2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for futher copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/seq_file.h>
30 * RAID10 provides a combination of RAID0 and RAID1 functionality.
31 * The layout of data is defined by
34 * near_copies (stored in low byte of layout)
35 * far_copies (stored in second byte of layout)
36 * far_offset (stored in bit 16 of layout )
38 * The data to be stored is divided into chunks using chunksize.
39 * Each device is divided into far_copies sections.
40 * In each section, chunks are laid out in a style similar to raid0, but
41 * near_copies copies of each chunk is stored (each on a different drive).
42 * The starting device for each section is offset near_copies from the starting
43 * device of the previous section.
44 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
46 * near_copies and far_copies must be at least one, and their product is at most
49 * If far_offset is true, then the far_copies are handled a bit differently.
50 * The copies are still in different stripes, but instead of be very far apart
51 * on disk, there are adjacent stripes.
55 * Number of guaranteed r10bios in case of extreme VM load:
57 #define NR_RAID10_BIOS 256
59 static void unplug_slaves(mddev_t
*mddev
);
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
)
68 int size
= offsetof(struct r10bio_s
, devs
[conf
->copies
]);
70 /* allocate a r10bio with room for raid_disks entries in the bios array */
71 r10_bio
= kzalloc(size
, gfp_flags
);
72 if (!r10_bio
&& conf
->mddev
)
73 unplug_slaves(conf
->mddev
);
78 static void r10bio_pool_free(void *r10_bio
, void *data
)
83 /* Maximum size of each resync request */
84 #define RESYNC_BLOCK_SIZE (64*1024)
85 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
86 /* amount of memory to reserve for resync requests */
87 #define RESYNC_WINDOW (1024*1024)
88 /* maximum number of concurrent requests, memory permitting */
89 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
92 * When performing a resync, we need to read and compare, so
93 * we need as many pages are there are copies.
94 * When performing a recovery, we need 2 bios, one for read,
95 * one for write (we recover only one drive per r10buf)
98 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
107 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
109 unplug_slaves(conf
->mddev
);
113 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
114 nalloc
= conf
->copies
; /* resync */
116 nalloc
= 2; /* recovery */
121 for (j
= nalloc
; j
-- ; ) {
122 bio
= bio_alloc(gfp_flags
, RESYNC_PAGES
);
125 r10_bio
->devs
[j
].bio
= bio
;
128 * Allocate RESYNC_PAGES data pages and attach them
131 for (j
= 0 ; j
< nalloc
; j
++) {
132 bio
= r10_bio
->devs
[j
].bio
;
133 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
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
);
255 static void raid10_end_read_request(struct bio
*bio
, int error
)
257 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
258 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
260 conf_t
*conf
= r10_bio
->mddev
->private;
263 slot
= r10_bio
->read_slot
;
264 dev
= r10_bio
->devs
[slot
].devnum
;
266 * this branch is our 'one mirror IO has finished' event handler:
268 update_head_pos(slot
, r10_bio
);
272 * Set R10BIO_Uptodate in our master bio, so that
273 * we will return a good error code to the higher
274 * levels even if IO on some other mirrored buffer fails.
276 * The 'master' represents the composite IO operation to
277 * user-side. So if something waits for IO, then it will
278 * wait for the 'master' bio.
280 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
281 raid_end_bio_io(r10_bio
);
286 char b
[BDEVNAME_SIZE
];
287 if (printk_ratelimit())
288 printk(KERN_ERR
"raid10: %s: rescheduling sector %llu\n",
289 bdevname(conf
->mirrors
[dev
].rdev
->bdev
,b
), (unsigned long long)r10_bio
->sector
);
290 reschedule_retry(r10_bio
);
293 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
296 static void raid10_end_write_request(struct bio
*bio
, int error
)
298 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
299 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
301 conf_t
*conf
= r10_bio
->mddev
->private;
303 for (slot
= 0; slot
< conf
->copies
; slot
++)
304 if (r10_bio
->devs
[slot
].bio
== bio
)
306 dev
= r10_bio
->devs
[slot
].devnum
;
309 * this branch is our 'one mirror IO has finished' event handler:
312 md_error(r10_bio
->mddev
, conf
->mirrors
[dev
].rdev
);
313 /* an I/O failed, we can't clear the bitmap */
314 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
317 * Set R10BIO_Uptodate in our master bio, so that
318 * we will return a good error code for to the higher
319 * levels even if IO on some other mirrored buffer fails.
321 * The 'master' represents the composite IO operation to
322 * user-side. So if something waits for IO, then it will
323 * wait for the 'master' bio.
325 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
327 update_head_pos(slot
, r10_bio
);
331 * Let's see if all mirrored write operations have finished
334 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
335 /* clear the bitmap if all writes complete successfully */
336 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
338 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
340 md_write_end(r10_bio
->mddev
);
341 raid_end_bio_io(r10_bio
);
344 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
349 * RAID10 layout manager
350 * Aswell as the chunksize and raid_disks count, there are two
351 * parameters: near_copies and far_copies.
352 * near_copies * far_copies must be <= raid_disks.
353 * Normally one of these will be 1.
354 * If both are 1, we get raid0.
355 * If near_copies == raid_disks, we get raid1.
357 * Chunks are layed out in raid0 style with near_copies copies of the
358 * first chunk, followed by near_copies copies of the next chunk and
360 * If far_copies > 1, then after 1/far_copies of the array has been assigned
361 * as described above, we start again with a device offset of near_copies.
362 * So we effectively have another copy of the whole array further down all
363 * the drives, but with blocks on different drives.
364 * With this layout, and block is never stored twice on the one device.
366 * raid10_find_phys finds the sector offset of a given virtual sector
367 * on each device that it is on.
369 * raid10_find_virt does the reverse mapping, from a device and a
370 * sector offset to a virtual address
373 static void raid10_find_phys(conf_t
*conf
, r10bio_t
*r10bio
)
383 /* now calculate first sector/dev */
384 chunk
= r10bio
->sector
>> conf
->chunk_shift
;
385 sector
= r10bio
->sector
& conf
->chunk_mask
;
387 chunk
*= conf
->near_copies
;
389 dev
= sector_div(stripe
, conf
->raid_disks
);
390 if (conf
->far_offset
)
391 stripe
*= conf
->far_copies
;
393 sector
+= stripe
<< conf
->chunk_shift
;
395 /* and calculate all the others */
396 for (n
=0; n
< conf
->near_copies
; n
++) {
399 r10bio
->devs
[slot
].addr
= sector
;
400 r10bio
->devs
[slot
].devnum
= d
;
403 for (f
= 1; f
< conf
->far_copies
; f
++) {
404 d
+= conf
->near_copies
;
405 if (d
>= conf
->raid_disks
)
406 d
-= conf
->raid_disks
;
408 r10bio
->devs
[slot
].devnum
= d
;
409 r10bio
->devs
[slot
].addr
= s
;
413 if (dev
>= conf
->raid_disks
) {
415 sector
+= (conf
->chunk_mask
+ 1);
418 BUG_ON(slot
!= conf
->copies
);
421 static sector_t
raid10_find_virt(conf_t
*conf
, sector_t sector
, int dev
)
423 sector_t offset
, chunk
, vchunk
;
425 offset
= sector
& conf
->chunk_mask
;
426 if (conf
->far_offset
) {
428 chunk
= sector
>> conf
->chunk_shift
;
429 fc
= sector_div(chunk
, conf
->far_copies
);
430 dev
-= fc
* conf
->near_copies
;
432 dev
+= conf
->raid_disks
;
434 while (sector
>= conf
->stride
) {
435 sector
-= conf
->stride
;
436 if (dev
< conf
->near_copies
)
437 dev
+= conf
->raid_disks
- conf
->near_copies
;
439 dev
-= conf
->near_copies
;
441 chunk
= sector
>> conf
->chunk_shift
;
443 vchunk
= chunk
* conf
->raid_disks
+ dev
;
444 sector_div(vchunk
, conf
->near_copies
);
445 return (vchunk
<< conf
->chunk_shift
) + offset
;
449 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
451 * @bvm: properties of new bio
452 * @biovec: the request that could be merged to it.
454 * Return amount of bytes we can accept at this offset
455 * If near_copies == raid_disk, there are no striping issues,
456 * but in that case, the function isn't called at all.
458 static int raid10_mergeable_bvec(struct request_queue
*q
,
459 struct bvec_merge_data
*bvm
,
460 struct bio_vec
*biovec
)
462 mddev_t
*mddev
= q
->queuedata
;
463 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
465 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
466 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
468 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
469 if (max
< 0) max
= 0; /* bio_add cannot handle a negative return */
470 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
471 return biovec
->bv_len
;
477 * This routine returns the disk from which the requested read should
478 * be done. There is a per-array 'next expected sequential IO' sector
479 * number - if this matches on the next IO then we use the last disk.
480 * There is also a per-disk 'last know head position' sector that is
481 * maintained from IRQ contexts, both the normal and the resync IO
482 * completion handlers update this position correctly. If there is no
483 * perfect sequential match then we pick the disk whose head is closest.
485 * If there are 2 mirrors in the same 2 devices, performance degrades
486 * because position is mirror, not device based.
488 * The rdev for the device selected will have nr_pending incremented.
492 * FIXME: possibly should rethink readbalancing and do it differently
493 * depending on near_copies / far_copies geometry.
495 static int read_balance(conf_t
*conf
, r10bio_t
*r10_bio
)
497 const sector_t this_sector
= r10_bio
->sector
;
498 int disk
, slot
, nslot
;
499 const int sectors
= r10_bio
->sectors
;
500 sector_t new_distance
, current_distance
;
503 raid10_find_phys(conf
, r10_bio
);
506 * Check if we can balance. We can balance on the whole
507 * device if no resync is going on (recovery is ok), or below
508 * the resync window. We take the first readable disk when
509 * above the resync window.
511 if (conf
->mddev
->recovery_cp
< MaxSector
512 && (this_sector
+ sectors
>= conf
->next_resync
)) {
513 /* make sure that disk is operational */
515 disk
= r10_bio
->devs
[slot
].devnum
;
517 while ((rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
518 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
519 !test_bit(In_sync
, &rdev
->flags
)) {
521 if (slot
== conf
->copies
) {
526 disk
= r10_bio
->devs
[slot
].devnum
;
532 /* make sure the disk is operational */
534 disk
= r10_bio
->devs
[slot
].devnum
;
535 while ((rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
536 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
537 !test_bit(In_sync
, &rdev
->flags
)) {
539 if (slot
== conf
->copies
) {
543 disk
= r10_bio
->devs
[slot
].devnum
;
547 current_distance
= abs(r10_bio
->devs
[slot
].addr
-
548 conf
->mirrors
[disk
].head_position
);
550 /* Find the disk whose head is closest,
551 * or - for far > 1 - find the closest to partition beginning */
553 for (nslot
= slot
; nslot
< conf
->copies
; nslot
++) {
554 int ndisk
= r10_bio
->devs
[nslot
].devnum
;
557 if ((rdev
=rcu_dereference(conf
->mirrors
[ndisk
].rdev
)) == NULL
||
558 r10_bio
->devs
[nslot
].bio
== IO_BLOCKED
||
559 !test_bit(In_sync
, &rdev
->flags
))
562 /* This optimisation is debatable, and completely destroys
563 * sequential read speed for 'far copies' arrays. So only
564 * keep it for 'near' arrays, and review those later.
566 if (conf
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
)) {
572 /* for far > 1 always use the lowest address */
573 if (conf
->far_copies
> 1)
574 new_distance
= r10_bio
->devs
[nslot
].addr
;
576 new_distance
= abs(r10_bio
->devs
[nslot
].addr
-
577 conf
->mirrors
[ndisk
].head_position
);
578 if (new_distance
< current_distance
) {
579 current_distance
= new_distance
;
586 r10_bio
->read_slot
= slot
;
587 /* conf->next_seq_sect = this_sector + sectors;*/
589 if (disk
>= 0 && (rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
))!= NULL
)
590 atomic_inc(&conf
->mirrors
[disk
].rdev
->nr_pending
);
598 static void unplug_slaves(mddev_t
*mddev
)
600 conf_t
*conf
= mddev
->private;
604 for (i
=0; i
<mddev
->raid_disks
; i
++) {
605 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
606 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
607 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
609 atomic_inc(&rdev
->nr_pending
);
614 rdev_dec_pending(rdev
, mddev
);
621 static void raid10_unplug(struct request_queue
*q
)
623 mddev_t
*mddev
= q
->queuedata
;
625 unplug_slaves(q
->queuedata
);
626 md_wakeup_thread(mddev
->thread
);
629 static int raid10_congested(void *data
, int bits
)
631 mddev_t
*mddev
= data
;
632 conf_t
*conf
= mddev
->private;
635 if (mddev_congested(mddev
, bits
))
638 for (i
= 0; i
< mddev
->raid_disks
&& ret
== 0; i
++) {
639 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
640 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
641 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
643 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
650 static int flush_pending_writes(conf_t
*conf
)
652 /* Any writes that have been queued but are awaiting
653 * bitmap updates get flushed here.
654 * We return 1 if any requests were actually submitted.
658 spin_lock_irq(&conf
->device_lock
);
660 if (conf
->pending_bio_list
.head
) {
662 bio
= bio_list_get(&conf
->pending_bio_list
);
663 blk_remove_plug(conf
->mddev
->queue
);
664 spin_unlock_irq(&conf
->device_lock
);
665 /* flush any pending bitmap writes to disk
666 * before proceeding w/ I/O */
667 bitmap_unplug(conf
->mddev
->bitmap
);
669 while (bio
) { /* submit pending writes */
670 struct bio
*next
= bio
->bi_next
;
672 generic_make_request(bio
);
677 spin_unlock_irq(&conf
->device_lock
);
681 * Sometimes we need to suspend IO while we do something else,
682 * either some resync/recovery, or reconfigure the array.
683 * To do this we raise a 'barrier'.
684 * The 'barrier' is a counter that can be raised multiple times
685 * to count how many activities are happening which preclude
687 * We can only raise the barrier if there is no pending IO.
688 * i.e. if nr_pending == 0.
689 * We choose only to raise the barrier if no-one is waiting for the
690 * barrier to go down. This means that as soon as an IO request
691 * is ready, no other operations which require a barrier will start
692 * until the IO request has had a chance.
694 * So: regular IO calls 'wait_barrier'. When that returns there
695 * is no backgroup IO happening, It must arrange to call
696 * allow_barrier when it has finished its IO.
697 * backgroup IO calls must call raise_barrier. Once that returns
698 * there is no normal IO happeing. It must arrange to call
699 * lower_barrier when the particular background IO completes.
702 static void raise_barrier(conf_t
*conf
, int force
)
704 BUG_ON(force
&& !conf
->barrier
);
705 spin_lock_irq(&conf
->resync_lock
);
707 /* Wait until no block IO is waiting (unless 'force') */
708 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
710 raid10_unplug(conf
->mddev
->queue
));
712 /* block any new IO from starting */
715 /* No wait for all pending IO to complete */
716 wait_event_lock_irq(conf
->wait_barrier
,
717 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
719 raid10_unplug(conf
->mddev
->queue
));
721 spin_unlock_irq(&conf
->resync_lock
);
724 static void lower_barrier(conf_t
*conf
)
727 spin_lock_irqsave(&conf
->resync_lock
, flags
);
729 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
730 wake_up(&conf
->wait_barrier
);
733 static void wait_barrier(conf_t
*conf
)
735 spin_lock_irq(&conf
->resync_lock
);
738 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
740 raid10_unplug(conf
->mddev
->queue
));
744 spin_unlock_irq(&conf
->resync_lock
);
747 static void allow_barrier(conf_t
*conf
)
750 spin_lock_irqsave(&conf
->resync_lock
, flags
);
752 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
753 wake_up(&conf
->wait_barrier
);
756 static void freeze_array(conf_t
*conf
)
758 /* stop syncio and normal IO and wait for everything to
760 * We increment barrier and nr_waiting, and then
761 * wait until nr_pending match nr_queued+1
762 * This is called in the context of one normal IO request
763 * that has failed. Thus any sync request that might be pending
764 * will be blocked by nr_pending, and we need to wait for
765 * pending IO requests to complete or be queued for re-try.
766 * Thus the number queued (nr_queued) plus this request (1)
767 * must match the number of pending IOs (nr_pending) before
770 spin_lock_irq(&conf
->resync_lock
);
773 wait_event_lock_irq(conf
->wait_barrier
,
774 conf
->nr_pending
== conf
->nr_queued
+1,
776 ({ flush_pending_writes(conf
);
777 raid10_unplug(conf
->mddev
->queue
); }));
778 spin_unlock_irq(&conf
->resync_lock
);
781 static void unfreeze_array(conf_t
*conf
)
783 /* reverse the effect of the freeze */
784 spin_lock_irq(&conf
->resync_lock
);
787 wake_up(&conf
->wait_barrier
);
788 spin_unlock_irq(&conf
->resync_lock
);
791 static int make_request(struct request_queue
*q
, struct bio
* bio
)
793 mddev_t
*mddev
= q
->queuedata
;
794 conf_t
*conf
= mddev
->private;
795 mirror_info_t
*mirror
;
797 struct bio
*read_bio
;
800 int chunk_sects
= conf
->chunk_mask
+ 1;
801 const int rw
= bio_data_dir(bio
);
802 const bool do_sync
= bio_rw_flagged(bio
, BIO_RW_SYNCIO
);
805 mdk_rdev_t
*blocked_rdev
;
807 if (unlikely(bio_rw_flagged(bio
, BIO_RW_BARRIER
))) {
808 md_barrier_request(mddev
, bio
);
812 /* If this request crosses a chunk boundary, we need to
813 * split it. This will only happen for 1 PAGE (or less) requests.
815 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
817 conf
->near_copies
< conf
->raid_disks
)) {
819 /* Sanity check -- queue functions should prevent this happening */
820 if (bio
->bi_vcnt
!= 1 ||
823 /* This is a one page bio that upper layers
824 * refuse to split for us, so we need to split it.
827 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
829 /* Each of these 'make_request' calls will call 'wait_barrier'.
830 * If the first succeeds but the second blocks due to the resync
831 * thread raising the barrier, we will deadlock because the
832 * IO to the underlying device will be queued in generic_make_request
833 * and will never complete, so will never reduce nr_pending.
834 * So increment nr_waiting here so no new raise_barriers will
835 * succeed, and so the second wait_barrier cannot block.
837 spin_lock_irq(&conf
->resync_lock
);
839 spin_unlock_irq(&conf
->resync_lock
);
841 if (make_request(q
, &bp
->bio1
))
842 generic_make_request(&bp
->bio1
);
843 if (make_request(q
, &bp
->bio2
))
844 generic_make_request(&bp
->bio2
);
846 spin_lock_irq(&conf
->resync_lock
);
848 wake_up(&conf
->wait_barrier
);
849 spin_unlock_irq(&conf
->resync_lock
);
851 bio_pair_release(bp
);
854 printk("raid10_make_request bug: can't convert block across chunks"
855 " or bigger than %dk %llu %d\n", chunk_sects
/2,
856 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
862 md_write_start(mddev
, bio
);
865 * Register the new request and wait if the reconstruction
866 * thread has put up a bar for new requests.
867 * Continue immediately if no resync is active currently.
871 cpu
= part_stat_lock();
872 part_stat_inc(cpu
, &mddev
->gendisk
->part0
, ios
[rw
]);
873 part_stat_add(cpu
, &mddev
->gendisk
->part0
, sectors
[rw
],
877 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
879 r10_bio
->master_bio
= bio
;
880 r10_bio
->sectors
= bio
->bi_size
>> 9;
882 r10_bio
->mddev
= mddev
;
883 r10_bio
->sector
= bio
->bi_sector
;
888 * read balancing logic:
890 int disk
= read_balance(conf
, r10_bio
);
891 int slot
= r10_bio
->read_slot
;
893 raid_end_bio_io(r10_bio
);
896 mirror
= conf
->mirrors
+ disk
;
898 read_bio
= bio_clone(bio
, GFP_NOIO
);
900 r10_bio
->devs
[slot
].bio
= read_bio
;
902 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
903 mirror
->rdev
->data_offset
;
904 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
905 read_bio
->bi_end_io
= raid10_end_read_request
;
906 read_bio
->bi_rw
= READ
| (do_sync
<< BIO_RW_SYNCIO
);
907 read_bio
->bi_private
= r10_bio
;
909 generic_make_request(read_bio
);
916 /* first select target devices under rcu_lock and
917 * inc refcount on their rdev. Record them by setting
920 raid10_find_phys(conf
, r10_bio
);
924 for (i
= 0; i
< conf
->copies
; i
++) {
925 int d
= r10_bio
->devs
[i
].devnum
;
926 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
927 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
928 atomic_inc(&rdev
->nr_pending
);
932 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
933 atomic_inc(&rdev
->nr_pending
);
934 r10_bio
->devs
[i
].bio
= bio
;
936 r10_bio
->devs
[i
].bio
= NULL
;
937 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
942 if (unlikely(blocked_rdev
)) {
943 /* Have to wait for this device to get unblocked, then retry */
947 for (j
= 0; j
< i
; j
++)
948 if (r10_bio
->devs
[j
].bio
) {
949 d
= r10_bio
->devs
[j
].devnum
;
950 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
953 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
958 atomic_set(&r10_bio
->remaining
, 0);
961 for (i
= 0; i
< conf
->copies
; i
++) {
963 int d
= r10_bio
->devs
[i
].devnum
;
964 if (!r10_bio
->devs
[i
].bio
)
967 mbio
= bio_clone(bio
, GFP_NOIO
);
968 r10_bio
->devs
[i
].bio
= mbio
;
970 mbio
->bi_sector
= r10_bio
->devs
[i
].addr
+
971 conf
->mirrors
[d
].rdev
->data_offset
;
972 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
973 mbio
->bi_end_io
= raid10_end_write_request
;
974 mbio
->bi_rw
= WRITE
| (do_sync
<< BIO_RW_SYNCIO
);
975 mbio
->bi_private
= r10_bio
;
977 atomic_inc(&r10_bio
->remaining
);
978 bio_list_add(&bl
, mbio
);
981 if (unlikely(!atomic_read(&r10_bio
->remaining
))) {
982 /* the array is dead */
984 raid_end_bio_io(r10_bio
);
988 bitmap_startwrite(mddev
->bitmap
, bio
->bi_sector
, r10_bio
->sectors
, 0);
989 spin_lock_irqsave(&conf
->device_lock
, flags
);
990 bio_list_merge(&conf
->pending_bio_list
, &bl
);
991 blk_plug_device(mddev
->queue
);
992 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
994 /* In case raid10d snuck in to freeze_array */
995 wake_up(&conf
->wait_barrier
);
998 md_wakeup_thread(mddev
->thread
);
1003 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
1005 conf_t
*conf
= mddev
->private;
1008 if (conf
->near_copies
< conf
->raid_disks
)
1009 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1010 if (conf
->near_copies
> 1)
1011 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
1012 if (conf
->far_copies
> 1) {
1013 if (conf
->far_offset
)
1014 seq_printf(seq
, " %d offset-copies", conf
->far_copies
);
1016 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
1018 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1019 conf
->raid_disks
- mddev
->degraded
);
1020 for (i
= 0; i
< conf
->raid_disks
; i
++)
1021 seq_printf(seq
, "%s",
1022 conf
->mirrors
[i
].rdev
&&
1023 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1024 seq_printf(seq
, "]");
1027 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1029 char b
[BDEVNAME_SIZE
];
1030 conf_t
*conf
= mddev
->private;
1033 * If it is not operational, then we have already marked it as dead
1034 * else if it is the last working disks, ignore the error, let the
1035 * next level up know.
1036 * else mark the drive as failed
1038 if (test_bit(In_sync
, &rdev
->flags
)
1039 && conf
->raid_disks
-mddev
->degraded
== 1)
1041 * Don't fail the drive, just return an IO error.
1042 * The test should really be more sophisticated than
1043 * "working_disks == 1", but it isn't critical, and
1044 * can wait until we do more sophisticated "is the drive
1045 * really dead" tests...
1048 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1049 unsigned long flags
;
1050 spin_lock_irqsave(&conf
->device_lock
, flags
);
1052 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1054 * if recovery is running, make sure it aborts.
1056 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1058 set_bit(Faulty
, &rdev
->flags
);
1059 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1060 printk(KERN_ALERT
"raid10: Disk failure on %s, disabling device.\n"
1061 "raid10: Operation continuing on %d devices.\n",
1062 bdevname(rdev
->bdev
,b
), conf
->raid_disks
- mddev
->degraded
);
1065 static void print_conf(conf_t
*conf
)
1070 printk("RAID10 conf printout:\n");
1072 printk("(!conf)\n");
1075 printk(" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1078 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1079 char b
[BDEVNAME_SIZE
];
1080 tmp
= conf
->mirrors
+ i
;
1082 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1083 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1084 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1085 bdevname(tmp
->rdev
->bdev
,b
));
1089 static void close_sync(conf_t
*conf
)
1092 allow_barrier(conf
);
1094 mempool_destroy(conf
->r10buf_pool
);
1095 conf
->r10buf_pool
= NULL
;
1098 /* check if there are enough drives for
1099 * every block to appear on atleast one
1101 static int enough(conf_t
*conf
)
1106 int n
= conf
->copies
;
1109 if (conf
->mirrors
[first
].rdev
)
1111 first
= (first
+1) % conf
->raid_disks
;
1115 } while (first
!= 0);
1119 static int raid10_spare_active(mddev_t
*mddev
)
1122 conf_t
*conf
= mddev
->private;
1126 * Find all non-in_sync disks within the RAID10 configuration
1127 * and mark them in_sync
1129 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1130 tmp
= conf
->mirrors
+ i
;
1132 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1133 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1134 unsigned long flags
;
1135 spin_lock_irqsave(&conf
->device_lock
, flags
);
1137 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1146 static int raid10_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1148 conf_t
*conf
= mddev
->private;
1153 int last
= mddev
->raid_disks
- 1;
1155 if (mddev
->recovery_cp
< MaxSector
)
1156 /* only hot-add to in-sync arrays, as recovery is
1157 * very different from resync
1163 if (rdev
->raid_disk
>= 0)
1164 first
= last
= rdev
->raid_disk
;
1166 if (rdev
->saved_raid_disk
>= 0 &&
1167 rdev
->saved_raid_disk
>= first
&&
1168 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1169 mirror
= rdev
->saved_raid_disk
;
1172 for ( ; mirror
<= last
; mirror
++)
1173 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1175 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1176 rdev
->data_offset
<< 9);
1177 /* as we don't honour merge_bvec_fn, we must
1178 * never risk violating it, so limit
1179 * ->max_segments to one lying with a single
1180 * page, as a one page request is never in
1183 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1184 blk_queue_max_segments(mddev
->queue
, 1);
1185 blk_queue_segment_boundary(mddev
->queue
,
1186 PAGE_CACHE_SIZE
- 1);
1189 p
->head_position
= 0;
1190 rdev
->raid_disk
= mirror
;
1192 if (rdev
->saved_raid_disk
!= mirror
)
1194 rcu_assign_pointer(p
->rdev
, rdev
);
1198 md_integrity_add_rdev(rdev
, mddev
);
1203 static int raid10_remove_disk(mddev_t
*mddev
, int number
)
1205 conf_t
*conf
= mddev
->private;
1208 mirror_info_t
*p
= conf
->mirrors
+ number
;
1213 if (test_bit(In_sync
, &rdev
->flags
) ||
1214 atomic_read(&rdev
->nr_pending
)) {
1218 /* Only remove faulty devices in recovery
1221 if (!test_bit(Faulty
, &rdev
->flags
) &&
1228 if (atomic_read(&rdev
->nr_pending
)) {
1229 /* lost the race, try later */
1234 md_integrity_register(mddev
);
1243 static void end_sync_read(struct bio
*bio
, int error
)
1245 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1246 conf_t
*conf
= r10_bio
->mddev
->private;
1249 for (i
=0; i
<conf
->copies
; i
++)
1250 if (r10_bio
->devs
[i
].bio
== bio
)
1252 BUG_ON(i
== conf
->copies
);
1253 update_head_pos(i
, r10_bio
);
1254 d
= r10_bio
->devs
[i
].devnum
;
1256 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1257 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1259 atomic_add(r10_bio
->sectors
,
1260 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1261 if (!test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
1262 md_error(r10_bio
->mddev
,
1263 conf
->mirrors
[d
].rdev
);
1266 /* for reconstruct, we always reschedule after a read.
1267 * for resync, only after all reads
1269 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1270 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1271 atomic_dec_and_test(&r10_bio
->remaining
)) {
1272 /* we have read all the blocks,
1273 * do the comparison in process context in raid10d
1275 reschedule_retry(r10_bio
);
1279 static void end_sync_write(struct bio
*bio
, int error
)
1281 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1282 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1283 mddev_t
*mddev
= r10_bio
->mddev
;
1284 conf_t
*conf
= mddev
->private;
1287 for (i
= 0; i
< conf
->copies
; i
++)
1288 if (r10_bio
->devs
[i
].bio
== bio
)
1290 d
= r10_bio
->devs
[i
].devnum
;
1293 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1295 update_head_pos(i
, r10_bio
);
1297 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1298 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1299 if (r10_bio
->master_bio
== NULL
) {
1300 /* the primary of several recovery bios */
1301 sector_t s
= r10_bio
->sectors
;
1303 md_done_sync(mddev
, s
, 1);
1306 r10bio_t
*r10_bio2
= (r10bio_t
*)r10_bio
->master_bio
;
1314 * Note: sync and recover and handled very differently for raid10
1315 * This code is for resync.
1316 * For resync, we read through virtual addresses and read all blocks.
1317 * If there is any error, we schedule a write. The lowest numbered
1318 * drive is authoritative.
1319 * However requests come for physical address, so we need to map.
1320 * For every physical address there are raid_disks/copies virtual addresses,
1321 * which is always are least one, but is not necessarly an integer.
1322 * This means that a physical address can span multiple chunks, so we may
1323 * have to submit multiple io requests for a single sync request.
1326 * We check if all blocks are in-sync and only write to blocks that
1329 static void sync_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1331 conf_t
*conf
= mddev
->private;
1333 struct bio
*tbio
, *fbio
;
1335 atomic_set(&r10_bio
->remaining
, 1);
1337 /* find the first device with a block */
1338 for (i
=0; i
<conf
->copies
; i
++)
1339 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1342 if (i
== conf
->copies
)
1346 fbio
= r10_bio
->devs
[i
].bio
;
1348 /* now find blocks with errors */
1349 for (i
=0 ; i
< conf
->copies
; i
++) {
1351 int vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1353 tbio
= r10_bio
->devs
[i
].bio
;
1355 if (tbio
->bi_end_io
!= end_sync_read
)
1359 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1360 /* We know that the bi_io_vec layout is the same for
1361 * both 'first' and 'i', so we just compare them.
1362 * All vec entries are PAGE_SIZE;
1364 for (j
= 0; j
< vcnt
; j
++)
1365 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1366 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1371 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1373 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1374 /* Don't fix anything. */
1376 /* Ok, we need to write this bio
1377 * First we need to fixup bv_offset, bv_len and
1378 * bi_vecs, as the read request might have corrupted these
1380 tbio
->bi_vcnt
= vcnt
;
1381 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1383 tbio
->bi_phys_segments
= 0;
1384 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1385 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1386 tbio
->bi_next
= NULL
;
1387 tbio
->bi_rw
= WRITE
;
1388 tbio
->bi_private
= r10_bio
;
1389 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1391 for (j
=0; j
< vcnt
; j
++) {
1392 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1393 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1395 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1396 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1399 tbio
->bi_end_io
= end_sync_write
;
1401 d
= r10_bio
->devs
[i
].devnum
;
1402 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1403 atomic_inc(&r10_bio
->remaining
);
1404 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1406 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1407 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1408 generic_make_request(tbio
);
1412 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1413 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1419 * Now for the recovery code.
1420 * Recovery happens across physical sectors.
1421 * We recover all non-is_sync drives by finding the virtual address of
1422 * each, and then choose a working drive that also has that virt address.
1423 * There is a separate r10_bio for each non-in_sync drive.
1424 * Only the first two slots are in use. The first for reading,
1425 * The second for writing.
1429 static void recovery_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1431 conf_t
*conf
= mddev
->private;
1433 struct bio
*bio
, *wbio
;
1436 /* move the pages across to the second bio
1437 * and submit the write request
1439 bio
= r10_bio
->devs
[0].bio
;
1440 wbio
= r10_bio
->devs
[1].bio
;
1441 for (i
=0; i
< wbio
->bi_vcnt
; i
++) {
1442 struct page
*p
= bio
->bi_io_vec
[i
].bv_page
;
1443 bio
->bi_io_vec
[i
].bv_page
= wbio
->bi_io_vec
[i
].bv_page
;
1444 wbio
->bi_io_vec
[i
].bv_page
= p
;
1446 d
= r10_bio
->devs
[1].devnum
;
1448 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1449 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1450 if (test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
1451 generic_make_request(wbio
);
1453 bio_endio(wbio
, -EIO
);
1458 * Used by fix_read_error() to decay the per rdev read_errors.
1459 * We halve the read error count for every hour that has elapsed
1460 * since the last recorded read error.
1463 static void check_decay_read_errors(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1465 struct timespec cur_time_mon
;
1466 unsigned long hours_since_last
;
1467 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
1469 ktime_get_ts(&cur_time_mon
);
1471 if (rdev
->last_read_error
.tv_sec
== 0 &&
1472 rdev
->last_read_error
.tv_nsec
== 0) {
1473 /* first time we've seen a read error */
1474 rdev
->last_read_error
= cur_time_mon
;
1478 hours_since_last
= (cur_time_mon
.tv_sec
-
1479 rdev
->last_read_error
.tv_sec
) / 3600;
1481 rdev
->last_read_error
= cur_time_mon
;
1484 * if hours_since_last is > the number of bits in read_errors
1485 * just set read errors to 0. We do this to avoid
1486 * overflowing the shift of read_errors by hours_since_last.
1488 if (hours_since_last
>= 8 * sizeof(read_errors
))
1489 atomic_set(&rdev
->read_errors
, 0);
1491 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
1495 * This is a kernel thread which:
1497 * 1. Retries failed read operations on working mirrors.
1498 * 2. Updates the raid superblock when problems encounter.
1499 * 3. Performs writes following reads for array synchronising.
1502 static void fix_read_error(conf_t
*conf
, mddev_t
*mddev
, r10bio_t
*r10_bio
)
1504 int sect
= 0; /* Offset from r10_bio->sector */
1505 int sectors
= r10_bio
->sectors
;
1507 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
1508 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1511 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1512 if (rdev
) { /* If rdev is not NULL */
1513 char b
[BDEVNAME_SIZE
];
1514 int cur_read_error_count
= 0;
1516 bdevname(rdev
->bdev
, b
);
1518 if (test_bit(Faulty
, &rdev
->flags
)) {
1520 /* drive has already been failed, just ignore any
1521 more fix_read_error() attempts */
1525 check_decay_read_errors(mddev
, rdev
);
1526 atomic_inc(&rdev
->read_errors
);
1527 cur_read_error_count
= atomic_read(&rdev
->read_errors
);
1528 if (cur_read_error_count
> max_read_errors
) {
1531 "raid10: %s: Raid device exceeded "
1532 "read_error threshold "
1533 "[cur %d:max %d]\n",
1534 b
, cur_read_error_count
, max_read_errors
);
1536 "raid10: %s: Failing raid "
1538 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1546 int sl
= r10_bio
->read_slot
;
1550 if (s
> (PAGE_SIZE
>>9))
1555 d
= r10_bio
->devs
[sl
].devnum
;
1556 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1558 test_bit(In_sync
, &rdev
->flags
)) {
1559 atomic_inc(&rdev
->nr_pending
);
1561 success
= sync_page_io(rdev
->bdev
,
1562 r10_bio
->devs
[sl
].addr
+
1563 sect
+ rdev
->data_offset
,
1565 conf
->tmppage
, READ
);
1566 rdev_dec_pending(rdev
, mddev
);
1572 if (sl
== conf
->copies
)
1574 } while (!success
&& sl
!= r10_bio
->read_slot
);
1578 /* Cannot read from anywhere -- bye bye array */
1579 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1580 md_error(mddev
, conf
->mirrors
[dn
].rdev
);
1585 /* write it back and re-read */
1587 while (sl
!= r10_bio
->read_slot
) {
1588 char b
[BDEVNAME_SIZE
];
1593 d
= r10_bio
->devs
[sl
].devnum
;
1594 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1596 test_bit(In_sync
, &rdev
->flags
)) {
1597 atomic_inc(&rdev
->nr_pending
);
1599 atomic_add(s
, &rdev
->corrected_errors
);
1600 if (sync_page_io(rdev
->bdev
,
1601 r10_bio
->devs
[sl
].addr
+
1602 sect
+ rdev
->data_offset
,
1603 s
<<9, conf
->tmppage
, WRITE
)
1605 /* Well, this device is dead */
1607 "raid10:%s: read correction "
1609 " (%d sectors at %llu on %s)\n",
1611 (unsigned long long)(sect
+
1613 bdevname(rdev
->bdev
, b
));
1614 printk(KERN_NOTICE
"raid10:%s: failing "
1616 bdevname(rdev
->bdev
, b
));
1617 md_error(mddev
, rdev
);
1619 rdev_dec_pending(rdev
, mddev
);
1624 while (sl
!= r10_bio
->read_slot
) {
1629 d
= r10_bio
->devs
[sl
].devnum
;
1630 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1632 test_bit(In_sync
, &rdev
->flags
)) {
1633 char b
[BDEVNAME_SIZE
];
1634 atomic_inc(&rdev
->nr_pending
);
1636 if (sync_page_io(rdev
->bdev
,
1637 r10_bio
->devs
[sl
].addr
+
1638 sect
+ rdev
->data_offset
,
1639 s
<<9, conf
->tmppage
,
1641 /* Well, this device is dead */
1643 "raid10:%s: unable to read back "
1645 " (%d sectors at %llu on %s)\n",
1647 (unsigned long long)(sect
+
1649 bdevname(rdev
->bdev
, b
));
1650 printk(KERN_NOTICE
"raid10:%s: failing drive\n",
1651 bdevname(rdev
->bdev
, b
));
1653 md_error(mddev
, rdev
);
1656 "raid10:%s: read error corrected"
1657 " (%d sectors at %llu on %s)\n",
1659 (unsigned long long)(sect
+
1661 bdevname(rdev
->bdev
, b
));
1664 rdev_dec_pending(rdev
, mddev
);
1675 static void raid10d(mddev_t
*mddev
)
1679 unsigned long flags
;
1680 conf_t
*conf
= mddev
->private;
1681 struct list_head
*head
= &conf
->retry_list
;
1685 md_check_recovery(mddev
);
1688 char b
[BDEVNAME_SIZE
];
1690 unplug
+= flush_pending_writes(conf
);
1692 spin_lock_irqsave(&conf
->device_lock
, flags
);
1693 if (list_empty(head
)) {
1694 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1697 r10_bio
= list_entry(head
->prev
, r10bio_t
, retry_list
);
1698 list_del(head
->prev
);
1700 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1702 mddev
= r10_bio
->mddev
;
1703 conf
= mddev
->private;
1704 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
)) {
1705 sync_request_write(mddev
, r10_bio
);
1707 } else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
1708 recovery_request_write(mddev
, r10_bio
);
1712 /* we got a read error. Maybe the drive is bad. Maybe just
1713 * the block and we can fix it.
1714 * We freeze all other IO, and try reading the block from
1715 * other devices. When we find one, we re-write
1716 * and check it that fixes the read error.
1717 * This is all done synchronously while the array is
1720 if (mddev
->ro
== 0) {
1722 fix_read_error(conf
, mddev
, r10_bio
);
1723 unfreeze_array(conf
);
1726 bio
= r10_bio
->devs
[r10_bio
->read_slot
].bio
;
1727 r10_bio
->devs
[r10_bio
->read_slot
].bio
=
1728 mddev
->ro
? IO_BLOCKED
: NULL
;
1729 mirror
= read_balance(conf
, r10_bio
);
1731 printk(KERN_ALERT
"raid10: %s: unrecoverable I/O"
1732 " read error for block %llu\n",
1733 bdevname(bio
->bi_bdev
,b
),
1734 (unsigned long long)r10_bio
->sector
);
1735 raid_end_bio_io(r10_bio
);
1738 const bool do_sync
= bio_rw_flagged(r10_bio
->master_bio
, BIO_RW_SYNCIO
);
1740 rdev
= conf
->mirrors
[mirror
].rdev
;
1741 if (printk_ratelimit())
1742 printk(KERN_ERR
"raid10: %s: redirecting sector %llu to"
1743 " another mirror\n",
1744 bdevname(rdev
->bdev
,b
),
1745 (unsigned long long)r10_bio
->sector
);
1746 bio
= bio_clone(r10_bio
->master_bio
, GFP_NOIO
);
1747 r10_bio
->devs
[r10_bio
->read_slot
].bio
= bio
;
1748 bio
->bi_sector
= r10_bio
->devs
[r10_bio
->read_slot
].addr
1749 + rdev
->data_offset
;
1750 bio
->bi_bdev
= rdev
->bdev
;
1751 bio
->bi_rw
= READ
| (do_sync
<< BIO_RW_SYNCIO
);
1752 bio
->bi_private
= r10_bio
;
1753 bio
->bi_end_io
= raid10_end_read_request
;
1755 generic_make_request(bio
);
1761 unplug_slaves(mddev
);
1765 static int init_resync(conf_t
*conf
)
1769 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1770 BUG_ON(conf
->r10buf_pool
);
1771 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
1772 if (!conf
->r10buf_pool
)
1774 conf
->next_resync
= 0;
1779 * perform a "sync" on one "block"
1781 * We need to make sure that no normal I/O request - particularly write
1782 * requests - conflict with active sync requests.
1784 * This is achieved by tracking pending requests and a 'barrier' concept
1785 * that can be installed to exclude normal IO requests.
1787 * Resync and recovery are handled very differently.
1788 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1790 * For resync, we iterate over virtual addresses, read all copies,
1791 * and update if there are differences. If only one copy is live,
1793 * For recovery, we iterate over physical addresses, read a good
1794 * value for each non-in_sync drive, and over-write.
1796 * So, for recovery we may have several outstanding complex requests for a
1797 * given address, one for each out-of-sync device. We model this by allocating
1798 * a number of r10_bio structures, one for each out-of-sync device.
1799 * As we setup these structures, we collect all bio's together into a list
1800 * which we then process collectively to add pages, and then process again
1801 * to pass to generic_make_request.
1803 * The r10_bio structures are linked using a borrowed master_bio pointer.
1804 * This link is counted in ->remaining. When the r10_bio that points to NULL
1805 * has its remaining count decremented to 0, the whole complex operation
1810 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
1812 conf_t
*conf
= mddev
->private;
1814 struct bio
*biolist
= NULL
, *bio
;
1815 sector_t max_sector
, nr_sectors
;
1821 sector_t sectors_skipped
= 0;
1822 int chunks_skipped
= 0;
1824 if (!conf
->r10buf_pool
)
1825 if (init_resync(conf
))
1829 max_sector
= mddev
->dev_sectors
;
1830 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1831 max_sector
= mddev
->resync_max_sectors
;
1832 if (sector_nr
>= max_sector
) {
1833 /* If we aborted, we need to abort the
1834 * sync on the 'current' bitmap chucks (there can
1835 * be several when recovering multiple devices).
1836 * as we may have started syncing it but not finished.
1837 * We can find the current address in
1838 * mddev->curr_resync, but for recovery,
1839 * we need to convert that to several
1840 * virtual addresses.
1842 if (mddev
->curr_resync
< max_sector
) { /* aborted */
1843 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1844 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
1846 else for (i
=0; i
<conf
->raid_disks
; i
++) {
1848 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
1849 bitmap_end_sync(mddev
->bitmap
, sect
,
1852 } else /* completed sync */
1855 bitmap_close_sync(mddev
->bitmap
);
1858 return sectors_skipped
;
1860 if (chunks_skipped
>= conf
->raid_disks
) {
1861 /* if there has been nothing to do on any drive,
1862 * then there is nothing to do at all..
1865 return (max_sector
- sector_nr
) + sectors_skipped
;
1868 if (max_sector
> mddev
->resync_max
)
1869 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
1871 /* make sure whole request will fit in a chunk - if chunks
1874 if (conf
->near_copies
< conf
->raid_disks
&&
1875 max_sector
> (sector_nr
| conf
->chunk_mask
))
1876 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
1878 * If there is non-resync activity waiting for us then
1879 * put in a delay to throttle resync.
1881 if (!go_faster
&& conf
->nr_waiting
)
1882 msleep_interruptible(1000);
1884 /* Again, very different code for resync and recovery.
1885 * Both must result in an r10bio with a list of bios that
1886 * have bi_end_io, bi_sector, bi_bdev set,
1887 * and bi_private set to the r10bio.
1888 * For recovery, we may actually create several r10bios
1889 * with 2 bios in each, that correspond to the bios in the main one.
1890 * In this case, the subordinate r10bios link back through a
1891 * borrowed master_bio pointer, and the counter in the master
1892 * includes a ref from each subordinate.
1894 /* First, we decide what to do and set ->bi_end_io
1895 * To end_sync_read if we want to read, and
1896 * end_sync_write if we will want to write.
1899 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
1900 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
1901 /* recovery... the complicated one */
1905 for (i
=0 ; i
<conf
->raid_disks
; i
++)
1906 if (conf
->mirrors
[i
].rdev
&&
1907 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
)) {
1908 int still_degraded
= 0;
1909 /* want to reconstruct this device */
1910 r10bio_t
*rb2
= r10_bio
;
1911 sector_t sect
= raid10_find_virt(conf
, sector_nr
, i
);
1913 /* Unless we are doing a full sync, we only need
1914 * to recover the block if it is set in the bitmap
1916 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1918 if (sync_blocks
< max_sync
)
1919 max_sync
= sync_blocks
;
1922 /* yep, skip the sync_blocks here, but don't assume
1923 * that there will never be anything to do here
1925 chunks_skipped
= -1;
1929 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1930 raise_barrier(conf
, rb2
!= NULL
);
1931 atomic_set(&r10_bio
->remaining
, 0);
1933 r10_bio
->master_bio
= (struct bio
*)rb2
;
1935 atomic_inc(&rb2
->remaining
);
1936 r10_bio
->mddev
= mddev
;
1937 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
1938 r10_bio
->sector
= sect
;
1940 raid10_find_phys(conf
, r10_bio
);
1942 /* Need to check if the array will still be
1945 for (j
=0; j
<conf
->raid_disks
; j
++)
1946 if (conf
->mirrors
[j
].rdev
== NULL
||
1947 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
1952 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1953 &sync_blocks
, still_degraded
);
1955 for (j
=0; j
<conf
->copies
;j
++) {
1956 int d
= r10_bio
->devs
[j
].devnum
;
1957 if (conf
->mirrors
[d
].rdev
&&
1958 test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
)) {
1959 /* This is where we read from */
1960 bio
= r10_bio
->devs
[0].bio
;
1961 bio
->bi_next
= biolist
;
1963 bio
->bi_private
= r10_bio
;
1964 bio
->bi_end_io
= end_sync_read
;
1966 bio
->bi_sector
= r10_bio
->devs
[j
].addr
+
1967 conf
->mirrors
[d
].rdev
->data_offset
;
1968 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1969 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1970 atomic_inc(&r10_bio
->remaining
);
1971 /* and we write to 'i' */
1973 for (k
=0; k
<conf
->copies
; k
++)
1974 if (r10_bio
->devs
[k
].devnum
== i
)
1976 BUG_ON(k
== conf
->copies
);
1977 bio
= r10_bio
->devs
[1].bio
;
1978 bio
->bi_next
= biolist
;
1980 bio
->bi_private
= r10_bio
;
1981 bio
->bi_end_io
= end_sync_write
;
1983 bio
->bi_sector
= r10_bio
->devs
[k
].addr
+
1984 conf
->mirrors
[i
].rdev
->data_offset
;
1985 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1987 r10_bio
->devs
[0].devnum
= d
;
1988 r10_bio
->devs
[1].devnum
= i
;
1993 if (j
== conf
->copies
) {
1994 /* Cannot recover, so abort the recovery */
1997 atomic_dec(&rb2
->remaining
);
1999 if (!test_and_set_bit(MD_RECOVERY_INTR
,
2001 printk(KERN_INFO
"raid10: %s: insufficient working devices for recovery.\n",
2006 if (biolist
== NULL
) {
2008 r10bio_t
*rb2
= r10_bio
;
2009 r10_bio
= (r10bio_t
*) rb2
->master_bio
;
2010 rb2
->master_bio
= NULL
;
2016 /* resync. Schedule a read for every block at this virt offset */
2019 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
2021 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2022 &sync_blocks
, mddev
->degraded
) &&
2023 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2024 /* We can skip this block */
2026 return sync_blocks
+ sectors_skipped
;
2028 if (sync_blocks
< max_sync
)
2029 max_sync
= sync_blocks
;
2030 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2032 r10_bio
->mddev
= mddev
;
2033 atomic_set(&r10_bio
->remaining
, 0);
2034 raise_barrier(conf
, 0);
2035 conf
->next_resync
= sector_nr
;
2037 r10_bio
->master_bio
= NULL
;
2038 r10_bio
->sector
= sector_nr
;
2039 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
2040 raid10_find_phys(conf
, r10_bio
);
2041 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
2043 for (i
=0; i
<conf
->copies
; i
++) {
2044 int d
= r10_bio
->devs
[i
].devnum
;
2045 bio
= r10_bio
->devs
[i
].bio
;
2046 bio
->bi_end_io
= NULL
;
2047 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2048 if (conf
->mirrors
[d
].rdev
== NULL
||
2049 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
2051 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2052 atomic_inc(&r10_bio
->remaining
);
2053 bio
->bi_next
= biolist
;
2055 bio
->bi_private
= r10_bio
;
2056 bio
->bi_end_io
= end_sync_read
;
2058 bio
->bi_sector
= r10_bio
->devs
[i
].addr
+
2059 conf
->mirrors
[d
].rdev
->data_offset
;
2060 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2065 for (i
=0; i
<conf
->copies
; i
++) {
2066 int d
= r10_bio
->devs
[i
].devnum
;
2067 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
2068 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
2076 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2078 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2080 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
2083 bio
->bi_phys_segments
= 0;
2088 if (sector_nr
+ max_sync
< max_sector
)
2089 max_sector
= sector_nr
+ max_sync
;
2092 int len
= PAGE_SIZE
;
2094 if (sector_nr
+ (len
>>9) > max_sector
)
2095 len
= (max_sector
- sector_nr
) << 9;
2098 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2099 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2100 if (bio_add_page(bio
, page
, len
, 0) == 0) {
2103 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2104 for (bio2
= biolist
; bio2
&& bio2
!= bio
; bio2
= bio2
->bi_next
) {
2105 /* remove last page from this bio */
2107 bio2
->bi_size
-= len
;
2108 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
2114 nr_sectors
+= len
>>9;
2115 sector_nr
+= len
>>9;
2116 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
2118 r10_bio
->sectors
= nr_sectors
;
2122 biolist
= biolist
->bi_next
;
2124 bio
->bi_next
= NULL
;
2125 r10_bio
= bio
->bi_private
;
2126 r10_bio
->sectors
= nr_sectors
;
2128 if (bio
->bi_end_io
== end_sync_read
) {
2129 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2130 generic_make_request(bio
);
2134 if (sectors_skipped
)
2135 /* pretend they weren't skipped, it makes
2136 * no important difference in this case
2138 md_done_sync(mddev
, sectors_skipped
, 1);
2140 return sectors_skipped
+ nr_sectors
;
2142 /* There is nowhere to write, so all non-sync
2143 * drives must be failed, so try the next chunk...
2145 if (sector_nr
+ max_sync
< max_sector
)
2146 max_sector
= sector_nr
+ max_sync
;
2148 sectors_skipped
+= (max_sector
- sector_nr
);
2150 sector_nr
= max_sector
;
2155 raid10_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
2158 conf_t
*conf
= mddev
->private;
2161 raid_disks
= mddev
->raid_disks
;
2163 sectors
= mddev
->dev_sectors
;
2165 size
= sectors
>> conf
->chunk_shift
;
2166 sector_div(size
, conf
->far_copies
);
2167 size
= size
* raid_disks
;
2168 sector_div(size
, conf
->near_copies
);
2170 return size
<< conf
->chunk_shift
;
2173 static int run(mddev_t
*mddev
)
2176 int i
, disk_idx
, chunk_size
;
2177 mirror_info_t
*disk
;
2180 sector_t stride
, size
;
2182 if (mddev
->chunk_sectors
< (PAGE_SIZE
>> 9) ||
2183 !is_power_of_2(mddev
->chunk_sectors
)) {
2184 printk(KERN_ERR
"md/raid10: chunk size must be "
2185 "at least PAGE_SIZE(%ld) and be a power of 2.\n", PAGE_SIZE
);
2189 nc
= mddev
->layout
& 255;
2190 fc
= (mddev
->layout
>> 8) & 255;
2191 fo
= mddev
->layout
& (1<<16);
2192 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
2193 (mddev
->layout
>> 17)) {
2194 printk(KERN_ERR
"raid10: %s: unsupported raid10 layout: 0x%8x\n",
2195 mdname(mddev
), mddev
->layout
);
2199 * copy the already verified devices into our private RAID10
2200 * bookkeeping area. [whatever we allocate in run(),
2201 * should be freed in stop()]
2203 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
2204 mddev
->private = conf
;
2206 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
2210 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
2212 if (!conf
->mirrors
) {
2213 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
2218 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2222 conf
->raid_disks
= mddev
->raid_disks
;
2223 conf
->near_copies
= nc
;
2224 conf
->far_copies
= fc
;
2225 conf
->copies
= nc
*fc
;
2226 conf
->far_offset
= fo
;
2227 conf
->chunk_mask
= mddev
->chunk_sectors
- 1;
2228 conf
->chunk_shift
= ffz(~mddev
->chunk_sectors
);
2229 size
= mddev
->dev_sectors
>> conf
->chunk_shift
;
2230 sector_div(size
, fc
);
2231 size
= size
* conf
->raid_disks
;
2232 sector_div(size
, nc
);
2233 /* 'size' is now the number of chunks in the array */
2234 /* calculate "used chunks per device" in 'stride' */
2235 stride
= size
* conf
->copies
;
2237 /* We need to round up when dividing by raid_disks to
2238 * get the stride size.
2240 stride
+= conf
->raid_disks
- 1;
2241 sector_div(stride
, conf
->raid_disks
);
2242 mddev
->dev_sectors
= stride
<< conf
->chunk_shift
;
2247 sector_div(stride
, fc
);
2248 conf
->stride
= stride
<< conf
->chunk_shift
;
2250 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
2251 r10bio_pool_free
, conf
);
2252 if (!conf
->r10bio_pool
) {
2253 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
2258 conf
->mddev
= mddev
;
2259 spin_lock_init(&conf
->device_lock
);
2260 mddev
->queue
->queue_lock
= &conf
->device_lock
;
2262 chunk_size
= mddev
->chunk_sectors
<< 9;
2263 blk_queue_io_min(mddev
->queue
, chunk_size
);
2264 if (conf
->raid_disks
% conf
->near_copies
)
2265 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->raid_disks
);
2267 blk_queue_io_opt(mddev
->queue
, chunk_size
*
2268 (conf
->raid_disks
/ conf
->near_copies
));
2270 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
2271 disk_idx
= rdev
->raid_disk
;
2272 if (disk_idx
>= mddev
->raid_disks
2275 disk
= conf
->mirrors
+ disk_idx
;
2278 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2279 rdev
->data_offset
<< 9);
2280 /* as we don't honour merge_bvec_fn, we must never risk
2281 * violating it, so limit max_segments to 1 lying
2282 * within a single page.
2284 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
2285 blk_queue_max_segments(mddev
->queue
, 1);
2286 blk_queue_segment_boundary(mddev
->queue
,
2287 PAGE_CACHE_SIZE
- 1);
2290 disk
->head_position
= 0;
2292 INIT_LIST_HEAD(&conf
->retry_list
);
2294 spin_lock_init(&conf
->resync_lock
);
2295 init_waitqueue_head(&conf
->wait_barrier
);
2297 /* need to check that every block has at least one working mirror */
2298 if (!enough(conf
)) {
2299 printk(KERN_ERR
"raid10: not enough operational mirrors for %s\n",
2304 mddev
->degraded
= 0;
2305 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2307 disk
= conf
->mirrors
+ i
;
2310 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2311 disk
->head_position
= 0;
2319 mddev
->thread
= md_register_thread(raid10d
, mddev
, NULL
);
2320 if (!mddev
->thread
) {
2322 "raid10: couldn't allocate thread for %s\n",
2327 if (mddev
->recovery_cp
!= MaxSector
)
2328 printk(KERN_NOTICE
"raid10: %s is not clean"
2329 " -- starting background reconstruction\n",
2332 "raid10: raid set %s active with %d out of %d devices\n",
2333 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2336 * Ok, everything is just fine now
2338 md_set_array_sectors(mddev
, raid10_size(mddev
, 0, 0));
2339 mddev
->resync_max_sectors
= raid10_size(mddev
, 0, 0);
2341 mddev
->queue
->unplug_fn
= raid10_unplug
;
2342 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
2343 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2345 /* Calculate max read-ahead size.
2346 * We need to readahead at least twice a whole stripe....
2350 int stripe
= conf
->raid_disks
*
2351 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
2352 stripe
/= conf
->near_copies
;
2353 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
2354 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
2357 if (conf
->near_copies
< mddev
->raid_disks
)
2358 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
2359 md_integrity_register(mddev
);
2363 if (conf
->r10bio_pool
)
2364 mempool_destroy(conf
->r10bio_pool
);
2365 safe_put_page(conf
->tmppage
);
2366 kfree(conf
->mirrors
);
2368 mddev
->private = NULL
;
2373 static int stop(mddev_t
*mddev
)
2375 conf_t
*conf
= mddev
->private;
2377 raise_barrier(conf
, 0);
2378 lower_barrier(conf
);
2380 md_unregister_thread(mddev
->thread
);
2381 mddev
->thread
= NULL
;
2382 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
2383 if (conf
->r10bio_pool
)
2384 mempool_destroy(conf
->r10bio_pool
);
2385 kfree(conf
->mirrors
);
2387 mddev
->private = NULL
;
2391 static void raid10_quiesce(mddev_t
*mddev
, int state
)
2393 conf_t
*conf
= mddev
->private;
2397 raise_barrier(conf
, 0);
2400 lower_barrier(conf
);
2405 static struct mdk_personality raid10_personality
=
2409 .owner
= THIS_MODULE
,
2410 .make_request
= make_request
,
2414 .error_handler
= error
,
2415 .hot_add_disk
= raid10_add_disk
,
2416 .hot_remove_disk
= raid10_remove_disk
,
2417 .spare_active
= raid10_spare_active
,
2418 .sync_request
= sync_request
,
2419 .quiesce
= raid10_quiesce
,
2420 .size
= raid10_size
,
2423 static int __init
raid_init(void)
2425 return register_md_personality(&raid10_personality
);
2428 static void raid_exit(void)
2430 unregister_md_personality(&raid10_personality
);
2433 module_init(raid_init
);
2434 module_exit(raid_exit
);
2435 MODULE_LICENSE("GPL");
2436 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2437 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2438 MODULE_ALIAS("md-raid10");
2439 MODULE_ALIAS("md-level-10");