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 futher 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/raid/raid10.h>
24 * RAID10 provides a combination of RAID0 and RAID1 functionality.
25 * The layout of data is defined by
28 * near_copies (stored in low byte of layout)
29 * far_copies (stored in second byte of layout)
31 * The data to be stored is divided into chunks using chunksize.
32 * Each device is divided into far_copies sections.
33 * In each section, chunks are laid out in a style similar to raid0, but
34 * near_copies copies of each chunk is stored (each on a different drive).
35 * The starting device for each section is offset near_copies from the starting
36 * device of the previous section.
37 * Thus there are (near_copies*far_copies) of each chunk, and each is on a different
39 * near_copies and far_copies must be at least one, and their product is at most
44 * Number of guaranteed r10bios in case of extreme VM load:
46 #define NR_RAID10_BIOS 256
48 static void unplug_slaves(mddev_t
*mddev
);
50 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
54 int size
= offsetof(struct r10bio_s
, devs
[conf
->copies
]);
56 /* allocate a r10bio with room for raid_disks entries in the bios array */
57 r10_bio
= kmalloc(size
, gfp_flags
);
59 memset(r10_bio
, 0, size
);
61 unplug_slaves(conf
->mddev
);
66 static void r10bio_pool_free(void *r10_bio
, void *data
)
71 #define RESYNC_BLOCK_SIZE (64*1024)
72 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
73 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
74 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
75 #define RESYNC_WINDOW (2048*1024)
78 * When performing a resync, we need to read and compare, so
79 * we need as many pages are there are copies.
80 * When performing a recovery, we need 2 bios, one for read,
81 * one for write (we recover only one drive per r10buf)
84 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
93 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
95 unplug_slaves(conf
->mddev
);
99 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
100 nalloc
= conf
->copies
; /* resync */
102 nalloc
= 2; /* recovery */
107 for (j
= nalloc
; j
-- ; ) {
108 bio
= bio_alloc(gfp_flags
, RESYNC_PAGES
);
111 r10_bio
->devs
[j
].bio
= bio
;
114 * Allocate RESYNC_PAGES data pages and attach them
117 for (j
= 0 ; j
< nalloc
; j
++) {
118 bio
= r10_bio
->devs
[j
].bio
;
119 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
120 page
= alloc_page(gfp_flags
);
124 bio
->bi_io_vec
[i
].bv_page
= page
;
132 __free_page(bio
->bi_io_vec
[i
-1].bv_page
);
134 for (i
= 0; i
< RESYNC_PAGES
; i
++)
135 __free_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
138 while ( ++j
< nalloc
)
139 bio_put(r10_bio
->devs
[j
].bio
);
140 r10bio_pool_free(r10_bio
, conf
);
144 static void r10buf_pool_free(void *__r10_bio
, void *data
)
148 r10bio_t
*r10bio
= __r10_bio
;
151 for (j
=0; j
< conf
->copies
; j
++) {
152 struct bio
*bio
= r10bio
->devs
[j
].bio
;
154 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
155 __free_page(bio
->bi_io_vec
[i
].bv_page
);
156 bio
->bi_io_vec
[i
].bv_page
= NULL
;
161 r10bio_pool_free(r10bio
, conf
);
164 static void put_all_bios(conf_t
*conf
, r10bio_t
*r10_bio
)
168 for (i
= 0; i
< conf
->copies
; i
++) {
169 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
176 static inline void free_r10bio(r10bio_t
*r10_bio
)
180 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
183 * Wake up any possible resync thread that waits for the device
186 spin_lock_irqsave(&conf
->resync_lock
, flags
);
187 if (!--conf
->nr_pending
) {
188 wake_up(&conf
->wait_idle
);
189 wake_up(&conf
->wait_resume
);
191 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
193 put_all_bios(conf
, r10_bio
);
194 mempool_free(r10_bio
, conf
->r10bio_pool
);
197 static inline void put_buf(r10bio_t
*r10_bio
)
199 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
202 mempool_free(r10_bio
, conf
->r10buf_pool
);
204 spin_lock_irqsave(&conf
->resync_lock
, flags
);
208 wake_up(&conf
->wait_resume
);
209 wake_up(&conf
->wait_idle
);
211 if (!--conf
->nr_pending
) {
212 wake_up(&conf
->wait_idle
);
213 wake_up(&conf
->wait_resume
);
215 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
218 static void reschedule_retry(r10bio_t
*r10_bio
)
221 mddev_t
*mddev
= r10_bio
->mddev
;
222 conf_t
*conf
= mddev_to_conf(mddev
);
224 spin_lock_irqsave(&conf
->device_lock
, flags
);
225 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
226 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
228 md_wakeup_thread(mddev
->thread
);
232 * raid_end_bio_io() is called when we have finished servicing a mirrored
233 * operation and are ready to return a success/failure code to the buffer
236 static void raid_end_bio_io(r10bio_t
*r10_bio
)
238 struct bio
*bio
= r10_bio
->master_bio
;
240 bio_endio(bio
, bio
->bi_size
,
241 test_bit(R10BIO_Uptodate
, &r10_bio
->state
) ? 0 : -EIO
);
242 free_r10bio(r10_bio
);
246 * Update disk head position estimator based on IRQ completion info.
248 static inline void update_head_pos(int slot
, r10bio_t
*r10_bio
)
250 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
252 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
253 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
256 static int raid10_end_read_request(struct bio
*bio
, unsigned int bytes_done
, int error
)
258 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
259 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
261 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
266 slot
= r10_bio
->read_slot
;
267 dev
= r10_bio
->devs
[slot
].devnum
;
269 * this branch is our 'one mirror IO has finished' event handler:
272 md_error(r10_bio
->mddev
, conf
->mirrors
[dev
].rdev
);
275 * Set R10BIO_Uptodate in our master bio, so that
276 * we will return a good error code to the higher
277 * levels even if IO on some other mirrored buffer fails.
279 * The 'master' represents the composite IO operation to
280 * user-side. So if something waits for IO, then it will
281 * wait for the 'master' bio.
283 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
285 update_head_pos(slot
, r10_bio
);
288 * we have only one bio on the read side
291 raid_end_bio_io(r10_bio
);
296 char b
[BDEVNAME_SIZE
];
297 if (printk_ratelimit())
298 printk(KERN_ERR
"raid10: %s: rescheduling sector %llu\n",
299 bdevname(conf
->mirrors
[dev
].rdev
->bdev
,b
), (unsigned long long)r10_bio
->sector
);
300 reschedule_retry(r10_bio
);
303 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
307 static int raid10_end_write_request(struct bio
*bio
, unsigned int bytes_done
, int error
)
309 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
310 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
312 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
317 for (slot
= 0; slot
< conf
->copies
; slot
++)
318 if (r10_bio
->devs
[slot
].bio
== bio
)
320 dev
= r10_bio
->devs
[slot
].devnum
;
323 * this branch is our 'one mirror IO has finished' event handler:
326 md_error(r10_bio
->mddev
, conf
->mirrors
[dev
].rdev
);
329 * Set R10BIO_Uptodate in our master bio, so that
330 * we will return a good error code for to the higher
331 * levels even if IO on some other mirrored buffer fails.
333 * The 'master' represents the composite IO operation to
334 * user-side. So if something waits for IO, then it will
335 * wait for the 'master' bio.
337 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
339 update_head_pos(slot
, r10_bio
);
343 * Let's see if all mirrored write operations have finished
346 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
347 md_write_end(r10_bio
->mddev
);
348 raid_end_bio_io(r10_bio
);
351 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
357 * RAID10 layout manager
358 * Aswell as the chunksize and raid_disks count, there are two
359 * parameters: near_copies and far_copies.
360 * near_copies * far_copies must be <= raid_disks.
361 * Normally one of these will be 1.
362 * If both are 1, we get raid0.
363 * If near_copies == raid_disks, we get raid1.
365 * Chunks are layed out in raid0 style with near_copies copies of the
366 * first chunk, followed by near_copies copies of the next chunk and
368 * If far_copies > 1, then after 1/far_copies of the array has been assigned
369 * as described above, we start again with a device offset of near_copies.
370 * So we effectively have another copy of the whole array further down all
371 * the drives, but with blocks on different drives.
372 * With this layout, and block is never stored twice on the one device.
374 * raid10_find_phys finds the sector offset of a given virtual sector
375 * on each device that it is on. If a block isn't on a device,
376 * that entry in the array is set to MaxSector.
378 * raid10_find_virt does the reverse mapping, from a device and a
379 * sector offset to a virtual address
382 static void raid10_find_phys(conf_t
*conf
, r10bio_t
*r10bio
)
392 /* now calculate first sector/dev */
393 chunk
= r10bio
->sector
>> conf
->chunk_shift
;
394 sector
= r10bio
->sector
& conf
->chunk_mask
;
396 chunk
*= conf
->near_copies
;
398 dev
= sector_div(stripe
, conf
->raid_disks
);
400 sector
+= stripe
<< conf
->chunk_shift
;
402 /* and calculate all the others */
403 for (n
=0; n
< conf
->near_copies
; n
++) {
406 r10bio
->devs
[slot
].addr
= sector
;
407 r10bio
->devs
[slot
].devnum
= d
;
410 for (f
= 1; f
< conf
->far_copies
; f
++) {
411 d
+= conf
->near_copies
;
412 if (d
>= conf
->raid_disks
)
413 d
-= conf
->raid_disks
;
415 r10bio
->devs
[slot
].devnum
= d
;
416 r10bio
->devs
[slot
].addr
= s
;
420 if (dev
>= conf
->raid_disks
) {
422 sector
+= (conf
->chunk_mask
+ 1);
425 BUG_ON(slot
!= conf
->copies
);
428 static sector_t
raid10_find_virt(conf_t
*conf
, sector_t sector
, int dev
)
430 sector_t offset
, chunk
, vchunk
;
432 while (sector
> conf
->stride
) {
433 sector
-= conf
->stride
;
434 if (dev
< conf
->near_copies
)
435 dev
+= conf
->raid_disks
- conf
->near_copies
;
437 dev
-= conf
->near_copies
;
440 offset
= sector
& conf
->chunk_mask
;
441 chunk
= sector
>> conf
->chunk_shift
;
442 vchunk
= chunk
* conf
->raid_disks
+ dev
;
443 sector_div(vchunk
, conf
->near_copies
);
444 return (vchunk
<< conf
->chunk_shift
) + offset
;
448 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
450 * @bio: the buffer head that's been built up so far
451 * @biovec: the request that could be merged to it.
453 * Return amount of bytes we can accept at this offset
454 * If near_copies == raid_disk, there are no striping issues,
455 * but in that case, the function isn't called at all.
457 static int raid10_mergeable_bvec(request_queue_t
*q
, struct bio
*bio
,
458 struct bio_vec
*bio_vec
)
460 mddev_t
*mddev
= q
->queuedata
;
461 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
463 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
464 unsigned int bio_sectors
= bio
->bi_size
>> 9;
466 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
467 if (max
< 0) max
= 0; /* bio_add cannot handle a negative return */
468 if (max
<= bio_vec
->bv_len
&& bio_sectors
== 0)
469 return bio_vec
->bv_len
;
475 * This routine returns the disk from which the requested read should
476 * be done. There is a per-array 'next expected sequential IO' sector
477 * number - if this matches on the next IO then we use the last disk.
478 * There is also a per-disk 'last know head position' sector that is
479 * maintained from IRQ contexts, both the normal and the resync IO
480 * completion handlers update this position correctly. If there is no
481 * perfect sequential match then we pick the disk whose head is closest.
483 * If there are 2 mirrors in the same 2 devices, performance degrades
484 * because position is mirror, not device based.
486 * The rdev for the device selected will have nr_pending incremented.
490 * FIXME: possibly should rethink readbalancing and do it differently
491 * depending on near_copies / far_copies geometry.
493 static int read_balance(conf_t
*conf
, r10bio_t
*r10_bio
)
495 const unsigned long this_sector
= r10_bio
->sector
;
496 int disk
, slot
, nslot
;
497 const int sectors
= r10_bio
->sectors
;
498 sector_t new_distance
, current_distance
;
501 raid10_find_phys(conf
, r10_bio
);
504 * Check if we can balance. We can balance on the whole
505 * device if no resync is going on, or below the resync window.
506 * We take the first readable disk when above the resync window.
508 if (conf
->mddev
->recovery_cp
< MaxSector
509 && (this_sector
+ sectors
>= conf
->next_resync
)) {
510 /* make sure that disk is operational */
512 disk
= r10_bio
->devs
[slot
].devnum
;
514 while ((rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
515 !test_bit(In_sync
, &rdev
->flags
)) {
517 if (slot
== conf
->copies
) {
522 disk
= r10_bio
->devs
[slot
].devnum
;
528 /* make sure the disk is operational */
530 disk
= r10_bio
->devs
[slot
].devnum
;
531 while ((rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
532 !test_bit(In_sync
, &rdev
->flags
)) {
534 if (slot
== conf
->copies
) {
538 disk
= r10_bio
->devs
[slot
].devnum
;
542 current_distance
= abs(r10_bio
->devs
[slot
].addr
-
543 conf
->mirrors
[disk
].head_position
);
545 /* Find the disk whose head is closest */
547 for (nslot
= slot
; nslot
< conf
->copies
; nslot
++) {
548 int ndisk
= r10_bio
->devs
[nslot
].devnum
;
551 if ((rdev
=rcu_dereference(conf
->mirrors
[ndisk
].rdev
)) == NULL
||
552 !test_bit(In_sync
, &rdev
->flags
))
555 /* This optimisation is debatable, and completely destroys
556 * sequential read speed for 'far copies' arrays. So only
557 * keep it for 'near' arrays, and review those later.
559 if (conf
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
)) {
564 new_distance
= abs(r10_bio
->devs
[nslot
].addr
-
565 conf
->mirrors
[ndisk
].head_position
);
566 if (new_distance
< current_distance
) {
567 current_distance
= new_distance
;
574 r10_bio
->read_slot
= slot
;
575 /* conf->next_seq_sect = this_sector + sectors;*/
577 if (disk
>= 0 && (rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
))!= NULL
)
578 atomic_inc(&conf
->mirrors
[disk
].rdev
->nr_pending
);
584 static void unplug_slaves(mddev_t
*mddev
)
586 conf_t
*conf
= mddev_to_conf(mddev
);
590 for (i
=0; i
<mddev
->raid_disks
; i
++) {
591 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
592 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
593 request_queue_t
*r_queue
= bdev_get_queue(rdev
->bdev
);
595 atomic_inc(&rdev
->nr_pending
);
598 if (r_queue
->unplug_fn
)
599 r_queue
->unplug_fn(r_queue
);
601 rdev_dec_pending(rdev
, mddev
);
608 static void raid10_unplug(request_queue_t
*q
)
610 unplug_slaves(q
->queuedata
);
613 static int raid10_issue_flush(request_queue_t
*q
, struct gendisk
*disk
,
614 sector_t
*error_sector
)
616 mddev_t
*mddev
= q
->queuedata
;
617 conf_t
*conf
= mddev_to_conf(mddev
);
621 for (i
=0; i
<mddev
->raid_disks
&& ret
== 0; i
++) {
622 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
623 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
624 struct block_device
*bdev
= rdev
->bdev
;
625 request_queue_t
*r_queue
= bdev_get_queue(bdev
);
627 if (!r_queue
->issue_flush_fn
)
630 atomic_inc(&rdev
->nr_pending
);
632 ret
= r_queue
->issue_flush_fn(r_queue
, bdev
->bd_disk
,
634 rdev_dec_pending(rdev
, mddev
);
644 * Throttle resync depth, so that we can both get proper overlapping of
645 * requests, but are still able to handle normal requests quickly.
647 #define RESYNC_DEPTH 32
649 static void device_barrier(conf_t
*conf
, sector_t sect
)
651 spin_lock_irq(&conf
->resync_lock
);
652 wait_event_lock_irq(conf
->wait_idle
, !waitqueue_active(&conf
->wait_resume
),
653 conf
->resync_lock
, unplug_slaves(conf
->mddev
));
655 if (!conf
->barrier
++) {
656 wait_event_lock_irq(conf
->wait_idle
, !conf
->nr_pending
,
657 conf
->resync_lock
, unplug_slaves(conf
->mddev
));
658 if (conf
->nr_pending
)
661 wait_event_lock_irq(conf
->wait_resume
, conf
->barrier
< RESYNC_DEPTH
,
662 conf
->resync_lock
, unplug_slaves(conf
->mddev
));
663 conf
->next_resync
= sect
;
664 spin_unlock_irq(&conf
->resync_lock
);
667 static int make_request(request_queue_t
*q
, struct bio
* bio
)
669 mddev_t
*mddev
= q
->queuedata
;
670 conf_t
*conf
= mddev_to_conf(mddev
);
671 mirror_info_t
*mirror
;
673 struct bio
*read_bio
;
675 int chunk_sects
= conf
->chunk_mask
+ 1;
676 const int rw
= bio_data_dir(bio
);
678 if (unlikely(bio_barrier(bio
))) {
679 bio_endio(bio
, bio
->bi_size
, -EOPNOTSUPP
);
683 /* If this request crosses a chunk boundary, we need to
684 * split it. This will only happen for 1 PAGE (or less) requests.
686 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
688 conf
->near_copies
< conf
->raid_disks
)) {
690 /* Sanity check -- queue functions should prevent this happening */
691 if (bio
->bi_vcnt
!= 1 ||
694 /* This is a one page bio that upper layers
695 * refuse to split for us, so we need to split it.
697 bp
= bio_split(bio
, bio_split_pool
,
698 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
699 if (make_request(q
, &bp
->bio1
))
700 generic_make_request(&bp
->bio1
);
701 if (make_request(q
, &bp
->bio2
))
702 generic_make_request(&bp
->bio2
);
704 bio_pair_release(bp
);
707 printk("raid10_make_request bug: can't convert block across chunks"
708 " or bigger than %dk %llu %d\n", chunk_sects
/2,
709 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
711 bio_io_error(bio
, bio
->bi_size
);
715 md_write_start(mddev
, bio
);
718 * Register the new request and wait if the reconstruction
719 * thread has put up a bar for new requests.
720 * Continue immediately if no resync is active currently.
722 spin_lock_irq(&conf
->resync_lock
);
723 wait_event_lock_irq(conf
->wait_resume
, !conf
->barrier
, conf
->resync_lock
, );
725 spin_unlock_irq(&conf
->resync_lock
);
727 disk_stat_inc(mddev
->gendisk
, ios
[rw
]);
728 disk_stat_add(mddev
->gendisk
, sectors
[rw
], bio_sectors(bio
));
730 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
732 r10_bio
->master_bio
= bio
;
733 r10_bio
->sectors
= bio
->bi_size
>> 9;
735 r10_bio
->mddev
= mddev
;
736 r10_bio
->sector
= bio
->bi_sector
;
740 * read balancing logic:
742 int disk
= read_balance(conf
, r10_bio
);
743 int slot
= r10_bio
->read_slot
;
745 raid_end_bio_io(r10_bio
);
748 mirror
= conf
->mirrors
+ disk
;
750 read_bio
= bio_clone(bio
, GFP_NOIO
);
752 r10_bio
->devs
[slot
].bio
= read_bio
;
754 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
755 mirror
->rdev
->data_offset
;
756 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
757 read_bio
->bi_end_io
= raid10_end_read_request
;
758 read_bio
->bi_rw
= READ
;
759 read_bio
->bi_private
= r10_bio
;
761 generic_make_request(read_bio
);
768 /* first select target devices under spinlock and
769 * inc refcount on their rdev. Record them by setting
772 raid10_find_phys(conf
, r10_bio
);
774 for (i
= 0; i
< conf
->copies
; i
++) {
775 int d
= r10_bio
->devs
[i
].devnum
;
776 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
778 !test_bit(Faulty
, &rdev
->flags
)) {
779 atomic_inc(&rdev
->nr_pending
);
780 r10_bio
->devs
[i
].bio
= bio
;
782 r10_bio
->devs
[i
].bio
= NULL
;
786 atomic_set(&r10_bio
->remaining
, 1);
788 for (i
= 0; i
< conf
->copies
; i
++) {
790 int d
= r10_bio
->devs
[i
].devnum
;
791 if (!r10_bio
->devs
[i
].bio
)
794 mbio
= bio_clone(bio
, GFP_NOIO
);
795 r10_bio
->devs
[i
].bio
= mbio
;
797 mbio
->bi_sector
= r10_bio
->devs
[i
].addr
+
798 conf
->mirrors
[d
].rdev
->data_offset
;
799 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
800 mbio
->bi_end_io
= raid10_end_write_request
;
802 mbio
->bi_private
= r10_bio
;
804 atomic_inc(&r10_bio
->remaining
);
805 generic_make_request(mbio
);
808 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
810 raid_end_bio_io(r10_bio
);
816 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
818 conf_t
*conf
= mddev_to_conf(mddev
);
821 if (conf
->near_copies
< conf
->raid_disks
)
822 seq_printf(seq
, " %dK chunks", mddev
->chunk_size
/1024);
823 if (conf
->near_copies
> 1)
824 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
825 if (conf
->far_copies
> 1)
826 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
828 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
829 conf
->working_disks
);
830 for (i
= 0; i
< conf
->raid_disks
; i
++)
831 seq_printf(seq
, "%s",
832 conf
->mirrors
[i
].rdev
&&
833 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
834 seq_printf(seq
, "]");
837 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
839 char b
[BDEVNAME_SIZE
];
840 conf_t
*conf
= mddev_to_conf(mddev
);
843 * If it is not operational, then we have already marked it as dead
844 * else if it is the last working disks, ignore the error, let the
845 * next level up know.
846 * else mark the drive as failed
848 if (test_bit(In_sync
, &rdev
->flags
)
849 && conf
->working_disks
== 1)
851 * Don't fail the drive, just return an IO error.
852 * The test should really be more sophisticated than
853 * "working_disks == 1", but it isn't critical, and
854 * can wait until we do more sophisticated "is the drive
855 * really dead" tests...
858 if (test_bit(In_sync
, &rdev
->flags
)) {
860 conf
->working_disks
--;
862 * if recovery is running, make sure it aborts.
864 set_bit(MD_RECOVERY_ERR
, &mddev
->recovery
);
866 clear_bit(In_sync
, &rdev
->flags
);
867 set_bit(Faulty
, &rdev
->flags
);
869 printk(KERN_ALERT
"raid10: Disk failure on %s, disabling device. \n"
870 " Operation continuing on %d devices\n",
871 bdevname(rdev
->bdev
,b
), conf
->working_disks
);
874 static void print_conf(conf_t
*conf
)
879 printk("RAID10 conf printout:\n");
884 printk(" --- wd:%d rd:%d\n", conf
->working_disks
,
887 for (i
= 0; i
< conf
->raid_disks
; i
++) {
888 char b
[BDEVNAME_SIZE
];
889 tmp
= conf
->mirrors
+ i
;
891 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
892 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
893 !test_bit(Faulty
, &tmp
->rdev
->flags
),
894 bdevname(tmp
->rdev
->bdev
,b
));
898 static void close_sync(conf_t
*conf
)
900 spin_lock_irq(&conf
->resync_lock
);
901 wait_event_lock_irq(conf
->wait_resume
, !conf
->barrier
,
902 conf
->resync_lock
, unplug_slaves(conf
->mddev
));
903 spin_unlock_irq(&conf
->resync_lock
);
905 if (conf
->barrier
) BUG();
906 if (waitqueue_active(&conf
->wait_idle
)) BUG();
908 mempool_destroy(conf
->r10buf_pool
);
909 conf
->r10buf_pool
= NULL
;
912 /* check if there are enough drives for
913 * every block to appear on atleast one
915 static int enough(conf_t
*conf
)
920 int n
= conf
->copies
;
923 if (conf
->mirrors
[first
].rdev
)
925 first
= (first
+1) % conf
->raid_disks
;
929 } while (first
!= 0);
933 static int raid10_spare_active(mddev_t
*mddev
)
936 conf_t
*conf
= mddev
->private;
940 * Find all non-in_sync disks within the RAID10 configuration
941 * and mark them in_sync
943 for (i
= 0; i
< conf
->raid_disks
; i
++) {
944 tmp
= conf
->mirrors
+ i
;
946 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
947 && !test_bit(In_sync
, &tmp
->rdev
->flags
)) {
948 conf
->working_disks
++;
950 set_bit(In_sync
, &tmp
->rdev
->flags
);
959 static int raid10_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
961 conf_t
*conf
= mddev
->private;
966 if (mddev
->recovery_cp
< MaxSector
)
967 /* only hot-add to in-sync arrays, as recovery is
968 * very different from resync
974 for (mirror
=0; mirror
< mddev
->raid_disks
; mirror
++)
975 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
977 blk_queue_stack_limits(mddev
->queue
,
978 rdev
->bdev
->bd_disk
->queue
);
979 /* as we don't honour merge_bvec_fn, we must never risk
980 * violating it, so limit ->max_sector to one PAGE, as
981 * a one page request is never in violation.
983 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
984 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
985 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
987 p
->head_position
= 0;
988 rdev
->raid_disk
= mirror
;
990 rcu_assign_pointer(p
->rdev
, rdev
);
998 static int raid10_remove_disk(mddev_t
*mddev
, int number
)
1000 conf_t
*conf
= mddev
->private;
1003 mirror_info_t
*p
= conf
->mirrors
+ number
;
1008 if (test_bit(In_sync
, &rdev
->flags
) ||
1009 atomic_read(&rdev
->nr_pending
)) {
1015 if (atomic_read(&rdev
->nr_pending
)) {
1016 /* lost the race, try later */
1028 static int end_sync_read(struct bio
*bio
, unsigned int bytes_done
, int error
)
1030 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1031 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1032 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
1038 for (i
=0; i
<conf
->copies
; i
++)
1039 if (r10_bio
->devs
[i
].bio
== bio
)
1041 if (i
== conf
->copies
)
1043 update_head_pos(i
, r10_bio
);
1044 d
= r10_bio
->devs
[i
].devnum
;
1046 md_error(r10_bio
->mddev
,
1047 conf
->mirrors
[d
].rdev
);
1049 /* for reconstruct, we always reschedule after a read.
1050 * for resync, only after all reads
1052 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1053 atomic_dec_and_test(&r10_bio
->remaining
)) {
1054 /* we have read all the blocks,
1055 * do the comparison in process context in raid10d
1057 reschedule_retry(r10_bio
);
1059 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1063 static int end_sync_write(struct bio
*bio
, unsigned int bytes_done
, int error
)
1065 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1066 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1067 mddev_t
*mddev
= r10_bio
->mddev
;
1068 conf_t
*conf
= mddev_to_conf(mddev
);
1074 for (i
= 0; i
< conf
->copies
; i
++)
1075 if (r10_bio
->devs
[i
].bio
== bio
)
1077 d
= r10_bio
->devs
[i
].devnum
;
1080 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1081 update_head_pos(i
, r10_bio
);
1083 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1084 if (r10_bio
->master_bio
== NULL
) {
1085 /* the primary of several recovery bios */
1086 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1090 r10bio_t
*r10_bio2
= (r10bio_t
*)r10_bio
->master_bio
;
1095 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1100 * Note: sync and recover and handled very differently for raid10
1101 * This code is for resync.
1102 * For resync, we read through virtual addresses and read all blocks.
1103 * If there is any error, we schedule a write. The lowest numbered
1104 * drive is authoritative.
1105 * However requests come for physical address, so we need to map.
1106 * For every physical address there are raid_disks/copies virtual addresses,
1107 * which is always are least one, but is not necessarly an integer.
1108 * This means that a physical address can span multiple chunks, so we may
1109 * have to submit multiple io requests for a single sync request.
1112 * We check if all blocks are in-sync and only write to blocks that
1115 static void sync_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1117 conf_t
*conf
= mddev_to_conf(mddev
);
1119 struct bio
*tbio
, *fbio
;
1121 atomic_set(&r10_bio
->remaining
, 1);
1123 /* find the first device with a block */
1124 for (i
=0; i
<conf
->copies
; i
++)
1125 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1128 if (i
== conf
->copies
)
1132 fbio
= r10_bio
->devs
[i
].bio
;
1134 /* now find blocks with errors */
1135 for (i
=first
+1 ; i
< conf
->copies
; i
++) {
1138 if (!test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1140 /* We know that the bi_io_vec layout is the same for
1141 * both 'first' and 'i', so we just compare them.
1142 * All vec entries are PAGE_SIZE;
1144 tbio
= r10_bio
->devs
[i
].bio
;
1145 vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1146 for (j
= 0; j
< vcnt
; j
++)
1147 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1148 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1153 /* Ok, we need to write this bio
1154 * First we need to fixup bv_offset, bv_len and
1155 * bi_vecs, as the read request might have corrupted these
1157 tbio
->bi_vcnt
= vcnt
;
1158 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1160 tbio
->bi_phys_segments
= 0;
1161 tbio
->bi_hw_segments
= 0;
1162 tbio
->bi_hw_front_size
= 0;
1163 tbio
->bi_hw_back_size
= 0;
1164 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1165 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1166 tbio
->bi_next
= NULL
;
1167 tbio
->bi_rw
= WRITE
;
1168 tbio
->bi_private
= r10_bio
;
1169 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1171 for (j
=0; j
< vcnt
; j
++) {
1172 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1173 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1175 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1176 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1179 tbio
->bi_end_io
= end_sync_write
;
1181 d
= r10_bio
->devs
[i
].devnum
;
1182 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1183 atomic_inc(&r10_bio
->remaining
);
1184 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1186 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1187 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1188 generic_make_request(tbio
);
1192 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1193 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1199 * Now for the recovery code.
1200 * Recovery happens across physical sectors.
1201 * We recover all non-is_sync drives by finding the virtual address of
1202 * each, and then choose a working drive that also has that virt address.
1203 * There is a separate r10_bio for each non-in_sync drive.
1204 * Only the first two slots are in use. The first for reading,
1205 * The second for writing.
1209 static void recovery_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1211 conf_t
*conf
= mddev_to_conf(mddev
);
1213 struct bio
*bio
, *wbio
;
1216 /* move the pages across to the second bio
1217 * and submit the write request
1219 bio
= r10_bio
->devs
[0].bio
;
1220 wbio
= r10_bio
->devs
[1].bio
;
1221 for (i
=0; i
< wbio
->bi_vcnt
; i
++) {
1222 struct page
*p
= bio
->bi_io_vec
[i
].bv_page
;
1223 bio
->bi_io_vec
[i
].bv_page
= wbio
->bi_io_vec
[i
].bv_page
;
1224 wbio
->bi_io_vec
[i
].bv_page
= p
;
1226 d
= r10_bio
->devs
[1].devnum
;
1228 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1229 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1230 generic_make_request(wbio
);
1235 * This is a kernel thread which:
1237 * 1. Retries failed read operations on working mirrors.
1238 * 2. Updates the raid superblock when problems encounter.
1239 * 3. Performs writes following reads for array syncronising.
1242 static void raid10d(mddev_t
*mddev
)
1246 unsigned long flags
;
1247 conf_t
*conf
= mddev_to_conf(mddev
);
1248 struct list_head
*head
= &conf
->retry_list
;
1252 md_check_recovery(mddev
);
1255 char b
[BDEVNAME_SIZE
];
1256 spin_lock_irqsave(&conf
->device_lock
, flags
);
1257 if (list_empty(head
))
1259 r10_bio
= list_entry(head
->prev
, r10bio_t
, retry_list
);
1260 list_del(head
->prev
);
1261 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1263 mddev
= r10_bio
->mddev
;
1264 conf
= mddev_to_conf(mddev
);
1265 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
)) {
1266 sync_request_write(mddev
, r10_bio
);
1268 } else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
1269 recovery_request_write(mddev
, r10_bio
);
1273 bio
= r10_bio
->devs
[r10_bio
->read_slot
].bio
;
1274 r10_bio
->devs
[r10_bio
->read_slot
].bio
= NULL
;
1276 mirror
= read_balance(conf
, r10_bio
);
1278 printk(KERN_ALERT
"raid10: %s: unrecoverable I/O"
1279 " read error for block %llu\n",
1280 bdevname(bio
->bi_bdev
,b
),
1281 (unsigned long long)r10_bio
->sector
);
1282 raid_end_bio_io(r10_bio
);
1284 rdev
= conf
->mirrors
[mirror
].rdev
;
1285 if (printk_ratelimit())
1286 printk(KERN_ERR
"raid10: %s: redirecting sector %llu to"
1287 " another mirror\n",
1288 bdevname(rdev
->bdev
,b
),
1289 (unsigned long long)r10_bio
->sector
);
1290 bio
= bio_clone(r10_bio
->master_bio
, GFP_NOIO
);
1291 r10_bio
->devs
[r10_bio
->read_slot
].bio
= bio
;
1292 bio
->bi_sector
= r10_bio
->devs
[r10_bio
->read_slot
].addr
1293 + rdev
->data_offset
;
1294 bio
->bi_bdev
= rdev
->bdev
;
1296 bio
->bi_private
= r10_bio
;
1297 bio
->bi_end_io
= raid10_end_read_request
;
1299 generic_make_request(bio
);
1303 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1305 unplug_slaves(mddev
);
1309 static int init_resync(conf_t
*conf
)
1313 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1314 if (conf
->r10buf_pool
)
1316 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
1317 if (!conf
->r10buf_pool
)
1319 conf
->next_resync
= 0;
1324 * perform a "sync" on one "block"
1326 * We need to make sure that no normal I/O request - particularly write
1327 * requests - conflict with active sync requests.
1329 * This is achieved by tracking pending requests and a 'barrier' concept
1330 * that can be installed to exclude normal IO requests.
1332 * Resync and recovery are handled very differently.
1333 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1335 * For resync, we iterate over virtual addresses, read all copies,
1336 * and update if there are differences. If only one copy is live,
1338 * For recovery, we iterate over physical addresses, read a good
1339 * value for each non-in_sync drive, and over-write.
1341 * So, for recovery we may have several outstanding complex requests for a
1342 * given address, one for each out-of-sync device. We model this by allocating
1343 * a number of r10_bio structures, one for each out-of-sync device.
1344 * As we setup these structures, we collect all bio's together into a list
1345 * which we then process collectively to add pages, and then process again
1346 * to pass to generic_make_request.
1348 * The r10_bio structures are linked using a borrowed master_bio pointer.
1349 * This link is counted in ->remaining. When the r10_bio that points to NULL
1350 * has its remaining count decremented to 0, the whole complex operation
1355 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
1357 conf_t
*conf
= mddev_to_conf(mddev
);
1359 struct bio
*biolist
= NULL
, *bio
;
1360 sector_t max_sector
, nr_sectors
;
1364 sector_t sectors_skipped
= 0;
1365 int chunks_skipped
= 0;
1367 if (!conf
->r10buf_pool
)
1368 if (init_resync(conf
))
1372 max_sector
= mddev
->size
<< 1;
1373 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1374 max_sector
= mddev
->resync_max_sectors
;
1375 if (sector_nr
>= max_sector
) {
1378 return sectors_skipped
;
1380 if (chunks_skipped
>= conf
->raid_disks
) {
1381 /* if there has been nothing to do on any drive,
1382 * then there is nothing to do at all..
1385 return (max_sector
- sector_nr
) + sectors_skipped
;
1388 /* make sure whole request will fit in a chunk - if chunks
1391 if (conf
->near_copies
< conf
->raid_disks
&&
1392 max_sector
> (sector_nr
| conf
->chunk_mask
))
1393 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
1395 * If there is non-resync activity waiting for us then
1396 * put in a delay to throttle resync.
1398 if (!go_faster
&& waitqueue_active(&conf
->wait_resume
))
1399 msleep_interruptible(1000);
1400 device_barrier(conf
, sector_nr
+ RESYNC_SECTORS
);
1402 /* Again, very different code for resync and recovery.
1403 * Both must result in an r10bio with a list of bios that
1404 * have bi_end_io, bi_sector, bi_bdev set,
1405 * and bi_private set to the r10bio.
1406 * For recovery, we may actually create several r10bios
1407 * with 2 bios in each, that correspond to the bios in the main one.
1408 * In this case, the subordinate r10bios link back through a
1409 * borrowed master_bio pointer, and the counter in the master
1410 * includes a ref from each subordinate.
1412 /* First, we decide what to do and set ->bi_end_io
1413 * To end_sync_read if we want to read, and
1414 * end_sync_write if we will want to write.
1417 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
1418 /* recovery... the complicated one */
1422 for (i
=0 ; i
<conf
->raid_disks
; i
++)
1423 if (conf
->mirrors
[i
].rdev
&&
1424 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
)) {
1425 /* want to reconstruct this device */
1426 r10bio_t
*rb2
= r10_bio
;
1428 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1429 spin_lock_irq(&conf
->resync_lock
);
1431 if (rb2
) conf
->barrier
++;
1432 spin_unlock_irq(&conf
->resync_lock
);
1433 atomic_set(&r10_bio
->remaining
, 0);
1435 r10_bio
->master_bio
= (struct bio
*)rb2
;
1437 atomic_inc(&rb2
->remaining
);
1438 r10_bio
->mddev
= mddev
;
1439 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
1440 r10_bio
->sector
= raid10_find_virt(conf
, sector_nr
, i
);
1441 raid10_find_phys(conf
, r10_bio
);
1442 for (j
=0; j
<conf
->copies
;j
++) {
1443 int d
= r10_bio
->devs
[j
].devnum
;
1444 if (conf
->mirrors
[d
].rdev
&&
1445 test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
)) {
1446 /* This is where we read from */
1447 bio
= r10_bio
->devs
[0].bio
;
1448 bio
->bi_next
= biolist
;
1450 bio
->bi_private
= r10_bio
;
1451 bio
->bi_end_io
= end_sync_read
;
1453 bio
->bi_sector
= r10_bio
->devs
[j
].addr
+
1454 conf
->mirrors
[d
].rdev
->data_offset
;
1455 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1456 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1457 atomic_inc(&r10_bio
->remaining
);
1458 /* and we write to 'i' */
1460 for (k
=0; k
<conf
->copies
; k
++)
1461 if (r10_bio
->devs
[k
].devnum
== i
)
1463 bio
= r10_bio
->devs
[1].bio
;
1464 bio
->bi_next
= biolist
;
1466 bio
->bi_private
= r10_bio
;
1467 bio
->bi_end_io
= end_sync_write
;
1469 bio
->bi_sector
= r10_bio
->devs
[k
].addr
+
1470 conf
->mirrors
[i
].rdev
->data_offset
;
1471 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1473 r10_bio
->devs
[0].devnum
= d
;
1474 r10_bio
->devs
[1].devnum
= i
;
1479 if (j
== conf
->copies
) {
1480 /* Cannot recover, so abort the recovery */
1483 if (!test_and_set_bit(MD_RECOVERY_ERR
, &mddev
->recovery
))
1484 printk(KERN_INFO
"raid10: %s: insufficient working devices for recovery.\n",
1489 if (biolist
== NULL
) {
1491 r10bio_t
*rb2
= r10_bio
;
1492 r10_bio
= (r10bio_t
*) rb2
->master_bio
;
1493 rb2
->master_bio
= NULL
;
1499 /* resync. Schedule a read for every block at this virt offset */
1501 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1503 spin_lock_irq(&conf
->resync_lock
);
1505 spin_unlock_irq(&conf
->resync_lock
);
1507 r10_bio
->mddev
= mddev
;
1508 atomic_set(&r10_bio
->remaining
, 0);
1510 r10_bio
->master_bio
= NULL
;
1511 r10_bio
->sector
= sector_nr
;
1512 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
1513 raid10_find_phys(conf
, r10_bio
);
1514 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
1516 for (i
=0; i
<conf
->copies
; i
++) {
1517 int d
= r10_bio
->devs
[i
].devnum
;
1518 bio
= r10_bio
->devs
[i
].bio
;
1519 bio
->bi_end_io
= NULL
;
1520 if (conf
->mirrors
[d
].rdev
== NULL
||
1521 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
1523 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1524 atomic_inc(&r10_bio
->remaining
);
1525 bio
->bi_next
= biolist
;
1527 bio
->bi_private
= r10_bio
;
1528 bio
->bi_end_io
= end_sync_read
;
1530 bio
->bi_sector
= r10_bio
->devs
[i
].addr
+
1531 conf
->mirrors
[d
].rdev
->data_offset
;
1532 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1537 for (i
=0; i
<conf
->copies
; i
++) {
1538 int d
= r10_bio
->devs
[i
].devnum
;
1539 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
1540 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1548 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1550 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1552 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
1555 bio
->bi_phys_segments
= 0;
1556 bio
->bi_hw_segments
= 0;
1563 int len
= PAGE_SIZE
;
1565 if (sector_nr
+ (len
>>9) > max_sector
)
1566 len
= (max_sector
- sector_nr
) << 9;
1569 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1570 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
1571 if (bio_add_page(bio
, page
, len
, 0) == 0) {
1574 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
1575 for (bio2
= biolist
; bio2
&& bio2
!= bio
; bio2
= bio2
->bi_next
) {
1576 /* remove last page from this bio */
1578 bio2
->bi_size
-= len
;
1579 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
1585 nr_sectors
+= len
>>9;
1586 sector_nr
+= len
>>9;
1587 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
1589 r10_bio
->sectors
= nr_sectors
;
1593 biolist
= biolist
->bi_next
;
1595 bio
->bi_next
= NULL
;
1596 r10_bio
= bio
->bi_private
;
1597 r10_bio
->sectors
= nr_sectors
;
1599 if (bio
->bi_end_io
== end_sync_read
) {
1600 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1601 generic_make_request(bio
);
1605 if (sectors_skipped
)
1606 /* pretend they weren't skipped, it makes
1607 * no important difference in this case
1609 md_done_sync(mddev
, sectors_skipped
, 1);
1611 return sectors_skipped
+ nr_sectors
;
1613 /* There is nowhere to write, so all non-sync
1614 * drives must be failed, so try the next chunk...
1617 sector_t sec
= max_sector
- sector_nr
;
1618 sectors_skipped
+= sec
;
1620 sector_nr
= max_sector
;
1625 static int run(mddev_t
*mddev
)
1629 mirror_info_t
*disk
;
1631 struct list_head
*tmp
;
1633 sector_t stride
, size
;
1635 if (mddev
->level
!= 10) {
1636 printk(KERN_ERR
"raid10: %s: raid level not set correctly... (%d)\n",
1637 mdname(mddev
), mddev
->level
);
1640 nc
= mddev
->layout
& 255;
1641 fc
= (mddev
->layout
>> 8) & 255;
1642 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
1643 (mddev
->layout
>> 16)) {
1644 printk(KERN_ERR
"raid10: %s: unsupported raid10 layout: 0x%8x\n",
1645 mdname(mddev
), mddev
->layout
);
1649 * copy the already verified devices into our private RAID10
1650 * bookkeeping area. [whatever we allocate in run(),
1651 * should be freed in stop()]
1653 conf
= kmalloc(sizeof(conf_t
), GFP_KERNEL
);
1654 mddev
->private = conf
;
1656 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
1660 memset(conf
, 0, sizeof(*conf
));
1661 conf
->mirrors
= kmalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
1663 if (!conf
->mirrors
) {
1664 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
1668 memset(conf
->mirrors
, 0, sizeof(struct mirror_info
)*mddev
->raid_disks
);
1670 conf
->near_copies
= nc
;
1671 conf
->far_copies
= fc
;
1672 conf
->copies
= nc
*fc
;
1673 conf
->chunk_mask
= (sector_t
)(mddev
->chunk_size
>>9)-1;
1674 conf
->chunk_shift
= ffz(~mddev
->chunk_size
) - 9;
1675 stride
= mddev
->size
>> (conf
->chunk_shift
-1);
1676 sector_div(stride
, fc
);
1677 conf
->stride
= stride
<< conf
->chunk_shift
;
1679 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
1680 r10bio_pool_free
, conf
);
1681 if (!conf
->r10bio_pool
) {
1682 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
1687 ITERATE_RDEV(mddev
, rdev
, tmp
) {
1688 disk_idx
= rdev
->raid_disk
;
1689 if (disk_idx
>= mddev
->raid_disks
1692 disk
= conf
->mirrors
+ disk_idx
;
1696 blk_queue_stack_limits(mddev
->queue
,
1697 rdev
->bdev
->bd_disk
->queue
);
1698 /* as we don't honour merge_bvec_fn, we must never risk
1699 * violating it, so limit ->max_sector to one PAGE, as
1700 * a one page request is never in violation.
1702 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
1703 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
1704 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
1706 disk
->head_position
= 0;
1707 if (!test_bit(Faulty
, &rdev
->flags
) && test_bit(In_sync
, &rdev
->flags
))
1708 conf
->working_disks
++;
1710 conf
->raid_disks
= mddev
->raid_disks
;
1711 conf
->mddev
= mddev
;
1712 spin_lock_init(&conf
->device_lock
);
1713 INIT_LIST_HEAD(&conf
->retry_list
);
1715 spin_lock_init(&conf
->resync_lock
);
1716 init_waitqueue_head(&conf
->wait_idle
);
1717 init_waitqueue_head(&conf
->wait_resume
);
1719 /* need to check that every block has at least one working mirror */
1720 if (!enough(conf
)) {
1721 printk(KERN_ERR
"raid10: not enough operational mirrors for %s\n",
1726 mddev
->degraded
= 0;
1727 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1729 disk
= conf
->mirrors
+ i
;
1732 disk
->head_position
= 0;
1738 mddev
->thread
= md_register_thread(raid10d
, mddev
, "%s_raid10");
1739 if (!mddev
->thread
) {
1741 "raid10: couldn't allocate thread for %s\n",
1747 "raid10: raid set %s active with %d out of %d devices\n",
1748 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
1751 * Ok, everything is just fine now
1753 size
= conf
->stride
* conf
->raid_disks
;
1754 sector_div(size
, conf
->near_copies
);
1755 mddev
->array_size
= size
/2;
1756 mddev
->resync_max_sectors
= size
;
1758 mddev
->queue
->unplug_fn
= raid10_unplug
;
1759 mddev
->queue
->issue_flush_fn
= raid10_issue_flush
;
1761 /* Calculate max read-ahead size.
1762 * We need to readahead at least twice a whole stripe....
1766 int stripe
= conf
->raid_disks
* mddev
->chunk_size
/ PAGE_CACHE_SIZE
;
1767 stripe
/= conf
->near_copies
;
1768 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
1769 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
1772 if (conf
->near_copies
< mddev
->raid_disks
)
1773 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
1777 if (conf
->r10bio_pool
)
1778 mempool_destroy(conf
->r10bio_pool
);
1779 kfree(conf
->mirrors
);
1781 mddev
->private = NULL
;
1786 static int stop(mddev_t
*mddev
)
1788 conf_t
*conf
= mddev_to_conf(mddev
);
1790 md_unregister_thread(mddev
->thread
);
1791 mddev
->thread
= NULL
;
1792 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
1793 if (conf
->r10bio_pool
)
1794 mempool_destroy(conf
->r10bio_pool
);
1795 kfree(conf
->mirrors
);
1797 mddev
->private = NULL
;
1802 static mdk_personality_t raid10_personality
=
1805 .owner
= THIS_MODULE
,
1806 .make_request
= make_request
,
1810 .error_handler
= error
,
1811 .hot_add_disk
= raid10_add_disk
,
1812 .hot_remove_disk
= raid10_remove_disk
,
1813 .spare_active
= raid10_spare_active
,
1814 .sync_request
= sync_request
,
1817 static int __init
raid_init(void)
1819 return register_md_personality(RAID10
, &raid10_personality
);
1822 static void raid_exit(void)
1824 unregister_md_personality(RAID10
);
1827 module_init(raid_init
);
1828 module_exit(raid_exit
);
1829 MODULE_LICENSE("GPL");
1830 MODULE_ALIAS("md-personality-9"); /* RAID10 */