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 static int raid10_congested(void *data
, int bits
)
858 struct mddev
*mddev
= data
;
859 struct r10conf
*conf
= mddev
->private;
862 if ((bits
& (1 << BDI_async_congested
)) &&
863 conf
->pending_count
>= max_queued_requests
)
866 if (mddev_congested(mddev
, bits
))
870 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
873 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
874 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
875 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
877 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
884 static void flush_pending_writes(struct r10conf
*conf
)
886 /* Any writes that have been queued but are awaiting
887 * bitmap updates get flushed here.
889 spin_lock_irq(&conf
->device_lock
);
891 if (conf
->pending_bio_list
.head
) {
893 bio
= bio_list_get(&conf
->pending_bio_list
);
894 conf
->pending_count
= 0;
895 spin_unlock_irq(&conf
->device_lock
);
896 /* flush any pending bitmap writes to disk
897 * before proceeding w/ I/O */
898 bitmap_unplug(conf
->mddev
->bitmap
);
899 wake_up(&conf
->wait_barrier
);
901 while (bio
) { /* submit pending writes */
902 struct bio
*next
= bio
->bi_next
;
904 generic_make_request(bio
);
908 spin_unlock_irq(&conf
->device_lock
);
912 * Sometimes we need to suspend IO while we do something else,
913 * either some resync/recovery, or reconfigure the array.
914 * To do this we raise a 'barrier'.
915 * The 'barrier' is a counter that can be raised multiple times
916 * to count how many activities are happening which preclude
918 * We can only raise the barrier if there is no pending IO.
919 * i.e. if nr_pending == 0.
920 * We choose only to raise the barrier if no-one is waiting for the
921 * barrier to go down. This means that as soon as an IO request
922 * is ready, no other operations which require a barrier will start
923 * until the IO request has had a chance.
925 * So: regular IO calls 'wait_barrier'. When that returns there
926 * is no backgroup IO happening, It must arrange to call
927 * allow_barrier when it has finished its IO.
928 * backgroup IO calls must call raise_barrier. Once that returns
929 * there is no normal IO happeing. It must arrange to call
930 * lower_barrier when the particular background IO completes.
933 static void raise_barrier(struct r10conf
*conf
, int force
)
935 BUG_ON(force
&& !conf
->barrier
);
936 spin_lock_irq(&conf
->resync_lock
);
938 /* Wait until no block IO is waiting (unless 'force') */
939 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
940 conf
->resync_lock
, );
942 /* block any new IO from starting */
945 /* Now wait for all pending IO to complete */
946 wait_event_lock_irq(conf
->wait_barrier
,
947 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
948 conf
->resync_lock
, );
950 spin_unlock_irq(&conf
->resync_lock
);
953 static void lower_barrier(struct r10conf
*conf
)
956 spin_lock_irqsave(&conf
->resync_lock
, flags
);
958 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
959 wake_up(&conf
->wait_barrier
);
962 static void wait_barrier(struct r10conf
*conf
)
964 spin_lock_irq(&conf
->resync_lock
);
967 /* Wait for the barrier to drop.
968 * However if there are already pending
969 * requests (preventing the barrier from
970 * rising completely), and the
971 * pre-process bio queue isn't empty,
972 * then don't wait, as we need to empty
973 * that queue to get the nr_pending
976 wait_event_lock_irq(conf
->wait_barrier
,
980 !bio_list_empty(current
->bio_list
)),
986 spin_unlock_irq(&conf
->resync_lock
);
989 static void allow_barrier(struct r10conf
*conf
)
992 spin_lock_irqsave(&conf
->resync_lock
, flags
);
994 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
995 wake_up(&conf
->wait_barrier
);
998 static void freeze_array(struct r10conf
*conf
)
1000 /* stop syncio and normal IO and wait for everything to
1002 * We increment barrier and nr_waiting, and then
1003 * wait until nr_pending match nr_queued+1
1004 * This is called in the context of one normal IO request
1005 * that has failed. Thus any sync request that might be pending
1006 * will be blocked by nr_pending, and we need to wait for
1007 * pending IO requests to complete or be queued for re-try.
1008 * Thus the number queued (nr_queued) plus this request (1)
1009 * must match the number of pending IOs (nr_pending) before
1012 spin_lock_irq(&conf
->resync_lock
);
1015 wait_event_lock_irq(conf
->wait_barrier
,
1016 conf
->nr_pending
== conf
->nr_queued
+1,
1018 flush_pending_writes(conf
));
1020 spin_unlock_irq(&conf
->resync_lock
);
1023 static void unfreeze_array(struct r10conf
*conf
)
1025 /* reverse the effect of the freeze */
1026 spin_lock_irq(&conf
->resync_lock
);
1029 wake_up(&conf
->wait_barrier
);
1030 spin_unlock_irq(&conf
->resync_lock
);
1033 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1034 struct md_rdev
*rdev
)
1036 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1037 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1038 return rdev
->data_offset
;
1040 return rdev
->new_data_offset
;
1043 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1045 struct r10conf
*conf
= mddev
->private;
1046 struct r10bio
*r10_bio
;
1047 struct bio
*read_bio
;
1049 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1050 int chunk_sects
= chunk_mask
+ 1;
1051 const int rw
= bio_data_dir(bio
);
1052 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1053 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1054 unsigned long flags
;
1055 struct md_rdev
*blocked_rdev
;
1056 int sectors_handled
;
1060 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1061 md_flush_request(mddev
, bio
);
1065 /* If this request crosses a chunk boundary, we need to
1066 * split it. This will only happen for 1 PAGE (or less) requests.
1068 if (unlikely((bio
->bi_sector
& chunk_mask
) + (bio
->bi_size
>> 9)
1070 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1071 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1072 struct bio_pair
*bp
;
1073 /* Sanity check -- queue functions should prevent this happening */
1074 if (bio
->bi_vcnt
!= 1 ||
1077 /* This is a one page bio that upper layers
1078 * refuse to split for us, so we need to split it.
1081 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
1083 /* Each of these 'make_request' calls will call 'wait_barrier'.
1084 * If the first succeeds but the second blocks due to the resync
1085 * thread raising the barrier, we will deadlock because the
1086 * IO to the underlying device will be queued in generic_make_request
1087 * and will never complete, so will never reduce nr_pending.
1088 * So increment nr_waiting here so no new raise_barriers will
1089 * succeed, and so the second wait_barrier cannot block.
1091 spin_lock_irq(&conf
->resync_lock
);
1093 spin_unlock_irq(&conf
->resync_lock
);
1095 make_request(mddev
, &bp
->bio1
);
1096 make_request(mddev
, &bp
->bio2
);
1098 spin_lock_irq(&conf
->resync_lock
);
1100 wake_up(&conf
->wait_barrier
);
1101 spin_unlock_irq(&conf
->resync_lock
);
1103 bio_pair_release(bp
);
1106 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1107 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1108 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
1114 md_write_start(mddev
, bio
);
1117 * Register the new request and wait if the reconstruction
1118 * thread has put up a bar for new requests.
1119 * Continue immediately if no resync is active currently.
1123 sectors
= bio
->bi_size
>> 9;
1124 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1125 bio
->bi_sector
< conf
->reshape_progress
&&
1126 bio
->bi_sector
+ sectors
> conf
->reshape_progress
) {
1127 /* IO spans the reshape position. Need to wait for
1130 allow_barrier(conf
);
1131 wait_event(conf
->wait_barrier
,
1132 conf
->reshape_progress
<= bio
->bi_sector
||
1133 conf
->reshape_progress
>= bio
->bi_sector
+ sectors
);
1136 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1137 bio_data_dir(bio
) == WRITE
&&
1138 (mddev
->reshape_backwards
1139 ? (bio
->bi_sector
< conf
->reshape_safe
&&
1140 bio
->bi_sector
+ sectors
> conf
->reshape_progress
)
1141 : (bio
->bi_sector
+ sectors
> conf
->reshape_safe
&&
1142 bio
->bi_sector
< conf
->reshape_progress
))) {
1143 /* Need to update reshape_position in metadata */
1144 mddev
->reshape_position
= conf
->reshape_progress
;
1145 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1146 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1147 md_wakeup_thread(mddev
->thread
);
1148 wait_event(mddev
->sb_wait
,
1149 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1151 conf
->reshape_safe
= mddev
->reshape_position
;
1154 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1156 r10_bio
->master_bio
= bio
;
1157 r10_bio
->sectors
= sectors
;
1159 r10_bio
->mddev
= mddev
;
1160 r10_bio
->sector
= bio
->bi_sector
;
1163 /* We might need to issue multiple reads to different
1164 * devices if there are bad blocks around, so we keep
1165 * track of the number of reads in bio->bi_phys_segments.
1166 * If this is 0, there is only one r10_bio and no locking
1167 * will be needed when the request completes. If it is
1168 * non-zero, then it is the number of not-completed requests.
1170 bio
->bi_phys_segments
= 0;
1171 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1175 * read balancing logic:
1177 struct md_rdev
*rdev
;
1181 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1183 raid_end_bio_io(r10_bio
);
1186 slot
= r10_bio
->read_slot
;
1188 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1189 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1192 r10_bio
->devs
[slot
].bio
= read_bio
;
1193 r10_bio
->devs
[slot
].rdev
= rdev
;
1195 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1196 choose_data_offset(r10_bio
, rdev
);
1197 read_bio
->bi_bdev
= rdev
->bdev
;
1198 read_bio
->bi_end_io
= raid10_end_read_request
;
1199 read_bio
->bi_rw
= READ
| do_sync
;
1200 read_bio
->bi_private
= r10_bio
;
1202 if (max_sectors
< r10_bio
->sectors
) {
1203 /* Could not read all from this device, so we will
1204 * need another r10_bio.
1206 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1208 r10_bio
->sectors
= max_sectors
;
1209 spin_lock_irq(&conf
->device_lock
);
1210 if (bio
->bi_phys_segments
== 0)
1211 bio
->bi_phys_segments
= 2;
1213 bio
->bi_phys_segments
++;
1214 spin_unlock(&conf
->device_lock
);
1215 /* Cannot call generic_make_request directly
1216 * as that will be queued in __generic_make_request
1217 * and subsequent mempool_alloc might block
1218 * waiting for it. so hand bio over to raid10d.
1220 reschedule_retry(r10_bio
);
1222 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1224 r10_bio
->master_bio
= bio
;
1225 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
1228 r10_bio
->mddev
= mddev
;
1229 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1232 generic_make_request(read_bio
);
1239 if (conf
->pending_count
>= max_queued_requests
) {
1240 md_wakeup_thread(mddev
->thread
);
1241 wait_event(conf
->wait_barrier
,
1242 conf
->pending_count
< max_queued_requests
);
1244 /* first select target devices under rcu_lock and
1245 * inc refcount on their rdev. Record them by setting
1247 * If there are known/acknowledged bad blocks on any device
1248 * on which we have seen a write error, we want to avoid
1249 * writing to those blocks. This potentially requires several
1250 * writes to write around the bad blocks. Each set of writes
1251 * gets its own r10_bio with a set of bios attached. The number
1252 * of r10_bios is recored in bio->bi_phys_segments just as with
1256 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1257 raid10_find_phys(conf
, r10_bio
);
1259 blocked_rdev
= NULL
;
1261 max_sectors
= r10_bio
->sectors
;
1263 for (i
= 0; i
< conf
->copies
; i
++) {
1264 int d
= r10_bio
->devs
[i
].devnum
;
1265 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1266 struct md_rdev
*rrdev
= rcu_dereference(
1267 conf
->mirrors
[d
].replacement
);
1270 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1271 atomic_inc(&rdev
->nr_pending
);
1272 blocked_rdev
= rdev
;
1275 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1276 atomic_inc(&rrdev
->nr_pending
);
1277 blocked_rdev
= rrdev
;
1280 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1281 || test_bit(Unmerged
, &rrdev
->flags
)))
1284 r10_bio
->devs
[i
].bio
= NULL
;
1285 r10_bio
->devs
[i
].repl_bio
= NULL
;
1286 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
1287 test_bit(Unmerged
, &rdev
->flags
)) {
1288 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1291 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1293 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1297 is_bad
= is_badblock(rdev
, dev_sector
,
1299 &first_bad
, &bad_sectors
);
1301 /* Mustn't write here until the bad block
1304 atomic_inc(&rdev
->nr_pending
);
1305 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1306 blocked_rdev
= rdev
;
1309 if (is_bad
&& first_bad
<= dev_sector
) {
1310 /* Cannot write here at all */
1311 bad_sectors
-= (dev_sector
- first_bad
);
1312 if (bad_sectors
< max_sectors
)
1313 /* Mustn't write more than bad_sectors
1314 * to other devices yet
1316 max_sectors
= bad_sectors
;
1317 /* We don't set R10BIO_Degraded as that
1318 * only applies if the disk is missing,
1319 * so it might be re-added, and we want to
1320 * know to recover this chunk.
1321 * In this case the device is here, and the
1322 * fact that this chunk is not in-sync is
1323 * recorded in the bad block log.
1328 int good_sectors
= first_bad
- dev_sector
;
1329 if (good_sectors
< max_sectors
)
1330 max_sectors
= good_sectors
;
1333 r10_bio
->devs
[i
].bio
= bio
;
1334 atomic_inc(&rdev
->nr_pending
);
1336 r10_bio
->devs
[i
].repl_bio
= bio
;
1337 atomic_inc(&rrdev
->nr_pending
);
1342 if (unlikely(blocked_rdev
)) {
1343 /* Have to wait for this device to get unblocked, then retry */
1347 for (j
= 0; j
< i
; j
++) {
1348 if (r10_bio
->devs
[j
].bio
) {
1349 d
= r10_bio
->devs
[j
].devnum
;
1350 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1352 if (r10_bio
->devs
[j
].repl_bio
) {
1353 struct md_rdev
*rdev
;
1354 d
= r10_bio
->devs
[j
].devnum
;
1355 rdev
= conf
->mirrors
[d
].replacement
;
1357 /* Race with remove_disk */
1359 rdev
= conf
->mirrors
[d
].rdev
;
1361 rdev_dec_pending(rdev
, mddev
);
1364 allow_barrier(conf
);
1365 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1370 if (max_sectors
< r10_bio
->sectors
) {
1371 /* We are splitting this into multiple parts, so
1372 * we need to prepare for allocating another r10_bio.
1374 r10_bio
->sectors
= max_sectors
;
1375 spin_lock_irq(&conf
->device_lock
);
1376 if (bio
->bi_phys_segments
== 0)
1377 bio
->bi_phys_segments
= 2;
1379 bio
->bi_phys_segments
++;
1380 spin_unlock_irq(&conf
->device_lock
);
1382 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1384 atomic_set(&r10_bio
->remaining
, 1);
1385 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1387 for (i
= 0; i
< conf
->copies
; i
++) {
1389 int d
= r10_bio
->devs
[i
].devnum
;
1390 if (!r10_bio
->devs
[i
].bio
)
1393 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1394 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1396 r10_bio
->devs
[i
].bio
= mbio
;
1398 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1399 choose_data_offset(r10_bio
,
1400 conf
->mirrors
[d
].rdev
));
1401 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1402 mbio
->bi_end_io
= raid10_end_write_request
;
1403 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
1404 mbio
->bi_private
= r10_bio
;
1406 atomic_inc(&r10_bio
->remaining
);
1407 spin_lock_irqsave(&conf
->device_lock
, flags
);
1408 bio_list_add(&conf
->pending_bio_list
, mbio
);
1409 conf
->pending_count
++;
1410 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1411 if (!mddev_check_plugged(mddev
))
1412 md_wakeup_thread(mddev
->thread
);
1414 if (!r10_bio
->devs
[i
].repl_bio
)
1417 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1418 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1420 r10_bio
->devs
[i
].repl_bio
= mbio
;
1422 /* We are actively writing to the original device
1423 * so it cannot disappear, so the replacement cannot
1426 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1429 conf
->mirrors
[d
].replacement
));
1430 mbio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
1431 mbio
->bi_end_io
= raid10_end_write_request
;
1432 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
;
1433 mbio
->bi_private
= r10_bio
;
1435 atomic_inc(&r10_bio
->remaining
);
1436 spin_lock_irqsave(&conf
->device_lock
, flags
);
1437 bio_list_add(&conf
->pending_bio_list
, mbio
);
1438 conf
->pending_count
++;
1439 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1440 if (!mddev_check_plugged(mddev
))
1441 md_wakeup_thread(mddev
->thread
);
1444 /* Don't remove the bias on 'remaining' (one_write_done) until
1445 * after checking if we need to go around again.
1448 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1449 one_write_done(r10_bio
);
1450 /* We need another r10_bio. It has already been counted
1451 * in bio->bi_phys_segments.
1453 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1455 r10_bio
->master_bio
= bio
;
1456 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1458 r10_bio
->mddev
= mddev
;
1459 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1463 one_write_done(r10_bio
);
1465 /* In case raid10d snuck in to freeze_array */
1466 wake_up(&conf
->wait_barrier
);
1469 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1471 struct r10conf
*conf
= mddev
->private;
1474 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1475 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1476 if (conf
->geo
.near_copies
> 1)
1477 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1478 if (conf
->geo
.far_copies
> 1) {
1479 if (conf
->geo
.far_offset
)
1480 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1482 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1484 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1485 conf
->geo
.raid_disks
- mddev
->degraded
);
1486 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1487 seq_printf(seq
, "%s",
1488 conf
->mirrors
[i
].rdev
&&
1489 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1490 seq_printf(seq
, "]");
1493 /* check if there are enough drives for
1494 * every block to appear on atleast one.
1495 * Don't consider the device numbered 'ignore'
1496 * as we might be about to remove it.
1498 static int _enough(struct r10conf
*conf
, struct geom
*geo
, int ignore
)
1503 int n
= conf
->copies
;
1506 if (conf
->mirrors
[first
].rdev
&&
1509 first
= (first
+1) % geo
->raid_disks
;
1513 } while (first
!= 0);
1517 static int enough(struct r10conf
*conf
, int ignore
)
1519 return _enough(conf
, &conf
->geo
, ignore
) &&
1520 _enough(conf
, &conf
->prev
, ignore
);
1523 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1525 char b
[BDEVNAME_SIZE
];
1526 struct r10conf
*conf
= mddev
->private;
1529 * If it is not operational, then we have already marked it as dead
1530 * else if it is the last working disks, ignore the error, let the
1531 * next level up know.
1532 * else mark the drive as failed
1534 if (test_bit(In_sync
, &rdev
->flags
)
1535 && !enough(conf
, rdev
->raid_disk
))
1537 * Don't fail the drive, just return an IO error.
1540 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1541 unsigned long flags
;
1542 spin_lock_irqsave(&conf
->device_lock
, flags
);
1544 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1546 * if recovery is running, make sure it aborts.
1548 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1550 set_bit(Blocked
, &rdev
->flags
);
1551 set_bit(Faulty
, &rdev
->flags
);
1552 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1554 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1555 "md/raid10:%s: Operation continuing on %d devices.\n",
1556 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1557 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1560 static void print_conf(struct r10conf
*conf
)
1563 struct raid10_info
*tmp
;
1565 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1567 printk(KERN_DEBUG
"(!conf)\n");
1570 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1571 conf
->geo
.raid_disks
);
1573 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1574 char b
[BDEVNAME_SIZE
];
1575 tmp
= conf
->mirrors
+ i
;
1577 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1578 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1579 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1580 bdevname(tmp
->rdev
->bdev
,b
));
1584 static void close_sync(struct r10conf
*conf
)
1587 allow_barrier(conf
);
1589 mempool_destroy(conf
->r10buf_pool
);
1590 conf
->r10buf_pool
= NULL
;
1593 static int raid10_spare_active(struct mddev
*mddev
)
1596 struct r10conf
*conf
= mddev
->private;
1597 struct raid10_info
*tmp
;
1599 unsigned long flags
;
1602 * Find all non-in_sync disks within the RAID10 configuration
1603 * and mark them in_sync
1605 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1606 tmp
= conf
->mirrors
+ i
;
1607 if (tmp
->replacement
1608 && tmp
->replacement
->recovery_offset
== MaxSector
1609 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1610 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1611 /* Replacement has just become active */
1613 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1616 /* Replaced device not technically faulty,
1617 * but we need to be sure it gets removed
1618 * and never re-added.
1620 set_bit(Faulty
, &tmp
->rdev
->flags
);
1621 sysfs_notify_dirent_safe(
1622 tmp
->rdev
->sysfs_state
);
1624 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1625 } else if (tmp
->rdev
1626 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1627 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1629 sysfs_notify_dirent(tmp
->rdev
->sysfs_state
);
1632 spin_lock_irqsave(&conf
->device_lock
, flags
);
1633 mddev
->degraded
-= count
;
1634 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1641 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1643 struct r10conf
*conf
= mddev
->private;
1647 int last
= conf
->geo
.raid_disks
- 1;
1648 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1650 if (mddev
->recovery_cp
< MaxSector
)
1651 /* only hot-add to in-sync arrays, as recovery is
1652 * very different from resync
1655 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, &conf
->prev
, -1))
1658 if (rdev
->raid_disk
>= 0)
1659 first
= last
= rdev
->raid_disk
;
1661 if (q
->merge_bvec_fn
) {
1662 set_bit(Unmerged
, &rdev
->flags
);
1663 mddev
->merge_check_needed
= 1;
1666 if (rdev
->saved_raid_disk
>= first
&&
1667 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1668 mirror
= rdev
->saved_raid_disk
;
1671 for ( ; mirror
<= last
; mirror
++) {
1672 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1673 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1676 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1677 p
->replacement
!= NULL
)
1679 clear_bit(In_sync
, &rdev
->flags
);
1680 set_bit(Replacement
, &rdev
->flags
);
1681 rdev
->raid_disk
= mirror
;
1683 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1684 rdev
->data_offset
<< 9);
1686 rcu_assign_pointer(p
->replacement
, rdev
);
1690 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1691 rdev
->data_offset
<< 9);
1693 p
->head_position
= 0;
1694 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1695 rdev
->raid_disk
= mirror
;
1697 if (rdev
->saved_raid_disk
!= mirror
)
1699 rcu_assign_pointer(p
->rdev
, rdev
);
1702 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1703 /* Some requests might not have seen this new
1704 * merge_bvec_fn. We must wait for them to complete
1705 * before merging the device fully.
1706 * First we make sure any code which has tested
1707 * our function has submitted the request, then
1708 * we wait for all outstanding requests to complete.
1710 synchronize_sched();
1711 raise_barrier(conf
, 0);
1712 lower_barrier(conf
);
1713 clear_bit(Unmerged
, &rdev
->flags
);
1715 md_integrity_add_rdev(rdev
, mddev
);
1720 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1722 struct r10conf
*conf
= mddev
->private;
1724 int number
= rdev
->raid_disk
;
1725 struct md_rdev
**rdevp
;
1726 struct raid10_info
*p
= conf
->mirrors
+ number
;
1729 if (rdev
== p
->rdev
)
1731 else if (rdev
== p
->replacement
)
1732 rdevp
= &p
->replacement
;
1736 if (test_bit(In_sync
, &rdev
->flags
) ||
1737 atomic_read(&rdev
->nr_pending
)) {
1741 /* Only remove faulty devices if recovery
1744 if (!test_bit(Faulty
, &rdev
->flags
) &&
1745 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1746 (!p
->replacement
|| p
->replacement
== rdev
) &&
1747 number
< conf
->geo
.raid_disks
&&
1754 if (atomic_read(&rdev
->nr_pending
)) {
1755 /* lost the race, try later */
1759 } else if (p
->replacement
) {
1760 /* We must have just cleared 'rdev' */
1761 p
->rdev
= p
->replacement
;
1762 clear_bit(Replacement
, &p
->replacement
->flags
);
1763 smp_mb(); /* Make sure other CPUs may see both as identical
1764 * but will never see neither -- if they are careful.
1766 p
->replacement
= NULL
;
1767 clear_bit(WantReplacement
, &rdev
->flags
);
1769 /* We might have just remove the Replacement as faulty
1770 * Clear the flag just in case
1772 clear_bit(WantReplacement
, &rdev
->flags
);
1774 err
= md_integrity_register(mddev
);
1783 static void end_sync_read(struct bio
*bio
, int error
)
1785 struct r10bio
*r10_bio
= bio
->bi_private
;
1786 struct r10conf
*conf
= r10_bio
->mddev
->private;
1789 if (bio
== r10_bio
->master_bio
) {
1790 /* this is a reshape read */
1791 d
= r10_bio
->read_slot
; /* really the read dev */
1793 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1795 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1796 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1798 /* The write handler will notice the lack of
1799 * R10BIO_Uptodate and record any errors etc
1801 atomic_add(r10_bio
->sectors
,
1802 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1804 /* for reconstruct, we always reschedule after a read.
1805 * for resync, only after all reads
1807 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1808 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1809 atomic_dec_and_test(&r10_bio
->remaining
)) {
1810 /* we have read all the blocks,
1811 * do the comparison in process context in raid10d
1813 reschedule_retry(r10_bio
);
1817 static void end_sync_request(struct r10bio
*r10_bio
)
1819 struct mddev
*mddev
= r10_bio
->mddev
;
1821 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1822 if (r10_bio
->master_bio
== NULL
) {
1823 /* the primary of several recovery bios */
1824 sector_t s
= r10_bio
->sectors
;
1825 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1826 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1827 reschedule_retry(r10_bio
);
1830 md_done_sync(mddev
, s
, 1);
1833 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1834 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1835 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1836 reschedule_retry(r10_bio
);
1844 static void end_sync_write(struct bio
*bio
, int error
)
1846 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1847 struct r10bio
*r10_bio
= bio
->bi_private
;
1848 struct mddev
*mddev
= r10_bio
->mddev
;
1849 struct r10conf
*conf
= mddev
->private;
1855 struct md_rdev
*rdev
= NULL
;
1857 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1859 rdev
= conf
->mirrors
[d
].replacement
;
1861 rdev
= conf
->mirrors
[d
].rdev
;
1865 md_error(mddev
, rdev
);
1867 set_bit(WriteErrorSeen
, &rdev
->flags
);
1868 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1869 set_bit(MD_RECOVERY_NEEDED
,
1870 &rdev
->mddev
->recovery
);
1871 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1873 } else if (is_badblock(rdev
,
1874 r10_bio
->devs
[slot
].addr
,
1876 &first_bad
, &bad_sectors
))
1877 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1879 rdev_dec_pending(rdev
, mddev
);
1881 end_sync_request(r10_bio
);
1885 * Note: sync and recover and handled very differently for raid10
1886 * This code is for resync.
1887 * For resync, we read through virtual addresses and read all blocks.
1888 * If there is any error, we schedule a write. The lowest numbered
1889 * drive is authoritative.
1890 * However requests come for physical address, so we need to map.
1891 * For every physical address there are raid_disks/copies virtual addresses,
1892 * which is always are least one, but is not necessarly an integer.
1893 * This means that a physical address can span multiple chunks, so we may
1894 * have to submit multiple io requests for a single sync request.
1897 * We check if all blocks are in-sync and only write to blocks that
1900 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1902 struct r10conf
*conf
= mddev
->private;
1904 struct bio
*tbio
, *fbio
;
1907 atomic_set(&r10_bio
->remaining
, 1);
1909 /* find the first device with a block */
1910 for (i
=0; i
<conf
->copies
; i
++)
1911 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1914 if (i
== conf
->copies
)
1918 fbio
= r10_bio
->devs
[i
].bio
;
1920 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
1921 /* now find blocks with errors */
1922 for (i
=0 ; i
< conf
->copies
; i
++) {
1925 tbio
= r10_bio
->devs
[i
].bio
;
1927 if (tbio
->bi_end_io
!= end_sync_read
)
1931 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1932 /* We know that the bi_io_vec layout is the same for
1933 * both 'first' and 'i', so we just compare them.
1934 * All vec entries are PAGE_SIZE;
1936 for (j
= 0; j
< vcnt
; j
++)
1937 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1938 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1939 fbio
->bi_io_vec
[j
].bv_len
))
1943 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1944 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1945 /* Don't fix anything. */
1948 /* Ok, we need to write this bio, either to correct an
1949 * inconsistency or to correct an unreadable block.
1950 * First we need to fixup bv_offset, bv_len and
1951 * bi_vecs, as the read request might have corrupted these
1953 tbio
->bi_vcnt
= vcnt
;
1954 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1956 tbio
->bi_phys_segments
= 0;
1957 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1958 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1959 tbio
->bi_next
= NULL
;
1960 tbio
->bi_rw
= WRITE
;
1961 tbio
->bi_private
= r10_bio
;
1962 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1964 for (j
=0; j
< vcnt
; j
++) {
1965 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1966 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1968 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1969 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1972 tbio
->bi_end_io
= end_sync_write
;
1974 d
= r10_bio
->devs
[i
].devnum
;
1975 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1976 atomic_inc(&r10_bio
->remaining
);
1977 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1979 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
1980 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1981 generic_make_request(tbio
);
1984 /* Now write out to any replacement devices
1987 for (i
= 0; i
< conf
->copies
; i
++) {
1990 tbio
= r10_bio
->devs
[i
].repl_bio
;
1991 if (!tbio
|| !tbio
->bi_end_io
)
1993 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
1994 && r10_bio
->devs
[i
].bio
!= fbio
)
1995 for (j
= 0; j
< vcnt
; j
++)
1996 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1997 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1999 d
= r10_bio
->devs
[i
].devnum
;
2000 atomic_inc(&r10_bio
->remaining
);
2001 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2002 tbio
->bi_size
>> 9);
2003 generic_make_request(tbio
);
2007 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2008 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2014 * Now for the recovery code.
2015 * Recovery happens across physical sectors.
2016 * We recover all non-is_sync drives by finding the virtual address of
2017 * each, and then choose a working drive that also has that virt address.
2018 * There is a separate r10_bio for each non-in_sync drive.
2019 * Only the first two slots are in use. The first for reading,
2020 * The second for writing.
2023 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2025 /* We got a read error during recovery.
2026 * We repeat the read in smaller page-sized sections.
2027 * If a read succeeds, write it to the new device or record
2028 * a bad block if we cannot.
2029 * If a read fails, record a bad block on both old and
2032 struct mddev
*mddev
= r10_bio
->mddev
;
2033 struct r10conf
*conf
= mddev
->private;
2034 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2036 int sectors
= r10_bio
->sectors
;
2038 int dr
= r10_bio
->devs
[0].devnum
;
2039 int dw
= r10_bio
->devs
[1].devnum
;
2043 struct md_rdev
*rdev
;
2047 if (s
> (PAGE_SIZE
>>9))
2050 rdev
= conf
->mirrors
[dr
].rdev
;
2051 addr
= r10_bio
->devs
[0].addr
+ sect
,
2052 ok
= sync_page_io(rdev
,
2055 bio
->bi_io_vec
[idx
].bv_page
,
2058 rdev
= conf
->mirrors
[dw
].rdev
;
2059 addr
= r10_bio
->devs
[1].addr
+ sect
;
2060 ok
= sync_page_io(rdev
,
2063 bio
->bi_io_vec
[idx
].bv_page
,
2066 set_bit(WriteErrorSeen
, &rdev
->flags
);
2067 if (!test_and_set_bit(WantReplacement
,
2069 set_bit(MD_RECOVERY_NEEDED
,
2070 &rdev
->mddev
->recovery
);
2074 /* We don't worry if we cannot set a bad block -
2075 * it really is bad so there is no loss in not
2078 rdev_set_badblocks(rdev
, addr
, s
, 0);
2080 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2081 /* need bad block on destination too */
2082 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2083 addr
= r10_bio
->devs
[1].addr
+ sect
;
2084 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2086 /* just abort the recovery */
2088 "md/raid10:%s: recovery aborted"
2089 " due to read error\n",
2092 conf
->mirrors
[dw
].recovery_disabled
2093 = mddev
->recovery_disabled
;
2094 set_bit(MD_RECOVERY_INTR
,
2107 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2109 struct r10conf
*conf
= mddev
->private;
2111 struct bio
*wbio
, *wbio2
;
2113 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2114 fix_recovery_read_error(r10_bio
);
2115 end_sync_request(r10_bio
);
2120 * share the pages with the first bio
2121 * and submit the write request
2123 d
= r10_bio
->devs
[1].devnum
;
2124 wbio
= r10_bio
->devs
[1].bio
;
2125 wbio2
= r10_bio
->devs
[1].repl_bio
;
2126 if (wbio
->bi_end_io
) {
2127 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2128 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
2129 generic_make_request(wbio
);
2131 if (wbio2
&& wbio2
->bi_end_io
) {
2132 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2133 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2134 wbio2
->bi_size
>> 9);
2135 generic_make_request(wbio2
);
2141 * Used by fix_read_error() to decay the per rdev read_errors.
2142 * We halve the read error count for every hour that has elapsed
2143 * since the last recorded read error.
2146 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2148 struct timespec cur_time_mon
;
2149 unsigned long hours_since_last
;
2150 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2152 ktime_get_ts(&cur_time_mon
);
2154 if (rdev
->last_read_error
.tv_sec
== 0 &&
2155 rdev
->last_read_error
.tv_nsec
== 0) {
2156 /* first time we've seen a read error */
2157 rdev
->last_read_error
= cur_time_mon
;
2161 hours_since_last
= (cur_time_mon
.tv_sec
-
2162 rdev
->last_read_error
.tv_sec
) / 3600;
2164 rdev
->last_read_error
= cur_time_mon
;
2167 * if hours_since_last is > the number of bits in read_errors
2168 * just set read errors to 0. We do this to avoid
2169 * overflowing the shift of read_errors by hours_since_last.
2171 if (hours_since_last
>= 8 * sizeof(read_errors
))
2172 atomic_set(&rdev
->read_errors
, 0);
2174 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2177 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2178 int sectors
, struct page
*page
, int rw
)
2183 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2184 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2186 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2190 set_bit(WriteErrorSeen
, &rdev
->flags
);
2191 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2192 set_bit(MD_RECOVERY_NEEDED
,
2193 &rdev
->mddev
->recovery
);
2195 /* need to record an error - either for the block or the device */
2196 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2197 md_error(rdev
->mddev
, rdev
);
2202 * This is a kernel thread which:
2204 * 1. Retries failed read operations on working mirrors.
2205 * 2. Updates the raid superblock when problems encounter.
2206 * 3. Performs writes following reads for array synchronising.
2209 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2211 int sect
= 0; /* Offset from r10_bio->sector */
2212 int sectors
= r10_bio
->sectors
;
2213 struct md_rdev
*rdev
;
2214 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2215 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2217 /* still own a reference to this rdev, so it cannot
2218 * have been cleared recently.
2220 rdev
= conf
->mirrors
[d
].rdev
;
2222 if (test_bit(Faulty
, &rdev
->flags
))
2223 /* drive has already been failed, just ignore any
2224 more fix_read_error() attempts */
2227 check_decay_read_errors(mddev
, rdev
);
2228 atomic_inc(&rdev
->read_errors
);
2229 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2230 char b
[BDEVNAME_SIZE
];
2231 bdevname(rdev
->bdev
, b
);
2234 "md/raid10:%s: %s: Raid device exceeded "
2235 "read_error threshold [cur %d:max %d]\n",
2237 atomic_read(&rdev
->read_errors
), max_read_errors
);
2239 "md/raid10:%s: %s: Failing raid device\n",
2241 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2242 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2248 int sl
= r10_bio
->read_slot
;
2252 if (s
> (PAGE_SIZE
>>9))
2260 d
= r10_bio
->devs
[sl
].devnum
;
2261 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2263 !test_bit(Unmerged
, &rdev
->flags
) &&
2264 test_bit(In_sync
, &rdev
->flags
) &&
2265 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2266 &first_bad
, &bad_sectors
) == 0) {
2267 atomic_inc(&rdev
->nr_pending
);
2269 success
= sync_page_io(rdev
,
2270 r10_bio
->devs
[sl
].addr
+
2273 conf
->tmppage
, READ
, false);
2274 rdev_dec_pending(rdev
, mddev
);
2280 if (sl
== conf
->copies
)
2282 } while (!success
&& sl
!= r10_bio
->read_slot
);
2286 /* Cannot read from anywhere, just mark the block
2287 * as bad on the first device to discourage future
2290 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2291 rdev
= conf
->mirrors
[dn
].rdev
;
2293 if (!rdev_set_badblocks(
2295 r10_bio
->devs
[r10_bio
->read_slot
].addr
2298 md_error(mddev
, rdev
);
2299 r10_bio
->devs
[r10_bio
->read_slot
].bio
2306 /* write it back and re-read */
2308 while (sl
!= r10_bio
->read_slot
) {
2309 char b
[BDEVNAME_SIZE
];
2314 d
= r10_bio
->devs
[sl
].devnum
;
2315 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2317 test_bit(Unmerged
, &rdev
->flags
) ||
2318 !test_bit(In_sync
, &rdev
->flags
))
2321 atomic_inc(&rdev
->nr_pending
);
2323 if (r10_sync_page_io(rdev
,
2324 r10_bio
->devs
[sl
].addr
+
2326 s
, conf
->tmppage
, WRITE
)
2328 /* Well, this device is dead */
2330 "md/raid10:%s: read correction "
2332 " (%d sectors at %llu on %s)\n",
2334 (unsigned long long)(
2336 choose_data_offset(r10_bio
,
2338 bdevname(rdev
->bdev
, b
));
2339 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2342 bdevname(rdev
->bdev
, b
));
2344 rdev_dec_pending(rdev
, mddev
);
2348 while (sl
!= r10_bio
->read_slot
) {
2349 char b
[BDEVNAME_SIZE
];
2354 d
= r10_bio
->devs
[sl
].devnum
;
2355 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2357 !test_bit(In_sync
, &rdev
->flags
))
2360 atomic_inc(&rdev
->nr_pending
);
2362 switch (r10_sync_page_io(rdev
,
2363 r10_bio
->devs
[sl
].addr
+
2368 /* Well, this device is dead */
2370 "md/raid10:%s: unable to read back "
2372 " (%d sectors at %llu on %s)\n",
2374 (unsigned long long)(
2376 choose_data_offset(r10_bio
, rdev
)),
2377 bdevname(rdev
->bdev
, b
));
2378 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2381 bdevname(rdev
->bdev
, b
));
2385 "md/raid10:%s: read error corrected"
2386 " (%d sectors at %llu on %s)\n",
2388 (unsigned long long)(
2390 choose_data_offset(r10_bio
, rdev
)),
2391 bdevname(rdev
->bdev
, b
));
2392 atomic_add(s
, &rdev
->corrected_errors
);
2395 rdev_dec_pending(rdev
, mddev
);
2405 static void bi_complete(struct bio
*bio
, int error
)
2407 complete((struct completion
*)bio
->bi_private
);
2410 static int submit_bio_wait(int rw
, struct bio
*bio
)
2412 struct completion event
;
2415 init_completion(&event
);
2416 bio
->bi_private
= &event
;
2417 bio
->bi_end_io
= bi_complete
;
2418 submit_bio(rw
, bio
);
2419 wait_for_completion(&event
);
2421 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2424 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2426 struct bio
*bio
= r10_bio
->master_bio
;
2427 struct mddev
*mddev
= r10_bio
->mddev
;
2428 struct r10conf
*conf
= mddev
->private;
2429 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2430 /* bio has the data to be written to slot 'i' where
2431 * we just recently had a write error.
2432 * We repeatedly clone the bio and trim down to one block,
2433 * then try the write. Where the write fails we record
2435 * It is conceivable that the bio doesn't exactly align with
2436 * blocks. We must handle this.
2438 * We currently own a reference to the rdev.
2444 int sect_to_write
= r10_bio
->sectors
;
2447 if (rdev
->badblocks
.shift
< 0)
2450 block_sectors
= 1 << rdev
->badblocks
.shift
;
2451 sector
= r10_bio
->sector
;
2452 sectors
= ((r10_bio
->sector
+ block_sectors
)
2453 & ~(sector_t
)(block_sectors
- 1))
2456 while (sect_to_write
) {
2458 if (sectors
> sect_to_write
)
2459 sectors
= sect_to_write
;
2460 /* Write at 'sector' for 'sectors' */
2461 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2462 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2463 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2464 choose_data_offset(r10_bio
, rdev
) +
2465 (sector
- r10_bio
->sector
));
2466 wbio
->bi_bdev
= rdev
->bdev
;
2467 if (submit_bio_wait(WRITE
, wbio
) == 0)
2469 ok
= rdev_set_badblocks(rdev
, sector
,
2474 sect_to_write
-= sectors
;
2476 sectors
= block_sectors
;
2481 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2483 int slot
= r10_bio
->read_slot
;
2485 struct r10conf
*conf
= mddev
->private;
2486 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2487 char b
[BDEVNAME_SIZE
];
2488 unsigned long do_sync
;
2491 /* we got a read error. Maybe the drive is bad. Maybe just
2492 * the block and we can fix it.
2493 * We freeze all other IO, and try reading the block from
2494 * other devices. When we find one, we re-write
2495 * and check it that fixes the read error.
2496 * This is all done synchronously while the array is
2499 bio
= r10_bio
->devs
[slot
].bio
;
2500 bdevname(bio
->bi_bdev
, b
);
2502 r10_bio
->devs
[slot
].bio
= NULL
;
2504 if (mddev
->ro
== 0) {
2506 fix_read_error(conf
, mddev
, r10_bio
);
2507 unfreeze_array(conf
);
2509 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2511 rdev_dec_pending(rdev
, mddev
);
2514 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2516 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2517 " read error for block %llu\n",
2519 (unsigned long long)r10_bio
->sector
);
2520 raid_end_bio_io(r10_bio
);
2524 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2525 slot
= r10_bio
->read_slot
;
2528 "md/raid10:%s: %s: redirecting "
2529 "sector %llu to another mirror\n",
2531 bdevname(rdev
->bdev
, b
),
2532 (unsigned long long)r10_bio
->sector
);
2533 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2536 r10_bio
->sector
- bio
->bi_sector
,
2538 r10_bio
->devs
[slot
].bio
= bio
;
2539 r10_bio
->devs
[slot
].rdev
= rdev
;
2540 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2541 + choose_data_offset(r10_bio
, rdev
);
2542 bio
->bi_bdev
= rdev
->bdev
;
2543 bio
->bi_rw
= READ
| do_sync
;
2544 bio
->bi_private
= r10_bio
;
2545 bio
->bi_end_io
= raid10_end_read_request
;
2546 if (max_sectors
< r10_bio
->sectors
) {
2547 /* Drat - have to split this up more */
2548 struct bio
*mbio
= r10_bio
->master_bio
;
2549 int sectors_handled
=
2550 r10_bio
->sector
+ max_sectors
2552 r10_bio
->sectors
= max_sectors
;
2553 spin_lock_irq(&conf
->device_lock
);
2554 if (mbio
->bi_phys_segments
== 0)
2555 mbio
->bi_phys_segments
= 2;
2557 mbio
->bi_phys_segments
++;
2558 spin_unlock_irq(&conf
->device_lock
);
2559 generic_make_request(bio
);
2561 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2563 r10_bio
->master_bio
= mbio
;
2564 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2567 set_bit(R10BIO_ReadError
,
2569 r10_bio
->mddev
= mddev
;
2570 r10_bio
->sector
= mbio
->bi_sector
2575 generic_make_request(bio
);
2578 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2580 /* Some sort of write request has finished and it
2581 * succeeded in writing where we thought there was a
2582 * bad block. So forget the bad block.
2583 * Or possibly if failed and we need to record
2587 struct md_rdev
*rdev
;
2589 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2590 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2591 for (m
= 0; m
< conf
->copies
; m
++) {
2592 int dev
= r10_bio
->devs
[m
].devnum
;
2593 rdev
= conf
->mirrors
[dev
].rdev
;
2594 if (r10_bio
->devs
[m
].bio
== NULL
)
2596 if (test_bit(BIO_UPTODATE
,
2597 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2598 rdev_clear_badblocks(
2600 r10_bio
->devs
[m
].addr
,
2601 r10_bio
->sectors
, 0);
2603 if (!rdev_set_badblocks(
2605 r10_bio
->devs
[m
].addr
,
2606 r10_bio
->sectors
, 0))
2607 md_error(conf
->mddev
, rdev
);
2609 rdev
= conf
->mirrors
[dev
].replacement
;
2610 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2612 if (test_bit(BIO_UPTODATE
,
2613 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2614 rdev_clear_badblocks(
2616 r10_bio
->devs
[m
].addr
,
2617 r10_bio
->sectors
, 0);
2619 if (!rdev_set_badblocks(
2621 r10_bio
->devs
[m
].addr
,
2622 r10_bio
->sectors
, 0))
2623 md_error(conf
->mddev
, rdev
);
2628 for (m
= 0; m
< conf
->copies
; m
++) {
2629 int dev
= r10_bio
->devs
[m
].devnum
;
2630 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2631 rdev
= conf
->mirrors
[dev
].rdev
;
2632 if (bio
== IO_MADE_GOOD
) {
2633 rdev_clear_badblocks(
2635 r10_bio
->devs
[m
].addr
,
2636 r10_bio
->sectors
, 0);
2637 rdev_dec_pending(rdev
, conf
->mddev
);
2638 } else if (bio
!= NULL
&&
2639 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2640 if (!narrow_write_error(r10_bio
, m
)) {
2641 md_error(conf
->mddev
, rdev
);
2642 set_bit(R10BIO_Degraded
,
2645 rdev_dec_pending(rdev
, conf
->mddev
);
2647 bio
= r10_bio
->devs
[m
].repl_bio
;
2648 rdev
= conf
->mirrors
[dev
].replacement
;
2649 if (rdev
&& bio
== IO_MADE_GOOD
) {
2650 rdev_clear_badblocks(
2652 r10_bio
->devs
[m
].addr
,
2653 r10_bio
->sectors
, 0);
2654 rdev_dec_pending(rdev
, conf
->mddev
);
2657 if (test_bit(R10BIO_WriteError
,
2659 close_write(r10_bio
);
2660 raid_end_bio_io(r10_bio
);
2664 static void raid10d(struct mddev
*mddev
)
2666 struct r10bio
*r10_bio
;
2667 unsigned long flags
;
2668 struct r10conf
*conf
= mddev
->private;
2669 struct list_head
*head
= &conf
->retry_list
;
2670 struct blk_plug plug
;
2672 md_check_recovery(mddev
);
2674 blk_start_plug(&plug
);
2677 if (atomic_read(&mddev
->plug_cnt
) == 0)
2678 flush_pending_writes(conf
);
2680 spin_lock_irqsave(&conf
->device_lock
, flags
);
2681 if (list_empty(head
)) {
2682 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2685 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2686 list_del(head
->prev
);
2688 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2690 mddev
= r10_bio
->mddev
;
2691 conf
= mddev
->private;
2692 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2693 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2694 handle_write_completed(conf
, r10_bio
);
2695 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2696 reshape_request_write(mddev
, r10_bio
);
2697 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2698 sync_request_write(mddev
, r10_bio
);
2699 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2700 recovery_request_write(mddev
, r10_bio
);
2701 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2702 handle_read_error(mddev
, r10_bio
);
2704 /* just a partial read to be scheduled from a
2707 int slot
= r10_bio
->read_slot
;
2708 generic_make_request(r10_bio
->devs
[slot
].bio
);
2712 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2713 md_check_recovery(mddev
);
2715 blk_finish_plug(&plug
);
2719 static int init_resync(struct r10conf
*conf
)
2724 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2725 BUG_ON(conf
->r10buf_pool
);
2726 conf
->have_replacement
= 0;
2727 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2728 if (conf
->mirrors
[i
].replacement
)
2729 conf
->have_replacement
= 1;
2730 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2731 if (!conf
->r10buf_pool
)
2733 conf
->next_resync
= 0;
2738 * perform a "sync" on one "block"
2740 * We need to make sure that no normal I/O request - particularly write
2741 * requests - conflict with active sync requests.
2743 * This is achieved by tracking pending requests and a 'barrier' concept
2744 * that can be installed to exclude normal IO requests.
2746 * Resync and recovery are handled very differently.
2747 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2749 * For resync, we iterate over virtual addresses, read all copies,
2750 * and update if there are differences. If only one copy is live,
2752 * For recovery, we iterate over physical addresses, read a good
2753 * value for each non-in_sync drive, and over-write.
2755 * So, for recovery we may have several outstanding complex requests for a
2756 * given address, one for each out-of-sync device. We model this by allocating
2757 * a number of r10_bio structures, one for each out-of-sync device.
2758 * As we setup these structures, we collect all bio's together into a list
2759 * which we then process collectively to add pages, and then process again
2760 * to pass to generic_make_request.
2762 * The r10_bio structures are linked using a borrowed master_bio pointer.
2763 * This link is counted in ->remaining. When the r10_bio that points to NULL
2764 * has its remaining count decremented to 0, the whole complex operation
2769 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2770 int *skipped
, int go_faster
)
2772 struct r10conf
*conf
= mddev
->private;
2773 struct r10bio
*r10_bio
;
2774 struct bio
*biolist
= NULL
, *bio
;
2775 sector_t max_sector
, nr_sectors
;
2778 sector_t sync_blocks
;
2779 sector_t sectors_skipped
= 0;
2780 int chunks_skipped
= 0;
2781 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2783 if (!conf
->r10buf_pool
)
2784 if (init_resync(conf
))
2788 max_sector
= mddev
->dev_sectors
;
2789 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2790 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2791 max_sector
= mddev
->resync_max_sectors
;
2792 if (sector_nr
>= max_sector
) {
2793 /* If we aborted, we need to abort the
2794 * sync on the 'current' bitmap chucks (there can
2795 * be several when recovering multiple devices).
2796 * as we may have started syncing it but not finished.
2797 * We can find the current address in
2798 * mddev->curr_resync, but for recovery,
2799 * we need to convert that to several
2800 * virtual addresses.
2802 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2807 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2808 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2809 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2811 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2813 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2814 bitmap_end_sync(mddev
->bitmap
, sect
,
2818 /* completed sync */
2819 if ((!mddev
->bitmap
|| conf
->fullsync
)
2820 && conf
->have_replacement
2821 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2822 /* Completed a full sync so the replacements
2823 * are now fully recovered.
2825 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2826 if (conf
->mirrors
[i
].replacement
)
2827 conf
->mirrors
[i
].replacement
2833 bitmap_close_sync(mddev
->bitmap
);
2836 return sectors_skipped
;
2839 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2840 return reshape_request(mddev
, sector_nr
, skipped
);
2842 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2843 /* if there has been nothing to do on any drive,
2844 * then there is nothing to do at all..
2847 return (max_sector
- sector_nr
) + sectors_skipped
;
2850 if (max_sector
> mddev
->resync_max
)
2851 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2853 /* make sure whole request will fit in a chunk - if chunks
2856 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2857 max_sector
> (sector_nr
| chunk_mask
))
2858 max_sector
= (sector_nr
| chunk_mask
) + 1;
2860 * If there is non-resync activity waiting for us then
2861 * put in a delay to throttle resync.
2863 if (!go_faster
&& conf
->nr_waiting
)
2864 msleep_interruptible(1000);
2866 /* Again, very different code for resync and recovery.
2867 * Both must result in an r10bio with a list of bios that
2868 * have bi_end_io, bi_sector, bi_bdev set,
2869 * and bi_private set to the r10bio.
2870 * For recovery, we may actually create several r10bios
2871 * with 2 bios in each, that correspond to the bios in the main one.
2872 * In this case, the subordinate r10bios link back through a
2873 * borrowed master_bio pointer, and the counter in the master
2874 * includes a ref from each subordinate.
2876 /* First, we decide what to do and set ->bi_end_io
2877 * To end_sync_read if we want to read, and
2878 * end_sync_write if we will want to write.
2881 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2882 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2883 /* recovery... the complicated one */
2887 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2893 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2895 if ((mirror
->rdev
== NULL
||
2896 test_bit(In_sync
, &mirror
->rdev
->flags
))
2898 (mirror
->replacement
== NULL
||
2900 &mirror
->replacement
->flags
)))
2904 /* want to reconstruct this device */
2906 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2907 if (sect
>= mddev
->resync_max_sectors
) {
2908 /* last stripe is not complete - don't
2909 * try to recover this sector.
2913 /* Unless we are doing a full sync, or a replacement
2914 * we only need to recover the block if it is set in
2917 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2919 if (sync_blocks
< max_sync
)
2920 max_sync
= sync_blocks
;
2922 mirror
->replacement
== NULL
&&
2924 /* yep, skip the sync_blocks here, but don't assume
2925 * that there will never be anything to do here
2927 chunks_skipped
= -1;
2931 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2932 raise_barrier(conf
, rb2
!= NULL
);
2933 atomic_set(&r10_bio
->remaining
, 0);
2935 r10_bio
->master_bio
= (struct bio
*)rb2
;
2937 atomic_inc(&rb2
->remaining
);
2938 r10_bio
->mddev
= mddev
;
2939 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
2940 r10_bio
->sector
= sect
;
2942 raid10_find_phys(conf
, r10_bio
);
2944 /* Need to check if the array will still be
2947 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
2948 if (conf
->mirrors
[j
].rdev
== NULL
||
2949 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
2954 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2955 &sync_blocks
, still_degraded
);
2958 for (j
=0; j
<conf
->copies
;j
++) {
2960 int d
= r10_bio
->devs
[j
].devnum
;
2961 sector_t from_addr
, to_addr
;
2962 struct md_rdev
*rdev
;
2963 sector_t sector
, first_bad
;
2965 if (!conf
->mirrors
[d
].rdev
||
2966 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
2968 /* This is where we read from */
2970 rdev
= conf
->mirrors
[d
].rdev
;
2971 sector
= r10_bio
->devs
[j
].addr
;
2973 if (is_badblock(rdev
, sector
, max_sync
,
2974 &first_bad
, &bad_sectors
)) {
2975 if (first_bad
> sector
)
2976 max_sync
= first_bad
- sector
;
2978 bad_sectors
-= (sector
2980 if (max_sync
> bad_sectors
)
2981 max_sync
= bad_sectors
;
2985 bio
= r10_bio
->devs
[0].bio
;
2986 bio
->bi_next
= biolist
;
2988 bio
->bi_private
= r10_bio
;
2989 bio
->bi_end_io
= end_sync_read
;
2991 from_addr
= r10_bio
->devs
[j
].addr
;
2992 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
2993 bio
->bi_bdev
= rdev
->bdev
;
2994 atomic_inc(&rdev
->nr_pending
);
2995 /* and we write to 'i' (if not in_sync) */
2997 for (k
=0; k
<conf
->copies
; k
++)
2998 if (r10_bio
->devs
[k
].devnum
== i
)
3000 BUG_ON(k
== conf
->copies
);
3001 to_addr
= r10_bio
->devs
[k
].addr
;
3002 r10_bio
->devs
[0].devnum
= d
;
3003 r10_bio
->devs
[0].addr
= from_addr
;
3004 r10_bio
->devs
[1].devnum
= i
;
3005 r10_bio
->devs
[1].addr
= to_addr
;
3007 rdev
= mirror
->rdev
;
3008 if (!test_bit(In_sync
, &rdev
->flags
)) {
3009 bio
= r10_bio
->devs
[1].bio
;
3010 bio
->bi_next
= biolist
;
3012 bio
->bi_private
= r10_bio
;
3013 bio
->bi_end_io
= end_sync_write
;
3015 bio
->bi_sector
= to_addr
3016 + rdev
->data_offset
;
3017 bio
->bi_bdev
= rdev
->bdev
;
3018 atomic_inc(&r10_bio
->remaining
);
3020 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3022 /* and maybe write to replacement */
3023 bio
= r10_bio
->devs
[1].repl_bio
;
3025 bio
->bi_end_io
= NULL
;
3026 rdev
= mirror
->replacement
;
3027 /* Note: if rdev != NULL, then bio
3028 * cannot be NULL as r10buf_pool_alloc will
3029 * have allocated it.
3030 * So the second test here is pointless.
3031 * But it keeps semantic-checkers happy, and
3032 * this comment keeps human reviewers
3035 if (rdev
== NULL
|| bio
== NULL
||
3036 test_bit(Faulty
, &rdev
->flags
))
3038 bio
->bi_next
= biolist
;
3040 bio
->bi_private
= r10_bio
;
3041 bio
->bi_end_io
= end_sync_write
;
3043 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3044 bio
->bi_bdev
= rdev
->bdev
;
3045 atomic_inc(&r10_bio
->remaining
);
3048 if (j
== conf
->copies
) {
3049 /* Cannot recover, so abort the recovery or
3050 * record a bad block */
3053 atomic_dec(&rb2
->remaining
);
3056 /* problem is that there are bad blocks
3057 * on other device(s)
3060 for (k
= 0; k
< conf
->copies
; k
++)
3061 if (r10_bio
->devs
[k
].devnum
== i
)
3063 if (!test_bit(In_sync
,
3064 &mirror
->rdev
->flags
)
3065 && !rdev_set_badblocks(
3067 r10_bio
->devs
[k
].addr
,
3070 if (mirror
->replacement
&&
3071 !rdev_set_badblocks(
3072 mirror
->replacement
,
3073 r10_bio
->devs
[k
].addr
,
3078 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3080 printk(KERN_INFO
"md/raid10:%s: insufficient "
3081 "working devices for recovery.\n",
3083 mirror
->recovery_disabled
3084 = mddev
->recovery_disabled
;
3089 if (biolist
== NULL
) {
3091 struct r10bio
*rb2
= r10_bio
;
3092 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3093 rb2
->master_bio
= NULL
;
3099 /* resync. Schedule a read for every block at this virt offset */
3102 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3104 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3105 &sync_blocks
, mddev
->degraded
) &&
3106 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3107 &mddev
->recovery
)) {
3108 /* We can skip this block */
3110 return sync_blocks
+ sectors_skipped
;
3112 if (sync_blocks
< max_sync
)
3113 max_sync
= sync_blocks
;
3114 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3116 r10_bio
->mddev
= mddev
;
3117 atomic_set(&r10_bio
->remaining
, 0);
3118 raise_barrier(conf
, 0);
3119 conf
->next_resync
= sector_nr
;
3121 r10_bio
->master_bio
= NULL
;
3122 r10_bio
->sector
= sector_nr
;
3123 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3124 raid10_find_phys(conf
, r10_bio
);
3125 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3127 for (i
= 0; i
< conf
->copies
; i
++) {
3128 int d
= r10_bio
->devs
[i
].devnum
;
3129 sector_t first_bad
, sector
;
3132 if (r10_bio
->devs
[i
].repl_bio
)
3133 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3135 bio
= r10_bio
->devs
[i
].bio
;
3136 bio
->bi_end_io
= NULL
;
3137 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3138 if (conf
->mirrors
[d
].rdev
== NULL
||
3139 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3141 sector
= r10_bio
->devs
[i
].addr
;
3142 if (is_badblock(conf
->mirrors
[d
].rdev
,
3144 &first_bad
, &bad_sectors
)) {
3145 if (first_bad
> sector
)
3146 max_sync
= first_bad
- sector
;
3148 bad_sectors
-= (sector
- first_bad
);
3149 if (max_sync
> bad_sectors
)
3150 max_sync
= max_sync
;
3154 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3155 atomic_inc(&r10_bio
->remaining
);
3156 bio
->bi_next
= biolist
;
3158 bio
->bi_private
= r10_bio
;
3159 bio
->bi_end_io
= end_sync_read
;
3161 bio
->bi_sector
= sector
+
3162 conf
->mirrors
[d
].rdev
->data_offset
;
3163 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3166 if (conf
->mirrors
[d
].replacement
== NULL
||
3168 &conf
->mirrors
[d
].replacement
->flags
))
3171 /* Need to set up for writing to the replacement */
3172 bio
= r10_bio
->devs
[i
].repl_bio
;
3173 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3175 sector
= r10_bio
->devs
[i
].addr
;
3176 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3177 bio
->bi_next
= biolist
;
3179 bio
->bi_private
= r10_bio
;
3180 bio
->bi_end_io
= end_sync_write
;
3182 bio
->bi_sector
= sector
+
3183 conf
->mirrors
[d
].replacement
->data_offset
;
3184 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3189 for (i
=0; i
<conf
->copies
; i
++) {
3190 int d
= r10_bio
->devs
[i
].devnum
;
3191 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3192 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3194 if (r10_bio
->devs
[i
].repl_bio
&&
3195 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3197 conf
->mirrors
[d
].replacement
,
3206 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3208 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
3210 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
3213 bio
->bi_phys_segments
= 0;
3218 if (sector_nr
+ max_sync
< max_sector
)
3219 max_sector
= sector_nr
+ max_sync
;
3222 int len
= PAGE_SIZE
;
3223 if (sector_nr
+ (len
>>9) > max_sector
)
3224 len
= (max_sector
- sector_nr
) << 9;
3227 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3229 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3230 if (bio_add_page(bio
, page
, len
, 0))
3234 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3235 for (bio2
= biolist
;
3236 bio2
&& bio2
!= bio
;
3237 bio2
= bio2
->bi_next
) {
3238 /* remove last page from this bio */
3240 bio2
->bi_size
-= len
;
3241 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3245 nr_sectors
+= len
>>9;
3246 sector_nr
+= len
>>9;
3247 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3249 r10_bio
->sectors
= nr_sectors
;
3253 biolist
= biolist
->bi_next
;
3255 bio
->bi_next
= NULL
;
3256 r10_bio
= bio
->bi_private
;
3257 r10_bio
->sectors
= nr_sectors
;
3259 if (bio
->bi_end_io
== end_sync_read
) {
3260 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3261 generic_make_request(bio
);
3265 if (sectors_skipped
)
3266 /* pretend they weren't skipped, it makes
3267 * no important difference in this case
3269 md_done_sync(mddev
, sectors_skipped
, 1);
3271 return sectors_skipped
+ nr_sectors
;
3273 /* There is nowhere to write, so all non-sync
3274 * drives must be failed or in resync, all drives
3275 * have a bad block, so try the next chunk...
3277 if (sector_nr
+ max_sync
< max_sector
)
3278 max_sector
= sector_nr
+ max_sync
;
3280 sectors_skipped
+= (max_sector
- sector_nr
);
3282 sector_nr
= max_sector
;
3287 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3290 struct r10conf
*conf
= mddev
->private;
3293 raid_disks
= min(conf
->geo
.raid_disks
,
3294 conf
->prev
.raid_disks
);
3296 sectors
= conf
->dev_sectors
;
3298 size
= sectors
>> conf
->geo
.chunk_shift
;
3299 sector_div(size
, conf
->geo
.far_copies
);
3300 size
= size
* raid_disks
;
3301 sector_div(size
, conf
->geo
.near_copies
);
3303 return size
<< conf
->geo
.chunk_shift
;
3306 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3308 /* Calculate the number of sectors-per-device that will
3309 * actually be used, and set conf->dev_sectors and
3313 size
= size
>> conf
->geo
.chunk_shift
;
3314 sector_div(size
, conf
->geo
.far_copies
);
3315 size
= size
* conf
->geo
.raid_disks
;
3316 sector_div(size
, conf
->geo
.near_copies
);
3317 /* 'size' is now the number of chunks in the array */
3318 /* calculate "used chunks per device" */
3319 size
= size
* conf
->copies
;
3321 /* We need to round up when dividing by raid_disks to
3322 * get the stride size.
3324 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3326 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3328 if (conf
->geo
.far_offset
)
3329 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3331 sector_div(size
, conf
->geo
.far_copies
);
3332 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3336 enum geo_type
{geo_new
, geo_old
, geo_start
};
3337 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3340 int layout
, chunk
, disks
;
3343 layout
= mddev
->layout
;
3344 chunk
= mddev
->chunk_sectors
;
3345 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3348 layout
= mddev
->new_layout
;
3349 chunk
= mddev
->new_chunk_sectors
;
3350 disks
= mddev
->raid_disks
;
3352 default: /* avoid 'may be unused' warnings */
3353 case geo_start
: /* new when starting reshape - raid_disks not
3355 layout
= mddev
->new_layout
;
3356 chunk
= mddev
->new_chunk_sectors
;
3357 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3362 if (chunk
< (PAGE_SIZE
>> 9) ||
3363 !is_power_of_2(chunk
))
3366 fc
= (layout
>> 8) & 255;
3367 fo
= layout
& (1<<16);
3368 geo
->raid_disks
= disks
;
3369 geo
->near_copies
= nc
;
3370 geo
->far_copies
= fc
;
3371 geo
->far_offset
= fo
;
3372 geo
->chunk_mask
= chunk
- 1;
3373 geo
->chunk_shift
= ffz(~chunk
);
3377 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3379 struct r10conf
*conf
= NULL
;
3384 copies
= setup_geo(&geo
, mddev
, geo_new
);
3387 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3388 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3389 mdname(mddev
), PAGE_SIZE
);
3393 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3394 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3395 mdname(mddev
), mddev
->new_layout
);
3400 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3404 /* FIXME calc properly */
3405 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3406 max(0,mddev
->delta_disks
)),
3411 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3416 conf
->copies
= copies
;
3417 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3418 r10bio_pool_free
, conf
);
3419 if (!conf
->r10bio_pool
)
3422 calc_sectors(conf
, mddev
->dev_sectors
);
3423 if (mddev
->reshape_position
== MaxSector
) {
3424 conf
->prev
= conf
->geo
;
3425 conf
->reshape_progress
= MaxSector
;
3427 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3431 conf
->reshape_progress
= mddev
->reshape_position
;
3432 if (conf
->prev
.far_offset
)
3433 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3435 /* far_copies must be 1 */
3436 conf
->prev
.stride
= conf
->dev_sectors
;
3438 spin_lock_init(&conf
->device_lock
);
3439 INIT_LIST_HEAD(&conf
->retry_list
);
3441 spin_lock_init(&conf
->resync_lock
);
3442 init_waitqueue_head(&conf
->wait_barrier
);
3444 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3448 conf
->mddev
= mddev
;
3453 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3456 if (conf
->r10bio_pool
)
3457 mempool_destroy(conf
->r10bio_pool
);
3458 kfree(conf
->mirrors
);
3459 safe_put_page(conf
->tmppage
);
3462 return ERR_PTR(err
);
3465 static int run(struct mddev
*mddev
)
3467 struct r10conf
*conf
;
3468 int i
, disk_idx
, chunk_size
;
3469 struct raid10_info
*disk
;
3470 struct md_rdev
*rdev
;
3472 sector_t min_offset_diff
= 0;
3475 if (mddev
->private == NULL
) {
3476 conf
= setup_conf(mddev
);
3478 return PTR_ERR(conf
);
3479 mddev
->private = conf
;
3481 conf
= mddev
->private;
3485 mddev
->thread
= conf
->thread
;
3486 conf
->thread
= NULL
;
3488 chunk_size
= mddev
->chunk_sectors
<< 9;
3489 blk_queue_io_min(mddev
->queue
, chunk_size
);
3490 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3491 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3493 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3494 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3496 rdev_for_each(rdev
, mddev
) {
3498 struct request_queue
*q
;
3500 disk_idx
= rdev
->raid_disk
;
3503 if (disk_idx
>= conf
->geo
.raid_disks
&&
3504 disk_idx
>= conf
->prev
.raid_disks
)
3506 disk
= conf
->mirrors
+ disk_idx
;
3508 if (test_bit(Replacement
, &rdev
->flags
)) {
3509 if (disk
->replacement
)
3511 disk
->replacement
= rdev
;
3517 q
= bdev_get_queue(rdev
->bdev
);
3518 if (q
->merge_bvec_fn
)
3519 mddev
->merge_check_needed
= 1;
3520 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3521 if (!mddev
->reshape_backwards
)
3525 if (first
|| diff
< min_offset_diff
)
3526 min_offset_diff
= diff
;
3528 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3529 rdev
->data_offset
<< 9);
3531 disk
->head_position
= 0;
3534 /* need to check that every block has at least one working mirror */
3535 if (!enough(conf
, -1)) {
3536 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3541 if (conf
->reshape_progress
!= MaxSector
) {
3542 /* must ensure that shape change is supported */
3543 if (conf
->geo
.far_copies
!= 1 &&
3544 conf
->geo
.far_offset
== 0)
3546 if (conf
->prev
.far_copies
!= 1 &&
3547 conf
->geo
.far_offset
== 0)
3551 mddev
->degraded
= 0;
3553 i
< conf
->geo
.raid_disks
3554 || i
< conf
->prev
.raid_disks
;
3557 disk
= conf
->mirrors
+ i
;
3559 if (!disk
->rdev
&& disk
->replacement
) {
3560 /* The replacement is all we have - use it */
3561 disk
->rdev
= disk
->replacement
;
3562 disk
->replacement
= NULL
;
3563 clear_bit(Replacement
, &disk
->rdev
->flags
);
3567 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3568 disk
->head_position
= 0;
3573 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3576 if (mddev
->recovery_cp
!= MaxSector
)
3577 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3578 " -- starting background reconstruction\n",
3581 "md/raid10:%s: active with %d out of %d devices\n",
3582 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3583 conf
->geo
.raid_disks
);
3585 * Ok, everything is just fine now
3587 mddev
->dev_sectors
= conf
->dev_sectors
;
3588 size
= raid10_size(mddev
, 0, 0);
3589 md_set_array_sectors(mddev
, size
);
3590 mddev
->resync_max_sectors
= size
;
3592 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3593 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3595 /* Calculate max read-ahead size.
3596 * We need to readahead at least twice a whole stripe....
3600 int stripe
= conf
->geo
.raid_disks
*
3601 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3602 stripe
/= conf
->geo
.near_copies
;
3603 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3604 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3607 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3609 if (md_integrity_register(mddev
))
3612 if (conf
->reshape_progress
!= MaxSector
) {
3613 unsigned long before_length
, after_length
;
3615 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3616 conf
->prev
.far_copies
);
3617 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3618 conf
->geo
.far_copies
);
3620 if (max(before_length
, after_length
) > min_offset_diff
) {
3621 /* This cannot work */
3622 printk("md/raid10: offset difference not enough to continue reshape\n");
3625 conf
->offset_diff
= min_offset_diff
;
3627 conf
->reshape_safe
= conf
->reshape_progress
;
3628 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3629 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3630 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3631 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3632 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3639 md_unregister_thread(&mddev
->thread
);
3640 if (conf
->r10bio_pool
)
3641 mempool_destroy(conf
->r10bio_pool
);
3642 safe_put_page(conf
->tmppage
);
3643 kfree(conf
->mirrors
);
3645 mddev
->private = NULL
;
3650 static int stop(struct mddev
*mddev
)
3652 struct r10conf
*conf
= mddev
->private;
3654 raise_barrier(conf
, 0);
3655 lower_barrier(conf
);
3657 md_unregister_thread(&mddev
->thread
);
3658 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
3659 if (conf
->r10bio_pool
)
3660 mempool_destroy(conf
->r10bio_pool
);
3661 kfree(conf
->mirrors
);
3663 mddev
->private = NULL
;
3667 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3669 struct r10conf
*conf
= mddev
->private;
3673 raise_barrier(conf
, 0);
3676 lower_barrier(conf
);
3681 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3683 /* Resize of 'far' arrays is not supported.
3684 * For 'near' and 'offset' arrays we can set the
3685 * number of sectors used to be an appropriate multiple
3686 * of the chunk size.
3687 * For 'offset', this is far_copies*chunksize.
3688 * For 'near' the multiplier is the LCM of
3689 * near_copies and raid_disks.
3690 * So if far_copies > 1 && !far_offset, fail.
3691 * Else find LCM(raid_disks, near_copy)*far_copies and
3692 * multiply by chunk_size. Then round to this number.
3693 * This is mostly done by raid10_size()
3695 struct r10conf
*conf
= mddev
->private;
3696 sector_t oldsize
, size
;
3698 if (mddev
->reshape_position
!= MaxSector
)
3701 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3704 oldsize
= raid10_size(mddev
, 0, 0);
3705 size
= raid10_size(mddev
, sectors
, 0);
3706 if (mddev
->external_size
&&
3707 mddev
->array_sectors
> size
)
3709 if (mddev
->bitmap
) {
3710 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3714 md_set_array_sectors(mddev
, size
);
3715 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3716 revalidate_disk(mddev
->gendisk
);
3717 if (sectors
> mddev
->dev_sectors
&&
3718 mddev
->recovery_cp
> oldsize
) {
3719 mddev
->recovery_cp
= oldsize
;
3720 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3722 calc_sectors(conf
, sectors
);
3723 mddev
->dev_sectors
= conf
->dev_sectors
;
3724 mddev
->resync_max_sectors
= size
;
3728 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3730 struct md_rdev
*rdev
;
3731 struct r10conf
*conf
;
3733 if (mddev
->degraded
> 0) {
3734 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3736 return ERR_PTR(-EINVAL
);
3739 /* Set new parameters */
3740 mddev
->new_level
= 10;
3741 /* new layout: far_copies = 1, near_copies = 2 */
3742 mddev
->new_layout
= (1<<8) + 2;
3743 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3744 mddev
->delta_disks
= mddev
->raid_disks
;
3745 mddev
->raid_disks
*= 2;
3746 /* make sure it will be not marked as dirty */
3747 mddev
->recovery_cp
= MaxSector
;
3749 conf
= setup_conf(mddev
);
3750 if (!IS_ERR(conf
)) {
3751 rdev_for_each(rdev
, mddev
)
3752 if (rdev
->raid_disk
>= 0)
3753 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3760 static void *raid10_takeover(struct mddev
*mddev
)
3762 struct r0conf
*raid0_conf
;
3764 /* raid10 can take over:
3765 * raid0 - providing it has only two drives
3767 if (mddev
->level
== 0) {
3768 /* for raid0 takeover only one zone is supported */
3769 raid0_conf
= mddev
->private;
3770 if (raid0_conf
->nr_strip_zones
> 1) {
3771 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3772 " with more than one zone.\n",
3774 return ERR_PTR(-EINVAL
);
3776 return raid10_takeover_raid0(mddev
);
3778 return ERR_PTR(-EINVAL
);
3781 static int raid10_check_reshape(struct mddev
*mddev
)
3783 /* Called when there is a request to change
3784 * - layout (to ->new_layout)
3785 * - chunk size (to ->new_chunk_sectors)
3786 * - raid_disks (by delta_disks)
3787 * or when trying to restart a reshape that was ongoing.
3789 * We need to validate the request and possibly allocate
3790 * space if that might be an issue later.
3792 * Currently we reject any reshape of a 'far' mode array,
3793 * allow chunk size to change if new is generally acceptable,
3794 * allow raid_disks to increase, and allow
3795 * a switch between 'near' mode and 'offset' mode.
3797 struct r10conf
*conf
= mddev
->private;
3800 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3803 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3804 /* mustn't change number of copies */
3806 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3807 /* Cannot switch to 'far' mode */
3810 if (mddev
->array_sectors
& geo
.chunk_mask
)
3811 /* not factor of array size */
3814 if (!enough(conf
, -1))
3817 kfree(conf
->mirrors_new
);
3818 conf
->mirrors_new
= NULL
;
3819 if (mddev
->delta_disks
> 0) {
3820 /* allocate new 'mirrors' list */
3821 conf
->mirrors_new
= kzalloc(
3822 sizeof(struct raid10_info
)
3823 *(mddev
->raid_disks
+
3824 mddev
->delta_disks
),
3826 if (!conf
->mirrors_new
)
3833 * Need to check if array has failed when deciding whether to:
3835 * - remove non-faulty devices
3838 * This determination is simple when no reshape is happening.
3839 * However if there is a reshape, we need to carefully check
3840 * both the before and after sections.
3841 * This is because some failed devices may only affect one
3842 * of the two sections, and some non-in_sync devices may
3843 * be insync in the section most affected by failed devices.
3845 static int calc_degraded(struct r10conf
*conf
)
3847 int degraded
, degraded2
;
3852 /* 'prev' section first */
3853 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
3854 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3855 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3857 else if (!test_bit(In_sync
, &rdev
->flags
))
3858 /* When we can reduce the number of devices in
3859 * an array, this might not contribute to
3860 * 'degraded'. It does now.
3865 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
3869 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3870 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3871 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3873 else if (!test_bit(In_sync
, &rdev
->flags
)) {
3874 /* If reshape is increasing the number of devices,
3875 * this section has already been recovered, so
3876 * it doesn't contribute to degraded.
3879 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
3884 if (degraded2
> degraded
)
3889 static int raid10_start_reshape(struct mddev
*mddev
)
3891 /* A 'reshape' has been requested. This commits
3892 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3893 * This also checks if there are enough spares and adds them
3895 * We currently require enough spares to make the final
3896 * array non-degraded. We also require that the difference
3897 * between old and new data_offset - on each device - is
3898 * enough that we never risk over-writing.
3901 unsigned long before_length
, after_length
;
3902 sector_t min_offset_diff
= 0;
3905 struct r10conf
*conf
= mddev
->private;
3906 struct md_rdev
*rdev
;
3910 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
3913 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
3916 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3917 conf
->prev
.far_copies
);
3918 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3919 conf
->geo
.far_copies
);
3921 rdev_for_each(rdev
, mddev
) {
3922 if (!test_bit(In_sync
, &rdev
->flags
)
3923 && !test_bit(Faulty
, &rdev
->flags
))
3925 if (rdev
->raid_disk
>= 0) {
3926 long long diff
= (rdev
->new_data_offset
3927 - rdev
->data_offset
);
3928 if (!mddev
->reshape_backwards
)
3932 if (first
|| diff
< min_offset_diff
)
3933 min_offset_diff
= diff
;
3937 if (max(before_length
, after_length
) > min_offset_diff
)
3940 if (spares
< mddev
->delta_disks
)
3943 conf
->offset_diff
= min_offset_diff
;
3944 spin_lock_irq(&conf
->device_lock
);
3945 if (conf
->mirrors_new
) {
3946 memcpy(conf
->mirrors_new
, conf
->mirrors
,
3947 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
3949 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
3950 conf
->mirrors_old
= conf
->mirrors
;
3951 conf
->mirrors
= conf
->mirrors_new
;
3952 conf
->mirrors_new
= NULL
;
3954 setup_geo(&conf
->geo
, mddev
, geo_start
);
3956 if (mddev
->reshape_backwards
) {
3957 sector_t size
= raid10_size(mddev
, 0, 0);
3958 if (size
< mddev
->array_sectors
) {
3959 spin_unlock_irq(&conf
->device_lock
);
3960 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
3964 mddev
->resync_max_sectors
= size
;
3965 conf
->reshape_progress
= size
;
3967 conf
->reshape_progress
= 0;
3968 spin_unlock_irq(&conf
->device_lock
);
3970 if (mddev
->delta_disks
&& mddev
->bitmap
) {
3971 ret
= bitmap_resize(mddev
->bitmap
,
3972 raid10_size(mddev
, 0,
3973 conf
->geo
.raid_disks
),
3978 if (mddev
->delta_disks
> 0) {
3979 rdev_for_each(rdev
, mddev
)
3980 if (rdev
->raid_disk
< 0 &&
3981 !test_bit(Faulty
, &rdev
->flags
)) {
3982 if (raid10_add_disk(mddev
, rdev
) == 0) {
3983 if (rdev
->raid_disk
>=
3984 conf
->prev
.raid_disks
)
3985 set_bit(In_sync
, &rdev
->flags
);
3987 rdev
->recovery_offset
= 0;
3989 if (sysfs_link_rdev(mddev
, rdev
))
3990 /* Failure here is OK */;
3992 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
3993 && !test_bit(Faulty
, &rdev
->flags
)) {
3994 /* This is a spare that was manually added */
3995 set_bit(In_sync
, &rdev
->flags
);
3998 /* When a reshape changes the number of devices,
3999 * ->degraded is measured against the larger of the
4000 * pre and post numbers.
4002 spin_lock_irq(&conf
->device_lock
);
4003 mddev
->degraded
= calc_degraded(conf
);
4004 spin_unlock_irq(&conf
->device_lock
);
4005 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4006 mddev
->reshape_position
= conf
->reshape_progress
;
4007 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4009 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4010 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4011 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4012 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4014 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4016 if (!mddev
->sync_thread
) {
4020 conf
->reshape_checkpoint
= jiffies
;
4021 md_wakeup_thread(mddev
->sync_thread
);
4022 md_new_event(mddev
);
4026 mddev
->recovery
= 0;
4027 spin_lock_irq(&conf
->device_lock
);
4028 conf
->geo
= conf
->prev
;
4029 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4030 rdev_for_each(rdev
, mddev
)
4031 rdev
->new_data_offset
= rdev
->data_offset
;
4033 conf
->reshape_progress
= MaxSector
;
4034 mddev
->reshape_position
= MaxSector
;
4035 spin_unlock_irq(&conf
->device_lock
);
4039 /* Calculate the last device-address that could contain
4040 * any block from the chunk that includes the array-address 's'
4041 * and report the next address.
4042 * i.e. the address returned will be chunk-aligned and after
4043 * any data that is in the chunk containing 's'.
4045 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4047 s
= (s
| geo
->chunk_mask
) + 1;
4048 s
>>= geo
->chunk_shift
;
4049 s
*= geo
->near_copies
;
4050 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4051 s
*= geo
->far_copies
;
4052 s
<<= geo
->chunk_shift
;
4056 /* Calculate the first device-address that could contain
4057 * any block from the chunk that includes the array-address 's'.
4058 * This too will be the start of a chunk
4060 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4062 s
>>= geo
->chunk_shift
;
4063 s
*= geo
->near_copies
;
4064 sector_div(s
, geo
->raid_disks
);
4065 s
*= geo
->far_copies
;
4066 s
<<= geo
->chunk_shift
;
4070 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4073 /* We simply copy at most one chunk (smallest of old and new)
4074 * at a time, possibly less if that exceeds RESYNC_PAGES,
4075 * or we hit a bad block or something.
4076 * This might mean we pause for normal IO in the middle of
4077 * a chunk, but that is not a problem was mddev->reshape_position
4078 * can record any location.
4080 * If we will want to write to a location that isn't
4081 * yet recorded as 'safe' (i.e. in metadata on disk) then
4082 * we need to flush all reshape requests and update the metadata.
4084 * When reshaping forwards (e.g. to more devices), we interpret
4085 * 'safe' as the earliest block which might not have been copied
4086 * down yet. We divide this by previous stripe size and multiply
4087 * by previous stripe length to get lowest device offset that we
4088 * cannot write to yet.
4089 * We interpret 'sector_nr' as an address that we want to write to.
4090 * From this we use last_device_address() to find where we might
4091 * write to, and first_device_address on the 'safe' position.
4092 * If this 'next' write position is after the 'safe' position,
4093 * we must update the metadata to increase the 'safe' position.
4095 * When reshaping backwards, we round in the opposite direction
4096 * and perform the reverse test: next write position must not be
4097 * less than current safe position.
4099 * In all this the minimum difference in data offsets
4100 * (conf->offset_diff - always positive) allows a bit of slack,
4101 * so next can be after 'safe', but not by more than offset_disk
4103 * We need to prepare all the bios here before we start any IO
4104 * to ensure the size we choose is acceptable to all devices.
4105 * The means one for each copy for write-out and an extra one for
4107 * We store the read-in bio in ->master_bio and the others in
4108 * ->devs[x].bio and ->devs[x].repl_bio.
4110 struct r10conf
*conf
= mddev
->private;
4111 struct r10bio
*r10_bio
;
4112 sector_t next
, safe
, last
;
4116 struct md_rdev
*rdev
;
4119 struct bio
*bio
, *read_bio
;
4120 int sectors_done
= 0;
4122 if (sector_nr
== 0) {
4123 /* If restarting in the middle, skip the initial sectors */
4124 if (mddev
->reshape_backwards
&&
4125 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4126 sector_nr
= (raid10_size(mddev
, 0, 0)
4127 - conf
->reshape_progress
);
4128 } else if (!mddev
->reshape_backwards
&&
4129 conf
->reshape_progress
> 0)
4130 sector_nr
= conf
->reshape_progress
;
4132 mddev
->curr_resync_completed
= sector_nr
;
4133 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4139 /* We don't use sector_nr to track where we are up to
4140 * as that doesn't work well for ->reshape_backwards.
4141 * So just use ->reshape_progress.
4143 if (mddev
->reshape_backwards
) {
4144 /* 'next' is the earliest device address that we might
4145 * write to for this chunk in the new layout
4147 next
= first_dev_address(conf
->reshape_progress
- 1,
4150 /* 'safe' is the last device address that we might read from
4151 * in the old layout after a restart
4153 safe
= last_dev_address(conf
->reshape_safe
- 1,
4156 if (next
+ conf
->offset_diff
< safe
)
4159 last
= conf
->reshape_progress
- 1;
4160 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4161 & conf
->prev
.chunk_mask
);
4162 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4163 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4165 /* 'next' is after the last device address that we
4166 * might write to for this chunk in the new layout
4168 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4170 /* 'safe' is the earliest device address that we might
4171 * read from in the old layout after a restart
4173 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4175 /* Need to update metadata if 'next' might be beyond 'safe'
4176 * as that would possibly corrupt data
4178 if (next
> safe
+ conf
->offset_diff
)
4181 sector_nr
= conf
->reshape_progress
;
4182 last
= sector_nr
| (conf
->geo
.chunk_mask
4183 & conf
->prev
.chunk_mask
);
4185 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4186 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4190 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4191 /* Need to update reshape_position in metadata */
4193 mddev
->reshape_position
= conf
->reshape_progress
;
4194 if (mddev
->reshape_backwards
)
4195 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4196 - conf
->reshape_progress
;
4198 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4199 conf
->reshape_checkpoint
= jiffies
;
4200 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4201 md_wakeup_thread(mddev
->thread
);
4202 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4203 kthread_should_stop());
4204 conf
->reshape_safe
= mddev
->reshape_position
;
4205 allow_barrier(conf
);
4209 /* Now schedule reads for blocks from sector_nr to last */
4210 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4211 raise_barrier(conf
, sectors_done
!= 0);
4212 atomic_set(&r10_bio
->remaining
, 0);
4213 r10_bio
->mddev
= mddev
;
4214 r10_bio
->sector
= sector_nr
;
4215 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4216 r10_bio
->sectors
= last
- sector_nr
+ 1;
4217 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4218 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4221 /* Cannot read from here, so need to record bad blocks
4222 * on all the target devices.
4225 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4226 return sectors_done
;
4229 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4231 read_bio
->bi_bdev
= rdev
->bdev
;
4232 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4233 + rdev
->data_offset
);
4234 read_bio
->bi_private
= r10_bio
;
4235 read_bio
->bi_end_io
= end_sync_read
;
4236 read_bio
->bi_rw
= READ
;
4237 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4238 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4239 read_bio
->bi_vcnt
= 0;
4240 read_bio
->bi_idx
= 0;
4241 read_bio
->bi_size
= 0;
4242 r10_bio
->master_bio
= read_bio
;
4243 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4245 /* Now find the locations in the new layout */
4246 __raid10_find_phys(&conf
->geo
, r10_bio
);
4249 read_bio
->bi_next
= NULL
;
4251 for (s
= 0; s
< conf
->copies
*2; s
++) {
4253 int d
= r10_bio
->devs
[s
/2].devnum
;
4254 struct md_rdev
*rdev2
;
4256 rdev2
= conf
->mirrors
[d
].replacement
;
4257 b
= r10_bio
->devs
[s
/2].repl_bio
;
4259 rdev2
= conf
->mirrors
[d
].rdev
;
4260 b
= r10_bio
->devs
[s
/2].bio
;
4262 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4264 b
->bi_bdev
= rdev2
->bdev
;
4265 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4266 b
->bi_private
= r10_bio
;
4267 b
->bi_end_io
= end_reshape_write
;
4269 b
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4270 b
->bi_flags
|= 1 << BIO_UPTODATE
;
4278 /* Now add as many pages as possible to all of these bios. */
4281 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4282 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4283 int len
= (max_sectors
- s
) << 9;
4284 if (len
> PAGE_SIZE
)
4286 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4288 if (bio_add_page(bio
, page
, len
, 0))
4291 /* Didn't fit, must stop */
4293 bio2
&& bio2
!= bio
;
4294 bio2
= bio2
->bi_next
) {
4295 /* Remove last page from this bio */
4297 bio2
->bi_size
-= len
;
4298 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4302 sector_nr
+= len
>> 9;
4303 nr_sectors
+= len
>> 9;
4306 r10_bio
->sectors
= nr_sectors
;
4308 /* Now submit the read */
4309 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4310 atomic_inc(&r10_bio
->remaining
);
4311 read_bio
->bi_next
= NULL
;
4312 generic_make_request(read_bio
);
4313 sector_nr
+= nr_sectors
;
4314 sectors_done
+= nr_sectors
;
4315 if (sector_nr
<= last
)
4318 /* Now that we have done the whole section we can
4319 * update reshape_progress
4321 if (mddev
->reshape_backwards
)
4322 conf
->reshape_progress
-= sectors_done
;
4324 conf
->reshape_progress
+= sectors_done
;
4326 return sectors_done
;
4329 static void end_reshape_request(struct r10bio
*r10_bio
);
4330 static int handle_reshape_read_error(struct mddev
*mddev
,
4331 struct r10bio
*r10_bio
);
4332 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4334 /* Reshape read completed. Hopefully we have a block
4336 * If we got a read error then we do sync 1-page reads from
4337 * elsewhere until we find the data - or give up.
4339 struct r10conf
*conf
= mddev
->private;
4342 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4343 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4344 /* Reshape has been aborted */
4345 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4349 /* We definitely have the data in the pages, schedule the
4352 atomic_set(&r10_bio
->remaining
, 1);
4353 for (s
= 0; s
< conf
->copies
*2; s
++) {
4355 int d
= r10_bio
->devs
[s
/2].devnum
;
4356 struct md_rdev
*rdev
;
4358 rdev
= conf
->mirrors
[d
].replacement
;
4359 b
= r10_bio
->devs
[s
/2].repl_bio
;
4361 rdev
= conf
->mirrors
[d
].rdev
;
4362 b
= r10_bio
->devs
[s
/2].bio
;
4364 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4366 atomic_inc(&rdev
->nr_pending
);
4367 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4368 atomic_inc(&r10_bio
->remaining
);
4370 generic_make_request(b
);
4372 end_reshape_request(r10_bio
);
4375 static void end_reshape(struct r10conf
*conf
)
4377 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4380 spin_lock_irq(&conf
->device_lock
);
4381 conf
->prev
= conf
->geo
;
4382 md_finish_reshape(conf
->mddev
);
4384 conf
->reshape_progress
= MaxSector
;
4385 spin_unlock_irq(&conf
->device_lock
);
4387 /* read-ahead size must cover two whole stripes, which is
4388 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4390 if (conf
->mddev
->queue
) {
4391 int stripe
= conf
->geo
.raid_disks
*
4392 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4393 stripe
/= conf
->geo
.near_copies
;
4394 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4395 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4401 static int handle_reshape_read_error(struct mddev
*mddev
,
4402 struct r10bio
*r10_bio
)
4404 /* Use sync reads to get the blocks from somewhere else */
4405 int sectors
= r10_bio
->sectors
;
4407 struct r10conf
*conf
= mddev
->private;
4410 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4412 r10b
.sector
= r10_bio
->sector
;
4413 __raid10_find_phys(&conf
->prev
, &r10b
);
4418 int first_slot
= slot
;
4420 if (s
> (PAGE_SIZE
>> 9))
4424 int d
= r10b
.devs
[slot
].devnum
;
4425 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4428 test_bit(Faulty
, &rdev
->flags
) ||
4429 !test_bit(In_sync
, &rdev
->flags
))
4432 addr
= r10b
.devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4433 success
= sync_page_io(rdev
,
4442 if (slot
>= conf
->copies
)
4444 if (slot
== first_slot
)
4448 /* couldn't read this block, must give up */
4449 set_bit(MD_RECOVERY_INTR
,
4459 static void end_reshape_write(struct bio
*bio
, int error
)
4461 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4462 struct r10bio
*r10_bio
= bio
->bi_private
;
4463 struct mddev
*mddev
= r10_bio
->mddev
;
4464 struct r10conf
*conf
= mddev
->private;
4468 struct md_rdev
*rdev
= NULL
;
4470 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4472 rdev
= conf
->mirrors
[d
].replacement
;
4475 rdev
= conf
->mirrors
[d
].rdev
;
4479 /* FIXME should record badblock */
4480 md_error(mddev
, rdev
);
4483 rdev_dec_pending(rdev
, mddev
);
4484 end_reshape_request(r10_bio
);
4487 static void end_reshape_request(struct r10bio
*r10_bio
)
4489 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4491 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4492 bio_put(r10_bio
->master_bio
);
4496 static void raid10_finish_reshape(struct mddev
*mddev
)
4498 struct r10conf
*conf
= mddev
->private;
4500 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4503 if (mddev
->delta_disks
> 0) {
4504 sector_t size
= raid10_size(mddev
, 0, 0);
4505 md_set_array_sectors(mddev
, size
);
4506 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4507 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4508 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4510 mddev
->resync_max_sectors
= size
;
4511 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4512 revalidate_disk(mddev
->gendisk
);
4515 for (d
= conf
->geo
.raid_disks
;
4516 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4518 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4520 clear_bit(In_sync
, &rdev
->flags
);
4521 rdev
= conf
->mirrors
[d
].replacement
;
4523 clear_bit(In_sync
, &rdev
->flags
);
4526 mddev
->layout
= mddev
->new_layout
;
4527 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4528 mddev
->reshape_position
= MaxSector
;
4529 mddev
->delta_disks
= 0;
4530 mddev
->reshape_backwards
= 0;
4533 static struct md_personality raid10_personality
=
4537 .owner
= THIS_MODULE
,
4538 .make_request
= make_request
,
4542 .error_handler
= error
,
4543 .hot_add_disk
= raid10_add_disk
,
4544 .hot_remove_disk
= raid10_remove_disk
,
4545 .spare_active
= raid10_spare_active
,
4546 .sync_request
= sync_request
,
4547 .quiesce
= raid10_quiesce
,
4548 .size
= raid10_size
,
4549 .resize
= raid10_resize
,
4550 .takeover
= raid10_takeover
,
4551 .check_reshape
= raid10_check_reshape
,
4552 .start_reshape
= raid10_start_reshape
,
4553 .finish_reshape
= raid10_finish_reshape
,
4556 static int __init
raid_init(void)
4558 return register_md_personality(&raid10_personality
);
4561 static void raid_exit(void)
4563 unregister_md_personality(&raid10_personality
);
4566 module_init(raid_init
);
4567 module_exit(raid_exit
);
4568 MODULE_LICENSE("GPL");
4569 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4570 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4571 MODULE_ALIAS("md-raid10");
4572 MODULE_ALIAS("md-level-10");
4574 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);