2 * raid1.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * RAID-1 management functions.
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
45 * Number of guaranteed r1bios in case of extreme VM load:
47 #define NR_RAID1_BIOS 256
49 /* when we get a read error on a read-only array, we redirect to another
50 * device without failing the first device, or trying to over-write to
51 * correct the read error. To keep track of bad blocks on a per-bio
52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
54 #define IO_BLOCKED ((struct bio *)1)
55 /* When we successfully write to a known bad-block, we need to remove the
56 * bad-block marking which must be done from process context. So we record
57 * the success by setting devs[n].bio to IO_MADE_GOOD
59 #define IO_MADE_GOOD ((struct bio *)2)
61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
63 /* When there are this many requests queue to be written by
64 * the raid1 thread, we become 'congested' to provide back-pressure
67 static int max_queued_requests
= 1024;
69 static void allow_barrier(struct r1conf
*conf
, sector_t start_next_window
,
71 static void lower_barrier(struct r1conf
*conf
);
73 static void * r1bio_pool_alloc(gfp_t gfp_flags
, void *data
)
75 struct pool_info
*pi
= data
;
76 int size
= offsetof(struct r1bio
, bios
[pi
->raid_disks
]);
78 /* allocate a r1bio with room for raid_disks entries in the bios array */
79 return kzalloc(size
, gfp_flags
);
82 static void r1bio_pool_free(void *r1_bio
, void *data
)
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_DEPTH 32
89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
93 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
95 static void * r1buf_pool_alloc(gfp_t gfp_flags
, void *data
)
97 struct pool_info
*pi
= data
;
102 r1_bio
= r1bio_pool_alloc(gfp_flags
, pi
);
107 * Allocate bios : 1 for reading, n-1 for writing
109 for (j
= pi
->raid_disks
; j
-- ; ) {
110 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
113 r1_bio
->bios
[j
] = bio
;
116 * Allocate RESYNC_PAGES data pages and attach them to
118 * If this is a user-requested check/repair, allocate
119 * RESYNC_PAGES for each bio.
121 if (test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
))
126 bio
= r1_bio
->bios
[j
];
127 bio
->bi_vcnt
= RESYNC_PAGES
;
129 if (bio_alloc_pages(bio
, gfp_flags
))
132 /* If not user-requests, copy the page pointers to all bios */
133 if (!test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
)) {
134 for (i
=0; i
<RESYNC_PAGES
; i
++)
135 for (j
=1; j
<pi
->raid_disks
; j
++)
136 r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
=
137 r1_bio
->bios
[0]->bi_io_vec
[i
].bv_page
;
140 r1_bio
->master_bio
= NULL
;
145 while (++j
< pi
->raid_disks
)
146 bio_put(r1_bio
->bios
[j
]);
147 r1bio_pool_free(r1_bio
, data
);
151 static void r1buf_pool_free(void *__r1_bio
, void *data
)
153 struct pool_info
*pi
= data
;
155 struct r1bio
*r1bio
= __r1_bio
;
157 for (i
= 0; i
< RESYNC_PAGES
; i
++)
158 for (j
= pi
->raid_disks
; j
-- ;) {
160 r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
!=
161 r1bio
->bios
[0]->bi_io_vec
[i
].bv_page
)
162 safe_put_page(r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
164 for (i
=0 ; i
< pi
->raid_disks
; i
++)
165 bio_put(r1bio
->bios
[i
]);
167 r1bio_pool_free(r1bio
, data
);
170 static void put_all_bios(struct r1conf
*conf
, struct r1bio
*r1_bio
)
174 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
175 struct bio
**bio
= r1_bio
->bios
+ i
;
176 if (!BIO_SPECIAL(*bio
))
182 static void free_r1bio(struct r1bio
*r1_bio
)
184 struct r1conf
*conf
= r1_bio
->mddev
->private;
186 put_all_bios(conf
, r1_bio
);
187 mempool_free(r1_bio
, conf
->r1bio_pool
);
190 static void put_buf(struct r1bio
*r1_bio
)
192 struct r1conf
*conf
= r1_bio
->mddev
->private;
195 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
196 struct bio
*bio
= r1_bio
->bios
[i
];
198 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
201 mempool_free(r1_bio
, conf
->r1buf_pool
);
206 static void reschedule_retry(struct r1bio
*r1_bio
)
209 struct mddev
*mddev
= r1_bio
->mddev
;
210 struct r1conf
*conf
= mddev
->private;
212 spin_lock_irqsave(&conf
->device_lock
, flags
);
213 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
215 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
217 wake_up(&conf
->wait_barrier
);
218 md_wakeup_thread(mddev
->thread
);
222 * raid_end_bio_io() is called when we have finished servicing a mirrored
223 * operation and are ready to return a success/failure code to the buffer
226 static void call_bio_endio(struct r1bio
*r1_bio
)
228 struct bio
*bio
= r1_bio
->master_bio
;
230 struct r1conf
*conf
= r1_bio
->mddev
->private;
231 sector_t start_next_window
= r1_bio
->start_next_window
;
232 sector_t bi_sector
= bio
->bi_sector
;
234 if (bio
->bi_phys_segments
) {
236 spin_lock_irqsave(&conf
->device_lock
, flags
);
237 bio
->bi_phys_segments
--;
238 done
= (bio
->bi_phys_segments
== 0);
239 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
241 * make_request() might be waiting for
242 * bi_phys_segments to decrease
244 wake_up(&conf
->wait_barrier
);
248 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
249 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
253 * Wake up any possible resync thread that waits for the device
256 allow_barrier(conf
, start_next_window
, bi_sector
);
260 static void raid_end_bio_io(struct r1bio
*r1_bio
)
262 struct bio
*bio
= r1_bio
->master_bio
;
264 /* if nobody has done the final endio yet, do it now */
265 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
266 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
267 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
268 (unsigned long long) bio
->bi_sector
,
269 (unsigned long long) bio
->bi_sector
+
270 bio_sectors(bio
) - 1);
272 call_bio_endio(r1_bio
);
278 * Update disk head position estimator based on IRQ completion info.
280 static inline void update_head_pos(int disk
, struct r1bio
*r1_bio
)
282 struct r1conf
*conf
= r1_bio
->mddev
->private;
284 conf
->mirrors
[disk
].head_position
=
285 r1_bio
->sector
+ (r1_bio
->sectors
);
289 * Find the disk number which triggered given bio
291 static int find_bio_disk(struct r1bio
*r1_bio
, struct bio
*bio
)
294 struct r1conf
*conf
= r1_bio
->mddev
->private;
295 int raid_disks
= conf
->raid_disks
;
297 for (mirror
= 0; mirror
< raid_disks
* 2; mirror
++)
298 if (r1_bio
->bios
[mirror
] == bio
)
301 BUG_ON(mirror
== raid_disks
* 2);
302 update_head_pos(mirror
, r1_bio
);
307 static void raid1_end_read_request(struct bio
*bio
, int error
)
309 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
310 struct r1bio
*r1_bio
= bio
->bi_private
;
312 struct r1conf
*conf
= r1_bio
->mddev
->private;
314 mirror
= r1_bio
->read_disk
;
316 * this branch is our 'one mirror IO has finished' event handler:
318 update_head_pos(mirror
, r1_bio
);
321 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
323 /* If all other devices have failed, we want to return
324 * the error upwards rather than fail the last device.
325 * Here we redefine "uptodate" to mean "Don't want to retry"
328 spin_lock_irqsave(&conf
->device_lock
, flags
);
329 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
330 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
331 !test_bit(Faulty
, &conf
->mirrors
[mirror
].rdev
->flags
)))
333 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
337 raid_end_bio_io(r1_bio
);
338 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
343 char b
[BDEVNAME_SIZE
];
345 KERN_ERR
"md/raid1:%s: %s: "
346 "rescheduling sector %llu\n",
348 bdevname(conf
->mirrors
[mirror
].rdev
->bdev
,
350 (unsigned long long)r1_bio
->sector
);
351 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
352 reschedule_retry(r1_bio
);
353 /* don't drop the reference on read_disk yet */
357 static void close_write(struct r1bio
*r1_bio
)
359 /* it really is the end of this request */
360 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
361 /* free extra copy of the data pages */
362 int i
= r1_bio
->behind_page_count
;
364 safe_put_page(r1_bio
->behind_bvecs
[i
].bv_page
);
365 kfree(r1_bio
->behind_bvecs
);
366 r1_bio
->behind_bvecs
= NULL
;
368 /* clear the bitmap if all writes complete successfully */
369 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
371 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
372 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
373 md_write_end(r1_bio
->mddev
);
376 static void r1_bio_write_done(struct r1bio
*r1_bio
)
378 if (!atomic_dec_and_test(&r1_bio
->remaining
))
381 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
382 reschedule_retry(r1_bio
);
385 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
))
386 reschedule_retry(r1_bio
);
388 raid_end_bio_io(r1_bio
);
392 static void raid1_end_write_request(struct bio
*bio
, int error
)
394 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
395 struct r1bio
*r1_bio
= bio
->bi_private
;
396 int mirror
, behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
397 struct r1conf
*conf
= r1_bio
->mddev
->private;
398 struct bio
*to_put
= NULL
;
400 mirror
= find_bio_disk(r1_bio
, bio
);
403 * 'one mirror IO has finished' event handler:
406 set_bit(WriteErrorSeen
,
407 &conf
->mirrors
[mirror
].rdev
->flags
);
408 if (!test_and_set_bit(WantReplacement
,
409 &conf
->mirrors
[mirror
].rdev
->flags
))
410 set_bit(MD_RECOVERY_NEEDED
, &
411 conf
->mddev
->recovery
);
413 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
416 * Set R1BIO_Uptodate in our master bio, so that we
417 * will return a good error code for to the higher
418 * levels even if IO on some other mirrored buffer
421 * The 'master' represents the composite IO operation
422 * to user-side. So if something waits for IO, then it
423 * will wait for the 'master' bio.
428 r1_bio
->bios
[mirror
] = NULL
;
431 * Do not set R1BIO_Uptodate if the current device is
432 * rebuilding or Faulty. This is because we cannot use
433 * such device for properly reading the data back (we could
434 * potentially use it, if the current write would have felt
435 * before rdev->recovery_offset, but for simplicity we don't
438 if (test_bit(In_sync
, &conf
->mirrors
[mirror
].rdev
->flags
) &&
439 !test_bit(Faulty
, &conf
->mirrors
[mirror
].rdev
->flags
))
440 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
442 /* Maybe we can clear some bad blocks. */
443 if (is_badblock(conf
->mirrors
[mirror
].rdev
,
444 r1_bio
->sector
, r1_bio
->sectors
,
445 &first_bad
, &bad_sectors
)) {
446 r1_bio
->bios
[mirror
] = IO_MADE_GOOD
;
447 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
452 if (test_bit(WriteMostly
, &conf
->mirrors
[mirror
].rdev
->flags
))
453 atomic_dec(&r1_bio
->behind_remaining
);
456 * In behind mode, we ACK the master bio once the I/O
457 * has safely reached all non-writemostly
458 * disks. Setting the Returned bit ensures that this
459 * gets done only once -- we don't ever want to return
460 * -EIO here, instead we'll wait
462 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
463 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
464 /* Maybe we can return now */
465 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
466 struct bio
*mbio
= r1_bio
->master_bio
;
467 pr_debug("raid1: behind end write sectors"
469 (unsigned long long) mbio
->bi_sector
,
470 (unsigned long long) mbio
->bi_sector
+
471 bio_sectors(mbio
) - 1);
472 call_bio_endio(r1_bio
);
476 if (r1_bio
->bios
[mirror
] == NULL
)
477 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
,
481 * Let's see if all mirrored write operations have finished
484 r1_bio_write_done(r1_bio
);
492 * This routine returns the disk from which the requested read should
493 * be done. There is a per-array 'next expected sequential IO' sector
494 * number - if this matches on the next IO then we use the last disk.
495 * There is also a per-disk 'last know head position' sector that is
496 * maintained from IRQ contexts, both the normal and the resync IO
497 * completion handlers update this position correctly. If there is no
498 * perfect sequential match then we pick the disk whose head is closest.
500 * If there are 2 mirrors in the same 2 devices, performance degrades
501 * because position is mirror, not device based.
503 * The rdev for the device selected will have nr_pending incremented.
505 static int read_balance(struct r1conf
*conf
, struct r1bio
*r1_bio
, int *max_sectors
)
507 const sector_t this_sector
= r1_bio
->sector
;
509 int best_good_sectors
;
510 int best_disk
, best_dist_disk
, best_pending_disk
;
514 unsigned int min_pending
;
515 struct md_rdev
*rdev
;
517 int choose_next_idle
;
521 * Check if we can balance. We can balance on the whole
522 * device if no resync is going on, or below the resync window.
523 * We take the first readable disk when above the resync window.
526 sectors
= r1_bio
->sectors
;
529 best_dist
= MaxSector
;
530 best_pending_disk
= -1;
531 min_pending
= UINT_MAX
;
532 best_good_sectors
= 0;
534 choose_next_idle
= 0;
536 if (conf
->mddev
->recovery_cp
< MaxSector
&&
537 (this_sector
+ sectors
>= conf
->next_resync
))
542 for (disk
= 0 ; disk
< conf
->raid_disks
* 2 ; disk
++) {
546 unsigned int pending
;
549 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
550 if (r1_bio
->bios
[disk
] == IO_BLOCKED
552 || test_bit(Unmerged
, &rdev
->flags
)
553 || test_bit(Faulty
, &rdev
->flags
))
555 if (!test_bit(In_sync
, &rdev
->flags
) &&
556 rdev
->recovery_offset
< this_sector
+ sectors
)
558 if (test_bit(WriteMostly
, &rdev
->flags
)) {
559 /* Don't balance among write-mostly, just
560 * use the first as a last resort */
562 if (is_badblock(rdev
, this_sector
, sectors
,
563 &first_bad
, &bad_sectors
)) {
564 if (first_bad
< this_sector
)
565 /* Cannot use this */
567 best_good_sectors
= first_bad
- this_sector
;
569 best_good_sectors
= sectors
;
574 /* This is a reasonable device to use. It might
577 if (is_badblock(rdev
, this_sector
, sectors
,
578 &first_bad
, &bad_sectors
)) {
579 if (best_dist
< MaxSector
)
580 /* already have a better device */
582 if (first_bad
<= this_sector
) {
583 /* cannot read here. If this is the 'primary'
584 * device, then we must not read beyond
585 * bad_sectors from another device..
587 bad_sectors
-= (this_sector
- first_bad
);
588 if (choose_first
&& sectors
> bad_sectors
)
589 sectors
= bad_sectors
;
590 if (best_good_sectors
> sectors
)
591 best_good_sectors
= sectors
;
594 sector_t good_sectors
= first_bad
- this_sector
;
595 if (good_sectors
> best_good_sectors
) {
596 best_good_sectors
= good_sectors
;
604 best_good_sectors
= sectors
;
606 nonrot
= blk_queue_nonrot(bdev_get_queue(rdev
->bdev
));
607 has_nonrot_disk
|= nonrot
;
608 pending
= atomic_read(&rdev
->nr_pending
);
609 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
614 /* Don't change to another disk for sequential reads */
615 if (conf
->mirrors
[disk
].next_seq_sect
== this_sector
617 int opt_iosize
= bdev_io_opt(rdev
->bdev
) >> 9;
618 struct raid1_info
*mirror
= &conf
->mirrors
[disk
];
622 * If buffered sequential IO size exceeds optimal
623 * iosize, check if there is idle disk. If yes, choose
624 * the idle disk. read_balance could already choose an
625 * idle disk before noticing it's a sequential IO in
626 * this disk. This doesn't matter because this disk
627 * will idle, next time it will be utilized after the
628 * first disk has IO size exceeds optimal iosize. In
629 * this way, iosize of the first disk will be optimal
630 * iosize at least. iosize of the second disk might be
631 * small, but not a big deal since when the second disk
632 * starts IO, the first disk is likely still busy.
634 if (nonrot
&& opt_iosize
> 0 &&
635 mirror
->seq_start
!= MaxSector
&&
636 mirror
->next_seq_sect
> opt_iosize
&&
637 mirror
->next_seq_sect
- opt_iosize
>=
639 choose_next_idle
= 1;
644 /* If device is idle, use it */
650 if (choose_next_idle
)
653 if (min_pending
> pending
) {
654 min_pending
= pending
;
655 best_pending_disk
= disk
;
658 if (dist
< best_dist
) {
660 best_dist_disk
= disk
;
665 * If all disks are rotational, choose the closest disk. If any disk is
666 * non-rotational, choose the disk with less pending request even the
667 * disk is rotational, which might/might not be optimal for raids with
668 * mixed ratation/non-rotational disks depending on workload.
670 if (best_disk
== -1) {
672 best_disk
= best_pending_disk
;
674 best_disk
= best_dist_disk
;
677 if (best_disk
>= 0) {
678 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
681 atomic_inc(&rdev
->nr_pending
);
682 if (test_bit(Faulty
, &rdev
->flags
)) {
683 /* cannot risk returning a device that failed
684 * before we inc'ed nr_pending
686 rdev_dec_pending(rdev
, conf
->mddev
);
689 sectors
= best_good_sectors
;
691 if (conf
->mirrors
[best_disk
].next_seq_sect
!= this_sector
)
692 conf
->mirrors
[best_disk
].seq_start
= this_sector
;
694 conf
->mirrors
[best_disk
].next_seq_sect
= this_sector
+ sectors
;
697 *max_sectors
= sectors
;
702 static int raid1_mergeable_bvec(struct request_queue
*q
,
703 struct bvec_merge_data
*bvm
,
704 struct bio_vec
*biovec
)
706 struct mddev
*mddev
= q
->queuedata
;
707 struct r1conf
*conf
= mddev
->private;
708 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
709 int max
= biovec
->bv_len
;
711 if (mddev
->merge_check_needed
) {
714 for (disk
= 0; disk
< conf
->raid_disks
* 2; disk
++) {
715 struct md_rdev
*rdev
= rcu_dereference(
716 conf
->mirrors
[disk
].rdev
);
717 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
718 struct request_queue
*q
=
719 bdev_get_queue(rdev
->bdev
);
720 if (q
->merge_bvec_fn
) {
721 bvm
->bi_sector
= sector
+
723 bvm
->bi_bdev
= rdev
->bdev
;
724 max
= min(max
, q
->merge_bvec_fn(
735 int md_raid1_congested(struct mddev
*mddev
, int bits
)
737 struct r1conf
*conf
= mddev
->private;
740 if ((bits
& (1 << BDI_async_congested
)) &&
741 conf
->pending_count
>= max_queued_requests
)
745 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
746 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
747 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
748 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
752 /* Note the '|| 1' - when read_balance prefers
753 * non-congested targets, it can be removed
755 if ((bits
& (1<<BDI_async_congested
)) || 1)
756 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
758 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
764 EXPORT_SYMBOL_GPL(md_raid1_congested
);
766 static int raid1_congested(void *data
, int bits
)
768 struct mddev
*mddev
= data
;
770 return mddev_congested(mddev
, bits
) ||
771 md_raid1_congested(mddev
, bits
);
774 static void flush_pending_writes(struct r1conf
*conf
)
776 /* Any writes that have been queued but are awaiting
777 * bitmap updates get flushed here.
779 spin_lock_irq(&conf
->device_lock
);
781 if (conf
->pending_bio_list
.head
) {
783 bio
= bio_list_get(&conf
->pending_bio_list
);
784 conf
->pending_count
= 0;
785 spin_unlock_irq(&conf
->device_lock
);
786 /* flush any pending bitmap writes to
787 * disk before proceeding w/ I/O */
788 bitmap_unplug(conf
->mddev
->bitmap
);
789 wake_up(&conf
->wait_barrier
);
791 while (bio
) { /* submit pending writes */
792 struct bio
*next
= bio
->bi_next
;
794 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
795 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
799 generic_make_request(bio
);
803 spin_unlock_irq(&conf
->device_lock
);
807 * Sometimes we need to suspend IO while we do something else,
808 * either some resync/recovery, or reconfigure the array.
809 * To do this we raise a 'barrier'.
810 * The 'barrier' is a counter that can be raised multiple times
811 * to count how many activities are happening which preclude
813 * We can only raise the barrier if there is no pending IO.
814 * i.e. if nr_pending == 0.
815 * We choose only to raise the barrier if no-one is waiting for the
816 * barrier to go down. This means that as soon as an IO request
817 * is ready, no other operations which require a barrier will start
818 * until the IO request has had a chance.
820 * So: regular IO calls 'wait_barrier'. When that returns there
821 * is no backgroup IO happening, It must arrange to call
822 * allow_barrier when it has finished its IO.
823 * backgroup IO calls must call raise_barrier. Once that returns
824 * there is no normal IO happeing. It must arrange to call
825 * lower_barrier when the particular background IO completes.
827 static void raise_barrier(struct r1conf
*conf
)
829 spin_lock_irq(&conf
->resync_lock
);
831 /* Wait until no block IO is waiting */
832 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
835 /* block any new IO from starting */
838 /* For these conditions we must wait:
839 * A: while the array is in frozen state
840 * B: while barrier >= RESYNC_DEPTH, meaning resync reach
841 * the max count which allowed.
842 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
843 * next resync will reach to the window which normal bios are
846 wait_event_lock_irq(conf
->wait_barrier
,
847 !conf
->array_frozen
&&
848 conf
->barrier
< RESYNC_DEPTH
&&
849 (conf
->start_next_window
>=
850 conf
->next_resync
+ RESYNC_SECTORS
),
853 spin_unlock_irq(&conf
->resync_lock
);
856 static void lower_barrier(struct r1conf
*conf
)
859 BUG_ON(conf
->barrier
<= 0);
860 spin_lock_irqsave(&conf
->resync_lock
, flags
);
862 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
863 wake_up(&conf
->wait_barrier
);
866 static bool need_to_wait_for_sync(struct r1conf
*conf
, struct bio
*bio
)
870 if (conf
->array_frozen
|| !bio
)
872 else if (conf
->barrier
&& bio_data_dir(bio
) == WRITE
) {
873 if (conf
->next_resync
< RESYNC_WINDOW_SECTORS
)
875 else if ((conf
->next_resync
- RESYNC_WINDOW_SECTORS
876 >= bio_end_sector(bio
)) ||
877 (conf
->next_resync
+ NEXT_NORMALIO_DISTANCE
887 static sector_t
wait_barrier(struct r1conf
*conf
, struct bio
*bio
)
891 spin_lock_irq(&conf
->resync_lock
);
892 if (need_to_wait_for_sync(conf
, bio
)) {
894 /* Wait for the barrier to drop.
895 * However if there are already pending
896 * requests (preventing the barrier from
897 * rising completely), and the
898 * pre-process bio queue isn't empty,
899 * then don't wait, as we need to empty
900 * that queue to get the nr_pending
903 wait_event_lock_irq(conf
->wait_barrier
,
904 !conf
->array_frozen
&&
906 ((conf
->start_next_window
<
907 conf
->next_resync
+ RESYNC_SECTORS
) &&
909 !bio_list_empty(current
->bio_list
))),
914 if (bio
&& bio_data_dir(bio
) == WRITE
) {
915 if (conf
->next_resync
+ NEXT_NORMALIO_DISTANCE
917 if (conf
->start_next_window
== MaxSector
)
918 conf
->start_next_window
=
920 NEXT_NORMALIO_DISTANCE
;
922 if ((conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
)
924 conf
->next_window_requests
++;
926 conf
->current_window_requests
++;
928 if (bio
->bi_sector
>= conf
->start_next_window
)
929 sector
= conf
->start_next_window
;
933 spin_unlock_irq(&conf
->resync_lock
);
937 static void allow_barrier(struct r1conf
*conf
, sector_t start_next_window
,
942 spin_lock_irqsave(&conf
->resync_lock
, flags
);
944 if (start_next_window
) {
945 if (start_next_window
== conf
->start_next_window
) {
946 if (conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
948 conf
->next_window_requests
--;
950 conf
->current_window_requests
--;
952 conf
->current_window_requests
--;
954 if (!conf
->current_window_requests
) {
955 if (conf
->next_window_requests
) {
956 conf
->current_window_requests
=
957 conf
->next_window_requests
;
958 conf
->next_window_requests
= 0;
959 conf
->start_next_window
+=
960 NEXT_NORMALIO_DISTANCE
;
962 conf
->start_next_window
= MaxSector
;
965 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
966 wake_up(&conf
->wait_barrier
);
969 static void freeze_array(struct r1conf
*conf
, int extra
)
971 /* stop syncio and normal IO and wait for everything to
973 * We wait until nr_pending match nr_queued+extra
974 * This is called in the context of one normal IO request
975 * that has failed. Thus any sync request that might be pending
976 * will be blocked by nr_pending, and we need to wait for
977 * pending IO requests to complete or be queued for re-try.
978 * Thus the number queued (nr_queued) plus this request (extra)
979 * must match the number of pending IOs (nr_pending) before
982 spin_lock_irq(&conf
->resync_lock
);
983 conf
->array_frozen
= 1;
984 wait_event_lock_irq_cmd(conf
->wait_barrier
,
985 conf
->nr_pending
== conf
->nr_queued
+extra
,
987 flush_pending_writes(conf
));
988 spin_unlock_irq(&conf
->resync_lock
);
990 static void unfreeze_array(struct r1conf
*conf
)
992 /* reverse the effect of the freeze */
993 spin_lock_irq(&conf
->resync_lock
);
994 conf
->array_frozen
= 0;
995 wake_up(&conf
->wait_barrier
);
996 spin_unlock_irq(&conf
->resync_lock
);
1000 /* duplicate the data pages for behind I/O
1002 static void alloc_behind_pages(struct bio
*bio
, struct r1bio
*r1_bio
)
1005 struct bio_vec
*bvec
;
1006 struct bio_vec
*bvecs
= kzalloc(bio
->bi_vcnt
* sizeof(struct bio_vec
),
1008 if (unlikely(!bvecs
))
1011 bio_for_each_segment_all(bvec
, bio
, i
) {
1013 bvecs
[i
].bv_page
= alloc_page(GFP_NOIO
);
1014 if (unlikely(!bvecs
[i
].bv_page
))
1016 memcpy(kmap(bvecs
[i
].bv_page
) + bvec
->bv_offset
,
1017 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
1018 kunmap(bvecs
[i
].bv_page
);
1019 kunmap(bvec
->bv_page
);
1021 r1_bio
->behind_bvecs
= bvecs
;
1022 r1_bio
->behind_page_count
= bio
->bi_vcnt
;
1023 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
1027 for (i
= 0; i
< bio
->bi_vcnt
; i
++)
1028 if (bvecs
[i
].bv_page
)
1029 put_page(bvecs
[i
].bv_page
);
1031 pr_debug("%dB behind alloc failed, doing sync I/O\n", bio
->bi_size
);
1034 struct raid1_plug_cb
{
1035 struct blk_plug_cb cb
;
1036 struct bio_list pending
;
1040 static void raid1_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1042 struct raid1_plug_cb
*plug
= container_of(cb
, struct raid1_plug_cb
,
1044 struct mddev
*mddev
= plug
->cb
.data
;
1045 struct r1conf
*conf
= mddev
->private;
1048 if (from_schedule
|| current
->bio_list
) {
1049 spin_lock_irq(&conf
->device_lock
);
1050 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1051 conf
->pending_count
+= plug
->pending_cnt
;
1052 spin_unlock_irq(&conf
->device_lock
);
1053 wake_up(&conf
->wait_barrier
);
1054 md_wakeup_thread(mddev
->thread
);
1059 /* we aren't scheduling, so we can do the write-out directly. */
1060 bio
= bio_list_get(&plug
->pending
);
1061 bitmap_unplug(mddev
->bitmap
);
1062 wake_up(&conf
->wait_barrier
);
1064 while (bio
) { /* submit pending writes */
1065 struct bio
*next
= bio
->bi_next
;
1066 bio
->bi_next
= NULL
;
1067 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
1068 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1069 /* Just ignore it */
1072 generic_make_request(bio
);
1078 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1080 struct r1conf
*conf
= mddev
->private;
1081 struct raid1_info
*mirror
;
1082 struct r1bio
*r1_bio
;
1083 struct bio
*read_bio
;
1085 struct bitmap
*bitmap
;
1086 unsigned long flags
;
1087 const int rw
= bio_data_dir(bio
);
1088 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1089 const unsigned long do_flush_fua
= (bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
));
1090 const unsigned long do_discard
= (bio
->bi_rw
1091 & (REQ_DISCARD
| REQ_SECURE
));
1092 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1093 struct md_rdev
*blocked_rdev
;
1094 struct blk_plug_cb
*cb
;
1095 struct raid1_plug_cb
*plug
= NULL
;
1097 int sectors_handled
;
1099 sector_t start_next_window
;
1102 * Register the new request and wait if the reconstruction
1103 * thread has put up a bar for new requests.
1104 * Continue immediately if no resync is active currently.
1107 md_write_start(mddev
, bio
); /* wait on superblock update early */
1109 if (bio_data_dir(bio
) == WRITE
&&
1110 bio_end_sector(bio
) > mddev
->suspend_lo
&&
1111 bio
->bi_sector
< mddev
->suspend_hi
) {
1112 /* As the suspend_* range is controlled by
1113 * userspace, we want an interruptible
1118 flush_signals(current
);
1119 prepare_to_wait(&conf
->wait_barrier
,
1120 &w
, TASK_INTERRUPTIBLE
);
1121 if (bio_end_sector(bio
) <= mddev
->suspend_lo
||
1122 bio
->bi_sector
>= mddev
->suspend_hi
)
1126 finish_wait(&conf
->wait_barrier
, &w
);
1129 start_next_window
= wait_barrier(conf
, bio
);
1131 bitmap
= mddev
->bitmap
;
1134 * make_request() can abort the operation when READA is being
1135 * used and no empty request is available.
1138 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1140 r1_bio
->master_bio
= bio
;
1141 r1_bio
->sectors
= bio_sectors(bio
);
1143 r1_bio
->mddev
= mddev
;
1144 r1_bio
->sector
= bio
->bi_sector
;
1146 /* We might need to issue multiple reads to different
1147 * devices if there are bad blocks around, so we keep
1148 * track of the number of reads in bio->bi_phys_segments.
1149 * If this is 0, there is only one r1_bio and no locking
1150 * will be needed when requests complete. If it is
1151 * non-zero, then it is the number of not-completed requests.
1153 bio
->bi_phys_segments
= 0;
1154 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1158 * read balancing logic:
1163 rdisk
= read_balance(conf
, r1_bio
, &max_sectors
);
1166 /* couldn't find anywhere to read from */
1167 raid_end_bio_io(r1_bio
);
1170 mirror
= conf
->mirrors
+ rdisk
;
1172 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
1174 /* Reading from a write-mostly device must
1175 * take care not to over-take any writes
1178 wait_event(bitmap
->behind_wait
,
1179 atomic_read(&bitmap
->behind_writes
) == 0);
1181 r1_bio
->read_disk
= rdisk
;
1183 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1184 bio_trim(read_bio
, r1_bio
->sector
- bio
->bi_sector
,
1187 r1_bio
->bios
[rdisk
] = read_bio
;
1189 read_bio
->bi_sector
= r1_bio
->sector
+ mirror
->rdev
->data_offset
;
1190 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
1191 read_bio
->bi_end_io
= raid1_end_read_request
;
1192 read_bio
->bi_rw
= READ
| do_sync
;
1193 read_bio
->bi_private
= r1_bio
;
1195 if (max_sectors
< r1_bio
->sectors
) {
1196 /* could not read all from this device, so we will
1197 * need another r1_bio.
1200 sectors_handled
= (r1_bio
->sector
+ max_sectors
1202 r1_bio
->sectors
= max_sectors
;
1203 spin_lock_irq(&conf
->device_lock
);
1204 if (bio
->bi_phys_segments
== 0)
1205 bio
->bi_phys_segments
= 2;
1207 bio
->bi_phys_segments
++;
1208 spin_unlock_irq(&conf
->device_lock
);
1209 /* Cannot call generic_make_request directly
1210 * as that will be queued in __make_request
1211 * and subsequent mempool_alloc might block waiting
1212 * for it. So hand bio over to raid1d.
1214 reschedule_retry(r1_bio
);
1216 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1218 r1_bio
->master_bio
= bio
;
1219 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1221 r1_bio
->mddev
= mddev
;
1222 r1_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1225 generic_make_request(read_bio
);
1232 if (conf
->pending_count
>= max_queued_requests
) {
1233 md_wakeup_thread(mddev
->thread
);
1234 wait_event(conf
->wait_barrier
,
1235 conf
->pending_count
< max_queued_requests
);
1237 /* first select target devices under rcu_lock and
1238 * inc refcount on their rdev. Record them by setting
1240 * If there are known/acknowledged bad blocks on any device on
1241 * which we have seen a write error, we want to avoid writing those
1243 * This potentially requires several writes to write around
1244 * the bad blocks. Each set of writes gets it's own r1bio
1245 * with a set of bios attached.
1248 disks
= conf
->raid_disks
* 2;
1250 r1_bio
->start_next_window
= start_next_window
;
1251 blocked_rdev
= NULL
;
1253 max_sectors
= r1_bio
->sectors
;
1254 for (i
= 0; i
< disks
; i
++) {
1255 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1256 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1257 atomic_inc(&rdev
->nr_pending
);
1258 blocked_rdev
= rdev
;
1261 r1_bio
->bios
[i
] = NULL
;
1262 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)
1263 || test_bit(Unmerged
, &rdev
->flags
)) {
1264 if (i
< conf
->raid_disks
)
1265 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1269 atomic_inc(&rdev
->nr_pending
);
1270 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1275 is_bad
= is_badblock(rdev
, r1_bio
->sector
,
1277 &first_bad
, &bad_sectors
);
1279 /* mustn't write here until the bad block is
1281 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1282 blocked_rdev
= rdev
;
1285 if (is_bad
&& first_bad
<= r1_bio
->sector
) {
1286 /* Cannot write here at all */
1287 bad_sectors
-= (r1_bio
->sector
- first_bad
);
1288 if (bad_sectors
< max_sectors
)
1289 /* mustn't write more than bad_sectors
1290 * to other devices yet
1292 max_sectors
= bad_sectors
;
1293 rdev_dec_pending(rdev
, mddev
);
1294 /* We don't set R1BIO_Degraded as that
1295 * only applies if the disk is
1296 * missing, so it might be re-added,
1297 * and we want to know to recover this
1299 * In this case the device is here,
1300 * and the fact that this chunk is not
1301 * in-sync is recorded in the bad
1307 int good_sectors
= first_bad
- r1_bio
->sector
;
1308 if (good_sectors
< max_sectors
)
1309 max_sectors
= good_sectors
;
1312 r1_bio
->bios
[i
] = bio
;
1316 if (unlikely(blocked_rdev
)) {
1317 /* Wait for this device to become unblocked */
1319 sector_t old
= start_next_window
;
1321 for (j
= 0; j
< i
; j
++)
1322 if (r1_bio
->bios
[j
])
1323 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
1325 allow_barrier(conf
, start_next_window
, bio
->bi_sector
);
1326 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1327 start_next_window
= wait_barrier(conf
, bio
);
1329 * We must make sure the multi r1bios of bio have
1330 * the same value of bi_phys_segments
1332 if (bio
->bi_phys_segments
&& old
&&
1333 old
!= start_next_window
)
1334 /* Wait for the former r1bio(s) to complete */
1335 wait_event(conf
->wait_barrier
,
1336 bio
->bi_phys_segments
== 1);
1340 if (max_sectors
< r1_bio
->sectors
) {
1341 /* We are splitting this write into multiple parts, so
1342 * we need to prepare for allocating another r1_bio.
1344 r1_bio
->sectors
= max_sectors
;
1345 spin_lock_irq(&conf
->device_lock
);
1346 if (bio
->bi_phys_segments
== 0)
1347 bio
->bi_phys_segments
= 2;
1349 bio
->bi_phys_segments
++;
1350 spin_unlock_irq(&conf
->device_lock
);
1352 sectors_handled
= r1_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1354 atomic_set(&r1_bio
->remaining
, 1);
1355 atomic_set(&r1_bio
->behind_remaining
, 0);
1358 for (i
= 0; i
< disks
; i
++) {
1360 if (!r1_bio
->bios
[i
])
1363 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1364 bio_trim(mbio
, r1_bio
->sector
- bio
->bi_sector
, max_sectors
);
1368 * Not if there are too many, or cannot
1369 * allocate memory, or a reader on WriteMostly
1370 * is waiting for behind writes to flush */
1372 (atomic_read(&bitmap
->behind_writes
)
1373 < mddev
->bitmap_info
.max_write_behind
) &&
1374 !waitqueue_active(&bitmap
->behind_wait
))
1375 alloc_behind_pages(mbio
, r1_bio
);
1377 bitmap_startwrite(bitmap
, r1_bio
->sector
,
1379 test_bit(R1BIO_BehindIO
,
1383 if (r1_bio
->behind_bvecs
) {
1384 struct bio_vec
*bvec
;
1388 * We trimmed the bio, so _all is legit
1390 bio_for_each_segment_all(bvec
, mbio
, j
)
1391 bvec
->bv_page
= r1_bio
->behind_bvecs
[j
].bv_page
;
1392 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
1393 atomic_inc(&r1_bio
->behind_remaining
);
1396 r1_bio
->bios
[i
] = mbio
;
1398 mbio
->bi_sector
= (r1_bio
->sector
+
1399 conf
->mirrors
[i
].rdev
->data_offset
);
1400 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1401 mbio
->bi_end_io
= raid1_end_write_request
;
1403 WRITE
| do_flush_fua
| do_sync
| do_discard
| do_same
;
1404 mbio
->bi_private
= r1_bio
;
1406 atomic_inc(&r1_bio
->remaining
);
1408 cb
= blk_check_plugged(raid1_unplug
, mddev
, sizeof(*plug
));
1410 plug
= container_of(cb
, struct raid1_plug_cb
, cb
);
1413 spin_lock_irqsave(&conf
->device_lock
, flags
);
1415 bio_list_add(&plug
->pending
, mbio
);
1416 plug
->pending_cnt
++;
1418 bio_list_add(&conf
->pending_bio_list
, mbio
);
1419 conf
->pending_count
++;
1421 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1423 md_wakeup_thread(mddev
->thread
);
1425 /* Mustn't call r1_bio_write_done before this next test,
1426 * as it could result in the bio being freed.
1428 if (sectors_handled
< bio_sectors(bio
)) {
1429 r1_bio_write_done(r1_bio
);
1430 /* We need another r1_bio. It has already been counted
1431 * in bio->bi_phys_segments
1433 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1434 r1_bio
->master_bio
= bio
;
1435 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1437 r1_bio
->mddev
= mddev
;
1438 r1_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1442 r1_bio_write_done(r1_bio
);
1444 /* In case raid1d snuck in to freeze_array */
1445 wake_up(&conf
->wait_barrier
);
1448 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1450 struct r1conf
*conf
= mddev
->private;
1453 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1454 conf
->raid_disks
- mddev
->degraded
);
1456 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1457 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1458 seq_printf(seq
, "%s",
1459 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1462 seq_printf(seq
, "]");
1466 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1468 char b
[BDEVNAME_SIZE
];
1469 struct r1conf
*conf
= mddev
->private;
1472 * If it is not operational, then we have already marked it as dead
1473 * else if it is the last working disks, ignore the error, let the
1474 * next level up know.
1475 * else mark the drive as failed
1477 if (test_bit(In_sync
, &rdev
->flags
)
1478 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1480 * Don't fail the drive, act as though we were just a
1481 * normal single drive.
1482 * However don't try a recovery from this drive as
1483 * it is very likely to fail.
1485 conf
->recovery_disabled
= mddev
->recovery_disabled
;
1488 set_bit(Blocked
, &rdev
->flags
);
1489 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1490 unsigned long flags
;
1491 spin_lock_irqsave(&conf
->device_lock
, flags
);
1493 set_bit(Faulty
, &rdev
->flags
);
1494 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1496 * if recovery is running, make sure it aborts.
1498 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1500 set_bit(Faulty
, &rdev
->flags
);
1501 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1503 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1504 "md/raid1:%s: Operation continuing on %d devices.\n",
1505 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1506 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1509 static void print_conf(struct r1conf
*conf
)
1513 printk(KERN_DEBUG
"RAID1 conf printout:\n");
1515 printk(KERN_DEBUG
"(!conf)\n");
1518 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1522 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1523 char b
[BDEVNAME_SIZE
];
1524 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1526 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1527 i
, !test_bit(In_sync
, &rdev
->flags
),
1528 !test_bit(Faulty
, &rdev
->flags
),
1529 bdevname(rdev
->bdev
,b
));
1534 static void close_sync(struct r1conf
*conf
)
1536 wait_barrier(conf
, NULL
);
1537 allow_barrier(conf
, 0, 0);
1539 mempool_destroy(conf
->r1buf_pool
);
1540 conf
->r1buf_pool
= NULL
;
1542 conf
->next_resync
= 0;
1543 conf
->start_next_window
= MaxSector
;
1546 static int raid1_spare_active(struct mddev
*mddev
)
1549 struct r1conf
*conf
= mddev
->private;
1551 unsigned long flags
;
1554 * Find all failed disks within the RAID1 configuration
1555 * and mark them readable.
1556 * Called under mddev lock, so rcu protection not needed.
1558 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1559 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1560 struct md_rdev
*repl
= conf
->mirrors
[conf
->raid_disks
+ i
].rdev
;
1562 && repl
->recovery_offset
== MaxSector
1563 && !test_bit(Faulty
, &repl
->flags
)
1564 && !test_and_set_bit(In_sync
, &repl
->flags
)) {
1565 /* replacement has just become active */
1567 !test_and_clear_bit(In_sync
, &rdev
->flags
))
1570 /* Replaced device not technically
1571 * faulty, but we need to be sure
1572 * it gets removed and never re-added
1574 set_bit(Faulty
, &rdev
->flags
);
1575 sysfs_notify_dirent_safe(
1580 && rdev
->recovery_offset
== MaxSector
1581 && !test_bit(Faulty
, &rdev
->flags
)
1582 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1584 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
1587 spin_lock_irqsave(&conf
->device_lock
, flags
);
1588 mddev
->degraded
-= count
;
1589 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1596 static int raid1_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1598 struct r1conf
*conf
= mddev
->private;
1601 struct raid1_info
*p
;
1603 int last
= conf
->raid_disks
- 1;
1604 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1606 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
1609 if (rdev
->raid_disk
>= 0)
1610 first
= last
= rdev
->raid_disk
;
1612 if (q
->merge_bvec_fn
) {
1613 set_bit(Unmerged
, &rdev
->flags
);
1614 mddev
->merge_check_needed
= 1;
1617 for (mirror
= first
; mirror
<= last
; mirror
++) {
1618 p
= conf
->mirrors
+mirror
;
1622 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1623 rdev
->data_offset
<< 9);
1625 p
->head_position
= 0;
1626 rdev
->raid_disk
= mirror
;
1628 /* As all devices are equivalent, we don't need a full recovery
1629 * if this was recently any drive of the array
1631 if (rdev
->saved_raid_disk
< 0)
1633 rcu_assign_pointer(p
->rdev
, rdev
);
1636 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
1637 p
[conf
->raid_disks
].rdev
== NULL
) {
1638 /* Add this device as a replacement */
1639 clear_bit(In_sync
, &rdev
->flags
);
1640 set_bit(Replacement
, &rdev
->flags
);
1641 rdev
->raid_disk
= mirror
;
1644 rcu_assign_pointer(p
[conf
->raid_disks
].rdev
, rdev
);
1648 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1649 /* Some requests might not have seen this new
1650 * merge_bvec_fn. We must wait for them to complete
1651 * before merging the device fully.
1652 * First we make sure any code which has tested
1653 * our function has submitted the request, then
1654 * we wait for all outstanding requests to complete.
1656 synchronize_sched();
1657 freeze_array(conf
, 0);
1658 unfreeze_array(conf
);
1659 clear_bit(Unmerged
, &rdev
->flags
);
1661 md_integrity_add_rdev(rdev
, mddev
);
1662 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1663 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1668 static int raid1_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1670 struct r1conf
*conf
= mddev
->private;
1672 int number
= rdev
->raid_disk
;
1673 struct raid1_info
*p
= conf
->mirrors
+ number
;
1675 if (rdev
!= p
->rdev
)
1676 p
= conf
->mirrors
+ conf
->raid_disks
+ number
;
1679 if (rdev
== p
->rdev
) {
1680 if (test_bit(In_sync
, &rdev
->flags
) ||
1681 atomic_read(&rdev
->nr_pending
)) {
1685 /* Only remove non-faulty devices if recovery
1688 if (!test_bit(Faulty
, &rdev
->flags
) &&
1689 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
1690 mddev
->degraded
< conf
->raid_disks
) {
1696 if (atomic_read(&rdev
->nr_pending
)) {
1697 /* lost the race, try later */
1701 } else if (conf
->mirrors
[conf
->raid_disks
+ number
].rdev
) {
1702 /* We just removed a device that is being replaced.
1703 * Move down the replacement. We drain all IO before
1704 * doing this to avoid confusion.
1706 struct md_rdev
*repl
=
1707 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
;
1708 freeze_array(conf
, 0);
1709 clear_bit(Replacement
, &repl
->flags
);
1711 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
= NULL
;
1712 unfreeze_array(conf
);
1713 clear_bit(WantReplacement
, &rdev
->flags
);
1715 clear_bit(WantReplacement
, &rdev
->flags
);
1716 err
= md_integrity_register(mddev
);
1725 static void end_sync_read(struct bio
*bio
, int error
)
1727 struct r1bio
*r1_bio
= bio
->bi_private
;
1729 update_head_pos(r1_bio
->read_disk
, r1_bio
);
1732 * we have read a block, now it needs to be re-written,
1733 * or re-read if the read failed.
1734 * We don't do much here, just schedule handling by raid1d
1736 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1737 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1739 if (atomic_dec_and_test(&r1_bio
->remaining
))
1740 reschedule_retry(r1_bio
);
1743 static void end_sync_write(struct bio
*bio
, int error
)
1745 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1746 struct r1bio
*r1_bio
= bio
->bi_private
;
1747 struct mddev
*mddev
= r1_bio
->mddev
;
1748 struct r1conf
*conf
= mddev
->private;
1753 mirror
= find_bio_disk(r1_bio
, bio
);
1756 sector_t sync_blocks
= 0;
1757 sector_t s
= r1_bio
->sector
;
1758 long sectors_to_go
= r1_bio
->sectors
;
1759 /* make sure these bits doesn't get cleared. */
1761 bitmap_end_sync(mddev
->bitmap
, s
,
1764 sectors_to_go
-= sync_blocks
;
1765 } while (sectors_to_go
> 0);
1766 set_bit(WriteErrorSeen
,
1767 &conf
->mirrors
[mirror
].rdev
->flags
);
1768 if (!test_and_set_bit(WantReplacement
,
1769 &conf
->mirrors
[mirror
].rdev
->flags
))
1770 set_bit(MD_RECOVERY_NEEDED
, &
1772 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
1773 } else if (is_badblock(conf
->mirrors
[mirror
].rdev
,
1776 &first_bad
, &bad_sectors
) &&
1777 !is_badblock(conf
->mirrors
[r1_bio
->read_disk
].rdev
,
1780 &first_bad
, &bad_sectors
)
1782 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
1784 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1785 int s
= r1_bio
->sectors
;
1786 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
1787 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1788 reschedule_retry(r1_bio
);
1791 md_done_sync(mddev
, s
, uptodate
);
1796 static int r1_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1797 int sectors
, struct page
*page
, int rw
)
1799 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
1803 set_bit(WriteErrorSeen
, &rdev
->flags
);
1804 if (!test_and_set_bit(WantReplacement
,
1806 set_bit(MD_RECOVERY_NEEDED
, &
1807 rdev
->mddev
->recovery
);
1809 /* need to record an error - either for the block or the device */
1810 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1811 md_error(rdev
->mddev
, rdev
);
1815 static int fix_sync_read_error(struct r1bio
*r1_bio
)
1817 /* Try some synchronous reads of other devices to get
1818 * good data, much like with normal read errors. Only
1819 * read into the pages we already have so we don't
1820 * need to re-issue the read request.
1821 * We don't need to freeze the array, because being in an
1822 * active sync request, there is no normal IO, and
1823 * no overlapping syncs.
1824 * We don't need to check is_badblock() again as we
1825 * made sure that anything with a bad block in range
1826 * will have bi_end_io clear.
1828 struct mddev
*mddev
= r1_bio
->mddev
;
1829 struct r1conf
*conf
= mddev
->private;
1830 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1831 sector_t sect
= r1_bio
->sector
;
1832 int sectors
= r1_bio
->sectors
;
1837 int d
= r1_bio
->read_disk
;
1839 struct md_rdev
*rdev
;
1842 if (s
> (PAGE_SIZE
>>9))
1845 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1846 /* No rcu protection needed here devices
1847 * can only be removed when no resync is
1848 * active, and resync is currently active
1850 rdev
= conf
->mirrors
[d
].rdev
;
1851 if (sync_page_io(rdev
, sect
, s
<<9,
1852 bio
->bi_io_vec
[idx
].bv_page
,
1859 if (d
== conf
->raid_disks
* 2)
1861 } while (!success
&& d
!= r1_bio
->read_disk
);
1864 char b
[BDEVNAME_SIZE
];
1866 /* Cannot read from anywhere, this block is lost.
1867 * Record a bad block on each device. If that doesn't
1868 * work just disable and interrupt the recovery.
1869 * Don't fail devices as that won't really help.
1871 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O read error"
1872 " for block %llu\n",
1874 bdevname(bio
->bi_bdev
, b
),
1875 (unsigned long long)r1_bio
->sector
);
1876 for (d
= 0; d
< conf
->raid_disks
* 2; d
++) {
1877 rdev
= conf
->mirrors
[d
].rdev
;
1878 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
1880 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
1884 conf
->recovery_disabled
=
1885 mddev
->recovery_disabled
;
1886 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1887 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1899 /* write it back and re-read */
1900 while (d
!= r1_bio
->read_disk
) {
1902 d
= conf
->raid_disks
* 2;
1904 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1906 rdev
= conf
->mirrors
[d
].rdev
;
1907 if (r1_sync_page_io(rdev
, sect
, s
,
1908 bio
->bi_io_vec
[idx
].bv_page
,
1910 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
1911 rdev_dec_pending(rdev
, mddev
);
1915 while (d
!= r1_bio
->read_disk
) {
1917 d
= conf
->raid_disks
* 2;
1919 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1921 rdev
= conf
->mirrors
[d
].rdev
;
1922 if (r1_sync_page_io(rdev
, sect
, s
,
1923 bio
->bi_io_vec
[idx
].bv_page
,
1925 atomic_add(s
, &rdev
->corrected_errors
);
1931 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1932 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1936 static int process_checks(struct r1bio
*r1_bio
)
1938 /* We have read all readable devices. If we haven't
1939 * got the block, then there is no hope left.
1940 * If we have, then we want to do a comparison
1941 * and skip the write if everything is the same.
1942 * If any blocks failed to read, then we need to
1943 * attempt an over-write
1945 struct mddev
*mddev
= r1_bio
->mddev
;
1946 struct r1conf
*conf
= mddev
->private;
1951 /* Fix variable parts of all bios */
1952 vcnt
= (r1_bio
->sectors
+ PAGE_SIZE
/ 512 - 1) >> (PAGE_SHIFT
- 9);
1953 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
1956 struct bio
*b
= r1_bio
->bios
[i
];
1957 if (b
->bi_end_io
!= end_sync_read
)
1959 /* fixup the bio for reuse */
1962 b
->bi_size
= r1_bio
->sectors
<< 9;
1963 b
->bi_sector
= r1_bio
->sector
+
1964 conf
->mirrors
[i
].rdev
->data_offset
;
1965 b
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1966 b
->bi_end_io
= end_sync_read
;
1967 b
->bi_private
= r1_bio
;
1970 for (j
= 0; j
< vcnt
; j
++) {
1972 bi
= &b
->bi_io_vec
[j
];
1974 if (size
> PAGE_SIZE
)
1975 bi
->bv_len
= PAGE_SIZE
;
1981 for (primary
= 0; primary
< conf
->raid_disks
* 2; primary
++)
1982 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
1983 test_bit(BIO_UPTODATE
, &r1_bio
->bios
[primary
]->bi_flags
)) {
1984 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
1985 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
1988 r1_bio
->read_disk
= primary
;
1989 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
1991 struct bio
*pbio
= r1_bio
->bios
[primary
];
1992 struct bio
*sbio
= r1_bio
->bios
[i
];
1994 if (sbio
->bi_end_io
!= end_sync_read
)
1997 if (test_bit(BIO_UPTODATE
, &sbio
->bi_flags
)) {
1998 for (j
= vcnt
; j
-- ; ) {
2000 p
= pbio
->bi_io_vec
[j
].bv_page
;
2001 s
= sbio
->bi_io_vec
[j
].bv_page
;
2002 if (memcmp(page_address(p
),
2004 sbio
->bi_io_vec
[j
].bv_len
))
2010 atomic64_add(r1_bio
->sectors
, &mddev
->resync_mismatches
);
2011 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
2012 && test_bit(BIO_UPTODATE
, &sbio
->bi_flags
))) {
2013 /* No need to write to this device. */
2014 sbio
->bi_end_io
= NULL
;
2015 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
2019 bio_copy_data(sbio
, pbio
);
2024 static void sync_request_write(struct mddev
*mddev
, struct r1bio
*r1_bio
)
2026 struct r1conf
*conf
= mddev
->private;
2028 int disks
= conf
->raid_disks
* 2;
2029 struct bio
*bio
, *wbio
;
2031 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2033 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
2034 /* ouch - failed to read all of that. */
2035 if (!fix_sync_read_error(r1_bio
))
2038 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2039 if (process_checks(r1_bio
) < 0)
2044 atomic_set(&r1_bio
->remaining
, 1);
2045 for (i
= 0; i
< disks
; i
++) {
2046 wbio
= r1_bio
->bios
[i
];
2047 if (wbio
->bi_end_io
== NULL
||
2048 (wbio
->bi_end_io
== end_sync_read
&&
2049 (i
== r1_bio
->read_disk
||
2050 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
2053 wbio
->bi_rw
= WRITE
;
2054 wbio
->bi_end_io
= end_sync_write
;
2055 atomic_inc(&r1_bio
->remaining
);
2056 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, bio_sectors(wbio
));
2058 generic_make_request(wbio
);
2061 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
2062 /* if we're here, all write(s) have completed, so clean up */
2063 int s
= r1_bio
->sectors
;
2064 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2065 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2066 reschedule_retry(r1_bio
);
2069 md_done_sync(mddev
, s
, 1);
2075 * This is a kernel thread which:
2077 * 1. Retries failed read operations on working mirrors.
2078 * 2. Updates the raid superblock when problems encounter.
2079 * 3. Performs writes following reads for array synchronising.
2082 static void fix_read_error(struct r1conf
*conf
, int read_disk
,
2083 sector_t sect
, int sectors
)
2085 struct mddev
*mddev
= conf
->mddev
;
2091 struct md_rdev
*rdev
;
2093 if (s
> (PAGE_SIZE
>>9))
2097 /* Note: no rcu protection needed here
2098 * as this is synchronous in the raid1d thread
2099 * which is the thread that might remove
2100 * a device. If raid1d ever becomes multi-threaded....
2105 rdev
= conf
->mirrors
[d
].rdev
;
2107 (test_bit(In_sync
, &rdev
->flags
) ||
2108 (!test_bit(Faulty
, &rdev
->flags
) &&
2109 rdev
->recovery_offset
>= sect
+ s
)) &&
2110 is_badblock(rdev
, sect
, s
,
2111 &first_bad
, &bad_sectors
) == 0 &&
2112 sync_page_io(rdev
, sect
, s
<<9,
2113 conf
->tmppage
, READ
, false))
2117 if (d
== conf
->raid_disks
* 2)
2120 } while (!success
&& d
!= read_disk
);
2123 /* Cannot read from anywhere - mark it bad */
2124 struct md_rdev
*rdev
= conf
->mirrors
[read_disk
].rdev
;
2125 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
2126 md_error(mddev
, rdev
);
2129 /* write it back and re-read */
2131 while (d
!= read_disk
) {
2133 d
= conf
->raid_disks
* 2;
2135 rdev
= conf
->mirrors
[d
].rdev
;
2137 test_bit(In_sync
, &rdev
->flags
))
2138 r1_sync_page_io(rdev
, sect
, s
,
2139 conf
->tmppage
, WRITE
);
2142 while (d
!= read_disk
) {
2143 char b
[BDEVNAME_SIZE
];
2145 d
= conf
->raid_disks
* 2;
2147 rdev
= conf
->mirrors
[d
].rdev
;
2149 test_bit(In_sync
, &rdev
->flags
)) {
2150 if (r1_sync_page_io(rdev
, sect
, s
,
2151 conf
->tmppage
, READ
)) {
2152 atomic_add(s
, &rdev
->corrected_errors
);
2154 "md/raid1:%s: read error corrected "
2155 "(%d sectors at %llu on %s)\n",
2157 (unsigned long long)(sect
+
2159 bdevname(rdev
->bdev
, b
));
2168 static int narrow_write_error(struct r1bio
*r1_bio
, int i
)
2170 struct mddev
*mddev
= r1_bio
->mddev
;
2171 struct r1conf
*conf
= mddev
->private;
2172 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2174 /* bio has the data to be written to device 'i' where
2175 * we just recently had a write error.
2176 * We repeatedly clone the bio and trim down to one block,
2177 * then try the write. Where the write fails we record
2179 * It is conceivable that the bio doesn't exactly align with
2180 * blocks. We must handle this somehow.
2182 * We currently own a reference on the rdev.
2188 int sect_to_write
= r1_bio
->sectors
;
2191 if (rdev
->badblocks
.shift
< 0)
2194 block_sectors
= 1 << rdev
->badblocks
.shift
;
2195 sector
= r1_bio
->sector
;
2196 sectors
= ((sector
+ block_sectors
)
2197 & ~(sector_t
)(block_sectors
- 1))
2200 while (sect_to_write
) {
2202 if (sectors
> sect_to_write
)
2203 sectors
= sect_to_write
;
2204 /* Write at 'sector' for 'sectors'*/
2206 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
2207 unsigned vcnt
= r1_bio
->behind_page_count
;
2208 struct bio_vec
*vec
= r1_bio
->behind_bvecs
;
2210 while (!vec
->bv_page
) {
2215 wbio
= bio_alloc_mddev(GFP_NOIO
, vcnt
, mddev
);
2216 memcpy(wbio
->bi_io_vec
, vec
, vcnt
* sizeof(struct bio_vec
));
2218 wbio
->bi_vcnt
= vcnt
;
2220 wbio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2223 wbio
->bi_rw
= WRITE
;
2224 wbio
->bi_sector
= r1_bio
->sector
;
2225 wbio
->bi_size
= r1_bio
->sectors
<< 9;
2227 bio_trim(wbio
, sector
- r1_bio
->sector
, sectors
);
2228 wbio
->bi_sector
+= rdev
->data_offset
;
2229 wbio
->bi_bdev
= rdev
->bdev
;
2230 if (submit_bio_wait(WRITE
, wbio
) == 0)
2232 ok
= rdev_set_badblocks(rdev
, sector
,
2237 sect_to_write
-= sectors
;
2239 sectors
= block_sectors
;
2244 static void handle_sync_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2247 int s
= r1_bio
->sectors
;
2248 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++) {
2249 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2250 struct bio
*bio
= r1_bio
->bios
[m
];
2251 if (bio
->bi_end_io
== NULL
)
2253 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
2254 test_bit(R1BIO_MadeGood
, &r1_bio
->state
)) {
2255 rdev_clear_badblocks(rdev
, r1_bio
->sector
, s
, 0);
2257 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
2258 test_bit(R1BIO_WriteError
, &r1_bio
->state
)) {
2259 if (!rdev_set_badblocks(rdev
, r1_bio
->sector
, s
, 0))
2260 md_error(conf
->mddev
, rdev
);
2264 md_done_sync(conf
->mddev
, s
, 1);
2267 static void handle_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2270 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++)
2271 if (r1_bio
->bios
[m
] == IO_MADE_GOOD
) {
2272 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2273 rdev_clear_badblocks(rdev
,
2275 r1_bio
->sectors
, 0);
2276 rdev_dec_pending(rdev
, conf
->mddev
);
2277 } else if (r1_bio
->bios
[m
] != NULL
) {
2278 /* This drive got a write error. We need to
2279 * narrow down and record precise write
2282 if (!narrow_write_error(r1_bio
, m
)) {
2283 md_error(conf
->mddev
,
2284 conf
->mirrors
[m
].rdev
);
2285 /* an I/O failed, we can't clear the bitmap */
2286 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2288 rdev_dec_pending(conf
->mirrors
[m
].rdev
,
2291 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2292 close_write(r1_bio
);
2293 raid_end_bio_io(r1_bio
);
2296 static void handle_read_error(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2300 struct mddev
*mddev
= conf
->mddev
;
2302 char b
[BDEVNAME_SIZE
];
2303 struct md_rdev
*rdev
;
2305 clear_bit(R1BIO_ReadError
, &r1_bio
->state
);
2306 /* we got a read error. Maybe the drive is bad. Maybe just
2307 * the block and we can fix it.
2308 * We freeze all other IO, and try reading the block from
2309 * other devices. When we find one, we re-write
2310 * and check it that fixes the read error.
2311 * This is all done synchronously while the array is
2314 if (mddev
->ro
== 0) {
2315 freeze_array(conf
, 1);
2316 fix_read_error(conf
, r1_bio
->read_disk
,
2317 r1_bio
->sector
, r1_bio
->sectors
);
2318 unfreeze_array(conf
);
2320 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
2321 rdev_dec_pending(conf
->mirrors
[r1_bio
->read_disk
].rdev
, conf
->mddev
);
2323 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2324 bdevname(bio
->bi_bdev
, b
);
2326 disk
= read_balance(conf
, r1_bio
, &max_sectors
);
2328 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O"
2329 " read error for block %llu\n",
2330 mdname(mddev
), b
, (unsigned long long)r1_bio
->sector
);
2331 raid_end_bio_io(r1_bio
);
2333 const unsigned long do_sync
2334 = r1_bio
->master_bio
->bi_rw
& REQ_SYNC
;
2336 r1_bio
->bios
[r1_bio
->read_disk
] =
2337 mddev
->ro
? IO_BLOCKED
: NULL
;
2340 r1_bio
->read_disk
= disk
;
2341 bio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2342 bio_trim(bio
, r1_bio
->sector
- bio
->bi_sector
, max_sectors
);
2343 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
2344 rdev
= conf
->mirrors
[disk
].rdev
;
2345 printk_ratelimited(KERN_ERR
2346 "md/raid1:%s: redirecting sector %llu"
2347 " to other mirror: %s\n",
2349 (unsigned long long)r1_bio
->sector
,
2350 bdevname(rdev
->bdev
, b
));
2351 bio
->bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
2352 bio
->bi_bdev
= rdev
->bdev
;
2353 bio
->bi_end_io
= raid1_end_read_request
;
2354 bio
->bi_rw
= READ
| do_sync
;
2355 bio
->bi_private
= r1_bio
;
2356 if (max_sectors
< r1_bio
->sectors
) {
2357 /* Drat - have to split this up more */
2358 struct bio
*mbio
= r1_bio
->master_bio
;
2359 int sectors_handled
= (r1_bio
->sector
+ max_sectors
2361 r1_bio
->sectors
= max_sectors
;
2362 spin_lock_irq(&conf
->device_lock
);
2363 if (mbio
->bi_phys_segments
== 0)
2364 mbio
->bi_phys_segments
= 2;
2366 mbio
->bi_phys_segments
++;
2367 spin_unlock_irq(&conf
->device_lock
);
2368 generic_make_request(bio
);
2371 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
2373 r1_bio
->master_bio
= mbio
;
2374 r1_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2376 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
2377 r1_bio
->mddev
= mddev
;
2378 r1_bio
->sector
= mbio
->bi_sector
+ sectors_handled
;
2382 generic_make_request(bio
);
2386 static void raid1d(struct md_thread
*thread
)
2388 struct mddev
*mddev
= thread
->mddev
;
2389 struct r1bio
*r1_bio
;
2390 unsigned long flags
;
2391 struct r1conf
*conf
= mddev
->private;
2392 struct list_head
*head
= &conf
->retry_list
;
2393 struct blk_plug plug
;
2395 md_check_recovery(mddev
);
2397 blk_start_plug(&plug
);
2400 flush_pending_writes(conf
);
2402 spin_lock_irqsave(&conf
->device_lock
, flags
);
2403 if (list_empty(head
)) {
2404 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2407 r1_bio
= list_entry(head
->prev
, struct r1bio
, retry_list
);
2408 list_del(head
->prev
);
2410 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2412 mddev
= r1_bio
->mddev
;
2413 conf
= mddev
->private;
2414 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
2415 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2416 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2417 handle_sync_write_finished(conf
, r1_bio
);
2419 sync_request_write(mddev
, r1_bio
);
2420 } else if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2421 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2422 handle_write_finished(conf
, r1_bio
);
2423 else if (test_bit(R1BIO_ReadError
, &r1_bio
->state
))
2424 handle_read_error(conf
, r1_bio
);
2426 /* just a partial read to be scheduled from separate
2429 generic_make_request(r1_bio
->bios
[r1_bio
->read_disk
]);
2432 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2433 md_check_recovery(mddev
);
2435 blk_finish_plug(&plug
);
2439 static int init_resync(struct r1conf
*conf
)
2443 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2444 BUG_ON(conf
->r1buf_pool
);
2445 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
2447 if (!conf
->r1buf_pool
)
2449 conf
->next_resync
= 0;
2454 * perform a "sync" on one "block"
2456 * We need to make sure that no normal I/O request - particularly write
2457 * requests - conflict with active sync requests.
2459 * This is achieved by tracking pending requests and a 'barrier' concept
2460 * that can be installed to exclude normal IO requests.
2463 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
2465 struct r1conf
*conf
= mddev
->private;
2466 struct r1bio
*r1_bio
;
2468 sector_t max_sector
, nr_sectors
;
2472 int write_targets
= 0, read_targets
= 0;
2473 sector_t sync_blocks
;
2474 int still_degraded
= 0;
2475 int good_sectors
= RESYNC_SECTORS
;
2476 int min_bad
= 0; /* number of sectors that are bad in all devices */
2478 if (!conf
->r1buf_pool
)
2479 if (init_resync(conf
))
2482 max_sector
= mddev
->dev_sectors
;
2483 if (sector_nr
>= max_sector
) {
2484 /* If we aborted, we need to abort the
2485 * sync on the 'current' bitmap chunk (there will
2486 * only be one in raid1 resync.
2487 * We can find the current addess in mddev->curr_resync
2489 if (mddev
->curr_resync
< max_sector
) /* aborted */
2490 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2492 else /* completed sync */
2495 bitmap_close_sync(mddev
->bitmap
);
2500 if (mddev
->bitmap
== NULL
&&
2501 mddev
->recovery_cp
== MaxSector
&&
2502 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2503 conf
->fullsync
== 0) {
2505 return max_sector
- sector_nr
;
2507 /* before building a request, check if we can skip these blocks..
2508 * This call the bitmap_start_sync doesn't actually record anything
2510 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
2511 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2512 /* We can skip this block, and probably several more */
2517 * If there is non-resync activity waiting for a turn,
2518 * and resync is going fast enough,
2519 * then let it though before starting on this new sync request.
2521 if (!go_faster
&& conf
->nr_waiting
)
2522 msleep_interruptible(1000);
2524 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
2525 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
2526 raise_barrier(conf
);
2528 conf
->next_resync
= sector_nr
;
2532 * If we get a correctably read error during resync or recovery,
2533 * we might want to read from a different device. So we
2534 * flag all drives that could conceivably be read from for READ,
2535 * and any others (which will be non-In_sync devices) for WRITE.
2536 * If a read fails, we try reading from something else for which READ
2540 r1_bio
->mddev
= mddev
;
2541 r1_bio
->sector
= sector_nr
;
2543 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
2545 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2546 struct md_rdev
*rdev
;
2547 bio
= r1_bio
->bios
[i
];
2550 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
2552 test_bit(Faulty
, &rdev
->flags
)) {
2553 if (i
< conf
->raid_disks
)
2555 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
2557 bio
->bi_end_io
= end_sync_write
;
2560 /* may need to read from here */
2561 sector_t first_bad
= MaxSector
;
2564 if (is_badblock(rdev
, sector_nr
, good_sectors
,
2565 &first_bad
, &bad_sectors
)) {
2566 if (first_bad
> sector_nr
)
2567 good_sectors
= first_bad
- sector_nr
;
2569 bad_sectors
-= (sector_nr
- first_bad
);
2571 min_bad
> bad_sectors
)
2572 min_bad
= bad_sectors
;
2575 if (sector_nr
< first_bad
) {
2576 if (test_bit(WriteMostly
, &rdev
->flags
)) {
2584 bio
->bi_end_io
= end_sync_read
;
2586 } else if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
2587 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2588 !test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)) {
2590 * The device is suitable for reading (InSync),
2591 * but has bad block(s) here. Let's try to correct them,
2592 * if we are doing resync or repair. Otherwise, leave
2593 * this device alone for this sync request.
2596 bio
->bi_end_io
= end_sync_write
;
2600 if (bio
->bi_end_io
) {
2601 atomic_inc(&rdev
->nr_pending
);
2602 bio
->bi_sector
= sector_nr
+ rdev
->data_offset
;
2603 bio
->bi_bdev
= rdev
->bdev
;
2604 bio
->bi_private
= r1_bio
;
2610 r1_bio
->read_disk
= disk
;
2612 if (read_targets
== 0 && min_bad
> 0) {
2613 /* These sectors are bad on all InSync devices, so we
2614 * need to mark them bad on all write targets
2617 for (i
= 0 ; i
< conf
->raid_disks
* 2 ; i
++)
2618 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_write
) {
2619 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2620 ok
= rdev_set_badblocks(rdev
, sector_nr
,
2624 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2629 /* Cannot record the badblocks, so need to
2631 * If there are multiple read targets, could just
2632 * fail the really bad ones ???
2634 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2635 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2641 if (min_bad
> 0 && min_bad
< good_sectors
) {
2642 /* only resync enough to reach the next bad->good
2644 good_sectors
= min_bad
;
2647 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
2648 /* extra read targets are also write targets */
2649 write_targets
+= read_targets
-1;
2651 if (write_targets
== 0 || read_targets
== 0) {
2652 /* There is nowhere to write, so all non-sync
2653 * drives must be failed - so we are finished
2657 max_sector
= sector_nr
+ min_bad
;
2658 rv
= max_sector
- sector_nr
;
2664 if (max_sector
> mddev
->resync_max
)
2665 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2666 if (max_sector
> sector_nr
+ good_sectors
)
2667 max_sector
= sector_nr
+ good_sectors
;
2672 int len
= PAGE_SIZE
;
2673 if (sector_nr
+ (len
>>9) > max_sector
)
2674 len
= (max_sector
- sector_nr
) << 9;
2677 if (sync_blocks
== 0) {
2678 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2679 &sync_blocks
, still_degraded
) &&
2681 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2683 BUG_ON(sync_blocks
< (PAGE_SIZE
>>9));
2684 if ((len
>> 9) > sync_blocks
)
2685 len
= sync_blocks
<<9;
2688 for (i
= 0 ; i
< conf
->raid_disks
* 2; i
++) {
2689 bio
= r1_bio
->bios
[i
];
2690 if (bio
->bi_end_io
) {
2691 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2692 if (bio_add_page(bio
, page
, len
, 0) == 0) {
2694 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2697 bio
= r1_bio
->bios
[i
];
2698 if (bio
->bi_end_io
==NULL
)
2700 /* remove last page from this bio */
2702 bio
->bi_size
-= len
;
2703 bio
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
2709 nr_sectors
+= len
>>9;
2710 sector_nr
+= len
>>9;
2711 sync_blocks
-= (len
>>9);
2712 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
2714 r1_bio
->sectors
= nr_sectors
;
2716 /* For a user-requested sync, we read all readable devices and do a
2719 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2720 atomic_set(&r1_bio
->remaining
, read_targets
);
2721 for (i
= 0; i
< conf
->raid_disks
* 2 && read_targets
; i
++) {
2722 bio
= r1_bio
->bios
[i
];
2723 if (bio
->bi_end_io
== end_sync_read
) {
2725 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2726 generic_make_request(bio
);
2730 atomic_set(&r1_bio
->remaining
, 1);
2731 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2732 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2733 generic_make_request(bio
);
2739 static sector_t
raid1_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2744 return mddev
->dev_sectors
;
2747 static struct r1conf
*setup_conf(struct mddev
*mddev
)
2749 struct r1conf
*conf
;
2751 struct raid1_info
*disk
;
2752 struct md_rdev
*rdev
;
2755 conf
= kzalloc(sizeof(struct r1conf
), GFP_KERNEL
);
2759 conf
->mirrors
= kzalloc(sizeof(struct raid1_info
)
2760 * mddev
->raid_disks
* 2,
2765 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2769 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
2770 if (!conf
->poolinfo
)
2772 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
* 2;
2773 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2776 if (!conf
->r1bio_pool
)
2779 conf
->poolinfo
->mddev
= mddev
;
2782 spin_lock_init(&conf
->device_lock
);
2783 rdev_for_each(rdev
, mddev
) {
2784 struct request_queue
*q
;
2785 int disk_idx
= rdev
->raid_disk
;
2786 if (disk_idx
>= mddev
->raid_disks
2789 if (test_bit(Replacement
, &rdev
->flags
))
2790 disk
= conf
->mirrors
+ mddev
->raid_disks
+ disk_idx
;
2792 disk
= conf
->mirrors
+ disk_idx
;
2797 q
= bdev_get_queue(rdev
->bdev
);
2798 if (q
->merge_bvec_fn
)
2799 mddev
->merge_check_needed
= 1;
2801 disk
->head_position
= 0;
2802 disk
->seq_start
= MaxSector
;
2804 conf
->raid_disks
= mddev
->raid_disks
;
2805 conf
->mddev
= mddev
;
2806 INIT_LIST_HEAD(&conf
->retry_list
);
2808 spin_lock_init(&conf
->resync_lock
);
2809 init_waitqueue_head(&conf
->wait_barrier
);
2811 bio_list_init(&conf
->pending_bio_list
);
2812 conf
->pending_count
= 0;
2813 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
2815 conf
->start_next_window
= MaxSector
;
2816 conf
->current_window_requests
= conf
->next_window_requests
= 0;
2819 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2821 disk
= conf
->mirrors
+ i
;
2823 if (i
< conf
->raid_disks
&&
2824 disk
[conf
->raid_disks
].rdev
) {
2825 /* This slot has a replacement. */
2827 /* No original, just make the replacement
2828 * a recovering spare
2831 disk
[conf
->raid_disks
].rdev
;
2832 disk
[conf
->raid_disks
].rdev
= NULL
;
2833 } else if (!test_bit(In_sync
, &disk
->rdev
->flags
))
2834 /* Original is not in_sync - bad */
2839 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2840 disk
->head_position
= 0;
2842 (disk
->rdev
->saved_raid_disk
< 0))
2848 conf
->thread
= md_register_thread(raid1d
, mddev
, "raid1");
2849 if (!conf
->thread
) {
2851 "md/raid1:%s: couldn't allocate thread\n",
2860 if (conf
->r1bio_pool
)
2861 mempool_destroy(conf
->r1bio_pool
);
2862 kfree(conf
->mirrors
);
2863 safe_put_page(conf
->tmppage
);
2864 kfree(conf
->poolinfo
);
2867 return ERR_PTR(err
);
2870 static int stop(struct mddev
*mddev
);
2871 static int run(struct mddev
*mddev
)
2873 struct r1conf
*conf
;
2875 struct md_rdev
*rdev
;
2877 bool discard_supported
= false;
2879 if (mddev
->level
!= 1) {
2880 printk(KERN_ERR
"md/raid1:%s: raid level not set to mirroring (%d)\n",
2881 mdname(mddev
), mddev
->level
);
2884 if (mddev
->reshape_position
!= MaxSector
) {
2885 printk(KERN_ERR
"md/raid1:%s: reshape_position set but not supported\n",
2890 * copy the already verified devices into our private RAID1
2891 * bookkeeping area. [whatever we allocate in run(),
2892 * should be freed in stop()]
2894 if (mddev
->private == NULL
)
2895 conf
= setup_conf(mddev
);
2897 conf
= mddev
->private;
2900 return PTR_ERR(conf
);
2903 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
2905 rdev_for_each(rdev
, mddev
) {
2906 if (!mddev
->gendisk
)
2908 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2909 rdev
->data_offset
<< 9);
2910 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
2911 discard_supported
= true;
2914 mddev
->degraded
= 0;
2915 for (i
=0; i
< conf
->raid_disks
; i
++)
2916 if (conf
->mirrors
[i
].rdev
== NULL
||
2917 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
2918 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2921 if (conf
->raid_disks
- mddev
->degraded
== 1)
2922 mddev
->recovery_cp
= MaxSector
;
2924 if (mddev
->recovery_cp
!= MaxSector
)
2925 printk(KERN_NOTICE
"md/raid1:%s: not clean"
2926 " -- starting background reconstruction\n",
2929 "md/raid1:%s: active with %d out of %d mirrors\n",
2930 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2934 * Ok, everything is just fine now
2936 mddev
->thread
= conf
->thread
;
2937 conf
->thread
= NULL
;
2938 mddev
->private = conf
;
2940 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
2943 mddev
->queue
->backing_dev_info
.congested_fn
= raid1_congested
;
2944 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2945 blk_queue_merge_bvec(mddev
->queue
, raid1_mergeable_bvec
);
2947 if (discard_supported
)
2948 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
2951 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
2955 ret
= md_integrity_register(mddev
);
2961 static int stop(struct mddev
*mddev
)
2963 struct r1conf
*conf
= mddev
->private;
2964 struct bitmap
*bitmap
= mddev
->bitmap
;
2966 /* wait for behind writes to complete */
2967 if (bitmap
&& atomic_read(&bitmap
->behind_writes
) > 0) {
2968 printk(KERN_INFO
"md/raid1:%s: behind writes in progress - waiting to stop.\n",
2970 /* need to kick something here to make sure I/O goes? */
2971 wait_event(bitmap
->behind_wait
,
2972 atomic_read(&bitmap
->behind_writes
) == 0);
2975 freeze_array(conf
, 0);
2976 unfreeze_array(conf
);
2978 md_unregister_thread(&mddev
->thread
);
2979 if (conf
->r1bio_pool
)
2980 mempool_destroy(conf
->r1bio_pool
);
2981 kfree(conf
->mirrors
);
2982 safe_put_page(conf
->tmppage
);
2983 kfree(conf
->poolinfo
);
2985 mddev
->private = NULL
;
2989 static int raid1_resize(struct mddev
*mddev
, sector_t sectors
)
2991 /* no resync is happening, and there is enough space
2992 * on all devices, so we can resize.
2993 * We need to make sure resync covers any new space.
2994 * If the array is shrinking we should possibly wait until
2995 * any io in the removed space completes, but it hardly seems
2998 sector_t newsize
= raid1_size(mddev
, sectors
, 0);
2999 if (mddev
->external_size
&&
3000 mddev
->array_sectors
> newsize
)
3002 if (mddev
->bitmap
) {
3003 int ret
= bitmap_resize(mddev
->bitmap
, newsize
, 0, 0);
3007 md_set_array_sectors(mddev
, newsize
);
3008 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3009 revalidate_disk(mddev
->gendisk
);
3010 if (sectors
> mddev
->dev_sectors
&&
3011 mddev
->recovery_cp
> mddev
->dev_sectors
) {
3012 mddev
->recovery_cp
= mddev
->dev_sectors
;
3013 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3015 mddev
->dev_sectors
= sectors
;
3016 mddev
->resync_max_sectors
= sectors
;
3020 static int raid1_reshape(struct mddev
*mddev
)
3023 * 1/ resize the r1bio_pool
3024 * 2/ resize conf->mirrors
3026 * We allocate a new r1bio_pool if we can.
3027 * Then raise a device barrier and wait until all IO stops.
3028 * Then resize conf->mirrors and swap in the new r1bio pool.
3030 * At the same time, we "pack" the devices so that all the missing
3031 * devices have the higher raid_disk numbers.
3033 mempool_t
*newpool
, *oldpool
;
3034 struct pool_info
*newpoolinfo
;
3035 struct raid1_info
*newmirrors
;
3036 struct r1conf
*conf
= mddev
->private;
3037 int cnt
, raid_disks
;
3038 unsigned long flags
;
3041 /* Cannot change chunk_size, layout, or level */
3042 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
3043 mddev
->layout
!= mddev
->new_layout
||
3044 mddev
->level
!= mddev
->new_level
) {
3045 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3046 mddev
->new_layout
= mddev
->layout
;
3047 mddev
->new_level
= mddev
->level
;
3051 err
= md_allow_write(mddev
);
3055 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3057 if (raid_disks
< conf
->raid_disks
) {
3059 for (d
= 0; d
< conf
->raid_disks
; d
++)
3060 if (conf
->mirrors
[d
].rdev
)
3062 if (cnt
> raid_disks
)
3066 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
3069 newpoolinfo
->mddev
= mddev
;
3070 newpoolinfo
->raid_disks
= raid_disks
* 2;
3072 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
3073 r1bio_pool_free
, newpoolinfo
);
3078 newmirrors
= kzalloc(sizeof(struct raid1_info
) * raid_disks
* 2,
3082 mempool_destroy(newpool
);
3086 freeze_array(conf
, 0);
3088 /* ok, everything is stopped */
3089 oldpool
= conf
->r1bio_pool
;
3090 conf
->r1bio_pool
= newpool
;
3092 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
3093 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
3094 if (rdev
&& rdev
->raid_disk
!= d2
) {
3095 sysfs_unlink_rdev(mddev
, rdev
);
3096 rdev
->raid_disk
= d2
;
3097 sysfs_unlink_rdev(mddev
, rdev
);
3098 if (sysfs_link_rdev(mddev
, rdev
))
3100 "md/raid1:%s: cannot register rd%d\n",
3101 mdname(mddev
), rdev
->raid_disk
);
3104 newmirrors
[d2
++].rdev
= rdev
;
3106 kfree(conf
->mirrors
);
3107 conf
->mirrors
= newmirrors
;
3108 kfree(conf
->poolinfo
);
3109 conf
->poolinfo
= newpoolinfo
;
3111 spin_lock_irqsave(&conf
->device_lock
, flags
);
3112 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
3113 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3114 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
3115 mddev
->delta_disks
= 0;
3117 unfreeze_array(conf
);
3119 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3120 md_wakeup_thread(mddev
->thread
);
3122 mempool_destroy(oldpool
);
3126 static void raid1_quiesce(struct mddev
*mddev
, int state
)
3128 struct r1conf
*conf
= mddev
->private;
3131 case 2: /* wake for suspend */
3132 wake_up(&conf
->wait_barrier
);
3135 freeze_array(conf
, 0);
3138 unfreeze_array(conf
);
3143 static void *raid1_takeover(struct mddev
*mddev
)
3145 /* raid1 can take over:
3146 * raid5 with 2 devices, any layout or chunk size
3148 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
3149 struct r1conf
*conf
;
3150 mddev
->new_level
= 1;
3151 mddev
->new_layout
= 0;
3152 mddev
->new_chunk_sectors
= 0;
3153 conf
= setup_conf(mddev
);
3155 /* Array must appear to be quiesced */
3156 conf
->array_frozen
= 1;
3159 return ERR_PTR(-EINVAL
);
3162 static struct md_personality raid1_personality
=
3166 .owner
= THIS_MODULE
,
3167 .make_request
= make_request
,
3171 .error_handler
= error
,
3172 .hot_add_disk
= raid1_add_disk
,
3173 .hot_remove_disk
= raid1_remove_disk
,
3174 .spare_active
= raid1_spare_active
,
3175 .sync_request
= sync_request
,
3176 .resize
= raid1_resize
,
3178 .check_reshape
= raid1_reshape
,
3179 .quiesce
= raid1_quiesce
,
3180 .takeover
= raid1_takeover
,
3183 static int __init
raid_init(void)
3185 return register_md_personality(&raid1_personality
);
3188 static void raid_exit(void)
3190 unregister_md_personality(&raid1_personality
);
3193 module_init(raid_init
);
3194 module_exit(raid_exit
);
3195 MODULE_LICENSE("GPL");
3196 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3197 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3198 MODULE_ALIAS("md-raid1");
3199 MODULE_ALIAS("md-level-1");
3201 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);