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 CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
94 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
95 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
97 static void * r1buf_pool_alloc(gfp_t gfp_flags
, void *data
)
99 struct pool_info
*pi
= data
;
100 struct r1bio
*r1_bio
;
105 r1_bio
= r1bio_pool_alloc(gfp_flags
, pi
);
110 * Allocate bios : 1 for reading, n-1 for writing
112 for (j
= pi
->raid_disks
; j
-- ; ) {
113 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
116 r1_bio
->bios
[j
] = bio
;
119 * Allocate RESYNC_PAGES data pages and attach them to
121 * If this is a user-requested check/repair, allocate
122 * RESYNC_PAGES for each bio.
124 if (test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
))
125 need_pages
= pi
->raid_disks
;
128 for (j
= 0; j
< need_pages
; j
++) {
129 bio
= r1_bio
->bios
[j
];
130 bio
->bi_vcnt
= RESYNC_PAGES
;
132 if (bio_alloc_pages(bio
, gfp_flags
))
135 /* If not user-requests, copy the page pointers to all bios */
136 if (!test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
)) {
137 for (i
=0; i
<RESYNC_PAGES
; i
++)
138 for (j
=1; j
<pi
->raid_disks
; j
++)
139 r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
=
140 r1_bio
->bios
[0]->bi_io_vec
[i
].bv_page
;
143 r1_bio
->master_bio
= NULL
;
151 bio_for_each_segment_all(bv
, r1_bio
->bios
[j
], i
)
152 __free_page(bv
->bv_page
);
156 while (++j
< pi
->raid_disks
)
157 bio_put(r1_bio
->bios
[j
]);
158 r1bio_pool_free(r1_bio
, data
);
162 static void r1buf_pool_free(void *__r1_bio
, void *data
)
164 struct pool_info
*pi
= data
;
166 struct r1bio
*r1bio
= __r1_bio
;
168 for (i
= 0; i
< RESYNC_PAGES
; i
++)
169 for (j
= pi
->raid_disks
; j
-- ;) {
171 r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
!=
172 r1bio
->bios
[0]->bi_io_vec
[i
].bv_page
)
173 safe_put_page(r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
175 for (i
=0 ; i
< pi
->raid_disks
; i
++)
176 bio_put(r1bio
->bios
[i
]);
178 r1bio_pool_free(r1bio
, data
);
181 static void put_all_bios(struct r1conf
*conf
, struct r1bio
*r1_bio
)
185 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
186 struct bio
**bio
= r1_bio
->bios
+ i
;
187 if (!BIO_SPECIAL(*bio
))
193 static void free_r1bio(struct r1bio
*r1_bio
)
195 struct r1conf
*conf
= r1_bio
->mddev
->private;
197 put_all_bios(conf
, r1_bio
);
198 mempool_free(r1_bio
, conf
->r1bio_pool
);
201 static void put_buf(struct r1bio
*r1_bio
)
203 struct r1conf
*conf
= r1_bio
->mddev
->private;
206 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
207 struct bio
*bio
= r1_bio
->bios
[i
];
209 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
212 mempool_free(r1_bio
, conf
->r1buf_pool
);
217 static void reschedule_retry(struct r1bio
*r1_bio
)
220 struct mddev
*mddev
= r1_bio
->mddev
;
221 struct r1conf
*conf
= mddev
->private;
223 spin_lock_irqsave(&conf
->device_lock
, flags
);
224 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
226 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
228 wake_up(&conf
->wait_barrier
);
229 md_wakeup_thread(mddev
->thread
);
233 * raid_end_bio_io() is called when we have finished servicing a mirrored
234 * operation and are ready to return a success/failure code to the buffer
237 static void call_bio_endio(struct r1bio
*r1_bio
)
239 struct bio
*bio
= r1_bio
->master_bio
;
241 struct r1conf
*conf
= r1_bio
->mddev
->private;
242 sector_t start_next_window
= r1_bio
->start_next_window
;
243 sector_t bi_sector
= bio
->bi_iter
.bi_sector
;
245 if (bio
->bi_phys_segments
) {
247 spin_lock_irqsave(&conf
->device_lock
, flags
);
248 bio
->bi_phys_segments
--;
249 done
= (bio
->bi_phys_segments
== 0);
250 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
252 * make_request() might be waiting for
253 * bi_phys_segments to decrease
255 wake_up(&conf
->wait_barrier
);
259 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
260 bio
->bi_error
= -EIO
;
265 * Wake up any possible resync thread that waits for the device
268 allow_barrier(conf
, start_next_window
, bi_sector
);
272 static void raid_end_bio_io(struct r1bio
*r1_bio
)
274 struct bio
*bio
= r1_bio
->master_bio
;
276 /* if nobody has done the final endio yet, do it now */
277 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
278 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
279 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
280 (unsigned long long) bio
->bi_iter
.bi_sector
,
281 (unsigned long long) bio_end_sector(bio
) - 1);
283 call_bio_endio(r1_bio
);
289 * Update disk head position estimator based on IRQ completion info.
291 static inline void update_head_pos(int disk
, struct r1bio
*r1_bio
)
293 struct r1conf
*conf
= r1_bio
->mddev
->private;
295 conf
->mirrors
[disk
].head_position
=
296 r1_bio
->sector
+ (r1_bio
->sectors
);
300 * Find the disk number which triggered given bio
302 static int find_bio_disk(struct r1bio
*r1_bio
, struct bio
*bio
)
305 struct r1conf
*conf
= r1_bio
->mddev
->private;
306 int raid_disks
= conf
->raid_disks
;
308 for (mirror
= 0; mirror
< raid_disks
* 2; mirror
++)
309 if (r1_bio
->bios
[mirror
] == bio
)
312 BUG_ON(mirror
== raid_disks
* 2);
313 update_head_pos(mirror
, r1_bio
);
318 static void raid1_end_read_request(struct bio
*bio
)
320 int uptodate
= !bio
->bi_error
;
321 struct r1bio
*r1_bio
= bio
->bi_private
;
322 struct r1conf
*conf
= r1_bio
->mddev
->private;
323 struct md_rdev
*rdev
= conf
->mirrors
[r1_bio
->read_disk
].rdev
;
326 * this branch is our 'one mirror IO has finished' event handler:
328 update_head_pos(r1_bio
->read_disk
, r1_bio
);
331 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
333 /* If all other devices have failed, we want to return
334 * the error upwards rather than fail the last device.
335 * Here we redefine "uptodate" to mean "Don't want to retry"
338 spin_lock_irqsave(&conf
->device_lock
, flags
);
339 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
340 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
341 test_bit(In_sync
, &rdev
->flags
)))
343 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
347 raid_end_bio_io(r1_bio
);
348 rdev_dec_pending(rdev
, conf
->mddev
);
353 char b
[BDEVNAME_SIZE
];
355 KERN_ERR
"md/raid1:%s: %s: "
356 "rescheduling sector %llu\n",
360 (unsigned long long)r1_bio
->sector
);
361 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
362 reschedule_retry(r1_bio
);
363 /* don't drop the reference on read_disk yet */
367 static void close_write(struct r1bio
*r1_bio
)
369 /* it really is the end of this request */
370 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
371 /* free extra copy of the data pages */
372 int i
= r1_bio
->behind_page_count
;
374 safe_put_page(r1_bio
->behind_bvecs
[i
].bv_page
);
375 kfree(r1_bio
->behind_bvecs
);
376 r1_bio
->behind_bvecs
= NULL
;
378 /* clear the bitmap if all writes complete successfully */
379 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
381 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
382 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
383 md_write_end(r1_bio
->mddev
);
386 static void r1_bio_write_done(struct r1bio
*r1_bio
)
388 if (!atomic_dec_and_test(&r1_bio
->remaining
))
391 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
392 reschedule_retry(r1_bio
);
395 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
))
396 reschedule_retry(r1_bio
);
398 raid_end_bio_io(r1_bio
);
402 static void raid1_end_write_request(struct bio
*bio
)
404 struct r1bio
*r1_bio
= bio
->bi_private
;
405 int behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
406 struct r1conf
*conf
= r1_bio
->mddev
->private;
407 struct bio
*to_put
= NULL
;
408 int mirror
= find_bio_disk(r1_bio
, bio
);
409 struct md_rdev
*rdev
= conf
->mirrors
[mirror
].rdev
;
412 * 'one mirror IO has finished' event handler:
415 set_bit(WriteErrorSeen
, &rdev
->flags
);
416 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
417 set_bit(MD_RECOVERY_NEEDED
, &
418 conf
->mddev
->recovery
);
420 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
423 * Set R1BIO_Uptodate in our master bio, so that we
424 * will return a good error code for to the higher
425 * levels even if IO on some other mirrored buffer
428 * The 'master' represents the composite IO operation
429 * to user-side. So if something waits for IO, then it
430 * will wait for the 'master' bio.
435 r1_bio
->bios
[mirror
] = NULL
;
438 * Do not set R1BIO_Uptodate if the current device is
439 * rebuilding or Faulty. This is because we cannot use
440 * such device for properly reading the data back (we could
441 * potentially use it, if the current write would have felt
442 * before rdev->recovery_offset, but for simplicity we don't
445 if (test_bit(In_sync
, &rdev
->flags
) &&
446 !test_bit(Faulty
, &rdev
->flags
))
447 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
449 /* Maybe we can clear some bad blocks. */
450 if (is_badblock(rdev
, r1_bio
->sector
, r1_bio
->sectors
,
451 &first_bad
, &bad_sectors
)) {
452 r1_bio
->bios
[mirror
] = IO_MADE_GOOD
;
453 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
458 if (test_bit(WriteMostly
, &rdev
->flags
))
459 atomic_dec(&r1_bio
->behind_remaining
);
462 * In behind mode, we ACK the master bio once the I/O
463 * has safely reached all non-writemostly
464 * disks. Setting the Returned bit ensures that this
465 * gets done only once -- we don't ever want to return
466 * -EIO here, instead we'll wait
468 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
469 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
470 /* Maybe we can return now */
471 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
472 struct bio
*mbio
= r1_bio
->master_bio
;
473 pr_debug("raid1: behind end write sectors"
475 (unsigned long long) mbio
->bi_iter
.bi_sector
,
476 (unsigned long long) bio_end_sector(mbio
) - 1);
477 call_bio_endio(r1_bio
);
481 if (r1_bio
->bios
[mirror
] == NULL
)
482 rdev_dec_pending(rdev
, conf
->mddev
);
485 * Let's see if all mirrored write operations have finished
488 r1_bio_write_done(r1_bio
);
495 * This routine returns the disk from which the requested read should
496 * be done. There is a per-array 'next expected sequential IO' sector
497 * number - if this matches on the next IO then we use the last disk.
498 * There is also a per-disk 'last know head position' sector that is
499 * maintained from IRQ contexts, both the normal and the resync IO
500 * completion handlers update this position correctly. If there is no
501 * perfect sequential match then we pick the disk whose head is closest.
503 * If there are 2 mirrors in the same 2 devices, performance degrades
504 * because position is mirror, not device based.
506 * The rdev for the device selected will have nr_pending incremented.
508 static int read_balance(struct r1conf
*conf
, struct r1bio
*r1_bio
, int *max_sectors
)
510 const sector_t this_sector
= r1_bio
->sector
;
512 int best_good_sectors
;
513 int best_disk
, best_dist_disk
, best_pending_disk
;
517 unsigned int min_pending
;
518 struct md_rdev
*rdev
;
520 int choose_next_idle
;
524 * Check if we can balance. We can balance on the whole
525 * device if no resync is going on, or below the resync window.
526 * We take the first readable disk when above the resync window.
529 sectors
= r1_bio
->sectors
;
532 best_dist
= MaxSector
;
533 best_pending_disk
= -1;
534 min_pending
= UINT_MAX
;
535 best_good_sectors
= 0;
537 choose_next_idle
= 0;
539 if ((conf
->mddev
->recovery_cp
< this_sector
+ sectors
) ||
540 (mddev_is_clustered(conf
->mddev
) &&
541 md_cluster_ops
->area_resyncing(conf
->mddev
, READ
, this_sector
,
542 this_sector
+ sectors
)))
547 for (disk
= 0 ; disk
< conf
->raid_disks
* 2 ; disk
++) {
551 unsigned int pending
;
554 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
555 if (r1_bio
->bios
[disk
] == IO_BLOCKED
557 || test_bit(Faulty
, &rdev
->flags
))
559 if (!test_bit(In_sync
, &rdev
->flags
) &&
560 rdev
->recovery_offset
< this_sector
+ sectors
)
562 if (test_bit(WriteMostly
, &rdev
->flags
)) {
563 /* Don't balance among write-mostly, just
564 * use the first as a last resort */
565 if (best_dist_disk
< 0) {
566 if (is_badblock(rdev
, this_sector
, sectors
,
567 &first_bad
, &bad_sectors
)) {
568 if (first_bad
<= this_sector
)
569 /* Cannot use this */
571 best_good_sectors
= first_bad
- this_sector
;
573 best_good_sectors
= sectors
;
574 best_dist_disk
= disk
;
575 best_pending_disk
= disk
;
579 /* This is a reasonable device to use. It might
582 if (is_badblock(rdev
, this_sector
, sectors
,
583 &first_bad
, &bad_sectors
)) {
584 if (best_dist
< MaxSector
)
585 /* already have a better device */
587 if (first_bad
<= this_sector
) {
588 /* cannot read here. If this is the 'primary'
589 * device, then we must not read beyond
590 * bad_sectors from another device..
592 bad_sectors
-= (this_sector
- first_bad
);
593 if (choose_first
&& sectors
> bad_sectors
)
594 sectors
= bad_sectors
;
595 if (best_good_sectors
> sectors
)
596 best_good_sectors
= sectors
;
599 sector_t good_sectors
= first_bad
- this_sector
;
600 if (good_sectors
> best_good_sectors
) {
601 best_good_sectors
= good_sectors
;
609 best_good_sectors
= sectors
;
611 nonrot
= blk_queue_nonrot(bdev_get_queue(rdev
->bdev
));
612 has_nonrot_disk
|= nonrot
;
613 pending
= atomic_read(&rdev
->nr_pending
);
614 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
619 /* Don't change to another disk for sequential reads */
620 if (conf
->mirrors
[disk
].next_seq_sect
== this_sector
622 int opt_iosize
= bdev_io_opt(rdev
->bdev
) >> 9;
623 struct raid1_info
*mirror
= &conf
->mirrors
[disk
];
627 * If buffered sequential IO size exceeds optimal
628 * iosize, check if there is idle disk. If yes, choose
629 * the idle disk. read_balance could already choose an
630 * idle disk before noticing it's a sequential IO in
631 * this disk. This doesn't matter because this disk
632 * will idle, next time it will be utilized after the
633 * first disk has IO size exceeds optimal iosize. In
634 * this way, iosize of the first disk will be optimal
635 * iosize at least. iosize of the second disk might be
636 * small, but not a big deal since when the second disk
637 * starts IO, the first disk is likely still busy.
639 if (nonrot
&& opt_iosize
> 0 &&
640 mirror
->seq_start
!= MaxSector
&&
641 mirror
->next_seq_sect
> opt_iosize
&&
642 mirror
->next_seq_sect
- opt_iosize
>=
644 choose_next_idle
= 1;
649 /* If device is idle, use it */
655 if (choose_next_idle
)
658 if (min_pending
> pending
) {
659 min_pending
= pending
;
660 best_pending_disk
= disk
;
663 if (dist
< best_dist
) {
665 best_dist_disk
= disk
;
670 * If all disks are rotational, choose the closest disk. If any disk is
671 * non-rotational, choose the disk with less pending request even the
672 * disk is rotational, which might/might not be optimal for raids with
673 * mixed ratation/non-rotational disks depending on workload.
675 if (best_disk
== -1) {
677 best_disk
= best_pending_disk
;
679 best_disk
= best_dist_disk
;
682 if (best_disk
>= 0) {
683 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
686 atomic_inc(&rdev
->nr_pending
);
687 sectors
= best_good_sectors
;
689 if (conf
->mirrors
[best_disk
].next_seq_sect
!= this_sector
)
690 conf
->mirrors
[best_disk
].seq_start
= this_sector
;
692 conf
->mirrors
[best_disk
].next_seq_sect
= this_sector
+ sectors
;
695 *max_sectors
= sectors
;
700 static int raid1_congested(struct mddev
*mddev
, int bits
)
702 struct r1conf
*conf
= mddev
->private;
705 if ((bits
& (1 << WB_async_congested
)) &&
706 conf
->pending_count
>= max_queued_requests
)
710 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
711 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
712 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
713 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
717 /* Note the '|| 1' - when read_balance prefers
718 * non-congested targets, it can be removed
720 if ((bits
& (1 << WB_async_congested
)) || 1)
721 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
723 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
730 static void flush_pending_writes(struct r1conf
*conf
)
732 /* Any writes that have been queued but are awaiting
733 * bitmap updates get flushed here.
735 spin_lock_irq(&conf
->device_lock
);
737 if (conf
->pending_bio_list
.head
) {
739 bio
= bio_list_get(&conf
->pending_bio_list
);
740 conf
->pending_count
= 0;
741 spin_unlock_irq(&conf
->device_lock
);
742 /* flush any pending bitmap writes to
743 * disk before proceeding w/ I/O */
744 bitmap_unplug(conf
->mddev
->bitmap
);
745 wake_up(&conf
->wait_barrier
);
747 while (bio
) { /* submit pending writes */
748 struct bio
*next
= bio
->bi_next
;
750 if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
751 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
755 generic_make_request(bio
);
759 spin_unlock_irq(&conf
->device_lock
);
763 * Sometimes we need to suspend IO while we do something else,
764 * either some resync/recovery, or reconfigure the array.
765 * To do this we raise a 'barrier'.
766 * The 'barrier' is a counter that can be raised multiple times
767 * to count how many activities are happening which preclude
769 * We can only raise the barrier if there is no pending IO.
770 * i.e. if nr_pending == 0.
771 * We choose only to raise the barrier if no-one is waiting for the
772 * barrier to go down. This means that as soon as an IO request
773 * is ready, no other operations which require a barrier will start
774 * until the IO request has had a chance.
776 * So: regular IO calls 'wait_barrier'. When that returns there
777 * is no backgroup IO happening, It must arrange to call
778 * allow_barrier when it has finished its IO.
779 * backgroup IO calls must call raise_barrier. Once that returns
780 * there is no normal IO happeing. It must arrange to call
781 * lower_barrier when the particular background IO completes.
783 static void raise_barrier(struct r1conf
*conf
, sector_t sector_nr
)
785 spin_lock_irq(&conf
->resync_lock
);
787 /* Wait until no block IO is waiting */
788 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
791 /* block any new IO from starting */
793 conf
->next_resync
= sector_nr
;
795 /* For these conditions we must wait:
796 * A: while the array is in frozen state
797 * B: while barrier >= RESYNC_DEPTH, meaning resync reach
798 * the max count which allowed.
799 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
800 * next resync will reach to the window which normal bios are
802 * D: while there are any active requests in the current window.
804 wait_event_lock_irq(conf
->wait_barrier
,
805 !conf
->array_frozen
&&
806 conf
->barrier
< RESYNC_DEPTH
&&
807 conf
->current_window_requests
== 0 &&
808 (conf
->start_next_window
>=
809 conf
->next_resync
+ RESYNC_SECTORS
),
813 spin_unlock_irq(&conf
->resync_lock
);
816 static void lower_barrier(struct r1conf
*conf
)
819 BUG_ON(conf
->barrier
<= 0);
820 spin_lock_irqsave(&conf
->resync_lock
, flags
);
823 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
824 wake_up(&conf
->wait_barrier
);
827 static bool need_to_wait_for_sync(struct r1conf
*conf
, struct bio
*bio
)
831 if (conf
->array_frozen
|| !bio
)
833 else if (conf
->barrier
&& bio_data_dir(bio
) == WRITE
) {
834 if ((conf
->mddev
->curr_resync_completed
835 >= bio_end_sector(bio
)) ||
836 (conf
->next_resync
+ NEXT_NORMALIO_DISTANCE
837 <= bio
->bi_iter
.bi_sector
))
846 static sector_t
wait_barrier(struct r1conf
*conf
, struct bio
*bio
)
850 spin_lock_irq(&conf
->resync_lock
);
851 if (need_to_wait_for_sync(conf
, bio
)) {
853 /* Wait for the barrier to drop.
854 * However if there are already pending
855 * requests (preventing the barrier from
856 * rising completely), and the
857 * per-process bio queue isn't empty,
858 * then don't wait, as we need to empty
859 * that queue to allow conf->start_next_window
862 wait_event_lock_irq(conf
->wait_barrier
,
863 !conf
->array_frozen
&&
865 ((conf
->start_next_window
<
866 conf
->next_resync
+ RESYNC_SECTORS
) &&
868 !bio_list_empty(current
->bio_list
))),
873 if (bio
&& bio_data_dir(bio
) == WRITE
) {
874 if (bio
->bi_iter
.bi_sector
>= conf
->next_resync
) {
875 if (conf
->start_next_window
== MaxSector
)
876 conf
->start_next_window
=
878 NEXT_NORMALIO_DISTANCE
;
880 if ((conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
)
881 <= bio
->bi_iter
.bi_sector
)
882 conf
->next_window_requests
++;
884 conf
->current_window_requests
++;
885 sector
= conf
->start_next_window
;
890 spin_unlock_irq(&conf
->resync_lock
);
894 static void allow_barrier(struct r1conf
*conf
, sector_t start_next_window
,
899 spin_lock_irqsave(&conf
->resync_lock
, flags
);
901 if (start_next_window
) {
902 if (start_next_window
== conf
->start_next_window
) {
903 if (conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
905 conf
->next_window_requests
--;
907 conf
->current_window_requests
--;
909 conf
->current_window_requests
--;
911 if (!conf
->current_window_requests
) {
912 if (conf
->next_window_requests
) {
913 conf
->current_window_requests
=
914 conf
->next_window_requests
;
915 conf
->next_window_requests
= 0;
916 conf
->start_next_window
+=
917 NEXT_NORMALIO_DISTANCE
;
919 conf
->start_next_window
= MaxSector
;
922 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
923 wake_up(&conf
->wait_barrier
);
926 static void freeze_array(struct r1conf
*conf
, int extra
)
928 /* stop syncio and normal IO and wait for everything to
930 * We wait until nr_pending match nr_queued+extra
931 * This is called in the context of one normal IO request
932 * that has failed. Thus any sync request that might be pending
933 * will be blocked by nr_pending, and we need to wait for
934 * pending IO requests to complete or be queued for re-try.
935 * Thus the number queued (nr_queued) plus this request (extra)
936 * must match the number of pending IOs (nr_pending) before
939 spin_lock_irq(&conf
->resync_lock
);
940 conf
->array_frozen
= 1;
941 wait_event_lock_irq_cmd(conf
->wait_barrier
,
942 conf
->nr_pending
== conf
->nr_queued
+extra
,
944 flush_pending_writes(conf
));
945 spin_unlock_irq(&conf
->resync_lock
);
947 static void unfreeze_array(struct r1conf
*conf
)
949 /* reverse the effect of the freeze */
950 spin_lock_irq(&conf
->resync_lock
);
951 conf
->array_frozen
= 0;
952 wake_up(&conf
->wait_barrier
);
953 spin_unlock_irq(&conf
->resync_lock
);
956 /* duplicate the data pages for behind I/O
958 static void alloc_behind_pages(struct bio
*bio
, struct r1bio
*r1_bio
)
961 struct bio_vec
*bvec
;
962 struct bio_vec
*bvecs
= kzalloc(bio
->bi_vcnt
* sizeof(struct bio_vec
),
964 if (unlikely(!bvecs
))
967 bio_for_each_segment_all(bvec
, bio
, i
) {
969 bvecs
[i
].bv_page
= alloc_page(GFP_NOIO
);
970 if (unlikely(!bvecs
[i
].bv_page
))
972 memcpy(kmap(bvecs
[i
].bv_page
) + bvec
->bv_offset
,
973 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
974 kunmap(bvecs
[i
].bv_page
);
975 kunmap(bvec
->bv_page
);
977 r1_bio
->behind_bvecs
= bvecs
;
978 r1_bio
->behind_page_count
= bio
->bi_vcnt
;
979 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
983 for (i
= 0; i
< bio
->bi_vcnt
; i
++)
984 if (bvecs
[i
].bv_page
)
985 put_page(bvecs
[i
].bv_page
);
987 pr_debug("%dB behind alloc failed, doing sync I/O\n",
988 bio
->bi_iter
.bi_size
);
991 struct raid1_plug_cb
{
992 struct blk_plug_cb cb
;
993 struct bio_list pending
;
997 static void raid1_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
999 struct raid1_plug_cb
*plug
= container_of(cb
, struct raid1_plug_cb
,
1001 struct mddev
*mddev
= plug
->cb
.data
;
1002 struct r1conf
*conf
= mddev
->private;
1005 if (from_schedule
|| current
->bio_list
) {
1006 spin_lock_irq(&conf
->device_lock
);
1007 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1008 conf
->pending_count
+= plug
->pending_cnt
;
1009 spin_unlock_irq(&conf
->device_lock
);
1010 wake_up(&conf
->wait_barrier
);
1011 md_wakeup_thread(mddev
->thread
);
1016 /* we aren't scheduling, so we can do the write-out directly. */
1017 bio
= bio_list_get(&plug
->pending
);
1018 bitmap_unplug(mddev
->bitmap
);
1019 wake_up(&conf
->wait_barrier
);
1021 while (bio
) { /* submit pending writes */
1022 struct bio
*next
= bio
->bi_next
;
1023 bio
->bi_next
= NULL
;
1024 if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
1025 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1026 /* Just ignore it */
1029 generic_make_request(bio
);
1035 static void raid1_make_request(struct mddev
*mddev
, struct bio
* bio
)
1037 struct r1conf
*conf
= mddev
->private;
1038 struct raid1_info
*mirror
;
1039 struct r1bio
*r1_bio
;
1040 struct bio
*read_bio
;
1042 struct bitmap
*bitmap
;
1043 unsigned long flags
;
1044 const int op
= bio_op(bio
);
1045 const int rw
= bio_data_dir(bio
);
1046 const unsigned long do_sync
= (bio
->bi_opf
& REQ_SYNC
);
1047 const unsigned long do_flush_fua
= (bio
->bi_opf
&
1048 (REQ_PREFLUSH
| REQ_FUA
));
1049 struct md_rdev
*blocked_rdev
;
1050 struct blk_plug_cb
*cb
;
1051 struct raid1_plug_cb
*plug
= NULL
;
1053 int sectors_handled
;
1055 sector_t start_next_window
;
1058 * Register the new request and wait if the reconstruction
1059 * thread has put up a bar for new requests.
1060 * Continue immediately if no resync is active currently.
1063 md_write_start(mddev
, bio
); /* wait on superblock update early */
1065 if (bio_data_dir(bio
) == WRITE
&&
1066 ((bio_end_sector(bio
) > mddev
->suspend_lo
&&
1067 bio
->bi_iter
.bi_sector
< mddev
->suspend_hi
) ||
1068 (mddev_is_clustered(mddev
) &&
1069 md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1070 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
))))) {
1071 /* As the suspend_* range is controlled by
1072 * userspace, we want an interruptible
1077 flush_signals(current
);
1078 prepare_to_wait(&conf
->wait_barrier
,
1079 &w
, TASK_INTERRUPTIBLE
);
1080 if (bio_end_sector(bio
) <= mddev
->suspend_lo
||
1081 bio
->bi_iter
.bi_sector
>= mddev
->suspend_hi
||
1082 (mddev_is_clustered(mddev
) &&
1083 !md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1084 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
))))
1088 finish_wait(&conf
->wait_barrier
, &w
);
1091 start_next_window
= wait_barrier(conf
, bio
);
1093 bitmap
= mddev
->bitmap
;
1096 * make_request() can abort the operation when read-ahead is being
1097 * used and no empty request is available.
1100 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1102 r1_bio
->master_bio
= bio
;
1103 r1_bio
->sectors
= bio_sectors(bio
);
1105 r1_bio
->mddev
= mddev
;
1106 r1_bio
->sector
= bio
->bi_iter
.bi_sector
;
1108 /* We might need to issue multiple reads to different
1109 * devices if there are bad blocks around, so we keep
1110 * track of the number of reads in bio->bi_phys_segments.
1111 * If this is 0, there is only one r1_bio and no locking
1112 * will be needed when requests complete. If it is
1113 * non-zero, then it is the number of not-completed requests.
1115 bio
->bi_phys_segments
= 0;
1116 bio_clear_flag(bio
, BIO_SEG_VALID
);
1120 * read balancing logic:
1125 rdisk
= read_balance(conf
, r1_bio
, &max_sectors
);
1128 /* couldn't find anywhere to read from */
1129 raid_end_bio_io(r1_bio
);
1132 mirror
= conf
->mirrors
+ rdisk
;
1134 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
1136 /* Reading from a write-mostly device must
1137 * take care not to over-take any writes
1140 wait_event(bitmap
->behind_wait
,
1141 atomic_read(&bitmap
->behind_writes
) == 0);
1143 r1_bio
->read_disk
= rdisk
;
1144 r1_bio
->start_next_window
= 0;
1146 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1147 bio_trim(read_bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
1150 r1_bio
->bios
[rdisk
] = read_bio
;
1152 read_bio
->bi_iter
.bi_sector
= r1_bio
->sector
+
1153 mirror
->rdev
->data_offset
;
1154 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
1155 read_bio
->bi_end_io
= raid1_end_read_request
;
1156 bio_set_op_attrs(read_bio
, op
, do_sync
);
1157 read_bio
->bi_private
= r1_bio
;
1159 if (max_sectors
< r1_bio
->sectors
) {
1160 /* could not read all from this device, so we will
1161 * need another r1_bio.
1164 sectors_handled
= (r1_bio
->sector
+ max_sectors
1165 - bio
->bi_iter
.bi_sector
);
1166 r1_bio
->sectors
= max_sectors
;
1167 spin_lock_irq(&conf
->device_lock
);
1168 if (bio
->bi_phys_segments
== 0)
1169 bio
->bi_phys_segments
= 2;
1171 bio
->bi_phys_segments
++;
1172 spin_unlock_irq(&conf
->device_lock
);
1173 /* Cannot call generic_make_request directly
1174 * as that will be queued in __make_request
1175 * and subsequent mempool_alloc might block waiting
1176 * for it. So hand bio over to raid1d.
1178 reschedule_retry(r1_bio
);
1180 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1182 r1_bio
->master_bio
= bio
;
1183 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1185 r1_bio
->mddev
= mddev
;
1186 r1_bio
->sector
= bio
->bi_iter
.bi_sector
+
1190 generic_make_request(read_bio
);
1197 if (conf
->pending_count
>= max_queued_requests
) {
1198 md_wakeup_thread(mddev
->thread
);
1199 wait_event(conf
->wait_barrier
,
1200 conf
->pending_count
< max_queued_requests
);
1202 /* first select target devices under rcu_lock and
1203 * inc refcount on their rdev. Record them by setting
1205 * If there are known/acknowledged bad blocks on any device on
1206 * which we have seen a write error, we want to avoid writing those
1208 * This potentially requires several writes to write around
1209 * the bad blocks. Each set of writes gets it's own r1bio
1210 * with a set of bios attached.
1213 disks
= conf
->raid_disks
* 2;
1215 r1_bio
->start_next_window
= start_next_window
;
1216 blocked_rdev
= NULL
;
1218 max_sectors
= r1_bio
->sectors
;
1219 for (i
= 0; i
< disks
; i
++) {
1220 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1221 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1222 atomic_inc(&rdev
->nr_pending
);
1223 blocked_rdev
= rdev
;
1226 r1_bio
->bios
[i
] = NULL
;
1227 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
1228 if (i
< conf
->raid_disks
)
1229 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1233 atomic_inc(&rdev
->nr_pending
);
1234 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1239 is_bad
= is_badblock(rdev
, r1_bio
->sector
,
1241 &first_bad
, &bad_sectors
);
1243 /* mustn't write here until the bad block is
1245 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1246 blocked_rdev
= rdev
;
1249 if (is_bad
&& first_bad
<= r1_bio
->sector
) {
1250 /* Cannot write here at all */
1251 bad_sectors
-= (r1_bio
->sector
- first_bad
);
1252 if (bad_sectors
< max_sectors
)
1253 /* mustn't write more than bad_sectors
1254 * to other devices yet
1256 max_sectors
= bad_sectors
;
1257 rdev_dec_pending(rdev
, mddev
);
1258 /* We don't set R1BIO_Degraded as that
1259 * only applies if the disk is
1260 * missing, so it might be re-added,
1261 * and we want to know to recover this
1263 * In this case the device is here,
1264 * and the fact that this chunk is not
1265 * in-sync is recorded in the bad
1271 int good_sectors
= first_bad
- r1_bio
->sector
;
1272 if (good_sectors
< max_sectors
)
1273 max_sectors
= good_sectors
;
1276 r1_bio
->bios
[i
] = bio
;
1280 if (unlikely(blocked_rdev
)) {
1281 /* Wait for this device to become unblocked */
1283 sector_t old
= start_next_window
;
1285 for (j
= 0; j
< i
; j
++)
1286 if (r1_bio
->bios
[j
])
1287 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
1289 allow_barrier(conf
, start_next_window
, bio
->bi_iter
.bi_sector
);
1290 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1291 start_next_window
= wait_barrier(conf
, bio
);
1293 * We must make sure the multi r1bios of bio have
1294 * the same value of bi_phys_segments
1296 if (bio
->bi_phys_segments
&& old
&&
1297 old
!= start_next_window
)
1298 /* Wait for the former r1bio(s) to complete */
1299 wait_event(conf
->wait_barrier
,
1300 bio
->bi_phys_segments
== 1);
1304 if (max_sectors
< r1_bio
->sectors
) {
1305 /* We are splitting this write into multiple parts, so
1306 * we need to prepare for allocating another r1_bio.
1308 r1_bio
->sectors
= max_sectors
;
1309 spin_lock_irq(&conf
->device_lock
);
1310 if (bio
->bi_phys_segments
== 0)
1311 bio
->bi_phys_segments
= 2;
1313 bio
->bi_phys_segments
++;
1314 spin_unlock_irq(&conf
->device_lock
);
1316 sectors_handled
= r1_bio
->sector
+ max_sectors
- bio
->bi_iter
.bi_sector
;
1318 atomic_set(&r1_bio
->remaining
, 1);
1319 atomic_set(&r1_bio
->behind_remaining
, 0);
1322 for (i
= 0; i
< disks
; i
++) {
1324 if (!r1_bio
->bios
[i
])
1327 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1328 bio_trim(mbio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
1332 * Not if there are too many, or cannot
1333 * allocate memory, or a reader on WriteMostly
1334 * is waiting for behind writes to flush */
1336 (atomic_read(&bitmap
->behind_writes
)
1337 < mddev
->bitmap_info
.max_write_behind
) &&
1338 !waitqueue_active(&bitmap
->behind_wait
))
1339 alloc_behind_pages(mbio
, r1_bio
);
1341 bitmap_startwrite(bitmap
, r1_bio
->sector
,
1343 test_bit(R1BIO_BehindIO
,
1347 if (r1_bio
->behind_bvecs
) {
1348 struct bio_vec
*bvec
;
1352 * We trimmed the bio, so _all is legit
1354 bio_for_each_segment_all(bvec
, mbio
, j
)
1355 bvec
->bv_page
= r1_bio
->behind_bvecs
[j
].bv_page
;
1356 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
1357 atomic_inc(&r1_bio
->behind_remaining
);
1360 r1_bio
->bios
[i
] = mbio
;
1362 mbio
->bi_iter
.bi_sector
= (r1_bio
->sector
+
1363 conf
->mirrors
[i
].rdev
->data_offset
);
1364 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1365 mbio
->bi_end_io
= raid1_end_write_request
;
1366 bio_set_op_attrs(mbio
, op
, do_flush_fua
| do_sync
);
1367 mbio
->bi_private
= r1_bio
;
1369 atomic_inc(&r1_bio
->remaining
);
1371 cb
= blk_check_plugged(raid1_unplug
, mddev
, sizeof(*plug
));
1373 plug
= container_of(cb
, struct raid1_plug_cb
, cb
);
1376 spin_lock_irqsave(&conf
->device_lock
, flags
);
1378 bio_list_add(&plug
->pending
, mbio
);
1379 plug
->pending_cnt
++;
1381 bio_list_add(&conf
->pending_bio_list
, mbio
);
1382 conf
->pending_count
++;
1384 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1386 md_wakeup_thread(mddev
->thread
);
1388 /* Mustn't call r1_bio_write_done before this next test,
1389 * as it could result in the bio being freed.
1391 if (sectors_handled
< bio_sectors(bio
)) {
1392 r1_bio_write_done(r1_bio
);
1393 /* We need another r1_bio. It has already been counted
1394 * in bio->bi_phys_segments
1396 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1397 r1_bio
->master_bio
= bio
;
1398 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1400 r1_bio
->mddev
= mddev
;
1401 r1_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1405 r1_bio_write_done(r1_bio
);
1407 /* In case raid1d snuck in to freeze_array */
1408 wake_up(&conf
->wait_barrier
);
1411 static void raid1_status(struct seq_file
*seq
, struct mddev
*mddev
)
1413 struct r1conf
*conf
= mddev
->private;
1416 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1417 conf
->raid_disks
- mddev
->degraded
);
1419 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1420 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1421 seq_printf(seq
, "%s",
1422 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1425 seq_printf(seq
, "]");
1428 static void raid1_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1430 char b
[BDEVNAME_SIZE
];
1431 struct r1conf
*conf
= mddev
->private;
1432 unsigned long flags
;
1435 * If it is not operational, then we have already marked it as dead
1436 * else if it is the last working disks, ignore the error, let the
1437 * next level up know.
1438 * else mark the drive as failed
1440 if (test_bit(In_sync
, &rdev
->flags
)
1441 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1443 * Don't fail the drive, act as though we were just a
1444 * normal single drive.
1445 * However don't try a recovery from this drive as
1446 * it is very likely to fail.
1448 conf
->recovery_disabled
= mddev
->recovery_disabled
;
1451 set_bit(Blocked
, &rdev
->flags
);
1452 spin_lock_irqsave(&conf
->device_lock
, flags
);
1453 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1455 set_bit(Faulty
, &rdev
->flags
);
1457 set_bit(Faulty
, &rdev
->flags
);
1458 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1460 * if recovery is running, make sure it aborts.
1462 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1463 set_mask_bits(&mddev
->flags
, 0,
1464 BIT(MD_CHANGE_DEVS
) | BIT(MD_CHANGE_PENDING
));
1466 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1467 "md/raid1:%s: Operation continuing on %d devices.\n",
1468 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1469 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1472 static void print_conf(struct r1conf
*conf
)
1476 printk(KERN_DEBUG
"RAID1 conf printout:\n");
1478 printk(KERN_DEBUG
"(!conf)\n");
1481 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1485 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1486 char b
[BDEVNAME_SIZE
];
1487 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1489 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1490 i
, !test_bit(In_sync
, &rdev
->flags
),
1491 !test_bit(Faulty
, &rdev
->flags
),
1492 bdevname(rdev
->bdev
,b
));
1497 static void close_sync(struct r1conf
*conf
)
1499 wait_barrier(conf
, NULL
);
1500 allow_barrier(conf
, 0, 0);
1502 mempool_destroy(conf
->r1buf_pool
);
1503 conf
->r1buf_pool
= NULL
;
1505 spin_lock_irq(&conf
->resync_lock
);
1506 conf
->next_resync
= MaxSector
- 2 * NEXT_NORMALIO_DISTANCE
;
1507 conf
->start_next_window
= MaxSector
;
1508 conf
->current_window_requests
+=
1509 conf
->next_window_requests
;
1510 conf
->next_window_requests
= 0;
1511 spin_unlock_irq(&conf
->resync_lock
);
1514 static int raid1_spare_active(struct mddev
*mddev
)
1517 struct r1conf
*conf
= mddev
->private;
1519 unsigned long flags
;
1522 * Find all failed disks within the RAID1 configuration
1523 * and mark them readable.
1524 * Called under mddev lock, so rcu protection not needed.
1525 * device_lock used to avoid races with raid1_end_read_request
1526 * which expects 'In_sync' flags and ->degraded to be consistent.
1528 spin_lock_irqsave(&conf
->device_lock
, flags
);
1529 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1530 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1531 struct md_rdev
*repl
= conf
->mirrors
[conf
->raid_disks
+ i
].rdev
;
1533 && !test_bit(Candidate
, &repl
->flags
)
1534 && repl
->recovery_offset
== MaxSector
1535 && !test_bit(Faulty
, &repl
->flags
)
1536 && !test_and_set_bit(In_sync
, &repl
->flags
)) {
1537 /* replacement has just become active */
1539 !test_and_clear_bit(In_sync
, &rdev
->flags
))
1542 /* Replaced device not technically
1543 * faulty, but we need to be sure
1544 * it gets removed and never re-added
1546 set_bit(Faulty
, &rdev
->flags
);
1547 sysfs_notify_dirent_safe(
1552 && rdev
->recovery_offset
== MaxSector
1553 && !test_bit(Faulty
, &rdev
->flags
)
1554 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1556 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
1559 mddev
->degraded
-= count
;
1560 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1566 static int raid1_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1568 struct r1conf
*conf
= mddev
->private;
1571 struct raid1_info
*p
;
1573 int last
= conf
->raid_disks
- 1;
1575 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
1578 if (md_integrity_add_rdev(rdev
, mddev
))
1581 if (rdev
->raid_disk
>= 0)
1582 first
= last
= rdev
->raid_disk
;
1585 * find the disk ... but prefer rdev->saved_raid_disk
1588 if (rdev
->saved_raid_disk
>= 0 &&
1589 rdev
->saved_raid_disk
>= first
&&
1590 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1591 first
= last
= rdev
->saved_raid_disk
;
1593 for (mirror
= first
; mirror
<= last
; mirror
++) {
1594 p
= conf
->mirrors
+mirror
;
1598 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1599 rdev
->data_offset
<< 9);
1601 p
->head_position
= 0;
1602 rdev
->raid_disk
= mirror
;
1604 /* As all devices are equivalent, we don't need a full recovery
1605 * if this was recently any drive of the array
1607 if (rdev
->saved_raid_disk
< 0)
1609 rcu_assign_pointer(p
->rdev
, rdev
);
1612 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
1613 p
[conf
->raid_disks
].rdev
== NULL
) {
1614 /* Add this device as a replacement */
1615 clear_bit(In_sync
, &rdev
->flags
);
1616 set_bit(Replacement
, &rdev
->flags
);
1617 rdev
->raid_disk
= mirror
;
1620 rcu_assign_pointer(p
[conf
->raid_disks
].rdev
, rdev
);
1624 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1625 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1630 static int raid1_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1632 struct r1conf
*conf
= mddev
->private;
1634 int number
= rdev
->raid_disk
;
1635 struct raid1_info
*p
= conf
->mirrors
+ number
;
1637 if (rdev
!= p
->rdev
)
1638 p
= conf
->mirrors
+ conf
->raid_disks
+ number
;
1641 if (rdev
== p
->rdev
) {
1642 if (test_bit(In_sync
, &rdev
->flags
) ||
1643 atomic_read(&rdev
->nr_pending
)) {
1647 /* Only remove non-faulty devices if recovery
1650 if (!test_bit(Faulty
, &rdev
->flags
) &&
1651 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
1652 mddev
->degraded
< conf
->raid_disks
) {
1657 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
1659 if (atomic_read(&rdev
->nr_pending
)) {
1660 /* lost the race, try later */
1666 if (conf
->mirrors
[conf
->raid_disks
+ number
].rdev
) {
1667 /* We just removed a device that is being replaced.
1668 * Move down the replacement. We drain all IO before
1669 * doing this to avoid confusion.
1671 struct md_rdev
*repl
=
1672 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
;
1673 freeze_array(conf
, 0);
1674 clear_bit(Replacement
, &repl
->flags
);
1676 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
= NULL
;
1677 unfreeze_array(conf
);
1678 clear_bit(WantReplacement
, &rdev
->flags
);
1680 clear_bit(WantReplacement
, &rdev
->flags
);
1681 err
= md_integrity_register(mddev
);
1689 static void end_sync_read(struct bio
*bio
)
1691 struct r1bio
*r1_bio
= bio
->bi_private
;
1693 update_head_pos(r1_bio
->read_disk
, r1_bio
);
1696 * we have read a block, now it needs to be re-written,
1697 * or re-read if the read failed.
1698 * We don't do much here, just schedule handling by raid1d
1701 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1703 if (atomic_dec_and_test(&r1_bio
->remaining
))
1704 reschedule_retry(r1_bio
);
1707 static void end_sync_write(struct bio
*bio
)
1709 int uptodate
= !bio
->bi_error
;
1710 struct r1bio
*r1_bio
= bio
->bi_private
;
1711 struct mddev
*mddev
= r1_bio
->mddev
;
1712 struct r1conf
*conf
= mddev
->private;
1715 struct md_rdev
*rdev
= conf
->mirrors
[find_bio_disk(r1_bio
, bio
)].rdev
;
1718 sector_t sync_blocks
= 0;
1719 sector_t s
= r1_bio
->sector
;
1720 long sectors_to_go
= r1_bio
->sectors
;
1721 /* make sure these bits doesn't get cleared. */
1723 bitmap_end_sync(mddev
->bitmap
, s
,
1726 sectors_to_go
-= sync_blocks
;
1727 } while (sectors_to_go
> 0);
1728 set_bit(WriteErrorSeen
, &rdev
->flags
);
1729 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1730 set_bit(MD_RECOVERY_NEEDED
, &
1732 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
1733 } else if (is_badblock(rdev
, r1_bio
->sector
, r1_bio
->sectors
,
1734 &first_bad
, &bad_sectors
) &&
1735 !is_badblock(conf
->mirrors
[r1_bio
->read_disk
].rdev
,
1738 &first_bad
, &bad_sectors
)
1740 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
1742 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1743 int s
= r1_bio
->sectors
;
1744 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
1745 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1746 reschedule_retry(r1_bio
);
1749 md_done_sync(mddev
, s
, uptodate
);
1754 static int r1_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1755 int sectors
, struct page
*page
, int rw
)
1757 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, 0, false))
1761 set_bit(WriteErrorSeen
, &rdev
->flags
);
1762 if (!test_and_set_bit(WantReplacement
,
1764 set_bit(MD_RECOVERY_NEEDED
, &
1765 rdev
->mddev
->recovery
);
1767 /* need to record an error - either for the block or the device */
1768 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1769 md_error(rdev
->mddev
, rdev
);
1773 static int fix_sync_read_error(struct r1bio
*r1_bio
)
1775 /* Try some synchronous reads of other devices to get
1776 * good data, much like with normal read errors. Only
1777 * read into the pages we already have so we don't
1778 * need to re-issue the read request.
1779 * We don't need to freeze the array, because being in an
1780 * active sync request, there is no normal IO, and
1781 * no overlapping syncs.
1782 * We don't need to check is_badblock() again as we
1783 * made sure that anything with a bad block in range
1784 * will have bi_end_io clear.
1786 struct mddev
*mddev
= r1_bio
->mddev
;
1787 struct r1conf
*conf
= mddev
->private;
1788 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1789 sector_t sect
= r1_bio
->sector
;
1790 int sectors
= r1_bio
->sectors
;
1795 int d
= r1_bio
->read_disk
;
1797 struct md_rdev
*rdev
;
1800 if (s
> (PAGE_SIZE
>>9))
1803 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1804 /* No rcu protection needed here devices
1805 * can only be removed when no resync is
1806 * active, and resync is currently active
1808 rdev
= conf
->mirrors
[d
].rdev
;
1809 if (sync_page_io(rdev
, sect
, s
<<9,
1810 bio
->bi_io_vec
[idx
].bv_page
,
1811 REQ_OP_READ
, 0, false)) {
1817 if (d
== conf
->raid_disks
* 2)
1819 } while (!success
&& d
!= r1_bio
->read_disk
);
1822 char b
[BDEVNAME_SIZE
];
1824 /* Cannot read from anywhere, this block is lost.
1825 * Record a bad block on each device. If that doesn't
1826 * work just disable and interrupt the recovery.
1827 * Don't fail devices as that won't really help.
1829 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O read error"
1830 " for block %llu\n",
1832 bdevname(bio
->bi_bdev
, b
),
1833 (unsigned long long)r1_bio
->sector
);
1834 for (d
= 0; d
< conf
->raid_disks
* 2; d
++) {
1835 rdev
= conf
->mirrors
[d
].rdev
;
1836 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
1838 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
1842 conf
->recovery_disabled
=
1843 mddev
->recovery_disabled
;
1844 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1845 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1857 /* write it back and re-read */
1858 while (d
!= r1_bio
->read_disk
) {
1860 d
= conf
->raid_disks
* 2;
1862 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1864 rdev
= conf
->mirrors
[d
].rdev
;
1865 if (r1_sync_page_io(rdev
, sect
, s
,
1866 bio
->bi_io_vec
[idx
].bv_page
,
1868 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
1869 rdev_dec_pending(rdev
, mddev
);
1873 while (d
!= r1_bio
->read_disk
) {
1875 d
= conf
->raid_disks
* 2;
1877 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1879 rdev
= conf
->mirrors
[d
].rdev
;
1880 if (r1_sync_page_io(rdev
, sect
, s
,
1881 bio
->bi_io_vec
[idx
].bv_page
,
1883 atomic_add(s
, &rdev
->corrected_errors
);
1889 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1894 static void process_checks(struct r1bio
*r1_bio
)
1896 /* We have read all readable devices. If we haven't
1897 * got the block, then there is no hope left.
1898 * If we have, then we want to do a comparison
1899 * and skip the write if everything is the same.
1900 * If any blocks failed to read, then we need to
1901 * attempt an over-write
1903 struct mddev
*mddev
= r1_bio
->mddev
;
1904 struct r1conf
*conf
= mddev
->private;
1909 /* Fix variable parts of all bios */
1910 vcnt
= (r1_bio
->sectors
+ PAGE_SIZE
/ 512 - 1) >> (PAGE_SHIFT
- 9);
1911 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
1915 struct bio
*b
= r1_bio
->bios
[i
];
1916 if (b
->bi_end_io
!= end_sync_read
)
1918 /* fixup the bio for reuse, but preserve errno */
1919 error
= b
->bi_error
;
1921 b
->bi_error
= error
;
1923 b
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
1924 b
->bi_iter
.bi_sector
= r1_bio
->sector
+
1925 conf
->mirrors
[i
].rdev
->data_offset
;
1926 b
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1927 b
->bi_end_io
= end_sync_read
;
1928 b
->bi_private
= r1_bio
;
1930 size
= b
->bi_iter
.bi_size
;
1931 for (j
= 0; j
< vcnt
; j
++) {
1933 bi
= &b
->bi_io_vec
[j
];
1935 if (size
> PAGE_SIZE
)
1936 bi
->bv_len
= PAGE_SIZE
;
1942 for (primary
= 0; primary
< conf
->raid_disks
* 2; primary
++)
1943 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
1944 !r1_bio
->bios
[primary
]->bi_error
) {
1945 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
1946 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
1949 r1_bio
->read_disk
= primary
;
1950 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
1952 struct bio
*pbio
= r1_bio
->bios
[primary
];
1953 struct bio
*sbio
= r1_bio
->bios
[i
];
1954 int error
= sbio
->bi_error
;
1956 if (sbio
->bi_end_io
!= end_sync_read
)
1958 /* Now we can 'fixup' the error value */
1962 for (j
= vcnt
; j
-- ; ) {
1964 p
= pbio
->bi_io_vec
[j
].bv_page
;
1965 s
= sbio
->bi_io_vec
[j
].bv_page
;
1966 if (memcmp(page_address(p
),
1968 sbio
->bi_io_vec
[j
].bv_len
))
1974 atomic64_add(r1_bio
->sectors
, &mddev
->resync_mismatches
);
1975 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
1977 /* No need to write to this device. */
1978 sbio
->bi_end_io
= NULL
;
1979 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
1983 bio_copy_data(sbio
, pbio
);
1987 static void sync_request_write(struct mddev
*mddev
, struct r1bio
*r1_bio
)
1989 struct r1conf
*conf
= mddev
->private;
1991 int disks
= conf
->raid_disks
* 2;
1992 struct bio
*bio
, *wbio
;
1994 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1996 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
1997 /* ouch - failed to read all of that. */
1998 if (!fix_sync_read_error(r1_bio
))
2001 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2002 process_checks(r1_bio
);
2007 atomic_set(&r1_bio
->remaining
, 1);
2008 for (i
= 0; i
< disks
; i
++) {
2009 wbio
= r1_bio
->bios
[i
];
2010 if (wbio
->bi_end_io
== NULL
||
2011 (wbio
->bi_end_io
== end_sync_read
&&
2012 (i
== r1_bio
->read_disk
||
2013 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
2016 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2017 wbio
->bi_end_io
= end_sync_write
;
2018 atomic_inc(&r1_bio
->remaining
);
2019 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, bio_sectors(wbio
));
2021 generic_make_request(wbio
);
2024 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
2025 /* if we're here, all write(s) have completed, so clean up */
2026 int s
= r1_bio
->sectors
;
2027 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2028 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2029 reschedule_retry(r1_bio
);
2032 md_done_sync(mddev
, s
, 1);
2038 * This is a kernel thread which:
2040 * 1. Retries failed read operations on working mirrors.
2041 * 2. Updates the raid superblock when problems encounter.
2042 * 3. Performs writes following reads for array synchronising.
2045 static void fix_read_error(struct r1conf
*conf
, int read_disk
,
2046 sector_t sect
, int sectors
)
2048 struct mddev
*mddev
= conf
->mddev
;
2054 struct md_rdev
*rdev
;
2056 if (s
> (PAGE_SIZE
>>9))
2064 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2066 (test_bit(In_sync
, &rdev
->flags
) ||
2067 (!test_bit(Faulty
, &rdev
->flags
) &&
2068 rdev
->recovery_offset
>= sect
+ s
)) &&
2069 is_badblock(rdev
, sect
, s
,
2070 &first_bad
, &bad_sectors
) == 0) {
2071 atomic_inc(&rdev
->nr_pending
);
2073 if (sync_page_io(rdev
, sect
, s
<<9,
2074 conf
->tmppage
, REQ_OP_READ
, 0, false))
2076 rdev_dec_pending(rdev
, mddev
);
2082 if (d
== conf
->raid_disks
* 2)
2084 } while (!success
&& d
!= read_disk
);
2087 /* Cannot read from anywhere - mark it bad */
2088 struct md_rdev
*rdev
= conf
->mirrors
[read_disk
].rdev
;
2089 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
2090 md_error(mddev
, rdev
);
2093 /* write it back and re-read */
2095 while (d
!= read_disk
) {
2097 d
= conf
->raid_disks
* 2;
2100 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2102 !test_bit(Faulty
, &rdev
->flags
)) {
2103 atomic_inc(&rdev
->nr_pending
);
2105 r1_sync_page_io(rdev
, sect
, s
,
2106 conf
->tmppage
, WRITE
);
2107 rdev_dec_pending(rdev
, mddev
);
2112 while (d
!= read_disk
) {
2113 char b
[BDEVNAME_SIZE
];
2115 d
= conf
->raid_disks
* 2;
2118 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2120 !test_bit(Faulty
, &rdev
->flags
)) {
2121 atomic_inc(&rdev
->nr_pending
);
2123 if (r1_sync_page_io(rdev
, sect
, s
,
2124 conf
->tmppage
, READ
)) {
2125 atomic_add(s
, &rdev
->corrected_errors
);
2127 "md/raid1:%s: read error corrected "
2128 "(%d sectors at %llu on %s)\n",
2130 (unsigned long long)(sect
+
2132 bdevname(rdev
->bdev
, b
));
2134 rdev_dec_pending(rdev
, mddev
);
2143 static int narrow_write_error(struct r1bio
*r1_bio
, int i
)
2145 struct mddev
*mddev
= r1_bio
->mddev
;
2146 struct r1conf
*conf
= mddev
->private;
2147 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2149 /* bio has the data to be written to device 'i' where
2150 * we just recently had a write error.
2151 * We repeatedly clone the bio and trim down to one block,
2152 * then try the write. Where the write fails we record
2154 * It is conceivable that the bio doesn't exactly align with
2155 * blocks. We must handle this somehow.
2157 * We currently own a reference on the rdev.
2163 int sect_to_write
= r1_bio
->sectors
;
2166 if (rdev
->badblocks
.shift
< 0)
2169 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2170 bdev_logical_block_size(rdev
->bdev
) >> 9);
2171 sector
= r1_bio
->sector
;
2172 sectors
= ((sector
+ block_sectors
)
2173 & ~(sector_t
)(block_sectors
- 1))
2176 while (sect_to_write
) {
2178 if (sectors
> sect_to_write
)
2179 sectors
= sect_to_write
;
2180 /* Write at 'sector' for 'sectors'*/
2182 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
2183 unsigned vcnt
= r1_bio
->behind_page_count
;
2184 struct bio_vec
*vec
= r1_bio
->behind_bvecs
;
2186 while (!vec
->bv_page
) {
2191 wbio
= bio_alloc_mddev(GFP_NOIO
, vcnt
, mddev
);
2192 memcpy(wbio
->bi_io_vec
, vec
, vcnt
* sizeof(struct bio_vec
));
2194 wbio
->bi_vcnt
= vcnt
;
2196 wbio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2199 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2200 wbio
->bi_iter
.bi_sector
= r1_bio
->sector
;
2201 wbio
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
2203 bio_trim(wbio
, sector
- r1_bio
->sector
, sectors
);
2204 wbio
->bi_iter
.bi_sector
+= rdev
->data_offset
;
2205 wbio
->bi_bdev
= rdev
->bdev
;
2207 if (submit_bio_wait(wbio
) < 0)
2209 ok
= rdev_set_badblocks(rdev
, sector
,
2214 sect_to_write
-= sectors
;
2216 sectors
= block_sectors
;
2221 static void handle_sync_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2224 int s
= r1_bio
->sectors
;
2225 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++) {
2226 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2227 struct bio
*bio
= r1_bio
->bios
[m
];
2228 if (bio
->bi_end_io
== NULL
)
2230 if (!bio
->bi_error
&&
2231 test_bit(R1BIO_MadeGood
, &r1_bio
->state
)) {
2232 rdev_clear_badblocks(rdev
, r1_bio
->sector
, s
, 0);
2234 if (bio
->bi_error
&&
2235 test_bit(R1BIO_WriteError
, &r1_bio
->state
)) {
2236 if (!rdev_set_badblocks(rdev
, r1_bio
->sector
, s
, 0))
2237 md_error(conf
->mddev
, rdev
);
2241 md_done_sync(conf
->mddev
, s
, 1);
2244 static void handle_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2248 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++)
2249 if (r1_bio
->bios
[m
] == IO_MADE_GOOD
) {
2250 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2251 rdev_clear_badblocks(rdev
,
2253 r1_bio
->sectors
, 0);
2254 rdev_dec_pending(rdev
, conf
->mddev
);
2255 } else if (r1_bio
->bios
[m
] != NULL
) {
2256 /* This drive got a write error. We need to
2257 * narrow down and record precise write
2261 if (!narrow_write_error(r1_bio
, m
)) {
2262 md_error(conf
->mddev
,
2263 conf
->mirrors
[m
].rdev
);
2264 /* an I/O failed, we can't clear the bitmap */
2265 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2267 rdev_dec_pending(conf
->mirrors
[m
].rdev
,
2271 spin_lock_irq(&conf
->device_lock
);
2272 list_add(&r1_bio
->retry_list
, &conf
->bio_end_io_list
);
2274 spin_unlock_irq(&conf
->device_lock
);
2275 md_wakeup_thread(conf
->mddev
->thread
);
2277 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2278 close_write(r1_bio
);
2279 raid_end_bio_io(r1_bio
);
2283 static void handle_read_error(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2287 struct mddev
*mddev
= conf
->mddev
;
2289 char b
[BDEVNAME_SIZE
];
2290 struct md_rdev
*rdev
;
2292 clear_bit(R1BIO_ReadError
, &r1_bio
->state
);
2293 /* we got a read error. Maybe the drive is bad. Maybe just
2294 * the block and we can fix it.
2295 * We freeze all other IO, and try reading the block from
2296 * other devices. When we find one, we re-write
2297 * and check it that fixes the read error.
2298 * This is all done synchronously while the array is
2301 if (mddev
->ro
== 0) {
2302 freeze_array(conf
, 1);
2303 fix_read_error(conf
, r1_bio
->read_disk
,
2304 r1_bio
->sector
, r1_bio
->sectors
);
2305 unfreeze_array(conf
);
2307 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
2308 rdev_dec_pending(conf
->mirrors
[r1_bio
->read_disk
].rdev
, conf
->mddev
);
2310 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2311 bdevname(bio
->bi_bdev
, b
);
2313 disk
= read_balance(conf
, r1_bio
, &max_sectors
);
2315 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O"
2316 " read error for block %llu\n",
2317 mdname(mddev
), b
, (unsigned long long)r1_bio
->sector
);
2318 raid_end_bio_io(r1_bio
);
2320 const unsigned long do_sync
2321 = r1_bio
->master_bio
->bi_opf
& REQ_SYNC
;
2323 r1_bio
->bios
[r1_bio
->read_disk
] =
2324 mddev
->ro
? IO_BLOCKED
: NULL
;
2327 r1_bio
->read_disk
= disk
;
2328 bio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2329 bio_trim(bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
2331 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
2332 rdev
= conf
->mirrors
[disk
].rdev
;
2333 printk_ratelimited(KERN_ERR
2334 "md/raid1:%s: redirecting sector %llu"
2335 " to other mirror: %s\n",
2337 (unsigned long long)r1_bio
->sector
,
2338 bdevname(rdev
->bdev
, b
));
2339 bio
->bi_iter
.bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
2340 bio
->bi_bdev
= rdev
->bdev
;
2341 bio
->bi_end_io
= raid1_end_read_request
;
2342 bio_set_op_attrs(bio
, REQ_OP_READ
, do_sync
);
2343 bio
->bi_private
= r1_bio
;
2344 if (max_sectors
< r1_bio
->sectors
) {
2345 /* Drat - have to split this up more */
2346 struct bio
*mbio
= r1_bio
->master_bio
;
2347 int sectors_handled
= (r1_bio
->sector
+ max_sectors
2348 - mbio
->bi_iter
.bi_sector
);
2349 r1_bio
->sectors
= max_sectors
;
2350 spin_lock_irq(&conf
->device_lock
);
2351 if (mbio
->bi_phys_segments
== 0)
2352 mbio
->bi_phys_segments
= 2;
2354 mbio
->bi_phys_segments
++;
2355 spin_unlock_irq(&conf
->device_lock
);
2356 generic_make_request(bio
);
2359 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
2361 r1_bio
->master_bio
= mbio
;
2362 r1_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2364 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
2365 r1_bio
->mddev
= mddev
;
2366 r1_bio
->sector
= mbio
->bi_iter
.bi_sector
+
2371 generic_make_request(bio
);
2375 static void raid1d(struct md_thread
*thread
)
2377 struct mddev
*mddev
= thread
->mddev
;
2378 struct r1bio
*r1_bio
;
2379 unsigned long flags
;
2380 struct r1conf
*conf
= mddev
->private;
2381 struct list_head
*head
= &conf
->retry_list
;
2382 struct blk_plug plug
;
2384 md_check_recovery(mddev
);
2386 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2387 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2389 spin_lock_irqsave(&conf
->device_lock
, flags
);
2390 if (!test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2391 while (!list_empty(&conf
->bio_end_io_list
)) {
2392 list_move(conf
->bio_end_io_list
.prev
, &tmp
);
2396 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2397 while (!list_empty(&tmp
)) {
2398 r1_bio
= list_first_entry(&tmp
, struct r1bio
,
2400 list_del(&r1_bio
->retry_list
);
2401 if (mddev
->degraded
)
2402 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2403 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2404 close_write(r1_bio
);
2405 raid_end_bio_io(r1_bio
);
2409 blk_start_plug(&plug
);
2412 flush_pending_writes(conf
);
2414 spin_lock_irqsave(&conf
->device_lock
, flags
);
2415 if (list_empty(head
)) {
2416 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2419 r1_bio
= list_entry(head
->prev
, struct r1bio
, retry_list
);
2420 list_del(head
->prev
);
2422 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2424 mddev
= r1_bio
->mddev
;
2425 conf
= mddev
->private;
2426 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
2427 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2428 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2429 handle_sync_write_finished(conf
, r1_bio
);
2431 sync_request_write(mddev
, r1_bio
);
2432 } else if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2433 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2434 handle_write_finished(conf
, r1_bio
);
2435 else if (test_bit(R1BIO_ReadError
, &r1_bio
->state
))
2436 handle_read_error(conf
, r1_bio
);
2438 /* just a partial read to be scheduled from separate
2441 generic_make_request(r1_bio
->bios
[r1_bio
->read_disk
]);
2444 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2445 md_check_recovery(mddev
);
2447 blk_finish_plug(&plug
);
2450 static int init_resync(struct r1conf
*conf
)
2454 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2455 BUG_ON(conf
->r1buf_pool
);
2456 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
2458 if (!conf
->r1buf_pool
)
2460 conf
->next_resync
= 0;
2465 * perform a "sync" on one "block"
2467 * We need to make sure that no normal I/O request - particularly write
2468 * requests - conflict with active sync requests.
2470 * This is achieved by tracking pending requests and a 'barrier' concept
2471 * that can be installed to exclude normal IO requests.
2474 static sector_t
raid1_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2477 struct r1conf
*conf
= mddev
->private;
2478 struct r1bio
*r1_bio
;
2480 sector_t max_sector
, nr_sectors
;
2484 int write_targets
= 0, read_targets
= 0;
2485 sector_t sync_blocks
;
2486 int still_degraded
= 0;
2487 int good_sectors
= RESYNC_SECTORS
;
2488 int min_bad
= 0; /* number of sectors that are bad in all devices */
2490 if (!conf
->r1buf_pool
)
2491 if (init_resync(conf
))
2494 max_sector
= mddev
->dev_sectors
;
2495 if (sector_nr
>= max_sector
) {
2496 /* If we aborted, we need to abort the
2497 * sync on the 'current' bitmap chunk (there will
2498 * only be one in raid1 resync.
2499 * We can find the current addess in mddev->curr_resync
2501 if (mddev
->curr_resync
< max_sector
) /* aborted */
2502 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2504 else /* completed sync */
2507 bitmap_close_sync(mddev
->bitmap
);
2510 if (mddev_is_clustered(mddev
)) {
2511 conf
->cluster_sync_low
= 0;
2512 conf
->cluster_sync_high
= 0;
2517 if (mddev
->bitmap
== NULL
&&
2518 mddev
->recovery_cp
== MaxSector
&&
2519 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2520 conf
->fullsync
== 0) {
2522 return max_sector
- sector_nr
;
2524 /* before building a request, check if we can skip these blocks..
2525 * This call the bitmap_start_sync doesn't actually record anything
2527 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
2528 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2529 /* We can skip this block, and probably several more */
2535 * If there is non-resync activity waiting for a turn, then let it
2536 * though before starting on this new sync request.
2538 if (conf
->nr_waiting
)
2539 schedule_timeout_uninterruptible(1);
2541 /* we are incrementing sector_nr below. To be safe, we check against
2542 * sector_nr + two times RESYNC_SECTORS
2545 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
,
2546 mddev_is_clustered(mddev
) && (sector_nr
+ 2 * RESYNC_SECTORS
> conf
->cluster_sync_high
));
2547 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
2549 raise_barrier(conf
, sector_nr
);
2553 * If we get a correctably read error during resync or recovery,
2554 * we might want to read from a different device. So we
2555 * flag all drives that could conceivably be read from for READ,
2556 * and any others (which will be non-In_sync devices) for WRITE.
2557 * If a read fails, we try reading from something else for which READ
2561 r1_bio
->mddev
= mddev
;
2562 r1_bio
->sector
= sector_nr
;
2564 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
2566 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2567 struct md_rdev
*rdev
;
2568 bio
= r1_bio
->bios
[i
];
2571 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
2573 test_bit(Faulty
, &rdev
->flags
)) {
2574 if (i
< conf
->raid_disks
)
2576 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
2577 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
2578 bio
->bi_end_io
= end_sync_write
;
2581 /* may need to read from here */
2582 sector_t first_bad
= MaxSector
;
2585 if (is_badblock(rdev
, sector_nr
, good_sectors
,
2586 &first_bad
, &bad_sectors
)) {
2587 if (first_bad
> sector_nr
)
2588 good_sectors
= first_bad
- sector_nr
;
2590 bad_sectors
-= (sector_nr
- first_bad
);
2592 min_bad
> bad_sectors
)
2593 min_bad
= bad_sectors
;
2596 if (sector_nr
< first_bad
) {
2597 if (test_bit(WriteMostly
, &rdev
->flags
)) {
2604 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
2605 bio
->bi_end_io
= end_sync_read
;
2607 } else if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
2608 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2609 !test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)) {
2611 * The device is suitable for reading (InSync),
2612 * but has bad block(s) here. Let's try to correct them,
2613 * if we are doing resync or repair. Otherwise, leave
2614 * this device alone for this sync request.
2616 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
2617 bio
->bi_end_io
= end_sync_write
;
2621 if (bio
->bi_end_io
) {
2622 atomic_inc(&rdev
->nr_pending
);
2623 bio
->bi_iter
.bi_sector
= sector_nr
+ rdev
->data_offset
;
2624 bio
->bi_bdev
= rdev
->bdev
;
2625 bio
->bi_private
= r1_bio
;
2631 r1_bio
->read_disk
= disk
;
2633 if (read_targets
== 0 && min_bad
> 0) {
2634 /* These sectors are bad on all InSync devices, so we
2635 * need to mark them bad on all write targets
2638 for (i
= 0 ; i
< conf
->raid_disks
* 2 ; i
++)
2639 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_write
) {
2640 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2641 ok
= rdev_set_badblocks(rdev
, sector_nr
,
2645 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2650 /* Cannot record the badblocks, so need to
2652 * If there are multiple read targets, could just
2653 * fail the really bad ones ???
2655 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2656 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2662 if (min_bad
> 0 && min_bad
< good_sectors
) {
2663 /* only resync enough to reach the next bad->good
2665 good_sectors
= min_bad
;
2668 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
2669 /* extra read targets are also write targets */
2670 write_targets
+= read_targets
-1;
2672 if (write_targets
== 0 || read_targets
== 0) {
2673 /* There is nowhere to write, so all non-sync
2674 * drives must be failed - so we are finished
2678 max_sector
= sector_nr
+ min_bad
;
2679 rv
= max_sector
- sector_nr
;
2685 if (max_sector
> mddev
->resync_max
)
2686 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2687 if (max_sector
> sector_nr
+ good_sectors
)
2688 max_sector
= sector_nr
+ good_sectors
;
2693 int len
= PAGE_SIZE
;
2694 if (sector_nr
+ (len
>>9) > max_sector
)
2695 len
= (max_sector
- sector_nr
) << 9;
2698 if (sync_blocks
== 0) {
2699 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2700 &sync_blocks
, still_degraded
) &&
2702 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2704 if ((len
>> 9) > sync_blocks
)
2705 len
= sync_blocks
<<9;
2708 for (i
= 0 ; i
< conf
->raid_disks
* 2; i
++) {
2709 bio
= r1_bio
->bios
[i
];
2710 if (bio
->bi_end_io
) {
2711 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2712 if (bio_add_page(bio
, page
, len
, 0) == 0) {
2714 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2717 bio
= r1_bio
->bios
[i
];
2718 if (bio
->bi_end_io
==NULL
)
2720 /* remove last page from this bio */
2722 bio
->bi_iter
.bi_size
-= len
;
2723 bio_clear_flag(bio
, BIO_SEG_VALID
);
2729 nr_sectors
+= len
>>9;
2730 sector_nr
+= len
>>9;
2731 sync_blocks
-= (len
>>9);
2732 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
2734 r1_bio
->sectors
= nr_sectors
;
2736 if (mddev_is_clustered(mddev
) &&
2737 conf
->cluster_sync_high
< sector_nr
+ nr_sectors
) {
2738 conf
->cluster_sync_low
= mddev
->curr_resync_completed
;
2739 conf
->cluster_sync_high
= conf
->cluster_sync_low
+ CLUSTER_RESYNC_WINDOW_SECTORS
;
2740 /* Send resync message */
2741 md_cluster_ops
->resync_info_update(mddev
,
2742 conf
->cluster_sync_low
,
2743 conf
->cluster_sync_high
);
2746 /* For a user-requested sync, we read all readable devices and do a
2749 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2750 atomic_set(&r1_bio
->remaining
, read_targets
);
2751 for (i
= 0; i
< conf
->raid_disks
* 2 && read_targets
; i
++) {
2752 bio
= r1_bio
->bios
[i
];
2753 if (bio
->bi_end_io
== end_sync_read
) {
2755 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2756 generic_make_request(bio
);
2760 atomic_set(&r1_bio
->remaining
, 1);
2761 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2762 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2763 generic_make_request(bio
);
2769 static sector_t
raid1_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2774 return mddev
->dev_sectors
;
2777 static struct r1conf
*setup_conf(struct mddev
*mddev
)
2779 struct r1conf
*conf
;
2781 struct raid1_info
*disk
;
2782 struct md_rdev
*rdev
;
2785 conf
= kzalloc(sizeof(struct r1conf
), GFP_KERNEL
);
2789 conf
->mirrors
= kzalloc(sizeof(struct raid1_info
)
2790 * mddev
->raid_disks
* 2,
2795 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2799 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
2800 if (!conf
->poolinfo
)
2802 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
* 2;
2803 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2806 if (!conf
->r1bio_pool
)
2809 conf
->poolinfo
->mddev
= mddev
;
2812 spin_lock_init(&conf
->device_lock
);
2813 rdev_for_each(rdev
, mddev
) {
2814 struct request_queue
*q
;
2815 int disk_idx
= rdev
->raid_disk
;
2816 if (disk_idx
>= mddev
->raid_disks
2819 if (test_bit(Replacement
, &rdev
->flags
))
2820 disk
= conf
->mirrors
+ mddev
->raid_disks
+ disk_idx
;
2822 disk
= conf
->mirrors
+ disk_idx
;
2827 q
= bdev_get_queue(rdev
->bdev
);
2829 disk
->head_position
= 0;
2830 disk
->seq_start
= MaxSector
;
2832 conf
->raid_disks
= mddev
->raid_disks
;
2833 conf
->mddev
= mddev
;
2834 INIT_LIST_HEAD(&conf
->retry_list
);
2835 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
2837 spin_lock_init(&conf
->resync_lock
);
2838 init_waitqueue_head(&conf
->wait_barrier
);
2840 bio_list_init(&conf
->pending_bio_list
);
2841 conf
->pending_count
= 0;
2842 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
2844 conf
->start_next_window
= MaxSector
;
2845 conf
->current_window_requests
= conf
->next_window_requests
= 0;
2848 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2850 disk
= conf
->mirrors
+ i
;
2852 if (i
< conf
->raid_disks
&&
2853 disk
[conf
->raid_disks
].rdev
) {
2854 /* This slot has a replacement. */
2856 /* No original, just make the replacement
2857 * a recovering spare
2860 disk
[conf
->raid_disks
].rdev
;
2861 disk
[conf
->raid_disks
].rdev
= NULL
;
2862 } else if (!test_bit(In_sync
, &disk
->rdev
->flags
))
2863 /* Original is not in_sync - bad */
2868 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2869 disk
->head_position
= 0;
2871 (disk
->rdev
->saved_raid_disk
< 0))
2877 conf
->thread
= md_register_thread(raid1d
, mddev
, "raid1");
2878 if (!conf
->thread
) {
2880 "md/raid1:%s: couldn't allocate thread\n",
2889 mempool_destroy(conf
->r1bio_pool
);
2890 kfree(conf
->mirrors
);
2891 safe_put_page(conf
->tmppage
);
2892 kfree(conf
->poolinfo
);
2895 return ERR_PTR(err
);
2898 static void raid1_free(struct mddev
*mddev
, void *priv
);
2899 static int raid1_run(struct mddev
*mddev
)
2901 struct r1conf
*conf
;
2903 struct md_rdev
*rdev
;
2905 bool discard_supported
= false;
2907 if (mddev
->level
!= 1) {
2908 printk(KERN_ERR
"md/raid1:%s: raid level not set to mirroring (%d)\n",
2909 mdname(mddev
), mddev
->level
);
2912 if (mddev
->reshape_position
!= MaxSector
) {
2913 printk(KERN_ERR
"md/raid1:%s: reshape_position set but not supported\n",
2918 * copy the already verified devices into our private RAID1
2919 * bookkeeping area. [whatever we allocate in run(),
2920 * should be freed in raid1_free()]
2922 if (mddev
->private == NULL
)
2923 conf
= setup_conf(mddev
);
2925 conf
= mddev
->private;
2928 return PTR_ERR(conf
);
2931 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
2933 rdev_for_each(rdev
, mddev
) {
2934 if (!mddev
->gendisk
)
2936 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2937 rdev
->data_offset
<< 9);
2938 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
2939 discard_supported
= true;
2942 mddev
->degraded
= 0;
2943 for (i
=0; i
< conf
->raid_disks
; i
++)
2944 if (conf
->mirrors
[i
].rdev
== NULL
||
2945 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
2946 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2949 if (conf
->raid_disks
- mddev
->degraded
== 1)
2950 mddev
->recovery_cp
= MaxSector
;
2952 if (mddev
->recovery_cp
!= MaxSector
)
2953 printk(KERN_NOTICE
"md/raid1:%s: not clean"
2954 " -- starting background reconstruction\n",
2957 "md/raid1:%s: active with %d out of %d mirrors\n",
2958 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2962 * Ok, everything is just fine now
2964 mddev
->thread
= conf
->thread
;
2965 conf
->thread
= NULL
;
2966 mddev
->private = conf
;
2968 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
2971 if (discard_supported
)
2972 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
2975 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
2979 ret
= md_integrity_register(mddev
);
2981 md_unregister_thread(&mddev
->thread
);
2982 raid1_free(mddev
, conf
);
2987 static void raid1_free(struct mddev
*mddev
, void *priv
)
2989 struct r1conf
*conf
= priv
;
2991 mempool_destroy(conf
->r1bio_pool
);
2992 kfree(conf
->mirrors
);
2993 safe_put_page(conf
->tmppage
);
2994 kfree(conf
->poolinfo
);
2998 static int raid1_resize(struct mddev
*mddev
, sector_t sectors
)
3000 /* no resync is happening, and there is enough space
3001 * on all devices, so we can resize.
3002 * We need to make sure resync covers any new space.
3003 * If the array is shrinking we should possibly wait until
3004 * any io in the removed space completes, but it hardly seems
3007 sector_t newsize
= raid1_size(mddev
, sectors
, 0);
3008 if (mddev
->external_size
&&
3009 mddev
->array_sectors
> newsize
)
3011 if (mddev
->bitmap
) {
3012 int ret
= bitmap_resize(mddev
->bitmap
, newsize
, 0, 0);
3016 md_set_array_sectors(mddev
, newsize
);
3017 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3018 revalidate_disk(mddev
->gendisk
);
3019 if (sectors
> mddev
->dev_sectors
&&
3020 mddev
->recovery_cp
> mddev
->dev_sectors
) {
3021 mddev
->recovery_cp
= mddev
->dev_sectors
;
3022 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3024 mddev
->dev_sectors
= sectors
;
3025 mddev
->resync_max_sectors
= sectors
;
3029 static int raid1_reshape(struct mddev
*mddev
)
3032 * 1/ resize the r1bio_pool
3033 * 2/ resize conf->mirrors
3035 * We allocate a new r1bio_pool if we can.
3036 * Then raise a device barrier and wait until all IO stops.
3037 * Then resize conf->mirrors and swap in the new r1bio pool.
3039 * At the same time, we "pack" the devices so that all the missing
3040 * devices have the higher raid_disk numbers.
3042 mempool_t
*newpool
, *oldpool
;
3043 struct pool_info
*newpoolinfo
;
3044 struct raid1_info
*newmirrors
;
3045 struct r1conf
*conf
= mddev
->private;
3046 int cnt
, raid_disks
;
3047 unsigned long flags
;
3050 /* Cannot change chunk_size, layout, or level */
3051 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
3052 mddev
->layout
!= mddev
->new_layout
||
3053 mddev
->level
!= mddev
->new_level
) {
3054 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3055 mddev
->new_layout
= mddev
->layout
;
3056 mddev
->new_level
= mddev
->level
;
3060 if (!mddev_is_clustered(mddev
)) {
3061 err
= md_allow_write(mddev
);
3066 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3068 if (raid_disks
< conf
->raid_disks
) {
3070 for (d
= 0; d
< conf
->raid_disks
; d
++)
3071 if (conf
->mirrors
[d
].rdev
)
3073 if (cnt
> raid_disks
)
3077 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
3080 newpoolinfo
->mddev
= mddev
;
3081 newpoolinfo
->raid_disks
= raid_disks
* 2;
3083 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
3084 r1bio_pool_free
, newpoolinfo
);
3089 newmirrors
= kzalloc(sizeof(struct raid1_info
) * raid_disks
* 2,
3093 mempool_destroy(newpool
);
3097 freeze_array(conf
, 0);
3099 /* ok, everything is stopped */
3100 oldpool
= conf
->r1bio_pool
;
3101 conf
->r1bio_pool
= newpool
;
3103 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
3104 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
3105 if (rdev
&& rdev
->raid_disk
!= d2
) {
3106 sysfs_unlink_rdev(mddev
, rdev
);
3107 rdev
->raid_disk
= d2
;
3108 sysfs_unlink_rdev(mddev
, rdev
);
3109 if (sysfs_link_rdev(mddev
, rdev
))
3111 "md/raid1:%s: cannot register rd%d\n",
3112 mdname(mddev
), rdev
->raid_disk
);
3115 newmirrors
[d2
++].rdev
= rdev
;
3117 kfree(conf
->mirrors
);
3118 conf
->mirrors
= newmirrors
;
3119 kfree(conf
->poolinfo
);
3120 conf
->poolinfo
= newpoolinfo
;
3122 spin_lock_irqsave(&conf
->device_lock
, flags
);
3123 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
3124 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3125 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
3126 mddev
->delta_disks
= 0;
3128 unfreeze_array(conf
);
3130 set_bit(MD_RECOVERY_RECOVER
, &mddev
->recovery
);
3131 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3132 md_wakeup_thread(mddev
->thread
);
3134 mempool_destroy(oldpool
);
3138 static void raid1_quiesce(struct mddev
*mddev
, int state
)
3140 struct r1conf
*conf
= mddev
->private;
3143 case 2: /* wake for suspend */
3144 wake_up(&conf
->wait_barrier
);
3147 freeze_array(conf
, 0);
3150 unfreeze_array(conf
);
3155 static void *raid1_takeover(struct mddev
*mddev
)
3157 /* raid1 can take over:
3158 * raid5 with 2 devices, any layout or chunk size
3160 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
3161 struct r1conf
*conf
;
3162 mddev
->new_level
= 1;
3163 mddev
->new_layout
= 0;
3164 mddev
->new_chunk_sectors
= 0;
3165 conf
= setup_conf(mddev
);
3167 /* Array must appear to be quiesced */
3168 conf
->array_frozen
= 1;
3171 return ERR_PTR(-EINVAL
);
3174 static struct md_personality raid1_personality
=
3178 .owner
= THIS_MODULE
,
3179 .make_request
= raid1_make_request
,
3182 .status
= raid1_status
,
3183 .error_handler
= raid1_error
,
3184 .hot_add_disk
= raid1_add_disk
,
3185 .hot_remove_disk
= raid1_remove_disk
,
3186 .spare_active
= raid1_spare_active
,
3187 .sync_request
= raid1_sync_request
,
3188 .resize
= raid1_resize
,
3190 .check_reshape
= raid1_reshape
,
3191 .quiesce
= raid1_quiesce
,
3192 .takeover
= raid1_takeover
,
3193 .congested
= raid1_congested
,
3196 static int __init
raid_init(void)
3198 return register_md_personality(&raid1_personality
);
3201 static void raid_exit(void)
3203 unregister_md_personality(&raid1_personality
);
3206 module_init(raid_init
);
3207 module_exit(raid_exit
);
3208 MODULE_LICENSE("GPL");
3209 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3210 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3211 MODULE_ALIAS("md-raid1");
3212 MODULE_ALIAS("md-level-1");
3214 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);