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 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1060 struct r10conf
*conf
= mddev
->private;
1061 struct r10bio
*r10_bio
;
1062 struct bio
*read_bio
;
1064 sector_t chunk_mask
= (conf
->geo
.chunk_mask
& conf
->prev
.chunk_mask
);
1065 int chunk_sects
= chunk_mask
+ 1;
1066 const int rw
= bio_data_dir(bio
);
1067 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1068 const unsigned long do_fua
= (bio
->bi_rw
& REQ_FUA
);
1069 const unsigned long do_discard
= (bio
->bi_rw
1070 & (REQ_DISCARD
| REQ_SECURE
));
1071 unsigned long flags
;
1072 struct md_rdev
*blocked_rdev
;
1073 int sectors_handled
;
1077 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)) {
1078 md_flush_request(mddev
, bio
);
1082 /* If this request crosses a chunk boundary, we need to
1083 * split it. This will only happen for 1 PAGE (or less) requests.
1085 if (unlikely((bio
->bi_sector
& chunk_mask
) + (bio
->bi_size
>> 9)
1087 && (conf
->geo
.near_copies
< conf
->geo
.raid_disks
1088 || conf
->prev
.near_copies
< conf
->prev
.raid_disks
))) {
1089 struct bio_pair
*bp
;
1090 /* Sanity check -- queue functions should prevent this happening */
1091 if ((bio
->bi_vcnt
!= 1 && bio
->bi_vcnt
!= 0) ||
1094 /* This is a one page bio that upper layers
1095 * refuse to split for us, so we need to split it.
1098 chunk_sects
- (bio
->bi_sector
& (chunk_sects
- 1)) );
1100 /* Each of these 'make_request' calls will call 'wait_barrier'.
1101 * If the first succeeds but the second blocks due to the resync
1102 * thread raising the barrier, we will deadlock because the
1103 * IO to the underlying device will be queued in generic_make_request
1104 * and will never complete, so will never reduce nr_pending.
1105 * So increment nr_waiting here so no new raise_barriers will
1106 * succeed, and so the second wait_barrier cannot block.
1108 spin_lock_irq(&conf
->resync_lock
);
1110 spin_unlock_irq(&conf
->resync_lock
);
1112 make_request(mddev
, &bp
->bio1
);
1113 make_request(mddev
, &bp
->bio2
);
1115 spin_lock_irq(&conf
->resync_lock
);
1117 wake_up(&conf
->wait_barrier
);
1118 spin_unlock_irq(&conf
->resync_lock
);
1120 bio_pair_release(bp
);
1123 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1124 " or bigger than %dk %llu %d\n", mdname(mddev
), chunk_sects
/2,
1125 (unsigned long long)bio
->bi_sector
, bio
->bi_size
>> 10);
1131 md_write_start(mddev
, bio
);
1134 * Register the new request and wait if the reconstruction
1135 * thread has put up a bar for new requests.
1136 * Continue immediately if no resync is active currently.
1140 sectors
= bio
->bi_size
>> 9;
1141 while (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1142 bio
->bi_sector
< conf
->reshape_progress
&&
1143 bio
->bi_sector
+ sectors
> conf
->reshape_progress
) {
1144 /* IO spans the reshape position. Need to wait for
1147 allow_barrier(conf
);
1148 wait_event(conf
->wait_barrier
,
1149 conf
->reshape_progress
<= bio
->bi_sector
||
1150 conf
->reshape_progress
>= bio
->bi_sector
+ sectors
);
1153 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
1154 bio_data_dir(bio
) == WRITE
&&
1155 (mddev
->reshape_backwards
1156 ? (bio
->bi_sector
< conf
->reshape_safe
&&
1157 bio
->bi_sector
+ sectors
> conf
->reshape_progress
)
1158 : (bio
->bi_sector
+ sectors
> conf
->reshape_safe
&&
1159 bio
->bi_sector
< conf
->reshape_progress
))) {
1160 /* Need to update reshape_position in metadata */
1161 mddev
->reshape_position
= conf
->reshape_progress
;
1162 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1163 set_bit(MD_CHANGE_PENDING
, &mddev
->flags
);
1164 md_wakeup_thread(mddev
->thread
);
1165 wait_event(mddev
->sb_wait
,
1166 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
));
1168 conf
->reshape_safe
= mddev
->reshape_position
;
1171 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1173 r10_bio
->master_bio
= bio
;
1174 r10_bio
->sectors
= sectors
;
1176 r10_bio
->mddev
= mddev
;
1177 r10_bio
->sector
= bio
->bi_sector
;
1180 /* We might need to issue multiple reads to different
1181 * devices if there are bad blocks around, so we keep
1182 * track of the number of reads in bio->bi_phys_segments.
1183 * If this is 0, there is only one r10_bio and no locking
1184 * will be needed when the request completes. If it is
1185 * non-zero, then it is the number of not-completed requests.
1187 bio
->bi_phys_segments
= 0;
1188 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1192 * read balancing logic:
1194 struct md_rdev
*rdev
;
1198 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
1200 raid_end_bio_io(r10_bio
);
1203 slot
= r10_bio
->read_slot
;
1205 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1206 md_trim_bio(read_bio
, r10_bio
->sector
- bio
->bi_sector
,
1209 r10_bio
->devs
[slot
].bio
= read_bio
;
1210 r10_bio
->devs
[slot
].rdev
= rdev
;
1212 read_bio
->bi_sector
= r10_bio
->devs
[slot
].addr
+
1213 choose_data_offset(r10_bio
, rdev
);
1214 read_bio
->bi_bdev
= rdev
->bdev
;
1215 read_bio
->bi_end_io
= raid10_end_read_request
;
1216 read_bio
->bi_rw
= READ
| do_sync
;
1217 read_bio
->bi_private
= r10_bio
;
1219 if (max_sectors
< r10_bio
->sectors
) {
1220 /* Could not read all from this device, so we will
1221 * need another r10_bio.
1223 sectors_handled
= (r10_bio
->sectors
+ max_sectors
1225 r10_bio
->sectors
= max_sectors
;
1226 spin_lock_irq(&conf
->device_lock
);
1227 if (bio
->bi_phys_segments
== 0)
1228 bio
->bi_phys_segments
= 2;
1230 bio
->bi_phys_segments
++;
1231 spin_unlock(&conf
->device_lock
);
1232 /* Cannot call generic_make_request directly
1233 * as that will be queued in __generic_make_request
1234 * and subsequent mempool_alloc might block
1235 * waiting for it. so hand bio over to raid10d.
1237 reschedule_retry(r10_bio
);
1239 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1241 r10_bio
->master_bio
= bio
;
1242 r10_bio
->sectors
= ((bio
->bi_size
>> 9)
1245 r10_bio
->mddev
= mddev
;
1246 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1249 generic_make_request(read_bio
);
1256 if (conf
->pending_count
>= max_queued_requests
) {
1257 md_wakeup_thread(mddev
->thread
);
1258 wait_event(conf
->wait_barrier
,
1259 conf
->pending_count
< max_queued_requests
);
1261 /* first select target devices under rcu_lock and
1262 * inc refcount on their rdev. Record them by setting
1264 * If there are known/acknowledged bad blocks on any device
1265 * on which we have seen a write error, we want to avoid
1266 * writing to those blocks. This potentially requires several
1267 * writes to write around the bad blocks. Each set of writes
1268 * gets its own r10_bio with a set of bios attached. The number
1269 * of r10_bios is recored in bio->bi_phys_segments just as with
1273 r10_bio
->read_slot
= -1; /* make sure repl_bio gets freed */
1274 raid10_find_phys(conf
, r10_bio
);
1276 blocked_rdev
= NULL
;
1278 max_sectors
= r10_bio
->sectors
;
1280 for (i
= 0; i
< conf
->copies
; i
++) {
1281 int d
= r10_bio
->devs
[i
].devnum
;
1282 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
1283 struct md_rdev
*rrdev
= rcu_dereference(
1284 conf
->mirrors
[d
].replacement
);
1287 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1288 atomic_inc(&rdev
->nr_pending
);
1289 blocked_rdev
= rdev
;
1292 if (rrdev
&& unlikely(test_bit(Blocked
, &rrdev
->flags
))) {
1293 atomic_inc(&rrdev
->nr_pending
);
1294 blocked_rdev
= rrdev
;
1297 if (rrdev
&& (test_bit(Faulty
, &rrdev
->flags
)
1298 || test_bit(Unmerged
, &rrdev
->flags
)))
1301 r10_bio
->devs
[i
].bio
= NULL
;
1302 r10_bio
->devs
[i
].repl_bio
= NULL
;
1303 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
1304 test_bit(Unmerged
, &rdev
->flags
)) {
1305 set_bit(R10BIO_Degraded
, &r10_bio
->state
);
1308 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1310 sector_t dev_sector
= r10_bio
->devs
[i
].addr
;
1314 is_bad
= is_badblock(rdev
, dev_sector
,
1316 &first_bad
, &bad_sectors
);
1318 /* Mustn't write here until the bad block
1321 atomic_inc(&rdev
->nr_pending
);
1322 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1323 blocked_rdev
= rdev
;
1326 if (is_bad
&& first_bad
<= dev_sector
) {
1327 /* Cannot write here at all */
1328 bad_sectors
-= (dev_sector
- first_bad
);
1329 if (bad_sectors
< max_sectors
)
1330 /* Mustn't write more than bad_sectors
1331 * to other devices yet
1333 max_sectors
= bad_sectors
;
1334 /* We don't set R10BIO_Degraded as that
1335 * only applies if the disk is missing,
1336 * so it might be re-added, and we want to
1337 * know to recover this chunk.
1338 * In this case the device is here, and the
1339 * fact that this chunk is not in-sync is
1340 * recorded in the bad block log.
1345 int good_sectors
= first_bad
- dev_sector
;
1346 if (good_sectors
< max_sectors
)
1347 max_sectors
= good_sectors
;
1350 r10_bio
->devs
[i
].bio
= bio
;
1351 atomic_inc(&rdev
->nr_pending
);
1353 r10_bio
->devs
[i
].repl_bio
= bio
;
1354 atomic_inc(&rrdev
->nr_pending
);
1359 if (unlikely(blocked_rdev
)) {
1360 /* Have to wait for this device to get unblocked, then retry */
1364 for (j
= 0; j
< i
; j
++) {
1365 if (r10_bio
->devs
[j
].bio
) {
1366 d
= r10_bio
->devs
[j
].devnum
;
1367 rdev_dec_pending(conf
->mirrors
[d
].rdev
, mddev
);
1369 if (r10_bio
->devs
[j
].repl_bio
) {
1370 struct md_rdev
*rdev
;
1371 d
= r10_bio
->devs
[j
].devnum
;
1372 rdev
= conf
->mirrors
[d
].replacement
;
1374 /* Race with remove_disk */
1376 rdev
= conf
->mirrors
[d
].rdev
;
1378 rdev_dec_pending(rdev
, mddev
);
1381 allow_barrier(conf
);
1382 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1387 if (max_sectors
< r10_bio
->sectors
) {
1388 /* We are splitting this into multiple parts, so
1389 * we need to prepare for allocating another r10_bio.
1391 r10_bio
->sectors
= max_sectors
;
1392 spin_lock_irq(&conf
->device_lock
);
1393 if (bio
->bi_phys_segments
== 0)
1394 bio
->bi_phys_segments
= 2;
1396 bio
->bi_phys_segments
++;
1397 spin_unlock_irq(&conf
->device_lock
);
1399 sectors_handled
= r10_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1401 atomic_set(&r10_bio
->remaining
, 1);
1402 bitmap_startwrite(mddev
->bitmap
, r10_bio
->sector
, r10_bio
->sectors
, 0);
1404 for (i
= 0; i
< conf
->copies
; i
++) {
1406 int d
= r10_bio
->devs
[i
].devnum
;
1407 if (!r10_bio
->devs
[i
].bio
)
1410 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1411 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1413 r10_bio
->devs
[i
].bio
= mbio
;
1415 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1416 choose_data_offset(r10_bio
,
1417 conf
->mirrors
[d
].rdev
));
1418 mbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
1419 mbio
->bi_end_io
= raid10_end_write_request
;
1420 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
| do_discard
;
1421 mbio
->bi_private
= r10_bio
;
1423 atomic_inc(&r10_bio
->remaining
);
1424 spin_lock_irqsave(&conf
->device_lock
, flags
);
1425 bio_list_add(&conf
->pending_bio_list
, mbio
);
1426 conf
->pending_count
++;
1427 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1428 if (!mddev_check_plugged(mddev
))
1429 md_wakeup_thread(mddev
->thread
);
1431 if (!r10_bio
->devs
[i
].repl_bio
)
1434 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1435 md_trim_bio(mbio
, r10_bio
->sector
- bio
->bi_sector
,
1437 r10_bio
->devs
[i
].repl_bio
= mbio
;
1439 /* We are actively writing to the original device
1440 * so it cannot disappear, so the replacement cannot
1443 mbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
1446 conf
->mirrors
[d
].replacement
));
1447 mbio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
1448 mbio
->bi_end_io
= raid10_end_write_request
;
1449 mbio
->bi_rw
= WRITE
| do_sync
| do_fua
| do_discard
;
1450 mbio
->bi_private
= r10_bio
;
1452 atomic_inc(&r10_bio
->remaining
);
1453 spin_lock_irqsave(&conf
->device_lock
, flags
);
1454 bio_list_add(&conf
->pending_bio_list
, mbio
);
1455 conf
->pending_count
++;
1456 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1457 if (!mddev_check_plugged(mddev
))
1458 md_wakeup_thread(mddev
->thread
);
1461 /* Don't remove the bias on 'remaining' (one_write_done) until
1462 * after checking if we need to go around again.
1465 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1466 one_write_done(r10_bio
);
1467 /* We need another r10_bio. It has already been counted
1468 * in bio->bi_phys_segments.
1470 r10_bio
= mempool_alloc(conf
->r10bio_pool
, GFP_NOIO
);
1472 r10_bio
->master_bio
= bio
;
1473 r10_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1475 r10_bio
->mddev
= mddev
;
1476 r10_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1480 one_write_done(r10_bio
);
1482 /* In case raid10d snuck in to freeze_array */
1483 wake_up(&conf
->wait_barrier
);
1486 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1488 struct r10conf
*conf
= mddev
->private;
1491 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
)
1492 seq_printf(seq
, " %dK chunks", mddev
->chunk_sectors
/ 2);
1493 if (conf
->geo
.near_copies
> 1)
1494 seq_printf(seq
, " %d near-copies", conf
->geo
.near_copies
);
1495 if (conf
->geo
.far_copies
> 1) {
1496 if (conf
->geo
.far_offset
)
1497 seq_printf(seq
, " %d offset-copies", conf
->geo
.far_copies
);
1499 seq_printf(seq
, " %d far-copies", conf
->geo
.far_copies
);
1501 seq_printf(seq
, " [%d/%d] [", conf
->geo
.raid_disks
,
1502 conf
->geo
.raid_disks
- mddev
->degraded
);
1503 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
1504 seq_printf(seq
, "%s",
1505 conf
->mirrors
[i
].rdev
&&
1506 test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ? "U" : "_");
1507 seq_printf(seq
, "]");
1510 /* check if there are enough drives for
1511 * every block to appear on atleast one.
1512 * Don't consider the device numbered 'ignore'
1513 * as we might be about to remove it.
1515 static int _enough(struct r10conf
*conf
, struct geom
*geo
, int ignore
)
1520 int n
= conf
->copies
;
1523 if (conf
->mirrors
[first
].rdev
&&
1526 first
= (first
+1) % geo
->raid_disks
;
1530 } while (first
!= 0);
1534 static int enough(struct r10conf
*conf
, int ignore
)
1536 return _enough(conf
, &conf
->geo
, ignore
) &&
1537 _enough(conf
, &conf
->prev
, ignore
);
1540 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1542 char b
[BDEVNAME_SIZE
];
1543 struct r10conf
*conf
= mddev
->private;
1546 * If it is not operational, then we have already marked it as dead
1547 * else if it is the last working disks, ignore the error, let the
1548 * next level up know.
1549 * else mark the drive as failed
1551 if (test_bit(In_sync
, &rdev
->flags
)
1552 && !enough(conf
, rdev
->raid_disk
))
1554 * Don't fail the drive, just return an IO error.
1557 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1558 unsigned long flags
;
1559 spin_lock_irqsave(&conf
->device_lock
, flags
);
1561 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1563 * if recovery is running, make sure it aborts.
1565 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1567 set_bit(Blocked
, &rdev
->flags
);
1568 set_bit(Faulty
, &rdev
->flags
);
1569 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1571 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1572 "md/raid10:%s: Operation continuing on %d devices.\n",
1573 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1574 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
);
1577 static void print_conf(struct r10conf
*conf
)
1580 struct raid10_info
*tmp
;
1582 printk(KERN_DEBUG
"RAID10 conf printout:\n");
1584 printk(KERN_DEBUG
"(!conf)\n");
1587 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->geo
.raid_disks
- conf
->mddev
->degraded
,
1588 conf
->geo
.raid_disks
);
1590 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1591 char b
[BDEVNAME_SIZE
];
1592 tmp
= conf
->mirrors
+ i
;
1594 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1595 i
, !test_bit(In_sync
, &tmp
->rdev
->flags
),
1596 !test_bit(Faulty
, &tmp
->rdev
->flags
),
1597 bdevname(tmp
->rdev
->bdev
,b
));
1601 static void close_sync(struct r10conf
*conf
)
1604 allow_barrier(conf
);
1606 mempool_destroy(conf
->r10buf_pool
);
1607 conf
->r10buf_pool
= NULL
;
1610 static int raid10_spare_active(struct mddev
*mddev
)
1613 struct r10conf
*conf
= mddev
->private;
1614 struct raid10_info
*tmp
;
1616 unsigned long flags
;
1619 * Find all non-in_sync disks within the RAID10 configuration
1620 * and mark them in_sync
1622 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
1623 tmp
= conf
->mirrors
+ i
;
1624 if (tmp
->replacement
1625 && tmp
->replacement
->recovery_offset
== MaxSector
1626 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
1627 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
1628 /* Replacement has just become active */
1630 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
1633 /* Replaced device not technically faulty,
1634 * but we need to be sure it gets removed
1635 * and never re-added.
1637 set_bit(Faulty
, &tmp
->rdev
->flags
);
1638 sysfs_notify_dirent_safe(
1639 tmp
->rdev
->sysfs_state
);
1641 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
1642 } else if (tmp
->rdev
1643 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
1644 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
1646 sysfs_notify_dirent(tmp
->rdev
->sysfs_state
);
1649 spin_lock_irqsave(&conf
->device_lock
, flags
);
1650 mddev
->degraded
-= count
;
1651 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1658 static int raid10_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1660 struct r10conf
*conf
= mddev
->private;
1664 int last
= conf
->geo
.raid_disks
- 1;
1665 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1667 if (mddev
->recovery_cp
< MaxSector
)
1668 /* only hot-add to in-sync arrays, as recovery is
1669 * very different from resync
1672 if (rdev
->saved_raid_disk
< 0 && !_enough(conf
, &conf
->prev
, -1))
1675 if (rdev
->raid_disk
>= 0)
1676 first
= last
= rdev
->raid_disk
;
1678 if (q
->merge_bvec_fn
) {
1679 set_bit(Unmerged
, &rdev
->flags
);
1680 mddev
->merge_check_needed
= 1;
1683 if (rdev
->saved_raid_disk
>= first
&&
1684 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1685 mirror
= rdev
->saved_raid_disk
;
1688 for ( ; mirror
<= last
; mirror
++) {
1689 struct raid10_info
*p
= &conf
->mirrors
[mirror
];
1690 if (p
->recovery_disabled
== mddev
->recovery_disabled
)
1693 if (!test_bit(WantReplacement
, &p
->rdev
->flags
) ||
1694 p
->replacement
!= NULL
)
1696 clear_bit(In_sync
, &rdev
->flags
);
1697 set_bit(Replacement
, &rdev
->flags
);
1698 rdev
->raid_disk
= mirror
;
1700 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1701 rdev
->data_offset
<< 9);
1703 rcu_assign_pointer(p
->replacement
, rdev
);
1707 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1708 rdev
->data_offset
<< 9);
1710 p
->head_position
= 0;
1711 p
->recovery_disabled
= mddev
->recovery_disabled
- 1;
1712 rdev
->raid_disk
= mirror
;
1714 if (rdev
->saved_raid_disk
!= mirror
)
1716 rcu_assign_pointer(p
->rdev
, rdev
);
1719 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1720 /* Some requests might not have seen this new
1721 * merge_bvec_fn. We must wait for them to complete
1722 * before merging the device fully.
1723 * First we make sure any code which has tested
1724 * our function has submitted the request, then
1725 * we wait for all outstanding requests to complete.
1727 synchronize_sched();
1728 raise_barrier(conf
, 0);
1729 lower_barrier(conf
);
1730 clear_bit(Unmerged
, &rdev
->flags
);
1732 md_integrity_add_rdev(rdev
, mddev
);
1733 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1734 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1740 static int raid10_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1742 struct r10conf
*conf
= mddev
->private;
1744 int number
= rdev
->raid_disk
;
1745 struct md_rdev
**rdevp
;
1746 struct raid10_info
*p
= conf
->mirrors
+ number
;
1749 if (rdev
== p
->rdev
)
1751 else if (rdev
== p
->replacement
)
1752 rdevp
= &p
->replacement
;
1756 if (test_bit(In_sync
, &rdev
->flags
) ||
1757 atomic_read(&rdev
->nr_pending
)) {
1761 /* Only remove faulty devices if recovery
1764 if (!test_bit(Faulty
, &rdev
->flags
) &&
1765 mddev
->recovery_disabled
!= p
->recovery_disabled
&&
1766 (!p
->replacement
|| p
->replacement
== rdev
) &&
1767 number
< conf
->geo
.raid_disks
&&
1774 if (atomic_read(&rdev
->nr_pending
)) {
1775 /* lost the race, try later */
1779 } else if (p
->replacement
) {
1780 /* We must have just cleared 'rdev' */
1781 p
->rdev
= p
->replacement
;
1782 clear_bit(Replacement
, &p
->replacement
->flags
);
1783 smp_mb(); /* Make sure other CPUs may see both as identical
1784 * but will never see neither -- if they are careful.
1786 p
->replacement
= NULL
;
1787 clear_bit(WantReplacement
, &rdev
->flags
);
1789 /* We might have just remove the Replacement as faulty
1790 * Clear the flag just in case
1792 clear_bit(WantReplacement
, &rdev
->flags
);
1794 err
= md_integrity_register(mddev
);
1803 static void end_sync_read(struct bio
*bio
, int error
)
1805 struct r10bio
*r10_bio
= bio
->bi_private
;
1806 struct r10conf
*conf
= r10_bio
->mddev
->private;
1809 if (bio
== r10_bio
->master_bio
) {
1810 /* this is a reshape read */
1811 d
= r10_bio
->read_slot
; /* really the read dev */
1813 d
= find_bio_disk(conf
, r10_bio
, bio
, NULL
, NULL
);
1815 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1816 set_bit(R10BIO_Uptodate
, &r10_bio
->state
);
1818 /* The write handler will notice the lack of
1819 * R10BIO_Uptodate and record any errors etc
1821 atomic_add(r10_bio
->sectors
,
1822 &conf
->mirrors
[d
].rdev
->corrected_errors
);
1824 /* for reconstruct, we always reschedule after a read.
1825 * for resync, only after all reads
1827 rdev_dec_pending(conf
->mirrors
[d
].rdev
, conf
->mddev
);
1828 if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
) ||
1829 atomic_dec_and_test(&r10_bio
->remaining
)) {
1830 /* we have read all the blocks,
1831 * do the comparison in process context in raid10d
1833 reschedule_retry(r10_bio
);
1837 static void end_sync_request(struct r10bio
*r10_bio
)
1839 struct mddev
*mddev
= r10_bio
->mddev
;
1841 while (atomic_dec_and_test(&r10_bio
->remaining
)) {
1842 if (r10_bio
->master_bio
== NULL
) {
1843 /* the primary of several recovery bios */
1844 sector_t s
= r10_bio
->sectors
;
1845 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1846 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1847 reschedule_retry(r10_bio
);
1850 md_done_sync(mddev
, s
, 1);
1853 struct r10bio
*r10_bio2
= (struct r10bio
*)r10_bio
->master_bio
;
1854 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
1855 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
1856 reschedule_retry(r10_bio
);
1864 static void end_sync_write(struct bio
*bio
, int error
)
1866 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1867 struct r10bio
*r10_bio
= bio
->bi_private
;
1868 struct mddev
*mddev
= r10_bio
->mddev
;
1869 struct r10conf
*conf
= mddev
->private;
1875 struct md_rdev
*rdev
= NULL
;
1877 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
1879 rdev
= conf
->mirrors
[d
].replacement
;
1881 rdev
= conf
->mirrors
[d
].rdev
;
1885 md_error(mddev
, rdev
);
1887 set_bit(WriteErrorSeen
, &rdev
->flags
);
1888 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1889 set_bit(MD_RECOVERY_NEEDED
,
1890 &rdev
->mddev
->recovery
);
1891 set_bit(R10BIO_WriteError
, &r10_bio
->state
);
1893 } else if (is_badblock(rdev
,
1894 r10_bio
->devs
[slot
].addr
,
1896 &first_bad
, &bad_sectors
))
1897 set_bit(R10BIO_MadeGood
, &r10_bio
->state
);
1899 rdev_dec_pending(rdev
, mddev
);
1901 end_sync_request(r10_bio
);
1905 * Note: sync and recover and handled very differently for raid10
1906 * This code is for resync.
1907 * For resync, we read through virtual addresses and read all blocks.
1908 * If there is any error, we schedule a write. The lowest numbered
1909 * drive is authoritative.
1910 * However requests come for physical address, so we need to map.
1911 * For every physical address there are raid_disks/copies virtual addresses,
1912 * which is always are least one, but is not necessarly an integer.
1913 * This means that a physical address can span multiple chunks, so we may
1914 * have to submit multiple io requests for a single sync request.
1917 * We check if all blocks are in-sync and only write to blocks that
1920 static void sync_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
1922 struct r10conf
*conf
= mddev
->private;
1924 struct bio
*tbio
, *fbio
;
1927 atomic_set(&r10_bio
->remaining
, 1);
1929 /* find the first device with a block */
1930 for (i
=0; i
<conf
->copies
; i
++)
1931 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
))
1934 if (i
== conf
->copies
)
1938 fbio
= r10_bio
->devs
[i
].bio
;
1940 vcnt
= (r10_bio
->sectors
+ (PAGE_SIZE
>> 9) - 1) >> (PAGE_SHIFT
- 9);
1941 /* now find blocks with errors */
1942 for (i
=0 ; i
< conf
->copies
; i
++) {
1945 tbio
= r10_bio
->devs
[i
].bio
;
1947 if (tbio
->bi_end_io
!= end_sync_read
)
1951 if (test_bit(BIO_UPTODATE
, &r10_bio
->devs
[i
].bio
->bi_flags
)) {
1952 /* We know that the bi_io_vec layout is the same for
1953 * both 'first' and 'i', so we just compare them.
1954 * All vec entries are PAGE_SIZE;
1956 for (j
= 0; j
< vcnt
; j
++)
1957 if (memcmp(page_address(fbio
->bi_io_vec
[j
].bv_page
),
1958 page_address(tbio
->bi_io_vec
[j
].bv_page
),
1959 fbio
->bi_io_vec
[j
].bv_len
))
1963 mddev
->resync_mismatches
+= r10_bio
->sectors
;
1964 if (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
))
1965 /* Don't fix anything. */
1968 /* Ok, we need to write this bio, either to correct an
1969 * inconsistency or to correct an unreadable block.
1970 * First we need to fixup bv_offset, bv_len and
1971 * bi_vecs, as the read request might have corrupted these
1973 tbio
->bi_vcnt
= vcnt
;
1974 tbio
->bi_size
= r10_bio
->sectors
<< 9;
1976 tbio
->bi_phys_segments
= 0;
1977 tbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1978 tbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1979 tbio
->bi_next
= NULL
;
1980 tbio
->bi_rw
= WRITE
;
1981 tbio
->bi_private
= r10_bio
;
1982 tbio
->bi_sector
= r10_bio
->devs
[i
].addr
;
1984 for (j
=0; j
< vcnt
; j
++) {
1985 tbio
->bi_io_vec
[j
].bv_offset
= 0;
1986 tbio
->bi_io_vec
[j
].bv_len
= PAGE_SIZE
;
1988 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
1989 page_address(fbio
->bi_io_vec
[j
].bv_page
),
1992 tbio
->bi_end_io
= end_sync_write
;
1994 d
= r10_bio
->devs
[i
].devnum
;
1995 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
1996 atomic_inc(&r10_bio
->remaining
);
1997 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, tbio
->bi_size
>> 9);
1999 tbio
->bi_sector
+= conf
->mirrors
[d
].rdev
->data_offset
;
2000 tbio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
2001 generic_make_request(tbio
);
2004 /* Now write out to any replacement devices
2007 for (i
= 0; i
< conf
->copies
; i
++) {
2010 tbio
= r10_bio
->devs
[i
].repl_bio
;
2011 if (!tbio
|| !tbio
->bi_end_io
)
2013 if (r10_bio
->devs
[i
].bio
->bi_end_io
!= end_sync_write
2014 && r10_bio
->devs
[i
].bio
!= fbio
)
2015 for (j
= 0; j
< vcnt
; j
++)
2016 memcpy(page_address(tbio
->bi_io_vec
[j
].bv_page
),
2017 page_address(fbio
->bi_io_vec
[j
].bv_page
),
2019 d
= r10_bio
->devs
[i
].devnum
;
2020 atomic_inc(&r10_bio
->remaining
);
2021 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2022 tbio
->bi_size
>> 9);
2023 generic_make_request(tbio
);
2027 if (atomic_dec_and_test(&r10_bio
->remaining
)) {
2028 md_done_sync(mddev
, r10_bio
->sectors
, 1);
2034 * Now for the recovery code.
2035 * Recovery happens across physical sectors.
2036 * We recover all non-is_sync drives by finding the virtual address of
2037 * each, and then choose a working drive that also has that virt address.
2038 * There is a separate r10_bio for each non-in_sync drive.
2039 * Only the first two slots are in use. The first for reading,
2040 * The second for writing.
2043 static void fix_recovery_read_error(struct r10bio
*r10_bio
)
2045 /* We got a read error during recovery.
2046 * We repeat the read in smaller page-sized sections.
2047 * If a read succeeds, write it to the new device or record
2048 * a bad block if we cannot.
2049 * If a read fails, record a bad block on both old and
2052 struct mddev
*mddev
= r10_bio
->mddev
;
2053 struct r10conf
*conf
= mddev
->private;
2054 struct bio
*bio
= r10_bio
->devs
[0].bio
;
2056 int sectors
= r10_bio
->sectors
;
2058 int dr
= r10_bio
->devs
[0].devnum
;
2059 int dw
= r10_bio
->devs
[1].devnum
;
2063 struct md_rdev
*rdev
;
2067 if (s
> (PAGE_SIZE
>>9))
2070 rdev
= conf
->mirrors
[dr
].rdev
;
2071 addr
= r10_bio
->devs
[0].addr
+ sect
,
2072 ok
= sync_page_io(rdev
,
2075 bio
->bi_io_vec
[idx
].bv_page
,
2078 rdev
= conf
->mirrors
[dw
].rdev
;
2079 addr
= r10_bio
->devs
[1].addr
+ sect
;
2080 ok
= sync_page_io(rdev
,
2083 bio
->bi_io_vec
[idx
].bv_page
,
2086 set_bit(WriteErrorSeen
, &rdev
->flags
);
2087 if (!test_and_set_bit(WantReplacement
,
2089 set_bit(MD_RECOVERY_NEEDED
,
2090 &rdev
->mddev
->recovery
);
2094 /* We don't worry if we cannot set a bad block -
2095 * it really is bad so there is no loss in not
2098 rdev_set_badblocks(rdev
, addr
, s
, 0);
2100 if (rdev
!= conf
->mirrors
[dw
].rdev
) {
2101 /* need bad block on destination too */
2102 struct md_rdev
*rdev2
= conf
->mirrors
[dw
].rdev
;
2103 addr
= r10_bio
->devs
[1].addr
+ sect
;
2104 ok
= rdev_set_badblocks(rdev2
, addr
, s
, 0);
2106 /* just abort the recovery */
2108 "md/raid10:%s: recovery aborted"
2109 " due to read error\n",
2112 conf
->mirrors
[dw
].recovery_disabled
2113 = mddev
->recovery_disabled
;
2114 set_bit(MD_RECOVERY_INTR
,
2127 static void recovery_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2129 struct r10conf
*conf
= mddev
->private;
2131 struct bio
*wbio
, *wbio2
;
2133 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
)) {
2134 fix_recovery_read_error(r10_bio
);
2135 end_sync_request(r10_bio
);
2140 * share the pages with the first bio
2141 * and submit the write request
2143 d
= r10_bio
->devs
[1].devnum
;
2144 wbio
= r10_bio
->devs
[1].bio
;
2145 wbio2
= r10_bio
->devs
[1].repl_bio
;
2146 if (wbio
->bi_end_io
) {
2147 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
2148 md_sync_acct(conf
->mirrors
[d
].rdev
->bdev
, wbio
->bi_size
>> 9);
2149 generic_make_request(wbio
);
2151 if (wbio2
&& wbio2
->bi_end_io
) {
2152 atomic_inc(&conf
->mirrors
[d
].replacement
->nr_pending
);
2153 md_sync_acct(conf
->mirrors
[d
].replacement
->bdev
,
2154 wbio2
->bi_size
>> 9);
2155 generic_make_request(wbio2
);
2161 * Used by fix_read_error() to decay the per rdev read_errors.
2162 * We halve the read error count for every hour that has elapsed
2163 * since the last recorded read error.
2166 static void check_decay_read_errors(struct mddev
*mddev
, struct md_rdev
*rdev
)
2168 struct timespec cur_time_mon
;
2169 unsigned long hours_since_last
;
2170 unsigned int read_errors
= atomic_read(&rdev
->read_errors
);
2172 ktime_get_ts(&cur_time_mon
);
2174 if (rdev
->last_read_error
.tv_sec
== 0 &&
2175 rdev
->last_read_error
.tv_nsec
== 0) {
2176 /* first time we've seen a read error */
2177 rdev
->last_read_error
= cur_time_mon
;
2181 hours_since_last
= (cur_time_mon
.tv_sec
-
2182 rdev
->last_read_error
.tv_sec
) / 3600;
2184 rdev
->last_read_error
= cur_time_mon
;
2187 * if hours_since_last is > the number of bits in read_errors
2188 * just set read errors to 0. We do this to avoid
2189 * overflowing the shift of read_errors by hours_since_last.
2191 if (hours_since_last
>= 8 * sizeof(read_errors
))
2192 atomic_set(&rdev
->read_errors
, 0);
2194 atomic_set(&rdev
->read_errors
, read_errors
>> hours_since_last
);
2197 static int r10_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
2198 int sectors
, struct page
*page
, int rw
)
2203 if (is_badblock(rdev
, sector
, sectors
, &first_bad
, &bad_sectors
)
2204 && (rw
== READ
|| test_bit(WriteErrorSeen
, &rdev
->flags
)))
2206 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
2210 set_bit(WriteErrorSeen
, &rdev
->flags
);
2211 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2212 set_bit(MD_RECOVERY_NEEDED
,
2213 &rdev
->mddev
->recovery
);
2215 /* need to record an error - either for the block or the device */
2216 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
2217 md_error(rdev
->mddev
, rdev
);
2222 * This is a kernel thread which:
2224 * 1. Retries failed read operations on working mirrors.
2225 * 2. Updates the raid superblock when problems encounter.
2226 * 3. Performs writes following reads for array synchronising.
2229 static void fix_read_error(struct r10conf
*conf
, struct mddev
*mddev
, struct r10bio
*r10_bio
)
2231 int sect
= 0; /* Offset from r10_bio->sector */
2232 int sectors
= r10_bio
->sectors
;
2233 struct md_rdev
*rdev
;
2234 int max_read_errors
= atomic_read(&mddev
->max_corr_read_errors
);
2235 int d
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2237 /* still own a reference to this rdev, so it cannot
2238 * have been cleared recently.
2240 rdev
= conf
->mirrors
[d
].rdev
;
2242 if (test_bit(Faulty
, &rdev
->flags
))
2243 /* drive has already been failed, just ignore any
2244 more fix_read_error() attempts */
2247 check_decay_read_errors(mddev
, rdev
);
2248 atomic_inc(&rdev
->read_errors
);
2249 if (atomic_read(&rdev
->read_errors
) > max_read_errors
) {
2250 char b
[BDEVNAME_SIZE
];
2251 bdevname(rdev
->bdev
, b
);
2254 "md/raid10:%s: %s: Raid device exceeded "
2255 "read_error threshold [cur %d:max %d]\n",
2257 atomic_read(&rdev
->read_errors
), max_read_errors
);
2259 "md/raid10:%s: %s: Failing raid device\n",
2261 md_error(mddev
, conf
->mirrors
[d
].rdev
);
2262 r10_bio
->devs
[r10_bio
->read_slot
].bio
= IO_BLOCKED
;
2268 int sl
= r10_bio
->read_slot
;
2272 if (s
> (PAGE_SIZE
>>9))
2280 d
= r10_bio
->devs
[sl
].devnum
;
2281 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2283 !test_bit(Unmerged
, &rdev
->flags
) &&
2284 test_bit(In_sync
, &rdev
->flags
) &&
2285 is_badblock(rdev
, r10_bio
->devs
[sl
].addr
+ sect
, s
,
2286 &first_bad
, &bad_sectors
) == 0) {
2287 atomic_inc(&rdev
->nr_pending
);
2289 success
= sync_page_io(rdev
,
2290 r10_bio
->devs
[sl
].addr
+
2293 conf
->tmppage
, READ
, false);
2294 rdev_dec_pending(rdev
, mddev
);
2300 if (sl
== conf
->copies
)
2302 } while (!success
&& sl
!= r10_bio
->read_slot
);
2306 /* Cannot read from anywhere, just mark the block
2307 * as bad on the first device to discourage future
2310 int dn
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
2311 rdev
= conf
->mirrors
[dn
].rdev
;
2313 if (!rdev_set_badblocks(
2315 r10_bio
->devs
[r10_bio
->read_slot
].addr
2318 md_error(mddev
, rdev
);
2319 r10_bio
->devs
[r10_bio
->read_slot
].bio
2326 /* write it back and re-read */
2328 while (sl
!= r10_bio
->read_slot
) {
2329 char b
[BDEVNAME_SIZE
];
2334 d
= r10_bio
->devs
[sl
].devnum
;
2335 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2337 test_bit(Unmerged
, &rdev
->flags
) ||
2338 !test_bit(In_sync
, &rdev
->flags
))
2341 atomic_inc(&rdev
->nr_pending
);
2343 if (r10_sync_page_io(rdev
,
2344 r10_bio
->devs
[sl
].addr
+
2346 s
, conf
->tmppage
, WRITE
)
2348 /* Well, this device is dead */
2350 "md/raid10:%s: read correction "
2352 " (%d sectors at %llu on %s)\n",
2354 (unsigned long long)(
2356 choose_data_offset(r10_bio
,
2358 bdevname(rdev
->bdev
, b
));
2359 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2362 bdevname(rdev
->bdev
, b
));
2364 rdev_dec_pending(rdev
, mddev
);
2368 while (sl
!= r10_bio
->read_slot
) {
2369 char b
[BDEVNAME_SIZE
];
2374 d
= r10_bio
->devs
[sl
].devnum
;
2375 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2377 !test_bit(In_sync
, &rdev
->flags
))
2380 atomic_inc(&rdev
->nr_pending
);
2382 switch (r10_sync_page_io(rdev
,
2383 r10_bio
->devs
[sl
].addr
+
2388 /* Well, this device is dead */
2390 "md/raid10:%s: unable to read back "
2392 " (%d sectors at %llu on %s)\n",
2394 (unsigned long long)(
2396 choose_data_offset(r10_bio
, rdev
)),
2397 bdevname(rdev
->bdev
, b
));
2398 printk(KERN_NOTICE
"md/raid10:%s: %s: failing "
2401 bdevname(rdev
->bdev
, b
));
2405 "md/raid10:%s: read error corrected"
2406 " (%d sectors at %llu on %s)\n",
2408 (unsigned long long)(
2410 choose_data_offset(r10_bio
, rdev
)),
2411 bdevname(rdev
->bdev
, b
));
2412 atomic_add(s
, &rdev
->corrected_errors
);
2415 rdev_dec_pending(rdev
, mddev
);
2425 static void bi_complete(struct bio
*bio
, int error
)
2427 complete((struct completion
*)bio
->bi_private
);
2430 static int submit_bio_wait(int rw
, struct bio
*bio
)
2432 struct completion event
;
2435 init_completion(&event
);
2436 bio
->bi_private
= &event
;
2437 bio
->bi_end_io
= bi_complete
;
2438 submit_bio(rw
, bio
);
2439 wait_for_completion(&event
);
2441 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
2444 static int narrow_write_error(struct r10bio
*r10_bio
, int i
)
2446 struct bio
*bio
= r10_bio
->master_bio
;
2447 struct mddev
*mddev
= r10_bio
->mddev
;
2448 struct r10conf
*conf
= mddev
->private;
2449 struct md_rdev
*rdev
= conf
->mirrors
[r10_bio
->devs
[i
].devnum
].rdev
;
2450 /* bio has the data to be written to slot 'i' where
2451 * we just recently had a write error.
2452 * We repeatedly clone the bio and trim down to one block,
2453 * then try the write. Where the write fails we record
2455 * It is conceivable that the bio doesn't exactly align with
2456 * blocks. We must handle this.
2458 * We currently own a reference to the rdev.
2464 int sect_to_write
= r10_bio
->sectors
;
2467 if (rdev
->badblocks
.shift
< 0)
2470 block_sectors
= 1 << rdev
->badblocks
.shift
;
2471 sector
= r10_bio
->sector
;
2472 sectors
= ((r10_bio
->sector
+ block_sectors
)
2473 & ~(sector_t
)(block_sectors
- 1))
2476 while (sect_to_write
) {
2478 if (sectors
> sect_to_write
)
2479 sectors
= sect_to_write
;
2480 /* Write at 'sector' for 'sectors' */
2481 wbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
2482 md_trim_bio(wbio
, sector
- bio
->bi_sector
, sectors
);
2483 wbio
->bi_sector
= (r10_bio
->devs
[i
].addr
+
2484 choose_data_offset(r10_bio
, rdev
) +
2485 (sector
- r10_bio
->sector
));
2486 wbio
->bi_bdev
= rdev
->bdev
;
2487 if (submit_bio_wait(WRITE
, wbio
) == 0)
2489 ok
= rdev_set_badblocks(rdev
, sector
,
2494 sect_to_write
-= sectors
;
2496 sectors
= block_sectors
;
2501 static void handle_read_error(struct mddev
*mddev
, struct r10bio
*r10_bio
)
2503 int slot
= r10_bio
->read_slot
;
2505 struct r10conf
*conf
= mddev
->private;
2506 struct md_rdev
*rdev
= r10_bio
->devs
[slot
].rdev
;
2507 char b
[BDEVNAME_SIZE
];
2508 unsigned long do_sync
;
2511 /* we got a read error. Maybe the drive is bad. Maybe just
2512 * the block and we can fix it.
2513 * We freeze all other IO, and try reading the block from
2514 * other devices. When we find one, we re-write
2515 * and check it that fixes the read error.
2516 * This is all done synchronously while the array is
2519 bio
= r10_bio
->devs
[slot
].bio
;
2520 bdevname(bio
->bi_bdev
, b
);
2522 r10_bio
->devs
[slot
].bio
= NULL
;
2524 if (mddev
->ro
== 0) {
2526 fix_read_error(conf
, mddev
, r10_bio
);
2527 unfreeze_array(conf
);
2529 r10_bio
->devs
[slot
].bio
= IO_BLOCKED
;
2531 rdev_dec_pending(rdev
, mddev
);
2534 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
2536 printk(KERN_ALERT
"md/raid10:%s: %s: unrecoverable I/O"
2537 " read error for block %llu\n",
2539 (unsigned long long)r10_bio
->sector
);
2540 raid_end_bio_io(r10_bio
);
2544 do_sync
= (r10_bio
->master_bio
->bi_rw
& REQ_SYNC
);
2545 slot
= r10_bio
->read_slot
;
2548 "md/raid10:%s: %s: redirecting "
2549 "sector %llu to another mirror\n",
2551 bdevname(rdev
->bdev
, b
),
2552 (unsigned long long)r10_bio
->sector
);
2553 bio
= bio_clone_mddev(r10_bio
->master_bio
,
2556 r10_bio
->sector
- bio
->bi_sector
,
2558 r10_bio
->devs
[slot
].bio
= bio
;
2559 r10_bio
->devs
[slot
].rdev
= rdev
;
2560 bio
->bi_sector
= r10_bio
->devs
[slot
].addr
2561 + choose_data_offset(r10_bio
, rdev
);
2562 bio
->bi_bdev
= rdev
->bdev
;
2563 bio
->bi_rw
= READ
| do_sync
;
2564 bio
->bi_private
= r10_bio
;
2565 bio
->bi_end_io
= raid10_end_read_request
;
2566 if (max_sectors
< r10_bio
->sectors
) {
2567 /* Drat - have to split this up more */
2568 struct bio
*mbio
= r10_bio
->master_bio
;
2569 int sectors_handled
=
2570 r10_bio
->sector
+ max_sectors
2572 r10_bio
->sectors
= max_sectors
;
2573 spin_lock_irq(&conf
->device_lock
);
2574 if (mbio
->bi_phys_segments
== 0)
2575 mbio
->bi_phys_segments
= 2;
2577 mbio
->bi_phys_segments
++;
2578 spin_unlock_irq(&conf
->device_lock
);
2579 generic_make_request(bio
);
2581 r10_bio
= mempool_alloc(conf
->r10bio_pool
,
2583 r10_bio
->master_bio
= mbio
;
2584 r10_bio
->sectors
= (mbio
->bi_size
>> 9)
2587 set_bit(R10BIO_ReadError
,
2589 r10_bio
->mddev
= mddev
;
2590 r10_bio
->sector
= mbio
->bi_sector
2595 generic_make_request(bio
);
2598 static void handle_write_completed(struct r10conf
*conf
, struct r10bio
*r10_bio
)
2600 /* Some sort of write request has finished and it
2601 * succeeded in writing where we thought there was a
2602 * bad block. So forget the bad block.
2603 * Or possibly if failed and we need to record
2607 struct md_rdev
*rdev
;
2609 if (test_bit(R10BIO_IsSync
, &r10_bio
->state
) ||
2610 test_bit(R10BIO_IsRecover
, &r10_bio
->state
)) {
2611 for (m
= 0; m
< conf
->copies
; m
++) {
2612 int dev
= r10_bio
->devs
[m
].devnum
;
2613 rdev
= conf
->mirrors
[dev
].rdev
;
2614 if (r10_bio
->devs
[m
].bio
== NULL
)
2616 if (test_bit(BIO_UPTODATE
,
2617 &r10_bio
->devs
[m
].bio
->bi_flags
)) {
2618 rdev_clear_badblocks(
2620 r10_bio
->devs
[m
].addr
,
2621 r10_bio
->sectors
, 0);
2623 if (!rdev_set_badblocks(
2625 r10_bio
->devs
[m
].addr
,
2626 r10_bio
->sectors
, 0))
2627 md_error(conf
->mddev
, rdev
);
2629 rdev
= conf
->mirrors
[dev
].replacement
;
2630 if (r10_bio
->devs
[m
].repl_bio
== NULL
)
2632 if (test_bit(BIO_UPTODATE
,
2633 &r10_bio
->devs
[m
].repl_bio
->bi_flags
)) {
2634 rdev_clear_badblocks(
2636 r10_bio
->devs
[m
].addr
,
2637 r10_bio
->sectors
, 0);
2639 if (!rdev_set_badblocks(
2641 r10_bio
->devs
[m
].addr
,
2642 r10_bio
->sectors
, 0))
2643 md_error(conf
->mddev
, rdev
);
2648 for (m
= 0; m
< conf
->copies
; m
++) {
2649 int dev
= r10_bio
->devs
[m
].devnum
;
2650 struct bio
*bio
= r10_bio
->devs
[m
].bio
;
2651 rdev
= conf
->mirrors
[dev
].rdev
;
2652 if (bio
== IO_MADE_GOOD
) {
2653 rdev_clear_badblocks(
2655 r10_bio
->devs
[m
].addr
,
2656 r10_bio
->sectors
, 0);
2657 rdev_dec_pending(rdev
, conf
->mddev
);
2658 } else if (bio
!= NULL
&&
2659 !test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2660 if (!narrow_write_error(r10_bio
, m
)) {
2661 md_error(conf
->mddev
, rdev
);
2662 set_bit(R10BIO_Degraded
,
2665 rdev_dec_pending(rdev
, conf
->mddev
);
2667 bio
= r10_bio
->devs
[m
].repl_bio
;
2668 rdev
= conf
->mirrors
[dev
].replacement
;
2669 if (rdev
&& bio
== IO_MADE_GOOD
) {
2670 rdev_clear_badblocks(
2672 r10_bio
->devs
[m
].addr
,
2673 r10_bio
->sectors
, 0);
2674 rdev_dec_pending(rdev
, conf
->mddev
);
2677 if (test_bit(R10BIO_WriteError
,
2679 close_write(r10_bio
);
2680 raid_end_bio_io(r10_bio
);
2684 static void raid10d(struct mddev
*mddev
)
2686 struct r10bio
*r10_bio
;
2687 unsigned long flags
;
2688 struct r10conf
*conf
= mddev
->private;
2689 struct list_head
*head
= &conf
->retry_list
;
2690 struct blk_plug plug
;
2692 md_check_recovery(mddev
);
2694 blk_start_plug(&plug
);
2697 flush_pending_writes(conf
);
2699 spin_lock_irqsave(&conf
->device_lock
, flags
);
2700 if (list_empty(head
)) {
2701 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2704 r10_bio
= list_entry(head
->prev
, struct r10bio
, retry_list
);
2705 list_del(head
->prev
);
2707 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2709 mddev
= r10_bio
->mddev
;
2710 conf
= mddev
->private;
2711 if (test_bit(R10BIO_MadeGood
, &r10_bio
->state
) ||
2712 test_bit(R10BIO_WriteError
, &r10_bio
->state
))
2713 handle_write_completed(conf
, r10_bio
);
2714 else if (test_bit(R10BIO_IsReshape
, &r10_bio
->state
))
2715 reshape_request_write(mddev
, r10_bio
);
2716 else if (test_bit(R10BIO_IsSync
, &r10_bio
->state
))
2717 sync_request_write(mddev
, r10_bio
);
2718 else if (test_bit(R10BIO_IsRecover
, &r10_bio
->state
))
2719 recovery_request_write(mddev
, r10_bio
);
2720 else if (test_bit(R10BIO_ReadError
, &r10_bio
->state
))
2721 handle_read_error(mddev
, r10_bio
);
2723 /* just a partial read to be scheduled from a
2726 int slot
= r10_bio
->read_slot
;
2727 generic_make_request(r10_bio
->devs
[slot
].bio
);
2731 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2732 md_check_recovery(mddev
);
2734 blk_finish_plug(&plug
);
2738 static int init_resync(struct r10conf
*conf
)
2743 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2744 BUG_ON(conf
->r10buf_pool
);
2745 conf
->have_replacement
= 0;
2746 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2747 if (conf
->mirrors
[i
].replacement
)
2748 conf
->have_replacement
= 1;
2749 conf
->r10buf_pool
= mempool_create(buffs
, r10buf_pool_alloc
, r10buf_pool_free
, conf
);
2750 if (!conf
->r10buf_pool
)
2752 conf
->next_resync
= 0;
2757 * perform a "sync" on one "block"
2759 * We need to make sure that no normal I/O request - particularly write
2760 * requests - conflict with active sync requests.
2762 * This is achieved by tracking pending requests and a 'barrier' concept
2763 * that can be installed to exclude normal IO requests.
2765 * Resync and recovery are handled very differently.
2766 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2768 * For resync, we iterate over virtual addresses, read all copies,
2769 * and update if there are differences. If only one copy is live,
2771 * For recovery, we iterate over physical addresses, read a good
2772 * value for each non-in_sync drive, and over-write.
2774 * So, for recovery we may have several outstanding complex requests for a
2775 * given address, one for each out-of-sync device. We model this by allocating
2776 * a number of r10_bio structures, one for each out-of-sync device.
2777 * As we setup these structures, we collect all bio's together into a list
2778 * which we then process collectively to add pages, and then process again
2779 * to pass to generic_make_request.
2781 * The r10_bio structures are linked using a borrowed master_bio pointer.
2782 * This link is counted in ->remaining. When the r10_bio that points to NULL
2783 * has its remaining count decremented to 0, the whole complex operation
2788 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2789 int *skipped
, int go_faster
)
2791 struct r10conf
*conf
= mddev
->private;
2792 struct r10bio
*r10_bio
;
2793 struct bio
*biolist
= NULL
, *bio
;
2794 sector_t max_sector
, nr_sectors
;
2797 sector_t sync_blocks
;
2798 sector_t sectors_skipped
= 0;
2799 int chunks_skipped
= 0;
2800 sector_t chunk_mask
= conf
->geo
.chunk_mask
;
2802 if (!conf
->r10buf_pool
)
2803 if (init_resync(conf
))
2807 max_sector
= mddev
->dev_sectors
;
2808 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) ||
2809 test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2810 max_sector
= mddev
->resync_max_sectors
;
2811 if (sector_nr
>= max_sector
) {
2812 /* If we aborted, we need to abort the
2813 * sync on the 'current' bitmap chucks (there can
2814 * be several when recovering multiple devices).
2815 * as we may have started syncing it but not finished.
2816 * We can find the current address in
2817 * mddev->curr_resync, but for recovery,
2818 * we need to convert that to several
2819 * virtual addresses.
2821 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
2826 if (mddev
->curr_resync
< max_sector
) { /* aborted */
2827 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))
2828 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2830 else for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
2832 raid10_find_virt(conf
, mddev
->curr_resync
, i
);
2833 bitmap_end_sync(mddev
->bitmap
, sect
,
2837 /* completed sync */
2838 if ((!mddev
->bitmap
|| conf
->fullsync
)
2839 && conf
->have_replacement
2840 && test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2841 /* Completed a full sync so the replacements
2842 * are now fully recovered.
2844 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++)
2845 if (conf
->mirrors
[i
].replacement
)
2846 conf
->mirrors
[i
].replacement
2852 bitmap_close_sync(mddev
->bitmap
);
2855 return sectors_skipped
;
2858 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
2859 return reshape_request(mddev
, sector_nr
, skipped
);
2861 if (chunks_skipped
>= conf
->geo
.raid_disks
) {
2862 /* if there has been nothing to do on any drive,
2863 * then there is nothing to do at all..
2866 return (max_sector
- sector_nr
) + sectors_skipped
;
2869 if (max_sector
> mddev
->resync_max
)
2870 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2872 /* make sure whole request will fit in a chunk - if chunks
2875 if (conf
->geo
.near_copies
< conf
->geo
.raid_disks
&&
2876 max_sector
> (sector_nr
| chunk_mask
))
2877 max_sector
= (sector_nr
| chunk_mask
) + 1;
2879 * If there is non-resync activity waiting for us then
2880 * put in a delay to throttle resync.
2882 if (!go_faster
&& conf
->nr_waiting
)
2883 msleep_interruptible(1000);
2885 /* Again, very different code for resync and recovery.
2886 * Both must result in an r10bio with a list of bios that
2887 * have bi_end_io, bi_sector, bi_bdev set,
2888 * and bi_private set to the r10bio.
2889 * For recovery, we may actually create several r10bios
2890 * with 2 bios in each, that correspond to the bios in the main one.
2891 * In this case, the subordinate r10bios link back through a
2892 * borrowed master_bio pointer, and the counter in the master
2893 * includes a ref from each subordinate.
2895 /* First, we decide what to do and set ->bi_end_io
2896 * To end_sync_read if we want to read, and
2897 * end_sync_write if we will want to write.
2900 max_sync
= RESYNC_PAGES
<< (PAGE_SHIFT
-9);
2901 if (!test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
2902 /* recovery... the complicated one */
2906 for (i
= 0 ; i
< conf
->geo
.raid_disks
; i
++) {
2912 struct raid10_info
*mirror
= &conf
->mirrors
[i
];
2914 if ((mirror
->rdev
== NULL
||
2915 test_bit(In_sync
, &mirror
->rdev
->flags
))
2917 (mirror
->replacement
== NULL
||
2919 &mirror
->replacement
->flags
)))
2923 /* want to reconstruct this device */
2925 sect
= raid10_find_virt(conf
, sector_nr
, i
);
2926 if (sect
>= mddev
->resync_max_sectors
) {
2927 /* last stripe is not complete - don't
2928 * try to recover this sector.
2932 /* Unless we are doing a full sync, or a replacement
2933 * we only need to recover the block if it is set in
2936 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2938 if (sync_blocks
< max_sync
)
2939 max_sync
= sync_blocks
;
2941 mirror
->replacement
== NULL
&&
2943 /* yep, skip the sync_blocks here, but don't assume
2944 * that there will never be anything to do here
2946 chunks_skipped
= -1;
2950 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
2951 raise_barrier(conf
, rb2
!= NULL
);
2952 atomic_set(&r10_bio
->remaining
, 0);
2954 r10_bio
->master_bio
= (struct bio
*)rb2
;
2956 atomic_inc(&rb2
->remaining
);
2957 r10_bio
->mddev
= mddev
;
2958 set_bit(R10BIO_IsRecover
, &r10_bio
->state
);
2959 r10_bio
->sector
= sect
;
2961 raid10_find_phys(conf
, r10_bio
);
2963 /* Need to check if the array will still be
2966 for (j
= 0; j
< conf
->geo
.raid_disks
; j
++)
2967 if (conf
->mirrors
[j
].rdev
== NULL
||
2968 test_bit(Faulty
, &conf
->mirrors
[j
].rdev
->flags
)) {
2973 must_sync
= bitmap_start_sync(mddev
->bitmap
, sect
,
2974 &sync_blocks
, still_degraded
);
2977 for (j
=0; j
<conf
->copies
;j
++) {
2979 int d
= r10_bio
->devs
[j
].devnum
;
2980 sector_t from_addr
, to_addr
;
2981 struct md_rdev
*rdev
;
2982 sector_t sector
, first_bad
;
2984 if (!conf
->mirrors
[d
].rdev
||
2985 !test_bit(In_sync
, &conf
->mirrors
[d
].rdev
->flags
))
2987 /* This is where we read from */
2989 rdev
= conf
->mirrors
[d
].rdev
;
2990 sector
= r10_bio
->devs
[j
].addr
;
2992 if (is_badblock(rdev
, sector
, max_sync
,
2993 &first_bad
, &bad_sectors
)) {
2994 if (first_bad
> sector
)
2995 max_sync
= first_bad
- sector
;
2997 bad_sectors
-= (sector
2999 if (max_sync
> bad_sectors
)
3000 max_sync
= bad_sectors
;
3004 bio
= r10_bio
->devs
[0].bio
;
3005 bio
->bi_next
= biolist
;
3007 bio
->bi_private
= r10_bio
;
3008 bio
->bi_end_io
= end_sync_read
;
3010 from_addr
= r10_bio
->devs
[j
].addr
;
3011 bio
->bi_sector
= from_addr
+ rdev
->data_offset
;
3012 bio
->bi_bdev
= rdev
->bdev
;
3013 atomic_inc(&rdev
->nr_pending
);
3014 /* and we write to 'i' (if not in_sync) */
3016 for (k
=0; k
<conf
->copies
; k
++)
3017 if (r10_bio
->devs
[k
].devnum
== i
)
3019 BUG_ON(k
== conf
->copies
);
3020 to_addr
= r10_bio
->devs
[k
].addr
;
3021 r10_bio
->devs
[0].devnum
= d
;
3022 r10_bio
->devs
[0].addr
= from_addr
;
3023 r10_bio
->devs
[1].devnum
= i
;
3024 r10_bio
->devs
[1].addr
= to_addr
;
3026 rdev
= mirror
->rdev
;
3027 if (!test_bit(In_sync
, &rdev
->flags
)) {
3028 bio
= r10_bio
->devs
[1].bio
;
3029 bio
->bi_next
= biolist
;
3031 bio
->bi_private
= r10_bio
;
3032 bio
->bi_end_io
= end_sync_write
;
3034 bio
->bi_sector
= to_addr
3035 + rdev
->data_offset
;
3036 bio
->bi_bdev
= rdev
->bdev
;
3037 atomic_inc(&r10_bio
->remaining
);
3039 r10_bio
->devs
[1].bio
->bi_end_io
= NULL
;
3041 /* and maybe write to replacement */
3042 bio
= r10_bio
->devs
[1].repl_bio
;
3044 bio
->bi_end_io
= NULL
;
3045 rdev
= mirror
->replacement
;
3046 /* Note: if rdev != NULL, then bio
3047 * cannot be NULL as r10buf_pool_alloc will
3048 * have allocated it.
3049 * So the second test here is pointless.
3050 * But it keeps semantic-checkers happy, and
3051 * this comment keeps human reviewers
3054 if (rdev
== NULL
|| bio
== NULL
||
3055 test_bit(Faulty
, &rdev
->flags
))
3057 bio
->bi_next
= biolist
;
3059 bio
->bi_private
= r10_bio
;
3060 bio
->bi_end_io
= end_sync_write
;
3062 bio
->bi_sector
= to_addr
+ rdev
->data_offset
;
3063 bio
->bi_bdev
= rdev
->bdev
;
3064 atomic_inc(&r10_bio
->remaining
);
3067 if (j
== conf
->copies
) {
3068 /* Cannot recover, so abort the recovery or
3069 * record a bad block */
3072 atomic_dec(&rb2
->remaining
);
3075 /* problem is that there are bad blocks
3076 * on other device(s)
3079 for (k
= 0; k
< conf
->copies
; k
++)
3080 if (r10_bio
->devs
[k
].devnum
== i
)
3082 if (!test_bit(In_sync
,
3083 &mirror
->rdev
->flags
)
3084 && !rdev_set_badblocks(
3086 r10_bio
->devs
[k
].addr
,
3089 if (mirror
->replacement
&&
3090 !rdev_set_badblocks(
3091 mirror
->replacement
,
3092 r10_bio
->devs
[k
].addr
,
3097 if (!test_and_set_bit(MD_RECOVERY_INTR
,
3099 printk(KERN_INFO
"md/raid10:%s: insufficient "
3100 "working devices for recovery.\n",
3102 mirror
->recovery_disabled
3103 = mddev
->recovery_disabled
;
3108 if (biolist
== NULL
) {
3110 struct r10bio
*rb2
= r10_bio
;
3111 r10_bio
= (struct r10bio
*) rb2
->master_bio
;
3112 rb2
->master_bio
= NULL
;
3118 /* resync. Schedule a read for every block at this virt offset */
3121 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3123 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
3124 &sync_blocks
, mddev
->degraded
) &&
3125 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
,
3126 &mddev
->recovery
)) {
3127 /* We can skip this block */
3129 return sync_blocks
+ sectors_skipped
;
3131 if (sync_blocks
< max_sync
)
3132 max_sync
= sync_blocks
;
3133 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
3135 r10_bio
->mddev
= mddev
;
3136 atomic_set(&r10_bio
->remaining
, 0);
3137 raise_barrier(conf
, 0);
3138 conf
->next_resync
= sector_nr
;
3140 r10_bio
->master_bio
= NULL
;
3141 r10_bio
->sector
= sector_nr
;
3142 set_bit(R10BIO_IsSync
, &r10_bio
->state
);
3143 raid10_find_phys(conf
, r10_bio
);
3144 r10_bio
->sectors
= (sector_nr
| chunk_mask
) - sector_nr
+ 1;
3146 for (i
= 0; i
< conf
->copies
; i
++) {
3147 int d
= r10_bio
->devs
[i
].devnum
;
3148 sector_t first_bad
, sector
;
3151 if (r10_bio
->devs
[i
].repl_bio
)
3152 r10_bio
->devs
[i
].repl_bio
->bi_end_io
= NULL
;
3154 bio
= r10_bio
->devs
[i
].bio
;
3155 bio
->bi_end_io
= NULL
;
3156 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3157 if (conf
->mirrors
[d
].rdev
== NULL
||
3158 test_bit(Faulty
, &conf
->mirrors
[d
].rdev
->flags
))
3160 sector
= r10_bio
->devs
[i
].addr
;
3161 if (is_badblock(conf
->mirrors
[d
].rdev
,
3163 &first_bad
, &bad_sectors
)) {
3164 if (first_bad
> sector
)
3165 max_sync
= first_bad
- sector
;
3167 bad_sectors
-= (sector
- first_bad
);
3168 if (max_sync
> bad_sectors
)
3169 max_sync
= max_sync
;
3173 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3174 atomic_inc(&r10_bio
->remaining
);
3175 bio
->bi_next
= biolist
;
3177 bio
->bi_private
= r10_bio
;
3178 bio
->bi_end_io
= end_sync_read
;
3180 bio
->bi_sector
= sector
+
3181 conf
->mirrors
[d
].rdev
->data_offset
;
3182 bio
->bi_bdev
= conf
->mirrors
[d
].rdev
->bdev
;
3185 if (conf
->mirrors
[d
].replacement
== NULL
||
3187 &conf
->mirrors
[d
].replacement
->flags
))
3190 /* Need to set up for writing to the replacement */
3191 bio
= r10_bio
->devs
[i
].repl_bio
;
3192 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
3194 sector
= r10_bio
->devs
[i
].addr
;
3195 atomic_inc(&conf
->mirrors
[d
].rdev
->nr_pending
);
3196 bio
->bi_next
= biolist
;
3198 bio
->bi_private
= r10_bio
;
3199 bio
->bi_end_io
= end_sync_write
;
3201 bio
->bi_sector
= sector
+
3202 conf
->mirrors
[d
].replacement
->data_offset
;
3203 bio
->bi_bdev
= conf
->mirrors
[d
].replacement
->bdev
;
3208 for (i
=0; i
<conf
->copies
; i
++) {
3209 int d
= r10_bio
->devs
[i
].devnum
;
3210 if (r10_bio
->devs
[i
].bio
->bi_end_io
)
3211 rdev_dec_pending(conf
->mirrors
[d
].rdev
,
3213 if (r10_bio
->devs
[i
].repl_bio
&&
3214 r10_bio
->devs
[i
].repl_bio
->bi_end_io
)
3216 conf
->mirrors
[d
].replacement
,
3225 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3227 bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
3229 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
3232 bio
->bi_phys_segments
= 0;
3237 if (sector_nr
+ max_sync
< max_sector
)
3238 max_sector
= sector_nr
+ max_sync
;
3241 int len
= PAGE_SIZE
;
3242 if (sector_nr
+ (len
>>9) > max_sector
)
3243 len
= (max_sector
- sector_nr
) << 9;
3246 for (bio
= biolist
; bio
; bio
=bio
->bi_next
) {
3248 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
3249 if (bio_add_page(bio
, page
, len
, 0))
3253 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
3254 for (bio2
= biolist
;
3255 bio2
&& bio2
!= bio
;
3256 bio2
= bio2
->bi_next
) {
3257 /* remove last page from this bio */
3259 bio2
->bi_size
-= len
;
3260 bio2
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
3264 nr_sectors
+= len
>>9;
3265 sector_nr
+= len
>>9;
3266 } while (biolist
->bi_vcnt
< RESYNC_PAGES
);
3268 r10_bio
->sectors
= nr_sectors
;
3272 biolist
= biolist
->bi_next
;
3274 bio
->bi_next
= NULL
;
3275 r10_bio
= bio
->bi_private
;
3276 r10_bio
->sectors
= nr_sectors
;
3278 if (bio
->bi_end_io
== end_sync_read
) {
3279 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
3280 generic_make_request(bio
);
3284 if (sectors_skipped
)
3285 /* pretend they weren't skipped, it makes
3286 * no important difference in this case
3288 md_done_sync(mddev
, sectors_skipped
, 1);
3290 return sectors_skipped
+ nr_sectors
;
3292 /* There is nowhere to write, so all non-sync
3293 * drives must be failed or in resync, all drives
3294 * have a bad block, so try the next chunk...
3296 if (sector_nr
+ max_sync
< max_sector
)
3297 max_sector
= sector_nr
+ max_sync
;
3299 sectors_skipped
+= (max_sector
- sector_nr
);
3301 sector_nr
= max_sector
;
3306 raid10_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
3309 struct r10conf
*conf
= mddev
->private;
3312 raid_disks
= min(conf
->geo
.raid_disks
,
3313 conf
->prev
.raid_disks
);
3315 sectors
= conf
->dev_sectors
;
3317 size
= sectors
>> conf
->geo
.chunk_shift
;
3318 sector_div(size
, conf
->geo
.far_copies
);
3319 size
= size
* raid_disks
;
3320 sector_div(size
, conf
->geo
.near_copies
);
3322 return size
<< conf
->geo
.chunk_shift
;
3325 static void calc_sectors(struct r10conf
*conf
, sector_t size
)
3327 /* Calculate the number of sectors-per-device that will
3328 * actually be used, and set conf->dev_sectors and
3332 size
= size
>> conf
->geo
.chunk_shift
;
3333 sector_div(size
, conf
->geo
.far_copies
);
3334 size
= size
* conf
->geo
.raid_disks
;
3335 sector_div(size
, conf
->geo
.near_copies
);
3336 /* 'size' is now the number of chunks in the array */
3337 /* calculate "used chunks per device" */
3338 size
= size
* conf
->copies
;
3340 /* We need to round up when dividing by raid_disks to
3341 * get the stride size.
3343 size
= DIV_ROUND_UP_SECTOR_T(size
, conf
->geo
.raid_disks
);
3345 conf
->dev_sectors
= size
<< conf
->geo
.chunk_shift
;
3347 if (conf
->geo
.far_offset
)
3348 conf
->geo
.stride
= 1 << conf
->geo
.chunk_shift
;
3350 sector_div(size
, conf
->geo
.far_copies
);
3351 conf
->geo
.stride
= size
<< conf
->geo
.chunk_shift
;
3355 enum geo_type
{geo_new
, geo_old
, geo_start
};
3356 static int setup_geo(struct geom
*geo
, struct mddev
*mddev
, enum geo_type
new)
3359 int layout
, chunk
, disks
;
3362 layout
= mddev
->layout
;
3363 chunk
= mddev
->chunk_sectors
;
3364 disks
= mddev
->raid_disks
- mddev
->delta_disks
;
3367 layout
= mddev
->new_layout
;
3368 chunk
= mddev
->new_chunk_sectors
;
3369 disks
= mddev
->raid_disks
;
3371 default: /* avoid 'may be unused' warnings */
3372 case geo_start
: /* new when starting reshape - raid_disks not
3374 layout
= mddev
->new_layout
;
3375 chunk
= mddev
->new_chunk_sectors
;
3376 disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3381 if (chunk
< (PAGE_SIZE
>> 9) ||
3382 !is_power_of_2(chunk
))
3385 fc
= (layout
>> 8) & 255;
3386 fo
= layout
& (1<<16);
3387 geo
->raid_disks
= disks
;
3388 geo
->near_copies
= nc
;
3389 geo
->far_copies
= fc
;
3390 geo
->far_offset
= fo
;
3391 geo
->chunk_mask
= chunk
- 1;
3392 geo
->chunk_shift
= ffz(~chunk
);
3396 static struct r10conf
*setup_conf(struct mddev
*mddev
)
3398 struct r10conf
*conf
= NULL
;
3403 copies
= setup_geo(&geo
, mddev
, geo_new
);
3406 printk(KERN_ERR
"md/raid10:%s: chunk size must be "
3407 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3408 mdname(mddev
), PAGE_SIZE
);
3412 if (copies
< 2 || copies
> mddev
->raid_disks
) {
3413 printk(KERN_ERR
"md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3414 mdname(mddev
), mddev
->new_layout
);
3419 conf
= kzalloc(sizeof(struct r10conf
), GFP_KERNEL
);
3423 /* FIXME calc properly */
3424 conf
->mirrors
= kzalloc(sizeof(struct raid10_info
)*(mddev
->raid_disks
+
3425 max(0,mddev
->delta_disks
)),
3430 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3435 conf
->copies
= copies
;
3436 conf
->r10bio_pool
= mempool_create(NR_RAID10_BIOS
, r10bio_pool_alloc
,
3437 r10bio_pool_free
, conf
);
3438 if (!conf
->r10bio_pool
)
3441 calc_sectors(conf
, mddev
->dev_sectors
);
3442 if (mddev
->reshape_position
== MaxSector
) {
3443 conf
->prev
= conf
->geo
;
3444 conf
->reshape_progress
= MaxSector
;
3446 if (setup_geo(&conf
->prev
, mddev
, geo_old
) != conf
->copies
) {
3450 conf
->reshape_progress
= mddev
->reshape_position
;
3451 if (conf
->prev
.far_offset
)
3452 conf
->prev
.stride
= 1 << conf
->prev
.chunk_shift
;
3454 /* far_copies must be 1 */
3455 conf
->prev
.stride
= conf
->dev_sectors
;
3457 spin_lock_init(&conf
->device_lock
);
3458 INIT_LIST_HEAD(&conf
->retry_list
);
3460 spin_lock_init(&conf
->resync_lock
);
3461 init_waitqueue_head(&conf
->wait_barrier
);
3463 conf
->thread
= md_register_thread(raid10d
, mddev
, "raid10");
3467 conf
->mddev
= mddev
;
3472 printk(KERN_ERR
"md/raid10:%s: couldn't allocate memory.\n",
3475 if (conf
->r10bio_pool
)
3476 mempool_destroy(conf
->r10bio_pool
);
3477 kfree(conf
->mirrors
);
3478 safe_put_page(conf
->tmppage
);
3481 return ERR_PTR(err
);
3484 static int run(struct mddev
*mddev
)
3486 struct r10conf
*conf
;
3487 int i
, disk_idx
, chunk_size
;
3488 struct raid10_info
*disk
;
3489 struct md_rdev
*rdev
;
3491 sector_t min_offset_diff
= 0;
3493 bool discard_supported
= false;
3495 if (mddev
->private == NULL
) {
3496 conf
= setup_conf(mddev
);
3498 return PTR_ERR(conf
);
3499 mddev
->private = conf
;
3501 conf
= mddev
->private;
3505 mddev
->thread
= conf
->thread
;
3506 conf
->thread
= NULL
;
3508 chunk_size
= mddev
->chunk_sectors
<< 9;
3510 blk_queue_max_discard_sectors(mddev
->queue
,
3511 mddev
->chunk_sectors
);
3512 blk_queue_io_min(mddev
->queue
, chunk_size
);
3513 if (conf
->geo
.raid_disks
% conf
->geo
.near_copies
)
3514 blk_queue_io_opt(mddev
->queue
, chunk_size
* conf
->geo
.raid_disks
);
3516 blk_queue_io_opt(mddev
->queue
, chunk_size
*
3517 (conf
->geo
.raid_disks
/ conf
->geo
.near_copies
));
3520 rdev_for_each(rdev
, mddev
) {
3522 struct request_queue
*q
;
3524 disk_idx
= rdev
->raid_disk
;
3527 if (disk_idx
>= conf
->geo
.raid_disks
&&
3528 disk_idx
>= conf
->prev
.raid_disks
)
3530 disk
= conf
->mirrors
+ disk_idx
;
3532 if (test_bit(Replacement
, &rdev
->flags
)) {
3533 if (disk
->replacement
)
3535 disk
->replacement
= rdev
;
3541 q
= bdev_get_queue(rdev
->bdev
);
3542 if (q
->merge_bvec_fn
)
3543 mddev
->merge_check_needed
= 1;
3544 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
3545 if (!mddev
->reshape_backwards
)
3549 if (first
|| diff
< min_offset_diff
)
3550 min_offset_diff
= diff
;
3553 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3554 rdev
->data_offset
<< 9);
3556 disk
->head_position
= 0;
3558 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3559 discard_supported
= true;
3562 if (discard_supported
)
3563 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
3565 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
3567 /* need to check that every block has at least one working mirror */
3568 if (!enough(conf
, -1)) {
3569 printk(KERN_ERR
"md/raid10:%s: not enough operational mirrors.\n",
3574 if (conf
->reshape_progress
!= MaxSector
) {
3575 /* must ensure that shape change is supported */
3576 if (conf
->geo
.far_copies
!= 1 &&
3577 conf
->geo
.far_offset
== 0)
3579 if (conf
->prev
.far_copies
!= 1 &&
3580 conf
->geo
.far_offset
== 0)
3584 mddev
->degraded
= 0;
3586 i
< conf
->geo
.raid_disks
3587 || i
< conf
->prev
.raid_disks
;
3590 disk
= conf
->mirrors
+ i
;
3592 if (!disk
->rdev
&& disk
->replacement
) {
3593 /* The replacement is all we have - use it */
3594 disk
->rdev
= disk
->replacement
;
3595 disk
->replacement
= NULL
;
3596 clear_bit(Replacement
, &disk
->rdev
->flags
);
3600 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3601 disk
->head_position
= 0;
3606 disk
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3609 if (mddev
->recovery_cp
!= MaxSector
)
3610 printk(KERN_NOTICE
"md/raid10:%s: not clean"
3611 " -- starting background reconstruction\n",
3614 "md/raid10:%s: active with %d out of %d devices\n",
3615 mdname(mddev
), conf
->geo
.raid_disks
- mddev
->degraded
,
3616 conf
->geo
.raid_disks
);
3618 * Ok, everything is just fine now
3620 mddev
->dev_sectors
= conf
->dev_sectors
;
3621 size
= raid10_size(mddev
, 0, 0);
3622 md_set_array_sectors(mddev
, size
);
3623 mddev
->resync_max_sectors
= size
;
3626 int stripe
= conf
->geo
.raid_disks
*
3627 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
3628 mddev
->queue
->backing_dev_info
.congested_fn
= raid10_congested
;
3629 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
3631 /* Calculate max read-ahead size.
3632 * We need to readahead at least twice a whole stripe....
3635 stripe
/= conf
->geo
.near_copies
;
3636 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
3637 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
3638 blk_queue_merge_bvec(mddev
->queue
, raid10_mergeable_bvec
);
3642 if (md_integrity_register(mddev
))
3645 if (conf
->reshape_progress
!= MaxSector
) {
3646 unsigned long before_length
, after_length
;
3648 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3649 conf
->prev
.far_copies
);
3650 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3651 conf
->geo
.far_copies
);
3653 if (max(before_length
, after_length
) > min_offset_diff
) {
3654 /* This cannot work */
3655 printk("md/raid10: offset difference not enough to continue reshape\n");
3658 conf
->offset_diff
= min_offset_diff
;
3660 conf
->reshape_safe
= conf
->reshape_progress
;
3661 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
3662 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
3663 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
3664 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
3665 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
3672 md_unregister_thread(&mddev
->thread
);
3673 if (conf
->r10bio_pool
)
3674 mempool_destroy(conf
->r10bio_pool
);
3675 safe_put_page(conf
->tmppage
);
3676 kfree(conf
->mirrors
);
3678 mddev
->private = NULL
;
3683 static int stop(struct mddev
*mddev
)
3685 struct r10conf
*conf
= mddev
->private;
3687 raise_barrier(conf
, 0);
3688 lower_barrier(conf
);
3690 md_unregister_thread(&mddev
->thread
);
3692 /* the unplug fn references 'conf'*/
3693 blk_sync_queue(mddev
->queue
);
3695 if (conf
->r10bio_pool
)
3696 mempool_destroy(conf
->r10bio_pool
);
3697 kfree(conf
->mirrors
);
3699 mddev
->private = NULL
;
3703 static void raid10_quiesce(struct mddev
*mddev
, int state
)
3705 struct r10conf
*conf
= mddev
->private;
3709 raise_barrier(conf
, 0);
3712 lower_barrier(conf
);
3717 static int raid10_resize(struct mddev
*mddev
, sector_t sectors
)
3719 /* Resize of 'far' arrays is not supported.
3720 * For 'near' and 'offset' arrays we can set the
3721 * number of sectors used to be an appropriate multiple
3722 * of the chunk size.
3723 * For 'offset', this is far_copies*chunksize.
3724 * For 'near' the multiplier is the LCM of
3725 * near_copies and raid_disks.
3726 * So if far_copies > 1 && !far_offset, fail.
3727 * Else find LCM(raid_disks, near_copy)*far_copies and
3728 * multiply by chunk_size. Then round to this number.
3729 * This is mostly done by raid10_size()
3731 struct r10conf
*conf
= mddev
->private;
3732 sector_t oldsize
, size
;
3734 if (mddev
->reshape_position
!= MaxSector
)
3737 if (conf
->geo
.far_copies
> 1 && !conf
->geo
.far_offset
)
3740 oldsize
= raid10_size(mddev
, 0, 0);
3741 size
= raid10_size(mddev
, sectors
, 0);
3742 if (mddev
->external_size
&&
3743 mddev
->array_sectors
> size
)
3745 if (mddev
->bitmap
) {
3746 int ret
= bitmap_resize(mddev
->bitmap
, size
, 0, 0);
3750 md_set_array_sectors(mddev
, size
);
3751 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3752 revalidate_disk(mddev
->gendisk
);
3753 if (sectors
> mddev
->dev_sectors
&&
3754 mddev
->recovery_cp
> oldsize
) {
3755 mddev
->recovery_cp
= oldsize
;
3756 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3758 calc_sectors(conf
, sectors
);
3759 mddev
->dev_sectors
= conf
->dev_sectors
;
3760 mddev
->resync_max_sectors
= size
;
3764 static void *raid10_takeover_raid0(struct mddev
*mddev
)
3766 struct md_rdev
*rdev
;
3767 struct r10conf
*conf
;
3769 if (mddev
->degraded
> 0) {
3770 printk(KERN_ERR
"md/raid10:%s: Error: degraded raid0!\n",
3772 return ERR_PTR(-EINVAL
);
3775 /* Set new parameters */
3776 mddev
->new_level
= 10;
3777 /* new layout: far_copies = 1, near_copies = 2 */
3778 mddev
->new_layout
= (1<<8) + 2;
3779 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3780 mddev
->delta_disks
= mddev
->raid_disks
;
3781 mddev
->raid_disks
*= 2;
3782 /* make sure it will be not marked as dirty */
3783 mddev
->recovery_cp
= MaxSector
;
3785 conf
= setup_conf(mddev
);
3786 if (!IS_ERR(conf
)) {
3787 rdev_for_each(rdev
, mddev
)
3788 if (rdev
->raid_disk
>= 0)
3789 rdev
->new_raid_disk
= rdev
->raid_disk
* 2;
3796 static void *raid10_takeover(struct mddev
*mddev
)
3798 struct r0conf
*raid0_conf
;
3800 /* raid10 can take over:
3801 * raid0 - providing it has only two drives
3803 if (mddev
->level
== 0) {
3804 /* for raid0 takeover only one zone is supported */
3805 raid0_conf
= mddev
->private;
3806 if (raid0_conf
->nr_strip_zones
> 1) {
3807 printk(KERN_ERR
"md/raid10:%s: cannot takeover raid 0"
3808 " with more than one zone.\n",
3810 return ERR_PTR(-EINVAL
);
3812 return raid10_takeover_raid0(mddev
);
3814 return ERR_PTR(-EINVAL
);
3817 static int raid10_check_reshape(struct mddev
*mddev
)
3819 /* Called when there is a request to change
3820 * - layout (to ->new_layout)
3821 * - chunk size (to ->new_chunk_sectors)
3822 * - raid_disks (by delta_disks)
3823 * or when trying to restart a reshape that was ongoing.
3825 * We need to validate the request and possibly allocate
3826 * space if that might be an issue later.
3828 * Currently we reject any reshape of a 'far' mode array,
3829 * allow chunk size to change if new is generally acceptable,
3830 * allow raid_disks to increase, and allow
3831 * a switch between 'near' mode and 'offset' mode.
3833 struct r10conf
*conf
= mddev
->private;
3836 if (conf
->geo
.far_copies
!= 1 && !conf
->geo
.far_offset
)
3839 if (setup_geo(&geo
, mddev
, geo_start
) != conf
->copies
)
3840 /* mustn't change number of copies */
3842 if (geo
.far_copies
> 1 && !geo
.far_offset
)
3843 /* Cannot switch to 'far' mode */
3846 if (mddev
->array_sectors
& geo
.chunk_mask
)
3847 /* not factor of array size */
3850 if (!enough(conf
, -1))
3853 kfree(conf
->mirrors_new
);
3854 conf
->mirrors_new
= NULL
;
3855 if (mddev
->delta_disks
> 0) {
3856 /* allocate new 'mirrors' list */
3857 conf
->mirrors_new
= kzalloc(
3858 sizeof(struct raid10_info
)
3859 *(mddev
->raid_disks
+
3860 mddev
->delta_disks
),
3862 if (!conf
->mirrors_new
)
3869 * Need to check if array has failed when deciding whether to:
3871 * - remove non-faulty devices
3874 * This determination is simple when no reshape is happening.
3875 * However if there is a reshape, we need to carefully check
3876 * both the before and after sections.
3877 * This is because some failed devices may only affect one
3878 * of the two sections, and some non-in_sync devices may
3879 * be insync in the section most affected by failed devices.
3881 static int calc_degraded(struct r10conf
*conf
)
3883 int degraded
, degraded2
;
3888 /* 'prev' section first */
3889 for (i
= 0; i
< conf
->prev
.raid_disks
; i
++) {
3890 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3891 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3893 else if (!test_bit(In_sync
, &rdev
->flags
))
3894 /* When we can reduce the number of devices in
3895 * an array, this might not contribute to
3896 * 'degraded'. It does now.
3901 if (conf
->geo
.raid_disks
== conf
->prev
.raid_disks
)
3905 for (i
= 0; i
< conf
->geo
.raid_disks
; i
++) {
3906 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
3907 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
3909 else if (!test_bit(In_sync
, &rdev
->flags
)) {
3910 /* If reshape is increasing the number of devices,
3911 * this section has already been recovered, so
3912 * it doesn't contribute to degraded.
3915 if (conf
->geo
.raid_disks
<= conf
->prev
.raid_disks
)
3920 if (degraded2
> degraded
)
3925 static int raid10_start_reshape(struct mddev
*mddev
)
3927 /* A 'reshape' has been requested. This commits
3928 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3929 * This also checks if there are enough spares and adds them
3931 * We currently require enough spares to make the final
3932 * array non-degraded. We also require that the difference
3933 * between old and new data_offset - on each device - is
3934 * enough that we never risk over-writing.
3937 unsigned long before_length
, after_length
;
3938 sector_t min_offset_diff
= 0;
3941 struct r10conf
*conf
= mddev
->private;
3942 struct md_rdev
*rdev
;
3946 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
3949 if (setup_geo(&new, mddev
, geo_start
) != conf
->copies
)
3952 before_length
= ((1 << conf
->prev
.chunk_shift
) *
3953 conf
->prev
.far_copies
);
3954 after_length
= ((1 << conf
->geo
.chunk_shift
) *
3955 conf
->geo
.far_copies
);
3957 rdev_for_each(rdev
, mddev
) {
3958 if (!test_bit(In_sync
, &rdev
->flags
)
3959 && !test_bit(Faulty
, &rdev
->flags
))
3961 if (rdev
->raid_disk
>= 0) {
3962 long long diff
= (rdev
->new_data_offset
3963 - rdev
->data_offset
);
3964 if (!mddev
->reshape_backwards
)
3968 if (first
|| diff
< min_offset_diff
)
3969 min_offset_diff
= diff
;
3973 if (max(before_length
, after_length
) > min_offset_diff
)
3976 if (spares
< mddev
->delta_disks
)
3979 conf
->offset_diff
= min_offset_diff
;
3980 spin_lock_irq(&conf
->device_lock
);
3981 if (conf
->mirrors_new
) {
3982 memcpy(conf
->mirrors_new
, conf
->mirrors
,
3983 sizeof(struct raid10_info
)*conf
->prev
.raid_disks
);
3985 kfree(conf
->mirrors_old
); /* FIXME and elsewhere */
3986 conf
->mirrors_old
= conf
->mirrors
;
3987 conf
->mirrors
= conf
->mirrors_new
;
3988 conf
->mirrors_new
= NULL
;
3990 setup_geo(&conf
->geo
, mddev
, geo_start
);
3992 if (mddev
->reshape_backwards
) {
3993 sector_t size
= raid10_size(mddev
, 0, 0);
3994 if (size
< mddev
->array_sectors
) {
3995 spin_unlock_irq(&conf
->device_lock
);
3996 printk(KERN_ERR
"md/raid10:%s: array size must be reduce before number of disks\n",
4000 mddev
->resync_max_sectors
= size
;
4001 conf
->reshape_progress
= size
;
4003 conf
->reshape_progress
= 0;
4004 spin_unlock_irq(&conf
->device_lock
);
4006 if (mddev
->delta_disks
&& mddev
->bitmap
) {
4007 ret
= bitmap_resize(mddev
->bitmap
,
4008 raid10_size(mddev
, 0,
4009 conf
->geo
.raid_disks
),
4014 if (mddev
->delta_disks
> 0) {
4015 rdev_for_each(rdev
, mddev
)
4016 if (rdev
->raid_disk
< 0 &&
4017 !test_bit(Faulty
, &rdev
->flags
)) {
4018 if (raid10_add_disk(mddev
, rdev
) == 0) {
4019 if (rdev
->raid_disk
>=
4020 conf
->prev
.raid_disks
)
4021 set_bit(In_sync
, &rdev
->flags
);
4023 rdev
->recovery_offset
= 0;
4025 if (sysfs_link_rdev(mddev
, rdev
))
4026 /* Failure here is OK */;
4028 } else if (rdev
->raid_disk
>= conf
->prev
.raid_disks
4029 && !test_bit(Faulty
, &rdev
->flags
)) {
4030 /* This is a spare that was manually added */
4031 set_bit(In_sync
, &rdev
->flags
);
4034 /* When a reshape changes the number of devices,
4035 * ->degraded is measured against the larger of the
4036 * pre and post numbers.
4038 spin_lock_irq(&conf
->device_lock
);
4039 mddev
->degraded
= calc_degraded(conf
);
4040 spin_unlock_irq(&conf
->device_lock
);
4041 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4042 mddev
->reshape_position
= conf
->reshape_progress
;
4043 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4045 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4046 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4047 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4048 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4050 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4052 if (!mddev
->sync_thread
) {
4056 conf
->reshape_checkpoint
= jiffies
;
4057 md_wakeup_thread(mddev
->sync_thread
);
4058 md_new_event(mddev
);
4062 mddev
->recovery
= 0;
4063 spin_lock_irq(&conf
->device_lock
);
4064 conf
->geo
= conf
->prev
;
4065 mddev
->raid_disks
= conf
->geo
.raid_disks
;
4066 rdev_for_each(rdev
, mddev
)
4067 rdev
->new_data_offset
= rdev
->data_offset
;
4069 conf
->reshape_progress
= MaxSector
;
4070 mddev
->reshape_position
= MaxSector
;
4071 spin_unlock_irq(&conf
->device_lock
);
4075 /* Calculate the last device-address that could contain
4076 * any block from the chunk that includes the array-address 's'
4077 * and report the next address.
4078 * i.e. the address returned will be chunk-aligned and after
4079 * any data that is in the chunk containing 's'.
4081 static sector_t
last_dev_address(sector_t s
, struct geom
*geo
)
4083 s
= (s
| geo
->chunk_mask
) + 1;
4084 s
>>= geo
->chunk_shift
;
4085 s
*= geo
->near_copies
;
4086 s
= DIV_ROUND_UP_SECTOR_T(s
, geo
->raid_disks
);
4087 s
*= geo
->far_copies
;
4088 s
<<= geo
->chunk_shift
;
4092 /* Calculate the first device-address that could contain
4093 * any block from the chunk that includes the array-address 's'.
4094 * This too will be the start of a chunk
4096 static sector_t
first_dev_address(sector_t s
, struct geom
*geo
)
4098 s
>>= geo
->chunk_shift
;
4099 s
*= geo
->near_copies
;
4100 sector_div(s
, geo
->raid_disks
);
4101 s
*= geo
->far_copies
;
4102 s
<<= geo
->chunk_shift
;
4106 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
,
4109 /* We simply copy at most one chunk (smallest of old and new)
4110 * at a time, possibly less if that exceeds RESYNC_PAGES,
4111 * or we hit a bad block or something.
4112 * This might mean we pause for normal IO in the middle of
4113 * a chunk, but that is not a problem was mddev->reshape_position
4114 * can record any location.
4116 * If we will want to write to a location that isn't
4117 * yet recorded as 'safe' (i.e. in metadata on disk) then
4118 * we need to flush all reshape requests and update the metadata.
4120 * When reshaping forwards (e.g. to more devices), we interpret
4121 * 'safe' as the earliest block which might not have been copied
4122 * down yet. We divide this by previous stripe size and multiply
4123 * by previous stripe length to get lowest device offset that we
4124 * cannot write to yet.
4125 * We interpret 'sector_nr' as an address that we want to write to.
4126 * From this we use last_device_address() to find where we might
4127 * write to, and first_device_address on the 'safe' position.
4128 * If this 'next' write position is after the 'safe' position,
4129 * we must update the metadata to increase the 'safe' position.
4131 * When reshaping backwards, we round in the opposite direction
4132 * and perform the reverse test: next write position must not be
4133 * less than current safe position.
4135 * In all this the minimum difference in data offsets
4136 * (conf->offset_diff - always positive) allows a bit of slack,
4137 * so next can be after 'safe', but not by more than offset_disk
4139 * We need to prepare all the bios here before we start any IO
4140 * to ensure the size we choose is acceptable to all devices.
4141 * The means one for each copy for write-out and an extra one for
4143 * We store the read-in bio in ->master_bio and the others in
4144 * ->devs[x].bio and ->devs[x].repl_bio.
4146 struct r10conf
*conf
= mddev
->private;
4147 struct r10bio
*r10_bio
;
4148 sector_t next
, safe
, last
;
4152 struct md_rdev
*rdev
;
4155 struct bio
*bio
, *read_bio
;
4156 int sectors_done
= 0;
4158 if (sector_nr
== 0) {
4159 /* If restarting in the middle, skip the initial sectors */
4160 if (mddev
->reshape_backwards
&&
4161 conf
->reshape_progress
< raid10_size(mddev
, 0, 0)) {
4162 sector_nr
= (raid10_size(mddev
, 0, 0)
4163 - conf
->reshape_progress
);
4164 } else if (!mddev
->reshape_backwards
&&
4165 conf
->reshape_progress
> 0)
4166 sector_nr
= conf
->reshape_progress
;
4168 mddev
->curr_resync_completed
= sector_nr
;
4169 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4175 /* We don't use sector_nr to track where we are up to
4176 * as that doesn't work well for ->reshape_backwards.
4177 * So just use ->reshape_progress.
4179 if (mddev
->reshape_backwards
) {
4180 /* 'next' is the earliest device address that we might
4181 * write to for this chunk in the new layout
4183 next
= first_dev_address(conf
->reshape_progress
- 1,
4186 /* 'safe' is the last device address that we might read from
4187 * in the old layout after a restart
4189 safe
= last_dev_address(conf
->reshape_safe
- 1,
4192 if (next
+ conf
->offset_diff
< safe
)
4195 last
= conf
->reshape_progress
- 1;
4196 sector_nr
= last
& ~(sector_t
)(conf
->geo
.chunk_mask
4197 & conf
->prev
.chunk_mask
);
4198 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 < last
)
4199 sector_nr
= last
+ 1 - RESYNC_BLOCK_SIZE
/512;
4201 /* 'next' is after the last device address that we
4202 * might write to for this chunk in the new layout
4204 next
= last_dev_address(conf
->reshape_progress
, &conf
->geo
);
4206 /* 'safe' is the earliest device address that we might
4207 * read from in the old layout after a restart
4209 safe
= first_dev_address(conf
->reshape_safe
, &conf
->prev
);
4211 /* Need to update metadata if 'next' might be beyond 'safe'
4212 * as that would possibly corrupt data
4214 if (next
> safe
+ conf
->offset_diff
)
4217 sector_nr
= conf
->reshape_progress
;
4218 last
= sector_nr
| (conf
->geo
.chunk_mask
4219 & conf
->prev
.chunk_mask
);
4221 if (sector_nr
+ RESYNC_BLOCK_SIZE
/512 <= last
)
4222 last
= sector_nr
+ RESYNC_BLOCK_SIZE
/512 - 1;
4226 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4227 /* Need to update reshape_position in metadata */
4229 mddev
->reshape_position
= conf
->reshape_progress
;
4230 if (mddev
->reshape_backwards
)
4231 mddev
->curr_resync_completed
= raid10_size(mddev
, 0, 0)
4232 - conf
->reshape_progress
;
4234 mddev
->curr_resync_completed
= conf
->reshape_progress
;
4235 conf
->reshape_checkpoint
= jiffies
;
4236 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4237 md_wakeup_thread(mddev
->thread
);
4238 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4239 kthread_should_stop());
4240 conf
->reshape_safe
= mddev
->reshape_position
;
4241 allow_barrier(conf
);
4245 /* Now schedule reads for blocks from sector_nr to last */
4246 r10_bio
= mempool_alloc(conf
->r10buf_pool
, GFP_NOIO
);
4247 raise_barrier(conf
, sectors_done
!= 0);
4248 atomic_set(&r10_bio
->remaining
, 0);
4249 r10_bio
->mddev
= mddev
;
4250 r10_bio
->sector
= sector_nr
;
4251 set_bit(R10BIO_IsReshape
, &r10_bio
->state
);
4252 r10_bio
->sectors
= last
- sector_nr
+ 1;
4253 rdev
= read_balance(conf
, r10_bio
, &max_sectors
);
4254 BUG_ON(!test_bit(R10BIO_Previous
, &r10_bio
->state
));
4257 /* Cannot read from here, so need to record bad blocks
4258 * on all the target devices.
4261 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
4262 return sectors_done
;
4265 read_bio
= bio_alloc_mddev(GFP_KERNEL
, RESYNC_PAGES
, mddev
);
4267 read_bio
->bi_bdev
= rdev
->bdev
;
4268 read_bio
->bi_sector
= (r10_bio
->devs
[r10_bio
->read_slot
].addr
4269 + rdev
->data_offset
);
4270 read_bio
->bi_private
= r10_bio
;
4271 read_bio
->bi_end_io
= end_sync_read
;
4272 read_bio
->bi_rw
= READ
;
4273 read_bio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4274 read_bio
->bi_flags
|= 1 << BIO_UPTODATE
;
4275 read_bio
->bi_vcnt
= 0;
4276 read_bio
->bi_idx
= 0;
4277 read_bio
->bi_size
= 0;
4278 r10_bio
->master_bio
= read_bio
;
4279 r10_bio
->read_slot
= r10_bio
->devs
[r10_bio
->read_slot
].devnum
;
4281 /* Now find the locations in the new layout */
4282 __raid10_find_phys(&conf
->geo
, r10_bio
);
4285 read_bio
->bi_next
= NULL
;
4287 for (s
= 0; s
< conf
->copies
*2; s
++) {
4289 int d
= r10_bio
->devs
[s
/2].devnum
;
4290 struct md_rdev
*rdev2
;
4292 rdev2
= conf
->mirrors
[d
].replacement
;
4293 b
= r10_bio
->devs
[s
/2].repl_bio
;
4295 rdev2
= conf
->mirrors
[d
].rdev
;
4296 b
= r10_bio
->devs
[s
/2].bio
;
4298 if (!rdev2
|| test_bit(Faulty
, &rdev2
->flags
))
4300 b
->bi_bdev
= rdev2
->bdev
;
4301 b
->bi_sector
= r10_bio
->devs
[s
/2].addr
+ rdev2
->new_data_offset
;
4302 b
->bi_private
= r10_bio
;
4303 b
->bi_end_io
= end_reshape_write
;
4305 b
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
4306 b
->bi_flags
|= 1 << BIO_UPTODATE
;
4314 /* Now add as many pages as possible to all of these bios. */
4317 for (s
= 0 ; s
< max_sectors
; s
+= PAGE_SIZE
>> 9) {
4318 struct page
*page
= r10_bio
->devs
[0].bio
->bi_io_vec
[s
/(PAGE_SIZE
>>9)].bv_page
;
4319 int len
= (max_sectors
- s
) << 9;
4320 if (len
> PAGE_SIZE
)
4322 for (bio
= blist
; bio
; bio
= bio
->bi_next
) {
4324 if (bio_add_page(bio
, page
, len
, 0))
4327 /* Didn't fit, must stop */
4329 bio2
&& bio2
!= bio
;
4330 bio2
= bio2
->bi_next
) {
4331 /* Remove last page from this bio */
4333 bio2
->bi_size
-= len
;
4334 bio2
->bi_flags
&= ~(1<<BIO_SEG_VALID
);
4338 sector_nr
+= len
>> 9;
4339 nr_sectors
+= len
>> 9;
4342 r10_bio
->sectors
= nr_sectors
;
4344 /* Now submit the read */
4345 md_sync_acct(read_bio
->bi_bdev
, r10_bio
->sectors
);
4346 atomic_inc(&r10_bio
->remaining
);
4347 read_bio
->bi_next
= NULL
;
4348 generic_make_request(read_bio
);
4349 sector_nr
+= nr_sectors
;
4350 sectors_done
+= nr_sectors
;
4351 if (sector_nr
<= last
)
4354 /* Now that we have done the whole section we can
4355 * update reshape_progress
4357 if (mddev
->reshape_backwards
)
4358 conf
->reshape_progress
-= sectors_done
;
4360 conf
->reshape_progress
+= sectors_done
;
4362 return sectors_done
;
4365 static void end_reshape_request(struct r10bio
*r10_bio
);
4366 static int handle_reshape_read_error(struct mddev
*mddev
,
4367 struct r10bio
*r10_bio
);
4368 static void reshape_request_write(struct mddev
*mddev
, struct r10bio
*r10_bio
)
4370 /* Reshape read completed. Hopefully we have a block
4372 * If we got a read error then we do sync 1-page reads from
4373 * elsewhere until we find the data - or give up.
4375 struct r10conf
*conf
= mddev
->private;
4378 if (!test_bit(R10BIO_Uptodate
, &r10_bio
->state
))
4379 if (handle_reshape_read_error(mddev
, r10_bio
) < 0) {
4380 /* Reshape has been aborted */
4381 md_done_sync(mddev
, r10_bio
->sectors
, 0);
4385 /* We definitely have the data in the pages, schedule the
4388 atomic_set(&r10_bio
->remaining
, 1);
4389 for (s
= 0; s
< conf
->copies
*2; s
++) {
4391 int d
= r10_bio
->devs
[s
/2].devnum
;
4392 struct md_rdev
*rdev
;
4394 rdev
= conf
->mirrors
[d
].replacement
;
4395 b
= r10_bio
->devs
[s
/2].repl_bio
;
4397 rdev
= conf
->mirrors
[d
].rdev
;
4398 b
= r10_bio
->devs
[s
/2].bio
;
4400 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
4402 atomic_inc(&rdev
->nr_pending
);
4403 md_sync_acct(b
->bi_bdev
, r10_bio
->sectors
);
4404 atomic_inc(&r10_bio
->remaining
);
4406 generic_make_request(b
);
4408 end_reshape_request(r10_bio
);
4411 static void end_reshape(struct r10conf
*conf
)
4413 if (test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
))
4416 spin_lock_irq(&conf
->device_lock
);
4417 conf
->prev
= conf
->geo
;
4418 md_finish_reshape(conf
->mddev
);
4420 conf
->reshape_progress
= MaxSector
;
4421 spin_unlock_irq(&conf
->device_lock
);
4423 /* read-ahead size must cover two whole stripes, which is
4424 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4426 if (conf
->mddev
->queue
) {
4427 int stripe
= conf
->geo
.raid_disks
*
4428 ((conf
->mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4429 stripe
/= conf
->geo
.near_copies
;
4430 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4431 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4437 static int handle_reshape_read_error(struct mddev
*mddev
,
4438 struct r10bio
*r10_bio
)
4440 /* Use sync reads to get the blocks from somewhere else */
4441 int sectors
= r10_bio
->sectors
;
4442 struct r10conf
*conf
= mddev
->private;
4444 struct r10bio r10_bio
;
4445 struct r10dev devs
[conf
->copies
];
4447 struct r10bio
*r10b
= &on_stack
.r10_bio
;
4450 struct bio_vec
*bvec
= r10_bio
->master_bio
->bi_io_vec
;
4452 r10b
->sector
= r10_bio
->sector
;
4453 __raid10_find_phys(&conf
->prev
, r10b
);
4458 int first_slot
= slot
;
4460 if (s
> (PAGE_SIZE
>> 9))
4464 int d
= r10b
->devs
[slot
].devnum
;
4465 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4468 test_bit(Faulty
, &rdev
->flags
) ||
4469 !test_bit(In_sync
, &rdev
->flags
))
4472 addr
= r10b
->devs
[slot
].addr
+ idx
* PAGE_SIZE
;
4473 success
= sync_page_io(rdev
,
4482 if (slot
>= conf
->copies
)
4484 if (slot
== first_slot
)
4488 /* couldn't read this block, must give up */
4489 set_bit(MD_RECOVERY_INTR
,
4499 static void end_reshape_write(struct bio
*bio
, int error
)
4501 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
4502 struct r10bio
*r10_bio
= bio
->bi_private
;
4503 struct mddev
*mddev
= r10_bio
->mddev
;
4504 struct r10conf
*conf
= mddev
->private;
4508 struct md_rdev
*rdev
= NULL
;
4510 d
= find_bio_disk(conf
, r10_bio
, bio
, &slot
, &repl
);
4512 rdev
= conf
->mirrors
[d
].replacement
;
4515 rdev
= conf
->mirrors
[d
].rdev
;
4519 /* FIXME should record badblock */
4520 md_error(mddev
, rdev
);
4523 rdev_dec_pending(rdev
, mddev
);
4524 end_reshape_request(r10_bio
);
4527 static void end_reshape_request(struct r10bio
*r10_bio
)
4529 if (!atomic_dec_and_test(&r10_bio
->remaining
))
4531 md_done_sync(r10_bio
->mddev
, r10_bio
->sectors
, 1);
4532 bio_put(r10_bio
->master_bio
);
4536 static void raid10_finish_reshape(struct mddev
*mddev
)
4538 struct r10conf
*conf
= mddev
->private;
4540 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
4543 if (mddev
->delta_disks
> 0) {
4544 sector_t size
= raid10_size(mddev
, 0, 0);
4545 md_set_array_sectors(mddev
, size
);
4546 if (mddev
->recovery_cp
> mddev
->resync_max_sectors
) {
4547 mddev
->recovery_cp
= mddev
->resync_max_sectors
;
4548 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4550 mddev
->resync_max_sectors
= size
;
4551 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4552 revalidate_disk(mddev
->gendisk
);
4555 for (d
= conf
->geo
.raid_disks
;
4556 d
< conf
->geo
.raid_disks
- mddev
->delta_disks
;
4558 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
4560 clear_bit(In_sync
, &rdev
->flags
);
4561 rdev
= conf
->mirrors
[d
].replacement
;
4563 clear_bit(In_sync
, &rdev
->flags
);
4566 mddev
->layout
= mddev
->new_layout
;
4567 mddev
->chunk_sectors
= 1 << conf
->geo
.chunk_shift
;
4568 mddev
->reshape_position
= MaxSector
;
4569 mddev
->delta_disks
= 0;
4570 mddev
->reshape_backwards
= 0;
4573 static struct md_personality raid10_personality
=
4577 .owner
= THIS_MODULE
,
4578 .make_request
= make_request
,
4582 .error_handler
= error
,
4583 .hot_add_disk
= raid10_add_disk
,
4584 .hot_remove_disk
= raid10_remove_disk
,
4585 .spare_active
= raid10_spare_active
,
4586 .sync_request
= sync_request
,
4587 .quiesce
= raid10_quiesce
,
4588 .size
= raid10_size
,
4589 .resize
= raid10_resize
,
4590 .takeover
= raid10_takeover
,
4591 .check_reshape
= raid10_check_reshape
,
4592 .start_reshape
= raid10_start_reshape
,
4593 .finish_reshape
= raid10_finish_reshape
,
4596 static int __init
raid_init(void)
4598 return register_md_personality(&raid10_personality
);
4601 static void raid_exit(void)
4603 unregister_md_personality(&raid10_personality
);
4606 module_init(raid_init
);
4607 module_exit(raid_exit
);
4608 MODULE_LICENSE("GPL");
4609 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4610 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4611 MODULE_ALIAS("md-raid10");
4612 MODULE_ALIAS("md-level-10");
4614 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);