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 "dm-bio-list.h"
22 #include <linux/raid/raid10.h>
23 #include <linux/raid/bitmap.h>
26 * RAID10 provides a combination of RAID0 and RAID1 functionality.
27 * The layout of data is defined by
30 * near_copies (stored in low byte of layout)
31 * far_copies (stored in second byte of layout)
33 * The data to be stored is divided into chunks using chunksize.
34 * Each device is divided into far_copies sections.
35 * In each section, chunks are laid out in a style similar to raid0, but
36 * near_copies copies of each chunk is stored (each on a different drive).
37 * The starting device for each section is offset near_copies from the starting
38 * device of the previous section.
39 * Thus there are (near_copies*far_copies) of each chunk, and each is on a different
41 * near_copies and far_copies must be at least one, and their product is at most
46 * Number of guaranteed r10bios in case of extreme VM load:
48 #define NR_RAID10_BIOS 256
50 static void unplug_slaves(mddev_t
*mddev
);
52 static void allow_barrier(conf_t
*conf
);
53 static void lower_barrier(conf_t
*conf
);
55 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
59 int size
= offsetof(struct r10bio_s
, devs
[conf
->copies
]);
61 /* allocate a r10bio with room for raid_disks entries in the bios array */
62 r10_bio
= kzalloc(size
, gfp_flags
);
64 unplug_slaves(conf
->mddev
);
69 static void r10bio_pool_free(void *r10_bio
, void *data
)
74 #define RESYNC_BLOCK_SIZE (64*1024)
75 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
76 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
77 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
78 #define RESYNC_WINDOW (2048*1024)
81 * When performing a resync, we need to read and compare, so
82 * we need as many pages are there are copies.
83 * When performing a recovery, we need 2 bios, one for read,
84 * one for write (we recover only one drive per r10buf)
87 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
96 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
98 unplug_slaves(conf
->mddev
);
102 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
103 nalloc
= conf
->copies
; /* resync */
105 nalloc
= 2; /* recovery */
110 for (j
= nalloc
; j
-- ; ) {
111 bio
= bio_alloc(gfp_flags
, RESYNC_PAGES
);
114 r10_bio
->devs
[j
].bio
= bio
;
117 * Allocate RESYNC_PAGES data pages and attach them
120 for (j
= 0 ; j
< nalloc
; j
++) {
121 bio
= r10_bio
->devs
[j
].bio
;
122 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
123 page
= alloc_page(gfp_flags
);
127 bio
->bi_io_vec
[i
].bv_page
= page
;
135 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
137 for (i
= 0; i
< RESYNC_PAGES
; i
++)
138 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
141 while ( ++j
< nalloc
)
142 bio_put(r10_bio
->devs
[j
].bio
);
143 r10bio_pool_free(r10_bio
, conf
);
147 static void r10buf_pool_free(void *__r10_bio
, void *data
)
151 r10bio_t
*r10bio
= __r10_bio
;
154 for (j
=0; j
< conf
->copies
; j
++) {
155 struct bio
*bio
= r10bio
->devs
[j
].bio
;
157 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
158 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
159 bio
->bi_io_vec
[i
].bv_page
= NULL
;
164 r10bio_pool_free(r10bio
, conf
);
167 static void put_all_bios(conf_t
*conf
, r10bio_t
*r10_bio
)
171 for (i
= 0; i
< conf
->copies
; i
++) {
172 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
173 if (*bio
&& *bio
!= IO_BLOCKED
)
179 static inline void free_r10bio(r10bio_t
*r10_bio
)
181 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
184 * Wake up any possible resync thread that waits for the device
189 put_all_bios(conf
, r10_bio
);
190 mempool_free(r10_bio
, conf
->r10bio_pool
);
193 static inline void put_buf(r10bio_t
*r10_bio
)
195 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
197 mempool_free(r10_bio
, conf
->r10buf_pool
);
202 static void reschedule_retry(r10bio_t
*r10_bio
)
205 mddev_t
*mddev
= r10_bio
->mddev
;
206 conf_t
*conf
= mddev_to_conf(mddev
);
208 spin_lock_irqsave(&conf
->device_lock
, flags
);
209 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
211 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
213 md_wakeup_thread(mddev
->thread
);
217 * raid_end_bio_io() is called when we have finished servicing a mirrored
218 * operation and are ready to return a success/failure code to the buffer
221 static void raid_end_bio_io(r10bio_t
*r10_bio
)
223 struct bio
*bio
= r10_bio
->master_bio
;
225 bio_endio(bio
, bio
->bi_size
,
226 test_bit(R10BIO_Uptodate
, &r10_bio
->state
) ? 0 : -EIO
);
227 free_r10bio(r10_bio
);
231 * Update disk head position estimator based on IRQ completion info.
233 static inline void update_head_pos(int slot
, r10bio_t
*r10_bio
)
235 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
237 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
238 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
241 static int raid10_end_read_request(struct bio
*bio
, unsigned int bytes_done
, int error
)
243 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
244 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
246 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
251 slot
= r10_bio
->read_slot
;
252 dev
= r10_bio
->devs
[slot
].devnum
;
254 * this branch is our 'one mirror IO has finished' event handler:
256 update_head_pos(slot
, r10_bio
);
260 * Set R10BIO_Uptodate in our master bio, so that
261 * we will return a good error code to the higher
262 * levels even if IO on some other mirrored buffer fails.
264 * The 'master' represents the composite IO operation to
265 * user-side. So if something waits for IO, then it will
266 * wait for the 'master' bio.
268 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
269 raid_end_bio_io(r10_bio
);
274 char b
[BDEVNAME_SIZE
];
275 if (printk_ratelimit())
276 printk(KERN_ERR
"raid10: %s: rescheduling sector %llu\n",
277 bdevname(conf
->mirrors
[dev
].rdev
->bdev
,b
), (unsigned long long)r10_bio
->sector
);
278 reschedule_retry(r10_bio
);
281 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
285 static int raid10_end_write_request(struct bio
*bio
, unsigned int bytes_done
, int error
)
287 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
288 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
290 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
295 for (slot
= 0; slot
< conf
->copies
; slot
++)
296 if (r10_bio
->devs
[slot
].bio
== bio
)
298 dev
= r10_bio
->devs
[slot
].devnum
;
301 * this branch is our 'one mirror IO has finished' event handler:
304 md_error(r10_bio
->mddev
, conf
->mirrors
[dev
].rdev
);
305 /* an I/O failed, we can't clear the bitmap */
306 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
309 * Set R10BIO_Uptodate in our master bio, so that
310 * we will return a good error code for to the higher
311 * levels even if IO on some other mirrored buffer fails.
313 * The 'master' represents the composite IO operation to
314 * user-side. So if something waits for IO, then it will
315 * wait for the 'master' bio.
317 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
319 update_head_pos(slot
, r10_bio
);
323 * Let's see if all mirrored write operations have finished
326 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
327 /* clear the bitmap if all writes complete successfully */
328 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
330 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
332 md_write_end(r10_bio
->mddev
);
333 raid_end_bio_io(r10_bio
);
336 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
342 * RAID10 layout manager
343 * Aswell as the chunksize and raid_disks count, there are two
344 * parameters: near_copies and far_copies.
345 * near_copies * far_copies must be <= raid_disks.
346 * Normally one of these will be 1.
347 * If both are 1, we get raid0.
348 * If near_copies == raid_disks, we get raid1.
350 * Chunks are layed out in raid0 style with near_copies copies of the
351 * first chunk, followed by near_copies copies of the next chunk and
353 * If far_copies > 1, then after 1/far_copies of the array has been assigned
354 * as described above, we start again with a device offset of near_copies.
355 * So we effectively have another copy of the whole array further down all
356 * the drives, but with blocks on different drives.
357 * With this layout, and block is never stored twice on the one device.
359 * raid10_find_phys finds the sector offset of a given virtual sector
360 * on each device that it is on. If a block isn't on a device,
361 * that entry in the array is set to MaxSector.
363 * raid10_find_virt does the reverse mapping, from a device and a
364 * sector offset to a virtual address
367 static void raid10_find_phys(conf_t
*conf
, r10bio_t
*r10bio
)
377 /* now calculate first sector/dev */
378 chunk
= r10bio
->sector
>> conf
->chunk_shift
;
379 sector
= r10bio
->sector
& conf
->chunk_mask
;
381 chunk
*= conf
->near_copies
;
383 dev
= sector_div(stripe
, conf
->raid_disks
);
385 sector
+= stripe
<< conf
->chunk_shift
;
387 /* and calculate all the others */
388 for (n
=0; n
< conf
->near_copies
; n
++) {
391 r10bio
->devs
[slot
].addr
= sector
;
392 r10bio
->devs
[slot
].devnum
= d
;
395 for (f
= 1; f
< conf
->far_copies
; f
++) {
396 d
+= conf
->near_copies
;
397 if (d
>= conf
->raid_disks
)
398 d
-= conf
->raid_disks
;
400 r10bio
->devs
[slot
].devnum
= d
;
401 r10bio
->devs
[slot
].addr
= s
;
405 if (dev
>= conf
->raid_disks
) {
407 sector
+= (conf
->chunk_mask
+ 1);
410 BUG_ON(slot
!= conf
->copies
);
413 static sector_t
raid10_find_virt(conf_t
*conf
, sector_t sector
, int dev
)
415 sector_t offset
, chunk
, vchunk
;
417 while (sector
> conf
->stride
) {
418 sector
-= conf
->stride
;
419 if (dev
< conf
->near_copies
)
420 dev
+= conf
->raid_disks
- conf
->near_copies
;
422 dev
-= conf
->near_copies
;
425 offset
= sector
& conf
->chunk_mask
;
426 chunk
= sector
>> conf
->chunk_shift
;
427 vchunk
= chunk
* conf
->raid_disks
+ dev
;
428 sector_div(vchunk
, conf
->near_copies
);
429 return (vchunk
<< conf
->chunk_shift
) + offset
;
433 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
435 * @bio: the buffer head that's been built up so far
436 * @biovec: the request that could be merged to it.
438 * Return amount of bytes we can accept at this offset
439 * If near_copies == raid_disk, there are no striping issues,
440 * but in that case, the function isn't called at all.
442 static int raid10_mergeable_bvec(request_queue_t
*q
, struct bio
*bio
,
443 struct bio_vec
*bio_vec
)
445 mddev_t
*mddev
= q
->queuedata
;
446 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
448 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
449 unsigned int bio_sectors
= bio
->bi_size
>> 9;
451 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
452 if (max
< 0) max
= 0; /* bio_add cannot handle a negative return */
453 if (max
<= bio_vec
->bv_len
&& bio_sectors
== 0)
454 return bio_vec
->bv_len
;
460 * This routine returns the disk from which the requested read should
461 * be done. There is a per-array 'next expected sequential IO' sector
462 * number - if this matches on the next IO then we use the last disk.
463 * There is also a per-disk 'last know head position' sector that is
464 * maintained from IRQ contexts, both the normal and the resync IO
465 * completion handlers update this position correctly. If there is no
466 * perfect sequential match then we pick the disk whose head is closest.
468 * If there are 2 mirrors in the same 2 devices, performance degrades
469 * because position is mirror, not device based.
471 * The rdev for the device selected will have nr_pending incremented.
475 * FIXME: possibly should rethink readbalancing and do it differently
476 * depending on near_copies / far_copies geometry.
478 static int read_balance(conf_t
*conf
, r10bio_t
*r10_bio
)
480 const unsigned long this_sector
= r10_bio
->sector
;
481 int disk
, slot
, nslot
;
482 const int sectors
= r10_bio
->sectors
;
483 sector_t new_distance
, current_distance
;
486 raid10_find_phys(conf
, r10_bio
);
489 * Check if we can balance. We can balance on the whole
490 * device if no resync is going on (recovery is ok), or below
491 * the resync window. We take the first readable disk when
492 * above the resync window.
494 if (conf
->mddev
->recovery_cp
< MaxSector
495 && (this_sector
+ sectors
>= conf
->next_resync
)) {
496 /* make sure that disk is operational */
498 disk
= r10_bio
->devs
[slot
].devnum
;
500 while ((rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
501 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
502 !test_bit(In_sync
, &rdev
->flags
)) {
504 if (slot
== conf
->copies
) {
509 disk
= r10_bio
->devs
[slot
].devnum
;
515 /* make sure the disk is operational */
517 disk
= r10_bio
->devs
[slot
].devnum
;
518 while ((rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
519 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
520 !test_bit(In_sync
, &rdev
->flags
)) {
522 if (slot
== conf
->copies
) {
526 disk
= r10_bio
->devs
[slot
].devnum
;
530 current_distance
= abs(r10_bio
->devs
[slot
].addr
-
531 conf
->mirrors
[disk
].head_position
);
533 /* Find the disk whose head is closest */
535 for (nslot
= slot
; nslot
< conf
->copies
; nslot
++) {
536 int ndisk
= r10_bio
->devs
[nslot
].devnum
;
539 if ((rdev
=rcu_dereference(conf
->mirrors
[ndisk
].rdev
)) == NULL
||
540 r10_bio
->devs
[nslot
].bio
== IO_BLOCKED
||
541 !test_bit(In_sync
, &rdev
->flags
))
544 /* This optimisation is debatable, and completely destroys
545 * sequential read speed for 'far copies' arrays. So only
546 * keep it for 'near' arrays, and review those later.
548 if (conf
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
)) {
553 new_distance
= abs(r10_bio
->devs
[nslot
].addr
-
554 conf
->mirrors
[ndisk
].head_position
);
555 if (new_distance
< current_distance
) {
556 current_distance
= new_distance
;
563 r10_bio
->read_slot
= slot
;
564 /* conf->next_seq_sect = this_sector + sectors;*/
566 if (disk
>= 0 && (rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
))!= NULL
)
567 atomic_inc(&conf
->mirrors
[disk
].rdev
->nr_pending
);
573 static void unplug_slaves(mddev_t
*mddev
)
575 conf_t
*conf
= mddev_to_conf(mddev
);
579 for (i
=0; i
<mddev
->raid_disks
; i
++) {
580 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
581 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
582 request_queue_t
*r_queue
= bdev_get_queue(rdev
->bdev
);
584 atomic_inc(&rdev
->nr_pending
);
587 if (r_queue
->unplug_fn
)
588 r_queue
->unplug_fn(r_queue
);
590 rdev_dec_pending(rdev
, mddev
);
597 static void raid10_unplug(request_queue_t
*q
)
599 mddev_t
*mddev
= q
->queuedata
;
601 unplug_slaves(q
->queuedata
);
602 md_wakeup_thread(mddev
->thread
);
605 static int raid10_issue_flush(request_queue_t
*q
, struct gendisk
*disk
,
606 sector_t
*error_sector
)
608 mddev_t
*mddev
= q
->queuedata
;
609 conf_t
*conf
= mddev_to_conf(mddev
);
613 for (i
=0; i
<mddev
->raid_disks
&& ret
== 0; i
++) {
614 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
615 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
616 struct block_device
*bdev
= rdev
->bdev
;
617 request_queue_t
*r_queue
= bdev_get_queue(bdev
);
619 if (!r_queue
->issue_flush_fn
)
622 atomic_inc(&rdev
->nr_pending
);
624 ret
= r_queue
->issue_flush_fn(r_queue
, bdev
->bd_disk
,
626 rdev_dec_pending(rdev
, mddev
);
636 * Sometimes we need to suspend IO while we do something else,
637 * either some resync/recovery, or reconfigure the array.
638 * To do this we raise a 'barrier'.
639 * The 'barrier' is a counter that can be raised multiple times
640 * to count how many activities are happening which preclude
642 * We can only raise the barrier if there is no pending IO.
643 * i.e. if nr_pending == 0.
644 * We choose only to raise the barrier if no-one is waiting for the
645 * barrier to go down. This means that as soon as an IO request
646 * is ready, no other operations which require a barrier will start
647 * until the IO request has had a chance.
649 * So: regular IO calls 'wait_barrier'. When that returns there
650 * is no backgroup IO happening, It must arrange to call
651 * allow_barrier when it has finished its IO.
652 * backgroup IO calls must call raise_barrier. Once that returns
653 * there is no normal IO happeing. It must arrange to call
654 * lower_barrier when the particular background IO completes.
656 #define RESYNC_DEPTH 32
658 static void raise_barrier(conf_t
*conf
, int force
)
660 BUG_ON(force
&& !conf
->barrier
);
661 spin_lock_irq(&conf
->resync_lock
);
663 /* Wait until no block IO is waiting (unless 'force') */
664 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
666 raid10_unplug(conf
->mddev
->queue
));
668 /* block any new IO from starting */
671 /* No wait for all pending IO to complete */
672 wait_event_lock_irq(conf
->wait_barrier
,
673 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
675 raid10_unplug(conf
->mddev
->queue
));
677 spin_unlock_irq(&conf
->resync_lock
);
680 static void lower_barrier(conf_t
*conf
)
683 spin_lock_irqsave(&conf
->resync_lock
, flags
);
685 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
686 wake_up(&conf
->wait_barrier
);
689 static void wait_barrier(conf_t
*conf
)
691 spin_lock_irq(&conf
->resync_lock
);
694 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
696 raid10_unplug(conf
->mddev
->queue
));
700 spin_unlock_irq(&conf
->resync_lock
);
703 static void allow_barrier(conf_t
*conf
)
706 spin_lock_irqsave(&conf
->resync_lock
, flags
);
708 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
709 wake_up(&conf
->wait_barrier
);
712 static void freeze_array(conf_t
*conf
)
714 /* stop syncio and normal IO and wait for everything to
716 * We increment barrier and nr_waiting, and then
717 * wait until barrier+nr_pending match nr_queued+2
719 spin_lock_irq(&conf
->resync_lock
);
722 wait_event_lock_irq(conf
->wait_barrier
,
723 conf
->barrier
+conf
->nr_pending
== conf
->nr_queued
+2,
725 raid10_unplug(conf
->mddev
->queue
));
726 spin_unlock_irq(&conf
->resync_lock
);
729 static void unfreeze_array(conf_t
*conf
)
731 /* reverse the effect of the freeze */
732 spin_lock_irq(&conf
->resync_lock
);
735 wake_up(&conf
->wait_barrier
);
736 spin_unlock_irq(&conf
->resync_lock
);
739 static int make_request(request_queue_t
*q
, struct bio
* bio
)
741 mddev_t
*mddev
= q
->queuedata
;
742 conf_t
*conf
= mddev_to_conf(mddev
);
743 mirror_info_t
*mirror
;
745 struct bio
*read_bio
;
747 int chunk_sects
= conf
->chunk_mask
+ 1;
748 const int rw
= bio_data_dir(bio
);
752 if (unlikely(bio_barrier(bio
))) {
753 bio_endio(bio
, bio
->bi_size
, -EOPNOTSUPP
);
757 /* If this request crosses a chunk boundary, we need to
758 * split it. This will only happen for 1 PAGE (or less) requests.
760 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
762 conf
->near_copies
< conf
->raid_disks
)) {
764 /* Sanity check -- queue functions should prevent this happening */
765 if (bio
->bi_vcnt
!= 1 ||
768 /* This is a one page bio that upper layers
769 * refuse to split for us, so we need to split it.
771 bp
= bio_split(bio
, bio_split_pool
,
772 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
773 if (make_request(q
, &bp
->bio1
))
774 generic_make_request(&bp
->bio1
);
775 if (make_request(q
, &bp
->bio2
))
776 generic_make_request(&bp
->bio2
);
778 bio_pair_release(bp
);
781 printk("raid10_make_request bug: can't convert block across chunks"
782 " or bigger than %dk %llu %d\n", chunk_sects
/2,
783 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
785 bio_io_error(bio
, bio
->bi_size
);
789 md_write_start(mddev
, bio
);
792 * Register the new request and wait if the reconstruction
793 * thread has put up a bar for new requests.
794 * Continue immediately if no resync is active currently.
798 disk_stat_inc(mddev
->gendisk
, ios
[rw
]);
799 disk_stat_add(mddev
->gendisk
, sectors
[rw
], bio_sectors(bio
));
801 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
803 r10_bio
->master_bio
= bio
;
804 r10_bio
->sectors
= bio
->bi_size
>> 9;
806 r10_bio
->mddev
= mddev
;
807 r10_bio
->sector
= bio
->bi_sector
;
812 * read balancing logic:
814 int disk
= read_balance(conf
, r10_bio
);
815 int slot
= r10_bio
->read_slot
;
817 raid_end_bio_io(r10_bio
);
820 mirror
= conf
->mirrors
+ disk
;
822 read_bio
= bio_clone(bio
, GFP_NOIO
);
824 r10_bio
->devs
[slot
].bio
= read_bio
;
826 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
827 mirror
->rdev
->data_offset
;
828 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
829 read_bio
->bi_end_io
= raid10_end_read_request
;
830 read_bio
->bi_rw
= READ
;
831 read_bio
->bi_private
= r10_bio
;
833 generic_make_request(read_bio
);
840 /* first select target devices under spinlock and
841 * inc refcount on their rdev. Record them by setting
844 raid10_find_phys(conf
, r10_bio
);
846 for (i
= 0; i
< conf
->copies
; i
++) {
847 int d
= r10_bio
->devs
[i
].devnum
;
848 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
850 !test_bit(Faulty
, &rdev
->flags
)) {
851 atomic_inc(&rdev
->nr_pending
);
852 r10_bio
->devs
[i
].bio
= bio
;
854 r10_bio
->devs
[i
].bio
= NULL
;
855 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
860 atomic_set(&r10_bio
->remaining
, 0);
863 for (i
= 0; i
< conf
->copies
; i
++) {
865 int d
= r10_bio
->devs
[i
].devnum
;
866 if (!r10_bio
->devs
[i
].bio
)
869 mbio
= bio_clone(bio
, GFP_NOIO
);
870 r10_bio
->devs
[i
].bio
= mbio
;
872 mbio
->bi_sector
= r10_bio
->devs
[i
].addr
+
873 conf
->mirrors
[d
].rdev
->data_offset
;
874 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
875 mbio
->bi_end_io
= raid10_end_write_request
;
877 mbio
->bi_private
= r10_bio
;
879 atomic_inc(&r10_bio
->remaining
);
880 bio_list_add(&bl
, mbio
);
883 bitmap_startwrite(mddev
->bitmap
, bio
->bi_sector
, r10_bio
->sectors
, 0);
884 spin_lock_irqsave(&conf
->device_lock
, flags
);
885 bio_list_merge(&conf
->pending_bio_list
, &bl
);
886 blk_plug_device(mddev
->queue
);
887 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
892 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
894 conf_t
*conf
= mddev_to_conf(mddev
);
897 if (conf
->near_copies
< conf
->raid_disks
)
898 seq_printf(seq
, " %dK chunks", mddev
->chunk_size
/1024);
899 if (conf
->near_copies
> 1)
900 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
901 if (conf
->far_copies
> 1)
902 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
904 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
905 conf
->working_disks
);
906 for (i
= 0; i
< conf
->raid_disks
; i
++)
907 seq_printf(seq
, "%s",
908 conf
->mirrors
[i
].rdev
&&
909 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
910 seq_printf(seq
, "]");
913 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
915 char b
[BDEVNAME_SIZE
];
916 conf_t
*conf
= mddev_to_conf(mddev
);
919 * If it is not operational, then we have already marked it as dead
920 * else if it is the last working disks, ignore the error, let the
921 * next level up know.
922 * else mark the drive as failed
924 if (test_bit(In_sync
, &rdev
->flags
)
925 && conf
->working_disks
== 1)
927 * Don't fail the drive, just return an IO error.
928 * The test should really be more sophisticated than
929 * "working_disks == 1", but it isn't critical, and
930 * can wait until we do more sophisticated "is the drive
931 * really dead" tests...
934 if (test_bit(In_sync
, &rdev
->flags
)) {
936 conf
->working_disks
--;
938 * if recovery is running, make sure it aborts.
940 set_bit(MD_RECOVERY_ERR
, &mddev
->recovery
);
942 clear_bit(In_sync
, &rdev
->flags
);
943 set_bit(Faulty
, &rdev
->flags
);
945 printk(KERN_ALERT
"raid10: Disk failure on %s, disabling device. \n"
946 " Operation continuing on %d devices\n",
947 bdevname(rdev
->bdev
,b
), conf
->working_disks
);
950 static void print_conf(conf_t
*conf
)
955 printk("RAID10 conf printout:\n");
960 printk(" --- wd:%d rd:%d\n", conf
->working_disks
,
963 for (i
= 0; i
< conf
->raid_disks
; i
++) {
964 char b
[BDEVNAME_SIZE
];
965 tmp
= conf
->mirrors
+ i
;
967 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
968 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
969 !test_bit(Faulty
, &tmp
->rdev
->flags
),
970 bdevname(tmp
->rdev
->bdev
,b
));
974 static void close_sync(conf_t
*conf
)
979 mempool_destroy(conf
->r10buf_pool
);
980 conf
->r10buf_pool
= NULL
;
983 /* check if there are enough drives for
984 * every block to appear on atleast one
986 static int enough(conf_t
*conf
)
991 int n
= conf
->copies
;
994 if (conf
->mirrors
[first
].rdev
)
996 first
= (first
+1) % conf
->raid_disks
;
1000 } while (first
!= 0);
1004 static int raid10_spare_active(mddev_t
*mddev
)
1007 conf_t
*conf
= mddev
->private;
1011 * Find all non-in_sync disks within the RAID10 configuration
1012 * and mark them in_sync
1014 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1015 tmp
= conf
->mirrors
+ i
;
1017 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1018 && !test_bit(In_sync
, &tmp
->rdev
->flags
)) {
1019 conf
->working_disks
++;
1021 set_bit(In_sync
, &tmp
->rdev
->flags
);
1030 static int raid10_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1032 conf_t
*conf
= mddev
->private;
1037 if (mddev
->recovery_cp
< MaxSector
)
1038 /* only hot-add to in-sync arrays, as recovery is
1039 * very different from resync
1045 if (rdev
->saved_raid_disk
>= 0 &&
1046 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1047 mirror
= rdev
->saved_raid_disk
;
1050 for ( ; mirror
< mddev
->raid_disks
; mirror
++)
1051 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1053 blk_queue_stack_limits(mddev
->queue
,
1054 rdev
->bdev
->bd_disk
->queue
);
1055 /* as we don't honour merge_bvec_fn, we must never risk
1056 * violating it, so limit ->max_sector to one PAGE, as
1057 * a one page request is never in violation.
1059 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
1060 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
1061 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
1063 p
->head_position
= 0;
1064 rdev
->raid_disk
= mirror
;
1066 if (rdev
->saved_raid_disk
!= mirror
)
1068 rcu_assign_pointer(p
->rdev
, rdev
);
1076 static int raid10_remove_disk(mddev_t
*mddev
, int number
)
1078 conf_t
*conf
= mddev
->private;
1081 mirror_info_t
*p
= conf
->mirrors
+ number
;
1086 if (test_bit(In_sync
, &rdev
->flags
) ||
1087 atomic_read(&rdev
->nr_pending
)) {
1093 if (atomic_read(&rdev
->nr_pending
)) {
1094 /* lost the race, try later */
1106 static int end_sync_read(struct bio
*bio
, unsigned int bytes_done
, int error
)
1108 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1109 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
1115 for (i
=0; i
<conf
->copies
; i
++)
1116 if (r10_bio
->devs
[i
].bio
== bio
)
1118 if (i
== conf
->copies
)
1120 update_head_pos(i
, r10_bio
);
1121 d
= r10_bio
->devs
[i
].devnum
;
1123 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1124 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1126 atomic_add(r10_bio
->sectors
,
1127 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1128 if (!test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
1129 md_error(r10_bio
->mddev
,
1130 conf
->mirrors
[d
].rdev
);
1133 /* for reconstruct, we always reschedule after a read.
1134 * for resync, only after all reads
1136 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1137 atomic_dec_and_test(&r10_bio
->remaining
)) {
1138 /* we have read all the blocks,
1139 * do the comparison in process context in raid10d
1141 reschedule_retry(r10_bio
);
1143 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1147 static int end_sync_write(struct bio
*bio
, unsigned int bytes_done
, int error
)
1149 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1150 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1151 mddev_t
*mddev
= r10_bio
->mddev
;
1152 conf_t
*conf
= mddev_to_conf(mddev
);
1158 for (i
= 0; i
< conf
->copies
; i
++)
1159 if (r10_bio
->devs
[i
].bio
== bio
)
1161 d
= r10_bio
->devs
[i
].devnum
;
1164 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1165 update_head_pos(i
, r10_bio
);
1167 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1168 if (r10_bio
->master_bio
== NULL
) {
1169 /* the primary of several recovery bios */
1170 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1174 r10bio_t
*r10_bio2
= (r10bio_t
*)r10_bio
->master_bio
;
1179 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1184 * Note: sync and recover and handled very differently for raid10
1185 * This code is for resync.
1186 * For resync, we read through virtual addresses and read all blocks.
1187 * If there is any error, we schedule a write. The lowest numbered
1188 * drive is authoritative.
1189 * However requests come for physical address, so we need to map.
1190 * For every physical address there are raid_disks/copies virtual addresses,
1191 * which is always are least one, but is not necessarly an integer.
1192 * This means that a physical address can span multiple chunks, so we may
1193 * have to submit multiple io requests for a single sync request.
1196 * We check if all blocks are in-sync and only write to blocks that
1199 static void sync_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1201 conf_t
*conf
= mddev_to_conf(mddev
);
1203 struct bio
*tbio
, *fbio
;
1205 atomic_set(&r10_bio
->remaining
, 1);
1207 /* find the first device with a block */
1208 for (i
=0; i
<conf
->copies
; i
++)
1209 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1212 if (i
== conf
->copies
)
1216 fbio
= r10_bio
->devs
[i
].bio
;
1218 /* now find blocks with errors */
1219 for (i
=0 ; i
< conf
->copies
; i
++) {
1221 int vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1223 tbio
= r10_bio
->devs
[i
].bio
;
1225 if (tbio
->bi_end_io
!= end_sync_read
)
1229 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1230 /* We know that the bi_io_vec layout is the same for
1231 * both 'first' and 'i', so we just compare them.
1232 * All vec entries are PAGE_SIZE;
1234 for (j
= 0; j
< vcnt
; j
++)
1235 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1236 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1241 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1243 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1244 /* Don't fix anything. */
1246 /* Ok, we need to write this bio
1247 * First we need to fixup bv_offset, bv_len and
1248 * bi_vecs, as the read request might have corrupted these
1250 tbio
->bi_vcnt
= vcnt
;
1251 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1253 tbio
->bi_phys_segments
= 0;
1254 tbio
->bi_hw_segments
= 0;
1255 tbio
->bi_hw_front_size
= 0;
1256 tbio
->bi_hw_back_size
= 0;
1257 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1258 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1259 tbio
->bi_next
= NULL
;
1260 tbio
->bi_rw
= WRITE
;
1261 tbio
->bi_private
= r10_bio
;
1262 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1264 for (j
=0; j
< vcnt
; j
++) {
1265 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1266 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1268 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1269 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1272 tbio
->bi_end_io
= end_sync_write
;
1274 d
= r10_bio
->devs
[i
].devnum
;
1275 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1276 atomic_inc(&r10_bio
->remaining
);
1277 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1279 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1280 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1281 generic_make_request(tbio
);
1285 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1286 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1292 * Now for the recovery code.
1293 * Recovery happens across physical sectors.
1294 * We recover all non-is_sync drives by finding the virtual address of
1295 * each, and then choose a working drive that also has that virt address.
1296 * There is a separate r10_bio for each non-in_sync drive.
1297 * Only the first two slots are in use. The first for reading,
1298 * The second for writing.
1302 static void recovery_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1304 conf_t
*conf
= mddev_to_conf(mddev
);
1306 struct bio
*bio
, *wbio
;
1309 /* move the pages across to the second bio
1310 * and submit the write request
1312 bio
= r10_bio
->devs
[0].bio
;
1313 wbio
= r10_bio
->devs
[1].bio
;
1314 for (i
=0; i
< wbio
->bi_vcnt
; i
++) {
1315 struct page
*p
= bio
->bi_io_vec
[i
].bv_page
;
1316 bio
->bi_io_vec
[i
].bv_page
= wbio
->bi_io_vec
[i
].bv_page
;
1317 wbio
->bi_io_vec
[i
].bv_page
= p
;
1319 d
= r10_bio
->devs
[1].devnum
;
1321 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1322 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1323 if (test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
1324 generic_make_request(wbio
);
1326 bio_endio(wbio
, wbio
->bi_size
, -EIO
);
1331 * This is a kernel thread which:
1333 * 1. Retries failed read operations on working mirrors.
1334 * 2. Updates the raid superblock when problems encounter.
1335 * 3. Performs writes following reads for array syncronising.
1338 static void raid10d(mddev_t
*mddev
)
1342 unsigned long flags
;
1343 conf_t
*conf
= mddev_to_conf(mddev
);
1344 struct list_head
*head
= &conf
->retry_list
;
1348 md_check_recovery(mddev
);
1351 char b
[BDEVNAME_SIZE
];
1352 spin_lock_irqsave(&conf
->device_lock
, flags
);
1354 if (conf
->pending_bio_list
.head
) {
1355 bio
= bio_list_get(&conf
->pending_bio_list
);
1356 blk_remove_plug(mddev
->queue
);
1357 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1358 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1359 if (bitmap_unplug(mddev
->bitmap
) != 0)
1360 printk("%s: bitmap file write failed!\n", mdname(mddev
));
1362 while (bio
) { /* submit pending writes */
1363 struct bio
*next
= bio
->bi_next
;
1364 bio
->bi_next
= NULL
;
1365 generic_make_request(bio
);
1373 if (list_empty(head
))
1375 r10_bio
= list_entry(head
->prev
, r10bio_t
, retry_list
);
1376 list_del(head
->prev
);
1378 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1380 mddev
= r10_bio
->mddev
;
1381 conf
= mddev_to_conf(mddev
);
1382 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
)) {
1383 sync_request_write(mddev
, r10_bio
);
1385 } else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
1386 recovery_request_write(mddev
, r10_bio
);
1390 /* we got a read error. Maybe the drive is bad. Maybe just
1391 * the block and we can fix it.
1392 * We freeze all other IO, and try reading the block from
1393 * other devices. When we find one, we re-write
1394 * and check it that fixes the read error.
1395 * This is all done synchronously while the array is
1398 int sect
= 0; /* Offset from r10_bio->sector */
1399 int sectors
= r10_bio
->sectors
;
1401 if (mddev
->ro
== 0) while(sectors
) {
1403 int sl
= r10_bio
->read_slot
;
1406 if (s
> (PAGE_SIZE
>>9))
1410 int d
= r10_bio
->devs
[sl
].devnum
;
1411 rdev
= conf
->mirrors
[d
].rdev
;
1413 test_bit(In_sync
, &rdev
->flags
) &&
1414 sync_page_io(rdev
->bdev
,
1415 r10_bio
->devs
[sl
].addr
+
1416 sect
+ rdev
->data_offset
,
1418 conf
->tmppage
, READ
))
1422 if (sl
== conf
->copies
)
1425 } while (!success
&& sl
!= r10_bio
->read_slot
);
1429 /* write it back and re-read */
1430 while (sl
!= r10_bio
->read_slot
) {
1435 d
= r10_bio
->devs
[sl
].devnum
;
1436 rdev
= conf
->mirrors
[d
].rdev
;
1437 atomic_add(s
, &rdev
->corrected_errors
);
1439 test_bit(In_sync
, &rdev
->flags
)) {
1440 if (sync_page_io(rdev
->bdev
,
1441 r10_bio
->devs
[sl
].addr
+
1442 sect
+ rdev
->data_offset
,
1443 s
<<9, conf
->tmppage
, WRITE
) == 0)
1444 /* Well, this device is dead */
1445 md_error(mddev
, rdev
);
1449 while (sl
!= r10_bio
->read_slot
) {
1454 d
= r10_bio
->devs
[sl
].devnum
;
1455 rdev
= conf
->mirrors
[d
].rdev
;
1457 test_bit(In_sync
, &rdev
->flags
)) {
1458 if (sync_page_io(rdev
->bdev
,
1459 r10_bio
->devs
[sl
].addr
+
1460 sect
+ rdev
->data_offset
,
1461 s
<<9, conf
->tmppage
, READ
) == 0)
1462 /* Well, this device is dead */
1463 md_error(mddev
, rdev
);
1467 /* Cannot read from anywhere -- bye bye array */
1468 md_error(mddev
, conf
->mirrors
[r10_bio
->devs
[r10_bio
->read_slot
].devnum
].rdev
);
1475 unfreeze_array(conf
);
1477 bio
= r10_bio
->devs
[r10_bio
->read_slot
].bio
;
1478 r10_bio
->devs
[r10_bio
->read_slot
].bio
=
1479 mddev
->ro
? IO_BLOCKED
: NULL
;
1481 mirror
= read_balance(conf
, r10_bio
);
1483 printk(KERN_ALERT
"raid10: %s: unrecoverable I/O"
1484 " read error for block %llu\n",
1485 bdevname(bio
->bi_bdev
,b
),
1486 (unsigned long long)r10_bio
->sector
);
1487 raid_end_bio_io(r10_bio
);
1489 rdev
= conf
->mirrors
[mirror
].rdev
;
1490 if (printk_ratelimit())
1491 printk(KERN_ERR
"raid10: %s: redirecting sector %llu to"
1492 " another mirror\n",
1493 bdevname(rdev
->bdev
,b
),
1494 (unsigned long long)r10_bio
->sector
);
1495 bio
= bio_clone(r10_bio
->master_bio
, GFP_NOIO
);
1496 r10_bio
->devs
[r10_bio
->read_slot
].bio
= bio
;
1497 bio
->bi_sector
= r10_bio
->devs
[r10_bio
->read_slot
].addr
1498 + rdev
->data_offset
;
1499 bio
->bi_bdev
= rdev
->bdev
;
1501 bio
->bi_private
= r10_bio
;
1502 bio
->bi_end_io
= raid10_end_read_request
;
1504 generic_make_request(bio
);
1508 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1510 unplug_slaves(mddev
);
1514 static int init_resync(conf_t
*conf
)
1518 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1519 if (conf
->r10buf_pool
)
1521 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
1522 if (!conf
->r10buf_pool
)
1524 conf
->next_resync
= 0;
1529 * perform a "sync" on one "block"
1531 * We need to make sure that no normal I/O request - particularly write
1532 * requests - conflict with active sync requests.
1534 * This is achieved by tracking pending requests and a 'barrier' concept
1535 * that can be installed to exclude normal IO requests.
1537 * Resync and recovery are handled very differently.
1538 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1540 * For resync, we iterate over virtual addresses, read all copies,
1541 * and update if there are differences. If only one copy is live,
1543 * For recovery, we iterate over physical addresses, read a good
1544 * value for each non-in_sync drive, and over-write.
1546 * So, for recovery we may have several outstanding complex requests for a
1547 * given address, one for each out-of-sync device. We model this by allocating
1548 * a number of r10_bio structures, one for each out-of-sync device.
1549 * As we setup these structures, we collect all bio's together into a list
1550 * which we then process collectively to add pages, and then process again
1551 * to pass to generic_make_request.
1553 * The r10_bio structures are linked using a borrowed master_bio pointer.
1554 * This link is counted in ->remaining. When the r10_bio that points to NULL
1555 * has its remaining count decremented to 0, the whole complex operation
1560 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
1562 conf_t
*conf
= mddev_to_conf(mddev
);
1564 struct bio
*biolist
= NULL
, *bio
;
1565 sector_t max_sector
, nr_sectors
;
1571 sector_t sectors_skipped
= 0;
1572 int chunks_skipped
= 0;
1574 if (!conf
->r10buf_pool
)
1575 if (init_resync(conf
))
1579 max_sector
= mddev
->size
<< 1;
1580 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1581 max_sector
= mddev
->resync_max_sectors
;
1582 if (sector_nr
>= max_sector
) {
1583 /* If we aborted, we need to abort the
1584 * sync on the 'current' bitmap chucks (there can
1585 * be several when recovering multiple devices).
1586 * as we may have started syncing it but not finished.
1587 * We can find the current address in
1588 * mddev->curr_resync, but for recovery,
1589 * we need to convert that to several
1590 * virtual addresses.
1592 if (mddev
->curr_resync
< max_sector
) { /* aborted */
1593 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1594 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
1596 else for (i
=0; i
<conf
->raid_disks
; i
++) {
1598 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
1599 bitmap_end_sync(mddev
->bitmap
, sect
,
1602 } else /* completed sync */
1605 bitmap_close_sync(mddev
->bitmap
);
1608 return sectors_skipped
;
1610 if (chunks_skipped
>= conf
->raid_disks
) {
1611 /* if there has been nothing to do on any drive,
1612 * then there is nothing to do at all..
1615 return (max_sector
- sector_nr
) + sectors_skipped
;
1618 /* make sure whole request will fit in a chunk - if chunks
1621 if (conf
->near_copies
< conf
->raid_disks
&&
1622 max_sector
> (sector_nr
| conf
->chunk_mask
))
1623 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
1625 * If there is non-resync activity waiting for us then
1626 * put in a delay to throttle resync.
1628 if (!go_faster
&& conf
->nr_waiting
)
1629 msleep_interruptible(1000);
1631 /* Again, very different code for resync and recovery.
1632 * Both must result in an r10bio with a list of bios that
1633 * have bi_end_io, bi_sector, bi_bdev set,
1634 * and bi_private set to the r10bio.
1635 * For recovery, we may actually create several r10bios
1636 * with 2 bios in each, that correspond to the bios in the main one.
1637 * In this case, the subordinate r10bios link back through a
1638 * borrowed master_bio pointer, and the counter in the master
1639 * includes a ref from each subordinate.
1641 /* First, we decide what to do and set ->bi_end_io
1642 * To end_sync_read if we want to read, and
1643 * end_sync_write if we will want to write.
1646 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
1647 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
1648 /* recovery... the complicated one */
1652 for (i
=0 ; i
<conf
->raid_disks
; i
++)
1653 if (conf
->mirrors
[i
].rdev
&&
1654 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
)) {
1655 int still_degraded
= 0;
1656 /* want to reconstruct this device */
1657 r10bio_t
*rb2
= r10_bio
;
1658 sector_t sect
= raid10_find_virt(conf
, sector_nr
, i
);
1660 /* Unless we are doing a full sync, we only need
1661 * to recover the block if it is set in the bitmap
1663 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1665 if (sync_blocks
< max_sync
)
1666 max_sync
= sync_blocks
;
1669 /* yep, skip the sync_blocks here, but don't assume
1670 * that there will never be anything to do here
1672 chunks_skipped
= -1;
1676 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1677 raise_barrier(conf
, rb2
!= NULL
);
1678 atomic_set(&r10_bio
->remaining
, 0);
1680 r10_bio
->master_bio
= (struct bio
*)rb2
;
1682 atomic_inc(&rb2
->remaining
);
1683 r10_bio
->mddev
= mddev
;
1684 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
1685 r10_bio
->sector
= sect
;
1687 raid10_find_phys(conf
, r10_bio
);
1688 /* Need to check if this section will still be
1691 for (j
=0; j
<conf
->copies
;j
++) {
1692 int d
= r10_bio
->devs
[j
].devnum
;
1693 if (conf
->mirrors
[d
].rdev
== NULL
||
1694 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
)) {
1699 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1700 &sync_blocks
, still_degraded
);
1702 for (j
=0; j
<conf
->copies
;j
++) {
1703 int d
= r10_bio
->devs
[j
].devnum
;
1704 if (conf
->mirrors
[d
].rdev
&&
1705 test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
)) {
1706 /* This is where we read from */
1707 bio
= r10_bio
->devs
[0].bio
;
1708 bio
->bi_next
= biolist
;
1710 bio
->bi_private
= r10_bio
;
1711 bio
->bi_end_io
= end_sync_read
;
1713 bio
->bi_sector
= r10_bio
->devs
[j
].addr
+
1714 conf
->mirrors
[d
].rdev
->data_offset
;
1715 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1716 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1717 atomic_inc(&r10_bio
->remaining
);
1718 /* and we write to 'i' */
1720 for (k
=0; k
<conf
->copies
; k
++)
1721 if (r10_bio
->devs
[k
].devnum
== i
)
1723 bio
= r10_bio
->devs
[1].bio
;
1724 bio
->bi_next
= biolist
;
1726 bio
->bi_private
= r10_bio
;
1727 bio
->bi_end_io
= end_sync_write
;
1729 bio
->bi_sector
= r10_bio
->devs
[k
].addr
+
1730 conf
->mirrors
[i
].rdev
->data_offset
;
1731 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1733 r10_bio
->devs
[0].devnum
= d
;
1734 r10_bio
->devs
[1].devnum
= i
;
1739 if (j
== conf
->copies
) {
1740 /* Cannot recover, so abort the recovery */
1743 if (!test_and_set_bit(MD_RECOVERY_ERR
, &mddev
->recovery
))
1744 printk(KERN_INFO
"raid10: %s: insufficient working devices for recovery.\n",
1749 if (biolist
== NULL
) {
1751 r10bio_t
*rb2
= r10_bio
;
1752 r10_bio
= (r10bio_t
*) rb2
->master_bio
;
1753 rb2
->master_bio
= NULL
;
1759 /* resync. Schedule a read for every block at this virt offset */
1762 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
1763 &sync_blocks
, mddev
->degraded
) &&
1764 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1765 /* We can skip this block */
1767 return sync_blocks
+ sectors_skipped
;
1769 if (sync_blocks
< max_sync
)
1770 max_sync
= sync_blocks
;
1771 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1773 r10_bio
->mddev
= mddev
;
1774 atomic_set(&r10_bio
->remaining
, 0);
1775 raise_barrier(conf
, 0);
1776 conf
->next_resync
= sector_nr
;
1778 r10_bio
->master_bio
= NULL
;
1779 r10_bio
->sector
= sector_nr
;
1780 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
1781 raid10_find_phys(conf
, r10_bio
);
1782 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
1784 for (i
=0; i
<conf
->copies
; i
++) {
1785 int d
= r10_bio
->devs
[i
].devnum
;
1786 bio
= r10_bio
->devs
[i
].bio
;
1787 bio
->bi_end_io
= NULL
;
1788 if (conf
->mirrors
[d
].rdev
== NULL
||
1789 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
1791 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1792 atomic_inc(&r10_bio
->remaining
);
1793 bio
->bi_next
= biolist
;
1795 bio
->bi_private
= r10_bio
;
1796 bio
->bi_end_io
= end_sync_read
;
1798 bio
->bi_sector
= r10_bio
->devs
[i
].addr
+
1799 conf
->mirrors
[d
].rdev
->data_offset
;
1800 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1805 for (i
=0; i
<conf
->copies
; i
++) {
1806 int d
= r10_bio
->devs
[i
].devnum
;
1807 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
1808 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1816 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1818 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1820 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
1823 bio
->bi_phys_segments
= 0;
1824 bio
->bi_hw_segments
= 0;
1829 if (sector_nr
+ max_sync
< max_sector
)
1830 max_sector
= sector_nr
+ max_sync
;
1833 int len
= PAGE_SIZE
;
1835 if (sector_nr
+ (len
>>9) > max_sector
)
1836 len
= (max_sector
- sector_nr
) << 9;
1839 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1840 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
1841 if (bio_add_page(bio
, page
, len
, 0) == 0) {
1844 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
1845 for (bio2
= biolist
; bio2
&& bio2
!= bio
; bio2
= bio2
->bi_next
) {
1846 /* remove last page from this bio */
1848 bio2
->bi_size
-= len
;
1849 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
1855 nr_sectors
+= len
>>9;
1856 sector_nr
+= len
>>9;
1857 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
1859 r10_bio
->sectors
= nr_sectors
;
1863 biolist
= biolist
->bi_next
;
1865 bio
->bi_next
= NULL
;
1866 r10_bio
= bio
->bi_private
;
1867 r10_bio
->sectors
= nr_sectors
;
1869 if (bio
->bi_end_io
== end_sync_read
) {
1870 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1871 generic_make_request(bio
);
1875 if (sectors_skipped
)
1876 /* pretend they weren't skipped, it makes
1877 * no important difference in this case
1879 md_done_sync(mddev
, sectors_skipped
, 1);
1881 return sectors_skipped
+ nr_sectors
;
1883 /* There is nowhere to write, so all non-sync
1884 * drives must be failed, so try the next chunk...
1887 sector_t sec
= max_sector
- sector_nr
;
1888 sectors_skipped
+= sec
;
1890 sector_nr
= max_sector
;
1895 static int run(mddev_t
*mddev
)
1899 mirror_info_t
*disk
;
1901 struct list_head
*tmp
;
1903 sector_t stride
, size
;
1905 if (mddev
->chunk_size
== 0) {
1906 printk(KERN_ERR
"md/raid10: non-zero chunk size required.\n");
1910 nc
= mddev
->layout
& 255;
1911 fc
= (mddev
->layout
>> 8) & 255;
1912 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
1913 (mddev
->layout
>> 16)) {
1914 printk(KERN_ERR
"raid10: %s: unsupported raid10 layout: 0x%8x\n",
1915 mdname(mddev
), mddev
->layout
);
1919 * copy the already verified devices into our private RAID10
1920 * bookkeeping area. [whatever we allocate in run(),
1921 * should be freed in stop()]
1923 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
1924 mddev
->private = conf
;
1926 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
1930 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
1932 if (!conf
->mirrors
) {
1933 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
1938 conf
->tmppage
= alloc_page(GFP_KERNEL
);
1942 conf
->near_copies
= nc
;
1943 conf
->far_copies
= fc
;
1944 conf
->copies
= nc
*fc
;
1945 conf
->chunk_mask
= (sector_t
)(mddev
->chunk_size
>>9)-1;
1946 conf
->chunk_shift
= ffz(~mddev
->chunk_size
) - 9;
1947 stride
= mddev
->size
>> (conf
->chunk_shift
-1);
1948 sector_div(stride
, fc
);
1949 conf
->stride
= stride
<< conf
->chunk_shift
;
1951 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
1952 r10bio_pool_free
, conf
);
1953 if (!conf
->r10bio_pool
) {
1954 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
1959 ITERATE_RDEV(mddev
, rdev
, tmp
) {
1960 disk_idx
= rdev
->raid_disk
;
1961 if (disk_idx
>= mddev
->raid_disks
1964 disk
= conf
->mirrors
+ disk_idx
;
1968 blk_queue_stack_limits(mddev
->queue
,
1969 rdev
->bdev
->bd_disk
->queue
);
1970 /* as we don't honour merge_bvec_fn, we must never risk
1971 * violating it, so limit ->max_sector to one PAGE, as
1972 * a one page request is never in violation.
1974 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
1975 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
1976 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
1978 disk
->head_position
= 0;
1979 if (!test_bit(Faulty
, &rdev
->flags
) && test_bit(In_sync
, &rdev
->flags
))
1980 conf
->working_disks
++;
1982 conf
->raid_disks
= mddev
->raid_disks
;
1983 conf
->mddev
= mddev
;
1984 spin_lock_init(&conf
->device_lock
);
1985 INIT_LIST_HEAD(&conf
->retry_list
);
1987 spin_lock_init(&conf
->resync_lock
);
1988 init_waitqueue_head(&conf
->wait_barrier
);
1990 /* need to check that every block has at least one working mirror */
1991 if (!enough(conf
)) {
1992 printk(KERN_ERR
"raid10: not enough operational mirrors for %s\n",
1997 mddev
->degraded
= 0;
1998 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2000 disk
= conf
->mirrors
+ i
;
2003 disk
->head_position
= 0;
2009 mddev
->thread
= md_register_thread(raid10d
, mddev
, "%s_raid10");
2010 if (!mddev
->thread
) {
2012 "raid10: couldn't allocate thread for %s\n",
2018 "raid10: raid set %s active with %d out of %d devices\n",
2019 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2022 * Ok, everything is just fine now
2024 size
= conf
->stride
* conf
->raid_disks
;
2025 sector_div(size
, conf
->near_copies
);
2026 mddev
->array_size
= size
/2;
2027 mddev
->resync_max_sectors
= size
;
2029 mddev
->queue
->unplug_fn
= raid10_unplug
;
2030 mddev
->queue
->issue_flush_fn
= raid10_issue_flush
;
2032 /* Calculate max read-ahead size.
2033 * We need to readahead at least twice a whole stripe....
2037 int stripe
= conf
->raid_disks
* mddev
->chunk_size
/ PAGE_SIZE
;
2038 stripe
/= conf
->near_copies
;
2039 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
2040 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
2043 if (conf
->near_copies
< mddev
->raid_disks
)
2044 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
2048 if (conf
->r10bio_pool
)
2049 mempool_destroy(conf
->r10bio_pool
);
2050 safe_put_page(conf
->tmppage
);
2051 kfree(conf
->mirrors
);
2053 mddev
->private = NULL
;
2058 static int stop(mddev_t
*mddev
)
2060 conf_t
*conf
= mddev_to_conf(mddev
);
2062 md_unregister_thread(mddev
->thread
);
2063 mddev
->thread
= NULL
;
2064 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
2065 if (conf
->r10bio_pool
)
2066 mempool_destroy(conf
->r10bio_pool
);
2067 kfree(conf
->mirrors
);
2069 mddev
->private = NULL
;
2073 static void raid10_quiesce(mddev_t
*mddev
, int state
)
2075 conf_t
*conf
= mddev_to_conf(mddev
);
2079 raise_barrier(conf
, 0);
2082 lower_barrier(conf
);
2085 if (mddev
->thread
) {
2087 mddev
->thread
->timeout
= mddev
->bitmap
->daemon_sleep
* HZ
;
2089 mddev
->thread
->timeout
= MAX_SCHEDULE_TIMEOUT
;
2090 md_wakeup_thread(mddev
->thread
);
2094 static struct mdk_personality raid10_personality
=
2098 .owner
= THIS_MODULE
,
2099 .make_request
= make_request
,
2103 .error_handler
= error
,
2104 .hot_add_disk
= raid10_add_disk
,
2105 .hot_remove_disk
= raid10_remove_disk
,
2106 .spare_active
= raid10_spare_active
,
2107 .sync_request
= sync_request
,
2108 .quiesce
= raid10_quiesce
,
2111 static int __init
raid_init(void)
2113 return register_md_personality(&raid10_personality
);
2116 static void raid_exit(void)
2118 unregister_md_personality(&raid10_personality
);
2121 module_init(raid_init
);
2122 module_exit(raid_exit
);
2123 MODULE_LICENSE("GPL");
2124 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2125 MODULE_ALIAS("md-raid10");
2126 MODULE_ALIAS("md-level-10");