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
);
711 sectors
= r10_bio
->sectors
;
714 best_dist
= MaxSector
;
715 best_good_sectors
= 0;
718 * Check if we can balance. We can balance on the whole
719 * device if no resync is going on (recovery is ok), or below
720 * the resync window. We take the first readable disk when
721 * above the resync window.
723 if (conf
->mddev
->recovery_cp
< MaxSector
724 && (this_sector
+ sectors
>= conf
->next_resync
))
727 for (slot
= 0; slot
< conf
->copies
; slot
++) {
732 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
734 disk
= r10_bio
->devs
[slot
].devnum
;
735 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
736 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
737 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
738 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
740 test_bit(Faulty
, &rdev
->flags
))
742 if (!test_bit(In_sync
, &rdev
->flags
) &&
743 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
746 dev_sector
= r10_bio
->devs
[slot
].addr
;
747 if (is_badblock(rdev
, dev_sector
, sectors
,
748 &first_bad
, &bad_sectors
)) {
749 if (best_dist
< MaxSector
)
750 /* Already have a better slot */
752 if (first_bad
<= dev_sector
) {
753 /* Cannot read here. If this is the
754 * 'primary' device, then we must not read
755 * beyond 'bad_sectors' from another device.
757 bad_sectors
-= (dev_sector
- first_bad
);
758 if (!do_balance
&& sectors
> bad_sectors
)
759 sectors
= bad_sectors
;
760 if (best_good_sectors
> sectors
)
761 best_good_sectors
= sectors
;
763 sector_t good_sectors
=
764 first_bad
- dev_sector
;
765 if (good_sectors
> best_good_sectors
) {
766 best_good_sectors
= good_sectors
;
771 /* Must read from here */
776 best_good_sectors
= sectors
;
781 /* This optimisation is debatable, and completely destroys
782 * sequential read speed for 'far copies' arrays. So only
783 * keep it for 'near' arrays, and review those later.
785 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
788 /* for far > 1 always use the lowest address */
789 if (geo
->far_copies
> 1)
790 new_distance
= r10_bio
->devs
[slot
].addr
;
792 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
793 conf
->mirrors
[disk
].head_position
);
794 if (new_distance
< best_dist
) {
795 best_dist
= new_distance
;
800 if (slot
>= conf
->copies
) {
806 atomic_inc(&rdev
->nr_pending
);
807 if (test_bit(Faulty
, &rdev
->flags
)) {
808 /* Cannot risk returning a device that failed
809 * before we inc'ed nr_pending
811 rdev_dec_pending(rdev
, conf
->mddev
);
814 r10_bio
->read_slot
= slot
;
818 *max_sectors
= best_good_sectors
;
823 static int raid10_congested(struct mddev
*mddev
, int bits
)
825 struct r10conf
*conf
= mddev
->private;
828 if ((bits
& (1 << WB_async_congested
)) &&
829 conf
->pending_count
>= max_queued_requests
)
834 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
837 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
838 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
839 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
841 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
848 static void flush_pending_writes(struct r10conf
*conf
)
850 /* Any writes that have been queued but are awaiting
851 * bitmap updates get flushed here.
853 spin_lock_irq(&conf
->device_lock
);
855 if (conf
->pending_bio_list
.head
) {
857 bio
= bio_list_get(&conf
->pending_bio_list
);
858 conf
->pending_count
= 0;
859 spin_unlock_irq(&conf
->device_lock
);
860 /* flush any pending bitmap writes to disk
861 * before proceeding w/ I/O */
862 bitmap_unplug(conf
->mddev
->bitmap
);
863 wake_up(&conf
->wait_barrier
);
865 while (bio
) { /* submit pending writes */
866 struct bio
*next
= bio
->bi_next
;
868 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
869 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
873 generic_make_request(bio
);
877 spin_unlock_irq(&conf
->device_lock
);
881 * Sometimes we need to suspend IO while we do something else,
882 * either some resync/recovery, or reconfigure the array.
883 * To do this we raise a 'barrier'.
884 * The 'barrier' is a counter that can be raised multiple times
885 * to count how many activities are happening which preclude
887 * We can only raise the barrier if there is no pending IO.
888 * i.e. if nr_pending == 0.
889 * We choose only to raise the barrier if no-one is waiting for the
890 * barrier to go down. This means that as soon as an IO request
891 * is ready, no other operations which require a barrier will start
892 * until the IO request has had a chance.
894 * So: regular IO calls 'wait_barrier'. When that returns there
895 * is no backgroup IO happening, It must arrange to call
896 * allow_barrier when it has finished its IO.
897 * backgroup IO calls must call raise_barrier. Once that returns
898 * there is no normal IO happeing. It must arrange to call
899 * lower_barrier when the particular background IO completes.
902 static void raise_barrier(struct r10conf
*conf
, int force
)
904 BUG_ON(force
&& !conf
->barrier
);
905 spin_lock_irq(&conf
->resync_lock
);
907 /* Wait until no block IO is waiting (unless 'force') */
908 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
911 /* block any new IO from starting */
914 /* Now wait for all pending IO to complete */
915 wait_event_lock_irq(conf
->wait_barrier
,
916 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
919 spin_unlock_irq(&conf
->resync_lock
);
922 static void lower_barrier(struct r10conf
*conf
)
925 spin_lock_irqsave(&conf
->resync_lock
, flags
);
927 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
928 wake_up(&conf
->wait_barrier
);
931 static void wait_barrier(struct r10conf
*conf
)
933 spin_lock_irq(&conf
->resync_lock
);
936 /* Wait for the barrier to drop.
937 * However if there are already pending
938 * requests (preventing the barrier from
939 * rising completely), and the
940 * pre-process bio queue isn't empty,
941 * then don't wait, as we need to empty
942 * that queue to get the nr_pending
945 wait_event_lock_irq(conf
->wait_barrier
,
949 !bio_list_empty(current
->bio_list
)),
954 spin_unlock_irq(&conf
->resync_lock
);
957 static void allow_barrier(struct r10conf
*conf
)
960 spin_lock_irqsave(&conf
->resync_lock
, flags
);
962 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
963 wake_up(&conf
->wait_barrier
);
966 static void freeze_array(struct r10conf
*conf
, int extra
)
968 /* stop syncio and normal IO and wait for everything to
970 * We increment barrier and nr_waiting, and then
971 * wait until nr_pending match nr_queued+extra
972 * This is called in the context of one normal IO request
973 * that has failed. Thus any sync request that might be pending
974 * will be blocked by nr_pending, and we need to wait for
975 * pending IO requests to complete or be queued for re-try.
976 * Thus the number queued (nr_queued) plus this request (extra)
977 * must match the number of pending IOs (nr_pending) before
980 spin_lock_irq(&conf
->resync_lock
);
983 wait_event_lock_irq_cmd(conf
->wait_barrier
,
984 conf
->nr_pending
== conf
->nr_queued
+extra
,
986 flush_pending_writes(conf
));
988 spin_unlock_irq(&conf
->resync_lock
);
991 static void unfreeze_array(struct r10conf
*conf
)
993 /* reverse the effect of the freeze */
994 spin_lock_irq(&conf
->resync_lock
);
997 wake_up(&conf
->wait_barrier
);
998 spin_unlock_irq(&conf
->resync_lock
);
1001 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1002 struct md_rdev
*rdev
)
1004 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1005 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1006 return rdev
->data_offset
;
1008 return rdev
->new_data_offset
;
1011 struct raid10_plug_cb
{
1012 struct blk_plug_cb cb
;
1013 struct bio_list pending
;
1017 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1019 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1021 struct mddev
*mddev
= plug
->cb
.data
;
1022 struct r10conf
*conf
= mddev
->private;
1025 if (from_schedule
|| current
->bio_list
) {
1026 spin_lock_irq(&conf
->device_lock
);
1027 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1028 conf
->pending_count
+= plug
->pending_cnt
;
1029 spin_unlock_irq(&conf
->device_lock
);
1030 wake_up(&conf
->wait_barrier
);
1031 md_wakeup_thread(mddev
->thread
);
1036 /* we aren't scheduling, so we can do the write-out directly. */
1037 bio
= bio_list_get(&plug
->pending
);
1038 bitmap_unplug(mddev
->bitmap
);
1039 wake_up(&conf
->wait_barrier
);
1041 while (bio
) { /* submit pending writes */
1042 struct bio
*next
= bio
->bi_next
;
1043 bio
->bi_next
= NULL
;
1044 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
1045 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1046 /* Just ignore it */
1049 generic_make_request(bio
);
1055 static void __make_request(struct mddev
*mddev
, struct bio
*bio
)
1057 struct r10conf
*conf
= mddev
->private;
1058 struct r10bio
*r10_bio
;
1059 struct bio
*read_bio
;
1061 const int rw
= bio_data_dir(bio
);
1062 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1063 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1064 const unsigned long do_discard
= (bio
->bi_rw
1065 & (REQ_DISCARD
| REQ_SECURE
));
1066 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1067 unsigned long flags
;
1068 struct md_rdev
*blocked_rdev
;
1069 struct blk_plug_cb
*cb
;
1070 struct raid10_plug_cb
*plug
= NULL
;
1071 int sectors_handled
;
1076 * Register the new request and wait if the reconstruction
1077 * thread has put up a bar for new requests.
1078 * Continue immediately if no resync is active currently.
1082 sectors
= bio_sectors(bio
);
1083 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1084 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1085 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1086 /* IO spans the reshape position. Need to wait for
1089 allow_barrier(conf
);
1090 wait_event(conf
->wait_barrier
,
1091 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1092 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1096 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1097 bio_data_dir(bio
) == WRITE
&&
1098 (mddev
->reshape_backwards
1099 ? (bio
->bi_iter
.bi_sector
< conf
->reshape_safe
&&
1100 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
)
1101 : (bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_safe
&&
1102 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
))) {
1103 /* Need to update reshape_position in metadata */
1104 mddev
->reshape_position
= conf
->reshape_progress
;
1105 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1106 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1107 md_wakeup_thread(mddev
->thread
);
1108 wait_event(mddev
->sb_wait
,
1109 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1111 conf
->reshape_safe
= mddev
->reshape_position
;
1114 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1116 r10_bio
->master_bio
= bio
;
1117 r10_bio
->sectors
= sectors
;
1119 r10_bio
->mddev
= mddev
;
1120 r10_bio
->sector
= bio
->bi_iter
.bi_sector
;
1123 /* We might need to issue multiple reads to different
1124 * devices if there are bad blocks around, so we keep
1125 * track of the number of reads in bio->bi_phys_segments.
1126 * If this is 0, there is only one r10_bio and no locking
1127 * will be needed when the request completes. If it is
1128 * non-zero, then it is the number of not-completed requests.
1130 bio
->bi_phys_segments
= 0;
1131 bio_clear_flag(bio
, BIO_SEG_VALID
);
1135 * read balancing logic:
1137 struct md_rdev
*rdev
;
1141 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1143 raid_end_bio_io(r10_bio
);
1146 slot
= r10_bio
->read_slot
;
1148 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1149 bio_trim(read_bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1152 r10_bio
->devs
[slot
].bio
= read_bio
;
1153 r10_bio
->devs
[slot
].rdev
= rdev
;
1155 read_bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
+
1156 choose_data_offset(r10_bio
, rdev
);
1157 read_bio
->bi_bdev
= rdev
->bdev
;
1158 read_bio
->bi_end_io
= raid10_end_read_request
;
1159 read_bio
->bi_rw
= READ
| do_sync
;
1160 read_bio
->bi_private
= r10_bio
;
1162 if (max_sectors
< r10_bio
->sectors
) {
1163 /* Could not read all from this device, so we will
1164 * need another r10_bio.
1166 sectors_handled
= (r10_bio
->sector
+ max_sectors
1167 - bio
->bi_iter
.bi_sector
);
1168 r10_bio
->sectors
= max_sectors
;
1169 spin_lock_irq(&conf
->device_lock
);
1170 if (bio
->bi_phys_segments
== 0)
1171 bio
->bi_phys_segments
= 2;
1173 bio
->bi_phys_segments
++;
1174 spin_unlock_irq(&conf
->device_lock
);
1175 /* Cannot call generic_make_request directly
1176 * as that will be queued in __generic_make_request
1177 * and subsequent mempool_alloc might block
1178 * waiting for it. so hand bio over to raid10d.
1180 reschedule_retry(r10_bio
);
1182 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1184 r10_bio
->master_bio
= bio
;
1185 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1187 r10_bio
->mddev
= mddev
;
1188 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+
1192 generic_make_request(read_bio
);
1199 if (conf
->pending_count
>= max_queued_requests
) {
1200 md_wakeup_thread(mddev
->thread
);
1201 wait_event(conf
->wait_barrier
,
1202 conf
->pending_count
< max_queued_requests
);
1204 /* first select target devices under rcu_lock and
1205 * inc refcount on their rdev. Record them by setting
1207 * If there are known/acknowledged bad blocks on any device
1208 * on which we have seen a write error, we want to avoid
1209 * writing to those blocks. This potentially requires several
1210 * writes to write around the bad blocks. Each set of writes
1211 * gets its own r10_bio with a set of bios attached. The number
1212 * of r10_bios is recored in bio->bi_phys_segments just as with
1216 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1217 raid10_find_phys(conf
, r10_bio
);
1219 blocked_rdev
= NULL
;
1221 max_sectors
= r10_bio
->sectors
;
1223 for (i
= 0; i
< conf
->copies
; i
++) {
1224 int d
= r10_bio
->devs
[i
].devnum
;
1225 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1226 struct md_rdev
*rrdev
= rcu_dereference(
1227 conf
->mirrors
[d
].replacement
);
1230 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1231 atomic_inc(&rdev
->nr_pending
);
1232 blocked_rdev
= rdev
;
1235 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1236 atomic_inc(&rrdev
->nr_pending
);
1237 blocked_rdev
= rrdev
;
1240 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)))
1242 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)))
1245 r10_bio
->devs
[i
].bio
= NULL
;
1246 r10_bio
->devs
[i
].repl_bio
= NULL
;
1248 if (!rdev
&& !rrdev
) {
1249 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1252 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1254 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1258 is_bad
= is_badblock(rdev
, dev_sector
,
1260 &first_bad
, &bad_sectors
);
1262 /* Mustn't write here until the bad block
1265 atomic_inc(&rdev
->nr_pending
);
1266 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1267 blocked_rdev
= rdev
;
1270 if (is_bad
&& first_bad
<= dev_sector
) {
1271 /* Cannot write here at all */
1272 bad_sectors
-= (dev_sector
- first_bad
);
1273 if (bad_sectors
< max_sectors
)
1274 /* Mustn't write more than bad_sectors
1275 * to other devices yet
1277 max_sectors
= bad_sectors
;
1278 /* We don't set R10BIO_Degraded as that
1279 * only applies if the disk is missing,
1280 * so it might be re-added, and we want to
1281 * know to recover this chunk.
1282 * In this case the device is here, and the
1283 * fact that this chunk is not in-sync is
1284 * recorded in the bad block log.
1289 int good_sectors
= first_bad
- dev_sector
;
1290 if (good_sectors
< max_sectors
)
1291 max_sectors
= good_sectors
;
1295 r10_bio
->devs
[i
].bio
= bio
;
1296 atomic_inc(&rdev
->nr_pending
);
1299 r10_bio
->devs
[i
].repl_bio
= bio
;
1300 atomic_inc(&rrdev
->nr_pending
);
1305 if (unlikely(blocked_rdev
)) {
1306 /* Have to wait for this device to get unblocked, then retry */
1310 for (j
= 0; j
< i
; j
++) {
1311 if (r10_bio
->devs
[j
].bio
) {
1312 d
= r10_bio
->devs
[j
].devnum
;
1313 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1315 if (r10_bio
->devs
[j
].repl_bio
) {
1316 struct md_rdev
*rdev
;
1317 d
= r10_bio
->devs
[j
].devnum
;
1318 rdev
= conf
->mirrors
[d
].replacement
;
1320 /* Race with remove_disk */
1322 rdev
= conf
->mirrors
[d
].rdev
;
1324 rdev_dec_pending(rdev
, mddev
);
1327 allow_barrier(conf
);
1328 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1333 if (max_sectors
< r10_bio
->sectors
) {
1334 /* We are splitting this into multiple parts, so
1335 * we need to prepare for allocating another r10_bio.
1337 r10_bio
->sectors
= max_sectors
;
1338 spin_lock_irq(&conf
->device_lock
);
1339 if (bio
->bi_phys_segments
== 0)
1340 bio
->bi_phys_segments
= 2;
1342 bio
->bi_phys_segments
++;
1343 spin_unlock_irq(&conf
->device_lock
);
1345 sectors_handled
= r10_bio
->sector
+ max_sectors
-
1346 bio
->bi_iter
.bi_sector
;
1348 atomic_set(&r10_bio
->remaining
, 1);
1349 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1351 for (i
= 0; i
< conf
->copies
; i
++) {
1353 int d
= r10_bio
->devs
[i
].devnum
;
1354 if (r10_bio
->devs
[i
].bio
) {
1355 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1356 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1357 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1359 r10_bio
->devs
[i
].bio
= mbio
;
1361 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1362 choose_data_offset(r10_bio
,
1364 mbio
->bi_bdev
= rdev
->bdev
;
1365 mbio
->bi_end_io
= raid10_end_write_request
;
1367 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1368 mbio
->bi_private
= r10_bio
;
1370 atomic_inc(&r10_bio
->remaining
);
1372 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1375 plug
= container_of(cb
, struct raid10_plug_cb
,
1379 spin_lock_irqsave(&conf
->device_lock
, flags
);
1381 bio_list_add(&plug
->pending
, mbio
);
1382 plug
->pending_cnt
++;
1384 bio_list_add(&conf
->pending_bio_list
, mbio
);
1385 conf
->pending_count
++;
1387 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1389 md_wakeup_thread(mddev
->thread
);
1392 if (r10_bio
->devs
[i
].repl_bio
) {
1393 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1395 /* Replacement just got moved to main 'rdev' */
1397 rdev
= conf
->mirrors
[d
].rdev
;
1399 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1400 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1402 r10_bio
->devs
[i
].repl_bio
= mbio
;
1404 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1407 mbio
->bi_bdev
= rdev
->bdev
;
1408 mbio
->bi_end_io
= raid10_end_write_request
;
1410 WRITE
| do_sync
| do_fua
| do_discard
| do_same
;
1411 mbio
->bi_private
= r10_bio
;
1413 atomic_inc(&r10_bio
->remaining
);
1414 spin_lock_irqsave(&conf
->device_lock
, flags
);
1415 bio_list_add(&conf
->pending_bio_list
, mbio
);
1416 conf
->pending_count
++;
1417 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1418 if (!mddev_check_plugged(mddev
))
1419 md_wakeup_thread(mddev
->thread
);
1423 /* Don't remove the bias on 'remaining' (one_write_done) until
1424 * after checking if we need to go around again.
1427 if (sectors_handled
< bio_sectors(bio
)) {
1428 one_write_done(r10_bio
);
1429 /* We need another r10_bio. It has already been counted
1430 * in bio->bi_phys_segments.
1432 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1434 r10_bio
->master_bio
= bio
;
1435 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1437 r10_bio
->mddev
= mddev
;
1438 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1442 one_write_done(r10_bio
);
1445 static void make_request(struct mddev
*mddev
, struct bio
*bio
)
1447 struct r10conf
*conf
= mddev
->private;
1448 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1449 int chunk_sects
= chunk_mask
+ 1;
1453 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1454 md_flush_request(mddev
, bio
);
1458 md_write_start(mddev
, bio
);
1463 * If this request crosses a chunk boundary, we need to split
1466 if (unlikely((bio
->bi_iter
.bi_sector
& chunk_mask
) +
1467 bio_sectors(bio
) > chunk_sects
1468 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1469 || conf
->prev
.near_copies
<
1470 conf
->prev
.raid_disks
))) {
1471 split
= bio_split(bio
, chunk_sects
-
1472 (bio
->bi_iter
.bi_sector
&
1474 GFP_NOIO
, fs_bio_set
);
1475 bio_chain(split
, bio
);
1480 __make_request(mddev
, split
);
1481 } while (split
!= bio
);
1483 /* In case raid10d snuck in to freeze_array */
1484 wake_up(&conf
->wait_barrier
);
1487 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1489 struct r10conf
*conf
= mddev
->private;
1492 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1493 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1494 if (conf
->geo
.near_copies
> 1)
1495 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1496 if (conf
->geo
.far_copies
> 1) {
1497 if (conf
->geo
.far_offset
)
1498 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1500 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1501 if (conf
->geo
.far_set_size
!= conf
->geo
.raid_disks
)
1502 seq_printf(seq
, " %d devices per set", conf
->geo
.far_set_size
);
1504 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1505 conf
->geo
.raid_disks
- mddev
->degraded
);
1506 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1507 seq_printf(seq
, "%s",
1508 conf
->mirrors
[i
].rdev
&&
1509 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1510 seq_printf(seq
, "]");
1513 /* check if there are enough drives for
1514 * every block to appear on atleast one.
1515 * Don't consider the device numbered 'ignore'
1516 * as we might be about to remove it.
1518 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1524 disks
= conf
->prev
.raid_disks
;
1525 ncopies
= conf
->prev
.near_copies
;
1527 disks
= conf
->geo
.raid_disks
;
1528 ncopies
= conf
->geo
.near_copies
;
1533 int n
= conf
->copies
;
1537 struct md_rdev
*rdev
;
1538 if (this != ignore
&&
1539 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1540 test_bit(In_sync
, &rdev
->flags
))
1542 this = (this+1) % disks
;
1546 first
= (first
+ ncopies
) % disks
;
1547 } while (first
!= 0);
1554 static int enough(struct r10conf
*conf
, int ignore
)
1556 /* when calling 'enough', both 'prev' and 'geo' must
1558 * This is ensured if ->reconfig_mutex or ->device_lock
1561 return _enough(conf
, 0, ignore
) &&
1562 _enough(conf
, 1, ignore
);
1565 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1567 char b
[BDEVNAME_SIZE
];
1568 struct r10conf
*conf
= mddev
->private;
1569 unsigned long flags
;
1572 * If it is not operational, then we have already marked it as dead
1573 * else if it is the last working disks, ignore the error, let the
1574 * next level up know.
1575 * else mark the drive as failed
1577 spin_lock_irqsave(&conf
->device_lock
, flags
);
1578 if (test_bit(In_sync
, &rdev
->flags
)
1579 && !enough(conf
, rdev
->raid_disk
)) {
1581 * Don't fail the drive, just return an IO error.
1583 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1586 if (test_and_clear_bit(In_sync
, &rdev
->flags
))
1589 * If recovery is running, make sure it aborts.
1591 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1592 set_bit(Blocked
, &rdev
->flags
);
1593 set_bit(Faulty
, &rdev
->flags
);
1594 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1595 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1596 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1598 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1599 "md/raid10:%s: Operation continuing on %d devices.\n",
1600 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1601 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1604 static void print_conf(struct r10conf
*conf
)
1607 struct raid10_info
*tmp
;
1609 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1611 printk(KERN_DEBUG
"(!conf)\n");
1614 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1615 conf
->geo
.raid_disks
);
1617 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1618 char b
[BDEVNAME_SIZE
];
1619 tmp
= conf
->mirrors
+ i
;
1621 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1622 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1623 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1624 bdevname(tmp
->rdev
->bdev
,b
));
1628 static void close_sync(struct r10conf
*conf
)
1631 allow_barrier(conf
);
1633 mempool_destroy(conf
->r10buf_pool
);
1634 conf
->r10buf_pool
= NULL
;
1637 static int raid10_spare_active(struct mddev
*mddev
)
1640 struct r10conf
*conf
= mddev
->private;
1641 struct raid10_info
*tmp
;
1643 unsigned long flags
;
1646 * Find all non-in_sync disks within the RAID10 configuration
1647 * and mark them in_sync
1649 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1650 tmp
= conf
->mirrors
+ i
;
1651 if (tmp
->replacement
1652 && tmp
->replacement
->recovery_offset
== MaxSector
1653 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1654 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1655 /* Replacement has just become active */
1657 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1660 /* Replaced device not technically faulty,
1661 * but we need to be sure it gets removed
1662 * and never re-added.
1664 set_bit(Faulty
, &tmp
->rdev
->flags
);
1665 sysfs_notify_dirent_safe(
1666 tmp
->rdev
->sysfs_state
);
1668 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1669 } else if (tmp
->rdev
1670 && tmp
->rdev
->recovery_offset
== MaxSector
1671 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1672 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1674 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1677 spin_lock_irqsave(&conf
->device_lock
, flags
);
1678 mddev
->degraded
-= count
;
1679 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1685 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1687 struct r10conf
*conf
= mddev
->private;
1691 int last
= conf
->geo
.raid_disks
- 1;
1693 if (mddev
->recovery_cp
< MaxSector
)
1694 /* only hot-add to in-sync arrays, as recovery is
1695 * very different from resync
1698 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1701 if (rdev
->raid_disk
>= 0)
1702 first
= last
= rdev
->raid_disk
;
1704 if (rdev
->saved_raid_disk
>= first
&&
1705 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1706 mirror
= rdev
->saved_raid_disk
;
1709 for ( ; mirror
<= last
; mirror
++) {
1710 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1711 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1714 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1715 p
->replacement
!= NULL
)
1717 clear_bit(In_sync
, &rdev
->flags
);
1718 set_bit(Replacement
, &rdev
->flags
);
1719 rdev
->raid_disk
= mirror
;
1722 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1723 rdev
->data_offset
<< 9);
1725 rcu_assign_pointer(p
->replacement
, rdev
);
1730 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1731 rdev
->data_offset
<< 9);
1733 p
->head_position
= 0;
1734 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1735 rdev
->raid_disk
= mirror
;
1737 if (rdev
->saved_raid_disk
!= mirror
)
1739 rcu_assign_pointer(p
->rdev
, rdev
);
1742 mddev_suspend(mddev
);
1743 md_integrity_add_rdev(rdev
, mddev
);
1744 mddev_resume(mddev
);
1745 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1746 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1752 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1754 struct r10conf
*conf
= mddev
->private;
1756 int number
= rdev
->raid_disk
;
1757 struct md_rdev
**rdevp
;
1758 struct raid10_info
*p
= conf
->mirrors
+ number
;
1761 if (rdev
== p
->rdev
)
1763 else if (rdev
== p
->replacement
)
1764 rdevp
= &p
->replacement
;
1768 if (test_bit(In_sync
, &rdev
->flags
) ||
1769 atomic_read(&rdev
->nr_pending
)) {
1773 /* Only remove faulty devices if recovery
1776 if (!test_bit(Faulty
, &rdev
->flags
) &&
1777 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1778 (!p
->replacement
|| p
->replacement
== rdev
) &&
1779 number
< conf
->geo
.raid_disks
&&
1786 if (atomic_read(&rdev
->nr_pending
)) {
1787 /* lost the race, try later */
1791 } else if (p
->replacement
) {
1792 /* We must have just cleared 'rdev' */
1793 p
->rdev
= p
->replacement
;
1794 clear_bit(Replacement
, &p
->replacement
->flags
);
1795 smp_mb(); /* Make sure other CPUs may see both as identical
1796 * but will never see neither -- if they are careful.
1798 p
->replacement
= NULL
;
1799 clear_bit(WantReplacement
, &rdev
->flags
);
1801 /* We might have just remove the Replacement as faulty
1802 * Clear the flag just in case
1804 clear_bit(WantReplacement
, &rdev
->flags
);
1806 err
= md_integrity_register(mddev
);
1814 static void end_sync_read(struct bio
*bio
)
1816 struct r10bio
*r10_bio
= bio
->bi_private
;
1817 struct r10conf
*conf
= r10_bio
->mddev
->private;
1820 if (bio
== r10_bio
->master_bio
) {
1821 /* this is a reshape read */
1822 d
= r10_bio
->read_slot
; /* really the read dev */
1824 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1827 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1829 /* The write handler will notice the lack of
1830 * R10BIO_Uptodate and record any errors etc
1832 atomic_add(r10_bio
->sectors
,
1833 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1835 /* for reconstruct, we always reschedule after a read.
1836 * for resync, only after all reads
1838 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1839 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1840 atomic_dec_and_test(&r10_bio
->remaining
)) {
1841 /* we have read all the blocks,
1842 * do the comparison in process context in raid10d
1844 reschedule_retry(r10_bio
);
1848 static void end_sync_request(struct r10bio
*r10_bio
)
1850 struct mddev
*mddev
= r10_bio
->mddev
;
1852 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1853 if (r10_bio
->master_bio
== NULL
) {
1854 /* the primary of several recovery bios */
1855 sector_t s
= r10_bio
->sectors
;
1856 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1857 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1858 reschedule_retry(r10_bio
);
1861 md_done_sync(mddev
, s
, 1);
1864 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1865 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1866 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1867 reschedule_retry(r10_bio
);
1875 static void end_sync_write(struct bio
*bio
)
1877 struct r10bio
*r10_bio
= bio
->bi_private
;
1878 struct mddev
*mddev
= r10_bio
->mddev
;
1879 struct r10conf
*conf
= mddev
->private;
1885 struct md_rdev
*rdev
= NULL
;
1887 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1889 rdev
= conf
->mirrors
[d
].replacement
;
1891 rdev
= conf
->mirrors
[d
].rdev
;
1893 if (bio
->bi_error
) {
1895 md_error(mddev
, rdev
);
1897 set_bit(WriteErrorSeen
, &rdev
->flags
);
1898 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1899 set_bit(MD_RECOVERY_NEEDED
,
1900 &rdev
->mddev
->recovery
);
1901 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1903 } else if (is_badblock(rdev
,
1904 r10_bio
->devs
[slot
].addr
,
1906 &first_bad
, &bad_sectors
))
1907 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1909 rdev_dec_pending(rdev
, mddev
);
1911 end_sync_request(r10_bio
);
1915 * Note: sync and recover and handled very differently for raid10
1916 * This code is for resync.
1917 * For resync, we read through virtual addresses and read all blocks.
1918 * If there is any error, we schedule a write. The lowest numbered
1919 * drive is authoritative.
1920 * However requests come for physical address, so we need to map.
1921 * For every physical address there are raid_disks/copies virtual addresses,
1922 * which is always are least one, but is not necessarly an integer.
1923 * This means that a physical address can span multiple chunks, so we may
1924 * have to submit multiple io requests for a single sync request.
1927 * We check if all blocks are in-sync and only write to blocks that
1930 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1932 struct r10conf
*conf
= mddev
->private;
1934 struct bio
*tbio
, *fbio
;
1937 atomic_set(&r10_bio
->remaining
, 1);
1939 /* find the first device with a block */
1940 for (i
=0; i
<conf
->copies
; i
++)
1941 if (!r10_bio
->devs
[i
].bio
->bi_error
)
1944 if (i
== conf
->copies
)
1948 fbio
= r10_bio
->devs
[i
].bio
;
1950 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
1951 /* now find blocks with errors */
1952 for (i
=0 ; i
< conf
->copies
; i
++) {
1955 tbio
= r10_bio
->devs
[i
].bio
;
1957 if (tbio
->bi_end_io
!= end_sync_read
)
1961 if (!r10_bio
->devs
[i
].bio
->bi_error
) {
1962 /* We know that the bi_io_vec layout is the same for
1963 * both 'first' and 'i', so we just compare them.
1964 * All vec entries are PAGE_SIZE;
1966 int sectors
= r10_bio
->sectors
;
1967 for (j
= 0; j
< vcnt
; j
++) {
1968 int len
= PAGE_SIZE
;
1969 if (sectors
< (len
/ 512))
1970 len
= sectors
* 512;
1971 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1972 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1979 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
1980 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1981 /* Don't fix anything. */
1984 /* Ok, we need to write this bio, either to correct an
1985 * inconsistency or to correct an unreadable block.
1986 * First we need to fixup bv_offset, bv_len and
1987 * bi_vecs, as the read request might have corrupted these
1991 tbio
->bi_vcnt
= vcnt
;
1992 tbio
->bi_iter
.bi_size
= r10_bio
->sectors
<< 9;
1993 tbio
->bi_rw
= WRITE
;
1994 tbio
->bi_private
= r10_bio
;
1995 tbio
->bi_iter
.bi_sector
= r10_bio
->devs
[i
].addr
;
1996 tbio
->bi_end_io
= end_sync_write
;
1998 bio_copy_data(tbio
, fbio
);
2000 d
= r10_bio
->devs
[i
].devnum
;
2001 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2002 atomic_inc(&r10_bio
->remaining
);
2003 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2005 tbio
->bi_iter
.bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2006 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2007 generic_make_request(tbio
);
2010 /* Now write out to any replacement devices
2013 for (i
= 0; i
< conf
->copies
; i
++) {
2016 tbio
= r10_bio
->devs
[i
].repl_bio
;
2017 if (!tbio
|| !tbio
->bi_end_io
)
2019 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2020 && r10_bio
->devs
[i
].bio
!= fbio
)
2021 bio_copy_data(tbio
, fbio
);
2022 d
= r10_bio
->devs
[i
].devnum
;
2023 atomic_inc(&r10_bio
->remaining
);
2024 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2026 generic_make_request(tbio
);
2030 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2031 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2037 * Now for the recovery code.
2038 * Recovery happens across physical sectors.
2039 * We recover all non-is_sync drives by finding the virtual address of
2040 * each, and then choose a working drive that also has that virt address.
2041 * There is a separate r10_bio for each non-in_sync drive.
2042 * Only the first two slots are in use. The first for reading,
2043 * The second for writing.
2046 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2048 /* We got a read error during recovery.
2049 * We repeat the read in smaller page-sized sections.
2050 * If a read succeeds, write it to the new device or record
2051 * a bad block if we cannot.
2052 * If a read fails, record a bad block on both old and
2055 struct mddev
*mddev
= r10_bio
->mddev
;
2056 struct r10conf
*conf
= mddev
->private;
2057 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2059 int sectors
= r10_bio
->sectors
;
2061 int dr
= r10_bio
->devs
[0].devnum
;
2062 int dw
= r10_bio
->devs
[1].devnum
;
2066 struct md_rdev
*rdev
;
2070 if (s
> (PAGE_SIZE
>>9))
2073 rdev
= conf
->mirrors
[dr
].rdev
;
2074 addr
= r10_bio
->devs
[0].addr
+ sect
,
2075 ok
= sync_page_io(rdev
,
2078 bio
->bi_io_vec
[idx
].bv_page
,
2081 rdev
= conf
->mirrors
[dw
].rdev
;
2082 addr
= r10_bio
->devs
[1].addr
+ sect
;
2083 ok
= sync_page_io(rdev
,
2086 bio
->bi_io_vec
[idx
].bv_page
,
2089 set_bit(WriteErrorSeen
, &rdev
->flags
);
2090 if (!test_and_set_bit(WantReplacement
,
2092 set_bit(MD_RECOVERY_NEEDED
,
2093 &rdev
->mddev
->recovery
);
2097 /* We don't worry if we cannot set a bad block -
2098 * it really is bad so there is no loss in not
2101 rdev_set_badblocks(rdev
, addr
, s
, 0);
2103 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2104 /* need bad block on destination too */
2105 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2106 addr
= r10_bio
->devs
[1].addr
+ sect
;
2107 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2109 /* just abort the recovery */
2111 "md/raid10:%s: recovery aborted"
2112 " due to read error\n",
2115 conf
->mirrors
[dw
].recovery_disabled
2116 = mddev
->recovery_disabled
;
2117 set_bit(MD_RECOVERY_INTR
,
2130 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2132 struct r10conf
*conf
= mddev
->private;
2134 struct bio
*wbio
, *wbio2
;
2136 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2137 fix_recovery_read_error(r10_bio
);
2138 end_sync_request(r10_bio
);
2143 * share the pages with the first bio
2144 * and submit the write request
2146 d
= r10_bio
->devs
[1].devnum
;
2147 wbio
= r10_bio
->devs
[1].bio
;
2148 wbio2
= r10_bio
->devs
[1].repl_bio
;
2149 /* Need to test wbio2->bi_end_io before we call
2150 * generic_make_request as if the former is NULL,
2151 * the latter is free to free wbio2.
2153 if (wbio2
&& !wbio2
->bi_end_io
)
2155 if (wbio
->bi_end_io
) {
2156 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2157 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2158 generic_make_request(wbio
);
2161 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2162 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2163 bio_sectors(wbio2
));
2164 generic_make_request(wbio2
);
2169 * Used by fix_read_error() to decay the per rdev read_errors.
2170 * We halve the read error count for every hour that has elapsed
2171 * since the last recorded read error.
2174 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2176 struct timespec cur_time_mon
;
2177 unsigned long hours_since_last
;
2178 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2180 ktime_get_ts(&cur_time_mon
);
2182 if (rdev
->last_read_error
.tv_sec
== 0 &&
2183 rdev
->last_read_error
.tv_nsec
== 0) {
2184 /* first time we've seen a read error */
2185 rdev
->last_read_error
= cur_time_mon
;
2189 hours_since_last
= (cur_time_mon
.tv_sec
-
2190 rdev
->last_read_error
.tv_sec
) / 3600;
2192 rdev
->last_read_error
= cur_time_mon
;
2195 * if hours_since_last is > the number of bits in read_errors
2196 * just set read errors to 0. We do this to avoid
2197 * overflowing the shift of read_errors by hours_since_last.
2199 if (hours_since_last
>= 8 * sizeof(read_errors
))
2200 atomic_set(&rdev
->read_errors
, 0);
2202 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2205 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2206 int sectors
, struct page
*page
, int rw
)
2211 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2212 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2214 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2218 set_bit(WriteErrorSeen
, &rdev
->flags
);
2219 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2220 set_bit(MD_RECOVERY_NEEDED
,
2221 &rdev
->mddev
->recovery
);
2223 /* need to record an error - either for the block or the device */
2224 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2225 md_error(rdev
->mddev
, rdev
);
2230 * This is a kernel thread which:
2232 * 1. Retries failed read operations on working mirrors.
2233 * 2. Updates the raid superblock when problems encounter.
2234 * 3. Performs writes following reads for array synchronising.
2237 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2239 int sect
= 0; /* Offset from r10_bio->sector */
2240 int sectors
= r10_bio
->sectors
;
2241 struct md_rdev
*rdev
;
2242 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2243 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2245 /* still own a reference to this rdev, so it cannot
2246 * have been cleared recently.
2248 rdev
= conf
->mirrors
[d
].rdev
;
2250 if (test_bit(Faulty
, &rdev
->flags
))
2251 /* drive has already been failed, just ignore any
2252 more fix_read_error() attempts */
2255 check_decay_read_errors(mddev
, rdev
);
2256 atomic_inc(&rdev
->read_errors
);
2257 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2258 char b
[BDEVNAME_SIZE
];
2259 bdevname(rdev
->bdev
, b
);
2262 "md/raid10:%s: %s: Raid device exceeded "
2263 "read_error threshold [cur %d:max %d]\n",
2265 atomic_read(&rdev
->read_errors
), max_read_errors
);
2267 "md/raid10:%s: %s: Failing raid device\n",
2269 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2270 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2276 int sl
= r10_bio
->read_slot
;
2280 if (s
> (PAGE_SIZE
>>9))
2288 d
= r10_bio
->devs
[sl
].devnum
;
2289 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2291 test_bit(In_sync
, &rdev
->flags
) &&
2292 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2293 &first_bad
, &bad_sectors
) == 0) {
2294 atomic_inc(&rdev
->nr_pending
);
2296 success
= sync_page_io(rdev
,
2297 r10_bio
->devs
[sl
].addr
+
2300 conf
->tmppage
, READ
, 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(In_sync
, &rdev
->flags
))
2347 atomic_inc(&rdev
->nr_pending
);
2349 if (r10_sync_page_io(rdev
,
2350 r10_bio
->devs
[sl
].addr
+
2352 s
, conf
->tmppage
, WRITE
)
2354 /* Well, this device is dead */
2356 "md/raid10:%s: read correction "
2358 " (%d sectors at %llu on %s)\n",
2360 (unsigned long long)(
2362 choose_data_offset(r10_bio
,
2364 bdevname(rdev
->bdev
, b
));
2365 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2368 bdevname(rdev
->bdev
, b
));
2370 rdev_dec_pending(rdev
, mddev
);
2374 while (sl
!= r10_bio
->read_slot
) {
2375 char b
[BDEVNAME_SIZE
];
2380 d
= r10_bio
->devs
[sl
].devnum
;
2381 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2383 !test_bit(In_sync
, &rdev
->flags
))
2386 atomic_inc(&rdev
->nr_pending
);
2388 switch (r10_sync_page_io(rdev
,
2389 r10_bio
->devs
[sl
].addr
+
2394 /* Well, this device is dead */
2396 "md/raid10:%s: unable to read back "
2398 " (%d sectors at %llu on %s)\n",
2400 (unsigned long long)(
2402 choose_data_offset(r10_bio
, rdev
)),
2403 bdevname(rdev
->bdev
, b
));
2404 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2407 bdevname(rdev
->bdev
, b
));
2411 "md/raid10:%s: read error corrected"
2412 " (%d sectors at %llu on %s)\n",
2414 (unsigned long long)(
2416 choose_data_offset(r10_bio
, rdev
)),
2417 bdevname(rdev
->bdev
, b
));
2418 atomic_add(s
, &rdev
->corrected_errors
);
2421 rdev_dec_pending(rdev
, mddev
);
2431 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2433 struct bio
*bio
= r10_bio
->master_bio
;
2434 struct mddev
*mddev
= r10_bio
->mddev
;
2435 struct r10conf
*conf
= mddev
->private;
2436 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2437 /* bio has the data to be written to slot 'i' where
2438 * we just recently had a write error.
2439 * We repeatedly clone the bio and trim down to one block,
2440 * then try the write. Where the write fails we record
2442 * It is conceivable that the bio doesn't exactly align with
2443 * blocks. We must handle this.
2445 * We currently own a reference to the rdev.
2451 int sect_to_write
= r10_bio
->sectors
;
2454 if (rdev
->badblocks
.shift
< 0)
2457 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2458 bdev_logical_block_size(rdev
->bdev
) >> 9);
2459 sector
= r10_bio
->sector
;
2460 sectors
= ((r10_bio
->sector
+ block_sectors
)
2461 & ~(sector_t
)(block_sectors
- 1))
2464 while (sect_to_write
) {
2466 if (sectors
> sect_to_write
)
2467 sectors
= sect_to_write
;
2468 /* Write at 'sector' for 'sectors' */
2469 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2470 bio_trim(wbio
, sector
- bio
->bi_iter
.bi_sector
, sectors
);
2471 wbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
2472 choose_data_offset(r10_bio
, rdev
) +
2473 (sector
- r10_bio
->sector
));
2474 wbio
->bi_bdev
= rdev
->bdev
;
2475 if (submit_bio_wait(WRITE
, wbio
) < 0)
2477 ok
= rdev_set_badblocks(rdev
, sector
,
2482 sect_to_write
-= sectors
;
2484 sectors
= block_sectors
;
2489 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2491 int slot
= r10_bio
->read_slot
;
2493 struct r10conf
*conf
= mddev
->private;
2494 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2495 char b
[BDEVNAME_SIZE
];
2496 unsigned long do_sync
;
2499 /* we got a read error. Maybe the drive is bad. Maybe just
2500 * the block and we can fix it.
2501 * We freeze all other IO, and try reading the block from
2502 * other devices. When we find one, we re-write
2503 * and check it that fixes the read error.
2504 * This is all done synchronously while the array is
2507 bio
= r10_bio
->devs
[slot
].bio
;
2508 bdevname(bio
->bi_bdev
, b
);
2510 r10_bio
->devs
[slot
].bio
= NULL
;
2512 if (mddev
->ro
== 0) {
2513 freeze_array(conf
, 1);
2514 fix_read_error(conf
, mddev
, r10_bio
);
2515 unfreeze_array(conf
);
2517 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2519 rdev_dec_pending(rdev
, mddev
);
2522 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2524 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2525 " read error for block %llu\n",
2527 (unsigned long long)r10_bio
->sector
);
2528 raid_end_bio_io(r10_bio
);
2532 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2533 slot
= r10_bio
->read_slot
;
2536 "md/raid10:%s: %s: redirecting "
2537 "sector %llu to another mirror\n",
2539 bdevname(rdev
->bdev
, b
),
2540 (unsigned long long)r10_bio
->sector
);
2541 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2543 bio_trim(bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
2544 r10_bio
->devs
[slot
].bio
= bio
;
2545 r10_bio
->devs
[slot
].rdev
= rdev
;
2546 bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
2547 + choose_data_offset(r10_bio
, rdev
);
2548 bio
->bi_bdev
= rdev
->bdev
;
2549 bio
->bi_rw
= READ
| do_sync
;
2550 bio
->bi_private
= r10_bio
;
2551 bio
->bi_end_io
= raid10_end_read_request
;
2552 if (max_sectors
< r10_bio
->sectors
) {
2553 /* Drat - have to split this up more */
2554 struct bio
*mbio
= r10_bio
->master_bio
;
2555 int sectors_handled
=
2556 r10_bio
->sector
+ max_sectors
2557 - mbio
->bi_iter
.bi_sector
;
2558 r10_bio
->sectors
= max_sectors
;
2559 spin_lock_irq(&conf
->device_lock
);
2560 if (mbio
->bi_phys_segments
== 0)
2561 mbio
->bi_phys_segments
= 2;
2563 mbio
->bi_phys_segments
++;
2564 spin_unlock_irq(&conf
->device_lock
);
2565 generic_make_request(bio
);
2567 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2569 r10_bio
->master_bio
= mbio
;
2570 r10_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2572 set_bit(R10BIO_ReadError
,
2574 r10_bio
->mddev
= mddev
;
2575 r10_bio
->sector
= mbio
->bi_iter
.bi_sector
2580 generic_make_request(bio
);
2583 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2585 /* Some sort of write request has finished and it
2586 * succeeded in writing where we thought there was a
2587 * bad block. So forget the bad block.
2588 * Or possibly if failed and we need to record
2592 struct md_rdev
*rdev
;
2594 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2595 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2596 for (m
= 0; m
< conf
->copies
; m
++) {
2597 int dev
= r10_bio
->devs
[m
].devnum
;
2598 rdev
= conf
->mirrors
[dev
].rdev
;
2599 if (r10_bio
->devs
[m
].bio
== NULL
)
2601 if (!r10_bio
->devs
[m
].bio
->bi_error
) {
2602 rdev_clear_badblocks(
2604 r10_bio
->devs
[m
].addr
,
2605 r10_bio
->sectors
, 0);
2607 if (!rdev_set_badblocks(
2609 r10_bio
->devs
[m
].addr
,
2610 r10_bio
->sectors
, 0))
2611 md_error(conf
->mddev
, rdev
);
2613 rdev
= conf
->mirrors
[dev
].replacement
;
2614 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2617 if (!r10_bio
->devs
[m
].repl_bio
->bi_error
) {
2618 rdev_clear_badblocks(
2620 r10_bio
->devs
[m
].addr
,
2621 r10_bio
->sectors
, 0);
2623 if (!rdev_set_badblocks(
2625 r10_bio
->devs
[m
].addr
,
2626 r10_bio
->sectors
, 0))
2627 md_error(conf
->mddev
, rdev
);
2633 for (m
= 0; m
< conf
->copies
; m
++) {
2634 int dev
= r10_bio
->devs
[m
].devnum
;
2635 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2636 rdev
= conf
->mirrors
[dev
].rdev
;
2637 if (bio
== IO_MADE_GOOD
) {
2638 rdev_clear_badblocks(
2640 r10_bio
->devs
[m
].addr
,
2641 r10_bio
->sectors
, 0);
2642 rdev_dec_pending(rdev
, conf
->mddev
);
2643 } else if (bio
!= NULL
&& bio
->bi_error
) {
2645 if (!narrow_write_error(r10_bio
, m
)) {
2646 md_error(conf
->mddev
, rdev
);
2647 set_bit(R10BIO_Degraded
,
2650 rdev_dec_pending(rdev
, conf
->mddev
);
2652 bio
= r10_bio
->devs
[m
].repl_bio
;
2653 rdev
= conf
->mirrors
[dev
].replacement
;
2654 if (rdev
&& bio
== IO_MADE_GOOD
) {
2655 rdev_clear_badblocks(
2657 r10_bio
->devs
[m
].addr
,
2658 r10_bio
->sectors
, 0);
2659 rdev_dec_pending(rdev
, conf
->mddev
);
2663 spin_lock_irq(&conf
->device_lock
);
2664 list_add(&r10_bio
->retry_list
, &conf
->bio_end_io_list
);
2665 spin_unlock_irq(&conf
->device_lock
);
2666 md_wakeup_thread(conf
->mddev
->thread
);
2668 if (test_bit(R10BIO_WriteError
,
2670 close_write(r10_bio
);
2671 raid_end_bio_io(r10_bio
);
2676 static void raid10d(struct md_thread
*thread
)
2678 struct mddev
*mddev
= thread
->mddev
;
2679 struct r10bio
*r10_bio
;
2680 unsigned long flags
;
2681 struct r10conf
*conf
= mddev
->private;
2682 struct list_head
*head
= &conf
->retry_list
;
2683 struct blk_plug plug
;
2685 md_check_recovery(mddev
);
2687 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2688 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2690 spin_lock_irqsave(&conf
->device_lock
, flags
);
2691 if (!test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2692 list_add(&tmp
, &conf
->bio_end_io_list
);
2693 list_del_init(&conf
->bio_end_io_list
);
2695 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2696 while (!list_empty(&tmp
)) {
2697 r10_bio
= list_first_entry(&tmp
, struct r10bio
,
2699 list_del(&r10_bio
->retry_list
);
2700 if (mddev
->degraded
)
2701 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
2703 if (test_bit(R10BIO_WriteError
,
2705 close_write(r10_bio
);
2706 raid_end_bio_io(r10_bio
);
2710 blk_start_plug(&plug
);
2713 flush_pending_writes(conf
);
2715 spin_lock_irqsave(&conf
->device_lock
, flags
);
2716 if (list_empty(head
)) {
2717 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2720 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2721 list_del(head
->prev
);
2723 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2725 mddev
= r10_bio
->mddev
;
2726 conf
= mddev
->private;
2727 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2728 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2729 handle_write_completed(conf
, r10_bio
);
2730 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2731 reshape_request_write(mddev
, r10_bio
);
2732 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2733 sync_request_write(mddev
, r10_bio
);
2734 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2735 recovery_request_write(mddev
, r10_bio
);
2736 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2737 handle_read_error(mddev
, r10_bio
);
2739 /* just a partial read to be scheduled from a
2742 int slot
= r10_bio
->read_slot
;
2743 generic_make_request(r10_bio
->devs
[slot
].bio
);
2747 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2748 md_check_recovery(mddev
);
2750 blk_finish_plug(&plug
);
2753 static int init_resync(struct r10conf
*conf
)
2758 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2759 BUG_ON(conf
->r10buf_pool
);
2760 conf
->have_replacement
= 0;
2761 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2762 if (conf
->mirrors
[i
].replacement
)
2763 conf
->have_replacement
= 1;
2764 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2765 if (!conf
->r10buf_pool
)
2767 conf
->next_resync
= 0;
2772 * perform a "sync" on one "block"
2774 * We need to make sure that no normal I/O request - particularly write
2775 * requests - conflict with active sync requests.
2777 * This is achieved by tracking pending requests and a 'barrier' concept
2778 * that can be installed to exclude normal IO requests.
2780 * Resync and recovery are handled very differently.
2781 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2783 * For resync, we iterate over virtual addresses, read all copies,
2784 * and update if there are differences. If only one copy is live,
2786 * For recovery, we iterate over physical addresses, read a good
2787 * value for each non-in_sync drive, and over-write.
2789 * So, for recovery we may have several outstanding complex requests for a
2790 * given address, one for each out-of-sync device. We model this by allocating
2791 * a number of r10_bio structures, one for each out-of-sync device.
2792 * As we setup these structures, we collect all bio's together into a list
2793 * which we then process collectively to add pages, and then process again
2794 * to pass to generic_make_request.
2796 * The r10_bio structures are linked using a borrowed master_bio pointer.
2797 * This link is counted in ->remaining. When the r10_bio that points to NULL
2798 * has its remaining count decremented to 0, the whole complex operation
2803 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2806 struct r10conf
*conf
= mddev
->private;
2807 struct r10bio
*r10_bio
;
2808 struct bio
*biolist
= NULL
, *bio
;
2809 sector_t max_sector
, nr_sectors
;
2812 sector_t sync_blocks
;
2813 sector_t sectors_skipped
= 0;
2814 int chunks_skipped
= 0;
2815 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2817 if (!conf
->r10buf_pool
)
2818 if (init_resync(conf
))
2822 * Allow skipping a full rebuild for incremental assembly
2823 * of a clean array, like RAID1 does.
2825 if (mddev
->bitmap
== NULL
&&
2826 mddev
->recovery_cp
== MaxSector
&&
2827 mddev
->reshape_position
== MaxSector
&&
2828 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2829 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2830 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2831 conf
->fullsync
== 0) {
2833 return mddev
->dev_sectors
- sector_nr
;
2837 max_sector
= mddev
->dev_sectors
;
2838 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2839 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2840 max_sector
= mddev
->resync_max_sectors
;
2841 if (sector_nr
>= max_sector
) {
2842 /* If we aborted, we need to abort the
2843 * sync on the 'current' bitmap chucks (there can
2844 * be several when recovering multiple devices).
2845 * as we may have started syncing it but not finished.
2846 * We can find the current address in
2847 * mddev->curr_resync, but for recovery,
2848 * we need to convert that to several
2849 * virtual addresses.
2851 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2857 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2858 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2859 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2861 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2863 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2864 bitmap_end_sync(mddev
->bitmap
, sect
,
2868 /* completed sync */
2869 if ((!mddev
->bitmap
|| conf
->fullsync
)
2870 && conf
->have_replacement
2871 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2872 /* Completed a full sync so the replacements
2873 * are now fully recovered.
2875 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2876 if (conf
->mirrors
[i
].replacement
)
2877 conf
->mirrors
[i
].replacement
2883 bitmap_close_sync(mddev
->bitmap
);
2886 return sectors_skipped
;
2889 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2890 return reshape_request(mddev
, sector_nr
, skipped
);
2892 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2893 /* if there has been nothing to do on any drive,
2894 * then there is nothing to do at all..
2897 return (max_sector
- sector_nr
) + sectors_skipped
;
2900 if (max_sector
> mddev
->resync_max
)
2901 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2903 /* make sure whole request will fit in a chunk - if chunks
2906 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2907 max_sector
> (sector_nr
| chunk_mask
))
2908 max_sector
= (sector_nr
| chunk_mask
) + 1;
2910 /* Again, very different code for resync and recovery.
2911 * Both must result in an r10bio with a list of bios that
2912 * have bi_end_io, bi_sector, bi_bdev set,
2913 * and bi_private set to the r10bio.
2914 * For recovery, we may actually create several r10bios
2915 * with 2 bios in each, that correspond to the bios in the main one.
2916 * In this case, the subordinate r10bios link back through a
2917 * borrowed master_bio pointer, and the counter in the master
2918 * includes a ref from each subordinate.
2920 /* First, we decide what to do and set ->bi_end_io
2921 * To end_sync_read if we want to read, and
2922 * end_sync_write if we will want to write.
2925 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2926 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2927 /* recovery... the complicated one */
2931 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2937 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2939 if ((mirror
->rdev
== NULL
||
2940 test_bit(In_sync
, &mirror
->rdev
->flags
))
2942 (mirror
->replacement
== NULL
||
2944 &mirror
->replacement
->flags
)))
2948 /* want to reconstruct this device */
2950 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2951 if (sect
>= mddev
->resync_max_sectors
) {
2952 /* last stripe is not complete - don't
2953 * try to recover this sector.
2957 /* Unless we are doing a full sync, or a replacement
2958 * we only need to recover the block if it is set in
2961 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2963 if (sync_blocks
< max_sync
)
2964 max_sync
= sync_blocks
;
2966 mirror
->replacement
== NULL
&&
2968 /* yep, skip the sync_blocks here, but don't assume
2969 * that there will never be anything to do here
2971 chunks_skipped
= -1;
2975 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2977 raise_barrier(conf
, rb2
!= NULL
);
2978 atomic_set(&r10_bio
->remaining
, 0);
2980 r10_bio
->master_bio
= (struct bio
*)rb2
;
2982 atomic_inc(&rb2
->remaining
);
2983 r10_bio
->mddev
= mddev
;
2984 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
2985 r10_bio
->sector
= sect
;
2987 raid10_find_phys(conf
, r10_bio
);
2989 /* Need to check if the array will still be
2992 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
2993 if (conf
->mirrors
[j
].rdev
== NULL
||
2994 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
2999 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3000 &sync_blocks
, still_degraded
);
3003 for (j
=0; j
<conf
->copies
;j
++) {
3005 int d
= r10_bio
->devs
[j
].devnum
;
3006 sector_t from_addr
, to_addr
;
3007 struct md_rdev
*rdev
;
3008 sector_t sector
, first_bad
;
3010 if (!conf
->mirrors
[d
].rdev
||
3011 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3013 /* This is where we read from */
3015 rdev
= conf
->mirrors
[d
].rdev
;
3016 sector
= r10_bio
->devs
[j
].addr
;
3018 if (is_badblock(rdev
, sector
, max_sync
,
3019 &first_bad
, &bad_sectors
)) {
3020 if (first_bad
> sector
)
3021 max_sync
= first_bad
- sector
;
3023 bad_sectors
-= (sector
3025 if (max_sync
> bad_sectors
)
3026 max_sync
= bad_sectors
;
3030 bio
= r10_bio
->devs
[0].bio
;
3032 bio
->bi_next
= biolist
;
3034 bio
->bi_private
= r10_bio
;
3035 bio
->bi_end_io
= end_sync_read
;
3037 from_addr
= r10_bio
->devs
[j
].addr
;
3038 bio
->bi_iter
.bi_sector
= from_addr
+
3040 bio
->bi_bdev
= rdev
->bdev
;
3041 atomic_inc(&rdev
->nr_pending
);
3042 /* and we write to 'i' (if not in_sync) */
3044 for (k
=0; k
<conf
->copies
; k
++)
3045 if (r10_bio
->devs
[k
].devnum
== i
)
3047 BUG_ON(k
== conf
->copies
);
3048 to_addr
= r10_bio
->devs
[k
].addr
;
3049 r10_bio
->devs
[0].devnum
= d
;
3050 r10_bio
->devs
[0].addr
= from_addr
;
3051 r10_bio
->devs
[1].devnum
= i
;
3052 r10_bio
->devs
[1].addr
= to_addr
;
3054 rdev
= mirror
->rdev
;
3055 if (!test_bit(In_sync
, &rdev
->flags
)) {
3056 bio
= r10_bio
->devs
[1].bio
;
3058 bio
->bi_next
= biolist
;
3060 bio
->bi_private
= r10_bio
;
3061 bio
->bi_end_io
= end_sync_write
;
3063 bio
->bi_iter
.bi_sector
= to_addr
3064 + rdev
->data_offset
;
3065 bio
->bi_bdev
= rdev
->bdev
;
3066 atomic_inc(&r10_bio
->remaining
);
3068 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3070 /* and maybe write to replacement */
3071 bio
= r10_bio
->devs
[1].repl_bio
;
3073 bio
->bi_end_io
= NULL
;
3074 rdev
= mirror
->replacement
;
3075 /* Note: if rdev != NULL, then bio
3076 * cannot be NULL as r10buf_pool_alloc will
3077 * have allocated it.
3078 * So the second test here is pointless.
3079 * But it keeps semantic-checkers happy, and
3080 * this comment keeps human reviewers
3083 if (rdev
== NULL
|| bio
== NULL
||
3084 test_bit(Faulty
, &rdev
->flags
))
3087 bio
->bi_next
= biolist
;
3089 bio
->bi_private
= r10_bio
;
3090 bio
->bi_end_io
= end_sync_write
;
3092 bio
->bi_iter
.bi_sector
= to_addr
+
3094 bio
->bi_bdev
= rdev
->bdev
;
3095 atomic_inc(&r10_bio
->remaining
);
3098 if (j
== conf
->copies
) {
3099 /* Cannot recover, so abort the recovery or
3100 * record a bad block */
3102 /* problem is that there are bad blocks
3103 * on other device(s)
3106 for (k
= 0; k
< conf
->copies
; k
++)
3107 if (r10_bio
->devs
[k
].devnum
== i
)
3109 if (!test_bit(In_sync
,
3110 &mirror
->rdev
->flags
)
3111 && !rdev_set_badblocks(
3113 r10_bio
->devs
[k
].addr
,
3116 if (mirror
->replacement
&&
3117 !rdev_set_badblocks(
3118 mirror
->replacement
,
3119 r10_bio
->devs
[k
].addr
,
3124 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3126 printk(KERN_INFO
"md/raid10:%s: insufficient "
3127 "working devices for recovery.\n",
3129 mirror
->recovery_disabled
3130 = mddev
->recovery_disabled
;
3134 atomic_dec(&rb2
->remaining
);
3139 if (biolist
== NULL
) {
3141 struct r10bio
*rb2
= r10_bio
;
3142 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3143 rb2
->master_bio
= NULL
;
3149 /* resync. Schedule a read for every block at this virt offset */
3152 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, 0);
3154 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3155 &sync_blocks
, mddev
->degraded
) &&
3156 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3157 &mddev
->recovery
)) {
3158 /* We can skip this block */
3160 return sync_blocks
+ sectors_skipped
;
3162 if (sync_blocks
< max_sync
)
3163 max_sync
= sync_blocks
;
3164 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3167 r10_bio
->mddev
= mddev
;
3168 atomic_set(&r10_bio
->remaining
, 0);
3169 raise_barrier(conf
, 0);
3170 conf
->next_resync
= sector_nr
;
3172 r10_bio
->master_bio
= NULL
;
3173 r10_bio
->sector
= sector_nr
;
3174 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3175 raid10_find_phys(conf
, r10_bio
);
3176 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3178 for (i
= 0; i
< conf
->copies
; i
++) {
3179 int d
= r10_bio
->devs
[i
].devnum
;
3180 sector_t first_bad
, sector
;
3183 if (r10_bio
->devs
[i
].repl_bio
)
3184 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3186 bio
= r10_bio
->devs
[i
].bio
;
3188 bio
->bi_error
= -EIO
;
3189 if (conf
->mirrors
[d
].rdev
== NULL
||
3190 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3192 sector
= r10_bio
->devs
[i
].addr
;
3193 if (is_badblock(conf
->mirrors
[d
].rdev
,
3195 &first_bad
, &bad_sectors
)) {
3196 if (first_bad
> sector
)
3197 max_sync
= first_bad
- sector
;
3199 bad_sectors
-= (sector
- first_bad
);
3200 if (max_sync
> bad_sectors
)
3201 max_sync
= bad_sectors
;
3205 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3206 atomic_inc(&r10_bio
->remaining
);
3207 bio
->bi_next
= biolist
;
3209 bio
->bi_private
= r10_bio
;
3210 bio
->bi_end_io
= end_sync_read
;
3212 bio
->bi_iter
.bi_sector
= sector
+
3213 conf
->mirrors
[d
].rdev
->data_offset
;
3214 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3217 if (conf
->mirrors
[d
].replacement
== NULL
||
3219 &conf
->mirrors
[d
].replacement
->flags
))
3222 /* Need to set up for writing to the replacement */
3223 bio
= r10_bio
->devs
[i
].repl_bio
;
3225 bio
->bi_error
= -EIO
;
3227 sector
= r10_bio
->devs
[i
].addr
;
3228 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3229 bio
->bi_next
= biolist
;
3231 bio
->bi_private
= r10_bio
;
3232 bio
->bi_end_io
= end_sync_write
;
3234 bio
->bi_iter
.bi_sector
= sector
+
3235 conf
->mirrors
[d
].replacement
->data_offset
;
3236 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3241 for (i
=0; i
<conf
->copies
; i
++) {
3242 int d
= r10_bio
->devs
[i
].devnum
;
3243 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3244 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3246 if (r10_bio
->devs
[i
].repl_bio
&&
3247 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3249 conf
->mirrors
[d
].replacement
,
3259 if (sector_nr
+ max_sync
< max_sector
)
3260 max_sector
= sector_nr
+ max_sync
;
3263 int len
= PAGE_SIZE
;
3264 if (sector_nr
+ (len
>>9) > max_sector
)
3265 len
= (max_sector
- sector_nr
) << 9;
3268 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3270 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3271 if (bio_add_page(bio
, page
, len
, 0))
3275 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3276 for (bio2
= biolist
;
3277 bio2
&& bio2
!= bio
;
3278 bio2
= bio2
->bi_next
) {
3279 /* remove last page from this bio */
3281 bio2
->bi_iter
.bi_size
-= len
;
3282 bio_clear_flag(bio2
, BIO_SEG_VALID
);
3286 nr_sectors
+= len
>>9;
3287 sector_nr
+= len
>>9;
3288 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3290 r10_bio
->sectors
= nr_sectors
;
3294 biolist
= biolist
->bi_next
;
3296 bio
->bi_next
= NULL
;
3297 r10_bio
= bio
->bi_private
;
3298 r10_bio
->sectors
= nr_sectors
;
3300 if (bio
->bi_end_io
== end_sync_read
) {
3301 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3303 generic_make_request(bio
);
3307 if (sectors_skipped
)
3308 /* pretend they weren't skipped, it makes
3309 * no important difference in this case
3311 md_done_sync(mddev
, sectors_skipped
, 1);
3313 return sectors_skipped
+ nr_sectors
;
3315 /* There is nowhere to write, so all non-sync
3316 * drives must be failed or in resync, all drives
3317 * have a bad block, so try the next chunk...
3319 if (sector_nr
+ max_sync
< max_sector
)
3320 max_sector
= sector_nr
+ max_sync
;
3322 sectors_skipped
+= (max_sector
- sector_nr
);
3324 sector_nr
= max_sector
;
3329 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3332 struct r10conf
*conf
= mddev
->private;
3335 raid_disks
= min(conf
->geo
.raid_disks
,
3336 conf
->prev
.raid_disks
);
3338 sectors
= conf
->dev_sectors
;
3340 size
= sectors
>> conf
->geo
.chunk_shift
;
3341 sector_div(size
, conf
->geo
.far_copies
);
3342 size
= size
* raid_disks
;
3343 sector_div(size
, conf
->geo
.near_copies
);
3345 return size
<< conf
->geo
.chunk_shift
;
3348 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3350 /* Calculate the number of sectors-per-device that will
3351 * actually be used, and set conf->dev_sectors and
3355 size
= size
>> conf
->geo
.chunk_shift
;
3356 sector_div(size
, conf
->geo
.far_copies
);
3357 size
= size
* conf
->geo
.raid_disks
;
3358 sector_div(size
, conf
->geo
.near_copies
);
3359 /* 'size' is now the number of chunks in the array */
3360 /* calculate "used chunks per device" */
3361 size
= size
* conf
->copies
;
3363 /* We need to round up when dividing by raid_disks to
3364 * get the stride size.
3366 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3368 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3370 if (conf
->geo
.far_offset
)
3371 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3373 sector_div(size
, conf
->geo
.far_copies
);
3374 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3378 enum geo_type
{geo_new
, geo_old
, geo_start
};
3379 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3382 int layout
, chunk
, disks
;
3385 layout
= mddev
->layout
;
3386 chunk
= mddev
->chunk_sectors
;
3387 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3390 layout
= mddev
->new_layout
;
3391 chunk
= mddev
->new_chunk_sectors
;
3392 disks
= mddev
->raid_disks
;
3394 default: /* avoid 'may be unused' warnings */
3395 case geo_start
: /* new when starting reshape - raid_disks not
3397 layout
= mddev
->new_layout
;
3398 chunk
= mddev
->new_chunk_sectors
;
3399 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3404 if (chunk
< (PAGE_SIZE
>> 9) ||
3405 !is_power_of_2(chunk
))
3408 fc
= (layout
>> 8) & 255;
3409 fo
= layout
& (1<<16);
3410 geo
->raid_disks
= disks
;
3411 geo
->near_copies
= nc
;
3412 geo
->far_copies
= fc
;
3413 geo
->far_offset
= fo
;
3414 switch (layout
>> 17) {
3415 case 0: /* original layout. simple but not always optimal */
3416 geo
->far_set_size
= disks
;
3418 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3419 * actually using this, but leave code here just in case.*/
3420 geo
->far_set_size
= disks
/fc
;
3421 WARN(geo
->far_set_size
< fc
,
3422 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3424 case 2: /* "improved" layout fixed to match documentation */
3425 geo
->far_set_size
= fc
* nc
;
3427 default: /* Not a valid layout */
3430 geo
->chunk_mask
= chunk
- 1;
3431 geo
->chunk_shift
= ffz(~chunk
);
3435 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3437 struct r10conf
*conf
= NULL
;
3442 copies
= setup_geo(&geo
, mddev
, geo_new
);
3445 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3446 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3447 mdname(mddev
), PAGE_SIZE
);
3451 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3452 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3453 mdname(mddev
), mddev
->new_layout
);
3458 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3462 /* FIXME calc properly */
3463 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3464 max(0,-mddev
->delta_disks
)),
3469 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3474 conf
->copies
= copies
;
3475 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3476 r10bio_pool_free
, conf
);
3477 if (!conf
->r10bio_pool
)
3480 calc_sectors(conf
, mddev
->dev_sectors
);
3481 if (mddev
->reshape_position
== MaxSector
) {
3482 conf
->prev
= conf
->geo
;
3483 conf
->reshape_progress
= MaxSector
;
3485 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3489 conf
->reshape_progress
= mddev
->reshape_position
;
3490 if (conf
->prev
.far_offset
)
3491 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3493 /* far_copies must be 1 */
3494 conf
->prev
.stride
= conf
->dev_sectors
;
3496 conf
->reshape_safe
= conf
->reshape_progress
;
3497 spin_lock_init(&conf
->device_lock
);
3498 INIT_LIST_HEAD(&conf
->retry_list
);
3499 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
3501 spin_lock_init(&conf
->resync_lock
);
3502 init_waitqueue_head(&conf
->wait_barrier
);
3504 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3508 conf
->mddev
= mddev
;
3513 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3516 mempool_destroy(conf
->r10bio_pool
);
3517 kfree(conf
->mirrors
);
3518 safe_put_page(conf
->tmppage
);
3521 return ERR_PTR(err
);
3524 static int run(struct mddev
*mddev
)
3526 struct r10conf
*conf
;
3527 int i
, disk_idx
, chunk_size
;
3528 struct raid10_info
*disk
;
3529 struct md_rdev
*rdev
;
3531 sector_t min_offset_diff
= 0;
3533 bool discard_supported
= false;
3535 if (mddev
->private == NULL
) {
3536 conf
= setup_conf(mddev
);
3538 return PTR_ERR(conf
);
3539 mddev
->private = conf
;
3541 conf
= mddev
->private;
3545 mddev
->thread
= conf
->thread
;
3546 conf
->thread
= NULL
;
3548 chunk_size
= mddev
->chunk_sectors
<< 9;
3550 blk_queue_max_discard_sectors(mddev
->queue
,
3551 mddev
->chunk_sectors
);
3552 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3553 blk_queue_io_min(mddev
->queue
, chunk_size
);
3554 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3555 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3557 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3558 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3561 rdev_for_each(rdev
, mddev
) {
3563 struct request_queue
*q
;
3565 disk_idx
= rdev
->raid_disk
;
3568 if (disk_idx
>= conf
->geo
.raid_disks
&&
3569 disk_idx
>= conf
->prev
.raid_disks
)
3571 disk
= conf
->mirrors
+ disk_idx
;
3573 if (test_bit(Replacement
, &rdev
->flags
)) {
3574 if (disk
->replacement
)
3576 disk
->replacement
= rdev
;
3582 q
= bdev_get_queue(rdev
->bdev
);
3583 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3584 if (!mddev
->reshape_backwards
)
3588 if (first
|| diff
< min_offset_diff
)
3589 min_offset_diff
= diff
;
3592 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3593 rdev
->data_offset
<< 9);
3595 disk
->head_position
= 0;
3597 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3598 discard_supported
= true;
3602 if (discard_supported
)
3603 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3606 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3609 /* need to check that every block has at least one working mirror */
3610 if (!enough(conf
, -1)) {
3611 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3616 if (conf
->reshape_progress
!= MaxSector
) {
3617 /* must ensure that shape change is supported */
3618 if (conf
->geo
.far_copies
!= 1 &&
3619 conf
->geo
.far_offset
== 0)
3621 if (conf
->prev
.far_copies
!= 1 &&
3622 conf
->prev
.far_offset
== 0)
3626 mddev
->degraded
= 0;
3628 i
< conf
->geo
.raid_disks
3629 || i
< conf
->prev
.raid_disks
;
3632 disk
= conf
->mirrors
+ i
;
3634 if (!disk
->rdev
&& disk
->replacement
) {
3635 /* The replacement is all we have - use it */
3636 disk
->rdev
= disk
->replacement
;
3637 disk
->replacement
= NULL
;
3638 clear_bit(Replacement
, &disk
->rdev
->flags
);
3642 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3643 disk
->head_position
= 0;
3646 disk
->rdev
->saved_raid_disk
< 0)
3649 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3652 if (mddev
->recovery_cp
!= MaxSector
)
3653 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3654 " -- starting background reconstruction\n",
3657 "md/raid10:%s: active with %d out of %d devices\n",
3658 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3659 conf
->geo
.raid_disks
);
3661 * Ok, everything is just fine now
3663 mddev
->dev_sectors
= conf
->dev_sectors
;
3664 size
= raid10_size(mddev
, 0, 0);
3665 md_set_array_sectors(mddev
, size
);
3666 mddev
->resync_max_sectors
= size
;
3669 int stripe
= conf
->geo
.raid_disks
*
3670 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3672 /* Calculate max read-ahead size.
3673 * We need to readahead at least twice a whole stripe....
3676 stripe
/= conf
->geo
.near_copies
;
3677 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3678 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3681 if (md_integrity_register(mddev
))
3684 if (conf
->reshape_progress
!= MaxSector
) {
3685 unsigned long before_length
, after_length
;
3687 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3688 conf
->prev
.far_copies
);
3689 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3690 conf
->geo
.far_copies
);
3692 if (max(before_length
, after_length
) > min_offset_diff
) {
3693 /* This cannot work */
3694 printk("md/raid10: offset difference not enough to continue reshape\n");
3697 conf
->offset_diff
= min_offset_diff
;
3699 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3700 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3701 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3702 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3703 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3710 md_unregister_thread(&mddev
->thread
);
3711 mempool_destroy(conf
->r10bio_pool
);
3712 safe_put_page(conf
->tmppage
);
3713 kfree(conf
->mirrors
);
3715 mddev
->private = NULL
;
3720 static void raid10_free(struct mddev
*mddev
, void *priv
)
3722 struct r10conf
*conf
= priv
;
3724 mempool_destroy(conf
->r10bio_pool
);
3725 safe_put_page(conf
->tmppage
);
3726 kfree(conf
->mirrors
);
3727 kfree(conf
->mirrors_old
);
3728 kfree(conf
->mirrors_new
);
3732 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3734 struct r10conf
*conf
= mddev
->private;
3738 raise_barrier(conf
, 0);
3741 lower_barrier(conf
);
3746 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3748 /* Resize of 'far' arrays is not supported.
3749 * For 'near' and 'offset' arrays we can set the
3750 * number of sectors used to be an appropriate multiple
3751 * of the chunk size.
3752 * For 'offset', this is far_copies*chunksize.
3753 * For 'near' the multiplier is the LCM of
3754 * near_copies and raid_disks.
3755 * So if far_copies > 1 && !far_offset, fail.
3756 * Else find LCM(raid_disks, near_copy)*far_copies and
3757 * multiply by chunk_size. Then round to this number.
3758 * This is mostly done by raid10_size()
3760 struct r10conf
*conf
= mddev
->private;
3761 sector_t oldsize
, size
;
3763 if (mddev
->reshape_position
!= MaxSector
)
3766 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3769 oldsize
= raid10_size(mddev
, 0, 0);
3770 size
= raid10_size(mddev
, sectors
, 0);
3771 if (mddev
->external_size
&&
3772 mddev
->array_sectors
> size
)
3774 if (mddev
->bitmap
) {
3775 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3779 md_set_array_sectors(mddev
, size
);
3780 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3781 revalidate_disk(mddev
->gendisk
);
3782 if (sectors
> mddev
->dev_sectors
&&
3783 mddev
->recovery_cp
> oldsize
) {
3784 mddev
->recovery_cp
= oldsize
;
3785 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3787 calc_sectors(conf
, sectors
);
3788 mddev
->dev_sectors
= conf
->dev_sectors
;
3789 mddev
->resync_max_sectors
= size
;
3793 static void *raid10_takeover_raid0(struct mddev
*mddev
, sector_t size
, int devs
)
3795 struct md_rdev
*rdev
;
3796 struct r10conf
*conf
;
3798 if (mddev
->degraded
> 0) {
3799 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3801 return ERR_PTR(-EINVAL
);
3803 sector_div(size
, devs
);
3805 /* Set new parameters */
3806 mddev
->new_level
= 10;
3807 /* new layout: far_copies = 1, near_copies = 2 */
3808 mddev
->new_layout
= (1<<8) + 2;
3809 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3810 mddev
->delta_disks
= mddev
->raid_disks
;
3811 mddev
->raid_disks
*= 2;
3812 /* make sure it will be not marked as dirty */
3813 mddev
->recovery_cp
= MaxSector
;
3814 mddev
->dev_sectors
= size
;
3816 conf
= setup_conf(mddev
);
3817 if (!IS_ERR(conf
)) {
3818 rdev_for_each(rdev
, mddev
)
3819 if (rdev
->raid_disk
>= 0) {
3820 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3821 rdev
->sectors
= size
;
3829 static void *raid10_takeover(struct mddev
*mddev
)
3831 struct r0conf
*raid0_conf
;
3833 /* raid10 can take over:
3834 * raid0 - providing it has only two drives
3836 if (mddev
->level
== 0) {
3837 /* for raid0 takeover only one zone is supported */
3838 raid0_conf
= mddev
->private;
3839 if (raid0_conf
->nr_strip_zones
> 1) {
3840 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3841 " with more than one zone.\n",
3843 return ERR_PTR(-EINVAL
);
3845 return raid10_takeover_raid0(mddev
,
3846 raid0_conf
->strip_zone
->zone_end
,
3847 raid0_conf
->strip_zone
->nb_dev
);
3849 return ERR_PTR(-EINVAL
);
3852 static int raid10_check_reshape(struct mddev
*mddev
)
3854 /* Called when there is a request to change
3855 * - layout (to ->new_layout)
3856 * - chunk size (to ->new_chunk_sectors)
3857 * - raid_disks (by delta_disks)
3858 * or when trying to restart a reshape that was ongoing.
3860 * We need to validate the request and possibly allocate
3861 * space if that might be an issue later.
3863 * Currently we reject any reshape of a 'far' mode array,
3864 * allow chunk size to change if new is generally acceptable,
3865 * allow raid_disks to increase, and allow
3866 * a switch between 'near' mode and 'offset' mode.
3868 struct r10conf
*conf
= mddev
->private;
3871 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3874 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3875 /* mustn't change number of copies */
3877 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3878 /* Cannot switch to 'far' mode */
3881 if (mddev
->array_sectors
& geo
.chunk_mask
)
3882 /* not factor of array size */
3885 if (!enough(conf
, -1))
3888 kfree(conf
->mirrors_new
);
3889 conf
->mirrors_new
= NULL
;
3890 if (mddev
->delta_disks
> 0) {
3891 /* allocate new 'mirrors' list */
3892 conf
->mirrors_new
= kzalloc(
3893 sizeof(struct raid10_info
)
3894 *(mddev
->raid_disks
+
3895 mddev
->delta_disks
),
3897 if (!conf
->mirrors_new
)
3904 * Need to check if array has failed when deciding whether to:
3906 * - remove non-faulty devices
3909 * This determination is simple when no reshape is happening.
3910 * However if there is a reshape, we need to carefully check
3911 * both the before and after sections.
3912 * This is because some failed devices may only affect one
3913 * of the two sections, and some non-in_sync devices may
3914 * be insync in the section most affected by failed devices.
3916 static int calc_degraded(struct r10conf
*conf
)
3918 int degraded
, degraded2
;
3923 /* 'prev' section first */
3924 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
3925 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3926 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3928 else if (!test_bit(In_sync
, &rdev
->flags
))
3929 /* When we can reduce the number of devices in
3930 * an array, this might not contribute to
3931 * 'degraded'. It does now.
3936 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
3940 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3941 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3942 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3944 else if (!test_bit(In_sync
, &rdev
->flags
)) {
3945 /* If reshape is increasing the number of devices,
3946 * this section has already been recovered, so
3947 * it doesn't contribute to degraded.
3950 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
3955 if (degraded2
> degraded
)
3960 static int raid10_start_reshape(struct mddev
*mddev
)
3962 /* A 'reshape' has been requested. This commits
3963 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3964 * This also checks if there are enough spares and adds them
3966 * We currently require enough spares to make the final
3967 * array non-degraded. We also require that the difference
3968 * between old and new data_offset - on each device - is
3969 * enough that we never risk over-writing.
3972 unsigned long before_length
, after_length
;
3973 sector_t min_offset_diff
= 0;
3976 struct r10conf
*conf
= mddev
->private;
3977 struct md_rdev
*rdev
;
3981 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
3984 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
3987 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3988 conf
->prev
.far_copies
);
3989 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3990 conf
->geo
.far_copies
);
3992 rdev_for_each(rdev
, mddev
) {
3993 if (!test_bit(In_sync
, &rdev
->flags
)
3994 && !test_bit(Faulty
, &rdev
->flags
))
3996 if (rdev
->raid_disk
>= 0) {
3997 long long diff
= (rdev
->new_data_offset
3998 - rdev
->data_offset
);
3999 if (!mddev
->reshape_backwards
)
4003 if (first
|| diff
< min_offset_diff
)
4004 min_offset_diff
= diff
;
4008 if (max(before_length
, after_length
) > min_offset_diff
)
4011 if (spares
< mddev
->delta_disks
)
4014 conf
->offset_diff
= min_offset_diff
;
4015 spin_lock_irq(&conf
->device_lock
);
4016 if (conf
->mirrors_new
) {
4017 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4018 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4020 kfree(conf
->mirrors_old
);
4021 conf
->mirrors_old
= conf
->mirrors
;
4022 conf
->mirrors
= conf
->mirrors_new
;
4023 conf
->mirrors_new
= NULL
;
4025 setup_geo(&conf
->geo
, mddev
, geo_start
);
4027 if (mddev
->reshape_backwards
) {
4028 sector_t size
= raid10_size(mddev
, 0, 0);
4029 if (size
< mddev
->array_sectors
) {
4030 spin_unlock_irq(&conf
->device_lock
);
4031 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4035 mddev
->resync_max_sectors
= size
;
4036 conf
->reshape_progress
= size
;
4038 conf
->reshape_progress
= 0;
4039 conf
->reshape_safe
= conf
->reshape_progress
;
4040 spin_unlock_irq(&conf
->device_lock
);
4042 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4043 ret
= bitmap_resize(mddev
->bitmap
,
4044 raid10_size(mddev
, 0,
4045 conf
->geo
.raid_disks
),
4050 if (mddev
->delta_disks
> 0) {
4051 rdev_for_each(rdev
, mddev
)
4052 if (rdev
->raid_disk
< 0 &&
4053 !test_bit(Faulty
, &rdev
->flags
)) {
4054 if (raid10_add_disk(mddev
, rdev
) == 0) {
4055 if (rdev
->raid_disk
>=
4056 conf
->prev
.raid_disks
)
4057 set_bit(In_sync
, &rdev
->flags
);
4059 rdev
->recovery_offset
= 0;
4061 if (sysfs_link_rdev(mddev
, rdev
))
4062 /* Failure here is OK */;
4064 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4065 && !test_bit(Faulty
, &rdev
->flags
)) {
4066 /* This is a spare that was manually added */
4067 set_bit(In_sync
, &rdev
->flags
);
4070 /* When a reshape changes the number of devices,
4071 * ->degraded is measured against the larger of the
4072 * pre and post numbers.
4074 spin_lock_irq(&conf
->device_lock
);
4075 mddev
->degraded
= calc_degraded(conf
);
4076 spin_unlock_irq(&conf
->device_lock
);
4077 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4078 mddev
->reshape_position
= conf
->reshape_progress
;
4079 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4081 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4082 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4083 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
4084 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4085 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4087 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4089 if (!mddev
->sync_thread
) {
4093 conf
->reshape_checkpoint
= jiffies
;
4094 md_wakeup_thread(mddev
->sync_thread
);
4095 md_new_event(mddev
);
4099 mddev
->recovery
= 0;
4100 spin_lock_irq(&conf
->device_lock
);
4101 conf
->geo
= conf
->prev
;
4102 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4103 rdev_for_each(rdev
, mddev
)
4104 rdev
->new_data_offset
= rdev
->data_offset
;
4106 conf
->reshape_progress
= MaxSector
;
4107 conf
->reshape_safe
= MaxSector
;
4108 mddev
->reshape_position
= MaxSector
;
4109 spin_unlock_irq(&conf
->device_lock
);
4113 /* Calculate the last device-address that could contain
4114 * any block from the chunk that includes the array-address 's'
4115 * and report the next address.
4116 * i.e. the address returned will be chunk-aligned and after
4117 * any data that is in the chunk containing 's'.
4119 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4121 s
= (s
| geo
->chunk_mask
) + 1;
4122 s
>>= geo
->chunk_shift
;
4123 s
*= geo
->near_copies
;
4124 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4125 s
*= geo
->far_copies
;
4126 s
<<= geo
->chunk_shift
;
4130 /* Calculate the first device-address that could contain
4131 * any block from the chunk that includes the array-address 's'.
4132 * This too will be the start of a chunk
4134 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4136 s
>>= geo
->chunk_shift
;
4137 s
*= geo
->near_copies
;
4138 sector_div(s
, geo
->raid_disks
);
4139 s
*= geo
->far_copies
;
4140 s
<<= geo
->chunk_shift
;
4144 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4147 /* We simply copy at most one chunk (smallest of old and new)
4148 * at a time, possibly less if that exceeds RESYNC_PAGES,
4149 * or we hit a bad block or something.
4150 * This might mean we pause for normal IO in the middle of
4151 * a chunk, but that is not a problem as mddev->reshape_position
4152 * can record any location.
4154 * If we will want to write to a location that isn't
4155 * yet recorded as 'safe' (i.e. in metadata on disk) then
4156 * we need to flush all reshape requests and update the metadata.
4158 * When reshaping forwards (e.g. to more devices), we interpret
4159 * 'safe' as the earliest block which might not have been copied
4160 * down yet. We divide this by previous stripe size and multiply
4161 * by previous stripe length to get lowest device offset that we
4162 * cannot write to yet.
4163 * We interpret 'sector_nr' as an address that we want to write to.
4164 * From this we use last_device_address() to find where we might
4165 * write to, and first_device_address on the 'safe' position.
4166 * If this 'next' write position is after the 'safe' position,
4167 * we must update the metadata to increase the 'safe' position.
4169 * When reshaping backwards, we round in the opposite direction
4170 * and perform the reverse test: next write position must not be
4171 * less than current safe position.
4173 * In all this the minimum difference in data offsets
4174 * (conf->offset_diff - always positive) allows a bit of slack,
4175 * so next can be after 'safe', but not by more than offset_diff
4177 * We need to prepare all the bios here before we start any IO
4178 * to ensure the size we choose is acceptable to all devices.
4179 * The means one for each copy for write-out and an extra one for
4181 * We store the read-in bio in ->master_bio and the others in
4182 * ->devs[x].bio and ->devs[x].repl_bio.
4184 struct r10conf
*conf
= mddev
->private;
4185 struct r10bio
*r10_bio
;
4186 sector_t next
, safe
, last
;
4190 struct md_rdev
*rdev
;
4193 struct bio
*bio
, *read_bio
;
4194 int sectors_done
= 0;
4196 if (sector_nr
== 0) {
4197 /* If restarting in the middle, skip the initial sectors */
4198 if (mddev
->reshape_backwards
&&
4199 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4200 sector_nr
= (raid10_size(mddev
, 0, 0)
4201 - conf
->reshape_progress
);
4202 } else if (!mddev
->reshape_backwards
&&
4203 conf
->reshape_progress
> 0)
4204 sector_nr
= conf
->reshape_progress
;
4206 mddev
->curr_resync_completed
= sector_nr
;
4207 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4213 /* We don't use sector_nr to track where we are up to
4214 * as that doesn't work well for ->reshape_backwards.
4215 * So just use ->reshape_progress.
4217 if (mddev
->reshape_backwards
) {
4218 /* 'next' is the earliest device address that we might
4219 * write to for this chunk in the new layout
4221 next
= first_dev_address(conf
->reshape_progress
- 1,
4224 /* 'safe' is the last device address that we might read from
4225 * in the old layout after a restart
4227 safe
= last_dev_address(conf
->reshape_safe
- 1,
4230 if (next
+ conf
->offset_diff
< safe
)
4233 last
= conf
->reshape_progress
- 1;
4234 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4235 & conf
->prev
.chunk_mask
);
4236 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4237 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4239 /* 'next' is after the last device address that we
4240 * might write to for this chunk in the new layout
4242 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4244 /* 'safe' is the earliest device address that we might
4245 * read from in the old layout after a restart
4247 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4249 /* Need to update metadata if 'next' might be beyond 'safe'
4250 * as that would possibly corrupt data
4252 if (next
> safe
+ conf
->offset_diff
)
4255 sector_nr
= conf
->reshape_progress
;
4256 last
= sector_nr
| (conf
->geo
.chunk_mask
4257 & conf
->prev
.chunk_mask
);
4259 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4260 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4264 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4265 /* Need to update reshape_position in metadata */
4267 mddev
->reshape_position
= conf
->reshape_progress
;
4268 if (mddev
->reshape_backwards
)
4269 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4270 - conf
->reshape_progress
;
4272 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4273 conf
->reshape_checkpoint
= jiffies
;
4274 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4275 md_wakeup_thread(mddev
->thread
);
4276 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4277 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4278 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4279 allow_barrier(conf
);
4280 return sectors_done
;
4282 conf
->reshape_safe
= mddev
->reshape_position
;
4283 allow_barrier(conf
);
4287 /* Now schedule reads for blocks from sector_nr to last */
4288 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4290 raise_barrier(conf
, sectors_done
!= 0);
4291 atomic_set(&r10_bio
->remaining
, 0);
4292 r10_bio
->mddev
= mddev
;
4293 r10_bio
->sector
= sector_nr
;
4294 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4295 r10_bio
->sectors
= last
- sector_nr
+ 1;
4296 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4297 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4300 /* Cannot read from here, so need to record bad blocks
4301 * on all the target devices.
4304 mempool_free(r10_bio
, conf
->r10buf_pool
);
4305 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4306 return sectors_done
;
4309 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4311 read_bio
->bi_bdev
= rdev
->bdev
;
4312 read_bio
->bi_iter
.bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4313 + rdev
->data_offset
);
4314 read_bio
->bi_private
= r10_bio
;
4315 read_bio
->bi_end_io
= end_sync_read
;
4316 read_bio
->bi_rw
= READ
;
4317 read_bio
->bi_flags
&= (~0UL << BIO_RESET_BITS
);
4318 read_bio
->bi_error
= 0;
4319 read_bio
->bi_vcnt
= 0;
4320 read_bio
->bi_iter
.bi_size
= 0;
4321 r10_bio
->master_bio
= read_bio
;
4322 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4324 /* Now find the locations in the new layout */
4325 __raid10_find_phys(&conf
->geo
, r10_bio
);
4328 read_bio
->bi_next
= NULL
;
4330 for (s
= 0; s
< conf
->copies
*2; s
++) {
4332 int d
= r10_bio
->devs
[s
/2].devnum
;
4333 struct md_rdev
*rdev2
;
4335 rdev2
= conf
->mirrors
[d
].replacement
;
4336 b
= r10_bio
->devs
[s
/2].repl_bio
;
4338 rdev2
= conf
->mirrors
[d
].rdev
;
4339 b
= r10_bio
->devs
[s
/2].bio
;
4341 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4345 b
->bi_bdev
= rdev2
->bdev
;
4346 b
->bi_iter
.bi_sector
= r10_bio
->devs
[s
/2].addr
+
4347 rdev2
->new_data_offset
;
4348 b
->bi_private
= r10_bio
;
4349 b
->bi_end_io
= end_reshape_write
;
4355 /* Now add as many pages as possible to all of these bios. */
4358 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4359 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4360 int len
= (max_sectors
- s
) << 9;
4361 if (len
> PAGE_SIZE
)
4363 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4365 if (bio_add_page(bio
, page
, len
, 0))
4368 /* Didn't fit, must stop */
4370 bio2
&& bio2
!= bio
;
4371 bio2
= bio2
->bi_next
) {
4372 /* Remove last page from this bio */
4374 bio2
->bi_iter
.bi_size
-= len
;
4375 bio_clear_flag(bio2
, BIO_SEG_VALID
);
4379 sector_nr
+= len
>> 9;
4380 nr_sectors
+= len
>> 9;
4383 r10_bio
->sectors
= nr_sectors
;
4385 /* Now submit the read */
4386 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4387 atomic_inc(&r10_bio
->remaining
);
4388 read_bio
->bi_next
= NULL
;
4389 generic_make_request(read_bio
);
4390 sector_nr
+= nr_sectors
;
4391 sectors_done
+= nr_sectors
;
4392 if (sector_nr
<= last
)
4395 /* Now that we have done the whole section we can
4396 * update reshape_progress
4398 if (mddev
->reshape_backwards
)
4399 conf
->reshape_progress
-= sectors_done
;
4401 conf
->reshape_progress
+= sectors_done
;
4403 return sectors_done
;
4406 static void end_reshape_request(struct r10bio
*r10_bio
);
4407 static int handle_reshape_read_error(struct mddev
*mddev
,
4408 struct r10bio
*r10_bio
);
4409 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4411 /* Reshape read completed. Hopefully we have a block
4413 * If we got a read error then we do sync 1-page reads from
4414 * elsewhere until we find the data - or give up.
4416 struct r10conf
*conf
= mddev
->private;
4419 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4420 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4421 /* Reshape has been aborted */
4422 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4426 /* We definitely have the data in the pages, schedule the
4429 atomic_set(&r10_bio
->remaining
, 1);
4430 for (s
= 0; s
< conf
->copies
*2; s
++) {
4432 int d
= r10_bio
->devs
[s
/2].devnum
;
4433 struct md_rdev
*rdev
;
4435 rdev
= conf
->mirrors
[d
].replacement
;
4436 b
= r10_bio
->devs
[s
/2].repl_bio
;
4438 rdev
= conf
->mirrors
[d
].rdev
;
4439 b
= r10_bio
->devs
[s
/2].bio
;
4441 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4443 atomic_inc(&rdev
->nr_pending
);
4444 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4445 atomic_inc(&r10_bio
->remaining
);
4447 generic_make_request(b
);
4449 end_reshape_request(r10_bio
);
4452 static void end_reshape(struct r10conf
*conf
)
4454 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4457 spin_lock_irq(&conf
->device_lock
);
4458 conf
->prev
= conf
->geo
;
4459 md_finish_reshape(conf
->mddev
);
4461 conf
->reshape_progress
= MaxSector
;
4462 conf
->reshape_safe
= MaxSector
;
4463 spin_unlock_irq(&conf
->device_lock
);
4465 /* read-ahead size must cover two whole stripes, which is
4466 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4468 if (conf
->mddev
->queue
) {
4469 int stripe
= conf
->geo
.raid_disks
*
4470 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4471 stripe
/= conf
->geo
.near_copies
;
4472 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4473 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4478 static int handle_reshape_read_error(struct mddev
*mddev
,
4479 struct r10bio
*r10_bio
)
4481 /* Use sync reads to get the blocks from somewhere else */
4482 int sectors
= r10_bio
->sectors
;
4483 struct r10conf
*conf
= mddev
->private;
4485 struct r10bio r10_bio
;
4486 struct r10dev devs
[conf
->copies
];
4488 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4491 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4493 r10b
->sector
= r10_bio
->sector
;
4494 __raid10_find_phys(&conf
->prev
, r10b
);
4499 int first_slot
= slot
;
4501 if (s
> (PAGE_SIZE
>> 9))
4505 int d
= r10b
->devs
[slot
].devnum
;
4506 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4509 test_bit(Faulty
, &rdev
->flags
) ||
4510 !test_bit(In_sync
, &rdev
->flags
))
4513 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4514 success
= sync_page_io(rdev
,
4523 if (slot
>= conf
->copies
)
4525 if (slot
== first_slot
)
4529 /* couldn't read this block, must give up */
4530 set_bit(MD_RECOVERY_INTR
,
4540 static void end_reshape_write(struct bio
*bio
)
4542 struct r10bio
*r10_bio
= bio
->bi_private
;
4543 struct mddev
*mddev
= r10_bio
->mddev
;
4544 struct r10conf
*conf
= mddev
->private;
4548 struct md_rdev
*rdev
= NULL
;
4550 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4552 rdev
= conf
->mirrors
[d
].replacement
;
4555 rdev
= conf
->mirrors
[d
].rdev
;
4558 if (bio
->bi_error
) {
4559 /* FIXME should record badblock */
4560 md_error(mddev
, rdev
);
4563 rdev_dec_pending(rdev
, mddev
);
4564 end_reshape_request(r10_bio
);
4567 static void end_reshape_request(struct r10bio
*r10_bio
)
4569 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4571 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4572 bio_put(r10_bio
->master_bio
);
4576 static void raid10_finish_reshape(struct mddev
*mddev
)
4578 struct r10conf
*conf
= mddev
->private;
4580 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4583 if (mddev
->delta_disks
> 0) {
4584 sector_t size
= raid10_size(mddev
, 0, 0);
4585 md_set_array_sectors(mddev
, size
);
4586 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4587 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4588 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4590 mddev
->resync_max_sectors
= size
;
4591 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4592 revalidate_disk(mddev
->gendisk
);
4595 for (d
= conf
->geo
.raid_disks
;
4596 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4598 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4600 clear_bit(In_sync
, &rdev
->flags
);
4601 rdev
= conf
->mirrors
[d
].replacement
;
4603 clear_bit(In_sync
, &rdev
->flags
);
4606 mddev
->layout
= mddev
->new_layout
;
4607 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4608 mddev
->reshape_position
= MaxSector
;
4609 mddev
->delta_disks
= 0;
4610 mddev
->reshape_backwards
= 0;
4613 static struct md_personality raid10_personality
=
4617 .owner
= THIS_MODULE
,
4618 .make_request
= make_request
,
4620 .free
= raid10_free
,
4622 .error_handler
= error
,
4623 .hot_add_disk
= raid10_add_disk
,
4624 .hot_remove_disk
= raid10_remove_disk
,
4625 .spare_active
= raid10_spare_active
,
4626 .sync_request
= sync_request
,
4627 .quiesce
= raid10_quiesce
,
4628 .size
= raid10_size
,
4629 .resize
= raid10_resize
,
4630 .takeover
= raid10_takeover
,
4631 .check_reshape
= raid10_check_reshape
,
4632 .start_reshape
= raid10_start_reshape
,
4633 .finish_reshape
= raid10_finish_reshape
,
4634 .congested
= raid10_congested
,
4637 static int __init
raid_init(void)
4639 return register_md_personality(&raid10_personality
);
4642 static void raid_exit(void)
4644 unregister_md_personality(&raid10_personality
);
4647 module_init(raid_init
);
4648 module_exit(raid_exit
);
4649 MODULE_LICENSE("GPL");
4650 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4651 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4652 MODULE_ALIAS("md-raid10");
4653 MODULE_ALIAS("md-level-10");
4655 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);