2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for futher copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include "dm-bio-list.h"
22 #include <linux/raid/raid10.h>
23 #include <linux/raid/bitmap.h>
26 * RAID10 provides a combination of RAID0 and RAID1 functionality.
27 * The layout of data is defined by
30 * near_copies (stored in low byte of layout)
31 * far_copies (stored in second byte of layout)
32 * far_offset (stored in bit 16 of layout )
34 * The data to be stored is divided into chunks using chunksize.
35 * Each device is divided into far_copies sections.
36 * In each section, chunks are laid out in a style similar to raid0, but
37 * near_copies copies of each chunk is stored (each on a different drive).
38 * The starting device for each section is offset near_copies from the starting
39 * device of the previous section.
40 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
42 * near_copies and far_copies must be at least one, and their product is at most
45 * If far_offset is true, then the far_copies are handled a bit differently.
46 * The copies are still in different stripes, but instead of be very far apart
47 * on disk, there are adjacent stripes.
51 * Number of guaranteed r10bios in case of extreme VM load:
53 #define NR_RAID10_BIOS 256
55 static void unplug_slaves(mddev_t
*mddev
);
57 static void allow_barrier(conf_t
*conf
);
58 static void lower_barrier(conf_t
*conf
);
60 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
64 int size
= offsetof(struct r10bio_s
, devs
[conf
->copies
]);
66 /* allocate a r10bio with room for raid_disks entries in the bios array */
67 r10_bio
= kzalloc(size
, gfp_flags
);
69 unplug_slaves(conf
->mddev
);
74 static void r10bio_pool_free(void *r10_bio
, void *data
)
79 /* Maximum size of each resync request */
80 #define RESYNC_BLOCK_SIZE (64*1024)
81 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
82 /* amount of memory to reserve for resync requests */
83 #define RESYNC_WINDOW (1024*1024)
84 /* maximum number of concurrent requests, memory permitting */
85 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
88 * When performing a resync, we need to read and compare, so
89 * we need as many pages are there are copies.
90 * When performing a recovery, we need 2 bios, one for read,
91 * one for write (we recover only one drive per r10buf)
94 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
103 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
105 unplug_slaves(conf
->mddev
);
109 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
110 nalloc
= conf
->copies
; /* resync */
112 nalloc
= 2; /* recovery */
117 for (j
= nalloc
; j
-- ; ) {
118 bio
= bio_alloc(gfp_flags
, RESYNC_PAGES
);
121 r10_bio
->devs
[j
].bio
= bio
;
124 * Allocate RESYNC_PAGES data pages and attach them
127 for (j
= 0 ; j
< nalloc
; j
++) {
128 bio
= r10_bio
->devs
[j
].bio
;
129 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
130 page
= alloc_page(gfp_flags
);
134 bio
->bi_io_vec
[i
].bv_page
= page
;
142 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
144 for (i
= 0; i
< RESYNC_PAGES
; i
++)
145 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
148 while ( ++j
< nalloc
)
149 bio_put(r10_bio
->devs
[j
].bio
);
150 r10bio_pool_free(r10_bio
, conf
);
154 static void r10buf_pool_free(void *__r10_bio
, void *data
)
158 r10bio_t
*r10bio
= __r10_bio
;
161 for (j
=0; j
< conf
->copies
; j
++) {
162 struct bio
*bio
= r10bio
->devs
[j
].bio
;
164 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
165 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
166 bio
->bi_io_vec
[i
].bv_page
= NULL
;
171 r10bio_pool_free(r10bio
, conf
);
174 static void put_all_bios(conf_t
*conf
, r10bio_t
*r10_bio
)
178 for (i
= 0; i
< conf
->copies
; i
++) {
179 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
180 if (*bio
&& *bio
!= IO_BLOCKED
)
186 static void free_r10bio(r10bio_t
*r10_bio
)
188 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
191 * Wake up any possible resync thread that waits for the device
196 put_all_bios(conf
, r10_bio
);
197 mempool_free(r10_bio
, conf
->r10bio_pool
);
200 static void put_buf(r10bio_t
*r10_bio
)
202 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
204 mempool_free(r10_bio
, conf
->r10buf_pool
);
209 static void reschedule_retry(r10bio_t
*r10_bio
)
212 mddev_t
*mddev
= r10_bio
->mddev
;
213 conf_t
*conf
= mddev_to_conf(mddev
);
215 spin_lock_irqsave(&conf
->device_lock
, flags
);
216 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
218 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
220 /* wake up frozen array... */
221 wake_up(&conf
->wait_barrier
);
223 md_wakeup_thread(mddev
->thread
);
227 * raid_end_bio_io() is called when we have finished servicing a mirrored
228 * operation and are ready to return a success/failure code to the buffer
231 static void raid_end_bio_io(r10bio_t
*r10_bio
)
233 struct bio
*bio
= r10_bio
->master_bio
;
236 test_bit(R10BIO_Uptodate
, &r10_bio
->state
) ? 0 : -EIO
);
237 free_r10bio(r10_bio
);
241 * Update disk head position estimator based on IRQ completion info.
243 static inline void update_head_pos(int slot
, r10bio_t
*r10_bio
)
245 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
247 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
248 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
251 static void raid10_end_read_request(struct bio
*bio
, int error
)
253 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
254 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
256 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
259 slot
= r10_bio
->read_slot
;
260 dev
= r10_bio
->devs
[slot
].devnum
;
262 * this branch is our 'one mirror IO has finished' event handler:
264 update_head_pos(slot
, r10_bio
);
268 * Set R10BIO_Uptodate in our master bio, so that
269 * we will return a good error code to the higher
270 * levels even if IO on some other mirrored buffer fails.
272 * The 'master' represents the composite IO operation to
273 * user-side. So if something waits for IO, then it will
274 * wait for the 'master' bio.
276 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
277 raid_end_bio_io(r10_bio
);
282 char b
[BDEVNAME_SIZE
];
283 if (printk_ratelimit())
284 printk(KERN_ERR
"raid10: %s: rescheduling sector %llu\n",
285 bdevname(conf
->mirrors
[dev
].rdev
->bdev
,b
), (unsigned long long)r10_bio
->sector
);
286 reschedule_retry(r10_bio
);
289 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
292 static void raid10_end_write_request(struct bio
*bio
, int error
)
294 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
295 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
297 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
299 for (slot
= 0; slot
< conf
->copies
; slot
++)
300 if (r10_bio
->devs
[slot
].bio
== bio
)
302 dev
= r10_bio
->devs
[slot
].devnum
;
305 * this branch is our 'one mirror IO has finished' event handler:
308 md_error(r10_bio
->mddev
, conf
->mirrors
[dev
].rdev
);
309 /* an I/O failed, we can't clear the bitmap */
310 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
313 * Set R10BIO_Uptodate in our master bio, so that
314 * we will return a good error code for to the higher
315 * levels even if IO on some other mirrored buffer fails.
317 * The 'master' represents the composite IO operation to
318 * user-side. So if something waits for IO, then it will
319 * wait for the 'master' bio.
321 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
323 update_head_pos(slot
, r10_bio
);
327 * Let's see if all mirrored write operations have finished
330 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
331 /* clear the bitmap if all writes complete successfully */
332 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
334 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
336 md_write_end(r10_bio
->mddev
);
337 raid_end_bio_io(r10_bio
);
340 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
345 * RAID10 layout manager
346 * Aswell as the chunksize and raid_disks count, there are two
347 * parameters: near_copies and far_copies.
348 * near_copies * far_copies must be <= raid_disks.
349 * Normally one of these will be 1.
350 * If both are 1, we get raid0.
351 * If near_copies == raid_disks, we get raid1.
353 * Chunks are layed out in raid0 style with near_copies copies of the
354 * first chunk, followed by near_copies copies of the next chunk and
356 * If far_copies > 1, then after 1/far_copies of the array has been assigned
357 * as described above, we start again with a device offset of near_copies.
358 * So we effectively have another copy of the whole array further down all
359 * the drives, but with blocks on different drives.
360 * With this layout, and block is never stored twice on the one device.
362 * raid10_find_phys finds the sector offset of a given virtual sector
363 * on each device that it is on.
365 * raid10_find_virt does the reverse mapping, from a device and a
366 * sector offset to a virtual address
369 static void raid10_find_phys(conf_t
*conf
, r10bio_t
*r10bio
)
379 /* now calculate first sector/dev */
380 chunk
= r10bio
->sector
>> conf
->chunk_shift
;
381 sector
= r10bio
->sector
& conf
->chunk_mask
;
383 chunk
*= conf
->near_copies
;
385 dev
= sector_div(stripe
, conf
->raid_disks
);
386 if (conf
->far_offset
)
387 stripe
*= conf
->far_copies
;
389 sector
+= stripe
<< conf
->chunk_shift
;
391 /* and calculate all the others */
392 for (n
=0; n
< conf
->near_copies
; n
++) {
395 r10bio
->devs
[slot
].addr
= sector
;
396 r10bio
->devs
[slot
].devnum
= d
;
399 for (f
= 1; f
< conf
->far_copies
; f
++) {
400 d
+= conf
->near_copies
;
401 if (d
>= conf
->raid_disks
)
402 d
-= conf
->raid_disks
;
404 r10bio
->devs
[slot
].devnum
= d
;
405 r10bio
->devs
[slot
].addr
= s
;
409 if (dev
>= conf
->raid_disks
) {
411 sector
+= (conf
->chunk_mask
+ 1);
414 BUG_ON(slot
!= conf
->copies
);
417 static sector_t
raid10_find_virt(conf_t
*conf
, sector_t sector
, int dev
)
419 sector_t offset
, chunk
, vchunk
;
421 offset
= sector
& conf
->chunk_mask
;
422 if (conf
->far_offset
) {
424 chunk
= sector
>> conf
->chunk_shift
;
425 fc
= sector_div(chunk
, conf
->far_copies
);
426 dev
-= fc
* conf
->near_copies
;
428 dev
+= conf
->raid_disks
;
430 while (sector
>= conf
->stride
) {
431 sector
-= conf
->stride
;
432 if (dev
< conf
->near_copies
)
433 dev
+= conf
->raid_disks
- conf
->near_copies
;
435 dev
-= conf
->near_copies
;
437 chunk
= sector
>> conf
->chunk_shift
;
439 vchunk
= chunk
* conf
->raid_disks
+ dev
;
440 sector_div(vchunk
, conf
->near_copies
);
441 return (vchunk
<< conf
->chunk_shift
) + offset
;
445 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
447 * @bvm: properties of new bio
448 * @biovec: the request that could be merged to it.
450 * Return amount of bytes we can accept at this offset
451 * If near_copies == raid_disk, there are no striping issues,
452 * but in that case, the function isn't called at all.
454 static int raid10_mergeable_bvec(struct request_queue
*q
,
455 struct bvec_merge_data
*bvm
,
456 struct bio_vec
*biovec
)
458 mddev_t
*mddev
= q
->queuedata
;
459 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
461 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
462 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
464 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
465 if (max
< 0) max
= 0; /* bio_add cannot handle a negative return */
466 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
467 return biovec
->bv_len
;
473 * This routine returns the disk from which the requested read should
474 * be done. There is a per-array 'next expected sequential IO' sector
475 * number - if this matches on the next IO then we use the last disk.
476 * There is also a per-disk 'last know head position' sector that is
477 * maintained from IRQ contexts, both the normal and the resync IO
478 * completion handlers update this position correctly. If there is no
479 * perfect sequential match then we pick the disk whose head is closest.
481 * If there are 2 mirrors in the same 2 devices, performance degrades
482 * because position is mirror, not device based.
484 * The rdev for the device selected will have nr_pending incremented.
488 * FIXME: possibly should rethink readbalancing and do it differently
489 * depending on near_copies / far_copies geometry.
491 static int read_balance(conf_t
*conf
, r10bio_t
*r10_bio
)
493 const unsigned long this_sector
= r10_bio
->sector
;
494 int disk
, slot
, nslot
;
495 const int sectors
= r10_bio
->sectors
;
496 sector_t new_distance
, current_distance
;
499 raid10_find_phys(conf
, r10_bio
);
502 * Check if we can balance. We can balance on the whole
503 * device if no resync is going on (recovery is ok), or below
504 * the resync window. We take the first readable disk when
505 * above the resync window.
507 if (conf
->mddev
->recovery_cp
< MaxSector
508 && (this_sector
+ sectors
>= conf
->next_resync
)) {
509 /* make sure that disk is operational */
511 disk
= r10_bio
->devs
[slot
].devnum
;
513 while ((rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
514 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
515 !test_bit(In_sync
, &rdev
->flags
)) {
517 if (slot
== conf
->copies
) {
522 disk
= r10_bio
->devs
[slot
].devnum
;
528 /* make sure the disk is operational */
530 disk
= r10_bio
->devs
[slot
].devnum
;
531 while ((rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
532 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
533 !test_bit(In_sync
, &rdev
->flags
)) {
535 if (slot
== conf
->copies
) {
539 disk
= r10_bio
->devs
[slot
].devnum
;
543 current_distance
= abs(r10_bio
->devs
[slot
].addr
-
544 conf
->mirrors
[disk
].head_position
);
546 /* Find the disk whose head is closest,
547 * or - for far > 1 - find the closest to partition beginning */
549 for (nslot
= slot
; nslot
< conf
->copies
; nslot
++) {
550 int ndisk
= r10_bio
->devs
[nslot
].devnum
;
553 if ((rdev
=rcu_dereference(conf
->mirrors
[ndisk
].rdev
)) == NULL
||
554 r10_bio
->devs
[nslot
].bio
== IO_BLOCKED
||
555 !test_bit(In_sync
, &rdev
->flags
))
558 /* This optimisation is debatable, and completely destroys
559 * sequential read speed for 'far copies' arrays. So only
560 * keep it for 'near' arrays, and review those later.
562 if (conf
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
)) {
568 /* for far > 1 always use the lowest address */
569 if (conf
->far_copies
> 1)
570 new_distance
= r10_bio
->devs
[nslot
].addr
;
572 new_distance
= abs(r10_bio
->devs
[nslot
].addr
-
573 conf
->mirrors
[ndisk
].head_position
);
574 if (new_distance
< current_distance
) {
575 current_distance
= new_distance
;
582 r10_bio
->read_slot
= slot
;
583 /* conf->next_seq_sect = this_sector + sectors;*/
585 if (disk
>= 0 && (rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
))!= NULL
)
586 atomic_inc(&conf
->mirrors
[disk
].rdev
->nr_pending
);
594 static void unplug_slaves(mddev_t
*mddev
)
596 conf_t
*conf
= mddev_to_conf(mddev
);
600 for (i
=0; i
<mddev
->raid_disks
; i
++) {
601 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
602 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
603 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
605 atomic_inc(&rdev
->nr_pending
);
610 rdev_dec_pending(rdev
, mddev
);
617 static void raid10_unplug(struct request_queue
*q
)
619 mddev_t
*mddev
= q
->queuedata
;
621 unplug_slaves(q
->queuedata
);
622 md_wakeup_thread(mddev
->thread
);
625 static int raid10_congested(void *data
, int bits
)
627 mddev_t
*mddev
= data
;
628 conf_t
*conf
= mddev_to_conf(mddev
);
632 for (i
= 0; i
< mddev
->raid_disks
&& ret
== 0; i
++) {
633 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
634 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
635 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
637 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
644 static int flush_pending_writes(conf_t
*conf
)
646 /* Any writes that have been queued but are awaiting
647 * bitmap updates get flushed here.
648 * We return 1 if any requests were actually submitted.
652 spin_lock_irq(&conf
->device_lock
);
654 if (conf
->pending_bio_list
.head
) {
656 bio
= bio_list_get(&conf
->pending_bio_list
);
657 blk_remove_plug(conf
->mddev
->queue
);
658 spin_unlock_irq(&conf
->device_lock
);
659 /* flush any pending bitmap writes to disk
660 * before proceeding w/ I/O */
661 bitmap_unplug(conf
->mddev
->bitmap
);
663 while (bio
) { /* submit pending writes */
664 struct bio
*next
= bio
->bi_next
;
666 generic_make_request(bio
);
671 spin_unlock_irq(&conf
->device_lock
);
675 * Sometimes we need to suspend IO while we do something else,
676 * either some resync/recovery, or reconfigure the array.
677 * To do this we raise a 'barrier'.
678 * The 'barrier' is a counter that can be raised multiple times
679 * to count how many activities are happening which preclude
681 * We can only raise the barrier if there is no pending IO.
682 * i.e. if nr_pending == 0.
683 * We choose only to raise the barrier if no-one is waiting for the
684 * barrier to go down. This means that as soon as an IO request
685 * is ready, no other operations which require a barrier will start
686 * until the IO request has had a chance.
688 * So: regular IO calls 'wait_barrier'. When that returns there
689 * is no backgroup IO happening, It must arrange to call
690 * allow_barrier when it has finished its IO.
691 * backgroup IO calls must call raise_barrier. Once that returns
692 * there is no normal IO happeing. It must arrange to call
693 * lower_barrier when the particular background IO completes.
696 static void raise_barrier(conf_t
*conf
, int force
)
698 BUG_ON(force
&& !conf
->barrier
);
699 spin_lock_irq(&conf
->resync_lock
);
701 /* Wait until no block IO is waiting (unless 'force') */
702 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
704 raid10_unplug(conf
->mddev
->queue
));
706 /* block any new IO from starting */
709 /* No wait for all pending IO to complete */
710 wait_event_lock_irq(conf
->wait_barrier
,
711 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
713 raid10_unplug(conf
->mddev
->queue
));
715 spin_unlock_irq(&conf
->resync_lock
);
718 static void lower_barrier(conf_t
*conf
)
721 spin_lock_irqsave(&conf
->resync_lock
, flags
);
723 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
724 wake_up(&conf
->wait_barrier
);
727 static void wait_barrier(conf_t
*conf
)
729 spin_lock_irq(&conf
->resync_lock
);
732 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
734 raid10_unplug(conf
->mddev
->queue
));
738 spin_unlock_irq(&conf
->resync_lock
);
741 static void allow_barrier(conf_t
*conf
)
744 spin_lock_irqsave(&conf
->resync_lock
, flags
);
746 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
747 wake_up(&conf
->wait_barrier
);
750 static void freeze_array(conf_t
*conf
)
752 /* stop syncio and normal IO and wait for everything to
754 * We increment barrier and nr_waiting, and then
755 * wait until nr_pending match nr_queued+1
756 * This is called in the context of one normal IO request
757 * that has failed. Thus any sync request that might be pending
758 * will be blocked by nr_pending, and we need to wait for
759 * pending IO requests to complete or be queued for re-try.
760 * Thus the number queued (nr_queued) plus this request (1)
761 * must match the number of pending IOs (nr_pending) before
764 spin_lock_irq(&conf
->resync_lock
);
767 wait_event_lock_irq(conf
->wait_barrier
,
768 conf
->nr_pending
== conf
->nr_queued
+1,
770 ({ flush_pending_writes(conf
);
771 raid10_unplug(conf
->mddev
->queue
); }));
772 spin_unlock_irq(&conf
->resync_lock
);
775 static void unfreeze_array(conf_t
*conf
)
777 /* reverse the effect of the freeze */
778 spin_lock_irq(&conf
->resync_lock
);
781 wake_up(&conf
->wait_barrier
);
782 spin_unlock_irq(&conf
->resync_lock
);
785 static int make_request(struct request_queue
*q
, struct bio
* bio
)
787 mddev_t
*mddev
= q
->queuedata
;
788 conf_t
*conf
= mddev_to_conf(mddev
);
789 mirror_info_t
*mirror
;
791 struct bio
*read_bio
;
794 int chunk_sects
= conf
->chunk_mask
+ 1;
795 const int rw
= bio_data_dir(bio
);
796 const int do_sync
= bio_sync(bio
);
799 mdk_rdev_t
*blocked_rdev
;
801 if (unlikely(bio_barrier(bio
))) {
802 bio_endio(bio
, -EOPNOTSUPP
);
806 /* If this request crosses a chunk boundary, we need to
807 * split it. This will only happen for 1 PAGE (or less) requests.
809 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
811 conf
->near_copies
< conf
->raid_disks
)) {
813 /* Sanity check -- queue functions should prevent this happening */
814 if (bio
->bi_vcnt
!= 1 ||
817 /* This is a one page bio that upper layers
818 * refuse to split for us, so we need to split it.
820 bp
= bio_split(bio
, bio_split_pool
,
821 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
822 if (make_request(q
, &bp
->bio1
))
823 generic_make_request(&bp
->bio1
);
824 if (make_request(q
, &bp
->bio2
))
825 generic_make_request(&bp
->bio2
);
827 bio_pair_release(bp
);
830 printk("raid10_make_request bug: can't convert block across chunks"
831 " or bigger than %dk %llu %d\n", chunk_sects
/2,
832 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
838 md_write_start(mddev
, bio
);
841 * Register the new request and wait if the reconstruction
842 * thread has put up a bar for new requests.
843 * Continue immediately if no resync is active currently.
847 cpu
= disk_stat_lock();
848 disk_stat_inc(cpu
, mddev
->gendisk
, ios
[rw
]);
849 disk_stat_add(cpu
, mddev
->gendisk
, sectors
[rw
], bio_sectors(bio
));
852 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
854 r10_bio
->master_bio
= bio
;
855 r10_bio
->sectors
= bio
->bi_size
>> 9;
857 r10_bio
->mddev
= mddev
;
858 r10_bio
->sector
= bio
->bi_sector
;
863 * read balancing logic:
865 int disk
= read_balance(conf
, r10_bio
);
866 int slot
= r10_bio
->read_slot
;
868 raid_end_bio_io(r10_bio
);
871 mirror
= conf
->mirrors
+ disk
;
873 read_bio
= bio_clone(bio
, GFP_NOIO
);
875 r10_bio
->devs
[slot
].bio
= read_bio
;
877 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
878 mirror
->rdev
->data_offset
;
879 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
880 read_bio
->bi_end_io
= raid10_end_read_request
;
881 read_bio
->bi_rw
= READ
| do_sync
;
882 read_bio
->bi_private
= r10_bio
;
884 generic_make_request(read_bio
);
891 /* first select target devices under rcu_lock and
892 * inc refcount on their rdev. Record them by setting
895 raid10_find_phys(conf
, r10_bio
);
899 for (i
= 0; i
< conf
->copies
; i
++) {
900 int d
= r10_bio
->devs
[i
].devnum
;
901 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
902 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
903 atomic_inc(&rdev
->nr_pending
);
907 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
908 atomic_inc(&rdev
->nr_pending
);
909 r10_bio
->devs
[i
].bio
= bio
;
911 r10_bio
->devs
[i
].bio
= NULL
;
912 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
917 if (unlikely(blocked_rdev
)) {
918 /* Have to wait for this device to get unblocked, then retry */
922 for (j
= 0; j
< i
; j
++)
923 if (r10_bio
->devs
[j
].bio
) {
924 d
= r10_bio
->devs
[j
].devnum
;
925 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
928 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
933 atomic_set(&r10_bio
->remaining
, 0);
936 for (i
= 0; i
< conf
->copies
; i
++) {
938 int d
= r10_bio
->devs
[i
].devnum
;
939 if (!r10_bio
->devs
[i
].bio
)
942 mbio
= bio_clone(bio
, GFP_NOIO
);
943 r10_bio
->devs
[i
].bio
= mbio
;
945 mbio
->bi_sector
= r10_bio
->devs
[i
].addr
+
946 conf
->mirrors
[d
].rdev
->data_offset
;
947 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
948 mbio
->bi_end_io
= raid10_end_write_request
;
949 mbio
->bi_rw
= WRITE
| do_sync
;
950 mbio
->bi_private
= r10_bio
;
952 atomic_inc(&r10_bio
->remaining
);
953 bio_list_add(&bl
, mbio
);
956 if (unlikely(!atomic_read(&r10_bio
->remaining
))) {
957 /* the array is dead */
959 raid_end_bio_io(r10_bio
);
963 bitmap_startwrite(mddev
->bitmap
, bio
->bi_sector
, r10_bio
->sectors
, 0);
964 spin_lock_irqsave(&conf
->device_lock
, flags
);
965 bio_list_merge(&conf
->pending_bio_list
, &bl
);
966 blk_plug_device(mddev
->queue
);
967 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
969 /* In case raid10d snuck in to freeze_array */
970 wake_up(&conf
->wait_barrier
);
973 md_wakeup_thread(mddev
->thread
);
978 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
980 conf_t
*conf
= mddev_to_conf(mddev
);
983 if (conf
->near_copies
< conf
->raid_disks
)
984 seq_printf(seq
, " %dK chunks", mddev
->chunk_size
/1024);
985 if (conf
->near_copies
> 1)
986 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
987 if (conf
->far_copies
> 1) {
988 if (conf
->far_offset
)
989 seq_printf(seq
, " %d offset-copies", conf
->far_copies
);
991 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
993 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
994 conf
->raid_disks
- mddev
->degraded
);
995 for (i
= 0; i
< conf
->raid_disks
; i
++)
996 seq_printf(seq
, "%s",
997 conf
->mirrors
[i
].rdev
&&
998 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
999 seq_printf(seq
, "]");
1002 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1004 char b
[BDEVNAME_SIZE
];
1005 conf_t
*conf
= mddev_to_conf(mddev
);
1008 * If it is not operational, then we have already marked it as dead
1009 * else if it is the last working disks, ignore the error, let the
1010 * next level up know.
1011 * else mark the drive as failed
1013 if (test_bit(In_sync
, &rdev
->flags
)
1014 && conf
->raid_disks
-mddev
->degraded
== 1)
1016 * Don't fail the drive, just return an IO error.
1017 * The test should really be more sophisticated than
1018 * "working_disks == 1", but it isn't critical, and
1019 * can wait until we do more sophisticated "is the drive
1020 * really dead" tests...
1023 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1024 unsigned long flags
;
1025 spin_lock_irqsave(&conf
->device_lock
, flags
);
1027 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1029 * if recovery is running, make sure it aborts.
1031 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1033 set_bit(Faulty
, &rdev
->flags
);
1034 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1035 printk(KERN_ALERT
"raid10: Disk failure on %s, disabling device.\n"
1036 "raid10: Operation continuing on %d devices.\n",
1037 bdevname(rdev
->bdev
,b
), conf
->raid_disks
- mddev
->degraded
);
1040 static void print_conf(conf_t
*conf
)
1045 printk("RAID10 conf printout:\n");
1047 printk("(!conf)\n");
1050 printk(" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1053 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1054 char b
[BDEVNAME_SIZE
];
1055 tmp
= conf
->mirrors
+ i
;
1057 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1058 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1059 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1060 bdevname(tmp
->rdev
->bdev
,b
));
1064 static void close_sync(conf_t
*conf
)
1067 allow_barrier(conf
);
1069 mempool_destroy(conf
->r10buf_pool
);
1070 conf
->r10buf_pool
= NULL
;
1073 /* check if there are enough drives for
1074 * every block to appear on atleast one
1076 static int enough(conf_t
*conf
)
1081 int n
= conf
->copies
;
1084 if (conf
->mirrors
[first
].rdev
)
1086 first
= (first
+1) % conf
->raid_disks
;
1090 } while (first
!= 0);
1094 static int raid10_spare_active(mddev_t
*mddev
)
1097 conf_t
*conf
= mddev
->private;
1101 * Find all non-in_sync disks within the RAID10 configuration
1102 * and mark them in_sync
1104 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1105 tmp
= conf
->mirrors
+ i
;
1107 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1108 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1109 unsigned long flags
;
1110 spin_lock_irqsave(&conf
->device_lock
, flags
);
1112 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1121 static int raid10_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1123 conf_t
*conf
= mddev
->private;
1128 int last
= mddev
->raid_disks
- 1;
1130 if (mddev
->recovery_cp
< MaxSector
)
1131 /* only hot-add to in-sync arrays, as recovery is
1132 * very different from resync
1138 if (rdev
->raid_disk
)
1139 first
= last
= rdev
->raid_disk
;
1141 if (rdev
->saved_raid_disk
>= 0 &&
1142 rdev
->saved_raid_disk
>= first
&&
1143 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1144 mirror
= rdev
->saved_raid_disk
;
1147 for ( ; mirror
<= last
; mirror
++)
1148 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1150 blk_queue_stack_limits(mddev
->queue
,
1151 rdev
->bdev
->bd_disk
->queue
);
1152 /* as we don't honour merge_bvec_fn, we must never risk
1153 * violating it, so limit ->max_sector to one PAGE, as
1154 * a one page request is never in violation.
1156 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
1157 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
1158 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
1160 p
->head_position
= 0;
1161 rdev
->raid_disk
= mirror
;
1163 if (rdev
->saved_raid_disk
!= mirror
)
1165 rcu_assign_pointer(p
->rdev
, rdev
);
1173 static int raid10_remove_disk(mddev_t
*mddev
, int number
)
1175 conf_t
*conf
= mddev
->private;
1178 mirror_info_t
*p
= conf
->mirrors
+ number
;
1183 if (test_bit(In_sync
, &rdev
->flags
) ||
1184 atomic_read(&rdev
->nr_pending
)) {
1188 /* Only remove faulty devices in recovery
1191 if (!test_bit(Faulty
, &rdev
->flags
) &&
1198 if (atomic_read(&rdev
->nr_pending
)) {
1199 /* lost the race, try later */
1211 static void end_sync_read(struct bio
*bio
, int error
)
1213 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1214 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
1217 for (i
=0; i
<conf
->copies
; i
++)
1218 if (r10_bio
->devs
[i
].bio
== bio
)
1220 BUG_ON(i
== conf
->copies
);
1221 update_head_pos(i
, r10_bio
);
1222 d
= r10_bio
->devs
[i
].devnum
;
1224 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1225 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1227 atomic_add(r10_bio
->sectors
,
1228 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1229 if (!test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
1230 md_error(r10_bio
->mddev
,
1231 conf
->mirrors
[d
].rdev
);
1234 /* for reconstruct, we always reschedule after a read.
1235 * for resync, only after all reads
1237 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1238 atomic_dec_and_test(&r10_bio
->remaining
)) {
1239 /* we have read all the blocks,
1240 * do the comparison in process context in raid10d
1242 reschedule_retry(r10_bio
);
1244 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1247 static void end_sync_write(struct bio
*bio
, int error
)
1249 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1250 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1251 mddev_t
*mddev
= r10_bio
->mddev
;
1252 conf_t
*conf
= mddev_to_conf(mddev
);
1255 for (i
= 0; i
< conf
->copies
; i
++)
1256 if (r10_bio
->devs
[i
].bio
== bio
)
1258 d
= r10_bio
->devs
[i
].devnum
;
1261 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1263 update_head_pos(i
, r10_bio
);
1265 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1266 if (r10_bio
->master_bio
== NULL
) {
1267 /* the primary of several recovery bios */
1268 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1272 r10bio_t
*r10_bio2
= (r10bio_t
*)r10_bio
->master_bio
;
1277 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1281 * Note: sync and recover and handled very differently for raid10
1282 * This code is for resync.
1283 * For resync, we read through virtual addresses and read all blocks.
1284 * If there is any error, we schedule a write. The lowest numbered
1285 * drive is authoritative.
1286 * However requests come for physical address, so we need to map.
1287 * For every physical address there are raid_disks/copies virtual addresses,
1288 * which is always are least one, but is not necessarly an integer.
1289 * This means that a physical address can span multiple chunks, so we may
1290 * have to submit multiple io requests for a single sync request.
1293 * We check if all blocks are in-sync and only write to blocks that
1296 static void sync_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1298 conf_t
*conf
= mddev_to_conf(mddev
);
1300 struct bio
*tbio
, *fbio
;
1302 atomic_set(&r10_bio
->remaining
, 1);
1304 /* find the first device with a block */
1305 for (i
=0; i
<conf
->copies
; i
++)
1306 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1309 if (i
== conf
->copies
)
1313 fbio
= r10_bio
->devs
[i
].bio
;
1315 /* now find blocks with errors */
1316 for (i
=0 ; i
< conf
->copies
; i
++) {
1318 int vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1320 tbio
= r10_bio
->devs
[i
].bio
;
1322 if (tbio
->bi_end_io
!= end_sync_read
)
1326 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1327 /* We know that the bi_io_vec layout is the same for
1328 * both 'first' and 'i', so we just compare them.
1329 * All vec entries are PAGE_SIZE;
1331 for (j
= 0; j
< vcnt
; j
++)
1332 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1333 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1338 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1340 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1341 /* Don't fix anything. */
1343 /* Ok, we need to write this bio
1344 * First we need to fixup bv_offset, bv_len and
1345 * bi_vecs, as the read request might have corrupted these
1347 tbio
->bi_vcnt
= vcnt
;
1348 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1350 tbio
->bi_phys_segments
= 0;
1351 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1352 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1353 tbio
->bi_next
= NULL
;
1354 tbio
->bi_rw
= WRITE
;
1355 tbio
->bi_private
= r10_bio
;
1356 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1358 for (j
=0; j
< vcnt
; j
++) {
1359 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1360 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1362 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1363 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1366 tbio
->bi_end_io
= end_sync_write
;
1368 d
= r10_bio
->devs
[i
].devnum
;
1369 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1370 atomic_inc(&r10_bio
->remaining
);
1371 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1373 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1374 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1375 generic_make_request(tbio
);
1379 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1380 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1386 * Now for the recovery code.
1387 * Recovery happens across physical sectors.
1388 * We recover all non-is_sync drives by finding the virtual address of
1389 * each, and then choose a working drive that also has that virt address.
1390 * There is a separate r10_bio for each non-in_sync drive.
1391 * Only the first two slots are in use. The first for reading,
1392 * The second for writing.
1396 static void recovery_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1398 conf_t
*conf
= mddev_to_conf(mddev
);
1400 struct bio
*bio
, *wbio
;
1403 /* move the pages across to the second bio
1404 * and submit the write request
1406 bio
= r10_bio
->devs
[0].bio
;
1407 wbio
= r10_bio
->devs
[1].bio
;
1408 for (i
=0; i
< wbio
->bi_vcnt
; i
++) {
1409 struct page
*p
= bio
->bi_io_vec
[i
].bv_page
;
1410 bio
->bi_io_vec
[i
].bv_page
= wbio
->bi_io_vec
[i
].bv_page
;
1411 wbio
->bi_io_vec
[i
].bv_page
= p
;
1413 d
= r10_bio
->devs
[1].devnum
;
1415 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1416 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1417 if (test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
1418 generic_make_request(wbio
);
1420 bio_endio(wbio
, -EIO
);
1425 * This is a kernel thread which:
1427 * 1. Retries failed read operations on working mirrors.
1428 * 2. Updates the raid superblock when problems encounter.
1429 * 3. Performs writes following reads for array synchronising.
1432 static void fix_read_error(conf_t
*conf
, mddev_t
*mddev
, r10bio_t
*r10_bio
)
1434 int sect
= 0; /* Offset from r10_bio->sector */
1435 int sectors
= r10_bio
->sectors
;
1439 int sl
= r10_bio
->read_slot
;
1443 if (s
> (PAGE_SIZE
>>9))
1448 int d
= r10_bio
->devs
[sl
].devnum
;
1449 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1451 test_bit(In_sync
, &rdev
->flags
)) {
1452 atomic_inc(&rdev
->nr_pending
);
1454 success
= sync_page_io(rdev
->bdev
,
1455 r10_bio
->devs
[sl
].addr
+
1456 sect
+ rdev
->data_offset
,
1458 conf
->tmppage
, READ
);
1459 rdev_dec_pending(rdev
, mddev
);
1465 if (sl
== conf
->copies
)
1467 } while (!success
&& sl
!= r10_bio
->read_slot
);
1471 /* Cannot read from anywhere -- bye bye array */
1472 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1473 md_error(mddev
, conf
->mirrors
[dn
].rdev
);
1478 /* write it back and re-read */
1480 while (sl
!= r10_bio
->read_slot
) {
1485 d
= r10_bio
->devs
[sl
].devnum
;
1486 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1488 test_bit(In_sync
, &rdev
->flags
)) {
1489 atomic_inc(&rdev
->nr_pending
);
1491 atomic_add(s
, &rdev
->corrected_errors
);
1492 if (sync_page_io(rdev
->bdev
,
1493 r10_bio
->devs
[sl
].addr
+
1494 sect
+ rdev
->data_offset
,
1495 s
<<9, conf
->tmppage
, WRITE
)
1497 /* Well, this device is dead */
1498 md_error(mddev
, rdev
);
1499 rdev_dec_pending(rdev
, mddev
);
1504 while (sl
!= r10_bio
->read_slot
) {
1509 d
= r10_bio
->devs
[sl
].devnum
;
1510 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1512 test_bit(In_sync
, &rdev
->flags
)) {
1513 char b
[BDEVNAME_SIZE
];
1514 atomic_inc(&rdev
->nr_pending
);
1516 if (sync_page_io(rdev
->bdev
,
1517 r10_bio
->devs
[sl
].addr
+
1518 sect
+ rdev
->data_offset
,
1519 s
<<9, conf
->tmppage
, READ
) == 0)
1520 /* Well, this device is dead */
1521 md_error(mddev
, rdev
);
1524 "raid10:%s: read error corrected"
1525 " (%d sectors at %llu on %s)\n",
1527 (unsigned long long)(sect
+
1529 bdevname(rdev
->bdev
, b
));
1531 rdev_dec_pending(rdev
, mddev
);
1542 static void raid10d(mddev_t
*mddev
)
1546 unsigned long flags
;
1547 conf_t
*conf
= mddev_to_conf(mddev
);
1548 struct list_head
*head
= &conf
->retry_list
;
1552 md_check_recovery(mddev
);
1555 char b
[BDEVNAME_SIZE
];
1557 unplug
+= flush_pending_writes(conf
);
1559 spin_lock_irqsave(&conf
->device_lock
, flags
);
1560 if (list_empty(head
)) {
1561 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1564 r10_bio
= list_entry(head
->prev
, r10bio_t
, retry_list
);
1565 list_del(head
->prev
);
1567 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1569 mddev
= r10_bio
->mddev
;
1570 conf
= mddev_to_conf(mddev
);
1571 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
)) {
1572 sync_request_write(mddev
, r10_bio
);
1574 } else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
1575 recovery_request_write(mddev
, r10_bio
);
1579 /* we got a read error. Maybe the drive is bad. Maybe just
1580 * the block and we can fix it.
1581 * We freeze all other IO, and try reading the block from
1582 * other devices. When we find one, we re-write
1583 * and check it that fixes the read error.
1584 * This is all done synchronously while the array is
1587 if (mddev
->ro
== 0) {
1589 fix_read_error(conf
, mddev
, r10_bio
);
1590 unfreeze_array(conf
);
1593 bio
= r10_bio
->devs
[r10_bio
->read_slot
].bio
;
1594 r10_bio
->devs
[r10_bio
->read_slot
].bio
=
1595 mddev
->ro
? IO_BLOCKED
: NULL
;
1596 mirror
= read_balance(conf
, r10_bio
);
1598 printk(KERN_ALERT
"raid10: %s: unrecoverable I/O"
1599 " read error for block %llu\n",
1600 bdevname(bio
->bi_bdev
,b
),
1601 (unsigned long long)r10_bio
->sector
);
1602 raid_end_bio_io(r10_bio
);
1605 const int do_sync
= bio_sync(r10_bio
->master_bio
);
1607 rdev
= conf
->mirrors
[mirror
].rdev
;
1608 if (printk_ratelimit())
1609 printk(KERN_ERR
"raid10: %s: redirecting sector %llu to"
1610 " another mirror\n",
1611 bdevname(rdev
->bdev
,b
),
1612 (unsigned long long)r10_bio
->sector
);
1613 bio
= bio_clone(r10_bio
->master_bio
, GFP_NOIO
);
1614 r10_bio
->devs
[r10_bio
->read_slot
].bio
= bio
;
1615 bio
->bi_sector
= r10_bio
->devs
[r10_bio
->read_slot
].addr
1616 + rdev
->data_offset
;
1617 bio
->bi_bdev
= rdev
->bdev
;
1618 bio
->bi_rw
= READ
| do_sync
;
1619 bio
->bi_private
= r10_bio
;
1620 bio
->bi_end_io
= raid10_end_read_request
;
1622 generic_make_request(bio
);
1627 unplug_slaves(mddev
);
1631 static int init_resync(conf_t
*conf
)
1635 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1636 BUG_ON(conf
->r10buf_pool
);
1637 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
1638 if (!conf
->r10buf_pool
)
1640 conf
->next_resync
= 0;
1645 * perform a "sync" on one "block"
1647 * We need to make sure that no normal I/O request - particularly write
1648 * requests - conflict with active sync requests.
1650 * This is achieved by tracking pending requests and a 'barrier' concept
1651 * that can be installed to exclude normal IO requests.
1653 * Resync and recovery are handled very differently.
1654 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1656 * For resync, we iterate over virtual addresses, read all copies,
1657 * and update if there are differences. If only one copy is live,
1659 * For recovery, we iterate over physical addresses, read a good
1660 * value for each non-in_sync drive, and over-write.
1662 * So, for recovery we may have several outstanding complex requests for a
1663 * given address, one for each out-of-sync device. We model this by allocating
1664 * a number of r10_bio structures, one for each out-of-sync device.
1665 * As we setup these structures, we collect all bio's together into a list
1666 * which we then process collectively to add pages, and then process again
1667 * to pass to generic_make_request.
1669 * The r10_bio structures are linked using a borrowed master_bio pointer.
1670 * This link is counted in ->remaining. When the r10_bio that points to NULL
1671 * has its remaining count decremented to 0, the whole complex operation
1676 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
1678 conf_t
*conf
= mddev_to_conf(mddev
);
1680 struct bio
*biolist
= NULL
, *bio
;
1681 sector_t max_sector
, nr_sectors
;
1687 sector_t sectors_skipped
= 0;
1688 int chunks_skipped
= 0;
1690 if (!conf
->r10buf_pool
)
1691 if (init_resync(conf
))
1695 max_sector
= mddev
->size
<< 1;
1696 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1697 max_sector
= mddev
->resync_max_sectors
;
1698 if (sector_nr
>= max_sector
) {
1699 /* If we aborted, we need to abort the
1700 * sync on the 'current' bitmap chucks (there can
1701 * be several when recovering multiple devices).
1702 * as we may have started syncing it but not finished.
1703 * We can find the current address in
1704 * mddev->curr_resync, but for recovery,
1705 * we need to convert that to several
1706 * virtual addresses.
1708 if (mddev
->curr_resync
< max_sector
) { /* aborted */
1709 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1710 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
1712 else for (i
=0; i
<conf
->raid_disks
; i
++) {
1714 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
1715 bitmap_end_sync(mddev
->bitmap
, sect
,
1718 } else /* completed sync */
1721 bitmap_close_sync(mddev
->bitmap
);
1724 return sectors_skipped
;
1726 if (chunks_skipped
>= conf
->raid_disks
) {
1727 /* if there has been nothing to do on any drive,
1728 * then there is nothing to do at all..
1731 return (max_sector
- sector_nr
) + sectors_skipped
;
1734 if (max_sector
> mddev
->resync_max
)
1735 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
1737 /* make sure whole request will fit in a chunk - if chunks
1740 if (conf
->near_copies
< conf
->raid_disks
&&
1741 max_sector
> (sector_nr
| conf
->chunk_mask
))
1742 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
1744 * If there is non-resync activity waiting for us then
1745 * put in a delay to throttle resync.
1747 if (!go_faster
&& conf
->nr_waiting
)
1748 msleep_interruptible(1000);
1750 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
1752 /* Again, very different code for resync and recovery.
1753 * Both must result in an r10bio with a list of bios that
1754 * have bi_end_io, bi_sector, bi_bdev set,
1755 * and bi_private set to the r10bio.
1756 * For recovery, we may actually create several r10bios
1757 * with 2 bios in each, that correspond to the bios in the main one.
1758 * In this case, the subordinate r10bios link back through a
1759 * borrowed master_bio pointer, and the counter in the master
1760 * includes a ref from each subordinate.
1762 /* First, we decide what to do and set ->bi_end_io
1763 * To end_sync_read if we want to read, and
1764 * end_sync_write if we will want to write.
1767 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
1768 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
1769 /* recovery... the complicated one */
1773 for (i
=0 ; i
<conf
->raid_disks
; i
++)
1774 if (conf
->mirrors
[i
].rdev
&&
1775 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
)) {
1776 int still_degraded
= 0;
1777 /* want to reconstruct this device */
1778 r10bio_t
*rb2
= r10_bio
;
1779 sector_t sect
= raid10_find_virt(conf
, sector_nr
, i
);
1781 /* Unless we are doing a full sync, we only need
1782 * to recover the block if it is set in the bitmap
1784 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1786 if (sync_blocks
< max_sync
)
1787 max_sync
= sync_blocks
;
1790 /* yep, skip the sync_blocks here, but don't assume
1791 * that there will never be anything to do here
1793 chunks_skipped
= -1;
1797 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1798 raise_barrier(conf
, rb2
!= NULL
);
1799 atomic_set(&r10_bio
->remaining
, 0);
1801 r10_bio
->master_bio
= (struct bio
*)rb2
;
1803 atomic_inc(&rb2
->remaining
);
1804 r10_bio
->mddev
= mddev
;
1805 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
1806 r10_bio
->sector
= sect
;
1808 raid10_find_phys(conf
, r10_bio
);
1809 /* Need to check if this section will still be
1812 for (j
=0; j
<conf
->copies
;j
++) {
1813 int d
= r10_bio
->devs
[j
].devnum
;
1814 if (conf
->mirrors
[d
].rdev
== NULL
||
1815 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
)) {
1820 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1821 &sync_blocks
, still_degraded
);
1823 for (j
=0; j
<conf
->copies
;j
++) {
1824 int d
= r10_bio
->devs
[j
].devnum
;
1825 if (conf
->mirrors
[d
].rdev
&&
1826 test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
)) {
1827 /* This is where we read from */
1828 bio
= r10_bio
->devs
[0].bio
;
1829 bio
->bi_next
= biolist
;
1831 bio
->bi_private
= r10_bio
;
1832 bio
->bi_end_io
= end_sync_read
;
1834 bio
->bi_sector
= r10_bio
->devs
[j
].addr
+
1835 conf
->mirrors
[d
].rdev
->data_offset
;
1836 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1837 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1838 atomic_inc(&r10_bio
->remaining
);
1839 /* and we write to 'i' */
1841 for (k
=0; k
<conf
->copies
; k
++)
1842 if (r10_bio
->devs
[k
].devnum
== i
)
1844 BUG_ON(k
== conf
->copies
);
1845 bio
= r10_bio
->devs
[1].bio
;
1846 bio
->bi_next
= biolist
;
1848 bio
->bi_private
= r10_bio
;
1849 bio
->bi_end_io
= end_sync_write
;
1851 bio
->bi_sector
= r10_bio
->devs
[k
].addr
+
1852 conf
->mirrors
[i
].rdev
->data_offset
;
1853 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1855 r10_bio
->devs
[0].devnum
= d
;
1856 r10_bio
->devs
[1].devnum
= i
;
1861 if (j
== conf
->copies
) {
1862 /* Cannot recover, so abort the recovery */
1865 atomic_dec(&rb2
->remaining
);
1867 if (!test_and_set_bit(MD_RECOVERY_INTR
,
1869 printk(KERN_INFO
"raid10: %s: insufficient working devices for recovery.\n",
1874 if (biolist
== NULL
) {
1876 r10bio_t
*rb2
= r10_bio
;
1877 r10_bio
= (r10bio_t
*) rb2
->master_bio
;
1878 rb2
->master_bio
= NULL
;
1884 /* resync. Schedule a read for every block at this virt offset */
1887 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
1888 &sync_blocks
, mddev
->degraded
) &&
1889 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1890 /* We can skip this block */
1892 return sync_blocks
+ sectors_skipped
;
1894 if (sync_blocks
< max_sync
)
1895 max_sync
= sync_blocks
;
1896 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1898 r10_bio
->mddev
= mddev
;
1899 atomic_set(&r10_bio
->remaining
, 0);
1900 raise_barrier(conf
, 0);
1901 conf
->next_resync
= sector_nr
;
1903 r10_bio
->master_bio
= NULL
;
1904 r10_bio
->sector
= sector_nr
;
1905 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
1906 raid10_find_phys(conf
, r10_bio
);
1907 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
1909 for (i
=0; i
<conf
->copies
; i
++) {
1910 int d
= r10_bio
->devs
[i
].devnum
;
1911 bio
= r10_bio
->devs
[i
].bio
;
1912 bio
->bi_end_io
= NULL
;
1913 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1914 if (conf
->mirrors
[d
].rdev
== NULL
||
1915 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
1917 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1918 atomic_inc(&r10_bio
->remaining
);
1919 bio
->bi_next
= biolist
;
1921 bio
->bi_private
= r10_bio
;
1922 bio
->bi_end_io
= end_sync_read
;
1924 bio
->bi_sector
= r10_bio
->devs
[i
].addr
+
1925 conf
->mirrors
[d
].rdev
->data_offset
;
1926 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1931 for (i
=0; i
<conf
->copies
; i
++) {
1932 int d
= r10_bio
->devs
[i
].devnum
;
1933 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
1934 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1942 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1944 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1946 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
1949 bio
->bi_phys_segments
= 0;
1954 if (sector_nr
+ max_sync
< max_sector
)
1955 max_sector
= sector_nr
+ max_sync
;
1958 int len
= PAGE_SIZE
;
1960 if (sector_nr
+ (len
>>9) > max_sector
)
1961 len
= (max_sector
- sector_nr
) << 9;
1964 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1965 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
1966 if (bio_add_page(bio
, page
, len
, 0) == 0) {
1969 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
1970 for (bio2
= biolist
; bio2
&& bio2
!= bio
; bio2
= bio2
->bi_next
) {
1971 /* remove last page from this bio */
1973 bio2
->bi_size
-= len
;
1974 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
1980 nr_sectors
+= len
>>9;
1981 sector_nr
+= len
>>9;
1982 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
1984 r10_bio
->sectors
= nr_sectors
;
1988 biolist
= biolist
->bi_next
;
1990 bio
->bi_next
= NULL
;
1991 r10_bio
= bio
->bi_private
;
1992 r10_bio
->sectors
= nr_sectors
;
1994 if (bio
->bi_end_io
== end_sync_read
) {
1995 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1996 generic_make_request(bio
);
2000 if (sectors_skipped
)
2001 /* pretend they weren't skipped, it makes
2002 * no important difference in this case
2004 md_done_sync(mddev
, sectors_skipped
, 1);
2006 return sectors_skipped
+ nr_sectors
;
2008 /* There is nowhere to write, so all non-sync
2009 * drives must be failed, so try the next chunk...
2012 sector_t sec
= max_sector
- sector_nr
;
2013 sectors_skipped
+= sec
;
2015 sector_nr
= max_sector
;
2020 static int run(mddev_t
*mddev
)
2024 mirror_info_t
*disk
;
2026 struct list_head
*tmp
;
2028 sector_t stride
, size
;
2030 if (mddev
->chunk_size
== 0) {
2031 printk(KERN_ERR
"md/raid10: non-zero chunk size required.\n");
2035 nc
= mddev
->layout
& 255;
2036 fc
= (mddev
->layout
>> 8) & 255;
2037 fo
= mddev
->layout
& (1<<16);
2038 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
2039 (mddev
->layout
>> 17)) {
2040 printk(KERN_ERR
"raid10: %s: unsupported raid10 layout: 0x%8x\n",
2041 mdname(mddev
), mddev
->layout
);
2045 * copy the already verified devices into our private RAID10
2046 * bookkeeping area. [whatever we allocate in run(),
2047 * should be freed in stop()]
2049 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
2050 mddev
->private = conf
;
2052 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
2056 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
2058 if (!conf
->mirrors
) {
2059 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
2064 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2068 conf
->mddev
= mddev
;
2069 conf
->raid_disks
= mddev
->raid_disks
;
2070 conf
->near_copies
= nc
;
2071 conf
->far_copies
= fc
;
2072 conf
->copies
= nc
*fc
;
2073 conf
->far_offset
= fo
;
2074 conf
->chunk_mask
= (sector_t
)(mddev
->chunk_size
>>9)-1;
2075 conf
->chunk_shift
= ffz(~mddev
->chunk_size
) - 9;
2076 size
= mddev
->size
>> (conf
->chunk_shift
-1);
2077 sector_div(size
, fc
);
2078 size
= size
* conf
->raid_disks
;
2079 sector_div(size
, nc
);
2080 /* 'size' is now the number of chunks in the array */
2081 /* calculate "used chunks per device" in 'stride' */
2082 stride
= size
* conf
->copies
;
2084 /* We need to round up when dividing by raid_disks to
2085 * get the stride size.
2087 stride
+= conf
->raid_disks
- 1;
2088 sector_div(stride
, conf
->raid_disks
);
2089 mddev
->size
= stride
<< (conf
->chunk_shift
-1);
2094 sector_div(stride
, fc
);
2095 conf
->stride
= stride
<< conf
->chunk_shift
;
2097 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
2098 r10bio_pool_free
, conf
);
2099 if (!conf
->r10bio_pool
) {
2100 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
2105 spin_lock_init(&conf
->device_lock
);
2106 mddev
->queue
->queue_lock
= &conf
->device_lock
;
2108 rdev_for_each(rdev
, tmp
, mddev
) {
2109 disk_idx
= rdev
->raid_disk
;
2110 if (disk_idx
>= mddev
->raid_disks
2113 disk
= conf
->mirrors
+ disk_idx
;
2117 blk_queue_stack_limits(mddev
->queue
,
2118 rdev
->bdev
->bd_disk
->queue
);
2119 /* as we don't honour merge_bvec_fn, we must never risk
2120 * violating it, so limit ->max_sector to one PAGE, as
2121 * a one page request is never in violation.
2123 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
2124 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
2125 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
2127 disk
->head_position
= 0;
2129 INIT_LIST_HEAD(&conf
->retry_list
);
2131 spin_lock_init(&conf
->resync_lock
);
2132 init_waitqueue_head(&conf
->wait_barrier
);
2134 /* need to check that every block has at least one working mirror */
2135 if (!enough(conf
)) {
2136 printk(KERN_ERR
"raid10: not enough operational mirrors for %s\n",
2141 mddev
->degraded
= 0;
2142 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2144 disk
= conf
->mirrors
+ i
;
2147 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2148 disk
->head_position
= 0;
2156 mddev
->thread
= md_register_thread(raid10d
, mddev
, "%s_raid10");
2157 if (!mddev
->thread
) {
2159 "raid10: couldn't allocate thread for %s\n",
2165 "raid10: raid set %s active with %d out of %d devices\n",
2166 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2169 * Ok, everything is just fine now
2171 mddev
->array_sectors
= size
<< conf
->chunk_shift
;
2172 mddev
->resync_max_sectors
= size
<< conf
->chunk_shift
;
2174 mddev
->queue
->unplug_fn
= raid10_unplug
;
2175 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
2176 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2178 /* Calculate max read-ahead size.
2179 * We need to readahead at least twice a whole stripe....
2183 int stripe
= conf
->raid_disks
* (mddev
->chunk_size
/ PAGE_SIZE
);
2184 stripe
/= conf
->near_copies
;
2185 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
2186 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
2189 if (conf
->near_copies
< mddev
->raid_disks
)
2190 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
2194 if (conf
->r10bio_pool
)
2195 mempool_destroy(conf
->r10bio_pool
);
2196 safe_put_page(conf
->tmppage
);
2197 kfree(conf
->mirrors
);
2199 mddev
->private = NULL
;
2204 static int stop(mddev_t
*mddev
)
2206 conf_t
*conf
= mddev_to_conf(mddev
);
2208 md_unregister_thread(mddev
->thread
);
2209 mddev
->thread
= NULL
;
2210 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
2211 if (conf
->r10bio_pool
)
2212 mempool_destroy(conf
->r10bio_pool
);
2213 kfree(conf
->mirrors
);
2215 mddev
->private = NULL
;
2219 static void raid10_quiesce(mddev_t
*mddev
, int state
)
2221 conf_t
*conf
= mddev_to_conf(mddev
);
2225 raise_barrier(conf
, 0);
2228 lower_barrier(conf
);
2231 if (mddev
->thread
) {
2233 mddev
->thread
->timeout
= mddev
->bitmap
->daemon_sleep
* HZ
;
2235 mddev
->thread
->timeout
= MAX_SCHEDULE_TIMEOUT
;
2236 md_wakeup_thread(mddev
->thread
);
2240 static struct mdk_personality raid10_personality
=
2244 .owner
= THIS_MODULE
,
2245 .make_request
= make_request
,
2249 .error_handler
= error
,
2250 .hot_add_disk
= raid10_add_disk
,
2251 .hot_remove_disk
= raid10_remove_disk
,
2252 .spare_active
= raid10_spare_active
,
2253 .sync_request
= sync_request
,
2254 .quiesce
= raid10_quiesce
,
2257 static int __init
raid_init(void)
2259 return register_md_personality(&raid10_personality
);
2262 static void raid_exit(void)
2264 unregister_md_personality(&raid10_personality
);
2267 module_init(raid_init
);
2268 module_exit(raid_exit
);
2269 MODULE_LICENSE("GPL");
2270 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2271 MODULE_ALIAS("md-raid10");
2272 MODULE_ALIAS("md-level-10");