2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
42 * The data to be stored is divided into chunks using chunksize.
43 * Each device is divided into far_copies sections.
44 * In each section, chunks are laid out in a style similar to raid0, but
45 * near_copies copies of each chunk is stored (each on a different drive).
46 * The starting device for each section is offset near_copies from the starting
47 * device of the previous section.
48 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
50 * near_copies and far_copies must be at least one, and their product is at most
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of be very far apart
55 * on disk, there are adjacent stripes.
59 * Number of guaranteed r10bios in case of extreme VM load:
61 #define NR_RAID10_BIOS 256
63 /* when we get a read error on a read-only array, we redirect to another
64 * device without failing the first device, or trying to over-write to
65 * correct the read error. To keep track of bad blocks on a per-bio
66 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
68 #define IO_BLOCKED ((struct bio *)1)
69 /* When we successfully write to a known bad-block, we need to remove the
70 * bad-block marking which must be done from process context. So we record
71 * the success by setting devs[n].bio to IO_MADE_GOOD
73 #define IO_MADE_GOOD ((struct bio *)2)
75 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
77 /* When there are this many requests queued to be written by
78 * the raid10 thread, we become 'congested' to provide back-pressure
81 static int max_queued_requests
= 1024;
83 static void allow_barrier(struct r10conf
*conf
);
84 static void lower_barrier(struct r10conf
*conf
);
85 static int enough(struct r10conf
*conf
, int ignore
);
86 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
88 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
);
89 static void end_reshape_write(struct bio
*bio
, int error
);
90 static void end_reshape(struct r10conf
*conf
);
92 static void * r10bio_pool_alloc(gfp_t gfp_flags
, void *data
)
94 struct r10conf
*conf
= data
;
95 int size
= offsetof(struct r10bio
, devs
[conf
->copies
]);
97 /* allocate a r10bio with room for raid_disks entries in the
99 return kzalloc(size
, gfp_flags
);
102 static void r10bio_pool_free(void *r10_bio
, void *data
)
107 /* Maximum size of each resync request */
108 #define RESYNC_BLOCK_SIZE (64*1024)
109 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
110 /* amount of memory to reserve for resync requests */
111 #define RESYNC_WINDOW (1024*1024)
112 /* maximum number of concurrent requests, memory permitting */
113 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
116 * When performing a resync, we need to read and compare, so
117 * we need as many pages are there are copies.
118 * When performing a recovery, we need 2 bios, one for read,
119 * one for write (we recover only one drive per r10buf)
122 static void * r10buf_pool_alloc(gfp_t gfp_flags
, void *data
)
124 struct r10conf
*conf
= data
;
126 struct r10bio
*r10_bio
;
131 r10_bio
= r10bio_pool_alloc(gfp_flags
, conf
);
135 if (test_bit(MD_RECOVERY_SYNC
, &conf
->mddev
->recovery
) ||
136 test_bit(MD_RECOVERY_RESHAPE
, &conf
->mddev
->recovery
))
137 nalloc
= conf
->copies
; /* resync */
139 nalloc
= 2; /* recovery */
144 for (j
= nalloc
; j
-- ; ) {
145 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
148 r10_bio
->devs
[j
].bio
= bio
;
149 if (!conf
->have_replacement
)
151 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
154 r10_bio
->devs
[j
].repl_bio
= bio
;
157 * Allocate RESYNC_PAGES data pages and attach them
160 for (j
= 0 ; j
< nalloc
; j
++) {
161 struct bio
*rbio
= r10_bio
->devs
[j
].repl_bio
;
162 bio
= r10_bio
->devs
[j
].bio
;
163 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
164 if (j
> 0 && !test_bit(MD_RECOVERY_SYNC
,
165 &conf
->mddev
->recovery
)) {
166 /* we can share bv_page's during recovery
168 struct bio
*rbio
= r10_bio
->devs
[0].bio
;
169 page
= rbio
->bi_io_vec
[i
].bv_page
;
172 page
= alloc_page(gfp_flags
);
176 bio
->bi_io_vec
[i
].bv_page
= page
;
178 rbio
->bi_io_vec
[i
].bv_page
= page
;
186 safe_put_page(bio
->bi_io_vec
[i
-1].bv_page
);
188 for (i
= 0; i
< RESYNC_PAGES
; i
++)
189 safe_put_page(r10_bio
->devs
[j
].bio
->bi_io_vec
[i
].bv_page
);
192 for ( ; j
< nalloc
; j
++) {
193 if (r10_bio
->devs
[j
].bio
)
194 bio_put(r10_bio
->devs
[j
].bio
);
195 if (r10_bio
->devs
[j
].repl_bio
)
196 bio_put(r10_bio
->devs
[j
].repl_bio
);
198 r10bio_pool_free(r10_bio
, conf
);
202 static void r10buf_pool_free(void *__r10_bio
, void *data
)
205 struct r10conf
*conf
= data
;
206 struct r10bio
*r10bio
= __r10_bio
;
209 for (j
=0; j
< conf
->copies
; j
++) {
210 struct bio
*bio
= r10bio
->devs
[j
].bio
;
212 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
213 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
214 bio
->bi_io_vec
[i
].bv_page
= NULL
;
218 bio
= r10bio
->devs
[j
].repl_bio
;
222 r10bio_pool_free(r10bio
, conf
);
225 static void put_all_bios(struct r10conf
*conf
, struct r10bio
*r10_bio
)
229 for (i
= 0; i
< conf
->copies
; i
++) {
230 struct bio
**bio
= & r10_bio
->devs
[i
].bio
;
231 if (!BIO_SPECIAL(*bio
))
234 bio
= &r10_bio
->devs
[i
].repl_bio
;
235 if (r10_bio
->read_slot
< 0 && !BIO_SPECIAL(*bio
))
241 static void free_r10bio(struct r10bio
*r10_bio
)
243 struct r10conf
*conf
= r10_bio
->mddev
->private;
245 put_all_bios(conf
, r10_bio
);
246 mempool_free(r10_bio
, conf
->r10bio_pool
);
249 static void put_buf(struct r10bio
*r10_bio
)
251 struct r10conf
*conf
= r10_bio
->mddev
->private;
253 mempool_free(r10_bio
, conf
->r10buf_pool
);
258 static void reschedule_retry(struct r10bio
*r10_bio
)
261 struct mddev
*mddev
= r10_bio
->mddev
;
262 struct r10conf
*conf
= mddev
->private;
264 spin_lock_irqsave(&conf
->device_lock
, flags
);
265 list_add(&r10_bio
->retry_list
, &conf
->retry_list
);
267 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
269 /* wake up frozen array... */
270 wake_up(&conf
->wait_barrier
);
272 md_wakeup_thread(mddev
->thread
);
276 * raid_end_bio_io() is called when we have finished servicing a mirrored
277 * operation and are ready to return a success/failure code to the buffer
280 static void raid_end_bio_io(struct r10bio
*r10_bio
)
282 struct bio
*bio
= r10_bio
->master_bio
;
284 struct r10conf
*conf
= r10_bio
->mddev
->private;
286 if (bio
->bi_phys_segments
) {
288 spin_lock_irqsave(&conf
->device_lock
, flags
);
289 bio
->bi_phys_segments
--;
290 done
= (bio
->bi_phys_segments
== 0);
291 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
294 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
295 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
299 * Wake up any possible resync thread that waits for the device
304 free_r10bio(r10_bio
);
308 * Update disk head position estimator based on IRQ completion info.
310 static inline void update_head_pos(int slot
, struct r10bio
*r10_bio
)
312 struct r10conf
*conf
= r10_bio
->mddev
->private;
314 conf
->mirrors
[r10_bio
->devs
[slot
].devnum
].head_position
=
315 r10_bio
->devs
[slot
].addr
+ (r10_bio
->sectors
);
319 * Find the disk number which triggered given bio
321 static int find_bio_disk(struct r10conf
*conf
, struct r10bio
*r10_bio
,
322 struct bio
*bio
, int *slotp
, int *replp
)
327 for (slot
= 0; slot
< conf
->copies
; slot
++) {
328 if (r10_bio
->devs
[slot
].bio
== bio
)
330 if (r10_bio
->devs
[slot
].repl_bio
== bio
) {
336 BUG_ON(slot
== conf
->copies
);
337 update_head_pos(slot
, r10_bio
);
343 return r10_bio
->devs
[slot
].devnum
;
346 static void raid10_end_read_request(struct bio
*bio
, int error
)
348 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
349 struct r10bio
*r10_bio
= bio
->bi_private
;
351 struct md_rdev
*rdev
;
352 struct r10conf
*conf
= r10_bio
->mddev
->private;
355 slot
= r10_bio
->read_slot
;
356 dev
= r10_bio
->devs
[slot
].devnum
;
357 rdev
= r10_bio
->devs
[slot
].rdev
;
359 * this branch is our 'one mirror IO has finished' event handler:
361 update_head_pos(slot
, r10_bio
);
365 * Set R10BIO_Uptodate in our master bio, so that
366 * we will return a good error code to the higher
367 * levels even if IO on some other mirrored buffer fails.
369 * The 'master' represents the composite IO operation to
370 * user-side. So if something waits for IO, then it will
371 * wait for the 'master' bio.
373 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
375 /* If all other devices that store this block have
376 * failed, we want to return the error upwards rather
377 * than fail the last device. Here we redefine
378 * "uptodate" to mean "Don't want to retry"
381 spin_lock_irqsave(&conf
->device_lock
, flags
);
382 if (!enough(conf
, rdev
->raid_disk
))
384 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
387 raid_end_bio_io(r10_bio
);
388 rdev_dec_pending(rdev
, conf
->mddev
);
391 * oops, read error - keep the refcount on the rdev
393 char b
[BDEVNAME_SIZE
];
394 printk_ratelimited(KERN_ERR
395 "md/raid10:%s: %s: rescheduling sector %llu\n",
397 bdevname(rdev
->bdev
, b
),
398 (unsigned long long)r10_bio
->sector
);
399 set_bit(R10BIO_ReadError
, &r10_bio
->state
);
400 reschedule_retry(r10_bio
);
404 static void close_write(struct r10bio
*r10_bio
)
406 /* clear the bitmap if all writes complete successfully */
407 bitmap_endwrite(r10_bio
->mddev
->bitmap
, r10_bio
->sector
,
409 !test_bit(R10BIO_Degraded
, &r10_bio
->state
),
411 md_write_end(r10_bio
->mddev
);
414 static void one_write_done(struct r10bio
*r10_bio
)
416 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
417 if (test_bit(R10BIO_WriteError
, &r10_bio
->state
))
418 reschedule_retry(r10_bio
);
420 close_write(r10_bio
);
421 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
))
422 reschedule_retry(r10_bio
);
424 raid_end_bio_io(r10_bio
);
429 static void raid10_end_write_request(struct bio
*bio
, int error
)
431 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
432 struct r10bio
*r10_bio
= bio
->bi_private
;
435 struct r10conf
*conf
= r10_bio
->mddev
->private;
437 struct md_rdev
*rdev
= NULL
;
439 dev
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
442 rdev
= conf
->mirrors
[dev
].replacement
;
446 rdev
= conf
->mirrors
[dev
].rdev
;
449 * this branch is our 'one mirror IO has finished' event handler:
453 /* Never record new bad blocks to replacement,
456 md_error(rdev
->mddev
, rdev
);
458 set_bit(WriteErrorSeen
, &rdev
->flags
);
459 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
460 set_bit(MD_RECOVERY_NEEDED
,
461 &rdev
->mddev
->recovery
);
462 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
467 * Set R10BIO_Uptodate in our master bio, so that
468 * we will return a good error code for to the higher
469 * levels even if IO on some other mirrored buffer fails.
471 * The 'master' represents the composite IO operation to
472 * user-side. So if something waits for IO, then it will
473 * wait for the 'master' bio.
478 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
480 /* Maybe we can clear some bad blocks. */
481 if (is_badblock(rdev
,
482 r10_bio
->devs
[slot
].addr
,
484 &first_bad
, &bad_sectors
)) {
487 r10_bio
->devs
[slot
].repl_bio
= IO_MADE_GOOD
;
489 r10_bio
->devs
[slot
].bio
= IO_MADE_GOOD
;
491 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
497 * Let's see if all mirrored write operations have finished
500 one_write_done(r10_bio
);
502 rdev_dec_pending(conf
->mirrors
[dev
].rdev
, conf
->mddev
);
506 * RAID10 layout manager
507 * As well as the chunksize and raid_disks count, there are two
508 * parameters: near_copies and far_copies.
509 * near_copies * far_copies must be <= raid_disks.
510 * Normally one of these will be 1.
511 * If both are 1, we get raid0.
512 * If near_copies == raid_disks, we get raid1.
514 * Chunks are laid out in raid0 style with near_copies copies of the
515 * first chunk, followed by near_copies copies of the next chunk and
517 * If far_copies > 1, then after 1/far_copies of the array has been assigned
518 * as described above, we start again with a device offset of near_copies.
519 * So we effectively have another copy of the whole array further down all
520 * the drives, but with blocks on different drives.
521 * With this layout, and block is never stored twice on the one device.
523 * raid10_find_phys finds the sector offset of a given virtual sector
524 * on each device that it is on.
526 * raid10_find_virt does the reverse mapping, from a device and a
527 * sector offset to a virtual address
530 static void __raid10_find_phys(struct geom
*geo
, struct r10bio
*r10bio
)
539 /* now calculate first sector/dev */
540 chunk
= r10bio
->sector
>> geo
->chunk_shift
;
541 sector
= r10bio
->sector
& geo
->chunk_mask
;
543 chunk
*= geo
->near_copies
;
545 dev
= sector_div(stripe
, geo
->raid_disks
);
547 stripe
*= geo
->far_copies
;
549 sector
+= stripe
<< geo
->chunk_shift
;
551 /* and calculate all the others */
552 for (n
= 0; n
< geo
->near_copies
; n
++) {
555 r10bio
->devs
[slot
].addr
= sector
;
556 r10bio
->devs
[slot
].devnum
= d
;
559 for (f
= 1; f
< geo
->far_copies
; f
++) {
560 d
+= geo
->near_copies
;
561 if (d
>= geo
->raid_disks
)
562 d
-= geo
->raid_disks
;
564 r10bio
->devs
[slot
].devnum
= d
;
565 r10bio
->devs
[slot
].addr
= s
;
569 if (dev
>= geo
->raid_disks
) {
571 sector
+= (geo
->chunk_mask
+ 1);
576 static void raid10_find_phys(struct r10conf
*conf
, struct r10bio
*r10bio
)
578 struct geom
*geo
= &conf
->geo
;
580 if (conf
->reshape_progress
!= MaxSector
&&
581 ((r10bio
->sector
>= conf
->reshape_progress
) !=
582 conf
->mddev
->reshape_backwards
)) {
583 set_bit(R10BIO_Previous
, &r10bio
->state
);
586 clear_bit(R10BIO_Previous
, &r10bio
->state
);
588 __raid10_find_phys(geo
, r10bio
);
591 static sector_t
raid10_find_virt(struct r10conf
*conf
, sector_t sector
, int dev
)
593 sector_t offset
, chunk
, vchunk
;
594 /* Never use conf->prev as this is only called during resync
595 * or recovery, so reshape isn't happening
597 struct geom
*geo
= &conf
->geo
;
599 offset
= sector
& geo
->chunk_mask
;
600 if (geo
->far_offset
) {
602 chunk
= sector
>> geo
->chunk_shift
;
603 fc
= sector_div(chunk
, geo
->far_copies
);
604 dev
-= fc
* geo
->near_copies
;
606 dev
+= geo
->raid_disks
;
608 while (sector
>= geo
->stride
) {
609 sector
-= geo
->stride
;
610 if (dev
< geo
->near_copies
)
611 dev
+= geo
->raid_disks
- geo
->near_copies
;
613 dev
-= geo
->near_copies
;
615 chunk
= sector
>> geo
->chunk_shift
;
617 vchunk
= chunk
* geo
->raid_disks
+ dev
;
618 sector_div(vchunk
, geo
->near_copies
);
619 return (vchunk
<< geo
->chunk_shift
) + offset
;
623 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
625 * @bvm: properties of new bio
626 * @biovec: the request that could be merged to it.
628 * Return amount of bytes we can accept at this offset
629 * This requires checking for end-of-chunk if near_copies != raid_disks,
630 * and for subordinate merge_bvec_fns if merge_check_needed.
632 static int raid10_mergeable_bvec(struct request_queue
*q
,
633 struct bvec_merge_data
*bvm
,
634 struct bio_vec
*biovec
)
636 struct mddev
*mddev
= q
->queuedata
;
637 struct r10conf
*conf
= mddev
->private;
638 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
640 unsigned int chunk_sectors
;
641 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
642 struct geom
*geo
= &conf
->geo
;
644 chunk_sectors
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
) + 1;
645 if (conf
->reshape_progress
!= MaxSector
&&
646 ((sector
>= conf
->reshape_progress
) !=
647 conf
->mddev
->reshape_backwards
))
650 if (geo
->near_copies
< geo
->raid_disks
) {
651 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1))
652 + bio_sectors
)) << 9;
654 /* bio_add cannot handle a negative return */
656 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
657 return biovec
->bv_len
;
659 max
= biovec
->bv_len
;
661 if (mddev
->merge_check_needed
) {
663 struct r10bio r10_bio
;
664 struct r10dev devs
[conf
->copies
];
666 struct r10bio
*r10_bio
= &on_stack
.r10_bio
;
668 if (conf
->reshape_progress
!= MaxSector
) {
669 /* Cannot give any guidance during reshape */
670 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
671 return biovec
->bv_len
;
674 r10_bio
->sector
= sector
;
675 raid10_find_phys(conf
, r10_bio
);
677 for (s
= 0; s
< conf
->copies
; s
++) {
678 int disk
= r10_bio
->devs
[s
].devnum
;
679 struct md_rdev
*rdev
= rcu_dereference(
680 conf
->mirrors
[disk
].rdev
);
681 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
682 struct request_queue
*q
=
683 bdev_get_queue(rdev
->bdev
);
684 if (q
->merge_bvec_fn
) {
685 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
687 bvm
->bi_bdev
= rdev
->bdev
;
688 max
= min(max
, q
->merge_bvec_fn(
692 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
693 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
694 struct request_queue
*q
=
695 bdev_get_queue(rdev
->bdev
);
696 if (q
->merge_bvec_fn
) {
697 bvm
->bi_sector
= r10_bio
->devs
[s
].addr
699 bvm
->bi_bdev
= rdev
->bdev
;
700 max
= min(max
, q
->merge_bvec_fn(
711 * This routine returns the disk from which the requested read should
712 * be done. There is a per-array 'next expected sequential IO' sector
713 * number - if this matches on the next IO then we use the last disk.
714 * There is also a per-disk 'last know head position' sector that is
715 * maintained from IRQ contexts, both the normal and the resync IO
716 * completion handlers update this position correctly. If there is no
717 * perfect sequential match then we pick the disk whose head is closest.
719 * If there are 2 mirrors in the same 2 devices, performance degrades
720 * because position is mirror, not device based.
722 * The rdev for the device selected will have nr_pending incremented.
726 * FIXME: possibly should rethink readbalancing and do it differently
727 * depending on near_copies / far_copies geometry.
729 static struct md_rdev
*read_balance(struct r10conf
*conf
,
730 struct r10bio
*r10_bio
,
733 const sector_t this_sector
= r10_bio
->sector
;
735 int sectors
= r10_bio
->sectors
;
736 int best_good_sectors
;
737 sector_t new_distance
, best_dist
;
738 struct md_rdev
*best_rdev
, *rdev
= NULL
;
741 struct geom
*geo
= &conf
->geo
;
743 raid10_find_phys(conf
, r10_bio
);
746 sectors
= r10_bio
->sectors
;
749 best_dist
= MaxSector
;
750 best_good_sectors
= 0;
753 * Check if we can balance. We can balance on the whole
754 * device if no resync is going on (recovery is ok), or below
755 * the resync window. We take the first readable disk when
756 * above the resync window.
758 if (conf
->mddev
->recovery_cp
< MaxSector
759 && (this_sector
+ sectors
>= conf
->next_resync
))
762 for (slot
= 0; slot
< conf
->copies
; slot
++) {
767 if (r10_bio
->devs
[slot
].bio
== IO_BLOCKED
)
769 disk
= r10_bio
->devs
[slot
].devnum
;
770 rdev
= rcu_dereference(conf
->mirrors
[disk
].replacement
);
771 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
) ||
772 test_bit(Unmerged
, &rdev
->flags
) ||
773 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
774 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
776 test_bit(Faulty
, &rdev
->flags
) ||
777 test_bit(Unmerged
, &rdev
->flags
))
779 if (!test_bit(In_sync
, &rdev
->flags
) &&
780 r10_bio
->devs
[slot
].addr
+ sectors
> rdev
->recovery_offset
)
783 dev_sector
= r10_bio
->devs
[slot
].addr
;
784 if (is_badblock(rdev
, dev_sector
, sectors
,
785 &first_bad
, &bad_sectors
)) {
786 if (best_dist
< MaxSector
)
787 /* Already have a better slot */
789 if (first_bad
<= dev_sector
) {
790 /* Cannot read here. If this is the
791 * 'primary' device, then we must not read
792 * beyond 'bad_sectors' from another device.
794 bad_sectors
-= (dev_sector
- first_bad
);
795 if (!do_balance
&& sectors
> bad_sectors
)
796 sectors
= bad_sectors
;
797 if (best_good_sectors
> sectors
)
798 best_good_sectors
= sectors
;
800 sector_t good_sectors
=
801 first_bad
- dev_sector
;
802 if (good_sectors
> best_good_sectors
) {
803 best_good_sectors
= good_sectors
;
808 /* Must read from here */
813 best_good_sectors
= sectors
;
818 /* This optimisation is debatable, and completely destroys
819 * sequential read speed for 'far copies' arrays. So only
820 * keep it for 'near' arrays, and review those later.
822 if (geo
->near_copies
> 1 && !atomic_read(&rdev
->nr_pending
))
825 /* for far > 1 always use the lowest address */
826 if (geo
->far_copies
> 1)
827 new_distance
= r10_bio
->devs
[slot
].addr
;
829 new_distance
= abs(r10_bio
->devs
[slot
].addr
-
830 conf
->mirrors
[disk
].head_position
);
831 if (new_distance
< best_dist
) {
832 best_dist
= new_distance
;
837 if (slot
>= conf
->copies
) {
843 atomic_inc(&rdev
->nr_pending
);
844 if (test_bit(Faulty
, &rdev
->flags
)) {
845 /* Cannot risk returning a device that failed
846 * before we inc'ed nr_pending
848 rdev_dec_pending(rdev
, conf
->mddev
);
851 r10_bio
->read_slot
= slot
;
855 *max_sectors
= best_good_sectors
;
860 int md_raid10_congested(struct mddev
*mddev
, int bits
)
862 struct r10conf
*conf
= mddev
->private;
865 if ((bits
& (1 << BDI_async_congested
)) &&
866 conf
->pending_count
>= max_queued_requests
)
871 (i
< conf
->geo
.raid_disks
|| i
< conf
->prev
.raid_disks
)
874 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
875 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
876 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
878 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
884 EXPORT_SYMBOL_GPL(md_raid10_congested
);
886 static int raid10_congested(void *data
, int bits
)
888 struct mddev
*mddev
= data
;
890 return mddev_congested(mddev
, bits
) ||
891 md_raid10_congested(mddev
, bits
);
894 static void flush_pending_writes(struct r10conf
*conf
)
896 /* Any writes that have been queued but are awaiting
897 * bitmap updates get flushed here.
899 spin_lock_irq(&conf
->device_lock
);
901 if (conf
->pending_bio_list
.head
) {
903 bio
= bio_list_get(&conf
->pending_bio_list
);
904 conf
->pending_count
= 0;
905 spin_unlock_irq(&conf
->device_lock
);
906 /* flush any pending bitmap writes to disk
907 * before proceeding w/ I/O */
908 bitmap_unplug(conf
->mddev
->bitmap
);
909 wake_up(&conf
->wait_barrier
);
911 while (bio
) { /* submit pending writes */
912 struct bio
*next
= bio
->bi_next
;
914 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
915 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
919 generic_make_request(bio
);
923 spin_unlock_irq(&conf
->device_lock
);
927 * Sometimes we need to suspend IO while we do something else,
928 * either some resync/recovery, or reconfigure the array.
929 * To do this we raise a 'barrier'.
930 * The 'barrier' is a counter that can be raised multiple times
931 * to count how many activities are happening which preclude
933 * We can only raise the barrier if there is no pending IO.
934 * i.e. if nr_pending == 0.
935 * We choose only to raise the barrier if no-one is waiting for the
936 * barrier to go down. This means that as soon as an IO request
937 * is ready, no other operations which require a barrier will start
938 * until the IO request has had a chance.
940 * So: regular IO calls 'wait_barrier'. When that returns there
941 * is no backgroup IO happening, It must arrange to call
942 * allow_barrier when it has finished its IO.
943 * backgroup IO calls must call raise_barrier. Once that returns
944 * there is no normal IO happeing. It must arrange to call
945 * lower_barrier when the particular background IO completes.
948 static void raise_barrier(struct r10conf
*conf
, int force
)
950 BUG_ON(force
&& !conf
->barrier
);
951 spin_lock_irq(&conf
->resync_lock
);
953 /* Wait until no block IO is waiting (unless 'force') */
954 wait_event_lock_irq(conf
->wait_barrier
, force
|| !conf
->nr_waiting
,
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
) {
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 (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1338 || test_bit(Unmerged
, &rrdev
->flags
)))
1341 r10_bio
->devs
[i
].bio
= NULL
;
1342 r10_bio
->devs
[i
].repl_bio
= NULL
;
1343 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
1344 test_bit(Unmerged
, &rdev
->flags
)) {
1345 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1348 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1350 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1354 is_bad
= is_badblock(rdev
, dev_sector
,
1356 &first_bad
, &bad_sectors
);
1358 /* Mustn't write here until the bad block
1361 atomic_inc(&rdev
->nr_pending
);
1362 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1363 blocked_rdev
= rdev
;
1366 if (is_bad
&& first_bad
<= dev_sector
) {
1367 /* Cannot write here at all */
1368 bad_sectors
-= (dev_sector
- first_bad
);
1369 if (bad_sectors
< max_sectors
)
1370 /* Mustn't write more than bad_sectors
1371 * to other devices yet
1373 max_sectors
= bad_sectors
;
1374 /* We don't set R10BIO_Degraded as that
1375 * only applies if the disk is missing,
1376 * so it might be re-added, and we want to
1377 * know to recover this chunk.
1378 * In this case the device is here, and the
1379 * fact that this chunk is not in-sync is
1380 * recorded in the bad block log.
1385 int good_sectors
= first_bad
- dev_sector
;
1386 if (good_sectors
< max_sectors
)
1387 max_sectors
= good_sectors
;
1390 r10_bio
->devs
[i
].bio
= bio
;
1391 atomic_inc(&rdev
->nr_pending
);
1393 r10_bio
->devs
[i
].repl_bio
= bio
;
1394 atomic_inc(&rrdev
->nr_pending
);
1399 if (unlikely(blocked_rdev
)) {
1400 /* Have to wait for this device to get unblocked, then retry */
1404 for (j
= 0; j
< i
; j
++) {
1405 if (r10_bio
->devs
[j
].bio
) {
1406 d
= r10_bio
->devs
[j
].devnum
;
1407 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1409 if (r10_bio
->devs
[j
].repl_bio
) {
1410 struct md_rdev
*rdev
;
1411 d
= r10_bio
->devs
[j
].devnum
;
1412 rdev
= conf
->mirrors
[d
].replacement
;
1414 /* Race with remove_disk */
1416 rdev
= conf
->mirrors
[d
].rdev
;
1418 rdev_dec_pending(rdev
, mddev
);
1421 allow_barrier(conf
);
1422 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1427 if (max_sectors
< r10_bio
->sectors
) {
1428 /* We are splitting this into multiple parts, so
1429 * we need to prepare for allocating another r10_bio.
1431 r10_bio
->sectors
= max_sectors
;
1432 spin_lock_irq(&conf
->device_lock
);
1433 if (bio
->bi_phys_segments
== 0)
1434 bio
->bi_phys_segments
= 2;
1436 bio
->bi_phys_segments
++;
1437 spin_unlock_irq(&conf
->device_lock
);
1439 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1441 atomic_set(&r10_bio
->remaining
, 1);
1442 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1444 for (i
= 0; i
< conf
->copies
; i
++) {
1446 int d
= r10_bio
->devs
[i
].devnum
;
1447 if (!r10_bio
->devs
[i
].bio
)
1450 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1451 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1453 r10_bio
->devs
[i
].bio
= mbio
;
1455 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1456 choose_data_offset(r10_bio
,
1457 conf
->mirrors
[d
].rdev
));
1458 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1459 mbio
->bi_end_io
= raid10_end_write_request
;
1460 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
| do_discard
;
1461 mbio
->bi_private
= r10_bio
;
1463 atomic_inc(&r10_bio
->remaining
);
1465 cb
= blk_check_plugged(raid10_unplug
, mddev
, sizeof(*plug
));
1467 plug
= container_of(cb
, struct raid10_plug_cb
, cb
);
1470 spin_lock_irqsave(&conf
->device_lock
, flags
);
1472 bio_list_add(&plug
->pending
, mbio
);
1473 plug
->pending_cnt
++;
1475 bio_list_add(&conf
->pending_bio_list
, mbio
);
1476 conf
->pending_count
++;
1478 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1480 md_wakeup_thread(mddev
->thread
);
1482 if (!r10_bio
->devs
[i
].repl_bio
)
1485 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1486 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1488 r10_bio
->devs
[i
].repl_bio
= mbio
;
1490 /* We are actively writing to the original device
1491 * so it cannot disappear, so the replacement cannot
1494 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1497 conf
->mirrors
[d
].replacement
));
1498 mbio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
1499 mbio
->bi_end_io
= raid10_end_write_request
;
1500 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
| do_discard
;
1501 mbio
->bi_private
= r10_bio
;
1503 atomic_inc(&r10_bio
->remaining
);
1504 spin_lock_irqsave(&conf
->device_lock
, flags
);
1505 bio_list_add(&conf
->pending_bio_list
, mbio
);
1506 conf
->pending_count
++;
1507 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1508 if (!mddev_check_plugged(mddev
))
1509 md_wakeup_thread(mddev
->thread
);
1512 /* Don't remove the bias on 'remaining' (one_write_done) until
1513 * after checking if we need to go around again.
1516 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1517 one_write_done(r10_bio
);
1518 /* We need another r10_bio. It has already been counted
1519 * in bio->bi_phys_segments.
1521 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1523 r10_bio
->master_bio
= bio
;
1524 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1526 r10_bio
->mddev
= mddev
;
1527 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1531 one_write_done(r10_bio
);
1533 /* In case raid10d snuck in to freeze_array */
1534 wake_up(&conf
->wait_barrier
);
1537 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1539 struct r10conf
*conf
= mddev
->private;
1542 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1543 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1544 if (conf
->geo
.near_copies
> 1)
1545 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1546 if (conf
->geo
.far_copies
> 1) {
1547 if (conf
->geo
.far_offset
)
1548 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1550 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1552 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1553 conf
->geo
.raid_disks
- mddev
->degraded
);
1554 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1555 seq_printf(seq
, "%s",
1556 conf
->mirrors
[i
].rdev
&&
1557 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1558 seq_printf(seq
, "]");
1561 /* check if there are enough drives for
1562 * every block to appear on atleast one.
1563 * Don't consider the device numbered 'ignore'
1564 * as we might be about to remove it.
1566 static int _enough(struct r10conf
*conf
, struct geom
*geo
, int ignore
)
1571 int n
= conf
->copies
;
1574 if (conf
->mirrors
[first
].rdev
&&
1577 first
= (first
+1) % geo
->raid_disks
;
1581 } while (first
!= 0);
1585 static int enough(struct r10conf
*conf
, int ignore
)
1587 return _enough(conf
, &conf
->geo
, ignore
) &&
1588 _enough(conf
, &conf
->prev
, ignore
);
1591 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1593 char b
[BDEVNAME_SIZE
];
1594 struct r10conf
*conf
= mddev
->private;
1597 * If it is not operational, then we have already marked it as dead
1598 * else if it is the last working disks, ignore the error, let the
1599 * next level up know.
1600 * else mark the drive as failed
1602 if (test_bit(In_sync
, &rdev
->flags
)
1603 && !enough(conf
, rdev
->raid_disk
))
1605 * Don't fail the drive, just return an IO error.
1608 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1609 unsigned long flags
;
1610 spin_lock_irqsave(&conf
->device_lock
, flags
);
1612 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1614 * if recovery is running, make sure it aborts.
1616 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1618 set_bit(Blocked
, &rdev
->flags
);
1619 set_bit(Faulty
, &rdev
->flags
);
1620 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1622 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1623 "md/raid10:%s: Operation continuing on %d devices.\n",
1624 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1625 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1628 static void print_conf(struct r10conf
*conf
)
1631 struct raid10_info
*tmp
;
1633 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1635 printk(KERN_DEBUG
"(!conf)\n");
1638 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1639 conf
->geo
.raid_disks
);
1641 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1642 char b
[BDEVNAME_SIZE
];
1643 tmp
= conf
->mirrors
+ i
;
1645 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1646 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1647 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1648 bdevname(tmp
->rdev
->bdev
,b
));
1652 static void close_sync(struct r10conf
*conf
)
1655 allow_barrier(conf
);
1657 mempool_destroy(conf
->r10buf_pool
);
1658 conf
->r10buf_pool
= NULL
;
1661 static int raid10_spare_active(struct mddev
*mddev
)
1664 struct r10conf
*conf
= mddev
->private;
1665 struct raid10_info
*tmp
;
1667 unsigned long flags
;
1670 * Find all non-in_sync disks within the RAID10 configuration
1671 * and mark them in_sync
1673 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1674 tmp
= conf
->mirrors
+ i
;
1675 if (tmp
->replacement
1676 && tmp
->replacement
->recovery_offset
== MaxSector
1677 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1678 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1679 /* Replacement has just become active */
1681 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1684 /* Replaced device not technically faulty,
1685 * but we need to be sure it gets removed
1686 * and never re-added.
1688 set_bit(Faulty
, &tmp
->rdev
->flags
);
1689 sysfs_notify_dirent_safe(
1690 tmp
->rdev
->sysfs_state
);
1692 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1693 } else if (tmp
->rdev
1694 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1695 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1697 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
1700 spin_lock_irqsave(&conf
->device_lock
, flags
);
1701 mddev
->degraded
-= count
;
1702 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1709 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1711 struct r10conf
*conf
= mddev
->private;
1715 int last
= conf
->geo
.raid_disks
- 1;
1716 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1718 if (mddev
->recovery_cp
< MaxSector
)
1719 /* only hot-add to in-sync arrays, as recovery is
1720 * very different from resync
1723 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, &conf
->prev
, -1))
1726 if (rdev
->raid_disk
>= 0)
1727 first
= last
= rdev
->raid_disk
;
1729 if (q
->merge_bvec_fn
) {
1730 set_bit(Unmerged
, &rdev
->flags
);
1731 mddev
->merge_check_needed
= 1;
1734 if (rdev
->saved_raid_disk
>= first
&&
1735 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1736 mirror
= rdev
->saved_raid_disk
;
1739 for ( ; mirror
<= last
; mirror
++) {
1740 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1741 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1744 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1745 p
->replacement
!= NULL
)
1747 clear_bit(In_sync
, &rdev
->flags
);
1748 set_bit(Replacement
, &rdev
->flags
);
1749 rdev
->raid_disk
= mirror
;
1751 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1752 rdev
->data_offset
<< 9);
1754 rcu_assign_pointer(p
->replacement
, rdev
);
1758 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1759 rdev
->data_offset
<< 9);
1761 p
->head_position
= 0;
1762 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1763 rdev
->raid_disk
= mirror
;
1765 if (rdev
->saved_raid_disk
!= mirror
)
1767 rcu_assign_pointer(p
->rdev
, rdev
);
1770 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1771 /* Some requests might not have seen this new
1772 * merge_bvec_fn. We must wait for them to complete
1773 * before merging the device fully.
1774 * First we make sure any code which has tested
1775 * our function has submitted the request, then
1776 * we wait for all outstanding requests to complete.
1778 synchronize_sched();
1779 raise_barrier(conf
, 0);
1780 lower_barrier(conf
);
1781 clear_bit(Unmerged
, &rdev
->flags
);
1783 md_integrity_add_rdev(rdev
, mddev
);
1784 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1785 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1791 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1793 struct r10conf
*conf
= mddev
->private;
1795 int number
= rdev
->raid_disk
;
1796 struct md_rdev
**rdevp
;
1797 struct raid10_info
*p
= conf
->mirrors
+ number
;
1800 if (rdev
== p
->rdev
)
1802 else if (rdev
== p
->replacement
)
1803 rdevp
= &p
->replacement
;
1807 if (test_bit(In_sync
, &rdev
->flags
) ||
1808 atomic_read(&rdev
->nr_pending
)) {
1812 /* Only remove faulty devices if recovery
1815 if (!test_bit(Faulty
, &rdev
->flags
) &&
1816 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1817 (!p
->replacement
|| p
->replacement
== rdev
) &&
1818 number
< conf
->geo
.raid_disks
&&
1825 if (atomic_read(&rdev
->nr_pending
)) {
1826 /* lost the race, try later */
1830 } else if (p
->replacement
) {
1831 /* We must have just cleared 'rdev' */
1832 p
->rdev
= p
->replacement
;
1833 clear_bit(Replacement
, &p
->replacement
->flags
);
1834 smp_mb(); /* Make sure other CPUs may see both as identical
1835 * but will never see neither -- if they are careful.
1837 p
->replacement
= NULL
;
1838 clear_bit(WantReplacement
, &rdev
->flags
);
1840 /* We might have just remove the Replacement as faulty
1841 * Clear the flag just in case
1843 clear_bit(WantReplacement
, &rdev
->flags
);
1845 err
= md_integrity_register(mddev
);
1854 static void end_sync_read(struct bio
*bio
, int error
)
1856 struct r10bio
*r10_bio
= bio
->bi_private
;
1857 struct r10conf
*conf
= r10_bio
->mddev
->private;
1860 if (bio
== r10_bio
->master_bio
) {
1861 /* this is a reshape read */
1862 d
= r10_bio
->read_slot
; /* really the read dev */
1864 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1866 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1867 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1869 /* The write handler will notice the lack of
1870 * R10BIO_Uptodate and record any errors etc
1872 atomic_add(r10_bio
->sectors
,
1873 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1875 /* for reconstruct, we always reschedule after a read.
1876 * for resync, only after all reads
1878 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1879 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1880 atomic_dec_and_test(&r10_bio
->remaining
)) {
1881 /* we have read all the blocks,
1882 * do the comparison in process context in raid10d
1884 reschedule_retry(r10_bio
);
1888 static void end_sync_request(struct r10bio
*r10_bio
)
1890 struct mddev
*mddev
= r10_bio
->mddev
;
1892 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1893 if (r10_bio
->master_bio
== NULL
) {
1894 /* the primary of several recovery bios */
1895 sector_t s
= r10_bio
->sectors
;
1896 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1897 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1898 reschedule_retry(r10_bio
);
1901 md_done_sync(mddev
, s
, 1);
1904 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1905 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1906 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1907 reschedule_retry(r10_bio
);
1915 static void end_sync_write(struct bio
*bio
, int error
)
1917 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1918 struct r10bio
*r10_bio
= bio
->bi_private
;
1919 struct mddev
*mddev
= r10_bio
->mddev
;
1920 struct r10conf
*conf
= mddev
->private;
1926 struct md_rdev
*rdev
= NULL
;
1928 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1930 rdev
= conf
->mirrors
[d
].replacement
;
1932 rdev
= conf
->mirrors
[d
].rdev
;
1936 md_error(mddev
, rdev
);
1938 set_bit(WriteErrorSeen
, &rdev
->flags
);
1939 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1940 set_bit(MD_RECOVERY_NEEDED
,
1941 &rdev
->mddev
->recovery
);
1942 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1944 } else if (is_badblock(rdev
,
1945 r10_bio
->devs
[slot
].addr
,
1947 &first_bad
, &bad_sectors
))
1948 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1950 rdev_dec_pending(rdev
, mddev
);
1952 end_sync_request(r10_bio
);
1956 * Note: sync and recover and handled very differently for raid10
1957 * This code is for resync.
1958 * For resync, we read through virtual addresses and read all blocks.
1959 * If there is any error, we schedule a write. The lowest numbered
1960 * drive is authoritative.
1961 * However requests come for physical address, so we need to map.
1962 * For every physical address there are raid_disks/copies virtual addresses,
1963 * which is always are least one, but is not necessarly an integer.
1964 * This means that a physical address can span multiple chunks, so we may
1965 * have to submit multiple io requests for a single sync request.
1968 * We check if all blocks are in-sync and only write to blocks that
1971 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1973 struct r10conf
*conf
= mddev
->private;
1975 struct bio
*tbio
, *fbio
;
1978 atomic_set(&r10_bio
->remaining
, 1);
1980 /* find the first device with a block */
1981 for (i
=0; i
<conf
->copies
; i
++)
1982 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1985 if (i
== conf
->copies
)
1989 fbio
= r10_bio
->devs
[i
].bio
;
1991 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
1992 /* now find blocks with errors */
1993 for (i
=0 ; i
< conf
->copies
; i
++) {
1996 tbio
= r10_bio
->devs
[i
].bio
;
1998 if (tbio
->bi_end_io
!= end_sync_read
)
2002 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
2003 /* We know that the bi_io_vec layout is the same for
2004 * both 'first' and 'i', so we just compare them.
2005 * All vec entries are PAGE_SIZE;
2007 for (j
= 0; j
< vcnt
; j
++)
2008 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
2009 page_address(tbio
->bi_io_vec
[j
].bv_page
),
2010 fbio
->bi_io_vec
[j
].bv_len
))
2014 atomic64_add(r10_bio
->sectors
, &mddev
->resync_mismatches
);
2015 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
2016 /* Don't fix anything. */
2019 /* Ok, we need to write this bio, either to correct an
2020 * inconsistency or to correct an unreadable block.
2021 * First we need to fixup bv_offset, bv_len and
2022 * bi_vecs, as the read request might have corrupted these
2024 tbio
->bi_vcnt
= vcnt
;
2025 tbio
->bi_size
= r10_bio
->sectors
<< 9;
2027 tbio
->bi_phys_segments
= 0;
2028 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
2029 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
2030 tbio
->bi_next
= NULL
;
2031 tbio
->bi_rw
= WRITE
;
2032 tbio
->bi_private
= r10_bio
;
2033 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
2035 for (j
=0; j
< vcnt
; j
++) {
2036 tbio
->bi_io_vec
[j
].bv_offset
= 0;
2037 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
2039 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2040 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2043 tbio
->bi_end_io
= end_sync_write
;
2045 d
= r10_bio
->devs
[i
].devnum
;
2046 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2047 atomic_inc(&r10_bio
->remaining
);
2048 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
2050 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2051 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2052 generic_make_request(tbio
);
2055 /* Now write out to any replacement devices
2058 for (i
= 0; i
< conf
->copies
; i
++) {
2061 tbio
= r10_bio
->devs
[i
].repl_bio
;
2062 if (!tbio
|| !tbio
->bi_end_io
)
2064 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2065 && r10_bio
->devs
[i
].bio
!= fbio
)
2066 for (j
= 0; j
< vcnt
; j
++)
2067 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2068 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2070 d
= r10_bio
->devs
[i
].devnum
;
2071 atomic_inc(&r10_bio
->remaining
);
2072 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2073 tbio
->bi_size
>> 9);
2074 generic_make_request(tbio
);
2078 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2079 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2085 * Now for the recovery code.
2086 * Recovery happens across physical sectors.
2087 * We recover all non-is_sync drives by finding the virtual address of
2088 * each, and then choose a working drive that also has that virt address.
2089 * There is a separate r10_bio for each non-in_sync drive.
2090 * Only the first two slots are in use. The first for reading,
2091 * The second for writing.
2094 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2096 /* We got a read error during recovery.
2097 * We repeat the read in smaller page-sized sections.
2098 * If a read succeeds, write it to the new device or record
2099 * a bad block if we cannot.
2100 * If a read fails, record a bad block on both old and
2103 struct mddev
*mddev
= r10_bio
->mddev
;
2104 struct r10conf
*conf
= mddev
->private;
2105 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2107 int sectors
= r10_bio
->sectors
;
2109 int dr
= r10_bio
->devs
[0].devnum
;
2110 int dw
= r10_bio
->devs
[1].devnum
;
2114 struct md_rdev
*rdev
;
2118 if (s
> (PAGE_SIZE
>>9))
2121 rdev
= conf
->mirrors
[dr
].rdev
;
2122 addr
= r10_bio
->devs
[0].addr
+ sect
,
2123 ok
= sync_page_io(rdev
,
2126 bio
->bi_io_vec
[idx
].bv_page
,
2129 rdev
= conf
->mirrors
[dw
].rdev
;
2130 addr
= r10_bio
->devs
[1].addr
+ sect
;
2131 ok
= sync_page_io(rdev
,
2134 bio
->bi_io_vec
[idx
].bv_page
,
2137 set_bit(WriteErrorSeen
, &rdev
->flags
);
2138 if (!test_and_set_bit(WantReplacement
,
2140 set_bit(MD_RECOVERY_NEEDED
,
2141 &rdev
->mddev
->recovery
);
2145 /* We don't worry if we cannot set a bad block -
2146 * it really is bad so there is no loss in not
2149 rdev_set_badblocks(rdev
, addr
, s
, 0);
2151 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2152 /* need bad block on destination too */
2153 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2154 addr
= r10_bio
->devs
[1].addr
+ sect
;
2155 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2157 /* just abort the recovery */
2159 "md/raid10:%s: recovery aborted"
2160 " due to read error\n",
2163 conf
->mirrors
[dw
].recovery_disabled
2164 = mddev
->recovery_disabled
;
2165 set_bit(MD_RECOVERY_INTR
,
2178 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2180 struct r10conf
*conf
= mddev
->private;
2182 struct bio
*wbio
, *wbio2
;
2184 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2185 fix_recovery_read_error(r10_bio
);
2186 end_sync_request(r10_bio
);
2191 * share the pages with the first bio
2192 * and submit the write request
2194 d
= r10_bio
->devs
[1].devnum
;
2195 wbio
= r10_bio
->devs
[1].bio
;
2196 wbio2
= r10_bio
->devs
[1].repl_bio
;
2197 if (wbio
->bi_end_io
) {
2198 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2199 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
2200 generic_make_request(wbio
);
2202 if (wbio2
&& wbio2
->bi_end_io
) {
2203 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2204 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2205 wbio2
->bi_size
>> 9);
2206 generic_make_request(wbio2
);
2212 * Used by fix_read_error() to decay the per rdev read_errors.
2213 * We halve the read error count for every hour that has elapsed
2214 * since the last recorded read error.
2217 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2219 struct timespec cur_time_mon
;
2220 unsigned long hours_since_last
;
2221 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2223 ktime_get_ts(&cur_time_mon
);
2225 if (rdev
->last_read_error
.tv_sec
== 0 &&
2226 rdev
->last_read_error
.tv_nsec
== 0) {
2227 /* first time we've seen a read error */
2228 rdev
->last_read_error
= cur_time_mon
;
2232 hours_since_last
= (cur_time_mon
.tv_sec
-
2233 rdev
->last_read_error
.tv_sec
) / 3600;
2235 rdev
->last_read_error
= cur_time_mon
;
2238 * if hours_since_last is > the number of bits in read_errors
2239 * just set read errors to 0. We do this to avoid
2240 * overflowing the shift of read_errors by hours_since_last.
2242 if (hours_since_last
>= 8 * sizeof(read_errors
))
2243 atomic_set(&rdev
->read_errors
, 0);
2245 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2248 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2249 int sectors
, struct page
*page
, int rw
)
2254 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2255 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2257 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2261 set_bit(WriteErrorSeen
, &rdev
->flags
);
2262 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2263 set_bit(MD_RECOVERY_NEEDED
,
2264 &rdev
->mddev
->recovery
);
2266 /* need to record an error - either for the block or the device */
2267 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2268 md_error(rdev
->mddev
, rdev
);
2273 * This is a kernel thread which:
2275 * 1. Retries failed read operations on working mirrors.
2276 * 2. Updates the raid superblock when problems encounter.
2277 * 3. Performs writes following reads for array synchronising.
2280 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2282 int sect
= 0; /* Offset from r10_bio->sector */
2283 int sectors
= r10_bio
->sectors
;
2284 struct md_rdev
*rdev
;
2285 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2286 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2288 /* still own a reference to this rdev, so it cannot
2289 * have been cleared recently.
2291 rdev
= conf
->mirrors
[d
].rdev
;
2293 if (test_bit(Faulty
, &rdev
->flags
))
2294 /* drive has already been failed, just ignore any
2295 more fix_read_error() attempts */
2298 check_decay_read_errors(mddev
, rdev
);
2299 atomic_inc(&rdev
->read_errors
);
2300 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2301 char b
[BDEVNAME_SIZE
];
2302 bdevname(rdev
->bdev
, b
);
2305 "md/raid10:%s: %s: Raid device exceeded "
2306 "read_error threshold [cur %d:max %d]\n",
2308 atomic_read(&rdev
->read_errors
), max_read_errors
);
2310 "md/raid10:%s: %s: Failing raid device\n",
2312 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2313 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2319 int sl
= r10_bio
->read_slot
;
2323 if (s
> (PAGE_SIZE
>>9))
2331 d
= r10_bio
->devs
[sl
].devnum
;
2332 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2334 !test_bit(Unmerged
, &rdev
->flags
) &&
2335 test_bit(In_sync
, &rdev
->flags
) &&
2336 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2337 &first_bad
, &bad_sectors
) == 0) {
2338 atomic_inc(&rdev
->nr_pending
);
2340 success
= sync_page_io(rdev
,
2341 r10_bio
->devs
[sl
].addr
+
2344 conf
->tmppage
, READ
, false);
2345 rdev_dec_pending(rdev
, mddev
);
2351 if (sl
== conf
->copies
)
2353 } while (!success
&& sl
!= r10_bio
->read_slot
);
2357 /* Cannot read from anywhere, just mark the block
2358 * as bad on the first device to discourage future
2361 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2362 rdev
= conf
->mirrors
[dn
].rdev
;
2364 if (!rdev_set_badblocks(
2366 r10_bio
->devs
[r10_bio
->read_slot
].addr
2369 md_error(mddev
, rdev
);
2370 r10_bio
->devs
[r10_bio
->read_slot
].bio
2377 /* write it back and re-read */
2379 while (sl
!= r10_bio
->read_slot
) {
2380 char b
[BDEVNAME_SIZE
];
2385 d
= r10_bio
->devs
[sl
].devnum
;
2386 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2388 test_bit(Unmerged
, &rdev
->flags
) ||
2389 !test_bit(In_sync
, &rdev
->flags
))
2392 atomic_inc(&rdev
->nr_pending
);
2394 if (r10_sync_page_io(rdev
,
2395 r10_bio
->devs
[sl
].addr
+
2397 s
, conf
->tmppage
, WRITE
)
2399 /* Well, this device is dead */
2401 "md/raid10:%s: read correction "
2403 " (%d sectors at %llu on %s)\n",
2405 (unsigned long long)(
2407 choose_data_offset(r10_bio
,
2409 bdevname(rdev
->bdev
, b
));
2410 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2413 bdevname(rdev
->bdev
, b
));
2415 rdev_dec_pending(rdev
, mddev
);
2419 while (sl
!= r10_bio
->read_slot
) {
2420 char b
[BDEVNAME_SIZE
];
2425 d
= r10_bio
->devs
[sl
].devnum
;
2426 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2428 !test_bit(In_sync
, &rdev
->flags
))
2431 atomic_inc(&rdev
->nr_pending
);
2433 switch (r10_sync_page_io(rdev
,
2434 r10_bio
->devs
[sl
].addr
+
2439 /* Well, this device is dead */
2441 "md/raid10:%s: unable to read back "
2443 " (%d sectors at %llu on %s)\n",
2445 (unsigned long long)(
2447 choose_data_offset(r10_bio
, rdev
)),
2448 bdevname(rdev
->bdev
, b
));
2449 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2452 bdevname(rdev
->bdev
, b
));
2456 "md/raid10:%s: read error corrected"
2457 " (%d sectors at %llu on %s)\n",
2459 (unsigned long long)(
2461 choose_data_offset(r10_bio
, rdev
)),
2462 bdevname(rdev
->bdev
, b
));
2463 atomic_add(s
, &rdev
->corrected_errors
);
2466 rdev_dec_pending(rdev
, mddev
);
2476 static void bi_complete(struct bio
*bio
, int error
)
2478 complete((struct completion
*)bio
->bi_private
);
2481 static int submit_bio_wait(int rw
, struct bio
*bio
)
2483 struct completion event
;
2486 init_completion(&event
);
2487 bio
->bi_private
= &event
;
2488 bio
->bi_end_io
= bi_complete
;
2489 submit_bio(rw
, bio
);
2490 wait_for_completion(&event
);
2492 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2495 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2497 struct bio
*bio
= r10_bio
->master_bio
;
2498 struct mddev
*mddev
= r10_bio
->mddev
;
2499 struct r10conf
*conf
= mddev
->private;
2500 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2501 /* bio has the data to be written to slot 'i' where
2502 * we just recently had a write error.
2503 * We repeatedly clone the bio and trim down to one block,
2504 * then try the write. Where the write fails we record
2506 * It is conceivable that the bio doesn't exactly align with
2507 * blocks. We must handle this.
2509 * We currently own a reference to the rdev.
2515 int sect_to_write
= r10_bio
->sectors
;
2518 if (rdev
->badblocks
.shift
< 0)
2521 block_sectors
= 1 << rdev
->badblocks
.shift
;
2522 sector
= r10_bio
->sector
;
2523 sectors
= ((r10_bio
->sector
+ block_sectors
)
2524 & ~(sector_t
)(block_sectors
- 1))
2527 while (sect_to_write
) {
2529 if (sectors
> sect_to_write
)
2530 sectors
= sect_to_write
;
2531 /* Write at 'sector' for 'sectors' */
2532 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2533 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2534 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2535 choose_data_offset(r10_bio
, rdev
) +
2536 (sector
- r10_bio
->sector
));
2537 wbio
->bi_bdev
= rdev
->bdev
;
2538 if (submit_bio_wait(WRITE
, wbio
) == 0)
2540 ok
= rdev_set_badblocks(rdev
, sector
,
2545 sect_to_write
-= sectors
;
2547 sectors
= block_sectors
;
2552 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2554 int slot
= r10_bio
->read_slot
;
2556 struct r10conf
*conf
= mddev
->private;
2557 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2558 char b
[BDEVNAME_SIZE
];
2559 unsigned long do_sync
;
2562 /* we got a read error. Maybe the drive is bad. Maybe just
2563 * the block and we can fix it.
2564 * We freeze all other IO, and try reading the block from
2565 * other devices. When we find one, we re-write
2566 * and check it that fixes the read error.
2567 * This is all done synchronously while the array is
2570 bio
= r10_bio
->devs
[slot
].bio
;
2571 bdevname(bio
->bi_bdev
, b
);
2573 r10_bio
->devs
[slot
].bio
= NULL
;
2575 if (mddev
->ro
== 0) {
2577 fix_read_error(conf
, mddev
, r10_bio
);
2578 unfreeze_array(conf
);
2580 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2582 rdev_dec_pending(rdev
, mddev
);
2585 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2587 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2588 " read error for block %llu\n",
2590 (unsigned long long)r10_bio
->sector
);
2591 raid_end_bio_io(r10_bio
);
2595 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2596 slot
= r10_bio
->read_slot
;
2599 "md/raid10:%s: %s: redirecting "
2600 "sector %llu to another mirror\n",
2602 bdevname(rdev
->bdev
, b
),
2603 (unsigned long long)r10_bio
->sector
);
2604 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2607 r10_bio
->sector
- bio
->bi_sector
,
2609 r10_bio
->devs
[slot
].bio
= bio
;
2610 r10_bio
->devs
[slot
].rdev
= rdev
;
2611 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2612 + choose_data_offset(r10_bio
, rdev
);
2613 bio
->bi_bdev
= rdev
->bdev
;
2614 bio
->bi_rw
= READ
| do_sync
;
2615 bio
->bi_private
= r10_bio
;
2616 bio
->bi_end_io
= raid10_end_read_request
;
2617 if (max_sectors
< r10_bio
->sectors
) {
2618 /* Drat - have to split this up more */
2619 struct bio
*mbio
= r10_bio
->master_bio
;
2620 int sectors_handled
=
2621 r10_bio
->sector
+ max_sectors
2623 r10_bio
->sectors
= max_sectors
;
2624 spin_lock_irq(&conf
->device_lock
);
2625 if (mbio
->bi_phys_segments
== 0)
2626 mbio
->bi_phys_segments
= 2;
2628 mbio
->bi_phys_segments
++;
2629 spin_unlock_irq(&conf
->device_lock
);
2630 generic_make_request(bio
);
2632 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2634 r10_bio
->master_bio
= mbio
;
2635 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2638 set_bit(R10BIO_ReadError
,
2640 r10_bio
->mddev
= mddev
;
2641 r10_bio
->sector
= mbio
->bi_sector
2646 generic_make_request(bio
);
2649 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2651 /* Some sort of write request has finished and it
2652 * succeeded in writing where we thought there was a
2653 * bad block. So forget the bad block.
2654 * Or possibly if failed and we need to record
2658 struct md_rdev
*rdev
;
2660 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2661 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2662 for (m
= 0; m
< conf
->copies
; m
++) {
2663 int dev
= r10_bio
->devs
[m
].devnum
;
2664 rdev
= conf
->mirrors
[dev
].rdev
;
2665 if (r10_bio
->devs
[m
].bio
== NULL
)
2667 if (test_bit(BIO_UPTODATE
,
2668 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2669 rdev_clear_badblocks(
2671 r10_bio
->devs
[m
].addr
,
2672 r10_bio
->sectors
, 0);
2674 if (!rdev_set_badblocks(
2676 r10_bio
->devs
[m
].addr
,
2677 r10_bio
->sectors
, 0))
2678 md_error(conf
->mddev
, rdev
);
2680 rdev
= conf
->mirrors
[dev
].replacement
;
2681 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2683 if (test_bit(BIO_UPTODATE
,
2684 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2685 rdev_clear_badblocks(
2687 r10_bio
->devs
[m
].addr
,
2688 r10_bio
->sectors
, 0);
2690 if (!rdev_set_badblocks(
2692 r10_bio
->devs
[m
].addr
,
2693 r10_bio
->sectors
, 0))
2694 md_error(conf
->mddev
, rdev
);
2699 for (m
= 0; m
< conf
->copies
; m
++) {
2700 int dev
= r10_bio
->devs
[m
].devnum
;
2701 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2702 rdev
= conf
->mirrors
[dev
].rdev
;
2703 if (bio
== IO_MADE_GOOD
) {
2704 rdev_clear_badblocks(
2706 r10_bio
->devs
[m
].addr
,
2707 r10_bio
->sectors
, 0);
2708 rdev_dec_pending(rdev
, conf
->mddev
);
2709 } else if (bio
!= NULL
&&
2710 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2711 if (!narrow_write_error(r10_bio
, m
)) {
2712 md_error(conf
->mddev
, rdev
);
2713 set_bit(R10BIO_Degraded
,
2716 rdev_dec_pending(rdev
, conf
->mddev
);
2718 bio
= r10_bio
->devs
[m
].repl_bio
;
2719 rdev
= conf
->mirrors
[dev
].replacement
;
2720 if (rdev
&& bio
== IO_MADE_GOOD
) {
2721 rdev_clear_badblocks(
2723 r10_bio
->devs
[m
].addr
,
2724 r10_bio
->sectors
, 0);
2725 rdev_dec_pending(rdev
, conf
->mddev
);
2728 if (test_bit(R10BIO_WriteError
,
2730 close_write(r10_bio
);
2731 raid_end_bio_io(r10_bio
);
2735 static void raid10d(struct md_thread
*thread
)
2737 struct mddev
*mddev
= thread
->mddev
;
2738 struct r10bio
*r10_bio
;
2739 unsigned long flags
;
2740 struct r10conf
*conf
= mddev
->private;
2741 struct list_head
*head
= &conf
->retry_list
;
2742 struct blk_plug plug
;
2744 md_check_recovery(mddev
);
2746 blk_start_plug(&plug
);
2749 flush_pending_writes(conf
);
2751 spin_lock_irqsave(&conf
->device_lock
, flags
);
2752 if (list_empty(head
)) {
2753 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2756 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2757 list_del(head
->prev
);
2759 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2761 mddev
= r10_bio
->mddev
;
2762 conf
= mddev
->private;
2763 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2764 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2765 handle_write_completed(conf
, r10_bio
);
2766 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2767 reshape_request_write(mddev
, r10_bio
);
2768 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2769 sync_request_write(mddev
, r10_bio
);
2770 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2771 recovery_request_write(mddev
, r10_bio
);
2772 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2773 handle_read_error(mddev
, r10_bio
);
2775 /* just a partial read to be scheduled from a
2778 int slot
= r10_bio
->read_slot
;
2779 generic_make_request(r10_bio
->devs
[slot
].bio
);
2783 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2784 md_check_recovery(mddev
);
2786 blk_finish_plug(&plug
);
2790 static int init_resync(struct r10conf
*conf
)
2795 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2796 BUG_ON(conf
->r10buf_pool
);
2797 conf
->have_replacement
= 0;
2798 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2799 if (conf
->mirrors
[i
].replacement
)
2800 conf
->have_replacement
= 1;
2801 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2802 if (!conf
->r10buf_pool
)
2804 conf
->next_resync
= 0;
2809 * perform a "sync" on one "block"
2811 * We need to make sure that no normal I/O request - particularly write
2812 * requests - conflict with active sync requests.
2814 * This is achieved by tracking pending requests and a 'barrier' concept
2815 * that can be installed to exclude normal IO requests.
2817 * Resync and recovery are handled very differently.
2818 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2820 * For resync, we iterate over virtual addresses, read all copies,
2821 * and update if there are differences. If only one copy is live,
2823 * For recovery, we iterate over physical addresses, read a good
2824 * value for each non-in_sync drive, and over-write.
2826 * So, for recovery we may have several outstanding complex requests for a
2827 * given address, one for each out-of-sync device. We model this by allocating
2828 * a number of r10_bio structures, one for each out-of-sync device.
2829 * As we setup these structures, we collect all bio's together into a list
2830 * which we then process collectively to add pages, and then process again
2831 * to pass to generic_make_request.
2833 * The r10_bio structures are linked using a borrowed master_bio pointer.
2834 * This link is counted in ->remaining. When the r10_bio that points to NULL
2835 * has its remaining count decremented to 0, the whole complex operation
2840 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2841 int *skipped
, int go_faster
)
2843 struct r10conf
*conf
= mddev
->private;
2844 struct r10bio
*r10_bio
;
2845 struct bio
*biolist
= NULL
, *bio
;
2846 sector_t max_sector
, nr_sectors
;
2849 sector_t sync_blocks
;
2850 sector_t sectors_skipped
= 0;
2851 int chunks_skipped
= 0;
2852 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2854 if (!conf
->r10buf_pool
)
2855 if (init_resync(conf
))
2859 max_sector
= mddev
->dev_sectors
;
2860 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2861 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2862 max_sector
= mddev
->resync_max_sectors
;
2863 if (sector_nr
>= max_sector
) {
2864 /* If we aborted, we need to abort the
2865 * sync on the 'current' bitmap chucks (there can
2866 * be several when recovering multiple devices).
2867 * as we may have started syncing it but not finished.
2868 * We can find the current address in
2869 * mddev->curr_resync, but for recovery,
2870 * we need to convert that to several
2871 * virtual addresses.
2873 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2878 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2879 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2880 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2882 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2884 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2885 bitmap_end_sync(mddev
->bitmap
, sect
,
2889 /* completed sync */
2890 if ((!mddev
->bitmap
|| conf
->fullsync
)
2891 && conf
->have_replacement
2892 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2893 /* Completed a full sync so the replacements
2894 * are now fully recovered.
2896 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2897 if (conf
->mirrors
[i
].replacement
)
2898 conf
->mirrors
[i
].replacement
2904 bitmap_close_sync(mddev
->bitmap
);
2907 return sectors_skipped
;
2910 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2911 return reshape_request(mddev
, sector_nr
, skipped
);
2913 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2914 /* if there has been nothing to do on any drive,
2915 * then there is nothing to do at all..
2918 return (max_sector
- sector_nr
) + sectors_skipped
;
2921 if (max_sector
> mddev
->resync_max
)
2922 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2924 /* make sure whole request will fit in a chunk - if chunks
2927 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2928 max_sector
> (sector_nr
| chunk_mask
))
2929 max_sector
= (sector_nr
| chunk_mask
) + 1;
2931 * If there is non-resync activity waiting for us then
2932 * put in a delay to throttle resync.
2934 if (!go_faster
&& conf
->nr_waiting
)
2935 msleep_interruptible(1000);
2937 /* Again, very different code for resync and recovery.
2938 * Both must result in an r10bio with a list of bios that
2939 * have bi_end_io, bi_sector, bi_bdev set,
2940 * and bi_private set to the r10bio.
2941 * For recovery, we may actually create several r10bios
2942 * with 2 bios in each, that correspond to the bios in the main one.
2943 * In this case, the subordinate r10bios link back through a
2944 * borrowed master_bio pointer, and the counter in the master
2945 * includes a ref from each subordinate.
2947 /* First, we decide what to do and set ->bi_end_io
2948 * To end_sync_read if we want to read, and
2949 * end_sync_write if we will want to write.
2952 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2953 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2954 /* recovery... the complicated one */
2958 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2964 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2966 if ((mirror
->rdev
== NULL
||
2967 test_bit(In_sync
, &mirror
->rdev
->flags
))
2969 (mirror
->replacement
== NULL
||
2971 &mirror
->replacement
->flags
)))
2975 /* want to reconstruct this device */
2977 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2978 if (sect
>= mddev
->resync_max_sectors
) {
2979 /* last stripe is not complete - don't
2980 * try to recover this sector.
2984 /* Unless we are doing a full sync, or a replacement
2985 * we only need to recover the block if it is set in
2988 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2990 if (sync_blocks
< max_sync
)
2991 max_sync
= sync_blocks
;
2993 mirror
->replacement
== NULL
&&
2995 /* yep, skip the sync_blocks here, but don't assume
2996 * that there will never be anything to do here
2998 chunks_skipped
= -1;
3002 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3003 raise_barrier(conf
, rb2
!= NULL
);
3004 atomic_set(&r10_bio
->remaining
, 0);
3006 r10_bio
->master_bio
= (struct bio
*)rb2
;
3008 atomic_inc(&rb2
->remaining
);
3009 r10_bio
->mddev
= mddev
;
3010 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
3011 r10_bio
->sector
= sect
;
3013 raid10_find_phys(conf
, r10_bio
);
3015 /* Need to check if the array will still be
3018 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
3019 if (conf
->mirrors
[j
].rdev
== NULL
||
3020 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
3025 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
3026 &sync_blocks
, still_degraded
);
3029 for (j
=0; j
<conf
->copies
;j
++) {
3031 int d
= r10_bio
->devs
[j
].devnum
;
3032 sector_t from_addr
, to_addr
;
3033 struct md_rdev
*rdev
;
3034 sector_t sector
, first_bad
;
3036 if (!conf
->mirrors
[d
].rdev
||
3037 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
3039 /* This is where we read from */
3041 rdev
= conf
->mirrors
[d
].rdev
;
3042 sector
= r10_bio
->devs
[j
].addr
;
3044 if (is_badblock(rdev
, sector
, max_sync
,
3045 &first_bad
, &bad_sectors
)) {
3046 if (first_bad
> sector
)
3047 max_sync
= first_bad
- sector
;
3049 bad_sectors
-= (sector
3051 if (max_sync
> bad_sectors
)
3052 max_sync
= bad_sectors
;
3056 bio
= r10_bio
->devs
[0].bio
;
3057 bio
->bi_next
= biolist
;
3059 bio
->bi_private
= r10_bio
;
3060 bio
->bi_end_io
= end_sync_read
;
3062 from_addr
= r10_bio
->devs
[j
].addr
;
3063 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
3064 bio
->bi_bdev
= rdev
->bdev
;
3065 atomic_inc(&rdev
->nr_pending
);
3066 /* and we write to 'i' (if not in_sync) */
3068 for (k
=0; k
<conf
->copies
; k
++)
3069 if (r10_bio
->devs
[k
].devnum
== i
)
3071 BUG_ON(k
== conf
->copies
);
3072 to_addr
= r10_bio
->devs
[k
].addr
;
3073 r10_bio
->devs
[0].devnum
= d
;
3074 r10_bio
->devs
[0].addr
= from_addr
;
3075 r10_bio
->devs
[1].devnum
= i
;
3076 r10_bio
->devs
[1].addr
= to_addr
;
3078 rdev
= mirror
->rdev
;
3079 if (!test_bit(In_sync
, &rdev
->flags
)) {
3080 bio
= r10_bio
->devs
[1].bio
;
3081 bio
->bi_next
= biolist
;
3083 bio
->bi_private
= r10_bio
;
3084 bio
->bi_end_io
= end_sync_write
;
3086 bio
->bi_sector
= to_addr
3087 + rdev
->data_offset
;
3088 bio
->bi_bdev
= rdev
->bdev
;
3089 atomic_inc(&r10_bio
->remaining
);
3091 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3093 /* and maybe write to replacement */
3094 bio
= r10_bio
->devs
[1].repl_bio
;
3096 bio
->bi_end_io
= NULL
;
3097 rdev
= mirror
->replacement
;
3098 /* Note: if rdev != NULL, then bio
3099 * cannot be NULL as r10buf_pool_alloc will
3100 * have allocated it.
3101 * So the second test here is pointless.
3102 * But it keeps semantic-checkers happy, and
3103 * this comment keeps human reviewers
3106 if (rdev
== NULL
|| bio
== NULL
||
3107 test_bit(Faulty
, &rdev
->flags
))
3109 bio
->bi_next
= biolist
;
3111 bio
->bi_private
= r10_bio
;
3112 bio
->bi_end_io
= end_sync_write
;
3114 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3115 bio
->bi_bdev
= rdev
->bdev
;
3116 atomic_inc(&r10_bio
->remaining
);
3119 if (j
== conf
->copies
) {
3120 /* Cannot recover, so abort the recovery or
3121 * record a bad block */
3124 atomic_dec(&rb2
->remaining
);
3127 /* problem is that there are bad blocks
3128 * on other device(s)
3131 for (k
= 0; k
< conf
->copies
; k
++)
3132 if (r10_bio
->devs
[k
].devnum
== i
)
3134 if (!test_bit(In_sync
,
3135 &mirror
->rdev
->flags
)
3136 && !rdev_set_badblocks(
3138 r10_bio
->devs
[k
].addr
,
3141 if (mirror
->replacement
&&
3142 !rdev_set_badblocks(
3143 mirror
->replacement
,
3144 r10_bio
->devs
[k
].addr
,
3149 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3151 printk(KERN_INFO
"md/raid10:%s: insufficient "
3152 "working devices for recovery.\n",
3154 mirror
->recovery_disabled
3155 = mddev
->recovery_disabled
;
3160 if (biolist
== NULL
) {
3162 struct r10bio
*rb2
= r10_bio
;
3163 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3164 rb2
->master_bio
= NULL
;
3170 /* resync. Schedule a read for every block at this virt offset */
3173 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3175 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3176 &sync_blocks
, mddev
->degraded
) &&
3177 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3178 &mddev
->recovery
)) {
3179 /* We can skip this block */
3181 return sync_blocks
+ sectors_skipped
;
3183 if (sync_blocks
< max_sync
)
3184 max_sync
= sync_blocks
;
3185 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3187 r10_bio
->mddev
= mddev
;
3188 atomic_set(&r10_bio
->remaining
, 0);
3189 raise_barrier(conf
, 0);
3190 conf
->next_resync
= sector_nr
;
3192 r10_bio
->master_bio
= NULL
;
3193 r10_bio
->sector
= sector_nr
;
3194 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3195 raid10_find_phys(conf
, r10_bio
);
3196 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3198 for (i
= 0; i
< conf
->copies
; i
++) {
3199 int d
= r10_bio
->devs
[i
].devnum
;
3200 sector_t first_bad
, sector
;
3203 if (r10_bio
->devs
[i
].repl_bio
)
3204 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3206 bio
= r10_bio
->devs
[i
].bio
;
3207 bio
->bi_end_io
= NULL
;
3208 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3209 if (conf
->mirrors
[d
].rdev
== NULL
||
3210 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3212 sector
= r10_bio
->devs
[i
].addr
;
3213 if (is_badblock(conf
->mirrors
[d
].rdev
,
3215 &first_bad
, &bad_sectors
)) {
3216 if (first_bad
> sector
)
3217 max_sync
= first_bad
- sector
;
3219 bad_sectors
-= (sector
- first_bad
);
3220 if (max_sync
> bad_sectors
)
3221 max_sync
= max_sync
;
3225 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3226 atomic_inc(&r10_bio
->remaining
);
3227 bio
->bi_next
= biolist
;
3229 bio
->bi_private
= r10_bio
;
3230 bio
->bi_end_io
= end_sync_read
;
3232 bio
->bi_sector
= sector
+
3233 conf
->mirrors
[d
].rdev
->data_offset
;
3234 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3237 if (conf
->mirrors
[d
].replacement
== NULL
||
3239 &conf
->mirrors
[d
].replacement
->flags
))
3242 /* Need to set up for writing to the replacement */
3243 bio
= r10_bio
->devs
[i
].repl_bio
;
3244 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3246 sector
= r10_bio
->devs
[i
].addr
;
3247 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3248 bio
->bi_next
= biolist
;
3250 bio
->bi_private
= r10_bio
;
3251 bio
->bi_end_io
= end_sync_write
;
3253 bio
->bi_sector
= sector
+
3254 conf
->mirrors
[d
].replacement
->data_offset
;
3255 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3260 for (i
=0; i
<conf
->copies
; i
++) {
3261 int d
= r10_bio
->devs
[i
].devnum
;
3262 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3263 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3265 if (r10_bio
->devs
[i
].repl_bio
&&
3266 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3268 conf
->mirrors
[d
].replacement
,
3277 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3279 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
3281 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
3284 bio
->bi_phys_segments
= 0;
3289 if (sector_nr
+ max_sync
< max_sector
)
3290 max_sector
= sector_nr
+ max_sync
;
3293 int len
= PAGE_SIZE
;
3294 if (sector_nr
+ (len
>>9) > max_sector
)
3295 len
= (max_sector
- sector_nr
) << 9;
3298 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3300 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3301 if (bio_add_page(bio
, page
, len
, 0))
3305 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3306 for (bio2
= biolist
;
3307 bio2
&& bio2
!= bio
;
3308 bio2
= bio2
->bi_next
) {
3309 /* remove last page from this bio */
3311 bio2
->bi_size
-= len
;
3312 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3316 nr_sectors
+= len
>>9;
3317 sector_nr
+= len
>>9;
3318 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3320 r10_bio
->sectors
= nr_sectors
;
3324 biolist
= biolist
->bi_next
;
3326 bio
->bi_next
= NULL
;
3327 r10_bio
= bio
->bi_private
;
3328 r10_bio
->sectors
= nr_sectors
;
3330 if (bio
->bi_end_io
== end_sync_read
) {
3331 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3332 generic_make_request(bio
);
3336 if (sectors_skipped
)
3337 /* pretend they weren't skipped, it makes
3338 * no important difference in this case
3340 md_done_sync(mddev
, sectors_skipped
, 1);
3342 return sectors_skipped
+ nr_sectors
;
3344 /* There is nowhere to write, so all non-sync
3345 * drives must be failed or in resync, all drives
3346 * have a bad block, so try the next chunk...
3348 if (sector_nr
+ max_sync
< max_sector
)
3349 max_sector
= sector_nr
+ max_sync
;
3351 sectors_skipped
+= (max_sector
- sector_nr
);
3353 sector_nr
= max_sector
;
3358 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3361 struct r10conf
*conf
= mddev
->private;
3364 raid_disks
= min(conf
->geo
.raid_disks
,
3365 conf
->prev
.raid_disks
);
3367 sectors
= conf
->dev_sectors
;
3369 size
= sectors
>> conf
->geo
.chunk_shift
;
3370 sector_div(size
, conf
->geo
.far_copies
);
3371 size
= size
* raid_disks
;
3372 sector_div(size
, conf
->geo
.near_copies
);
3374 return size
<< conf
->geo
.chunk_shift
;
3377 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3379 /* Calculate the number of sectors-per-device that will
3380 * actually be used, and set conf->dev_sectors and
3384 size
= size
>> conf
->geo
.chunk_shift
;
3385 sector_div(size
, conf
->geo
.far_copies
);
3386 size
= size
* conf
->geo
.raid_disks
;
3387 sector_div(size
, conf
->geo
.near_copies
);
3388 /* 'size' is now the number of chunks in the array */
3389 /* calculate "used chunks per device" */
3390 size
= size
* conf
->copies
;
3392 /* We need to round up when dividing by raid_disks to
3393 * get the stride size.
3395 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3397 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3399 if (conf
->geo
.far_offset
)
3400 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3402 sector_div(size
, conf
->geo
.far_copies
);
3403 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3407 enum geo_type
{geo_new
, geo_old
, geo_start
};
3408 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3411 int layout
, chunk
, disks
;
3414 layout
= mddev
->layout
;
3415 chunk
= mddev
->chunk_sectors
;
3416 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3419 layout
= mddev
->new_layout
;
3420 chunk
= mddev
->new_chunk_sectors
;
3421 disks
= mddev
->raid_disks
;
3423 default: /* avoid 'may be unused' warnings */
3424 case geo_start
: /* new when starting reshape - raid_disks not
3426 layout
= mddev
->new_layout
;
3427 chunk
= mddev
->new_chunk_sectors
;
3428 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3433 if (chunk
< (PAGE_SIZE
>> 9) ||
3434 !is_power_of_2(chunk
))
3437 fc
= (layout
>> 8) & 255;
3438 fo
= layout
& (1<<16);
3439 geo
->raid_disks
= disks
;
3440 geo
->near_copies
= nc
;
3441 geo
->far_copies
= fc
;
3442 geo
->far_offset
= fo
;
3443 geo
->chunk_mask
= chunk
- 1;
3444 geo
->chunk_shift
= ffz(~chunk
);
3448 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3450 struct r10conf
*conf
= NULL
;
3455 copies
= setup_geo(&geo
, mddev
, geo_new
);
3458 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3459 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3460 mdname(mddev
), PAGE_SIZE
);
3464 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3465 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3466 mdname(mddev
), mddev
->new_layout
);
3471 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3475 /* FIXME calc properly */
3476 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3477 max(0,mddev
->delta_disks
)),
3482 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3487 conf
->copies
= copies
;
3488 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3489 r10bio_pool_free
, conf
);
3490 if (!conf
->r10bio_pool
)
3493 calc_sectors(conf
, mddev
->dev_sectors
);
3494 if (mddev
->reshape_position
== MaxSector
) {
3495 conf
->prev
= conf
->geo
;
3496 conf
->reshape_progress
= MaxSector
;
3498 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3502 conf
->reshape_progress
= mddev
->reshape_position
;
3503 if (conf
->prev
.far_offset
)
3504 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3506 /* far_copies must be 1 */
3507 conf
->prev
.stride
= conf
->dev_sectors
;
3509 spin_lock_init(&conf
->device_lock
);
3510 INIT_LIST_HEAD(&conf
->retry_list
);
3512 spin_lock_init(&conf
->resync_lock
);
3513 init_waitqueue_head(&conf
->wait_barrier
);
3515 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3519 conf
->mddev
= mddev
;
3524 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3527 if (conf
->r10bio_pool
)
3528 mempool_destroy(conf
->r10bio_pool
);
3529 kfree(conf
->mirrors
);
3530 safe_put_page(conf
->tmppage
);
3533 return ERR_PTR(err
);
3536 static int run(struct mddev
*mddev
)
3538 struct r10conf
*conf
;
3539 int i
, disk_idx
, chunk_size
;
3540 struct raid10_info
*disk
;
3541 struct md_rdev
*rdev
;
3543 sector_t min_offset_diff
= 0;
3545 bool discard_supported
= false;
3547 if (mddev
->private == NULL
) {
3548 conf
= setup_conf(mddev
);
3550 return PTR_ERR(conf
);
3551 mddev
->private = conf
;
3553 conf
= mddev
->private;
3557 mddev
->thread
= conf
->thread
;
3558 conf
->thread
= NULL
;
3560 chunk_size
= mddev
->chunk_sectors
<< 9;
3562 blk_queue_max_discard_sectors(mddev
->queue
,
3563 mddev
->chunk_sectors
);
3564 blk_queue_io_min(mddev
->queue
, chunk_size
);
3565 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3566 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3568 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3569 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3572 rdev_for_each(rdev
, mddev
) {
3574 struct request_queue
*q
;
3576 disk_idx
= rdev
->raid_disk
;
3579 if (disk_idx
>= conf
->geo
.raid_disks
&&
3580 disk_idx
>= conf
->prev
.raid_disks
)
3582 disk
= conf
->mirrors
+ disk_idx
;
3584 if (test_bit(Replacement
, &rdev
->flags
)) {
3585 if (disk
->replacement
)
3587 disk
->replacement
= rdev
;
3593 q
= bdev_get_queue(rdev
->bdev
);
3594 if (q
->merge_bvec_fn
)
3595 mddev
->merge_check_needed
= 1;
3596 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3597 if (!mddev
->reshape_backwards
)
3601 if (first
|| diff
< min_offset_diff
)
3602 min_offset_diff
= diff
;
3605 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3606 rdev
->data_offset
<< 9);
3608 disk
->head_position
= 0;
3610 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3611 discard_supported
= true;
3614 if (discard_supported
)
3615 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
3617 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
3619 /* need to check that every block has at least one working mirror */
3620 if (!enough(conf
, -1)) {
3621 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3626 if (conf
->reshape_progress
!= MaxSector
) {
3627 /* must ensure that shape change is supported */
3628 if (conf
->geo
.far_copies
!= 1 &&
3629 conf
->geo
.far_offset
== 0)
3631 if (conf
->prev
.far_copies
!= 1 &&
3632 conf
->geo
.far_offset
== 0)
3636 mddev
->degraded
= 0;
3638 i
< conf
->geo
.raid_disks
3639 || i
< conf
->prev
.raid_disks
;
3642 disk
= conf
->mirrors
+ i
;
3644 if (!disk
->rdev
&& disk
->replacement
) {
3645 /* The replacement is all we have - use it */
3646 disk
->rdev
= disk
->replacement
;
3647 disk
->replacement
= NULL
;
3648 clear_bit(Replacement
, &disk
->rdev
->flags
);
3652 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3653 disk
->head_position
= 0;
3658 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3661 if (mddev
->recovery_cp
!= MaxSector
)
3662 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3663 " -- starting background reconstruction\n",
3666 "md/raid10:%s: active with %d out of %d devices\n",
3667 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3668 conf
->geo
.raid_disks
);
3670 * Ok, everything is just fine now
3672 mddev
->dev_sectors
= conf
->dev_sectors
;
3673 size
= raid10_size(mddev
, 0, 0);
3674 md_set_array_sectors(mddev
, size
);
3675 mddev
->resync_max_sectors
= size
;
3678 int stripe
= conf
->geo
.raid_disks
*
3679 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3680 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3681 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3683 /* Calculate max read-ahead size.
3684 * We need to readahead at least twice a whole stripe....
3687 stripe
/= conf
->geo
.near_copies
;
3688 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3689 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3690 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3694 if (md_integrity_register(mddev
))
3697 if (conf
->reshape_progress
!= MaxSector
) {
3698 unsigned long before_length
, after_length
;
3700 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3701 conf
->prev
.far_copies
);
3702 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3703 conf
->geo
.far_copies
);
3705 if (max(before_length
, after_length
) > min_offset_diff
) {
3706 /* This cannot work */
3707 printk("md/raid10: offset difference not enough to continue reshape\n");
3710 conf
->offset_diff
= min_offset_diff
;
3712 conf
->reshape_safe
= conf
->reshape_progress
;
3713 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3714 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3715 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3716 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3717 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3724 md_unregister_thread(&mddev
->thread
);
3725 if (conf
->r10bio_pool
)
3726 mempool_destroy(conf
->r10bio_pool
);
3727 safe_put_page(conf
->tmppage
);
3728 kfree(conf
->mirrors
);
3730 mddev
->private = NULL
;
3735 static int stop(struct mddev
*mddev
)
3737 struct r10conf
*conf
= mddev
->private;
3739 raise_barrier(conf
, 0);
3740 lower_barrier(conf
);
3742 md_unregister_thread(&mddev
->thread
);
3744 /* the unplug fn references 'conf'*/
3745 blk_sync_queue(mddev
->queue
);
3747 if (conf
->r10bio_pool
)
3748 mempool_destroy(conf
->r10bio_pool
);
3749 kfree(conf
->mirrors
);
3751 mddev
->private = NULL
;
3755 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3757 struct r10conf
*conf
= mddev
->private;
3761 raise_barrier(conf
, 0);
3764 lower_barrier(conf
);
3769 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3771 /* Resize of 'far' arrays is not supported.
3772 * For 'near' and 'offset' arrays we can set the
3773 * number of sectors used to be an appropriate multiple
3774 * of the chunk size.
3775 * For 'offset', this is far_copies*chunksize.
3776 * For 'near' the multiplier is the LCM of
3777 * near_copies and raid_disks.
3778 * So if far_copies > 1 && !far_offset, fail.
3779 * Else find LCM(raid_disks, near_copy)*far_copies and
3780 * multiply by chunk_size. Then round to this number.
3781 * This is mostly done by raid10_size()
3783 struct r10conf
*conf
= mddev
->private;
3784 sector_t oldsize
, size
;
3786 if (mddev
->reshape_position
!= MaxSector
)
3789 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3792 oldsize
= raid10_size(mddev
, 0, 0);
3793 size
= raid10_size(mddev
, sectors
, 0);
3794 if (mddev
->external_size
&&
3795 mddev
->array_sectors
> size
)
3797 if (mddev
->bitmap
) {
3798 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3802 md_set_array_sectors(mddev
, size
);
3803 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3804 revalidate_disk(mddev
->gendisk
);
3805 if (sectors
> mddev
->dev_sectors
&&
3806 mddev
->recovery_cp
> oldsize
) {
3807 mddev
->recovery_cp
= oldsize
;
3808 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3810 calc_sectors(conf
, sectors
);
3811 mddev
->dev_sectors
= conf
->dev_sectors
;
3812 mddev
->resync_max_sectors
= size
;
3816 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3818 struct md_rdev
*rdev
;
3819 struct r10conf
*conf
;
3821 if (mddev
->degraded
> 0) {
3822 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3824 return ERR_PTR(-EINVAL
);
3827 /* Set new parameters */
3828 mddev
->new_level
= 10;
3829 /* new layout: far_copies = 1, near_copies = 2 */
3830 mddev
->new_layout
= (1<<8) + 2;
3831 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3832 mddev
->delta_disks
= mddev
->raid_disks
;
3833 mddev
->raid_disks
*= 2;
3834 /* make sure it will be not marked as dirty */
3835 mddev
->recovery_cp
= MaxSector
;
3837 conf
= setup_conf(mddev
);
3838 if (!IS_ERR(conf
)) {
3839 rdev_for_each(rdev
, mddev
)
3840 if (rdev
->raid_disk
>= 0)
3841 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3848 static void *raid10_takeover(struct mddev
*mddev
)
3850 struct r0conf
*raid0_conf
;
3852 /* raid10 can take over:
3853 * raid0 - providing it has only two drives
3855 if (mddev
->level
== 0) {
3856 /* for raid0 takeover only one zone is supported */
3857 raid0_conf
= mddev
->private;
3858 if (raid0_conf
->nr_strip_zones
> 1) {
3859 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3860 " with more than one zone.\n",
3862 return ERR_PTR(-EINVAL
);
3864 return raid10_takeover_raid0(mddev
);
3866 return ERR_PTR(-EINVAL
);
3869 static int raid10_check_reshape(struct mddev
*mddev
)
3871 /* Called when there is a request to change
3872 * - layout (to ->new_layout)
3873 * - chunk size (to ->new_chunk_sectors)
3874 * - raid_disks (by delta_disks)
3875 * or when trying to restart a reshape that was ongoing.
3877 * We need to validate the request and possibly allocate
3878 * space if that might be an issue later.
3880 * Currently we reject any reshape of a 'far' mode array,
3881 * allow chunk size to change if new is generally acceptable,
3882 * allow raid_disks to increase, and allow
3883 * a switch between 'near' mode and 'offset' mode.
3885 struct r10conf
*conf
= mddev
->private;
3888 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3891 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3892 /* mustn't change number of copies */
3894 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3895 /* Cannot switch to 'far' mode */
3898 if (mddev
->array_sectors
& geo
.chunk_mask
)
3899 /* not factor of array size */
3902 if (!enough(conf
, -1))
3905 kfree(conf
->mirrors_new
);
3906 conf
->mirrors_new
= NULL
;
3907 if (mddev
->delta_disks
> 0) {
3908 /* allocate new 'mirrors' list */
3909 conf
->mirrors_new
= kzalloc(
3910 sizeof(struct raid10_info
)
3911 *(mddev
->raid_disks
+
3912 mddev
->delta_disks
),
3914 if (!conf
->mirrors_new
)
3921 * Need to check if array has failed when deciding whether to:
3923 * - remove non-faulty devices
3926 * This determination is simple when no reshape is happening.
3927 * However if there is a reshape, we need to carefully check
3928 * both the before and after sections.
3929 * This is because some failed devices may only affect one
3930 * of the two sections, and some non-in_sync devices may
3931 * be insync in the section most affected by failed devices.
3933 static int calc_degraded(struct r10conf
*conf
)
3935 int degraded
, degraded2
;
3940 /* 'prev' section first */
3941 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
3942 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3943 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3945 else if (!test_bit(In_sync
, &rdev
->flags
))
3946 /* When we can reduce the number of devices in
3947 * an array, this might not contribute to
3948 * 'degraded'. It does now.
3953 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
3957 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3958 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3959 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3961 else if (!test_bit(In_sync
, &rdev
->flags
)) {
3962 /* If reshape is increasing the number of devices,
3963 * this section has already been recovered, so
3964 * it doesn't contribute to degraded.
3967 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
3972 if (degraded2
> degraded
)
3977 static int raid10_start_reshape(struct mddev
*mddev
)
3979 /* A 'reshape' has been requested. This commits
3980 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3981 * This also checks if there are enough spares and adds them
3983 * We currently require enough spares to make the final
3984 * array non-degraded. We also require that the difference
3985 * between old and new data_offset - on each device - is
3986 * enough that we never risk over-writing.
3989 unsigned long before_length
, after_length
;
3990 sector_t min_offset_diff
= 0;
3993 struct r10conf
*conf
= mddev
->private;
3994 struct md_rdev
*rdev
;
3998 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4001 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
4004 before_length
= ((1 << conf
->prev
.chunk_shift
) *
4005 conf
->prev
.far_copies
);
4006 after_length
= ((1 << conf
->geo
.chunk_shift
) *
4007 conf
->geo
.far_copies
);
4009 rdev_for_each(rdev
, mddev
) {
4010 if (!test_bit(In_sync
, &rdev
->flags
)
4011 && !test_bit(Faulty
, &rdev
->flags
))
4013 if (rdev
->raid_disk
>= 0) {
4014 long long diff
= (rdev
->new_data_offset
4015 - rdev
->data_offset
);
4016 if (!mddev
->reshape_backwards
)
4020 if (first
|| diff
< min_offset_diff
)
4021 min_offset_diff
= diff
;
4025 if (max(before_length
, after_length
) > min_offset_diff
)
4028 if (spares
< mddev
->delta_disks
)
4031 conf
->offset_diff
= min_offset_diff
;
4032 spin_lock_irq(&conf
->device_lock
);
4033 if (conf
->mirrors_new
) {
4034 memcpy(conf
->mirrors_new
, conf
->mirrors
,
4035 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
4037 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
4038 conf
->mirrors_old
= conf
->mirrors
;
4039 conf
->mirrors
= conf
->mirrors_new
;
4040 conf
->mirrors_new
= NULL
;
4042 setup_geo(&conf
->geo
, mddev
, geo_start
);
4044 if (mddev
->reshape_backwards
) {
4045 sector_t size
= raid10_size(mddev
, 0, 0);
4046 if (size
< mddev
->array_sectors
) {
4047 spin_unlock_irq(&conf
->device_lock
);
4048 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4052 mddev
->resync_max_sectors
= size
;
4053 conf
->reshape_progress
= size
;
4055 conf
->reshape_progress
= 0;
4056 spin_unlock_irq(&conf
->device_lock
);
4058 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4059 ret
= bitmap_resize(mddev
->bitmap
,
4060 raid10_size(mddev
, 0,
4061 conf
->geo
.raid_disks
),
4066 if (mddev
->delta_disks
> 0) {
4067 rdev_for_each(rdev
, mddev
)
4068 if (rdev
->raid_disk
< 0 &&
4069 !test_bit(Faulty
, &rdev
->flags
)) {
4070 if (raid10_add_disk(mddev
, rdev
) == 0) {
4071 if (rdev
->raid_disk
>=
4072 conf
->prev
.raid_disks
)
4073 set_bit(In_sync
, &rdev
->flags
);
4075 rdev
->recovery_offset
= 0;
4077 if (sysfs_link_rdev(mddev
, rdev
))
4078 /* Failure here is OK */;
4080 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4081 && !test_bit(Faulty
, &rdev
->flags
)) {
4082 /* This is a spare that was manually added */
4083 set_bit(In_sync
, &rdev
->flags
);
4086 /* When a reshape changes the number of devices,
4087 * ->degraded is measured against the larger of the
4088 * pre and post numbers.
4090 spin_lock_irq(&conf
->device_lock
);
4091 mddev
->degraded
= calc_degraded(conf
);
4092 spin_unlock_irq(&conf
->device_lock
);
4093 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4094 mddev
->reshape_position
= conf
->reshape_progress
;
4095 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4097 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4098 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4099 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4100 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4102 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4104 if (!mddev
->sync_thread
) {
4108 conf
->reshape_checkpoint
= jiffies
;
4109 md_wakeup_thread(mddev
->sync_thread
);
4110 md_new_event(mddev
);
4114 mddev
->recovery
= 0;
4115 spin_lock_irq(&conf
->device_lock
);
4116 conf
->geo
= conf
->prev
;
4117 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4118 rdev_for_each(rdev
, mddev
)
4119 rdev
->new_data_offset
= rdev
->data_offset
;
4121 conf
->reshape_progress
= MaxSector
;
4122 mddev
->reshape_position
= MaxSector
;
4123 spin_unlock_irq(&conf
->device_lock
);
4127 /* Calculate the last device-address that could contain
4128 * any block from the chunk that includes the array-address 's'
4129 * and report the next address.
4130 * i.e. the address returned will be chunk-aligned and after
4131 * any data that is in the chunk containing 's'.
4133 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4135 s
= (s
| geo
->chunk_mask
) + 1;
4136 s
>>= geo
->chunk_shift
;
4137 s
*= geo
->near_copies
;
4138 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4139 s
*= geo
->far_copies
;
4140 s
<<= geo
->chunk_shift
;
4144 /* Calculate the first device-address that could contain
4145 * any block from the chunk that includes the array-address 's'.
4146 * This too will be the start of a chunk
4148 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4150 s
>>= geo
->chunk_shift
;
4151 s
*= geo
->near_copies
;
4152 sector_div(s
, geo
->raid_disks
);
4153 s
*= geo
->far_copies
;
4154 s
<<= geo
->chunk_shift
;
4158 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4161 /* We simply copy at most one chunk (smallest of old and new)
4162 * at a time, possibly less if that exceeds RESYNC_PAGES,
4163 * or we hit a bad block or something.
4164 * This might mean we pause for normal IO in the middle of
4165 * a chunk, but that is not a problem was mddev->reshape_position
4166 * can record any location.
4168 * If we will want to write to a location that isn't
4169 * yet recorded as 'safe' (i.e. in metadata on disk) then
4170 * we need to flush all reshape requests and update the metadata.
4172 * When reshaping forwards (e.g. to more devices), we interpret
4173 * 'safe' as the earliest block which might not have been copied
4174 * down yet. We divide this by previous stripe size and multiply
4175 * by previous stripe length to get lowest device offset that we
4176 * cannot write to yet.
4177 * We interpret 'sector_nr' as an address that we want to write to.
4178 * From this we use last_device_address() to find where we might
4179 * write to, and first_device_address on the 'safe' position.
4180 * If this 'next' write position is after the 'safe' position,
4181 * we must update the metadata to increase the 'safe' position.
4183 * When reshaping backwards, we round in the opposite direction
4184 * and perform the reverse test: next write position must not be
4185 * less than current safe position.
4187 * In all this the minimum difference in data offsets
4188 * (conf->offset_diff - always positive) allows a bit of slack,
4189 * so next can be after 'safe', but not by more than offset_disk
4191 * We need to prepare all the bios here before we start any IO
4192 * to ensure the size we choose is acceptable to all devices.
4193 * The means one for each copy for write-out and an extra one for
4195 * We store the read-in bio in ->master_bio and the others in
4196 * ->devs[x].bio and ->devs[x].repl_bio.
4198 struct r10conf
*conf
= mddev
->private;
4199 struct r10bio
*r10_bio
;
4200 sector_t next
, safe
, last
;
4204 struct md_rdev
*rdev
;
4207 struct bio
*bio
, *read_bio
;
4208 int sectors_done
= 0;
4210 if (sector_nr
== 0) {
4211 /* If restarting in the middle, skip the initial sectors */
4212 if (mddev
->reshape_backwards
&&
4213 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4214 sector_nr
= (raid10_size(mddev
, 0, 0)
4215 - conf
->reshape_progress
);
4216 } else if (!mddev
->reshape_backwards
&&
4217 conf
->reshape_progress
> 0)
4218 sector_nr
= conf
->reshape_progress
;
4220 mddev
->curr_resync_completed
= sector_nr
;
4221 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4227 /* We don't use sector_nr to track where we are up to
4228 * as that doesn't work well for ->reshape_backwards.
4229 * So just use ->reshape_progress.
4231 if (mddev
->reshape_backwards
) {
4232 /* 'next' is the earliest device address that we might
4233 * write to for this chunk in the new layout
4235 next
= first_dev_address(conf
->reshape_progress
- 1,
4238 /* 'safe' is the last device address that we might read from
4239 * in the old layout after a restart
4241 safe
= last_dev_address(conf
->reshape_safe
- 1,
4244 if (next
+ conf
->offset_diff
< safe
)
4247 last
= conf
->reshape_progress
- 1;
4248 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4249 & conf
->prev
.chunk_mask
);
4250 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4251 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4253 /* 'next' is after the last device address that we
4254 * might write to for this chunk in the new layout
4256 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4258 /* 'safe' is the earliest device address that we might
4259 * read from in the old layout after a restart
4261 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4263 /* Need to update metadata if 'next' might be beyond 'safe'
4264 * as that would possibly corrupt data
4266 if (next
> safe
+ conf
->offset_diff
)
4269 sector_nr
= conf
->reshape_progress
;
4270 last
= sector_nr
| (conf
->geo
.chunk_mask
4271 & conf
->prev
.chunk_mask
);
4273 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4274 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4278 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4279 /* Need to update reshape_position in metadata */
4281 mddev
->reshape_position
= conf
->reshape_progress
;
4282 if (mddev
->reshape_backwards
)
4283 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4284 - conf
->reshape_progress
;
4286 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4287 conf
->reshape_checkpoint
= jiffies
;
4288 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4289 md_wakeup_thread(mddev
->thread
);
4290 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4291 kthread_should_stop());
4292 conf
->reshape_safe
= mddev
->reshape_position
;
4293 allow_barrier(conf
);
4297 /* Now schedule reads for blocks from sector_nr to last */
4298 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4299 raise_barrier(conf
, sectors_done
!= 0);
4300 atomic_set(&r10_bio
->remaining
, 0);
4301 r10_bio
->mddev
= mddev
;
4302 r10_bio
->sector
= sector_nr
;
4303 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4304 r10_bio
->sectors
= last
- sector_nr
+ 1;
4305 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4306 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4309 /* Cannot read from here, so need to record bad blocks
4310 * on all the target devices.
4313 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4314 return sectors_done
;
4317 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4319 read_bio
->bi_bdev
= rdev
->bdev
;
4320 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4321 + rdev
->data_offset
);
4322 read_bio
->bi_private
= r10_bio
;
4323 read_bio
->bi_end_io
= end_sync_read
;
4324 read_bio
->bi_rw
= READ
;
4325 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4326 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4327 read_bio
->bi_vcnt
= 0;
4328 read_bio
->bi_idx
= 0;
4329 read_bio
->bi_size
= 0;
4330 r10_bio
->master_bio
= read_bio
;
4331 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4333 /* Now find the locations in the new layout */
4334 __raid10_find_phys(&conf
->geo
, r10_bio
);
4337 read_bio
->bi_next
= NULL
;
4339 for (s
= 0; s
< conf
->copies
*2; s
++) {
4341 int d
= r10_bio
->devs
[s
/2].devnum
;
4342 struct md_rdev
*rdev2
;
4344 rdev2
= conf
->mirrors
[d
].replacement
;
4345 b
= r10_bio
->devs
[s
/2].repl_bio
;
4347 rdev2
= conf
->mirrors
[d
].rdev
;
4348 b
= r10_bio
->devs
[s
/2].bio
;
4350 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4352 b
->bi_bdev
= rdev2
->bdev
;
4353 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4354 b
->bi_private
= r10_bio
;
4355 b
->bi_end_io
= end_reshape_write
;
4357 b
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4358 b
->bi_flags
|= 1 << BIO_UPTODATE
;
4366 /* Now add as many pages as possible to all of these bios. */
4369 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4370 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4371 int len
= (max_sectors
- s
) << 9;
4372 if (len
> PAGE_SIZE
)
4374 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4376 if (bio_add_page(bio
, page
, len
, 0))
4379 /* Didn't fit, must stop */
4381 bio2
&& bio2
!= bio
;
4382 bio2
= bio2
->bi_next
) {
4383 /* Remove last page from this bio */
4385 bio2
->bi_size
-= len
;
4386 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4390 sector_nr
+= len
>> 9;
4391 nr_sectors
+= len
>> 9;
4394 r10_bio
->sectors
= nr_sectors
;
4396 /* Now submit the read */
4397 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4398 atomic_inc(&r10_bio
->remaining
);
4399 read_bio
->bi_next
= NULL
;
4400 generic_make_request(read_bio
);
4401 sector_nr
+= nr_sectors
;
4402 sectors_done
+= nr_sectors
;
4403 if (sector_nr
<= last
)
4406 /* Now that we have done the whole section we can
4407 * update reshape_progress
4409 if (mddev
->reshape_backwards
)
4410 conf
->reshape_progress
-= sectors_done
;
4412 conf
->reshape_progress
+= sectors_done
;
4414 return sectors_done
;
4417 static void end_reshape_request(struct r10bio
*r10_bio
);
4418 static int handle_reshape_read_error(struct mddev
*mddev
,
4419 struct r10bio
*r10_bio
);
4420 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4422 /* Reshape read completed. Hopefully we have a block
4424 * If we got a read error then we do sync 1-page reads from
4425 * elsewhere until we find the data - or give up.
4427 struct r10conf
*conf
= mddev
->private;
4430 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4431 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4432 /* Reshape has been aborted */
4433 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4437 /* We definitely have the data in the pages, schedule the
4440 atomic_set(&r10_bio
->remaining
, 1);
4441 for (s
= 0; s
< conf
->copies
*2; s
++) {
4443 int d
= r10_bio
->devs
[s
/2].devnum
;
4444 struct md_rdev
*rdev
;
4446 rdev
= conf
->mirrors
[d
].replacement
;
4447 b
= r10_bio
->devs
[s
/2].repl_bio
;
4449 rdev
= conf
->mirrors
[d
].rdev
;
4450 b
= r10_bio
->devs
[s
/2].bio
;
4452 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4454 atomic_inc(&rdev
->nr_pending
);
4455 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4456 atomic_inc(&r10_bio
->remaining
);
4458 generic_make_request(b
);
4460 end_reshape_request(r10_bio
);
4463 static void end_reshape(struct r10conf
*conf
)
4465 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4468 spin_lock_irq(&conf
->device_lock
);
4469 conf
->prev
= conf
->geo
;
4470 md_finish_reshape(conf
->mddev
);
4472 conf
->reshape_progress
= MaxSector
;
4473 spin_unlock_irq(&conf
->device_lock
);
4475 /* read-ahead size must cover two whole stripes, which is
4476 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4478 if (conf
->mddev
->queue
) {
4479 int stripe
= conf
->geo
.raid_disks
*
4480 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4481 stripe
/= conf
->geo
.near_copies
;
4482 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4483 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4489 static int handle_reshape_read_error(struct mddev
*mddev
,
4490 struct r10bio
*r10_bio
)
4492 /* Use sync reads to get the blocks from somewhere else */
4493 int sectors
= r10_bio
->sectors
;
4494 struct r10conf
*conf
= mddev
->private;
4496 struct r10bio r10_bio
;
4497 struct r10dev devs
[conf
->copies
];
4499 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4502 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4504 r10b
->sector
= r10_bio
->sector
;
4505 __raid10_find_phys(&conf
->prev
, r10b
);
4510 int first_slot
= slot
;
4512 if (s
> (PAGE_SIZE
>> 9))
4516 int d
= r10b
->devs
[slot
].devnum
;
4517 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4520 test_bit(Faulty
, &rdev
->flags
) ||
4521 !test_bit(In_sync
, &rdev
->flags
))
4524 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4525 success
= sync_page_io(rdev
,
4534 if (slot
>= conf
->copies
)
4536 if (slot
== first_slot
)
4540 /* couldn't read this block, must give up */
4541 set_bit(MD_RECOVERY_INTR
,
4551 static void end_reshape_write(struct bio
*bio
, int error
)
4553 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4554 struct r10bio
*r10_bio
= bio
->bi_private
;
4555 struct mddev
*mddev
= r10_bio
->mddev
;
4556 struct r10conf
*conf
= mddev
->private;
4560 struct md_rdev
*rdev
= NULL
;
4562 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4564 rdev
= conf
->mirrors
[d
].replacement
;
4567 rdev
= conf
->mirrors
[d
].rdev
;
4571 /* FIXME should record badblock */
4572 md_error(mddev
, rdev
);
4575 rdev_dec_pending(rdev
, mddev
);
4576 end_reshape_request(r10_bio
);
4579 static void end_reshape_request(struct r10bio
*r10_bio
)
4581 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4583 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4584 bio_put(r10_bio
->master_bio
);
4588 static void raid10_finish_reshape(struct mddev
*mddev
)
4590 struct r10conf
*conf
= mddev
->private;
4592 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4595 if (mddev
->delta_disks
> 0) {
4596 sector_t size
= raid10_size(mddev
, 0, 0);
4597 md_set_array_sectors(mddev
, size
);
4598 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4599 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4600 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4602 mddev
->resync_max_sectors
= size
;
4603 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4604 revalidate_disk(mddev
->gendisk
);
4607 for (d
= conf
->geo
.raid_disks
;
4608 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4610 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4612 clear_bit(In_sync
, &rdev
->flags
);
4613 rdev
= conf
->mirrors
[d
].replacement
;
4615 clear_bit(In_sync
, &rdev
->flags
);
4618 mddev
->layout
= mddev
->new_layout
;
4619 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4620 mddev
->reshape_position
= MaxSector
;
4621 mddev
->delta_disks
= 0;
4622 mddev
->reshape_backwards
= 0;
4625 static struct md_personality raid10_personality
=
4629 .owner
= THIS_MODULE
,
4630 .make_request
= make_request
,
4634 .error_handler
= error
,
4635 .hot_add_disk
= raid10_add_disk
,
4636 .hot_remove_disk
= raid10_remove_disk
,
4637 .spare_active
= raid10_spare_active
,
4638 .sync_request
= sync_request
,
4639 .quiesce
= raid10_quiesce
,
4640 .size
= raid10_size
,
4641 .resize
= raid10_resize
,
4642 .takeover
= raid10_takeover
,
4643 .check_reshape
= raid10_check_reshape
,
4644 .start_reshape
= raid10_start_reshape
,
4645 .finish_reshape
= raid10_finish_reshape
,
4648 static int __init
raid_init(void)
4650 return register_md_personality(&raid10_personality
);
4653 static void raid_exit(void)
4655 unregister_md_personality(&raid10_personality
);
4658 module_init(raid_init
);
4659 module_exit(raid_exit
);
4660 MODULE_LICENSE("GPL");
4661 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4662 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4663 MODULE_ALIAS("md-raid10");
4664 MODULE_ALIAS("md-level-10");
4666 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);