier is the LCM of
3798 * near_copies and raid_disks.
3799 * So if far_copies > 1 && !far_offset, fail.
3800 * Else find LCM(raid_disks, near_copy)*far_copies and
3801 * multiply by chunk_size. Then round to this number.
3802 * This is mostly done by raid10_size()
3804 struct r10conf
*conf
= mddev
->private;
3805 sector_t oldsize
, size
;
3807 if (mddev
->reshape_position
!= MaxSector
)
3810 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3813 oldsize
= raid10_size(mddev
, 0, 0);
3814 size
= raid10_size(mddev
, sectors
, 0);
3815 if (mddev
->external_size
&&
3816 mddev
->array_sectors
> size
)
3818 if (mddev
->bitmap
) {
3819 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3823 md_set_array_sectors(mddev
, size
);
3825 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3826 revalidate_disk(mddev
->gendisk
);
3828 if (sectors
> mddev
->dev_sectors
&&
3829 mddev
->recovery_cp
> oldsize
) {
3830 mddev
->recovery_cp
= oldsize
;
3831 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3833 calc_sectors(conf
, sectors
);
3834 mddev
->dev_sectors
= conf
->dev_sectors
;
3835 mddev
->resync_max_sectors
= size
;
3839 static void *raid10_takeover_raid0(struct mddev
*mddev
, sector_t size
, int devs
)
3841 struct md_rdev
*rdev
;
3842 struct r10conf
*conf
;
3844 if (mddev
->degraded
> 0) {
3845 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3847 return ERR_PTR(-EINVAL
);
3849 sector_div(size
, devs
);
3851 /* Set new parameters */
3852 mddev
->new_level
= 10;
3853 /* new layout: far_copies = 1, near_copies = 2 */
3854 mddev
->new_layout
= (1<<8) + 2;
3855 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3856 mddev
->delta_disks
= mddev
->raid_disks
;
3857 mddev
->raid_disks
*= 2;
3858 /* make sure it will be not marked as dirty */
3859 mddev
->recovery_cp
= MaxSector
;
3860 mddev
->dev_sectors
= size
;
3862 conf
= setup_conf(mddev
);
3863 if (!IS_ERR(conf
)) {
3864 rdev_for_each(rdev
, mddev
)
3865 if (rdev
->raid_disk
>= 0) {
3866 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3867 rdev
->sectors
= size
;
3875 static void *raid10_takeover(struct mddev
*mddev
)
3877 struct r0conf
*raid0_conf
;
3879 /* raid10 can take over:
3880 * raid0 - providing it has only two drives
3882 if (mddev
->level
== 0) {
3883 /* for raid0 takeover only one zone is supported */
3884 raid0_conf
= mddev
->private;
3885 if (raid0_conf
->nr_strip_zones
> 1) {
3886 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3887 " with more than one zone.\n",
3889 return ERR_PTR(-EINVAL
);
3891 return raid10_takeover_raid0(mddev
,
3892 raid0_conf
->strip_zone
->zone_end
,
3893 raid0_conf
->strip_zone
->nb_dev
);
3895 return ERR_PTR(-EINVAL
);
3898 static int raid10_check_reshape(struct mddev
*mddev
)
3900 /* Called when there is a request to change
3901 * - layout (to ->new_layout)
3902 * - chunk size (to ->new_chunk_sectors)
3903 * - raid_disks (by delta_disks)
3904 * or when trying to restart a reshape that was ongoing.
3906 * We need to validate the request and possibly allocate
3907 * space if that might be an issue later.
3909 * Currently we reject any reshape of a 'far' mode array,
3910 * allow chunk size to change if new is generally acceptable,
3911 * allow raid_disks to increase, and allow
3912 * a switch between 'near' mode and 'offset' mode.
3914 struct r10conf
*conf
= mddev
->private;
3917 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3920 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3921 /* mustn't change number of copies */
3923 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3924 /* Cannot switch to 'far' mode */
3927 if (mddev
->array_sectors
& geo
.chunk_mask
)
3928 /* not factor of array size */
3931 if (!enough(conf
, -1))
3934 kfree(conf
->mirrors_new
);
3935 conf
->mirrors_new
= NULL
;
3936 if (mddev
->delta_disks
> 0) {
3937 /* allocate new 'mirrors' list */
3938 conf
->mirrors_new
= kzalloc(
3939 sizeof(struct raid10_info
)
3940 *(mddev
->raid_disks
+
3941 mddev
->delta_disks
),
3943 if (!conf
->mirrors_new
)
3950 * Need to check if array has failed when deciding whether to:
3952 * - remove non-faulty devices
3955 * This determination is simple when no reshape is happening.
3956 * However if there is a reshape, we need to carefully check
3957 * both the before and after sections.
3958 * This is because some failed devices may only affect one
3959 * of the two sections, and some non-in_sync devices may
3960 * be insync in the section most affected by failed devices.
3962 static int calc_degraded(struct r10conf
*conf
)
3964 int degraded
, degraded2
;
3969 /* 'prev' section first */
3970 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
3971 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3972 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3974 else if (!test_bit(In_sync
, &rdev
->flags
))
3975 /* When we can reduce the number of devices in
3976 * an array, this might not contribute to
3977 * 'degraded'. It does now.
3982 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
3986 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3987 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3988 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3990 else if (!test_bit(In_sync
, &rdev
->flags
)) {
3991 /* If reshape is increasing the number of devices,
3992 * this section has already been recovered, so
3993 * it doesn't contribute to degraded.
3996 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
4001 if (degraded2
> degraded
)
4006 static int raid10_start_reshape(struct mddev
*mddev
)
4008 /* A 'reshape' has been requested. This commits
4009 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4010 * This also checks if there are enough spares and adds them
4012 * We currently require enough spares to make the final
4013 * array non-degraded. We also require that the difference
4014 * between old and new data_offset - on each device - is
4015 * enough that we never risk over-writing.
4018 unsigned long before_length
, after_length
;
4019 sector_t min_offset_diff
= 0;
4022 struct r10conf
*conf
= mddev
->private;
4023 struct md_rdev
*rdev
;
4027 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4030 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4033 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4034 conf
->prev
.far_copies
);
4035 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4036 conf
->geo
.far_copies
);
4038 rdev_for_each(rdev
, mddev
) {
4039 if (!test_bit(In_sync
, &rdev
->flags
)
4040 && !test_bit(Faulty
, &rdev
->flags
))
4042 if (rdev
->raid_disk
>= 0) {
4043 long long diff
= (rdev
->new_data_offset
4044 - rdev
->data_offset
);
4045 if (!mddev
->reshape_backwards
)
4049 if (first
|| diff
< min_offset_diff
)
4050 min_offset_diff
= diff
;
4054 if (max(before_length
, after_length
) > min_offset_diff
)
4057 if (spares
< mddev
->delta_disks
)
4060 conf
->offset_diff
= min_offset_diff
;
4061 spin_lock_irq(&conf
->device_lock
);
4062 if (conf
->mirrors_new
) {
4063 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4064 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4066 kfree(conf
->mirrors_old
);
4067 conf
->mirrors_old
= conf
->mirrors
;
4068 conf
->mirrors
= conf
->mirrors_new
;
4069 conf
->mirrors_new
= NULL
;
4071 setup_geo(&conf
->geo
, mddev
, geo_start
);
4073 if (mddev
->reshape_backwards
) {
4074 sector_t size
= raid10_size(mddev
, 0, 0);
4075 if (size
< mddev
->array_sectors
) {
4076 spin_unlock_irq(&conf
->device_lock
);
4077 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4081 mddev
->resync_max_sectors
= size
;
4082 conf
->reshape_progress
= size
;
4084 conf
->reshape_progress
= 0;
4085 conf
->reshape_safe
= conf
->reshape_progress
;
4086 spin_unlock_irq(&conf
->device_lock
);
4088 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4089 ret
= bitmap_resize(mddev
->bitmap
,
4090 raid10_size(mddev
, 0,
4091 conf
->geo
.raid_disks
),
4096 if (mddev
->delta_disks
> 0) {
4097 rdev_for_each(rdev
, mddev
)
4098 if (rdev
->raid_disk
< 0 &&
4099 !test_bit(Faulty
, &rdev
->flags
)) {
4100 if (raid10_add_disk(mddev
, rdev
) == 0) {
4101 if (rdev
->raid_disk
>=
4102 conf
->prev
.raid_disks
)
4103 set_bit(In_sync
, &rdev
->flags
);
4105 rdev
->recovery_offset
= 0;
4107 if (sysfs_link_rdev(mddev
, rdev
))
4108 /* Failure here is OK */;
4110 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4111 && !test_bit(Faulty
, &rdev
->flags
)) {
4112 /* This is a spare that was manually added */
4113 set_bit(In_sync
, &rdev
->flags
);
4116 /* When a reshape changes the number of devices,
4117 * ->degraded is measured against the larger of the
4118 * pre and post numbers.
4120 spin_lock_irq(&conf
->device_lock
);
4121 mddev
->degraded
= calc_degraded(conf
);
4122 spin_unlock_irq(&conf
->device_lock
);
4123 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4124 mddev
->reshape_position
= conf
->reshape_progress
;
4125 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4127 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4128 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4129 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
4130 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4131 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4133 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4135 if (!mddev
->sync_thread
) {
4139 conf
->reshape_checkpoint
= jiffies
;
4140 md_wakeup_thread(mddev
->sync_thread
);
4141 md_new_event(mddev
);
4145 mddev
->recovery
= 0;
4146 spin_lock_irq(&conf
->device_lock
);
4147 conf
->geo
= conf
->prev
;
4148 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4149 rdev_for_each(rdev
, mddev
)
4150 rdev
->new_data_offset
= rdev
->data_offset
;
4152 conf
->reshape_progress
= MaxSector
;
4153 conf
->reshape_safe
= MaxSector
;
4154 mddev
->reshape_position
= MaxSector
;
4155 spin_unlock_irq(&conf
->device_lock
);
4159 /* Calculate the last device-address that could contain
4160 * any block from the chunk that includes the array-address 's'
4161 * and report the next address.
4162 * i.e. the address returned will be chunk-aligned and after
4163 * any data that is in the chunk containing 's'.
4165 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4167 s
= (s
| geo
->chunk_mask
) + 1;
4168 s
>>= geo
->chunk_shift
;
4169 s
*= geo
->near_copies
;
4170 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4171 s
*= geo
->far_copies
;
4172 s
<<= geo
->chunk_shift
;
4176 /* Calculate the first device-address that could contain
4177 * any block from the chunk that includes the array-address 's'.
4178 * This too will be the start of a chunk
4180 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4182 s
>>= geo
->chunk_shift
;
4183 s
*= geo
->near_copies
;
4184 sector_div(s
, geo
->raid_disks
);
4185 s
*= geo
->far_copies
;
4186 s
<<= geo
->chunk_shift
;
4190 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4193 /* We simply copy at most one chunk (smallest of old and new)
4194 * at a time, possibly less if that exceeds RESYNC_PAGES,
4195 * or we hit a bad block or something.
4196 * This might mean we pause for normal IO in the middle of
4197 * a chunk, but that is not a problem as mddev->reshape_position
4198 * can record any location.
4200 * If we will want to write to a location that isn't
4201 * yet recorded as 'safe' (i.e. in metadata on disk) then
4202 * we need to flush all reshape requests and update the metadata.
4204 * When reshaping forwards (e.g. to more devices), we interpret
4205 * 'safe' as the earliest block which might not have been copied
4206 * down yet. We divide this by previous stripe size and multiply
4207 * by previous stripe length to get lowest device offset that we
4208 * cannot write to yet.
4209 * We interpret 'sector_nr' as an address that we want to write to.
4210 * From this we use last_device_address() to find where we might
4211 * write to, and first_device_address on the 'safe' position.
4212 * If this 'next' write position is after the 'safe' position,
4213 * we must update the metadata to increase the 'safe' position.
4215 * When reshaping backwards, we round in the opposite direction
4216 * and perform the reverse test: next write position must not be
4217 * less than current safe position.
4219 * In all this the minimum difference in data offsets
4220 * (conf->offset_diff - always positive) allows a bit of slack,
4221 * so next can be after 'safe', but not by more than offset_diff
4223 * We need to prepare all the bios here before we start any IO
4224 * to ensure the size we choose is acceptable to all devices.
4225 * The means one for each copy for write-out and an extra one for
4227 * We store the read-in bio in ->master_bio and the others in
4228 * ->devs[x].bio and ->devs[x].repl_bio.
4230 struct r10conf
*conf
= mddev
->private;
4231 struct r10bio
*r10_bio
;
4232 sector_t next
, safe
, last
;
4236 struct md_rdev
*rdev
;
4239 struct bio
*bio
, *read_bio
;
4240 int sectors_done
= 0;
4242 if (sector_nr
== 0) {
4243 /* If restarting in the middle, skip the initial sectors */
4244 if (mddev
->reshape_backwards
&&
4245 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4246 sector_nr
= (raid10_size(mddev
, 0, 0)
4247 - conf
->reshape_progress
);
4248 } else if (!mddev
->reshape_backwards
&&
4249 conf
->reshape_progress
> 0)
4250 sector_nr
= conf
->reshape_progress
;
4252 mddev
->curr_resync_completed
= sector_nr
;
4253 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4259 /* We don't use sector_nr to track where we are up to
4260 * as that doesn't work well for ->reshape_backwards.
4261 * So just use ->reshape_progress.
4263 if (mddev
->reshape_backwards
) {
4264 /* 'next' is the earliest device address that we might
4265 * write to for this chunk in the new layout
4267 next
= first_dev_address(conf
->reshape_progress
- 1,
4270 /* 'safe' is the last device address that we might read from
4271 * in the old layout after a restart
4273 safe
= last_dev_address(conf
->reshape_safe
- 1,
4276 if (next
+ conf
->offset_diff
< safe
)
4279 last
= conf
->reshape_progress
- 1;
4280 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4281 & conf
->prev
.chunk_mask
);
4282 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4283 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4285 /* 'next' is after the last device address that we
4286 * might write to for this chunk in the new layout
4288 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4290 /* 'safe' is the earliest device address that we might
4291 * read from in the old layout after a restart
4293 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4295 /* Need to update metadata if 'next' might be beyond 'safe'
4296 * as that would possibly corrupt data
4298 if (next
> safe
+ conf
->offset_diff
)
4301 sector_nr
= conf
->reshape_progress
;
4302 last
= sector_nr
| (conf
->geo
.chunk_mask
4303 & conf
->prev
.chunk_mask
);
4305 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4306 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4310 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4311 /* Need to update reshape_position in metadata */
4313 mddev
->reshape_position
= conf
->reshape_progress
;
4314 if (mddev
->reshape_backwards
)
4315 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4316 - conf
->reshape_progress
;
4318 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4319 conf
->reshape_checkpoint
= jiffies
;
4320 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4321 md_wakeup_thread(mddev
->thread
);
4322 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4323 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4324 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4325 allow_barrier(conf
);
4326 return sectors_done
;
4328 conf
->reshape_safe
= mddev
->reshape_position
;
4329 allow_barrier(conf
);
4333 /* Now schedule reads for blocks from sector_nr to last */
4334 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4336 raise_barrier(conf
, sectors_done
!= 0);
4337 atomic_set(&r10_bio
->remaining
, 0);
4338 r10_bio
->mddev
= mddev
;
4339 r10_bio
->sector
= sector_nr
;
4340 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4341 r10_bio
->sectors
= last
- sector_nr
+ 1;
4342 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4343 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4346 /* Cannot read from here, so need to record bad blocks
4347 * on all the target devices.
4350 mempool_free(r10_bio
, conf
->r10buf_pool
);
4351 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4352 return sectors_done
;
4355 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4357 read_bio
->bi_bdev
= rdev
->bdev
;
4358 read_bio
->bi_iter
.bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4359 + rdev
->data_offset
);
4360 read_bio
->bi_private
= r10_bio
;
4361 read_bio
->bi_end_io
= end_sync_read
;
4362 bio_set_op_attrs(read_bio
, REQ_OP_READ
, 0);
4363 read_bio
->bi_flags
&= (~0UL << BIO_RESET_BITS
);
4364 read_bio
->bi_error
= 0;
4365 read_bio
->bi_vcnt
= 0;
4366 read_bio
->bi_iter
.bi_size
= 0;
4367 r10_bio
->master_bio
= read_bio
;
4368 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4370 /* Now find the locations in the new layout */
4371 __raid10_find_phys(&conf
->geo
, r10_bio
);
4374 read_bio
->bi_next
= NULL
;
4377 for (s
= 0; s
< conf
->copies
*2; s
++) {
4379 int d
= r10_bio
->devs
[s
/2].devnum
;
4380 struct md_rdev
*rdev2
;
4382 rdev2
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4383 b
= r10_bio
->devs
[s
/2].repl_bio
;
4385 rdev2
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4386 b
= r10_bio
->devs
[s
/2].bio
;
4388 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4392 b
->bi_bdev
= rdev2
->bdev
;
4393 b
->bi_iter
.bi_sector
= r10_bio
->devs
[s
/2].addr
+
4394 rdev2
->new_data_offset
;
4395 b
->bi_private
= r10_bio
;
4396 b
->bi_end_io
= end_reshape_write
;
4397 bio_set_op_attrs(b
, REQ_OP_WRITE
, 0);
4402 /* Now add as many pages as possible to all of these bios. */
4405 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4406 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4407 int len
= (max_sectors
- s
) << 9;
4408 if (len
> PAGE_SIZE
)
4410 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4412 if (bio_add_page(bio
, page
, len
, 0))
4415 /* Didn't fit, must stop */
4417 bio2
&& bio2
!= bio
;
4418 bio2
= bio2
->bi_next
) {
4419 /* Remove last page from this bio */
4421 bio2
->bi_iter
.bi_size
-= len
;
4422 bio_clear_flag(bio2
, BIO_SEG_VALID
);
4426 sector_nr
+= len
>> 9;
4427 nr_sectors
+= len
>> 9;
4431 r10_bio
->sectors
= nr_sectors
;
4433 /* Now submit the read */
4434 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4435 atomic_inc(&r10_bio
->remaining
);
4436 read_bio
->bi_next
= NULL
;
4437 generic_make_request(read_bio
);
4438 sector_nr
+= nr_sectors
;
4439 sectors_done
+= nr_sectors
;
4440 if (sector_nr
<= last
)
4443 /* Now that we have done the whole section we can
4444 * update reshape_progress
4446 if (mddev
->reshape_backwards
)
4447 conf
->reshape_progress
-= sectors_done
;
4449 conf
->reshape_progress
+= sectors_done
;
4451 return sectors_done
;
4454 static void end_reshape_request(struct r10bio
*r10_bio
);
4455 static int handle_reshape_read_error(struct mddev
*mddev
,
4456 struct r10bio
*r10_bio
);
4457 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4459 /* Reshape read completed. Hopefully we have a block
4461 * If we got a read error then we do sync 1-page reads from
4462 * elsewhere until we find the data - or give up.
4464 struct r10conf
*conf
= mddev
->private;
4467 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4468 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4469 /* Reshape has been aborted */
4470 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4474 /* We definitely have the data in the pages, schedule the
4477 atomic_set(&r10_bio
->remaining
, 1);
4478 for (s
= 0; s
< conf
->copies
*2; s
++) {
4480 int d
= r10_bio
->devs
[s
/2].devnum
;
4481 struct md_rdev
*rdev
;
4484 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4485 b
= r10_bio
->devs
[s
/2].repl_bio
;
4487 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4488 b
= r10_bio
->devs
[s
/2].bio
;
4490 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
4494 atomic_inc(&rdev
->nr_pending
);
4496 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4497 atomic_inc(&r10_bio
->remaining
);
4499 generic_make_request(b
);
4501 end_reshape_request(r10_bio
);
4504 static void end_reshape(struct r10conf
*conf
)
4506 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4509 spin_lock_irq(&conf
->device_lock
);
4510 conf
->prev
= conf
->geo
;
4511 md_finish_reshape(conf
->mddev
);
4513 conf
->reshape_progress
= MaxSector
;
4514 conf
->reshape_safe
= MaxSector
;
4515 spin_unlock_irq(&conf
->device_lock
);
4517 /* read-ahead size must cover two whole stripes, which is
4518 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4520 if (conf
->mddev
->queue
) {
4521 int stripe
= conf
->geo
.raid_disks
*
4522 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4523 stripe
/= conf
->geo
.near_copies
;
4524 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4525 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4530 static int handle_reshape_read_error(struct mddev
*mddev
,
4531 struct r10bio
*r10_bio
)
4533 /* Use sync reads to get the blocks from somewhere else */
4534 int sectors
= r10_bio
->sectors
;
4535 struct r10conf
*conf
= mddev
->private;
4537 struct r10bio r10_bio
;
4538 struct r10dev devs
[conf
->copies
];
4540 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4543 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4545 r10b
->sector
= r10_bio
->sector
;
4546 __raid10_find_phys(&conf
->prev
, r10b
);
4551 int first_slot
= slot
;
4553 if (s
> (PAGE_SIZE
>> 9))
4558 int d
= r10b
->devs
[slot
].devnum
;
4559 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4562 test_bit(Faulty
, &rdev
->flags
) ||
4563 !test_bit(In_sync
, &rdev
->flags
))
4566 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4567 atomic_inc(&rdev
->nr_pending
);
4569 success
= sync_page_io(rdev
,
4573 REQ_OP_READ
, 0, false);
4574 rdev_dec_pending(rdev
, mddev
);
4580 if (slot
>= conf
->copies
)
4582 if (slot
== first_slot
)
4587 /* couldn't read this block, must give up */
4588 set_bit(MD_RECOVERY_INTR
,
4598 static void end_reshape_write(struct bio
*bio
)
4600 struct r10bio
*r10_bio
= bio
->bi_private
;
4601 struct mddev
*mddev
= r10_bio
->mddev
;
4602 struct r10conf
*conf
= mddev
->private;
4606 struct md_rdev
*rdev
= NULL
;
4608 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4610 rdev
= conf
->mirrors
[d
].replacement
;
4613 rdev
= conf
->mirrors
[d
].rdev
;
4616 if (bio
->bi_error
) {
4617 /* FIXME should record badblock */
4618 md_error(mddev
, rdev
);
4621 rdev_dec_pending(rdev
, mddev
);
4622 end_reshape_request(r10_bio
);
4625 static void end_reshape_request(struct r10bio
*r10_bio
)
4627 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4629 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4630 bio_put(r10_bio
->master_bio
);
4634 static void raid10_finish_reshape(struct mddev
*mddev
)
4636 struct r10conf
*conf
= mddev
->private;
4638 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4641 if (mddev
->delta_disks
> 0) {
4642 sector_t size
= raid10_size(mddev
, 0, 0);
4643 md_set_array_sectors(mddev
, size
);
4644 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4645 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4646 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4648 mddev
->resync_max_sectors
= size
;
4650 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4651 revalidate_disk(mddev
->gendisk
);
4656 for (d
= conf
->geo
.raid_disks
;
4657 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4659 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
4661 clear_bit(In_sync
, &rdev
->flags
);
4662 rdev
= rcu_dereference(conf
->mirrors
[d
].replacement
);
4664 clear_bit(In_sync
, &rdev
->flags
);
4668 mddev
->layout
= mddev
->new_layout
;
4669 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4670 mddev
->reshape_position
= MaxSector
;
4671 mddev
->delta_disks
= 0;
4672 mddev
->reshape_backwards
= 0;
4675 static struct md_personality raid10_personality
=
4679 .owner
= THIS_MODULE
,
4680 .make_request
= raid10_make_request
,
4682 .free
= raid10_free
,
4683 .status
= raid10_status
,
4684 .error_handler
= raid10_error
,
4685 .hot_add_disk
= raid10_add_disk
,
4686 .hot_remove_disk
= raid10_remove_disk
,
4687 .spare_active
= raid10_spare_active
,
4688 .sync_request
= raid10_sync_request
,
4689 .quiesce
= raid10_quiesce
,
4690 .size
= raid10_size
,
4691 .resize
= raid10_resize
,
4692 .takeover
= raid10_takeover
,
4693 .check_reshape
= raid10_check_reshape
,
4694 .start_reshape
= raid10_start_reshape
,
4695 .finish_reshape
= raid10_finish_reshape
,
4696 .congested
= raid10_congested
,
4699 static int __init
raid_init(void)
4701 return register_md_personality(&raid10_personality
);
4704 static void raid_exit(void)
4706 unregister_md_personality(&raid10_personality
);
4709 module_init(raid_init
);
4710 module_exit(raid_exit
);
4711 MODULE_LICENSE("GPL");
4712 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4713 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4714 MODULE_ALIAS("md-raid10");
4715 MODULE_ALIAS("md-level-10");
4717 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);