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 #define RESYNC_BLOCK_SIZE (64*1024)
80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 #define RESYNC_WINDOW (2048*1024)
86 * When performing a resync, we need to read and compare, so
87 * we need as many pages are there are copies.
88 * When performing a recovery, we need 2 bios, one for read,
89 * one for write (we recover only one drive per r10buf)
92 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
101 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
103 unplug_slaves(conf
->mddev
);
107 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
108 nalloc
= conf
->copies
; /* resync */
110 nalloc
= 2; /* recovery */
115 for (j
= nalloc
; j
-- ; ) {
116 bio
= bio_alloc(gfp_flags
, RESYNC_PAGES
);
119 r10_bio
->devs
[j
].bio
= bio
;
122 * Allocate RESYNC_PAGES data pages and attach them
125 for (j
= 0 ; j
< nalloc
; j
++) {
126 bio
= r10_bio
->devs
[j
].bio
;
127 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
128 page
= alloc_page(gfp_flags
);
132 bio
->bi_io_vec
[i
].bv_page
= page
;
140 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
142 for (i
= 0; i
< RESYNC_PAGES
; i
++)
143 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
146 while ( ++j
< nalloc
)
147 bio_put(r10_bio
->devs
[j
].bio
);
148 r10bio_pool_free(r10_bio
, conf
);
152 static void r10buf_pool_free(void *__r10_bio
, void *data
)
156 r10bio_t
*r10bio
= __r10_bio
;
159 for (j
=0; j
< conf
->copies
; j
++) {
160 struct bio
*bio
= r10bio
->devs
[j
].bio
;
162 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
163 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
164 bio
->bi_io_vec
[i
].bv_page
= NULL
;
169 r10bio_pool_free(r10bio
, conf
);
172 static void put_all_bios(conf_t
*conf
, r10bio_t
*r10_bio
)
176 for (i
= 0; i
< conf
->copies
; i
++) {
177 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
178 if (*bio
&& *bio
!= IO_BLOCKED
)
184 static void free_r10bio(r10bio_t
*r10_bio
)
186 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
189 * Wake up any possible resync thread that waits for the device
194 put_all_bios(conf
, r10_bio
);
195 mempool_free(r10_bio
, conf
->r10bio_pool
);
198 static void put_buf(r10bio_t
*r10_bio
)
200 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
202 mempool_free(r10_bio
, conf
->r10buf_pool
);
207 static void reschedule_retry(r10bio_t
*r10_bio
)
210 mddev_t
*mddev
= r10_bio
->mddev
;
211 conf_t
*conf
= mddev_to_conf(mddev
);
213 spin_lock_irqsave(&conf
->device_lock
, flags
);
214 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
216 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
218 md_wakeup_thread(mddev
->thread
);
222 * raid_end_bio_io() is called when we have finished servicing a mirrored
223 * operation and are ready to return a success/failure code to the buffer
226 static void raid_end_bio_io(r10bio_t
*r10_bio
)
228 struct bio
*bio
= r10_bio
->master_bio
;
230 bio_endio(bio
, bio
->bi_size
,
231 test_bit(R10BIO_Uptodate
, &r10_bio
->state
) ? 0 : -EIO
);
232 free_r10bio(r10_bio
);
236 * Update disk head position estimator based on IRQ completion info.
238 static inline void update_head_pos(int slot
, r10bio_t
*r10_bio
)
240 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
242 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
243 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
246 static int raid10_end_read_request(struct bio
*bio
, unsigned int bytes_done
, int error
)
248 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
249 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
251 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
256 slot
= r10_bio
->read_slot
;
257 dev
= r10_bio
->devs
[slot
].devnum
;
259 * this branch is our 'one mirror IO has finished' event handler:
261 update_head_pos(slot
, r10_bio
);
265 * Set R10BIO_Uptodate in our master bio, so that
266 * we will return a good error code to the higher
267 * levels even if IO on some other mirrored buffer fails.
269 * The 'master' represents the composite IO operation to
270 * user-side. So if something waits for IO, then it will
271 * wait for the 'master' bio.
273 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
274 raid_end_bio_io(r10_bio
);
279 char b
[BDEVNAME_SIZE
];
280 if (printk_ratelimit())
281 printk(KERN_ERR
"raid10: %s: rescheduling sector %llu\n",
282 bdevname(conf
->mirrors
[dev
].rdev
->bdev
,b
), (unsigned long long)r10_bio
->sector
);
283 reschedule_retry(r10_bio
);
286 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
290 static int raid10_end_write_request(struct bio
*bio
, unsigned int bytes_done
, int error
)
292 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
293 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
295 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
300 for (slot
= 0; slot
< conf
->copies
; slot
++)
301 if (r10_bio
->devs
[slot
].bio
== bio
)
303 dev
= r10_bio
->devs
[slot
].devnum
;
306 * this branch is our 'one mirror IO has finished' event handler:
309 md_error(r10_bio
->mddev
, conf
->mirrors
[dev
].rdev
);
310 /* an I/O failed, we can't clear the bitmap */
311 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
314 * Set R10BIO_Uptodate in our master bio, so that
315 * we will return a good error code for to the higher
316 * levels even if IO on some other mirrored buffer fails.
318 * The 'master' represents the composite IO operation to
319 * user-side. So if something waits for IO, then it will
320 * wait for the 'master' bio.
322 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
324 update_head_pos(slot
, r10_bio
);
328 * Let's see if all mirrored write operations have finished
331 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
335 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
337 md_write_end(r10_bio
->mddev
);
338 raid_end_bio_io(r10_bio
);
341 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
347 * RAID10 layout manager
348 * Aswell as the chunksize and raid_disks count, there are two
349 * parameters: near_copies and far_copies.
350 * near_copies * far_copies must be <= raid_disks.
351 * Normally one of these will be 1.
352 * If both are 1, we get raid0.
353 * If near_copies == raid_disks, we get raid1.
355 * Chunks are layed out in raid0 style with near_copies copies of the
356 * first chunk, followed by near_copies copies of the next chunk and
358 * If far_copies > 1, then after 1/far_copies of the array has been assigned
359 * as described above, we start again with a device offset of near_copies.
360 * So we effectively have another copy of the whole array further down all
361 * the drives, but with blocks on different drives.
362 * With this layout, and block is never stored twice on the one device.
364 * raid10_find_phys finds the sector offset of a given virtual sector
365 * on each device that it is on.
367 * raid10_find_virt does the reverse mapping, from a device and a
368 * sector offset to a virtual address
371 static void raid10_find_phys(conf_t
*conf
, r10bio_t
*r10bio
)
381 /* now calculate first sector/dev */
382 chunk
= r10bio
->sector
>> conf
->chunk_shift
;
383 sector
= r10bio
->sector
& conf
->chunk_mask
;
385 chunk
*= conf
->near_copies
;
387 dev
= sector_div(stripe
, conf
->raid_disks
);
388 if (conf
->far_offset
)
389 stripe
*= conf
->far_copies
;
391 sector
+= stripe
<< conf
->chunk_shift
;
393 /* and calculate all the others */
394 for (n
=0; n
< conf
->near_copies
; n
++) {
397 r10bio
->devs
[slot
].addr
= sector
;
398 r10bio
->devs
[slot
].devnum
= d
;
401 for (f
= 1; f
< conf
->far_copies
; f
++) {
402 d
+= conf
->near_copies
;
403 if (d
>= conf
->raid_disks
)
404 d
-= conf
->raid_disks
;
406 r10bio
->devs
[slot
].devnum
= d
;
407 r10bio
->devs
[slot
].addr
= s
;
411 if (dev
>= conf
->raid_disks
) {
413 sector
+= (conf
->chunk_mask
+ 1);
416 BUG_ON(slot
!= conf
->copies
);
419 static sector_t
raid10_find_virt(conf_t
*conf
, sector_t sector
, int dev
)
421 sector_t offset
, chunk
, vchunk
;
423 offset
= sector
& conf
->chunk_mask
;
424 if (conf
->far_offset
) {
426 chunk
= sector
>> conf
->chunk_shift
;
427 fc
= sector_div(chunk
, conf
->far_copies
);
428 dev
-= fc
* conf
->near_copies
;
430 dev
+= conf
->raid_disks
;
432 while (sector
> conf
->stride
) {
433 sector
-= conf
->stride
;
434 if (dev
< conf
->near_copies
)
435 dev
+= conf
->raid_disks
- conf
->near_copies
;
437 dev
-= conf
->near_copies
;
439 chunk
= sector
>> conf
->chunk_shift
;
441 vchunk
= chunk
* conf
->raid_disks
+ dev
;
442 sector_div(vchunk
, conf
->near_copies
);
443 return (vchunk
<< conf
->chunk_shift
) + offset
;
447 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
449 * @bio: the buffer head that's been built up so far
450 * @biovec: the request that could be merged to it.
452 * Return amount of bytes we can accept at this offset
453 * If near_copies == raid_disk, there are no striping issues,
454 * but in that case, the function isn't called at all.
456 static int raid10_mergeable_bvec(request_queue_t
*q
, struct bio
*bio
,
457 struct bio_vec
*bio_vec
)
459 mddev_t
*mddev
= q
->queuedata
;
460 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
462 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
463 unsigned int bio_sectors
= bio
->bi_size
>> 9;
465 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
466 if (max
< 0) max
= 0; /* bio_add cannot handle a negative return */
467 if (max
<= bio_vec
->bv_len
&& bio_sectors
== 0)
468 return bio_vec
->bv_len
;
474 * This routine returns the disk from which the requested read should
475 * be done. There is a per-array 'next expected sequential IO' sector
476 * number - if this matches on the next IO then we use the last disk.
477 * There is also a per-disk 'last know head position' sector that is
478 * maintained from IRQ contexts, both the normal and the resync IO
479 * completion handlers update this position correctly. If there is no
480 * perfect sequential match then we pick the disk whose head is closest.
482 * If there are 2 mirrors in the same 2 devices, performance degrades
483 * because position is mirror, not device based.
485 * The rdev for the device selected will have nr_pending incremented.
489 * FIXME: possibly should rethink readbalancing and do it differently
490 * depending on near_copies / far_copies geometry.
492 static int read_balance(conf_t
*conf
, r10bio_t
*r10_bio
)
494 const unsigned long this_sector
= r10_bio
->sector
;
495 int disk
, slot
, nslot
;
496 const int sectors
= r10_bio
->sectors
;
497 sector_t new_distance
, current_distance
;
500 raid10_find_phys(conf
, r10_bio
);
503 * Check if we can balance. We can balance on the whole
504 * device if no resync is going on (recovery is ok), or below
505 * the resync window. We take the first readable disk when
506 * above the resync window.
508 if (conf
->mddev
->recovery_cp
< MaxSector
509 && (this_sector
+ sectors
>= conf
->next_resync
)) {
510 /* make sure that disk is operational */
512 disk
= r10_bio
->devs
[slot
].devnum
;
514 while ((rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
515 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
516 !test_bit(In_sync
, &rdev
->flags
)) {
518 if (slot
== conf
->copies
) {
523 disk
= r10_bio
->devs
[slot
].devnum
;
529 /* make sure the disk is operational */
531 disk
= r10_bio
->devs
[slot
].devnum
;
532 while ((rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
533 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
534 !test_bit(In_sync
, &rdev
->flags
)) {
536 if (slot
== conf
->copies
) {
540 disk
= r10_bio
->devs
[slot
].devnum
;
544 current_distance
= abs(r10_bio
->devs
[slot
].addr
-
545 conf
->mirrors
[disk
].head_position
);
547 /* Find the disk whose head is closest */
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
)) {
567 new_distance
= abs(r10_bio
->devs
[nslot
].addr
-
568 conf
->mirrors
[ndisk
].head_position
);
569 if (new_distance
< current_distance
) {
570 current_distance
= new_distance
;
577 r10_bio
->read_slot
= slot
;
578 /* conf->next_seq_sect = this_sector + sectors;*/
580 if (disk
>= 0 && (rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
))!= NULL
)
581 atomic_inc(&conf
->mirrors
[disk
].rdev
->nr_pending
);
589 static void unplug_slaves(mddev_t
*mddev
)
591 conf_t
*conf
= mddev_to_conf(mddev
);
595 for (i
=0; i
<mddev
->raid_disks
; i
++) {
596 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
597 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
598 request_queue_t
*r_queue
= bdev_get_queue(rdev
->bdev
);
600 atomic_inc(&rdev
->nr_pending
);
603 if (r_queue
->unplug_fn
)
604 r_queue
->unplug_fn(r_queue
);
606 rdev_dec_pending(rdev
, mddev
);
613 static void raid10_unplug(request_queue_t
*q
)
615 mddev_t
*mddev
= q
->queuedata
;
617 unplug_slaves(q
->queuedata
);
618 md_wakeup_thread(mddev
->thread
);
621 static int raid10_issue_flush(request_queue_t
*q
, struct gendisk
*disk
,
622 sector_t
*error_sector
)
624 mddev_t
*mddev
= q
->queuedata
;
625 conf_t
*conf
= mddev_to_conf(mddev
);
629 for (i
=0; i
<mddev
->raid_disks
&& ret
== 0; i
++) {
630 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
631 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
632 struct block_device
*bdev
= rdev
->bdev
;
633 request_queue_t
*r_queue
= bdev_get_queue(bdev
);
635 if (!r_queue
->issue_flush_fn
)
638 atomic_inc(&rdev
->nr_pending
);
640 ret
= r_queue
->issue_flush_fn(r_queue
, bdev
->bd_disk
,
642 rdev_dec_pending(rdev
, mddev
);
651 static int raid10_congested(void *data
, int bits
)
653 mddev_t
*mddev
= data
;
654 conf_t
*conf
= mddev_to_conf(mddev
);
658 for (i
= 0; i
< mddev
->raid_disks
&& ret
== 0; i
++) {
659 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
660 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
661 request_queue_t
*q
= bdev_get_queue(rdev
->bdev
);
663 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
672 * Sometimes we need to suspend IO while we do something else,
673 * either some resync/recovery, or reconfigure the array.
674 * To do this we raise a 'barrier'.
675 * The 'barrier' is a counter that can be raised multiple times
676 * to count how many activities are happening which preclude
678 * We can only raise the barrier if there is no pending IO.
679 * i.e. if nr_pending == 0.
680 * We choose only to raise the barrier if no-one is waiting for the
681 * barrier to go down. This means that as soon as an IO request
682 * is ready, no other operations which require a barrier will start
683 * until the IO request has had a chance.
685 * So: regular IO calls 'wait_barrier'. When that returns there
686 * is no backgroup IO happening, It must arrange to call
687 * allow_barrier when it has finished its IO.
688 * backgroup IO calls must call raise_barrier. Once that returns
689 * there is no normal IO happeing. It must arrange to call
690 * lower_barrier when the particular background IO completes.
692 #define RESYNC_DEPTH 32
694 static void raise_barrier(conf_t
*conf
, int force
)
696 BUG_ON(force
&& !conf
->barrier
);
697 spin_lock_irq(&conf
->resync_lock
);
699 /* Wait until no block IO is waiting (unless 'force') */
700 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
702 raid10_unplug(conf
->mddev
->queue
));
704 /* block any new IO from starting */
707 /* No wait for all pending IO to complete */
708 wait_event_lock_irq(conf
->wait_barrier
,
709 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
711 raid10_unplug(conf
->mddev
->queue
));
713 spin_unlock_irq(&conf
->resync_lock
);
716 static void lower_barrier(conf_t
*conf
)
719 spin_lock_irqsave(&conf
->resync_lock
, flags
);
721 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
722 wake_up(&conf
->wait_barrier
);
725 static void wait_barrier(conf_t
*conf
)
727 spin_lock_irq(&conf
->resync_lock
);
730 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
732 raid10_unplug(conf
->mddev
->queue
));
736 spin_unlock_irq(&conf
->resync_lock
);
739 static void allow_barrier(conf_t
*conf
)
742 spin_lock_irqsave(&conf
->resync_lock
, flags
);
744 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
745 wake_up(&conf
->wait_barrier
);
748 static void freeze_array(conf_t
*conf
)
750 /* stop syncio and normal IO and wait for everything to
752 * We increment barrier and nr_waiting, and then
753 * wait until barrier+nr_pending match nr_queued+2
755 spin_lock_irq(&conf
->resync_lock
);
758 wait_event_lock_irq(conf
->wait_barrier
,
759 conf
->barrier
+conf
->nr_pending
== conf
->nr_queued
+2,
761 raid10_unplug(conf
->mddev
->queue
));
762 spin_unlock_irq(&conf
->resync_lock
);
765 static void unfreeze_array(conf_t
*conf
)
767 /* reverse the effect of the freeze */
768 spin_lock_irq(&conf
->resync_lock
);
771 wake_up(&conf
->wait_barrier
);
772 spin_unlock_irq(&conf
->resync_lock
);
775 static int make_request(request_queue_t
*q
, struct bio
* bio
)
777 mddev_t
*mddev
= q
->queuedata
;
778 conf_t
*conf
= mddev_to_conf(mddev
);
779 mirror_info_t
*mirror
;
781 struct bio
*read_bio
;
783 int chunk_sects
= conf
->chunk_mask
+ 1;
784 const int rw
= bio_data_dir(bio
);
788 if (unlikely(bio_barrier(bio
))) {
789 bio_endio(bio
, bio
->bi_size
, -EOPNOTSUPP
);
793 /* If this request crosses a chunk boundary, we need to
794 * split it. This will only happen for 1 PAGE (or less) requests.
796 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
798 conf
->near_copies
< conf
->raid_disks
)) {
800 /* Sanity check -- queue functions should prevent this happening */
801 if (bio
->bi_vcnt
!= 1 ||
804 /* This is a one page bio that upper layers
805 * refuse to split for us, so we need to split it.
807 bp
= bio_split(bio
, bio_split_pool
,
808 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
809 if (make_request(q
, &bp
->bio1
))
810 generic_make_request(&bp
->bio1
);
811 if (make_request(q
, &bp
->bio2
))
812 generic_make_request(&bp
->bio2
);
814 bio_pair_release(bp
);
817 printk("raid10_make_request bug: can't convert block across chunks"
818 " or bigger than %dk %llu %d\n", chunk_sects
/2,
819 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
821 bio_io_error(bio
, bio
->bi_size
);
825 md_write_start(mddev
, bio
);
828 * Register the new request and wait if the reconstruction
829 * thread has put up a bar for new requests.
830 * Continue immediately if no resync is active currently.
834 disk_stat_inc(mddev
->gendisk
, ios
[rw
]);
835 disk_stat_add(mddev
->gendisk
, sectors
[rw
], bio_sectors(bio
));
837 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
839 r10_bio
->master_bio
= bio
;
840 r10_bio
->sectors
= bio
->bi_size
>> 9;
842 r10_bio
->mddev
= mddev
;
843 r10_bio
->sector
= bio
->bi_sector
;
848 * read balancing logic:
850 int disk
= read_balance(conf
, r10_bio
);
851 int slot
= r10_bio
->read_slot
;
853 raid_end_bio_io(r10_bio
);
856 mirror
= conf
->mirrors
+ disk
;
858 read_bio
= bio_clone(bio
, GFP_NOIO
);
860 r10_bio
->devs
[slot
].bio
= read_bio
;
862 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
863 mirror
->rdev
->data_offset
;
864 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
865 read_bio
->bi_end_io
= raid10_end_read_request
;
866 read_bio
->bi_rw
= READ
;
867 read_bio
->bi_private
= r10_bio
;
869 generic_make_request(read_bio
);
876 /* first select target devices under spinlock and
877 * inc refcount on their rdev. Record them by setting
880 raid10_find_phys(conf
, r10_bio
);
882 for (i
= 0; i
< conf
->copies
; i
++) {
883 int d
= r10_bio
->devs
[i
].devnum
;
884 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
886 !test_bit(Faulty
, &rdev
->flags
)) {
887 atomic_inc(&rdev
->nr_pending
);
888 r10_bio
->devs
[i
].bio
= bio
;
890 r10_bio
->devs
[i
].bio
= NULL
;
891 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
896 atomic_set(&r10_bio
->remaining
, 0);
899 for (i
= 0; i
< conf
->copies
; i
++) {
901 int d
= r10_bio
->devs
[i
].devnum
;
902 if (!r10_bio
->devs
[i
].bio
)
905 mbio
= bio_clone(bio
, GFP_NOIO
);
906 r10_bio
->devs
[i
].bio
= mbio
;
908 mbio
->bi_sector
= r10_bio
->devs
[i
].addr
+
909 conf
->mirrors
[d
].rdev
->data_offset
;
910 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
911 mbio
->bi_end_io
= raid10_end_write_request
;
913 mbio
->bi_private
= r10_bio
;
915 atomic_inc(&r10_bio
->remaining
);
916 bio_list_add(&bl
, mbio
);
919 bitmap_startwrite(mddev
->bitmap
, bio
->bi_sector
, r10_bio
->sectors
, 0);
920 spin_lock_irqsave(&conf
->device_lock
, flags
);
921 bio_list_merge(&conf
->pending_bio_list
, &bl
);
922 blk_plug_device(mddev
->queue
);
923 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
928 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
930 conf_t
*conf
= mddev_to_conf(mddev
);
933 if (conf
->near_copies
< conf
->raid_disks
)
934 seq_printf(seq
, " %dK chunks", mddev
->chunk_size
/1024);
935 if (conf
->near_copies
> 1)
936 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
937 if (conf
->far_copies
> 1) {
938 if (conf
->far_offset
)
939 seq_printf(seq
, " %d offset-copies", conf
->far_copies
);
941 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
943 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
944 conf
->raid_disks
- mddev
->degraded
);
945 for (i
= 0; i
< conf
->raid_disks
; i
++)
946 seq_printf(seq
, "%s",
947 conf
->mirrors
[i
].rdev
&&
948 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
949 seq_printf(seq
, "]");
952 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
954 char b
[BDEVNAME_SIZE
];
955 conf_t
*conf
= mddev_to_conf(mddev
);
958 * If it is not operational, then we have already marked it as dead
959 * else if it is the last working disks, ignore the error, let the
960 * next level up know.
961 * else mark the drive as failed
963 if (test_bit(In_sync
, &rdev
->flags
)
964 && conf
->raid_disks
-mddev
->degraded
== 1)
966 * Don't fail the drive, just return an IO error.
967 * The test should really be more sophisticated than
968 * "working_disks == 1", but it isn't critical, and
969 * can wait until we do more sophisticated "is the drive
970 * really dead" tests...
973 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
975 spin_lock_irqsave(&conf
->device_lock
, flags
);
977 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
979 * if recovery is running, make sure it aborts.
981 set_bit(MD_RECOVERY_ERR
, &mddev
->recovery
);
983 set_bit(Faulty
, &rdev
->flags
);
984 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
985 printk(KERN_ALERT
"raid10: Disk failure on %s, disabling device. \n"
986 " Operation continuing on %d devices\n",
987 bdevname(rdev
->bdev
,b
), conf
->raid_disks
- mddev
->degraded
);
990 static void print_conf(conf_t
*conf
)
995 printk("RAID10 conf printout:\n");
1000 printk(" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1003 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1004 char b
[BDEVNAME_SIZE
];
1005 tmp
= conf
->mirrors
+ i
;
1007 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1008 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1009 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1010 bdevname(tmp
->rdev
->bdev
,b
));
1014 static void close_sync(conf_t
*conf
)
1017 allow_barrier(conf
);
1019 mempool_destroy(conf
->r10buf_pool
);
1020 conf
->r10buf_pool
= NULL
;
1023 /* check if there are enough drives for
1024 * every block to appear on atleast one
1026 static int enough(conf_t
*conf
)
1031 int n
= conf
->copies
;
1034 if (conf
->mirrors
[first
].rdev
)
1036 first
= (first
+1) % conf
->raid_disks
;
1040 } while (first
!= 0);
1044 static int raid10_spare_active(mddev_t
*mddev
)
1047 conf_t
*conf
= mddev
->private;
1051 * Find all non-in_sync disks within the RAID10 configuration
1052 * and mark them in_sync
1054 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1055 tmp
= conf
->mirrors
+ i
;
1057 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1058 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1059 unsigned long flags
;
1060 spin_lock_irqsave(&conf
->device_lock
, flags
);
1062 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1071 static int raid10_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1073 conf_t
*conf
= mddev
->private;
1078 if (mddev
->recovery_cp
< MaxSector
)
1079 /* only hot-add to in-sync arrays, as recovery is
1080 * very different from resync
1086 if (rdev
->saved_raid_disk
>= 0 &&
1087 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1088 mirror
= rdev
->saved_raid_disk
;
1091 for ( ; mirror
< mddev
->raid_disks
; mirror
++)
1092 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1094 blk_queue_stack_limits(mddev
->queue
,
1095 rdev
->bdev
->bd_disk
->queue
);
1096 /* as we don't honour merge_bvec_fn, we must never risk
1097 * violating it, so limit ->max_sector to one PAGE, as
1098 * a one page request is never in violation.
1100 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
1101 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
1102 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
1104 p
->head_position
= 0;
1105 rdev
->raid_disk
= mirror
;
1107 if (rdev
->saved_raid_disk
!= mirror
)
1109 rcu_assign_pointer(p
->rdev
, rdev
);
1117 static int raid10_remove_disk(mddev_t
*mddev
, int number
)
1119 conf_t
*conf
= mddev
->private;
1122 mirror_info_t
*p
= conf
->mirrors
+ number
;
1127 if (test_bit(In_sync
, &rdev
->flags
) ||
1128 atomic_read(&rdev
->nr_pending
)) {
1134 if (atomic_read(&rdev
->nr_pending
)) {
1135 /* lost the race, try later */
1147 static int end_sync_read(struct bio
*bio
, unsigned int bytes_done
, int error
)
1149 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1150 conf_t
*conf
= mddev_to_conf(r10_bio
->mddev
);
1156 for (i
=0; i
<conf
->copies
; i
++)
1157 if (r10_bio
->devs
[i
].bio
== bio
)
1159 BUG_ON(i
== conf
->copies
);
1160 update_head_pos(i
, r10_bio
);
1161 d
= r10_bio
->devs
[i
].devnum
;
1163 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1164 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1166 atomic_add(r10_bio
->sectors
,
1167 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1168 if (!test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
1169 md_error(r10_bio
->mddev
,
1170 conf
->mirrors
[d
].rdev
);
1173 /* for reconstruct, we always reschedule after a read.
1174 * for resync, only after all reads
1176 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1177 atomic_dec_and_test(&r10_bio
->remaining
)) {
1178 /* we have read all the blocks,
1179 * do the comparison in process context in raid10d
1181 reschedule_retry(r10_bio
);
1183 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1187 static int end_sync_write(struct bio
*bio
, unsigned int bytes_done
, int error
)
1189 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1190 r10bio_t
* r10_bio
= (r10bio_t
*)(bio
->bi_private
);
1191 mddev_t
*mddev
= r10_bio
->mddev
;
1192 conf_t
*conf
= mddev_to_conf(mddev
);
1198 for (i
= 0; i
< conf
->copies
; i
++)
1199 if (r10_bio
->devs
[i
].bio
== bio
)
1201 d
= r10_bio
->devs
[i
].devnum
;
1204 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1205 update_head_pos(i
, r10_bio
);
1207 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1208 if (r10_bio
->master_bio
== NULL
) {
1209 /* the primary of several recovery bios */
1210 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1214 r10bio_t
*r10_bio2
= (r10bio_t
*)r10_bio
->master_bio
;
1219 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1224 * Note: sync and recover and handled very differently for raid10
1225 * This code is for resync.
1226 * For resync, we read through virtual addresses and read all blocks.
1227 * If there is any error, we schedule a write. The lowest numbered
1228 * drive is authoritative.
1229 * However requests come for physical address, so we need to map.
1230 * For every physical address there are raid_disks/copies virtual addresses,
1231 * which is always are least one, but is not necessarly an integer.
1232 * This means that a physical address can span multiple chunks, so we may
1233 * have to submit multiple io requests for a single sync request.
1236 * We check if all blocks are in-sync and only write to blocks that
1239 static void sync_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1241 conf_t
*conf
= mddev_to_conf(mddev
);
1243 struct bio
*tbio
, *fbio
;
1245 atomic_set(&r10_bio
->remaining
, 1);
1247 /* find the first device with a block */
1248 for (i
=0; i
<conf
->copies
; i
++)
1249 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1252 if (i
== conf
->copies
)
1256 fbio
= r10_bio
->devs
[i
].bio
;
1258 /* now find blocks with errors */
1259 for (i
=0 ; i
< conf
->copies
; i
++) {
1261 int vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1263 tbio
= r10_bio
->devs
[i
].bio
;
1265 if (tbio
->bi_end_io
!= end_sync_read
)
1269 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1270 /* We know that the bi_io_vec layout is the same for
1271 * both 'first' and 'i', so we just compare them.
1272 * All vec entries are PAGE_SIZE;
1274 for (j
= 0; j
< vcnt
; j
++)
1275 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1276 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1281 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1283 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1284 /* Don't fix anything. */
1286 /* Ok, we need to write this bio
1287 * First we need to fixup bv_offset, bv_len and
1288 * bi_vecs, as the read request might have corrupted these
1290 tbio
->bi_vcnt
= vcnt
;
1291 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1293 tbio
->bi_phys_segments
= 0;
1294 tbio
->bi_hw_segments
= 0;
1295 tbio
->bi_hw_front_size
= 0;
1296 tbio
->bi_hw_back_size
= 0;
1297 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1298 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1299 tbio
->bi_next
= NULL
;
1300 tbio
->bi_rw
= WRITE
;
1301 tbio
->bi_private
= r10_bio
;
1302 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1304 for (j
=0; j
< vcnt
; j
++) {
1305 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1306 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1308 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1309 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1312 tbio
->bi_end_io
= end_sync_write
;
1314 d
= r10_bio
->devs
[i
].devnum
;
1315 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1316 atomic_inc(&r10_bio
->remaining
);
1317 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1319 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1320 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1321 generic_make_request(tbio
);
1325 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1326 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1332 * Now for the recovery code.
1333 * Recovery happens across physical sectors.
1334 * We recover all non-is_sync drives by finding the virtual address of
1335 * each, and then choose a working drive that also has that virt address.
1336 * There is a separate r10_bio for each non-in_sync drive.
1337 * Only the first two slots are in use. The first for reading,
1338 * The second for writing.
1342 static void recovery_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1344 conf_t
*conf
= mddev_to_conf(mddev
);
1346 struct bio
*bio
, *wbio
;
1349 /* move the pages across to the second bio
1350 * and submit the write request
1352 bio
= r10_bio
->devs
[0].bio
;
1353 wbio
= r10_bio
->devs
[1].bio
;
1354 for (i
=0; i
< wbio
->bi_vcnt
; i
++) {
1355 struct page
*p
= bio
->bi_io_vec
[i
].bv_page
;
1356 bio
->bi_io_vec
[i
].bv_page
= wbio
->bi_io_vec
[i
].bv_page
;
1357 wbio
->bi_io_vec
[i
].bv_page
= p
;
1359 d
= r10_bio
->devs
[1].devnum
;
1361 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1362 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1363 if (test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
1364 generic_make_request(wbio
);
1366 bio_endio(wbio
, wbio
->bi_size
, -EIO
);
1371 * This is a kernel thread which:
1373 * 1. Retries failed read operations on working mirrors.
1374 * 2. Updates the raid superblock when problems encounter.
1375 * 3. Performs writes following reads for array synchronising.
1378 static void fix_read_error(conf_t
*conf
, mddev_t
*mddev
, r10bio_t
*r10_bio
)
1380 int sect
= 0; /* Offset from r10_bio->sector */
1381 int sectors
= r10_bio
->sectors
;
1385 int sl
= r10_bio
->read_slot
;
1389 if (s
> (PAGE_SIZE
>>9))
1394 int d
= r10_bio
->devs
[sl
].devnum
;
1395 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1397 test_bit(In_sync
, &rdev
->flags
)) {
1398 atomic_inc(&rdev
->nr_pending
);
1400 success
= sync_page_io(rdev
->bdev
,
1401 r10_bio
->devs
[sl
].addr
+
1402 sect
+ rdev
->data_offset
,
1404 conf
->tmppage
, READ
);
1405 rdev_dec_pending(rdev
, mddev
);
1411 if (sl
== conf
->copies
)
1413 } while (!success
&& sl
!= r10_bio
->read_slot
);
1417 /* Cannot read from anywhere -- bye bye array */
1418 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1419 md_error(mddev
, conf
->mirrors
[dn
].rdev
);
1424 /* write it back and re-read */
1426 while (sl
!= r10_bio
->read_slot
) {
1431 d
= r10_bio
->devs
[sl
].devnum
;
1432 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1434 test_bit(In_sync
, &rdev
->flags
)) {
1435 atomic_inc(&rdev
->nr_pending
);
1437 atomic_add(s
, &rdev
->corrected_errors
);
1438 if (sync_page_io(rdev
->bdev
,
1439 r10_bio
->devs
[sl
].addr
+
1440 sect
+ rdev
->data_offset
,
1441 s
<<9, conf
->tmppage
, WRITE
)
1443 /* Well, this device is dead */
1444 md_error(mddev
, rdev
);
1445 rdev_dec_pending(rdev
, mddev
);
1450 while (sl
!= r10_bio
->read_slot
) {
1455 d
= r10_bio
->devs
[sl
].devnum
;
1456 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1458 test_bit(In_sync
, &rdev
->flags
)) {
1459 char b
[BDEVNAME_SIZE
];
1460 atomic_inc(&rdev
->nr_pending
);
1462 if (sync_page_io(rdev
->bdev
,
1463 r10_bio
->devs
[sl
].addr
+
1464 sect
+ rdev
->data_offset
,
1465 s
<<9, conf
->tmppage
, READ
) == 0)
1466 /* Well, this device is dead */
1467 md_error(mddev
, rdev
);
1470 "raid10:%s: read error corrected"
1471 " (%d sectors at %llu on %s)\n",
1473 (unsigned long long)sect
+
1475 bdevname(rdev
->bdev
, b
));
1477 rdev_dec_pending(rdev
, mddev
);
1488 static void raid10d(mddev_t
*mddev
)
1492 unsigned long flags
;
1493 conf_t
*conf
= mddev_to_conf(mddev
);
1494 struct list_head
*head
= &conf
->retry_list
;
1498 md_check_recovery(mddev
);
1501 char b
[BDEVNAME_SIZE
];
1502 spin_lock_irqsave(&conf
->device_lock
, flags
);
1504 if (conf
->pending_bio_list
.head
) {
1505 bio
= bio_list_get(&conf
->pending_bio_list
);
1506 blk_remove_plug(mddev
->queue
);
1507 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1508 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1509 if (bitmap_unplug(mddev
->bitmap
) != 0)
1510 printk("%s: bitmap file write failed!\n", mdname(mddev
));
1512 while (bio
) { /* submit pending writes */
1513 struct bio
*next
= bio
->bi_next
;
1514 bio
->bi_next
= NULL
;
1515 generic_make_request(bio
);
1523 if (list_empty(head
))
1525 r10_bio
= list_entry(head
->prev
, r10bio_t
, retry_list
);
1526 list_del(head
->prev
);
1528 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1530 mddev
= r10_bio
->mddev
;
1531 conf
= mddev_to_conf(mddev
);
1532 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
)) {
1533 sync_request_write(mddev
, r10_bio
);
1535 } else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
1536 recovery_request_write(mddev
, r10_bio
);
1540 /* we got a read error. Maybe the drive is bad. Maybe just
1541 * the block and we can fix it.
1542 * We freeze all other IO, and try reading the block from
1543 * other devices. When we find one, we re-write
1544 * and check it that fixes the read error.
1545 * This is all done synchronously while the array is
1548 if (mddev
->ro
== 0) {
1550 fix_read_error(conf
, mddev
, r10_bio
);
1551 unfreeze_array(conf
);
1554 bio
= r10_bio
->devs
[r10_bio
->read_slot
].bio
;
1555 r10_bio
->devs
[r10_bio
->read_slot
].bio
=
1556 mddev
->ro
? IO_BLOCKED
: NULL
;
1558 mirror
= read_balance(conf
, r10_bio
);
1560 printk(KERN_ALERT
"raid10: %s: unrecoverable I/O"
1561 " read error for block %llu\n",
1562 bdevname(bio
->bi_bdev
,b
),
1563 (unsigned long long)r10_bio
->sector
);
1564 raid_end_bio_io(r10_bio
);
1566 rdev
= conf
->mirrors
[mirror
].rdev
;
1567 if (printk_ratelimit())
1568 printk(KERN_ERR
"raid10: %s: redirecting sector %llu to"
1569 " another mirror\n",
1570 bdevname(rdev
->bdev
,b
),
1571 (unsigned long long)r10_bio
->sector
);
1572 bio
= bio_clone(r10_bio
->master_bio
, GFP_NOIO
);
1573 r10_bio
->devs
[r10_bio
->read_slot
].bio
= bio
;
1574 bio
->bi_sector
= r10_bio
->devs
[r10_bio
->read_slot
].addr
1575 + rdev
->data_offset
;
1576 bio
->bi_bdev
= rdev
->bdev
;
1578 bio
->bi_private
= r10_bio
;
1579 bio
->bi_end_io
= raid10_end_read_request
;
1581 generic_make_request(bio
);
1585 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1587 unplug_slaves(mddev
);
1591 static int init_resync(conf_t
*conf
)
1595 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1596 BUG_ON(conf
->r10buf_pool
);
1597 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
1598 if (!conf
->r10buf_pool
)
1600 conf
->next_resync
= 0;
1605 * perform a "sync" on one "block"
1607 * We need to make sure that no normal I/O request - particularly write
1608 * requests - conflict with active sync requests.
1610 * This is achieved by tracking pending requests and a 'barrier' concept
1611 * that can be installed to exclude normal IO requests.
1613 * Resync and recovery are handled very differently.
1614 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1616 * For resync, we iterate over virtual addresses, read all copies,
1617 * and update if there are differences. If only one copy is live,
1619 * For recovery, we iterate over physical addresses, read a good
1620 * value for each non-in_sync drive, and over-write.
1622 * So, for recovery we may have several outstanding complex requests for a
1623 * given address, one for each out-of-sync device. We model this by allocating
1624 * a number of r10_bio structures, one for each out-of-sync device.
1625 * As we setup these structures, we collect all bio's together into a list
1626 * which we then process collectively to add pages, and then process again
1627 * to pass to generic_make_request.
1629 * The r10_bio structures are linked using a borrowed master_bio pointer.
1630 * This link is counted in ->remaining. When the r10_bio that points to NULL
1631 * has its remaining count decremented to 0, the whole complex operation
1636 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
1638 conf_t
*conf
= mddev_to_conf(mddev
);
1640 struct bio
*biolist
= NULL
, *bio
;
1641 sector_t max_sector
, nr_sectors
;
1647 sector_t sectors_skipped
= 0;
1648 int chunks_skipped
= 0;
1650 if (!conf
->r10buf_pool
)
1651 if (init_resync(conf
))
1655 max_sector
= mddev
->size
<< 1;
1656 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1657 max_sector
= mddev
->resync_max_sectors
;
1658 if (sector_nr
>= max_sector
) {
1659 /* If we aborted, we need to abort the
1660 * sync on the 'current' bitmap chucks (there can
1661 * be several when recovering multiple devices).
1662 * as we may have started syncing it but not finished.
1663 * We can find the current address in
1664 * mddev->curr_resync, but for recovery,
1665 * we need to convert that to several
1666 * virtual addresses.
1668 if (mddev
->curr_resync
< max_sector
) { /* aborted */
1669 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1670 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
1672 else for (i
=0; i
<conf
->raid_disks
; i
++) {
1674 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
1675 bitmap_end_sync(mddev
->bitmap
, sect
,
1678 } else /* completed sync */
1681 bitmap_close_sync(mddev
->bitmap
);
1684 return sectors_skipped
;
1686 if (chunks_skipped
>= conf
->raid_disks
) {
1687 /* if there has been nothing to do on any drive,
1688 * then there is nothing to do at all..
1691 return (max_sector
- sector_nr
) + sectors_skipped
;
1694 /* make sure whole request will fit in a chunk - if chunks
1697 if (conf
->near_copies
< conf
->raid_disks
&&
1698 max_sector
> (sector_nr
| conf
->chunk_mask
))
1699 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
1701 * If there is non-resync activity waiting for us then
1702 * put in a delay to throttle resync.
1704 if (!go_faster
&& conf
->nr_waiting
)
1705 msleep_interruptible(1000);
1707 /* Again, very different code for resync and recovery.
1708 * Both must result in an r10bio with a list of bios that
1709 * have bi_end_io, bi_sector, bi_bdev set,
1710 * and bi_private set to the r10bio.
1711 * For recovery, we may actually create several r10bios
1712 * with 2 bios in each, that correspond to the bios in the main one.
1713 * In this case, the subordinate r10bios link back through a
1714 * borrowed master_bio pointer, and the counter in the master
1715 * includes a ref from each subordinate.
1717 /* First, we decide what to do and set ->bi_end_io
1718 * To end_sync_read if we want to read, and
1719 * end_sync_write if we will want to write.
1722 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
1723 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
1724 /* recovery... the complicated one */
1728 for (i
=0 ; i
<conf
->raid_disks
; i
++)
1729 if (conf
->mirrors
[i
].rdev
&&
1730 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
)) {
1731 int still_degraded
= 0;
1732 /* want to reconstruct this device */
1733 r10bio_t
*rb2
= r10_bio
;
1734 sector_t sect
= raid10_find_virt(conf
, sector_nr
, i
);
1736 /* Unless we are doing a full sync, we only need
1737 * to recover the block if it is set in the bitmap
1739 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1741 if (sync_blocks
< max_sync
)
1742 max_sync
= sync_blocks
;
1745 /* yep, skip the sync_blocks here, but don't assume
1746 * that there will never be anything to do here
1748 chunks_skipped
= -1;
1752 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1753 raise_barrier(conf
, rb2
!= NULL
);
1754 atomic_set(&r10_bio
->remaining
, 0);
1756 r10_bio
->master_bio
= (struct bio
*)rb2
;
1758 atomic_inc(&rb2
->remaining
);
1759 r10_bio
->mddev
= mddev
;
1760 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
1761 r10_bio
->sector
= sect
;
1763 raid10_find_phys(conf
, r10_bio
);
1764 /* Need to check if this section will still be
1767 for (j
=0; j
<conf
->copies
;j
++) {
1768 int d
= r10_bio
->devs
[j
].devnum
;
1769 if (conf
->mirrors
[d
].rdev
== NULL
||
1770 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
)) {
1775 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1776 &sync_blocks
, still_degraded
);
1778 for (j
=0; j
<conf
->copies
;j
++) {
1779 int d
= r10_bio
->devs
[j
].devnum
;
1780 if (conf
->mirrors
[d
].rdev
&&
1781 test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
)) {
1782 /* This is where we read from */
1783 bio
= r10_bio
->devs
[0].bio
;
1784 bio
->bi_next
= biolist
;
1786 bio
->bi_private
= r10_bio
;
1787 bio
->bi_end_io
= end_sync_read
;
1789 bio
->bi_sector
= r10_bio
->devs
[j
].addr
+
1790 conf
->mirrors
[d
].rdev
->data_offset
;
1791 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1792 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1793 atomic_inc(&r10_bio
->remaining
);
1794 /* and we write to 'i' */
1796 for (k
=0; k
<conf
->copies
; k
++)
1797 if (r10_bio
->devs
[k
].devnum
== i
)
1799 bio
= r10_bio
->devs
[1].bio
;
1800 bio
->bi_next
= biolist
;
1802 bio
->bi_private
= r10_bio
;
1803 bio
->bi_end_io
= end_sync_write
;
1805 bio
->bi_sector
= r10_bio
->devs
[k
].addr
+
1806 conf
->mirrors
[i
].rdev
->data_offset
;
1807 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1809 r10_bio
->devs
[0].devnum
= d
;
1810 r10_bio
->devs
[1].devnum
= i
;
1815 if (j
== conf
->copies
) {
1816 /* Cannot recover, so abort the recovery */
1819 if (!test_and_set_bit(MD_RECOVERY_ERR
, &mddev
->recovery
))
1820 printk(KERN_INFO
"raid10: %s: insufficient working devices for recovery.\n",
1825 if (biolist
== NULL
) {
1827 r10bio_t
*rb2
= r10_bio
;
1828 r10_bio
= (r10bio_t
*) rb2
->master_bio
;
1829 rb2
->master_bio
= NULL
;
1835 /* resync. Schedule a read for every block at this virt offset */
1838 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
1839 &sync_blocks
, mddev
->degraded
) &&
1840 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1841 /* We can skip this block */
1843 return sync_blocks
+ sectors_skipped
;
1845 if (sync_blocks
< max_sync
)
1846 max_sync
= sync_blocks
;
1847 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1849 r10_bio
->mddev
= mddev
;
1850 atomic_set(&r10_bio
->remaining
, 0);
1851 raise_barrier(conf
, 0);
1852 conf
->next_resync
= sector_nr
;
1854 r10_bio
->master_bio
= NULL
;
1855 r10_bio
->sector
= sector_nr
;
1856 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
1857 raid10_find_phys(conf
, r10_bio
);
1858 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
1860 for (i
=0; i
<conf
->copies
; i
++) {
1861 int d
= r10_bio
->devs
[i
].devnum
;
1862 bio
= r10_bio
->devs
[i
].bio
;
1863 bio
->bi_end_io
= NULL
;
1864 if (conf
->mirrors
[d
].rdev
== NULL
||
1865 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
1867 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1868 atomic_inc(&r10_bio
->remaining
);
1869 bio
->bi_next
= biolist
;
1871 bio
->bi_private
= r10_bio
;
1872 bio
->bi_end_io
= end_sync_read
;
1874 bio
->bi_sector
= r10_bio
->devs
[i
].addr
+
1875 conf
->mirrors
[d
].rdev
->data_offset
;
1876 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1881 for (i
=0; i
<conf
->copies
; i
++) {
1882 int d
= r10_bio
->devs
[i
].devnum
;
1883 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
1884 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1892 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1894 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1896 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
1899 bio
->bi_phys_segments
= 0;
1900 bio
->bi_hw_segments
= 0;
1905 if (sector_nr
+ max_sync
< max_sector
)
1906 max_sector
= sector_nr
+ max_sync
;
1909 int len
= PAGE_SIZE
;
1911 if (sector_nr
+ (len
>>9) > max_sector
)
1912 len
= (max_sector
- sector_nr
) << 9;
1915 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
1916 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
1917 if (bio_add_page(bio
, page
, len
, 0) == 0) {
1920 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
1921 for (bio2
= biolist
; bio2
&& bio2
!= bio
; bio2
= bio2
->bi_next
) {
1922 /* remove last page from this bio */
1924 bio2
->bi_size
-= len
;
1925 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
1931 nr_sectors
+= len
>>9;
1932 sector_nr
+= len
>>9;
1933 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
1935 r10_bio
->sectors
= nr_sectors
;
1939 biolist
= biolist
->bi_next
;
1941 bio
->bi_next
= NULL
;
1942 r10_bio
= bio
->bi_private
;
1943 r10_bio
->sectors
= nr_sectors
;
1945 if (bio
->bi_end_io
== end_sync_read
) {
1946 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1947 generic_make_request(bio
);
1951 if (sectors_skipped
)
1952 /* pretend they weren't skipped, it makes
1953 * no important difference in this case
1955 md_done_sync(mddev
, sectors_skipped
, 1);
1957 return sectors_skipped
+ nr_sectors
;
1959 /* There is nowhere to write, so all non-sync
1960 * drives must be failed, so try the next chunk...
1963 sector_t sec
= max_sector
- sector_nr
;
1964 sectors_skipped
+= sec
;
1966 sector_nr
= max_sector
;
1971 static int run(mddev_t
*mddev
)
1975 mirror_info_t
*disk
;
1977 struct list_head
*tmp
;
1979 sector_t stride
, size
;
1981 if (mddev
->chunk_size
== 0) {
1982 printk(KERN_ERR
"md/raid10: non-zero chunk size required.\n");
1986 nc
= mddev
->layout
& 255;
1987 fc
= (mddev
->layout
>> 8) & 255;
1988 fo
= mddev
->layout
& (1<<16);
1989 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
1990 (mddev
->layout
>> 17)) {
1991 printk(KERN_ERR
"raid10: %s: unsupported raid10 layout: 0x%8x\n",
1992 mdname(mddev
), mddev
->layout
);
1996 * copy the already verified devices into our private RAID10
1997 * bookkeeping area. [whatever we allocate in run(),
1998 * should be freed in stop()]
2000 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
2001 mddev
->private = conf
;
2003 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
2007 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
2009 if (!conf
->mirrors
) {
2010 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
2015 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2019 conf
->near_copies
= nc
;
2020 conf
->far_copies
= fc
;
2021 conf
->copies
= nc
*fc
;
2022 conf
->far_offset
= fo
;
2023 conf
->chunk_mask
= (sector_t
)(mddev
->chunk_size
>>9)-1;
2024 conf
->chunk_shift
= ffz(~mddev
->chunk_size
) - 9;
2026 conf
->stride
= 1 << conf
->chunk_shift
;
2028 stride
= mddev
->size
>> (conf
->chunk_shift
-1);
2029 sector_div(stride
, fc
);
2030 conf
->stride
= stride
<< conf
->chunk_shift
;
2032 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
2033 r10bio_pool_free
, conf
);
2034 if (!conf
->r10bio_pool
) {
2035 printk(KERN_ERR
"raid10: couldn't allocate memory for %s\n",
2040 ITERATE_RDEV(mddev
, rdev
, tmp
) {
2041 disk_idx
= rdev
->raid_disk
;
2042 if (disk_idx
>= mddev
->raid_disks
2045 disk
= conf
->mirrors
+ disk_idx
;
2049 blk_queue_stack_limits(mddev
->queue
,
2050 rdev
->bdev
->bd_disk
->queue
);
2051 /* as we don't honour merge_bvec_fn, we must never risk
2052 * violating it, so limit ->max_sector to one PAGE, as
2053 * a one page request is never in violation.
2055 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
2056 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
2057 mddev
->queue
->max_sectors
= (PAGE_SIZE
>>9);
2059 disk
->head_position
= 0;
2061 conf
->raid_disks
= mddev
->raid_disks
;
2062 conf
->mddev
= mddev
;
2063 spin_lock_init(&conf
->device_lock
);
2064 INIT_LIST_HEAD(&conf
->retry_list
);
2066 spin_lock_init(&conf
->resync_lock
);
2067 init_waitqueue_head(&conf
->wait_barrier
);
2069 /* need to check that every block has at least one working mirror */
2070 if (!enough(conf
)) {
2071 printk(KERN_ERR
"raid10: not enough operational mirrors for %s\n",
2076 mddev
->degraded
= 0;
2077 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2079 disk
= conf
->mirrors
+ i
;
2082 !test_bit(In_sync
, &rdev
->flags
)) {
2083 disk
->head_position
= 0;
2089 mddev
->thread
= md_register_thread(raid10d
, mddev
, "%s_raid10");
2090 if (!mddev
->thread
) {
2092 "raid10: couldn't allocate thread for %s\n",
2098 "raid10: raid set %s active with %d out of %d devices\n",
2099 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2102 * Ok, everything is just fine now
2104 if (conf
->far_offset
) {
2105 size
= mddev
->size
>> (conf
->chunk_shift
-1);
2106 size
*= conf
->raid_disks
;
2107 size
<<= conf
->chunk_shift
;
2108 sector_div(size
, conf
->far_copies
);
2110 size
= conf
->stride
* conf
->raid_disks
;
2111 sector_div(size
, conf
->near_copies
);
2112 mddev
->array_size
= size
/2;
2113 mddev
->resync_max_sectors
= size
;
2115 mddev
->queue
->unplug_fn
= raid10_unplug
;
2116 mddev
->queue
->issue_flush_fn
= raid10_issue_flush
;
2117 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
2118 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2120 /* Calculate max read-ahead size.
2121 * We need to readahead at least twice a whole stripe....
2125 int stripe
= conf
->raid_disks
* (mddev
->chunk_size
/ PAGE_SIZE
);
2126 stripe
/= conf
->near_copies
;
2127 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
2128 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
2131 if (conf
->near_copies
< mddev
->raid_disks
)
2132 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
2136 if (conf
->r10bio_pool
)
2137 mempool_destroy(conf
->r10bio_pool
);
2138 safe_put_page(conf
->tmppage
);
2139 kfree(conf
->mirrors
);
2141 mddev
->private = NULL
;
2146 static int stop(mddev_t
*mddev
)
2148 conf_t
*conf
= mddev_to_conf(mddev
);
2150 md_unregister_thread(mddev
->thread
);
2151 mddev
->thread
= NULL
;
2152 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
2153 if (conf
->r10bio_pool
)
2154 mempool_destroy(conf
->r10bio_pool
);
2155 kfree(conf
->mirrors
);
2157 mddev
->private = NULL
;
2161 static void raid10_quiesce(mddev_t
*mddev
, int state
)
2163 conf_t
*conf
= mddev_to_conf(mddev
);
2167 raise_barrier(conf
, 0);
2170 lower_barrier(conf
);
2173 if (mddev
->thread
) {
2175 mddev
->thread
->timeout
= mddev
->bitmap
->daemon_sleep
* HZ
;
2177 mddev
->thread
->timeout
= MAX_SCHEDULE_TIMEOUT
;
2178 md_wakeup_thread(mddev
->thread
);
2182 static struct mdk_personality raid10_personality
=
2186 .owner
= THIS_MODULE
,
2187 .make_request
= make_request
,
2191 .error_handler
= error
,
2192 .hot_add_disk
= raid10_add_disk
,
2193 .hot_remove_disk
= raid10_remove_disk
,
2194 .spare_active
= raid10_spare_active
,
2195 .sync_request
= sync_request
,
2196 .quiesce
= raid10_quiesce
,
2199 static int __init
raid_init(void)
2201 return register_md_personality(&raid10_personality
);
2204 static void raid_exit(void)
2206 unregister_md_personality(&raid10_personality
);
2209 module_init(raid_init
);
2210 module_exit(raid_exit
);
2211 MODULE_LICENSE("GPL");
2212 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2213 MODULE_ALIAS("md-raid10");
2214 MODULE_ALIAS("md-level-10");