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
) {
663 struct r10bio r10_bio
;
664 struct r10dev devs
[conf
->copies
];
666 struct r10bio
*r10_bio
= &on_stack
.r10_bio
;
668 if (conf
->reshape_progress
!= MaxSector
) {
669 /* Cannot give any guidance during reshape */
670 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
671 return biovec
->bv_len
;
674 r10_bio
->sector
= sector
;
675 raid10_find_phys(conf
, r10_bio
);
677 for (s
= 0; s
< conf
->copies
; s
++) {
678 int disk
= r10_bio
->devs
[s
].devnum
;
679 struct md_rdev
*rdev
= rcu_dereference(
680 conf
->mirrors
[disk
].rdev
);
681 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
682 struct request_queue
*q
=
683 bdev_get_queue(rdev
->bdev
);
684 if (q
->merge_bvec_fn
) {
685 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
687 bvm
->bi_bdev
= rdev
->bdev
;
688 max
= min(max
, q
->merge_bvec_fn(
692 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
693 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
694 struct request_queue
*q
=
695 bdev_get_queue(rdev
->bdev
);
696 if (q
->merge_bvec_fn
) {
697 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
699 bvm
->bi_bdev
= rdev
->bdev
;
700 max
= min(max
, q
->merge_bvec_fn(
711 * This routine returns the disk from which the requested read should
712 * be done. There is a per-array 'next expected sequential IO' sector
713 * number - if this matches on the next IO then we use the last disk.
714 * There is also a per-disk 'last know head position' sector that is
715 * maintained from IRQ contexts, both the normal and the resync IO
716 * completion handlers update this position correctly. If there is no
717 * perfect sequential match then we pick the disk whose head is closest.
719 * If there are 2 mirrors in the same 2 devices, performance degrades
720 * because position is mirror, not device based.
722 * The rdev for the device selected will have nr_pending incremented.
726 * FIXME: possibly should rethink readbalancing and do it differently
727 * depending on near_copies / far_copies geometry.
729 static struct md_rdev
*read_balance(struct r10conf
*conf
,
730 struct r10bio
*r10_bio
,
733 const sector_t this_sector
= r10_bio
->sector
;
735 int sectors
= r10_bio
->sectors
;
736 int best_good_sectors
;
737 sector_t new_distance
, best_dist
;
738 struct md_rdev
*best_rdev
, *rdev
= NULL
;
741 struct geom
*geo
= &conf
->geo
;
743 raid10_find_phys(conf
, r10_bio
);
746 sectors
= r10_bio
->sectors
;
749 best_dist
= MaxSector
;
750 best_good_sectors
= 0;
753 * Check if we can balance. We can balance on the whole
754 * device if no resync is going on (recovery is ok), or below
755 * the resync window. We take the first readable disk when
756 * above the resync window.
758 if (conf
->mddev
->recovery_cp
< MaxSector
759 && (this_sector
+ sectors
>= conf
->next_resync
))
762 for (slot
= 0; slot
< conf
->copies
; slot
++) {
767 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
769 disk
= r10_bio
->devs
[slot
].devnum
;
770 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
771 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
772 test_bit(Unmerged
, &rdev
->flags
) ||
773 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
774 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
776 test_bit(Faulty
, &rdev
->flags
) ||
777 test_bit(Unmerged
, &rdev
->flags
))
779 if (!test_bit(In_sync
, &rdev
->flags
) &&
780 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
783 dev_sector
= r10_bio
->devs
[slot
].addr
;
784 if (is_badblock(rdev
, dev_sector
, sectors
,
785 &first_bad
, &bad_sectors
)) {
786 if (best_dist
< MaxSector
)
787 /* Already have a better slot */
789 if (first_bad
<= dev_sector
) {
790 /* Cannot read here. If this is the
791 * 'primary' device, then we must not read
792 * beyond 'bad_sectors' from another device.
794 bad_sectors
-= (dev_sector
- first_bad
);
795 if (!do_balance
&& sectors
> bad_sectors
)
796 sectors
= bad_sectors
;
797 if (best_good_sectors
> sectors
)
798 best_good_sectors
= sectors
;
800 sector_t good_sectors
=
801 first_bad
- dev_sector
;
802 if (good_sectors
> best_good_sectors
) {
803 best_good_sectors
= good_sectors
;
808 /* Must read from here */
813 best_good_sectors
= sectors
;
818 /* This optimisation is debatable, and completely destroys
819 * sequential read speed for 'far copies' arrays. So only
820 * keep it for 'near' arrays, and review those later.
822 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
825 /* for far > 1 always use the lowest address */
826 if (geo
->far_copies
> 1)
827 new_distance
= r10_bio
->devs
[slot
].addr
;
829 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
830 conf
->mirrors
[disk
].head_position
);
831 if (new_distance
< best_dist
) {
832 best_dist
= new_distance
;
837 if (slot
>= conf
->copies
) {
843 atomic_inc(&rdev
->nr_pending
);
844 if (test_bit(Faulty
, &rdev
->flags
)) {
845 /* Cannot risk returning a device that failed
846 * before we inc'ed nr_pending
848 rdev_dec_pending(rdev
, conf
->mddev
);
851 r10_bio
->read_slot
= slot
;
855 *max_sectors
= best_good_sectors
;
860 int md_raid10_congested(struct mddev
*mddev
, int bits
)
862 struct r10conf
*conf
= mddev
->private;
865 if ((bits
& (1 << BDI_async_congested
)) &&
866 conf
->pending_count
>= max_queued_requests
)
871 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
874 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
875 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
876 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
878 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
884 EXPORT_SYMBOL_GPL(md_raid10_congested
);
886 static int raid10_congested(void *data
, int bits
)
888 struct mddev
*mddev
= data
;
890 return mddev_congested(mddev
, bits
) ||
891 md_raid10_congested(mddev
, bits
);
894 static void flush_pending_writes(struct r10conf
*conf
)
896 /* Any writes that have been queued but are awaiting
897 * bitmap updates get flushed here.
899 spin_lock_irq(&conf
->device_lock
);
901 if (conf
->pending_bio_list
.head
) {
903 bio
= bio_list_get(&conf
->pending_bio_list
);
904 conf
->pending_count
= 0;
905 spin_unlock_irq(&conf
->device_lock
);
906 /* flush any pending bitmap writes to disk
907 * before proceeding w/ I/O */
908 bitmap_unplug(conf
->mddev
->bitmap
);
909 wake_up(&conf
->wait_barrier
);
911 while (bio
) { /* submit pending writes */
912 struct bio
*next
= bio
->bi_next
;
914 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
915 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
919 generic_make_request(bio
);
923 spin_unlock_irq(&conf
->device_lock
);
927 * Sometimes we need to suspend IO while we do something else,
928 * either some resync/recovery, or reconfigure the array.
929 * To do this we raise a 'barrier'.
930 * The 'barrier' is a counter that can be raised multiple times
931 * to count how many activities are happening which preclude
933 * We can only raise the barrier if there is no pending IO.
934 * i.e. if nr_pending == 0.
935 * We choose only to raise the barrier if no-one is waiting for the
936 * barrier to go down. This means that as soon as an IO request
937 * is ready, no other operations which require a barrier will start
938 * until the IO request has had a chance.
940 * So: regular IO calls 'wait_barrier'. When that returns there
941 * is no backgroup IO happening, It must arrange to call
942 * allow_barrier when it has finished its IO.
943 * backgroup IO calls must call raise_barrier. Once that returns
944 * there is no normal IO happeing. It must arrange to call
945 * lower_barrier when the particular background IO completes.
948 static void raise_barrier(struct r10conf
*conf
, int force
)
950 BUG_ON(force
&& !conf
->barrier
);
951 spin_lock_irq(&conf
->resync_lock
);
953 /* Wait until no block IO is waiting (unless 'force') */
954 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
957 /* block any new IO from starting */
960 /* Now wait for all pending IO to complete */
961 wait_event_lock_irq(conf
->wait_barrier
,
962 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
965 spin_unlock_irq(&conf
->resync_lock
);
968 static void lower_barrier(struct r10conf
*conf
)
971 spin_lock_irqsave(&conf
->resync_lock
, flags
);
973 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
974 wake_up(&conf
->wait_barrier
);
977 static void wait_barrier(struct r10conf
*conf
)
979 spin_lock_irq(&conf
->resync_lock
);
982 /* Wait for the barrier to drop.
983 * However if there are already pending
984 * requests (preventing the barrier from
985 * rising completely), and the
986 * pre-process bio queue isn't empty,
987 * then don't wait, as we need to empty
988 * that queue to get the nr_pending
991 wait_event_lock_irq(conf
->wait_barrier
,
995 !bio_list_empty(current
->bio_list
)),
1000 spin_unlock_irq(&conf
->resync_lock
);
1003 static void allow_barrier(struct r10conf
*conf
)
1005 unsigned long flags
;
1006 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1008 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1009 wake_up(&conf
->wait_barrier
);
1012 static void freeze_array(struct r10conf
*conf
)
1014 /* stop syncio and normal IO and wait for everything to
1016 * We increment barrier and nr_waiting, and then
1017 * wait until nr_pending match nr_queued+1
1018 * This is called in the context of one normal IO request
1019 * that has failed. Thus any sync request that might be pending
1020 * will be blocked by nr_pending, and we need to wait for
1021 * pending IO requests to complete or be queued for re-try.
1022 * Thus the number queued (nr_queued) plus this request (1)
1023 * must match the number of pending IOs (nr_pending) before
1026 spin_lock_irq(&conf
->resync_lock
);
1029 wait_event_lock_irq_cmd(conf
->wait_barrier
,
1030 conf
->nr_pending
== conf
->nr_queued
+1,
1032 flush_pending_writes(conf
));
1034 spin_unlock_irq(&conf
->resync_lock
);
1037 static void unfreeze_array(struct r10conf
*conf
)
1039 /* reverse the effect of the freeze */
1040 spin_lock_irq(&conf
->resync_lock
);
1043 wake_up(&conf
->wait_barrier
);
1044 spin_unlock_irq(&conf
->resync_lock
);
1047 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1048 struct md_rdev
*rdev
)
1050 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1051 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1052 return rdev
->data_offset
;
1054 return rdev
->new_data_offset
;
1057 struct raid10_plug_cb
{
1058 struct blk_plug_cb cb
;
1059 struct bio_list pending
;
1063 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1065 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1067 struct mddev
*mddev
= plug
->cb
.data
;
1068 struct r10conf
*conf
= mddev
->private;
1071 if (from_schedule
) {
1072 spin_lock_irq(&conf
->device_lock
);
1073 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1074 conf
->pending_count
+= plug
->pending_cnt
;
1075 spin_unlock_irq(&conf
->device_lock
);
1076 md_wakeup_thread(mddev
->thread
);
1081 /* we aren't scheduling, so we can do the write-out directly. */
1082 bio
= bio_list_get(&plug
->pending
);
1083 bitmap_unplug(mddev
->bitmap
);
1084 wake_up(&conf
->wait_barrier
);
1086 while (bio
) { /* submit pending writes */
1087 struct bio
*next
= bio
->bi_next
;
1088 bio
->bi_next
= NULL
;
1089 generic_make_request(bio
);
1095 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1097 struct r10conf
*conf
= mddev
->private;
1098 struct r10bio
*r10_bio
;
1099 struct bio
*read_bio
;
1101 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1102 int chunk_sects
= chunk_mask
+ 1;
1103 const int rw
= bio_data_dir(bio
);
1104 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1105 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1106 const unsigned long do_discard
= (bio
->bi_rw
1107 & (REQ_DISCARD
| REQ_SECURE
));
1108 unsigned long flags
;
1109 struct md_rdev
*blocked_rdev
;
1110 struct blk_plug_cb
*cb
;
1111 struct raid10_plug_cb
*plug
= NULL
;
1112 int sectors_handled
;
1116 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1117 md_flush_request(mddev
, bio
);
1121 /* If this request crosses a chunk boundary, we need to
1122 * split it. This will only happen for 1 PAGE (or less) requests.
1124 if (unlikely((bio
->bi_sector
& chunk_mask
) + (bio
->bi_size
>> 9)
1126 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1127 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1128 struct bio_pair
*bp
;
1129 /* Sanity check -- queue functions should prevent this happening */
1130 if ((bio
->bi_vcnt
!= 1 && bio
->bi_vcnt
!= 0) ||
1133 /* This is a one page bio that upper layers
1134 * refuse to split for us, so we need to split it.
1137 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
1139 /* Each of these 'make_request' calls will call 'wait_barrier'.
1140 * If the first succeeds but the second blocks due to the resync
1141 * thread raising the barrier, we will deadlock because the
1142 * IO to the underlying device will be queued in generic_make_request
1143 * and will never complete, so will never reduce nr_pending.
1144 * So increment nr_waiting here so no new raise_barriers will
1145 * succeed, and so the second wait_barrier cannot block.
1147 spin_lock_irq(&conf
->resync_lock
);
1149 spin_unlock_irq(&conf
->resync_lock
);
1151 make_request(mddev
, &bp
->bio1
);
1152 make_request(mddev
, &bp
->bio2
);
1154 spin_lock_irq(&conf
->resync_lock
);
1156 wake_up(&conf
->wait_barrier
);
1157 spin_unlock_irq(&conf
->resync_lock
);
1159 bio_pair_release(bp
);
1162 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1163 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1164 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
1170 md_write_start(mddev
, bio
);
1173 * Register the new request and wait if the reconstruction
1174 * thread has put up a bar for new requests.
1175 * Continue immediately if no resync is active currently.
1179 sectors
= bio
->bi_size
>> 9;
1180 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1181 bio
->bi_sector
< conf
->reshape_progress
&&
1182 bio
->bi_sector
+ sectors
> conf
->reshape_progress
) {
1183 /* IO spans the reshape position. Need to wait for
1186 allow_barrier(conf
);
1187 wait_event(conf
->wait_barrier
,
1188 conf
->reshape_progress
<= bio
->bi_sector
||
1189 conf
->reshape_progress
>= bio
->bi_sector
+ sectors
);
1192 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1193 bio_data_dir(bio
) == WRITE
&&
1194 (mddev
->reshape_backwards
1195 ? (bio
->bi_sector
< conf
->reshape_safe
&&
1196 bio
->bi_sector
+ sectors
> conf
->reshape_progress
)
1197 : (bio
->bi_sector
+ sectors
> conf
->reshape_safe
&&
1198 bio
->bi_sector
< conf
->reshape_progress
))) {
1199 /* Need to update reshape_position in metadata */
1200 mddev
->reshape_position
= conf
->reshape_progress
;
1201 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1202 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1203 md_wakeup_thread(mddev
->thread
);
1204 wait_event(mddev
->sb_wait
,
1205 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1207 conf
->reshape_safe
= mddev
->reshape_position
;
1210 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1212 r10_bio
->master_bio
= bio
;
1213 r10_bio
->sectors
= sectors
;
1215 r10_bio
->mddev
= mddev
;
1216 r10_bio
->sector
= bio
->bi_sector
;
1219 /* We might need to issue multiple reads to different
1220 * devices if there are bad blocks around, so we keep
1221 * track of the number of reads in bio->bi_phys_segments.
1222 * If this is 0, there is only one r10_bio and no locking
1223 * will be needed when the request completes. If it is
1224 * non-zero, then it is the number of not-completed requests.
1226 bio
->bi_phys_segments
= 0;
1227 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1231 * read balancing logic:
1233 struct md_rdev
*rdev
;
1237 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1239 raid_end_bio_io(r10_bio
);
1242 slot
= r10_bio
->read_slot
;
1244 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1245 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1248 r10_bio
->devs
[slot
].bio
= read_bio
;
1249 r10_bio
->devs
[slot
].rdev
= rdev
;
1251 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1252 choose_data_offset(r10_bio
, rdev
);
1253 read_bio
->bi_bdev
= rdev
->bdev
;
1254 read_bio
->bi_end_io
= raid10_end_read_request
;
1255 read_bio
->bi_rw
= READ
| do_sync
;
1256 read_bio
->bi_private
= r10_bio
;
1258 if (max_sectors
< r10_bio
->sectors
) {
1259 /* Could not read all from this device, so we will
1260 * need another r10_bio.
1262 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1264 r10_bio
->sectors
= max_sectors
;
1265 spin_lock_irq(&conf
->device_lock
);
1266 if (bio
->bi_phys_segments
== 0)
1267 bio
->bi_phys_segments
= 2;
1269 bio
->bi_phys_segments
++;
1270 spin_unlock(&conf
->device_lock
);
1271 /* Cannot call generic_make_request directly
1272 * as that will be queued in __generic_make_request
1273 * and subsequent mempool_alloc might block
1274 * waiting for it. so hand bio over to raid10d.
1276 reschedule_retry(r10_bio
);
1278 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1280 r10_bio
->master_bio
= bio
;
1281 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
1284 r10_bio
->mddev
= mddev
;
1285 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1288 generic_make_request(read_bio
);
1295 if (conf
->pending_count
>= max_queued_requests
) {
1296 md_wakeup_thread(mddev
->thread
);
1297 wait_event(conf
->wait_barrier
,
1298 conf
->pending_count
< max_queued_requests
);
1300 /* first select target devices under rcu_lock and
1301 * inc refcount on their rdev. Record them by setting
1303 * If there are known/acknowledged bad blocks on any device
1304 * on which we have seen a write error, we want to avoid
1305 * writing to those blocks. This potentially requires several
1306 * writes to write around the bad blocks. Each set of writes
1307 * gets its own r10_bio with a set of bios attached. The number
1308 * of r10_bios is recored in bio->bi_phys_segments just as with
1312 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1313 raid10_find_phys(conf
, r10_bio
);
1315 blocked_rdev
= NULL
;
1317 max_sectors
= r10_bio
->sectors
;
1319 for (i
= 0; i
< conf
->copies
; i
++) {
1320 int d
= r10_bio
->devs
[i
].devnum
;
1321 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1322 struct md_rdev
*rrdev
= rcu_dereference(
1323 conf
->mirrors
[d
].replacement
);
1326 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1327 atomic_inc(&rdev
->nr_pending
);
1328 blocked_rdev
= rdev
;
1331 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1332 atomic_inc(&rrdev
->nr_pending
);
1333 blocked_rdev
= rrdev
;
1336 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1337 || test_bit(Unmerged
, &rrdev
->flags
)))
1340 r10_bio
->devs
[i
].bio
= NULL
;
1341 r10_bio
->devs
[i
].repl_bio
= NULL
;
1342 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
1343 test_bit(Unmerged
, &rdev
->flags
)) {
1344 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1347 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1349 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1353 is_bad
= is_badblock(rdev
, dev_sector
,
1355 &first_bad
, &bad_sectors
);
1357 /* Mustn't write here until the bad block
1360 atomic_inc(&rdev
->nr_pending
);
1361 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1362 blocked_rdev
= rdev
;
1365 if (is_bad
&& first_bad
<= dev_sector
) {
1366 /* Cannot write here at all */
1367 bad_sectors
-= (dev_sector
- first_bad
);
1368 if (bad_sectors
< max_sectors
)
1369 /* Mustn't write more than bad_sectors
1370 * to other devices yet
1372 max_sectors
= bad_sectors
;
1373 /* We don't set R10BIO_Degraded as that
1374 * only applies if the disk is missing,
1375 * so it might be re-added, and we want to
1376 * know to recover this chunk.
1377 * In this case the device is here, and the
1378 * fact that this chunk is not in-sync is
1379 * recorded in the bad block log.
1384 int good_sectors
= first_bad
- dev_sector
;
1385 if (good_sectors
< max_sectors
)
1386 max_sectors
= good_sectors
;
1389 r10_bio
->devs
[i
].bio
= bio
;
1390 atomic_inc(&rdev
->nr_pending
);
1392 r10_bio
->devs
[i
].repl_bio
= bio
;
1393 atomic_inc(&rrdev
->nr_pending
);
1398 if (unlikely(blocked_rdev
)) {
1399 /* Have to wait for this device to get unblocked, then retry */
1403 for (j
= 0; j
< i
; j
++) {
1404 if (r10_bio
->devs
[j
].bio
) {
1405 d
= r10_bio
->devs
[j
].devnum
;
1406 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1408 if (r10_bio
->devs
[j
].repl_bio
) {
1409 struct md_rdev
*rdev
;
1410 d
= r10_bio
->devs
[j
].devnum
;
1411 rdev
= conf
->mirrors
[d
].replacement
;
1413 /* Race with remove_disk */
1415 rdev
= conf
->mirrors
[d
].rdev
;
1417 rdev_dec_pending(rdev
, mddev
);
1420 allow_barrier(conf
);
1421 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1426 if (max_sectors
< r10_bio
->sectors
) {
1427 /* We are splitting this into multiple parts, so
1428 * we need to prepare for allocating another r10_bio.
1430 r10_bio
->sectors
= max_sectors
;
1431 spin_lock_irq(&conf
->device_lock
);
1432 if (bio
->bi_phys_segments
== 0)
1433 bio
->bi_phys_segments
= 2;
1435 bio
->bi_phys_segments
++;
1436 spin_unlock_irq(&conf
->device_lock
);
1438 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1440 atomic_set(&r10_bio
->remaining
, 1);
1441 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1443 for (i
= 0; i
< conf
->copies
; i
++) {
1445 int d
= r10_bio
->devs
[i
].devnum
;
1446 if (!r10_bio
->devs
[i
].bio
)
1449 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1450 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1452 r10_bio
->devs
[i
].bio
= mbio
;
1454 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1455 choose_data_offset(r10_bio
,
1456 conf
->mirrors
[d
].rdev
));
1457 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1458 mbio
->bi_end_io
= raid10_end_write_request
;
1459 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
| do_discard
;
1460 mbio
->bi_private
= r10_bio
;
1462 atomic_inc(&r10_bio
->remaining
);
1464 cb
= blk_check_plugged(raid10_unplug
, mddev
, sizeof(*plug
));
1466 plug
= container_of(cb
, struct raid10_plug_cb
, cb
);
1469 spin_lock_irqsave(&conf
->device_lock
, flags
);
1471 bio_list_add(&plug
->pending
, mbio
);
1472 plug
->pending_cnt
++;
1474 bio_list_add(&conf
->pending_bio_list
, mbio
);
1475 conf
->pending_count
++;
1477 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1479 md_wakeup_thread(mddev
->thread
);
1481 if (!r10_bio
->devs
[i
].repl_bio
)
1484 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1485 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1487 r10_bio
->devs
[i
].repl_bio
= mbio
;
1489 /* We are actively writing to the original device
1490 * so it cannot disappear, so the replacement cannot
1493 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1496 conf
->mirrors
[d
].replacement
));
1497 mbio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
1498 mbio
->bi_end_io
= raid10_end_write_request
;
1499 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
| do_discard
;
1500 mbio
->bi_private
= r10_bio
;
1502 atomic_inc(&r10_bio
->remaining
);
1503 spin_lock_irqsave(&conf
->device_lock
, flags
);
1504 bio_list_add(&conf
->pending_bio_list
, mbio
);
1505 conf
->pending_count
++;
1506 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1507 if (!mddev_check_plugged(mddev
))
1508 md_wakeup_thread(mddev
->thread
);
1511 /* Don't remove the bias on 'remaining' (one_write_done) until
1512 * after checking if we need to go around again.
1515 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1516 one_write_done(r10_bio
);
1517 /* We need another r10_bio. It has already been counted
1518 * in bio->bi_phys_segments.
1520 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1522 r10_bio
->master_bio
= bio
;
1523 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1525 r10_bio
->mddev
= mddev
;
1526 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1530 one_write_done(r10_bio
);
1532 /* In case raid10d snuck in to freeze_array */
1533 wake_up(&conf
->wait_barrier
);
1536 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1538 struct r10conf
*conf
= mddev
->private;
1541 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1542 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1543 if (conf
->geo
.near_copies
> 1)
1544 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1545 if (conf
->geo
.far_copies
> 1) {
1546 if (conf
->geo
.far_offset
)
1547 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1549 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1551 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1552 conf
->geo
.raid_disks
- mddev
->degraded
);
1553 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1554 seq_printf(seq
, "%s",
1555 conf
->mirrors
[i
].rdev
&&
1556 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1557 seq_printf(seq
, "]");
1560 /* check if there are enough drives for
1561 * every block to appear on atleast one.
1562 * Don't consider the device numbered 'ignore'
1563 * as we might be about to remove it.
1565 static int _enough(struct r10conf
*conf
, struct geom
*geo
, int ignore
)
1570 int n
= conf
->copies
;
1574 if (conf
->mirrors
[this].rdev
&&
1577 this = (this+1) % geo
->raid_disks
;
1581 first
= (first
+ geo
->near_copies
) % geo
->raid_disks
;
1582 } while (first
!= 0);
1586 static int enough(struct r10conf
*conf
, int ignore
)
1588 return _enough(conf
, &conf
->geo
, ignore
) &&
1589 _enough(conf
, &conf
->prev
, ignore
);
1592 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1594 char b
[BDEVNAME_SIZE
];
1595 struct r10conf
*conf
= mddev
->private;
1598 * If it is not operational, then we have already marked it as dead
1599 * else if it is the last working disks, ignore the error, let the
1600 * next level up know.
1601 * else mark the drive as failed
1603 if (test_bit(In_sync
, &rdev
->flags
)
1604 && !enough(conf
, rdev
->raid_disk
))
1606 * Don't fail the drive, just return an IO error.
1609 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1610 unsigned long flags
;
1611 spin_lock_irqsave(&conf
->device_lock
, flags
);
1613 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1615 * if recovery is running, make sure it aborts.
1617 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1619 set_bit(Blocked
, &rdev
->flags
);
1620 set_bit(Faulty
, &rdev
->flags
);
1621 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1623 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1624 "md/raid10:%s: Operation continuing on %d devices.\n",
1625 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1626 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1629 static void print_conf(struct r10conf
*conf
)
1632 struct raid10_info
*tmp
;
1634 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1636 printk(KERN_DEBUG
"(!conf)\n");
1639 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1640 conf
->geo
.raid_disks
);
1642 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1643 char b
[BDEVNAME_SIZE
];
1644 tmp
= conf
->mirrors
+ i
;
1646 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1647 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1648 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1649 bdevname(tmp
->rdev
->bdev
,b
));
1653 static void close_sync(struct r10conf
*conf
)
1656 allow_barrier(conf
);
1658 mempool_destroy(conf
->r10buf_pool
);
1659 conf
->r10buf_pool
= NULL
;
1662 static int raid10_spare_active(struct mddev
*mddev
)
1665 struct r10conf
*conf
= mddev
->private;
1666 struct raid10_info
*tmp
;
1668 unsigned long flags
;
1671 * Find all non-in_sync disks within the RAID10 configuration
1672 * and mark them in_sync
1674 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1675 tmp
= conf
->mirrors
+ i
;
1676 if (tmp
->replacement
1677 && tmp
->replacement
->recovery_offset
== MaxSector
1678 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1679 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1680 /* Replacement has just become active */
1682 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1685 /* Replaced device not technically faulty,
1686 * but we need to be sure it gets removed
1687 * and never re-added.
1689 set_bit(Faulty
, &tmp
->rdev
->flags
);
1690 sysfs_notify_dirent_safe(
1691 tmp
->rdev
->sysfs_state
);
1693 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1694 } else if (tmp
->rdev
1695 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1696 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1698 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1701 spin_lock_irqsave(&conf
->device_lock
, flags
);
1702 mddev
->degraded
-= count
;
1703 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1710 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1712 struct r10conf
*conf
= mddev
->private;
1716 int last
= conf
->geo
.raid_disks
- 1;
1717 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1719 if (mddev
->recovery_cp
< MaxSector
)
1720 /* only hot-add to in-sync arrays, as recovery is
1721 * very different from resync
1724 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, &conf
->prev
, -1))
1727 if (rdev
->raid_disk
>= 0)
1728 first
= last
= rdev
->raid_disk
;
1730 if (q
->merge_bvec_fn
) {
1731 set_bit(Unmerged
, &rdev
->flags
);
1732 mddev
->merge_check_needed
= 1;
1735 if (rdev
->saved_raid_disk
>= first
&&
1736 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1737 mirror
= rdev
->saved_raid_disk
;
1740 for ( ; mirror
<= last
; mirror
++) {
1741 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1742 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1745 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1746 p
->replacement
!= NULL
)
1748 clear_bit(In_sync
, &rdev
->flags
);
1749 set_bit(Replacement
, &rdev
->flags
);
1750 rdev
->raid_disk
= mirror
;
1752 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1753 rdev
->data_offset
<< 9);
1755 rcu_assign_pointer(p
->replacement
, rdev
);
1759 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1760 rdev
->data_offset
<< 9);
1762 p
->head_position
= 0;
1763 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1764 rdev
->raid_disk
= mirror
;
1766 if (rdev
->saved_raid_disk
!= mirror
)
1768 rcu_assign_pointer(p
->rdev
, rdev
);
1771 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1772 /* Some requests might not have seen this new
1773 * merge_bvec_fn. We must wait for them to complete
1774 * before merging the device fully.
1775 * First we make sure any code which has tested
1776 * our function has submitted the request, then
1777 * we wait for all outstanding requests to complete.
1779 synchronize_sched();
1780 raise_barrier(conf
, 0);
1781 lower_barrier(conf
);
1782 clear_bit(Unmerged
, &rdev
->flags
);
1784 md_integrity_add_rdev(rdev
, mddev
);
1785 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1786 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1792 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1794 struct r10conf
*conf
= mddev
->private;
1796 int number
= rdev
->raid_disk
;
1797 struct md_rdev
**rdevp
;
1798 struct raid10_info
*p
= conf
->mirrors
+ number
;
1801 if (rdev
== p
->rdev
)
1803 else if (rdev
== p
->replacement
)
1804 rdevp
= &p
->replacement
;
1808 if (test_bit(In_sync
, &rdev
->flags
) ||
1809 atomic_read(&rdev
->nr_pending
)) {
1813 /* Only remove faulty devices if recovery
1816 if (!test_bit(Faulty
, &rdev
->flags
) &&
1817 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1818 (!p
->replacement
|| p
->replacement
== rdev
) &&
1819 number
< conf
->geo
.raid_disks
&&
1826 if (atomic_read(&rdev
->nr_pending
)) {
1827 /* lost the race, try later */
1831 } else if (p
->replacement
) {
1832 /* We must have just cleared 'rdev' */
1833 p
->rdev
= p
->replacement
;
1834 clear_bit(Replacement
, &p
->replacement
->flags
);
1835 smp_mb(); /* Make sure other CPUs may see both as identical
1836 * but will never see neither -- if they are careful.
1838 p
->replacement
= NULL
;
1839 clear_bit(WantReplacement
, &rdev
->flags
);
1841 /* We might have just remove the Replacement as faulty
1842 * Clear the flag just in case
1844 clear_bit(WantReplacement
, &rdev
->flags
);
1846 err
= md_integrity_register(mddev
);
1855 static void end_sync_read(struct bio
*bio
, int error
)
1857 struct r10bio
*r10_bio
= bio
->bi_private
;
1858 struct r10conf
*conf
= r10_bio
->mddev
->private;
1861 if (bio
== r10_bio
->master_bio
) {
1862 /* this is a reshape read */
1863 d
= r10_bio
->read_slot
; /* really the read dev */
1865 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1867 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1868 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1870 /* The write handler will notice the lack of
1871 * R10BIO_Uptodate and record any errors etc
1873 atomic_add(r10_bio
->sectors
,
1874 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1876 /* for reconstruct, we always reschedule after a read.
1877 * for resync, only after all reads
1879 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1880 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1881 atomic_dec_and_test(&r10_bio
->remaining
)) {
1882 /* we have read all the blocks,
1883 * do the comparison in process context in raid10d
1885 reschedule_retry(r10_bio
);
1889 static void end_sync_request(struct r10bio
*r10_bio
)
1891 struct mddev
*mddev
= r10_bio
->mddev
;
1893 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1894 if (r10_bio
->master_bio
== NULL
) {
1895 /* the primary of several recovery bios */
1896 sector_t s
= r10_bio
->sectors
;
1897 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1898 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1899 reschedule_retry(r10_bio
);
1902 md_done_sync(mddev
, s
, 1);
1905 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1906 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1907 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1908 reschedule_retry(r10_bio
);
1916 static void end_sync_write(struct bio
*bio
, int error
)
1918 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1919 struct r10bio
*r10_bio
= bio
->bi_private
;
1920 struct mddev
*mddev
= r10_bio
->mddev
;
1921 struct r10conf
*conf
= mddev
->private;
1927 struct md_rdev
*rdev
= NULL
;
1929 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1931 rdev
= conf
->mirrors
[d
].replacement
;
1933 rdev
= conf
->mirrors
[d
].rdev
;
1937 md_error(mddev
, rdev
);
1939 set_bit(WriteErrorSeen
, &rdev
->flags
);
1940 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1941 set_bit(MD_RECOVERY_NEEDED
,
1942 &rdev
->mddev
->recovery
);
1943 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1945 } else if (is_badblock(rdev
,
1946 r10_bio
->devs
[slot
].addr
,
1948 &first_bad
, &bad_sectors
))
1949 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1951 rdev_dec_pending(rdev
, mddev
);
1953 end_sync_request(r10_bio
);
1957 * Note: sync and recover and handled very differently for raid10
1958 * This code is for resync.
1959 * For resync, we read through virtual addresses and read all blocks.
1960 * If there is any error, we schedule a write. The lowest numbered
1961 * drive is authoritative.
1962 * However requests come for physical address, so we need to map.
1963 * For every physical address there are raid_disks/copies virtual addresses,
1964 * which is always are least one, but is not necessarly an integer.
1965 * This means that a physical address can span multiple chunks, so we may
1966 * have to submit multiple io requests for a single sync request.
1969 * We check if all blocks are in-sync and only write to blocks that
1972 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1974 struct r10conf
*conf
= mddev
->private;
1976 struct bio
*tbio
, *fbio
;
1979 atomic_set(&r10_bio
->remaining
, 1);
1981 /* find the first device with a block */
1982 for (i
=0; i
<conf
->copies
; i
++)
1983 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1986 if (i
== conf
->copies
)
1990 fbio
= r10_bio
->devs
[i
].bio
;
1992 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
1993 /* now find blocks with errors */
1994 for (i
=0 ; i
< conf
->copies
; i
++) {
1997 tbio
= r10_bio
->devs
[i
].bio
;
1999 if (tbio
->bi_end_io
!= end_sync_read
)
2003 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2004 /* We know that the bi_io_vec layout is the same for
2005 * both 'first' and 'i', so we just compare them.
2006 * All vec entries are PAGE_SIZE;
2008 for (j
= 0; j
< vcnt
; j
++)
2009 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2010 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2011 fbio
->bi_io_vec
[j
].bv_len
))
2015 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2016 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2017 /* Don't fix anything. */
2020 /* Ok, we need to write this bio, either to correct an
2021 * inconsistency or to correct an unreadable block.
2022 * First we need to fixup bv_offset, bv_len and
2023 * bi_vecs, as the read request might have corrupted these
2025 tbio
->bi_vcnt
= vcnt
;
2026 tbio
->bi_size
= r10_bio
->sectors
<< 9;
2028 tbio
->bi_phys_segments
= 0;
2029 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2030 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
2031 tbio
->bi_next
= NULL
;
2032 tbio
->bi_rw
= WRITE
;
2033 tbio
->bi_private
= r10_bio
;
2034 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
2036 for (j
=0; j
< vcnt
; j
++) {
2037 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2038 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2040 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2041 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2044 tbio
->bi_end_io
= end_sync_write
;
2046 d
= r10_bio
->devs
[i
].devnum
;
2047 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2048 atomic_inc(&r10_bio
->remaining
);
2049 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
2051 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2052 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2053 generic_make_request(tbio
);
2056 /* Now write out to any replacement devices
2059 for (i
= 0; i
< conf
->copies
; i
++) {
2062 tbio
= r10_bio
->devs
[i
].repl_bio
;
2063 if (!tbio
|| !tbio
->bi_end_io
)
2065 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2066 && r10_bio
->devs
[i
].bio
!= fbio
)
2067 for (j
= 0; j
< vcnt
; j
++)
2068 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2069 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2071 d
= r10_bio
->devs
[i
].devnum
;
2072 atomic_inc(&r10_bio
->remaining
);
2073 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2074 tbio
->bi_size
>> 9);
2075 generic_make_request(tbio
);
2079 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2080 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2086 * Now for the recovery code.
2087 * Recovery happens across physical sectors.
2088 * We recover all non-is_sync drives by finding the virtual address of
2089 * each, and then choose a working drive that also has that virt address.
2090 * There is a separate r10_bio for each non-in_sync drive.
2091 * Only the first two slots are in use. The first for reading,
2092 * The second for writing.
2095 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2097 /* We got a read error during recovery.
2098 * We repeat the read in smaller page-sized sections.
2099 * If a read succeeds, write it to the new device or record
2100 * a bad block if we cannot.
2101 * If a read fails, record a bad block on both old and
2104 struct mddev
*mddev
= r10_bio
->mddev
;
2105 struct r10conf
*conf
= mddev
->private;
2106 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2108 int sectors
= r10_bio
->sectors
;
2110 int dr
= r10_bio
->devs
[0].devnum
;
2111 int dw
= r10_bio
->devs
[1].devnum
;
2115 struct md_rdev
*rdev
;
2119 if (s
> (PAGE_SIZE
>>9))
2122 rdev
= conf
->mirrors
[dr
].rdev
;
2123 addr
= r10_bio
->devs
[0].addr
+ sect
,
2124 ok
= sync_page_io(rdev
,
2127 bio
->bi_io_vec
[idx
].bv_page
,
2130 rdev
= conf
->mirrors
[dw
].rdev
;
2131 addr
= r10_bio
->devs
[1].addr
+ sect
;
2132 ok
= sync_page_io(rdev
,
2135 bio
->bi_io_vec
[idx
].bv_page
,
2138 set_bit(WriteErrorSeen
, &rdev
->flags
);
2139 if (!test_and_set_bit(WantReplacement
,
2141 set_bit(MD_RECOVERY_NEEDED
,
2142 &rdev
->mddev
->recovery
);
2146 /* We don't worry if we cannot set a bad block -
2147 * it really is bad so there is no loss in not
2150 rdev_set_badblocks(rdev
, addr
, s
, 0);
2152 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2153 /* need bad block on destination too */
2154 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2155 addr
= r10_bio
->devs
[1].addr
+ sect
;
2156 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2158 /* just abort the recovery */
2160 "md/raid10:%s: recovery aborted"
2161 " due to read error\n",
2164 conf
->mirrors
[dw
].recovery_disabled
2165 = mddev
->recovery_disabled
;
2166 set_bit(MD_RECOVERY_INTR
,
2179 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2181 struct r10conf
*conf
= mddev
->private;
2183 struct bio
*wbio
, *wbio2
;
2185 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2186 fix_recovery_read_error(r10_bio
);
2187 end_sync_request(r10_bio
);
2192 * share the pages with the first bio
2193 * and submit the write request
2195 d
= r10_bio
->devs
[1].devnum
;
2196 wbio
= r10_bio
->devs
[1].bio
;
2197 wbio2
= r10_bio
->devs
[1].repl_bio
;
2198 if (wbio
->bi_end_io
) {
2199 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2200 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
2201 generic_make_request(wbio
);
2203 if (wbio2
&& wbio2
->bi_end_io
) {
2204 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2205 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2206 wbio2
->bi_size
>> 9);
2207 generic_make_request(wbio2
);
2213 * Used by fix_read_error() to decay the per rdev read_errors.
2214 * We halve the read error count for every hour that has elapsed
2215 * since the last recorded read error.
2218 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2220 struct timespec cur_time_mon
;
2221 unsigned long hours_since_last
;
2222 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2224 ktime_get_ts(&cur_time_mon
);
2226 if (rdev
->last_read_error
.tv_sec
== 0 &&
2227 rdev
->last_read_error
.tv_nsec
== 0) {
2228 /* first time we've seen a read error */
2229 rdev
->last_read_error
= cur_time_mon
;
2233 hours_since_last
= (cur_time_mon
.tv_sec
-
2234 rdev
->last_read_error
.tv_sec
) / 3600;
2236 rdev
->last_read_error
= cur_time_mon
;
2239 * if hours_since_last is > the number of bits in read_errors
2240 * just set read errors to 0. We do this to avoid
2241 * overflowing the shift of read_errors by hours_since_last.
2243 if (hours_since_last
>= 8 * sizeof(read_errors
))
2244 atomic_set(&rdev
->read_errors
, 0);
2246 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2249 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2250 int sectors
, struct page
*page
, int rw
)
2255 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2256 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2258 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2262 set_bit(WriteErrorSeen
, &rdev
->flags
);
2263 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2264 set_bit(MD_RECOVERY_NEEDED
,
2265 &rdev
->mddev
->recovery
);
2267 /* need to record an error - either for the block or the device */
2268 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2269 md_error(rdev
->mddev
, rdev
);
2274 * This is a kernel thread which:
2276 * 1. Retries failed read operations on working mirrors.
2277 * 2. Updates the raid superblock when problems encounter.
2278 * 3. Performs writes following reads for array synchronising.
2281 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2283 int sect
= 0; /* Offset from r10_bio->sector */
2284 int sectors
= r10_bio
->sectors
;
2285 struct md_rdev
*rdev
;
2286 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2287 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2289 /* still own a reference to this rdev, so it cannot
2290 * have been cleared recently.
2292 rdev
= conf
->mirrors
[d
].rdev
;
2294 if (test_bit(Faulty
, &rdev
->flags
))
2295 /* drive has already been failed, just ignore any
2296 more fix_read_error() attempts */
2299 check_decay_read_errors(mddev
, rdev
);
2300 atomic_inc(&rdev
->read_errors
);
2301 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2302 char b
[BDEVNAME_SIZE
];
2303 bdevname(rdev
->bdev
, b
);
2306 "md/raid10:%s: %s: Raid device exceeded "
2307 "read_error threshold [cur %d:max %d]\n",
2309 atomic_read(&rdev
->read_errors
), max_read_errors
);
2311 "md/raid10:%s: %s: Failing raid device\n",
2313 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2314 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2320 int sl
= r10_bio
->read_slot
;
2324 if (s
> (PAGE_SIZE
>>9))
2332 d
= r10_bio
->devs
[sl
].devnum
;
2333 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2335 !test_bit(Unmerged
, &rdev
->flags
) &&
2336 test_bit(In_sync
, &rdev
->flags
) &&
2337 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2338 &first_bad
, &bad_sectors
) == 0) {
2339 atomic_inc(&rdev
->nr_pending
);
2341 success
= sync_page_io(rdev
,
2342 r10_bio
->devs
[sl
].addr
+
2345 conf
->tmppage
, READ
, false);
2346 rdev_dec_pending(rdev
, mddev
);
2352 if (sl
== conf
->copies
)
2354 } while (!success
&& sl
!= r10_bio
->read_slot
);
2358 /* Cannot read from anywhere, just mark the block
2359 * as bad on the first device to discourage future
2362 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2363 rdev
= conf
->mirrors
[dn
].rdev
;
2365 if (!rdev_set_badblocks(
2367 r10_bio
->devs
[r10_bio
->read_slot
].addr
2370 md_error(mddev
, rdev
);
2371 r10_bio
->devs
[r10_bio
->read_slot
].bio
2378 /* write it back and re-read */
2380 while (sl
!= r10_bio
->read_slot
) {
2381 char b
[BDEVNAME_SIZE
];
2386 d
= r10_bio
->devs
[sl
].devnum
;
2387 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2389 test_bit(Unmerged
, &rdev
->flags
) ||
2390 !test_bit(In_sync
, &rdev
->flags
))
2393 atomic_inc(&rdev
->nr_pending
);
2395 if (r10_sync_page_io(rdev
,
2396 r10_bio
->devs
[sl
].addr
+
2398 s
, conf
->tmppage
, WRITE
)
2400 /* Well, this device is dead */
2402 "md/raid10:%s: read correction "
2404 " (%d sectors at %llu on %s)\n",
2406 (unsigned long long)(
2408 choose_data_offset(r10_bio
,
2410 bdevname(rdev
->bdev
, b
));
2411 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2414 bdevname(rdev
->bdev
, b
));
2416 rdev_dec_pending(rdev
, mddev
);
2420 while (sl
!= r10_bio
->read_slot
) {
2421 char b
[BDEVNAME_SIZE
];
2426 d
= r10_bio
->devs
[sl
].devnum
;
2427 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2429 !test_bit(In_sync
, &rdev
->flags
))
2432 atomic_inc(&rdev
->nr_pending
);
2434 switch (r10_sync_page_io(rdev
,
2435 r10_bio
->devs
[sl
].addr
+
2440 /* Well, this device is dead */
2442 "md/raid10:%s: unable to read back "
2444 " (%d sectors at %llu on %s)\n",
2446 (unsigned long long)(
2448 choose_data_offset(r10_bio
, rdev
)),
2449 bdevname(rdev
->bdev
, b
));
2450 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2453 bdevname(rdev
->bdev
, b
));
2457 "md/raid10:%s: read error corrected"
2458 " (%d sectors at %llu on %s)\n",
2460 (unsigned long long)(
2462 choose_data_offset(r10_bio
, rdev
)),
2463 bdevname(rdev
->bdev
, b
));
2464 atomic_add(s
, &rdev
->corrected_errors
);
2467 rdev_dec_pending(rdev
, mddev
);
2477 static void bi_complete(struct bio
*bio
, int error
)
2479 complete((struct completion
*)bio
->bi_private
);
2482 static int submit_bio_wait(int rw
, struct bio
*bio
)
2484 struct completion event
;
2487 init_completion(&event
);
2488 bio
->bi_private
= &event
;
2489 bio
->bi_end_io
= bi_complete
;
2490 submit_bio(rw
, bio
);
2491 wait_for_completion(&event
);
2493 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2496 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2498 struct bio
*bio
= r10_bio
->master_bio
;
2499 struct mddev
*mddev
= r10_bio
->mddev
;
2500 struct r10conf
*conf
= mddev
->private;
2501 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2502 /* bio has the data to be written to slot 'i' where
2503 * we just recently had a write error.
2504 * We repeatedly clone the bio and trim down to one block,
2505 * then try the write. Where the write fails we record
2507 * It is conceivable that the bio doesn't exactly align with
2508 * blocks. We must handle this.
2510 * We currently own a reference to the rdev.
2516 int sect_to_write
= r10_bio
->sectors
;
2519 if (rdev
->badblocks
.shift
< 0)
2522 block_sectors
= 1 << rdev
->badblocks
.shift
;
2523 sector
= r10_bio
->sector
;
2524 sectors
= ((r10_bio
->sector
+ block_sectors
)
2525 & ~(sector_t
)(block_sectors
- 1))
2528 while (sect_to_write
) {
2530 if (sectors
> sect_to_write
)
2531 sectors
= sect_to_write
;
2532 /* Write at 'sector' for 'sectors' */
2533 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2534 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2535 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2536 choose_data_offset(r10_bio
, rdev
) +
2537 (sector
- r10_bio
->sector
));
2538 wbio
->bi_bdev
= rdev
->bdev
;
2539 if (submit_bio_wait(WRITE
, wbio
) == 0)
2541 ok
= rdev_set_badblocks(rdev
, sector
,
2546 sect_to_write
-= sectors
;
2548 sectors
= block_sectors
;
2553 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2555 int slot
= r10_bio
->read_slot
;
2557 struct r10conf
*conf
= mddev
->private;
2558 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2559 char b
[BDEVNAME_SIZE
];
2560 unsigned long do_sync
;
2563 /* we got a read error. Maybe the drive is bad. Maybe just
2564 * the block and we can fix it.
2565 * We freeze all other IO, and try reading the block from
2566 * other devices. When we find one, we re-write
2567 * and check it that fixes the read error.
2568 * This is all done synchronously while the array is
2571 bio
= r10_bio
->devs
[slot
].bio
;
2572 bdevname(bio
->bi_bdev
, b
);
2574 r10_bio
->devs
[slot
].bio
= NULL
;
2576 if (mddev
->ro
== 0) {
2578 fix_read_error(conf
, mddev
, r10_bio
);
2579 unfreeze_array(conf
);
2581 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2583 rdev_dec_pending(rdev
, mddev
);
2586 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2588 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2589 " read error for block %llu\n",
2591 (unsigned long long)r10_bio
->sector
);
2592 raid_end_bio_io(r10_bio
);
2596 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2597 slot
= r10_bio
->read_slot
;
2600 "md/raid10:%s: %s: redirecting "
2601 "sector %llu to another mirror\n",
2603 bdevname(rdev
->bdev
, b
),
2604 (unsigned long long)r10_bio
->sector
);
2605 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2608 r10_bio
->sector
- bio
->bi_sector
,
2610 r10_bio
->devs
[slot
].bio
= bio
;
2611 r10_bio
->devs
[slot
].rdev
= rdev
;
2612 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2613 + choose_data_offset(r10_bio
, rdev
);
2614 bio
->bi_bdev
= rdev
->bdev
;
2615 bio
->bi_rw
= READ
| do_sync
;
2616 bio
->bi_private
= r10_bio
;
2617 bio
->bi_end_io
= raid10_end_read_request
;
2618 if (max_sectors
< r10_bio
->sectors
) {
2619 /* Drat - have to split this up more */
2620 struct bio
*mbio
= r10_bio
->master_bio
;
2621 int sectors_handled
=
2622 r10_bio
->sector
+ max_sectors
2624 r10_bio
->sectors
= max_sectors
;
2625 spin_lock_irq(&conf
->device_lock
);
2626 if (mbio
->bi_phys_segments
== 0)
2627 mbio
->bi_phys_segments
= 2;
2629 mbio
->bi_phys_segments
++;
2630 spin_unlock_irq(&conf
->device_lock
);
2631 generic_make_request(bio
);
2633 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2635 r10_bio
->master_bio
= mbio
;
2636 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2639 set_bit(R10BIO_ReadError
,
2641 r10_bio
->mddev
= mddev
;
2642 r10_bio
->sector
= mbio
->bi_sector
2647 generic_make_request(bio
);
2650 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2652 /* Some sort of write request has finished and it
2653 * succeeded in writing where we thought there was a
2654 * bad block. So forget the bad block.
2655 * Or possibly if failed and we need to record
2659 struct md_rdev
*rdev
;
2661 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2662 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2663 for (m
= 0; m
< conf
->copies
; m
++) {
2664 int dev
= r10_bio
->devs
[m
].devnum
;
2665 rdev
= conf
->mirrors
[dev
].rdev
;
2666 if (r10_bio
->devs
[m
].bio
== NULL
)
2668 if (test_bit(BIO_UPTODATE
,
2669 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2670 rdev_clear_badblocks(
2672 r10_bio
->devs
[m
].addr
,
2673 r10_bio
->sectors
, 0);
2675 if (!rdev_set_badblocks(
2677 r10_bio
->devs
[m
].addr
,
2678 r10_bio
->sectors
, 0))
2679 md_error(conf
->mddev
, rdev
);
2681 rdev
= conf
->mirrors
[dev
].replacement
;
2682 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2684 if (test_bit(BIO_UPTODATE
,
2685 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2686 rdev_clear_badblocks(
2688 r10_bio
->devs
[m
].addr
,
2689 r10_bio
->sectors
, 0);
2691 if (!rdev_set_badblocks(
2693 r10_bio
->devs
[m
].addr
,
2694 r10_bio
->sectors
, 0))
2695 md_error(conf
->mddev
, rdev
);
2700 for (m
= 0; m
< conf
->copies
; m
++) {
2701 int dev
= r10_bio
->devs
[m
].devnum
;
2702 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2703 rdev
= conf
->mirrors
[dev
].rdev
;
2704 if (bio
== IO_MADE_GOOD
) {
2705 rdev_clear_badblocks(
2707 r10_bio
->devs
[m
].addr
,
2708 r10_bio
->sectors
, 0);
2709 rdev_dec_pending(rdev
, conf
->mddev
);
2710 } else if (bio
!= NULL
&&
2711 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2712 if (!narrow_write_error(r10_bio
, m
)) {
2713 md_error(conf
->mddev
, rdev
);
2714 set_bit(R10BIO_Degraded
,
2717 rdev_dec_pending(rdev
, conf
->mddev
);
2719 bio
= r10_bio
->devs
[m
].repl_bio
;
2720 rdev
= conf
->mirrors
[dev
].replacement
;
2721 if (rdev
&& bio
== IO_MADE_GOOD
) {
2722 rdev_clear_badblocks(
2724 r10_bio
->devs
[m
].addr
,
2725 r10_bio
->sectors
, 0);
2726 rdev_dec_pending(rdev
, conf
->mddev
);
2729 if (test_bit(R10BIO_WriteError
,
2731 close_write(r10_bio
);
2732 raid_end_bio_io(r10_bio
);
2736 static void raid10d(struct md_thread
*thread
)
2738 struct mddev
*mddev
= thread
->mddev
;
2739 struct r10bio
*r10_bio
;
2740 unsigned long flags
;
2741 struct r10conf
*conf
= mddev
->private;
2742 struct list_head
*head
= &conf
->retry_list
;
2743 struct blk_plug plug
;
2745 md_check_recovery(mddev
);
2747 blk_start_plug(&plug
);
2750 flush_pending_writes(conf
);
2752 spin_lock_irqsave(&conf
->device_lock
, flags
);
2753 if (list_empty(head
)) {
2754 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2757 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2758 list_del(head
->prev
);
2760 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2762 mddev
= r10_bio
->mddev
;
2763 conf
= mddev
->private;
2764 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2765 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2766 handle_write_completed(conf
, r10_bio
);
2767 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2768 reshape_request_write(mddev
, r10_bio
);
2769 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2770 sync_request_write(mddev
, r10_bio
);
2771 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2772 recovery_request_write(mddev
, r10_bio
);
2773 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2774 handle_read_error(mddev
, r10_bio
);
2776 /* just a partial read to be scheduled from a
2779 int slot
= r10_bio
->read_slot
;
2780 generic_make_request(r10_bio
->devs
[slot
].bio
);
2784 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2785 md_check_recovery(mddev
);
2787 blk_finish_plug(&plug
);
2791 static int init_resync(struct r10conf
*conf
)
2796 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2797 BUG_ON(conf
->r10buf_pool
);
2798 conf
->have_replacement
= 0;
2799 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2800 if (conf
->mirrors
[i
].replacement
)
2801 conf
->have_replacement
= 1;
2802 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2803 if (!conf
->r10buf_pool
)
2805 conf
->next_resync
= 0;
2810 * perform a "sync" on one "block"
2812 * We need to make sure that no normal I/O request - particularly write
2813 * requests - conflict with active sync requests.
2815 * This is achieved by tracking pending requests and a 'barrier' concept
2816 * that can be installed to exclude normal IO requests.
2818 * Resync and recovery are handled very differently.
2819 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2821 * For resync, we iterate over virtual addresses, read all copies,
2822 * and update if there are differences. If only one copy is live,
2824 * For recovery, we iterate over physical addresses, read a good
2825 * value for each non-in_sync drive, and over-write.
2827 * So, for recovery we may have several outstanding complex requests for a
2828 * given address, one for each out-of-sync device. We model this by allocating
2829 * a number of r10_bio structures, one for each out-of-sync device.
2830 * As we setup these structures, we collect all bio's together into a list
2831 * which we then process collectively to add pages, and then process again
2832 * to pass to generic_make_request.
2834 * The r10_bio structures are linked using a borrowed master_bio pointer.
2835 * This link is counted in ->remaining. When the r10_bio that points to NULL
2836 * has its remaining count decremented to 0, the whole complex operation
2841 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2842 int *skipped
, int go_faster
)
2844 struct r10conf
*conf
= mddev
->private;
2845 struct r10bio
*r10_bio
;
2846 struct bio
*biolist
= NULL
, *bio
;
2847 sector_t max_sector
, nr_sectors
;
2850 sector_t sync_blocks
;
2851 sector_t sectors_skipped
= 0;
2852 int chunks_skipped
= 0;
2853 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2855 if (!conf
->r10buf_pool
)
2856 if (init_resync(conf
))
2860 max_sector
= mddev
->dev_sectors
;
2861 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2862 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2863 max_sector
= mddev
->resync_max_sectors
;
2864 if (sector_nr
>= max_sector
) {
2865 /* If we aborted, we need to abort the
2866 * sync on the 'current' bitmap chucks (there can
2867 * be several when recovering multiple devices).
2868 * as we may have started syncing it but not finished.
2869 * We can find the current address in
2870 * mddev->curr_resync, but for recovery,
2871 * we need to convert that to several
2872 * virtual addresses.
2874 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2879 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2880 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2881 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2883 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2885 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2886 bitmap_end_sync(mddev
->bitmap
, sect
,
2890 /* completed sync */
2891 if ((!mddev
->bitmap
|| conf
->fullsync
)
2892 && conf
->have_replacement
2893 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2894 /* Completed a full sync so the replacements
2895 * are now fully recovered.
2897 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2898 if (conf
->mirrors
[i
].replacement
)
2899 conf
->mirrors
[i
].replacement
2905 bitmap_close_sync(mddev
->bitmap
);
2908 return sectors_skipped
;
2911 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2912 return reshape_request(mddev
, sector_nr
, skipped
);
2914 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2915 /* if there has been nothing to do on any drive,
2916 * then there is nothing to do at all..
2919 return (max_sector
- sector_nr
) + sectors_skipped
;
2922 if (max_sector
> mddev
->resync_max
)
2923 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2925 /* make sure whole request will fit in a chunk - if chunks
2928 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2929 max_sector
> (sector_nr
| chunk_mask
))
2930 max_sector
= (sector_nr
| chunk_mask
) + 1;
2932 * If there is non-resync activity waiting for us then
2933 * put in a delay to throttle resync.
2935 if (!go_faster
&& conf
->nr_waiting
)
2936 msleep_interruptible(1000);
2938 /* Again, very different code for resync and recovery.
2939 * Both must result in an r10bio with a list of bios that
2940 * have bi_end_io, bi_sector, bi_bdev set,
2941 * and bi_private set to the r10bio.
2942 * For recovery, we may actually create several r10bios
2943 * with 2 bios in each, that correspond to the bios in the main one.
2944 * In this case, the subordinate r10bios link back through a
2945 * borrowed master_bio pointer, and the counter in the master
2946 * includes a ref from each subordinate.
2948 /* First, we decide what to do and set ->bi_end_io
2949 * To end_sync_read if we want to read, and
2950 * end_sync_write if we will want to write.
2953 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2954 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2955 /* recovery... the complicated one */
2959 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2965 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2967 if ((mirror
->rdev
== NULL
||
2968 test_bit(In_sync
, &mirror
->rdev
->flags
))
2970 (mirror
->replacement
== NULL
||
2972 &mirror
->replacement
->flags
)))
2976 /* want to reconstruct this device */
2978 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2979 if (sect
>= mddev
->resync_max_sectors
) {
2980 /* last stripe is not complete - don't
2981 * try to recover this sector.
2985 /* Unless we are doing a full sync, or a replacement
2986 * we only need to recover the block if it is set in
2989 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2991 if (sync_blocks
< max_sync
)
2992 max_sync
= sync_blocks
;
2994 mirror
->replacement
== NULL
&&
2996 /* yep, skip the sync_blocks here, but don't assume
2997 * that there will never be anything to do here
2999 chunks_skipped
= -1;
3003 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3004 raise_barrier(conf
, rb2
!= NULL
);
3005 atomic_set(&r10_bio
->remaining
, 0);
3007 r10_bio
->master_bio
= (struct bio
*)rb2
;
3009 atomic_inc(&rb2
->remaining
);
3010 r10_bio
->mddev
= mddev
;
3011 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3012 r10_bio
->sector
= sect
;
3014 raid10_find_phys(conf
, r10_bio
);
3016 /* Need to check if the array will still be
3019 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3020 if (conf
->mirrors
[j
].rdev
== NULL
||
3021 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3026 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3027 &sync_blocks
, still_degraded
);
3030 for (j
=0; j
<conf
->copies
;j
++) {
3032 int d
= r10_bio
->devs
[j
].devnum
;
3033 sector_t from_addr
, to_addr
;
3034 struct md_rdev
*rdev
;
3035 sector_t sector
, first_bad
;
3037 if (!conf
->mirrors
[d
].rdev
||
3038 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3040 /* This is where we read from */
3042 rdev
= conf
->mirrors
[d
].rdev
;
3043 sector
= r10_bio
->devs
[j
].addr
;
3045 if (is_badblock(rdev
, sector
, max_sync
,
3046 &first_bad
, &bad_sectors
)) {
3047 if (first_bad
> sector
)
3048 max_sync
= first_bad
- sector
;
3050 bad_sectors
-= (sector
3052 if (max_sync
> bad_sectors
)
3053 max_sync
= bad_sectors
;
3057 bio
= r10_bio
->devs
[0].bio
;
3058 bio
->bi_next
= biolist
;
3060 bio
->bi_private
= r10_bio
;
3061 bio
->bi_end_io
= end_sync_read
;
3063 from_addr
= r10_bio
->devs
[j
].addr
;
3064 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
3065 bio
->bi_bdev
= rdev
->bdev
;
3066 atomic_inc(&rdev
->nr_pending
);
3067 /* and we write to 'i' (if not in_sync) */
3069 for (k
=0; k
<conf
->copies
; k
++)
3070 if (r10_bio
->devs
[k
].devnum
== i
)
3072 BUG_ON(k
== conf
->copies
);
3073 to_addr
= r10_bio
->devs
[k
].addr
;
3074 r10_bio
->devs
[0].devnum
= d
;
3075 r10_bio
->devs
[0].addr
= from_addr
;
3076 r10_bio
->devs
[1].devnum
= i
;
3077 r10_bio
->devs
[1].addr
= to_addr
;
3079 rdev
= mirror
->rdev
;
3080 if (!test_bit(In_sync
, &rdev
->flags
)) {
3081 bio
= r10_bio
->devs
[1].bio
;
3082 bio
->bi_next
= biolist
;
3084 bio
->bi_private
= r10_bio
;
3085 bio
->bi_end_io
= end_sync_write
;
3087 bio
->bi_sector
= to_addr
3088 + rdev
->data_offset
;
3089 bio
->bi_bdev
= rdev
->bdev
;
3090 atomic_inc(&r10_bio
->remaining
);
3092 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3094 /* and maybe write to replacement */
3095 bio
= r10_bio
->devs
[1].repl_bio
;
3097 bio
->bi_end_io
= NULL
;
3098 rdev
= mirror
->replacement
;
3099 /* Note: if rdev != NULL, then bio
3100 * cannot be NULL as r10buf_pool_alloc will
3101 * have allocated it.
3102 * So the second test here is pointless.
3103 * But it keeps semantic-checkers happy, and
3104 * this comment keeps human reviewers
3107 if (rdev
== NULL
|| bio
== NULL
||
3108 test_bit(Faulty
, &rdev
->flags
))
3110 bio
->bi_next
= biolist
;
3112 bio
->bi_private
= r10_bio
;
3113 bio
->bi_end_io
= end_sync_write
;
3115 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3116 bio
->bi_bdev
= rdev
->bdev
;
3117 atomic_inc(&r10_bio
->remaining
);
3120 if (j
== conf
->copies
) {
3121 /* Cannot recover, so abort the recovery or
3122 * record a bad block */
3125 atomic_dec(&rb2
->remaining
);
3128 /* problem is that there are bad blocks
3129 * on other device(s)
3132 for (k
= 0; k
< conf
->copies
; k
++)
3133 if (r10_bio
->devs
[k
].devnum
== i
)
3135 if (!test_bit(In_sync
,
3136 &mirror
->rdev
->flags
)
3137 && !rdev_set_badblocks(
3139 r10_bio
->devs
[k
].addr
,
3142 if (mirror
->replacement
&&
3143 !rdev_set_badblocks(
3144 mirror
->replacement
,
3145 r10_bio
->devs
[k
].addr
,
3150 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3152 printk(KERN_INFO
"md/raid10:%s: insufficient "
3153 "working devices for recovery.\n",
3155 mirror
->recovery_disabled
3156 = mddev
->recovery_disabled
;
3161 if (biolist
== NULL
) {
3163 struct r10bio
*rb2
= r10_bio
;
3164 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3165 rb2
->master_bio
= NULL
;
3171 /* resync. Schedule a read for every block at this virt offset */
3174 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3176 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3177 &sync_blocks
, mddev
->degraded
) &&
3178 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3179 &mddev
->recovery
)) {
3180 /* We can skip this block */
3182 return sync_blocks
+ sectors_skipped
;
3184 if (sync_blocks
< max_sync
)
3185 max_sync
= sync_blocks
;
3186 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3188 r10_bio
->mddev
= mddev
;
3189 atomic_set(&r10_bio
->remaining
, 0);
3190 raise_barrier(conf
, 0);
3191 conf
->next_resync
= sector_nr
;
3193 r10_bio
->master_bio
= NULL
;
3194 r10_bio
->sector
= sector_nr
;
3195 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3196 raid10_find_phys(conf
, r10_bio
);
3197 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3199 for (i
= 0; i
< conf
->copies
; i
++) {
3200 int d
= r10_bio
->devs
[i
].devnum
;
3201 sector_t first_bad
, sector
;
3204 if (r10_bio
->devs
[i
].repl_bio
)
3205 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3207 bio
= r10_bio
->devs
[i
].bio
;
3208 bio
->bi_end_io
= NULL
;
3209 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3210 if (conf
->mirrors
[d
].rdev
== NULL
||
3211 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3213 sector
= r10_bio
->devs
[i
].addr
;
3214 if (is_badblock(conf
->mirrors
[d
].rdev
,
3216 &first_bad
, &bad_sectors
)) {
3217 if (first_bad
> sector
)
3218 max_sync
= first_bad
- sector
;
3220 bad_sectors
-= (sector
- first_bad
);
3221 if (max_sync
> bad_sectors
)
3222 max_sync
= bad_sectors
;
3226 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3227 atomic_inc(&r10_bio
->remaining
);
3228 bio
->bi_next
= biolist
;
3230 bio
->bi_private
= r10_bio
;
3231 bio
->bi_end_io
= end_sync_read
;
3233 bio
->bi_sector
= sector
+
3234 conf
->mirrors
[d
].rdev
->data_offset
;
3235 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3238 if (conf
->mirrors
[d
].replacement
== NULL
||
3240 &conf
->mirrors
[d
].replacement
->flags
))
3243 /* Need to set up for writing to the replacement */
3244 bio
= r10_bio
->devs
[i
].repl_bio
;
3245 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3247 sector
= r10_bio
->devs
[i
].addr
;
3248 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3249 bio
->bi_next
= biolist
;
3251 bio
->bi_private
= r10_bio
;
3252 bio
->bi_end_io
= end_sync_write
;
3254 bio
->bi_sector
= sector
+
3255 conf
->mirrors
[d
].replacement
->data_offset
;
3256 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3261 for (i
=0; i
<conf
->copies
; i
++) {
3262 int d
= r10_bio
->devs
[i
].devnum
;
3263 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3264 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3266 if (r10_bio
->devs
[i
].repl_bio
&&
3267 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3269 conf
->mirrors
[d
].replacement
,
3278 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3280 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
3282 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
3285 bio
->bi_phys_segments
= 0;
3290 if (sector_nr
+ max_sync
< max_sector
)
3291 max_sector
= sector_nr
+ max_sync
;
3294 int len
= PAGE_SIZE
;
3295 if (sector_nr
+ (len
>>9) > max_sector
)
3296 len
= (max_sector
- sector_nr
) << 9;
3299 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3301 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3302 if (bio_add_page(bio
, page
, len
, 0))
3306 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3307 for (bio2
= biolist
;
3308 bio2
&& bio2
!= bio
;
3309 bio2
= bio2
->bi_next
) {
3310 /* remove last page from this bio */
3312 bio2
->bi_size
-= len
;
3313 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3317 nr_sectors
+= len
>>9;
3318 sector_nr
+= len
>>9;
3319 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3321 r10_bio
->sectors
= nr_sectors
;
3325 biolist
= biolist
->bi_next
;
3327 bio
->bi_next
= NULL
;
3328 r10_bio
= bio
->bi_private
;
3329 r10_bio
->sectors
= nr_sectors
;
3331 if (bio
->bi_end_io
== end_sync_read
) {
3332 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3333 generic_make_request(bio
);
3337 if (sectors_skipped
)
3338 /* pretend they weren't skipped, it makes
3339 * no important difference in this case
3341 md_done_sync(mddev
, sectors_skipped
, 1);
3343 return sectors_skipped
+ nr_sectors
;
3345 /* There is nowhere to write, so all non-sync
3346 * drives must be failed or in resync, all drives
3347 * have a bad block, so try the next chunk...
3349 if (sector_nr
+ max_sync
< max_sector
)
3350 max_sector
= sector_nr
+ max_sync
;
3352 sectors_skipped
+= (max_sector
- sector_nr
);
3354 sector_nr
= max_sector
;
3359 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3362 struct r10conf
*conf
= mddev
->private;
3365 raid_disks
= min(conf
->geo
.raid_disks
,
3366 conf
->prev
.raid_disks
);
3368 sectors
= conf
->dev_sectors
;
3370 size
= sectors
>> conf
->geo
.chunk_shift
;
3371 sector_div(size
, conf
->geo
.far_copies
);
3372 size
= size
* raid_disks
;
3373 sector_div(size
, conf
->geo
.near_copies
);
3375 return size
<< conf
->geo
.chunk_shift
;
3378 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3380 /* Calculate the number of sectors-per-device that will
3381 * actually be used, and set conf->dev_sectors and
3385 size
= size
>> conf
->geo
.chunk_shift
;
3386 sector_div(size
, conf
->geo
.far_copies
);
3387 size
= size
* conf
->geo
.raid_disks
;
3388 sector_div(size
, conf
->geo
.near_copies
);
3389 /* 'size' is now the number of chunks in the array */
3390 /* calculate "used chunks per device" */
3391 size
= size
* conf
->copies
;
3393 /* We need to round up when dividing by raid_disks to
3394 * get the stride size.
3396 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3398 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3400 if (conf
->geo
.far_offset
)
3401 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3403 sector_div(size
, conf
->geo
.far_copies
);
3404 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3408 enum geo_type
{geo_new
, geo_old
, geo_start
};
3409 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3412 int layout
, chunk
, disks
;
3415 layout
= mddev
->layout
;
3416 chunk
= mddev
->chunk_sectors
;
3417 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3420 layout
= mddev
->new_layout
;
3421 chunk
= mddev
->new_chunk_sectors
;
3422 disks
= mddev
->raid_disks
;
3424 default: /* avoid 'may be unused' warnings */
3425 case geo_start
: /* new when starting reshape - raid_disks not
3427 layout
= mddev
->new_layout
;
3428 chunk
= mddev
->new_chunk_sectors
;
3429 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3434 if (chunk
< (PAGE_SIZE
>> 9) ||
3435 !is_power_of_2(chunk
))
3438 fc
= (layout
>> 8) & 255;
3439 fo
= layout
& (1<<16);
3440 geo
->raid_disks
= disks
;
3441 geo
->near_copies
= nc
;
3442 geo
->far_copies
= fc
;
3443 geo
->far_offset
= fo
;
3444 geo
->chunk_mask
= chunk
- 1;
3445 geo
->chunk_shift
= ffz(~chunk
);
3449 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3451 struct r10conf
*conf
= NULL
;
3456 copies
= setup_geo(&geo
, mddev
, geo_new
);
3459 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3460 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3461 mdname(mddev
), PAGE_SIZE
);
3465 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3466 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3467 mdname(mddev
), mddev
->new_layout
);
3472 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3476 /* FIXME calc properly */
3477 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3478 max(0,mddev
->delta_disks
)),
3483 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3488 conf
->copies
= copies
;
3489 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3490 r10bio_pool_free
, conf
);
3491 if (!conf
->r10bio_pool
)
3494 calc_sectors(conf
, mddev
->dev_sectors
);
3495 if (mddev
->reshape_position
== MaxSector
) {
3496 conf
->prev
= conf
->geo
;
3497 conf
->reshape_progress
= MaxSector
;
3499 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3503 conf
->reshape_progress
= mddev
->reshape_position
;
3504 if (conf
->prev
.far_offset
)
3505 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3507 /* far_copies must be 1 */
3508 conf
->prev
.stride
= conf
->dev_sectors
;
3510 spin_lock_init(&conf
->device_lock
);
3511 INIT_LIST_HEAD(&conf
->retry_list
);
3513 spin_lock_init(&conf
->resync_lock
);
3514 init_waitqueue_head(&conf
->wait_barrier
);
3516 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3520 conf
->mddev
= mddev
;
3525 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3528 if (conf
->r10bio_pool
)
3529 mempool_destroy(conf
->r10bio_pool
);
3530 kfree(conf
->mirrors
);
3531 safe_put_page(conf
->tmppage
);
3534 return ERR_PTR(err
);
3537 static int run(struct mddev
*mddev
)
3539 struct r10conf
*conf
;
3540 int i
, disk_idx
, chunk_size
;
3541 struct raid10_info
*disk
;
3542 struct md_rdev
*rdev
;
3544 sector_t min_offset_diff
= 0;
3546 bool discard_supported
= false;
3548 if (mddev
->private == NULL
) {
3549 conf
= setup_conf(mddev
);
3551 return PTR_ERR(conf
);
3552 mddev
->private = conf
;
3554 conf
= mddev
->private;
3558 mddev
->thread
= conf
->thread
;
3559 conf
->thread
= NULL
;
3561 chunk_size
= mddev
->chunk_sectors
<< 9;
3563 blk_queue_max_discard_sectors(mddev
->queue
,
3564 mddev
->chunk_sectors
);
3565 blk_queue_io_min(mddev
->queue
, chunk_size
);
3566 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3567 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3569 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3570 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3573 rdev_for_each(rdev
, mddev
) {
3575 struct request_queue
*q
;
3577 disk_idx
= rdev
->raid_disk
;
3580 if (disk_idx
>= conf
->geo
.raid_disks
&&
3581 disk_idx
>= conf
->prev
.raid_disks
)
3583 disk
= conf
->mirrors
+ disk_idx
;
3585 if (test_bit(Replacement
, &rdev
->flags
)) {
3586 if (disk
->replacement
)
3588 disk
->replacement
= rdev
;
3594 q
= bdev_get_queue(rdev
->bdev
);
3595 if (q
->merge_bvec_fn
)
3596 mddev
->merge_check_needed
= 1;
3597 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3598 if (!mddev
->reshape_backwards
)
3602 if (first
|| diff
< min_offset_diff
)
3603 min_offset_diff
= diff
;
3606 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3607 rdev
->data_offset
<< 9);
3609 disk
->head_position
= 0;
3611 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3612 discard_supported
= true;
3615 if (discard_supported
)
3616 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
3618 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
3620 /* need to check that every block has at least one working mirror */
3621 if (!enough(conf
, -1)) {
3622 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3627 if (conf
->reshape_progress
!= MaxSector
) {
3628 /* must ensure that shape change is supported */
3629 if (conf
->geo
.far_copies
!= 1 &&
3630 conf
->geo
.far_offset
== 0)
3632 if (conf
->prev
.far_copies
!= 1 &&
3633 conf
->geo
.far_offset
== 0)
3637 mddev
->degraded
= 0;
3639 i
< conf
->geo
.raid_disks
3640 || i
< conf
->prev
.raid_disks
;
3643 disk
= conf
->mirrors
+ i
;
3645 if (!disk
->rdev
&& disk
->replacement
) {
3646 /* The replacement is all we have - use it */
3647 disk
->rdev
= disk
->replacement
;
3648 disk
->replacement
= NULL
;
3649 clear_bit(Replacement
, &disk
->rdev
->flags
);
3653 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3654 disk
->head_position
= 0;
3659 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3662 if (mddev
->recovery_cp
!= MaxSector
)
3663 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3664 " -- starting background reconstruction\n",
3667 "md/raid10:%s: active with %d out of %d devices\n",
3668 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3669 conf
->geo
.raid_disks
);
3671 * Ok, everything is just fine now
3673 mddev
->dev_sectors
= conf
->dev_sectors
;
3674 size
= raid10_size(mddev
, 0, 0);
3675 md_set_array_sectors(mddev
, size
);
3676 mddev
->resync_max_sectors
= size
;
3679 int stripe
= conf
->geo
.raid_disks
*
3680 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3681 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3682 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3684 /* Calculate max read-ahead size.
3685 * We need to readahead at least twice a whole stripe....
3688 stripe
/= conf
->geo
.near_copies
;
3689 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3690 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3691 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3695 if (md_integrity_register(mddev
))
3698 if (conf
->reshape_progress
!= MaxSector
) {
3699 unsigned long before_length
, after_length
;
3701 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3702 conf
->prev
.far_copies
);
3703 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3704 conf
->geo
.far_copies
);
3706 if (max(before_length
, after_length
) > min_offset_diff
) {
3707 /* This cannot work */
3708 printk("md/raid10: offset difference not enough to continue reshape\n");
3711 conf
->offset_diff
= min_offset_diff
;
3713 conf
->reshape_safe
= conf
->reshape_progress
;
3714 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3715 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3716 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3717 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3718 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3725 md_unregister_thread(&mddev
->thread
);
3726 if (conf
->r10bio_pool
)
3727 mempool_destroy(conf
->r10bio_pool
);
3728 safe_put_page(conf
->tmppage
);
3729 kfree(conf
->mirrors
);
3731 mddev
->private = NULL
;
3736 static int stop(struct mddev
*mddev
)
3738 struct r10conf
*conf
= mddev
->private;
3740 raise_barrier(conf
, 0);
3741 lower_barrier(conf
);
3743 md_unregister_thread(&mddev
->thread
);
3745 /* the unplug fn references 'conf'*/
3746 blk_sync_queue(mddev
->queue
);
3748 if (conf
->r10bio_pool
)
3749 mempool_destroy(conf
->r10bio_pool
);
3750 kfree(conf
->mirrors
);
3752 mddev
->private = NULL
;
3756 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3758 struct r10conf
*conf
= mddev
->private;
3762 raise_barrier(conf
, 0);
3765 lower_barrier(conf
);
3770 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3772 /* Resize of 'far' arrays is not supported.
3773 * For 'near' and 'offset' arrays we can set the
3774 * number of sectors used to be an appropriate multiple
3775 * of the chunk size.
3776 * For 'offset', this is far_copies*chunksize.
3777 * For 'near' the multiplier is the LCM of
3778 * near_copies and raid_disks.
3779 * So if far_copies > 1 && !far_offset, fail.
3780 * Else find LCM(raid_disks, near_copy)*far_copies and
3781 * multiply by chunk_size. Then round to this number.
3782 * This is mostly done by raid10_size()
3784 struct r10conf
*conf
= mddev
->private;
3785 sector_t oldsize
, size
;
3787 if (mddev
->reshape_position
!= MaxSector
)
3790 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3793 oldsize
= raid10_size(mddev
, 0, 0);
3794 size
= raid10_size(mddev
, sectors
, 0);
3795 if (mddev
->external_size
&&
3796 mddev
->array_sectors
> size
)
3798 if (mddev
->bitmap
) {
3799 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3803 md_set_array_sectors(mddev
, size
);
3804 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3805 revalidate_disk(mddev
->gendisk
);
3806 if (sectors
> mddev
->dev_sectors
&&
3807 mddev
->recovery_cp
> oldsize
) {
3808 mddev
->recovery_cp
= oldsize
;
3809 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3811 calc_sectors(conf
, sectors
);
3812 mddev
->dev_sectors
= conf
->dev_sectors
;
3813 mddev
->resync_max_sectors
= size
;
3817 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3819 struct md_rdev
*rdev
;
3820 struct r10conf
*conf
;
3822 if (mddev
->degraded
> 0) {
3823 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3825 return ERR_PTR(-EINVAL
);
3828 /* Set new parameters */
3829 mddev
->new_level
= 10;
3830 /* new layout: far_copies = 1, near_copies = 2 */
3831 mddev
->new_layout
= (1<<8) + 2;
3832 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3833 mddev
->delta_disks
= mddev
->raid_disks
;
3834 mddev
->raid_disks
*= 2;
3835 /* make sure it will be not marked as dirty */
3836 mddev
->recovery_cp
= MaxSector
;
3838 conf
= setup_conf(mddev
);
3839 if (!IS_ERR(conf
)) {
3840 rdev_for_each(rdev
, mddev
)
3841 if (rdev
->raid_disk
>= 0)
3842 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3849 static void *raid10_takeover(struct mddev
*mddev
)
3851 struct r0conf
*raid0_conf
;
3853 /* raid10 can take over:
3854 * raid0 - providing it has only two drives
3856 if (mddev
->level
== 0) {
3857 /* for raid0 takeover only one zone is supported */
3858 raid0_conf
= mddev
->private;
3859 if (raid0_conf
->nr_strip_zones
> 1) {
3860 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3861 " with more than one zone.\n",
3863 return ERR_PTR(-EINVAL
);
3865 return raid10_takeover_raid0(mddev
);
3867 return ERR_PTR(-EINVAL
);
3870 static int raid10_check_reshape(struct mddev
*mddev
)
3872 /* Called when there is a request to change
3873 * - layout (to ->new_layout)
3874 * - chunk size (to ->new_chunk_sectors)
3875 * - raid_disks (by delta_disks)
3876 * or when trying to restart a reshape that was ongoing.
3878 * We need to validate the request and possibly allocate
3879 * space if that might be an issue later.
3881 * Currently we reject any reshape of a 'far' mode array,
3882 * allow chunk size to change if new is generally acceptable,
3883 * allow raid_disks to increase, and allow
3884 * a switch between 'near' mode and 'offset' mode.
3886 struct r10conf
*conf
= mddev
->private;
3889 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3892 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3893 /* mustn't change number of copies */
3895 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3896 /* Cannot switch to 'far' mode */
3899 if (mddev
->array_sectors
& geo
.chunk_mask
)
3900 /* not factor of array size */
3903 if (!enough(conf
, -1))
3906 kfree(conf
->mirrors_new
);
3907 conf
->mirrors_new
= NULL
;
3908 if (mddev
->delta_disks
> 0) {
3909 /* allocate new 'mirrors' list */
3910 conf
->mirrors_new
= kzalloc(
3911 sizeof(struct raid10_info
)
3912 *(mddev
->raid_disks
+
3913 mddev
->delta_disks
),
3915 if (!conf
->mirrors_new
)
3922 * Need to check if array has failed when deciding whether to:
3924 * - remove non-faulty devices
3927 * This determination is simple when no reshape is happening.
3928 * However if there is a reshape, we need to carefully check
3929 * both the before and after sections.
3930 * This is because some failed devices may only affect one
3931 * of the two sections, and some non-in_sync devices may
3932 * be insync in the section most affected by failed devices.
3934 static int calc_degraded(struct r10conf
*conf
)
3936 int degraded
, degraded2
;
3941 /* 'prev' section first */
3942 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
3943 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3944 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3946 else if (!test_bit(In_sync
, &rdev
->flags
))
3947 /* When we can reduce the number of devices in
3948 * an array, this might not contribute to
3949 * 'degraded'. It does now.
3954 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
3958 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3959 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3960 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3962 else if (!test_bit(In_sync
, &rdev
->flags
)) {
3963 /* If reshape is increasing the number of devices,
3964 * this section has already been recovered, so
3965 * it doesn't contribute to degraded.
3968 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
3973 if (degraded2
> degraded
)
3978 static int raid10_start_reshape(struct mddev
*mddev
)
3980 /* A 'reshape' has been requested. This commits
3981 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3982 * This also checks if there are enough spares and adds them
3984 * We currently require enough spares to make the final
3985 * array non-degraded. We also require that the difference
3986 * between old and new data_offset - on each device - is
3987 * enough that we never risk over-writing.
3990 unsigned long before_length
, after_length
;
3991 sector_t min_offset_diff
= 0;
3994 struct r10conf
*conf
= mddev
->private;
3995 struct md_rdev
*rdev
;
3999 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4002 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4005 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4006 conf
->prev
.far_copies
);
4007 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4008 conf
->geo
.far_copies
);
4010 rdev_for_each(rdev
, mddev
) {
4011 if (!test_bit(In_sync
, &rdev
->flags
)
4012 && !test_bit(Faulty
, &rdev
->flags
))
4014 if (rdev
->raid_disk
>= 0) {
4015 long long diff
= (rdev
->new_data_offset
4016 - rdev
->data_offset
);
4017 if (!mddev
->reshape_backwards
)
4021 if (first
|| diff
< min_offset_diff
)
4022 min_offset_diff
= diff
;
4026 if (max(before_length
, after_length
) > min_offset_diff
)
4029 if (spares
< mddev
->delta_disks
)
4032 conf
->offset_diff
= min_offset_diff
;
4033 spin_lock_irq(&conf
->device_lock
);
4034 if (conf
->mirrors_new
) {
4035 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4036 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4038 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
4039 conf
->mirrors_old
= conf
->mirrors
;
4040 conf
->mirrors
= conf
->mirrors_new
;
4041 conf
->mirrors_new
= NULL
;
4043 setup_geo(&conf
->geo
, mddev
, geo_start
);
4045 if (mddev
->reshape_backwards
) {
4046 sector_t size
= raid10_size(mddev
, 0, 0);
4047 if (size
< mddev
->array_sectors
) {
4048 spin_unlock_irq(&conf
->device_lock
);
4049 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4053 mddev
->resync_max_sectors
= size
;
4054 conf
->reshape_progress
= size
;
4056 conf
->reshape_progress
= 0;
4057 spin_unlock_irq(&conf
->device_lock
);
4059 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4060 ret
= bitmap_resize(mddev
->bitmap
,
4061 raid10_size(mddev
, 0,
4062 conf
->geo
.raid_disks
),
4067 if (mddev
->delta_disks
> 0) {
4068 rdev_for_each(rdev
, mddev
)
4069 if (rdev
->raid_disk
< 0 &&
4070 !test_bit(Faulty
, &rdev
->flags
)) {
4071 if (raid10_add_disk(mddev
, rdev
) == 0) {
4072 if (rdev
->raid_disk
>=
4073 conf
->prev
.raid_disks
)
4074 set_bit(In_sync
, &rdev
->flags
);
4076 rdev
->recovery_offset
= 0;
4078 if (sysfs_link_rdev(mddev
, rdev
))
4079 /* Failure here is OK */;
4081 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4082 && !test_bit(Faulty
, &rdev
->flags
)) {
4083 /* This is a spare that was manually added */
4084 set_bit(In_sync
, &rdev
->flags
);
4087 /* When a reshape changes the number of devices,
4088 * ->degraded is measured against the larger of the
4089 * pre and post numbers.
4091 spin_lock_irq(&conf
->device_lock
);
4092 mddev
->degraded
= calc_degraded(conf
);
4093 spin_unlock_irq(&conf
->device_lock
);
4094 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4095 mddev
->reshape_position
= conf
->reshape_progress
;
4096 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4098 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4099 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4100 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4101 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4103 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4105 if (!mddev
->sync_thread
) {
4109 conf
->reshape_checkpoint
= jiffies
;
4110 md_wakeup_thread(mddev
->sync_thread
);
4111 md_new_event(mddev
);
4115 mddev
->recovery
= 0;
4116 spin_lock_irq(&conf
->device_lock
);
4117 conf
->geo
= conf
->prev
;
4118 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4119 rdev_for_each(rdev
, mddev
)
4120 rdev
->new_data_offset
= rdev
->data_offset
;
4122 conf
->reshape_progress
= MaxSector
;
4123 mddev
->reshape_position
= MaxSector
;
4124 spin_unlock_irq(&conf
->device_lock
);
4128 /* Calculate the last device-address that could contain
4129 * any block from the chunk that includes the array-address 's'
4130 * and report the next address.
4131 * i.e. the address returned will be chunk-aligned and after
4132 * any data that is in the chunk containing 's'.
4134 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4136 s
= (s
| geo
->chunk_mask
) + 1;
4137 s
>>= geo
->chunk_shift
;
4138 s
*= geo
->near_copies
;
4139 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4140 s
*= geo
->far_copies
;
4141 s
<<= geo
->chunk_shift
;
4145 /* Calculate the first device-address that could contain
4146 * any block from the chunk that includes the array-address 's'.
4147 * This too will be the start of a chunk
4149 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4151 s
>>= geo
->chunk_shift
;
4152 s
*= geo
->near_copies
;
4153 sector_div(s
, geo
->raid_disks
);
4154 s
*= geo
->far_copies
;
4155 s
<<= geo
->chunk_shift
;
4159 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4162 /* We simply copy at most one chunk (smallest of old and new)
4163 * at a time, possibly less if that exceeds RESYNC_PAGES,
4164 * or we hit a bad block or something.
4165 * This might mean we pause for normal IO in the middle of
4166 * a chunk, but that is not a problem was mddev->reshape_position
4167 * can record any location.
4169 * If we will want to write to a location that isn't
4170 * yet recorded as 'safe' (i.e. in metadata on disk) then
4171 * we need to flush all reshape requests and update the metadata.
4173 * When reshaping forwards (e.g. to more devices), we interpret
4174 * 'safe' as the earliest block which might not have been copied
4175 * down yet. We divide this by previous stripe size and multiply
4176 * by previous stripe length to get lowest device offset that we
4177 * cannot write to yet.
4178 * We interpret 'sector_nr' as an address that we want to write to.
4179 * From this we use last_device_address() to find where we might
4180 * write to, and first_device_address on the 'safe' position.
4181 * If this 'next' write position is after the 'safe' position,
4182 * we must update the metadata to increase the 'safe' position.
4184 * When reshaping backwards, we round in the opposite direction
4185 * and perform the reverse test: next write position must not be
4186 * less than current safe position.
4188 * In all this the minimum difference in data offsets
4189 * (conf->offset_diff - always positive) allows a bit of slack,
4190 * so next can be after 'safe', but not by more than offset_disk
4192 * We need to prepare all the bios here before we start any IO
4193 * to ensure the size we choose is acceptable to all devices.
4194 * The means one for each copy for write-out and an extra one for
4196 * We store the read-in bio in ->master_bio and the others in
4197 * ->devs[x].bio and ->devs[x].repl_bio.
4199 struct r10conf
*conf
= mddev
->private;
4200 struct r10bio
*r10_bio
;
4201 sector_t next
, safe
, last
;
4205 struct md_rdev
*rdev
;
4208 struct bio
*bio
, *read_bio
;
4209 int sectors_done
= 0;
4211 if (sector_nr
== 0) {
4212 /* If restarting in the middle, skip the initial sectors */
4213 if (mddev
->reshape_backwards
&&
4214 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4215 sector_nr
= (raid10_size(mddev
, 0, 0)
4216 - conf
->reshape_progress
);
4217 } else if (!mddev
->reshape_backwards
&&
4218 conf
->reshape_progress
> 0)
4219 sector_nr
= conf
->reshape_progress
;
4221 mddev
->curr_resync_completed
= sector_nr
;
4222 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4228 /* We don't use sector_nr to track where we are up to
4229 * as that doesn't work well for ->reshape_backwards.
4230 * So just use ->reshape_progress.
4232 if (mddev
->reshape_backwards
) {
4233 /* 'next' is the earliest device address that we might
4234 * write to for this chunk in the new layout
4236 next
= first_dev_address(conf
->reshape_progress
- 1,
4239 /* 'safe' is the last device address that we might read from
4240 * in the old layout after a restart
4242 safe
= last_dev_address(conf
->reshape_safe
- 1,
4245 if (next
+ conf
->offset_diff
< safe
)
4248 last
= conf
->reshape_progress
- 1;
4249 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4250 & conf
->prev
.chunk_mask
);
4251 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4252 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4254 /* 'next' is after the last device address that we
4255 * might write to for this chunk in the new layout
4257 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4259 /* 'safe' is the earliest device address that we might
4260 * read from in the old layout after a restart
4262 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4264 /* Need to update metadata if 'next' might be beyond 'safe'
4265 * as that would possibly corrupt data
4267 if (next
> safe
+ conf
->offset_diff
)
4270 sector_nr
= conf
->reshape_progress
;
4271 last
= sector_nr
| (conf
->geo
.chunk_mask
4272 & conf
->prev
.chunk_mask
);
4274 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4275 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4279 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4280 /* Need to update reshape_position in metadata */
4282 mddev
->reshape_position
= conf
->reshape_progress
;
4283 if (mddev
->reshape_backwards
)
4284 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4285 - conf
->reshape_progress
;
4287 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4288 conf
->reshape_checkpoint
= jiffies
;
4289 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4290 md_wakeup_thread(mddev
->thread
);
4291 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4292 kthread_should_stop());
4293 conf
->reshape_safe
= mddev
->reshape_position
;
4294 allow_barrier(conf
);
4298 /* Now schedule reads for blocks from sector_nr to last */
4299 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4300 raise_barrier(conf
, sectors_done
!= 0);
4301 atomic_set(&r10_bio
->remaining
, 0);
4302 r10_bio
->mddev
= mddev
;
4303 r10_bio
->sector
= sector_nr
;
4304 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4305 r10_bio
->sectors
= last
- sector_nr
+ 1;
4306 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4307 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4310 /* Cannot read from here, so need to record bad blocks
4311 * on all the target devices.
4314 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4315 return sectors_done
;
4318 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4320 read_bio
->bi_bdev
= rdev
->bdev
;
4321 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4322 + rdev
->data_offset
);
4323 read_bio
->bi_private
= r10_bio
;
4324 read_bio
->bi_end_io
= end_sync_read
;
4325 read_bio
->bi_rw
= READ
;
4326 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4327 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4328 read_bio
->bi_vcnt
= 0;
4329 read_bio
->bi_idx
= 0;
4330 read_bio
->bi_size
= 0;
4331 r10_bio
->master_bio
= read_bio
;
4332 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4334 /* Now find the locations in the new layout */
4335 __raid10_find_phys(&conf
->geo
, r10_bio
);
4338 read_bio
->bi_next
= NULL
;
4340 for (s
= 0; s
< conf
->copies
*2; s
++) {
4342 int d
= r10_bio
->devs
[s
/2].devnum
;
4343 struct md_rdev
*rdev2
;
4345 rdev2
= conf
->mirrors
[d
].replacement
;
4346 b
= r10_bio
->devs
[s
/2].repl_bio
;
4348 rdev2
= conf
->mirrors
[d
].rdev
;
4349 b
= r10_bio
->devs
[s
/2].bio
;
4351 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4353 b
->bi_bdev
= rdev2
->bdev
;
4354 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4355 b
->bi_private
= r10_bio
;
4356 b
->bi_end_io
= end_reshape_write
;
4358 b
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4359 b
->bi_flags
|= 1 << BIO_UPTODATE
;
4367 /* Now add as many pages as possible to all of these bios. */
4370 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4371 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4372 int len
= (max_sectors
- s
) << 9;
4373 if (len
> PAGE_SIZE
)
4375 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4377 if (bio_add_page(bio
, page
, len
, 0))
4380 /* Didn't fit, must stop */
4382 bio2
&& bio2
!= bio
;
4383 bio2
= bio2
->bi_next
) {
4384 /* Remove last page from this bio */
4386 bio2
->bi_size
-= len
;
4387 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4391 sector_nr
+= len
>> 9;
4392 nr_sectors
+= len
>> 9;
4395 r10_bio
->sectors
= nr_sectors
;
4397 /* Now submit the read */
4398 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4399 atomic_inc(&r10_bio
->remaining
);
4400 read_bio
->bi_next
= NULL
;
4401 generic_make_request(read_bio
);
4402 sector_nr
+= nr_sectors
;
4403 sectors_done
+= nr_sectors
;
4404 if (sector_nr
<= last
)
4407 /* Now that we have done the whole section we can
4408 * update reshape_progress
4410 if (mddev
->reshape_backwards
)
4411 conf
->reshape_progress
-= sectors_done
;
4413 conf
->reshape_progress
+= sectors_done
;
4415 return sectors_done
;
4418 static void end_reshape_request(struct r10bio
*r10_bio
);
4419 static int handle_reshape_read_error(struct mddev
*mddev
,
4420 struct r10bio
*r10_bio
);
4421 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4423 /* Reshape read completed. Hopefully we have a block
4425 * If we got a read error then we do sync 1-page reads from
4426 * elsewhere until we find the data - or give up.
4428 struct r10conf
*conf
= mddev
->private;
4431 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4432 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4433 /* Reshape has been aborted */
4434 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4438 /* We definitely have the data in the pages, schedule the
4441 atomic_set(&r10_bio
->remaining
, 1);
4442 for (s
= 0; s
< conf
->copies
*2; s
++) {
4444 int d
= r10_bio
->devs
[s
/2].devnum
;
4445 struct md_rdev
*rdev
;
4447 rdev
= conf
->mirrors
[d
].replacement
;
4448 b
= r10_bio
->devs
[s
/2].repl_bio
;
4450 rdev
= conf
->mirrors
[d
].rdev
;
4451 b
= r10_bio
->devs
[s
/2].bio
;
4453 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4455 atomic_inc(&rdev
->nr_pending
);
4456 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4457 atomic_inc(&r10_bio
->remaining
);
4459 generic_make_request(b
);
4461 end_reshape_request(r10_bio
);
4464 static void end_reshape(struct r10conf
*conf
)
4466 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4469 spin_lock_irq(&conf
->device_lock
);
4470 conf
->prev
= conf
->geo
;
4471 md_finish_reshape(conf
->mddev
);
4473 conf
->reshape_progress
= MaxSector
;
4474 spin_unlock_irq(&conf
->device_lock
);
4476 /* read-ahead size must cover two whole stripes, which is
4477 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4479 if (conf
->mddev
->queue
) {
4480 int stripe
= conf
->geo
.raid_disks
*
4481 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4482 stripe
/= conf
->geo
.near_copies
;
4483 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4484 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4490 static int handle_reshape_read_error(struct mddev
*mddev
,
4491 struct r10bio
*r10_bio
)
4493 /* Use sync reads to get the blocks from somewhere else */
4494 int sectors
= r10_bio
->sectors
;
4495 struct r10conf
*conf
= mddev
->private;
4497 struct r10bio r10_bio
;
4498 struct r10dev devs
[conf
->copies
];
4500 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4503 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4505 r10b
->sector
= r10_bio
->sector
;
4506 __raid10_find_phys(&conf
->prev
, r10b
);
4511 int first_slot
= slot
;
4513 if (s
> (PAGE_SIZE
>> 9))
4517 int d
= r10b
->devs
[slot
].devnum
;
4518 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4521 test_bit(Faulty
, &rdev
->flags
) ||
4522 !test_bit(In_sync
, &rdev
->flags
))
4525 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4526 success
= sync_page_io(rdev
,
4535 if (slot
>= conf
->copies
)
4537 if (slot
== first_slot
)
4541 /* couldn't read this block, must give up */
4542 set_bit(MD_RECOVERY_INTR
,
4552 static void end_reshape_write(struct bio
*bio
, int error
)
4554 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4555 struct r10bio
*r10_bio
= bio
->bi_private
;
4556 struct mddev
*mddev
= r10_bio
->mddev
;
4557 struct r10conf
*conf
= mddev
->private;
4561 struct md_rdev
*rdev
= NULL
;
4563 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4565 rdev
= conf
->mirrors
[d
].replacement
;
4568 rdev
= conf
->mirrors
[d
].rdev
;
4572 /* FIXME should record badblock */
4573 md_error(mddev
, rdev
);
4576 rdev_dec_pending(rdev
, mddev
);
4577 end_reshape_request(r10_bio
);
4580 static void end_reshape_request(struct r10bio
*r10_bio
)
4582 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4584 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4585 bio_put(r10_bio
->master_bio
);
4589 static void raid10_finish_reshape(struct mddev
*mddev
)
4591 struct r10conf
*conf
= mddev
->private;
4593 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4596 if (mddev
->delta_disks
> 0) {
4597 sector_t size
= raid10_size(mddev
, 0, 0);
4598 md_set_array_sectors(mddev
, size
);
4599 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4600 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4601 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4603 mddev
->resync_max_sectors
= size
;
4604 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4605 revalidate_disk(mddev
->gendisk
);
4608 for (d
= conf
->geo
.raid_disks
;
4609 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4611 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4613 clear_bit(In_sync
, &rdev
->flags
);
4614 rdev
= conf
->mirrors
[d
].replacement
;
4616 clear_bit(In_sync
, &rdev
->flags
);
4619 mddev
->layout
= mddev
->new_layout
;
4620 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4621 mddev
->reshape_position
= MaxSector
;
4622 mddev
->delta_disks
= 0;
4623 mddev
->reshape_backwards
= 0;
4626 static struct md_personality raid10_personality
=
4630 .owner
= THIS_MODULE
,
4631 .make_request
= make_request
,
4635 .error_handler
= error
,
4636 .hot_add_disk
= raid10_add_disk
,
4637 .hot_remove_disk
= raid10_remove_disk
,
4638 .spare_active
= raid10_spare_active
,
4639 .sync_request
= sync_request
,
4640 .quiesce
= raid10_quiesce
,
4641 .size
= raid10_size
,
4642 .resize
= raid10_resize
,
4643 .takeover
= raid10_takeover
,
4644 .check_reshape
= raid10_check_reshape
,
4645 .start_reshape
= raid10_start_reshape
,
4646 .finish_reshape
= raid10_finish_reshape
,
4649 static int __init
raid_init(void)
4651 return register_md_personality(&raid10_personality
);
4654 static void raid_exit(void)
4656 unregister_md_personality(&raid10_personality
);
4659 module_init(raid_init
);
4660 module_exit(raid_exit
);
4661 MODULE_LICENSE("GPL");
4662 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4663 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4664 MODULE_ALIAS("md-raid10");
4665 MODULE_ALIAS("md-level-10");
4667 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);