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>
33 * RAID10 provides a combination of RAID0 and RAID1 functionality.
34 * The layout of data is defined by
37 * near_copies (stored in low byte of layout)
38 * far_copies (stored in second byte of layout)
39 * far_offset (stored in bit 16 of layout )
41 * The data to be stored is divided into chunks using chunksize.
42 * Each device is divided into far_copies sections.
43 * In each section, chunks are laid out in a style similar to raid0, but
44 * near_copies copies of each chunk is stored (each on a different drive).
45 * The starting device for each section is offset near_copies from the starting
46 * device of the previous section.
47 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
49 * near_copies and far_copies must be at least one, and their product is at most
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of be very far apart
54 * on disk, there are adjacent stripes.
58 * Number of guaranteed r10bios in case of extreme VM load:
60 #define NR_RAID10_BIOS 256
62 /* When there are this many requests queue to be written by
63 * the raid10 thread, we become 'congested' to provide back-pressure
66 static int max_queued_requests
= 1024;
68 static void allow_barrier(struct r10conf
*conf
);
69 static void lower_barrier(struct r10conf
*conf
);
71 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
73 struct r10conf
*conf
= data
;
74 int size
= offsetof(struct r10bio
, devs
[conf
->copies
]);
76 /* allocate a r10bio with room for raid_disks entries in the
78 return kzalloc(size
, gfp_flags
);
81 static void r10bio_pool_free(void *r10_bio
, void *data
)
86 /* Maximum size of each resync request */
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
89 /* amount of memory to reserve for resync requests */
90 #define RESYNC_WINDOW (1024*1024)
91 /* maximum number of concurrent requests, memory permitting */
92 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
95 * When performing a resync, we need to read and compare, so
96 * we need as many pages are there are copies.
97 * When performing a recovery, we need 2 bios, one for read,
98 * one for write (we recover only one drive per r10buf)
101 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
103 struct r10conf
*conf
= data
;
105 struct r10bio
*r10_bio
;
110 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
114 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
115 nalloc
= conf
->copies
; /* resync */
117 nalloc
= 2; /* recovery */
122 for (j
= nalloc
; j
-- ; ) {
123 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
126 r10_bio
->devs
[j
].bio
= bio
;
127 if (!conf
->have_replacement
)
129 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
132 r10_bio
->devs
[j
].repl_bio
= bio
;
135 * Allocate RESYNC_PAGES data pages and attach them
138 for (j
= 0 ; j
< nalloc
; j
++) {
139 struct bio
*rbio
= r10_bio
->devs
[j
].repl_bio
;
140 bio
= r10_bio
->devs
[j
].bio
;
141 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
142 if (j
== 1 && !test_bit(MD_RECOVERY_SYNC
,
143 &conf
->mddev
->recovery
)) {
144 /* we can share bv_page's during recovery */
145 struct bio
*rbio
= r10_bio
->devs
[0].bio
;
146 page
= rbio
->bi_io_vec
[i
].bv_page
;
149 page
= alloc_page(gfp_flags
);
153 bio
->bi_io_vec
[i
].bv_page
= page
;
155 rbio
->bi_io_vec
[i
].bv_page
= page
;
163 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
165 for (i
= 0; i
< RESYNC_PAGES
; i
++)
166 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
169 while (++j
< nalloc
) {
170 bio_put(r10_bio
->devs
[j
].bio
);
171 if (r10_bio
->devs
[j
].repl_bio
)
172 bio_put(r10_bio
->devs
[j
].repl_bio
);
174 r10bio_pool_free(r10_bio
, conf
);
178 static void r10buf_pool_free(void *__r10_bio
, void *data
)
181 struct r10conf
*conf
= data
;
182 struct r10bio
*r10bio
= __r10_bio
;
185 for (j
=0; j
< conf
->copies
; j
++) {
186 struct bio
*bio
= r10bio
->devs
[j
].bio
;
188 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
189 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
190 bio
->bi_io_vec
[i
].bv_page
= NULL
;
194 bio
= r10bio
->devs
[j
].repl_bio
;
198 r10bio_pool_free(r10bio
, conf
);
201 static void put_all_bios(struct r10conf
*conf
, struct r10bio
*r10_bio
)
205 for (i
= 0; i
< conf
->copies
; i
++) {
206 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
207 if (!BIO_SPECIAL(*bio
))
210 bio
= &r10_bio
->devs
[i
].repl_bio
;
211 if (r10_bio
->read_slot
< 0 && !BIO_SPECIAL(*bio
))
217 static void free_r10bio(struct r10bio
*r10_bio
)
219 struct r10conf
*conf
= r10_bio
->mddev
->private;
221 put_all_bios(conf
, r10_bio
);
222 mempool_free(r10_bio
, conf
->r10bio_pool
);
225 static void put_buf(struct r10bio
*r10_bio
)
227 struct r10conf
*conf
= r10_bio
->mddev
->private;
229 mempool_free(r10_bio
, conf
->r10buf_pool
);
234 static void reschedule_retry(struct r10bio
*r10_bio
)
237 struct mddev
*mddev
= r10_bio
->mddev
;
238 struct r10conf
*conf
= mddev
->private;
240 spin_lock_irqsave(&conf
->device_lock
, flags
);
241 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
243 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
245 /* wake up frozen array... */
246 wake_up(&conf
->wait_barrier
);
248 md_wakeup_thread(mddev
->thread
);
252 * raid_end_bio_io() is called when we have finished servicing a mirrored
253 * operation and are ready to return a success/failure code to the buffer
256 static void raid_end_bio_io(struct r10bio
*r10_bio
)
258 struct bio
*bio
= r10_bio
->master_bio
;
260 struct r10conf
*conf
= r10_bio
->mddev
->private;
262 if (bio
->bi_phys_segments
) {
264 spin_lock_irqsave(&conf
->device_lock
, flags
);
265 bio
->bi_phys_segments
--;
266 done
= (bio
->bi_phys_segments
== 0);
267 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
270 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
271 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
275 * Wake up any possible resync thread that waits for the device
280 free_r10bio(r10_bio
);
284 * Update disk head position estimator based on IRQ completion info.
286 static inline void update_head_pos(int slot
, struct r10bio
*r10_bio
)
288 struct r10conf
*conf
= r10_bio
->mddev
->private;
290 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
291 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
295 * Find the disk number which triggered given bio
297 static int find_bio_disk(struct r10conf
*conf
, struct r10bio
*r10_bio
,
298 struct bio
*bio
, int *slotp
, int *replp
)
303 for (slot
= 0; slot
< conf
->copies
; slot
++) {
304 if (r10_bio
->devs
[slot
].bio
== bio
)
306 if (r10_bio
->devs
[slot
].repl_bio
== bio
) {
312 BUG_ON(slot
== conf
->copies
);
313 update_head_pos(slot
, r10_bio
);
319 return r10_bio
->devs
[slot
].devnum
;
322 static void raid10_end_read_request(struct bio
*bio
, int error
)
324 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
325 struct r10bio
*r10_bio
= bio
->bi_private
;
327 struct md_rdev
*rdev
;
328 struct r10conf
*conf
= r10_bio
->mddev
->private;
331 slot
= r10_bio
->read_slot
;
332 dev
= r10_bio
->devs
[slot
].devnum
;
333 rdev
= r10_bio
->devs
[slot
].rdev
;
335 * this branch is our 'one mirror IO has finished' event handler:
337 update_head_pos(slot
, r10_bio
);
341 * Set R10BIO_Uptodate in our master bio, so that
342 * we will return a good error code to the higher
343 * levels even if IO on some other mirrored buffer fails.
345 * The 'master' represents the composite IO operation to
346 * user-side. So if something waits for IO, then it will
347 * wait for the 'master' bio.
349 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
350 raid_end_bio_io(r10_bio
);
351 rdev_dec_pending(rdev
, conf
->mddev
);
354 * oops, read error - keep the refcount on the rdev
356 char b
[BDEVNAME_SIZE
];
357 printk_ratelimited(KERN_ERR
358 "md/raid10:%s: %s: rescheduling sector %llu\n",
360 bdevname(rdev
->bdev
, b
),
361 (unsigned long long)r10_bio
->sector
);
362 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
363 reschedule_retry(r10_bio
);
367 static void close_write(struct r10bio
*r10_bio
)
369 /* clear the bitmap if all writes complete successfully */
370 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
372 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
374 md_write_end(r10_bio
->mddev
);
377 static void one_write_done(struct r10bio
*r10_bio
)
379 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
380 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
381 reschedule_retry(r10_bio
);
383 close_write(r10_bio
);
384 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
385 reschedule_retry(r10_bio
);
387 raid_end_bio_io(r10_bio
);
392 static void raid10_end_write_request(struct bio
*bio
, int error
)
394 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
395 struct r10bio
*r10_bio
= bio
->bi_private
;
398 struct r10conf
*conf
= r10_bio
->mddev
->private;
400 struct md_rdev
*rdev
;
402 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
405 rdev
= conf
->mirrors
[dev
].replacement
;
407 rdev
= conf
->mirrors
[dev
].rdev
;
409 * this branch is our 'one mirror IO has finished' event handler:
413 /* Never record new bad blocks to replacement,
416 md_error(rdev
->mddev
, rdev
);
418 set_bit(WriteErrorSeen
, &rdev
->flags
);
419 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
424 * Set R10BIO_Uptodate in our master bio, so that
425 * we will return a good error code for to the higher
426 * levels even if IO on some other mirrored buffer fails.
428 * The 'master' represents the composite IO operation to
429 * user-side. So if something waits for IO, then it will
430 * wait for the 'master' bio.
435 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
437 /* Maybe we can clear some bad blocks. */
438 if (is_badblock(rdev
,
439 r10_bio
->devs
[slot
].addr
,
441 &first_bad
, &bad_sectors
)) {
444 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
446 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
448 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
454 * Let's see if all mirrored write operations have finished
457 one_write_done(r10_bio
);
459 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
463 * RAID10 layout manager
464 * As well as the chunksize and raid_disks count, there are two
465 * parameters: near_copies and far_copies.
466 * near_copies * far_copies must be <= raid_disks.
467 * Normally one of these will be 1.
468 * If both are 1, we get raid0.
469 * If near_copies == raid_disks, we get raid1.
471 * Chunks are laid out in raid0 style with near_copies copies of the
472 * first chunk, followed by near_copies copies of the next chunk and
474 * If far_copies > 1, then after 1/far_copies of the array has been assigned
475 * as described above, we start again with a device offset of near_copies.
476 * So we effectively have another copy of the whole array further down all
477 * the drives, but with blocks on different drives.
478 * With this layout, and block is never stored twice on the one device.
480 * raid10_find_phys finds the sector offset of a given virtual sector
481 * on each device that it is on.
483 * raid10_find_virt does the reverse mapping, from a device and a
484 * sector offset to a virtual address
487 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
497 /* now calculate first sector/dev */
498 chunk
= r10bio
->sector
>> conf
->chunk_shift
;
499 sector
= r10bio
->sector
& conf
->chunk_mask
;
501 chunk
*= conf
->near_copies
;
503 dev
= sector_div(stripe
, conf
->raid_disks
);
504 if (conf
->far_offset
)
505 stripe
*= conf
->far_copies
;
507 sector
+= stripe
<< conf
->chunk_shift
;
509 /* and calculate all the others */
510 for (n
=0; n
< conf
->near_copies
; n
++) {
513 r10bio
->devs
[slot
].addr
= sector
;
514 r10bio
->devs
[slot
].devnum
= d
;
517 for (f
= 1; f
< conf
->far_copies
; f
++) {
518 d
+= conf
->near_copies
;
519 if (d
>= conf
->raid_disks
)
520 d
-= conf
->raid_disks
;
522 r10bio
->devs
[slot
].devnum
= d
;
523 r10bio
->devs
[slot
].addr
= s
;
527 if (dev
>= conf
->raid_disks
) {
529 sector
+= (conf
->chunk_mask
+ 1);
532 BUG_ON(slot
!= conf
->copies
);
535 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
537 sector_t offset
, chunk
, vchunk
;
539 offset
= sector
& conf
->chunk_mask
;
540 if (conf
->far_offset
) {
542 chunk
= sector
>> conf
->chunk_shift
;
543 fc
= sector_div(chunk
, conf
->far_copies
);
544 dev
-= fc
* conf
->near_copies
;
546 dev
+= conf
->raid_disks
;
548 while (sector
>= conf
->stride
) {
549 sector
-= conf
->stride
;
550 if (dev
< conf
->near_copies
)
551 dev
+= conf
->raid_disks
- conf
->near_copies
;
553 dev
-= conf
->near_copies
;
555 chunk
= sector
>> conf
->chunk_shift
;
557 vchunk
= chunk
* conf
->raid_disks
+ dev
;
558 sector_div(vchunk
, conf
->near_copies
);
559 return (vchunk
<< conf
->chunk_shift
) + offset
;
563 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
565 * @bvm: properties of new bio
566 * @biovec: the request that could be merged to it.
568 * Return amount of bytes we can accept at this offset
569 * If near_copies == raid_disk, there are no striping issues,
570 * but in that case, the function isn't called at all.
572 static int raid10_mergeable_bvec(struct request_queue
*q
,
573 struct bvec_merge_data
*bvm
,
574 struct bio_vec
*biovec
)
576 struct mddev
*mddev
= q
->queuedata
;
577 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
579 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
580 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
582 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
583 if (max
< 0) max
= 0; /* bio_add cannot handle a negative return */
584 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
585 return biovec
->bv_len
;
591 * This routine returns the disk from which the requested read should
592 * be done. There is a per-array 'next expected sequential IO' sector
593 * number - if this matches on the next IO then we use the last disk.
594 * There is also a per-disk 'last know head position' sector that is
595 * maintained from IRQ contexts, both the normal and the resync IO
596 * completion handlers update this position correctly. If there is no
597 * perfect sequential match then we pick the disk whose head is closest.
599 * If there are 2 mirrors in the same 2 devices, performance degrades
600 * because position is mirror, not device based.
602 * The rdev for the device selected will have nr_pending incremented.
606 * FIXME: possibly should rethink readbalancing and do it differently
607 * depending on near_copies / far_copies geometry.
609 static struct md_rdev
*read_balance(struct r10conf
*conf
,
610 struct r10bio
*r10_bio
,
613 const sector_t this_sector
= r10_bio
->sector
;
615 int sectors
= r10_bio
->sectors
;
616 int best_good_sectors
;
617 sector_t new_distance
, best_dist
;
618 struct md_rdev
*rdev
, *best_rdev
;
622 raid10_find_phys(conf
, r10_bio
);
625 sectors
= r10_bio
->sectors
;
628 best_dist
= MaxSector
;
629 best_good_sectors
= 0;
632 * Check if we can balance. We can balance on the whole
633 * device if no resync is going on (recovery is ok), or below
634 * the resync window. We take the first readable disk when
635 * above the resync window.
637 if (conf
->mddev
->recovery_cp
< MaxSector
638 && (this_sector
+ sectors
>= conf
->next_resync
))
641 for (slot
= 0; slot
< conf
->copies
; slot
++) {
646 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
648 disk
= r10_bio
->devs
[slot
].devnum
;
649 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
650 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
651 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
652 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
655 if (test_bit(Faulty
, &rdev
->flags
))
657 if (!test_bit(In_sync
, &rdev
->flags
) &&
658 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
661 dev_sector
= r10_bio
->devs
[slot
].addr
;
662 if (is_badblock(rdev
, dev_sector
, sectors
,
663 &first_bad
, &bad_sectors
)) {
664 if (best_dist
< MaxSector
)
665 /* Already have a better slot */
667 if (first_bad
<= dev_sector
) {
668 /* Cannot read here. If this is the
669 * 'primary' device, then we must not read
670 * beyond 'bad_sectors' from another device.
672 bad_sectors
-= (dev_sector
- first_bad
);
673 if (!do_balance
&& sectors
> bad_sectors
)
674 sectors
= bad_sectors
;
675 if (best_good_sectors
> sectors
)
676 best_good_sectors
= sectors
;
678 sector_t good_sectors
=
679 first_bad
- dev_sector
;
680 if (good_sectors
> best_good_sectors
) {
681 best_good_sectors
= good_sectors
;
686 /* Must read from here */
691 best_good_sectors
= sectors
;
696 /* This optimisation is debatable, and completely destroys
697 * sequential read speed for 'far copies' arrays. So only
698 * keep it for 'near' arrays, and review those later.
700 if (conf
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
703 /* for far > 1 always use the lowest address */
704 if (conf
->far_copies
> 1)
705 new_distance
= r10_bio
->devs
[slot
].addr
;
707 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
708 conf
->mirrors
[disk
].head_position
);
709 if (new_distance
< best_dist
) {
710 best_dist
= new_distance
;
715 if (slot
>= conf
->copies
) {
721 atomic_inc(&rdev
->nr_pending
);
722 if (test_bit(Faulty
, &rdev
->flags
)) {
723 /* Cannot risk returning a device that failed
724 * before we inc'ed nr_pending
726 rdev_dec_pending(rdev
, conf
->mddev
);
729 r10_bio
->read_slot
= slot
;
733 *max_sectors
= best_good_sectors
;
738 static int raid10_congested(void *data
, int bits
)
740 struct mddev
*mddev
= data
;
741 struct r10conf
*conf
= mddev
->private;
744 if ((bits
& (1 << BDI_async_congested
)) &&
745 conf
->pending_count
>= max_queued_requests
)
748 if (mddev_congested(mddev
, bits
))
751 for (i
= 0; i
< conf
->raid_disks
&& ret
== 0; i
++) {
752 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
753 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
754 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
756 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
763 static void flush_pending_writes(struct r10conf
*conf
)
765 /* Any writes that have been queued but are awaiting
766 * bitmap updates get flushed here.
768 spin_lock_irq(&conf
->device_lock
);
770 if (conf
->pending_bio_list
.head
) {
772 bio
= bio_list_get(&conf
->pending_bio_list
);
773 conf
->pending_count
= 0;
774 spin_unlock_irq(&conf
->device_lock
);
775 /* flush any pending bitmap writes to disk
776 * before proceeding w/ I/O */
777 bitmap_unplug(conf
->mddev
->bitmap
);
778 wake_up(&conf
->wait_barrier
);
780 while (bio
) { /* submit pending writes */
781 struct bio
*next
= bio
->bi_next
;
783 generic_make_request(bio
);
787 spin_unlock_irq(&conf
->device_lock
);
791 * Sometimes we need to suspend IO while we do something else,
792 * either some resync/recovery, or reconfigure the array.
793 * To do this we raise a 'barrier'.
794 * The 'barrier' is a counter that can be raised multiple times
795 * to count how many activities are happening which preclude
797 * We can only raise the barrier if there is no pending IO.
798 * i.e. if nr_pending == 0.
799 * We choose only to raise the barrier if no-one is waiting for the
800 * barrier to go down. This means that as soon as an IO request
801 * is ready, no other operations which require a barrier will start
802 * until the IO request has had a chance.
804 * So: regular IO calls 'wait_barrier'. When that returns there
805 * is no backgroup IO happening, It must arrange to call
806 * allow_barrier when it has finished its IO.
807 * backgroup IO calls must call raise_barrier. Once that returns
808 * there is no normal IO happeing. It must arrange to call
809 * lower_barrier when the particular background IO completes.
812 static void raise_barrier(struct r10conf
*conf
, int force
)
814 BUG_ON(force
&& !conf
->barrier
);
815 spin_lock_irq(&conf
->resync_lock
);
817 /* Wait until no block IO is waiting (unless 'force') */
818 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
819 conf
->resync_lock
, );
821 /* block any new IO from starting */
824 /* Now wait for all pending IO to complete */
825 wait_event_lock_irq(conf
->wait_barrier
,
826 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
827 conf
->resync_lock
, );
829 spin_unlock_irq(&conf
->resync_lock
);
832 static void lower_barrier(struct r10conf
*conf
)
835 spin_lock_irqsave(&conf
->resync_lock
, flags
);
837 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
838 wake_up(&conf
->wait_barrier
);
841 static void wait_barrier(struct r10conf
*conf
)
843 spin_lock_irq(&conf
->resync_lock
);
846 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
852 spin_unlock_irq(&conf
->resync_lock
);
855 static void allow_barrier(struct r10conf
*conf
)
858 spin_lock_irqsave(&conf
->resync_lock
, flags
);
860 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
861 wake_up(&conf
->wait_barrier
);
864 static void freeze_array(struct r10conf
*conf
)
866 /* stop syncio and normal IO and wait for everything to
868 * We increment barrier and nr_waiting, and then
869 * wait until nr_pending match nr_queued+1
870 * This is called in the context of one normal IO request
871 * that has failed. Thus any sync request that might be pending
872 * will be blocked by nr_pending, and we need to wait for
873 * pending IO requests to complete or be queued for re-try.
874 * Thus the number queued (nr_queued) plus this request (1)
875 * must match the number of pending IOs (nr_pending) before
878 spin_lock_irq(&conf
->resync_lock
);
881 wait_event_lock_irq(conf
->wait_barrier
,
882 conf
->nr_pending
== conf
->nr_queued
+1,
884 flush_pending_writes(conf
));
886 spin_unlock_irq(&conf
->resync_lock
);
889 static void unfreeze_array(struct r10conf
*conf
)
891 /* reverse the effect of the freeze */
892 spin_lock_irq(&conf
->resync_lock
);
895 wake_up(&conf
->wait_barrier
);
896 spin_unlock_irq(&conf
->resync_lock
);
899 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
901 struct r10conf
*conf
= mddev
->private;
902 struct r10bio
*r10_bio
;
903 struct bio
*read_bio
;
905 int chunk_sects
= conf
->chunk_mask
+ 1;
906 const int rw
= bio_data_dir(bio
);
907 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
908 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
910 struct md_rdev
*blocked_rdev
;
915 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
916 md_flush_request(mddev
, bio
);
920 /* If this request crosses a chunk boundary, we need to
921 * split it. This will only happen for 1 PAGE (or less) requests.
923 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
925 conf
->near_copies
< conf
->raid_disks
)) {
927 /* Sanity check -- queue functions should prevent this happening */
928 if (bio
->bi_vcnt
!= 1 ||
931 /* This is a one page bio that upper layers
932 * refuse to split for us, so we need to split it.
935 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
937 /* Each of these 'make_request' calls will call 'wait_barrier'.
938 * If the first succeeds but the second blocks due to the resync
939 * thread raising the barrier, we will deadlock because the
940 * IO to the underlying device will be queued in generic_make_request
941 * and will never complete, so will never reduce nr_pending.
942 * So increment nr_waiting here so no new raise_barriers will
943 * succeed, and so the second wait_barrier cannot block.
945 spin_lock_irq(&conf
->resync_lock
);
947 spin_unlock_irq(&conf
->resync_lock
);
949 make_request(mddev
, &bp
->bio1
);
950 make_request(mddev
, &bp
->bio2
);
952 spin_lock_irq(&conf
->resync_lock
);
954 wake_up(&conf
->wait_barrier
);
955 spin_unlock_irq(&conf
->resync_lock
);
957 bio_pair_release(bp
);
960 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
961 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
962 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
968 md_write_start(mddev
, bio
);
971 * Register the new request and wait if the reconstruction
972 * thread has put up a bar for new requests.
973 * Continue immediately if no resync is active currently.
977 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
979 r10_bio
->master_bio
= bio
;
980 r10_bio
->sectors
= bio
->bi_size
>> 9;
982 r10_bio
->mddev
= mddev
;
983 r10_bio
->sector
= bio
->bi_sector
;
986 /* We might need to issue multiple reads to different
987 * devices if there are bad blocks around, so we keep
988 * track of the number of reads in bio->bi_phys_segments.
989 * If this is 0, there is only one r10_bio and no locking
990 * will be needed when the request completes. If it is
991 * non-zero, then it is the number of not-completed requests.
993 bio
->bi_phys_segments
= 0;
994 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
998 * read balancing logic:
1000 struct md_rdev
*rdev
;
1004 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1006 raid_end_bio_io(r10_bio
);
1009 slot
= r10_bio
->read_slot
;
1011 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1012 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1015 r10_bio
->devs
[slot
].bio
= read_bio
;
1016 r10_bio
->devs
[slot
].rdev
= rdev
;
1018 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1020 read_bio
->bi_bdev
= rdev
->bdev
;
1021 read_bio
->bi_end_io
= raid10_end_read_request
;
1022 read_bio
->bi_rw
= READ
| do_sync
;
1023 read_bio
->bi_private
= r10_bio
;
1025 if (max_sectors
< r10_bio
->sectors
) {
1026 /* Could not read all from this device, so we will
1027 * need another r10_bio.
1029 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1031 r10_bio
->sectors
= max_sectors
;
1032 spin_lock_irq(&conf
->device_lock
);
1033 if (bio
->bi_phys_segments
== 0)
1034 bio
->bi_phys_segments
= 2;
1036 bio
->bi_phys_segments
++;
1037 spin_unlock(&conf
->device_lock
);
1038 /* Cannot call generic_make_request directly
1039 * as that will be queued in __generic_make_request
1040 * and subsequent mempool_alloc might block
1041 * waiting for it. so hand bio over to raid10d.
1043 reschedule_retry(r10_bio
);
1045 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1047 r10_bio
->master_bio
= bio
;
1048 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
1051 r10_bio
->mddev
= mddev
;
1052 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1055 generic_make_request(read_bio
);
1062 if (conf
->pending_count
>= max_queued_requests
) {
1063 md_wakeup_thread(mddev
->thread
);
1064 wait_event(conf
->wait_barrier
,
1065 conf
->pending_count
< max_queued_requests
);
1067 /* first select target devices under rcu_lock and
1068 * inc refcount on their rdev. Record them by setting
1070 * If there are known/acknowledged bad blocks on any device
1071 * on which we have seen a write error, we want to avoid
1072 * writing to those blocks. This potentially requires several
1073 * writes to write around the bad blocks. Each set of writes
1074 * gets its own r10_bio with a set of bios attached. The number
1075 * of r10_bios is recored in bio->bi_phys_segments just as with
1078 plugged
= mddev_check_plugged(mddev
);
1080 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1081 raid10_find_phys(conf
, r10_bio
);
1083 blocked_rdev
= NULL
;
1085 max_sectors
= r10_bio
->sectors
;
1087 for (i
= 0; i
< conf
->copies
; i
++) {
1088 int d
= r10_bio
->devs
[i
].devnum
;
1089 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1090 struct md_rdev
*rrdev
= rcu_dereference(
1091 conf
->mirrors
[d
].replacement
);
1092 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1093 atomic_inc(&rdev
->nr_pending
);
1094 blocked_rdev
= rdev
;
1097 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1098 atomic_inc(&rrdev
->nr_pending
);
1099 blocked_rdev
= rrdev
;
1102 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
1105 r10_bio
->devs
[i
].bio
= NULL
;
1106 r10_bio
->devs
[i
].repl_bio
= NULL
;
1107 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
1108 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1111 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1113 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1117 is_bad
= is_badblock(rdev
, dev_sector
,
1119 &first_bad
, &bad_sectors
);
1121 /* Mustn't write here until the bad block
1124 atomic_inc(&rdev
->nr_pending
);
1125 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1126 blocked_rdev
= rdev
;
1129 if (is_bad
&& first_bad
<= dev_sector
) {
1130 /* Cannot write here at all */
1131 bad_sectors
-= (dev_sector
- first_bad
);
1132 if (bad_sectors
< max_sectors
)
1133 /* Mustn't write more than bad_sectors
1134 * to other devices yet
1136 max_sectors
= bad_sectors
;
1137 /* We don't set R10BIO_Degraded as that
1138 * only applies if the disk is missing,
1139 * so it might be re-added, and we want to
1140 * know to recover this chunk.
1141 * In this case the device is here, and the
1142 * fact that this chunk is not in-sync is
1143 * recorded in the bad block log.
1148 int good_sectors
= first_bad
- dev_sector
;
1149 if (good_sectors
< max_sectors
)
1150 max_sectors
= good_sectors
;
1153 r10_bio
->devs
[i
].bio
= bio
;
1154 atomic_inc(&rdev
->nr_pending
);
1156 r10_bio
->devs
[i
].repl_bio
= bio
;
1157 atomic_inc(&rrdev
->nr_pending
);
1162 if (unlikely(blocked_rdev
)) {
1163 /* Have to wait for this device to get unblocked, then retry */
1167 for (j
= 0; j
< i
; j
++) {
1168 if (r10_bio
->devs
[j
].bio
) {
1169 d
= r10_bio
->devs
[j
].devnum
;
1170 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1172 if (r10_bio
->devs
[j
].repl_bio
) {
1173 d
= r10_bio
->devs
[j
].devnum
;
1175 conf
->mirrors
[d
].replacement
, mddev
);
1178 allow_barrier(conf
);
1179 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1184 if (max_sectors
< r10_bio
->sectors
) {
1185 /* We are splitting this into multiple parts, so
1186 * we need to prepare for allocating another r10_bio.
1188 r10_bio
->sectors
= max_sectors
;
1189 spin_lock_irq(&conf
->device_lock
);
1190 if (bio
->bi_phys_segments
== 0)
1191 bio
->bi_phys_segments
= 2;
1193 bio
->bi_phys_segments
++;
1194 spin_unlock_irq(&conf
->device_lock
);
1196 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1198 atomic_set(&r10_bio
->remaining
, 1);
1199 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1201 for (i
= 0; i
< conf
->copies
; i
++) {
1203 int d
= r10_bio
->devs
[i
].devnum
;
1204 if (!r10_bio
->devs
[i
].bio
)
1207 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1208 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1210 r10_bio
->devs
[i
].bio
= mbio
;
1212 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1213 conf
->mirrors
[d
].rdev
->data_offset
);
1214 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1215 mbio
->bi_end_io
= raid10_end_write_request
;
1216 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
1217 mbio
->bi_private
= r10_bio
;
1219 atomic_inc(&r10_bio
->remaining
);
1220 spin_lock_irqsave(&conf
->device_lock
, flags
);
1221 bio_list_add(&conf
->pending_bio_list
, mbio
);
1222 conf
->pending_count
++;
1223 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1225 if (!r10_bio
->devs
[i
].repl_bio
)
1228 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1229 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1231 r10_bio
->devs
[i
].repl_bio
= mbio
;
1233 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1234 conf
->mirrors
[d
].replacement
->data_offset
);
1235 mbio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
1236 mbio
->bi_end_io
= raid10_end_write_request
;
1237 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
1238 mbio
->bi_private
= r10_bio
;
1240 atomic_inc(&r10_bio
->remaining
);
1241 spin_lock_irqsave(&conf
->device_lock
, flags
);
1242 bio_list_add(&conf
->pending_bio_list
, mbio
);
1243 conf
->pending_count
++;
1244 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1247 /* Don't remove the bias on 'remaining' (one_write_done) until
1248 * after checking if we need to go around again.
1251 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1252 one_write_done(r10_bio
);
1253 /* We need another r10_bio. It has already been counted
1254 * in bio->bi_phys_segments.
1256 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1258 r10_bio
->master_bio
= bio
;
1259 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1261 r10_bio
->mddev
= mddev
;
1262 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1266 one_write_done(r10_bio
);
1268 /* In case raid10d snuck in to freeze_array */
1269 wake_up(&conf
->wait_barrier
);
1271 if (do_sync
|| !mddev
->bitmap
|| !plugged
)
1272 md_wakeup_thread(mddev
->thread
);
1275 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1277 struct r10conf
*conf
= mddev
->private;
1280 if (conf
->near_copies
< conf
->raid_disks
)
1281 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1282 if (conf
->near_copies
> 1)
1283 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
1284 if (conf
->far_copies
> 1) {
1285 if (conf
->far_offset
)
1286 seq_printf(seq
, " %d offset-copies", conf
->far_copies
);
1288 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
1290 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1291 conf
->raid_disks
- mddev
->degraded
);
1292 for (i
= 0; i
< conf
->raid_disks
; i
++)
1293 seq_printf(seq
, "%s",
1294 conf
->mirrors
[i
].rdev
&&
1295 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1296 seq_printf(seq
, "]");
1299 /* check if there are enough drives for
1300 * every block to appear on atleast one.
1301 * Don't consider the device numbered 'ignore'
1302 * as we might be about to remove it.
1304 static int enough(struct r10conf
*conf
, int ignore
)
1309 int n
= conf
->copies
;
1312 if (conf
->mirrors
[first
].rdev
&&
1315 first
= (first
+1) % conf
->raid_disks
;
1319 } while (first
!= 0);
1323 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1325 char b
[BDEVNAME_SIZE
];
1326 struct r10conf
*conf
= mddev
->private;
1329 * If it is not operational, then we have already marked it as dead
1330 * else if it is the last working disks, ignore the error, let the
1331 * next level up know.
1332 * else mark the drive as failed
1334 if (test_bit(In_sync
, &rdev
->flags
)
1335 && !enough(conf
, rdev
->raid_disk
))
1337 * Don't fail the drive, just return an IO error.
1340 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1341 unsigned long flags
;
1342 spin_lock_irqsave(&conf
->device_lock
, flags
);
1344 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1346 * if recovery is running, make sure it aborts.
1348 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1350 set_bit(Blocked
, &rdev
->flags
);
1351 set_bit(Faulty
, &rdev
->flags
);
1352 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1354 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1355 "md/raid10:%s: Operation continuing on %d devices.\n",
1356 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1357 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1360 static void print_conf(struct r10conf
*conf
)
1363 struct mirror_info
*tmp
;
1365 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1367 printk(KERN_DEBUG
"(!conf)\n");
1370 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1373 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1374 char b
[BDEVNAME_SIZE
];
1375 tmp
= conf
->mirrors
+ i
;
1377 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1378 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1379 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1380 bdevname(tmp
->rdev
->bdev
,b
));
1384 static void close_sync(struct r10conf
*conf
)
1387 allow_barrier(conf
);
1389 mempool_destroy(conf
->r10buf_pool
);
1390 conf
->r10buf_pool
= NULL
;
1393 static int raid10_spare_active(struct mddev
*mddev
)
1396 struct r10conf
*conf
= mddev
->private;
1397 struct mirror_info
*tmp
;
1399 unsigned long flags
;
1402 * Find all non-in_sync disks within the RAID10 configuration
1403 * and mark them in_sync
1405 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1406 tmp
= conf
->mirrors
+ i
;
1408 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1409 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1411 sysfs_notify_dirent(tmp
->rdev
->sysfs_state
);
1414 spin_lock_irqsave(&conf
->device_lock
, flags
);
1415 mddev
->degraded
-= count
;
1416 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1423 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1425 struct r10conf
*conf
= mddev
->private;
1429 int last
= conf
->raid_disks
- 1;
1431 if (mddev
->recovery_cp
< MaxSector
)
1432 /* only hot-add to in-sync arrays, as recovery is
1433 * very different from resync
1436 if (!enough(conf
, -1))
1439 if (rdev
->raid_disk
>= 0)
1440 first
= last
= rdev
->raid_disk
;
1442 if (rdev
->saved_raid_disk
>= first
&&
1443 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1444 mirror
= rdev
->saved_raid_disk
;
1447 for ( ; mirror
<= last
; mirror
++) {
1448 struct mirror_info
*p
= &conf
->mirrors
[mirror
];
1449 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1454 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1455 rdev
->data_offset
<< 9);
1456 /* as we don't honour merge_bvec_fn, we must
1457 * never risk violating it, so limit
1458 * ->max_segments to one lying with a single
1459 * page, as a one page request is never in
1462 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1463 blk_queue_max_segments(mddev
->queue
, 1);
1464 blk_queue_segment_boundary(mddev
->queue
,
1465 PAGE_CACHE_SIZE
- 1);
1468 p
->head_position
= 0;
1469 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1470 rdev
->raid_disk
= mirror
;
1472 if (rdev
->saved_raid_disk
!= mirror
)
1474 rcu_assign_pointer(p
->rdev
, rdev
);
1478 md_integrity_add_rdev(rdev
, mddev
);
1483 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1485 struct r10conf
*conf
= mddev
->private;
1487 int number
= rdev
->raid_disk
;
1488 struct md_rdev
**rdevp
;
1489 struct mirror_info
*p
= conf
->mirrors
+ number
;
1492 if (rdev
== p
->rdev
)
1494 else if (rdev
== p
->replacement
)
1495 rdevp
= &p
->replacement
;
1499 if (test_bit(In_sync
, &rdev
->flags
) ||
1500 atomic_read(&rdev
->nr_pending
)) {
1504 /* Only remove faulty devices if recovery
1507 if (!test_bit(Faulty
, &rdev
->flags
) &&
1508 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1515 if (atomic_read(&rdev
->nr_pending
)) {
1516 /* lost the race, try later */
1521 err
= md_integrity_register(mddev
);
1530 static void end_sync_read(struct bio
*bio
, int error
)
1532 struct r10bio
*r10_bio
= bio
->bi_private
;
1533 struct r10conf
*conf
= r10_bio
->mddev
->private;
1536 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1538 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1539 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1541 /* The write handler will notice the lack of
1542 * R10BIO_Uptodate and record any errors etc
1544 atomic_add(r10_bio
->sectors
,
1545 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1547 /* for reconstruct, we always reschedule after a read.
1548 * for resync, only after all reads
1550 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1551 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1552 atomic_dec_and_test(&r10_bio
->remaining
)) {
1553 /* we have read all the blocks,
1554 * do the comparison in process context in raid10d
1556 reschedule_retry(r10_bio
);
1560 static void end_sync_request(struct r10bio
*r10_bio
)
1562 struct mddev
*mddev
= r10_bio
->mddev
;
1564 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1565 if (r10_bio
->master_bio
== NULL
) {
1566 /* the primary of several recovery bios */
1567 sector_t s
= r10_bio
->sectors
;
1568 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1569 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1570 reschedule_retry(r10_bio
);
1573 md_done_sync(mddev
, s
, 1);
1576 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1577 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1578 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1579 reschedule_retry(r10_bio
);
1587 static void end_sync_write(struct bio
*bio
, int error
)
1589 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1590 struct r10bio
*r10_bio
= bio
->bi_private
;
1591 struct mddev
*mddev
= r10_bio
->mddev
;
1592 struct r10conf
*conf
= mddev
->private;
1598 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, NULL
);
1601 set_bit(WriteErrorSeen
, &conf
->mirrors
[d
].rdev
->flags
);
1602 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1603 } else if (is_badblock(conf
->mirrors
[d
].rdev
,
1604 r10_bio
->devs
[slot
].addr
,
1606 &first_bad
, &bad_sectors
))
1607 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1609 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1611 end_sync_request(r10_bio
);
1615 * Note: sync and recover and handled very differently for raid10
1616 * This code is for resync.
1617 * For resync, we read through virtual addresses and read all blocks.
1618 * If there is any error, we schedule a write. The lowest numbered
1619 * drive is authoritative.
1620 * However requests come for physical address, so we need to map.
1621 * For every physical address there are raid_disks/copies virtual addresses,
1622 * which is always are least one, but is not necessarly an integer.
1623 * This means that a physical address can span multiple chunks, so we may
1624 * have to submit multiple io requests for a single sync request.
1627 * We check if all blocks are in-sync and only write to blocks that
1630 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1632 struct r10conf
*conf
= mddev
->private;
1634 struct bio
*tbio
, *fbio
;
1636 atomic_set(&r10_bio
->remaining
, 1);
1638 /* find the first device with a block */
1639 for (i
=0; i
<conf
->copies
; i
++)
1640 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1643 if (i
== conf
->copies
)
1647 fbio
= r10_bio
->devs
[i
].bio
;
1649 /* now find blocks with errors */
1650 for (i
=0 ; i
< conf
->copies
; i
++) {
1652 int vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1654 tbio
= r10_bio
->devs
[i
].bio
;
1656 if (tbio
->bi_end_io
!= end_sync_read
)
1660 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1661 /* We know that the bi_io_vec layout is the same for
1662 * both 'first' and 'i', so we just compare them.
1663 * All vec entries are PAGE_SIZE;
1665 for (j
= 0; j
< vcnt
; j
++)
1666 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1667 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1672 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1673 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1674 /* Don't fix anything. */
1677 /* Ok, we need to write this bio, either to correct an
1678 * inconsistency or to correct an unreadable block.
1679 * First we need to fixup bv_offset, bv_len and
1680 * bi_vecs, as the read request might have corrupted these
1682 tbio
->bi_vcnt
= vcnt
;
1683 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1685 tbio
->bi_phys_segments
= 0;
1686 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1687 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1688 tbio
->bi_next
= NULL
;
1689 tbio
->bi_rw
= WRITE
;
1690 tbio
->bi_private
= r10_bio
;
1691 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1693 for (j
=0; j
< vcnt
; j
++) {
1694 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1695 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1697 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1698 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1701 tbio
->bi_end_io
= end_sync_write
;
1703 d
= r10_bio
->devs
[i
].devnum
;
1704 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1705 atomic_inc(&r10_bio
->remaining
);
1706 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1708 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1709 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1710 generic_make_request(tbio
);
1714 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1715 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1721 * Now for the recovery code.
1722 * Recovery happens across physical sectors.
1723 * We recover all non-is_sync drives by finding the virtual address of
1724 * each, and then choose a working drive that also has that virt address.
1725 * There is a separate r10_bio for each non-in_sync drive.
1726 * Only the first two slots are in use. The first for reading,
1727 * The second for writing.
1730 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
1732 /* We got a read error during recovery.
1733 * We repeat the read in smaller page-sized sections.
1734 * If a read succeeds, write it to the new device or record
1735 * a bad block if we cannot.
1736 * If a read fails, record a bad block on both old and
1739 struct mddev
*mddev
= r10_bio
->mddev
;
1740 struct r10conf
*conf
= mddev
->private;
1741 struct bio
*bio
= r10_bio
->devs
[0].bio
;
1743 int sectors
= r10_bio
->sectors
;
1745 int dr
= r10_bio
->devs
[0].devnum
;
1746 int dw
= r10_bio
->devs
[1].devnum
;
1750 struct md_rdev
*rdev
;
1754 if (s
> (PAGE_SIZE
>>9))
1757 rdev
= conf
->mirrors
[dr
].rdev
;
1758 addr
= r10_bio
->devs
[0].addr
+ sect
,
1759 ok
= sync_page_io(rdev
,
1762 bio
->bi_io_vec
[idx
].bv_page
,
1765 rdev
= conf
->mirrors
[dw
].rdev
;
1766 addr
= r10_bio
->devs
[1].addr
+ sect
;
1767 ok
= sync_page_io(rdev
,
1770 bio
->bi_io_vec
[idx
].bv_page
,
1773 set_bit(WriteErrorSeen
, &rdev
->flags
);
1776 /* We don't worry if we cannot set a bad block -
1777 * it really is bad so there is no loss in not
1780 rdev_set_badblocks(rdev
, addr
, s
, 0);
1782 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
1783 /* need bad block on destination too */
1784 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
1785 addr
= r10_bio
->devs
[1].addr
+ sect
;
1786 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
1788 /* just abort the recovery */
1790 "md/raid10:%s: recovery aborted"
1791 " due to read error\n",
1794 conf
->mirrors
[dw
].recovery_disabled
1795 = mddev
->recovery_disabled
;
1796 set_bit(MD_RECOVERY_INTR
,
1809 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1811 struct r10conf
*conf
= mddev
->private;
1815 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
1816 fix_recovery_read_error(r10_bio
);
1817 end_sync_request(r10_bio
);
1822 * share the pages with the first bio
1823 * and submit the write request
1825 wbio
= r10_bio
->devs
[1].bio
;
1826 d
= r10_bio
->devs
[1].devnum
;
1828 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1829 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1830 generic_make_request(wbio
);
1835 * Used by fix_read_error() to decay the per rdev read_errors.
1836 * We halve the read error count for every hour that has elapsed
1837 * since the last recorded read error.
1840 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
1842 struct timespec cur_time_mon
;
1843 unsigned long hours_since_last
;
1844 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
1846 ktime_get_ts(&cur_time_mon
);
1848 if (rdev
->last_read_error
.tv_sec
== 0 &&
1849 rdev
->last_read_error
.tv_nsec
== 0) {
1850 /* first time we've seen a read error */
1851 rdev
->last_read_error
= cur_time_mon
;
1855 hours_since_last
= (cur_time_mon
.tv_sec
-
1856 rdev
->last_read_error
.tv_sec
) / 3600;
1858 rdev
->last_read_error
= cur_time_mon
;
1861 * if hours_since_last is > the number of bits in read_errors
1862 * just set read errors to 0. We do this to avoid
1863 * overflowing the shift of read_errors by hours_since_last.
1865 if (hours_since_last
>= 8 * sizeof(read_errors
))
1866 atomic_set(&rdev
->read_errors
, 0);
1868 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
1871 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1872 int sectors
, struct page
*page
, int rw
)
1877 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
1878 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
1880 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
1884 set_bit(WriteErrorSeen
, &rdev
->flags
);
1885 /* need to record an error - either for the block or the device */
1886 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1887 md_error(rdev
->mddev
, rdev
);
1892 * This is a kernel thread which:
1894 * 1. Retries failed read operations on working mirrors.
1895 * 2. Updates the raid superblock when problems encounter.
1896 * 3. Performs writes following reads for array synchronising.
1899 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
1901 int sect
= 0; /* Offset from r10_bio->sector */
1902 int sectors
= r10_bio
->sectors
;
1903 struct md_rdev
*rdev
;
1904 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
1905 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1907 /* still own a reference to this rdev, so it cannot
1908 * have been cleared recently.
1910 rdev
= conf
->mirrors
[d
].rdev
;
1912 if (test_bit(Faulty
, &rdev
->flags
))
1913 /* drive has already been failed, just ignore any
1914 more fix_read_error() attempts */
1917 check_decay_read_errors(mddev
, rdev
);
1918 atomic_inc(&rdev
->read_errors
);
1919 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
1920 char b
[BDEVNAME_SIZE
];
1921 bdevname(rdev
->bdev
, b
);
1924 "md/raid10:%s: %s: Raid device exceeded "
1925 "read_error threshold [cur %d:max %d]\n",
1927 atomic_read(&rdev
->read_errors
), max_read_errors
);
1929 "md/raid10:%s: %s: Failing raid device\n",
1931 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1937 int sl
= r10_bio
->read_slot
;
1941 if (s
> (PAGE_SIZE
>>9))
1949 d
= r10_bio
->devs
[sl
].devnum
;
1950 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1952 test_bit(In_sync
, &rdev
->flags
) &&
1953 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
1954 &first_bad
, &bad_sectors
) == 0) {
1955 atomic_inc(&rdev
->nr_pending
);
1957 success
= sync_page_io(rdev
,
1958 r10_bio
->devs
[sl
].addr
+
1961 conf
->tmppage
, READ
, false);
1962 rdev_dec_pending(rdev
, mddev
);
1968 if (sl
== conf
->copies
)
1970 } while (!success
&& sl
!= r10_bio
->read_slot
);
1974 /* Cannot read from anywhere, just mark the block
1975 * as bad on the first device to discourage future
1978 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1979 rdev
= conf
->mirrors
[dn
].rdev
;
1981 if (!rdev_set_badblocks(
1983 r10_bio
->devs
[r10_bio
->read_slot
].addr
1986 md_error(mddev
, rdev
);
1991 /* write it back and re-read */
1993 while (sl
!= r10_bio
->read_slot
) {
1994 char b
[BDEVNAME_SIZE
];
1999 d
= r10_bio
->devs
[sl
].devnum
;
2000 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2002 !test_bit(In_sync
, &rdev
->flags
))
2005 atomic_inc(&rdev
->nr_pending
);
2007 if (r10_sync_page_io(rdev
,
2008 r10_bio
->devs
[sl
].addr
+
2010 s
<<9, conf
->tmppage
, WRITE
)
2012 /* Well, this device is dead */
2014 "md/raid10:%s: read correction "
2016 " (%d sectors at %llu on %s)\n",
2018 (unsigned long long)(
2019 sect
+ rdev
->data_offset
),
2020 bdevname(rdev
->bdev
, b
));
2021 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2024 bdevname(rdev
->bdev
, b
));
2026 rdev_dec_pending(rdev
, mddev
);
2030 while (sl
!= r10_bio
->read_slot
) {
2031 char b
[BDEVNAME_SIZE
];
2036 d
= r10_bio
->devs
[sl
].devnum
;
2037 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2039 !test_bit(In_sync
, &rdev
->flags
))
2042 atomic_inc(&rdev
->nr_pending
);
2044 switch (r10_sync_page_io(rdev
,
2045 r10_bio
->devs
[sl
].addr
+
2047 s
<<9, conf
->tmppage
,
2050 /* Well, this device is dead */
2052 "md/raid10:%s: unable to read back "
2054 " (%d sectors at %llu on %s)\n",
2056 (unsigned long long)(
2057 sect
+ rdev
->data_offset
),
2058 bdevname(rdev
->bdev
, b
));
2059 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2062 bdevname(rdev
->bdev
, b
));
2066 "md/raid10:%s: read error corrected"
2067 " (%d sectors at %llu on %s)\n",
2069 (unsigned long long)(
2070 sect
+ rdev
->data_offset
),
2071 bdevname(rdev
->bdev
, b
));
2072 atomic_add(s
, &rdev
->corrected_errors
);
2075 rdev_dec_pending(rdev
, mddev
);
2085 static void bi_complete(struct bio
*bio
, int error
)
2087 complete((struct completion
*)bio
->bi_private
);
2090 static int submit_bio_wait(int rw
, struct bio
*bio
)
2092 struct completion event
;
2095 init_completion(&event
);
2096 bio
->bi_private
= &event
;
2097 bio
->bi_end_io
= bi_complete
;
2098 submit_bio(rw
, bio
);
2099 wait_for_completion(&event
);
2101 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2104 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2106 struct bio
*bio
= r10_bio
->master_bio
;
2107 struct mddev
*mddev
= r10_bio
->mddev
;
2108 struct r10conf
*conf
= mddev
->private;
2109 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2110 /* bio has the data to be written to slot 'i' where
2111 * we just recently had a write error.
2112 * We repeatedly clone the bio and trim down to one block,
2113 * then try the write. Where the write fails we record
2115 * It is conceivable that the bio doesn't exactly align with
2116 * blocks. We must handle this.
2118 * We currently own a reference to the rdev.
2124 int sect_to_write
= r10_bio
->sectors
;
2127 if (rdev
->badblocks
.shift
< 0)
2130 block_sectors
= 1 << rdev
->badblocks
.shift
;
2131 sector
= r10_bio
->sector
;
2132 sectors
= ((r10_bio
->sector
+ block_sectors
)
2133 & ~(sector_t
)(block_sectors
- 1))
2136 while (sect_to_write
) {
2138 if (sectors
> sect_to_write
)
2139 sectors
= sect_to_write
;
2140 /* Write at 'sector' for 'sectors' */
2141 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2142 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2143 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2145 (sector
- r10_bio
->sector
));
2146 wbio
->bi_bdev
= rdev
->bdev
;
2147 if (submit_bio_wait(WRITE
, wbio
) == 0)
2149 ok
= rdev_set_badblocks(rdev
, sector
,
2154 sect_to_write
-= sectors
;
2156 sectors
= block_sectors
;
2161 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2163 int slot
= r10_bio
->read_slot
;
2165 struct r10conf
*conf
= mddev
->private;
2166 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2167 char b
[BDEVNAME_SIZE
];
2168 unsigned long do_sync
;
2171 /* we got a read error. Maybe the drive is bad. Maybe just
2172 * the block and we can fix it.
2173 * We freeze all other IO, and try reading the block from
2174 * other devices. When we find one, we re-write
2175 * and check it that fixes the read error.
2176 * This is all done synchronously while the array is
2179 if (mddev
->ro
== 0) {
2181 fix_read_error(conf
, mddev
, r10_bio
);
2182 unfreeze_array(conf
);
2184 rdev_dec_pending(rdev
, mddev
);
2186 bio
= r10_bio
->devs
[slot
].bio
;
2187 bdevname(bio
->bi_bdev
, b
);
2188 r10_bio
->devs
[slot
].bio
=
2189 mddev
->ro
? IO_BLOCKED
: NULL
;
2191 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2193 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2194 " read error for block %llu\n",
2196 (unsigned long long)r10_bio
->sector
);
2197 raid_end_bio_io(r10_bio
);
2202 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2205 slot
= r10_bio
->read_slot
;
2208 "md/raid10:%s: %s: redirecting"
2209 "sector %llu to another mirror\n",
2211 bdevname(rdev
->bdev
, b
),
2212 (unsigned long long)r10_bio
->sector
);
2213 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2216 r10_bio
->sector
- bio
->bi_sector
,
2218 r10_bio
->devs
[slot
].bio
= bio
;
2219 r10_bio
->devs
[slot
].rdev
= rdev
;
2220 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2221 + rdev
->data_offset
;
2222 bio
->bi_bdev
= rdev
->bdev
;
2223 bio
->bi_rw
= READ
| do_sync
;
2224 bio
->bi_private
= r10_bio
;
2225 bio
->bi_end_io
= raid10_end_read_request
;
2226 if (max_sectors
< r10_bio
->sectors
) {
2227 /* Drat - have to split this up more */
2228 struct bio
*mbio
= r10_bio
->master_bio
;
2229 int sectors_handled
=
2230 r10_bio
->sector
+ max_sectors
2232 r10_bio
->sectors
= max_sectors
;
2233 spin_lock_irq(&conf
->device_lock
);
2234 if (mbio
->bi_phys_segments
== 0)
2235 mbio
->bi_phys_segments
= 2;
2237 mbio
->bi_phys_segments
++;
2238 spin_unlock_irq(&conf
->device_lock
);
2239 generic_make_request(bio
);
2242 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2244 r10_bio
->master_bio
= mbio
;
2245 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2248 set_bit(R10BIO_ReadError
,
2250 r10_bio
->mddev
= mddev
;
2251 r10_bio
->sector
= mbio
->bi_sector
2256 generic_make_request(bio
);
2259 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2261 /* Some sort of write request has finished and it
2262 * succeeded in writing where we thought there was a
2263 * bad block. So forget the bad block.
2264 * Or possibly if failed and we need to record
2268 struct md_rdev
*rdev
;
2270 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2271 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2272 for (m
= 0; m
< conf
->copies
; m
++) {
2273 int dev
= r10_bio
->devs
[m
].devnum
;
2274 rdev
= conf
->mirrors
[dev
].rdev
;
2275 if (r10_bio
->devs
[m
].bio
== NULL
)
2277 if (test_bit(BIO_UPTODATE
,
2278 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2279 rdev_clear_badblocks(
2281 r10_bio
->devs
[m
].addr
,
2284 if (!rdev_set_badblocks(
2286 r10_bio
->devs
[m
].addr
,
2287 r10_bio
->sectors
, 0))
2288 md_error(conf
->mddev
, rdev
);
2293 for (m
= 0; m
< conf
->copies
; m
++) {
2294 int dev
= r10_bio
->devs
[m
].devnum
;
2295 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2296 rdev
= conf
->mirrors
[dev
].rdev
;
2297 if (bio
== IO_MADE_GOOD
) {
2298 rdev_clear_badblocks(
2300 r10_bio
->devs
[m
].addr
,
2302 rdev_dec_pending(rdev
, conf
->mddev
);
2303 } else if (bio
!= NULL
&&
2304 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2305 if (!narrow_write_error(r10_bio
, m
)) {
2306 md_error(conf
->mddev
, rdev
);
2307 set_bit(R10BIO_Degraded
,
2310 rdev_dec_pending(rdev
, conf
->mddev
);
2312 bio
= r10_bio
->devs
[m
].repl_bio
;
2313 rdev
= conf
->mirrors
[dev
].replacement
;
2314 if (bio
== IO_MADE_GOOD
) {
2315 rdev_clear_badblocks(
2317 r10_bio
->devs
[m
].addr
,
2319 rdev_dec_pending(rdev
, conf
->mddev
);
2322 if (test_bit(R10BIO_WriteError
,
2324 close_write(r10_bio
);
2325 raid_end_bio_io(r10_bio
);
2329 static void raid10d(struct mddev
*mddev
)
2331 struct r10bio
*r10_bio
;
2332 unsigned long flags
;
2333 struct r10conf
*conf
= mddev
->private;
2334 struct list_head
*head
= &conf
->retry_list
;
2335 struct blk_plug plug
;
2337 md_check_recovery(mddev
);
2339 blk_start_plug(&plug
);
2342 flush_pending_writes(conf
);
2344 spin_lock_irqsave(&conf
->device_lock
, flags
);
2345 if (list_empty(head
)) {
2346 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2349 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2350 list_del(head
->prev
);
2352 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2354 mddev
= r10_bio
->mddev
;
2355 conf
= mddev
->private;
2356 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2357 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2358 handle_write_completed(conf
, r10_bio
);
2359 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2360 sync_request_write(mddev
, r10_bio
);
2361 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2362 recovery_request_write(mddev
, r10_bio
);
2363 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2364 handle_read_error(mddev
, r10_bio
);
2366 /* just a partial read to be scheduled from a
2369 int slot
= r10_bio
->read_slot
;
2370 generic_make_request(r10_bio
->devs
[slot
].bio
);
2374 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2375 md_check_recovery(mddev
);
2377 blk_finish_plug(&plug
);
2381 static int init_resync(struct r10conf
*conf
)
2386 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2387 BUG_ON(conf
->r10buf_pool
);
2388 conf
->have_replacement
= 0;
2389 for (i
= 0; i
< conf
->raid_disks
; i
++)
2390 if (conf
->mirrors
[i
].replacement
)
2391 conf
->have_replacement
= 1;
2392 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2393 if (!conf
->r10buf_pool
)
2395 conf
->next_resync
= 0;
2400 * perform a "sync" on one "block"
2402 * We need to make sure that no normal I/O request - particularly write
2403 * requests - conflict with active sync requests.
2405 * This is achieved by tracking pending requests and a 'barrier' concept
2406 * that can be installed to exclude normal IO requests.
2408 * Resync and recovery are handled very differently.
2409 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2411 * For resync, we iterate over virtual addresses, read all copies,
2412 * and update if there are differences. If only one copy is live,
2414 * For recovery, we iterate over physical addresses, read a good
2415 * value for each non-in_sync drive, and over-write.
2417 * So, for recovery we may have several outstanding complex requests for a
2418 * given address, one for each out-of-sync device. We model this by allocating
2419 * a number of r10_bio structures, one for each out-of-sync device.
2420 * As we setup these structures, we collect all bio's together into a list
2421 * which we then process collectively to add pages, and then process again
2422 * to pass to generic_make_request.
2424 * The r10_bio structures are linked using a borrowed master_bio pointer.
2425 * This link is counted in ->remaining. When the r10_bio that points to NULL
2426 * has its remaining count decremented to 0, the whole complex operation
2431 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2432 int *skipped
, int go_faster
)
2434 struct r10conf
*conf
= mddev
->private;
2435 struct r10bio
*r10_bio
;
2436 struct bio
*biolist
= NULL
, *bio
;
2437 sector_t max_sector
, nr_sectors
;
2440 sector_t sync_blocks
;
2441 sector_t sectors_skipped
= 0;
2442 int chunks_skipped
= 0;
2444 if (!conf
->r10buf_pool
)
2445 if (init_resync(conf
))
2449 max_sector
= mddev
->dev_sectors
;
2450 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2451 max_sector
= mddev
->resync_max_sectors
;
2452 if (sector_nr
>= max_sector
) {
2453 /* If we aborted, we need to abort the
2454 * sync on the 'current' bitmap chucks (there can
2455 * be several when recovering multiple devices).
2456 * as we may have started syncing it but not finished.
2457 * We can find the current address in
2458 * mddev->curr_resync, but for recovery,
2459 * we need to convert that to several
2460 * virtual addresses.
2462 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2463 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2464 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2466 else for (i
=0; i
<conf
->raid_disks
; i
++) {
2468 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2469 bitmap_end_sync(mddev
->bitmap
, sect
,
2472 } else /* completed sync */
2475 bitmap_close_sync(mddev
->bitmap
);
2478 return sectors_skipped
;
2480 if (chunks_skipped
>= conf
->raid_disks
) {
2481 /* if there has been nothing to do on any drive,
2482 * then there is nothing to do at all..
2485 return (max_sector
- sector_nr
) + sectors_skipped
;
2488 if (max_sector
> mddev
->resync_max
)
2489 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2491 /* make sure whole request will fit in a chunk - if chunks
2494 if (conf
->near_copies
< conf
->raid_disks
&&
2495 max_sector
> (sector_nr
| conf
->chunk_mask
))
2496 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
2498 * If there is non-resync activity waiting for us then
2499 * put in a delay to throttle resync.
2501 if (!go_faster
&& conf
->nr_waiting
)
2502 msleep_interruptible(1000);
2504 /* Again, very different code for resync and recovery.
2505 * Both must result in an r10bio with a list of bios that
2506 * have bi_end_io, bi_sector, bi_bdev set,
2507 * and bi_private set to the r10bio.
2508 * For recovery, we may actually create several r10bios
2509 * with 2 bios in each, that correspond to the bios in the main one.
2510 * In this case, the subordinate r10bios link back through a
2511 * borrowed master_bio pointer, and the counter in the master
2512 * includes a ref from each subordinate.
2514 /* First, we decide what to do and set ->bi_end_io
2515 * To end_sync_read if we want to read, and
2516 * end_sync_write if we will want to write.
2519 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2520 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2521 /* recovery... the complicated one */
2525 for (i
=0 ; i
<conf
->raid_disks
; i
++) {
2532 if (conf
->mirrors
[i
].rdev
== NULL
||
2533 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
))
2537 /* want to reconstruct this device */
2539 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2540 /* Unless we are doing a full sync, we only need
2541 * to recover the block if it is set in the bitmap
2543 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2545 if (sync_blocks
< max_sync
)
2546 max_sync
= sync_blocks
;
2549 /* yep, skip the sync_blocks here, but don't assume
2550 * that there will never be anything to do here
2552 chunks_skipped
= -1;
2556 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2557 raise_barrier(conf
, rb2
!= NULL
);
2558 atomic_set(&r10_bio
->remaining
, 0);
2560 r10_bio
->master_bio
= (struct bio
*)rb2
;
2562 atomic_inc(&rb2
->remaining
);
2563 r10_bio
->mddev
= mddev
;
2564 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
2565 r10_bio
->sector
= sect
;
2567 raid10_find_phys(conf
, r10_bio
);
2569 /* Need to check if the array will still be
2572 for (j
=0; j
<conf
->raid_disks
; j
++)
2573 if (conf
->mirrors
[j
].rdev
== NULL
||
2574 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
2579 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2580 &sync_blocks
, still_degraded
);
2583 for (j
=0; j
<conf
->copies
;j
++) {
2585 int d
= r10_bio
->devs
[j
].devnum
;
2586 sector_t from_addr
, to_addr
;
2587 struct md_rdev
*rdev
;
2588 sector_t sector
, first_bad
;
2590 if (!conf
->mirrors
[d
].rdev
||
2591 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
2593 /* This is where we read from */
2595 rdev
= conf
->mirrors
[d
].rdev
;
2596 sector
= r10_bio
->devs
[j
].addr
;
2598 if (is_badblock(rdev
, sector
, max_sync
,
2599 &first_bad
, &bad_sectors
)) {
2600 if (first_bad
> sector
)
2601 max_sync
= first_bad
- sector
;
2603 bad_sectors
-= (sector
2605 if (max_sync
> bad_sectors
)
2606 max_sync
= bad_sectors
;
2610 bio
= r10_bio
->devs
[0].bio
;
2611 bio
->bi_next
= biolist
;
2613 bio
->bi_private
= r10_bio
;
2614 bio
->bi_end_io
= end_sync_read
;
2616 from_addr
= r10_bio
->devs
[j
].addr
;
2617 bio
->bi_sector
= from_addr
+
2618 conf
->mirrors
[d
].rdev
->data_offset
;
2619 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2620 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2621 atomic_inc(&r10_bio
->remaining
);
2622 /* and we write to 'i' */
2624 for (k
=0; k
<conf
->copies
; k
++)
2625 if (r10_bio
->devs
[k
].devnum
== i
)
2627 BUG_ON(k
== conf
->copies
);
2628 bio
= r10_bio
->devs
[1].bio
;
2629 bio
->bi_next
= biolist
;
2631 bio
->bi_private
= r10_bio
;
2632 bio
->bi_end_io
= end_sync_write
;
2634 to_addr
= r10_bio
->devs
[k
].addr
;
2635 bio
->bi_sector
= to_addr
+
2636 conf
->mirrors
[i
].rdev
->data_offset
;
2637 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
2639 r10_bio
->devs
[0].devnum
= d
;
2640 r10_bio
->devs
[0].addr
= from_addr
;
2641 r10_bio
->devs
[1].devnum
= i
;
2642 r10_bio
->devs
[1].addr
= to_addr
;
2646 if (j
== conf
->copies
) {
2647 /* Cannot recover, so abort the recovery or
2648 * record a bad block */
2651 atomic_dec(&rb2
->remaining
);
2654 /* problem is that there are bad blocks
2655 * on other device(s)
2658 for (k
= 0; k
< conf
->copies
; k
++)
2659 if (r10_bio
->devs
[k
].devnum
== i
)
2661 if (!rdev_set_badblocks(
2662 conf
->mirrors
[i
].rdev
,
2663 r10_bio
->devs
[k
].addr
,
2668 if (!test_and_set_bit(MD_RECOVERY_INTR
,
2670 printk(KERN_INFO
"md/raid10:%s: insufficient "
2671 "working devices for recovery.\n",
2673 conf
->mirrors
[i
].recovery_disabled
2674 = mddev
->recovery_disabled
;
2679 if (biolist
== NULL
) {
2681 struct r10bio
*rb2
= r10_bio
;
2682 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
2683 rb2
->master_bio
= NULL
;
2689 /* resync. Schedule a read for every block at this virt offset */
2692 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
2694 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2695 &sync_blocks
, mddev
->degraded
) &&
2696 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
2697 &mddev
->recovery
)) {
2698 /* We can skip this block */
2700 return sync_blocks
+ sectors_skipped
;
2702 if (sync_blocks
< max_sync
)
2703 max_sync
= sync_blocks
;
2704 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2706 r10_bio
->mddev
= mddev
;
2707 atomic_set(&r10_bio
->remaining
, 0);
2708 raise_barrier(conf
, 0);
2709 conf
->next_resync
= sector_nr
;
2711 r10_bio
->master_bio
= NULL
;
2712 r10_bio
->sector
= sector_nr
;
2713 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
2714 raid10_find_phys(conf
, r10_bio
);
2715 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
2717 for (i
=0; i
<conf
->copies
; i
++) {
2718 int d
= r10_bio
->devs
[i
].devnum
;
2719 sector_t first_bad
, sector
;
2722 bio
= r10_bio
->devs
[i
].bio
;
2723 bio
->bi_end_io
= NULL
;
2724 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2725 if (conf
->mirrors
[d
].rdev
== NULL
||
2726 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
2728 sector
= r10_bio
->devs
[i
].addr
;
2729 if (is_badblock(conf
->mirrors
[d
].rdev
,
2731 &first_bad
, &bad_sectors
)) {
2732 if (first_bad
> sector
)
2733 max_sync
= first_bad
- sector
;
2735 bad_sectors
-= (sector
- first_bad
);
2736 if (max_sync
> bad_sectors
)
2737 max_sync
= max_sync
;
2741 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2742 atomic_inc(&r10_bio
->remaining
);
2743 bio
->bi_next
= biolist
;
2745 bio
->bi_private
= r10_bio
;
2746 bio
->bi_end_io
= end_sync_read
;
2748 bio
->bi_sector
= sector
+
2749 conf
->mirrors
[d
].rdev
->data_offset
;
2750 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2755 for (i
=0; i
<conf
->copies
; i
++) {
2756 int d
= r10_bio
->devs
[i
].devnum
;
2757 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
2758 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
2767 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2769 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2771 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
2774 bio
->bi_phys_segments
= 0;
2779 if (sector_nr
+ max_sync
< max_sector
)
2780 max_sector
= sector_nr
+ max_sync
;
2783 int len
= PAGE_SIZE
;
2784 if (sector_nr
+ (len
>>9) > max_sector
)
2785 len
= (max_sector
- sector_nr
) << 9;
2788 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2790 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2791 if (bio_add_page(bio
, page
, len
, 0))
2795 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2796 for (bio2
= biolist
;
2797 bio2
&& bio2
!= bio
;
2798 bio2
= bio2
->bi_next
) {
2799 /* remove last page from this bio */
2801 bio2
->bi_size
-= len
;
2802 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
2806 nr_sectors
+= len
>>9;
2807 sector_nr
+= len
>>9;
2808 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
2810 r10_bio
->sectors
= nr_sectors
;
2814 biolist
= biolist
->bi_next
;
2816 bio
->bi_next
= NULL
;
2817 r10_bio
= bio
->bi_private
;
2818 r10_bio
->sectors
= nr_sectors
;
2820 if (bio
->bi_end_io
== end_sync_read
) {
2821 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2822 generic_make_request(bio
);
2826 if (sectors_skipped
)
2827 /* pretend they weren't skipped, it makes
2828 * no important difference in this case
2830 md_done_sync(mddev
, sectors_skipped
, 1);
2832 return sectors_skipped
+ nr_sectors
;
2834 /* There is nowhere to write, so all non-sync
2835 * drives must be failed or in resync, all drives
2836 * have a bad block, so try the next chunk...
2838 if (sector_nr
+ max_sync
< max_sector
)
2839 max_sector
= sector_nr
+ max_sync
;
2841 sectors_skipped
+= (max_sector
- sector_nr
);
2843 sector_nr
= max_sector
;
2848 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2851 struct r10conf
*conf
= mddev
->private;
2854 raid_disks
= conf
->raid_disks
;
2856 sectors
= conf
->dev_sectors
;
2858 size
= sectors
>> conf
->chunk_shift
;
2859 sector_div(size
, conf
->far_copies
);
2860 size
= size
* raid_disks
;
2861 sector_div(size
, conf
->near_copies
);
2863 return size
<< conf
->chunk_shift
;
2867 static struct r10conf
*setup_conf(struct mddev
*mddev
)
2869 struct r10conf
*conf
= NULL
;
2871 sector_t stride
, size
;
2874 if (mddev
->new_chunk_sectors
< (PAGE_SIZE
>> 9) ||
2875 !is_power_of_2(mddev
->new_chunk_sectors
)) {
2876 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
2877 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2878 mdname(mddev
), PAGE_SIZE
);
2882 nc
= mddev
->new_layout
& 255;
2883 fc
= (mddev
->new_layout
>> 8) & 255;
2884 fo
= mddev
->new_layout
& (1<<16);
2886 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
2887 (mddev
->new_layout
>> 17)) {
2888 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2889 mdname(mddev
), mddev
->new_layout
);
2894 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
2898 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
2903 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2908 conf
->raid_disks
= mddev
->raid_disks
;
2909 conf
->near_copies
= nc
;
2910 conf
->far_copies
= fc
;
2911 conf
->copies
= nc
*fc
;
2912 conf
->far_offset
= fo
;
2913 conf
->chunk_mask
= mddev
->new_chunk_sectors
- 1;
2914 conf
->chunk_shift
= ffz(~mddev
->new_chunk_sectors
);
2916 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
2917 r10bio_pool_free
, conf
);
2918 if (!conf
->r10bio_pool
)
2921 size
= mddev
->dev_sectors
>> conf
->chunk_shift
;
2922 sector_div(size
, fc
);
2923 size
= size
* conf
->raid_disks
;
2924 sector_div(size
, nc
);
2925 /* 'size' is now the number of chunks in the array */
2926 /* calculate "used chunks per device" in 'stride' */
2927 stride
= size
* conf
->copies
;
2929 /* We need to round up when dividing by raid_disks to
2930 * get the stride size.
2932 stride
+= conf
->raid_disks
- 1;
2933 sector_div(stride
, conf
->raid_disks
);
2935 conf
->dev_sectors
= stride
<< conf
->chunk_shift
;
2940 sector_div(stride
, fc
);
2941 conf
->stride
= stride
<< conf
->chunk_shift
;
2944 spin_lock_init(&conf
->device_lock
);
2945 INIT_LIST_HEAD(&conf
->retry_list
);
2947 spin_lock_init(&conf
->resync_lock
);
2948 init_waitqueue_head(&conf
->wait_barrier
);
2950 conf
->thread
= md_register_thread(raid10d
, mddev
, NULL
);
2954 conf
->mddev
= mddev
;
2958 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
2961 if (conf
->r10bio_pool
)
2962 mempool_destroy(conf
->r10bio_pool
);
2963 kfree(conf
->mirrors
);
2964 safe_put_page(conf
->tmppage
);
2967 return ERR_PTR(err
);
2970 static int run(struct mddev
*mddev
)
2972 struct r10conf
*conf
;
2973 int i
, disk_idx
, chunk_size
;
2974 struct mirror_info
*disk
;
2975 struct md_rdev
*rdev
;
2979 * copy the already verified devices into our private RAID10
2980 * bookkeeping area. [whatever we allocate in run(),
2981 * should be freed in stop()]
2984 if (mddev
->private == NULL
) {
2985 conf
= setup_conf(mddev
);
2987 return PTR_ERR(conf
);
2988 mddev
->private = conf
;
2990 conf
= mddev
->private;
2994 mddev
->thread
= conf
->thread
;
2995 conf
->thread
= NULL
;
2997 chunk_size
= mddev
->chunk_sectors
<< 9;
2998 blk_queue_io_min(mddev
->queue
, chunk_size
);
2999 if (conf
->raid_disks
% conf
->near_copies
)
3000 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->raid_disks
);
3002 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3003 (conf
->raid_disks
/ conf
->near_copies
));
3005 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
3007 disk_idx
= rdev
->raid_disk
;
3008 if (disk_idx
>= conf
->raid_disks
3011 disk
= conf
->mirrors
+ disk_idx
;
3014 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3015 rdev
->data_offset
<< 9);
3016 /* as we don't honour merge_bvec_fn, we must never risk
3017 * violating it, so limit max_segments to 1 lying
3018 * within a single page.
3020 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
3021 blk_queue_max_segments(mddev
->queue
, 1);
3022 blk_queue_segment_boundary(mddev
->queue
,
3023 PAGE_CACHE_SIZE
- 1);
3026 disk
->head_position
= 0;
3028 /* need to check that every block has at least one working mirror */
3029 if (!enough(conf
, -1)) {
3030 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3035 mddev
->degraded
= 0;
3036 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3038 disk
= conf
->mirrors
+ i
;
3041 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3042 disk
->head_position
= 0;
3047 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3050 if (mddev
->recovery_cp
!= MaxSector
)
3051 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3052 " -- starting background reconstruction\n",
3055 "md/raid10:%s: active with %d out of %d devices\n",
3056 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
,
3059 * Ok, everything is just fine now
3061 mddev
->dev_sectors
= conf
->dev_sectors
;
3062 size
= raid10_size(mddev
, 0, 0);
3063 md_set_array_sectors(mddev
, size
);
3064 mddev
->resync_max_sectors
= size
;
3066 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3067 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3069 /* Calculate max read-ahead size.
3070 * We need to readahead at least twice a whole stripe....
3074 int stripe
= conf
->raid_disks
*
3075 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3076 stripe
/= conf
->near_copies
;
3077 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
3078 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
3081 if (conf
->near_copies
< conf
->raid_disks
)
3082 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3084 if (md_integrity_register(mddev
))
3090 md_unregister_thread(&mddev
->thread
);
3091 if (conf
->r10bio_pool
)
3092 mempool_destroy(conf
->r10bio_pool
);
3093 safe_put_page(conf
->tmppage
);
3094 kfree(conf
->mirrors
);
3096 mddev
->private = NULL
;
3101 static int stop(struct mddev
*mddev
)
3103 struct r10conf
*conf
= mddev
->private;
3105 raise_barrier(conf
, 0);
3106 lower_barrier(conf
);
3108 md_unregister_thread(&mddev
->thread
);
3109 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
3110 if (conf
->r10bio_pool
)
3111 mempool_destroy(conf
->r10bio_pool
);
3112 kfree(conf
->mirrors
);
3114 mddev
->private = NULL
;
3118 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3120 struct r10conf
*conf
= mddev
->private;
3124 raise_barrier(conf
, 0);
3127 lower_barrier(conf
);
3132 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3134 struct md_rdev
*rdev
;
3135 struct r10conf
*conf
;
3137 if (mddev
->degraded
> 0) {
3138 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3140 return ERR_PTR(-EINVAL
);
3143 /* Set new parameters */
3144 mddev
->new_level
= 10;
3145 /* new layout: far_copies = 1, near_copies = 2 */
3146 mddev
->new_layout
= (1<<8) + 2;
3147 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3148 mddev
->delta_disks
= mddev
->raid_disks
;
3149 mddev
->raid_disks
*= 2;
3150 /* make sure it will be not marked as dirty */
3151 mddev
->recovery_cp
= MaxSector
;
3153 conf
= setup_conf(mddev
);
3154 if (!IS_ERR(conf
)) {
3155 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
3156 if (rdev
->raid_disk
>= 0)
3157 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3164 static void *raid10_takeover(struct mddev
*mddev
)
3166 struct r0conf
*raid0_conf
;
3168 /* raid10 can take over:
3169 * raid0 - providing it has only two drives
3171 if (mddev
->level
== 0) {
3172 /* for raid0 takeover only one zone is supported */
3173 raid0_conf
= mddev
->private;
3174 if (raid0_conf
->nr_strip_zones
> 1) {
3175 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3176 " with more than one zone.\n",
3178 return ERR_PTR(-EINVAL
);
3180 return raid10_takeover_raid0(mddev
);
3182 return ERR_PTR(-EINVAL
);
3185 static struct md_personality raid10_personality
=
3189 .owner
= THIS_MODULE
,
3190 .make_request
= make_request
,
3194 .error_handler
= error
,
3195 .hot_add_disk
= raid10_add_disk
,
3196 .hot_remove_disk
= raid10_remove_disk
,
3197 .spare_active
= raid10_spare_active
,
3198 .sync_request
= sync_request
,
3199 .quiesce
= raid10_quiesce
,
3200 .size
= raid10_size
,
3201 .takeover
= raid10_takeover
,
3204 static int __init
raid_init(void)
3206 return register_md_personality(&raid10_personality
);
3209 static void raid_exit(void)
3211 unregister_md_personality(&raid10_personality
);
3214 module_init(raid_init
);
3215 module_exit(raid_exit
);
3216 MODULE_LICENSE("GPL");
3217 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3218 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3219 MODULE_ALIAS("md-raid10");
3220 MODULE_ALIAS("md-level-10");
3222 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);