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(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
,
955 conf
->resync_lock
, );
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
,
963 conf
->resync_lock
, );
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
)),
1001 spin_unlock_irq(&conf
->resync_lock
);
1004 static void allow_barrier(struct r10conf
*conf
)
1006 unsigned long flags
;
1007 spin_lock_irqsave(&conf
->resync_lock
, flags
);
1009 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
1010 wake_up(&conf
->wait_barrier
);
1013 static void freeze_array(struct r10conf
*conf
)
1015 /* stop syncio and normal IO and wait for everything to
1017 * We increment barrier and nr_waiting, and then
1018 * wait until nr_pending match nr_queued+1
1019 * This is called in the context of one normal IO request
1020 * that has failed. Thus any sync request that might be pending
1021 * will be blocked by nr_pending, and we need to wait for
1022 * pending IO requests to complete or be queued for re-try.
1023 * Thus the number queued (nr_queued) plus this request (1)
1024 * must match the number of pending IOs (nr_pending) before
1027 spin_lock_irq(&conf
->resync_lock
);
1030 wait_event_lock_irq(conf
->wait_barrier
,
1031 conf
->nr_pending
== conf
->nr_queued
+1,
1033 flush_pending_writes(conf
));
1035 spin_unlock_irq(&conf
->resync_lock
);
1038 static void unfreeze_array(struct r10conf
*conf
)
1040 /* reverse the effect of the freeze */
1041 spin_lock_irq(&conf
->resync_lock
);
1044 wake_up(&conf
->wait_barrier
);
1045 spin_unlock_irq(&conf
->resync_lock
);
1048 static sector_t
choose_data_offset(struct r10bio
*r10_bio
,
1049 struct md_rdev
*rdev
)
1051 if (!test_bit(MD_RECOVERY_RESHAPE
, &rdev
->mddev
->recovery
) ||
1052 test_bit(R10BIO_Previous
, &r10_bio
->state
))
1053 return rdev
->data_offset
;
1055 return rdev
->new_data_offset
;
1058 struct raid10_plug_cb
{
1059 struct blk_plug_cb cb
;
1060 struct bio_list pending
;
1064 static void raid10_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1066 struct raid10_plug_cb
*plug
= container_of(cb
, struct raid10_plug_cb
,
1068 struct mddev
*mddev
= plug
->cb
.data
;
1069 struct r10conf
*conf
= mddev
->private;
1072 if (from_schedule
|| current
->bio_list
) {
1073 spin_lock_irq(&conf
->device_lock
);
1074 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1075 conf
->pending_count
+= plug
->pending_cnt
;
1076 spin_unlock_irq(&conf
->device_lock
);
1077 md_wakeup_thread(mddev
->thread
);
1082 /* we aren't scheduling, so we can do the write-out directly. */
1083 bio
= bio_list_get(&plug
->pending
);
1084 bitmap_unplug(mddev
->bitmap
);
1085 wake_up(&conf
->wait_barrier
);
1087 while (bio
) { /* submit pending writes */
1088 struct bio
*next
= bio
->bi_next
;
1089 bio
->bi_next
= NULL
;
1090 generic_make_request(bio
);
1096 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1098 struct r10conf
*conf
= mddev
->private;
1099 struct r10bio
*r10_bio
;
1100 struct bio
*read_bio
;
1102 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1103 int chunk_sects
= chunk_mask
+ 1;
1104 const int rw
= bio_data_dir(bio
);
1105 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1106 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1107 const unsigned long do_discard
= (bio
->bi_rw
1108 & (REQ_DISCARD
| REQ_SECURE
));
1109 unsigned long flags
;
1110 struct md_rdev
*blocked_rdev
;
1111 struct blk_plug_cb
*cb
;
1112 struct raid10_plug_cb
*plug
= NULL
;
1113 int sectors_handled
;
1117 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1118 md_flush_request(mddev
, bio
);
1122 /* If this request crosses a chunk boundary, we need to
1123 * split it. This will only happen for 1 PAGE (or less) requests.
1125 if (unlikely((bio
->bi_sector
& chunk_mask
) + (bio
->bi_size
>> 9)
1127 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1128 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1129 struct bio_pair
*bp
;
1130 /* Sanity check -- queue functions should prevent this happening */
1131 if ((bio
->bi_vcnt
!= 1 && bio
->bi_vcnt
!= 0) ||
1134 /* This is a one page bio that upper layers
1135 * refuse to split for us, so we need to split it.
1138 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
1140 /* Each of these 'make_request' calls will call 'wait_barrier'.
1141 * If the first succeeds but the second blocks due to the resync
1142 * thread raising the barrier, we will deadlock because the
1143 * IO to the underlying device will be queued in generic_make_request
1144 * and will never complete, so will never reduce nr_pending.
1145 * So increment nr_waiting here so no new raise_barriers will
1146 * succeed, and so the second wait_barrier cannot block.
1148 spin_lock_irq(&conf
->resync_lock
);
1150 spin_unlock_irq(&conf
->resync_lock
);
1152 make_request(mddev
, &bp
->bio1
);
1153 make_request(mddev
, &bp
->bio2
);
1155 spin_lock_irq(&conf
->resync_lock
);
1157 wake_up(&conf
->wait_barrier
);
1158 spin_unlock_irq(&conf
->resync_lock
);
1160 bio_pair_release(bp
);
1163 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1164 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1165 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
1171 md_write_start(mddev
, bio
);
1174 * Register the new request and wait if the reconstruction
1175 * thread has put up a bar for new requests.
1176 * Continue immediately if no resync is active currently.
1180 sectors
= bio
->bi_size
>> 9;
1181 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1182 bio
->bi_sector
< conf
->reshape_progress
&&
1183 bio
->bi_sector
+ sectors
> conf
->reshape_progress
) {
1184 /* IO spans the reshape position. Need to wait for
1187 allow_barrier(conf
);
1188 wait_event(conf
->wait_barrier
,
1189 conf
->reshape_progress
<= bio
->bi_sector
||
1190 conf
->reshape_progress
>= bio
->bi_sector
+ sectors
);
1193 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1194 bio_data_dir(bio
) == WRITE
&&
1195 (mddev
->reshape_backwards
1196 ? (bio
->bi_sector
< conf
->reshape_safe
&&
1197 bio
->bi_sector
+ sectors
> conf
->reshape_progress
)
1198 : (bio
->bi_sector
+ sectors
> conf
->reshape_safe
&&
1199 bio
->bi_sector
< conf
->reshape_progress
))) {
1200 /* Need to update reshape_position in metadata */
1201 mddev
->reshape_position
= conf
->reshape_progress
;
1202 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1203 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1204 md_wakeup_thread(mddev
->thread
);
1205 wait_event(mddev
->sb_wait
,
1206 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1208 conf
->reshape_safe
= mddev
->reshape_position
;
1211 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1213 r10_bio
->master_bio
= bio
;
1214 r10_bio
->sectors
= sectors
;
1216 r10_bio
->mddev
= mddev
;
1217 r10_bio
->sector
= bio
->bi_sector
;
1220 /* We might need to issue multiple reads to different
1221 * devices if there are bad blocks around, so we keep
1222 * track of the number of reads in bio->bi_phys_segments.
1223 * If this is 0, there is only one r10_bio and no locking
1224 * will be needed when the request completes. If it is
1225 * non-zero, then it is the number of not-completed requests.
1227 bio
->bi_phys_segments
= 0;
1228 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1232 * read balancing logic:
1234 struct md_rdev
*rdev
;
1238 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1240 raid_end_bio_io(r10_bio
);
1243 slot
= r10_bio
->read_slot
;
1245 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1246 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1249 r10_bio
->devs
[slot
].bio
= read_bio
;
1250 r10_bio
->devs
[slot
].rdev
= rdev
;
1252 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1253 choose_data_offset(r10_bio
, rdev
);
1254 read_bio
->bi_bdev
= rdev
->bdev
;
1255 read_bio
->bi_end_io
= raid10_end_read_request
;
1256 read_bio
->bi_rw
= READ
| do_sync
;
1257 read_bio
->bi_private
= r10_bio
;
1259 if (max_sectors
< r10_bio
->sectors
) {
1260 /* Could not read all from this device, so we will
1261 * need another r10_bio.
1263 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1265 r10_bio
->sectors
= max_sectors
;
1266 spin_lock_irq(&conf
->device_lock
);
1267 if (bio
->bi_phys_segments
== 0)
1268 bio
->bi_phys_segments
= 2;
1270 bio
->bi_phys_segments
++;
1271 spin_unlock(&conf
->device_lock
);
1272 /* Cannot call generic_make_request directly
1273 * as that will be queued in __generic_make_request
1274 * and subsequent mempool_alloc might block
1275 * waiting for it. so hand bio over to raid10d.
1277 reschedule_retry(r10_bio
);
1279 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1281 r10_bio
->master_bio
= bio
;
1282 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
1285 r10_bio
->mddev
= mddev
;
1286 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1289 generic_make_request(read_bio
);
1296 if (conf
->pending_count
>= max_queued_requests
) {
1297 md_wakeup_thread(mddev
->thread
);
1298 wait_event(conf
->wait_barrier
,
1299 conf
->pending_count
< max_queued_requests
);
1301 /* first select target devices under rcu_lock and
1302 * inc refcount on their rdev. Record them by setting
1304 * If there are known/acknowledged bad blocks on any device
1305 * on which we have seen a write error, we want to avoid
1306 * writing to those blocks. This potentially requires several
1307 * writes to write around the bad blocks. Each set of writes
1308 * gets its own r10_bio with a set of bios attached. The number
1309 * of r10_bios is recored in bio->bi_phys_segments just as with
1313 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1314 raid10_find_phys(conf
, r10_bio
);
1316 blocked_rdev
= NULL
;
1318 max_sectors
= r10_bio
->sectors
;
1320 for (i
= 0; i
< conf
->copies
; i
++) {
1321 int d
= r10_bio
->devs
[i
].devnum
;
1322 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1323 struct md_rdev
*rrdev
= rcu_dereference(
1324 conf
->mirrors
[d
].replacement
);
1327 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1328 atomic_inc(&rdev
->nr_pending
);
1329 blocked_rdev
= rdev
;
1332 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1333 atomic_inc(&rrdev
->nr_pending
);
1334 blocked_rdev
= rrdev
;
1337 if (rdev
&& (test_bit(Faulty
, &rdev
->flags
)
1338 || test_bit(Unmerged
, &rdev
->flags
)))
1340 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1341 || test_bit(Unmerged
, &rrdev
->flags
)))
1344 r10_bio
->devs
[i
].bio
= NULL
;
1345 r10_bio
->devs
[i
].repl_bio
= NULL
;
1347 if (!rdev
&& !rrdev
) {
1348 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1351 if (rdev
&& test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1353 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1357 is_bad
= is_badblock(rdev
, dev_sector
,
1359 &first_bad
, &bad_sectors
);
1361 /* Mustn't write here until the bad block
1364 atomic_inc(&rdev
->nr_pending
);
1365 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1366 blocked_rdev
= rdev
;
1369 if (is_bad
&& first_bad
<= dev_sector
) {
1370 /* Cannot write here at all */
1371 bad_sectors
-= (dev_sector
- first_bad
);
1372 if (bad_sectors
< max_sectors
)
1373 /* Mustn't write more than bad_sectors
1374 * to other devices yet
1376 max_sectors
= bad_sectors
;
1377 /* We don't set R10BIO_Degraded as that
1378 * only applies if the disk is missing,
1379 * so it might be re-added, and we want to
1380 * know to recover this chunk.
1381 * In this case the device is here, and the
1382 * fact that this chunk is not in-sync is
1383 * recorded in the bad block log.
1388 int good_sectors
= first_bad
- dev_sector
;
1389 if (good_sectors
< max_sectors
)
1390 max_sectors
= good_sectors
;
1394 r10_bio
->devs
[i
].bio
= bio
;
1395 atomic_inc(&rdev
->nr_pending
);
1398 r10_bio
->devs
[i
].repl_bio
= bio
;
1399 atomic_inc(&rrdev
->nr_pending
);
1404 if (unlikely(blocked_rdev
)) {
1405 /* Have to wait for this device to get unblocked, then retry */
1409 for (j
= 0; j
< i
; j
++) {
1410 if (r10_bio
->devs
[j
].bio
) {
1411 d
= r10_bio
->devs
[j
].devnum
;
1412 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1414 if (r10_bio
->devs
[j
].repl_bio
) {
1415 struct md_rdev
*rdev
;
1416 d
= r10_bio
->devs
[j
].devnum
;
1417 rdev
= conf
->mirrors
[d
].replacement
;
1419 /* Race with remove_disk */
1421 rdev
= conf
->mirrors
[d
].rdev
;
1423 rdev_dec_pending(rdev
, mddev
);
1426 allow_barrier(conf
);
1427 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1432 if (max_sectors
< r10_bio
->sectors
) {
1433 /* We are splitting this into multiple parts, so
1434 * we need to prepare for allocating another r10_bio.
1436 r10_bio
->sectors
= max_sectors
;
1437 spin_lock_irq(&conf
->device_lock
);
1438 if (bio
->bi_phys_segments
== 0)
1439 bio
->bi_phys_segments
= 2;
1441 bio
->bi_phys_segments
++;
1442 spin_unlock_irq(&conf
->device_lock
);
1444 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1446 atomic_set(&r10_bio
->remaining
, 1);
1447 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1449 for (i
= 0; i
< conf
->copies
; i
++) {
1451 int d
= r10_bio
->devs
[i
].devnum
;
1452 if (r10_bio
->devs
[i
].bio
) {
1453 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
1454 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1455 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1457 r10_bio
->devs
[i
].bio
= mbio
;
1459 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1460 choose_data_offset(r10_bio
,
1462 mbio
->bi_bdev
= rdev
->bdev
;
1463 mbio
->bi_end_io
= raid10_end_write_request
;
1464 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
| do_discard
;
1465 mbio
->bi_private
= r10_bio
;
1467 atomic_inc(&r10_bio
->remaining
);
1469 cb
= blk_check_plugged(raid10_unplug
, mddev
,
1472 plug
= container_of(cb
, struct raid10_plug_cb
,
1476 spin_lock_irqsave(&conf
->device_lock
, flags
);
1478 bio_list_add(&plug
->pending
, mbio
);
1479 plug
->pending_cnt
++;
1481 bio_list_add(&conf
->pending_bio_list
, mbio
);
1482 conf
->pending_count
++;
1484 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1486 md_wakeup_thread(mddev
->thread
);
1489 if (r10_bio
->devs
[i
].repl_bio
) {
1490 struct md_rdev
*rdev
= conf
->mirrors
[d
].replacement
;
1492 /* Replacement just got moved to main 'rdev' */
1494 rdev
= conf
->mirrors
[d
].rdev
;
1496 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1497 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1499 r10_bio
->devs
[i
].repl_bio
= mbio
;
1501 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1504 mbio
->bi_bdev
= rdev
->bdev
;
1505 mbio
->bi_end_io
= raid10_end_write_request
;
1506 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
| do_discard
;
1507 mbio
->bi_private
= r10_bio
;
1509 atomic_inc(&r10_bio
->remaining
);
1510 spin_lock_irqsave(&conf
->device_lock
, flags
);
1511 bio_list_add(&conf
->pending_bio_list
, mbio
);
1512 conf
->pending_count
++;
1513 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1514 if (!mddev_check_plugged(mddev
))
1515 md_wakeup_thread(mddev
->thread
);
1519 /* Don't remove the bias on 'remaining' (one_write_done) until
1520 * after checking if we need to go around again.
1523 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1524 one_write_done(r10_bio
);
1525 /* We need another r10_bio. It has already been counted
1526 * in bio->bi_phys_segments.
1528 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1530 r10_bio
->master_bio
= bio
;
1531 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1533 r10_bio
->mddev
= mddev
;
1534 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1538 one_write_done(r10_bio
);
1540 /* In case raid10d snuck in to freeze_array */
1541 wake_up(&conf
->wait_barrier
);
1544 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1546 struct r10conf
*conf
= mddev
->private;
1549 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1550 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1551 if (conf
->geo
.near_copies
> 1)
1552 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1553 if (conf
->geo
.far_copies
> 1) {
1554 if (conf
->geo
.far_offset
)
1555 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1557 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1559 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1560 conf
->geo
.raid_disks
- mddev
->degraded
);
1561 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1562 seq_printf(seq
, "%s",
1563 conf
->mirrors
[i
].rdev
&&
1564 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1565 seq_printf(seq
, "]");
1568 /* check if there are enough drives for
1569 * every block to appear on atleast one.
1570 * Don't consider the device numbered 'ignore'
1571 * as we might be about to remove it.
1573 static int _enough(struct r10conf
*conf
, struct geom
*geo
, int ignore
)
1578 int n
= conf
->copies
;
1582 if (conf
->mirrors
[this].rdev
&&
1585 this = (this+1) % geo
->raid_disks
;
1589 first
= (first
+ geo
->near_copies
) % geo
->raid_disks
;
1590 } while (first
!= 0);
1594 static int enough(struct r10conf
*conf
, int ignore
)
1596 return _enough(conf
, &conf
->geo
, ignore
) &&
1597 _enough(conf
, &conf
->prev
, ignore
);
1600 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1602 char b
[BDEVNAME_SIZE
];
1603 struct r10conf
*conf
= mddev
->private;
1606 * If it is not operational, then we have already marked it as dead
1607 * else if it is the last working disks, ignore the error, let the
1608 * next level up know.
1609 * else mark the drive as failed
1611 if (test_bit(In_sync
, &rdev
->flags
)
1612 && !enough(conf
, rdev
->raid_disk
))
1614 * Don't fail the drive, just return an IO error.
1617 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1618 unsigned long flags
;
1619 spin_lock_irqsave(&conf
->device_lock
, flags
);
1621 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1623 * if recovery is running, make sure it aborts.
1625 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1627 set_bit(Blocked
, &rdev
->flags
);
1628 set_bit(Faulty
, &rdev
->flags
);
1629 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1631 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1632 "md/raid10:%s: Operation continuing on %d devices.\n",
1633 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1634 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1637 static void print_conf(struct r10conf
*conf
)
1640 struct raid10_info
*tmp
;
1642 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1644 printk(KERN_DEBUG
"(!conf)\n");
1647 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1648 conf
->geo
.raid_disks
);
1650 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1651 char b
[BDEVNAME_SIZE
];
1652 tmp
= conf
->mirrors
+ i
;
1654 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1655 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1656 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1657 bdevname(tmp
->rdev
->bdev
,b
));
1661 static void close_sync(struct r10conf
*conf
)
1664 allow_barrier(conf
);
1666 mempool_destroy(conf
->r10buf_pool
);
1667 conf
->r10buf_pool
= NULL
;
1670 static int raid10_spare_active(struct mddev
*mddev
)
1673 struct r10conf
*conf
= mddev
->private;
1674 struct raid10_info
*tmp
;
1676 unsigned long flags
;
1679 * Find all non-in_sync disks within the RAID10 configuration
1680 * and mark them in_sync
1682 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1683 tmp
= conf
->mirrors
+ i
;
1684 if (tmp
->replacement
1685 && tmp
->replacement
->recovery_offset
== MaxSector
1686 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1687 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1688 /* Replacement has just become active */
1690 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1693 /* Replaced device not technically faulty,
1694 * but we need to be sure it gets removed
1695 * and never re-added.
1697 set_bit(Faulty
, &tmp
->rdev
->flags
);
1698 sysfs_notify_dirent_safe(
1699 tmp
->rdev
->sysfs_state
);
1701 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1702 } else if (tmp
->rdev
1703 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1704 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1706 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1709 spin_lock_irqsave(&conf
->device_lock
, flags
);
1710 mddev
->degraded
-= count
;
1711 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1718 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1720 struct r10conf
*conf
= mddev
->private;
1724 int last
= conf
->geo
.raid_disks
- 1;
1725 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1727 if (mddev
->recovery_cp
< MaxSector
)
1728 /* only hot-add to in-sync arrays, as recovery is
1729 * very different from resync
1732 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, &conf
->prev
, -1))
1735 if (rdev
->raid_disk
>= 0)
1736 first
= last
= rdev
->raid_disk
;
1738 if (q
->merge_bvec_fn
) {
1739 set_bit(Unmerged
, &rdev
->flags
);
1740 mddev
->merge_check_needed
= 1;
1743 if (rdev
->saved_raid_disk
>= first
&&
1744 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1745 mirror
= rdev
->saved_raid_disk
;
1748 for ( ; mirror
<= last
; mirror
++) {
1749 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1750 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1753 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1754 p
->replacement
!= NULL
)
1756 clear_bit(In_sync
, &rdev
->flags
);
1757 set_bit(Replacement
, &rdev
->flags
);
1758 rdev
->raid_disk
= mirror
;
1760 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1761 rdev
->data_offset
<< 9);
1763 rcu_assign_pointer(p
->replacement
, rdev
);
1767 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1768 rdev
->data_offset
<< 9);
1770 p
->head_position
= 0;
1771 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1772 rdev
->raid_disk
= mirror
;
1774 if (rdev
->saved_raid_disk
!= mirror
)
1776 rcu_assign_pointer(p
->rdev
, rdev
);
1779 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1780 /* Some requests might not have seen this new
1781 * merge_bvec_fn. We must wait for them to complete
1782 * before merging the device fully.
1783 * First we make sure any code which has tested
1784 * our function has submitted the request, then
1785 * we wait for all outstanding requests to complete.
1787 synchronize_sched();
1788 raise_barrier(conf
, 0);
1789 lower_barrier(conf
);
1790 clear_bit(Unmerged
, &rdev
->flags
);
1792 md_integrity_add_rdev(rdev
, mddev
);
1793 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1794 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1800 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1802 struct r10conf
*conf
= mddev
->private;
1804 int number
= rdev
->raid_disk
;
1805 struct md_rdev
**rdevp
;
1806 struct raid10_info
*p
= conf
->mirrors
+ number
;
1809 if (rdev
== p
->rdev
)
1811 else if (rdev
== p
->replacement
)
1812 rdevp
= &p
->replacement
;
1816 if (test_bit(In_sync
, &rdev
->flags
) ||
1817 atomic_read(&rdev
->nr_pending
)) {
1821 /* Only remove faulty devices if recovery
1824 if (!test_bit(Faulty
, &rdev
->flags
) &&
1825 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1826 (!p
->replacement
|| p
->replacement
== rdev
) &&
1827 number
< conf
->geo
.raid_disks
&&
1834 if (atomic_read(&rdev
->nr_pending
)) {
1835 /* lost the race, try later */
1839 } else if (p
->replacement
) {
1840 /* We must have just cleared 'rdev' */
1841 p
->rdev
= p
->replacement
;
1842 clear_bit(Replacement
, &p
->replacement
->flags
);
1843 smp_mb(); /* Make sure other CPUs may see both as identical
1844 * but will never see neither -- if they are careful.
1846 p
->replacement
= NULL
;
1847 clear_bit(WantReplacement
, &rdev
->flags
);
1849 /* We might have just remove the Replacement as faulty
1850 * Clear the flag just in case
1852 clear_bit(WantReplacement
, &rdev
->flags
);
1854 err
= md_integrity_register(mddev
);
1863 static void end_sync_read(struct bio
*bio
, int error
)
1865 struct r10bio
*r10_bio
= bio
->bi_private
;
1866 struct r10conf
*conf
= r10_bio
->mddev
->private;
1869 if (bio
== r10_bio
->master_bio
) {
1870 /* this is a reshape read */
1871 d
= r10_bio
->read_slot
; /* really the read dev */
1873 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1875 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1876 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1878 /* The write handler will notice the lack of
1879 * R10BIO_Uptodate and record any errors etc
1881 atomic_add(r10_bio
->sectors
,
1882 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1884 /* for reconstruct, we always reschedule after a read.
1885 * for resync, only after all reads
1887 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1888 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1889 atomic_dec_and_test(&r10_bio
->remaining
)) {
1890 /* we have read all the blocks,
1891 * do the comparison in process context in raid10d
1893 reschedule_retry(r10_bio
);
1897 static void end_sync_request(struct r10bio
*r10_bio
)
1899 struct mddev
*mddev
= r10_bio
->mddev
;
1901 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1902 if (r10_bio
->master_bio
== NULL
) {
1903 /* the primary of several recovery bios */
1904 sector_t s
= r10_bio
->sectors
;
1905 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1906 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1907 reschedule_retry(r10_bio
);
1910 md_done_sync(mddev
, s
, 1);
1913 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1914 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1915 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1916 reschedule_retry(r10_bio
);
1924 static void end_sync_write(struct bio
*bio
, int error
)
1926 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1927 struct r10bio
*r10_bio
= bio
->bi_private
;
1928 struct mddev
*mddev
= r10_bio
->mddev
;
1929 struct r10conf
*conf
= mddev
->private;
1935 struct md_rdev
*rdev
= NULL
;
1937 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1939 rdev
= conf
->mirrors
[d
].replacement
;
1941 rdev
= conf
->mirrors
[d
].rdev
;
1945 md_error(mddev
, rdev
);
1947 set_bit(WriteErrorSeen
, &rdev
->flags
);
1948 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1949 set_bit(MD_RECOVERY_NEEDED
,
1950 &rdev
->mddev
->recovery
);
1951 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1953 } else if (is_badblock(rdev
,
1954 r10_bio
->devs
[slot
].addr
,
1956 &first_bad
, &bad_sectors
))
1957 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1959 rdev_dec_pending(rdev
, mddev
);
1961 end_sync_request(r10_bio
);
1965 * Note: sync and recover and handled very differently for raid10
1966 * This code is for resync.
1967 * For resync, we read through virtual addresses and read all blocks.
1968 * If there is any error, we schedule a write. The lowest numbered
1969 * drive is authoritative.
1970 * However requests come for physical address, so we need to map.
1971 * For every physical address there are raid_disks/copies virtual addresses,
1972 * which is always are least one, but is not necessarly an integer.
1973 * This means that a physical address can span multiple chunks, so we may
1974 * have to submit multiple io requests for a single sync request.
1977 * We check if all blocks are in-sync and only write to blocks that
1980 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1982 struct r10conf
*conf
= mddev
->private;
1984 struct bio
*tbio
, *fbio
;
1987 atomic_set(&r10_bio
->remaining
, 1);
1989 /* find the first device with a block */
1990 for (i
=0; i
<conf
->copies
; i
++)
1991 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1994 if (i
== conf
->copies
)
1998 fbio
= r10_bio
->devs
[i
].bio
;
2000 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
2001 /* now find blocks with errors */
2002 for (i
=0 ; i
< conf
->copies
; i
++) {
2005 tbio
= r10_bio
->devs
[i
].bio
;
2007 if (tbio
->bi_end_io
!= end_sync_read
)
2011 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2012 /* We know that the bi_io_vec layout is the same for
2013 * both 'first' and 'i', so we just compare them.
2014 * All vec entries are PAGE_SIZE;
2016 for (j
= 0; j
< vcnt
; j
++)
2017 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2018 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2019 fbio
->bi_io_vec
[j
].bv_len
))
2023 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2024 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2025 /* Don't fix anything. */
2028 /* Ok, we need to write this bio, either to correct an
2029 * inconsistency or to correct an unreadable block.
2030 * First we need to fixup bv_offset, bv_len and
2031 * bi_vecs, as the read request might have corrupted these
2033 tbio
->bi_vcnt
= vcnt
;
2034 tbio
->bi_size
= r10_bio
->sectors
<< 9;
2036 tbio
->bi_phys_segments
= 0;
2037 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2038 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
2039 tbio
->bi_next
= NULL
;
2040 tbio
->bi_rw
= WRITE
;
2041 tbio
->bi_private
= r10_bio
;
2042 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
2044 for (j
=0; j
< vcnt
; j
++) {
2045 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2046 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2048 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2049 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2052 tbio
->bi_end_io
= end_sync_write
;
2054 d
= r10_bio
->devs
[i
].devnum
;
2055 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2056 atomic_inc(&r10_bio
->remaining
);
2057 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
2059 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2060 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2061 generic_make_request(tbio
);
2064 /* Now write out to any replacement devices
2067 for (i
= 0; i
< conf
->copies
; i
++) {
2070 tbio
= r10_bio
->devs
[i
].repl_bio
;
2071 if (!tbio
|| !tbio
->bi_end_io
)
2073 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2074 && r10_bio
->devs
[i
].bio
!= fbio
)
2075 for (j
= 0; j
< vcnt
; j
++)
2076 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2077 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2079 d
= r10_bio
->devs
[i
].devnum
;
2080 atomic_inc(&r10_bio
->remaining
);
2081 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2082 tbio
->bi_size
>> 9);
2083 generic_make_request(tbio
);
2087 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2088 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2094 * Now for the recovery code.
2095 * Recovery happens across physical sectors.
2096 * We recover all non-is_sync drives by finding the virtual address of
2097 * each, and then choose a working drive that also has that virt address.
2098 * There is a separate r10_bio for each non-in_sync drive.
2099 * Only the first two slots are in use. The first for reading,
2100 * The second for writing.
2103 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2105 /* We got a read error during recovery.
2106 * We repeat the read in smaller page-sized sections.
2107 * If a read succeeds, write it to the new device or record
2108 * a bad block if we cannot.
2109 * If a read fails, record a bad block on both old and
2112 struct mddev
*mddev
= r10_bio
->mddev
;
2113 struct r10conf
*conf
= mddev
->private;
2114 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2116 int sectors
= r10_bio
->sectors
;
2118 int dr
= r10_bio
->devs
[0].devnum
;
2119 int dw
= r10_bio
->devs
[1].devnum
;
2123 struct md_rdev
*rdev
;
2127 if (s
> (PAGE_SIZE
>>9))
2130 rdev
= conf
->mirrors
[dr
].rdev
;
2131 addr
= r10_bio
->devs
[0].addr
+ sect
,
2132 ok
= sync_page_io(rdev
,
2135 bio
->bi_io_vec
[idx
].bv_page
,
2138 rdev
= conf
->mirrors
[dw
].rdev
;
2139 addr
= r10_bio
->devs
[1].addr
+ sect
;
2140 ok
= sync_page_io(rdev
,
2143 bio
->bi_io_vec
[idx
].bv_page
,
2146 set_bit(WriteErrorSeen
, &rdev
->flags
);
2147 if (!test_and_set_bit(WantReplacement
,
2149 set_bit(MD_RECOVERY_NEEDED
,
2150 &rdev
->mddev
->recovery
);
2154 /* We don't worry if we cannot set a bad block -
2155 * it really is bad so there is no loss in not
2158 rdev_set_badblocks(rdev
, addr
, s
, 0);
2160 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2161 /* need bad block on destination too */
2162 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2163 addr
= r10_bio
->devs
[1].addr
+ sect
;
2164 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2166 /* just abort the recovery */
2168 "md/raid10:%s: recovery aborted"
2169 " due to read error\n",
2172 conf
->mirrors
[dw
].recovery_disabled
2173 = mddev
->recovery_disabled
;
2174 set_bit(MD_RECOVERY_INTR
,
2187 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2189 struct r10conf
*conf
= mddev
->private;
2191 struct bio
*wbio
, *wbio2
;
2193 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2194 fix_recovery_read_error(r10_bio
);
2195 end_sync_request(r10_bio
);
2200 * share the pages with the first bio
2201 * and submit the write request
2203 d
= r10_bio
->devs
[1].devnum
;
2204 wbio
= r10_bio
->devs
[1].bio
;
2205 wbio2
= r10_bio
->devs
[1].repl_bio
;
2206 if (wbio
->bi_end_io
) {
2207 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2208 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
2209 generic_make_request(wbio
);
2211 if (wbio2
&& wbio2
->bi_end_io
) {
2212 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2213 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2214 wbio2
->bi_size
>> 9);
2215 generic_make_request(wbio2
);
2221 * Used by fix_read_error() to decay the per rdev read_errors.
2222 * We halve the read error count for every hour that has elapsed
2223 * since the last recorded read error.
2226 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2228 struct timespec cur_time_mon
;
2229 unsigned long hours_since_last
;
2230 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2232 ktime_get_ts(&cur_time_mon
);
2234 if (rdev
->last_read_error
.tv_sec
== 0 &&
2235 rdev
->last_read_error
.tv_nsec
== 0) {
2236 /* first time we've seen a read error */
2237 rdev
->last_read_error
= cur_time_mon
;
2241 hours_since_last
= (cur_time_mon
.tv_sec
-
2242 rdev
->last_read_error
.tv_sec
) / 3600;
2244 rdev
->last_read_error
= cur_time_mon
;
2247 * if hours_since_last is > the number of bits in read_errors
2248 * just set read errors to 0. We do this to avoid
2249 * overflowing the shift of read_errors by hours_since_last.
2251 if (hours_since_last
>= 8 * sizeof(read_errors
))
2252 atomic_set(&rdev
->read_errors
, 0);
2254 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2257 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2258 int sectors
, struct page
*page
, int rw
)
2263 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2264 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2266 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2270 set_bit(WriteErrorSeen
, &rdev
->flags
);
2271 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2272 set_bit(MD_RECOVERY_NEEDED
,
2273 &rdev
->mddev
->recovery
);
2275 /* need to record an error - either for the block or the device */
2276 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2277 md_error(rdev
->mddev
, rdev
);
2282 * This is a kernel thread which:
2284 * 1. Retries failed read operations on working mirrors.
2285 * 2. Updates the raid superblock when problems encounter.
2286 * 3. Performs writes following reads for array synchronising.
2289 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2291 int sect
= 0; /* Offset from r10_bio->sector */
2292 int sectors
= r10_bio
->sectors
;
2293 struct md_rdev
*rdev
;
2294 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2295 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2297 /* still own a reference to this rdev, so it cannot
2298 * have been cleared recently.
2300 rdev
= conf
->mirrors
[d
].rdev
;
2302 if (test_bit(Faulty
, &rdev
->flags
))
2303 /* drive has already been failed, just ignore any
2304 more fix_read_error() attempts */
2307 check_decay_read_errors(mddev
, rdev
);
2308 atomic_inc(&rdev
->read_errors
);
2309 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2310 char b
[BDEVNAME_SIZE
];
2311 bdevname(rdev
->bdev
, b
);
2314 "md/raid10:%s: %s: Raid device exceeded "
2315 "read_error threshold [cur %d:max %d]\n",
2317 atomic_read(&rdev
->read_errors
), max_read_errors
);
2319 "md/raid10:%s: %s: Failing raid device\n",
2321 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2322 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2328 int sl
= r10_bio
->read_slot
;
2332 if (s
> (PAGE_SIZE
>>9))
2340 d
= r10_bio
->devs
[sl
].devnum
;
2341 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2343 !test_bit(Unmerged
, &rdev
->flags
) &&
2344 test_bit(In_sync
, &rdev
->flags
) &&
2345 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2346 &first_bad
, &bad_sectors
) == 0) {
2347 atomic_inc(&rdev
->nr_pending
);
2349 success
= sync_page_io(rdev
,
2350 r10_bio
->devs
[sl
].addr
+
2353 conf
->tmppage
, READ
, false);
2354 rdev_dec_pending(rdev
, mddev
);
2360 if (sl
== conf
->copies
)
2362 } while (!success
&& sl
!= r10_bio
->read_slot
);
2366 /* Cannot read from anywhere, just mark the block
2367 * as bad on the first device to discourage future
2370 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2371 rdev
= conf
->mirrors
[dn
].rdev
;
2373 if (!rdev_set_badblocks(
2375 r10_bio
->devs
[r10_bio
->read_slot
].addr
2378 md_error(mddev
, rdev
);
2379 r10_bio
->devs
[r10_bio
->read_slot
].bio
2386 /* write it back and re-read */
2388 while (sl
!= r10_bio
->read_slot
) {
2389 char b
[BDEVNAME_SIZE
];
2394 d
= r10_bio
->devs
[sl
].devnum
;
2395 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2397 test_bit(Unmerged
, &rdev
->flags
) ||
2398 !test_bit(In_sync
, &rdev
->flags
))
2401 atomic_inc(&rdev
->nr_pending
);
2403 if (r10_sync_page_io(rdev
,
2404 r10_bio
->devs
[sl
].addr
+
2406 s
, conf
->tmppage
, WRITE
)
2408 /* Well, this device is dead */
2410 "md/raid10:%s: read correction "
2412 " (%d sectors at %llu on %s)\n",
2414 (unsigned long long)(
2416 choose_data_offset(r10_bio
,
2418 bdevname(rdev
->bdev
, b
));
2419 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2422 bdevname(rdev
->bdev
, b
));
2424 rdev_dec_pending(rdev
, mddev
);
2428 while (sl
!= r10_bio
->read_slot
) {
2429 char b
[BDEVNAME_SIZE
];
2434 d
= r10_bio
->devs
[sl
].devnum
;
2435 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2437 !test_bit(In_sync
, &rdev
->flags
))
2440 atomic_inc(&rdev
->nr_pending
);
2442 switch (r10_sync_page_io(rdev
,
2443 r10_bio
->devs
[sl
].addr
+
2448 /* Well, this device is dead */
2450 "md/raid10:%s: unable to read back "
2452 " (%d sectors at %llu on %s)\n",
2454 (unsigned long long)(
2456 choose_data_offset(r10_bio
, rdev
)),
2457 bdevname(rdev
->bdev
, b
));
2458 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2461 bdevname(rdev
->bdev
, b
));
2465 "md/raid10:%s: read error corrected"
2466 " (%d sectors at %llu on %s)\n",
2468 (unsigned long long)(
2470 choose_data_offset(r10_bio
, rdev
)),
2471 bdevname(rdev
->bdev
, b
));
2472 atomic_add(s
, &rdev
->corrected_errors
);
2475 rdev_dec_pending(rdev
, mddev
);
2485 static void bi_complete(struct bio
*bio
, int error
)
2487 complete((struct completion
*)bio
->bi_private
);
2490 static int submit_bio_wait(int rw
, struct bio
*bio
)
2492 struct completion event
;
2495 init_completion(&event
);
2496 bio
->bi_private
= &event
;
2497 bio
->bi_end_io
= bi_complete
;
2498 submit_bio(rw
, bio
);
2499 wait_for_completion(&event
);
2501 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2504 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2506 struct bio
*bio
= r10_bio
->master_bio
;
2507 struct mddev
*mddev
= r10_bio
->mddev
;
2508 struct r10conf
*conf
= mddev
->private;
2509 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2510 /* bio has the data to be written to slot 'i' where
2511 * we just recently had a write error.
2512 * We repeatedly clone the bio and trim down to one block,
2513 * then try the write. Where the write fails we record
2515 * It is conceivable that the bio doesn't exactly align with
2516 * blocks. We must handle this.
2518 * We currently own a reference to the rdev.
2524 int sect_to_write
= r10_bio
->sectors
;
2527 if (rdev
->badblocks
.shift
< 0)
2530 block_sectors
= 1 << rdev
->badblocks
.shift
;
2531 sector
= r10_bio
->sector
;
2532 sectors
= ((r10_bio
->sector
+ block_sectors
)
2533 & ~(sector_t
)(block_sectors
- 1))
2536 while (sect_to_write
) {
2538 if (sectors
> sect_to_write
)
2539 sectors
= sect_to_write
;
2540 /* Write at 'sector' for 'sectors' */
2541 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2542 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2543 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2544 choose_data_offset(r10_bio
, rdev
) +
2545 (sector
- r10_bio
->sector
));
2546 wbio
->bi_bdev
= rdev
->bdev
;
2547 if (submit_bio_wait(WRITE
, wbio
) == 0)
2549 ok
= rdev_set_badblocks(rdev
, sector
,
2554 sect_to_write
-= sectors
;
2556 sectors
= block_sectors
;
2561 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2563 int slot
= r10_bio
->read_slot
;
2565 struct r10conf
*conf
= mddev
->private;
2566 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2567 char b
[BDEVNAME_SIZE
];
2568 unsigned long do_sync
;
2571 /* we got a read error. Maybe the drive is bad. Maybe just
2572 * the block and we can fix it.
2573 * We freeze all other IO, and try reading the block from
2574 * other devices. When we find one, we re-write
2575 * and check it that fixes the read error.
2576 * This is all done synchronously while the array is
2579 bio
= r10_bio
->devs
[slot
].bio
;
2580 bdevname(bio
->bi_bdev
, b
);
2582 r10_bio
->devs
[slot
].bio
= NULL
;
2584 if (mddev
->ro
== 0) {
2586 fix_read_error(conf
, mddev
, r10_bio
);
2587 unfreeze_array(conf
);
2589 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2591 rdev_dec_pending(rdev
, mddev
);
2594 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2596 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2597 " read error for block %llu\n",
2599 (unsigned long long)r10_bio
->sector
);
2600 raid_end_bio_io(r10_bio
);
2604 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2605 slot
= r10_bio
->read_slot
;
2608 "md/raid10:%s: %s: redirecting "
2609 "sector %llu to another mirror\n",
2611 bdevname(rdev
->bdev
, b
),
2612 (unsigned long long)r10_bio
->sector
);
2613 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2616 r10_bio
->sector
- bio
->bi_sector
,
2618 r10_bio
->devs
[slot
].bio
= bio
;
2619 r10_bio
->devs
[slot
].rdev
= rdev
;
2620 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2621 + choose_data_offset(r10_bio
, rdev
);
2622 bio
->bi_bdev
= rdev
->bdev
;
2623 bio
->bi_rw
= READ
| do_sync
;
2624 bio
->bi_private
= r10_bio
;
2625 bio
->bi_end_io
= raid10_end_read_request
;
2626 if (max_sectors
< r10_bio
->sectors
) {
2627 /* Drat - have to split this up more */
2628 struct bio
*mbio
= r10_bio
->master_bio
;
2629 int sectors_handled
=
2630 r10_bio
->sector
+ max_sectors
2632 r10_bio
->sectors
= max_sectors
;
2633 spin_lock_irq(&conf
->device_lock
);
2634 if (mbio
->bi_phys_segments
== 0)
2635 mbio
->bi_phys_segments
= 2;
2637 mbio
->bi_phys_segments
++;
2638 spin_unlock_irq(&conf
->device_lock
);
2639 generic_make_request(bio
);
2641 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2643 r10_bio
->master_bio
= mbio
;
2644 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2647 set_bit(R10BIO_ReadError
,
2649 r10_bio
->mddev
= mddev
;
2650 r10_bio
->sector
= mbio
->bi_sector
2655 generic_make_request(bio
);
2658 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2660 /* Some sort of write request has finished and it
2661 * succeeded in writing where we thought there was a
2662 * bad block. So forget the bad block.
2663 * Or possibly if failed and we need to record
2667 struct md_rdev
*rdev
;
2669 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2670 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2671 for (m
= 0; m
< conf
->copies
; m
++) {
2672 int dev
= r10_bio
->devs
[m
].devnum
;
2673 rdev
= conf
->mirrors
[dev
].rdev
;
2674 if (r10_bio
->devs
[m
].bio
== NULL
)
2676 if (test_bit(BIO_UPTODATE
,
2677 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2678 rdev_clear_badblocks(
2680 r10_bio
->devs
[m
].addr
,
2681 r10_bio
->sectors
, 0);
2683 if (!rdev_set_badblocks(
2685 r10_bio
->devs
[m
].addr
,
2686 r10_bio
->sectors
, 0))
2687 md_error(conf
->mddev
, rdev
);
2689 rdev
= conf
->mirrors
[dev
].replacement
;
2690 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2692 if (test_bit(BIO_UPTODATE
,
2693 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2694 rdev_clear_badblocks(
2696 r10_bio
->devs
[m
].addr
,
2697 r10_bio
->sectors
, 0);
2699 if (!rdev_set_badblocks(
2701 r10_bio
->devs
[m
].addr
,
2702 r10_bio
->sectors
, 0))
2703 md_error(conf
->mddev
, rdev
);
2708 for (m
= 0; m
< conf
->copies
; m
++) {
2709 int dev
= r10_bio
->devs
[m
].devnum
;
2710 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2711 rdev
= conf
->mirrors
[dev
].rdev
;
2712 if (bio
== IO_MADE_GOOD
) {
2713 rdev_clear_badblocks(
2715 r10_bio
->devs
[m
].addr
,
2716 r10_bio
->sectors
, 0);
2717 rdev_dec_pending(rdev
, conf
->mddev
);
2718 } else if (bio
!= NULL
&&
2719 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2720 if (!narrow_write_error(r10_bio
, m
)) {
2721 md_error(conf
->mddev
, rdev
);
2722 set_bit(R10BIO_Degraded
,
2725 rdev_dec_pending(rdev
, conf
->mddev
);
2727 bio
= r10_bio
->devs
[m
].repl_bio
;
2728 rdev
= conf
->mirrors
[dev
].replacement
;
2729 if (rdev
&& bio
== IO_MADE_GOOD
) {
2730 rdev_clear_badblocks(
2732 r10_bio
->devs
[m
].addr
,
2733 r10_bio
->sectors
, 0);
2734 rdev_dec_pending(rdev
, conf
->mddev
);
2737 if (test_bit(R10BIO_WriteError
,
2739 close_write(r10_bio
);
2740 raid_end_bio_io(r10_bio
);
2744 static void raid10d(struct md_thread
*thread
)
2746 struct mddev
*mddev
= thread
->mddev
;
2747 struct r10bio
*r10_bio
;
2748 unsigned long flags
;
2749 struct r10conf
*conf
= mddev
->private;
2750 struct list_head
*head
= &conf
->retry_list
;
2751 struct blk_plug plug
;
2753 md_check_recovery(mddev
);
2755 blk_start_plug(&plug
);
2758 flush_pending_writes(conf
);
2760 spin_lock_irqsave(&conf
->device_lock
, flags
);
2761 if (list_empty(head
)) {
2762 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2765 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2766 list_del(head
->prev
);
2768 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2770 mddev
= r10_bio
->mddev
;
2771 conf
= mddev
->private;
2772 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2773 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2774 handle_write_completed(conf
, r10_bio
);
2775 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2776 reshape_request_write(mddev
, r10_bio
);
2777 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2778 sync_request_write(mddev
, r10_bio
);
2779 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2780 recovery_request_write(mddev
, r10_bio
);
2781 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2782 handle_read_error(mddev
, r10_bio
);
2784 /* just a partial read to be scheduled from a
2787 int slot
= r10_bio
->read_slot
;
2788 generic_make_request(r10_bio
->devs
[slot
].bio
);
2792 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2793 md_check_recovery(mddev
);
2795 blk_finish_plug(&plug
);
2799 static int init_resync(struct r10conf
*conf
)
2804 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2805 BUG_ON(conf
->r10buf_pool
);
2806 conf
->have_replacement
= 0;
2807 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2808 if (conf
->mirrors
[i
].replacement
)
2809 conf
->have_replacement
= 1;
2810 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2811 if (!conf
->r10buf_pool
)
2813 conf
->next_resync
= 0;
2818 * perform a "sync" on one "block"
2820 * We need to make sure that no normal I/O request - particularly write
2821 * requests - conflict with active sync requests.
2823 * This is achieved by tracking pending requests and a 'barrier' concept
2824 * that can be installed to exclude normal IO requests.
2826 * Resync and recovery are handled very differently.
2827 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2829 * For resync, we iterate over virtual addresses, read all copies,
2830 * and update if there are differences. If only one copy is live,
2832 * For recovery, we iterate over physical addresses, read a good
2833 * value for each non-in_sync drive, and over-write.
2835 * So, for recovery we may have several outstanding complex requests for a
2836 * given address, one for each out-of-sync device. We model this by allocating
2837 * a number of r10_bio structures, one for each out-of-sync device.
2838 * As we setup these structures, we collect all bio's together into a list
2839 * which we then process collectively to add pages, and then process again
2840 * to pass to generic_make_request.
2842 * The r10_bio structures are linked using a borrowed master_bio pointer.
2843 * This link is counted in ->remaining. When the r10_bio that points to NULL
2844 * has its remaining count decremented to 0, the whole complex operation
2849 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2850 int *skipped
, int go_faster
)
2852 struct r10conf
*conf
= mddev
->private;
2853 struct r10bio
*r10_bio
;
2854 struct bio
*biolist
= NULL
, *bio
;
2855 sector_t max_sector
, nr_sectors
;
2858 sector_t sync_blocks
;
2859 sector_t sectors_skipped
= 0;
2860 int chunks_skipped
= 0;
2861 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2863 if (!conf
->r10buf_pool
)
2864 if (init_resync(conf
))
2868 max_sector
= mddev
->dev_sectors
;
2869 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2870 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2871 max_sector
= mddev
->resync_max_sectors
;
2872 if (sector_nr
>= max_sector
) {
2873 /* If we aborted, we need to abort the
2874 * sync on the 'current' bitmap chucks (there can
2875 * be several when recovering multiple devices).
2876 * as we may have started syncing it but not finished.
2877 * We can find the current address in
2878 * mddev->curr_resync, but for recovery,
2879 * we need to convert that to several
2880 * virtual addresses.
2882 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2887 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2888 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2889 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2891 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2893 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2894 bitmap_end_sync(mddev
->bitmap
, sect
,
2898 /* completed sync */
2899 if ((!mddev
->bitmap
|| conf
->fullsync
)
2900 && conf
->have_replacement
2901 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2902 /* Completed a full sync so the replacements
2903 * are now fully recovered.
2905 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2906 if (conf
->mirrors
[i
].replacement
)
2907 conf
->mirrors
[i
].replacement
2913 bitmap_close_sync(mddev
->bitmap
);
2916 return sectors_skipped
;
2919 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2920 return reshape_request(mddev
, sector_nr
, skipped
);
2922 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2923 /* if there has been nothing to do on any drive,
2924 * then there is nothing to do at all..
2927 return (max_sector
- sector_nr
) + sectors_skipped
;
2930 if (max_sector
> mddev
->resync_max
)
2931 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2933 /* make sure whole request will fit in a chunk - if chunks
2936 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2937 max_sector
> (sector_nr
| chunk_mask
))
2938 max_sector
= (sector_nr
| chunk_mask
) + 1;
2940 * If there is non-resync activity waiting for us then
2941 * put in a delay to throttle resync.
2943 if (!go_faster
&& conf
->nr_waiting
)
2944 msleep_interruptible(1000);
2946 /* Again, very different code for resync and recovery.
2947 * Both must result in an r10bio with a list of bios that
2948 * have bi_end_io, bi_sector, bi_bdev set,
2949 * and bi_private set to the r10bio.
2950 * For recovery, we may actually create several r10bios
2951 * with 2 bios in each, that correspond to the bios in the main one.
2952 * In this case, the subordinate r10bios link back through a
2953 * borrowed master_bio pointer, and the counter in the master
2954 * includes a ref from each subordinate.
2956 /* First, we decide what to do and set ->bi_end_io
2957 * To end_sync_read if we want to read, and
2958 * end_sync_write if we will want to write.
2961 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2962 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2963 /* recovery... the complicated one */
2967 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2973 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2975 if ((mirror
->rdev
== NULL
||
2976 test_bit(In_sync
, &mirror
->rdev
->flags
))
2978 (mirror
->replacement
== NULL
||
2980 &mirror
->replacement
->flags
)))
2984 /* want to reconstruct this device */
2986 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2987 if (sect
>= mddev
->resync_max_sectors
) {
2988 /* last stripe is not complete - don't
2989 * try to recover this sector.
2993 /* Unless we are doing a full sync, or a replacement
2994 * we only need to recover the block if it is set in
2997 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2999 if (sync_blocks
< max_sync
)
3000 max_sync
= sync_blocks
;
3002 mirror
->replacement
== NULL
&&
3004 /* yep, skip the sync_blocks here, but don't assume
3005 * that there will never be anything to do here
3007 chunks_skipped
= -1;
3011 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3012 raise_barrier(conf
, rb2
!= NULL
);
3013 atomic_set(&r10_bio
->remaining
, 0);
3015 r10_bio
->master_bio
= (struct bio
*)rb2
;
3017 atomic_inc(&rb2
->remaining
);
3018 r10_bio
->mddev
= mddev
;
3019 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3020 r10_bio
->sector
= sect
;
3022 raid10_find_phys(conf
, r10_bio
);
3024 /* Need to check if the array will still be
3027 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3028 if (conf
->mirrors
[j
].rdev
== NULL
||
3029 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3034 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3035 &sync_blocks
, still_degraded
);
3038 for (j
=0; j
<conf
->copies
;j
++) {
3040 int d
= r10_bio
->devs
[j
].devnum
;
3041 sector_t from_addr
, to_addr
;
3042 struct md_rdev
*rdev
;
3043 sector_t sector
, first_bad
;
3045 if (!conf
->mirrors
[d
].rdev
||
3046 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3048 /* This is where we read from */
3050 rdev
= conf
->mirrors
[d
].rdev
;
3051 sector
= r10_bio
->devs
[j
].addr
;
3053 if (is_badblock(rdev
, sector
, max_sync
,
3054 &first_bad
, &bad_sectors
)) {
3055 if (first_bad
> sector
)
3056 max_sync
= first_bad
- sector
;
3058 bad_sectors
-= (sector
3060 if (max_sync
> bad_sectors
)
3061 max_sync
= bad_sectors
;
3065 bio
= r10_bio
->devs
[0].bio
;
3066 bio
->bi_next
= biolist
;
3068 bio
->bi_private
= r10_bio
;
3069 bio
->bi_end_io
= end_sync_read
;
3071 from_addr
= r10_bio
->devs
[j
].addr
;
3072 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
3073 bio
->bi_bdev
= rdev
->bdev
;
3074 atomic_inc(&rdev
->nr_pending
);
3075 /* and we write to 'i' (if not in_sync) */
3077 for (k
=0; k
<conf
->copies
; k
++)
3078 if (r10_bio
->devs
[k
].devnum
== i
)
3080 BUG_ON(k
== conf
->copies
);
3081 to_addr
= r10_bio
->devs
[k
].addr
;
3082 r10_bio
->devs
[0].devnum
= d
;
3083 r10_bio
->devs
[0].addr
= from_addr
;
3084 r10_bio
->devs
[1].devnum
= i
;
3085 r10_bio
->devs
[1].addr
= to_addr
;
3087 rdev
= mirror
->rdev
;
3088 if (!test_bit(In_sync
, &rdev
->flags
)) {
3089 bio
= r10_bio
->devs
[1].bio
;
3090 bio
->bi_next
= biolist
;
3092 bio
->bi_private
= r10_bio
;
3093 bio
->bi_end_io
= end_sync_write
;
3095 bio
->bi_sector
= to_addr
3096 + rdev
->data_offset
;
3097 bio
->bi_bdev
= rdev
->bdev
;
3098 atomic_inc(&r10_bio
->remaining
);
3100 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3102 /* and maybe write to replacement */
3103 bio
= r10_bio
->devs
[1].repl_bio
;
3105 bio
->bi_end_io
= NULL
;
3106 rdev
= mirror
->replacement
;
3107 /* Note: if rdev != NULL, then bio
3108 * cannot be NULL as r10buf_pool_alloc will
3109 * have allocated it.
3110 * So the second test here is pointless.
3111 * But it keeps semantic-checkers happy, and
3112 * this comment keeps human reviewers
3115 if (rdev
== NULL
|| bio
== NULL
||
3116 test_bit(Faulty
, &rdev
->flags
))
3118 bio
->bi_next
= biolist
;
3120 bio
->bi_private
= r10_bio
;
3121 bio
->bi_end_io
= end_sync_write
;
3123 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3124 bio
->bi_bdev
= rdev
->bdev
;
3125 atomic_inc(&r10_bio
->remaining
);
3128 if (j
== conf
->copies
) {
3129 /* Cannot recover, so abort the recovery or
3130 * record a bad block */
3133 atomic_dec(&rb2
->remaining
);
3136 /* problem is that there are bad blocks
3137 * on other device(s)
3140 for (k
= 0; k
< conf
->copies
; k
++)
3141 if (r10_bio
->devs
[k
].devnum
== i
)
3143 if (!test_bit(In_sync
,
3144 &mirror
->rdev
->flags
)
3145 && !rdev_set_badblocks(
3147 r10_bio
->devs
[k
].addr
,
3150 if (mirror
->replacement
&&
3151 !rdev_set_badblocks(
3152 mirror
->replacement
,
3153 r10_bio
->devs
[k
].addr
,
3158 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3160 printk(KERN_INFO
"md/raid10:%s: insufficient "
3161 "working devices for recovery.\n",
3163 mirror
->recovery_disabled
3164 = mddev
->recovery_disabled
;
3169 if (biolist
== NULL
) {
3171 struct r10bio
*rb2
= r10_bio
;
3172 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3173 rb2
->master_bio
= NULL
;
3179 /* resync. Schedule a read for every block at this virt offset */
3182 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3184 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3185 &sync_blocks
, mddev
->degraded
) &&
3186 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3187 &mddev
->recovery
)) {
3188 /* We can skip this block */
3190 return sync_blocks
+ sectors_skipped
;
3192 if (sync_blocks
< max_sync
)
3193 max_sync
= sync_blocks
;
3194 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3196 r10_bio
->mddev
= mddev
;
3197 atomic_set(&r10_bio
->remaining
, 0);
3198 raise_barrier(conf
, 0);
3199 conf
->next_resync
= sector_nr
;
3201 r10_bio
->master_bio
= NULL
;
3202 r10_bio
->sector
= sector_nr
;
3203 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3204 raid10_find_phys(conf
, r10_bio
);
3205 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3207 for (i
= 0; i
< conf
->copies
; i
++) {
3208 int d
= r10_bio
->devs
[i
].devnum
;
3209 sector_t first_bad
, sector
;
3212 if (r10_bio
->devs
[i
].repl_bio
)
3213 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3215 bio
= r10_bio
->devs
[i
].bio
;
3216 bio
->bi_end_io
= NULL
;
3217 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3218 if (conf
->mirrors
[d
].rdev
== NULL
||
3219 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3221 sector
= r10_bio
->devs
[i
].addr
;
3222 if (is_badblock(conf
->mirrors
[d
].rdev
,
3224 &first_bad
, &bad_sectors
)) {
3225 if (first_bad
> sector
)
3226 max_sync
= first_bad
- sector
;
3228 bad_sectors
-= (sector
- first_bad
);
3229 if (max_sync
> bad_sectors
)
3230 max_sync
= bad_sectors
;
3234 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3235 atomic_inc(&r10_bio
->remaining
);
3236 bio
->bi_next
= biolist
;
3238 bio
->bi_private
= r10_bio
;
3239 bio
->bi_end_io
= end_sync_read
;
3241 bio
->bi_sector
= sector
+
3242 conf
->mirrors
[d
].rdev
->data_offset
;
3243 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3246 if (conf
->mirrors
[d
].replacement
== NULL
||
3248 &conf
->mirrors
[d
].replacement
->flags
))
3251 /* Need to set up for writing to the replacement */
3252 bio
= r10_bio
->devs
[i
].repl_bio
;
3253 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3255 sector
= r10_bio
->devs
[i
].addr
;
3256 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3257 bio
->bi_next
= biolist
;
3259 bio
->bi_private
= r10_bio
;
3260 bio
->bi_end_io
= end_sync_write
;
3262 bio
->bi_sector
= sector
+
3263 conf
->mirrors
[d
].replacement
->data_offset
;
3264 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3269 for (i
=0; i
<conf
->copies
; i
++) {
3270 int d
= r10_bio
->devs
[i
].devnum
;
3271 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3272 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3274 if (r10_bio
->devs
[i
].repl_bio
&&
3275 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3277 conf
->mirrors
[d
].replacement
,
3286 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3288 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
3290 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
3293 bio
->bi_phys_segments
= 0;
3298 if (sector_nr
+ max_sync
< max_sector
)
3299 max_sector
= sector_nr
+ max_sync
;
3302 int len
= PAGE_SIZE
;
3303 if (sector_nr
+ (len
>>9) > max_sector
)
3304 len
= (max_sector
- sector_nr
) << 9;
3307 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3309 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3310 if (bio_add_page(bio
, page
, len
, 0))
3314 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3315 for (bio2
= biolist
;
3316 bio2
&& bio2
!= bio
;
3317 bio2
= bio2
->bi_next
) {
3318 /* remove last page from this bio */
3320 bio2
->bi_size
-= len
;
3321 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3325 nr_sectors
+= len
>>9;
3326 sector_nr
+= len
>>9;
3327 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3329 r10_bio
->sectors
= nr_sectors
;
3333 biolist
= biolist
->bi_next
;
3335 bio
->bi_next
= NULL
;
3336 r10_bio
= bio
->bi_private
;
3337 r10_bio
->sectors
= nr_sectors
;
3339 if (bio
->bi_end_io
== end_sync_read
) {
3340 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3341 generic_make_request(bio
);
3345 if (sectors_skipped
)
3346 /* pretend they weren't skipped, it makes
3347 * no important difference in this case
3349 md_done_sync(mddev
, sectors_skipped
, 1);
3351 return sectors_skipped
+ nr_sectors
;
3353 /* There is nowhere to write, so all non-sync
3354 * drives must be failed or in resync, all drives
3355 * have a bad block, so try the next chunk...
3357 if (sector_nr
+ max_sync
< max_sector
)
3358 max_sector
= sector_nr
+ max_sync
;
3360 sectors_skipped
+= (max_sector
- sector_nr
);
3362 sector_nr
= max_sector
;
3367 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3370 struct r10conf
*conf
= mddev
->private;
3373 raid_disks
= min(conf
->geo
.raid_disks
,
3374 conf
->prev
.raid_disks
);
3376 sectors
= conf
->dev_sectors
;
3378 size
= sectors
>> conf
->geo
.chunk_shift
;
3379 sector_div(size
, conf
->geo
.far_copies
);
3380 size
= size
* raid_disks
;
3381 sector_div(size
, conf
->geo
.near_copies
);
3383 return size
<< conf
->geo
.chunk_shift
;
3386 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3388 /* Calculate the number of sectors-per-device that will
3389 * actually be used, and set conf->dev_sectors and
3393 size
= size
>> conf
->geo
.chunk_shift
;
3394 sector_div(size
, conf
->geo
.far_copies
);
3395 size
= size
* conf
->geo
.raid_disks
;
3396 sector_div(size
, conf
->geo
.near_copies
);
3397 /* 'size' is now the number of chunks in the array */
3398 /* calculate "used chunks per device" */
3399 size
= size
* conf
->copies
;
3401 /* We need to round up when dividing by raid_disks to
3402 * get the stride size.
3404 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3406 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3408 if (conf
->geo
.far_offset
)
3409 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3411 sector_div(size
, conf
->geo
.far_copies
);
3412 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3416 enum geo_type
{geo_new
, geo_old
, geo_start
};
3417 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3420 int layout
, chunk
, disks
;
3423 layout
= mddev
->layout
;
3424 chunk
= mddev
->chunk_sectors
;
3425 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3428 layout
= mddev
->new_layout
;
3429 chunk
= mddev
->new_chunk_sectors
;
3430 disks
= mddev
->raid_disks
;
3432 default: /* avoid 'may be unused' warnings */
3433 case geo_start
: /* new when starting reshape - raid_disks not
3435 layout
= mddev
->new_layout
;
3436 chunk
= mddev
->new_chunk_sectors
;
3437 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3442 if (chunk
< (PAGE_SIZE
>> 9) ||
3443 !is_power_of_2(chunk
))
3446 fc
= (layout
>> 8) & 255;
3447 fo
= layout
& (1<<16);
3448 geo
->raid_disks
= disks
;
3449 geo
->near_copies
= nc
;
3450 geo
->far_copies
= fc
;
3451 geo
->far_offset
= fo
;
3452 geo
->chunk_mask
= chunk
- 1;
3453 geo
->chunk_shift
= ffz(~chunk
);
3457 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3459 struct r10conf
*conf
= NULL
;
3464 copies
= setup_geo(&geo
, mddev
, geo_new
);
3467 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3468 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3469 mdname(mddev
), PAGE_SIZE
);
3473 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3474 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3475 mdname(mddev
), mddev
->new_layout
);
3480 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3484 /* FIXME calc properly */
3485 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3486 max(0,mddev
->delta_disks
)),
3491 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3496 conf
->copies
= copies
;
3497 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3498 r10bio_pool_free
, conf
);
3499 if (!conf
->r10bio_pool
)
3502 calc_sectors(conf
, mddev
->dev_sectors
);
3503 if (mddev
->reshape_position
== MaxSector
) {
3504 conf
->prev
= conf
->geo
;
3505 conf
->reshape_progress
= MaxSector
;
3507 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3511 conf
->reshape_progress
= mddev
->reshape_position
;
3512 if (conf
->prev
.far_offset
)
3513 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3515 /* far_copies must be 1 */
3516 conf
->prev
.stride
= conf
->dev_sectors
;
3518 spin_lock_init(&conf
->device_lock
);
3519 INIT_LIST_HEAD(&conf
->retry_list
);
3521 spin_lock_init(&conf
->resync_lock
);
3522 init_waitqueue_head(&conf
->wait_barrier
);
3524 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3528 conf
->mddev
= mddev
;
3533 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3536 if (conf
->r10bio_pool
)
3537 mempool_destroy(conf
->r10bio_pool
);
3538 kfree(conf
->mirrors
);
3539 safe_put_page(conf
->tmppage
);
3542 return ERR_PTR(err
);
3545 static int run(struct mddev
*mddev
)
3547 struct r10conf
*conf
;
3548 int i
, disk_idx
, chunk_size
;
3549 struct raid10_info
*disk
;
3550 struct md_rdev
*rdev
;
3552 sector_t min_offset_diff
= 0;
3554 bool discard_supported
= false;
3556 if (mddev
->private == NULL
) {
3557 conf
= setup_conf(mddev
);
3559 return PTR_ERR(conf
);
3560 mddev
->private = conf
;
3562 conf
= mddev
->private;
3566 mddev
->thread
= conf
->thread
;
3567 conf
->thread
= NULL
;
3569 chunk_size
= mddev
->chunk_sectors
<< 9;
3571 blk_queue_max_discard_sectors(mddev
->queue
,
3572 mddev
->chunk_sectors
);
3573 blk_queue_io_min(mddev
->queue
, chunk_size
);
3574 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3575 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3577 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3578 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3581 rdev_for_each(rdev
, mddev
) {
3583 struct request_queue
*q
;
3585 disk_idx
= rdev
->raid_disk
;
3588 if (disk_idx
>= conf
->geo
.raid_disks
&&
3589 disk_idx
>= conf
->prev
.raid_disks
)
3591 disk
= conf
->mirrors
+ disk_idx
;
3593 if (test_bit(Replacement
, &rdev
->flags
)) {
3594 if (disk
->replacement
)
3596 disk
->replacement
= rdev
;
3602 q
= bdev_get_queue(rdev
->bdev
);
3603 if (q
->merge_bvec_fn
)
3604 mddev
->merge_check_needed
= 1;
3605 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3606 if (!mddev
->reshape_backwards
)
3610 if (first
|| diff
< min_offset_diff
)
3611 min_offset_diff
= diff
;
3614 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3615 rdev
->data_offset
<< 9);
3617 disk
->head_position
= 0;
3619 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3620 discard_supported
= true;
3624 if (discard_supported
)
3625 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3628 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3631 /* need to check that every block has at least one working mirror */
3632 if (!enough(conf
, -1)) {
3633 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3638 if (conf
->reshape_progress
!= MaxSector
) {
3639 /* must ensure that shape change is supported */
3640 if (conf
->geo
.far_copies
!= 1 &&
3641 conf
->geo
.far_offset
== 0)
3643 if (conf
->prev
.far_copies
!= 1 &&
3644 conf
->geo
.far_offset
== 0)
3648 mddev
->degraded
= 0;
3650 i
< conf
->geo
.raid_disks
3651 || i
< conf
->prev
.raid_disks
;
3654 disk
= conf
->mirrors
+ i
;
3656 if (!disk
->rdev
&& disk
->replacement
) {
3657 /* The replacement is all we have - use it */
3658 disk
->rdev
= disk
->replacement
;
3659 disk
->replacement
= NULL
;
3660 clear_bit(Replacement
, &disk
->rdev
->flags
);
3664 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3665 disk
->head_position
= 0;
3670 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3673 if (mddev
->recovery_cp
!= MaxSector
)
3674 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3675 " -- starting background reconstruction\n",
3678 "md/raid10:%s: active with %d out of %d devices\n",
3679 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3680 conf
->geo
.raid_disks
);
3682 * Ok, everything is just fine now
3684 mddev
->dev_sectors
= conf
->dev_sectors
;
3685 size
= raid10_size(mddev
, 0, 0);
3686 md_set_array_sectors(mddev
, size
);
3687 mddev
->resync_max_sectors
= size
;
3690 int stripe
= conf
->geo
.raid_disks
*
3691 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3692 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3693 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3695 /* Calculate max read-ahead size.
3696 * We need to readahead at least twice a whole stripe....
3699 stripe
/= conf
->geo
.near_copies
;
3700 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3701 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3702 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3706 if (md_integrity_register(mddev
))
3709 if (conf
->reshape_progress
!= MaxSector
) {
3710 unsigned long before_length
, after_length
;
3712 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3713 conf
->prev
.far_copies
);
3714 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3715 conf
->geo
.far_copies
);
3717 if (max(before_length
, after_length
) > min_offset_diff
) {
3718 /* This cannot work */
3719 printk("md/raid10: offset difference not enough to continue reshape\n");
3722 conf
->offset_diff
= min_offset_diff
;
3724 conf
->reshape_safe
= conf
->reshape_progress
;
3725 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3726 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3727 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3728 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3729 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3736 md_unregister_thread(&mddev
->thread
);
3737 if (conf
->r10bio_pool
)
3738 mempool_destroy(conf
->r10bio_pool
);
3739 safe_put_page(conf
->tmppage
);
3740 kfree(conf
->mirrors
);
3742 mddev
->private = NULL
;
3747 static int stop(struct mddev
*mddev
)
3749 struct r10conf
*conf
= mddev
->private;
3751 raise_barrier(conf
, 0);
3752 lower_barrier(conf
);
3754 md_unregister_thread(&mddev
->thread
);
3756 /* the unplug fn references 'conf'*/
3757 blk_sync_queue(mddev
->queue
);
3759 if (conf
->r10bio_pool
)
3760 mempool_destroy(conf
->r10bio_pool
);
3761 kfree(conf
->mirrors
);
3763 mddev
->private = NULL
;
3767 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3769 struct r10conf
*conf
= mddev
->private;
3773 raise_barrier(conf
, 0);
3776 lower_barrier(conf
);
3781 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3783 /* Resize of 'far' arrays is not supported.
3784 * For 'near' and 'offset' arrays we can set the
3785 * number of sectors used to be an appropriate multiple
3786 * of the chunk size.
3787 * For 'offset', this is far_copies*chunksize.
3788 * For 'near' the multiplier is the LCM of
3789 * near_copies and raid_disks.
3790 * So if far_copies > 1 && !far_offset, fail.
3791 * Else find LCM(raid_disks, near_copy)*far_copies and
3792 * multiply by chunk_size. Then round to this number.
3793 * This is mostly done by raid10_size()
3795 struct r10conf
*conf
= mddev
->private;
3796 sector_t oldsize
, size
;
3798 if (mddev
->reshape_position
!= MaxSector
)
3801 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3804 oldsize
= raid10_size(mddev
, 0, 0);
3805 size
= raid10_size(mddev
, sectors
, 0);
3806 if (mddev
->external_size
&&
3807 mddev
->array_sectors
> size
)
3809 if (mddev
->bitmap
) {
3810 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3814 md_set_array_sectors(mddev
, size
);
3815 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3816 revalidate_disk(mddev
->gendisk
);
3817 if (sectors
> mddev
->dev_sectors
&&
3818 mddev
->recovery_cp
> oldsize
) {
3819 mddev
->recovery_cp
= oldsize
;
3820 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3822 calc_sectors(conf
, sectors
);
3823 mddev
->dev_sectors
= conf
->dev_sectors
;
3824 mddev
->resync_max_sectors
= size
;
3828 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3830 struct md_rdev
*rdev
;
3831 struct r10conf
*conf
;
3833 if (mddev
->degraded
> 0) {
3834 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3836 return ERR_PTR(-EINVAL
);
3839 /* Set new parameters */
3840 mddev
->new_level
= 10;
3841 /* new layout: far_copies = 1, near_copies = 2 */
3842 mddev
->new_layout
= (1<<8) + 2;
3843 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3844 mddev
->delta_disks
= mddev
->raid_disks
;
3845 mddev
->raid_disks
*= 2;
3846 /* make sure it will be not marked as dirty */
3847 mddev
->recovery_cp
= MaxSector
;
3849 conf
= setup_conf(mddev
);
3850 if (!IS_ERR(conf
)) {
3851 rdev_for_each(rdev
, mddev
)
3852 if (rdev
->raid_disk
>= 0)
3853 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3860 static void *raid10_takeover(struct mddev
*mddev
)
3862 struct r0conf
*raid0_conf
;
3864 /* raid10 can take over:
3865 * raid0 - providing it has only two drives
3867 if (mddev
->level
== 0) {
3868 /* for raid0 takeover only one zone is supported */
3869 raid0_conf
= mddev
->private;
3870 if (raid0_conf
->nr_strip_zones
> 1) {
3871 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3872 " with more than one zone.\n",
3874 return ERR_PTR(-EINVAL
);
3876 return raid10_takeover_raid0(mddev
);
3878 return ERR_PTR(-EINVAL
);
3881 static int raid10_check_reshape(struct mddev
*mddev
)
3883 /* Called when there is a request to change
3884 * - layout (to ->new_layout)
3885 * - chunk size (to ->new_chunk_sectors)
3886 * - raid_disks (by delta_disks)
3887 * or when trying to restart a reshape that was ongoing.
3889 * We need to validate the request and possibly allocate
3890 * space if that might be an issue later.
3892 * Currently we reject any reshape of a 'far' mode array,
3893 * allow chunk size to change if new is generally acceptable,
3894 * allow raid_disks to increase, and allow
3895 * a switch between 'near' mode and 'offset' mode.
3897 struct r10conf
*conf
= mddev
->private;
3900 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3903 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3904 /* mustn't change number of copies */
3906 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3907 /* Cannot switch to 'far' mode */
3910 if (mddev
->array_sectors
& geo
.chunk_mask
)
3911 /* not factor of array size */
3914 if (!enough(conf
, -1))
3917 kfree(conf
->mirrors_new
);
3918 conf
->mirrors_new
= NULL
;
3919 if (mddev
->delta_disks
> 0) {
3920 /* allocate new 'mirrors' list */
3921 conf
->mirrors_new
= kzalloc(
3922 sizeof(struct raid10_info
)
3923 *(mddev
->raid_disks
+
3924 mddev
->delta_disks
),
3926 if (!conf
->mirrors_new
)
3933 * Need to check if array has failed when deciding whether to:
3935 * - remove non-faulty devices
3938 * This determination is simple when no reshape is happening.
3939 * However if there is a reshape, we need to carefully check
3940 * both the before and after sections.
3941 * This is because some failed devices may only affect one
3942 * of the two sections, and some non-in_sync devices may
3943 * be insync in the section most affected by failed devices.
3945 static int calc_degraded(struct r10conf
*conf
)
3947 int degraded
, degraded2
;
3952 /* 'prev' section first */
3953 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
3954 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3955 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3957 else if (!test_bit(In_sync
, &rdev
->flags
))
3958 /* When we can reduce the number of devices in
3959 * an array, this might not contribute to
3960 * 'degraded'. It does now.
3965 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
3969 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3970 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3971 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3973 else if (!test_bit(In_sync
, &rdev
->flags
)) {
3974 /* If reshape is increasing the number of devices,
3975 * this section has already been recovered, so
3976 * it doesn't contribute to degraded.
3979 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
3984 if (degraded2
> degraded
)
3989 static int raid10_start_reshape(struct mddev
*mddev
)
3991 /* A 'reshape' has been requested. This commits
3992 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3993 * This also checks if there are enough spares and adds them
3995 * We currently require enough spares to make the final
3996 * array non-degraded. We also require that the difference
3997 * between old and new data_offset - on each device - is
3998 * enough that we never risk over-writing.
4001 unsigned long before_length
, after_length
;
4002 sector_t min_offset_diff
= 0;
4005 struct r10conf
*conf
= mddev
->private;
4006 struct md_rdev
*rdev
;
4010 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4013 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4016 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4017 conf
->prev
.far_copies
);
4018 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4019 conf
->geo
.far_copies
);
4021 rdev_for_each(rdev
, mddev
) {
4022 if (!test_bit(In_sync
, &rdev
->flags
)
4023 && !test_bit(Faulty
, &rdev
->flags
))
4025 if (rdev
->raid_disk
>= 0) {
4026 long long diff
= (rdev
->new_data_offset
4027 - rdev
->data_offset
);
4028 if (!mddev
->reshape_backwards
)
4032 if (first
|| diff
< min_offset_diff
)
4033 min_offset_diff
= diff
;
4037 if (max(before_length
, after_length
) > min_offset_diff
)
4040 if (spares
< mddev
->delta_disks
)
4043 conf
->offset_diff
= min_offset_diff
;
4044 spin_lock_irq(&conf
->device_lock
);
4045 if (conf
->mirrors_new
) {
4046 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4047 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4049 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
4050 conf
->mirrors_old
= conf
->mirrors
;
4051 conf
->mirrors
= conf
->mirrors_new
;
4052 conf
->mirrors_new
= NULL
;
4054 setup_geo(&conf
->geo
, mddev
, geo_start
);
4056 if (mddev
->reshape_backwards
) {
4057 sector_t size
= raid10_size(mddev
, 0, 0);
4058 if (size
< mddev
->array_sectors
) {
4059 spin_unlock_irq(&conf
->device_lock
);
4060 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4064 mddev
->resync_max_sectors
= size
;
4065 conf
->reshape_progress
= size
;
4067 conf
->reshape_progress
= 0;
4068 spin_unlock_irq(&conf
->device_lock
);
4070 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4071 ret
= bitmap_resize(mddev
->bitmap
,
4072 raid10_size(mddev
, 0,
4073 conf
->geo
.raid_disks
),
4078 if (mddev
->delta_disks
> 0) {
4079 rdev_for_each(rdev
, mddev
)
4080 if (rdev
->raid_disk
< 0 &&
4081 !test_bit(Faulty
, &rdev
->flags
)) {
4082 if (raid10_add_disk(mddev
, rdev
) == 0) {
4083 if (rdev
->raid_disk
>=
4084 conf
->prev
.raid_disks
)
4085 set_bit(In_sync
, &rdev
->flags
);
4087 rdev
->recovery_offset
= 0;
4089 if (sysfs_link_rdev(mddev
, rdev
))
4090 /* Failure here is OK */;
4092 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4093 && !test_bit(Faulty
, &rdev
->flags
)) {
4094 /* This is a spare that was manually added */
4095 set_bit(In_sync
, &rdev
->flags
);
4098 /* When a reshape changes the number of devices,
4099 * ->degraded is measured against the larger of the
4100 * pre and post numbers.
4102 spin_lock_irq(&conf
->device_lock
);
4103 mddev
->degraded
= calc_degraded(conf
);
4104 spin_unlock_irq(&conf
->device_lock
);
4105 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4106 mddev
->reshape_position
= conf
->reshape_progress
;
4107 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4109 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4110 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4111 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4112 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4114 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4116 if (!mddev
->sync_thread
) {
4120 conf
->reshape_checkpoint
= jiffies
;
4121 md_wakeup_thread(mddev
->sync_thread
);
4122 md_new_event(mddev
);
4126 mddev
->recovery
= 0;
4127 spin_lock_irq(&conf
->device_lock
);
4128 conf
->geo
= conf
->prev
;
4129 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4130 rdev_for_each(rdev
, mddev
)
4131 rdev
->new_data_offset
= rdev
->data_offset
;
4133 conf
->reshape_progress
= MaxSector
;
4134 mddev
->reshape_position
= MaxSector
;
4135 spin_unlock_irq(&conf
->device_lock
);
4139 /* Calculate the last device-address that could contain
4140 * any block from the chunk that includes the array-address 's'
4141 * and report the next address.
4142 * i.e. the address returned will be chunk-aligned and after
4143 * any data that is in the chunk containing 's'.
4145 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4147 s
= (s
| geo
->chunk_mask
) + 1;
4148 s
>>= geo
->chunk_shift
;
4149 s
*= geo
->near_copies
;
4150 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4151 s
*= geo
->far_copies
;
4152 s
<<= geo
->chunk_shift
;
4156 /* Calculate the first device-address that could contain
4157 * any block from the chunk that includes the array-address 's'.
4158 * This too will be the start of a chunk
4160 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4162 s
>>= geo
->chunk_shift
;
4163 s
*= geo
->near_copies
;
4164 sector_div(s
, geo
->raid_disks
);
4165 s
*= geo
->far_copies
;
4166 s
<<= geo
->chunk_shift
;
4170 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4173 /* We simply copy at most one chunk (smallest of old and new)
4174 * at a time, possibly less if that exceeds RESYNC_PAGES,
4175 * or we hit a bad block or something.
4176 * This might mean we pause for normal IO in the middle of
4177 * a chunk, but that is not a problem was mddev->reshape_position
4178 * can record any location.
4180 * If we will want to write to a location that isn't
4181 * yet recorded as 'safe' (i.e. in metadata on disk) then
4182 * we need to flush all reshape requests and update the metadata.
4184 * When reshaping forwards (e.g. to more devices), we interpret
4185 * 'safe' as the earliest block which might not have been copied
4186 * down yet. We divide this by previous stripe size and multiply
4187 * by previous stripe length to get lowest device offset that we
4188 * cannot write to yet.
4189 * We interpret 'sector_nr' as an address that we want to write to.
4190 * From this we use last_device_address() to find where we might
4191 * write to, and first_device_address on the 'safe' position.
4192 * If this 'next' write position is after the 'safe' position,
4193 * we must update the metadata to increase the 'safe' position.
4195 * When reshaping backwards, we round in the opposite direction
4196 * and perform the reverse test: next write position must not be
4197 * less than current safe position.
4199 * In all this the minimum difference in data offsets
4200 * (conf->offset_diff - always positive) allows a bit of slack,
4201 * so next can be after 'safe', but not by more than offset_disk
4203 * We need to prepare all the bios here before we start any IO
4204 * to ensure the size we choose is acceptable to all devices.
4205 * The means one for each copy for write-out and an extra one for
4207 * We store the read-in bio in ->master_bio and the others in
4208 * ->devs[x].bio and ->devs[x].repl_bio.
4210 struct r10conf
*conf
= mddev
->private;
4211 struct r10bio
*r10_bio
;
4212 sector_t next
, safe
, last
;
4216 struct md_rdev
*rdev
;
4219 struct bio
*bio
, *read_bio
;
4220 int sectors_done
= 0;
4222 if (sector_nr
== 0) {
4223 /* If restarting in the middle, skip the initial sectors */
4224 if (mddev
->reshape_backwards
&&
4225 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4226 sector_nr
= (raid10_size(mddev
, 0, 0)
4227 - conf
->reshape_progress
);
4228 } else if (!mddev
->reshape_backwards
&&
4229 conf
->reshape_progress
> 0)
4230 sector_nr
= conf
->reshape_progress
;
4232 mddev
->curr_resync_completed
= sector_nr
;
4233 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4239 /* We don't use sector_nr to track where we are up to
4240 * as that doesn't work well for ->reshape_backwards.
4241 * So just use ->reshape_progress.
4243 if (mddev
->reshape_backwards
) {
4244 /* 'next' is the earliest device address that we might
4245 * write to for this chunk in the new layout
4247 next
= first_dev_address(conf
->reshape_progress
- 1,
4250 /* 'safe' is the last device address that we might read from
4251 * in the old layout after a restart
4253 safe
= last_dev_address(conf
->reshape_safe
- 1,
4256 if (next
+ conf
->offset_diff
< safe
)
4259 last
= conf
->reshape_progress
- 1;
4260 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4261 & conf
->prev
.chunk_mask
);
4262 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4263 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4265 /* 'next' is after the last device address that we
4266 * might write to for this chunk in the new layout
4268 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4270 /* 'safe' is the earliest device address that we might
4271 * read from in the old layout after a restart
4273 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4275 /* Need to update metadata if 'next' might be beyond 'safe'
4276 * as that would possibly corrupt data
4278 if (next
> safe
+ conf
->offset_diff
)
4281 sector_nr
= conf
->reshape_progress
;
4282 last
= sector_nr
| (conf
->geo
.chunk_mask
4283 & conf
->prev
.chunk_mask
);
4285 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4286 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4290 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4291 /* Need to update reshape_position in metadata */
4293 mddev
->reshape_position
= conf
->reshape_progress
;
4294 if (mddev
->reshape_backwards
)
4295 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4296 - conf
->reshape_progress
;
4298 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4299 conf
->reshape_checkpoint
= jiffies
;
4300 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4301 md_wakeup_thread(mddev
->thread
);
4302 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4303 kthread_should_stop());
4304 conf
->reshape_safe
= mddev
->reshape_position
;
4305 allow_barrier(conf
);
4309 /* Now schedule reads for blocks from sector_nr to last */
4310 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4311 raise_barrier(conf
, sectors_done
!= 0);
4312 atomic_set(&r10_bio
->remaining
, 0);
4313 r10_bio
->mddev
= mddev
;
4314 r10_bio
->sector
= sector_nr
;
4315 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4316 r10_bio
->sectors
= last
- sector_nr
+ 1;
4317 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4318 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4321 /* Cannot read from here, so need to record bad blocks
4322 * on all the target devices.
4325 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4326 return sectors_done
;
4329 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4331 read_bio
->bi_bdev
= rdev
->bdev
;
4332 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4333 + rdev
->data_offset
);
4334 read_bio
->bi_private
= r10_bio
;
4335 read_bio
->bi_end_io
= end_sync_read
;
4336 read_bio
->bi_rw
= READ
;
4337 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4338 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4339 read_bio
->bi_vcnt
= 0;
4340 read_bio
->bi_idx
= 0;
4341 read_bio
->bi_size
= 0;
4342 r10_bio
->master_bio
= read_bio
;
4343 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4345 /* Now find the locations in the new layout */
4346 __raid10_find_phys(&conf
->geo
, r10_bio
);
4349 read_bio
->bi_next
= NULL
;
4351 for (s
= 0; s
< conf
->copies
*2; s
++) {
4353 int d
= r10_bio
->devs
[s
/2].devnum
;
4354 struct md_rdev
*rdev2
;
4356 rdev2
= conf
->mirrors
[d
].replacement
;
4357 b
= r10_bio
->devs
[s
/2].repl_bio
;
4359 rdev2
= conf
->mirrors
[d
].rdev
;
4360 b
= r10_bio
->devs
[s
/2].bio
;
4362 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4364 b
->bi_bdev
= rdev2
->bdev
;
4365 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4366 b
->bi_private
= r10_bio
;
4367 b
->bi_end_io
= end_reshape_write
;
4369 b
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4370 b
->bi_flags
|= 1 << BIO_UPTODATE
;
4378 /* Now add as many pages as possible to all of these bios. */
4381 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4382 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4383 int len
= (max_sectors
- s
) << 9;
4384 if (len
> PAGE_SIZE
)
4386 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4388 if (bio_add_page(bio
, page
, len
, 0))
4391 /* Didn't fit, must stop */
4393 bio2
&& bio2
!= bio
;
4394 bio2
= bio2
->bi_next
) {
4395 /* Remove last page from this bio */
4397 bio2
->bi_size
-= len
;
4398 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4402 sector_nr
+= len
>> 9;
4403 nr_sectors
+= len
>> 9;
4406 r10_bio
->sectors
= nr_sectors
;
4408 /* Now submit the read */
4409 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4410 atomic_inc(&r10_bio
->remaining
);
4411 read_bio
->bi_next
= NULL
;
4412 generic_make_request(read_bio
);
4413 sector_nr
+= nr_sectors
;
4414 sectors_done
+= nr_sectors
;
4415 if (sector_nr
<= last
)
4418 /* Now that we have done the whole section we can
4419 * update reshape_progress
4421 if (mddev
->reshape_backwards
)
4422 conf
->reshape_progress
-= sectors_done
;
4424 conf
->reshape_progress
+= sectors_done
;
4426 return sectors_done
;
4429 static void end_reshape_request(struct r10bio
*r10_bio
);
4430 static int handle_reshape_read_error(struct mddev
*mddev
,
4431 struct r10bio
*r10_bio
);
4432 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4434 /* Reshape read completed. Hopefully we have a block
4436 * If we got a read error then we do sync 1-page reads from
4437 * elsewhere until we find the data - or give up.
4439 struct r10conf
*conf
= mddev
->private;
4442 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4443 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4444 /* Reshape has been aborted */
4445 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4449 /* We definitely have the data in the pages, schedule the
4452 atomic_set(&r10_bio
->remaining
, 1);
4453 for (s
= 0; s
< conf
->copies
*2; s
++) {
4455 int d
= r10_bio
->devs
[s
/2].devnum
;
4456 struct md_rdev
*rdev
;
4458 rdev
= conf
->mirrors
[d
].replacement
;
4459 b
= r10_bio
->devs
[s
/2].repl_bio
;
4461 rdev
= conf
->mirrors
[d
].rdev
;
4462 b
= r10_bio
->devs
[s
/2].bio
;
4464 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4466 atomic_inc(&rdev
->nr_pending
);
4467 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4468 atomic_inc(&r10_bio
->remaining
);
4470 generic_make_request(b
);
4472 end_reshape_request(r10_bio
);
4475 static void end_reshape(struct r10conf
*conf
)
4477 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4480 spin_lock_irq(&conf
->device_lock
);
4481 conf
->prev
= conf
->geo
;
4482 md_finish_reshape(conf
->mddev
);
4484 conf
->reshape_progress
= MaxSector
;
4485 spin_unlock_irq(&conf
->device_lock
);
4487 /* read-ahead size must cover two whole stripes, which is
4488 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4490 if (conf
->mddev
->queue
) {
4491 int stripe
= conf
->geo
.raid_disks
*
4492 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4493 stripe
/= conf
->geo
.near_copies
;
4494 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4495 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4501 static int handle_reshape_read_error(struct mddev
*mddev
,
4502 struct r10bio
*r10_bio
)
4504 /* Use sync reads to get the blocks from somewhere else */
4505 int sectors
= r10_bio
->sectors
;
4506 struct r10conf
*conf
= mddev
->private;
4508 struct r10bio r10_bio
;
4509 struct r10dev devs
[conf
->copies
];
4511 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4514 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4516 r10b
->sector
= r10_bio
->sector
;
4517 __raid10_find_phys(&conf
->prev
, r10b
);
4522 int first_slot
= slot
;
4524 if (s
> (PAGE_SIZE
>> 9))
4528 int d
= r10b
->devs
[slot
].devnum
;
4529 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4532 test_bit(Faulty
, &rdev
->flags
) ||
4533 !test_bit(In_sync
, &rdev
->flags
))
4536 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4537 success
= sync_page_io(rdev
,
4546 if (slot
>= conf
->copies
)
4548 if (slot
== first_slot
)
4552 /* couldn't read this block, must give up */
4553 set_bit(MD_RECOVERY_INTR
,
4563 static void end_reshape_write(struct bio
*bio
, int error
)
4565 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4566 struct r10bio
*r10_bio
= bio
->bi_private
;
4567 struct mddev
*mddev
= r10_bio
->mddev
;
4568 struct r10conf
*conf
= mddev
->private;
4572 struct md_rdev
*rdev
= NULL
;
4574 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4576 rdev
= conf
->mirrors
[d
].replacement
;
4579 rdev
= conf
->mirrors
[d
].rdev
;
4583 /* FIXME should record badblock */
4584 md_error(mddev
, rdev
);
4587 rdev_dec_pending(rdev
, mddev
);
4588 end_reshape_request(r10_bio
);
4591 static void end_reshape_request(struct r10bio
*r10_bio
)
4593 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4595 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4596 bio_put(r10_bio
->master_bio
);
4600 static void raid10_finish_reshape(struct mddev
*mddev
)
4602 struct r10conf
*conf
= mddev
->private;
4604 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4607 if (mddev
->delta_disks
> 0) {
4608 sector_t size
= raid10_size(mddev
, 0, 0);
4609 md_set_array_sectors(mddev
, size
);
4610 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4611 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4612 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4614 mddev
->resync_max_sectors
= size
;
4615 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4616 revalidate_disk(mddev
->gendisk
);
4619 for (d
= conf
->geo
.raid_disks
;
4620 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4622 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4624 clear_bit(In_sync
, &rdev
->flags
);
4625 rdev
= conf
->mirrors
[d
].replacement
;
4627 clear_bit(In_sync
, &rdev
->flags
);
4630 mddev
->layout
= mddev
->new_layout
;
4631 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4632 mddev
->reshape_position
= MaxSector
;
4633 mddev
->delta_disks
= 0;
4634 mddev
->reshape_backwards
= 0;
4637 static struct md_personality raid10_personality
=
4641 .owner
= THIS_MODULE
,
4642 .make_request
= make_request
,
4646 .error_handler
= error
,
4647 .hot_add_disk
= raid10_add_disk
,
4648 .hot_remove_disk
= raid10_remove_disk
,
4649 .spare_active
= raid10_spare_active
,
4650 .sync_request
= sync_request
,
4651 .quiesce
= raid10_quiesce
,
4652 .size
= raid10_size
,
4653 .resize
= raid10_resize
,
4654 .takeover
= raid10_takeover
,
4655 .check_reshape
= raid10_check_reshape
,
4656 .start_reshape
= raid10_start_reshape
,
4657 .finish_reshape
= raid10_finish_reshape
,
4660 static int __init
raid_init(void)
4662 return register_md_personality(&raid10_personality
);
4665 static void raid_exit(void)
4667 unregister_md_personality(&raid10_personality
);
4670 module_init(raid_init
);
4671 module_exit(raid_exit
);
4672 MODULE_LICENSE("GPL");
4673 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4674 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4675 MODULE_ALIAS("md-raid10");
4676 MODULE_ALIAS("md-level-10");
4678 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);