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
);
70 static int enough(struct r10conf
*conf
, int ignore
);
72 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
74 struct r10conf
*conf
= data
;
75 int size
= offsetof(struct r10bio
, devs
[conf
->copies
]);
77 /* allocate a r10bio with room for raid_disks entries in the
79 return kzalloc(size
, gfp_flags
);
82 static void r10bio_pool_free(void *r10_bio
, void *data
)
87 /* Maximum size of each resync request */
88 #define RESYNC_BLOCK_SIZE (64*1024)
89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
90 /* amount of memory to reserve for resync requests */
91 #define RESYNC_WINDOW (1024*1024)
92 /* maximum number of concurrent requests, memory permitting */
93 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
96 * When performing a resync, we need to read and compare, so
97 * we need as many pages are there are copies.
98 * When performing a recovery, we need 2 bios, one for read,
99 * one for write (we recover only one drive per r10buf)
102 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
104 struct r10conf
*conf
= data
;
106 struct r10bio
*r10_bio
;
111 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
115 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
116 nalloc
= conf
->copies
; /* resync */
118 nalloc
= 2; /* recovery */
123 for (j
= nalloc
; j
-- ; ) {
124 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
127 r10_bio
->devs
[j
].bio
= bio
;
128 if (!conf
->have_replacement
)
130 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
133 r10_bio
->devs
[j
].repl_bio
= bio
;
136 * Allocate RESYNC_PAGES data pages and attach them
139 for (j
= 0 ; j
< nalloc
; j
++) {
140 struct bio
*rbio
= r10_bio
->devs
[j
].repl_bio
;
141 bio
= r10_bio
->devs
[j
].bio
;
142 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
143 if (j
== 1 && !test_bit(MD_RECOVERY_SYNC
,
144 &conf
->mddev
->recovery
)) {
145 /* we can share bv_page's during recovery */
146 struct bio
*rbio
= r10_bio
->devs
[0].bio
;
147 page
= rbio
->bi_io_vec
[i
].bv_page
;
150 page
= alloc_page(gfp_flags
);
154 bio
->bi_io_vec
[i
].bv_page
= page
;
156 rbio
->bi_io_vec
[i
].bv_page
= page
;
164 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
166 for (i
= 0; i
< RESYNC_PAGES
; i
++)
167 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
170 while (++j
< nalloc
) {
171 bio_put(r10_bio
->devs
[j
].bio
);
172 if (r10_bio
->devs
[j
].repl_bio
)
173 bio_put(r10_bio
->devs
[j
].repl_bio
);
175 r10bio_pool_free(r10_bio
, conf
);
179 static void r10buf_pool_free(void *__r10_bio
, void *data
)
182 struct r10conf
*conf
= data
;
183 struct r10bio
*r10bio
= __r10_bio
;
186 for (j
=0; j
< conf
->copies
; j
++) {
187 struct bio
*bio
= r10bio
->devs
[j
].bio
;
189 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
190 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
191 bio
->bi_io_vec
[i
].bv_page
= NULL
;
195 bio
= r10bio
->devs
[j
].repl_bio
;
199 r10bio_pool_free(r10bio
, conf
);
202 static void put_all_bios(struct r10conf
*conf
, struct r10bio
*r10_bio
)
206 for (i
= 0; i
< conf
->copies
; i
++) {
207 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
208 if (!BIO_SPECIAL(*bio
))
211 bio
= &r10_bio
->devs
[i
].repl_bio
;
212 if (r10_bio
->read_slot
< 0 && !BIO_SPECIAL(*bio
))
218 static void free_r10bio(struct r10bio
*r10_bio
)
220 struct r10conf
*conf
= r10_bio
->mddev
->private;
222 put_all_bios(conf
, r10_bio
);
223 mempool_free(r10_bio
, conf
->r10bio_pool
);
226 static void put_buf(struct r10bio
*r10_bio
)
228 struct r10conf
*conf
= r10_bio
->mddev
->private;
230 mempool_free(r10_bio
, conf
->r10buf_pool
);
235 static void reschedule_retry(struct r10bio
*r10_bio
)
238 struct mddev
*mddev
= r10_bio
->mddev
;
239 struct r10conf
*conf
= mddev
->private;
241 spin_lock_irqsave(&conf
->device_lock
, flags
);
242 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
244 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
246 /* wake up frozen array... */
247 wake_up(&conf
->wait_barrier
);
249 md_wakeup_thread(mddev
->thread
);
253 * raid_end_bio_io() is called when we have finished servicing a mirrored
254 * operation and are ready to return a success/failure code to the buffer
257 static void raid_end_bio_io(struct r10bio
*r10_bio
)
259 struct bio
*bio
= r10_bio
->master_bio
;
261 struct r10conf
*conf
= r10_bio
->mddev
->private;
263 if (bio
->bi_phys_segments
) {
265 spin_lock_irqsave(&conf
->device_lock
, flags
);
266 bio
->bi_phys_segments
--;
267 done
= (bio
->bi_phys_segments
== 0);
268 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
271 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
272 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
276 * Wake up any possible resync thread that waits for the device
281 free_r10bio(r10_bio
);
285 * Update disk head position estimator based on IRQ completion info.
287 static inline void update_head_pos(int slot
, struct r10bio
*r10_bio
)
289 struct r10conf
*conf
= r10_bio
->mddev
->private;
291 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
292 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
296 * Find the disk number which triggered given bio
298 static int find_bio_disk(struct r10conf
*conf
, struct r10bio
*r10_bio
,
299 struct bio
*bio
, int *slotp
, int *replp
)
304 for (slot
= 0; slot
< conf
->copies
; slot
++) {
305 if (r10_bio
->devs
[slot
].bio
== bio
)
307 if (r10_bio
->devs
[slot
].repl_bio
== bio
) {
313 BUG_ON(slot
== conf
->copies
);
314 update_head_pos(slot
, r10_bio
);
320 return r10_bio
->devs
[slot
].devnum
;
323 static void raid10_end_read_request(struct bio
*bio
, int error
)
325 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
326 struct r10bio
*r10_bio
= bio
->bi_private
;
328 struct md_rdev
*rdev
;
329 struct r10conf
*conf
= r10_bio
->mddev
->private;
332 slot
= r10_bio
->read_slot
;
333 dev
= r10_bio
->devs
[slot
].devnum
;
334 rdev
= r10_bio
->devs
[slot
].rdev
;
336 * this branch is our 'one mirror IO has finished' event handler:
338 update_head_pos(slot
, r10_bio
);
342 * Set R10BIO_Uptodate in our master bio, so that
343 * we will return a good error code to the higher
344 * levels even if IO on some other mirrored buffer fails.
346 * The 'master' represents the composite IO operation to
347 * user-side. So if something waits for IO, then it will
348 * wait for the 'master' bio.
350 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
352 /* If all other devices that store this block have
353 * failed, we want to return the error upwards rather
354 * than fail the last device. Here we redefine
355 * "uptodate" to mean "Don't want to retry"
358 spin_lock_irqsave(&conf
->device_lock
, flags
);
359 if (!enough(conf
, rdev
->raid_disk
))
361 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
364 raid_end_bio_io(r10_bio
);
365 rdev_dec_pending(rdev
, conf
->mddev
);
368 * oops, read error - keep the refcount on the rdev
370 char b
[BDEVNAME_SIZE
];
371 printk_ratelimited(KERN_ERR
372 "md/raid10:%s: %s: rescheduling sector %llu\n",
374 bdevname(rdev
->bdev
, b
),
375 (unsigned long long)r10_bio
->sector
);
376 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
377 reschedule_retry(r10_bio
);
381 static void close_write(struct r10bio
*r10_bio
)
383 /* clear the bitmap if all writes complete successfully */
384 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
386 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
388 md_write_end(r10_bio
->mddev
);
391 static void one_write_done(struct r10bio
*r10_bio
)
393 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
394 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
395 reschedule_retry(r10_bio
);
397 close_write(r10_bio
);
398 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
399 reschedule_retry(r10_bio
);
401 raid_end_bio_io(r10_bio
);
406 static void raid10_end_write_request(struct bio
*bio
, int error
)
408 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
409 struct r10bio
*r10_bio
= bio
->bi_private
;
412 struct r10conf
*conf
= r10_bio
->mddev
->private;
414 struct md_rdev
*rdev
= NULL
;
416 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
419 rdev
= conf
->mirrors
[dev
].replacement
;
423 rdev
= conf
->mirrors
[dev
].rdev
;
426 * this branch is our 'one mirror IO has finished' event handler:
430 /* Never record new bad blocks to replacement,
433 md_error(rdev
->mddev
, rdev
);
435 set_bit(WriteErrorSeen
, &rdev
->flags
);
436 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
437 set_bit(MD_RECOVERY_NEEDED
,
438 &rdev
->mddev
->recovery
);
439 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
444 * Set R10BIO_Uptodate in our master bio, so that
445 * we will return a good error code for to the higher
446 * levels even if IO on some other mirrored buffer fails.
448 * The 'master' represents the composite IO operation to
449 * user-side. So if something waits for IO, then it will
450 * wait for the 'master' bio.
455 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
457 /* Maybe we can clear some bad blocks. */
458 if (is_badblock(rdev
,
459 r10_bio
->devs
[slot
].addr
,
461 &first_bad
, &bad_sectors
)) {
464 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
466 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
468 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
474 * Let's see if all mirrored write operations have finished
477 one_write_done(r10_bio
);
479 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
483 * RAID10 layout manager
484 * As well as the chunksize and raid_disks count, there are two
485 * parameters: near_copies and far_copies.
486 * near_copies * far_copies must be <= raid_disks.
487 * Normally one of these will be 1.
488 * If both are 1, we get raid0.
489 * If near_copies == raid_disks, we get raid1.
491 * Chunks are laid out in raid0 style with near_copies copies of the
492 * first chunk, followed by near_copies copies of the next chunk and
494 * If far_copies > 1, then after 1/far_copies of the array has been assigned
495 * as described above, we start again with a device offset of near_copies.
496 * So we effectively have another copy of the whole array further down all
497 * the drives, but with blocks on different drives.
498 * With this layout, and block is never stored twice on the one device.
500 * raid10_find_phys finds the sector offset of a given virtual sector
501 * on each device that it is on.
503 * raid10_find_virt does the reverse mapping, from a device and a
504 * sector offset to a virtual address
507 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
516 /* now calculate first sector/dev */
517 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
518 sector
= r10bio
->sector
& geo
->chunk_mask
;
520 chunk
*= geo
->near_copies
;
522 dev
= sector_div(stripe
, geo
->raid_disks
);
524 stripe
*= geo
->far_copies
;
526 sector
+= stripe
<< geo
->chunk_shift
;
528 /* and calculate all the others */
529 for (n
= 0; n
< geo
->near_copies
; n
++) {
532 r10bio
->devs
[slot
].addr
= sector
;
533 r10bio
->devs
[slot
].devnum
= d
;
536 for (f
= 1; f
< geo
->far_copies
; f
++) {
537 d
+= geo
->near_copies
;
538 if (d
>= geo
->raid_disks
)
539 d
-= geo
->raid_disks
;
541 r10bio
->devs
[slot
].devnum
= d
;
542 r10bio
->devs
[slot
].addr
= s
;
546 if (dev
>= geo
->raid_disks
) {
548 sector
+= (geo
->chunk_mask
+ 1);
553 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
555 struct geom
*geo
= &conf
->geo
;
557 if (conf
->reshape_progress
!= MaxSector
&&
558 ((r10bio
->sector
>= conf
->reshape_progress
) !=
559 conf
->mddev
->reshape_backwards
)) {
560 set_bit(R10BIO_Previous
, &r10bio
->state
);
563 clear_bit(R10BIO_Previous
, &r10bio
->state
);
565 __raid10_find_phys(geo
, r10bio
);
568 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
570 sector_t offset
, chunk
, vchunk
;
571 /* Never use conf->prev as this is only called during resync
572 * or recovery, so reshape isn't happening
574 struct geom
*geo
= &conf
->geo
;
576 offset
= sector
& geo
->chunk_mask
;
577 if (geo
->far_offset
) {
579 chunk
= sector
>> geo
->chunk_shift
;
580 fc
= sector_div(chunk
, geo
->far_copies
);
581 dev
-= fc
* geo
->near_copies
;
583 dev
+= geo
->raid_disks
;
585 while (sector
>= geo
->stride
) {
586 sector
-= geo
->stride
;
587 if (dev
< geo
->near_copies
)
588 dev
+= geo
->raid_disks
- geo
->near_copies
;
590 dev
-= geo
->near_copies
;
592 chunk
= sector
>> geo
->chunk_shift
;
594 vchunk
= chunk
* geo
->raid_disks
+ dev
;
595 sector_div(vchunk
, geo
->near_copies
);
596 return (vchunk
<< geo
->chunk_shift
) + offset
;
600 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
602 * @bvm: properties of new bio
603 * @biovec: the request that could be merged to it.
605 * Return amount of bytes we can accept at this offset
606 * This requires checking for end-of-chunk if near_copies != raid_disks,
607 * and for subordinate merge_bvec_fns if merge_check_needed.
609 static int raid10_mergeable_bvec(struct request_queue
*q
,
610 struct bvec_merge_data
*bvm
,
611 struct bio_vec
*biovec
)
613 struct mddev
*mddev
= q
->queuedata
;
614 struct r10conf
*conf
= mddev
->private;
615 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
617 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
618 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
619 struct geom
*geo
= &conf
->geo
;
621 if (conf
->reshape_progress
!= MaxSector
&&
622 ((sector
>= conf
->reshape_progress
) !=
623 conf
->mddev
->reshape_backwards
))
626 if (geo
->near_copies
< geo
->raid_disks
) {
627 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1))
628 + bio_sectors
)) << 9;
630 /* bio_add cannot handle a negative return */
632 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
633 return biovec
->bv_len
;
635 max
= biovec
->bv_len
;
637 if (mddev
->merge_check_needed
) {
638 struct r10bio r10_bio
;
640 if (conf
->reshape_progress
!= MaxSector
) {
641 /* Cannot give any guidance during reshape */
642 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
643 return biovec
->bv_len
;
646 r10_bio
.sector
= sector
;
647 raid10_find_phys(conf
, &r10_bio
);
649 for (s
= 0; s
< conf
->copies
; s
++) {
650 int disk
= r10_bio
.devs
[s
].devnum
;
651 struct md_rdev
*rdev
= rcu_dereference(
652 conf
->mirrors
[disk
].rdev
);
653 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
654 struct request_queue
*q
=
655 bdev_get_queue(rdev
->bdev
);
656 if (q
->merge_bvec_fn
) {
657 bvm
->bi_sector
= r10_bio
.devs
[s
].addr
659 bvm
->bi_bdev
= rdev
->bdev
;
660 max
= min(max
, q
->merge_bvec_fn(
664 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
665 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
666 struct request_queue
*q
=
667 bdev_get_queue(rdev
->bdev
);
668 if (q
->merge_bvec_fn
) {
669 bvm
->bi_sector
= r10_bio
.devs
[s
].addr
671 bvm
->bi_bdev
= rdev
->bdev
;
672 max
= min(max
, q
->merge_bvec_fn(
683 * This routine returns the disk from which the requested read should
684 * be done. There is a per-array 'next expected sequential IO' sector
685 * number - if this matches on the next IO then we use the last disk.
686 * There is also a per-disk 'last know head position' sector that is
687 * maintained from IRQ contexts, both the normal and the resync IO
688 * completion handlers update this position correctly. If there is no
689 * perfect sequential match then we pick the disk whose head is closest.
691 * If there are 2 mirrors in the same 2 devices, performance degrades
692 * because position is mirror, not device based.
694 * The rdev for the device selected will have nr_pending incremented.
698 * FIXME: possibly should rethink readbalancing and do it differently
699 * depending on near_copies / far_copies geometry.
701 static struct md_rdev
*read_balance(struct r10conf
*conf
,
702 struct r10bio
*r10_bio
,
705 const sector_t this_sector
= r10_bio
->sector
;
707 int sectors
= r10_bio
->sectors
;
708 int best_good_sectors
;
709 sector_t new_distance
, best_dist
;
710 struct md_rdev
*rdev
, *best_rdev
;
713 struct geom
*geo
= &conf
->geo
;
715 raid10_find_phys(conf
, r10_bio
);
718 sectors
= r10_bio
->sectors
;
721 best_dist
= MaxSector
;
722 best_good_sectors
= 0;
725 * Check if we can balance. We can balance on the whole
726 * device if no resync is going on (recovery is ok), or below
727 * the resync window. We take the first readable disk when
728 * above the resync window.
730 if (conf
->mddev
->recovery_cp
< MaxSector
731 && (this_sector
+ sectors
>= conf
->next_resync
))
734 for (slot
= 0; slot
< conf
->copies
; slot
++) {
739 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
741 disk
= r10_bio
->devs
[slot
].devnum
;
742 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
743 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
744 test_bit(Unmerged
, &rdev
->flags
) ||
745 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
746 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
748 test_bit(Faulty
, &rdev
->flags
) ||
749 test_bit(Unmerged
, &rdev
->flags
))
751 if (!test_bit(In_sync
, &rdev
->flags
) &&
752 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
755 dev_sector
= r10_bio
->devs
[slot
].addr
;
756 if (is_badblock(rdev
, dev_sector
, sectors
,
757 &first_bad
, &bad_sectors
)) {
758 if (best_dist
< MaxSector
)
759 /* Already have a better slot */
761 if (first_bad
<= dev_sector
) {
762 /* Cannot read here. If this is the
763 * 'primary' device, then we must not read
764 * beyond 'bad_sectors' from another device.
766 bad_sectors
-= (dev_sector
- first_bad
);
767 if (!do_balance
&& sectors
> bad_sectors
)
768 sectors
= bad_sectors
;
769 if (best_good_sectors
> sectors
)
770 best_good_sectors
= sectors
;
772 sector_t good_sectors
=
773 first_bad
- dev_sector
;
774 if (good_sectors
> best_good_sectors
) {
775 best_good_sectors
= good_sectors
;
780 /* Must read from here */
785 best_good_sectors
= sectors
;
790 /* This optimisation is debatable, and completely destroys
791 * sequential read speed for 'far copies' arrays. So only
792 * keep it for 'near' arrays, and review those later.
794 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
797 /* for far > 1 always use the lowest address */
798 if (geo
->far_copies
> 1)
799 new_distance
= r10_bio
->devs
[slot
].addr
;
801 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
802 conf
->mirrors
[disk
].head_position
);
803 if (new_distance
< best_dist
) {
804 best_dist
= new_distance
;
809 if (slot
>= conf
->copies
) {
815 atomic_inc(&rdev
->nr_pending
);
816 if (test_bit(Faulty
, &rdev
->flags
)) {
817 /* Cannot risk returning a device that failed
818 * before we inc'ed nr_pending
820 rdev_dec_pending(rdev
, conf
->mddev
);
823 r10_bio
->read_slot
= slot
;
827 *max_sectors
= best_good_sectors
;
832 static int raid10_congested(void *data
, int bits
)
834 struct mddev
*mddev
= data
;
835 struct r10conf
*conf
= mddev
->private;
838 if ((bits
& (1 << BDI_async_congested
)) &&
839 conf
->pending_count
>= max_queued_requests
)
842 if (mddev_congested(mddev
, bits
))
846 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
849 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
850 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
851 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
853 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
860 static void flush_pending_writes(struct r10conf
*conf
)
862 /* Any writes that have been queued but are awaiting
863 * bitmap updates get flushed here.
865 spin_lock_irq(&conf
->device_lock
);
867 if (conf
->pending_bio_list
.head
) {
869 bio
= bio_list_get(&conf
->pending_bio_list
);
870 conf
->pending_count
= 0;
871 spin_unlock_irq(&conf
->device_lock
);
872 /* flush any pending bitmap writes to disk
873 * before proceeding w/ I/O */
874 bitmap_unplug(conf
->mddev
->bitmap
);
875 wake_up(&conf
->wait_barrier
);
877 while (bio
) { /* submit pending writes */
878 struct bio
*next
= bio
->bi_next
;
880 generic_make_request(bio
);
884 spin_unlock_irq(&conf
->device_lock
);
888 * Sometimes we need to suspend IO while we do something else,
889 * either some resync/recovery, or reconfigure the array.
890 * To do this we raise a 'barrier'.
891 * The 'barrier' is a counter that can be raised multiple times
892 * to count how many activities are happening which preclude
894 * We can only raise the barrier if there is no pending IO.
895 * i.e. if nr_pending == 0.
896 * We choose only to raise the barrier if no-one is waiting for the
897 * barrier to go down. This means that as soon as an IO request
898 * is ready, no other operations which require a barrier will start
899 * until the IO request has had a chance.
901 * So: regular IO calls 'wait_barrier'. When that returns there
902 * is no backgroup IO happening, It must arrange to call
903 * allow_barrier when it has finished its IO.
904 * backgroup IO calls must call raise_barrier. Once that returns
905 * there is no normal IO happeing. It must arrange to call
906 * lower_barrier when the particular background IO completes.
909 static void raise_barrier(struct r10conf
*conf
, int force
)
911 BUG_ON(force
&& !conf
->barrier
);
912 spin_lock_irq(&conf
->resync_lock
);
914 /* Wait until no block IO is waiting (unless 'force') */
915 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
916 conf
->resync_lock
, );
918 /* block any new IO from starting */
921 /* Now wait for all pending IO to complete */
922 wait_event_lock_irq(conf
->wait_barrier
,
923 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
924 conf
->resync_lock
, );
926 spin_unlock_irq(&conf
->resync_lock
);
929 static void lower_barrier(struct r10conf
*conf
)
932 spin_lock_irqsave(&conf
->resync_lock
, flags
);
934 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
935 wake_up(&conf
->wait_barrier
);
938 static void wait_barrier(struct r10conf
*conf
)
940 spin_lock_irq(&conf
->resync_lock
);
943 /* Wait for the barrier to drop.
944 * However if there are already pending
945 * requests (preventing the barrier from
946 * rising completely), and the
947 * pre-process bio queue isn't empty,
948 * then don't wait, as we need to empty
949 * that queue to get the nr_pending
952 wait_event_lock_irq(conf
->wait_barrier
,
956 !bio_list_empty(current
->bio_list
)),
962 spin_unlock_irq(&conf
->resync_lock
);
965 static void allow_barrier(struct r10conf
*conf
)
968 spin_lock_irqsave(&conf
->resync_lock
, flags
);
970 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
971 wake_up(&conf
->wait_barrier
);
974 static void freeze_array(struct r10conf
*conf
)
976 /* stop syncio and normal IO and wait for everything to
978 * We increment barrier and nr_waiting, and then
979 * wait until nr_pending match nr_queued+1
980 * This is called in the context of one normal IO request
981 * that has failed. Thus any sync request that might be pending
982 * will be blocked by nr_pending, and we need to wait for
983 * pending IO requests to complete or be queued for re-try.
984 * Thus the number queued (nr_queued) plus this request (1)
985 * must match the number of pending IOs (nr_pending) before
988 spin_lock_irq(&conf
->resync_lock
);
991 wait_event_lock_irq(conf
->wait_barrier
,
992 conf
->nr_pending
== conf
->nr_queued
+1,
994 flush_pending_writes(conf
));
996 spin_unlock_irq(&conf
->resync_lock
);
999 static void unfreeze_array(struct r10conf
*conf
)
1001 /* reverse the effect of the freeze */
1002 spin_lock_irq(&conf
->resync_lock
);
1005 wake_up(&conf
->wait_barrier
);
1006 spin_unlock_irq(&conf
->resync_lock
);
1009 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1010 struct md_rdev
*rdev
)
1012 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1013 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1014 return rdev
->data_offset
;
1016 return rdev
->new_data_offset
;
1019 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1021 struct r10conf
*conf
= mddev
->private;
1022 struct r10bio
*r10_bio
;
1023 struct bio
*read_bio
;
1025 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1026 int chunk_sects
= chunk_mask
+ 1;
1027 const int rw
= bio_data_dir(bio
);
1028 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1029 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1030 unsigned long flags
;
1031 struct md_rdev
*blocked_rdev
;
1033 int sectors_handled
;
1036 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1037 md_flush_request(mddev
, bio
);
1041 /* If this request crosses a chunk boundary, we need to
1042 * split it. This will only happen for 1 PAGE (or less) requests.
1044 if (unlikely((bio
->bi_sector
& chunk_mask
) + (bio
->bi_size
>> 9)
1046 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1047 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1048 struct bio_pair
*bp
;
1049 /* Sanity check -- queue functions should prevent this happening */
1050 if (bio
->bi_vcnt
!= 1 ||
1053 /* This is a one page bio that upper layers
1054 * refuse to split for us, so we need to split it.
1057 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
1059 /* Each of these 'make_request' calls will call 'wait_barrier'.
1060 * If the first succeeds but the second blocks due to the resync
1061 * thread raising the barrier, we will deadlock because the
1062 * IO to the underlying device will be queued in generic_make_request
1063 * and will never complete, so will never reduce nr_pending.
1064 * So increment nr_waiting here so no new raise_barriers will
1065 * succeed, and so the second wait_barrier cannot block.
1067 spin_lock_irq(&conf
->resync_lock
);
1069 spin_unlock_irq(&conf
->resync_lock
);
1071 make_request(mddev
, &bp
->bio1
);
1072 make_request(mddev
, &bp
->bio2
);
1074 spin_lock_irq(&conf
->resync_lock
);
1076 wake_up(&conf
->wait_barrier
);
1077 spin_unlock_irq(&conf
->resync_lock
);
1079 bio_pair_release(bp
);
1082 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1083 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1084 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
1090 md_write_start(mddev
, bio
);
1093 * Register the new request and wait if the reconstruction
1094 * thread has put up a bar for new requests.
1095 * Continue immediately if no resync is active currently.
1099 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1101 r10_bio
->master_bio
= bio
;
1102 r10_bio
->sectors
= bio
->bi_size
>> 9;
1104 r10_bio
->mddev
= mddev
;
1105 r10_bio
->sector
= bio
->bi_sector
;
1108 /* We might need to issue multiple reads to different
1109 * devices if there are bad blocks around, so we keep
1110 * track of the number of reads in bio->bi_phys_segments.
1111 * If this is 0, there is only one r10_bio and no locking
1112 * will be needed when the request completes. If it is
1113 * non-zero, then it is the number of not-completed requests.
1115 bio
->bi_phys_segments
= 0;
1116 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1120 * read balancing logic:
1122 struct md_rdev
*rdev
;
1126 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1128 raid_end_bio_io(r10_bio
);
1131 slot
= r10_bio
->read_slot
;
1133 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1134 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1137 r10_bio
->devs
[slot
].bio
= read_bio
;
1138 r10_bio
->devs
[slot
].rdev
= rdev
;
1140 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1141 choose_data_offset(r10_bio
, rdev
);
1142 read_bio
->bi_bdev
= rdev
->bdev
;
1143 read_bio
->bi_end_io
= raid10_end_read_request
;
1144 read_bio
->bi_rw
= READ
| do_sync
;
1145 read_bio
->bi_private
= r10_bio
;
1147 if (max_sectors
< r10_bio
->sectors
) {
1148 /* Could not read all from this device, so we will
1149 * need another r10_bio.
1151 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1153 r10_bio
->sectors
= max_sectors
;
1154 spin_lock_irq(&conf
->device_lock
);
1155 if (bio
->bi_phys_segments
== 0)
1156 bio
->bi_phys_segments
= 2;
1158 bio
->bi_phys_segments
++;
1159 spin_unlock(&conf
->device_lock
);
1160 /* Cannot call generic_make_request directly
1161 * as that will be queued in __generic_make_request
1162 * and subsequent mempool_alloc might block
1163 * waiting for it. so hand bio over to raid10d.
1165 reschedule_retry(r10_bio
);
1167 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1169 r10_bio
->master_bio
= bio
;
1170 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
1173 r10_bio
->mddev
= mddev
;
1174 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1177 generic_make_request(read_bio
);
1184 if (conf
->pending_count
>= max_queued_requests
) {
1185 md_wakeup_thread(mddev
->thread
);
1186 wait_event(conf
->wait_barrier
,
1187 conf
->pending_count
< max_queued_requests
);
1189 /* first select target devices under rcu_lock and
1190 * inc refcount on their rdev. Record them by setting
1192 * If there are known/acknowledged bad blocks on any device
1193 * on which we have seen a write error, we want to avoid
1194 * writing to those blocks. This potentially requires several
1195 * writes to write around the bad blocks. Each set of writes
1196 * gets its own r10_bio with a set of bios attached. The number
1197 * of r10_bios is recored in bio->bi_phys_segments just as with
1200 plugged
= mddev_check_plugged(mddev
);
1202 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1203 raid10_find_phys(conf
, r10_bio
);
1205 blocked_rdev
= NULL
;
1207 max_sectors
= r10_bio
->sectors
;
1209 for (i
= 0; i
< conf
->copies
; i
++) {
1210 int d
= r10_bio
->devs
[i
].devnum
;
1211 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1212 struct md_rdev
*rrdev
= rcu_dereference(
1213 conf
->mirrors
[d
].replacement
);
1216 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1217 atomic_inc(&rdev
->nr_pending
);
1218 blocked_rdev
= rdev
;
1221 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1222 atomic_inc(&rrdev
->nr_pending
);
1223 blocked_rdev
= rrdev
;
1226 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1227 || test_bit(Unmerged
, &rrdev
->flags
)))
1230 r10_bio
->devs
[i
].bio
= NULL
;
1231 r10_bio
->devs
[i
].repl_bio
= NULL
;
1232 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
1233 test_bit(Unmerged
, &rdev
->flags
)) {
1234 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1237 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1239 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1243 is_bad
= is_badblock(rdev
, dev_sector
,
1245 &first_bad
, &bad_sectors
);
1247 /* Mustn't write here until the bad block
1250 atomic_inc(&rdev
->nr_pending
);
1251 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1252 blocked_rdev
= rdev
;
1255 if (is_bad
&& first_bad
<= dev_sector
) {
1256 /* Cannot write here at all */
1257 bad_sectors
-= (dev_sector
- first_bad
);
1258 if (bad_sectors
< max_sectors
)
1259 /* Mustn't write more than bad_sectors
1260 * to other devices yet
1262 max_sectors
= bad_sectors
;
1263 /* We don't set R10BIO_Degraded as that
1264 * only applies if the disk is missing,
1265 * so it might be re-added, and we want to
1266 * know to recover this chunk.
1267 * In this case the device is here, and the
1268 * fact that this chunk is not in-sync is
1269 * recorded in the bad block log.
1274 int good_sectors
= first_bad
- dev_sector
;
1275 if (good_sectors
< max_sectors
)
1276 max_sectors
= good_sectors
;
1279 r10_bio
->devs
[i
].bio
= bio
;
1280 atomic_inc(&rdev
->nr_pending
);
1282 r10_bio
->devs
[i
].repl_bio
= bio
;
1283 atomic_inc(&rrdev
->nr_pending
);
1288 if (unlikely(blocked_rdev
)) {
1289 /* Have to wait for this device to get unblocked, then retry */
1293 for (j
= 0; j
< i
; j
++) {
1294 if (r10_bio
->devs
[j
].bio
) {
1295 d
= r10_bio
->devs
[j
].devnum
;
1296 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1298 if (r10_bio
->devs
[j
].repl_bio
) {
1299 struct md_rdev
*rdev
;
1300 d
= r10_bio
->devs
[j
].devnum
;
1301 rdev
= conf
->mirrors
[d
].replacement
;
1303 /* Race with remove_disk */
1305 rdev
= conf
->mirrors
[d
].rdev
;
1307 rdev_dec_pending(rdev
, mddev
);
1310 allow_barrier(conf
);
1311 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1316 if (max_sectors
< r10_bio
->sectors
) {
1317 /* We are splitting this into multiple parts, so
1318 * we need to prepare for allocating another r10_bio.
1320 r10_bio
->sectors
= max_sectors
;
1321 spin_lock_irq(&conf
->device_lock
);
1322 if (bio
->bi_phys_segments
== 0)
1323 bio
->bi_phys_segments
= 2;
1325 bio
->bi_phys_segments
++;
1326 spin_unlock_irq(&conf
->device_lock
);
1328 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1330 atomic_set(&r10_bio
->remaining
, 1);
1331 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1333 for (i
= 0; i
< conf
->copies
; i
++) {
1335 int d
= r10_bio
->devs
[i
].devnum
;
1336 if (!r10_bio
->devs
[i
].bio
)
1339 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1340 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1342 r10_bio
->devs
[i
].bio
= mbio
;
1344 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1345 choose_data_offset(r10_bio
,
1346 conf
->mirrors
[d
].rdev
));
1347 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1348 mbio
->bi_end_io
= raid10_end_write_request
;
1349 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
1350 mbio
->bi_private
= r10_bio
;
1352 atomic_inc(&r10_bio
->remaining
);
1353 spin_lock_irqsave(&conf
->device_lock
, flags
);
1354 bio_list_add(&conf
->pending_bio_list
, mbio
);
1355 conf
->pending_count
++;
1356 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1358 if (!r10_bio
->devs
[i
].repl_bio
)
1361 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1362 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1364 r10_bio
->devs
[i
].repl_bio
= mbio
;
1366 /* We are actively writing to the original device
1367 * so it cannot disappear, so the replacement cannot
1370 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1373 conf
->mirrors
[d
].replacement
));
1374 mbio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
1375 mbio
->bi_end_io
= raid10_end_write_request
;
1376 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
1377 mbio
->bi_private
= r10_bio
;
1379 atomic_inc(&r10_bio
->remaining
);
1380 spin_lock_irqsave(&conf
->device_lock
, flags
);
1381 bio_list_add(&conf
->pending_bio_list
, mbio
);
1382 conf
->pending_count
++;
1383 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1386 /* Don't remove the bias on 'remaining' (one_write_done) until
1387 * after checking if we need to go around again.
1390 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1391 one_write_done(r10_bio
);
1392 /* We need another r10_bio. It has already been counted
1393 * in bio->bi_phys_segments.
1395 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1397 r10_bio
->master_bio
= bio
;
1398 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1400 r10_bio
->mddev
= mddev
;
1401 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1405 one_write_done(r10_bio
);
1407 /* In case raid10d snuck in to freeze_array */
1408 wake_up(&conf
->wait_barrier
);
1410 if (do_sync
|| !mddev
->bitmap
|| !plugged
)
1411 md_wakeup_thread(mddev
->thread
);
1414 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1416 struct r10conf
*conf
= mddev
->private;
1419 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1420 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1421 if (conf
->geo
.near_copies
> 1)
1422 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1423 if (conf
->geo
.far_copies
> 1) {
1424 if (conf
->geo
.far_offset
)
1425 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1427 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1429 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1430 conf
->geo
.raid_disks
- mddev
->degraded
);
1431 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1432 seq_printf(seq
, "%s",
1433 conf
->mirrors
[i
].rdev
&&
1434 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1435 seq_printf(seq
, "]");
1438 /* check if there are enough drives for
1439 * every block to appear on atleast one.
1440 * Don't consider the device numbered 'ignore'
1441 * as we might be about to remove it.
1443 static int _enough(struct r10conf
*conf
, struct geom
*geo
, int ignore
)
1448 int n
= conf
->copies
;
1451 if (conf
->mirrors
[first
].rdev
&&
1454 first
= (first
+1) % geo
->raid_disks
;
1458 } while (first
!= 0);
1462 static int enough(struct r10conf
*conf
, int ignore
)
1464 return _enough(conf
, &conf
->geo
, ignore
) &&
1465 _enough(conf
, &conf
->prev
, ignore
);
1468 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1470 char b
[BDEVNAME_SIZE
];
1471 struct r10conf
*conf
= mddev
->private;
1474 * If it is not operational, then we have already marked it as dead
1475 * else if it is the last working disks, ignore the error, let the
1476 * next level up know.
1477 * else mark the drive as failed
1479 if (test_bit(In_sync
, &rdev
->flags
)
1480 && !enough(conf
, rdev
->raid_disk
))
1482 * Don't fail the drive, just return an IO error.
1485 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1486 unsigned long flags
;
1487 spin_lock_irqsave(&conf
->device_lock
, flags
);
1489 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1491 * if recovery is running, make sure it aborts.
1493 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1495 set_bit(Blocked
, &rdev
->flags
);
1496 set_bit(Faulty
, &rdev
->flags
);
1497 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1499 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1500 "md/raid10:%s: Operation continuing on %d devices.\n",
1501 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1502 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1505 static void print_conf(struct r10conf
*conf
)
1508 struct mirror_info
*tmp
;
1510 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1512 printk(KERN_DEBUG
"(!conf)\n");
1515 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1516 conf
->geo
.raid_disks
);
1518 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1519 char b
[BDEVNAME_SIZE
];
1520 tmp
= conf
->mirrors
+ i
;
1522 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1523 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1524 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1525 bdevname(tmp
->rdev
->bdev
,b
));
1529 static void close_sync(struct r10conf
*conf
)
1532 allow_barrier(conf
);
1534 mempool_destroy(conf
->r10buf_pool
);
1535 conf
->r10buf_pool
= NULL
;
1538 static int raid10_spare_active(struct mddev
*mddev
)
1541 struct r10conf
*conf
= mddev
->private;
1542 struct mirror_info
*tmp
;
1544 unsigned long flags
;
1547 * Find all non-in_sync disks within the RAID10 configuration
1548 * and mark them in_sync
1550 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1551 tmp
= conf
->mirrors
+ i
;
1552 if (tmp
->replacement
1553 && tmp
->replacement
->recovery_offset
== MaxSector
1554 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1555 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1556 /* Replacement has just become active */
1558 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1561 /* Replaced device not technically faulty,
1562 * but we need to be sure it gets removed
1563 * and never re-added.
1565 set_bit(Faulty
, &tmp
->rdev
->flags
);
1566 sysfs_notify_dirent_safe(
1567 tmp
->rdev
->sysfs_state
);
1569 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1570 } else if (tmp
->rdev
1571 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1572 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1574 sysfs_notify_dirent(tmp
->rdev
->sysfs_state
);
1577 spin_lock_irqsave(&conf
->device_lock
, flags
);
1578 mddev
->degraded
-= count
;
1579 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1586 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1588 struct r10conf
*conf
= mddev
->private;
1592 int last
= conf
->geo
.raid_disks
- 1;
1593 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1595 if (mddev
->recovery_cp
< MaxSector
)
1596 /* only hot-add to in-sync arrays, as recovery is
1597 * very different from resync
1600 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, &conf
->prev
, -1))
1603 if (rdev
->raid_disk
>= 0)
1604 first
= last
= rdev
->raid_disk
;
1606 if (q
->merge_bvec_fn
) {
1607 set_bit(Unmerged
, &rdev
->flags
);
1608 mddev
->merge_check_needed
= 1;
1611 if (rdev
->saved_raid_disk
>= first
&&
1612 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1613 mirror
= rdev
->saved_raid_disk
;
1616 for ( ; mirror
<= last
; mirror
++) {
1617 struct mirror_info
*p
= &conf
->mirrors
[mirror
];
1618 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1621 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1622 p
->replacement
!= NULL
)
1624 clear_bit(In_sync
, &rdev
->flags
);
1625 set_bit(Replacement
, &rdev
->flags
);
1626 rdev
->raid_disk
= mirror
;
1628 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1629 rdev
->data_offset
<< 9);
1631 rcu_assign_pointer(p
->replacement
, rdev
);
1635 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1636 rdev
->data_offset
<< 9);
1638 p
->head_position
= 0;
1639 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1640 rdev
->raid_disk
= mirror
;
1642 if (rdev
->saved_raid_disk
!= mirror
)
1644 rcu_assign_pointer(p
->rdev
, rdev
);
1647 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1648 /* Some requests might not have seen this new
1649 * merge_bvec_fn. We must wait for them to complete
1650 * before merging the device fully.
1651 * First we make sure any code which has tested
1652 * our function has submitted the request, then
1653 * we wait for all outstanding requests to complete.
1655 synchronize_sched();
1656 raise_barrier(conf
, 0);
1657 lower_barrier(conf
);
1658 clear_bit(Unmerged
, &rdev
->flags
);
1660 md_integrity_add_rdev(rdev
, mddev
);
1665 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1667 struct r10conf
*conf
= mddev
->private;
1669 int number
= rdev
->raid_disk
;
1670 struct md_rdev
**rdevp
;
1671 struct mirror_info
*p
= conf
->mirrors
+ number
;
1674 if (rdev
== p
->rdev
)
1676 else if (rdev
== p
->replacement
)
1677 rdevp
= &p
->replacement
;
1681 if (test_bit(In_sync
, &rdev
->flags
) ||
1682 atomic_read(&rdev
->nr_pending
)) {
1686 /* Only remove faulty devices if recovery
1689 if (!test_bit(Faulty
, &rdev
->flags
) &&
1690 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1691 (!p
->replacement
|| p
->replacement
== rdev
) &&
1698 if (atomic_read(&rdev
->nr_pending
)) {
1699 /* lost the race, try later */
1703 } else if (p
->replacement
) {
1704 /* We must have just cleared 'rdev' */
1705 p
->rdev
= p
->replacement
;
1706 clear_bit(Replacement
, &p
->replacement
->flags
);
1707 smp_mb(); /* Make sure other CPUs may see both as identical
1708 * but will never see neither -- if they are careful.
1710 p
->replacement
= NULL
;
1711 clear_bit(WantReplacement
, &rdev
->flags
);
1713 /* We might have just remove the Replacement as faulty
1714 * Clear the flag just in case
1716 clear_bit(WantReplacement
, &rdev
->flags
);
1718 err
= md_integrity_register(mddev
);
1727 static void end_sync_read(struct bio
*bio
, int error
)
1729 struct r10bio
*r10_bio
= bio
->bi_private
;
1730 struct r10conf
*conf
= r10_bio
->mddev
->private;
1733 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1735 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1736 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1738 /* The write handler will notice the lack of
1739 * R10BIO_Uptodate and record any errors etc
1741 atomic_add(r10_bio
->sectors
,
1742 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1744 /* for reconstruct, we always reschedule after a read.
1745 * for resync, only after all reads
1747 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1748 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1749 atomic_dec_and_test(&r10_bio
->remaining
)) {
1750 /* we have read all the blocks,
1751 * do the comparison in process context in raid10d
1753 reschedule_retry(r10_bio
);
1757 static void end_sync_request(struct r10bio
*r10_bio
)
1759 struct mddev
*mddev
= r10_bio
->mddev
;
1761 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1762 if (r10_bio
->master_bio
== NULL
) {
1763 /* the primary of several recovery bios */
1764 sector_t s
= r10_bio
->sectors
;
1765 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1766 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1767 reschedule_retry(r10_bio
);
1770 md_done_sync(mddev
, s
, 1);
1773 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1774 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1775 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1776 reschedule_retry(r10_bio
);
1784 static void end_sync_write(struct bio
*bio
, int error
)
1786 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1787 struct r10bio
*r10_bio
= bio
->bi_private
;
1788 struct mddev
*mddev
= r10_bio
->mddev
;
1789 struct r10conf
*conf
= mddev
->private;
1795 struct md_rdev
*rdev
= NULL
;
1797 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1799 rdev
= conf
->mirrors
[d
].replacement
;
1801 rdev
= conf
->mirrors
[d
].rdev
;
1805 md_error(mddev
, rdev
);
1807 set_bit(WriteErrorSeen
, &rdev
->flags
);
1808 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1809 set_bit(MD_RECOVERY_NEEDED
,
1810 &rdev
->mddev
->recovery
);
1811 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1813 } else if (is_badblock(rdev
,
1814 r10_bio
->devs
[slot
].addr
,
1816 &first_bad
, &bad_sectors
))
1817 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1819 rdev_dec_pending(rdev
, mddev
);
1821 end_sync_request(r10_bio
);
1825 * Note: sync and recover and handled very differently for raid10
1826 * This code is for resync.
1827 * For resync, we read through virtual addresses and read all blocks.
1828 * If there is any error, we schedule a write. The lowest numbered
1829 * drive is authoritative.
1830 * However requests come for physical address, so we need to map.
1831 * For every physical address there are raid_disks/copies virtual addresses,
1832 * which is always are least one, but is not necessarly an integer.
1833 * This means that a physical address can span multiple chunks, so we may
1834 * have to submit multiple io requests for a single sync request.
1837 * We check if all blocks are in-sync and only write to blocks that
1840 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1842 struct r10conf
*conf
= mddev
->private;
1844 struct bio
*tbio
, *fbio
;
1847 atomic_set(&r10_bio
->remaining
, 1);
1849 /* find the first device with a block */
1850 for (i
=0; i
<conf
->copies
; i
++)
1851 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1854 if (i
== conf
->copies
)
1858 fbio
= r10_bio
->devs
[i
].bio
;
1860 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
1861 /* now find blocks with errors */
1862 for (i
=0 ; i
< conf
->copies
; i
++) {
1865 tbio
= r10_bio
->devs
[i
].bio
;
1867 if (tbio
->bi_end_io
!= end_sync_read
)
1871 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1872 /* We know that the bi_io_vec layout is the same for
1873 * both 'first' and 'i', so we just compare them.
1874 * All vec entries are PAGE_SIZE;
1876 for (j
= 0; j
< vcnt
; j
++)
1877 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1878 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1879 fbio
->bi_io_vec
[j
].bv_len
))
1883 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1884 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1885 /* Don't fix anything. */
1888 /* Ok, we need to write this bio, either to correct an
1889 * inconsistency or to correct an unreadable block.
1890 * First we need to fixup bv_offset, bv_len and
1891 * bi_vecs, as the read request might have corrupted these
1893 tbio
->bi_vcnt
= vcnt
;
1894 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1896 tbio
->bi_phys_segments
= 0;
1897 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1898 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1899 tbio
->bi_next
= NULL
;
1900 tbio
->bi_rw
= WRITE
;
1901 tbio
->bi_private
= r10_bio
;
1902 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1904 for (j
=0; j
< vcnt
; j
++) {
1905 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1906 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1908 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1909 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1912 tbio
->bi_end_io
= end_sync_write
;
1914 d
= r10_bio
->devs
[i
].devnum
;
1915 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1916 atomic_inc(&r10_bio
->remaining
);
1917 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1919 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1920 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1921 generic_make_request(tbio
);
1924 /* Now write out to any replacement devices
1927 for (i
= 0; i
< conf
->copies
; i
++) {
1930 tbio
= r10_bio
->devs
[i
].repl_bio
;
1931 if (!tbio
|| !tbio
->bi_end_io
)
1933 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
1934 && r10_bio
->devs
[i
].bio
!= fbio
)
1935 for (j
= 0; j
< vcnt
; j
++)
1936 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1937 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1939 d
= r10_bio
->devs
[i
].devnum
;
1940 atomic_inc(&r10_bio
->remaining
);
1941 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
1942 tbio
->bi_size
>> 9);
1943 generic_make_request(tbio
);
1947 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1948 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1954 * Now for the recovery code.
1955 * Recovery happens across physical sectors.
1956 * We recover all non-is_sync drives by finding the virtual address of
1957 * each, and then choose a working drive that also has that virt address.
1958 * There is a separate r10_bio for each non-in_sync drive.
1959 * Only the first two slots are in use. The first for reading,
1960 * The second for writing.
1963 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
1965 /* We got a read error during recovery.
1966 * We repeat the read in smaller page-sized sections.
1967 * If a read succeeds, write it to the new device or record
1968 * a bad block if we cannot.
1969 * If a read fails, record a bad block on both old and
1972 struct mddev
*mddev
= r10_bio
->mddev
;
1973 struct r10conf
*conf
= mddev
->private;
1974 struct bio
*bio
= r10_bio
->devs
[0].bio
;
1976 int sectors
= r10_bio
->sectors
;
1978 int dr
= r10_bio
->devs
[0].devnum
;
1979 int dw
= r10_bio
->devs
[1].devnum
;
1983 struct md_rdev
*rdev
;
1987 if (s
> (PAGE_SIZE
>>9))
1990 rdev
= conf
->mirrors
[dr
].rdev
;
1991 addr
= r10_bio
->devs
[0].addr
+ sect
,
1992 ok
= sync_page_io(rdev
,
1995 bio
->bi_io_vec
[idx
].bv_page
,
1998 rdev
= conf
->mirrors
[dw
].rdev
;
1999 addr
= r10_bio
->devs
[1].addr
+ sect
;
2000 ok
= sync_page_io(rdev
,
2003 bio
->bi_io_vec
[idx
].bv_page
,
2006 set_bit(WriteErrorSeen
, &rdev
->flags
);
2007 if (!test_and_set_bit(WantReplacement
,
2009 set_bit(MD_RECOVERY_NEEDED
,
2010 &rdev
->mddev
->recovery
);
2014 /* We don't worry if we cannot set a bad block -
2015 * it really is bad so there is no loss in not
2018 rdev_set_badblocks(rdev
, addr
, s
, 0);
2020 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2021 /* need bad block on destination too */
2022 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2023 addr
= r10_bio
->devs
[1].addr
+ sect
;
2024 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2026 /* just abort the recovery */
2028 "md/raid10:%s: recovery aborted"
2029 " due to read error\n",
2032 conf
->mirrors
[dw
].recovery_disabled
2033 = mddev
->recovery_disabled
;
2034 set_bit(MD_RECOVERY_INTR
,
2047 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2049 struct r10conf
*conf
= mddev
->private;
2051 struct bio
*wbio
, *wbio2
;
2053 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2054 fix_recovery_read_error(r10_bio
);
2055 end_sync_request(r10_bio
);
2060 * share the pages with the first bio
2061 * and submit the write request
2063 d
= r10_bio
->devs
[1].devnum
;
2064 wbio
= r10_bio
->devs
[1].bio
;
2065 wbio2
= r10_bio
->devs
[1].repl_bio
;
2066 if (wbio
->bi_end_io
) {
2067 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2068 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
2069 generic_make_request(wbio
);
2071 if (wbio2
&& wbio2
->bi_end_io
) {
2072 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2073 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2074 wbio2
->bi_size
>> 9);
2075 generic_make_request(wbio2
);
2081 * Used by fix_read_error() to decay the per rdev read_errors.
2082 * We halve the read error count for every hour that has elapsed
2083 * since the last recorded read error.
2086 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2088 struct timespec cur_time_mon
;
2089 unsigned long hours_since_last
;
2090 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2092 ktime_get_ts(&cur_time_mon
);
2094 if (rdev
->last_read_error
.tv_sec
== 0 &&
2095 rdev
->last_read_error
.tv_nsec
== 0) {
2096 /* first time we've seen a read error */
2097 rdev
->last_read_error
= cur_time_mon
;
2101 hours_since_last
= (cur_time_mon
.tv_sec
-
2102 rdev
->last_read_error
.tv_sec
) / 3600;
2104 rdev
->last_read_error
= cur_time_mon
;
2107 * if hours_since_last is > the number of bits in read_errors
2108 * just set read errors to 0. We do this to avoid
2109 * overflowing the shift of read_errors by hours_since_last.
2111 if (hours_since_last
>= 8 * sizeof(read_errors
))
2112 atomic_set(&rdev
->read_errors
, 0);
2114 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2117 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2118 int sectors
, struct page
*page
, int rw
)
2123 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2124 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2126 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2130 set_bit(WriteErrorSeen
, &rdev
->flags
);
2131 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2132 set_bit(MD_RECOVERY_NEEDED
,
2133 &rdev
->mddev
->recovery
);
2135 /* need to record an error - either for the block or the device */
2136 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2137 md_error(rdev
->mddev
, rdev
);
2142 * This is a kernel thread which:
2144 * 1. Retries failed read operations on working mirrors.
2145 * 2. Updates the raid superblock when problems encounter.
2146 * 3. Performs writes following reads for array synchronising.
2149 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2151 int sect
= 0; /* Offset from r10_bio->sector */
2152 int sectors
= r10_bio
->sectors
;
2153 struct md_rdev
*rdev
;
2154 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2155 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2157 /* still own a reference to this rdev, so it cannot
2158 * have been cleared recently.
2160 rdev
= conf
->mirrors
[d
].rdev
;
2162 if (test_bit(Faulty
, &rdev
->flags
))
2163 /* drive has already been failed, just ignore any
2164 more fix_read_error() attempts */
2167 check_decay_read_errors(mddev
, rdev
);
2168 atomic_inc(&rdev
->read_errors
);
2169 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2170 char b
[BDEVNAME_SIZE
];
2171 bdevname(rdev
->bdev
, b
);
2174 "md/raid10:%s: %s: Raid device exceeded "
2175 "read_error threshold [cur %d:max %d]\n",
2177 atomic_read(&rdev
->read_errors
), max_read_errors
);
2179 "md/raid10:%s: %s: Failing raid device\n",
2181 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2182 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2188 int sl
= r10_bio
->read_slot
;
2192 if (s
> (PAGE_SIZE
>>9))
2200 d
= r10_bio
->devs
[sl
].devnum
;
2201 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2203 !test_bit(Unmerged
, &rdev
->flags
) &&
2204 test_bit(In_sync
, &rdev
->flags
) &&
2205 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2206 &first_bad
, &bad_sectors
) == 0) {
2207 atomic_inc(&rdev
->nr_pending
);
2209 success
= sync_page_io(rdev
,
2210 r10_bio
->devs
[sl
].addr
+
2213 conf
->tmppage
, READ
, false);
2214 rdev_dec_pending(rdev
, mddev
);
2220 if (sl
== conf
->copies
)
2222 } while (!success
&& sl
!= r10_bio
->read_slot
);
2226 /* Cannot read from anywhere, just mark the block
2227 * as bad on the first device to discourage future
2230 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2231 rdev
= conf
->mirrors
[dn
].rdev
;
2233 if (!rdev_set_badblocks(
2235 r10_bio
->devs
[r10_bio
->read_slot
].addr
2238 md_error(mddev
, rdev
);
2239 r10_bio
->devs
[r10_bio
->read_slot
].bio
2246 /* write it back and re-read */
2248 while (sl
!= r10_bio
->read_slot
) {
2249 char b
[BDEVNAME_SIZE
];
2254 d
= r10_bio
->devs
[sl
].devnum
;
2255 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2257 test_bit(Unmerged
, &rdev
->flags
) ||
2258 !test_bit(In_sync
, &rdev
->flags
))
2261 atomic_inc(&rdev
->nr_pending
);
2263 if (r10_sync_page_io(rdev
,
2264 r10_bio
->devs
[sl
].addr
+
2266 s
<<9, conf
->tmppage
, WRITE
)
2268 /* Well, this device is dead */
2270 "md/raid10:%s: read correction "
2272 " (%d sectors at %llu on %s)\n",
2274 (unsigned long long)(
2276 choose_data_offset(r10_bio
,
2278 bdevname(rdev
->bdev
, b
));
2279 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2282 bdevname(rdev
->bdev
, b
));
2284 rdev_dec_pending(rdev
, mddev
);
2288 while (sl
!= r10_bio
->read_slot
) {
2289 char b
[BDEVNAME_SIZE
];
2294 d
= r10_bio
->devs
[sl
].devnum
;
2295 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2297 !test_bit(In_sync
, &rdev
->flags
))
2300 atomic_inc(&rdev
->nr_pending
);
2302 switch (r10_sync_page_io(rdev
,
2303 r10_bio
->devs
[sl
].addr
+
2305 s
<<9, conf
->tmppage
,
2308 /* Well, this device is dead */
2310 "md/raid10:%s: unable to read back "
2312 " (%d sectors at %llu on %s)\n",
2314 (unsigned long long)(
2316 choose_data_offset(r10_bio
, rdev
)),
2317 bdevname(rdev
->bdev
, b
));
2318 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2321 bdevname(rdev
->bdev
, b
));
2325 "md/raid10:%s: read error corrected"
2326 " (%d sectors at %llu on %s)\n",
2328 (unsigned long long)(
2330 choose_data_offset(r10_bio
, rdev
)),
2331 bdevname(rdev
->bdev
, b
));
2332 atomic_add(s
, &rdev
->corrected_errors
);
2335 rdev_dec_pending(rdev
, mddev
);
2345 static void bi_complete(struct bio
*bio
, int error
)
2347 complete((struct completion
*)bio
->bi_private
);
2350 static int submit_bio_wait(int rw
, struct bio
*bio
)
2352 struct completion event
;
2355 init_completion(&event
);
2356 bio
->bi_private
= &event
;
2357 bio
->bi_end_io
= bi_complete
;
2358 submit_bio(rw
, bio
);
2359 wait_for_completion(&event
);
2361 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2364 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2366 struct bio
*bio
= r10_bio
->master_bio
;
2367 struct mddev
*mddev
= r10_bio
->mddev
;
2368 struct r10conf
*conf
= mddev
->private;
2369 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2370 /* bio has the data to be written to slot 'i' where
2371 * we just recently had a write error.
2372 * We repeatedly clone the bio and trim down to one block,
2373 * then try the write. Where the write fails we record
2375 * It is conceivable that the bio doesn't exactly align with
2376 * blocks. We must handle this.
2378 * We currently own a reference to the rdev.
2384 int sect_to_write
= r10_bio
->sectors
;
2387 if (rdev
->badblocks
.shift
< 0)
2390 block_sectors
= 1 << rdev
->badblocks
.shift
;
2391 sector
= r10_bio
->sector
;
2392 sectors
= ((r10_bio
->sector
+ block_sectors
)
2393 & ~(sector_t
)(block_sectors
- 1))
2396 while (sect_to_write
) {
2398 if (sectors
> sect_to_write
)
2399 sectors
= sect_to_write
;
2400 /* Write at 'sector' for 'sectors' */
2401 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2402 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2403 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2404 choose_data_offset(r10_bio
, rdev
) +
2405 (sector
- r10_bio
->sector
));
2406 wbio
->bi_bdev
= rdev
->bdev
;
2407 if (submit_bio_wait(WRITE
, wbio
) == 0)
2409 ok
= rdev_set_badblocks(rdev
, sector
,
2414 sect_to_write
-= sectors
;
2416 sectors
= block_sectors
;
2421 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2423 int slot
= r10_bio
->read_slot
;
2425 struct r10conf
*conf
= mddev
->private;
2426 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2427 char b
[BDEVNAME_SIZE
];
2428 unsigned long do_sync
;
2431 /* we got a read error. Maybe the drive is bad. Maybe just
2432 * the block and we can fix it.
2433 * We freeze all other IO, and try reading the block from
2434 * other devices. When we find one, we re-write
2435 * and check it that fixes the read error.
2436 * This is all done synchronously while the array is
2439 bio
= r10_bio
->devs
[slot
].bio
;
2440 bdevname(bio
->bi_bdev
, b
);
2442 r10_bio
->devs
[slot
].bio
= NULL
;
2444 if (mddev
->ro
== 0) {
2446 fix_read_error(conf
, mddev
, r10_bio
);
2447 unfreeze_array(conf
);
2449 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2451 rdev_dec_pending(rdev
, mddev
);
2454 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2456 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2457 " read error for block %llu\n",
2459 (unsigned long long)r10_bio
->sector
);
2460 raid_end_bio_io(r10_bio
);
2464 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2465 slot
= r10_bio
->read_slot
;
2468 "md/raid10:%s: %s: redirecting"
2469 "sector %llu to another mirror\n",
2471 bdevname(rdev
->bdev
, b
),
2472 (unsigned long long)r10_bio
->sector
);
2473 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2476 r10_bio
->sector
- bio
->bi_sector
,
2478 r10_bio
->devs
[slot
].bio
= bio
;
2479 r10_bio
->devs
[slot
].rdev
= rdev
;
2480 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2481 + choose_data_offset(r10_bio
, rdev
);
2482 bio
->bi_bdev
= rdev
->bdev
;
2483 bio
->bi_rw
= READ
| do_sync
;
2484 bio
->bi_private
= r10_bio
;
2485 bio
->bi_end_io
= raid10_end_read_request
;
2486 if (max_sectors
< r10_bio
->sectors
) {
2487 /* Drat - have to split this up more */
2488 struct bio
*mbio
= r10_bio
->master_bio
;
2489 int sectors_handled
=
2490 r10_bio
->sector
+ max_sectors
2492 r10_bio
->sectors
= max_sectors
;
2493 spin_lock_irq(&conf
->device_lock
);
2494 if (mbio
->bi_phys_segments
== 0)
2495 mbio
->bi_phys_segments
= 2;
2497 mbio
->bi_phys_segments
++;
2498 spin_unlock_irq(&conf
->device_lock
);
2499 generic_make_request(bio
);
2501 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2503 r10_bio
->master_bio
= mbio
;
2504 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2507 set_bit(R10BIO_ReadError
,
2509 r10_bio
->mddev
= mddev
;
2510 r10_bio
->sector
= mbio
->bi_sector
2515 generic_make_request(bio
);
2518 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2520 /* Some sort of write request has finished and it
2521 * succeeded in writing where we thought there was a
2522 * bad block. So forget the bad block.
2523 * Or possibly if failed and we need to record
2527 struct md_rdev
*rdev
;
2529 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2530 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2531 for (m
= 0; m
< conf
->copies
; m
++) {
2532 int dev
= r10_bio
->devs
[m
].devnum
;
2533 rdev
= conf
->mirrors
[dev
].rdev
;
2534 if (r10_bio
->devs
[m
].bio
== NULL
)
2536 if (test_bit(BIO_UPTODATE
,
2537 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2538 rdev_clear_badblocks(
2540 r10_bio
->devs
[m
].addr
,
2541 r10_bio
->sectors
, 0);
2543 if (!rdev_set_badblocks(
2545 r10_bio
->devs
[m
].addr
,
2546 r10_bio
->sectors
, 0))
2547 md_error(conf
->mddev
, rdev
);
2549 rdev
= conf
->mirrors
[dev
].replacement
;
2550 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2552 if (test_bit(BIO_UPTODATE
,
2553 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2554 rdev_clear_badblocks(
2556 r10_bio
->devs
[m
].addr
,
2557 r10_bio
->sectors
, 0);
2559 if (!rdev_set_badblocks(
2561 r10_bio
->devs
[m
].addr
,
2562 r10_bio
->sectors
, 0))
2563 md_error(conf
->mddev
, rdev
);
2568 for (m
= 0; m
< conf
->copies
; m
++) {
2569 int dev
= r10_bio
->devs
[m
].devnum
;
2570 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2571 rdev
= conf
->mirrors
[dev
].rdev
;
2572 if (bio
== IO_MADE_GOOD
) {
2573 rdev_clear_badblocks(
2575 r10_bio
->devs
[m
].addr
,
2576 r10_bio
->sectors
, 0);
2577 rdev_dec_pending(rdev
, conf
->mddev
);
2578 } else if (bio
!= NULL
&&
2579 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2580 if (!narrow_write_error(r10_bio
, m
)) {
2581 md_error(conf
->mddev
, rdev
);
2582 set_bit(R10BIO_Degraded
,
2585 rdev_dec_pending(rdev
, conf
->mddev
);
2587 bio
= r10_bio
->devs
[m
].repl_bio
;
2588 rdev
= conf
->mirrors
[dev
].replacement
;
2589 if (rdev
&& bio
== IO_MADE_GOOD
) {
2590 rdev_clear_badblocks(
2592 r10_bio
->devs
[m
].addr
,
2593 r10_bio
->sectors
, 0);
2594 rdev_dec_pending(rdev
, conf
->mddev
);
2597 if (test_bit(R10BIO_WriteError
,
2599 close_write(r10_bio
);
2600 raid_end_bio_io(r10_bio
);
2604 static void raid10d(struct mddev
*mddev
)
2606 struct r10bio
*r10_bio
;
2607 unsigned long flags
;
2608 struct r10conf
*conf
= mddev
->private;
2609 struct list_head
*head
= &conf
->retry_list
;
2610 struct blk_plug plug
;
2612 md_check_recovery(mddev
);
2614 blk_start_plug(&plug
);
2617 flush_pending_writes(conf
);
2619 spin_lock_irqsave(&conf
->device_lock
, flags
);
2620 if (list_empty(head
)) {
2621 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2624 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2625 list_del(head
->prev
);
2627 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2629 mddev
= r10_bio
->mddev
;
2630 conf
= mddev
->private;
2631 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2632 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2633 handle_write_completed(conf
, r10_bio
);
2634 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2635 sync_request_write(mddev
, r10_bio
);
2636 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2637 recovery_request_write(mddev
, r10_bio
);
2638 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2639 handle_read_error(mddev
, r10_bio
);
2641 /* just a partial read to be scheduled from a
2644 int slot
= r10_bio
->read_slot
;
2645 generic_make_request(r10_bio
->devs
[slot
].bio
);
2649 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2650 md_check_recovery(mddev
);
2652 blk_finish_plug(&plug
);
2656 static int init_resync(struct r10conf
*conf
)
2661 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2662 BUG_ON(conf
->r10buf_pool
);
2663 conf
->have_replacement
= 0;
2664 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2665 if (conf
->mirrors
[i
].replacement
)
2666 conf
->have_replacement
= 1;
2667 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2668 if (!conf
->r10buf_pool
)
2670 conf
->next_resync
= 0;
2675 * perform a "sync" on one "block"
2677 * We need to make sure that no normal I/O request - particularly write
2678 * requests - conflict with active sync requests.
2680 * This is achieved by tracking pending requests and a 'barrier' concept
2681 * that can be installed to exclude normal IO requests.
2683 * Resync and recovery are handled very differently.
2684 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2686 * For resync, we iterate over virtual addresses, read all copies,
2687 * and update if there are differences. If only one copy is live,
2689 * For recovery, we iterate over physical addresses, read a good
2690 * value for each non-in_sync drive, and over-write.
2692 * So, for recovery we may have several outstanding complex requests for a
2693 * given address, one for each out-of-sync device. We model this by allocating
2694 * a number of r10_bio structures, one for each out-of-sync device.
2695 * As we setup these structures, we collect all bio's together into a list
2696 * which we then process collectively to add pages, and then process again
2697 * to pass to generic_make_request.
2699 * The r10_bio structures are linked using a borrowed master_bio pointer.
2700 * This link is counted in ->remaining. When the r10_bio that points to NULL
2701 * has its remaining count decremented to 0, the whole complex operation
2706 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2707 int *skipped
, int go_faster
)
2709 struct r10conf
*conf
= mddev
->private;
2710 struct r10bio
*r10_bio
;
2711 struct bio
*biolist
= NULL
, *bio
;
2712 sector_t max_sector
, nr_sectors
;
2715 sector_t sync_blocks
;
2716 sector_t sectors_skipped
= 0;
2717 int chunks_skipped
= 0;
2718 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2720 if (!conf
->r10buf_pool
)
2721 if (init_resync(conf
))
2725 max_sector
= mddev
->dev_sectors
;
2726 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2727 max_sector
= mddev
->resync_max_sectors
;
2728 if (sector_nr
>= max_sector
) {
2729 /* If we aborted, we need to abort the
2730 * sync on the 'current' bitmap chucks (there can
2731 * be several when recovering multiple devices).
2732 * as we may have started syncing it but not finished.
2733 * We can find the current address in
2734 * mddev->curr_resync, but for recovery,
2735 * we need to convert that to several
2736 * virtual addresses.
2738 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2739 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2740 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2742 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2744 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2745 bitmap_end_sync(mddev
->bitmap
, sect
,
2749 /* completed sync */
2750 if ((!mddev
->bitmap
|| conf
->fullsync
)
2751 && conf
->have_replacement
2752 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2753 /* Completed a full sync so the replacements
2754 * are now fully recovered.
2756 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2757 if (conf
->mirrors
[i
].replacement
)
2758 conf
->mirrors
[i
].replacement
2764 bitmap_close_sync(mddev
->bitmap
);
2767 return sectors_skipped
;
2769 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2770 /* if there has been nothing to do on any drive,
2771 * then there is nothing to do at all..
2774 return (max_sector
- sector_nr
) + sectors_skipped
;
2777 if (max_sector
> mddev
->resync_max
)
2778 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2780 /* make sure whole request will fit in a chunk - if chunks
2783 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2784 max_sector
> (sector_nr
| chunk_mask
))
2785 max_sector
= (sector_nr
| chunk_mask
) + 1;
2787 * If there is non-resync activity waiting for us then
2788 * put in a delay to throttle resync.
2790 if (!go_faster
&& conf
->nr_waiting
)
2791 msleep_interruptible(1000);
2793 /* Again, very different code for resync and recovery.
2794 * Both must result in an r10bio with a list of bios that
2795 * have bi_end_io, bi_sector, bi_bdev set,
2796 * and bi_private set to the r10bio.
2797 * For recovery, we may actually create several r10bios
2798 * with 2 bios in each, that correspond to the bios in the main one.
2799 * In this case, the subordinate r10bios link back through a
2800 * borrowed master_bio pointer, and the counter in the master
2801 * includes a ref from each subordinate.
2803 /* First, we decide what to do and set ->bi_end_io
2804 * To end_sync_read if we want to read, and
2805 * end_sync_write if we will want to write.
2808 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2809 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2810 /* recovery... the complicated one */
2814 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2820 struct mirror_info
*mirror
= &conf
->mirrors
[i
];
2822 if ((mirror
->rdev
== NULL
||
2823 test_bit(In_sync
, &mirror
->rdev
->flags
))
2825 (mirror
->replacement
== NULL
||
2827 &mirror
->replacement
->flags
)))
2831 /* want to reconstruct this device */
2833 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2834 /* Unless we are doing a full sync, or a replacement
2835 * we only need to recover the block if it is set in
2838 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2840 if (sync_blocks
< max_sync
)
2841 max_sync
= sync_blocks
;
2843 mirror
->replacement
== NULL
&&
2845 /* yep, skip the sync_blocks here, but don't assume
2846 * that there will never be anything to do here
2848 chunks_skipped
= -1;
2852 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2853 raise_barrier(conf
, rb2
!= NULL
);
2854 atomic_set(&r10_bio
->remaining
, 0);
2856 r10_bio
->master_bio
= (struct bio
*)rb2
;
2858 atomic_inc(&rb2
->remaining
);
2859 r10_bio
->mddev
= mddev
;
2860 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
2861 r10_bio
->sector
= sect
;
2863 raid10_find_phys(conf
, r10_bio
);
2865 /* Need to check if the array will still be
2868 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
2869 if (conf
->mirrors
[j
].rdev
== NULL
||
2870 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
2875 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2876 &sync_blocks
, still_degraded
);
2879 for (j
=0; j
<conf
->copies
;j
++) {
2881 int d
= r10_bio
->devs
[j
].devnum
;
2882 sector_t from_addr
, to_addr
;
2883 struct md_rdev
*rdev
;
2884 sector_t sector
, first_bad
;
2886 if (!conf
->mirrors
[d
].rdev
||
2887 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
2889 /* This is where we read from */
2891 rdev
= conf
->mirrors
[d
].rdev
;
2892 sector
= r10_bio
->devs
[j
].addr
;
2894 if (is_badblock(rdev
, sector
, max_sync
,
2895 &first_bad
, &bad_sectors
)) {
2896 if (first_bad
> sector
)
2897 max_sync
= first_bad
- sector
;
2899 bad_sectors
-= (sector
2901 if (max_sync
> bad_sectors
)
2902 max_sync
= bad_sectors
;
2906 bio
= r10_bio
->devs
[0].bio
;
2907 bio
->bi_next
= biolist
;
2909 bio
->bi_private
= r10_bio
;
2910 bio
->bi_end_io
= end_sync_read
;
2912 from_addr
= r10_bio
->devs
[j
].addr
;
2913 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
2914 bio
->bi_bdev
= rdev
->bdev
;
2915 atomic_inc(&rdev
->nr_pending
);
2916 /* and we write to 'i' (if not in_sync) */
2918 for (k
=0; k
<conf
->copies
; k
++)
2919 if (r10_bio
->devs
[k
].devnum
== i
)
2921 BUG_ON(k
== conf
->copies
);
2922 to_addr
= r10_bio
->devs
[k
].addr
;
2923 r10_bio
->devs
[0].devnum
= d
;
2924 r10_bio
->devs
[0].addr
= from_addr
;
2925 r10_bio
->devs
[1].devnum
= i
;
2926 r10_bio
->devs
[1].addr
= to_addr
;
2928 rdev
= mirror
->rdev
;
2929 if (!test_bit(In_sync
, &rdev
->flags
)) {
2930 bio
= r10_bio
->devs
[1].bio
;
2931 bio
->bi_next
= biolist
;
2933 bio
->bi_private
= r10_bio
;
2934 bio
->bi_end_io
= end_sync_write
;
2936 bio
->bi_sector
= to_addr
2937 + rdev
->data_offset
;
2938 bio
->bi_bdev
= rdev
->bdev
;
2939 atomic_inc(&r10_bio
->remaining
);
2941 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
2943 /* and maybe write to replacement */
2944 bio
= r10_bio
->devs
[1].repl_bio
;
2946 bio
->bi_end_io
= NULL
;
2947 rdev
= mirror
->replacement
;
2948 /* Note: if rdev != NULL, then bio
2949 * cannot be NULL as r10buf_pool_alloc will
2950 * have allocated it.
2951 * So the second test here is pointless.
2952 * But it keeps semantic-checkers happy, and
2953 * this comment keeps human reviewers
2956 if (rdev
== NULL
|| bio
== NULL
||
2957 test_bit(Faulty
, &rdev
->flags
))
2959 bio
->bi_next
= biolist
;
2961 bio
->bi_private
= r10_bio
;
2962 bio
->bi_end_io
= end_sync_write
;
2964 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
2965 bio
->bi_bdev
= rdev
->bdev
;
2966 atomic_inc(&r10_bio
->remaining
);
2969 if (j
== conf
->copies
) {
2970 /* Cannot recover, so abort the recovery or
2971 * record a bad block */
2974 atomic_dec(&rb2
->remaining
);
2977 /* problem is that there are bad blocks
2978 * on other device(s)
2981 for (k
= 0; k
< conf
->copies
; k
++)
2982 if (r10_bio
->devs
[k
].devnum
== i
)
2984 if (!test_bit(In_sync
,
2985 &mirror
->rdev
->flags
)
2986 && !rdev_set_badblocks(
2988 r10_bio
->devs
[k
].addr
,
2991 if (mirror
->replacement
&&
2992 !rdev_set_badblocks(
2993 mirror
->replacement
,
2994 r10_bio
->devs
[k
].addr
,
2999 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3001 printk(KERN_INFO
"md/raid10:%s: insufficient "
3002 "working devices for recovery.\n",
3004 mirror
->recovery_disabled
3005 = mddev
->recovery_disabled
;
3010 if (biolist
== NULL
) {
3012 struct r10bio
*rb2
= r10_bio
;
3013 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3014 rb2
->master_bio
= NULL
;
3020 /* resync. Schedule a read for every block at this virt offset */
3023 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3025 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3026 &sync_blocks
, mddev
->degraded
) &&
3027 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3028 &mddev
->recovery
)) {
3029 /* We can skip this block */
3031 return sync_blocks
+ sectors_skipped
;
3033 if (sync_blocks
< max_sync
)
3034 max_sync
= sync_blocks
;
3035 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3037 r10_bio
->mddev
= mddev
;
3038 atomic_set(&r10_bio
->remaining
, 0);
3039 raise_barrier(conf
, 0);
3040 conf
->next_resync
= sector_nr
;
3042 r10_bio
->master_bio
= NULL
;
3043 r10_bio
->sector
= sector_nr
;
3044 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3045 raid10_find_phys(conf
, r10_bio
);
3046 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3048 for (i
= 0; i
< conf
->copies
; i
++) {
3049 int d
= r10_bio
->devs
[i
].devnum
;
3050 sector_t first_bad
, sector
;
3053 if (r10_bio
->devs
[i
].repl_bio
)
3054 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3056 bio
= r10_bio
->devs
[i
].bio
;
3057 bio
->bi_end_io
= NULL
;
3058 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3059 if (conf
->mirrors
[d
].rdev
== NULL
||
3060 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3062 sector
= r10_bio
->devs
[i
].addr
;
3063 if (is_badblock(conf
->mirrors
[d
].rdev
,
3065 &first_bad
, &bad_sectors
)) {
3066 if (first_bad
> sector
)
3067 max_sync
= first_bad
- sector
;
3069 bad_sectors
-= (sector
- first_bad
);
3070 if (max_sync
> bad_sectors
)
3071 max_sync
= max_sync
;
3075 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3076 atomic_inc(&r10_bio
->remaining
);
3077 bio
->bi_next
= biolist
;
3079 bio
->bi_private
= r10_bio
;
3080 bio
->bi_end_io
= end_sync_read
;
3082 bio
->bi_sector
= sector
+
3083 conf
->mirrors
[d
].rdev
->data_offset
;
3084 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3087 if (conf
->mirrors
[d
].replacement
== NULL
||
3089 &conf
->mirrors
[d
].replacement
->flags
))
3092 /* Need to set up for writing to the replacement */
3093 bio
= r10_bio
->devs
[i
].repl_bio
;
3094 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3096 sector
= r10_bio
->devs
[i
].addr
;
3097 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3098 bio
->bi_next
= biolist
;
3100 bio
->bi_private
= r10_bio
;
3101 bio
->bi_end_io
= end_sync_write
;
3103 bio
->bi_sector
= sector
+
3104 conf
->mirrors
[d
].replacement
->data_offset
;
3105 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3110 for (i
=0; i
<conf
->copies
; i
++) {
3111 int d
= r10_bio
->devs
[i
].devnum
;
3112 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3113 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3115 if (r10_bio
->devs
[i
].repl_bio
&&
3116 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3118 conf
->mirrors
[d
].replacement
,
3127 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3129 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
3131 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
3134 bio
->bi_phys_segments
= 0;
3139 if (sector_nr
+ max_sync
< max_sector
)
3140 max_sector
= sector_nr
+ max_sync
;
3143 int len
= PAGE_SIZE
;
3144 if (sector_nr
+ (len
>>9) > max_sector
)
3145 len
= (max_sector
- sector_nr
) << 9;
3148 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3150 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3151 if (bio_add_page(bio
, page
, len
, 0))
3155 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3156 for (bio2
= biolist
;
3157 bio2
&& bio2
!= bio
;
3158 bio2
= bio2
->bi_next
) {
3159 /* remove last page from this bio */
3161 bio2
->bi_size
-= len
;
3162 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3166 nr_sectors
+= len
>>9;
3167 sector_nr
+= len
>>9;
3168 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3170 r10_bio
->sectors
= nr_sectors
;
3174 biolist
= biolist
->bi_next
;
3176 bio
->bi_next
= NULL
;
3177 r10_bio
= bio
->bi_private
;
3178 r10_bio
->sectors
= nr_sectors
;
3180 if (bio
->bi_end_io
== end_sync_read
) {
3181 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3182 generic_make_request(bio
);
3186 if (sectors_skipped
)
3187 /* pretend they weren't skipped, it makes
3188 * no important difference in this case
3190 md_done_sync(mddev
, sectors_skipped
, 1);
3192 return sectors_skipped
+ nr_sectors
;
3194 /* There is nowhere to write, so all non-sync
3195 * drives must be failed or in resync, all drives
3196 * have a bad block, so try the next chunk...
3198 if (sector_nr
+ max_sync
< max_sector
)
3199 max_sector
= sector_nr
+ max_sync
;
3201 sectors_skipped
+= (max_sector
- sector_nr
);
3203 sector_nr
= max_sector
;
3208 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3211 struct r10conf
*conf
= mddev
->private;
3214 raid_disks
= conf
->geo
.raid_disks
;
3216 sectors
= conf
->dev_sectors
;
3218 size
= sectors
>> conf
->geo
.chunk_shift
;
3219 sector_div(size
, conf
->geo
.far_copies
);
3220 size
= size
* raid_disks
;
3221 sector_div(size
, conf
->geo
.near_copies
);
3223 return size
<< conf
->geo
.chunk_shift
;
3226 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3228 /* Calculate the number of sectors-per-device that will
3229 * actually be used, and set conf->dev_sectors and
3233 size
= size
>> conf
->geo
.chunk_shift
;
3234 sector_div(size
, conf
->geo
.far_copies
);
3235 size
= size
* conf
->geo
.raid_disks
;
3236 sector_div(size
, conf
->geo
.near_copies
);
3237 /* 'size' is now the number of chunks in the array */
3238 /* calculate "used chunks per device" */
3239 size
= size
* conf
->copies
;
3241 /* We need to round up when dividing by raid_disks to
3242 * get the stride size.
3244 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3246 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3248 if (conf
->geo
.far_offset
)
3249 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3251 sector_div(size
, conf
->geo
.far_copies
);
3252 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3256 enum geo_type
{geo_new
, geo_old
, geo_start
};
3257 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3260 int layout
, chunk
, disks
;
3263 layout
= mddev
->layout
;
3264 chunk
= mddev
->chunk_sectors
;
3265 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3268 layout
= mddev
->new_layout
;
3269 chunk
= mddev
->new_chunk_sectors
;
3270 disks
= mddev
->raid_disks
;
3272 default: /* avoid 'may be unused' warnings */
3273 case geo_start
: /* new when starting reshape - raid_disks not
3275 layout
= mddev
->new_layout
;
3276 chunk
= mddev
->new_chunk_sectors
;
3277 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3282 if (chunk
< (PAGE_SIZE
>> 9) ||
3283 !is_power_of_2(chunk
))
3286 fc
= (layout
>> 8) & 255;
3287 fo
= layout
& (1<<16);
3288 geo
->raid_disks
= disks
;
3289 geo
->near_copies
= nc
;
3290 geo
->far_copies
= fc
;
3291 geo
->far_offset
= fo
;
3292 geo
->chunk_mask
= chunk
- 1;
3293 geo
->chunk_shift
= ffz(~chunk
);
3297 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3299 struct r10conf
*conf
= NULL
;
3304 copies
= setup_geo(&geo
, mddev
, geo_new
);
3307 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3308 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3309 mdname(mddev
), PAGE_SIZE
);
3313 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3314 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3315 mdname(mddev
), mddev
->new_layout
);
3320 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3324 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
3329 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3334 conf
->copies
= copies
;
3335 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3336 r10bio_pool_free
, conf
);
3337 if (!conf
->r10bio_pool
)
3340 calc_sectors(conf
, mddev
->dev_sectors
);
3341 conf
->prev
= conf
->geo
;
3342 conf
->reshape_progress
= MaxSector
;
3344 spin_lock_init(&conf
->device_lock
);
3345 INIT_LIST_HEAD(&conf
->retry_list
);
3347 spin_lock_init(&conf
->resync_lock
);
3348 init_waitqueue_head(&conf
->wait_barrier
);
3350 conf
->thread
= md_register_thread(raid10d
, mddev
, NULL
);
3354 conf
->mddev
= mddev
;
3358 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3361 if (conf
->r10bio_pool
)
3362 mempool_destroy(conf
->r10bio_pool
);
3363 kfree(conf
->mirrors
);
3364 safe_put_page(conf
->tmppage
);
3367 return ERR_PTR(err
);
3370 static int run(struct mddev
*mddev
)
3372 struct r10conf
*conf
;
3373 int i
, disk_idx
, chunk_size
;
3374 struct mirror_info
*disk
;
3375 struct md_rdev
*rdev
;
3379 * copy the already verified devices into our private RAID10
3380 * bookkeeping area. [whatever we allocate in run(),
3381 * should be freed in stop()]
3384 if (mddev
->private == NULL
) {
3385 conf
= setup_conf(mddev
);
3387 return PTR_ERR(conf
);
3388 mddev
->private = conf
;
3390 conf
= mddev
->private;
3394 mddev
->thread
= conf
->thread
;
3395 conf
->thread
= NULL
;
3397 chunk_size
= mddev
->chunk_sectors
<< 9;
3398 blk_queue_io_min(mddev
->queue
, chunk_size
);
3399 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3400 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3402 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3403 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3405 rdev_for_each(rdev
, mddev
) {
3407 disk_idx
= rdev
->raid_disk
;
3410 if (disk_idx
>= conf
->geo
.raid_disks
&&
3411 disk_idx
>= conf
->prev
.raid_disks
)
3413 disk
= conf
->mirrors
+ disk_idx
;
3415 if (test_bit(Replacement
, &rdev
->flags
)) {
3416 if (disk
->replacement
)
3418 disk
->replacement
= rdev
;
3425 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3426 rdev
->data_offset
<< 9);
3428 disk
->head_position
= 0;
3430 /* need to check that every block has at least one working mirror */
3431 if (!enough(conf
, -1)) {
3432 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3437 mddev
->degraded
= 0;
3439 i
< conf
->geo
.raid_disks
3440 || i
< conf
->prev
.raid_disks
;
3443 disk
= conf
->mirrors
+ i
;
3445 if (!disk
->rdev
&& disk
->replacement
) {
3446 /* The replacement is all we have - use it */
3447 disk
->rdev
= disk
->replacement
;
3448 disk
->replacement
= NULL
;
3449 clear_bit(Replacement
, &disk
->rdev
->flags
);
3453 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3454 disk
->head_position
= 0;
3459 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3462 if (mddev
->recovery_cp
!= MaxSector
)
3463 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3464 " -- starting background reconstruction\n",
3467 "md/raid10:%s: active with %d out of %d devices\n",
3468 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3469 conf
->geo
.raid_disks
);
3471 * Ok, everything is just fine now
3473 mddev
->dev_sectors
= conf
->dev_sectors
;
3474 size
= raid10_size(mddev
, 0, 0);
3475 md_set_array_sectors(mddev
, size
);
3476 mddev
->resync_max_sectors
= size
;
3478 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3479 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3481 /* Calculate max read-ahead size.
3482 * We need to readahead at least twice a whole stripe....
3486 int stripe
= conf
->geo
.raid_disks
*
3487 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3488 stripe
/= conf
->geo
.near_copies
;
3489 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
3490 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
3493 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3495 if (md_integrity_register(mddev
))
3501 md_unregister_thread(&mddev
->thread
);
3502 if (conf
->r10bio_pool
)
3503 mempool_destroy(conf
->r10bio_pool
);
3504 safe_put_page(conf
->tmppage
);
3505 kfree(conf
->mirrors
);
3507 mddev
->private = NULL
;
3512 static int stop(struct mddev
*mddev
)
3514 struct r10conf
*conf
= mddev
->private;
3516 raise_barrier(conf
, 0);
3517 lower_barrier(conf
);
3519 md_unregister_thread(&mddev
->thread
);
3520 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
3521 if (conf
->r10bio_pool
)
3522 mempool_destroy(conf
->r10bio_pool
);
3523 kfree(conf
->mirrors
);
3525 mddev
->private = NULL
;
3529 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3531 struct r10conf
*conf
= mddev
->private;
3535 raise_barrier(conf
, 0);
3538 lower_barrier(conf
);
3543 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3545 /* Resize of 'far' arrays is not supported.
3546 * For 'near' and 'offset' arrays we can set the
3547 * number of sectors used to be an appropriate multiple
3548 * of the chunk size.
3549 * For 'offset', this is far_copies*chunksize.
3550 * For 'near' the multiplier is the LCM of
3551 * near_copies and raid_disks.
3552 * So if far_copies > 1 && !far_offset, fail.
3553 * Else find LCM(raid_disks, near_copy)*far_copies and
3554 * multiply by chunk_size. Then round to this number.
3555 * This is mostly done by raid10_size()
3557 struct r10conf
*conf
= mddev
->private;
3558 sector_t oldsize
, size
;
3560 if (mddev
->reshape_position
!= MaxSector
)
3563 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3566 oldsize
= raid10_size(mddev
, 0, 0);
3567 size
= raid10_size(mddev
, sectors
, 0);
3568 md_set_array_sectors(mddev
, size
);
3569 if (mddev
->array_sectors
> size
)
3571 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3572 revalidate_disk(mddev
->gendisk
);
3573 if (sectors
> mddev
->dev_sectors
&&
3574 mddev
->recovery_cp
> oldsize
) {
3575 mddev
->recovery_cp
= oldsize
;
3576 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3578 calc_sectors(conf
, sectors
);
3579 mddev
->dev_sectors
= conf
->dev_sectors
;
3580 mddev
->resync_max_sectors
= size
;
3584 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3586 struct md_rdev
*rdev
;
3587 struct r10conf
*conf
;
3589 if (mddev
->degraded
> 0) {
3590 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3592 return ERR_PTR(-EINVAL
);
3595 /* Set new parameters */
3596 mddev
->new_level
= 10;
3597 /* new layout: far_copies = 1, near_copies = 2 */
3598 mddev
->new_layout
= (1<<8) + 2;
3599 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3600 mddev
->delta_disks
= mddev
->raid_disks
;
3601 mddev
->raid_disks
*= 2;
3602 /* make sure it will be not marked as dirty */
3603 mddev
->recovery_cp
= MaxSector
;
3605 conf
= setup_conf(mddev
);
3606 if (!IS_ERR(conf
)) {
3607 rdev_for_each(rdev
, mddev
)
3608 if (rdev
->raid_disk
>= 0)
3609 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3616 static void *raid10_takeover(struct mddev
*mddev
)
3618 struct r0conf
*raid0_conf
;
3620 /* raid10 can take over:
3621 * raid0 - providing it has only two drives
3623 if (mddev
->level
== 0) {
3624 /* for raid0 takeover only one zone is supported */
3625 raid0_conf
= mddev
->private;
3626 if (raid0_conf
->nr_strip_zones
> 1) {
3627 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3628 " with more than one zone.\n",
3630 return ERR_PTR(-EINVAL
);
3632 return raid10_takeover_raid0(mddev
);
3634 return ERR_PTR(-EINVAL
);
3637 static struct md_personality raid10_personality
=
3641 .owner
= THIS_MODULE
,
3642 .make_request
= make_request
,
3646 .error_handler
= error
,
3647 .hot_add_disk
= raid10_add_disk
,
3648 .hot_remove_disk
= raid10_remove_disk
,
3649 .spare_active
= raid10_spare_active
,
3650 .sync_request
= sync_request
,
3651 .quiesce
= raid10_quiesce
,
3652 .size
= raid10_size
,
3653 .resize
= raid10_resize
,
3654 .takeover
= raid10_takeover
,
3657 static int __init
raid_init(void)
3659 return register_md_personality(&raid10_personality
);
3662 static void raid_exit(void)
3664 unregister_md_personality(&raid10_personality
);
3667 module_init(raid_init
);
3668 module_exit(raid_exit
);
3669 MODULE_LICENSE("GPL");
3670 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3671 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3672 MODULE_ALIAS("md-raid10");
3673 MODULE_ALIAS("md-level-10");
3675 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);