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_op(bio
) == REQ_OP_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_op(bio
) == REQ_OP_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 op
= bio_op(bio
);
1062 const int rw
= bio_data_dir(bio
);
1063 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1064 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1065 unsigned long flags
;
1066 struct md_rdev
*blocked_rdev
;
1067 struct blk_plug_cb
*cb
;
1068 struct raid10_plug_cb
*plug
= NULL
;
1069 int sectors_handled
;
1074 * Register the new request and wait if the reconstruction
1075 * thread has put up a bar for new requests.
1076 * Continue immediately if no resync is active currently.
1080 sectors
= bio_sectors(bio
);
1081 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1082 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
&&
1083 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
) {
1084 /* IO spans the reshape position. Need to wait for
1087 allow_barrier(conf
);
1088 wait_event(conf
->wait_barrier
,
1089 conf
->reshape_progress
<= bio
->bi_iter
.bi_sector
||
1090 conf
->reshape_progress
>= bio
->bi_iter
.bi_sector
+
1094 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1095 bio_data_dir(bio
) == WRITE
&&
1096 (mddev
->reshape_backwards
1097 ? (bio
->bi_iter
.bi_sector
< conf
->reshape_safe
&&
1098 bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_progress
)
1099 : (bio
->bi_iter
.bi_sector
+ sectors
> conf
->reshape_safe
&&
1100 bio
->bi_iter
.bi_sector
< conf
->reshape_progress
))) {
1101 /* Need to update reshape_position in metadata */
1102 mddev
->reshape_position
= conf
->reshape_progress
;
1103 set_mask_bits(&mddev
->flags
, 0,
1104 BIT(MD_CHANGE_DEVS
) | BIT(MD_CHANGE_PENDING
));
1105 md_wakeup_thread(mddev
->thread
);
1106 wait_event(mddev
->sb_wait
,
1107 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1109 conf
->reshape_safe
= mddev
->reshape_position
;
1112 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1114 r10_bio
->master_bio
= bio
;
1115 r10_bio
->sectors
= sectors
;
1117 r10_bio
->mddev
= mddev
;
1118 r10_bio
->sector
= bio
->bi_iter
.bi_sector
;
1121 /* We might need to issue multiple reads to different
1122 * devices if there are bad blocks around, so we keep
1123 * track of the number of reads in bio->bi_phys_segments.
1124 * If this is 0, there is only one r10_bio and no locking
1125 * will be needed when the request completes. If it is
1126 * non-zero, then it is the number of not-completed requests.
1128 bio
->bi_phys_segments
= 0;
1129 bio_clear_flag(bio
, BIO_SEG_VALID
);
1133 * read balancing logic:
1135 struct md_rdev
*rdev
;
1139 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1141 raid_end_bio_io(r10_bio
);
1144 slot
= r10_bio
->read_slot
;
1146 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1147 bio_trim(read_bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1150 r10_bio
->devs
[slot
].bio
= read_bio
;
1151 r10_bio
->devs
[slot
].rdev
= rdev
;
1153 read_bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
+
1154 choose_data_offset(r10_bio
, rdev
);
1155 read_bio
->bi_bdev
= rdev
->bdev
;
1156 read_bio
->bi_end_io
= raid10_end_read_request
;
1157 bio_set_op_attrs(read_bio
, op
, do_sync
);
1158 read_bio
->bi_private
= r10_bio
;
1160 if (max_sectors
< r10_bio
->sectors
) {
1161 /* Could not read all from this device, so we will
1162 * need another r10_bio.
1164 sectors_handled
= (r10_bio
->sector
+ max_sectors
1165 - bio
->bi_iter
.bi_sector
);
1166 r10_bio
->sectors
= max_sectors
;
1167 spin_lock_irq(&conf
->device_lock
);
1168 if (bio
->bi_phys_segments
== 0)
1169 bio
->bi_phys_segments
= 2;
1171 bio
->bi_phys_segments
++;
1172 spin_unlock_irq(&conf
->device_lock
);
1173 /* Cannot call generic_make_request directly
1174 * as that will be queued in __generic_make_request
1175 * and subsequent mempool_alloc might block
1176 * waiting for it. so hand bio over to raid10d.
1178 reschedule_retry(r10_bio
);
1180 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1182 r10_bio
->master_bio
= bio
;
1183 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1185 r10_bio
->mddev
= mddev
;
1186 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+
1190 generic_make_request(read_bio
);
1197 if (conf
->pending_count
>= max_queued_requests
) {
1198 md_wakeup_thread(mddev
->thread
);
1199 wait_event(conf
->wait_barrier
,
1200 conf
->pending_count
< max_queued_requests
);
1202 /* first select target devices under rcu_lock and
1203 * inc refcount on their rdev. Record them by setting
1205 * If there are known/acknowledged bad blocks on any device
1206 * on which we have seen a write error, we want to avoid
1207 * writing to those blocks. This potentially requires several
1208 * writes to write around the bad blocks. Each set of writes
1209 * gets its own r10_bio with a set of bios attached. The number
1210 * of r10_bios is recored in bio->bi_phys_segments just as with
1214 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1215 raid10_find_phys(conf
, r10_bio
);
1217 blocked_rdev
= NULL
;
1219 max_sectors
= r10_bio
->sectors
;
1221 for (i
= 0; i
< conf
->copies
; i
++) {
1222 int d
= r10_bio
->devs
[i
].devnum
;
1223 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1224 struct md_rdev
*rrdev
= rcu_dereference(
1225 conf
->mirrors
[d
].replacement
);
1228 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1229 atomic_inc(&rdev
->nr_pending
);
1230 blocked_rdev
= rdev
;
1233 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1234 atomic_inc(&rrdev
->nr_pending
);
1235 blocked_rdev
= rrdev
;
1238 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)))
1240 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)))
1243 r10_bio
->devs
[i
].bio
= NULL
;
1244 r10_bio
->devs
[i
].repl_bio
= NULL
;
1246 if (!rdev
&& !rrdev
) {
1247 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1250 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1252 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1256 is_bad
= is_badblock(rdev
, dev_sector
,
1258 &first_bad
, &bad_sectors
);
1260 /* Mustn't write here until the bad block
1263 atomic_inc(&rdev
->nr_pending
);
1264 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1265 blocked_rdev
= rdev
;
1268 if (is_bad
&& first_bad
<= dev_sector
) {
1269 /* Cannot write here at all */
1270 bad_sectors
-= (dev_sector
- first_bad
);
1271 if (bad_sectors
< max_sectors
)
1272 /* Mustn't write more than bad_sectors
1273 * to other devices yet
1275 max_sectors
= bad_sectors
;
1276 /* We don't set R10BIO_Degraded as that
1277 * only applies if the disk is missing,
1278 * so it might be re-added, and we want to
1279 * know to recover this chunk.
1280 * In this case the device is here, and the
1281 * fact that this chunk is not in-sync is
1282 * recorded in the bad block log.
1287 int good_sectors
= first_bad
- dev_sector
;
1288 if (good_sectors
< max_sectors
)
1289 max_sectors
= good_sectors
;
1293 r10_bio
->devs
[i
].bio
= bio
;
1294 atomic_inc(&rdev
->nr_pending
);
1297 r10_bio
->devs
[i
].repl_bio
= bio
;
1298 atomic_inc(&rrdev
->nr_pending
);
1303 if (unlikely(blocked_rdev
)) {
1304 /* Have to wait for this device to get unblocked, then retry */
1308 for (j
= 0; j
< i
; j
++) {
1309 if (r10_bio
->devs
[j
].bio
) {
1310 d
= r10_bio
->devs
[j
].devnum
;
1311 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1313 if (r10_bio
->devs
[j
].repl_bio
) {
1314 struct md_rdev
*rdev
;
1315 d
= r10_bio
->devs
[j
].devnum
;
1316 rdev
= conf
->mirrors
[d
].replacement
;
1318 /* Race with remove_disk */
1320 rdev
= conf
->mirrors
[d
].rdev
;
1322 rdev_dec_pending(rdev
, mddev
);
1325 allow_barrier(conf
);
1326 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1331 if (max_sectors
< r10_bio
->sectors
) {
1332 /* We are splitting this into multiple parts, so
1333 * we need to prepare for allocating another r10_bio.
1335 r10_bio
->sectors
= max_sectors
;
1336 spin_lock_irq(&conf
->device_lock
);
1337 if (bio
->bi_phys_segments
== 0)
1338 bio
->bi_phys_segments
= 2;
1340 bio
->bi_phys_segments
++;
1341 spin_unlock_irq(&conf
->device_lock
);
1343 sectors_handled
= r10_bio
->sector
+ max_sectors
-
1344 bio
->bi_iter
.bi_sector
;
1346 atomic_set(&r10_bio
->remaining
, 1);
1347 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1349 for (i
= 0; i
< conf
->copies
; i
++) {
1351 int d
= r10_bio
->devs
[i
].devnum
;
1352 if (r10_bio
->devs
[i
].bio
) {
1353 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1354 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1355 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1357 r10_bio
->devs
[i
].bio
= mbio
;
1359 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1360 choose_data_offset(r10_bio
,
1362 mbio
->bi_bdev
= rdev
->bdev
;
1363 mbio
->bi_end_io
= raid10_end_write_request
;
1364 bio_set_op_attrs(mbio
, op
, do_sync
| do_fua
);
1365 mbio
->bi_private
= r10_bio
;
1367 atomic_inc(&r10_bio
->remaining
);
1369 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1372 plug
= container_of(cb
, struct raid10_plug_cb
,
1376 spin_lock_irqsave(&conf
->device_lock
, flags
);
1378 bio_list_add(&plug
->pending
, mbio
);
1379 plug
->pending_cnt
++;
1381 bio_list_add(&conf
->pending_bio_list
, mbio
);
1382 conf
->pending_count
++;
1384 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1386 md_wakeup_thread(mddev
->thread
);
1389 if (r10_bio
->devs
[i
].repl_bio
) {
1390 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1392 /* Replacement just got moved to main 'rdev' */
1394 rdev
= conf
->mirrors
[d
].rdev
;
1396 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1397 bio_trim(mbio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
,
1399 r10_bio
->devs
[i
].repl_bio
= mbio
;
1401 mbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
1404 mbio
->bi_bdev
= rdev
->bdev
;
1405 mbio
->bi_end_io
= raid10_end_write_request
;
1406 bio_set_op_attrs(mbio
, op
, do_sync
| do_fua
);
1407 mbio
->bi_private
= r10_bio
;
1409 atomic_inc(&r10_bio
->remaining
);
1410 spin_lock_irqsave(&conf
->device_lock
, flags
);
1411 bio_list_add(&conf
->pending_bio_list
, mbio
);
1412 conf
->pending_count
++;
1413 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1414 if (!mddev_check_plugged(mddev
))
1415 md_wakeup_thread(mddev
->thread
);
1419 /* Don't remove the bias on 'remaining' (one_write_done) until
1420 * after checking if we need to go around again.
1423 if (sectors_handled
< bio_sectors(bio
)) {
1424 one_write_done(r10_bio
);
1425 /* We need another r10_bio. It has already been counted
1426 * in bio->bi_phys_segments.
1428 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1430 r10_bio
->master_bio
= bio
;
1431 r10_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1433 r10_bio
->mddev
= mddev
;
1434 r10_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1438 one_write_done(r10_bio
);
1441 static void raid10_make_request(struct mddev
*mddev
, struct bio
*bio
)
1443 struct r10conf
*conf
= mddev
->private;
1444 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1445 int chunk_sects
= chunk_mask
+ 1;
1449 if (unlikely(bio
->bi_rw
& REQ_PREFLUSH
)) {
1450 md_flush_request(mddev
, bio
);
1454 md_write_start(mddev
, bio
);
1459 * If this request crosses a chunk boundary, we need to split
1462 if (unlikely((bio
->bi_iter
.bi_sector
& chunk_mask
) +
1463 bio_sectors(bio
) > chunk_sects
1464 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1465 || conf
->prev
.near_copies
<
1466 conf
->prev
.raid_disks
))) {
1467 split
= bio_split(bio
, chunk_sects
-
1468 (bio
->bi_iter
.bi_sector
&
1470 GFP_NOIO
, fs_bio_set
);
1471 bio_chain(split
, bio
);
1476 __make_request(mddev
, split
);
1477 } while (split
!= bio
);
1479 /* In case raid10d snuck in to freeze_array */
1480 wake_up(&conf
->wait_barrier
);
1483 static void raid10_status(struct seq_file
*seq
, struct mddev
*mddev
)
1485 struct r10conf
*conf
= mddev
->private;
1488 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1489 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1490 if (conf
->geo
.near_copies
> 1)
1491 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1492 if (conf
->geo
.far_copies
> 1) {
1493 if (conf
->geo
.far_offset
)
1494 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1496 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1497 if (conf
->geo
.far_set_size
!= conf
->geo
.raid_disks
)
1498 seq_printf(seq
, " %d devices per set", conf
->geo
.far_set_size
);
1500 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1501 conf
->geo
.raid_disks
- mddev
->degraded
);
1502 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1503 seq_printf(seq
, "%s",
1504 conf
->mirrors
[i
].rdev
&&
1505 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1506 seq_printf(seq
, "]");
1509 /* check if there are enough drives for
1510 * every block to appear on atleast one.
1511 * Don't consider the device numbered 'ignore'
1512 * as we might be about to remove it.
1514 static int _enough(struct r10conf
*conf
, int previous
, int ignore
)
1520 disks
= conf
->prev
.raid_disks
;
1521 ncopies
= conf
->prev
.near_copies
;
1523 disks
= conf
->geo
.raid_disks
;
1524 ncopies
= conf
->geo
.near_copies
;
1529 int n
= conf
->copies
;
1533 struct md_rdev
*rdev
;
1534 if (this != ignore
&&
1535 (rdev
= rcu_dereference(conf
->mirrors
[this].rdev
)) &&
1536 test_bit(In_sync
, &rdev
->flags
))
1538 this = (this+1) % disks
;
1542 first
= (first
+ ncopies
) % disks
;
1543 } while (first
!= 0);
1550 static int enough(struct r10conf
*conf
, int ignore
)
1552 /* when calling 'enough', both 'prev' and 'geo' must
1554 * This is ensured if ->reconfig_mutex or ->device_lock
1557 return _enough(conf
, 0, ignore
) &&
1558 _enough(conf
, 1, ignore
);
1561 static void raid10_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1563 char b
[BDEVNAME_SIZE
];
1564 struct r10conf
*conf
= mddev
->private;
1565 unsigned long flags
;
1568 * If it is not operational, then we have already marked it as dead
1569 * else if it is the last working disks, ignore the error, let the
1570 * next level up know.
1571 * else mark the drive as failed
1573 spin_lock_irqsave(&conf
->device_lock
, flags
);
1574 if (test_bit(In_sync
, &rdev
->flags
)
1575 && !enough(conf
, rdev
->raid_disk
)) {
1577 * Don't fail the drive, just return an IO error.
1579 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1582 if (test_and_clear_bit(In_sync
, &rdev
->flags
))
1585 * If recovery is running, make sure it aborts.
1587 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1588 set_bit(Blocked
, &rdev
->flags
);
1589 set_bit(Faulty
, &rdev
->flags
);
1590 set_mask_bits(&mddev
->flags
, 0,
1591 BIT(MD_CHANGE_DEVS
) | BIT(MD_CHANGE_PENDING
));
1592 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1594 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1595 "md/raid10:%s: Operation continuing on %d devices.\n",
1596 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1597 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1600 static void print_conf(struct r10conf
*conf
)
1603 struct raid10_info
*tmp
;
1605 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1607 printk(KERN_DEBUG
"(!conf)\n");
1610 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1611 conf
->geo
.raid_disks
);
1613 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1614 char b
[BDEVNAME_SIZE
];
1615 tmp
= conf
->mirrors
+ i
;
1617 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1618 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1619 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1620 bdevname(tmp
->rdev
->bdev
,b
));
1624 static void close_sync(struct r10conf
*conf
)
1627 allow_barrier(conf
);
1629 mempool_destroy(conf
->r10buf_pool
);
1630 conf
->r10buf_pool
= NULL
;
1633 static int raid10_spare_active(struct mddev
*mddev
)
1636 struct r10conf
*conf
= mddev
->private;
1637 struct raid10_info
*tmp
;
1639 unsigned long flags
;
1642 * Find all non-in_sync disks within the RAID10 configuration
1643 * and mark them in_sync
1645 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1646 tmp
= conf
->mirrors
+ i
;
1647 if (tmp
->replacement
1648 && tmp
->replacement
->recovery_offset
== MaxSector
1649 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1650 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1651 /* Replacement has just become active */
1653 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1656 /* Replaced device not technically faulty,
1657 * but we need to be sure it gets removed
1658 * and never re-added.
1660 set_bit(Faulty
, &tmp
->rdev
->flags
);
1661 sysfs_notify_dirent_safe(
1662 tmp
->rdev
->sysfs_state
);
1664 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1665 } else if (tmp
->rdev
1666 && tmp
->rdev
->recovery_offset
== MaxSector
1667 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1668 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1670 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1673 spin_lock_irqsave(&conf
->device_lock
, flags
);
1674 mddev
->degraded
-= count
;
1675 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1681 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1683 struct r10conf
*conf
= mddev
->private;
1687 int last
= conf
->geo
.raid_disks
- 1;
1689 if (mddev
->recovery_cp
< MaxSector
)
1690 /* only hot-add to in-sync arrays, as recovery is
1691 * very different from resync
1694 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, 1, -1))
1697 if (md_integrity_add_rdev(rdev
, mddev
))
1700 if (rdev
->raid_disk
>= 0)
1701 first
= last
= rdev
->raid_disk
;
1703 if (rdev
->saved_raid_disk
>= first
&&
1704 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1705 mirror
= rdev
->saved_raid_disk
;
1708 for ( ; mirror
<= last
; mirror
++) {
1709 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1710 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1713 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1714 p
->replacement
!= NULL
)
1716 clear_bit(In_sync
, &rdev
->flags
);
1717 set_bit(Replacement
, &rdev
->flags
);
1718 rdev
->raid_disk
= mirror
;
1721 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1722 rdev
->data_offset
<< 9);
1724 rcu_assign_pointer(p
->replacement
, rdev
);
1729 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1730 rdev
->data_offset
<< 9);
1732 p
->head_position
= 0;
1733 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1734 rdev
->raid_disk
= mirror
;
1736 if (rdev
->saved_raid_disk
!= mirror
)
1738 rcu_assign_pointer(p
->rdev
, rdev
);
1741 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1742 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1748 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1750 struct r10conf
*conf
= mddev
->private;
1752 int number
= rdev
->raid_disk
;
1753 struct md_rdev
**rdevp
;
1754 struct raid10_info
*p
= conf
->mirrors
+ number
;
1757 if (rdev
== p
->rdev
)
1759 else if (rdev
== p
->replacement
)
1760 rdevp
= &p
->replacement
;
1764 if (test_bit(In_sync
, &rdev
->flags
) ||
1765 atomic_read(&rdev
->nr_pending
)) {
1769 /* Only remove faulty devices if recovery
1772 if (!test_bit(Faulty
, &rdev
->flags
) &&
1773 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1774 (!p
->replacement
|| p
->replacement
== rdev
) &&
1775 number
< conf
->geo
.raid_disks
&&
1782 if (atomic_read(&rdev
->nr_pending
)) {
1783 /* lost the race, try later */
1787 } else if (p
->replacement
) {
1788 /* We must have just cleared 'rdev' */
1789 p
->rdev
= p
->replacement
;
1790 clear_bit(Replacement
, &p
->replacement
->flags
);
1791 smp_mb(); /* Make sure other CPUs may see both as identical
1792 * but will never see neither -- if they are careful.
1794 p
->replacement
= NULL
;
1795 clear_bit(WantReplacement
, &rdev
->flags
);
1797 /* We might have just remove the Replacement as faulty
1798 * Clear the flag just in case
1800 clear_bit(WantReplacement
, &rdev
->flags
);
1802 err
= md_integrity_register(mddev
);
1810 static void end_sync_read(struct bio
*bio
)
1812 struct r10bio
*r10_bio
= bio
->bi_private
;
1813 struct r10conf
*conf
= r10_bio
->mddev
->private;
1816 if (bio
== r10_bio
->master_bio
) {
1817 /* this is a reshape read */
1818 d
= r10_bio
->read_slot
; /* really the read dev */
1820 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1823 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1825 /* The write handler will notice the lack of
1826 * R10BIO_Uptodate and record any errors etc
1828 atomic_add(r10_bio
->sectors
,
1829 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1831 /* for reconstruct, we always reschedule after a read.
1832 * for resync, only after all reads
1834 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1835 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1836 atomic_dec_and_test(&r10_bio
->remaining
)) {
1837 /* we have read all the blocks,
1838 * do the comparison in process context in raid10d
1840 reschedule_retry(r10_bio
);
1844 static void end_sync_request(struct r10bio
*r10_bio
)
1846 struct mddev
*mddev
= r10_bio
->mddev
;
1848 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1849 if (r10_bio
->master_bio
== NULL
) {
1850 /* the primary of several recovery bios */
1851 sector_t s
= r10_bio
->sectors
;
1852 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1853 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1854 reschedule_retry(r10_bio
);
1857 md_done_sync(mddev
, s
, 1);
1860 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1861 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1862 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1863 reschedule_retry(r10_bio
);
1871 static void end_sync_write(struct bio
*bio
)
1873 struct r10bio
*r10_bio
= bio
->bi_private
;
1874 struct mddev
*mddev
= r10_bio
->mddev
;
1875 struct r10conf
*conf
= mddev
->private;
1881 struct md_rdev
*rdev
= NULL
;
1883 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1885 rdev
= conf
->mirrors
[d
].replacement
;
1887 rdev
= conf
->mirrors
[d
].rdev
;
1889 if (bio
->bi_error
) {
1891 md_error(mddev
, rdev
);
1893 set_bit(WriteErrorSeen
, &rdev
->flags
);
1894 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1895 set_bit(MD_RECOVERY_NEEDED
,
1896 &rdev
->mddev
->recovery
);
1897 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1899 } else if (is_badblock(rdev
,
1900 r10_bio
->devs
[slot
].addr
,
1902 &first_bad
, &bad_sectors
))
1903 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1905 rdev_dec_pending(rdev
, mddev
);
1907 end_sync_request(r10_bio
);
1911 * Note: sync and recover and handled very differently for raid10
1912 * This code is for resync.
1913 * For resync, we read through virtual addresses and read all blocks.
1914 * If there is any error, we schedule a write. The lowest numbered
1915 * drive is authoritative.
1916 * However requests come for physical address, so we need to map.
1917 * For every physical address there are raid_disks/copies virtual addresses,
1918 * which is always are least one, but is not necessarly an integer.
1919 * This means that a physical address can span multiple chunks, so we may
1920 * have to submit multiple io requests for a single sync request.
1923 * We check if all blocks are in-sync and only write to blocks that
1926 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1928 struct r10conf
*conf
= mddev
->private;
1930 struct bio
*tbio
, *fbio
;
1933 atomic_set(&r10_bio
->remaining
, 1);
1935 /* find the first device with a block */
1936 for (i
=0; i
<conf
->copies
; i
++)
1937 if (!r10_bio
->devs
[i
].bio
->bi_error
)
1940 if (i
== conf
->copies
)
1944 fbio
= r10_bio
->devs
[i
].bio
;
1945 fbio
->bi_iter
.bi_size
= r10_bio
->sectors
<< 9;
1946 fbio
->bi_iter
.bi_idx
= 0;
1948 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
1949 /* now find blocks with errors */
1950 for (i
=0 ; i
< conf
->copies
; i
++) {
1953 tbio
= r10_bio
->devs
[i
].bio
;
1955 if (tbio
->bi_end_io
!= end_sync_read
)
1959 if (!r10_bio
->devs
[i
].bio
->bi_error
) {
1960 /* We know that the bi_io_vec layout is the same for
1961 * both 'first' and 'i', so we just compare them.
1962 * All vec entries are PAGE_SIZE;
1964 int sectors
= r10_bio
->sectors
;
1965 for (j
= 0; j
< vcnt
; j
++) {
1966 int len
= PAGE_SIZE
;
1967 if (sectors
< (len
/ 512))
1968 len
= sectors
* 512;
1969 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1970 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1977 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
1978 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1979 /* Don't fix anything. */
1982 /* Ok, we need to write this bio, either to correct an
1983 * inconsistency or to correct an unreadable block.
1984 * First we need to fixup bv_offset, bv_len and
1985 * bi_vecs, as the read request might have corrupted these
1989 tbio
->bi_vcnt
= vcnt
;
1990 tbio
->bi_iter
.bi_size
= fbio
->bi_iter
.bi_size
;
1991 tbio
->bi_private
= r10_bio
;
1992 tbio
->bi_iter
.bi_sector
= r10_bio
->devs
[i
].addr
;
1993 tbio
->bi_end_io
= end_sync_write
;
1994 bio_set_op_attrs(tbio
, REQ_OP_WRITE
, 0);
1996 bio_copy_data(tbio
, fbio
);
1998 d
= r10_bio
->devs
[i
].devnum
;
1999 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2000 atomic_inc(&r10_bio
->remaining
);
2001 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(tbio
));
2003 tbio
->bi_iter
.bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2004 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2005 generic_make_request(tbio
);
2008 /* Now write out to any replacement devices
2011 for (i
= 0; i
< conf
->copies
; i
++) {
2014 tbio
= r10_bio
->devs
[i
].repl_bio
;
2015 if (!tbio
|| !tbio
->bi_end_io
)
2017 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2018 && r10_bio
->devs
[i
].bio
!= fbio
)
2019 bio_copy_data(tbio
, fbio
);
2020 d
= r10_bio
->devs
[i
].devnum
;
2021 atomic_inc(&r10_bio
->remaining
);
2022 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2024 generic_make_request(tbio
);
2028 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2029 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2035 * Now for the recovery code.
2036 * Recovery happens across physical sectors.
2037 * We recover all non-is_sync drives by finding the virtual address of
2038 * each, and then choose a working drive that also has that virt address.
2039 * There is a separate r10_bio for each non-in_sync drive.
2040 * Only the first two slots are in use. The first for reading,
2041 * The second for writing.
2044 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2046 /* We got a read error during recovery.
2047 * We repeat the read in smaller page-sized sections.
2048 * If a read succeeds, write it to the new device or record
2049 * a bad block if we cannot.
2050 * If a read fails, record a bad block on both old and
2053 struct mddev
*mddev
= r10_bio
->mddev
;
2054 struct r10conf
*conf
= mddev
->private;
2055 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2057 int sectors
= r10_bio
->sectors
;
2059 int dr
= r10_bio
->devs
[0].devnum
;
2060 int dw
= r10_bio
->devs
[1].devnum
;
2064 struct md_rdev
*rdev
;
2068 if (s
> (PAGE_SIZE
>>9))
2071 rdev
= conf
->mirrors
[dr
].rdev
;
2072 addr
= r10_bio
->devs
[0].addr
+ sect
,
2073 ok
= sync_page_io(rdev
,
2076 bio
->bi_io_vec
[idx
].bv_page
,
2077 REQ_OP_READ
, 0, false);
2079 rdev
= conf
->mirrors
[dw
].rdev
;
2080 addr
= r10_bio
->devs
[1].addr
+ sect
;
2081 ok
= sync_page_io(rdev
,
2084 bio
->bi_io_vec
[idx
].bv_page
,
2085 REQ_OP_WRITE
, 0, false);
2087 set_bit(WriteErrorSeen
, &rdev
->flags
);
2088 if (!test_and_set_bit(WantReplacement
,
2090 set_bit(MD_RECOVERY_NEEDED
,
2091 &rdev
->mddev
->recovery
);
2095 /* We don't worry if we cannot set a bad block -
2096 * it really is bad so there is no loss in not
2099 rdev_set_badblocks(rdev
, addr
, s
, 0);
2101 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2102 /* need bad block on destination too */
2103 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2104 addr
= r10_bio
->devs
[1].addr
+ sect
;
2105 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2107 /* just abort the recovery */
2109 "md/raid10:%s: recovery aborted"
2110 " due to read error\n",
2113 conf
->mirrors
[dw
].recovery_disabled
2114 = mddev
->recovery_disabled
;
2115 set_bit(MD_RECOVERY_INTR
,
2128 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2130 struct r10conf
*conf
= mddev
->private;
2132 struct bio
*wbio
, *wbio2
;
2134 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2135 fix_recovery_read_error(r10_bio
);
2136 end_sync_request(r10_bio
);
2141 * share the pages with the first bio
2142 * and submit the write request
2144 d
= r10_bio
->devs
[1].devnum
;
2145 wbio
= r10_bio
->devs
[1].bio
;
2146 wbio2
= r10_bio
->devs
[1].repl_bio
;
2147 /* Need to test wbio2->bi_end_io before we call
2148 * generic_make_request as if the former is NULL,
2149 * the latter is free to free wbio2.
2151 if (wbio2
&& !wbio2
->bi_end_io
)
2153 if (wbio
->bi_end_io
) {
2154 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2155 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, bio_sectors(wbio
));
2156 generic_make_request(wbio
);
2159 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2160 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2161 bio_sectors(wbio2
));
2162 generic_make_request(wbio2
);
2167 * Used by fix_read_error() to decay the per rdev read_errors.
2168 * We halve the read error count for every hour that has elapsed
2169 * since the last recorded read error.
2172 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2174 struct timespec cur_time_mon
;
2175 unsigned long hours_since_last
;
2176 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2178 ktime_get_ts(&cur_time_mon
);
2180 if (rdev
->last_read_error
.tv_sec
== 0 &&
2181 rdev
->last_read_error
.tv_nsec
== 0) {
2182 /* first time we've seen a read error */
2183 rdev
->last_read_error
= cur_time_mon
;
2187 hours_since_last
= (cur_time_mon
.tv_sec
-
2188 rdev
->last_read_error
.tv_sec
) / 3600;
2190 rdev
->last_read_error
= cur_time_mon
;
2193 * if hours_since_last is > the number of bits in read_errors
2194 * just set read errors to 0. We do this to avoid
2195 * overflowing the shift of read_errors by hours_since_last.
2197 if (hours_since_last
>= 8 * sizeof(read_errors
))
2198 atomic_set(&rdev
->read_errors
, 0);
2200 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2203 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2204 int sectors
, struct page
*page
, int rw
)
2209 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2210 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2212 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, 0, false))
2216 set_bit(WriteErrorSeen
, &rdev
->flags
);
2217 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2218 set_bit(MD_RECOVERY_NEEDED
,
2219 &rdev
->mddev
->recovery
);
2221 /* need to record an error - either for the block or the device */
2222 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2223 md_error(rdev
->mddev
, rdev
);
2228 * This is a kernel thread which:
2230 * 1. Retries failed read operations on working mirrors.
2231 * 2. Updates the raid superblock when problems encounter.
2232 * 3. Performs writes following reads for array synchronising.
2235 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2237 int sect
= 0; /* Offset from r10_bio->sector */
2238 int sectors
= r10_bio
->sectors
;
2239 struct md_rdev
*rdev
;
2240 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2241 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2243 /* still own a reference to this rdev, so it cannot
2244 * have been cleared recently.
2246 rdev
= conf
->mirrors
[d
].rdev
;
2248 if (test_bit(Faulty
, &rdev
->flags
))
2249 /* drive has already been failed, just ignore any
2250 more fix_read_error() attempts */
2253 check_decay_read_errors(mddev
, rdev
);
2254 atomic_inc(&rdev
->read_errors
);
2255 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2256 char b
[BDEVNAME_SIZE
];
2257 bdevname(rdev
->bdev
, b
);
2260 "md/raid10:%s: %s: Raid device exceeded "
2261 "read_error threshold [cur %d:max %d]\n",
2263 atomic_read(&rdev
->read_errors
), max_read_errors
);
2265 "md/raid10:%s: %s: Failing raid device\n",
2267 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2268 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2274 int sl
= r10_bio
->read_slot
;
2278 if (s
> (PAGE_SIZE
>>9))
2286 d
= r10_bio
->devs
[sl
].devnum
;
2287 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2289 test_bit(In_sync
, &rdev
->flags
) &&
2290 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2291 &first_bad
, &bad_sectors
) == 0) {
2292 atomic_inc(&rdev
->nr_pending
);
2294 success
= sync_page_io(rdev
,
2295 r10_bio
->devs
[sl
].addr
+
2299 REQ_OP_READ
, 0, false);
2300 rdev_dec_pending(rdev
, mddev
);
2306 if (sl
== conf
->copies
)
2308 } while (!success
&& sl
!= r10_bio
->read_slot
);
2312 /* Cannot read from anywhere, just mark the block
2313 * as bad on the first device to discourage future
2316 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2317 rdev
= conf
->mirrors
[dn
].rdev
;
2319 if (!rdev_set_badblocks(
2321 r10_bio
->devs
[r10_bio
->read_slot
].addr
2324 md_error(mddev
, rdev
);
2325 r10_bio
->devs
[r10_bio
->read_slot
].bio
2332 /* write it back and re-read */
2334 while (sl
!= r10_bio
->read_slot
) {
2335 char b
[BDEVNAME_SIZE
];
2340 d
= r10_bio
->devs
[sl
].devnum
;
2341 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2343 !test_bit(In_sync
, &rdev
->flags
))
2346 atomic_inc(&rdev
->nr_pending
);
2348 if (r10_sync_page_io(rdev
,
2349 r10_bio
->devs
[sl
].addr
+
2351 s
, conf
->tmppage
, WRITE
)
2353 /* Well, this device is dead */
2355 "md/raid10:%s: read correction "
2357 " (%d sectors at %llu on %s)\n",
2359 (unsigned long long)(
2361 choose_data_offset(r10_bio
,
2363 bdevname(rdev
->bdev
, b
));
2364 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2367 bdevname(rdev
->bdev
, b
));
2369 rdev_dec_pending(rdev
, mddev
);
2373 while (sl
!= r10_bio
->read_slot
) {
2374 char b
[BDEVNAME_SIZE
];
2379 d
= r10_bio
->devs
[sl
].devnum
;
2380 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2382 !test_bit(In_sync
, &rdev
->flags
))
2385 atomic_inc(&rdev
->nr_pending
);
2387 switch (r10_sync_page_io(rdev
,
2388 r10_bio
->devs
[sl
].addr
+
2393 /* Well, this device is dead */
2395 "md/raid10:%s: unable to read back "
2397 " (%d sectors at %llu on %s)\n",
2399 (unsigned long long)(
2401 choose_data_offset(r10_bio
, rdev
)),
2402 bdevname(rdev
->bdev
, b
));
2403 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2406 bdevname(rdev
->bdev
, b
));
2410 "md/raid10:%s: read error corrected"
2411 " (%d sectors at %llu on %s)\n",
2413 (unsigned long long)(
2415 choose_data_offset(r10_bio
, rdev
)),
2416 bdevname(rdev
->bdev
, b
));
2417 atomic_add(s
, &rdev
->corrected_errors
);
2420 rdev_dec_pending(rdev
, mddev
);
2430 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2432 struct bio
*bio
= r10_bio
->master_bio
;
2433 struct mddev
*mddev
= r10_bio
->mddev
;
2434 struct r10conf
*conf
= mddev
->private;
2435 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2436 /* bio has the data to be written to slot 'i' where
2437 * we just recently had a write error.
2438 * We repeatedly clone the bio and trim down to one block,
2439 * then try the write. Where the write fails we record
2441 * It is conceivable that the bio doesn't exactly align with
2442 * blocks. We must handle this.
2444 * We currently own a reference to the rdev.
2450 int sect_to_write
= r10_bio
->sectors
;
2453 if (rdev
->badblocks
.shift
< 0)
2456 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2457 bdev_logical_block_size(rdev
->bdev
) >> 9);
2458 sector
= r10_bio
->sector
;
2459 sectors
= ((r10_bio
->sector
+ block_sectors
)
2460 & ~(sector_t
)(block_sectors
- 1))
2463 while (sect_to_write
) {
2465 if (sectors
> sect_to_write
)
2466 sectors
= sect_to_write
;
2467 /* Write at 'sector' for 'sectors' */
2468 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2469 bio_trim(wbio
, sector
- bio
->bi_iter
.bi_sector
, sectors
);
2470 wbio
->bi_iter
.bi_sector
= (r10_bio
->devs
[i
].addr
+
2471 choose_data_offset(r10_bio
, rdev
) +
2472 (sector
- r10_bio
->sector
));
2473 wbio
->bi_bdev
= rdev
->bdev
;
2474 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2476 if (submit_bio_wait(wbio
) < 0)
2478 ok
= rdev_set_badblocks(rdev
, sector
,
2483 sect_to_write
-= sectors
;
2485 sectors
= block_sectors
;
2490 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2492 int slot
= r10_bio
->read_slot
;
2494 struct r10conf
*conf
= mddev
->private;
2495 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2496 char b
[BDEVNAME_SIZE
];
2497 unsigned long do_sync
;
2500 /* we got a read error. Maybe the drive is bad. Maybe just
2501 * the block and we can fix it.
2502 * We freeze all other IO, and try reading the block from
2503 * other devices. When we find one, we re-write
2504 * and check it that fixes the read error.
2505 * This is all done synchronously while the array is
2508 bio
= r10_bio
->devs
[slot
].bio
;
2509 bdevname(bio
->bi_bdev
, b
);
2511 r10_bio
->devs
[slot
].bio
= NULL
;
2513 if (mddev
->ro
== 0) {
2514 freeze_array(conf
, 1);
2515 fix_read_error(conf
, mddev
, r10_bio
);
2516 unfreeze_array(conf
);
2518 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2520 rdev_dec_pending(rdev
, mddev
);
2523 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2525 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2526 " read error for block %llu\n",
2528 (unsigned long long)r10_bio
->sector
);
2529 raid_end_bio_io(r10_bio
);
2533 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2534 slot
= r10_bio
->read_slot
;
2537 "md/raid10:%s: %s: redirecting "
2538 "sector %llu to another mirror\n",
2540 bdevname(rdev
->bdev
, b
),
2541 (unsigned long long)r10_bio
->sector
);
2542 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2544 bio_trim(bio
, r10_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
2545 r10_bio
->devs
[slot
].bio
= bio
;
2546 r10_bio
->devs
[slot
].rdev
= rdev
;
2547 bio
->bi_iter
.bi_sector
= r10_bio
->devs
[slot
].addr
2548 + choose_data_offset(r10_bio
, rdev
);
2549 bio
->bi_bdev
= rdev
->bdev
;
2550 bio_set_op_attrs(bio
, REQ_OP_READ
, do_sync
);
2551 bio
->bi_private
= r10_bio
;
2552 bio
->bi_end_io
= raid10_end_read_request
;
2553 if (max_sectors
< r10_bio
->sectors
) {
2554 /* Drat - have to split this up more */
2555 struct bio
*mbio
= r10_bio
->master_bio
;
2556 int sectors_handled
=
2557 r10_bio
->sector
+ max_sectors
2558 - mbio
->bi_iter
.bi_sector
;
2559 r10_bio
->sectors
= max_sectors
;
2560 spin_lock_irq(&conf
->device_lock
);
2561 if (mbio
->bi_phys_segments
== 0)
2562 mbio
->bi_phys_segments
= 2;
2564 mbio
->bi_phys_segments
++;
2565 spin_unlock_irq(&conf
->device_lock
);
2566 generic_make_request(bio
);
2568 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2570 r10_bio
->master_bio
= mbio
;
2571 r10_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2573 set_bit(R10BIO_ReadError
,
2575 r10_bio
->mddev
= mddev
;
2576 r10_bio
->sector
= mbio
->bi_iter
.bi_sector
2581 generic_make_request(bio
);
2584 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2586 /* Some sort of write request has finished and it
2587 * succeeded in writing where we thought there was a
2588 * bad block. So forget the bad block.
2589 * Or possibly if failed and we need to record
2593 struct md_rdev
*rdev
;
2595 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2596 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2597 for (m
= 0; m
< conf
->copies
; m
++) {
2598 int dev
= r10_bio
->devs
[m
].devnum
;
2599 rdev
= conf
->mirrors
[dev
].rdev
;
2600 if (r10_bio
->devs
[m
].bio
== NULL
)
2602 if (!r10_bio
->devs
[m
].bio
->bi_error
) {
2603 rdev_clear_badblocks(
2605 r10_bio
->devs
[m
].addr
,
2606 r10_bio
->sectors
, 0);
2608 if (!rdev_set_badblocks(
2610 r10_bio
->devs
[m
].addr
,
2611 r10_bio
->sectors
, 0))
2612 md_error(conf
->mddev
, rdev
);
2614 rdev
= conf
->mirrors
[dev
].replacement
;
2615 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2618 if (!r10_bio
->devs
[m
].repl_bio
->bi_error
) {
2619 rdev_clear_badblocks(
2621 r10_bio
->devs
[m
].addr
,
2622 r10_bio
->sectors
, 0);
2624 if (!rdev_set_badblocks(
2626 r10_bio
->devs
[m
].addr
,
2627 r10_bio
->sectors
, 0))
2628 md_error(conf
->mddev
, rdev
);
2634 for (m
= 0; m
< conf
->copies
; m
++) {
2635 int dev
= r10_bio
->devs
[m
].devnum
;
2636 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2637 rdev
= conf
->mirrors
[dev
].rdev
;
2638 if (bio
== IO_MADE_GOOD
) {
2639 rdev_clear_badblocks(
2641 r10_bio
->devs
[m
].addr
,
2642 r10_bio
->sectors
, 0);
2643 rdev_dec_pending(rdev
, conf
->mddev
);
2644 } else if (bio
!= NULL
&& bio
->bi_error
) {
2646 if (!narrow_write_error(r10_bio
, m
)) {
2647 md_error(conf
->mddev
, rdev
);
2648 set_bit(R10BIO_Degraded
,
2651 rdev_dec_pending(rdev
, conf
->mddev
);
2653 bio
= r10_bio
->devs
[m
].repl_bio
;
2654 rdev
= conf
->mirrors
[dev
].replacement
;
2655 if (rdev
&& bio
== IO_MADE_GOOD
) {
2656 rdev_clear_badblocks(
2658 r10_bio
->devs
[m
].addr
,
2659 r10_bio
->sectors
, 0);
2660 rdev_dec_pending(rdev
, conf
->mddev
);
2664 spin_lock_irq(&conf
->device_lock
);
2665 list_add(&r10_bio
->retry_list
, &conf
->bio_end_io_list
);
2667 spin_unlock_irq(&conf
->device_lock
);
2668 md_wakeup_thread(conf
->mddev
->thread
);
2670 if (test_bit(R10BIO_WriteError
,
2672 close_write(r10_bio
);
2673 raid_end_bio_io(r10_bio
);
2678 static void raid10d(struct md_thread
*thread
)
2680 struct mddev
*mddev
= thread
->mddev
;
2681 struct r10bio
*r10_bio
;
2682 unsigned long flags
;
2683 struct r10conf
*conf
= mddev
->private;
2684 struct list_head
*head
= &conf
->retry_list
;
2685 struct blk_plug plug
;
2687 md_check_recovery(mddev
);
2689 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2690 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2692 spin_lock_irqsave(&conf
->device_lock
, flags
);
2693 if (!test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2694 while (!list_empty(&conf
->bio_end_io_list
)) {
2695 list_move(conf
->bio_end_io_list
.prev
, &tmp
);
2699 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2700 while (!list_empty(&tmp
)) {
2701 r10_bio
= list_first_entry(&tmp
, struct r10bio
,
2703 list_del(&r10_bio
->retry_list
);
2704 if (mddev
->degraded
)
2705 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
2707 if (test_bit(R10BIO_WriteError
,
2709 close_write(r10_bio
);
2710 raid_end_bio_io(r10_bio
);
2714 blk_start_plug(&plug
);
2717 flush_pending_writes(conf
);
2719 spin_lock_irqsave(&conf
->device_lock
, flags
);
2720 if (list_empty(head
)) {
2721 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2724 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2725 list_del(head
->prev
);
2727 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2729 mddev
= r10_bio
->mddev
;
2730 conf
= mddev
->private;
2731 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2732 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2733 handle_write_completed(conf
, r10_bio
);
2734 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2735 reshape_request_write(mddev
, r10_bio
);
2736 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2737 sync_request_write(mddev
, r10_bio
);
2738 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2739 recovery_request_write(mddev
, r10_bio
);
2740 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2741 handle_read_error(mddev
, r10_bio
);
2743 /* just a partial read to be scheduled from a
2746 int slot
= r10_bio
->read_slot
;
2747 generic_make_request(r10_bio
->devs
[slot
].bio
);
2751 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2752 md_check_recovery(mddev
);
2754 blk_finish_plug(&plug
);
2757 static int init_resync(struct r10conf
*conf
)
2762 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2763 BUG_ON(conf
->r10buf_pool
);
2764 conf
->have_replacement
= 0;
2765 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2766 if (conf
->mirrors
[i
].replacement
)
2767 conf
->have_replacement
= 1;
2768 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2769 if (!conf
->r10buf_pool
)
2771 conf
->next_resync
= 0;
2776 * perform a "sync" on one "block"
2778 * We need to make sure that no normal I/O request - particularly write
2779 * requests - conflict with active sync requests.
2781 * This is achieved by tracking pending requests and a 'barrier' concept
2782 * that can be installed to exclude normal IO requests.
2784 * Resync and recovery are handled very differently.
2785 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2787 * For resync, we iterate over virtual addresses, read all copies,
2788 * and update if there are differences. If only one copy is live,
2790 * For recovery, we iterate over physical addresses, read a good
2791 * value for each non-in_sync drive, and over-write.
2793 * So, for recovery we may have several outstanding complex requests for a
2794 * given address, one for each out-of-sync device. We model this by allocating
2795 * a number of r10_bio structures, one for each out-of-sync device.
2796 * As we setup these structures, we collect all bio's together into a list
2797 * which we then process collectively to add pages, and then process again
2798 * to pass to generic_make_request.
2800 * The r10_bio structures are linked using a borrowed master_bio pointer.
2801 * This link is counted in ->remaining. When the r10_bio that points to NULL
2802 * has its remaining count decremented to 0, the whole complex operation
2807 static sector_t
raid10_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2810 struct r10conf
*conf
= mddev
->private;
2811 struct r10bio
*r10_bio
;
2812 struct bio
*biolist
= NULL
, *bio
;
2813 sector_t max_sector
, nr_sectors
;
2816 sector_t sync_blocks
;
2817 sector_t sectors_skipped
= 0;
2818 int chunks_skipped
= 0;
2819 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2821 if (!conf
->r10buf_pool
)
2822 if (init_resync(conf
))
2826 * Allow skipping a full rebuild for incremental assembly
2827 * of a clean array, like RAID1 does.
2829 if (mddev
->bitmap
== NULL
&&
2830 mddev
->recovery_cp
== MaxSector
&&
2831 mddev
->reshape_position
== MaxSector
&&
2832 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2833 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2834 !test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
2835 conf
->fullsync
== 0) {
2837 return mddev
->dev_sectors
- sector_nr
;
2841 max_sector
= mddev
->dev_sectors
;
2842 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2843 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2844 max_sector
= mddev
->resync_max_sectors
;
2845 if (sector_nr
>= max_sector
) {
2846 /* If we aborted, we need to abort the
2847 * sync on the 'current' bitmap chucks (there can
2848 * be several when recovering multiple devices).
2849 * as we may have started syncing it but not finished.
2850 * We can find the current address in
2851 * mddev->curr_resync, but for recovery,
2852 * we need to convert that to several
2853 * virtual addresses.
2855 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2861 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2862 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2863 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2865 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2867 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2868 bitmap_end_sync(mddev
->bitmap
, sect
,
2872 /* completed sync */
2873 if ((!mddev
->bitmap
|| conf
->fullsync
)
2874 && conf
->have_replacement
2875 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2876 /* Completed a full sync so the replacements
2877 * are now fully recovered.
2879 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2880 if (conf
->mirrors
[i
].replacement
)
2881 conf
->mirrors
[i
].replacement
2887 bitmap_close_sync(mddev
->bitmap
);
2890 return sectors_skipped
;
2893 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2894 return reshape_request(mddev
, sector_nr
, skipped
);
2896 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2897 /* if there has been nothing to do on any drive,
2898 * then there is nothing to do at all..
2901 return (max_sector
- sector_nr
) + sectors_skipped
;
2904 if (max_sector
> mddev
->resync_max
)
2905 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2907 /* make sure whole request will fit in a chunk - if chunks
2910 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2911 max_sector
> (sector_nr
| chunk_mask
))
2912 max_sector
= (sector_nr
| chunk_mask
) + 1;
2914 /* Again, very different code for resync and recovery.
2915 * Both must result in an r10bio with a list of bios that
2916 * have bi_end_io, bi_sector, bi_bdev set,
2917 * and bi_private set to the r10bio.
2918 * For recovery, we may actually create several r10bios
2919 * with 2 bios in each, that correspond to the bios in the main one.
2920 * In this case, the subordinate r10bios link back through a
2921 * borrowed master_bio pointer, and the counter in the master
2922 * includes a ref from each subordinate.
2924 /* First, we decide what to do and set ->bi_end_io
2925 * To end_sync_read if we want to read, and
2926 * end_sync_write if we will want to write.
2929 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2930 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2931 /* recovery... the complicated one */
2935 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2941 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2943 if ((mirror
->rdev
== NULL
||
2944 test_bit(In_sync
, &mirror
->rdev
->flags
))
2946 (mirror
->replacement
== NULL
||
2948 &mirror
->replacement
->flags
)))
2952 /* want to reconstruct this device */
2954 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2955 if (sect
>= mddev
->resync_max_sectors
) {
2956 /* last stripe is not complete - don't
2957 * try to recover this sector.
2961 /* Unless we are doing a full sync, or a replacement
2962 * we only need to recover the block if it is set in
2965 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2967 if (sync_blocks
< max_sync
)
2968 max_sync
= sync_blocks
;
2970 mirror
->replacement
== NULL
&&
2972 /* yep, skip the sync_blocks here, but don't assume
2973 * that there will never be anything to do here
2975 chunks_skipped
= -1;
2979 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2981 raise_barrier(conf
, rb2
!= NULL
);
2982 atomic_set(&r10_bio
->remaining
, 0);
2984 r10_bio
->master_bio
= (struct bio
*)rb2
;
2986 atomic_inc(&rb2
->remaining
);
2987 r10_bio
->mddev
= mddev
;
2988 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
2989 r10_bio
->sector
= sect
;
2991 raid10_find_phys(conf
, r10_bio
);
2993 /* Need to check if the array will still be
2996 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
2997 if (conf
->mirrors
[j
].rdev
== NULL
||
2998 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3003 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3004 &sync_blocks
, still_degraded
);
3007 for (j
=0; j
<conf
->copies
;j
++) {
3009 int d
= r10_bio
->devs
[j
].devnum
;
3010 sector_t from_addr
, to_addr
;
3011 struct md_rdev
*rdev
;
3012 sector_t sector
, first_bad
;
3014 if (!conf
->mirrors
[d
].rdev
||
3015 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3017 /* This is where we read from */
3019 rdev
= conf
->mirrors
[d
].rdev
;
3020 sector
= r10_bio
->devs
[j
].addr
;
3022 if (is_badblock(rdev
, sector
, max_sync
,
3023 &first_bad
, &bad_sectors
)) {
3024 if (first_bad
> sector
)
3025 max_sync
= first_bad
- sector
;
3027 bad_sectors
-= (sector
3029 if (max_sync
> bad_sectors
)
3030 max_sync
= bad_sectors
;
3034 bio
= r10_bio
->devs
[0].bio
;
3036 bio
->bi_next
= biolist
;
3038 bio
->bi_private
= r10_bio
;
3039 bio
->bi_end_io
= end_sync_read
;
3040 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3041 from_addr
= r10_bio
->devs
[j
].addr
;
3042 bio
->bi_iter
.bi_sector
= from_addr
+
3044 bio
->bi_bdev
= rdev
->bdev
;
3045 atomic_inc(&rdev
->nr_pending
);
3046 /* and we write to 'i' (if not in_sync) */
3048 for (k
=0; k
<conf
->copies
; k
++)
3049 if (r10_bio
->devs
[k
].devnum
== i
)
3051 BUG_ON(k
== conf
->copies
);
3052 to_addr
= r10_bio
->devs
[k
].addr
;
3053 r10_bio
->devs
[0].devnum
= d
;
3054 r10_bio
->devs
[0].addr
= from_addr
;
3055 r10_bio
->devs
[1].devnum
= i
;
3056 r10_bio
->devs
[1].addr
= to_addr
;
3058 rdev
= mirror
->rdev
;
3059 if (!test_bit(In_sync
, &rdev
->flags
)) {
3060 bio
= r10_bio
->devs
[1].bio
;
3062 bio
->bi_next
= biolist
;
3064 bio
->bi_private
= r10_bio
;
3065 bio
->bi_end_io
= end_sync_write
;
3066 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3067 bio
->bi_iter
.bi_sector
= to_addr
3068 + rdev
->data_offset
;
3069 bio
->bi_bdev
= rdev
->bdev
;
3070 atomic_inc(&r10_bio
->remaining
);
3072 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3074 /* and maybe write to replacement */
3075 bio
= r10_bio
->devs
[1].repl_bio
;
3077 bio
->bi_end_io
= NULL
;
3078 rdev
= mirror
->replacement
;
3079 /* Note: if rdev != NULL, then bio
3080 * cannot be NULL as r10buf_pool_alloc will
3081 * have allocated it.
3082 * So the second test here is pointless.
3083 * But it keeps semantic-checkers happy, and
3084 * this comment keeps human reviewers
3087 if (rdev
== NULL
|| bio
== NULL
||
3088 test_bit(Faulty
, &rdev
->flags
))
3091 bio
->bi_next
= biolist
;
3093 bio
->bi_private
= r10_bio
;
3094 bio
->bi_end_io
= end_sync_write
;
3095 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3096 bio
->bi_iter
.bi_sector
= to_addr
+
3098 bio
->bi_bdev
= rdev
->bdev
;
3099 atomic_inc(&r10_bio
->remaining
);
3102 if (j
== conf
->copies
) {
3103 /* Cannot recover, so abort the recovery or
3104 * record a bad block */
3106 /* problem is that there are bad blocks
3107 * on other device(s)
3110 for (k
= 0; k
< conf
->copies
; k
++)
3111 if (r10_bio
->devs
[k
].devnum
== i
)
3113 if (!test_bit(In_sync
,
3114 &mirror
->rdev
->flags
)
3115 && !rdev_set_badblocks(
3117 r10_bio
->devs
[k
].addr
,
3120 if (mirror
->replacement
&&
3121 !rdev_set_badblocks(
3122 mirror
->replacement
,
3123 r10_bio
->devs
[k
].addr
,
3128 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3130 printk(KERN_INFO
"md/raid10:%s: insufficient "
3131 "working devices for recovery.\n",
3133 mirror
->recovery_disabled
3134 = mddev
->recovery_disabled
;
3138 atomic_dec(&rb2
->remaining
);
3143 if (biolist
== NULL
) {
3145 struct r10bio
*rb2
= r10_bio
;
3146 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3147 rb2
->master_bio
= NULL
;
3153 /* resync. Schedule a read for every block at this virt offset */
3156 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, 0);
3158 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3159 &sync_blocks
, mddev
->degraded
) &&
3160 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3161 &mddev
->recovery
)) {
3162 /* We can skip this block */
3164 return sync_blocks
+ sectors_skipped
;
3166 if (sync_blocks
< max_sync
)
3167 max_sync
= sync_blocks
;
3168 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3171 r10_bio
->mddev
= mddev
;
3172 atomic_set(&r10_bio
->remaining
, 0);
3173 raise_barrier(conf
, 0);
3174 conf
->next_resync
= sector_nr
;
3176 r10_bio
->master_bio
= NULL
;
3177 r10_bio
->sector
= sector_nr
;
3178 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3179 raid10_find_phys(conf
, r10_bio
);
3180 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3182 for (i
= 0; i
< conf
->copies
; i
++) {
3183 int d
= r10_bio
->devs
[i
].devnum
;
3184 sector_t first_bad
, sector
;
3187 if (r10_bio
->devs
[i
].repl_bio
)
3188 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3190 bio
= r10_bio
->devs
[i
].bio
;
3192 bio
->bi_error
= -EIO
;
3193 if (conf
->mirrors
[d
].rdev
== NULL
||
3194 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3196 sector
= r10_bio
->devs
[i
].addr
;
3197 if (is_badblock(conf
->mirrors
[d
].rdev
,
3199 &first_bad
, &bad_sectors
)) {
3200 if (first_bad
> sector
)
3201 max_sync
= first_bad
- sector
;
3203 bad_sectors
-= (sector
- first_bad
);
3204 if (max_sync
> bad_sectors
)
3205 max_sync
= bad_sectors
;
3209 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3210 atomic_inc(&r10_bio
->remaining
);
3211 bio
->bi_next
= biolist
;
3213 bio
->bi_private
= r10_bio
;
3214 bio
->bi_end_io
= end_sync_read
;
3215 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
3216 bio
->bi_iter
.bi_sector
= sector
+
3217 conf
->mirrors
[d
].rdev
->data_offset
;
3218 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3221 if (conf
->mirrors
[d
].replacement
== NULL
||
3223 &conf
->mirrors
[d
].replacement
->flags
))
3226 /* Need to set up for writing to the replacement */
3227 bio
= r10_bio
->devs
[i
].repl_bio
;
3229 bio
->bi_error
= -EIO
;
3231 sector
= r10_bio
->devs
[i
].addr
;
3232 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3233 bio
->bi_next
= biolist
;
3235 bio
->bi_private
= r10_bio
;
3236 bio
->bi_end_io
= end_sync_write
;
3237 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
3238 bio
->bi_iter
.bi_sector
= sector
+
3239 conf
->mirrors
[d
].replacement
->data_offset
;
3240 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3245 for (i
=0; i
<conf
->copies
; i
++) {
3246 int d
= r10_bio
->devs
[i
].devnum
;
3247 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3248 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3250 if (r10_bio
->devs
[i
].repl_bio
&&
3251 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3253 conf
->mirrors
[d
].replacement
,
3263 if (sector_nr
+ max_sync
< max_sector
)
3264 max_sector
= sector_nr
+ max_sync
;
3267 int len
= PAGE_SIZE
;
3268 if (sector_nr
+ (len
>>9) > max_sector
)
3269 len
= (max_sector
- sector_nr
) << 9;
3272 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3274 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3275 if (bio_add_page(bio
, page
, len
, 0))
3279 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3280 for (bio2
= biolist
;
3281 bio2
&& bio2
!= bio
;
3282 bio2
= bio2
->bi_next
) {
3283 /* remove last page from this bio */
3285 bio2
->bi_iter
.bi_size
-= len
;
3286 bio_clear_flag(bio2
, BIO_SEG_VALID
);
3290 nr_sectors
+= len
>>9;
3291 sector_nr
+= len
>>9;
3292 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3294 r10_bio
->sectors
= nr_sectors
;
3298 biolist
= biolist
->bi_next
;
3300 bio
->bi_next
= NULL
;
3301 r10_bio
= bio
->bi_private
;
3302 r10_bio
->sectors
= nr_sectors
;
3304 if (bio
->bi_end_io
== end_sync_read
) {
3305 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3307 generic_make_request(bio
);
3311 if (sectors_skipped
)
3312 /* pretend they weren't skipped, it makes
3313 * no important difference in this case
3315 md_done_sync(mddev
, sectors_skipped
, 1);
3317 return sectors_skipped
+ nr_sectors
;
3319 /* There is nowhere to write, so all non-sync
3320 * drives must be failed or in resync, all drives
3321 * have a bad block, so try the next chunk...
3323 if (sector_nr
+ max_sync
< max_sector
)
3324 max_sector
= sector_nr
+ max_sync
;
3326 sectors_skipped
+= (max_sector
- sector_nr
);
3328 sector_nr
= max_sector
;
3333 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3336 struct r10conf
*conf
= mddev
->private;
3339 raid_disks
= min(conf
->geo
.raid_disks
,
3340 conf
->prev
.raid_disks
);
3342 sectors
= conf
->dev_sectors
;
3344 size
= sectors
>> conf
->geo
.chunk_shift
;
3345 sector_div(size
, conf
->geo
.far_copies
);
3346 size
= size
* raid_disks
;
3347 sector_div(size
, conf
->geo
.near_copies
);
3349 return size
<< conf
->geo
.chunk_shift
;
3352 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3354 /* Calculate the number of sectors-per-device that will
3355 * actually be used, and set conf->dev_sectors and
3359 size
= size
>> conf
->geo
.chunk_shift
;
3360 sector_div(size
, conf
->geo
.far_copies
);
3361 size
= size
* conf
->geo
.raid_disks
;
3362 sector_div(size
, conf
->geo
.near_copies
);
3363 /* 'size' is now the number of chunks in the array */
3364 /* calculate "used chunks per device" */
3365 size
= size
* conf
->copies
;
3367 /* We need to round up when dividing by raid_disks to
3368 * get the stride size.
3370 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3372 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3374 if (conf
->geo
.far_offset
)
3375 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3377 sector_div(size
, conf
->geo
.far_copies
);
3378 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3382 enum geo_type
{geo_new
, geo_old
, geo_start
};
3383 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3386 int layout
, chunk
, disks
;
3389 layout
= mddev
->layout
;
3390 chunk
= mddev
->chunk_sectors
;
3391 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3394 layout
= mddev
->new_layout
;
3395 chunk
= mddev
->new_chunk_sectors
;
3396 disks
= mddev
->raid_disks
;
3398 default: /* avoid 'may be unused' warnings */
3399 case geo_start
: /* new when starting reshape - raid_disks not
3401 layout
= mddev
->new_layout
;
3402 chunk
= mddev
->new_chunk_sectors
;
3403 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3408 if (chunk
< (PAGE_SIZE
>> 9) ||
3409 !is_power_of_2(chunk
))
3412 fc
= (layout
>> 8) & 255;
3413 fo
= layout
& (1<<16);
3414 geo
->raid_disks
= disks
;
3415 geo
->near_copies
= nc
;
3416 geo
->far_copies
= fc
;
3417 geo
->far_offset
= fo
;
3418 switch (layout
>> 17) {
3419 case 0: /* original layout. simple but not always optimal */
3420 geo
->far_set_size
= disks
;
3422 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3423 * actually using this, but leave code here just in case.*/
3424 geo
->far_set_size
= disks
/fc
;
3425 WARN(geo
->far_set_size
< fc
,
3426 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3428 case 2: /* "improved" layout fixed to match documentation */
3429 geo
->far_set_size
= fc
* nc
;
3431 default: /* Not a valid layout */
3434 geo
->chunk_mask
= chunk
- 1;
3435 geo
->chunk_shift
= ffz(~chunk
);
3439 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3441 struct r10conf
*conf
= NULL
;
3446 copies
= setup_geo(&geo
, mddev
, geo_new
);
3449 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3450 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3451 mdname(mddev
), PAGE_SIZE
);
3455 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3456 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3457 mdname(mddev
), mddev
->new_layout
);
3462 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3466 /* FIXME calc properly */
3467 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3468 max(0,-mddev
->delta_disks
)),
3473 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3478 conf
->copies
= copies
;
3479 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3480 r10bio_pool_free
, conf
);
3481 if (!conf
->r10bio_pool
)
3484 calc_sectors(conf
, mddev
->dev_sectors
);
3485 if (mddev
->reshape_position
== MaxSector
) {
3486 conf
->prev
= conf
->geo
;
3487 conf
->reshape_progress
= MaxSector
;
3489 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3493 conf
->reshape_progress
= mddev
->reshape_position
;
3494 if (conf
->prev
.far_offset
)
3495 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3497 /* far_copies must be 1 */
3498 conf
->prev
.stride
= conf
->dev_sectors
;
3500 conf
->reshape_safe
= conf
->reshape_progress
;
3501 spin_lock_init(&conf
->device_lock
);
3502 INIT_LIST_HEAD(&conf
->retry_list
);
3503 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
3505 spin_lock_init(&conf
->resync_lock
);
3506 init_waitqueue_head(&conf
->wait_barrier
);
3508 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3512 conf
->mddev
= mddev
;
3517 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3520 mempool_destroy(conf
->r10bio_pool
);
3521 kfree(conf
->mirrors
);
3522 safe_put_page(conf
->tmppage
);
3525 return ERR_PTR(err
);
3528 static int raid10_run(struct mddev
*mddev
)
3530 struct r10conf
*conf
;
3531 int i
, disk_idx
, chunk_size
;
3532 struct raid10_info
*disk
;
3533 struct md_rdev
*rdev
;
3535 sector_t min_offset_diff
= 0;
3537 bool discard_supported
= false;
3539 if (mddev
->private == NULL
) {
3540 conf
= setup_conf(mddev
);
3542 return PTR_ERR(conf
);
3543 mddev
->private = conf
;
3545 conf
= mddev
->private;
3549 mddev
->thread
= conf
->thread
;
3550 conf
->thread
= NULL
;
3552 chunk_size
= mddev
->chunk_sectors
<< 9;
3554 blk_queue_max_discard_sectors(mddev
->queue
,
3555 mddev
->chunk_sectors
);
3556 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3557 blk_queue_io_min(mddev
->queue
, chunk_size
);
3558 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3559 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3561 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3562 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3565 rdev_for_each(rdev
, mddev
) {
3567 struct request_queue
*q
;
3569 disk_idx
= rdev
->raid_disk
;
3572 if (disk_idx
>= conf
->geo
.raid_disks
&&
3573 disk_idx
>= conf
->prev
.raid_disks
)
3575 disk
= conf
->mirrors
+ disk_idx
;
3577 if (test_bit(Replacement
, &rdev
->flags
)) {
3578 if (disk
->replacement
)
3580 disk
->replacement
= rdev
;
3586 q
= bdev_get_queue(rdev
->bdev
);
3587 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3588 if (!mddev
->reshape_backwards
)
3592 if (first
|| diff
< min_offset_diff
)
3593 min_offset_diff
= diff
;
3596 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3597 rdev
->data_offset
<< 9);
3599 disk
->head_position
= 0;
3601 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3602 discard_supported
= true;
3606 if (discard_supported
)
3607 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3610 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3613 /* need to check that every block has at least one working mirror */
3614 if (!enough(conf
, -1)) {
3615 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3620 if (conf
->reshape_progress
!= MaxSector
) {
3621 /* must ensure that shape change is supported */
3622 if (conf
->geo
.far_copies
!= 1 &&
3623 conf
->geo
.far_offset
== 0)
3625 if (conf
->prev
.far_copies
!= 1 &&
3626 conf
->prev
.far_offset
== 0)
3630 mddev
->degraded
= 0;
3632 i
< conf
->geo
.raid_disks
3633 || i
< conf
->prev
.raid_disks
;
3636 disk
= conf
->mirrors
+ i
;
3638 if (!disk
->rdev
&& disk
->replacement
) {
3639 /* The replacement is all we have - use it */
3640 disk
->rdev
= disk
->replacement
;
3641 disk
->replacement
= NULL
;
3642 clear_bit(Replacement
, &disk
->rdev
->flags
);
3646 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3647 disk
->head_position
= 0;
3650 disk
->rdev
->saved_raid_disk
< 0)
3653 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3656 if (mddev
->recovery_cp
!= MaxSector
)
3657 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3658 " -- starting background reconstruction\n",
3661 "md/raid10:%s: active with %d out of %d devices\n",
3662 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3663 conf
->geo
.raid_disks
);
3665 * Ok, everything is just fine now
3667 mddev
->dev_sectors
= conf
->dev_sectors
;
3668 size
= raid10_size(mddev
, 0, 0);
3669 md_set_array_sectors(mddev
, size
);
3670 mddev
->resync_max_sectors
= size
;
3673 int stripe
= conf
->geo
.raid_disks
*
3674 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3676 /* Calculate max read-ahead size.
3677 * We need to readahead at least twice a whole stripe....
3680 stripe
/= conf
->geo
.near_copies
;
3681 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3682 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3685 if (md_integrity_register(mddev
))
3688 if (conf
->reshape_progress
!= MaxSector
) {
3689 unsigned long before_length
, after_length
;
3691 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3692 conf
->prev
.far_copies
);
3693 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3694 conf
->geo
.far_copies
);
3696 if (max(before_length
, after_length
) > min_offset_diff
) {
3697 /* This cannot work */
3698 printk("md/raid10: offset difference not enough to continue reshape\n");
3701 conf
->offset_diff
= min_offset_diff
;
3703 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3704 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3705 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3706 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3707 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3714 md_unregister_thread(&mddev
->thread
);
3715 mempool_destroy(conf
->r10bio_pool
);
3716 safe_put_page(conf
->tmppage
);
3717 kfree(conf
->mirrors
);
3719 mddev
->private = NULL
;
3724 static void raid10_free(struct mddev
*mddev
, void *priv
)
3726 struct r10conf
*conf
= priv
;
3728 mempool_destroy(conf
->r10bio_pool
);
3729 safe_put_page(conf
->tmppage
);
3730 kfree(conf
->mirrors
);
3731 kfree(conf
->mirrors_old
);
3732 kfree(conf
->mirrors_new
);
3736 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3738 struct r10conf
*conf
= mddev
->private;
3742 raise_barrier(conf
, 0);
3745 lower_barrier(conf
);
3750 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3752 /* Resize of 'far' arrays is not supported.
3753 * For 'near' and 'offset' arrays we can set the
3754 * number of sectors used to be an appropriate multiple
3755 * of the chunk size.
3756 * For 'offset', this is far_copies*chunksize.
3757 * For 'near' the multiplier is the LCM of
3758 * near_copies and raid_disks.
3759 * So if far_copies > 1 && !far_offset, fail.
3760 * Else find LCM(raid_disks, near_copy)*far_copies and
3761 * multiply by chunk_size. Then round to this number.
3762 * This is mostly done by raid10_size()
3764 struct r10conf
*conf
= mddev
->private;
3765 sector_t oldsize
, size
;
3767 if (mddev
->reshape_position
!= MaxSector
)
3770 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3773 oldsize
= raid10_size(mddev
, 0, 0);
3774 size
= raid10_size(mddev
, sectors
, 0);
3775 if (mddev
->external_size
&&
3776 mddev
->array_sectors
> size
)
3778 if (mddev
->bitmap
) {
3779 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3783 md_set_array_sectors(mddev
, size
);
3785 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3786 revalidate_disk(mddev
->gendisk
);
3788 if (sectors
> mddev
->dev_sectors
&&
3789 mddev
->recovery_cp
> oldsize
) {
3790 mddev
->recovery_cp
= oldsize
;
3791 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3793 calc_sectors(conf
, sectors
);
3794 mddev
->dev_sectors
= conf
->dev_sectors
;
3795 mddev
->resync_max_sectors
= size
;
3799 static void *raid10_takeover_raid0(struct mddev
*mddev
, sector_t size
, int devs
)
3801 struct md_rdev
*rdev
;
3802 struct r10conf
*conf
;
3804 if (mddev
->degraded
> 0) {
3805 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3807 return ERR_PTR(-EINVAL
);
3809 sector_div(size
, devs
);
3811 /* Set new parameters */
3812 mddev
->new_level
= 10;
3813 /* new layout: far_copies = 1, near_copies = 2 */
3814 mddev
->new_layout
= (1<<8) + 2;
3815 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3816 mddev
->delta_disks
= mddev
->raid_disks
;
3817 mddev
->raid_disks
*= 2;
3818 /* make sure it will be not marked as dirty */
3819 mddev
->recovery_cp
= MaxSector
;
3820 mddev
->dev_sectors
= size
;
3822 conf
= setup_conf(mddev
);
3823 if (!IS_ERR(conf
)) {
3824 rdev_for_each(rdev
, mddev
)
3825 if (rdev
->raid_disk
>= 0) {
3826 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3827 rdev
->sectors
= size
;
3835 static void *raid10_takeover(struct mddev
*mddev
)
3837 struct r0conf
*raid0_conf
;
3839 /* raid10 can take over:
3840 * raid0 - providing it has only two drives
3842 if (mddev
->level
== 0) {
3843 /* for raid0 takeover only one zone is supported */
3844 raid0_conf
= mddev
->private;
3845 if (raid0_conf
->nr_strip_zones
> 1) {
3846 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3847 " with more than one zone.\n",
3849 return ERR_PTR(-EINVAL
);
3851 return raid10_takeover_raid0(mddev
,
3852 raid0_conf
->strip_zone
->zone_end
,
3853 raid0_conf
->strip_zone
->nb_dev
);
3855 return ERR_PTR(-EINVAL
);
3858 static int raid10_check_reshape(struct mddev
*mddev
)
3860 /* Called when there is a request to change
3861 * - layout (to ->new_layout)
3862 * - chunk size (to ->new_chunk_sectors)
3863 * - raid_disks (by delta_disks)
3864 * or when trying to restart a reshape that was ongoing.
3866 * We need to validate the request and possibly allocate
3867 * space if that might be an issue later.
3869 * Currently we reject any reshape of a 'far' mode array,
3870 * allow chunk size to change if new is generally acceptable,
3871 * allow raid_disks to increase, and allow
3872 * a switch between 'near' mode and 'offset' mode.
3874 struct r10conf
*conf
= mddev
->private;
3877 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3880 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3881 /* mustn't change number of copies */
3883 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3884 /* Cannot switch to 'far' mode */
3887 if (mddev
->array_sectors
& geo
.chunk_mask
)
3888 /* not factor of array size */
3891 if (!enough(conf
, -1))
3894 kfree(conf
->mirrors_new
);
3895 conf
->mirrors_new
= NULL
;
3896 if (mddev
->delta_disks
> 0) {
3897 /* allocate new 'mirrors' list */
3898 conf
->mirrors_new
= kzalloc(
3899 sizeof(struct raid10_info
)
3900 *(mddev
->raid_disks
+
3901 mddev
->delta_disks
),
3903 if (!conf
->mirrors_new
)
3910 * Need to check if array has failed when deciding whether to:
3912 * - remove non-faulty devices
3915 * This determination is simple when no reshape is happening.
3916 * However if there is a reshape, we need to carefully check
3917 * both the before and after sections.
3918 * This is because some failed devices may only affect one
3919 * of the two sections, and some non-in_sync devices may
3920 * be insync in the section most affected by failed devices.
3922 static int calc_degraded(struct r10conf
*conf
)
3924 int degraded
, degraded2
;
3929 /* 'prev' section first */
3930 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
3931 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3932 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3934 else if (!test_bit(In_sync
, &rdev
->flags
))
3935 /* When we can reduce the number of devices in
3936 * an array, this might not contribute to
3937 * 'degraded'. It does now.
3942 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
3946 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3947 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3948 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3950 else if (!test_bit(In_sync
, &rdev
->flags
)) {
3951 /* If reshape is increasing the number of devices,
3952 * this section has already been recovered, so
3953 * it doesn't contribute to degraded.
3956 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
3961 if (degraded2
> degraded
)
3966 static int raid10_start_reshape(struct mddev
*mddev
)
3968 /* A 'reshape' has been requested. This commits
3969 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3970 * This also checks if there are enough spares and adds them
3972 * We currently require enough spares to make the final
3973 * array non-degraded. We also require that the difference
3974 * between old and new data_offset - on each device - is
3975 * enough that we never risk over-writing.
3978 unsigned long before_length
, after_length
;
3979 sector_t min_offset_diff
= 0;
3982 struct r10conf
*conf
= mddev
->private;
3983 struct md_rdev
*rdev
;
3987 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
3990 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
3993 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3994 conf
->prev
.far_copies
);
3995 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3996 conf
->geo
.far_copies
);
3998 rdev_for_each(rdev
, mddev
) {
3999 if (!test_bit(In_sync
, &rdev
->flags
)
4000 && !test_bit(Faulty
, &rdev
->flags
))
4002 if (rdev
->raid_disk
>= 0) {
4003 long long diff
= (rdev
->new_data_offset
4004 - rdev
->data_offset
);
4005 if (!mddev
->reshape_backwards
)
4009 if (first
|| diff
< min_offset_diff
)
4010 min_offset_diff
= diff
;
4014 if (max(before_length
, after_length
) > min_offset_diff
)
4017 if (spares
< mddev
->delta_disks
)
4020 conf
->offset_diff
= min_offset_diff
;
4021 spin_lock_irq(&conf
->device_lock
);
4022 if (conf
->mirrors_new
) {
4023 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4024 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4026 kfree(conf
->mirrors_old
);
4027 conf
->mirrors_old
= conf
->mirrors
;
4028 conf
->mirrors
= conf
->mirrors_new
;
4029 conf
->mirrors_new
= NULL
;
4031 setup_geo(&conf
->geo
, mddev
, geo_start
);
4033 if (mddev
->reshape_backwards
) {
4034 sector_t size
= raid10_size(mddev
, 0, 0);
4035 if (size
< mddev
->array_sectors
) {
4036 spin_unlock_irq(&conf
->device_lock
);
4037 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4041 mddev
->resync_max_sectors
= size
;
4042 conf
->reshape_progress
= size
;
4044 conf
->reshape_progress
= 0;
4045 conf
->reshape_safe
= conf
->reshape_progress
;
4046 spin_unlock_irq(&conf
->device_lock
);
4048 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4049 ret
= bitmap_resize(mddev
->bitmap
,
4050 raid10_size(mddev
, 0,
4051 conf
->geo
.raid_disks
),
4056 if (mddev
->delta_disks
> 0) {
4057 rdev_for_each(rdev
, mddev
)
4058 if (rdev
->raid_disk
< 0 &&
4059 !test_bit(Faulty
, &rdev
->flags
)) {
4060 if (raid10_add_disk(mddev
, rdev
) == 0) {
4061 if (rdev
->raid_disk
>=
4062 conf
->prev
.raid_disks
)
4063 set_bit(In_sync
, &rdev
->flags
);
4065 rdev
->recovery_offset
= 0;
4067 if (sysfs_link_rdev(mddev
, rdev
))
4068 /* Failure here is OK */;
4070 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4071 && !test_bit(Faulty
, &rdev
->flags
)) {
4072 /* This is a spare that was manually added */
4073 set_bit(In_sync
, &rdev
->flags
);
4076 /* When a reshape changes the number of devices,
4077 * ->degraded is measured against the larger of the
4078 * pre and post numbers.
4080 spin_lock_irq(&conf
->device_lock
);
4081 mddev
->degraded
= calc_degraded(conf
);
4082 spin_unlock_irq(&conf
->device_lock
);
4083 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4084 mddev
->reshape_position
= conf
->reshape_progress
;
4085 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4087 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4088 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4089 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
4090 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4091 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4093 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4095 if (!mddev
->sync_thread
) {
4099 conf
->reshape_checkpoint
= jiffies
;
4100 md_wakeup_thread(mddev
->sync_thread
);
4101 md_new_event(mddev
);
4105 mddev
->recovery
= 0;
4106 spin_lock_irq(&conf
->device_lock
);
4107 conf
->geo
= conf
->prev
;
4108 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4109 rdev_for_each(rdev
, mddev
)
4110 rdev
->new_data_offset
= rdev
->data_offset
;
4112 conf
->reshape_progress
= MaxSector
;
4113 conf
->reshape_safe
= MaxSector
;
4114 mddev
->reshape_position
= MaxSector
;
4115 spin_unlock_irq(&conf
->device_lock
);
4119 /* Calculate the last device-address that could contain
4120 * any block from the chunk that includes the array-address 's'
4121 * and report the next address.
4122 * i.e. the address returned will be chunk-aligned and after
4123 * any data that is in the chunk containing 's'.
4125 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4127 s
= (s
| geo
->chunk_mask
) + 1;
4128 s
>>= geo
->chunk_shift
;
4129 s
*= geo
->near_copies
;
4130 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4131 s
*= geo
->far_copies
;
4132 s
<<= geo
->chunk_shift
;
4136 /* Calculate the first device-address that could contain
4137 * any block from the chunk that includes the array-address 's'.
4138 * This too will be the start of a chunk
4140 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4142 s
>>= geo
->chunk_shift
;
4143 s
*= geo
->near_copies
;
4144 sector_div(s
, geo
->raid_disks
);
4145 s
*= geo
->far_copies
;
4146 s
<<= geo
->chunk_shift
;
4150 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4153 /* We simply copy at most one chunk (smallest of old and new)
4154 * at a time, possibly less if that exceeds RESYNC_PAGES,
4155 * or we hit a bad block or something.
4156 * This might mean we pause for normal IO in the middle of
4157 * a chunk, but that is not a problem as mddev->reshape_position
4158 * can record any location.
4160 * If we will want to write to a location that isn't
4161 * yet recorded as 'safe' (i.e. in metadata on disk) then
4162 * we need to flush all reshape requests and update the metadata.
4164 * When reshaping forwards (e.g. to more devices), we interpret
4165 * 'safe' as the earliest block which might not have been copied
4166 * down yet. We divide this by previous stripe size and multiply
4167 * by previous stripe length to get lowest device offset that we
4168 * cannot write to yet.
4169 * We interpret 'sector_nr' as an address that we want to write to.
4170 * From this we use last_device_address() to find where we might
4171 * write to, and first_device_address on the 'safe' position.
4172 * If this 'next' write position is after the 'safe' position,
4173 * we must update the metadata to increase the 'safe' position.
4175 * When reshaping backwards, we round in the opposite direction
4176 * and perform the reverse test: next write position must not be
4177 * less than current safe position.
4179 * In all this the minimum difference in data offsets
4180 * (conf->offset_diff - always positive) allows a bit of slack,
4181 * so next can be after 'safe', but not by more than offset_diff
4183 * We need to prepare all the bios here before we start any IO
4184 * to ensure the size we choose is acceptable to all devices.
4185 * The means one for each copy for write-out and an extra one for
4187 * We store the read-in bio in ->master_bio and the others in
4188 * ->devs[x].bio and ->devs[x].repl_bio.
4190 struct r10conf
*conf
= mddev
->private;
4191 struct r10bio
*r10_bio
;
4192 sector_t next
, safe
, last
;
4196 struct md_rdev
*rdev
;
4199 struct bio
*bio
, *read_bio
;
4200 int sectors_done
= 0;
4202 if (sector_nr
== 0) {
4203 /* If restarting in the middle, skip the initial sectors */
4204 if (mddev
->reshape_backwards
&&
4205 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4206 sector_nr
= (raid10_size(mddev
, 0, 0)
4207 - conf
->reshape_progress
);
4208 } else if (!mddev
->reshape_backwards
&&
4209 conf
->reshape_progress
> 0)
4210 sector_nr
= conf
->reshape_progress
;
4212 mddev
->curr_resync_completed
= sector_nr
;
4213 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4219 /* We don't use sector_nr to track where we are up to
4220 * as that doesn't work well for ->reshape_backwards.
4221 * So just use ->reshape_progress.
4223 if (mddev
->reshape_backwards
) {
4224 /* 'next' is the earliest device address that we might
4225 * write to for this chunk in the new layout
4227 next
= first_dev_address(conf
->reshape_progress
- 1,
4230 /* 'safe' is the last device address that we might read from
4231 * in the old layout after a restart
4233 safe
= last_dev_address(conf
->reshape_safe
- 1,
4236 if (next
+ conf
->offset_diff
< safe
)
4239 last
= conf
->reshape_progress
- 1;
4240 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4241 & conf
->prev
.chunk_mask
);
4242 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4243 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4245 /* 'next' is after the last device address that we
4246 * might write to for this chunk in the new layout
4248 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4250 /* 'safe' is the earliest device address that we might
4251 * read from in the old layout after a restart
4253 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4255 /* Need to update metadata if 'next' might be beyond 'safe'
4256 * as that would possibly corrupt data
4258 if (next
> safe
+ conf
->offset_diff
)
4261 sector_nr
= conf
->reshape_progress
;
4262 last
= sector_nr
| (conf
->geo
.chunk_mask
4263 & conf
->prev
.chunk_mask
);
4265 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4266 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4270 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4271 /* Need to update reshape_position in metadata */
4273 mddev
->reshape_position
= conf
->reshape_progress
;
4274 if (mddev
->reshape_backwards
)
4275 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4276 - conf
->reshape_progress
;
4278 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4279 conf
->reshape_checkpoint
= jiffies
;
4280 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4281 md_wakeup_thread(mddev
->thread
);
4282 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4283 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
4284 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4285 allow_barrier(conf
);
4286 return sectors_done
;
4288 conf
->reshape_safe
= mddev
->reshape_position
;
4289 allow_barrier(conf
);
4293 /* Now schedule reads for blocks from sector_nr to last */
4294 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4296 raise_barrier(conf
, sectors_done
!= 0);
4297 atomic_set(&r10_bio
->remaining
, 0);
4298 r10_bio
->mddev
= mddev
;
4299 r10_bio
->sector
= sector_nr
;
4300 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4301 r10_bio
->sectors
= last
- sector_nr
+ 1;
4302 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4303 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4306 /* Cannot read from here, so need to record bad blocks
4307 * on all the target devices.
4310 mempool_free(r10_bio
, conf
->r10buf_pool
);
4311 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4312 return sectors_done
;
4315 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4317 read_bio
->bi_bdev
= rdev
->bdev
;
4318 read_bio
->bi_iter
.bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4319 + rdev
->data_offset
);
4320 read_bio
->bi_private
= r10_bio
;
4321 read_bio
->bi_end_io
= end_sync_read
;
4322 bio_set_op_attrs(read_bio
, REQ_OP_READ
, 0);
4323 read_bio
->bi_flags
&= (~0UL << BIO_RESET_BITS
);
4324 read_bio
->bi_error
= 0;
4325 read_bio
->bi_vcnt
= 0;
4326 read_bio
->bi_iter
.bi_size
= 0;
4327 r10_bio
->master_bio
= read_bio
;
4328 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4330 /* Now find the locations in the new layout */
4331 __raid10_find_phys(&conf
->geo
, r10_bio
);
4334 read_bio
->bi_next
= NULL
;
4336 for (s
= 0; s
< conf
->copies
*2; s
++) {
4338 int d
= r10_bio
->devs
[s
/2].devnum
;
4339 struct md_rdev
*rdev2
;
4341 rdev2
= conf
->mirrors
[d
].replacement
;
4342 b
= r10_bio
->devs
[s
/2].repl_bio
;
4344 rdev2
= conf
->mirrors
[d
].rdev
;
4345 b
= r10_bio
->devs
[s
/2].bio
;
4347 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4351 b
->bi_bdev
= rdev2
->bdev
;
4352 b
->bi_iter
.bi_sector
= r10_bio
->devs
[s
/2].addr
+
4353 rdev2
->new_data_offset
;
4354 b
->bi_private
= r10_bio
;
4355 b
->bi_end_io
= end_reshape_write
;
4356 bio_set_op_attrs(b
, REQ_OP_WRITE
, 0);
4361 /* Now add as many pages as possible to all of these bios. */
4364 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4365 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4366 int len
= (max_sectors
- s
) << 9;
4367 if (len
> PAGE_SIZE
)
4369 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4371 if (bio_add_page(bio
, page
, len
, 0))
4374 /* Didn't fit, must stop */
4376 bio2
&& bio2
!= bio
;
4377 bio2
= bio2
->bi_next
) {
4378 /* Remove last page from this bio */
4380 bio2
->bi_iter
.bi_size
-= len
;
4381 bio_clear_flag(bio2
, BIO_SEG_VALID
);
4385 sector_nr
+= len
>> 9;
4386 nr_sectors
+= len
>> 9;
4389 r10_bio
->sectors
= nr_sectors
;
4391 /* Now submit the read */
4392 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4393 atomic_inc(&r10_bio
->remaining
);
4394 read_bio
->bi_next
= NULL
;
4395 generic_make_request(read_bio
);
4396 sector_nr
+= nr_sectors
;
4397 sectors_done
+= nr_sectors
;
4398 if (sector_nr
<= last
)
4401 /* Now that we have done the whole section we can
4402 * update reshape_progress
4404 if (mddev
->reshape_backwards
)
4405 conf
->reshape_progress
-= sectors_done
;
4407 conf
->reshape_progress
+= sectors_done
;
4409 return sectors_done
;
4412 static void end_reshape_request(struct r10bio
*r10_bio
);
4413 static int handle_reshape_read_error(struct mddev
*mddev
,
4414 struct r10bio
*r10_bio
);
4415 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4417 /* Reshape read completed. Hopefully we have a block
4419 * If we got a read error then we do sync 1-page reads from
4420 * elsewhere until we find the data - or give up.
4422 struct r10conf
*conf
= mddev
->private;
4425 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4426 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4427 /* Reshape has been aborted */
4428 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4432 /* We definitely have the data in the pages, schedule the
4435 atomic_set(&r10_bio
->remaining
, 1);
4436 for (s
= 0; s
< conf
->copies
*2; s
++) {
4438 int d
= r10_bio
->devs
[s
/2].devnum
;
4439 struct md_rdev
*rdev
;
4441 rdev
= conf
->mirrors
[d
].replacement
;
4442 b
= r10_bio
->devs
[s
/2].repl_bio
;
4444 rdev
= conf
->mirrors
[d
].rdev
;
4445 b
= r10_bio
->devs
[s
/2].bio
;
4447 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4449 atomic_inc(&rdev
->nr_pending
);
4450 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4451 atomic_inc(&r10_bio
->remaining
);
4453 generic_make_request(b
);
4455 end_reshape_request(r10_bio
);
4458 static void end_reshape(struct r10conf
*conf
)
4460 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4463 spin_lock_irq(&conf
->device_lock
);
4464 conf
->prev
= conf
->geo
;
4465 md_finish_reshape(conf
->mddev
);
4467 conf
->reshape_progress
= MaxSector
;
4468 conf
->reshape_safe
= MaxSector
;
4469 spin_unlock_irq(&conf
->device_lock
);
4471 /* read-ahead size must cover two whole stripes, which is
4472 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4474 if (conf
->mddev
->queue
) {
4475 int stripe
= conf
->geo
.raid_disks
*
4476 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4477 stripe
/= conf
->geo
.near_copies
;
4478 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4479 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4484 static int handle_reshape_read_error(struct mddev
*mddev
,
4485 struct r10bio
*r10_bio
)
4487 /* Use sync reads to get the blocks from somewhere else */
4488 int sectors
= r10_bio
->sectors
;
4489 struct r10conf
*conf
= mddev
->private;
4491 struct r10bio r10_bio
;
4492 struct r10dev devs
[conf
->copies
];
4494 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4497 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4499 r10b
->sector
= r10_bio
->sector
;
4500 __raid10_find_phys(&conf
->prev
, r10b
);
4505 int first_slot
= slot
;
4507 if (s
> (PAGE_SIZE
>> 9))
4511 int d
= r10b
->devs
[slot
].devnum
;
4512 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4515 test_bit(Faulty
, &rdev
->flags
) ||
4516 !test_bit(In_sync
, &rdev
->flags
))
4519 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4520 success
= sync_page_io(rdev
,
4524 REQ_OP_READ
, 0, false);
4529 if (slot
>= conf
->copies
)
4531 if (slot
== first_slot
)
4535 /* couldn't read this block, must give up */
4536 set_bit(MD_RECOVERY_INTR
,
4546 static void end_reshape_write(struct bio
*bio
)
4548 struct r10bio
*r10_bio
= bio
->bi_private
;
4549 struct mddev
*mddev
= r10_bio
->mddev
;
4550 struct r10conf
*conf
= mddev
->private;
4554 struct md_rdev
*rdev
= NULL
;
4556 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4558 rdev
= conf
->mirrors
[d
].replacement
;
4561 rdev
= conf
->mirrors
[d
].rdev
;
4564 if (bio
->bi_error
) {
4565 /* FIXME should record badblock */
4566 md_error(mddev
, rdev
);
4569 rdev_dec_pending(rdev
, mddev
);
4570 end_reshape_request(r10_bio
);
4573 static void end_reshape_request(struct r10bio
*r10_bio
)
4575 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4577 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4578 bio_put(r10_bio
->master_bio
);
4582 static void raid10_finish_reshape(struct mddev
*mddev
)
4584 struct r10conf
*conf
= mddev
->private;
4586 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4589 if (mddev
->delta_disks
> 0) {
4590 sector_t size
= raid10_size(mddev
, 0, 0);
4591 md_set_array_sectors(mddev
, size
);
4592 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4593 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4594 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4596 mddev
->resync_max_sectors
= size
;
4598 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4599 revalidate_disk(mddev
->gendisk
);
4603 for (d
= conf
->geo
.raid_disks
;
4604 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4606 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4608 clear_bit(In_sync
, &rdev
->flags
);
4609 rdev
= conf
->mirrors
[d
].replacement
;
4611 clear_bit(In_sync
, &rdev
->flags
);
4614 mddev
->layout
= mddev
->new_layout
;
4615 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4616 mddev
->reshape_position
= MaxSector
;
4617 mddev
->delta_disks
= 0;
4618 mddev
->reshape_backwards
= 0;
4621 static struct md_personality raid10_personality
=
4625 .owner
= THIS_MODULE
,
4626 .make_request
= raid10_make_request
,
4628 .free
= raid10_free
,
4629 .status
= raid10_status
,
4630 .error_handler
= raid10_error
,
4631 .hot_add_disk
= raid10_add_disk
,
4632 .hot_remove_disk
= raid10_remove_disk
,
4633 .spare_active
= raid10_spare_active
,
4634 .sync_request
= raid10_sync_request
,
4635 .quiesce
= raid10_quiesce
,
4636 .size
= raid10_size
,
4637 .resize
= raid10_resize
,
4638 .takeover
= raid10_takeover
,
4639 .check_reshape
= raid10_check_reshape
,
4640 .start_reshape
= raid10_start_reshape
,
4641 .finish_reshape
= raid10_finish_reshape
,
4642 .congested
= raid10_congested
,
4645 static int __init
raid_init(void)
4647 return register_md_personality(&raid10_personality
);
4650 static void raid_exit(void)
4652 unregister_md_personality(&raid10_personality
);
4655 module_init(raid_init
);
4656 module_exit(raid_exit
);
4657 MODULE_LICENSE("GPL");
4658 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4659 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4660 MODULE_ALIAS("md-raid10");
4661 MODULE_ALIAS("md-level-10");
4663 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);