2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
42 * The data to be stored is divided into chunks using chunksize.
43 * Each device is divided into far_copies sections.
44 * In each section, chunks are laid out in a style similar to raid0, but
45 * near_copies copies of each chunk is stored (each on a different drive).
46 * The starting device for each section is offset near_copies from the starting
47 * device of the previous section.
48 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
50 * near_copies and far_copies must be at least one, and their product is at most
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of be very far apart
55 * on disk, there are adjacent stripes.
59 * Number of guaranteed r10bios in case of extreme VM load:
61 #define NR_RAID10_BIOS 256
63 /* when we get a read error on a read-only array, we redirect to another
64 * device without failing the first device, or trying to over-write to
65 * correct the read error. To keep track of bad blocks on a per-bio
66 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
68 #define IO_BLOCKED ((struct bio *)1)
69 /* When we successfully write to a known bad-block, we need to remove the
70 * bad-block marking which must be done from process context. So we record
71 * the success by setting devs[n].bio to IO_MADE_GOOD
73 #define IO_MADE_GOOD ((struct bio *)2)
75 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
77 /* When there are this many requests queued to be written by
78 * the raid10 thread, we become 'congested' to provide back-pressure
81 static int max_queued_requests
= 1024;
83 static void allow_barrier(struct r10conf
*conf
);
84 static void lower_barrier(struct r10conf
*conf
);
85 static int enough(struct r10conf
*conf
, int ignore
);
86 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
88 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
);
89 static void end_reshape_write(struct bio
*bio
, int error
);
90 static void end_reshape(struct r10conf
*conf
);
92 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
94 struct r10conf
*conf
= data
;
95 int size
= offsetof(struct r10bio
, devs
[conf
->copies
]);
97 /* allocate a r10bio with room for raid_disks entries in the
99 return kzalloc(size
, gfp_flags
);
102 static void r10bio_pool_free(void *r10_bio
, void *data
)
107 /* Maximum size of each resync request */
108 #define RESYNC_BLOCK_SIZE (64*1024)
109 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
110 /* amount of memory to reserve for resync requests */
111 #define RESYNC_WINDOW (1024*1024)
112 /* maximum number of concurrent requests, memory permitting */
113 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
116 * When performing a resync, we need to read and compare, so
117 * we need as many pages are there are copies.
118 * When performing a recovery, we need 2 bios, one for read,
119 * one for write (we recover only one drive per r10buf)
122 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
124 struct r10conf
*conf
= data
;
126 struct r10bio
*r10_bio
;
131 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
135 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
) ||
136 test_bit(MD_RECOVERY_RESHAPE
, &conf
->mddev
->recovery
))
137 nalloc
= conf
->copies
; /* resync */
139 nalloc
= 2; /* recovery */
144 for (j
= nalloc
; j
-- ; ) {
145 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
148 r10_bio
->devs
[j
].bio
= bio
;
149 if (!conf
->have_replacement
)
151 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
154 r10_bio
->devs
[j
].repl_bio
= bio
;
157 * Allocate RESYNC_PAGES data pages and attach them
160 for (j
= 0 ; j
< nalloc
; j
++) {
161 struct bio
*rbio
= r10_bio
->devs
[j
].repl_bio
;
162 bio
= r10_bio
->devs
[j
].bio
;
163 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
164 if (j
> 0 && !test_bit(MD_RECOVERY_SYNC
,
165 &conf
->mddev
->recovery
)) {
166 /* we can share bv_page's during recovery
168 struct bio
*rbio
= r10_bio
->devs
[0].bio
;
169 page
= rbio
->bi_io_vec
[i
].bv_page
;
172 page
= alloc_page(gfp_flags
);
176 bio
->bi_io_vec
[i
].bv_page
= page
;
178 rbio
->bi_io_vec
[i
].bv_page
= page
;
186 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
188 for (i
= 0; i
< RESYNC_PAGES
; i
++)
189 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
192 for ( ; j
< nalloc
; j
++) {
193 if (r10_bio
->devs
[j
].bio
)
194 bio_put(r10_bio
->devs
[j
].bio
);
195 if (r10_bio
->devs
[j
].repl_bio
)
196 bio_put(r10_bio
->devs
[j
].repl_bio
);
198 r10bio_pool_free(r10_bio
, conf
);
202 static void r10buf_pool_free(void *__r10_bio
, void *data
)
205 struct r10conf
*conf
= data
;
206 struct r10bio
*r10bio
= __r10_bio
;
209 for (j
=0; j
< conf
->copies
; j
++) {
210 struct bio
*bio
= r10bio
->devs
[j
].bio
;
212 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
213 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
214 bio
->bi_io_vec
[i
].bv_page
= NULL
;
218 bio
= r10bio
->devs
[j
].repl_bio
;
222 r10bio_pool_free(r10bio
, conf
);
225 static void put_all_bios(struct r10conf
*conf
, struct r10bio
*r10_bio
)
229 for (i
= 0; i
< conf
->copies
; i
++) {
230 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
231 if (!BIO_SPECIAL(*bio
))
234 bio
= &r10_bio
->devs
[i
].repl_bio
;
235 if (r10_bio
->read_slot
< 0 && !BIO_SPECIAL(*bio
))
241 static void free_r10bio(struct r10bio
*r10_bio
)
243 struct r10conf
*conf
= r10_bio
->mddev
->private;
245 put_all_bios(conf
, r10_bio
);
246 mempool_free(r10_bio
, conf
->r10bio_pool
);
249 static void put_buf(struct r10bio
*r10_bio
)
251 struct r10conf
*conf
= r10_bio
->mddev
->private;
253 mempool_free(r10_bio
, conf
->r10buf_pool
);
258 static void reschedule_retry(struct r10bio
*r10_bio
)
261 struct mddev
*mddev
= r10_bio
->mddev
;
262 struct r10conf
*conf
= mddev
->private;
264 spin_lock_irqsave(&conf
->device_lock
, flags
);
265 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
267 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
269 /* wake up frozen array... */
270 wake_up(&conf
->wait_barrier
);
272 md_wakeup_thread(mddev
->thread
);
276 * raid_end_bio_io() is called when we have finished servicing a mirrored
277 * operation and are ready to return a success/failure code to the buffer
280 static void raid_end_bio_io(struct r10bio
*r10_bio
)
282 struct bio
*bio
= r10_bio
->master_bio
;
284 struct r10conf
*conf
= r10_bio
->mddev
->private;
286 if (bio
->bi_phys_segments
) {
288 spin_lock_irqsave(&conf
->device_lock
, flags
);
289 bio
->bi_phys_segments
--;
290 done
= (bio
->bi_phys_segments
== 0);
291 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
294 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
295 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
299 * Wake up any possible resync thread that waits for the device
304 free_r10bio(r10_bio
);
308 * Update disk head position estimator based on IRQ completion info.
310 static inline void update_head_pos(int slot
, struct r10bio
*r10_bio
)
312 struct r10conf
*conf
= r10_bio
->mddev
->private;
314 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
315 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
319 * Find the disk number which triggered given bio
321 static int find_bio_disk(struct r10conf
*conf
, struct r10bio
*r10_bio
,
322 struct bio
*bio
, int *slotp
, int *replp
)
327 for (slot
= 0; slot
< conf
->copies
; slot
++) {
328 if (r10_bio
->devs
[slot
].bio
== bio
)
330 if (r10_bio
->devs
[slot
].repl_bio
== bio
) {
336 BUG_ON(slot
== conf
->copies
);
337 update_head_pos(slot
, r10_bio
);
343 return r10_bio
->devs
[slot
].devnum
;
346 static void raid10_end_read_request(struct bio
*bio
, int error
)
348 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
349 struct r10bio
*r10_bio
= bio
->bi_private
;
351 struct md_rdev
*rdev
;
352 struct r10conf
*conf
= r10_bio
->mddev
->private;
355 slot
= r10_bio
->read_slot
;
356 dev
= r10_bio
->devs
[slot
].devnum
;
357 rdev
= r10_bio
->devs
[slot
].rdev
;
359 * this branch is our 'one mirror IO has finished' event handler:
361 update_head_pos(slot
, r10_bio
);
365 * Set R10BIO_Uptodate in our master bio, so that
366 * we will return a good error code to the higher
367 * levels even if IO on some other mirrored buffer fails.
369 * The 'master' represents the composite IO operation to
370 * user-side. So if something waits for IO, then it will
371 * wait for the 'master' bio.
373 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
375 /* If all other devices that store this block have
376 * failed, we want to return the error upwards rather
377 * than fail the last device. Here we redefine
378 * "uptodate" to mean "Don't want to retry"
381 spin_lock_irqsave(&conf
->device_lock
, flags
);
382 if (!enough(conf
, rdev
->raid_disk
))
384 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
387 raid_end_bio_io(r10_bio
);
388 rdev_dec_pending(rdev
, conf
->mddev
);
391 * oops, read error - keep the refcount on the rdev
393 char b
[BDEVNAME_SIZE
];
394 printk_ratelimited(KERN_ERR
395 "md/raid10:%s: %s: rescheduling sector %llu\n",
397 bdevname(rdev
->bdev
, b
),
398 (unsigned long long)r10_bio
->sector
);
399 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
400 reschedule_retry(r10_bio
);
404 static void close_write(struct r10bio
*r10_bio
)
406 /* clear the bitmap if all writes complete successfully */
407 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
409 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
411 md_write_end(r10_bio
->mddev
);
414 static void one_write_done(struct r10bio
*r10_bio
)
416 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
417 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
418 reschedule_retry(r10_bio
);
420 close_write(r10_bio
);
421 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
422 reschedule_retry(r10_bio
);
424 raid_end_bio_io(r10_bio
);
429 static void raid10_end_write_request(struct bio
*bio
, int error
)
431 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
432 struct r10bio
*r10_bio
= bio
->bi_private
;
435 struct r10conf
*conf
= r10_bio
->mddev
->private;
437 struct md_rdev
*rdev
= NULL
;
439 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
442 rdev
= conf
->mirrors
[dev
].replacement
;
446 rdev
= conf
->mirrors
[dev
].rdev
;
449 * this branch is our 'one mirror IO has finished' event handler:
453 /* Never record new bad blocks to replacement,
456 md_error(rdev
->mddev
, rdev
);
458 set_bit(WriteErrorSeen
, &rdev
->flags
);
459 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
460 set_bit(MD_RECOVERY_NEEDED
,
461 &rdev
->mddev
->recovery
);
462 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
467 * Set R10BIO_Uptodate in our master bio, so that
468 * we will return a good error code for to the higher
469 * levels even if IO on some other mirrored buffer fails.
471 * The 'master' represents the composite IO operation to
472 * user-side. So if something waits for IO, then it will
473 * wait for the 'master' bio.
478 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
480 /* Maybe we can clear some bad blocks. */
481 if (is_badblock(rdev
,
482 r10_bio
->devs
[slot
].addr
,
484 &first_bad
, &bad_sectors
)) {
487 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
489 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
491 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
497 * Let's see if all mirrored write operations have finished
500 one_write_done(r10_bio
);
502 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
506 * RAID10 layout manager
507 * As well as the chunksize and raid_disks count, there are two
508 * parameters: near_copies and far_copies.
509 * near_copies * far_copies must be <= raid_disks.
510 * Normally one of these will be 1.
511 * If both are 1, we get raid0.
512 * If near_copies == raid_disks, we get raid1.
514 * Chunks are laid out in raid0 style with near_copies copies of the
515 * first chunk, followed by near_copies copies of the next chunk and
517 * If far_copies > 1, then after 1/far_copies of the array has been assigned
518 * as described above, we start again with a device offset of near_copies.
519 * So we effectively have another copy of the whole array further down all
520 * the drives, but with blocks on different drives.
521 * With this layout, and block is never stored twice on the one device.
523 * raid10_find_phys finds the sector offset of a given virtual sector
524 * on each device that it is on.
526 * raid10_find_virt does the reverse mapping, from a device and a
527 * sector offset to a virtual address
530 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
539 /* now calculate first sector/dev */
540 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
541 sector
= r10bio
->sector
& geo
->chunk_mask
;
543 chunk
*= geo
->near_copies
;
545 dev
= sector_div(stripe
, geo
->raid_disks
);
547 stripe
*= geo
->far_copies
;
549 sector
+= stripe
<< geo
->chunk_shift
;
551 /* and calculate all the others */
552 for (n
= 0; n
< geo
->near_copies
; n
++) {
555 r10bio
->devs
[slot
].addr
= sector
;
556 r10bio
->devs
[slot
].devnum
= d
;
559 for (f
= 1; f
< geo
->far_copies
; f
++) {
560 d
+= geo
->near_copies
;
561 if (d
>= geo
->raid_disks
)
562 d
-= geo
->raid_disks
;
564 r10bio
->devs
[slot
].devnum
= d
;
565 r10bio
->devs
[slot
].addr
= s
;
569 if (dev
>= geo
->raid_disks
) {
571 sector
+= (geo
->chunk_mask
+ 1);
576 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
578 struct geom
*geo
= &conf
->geo
;
580 if (conf
->reshape_progress
!= MaxSector
&&
581 ((r10bio
->sector
>= conf
->reshape_progress
) !=
582 conf
->mddev
->reshape_backwards
)) {
583 set_bit(R10BIO_Previous
, &r10bio
->state
);
586 clear_bit(R10BIO_Previous
, &r10bio
->state
);
588 __raid10_find_phys(geo
, r10bio
);
591 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
593 sector_t offset
, chunk
, vchunk
;
594 /* Never use conf->prev as this is only called during resync
595 * or recovery, so reshape isn't happening
597 struct geom
*geo
= &conf
->geo
;
599 offset
= sector
& geo
->chunk_mask
;
600 if (geo
->far_offset
) {
602 chunk
= sector
>> geo
->chunk_shift
;
603 fc
= sector_div(chunk
, geo
->far_copies
);
604 dev
-= fc
* geo
->near_copies
;
606 dev
+= geo
->raid_disks
;
608 while (sector
>= geo
->stride
) {
609 sector
-= geo
->stride
;
610 if (dev
< geo
->near_copies
)
611 dev
+= geo
->raid_disks
- geo
->near_copies
;
613 dev
-= geo
->near_copies
;
615 chunk
= sector
>> geo
->chunk_shift
;
617 vchunk
= chunk
* geo
->raid_disks
+ dev
;
618 sector_div(vchunk
, geo
->near_copies
);
619 return (vchunk
<< geo
->chunk_shift
) + offset
;
623 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
625 * @bvm: properties of new bio
626 * @biovec: the request that could be merged to it.
628 * Return amount of bytes we can accept at this offset
629 * This requires checking for end-of-chunk if near_copies != raid_disks,
630 * and for subordinate merge_bvec_fns if merge_check_needed.
632 static int raid10_mergeable_bvec(struct request_queue
*q
,
633 struct bvec_merge_data
*bvm
,
634 struct bio_vec
*biovec
)
636 struct mddev
*mddev
= q
->queuedata
;
637 struct r10conf
*conf
= mddev
->private;
638 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
640 unsigned int chunk_sectors
;
641 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
642 struct geom
*geo
= &conf
->geo
;
644 chunk_sectors
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
) + 1;
645 if (conf
->reshape_progress
!= MaxSector
&&
646 ((sector
>= conf
->reshape_progress
) !=
647 conf
->mddev
->reshape_backwards
))
650 if (geo
->near_copies
< geo
->raid_disks
) {
651 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1))
652 + bio_sectors
)) << 9;
654 /* bio_add cannot handle a negative return */
656 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
657 return biovec
->bv_len
;
659 max
= biovec
->bv_len
;
661 if (mddev
->merge_check_needed
) {
662 struct r10bio r10_bio
;
664 if (conf
->reshape_progress
!= MaxSector
) {
665 /* Cannot give any guidance during reshape */
666 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
667 return biovec
->bv_len
;
670 r10_bio
.sector
= sector
;
671 raid10_find_phys(conf
, &r10_bio
);
673 for (s
= 0; s
< conf
->copies
; s
++) {
674 int disk
= r10_bio
.devs
[s
].devnum
;
675 struct md_rdev
*rdev
= rcu_dereference(
676 conf
->mirrors
[disk
].rdev
);
677 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
678 struct request_queue
*q
=
679 bdev_get_queue(rdev
->bdev
);
680 if (q
->merge_bvec_fn
) {
681 bvm
->bi_sector
= r10_bio
.devs
[s
].addr
683 bvm
->bi_bdev
= rdev
->bdev
;
684 max
= min(max
, q
->merge_bvec_fn(
688 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
689 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
690 struct request_queue
*q
=
691 bdev_get_queue(rdev
->bdev
);
692 if (q
->merge_bvec_fn
) {
693 bvm
->bi_sector
= r10_bio
.devs
[s
].addr
695 bvm
->bi_bdev
= rdev
->bdev
;
696 max
= min(max
, q
->merge_bvec_fn(
707 * This routine returns the disk from which the requested read should
708 * be done. There is a per-array 'next expected sequential IO' sector
709 * number - if this matches on the next IO then we use the last disk.
710 * There is also a per-disk 'last know head position' sector that is
711 * maintained from IRQ contexts, both the normal and the resync IO
712 * completion handlers update this position correctly. If there is no
713 * perfect sequential match then we pick the disk whose head is closest.
715 * If there are 2 mirrors in the same 2 devices, performance degrades
716 * because position is mirror, not device based.
718 * The rdev for the device selected will have nr_pending incremented.
722 * FIXME: possibly should rethink readbalancing and do it differently
723 * depending on near_copies / far_copies geometry.
725 static struct md_rdev
*read_balance(struct r10conf
*conf
,
726 struct r10bio
*r10_bio
,
729 const sector_t this_sector
= r10_bio
->sector
;
731 int sectors
= r10_bio
->sectors
;
732 int best_good_sectors
;
733 sector_t new_distance
, best_dist
;
734 struct md_rdev
*best_rdev
, *rdev
= NULL
;
737 struct geom
*geo
= &conf
->geo
;
739 raid10_find_phys(conf
, r10_bio
);
742 sectors
= r10_bio
->sectors
;
745 best_dist
= MaxSector
;
746 best_good_sectors
= 0;
749 * Check if we can balance. We can balance on the whole
750 * device if no resync is going on (recovery is ok), or below
751 * the resync window. We take the first readable disk when
752 * above the resync window.
754 if (conf
->mddev
->recovery_cp
< MaxSector
755 && (this_sector
+ sectors
>= conf
->next_resync
))
758 for (slot
= 0; slot
< conf
->copies
; slot
++) {
763 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
765 disk
= r10_bio
->devs
[slot
].devnum
;
766 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
767 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
768 test_bit(Unmerged
, &rdev
->flags
) ||
769 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
770 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
772 test_bit(Faulty
, &rdev
->flags
) ||
773 test_bit(Unmerged
, &rdev
->flags
))
775 if (!test_bit(In_sync
, &rdev
->flags
) &&
776 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
779 dev_sector
= r10_bio
->devs
[slot
].addr
;
780 if (is_badblock(rdev
, dev_sector
, sectors
,
781 &first_bad
, &bad_sectors
)) {
782 if (best_dist
< MaxSector
)
783 /* Already have a better slot */
785 if (first_bad
<= dev_sector
) {
786 /* Cannot read here. If this is the
787 * 'primary' device, then we must not read
788 * beyond 'bad_sectors' from another device.
790 bad_sectors
-= (dev_sector
- first_bad
);
791 if (!do_balance
&& sectors
> bad_sectors
)
792 sectors
= bad_sectors
;
793 if (best_good_sectors
> sectors
)
794 best_good_sectors
= sectors
;
796 sector_t good_sectors
=
797 first_bad
- dev_sector
;
798 if (good_sectors
> best_good_sectors
) {
799 best_good_sectors
= good_sectors
;
804 /* Must read from here */
809 best_good_sectors
= sectors
;
814 /* This optimisation is debatable, and completely destroys
815 * sequential read speed for 'far copies' arrays. So only
816 * keep it for 'near' arrays, and review those later.
818 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
821 /* for far > 1 always use the lowest address */
822 if (geo
->far_copies
> 1)
823 new_distance
= r10_bio
->devs
[slot
].addr
;
825 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
826 conf
->mirrors
[disk
].head_position
);
827 if (new_distance
< best_dist
) {
828 best_dist
= new_distance
;
833 if (slot
>= conf
->copies
) {
839 atomic_inc(&rdev
->nr_pending
);
840 if (test_bit(Faulty
, &rdev
->flags
)) {
841 /* Cannot risk returning a device that failed
842 * before we inc'ed nr_pending
844 rdev_dec_pending(rdev
, conf
->mddev
);
847 r10_bio
->read_slot
= slot
;
851 *max_sectors
= best_good_sectors
;
856 int md_raid10_congested(struct mddev
*mddev
, int bits
)
858 struct r10conf
*conf
= mddev
->private;
861 if ((bits
& (1 << BDI_async_congested
)) &&
862 conf
->pending_count
>= max_queued_requests
)
867 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
870 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
871 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
872 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
874 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
880 EXPORT_SYMBOL_GPL(md_raid10_congested
);
882 static int raid10_congested(void *data
, int bits
)
884 struct mddev
*mddev
= data
;
886 return mddev_congested(mddev
, bits
) ||
887 md_raid10_congested(mddev
, bits
);
890 static void flush_pending_writes(struct r10conf
*conf
)
892 /* Any writes that have been queued but are awaiting
893 * bitmap updates get flushed here.
895 spin_lock_irq(&conf
->device_lock
);
897 if (conf
->pending_bio_list
.head
) {
899 bio
= bio_list_get(&conf
->pending_bio_list
);
900 conf
->pending_count
= 0;
901 spin_unlock_irq(&conf
->device_lock
);
902 /* flush any pending bitmap writes to disk
903 * before proceeding w/ I/O */
904 bitmap_unplug(conf
->mddev
->bitmap
);
905 wake_up(&conf
->wait_barrier
);
907 while (bio
) { /* submit pending writes */
908 struct bio
*next
= bio
->bi_next
;
910 generic_make_request(bio
);
914 spin_unlock_irq(&conf
->device_lock
);
918 * Sometimes we need to suspend IO while we do something else,
919 * either some resync/recovery, or reconfigure the array.
920 * To do this we raise a 'barrier'.
921 * The 'barrier' is a counter that can be raised multiple times
922 * to count how many activities are happening which preclude
924 * We can only raise the barrier if there is no pending IO.
925 * i.e. if nr_pending == 0.
926 * We choose only to raise the barrier if no-one is waiting for the
927 * barrier to go down. This means that as soon as an IO request
928 * is ready, no other operations which require a barrier will start
929 * until the IO request has had a chance.
931 * So: regular IO calls 'wait_barrier'. When that returns there
932 * is no backgroup IO happening, It must arrange to call
933 * allow_barrier when it has finished its IO.
934 * backgroup IO calls must call raise_barrier. Once that returns
935 * there is no normal IO happeing. It must arrange to call
936 * lower_barrier when the particular background IO completes.
939 static void raise_barrier(struct r10conf
*conf
, int force
)
941 BUG_ON(force
&& !conf
->barrier
);
942 spin_lock_irq(&conf
->resync_lock
);
944 /* Wait until no block IO is waiting (unless 'force') */
945 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
946 conf
->resync_lock
, );
948 /* block any new IO from starting */
951 /* Now wait for all pending IO to complete */
952 wait_event_lock_irq(conf
->wait_barrier
,
953 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
954 conf
->resync_lock
, );
956 spin_unlock_irq(&conf
->resync_lock
);
959 static void lower_barrier(struct r10conf
*conf
)
962 spin_lock_irqsave(&conf
->resync_lock
, flags
);
964 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
965 wake_up(&conf
->wait_barrier
);
968 static void wait_barrier(struct r10conf
*conf
)
970 spin_lock_irq(&conf
->resync_lock
);
973 /* Wait for the barrier to drop.
974 * However if there are already pending
975 * requests (preventing the barrier from
976 * rising completely), and the
977 * pre-process bio queue isn't empty,
978 * then don't wait, as we need to empty
979 * that queue to get the nr_pending
982 wait_event_lock_irq(conf
->wait_barrier
,
986 !bio_list_empty(current
->bio_list
)),
992 spin_unlock_irq(&conf
->resync_lock
);
995 static void allow_barrier(struct r10conf
*conf
)
998 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1000 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1001 wake_up(&conf
->wait_barrier
);
1004 static void freeze_array(struct r10conf
*conf
)
1006 /* stop syncio and normal IO and wait for everything to
1008 * We increment barrier and nr_waiting, and then
1009 * wait until nr_pending match nr_queued+1
1010 * This is called in the context of one normal IO request
1011 * that has failed. Thus any sync request that might be pending
1012 * will be blocked by nr_pending, and we need to wait for
1013 * pending IO requests to complete or be queued for re-try.
1014 * Thus the number queued (nr_queued) plus this request (1)
1015 * must match the number of pending IOs (nr_pending) before
1018 spin_lock_irq(&conf
->resync_lock
);
1021 wait_event_lock_irq(conf
->wait_barrier
,
1022 conf
->nr_pending
== conf
->nr_queued
+1,
1024 flush_pending_writes(conf
));
1026 spin_unlock_irq(&conf
->resync_lock
);
1029 static void unfreeze_array(struct r10conf
*conf
)
1031 /* reverse the effect of the freeze */
1032 spin_lock_irq(&conf
->resync_lock
);
1035 wake_up(&conf
->wait_barrier
);
1036 spin_unlock_irq(&conf
->resync_lock
);
1039 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1040 struct md_rdev
*rdev
)
1042 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1043 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1044 return rdev
->data_offset
;
1046 return rdev
->new_data_offset
;
1049 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1051 struct r10conf
*conf
= mddev
->private;
1052 struct r10bio
*r10_bio
;
1053 struct bio
*read_bio
;
1055 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1056 int chunk_sects
= chunk_mask
+ 1;
1057 const int rw
= bio_data_dir(bio
);
1058 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1059 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1060 unsigned long flags
;
1061 struct md_rdev
*blocked_rdev
;
1062 int sectors_handled
;
1066 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1067 md_flush_request(mddev
, bio
);
1071 /* If this request crosses a chunk boundary, we need to
1072 * split it. This will only happen for 1 PAGE (or less) requests.
1074 if (unlikely((bio
->bi_sector
& chunk_mask
) + (bio
->bi_size
>> 9)
1076 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1077 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1078 struct bio_pair
*bp
;
1079 /* Sanity check -- queue functions should prevent this happening */
1080 if (bio
->bi_vcnt
!= 1 ||
1083 /* This is a one page bio that upper layers
1084 * refuse to split for us, so we need to split it.
1087 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
1089 /* Each of these 'make_request' calls will call 'wait_barrier'.
1090 * If the first succeeds but the second blocks due to the resync
1091 * thread raising the barrier, we will deadlock because the
1092 * IO to the underlying device will be queued in generic_make_request
1093 * and will never complete, so will never reduce nr_pending.
1094 * So increment nr_waiting here so no new raise_barriers will
1095 * succeed, and so the second wait_barrier cannot block.
1097 spin_lock_irq(&conf
->resync_lock
);
1099 spin_unlock_irq(&conf
->resync_lock
);
1101 make_request(mddev
, &bp
->bio1
);
1102 make_request(mddev
, &bp
->bio2
);
1104 spin_lock_irq(&conf
->resync_lock
);
1106 wake_up(&conf
->wait_barrier
);
1107 spin_unlock_irq(&conf
->resync_lock
);
1109 bio_pair_release(bp
);
1112 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1113 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1114 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
1120 md_write_start(mddev
, bio
);
1123 * Register the new request and wait if the reconstruction
1124 * thread has put up a bar for new requests.
1125 * Continue immediately if no resync is active currently.
1129 sectors
= bio
->bi_size
>> 9;
1130 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1131 bio
->bi_sector
< conf
->reshape_progress
&&
1132 bio
->bi_sector
+ sectors
> conf
->reshape_progress
) {
1133 /* IO spans the reshape position. Need to wait for
1136 allow_barrier(conf
);
1137 wait_event(conf
->wait_barrier
,
1138 conf
->reshape_progress
<= bio
->bi_sector
||
1139 conf
->reshape_progress
>= bio
->bi_sector
+ sectors
);
1142 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1143 bio_data_dir(bio
) == WRITE
&&
1144 (mddev
->reshape_backwards
1145 ? (bio
->bi_sector
< conf
->reshape_safe
&&
1146 bio
->bi_sector
+ sectors
> conf
->reshape_progress
)
1147 : (bio
->bi_sector
+ sectors
> conf
->reshape_safe
&&
1148 bio
->bi_sector
< conf
->reshape_progress
))) {
1149 /* Need to update reshape_position in metadata */
1150 mddev
->reshape_position
= conf
->reshape_progress
;
1151 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1152 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1153 md_wakeup_thread(mddev
->thread
);
1154 wait_event(mddev
->sb_wait
,
1155 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1157 conf
->reshape_safe
= mddev
->reshape_position
;
1160 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1162 r10_bio
->master_bio
= bio
;
1163 r10_bio
->sectors
= sectors
;
1165 r10_bio
->mddev
= mddev
;
1166 r10_bio
->sector
= bio
->bi_sector
;
1169 /* We might need to issue multiple reads to different
1170 * devices if there are bad blocks around, so we keep
1171 * track of the number of reads in bio->bi_phys_segments.
1172 * If this is 0, there is only one r10_bio and no locking
1173 * will be needed when the request completes. If it is
1174 * non-zero, then it is the number of not-completed requests.
1176 bio
->bi_phys_segments
= 0;
1177 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1181 * read balancing logic:
1183 struct md_rdev
*rdev
;
1187 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1189 raid_end_bio_io(r10_bio
);
1192 slot
= r10_bio
->read_slot
;
1194 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1195 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1198 r10_bio
->devs
[slot
].bio
= read_bio
;
1199 r10_bio
->devs
[slot
].rdev
= rdev
;
1201 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1202 choose_data_offset(r10_bio
, rdev
);
1203 read_bio
->bi_bdev
= rdev
->bdev
;
1204 read_bio
->bi_end_io
= raid10_end_read_request
;
1205 read_bio
->bi_rw
= READ
| do_sync
;
1206 read_bio
->bi_private
= r10_bio
;
1208 if (max_sectors
< r10_bio
->sectors
) {
1209 /* Could not read all from this device, so we will
1210 * need another r10_bio.
1212 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1214 r10_bio
->sectors
= max_sectors
;
1215 spin_lock_irq(&conf
->device_lock
);
1216 if (bio
->bi_phys_segments
== 0)
1217 bio
->bi_phys_segments
= 2;
1219 bio
->bi_phys_segments
++;
1220 spin_unlock(&conf
->device_lock
);
1221 /* Cannot call generic_make_request directly
1222 * as that will be queued in __generic_make_request
1223 * and subsequent mempool_alloc might block
1224 * waiting for it. so hand bio over to raid10d.
1226 reschedule_retry(r10_bio
);
1228 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1230 r10_bio
->master_bio
= bio
;
1231 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
1234 r10_bio
->mddev
= mddev
;
1235 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1238 generic_make_request(read_bio
);
1245 if (conf
->pending_count
>= max_queued_requests
) {
1246 md_wakeup_thread(mddev
->thread
);
1247 wait_event(conf
->wait_barrier
,
1248 conf
->pending_count
< max_queued_requests
);
1250 /* first select target devices under rcu_lock and
1251 * inc refcount on their rdev. Record them by setting
1253 * If there are known/acknowledged bad blocks on any device
1254 * on which we have seen a write error, we want to avoid
1255 * writing to those blocks. This potentially requires several
1256 * writes to write around the bad blocks. Each set of writes
1257 * gets its own r10_bio with a set of bios attached. The number
1258 * of r10_bios is recored in bio->bi_phys_segments just as with
1262 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1263 raid10_find_phys(conf
, r10_bio
);
1265 blocked_rdev
= NULL
;
1267 max_sectors
= r10_bio
->sectors
;
1269 for (i
= 0; i
< conf
->copies
; i
++) {
1270 int d
= r10_bio
->devs
[i
].devnum
;
1271 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1272 struct md_rdev
*rrdev
= rcu_dereference(
1273 conf
->mirrors
[d
].replacement
);
1276 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1277 atomic_inc(&rdev
->nr_pending
);
1278 blocked_rdev
= rdev
;
1281 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1282 atomic_inc(&rrdev
->nr_pending
);
1283 blocked_rdev
= rrdev
;
1286 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1287 || test_bit(Unmerged
, &rrdev
->flags
)))
1290 r10_bio
->devs
[i
].bio
= NULL
;
1291 r10_bio
->devs
[i
].repl_bio
= NULL
;
1292 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
1293 test_bit(Unmerged
, &rdev
->flags
)) {
1294 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1297 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1299 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1303 is_bad
= is_badblock(rdev
, dev_sector
,
1305 &first_bad
, &bad_sectors
);
1307 /* Mustn't write here until the bad block
1310 atomic_inc(&rdev
->nr_pending
);
1311 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1312 blocked_rdev
= rdev
;
1315 if (is_bad
&& first_bad
<= dev_sector
) {
1316 /* Cannot write here at all */
1317 bad_sectors
-= (dev_sector
- first_bad
);
1318 if (bad_sectors
< max_sectors
)
1319 /* Mustn't write more than bad_sectors
1320 * to other devices yet
1322 max_sectors
= bad_sectors
;
1323 /* We don't set R10BIO_Degraded as that
1324 * only applies if the disk is missing,
1325 * so it might be re-added, and we want to
1326 * know to recover this chunk.
1327 * In this case the device is here, and the
1328 * fact that this chunk is not in-sync is
1329 * recorded in the bad block log.
1334 int good_sectors
= first_bad
- dev_sector
;
1335 if (good_sectors
< max_sectors
)
1336 max_sectors
= good_sectors
;
1339 r10_bio
->devs
[i
].bio
= bio
;
1340 atomic_inc(&rdev
->nr_pending
);
1342 r10_bio
->devs
[i
].repl_bio
= bio
;
1343 atomic_inc(&rrdev
->nr_pending
);
1348 if (unlikely(blocked_rdev
)) {
1349 /* Have to wait for this device to get unblocked, then retry */
1353 for (j
= 0; j
< i
; j
++) {
1354 if (r10_bio
->devs
[j
].bio
) {
1355 d
= r10_bio
->devs
[j
].devnum
;
1356 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1358 if (r10_bio
->devs
[j
].repl_bio
) {
1359 struct md_rdev
*rdev
;
1360 d
= r10_bio
->devs
[j
].devnum
;
1361 rdev
= conf
->mirrors
[d
].replacement
;
1363 /* Race with remove_disk */
1365 rdev
= conf
->mirrors
[d
].rdev
;
1367 rdev_dec_pending(rdev
, mddev
);
1370 allow_barrier(conf
);
1371 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1376 if (max_sectors
< r10_bio
->sectors
) {
1377 /* We are splitting this into multiple parts, so
1378 * we need to prepare for allocating another r10_bio.
1380 r10_bio
->sectors
= max_sectors
;
1381 spin_lock_irq(&conf
->device_lock
);
1382 if (bio
->bi_phys_segments
== 0)
1383 bio
->bi_phys_segments
= 2;
1385 bio
->bi_phys_segments
++;
1386 spin_unlock_irq(&conf
->device_lock
);
1388 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1390 atomic_set(&r10_bio
->remaining
, 1);
1391 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1393 for (i
= 0; i
< conf
->copies
; i
++) {
1395 int d
= r10_bio
->devs
[i
].devnum
;
1396 if (!r10_bio
->devs
[i
].bio
)
1399 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1400 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1402 r10_bio
->devs
[i
].bio
= mbio
;
1404 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1405 choose_data_offset(r10_bio
,
1406 conf
->mirrors
[d
].rdev
));
1407 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1408 mbio
->bi_end_io
= raid10_end_write_request
;
1409 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
1410 mbio
->bi_private
= r10_bio
;
1412 atomic_inc(&r10_bio
->remaining
);
1413 spin_lock_irqsave(&conf
->device_lock
, flags
);
1414 bio_list_add(&conf
->pending_bio_list
, mbio
);
1415 conf
->pending_count
++;
1416 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1417 if (!mddev_check_plugged(mddev
))
1418 md_wakeup_thread(mddev
->thread
);
1420 if (!r10_bio
->devs
[i
].repl_bio
)
1423 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1424 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1426 r10_bio
->devs
[i
].repl_bio
= mbio
;
1428 /* We are actively writing to the original device
1429 * so it cannot disappear, so the replacement cannot
1432 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1435 conf
->mirrors
[d
].replacement
));
1436 mbio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
1437 mbio
->bi_end_io
= raid10_end_write_request
;
1438 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
1439 mbio
->bi_private
= r10_bio
;
1441 atomic_inc(&r10_bio
->remaining
);
1442 spin_lock_irqsave(&conf
->device_lock
, flags
);
1443 bio_list_add(&conf
->pending_bio_list
, mbio
);
1444 conf
->pending_count
++;
1445 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1446 if (!mddev_check_plugged(mddev
))
1447 md_wakeup_thread(mddev
->thread
);
1450 /* Don't remove the bias on 'remaining' (one_write_done) until
1451 * after checking if we need to go around again.
1454 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1455 one_write_done(r10_bio
);
1456 /* We need another r10_bio. It has already been counted
1457 * in bio->bi_phys_segments.
1459 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1461 r10_bio
->master_bio
= bio
;
1462 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1464 r10_bio
->mddev
= mddev
;
1465 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1469 one_write_done(r10_bio
);
1471 /* In case raid10d snuck in to freeze_array */
1472 wake_up(&conf
->wait_barrier
);
1475 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1477 struct r10conf
*conf
= mddev
->private;
1480 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1481 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1482 if (conf
->geo
.near_copies
> 1)
1483 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1484 if (conf
->geo
.far_copies
> 1) {
1485 if (conf
->geo
.far_offset
)
1486 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1488 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1490 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1491 conf
->geo
.raid_disks
- mddev
->degraded
);
1492 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1493 seq_printf(seq
, "%s",
1494 conf
->mirrors
[i
].rdev
&&
1495 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1496 seq_printf(seq
, "]");
1499 /* check if there are enough drives for
1500 * every block to appear on atleast one.
1501 * Don't consider the device numbered 'ignore'
1502 * as we might be about to remove it.
1504 static int _enough(struct r10conf
*conf
, struct geom
*geo
, int ignore
)
1509 int n
= conf
->copies
;
1512 if (conf
->mirrors
[first
].rdev
&&
1515 first
= (first
+1) % geo
->raid_disks
;
1519 } while (first
!= 0);
1523 static int enough(struct r10conf
*conf
, int ignore
)
1525 return _enough(conf
, &conf
->geo
, ignore
) &&
1526 _enough(conf
, &conf
->prev
, ignore
);
1529 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1531 char b
[BDEVNAME_SIZE
];
1532 struct r10conf
*conf
= mddev
->private;
1535 * If it is not operational, then we have already marked it as dead
1536 * else if it is the last working disks, ignore the error, let the
1537 * next level up know.
1538 * else mark the drive as failed
1540 if (test_bit(In_sync
, &rdev
->flags
)
1541 && !enough(conf
, rdev
->raid_disk
))
1543 * Don't fail the drive, just return an IO error.
1546 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1547 unsigned long flags
;
1548 spin_lock_irqsave(&conf
->device_lock
, flags
);
1550 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1552 * if recovery is running, make sure it aborts.
1554 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1556 set_bit(Blocked
, &rdev
->flags
);
1557 set_bit(Faulty
, &rdev
->flags
);
1558 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1560 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1561 "md/raid10:%s: Operation continuing on %d devices.\n",
1562 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1563 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1566 static void print_conf(struct r10conf
*conf
)
1569 struct raid10_info
*tmp
;
1571 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1573 printk(KERN_DEBUG
"(!conf)\n");
1576 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1577 conf
->geo
.raid_disks
);
1579 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1580 char b
[BDEVNAME_SIZE
];
1581 tmp
= conf
->mirrors
+ i
;
1583 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1584 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1585 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1586 bdevname(tmp
->rdev
->bdev
,b
));
1590 static void close_sync(struct r10conf
*conf
)
1593 allow_barrier(conf
);
1595 mempool_destroy(conf
->r10buf_pool
);
1596 conf
->r10buf_pool
= NULL
;
1599 static int raid10_spare_active(struct mddev
*mddev
)
1602 struct r10conf
*conf
= mddev
->private;
1603 struct raid10_info
*tmp
;
1605 unsigned long flags
;
1608 * Find all non-in_sync disks within the RAID10 configuration
1609 * and mark them in_sync
1611 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1612 tmp
= conf
->mirrors
+ i
;
1613 if (tmp
->replacement
1614 && tmp
->replacement
->recovery_offset
== MaxSector
1615 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1616 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1617 /* Replacement has just become active */
1619 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1622 /* Replaced device not technically faulty,
1623 * but we need to be sure it gets removed
1624 * and never re-added.
1626 set_bit(Faulty
, &tmp
->rdev
->flags
);
1627 sysfs_notify_dirent_safe(
1628 tmp
->rdev
->sysfs_state
);
1630 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1631 } else if (tmp
->rdev
1632 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1633 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1635 sysfs_notify_dirent(tmp
->rdev
->sysfs_state
);
1638 spin_lock_irqsave(&conf
->device_lock
, flags
);
1639 mddev
->degraded
-= count
;
1640 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1647 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1649 struct r10conf
*conf
= mddev
->private;
1653 int last
= conf
->geo
.raid_disks
- 1;
1654 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1656 if (mddev
->recovery_cp
< MaxSector
)
1657 /* only hot-add to in-sync arrays, as recovery is
1658 * very different from resync
1661 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, &conf
->prev
, -1))
1664 if (rdev
->raid_disk
>= 0)
1665 first
= last
= rdev
->raid_disk
;
1667 if (q
->merge_bvec_fn
) {
1668 set_bit(Unmerged
, &rdev
->flags
);
1669 mddev
->merge_check_needed
= 1;
1672 if (rdev
->saved_raid_disk
>= first
&&
1673 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1674 mirror
= rdev
->saved_raid_disk
;
1677 for ( ; mirror
<= last
; mirror
++) {
1678 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1679 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1682 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1683 p
->replacement
!= NULL
)
1685 clear_bit(In_sync
, &rdev
->flags
);
1686 set_bit(Replacement
, &rdev
->flags
);
1687 rdev
->raid_disk
= mirror
;
1689 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1690 rdev
->data_offset
<< 9);
1692 rcu_assign_pointer(p
->replacement
, rdev
);
1696 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1697 rdev
->data_offset
<< 9);
1699 p
->head_position
= 0;
1700 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1701 rdev
->raid_disk
= mirror
;
1703 if (rdev
->saved_raid_disk
!= mirror
)
1705 rcu_assign_pointer(p
->rdev
, rdev
);
1708 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1709 /* Some requests might not have seen this new
1710 * merge_bvec_fn. We must wait for them to complete
1711 * before merging the device fully.
1712 * First we make sure any code which has tested
1713 * our function has submitted the request, then
1714 * we wait for all outstanding requests to complete.
1716 synchronize_sched();
1717 raise_barrier(conf
, 0);
1718 lower_barrier(conf
);
1719 clear_bit(Unmerged
, &rdev
->flags
);
1721 md_integrity_add_rdev(rdev
, mddev
);
1726 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1728 struct r10conf
*conf
= mddev
->private;
1730 int number
= rdev
->raid_disk
;
1731 struct md_rdev
**rdevp
;
1732 struct raid10_info
*p
= conf
->mirrors
+ number
;
1735 if (rdev
== p
->rdev
)
1737 else if (rdev
== p
->replacement
)
1738 rdevp
= &p
->replacement
;
1742 if (test_bit(In_sync
, &rdev
->flags
) ||
1743 atomic_read(&rdev
->nr_pending
)) {
1747 /* Only remove faulty devices if recovery
1750 if (!test_bit(Faulty
, &rdev
->flags
) &&
1751 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1752 (!p
->replacement
|| p
->replacement
== rdev
) &&
1753 number
< conf
->geo
.raid_disks
&&
1760 if (atomic_read(&rdev
->nr_pending
)) {
1761 /* lost the race, try later */
1765 } else if (p
->replacement
) {
1766 /* We must have just cleared 'rdev' */
1767 p
->rdev
= p
->replacement
;
1768 clear_bit(Replacement
, &p
->replacement
->flags
);
1769 smp_mb(); /* Make sure other CPUs may see both as identical
1770 * but will never see neither -- if they are careful.
1772 p
->replacement
= NULL
;
1773 clear_bit(WantReplacement
, &rdev
->flags
);
1775 /* We might have just remove the Replacement as faulty
1776 * Clear the flag just in case
1778 clear_bit(WantReplacement
, &rdev
->flags
);
1780 err
= md_integrity_register(mddev
);
1789 static void end_sync_read(struct bio
*bio
, int error
)
1791 struct r10bio
*r10_bio
= bio
->bi_private
;
1792 struct r10conf
*conf
= r10_bio
->mddev
->private;
1795 if (bio
== r10_bio
->master_bio
) {
1796 /* this is a reshape read */
1797 d
= r10_bio
->read_slot
; /* really the read dev */
1799 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1801 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1802 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1804 /* The write handler will notice the lack of
1805 * R10BIO_Uptodate and record any errors etc
1807 atomic_add(r10_bio
->sectors
,
1808 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1810 /* for reconstruct, we always reschedule after a read.
1811 * for resync, only after all reads
1813 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1814 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1815 atomic_dec_and_test(&r10_bio
->remaining
)) {
1816 /* we have read all the blocks,
1817 * do the comparison in process context in raid10d
1819 reschedule_retry(r10_bio
);
1823 static void end_sync_request(struct r10bio
*r10_bio
)
1825 struct mddev
*mddev
= r10_bio
->mddev
;
1827 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1828 if (r10_bio
->master_bio
== NULL
) {
1829 /* the primary of several recovery bios */
1830 sector_t s
= r10_bio
->sectors
;
1831 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1832 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1833 reschedule_retry(r10_bio
);
1836 md_done_sync(mddev
, s
, 1);
1839 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1840 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1841 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1842 reschedule_retry(r10_bio
);
1850 static void end_sync_write(struct bio
*bio
, int error
)
1852 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1853 struct r10bio
*r10_bio
= bio
->bi_private
;
1854 struct mddev
*mddev
= r10_bio
->mddev
;
1855 struct r10conf
*conf
= mddev
->private;
1861 struct md_rdev
*rdev
= NULL
;
1863 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1865 rdev
= conf
->mirrors
[d
].replacement
;
1867 rdev
= conf
->mirrors
[d
].rdev
;
1871 md_error(mddev
, rdev
);
1873 set_bit(WriteErrorSeen
, &rdev
->flags
);
1874 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1875 set_bit(MD_RECOVERY_NEEDED
,
1876 &rdev
->mddev
->recovery
);
1877 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1879 } else if (is_badblock(rdev
,
1880 r10_bio
->devs
[slot
].addr
,
1882 &first_bad
, &bad_sectors
))
1883 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1885 rdev_dec_pending(rdev
, mddev
);
1887 end_sync_request(r10_bio
);
1891 * Note: sync and recover and handled very differently for raid10
1892 * This code is for resync.
1893 * For resync, we read through virtual addresses and read all blocks.
1894 * If there is any error, we schedule a write. The lowest numbered
1895 * drive is authoritative.
1896 * However requests come for physical address, so we need to map.
1897 * For every physical address there are raid_disks/copies virtual addresses,
1898 * which is always are least one, but is not necessarly an integer.
1899 * This means that a physical address can span multiple chunks, so we may
1900 * have to submit multiple io requests for a single sync request.
1903 * We check if all blocks are in-sync and only write to blocks that
1906 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1908 struct r10conf
*conf
= mddev
->private;
1910 struct bio
*tbio
, *fbio
;
1913 atomic_set(&r10_bio
->remaining
, 1);
1915 /* find the first device with a block */
1916 for (i
=0; i
<conf
->copies
; i
++)
1917 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1920 if (i
== conf
->copies
)
1924 fbio
= r10_bio
->devs
[i
].bio
;
1926 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
1927 /* now find blocks with errors */
1928 for (i
=0 ; i
< conf
->copies
; i
++) {
1931 tbio
= r10_bio
->devs
[i
].bio
;
1933 if (tbio
->bi_end_io
!= end_sync_read
)
1937 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1938 /* We know that the bi_io_vec layout is the same for
1939 * both 'first' and 'i', so we just compare them.
1940 * All vec entries are PAGE_SIZE;
1942 for (j
= 0; j
< vcnt
; j
++)
1943 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1944 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1945 fbio
->bi_io_vec
[j
].bv_len
))
1949 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1950 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1951 /* Don't fix anything. */
1954 /* Ok, we need to write this bio, either to correct an
1955 * inconsistency or to correct an unreadable block.
1956 * First we need to fixup bv_offset, bv_len and
1957 * bi_vecs, as the read request might have corrupted these
1959 tbio
->bi_vcnt
= vcnt
;
1960 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1962 tbio
->bi_phys_segments
= 0;
1963 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1964 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1965 tbio
->bi_next
= NULL
;
1966 tbio
->bi_rw
= WRITE
;
1967 tbio
->bi_private
= r10_bio
;
1968 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1970 for (j
=0; j
< vcnt
; j
++) {
1971 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1972 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1974 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1975 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1978 tbio
->bi_end_io
= end_sync_write
;
1980 d
= r10_bio
->devs
[i
].devnum
;
1981 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1982 atomic_inc(&r10_bio
->remaining
);
1983 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1985 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1986 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1987 generic_make_request(tbio
);
1990 /* Now write out to any replacement devices
1993 for (i
= 0; i
< conf
->copies
; i
++) {
1996 tbio
= r10_bio
->devs
[i
].repl_bio
;
1997 if (!tbio
|| !tbio
->bi_end_io
)
1999 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2000 && r10_bio
->devs
[i
].bio
!= fbio
)
2001 for (j
= 0; j
< vcnt
; j
++)
2002 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2003 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2005 d
= r10_bio
->devs
[i
].devnum
;
2006 atomic_inc(&r10_bio
->remaining
);
2007 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2008 tbio
->bi_size
>> 9);
2009 generic_make_request(tbio
);
2013 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2014 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2020 * Now for the recovery code.
2021 * Recovery happens across physical sectors.
2022 * We recover all non-is_sync drives by finding the virtual address of
2023 * each, and then choose a working drive that also has that virt address.
2024 * There is a separate r10_bio for each non-in_sync drive.
2025 * Only the first two slots are in use. The first for reading,
2026 * The second for writing.
2029 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2031 /* We got a read error during recovery.
2032 * We repeat the read in smaller page-sized sections.
2033 * If a read succeeds, write it to the new device or record
2034 * a bad block if we cannot.
2035 * If a read fails, record a bad block on both old and
2038 struct mddev
*mddev
= r10_bio
->mddev
;
2039 struct r10conf
*conf
= mddev
->private;
2040 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2042 int sectors
= r10_bio
->sectors
;
2044 int dr
= r10_bio
->devs
[0].devnum
;
2045 int dw
= r10_bio
->devs
[1].devnum
;
2049 struct md_rdev
*rdev
;
2053 if (s
> (PAGE_SIZE
>>9))
2056 rdev
= conf
->mirrors
[dr
].rdev
;
2057 addr
= r10_bio
->devs
[0].addr
+ sect
,
2058 ok
= sync_page_io(rdev
,
2061 bio
->bi_io_vec
[idx
].bv_page
,
2064 rdev
= conf
->mirrors
[dw
].rdev
;
2065 addr
= r10_bio
->devs
[1].addr
+ sect
;
2066 ok
= sync_page_io(rdev
,
2069 bio
->bi_io_vec
[idx
].bv_page
,
2072 set_bit(WriteErrorSeen
, &rdev
->flags
);
2073 if (!test_and_set_bit(WantReplacement
,
2075 set_bit(MD_RECOVERY_NEEDED
,
2076 &rdev
->mddev
->recovery
);
2080 /* We don't worry if we cannot set a bad block -
2081 * it really is bad so there is no loss in not
2084 rdev_set_badblocks(rdev
, addr
, s
, 0);
2086 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2087 /* need bad block on destination too */
2088 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2089 addr
= r10_bio
->devs
[1].addr
+ sect
;
2090 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2092 /* just abort the recovery */
2094 "md/raid10:%s: recovery aborted"
2095 " due to read error\n",
2098 conf
->mirrors
[dw
].recovery_disabled
2099 = mddev
->recovery_disabled
;
2100 set_bit(MD_RECOVERY_INTR
,
2113 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2115 struct r10conf
*conf
= mddev
->private;
2117 struct bio
*wbio
, *wbio2
;
2119 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2120 fix_recovery_read_error(r10_bio
);
2121 end_sync_request(r10_bio
);
2126 * share the pages with the first bio
2127 * and submit the write request
2129 d
= r10_bio
->devs
[1].devnum
;
2130 wbio
= r10_bio
->devs
[1].bio
;
2131 wbio2
= r10_bio
->devs
[1].repl_bio
;
2132 if (wbio
->bi_end_io
) {
2133 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2134 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
2135 generic_make_request(wbio
);
2137 if (wbio2
&& wbio2
->bi_end_io
) {
2138 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2139 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2140 wbio2
->bi_size
>> 9);
2141 generic_make_request(wbio2
);
2147 * Used by fix_read_error() to decay the per rdev read_errors.
2148 * We halve the read error count for every hour that has elapsed
2149 * since the last recorded read error.
2152 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2154 struct timespec cur_time_mon
;
2155 unsigned long hours_since_last
;
2156 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2158 ktime_get_ts(&cur_time_mon
);
2160 if (rdev
->last_read_error
.tv_sec
== 0 &&
2161 rdev
->last_read_error
.tv_nsec
== 0) {
2162 /* first time we've seen a read error */
2163 rdev
->last_read_error
= cur_time_mon
;
2167 hours_since_last
= (cur_time_mon
.tv_sec
-
2168 rdev
->last_read_error
.tv_sec
) / 3600;
2170 rdev
->last_read_error
= cur_time_mon
;
2173 * if hours_since_last is > the number of bits in read_errors
2174 * just set read errors to 0. We do this to avoid
2175 * overflowing the shift of read_errors by hours_since_last.
2177 if (hours_since_last
>= 8 * sizeof(read_errors
))
2178 atomic_set(&rdev
->read_errors
, 0);
2180 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2183 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2184 int sectors
, struct page
*page
, int rw
)
2189 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2190 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2192 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2196 set_bit(WriteErrorSeen
, &rdev
->flags
);
2197 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2198 set_bit(MD_RECOVERY_NEEDED
,
2199 &rdev
->mddev
->recovery
);
2201 /* need to record an error - either for the block or the device */
2202 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2203 md_error(rdev
->mddev
, rdev
);
2208 * This is a kernel thread which:
2210 * 1. Retries failed read operations on working mirrors.
2211 * 2. Updates the raid superblock when problems encounter.
2212 * 3. Performs writes following reads for array synchronising.
2215 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2217 int sect
= 0; /* Offset from r10_bio->sector */
2218 int sectors
= r10_bio
->sectors
;
2219 struct md_rdev
*rdev
;
2220 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2221 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2223 /* still own a reference to this rdev, so it cannot
2224 * have been cleared recently.
2226 rdev
= conf
->mirrors
[d
].rdev
;
2228 if (test_bit(Faulty
, &rdev
->flags
))
2229 /* drive has already been failed, just ignore any
2230 more fix_read_error() attempts */
2233 check_decay_read_errors(mddev
, rdev
);
2234 atomic_inc(&rdev
->read_errors
);
2235 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2236 char b
[BDEVNAME_SIZE
];
2237 bdevname(rdev
->bdev
, b
);
2240 "md/raid10:%s: %s: Raid device exceeded "
2241 "read_error threshold [cur %d:max %d]\n",
2243 atomic_read(&rdev
->read_errors
), max_read_errors
);
2245 "md/raid10:%s: %s: Failing raid device\n",
2247 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2248 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2254 int sl
= r10_bio
->read_slot
;
2258 if (s
> (PAGE_SIZE
>>9))
2266 d
= r10_bio
->devs
[sl
].devnum
;
2267 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2269 !test_bit(Unmerged
, &rdev
->flags
) &&
2270 test_bit(In_sync
, &rdev
->flags
) &&
2271 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2272 &first_bad
, &bad_sectors
) == 0) {
2273 atomic_inc(&rdev
->nr_pending
);
2275 success
= sync_page_io(rdev
,
2276 r10_bio
->devs
[sl
].addr
+
2279 conf
->tmppage
, READ
, false);
2280 rdev_dec_pending(rdev
, mddev
);
2286 if (sl
== conf
->copies
)
2288 } while (!success
&& sl
!= r10_bio
->read_slot
);
2292 /* Cannot read from anywhere, just mark the block
2293 * as bad on the first device to discourage future
2296 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2297 rdev
= conf
->mirrors
[dn
].rdev
;
2299 if (!rdev_set_badblocks(
2301 r10_bio
->devs
[r10_bio
->read_slot
].addr
2304 md_error(mddev
, rdev
);
2305 r10_bio
->devs
[r10_bio
->read_slot
].bio
2312 /* write it back and re-read */
2314 while (sl
!= r10_bio
->read_slot
) {
2315 char b
[BDEVNAME_SIZE
];
2320 d
= r10_bio
->devs
[sl
].devnum
;
2321 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2323 test_bit(Unmerged
, &rdev
->flags
) ||
2324 !test_bit(In_sync
, &rdev
->flags
))
2327 atomic_inc(&rdev
->nr_pending
);
2329 if (r10_sync_page_io(rdev
,
2330 r10_bio
->devs
[sl
].addr
+
2332 s
, conf
->tmppage
, WRITE
)
2334 /* Well, this device is dead */
2336 "md/raid10:%s: read correction "
2338 " (%d sectors at %llu on %s)\n",
2340 (unsigned long long)(
2342 choose_data_offset(r10_bio
,
2344 bdevname(rdev
->bdev
, b
));
2345 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2348 bdevname(rdev
->bdev
, b
));
2350 rdev_dec_pending(rdev
, mddev
);
2354 while (sl
!= r10_bio
->read_slot
) {
2355 char b
[BDEVNAME_SIZE
];
2360 d
= r10_bio
->devs
[sl
].devnum
;
2361 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2363 !test_bit(In_sync
, &rdev
->flags
))
2366 atomic_inc(&rdev
->nr_pending
);
2368 switch (r10_sync_page_io(rdev
,
2369 r10_bio
->devs
[sl
].addr
+
2374 /* Well, this device is dead */
2376 "md/raid10:%s: unable to read back "
2378 " (%d sectors at %llu on %s)\n",
2380 (unsigned long long)(
2382 choose_data_offset(r10_bio
, rdev
)),
2383 bdevname(rdev
->bdev
, b
));
2384 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2387 bdevname(rdev
->bdev
, b
));
2391 "md/raid10:%s: read error corrected"
2392 " (%d sectors at %llu on %s)\n",
2394 (unsigned long long)(
2396 choose_data_offset(r10_bio
, rdev
)),
2397 bdevname(rdev
->bdev
, b
));
2398 atomic_add(s
, &rdev
->corrected_errors
);
2401 rdev_dec_pending(rdev
, mddev
);
2411 static void bi_complete(struct bio
*bio
, int error
)
2413 complete((struct completion
*)bio
->bi_private
);
2416 static int submit_bio_wait(int rw
, struct bio
*bio
)
2418 struct completion event
;
2421 init_completion(&event
);
2422 bio
->bi_private
= &event
;
2423 bio
->bi_end_io
= bi_complete
;
2424 submit_bio(rw
, bio
);
2425 wait_for_completion(&event
);
2427 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2430 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2432 struct bio
*bio
= r10_bio
->master_bio
;
2433 struct mddev
*mddev
= r10_bio
->mddev
;
2434 struct r10conf
*conf
= mddev
->private;
2435 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2436 /* bio has the data to be written to slot 'i' where
2437 * we just recently had a write error.
2438 * We repeatedly clone the bio and trim down to one block,
2439 * then try the write. Where the write fails we record
2441 * It is conceivable that the bio doesn't exactly align with
2442 * blocks. We must handle this.
2444 * We currently own a reference to the rdev.
2450 int sect_to_write
= r10_bio
->sectors
;
2453 if (rdev
->badblocks
.shift
< 0)
2456 block_sectors
= 1 << rdev
->badblocks
.shift
;
2457 sector
= r10_bio
->sector
;
2458 sectors
= ((r10_bio
->sector
+ block_sectors
)
2459 & ~(sector_t
)(block_sectors
- 1))
2462 while (sect_to_write
) {
2464 if (sectors
> sect_to_write
)
2465 sectors
= sect_to_write
;
2466 /* Write at 'sector' for 'sectors' */
2467 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2468 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2469 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2470 choose_data_offset(r10_bio
, rdev
) +
2471 (sector
- r10_bio
->sector
));
2472 wbio
->bi_bdev
= rdev
->bdev
;
2473 if (submit_bio_wait(WRITE
, wbio
) == 0)
2475 ok
= rdev_set_badblocks(rdev
, sector
,
2480 sect_to_write
-= sectors
;
2482 sectors
= block_sectors
;
2487 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2489 int slot
= r10_bio
->read_slot
;
2491 struct r10conf
*conf
= mddev
->private;
2492 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2493 char b
[BDEVNAME_SIZE
];
2494 unsigned long do_sync
;
2497 /* we got a read error. Maybe the drive is bad. Maybe just
2498 * the block and we can fix it.
2499 * We freeze all other IO, and try reading the block from
2500 * other devices. When we find one, we re-write
2501 * and check it that fixes the read error.
2502 * This is all done synchronously while the array is
2505 bio
= r10_bio
->devs
[slot
].bio
;
2506 bdevname(bio
->bi_bdev
, b
);
2508 r10_bio
->devs
[slot
].bio
= NULL
;
2510 if (mddev
->ro
== 0) {
2512 fix_read_error(conf
, mddev
, r10_bio
);
2513 unfreeze_array(conf
);
2515 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2517 rdev_dec_pending(rdev
, mddev
);
2520 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2522 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2523 " read error for block %llu\n",
2525 (unsigned long long)r10_bio
->sector
);
2526 raid_end_bio_io(r10_bio
);
2530 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2531 slot
= r10_bio
->read_slot
;
2534 "md/raid10:%s: %s: redirecting "
2535 "sector %llu to another mirror\n",
2537 bdevname(rdev
->bdev
, b
),
2538 (unsigned long long)r10_bio
->sector
);
2539 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2542 r10_bio
->sector
- bio
->bi_sector
,
2544 r10_bio
->devs
[slot
].bio
= bio
;
2545 r10_bio
->devs
[slot
].rdev
= rdev
;
2546 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2547 + choose_data_offset(r10_bio
, rdev
);
2548 bio
->bi_bdev
= rdev
->bdev
;
2549 bio
->bi_rw
= READ
| do_sync
;
2550 bio
->bi_private
= r10_bio
;
2551 bio
->bi_end_io
= raid10_end_read_request
;
2552 if (max_sectors
< r10_bio
->sectors
) {
2553 /* Drat - have to split this up more */
2554 struct bio
*mbio
= r10_bio
->master_bio
;
2555 int sectors_handled
=
2556 r10_bio
->sector
+ max_sectors
2558 r10_bio
->sectors
= max_sectors
;
2559 spin_lock_irq(&conf
->device_lock
);
2560 if (mbio
->bi_phys_segments
== 0)
2561 mbio
->bi_phys_segments
= 2;
2563 mbio
->bi_phys_segments
++;
2564 spin_unlock_irq(&conf
->device_lock
);
2565 generic_make_request(bio
);
2567 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2569 r10_bio
->master_bio
= mbio
;
2570 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2573 set_bit(R10BIO_ReadError
,
2575 r10_bio
->mddev
= mddev
;
2576 r10_bio
->sector
= mbio
->bi_sector
2581 generic_make_request(bio
);
2584 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2586 /* Some sort of write request has finished and it
2587 * succeeded in writing where we thought there was a
2588 * bad block. So forget the bad block.
2589 * Or possibly if failed and we need to record
2593 struct md_rdev
*rdev
;
2595 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2596 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2597 for (m
= 0; m
< conf
->copies
; m
++) {
2598 int dev
= r10_bio
->devs
[m
].devnum
;
2599 rdev
= conf
->mirrors
[dev
].rdev
;
2600 if (r10_bio
->devs
[m
].bio
== NULL
)
2602 if (test_bit(BIO_UPTODATE
,
2603 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2604 rdev_clear_badblocks(
2606 r10_bio
->devs
[m
].addr
,
2607 r10_bio
->sectors
, 0);
2609 if (!rdev_set_badblocks(
2611 r10_bio
->devs
[m
].addr
,
2612 r10_bio
->sectors
, 0))
2613 md_error(conf
->mddev
, rdev
);
2615 rdev
= conf
->mirrors
[dev
].replacement
;
2616 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2618 if (test_bit(BIO_UPTODATE
,
2619 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2620 rdev_clear_badblocks(
2622 r10_bio
->devs
[m
].addr
,
2623 r10_bio
->sectors
, 0);
2625 if (!rdev_set_badblocks(
2627 r10_bio
->devs
[m
].addr
,
2628 r10_bio
->sectors
, 0))
2629 md_error(conf
->mddev
, rdev
);
2634 for (m
= 0; m
< conf
->copies
; m
++) {
2635 int dev
= r10_bio
->devs
[m
].devnum
;
2636 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2637 rdev
= conf
->mirrors
[dev
].rdev
;
2638 if (bio
== IO_MADE_GOOD
) {
2639 rdev_clear_badblocks(
2641 r10_bio
->devs
[m
].addr
,
2642 r10_bio
->sectors
, 0);
2643 rdev_dec_pending(rdev
, conf
->mddev
);
2644 } else if (bio
!= NULL
&&
2645 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2646 if (!narrow_write_error(r10_bio
, m
)) {
2647 md_error(conf
->mddev
, rdev
);
2648 set_bit(R10BIO_Degraded
,
2651 rdev_dec_pending(rdev
, conf
->mddev
);
2653 bio
= r10_bio
->devs
[m
].repl_bio
;
2654 rdev
= conf
->mirrors
[dev
].replacement
;
2655 if (rdev
&& bio
== IO_MADE_GOOD
) {
2656 rdev_clear_badblocks(
2658 r10_bio
->devs
[m
].addr
,
2659 r10_bio
->sectors
, 0);
2660 rdev_dec_pending(rdev
, conf
->mddev
);
2663 if (test_bit(R10BIO_WriteError
,
2665 close_write(r10_bio
);
2666 raid_end_bio_io(r10_bio
);
2670 static void raid10d(struct mddev
*mddev
)
2672 struct r10bio
*r10_bio
;
2673 unsigned long flags
;
2674 struct r10conf
*conf
= mddev
->private;
2675 struct list_head
*head
= &conf
->retry_list
;
2676 struct blk_plug plug
;
2678 md_check_recovery(mddev
);
2680 blk_start_plug(&plug
);
2683 if (atomic_read(&mddev
->plug_cnt
) == 0)
2684 flush_pending_writes(conf
);
2686 spin_lock_irqsave(&conf
->device_lock
, flags
);
2687 if (list_empty(head
)) {
2688 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2691 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2692 list_del(head
->prev
);
2694 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2696 mddev
= r10_bio
->mddev
;
2697 conf
= mddev
->private;
2698 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2699 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2700 handle_write_completed(conf
, r10_bio
);
2701 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2702 reshape_request_write(mddev
, r10_bio
);
2703 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2704 sync_request_write(mddev
, r10_bio
);
2705 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2706 recovery_request_write(mddev
, r10_bio
);
2707 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2708 handle_read_error(mddev
, r10_bio
);
2710 /* just a partial read to be scheduled from a
2713 int slot
= r10_bio
->read_slot
;
2714 generic_make_request(r10_bio
->devs
[slot
].bio
);
2718 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2719 md_check_recovery(mddev
);
2721 blk_finish_plug(&plug
);
2725 static int init_resync(struct r10conf
*conf
)
2730 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2731 BUG_ON(conf
->r10buf_pool
);
2732 conf
->have_replacement
= 0;
2733 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2734 if (conf
->mirrors
[i
].replacement
)
2735 conf
->have_replacement
= 1;
2736 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2737 if (!conf
->r10buf_pool
)
2739 conf
->next_resync
= 0;
2744 * perform a "sync" on one "block"
2746 * We need to make sure that no normal I/O request - particularly write
2747 * requests - conflict with active sync requests.
2749 * This is achieved by tracking pending requests and a 'barrier' concept
2750 * that can be installed to exclude normal IO requests.
2752 * Resync and recovery are handled very differently.
2753 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2755 * For resync, we iterate over virtual addresses, read all copies,
2756 * and update if there are differences. If only one copy is live,
2758 * For recovery, we iterate over physical addresses, read a good
2759 * value for each non-in_sync drive, and over-write.
2761 * So, for recovery we may have several outstanding complex requests for a
2762 * given address, one for each out-of-sync device. We model this by allocating
2763 * a number of r10_bio structures, one for each out-of-sync device.
2764 * As we setup these structures, we collect all bio's together into a list
2765 * which we then process collectively to add pages, and then process again
2766 * to pass to generic_make_request.
2768 * The r10_bio structures are linked using a borrowed master_bio pointer.
2769 * This link is counted in ->remaining. When the r10_bio that points to NULL
2770 * has its remaining count decremented to 0, the whole complex operation
2775 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2776 int *skipped
, int go_faster
)
2778 struct r10conf
*conf
= mddev
->private;
2779 struct r10bio
*r10_bio
;
2780 struct bio
*biolist
= NULL
, *bio
;
2781 sector_t max_sector
, nr_sectors
;
2784 sector_t sync_blocks
;
2785 sector_t sectors_skipped
= 0;
2786 int chunks_skipped
= 0;
2787 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2789 if (!conf
->r10buf_pool
)
2790 if (init_resync(conf
))
2794 max_sector
= mddev
->dev_sectors
;
2795 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2796 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2797 max_sector
= mddev
->resync_max_sectors
;
2798 if (sector_nr
>= max_sector
) {
2799 /* If we aborted, we need to abort the
2800 * sync on the 'current' bitmap chucks (there can
2801 * be several when recovering multiple devices).
2802 * as we may have started syncing it but not finished.
2803 * We can find the current address in
2804 * mddev->curr_resync, but for recovery,
2805 * we need to convert that to several
2806 * virtual addresses.
2808 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2813 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2814 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2815 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2817 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2819 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2820 bitmap_end_sync(mddev
->bitmap
, sect
,
2824 /* completed sync */
2825 if ((!mddev
->bitmap
|| conf
->fullsync
)
2826 && conf
->have_replacement
2827 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2828 /* Completed a full sync so the replacements
2829 * are now fully recovered.
2831 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2832 if (conf
->mirrors
[i
].replacement
)
2833 conf
->mirrors
[i
].replacement
2839 bitmap_close_sync(mddev
->bitmap
);
2842 return sectors_skipped
;
2845 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2846 return reshape_request(mddev
, sector_nr
, skipped
);
2848 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2849 /* if there has been nothing to do on any drive,
2850 * then there is nothing to do at all..
2853 return (max_sector
- sector_nr
) + sectors_skipped
;
2856 if (max_sector
> mddev
->resync_max
)
2857 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2859 /* make sure whole request will fit in a chunk - if chunks
2862 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2863 max_sector
> (sector_nr
| chunk_mask
))
2864 max_sector
= (sector_nr
| chunk_mask
) + 1;
2866 * If there is non-resync activity waiting for us then
2867 * put in a delay to throttle resync.
2869 if (!go_faster
&& conf
->nr_waiting
)
2870 msleep_interruptible(1000);
2872 /* Again, very different code for resync and recovery.
2873 * Both must result in an r10bio with a list of bios that
2874 * have bi_end_io, bi_sector, bi_bdev set,
2875 * and bi_private set to the r10bio.
2876 * For recovery, we may actually create several r10bios
2877 * with 2 bios in each, that correspond to the bios in the main one.
2878 * In this case, the subordinate r10bios link back through a
2879 * borrowed master_bio pointer, and the counter in the master
2880 * includes a ref from each subordinate.
2882 /* First, we decide what to do and set ->bi_end_io
2883 * To end_sync_read if we want to read, and
2884 * end_sync_write if we will want to write.
2887 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2888 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2889 /* recovery... the complicated one */
2893 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2899 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2901 if ((mirror
->rdev
== NULL
||
2902 test_bit(In_sync
, &mirror
->rdev
->flags
))
2904 (mirror
->replacement
== NULL
||
2906 &mirror
->replacement
->flags
)))
2910 /* want to reconstruct this device */
2912 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2913 if (sect
>= mddev
->resync_max_sectors
) {
2914 /* last stripe is not complete - don't
2915 * try to recover this sector.
2919 /* Unless we are doing a full sync, or a replacement
2920 * we only need to recover the block if it is set in
2923 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2925 if (sync_blocks
< max_sync
)
2926 max_sync
= sync_blocks
;
2928 mirror
->replacement
== NULL
&&
2930 /* yep, skip the sync_blocks here, but don't assume
2931 * that there will never be anything to do here
2933 chunks_skipped
= -1;
2937 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2938 raise_barrier(conf
, rb2
!= NULL
);
2939 atomic_set(&r10_bio
->remaining
, 0);
2941 r10_bio
->master_bio
= (struct bio
*)rb2
;
2943 atomic_inc(&rb2
->remaining
);
2944 r10_bio
->mddev
= mddev
;
2945 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
2946 r10_bio
->sector
= sect
;
2948 raid10_find_phys(conf
, r10_bio
);
2950 /* Need to check if the array will still be
2953 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
2954 if (conf
->mirrors
[j
].rdev
== NULL
||
2955 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
2960 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2961 &sync_blocks
, still_degraded
);
2964 for (j
=0; j
<conf
->copies
;j
++) {
2966 int d
= r10_bio
->devs
[j
].devnum
;
2967 sector_t from_addr
, to_addr
;
2968 struct md_rdev
*rdev
;
2969 sector_t sector
, first_bad
;
2971 if (!conf
->mirrors
[d
].rdev
||
2972 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
2974 /* This is where we read from */
2976 rdev
= conf
->mirrors
[d
].rdev
;
2977 sector
= r10_bio
->devs
[j
].addr
;
2979 if (is_badblock(rdev
, sector
, max_sync
,
2980 &first_bad
, &bad_sectors
)) {
2981 if (first_bad
> sector
)
2982 max_sync
= first_bad
- sector
;
2984 bad_sectors
-= (sector
2986 if (max_sync
> bad_sectors
)
2987 max_sync
= bad_sectors
;
2991 bio
= r10_bio
->devs
[0].bio
;
2992 bio
->bi_next
= biolist
;
2994 bio
->bi_private
= r10_bio
;
2995 bio
->bi_end_io
= end_sync_read
;
2997 from_addr
= r10_bio
->devs
[j
].addr
;
2998 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
2999 bio
->bi_bdev
= rdev
->bdev
;
3000 atomic_inc(&rdev
->nr_pending
);
3001 /* and we write to 'i' (if not in_sync) */
3003 for (k
=0; k
<conf
->copies
; k
++)
3004 if (r10_bio
->devs
[k
].devnum
== i
)
3006 BUG_ON(k
== conf
->copies
);
3007 to_addr
= r10_bio
->devs
[k
].addr
;
3008 r10_bio
->devs
[0].devnum
= d
;
3009 r10_bio
->devs
[0].addr
= from_addr
;
3010 r10_bio
->devs
[1].devnum
= i
;
3011 r10_bio
->devs
[1].addr
= to_addr
;
3013 rdev
= mirror
->rdev
;
3014 if (!test_bit(In_sync
, &rdev
->flags
)) {
3015 bio
= r10_bio
->devs
[1].bio
;
3016 bio
->bi_next
= biolist
;
3018 bio
->bi_private
= r10_bio
;
3019 bio
->bi_end_io
= end_sync_write
;
3021 bio
->bi_sector
= to_addr
3022 + rdev
->data_offset
;
3023 bio
->bi_bdev
= rdev
->bdev
;
3024 atomic_inc(&r10_bio
->remaining
);
3026 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3028 /* and maybe write to replacement */
3029 bio
= r10_bio
->devs
[1].repl_bio
;
3031 bio
->bi_end_io
= NULL
;
3032 rdev
= mirror
->replacement
;
3033 /* Note: if rdev != NULL, then bio
3034 * cannot be NULL as r10buf_pool_alloc will
3035 * have allocated it.
3036 * So the second test here is pointless.
3037 * But it keeps semantic-checkers happy, and
3038 * this comment keeps human reviewers
3041 if (rdev
== NULL
|| bio
== NULL
||
3042 test_bit(Faulty
, &rdev
->flags
))
3044 bio
->bi_next
= biolist
;
3046 bio
->bi_private
= r10_bio
;
3047 bio
->bi_end_io
= end_sync_write
;
3049 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3050 bio
->bi_bdev
= rdev
->bdev
;
3051 atomic_inc(&r10_bio
->remaining
);
3054 if (j
== conf
->copies
) {
3055 /* Cannot recover, so abort the recovery or
3056 * record a bad block */
3059 atomic_dec(&rb2
->remaining
);
3062 /* problem is that there are bad blocks
3063 * on other device(s)
3066 for (k
= 0; k
< conf
->copies
; k
++)
3067 if (r10_bio
->devs
[k
].devnum
== i
)
3069 if (!test_bit(In_sync
,
3070 &mirror
->rdev
->flags
)
3071 && !rdev_set_badblocks(
3073 r10_bio
->devs
[k
].addr
,
3076 if (mirror
->replacement
&&
3077 !rdev_set_badblocks(
3078 mirror
->replacement
,
3079 r10_bio
->devs
[k
].addr
,
3084 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3086 printk(KERN_INFO
"md/raid10:%s: insufficient "
3087 "working devices for recovery.\n",
3089 mirror
->recovery_disabled
3090 = mddev
->recovery_disabled
;
3095 if (biolist
== NULL
) {
3097 struct r10bio
*rb2
= r10_bio
;
3098 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3099 rb2
->master_bio
= NULL
;
3105 /* resync. Schedule a read for every block at this virt offset */
3108 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3110 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3111 &sync_blocks
, mddev
->degraded
) &&
3112 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3113 &mddev
->recovery
)) {
3114 /* We can skip this block */
3116 return sync_blocks
+ sectors_skipped
;
3118 if (sync_blocks
< max_sync
)
3119 max_sync
= sync_blocks
;
3120 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3122 r10_bio
->mddev
= mddev
;
3123 atomic_set(&r10_bio
->remaining
, 0);
3124 raise_barrier(conf
, 0);
3125 conf
->next_resync
= sector_nr
;
3127 r10_bio
->master_bio
= NULL
;
3128 r10_bio
->sector
= sector_nr
;
3129 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3130 raid10_find_phys(conf
, r10_bio
);
3131 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3133 for (i
= 0; i
< conf
->copies
; i
++) {
3134 int d
= r10_bio
->devs
[i
].devnum
;
3135 sector_t first_bad
, sector
;
3138 if (r10_bio
->devs
[i
].repl_bio
)
3139 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3141 bio
= r10_bio
->devs
[i
].bio
;
3142 bio
->bi_end_io
= NULL
;
3143 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3144 if (conf
->mirrors
[d
].rdev
== NULL
||
3145 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3147 sector
= r10_bio
->devs
[i
].addr
;
3148 if (is_badblock(conf
->mirrors
[d
].rdev
,
3150 &first_bad
, &bad_sectors
)) {
3151 if (first_bad
> sector
)
3152 max_sync
= first_bad
- sector
;
3154 bad_sectors
-= (sector
- first_bad
);
3155 if (max_sync
> bad_sectors
)
3156 max_sync
= max_sync
;
3160 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3161 atomic_inc(&r10_bio
->remaining
);
3162 bio
->bi_next
= biolist
;
3164 bio
->bi_private
= r10_bio
;
3165 bio
->bi_end_io
= end_sync_read
;
3167 bio
->bi_sector
= sector
+
3168 conf
->mirrors
[d
].rdev
->data_offset
;
3169 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3172 if (conf
->mirrors
[d
].replacement
== NULL
||
3174 &conf
->mirrors
[d
].replacement
->flags
))
3177 /* Need to set up for writing to the replacement */
3178 bio
= r10_bio
->devs
[i
].repl_bio
;
3179 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3181 sector
= r10_bio
->devs
[i
].addr
;
3182 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3183 bio
->bi_next
= biolist
;
3185 bio
->bi_private
= r10_bio
;
3186 bio
->bi_end_io
= end_sync_write
;
3188 bio
->bi_sector
= sector
+
3189 conf
->mirrors
[d
].replacement
->data_offset
;
3190 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3195 for (i
=0; i
<conf
->copies
; i
++) {
3196 int d
= r10_bio
->devs
[i
].devnum
;
3197 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3198 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3200 if (r10_bio
->devs
[i
].repl_bio
&&
3201 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3203 conf
->mirrors
[d
].replacement
,
3212 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3214 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
3216 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
3219 bio
->bi_phys_segments
= 0;
3224 if (sector_nr
+ max_sync
< max_sector
)
3225 max_sector
= sector_nr
+ max_sync
;
3228 int len
= PAGE_SIZE
;
3229 if (sector_nr
+ (len
>>9) > max_sector
)
3230 len
= (max_sector
- sector_nr
) << 9;
3233 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3235 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3236 if (bio_add_page(bio
, page
, len
, 0))
3240 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3241 for (bio2
= biolist
;
3242 bio2
&& bio2
!= bio
;
3243 bio2
= bio2
->bi_next
) {
3244 /* remove last page from this bio */
3246 bio2
->bi_size
-= len
;
3247 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3251 nr_sectors
+= len
>>9;
3252 sector_nr
+= len
>>9;
3253 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3255 r10_bio
->sectors
= nr_sectors
;
3259 biolist
= biolist
->bi_next
;
3261 bio
->bi_next
= NULL
;
3262 r10_bio
= bio
->bi_private
;
3263 r10_bio
->sectors
= nr_sectors
;
3265 if (bio
->bi_end_io
== end_sync_read
) {
3266 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3267 generic_make_request(bio
);
3271 if (sectors_skipped
)
3272 /* pretend they weren't skipped, it makes
3273 * no important difference in this case
3275 md_done_sync(mddev
, sectors_skipped
, 1);
3277 return sectors_skipped
+ nr_sectors
;
3279 /* There is nowhere to write, so all non-sync
3280 * drives must be failed or in resync, all drives
3281 * have a bad block, so try the next chunk...
3283 if (sector_nr
+ max_sync
< max_sector
)
3284 max_sector
= sector_nr
+ max_sync
;
3286 sectors_skipped
+= (max_sector
- sector_nr
);
3288 sector_nr
= max_sector
;
3293 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3296 struct r10conf
*conf
= mddev
->private;
3299 raid_disks
= min(conf
->geo
.raid_disks
,
3300 conf
->prev
.raid_disks
);
3302 sectors
= conf
->dev_sectors
;
3304 size
= sectors
>> conf
->geo
.chunk_shift
;
3305 sector_div(size
, conf
->geo
.far_copies
);
3306 size
= size
* raid_disks
;
3307 sector_div(size
, conf
->geo
.near_copies
);
3309 return size
<< conf
->geo
.chunk_shift
;
3312 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3314 /* Calculate the number of sectors-per-device that will
3315 * actually be used, and set conf->dev_sectors and
3319 size
= size
>> conf
->geo
.chunk_shift
;
3320 sector_div(size
, conf
->geo
.far_copies
);
3321 size
= size
* conf
->geo
.raid_disks
;
3322 sector_div(size
, conf
->geo
.near_copies
);
3323 /* 'size' is now the number of chunks in the array */
3324 /* calculate "used chunks per device" */
3325 size
= size
* conf
->copies
;
3327 /* We need to round up when dividing by raid_disks to
3328 * get the stride size.
3330 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3332 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3334 if (conf
->geo
.far_offset
)
3335 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3337 sector_div(size
, conf
->geo
.far_copies
);
3338 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3342 enum geo_type
{geo_new
, geo_old
, geo_start
};
3343 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3346 int layout
, chunk
, disks
;
3349 layout
= mddev
->layout
;
3350 chunk
= mddev
->chunk_sectors
;
3351 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3354 layout
= mddev
->new_layout
;
3355 chunk
= mddev
->new_chunk_sectors
;
3356 disks
= mddev
->raid_disks
;
3358 default: /* avoid 'may be unused' warnings */
3359 case geo_start
: /* new when starting reshape - raid_disks not
3361 layout
= mddev
->new_layout
;
3362 chunk
= mddev
->new_chunk_sectors
;
3363 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3368 if (chunk
< (PAGE_SIZE
>> 9) ||
3369 !is_power_of_2(chunk
))
3372 fc
= (layout
>> 8) & 255;
3373 fo
= layout
& (1<<16);
3374 geo
->raid_disks
= disks
;
3375 geo
->near_copies
= nc
;
3376 geo
->far_copies
= fc
;
3377 geo
->far_offset
= fo
;
3378 geo
->chunk_mask
= chunk
- 1;
3379 geo
->chunk_shift
= ffz(~chunk
);
3383 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3385 struct r10conf
*conf
= NULL
;
3390 copies
= setup_geo(&geo
, mddev
, geo_new
);
3393 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3394 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3395 mdname(mddev
), PAGE_SIZE
);
3399 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3400 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3401 mdname(mddev
), mddev
->new_layout
);
3406 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3410 /* FIXME calc properly */
3411 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3412 max(0,mddev
->delta_disks
)),
3417 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3422 conf
->copies
= copies
;
3423 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3424 r10bio_pool_free
, conf
);
3425 if (!conf
->r10bio_pool
)
3428 calc_sectors(conf
, mddev
->dev_sectors
);
3429 if (mddev
->reshape_position
== MaxSector
) {
3430 conf
->prev
= conf
->geo
;
3431 conf
->reshape_progress
= MaxSector
;
3433 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3437 conf
->reshape_progress
= mddev
->reshape_position
;
3438 if (conf
->prev
.far_offset
)
3439 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3441 /* far_copies must be 1 */
3442 conf
->prev
.stride
= conf
->dev_sectors
;
3444 spin_lock_init(&conf
->device_lock
);
3445 INIT_LIST_HEAD(&conf
->retry_list
);
3447 spin_lock_init(&conf
->resync_lock
);
3448 init_waitqueue_head(&conf
->wait_barrier
);
3450 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3454 conf
->mddev
= mddev
;
3459 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3462 if (conf
->r10bio_pool
)
3463 mempool_destroy(conf
->r10bio_pool
);
3464 kfree(conf
->mirrors
);
3465 safe_put_page(conf
->tmppage
);
3468 return ERR_PTR(err
);
3471 static int run(struct mddev
*mddev
)
3473 struct r10conf
*conf
;
3474 int i
, disk_idx
, chunk_size
;
3475 struct raid10_info
*disk
;
3476 struct md_rdev
*rdev
;
3478 sector_t min_offset_diff
= 0;
3481 if (mddev
->private == NULL
) {
3482 conf
= setup_conf(mddev
);
3484 return PTR_ERR(conf
);
3485 mddev
->private = conf
;
3487 conf
= mddev
->private;
3491 mddev
->thread
= conf
->thread
;
3492 conf
->thread
= NULL
;
3494 chunk_size
= mddev
->chunk_sectors
<< 9;
3496 blk_queue_io_min(mddev
->queue
, chunk_size
);
3497 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3498 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3500 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3501 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3504 rdev_for_each(rdev
, mddev
) {
3506 struct request_queue
*q
;
3508 disk_idx
= rdev
->raid_disk
;
3511 if (disk_idx
>= conf
->geo
.raid_disks
&&
3512 disk_idx
>= conf
->prev
.raid_disks
)
3514 disk
= conf
->mirrors
+ disk_idx
;
3516 if (test_bit(Replacement
, &rdev
->flags
)) {
3517 if (disk
->replacement
)
3519 disk
->replacement
= rdev
;
3525 q
= bdev_get_queue(rdev
->bdev
);
3526 if (q
->merge_bvec_fn
)
3527 mddev
->merge_check_needed
= 1;
3528 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3529 if (!mddev
->reshape_backwards
)
3533 if (first
|| diff
< min_offset_diff
)
3534 min_offset_diff
= diff
;
3537 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3538 rdev
->data_offset
<< 9);
3540 disk
->head_position
= 0;
3543 /* need to check that every block has at least one working mirror */
3544 if (!enough(conf
, -1)) {
3545 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3550 if (conf
->reshape_progress
!= MaxSector
) {
3551 /* must ensure that shape change is supported */
3552 if (conf
->geo
.far_copies
!= 1 &&
3553 conf
->geo
.far_offset
== 0)
3555 if (conf
->prev
.far_copies
!= 1 &&
3556 conf
->geo
.far_offset
== 0)
3560 mddev
->degraded
= 0;
3562 i
< conf
->geo
.raid_disks
3563 || i
< conf
->prev
.raid_disks
;
3566 disk
= conf
->mirrors
+ i
;
3568 if (!disk
->rdev
&& disk
->replacement
) {
3569 /* The replacement is all we have - use it */
3570 disk
->rdev
= disk
->replacement
;
3571 disk
->replacement
= NULL
;
3572 clear_bit(Replacement
, &disk
->rdev
->flags
);
3576 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3577 disk
->head_position
= 0;
3582 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3585 if (mddev
->recovery_cp
!= MaxSector
)
3586 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3587 " -- starting background reconstruction\n",
3590 "md/raid10:%s: active with %d out of %d devices\n",
3591 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3592 conf
->geo
.raid_disks
);
3594 * Ok, everything is just fine now
3596 mddev
->dev_sectors
= conf
->dev_sectors
;
3597 size
= raid10_size(mddev
, 0, 0);
3598 md_set_array_sectors(mddev
, size
);
3599 mddev
->resync_max_sectors
= size
;
3602 int stripe
= conf
->geo
.raid_disks
*
3603 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3604 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3605 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3607 /* Calculate max read-ahead size.
3608 * We need to readahead at least twice a whole stripe....
3611 stripe
/= conf
->geo
.near_copies
;
3612 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3613 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3614 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3618 if (md_integrity_register(mddev
))
3621 if (conf
->reshape_progress
!= MaxSector
) {
3622 unsigned long before_length
, after_length
;
3624 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3625 conf
->prev
.far_copies
);
3626 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3627 conf
->geo
.far_copies
);
3629 if (max(before_length
, after_length
) > min_offset_diff
) {
3630 /* This cannot work */
3631 printk("md/raid10: offset difference not enough to continue reshape\n");
3634 conf
->offset_diff
= min_offset_diff
;
3636 conf
->reshape_safe
= conf
->reshape_progress
;
3637 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3638 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3639 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3640 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3641 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3648 md_unregister_thread(&mddev
->thread
);
3649 if (conf
->r10bio_pool
)
3650 mempool_destroy(conf
->r10bio_pool
);
3651 safe_put_page(conf
->tmppage
);
3652 kfree(conf
->mirrors
);
3654 mddev
->private = NULL
;
3659 static int stop(struct mddev
*mddev
)
3661 struct r10conf
*conf
= mddev
->private;
3663 raise_barrier(conf
, 0);
3664 lower_barrier(conf
);
3666 md_unregister_thread(&mddev
->thread
);
3668 /* the unplug fn references 'conf'*/
3669 blk_sync_queue(mddev
->queue
);
3671 if (conf
->r10bio_pool
)
3672 mempool_destroy(conf
->r10bio_pool
);
3673 kfree(conf
->mirrors
);
3675 mddev
->private = NULL
;
3679 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3681 struct r10conf
*conf
= mddev
->private;
3685 raise_barrier(conf
, 0);
3688 lower_barrier(conf
);
3693 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3695 /* Resize of 'far' arrays is not supported.
3696 * For 'near' and 'offset' arrays we can set the
3697 * number of sectors used to be an appropriate multiple
3698 * of the chunk size.
3699 * For 'offset', this is far_copies*chunksize.
3700 * For 'near' the multiplier is the LCM of
3701 * near_copies and raid_disks.
3702 * So if far_copies > 1 && !far_offset, fail.
3703 * Else find LCM(raid_disks, near_copy)*far_copies and
3704 * multiply by chunk_size. Then round to this number.
3705 * This is mostly done by raid10_size()
3707 struct r10conf
*conf
= mddev
->private;
3708 sector_t oldsize
, size
;
3710 if (mddev
->reshape_position
!= MaxSector
)
3713 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3716 oldsize
= raid10_size(mddev
, 0, 0);
3717 size
= raid10_size(mddev
, sectors
, 0);
3718 if (mddev
->external_size
&&
3719 mddev
->array_sectors
> size
)
3721 if (mddev
->bitmap
) {
3722 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3726 md_set_array_sectors(mddev
, size
);
3727 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3728 revalidate_disk(mddev
->gendisk
);
3729 if (sectors
> mddev
->dev_sectors
&&
3730 mddev
->recovery_cp
> oldsize
) {
3731 mddev
->recovery_cp
= oldsize
;
3732 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3734 calc_sectors(conf
, sectors
);
3735 mddev
->dev_sectors
= conf
->dev_sectors
;
3736 mddev
->resync_max_sectors
= size
;
3740 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3742 struct md_rdev
*rdev
;
3743 struct r10conf
*conf
;
3745 if (mddev
->degraded
> 0) {
3746 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3748 return ERR_PTR(-EINVAL
);
3751 /* Set new parameters */
3752 mddev
->new_level
= 10;
3753 /* new layout: far_copies = 1, near_copies = 2 */
3754 mddev
->new_layout
= (1<<8) + 2;
3755 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3756 mddev
->delta_disks
= mddev
->raid_disks
;
3757 mddev
->raid_disks
*= 2;
3758 /* make sure it will be not marked as dirty */
3759 mddev
->recovery_cp
= MaxSector
;
3761 conf
= setup_conf(mddev
);
3762 if (!IS_ERR(conf
)) {
3763 rdev_for_each(rdev
, mddev
)
3764 if (rdev
->raid_disk
>= 0)
3765 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3772 static void *raid10_takeover(struct mddev
*mddev
)
3774 struct r0conf
*raid0_conf
;
3776 /* raid10 can take over:
3777 * raid0 - providing it has only two drives
3779 if (mddev
->level
== 0) {
3780 /* for raid0 takeover only one zone is supported */
3781 raid0_conf
= mddev
->private;
3782 if (raid0_conf
->nr_strip_zones
> 1) {
3783 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3784 " with more than one zone.\n",
3786 return ERR_PTR(-EINVAL
);
3788 return raid10_takeover_raid0(mddev
);
3790 return ERR_PTR(-EINVAL
);
3793 static int raid10_check_reshape(struct mddev
*mddev
)
3795 /* Called when there is a request to change
3796 * - layout (to ->new_layout)
3797 * - chunk size (to ->new_chunk_sectors)
3798 * - raid_disks (by delta_disks)
3799 * or when trying to restart a reshape that was ongoing.
3801 * We need to validate the request and possibly allocate
3802 * space if that might be an issue later.
3804 * Currently we reject any reshape of a 'far' mode array,
3805 * allow chunk size to change if new is generally acceptable,
3806 * allow raid_disks to increase, and allow
3807 * a switch between 'near' mode and 'offset' mode.
3809 struct r10conf
*conf
= mddev
->private;
3812 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3815 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3816 /* mustn't change number of copies */
3818 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3819 /* Cannot switch to 'far' mode */
3822 if (mddev
->array_sectors
& geo
.chunk_mask
)
3823 /* not factor of array size */
3826 if (!enough(conf
, -1))
3829 kfree(conf
->mirrors_new
);
3830 conf
->mirrors_new
= NULL
;
3831 if (mddev
->delta_disks
> 0) {
3832 /* allocate new 'mirrors' list */
3833 conf
->mirrors_new
= kzalloc(
3834 sizeof(struct raid10_info
)
3835 *(mddev
->raid_disks
+
3836 mddev
->delta_disks
),
3838 if (!conf
->mirrors_new
)
3845 * Need to check if array has failed when deciding whether to:
3847 * - remove non-faulty devices
3850 * This determination is simple when no reshape is happening.
3851 * However if there is a reshape, we need to carefully check
3852 * both the before and after sections.
3853 * This is because some failed devices may only affect one
3854 * of the two sections, and some non-in_sync devices may
3855 * be insync in the section most affected by failed devices.
3857 static int calc_degraded(struct r10conf
*conf
)
3859 int degraded
, degraded2
;
3864 /* 'prev' section first */
3865 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
3866 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3867 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3869 else if (!test_bit(In_sync
, &rdev
->flags
))
3870 /* When we can reduce the number of devices in
3871 * an array, this might not contribute to
3872 * 'degraded'. It does now.
3877 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
3881 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3882 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3883 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3885 else if (!test_bit(In_sync
, &rdev
->flags
)) {
3886 /* If reshape is increasing the number of devices,
3887 * this section has already been recovered, so
3888 * it doesn't contribute to degraded.
3891 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
3896 if (degraded2
> degraded
)
3901 static int raid10_start_reshape(struct mddev
*mddev
)
3903 /* A 'reshape' has been requested. This commits
3904 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3905 * This also checks if there are enough spares and adds them
3907 * We currently require enough spares to make the final
3908 * array non-degraded. We also require that the difference
3909 * between old and new data_offset - on each device - is
3910 * enough that we never risk over-writing.
3913 unsigned long before_length
, after_length
;
3914 sector_t min_offset_diff
= 0;
3917 struct r10conf
*conf
= mddev
->private;
3918 struct md_rdev
*rdev
;
3922 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
3925 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
3928 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3929 conf
->prev
.far_copies
);
3930 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3931 conf
->geo
.far_copies
);
3933 rdev_for_each(rdev
, mddev
) {
3934 if (!test_bit(In_sync
, &rdev
->flags
)
3935 && !test_bit(Faulty
, &rdev
->flags
))
3937 if (rdev
->raid_disk
>= 0) {
3938 long long diff
= (rdev
->new_data_offset
3939 - rdev
->data_offset
);
3940 if (!mddev
->reshape_backwards
)
3944 if (first
|| diff
< min_offset_diff
)
3945 min_offset_diff
= diff
;
3949 if (max(before_length
, after_length
) > min_offset_diff
)
3952 if (spares
< mddev
->delta_disks
)
3955 conf
->offset_diff
= min_offset_diff
;
3956 spin_lock_irq(&conf
->device_lock
);
3957 if (conf
->mirrors_new
) {
3958 memcpy(conf
->mirrors_new
, conf
->mirrors
,
3959 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
3961 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
3962 conf
->mirrors_old
= conf
->mirrors
;
3963 conf
->mirrors
= conf
->mirrors_new
;
3964 conf
->mirrors_new
= NULL
;
3966 setup_geo(&conf
->geo
, mddev
, geo_start
);
3968 if (mddev
->reshape_backwards
) {
3969 sector_t size
= raid10_size(mddev
, 0, 0);
3970 if (size
< mddev
->array_sectors
) {
3971 spin_unlock_irq(&conf
->device_lock
);
3972 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
3976 mddev
->resync_max_sectors
= size
;
3977 conf
->reshape_progress
= size
;
3979 conf
->reshape_progress
= 0;
3980 spin_unlock_irq(&conf
->device_lock
);
3982 if (mddev
->delta_disks
&& mddev
->bitmap
) {
3983 ret
= bitmap_resize(mddev
->bitmap
,
3984 raid10_size(mddev
, 0,
3985 conf
->geo
.raid_disks
),
3990 if (mddev
->delta_disks
> 0) {
3991 rdev_for_each(rdev
, mddev
)
3992 if (rdev
->raid_disk
< 0 &&
3993 !test_bit(Faulty
, &rdev
->flags
)) {
3994 if (raid10_add_disk(mddev
, rdev
) == 0) {
3995 if (rdev
->raid_disk
>=
3996 conf
->prev
.raid_disks
)
3997 set_bit(In_sync
, &rdev
->flags
);
3999 rdev
->recovery_offset
= 0;
4001 if (sysfs_link_rdev(mddev
, rdev
))
4002 /* Failure here is OK */;
4004 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4005 && !test_bit(Faulty
, &rdev
->flags
)) {
4006 /* This is a spare that was manually added */
4007 set_bit(In_sync
, &rdev
->flags
);
4010 /* When a reshape changes the number of devices,
4011 * ->degraded is measured against the larger of the
4012 * pre and post numbers.
4014 spin_lock_irq(&conf
->device_lock
);
4015 mddev
->degraded
= calc_degraded(conf
);
4016 spin_unlock_irq(&conf
->device_lock
);
4017 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4018 mddev
->reshape_position
= conf
->reshape_progress
;
4019 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4021 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4022 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4023 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4024 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4026 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4028 if (!mddev
->sync_thread
) {
4032 conf
->reshape_checkpoint
= jiffies
;
4033 md_wakeup_thread(mddev
->sync_thread
);
4034 md_new_event(mddev
);
4038 mddev
->recovery
= 0;
4039 spin_lock_irq(&conf
->device_lock
);
4040 conf
->geo
= conf
->prev
;
4041 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4042 rdev_for_each(rdev
, mddev
)
4043 rdev
->new_data_offset
= rdev
->data_offset
;
4045 conf
->reshape_progress
= MaxSector
;
4046 mddev
->reshape_position
= MaxSector
;
4047 spin_unlock_irq(&conf
->device_lock
);
4051 /* Calculate the last device-address that could contain
4052 * any block from the chunk that includes the array-address 's'
4053 * and report the next address.
4054 * i.e. the address returned will be chunk-aligned and after
4055 * any data that is in the chunk containing 's'.
4057 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4059 s
= (s
| geo
->chunk_mask
) + 1;
4060 s
>>= geo
->chunk_shift
;
4061 s
*= geo
->near_copies
;
4062 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4063 s
*= geo
->far_copies
;
4064 s
<<= geo
->chunk_shift
;
4068 /* Calculate the first device-address that could contain
4069 * any block from the chunk that includes the array-address 's'.
4070 * This too will be the start of a chunk
4072 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4074 s
>>= geo
->chunk_shift
;
4075 s
*= geo
->near_copies
;
4076 sector_div(s
, geo
->raid_disks
);
4077 s
*= geo
->far_copies
;
4078 s
<<= geo
->chunk_shift
;
4082 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4085 /* We simply copy at most one chunk (smallest of old and new)
4086 * at a time, possibly less if that exceeds RESYNC_PAGES,
4087 * or we hit a bad block or something.
4088 * This might mean we pause for normal IO in the middle of
4089 * a chunk, but that is not a problem was mddev->reshape_position
4090 * can record any location.
4092 * If we will want to write to a location that isn't
4093 * yet recorded as 'safe' (i.e. in metadata on disk) then
4094 * we need to flush all reshape requests and update the metadata.
4096 * When reshaping forwards (e.g. to more devices), we interpret
4097 * 'safe' as the earliest block which might not have been copied
4098 * down yet. We divide this by previous stripe size and multiply
4099 * by previous stripe length to get lowest device offset that we
4100 * cannot write to yet.
4101 * We interpret 'sector_nr' as an address that we want to write to.
4102 * From this we use last_device_address() to find where we might
4103 * write to, and first_device_address on the 'safe' position.
4104 * If this 'next' write position is after the 'safe' position,
4105 * we must update the metadata to increase the 'safe' position.
4107 * When reshaping backwards, we round in the opposite direction
4108 * and perform the reverse test: next write position must not be
4109 * less than current safe position.
4111 * In all this the minimum difference in data offsets
4112 * (conf->offset_diff - always positive) allows a bit of slack,
4113 * so next can be after 'safe', but not by more than offset_disk
4115 * We need to prepare all the bios here before we start any IO
4116 * to ensure the size we choose is acceptable to all devices.
4117 * The means one for each copy for write-out and an extra one for
4119 * We store the read-in bio in ->master_bio and the others in
4120 * ->devs[x].bio and ->devs[x].repl_bio.
4122 struct r10conf
*conf
= mddev
->private;
4123 struct r10bio
*r10_bio
;
4124 sector_t next
, safe
, last
;
4128 struct md_rdev
*rdev
;
4131 struct bio
*bio
, *read_bio
;
4132 int sectors_done
= 0;
4134 if (sector_nr
== 0) {
4135 /* If restarting in the middle, skip the initial sectors */
4136 if (mddev
->reshape_backwards
&&
4137 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4138 sector_nr
= (raid10_size(mddev
, 0, 0)
4139 - conf
->reshape_progress
);
4140 } else if (!mddev
->reshape_backwards
&&
4141 conf
->reshape_progress
> 0)
4142 sector_nr
= conf
->reshape_progress
;
4144 mddev
->curr_resync_completed
= sector_nr
;
4145 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4151 /* We don't use sector_nr to track where we are up to
4152 * as that doesn't work well for ->reshape_backwards.
4153 * So just use ->reshape_progress.
4155 if (mddev
->reshape_backwards
) {
4156 /* 'next' is the earliest device address that we might
4157 * write to for this chunk in the new layout
4159 next
= first_dev_address(conf
->reshape_progress
- 1,
4162 /* 'safe' is the last device address that we might read from
4163 * in the old layout after a restart
4165 safe
= last_dev_address(conf
->reshape_safe
- 1,
4168 if (next
+ conf
->offset_diff
< safe
)
4171 last
= conf
->reshape_progress
- 1;
4172 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4173 & conf
->prev
.chunk_mask
);
4174 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4175 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4177 /* 'next' is after the last device address that we
4178 * might write to for this chunk in the new layout
4180 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4182 /* 'safe' is the earliest device address that we might
4183 * read from in the old layout after a restart
4185 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4187 /* Need to update metadata if 'next' might be beyond 'safe'
4188 * as that would possibly corrupt data
4190 if (next
> safe
+ conf
->offset_diff
)
4193 sector_nr
= conf
->reshape_progress
;
4194 last
= sector_nr
| (conf
->geo
.chunk_mask
4195 & conf
->prev
.chunk_mask
);
4197 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4198 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4202 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4203 /* Need to update reshape_position in metadata */
4205 mddev
->reshape_position
= conf
->reshape_progress
;
4206 if (mddev
->reshape_backwards
)
4207 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4208 - conf
->reshape_progress
;
4210 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4211 conf
->reshape_checkpoint
= jiffies
;
4212 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4213 md_wakeup_thread(mddev
->thread
);
4214 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4215 kthread_should_stop());
4216 conf
->reshape_safe
= mddev
->reshape_position
;
4217 allow_barrier(conf
);
4221 /* Now schedule reads for blocks from sector_nr to last */
4222 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4223 raise_barrier(conf
, sectors_done
!= 0);
4224 atomic_set(&r10_bio
->remaining
, 0);
4225 r10_bio
->mddev
= mddev
;
4226 r10_bio
->sector
= sector_nr
;
4227 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4228 r10_bio
->sectors
= last
- sector_nr
+ 1;
4229 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4230 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4233 /* Cannot read from here, so need to record bad blocks
4234 * on all the target devices.
4237 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4238 return sectors_done
;
4241 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4243 read_bio
->bi_bdev
= rdev
->bdev
;
4244 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4245 + rdev
->data_offset
);
4246 read_bio
->bi_private
= r10_bio
;
4247 read_bio
->bi_end_io
= end_sync_read
;
4248 read_bio
->bi_rw
= READ
;
4249 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4250 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4251 read_bio
->bi_vcnt
= 0;
4252 read_bio
->bi_idx
= 0;
4253 read_bio
->bi_size
= 0;
4254 r10_bio
->master_bio
= read_bio
;
4255 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4257 /* Now find the locations in the new layout */
4258 __raid10_find_phys(&conf
->geo
, r10_bio
);
4261 read_bio
->bi_next
= NULL
;
4263 for (s
= 0; s
< conf
->copies
*2; s
++) {
4265 int d
= r10_bio
->devs
[s
/2].devnum
;
4266 struct md_rdev
*rdev2
;
4268 rdev2
= conf
->mirrors
[d
].replacement
;
4269 b
= r10_bio
->devs
[s
/2].repl_bio
;
4271 rdev2
= conf
->mirrors
[d
].rdev
;
4272 b
= r10_bio
->devs
[s
/2].bio
;
4274 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4276 b
->bi_bdev
= rdev2
->bdev
;
4277 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4278 b
->bi_private
= r10_bio
;
4279 b
->bi_end_io
= end_reshape_write
;
4281 b
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4282 b
->bi_flags
|= 1 << BIO_UPTODATE
;
4290 /* Now add as many pages as possible to all of these bios. */
4293 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4294 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4295 int len
= (max_sectors
- s
) << 9;
4296 if (len
> PAGE_SIZE
)
4298 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4300 if (bio_add_page(bio
, page
, len
, 0))
4303 /* Didn't fit, must stop */
4305 bio2
&& bio2
!= bio
;
4306 bio2
= bio2
->bi_next
) {
4307 /* Remove last page from this bio */
4309 bio2
->bi_size
-= len
;
4310 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4314 sector_nr
+= len
>> 9;
4315 nr_sectors
+= len
>> 9;
4318 r10_bio
->sectors
= nr_sectors
;
4320 /* Now submit the read */
4321 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4322 atomic_inc(&r10_bio
->remaining
);
4323 read_bio
->bi_next
= NULL
;
4324 generic_make_request(read_bio
);
4325 sector_nr
+= nr_sectors
;
4326 sectors_done
+= nr_sectors
;
4327 if (sector_nr
<= last
)
4330 /* Now that we have done the whole section we can
4331 * update reshape_progress
4333 if (mddev
->reshape_backwards
)
4334 conf
->reshape_progress
-= sectors_done
;
4336 conf
->reshape_progress
+= sectors_done
;
4338 return sectors_done
;
4341 static void end_reshape_request(struct r10bio
*r10_bio
);
4342 static int handle_reshape_read_error(struct mddev
*mddev
,
4343 struct r10bio
*r10_bio
);
4344 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4346 /* Reshape read completed. Hopefully we have a block
4348 * If we got a read error then we do sync 1-page reads from
4349 * elsewhere until we find the data - or give up.
4351 struct r10conf
*conf
= mddev
->private;
4354 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4355 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4356 /* Reshape has been aborted */
4357 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4361 /* We definitely have the data in the pages, schedule the
4364 atomic_set(&r10_bio
->remaining
, 1);
4365 for (s
= 0; s
< conf
->copies
*2; s
++) {
4367 int d
= r10_bio
->devs
[s
/2].devnum
;
4368 struct md_rdev
*rdev
;
4370 rdev
= conf
->mirrors
[d
].replacement
;
4371 b
= r10_bio
->devs
[s
/2].repl_bio
;
4373 rdev
= conf
->mirrors
[d
].rdev
;
4374 b
= r10_bio
->devs
[s
/2].bio
;
4376 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4378 atomic_inc(&rdev
->nr_pending
);
4379 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4380 atomic_inc(&r10_bio
->remaining
);
4382 generic_make_request(b
);
4384 end_reshape_request(r10_bio
);
4387 static void end_reshape(struct r10conf
*conf
)
4389 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4392 spin_lock_irq(&conf
->device_lock
);
4393 conf
->prev
= conf
->geo
;
4394 md_finish_reshape(conf
->mddev
);
4396 conf
->reshape_progress
= MaxSector
;
4397 spin_unlock_irq(&conf
->device_lock
);
4399 /* read-ahead size must cover two whole stripes, which is
4400 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4402 if (conf
->mddev
->queue
) {
4403 int stripe
= conf
->geo
.raid_disks
*
4404 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4405 stripe
/= conf
->geo
.near_copies
;
4406 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4407 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4413 static int handle_reshape_read_error(struct mddev
*mddev
,
4414 struct r10bio
*r10_bio
)
4416 /* Use sync reads to get the blocks from somewhere else */
4417 int sectors
= r10_bio
->sectors
;
4419 struct r10conf
*conf
= mddev
->private;
4422 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4424 r10b
.sector
= r10_bio
->sector
;
4425 __raid10_find_phys(&conf
->prev
, &r10b
);
4430 int first_slot
= slot
;
4432 if (s
> (PAGE_SIZE
>> 9))
4436 int d
= r10b
.devs
[slot
].devnum
;
4437 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4440 test_bit(Faulty
, &rdev
->flags
) ||
4441 !test_bit(In_sync
, &rdev
->flags
))
4444 addr
= r10b
.devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4445 success
= sync_page_io(rdev
,
4454 if (slot
>= conf
->copies
)
4456 if (slot
== first_slot
)
4460 /* couldn't read this block, must give up */
4461 set_bit(MD_RECOVERY_INTR
,
4471 static void end_reshape_write(struct bio
*bio
, int error
)
4473 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4474 struct r10bio
*r10_bio
= bio
->bi_private
;
4475 struct mddev
*mddev
= r10_bio
->mddev
;
4476 struct r10conf
*conf
= mddev
->private;
4480 struct md_rdev
*rdev
= NULL
;
4482 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4484 rdev
= conf
->mirrors
[d
].replacement
;
4487 rdev
= conf
->mirrors
[d
].rdev
;
4491 /* FIXME should record badblock */
4492 md_error(mddev
, rdev
);
4495 rdev_dec_pending(rdev
, mddev
);
4496 end_reshape_request(r10_bio
);
4499 static void end_reshape_request(struct r10bio
*r10_bio
)
4501 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4503 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4504 bio_put(r10_bio
->master_bio
);
4508 static void raid10_finish_reshape(struct mddev
*mddev
)
4510 struct r10conf
*conf
= mddev
->private;
4512 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4515 if (mddev
->delta_disks
> 0) {
4516 sector_t size
= raid10_size(mddev
, 0, 0);
4517 md_set_array_sectors(mddev
, size
);
4518 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4519 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4520 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4522 mddev
->resync_max_sectors
= size
;
4523 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4524 revalidate_disk(mddev
->gendisk
);
4527 for (d
= conf
->geo
.raid_disks
;
4528 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4530 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4532 clear_bit(In_sync
, &rdev
->flags
);
4533 rdev
= conf
->mirrors
[d
].replacement
;
4535 clear_bit(In_sync
, &rdev
->flags
);
4538 mddev
->layout
= mddev
->new_layout
;
4539 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4540 mddev
->reshape_position
= MaxSector
;
4541 mddev
->delta_disks
= 0;
4542 mddev
->reshape_backwards
= 0;
4545 static struct md_personality raid10_personality
=
4549 .owner
= THIS_MODULE
,
4550 .make_request
= make_request
,
4554 .error_handler
= error
,
4555 .hot_add_disk
= raid10_add_disk
,
4556 .hot_remove_disk
= raid10_remove_disk
,
4557 .spare_active
= raid10_spare_active
,
4558 .sync_request
= sync_request
,
4559 .quiesce
= raid10_quiesce
,
4560 .size
= raid10_size
,
4561 .resize
= raid10_resize
,
4562 .takeover
= raid10_takeover
,
4563 .check_reshape
= raid10_check_reshape
,
4564 .start_reshape
= raid10_start_reshape
,
4565 .finish_reshape
= raid10_finish_reshape
,
4568 static int __init
raid_init(void)
4570 return register_md_personality(&raid10_personality
);
4573 static void raid_exit(void)
4575 unregister_md_personality(&raid10_personality
);
4578 module_init(raid_init
);
4579 module_exit(raid_exit
);
4580 MODULE_LICENSE("GPL");
4581 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4582 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4583 MODULE_ALIAS("md-raid10");
4584 MODULE_ALIAS("md-level-10");
4586 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);