2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/seq_file.h>
25 #include <linux/ratelimit.h>
32 * RAID10 provides a combination of RAID0 and RAID1 functionality.
33 * The layout of data is defined by
36 * near_copies (stored in low byte of layout)
37 * far_copies (stored in second byte of layout)
38 * far_offset (stored in bit 16 of layout )
40 * The data to be stored is divided into chunks using chunksize.
41 * Each device is divided into far_copies sections.
42 * In each section, chunks are laid out in a style similar to raid0, but
43 * near_copies copies of each chunk is stored (each on a different drive).
44 * The starting device for each section is offset near_copies from the starting
45 * device of the previous section.
46 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
48 * near_copies and far_copies must be at least one, and their product is at most
51 * If far_offset is true, then the far_copies are handled a bit differently.
52 * The copies are still in different stripes, but instead of be very far apart
53 * on disk, there are adjacent stripes.
57 * Number of guaranteed r10bios in case of extreme VM load:
59 #define NR_RAID10_BIOS 256
61 static void allow_barrier(conf_t
*conf
);
62 static void lower_barrier(conf_t
*conf
);
64 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
67 int size
= offsetof(struct r10bio_s
, devs
[conf
->copies
]);
69 /* allocate a r10bio with room for raid_disks entries in the bios array */
70 return kzalloc(size
, gfp_flags
);
73 static void r10bio_pool_free(void *r10_bio
, void *data
)
78 /* Maximum size of each resync request */
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
81 /* amount of memory to reserve for resync requests */
82 #define RESYNC_WINDOW (1024*1024)
83 /* maximum number of concurrent requests, memory permitting */
84 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
87 * When performing a resync, we need to read and compare, so
88 * we need as many pages are there are copies.
89 * When performing a recovery, we need 2 bios, one for read,
90 * one for write (we recover only one drive per r10buf)
93 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
102 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
106 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
107 nalloc
= conf
->copies
; /* resync */
109 nalloc
= 2; /* recovery */
114 for (j
= nalloc
; j
-- ; ) {
115 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
118 r10_bio
->devs
[j
].bio
= bio
;
121 * Allocate RESYNC_PAGES data pages and attach them
124 for (j
= 0 ; j
< nalloc
; j
++) {
125 bio
= r10_bio
->devs
[j
].bio
;
126 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
127 if (j
== 1 && !test_bit(MD_RECOVERY_SYNC
,
128 &conf
->mddev
->recovery
)) {
129 /* we can share bv_page's during recovery */
130 struct bio
*rbio
= r10_bio
->devs
[0].bio
;
131 page
= rbio
->bi_io_vec
[i
].bv_page
;
134 page
= alloc_page(gfp_flags
);
138 bio
->bi_io_vec
[i
].bv_page
= page
;
146 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
148 for (i
= 0; i
< RESYNC_PAGES
; i
++)
149 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
152 while ( ++j
< nalloc
)
153 bio_put(r10_bio
->devs
[j
].bio
);
154 r10bio_pool_free(r10_bio
, conf
);
158 static void r10buf_pool_free(void *__r10_bio
, void *data
)
162 r10bio_t
*r10bio
= __r10_bio
;
165 for (j
=0; j
< conf
->copies
; j
++) {
166 struct bio
*bio
= r10bio
->devs
[j
].bio
;
168 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
169 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
170 bio
->bi_io_vec
[i
].bv_page
= NULL
;
175 r10bio_pool_free(r10bio
, conf
);
178 static void put_all_bios(conf_t
*conf
, r10bio_t
*r10_bio
)
182 for (i
= 0; i
< conf
->copies
; i
++) {
183 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
184 if (*bio
&& *bio
!= IO_BLOCKED
)
190 static void free_r10bio(r10bio_t
*r10_bio
)
192 conf_t
*conf
= r10_bio
->mddev
->private;
195 * Wake up any possible resync thread that waits for the device
200 put_all_bios(conf
, r10_bio
);
201 mempool_free(r10_bio
, conf
->r10bio_pool
);
204 static void put_buf(r10bio_t
*r10_bio
)
206 conf_t
*conf
= r10_bio
->mddev
->private;
208 mempool_free(r10_bio
, conf
->r10buf_pool
);
213 static void reschedule_retry(r10bio_t
*r10_bio
)
216 mddev_t
*mddev
= r10_bio
->mddev
;
217 conf_t
*conf
= mddev
->private;
219 spin_lock_irqsave(&conf
->device_lock
, flags
);
220 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
222 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
224 /* wake up frozen array... */
225 wake_up(&conf
->wait_barrier
);
227 md_wakeup_thread(mddev
->thread
);
231 * raid_end_bio_io() is called when we have finished servicing a mirrored
232 * operation and are ready to return a success/failure code to the buffer
235 static void raid_end_bio_io(r10bio_t
*r10_bio
)
237 struct bio
*bio
= r10_bio
->master_bio
;
240 test_bit(R10BIO_Uptodate
, &r10_bio
->state
) ? 0 : -EIO
);
241 free_r10bio(r10_bio
);
245 * Update disk head position estimator based on IRQ completion info.
247 static inline void update_head_pos(int slot
, r10bio_t
*r10_bio
)
249 conf_t
*conf
= r10_bio
->mddev
->private;
251 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
252 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
256 * Find the disk number which triggered given bio
258 static int find_bio_disk(conf_t
*conf
, r10bio_t
*r10_bio
, struct bio
*bio
)
262 for (slot
= 0; slot
< conf
->copies
; slot
++)
263 if (r10_bio
->devs
[slot
].bio
== bio
)
266 BUG_ON(slot
== conf
->copies
);
267 update_head_pos(slot
, r10_bio
);
269 return r10_bio
->devs
[slot
].devnum
;
272 static void raid10_end_read_request(struct bio
*bio
, int error
)
274 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
275 r10bio_t
*r10_bio
= bio
->bi_private
;
277 conf_t
*conf
= r10_bio
->mddev
->private;
280 slot
= r10_bio
->read_slot
;
281 dev
= r10_bio
->devs
[slot
].devnum
;
283 * this branch is our 'one mirror IO has finished' event handler:
285 update_head_pos(slot
, r10_bio
);
289 * Set R10BIO_Uptodate in our master bio, so that
290 * we will return a good error code to the higher
291 * levels even if IO on some other mirrored buffer fails.
293 * The 'master' represents the composite IO operation to
294 * user-side. So if something waits for IO, then it will
295 * wait for the 'master' bio.
297 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
298 raid_end_bio_io(r10_bio
);
299 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
302 * oops, read error - keep the refcount on the rdev
304 char b
[BDEVNAME_SIZE
];
305 printk_ratelimited(KERN_ERR
306 "md/raid10:%s: %s: rescheduling sector %llu\n",
308 bdevname(conf
->mirrors
[dev
].rdev
->bdev
, b
),
309 (unsigned long long)r10_bio
->sector
);
310 reschedule_retry(r10_bio
);
314 static void raid10_end_write_request(struct bio
*bio
, int error
)
316 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
317 r10bio_t
*r10_bio
= bio
->bi_private
;
319 conf_t
*conf
= r10_bio
->mddev
->private;
321 dev
= find_bio_disk(conf
, r10_bio
, bio
);
324 * this branch is our 'one mirror IO has finished' event handler:
327 md_error(r10_bio
->mddev
, conf
->mirrors
[dev
].rdev
);
328 /* an I/O failed, we can't clear the bitmap */
329 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
332 * Set R10BIO_Uptodate in our master bio, so that
333 * we will return a good error code for to the higher
334 * levels even if IO on some other mirrored buffer fails.
336 * The 'master' represents the composite IO operation to
337 * user-side. So if something waits for IO, then it will
338 * wait for the 'master' bio.
340 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
344 * Let's see if all mirrored write operations have finished
347 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
348 /* clear the bitmap if all writes complete successfully */
349 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
351 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
353 md_write_end(r10_bio
->mddev
);
354 raid_end_bio_io(r10_bio
);
357 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
362 * RAID10 layout manager
363 * As well as the chunksize and raid_disks count, there are two
364 * parameters: near_copies and far_copies.
365 * near_copies * far_copies must be <= raid_disks.
366 * Normally one of these will be 1.
367 * If both are 1, we get raid0.
368 * If near_copies == raid_disks, we get raid1.
370 * Chunks are laid out in raid0 style with near_copies copies of the
371 * first chunk, followed by near_copies copies of the next chunk and
373 * If far_copies > 1, then after 1/far_copies of the array has been assigned
374 * as described above, we start again with a device offset of near_copies.
375 * So we effectively have another copy of the whole array further down all
376 * the drives, but with blocks on different drives.
377 * With this layout, and block is never stored twice on the one device.
379 * raid10_find_phys finds the sector offset of a given virtual sector
380 * on each device that it is on.
382 * raid10_find_virt does the reverse mapping, from a device and a
383 * sector offset to a virtual address
386 static void raid10_find_phys(conf_t
*conf
, r10bio_t
*r10bio
)
396 /* now calculate first sector/dev */
397 chunk
= r10bio
->sector
>> conf
->chunk_shift
;
398 sector
= r10bio
->sector
& conf
->chunk_mask
;
400 chunk
*= conf
->near_copies
;
402 dev
= sector_div(stripe
, conf
->raid_disks
);
403 if (conf
->far_offset
)
404 stripe
*= conf
->far_copies
;
406 sector
+= stripe
<< conf
->chunk_shift
;
408 /* and calculate all the others */
409 for (n
=0; n
< conf
->near_copies
; n
++) {
412 r10bio
->devs
[slot
].addr
= sector
;
413 r10bio
->devs
[slot
].devnum
= d
;
416 for (f
= 1; f
< conf
->far_copies
; f
++) {
417 d
+= conf
->near_copies
;
418 if (d
>= conf
->raid_disks
)
419 d
-= conf
->raid_disks
;
421 r10bio
->devs
[slot
].devnum
= d
;
422 r10bio
->devs
[slot
].addr
= s
;
426 if (dev
>= conf
->raid_disks
) {
428 sector
+= (conf
->chunk_mask
+ 1);
431 BUG_ON(slot
!= conf
->copies
);
434 static sector_t
raid10_find_virt(conf_t
*conf
, sector_t sector
, int dev
)
436 sector_t offset
, chunk
, vchunk
;
438 offset
= sector
& conf
->chunk_mask
;
439 if (conf
->far_offset
) {
441 chunk
= sector
>> conf
->chunk_shift
;
442 fc
= sector_div(chunk
, conf
->far_copies
);
443 dev
-= fc
* conf
->near_copies
;
445 dev
+= conf
->raid_disks
;
447 while (sector
>= conf
->stride
) {
448 sector
-= conf
->stride
;
449 if (dev
< conf
->near_copies
)
450 dev
+= conf
->raid_disks
- conf
->near_copies
;
452 dev
-= conf
->near_copies
;
454 chunk
= sector
>> conf
->chunk_shift
;
456 vchunk
= chunk
* conf
->raid_disks
+ dev
;
457 sector_div(vchunk
, conf
->near_copies
);
458 return (vchunk
<< conf
->chunk_shift
) + offset
;
462 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
464 * @bvm: properties of new bio
465 * @biovec: the request that could be merged to it.
467 * Return amount of bytes we can accept at this offset
468 * If near_copies == raid_disk, there are no striping issues,
469 * but in that case, the function isn't called at all.
471 static int raid10_mergeable_bvec(struct request_queue
*q
,
472 struct bvec_merge_data
*bvm
,
473 struct bio_vec
*biovec
)
475 mddev_t
*mddev
= q
->queuedata
;
476 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
478 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
479 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
481 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
482 if (max
< 0) max
= 0; /* bio_add cannot handle a negative return */
483 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
484 return biovec
->bv_len
;
490 * This routine returns the disk from which the requested read should
491 * be done. There is a per-array 'next expected sequential IO' sector
492 * number - if this matches on the next IO then we use the last disk.
493 * There is also a per-disk 'last know head position' sector that is
494 * maintained from IRQ contexts, both the normal and the resync IO
495 * completion handlers update this position correctly. If there is no
496 * perfect sequential match then we pick the disk whose head is closest.
498 * If there are 2 mirrors in the same 2 devices, performance degrades
499 * because position is mirror, not device based.
501 * The rdev for the device selected will have nr_pending incremented.
505 * FIXME: possibly should rethink readbalancing and do it differently
506 * depending on near_copies / far_copies geometry.
508 static int read_balance(conf_t
*conf
, r10bio_t
*r10_bio
)
510 const sector_t this_sector
= r10_bio
->sector
;
512 const int sectors
= r10_bio
->sectors
;
513 sector_t new_distance
, best_dist
;
518 raid10_find_phys(conf
, r10_bio
);
522 best_dist
= MaxSector
;
525 * Check if we can balance. We can balance on the whole
526 * device if no resync is going on (recovery is ok), or below
527 * the resync window. We take the first readable disk when
528 * above the resync window.
530 if (conf
->mddev
->recovery_cp
< MaxSector
531 && (this_sector
+ sectors
>= conf
->next_resync
))
534 for (slot
= 0; slot
< conf
->copies
; slot
++) {
535 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
537 disk
= r10_bio
->devs
[slot
].devnum
;
538 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
541 if (!test_bit(In_sync
, &rdev
->flags
))
547 /* This optimisation is debatable, and completely destroys
548 * sequential read speed for 'far copies' arrays. So only
549 * keep it for 'near' arrays, and review those later.
551 if (conf
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
554 /* for far > 1 always use the lowest address */
555 if (conf
->far_copies
> 1)
556 new_distance
= r10_bio
->devs
[slot
].addr
;
558 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
559 conf
->mirrors
[disk
].head_position
);
560 if (new_distance
< best_dist
) {
561 best_dist
= new_distance
;
565 if (slot
== conf
->copies
)
569 disk
= r10_bio
->devs
[slot
].devnum
;
570 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
573 atomic_inc(&rdev
->nr_pending
);
574 if (test_bit(Faulty
, &rdev
->flags
)) {
575 /* Cannot risk returning a device that failed
576 * before we inc'ed nr_pending
578 rdev_dec_pending(rdev
, conf
->mddev
);
581 r10_bio
->read_slot
= slot
;
589 static int raid10_congested(void *data
, int bits
)
591 mddev_t
*mddev
= data
;
592 conf_t
*conf
= mddev
->private;
595 if (mddev_congested(mddev
, bits
))
598 for (i
= 0; i
< conf
->raid_disks
&& ret
== 0; i
++) {
599 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
600 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
601 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
603 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
610 static void flush_pending_writes(conf_t
*conf
)
612 /* Any writes that have been queued but are awaiting
613 * bitmap updates get flushed here.
615 spin_lock_irq(&conf
->device_lock
);
617 if (conf
->pending_bio_list
.head
) {
619 bio
= bio_list_get(&conf
->pending_bio_list
);
620 spin_unlock_irq(&conf
->device_lock
);
621 /* flush any pending bitmap writes to disk
622 * before proceeding w/ I/O */
623 bitmap_unplug(conf
->mddev
->bitmap
);
625 while (bio
) { /* submit pending writes */
626 struct bio
*next
= bio
->bi_next
;
628 generic_make_request(bio
);
632 spin_unlock_irq(&conf
->device_lock
);
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.
657 static void raise_barrier(conf_t
*conf
, int force
)
659 BUG_ON(force
&& !conf
->barrier
);
660 spin_lock_irq(&conf
->resync_lock
);
662 /* Wait until no block IO is waiting (unless 'force') */
663 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
664 conf
->resync_lock
, );
666 /* block any new IO from starting */
669 /* Now wait for all pending IO to complete */
670 wait_event_lock_irq(conf
->wait_barrier
,
671 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
672 conf
->resync_lock
, );
674 spin_unlock_irq(&conf
->resync_lock
);
677 static void lower_barrier(conf_t
*conf
)
680 spin_lock_irqsave(&conf
->resync_lock
, flags
);
682 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
683 wake_up(&conf
->wait_barrier
);
686 static void wait_barrier(conf_t
*conf
)
688 spin_lock_irq(&conf
->resync_lock
);
691 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
697 spin_unlock_irq(&conf
->resync_lock
);
700 static void allow_barrier(conf_t
*conf
)
703 spin_lock_irqsave(&conf
->resync_lock
, flags
);
705 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
706 wake_up(&conf
->wait_barrier
);
709 static void freeze_array(conf_t
*conf
)
711 /* stop syncio and normal IO and wait for everything to
713 * We increment barrier and nr_waiting, and then
714 * wait until nr_pending match nr_queued+1
715 * This is called in the context of one normal IO request
716 * that has failed. Thus any sync request that might be pending
717 * will be blocked by nr_pending, and we need to wait for
718 * pending IO requests to complete or be queued for re-try.
719 * Thus the number queued (nr_queued) plus this request (1)
720 * must match the number of pending IOs (nr_pending) before
723 spin_lock_irq(&conf
->resync_lock
);
726 wait_event_lock_irq(conf
->wait_barrier
,
727 conf
->nr_pending
== conf
->nr_queued
+1,
729 flush_pending_writes(conf
));
731 spin_unlock_irq(&conf
->resync_lock
);
734 static void unfreeze_array(conf_t
*conf
)
736 /* reverse the effect of the freeze */
737 spin_lock_irq(&conf
->resync_lock
);
740 wake_up(&conf
->wait_barrier
);
741 spin_unlock_irq(&conf
->resync_lock
);
744 static int make_request(mddev_t
*mddev
, struct bio
* bio
)
746 conf_t
*conf
= mddev
->private;
747 mirror_info_t
*mirror
;
749 struct bio
*read_bio
;
751 int chunk_sects
= conf
->chunk_mask
+ 1;
752 const int rw
= bio_data_dir(bio
);
753 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
754 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
756 mdk_rdev_t
*blocked_rdev
;
759 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
760 md_flush_request(mddev
, bio
);
764 /* If this request crosses a chunk boundary, we need to
765 * split it. This will only happen for 1 PAGE (or less) requests.
767 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
769 conf
->near_copies
< conf
->raid_disks
)) {
771 /* Sanity check -- queue functions should prevent this happening */
772 if (bio
->bi_vcnt
!= 1 ||
775 /* This is a one page bio that upper layers
776 * refuse to split for us, so we need to split it.
779 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
781 /* Each of these 'make_request' calls will call 'wait_barrier'.
782 * If the first succeeds but the second blocks due to the resync
783 * thread raising the barrier, we will deadlock because the
784 * IO to the underlying device will be queued in generic_make_request
785 * and will never complete, so will never reduce nr_pending.
786 * So increment nr_waiting here so no new raise_barriers will
787 * succeed, and so the second wait_barrier cannot block.
789 spin_lock_irq(&conf
->resync_lock
);
791 spin_unlock_irq(&conf
->resync_lock
);
793 if (make_request(mddev
, &bp
->bio1
))
794 generic_make_request(&bp
->bio1
);
795 if (make_request(mddev
, &bp
->bio2
))
796 generic_make_request(&bp
->bio2
);
798 spin_lock_irq(&conf
->resync_lock
);
800 wake_up(&conf
->wait_barrier
);
801 spin_unlock_irq(&conf
->resync_lock
);
803 bio_pair_release(bp
);
806 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
807 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
808 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
814 md_write_start(mddev
, bio
);
817 * Register the new request and wait if the reconstruction
818 * thread has put up a bar for new requests.
819 * Continue immediately if no resync is active currently.
823 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
825 r10_bio
->master_bio
= bio
;
826 r10_bio
->sectors
= bio
->bi_size
>> 9;
828 r10_bio
->mddev
= mddev
;
829 r10_bio
->sector
= bio
->bi_sector
;
834 * read balancing logic:
836 int disk
= read_balance(conf
, r10_bio
);
837 int slot
= r10_bio
->read_slot
;
839 raid_end_bio_io(r10_bio
);
842 mirror
= conf
->mirrors
+ disk
;
844 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
846 r10_bio
->devs
[slot
].bio
= read_bio
;
848 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
849 mirror
->rdev
->data_offset
;
850 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
851 read_bio
->bi_end_io
= raid10_end_read_request
;
852 read_bio
->bi_rw
= READ
| do_sync
;
853 read_bio
->bi_private
= r10_bio
;
855 generic_make_request(read_bio
);
862 /* first select target devices under rcu_lock and
863 * inc refcount on their rdev. Record them by setting
866 plugged
= mddev_check_plugged(mddev
);
868 raid10_find_phys(conf
, r10_bio
);
872 for (i
= 0; i
< conf
->copies
; i
++) {
873 int d
= r10_bio
->devs
[i
].devnum
;
874 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
875 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
876 atomic_inc(&rdev
->nr_pending
);
880 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
881 atomic_inc(&rdev
->nr_pending
);
882 r10_bio
->devs
[i
].bio
= bio
;
884 r10_bio
->devs
[i
].bio
= NULL
;
885 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
890 if (unlikely(blocked_rdev
)) {
891 /* Have to wait for this device to get unblocked, then retry */
895 for (j
= 0; j
< i
; j
++)
896 if (r10_bio
->devs
[j
].bio
) {
897 d
= r10_bio
->devs
[j
].devnum
;
898 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
901 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
906 atomic_set(&r10_bio
->remaining
, 1);
907 bitmap_startwrite(mddev
->bitmap
, bio
->bi_sector
, r10_bio
->sectors
, 0);
909 for (i
= 0; i
< conf
->copies
; i
++) {
911 int d
= r10_bio
->devs
[i
].devnum
;
912 if (!r10_bio
->devs
[i
].bio
)
915 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
916 r10_bio
->devs
[i
].bio
= mbio
;
918 mbio
->bi_sector
= r10_bio
->devs
[i
].addr
+
919 conf
->mirrors
[d
].rdev
->data_offset
;
920 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
921 mbio
->bi_end_io
= raid10_end_write_request
;
922 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
923 mbio
->bi_private
= r10_bio
;
925 atomic_inc(&r10_bio
->remaining
);
926 spin_lock_irqsave(&conf
->device_lock
, flags
);
927 bio_list_add(&conf
->pending_bio_list
, mbio
);
928 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
931 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
932 /* This matches the end of raid10_end_write_request() */
933 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
935 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
938 raid_end_bio_io(r10_bio
);
941 /* In case raid10d snuck in to freeze_array */
942 wake_up(&conf
->wait_barrier
);
944 if (do_sync
|| !mddev
->bitmap
|| !plugged
)
945 md_wakeup_thread(mddev
->thread
);
949 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
951 conf_t
*conf
= mddev
->private;
954 if (conf
->near_copies
< conf
->raid_disks
)
955 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
956 if (conf
->near_copies
> 1)
957 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
958 if (conf
->far_copies
> 1) {
959 if (conf
->far_offset
)
960 seq_printf(seq
, " %d offset-copies", conf
->far_copies
);
962 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
964 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
965 conf
->raid_disks
- mddev
->degraded
);
966 for (i
= 0; i
< conf
->raid_disks
; i
++)
967 seq_printf(seq
, "%s",
968 conf
->mirrors
[i
].rdev
&&
969 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
970 seq_printf(seq
, "]");
973 /* check if there are enough drives for
974 * every block to appear on atleast one.
975 * Don't consider the device numbered 'ignore'
976 * as we might be about to remove it.
978 static int enough(conf_t
*conf
, int ignore
)
983 int n
= conf
->copies
;
986 if (conf
->mirrors
[first
].rdev
&&
989 first
= (first
+1) % conf
->raid_disks
;
993 } while (first
!= 0);
997 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
999 char b
[BDEVNAME_SIZE
];
1000 conf_t
*conf
= mddev
->private;
1003 * If it is not operational, then we have already marked it as dead
1004 * else if it is the last working disks, ignore the error, let the
1005 * next level up know.
1006 * else mark the drive as failed
1008 if (test_bit(In_sync
, &rdev
->flags
)
1009 && !enough(conf
, rdev
->raid_disk
))
1011 * Don't fail the drive, just return an IO error.
1014 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1015 unsigned long flags
;
1016 spin_lock_irqsave(&conf
->device_lock
, flags
);
1018 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1020 * if recovery is running, make sure it aborts.
1022 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1024 set_bit(Blocked
, &rdev
->flags
);
1025 set_bit(Faulty
, &rdev
->flags
);
1026 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1028 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1029 "md/raid10:%s: Operation continuing on %d devices.\n",
1030 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1031 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1034 static void print_conf(conf_t
*conf
)
1039 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1041 printk(KERN_DEBUG
"(!conf)\n");
1044 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1047 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1048 char b
[BDEVNAME_SIZE
];
1049 tmp
= conf
->mirrors
+ i
;
1051 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1052 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1053 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1054 bdevname(tmp
->rdev
->bdev
,b
));
1058 static void close_sync(conf_t
*conf
)
1061 allow_barrier(conf
);
1063 mempool_destroy(conf
->r10buf_pool
);
1064 conf
->r10buf_pool
= NULL
;
1067 static int raid10_spare_active(mddev_t
*mddev
)
1070 conf_t
*conf
= mddev
->private;
1073 unsigned long flags
;
1076 * Find all non-in_sync disks within the RAID10 configuration
1077 * and mark them in_sync
1079 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1080 tmp
= conf
->mirrors
+ i
;
1082 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1083 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1085 sysfs_notify_dirent(tmp
->rdev
->sysfs_state
);
1088 spin_lock_irqsave(&conf
->device_lock
, flags
);
1089 mddev
->degraded
-= count
;
1090 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1097 static int raid10_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1099 conf_t
*conf
= mddev
->private;
1103 int last
= conf
->raid_disks
- 1;
1105 if (rdev
->badblocks
.count
)
1108 if (mddev
->recovery_cp
< MaxSector
)
1109 /* only hot-add to in-sync arrays, as recovery is
1110 * very different from resync
1113 if (!enough(conf
, -1))
1116 if (rdev
->raid_disk
>= 0)
1117 first
= last
= rdev
->raid_disk
;
1119 if (rdev
->saved_raid_disk
>= first
&&
1120 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1121 mirror
= rdev
->saved_raid_disk
;
1124 for ( ; mirror
<= last
; mirror
++) {
1125 mirror_info_t
*p
= &conf
->mirrors
[mirror
];
1126 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1131 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1132 rdev
->data_offset
<< 9);
1133 /* as we don't honour merge_bvec_fn, we must
1134 * never risk violating it, so limit
1135 * ->max_segments to one lying with a single
1136 * page, as a one page request is never in
1139 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1140 blk_queue_max_segments(mddev
->queue
, 1);
1141 blk_queue_segment_boundary(mddev
->queue
,
1142 PAGE_CACHE_SIZE
- 1);
1145 p
->head_position
= 0;
1146 rdev
->raid_disk
= mirror
;
1148 if (rdev
->saved_raid_disk
!= mirror
)
1150 rcu_assign_pointer(p
->rdev
, rdev
);
1154 md_integrity_add_rdev(rdev
, mddev
);
1159 static int raid10_remove_disk(mddev_t
*mddev
, int number
)
1161 conf_t
*conf
= mddev
->private;
1164 mirror_info_t
*p
= conf
->mirrors
+ number
;
1169 if (test_bit(In_sync
, &rdev
->flags
) ||
1170 atomic_read(&rdev
->nr_pending
)) {
1174 /* Only remove faulty devices in recovery
1177 if (!test_bit(Faulty
, &rdev
->flags
) &&
1178 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1185 if (atomic_read(&rdev
->nr_pending
)) {
1186 /* lost the race, try later */
1191 err
= md_integrity_register(mddev
);
1200 static void end_sync_read(struct bio
*bio
, int error
)
1202 r10bio_t
*r10_bio
= bio
->bi_private
;
1203 conf_t
*conf
= r10_bio
->mddev
->private;
1206 d
= find_bio_disk(conf
, r10_bio
, bio
);
1208 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1209 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1211 atomic_add(r10_bio
->sectors
,
1212 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1213 if (!test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
1214 md_error(r10_bio
->mddev
,
1215 conf
->mirrors
[d
].rdev
);
1218 /* for reconstruct, we always reschedule after a read.
1219 * for resync, only after all reads
1221 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1222 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1223 atomic_dec_and_test(&r10_bio
->remaining
)) {
1224 /* we have read all the blocks,
1225 * do the comparison in process context in raid10d
1227 reschedule_retry(r10_bio
);
1231 static void end_sync_write(struct bio
*bio
, int error
)
1233 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1234 r10bio_t
*r10_bio
= bio
->bi_private
;
1235 mddev_t
*mddev
= r10_bio
->mddev
;
1236 conf_t
*conf
= mddev
->private;
1239 d
= find_bio_disk(conf
, r10_bio
, bio
);
1242 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1244 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1245 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1246 if (r10_bio
->master_bio
== NULL
) {
1247 /* the primary of several recovery bios */
1248 sector_t s
= r10_bio
->sectors
;
1250 md_done_sync(mddev
, s
, 1);
1253 r10bio_t
*r10_bio2
= (r10bio_t
*)r10_bio
->master_bio
;
1261 * Note: sync and recover and handled very differently for raid10
1262 * This code is for resync.
1263 * For resync, we read through virtual addresses and read all blocks.
1264 * If there is any error, we schedule a write. The lowest numbered
1265 * drive is authoritative.
1266 * However requests come for physical address, so we need to map.
1267 * For every physical address there are raid_disks/copies virtual addresses,
1268 * which is always are least one, but is not necessarly an integer.
1269 * This means that a physical address can span multiple chunks, so we may
1270 * have to submit multiple io requests for a single sync request.
1273 * We check if all blocks are in-sync and only write to blocks that
1276 static void sync_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1278 conf_t
*conf
= mddev
->private;
1280 struct bio
*tbio
, *fbio
;
1282 atomic_set(&r10_bio
->remaining
, 1);
1284 /* find the first device with a block */
1285 for (i
=0; i
<conf
->copies
; i
++)
1286 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1289 if (i
== conf
->copies
)
1293 fbio
= r10_bio
->devs
[i
].bio
;
1295 /* now find blocks with errors */
1296 for (i
=0 ; i
< conf
->copies
; i
++) {
1298 int vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1300 tbio
= r10_bio
->devs
[i
].bio
;
1302 if (tbio
->bi_end_io
!= end_sync_read
)
1306 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1307 /* We know that the bi_io_vec layout is the same for
1308 * both 'first' and 'i', so we just compare them.
1309 * All vec entries are PAGE_SIZE;
1311 for (j
= 0; j
< vcnt
; j
++)
1312 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1313 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1318 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1320 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1321 /* Don't fix anything. */
1323 /* Ok, we need to write this bio
1324 * First we need to fixup bv_offset, bv_len and
1325 * bi_vecs, as the read request might have corrupted these
1327 tbio
->bi_vcnt
= vcnt
;
1328 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1330 tbio
->bi_phys_segments
= 0;
1331 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1332 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1333 tbio
->bi_next
= NULL
;
1334 tbio
->bi_rw
= WRITE
;
1335 tbio
->bi_private
= r10_bio
;
1336 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1338 for (j
=0; j
< vcnt
; j
++) {
1339 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1340 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1342 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1343 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1346 tbio
->bi_end_io
= end_sync_write
;
1348 d
= r10_bio
->devs
[i
].devnum
;
1349 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1350 atomic_inc(&r10_bio
->remaining
);
1351 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1353 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1354 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1355 generic_make_request(tbio
);
1359 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1360 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1366 * Now for the recovery code.
1367 * Recovery happens across physical sectors.
1368 * We recover all non-is_sync drives by finding the virtual address of
1369 * each, and then choose a working drive that also has that virt address.
1370 * There is a separate r10_bio for each non-in_sync drive.
1371 * Only the first two slots are in use. The first for reading,
1372 * The second for writing.
1376 static void recovery_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1378 conf_t
*conf
= mddev
->private;
1383 * share the pages with the first bio
1384 * and submit the write request
1386 wbio
= r10_bio
->devs
[1].bio
;
1387 d
= r10_bio
->devs
[1].devnum
;
1389 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1390 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1391 if (test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
1392 generic_make_request(wbio
);
1395 "md/raid10:%s: recovery aborted due to read error\n",
1397 conf
->mirrors
[d
].recovery_disabled
= mddev
->recovery_disabled
;
1398 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1405 * Used by fix_read_error() to decay the per rdev read_errors.
1406 * We halve the read error count for every hour that has elapsed
1407 * since the last recorded read error.
1410 static void check_decay_read_errors(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1412 struct timespec cur_time_mon
;
1413 unsigned long hours_since_last
;
1414 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
1416 ktime_get_ts(&cur_time_mon
);
1418 if (rdev
->last_read_error
.tv_sec
== 0 &&
1419 rdev
->last_read_error
.tv_nsec
== 0) {
1420 /* first time we've seen a read error */
1421 rdev
->last_read_error
= cur_time_mon
;
1425 hours_since_last
= (cur_time_mon
.tv_sec
-
1426 rdev
->last_read_error
.tv_sec
) / 3600;
1428 rdev
->last_read_error
= cur_time_mon
;
1431 * if hours_since_last is > the number of bits in read_errors
1432 * just set read errors to 0. We do this to avoid
1433 * overflowing the shift of read_errors by hours_since_last.
1435 if (hours_since_last
>= 8 * sizeof(read_errors
))
1436 atomic_set(&rdev
->read_errors
, 0);
1438 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
1442 * This is a kernel thread which:
1444 * 1. Retries failed read operations on working mirrors.
1445 * 2. Updates the raid superblock when problems encounter.
1446 * 3. Performs writes following reads for array synchronising.
1449 static void fix_read_error(conf_t
*conf
, mddev_t
*mddev
, r10bio_t
*r10_bio
)
1451 int sect
= 0; /* Offset from r10_bio->sector */
1452 int sectors
= r10_bio
->sectors
;
1454 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
1455 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1457 /* still own a reference to this rdev, so it cannot
1458 * have been cleared recently.
1460 rdev
= conf
->mirrors
[d
].rdev
;
1462 if (test_bit(Faulty
, &rdev
->flags
))
1463 /* drive has already been failed, just ignore any
1464 more fix_read_error() attempts */
1467 check_decay_read_errors(mddev
, rdev
);
1468 atomic_inc(&rdev
->read_errors
);
1469 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
1470 char b
[BDEVNAME_SIZE
];
1471 bdevname(rdev
->bdev
, b
);
1474 "md/raid10:%s: %s: Raid device exceeded "
1475 "read_error threshold [cur %d:max %d]\n",
1477 atomic_read(&rdev
->read_errors
), max_read_errors
);
1479 "md/raid10:%s: %s: Failing raid device\n",
1481 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1487 int sl
= r10_bio
->read_slot
;
1491 if (s
> (PAGE_SIZE
>>9))
1496 d
= r10_bio
->devs
[sl
].devnum
;
1497 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1499 test_bit(In_sync
, &rdev
->flags
)) {
1500 atomic_inc(&rdev
->nr_pending
);
1502 success
= sync_page_io(rdev
,
1503 r10_bio
->devs
[sl
].addr
+
1506 conf
->tmppage
, READ
, false);
1507 rdev_dec_pending(rdev
, mddev
);
1513 if (sl
== conf
->copies
)
1515 } while (!success
&& sl
!= r10_bio
->read_slot
);
1519 /* Cannot read from anywhere -- bye bye array */
1520 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1521 md_error(mddev
, conf
->mirrors
[dn
].rdev
);
1526 /* write it back and re-read */
1528 while (sl
!= r10_bio
->read_slot
) {
1529 char b
[BDEVNAME_SIZE
];
1534 d
= r10_bio
->devs
[sl
].devnum
;
1535 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1537 !test_bit(In_sync
, &rdev
->flags
))
1540 atomic_inc(&rdev
->nr_pending
);
1542 if (sync_page_io(rdev
,
1543 r10_bio
->devs
[sl
].addr
+
1545 s
<<9, conf
->tmppage
, WRITE
, false)
1547 /* Well, this device is dead */
1549 "md/raid10:%s: read correction "
1551 " (%d sectors at %llu on %s)\n",
1553 (unsigned long long)(
1554 sect
+ rdev
->data_offset
),
1555 bdevname(rdev
->bdev
, b
));
1556 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
1559 bdevname(rdev
->bdev
, b
));
1560 md_error(mddev
, rdev
);
1562 rdev_dec_pending(rdev
, mddev
);
1566 while (sl
!= r10_bio
->read_slot
) {
1567 char b
[BDEVNAME_SIZE
];
1572 d
= r10_bio
->devs
[sl
].devnum
;
1573 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1575 !test_bit(In_sync
, &rdev
->flags
))
1578 atomic_inc(&rdev
->nr_pending
);
1580 if (sync_page_io(rdev
,
1581 r10_bio
->devs
[sl
].addr
+
1583 s
<<9, conf
->tmppage
,
1584 READ
, false) == 0) {
1585 /* Well, this device is dead */
1587 "md/raid10:%s: unable to read back "
1589 " (%d sectors at %llu on %s)\n",
1591 (unsigned long long)(
1592 sect
+ rdev
->data_offset
),
1593 bdevname(rdev
->bdev
, b
));
1594 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
1597 bdevname(rdev
->bdev
, b
));
1599 md_error(mddev
, rdev
);
1602 "md/raid10:%s: read error corrected"
1603 " (%d sectors at %llu on %s)\n",
1605 (unsigned long long)(
1606 sect
+ rdev
->data_offset
),
1607 bdevname(rdev
->bdev
, b
));
1608 atomic_add(s
, &rdev
->corrected_errors
);
1611 rdev_dec_pending(rdev
, mddev
);
1621 static void handle_read_error(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1623 int slot
= r10_bio
->read_slot
;
1624 int mirror
= r10_bio
->devs
[slot
].devnum
;
1626 conf_t
*conf
= mddev
->private;
1628 char b
[BDEVNAME_SIZE
];
1629 unsigned long do_sync
;
1631 /* we got a read error. Maybe the drive is bad. Maybe just
1632 * the block and we can fix it.
1633 * We freeze all other IO, and try reading the block from
1634 * other devices. When we find one, we re-write
1635 * and check it that fixes the read error.
1636 * This is all done synchronously while the array is
1639 if (mddev
->ro
== 0) {
1641 fix_read_error(conf
, mddev
, r10_bio
);
1642 unfreeze_array(conf
);
1644 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, mddev
);
1646 bio
= r10_bio
->devs
[slot
].bio
;
1647 r10_bio
->devs
[slot
].bio
=
1648 mddev
->ro
? IO_BLOCKED
: NULL
;
1649 mirror
= read_balance(conf
, r10_bio
);
1651 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
1652 " read error for block %llu\n",
1654 bdevname(bio
->bi_bdev
, b
),
1655 (unsigned long long)r10_bio
->sector
);
1656 raid_end_bio_io(r10_bio
);
1661 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
1663 slot
= r10_bio
->read_slot
;
1664 rdev
= conf
->mirrors
[mirror
].rdev
;
1667 "md/raid10:%s: %s: redirecting"
1668 "sector %llu to another mirror\n",
1670 bdevname(rdev
->bdev
, b
),
1671 (unsigned long long)r10_bio
->sector
);
1672 bio
= bio_clone_mddev(r10_bio
->master_bio
,
1674 r10_bio
->devs
[slot
].bio
= bio
;
1675 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
1676 + rdev
->data_offset
;
1677 bio
->bi_bdev
= rdev
->bdev
;
1678 bio
->bi_rw
= READ
| do_sync
;
1679 bio
->bi_private
= r10_bio
;
1680 bio
->bi_end_io
= raid10_end_read_request
;
1681 generic_make_request(bio
);
1684 static void raid10d(mddev_t
*mddev
)
1687 unsigned long flags
;
1688 conf_t
*conf
= mddev
->private;
1689 struct list_head
*head
= &conf
->retry_list
;
1690 struct blk_plug plug
;
1692 md_check_recovery(mddev
);
1694 blk_start_plug(&plug
);
1697 flush_pending_writes(conf
);
1699 spin_lock_irqsave(&conf
->device_lock
, flags
);
1700 if (list_empty(head
)) {
1701 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1704 r10_bio
= list_entry(head
->prev
, r10bio_t
, retry_list
);
1705 list_del(head
->prev
);
1707 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1709 mddev
= r10_bio
->mddev
;
1710 conf
= mddev
->private;
1711 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
1712 sync_request_write(mddev
, r10_bio
);
1713 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
1714 recovery_request_write(mddev
, r10_bio
);
1716 handle_read_error(mddev
, r10_bio
);
1719 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
1720 md_check_recovery(mddev
);
1722 blk_finish_plug(&plug
);
1726 static int init_resync(conf_t
*conf
)
1730 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1731 BUG_ON(conf
->r10buf_pool
);
1732 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
1733 if (!conf
->r10buf_pool
)
1735 conf
->next_resync
= 0;
1740 * perform a "sync" on one "block"
1742 * We need to make sure that no normal I/O request - particularly write
1743 * requests - conflict with active sync requests.
1745 * This is achieved by tracking pending requests and a 'barrier' concept
1746 * that can be installed to exclude normal IO requests.
1748 * Resync and recovery are handled very differently.
1749 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1751 * For resync, we iterate over virtual addresses, read all copies,
1752 * and update if there are differences. If only one copy is live,
1754 * For recovery, we iterate over physical addresses, read a good
1755 * value for each non-in_sync drive, and over-write.
1757 * So, for recovery we may have several outstanding complex requests for a
1758 * given address, one for each out-of-sync device. We model this by allocating
1759 * a number of r10_bio structures, one for each out-of-sync device.
1760 * As we setup these structures, we collect all bio's together into a list
1761 * which we then process collectively to add pages, and then process again
1762 * to pass to generic_make_request.
1764 * The r10_bio structures are linked using a borrowed master_bio pointer.
1765 * This link is counted in ->remaining. When the r10_bio that points to NULL
1766 * has its remaining count decremented to 0, the whole complex operation
1771 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
,
1772 int *skipped
, int go_faster
)
1774 conf_t
*conf
= mddev
->private;
1776 struct bio
*biolist
= NULL
, *bio
;
1777 sector_t max_sector
, nr_sectors
;
1780 sector_t sync_blocks
;
1782 sector_t sectors_skipped
= 0;
1783 int chunks_skipped
= 0;
1785 if (!conf
->r10buf_pool
)
1786 if (init_resync(conf
))
1790 max_sector
= mddev
->dev_sectors
;
1791 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1792 max_sector
= mddev
->resync_max_sectors
;
1793 if (sector_nr
>= max_sector
) {
1794 /* If we aborted, we need to abort the
1795 * sync on the 'current' bitmap chucks (there can
1796 * be several when recovering multiple devices).
1797 * as we may have started syncing it but not finished.
1798 * We can find the current address in
1799 * mddev->curr_resync, but for recovery,
1800 * we need to convert that to several
1801 * virtual addresses.
1803 if (mddev
->curr_resync
< max_sector
) { /* aborted */
1804 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1805 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
1807 else for (i
=0; i
<conf
->raid_disks
; i
++) {
1809 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
1810 bitmap_end_sync(mddev
->bitmap
, sect
,
1813 } else /* completed sync */
1816 bitmap_close_sync(mddev
->bitmap
);
1819 return sectors_skipped
;
1821 if (chunks_skipped
>= conf
->raid_disks
) {
1822 /* if there has been nothing to do on any drive,
1823 * then there is nothing to do at all..
1826 return (max_sector
- sector_nr
) + sectors_skipped
;
1829 if (max_sector
> mddev
->resync_max
)
1830 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
1832 /* make sure whole request will fit in a chunk - if chunks
1835 if (conf
->near_copies
< conf
->raid_disks
&&
1836 max_sector
> (sector_nr
| conf
->chunk_mask
))
1837 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
1839 * If there is non-resync activity waiting for us then
1840 * put in a delay to throttle resync.
1842 if (!go_faster
&& conf
->nr_waiting
)
1843 msleep_interruptible(1000);
1845 /* Again, very different code for resync and recovery.
1846 * Both must result in an r10bio with a list of bios that
1847 * have bi_end_io, bi_sector, bi_bdev set,
1848 * and bi_private set to the r10bio.
1849 * For recovery, we may actually create several r10bios
1850 * with 2 bios in each, that correspond to the bios in the main one.
1851 * In this case, the subordinate r10bios link back through a
1852 * borrowed master_bio pointer, and the counter in the master
1853 * includes a ref from each subordinate.
1855 /* First, we decide what to do and set ->bi_end_io
1856 * To end_sync_read if we want to read, and
1857 * end_sync_write if we will want to write.
1860 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
1861 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
1862 /* recovery... the complicated one */
1866 for (i
=0 ; i
<conf
->raid_disks
; i
++) {
1872 if (conf
->mirrors
[i
].rdev
== NULL
||
1873 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
))
1877 /* want to reconstruct this device */
1879 sect
= raid10_find_virt(conf
, sector_nr
, i
);
1880 /* Unless we are doing a full sync, we only need
1881 * to recover the block if it is set in the bitmap
1883 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1885 if (sync_blocks
< max_sync
)
1886 max_sync
= sync_blocks
;
1889 /* yep, skip the sync_blocks here, but don't assume
1890 * that there will never be anything to do here
1892 chunks_skipped
= -1;
1896 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1897 raise_barrier(conf
, rb2
!= NULL
);
1898 atomic_set(&r10_bio
->remaining
, 0);
1900 r10_bio
->master_bio
= (struct bio
*)rb2
;
1902 atomic_inc(&rb2
->remaining
);
1903 r10_bio
->mddev
= mddev
;
1904 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
1905 r10_bio
->sector
= sect
;
1907 raid10_find_phys(conf
, r10_bio
);
1909 /* Need to check if the array will still be
1912 for (j
=0; j
<conf
->raid_disks
; j
++)
1913 if (conf
->mirrors
[j
].rdev
== NULL
||
1914 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
1919 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1920 &sync_blocks
, still_degraded
);
1922 for (j
=0; j
<conf
->copies
;j
++) {
1923 int d
= r10_bio
->devs
[j
].devnum
;
1924 if (!conf
->mirrors
[d
].rdev
||
1925 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
1927 /* This is where we read from */
1928 bio
= r10_bio
->devs
[0].bio
;
1929 bio
->bi_next
= biolist
;
1931 bio
->bi_private
= r10_bio
;
1932 bio
->bi_end_io
= end_sync_read
;
1934 bio
->bi_sector
= r10_bio
->devs
[j
].addr
+
1935 conf
->mirrors
[d
].rdev
->data_offset
;
1936 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1937 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1938 atomic_inc(&r10_bio
->remaining
);
1939 /* and we write to 'i' */
1941 for (k
=0; k
<conf
->copies
; k
++)
1942 if (r10_bio
->devs
[k
].devnum
== i
)
1944 BUG_ON(k
== conf
->copies
);
1945 bio
= r10_bio
->devs
[1].bio
;
1946 bio
->bi_next
= biolist
;
1948 bio
->bi_private
= r10_bio
;
1949 bio
->bi_end_io
= end_sync_write
;
1951 bio
->bi_sector
= r10_bio
->devs
[k
].addr
+
1952 conf
->mirrors
[i
].rdev
->data_offset
;
1953 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1955 r10_bio
->devs
[0].devnum
= d
;
1956 r10_bio
->devs
[1].devnum
= i
;
1960 if (j
== conf
->copies
) {
1961 /* Cannot recover, so abort the recovery */
1964 atomic_dec(&rb2
->remaining
);
1966 if (!test_and_set_bit(MD_RECOVERY_INTR
,
1968 printk(KERN_INFO
"md/raid10:%s: insufficient "
1969 "working devices for recovery.\n",
1974 if (biolist
== NULL
) {
1976 r10bio_t
*rb2
= r10_bio
;
1977 r10_bio
= (r10bio_t
*) rb2
->master_bio
;
1978 rb2
->master_bio
= NULL
;
1984 /* resync. Schedule a read for every block at this virt offset */
1987 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
1989 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
1990 &sync_blocks
, mddev
->degraded
) &&
1991 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
1992 &mddev
->recovery
)) {
1993 /* We can skip this block */
1995 return sync_blocks
+ sectors_skipped
;
1997 if (sync_blocks
< max_sync
)
1998 max_sync
= sync_blocks
;
1999 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2001 r10_bio
->mddev
= mddev
;
2002 atomic_set(&r10_bio
->remaining
, 0);
2003 raise_barrier(conf
, 0);
2004 conf
->next_resync
= sector_nr
;
2006 r10_bio
->master_bio
= NULL
;
2007 r10_bio
->sector
= sector_nr
;
2008 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
2009 raid10_find_phys(conf
, r10_bio
);
2010 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
2012 for (i
=0; i
<conf
->copies
; i
++) {
2013 int d
= r10_bio
->devs
[i
].devnum
;
2014 bio
= r10_bio
->devs
[i
].bio
;
2015 bio
->bi_end_io
= NULL
;
2016 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2017 if (conf
->mirrors
[d
].rdev
== NULL
||
2018 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
2020 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2021 atomic_inc(&r10_bio
->remaining
);
2022 bio
->bi_next
= biolist
;
2024 bio
->bi_private
= r10_bio
;
2025 bio
->bi_end_io
= end_sync_read
;
2027 bio
->bi_sector
= r10_bio
->devs
[i
].addr
+
2028 conf
->mirrors
[d
].rdev
->data_offset
;
2029 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2034 for (i
=0; i
<conf
->copies
; i
++) {
2035 int d
= r10_bio
->devs
[i
].devnum
;
2036 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
2037 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
2046 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2048 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2050 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
2053 bio
->bi_phys_segments
= 0;
2058 if (sector_nr
+ max_sync
< max_sector
)
2059 max_sector
= sector_nr
+ max_sync
;
2062 int len
= PAGE_SIZE
;
2063 if (sector_nr
+ (len
>>9) > max_sector
)
2064 len
= (max_sector
- sector_nr
) << 9;
2067 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2069 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2070 if (bio_add_page(bio
, page
, len
, 0))
2074 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2075 for (bio2
= biolist
;
2076 bio2
&& bio2
!= bio
;
2077 bio2
= bio2
->bi_next
) {
2078 /* remove last page from this bio */
2080 bio2
->bi_size
-= len
;
2081 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
2085 nr_sectors
+= len
>>9;
2086 sector_nr
+= len
>>9;
2087 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
2089 r10_bio
->sectors
= nr_sectors
;
2093 biolist
= biolist
->bi_next
;
2095 bio
->bi_next
= NULL
;
2096 r10_bio
= bio
->bi_private
;
2097 r10_bio
->sectors
= nr_sectors
;
2099 if (bio
->bi_end_io
== end_sync_read
) {
2100 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2101 generic_make_request(bio
);
2105 if (sectors_skipped
)
2106 /* pretend they weren't skipped, it makes
2107 * no important difference in this case
2109 md_done_sync(mddev
, sectors_skipped
, 1);
2111 return sectors_skipped
+ nr_sectors
;
2113 /* There is nowhere to write, so all non-sync
2114 * drives must be failed, so try the next chunk...
2116 if (sector_nr
+ max_sync
< max_sector
)
2117 max_sector
= sector_nr
+ max_sync
;
2119 sectors_skipped
+= (max_sector
- sector_nr
);
2121 sector_nr
= max_sector
;
2126 raid10_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
2129 conf_t
*conf
= mddev
->private;
2132 raid_disks
= conf
->raid_disks
;
2134 sectors
= conf
->dev_sectors
;
2136 size
= sectors
>> conf
->chunk_shift
;
2137 sector_div(size
, conf
->far_copies
);
2138 size
= size
* raid_disks
;
2139 sector_div(size
, conf
->near_copies
);
2141 return size
<< conf
->chunk_shift
;
2145 static conf_t
*setup_conf(mddev_t
*mddev
)
2147 conf_t
*conf
= NULL
;
2149 sector_t stride
, size
;
2152 if (mddev
->new_chunk_sectors
< (PAGE_SIZE
>> 9) ||
2153 !is_power_of_2(mddev
->new_chunk_sectors
)) {
2154 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
2155 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2156 mdname(mddev
), PAGE_SIZE
);
2160 nc
= mddev
->new_layout
& 255;
2161 fc
= (mddev
->new_layout
>> 8) & 255;
2162 fo
= mddev
->new_layout
& (1<<16);
2164 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
2165 (mddev
->new_layout
>> 17)) {
2166 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2167 mdname(mddev
), mddev
->new_layout
);
2172 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
2176 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
2181 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2186 conf
->raid_disks
= mddev
->raid_disks
;
2187 conf
->near_copies
= nc
;
2188 conf
->far_copies
= fc
;
2189 conf
->copies
= nc
*fc
;
2190 conf
->far_offset
= fo
;
2191 conf
->chunk_mask
= mddev
->new_chunk_sectors
- 1;
2192 conf
->chunk_shift
= ffz(~mddev
->new_chunk_sectors
);
2194 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
2195 r10bio_pool_free
, conf
);
2196 if (!conf
->r10bio_pool
)
2199 size
= mddev
->dev_sectors
>> conf
->chunk_shift
;
2200 sector_div(size
, fc
);
2201 size
= size
* conf
->raid_disks
;
2202 sector_div(size
, nc
);
2203 /* 'size' is now the number of chunks in the array */
2204 /* calculate "used chunks per device" in 'stride' */
2205 stride
= size
* conf
->copies
;
2207 /* We need to round up when dividing by raid_disks to
2208 * get the stride size.
2210 stride
+= conf
->raid_disks
- 1;
2211 sector_div(stride
, conf
->raid_disks
);
2213 conf
->dev_sectors
= stride
<< conf
->chunk_shift
;
2218 sector_div(stride
, fc
);
2219 conf
->stride
= stride
<< conf
->chunk_shift
;
2222 spin_lock_init(&conf
->device_lock
);
2223 INIT_LIST_HEAD(&conf
->retry_list
);
2225 spin_lock_init(&conf
->resync_lock
);
2226 init_waitqueue_head(&conf
->wait_barrier
);
2228 conf
->thread
= md_register_thread(raid10d
, mddev
, NULL
);
2232 conf
->mddev
= mddev
;
2236 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
2239 if (conf
->r10bio_pool
)
2240 mempool_destroy(conf
->r10bio_pool
);
2241 kfree(conf
->mirrors
);
2242 safe_put_page(conf
->tmppage
);
2245 return ERR_PTR(err
);
2248 static int run(mddev_t
*mddev
)
2251 int i
, disk_idx
, chunk_size
;
2252 mirror_info_t
*disk
;
2257 * copy the already verified devices into our private RAID10
2258 * bookkeeping area. [whatever we allocate in run(),
2259 * should be freed in stop()]
2262 if (mddev
->private == NULL
) {
2263 conf
= setup_conf(mddev
);
2265 return PTR_ERR(conf
);
2266 mddev
->private = conf
;
2268 conf
= mddev
->private;
2272 mddev
->thread
= conf
->thread
;
2273 conf
->thread
= NULL
;
2275 chunk_size
= mddev
->chunk_sectors
<< 9;
2276 blk_queue_io_min(mddev
->queue
, chunk_size
);
2277 if (conf
->raid_disks
% conf
->near_copies
)
2278 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->raid_disks
);
2280 blk_queue_io_opt(mddev
->queue
, chunk_size
*
2281 (conf
->raid_disks
/ conf
->near_copies
));
2283 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
2285 if (rdev
->badblocks
.count
) {
2286 printk(KERN_ERR
"md/raid10: cannot handle bad blocks yet\n");
2289 disk_idx
= rdev
->raid_disk
;
2290 if (disk_idx
>= conf
->raid_disks
2293 disk
= conf
->mirrors
+ disk_idx
;
2296 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2297 rdev
->data_offset
<< 9);
2298 /* as we don't honour merge_bvec_fn, we must never risk
2299 * violating it, so limit max_segments to 1 lying
2300 * within a single page.
2302 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
2303 blk_queue_max_segments(mddev
->queue
, 1);
2304 blk_queue_segment_boundary(mddev
->queue
,
2305 PAGE_CACHE_SIZE
- 1);
2308 disk
->head_position
= 0;
2310 /* need to check that every block has at least one working mirror */
2311 if (!enough(conf
, -1)) {
2312 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
2317 mddev
->degraded
= 0;
2318 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2320 disk
= conf
->mirrors
+ i
;
2323 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2324 disk
->head_position
= 0;
2331 if (mddev
->recovery_cp
!= MaxSector
)
2332 printk(KERN_NOTICE
"md/raid10:%s: not clean"
2333 " -- starting background reconstruction\n",
2336 "md/raid10:%s: active with %d out of %d devices\n",
2337 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
,
2340 * Ok, everything is just fine now
2342 mddev
->dev_sectors
= conf
->dev_sectors
;
2343 size
= raid10_size(mddev
, 0, 0);
2344 md_set_array_sectors(mddev
, size
);
2345 mddev
->resync_max_sectors
= size
;
2347 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
2348 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2350 /* Calculate max read-ahead size.
2351 * We need to readahead at least twice a whole stripe....
2355 int stripe
= conf
->raid_disks
*
2356 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
2357 stripe
/= conf
->near_copies
;
2358 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
2359 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
2362 if (conf
->near_copies
< conf
->raid_disks
)
2363 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
2365 if (md_integrity_register(mddev
))
2371 md_unregister_thread(mddev
->thread
);
2372 if (conf
->r10bio_pool
)
2373 mempool_destroy(conf
->r10bio_pool
);
2374 safe_put_page(conf
->tmppage
);
2375 kfree(conf
->mirrors
);
2377 mddev
->private = NULL
;
2382 static int stop(mddev_t
*mddev
)
2384 conf_t
*conf
= mddev
->private;
2386 raise_barrier(conf
, 0);
2387 lower_barrier(conf
);
2389 md_unregister_thread(mddev
->thread
);
2390 mddev
->thread
= NULL
;
2391 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
2392 if (conf
->r10bio_pool
)
2393 mempool_destroy(conf
->r10bio_pool
);
2394 kfree(conf
->mirrors
);
2396 mddev
->private = NULL
;
2400 static void raid10_quiesce(mddev_t
*mddev
, int state
)
2402 conf_t
*conf
= mddev
->private;
2406 raise_barrier(conf
, 0);
2409 lower_barrier(conf
);
2414 static void *raid10_takeover_raid0(mddev_t
*mddev
)
2419 if (mddev
->degraded
> 0) {
2420 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
2422 return ERR_PTR(-EINVAL
);
2425 /* Set new parameters */
2426 mddev
->new_level
= 10;
2427 /* new layout: far_copies = 1, near_copies = 2 */
2428 mddev
->new_layout
= (1<<8) + 2;
2429 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
2430 mddev
->delta_disks
= mddev
->raid_disks
;
2431 mddev
->raid_disks
*= 2;
2432 /* make sure it will be not marked as dirty */
2433 mddev
->recovery_cp
= MaxSector
;
2435 conf
= setup_conf(mddev
);
2436 if (!IS_ERR(conf
)) {
2437 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
2438 if (rdev
->raid_disk
>= 0)
2439 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
2446 static void *raid10_takeover(mddev_t
*mddev
)
2448 struct raid0_private_data
*raid0_priv
;
2450 /* raid10 can take over:
2451 * raid0 - providing it has only two drives
2453 if (mddev
->level
== 0) {
2454 /* for raid0 takeover only one zone is supported */
2455 raid0_priv
= mddev
->private;
2456 if (raid0_priv
->nr_strip_zones
> 1) {
2457 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
2458 " with more than one zone.\n",
2460 return ERR_PTR(-EINVAL
);
2462 return raid10_takeover_raid0(mddev
);
2464 return ERR_PTR(-EINVAL
);
2467 static struct mdk_personality raid10_personality
=
2471 .owner
= THIS_MODULE
,
2472 .make_request
= make_request
,
2476 .error_handler
= error
,
2477 .hot_add_disk
= raid10_add_disk
,
2478 .hot_remove_disk
= raid10_remove_disk
,
2479 .spare_active
= raid10_spare_active
,
2480 .sync_request
= sync_request
,
2481 .quiesce
= raid10_quiesce
,
2482 .size
= raid10_size
,
2483 .takeover
= raid10_takeover
,
2486 static int __init
raid_init(void)
2488 return register_md_personality(&raid10_personality
);
2491 static void raid_exit(void)
2493 unregister_md_personality(&raid10_personality
);
2496 module_init(raid_init
);
2497 module_exit(raid_exit
);
2498 MODULE_LICENSE("GPL");
2499 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2500 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2501 MODULE_ALIAS("md-raid10");
2502 MODULE_ALIAS("md-level-10");