2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for futher copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/seq_file.h>
31 * RAID10 provides a combination of RAID0 and RAID1 functionality.
32 * The layout of data is defined by
35 * near_copies (stored in low byte of layout)
36 * far_copies (stored in second byte of layout)
37 * far_offset (stored in bit 16 of layout )
39 * The data to be stored is divided into chunks using chunksize.
40 * Each device is divided into far_copies sections.
41 * In each section, chunks are laid out in a style similar to raid0, but
42 * near_copies copies of each chunk is stored (each on a different drive).
43 * The starting device for each section is offset near_copies from the starting
44 * device of the previous section.
45 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
47 * near_copies and far_copies must be at least one, and their product is at most
50 * If far_offset is true, then the far_copies are handled a bit differently.
51 * The copies are still in different stripes, but instead of be very far apart
52 * on disk, there are adjacent stripes.
56 * Number of guaranteed r10bios in case of extreme VM load:
58 #define NR_RAID10_BIOS 256
60 static void unplug_slaves(mddev_t
*mddev
);
62 static void allow_barrier(conf_t
*conf
);
63 static void lower_barrier(conf_t
*conf
);
65 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
69 int size
= offsetof(struct r10bio_s
, devs
[conf
->copies
]);
71 /* allocate a r10bio with room for raid_disks entries in the bios array */
72 r10_bio
= kzalloc(size
, gfp_flags
);
73 if (!r10_bio
&& conf
->mddev
)
74 unplug_slaves(conf
->mddev
);
79 static void r10bio_pool_free(void *r10_bio
, void *data
)
84 /* Maximum size of each resync request */
85 #define RESYNC_BLOCK_SIZE (64*1024)
86 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
87 /* amount of memory to reserve for resync requests */
88 #define RESYNC_WINDOW (1024*1024)
89 /* maximum number of concurrent requests, memory permitting */
90 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
93 * When performing a resync, we need to read and compare, so
94 * we need as many pages are there are copies.
95 * When performing a recovery, we need 2 bios, one for read,
96 * one for write (we recover only one drive per r10buf)
99 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
108 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
110 unplug_slaves(conf
->mddev
);
114 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
115 nalloc
= conf
->copies
; /* resync */
117 nalloc
= 2; /* recovery */
122 for (j
= nalloc
; j
-- ; ) {
123 bio
= bio_alloc(gfp_flags
, RESYNC_PAGES
);
126 r10_bio
->devs
[j
].bio
= bio
;
129 * Allocate RESYNC_PAGES data pages and attach them
132 for (j
= 0 ; j
< nalloc
; j
++) {
133 bio
= r10_bio
->devs
[j
].bio
;
134 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
135 page
= alloc_page(gfp_flags
);
139 bio
->bi_io_vec
[i
].bv_page
= page
;
147 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
149 for (i
= 0; i
< RESYNC_PAGES
; i
++)
150 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
153 while ( ++j
< nalloc
)
154 bio_put(r10_bio
->devs
[j
].bio
);
155 r10bio_pool_free(r10_bio
, conf
);
159 static void r10buf_pool_free(void *__r10_bio
, void *data
)
163 r10bio_t
*r10bio
= __r10_bio
;
166 for (j
=0; j
< conf
->copies
; j
++) {
167 struct bio
*bio
= r10bio
->devs
[j
].bio
;
169 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
170 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
171 bio
->bi_io_vec
[i
].bv_page
= NULL
;
176 r10bio_pool_free(r10bio
, conf
);
179 static void put_all_bios(conf_t
*conf
, r10bio_t
*r10_bio
)
183 for (i
= 0; i
< conf
->copies
; i
++) {
184 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
185 if (*bio
&& *bio
!= IO_BLOCKED
)
191 static void free_r10bio(r10bio_t
*r10_bio
)
193 conf_t
*conf
= r10_bio
->mddev
->private;
196 * Wake up any possible resync thread that waits for the device
201 put_all_bios(conf
, r10_bio
);
202 mempool_free(r10_bio
, conf
->r10bio_pool
);
205 static void put_buf(r10bio_t
*r10_bio
)
207 conf_t
*conf
= r10_bio
->mddev
->private;
209 mempool_free(r10_bio
, conf
->r10buf_pool
);
214 static void reschedule_retry(r10bio_t
*r10_bio
)
217 mddev_t
*mddev
= r10_bio
->mddev
;
218 conf_t
*conf
= mddev
->private;
220 spin_lock_irqsave(&conf
->device_lock
, flags
);
221 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
223 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
225 /* wake up frozen array... */
226 wake_up(&conf
->wait_barrier
);
228 md_wakeup_thread(mddev
->thread
);
232 * raid_end_bio_io() is called when we have finished servicing a mirrored
233 * operation and are ready to return a success/failure code to the buffer
236 static void raid_end_bio_io(r10bio_t
*r10_bio
)
238 struct bio
*bio
= r10_bio
->master_bio
;
241 test_bit(R10BIO_Uptodate
, &r10_bio
->state
) ? 0 : -EIO
);
242 free_r10bio(r10_bio
);
246 * Update disk head position estimator based on IRQ completion info.
248 static inline void update_head_pos(int slot
, r10bio_t
*r10_bio
)
250 conf_t
*conf
= r10_bio
->mddev
->private;
252 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
253 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
256 static void raid10_end_read_request(struct bio
*bio
, int error
)
258 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
259 r10bio_t
*r10_bio
= bio
->bi_private
;
261 conf_t
*conf
= r10_bio
->mddev
->private;
264 slot
= r10_bio
->read_slot
;
265 dev
= r10_bio
->devs
[slot
].devnum
;
267 * this branch is our 'one mirror IO has finished' event handler:
269 update_head_pos(slot
, r10_bio
);
273 * Set R10BIO_Uptodate in our master bio, so that
274 * we will return a good error code to the higher
275 * levels even if IO on some other mirrored buffer fails.
277 * The 'master' represents the composite IO operation to
278 * user-side. So if something waits for IO, then it will
279 * wait for the 'master' bio.
281 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
282 raid_end_bio_io(r10_bio
);
287 char b
[BDEVNAME_SIZE
];
288 if (printk_ratelimit())
289 printk(KERN_ERR
"md/raid10:%s: %s: rescheduling sector %llu\n",
291 bdevname(conf
->mirrors
[dev
].rdev
->bdev
,b
), (unsigned long long)r10_bio
->sector
);
292 reschedule_retry(r10_bio
);
295 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
298 static void raid10_end_write_request(struct bio
*bio
, int error
)
300 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
301 r10bio_t
*r10_bio
= bio
->bi_private
;
303 conf_t
*conf
= r10_bio
->mddev
->private;
305 for (slot
= 0; slot
< conf
->copies
; slot
++)
306 if (r10_bio
->devs
[slot
].bio
== bio
)
308 dev
= r10_bio
->devs
[slot
].devnum
;
311 * this branch is our 'one mirror IO has finished' event handler:
314 md_error(r10_bio
->mddev
, conf
->mirrors
[dev
].rdev
);
315 /* an I/O failed, we can't clear the bitmap */
316 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
319 * Set R10BIO_Uptodate in our master bio, so that
320 * we will return a good error code for to the higher
321 * levels even if IO on some other mirrored buffer fails.
323 * The 'master' represents the composite IO operation to
324 * user-side. So if something waits for IO, then it will
325 * wait for the 'master' bio.
327 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
329 update_head_pos(slot
, r10_bio
);
333 * Let's see if all mirrored write operations have finished
336 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
337 /* clear the bitmap if all writes complete successfully */
338 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
340 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
342 md_write_end(r10_bio
->mddev
);
343 raid_end_bio_io(r10_bio
);
346 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
351 * RAID10 layout manager
352 * Aswell as the chunksize and raid_disks count, there are two
353 * parameters: near_copies and far_copies.
354 * near_copies * far_copies must be <= raid_disks.
355 * Normally one of these will be 1.
356 * If both are 1, we get raid0.
357 * If near_copies == raid_disks, we get raid1.
359 * Chunks are layed out in raid0 style with near_copies copies of the
360 * first chunk, followed by near_copies copies of the next chunk and
362 * If far_copies > 1, then after 1/far_copies of the array has been assigned
363 * as described above, we start again with a device offset of near_copies.
364 * So we effectively have another copy of the whole array further down all
365 * the drives, but with blocks on different drives.
366 * With this layout, and block is never stored twice on the one device.
368 * raid10_find_phys finds the sector offset of a given virtual sector
369 * on each device that it is on.
371 * raid10_find_virt does the reverse mapping, from a device and a
372 * sector offset to a virtual address
375 static void raid10_find_phys(conf_t
*conf
, r10bio_t
*r10bio
)
385 /* now calculate first sector/dev */
386 chunk
= r10bio
->sector
>> conf
->chunk_shift
;
387 sector
= r10bio
->sector
& conf
->chunk_mask
;
389 chunk
*= conf
->near_copies
;
391 dev
= sector_div(stripe
, conf
->raid_disks
);
392 if (conf
->far_offset
)
393 stripe
*= conf
->far_copies
;
395 sector
+= stripe
<< conf
->chunk_shift
;
397 /* and calculate all the others */
398 for (n
=0; n
< conf
->near_copies
; n
++) {
401 r10bio
->devs
[slot
].addr
= sector
;
402 r10bio
->devs
[slot
].devnum
= d
;
405 for (f
= 1; f
< conf
->far_copies
; f
++) {
406 d
+= conf
->near_copies
;
407 if (d
>= conf
->raid_disks
)
408 d
-= conf
->raid_disks
;
410 r10bio
->devs
[slot
].devnum
= d
;
411 r10bio
->devs
[slot
].addr
= s
;
415 if (dev
>= conf
->raid_disks
) {
417 sector
+= (conf
->chunk_mask
+ 1);
420 BUG_ON(slot
!= conf
->copies
);
423 static sector_t
raid10_find_virt(conf_t
*conf
, sector_t sector
, int dev
)
425 sector_t offset
, chunk
, vchunk
;
427 offset
= sector
& conf
->chunk_mask
;
428 if (conf
->far_offset
) {
430 chunk
= sector
>> conf
->chunk_shift
;
431 fc
= sector_div(chunk
, conf
->far_copies
);
432 dev
-= fc
* conf
->near_copies
;
434 dev
+= conf
->raid_disks
;
436 while (sector
>= conf
->stride
) {
437 sector
-= conf
->stride
;
438 if (dev
< conf
->near_copies
)
439 dev
+= conf
->raid_disks
- conf
->near_copies
;
441 dev
-= conf
->near_copies
;
443 chunk
= sector
>> conf
->chunk_shift
;
445 vchunk
= chunk
* conf
->raid_disks
+ dev
;
446 sector_div(vchunk
, conf
->near_copies
);
447 return (vchunk
<< conf
->chunk_shift
) + offset
;
451 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
453 * @bvm: properties of new bio
454 * @biovec: the request that could be merged to it.
456 * Return amount of bytes we can accept at this offset
457 * If near_copies == raid_disk, there are no striping issues,
458 * but in that case, the function isn't called at all.
460 static int raid10_mergeable_bvec(struct request_queue
*q
,
461 struct bvec_merge_data
*bvm
,
462 struct bio_vec
*biovec
)
464 mddev_t
*mddev
= q
->queuedata
;
465 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
467 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
468 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
470 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
471 if (max
< 0) max
= 0; /* bio_add cannot handle a negative return */
472 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
473 return biovec
->bv_len
;
479 * This routine returns the disk from which the requested read should
480 * be done. There is a per-array 'next expected sequential IO' sector
481 * number - if this matches on the next IO then we use the last disk.
482 * There is also a per-disk 'last know head position' sector that is
483 * maintained from IRQ contexts, both the normal and the resync IO
484 * completion handlers update this position correctly. If there is no
485 * perfect sequential match then we pick the disk whose head is closest.
487 * If there are 2 mirrors in the same 2 devices, performance degrades
488 * because position is mirror, not device based.
490 * The rdev for the device selected will have nr_pending incremented.
494 * FIXME: possibly should rethink readbalancing and do it differently
495 * depending on near_copies / far_copies geometry.
497 static int read_balance(conf_t
*conf
, r10bio_t
*r10_bio
)
499 const sector_t this_sector
= r10_bio
->sector
;
500 int disk
, slot
, nslot
;
501 const int sectors
= r10_bio
->sectors
;
502 sector_t new_distance
, current_distance
;
505 raid10_find_phys(conf
, r10_bio
);
508 * Check if we can balance. We can balance on the whole
509 * device if no resync is going on (recovery is ok), or below
510 * the resync window. We take the first readable disk when
511 * above the resync window.
513 if (conf
->mddev
->recovery_cp
< MaxSector
514 && (this_sector
+ sectors
>= conf
->next_resync
)) {
515 /* make sure that disk is operational */
517 disk
= r10_bio
->devs
[slot
].devnum
;
519 while ((rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
520 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
521 !test_bit(In_sync
, &rdev
->flags
)) {
523 if (slot
== conf
->copies
) {
528 disk
= r10_bio
->devs
[slot
].devnum
;
534 /* make sure the disk is operational */
536 disk
= r10_bio
->devs
[slot
].devnum
;
537 while ((rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
)) == NULL
||
538 r10_bio
->devs
[slot
].bio
== IO_BLOCKED
||
539 !test_bit(In_sync
, &rdev
->flags
)) {
541 if (slot
== conf
->copies
) {
545 disk
= r10_bio
->devs
[slot
].devnum
;
549 current_distance
= abs(r10_bio
->devs
[slot
].addr
-
550 conf
->mirrors
[disk
].head_position
);
552 /* Find the disk whose head is closest,
553 * or - for far > 1 - find the closest to partition beginning */
555 for (nslot
= slot
; nslot
< conf
->copies
; nslot
++) {
556 int ndisk
= r10_bio
->devs
[nslot
].devnum
;
559 if ((rdev
=rcu_dereference(conf
->mirrors
[ndisk
].rdev
)) == NULL
||
560 r10_bio
->devs
[nslot
].bio
== IO_BLOCKED
||
561 !test_bit(In_sync
, &rdev
->flags
))
564 /* This optimisation is debatable, and completely destroys
565 * sequential read speed for 'far copies' arrays. So only
566 * keep it for 'near' arrays, and review those later.
568 if (conf
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
)) {
574 /* for far > 1 always use the lowest address */
575 if (conf
->far_copies
> 1)
576 new_distance
= r10_bio
->devs
[nslot
].addr
;
578 new_distance
= abs(r10_bio
->devs
[nslot
].addr
-
579 conf
->mirrors
[ndisk
].head_position
);
580 if (new_distance
< current_distance
) {
581 current_distance
= new_distance
;
588 r10_bio
->read_slot
= slot
;
589 /* conf->next_seq_sect = this_sector + sectors;*/
591 if (disk
>= 0 && (rdev
=rcu_dereference(conf
->mirrors
[disk
].rdev
))!= NULL
)
592 atomic_inc(&conf
->mirrors
[disk
].rdev
->nr_pending
);
600 static void unplug_slaves(mddev_t
*mddev
)
602 conf_t
*conf
= mddev
->private;
606 for (i
=0; i
< conf
->raid_disks
; i
++) {
607 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
608 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
609 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
611 atomic_inc(&rdev
->nr_pending
);
616 rdev_dec_pending(rdev
, mddev
);
623 static void raid10_unplug(struct request_queue
*q
)
625 mddev_t
*mddev
= q
->queuedata
;
627 unplug_slaves(q
->queuedata
);
628 md_wakeup_thread(mddev
->thread
);
631 static int raid10_congested(void *data
, int bits
)
633 mddev_t
*mddev
= data
;
634 conf_t
*conf
= mddev
->private;
637 if (mddev_congested(mddev
, bits
))
640 for (i
= 0; i
< conf
->raid_disks
&& ret
== 0; i
++) {
641 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
642 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
643 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
645 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
652 static int flush_pending_writes(conf_t
*conf
)
654 /* Any writes that have been queued but are awaiting
655 * bitmap updates get flushed here.
656 * We return 1 if any requests were actually submitted.
660 spin_lock_irq(&conf
->device_lock
);
662 if (conf
->pending_bio_list
.head
) {
664 bio
= bio_list_get(&conf
->pending_bio_list
);
665 blk_remove_plug(conf
->mddev
->queue
);
666 spin_unlock_irq(&conf
->device_lock
);
667 /* flush any pending bitmap writes to disk
668 * before proceeding w/ I/O */
669 bitmap_unplug(conf
->mddev
->bitmap
);
671 while (bio
) { /* submit pending writes */
672 struct bio
*next
= bio
->bi_next
;
674 generic_make_request(bio
);
679 spin_unlock_irq(&conf
->device_lock
);
683 * Sometimes we need to suspend IO while we do something else,
684 * either some resync/recovery, or reconfigure the array.
685 * To do this we raise a 'barrier'.
686 * The 'barrier' is a counter that can be raised multiple times
687 * to count how many activities are happening which preclude
689 * We can only raise the barrier if there is no pending IO.
690 * i.e. if nr_pending == 0.
691 * We choose only to raise the barrier if no-one is waiting for the
692 * barrier to go down. This means that as soon as an IO request
693 * is ready, no other operations which require a barrier will start
694 * until the IO request has had a chance.
696 * So: regular IO calls 'wait_barrier'. When that returns there
697 * is no backgroup IO happening, It must arrange to call
698 * allow_barrier when it has finished its IO.
699 * backgroup IO calls must call raise_barrier. Once that returns
700 * there is no normal IO happeing. It must arrange to call
701 * lower_barrier when the particular background IO completes.
704 static void raise_barrier(conf_t
*conf
, int force
)
706 BUG_ON(force
&& !conf
->barrier
);
707 spin_lock_irq(&conf
->resync_lock
);
709 /* Wait until no block IO is waiting (unless 'force') */
710 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
712 raid10_unplug(conf
->mddev
->queue
));
714 /* block any new IO from starting */
717 /* No wait for all pending IO to complete */
718 wait_event_lock_irq(conf
->wait_barrier
,
719 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
721 raid10_unplug(conf
->mddev
->queue
));
723 spin_unlock_irq(&conf
->resync_lock
);
726 static void lower_barrier(conf_t
*conf
)
729 spin_lock_irqsave(&conf
->resync_lock
, flags
);
731 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
732 wake_up(&conf
->wait_barrier
);
735 static void wait_barrier(conf_t
*conf
)
737 spin_lock_irq(&conf
->resync_lock
);
740 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
742 raid10_unplug(conf
->mddev
->queue
));
746 spin_unlock_irq(&conf
->resync_lock
);
749 static void allow_barrier(conf_t
*conf
)
752 spin_lock_irqsave(&conf
->resync_lock
, flags
);
754 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
755 wake_up(&conf
->wait_barrier
);
758 static void freeze_array(conf_t
*conf
)
760 /* stop syncio and normal IO and wait for everything to
762 * We increment barrier and nr_waiting, and then
763 * wait until nr_pending match nr_queued+1
764 * This is called in the context of one normal IO request
765 * that has failed. Thus any sync request that might be pending
766 * will be blocked by nr_pending, and we need to wait for
767 * pending IO requests to complete or be queued for re-try.
768 * Thus the number queued (nr_queued) plus this request (1)
769 * must match the number of pending IOs (nr_pending) before
772 spin_lock_irq(&conf
->resync_lock
);
775 wait_event_lock_irq(conf
->wait_barrier
,
776 conf
->nr_pending
== conf
->nr_queued
+1,
778 ({ flush_pending_writes(conf
);
779 raid10_unplug(conf
->mddev
->queue
); }));
780 spin_unlock_irq(&conf
->resync_lock
);
783 static void unfreeze_array(conf_t
*conf
)
785 /* reverse the effect of the freeze */
786 spin_lock_irq(&conf
->resync_lock
);
789 wake_up(&conf
->wait_barrier
);
790 spin_unlock_irq(&conf
->resync_lock
);
793 static int make_request(mddev_t
*mddev
, struct bio
* bio
)
795 conf_t
*conf
= mddev
->private;
796 mirror_info_t
*mirror
;
798 struct bio
*read_bio
;
800 int chunk_sects
= conf
->chunk_mask
+ 1;
801 const int rw
= bio_data_dir(bio
);
802 const bool do_sync
= bio_rw_flagged(bio
, BIO_RW_SYNCIO
);
805 mdk_rdev_t
*blocked_rdev
;
807 if (unlikely(bio_rw_flagged(bio
, BIO_RW_BARRIER
))) {
808 md_barrier_request(mddev
, bio
);
812 /* If this request crosses a chunk boundary, we need to
813 * split it. This will only happen for 1 PAGE (or less) requests.
815 if (unlikely( (bio
->bi_sector
& conf
->chunk_mask
) + (bio
->bi_size
>> 9)
817 conf
->near_copies
< conf
->raid_disks
)) {
819 /* Sanity check -- queue functions should prevent this happening */
820 if (bio
->bi_vcnt
!= 1 ||
823 /* This is a one page bio that upper layers
824 * refuse to split for us, so we need to split it.
827 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
828 if (make_request(mddev
, &bp
->bio1
))
829 generic_make_request(&bp
->bio1
);
830 if (make_request(mddev
, &bp
->bio2
))
831 generic_make_request(&bp
->bio2
);
833 bio_pair_release(bp
);
836 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
837 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
838 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
844 md_write_start(mddev
, bio
);
847 * Register the new request and wait if the reconstruction
848 * thread has put up a bar for new requests.
849 * Continue immediately if no resync is active currently.
853 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
855 r10_bio
->master_bio
= bio
;
856 r10_bio
->sectors
= bio
->bi_size
>> 9;
858 r10_bio
->mddev
= mddev
;
859 r10_bio
->sector
= bio
->bi_sector
;
864 * read balancing logic:
866 int disk
= read_balance(conf
, r10_bio
);
867 int slot
= r10_bio
->read_slot
;
869 raid_end_bio_io(r10_bio
);
872 mirror
= conf
->mirrors
+ disk
;
874 read_bio
= bio_clone(bio
, GFP_NOIO
);
876 r10_bio
->devs
[slot
].bio
= read_bio
;
878 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
879 mirror
->rdev
->data_offset
;
880 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
881 read_bio
->bi_end_io
= raid10_end_read_request
;
882 read_bio
->bi_rw
= READ
| (do_sync
<< BIO_RW_SYNCIO
);
883 read_bio
->bi_private
= r10_bio
;
885 generic_make_request(read_bio
);
892 /* first select target devices under rcu_lock and
893 * inc refcount on their rdev. Record them by setting
896 raid10_find_phys(conf
, r10_bio
);
900 for (i
= 0; i
< conf
->copies
; i
++) {
901 int d
= r10_bio
->devs
[i
].devnum
;
902 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
903 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
904 atomic_inc(&rdev
->nr_pending
);
908 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
909 atomic_inc(&rdev
->nr_pending
);
910 r10_bio
->devs
[i
].bio
= bio
;
912 r10_bio
->devs
[i
].bio
= NULL
;
913 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
918 if (unlikely(blocked_rdev
)) {
919 /* Have to wait for this device to get unblocked, then retry */
923 for (j
= 0; j
< i
; j
++)
924 if (r10_bio
->devs
[j
].bio
) {
925 d
= r10_bio
->devs
[j
].devnum
;
926 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
929 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
934 atomic_set(&r10_bio
->remaining
, 0);
937 for (i
= 0; i
< conf
->copies
; i
++) {
939 int d
= r10_bio
->devs
[i
].devnum
;
940 if (!r10_bio
->devs
[i
].bio
)
943 mbio
= bio_clone(bio
, GFP_NOIO
);
944 r10_bio
->devs
[i
].bio
= mbio
;
946 mbio
->bi_sector
= r10_bio
->devs
[i
].addr
+
947 conf
->mirrors
[d
].rdev
->data_offset
;
948 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
949 mbio
->bi_end_io
= raid10_end_write_request
;
950 mbio
->bi_rw
= WRITE
| (do_sync
<< BIO_RW_SYNCIO
);
951 mbio
->bi_private
= r10_bio
;
953 atomic_inc(&r10_bio
->remaining
);
954 bio_list_add(&bl
, mbio
);
957 if (unlikely(!atomic_read(&r10_bio
->remaining
))) {
958 /* the array is dead */
960 raid_end_bio_io(r10_bio
);
964 bitmap_startwrite(mddev
->bitmap
, bio
->bi_sector
, r10_bio
->sectors
, 0);
965 spin_lock_irqsave(&conf
->device_lock
, flags
);
966 bio_list_merge(&conf
->pending_bio_list
, &bl
);
967 blk_plug_device(mddev
->queue
);
968 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
970 /* In case raid10d snuck in to freeze_array */
971 wake_up(&conf
->wait_barrier
);
974 md_wakeup_thread(mddev
->thread
);
979 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
981 conf_t
*conf
= mddev
->private;
984 if (conf
->near_copies
< conf
->raid_disks
)
985 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
986 if (conf
->near_copies
> 1)
987 seq_printf(seq
, " %d near-copies", conf
->near_copies
);
988 if (conf
->far_copies
> 1) {
989 if (conf
->far_offset
)
990 seq_printf(seq
, " %d offset-copies", conf
->far_copies
);
992 seq_printf(seq
, " %d far-copies", conf
->far_copies
);
994 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
995 conf
->raid_disks
- mddev
->degraded
);
996 for (i
= 0; i
< conf
->raid_disks
; i
++)
997 seq_printf(seq
, "%s",
998 conf
->mirrors
[i
].rdev
&&
999 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1000 seq_printf(seq
, "]");
1003 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1005 char b
[BDEVNAME_SIZE
];
1006 conf_t
*conf
= mddev
->private;
1009 * If it is not operational, then we have already marked it as dead
1010 * else if it is the last working disks, ignore the error, let the
1011 * next level up know.
1012 * else mark the drive as failed
1014 if (test_bit(In_sync
, &rdev
->flags
)
1015 && conf
->raid_disks
-mddev
->degraded
== 1)
1017 * Don't fail the drive, just return an IO error.
1018 * The test should really be more sophisticated than
1019 * "working_disks == 1", but it isn't critical, and
1020 * can wait until we do more sophisticated "is the drive
1021 * really dead" tests...
1024 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1025 unsigned long flags
;
1026 spin_lock_irqsave(&conf
->device_lock
, flags
);
1028 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1030 * if recovery is running, make sure it aborts.
1032 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1034 set_bit(Faulty
, &rdev
->flags
);
1035 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1036 printk(KERN_ALERT
"md/raid10:%s: Disk failure on %s, disabling device.\n"
1037 KERN_ALERT
"md/raid10:%s: Operation continuing on %d devices.\n",
1038 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1039 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1042 static void print_conf(conf_t
*conf
)
1047 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1049 printk(KERN_DEBUG
"(!conf)\n");
1052 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1055 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1056 char b
[BDEVNAME_SIZE
];
1057 tmp
= conf
->mirrors
+ i
;
1059 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1060 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1061 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1062 bdevname(tmp
->rdev
->bdev
,b
));
1066 static void close_sync(conf_t
*conf
)
1069 allow_barrier(conf
);
1071 mempool_destroy(conf
->r10buf_pool
);
1072 conf
->r10buf_pool
= NULL
;
1075 /* check if there are enough drives for
1076 * every block to appear on atleast one
1078 static int enough(conf_t
*conf
)
1083 int n
= conf
->copies
;
1086 if (conf
->mirrors
[first
].rdev
)
1088 first
= (first
+1) % conf
->raid_disks
;
1092 } while (first
!= 0);
1096 static int raid10_spare_active(mddev_t
*mddev
)
1099 conf_t
*conf
= mddev
->private;
1103 * Find all non-in_sync disks within the RAID10 configuration
1104 * and mark them in_sync
1106 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1107 tmp
= conf
->mirrors
+ i
;
1109 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1110 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1111 unsigned long flags
;
1112 spin_lock_irqsave(&conf
->device_lock
, flags
);
1114 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1123 static int raid10_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1125 conf_t
*conf
= mddev
->private;
1130 int last
= conf
->raid_disks
- 1;
1132 if (mddev
->recovery_cp
< MaxSector
)
1133 /* only hot-add to in-sync arrays, as recovery is
1134 * very different from resync
1140 if (rdev
->raid_disk
>= 0)
1141 first
= last
= rdev
->raid_disk
;
1143 if (rdev
->saved_raid_disk
>= 0 &&
1144 rdev
->saved_raid_disk
>= first
&&
1145 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1146 mirror
= rdev
->saved_raid_disk
;
1149 for ( ; mirror
<= last
; mirror
++)
1150 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1152 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1153 rdev
->data_offset
<< 9);
1154 /* as we don't honour merge_bvec_fn, we must
1155 * never risk violating it, so limit
1156 * ->max_segments to one lying with a single
1157 * page, as a one page request is never in
1160 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1161 blk_queue_max_segments(mddev
->queue
, 1);
1162 blk_queue_segment_boundary(mddev
->queue
,
1163 PAGE_CACHE_SIZE
- 1);
1166 p
->head_position
= 0;
1167 rdev
->raid_disk
= mirror
;
1169 if (rdev
->saved_raid_disk
!= mirror
)
1171 rcu_assign_pointer(p
->rdev
, rdev
);
1175 md_integrity_add_rdev(rdev
, mddev
);
1180 static int raid10_remove_disk(mddev_t
*mddev
, int number
)
1182 conf_t
*conf
= mddev
->private;
1185 mirror_info_t
*p
= conf
->mirrors
+ number
;
1190 if (test_bit(In_sync
, &rdev
->flags
) ||
1191 atomic_read(&rdev
->nr_pending
)) {
1195 /* Only remove faulty devices in recovery
1198 if (!test_bit(Faulty
, &rdev
->flags
) &&
1205 if (atomic_read(&rdev
->nr_pending
)) {
1206 /* lost the race, try later */
1211 md_integrity_register(mddev
);
1220 static void end_sync_read(struct bio
*bio
, int error
)
1222 r10bio_t
*r10_bio
= bio
->bi_private
;
1223 conf_t
*conf
= r10_bio
->mddev
->private;
1226 for (i
=0; i
<conf
->copies
; i
++)
1227 if (r10_bio
->devs
[i
].bio
== bio
)
1229 BUG_ON(i
== conf
->copies
);
1230 update_head_pos(i
, r10_bio
);
1231 d
= r10_bio
->devs
[i
].devnum
;
1233 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1234 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1236 atomic_add(r10_bio
->sectors
,
1237 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1238 if (!test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
))
1239 md_error(r10_bio
->mddev
,
1240 conf
->mirrors
[d
].rdev
);
1243 /* for reconstruct, we always reschedule after a read.
1244 * for resync, only after all reads
1246 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1247 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1248 atomic_dec_and_test(&r10_bio
->remaining
)) {
1249 /* we have read all the blocks,
1250 * do the comparison in process context in raid10d
1252 reschedule_retry(r10_bio
);
1256 static void end_sync_write(struct bio
*bio
, int error
)
1258 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1259 r10bio_t
*r10_bio
= bio
->bi_private
;
1260 mddev_t
*mddev
= r10_bio
->mddev
;
1261 conf_t
*conf
= mddev
->private;
1264 for (i
= 0; i
< conf
->copies
; i
++)
1265 if (r10_bio
->devs
[i
].bio
== bio
)
1267 d
= r10_bio
->devs
[i
].devnum
;
1270 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1272 update_head_pos(i
, r10_bio
);
1274 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1275 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1276 if (r10_bio
->master_bio
== NULL
) {
1277 /* the primary of several recovery bios */
1278 sector_t s
= r10_bio
->sectors
;
1280 md_done_sync(mddev
, s
, 1);
1283 r10bio_t
*r10_bio2
= (r10bio_t
*)r10_bio
->master_bio
;
1291 * Note: sync and recover and handled very differently for raid10
1292 * This code is for resync.
1293 * For resync, we read through virtual addresses and read all blocks.
1294 * If there is any error, we schedule a write. The lowest numbered
1295 * drive is authoritative.
1296 * However requests come for physical address, so we need to map.
1297 * For every physical address there are raid_disks/copies virtual addresses,
1298 * which is always are least one, but is not necessarly an integer.
1299 * This means that a physical address can span multiple chunks, so we may
1300 * have to submit multiple io requests for a single sync request.
1303 * We check if all blocks are in-sync and only write to blocks that
1306 static void sync_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1308 conf_t
*conf
= mddev
->private;
1310 struct bio
*tbio
, *fbio
;
1312 atomic_set(&r10_bio
->remaining
, 1);
1314 /* find the first device with a block */
1315 for (i
=0; i
<conf
->copies
; i
++)
1316 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1319 if (i
== conf
->copies
)
1323 fbio
= r10_bio
->devs
[i
].bio
;
1325 /* now find blocks with errors */
1326 for (i
=0 ; i
< conf
->copies
; i
++) {
1328 int vcnt
= r10_bio
->sectors
>> (PAGE_SHIFT
-9);
1330 tbio
= r10_bio
->devs
[i
].bio
;
1332 if (tbio
->bi_end_io
!= end_sync_read
)
1336 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1337 /* We know that the bi_io_vec layout is the same for
1338 * both 'first' and 'i', so we just compare them.
1339 * All vec entries are PAGE_SIZE;
1341 for (j
= 0; j
< vcnt
; j
++)
1342 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1343 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1348 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1350 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1351 /* Don't fix anything. */
1353 /* Ok, we need to write this bio
1354 * First we need to fixup bv_offset, bv_len and
1355 * bi_vecs, as the read request might have corrupted these
1357 tbio
->bi_vcnt
= vcnt
;
1358 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1360 tbio
->bi_phys_segments
= 0;
1361 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1362 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1363 tbio
->bi_next
= NULL
;
1364 tbio
->bi_rw
= WRITE
;
1365 tbio
->bi_private
= r10_bio
;
1366 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1368 for (j
=0; j
< vcnt
; j
++) {
1369 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1370 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1372 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1373 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1376 tbio
->bi_end_io
= end_sync_write
;
1378 d
= r10_bio
->devs
[i
].devnum
;
1379 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1380 atomic_inc(&r10_bio
->remaining
);
1381 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1383 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1384 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1385 generic_make_request(tbio
);
1389 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
1390 md_done_sync(mddev
, r10_bio
->sectors
, 1);
1396 * Now for the recovery code.
1397 * Recovery happens across physical sectors.
1398 * We recover all non-is_sync drives by finding the virtual address of
1399 * each, and then choose a working drive that also has that virt address.
1400 * There is a separate r10_bio for each non-in_sync drive.
1401 * Only the first two slots are in use. The first for reading,
1402 * The second for writing.
1406 static void recovery_request_write(mddev_t
*mddev
, r10bio_t
*r10_bio
)
1408 conf_t
*conf
= mddev
->private;
1410 struct bio
*bio
, *wbio
;
1413 /* move the pages across to the second bio
1414 * and submit the write request
1416 bio
= r10_bio
->devs
[0].bio
;
1417 wbio
= r10_bio
->devs
[1].bio
;
1418 for (i
=0; i
< wbio
->bi_vcnt
; i
++) {
1419 struct page
*p
= bio
->bi_io_vec
[i
].bv_page
;
1420 bio
->bi_io_vec
[i
].bv_page
= wbio
->bi_io_vec
[i
].bv_page
;
1421 wbio
->bi_io_vec
[i
].bv_page
= p
;
1423 d
= r10_bio
->devs
[1].devnum
;
1425 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1426 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
1427 if (test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
1428 generic_make_request(wbio
);
1430 bio_endio(wbio
, -EIO
);
1435 * Used by fix_read_error() to decay the per rdev read_errors.
1436 * We halve the read error count for every hour that has elapsed
1437 * since the last recorded read error.
1440 static void check_decay_read_errors(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1442 struct timespec cur_time_mon
;
1443 unsigned long hours_since_last
;
1444 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
1446 ktime_get_ts(&cur_time_mon
);
1448 if (rdev
->last_read_error
.tv_sec
== 0 &&
1449 rdev
->last_read_error
.tv_nsec
== 0) {
1450 /* first time we've seen a read error */
1451 rdev
->last_read_error
= cur_time_mon
;
1455 hours_since_last
= (cur_time_mon
.tv_sec
-
1456 rdev
->last_read_error
.tv_sec
) / 3600;
1458 rdev
->last_read_error
= cur_time_mon
;
1461 * if hours_since_last is > the number of bits in read_errors
1462 * just set read errors to 0. We do this to avoid
1463 * overflowing the shift of read_errors by hours_since_last.
1465 if (hours_since_last
>= 8 * sizeof(read_errors
))
1466 atomic_set(&rdev
->read_errors
, 0);
1468 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
1472 * This is a kernel thread which:
1474 * 1. Retries failed read operations on working mirrors.
1475 * 2. Updates the raid superblock when problems encounter.
1476 * 3. Performs writes following reads for array synchronising.
1479 static void fix_read_error(conf_t
*conf
, mddev_t
*mddev
, r10bio_t
*r10_bio
)
1481 int sect
= 0; /* Offset from r10_bio->sector */
1482 int sectors
= r10_bio
->sectors
;
1484 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
1485 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1488 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1489 if (rdev
) { /* If rdev is not NULL */
1490 char b
[BDEVNAME_SIZE
];
1491 int cur_read_error_count
= 0;
1493 bdevname(rdev
->bdev
, b
);
1495 if (test_bit(Faulty
, &rdev
->flags
)) {
1497 /* drive has already been failed, just ignore any
1498 more fix_read_error() attempts */
1502 check_decay_read_errors(mddev
, rdev
);
1503 atomic_inc(&rdev
->read_errors
);
1504 cur_read_error_count
= atomic_read(&rdev
->read_errors
);
1505 if (cur_read_error_count
> max_read_errors
) {
1508 "md/raid10:%s: %s: Raid device exceeded "
1509 "read_error threshold "
1510 "[cur %d:max %d]\n",
1512 b
, cur_read_error_count
, max_read_errors
);
1514 "md/raid10:%s: %s: Failing raid "
1515 "device\n", mdname(mddev
), b
);
1516 md_error(mddev
, conf
->mirrors
[d
].rdev
);
1524 int sl
= r10_bio
->read_slot
;
1528 if (s
> (PAGE_SIZE
>>9))
1533 d
= r10_bio
->devs
[sl
].devnum
;
1534 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1536 test_bit(In_sync
, &rdev
->flags
)) {
1537 atomic_inc(&rdev
->nr_pending
);
1539 success
= sync_page_io(rdev
->bdev
,
1540 r10_bio
->devs
[sl
].addr
+
1541 sect
+ rdev
->data_offset
,
1543 conf
->tmppage
, READ
);
1544 rdev_dec_pending(rdev
, mddev
);
1550 if (sl
== conf
->copies
)
1552 } while (!success
&& sl
!= r10_bio
->read_slot
);
1556 /* Cannot read from anywhere -- bye bye array */
1557 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
1558 md_error(mddev
, conf
->mirrors
[dn
].rdev
);
1563 /* write it back and re-read */
1565 while (sl
!= r10_bio
->read_slot
) {
1566 char b
[BDEVNAME_SIZE
];
1571 d
= r10_bio
->devs
[sl
].devnum
;
1572 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1574 test_bit(In_sync
, &rdev
->flags
)) {
1575 atomic_inc(&rdev
->nr_pending
);
1577 atomic_add(s
, &rdev
->corrected_errors
);
1578 if (sync_page_io(rdev
->bdev
,
1579 r10_bio
->devs
[sl
].addr
+
1580 sect
+ rdev
->data_offset
,
1581 s
<<9, conf
->tmppage
, WRITE
)
1583 /* Well, this device is dead */
1585 "md/raid10:%s: read correction "
1587 " (%d sectors at %llu on %s)\n",
1589 (unsigned long long)(sect
+
1591 bdevname(rdev
->bdev
, b
));
1592 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
1595 bdevname(rdev
->bdev
, b
));
1596 md_error(mddev
, rdev
);
1598 rdev_dec_pending(rdev
, mddev
);
1603 while (sl
!= r10_bio
->read_slot
) {
1608 d
= r10_bio
->devs
[sl
].devnum
;
1609 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1611 test_bit(In_sync
, &rdev
->flags
)) {
1612 char b
[BDEVNAME_SIZE
];
1613 atomic_inc(&rdev
->nr_pending
);
1615 if (sync_page_io(rdev
->bdev
,
1616 r10_bio
->devs
[sl
].addr
+
1617 sect
+ rdev
->data_offset
,
1618 s
<<9, conf
->tmppage
,
1620 /* Well, this device is dead */
1622 "md/raid10:%s: unable to read back "
1624 " (%d sectors at %llu on %s)\n",
1626 (unsigned long long)(sect
+
1628 bdevname(rdev
->bdev
, b
));
1629 printk(KERN_NOTICE
"md/raid10:%s: %s: failing drive\n",
1631 bdevname(rdev
->bdev
, b
));
1633 md_error(mddev
, rdev
);
1636 "md/raid10:%s: read error corrected"
1637 " (%d sectors at %llu on %s)\n",
1639 (unsigned long long)(sect
+
1641 bdevname(rdev
->bdev
, b
));
1644 rdev_dec_pending(rdev
, mddev
);
1655 static void raid10d(mddev_t
*mddev
)
1659 unsigned long flags
;
1660 conf_t
*conf
= mddev
->private;
1661 struct list_head
*head
= &conf
->retry_list
;
1665 md_check_recovery(mddev
);
1668 char b
[BDEVNAME_SIZE
];
1670 unplug
+= flush_pending_writes(conf
);
1672 spin_lock_irqsave(&conf
->device_lock
, flags
);
1673 if (list_empty(head
)) {
1674 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1677 r10_bio
= list_entry(head
->prev
, r10bio_t
, retry_list
);
1678 list_del(head
->prev
);
1680 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1682 mddev
= r10_bio
->mddev
;
1683 conf
= mddev
->private;
1684 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
)) {
1685 sync_request_write(mddev
, r10_bio
);
1687 } else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
1688 recovery_request_write(mddev
, r10_bio
);
1692 /* we got a read error. Maybe the drive is bad. Maybe just
1693 * the block and we can fix it.
1694 * We freeze all other IO, and try reading the block from
1695 * other devices. When we find one, we re-write
1696 * and check it that fixes the read error.
1697 * This is all done synchronously while the array is
1700 if (mddev
->ro
== 0) {
1702 fix_read_error(conf
, mddev
, r10_bio
);
1703 unfreeze_array(conf
);
1706 bio
= r10_bio
->devs
[r10_bio
->read_slot
].bio
;
1707 r10_bio
->devs
[r10_bio
->read_slot
].bio
=
1708 mddev
->ro
? IO_BLOCKED
: NULL
;
1709 mirror
= read_balance(conf
, r10_bio
);
1711 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
1712 " read error for block %llu\n",
1714 bdevname(bio
->bi_bdev
,b
),
1715 (unsigned long long)r10_bio
->sector
);
1716 raid_end_bio_io(r10_bio
);
1719 const bool do_sync
= bio_rw_flagged(r10_bio
->master_bio
, BIO_RW_SYNCIO
);
1721 rdev
= conf
->mirrors
[mirror
].rdev
;
1722 if (printk_ratelimit())
1723 printk(KERN_ERR
"md/raid10:%s: %s: redirecting sector %llu to"
1724 " another mirror\n",
1726 bdevname(rdev
->bdev
,b
),
1727 (unsigned long long)r10_bio
->sector
);
1728 bio
= bio_clone(r10_bio
->master_bio
, GFP_NOIO
);
1729 r10_bio
->devs
[r10_bio
->read_slot
].bio
= bio
;
1730 bio
->bi_sector
= r10_bio
->devs
[r10_bio
->read_slot
].addr
1731 + rdev
->data_offset
;
1732 bio
->bi_bdev
= rdev
->bdev
;
1733 bio
->bi_rw
= READ
| (do_sync
<< BIO_RW_SYNCIO
);
1734 bio
->bi_private
= r10_bio
;
1735 bio
->bi_end_io
= raid10_end_read_request
;
1737 generic_make_request(bio
);
1743 unplug_slaves(mddev
);
1747 static int init_resync(conf_t
*conf
)
1751 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1752 BUG_ON(conf
->r10buf_pool
);
1753 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
1754 if (!conf
->r10buf_pool
)
1756 conf
->next_resync
= 0;
1761 * perform a "sync" on one "block"
1763 * We need to make sure that no normal I/O request - particularly write
1764 * requests - conflict with active sync requests.
1766 * This is achieved by tracking pending requests and a 'barrier' concept
1767 * that can be installed to exclude normal IO requests.
1769 * Resync and recovery are handled very differently.
1770 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1772 * For resync, we iterate over virtual addresses, read all copies,
1773 * and update if there are differences. If only one copy is live,
1775 * For recovery, we iterate over physical addresses, read a good
1776 * value for each non-in_sync drive, and over-write.
1778 * So, for recovery we may have several outstanding complex requests for a
1779 * given address, one for each out-of-sync device. We model this by allocating
1780 * a number of r10_bio structures, one for each out-of-sync device.
1781 * As we setup these structures, we collect all bio's together into a list
1782 * which we then process collectively to add pages, and then process again
1783 * to pass to generic_make_request.
1785 * The r10_bio structures are linked using a borrowed master_bio pointer.
1786 * This link is counted in ->remaining. When the r10_bio that points to NULL
1787 * has its remaining count decremented to 0, the whole complex operation
1792 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
1794 conf_t
*conf
= mddev
->private;
1796 struct bio
*biolist
= NULL
, *bio
;
1797 sector_t max_sector
, nr_sectors
;
1803 sector_t sectors_skipped
= 0;
1804 int chunks_skipped
= 0;
1806 if (!conf
->r10buf_pool
)
1807 if (init_resync(conf
))
1811 max_sector
= mddev
->dev_sectors
;
1812 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1813 max_sector
= mddev
->resync_max_sectors
;
1814 if (sector_nr
>= max_sector
) {
1815 /* If we aborted, we need to abort the
1816 * sync on the 'current' bitmap chucks (there can
1817 * be several when recovering multiple devices).
1818 * as we may have started syncing it but not finished.
1819 * We can find the current address in
1820 * mddev->curr_resync, but for recovery,
1821 * we need to convert that to several
1822 * virtual addresses.
1824 if (mddev
->curr_resync
< max_sector
) { /* aborted */
1825 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
1826 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
1828 else for (i
=0; i
<conf
->raid_disks
; i
++) {
1830 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
1831 bitmap_end_sync(mddev
->bitmap
, sect
,
1834 } else /* completed sync */
1837 bitmap_close_sync(mddev
->bitmap
);
1840 return sectors_skipped
;
1842 if (chunks_skipped
>= conf
->raid_disks
) {
1843 /* if there has been nothing to do on any drive,
1844 * then there is nothing to do at all..
1847 return (max_sector
- sector_nr
) + sectors_skipped
;
1850 if (max_sector
> mddev
->resync_max
)
1851 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
1853 /* make sure whole request will fit in a chunk - if chunks
1856 if (conf
->near_copies
< conf
->raid_disks
&&
1857 max_sector
> (sector_nr
| conf
->chunk_mask
))
1858 max_sector
= (sector_nr
| conf
->chunk_mask
) + 1;
1860 * If there is non-resync activity waiting for us then
1861 * put in a delay to throttle resync.
1863 if (!go_faster
&& conf
->nr_waiting
)
1864 msleep_interruptible(1000);
1866 /* Again, very different code for resync and recovery.
1867 * Both must result in an r10bio with a list of bios that
1868 * have bi_end_io, bi_sector, bi_bdev set,
1869 * and bi_private set to the r10bio.
1870 * For recovery, we may actually create several r10bios
1871 * with 2 bios in each, that correspond to the bios in the main one.
1872 * In this case, the subordinate r10bios link back through a
1873 * borrowed master_bio pointer, and the counter in the master
1874 * includes a ref from each subordinate.
1876 /* First, we decide what to do and set ->bi_end_io
1877 * To end_sync_read if we want to read, and
1878 * end_sync_write if we will want to write.
1881 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
1882 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
1883 /* recovery... the complicated one */
1887 for (i
=0 ; i
<conf
->raid_disks
; i
++)
1888 if (conf
->mirrors
[i
].rdev
&&
1889 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
)) {
1890 int still_degraded
= 0;
1891 /* want to reconstruct this device */
1892 r10bio_t
*rb2
= r10_bio
;
1893 sector_t sect
= raid10_find_virt(conf
, sector_nr
, i
);
1895 /* Unless we are doing a full sync, we only need
1896 * to recover the block if it is set in the bitmap
1898 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1900 if (sync_blocks
< max_sync
)
1901 max_sync
= sync_blocks
;
1904 /* yep, skip the sync_blocks here, but don't assume
1905 * that there will never be anything to do here
1907 chunks_skipped
= -1;
1911 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
1912 raise_barrier(conf
, rb2
!= NULL
);
1913 atomic_set(&r10_bio
->remaining
, 0);
1915 r10_bio
->master_bio
= (struct bio
*)rb2
;
1917 atomic_inc(&rb2
->remaining
);
1918 r10_bio
->mddev
= mddev
;
1919 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
1920 r10_bio
->sector
= sect
;
1922 raid10_find_phys(conf
, r10_bio
);
1924 /* Need to check if the array will still be
1927 for (j
=0; j
<conf
->raid_disks
; j
++)
1928 if (conf
->mirrors
[j
].rdev
== NULL
||
1929 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
1934 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
1935 &sync_blocks
, still_degraded
);
1937 for (j
=0; j
<conf
->copies
;j
++) {
1938 int d
= r10_bio
->devs
[j
].devnum
;
1939 if (conf
->mirrors
[d
].rdev
&&
1940 test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
)) {
1941 /* This is where we read from */
1942 bio
= r10_bio
->devs
[0].bio
;
1943 bio
->bi_next
= biolist
;
1945 bio
->bi_private
= r10_bio
;
1946 bio
->bi_end_io
= end_sync_read
;
1948 bio
->bi_sector
= r10_bio
->devs
[j
].addr
+
1949 conf
->mirrors
[d
].rdev
->data_offset
;
1950 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1951 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1952 atomic_inc(&r10_bio
->remaining
);
1953 /* and we write to 'i' */
1955 for (k
=0; k
<conf
->copies
; k
++)
1956 if (r10_bio
->devs
[k
].devnum
== i
)
1958 BUG_ON(k
== conf
->copies
);
1959 bio
= r10_bio
->devs
[1].bio
;
1960 bio
->bi_next
= biolist
;
1962 bio
->bi_private
= r10_bio
;
1963 bio
->bi_end_io
= end_sync_write
;
1965 bio
->bi_sector
= r10_bio
->devs
[k
].addr
+
1966 conf
->mirrors
[i
].rdev
->data_offset
;
1967 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1969 r10_bio
->devs
[0].devnum
= d
;
1970 r10_bio
->devs
[1].devnum
= i
;
1975 if (j
== conf
->copies
) {
1976 /* Cannot recover, so abort the recovery */
1979 atomic_dec(&rb2
->remaining
);
1981 if (!test_and_set_bit(MD_RECOVERY_INTR
,
1983 printk(KERN_INFO
"md/raid10:%s: insufficient "
1984 "working devices for recovery.\n",
1989 if (biolist
== NULL
) {
1991 r10bio_t
*rb2
= r10_bio
;
1992 r10_bio
= (r10bio_t
*) rb2
->master_bio
;
1993 rb2
->master_bio
= NULL
;
1999 /* resync. Schedule a read for every block at this virt offset */
2002 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
2004 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2005 &sync_blocks
, mddev
->degraded
) &&
2006 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2007 /* We can skip this block */
2009 return sync_blocks
+ sectors_skipped
;
2011 if (sync_blocks
< max_sync
)
2012 max_sync
= sync_blocks
;
2013 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2015 r10_bio
->mddev
= mddev
;
2016 atomic_set(&r10_bio
->remaining
, 0);
2017 raise_barrier(conf
, 0);
2018 conf
->next_resync
= sector_nr
;
2020 r10_bio
->master_bio
= NULL
;
2021 r10_bio
->sector
= sector_nr
;
2022 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
2023 raid10_find_phys(conf
, r10_bio
);
2024 r10_bio
->sectors
= (sector_nr
| conf
->chunk_mask
) - sector_nr
+1;
2026 for (i
=0; i
<conf
->copies
; i
++) {
2027 int d
= r10_bio
->devs
[i
].devnum
;
2028 bio
= r10_bio
->devs
[i
].bio
;
2029 bio
->bi_end_io
= NULL
;
2030 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2031 if (conf
->mirrors
[d
].rdev
== NULL
||
2032 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
2034 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2035 atomic_inc(&r10_bio
->remaining
);
2036 bio
->bi_next
= biolist
;
2038 bio
->bi_private
= r10_bio
;
2039 bio
->bi_end_io
= end_sync_read
;
2041 bio
->bi_sector
= r10_bio
->devs
[i
].addr
+
2042 conf
->mirrors
[d
].rdev
->data_offset
;
2043 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2048 for (i
=0; i
<conf
->copies
; i
++) {
2049 int d
= r10_bio
->devs
[i
].devnum
;
2050 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
2051 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
2059 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2061 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2063 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
2066 bio
->bi_phys_segments
= 0;
2071 if (sector_nr
+ max_sync
< max_sector
)
2072 max_sector
= sector_nr
+ max_sync
;
2075 int len
= PAGE_SIZE
;
2077 if (sector_nr
+ (len
>>9) > max_sector
)
2078 len
= (max_sector
- sector_nr
) << 9;
2081 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
2082 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2083 if (bio_add_page(bio
, page
, len
, 0) == 0) {
2086 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2087 for (bio2
= biolist
; bio2
&& bio2
!= bio
; bio2
= bio2
->bi_next
) {
2088 /* remove last page from this bio */
2090 bio2
->bi_size
-= len
;
2091 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
2097 nr_sectors
+= len
>>9;
2098 sector_nr
+= len
>>9;
2099 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
2101 r10_bio
->sectors
= nr_sectors
;
2105 biolist
= biolist
->bi_next
;
2107 bio
->bi_next
= NULL
;
2108 r10_bio
= bio
->bi_private
;
2109 r10_bio
->sectors
= nr_sectors
;
2111 if (bio
->bi_end_io
== end_sync_read
) {
2112 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2113 generic_make_request(bio
);
2117 if (sectors_skipped
)
2118 /* pretend they weren't skipped, it makes
2119 * no important difference in this case
2121 md_done_sync(mddev
, sectors_skipped
, 1);
2123 return sectors_skipped
+ nr_sectors
;
2125 /* There is nowhere to write, so all non-sync
2126 * drives must be failed, so try the next chunk...
2128 if (sector_nr
+ max_sync
< max_sector
)
2129 max_sector
= sector_nr
+ max_sync
;
2131 sectors_skipped
+= (max_sector
- sector_nr
);
2133 sector_nr
= max_sector
;
2138 raid10_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
2141 conf_t
*conf
= mddev
->private;
2144 raid_disks
= conf
->raid_disks
;
2146 sectors
= conf
->dev_sectors
;
2148 size
= sectors
>> conf
->chunk_shift
;
2149 sector_div(size
, conf
->far_copies
);
2150 size
= size
* raid_disks
;
2151 sector_div(size
, conf
->near_copies
);
2153 return size
<< conf
->chunk_shift
;
2157 static conf_t
*setup_conf(mddev_t
*mddev
)
2159 conf_t
*conf
= NULL
;
2161 sector_t stride
, size
;
2164 if (mddev
->new_chunk_sectors
< (PAGE_SIZE
>> 9) ||
2165 !is_power_of_2(mddev
->new_chunk_sectors
)) {
2166 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
2167 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2168 mdname(mddev
), PAGE_SIZE
);
2172 nc
= mddev
->new_layout
& 255;
2173 fc
= (mddev
->new_layout
>> 8) & 255;
2174 fo
= mddev
->new_layout
& (1<<16);
2176 if ((nc
*fc
) <2 || (nc
*fc
) > mddev
->raid_disks
||
2177 (mddev
->new_layout
>> 17)) {
2178 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2179 mdname(mddev
), mddev
->new_layout
);
2184 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
2188 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
2193 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2198 conf
->raid_disks
= mddev
->raid_disks
;
2199 conf
->near_copies
= nc
;
2200 conf
->far_copies
= fc
;
2201 conf
->copies
= nc
*fc
;
2202 conf
->far_offset
= fo
;
2203 conf
->chunk_mask
= mddev
->new_chunk_sectors
- 1;
2204 conf
->chunk_shift
= ffz(~mddev
->new_chunk_sectors
);
2206 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
2207 r10bio_pool_free
, conf
);
2208 if (!conf
->r10bio_pool
)
2211 size
= mddev
->dev_sectors
>> conf
->chunk_shift
;
2212 sector_div(size
, fc
);
2213 size
= size
* conf
->raid_disks
;
2214 sector_div(size
, nc
);
2215 /* 'size' is now the number of chunks in the array */
2216 /* calculate "used chunks per device" in 'stride' */
2217 stride
= size
* conf
->copies
;
2219 /* We need to round up when dividing by raid_disks to
2220 * get the stride size.
2222 stride
+= conf
->raid_disks
- 1;
2223 sector_div(stride
, conf
->raid_disks
);
2225 conf
->dev_sectors
= stride
<< conf
->chunk_shift
;
2230 sector_div(stride
, fc
);
2231 conf
->stride
= stride
<< conf
->chunk_shift
;
2234 spin_lock_init(&conf
->device_lock
);
2235 INIT_LIST_HEAD(&conf
->retry_list
);
2237 spin_lock_init(&conf
->resync_lock
);
2238 init_waitqueue_head(&conf
->wait_barrier
);
2240 conf
->thread
= md_register_thread(raid10d
, mddev
, NULL
);
2244 conf
->mddev
= mddev
;
2248 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
2251 if (conf
->r10bio_pool
)
2252 mempool_destroy(conf
->r10bio_pool
);
2253 kfree(conf
->mirrors
);
2254 safe_put_page(conf
->tmppage
);
2257 return ERR_PTR(err
);
2260 static int run(mddev_t
*mddev
)
2263 int i
, disk_idx
, chunk_size
;
2264 mirror_info_t
*disk
;
2269 * copy the already verified devices into our private RAID10
2270 * bookkeeping area. [whatever we allocate in run(),
2271 * should be freed in stop()]
2274 if (mddev
->private == NULL
) {
2275 conf
= setup_conf(mddev
);
2277 return PTR_ERR(conf
);
2278 mddev
->private = conf
;
2280 conf
= mddev
->private;
2284 mddev
->queue
->queue_lock
= &conf
->device_lock
;
2286 mddev
->thread
= conf
->thread
;
2287 conf
->thread
= NULL
;
2289 chunk_size
= mddev
->chunk_sectors
<< 9;
2290 blk_queue_io_min(mddev
->queue
, chunk_size
);
2291 if (conf
->raid_disks
% conf
->near_copies
)
2292 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->raid_disks
);
2294 blk_queue_io_opt(mddev
->queue
, chunk_size
*
2295 (conf
->raid_disks
/ conf
->near_copies
));
2297 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
2298 disk_idx
= rdev
->raid_disk
;
2299 if (disk_idx
>= conf
->raid_disks
2302 disk
= conf
->mirrors
+ disk_idx
;
2305 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2306 rdev
->data_offset
<< 9);
2307 /* as we don't honour merge_bvec_fn, we must never risk
2308 * violating it, so limit max_segments to 1 lying
2309 * within a single page.
2311 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
2312 blk_queue_max_segments(mddev
->queue
, 1);
2313 blk_queue_segment_boundary(mddev
->queue
,
2314 PAGE_CACHE_SIZE
- 1);
2317 disk
->head_position
= 0;
2319 /* need to check that every block has at least one working mirror */
2320 if (!enough(conf
)) {
2321 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
2326 mddev
->degraded
= 0;
2327 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2329 disk
= conf
->mirrors
+ i
;
2332 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2333 disk
->head_position
= 0;
2340 if (mddev
->recovery_cp
!= MaxSector
)
2341 printk(KERN_NOTICE
"md/raid10:%s: not clean"
2342 " -- starting background reconstruction\n",
2345 "md/raid10:%s: active with %d out of %d devices\n",
2346 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
,
2349 * Ok, everything is just fine now
2351 mddev
->dev_sectors
= conf
->dev_sectors
;
2352 size
= raid10_size(mddev
, 0, 0);
2353 md_set_array_sectors(mddev
, size
);
2354 mddev
->resync_max_sectors
= size
;
2356 mddev
->queue
->unplug_fn
= raid10_unplug
;
2357 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
2358 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2360 /* Calculate max read-ahead size.
2361 * We need to readahead at least twice a whole stripe....
2365 int stripe
= conf
->raid_disks
*
2366 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
2367 stripe
/= conf
->near_copies
;
2368 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2* stripe
)
2369 mddev
->queue
->backing_dev_info
.ra_pages
= 2* stripe
;
2372 if (conf
->near_copies
< conf
->raid_disks
)
2373 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
2374 md_integrity_register(mddev
);
2378 if (conf
->r10bio_pool
)
2379 mempool_destroy(conf
->r10bio_pool
);
2380 safe_put_page(conf
->tmppage
);
2381 kfree(conf
->mirrors
);
2383 mddev
->private = NULL
;
2384 md_unregister_thread(mddev
->thread
);
2389 static int stop(mddev_t
*mddev
)
2391 conf_t
*conf
= mddev
->private;
2393 raise_barrier(conf
, 0);
2394 lower_barrier(conf
);
2396 md_unregister_thread(mddev
->thread
);
2397 mddev
->thread
= NULL
;
2398 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
2399 if (conf
->r10bio_pool
)
2400 mempool_destroy(conf
->r10bio_pool
);
2401 kfree(conf
->mirrors
);
2403 mddev
->private = NULL
;
2407 static void raid10_quiesce(mddev_t
*mddev
, int state
)
2409 conf_t
*conf
= mddev
->private;
2413 raise_barrier(conf
, 0);
2416 lower_barrier(conf
);
2421 static void *raid10_takeover_raid0(mddev_t
*mddev
)
2426 if (mddev
->degraded
> 0) {
2427 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
2429 return ERR_PTR(-EINVAL
);
2432 /* Set new parameters */
2433 mddev
->new_level
= 10;
2434 /* new layout: far_copies = 1, near_copies = 2 */
2435 mddev
->new_layout
= (1<<8) + 2;
2436 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
2437 mddev
->delta_disks
= mddev
->raid_disks
;
2438 mddev
->raid_disks
*= 2;
2439 /* make sure it will be not marked as dirty */
2440 mddev
->recovery_cp
= MaxSector
;
2442 conf
= setup_conf(mddev
);
2444 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
2445 if (rdev
->raid_disk
>= 0)
2446 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
2451 static void *raid10_takeover(mddev_t
*mddev
)
2453 struct raid0_private_data
*raid0_priv
;
2455 /* raid10 can take over:
2456 * raid0 - providing it has only two drives
2458 if (mddev
->level
== 0) {
2459 /* for raid0 takeover only one zone is supported */
2460 raid0_priv
= mddev
->private;
2461 if (raid0_priv
->nr_strip_zones
> 1) {
2462 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
2463 " with more than one zone.\n",
2465 return ERR_PTR(-EINVAL
);
2467 return raid10_takeover_raid0(mddev
);
2469 return ERR_PTR(-EINVAL
);
2472 static struct mdk_personality raid10_personality
=
2476 .owner
= THIS_MODULE
,
2477 .make_request
= make_request
,
2481 .error_handler
= error
,
2482 .hot_add_disk
= raid10_add_disk
,
2483 .hot_remove_disk
= raid10_remove_disk
,
2484 .spare_active
= raid10_spare_active
,
2485 .sync_request
= sync_request
,
2486 .quiesce
= raid10_quiesce
,
2487 .size
= raid10_size
,
2488 .takeover
= raid10_takeover
,
2491 static int __init
raid_init(void)
2493 return register_md_personality(&raid10_personality
);
2496 static void raid_exit(void)
2498 unregister_md_personality(&raid10_personality
);
2501 module_init(raid_init
);
2502 module_exit(raid_exit
);
2503 MODULE_LICENSE("GPL");
2504 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2505 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2506 MODULE_ALIAS("md-raid10");
2507 MODULE_ALIAS("md-level-10");