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_SPECIAL(*bio
))
190 static void free_r10bio(r10bio_t
*r10_bio
)
192 conf_t
*conf
= r10_bio
->mddev
->private;
194 put_all_bios(conf
, r10_bio
);
195 mempool_free(r10_bio
, conf
->r10bio_pool
);
198 static void put_buf(r10bio_t
*r10_bio
)
200 conf_t
*conf
= r10_bio
->mddev
->private;
202 mempool_free(r10_bio
, conf
->r10buf_pool
);
207 static void reschedule_retry(r10bio_t
*r10_bio
)
210 mddev_t
*mddev
= r10_bio
->mddev
;
211 conf_t
*conf
= mddev
->private;
213 spin_lock_irqsave(&conf
->device_lock
, flags
);
214 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
216 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
218 /* wake up frozen array... */
219 wake_up(&conf
->wait_barrier
);
221 md_wakeup_thread(mddev
->thread
);
225 * raid_end_bio_io() is called when we have finished servicing a mirrored
226 * operation and are ready to return a success/failure code to the buffer
229 static void raid_end_bio_io(r10bio_t
*r10_bio
)
231 struct bio
*bio
= r10_bio
->master_bio
;
233 conf_t
*conf
= r10_bio
->mddev
->private;
235 if (bio
->bi_phys_segments
) {
237 spin_lock_irqsave(&conf
->device_lock
, flags
);
238 bio
->bi_phys_segments
--;
239 done
= (bio
->bi_phys_segments
== 0);
240 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
243 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
244 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
248 * Wake up any possible resync thread that waits for the device
253 free_r10bio(r10_bio
);
257 * Update disk head position estimator based on IRQ completion info.
259 static inline void update_head_pos(int slot
, r10bio_t
*r10_bio
)
261 conf_t
*conf
= r10_bio
->mddev
->private;
263 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
264 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
268 * Find the disk number which triggered given bio
270 static int find_bio_disk(conf_t
*conf
, r10bio_t
*r10_bio
,
271 struct bio
*bio
, int *slotp
)
275 for (slot
= 0; slot
< conf
->copies
; slot
++)
276 if (r10_bio
->devs
[slot
].bio
== bio
)
279 BUG_ON(slot
== conf
->copies
);
280 update_head_pos(slot
, r10_bio
);
284 return r10_bio
->devs
[slot
].devnum
;
287 static void raid10_end_read_request(struct bio
*bio
, int error
)
289 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
290 r10bio_t
*r10_bio
= bio
->bi_private
;
292 conf_t
*conf
= r10_bio
->mddev
->private;
295 slot
= r10_bio
->read_slot
;
296 dev
= r10_bio
->devs
[slot
].devnum
;
298 * this branch is our 'one mirror IO has finished' event handler:
300 update_head_pos(slot
, r10_bio
);
304 * Set R10BIO_Uptodate in our master bio, so that
305 * we will return a good error code to the higher
306 * levels even if IO on some other mirrored buffer fails.
308 * The 'master' represents the composite IO operation to
309 * user-side. So if something waits for IO, then it will
310 * wait for the 'master' bio.
312 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
313 raid_end_bio_io(r10_bio
);
314 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
317 * oops, read error - keep the refcount on the rdev
319 char b
[BDEVNAME_SIZE
];
320 printk_ratelimited(KERN_ERR
321 "md/raid10:%s: %s: rescheduling sector %llu\n",
323 bdevname(conf
->mirrors
[dev
].rdev
->bdev
, b
),
324 (unsigned long long)r10_bio
->sector
);
325 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
326 reschedule_retry(r10_bio
);
330 static void close_write(r10bio_t
*r10_bio
)
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
335 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
337 md_write_end(r10_bio
->mddev
);
340 static void raid10_end_write_request(struct bio
*bio
, int error
)
342 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
343 r10bio_t
*r10_bio
= bio
->bi_private
;
346 conf_t
*conf
= r10_bio
->mddev
->private;
349 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
);
352 * this branch is our 'one mirror IO has finished' event handler:
355 set_bit(WriteErrorSeen
, &conf
->mirrors
[dev
].rdev
->flags
);
356 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
360 * Set R10BIO_Uptodate in our master bio, so that
361 * we will return a good error code for to the higher
362 * levels even if IO on some other mirrored buffer fails.
364 * The 'master' represents the composite IO operation to
365 * user-side. So if something waits for IO, then it will
366 * wait for the 'master' bio.
371 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
373 /* Maybe we can clear some bad blocks. */
374 if (is_badblock(conf
->mirrors
[dev
].rdev
,
375 r10_bio
->devs
[slot
].addr
,
377 &first_bad
, &bad_sectors
)) {
379 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
381 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
387 * Let's see if all mirrored write operations have finished
390 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
391 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
392 reschedule_retry(r10_bio
);
394 close_write(r10_bio
);
395 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
396 reschedule_retry(r10_bio
);
398 raid_end_bio_io(r10_bio
);
402 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
407 * RAID10 layout manager
408 * As well as the chunksize and raid_disks count, there are two
409 * parameters: near_copies and far_copies.
410 * near_copies * far_copies must be <= raid_disks.
411 * Normally one of these will be 1.
412 * If both are 1, we get raid0.
413 * If near_copies == raid_disks, we get raid1.
415 * Chunks are laid out in raid0 style with near_copies copies of the
416 * first chunk, followed by near_copies copies of the next chunk and
418 * If far_copies > 1, then after 1/far_copies of the array has been assigned
419 * as described above, we start again with a device offset of near_copies.
420 * So we effectively have another copy of the whole array further down all
421 * the drives, but with blocks on different drives.
422 * With this layout, and block is never stored twice on the one device.
424 * raid10_find_phys finds the sector offset of a given virtual sector
425 * on each device that it is on.
427 * raid10_find_virt does the reverse mapping, from a device and a
428 * sector offset to a virtual address
431 static void raid10_find_phys(conf_t
*conf
, r10bio_t
*r10bio
)
441 /* now calculate first sector/dev */
442 chunk
= r10bio
->sector
>> conf
->chunk_shift
;
443 sector
= r10bio
->sector
& conf
->chunk_mask
;
445 chunk
*= conf
->near_copies
;
447 dev
= sector_div(stripe
, conf
->raid_disks
);
448 if (conf
->far_offset
)
449 stripe
*= conf
->far_copies
;
451 sector
+= stripe
<< conf
->chunk_shift
;
453 /* and calculate all the others */
454 for (n
=0; n
< conf
->near_copies
; n
++) {
457 r10bio
->devs
[slot
].addr
= sector
;
458 r10bio
->devs
[slot
].devnum
= d
;
461 for (f
= 1; f
< conf
->far_copies
; f
++) {
462 d
+= conf
->near_copies
;
463 if (d
>= conf
->raid_disks
)
464 d
-= conf
->raid_disks
;
466 r10bio
->devs
[slot
].devnum
= d
;
467 r10bio
->devs
[slot
].addr
= s
;
471 if (dev
>= conf
->raid_disks
) {
473 sector
+= (conf
->chunk_mask
+ 1);
476 BUG_ON(slot
!= conf
->copies
);
479 static sector_t
raid10_find_virt(conf_t
*conf
, sector_t sector
, int dev
)
481 sector_t offset
, chunk
, vchunk
;
483 offset
= sector
& conf
->chunk_mask
;
484 if (conf
->far_offset
) {
486 chunk
= sector
>> conf
->chunk_shift
;
487 fc
= sector_div(chunk
, conf
->far_copies
);
488 dev
-= fc
* conf
->near_copies
;
490 dev
+= conf
->raid_disks
;
492 while (sector
>= conf
->stride
) {
493 sector
-= conf
->stride
;
494 if (dev
< conf
->near_copies
)
495 dev
+= conf
->raid_disks
- conf
->near_copies
;
497 dev
-= conf
->near_copies
;
499 chunk
= sector
>> conf
->chunk_shift
;
501 vchunk
= chunk
* conf
->raid_disks
+ dev
;
502 sector_div(vchunk
, conf
->near_copies
);
503 return (vchunk
<< conf
->chunk_shift
) + offset
;
507 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
509 * @bvm: properties of new bio
510 * @biovec: the request that could be merged to it.
512 * Return amount of bytes we can accept at this offset
513 * If near_copies == raid_disk, there are no striping issues,
514 * but in that case, the function isn't called at all.
516 static int raid10_mergeable_bvec(struct request_queue
*q
,
517 struct bvec_merge_data
*bvm
,
518 struct bio_vec
*biovec
)
520 mddev_t
*mddev
= q
->queuedata
;
521 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
523 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
524 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
526 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
527 if (max
< 0) max
= 0; /* bio_add cannot handle a negative return */
528 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
529 return biovec
->bv_len
;
535 * This routine returns the disk from which the requested read should
536 * be done. There is a per-array 'next expected sequential IO' sector
537 * number - if this matches on the next IO then we use the last disk.
538 * There is also a per-disk 'last know head position' sector that is
539 * maintained from IRQ contexts, both the normal and the resync IO
540 * completion handlers update this position correctly. If there is no
541 * perfect sequential match then we pick the disk whose head is closest.
543 * If there are 2 mirrors in the same 2 devices, performance degrades
544 * because position is mirror, not device based.
546 * The rdev for the device selected will have nr_pending incremented.
550 * FIXME: possibly should rethink readbalancing and do it differently
551 * depending on near_copies / far_copies geometry.
553 static int read_balance(conf_t
*conf
, r10bio_t
*r10_bio
, int *max_sectors
)
555 const sector_t this_sector
= r10_bio
->sector
;
557 int sectors
= r10_bio
->sectors
;
558 int best_good_sectors
;
559 sector_t new_distance
, best_dist
;
564 raid10_find_phys(conf
, r10_bio
);
567 sectors
= r10_bio
->sectors
;
569 best_dist
= MaxSector
;
570 best_good_sectors
= 0;
573 * Check if we can balance. We can balance on the whole
574 * device if no resync is going on (recovery is ok), or below
575 * the resync window. We take the first readable disk when
576 * above the resync window.
578 if (conf
->mddev
->recovery_cp
< MaxSector
579 && (this_sector
+ sectors
>= conf
->next_resync
))
582 for (slot
= 0; slot
< conf
->copies
; slot
++) {
587 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
589 disk
= r10_bio
->devs
[slot
].devnum
;
590 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
593 if (!test_bit(In_sync
, &rdev
->flags
))
596 dev_sector
= r10_bio
->devs
[slot
].addr
;
597 if (is_badblock(rdev
, dev_sector
, sectors
,
598 &first_bad
, &bad_sectors
)) {
599 if (best_dist
< MaxSector
)
600 /* Already have a better slot */
602 if (first_bad
<= dev_sector
) {
603 /* Cannot read here. If this is the
604 * 'primary' device, then we must not read
605 * beyond 'bad_sectors' from another device.
607 bad_sectors
-= (dev_sector
- first_bad
);
608 if (!do_balance
&& sectors
> bad_sectors
)
609 sectors
= bad_sectors
;
610 if (best_good_sectors
> sectors
)
611 best_good_sectors
= sectors
;
613 sector_t good_sectors
=
614 first_bad
- dev_sector
;
615 if (good_sectors
> best_good_sectors
) {
616 best_good_sectors
= good_sectors
;
620 /* Must read from here */
625 best_good_sectors
= sectors
;
630 /* This optimisation is debatable, and completely destroys
631 * sequential read speed for 'far copies' arrays. So only
632 * keep it for 'near' arrays, and review those later.
634 if (conf
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
637 /* for far > 1 always use the lowest address */
638 if (conf
->far_copies
> 1)
639 new_distance
= r10_bio
->devs
[slot
].addr
;
641 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
642 conf
->mirrors
[disk
].head_position
);
643 if (new_distance
< best_dist
) {
644 best_dist
= new_distance
;
648 if (slot
== conf
->copies
)
652 disk
= r10_bio
->devs
[slot
].devnum
;
653 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
656 atomic_inc(&rdev
->nr_pending
);
657 if (test_bit(Faulty
, &rdev
->flags
)) {
658 /* Cannot risk returning a device that failed
659 * before we inc'ed nr_pending
661 rdev_dec_pending(rdev
, conf
->mddev
);
664 r10_bio
->read_slot
= slot
;
668 *max_sectors
= best_good_sectors
;
673 static int raid10_congested(void *data
, int bits
)
675 mddev_t
*mddev
= data
;
676 conf_t
*conf
= mddev
->private;
679 if (mddev_congested(mddev
, bits
))
682 for (i
= 0; i
< conf
->raid_disks
&& ret
== 0; i
++) {
683 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
684 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
685 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
687 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
694 static void flush_pending_writes(conf_t
*conf
)
696 /* Any writes that have been queued but are awaiting
697 * bitmap updates get flushed here.
699 spin_lock_irq(&conf
->device_lock
);
701 if (conf
->pending_bio_list
.head
) {
703 bio
= bio_list_get(&conf
->pending_bio_list
);
704 spin_unlock_irq(&conf
->device_lock
);
705 /* flush any pending bitmap writes to disk
706 * before proceeding w/ I/O */
707 bitmap_unplug(conf
->mddev
->bitmap
);
709 while (bio
) { /* submit pending writes */
710 struct bio
*next
= bio
->bi_next
;
712 generic_make_request(bio
);
716 spin_unlock_irq(&conf
->device_lock
);
720 * Sometimes we need to suspend IO while we do something else,
721 * either some resync/recovery, or reconfigure the array.
722 * To do this we raise a 'barrier'.
723 * The 'barrier' is a counter that can be raised multiple times
724 * to count how many activities are happening which preclude
726 * We can only raise the barrier if there is no pending IO.
727 * i.e. if nr_pending == 0.
728 * We choose only to raise the barrier if no-one is waiting for the
729 * barrier to go down. This means that as soon as an IO request
730 * is ready, no other operations which require a barrier will start
731 * until the IO request has had a chance.
733 * So: regular IO calls 'wait_barrier'. When that returns there
734 * is no backgroup IO happening, It must arrange to call
735 * allow_barrier when it has finished its IO.
736 * backgroup IO calls must call raise_barrier. Once that returns
737 * there is no normal IO happeing. It must arrange to call
738 * lower_barrier when the particular background IO completes.
741 static void raise_barrier(conf_t
*conf
, int force
)
743 BUG_ON(force
&& !conf
->barrier
);
744 spin_lock_irq(&conf
->resync_lock
);
746 /* Wait until no block IO is waiting (unless 'force') */
747 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
748 conf
->resync_lock
, );
750 /* block any new IO from starting */
753 /* Now wait for all pending IO to complete */
754 wait_event_lock_irq(conf
->wait_barrier
,
755 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
756 conf
->resync_lock
, );
758 spin_unlock_irq(&conf
->resync_lock
);
761 static void lower_barrier(conf_t
*conf
)
764 spin_lock_irqsave(&conf
->resync_lock
, flags
);
766 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
767 wake_up(&conf
->wait_barrier
);
770 static void wait_barrier(conf_t
*conf
)
772 spin_lock_irq(&conf
->resync_lock
);
775 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
781 spin_unlock_irq(&conf
->resync_lock
);
784 static void allow_barrier(conf_t
*conf
)
787 spin_lock_irqsave(&conf
->resync_lock
, flags
);
789 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
790 wake_up(&conf
->wait_barrier
);
793 static void freeze_array(conf_t
*conf
)
795 /* stop syncio and normal IO and wait for everything to
797 * We increment barrier and nr_waiting, and then
798 * wait until nr_pending match nr_queued+1
799 * This is called in the context of one normal IO request
800 * that has failed. Thus any sync request that might be pending
801 * will be blocked by nr_pending, and we need to wait for
802 * pending IO requests to complete or be queued for re-try.
803 * Thus the number queued (nr_queued) plus this request (1)
804 * must match the number of pending IOs (nr_pending) before
807 spin_lock_irq(&conf
->resync_lock
);
810 wait_event_lock_irq(conf
->wait_barrier
,
811 conf
->nr_pending
== conf
->nr_queued
+1,
813 flush_pending_writes(conf
));
815 spin_unlock_irq(&conf
->resync_lock
);
818 static void unfreeze_array(conf_t
*conf
)
820 /* reverse the effect of the freeze */
821 spin_lock_irq(&conf
->resync_lock
);
824 wake_up(&conf
->wait_barrier
);
825 spin_unlock_irq(&conf
->resync_lock
);
828 static int make_request(mddev_t
*mddev
, struct bio
* bio
)
830 conf_t
*conf
= mddev
->private;
831 mirror_info_t
*mirror
;
833 struct bio
*read_bio
;
835 int chunk_sects
= conf
->chunk_mask
+ 1;
836 const int rw
= bio_data_dir(bio
);
837 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
838 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
840 mdk_rdev_t
*blocked_rdev
;
845 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
846 md_flush_request(mddev
, bio
);
850 /* If this request crosses a chunk boundary, we need to
851 * split it. This will only happen for 1 PAGE (or less) requests.
853 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
855 conf
->near_copies
< conf
->raid_disks
)) {
857 /* Sanity check -- queue functions should prevent this happening */
858 if (bio
->bi_vcnt
!= 1 ||
861 /* This is a one page bio that upper layers
862 * refuse to split for us, so we need to split it.
865 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
867 /* Each of these 'make_request' calls will call 'wait_barrier'.
868 * If the first succeeds but the second blocks due to the resync
869 * thread raising the barrier, we will deadlock because the
870 * IO to the underlying device will be queued in generic_make_request
871 * and will never complete, so will never reduce nr_pending.
872 * So increment nr_waiting here so no new raise_barriers will
873 * succeed, and so the second wait_barrier cannot block.
875 spin_lock_irq(&conf
->resync_lock
);
877 spin_unlock_irq(&conf
->resync_lock
);
879 if (make_request(mddev
, &bp
->bio1
))
880 generic_make_request(&bp
->bio1
);
881 if (make_request(mddev
, &bp
->bio2
))
882 generic_make_request(&bp
->bio2
);
884 spin_lock_irq(&conf
->resync_lock
);
886 wake_up(&conf
->wait_barrier
);
887 spin_unlock_irq(&conf
->resync_lock
);
889 bio_pair_release(bp
);
892 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
893 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
894 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
900 md_write_start(mddev
, bio
);
903 * Register the new request and wait if the reconstruction
904 * thread has put up a bar for new requests.
905 * Continue immediately if no resync is active currently.
909 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
911 r10_bio
->master_bio
= bio
;
912 r10_bio
->sectors
= bio
->bi_size
>> 9;
914 r10_bio
->mddev
= mddev
;
915 r10_bio
->sector
= bio
->bi_sector
;
918 /* We might need to issue multiple reads to different
919 * devices if there are bad blocks around, so we keep
920 * track of the number of reads in bio->bi_phys_segments.
921 * If this is 0, there is only one r10_bio and no locking
922 * will be needed when the request completes. If it is
923 * non-zero, then it is the number of not-completed requests.
925 bio
->bi_phys_segments
= 0;
926 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
930 * read balancing logic:
936 disk
= read_balance(conf
, r10_bio
, &max_sectors
);
937 slot
= r10_bio
->read_slot
;
939 raid_end_bio_io(r10_bio
);
942 mirror
= conf
->mirrors
+ disk
;
944 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
945 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
948 r10_bio
->devs
[slot
].bio
= read_bio
;
950 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
951 mirror
->rdev
->data_offset
;
952 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
953 read_bio
->bi_end_io
= raid10_end_read_request
;
954 read_bio
->bi_rw
= READ
| do_sync
;
955 read_bio
->bi_private
= r10_bio
;
957 if (max_sectors
< r10_bio
->sectors
) {
958 /* Could not read all from this device, so we will
959 * need another r10_bio.
961 sectors_handled
= (r10_bio
->sectors
+ max_sectors
963 r10_bio
->sectors
= max_sectors
;
964 spin_lock_irq(&conf
->device_lock
);
965 if (bio
->bi_phys_segments
== 0)
966 bio
->bi_phys_segments
= 2;
968 bio
->bi_phys_segments
++;
969 spin_unlock(&conf
->device_lock
);
970 /* Cannot call generic_make_request directly
971 * as that will be queued in __generic_make_request
972 * and subsequent mempool_alloc might block
973 * waiting for it. so hand bio over to raid10d.
975 reschedule_retry(r10_bio
);
977 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
979 r10_bio
->master_bio
= bio
;
980 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
983 r10_bio
->mddev
= mddev
;
984 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
987 generic_make_request(read_bio
);
994 /* first select target devices under rcu_lock and
995 * inc refcount on their rdev. Record them by setting
997 * If there are known/acknowledged bad blocks on any device
998 * on which we have seen a write error, we want to avoid
999 * writing to those blocks. This potentially requires several
1000 * writes to write around the bad blocks. Each set of writes
1001 * gets its own r10_bio with a set of bios attached. The number
1002 * of r10_bios is recored in bio->bi_phys_segments just as with
1005 plugged
= mddev_check_plugged(mddev
);
1007 raid10_find_phys(conf
, r10_bio
);
1009 blocked_rdev
= NULL
;
1011 max_sectors
= r10_bio
->sectors
;
1013 for (i
= 0; i
< conf
->copies
; i
++) {
1014 int d
= r10_bio
->devs
[i
].devnum
;
1015 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1016 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1017 atomic_inc(&rdev
->nr_pending
);
1018 blocked_rdev
= rdev
;
1021 r10_bio
->devs
[i
].bio
= NULL
;
1022 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
1023 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1026 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1028 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1032 is_bad
= is_badblock(rdev
, dev_sector
,
1034 &first_bad
, &bad_sectors
);
1036 /* Mustn't write here until the bad block
1039 atomic_inc(&rdev
->nr_pending
);
1040 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1041 blocked_rdev
= rdev
;
1044 if (is_bad
&& first_bad
<= dev_sector
) {
1045 /* Cannot write here at all */
1046 bad_sectors
-= (dev_sector
- first_bad
);
1047 if (bad_sectors
< max_sectors
)
1048 /* Mustn't write more than bad_sectors
1049 * to other devices yet
1051 max_sectors
= bad_sectors
;
1052 /* We don't set R10BIO_Degraded as that
1053 * only applies if the disk is missing,
1054 * so it might be re-added, and we want to
1055 * know to recover this chunk.
1056 * In this case the device is here, and the
1057 * fact that this chunk is not in-sync is
1058 * recorded in the bad block log.
1063 int good_sectors
= first_bad
- dev_sector
;
1064 if (good_sectors
< max_sectors
)
1065 max_sectors
= good_sectors
;
1068 r10_bio
->devs
[i
].bio
= bio
;
1069 atomic_inc(&rdev
->nr_pending
);
1073 if (unlikely(blocked_rdev
)) {
1074 /* Have to wait for this device to get unblocked, then retry */
1078 for (j
= 0; j
< i
; j
++)
1079 if (r10_bio
->devs
[j
].bio
) {
1080 d
= r10_bio
->devs
[j
].devnum
;
1081 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1083 allow_barrier(conf
);
1084 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1089 if (max_sectors
< r10_bio
->sectors
) {
1090 /* We are splitting this into multiple parts, so
1091 * we need to prepare for allocating another r10_bio.
1093 r10_bio
->sectors
= max_sectors
;
1094 spin_lock_irq(&conf
->device_lock
);
1095 if (bio
->bi_phys_segments
== 0)
1096 bio
->bi_phys_segments
= 2;
1098 bio
->bi_phys_segments
++;
1099 spin_unlock_irq(&conf
->device_lock
);
1101 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1103 atomic_set(&r10_bio
->remaining
, 1);
1104 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1106 for (i
= 0; i
< conf
->copies
; i
++) {
1108 int d
= r10_bio
->devs
[i
].devnum
;
1109 if (!r10_bio
->devs
[i
].bio
)
1112 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1113 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1115 r10_bio
->devs
[i
].bio
= mbio
;
1117 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1118 conf
->mirrors
[d
].rdev
->data_offset
);
1119 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1120 mbio
->bi_end_io
= raid10_end_write_request
;
1121 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
1122 mbio
->bi_private
= r10_bio
;
1124 atomic_inc(&r10_bio
->remaining
);
1125 spin_lock_irqsave(&conf
->device_lock
, flags
);
1126 bio_list_add(&conf
->pending_bio_list
, mbio
);
1127 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1130 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1131 /* This matches the end of raid10_end_write_request() */
1132 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
1134 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
1136 md_write_end(mddev
);
1137 raid_end_bio_io(r10_bio
);
1140 /* In case raid10d snuck in to freeze_array */
1141 wake_up(&conf
->wait_barrier
);
1143 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1144 /* We need another r10_bio. It has already been counted
1145 * in bio->bi_phys_segments.
1147 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1149 r10_bio
->master_bio
= bio
;
1150 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1152 r10_bio
->mddev
= mddev
;
1153 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1158 if (do_sync
|| !mddev
->bitmap
|| !plugged
)
1159 md_wakeup_thread(mddev
->thread
);
1163 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
1165 conf_t
*conf
= mddev
->private;
1168 if (conf
->near_copies
< conf
->raid_disks
)
1169 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1170 if (conf
->near_copies
> 1)
1171 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
1172 if (conf
->far_copies
> 1) {
1173 if (conf
->far_offset
)
1174 seq_printf(seq
, " %d offset-copies", conf
->far_copies
);
1176 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
1178 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1179 conf
->raid_disks
- mddev
->degraded
);
1180 for (i
= 0; i
< conf
->raid_disks
; i
++)
1181 seq_printf(seq
, "%s",
1182 conf
->mirrors
[i
].rdev
&&
1183 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1184 seq_printf(seq
, "]");
1187 /* check if there are enough drives for
1188 * every block to appear on atleast one.
1189 * Don't consider the device numbered 'ignore'
1190 * as we might be about to remove it.
1192 static int enough(conf_t
*conf
, int ignore
)
1197 int n
= conf
->copies
;
1200 if (conf
->mirrors
[first
].rdev
&&
1203 first
= (first
+1) % conf
->raid_disks
;
1207 } while (first
!= 0);
1211 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1213 char b
[BDEVNAME_SIZE
];
1214 conf_t
*conf
= mddev
->private;
1217 * If it is not operational, then we have already marked it as dead
1218 * else if it is the last working disks, ignore the error, let the
1219 * next level up know.
1220 * else mark the drive as failed
1222 if (test_bit(In_sync
, &rdev
->flags
)
1223 && !enough(conf
, rdev
->raid_disk
))
1225 * Don't fail the drive, just return an IO error.
1228 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1229 unsigned long flags
;
1230 spin_lock_irqsave(&conf
->device_lock
, flags
);
1232 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1234 * if recovery is running, make sure it aborts.
1236 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1238 set_bit(Blocked
, &rdev
->flags
);
1239 set_bit(Faulty
, &rdev
->flags
);
1240 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1242 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1243 "md/raid10:%s: Operation continuing on %d devices.\n",
1244 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1245 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1248 static void print_conf(conf_t
*conf
)
1253 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1255 printk(KERN_DEBUG
"(!conf)\n");
1258 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1261 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1262 char b
[BDEVNAME_SIZE
];
1263 tmp
= conf
->mirrors
+ i
;
1265 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1266 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1267 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1268 bdevname(tmp
->rdev
->bdev
,b
));
1272 static void close_sync(conf_t
*conf
)
1275 allow_barrier(conf
);
1277 mempool_destroy(conf
->r10buf_pool
);
1278 conf
->r10buf_pool
= NULL
;
1281 static int raid10_spare_active(mddev_t
*mddev
)
1284 conf_t
*conf
= mddev
->private;
1287 unsigned long flags
;
1290 * Find all non-in_sync disks within the RAID10 configuration
1291 * and mark them in_sync
1293 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1294 tmp
= conf
->mirrors
+ i
;
1296 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1297 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1299 sysfs_notify_dirent(tmp
->rdev
->sysfs_state
);
1302 spin_lock_irqsave(&conf
->device_lock
, flags
);
1303 mddev
->degraded
-= count
;
1304 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1311 static int raid10_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1313 conf_t
*conf
= mddev
->private;
1317 int last
= conf
->raid_disks
- 1;
1319 if (mddev
->recovery_cp
< MaxSector
)
1320 /* only hot-add to in-sync arrays, as recovery is
1321 * very different from resync
1324 if (!enough(conf
, -1))
1327 if (rdev
->raid_disk
>= 0)
1328 first
= last
= rdev
->raid_disk
;
1330 if (rdev
->saved_raid_disk
>= first
&&
1331 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1332 mirror
= rdev
->saved_raid_disk
;
1335 for ( ; mirror
<= last
; mirror
++) {
1336 mirror_info_t
*p
= &conf
->mirrors
[mirror
];
1337 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1342 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1343 rdev
->data_offset
<< 9);
1344 /* as we don't honour merge_bvec_fn, we must
1345 * never risk violating it, so limit
1346 * ->max_segments to one lying with a single
1347 * page, as a one page request is never in
1350 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1351 blk_queue_max_segments(mddev
->queue
, 1);
1352 blk_queue_segment_boundary(mddev
->queue
,
1353 PAGE_CACHE_SIZE
- 1);
1356 p
->head_position
= 0;
1357 rdev
->raid_disk
= mirror
;
1359 if (rdev
->saved_raid_disk
!= mirror
)
1361 rcu_assign_pointer(p
->rdev
, rdev
);
1365 md_integrity_add_rdev(rdev
, mddev
);
1370 static int raid10_remove_disk(mddev_t
*mddev
, int number
)
1372 conf_t
*conf
= mddev
->private;
1375 mirror_info_t
*p
= conf
->mirrors
+ number
;
1380 if (test_bit(In_sync
, &rdev
->flags
) ||
1381 atomic_read(&rdev
->nr_pending
)) {
1385 /* Only remove faulty devices in recovery
1388 if (!test_bit(Faulty
, &rdev
->flags
) &&
1389 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1396 if (atomic_read(&rdev
->nr_pending
)) {
1397 /* lost the race, try later */
1402 err
= md_integrity_register(mddev
);
1411 static void end_sync_read(struct bio
*bio
, int error
)
1413 r10bio_t
*r10_bio
= bio
->bi_private
;
1414 conf_t
*conf
= r10_bio
->mddev
->private;
1417 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
);
1419 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1420 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1422 /* The write handler will notice the lack of
1423 * R10BIO_Uptodate and record any errors etc
1425 atomic_add(r10_bio
->sectors
,
1426 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1428 /* for reconstruct, we always reschedule after a read.
1429 * for resync, only after all reads
1431 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1432 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1433 atomic_dec_and_test(&r10_bio
->remaining
)) {
1434 /* we have read all the blocks,
1435 * do the comparison in process context in raid10d
1437 reschedule_retry(r10_bio
);
1441 static void end_sync_request(r10bio_t
*r10_bio
)
1443 mddev_t
*mddev
= r10_bio
->mddev
;
1445 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1446 if (r10_bio
->master_bio
== NULL
) {
1447 /* the primary of several recovery bios */
1448 sector_t s
= r10_bio
->sectors
;
1449 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1450 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1451 reschedule_retry(r10_bio
);
1454 md_done_sync(mddev
, s
, 1);
1457 r10bio_t
*r10_bio2
= (r10bio_t
*)r10_bio
->master_bio
;
1458 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1459 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1460 reschedule_retry(r10_bio
);
1468 static void end_sync_write(struct bio
*bio
, int error
)
1470 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1471 r10bio_t
*r10_bio
= bio
->bi_private
;
1472 mddev_t
*mddev
= r10_bio
->mddev
;
1473 conf_t
*conf
= mddev
->private;
1479 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
);
1482 set_bit(WriteErrorSeen
, &conf
->mirrors
[d
].rdev
->flags
);
1483 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1484 } else if (is_badblock(conf
->mirrors
[d
].rdev
,
1485 r10_bio
->devs
[slot
].addr
,
1487 &first_bad
, &bad_sectors
))
1488 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1490 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1492 end_sync_request(r10_bio
);
1496 * Note: sync and recover and handled very differently for raid10
1497 * This code is for resync.
1498 * For resync, we read through virtual addresses and read all blocks.
1499 * If there is any error, we schedule a write. The lowest numbered
1500 * drive is authoritative.
1501 * However requests come for physical address, so we need to map.
1502 * For every physical address there are raid_disks/copies virtual addresses,
1503 * which is always are least one, but is not necessarly an integer.
1504 * This means that a physical address can span multiple chunks, so we may
1505 * have to submit multiple io requests for a single sync request.
1508 * We check if all blocks are in-sync and only write to blocks that
1511 static void sync_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1513 conf_t
*conf
= mddev
->private;
1515 struct bio
*tbio
, *fbio
;
1517 atomic_set(&r10_bio
->remaining
, 1);
1519 /* find the first device with a block */
1520 for (i
=0; i
<conf
->copies
; i
++)
1521 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1524 if (i
== conf
->copies
)
1528 fbio
= r10_bio
->devs
[i
].bio
;
1530 /* now find blocks with errors */
1531 for (i
=0 ; i
< conf
->copies
; i
++) {
1533 int vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1535 tbio
= r10_bio
->devs
[i
].bio
;
1537 if (tbio
->bi_end_io
!= end_sync_read
)
1541 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1542 /* We know that the bi_io_vec layout is the same for
1543 * both 'first' and 'i', so we just compare them.
1544 * All vec entries are PAGE_SIZE;
1546 for (j
= 0; j
< vcnt
; j
++)
1547 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1548 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1553 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1554 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1555 /* Don't fix anything. */
1558 /* Ok, we need to write this bio, either to correct an
1559 * inconsistency or to correct an unreadable block.
1560 * First we need to fixup bv_offset, bv_len and
1561 * bi_vecs, as the read request might have corrupted these
1563 tbio
->bi_vcnt
= vcnt
;
1564 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1566 tbio
->bi_phys_segments
= 0;
1567 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1568 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1569 tbio
->bi_next
= NULL
;
1570 tbio
->bi_rw
= WRITE
;
1571 tbio
->bi_private
= r10_bio
;
1572 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1574 for (j
=0; j
< vcnt
; j
++) {
1575 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1576 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1578 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1579 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1582 tbio
->bi_end_io
= end_sync_write
;
1584 d
= r10_bio
->devs
[i
].devnum
;
1585 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1586 atomic_inc(&r10_bio
->remaining
);
1587 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1589 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1590 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1591 generic_make_request(tbio
);
1595 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1596 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1602 * Now for the recovery code.
1603 * Recovery happens across physical sectors.
1604 * We recover all non-is_sync drives by finding the virtual address of
1605 * each, and then choose a working drive that also has that virt address.
1606 * There is a separate r10_bio for each non-in_sync drive.
1607 * Only the first two slots are in use. The first for reading,
1608 * The second for writing.
1611 static void fix_recovery_read_error(r10bio_t
*r10_bio
)
1613 /* We got a read error during recovery.
1614 * We repeat the read in smaller page-sized sections.
1615 * If a read succeeds, write it to the new device or record
1616 * a bad block if we cannot.
1617 * If a read fails, record a bad block on both old and
1620 mddev_t
*mddev
= r10_bio
->mddev
;
1621 conf_t
*conf
= mddev
->private;
1622 struct bio
*bio
= r10_bio
->devs
[0].bio
;
1624 int sectors
= r10_bio
->sectors
;
1626 int dr
= r10_bio
->devs
[0].devnum
;
1627 int dw
= r10_bio
->devs
[1].devnum
;
1635 if (s
> (PAGE_SIZE
>>9))
1638 rdev
= conf
->mirrors
[dr
].rdev
;
1639 addr
= r10_bio
->devs
[0].addr
+ sect
,
1640 ok
= sync_page_io(rdev
,
1643 bio
->bi_io_vec
[idx
].bv_page
,
1646 rdev
= conf
->mirrors
[dw
].rdev
;
1647 addr
= r10_bio
->devs
[1].addr
+ sect
;
1648 ok
= sync_page_io(rdev
,
1651 bio
->bi_io_vec
[idx
].bv_page
,
1654 set_bit(WriteErrorSeen
, &rdev
->flags
);
1657 /* We don't worry if we cannot set a bad block -
1658 * it really is bad so there is no loss in not
1661 rdev_set_badblocks(rdev
, addr
, s
, 0);
1663 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
1664 /* need bad block on destination too */
1665 mdk_rdev_t
*rdev2
= conf
->mirrors
[dw
].rdev
;
1666 addr
= r10_bio
->devs
[1].addr
+ sect
;
1667 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
1669 /* just abort the recovery */
1671 "md/raid10:%s: recovery aborted"
1672 " due to read error\n",
1675 conf
->mirrors
[dw
].recovery_disabled
1676 = mddev
->recovery_disabled
;
1677 set_bit(MD_RECOVERY_INTR
,
1690 static void recovery_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1692 conf_t
*conf
= mddev
->private;
1696 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
1697 fix_recovery_read_error(r10_bio
);
1698 end_sync_request(r10_bio
);
1703 * share the pages with the first bio
1704 * and submit the write request
1706 wbio
= r10_bio
->devs
[1].bio
;
1707 d
= r10_bio
->devs
[1].devnum
;
1709 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1710 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1711 generic_make_request(wbio
);
1716 * Used by fix_read_error() to decay the per rdev read_errors.
1717 * We halve the read error count for every hour that has elapsed
1718 * since the last recorded read error.
1721 static void check_decay_read_errors(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1723 struct timespec cur_time_mon
;
1724 unsigned long hours_since_last
;
1725 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
1727 ktime_get_ts(&cur_time_mon
);
1729 if (rdev
->last_read_error
.tv_sec
== 0 &&
1730 rdev
->last_read_error
.tv_nsec
== 0) {
1731 /* first time we've seen a read error */
1732 rdev
->last_read_error
= cur_time_mon
;
1736 hours_since_last
= (cur_time_mon
.tv_sec
-
1737 rdev
->last_read_error
.tv_sec
) / 3600;
1739 rdev
->last_read_error
= cur_time_mon
;
1742 * if hours_since_last is > the number of bits in read_errors
1743 * just set read errors to 0. We do this to avoid
1744 * overflowing the shift of read_errors by hours_since_last.
1746 if (hours_since_last
>= 8 * sizeof(read_errors
))
1747 atomic_set(&rdev
->read_errors
, 0);
1749 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
1753 * This is a kernel thread which:
1755 * 1. Retries failed read operations on working mirrors.
1756 * 2. Updates the raid superblock when problems encounter.
1757 * 3. Performs writes following reads for array synchronising.
1760 static void fix_read_error(conf_t
*conf
, mddev_t
*mddev
, r10bio_t
*r10_bio
)
1762 int sect
= 0; /* Offset from r10_bio->sector */
1763 int sectors
= r10_bio
->sectors
;
1765 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
1766 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1768 /* still own a reference to this rdev, so it cannot
1769 * have been cleared recently.
1771 rdev
= conf
->mirrors
[d
].rdev
;
1773 if (test_bit(Faulty
, &rdev
->flags
))
1774 /* drive has already been failed, just ignore any
1775 more fix_read_error() attempts */
1778 check_decay_read_errors(mddev
, rdev
);
1779 atomic_inc(&rdev
->read_errors
);
1780 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
1781 char b
[BDEVNAME_SIZE
];
1782 bdevname(rdev
->bdev
, b
);
1785 "md/raid10:%s: %s: Raid device exceeded "
1786 "read_error threshold [cur %d:max %d]\n",
1788 atomic_read(&rdev
->read_errors
), max_read_errors
);
1790 "md/raid10:%s: %s: Failing raid device\n",
1792 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1798 int sl
= r10_bio
->read_slot
;
1802 if (s
> (PAGE_SIZE
>>9))
1810 d
= r10_bio
->devs
[sl
].devnum
;
1811 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1813 test_bit(In_sync
, &rdev
->flags
) &&
1814 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
1815 &first_bad
, &bad_sectors
) == 0) {
1816 atomic_inc(&rdev
->nr_pending
);
1818 success
= sync_page_io(rdev
,
1819 r10_bio
->devs
[sl
].addr
+
1822 conf
->tmppage
, READ
, false);
1823 rdev_dec_pending(rdev
, mddev
);
1829 if (sl
== conf
->copies
)
1831 } while (!success
&& sl
!= r10_bio
->read_slot
);
1835 /* Cannot read from anywhere -- bye bye array */
1836 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1837 md_error(mddev
, conf
->mirrors
[dn
].rdev
);
1842 /* write it back and re-read */
1844 while (sl
!= r10_bio
->read_slot
) {
1845 char b
[BDEVNAME_SIZE
];
1850 d
= r10_bio
->devs
[sl
].devnum
;
1851 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1853 !test_bit(In_sync
, &rdev
->flags
))
1856 atomic_inc(&rdev
->nr_pending
);
1858 if (sync_page_io(rdev
,
1859 r10_bio
->devs
[sl
].addr
+
1861 s
<<9, conf
->tmppage
, WRITE
, false)
1863 /* Well, this device is dead */
1865 "md/raid10:%s: read correction "
1867 " (%d sectors at %llu on %s)\n",
1869 (unsigned long long)(
1870 sect
+ rdev
->data_offset
),
1871 bdevname(rdev
->bdev
, b
));
1872 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
1875 bdevname(rdev
->bdev
, b
));
1876 md_error(mddev
, rdev
);
1878 rdev_dec_pending(rdev
, mddev
);
1882 while (sl
!= r10_bio
->read_slot
) {
1883 char b
[BDEVNAME_SIZE
];
1888 d
= r10_bio
->devs
[sl
].devnum
;
1889 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1891 !test_bit(In_sync
, &rdev
->flags
))
1894 atomic_inc(&rdev
->nr_pending
);
1896 if (sync_page_io(rdev
,
1897 r10_bio
->devs
[sl
].addr
+
1899 s
<<9, conf
->tmppage
,
1900 READ
, false) == 0) {
1901 /* Well, this device is dead */
1903 "md/raid10:%s: unable to read back "
1905 " (%d sectors at %llu on %s)\n",
1907 (unsigned long long)(
1908 sect
+ rdev
->data_offset
),
1909 bdevname(rdev
->bdev
, b
));
1910 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
1913 bdevname(rdev
->bdev
, b
));
1915 md_error(mddev
, rdev
);
1918 "md/raid10:%s: read error corrected"
1919 " (%d sectors at %llu on %s)\n",
1921 (unsigned long long)(
1922 sect
+ rdev
->data_offset
),
1923 bdevname(rdev
->bdev
, b
));
1924 atomic_add(s
, &rdev
->corrected_errors
);
1927 rdev_dec_pending(rdev
, mddev
);
1937 static void bi_complete(struct bio
*bio
, int error
)
1939 complete((struct completion
*)bio
->bi_private
);
1942 static int submit_bio_wait(int rw
, struct bio
*bio
)
1944 struct completion event
;
1947 init_completion(&event
);
1948 bio
->bi_private
= &event
;
1949 bio
->bi_end_io
= bi_complete
;
1950 submit_bio(rw
, bio
);
1951 wait_for_completion(&event
);
1953 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1956 static int narrow_write_error(r10bio_t
*r10_bio
, int i
)
1958 struct bio
*bio
= r10_bio
->master_bio
;
1959 mddev_t
*mddev
= r10_bio
->mddev
;
1960 conf_t
*conf
= mddev
->private;
1961 mdk_rdev_t
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
1962 /* bio has the data to be written to slot 'i' where
1963 * we just recently had a write error.
1964 * We repeatedly clone the bio and trim down to one block,
1965 * then try the write. Where the write fails we record
1967 * It is conceivable that the bio doesn't exactly align with
1968 * blocks. We must handle this.
1970 * We currently own a reference to the rdev.
1976 int sect_to_write
= r10_bio
->sectors
;
1979 if (rdev
->badblocks
.shift
< 0)
1982 block_sectors
= 1 << rdev
->badblocks
.shift
;
1983 sector
= r10_bio
->sector
;
1984 sectors
= ((r10_bio
->sector
+ block_sectors
)
1985 & ~(sector_t
)(block_sectors
- 1))
1988 while (sect_to_write
) {
1990 if (sectors
> sect_to_write
)
1991 sectors
= sect_to_write
;
1992 /* Write at 'sector' for 'sectors' */
1993 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1994 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
1995 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1997 (sector
- r10_bio
->sector
));
1998 wbio
->bi_bdev
= rdev
->bdev
;
1999 if (submit_bio_wait(WRITE
, wbio
) == 0)
2001 ok
= rdev_set_badblocks(rdev
, sector
,
2006 sect_to_write
-= sectors
;
2008 sectors
= block_sectors
;
2013 static void handle_read_error(mddev_t
*mddev
, r10bio_t
*r10_bio
)
2015 int slot
= r10_bio
->read_slot
;
2016 int mirror
= r10_bio
->devs
[slot
].devnum
;
2018 conf_t
*conf
= mddev
->private;
2020 char b
[BDEVNAME_SIZE
];
2021 unsigned long do_sync
;
2024 /* we got a read error. Maybe the drive is bad. Maybe just
2025 * the block and we can fix it.
2026 * We freeze all other IO, and try reading the block from
2027 * other devices. When we find one, we re-write
2028 * and check it that fixes the read error.
2029 * This is all done synchronously while the array is
2032 if (mddev
->ro
== 0) {
2034 fix_read_error(conf
, mddev
, r10_bio
);
2035 unfreeze_array(conf
);
2037 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, mddev
);
2039 bio
= r10_bio
->devs
[slot
].bio
;
2040 bdevname(bio
->bi_bdev
, b
);
2041 r10_bio
->devs
[slot
].bio
=
2042 mddev
->ro
? IO_BLOCKED
: NULL
;
2044 mirror
= read_balance(conf
, r10_bio
, &max_sectors
);
2046 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2047 " read error for block %llu\n",
2049 (unsigned long long)r10_bio
->sector
);
2050 raid_end_bio_io(r10_bio
);
2055 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2058 slot
= r10_bio
->read_slot
;
2059 rdev
= conf
->mirrors
[mirror
].rdev
;
2062 "md/raid10:%s: %s: redirecting"
2063 "sector %llu to another mirror\n",
2065 bdevname(rdev
->bdev
, b
),
2066 (unsigned long long)r10_bio
->sector
);
2067 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2070 r10_bio
->sector
- bio
->bi_sector
,
2072 r10_bio
->devs
[slot
].bio
= bio
;
2073 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2074 + rdev
->data_offset
;
2075 bio
->bi_bdev
= rdev
->bdev
;
2076 bio
->bi_rw
= READ
| do_sync
;
2077 bio
->bi_private
= r10_bio
;
2078 bio
->bi_end_io
= raid10_end_read_request
;
2079 if (max_sectors
< r10_bio
->sectors
) {
2080 /* Drat - have to split this up more */
2081 struct bio
*mbio
= r10_bio
->master_bio
;
2082 int sectors_handled
=
2083 r10_bio
->sector
+ max_sectors
2085 r10_bio
->sectors
= max_sectors
;
2086 spin_lock_irq(&conf
->device_lock
);
2087 if (mbio
->bi_phys_segments
== 0)
2088 mbio
->bi_phys_segments
= 2;
2090 mbio
->bi_phys_segments
++;
2091 spin_unlock_irq(&conf
->device_lock
);
2092 generic_make_request(bio
);
2095 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2097 r10_bio
->master_bio
= mbio
;
2098 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2101 set_bit(R10BIO_ReadError
,
2103 r10_bio
->mddev
= mddev
;
2104 r10_bio
->sector
= mbio
->bi_sector
2109 generic_make_request(bio
);
2112 static void handle_write_completed(conf_t
*conf
, r10bio_t
*r10_bio
)
2114 /* Some sort of write request has finished and it
2115 * succeeded in writing where we thought there was a
2116 * bad block. So forget the bad block.
2117 * Or possibly if failed and we need to record
2123 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2124 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2125 for (m
= 0; m
< conf
->copies
; m
++) {
2126 int dev
= r10_bio
->devs
[m
].devnum
;
2127 rdev
= conf
->mirrors
[dev
].rdev
;
2128 if (r10_bio
->devs
[m
].bio
== NULL
)
2130 if (test_bit(BIO_UPTODATE
,
2131 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2132 rdev_clear_badblocks(
2134 r10_bio
->devs
[m
].addr
,
2137 if (!rdev_set_badblocks(
2139 r10_bio
->devs
[m
].addr
,
2140 r10_bio
->sectors
, 0))
2141 md_error(conf
->mddev
, rdev
);
2146 for (m
= 0; m
< conf
->copies
; m
++) {
2147 int dev
= r10_bio
->devs
[m
].devnum
;
2148 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2149 rdev
= conf
->mirrors
[dev
].rdev
;
2150 if (bio
== IO_MADE_GOOD
) {
2151 rdev_clear_badblocks(
2153 r10_bio
->devs
[m
].addr
,
2155 rdev_dec_pending(rdev
, conf
->mddev
);
2156 } else if (bio
!= NULL
&&
2157 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2158 if (!narrow_write_error(r10_bio
, m
)) {
2159 md_error(conf
->mddev
, rdev
);
2160 set_bit(R10BIO_Degraded
,
2163 rdev_dec_pending(rdev
, conf
->mddev
);
2166 if (test_bit(R10BIO_WriteError
,
2168 close_write(r10_bio
);
2169 raid_end_bio_io(r10_bio
);
2173 static void raid10d(mddev_t
*mddev
)
2176 unsigned long flags
;
2177 conf_t
*conf
= mddev
->private;
2178 struct list_head
*head
= &conf
->retry_list
;
2179 struct blk_plug plug
;
2181 md_check_recovery(mddev
);
2183 blk_start_plug(&plug
);
2186 flush_pending_writes(conf
);
2188 spin_lock_irqsave(&conf
->device_lock
, flags
);
2189 if (list_empty(head
)) {
2190 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2193 r10_bio
= list_entry(head
->prev
, r10bio_t
, retry_list
);
2194 list_del(head
->prev
);
2196 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2198 mddev
= r10_bio
->mddev
;
2199 conf
= mddev
->private;
2200 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2201 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2202 handle_write_completed(conf
, r10_bio
);
2203 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2204 sync_request_write(mddev
, r10_bio
);
2205 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2206 recovery_request_write(mddev
, r10_bio
);
2207 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2208 handle_read_error(mddev
, r10_bio
);
2210 /* just a partial read to be scheduled from a
2213 int slot
= r10_bio
->read_slot
;
2214 generic_make_request(r10_bio
->devs
[slot
].bio
);
2218 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2219 md_check_recovery(mddev
);
2221 blk_finish_plug(&plug
);
2225 static int init_resync(conf_t
*conf
)
2229 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2230 BUG_ON(conf
->r10buf_pool
);
2231 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2232 if (!conf
->r10buf_pool
)
2234 conf
->next_resync
= 0;
2239 * perform a "sync" on one "block"
2241 * We need to make sure that no normal I/O request - particularly write
2242 * requests - conflict with active sync requests.
2244 * This is achieved by tracking pending requests and a 'barrier' concept
2245 * that can be installed to exclude normal IO requests.
2247 * Resync and recovery are handled very differently.
2248 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2250 * For resync, we iterate over virtual addresses, read all copies,
2251 * and update if there are differences. If only one copy is live,
2253 * For recovery, we iterate over physical addresses, read a good
2254 * value for each non-in_sync drive, and over-write.
2256 * So, for recovery we may have several outstanding complex requests for a
2257 * given address, one for each out-of-sync device. We model this by allocating
2258 * a number of r10_bio structures, one for each out-of-sync device.
2259 * As we setup these structures, we collect all bio's together into a list
2260 * which we then process collectively to add pages, and then process again
2261 * to pass to generic_make_request.
2263 * The r10_bio structures are linked using a borrowed master_bio pointer.
2264 * This link is counted in ->remaining. When the r10_bio that points to NULL
2265 * has its remaining count decremented to 0, the whole complex operation
2270 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
,
2271 int *skipped
, int go_faster
)
2273 conf_t
*conf
= mddev
->private;
2275 struct bio
*biolist
= NULL
, *bio
;
2276 sector_t max_sector
, nr_sectors
;
2279 sector_t sync_blocks
;
2280 sector_t sectors_skipped
= 0;
2281 int chunks_skipped
= 0;
2283 if (!conf
->r10buf_pool
)
2284 if (init_resync(conf
))
2288 max_sector
= mddev
->dev_sectors
;
2289 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2290 max_sector
= mddev
->resync_max_sectors
;
2291 if (sector_nr
>= max_sector
) {
2292 /* If we aborted, we need to abort the
2293 * sync on the 'current' bitmap chucks (there can
2294 * be several when recovering multiple devices).
2295 * as we may have started syncing it but not finished.
2296 * We can find the current address in
2297 * mddev->curr_resync, but for recovery,
2298 * we need to convert that to several
2299 * virtual addresses.
2301 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2302 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2303 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2305 else for (i
=0; i
<conf
->raid_disks
; i
++) {
2307 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2308 bitmap_end_sync(mddev
->bitmap
, sect
,
2311 } else /* completed sync */
2314 bitmap_close_sync(mddev
->bitmap
);
2317 return sectors_skipped
;
2319 if (chunks_skipped
>= conf
->raid_disks
) {
2320 /* if there has been nothing to do on any drive,
2321 * then there is nothing to do at all..
2324 return (max_sector
- sector_nr
) + sectors_skipped
;
2327 if (max_sector
> mddev
->resync_max
)
2328 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2330 /* make sure whole request will fit in a chunk - if chunks
2333 if (conf
->near_copies
< conf
->raid_disks
&&
2334 max_sector
> (sector_nr
| conf
->chunk_mask
))
2335 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
2337 * If there is non-resync activity waiting for us then
2338 * put in a delay to throttle resync.
2340 if (!go_faster
&& conf
->nr_waiting
)
2341 msleep_interruptible(1000);
2343 /* Again, very different code for resync and recovery.
2344 * Both must result in an r10bio with a list of bios that
2345 * have bi_end_io, bi_sector, bi_bdev set,
2346 * and bi_private set to the r10bio.
2347 * For recovery, we may actually create several r10bios
2348 * with 2 bios in each, that correspond to the bios in the main one.
2349 * In this case, the subordinate r10bios link back through a
2350 * borrowed master_bio pointer, and the counter in the master
2351 * includes a ref from each subordinate.
2353 /* First, we decide what to do and set ->bi_end_io
2354 * To end_sync_read if we want to read, and
2355 * end_sync_write if we will want to write.
2358 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2359 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2360 /* recovery... the complicated one */
2364 for (i
=0 ; i
<conf
->raid_disks
; i
++) {
2371 if (conf
->mirrors
[i
].rdev
== NULL
||
2372 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
))
2376 /* want to reconstruct this device */
2378 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2379 /* Unless we are doing a full sync, we only need
2380 * to recover the block if it is set in the bitmap
2382 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2384 if (sync_blocks
< max_sync
)
2385 max_sync
= sync_blocks
;
2388 /* yep, skip the sync_blocks here, but don't assume
2389 * that there will never be anything to do here
2391 chunks_skipped
= -1;
2395 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2396 raise_barrier(conf
, rb2
!= NULL
);
2397 atomic_set(&r10_bio
->remaining
, 0);
2399 r10_bio
->master_bio
= (struct bio
*)rb2
;
2401 atomic_inc(&rb2
->remaining
);
2402 r10_bio
->mddev
= mddev
;
2403 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
2404 r10_bio
->sector
= sect
;
2406 raid10_find_phys(conf
, r10_bio
);
2408 /* Need to check if the array will still be
2411 for (j
=0; j
<conf
->raid_disks
; j
++)
2412 if (conf
->mirrors
[j
].rdev
== NULL
||
2413 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
2418 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2419 &sync_blocks
, still_degraded
);
2422 for (j
=0; j
<conf
->copies
;j
++) {
2424 int d
= r10_bio
->devs
[j
].devnum
;
2425 sector_t from_addr
, to_addr
;
2427 sector_t sector
, first_bad
;
2429 if (!conf
->mirrors
[d
].rdev
||
2430 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
2432 /* This is where we read from */
2434 rdev
= conf
->mirrors
[d
].rdev
;
2435 sector
= r10_bio
->devs
[j
].addr
;
2437 if (is_badblock(rdev
, sector
, max_sync
,
2438 &first_bad
, &bad_sectors
)) {
2439 if (first_bad
> sector
)
2440 max_sync
= first_bad
- sector
;
2442 bad_sectors
-= (sector
2444 if (max_sync
> bad_sectors
)
2445 max_sync
= bad_sectors
;
2449 bio
= r10_bio
->devs
[0].bio
;
2450 bio
->bi_next
= biolist
;
2452 bio
->bi_private
= r10_bio
;
2453 bio
->bi_end_io
= end_sync_read
;
2455 from_addr
= r10_bio
->devs
[j
].addr
;
2456 bio
->bi_sector
= from_addr
+
2457 conf
->mirrors
[d
].rdev
->data_offset
;
2458 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2459 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2460 atomic_inc(&r10_bio
->remaining
);
2461 /* and we write to 'i' */
2463 for (k
=0; k
<conf
->copies
; k
++)
2464 if (r10_bio
->devs
[k
].devnum
== i
)
2466 BUG_ON(k
== conf
->copies
);
2467 bio
= r10_bio
->devs
[1].bio
;
2468 bio
->bi_next
= biolist
;
2470 bio
->bi_private
= r10_bio
;
2471 bio
->bi_end_io
= end_sync_write
;
2473 to_addr
= r10_bio
->devs
[k
].addr
;
2474 bio
->bi_sector
= to_addr
+
2475 conf
->mirrors
[i
].rdev
->data_offset
;
2476 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
2478 r10_bio
->devs
[0].devnum
= d
;
2479 r10_bio
->devs
[0].addr
= from_addr
;
2480 r10_bio
->devs
[1].devnum
= i
;
2481 r10_bio
->devs
[1].addr
= to_addr
;
2485 if (j
== conf
->copies
) {
2486 /* Cannot recover, so abort the recovery or
2487 * record a bad block */
2490 atomic_dec(&rb2
->remaining
);
2493 /* problem is that there are bad blocks
2494 * on other device(s)
2497 for (k
= 0; k
< conf
->copies
; k
++)
2498 if (r10_bio
->devs
[k
].devnum
== i
)
2500 if (!rdev_set_badblocks(
2501 conf
->mirrors
[i
].rdev
,
2502 r10_bio
->devs
[k
].addr
,
2507 if (!test_and_set_bit(MD_RECOVERY_INTR
,
2509 printk(KERN_INFO
"md/raid10:%s: insufficient "
2510 "working devices for recovery.\n",
2512 conf
->mirrors
[i
].recovery_disabled
2513 = mddev
->recovery_disabled
;
2518 if (biolist
== NULL
) {
2520 r10bio_t
*rb2
= r10_bio
;
2521 r10_bio
= (r10bio_t
*) rb2
->master_bio
;
2522 rb2
->master_bio
= NULL
;
2528 /* resync. Schedule a read for every block at this virt offset */
2531 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
2533 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2534 &sync_blocks
, mddev
->degraded
) &&
2535 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
2536 &mddev
->recovery
)) {
2537 /* We can skip this block */
2539 return sync_blocks
+ sectors_skipped
;
2541 if (sync_blocks
< max_sync
)
2542 max_sync
= sync_blocks
;
2543 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2545 r10_bio
->mddev
= mddev
;
2546 atomic_set(&r10_bio
->remaining
, 0);
2547 raise_barrier(conf
, 0);
2548 conf
->next_resync
= sector_nr
;
2550 r10_bio
->master_bio
= NULL
;
2551 r10_bio
->sector
= sector_nr
;
2552 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
2553 raid10_find_phys(conf
, r10_bio
);
2554 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
2556 for (i
=0; i
<conf
->copies
; i
++) {
2557 int d
= r10_bio
->devs
[i
].devnum
;
2558 sector_t first_bad
, sector
;
2561 bio
= r10_bio
->devs
[i
].bio
;
2562 bio
->bi_end_io
= NULL
;
2563 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2564 if (conf
->mirrors
[d
].rdev
== NULL
||
2565 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
2567 sector
= r10_bio
->devs
[i
].addr
;
2568 if (is_badblock(conf
->mirrors
[d
].rdev
,
2570 &first_bad
, &bad_sectors
)) {
2571 if (first_bad
> sector
)
2572 max_sync
= first_bad
- sector
;
2574 bad_sectors
-= (sector
- first_bad
);
2575 if (max_sync
> bad_sectors
)
2576 max_sync
= max_sync
;
2580 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2581 atomic_inc(&r10_bio
->remaining
);
2582 bio
->bi_next
= biolist
;
2584 bio
->bi_private
= r10_bio
;
2585 bio
->bi_end_io
= end_sync_read
;
2587 bio
->bi_sector
= sector
+
2588 conf
->mirrors
[d
].rdev
->data_offset
;
2589 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2594 for (i
=0; i
<conf
->copies
; i
++) {
2595 int d
= r10_bio
->devs
[i
].devnum
;
2596 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
2597 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
2606 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2608 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2610 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
2613 bio
->bi_phys_segments
= 0;
2618 if (sector_nr
+ max_sync
< max_sector
)
2619 max_sector
= sector_nr
+ max_sync
;
2622 int len
= PAGE_SIZE
;
2623 if (sector_nr
+ (len
>>9) > max_sector
)
2624 len
= (max_sector
- sector_nr
) << 9;
2627 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2629 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2630 if (bio_add_page(bio
, page
, len
, 0))
2634 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2635 for (bio2
= biolist
;
2636 bio2
&& bio2
!= bio
;
2637 bio2
= bio2
->bi_next
) {
2638 /* remove last page from this bio */
2640 bio2
->bi_size
-= len
;
2641 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
2645 nr_sectors
+= len
>>9;
2646 sector_nr
+= len
>>9;
2647 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
2649 r10_bio
->sectors
= nr_sectors
;
2653 biolist
= biolist
->bi_next
;
2655 bio
->bi_next
= NULL
;
2656 r10_bio
= bio
->bi_private
;
2657 r10_bio
->sectors
= nr_sectors
;
2659 if (bio
->bi_end_io
== end_sync_read
) {
2660 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2661 generic_make_request(bio
);
2665 if (sectors_skipped
)
2666 /* pretend they weren't skipped, it makes
2667 * no important difference in this case
2669 md_done_sync(mddev
, sectors_skipped
, 1);
2671 return sectors_skipped
+ nr_sectors
;
2673 /* There is nowhere to write, so all non-sync
2674 * drives must be failed or in resync, all drives
2675 * have a bad block, so try the next chunk...
2677 if (sector_nr
+ max_sync
< max_sector
)
2678 max_sector
= sector_nr
+ max_sync
;
2680 sectors_skipped
+= (max_sector
- sector_nr
);
2682 sector_nr
= max_sector
;
2687 raid10_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
2690 conf_t
*conf
= mddev
->private;
2693 raid_disks
= conf
->raid_disks
;
2695 sectors
= conf
->dev_sectors
;
2697 size
= sectors
>> conf
->chunk_shift
;
2698 sector_div(size
, conf
->far_copies
);
2699 size
= size
* raid_disks
;
2700 sector_div(size
, conf
->near_copies
);
2702 return size
<< conf
->chunk_shift
;
2706 static conf_t
*setup_conf(mddev_t
*mddev
)
2708 conf_t
*conf
= NULL
;
2710 sector_t stride
, size
;
2713 if (mddev
->new_chunk_sectors
< (PAGE_SIZE
>> 9) ||
2714 !is_power_of_2(mddev
->new_chunk_sectors
)) {
2715 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
2716 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2717 mdname(mddev
), PAGE_SIZE
);
2721 nc
= mddev
->new_layout
& 255;
2722 fc
= (mddev
->new_layout
>> 8) & 255;
2723 fo
= mddev
->new_layout
& (1<<16);
2725 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
2726 (mddev
->new_layout
>> 17)) {
2727 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2728 mdname(mddev
), mddev
->new_layout
);
2733 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
2737 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
2742 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2747 conf
->raid_disks
= mddev
->raid_disks
;
2748 conf
->near_copies
= nc
;
2749 conf
->far_copies
= fc
;
2750 conf
->copies
= nc
*fc
;
2751 conf
->far_offset
= fo
;
2752 conf
->chunk_mask
= mddev
->new_chunk_sectors
- 1;
2753 conf
->chunk_shift
= ffz(~mddev
->new_chunk_sectors
);
2755 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
2756 r10bio_pool_free
, conf
);
2757 if (!conf
->r10bio_pool
)
2760 size
= mddev
->dev_sectors
>> conf
->chunk_shift
;
2761 sector_div(size
, fc
);
2762 size
= size
* conf
->raid_disks
;
2763 sector_div(size
, nc
);
2764 /* 'size' is now the number of chunks in the array */
2765 /* calculate "used chunks per device" in 'stride' */
2766 stride
= size
* conf
->copies
;
2768 /* We need to round up when dividing by raid_disks to
2769 * get the stride size.
2771 stride
+= conf
->raid_disks
- 1;
2772 sector_div(stride
, conf
->raid_disks
);
2774 conf
->dev_sectors
= stride
<< conf
->chunk_shift
;
2779 sector_div(stride
, fc
);
2780 conf
->stride
= stride
<< conf
->chunk_shift
;
2783 spin_lock_init(&conf
->device_lock
);
2784 INIT_LIST_HEAD(&conf
->retry_list
);
2786 spin_lock_init(&conf
->resync_lock
);
2787 init_waitqueue_head(&conf
->wait_barrier
);
2789 conf
->thread
= md_register_thread(raid10d
, mddev
, NULL
);
2793 conf
->mddev
= mddev
;
2797 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
2800 if (conf
->r10bio_pool
)
2801 mempool_destroy(conf
->r10bio_pool
);
2802 kfree(conf
->mirrors
);
2803 safe_put_page(conf
->tmppage
);
2806 return ERR_PTR(err
);
2809 static int run(mddev_t
*mddev
)
2812 int i
, disk_idx
, chunk_size
;
2813 mirror_info_t
*disk
;
2818 * copy the already verified devices into our private RAID10
2819 * bookkeeping area. [whatever we allocate in run(),
2820 * should be freed in stop()]
2823 if (mddev
->private == NULL
) {
2824 conf
= setup_conf(mddev
);
2826 return PTR_ERR(conf
);
2827 mddev
->private = conf
;
2829 conf
= mddev
->private;
2833 mddev
->thread
= conf
->thread
;
2834 conf
->thread
= NULL
;
2836 chunk_size
= mddev
->chunk_sectors
<< 9;
2837 blk_queue_io_min(mddev
->queue
, chunk_size
);
2838 if (conf
->raid_disks
% conf
->near_copies
)
2839 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->raid_disks
);
2841 blk_queue_io_opt(mddev
->queue
, chunk_size
*
2842 (conf
->raid_disks
/ conf
->near_copies
));
2844 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
2846 disk_idx
= rdev
->raid_disk
;
2847 if (disk_idx
>= conf
->raid_disks
2850 disk
= conf
->mirrors
+ disk_idx
;
2853 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2854 rdev
->data_offset
<< 9);
2855 /* as we don't honour merge_bvec_fn, we must never risk
2856 * violating it, so limit max_segments to 1 lying
2857 * within a single page.
2859 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
2860 blk_queue_max_segments(mddev
->queue
, 1);
2861 blk_queue_segment_boundary(mddev
->queue
,
2862 PAGE_CACHE_SIZE
- 1);
2865 disk
->head_position
= 0;
2867 /* need to check that every block has at least one working mirror */
2868 if (!enough(conf
, -1)) {
2869 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
2874 mddev
->degraded
= 0;
2875 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2877 disk
= conf
->mirrors
+ i
;
2880 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2881 disk
->head_position
= 0;
2888 if (mddev
->recovery_cp
!= MaxSector
)
2889 printk(KERN_NOTICE
"md/raid10:%s: not clean"
2890 " -- starting background reconstruction\n",
2893 "md/raid10:%s: active with %d out of %d devices\n",
2894 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
,
2897 * Ok, everything is just fine now
2899 mddev
->dev_sectors
= conf
->dev_sectors
;
2900 size
= raid10_size(mddev
, 0, 0);
2901 md_set_array_sectors(mddev
, size
);
2902 mddev
->resync_max_sectors
= size
;
2904 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
2905 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2907 /* Calculate max read-ahead size.
2908 * We need to readahead at least twice a whole stripe....
2912 int stripe
= conf
->raid_disks
*
2913 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
2914 stripe
/= conf
->near_copies
;
2915 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
2916 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
2919 if (conf
->near_copies
< conf
->raid_disks
)
2920 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
2922 if (md_integrity_register(mddev
))
2928 md_unregister_thread(mddev
->thread
);
2929 if (conf
->r10bio_pool
)
2930 mempool_destroy(conf
->r10bio_pool
);
2931 safe_put_page(conf
->tmppage
);
2932 kfree(conf
->mirrors
);
2934 mddev
->private = NULL
;
2939 static int stop(mddev_t
*mddev
)
2941 conf_t
*conf
= mddev
->private;
2943 raise_barrier(conf
, 0);
2944 lower_barrier(conf
);
2946 md_unregister_thread(mddev
->thread
);
2947 mddev
->thread
= NULL
;
2948 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
2949 if (conf
->r10bio_pool
)
2950 mempool_destroy(conf
->r10bio_pool
);
2951 kfree(conf
->mirrors
);
2953 mddev
->private = NULL
;
2957 static void raid10_quiesce(mddev_t
*mddev
, int state
)
2959 conf_t
*conf
= mddev
->private;
2963 raise_barrier(conf
, 0);
2966 lower_barrier(conf
);
2971 static void *raid10_takeover_raid0(mddev_t
*mddev
)
2976 if (mddev
->degraded
> 0) {
2977 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
2979 return ERR_PTR(-EINVAL
);
2982 /* Set new parameters */
2983 mddev
->new_level
= 10;
2984 /* new layout: far_copies = 1, near_copies = 2 */
2985 mddev
->new_layout
= (1<<8) + 2;
2986 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
2987 mddev
->delta_disks
= mddev
->raid_disks
;
2988 mddev
->raid_disks
*= 2;
2989 /* make sure it will be not marked as dirty */
2990 mddev
->recovery_cp
= MaxSector
;
2992 conf
= setup_conf(mddev
);
2993 if (!IS_ERR(conf
)) {
2994 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
2995 if (rdev
->raid_disk
>= 0)
2996 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3003 static void *raid10_takeover(mddev_t
*mddev
)
3005 struct raid0_private_data
*raid0_priv
;
3007 /* raid10 can take over:
3008 * raid0 - providing it has only two drives
3010 if (mddev
->level
== 0) {
3011 /* for raid0 takeover only one zone is supported */
3012 raid0_priv
= mddev
->private;
3013 if (raid0_priv
->nr_strip_zones
> 1) {
3014 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3015 " with more than one zone.\n",
3017 return ERR_PTR(-EINVAL
);
3019 return raid10_takeover_raid0(mddev
);
3021 return ERR_PTR(-EINVAL
);
3024 static struct mdk_personality raid10_personality
=
3028 .owner
= THIS_MODULE
,
3029 .make_request
= make_request
,
3033 .error_handler
= error
,
3034 .hot_add_disk
= raid10_add_disk
,
3035 .hot_remove_disk
= raid10_remove_disk
,
3036 .spare_active
= raid10_spare_active
,
3037 .sync_request
= sync_request
,
3038 .quiesce
= raid10_quiesce
,
3039 .size
= raid10_size
,
3040 .takeover
= raid10_takeover
,
3043 static int __init
raid_init(void)
3045 return register_md_personality(&raid10_personality
);
3048 static void raid_exit(void)
3050 unregister_md_personality(&raid10_personality
);
3053 module_init(raid_init
);
3054 module_exit(raid_exit
);
3055 MODULE_LICENSE("GPL");
3056 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3057 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3058 MODULE_ALIAS("md-raid10");
3059 MODULE_ALIAS("md-level-10");