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 there 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(int 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(int 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
;
500 raid10_find_phys(conf
, r10_bio
);
501 spin_lock_irq(&conf
->device_lock
);
503 * Check if we can balance. We can balance on the whole
504 * device if no resync is going on, or below the resync window.
505 * We take the first readable disk when above the resync window.
507 if (conf
->mddev
->recovery_cp
< MaxSector
508 && (this_sector
+ sectors
>= conf
->next_resync
)) {
509 /* make sure that disk is operational */
511 disk
= r10_bio
->devs
[slot
].devnum
;
513 while (!conf
->mirrors
[disk
].rdev
||
514 !conf
->mirrors
[disk
].rdev
->in_sync
) {
516 if (slot
== conf
->copies
) {
521 disk
= r10_bio
->devs
[slot
].devnum
;
527 /* make sure the disk is operational */
529 disk
= r10_bio
->devs
[slot
].devnum
;
530 while (!conf
->mirrors
[disk
].rdev
||
531 !conf
->mirrors
[disk
].rdev
->in_sync
) {
533 if (slot
== conf
->copies
) {
537 disk
= r10_bio
->devs
[slot
].devnum
;
541 current_distance
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
543 /* Find the disk whose head is closest */
545 for (nslot
= slot
; nslot
< conf
->copies
; nslot
++) {
546 int ndisk
= r10_bio
->devs
[nslot
].devnum
;
549 if (!conf
->mirrors
[ndisk
].rdev
||
550 !conf
->mirrors
[ndisk
].rdev
->in_sync
)
553 if (!atomic_read(&conf
->mirrors
[ndisk
].rdev
->nr_pending
)) {
558 new_distance
= abs(r10_bio
->devs
[nslot
].addr
-
559 conf
->mirrors
[ndisk
].head_position
);
560 if (new_distance
< current_distance
) {
561 current_distance
= new_distance
;
568 r10_bio
->read_slot
= slot
;
569 /* conf->next_seq_sect = this_sector + sectors;*/
571 if (disk
>= 0 && conf
->mirrors
[disk
].rdev
)
572 atomic_inc(&conf
->mirrors
[disk
].rdev
->nr_pending
);
573 spin_unlock_irq(&conf
->device_lock
);
578 static void unplug_slaves(mddev_t
*mddev
)
580 conf_t
*conf
= mddev_to_conf(mddev
);
584 spin_lock_irqsave(&conf
->device_lock
, flags
);
585 for (i
=0; i
<mddev
->raid_disks
; i
++) {
586 mdk_rdev_t
*rdev
= conf
->mirrors
[i
].rdev
;
587 if (rdev
&& atomic_read(&rdev
->nr_pending
)) {
588 request_queue_t
*r_queue
= bdev_get_queue(rdev
->bdev
);
590 atomic_inc(&rdev
->nr_pending
);
591 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
593 if (r_queue
->unplug_fn
)
594 r_queue
->unplug_fn(r_queue
);
596 spin_lock_irqsave(&conf
->device_lock
, flags
);
597 atomic_dec(&rdev
->nr_pending
);
600 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
602 static void raid10_unplug(request_queue_t
*q
)
604 unplug_slaves(q
->queuedata
);
607 static int raid10_issue_flush(request_queue_t
*q
, struct gendisk
*disk
,
608 sector_t
*error_sector
)
610 mddev_t
*mddev
= q
->queuedata
;
611 conf_t
*conf
= mddev_to_conf(mddev
);
615 spin_lock_irqsave(&conf
->device_lock
, flags
);
616 for (i
=0; i
<mddev
->raid_disks
; i
++) {
617 mdk_rdev_t
*rdev
= conf
->mirrors
[i
].rdev
;
618 if (rdev
&& !rdev
->faulty
) {
619 struct block_device
*bdev
= rdev
->bdev
;
620 request_queue_t
*r_queue
= bdev_get_queue(bdev
);
622 if (r_queue
->issue_flush_fn
) {
623 ret
= r_queue
->issue_flush_fn(r_queue
, bdev
->bd_disk
, error_sector
);
629 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
634 * Throttle resync depth, so that we can both get proper overlapping of
635 * requests, but are still able to handle normal requests quickly.
637 #define RESYNC_DEPTH 32
639 static void device_barrier(conf_t
*conf
, sector_t sect
)
641 spin_lock_irq(&conf
->resync_lock
);
642 wait_event_lock_irq(conf
->wait_idle
, !waitqueue_active(&conf
->wait_resume
),
643 conf
->resync_lock
, unplug_slaves(conf
->mddev
));
645 if (!conf
->barrier
++) {
646 wait_event_lock_irq(conf
->wait_idle
, !conf
->nr_pending
,
647 conf
->resync_lock
, unplug_slaves(conf
->mddev
));
648 if (conf
->nr_pending
)
651 wait_event_lock_irq(conf
->wait_resume
, conf
->barrier
< RESYNC_DEPTH
,
652 conf
->resync_lock
, unplug_slaves(conf
->mddev
));
653 conf
->next_resync
= sect
;
654 spin_unlock_irq(&conf
->resync_lock
);
657 static int make_request(request_queue_t
*q
, struct bio
* bio
)
659 mddev_t
*mddev
= q
->queuedata
;
660 conf_t
*conf
= mddev_to_conf(mddev
);
661 mirror_info_t
*mirror
;
663 struct bio
*read_bio
;
665 int chunk_sects
= conf
->chunk_mask
+ 1;
667 /* If this request crosses a chunk boundary, we need to
668 * split it. This will only happen for 1 PAGE (or less) requests.
670 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
672 conf
->near_copies
< conf
->raid_disks
)) {
674 /* Sanity check -- queue functions should prevent this happening */
675 if (bio
->bi_vcnt
!= 1 ||
678 /* This is a one page bio that upper layers
679 * refuse to split for us, so we need to split it.
681 bp
= bio_split(bio
, bio_split_pool
,
682 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
683 if (make_request(q
, &bp
->bio1
))
684 generic_make_request(&bp
->bio1
);
685 if (make_request(q
, &bp
->bio2
))
686 generic_make_request(&bp
->bio2
);
688 bio_pair_release(bp
);
691 printk("raid10_make_request bug: can't convert block across chunks"
692 " or bigger than %dk %llu %d\n", chunk_sects
/2,
693 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
695 bio_io_error(bio
, bio
->bi_size
);
700 * Register the new request and wait if the reconstruction
701 * thread has put up a bar for new requests.
702 * Continue immediately if no resync is active currently.
704 spin_lock_irq(&conf
->resync_lock
);
705 wait_event_lock_irq(conf
->wait_resume
, !conf
->barrier
, conf
->resync_lock
, );
707 spin_unlock_irq(&conf
->resync_lock
);
709 if (bio_data_dir(bio
)==WRITE
) {
710 disk_stat_inc(mddev
->gendisk
, writes
);
711 disk_stat_add(mddev
->gendisk
, write_sectors
, bio_sectors(bio
));
713 disk_stat_inc(mddev
->gendisk
, reads
);
714 disk_stat_add(mddev
->gendisk
, read_sectors
, bio_sectors(bio
));
717 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
719 r10_bio
->master_bio
= bio
;
720 r10_bio
->sectors
= bio
->bi_size
>> 9;
722 r10_bio
->mddev
= mddev
;
723 r10_bio
->sector
= bio
->bi_sector
;
725 if (bio_data_dir(bio
) == READ
) {
727 * read balancing logic:
729 int disk
= read_balance(conf
, r10_bio
);
730 int slot
= r10_bio
->read_slot
;
732 raid_end_bio_io(r10_bio
);
735 mirror
= conf
->mirrors
+ disk
;
737 read_bio
= bio_clone(bio
, GFP_NOIO
);
739 r10_bio
->devs
[slot
].bio
= read_bio
;
741 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
742 mirror
->rdev
->data_offset
;
743 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
744 read_bio
->bi_end_io
= raid10_end_read_request
;
745 read_bio
->bi_rw
= READ
;
746 read_bio
->bi_private
= r10_bio
;
748 generic_make_request(read_bio
);
755 /* first select target devices under spinlock and
756 * inc refcount on their rdev. Record them by setting
759 raid10_find_phys(conf
, r10_bio
);
760 spin_lock_irq(&conf
->device_lock
);
761 for (i
= 0; i
< conf
->copies
; i
++) {
762 int d
= r10_bio
->devs
[i
].devnum
;
763 if (conf
->mirrors
[d
].rdev
&&
764 !conf
->mirrors
[d
].rdev
->faulty
) {
765 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
766 r10_bio
->devs
[i
].bio
= bio
;
768 r10_bio
->devs
[i
].bio
= NULL
;
770 spin_unlock_irq(&conf
->device_lock
);
772 atomic_set(&r10_bio
->remaining
, 1);
773 md_write_start(mddev
);
774 for (i
= 0; i
< conf
->copies
; i
++) {
776 int d
= r10_bio
->devs
[i
].devnum
;
777 if (!r10_bio
->devs
[i
].bio
)
780 mbio
= bio_clone(bio
, GFP_NOIO
);
781 r10_bio
->devs
[i
].bio
= mbio
;
783 mbio
->bi_sector
= r10_bio
->devs
[i
].addr
+
784 conf
->mirrors
[d
].rdev
->data_offset
;
785 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
786 mbio
->bi_end_io
= raid10_end_write_request
;
788 mbio
->bi_private
= r10_bio
;
790 atomic_inc(&r10_bio
->remaining
);
791 generic_make_request(mbio
);
794 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
796 raid_end_bio_io(r10_bio
);
802 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
804 conf_t
*conf
= mddev_to_conf(mddev
);
807 if (conf
->near_copies
< conf
->raid_disks
)
808 seq_printf(seq
, " %dK chunks", mddev
->chunk_size
/1024);
809 if (conf
->near_copies
> 1)
810 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
811 if (conf
->far_copies
> 1)
812 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
814 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
815 conf
->working_disks
);
816 for (i
= 0; i
< conf
->raid_disks
; i
++)
817 seq_printf(seq
, "%s",
818 conf
->mirrors
[i
].rdev
&&
819 conf
->mirrors
[i
].rdev
->in_sync
? "U" : "_");
820 seq_printf(seq
, "]");
823 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
825 char b
[BDEVNAME_SIZE
];
826 conf_t
*conf
= mddev_to_conf(mddev
);
829 * If it is not operational, then we have already marked it as dead
830 * else if it is the last working disks, ignore the error, let the
831 * next level up know.
832 * else mark the drive as failed
835 && conf
->working_disks
== 1)
837 * Don't fail the drive, just return an IO error.
838 * The test should really be more sophisticated than
839 * "working_disks == 1", but it isn't critical, and
840 * can wait until we do more sophisticated "is the drive
841 * really dead" tests...
846 conf
->working_disks
--;
848 * if recovery is running, make sure it aborts.
850 set_bit(MD_RECOVERY_ERR
, &mddev
->recovery
);
855 printk(KERN_ALERT
"raid10: Disk failure on %s, disabling device. \n"
856 " Operation continuing on %d devices\n",
857 bdevname(rdev
->bdev
,b
), conf
->working_disks
);
860 static void print_conf(conf_t
*conf
)
865 printk("RAID10 conf printout:\n");
870 printk(" --- wd:%d rd:%d\n", conf
->working_disks
,
873 for (i
= 0; i
< conf
->raid_disks
; i
++) {
874 char b
[BDEVNAME_SIZE
];
875 tmp
= conf
->mirrors
+ i
;
877 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
878 i
, !tmp
->rdev
->in_sync
, !tmp
->rdev
->faulty
,
879 bdevname(tmp
->rdev
->bdev
,b
));
883 static void close_sync(conf_t
*conf
)
885 spin_lock_irq(&conf
->resync_lock
);
886 wait_event_lock_irq(conf
->wait_resume
, !conf
->barrier
,
887 conf
->resync_lock
, unplug_slaves(conf
->mddev
));
888 spin_unlock_irq(&conf
->resync_lock
);
890 if (conf
->barrier
) BUG();
891 if (waitqueue_active(&conf
->wait_idle
)) BUG();
893 mempool_destroy(conf
->r10buf_pool
);
894 conf
->r10buf_pool
= NULL
;
897 static int raid10_spare_active(mddev_t
*mddev
)
900 conf_t
*conf
= mddev
->private;
903 spin_lock_irq(&conf
->device_lock
);
905 * Find all non-in_sync disks within the RAID10 configuration
906 * and mark them in_sync
908 for (i
= 0; i
< conf
->raid_disks
; i
++) {
909 tmp
= conf
->mirrors
+ i
;
911 && !tmp
->rdev
->faulty
912 && !tmp
->rdev
->in_sync
) {
913 conf
->working_disks
++;
915 tmp
->rdev
->in_sync
= 1;
918 spin_unlock_irq(&conf
->device_lock
);
925 static int raid10_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
927 conf_t
*conf
= mddev
->private;
932 if (mddev
->recovery_cp
< MaxSector
)
933 /* only hot-add to in-sync arrays, as recovery is
934 * very different from resync
937 spin_lock_irq(&conf
->device_lock
);
938 for (mirror
=0; mirror
< mddev
->raid_disks
; mirror
++)
939 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
942 blk_queue_stack_limits(mddev
->queue
,
943 rdev
->bdev
->bd_disk
->queue
);
944 /* as we don't honour merge_bvec_fn, we must never risk
945 * violating it, so limit ->max_sector to one PAGE, as
946 * a one page request is never in violation.
948 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
949 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
950 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
952 p
->head_position
= 0;
953 rdev
->raid_disk
= mirror
;
957 spin_unlock_irq(&conf
->device_lock
);
963 static int raid10_remove_disk(mddev_t
*mddev
, int number
)
965 conf_t
*conf
= mddev
->private;
967 mirror_info_t
*p
= conf
->mirrors
+ number
;
970 spin_lock_irq(&conf
->device_lock
);
972 if (p
->rdev
->in_sync
||
973 atomic_read(&p
->rdev
->nr_pending
)) {
983 spin_unlock_irq(&conf
->device_lock
);
990 static int end_sync_read(struct bio
*bio
, unsigned int bytes_done
, int error
)
992 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
993 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
994 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
1000 for (i
=0; i
<conf
->copies
; i
++)
1001 if (r10_bio
->devs
[i
].bio
== bio
)
1003 if (i
== conf
->copies
)
1005 update_head_pos(i
, r10_bio
);
1006 d
= r10_bio
->devs
[i
].devnum
;
1008 md_error(r10_bio
->mddev
,
1009 conf
->mirrors
[d
].rdev
);
1011 /* for reconstruct, we always reschedule after a read.
1012 * for resync, only after all reads
1014 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1015 atomic_dec_and_test(&r10_bio
->remaining
)) {
1016 /* we have read all the blocks,
1017 * do the comparison in process context in raid10d
1019 reschedule_retry(r10_bio
);
1021 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1025 static int end_sync_write(struct bio
*bio
, unsigned int bytes_done
, int error
)
1027 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1028 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1029 mddev_t
*mddev
= r10_bio
->mddev
;
1030 conf_t
*conf
= mddev_to_conf(mddev
);
1036 for (i
= 0; i
< conf
->copies
; i
++)
1037 if (r10_bio
->devs
[i
].bio
== bio
)
1039 d
= r10_bio
->devs
[i
].devnum
;
1042 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1043 update_head_pos(i
, r10_bio
);
1045 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1046 if (r10_bio
->master_bio
== NULL
) {
1047 /* the primary of several recovery bios */
1048 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1052 r10bio_t
*r10_bio2
= (r10bio_t
*)r10_bio
->master_bio
;
1057 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1062 * Note: sync and recover and handled very differently for raid10
1063 * This code is for resync.
1064 * For resync, we read through virtual addresses and read all blocks.
1065 * If there is any error, we schedule a write. The lowest numbered
1066 * drive is authoritative.
1067 * However requests come for physical address, so we need to map.
1068 * For every physical address there are raid_disks/copies virtual addresses,
1069 * which is always are least one, but is not necessarly an integer.
1070 * This means that a physical address can span multiple chunks, so we may
1071 * have to submit multiple io requests for a single sync request.
1074 * We check if all blocks are in-sync and only write to blocks that
1077 static void sync_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1079 conf_t
*conf
= mddev_to_conf(mddev
);
1081 struct bio
*tbio
, *fbio
;
1083 atomic_set(&r10_bio
->remaining
, 1);
1085 /* find the first device with a block */
1086 for (i
=0; i
<conf
->copies
; i
++)
1087 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1090 if (i
== conf
->copies
)
1094 fbio
= r10_bio
->devs
[i
].bio
;
1096 /* now find blocks with errors */
1097 for (i
=first
+1 ; i
< conf
->copies
; i
++) {
1100 if (!test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1102 /* We know that the bi_io_vec layout is the same for
1103 * both 'first' and 'i', so we just compare them.
1104 * All vec entries are PAGE_SIZE;
1106 tbio
= r10_bio
->devs
[i
].bio
;
1107 vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1108 for (j
= 0; j
< vcnt
; j
++)
1109 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1110 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1115 /* Ok, we need to write this bio
1116 * First we need to fixup bv_offset, bv_len and
1117 * bi_vecs, as the read request might have corrupted these
1119 tbio
->bi_vcnt
= vcnt
;
1120 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1122 tbio
->bi_phys_segments
= 0;
1123 tbio
->bi_hw_segments
= 0;
1124 tbio
->bi_hw_front_size
= 0;
1125 tbio
->bi_hw_back_size
= 0;
1126 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1127 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1128 tbio
->bi_next
= NULL
;
1129 tbio
->bi_rw
= WRITE
;
1130 tbio
->bi_private
= r10_bio
;
1131 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1133 for (j
=0; j
< vcnt
; j
++) {
1134 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1135 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1137 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1138 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1141 tbio
->bi_end_io
= end_sync_write
;
1143 d
= r10_bio
->devs
[i
].devnum
;
1144 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1145 atomic_inc(&r10_bio
->remaining
);
1146 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1148 generic_make_request(tbio
);
1152 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1153 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1159 * Now for the recovery code.
1160 * Recovery happens across physical sectors.
1161 * We recover all non-is_sync drives by finding the virtual address of
1162 * each, and then choose a working drive that also has that virt address.
1163 * There is a separate r10_bio for each non-in_sync drive.
1164 * Only the first two slots are in use. The first for reading,
1165 * The second for writing.
1169 static void recovery_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1171 conf_t
*conf
= mddev_to_conf(mddev
);
1173 struct bio
*bio
, *wbio
;
1176 /* move the pages across to the second bio
1177 * and submit the write request
1179 bio
= r10_bio
->devs
[0].bio
;
1180 wbio
= r10_bio
->devs
[1].bio
;
1181 for (i
=0; i
< wbio
->bi_vcnt
; i
++) {
1182 struct page
*p
= bio
->bi_io_vec
[i
].bv_page
;
1183 bio
->bi_io_vec
[i
].bv_page
= wbio
->bi_io_vec
[i
].bv_page
;
1184 wbio
->bi_io_vec
[i
].bv_page
= p
;
1186 d
= r10_bio
->devs
[1].devnum
;
1188 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1189 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1190 generic_make_request(wbio
);
1195 * This is a kernel thread which:
1197 * 1. Retries failed read operations on working mirrors.
1198 * 2. Updates the raid superblock when problems encounter.
1199 * 3. Performs writes following reads for array syncronising.
1202 static void raid10d(mddev_t
*mddev
)
1206 unsigned long flags
;
1207 conf_t
*conf
= mddev_to_conf(mddev
);
1208 struct list_head
*head
= &conf
->retry_list
;
1212 md_check_recovery(mddev
);
1213 md_handle_safemode(mddev
);
1216 char b
[BDEVNAME_SIZE
];
1217 spin_lock_irqsave(&conf
->device_lock
, flags
);
1218 if (list_empty(head
))
1220 r10_bio
= list_entry(head
->prev
, r10bio_t
, retry_list
);
1221 list_del(head
->prev
);
1222 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1224 mddev
= r10_bio
->mddev
;
1225 conf
= mddev_to_conf(mddev
);
1226 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
)) {
1227 sync_request_write(mddev
, r10_bio
);
1229 } else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
1230 recovery_request_write(mddev
, r10_bio
);
1234 bio
= r10_bio
->devs
[r10_bio
->read_slot
].bio
;
1235 r10_bio
->devs
[r10_bio
->read_slot
].bio
= NULL
;
1236 mirror
= read_balance(conf
, r10_bio
);
1237 r10_bio
->devs
[r10_bio
->read_slot
].bio
= bio
;
1239 printk(KERN_ALERT
"raid10: %s: unrecoverable I/O"
1240 " read error for block %llu\n",
1241 bdevname(bio
->bi_bdev
,b
),
1242 (unsigned long long)r10_bio
->sector
);
1243 raid_end_bio_io(r10_bio
);
1245 rdev
= conf
->mirrors
[mirror
].rdev
;
1246 if (printk_ratelimit())
1247 printk(KERN_ERR
"raid10: %s: redirecting sector %llu to"
1248 " another mirror\n",
1249 bdevname(rdev
->bdev
,b
),
1250 (unsigned long long)r10_bio
->sector
);
1251 bio
->bi_bdev
= rdev
->bdev
;
1252 bio
->bi_sector
= r10_bio
->devs
[r10_bio
->read_slot
].addr
1253 + rdev
->data_offset
;
1254 bio
->bi_next
= NULL
;
1255 bio
->bi_flags
&= (1<<BIO_CLONED
);
1256 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
1258 bio
->bi_size
= r10_bio
->sectors
<< 9;
1261 generic_make_request(bio
);
1265 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1267 unplug_slaves(mddev
);
1271 static int init_resync(conf_t
*conf
)
1275 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1276 if (conf
->r10buf_pool
)
1278 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
1279 if (!conf
->r10buf_pool
)
1281 conf
->next_resync
= 0;
1286 * perform a "sync" on one "block"
1288 * We need to make sure that no normal I/O request - particularly write
1289 * requests - conflict with active sync requests.
1291 * This is achieved by tracking pending requests and a 'barrier' concept
1292 * that can be installed to exclude normal IO requests.
1294 * Resync and recovery are handled very differently.
1295 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1297 * For resync, we iterate over virtual addresses, read all copies,
1298 * and update if there are differences. If only one copy is live,
1300 * For recovery, we iterate over physical addresses, read a good
1301 * value for each non-in_sync drive, and over-write.
1303 * So, for recovery we may have several outstanding complex requests for a
1304 * given address, one for each out-of-sync device. We model this by allocating
1305 * a number of r10_bio structures, one for each out-of-sync device.
1306 * As we setup these structures, we collect all bio's together into a list
1307 * which we then process collectively to add pages, and then process again
1308 * to pass to generic_make_request.
1310 * The r10_bio structures are linked using a borrowed master_bio pointer.
1311 * This link is counted in ->remaining. When the r10_bio that points to NULL
1312 * has its remaining count decremented to 0, the whole complex operation
1317 static int sync_request(mddev_t
*mddev
, sector_t sector_nr
, int go_faster
)
1319 conf_t
*conf
= mddev_to_conf(mddev
);
1321 struct bio
*biolist
= NULL
, *bio
;
1322 sector_t max_sector
, nr_sectors
;
1326 sector_t sectors_skipped
= 0;
1327 int chunks_skipped
= 0;
1329 if (!conf
->r10buf_pool
)
1330 if (init_resync(conf
))
1334 max_sector
= mddev
->size
<< 1;
1335 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1336 max_sector
= mddev
->resync_max_sectors
;
1337 if (sector_nr
>= max_sector
) {
1339 return sectors_skipped
;
1341 if (chunks_skipped
>= conf
->raid_disks
) {
1342 /* if there has been nothing to do on any drive,
1343 * then there is nothing to do at all..
1345 sector_t sec
= max_sector
- sector_nr
;
1346 md_done_sync(mddev
, sec
, 1);
1347 return sec
+ sectors_skipped
;
1350 /* make sure whole request will fit in a chunk - if chunks
1353 if (conf
->near_copies
< conf
->raid_disks
&&
1354 max_sector
> (sector_nr
| conf
->chunk_mask
))
1355 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
1357 * If there is non-resync activity waiting for us then
1358 * put in a delay to throttle resync.
1360 if (!go_faster
&& waitqueue_active(&conf
->wait_resume
))
1361 schedule_timeout(HZ
);
1362 device_barrier(conf
, sector_nr
+ RESYNC_SECTORS
);
1364 /* Again, very different code for resync and recovery.
1365 * Both must result in an r10bio with a list of bios that
1366 * have bi_end_io, bi_sector, bi_bdev set,
1367 * and bi_private set to the r10bio.
1368 * For recovery, we may actually create several r10bios
1369 * with 2 bios in each, that correspond to the bios in the main one.
1370 * In this case, the subordinate r10bios link back through a
1371 * borrowed master_bio pointer, and the counter in the master
1372 * includes a ref from each subordinate.
1374 /* First, we decide what to do and set ->bi_end_io
1375 * To end_sync_read if we want to read, and
1376 * end_sync_write if we will want to write.
1379 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
1380 /* recovery... the complicated one */
1384 for (i
=0 ; i
<conf
->raid_disks
; i
++)
1385 if (conf
->mirrors
[i
].rdev
&&
1386 !conf
->mirrors
[i
].rdev
->in_sync
) {
1387 /* want to reconstruct this device */
1388 r10bio_t
*rb2
= r10_bio
;
1390 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1391 spin_lock_irq(&conf
->resync_lock
);
1393 if (rb2
) conf
->barrier
++;
1394 spin_unlock_irq(&conf
->resync_lock
);
1395 atomic_set(&r10_bio
->remaining
, 0);
1397 r10_bio
->master_bio
= (struct bio
*)rb2
;
1399 atomic_inc(&rb2
->remaining
);
1400 r10_bio
->mddev
= mddev
;
1401 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
1402 r10_bio
->sector
= raid10_find_virt(conf
, sector_nr
, i
);
1403 raid10_find_phys(conf
, r10_bio
);
1404 for (j
=0; j
<conf
->copies
;j
++) {
1405 int d
= r10_bio
->devs
[j
].devnum
;
1406 if (conf
->mirrors
[d
].rdev
&&
1407 conf
->mirrors
[d
].rdev
->in_sync
) {
1408 /* This is where we read from */
1409 bio
= r10_bio
->devs
[0].bio
;
1410 bio
->bi_next
= biolist
;
1412 bio
->bi_private
= r10_bio
;
1413 bio
->bi_end_io
= end_sync_read
;
1415 bio
->bi_sector
= r10_bio
->devs
[j
].addr
+
1416 conf
->mirrors
[d
].rdev
->data_offset
;
1417 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1418 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1419 atomic_inc(&r10_bio
->remaining
);
1420 /* and we write to 'i' */
1422 for (k
=0; k
<conf
->copies
; k
++)
1423 if (r10_bio
->devs
[k
].devnum
== i
)
1425 bio
= r10_bio
->devs
[1].bio
;
1426 bio
->bi_next
= biolist
;
1428 bio
->bi_private
= r10_bio
;
1429 bio
->bi_end_io
= end_sync_write
;
1431 bio
->bi_sector
= r10_bio
->devs
[k
].addr
+
1432 conf
->mirrors
[i
].rdev
->data_offset
;
1433 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1435 r10_bio
->devs
[0].devnum
= d
;
1436 r10_bio
->devs
[1].devnum
= i
;
1441 if (j
== conf
->copies
) {
1445 if (biolist
== NULL
) {
1447 r10bio_t
*rb2
= r10_bio
;
1448 r10_bio
= (r10bio_t
*) rb2
->master_bio
;
1449 rb2
->master_bio
= NULL
;
1455 /* resync. Schedule a read for every block at this virt offset */
1457 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1459 spin_lock_irq(&conf
->resync_lock
);
1461 spin_unlock_irq(&conf
->resync_lock
);
1463 r10_bio
->mddev
= mddev
;
1464 atomic_set(&r10_bio
->remaining
, 0);
1466 r10_bio
->master_bio
= NULL
;
1467 r10_bio
->sector
= sector_nr
;
1468 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
1469 raid10_find_phys(conf
, r10_bio
);
1470 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
1471 spin_lock_irq(&conf
->device_lock
);
1472 for (i
=0; i
<conf
->copies
; i
++) {
1473 int d
= r10_bio
->devs
[i
].devnum
;
1474 bio
= r10_bio
->devs
[i
].bio
;
1475 bio
->bi_end_io
= NULL
;
1476 if (conf
->mirrors
[d
].rdev
== NULL
||
1477 conf
->mirrors
[d
].rdev
->faulty
)
1479 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1480 atomic_inc(&r10_bio
->remaining
);
1481 bio
->bi_next
= biolist
;
1483 bio
->bi_private
= r10_bio
;
1484 bio
->bi_end_io
= end_sync_read
;
1486 bio
->bi_sector
= r10_bio
->devs
[i
].addr
+
1487 conf
->mirrors
[d
].rdev
->data_offset
;
1488 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1491 spin_unlock_irq(&conf
->device_lock
);
1493 for (i
=0; i
<conf
->copies
; i
++) {
1494 int d
= r10_bio
->devs
[i
].devnum
;
1495 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
1496 atomic_dec(&conf
->mirrors
[d
].rdev
->nr_pending
);
1503 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1505 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1507 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
1510 bio
->bi_phys_segments
= 0;
1511 bio
->bi_hw_segments
= 0;
1518 int len
= PAGE_SIZE
;
1520 if (sector_nr
+ (len
>>9) > max_sector
)
1521 len
= (max_sector
- sector_nr
) << 9;
1524 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1525 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
1526 if (bio_add_page(bio
, page
, len
, 0) == 0) {
1529 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
1530 for (bio2
= biolist
; bio2
&& bio2
!= bio
; bio2
= bio2
->bi_next
) {
1531 /* remove last page from this bio */
1533 bio2
->bi_size
-= len
;
1534 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
1540 nr_sectors
+= len
>>9;
1541 sector_nr
+= len
>>9;
1542 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
1544 r10_bio
->sectors
= nr_sectors
;
1548 biolist
= biolist
->bi_next
;
1550 bio
->bi_next
= NULL
;
1551 r10_bio
= bio
->bi_private
;
1552 r10_bio
->sectors
= nr_sectors
;
1554 if (bio
->bi_end_io
== end_sync_read
) {
1555 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1556 generic_make_request(bio
);
1562 /* There is nowhere to write, so all non-sync
1563 * drives must be failed, so try the next chunk...
1566 int sec
= max_sector
- sector_nr
;
1567 sectors_skipped
+= sec
;
1569 sector_nr
= max_sector
;
1570 md_done_sync(mddev
, sec
, 1);
1575 static int run(mddev_t
*mddev
)
1579 mirror_info_t
*disk
;
1581 struct list_head
*tmp
;
1583 sector_t stride
, size
;
1585 if (mddev
->level
!= 10) {
1586 printk(KERN_ERR
"raid10: %s: raid level not set correctly... (%d)\n",
1587 mdname(mddev
), mddev
->level
);
1590 nc
= mddev
->layout
& 255;
1591 fc
= (mddev
->layout
>> 8) & 255;
1592 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
1593 (mddev
->layout
>> 16)) {
1594 printk(KERN_ERR
"raid10: %s: unsupported raid10 layout: 0x%8x\n",
1595 mdname(mddev
), mddev
->layout
);
1599 * copy the already verified devices into our private RAID10
1600 * bookkeeping area. [whatever we allocate in run(),
1601 * should be freed in stop()]
1603 conf
= kmalloc(sizeof(conf_t
), GFP_KERNEL
);
1604 mddev
->private = conf
;
1606 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
1610 memset(conf
, 0, sizeof(*conf
));
1611 conf
->mirrors
= kmalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
1613 if (!conf
->mirrors
) {
1614 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
1618 memset(conf
->mirrors
, 0, sizeof(struct mirror_info
)*mddev
->raid_disks
);
1620 conf
->near_copies
= nc
;
1621 conf
->far_copies
= fc
;
1622 conf
->copies
= nc
*fc
;
1623 conf
->chunk_mask
= (sector_t
)(mddev
->chunk_size
>>9)-1;
1624 conf
->chunk_shift
= ffz(~mddev
->chunk_size
) - 9;
1625 stride
= mddev
->size
>> (conf
->chunk_shift
-1);
1626 sector_div(stride
, fc
);
1627 conf
->stride
= stride
<< conf
->chunk_shift
;
1629 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
1630 r10bio_pool_free
, conf
);
1631 if (!conf
->r10bio_pool
) {
1632 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
1636 mddev
->queue
->unplug_fn
= raid10_unplug
;
1638 mddev
->queue
->issue_flush_fn
= raid10_issue_flush
;
1640 ITERATE_RDEV(mddev
, rdev
, tmp
) {
1641 disk_idx
= rdev
->raid_disk
;
1642 if (disk_idx
>= mddev
->raid_disks
1645 disk
= conf
->mirrors
+ disk_idx
;
1649 blk_queue_stack_limits(mddev
->queue
,
1650 rdev
->bdev
->bd_disk
->queue
);
1651 /* as we don't honour merge_bvec_fn, we must never risk
1652 * violating it, so limit ->max_sector to one PAGE, as
1653 * a one page request is never in violation.
1655 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
1656 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
1657 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
1659 disk
->head_position
= 0;
1660 if (!rdev
->faulty
&& rdev
->in_sync
)
1661 conf
->working_disks
++;
1663 conf
->raid_disks
= mddev
->raid_disks
;
1664 conf
->mddev
= mddev
;
1665 conf
->device_lock
= SPIN_LOCK_UNLOCKED
;
1666 INIT_LIST_HEAD(&conf
->retry_list
);
1668 conf
->resync_lock
= SPIN_LOCK_UNLOCKED
;
1669 init_waitqueue_head(&conf
->wait_idle
);
1670 init_waitqueue_head(&conf
->wait_resume
);
1672 if (!conf
->working_disks
) {
1673 printk(KERN_ERR
"raid10: no operational mirrors for %s\n",
1678 mddev
->degraded
= 0;
1679 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1681 disk
= conf
->mirrors
+ i
;
1684 disk
->head_position
= 0;
1690 mddev
->thread
= md_register_thread(raid10d
, mddev
, "%s_raid10");
1691 if (!mddev
->thread
) {
1693 "raid10: couldn't allocate thread for %s\n",
1699 "raid10: raid set %s active with %d out of %d devices\n",
1700 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
1703 * Ok, everything is just fine now
1705 size
= conf
->stride
* conf
->raid_disks
;
1706 sector_div(size
, conf
->near_copies
);
1707 mddev
->array_size
= size
/2;
1708 mddev
->resync_max_sectors
= size
;
1710 /* Calculate max read-ahead size.
1711 * We need to readahead at least twice a whole stripe....
1715 int stripe
= conf
->raid_disks
* mddev
->chunk_size
/ PAGE_CACHE_SIZE
;
1716 stripe
/= conf
->near_copies
;
1717 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
1718 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
1721 if (conf
->near_copies
< mddev
->raid_disks
)
1722 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
1726 if (conf
->r10bio_pool
)
1727 mempool_destroy(conf
->r10bio_pool
);
1729 kfree(conf
->mirrors
);
1731 mddev
->private = NULL
;
1736 static int stop(mddev_t
*mddev
)
1738 conf_t
*conf
= mddev_to_conf(mddev
);
1740 md_unregister_thread(mddev
->thread
);
1741 mddev
->thread
= NULL
;
1742 if (conf
->r10bio_pool
)
1743 mempool_destroy(conf
->r10bio_pool
);
1745 kfree(conf
->mirrors
);
1747 mddev
->private = NULL
;
1752 static mdk_personality_t raid10_personality
=
1755 .owner
= THIS_MODULE
,
1756 .make_request
= make_request
,
1760 .error_handler
= error
,
1761 .hot_add_disk
= raid10_add_disk
,
1762 .hot_remove_disk
= raid10_remove_disk
,
1763 .spare_active
= raid10_spare_active
,
1764 .sync_request
= sync_request
,
1767 static int __init
raid_init(void)
1769 return register_md_personality(RAID10
, &raid10_personality
);
1772 static void raid_exit(void)
1774 unregister_md_personality(RAID10
);
1777 module_init(raid_init
);
1778 module_exit(raid_exit
);
1779 MODULE_LICENSE("GPL");
1780 MODULE_ALIAS("md-personality-9"); /* RAID10 */