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
;
103 r1_bio
= r1bio_pool_alloc(gfp_flags
, pi
);
108 * Allocate bios : 1 for reading, n-1 for writing
110 for (j
= pi
->raid_disks
; j
-- ; ) {
111 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
114 r1_bio
->bios
[j
] = bio
;
117 * Allocate RESYNC_PAGES data pages and attach them to
119 * If this is a user-requested check/repair, allocate
120 * RESYNC_PAGES for each bio.
122 if (test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
))
123 need_pages
= pi
->raid_disks
;
126 for (j
= 0; j
< need_pages
; j
++) {
127 bio
= r1_bio
->bios
[j
];
128 bio
->bi_vcnt
= RESYNC_PAGES
;
130 if (bio_alloc_pages(bio
, gfp_flags
))
133 /* If not user-requests, copy the page pointers to all bios */
134 if (!test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
)) {
135 for (i
=0; i
<RESYNC_PAGES
; i
++)
136 for (j
=1; j
<pi
->raid_disks
; j
++)
137 r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
=
138 r1_bio
->bios
[0]->bi_io_vec
[i
].bv_page
;
141 r1_bio
->master_bio
= NULL
;
149 bio_for_each_segment_all(bv
, r1_bio
->bios
[j
], i
)
150 __free_page(bv
->bv_page
);
154 while (++j
< pi
->raid_disks
)
155 bio_put(r1_bio
->bios
[j
]);
156 r1bio_pool_free(r1_bio
, data
);
160 static void r1buf_pool_free(void *__r1_bio
, void *data
)
162 struct pool_info
*pi
= data
;
164 struct r1bio
*r1bio
= __r1_bio
;
166 for (i
= 0; i
< RESYNC_PAGES
; i
++)
167 for (j
= pi
->raid_disks
; j
-- ;) {
169 r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
!=
170 r1bio
->bios
[0]->bi_io_vec
[i
].bv_page
)
171 safe_put_page(r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
173 for (i
=0 ; i
< pi
->raid_disks
; i
++)
174 bio_put(r1bio
->bios
[i
]);
176 r1bio_pool_free(r1bio
, data
);
179 static void put_all_bios(struct r1conf
*conf
, struct r1bio
*r1_bio
)
183 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
184 struct bio
**bio
= r1_bio
->bios
+ i
;
185 if (!BIO_SPECIAL(*bio
))
191 static void free_r1bio(struct r1bio
*r1_bio
)
193 struct r1conf
*conf
= r1_bio
->mddev
->private;
195 put_all_bios(conf
, r1_bio
);
196 mempool_free(r1_bio
, conf
->r1bio_pool
);
199 static void put_buf(struct r1bio
*r1_bio
)
201 struct r1conf
*conf
= r1_bio
->mddev
->private;
204 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
205 struct bio
*bio
= r1_bio
->bios
[i
];
207 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
210 mempool_free(r1_bio
, conf
->r1buf_pool
);
215 static void reschedule_retry(struct r1bio
*r1_bio
)
218 struct mddev
*mddev
= r1_bio
->mddev
;
219 struct r1conf
*conf
= mddev
->private;
221 spin_lock_irqsave(&conf
->device_lock
, flags
);
222 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
224 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
226 wake_up(&conf
->wait_barrier
);
227 md_wakeup_thread(mddev
->thread
);
231 * raid_end_bio_io() is called when we have finished servicing a mirrored
232 * operation and are ready to return a success/failure code to the buffer
235 static void call_bio_endio(struct r1bio
*r1_bio
)
237 struct bio
*bio
= r1_bio
->master_bio
;
239 struct r1conf
*conf
= r1_bio
->mddev
->private;
240 sector_t start_next_window
= r1_bio
->start_next_window
;
241 sector_t bi_sector
= bio
->bi_iter
.bi_sector
;
243 if (bio
->bi_phys_segments
) {
245 spin_lock_irqsave(&conf
->device_lock
, flags
);
246 bio
->bi_phys_segments
--;
247 done
= (bio
->bi_phys_segments
== 0);
248 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
250 * make_request() might be waiting for
251 * bi_phys_segments to decrease
253 wake_up(&conf
->wait_barrier
);
257 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
258 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
262 * Wake up any possible resync thread that waits for the device
265 allow_barrier(conf
, start_next_window
, bi_sector
);
269 static void raid_end_bio_io(struct r1bio
*r1_bio
)
271 struct bio
*bio
= r1_bio
->master_bio
;
273 /* if nobody has done the final endio yet, do it now */
274 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
275 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
276 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
277 (unsigned long long) bio
->bi_iter
.bi_sector
,
278 (unsigned long long) bio_end_sector(bio
) - 1);
280 call_bio_endio(r1_bio
);
286 * Update disk head position estimator based on IRQ completion info.
288 static inline void update_head_pos(int disk
, struct r1bio
*r1_bio
)
290 struct r1conf
*conf
= r1_bio
->mddev
->private;
292 conf
->mirrors
[disk
].head_position
=
293 r1_bio
->sector
+ (r1_bio
->sectors
);
297 * Find the disk number which triggered given bio
299 static int find_bio_disk(struct r1bio
*r1_bio
, struct bio
*bio
)
302 struct r1conf
*conf
= r1_bio
->mddev
->private;
303 int raid_disks
= conf
->raid_disks
;
305 for (mirror
= 0; mirror
< raid_disks
* 2; mirror
++)
306 if (r1_bio
->bios
[mirror
] == bio
)
309 BUG_ON(mirror
== raid_disks
* 2);
310 update_head_pos(mirror
, r1_bio
);
315 static void raid1_end_read_request(struct bio
*bio
, int error
)
317 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
318 struct r1bio
*r1_bio
= bio
->bi_private
;
320 struct r1conf
*conf
= r1_bio
->mddev
->private;
322 mirror
= r1_bio
->read_disk
;
324 * this branch is our 'one mirror IO has finished' event handler:
326 update_head_pos(mirror
, r1_bio
);
329 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
331 /* If all other devices have failed, we want to return
332 * the error upwards rather than fail the last device.
333 * Here we redefine "uptodate" to mean "Don't want to retry"
336 spin_lock_irqsave(&conf
->device_lock
, flags
);
337 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
338 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
339 test_bit(In_sync
, &conf
->mirrors
[mirror
].rdev
->flags
)))
341 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
345 raid_end_bio_io(r1_bio
);
346 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
351 char b
[BDEVNAME_SIZE
];
353 KERN_ERR
"md/raid1:%s: %s: "
354 "rescheduling sector %llu\n",
356 bdevname(conf
->mirrors
[mirror
].rdev
->bdev
,
358 (unsigned long long)r1_bio
->sector
);
359 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
360 reschedule_retry(r1_bio
);
361 /* don't drop the reference on read_disk yet */
365 static void close_write(struct r1bio
*r1_bio
)
367 /* it really is the end of this request */
368 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
369 /* free extra copy of the data pages */
370 int i
= r1_bio
->behind_page_count
;
372 safe_put_page(r1_bio
->behind_bvecs
[i
].bv_page
);
373 kfree(r1_bio
->behind_bvecs
);
374 r1_bio
->behind_bvecs
= NULL
;
376 /* clear the bitmap if all writes complete successfully */
377 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
379 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
380 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
381 md_write_end(r1_bio
->mddev
);
384 static void r1_bio_write_done(struct r1bio
*r1_bio
)
386 if (!atomic_dec_and_test(&r1_bio
->remaining
))
389 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
390 reschedule_retry(r1_bio
);
393 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
))
394 reschedule_retry(r1_bio
);
396 raid_end_bio_io(r1_bio
);
400 static void raid1_end_write_request(struct bio
*bio
, int error
)
402 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
403 struct r1bio
*r1_bio
= bio
->bi_private
;
404 int mirror
, behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
405 struct r1conf
*conf
= r1_bio
->mddev
->private;
406 struct bio
*to_put
= NULL
;
408 mirror
= find_bio_disk(r1_bio
, bio
);
411 * 'one mirror IO has finished' event handler:
414 set_bit(WriteErrorSeen
,
415 &conf
->mirrors
[mirror
].rdev
->flags
);
416 if (!test_and_set_bit(WantReplacement
,
417 &conf
->mirrors
[mirror
].rdev
->flags
))
418 set_bit(MD_RECOVERY_NEEDED
, &
419 conf
->mddev
->recovery
);
421 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
424 * Set R1BIO_Uptodate in our master bio, so that we
425 * will return a good error code for to the higher
426 * levels even if IO on some other mirrored buffer
429 * The 'master' represents the composite IO operation
430 * to user-side. So if something waits for IO, then it
431 * will wait for the 'master' bio.
436 r1_bio
->bios
[mirror
] = NULL
;
439 * Do not set R1BIO_Uptodate if the current device is
440 * rebuilding or Faulty. This is because we cannot use
441 * such device for properly reading the data back (we could
442 * potentially use it, if the current write would have felt
443 * before rdev->recovery_offset, but for simplicity we don't
446 if (test_bit(In_sync
, &conf
->mirrors
[mirror
].rdev
->flags
) &&
447 !test_bit(Faulty
, &conf
->mirrors
[mirror
].rdev
->flags
))
448 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
450 /* Maybe we can clear some bad blocks. */
451 if (is_badblock(conf
->mirrors
[mirror
].rdev
,
452 r1_bio
->sector
, r1_bio
->sectors
,
453 &first_bad
, &bad_sectors
)) {
454 r1_bio
->bios
[mirror
] = IO_MADE_GOOD
;
455 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
460 if (test_bit(WriteMostly
, &conf
->mirrors
[mirror
].rdev
->flags
))
461 atomic_dec(&r1_bio
->behind_remaining
);
464 * In behind mode, we ACK the master bio once the I/O
465 * has safely reached all non-writemostly
466 * disks. Setting the Returned bit ensures that this
467 * gets done only once -- we don't ever want to return
468 * -EIO here, instead we'll wait
470 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
471 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
472 /* Maybe we can return now */
473 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
474 struct bio
*mbio
= r1_bio
->master_bio
;
475 pr_debug("raid1: behind end write sectors"
477 (unsigned long long) mbio
->bi_iter
.bi_sector
,
478 (unsigned long long) bio_end_sector(mbio
) - 1);
479 call_bio_endio(r1_bio
);
483 if (r1_bio
->bios
[mirror
] == NULL
)
484 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
,
488 * Let's see if all mirrored write operations have finished
491 r1_bio_write_done(r1_bio
);
498 * This routine returns the disk from which the requested read should
499 * be done. There is a per-array 'next expected sequential IO' sector
500 * number - if this matches on the next IO then we use the last disk.
501 * There is also a per-disk 'last know head position' sector that is
502 * maintained from IRQ contexts, both the normal and the resync IO
503 * completion handlers update this position correctly. If there is no
504 * perfect sequential match then we pick the disk whose head is closest.
506 * If there are 2 mirrors in the same 2 devices, performance degrades
507 * because position is mirror, not device based.
509 * The rdev for the device selected will have nr_pending incremented.
511 static int read_balance(struct r1conf
*conf
, struct r1bio
*r1_bio
, int *max_sectors
)
513 const sector_t this_sector
= r1_bio
->sector
;
515 int best_good_sectors
;
516 int best_disk
, best_dist_disk
, best_pending_disk
;
520 unsigned int min_pending
;
521 struct md_rdev
*rdev
;
523 int choose_next_idle
;
527 * Check if we can balance. We can balance on the whole
528 * device if no resync is going on, or below the resync window.
529 * We take the first readable disk when above the resync window.
532 sectors
= r1_bio
->sectors
;
535 best_dist
= MaxSector
;
536 best_pending_disk
= -1;
537 min_pending
= UINT_MAX
;
538 best_good_sectors
= 0;
540 choose_next_idle
= 0;
542 if ((conf
->mddev
->recovery_cp
< this_sector
+ sectors
) ||
543 (mddev_is_clustered(conf
->mddev
) &&
544 md_cluster_ops
->area_resyncing(conf
->mddev
, READ
, this_sector
,
545 this_sector
+ sectors
)))
550 for (disk
= 0 ; disk
< conf
->raid_disks
* 2 ; disk
++) {
554 unsigned int pending
;
557 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
558 if (r1_bio
->bios
[disk
] == IO_BLOCKED
560 || test_bit(Unmerged
, &rdev
->flags
)
561 || test_bit(Faulty
, &rdev
->flags
))
563 if (!test_bit(In_sync
, &rdev
->flags
) &&
564 rdev
->recovery_offset
< this_sector
+ sectors
)
566 if (test_bit(WriteMostly
, &rdev
->flags
)) {
567 /* Don't balance among write-mostly, just
568 * use the first as a last resort */
569 if (best_dist_disk
< 0) {
570 if (is_badblock(rdev
, this_sector
, sectors
,
571 &first_bad
, &bad_sectors
)) {
572 if (first_bad
< this_sector
)
573 /* Cannot use this */
575 best_good_sectors
= first_bad
- this_sector
;
577 best_good_sectors
= sectors
;
578 best_dist_disk
= disk
;
579 best_pending_disk
= disk
;
583 /* This is a reasonable device to use. It might
586 if (is_badblock(rdev
, this_sector
, sectors
,
587 &first_bad
, &bad_sectors
)) {
588 if (best_dist
< MaxSector
)
589 /* already have a better device */
591 if (first_bad
<= this_sector
) {
592 /* cannot read here. If this is the 'primary'
593 * device, then we must not read beyond
594 * bad_sectors from another device..
596 bad_sectors
-= (this_sector
- first_bad
);
597 if (choose_first
&& sectors
> bad_sectors
)
598 sectors
= bad_sectors
;
599 if (best_good_sectors
> sectors
)
600 best_good_sectors
= sectors
;
603 sector_t good_sectors
= first_bad
- this_sector
;
604 if (good_sectors
> best_good_sectors
) {
605 best_good_sectors
= good_sectors
;
613 best_good_sectors
= sectors
;
615 nonrot
= blk_queue_nonrot(bdev_get_queue(rdev
->bdev
));
616 has_nonrot_disk
|= nonrot
;
617 pending
= atomic_read(&rdev
->nr_pending
);
618 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
623 /* Don't change to another disk for sequential reads */
624 if (conf
->mirrors
[disk
].next_seq_sect
== this_sector
626 int opt_iosize
= bdev_io_opt(rdev
->bdev
) >> 9;
627 struct raid1_info
*mirror
= &conf
->mirrors
[disk
];
631 * If buffered sequential IO size exceeds optimal
632 * iosize, check if there is idle disk. If yes, choose
633 * the idle disk. read_balance could already choose an
634 * idle disk before noticing it's a sequential IO in
635 * this disk. This doesn't matter because this disk
636 * will idle, next time it will be utilized after the
637 * first disk has IO size exceeds optimal iosize. In
638 * this way, iosize of the first disk will be optimal
639 * iosize at least. iosize of the second disk might be
640 * small, but not a big deal since when the second disk
641 * starts IO, the first disk is likely still busy.
643 if (nonrot
&& opt_iosize
> 0 &&
644 mirror
->seq_start
!= MaxSector
&&
645 mirror
->next_seq_sect
> opt_iosize
&&
646 mirror
->next_seq_sect
- opt_iosize
>=
648 choose_next_idle
= 1;
653 /* If device is idle, use it */
659 if (choose_next_idle
)
662 if (min_pending
> pending
) {
663 min_pending
= pending
;
664 best_pending_disk
= disk
;
667 if (dist
< best_dist
) {
669 best_dist_disk
= disk
;
674 * If all disks are rotational, choose the closest disk. If any disk is
675 * non-rotational, choose the disk with less pending request even the
676 * disk is rotational, which might/might not be optimal for raids with
677 * mixed ratation/non-rotational disks depending on workload.
679 if (best_disk
== -1) {
681 best_disk
= best_pending_disk
;
683 best_disk
= best_dist_disk
;
686 if (best_disk
>= 0) {
687 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
690 atomic_inc(&rdev
->nr_pending
);
691 if (test_bit(Faulty
, &rdev
->flags
)) {
692 /* cannot risk returning a device that failed
693 * before we inc'ed nr_pending
695 rdev_dec_pending(rdev
, conf
->mddev
);
698 sectors
= best_good_sectors
;
700 if (conf
->mirrors
[best_disk
].next_seq_sect
!= this_sector
)
701 conf
->mirrors
[best_disk
].seq_start
= this_sector
;
703 conf
->mirrors
[best_disk
].next_seq_sect
= this_sector
+ sectors
;
706 *max_sectors
= sectors
;
711 static int raid1_mergeable_bvec(struct mddev
*mddev
,
712 struct bvec_merge_data
*bvm
,
713 struct bio_vec
*biovec
)
715 struct r1conf
*conf
= mddev
->private;
716 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
717 int max
= biovec
->bv_len
;
719 if (mddev
->merge_check_needed
) {
722 for (disk
= 0; disk
< conf
->raid_disks
* 2; disk
++) {
723 struct md_rdev
*rdev
= rcu_dereference(
724 conf
->mirrors
[disk
].rdev
);
725 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
726 struct request_queue
*q
=
727 bdev_get_queue(rdev
->bdev
);
728 if (q
->merge_bvec_fn
) {
729 bvm
->bi_sector
= sector
+
731 bvm
->bi_bdev
= rdev
->bdev
;
732 max
= min(max
, q
->merge_bvec_fn(
743 static int raid1_congested(struct mddev
*mddev
, int bits
)
745 struct r1conf
*conf
= mddev
->private;
748 if ((bits
& (1 << WB_async_congested
)) &&
749 conf
->pending_count
>= max_queued_requests
)
753 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
754 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
755 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
756 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
760 /* Note the '|| 1' - when read_balance prefers
761 * non-congested targets, it can be removed
763 if ((bits
& (1 << WB_async_congested
)) || 1)
764 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
766 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
773 static void flush_pending_writes(struct r1conf
*conf
)
775 /* Any writes that have been queued but are awaiting
776 * bitmap updates get flushed here.
778 spin_lock_irq(&conf
->device_lock
);
780 if (conf
->pending_bio_list
.head
) {
782 bio
= bio_list_get(&conf
->pending_bio_list
);
783 conf
->pending_count
= 0;
784 spin_unlock_irq(&conf
->device_lock
);
785 /* flush any pending bitmap writes to
786 * disk before proceeding w/ I/O */
787 bitmap_unplug(conf
->mddev
->bitmap
);
788 wake_up(&conf
->wait_barrier
);
790 while (bio
) { /* submit pending writes */
791 struct bio
*next
= bio
->bi_next
;
793 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
794 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
798 generic_make_request(bio
);
802 spin_unlock_irq(&conf
->device_lock
);
806 * Sometimes we need to suspend IO while we do something else,
807 * either some resync/recovery, or reconfigure the array.
808 * To do this we raise a 'barrier'.
809 * The 'barrier' is a counter that can be raised multiple times
810 * to count how many activities are happening which preclude
812 * We can only raise the barrier if there is no pending IO.
813 * i.e. if nr_pending == 0.
814 * We choose only to raise the barrier if no-one is waiting for the
815 * barrier to go down. This means that as soon as an IO request
816 * is ready, no other operations which require a barrier will start
817 * until the IO request has had a chance.
819 * So: regular IO calls 'wait_barrier'. When that returns there
820 * is no backgroup IO happening, It must arrange to call
821 * allow_barrier when it has finished its IO.
822 * backgroup IO calls must call raise_barrier. Once that returns
823 * there is no normal IO happeing. It must arrange to call
824 * lower_barrier when the particular background IO completes.
826 static void raise_barrier(struct r1conf
*conf
, sector_t sector_nr
)
828 spin_lock_irq(&conf
->resync_lock
);
830 /* Wait until no block IO is waiting */
831 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
834 /* block any new IO from starting */
836 conf
->next_resync
= sector_nr
;
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
845 * D: while there are any active requests in the current window.
847 wait_event_lock_irq(conf
->wait_barrier
,
848 !conf
->array_frozen
&&
849 conf
->barrier
< RESYNC_DEPTH
&&
850 conf
->current_window_requests
== 0 &&
851 (conf
->start_next_window
>=
852 conf
->next_resync
+ RESYNC_SECTORS
),
856 spin_unlock_irq(&conf
->resync_lock
);
859 static void lower_barrier(struct r1conf
*conf
)
862 BUG_ON(conf
->barrier
<= 0);
863 spin_lock_irqsave(&conf
->resync_lock
, flags
);
866 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
867 wake_up(&conf
->wait_barrier
);
870 static bool need_to_wait_for_sync(struct r1conf
*conf
, struct bio
*bio
)
874 if (conf
->array_frozen
|| !bio
)
876 else if (conf
->barrier
&& bio_data_dir(bio
) == WRITE
) {
877 if ((conf
->mddev
->curr_resync_completed
878 >= bio_end_sector(bio
)) ||
879 (conf
->next_resync
+ NEXT_NORMALIO_DISTANCE
880 <= bio
->bi_iter
.bi_sector
))
889 static sector_t
wait_barrier(struct r1conf
*conf
, struct bio
*bio
)
893 spin_lock_irq(&conf
->resync_lock
);
894 if (need_to_wait_for_sync(conf
, bio
)) {
896 /* Wait for the barrier to drop.
897 * However if there are already pending
898 * requests (preventing the barrier from
899 * rising completely), and the
900 * per-process bio queue isn't empty,
901 * then don't wait, as we need to empty
902 * that queue to allow conf->start_next_window
905 wait_event_lock_irq(conf
->wait_barrier
,
906 !conf
->array_frozen
&&
908 ((conf
->start_next_window
<
909 conf
->next_resync
+ RESYNC_SECTORS
) &&
911 !bio_list_empty(current
->bio_list
))),
916 if (bio
&& bio_data_dir(bio
) == WRITE
) {
917 if (bio
->bi_iter
.bi_sector
>=
918 conf
->mddev
->curr_resync_completed
) {
919 if (conf
->start_next_window
== MaxSector
)
920 conf
->start_next_window
=
922 NEXT_NORMALIO_DISTANCE
;
924 if ((conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
)
925 <= bio
->bi_iter
.bi_sector
)
926 conf
->next_window_requests
++;
928 conf
->current_window_requests
++;
929 sector
= conf
->start_next_window
;
934 spin_unlock_irq(&conf
->resync_lock
);
938 static void allow_barrier(struct r1conf
*conf
, sector_t start_next_window
,
943 spin_lock_irqsave(&conf
->resync_lock
, flags
);
945 if (start_next_window
) {
946 if (start_next_window
== conf
->start_next_window
) {
947 if (conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
949 conf
->next_window_requests
--;
951 conf
->current_window_requests
--;
953 conf
->current_window_requests
--;
955 if (!conf
->current_window_requests
) {
956 if (conf
->next_window_requests
) {
957 conf
->current_window_requests
=
958 conf
->next_window_requests
;
959 conf
->next_window_requests
= 0;
960 conf
->start_next_window
+=
961 NEXT_NORMALIO_DISTANCE
;
963 conf
->start_next_window
= MaxSector
;
966 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
967 wake_up(&conf
->wait_barrier
);
970 static void freeze_array(struct r1conf
*conf
, int extra
)
972 /* stop syncio and normal IO and wait for everything to
974 * We wait until nr_pending match nr_queued+extra
975 * This is called in the context of one normal IO request
976 * that has failed. Thus any sync request that might be pending
977 * will be blocked by nr_pending, and we need to wait for
978 * pending IO requests to complete or be queued for re-try.
979 * Thus the number queued (nr_queued) plus this request (extra)
980 * must match the number of pending IOs (nr_pending) before
983 spin_lock_irq(&conf
->resync_lock
);
984 conf
->array_frozen
= 1;
985 wait_event_lock_irq_cmd(conf
->wait_barrier
,
986 conf
->nr_pending
== conf
->nr_queued
+extra
,
988 flush_pending_writes(conf
));
989 spin_unlock_irq(&conf
->resync_lock
);
991 static void unfreeze_array(struct r1conf
*conf
)
993 /* reverse the effect of the freeze */
994 spin_lock_irq(&conf
->resync_lock
);
995 conf
->array_frozen
= 0;
996 wake_up(&conf
->wait_barrier
);
997 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",
1032 bio
->bi_iter
.bi_size
);
1035 struct raid1_plug_cb
{
1036 struct blk_plug_cb cb
;
1037 struct bio_list pending
;
1041 static void raid1_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1043 struct raid1_plug_cb
*plug
= container_of(cb
, struct raid1_plug_cb
,
1045 struct mddev
*mddev
= plug
->cb
.data
;
1046 struct r1conf
*conf
= mddev
->private;
1049 if (from_schedule
|| current
->bio_list
) {
1050 spin_lock_irq(&conf
->device_lock
);
1051 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1052 conf
->pending_count
+= plug
->pending_cnt
;
1053 spin_unlock_irq(&conf
->device_lock
);
1054 wake_up(&conf
->wait_barrier
);
1055 md_wakeup_thread(mddev
->thread
);
1060 /* we aren't scheduling, so we can do the write-out directly. */
1061 bio
= bio_list_get(&plug
->pending
);
1062 bitmap_unplug(mddev
->bitmap
);
1063 wake_up(&conf
->wait_barrier
);
1065 while (bio
) { /* submit pending writes */
1066 struct bio
*next
= bio
->bi_next
;
1067 bio
->bi_next
= NULL
;
1068 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
1069 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1070 /* Just ignore it */
1073 generic_make_request(bio
);
1079 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1081 struct r1conf
*conf
= mddev
->private;
1082 struct raid1_info
*mirror
;
1083 struct r1bio
*r1_bio
;
1084 struct bio
*read_bio
;
1086 struct bitmap
*bitmap
;
1087 unsigned long flags
;
1088 const int rw
= bio_data_dir(bio
);
1089 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1090 const unsigned long do_flush_fua
= (bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
));
1091 const unsigned long do_discard
= (bio
->bi_rw
1092 & (REQ_DISCARD
| REQ_SECURE
));
1093 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1094 struct md_rdev
*blocked_rdev
;
1095 struct blk_plug_cb
*cb
;
1096 struct raid1_plug_cb
*plug
= NULL
;
1098 int sectors_handled
;
1100 sector_t start_next_window
;
1103 * Register the new request and wait if the reconstruction
1104 * thread has put up a bar for new requests.
1105 * Continue immediately if no resync is active currently.
1108 md_write_start(mddev
, bio
); /* wait on superblock update early */
1110 if (bio_data_dir(bio
) == WRITE
&&
1111 ((bio_end_sector(bio
) > mddev
->suspend_lo
&&
1112 bio
->bi_iter
.bi_sector
< mddev
->suspend_hi
) ||
1113 (mddev_is_clustered(mddev
) &&
1114 md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1115 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
))))) {
1116 /* As the suspend_* range is controlled by
1117 * userspace, we want an interruptible
1122 flush_signals(current
);
1123 prepare_to_wait(&conf
->wait_barrier
,
1124 &w
, TASK_INTERRUPTIBLE
);
1125 if (bio_end_sector(bio
) <= mddev
->suspend_lo
||
1126 bio
->bi_iter
.bi_sector
>= mddev
->suspend_hi
||
1127 (mddev_is_clustered(mddev
) &&
1128 !md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1129 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
))))
1133 finish_wait(&conf
->wait_barrier
, &w
);
1136 start_next_window
= wait_barrier(conf
, bio
);
1138 bitmap
= mddev
->bitmap
;
1141 * make_request() can abort the operation when READA is being
1142 * used and no empty request is available.
1145 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1147 r1_bio
->master_bio
= bio
;
1148 r1_bio
->sectors
= bio_sectors(bio
);
1150 r1_bio
->mddev
= mddev
;
1151 r1_bio
->sector
= bio
->bi_iter
.bi_sector
;
1153 /* We might need to issue multiple reads to different
1154 * devices if there are bad blocks around, so we keep
1155 * track of the number of reads in bio->bi_phys_segments.
1156 * If this is 0, there is only one r1_bio and no locking
1157 * will be needed when requests complete. If it is
1158 * non-zero, then it is the number of not-completed requests.
1160 bio
->bi_phys_segments
= 0;
1161 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1165 * read balancing logic:
1170 rdisk
= read_balance(conf
, r1_bio
, &max_sectors
);
1173 /* couldn't find anywhere to read from */
1174 raid_end_bio_io(r1_bio
);
1177 mirror
= conf
->mirrors
+ rdisk
;
1179 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
1181 /* Reading from a write-mostly device must
1182 * take care not to over-take any writes
1185 wait_event(bitmap
->behind_wait
,
1186 atomic_read(&bitmap
->behind_writes
) == 0);
1188 r1_bio
->read_disk
= rdisk
;
1189 r1_bio
->start_next_window
= 0;
1191 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1192 bio_trim(read_bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
1195 r1_bio
->bios
[rdisk
] = read_bio
;
1197 read_bio
->bi_iter
.bi_sector
= r1_bio
->sector
+
1198 mirror
->rdev
->data_offset
;
1199 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
1200 read_bio
->bi_end_io
= raid1_end_read_request
;
1201 read_bio
->bi_rw
= READ
| do_sync
;
1202 read_bio
->bi_private
= r1_bio
;
1204 if (max_sectors
< r1_bio
->sectors
) {
1205 /* could not read all from this device, so we will
1206 * need another r1_bio.
1209 sectors_handled
= (r1_bio
->sector
+ max_sectors
1210 - bio
->bi_iter
.bi_sector
);
1211 r1_bio
->sectors
= max_sectors
;
1212 spin_lock_irq(&conf
->device_lock
);
1213 if (bio
->bi_phys_segments
== 0)
1214 bio
->bi_phys_segments
= 2;
1216 bio
->bi_phys_segments
++;
1217 spin_unlock_irq(&conf
->device_lock
);
1218 /* Cannot call generic_make_request directly
1219 * as that will be queued in __make_request
1220 * and subsequent mempool_alloc might block waiting
1221 * for it. So hand bio over to raid1d.
1223 reschedule_retry(r1_bio
);
1225 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1227 r1_bio
->master_bio
= bio
;
1228 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1230 r1_bio
->mddev
= mddev
;
1231 r1_bio
->sector
= bio
->bi_iter
.bi_sector
+
1235 generic_make_request(read_bio
);
1242 if (conf
->pending_count
>= max_queued_requests
) {
1243 md_wakeup_thread(mddev
->thread
);
1244 wait_event(conf
->wait_barrier
,
1245 conf
->pending_count
< max_queued_requests
);
1247 /* first select target devices under rcu_lock and
1248 * inc refcount on their rdev. Record them by setting
1250 * If there are known/acknowledged bad blocks on any device on
1251 * which we have seen a write error, we want to avoid writing those
1253 * This potentially requires several writes to write around
1254 * the bad blocks. Each set of writes gets it's own r1bio
1255 * with a set of bios attached.
1258 disks
= conf
->raid_disks
* 2;
1260 r1_bio
->start_next_window
= start_next_window
;
1261 blocked_rdev
= NULL
;
1263 max_sectors
= r1_bio
->sectors
;
1264 for (i
= 0; i
< disks
; i
++) {
1265 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1266 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1267 atomic_inc(&rdev
->nr_pending
);
1268 blocked_rdev
= rdev
;
1271 r1_bio
->bios
[i
] = NULL
;
1272 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)
1273 || test_bit(Unmerged
, &rdev
->flags
)) {
1274 if (i
< conf
->raid_disks
)
1275 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1279 atomic_inc(&rdev
->nr_pending
);
1280 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1285 is_bad
= is_badblock(rdev
, r1_bio
->sector
,
1287 &first_bad
, &bad_sectors
);
1289 /* mustn't write here until the bad block is
1291 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1292 blocked_rdev
= rdev
;
1295 if (is_bad
&& first_bad
<= r1_bio
->sector
) {
1296 /* Cannot write here at all */
1297 bad_sectors
-= (r1_bio
->sector
- first_bad
);
1298 if (bad_sectors
< max_sectors
)
1299 /* mustn't write more than bad_sectors
1300 * to other devices yet
1302 max_sectors
= bad_sectors
;
1303 rdev_dec_pending(rdev
, mddev
);
1304 /* We don't set R1BIO_Degraded as that
1305 * only applies if the disk is
1306 * missing, so it might be re-added,
1307 * and we want to know to recover this
1309 * In this case the device is here,
1310 * and the fact that this chunk is not
1311 * in-sync is recorded in the bad
1317 int good_sectors
= first_bad
- r1_bio
->sector
;
1318 if (good_sectors
< max_sectors
)
1319 max_sectors
= good_sectors
;
1322 r1_bio
->bios
[i
] = bio
;
1326 if (unlikely(blocked_rdev
)) {
1327 /* Wait for this device to become unblocked */
1329 sector_t old
= start_next_window
;
1331 for (j
= 0; j
< i
; j
++)
1332 if (r1_bio
->bios
[j
])
1333 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
1335 allow_barrier(conf
, start_next_window
, bio
->bi_iter
.bi_sector
);
1336 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1337 start_next_window
= wait_barrier(conf
, bio
);
1339 * We must make sure the multi r1bios of bio have
1340 * the same value of bi_phys_segments
1342 if (bio
->bi_phys_segments
&& old
&&
1343 old
!= start_next_window
)
1344 /* Wait for the former r1bio(s) to complete */
1345 wait_event(conf
->wait_barrier
,
1346 bio
->bi_phys_segments
== 1);
1350 if (max_sectors
< r1_bio
->sectors
) {
1351 /* We are splitting this write into multiple parts, so
1352 * we need to prepare for allocating another r1_bio.
1354 r1_bio
->sectors
= max_sectors
;
1355 spin_lock_irq(&conf
->device_lock
);
1356 if (bio
->bi_phys_segments
== 0)
1357 bio
->bi_phys_segments
= 2;
1359 bio
->bi_phys_segments
++;
1360 spin_unlock_irq(&conf
->device_lock
);
1362 sectors_handled
= r1_bio
->sector
+ max_sectors
- bio
->bi_iter
.bi_sector
;
1364 atomic_set(&r1_bio
->remaining
, 1);
1365 atomic_set(&r1_bio
->behind_remaining
, 0);
1368 for (i
= 0; i
< disks
; i
++) {
1370 if (!r1_bio
->bios
[i
])
1373 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1374 bio_trim(mbio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
1378 * Not if there are too many, or cannot
1379 * allocate memory, or a reader on WriteMostly
1380 * is waiting for behind writes to flush */
1382 (atomic_read(&bitmap
->behind_writes
)
1383 < mddev
->bitmap_info
.max_write_behind
) &&
1384 !waitqueue_active(&bitmap
->behind_wait
))
1385 alloc_behind_pages(mbio
, r1_bio
);
1387 bitmap_startwrite(bitmap
, r1_bio
->sector
,
1389 test_bit(R1BIO_BehindIO
,
1393 if (r1_bio
->behind_bvecs
) {
1394 struct bio_vec
*bvec
;
1398 * We trimmed the bio, so _all is legit
1400 bio_for_each_segment_all(bvec
, mbio
, j
)
1401 bvec
->bv_page
= r1_bio
->behind_bvecs
[j
].bv_page
;
1402 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
1403 atomic_inc(&r1_bio
->behind_remaining
);
1406 r1_bio
->bios
[i
] = mbio
;
1408 mbio
->bi_iter
.bi_sector
= (r1_bio
->sector
+
1409 conf
->mirrors
[i
].rdev
->data_offset
);
1410 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1411 mbio
->bi_end_io
= raid1_end_write_request
;
1413 WRITE
| do_flush_fua
| do_sync
| do_discard
| do_same
;
1414 mbio
->bi_private
= r1_bio
;
1416 atomic_inc(&r1_bio
->remaining
);
1418 cb
= blk_check_plugged(raid1_unplug
, mddev
, sizeof(*plug
));
1420 plug
= container_of(cb
, struct raid1_plug_cb
, cb
);
1423 spin_lock_irqsave(&conf
->device_lock
, flags
);
1425 bio_list_add(&plug
->pending
, mbio
);
1426 plug
->pending_cnt
++;
1428 bio_list_add(&conf
->pending_bio_list
, mbio
);
1429 conf
->pending_count
++;
1431 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1433 md_wakeup_thread(mddev
->thread
);
1435 /* Mustn't call r1_bio_write_done before this next test,
1436 * as it could result in the bio being freed.
1438 if (sectors_handled
< bio_sectors(bio
)) {
1439 r1_bio_write_done(r1_bio
);
1440 /* We need another r1_bio. It has already been counted
1441 * in bio->bi_phys_segments
1443 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1444 r1_bio
->master_bio
= bio
;
1445 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1447 r1_bio
->mddev
= mddev
;
1448 r1_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1452 r1_bio_write_done(r1_bio
);
1454 /* In case raid1d snuck in to freeze_array */
1455 wake_up(&conf
->wait_barrier
);
1458 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1460 struct r1conf
*conf
= mddev
->private;
1463 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1464 conf
->raid_disks
- mddev
->degraded
);
1466 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1467 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1468 seq_printf(seq
, "%s",
1469 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1472 seq_printf(seq
, "]");
1475 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1477 char b
[BDEVNAME_SIZE
];
1478 struct r1conf
*conf
= mddev
->private;
1479 unsigned long flags
;
1482 * If it is not operational, then we have already marked it as dead
1483 * else if it is the last working disks, ignore the error, let the
1484 * next level up know.
1485 * else mark the drive as failed
1487 if (test_bit(In_sync
, &rdev
->flags
)
1488 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1490 * Don't fail the drive, act as though we were just a
1491 * normal single drive.
1492 * However don't try a recovery from this drive as
1493 * it is very likely to fail.
1495 conf
->recovery_disabled
= mddev
->recovery_disabled
;
1498 set_bit(Blocked
, &rdev
->flags
);
1499 spin_lock_irqsave(&conf
->device_lock
, flags
);
1500 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1502 set_bit(Faulty
, &rdev
->flags
);
1504 set_bit(Faulty
, &rdev
->flags
);
1505 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1507 * if recovery is running, make sure it aborts.
1509 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1510 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1512 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1513 "md/raid1:%s: Operation continuing on %d devices.\n",
1514 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1515 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1518 static void print_conf(struct r1conf
*conf
)
1522 printk(KERN_DEBUG
"RAID1 conf printout:\n");
1524 printk(KERN_DEBUG
"(!conf)\n");
1527 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1531 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1532 char b
[BDEVNAME_SIZE
];
1533 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1535 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1536 i
, !test_bit(In_sync
, &rdev
->flags
),
1537 !test_bit(Faulty
, &rdev
->flags
),
1538 bdevname(rdev
->bdev
,b
));
1543 static void close_sync(struct r1conf
*conf
)
1545 wait_barrier(conf
, NULL
);
1546 allow_barrier(conf
, 0, 0);
1548 mempool_destroy(conf
->r1buf_pool
);
1549 conf
->r1buf_pool
= NULL
;
1551 spin_lock_irq(&conf
->resync_lock
);
1552 conf
->next_resync
= 0;
1553 conf
->start_next_window
= MaxSector
;
1554 conf
->current_window_requests
+=
1555 conf
->next_window_requests
;
1556 conf
->next_window_requests
= 0;
1557 spin_unlock_irq(&conf
->resync_lock
);
1560 static int raid1_spare_active(struct mddev
*mddev
)
1563 struct r1conf
*conf
= mddev
->private;
1565 unsigned long flags
;
1568 * Find all failed disks within the RAID1 configuration
1569 * and mark them readable.
1570 * Called under mddev lock, so rcu protection not needed.
1571 * device_lock used to avoid races with raid1_end_read_request
1572 * which expects 'In_sync' flags and ->degraded to be consistent.
1574 spin_lock_irqsave(&conf
->device_lock
, flags
);
1575 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1576 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1577 struct md_rdev
*repl
= conf
->mirrors
[conf
->raid_disks
+ i
].rdev
;
1579 && !test_bit(Candidate
, &repl
->flags
)
1580 && repl
->recovery_offset
== MaxSector
1581 && !test_bit(Faulty
, &repl
->flags
)
1582 && !test_and_set_bit(In_sync
, &repl
->flags
)) {
1583 /* replacement has just become active */
1585 !test_and_clear_bit(In_sync
, &rdev
->flags
))
1588 /* Replaced device not technically
1589 * faulty, but we need to be sure
1590 * it gets removed and never re-added
1592 set_bit(Faulty
, &rdev
->flags
);
1593 sysfs_notify_dirent_safe(
1598 && rdev
->recovery_offset
== MaxSector
1599 && !test_bit(Faulty
, &rdev
->flags
)
1600 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1602 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
1605 mddev
->degraded
-= count
;
1606 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1612 static int raid1_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1614 struct r1conf
*conf
= mddev
->private;
1617 struct raid1_info
*p
;
1619 int last
= conf
->raid_disks
- 1;
1620 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1622 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
1625 if (rdev
->raid_disk
>= 0)
1626 first
= last
= rdev
->raid_disk
;
1628 if (q
->merge_bvec_fn
) {
1629 set_bit(Unmerged
, &rdev
->flags
);
1630 mddev
->merge_check_needed
= 1;
1633 for (mirror
= first
; mirror
<= last
; mirror
++) {
1634 p
= conf
->mirrors
+mirror
;
1638 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1639 rdev
->data_offset
<< 9);
1641 p
->head_position
= 0;
1642 rdev
->raid_disk
= mirror
;
1644 /* As all devices are equivalent, we don't need a full recovery
1645 * if this was recently any drive of the array
1647 if (rdev
->saved_raid_disk
< 0)
1649 rcu_assign_pointer(p
->rdev
, rdev
);
1652 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
1653 p
[conf
->raid_disks
].rdev
== NULL
) {
1654 /* Add this device as a replacement */
1655 clear_bit(In_sync
, &rdev
->flags
);
1656 set_bit(Replacement
, &rdev
->flags
);
1657 rdev
->raid_disk
= mirror
;
1660 rcu_assign_pointer(p
[conf
->raid_disks
].rdev
, rdev
);
1664 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1665 /* Some requests might not have seen this new
1666 * merge_bvec_fn. We must wait for them to complete
1667 * before merging the device fully.
1668 * First we make sure any code which has tested
1669 * our function has submitted the request, then
1670 * we wait for all outstanding requests to complete.
1672 synchronize_sched();
1673 freeze_array(conf
, 0);
1674 unfreeze_array(conf
);
1675 clear_bit(Unmerged
, &rdev
->flags
);
1677 md_integrity_add_rdev(rdev
, mddev
);
1678 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1679 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1684 static int raid1_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1686 struct r1conf
*conf
= mddev
->private;
1688 int number
= rdev
->raid_disk
;
1689 struct raid1_info
*p
= conf
->mirrors
+ number
;
1691 if (rdev
!= p
->rdev
)
1692 p
= conf
->mirrors
+ conf
->raid_disks
+ number
;
1695 if (rdev
== p
->rdev
) {
1696 if (test_bit(In_sync
, &rdev
->flags
) ||
1697 atomic_read(&rdev
->nr_pending
)) {
1701 /* Only remove non-faulty devices if recovery
1704 if (!test_bit(Faulty
, &rdev
->flags
) &&
1705 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
1706 mddev
->degraded
< conf
->raid_disks
) {
1712 if (atomic_read(&rdev
->nr_pending
)) {
1713 /* lost the race, try later */
1717 } else if (conf
->mirrors
[conf
->raid_disks
+ number
].rdev
) {
1718 /* We just removed a device that is being replaced.
1719 * Move down the replacement. We drain all IO before
1720 * doing this to avoid confusion.
1722 struct md_rdev
*repl
=
1723 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
;
1724 freeze_array(conf
, 0);
1725 clear_bit(Replacement
, &repl
->flags
);
1727 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
= NULL
;
1728 unfreeze_array(conf
);
1729 clear_bit(WantReplacement
, &rdev
->flags
);
1731 clear_bit(WantReplacement
, &rdev
->flags
);
1732 err
= md_integrity_register(mddev
);
1740 static void end_sync_read(struct bio
*bio
, int error
)
1742 struct r1bio
*r1_bio
= bio
->bi_private
;
1744 update_head_pos(r1_bio
->read_disk
, r1_bio
);
1747 * we have read a block, now it needs to be re-written,
1748 * or re-read if the read failed.
1749 * We don't do much here, just schedule handling by raid1d
1751 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1752 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1754 if (atomic_dec_and_test(&r1_bio
->remaining
))
1755 reschedule_retry(r1_bio
);
1758 static void end_sync_write(struct bio
*bio
, int error
)
1760 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1761 struct r1bio
*r1_bio
= bio
->bi_private
;
1762 struct mddev
*mddev
= r1_bio
->mddev
;
1763 struct r1conf
*conf
= mddev
->private;
1768 mirror
= find_bio_disk(r1_bio
, bio
);
1771 sector_t sync_blocks
= 0;
1772 sector_t s
= r1_bio
->sector
;
1773 long sectors_to_go
= r1_bio
->sectors
;
1774 /* make sure these bits doesn't get cleared. */
1776 bitmap_end_sync(mddev
->bitmap
, s
,
1779 sectors_to_go
-= sync_blocks
;
1780 } while (sectors_to_go
> 0);
1781 set_bit(WriteErrorSeen
,
1782 &conf
->mirrors
[mirror
].rdev
->flags
);
1783 if (!test_and_set_bit(WantReplacement
,
1784 &conf
->mirrors
[mirror
].rdev
->flags
))
1785 set_bit(MD_RECOVERY_NEEDED
, &
1787 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
1788 } else if (is_badblock(conf
->mirrors
[mirror
].rdev
,
1791 &first_bad
, &bad_sectors
) &&
1792 !is_badblock(conf
->mirrors
[r1_bio
->read_disk
].rdev
,
1795 &first_bad
, &bad_sectors
)
1797 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
1799 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1800 int s
= r1_bio
->sectors
;
1801 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
1802 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1803 reschedule_retry(r1_bio
);
1806 md_done_sync(mddev
, s
, uptodate
);
1811 static int r1_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1812 int sectors
, struct page
*page
, int rw
)
1814 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
1818 set_bit(WriteErrorSeen
, &rdev
->flags
);
1819 if (!test_and_set_bit(WantReplacement
,
1821 set_bit(MD_RECOVERY_NEEDED
, &
1822 rdev
->mddev
->recovery
);
1824 /* need to record an error - either for the block or the device */
1825 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1826 md_error(rdev
->mddev
, rdev
);
1830 static int fix_sync_read_error(struct r1bio
*r1_bio
)
1832 /* Try some synchronous reads of other devices to get
1833 * good data, much like with normal read errors. Only
1834 * read into the pages we already have so we don't
1835 * need to re-issue the read request.
1836 * We don't need to freeze the array, because being in an
1837 * active sync request, there is no normal IO, and
1838 * no overlapping syncs.
1839 * We don't need to check is_badblock() again as we
1840 * made sure that anything with a bad block in range
1841 * will have bi_end_io clear.
1843 struct mddev
*mddev
= r1_bio
->mddev
;
1844 struct r1conf
*conf
= mddev
->private;
1845 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1846 sector_t sect
= r1_bio
->sector
;
1847 int sectors
= r1_bio
->sectors
;
1852 int d
= r1_bio
->read_disk
;
1854 struct md_rdev
*rdev
;
1857 if (s
> (PAGE_SIZE
>>9))
1860 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1861 /* No rcu protection needed here devices
1862 * can only be removed when no resync is
1863 * active, and resync is currently active
1865 rdev
= conf
->mirrors
[d
].rdev
;
1866 if (sync_page_io(rdev
, sect
, s
<<9,
1867 bio
->bi_io_vec
[idx
].bv_page
,
1874 if (d
== conf
->raid_disks
* 2)
1876 } while (!success
&& d
!= r1_bio
->read_disk
);
1879 char b
[BDEVNAME_SIZE
];
1881 /* Cannot read from anywhere, this block is lost.
1882 * Record a bad block on each device. If that doesn't
1883 * work just disable and interrupt the recovery.
1884 * Don't fail devices as that won't really help.
1886 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O read error"
1887 " for block %llu\n",
1889 bdevname(bio
->bi_bdev
, b
),
1890 (unsigned long long)r1_bio
->sector
);
1891 for (d
= 0; d
< conf
->raid_disks
* 2; d
++) {
1892 rdev
= conf
->mirrors
[d
].rdev
;
1893 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
1895 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
1899 conf
->recovery_disabled
=
1900 mddev
->recovery_disabled
;
1901 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1902 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1914 /* write it back and re-read */
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 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
1926 rdev_dec_pending(rdev
, mddev
);
1930 while (d
!= r1_bio
->read_disk
) {
1932 d
= conf
->raid_disks
* 2;
1934 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1936 rdev
= conf
->mirrors
[d
].rdev
;
1937 if (r1_sync_page_io(rdev
, sect
, s
,
1938 bio
->bi_io_vec
[idx
].bv_page
,
1940 atomic_add(s
, &rdev
->corrected_errors
);
1946 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1947 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1951 static void process_checks(struct r1bio
*r1_bio
)
1953 /* We have read all readable devices. If we haven't
1954 * got the block, then there is no hope left.
1955 * If we have, then we want to do a comparison
1956 * and skip the write if everything is the same.
1957 * If any blocks failed to read, then we need to
1958 * attempt an over-write
1960 struct mddev
*mddev
= r1_bio
->mddev
;
1961 struct r1conf
*conf
= mddev
->private;
1966 /* Fix variable parts of all bios */
1967 vcnt
= (r1_bio
->sectors
+ PAGE_SIZE
/ 512 - 1) >> (PAGE_SHIFT
- 9);
1968 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
1972 struct bio
*b
= r1_bio
->bios
[i
];
1973 if (b
->bi_end_io
!= end_sync_read
)
1975 /* fixup the bio for reuse, but preserve BIO_UPTODATE */
1976 uptodate
= test_bit(BIO_UPTODATE
, &b
->bi_flags
);
1979 clear_bit(BIO_UPTODATE
, &b
->bi_flags
);
1981 b
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
1982 b
->bi_iter
.bi_sector
= r1_bio
->sector
+
1983 conf
->mirrors
[i
].rdev
->data_offset
;
1984 b
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1985 b
->bi_end_io
= end_sync_read
;
1986 b
->bi_private
= r1_bio
;
1988 size
= b
->bi_iter
.bi_size
;
1989 for (j
= 0; j
< vcnt
; j
++) {
1991 bi
= &b
->bi_io_vec
[j
];
1993 if (size
> PAGE_SIZE
)
1994 bi
->bv_len
= PAGE_SIZE
;
2000 for (primary
= 0; primary
< conf
->raid_disks
* 2; primary
++)
2001 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
2002 test_bit(BIO_UPTODATE
, &r1_bio
->bios
[primary
]->bi_flags
)) {
2003 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
2004 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
2007 r1_bio
->read_disk
= primary
;
2008 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2010 struct bio
*pbio
= r1_bio
->bios
[primary
];
2011 struct bio
*sbio
= r1_bio
->bios
[i
];
2012 int uptodate
= test_bit(BIO_UPTODATE
, &sbio
->bi_flags
);
2014 if (sbio
->bi_end_io
!= end_sync_read
)
2016 /* Now we can 'fixup' the BIO_UPTODATE flag */
2017 set_bit(BIO_UPTODATE
, &sbio
->bi_flags
);
2020 for (j
= vcnt
; j
-- ; ) {
2022 p
= pbio
->bi_io_vec
[j
].bv_page
;
2023 s
= sbio
->bi_io_vec
[j
].bv_page
;
2024 if (memcmp(page_address(p
),
2026 sbio
->bi_io_vec
[j
].bv_len
))
2032 atomic64_add(r1_bio
->sectors
, &mddev
->resync_mismatches
);
2033 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
2035 /* No need to write to this device. */
2036 sbio
->bi_end_io
= NULL
;
2037 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
2041 bio_copy_data(sbio
, pbio
);
2045 static void sync_request_write(struct mddev
*mddev
, struct r1bio
*r1_bio
)
2047 struct r1conf
*conf
= mddev
->private;
2049 int disks
= conf
->raid_disks
* 2;
2050 struct bio
*bio
, *wbio
;
2052 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2054 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
2055 /* ouch - failed to read all of that. */
2056 if (!fix_sync_read_error(r1_bio
))
2059 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2060 process_checks(r1_bio
);
2065 atomic_set(&r1_bio
->remaining
, 1);
2066 for (i
= 0; i
< disks
; i
++) {
2067 wbio
= r1_bio
->bios
[i
];
2068 if (wbio
->bi_end_io
== NULL
||
2069 (wbio
->bi_end_io
== end_sync_read
&&
2070 (i
== r1_bio
->read_disk
||
2071 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
2074 wbio
->bi_rw
= WRITE
;
2075 wbio
->bi_end_io
= end_sync_write
;
2076 atomic_inc(&r1_bio
->remaining
);
2077 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, bio_sectors(wbio
));
2079 generic_make_request(wbio
);
2082 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
2083 /* if we're here, all write(s) have completed, so clean up */
2084 int s
= r1_bio
->sectors
;
2085 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2086 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2087 reschedule_retry(r1_bio
);
2090 md_done_sync(mddev
, s
, 1);
2096 * This is a kernel thread which:
2098 * 1. Retries failed read operations on working mirrors.
2099 * 2. Updates the raid superblock when problems encounter.
2100 * 3. Performs writes following reads for array synchronising.
2103 static void fix_read_error(struct r1conf
*conf
, int read_disk
,
2104 sector_t sect
, int sectors
)
2106 struct mddev
*mddev
= conf
->mddev
;
2112 struct md_rdev
*rdev
;
2114 if (s
> (PAGE_SIZE
>>9))
2118 /* Note: no rcu protection needed here
2119 * as this is synchronous in the raid1d thread
2120 * which is the thread that might remove
2121 * a device. If raid1d ever becomes multi-threaded....
2126 rdev
= conf
->mirrors
[d
].rdev
;
2128 (test_bit(In_sync
, &rdev
->flags
) ||
2129 (!test_bit(Faulty
, &rdev
->flags
) &&
2130 rdev
->recovery_offset
>= sect
+ s
)) &&
2131 is_badblock(rdev
, sect
, s
,
2132 &first_bad
, &bad_sectors
) == 0 &&
2133 sync_page_io(rdev
, sect
, s
<<9,
2134 conf
->tmppage
, READ
, false))
2138 if (d
== conf
->raid_disks
* 2)
2141 } while (!success
&& d
!= read_disk
);
2144 /* Cannot read from anywhere - mark it bad */
2145 struct md_rdev
*rdev
= conf
->mirrors
[read_disk
].rdev
;
2146 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
2147 md_error(mddev
, rdev
);
2150 /* write it back and re-read */
2152 while (d
!= read_disk
) {
2154 d
= conf
->raid_disks
* 2;
2156 rdev
= conf
->mirrors
[d
].rdev
;
2158 !test_bit(Faulty
, &rdev
->flags
))
2159 r1_sync_page_io(rdev
, sect
, s
,
2160 conf
->tmppage
, WRITE
);
2163 while (d
!= read_disk
) {
2164 char b
[BDEVNAME_SIZE
];
2166 d
= conf
->raid_disks
* 2;
2168 rdev
= conf
->mirrors
[d
].rdev
;
2170 !test_bit(Faulty
, &rdev
->flags
)) {
2171 if (r1_sync_page_io(rdev
, sect
, s
,
2172 conf
->tmppage
, READ
)) {
2173 atomic_add(s
, &rdev
->corrected_errors
);
2175 "md/raid1:%s: read error corrected "
2176 "(%d sectors at %llu on %s)\n",
2178 (unsigned long long)(sect
+
2180 bdevname(rdev
->bdev
, b
));
2189 static int narrow_write_error(struct r1bio
*r1_bio
, int i
)
2191 struct mddev
*mddev
= r1_bio
->mddev
;
2192 struct r1conf
*conf
= mddev
->private;
2193 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2195 /* bio has the data to be written to device 'i' where
2196 * we just recently had a write error.
2197 * We repeatedly clone the bio and trim down to one block,
2198 * then try the write. Where the write fails we record
2200 * It is conceivable that the bio doesn't exactly align with
2201 * blocks. We must handle this somehow.
2203 * We currently own a reference on the rdev.
2209 int sect_to_write
= r1_bio
->sectors
;
2212 if (rdev
->badblocks
.shift
< 0)
2215 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2216 bdev_logical_block_size(rdev
->bdev
) >> 9);
2217 sector
= r1_bio
->sector
;
2218 sectors
= ((sector
+ block_sectors
)
2219 & ~(sector_t
)(block_sectors
- 1))
2222 while (sect_to_write
) {
2224 if (sectors
> sect_to_write
)
2225 sectors
= sect_to_write
;
2226 /* Write at 'sector' for 'sectors'*/
2228 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
2229 unsigned vcnt
= r1_bio
->behind_page_count
;
2230 struct bio_vec
*vec
= r1_bio
->behind_bvecs
;
2232 while (!vec
->bv_page
) {
2237 wbio
= bio_alloc_mddev(GFP_NOIO
, vcnt
, mddev
);
2238 memcpy(wbio
->bi_io_vec
, vec
, vcnt
* sizeof(struct bio_vec
));
2240 wbio
->bi_vcnt
= vcnt
;
2242 wbio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2245 wbio
->bi_rw
= WRITE
;
2246 wbio
->bi_iter
.bi_sector
= r1_bio
->sector
;
2247 wbio
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
2249 bio_trim(wbio
, sector
- r1_bio
->sector
, sectors
);
2250 wbio
->bi_iter
.bi_sector
+= rdev
->data_offset
;
2251 wbio
->bi_bdev
= rdev
->bdev
;
2252 if (submit_bio_wait(WRITE
, wbio
) == 0)
2254 ok
= rdev_set_badblocks(rdev
, sector
,
2259 sect_to_write
-= sectors
;
2261 sectors
= block_sectors
;
2266 static void handle_sync_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2269 int s
= r1_bio
->sectors
;
2270 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++) {
2271 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2272 struct bio
*bio
= r1_bio
->bios
[m
];
2273 if (bio
->bi_end_io
== NULL
)
2275 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
2276 test_bit(R1BIO_MadeGood
, &r1_bio
->state
)) {
2277 rdev_clear_badblocks(rdev
, r1_bio
->sector
, s
, 0);
2279 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
2280 test_bit(R1BIO_WriteError
, &r1_bio
->state
)) {
2281 if (!rdev_set_badblocks(rdev
, r1_bio
->sector
, s
, 0))
2282 md_error(conf
->mddev
, rdev
);
2286 md_done_sync(conf
->mddev
, s
, 1);
2289 static void handle_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2292 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++)
2293 if (r1_bio
->bios
[m
] == IO_MADE_GOOD
) {
2294 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2295 rdev_clear_badblocks(rdev
,
2297 r1_bio
->sectors
, 0);
2298 rdev_dec_pending(rdev
, conf
->mddev
);
2299 } else if (r1_bio
->bios
[m
] != NULL
) {
2300 /* This drive got a write error. We need to
2301 * narrow down and record precise write
2304 if (!narrow_write_error(r1_bio
, m
)) {
2305 md_error(conf
->mddev
,
2306 conf
->mirrors
[m
].rdev
);
2307 /* an I/O failed, we can't clear the bitmap */
2308 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2310 rdev_dec_pending(conf
->mirrors
[m
].rdev
,
2313 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2314 close_write(r1_bio
);
2315 raid_end_bio_io(r1_bio
);
2318 static void handle_read_error(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2322 struct mddev
*mddev
= conf
->mddev
;
2324 char b
[BDEVNAME_SIZE
];
2325 struct md_rdev
*rdev
;
2327 clear_bit(R1BIO_ReadError
, &r1_bio
->state
);
2328 /* we got a read error. Maybe the drive is bad. Maybe just
2329 * the block and we can fix it.
2330 * We freeze all other IO, and try reading the block from
2331 * other devices. When we find one, we re-write
2332 * and check it that fixes the read error.
2333 * This is all done synchronously while the array is
2336 if (mddev
->ro
== 0) {
2337 freeze_array(conf
, 1);
2338 fix_read_error(conf
, r1_bio
->read_disk
,
2339 r1_bio
->sector
, r1_bio
->sectors
);
2340 unfreeze_array(conf
);
2342 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
2343 rdev_dec_pending(conf
->mirrors
[r1_bio
->read_disk
].rdev
, conf
->mddev
);
2345 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2346 bdevname(bio
->bi_bdev
, b
);
2348 disk
= read_balance(conf
, r1_bio
, &max_sectors
);
2350 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O"
2351 " read error for block %llu\n",
2352 mdname(mddev
), b
, (unsigned long long)r1_bio
->sector
);
2353 raid_end_bio_io(r1_bio
);
2355 const unsigned long do_sync
2356 = r1_bio
->master_bio
->bi_rw
& REQ_SYNC
;
2358 r1_bio
->bios
[r1_bio
->read_disk
] =
2359 mddev
->ro
? IO_BLOCKED
: NULL
;
2362 r1_bio
->read_disk
= disk
;
2363 bio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2364 bio_trim(bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
2366 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
2367 rdev
= conf
->mirrors
[disk
].rdev
;
2368 printk_ratelimited(KERN_ERR
2369 "md/raid1:%s: redirecting sector %llu"
2370 " to other mirror: %s\n",
2372 (unsigned long long)r1_bio
->sector
,
2373 bdevname(rdev
->bdev
, b
));
2374 bio
->bi_iter
.bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
2375 bio
->bi_bdev
= rdev
->bdev
;
2376 bio
->bi_end_io
= raid1_end_read_request
;
2377 bio
->bi_rw
= READ
| do_sync
;
2378 bio
->bi_private
= r1_bio
;
2379 if (max_sectors
< r1_bio
->sectors
) {
2380 /* Drat - have to split this up more */
2381 struct bio
*mbio
= r1_bio
->master_bio
;
2382 int sectors_handled
= (r1_bio
->sector
+ max_sectors
2383 - mbio
->bi_iter
.bi_sector
);
2384 r1_bio
->sectors
= max_sectors
;
2385 spin_lock_irq(&conf
->device_lock
);
2386 if (mbio
->bi_phys_segments
== 0)
2387 mbio
->bi_phys_segments
= 2;
2389 mbio
->bi_phys_segments
++;
2390 spin_unlock_irq(&conf
->device_lock
);
2391 generic_make_request(bio
);
2394 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
2396 r1_bio
->master_bio
= mbio
;
2397 r1_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2399 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
2400 r1_bio
->mddev
= mddev
;
2401 r1_bio
->sector
= mbio
->bi_iter
.bi_sector
+
2406 generic_make_request(bio
);
2410 static void raid1d(struct md_thread
*thread
)
2412 struct mddev
*mddev
= thread
->mddev
;
2413 struct r1bio
*r1_bio
;
2414 unsigned long flags
;
2415 struct r1conf
*conf
= mddev
->private;
2416 struct list_head
*head
= &conf
->retry_list
;
2417 struct blk_plug plug
;
2419 md_check_recovery(mddev
);
2421 blk_start_plug(&plug
);
2424 flush_pending_writes(conf
);
2426 spin_lock_irqsave(&conf
->device_lock
, flags
);
2427 if (list_empty(head
)) {
2428 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2431 r1_bio
= list_entry(head
->prev
, struct r1bio
, retry_list
);
2432 list_del(head
->prev
);
2434 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2436 mddev
= r1_bio
->mddev
;
2437 conf
= mddev
->private;
2438 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
2439 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2440 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2441 handle_sync_write_finished(conf
, r1_bio
);
2443 sync_request_write(mddev
, r1_bio
);
2444 } else if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2445 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2446 handle_write_finished(conf
, r1_bio
);
2447 else if (test_bit(R1BIO_ReadError
, &r1_bio
->state
))
2448 handle_read_error(conf
, r1_bio
);
2450 /* just a partial read to be scheduled from separate
2453 generic_make_request(r1_bio
->bios
[r1_bio
->read_disk
]);
2456 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2457 md_check_recovery(mddev
);
2459 blk_finish_plug(&plug
);
2462 static int init_resync(struct r1conf
*conf
)
2466 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2467 BUG_ON(conf
->r1buf_pool
);
2468 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
2470 if (!conf
->r1buf_pool
)
2472 conf
->next_resync
= 0;
2477 * perform a "sync" on one "block"
2479 * We need to make sure that no normal I/O request - particularly write
2480 * requests - conflict with active sync requests.
2482 * This is achieved by tracking pending requests and a 'barrier' concept
2483 * that can be installed to exclude normal IO requests.
2486 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
2488 struct r1conf
*conf
= mddev
->private;
2489 struct r1bio
*r1_bio
;
2491 sector_t max_sector
, nr_sectors
;
2495 int write_targets
= 0, read_targets
= 0;
2496 sector_t sync_blocks
;
2497 int still_degraded
= 0;
2498 int good_sectors
= RESYNC_SECTORS
;
2499 int min_bad
= 0; /* number of sectors that are bad in all devices */
2501 if (!conf
->r1buf_pool
)
2502 if (init_resync(conf
))
2505 max_sector
= mddev
->dev_sectors
;
2506 if (sector_nr
>= max_sector
) {
2507 /* If we aborted, we need to abort the
2508 * sync on the 'current' bitmap chunk (there will
2509 * only be one in raid1 resync.
2510 * We can find the current addess in mddev->curr_resync
2512 if (mddev
->curr_resync
< max_sector
) /* aborted */
2513 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2515 else /* completed sync */
2518 bitmap_close_sync(mddev
->bitmap
);
2523 if (mddev
->bitmap
== NULL
&&
2524 mddev
->recovery_cp
== MaxSector
&&
2525 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2526 conf
->fullsync
== 0) {
2528 return max_sector
- sector_nr
;
2530 /* before building a request, check if we can skip these blocks..
2531 * This call the bitmap_start_sync doesn't actually record anything
2533 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
2534 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2535 /* We can skip this block, and probably several more */
2540 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
2541 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
2543 raise_barrier(conf
, sector_nr
);
2547 * If we get a correctably read error during resync or recovery,
2548 * we might want to read from a different device. So we
2549 * flag all drives that could conceivably be read from for READ,
2550 * and any others (which will be non-In_sync devices) for WRITE.
2551 * If a read fails, we try reading from something else for which READ
2555 r1_bio
->mddev
= mddev
;
2556 r1_bio
->sector
= sector_nr
;
2558 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
2560 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2561 struct md_rdev
*rdev
;
2562 bio
= r1_bio
->bios
[i
];
2565 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
2567 test_bit(Faulty
, &rdev
->flags
)) {
2568 if (i
< conf
->raid_disks
)
2570 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
2572 bio
->bi_end_io
= end_sync_write
;
2575 /* may need to read from here */
2576 sector_t first_bad
= MaxSector
;
2579 if (is_badblock(rdev
, sector_nr
, good_sectors
,
2580 &first_bad
, &bad_sectors
)) {
2581 if (first_bad
> sector_nr
)
2582 good_sectors
= first_bad
- sector_nr
;
2584 bad_sectors
-= (sector_nr
- first_bad
);
2586 min_bad
> bad_sectors
)
2587 min_bad
= bad_sectors
;
2590 if (sector_nr
< first_bad
) {
2591 if (test_bit(WriteMostly
, &rdev
->flags
)) {
2599 bio
->bi_end_io
= end_sync_read
;
2601 } else if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
2602 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2603 !test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)) {
2605 * The device is suitable for reading (InSync),
2606 * but has bad block(s) here. Let's try to correct them,
2607 * if we are doing resync or repair. Otherwise, leave
2608 * this device alone for this sync request.
2611 bio
->bi_end_io
= end_sync_write
;
2615 if (bio
->bi_end_io
) {
2616 atomic_inc(&rdev
->nr_pending
);
2617 bio
->bi_iter
.bi_sector
= sector_nr
+ rdev
->data_offset
;
2618 bio
->bi_bdev
= rdev
->bdev
;
2619 bio
->bi_private
= r1_bio
;
2625 r1_bio
->read_disk
= disk
;
2627 if (read_targets
== 0 && min_bad
> 0) {
2628 /* These sectors are bad on all InSync devices, so we
2629 * need to mark them bad on all write targets
2632 for (i
= 0 ; i
< conf
->raid_disks
* 2 ; i
++)
2633 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_write
) {
2634 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2635 ok
= rdev_set_badblocks(rdev
, sector_nr
,
2639 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2644 /* Cannot record the badblocks, so need to
2646 * If there are multiple read targets, could just
2647 * fail the really bad ones ???
2649 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2650 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2656 if (min_bad
> 0 && min_bad
< good_sectors
) {
2657 /* only resync enough to reach the next bad->good
2659 good_sectors
= min_bad
;
2662 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
2663 /* extra read targets are also write targets */
2664 write_targets
+= read_targets
-1;
2666 if (write_targets
== 0 || read_targets
== 0) {
2667 /* There is nowhere to write, so all non-sync
2668 * drives must be failed - so we are finished
2672 max_sector
= sector_nr
+ min_bad
;
2673 rv
= max_sector
- sector_nr
;
2679 if (max_sector
> mddev
->resync_max
)
2680 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2681 if (max_sector
> sector_nr
+ good_sectors
)
2682 max_sector
= sector_nr
+ good_sectors
;
2687 int len
= PAGE_SIZE
;
2688 if (sector_nr
+ (len
>>9) > max_sector
)
2689 len
= (max_sector
- sector_nr
) << 9;
2692 if (sync_blocks
== 0) {
2693 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2694 &sync_blocks
, still_degraded
) &&
2696 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2698 BUG_ON(sync_blocks
< (PAGE_SIZE
>>9));
2699 if ((len
>> 9) > sync_blocks
)
2700 len
= sync_blocks
<<9;
2703 for (i
= 0 ; i
< conf
->raid_disks
* 2; i
++) {
2704 bio
= r1_bio
->bios
[i
];
2705 if (bio
->bi_end_io
) {
2706 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2707 if (bio_add_page(bio
, page
, len
, 0) == 0) {
2709 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2712 bio
= r1_bio
->bios
[i
];
2713 if (bio
->bi_end_io
==NULL
)
2715 /* remove last page from this bio */
2717 bio
->bi_iter
.bi_size
-= len
;
2718 __clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
2724 nr_sectors
+= len
>>9;
2725 sector_nr
+= len
>>9;
2726 sync_blocks
-= (len
>>9);
2727 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
2729 r1_bio
->sectors
= nr_sectors
;
2731 /* For a user-requested sync, we read all readable devices and do a
2734 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2735 atomic_set(&r1_bio
->remaining
, read_targets
);
2736 for (i
= 0; i
< conf
->raid_disks
* 2 && read_targets
; i
++) {
2737 bio
= r1_bio
->bios
[i
];
2738 if (bio
->bi_end_io
== end_sync_read
) {
2740 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2741 generic_make_request(bio
);
2745 atomic_set(&r1_bio
->remaining
, 1);
2746 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2747 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2748 generic_make_request(bio
);
2754 static sector_t
raid1_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2759 return mddev
->dev_sectors
;
2762 static struct r1conf
*setup_conf(struct mddev
*mddev
)
2764 struct r1conf
*conf
;
2766 struct raid1_info
*disk
;
2767 struct md_rdev
*rdev
;
2770 conf
= kzalloc(sizeof(struct r1conf
), GFP_KERNEL
);
2774 conf
->mirrors
= kzalloc(sizeof(struct raid1_info
)
2775 * mddev
->raid_disks
* 2,
2780 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2784 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
2785 if (!conf
->poolinfo
)
2787 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
* 2;
2788 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2791 if (!conf
->r1bio_pool
)
2794 conf
->poolinfo
->mddev
= mddev
;
2797 spin_lock_init(&conf
->device_lock
);
2798 rdev_for_each(rdev
, mddev
) {
2799 struct request_queue
*q
;
2800 int disk_idx
= rdev
->raid_disk
;
2801 if (disk_idx
>= mddev
->raid_disks
2804 if (test_bit(Replacement
, &rdev
->flags
))
2805 disk
= conf
->mirrors
+ mddev
->raid_disks
+ disk_idx
;
2807 disk
= conf
->mirrors
+ disk_idx
;
2812 q
= bdev_get_queue(rdev
->bdev
);
2813 if (q
->merge_bvec_fn
)
2814 mddev
->merge_check_needed
= 1;
2816 disk
->head_position
= 0;
2817 disk
->seq_start
= MaxSector
;
2819 conf
->raid_disks
= mddev
->raid_disks
;
2820 conf
->mddev
= mddev
;
2821 INIT_LIST_HEAD(&conf
->retry_list
);
2823 spin_lock_init(&conf
->resync_lock
);
2824 init_waitqueue_head(&conf
->wait_barrier
);
2826 bio_list_init(&conf
->pending_bio_list
);
2827 conf
->pending_count
= 0;
2828 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
2830 conf
->start_next_window
= MaxSector
;
2831 conf
->current_window_requests
= conf
->next_window_requests
= 0;
2834 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2836 disk
= conf
->mirrors
+ i
;
2838 if (i
< conf
->raid_disks
&&
2839 disk
[conf
->raid_disks
].rdev
) {
2840 /* This slot has a replacement. */
2842 /* No original, just make the replacement
2843 * a recovering spare
2846 disk
[conf
->raid_disks
].rdev
;
2847 disk
[conf
->raid_disks
].rdev
= NULL
;
2848 } else if (!test_bit(In_sync
, &disk
->rdev
->flags
))
2849 /* Original is not in_sync - bad */
2854 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2855 disk
->head_position
= 0;
2857 (disk
->rdev
->saved_raid_disk
< 0))
2863 conf
->thread
= md_register_thread(raid1d
, mddev
, "raid1");
2864 if (!conf
->thread
) {
2866 "md/raid1:%s: couldn't allocate thread\n",
2875 if (conf
->r1bio_pool
)
2876 mempool_destroy(conf
->r1bio_pool
);
2877 kfree(conf
->mirrors
);
2878 safe_put_page(conf
->tmppage
);
2879 kfree(conf
->poolinfo
);
2882 return ERR_PTR(err
);
2885 static void raid1_free(struct mddev
*mddev
, void *priv
);
2886 static int run(struct mddev
*mddev
)
2888 struct r1conf
*conf
;
2890 struct md_rdev
*rdev
;
2892 bool discard_supported
= false;
2894 if (mddev
->level
!= 1) {
2895 printk(KERN_ERR
"md/raid1:%s: raid level not set to mirroring (%d)\n",
2896 mdname(mddev
), mddev
->level
);
2899 if (mddev
->reshape_position
!= MaxSector
) {
2900 printk(KERN_ERR
"md/raid1:%s: reshape_position set but not supported\n",
2905 * copy the already verified devices into our private RAID1
2906 * bookkeeping area. [whatever we allocate in run(),
2907 * should be freed in raid1_free()]
2909 if (mddev
->private == NULL
)
2910 conf
= setup_conf(mddev
);
2912 conf
= mddev
->private;
2915 return PTR_ERR(conf
);
2918 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
2920 rdev_for_each(rdev
, mddev
) {
2921 if (!mddev
->gendisk
)
2923 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2924 rdev
->data_offset
<< 9);
2925 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
2926 discard_supported
= true;
2929 mddev
->degraded
= 0;
2930 for (i
=0; i
< conf
->raid_disks
; i
++)
2931 if (conf
->mirrors
[i
].rdev
== NULL
||
2932 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
2933 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2936 if (conf
->raid_disks
- mddev
->degraded
== 1)
2937 mddev
->recovery_cp
= MaxSector
;
2939 if (mddev
->recovery_cp
!= MaxSector
)
2940 printk(KERN_NOTICE
"md/raid1:%s: not clean"
2941 " -- starting background reconstruction\n",
2944 "md/raid1:%s: active with %d out of %d mirrors\n",
2945 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2949 * Ok, everything is just fine now
2951 mddev
->thread
= conf
->thread
;
2952 conf
->thread
= NULL
;
2953 mddev
->private = conf
;
2955 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
2958 if (discard_supported
)
2959 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
2962 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
2966 ret
= md_integrity_register(mddev
);
2968 md_unregister_thread(&mddev
->thread
);
2969 raid1_free(mddev
, conf
);
2974 static void raid1_free(struct mddev
*mddev
, void *priv
)
2976 struct r1conf
*conf
= priv
;
2978 if (conf
->r1bio_pool
)
2979 mempool_destroy(conf
->r1bio_pool
);
2980 kfree(conf
->mirrors
);
2981 safe_put_page(conf
->tmppage
);
2982 kfree(conf
->poolinfo
);
2986 static int raid1_resize(struct mddev
*mddev
, sector_t sectors
)
2988 /* no resync is happening, and there is enough space
2989 * on all devices, so we can resize.
2990 * We need to make sure resync covers any new space.
2991 * If the array is shrinking we should possibly wait until
2992 * any io in the removed space completes, but it hardly seems
2995 sector_t newsize
= raid1_size(mddev
, sectors
, 0);
2996 if (mddev
->external_size
&&
2997 mddev
->array_sectors
> newsize
)
2999 if (mddev
->bitmap
) {
3000 int ret
= bitmap_resize(mddev
->bitmap
, newsize
, 0, 0);
3004 md_set_array_sectors(mddev
, newsize
);
3005 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3006 revalidate_disk(mddev
->gendisk
);
3007 if (sectors
> mddev
->dev_sectors
&&
3008 mddev
->recovery_cp
> mddev
->dev_sectors
) {
3009 mddev
->recovery_cp
= mddev
->dev_sectors
;
3010 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3012 mddev
->dev_sectors
= sectors
;
3013 mddev
->resync_max_sectors
= sectors
;
3017 static int raid1_reshape(struct mddev
*mddev
)
3020 * 1/ resize the r1bio_pool
3021 * 2/ resize conf->mirrors
3023 * We allocate a new r1bio_pool if we can.
3024 * Then raise a device barrier and wait until all IO stops.
3025 * Then resize conf->mirrors and swap in the new r1bio pool.
3027 * At the same time, we "pack" the devices so that all the missing
3028 * devices have the higher raid_disk numbers.
3030 mempool_t
*newpool
, *oldpool
;
3031 struct pool_info
*newpoolinfo
;
3032 struct raid1_info
*newmirrors
;
3033 struct r1conf
*conf
= mddev
->private;
3034 int cnt
, raid_disks
;
3035 unsigned long flags
;
3038 /* Cannot change chunk_size, layout, or level */
3039 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
3040 mddev
->layout
!= mddev
->new_layout
||
3041 mddev
->level
!= mddev
->new_level
) {
3042 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3043 mddev
->new_layout
= mddev
->layout
;
3044 mddev
->new_level
= mddev
->level
;
3048 err
= md_allow_write(mddev
);
3052 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3054 if (raid_disks
< conf
->raid_disks
) {
3056 for (d
= 0; d
< conf
->raid_disks
; d
++)
3057 if (conf
->mirrors
[d
].rdev
)
3059 if (cnt
> raid_disks
)
3063 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
3066 newpoolinfo
->mddev
= mddev
;
3067 newpoolinfo
->raid_disks
= raid_disks
* 2;
3069 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
3070 r1bio_pool_free
, newpoolinfo
);
3075 newmirrors
= kzalloc(sizeof(struct raid1_info
) * raid_disks
* 2,
3079 mempool_destroy(newpool
);
3083 freeze_array(conf
, 0);
3085 /* ok, everything is stopped */
3086 oldpool
= conf
->r1bio_pool
;
3087 conf
->r1bio_pool
= newpool
;
3089 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
3090 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
3091 if (rdev
&& rdev
->raid_disk
!= d2
) {
3092 sysfs_unlink_rdev(mddev
, rdev
);
3093 rdev
->raid_disk
= d2
;
3094 sysfs_unlink_rdev(mddev
, rdev
);
3095 if (sysfs_link_rdev(mddev
, rdev
))
3097 "md/raid1:%s: cannot register rd%d\n",
3098 mdname(mddev
), rdev
->raid_disk
);
3101 newmirrors
[d2
++].rdev
= rdev
;
3103 kfree(conf
->mirrors
);
3104 conf
->mirrors
= newmirrors
;
3105 kfree(conf
->poolinfo
);
3106 conf
->poolinfo
= newpoolinfo
;
3108 spin_lock_irqsave(&conf
->device_lock
, flags
);
3109 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
3110 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3111 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
3112 mddev
->delta_disks
= 0;
3114 unfreeze_array(conf
);
3116 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3117 md_wakeup_thread(mddev
->thread
);
3119 mempool_destroy(oldpool
);
3123 static void raid1_quiesce(struct mddev
*mddev
, int state
)
3125 struct r1conf
*conf
= mddev
->private;
3128 case 2: /* wake for suspend */
3129 wake_up(&conf
->wait_barrier
);
3132 freeze_array(conf
, 0);
3135 unfreeze_array(conf
);
3140 static void *raid1_takeover(struct mddev
*mddev
)
3142 /* raid1 can take over:
3143 * raid5 with 2 devices, any layout or chunk size
3145 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
3146 struct r1conf
*conf
;
3147 mddev
->new_level
= 1;
3148 mddev
->new_layout
= 0;
3149 mddev
->new_chunk_sectors
= 0;
3150 conf
= setup_conf(mddev
);
3152 /* Array must appear to be quiesced */
3153 conf
->array_frozen
= 1;
3156 return ERR_PTR(-EINVAL
);
3159 static struct md_personality raid1_personality
=
3163 .owner
= THIS_MODULE
,
3164 .make_request
= make_request
,
3168 .error_handler
= error
,
3169 .hot_add_disk
= raid1_add_disk
,
3170 .hot_remove_disk
= raid1_remove_disk
,
3171 .spare_active
= raid1_spare_active
,
3172 .sync_request
= sync_request
,
3173 .resize
= raid1_resize
,
3175 .check_reshape
= raid1_reshape
,
3176 .quiesce
= raid1_quiesce
,
3177 .takeover
= raid1_takeover
,
3178 .congested
= raid1_congested
,
3179 .mergeable_bvec
= raid1_mergeable_bvec
,
3182 static int __init
raid_init(void)
3184 return register_md_personality(&raid1_personality
);
3187 static void raid_exit(void)
3189 unregister_md_personality(&raid1_personality
);
3192 module_init(raid_init
);
3193 module_exit(raid_exit
);
3194 MODULE_LICENSE("GPL");
3195 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3196 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3197 MODULE_ALIAS("md-raid1");
3198 MODULE_ALIAS("md-level-1");
3200 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);