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>
40 #include <linux/sched/signal.h>
42 #include <trace/events/block.h>
48 #define UNSUPPORTED_MDDEV_FLAGS \
49 ((1L << MD_HAS_JOURNAL) | \
50 (1L << MD_JOURNAL_CLEAN) | \
54 * Number of guaranteed r1bios in case of extreme VM load:
56 #define NR_RAID1_BIOS 256
58 /* when we get a read error on a read-only array, we redirect to another
59 * device without failing the first device, or trying to over-write to
60 * correct the read error. To keep track of bad blocks on a per-bio
61 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
63 #define IO_BLOCKED ((struct bio *)1)
64 /* When we successfully write to a known bad-block, we need to remove the
65 * bad-block marking which must be done from process context. So we record
66 * the success by setting devs[n].bio to IO_MADE_GOOD
68 #define IO_MADE_GOOD ((struct bio *)2)
70 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
72 /* When there are this many requests queue to be written by
73 * the raid1 thread, we become 'congested' to provide back-pressure
76 static int max_queued_requests
= 1024;
78 static void allow_barrier(struct r1conf
*conf
, sector_t sector_nr
);
79 static void lower_barrier(struct r1conf
*conf
, sector_t sector_nr
);
81 #define raid1_log(md, fmt, args...) \
82 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
85 * 'strct resync_pages' stores actual pages used for doing the resync
86 * IO, and it is per-bio, so make .bi_private points to it.
88 static inline struct resync_pages
*get_resync_pages(struct bio
*bio
)
90 return bio
->bi_private
;
94 * for resync bio, r1bio pointer can be retrieved from the per-bio
95 * 'struct resync_pages'.
97 static inline struct r1bio
*get_resync_r1bio(struct bio
*bio
)
99 return get_resync_pages(bio
)->raid_bio
;
102 static void * r1bio_pool_alloc(gfp_t gfp_flags
, void *data
)
104 struct pool_info
*pi
= data
;
105 int size
= offsetof(struct r1bio
, bios
[pi
->raid_disks
]);
107 /* allocate a r1bio with room for raid_disks entries in the bios array */
108 return kzalloc(size
, gfp_flags
);
111 static void r1bio_pool_free(void *r1_bio
, void *data
)
116 #define RESYNC_DEPTH 32
117 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
118 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
119 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
120 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
121 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
123 static void * r1buf_pool_alloc(gfp_t gfp_flags
, void *data
)
125 struct pool_info
*pi
= data
;
126 struct r1bio
*r1_bio
;
130 struct resync_pages
*rps
;
132 r1_bio
= r1bio_pool_alloc(gfp_flags
, pi
);
136 rps
= kmalloc(sizeof(struct resync_pages
) * pi
->raid_disks
,
142 * Allocate bios : 1 for reading, n-1 for writing
144 for (j
= pi
->raid_disks
; j
-- ; ) {
145 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
148 r1_bio
->bios
[j
] = bio
;
151 * Allocate RESYNC_PAGES data pages and attach them to
153 * If this is a user-requested check/repair, allocate
154 * RESYNC_PAGES for each bio.
156 if (test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
))
157 need_pages
= pi
->raid_disks
;
160 for (j
= 0; j
< pi
->raid_disks
; j
++) {
161 struct resync_pages
*rp
= &rps
[j
];
163 bio
= r1_bio
->bios
[j
];
165 if (j
< need_pages
) {
166 if (resync_alloc_pages(rp
, gfp_flags
))
169 memcpy(rp
, &rps
[0], sizeof(*rp
));
170 resync_get_all_pages(rp
);
174 rp
->raid_bio
= r1_bio
;
175 bio
->bi_private
= rp
;
178 r1_bio
->master_bio
= NULL
;
184 resync_free_pages(&rps
[j
]);
187 while (++j
< pi
->raid_disks
)
188 bio_put(r1_bio
->bios
[j
]);
192 r1bio_pool_free(r1_bio
, data
);
196 static void r1buf_pool_free(void *__r1_bio
, void *data
)
198 struct pool_info
*pi
= data
;
200 struct r1bio
*r1bio
= __r1_bio
;
201 struct resync_pages
*rp
= NULL
;
203 for (i
= pi
->raid_disks
; i
--; ) {
204 rp
= get_resync_pages(r1bio
->bios
[i
]);
205 resync_free_pages(rp
);
206 bio_put(r1bio
->bios
[i
]);
209 /* resync pages array stored in the 1st bio's .bi_private */
212 r1bio_pool_free(r1bio
, data
);
215 static void put_all_bios(struct r1conf
*conf
, struct r1bio
*r1_bio
)
219 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
220 struct bio
**bio
= r1_bio
->bios
+ i
;
221 if (!BIO_SPECIAL(*bio
))
227 static void free_r1bio(struct r1bio
*r1_bio
)
229 struct r1conf
*conf
= r1_bio
->mddev
->private;
231 put_all_bios(conf
, r1_bio
);
232 mempool_free(r1_bio
, conf
->r1bio_pool
);
235 static void put_buf(struct r1bio
*r1_bio
)
237 struct r1conf
*conf
= r1_bio
->mddev
->private;
238 sector_t sect
= r1_bio
->sector
;
241 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
242 struct bio
*bio
= r1_bio
->bios
[i
];
244 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
247 mempool_free(r1_bio
, conf
->r1buf_pool
);
249 lower_barrier(conf
, sect
);
252 static void reschedule_retry(struct r1bio
*r1_bio
)
255 struct mddev
*mddev
= r1_bio
->mddev
;
256 struct r1conf
*conf
= mddev
->private;
259 idx
= sector_to_idx(r1_bio
->sector
);
260 spin_lock_irqsave(&conf
->device_lock
, flags
);
261 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
262 atomic_inc(&conf
->nr_queued
[idx
]);
263 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
265 wake_up(&conf
->wait_barrier
);
266 md_wakeup_thread(mddev
->thread
);
270 * raid_end_bio_io() is called when we have finished servicing a mirrored
271 * operation and are ready to return a success/failure code to the buffer
274 static void call_bio_endio(struct r1bio
*r1_bio
)
276 struct bio
*bio
= r1_bio
->master_bio
;
277 struct r1conf
*conf
= r1_bio
->mddev
->private;
279 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
280 bio
->bi_error
= -EIO
;
284 * Wake up any possible resync thread that waits for the device
287 allow_barrier(conf
, r1_bio
->sector
);
290 static void raid_end_bio_io(struct r1bio
*r1_bio
)
292 struct bio
*bio
= r1_bio
->master_bio
;
294 /* if nobody has done the final endio yet, do it now */
295 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
296 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
297 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
298 (unsigned long long) bio
->bi_iter
.bi_sector
,
299 (unsigned long long) bio_end_sector(bio
) - 1);
301 call_bio_endio(r1_bio
);
307 * Update disk head position estimator based on IRQ completion info.
309 static inline void update_head_pos(int disk
, struct r1bio
*r1_bio
)
311 struct r1conf
*conf
= r1_bio
->mddev
->private;
313 conf
->mirrors
[disk
].head_position
=
314 r1_bio
->sector
+ (r1_bio
->sectors
);
318 * Find the disk number which triggered given bio
320 static int find_bio_disk(struct r1bio
*r1_bio
, struct bio
*bio
)
323 struct r1conf
*conf
= r1_bio
->mddev
->private;
324 int raid_disks
= conf
->raid_disks
;
326 for (mirror
= 0; mirror
< raid_disks
* 2; mirror
++)
327 if (r1_bio
->bios
[mirror
] == bio
)
330 BUG_ON(mirror
== raid_disks
* 2);
331 update_head_pos(mirror
, r1_bio
);
336 static void raid1_end_read_request(struct bio
*bio
)
338 int uptodate
= !bio
->bi_error
;
339 struct r1bio
*r1_bio
= bio
->bi_private
;
340 struct r1conf
*conf
= r1_bio
->mddev
->private;
341 struct md_rdev
*rdev
= conf
->mirrors
[r1_bio
->read_disk
].rdev
;
344 * this branch is our 'one mirror IO has finished' event handler:
346 update_head_pos(r1_bio
->read_disk
, r1_bio
);
349 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
350 else if (test_bit(FailFast
, &rdev
->flags
) &&
351 test_bit(R1BIO_FailFast
, &r1_bio
->state
))
352 /* This was a fail-fast read so we definitely
356 /* If all other devices have failed, we want to return
357 * the error upwards rather than fail the last device.
358 * Here we redefine "uptodate" to mean "Don't want to retry"
361 spin_lock_irqsave(&conf
->device_lock
, flags
);
362 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
363 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
364 test_bit(In_sync
, &rdev
->flags
)))
366 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
370 raid_end_bio_io(r1_bio
);
371 rdev_dec_pending(rdev
, conf
->mddev
);
376 char b
[BDEVNAME_SIZE
];
377 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
379 bdevname(rdev
->bdev
, b
),
380 (unsigned long long)r1_bio
->sector
);
381 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
382 reschedule_retry(r1_bio
);
383 /* don't drop the reference on read_disk yet */
387 static void close_write(struct r1bio
*r1_bio
)
389 /* it really is the end of this request */
390 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
391 bio_free_pages(r1_bio
->behind_master_bio
);
392 bio_put(r1_bio
->behind_master_bio
);
393 r1_bio
->behind_master_bio
= NULL
;
395 /* clear the bitmap if all writes complete successfully */
396 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
398 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
399 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
400 md_write_end(r1_bio
->mddev
);
403 static void r1_bio_write_done(struct r1bio
*r1_bio
)
405 if (!atomic_dec_and_test(&r1_bio
->remaining
))
408 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
409 reschedule_retry(r1_bio
);
412 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
))
413 reschedule_retry(r1_bio
);
415 raid_end_bio_io(r1_bio
);
419 static void raid1_end_write_request(struct bio
*bio
)
421 struct r1bio
*r1_bio
= bio
->bi_private
;
422 int behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
423 struct r1conf
*conf
= r1_bio
->mddev
->private;
424 struct bio
*to_put
= NULL
;
425 int mirror
= find_bio_disk(r1_bio
, bio
);
426 struct md_rdev
*rdev
= conf
->mirrors
[mirror
].rdev
;
429 discard_error
= bio
->bi_error
&& bio_op(bio
) == REQ_OP_DISCARD
;
432 * 'one mirror IO has finished' event handler:
434 if (bio
->bi_error
&& !discard_error
) {
435 set_bit(WriteErrorSeen
, &rdev
->flags
);
436 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
437 set_bit(MD_RECOVERY_NEEDED
, &
438 conf
->mddev
->recovery
);
440 if (test_bit(FailFast
, &rdev
->flags
) &&
441 (bio
->bi_opf
& MD_FAILFAST
) &&
442 /* We never try FailFast to WriteMostly devices */
443 !test_bit(WriteMostly
, &rdev
->flags
)) {
444 md_error(r1_bio
->mddev
, rdev
);
445 if (!test_bit(Faulty
, &rdev
->flags
))
446 /* This is the only remaining device,
447 * We need to retry the write without
450 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
452 /* Finished with this branch */
453 r1_bio
->bios
[mirror
] = NULL
;
457 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
460 * Set R1BIO_Uptodate in our master bio, so that we
461 * will return a good error code for to the higher
462 * levels even if IO on some other mirrored buffer
465 * The 'master' represents the composite IO operation
466 * to user-side. So if something waits for IO, then it
467 * will wait for the 'master' bio.
472 r1_bio
->bios
[mirror
] = NULL
;
475 * Do not set R1BIO_Uptodate if the current device is
476 * rebuilding or Faulty. This is because we cannot use
477 * such device for properly reading the data back (we could
478 * potentially use it, if the current write would have felt
479 * before rdev->recovery_offset, but for simplicity we don't
482 if (test_bit(In_sync
, &rdev
->flags
) &&
483 !test_bit(Faulty
, &rdev
->flags
))
484 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
486 /* Maybe we can clear some bad blocks. */
487 if (is_badblock(rdev
, r1_bio
->sector
, r1_bio
->sectors
,
488 &first_bad
, &bad_sectors
) && !discard_error
) {
489 r1_bio
->bios
[mirror
] = IO_MADE_GOOD
;
490 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
495 /* we release behind master bio when all write are done */
496 if (r1_bio
->behind_master_bio
== bio
)
499 if (test_bit(WriteMostly
, &rdev
->flags
))
500 atomic_dec(&r1_bio
->behind_remaining
);
503 * In behind mode, we ACK the master bio once the I/O
504 * has safely reached all non-writemostly
505 * disks. Setting the Returned bit ensures that this
506 * gets done only once -- we don't ever want to return
507 * -EIO here, instead we'll wait
509 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
510 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
511 /* Maybe we can return now */
512 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
513 struct bio
*mbio
= r1_bio
->master_bio
;
514 pr_debug("raid1: behind end write sectors"
516 (unsigned long long) mbio
->bi_iter
.bi_sector
,
517 (unsigned long long) bio_end_sector(mbio
) - 1);
518 call_bio_endio(r1_bio
);
522 if (r1_bio
->bios
[mirror
] == NULL
)
523 rdev_dec_pending(rdev
, conf
->mddev
);
526 * Let's see if all mirrored write operations have finished
529 r1_bio_write_done(r1_bio
);
535 static sector_t
align_to_barrier_unit_end(sector_t start_sector
,
540 WARN_ON(sectors
== 0);
542 * len is the number of sectors from start_sector to end of the
543 * barrier unit which start_sector belongs to.
545 len
= round_up(start_sector
+ 1, BARRIER_UNIT_SECTOR_SIZE
) -
555 * This routine returns the disk from which the requested read should
556 * be done. There is a per-array 'next expected sequential IO' sector
557 * number - if this matches on the next IO then we use the last disk.
558 * There is also a per-disk 'last know head position' sector that is
559 * maintained from IRQ contexts, both the normal and the resync IO
560 * completion handlers update this position correctly. If there is no
561 * perfect sequential match then we pick the disk whose head is closest.
563 * If there are 2 mirrors in the same 2 devices, performance degrades
564 * because position is mirror, not device based.
566 * The rdev for the device selected will have nr_pending incremented.
568 static int read_balance(struct r1conf
*conf
, struct r1bio
*r1_bio
, int *max_sectors
)
570 const sector_t this_sector
= r1_bio
->sector
;
572 int best_good_sectors
;
573 int best_disk
, best_dist_disk
, best_pending_disk
;
577 unsigned int min_pending
;
578 struct md_rdev
*rdev
;
580 int choose_next_idle
;
584 * Check if we can balance. We can balance on the whole
585 * device if no resync is going on, or below the resync window.
586 * We take the first readable disk when above the resync window.
589 sectors
= r1_bio
->sectors
;
592 best_dist
= MaxSector
;
593 best_pending_disk
= -1;
594 min_pending
= UINT_MAX
;
595 best_good_sectors
= 0;
597 choose_next_idle
= 0;
598 clear_bit(R1BIO_FailFast
, &r1_bio
->state
);
600 if ((conf
->mddev
->recovery_cp
< this_sector
+ sectors
) ||
601 (mddev_is_clustered(conf
->mddev
) &&
602 md_cluster_ops
->area_resyncing(conf
->mddev
, READ
, this_sector
,
603 this_sector
+ sectors
)))
608 for (disk
= 0 ; disk
< conf
->raid_disks
* 2 ; disk
++) {
612 unsigned int pending
;
615 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
616 if (r1_bio
->bios
[disk
] == IO_BLOCKED
618 || test_bit(Faulty
, &rdev
->flags
))
620 if (!test_bit(In_sync
, &rdev
->flags
) &&
621 rdev
->recovery_offset
< this_sector
+ sectors
)
623 if (test_bit(WriteMostly
, &rdev
->flags
)) {
624 /* Don't balance among write-mostly, just
625 * use the first as a last resort */
626 if (best_dist_disk
< 0) {
627 if (is_badblock(rdev
, this_sector
, sectors
,
628 &first_bad
, &bad_sectors
)) {
629 if (first_bad
<= this_sector
)
630 /* Cannot use this */
632 best_good_sectors
= first_bad
- this_sector
;
634 best_good_sectors
= sectors
;
635 best_dist_disk
= disk
;
636 best_pending_disk
= disk
;
640 /* This is a reasonable device to use. It might
643 if (is_badblock(rdev
, this_sector
, sectors
,
644 &first_bad
, &bad_sectors
)) {
645 if (best_dist
< MaxSector
)
646 /* already have a better device */
648 if (first_bad
<= this_sector
) {
649 /* cannot read here. If this is the 'primary'
650 * device, then we must not read beyond
651 * bad_sectors from another device..
653 bad_sectors
-= (this_sector
- first_bad
);
654 if (choose_first
&& sectors
> bad_sectors
)
655 sectors
= bad_sectors
;
656 if (best_good_sectors
> sectors
)
657 best_good_sectors
= sectors
;
660 sector_t good_sectors
= first_bad
- this_sector
;
661 if (good_sectors
> best_good_sectors
) {
662 best_good_sectors
= good_sectors
;
670 best_good_sectors
= sectors
;
673 /* At least two disks to choose from so failfast is OK */
674 set_bit(R1BIO_FailFast
, &r1_bio
->state
);
676 nonrot
= blk_queue_nonrot(bdev_get_queue(rdev
->bdev
));
677 has_nonrot_disk
|= nonrot
;
678 pending
= atomic_read(&rdev
->nr_pending
);
679 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
684 /* Don't change to another disk for sequential reads */
685 if (conf
->mirrors
[disk
].next_seq_sect
== this_sector
687 int opt_iosize
= bdev_io_opt(rdev
->bdev
) >> 9;
688 struct raid1_info
*mirror
= &conf
->mirrors
[disk
];
692 * If buffered sequential IO size exceeds optimal
693 * iosize, check if there is idle disk. If yes, choose
694 * the idle disk. read_balance could already choose an
695 * idle disk before noticing it's a sequential IO in
696 * this disk. This doesn't matter because this disk
697 * will idle, next time it will be utilized after the
698 * first disk has IO size exceeds optimal iosize. In
699 * this way, iosize of the first disk will be optimal
700 * iosize at least. iosize of the second disk might be
701 * small, but not a big deal since when the second disk
702 * starts IO, the first disk is likely still busy.
704 if (nonrot
&& opt_iosize
> 0 &&
705 mirror
->seq_start
!= MaxSector
&&
706 mirror
->next_seq_sect
> opt_iosize
&&
707 mirror
->next_seq_sect
- opt_iosize
>=
709 choose_next_idle
= 1;
715 if (choose_next_idle
)
718 if (min_pending
> pending
) {
719 min_pending
= pending
;
720 best_pending_disk
= disk
;
723 if (dist
< best_dist
) {
725 best_dist_disk
= disk
;
730 * If all disks are rotational, choose the closest disk. If any disk is
731 * non-rotational, choose the disk with less pending request even the
732 * disk is rotational, which might/might not be optimal for raids with
733 * mixed ratation/non-rotational disks depending on workload.
735 if (best_disk
== -1) {
736 if (has_nonrot_disk
|| min_pending
== 0)
737 best_disk
= best_pending_disk
;
739 best_disk
= best_dist_disk
;
742 if (best_disk
>= 0) {
743 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
746 atomic_inc(&rdev
->nr_pending
);
747 sectors
= best_good_sectors
;
749 if (conf
->mirrors
[best_disk
].next_seq_sect
!= this_sector
)
750 conf
->mirrors
[best_disk
].seq_start
= this_sector
;
752 conf
->mirrors
[best_disk
].next_seq_sect
= this_sector
+ sectors
;
755 *max_sectors
= sectors
;
760 static int raid1_congested(struct mddev
*mddev
, int bits
)
762 struct r1conf
*conf
= mddev
->private;
765 if ((bits
& (1 << WB_async_congested
)) &&
766 conf
->pending_count
>= max_queued_requests
)
770 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
771 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
772 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
773 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
777 /* Note the '|| 1' - when read_balance prefers
778 * non-congested targets, it can be removed
780 if ((bits
& (1 << WB_async_congested
)) || 1)
781 ret
|= bdi_congested(q
->backing_dev_info
, bits
);
783 ret
&= bdi_congested(q
->backing_dev_info
, bits
);
790 static void flush_bio_list(struct r1conf
*conf
, struct bio
*bio
)
792 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
793 bitmap_unplug(conf
->mddev
->bitmap
);
794 wake_up(&conf
->wait_barrier
);
796 while (bio
) { /* submit pending writes */
797 struct bio
*next
= bio
->bi_next
;
798 struct md_rdev
*rdev
= (void*)bio
->bi_bdev
;
800 bio
->bi_bdev
= rdev
->bdev
;
801 if (test_bit(Faulty
, &rdev
->flags
)) {
802 bio
->bi_error
= -EIO
;
804 } else if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
805 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
809 generic_make_request(bio
);
814 static void flush_pending_writes(struct r1conf
*conf
)
816 /* Any writes that have been queued but are awaiting
817 * bitmap updates get flushed here.
819 spin_lock_irq(&conf
->device_lock
);
821 if (conf
->pending_bio_list
.head
) {
823 bio
= bio_list_get(&conf
->pending_bio_list
);
824 conf
->pending_count
= 0;
825 spin_unlock_irq(&conf
->device_lock
);
826 flush_bio_list(conf
, bio
);
828 spin_unlock_irq(&conf
->device_lock
);
832 * Sometimes we need to suspend IO while we do something else,
833 * either some resync/recovery, or reconfigure the array.
834 * To do this we raise a 'barrier'.
835 * The 'barrier' is a counter that can be raised multiple times
836 * to count how many activities are happening which preclude
838 * We can only raise the barrier if there is no pending IO.
839 * i.e. if nr_pending == 0.
840 * We choose only to raise the barrier if no-one is waiting for the
841 * barrier to go down. This means that as soon as an IO request
842 * is ready, no other operations which require a barrier will start
843 * until the IO request has had a chance.
845 * So: regular IO calls 'wait_barrier'. When that returns there
846 * is no backgroup IO happening, It must arrange to call
847 * allow_barrier when it has finished its IO.
848 * backgroup IO calls must call raise_barrier. Once that returns
849 * there is no normal IO happeing. It must arrange to call
850 * lower_barrier when the particular background IO completes.
852 static void raise_barrier(struct r1conf
*conf
, sector_t sector_nr
)
854 int idx
= sector_to_idx(sector_nr
);
856 spin_lock_irq(&conf
->resync_lock
);
858 /* Wait until no block IO is waiting */
859 wait_event_lock_irq(conf
->wait_barrier
,
860 !atomic_read(&conf
->nr_waiting
[idx
]),
863 /* block any new IO from starting */
864 atomic_inc(&conf
->barrier
[idx
]);
866 * In raise_barrier() we firstly increase conf->barrier[idx] then
867 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
868 * increase conf->nr_pending[idx] then check conf->barrier[idx].
869 * A memory barrier here to make sure conf->nr_pending[idx] won't
870 * be fetched before conf->barrier[idx] is increased. Otherwise
871 * there will be a race between raise_barrier() and _wait_barrier().
873 smp_mb__after_atomic();
875 /* For these conditions we must wait:
876 * A: while the array is in frozen state
877 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
878 * existing in corresponding I/O barrier bucket.
879 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
880 * max resync count which allowed on current I/O barrier bucket.
882 wait_event_lock_irq(conf
->wait_barrier
,
883 !conf
->array_frozen
&&
884 !atomic_read(&conf
->nr_pending
[idx
]) &&
885 atomic_read(&conf
->barrier
[idx
]) < RESYNC_DEPTH
,
888 atomic_inc(&conf
->nr_sync_pending
);
889 spin_unlock_irq(&conf
->resync_lock
);
892 static void lower_barrier(struct r1conf
*conf
, sector_t sector_nr
)
894 int idx
= sector_to_idx(sector_nr
);
896 BUG_ON(atomic_read(&conf
->barrier
[idx
]) <= 0);
898 atomic_dec(&conf
->barrier
[idx
]);
899 atomic_dec(&conf
->nr_sync_pending
);
900 wake_up(&conf
->wait_barrier
);
903 static void _wait_barrier(struct r1conf
*conf
, int idx
)
906 * We need to increase conf->nr_pending[idx] very early here,
907 * then raise_barrier() can be blocked when it waits for
908 * conf->nr_pending[idx] to be 0. Then we can avoid holding
909 * conf->resync_lock when there is no barrier raised in same
910 * barrier unit bucket. Also if the array is frozen, I/O
911 * should be blocked until array is unfrozen.
913 atomic_inc(&conf
->nr_pending
[idx
]);
915 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
916 * check conf->barrier[idx]. In raise_barrier() we firstly increase
917 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
918 * barrier is necessary here to make sure conf->barrier[idx] won't be
919 * fetched before conf->nr_pending[idx] is increased. Otherwise there
920 * will be a race between _wait_barrier() and raise_barrier().
922 smp_mb__after_atomic();
925 * Don't worry about checking two atomic_t variables at same time
926 * here. If during we check conf->barrier[idx], the array is
927 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
928 * 0, it is safe to return and make the I/O continue. Because the
929 * array is frozen, all I/O returned here will eventually complete
930 * or be queued, no race will happen. See code comment in
933 if (!READ_ONCE(conf
->array_frozen
) &&
934 !atomic_read(&conf
->barrier
[idx
]))
938 * After holding conf->resync_lock, conf->nr_pending[idx]
939 * should be decreased before waiting for barrier to drop.
940 * Otherwise, we may encounter a race condition because
941 * raise_barrer() might be waiting for conf->nr_pending[idx]
942 * to be 0 at same time.
944 spin_lock_irq(&conf
->resync_lock
);
945 atomic_inc(&conf
->nr_waiting
[idx
]);
946 atomic_dec(&conf
->nr_pending
[idx
]);
948 * In case freeze_array() is waiting for
949 * get_unqueued_pending() == extra
951 wake_up(&conf
->wait_barrier
);
952 /* Wait for the barrier in same barrier unit bucket to drop. */
953 wait_event_lock_irq(conf
->wait_barrier
,
954 !conf
->array_frozen
&&
955 !atomic_read(&conf
->barrier
[idx
]),
957 atomic_inc(&conf
->nr_pending
[idx
]);
958 atomic_dec(&conf
->nr_waiting
[idx
]);
959 spin_unlock_irq(&conf
->resync_lock
);
962 static void wait_read_barrier(struct r1conf
*conf
, sector_t sector_nr
)
964 int idx
= sector_to_idx(sector_nr
);
967 * Very similar to _wait_barrier(). The difference is, for read
968 * I/O we don't need wait for sync I/O, but if the whole array
969 * is frozen, the read I/O still has to wait until the array is
970 * unfrozen. Since there is no ordering requirement with
971 * conf->barrier[idx] here, memory barrier is unnecessary as well.
973 atomic_inc(&conf
->nr_pending
[idx
]);
975 if (!READ_ONCE(conf
->array_frozen
))
978 spin_lock_irq(&conf
->resync_lock
);
979 atomic_inc(&conf
->nr_waiting
[idx
]);
980 atomic_dec(&conf
->nr_pending
[idx
]);
982 * In case freeze_array() is waiting for
983 * get_unqueued_pending() == extra
985 wake_up(&conf
->wait_barrier
);
986 /* Wait for array to be unfrozen */
987 wait_event_lock_irq(conf
->wait_barrier
,
990 atomic_inc(&conf
->nr_pending
[idx
]);
991 atomic_dec(&conf
->nr_waiting
[idx
]);
992 spin_unlock_irq(&conf
->resync_lock
);
995 static void wait_barrier(struct r1conf
*conf
, sector_t sector_nr
)
997 int idx
= sector_to_idx(sector_nr
);
999 _wait_barrier(conf
, idx
);
1002 static void wait_all_barriers(struct r1conf
*conf
)
1006 for (idx
= 0; idx
< BARRIER_BUCKETS_NR
; idx
++)
1007 _wait_barrier(conf
, idx
);
1010 static void _allow_barrier(struct r1conf
*conf
, int idx
)
1012 atomic_dec(&conf
->nr_pending
[idx
]);
1013 wake_up(&conf
->wait_barrier
);
1016 static void allow_barrier(struct r1conf
*conf
, sector_t sector_nr
)
1018 int idx
= sector_to_idx(sector_nr
);
1020 _allow_barrier(conf
, idx
);
1023 static void allow_all_barriers(struct r1conf
*conf
)
1027 for (idx
= 0; idx
< BARRIER_BUCKETS_NR
; idx
++)
1028 _allow_barrier(conf
, idx
);
1031 /* conf->resync_lock should be held */
1032 static int get_unqueued_pending(struct r1conf
*conf
)
1036 ret
= atomic_read(&conf
->nr_sync_pending
);
1037 for (idx
= 0; idx
< BARRIER_BUCKETS_NR
; idx
++)
1038 ret
+= atomic_read(&conf
->nr_pending
[idx
]) -
1039 atomic_read(&conf
->nr_queued
[idx
]);
1044 static void freeze_array(struct r1conf
*conf
, int extra
)
1046 /* Stop sync I/O and normal I/O and wait for everything to
1048 * This is called in two situations:
1049 * 1) management command handlers (reshape, remove disk, quiesce).
1050 * 2) one normal I/O request failed.
1052 * After array_frozen is set to 1, new sync IO will be blocked at
1053 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1054 * or wait_read_barrier(). The flying I/Os will either complete or be
1055 * queued. When everything goes quite, there are only queued I/Os left.
1057 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1058 * barrier bucket index which this I/O request hits. When all sync and
1059 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1060 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1061 * in handle_read_error(), we may call freeze_array() before trying to
1062 * fix the read error. In this case, the error read I/O is not queued,
1063 * so get_unqueued_pending() == 1.
1065 * Therefore before this function returns, we need to wait until
1066 * get_unqueued_pendings(conf) gets equal to extra. For
1067 * normal I/O context, extra is 1, in rested situations extra is 0.
1069 spin_lock_irq(&conf
->resync_lock
);
1070 conf
->array_frozen
= 1;
1071 raid1_log(conf
->mddev
, "wait freeze");
1072 wait_event_lock_irq_cmd(
1074 get_unqueued_pending(conf
) == extra
,
1076 flush_pending_writes(conf
));
1077 spin_unlock_irq(&conf
->resync_lock
);
1079 static void unfreeze_array(struct r1conf
*conf
)
1081 /* reverse the effect of the freeze */
1082 spin_lock_irq(&conf
->resync_lock
);
1083 conf
->array_frozen
= 0;
1084 spin_unlock_irq(&conf
->resync_lock
);
1085 wake_up(&conf
->wait_barrier
);
1088 static struct bio
*alloc_behind_master_bio(struct r1bio
*r1_bio
,
1091 int size
= bio
->bi_iter
.bi_size
;
1092 unsigned vcnt
= (size
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1094 struct bio
*behind_bio
= NULL
;
1096 behind_bio
= bio_alloc_mddev(GFP_NOIO
, vcnt
, r1_bio
->mddev
);
1100 /* discard op, we don't support writezero/writesame yet */
1101 if (!bio_has_data(bio
))
1104 while (i
< vcnt
&& size
) {
1106 int len
= min_t(int, PAGE_SIZE
, size
);
1108 page
= alloc_page(GFP_NOIO
);
1109 if (unlikely(!page
))
1112 bio_add_page(behind_bio
, page
, len
, 0);
1118 bio_copy_data(behind_bio
, bio
);
1120 r1_bio
->behind_master_bio
= behind_bio
;;
1121 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
1126 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1127 bio
->bi_iter
.bi_size
);
1128 bio_free_pages(behind_bio
);
1133 struct raid1_plug_cb
{
1134 struct blk_plug_cb cb
;
1135 struct bio_list pending
;
1139 static void raid1_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1141 struct raid1_plug_cb
*plug
= container_of(cb
, struct raid1_plug_cb
,
1143 struct mddev
*mddev
= plug
->cb
.data
;
1144 struct r1conf
*conf
= mddev
->private;
1147 if (from_schedule
|| current
->bio_list
) {
1148 spin_lock_irq(&conf
->device_lock
);
1149 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1150 conf
->pending_count
+= plug
->pending_cnt
;
1151 spin_unlock_irq(&conf
->device_lock
);
1152 wake_up(&conf
->wait_barrier
);
1153 md_wakeup_thread(mddev
->thread
);
1158 /* we aren't scheduling, so we can do the write-out directly. */
1159 bio
= bio_list_get(&plug
->pending
);
1160 flush_bio_list(conf
, bio
);
1164 static void init_r1bio(struct r1bio
*r1_bio
, struct mddev
*mddev
, struct bio
*bio
)
1166 r1_bio
->master_bio
= bio
;
1167 r1_bio
->sectors
= bio_sectors(bio
);
1169 r1_bio
->mddev
= mddev
;
1170 r1_bio
->sector
= bio
->bi_iter
.bi_sector
;
1173 static inline struct r1bio
*
1174 alloc_r1bio(struct mddev
*mddev
, struct bio
*bio
)
1176 struct r1conf
*conf
= mddev
->private;
1177 struct r1bio
*r1_bio
;
1179 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1180 /* Ensure no bio records IO_BLOCKED */
1181 memset(r1_bio
->bios
, 0, conf
->raid_disks
* sizeof(r1_bio
->bios
[0]));
1182 init_r1bio(r1_bio
, mddev
, bio
);
1186 static void raid1_read_request(struct mddev
*mddev
, struct bio
*bio
,
1187 int max_read_sectors
, struct r1bio
*r1_bio
)
1189 struct r1conf
*conf
= mddev
->private;
1190 struct raid1_info
*mirror
;
1191 struct bio
*read_bio
;
1192 struct bitmap
*bitmap
= mddev
->bitmap
;
1193 const int op
= bio_op(bio
);
1194 const unsigned long do_sync
= (bio
->bi_opf
& REQ_SYNC
);
1197 bool print_msg
= !!r1_bio
;
1198 char b
[BDEVNAME_SIZE
];
1201 * If r1_bio is set, we are blocking the raid1d thread
1202 * so there is a tiny risk of deadlock. So ask for
1203 * emergency memory if needed.
1205 gfp_t gfp
= r1_bio
? (GFP_NOIO
| __GFP_HIGH
) : GFP_NOIO
;
1208 /* Need to get the block device name carefully */
1209 struct md_rdev
*rdev
;
1211 rdev
= rcu_dereference(conf
->mirrors
[r1_bio
->read_disk
].rdev
);
1213 bdevname(rdev
->bdev
, b
);
1220 * Still need barrier for READ in case that whole
1223 wait_read_barrier(conf
, bio
->bi_iter
.bi_sector
);
1226 r1_bio
= alloc_r1bio(mddev
, bio
);
1228 init_r1bio(r1_bio
, mddev
, bio
);
1229 r1_bio
->sectors
= max_read_sectors
;
1232 * make_request() can abort the operation when read-ahead is being
1233 * used and no empty request is available.
1235 rdisk
= read_balance(conf
, r1_bio
, &max_sectors
);
1238 /* couldn't find anywhere to read from */
1240 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1243 (unsigned long long)r1_bio
->sector
);
1245 raid_end_bio_io(r1_bio
);
1248 mirror
= conf
->mirrors
+ rdisk
;
1251 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1253 (unsigned long long)r1_bio
->sector
,
1254 bdevname(mirror
->rdev
->bdev
, b
));
1256 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
1259 * Reading from a write-mostly device must take care not to
1260 * over-take any writes that are 'behind'
1262 raid1_log(mddev
, "wait behind writes");
1263 wait_event(bitmap
->behind_wait
,
1264 atomic_read(&bitmap
->behind_writes
) == 0);
1267 if (max_sectors
< bio_sectors(bio
)) {
1268 struct bio
*split
= bio_split(bio
, max_sectors
,
1269 gfp
, conf
->bio_split
);
1270 bio_chain(split
, bio
);
1271 generic_make_request(bio
);
1273 r1_bio
->master_bio
= bio
;
1274 r1_bio
->sectors
= max_sectors
;
1277 r1_bio
->read_disk
= rdisk
;
1279 read_bio
= bio_clone_fast(bio
, gfp
, mddev
->bio_set
);
1281 r1_bio
->bios
[rdisk
] = read_bio
;
1283 read_bio
->bi_iter
.bi_sector
= r1_bio
->sector
+
1284 mirror
->rdev
->data_offset
;
1285 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
1286 read_bio
->bi_end_io
= raid1_end_read_request
;
1287 bio_set_op_attrs(read_bio
, op
, do_sync
);
1288 if (test_bit(FailFast
, &mirror
->rdev
->flags
) &&
1289 test_bit(R1BIO_FailFast
, &r1_bio
->state
))
1290 read_bio
->bi_opf
|= MD_FAILFAST
;
1291 read_bio
->bi_private
= r1_bio
;
1294 trace_block_bio_remap(bdev_get_queue(read_bio
->bi_bdev
),
1295 read_bio
, disk_devt(mddev
->gendisk
),
1298 generic_make_request(read_bio
);
1301 static void raid1_write_request(struct mddev
*mddev
, struct bio
*bio
,
1302 int max_write_sectors
)
1304 struct r1conf
*conf
= mddev
->private;
1305 struct r1bio
*r1_bio
;
1307 struct bitmap
*bitmap
= mddev
->bitmap
;
1308 unsigned long flags
;
1309 struct md_rdev
*blocked_rdev
;
1310 struct blk_plug_cb
*cb
;
1311 struct raid1_plug_cb
*plug
= NULL
;
1316 * Register the new request and wait if the reconstruction
1317 * thread has put up a bar for new requests.
1318 * Continue immediately if no resync is active currently.
1321 md_write_start(mddev
, bio
); /* wait on superblock update early */
1323 if ((bio_end_sector(bio
) > mddev
->suspend_lo
&&
1324 bio
->bi_iter
.bi_sector
< mddev
->suspend_hi
) ||
1325 (mddev_is_clustered(mddev
) &&
1326 md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1327 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
)))) {
1330 * As the suspend_* range is controlled by userspace, we want
1331 * an interruptible wait.
1335 flush_signals(current
);
1336 prepare_to_wait(&conf
->wait_barrier
,
1337 &w
, TASK_INTERRUPTIBLE
);
1338 if (bio_end_sector(bio
) <= mddev
->suspend_lo
||
1339 bio
->bi_iter
.bi_sector
>= mddev
->suspend_hi
||
1340 (mddev_is_clustered(mddev
) &&
1341 !md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1342 bio
->bi_iter
.bi_sector
,
1343 bio_end_sector(bio
))))
1347 finish_wait(&conf
->wait_barrier
, &w
);
1349 wait_barrier(conf
, bio
->bi_iter
.bi_sector
);
1351 r1_bio
= alloc_r1bio(mddev
, bio
);
1352 r1_bio
->sectors
= max_write_sectors
;
1354 if (conf
->pending_count
>= max_queued_requests
) {
1355 md_wakeup_thread(mddev
->thread
);
1356 raid1_log(mddev
, "wait queued");
1357 wait_event(conf
->wait_barrier
,
1358 conf
->pending_count
< max_queued_requests
);
1360 /* first select target devices under rcu_lock and
1361 * inc refcount on their rdev. Record them by setting
1363 * If there are known/acknowledged bad blocks on any device on
1364 * which we have seen a write error, we want to avoid writing those
1366 * This potentially requires several writes to write around
1367 * the bad blocks. Each set of writes gets it's own r1bio
1368 * with a set of bios attached.
1371 disks
= conf
->raid_disks
* 2;
1373 blocked_rdev
= NULL
;
1375 max_sectors
= r1_bio
->sectors
;
1376 for (i
= 0; i
< disks
; i
++) {
1377 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1378 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1379 atomic_inc(&rdev
->nr_pending
);
1380 blocked_rdev
= rdev
;
1383 r1_bio
->bios
[i
] = NULL
;
1384 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
1385 if (i
< conf
->raid_disks
)
1386 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1390 atomic_inc(&rdev
->nr_pending
);
1391 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1396 is_bad
= is_badblock(rdev
, r1_bio
->sector
, max_sectors
,
1397 &first_bad
, &bad_sectors
);
1399 /* mustn't write here until the bad block is
1401 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1402 blocked_rdev
= rdev
;
1405 if (is_bad
&& first_bad
<= r1_bio
->sector
) {
1406 /* Cannot write here at all */
1407 bad_sectors
-= (r1_bio
->sector
- first_bad
);
1408 if (bad_sectors
< max_sectors
)
1409 /* mustn't write more than bad_sectors
1410 * to other devices yet
1412 max_sectors
= bad_sectors
;
1413 rdev_dec_pending(rdev
, mddev
);
1414 /* We don't set R1BIO_Degraded as that
1415 * only applies if the disk is
1416 * missing, so it might be re-added,
1417 * and we want to know to recover this
1419 * In this case the device is here,
1420 * and the fact that this chunk is not
1421 * in-sync is recorded in the bad
1427 int good_sectors
= first_bad
- r1_bio
->sector
;
1428 if (good_sectors
< max_sectors
)
1429 max_sectors
= good_sectors
;
1432 r1_bio
->bios
[i
] = bio
;
1436 if (unlikely(blocked_rdev
)) {
1437 /* Wait for this device to become unblocked */
1440 for (j
= 0; j
< i
; j
++)
1441 if (r1_bio
->bios
[j
])
1442 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
1444 allow_barrier(conf
, bio
->bi_iter
.bi_sector
);
1445 raid1_log(mddev
, "wait rdev %d blocked", blocked_rdev
->raid_disk
);
1446 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1447 wait_barrier(conf
, bio
->bi_iter
.bi_sector
);
1451 if (max_sectors
< bio_sectors(bio
)) {
1452 struct bio
*split
= bio_split(bio
, max_sectors
,
1453 GFP_NOIO
, conf
->bio_split
);
1454 bio_chain(split
, bio
);
1455 generic_make_request(bio
);
1457 r1_bio
->master_bio
= bio
;
1458 r1_bio
->sectors
= max_sectors
;
1461 atomic_set(&r1_bio
->remaining
, 1);
1462 atomic_set(&r1_bio
->behind_remaining
, 0);
1466 for (i
= 0; i
< disks
; i
++) {
1467 struct bio
*mbio
= NULL
;
1468 if (!r1_bio
->bios
[i
])
1474 * Not if there are too many, or cannot
1475 * allocate memory, or a reader on WriteMostly
1476 * is waiting for behind writes to flush */
1478 (atomic_read(&bitmap
->behind_writes
)
1479 < mddev
->bitmap_info
.max_write_behind
) &&
1480 !waitqueue_active(&bitmap
->behind_wait
)) {
1481 mbio
= alloc_behind_master_bio(r1_bio
, bio
);
1484 bitmap_startwrite(bitmap
, r1_bio
->sector
,
1486 test_bit(R1BIO_BehindIO
,
1492 if (r1_bio
->behind_master_bio
)
1493 mbio
= bio_clone_fast(r1_bio
->behind_master_bio
,
1497 mbio
= bio_clone_fast(bio
, GFP_NOIO
, mddev
->bio_set
);
1500 if (r1_bio
->behind_master_bio
) {
1501 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
1502 atomic_inc(&r1_bio
->behind_remaining
);
1505 r1_bio
->bios
[i
] = mbio
;
1507 mbio
->bi_iter
.bi_sector
= (r1_bio
->sector
+
1508 conf
->mirrors
[i
].rdev
->data_offset
);
1509 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1510 mbio
->bi_end_io
= raid1_end_write_request
;
1511 mbio
->bi_opf
= bio_op(bio
) | (bio
->bi_opf
& (REQ_SYNC
| REQ_FUA
));
1512 if (test_bit(FailFast
, &conf
->mirrors
[i
].rdev
->flags
) &&
1513 !test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
) &&
1514 conf
->raid_disks
- mddev
->degraded
> 1)
1515 mbio
->bi_opf
|= MD_FAILFAST
;
1516 mbio
->bi_private
= r1_bio
;
1518 atomic_inc(&r1_bio
->remaining
);
1521 trace_block_bio_remap(bdev_get_queue(mbio
->bi_bdev
),
1522 mbio
, disk_devt(mddev
->gendisk
),
1524 /* flush_pending_writes() needs access to the rdev so...*/
1525 mbio
->bi_bdev
= (void*)conf
->mirrors
[i
].rdev
;
1527 cb
= blk_check_plugged(raid1_unplug
, mddev
, sizeof(*plug
));
1529 plug
= container_of(cb
, struct raid1_plug_cb
, cb
);
1532 spin_lock_irqsave(&conf
->device_lock
, flags
);
1534 bio_list_add(&plug
->pending
, mbio
);
1535 plug
->pending_cnt
++;
1537 bio_list_add(&conf
->pending_bio_list
, mbio
);
1538 conf
->pending_count
++;
1540 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1542 md_wakeup_thread(mddev
->thread
);
1545 r1_bio_write_done(r1_bio
);
1547 /* In case raid1d snuck in to freeze_array */
1548 wake_up(&conf
->wait_barrier
);
1551 static void raid1_make_request(struct mddev
*mddev
, struct bio
*bio
)
1555 if (unlikely(bio
->bi_opf
& REQ_PREFLUSH
)) {
1556 md_flush_request(mddev
, bio
);
1561 * There is a limit to the maximum size, but
1562 * the read/write handler might find a lower limit
1563 * due to bad blocks. To avoid multiple splits,
1564 * we pass the maximum number of sectors down
1565 * and let the lower level perform the split.
1567 sectors
= align_to_barrier_unit_end(
1568 bio
->bi_iter
.bi_sector
, bio_sectors(bio
));
1570 if (bio_data_dir(bio
) == READ
)
1571 raid1_read_request(mddev
, bio
, sectors
, NULL
);
1573 raid1_write_request(mddev
, bio
, sectors
);
1576 static void raid1_status(struct seq_file
*seq
, struct mddev
*mddev
)
1578 struct r1conf
*conf
= mddev
->private;
1581 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1582 conf
->raid_disks
- mddev
->degraded
);
1584 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1585 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1586 seq_printf(seq
, "%s",
1587 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1590 seq_printf(seq
, "]");
1593 static void raid1_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1595 char b
[BDEVNAME_SIZE
];
1596 struct r1conf
*conf
= mddev
->private;
1597 unsigned long flags
;
1600 * If it is not operational, then we have already marked it as dead
1601 * else if it is the last working disks, ignore the error, let the
1602 * next level up know.
1603 * else mark the drive as failed
1605 spin_lock_irqsave(&conf
->device_lock
, flags
);
1606 if (test_bit(In_sync
, &rdev
->flags
)
1607 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1609 * Don't fail the drive, act as though we were just a
1610 * normal single drive.
1611 * However don't try a recovery from this drive as
1612 * it is very likely to fail.
1614 conf
->recovery_disabled
= mddev
->recovery_disabled
;
1615 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1618 set_bit(Blocked
, &rdev
->flags
);
1619 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1621 set_bit(Faulty
, &rdev
->flags
);
1623 set_bit(Faulty
, &rdev
->flags
);
1624 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1626 * if recovery is running, make sure it aborts.
1628 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1629 set_mask_bits(&mddev
->sb_flags
, 0,
1630 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
1631 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1632 "md/raid1:%s: Operation continuing on %d devices.\n",
1633 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1634 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1637 static void print_conf(struct r1conf
*conf
)
1641 pr_debug("RAID1 conf printout:\n");
1643 pr_debug("(!conf)\n");
1646 pr_debug(" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1650 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1651 char b
[BDEVNAME_SIZE
];
1652 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1654 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1655 i
, !test_bit(In_sync
, &rdev
->flags
),
1656 !test_bit(Faulty
, &rdev
->flags
),
1657 bdevname(rdev
->bdev
,b
));
1662 static void close_sync(struct r1conf
*conf
)
1664 wait_all_barriers(conf
);
1665 allow_all_barriers(conf
);
1667 mempool_destroy(conf
->r1buf_pool
);
1668 conf
->r1buf_pool
= NULL
;
1671 static int raid1_spare_active(struct mddev
*mddev
)
1674 struct r1conf
*conf
= mddev
->private;
1676 unsigned long flags
;
1679 * Find all failed disks within the RAID1 configuration
1680 * and mark them readable.
1681 * Called under mddev lock, so rcu protection not needed.
1682 * device_lock used to avoid races with raid1_end_read_request
1683 * which expects 'In_sync' flags and ->degraded to be consistent.
1685 spin_lock_irqsave(&conf
->device_lock
, flags
);
1686 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1687 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1688 struct md_rdev
*repl
= conf
->mirrors
[conf
->raid_disks
+ i
].rdev
;
1690 && !test_bit(Candidate
, &repl
->flags
)
1691 && repl
->recovery_offset
== MaxSector
1692 && !test_bit(Faulty
, &repl
->flags
)
1693 && !test_and_set_bit(In_sync
, &repl
->flags
)) {
1694 /* replacement has just become active */
1696 !test_and_clear_bit(In_sync
, &rdev
->flags
))
1699 /* Replaced device not technically
1700 * faulty, but we need to be sure
1701 * it gets removed and never re-added
1703 set_bit(Faulty
, &rdev
->flags
);
1704 sysfs_notify_dirent_safe(
1709 && rdev
->recovery_offset
== MaxSector
1710 && !test_bit(Faulty
, &rdev
->flags
)
1711 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1713 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
1716 mddev
->degraded
-= count
;
1717 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1723 static int raid1_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1725 struct r1conf
*conf
= mddev
->private;
1728 struct raid1_info
*p
;
1730 int last
= conf
->raid_disks
- 1;
1732 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
1735 if (md_integrity_add_rdev(rdev
, mddev
))
1738 if (rdev
->raid_disk
>= 0)
1739 first
= last
= rdev
->raid_disk
;
1742 * find the disk ... but prefer rdev->saved_raid_disk
1745 if (rdev
->saved_raid_disk
>= 0 &&
1746 rdev
->saved_raid_disk
>= first
&&
1747 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1748 first
= last
= rdev
->saved_raid_disk
;
1750 for (mirror
= first
; mirror
<= last
; mirror
++) {
1751 p
= conf
->mirrors
+mirror
;
1755 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1756 rdev
->data_offset
<< 9);
1758 p
->head_position
= 0;
1759 rdev
->raid_disk
= mirror
;
1761 /* As all devices are equivalent, we don't need a full recovery
1762 * if this was recently any drive of the array
1764 if (rdev
->saved_raid_disk
< 0)
1766 rcu_assign_pointer(p
->rdev
, rdev
);
1769 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
1770 p
[conf
->raid_disks
].rdev
== NULL
) {
1771 /* Add this device as a replacement */
1772 clear_bit(In_sync
, &rdev
->flags
);
1773 set_bit(Replacement
, &rdev
->flags
);
1774 rdev
->raid_disk
= mirror
;
1777 rcu_assign_pointer(p
[conf
->raid_disks
].rdev
, rdev
);
1781 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1782 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1787 static int raid1_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1789 struct r1conf
*conf
= mddev
->private;
1791 int number
= rdev
->raid_disk
;
1792 struct raid1_info
*p
= conf
->mirrors
+ number
;
1794 if (rdev
!= p
->rdev
)
1795 p
= conf
->mirrors
+ conf
->raid_disks
+ number
;
1798 if (rdev
== p
->rdev
) {
1799 if (test_bit(In_sync
, &rdev
->flags
) ||
1800 atomic_read(&rdev
->nr_pending
)) {
1804 /* Only remove non-faulty devices if recovery
1807 if (!test_bit(Faulty
, &rdev
->flags
) &&
1808 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
1809 mddev
->degraded
< conf
->raid_disks
) {
1814 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
1816 if (atomic_read(&rdev
->nr_pending
)) {
1817 /* lost the race, try later */
1823 if (conf
->mirrors
[conf
->raid_disks
+ number
].rdev
) {
1824 /* We just removed a device that is being replaced.
1825 * Move down the replacement. We drain all IO before
1826 * doing this to avoid confusion.
1828 struct md_rdev
*repl
=
1829 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
;
1830 freeze_array(conf
, 0);
1831 clear_bit(Replacement
, &repl
->flags
);
1833 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
= NULL
;
1834 unfreeze_array(conf
);
1837 clear_bit(WantReplacement
, &rdev
->flags
);
1838 err
= md_integrity_register(mddev
);
1846 static void end_sync_read(struct bio
*bio
)
1848 struct r1bio
*r1_bio
= get_resync_r1bio(bio
);
1850 update_head_pos(r1_bio
->read_disk
, r1_bio
);
1853 * we have read a block, now it needs to be re-written,
1854 * or re-read if the read failed.
1855 * We don't do much here, just schedule handling by raid1d
1858 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1860 if (atomic_dec_and_test(&r1_bio
->remaining
))
1861 reschedule_retry(r1_bio
);
1864 static void end_sync_write(struct bio
*bio
)
1866 int uptodate
= !bio
->bi_error
;
1867 struct r1bio
*r1_bio
= get_resync_r1bio(bio
);
1868 struct mddev
*mddev
= r1_bio
->mddev
;
1869 struct r1conf
*conf
= mddev
->private;
1872 struct md_rdev
*rdev
= conf
->mirrors
[find_bio_disk(r1_bio
, bio
)].rdev
;
1875 sector_t sync_blocks
= 0;
1876 sector_t s
= r1_bio
->sector
;
1877 long sectors_to_go
= r1_bio
->sectors
;
1878 /* make sure these bits doesn't get cleared. */
1880 bitmap_end_sync(mddev
->bitmap
, s
,
1883 sectors_to_go
-= sync_blocks
;
1884 } while (sectors_to_go
> 0);
1885 set_bit(WriteErrorSeen
, &rdev
->flags
);
1886 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1887 set_bit(MD_RECOVERY_NEEDED
, &
1889 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
1890 } else if (is_badblock(rdev
, r1_bio
->sector
, r1_bio
->sectors
,
1891 &first_bad
, &bad_sectors
) &&
1892 !is_badblock(conf
->mirrors
[r1_bio
->read_disk
].rdev
,
1895 &first_bad
, &bad_sectors
)
1897 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
1899 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1900 int s
= r1_bio
->sectors
;
1901 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
1902 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1903 reschedule_retry(r1_bio
);
1906 md_done_sync(mddev
, s
, uptodate
);
1911 static int r1_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1912 int sectors
, struct page
*page
, int rw
)
1914 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, 0, false))
1918 set_bit(WriteErrorSeen
, &rdev
->flags
);
1919 if (!test_and_set_bit(WantReplacement
,
1921 set_bit(MD_RECOVERY_NEEDED
, &
1922 rdev
->mddev
->recovery
);
1924 /* need to record an error - either for the block or the device */
1925 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1926 md_error(rdev
->mddev
, rdev
);
1930 static int fix_sync_read_error(struct r1bio
*r1_bio
)
1932 /* Try some synchronous reads of other devices to get
1933 * good data, much like with normal read errors. Only
1934 * read into the pages we already have so we don't
1935 * need to re-issue the read request.
1936 * We don't need to freeze the array, because being in an
1937 * active sync request, there is no normal IO, and
1938 * no overlapping syncs.
1939 * We don't need to check is_badblock() again as we
1940 * made sure that anything with a bad block in range
1941 * will have bi_end_io clear.
1943 struct mddev
*mddev
= r1_bio
->mddev
;
1944 struct r1conf
*conf
= mddev
->private;
1945 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1946 struct page
**pages
= get_resync_pages(bio
)->pages
;
1947 sector_t sect
= r1_bio
->sector
;
1948 int sectors
= r1_bio
->sectors
;
1950 struct md_rdev
*rdev
;
1952 rdev
= conf
->mirrors
[r1_bio
->read_disk
].rdev
;
1953 if (test_bit(FailFast
, &rdev
->flags
)) {
1954 /* Don't try recovering from here - just fail it
1955 * ... unless it is the last working device of course */
1956 md_error(mddev
, rdev
);
1957 if (test_bit(Faulty
, &rdev
->flags
))
1958 /* Don't try to read from here, but make sure
1959 * put_buf does it's thing
1961 bio
->bi_end_io
= end_sync_write
;
1966 int d
= r1_bio
->read_disk
;
1970 if (s
> (PAGE_SIZE
>>9))
1973 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1974 /* No rcu protection needed here devices
1975 * can only be removed when no resync is
1976 * active, and resync is currently active
1978 rdev
= conf
->mirrors
[d
].rdev
;
1979 if (sync_page_io(rdev
, sect
, s
<<9,
1981 REQ_OP_READ
, 0, false)) {
1987 if (d
== conf
->raid_disks
* 2)
1989 } while (!success
&& d
!= r1_bio
->read_disk
);
1992 char b
[BDEVNAME_SIZE
];
1994 /* Cannot read from anywhere, this block is lost.
1995 * Record a bad block on each device. If that doesn't
1996 * work just disable and interrupt the recovery.
1997 * Don't fail devices as that won't really help.
1999 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2001 bdevname(bio
->bi_bdev
, b
),
2002 (unsigned long long)r1_bio
->sector
);
2003 for (d
= 0; d
< conf
->raid_disks
* 2; d
++) {
2004 rdev
= conf
->mirrors
[d
].rdev
;
2005 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
2007 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
2011 conf
->recovery_disabled
=
2012 mddev
->recovery_disabled
;
2013 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2014 md_done_sync(mddev
, r1_bio
->sectors
, 0);
2026 /* write it back and re-read */
2027 while (d
!= r1_bio
->read_disk
) {
2029 d
= conf
->raid_disks
* 2;
2031 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
2033 rdev
= conf
->mirrors
[d
].rdev
;
2034 if (r1_sync_page_io(rdev
, sect
, s
,
2037 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
2038 rdev_dec_pending(rdev
, mddev
);
2042 while (d
!= r1_bio
->read_disk
) {
2044 d
= conf
->raid_disks
* 2;
2046 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
2048 rdev
= conf
->mirrors
[d
].rdev
;
2049 if (r1_sync_page_io(rdev
, sect
, s
,
2052 atomic_add(s
, &rdev
->corrected_errors
);
2058 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
2063 static void process_checks(struct r1bio
*r1_bio
)
2065 /* We have read all readable devices. If we haven't
2066 * got the block, then there is no hope left.
2067 * If we have, then we want to do a comparison
2068 * and skip the write if everything is the same.
2069 * If any blocks failed to read, then we need to
2070 * attempt an over-write
2072 struct mddev
*mddev
= r1_bio
->mddev
;
2073 struct r1conf
*conf
= mddev
->private;
2078 /* Fix variable parts of all bios */
2079 vcnt
= (r1_bio
->sectors
+ PAGE_SIZE
/ 512 - 1) >> (PAGE_SHIFT
- 9);
2080 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2085 struct bio
*b
= r1_bio
->bios
[i
];
2086 struct resync_pages
*rp
= get_resync_pages(b
);
2087 if (b
->bi_end_io
!= end_sync_read
)
2089 /* fixup the bio for reuse, but preserve errno */
2090 error
= b
->bi_error
;
2092 b
->bi_error
= error
;
2094 b
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
2095 b
->bi_iter
.bi_sector
= r1_bio
->sector
+
2096 conf
->mirrors
[i
].rdev
->data_offset
;
2097 b
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
2098 b
->bi_end_io
= end_sync_read
;
2099 rp
->raid_bio
= r1_bio
;
2102 size
= b
->bi_iter
.bi_size
;
2103 bio_for_each_segment_all(bi
, b
, j
) {
2105 if (size
> PAGE_SIZE
)
2106 bi
->bv_len
= PAGE_SIZE
;
2112 for (primary
= 0; primary
< conf
->raid_disks
* 2; primary
++)
2113 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
2114 !r1_bio
->bios
[primary
]->bi_error
) {
2115 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
2116 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
2119 r1_bio
->read_disk
= primary
;
2120 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2122 struct bio
*pbio
= r1_bio
->bios
[primary
];
2123 struct bio
*sbio
= r1_bio
->bios
[i
];
2124 int error
= sbio
->bi_error
;
2125 struct page
**ppages
= get_resync_pages(pbio
)->pages
;
2126 struct page
**spages
= get_resync_pages(sbio
)->pages
;
2128 int page_len
[RESYNC_PAGES
] = { 0 };
2130 if (sbio
->bi_end_io
!= end_sync_read
)
2132 /* Now we can 'fixup' the error value */
2135 bio_for_each_segment_all(bi
, sbio
, j
)
2136 page_len
[j
] = bi
->bv_len
;
2139 for (j
= vcnt
; j
-- ; ) {
2140 if (memcmp(page_address(ppages
[j
]),
2141 page_address(spages
[j
]),
2148 atomic64_add(r1_bio
->sectors
, &mddev
->resync_mismatches
);
2149 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
2151 /* No need to write to this device. */
2152 sbio
->bi_end_io
= NULL
;
2153 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
2157 bio_copy_data(sbio
, pbio
);
2161 static void sync_request_write(struct mddev
*mddev
, struct r1bio
*r1_bio
)
2163 struct r1conf
*conf
= mddev
->private;
2165 int disks
= conf
->raid_disks
* 2;
2166 struct bio
*bio
, *wbio
;
2168 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2170 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
2171 /* ouch - failed to read all of that. */
2172 if (!fix_sync_read_error(r1_bio
))
2175 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2176 process_checks(r1_bio
);
2181 atomic_set(&r1_bio
->remaining
, 1);
2182 for (i
= 0; i
< disks
; i
++) {
2183 wbio
= r1_bio
->bios
[i
];
2184 if (wbio
->bi_end_io
== NULL
||
2185 (wbio
->bi_end_io
== end_sync_read
&&
2186 (i
== r1_bio
->read_disk
||
2187 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
2189 if (test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2192 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2193 if (test_bit(FailFast
, &conf
->mirrors
[i
].rdev
->flags
))
2194 wbio
->bi_opf
|= MD_FAILFAST
;
2196 wbio
->bi_end_io
= end_sync_write
;
2197 atomic_inc(&r1_bio
->remaining
);
2198 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, bio_sectors(wbio
));
2200 generic_make_request(wbio
);
2203 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
2204 /* if we're here, all write(s) have completed, so clean up */
2205 int s
= r1_bio
->sectors
;
2206 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2207 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2208 reschedule_retry(r1_bio
);
2211 md_done_sync(mddev
, s
, 1);
2217 * This is a kernel thread which:
2219 * 1. Retries failed read operations on working mirrors.
2220 * 2. Updates the raid superblock when problems encounter.
2221 * 3. Performs writes following reads for array synchronising.
2224 static void fix_read_error(struct r1conf
*conf
, int read_disk
,
2225 sector_t sect
, int sectors
)
2227 struct mddev
*mddev
= conf
->mddev
;
2233 struct md_rdev
*rdev
;
2235 if (s
> (PAGE_SIZE
>>9))
2243 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2245 (test_bit(In_sync
, &rdev
->flags
) ||
2246 (!test_bit(Faulty
, &rdev
->flags
) &&
2247 rdev
->recovery_offset
>= sect
+ s
)) &&
2248 is_badblock(rdev
, sect
, s
,
2249 &first_bad
, &bad_sectors
) == 0) {
2250 atomic_inc(&rdev
->nr_pending
);
2252 if (sync_page_io(rdev
, sect
, s
<<9,
2253 conf
->tmppage
, REQ_OP_READ
, 0, false))
2255 rdev_dec_pending(rdev
, mddev
);
2261 if (d
== conf
->raid_disks
* 2)
2263 } while (!success
&& d
!= read_disk
);
2266 /* Cannot read from anywhere - mark it bad */
2267 struct md_rdev
*rdev
= conf
->mirrors
[read_disk
].rdev
;
2268 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
2269 md_error(mddev
, rdev
);
2272 /* write it back and re-read */
2274 while (d
!= read_disk
) {
2276 d
= conf
->raid_disks
* 2;
2279 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2281 !test_bit(Faulty
, &rdev
->flags
)) {
2282 atomic_inc(&rdev
->nr_pending
);
2284 r1_sync_page_io(rdev
, sect
, s
,
2285 conf
->tmppage
, WRITE
);
2286 rdev_dec_pending(rdev
, mddev
);
2291 while (d
!= read_disk
) {
2292 char b
[BDEVNAME_SIZE
];
2294 d
= conf
->raid_disks
* 2;
2297 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2299 !test_bit(Faulty
, &rdev
->flags
)) {
2300 atomic_inc(&rdev
->nr_pending
);
2302 if (r1_sync_page_io(rdev
, sect
, s
,
2303 conf
->tmppage
, READ
)) {
2304 atomic_add(s
, &rdev
->corrected_errors
);
2305 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2307 (unsigned long long)(sect
+
2309 bdevname(rdev
->bdev
, b
));
2311 rdev_dec_pending(rdev
, mddev
);
2320 static int narrow_write_error(struct r1bio
*r1_bio
, int i
)
2322 struct mddev
*mddev
= r1_bio
->mddev
;
2323 struct r1conf
*conf
= mddev
->private;
2324 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2326 /* bio has the data to be written to device 'i' where
2327 * we just recently had a write error.
2328 * We repeatedly clone the bio and trim down to one block,
2329 * then try the write. Where the write fails we record
2331 * It is conceivable that the bio doesn't exactly align with
2332 * blocks. We must handle this somehow.
2334 * We currently own a reference on the rdev.
2340 int sect_to_write
= r1_bio
->sectors
;
2343 if (rdev
->badblocks
.shift
< 0)
2346 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2347 bdev_logical_block_size(rdev
->bdev
) >> 9);
2348 sector
= r1_bio
->sector
;
2349 sectors
= ((sector
+ block_sectors
)
2350 & ~(sector_t
)(block_sectors
- 1))
2353 while (sect_to_write
) {
2355 if (sectors
> sect_to_write
)
2356 sectors
= sect_to_write
;
2357 /* Write at 'sector' for 'sectors'*/
2359 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
2360 wbio
= bio_clone_fast(r1_bio
->behind_master_bio
,
2363 /* We really need a _all clone */
2364 wbio
->bi_iter
= (struct bvec_iter
){ 0 };
2366 wbio
= bio_clone_fast(r1_bio
->master_bio
, GFP_NOIO
,
2370 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2371 wbio
->bi_iter
.bi_sector
= r1_bio
->sector
;
2372 wbio
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
2374 bio_trim(wbio
, sector
- r1_bio
->sector
, sectors
);
2375 wbio
->bi_iter
.bi_sector
+= rdev
->data_offset
;
2376 wbio
->bi_bdev
= rdev
->bdev
;
2378 if (submit_bio_wait(wbio
) < 0)
2380 ok
= rdev_set_badblocks(rdev
, sector
,
2385 sect_to_write
-= sectors
;
2387 sectors
= block_sectors
;
2392 static void handle_sync_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2395 int s
= r1_bio
->sectors
;
2396 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++) {
2397 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2398 struct bio
*bio
= r1_bio
->bios
[m
];
2399 if (bio
->bi_end_io
== NULL
)
2401 if (!bio
->bi_error
&&
2402 test_bit(R1BIO_MadeGood
, &r1_bio
->state
)) {
2403 rdev_clear_badblocks(rdev
, r1_bio
->sector
, s
, 0);
2405 if (bio
->bi_error
&&
2406 test_bit(R1BIO_WriteError
, &r1_bio
->state
)) {
2407 if (!rdev_set_badblocks(rdev
, r1_bio
->sector
, s
, 0))
2408 md_error(conf
->mddev
, rdev
);
2412 md_done_sync(conf
->mddev
, s
, 1);
2415 static void handle_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2420 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++)
2421 if (r1_bio
->bios
[m
] == IO_MADE_GOOD
) {
2422 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2423 rdev_clear_badblocks(rdev
,
2425 r1_bio
->sectors
, 0);
2426 rdev_dec_pending(rdev
, conf
->mddev
);
2427 } else if (r1_bio
->bios
[m
] != NULL
) {
2428 /* This drive got a write error. We need to
2429 * narrow down and record precise write
2433 if (!narrow_write_error(r1_bio
, m
)) {
2434 md_error(conf
->mddev
,
2435 conf
->mirrors
[m
].rdev
);
2436 /* an I/O failed, we can't clear the bitmap */
2437 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2439 rdev_dec_pending(conf
->mirrors
[m
].rdev
,
2443 spin_lock_irq(&conf
->device_lock
);
2444 list_add(&r1_bio
->retry_list
, &conf
->bio_end_io_list
);
2445 idx
= sector_to_idx(r1_bio
->sector
);
2446 atomic_inc(&conf
->nr_queued
[idx
]);
2447 spin_unlock_irq(&conf
->device_lock
);
2449 * In case freeze_array() is waiting for condition
2450 * get_unqueued_pending() == extra to be true.
2452 wake_up(&conf
->wait_barrier
);
2453 md_wakeup_thread(conf
->mddev
->thread
);
2455 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2456 close_write(r1_bio
);
2457 raid_end_bio_io(r1_bio
);
2461 static void handle_read_error(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2463 struct mddev
*mddev
= conf
->mddev
;
2465 struct md_rdev
*rdev
;
2467 sector_t bio_sector
;
2469 clear_bit(R1BIO_ReadError
, &r1_bio
->state
);
2470 /* we got a read error. Maybe the drive is bad. Maybe just
2471 * the block and we can fix it.
2472 * We freeze all other IO, and try reading the block from
2473 * other devices. When we find one, we re-write
2474 * and check it that fixes the read error.
2475 * This is all done synchronously while the array is
2479 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2480 bio_dev
= bio
->bi_bdev
->bd_dev
;
2481 bio_sector
= conf
->mirrors
[r1_bio
->read_disk
].rdev
->data_offset
+ r1_bio
->sector
;
2483 r1_bio
->bios
[r1_bio
->read_disk
] = NULL
;
2485 rdev
= conf
->mirrors
[r1_bio
->read_disk
].rdev
;
2487 && !test_bit(FailFast
, &rdev
->flags
)) {
2488 freeze_array(conf
, 1);
2489 fix_read_error(conf
, r1_bio
->read_disk
,
2490 r1_bio
->sector
, r1_bio
->sectors
);
2491 unfreeze_array(conf
);
2493 r1_bio
->bios
[r1_bio
->read_disk
] = IO_BLOCKED
;
2496 rdev_dec_pending(rdev
, conf
->mddev
);
2497 allow_barrier(conf
, r1_bio
->sector
);
2498 bio
= r1_bio
->master_bio
;
2500 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2502 raid1_read_request(mddev
, bio
, r1_bio
->sectors
, r1_bio
);
2505 static void raid1d(struct md_thread
*thread
)
2507 struct mddev
*mddev
= thread
->mddev
;
2508 struct r1bio
*r1_bio
;
2509 unsigned long flags
;
2510 struct r1conf
*conf
= mddev
->private;
2511 struct list_head
*head
= &conf
->retry_list
;
2512 struct blk_plug plug
;
2515 md_check_recovery(mddev
);
2517 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2518 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
2520 spin_lock_irqsave(&conf
->device_lock
, flags
);
2521 if (!test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
))
2522 list_splice_init(&conf
->bio_end_io_list
, &tmp
);
2523 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2524 while (!list_empty(&tmp
)) {
2525 r1_bio
= list_first_entry(&tmp
, struct r1bio
,
2527 list_del(&r1_bio
->retry_list
);
2528 idx
= sector_to_idx(r1_bio
->sector
);
2529 atomic_dec(&conf
->nr_queued
[idx
]);
2530 if (mddev
->degraded
)
2531 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2532 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2533 close_write(r1_bio
);
2534 raid_end_bio_io(r1_bio
);
2538 blk_start_plug(&plug
);
2541 flush_pending_writes(conf
);
2543 spin_lock_irqsave(&conf
->device_lock
, flags
);
2544 if (list_empty(head
)) {
2545 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2548 r1_bio
= list_entry(head
->prev
, struct r1bio
, retry_list
);
2549 list_del(head
->prev
);
2550 idx
= sector_to_idx(r1_bio
->sector
);
2551 atomic_dec(&conf
->nr_queued
[idx
]);
2552 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2554 mddev
= r1_bio
->mddev
;
2555 conf
= mddev
->private;
2556 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
2557 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2558 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2559 handle_sync_write_finished(conf
, r1_bio
);
2561 sync_request_write(mddev
, r1_bio
);
2562 } else if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2563 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2564 handle_write_finished(conf
, r1_bio
);
2565 else if (test_bit(R1BIO_ReadError
, &r1_bio
->state
))
2566 handle_read_error(conf
, r1_bio
);
2571 if (mddev
->sb_flags
& ~(1<<MD_SB_CHANGE_PENDING
))
2572 md_check_recovery(mddev
);
2574 blk_finish_plug(&plug
);
2577 static int init_resync(struct r1conf
*conf
)
2581 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2582 BUG_ON(conf
->r1buf_pool
);
2583 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
2585 if (!conf
->r1buf_pool
)
2591 * perform a "sync" on one "block"
2593 * We need to make sure that no normal I/O request - particularly write
2594 * requests - conflict with active sync requests.
2596 * This is achieved by tracking pending requests and a 'barrier' concept
2597 * that can be installed to exclude normal IO requests.
2600 static sector_t
raid1_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2603 struct r1conf
*conf
= mddev
->private;
2604 struct r1bio
*r1_bio
;
2606 sector_t max_sector
, nr_sectors
;
2610 int write_targets
= 0, read_targets
= 0;
2611 sector_t sync_blocks
;
2612 int still_degraded
= 0;
2613 int good_sectors
= RESYNC_SECTORS
;
2614 int min_bad
= 0; /* number of sectors that are bad in all devices */
2615 int idx
= sector_to_idx(sector_nr
);
2617 if (!conf
->r1buf_pool
)
2618 if (init_resync(conf
))
2621 max_sector
= mddev
->dev_sectors
;
2622 if (sector_nr
>= max_sector
) {
2623 /* If we aborted, we need to abort the
2624 * sync on the 'current' bitmap chunk (there will
2625 * only be one in raid1 resync.
2626 * We can find the current addess in mddev->curr_resync
2628 if (mddev
->curr_resync
< max_sector
) /* aborted */
2629 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2631 else /* completed sync */
2634 bitmap_close_sync(mddev
->bitmap
);
2637 if (mddev_is_clustered(mddev
)) {
2638 conf
->cluster_sync_low
= 0;
2639 conf
->cluster_sync_high
= 0;
2644 if (mddev
->bitmap
== NULL
&&
2645 mddev
->recovery_cp
== MaxSector
&&
2646 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2647 conf
->fullsync
== 0) {
2649 return max_sector
- sector_nr
;
2651 /* before building a request, check if we can skip these blocks..
2652 * This call the bitmap_start_sync doesn't actually record anything
2654 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
2655 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2656 /* We can skip this block, and probably several more */
2662 * If there is non-resync activity waiting for a turn, then let it
2663 * though before starting on this new sync request.
2665 if (atomic_read(&conf
->nr_waiting
[idx
]))
2666 schedule_timeout_uninterruptible(1);
2668 /* we are incrementing sector_nr below. To be safe, we check against
2669 * sector_nr + two times RESYNC_SECTORS
2672 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
,
2673 mddev_is_clustered(mddev
) && (sector_nr
+ 2 * RESYNC_SECTORS
> conf
->cluster_sync_high
));
2674 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
2676 raise_barrier(conf
, sector_nr
);
2680 * If we get a correctably read error during resync or recovery,
2681 * we might want to read from a different device. So we
2682 * flag all drives that could conceivably be read from for READ,
2683 * and any others (which will be non-In_sync devices) for WRITE.
2684 * If a read fails, we try reading from something else for which READ
2688 r1_bio
->mddev
= mddev
;
2689 r1_bio
->sector
= sector_nr
;
2691 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
2692 /* make sure good_sectors won't go across barrier unit boundary */
2693 good_sectors
= align_to_barrier_unit_end(sector_nr
, good_sectors
);
2695 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2696 struct md_rdev
*rdev
;
2697 bio
= r1_bio
->bios
[i
];
2699 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
2701 test_bit(Faulty
, &rdev
->flags
)) {
2702 if (i
< conf
->raid_disks
)
2704 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
2705 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
2706 bio
->bi_end_io
= end_sync_write
;
2709 /* may need to read from here */
2710 sector_t first_bad
= MaxSector
;
2713 if (is_badblock(rdev
, sector_nr
, good_sectors
,
2714 &first_bad
, &bad_sectors
)) {
2715 if (first_bad
> sector_nr
)
2716 good_sectors
= first_bad
- sector_nr
;
2718 bad_sectors
-= (sector_nr
- first_bad
);
2720 min_bad
> bad_sectors
)
2721 min_bad
= bad_sectors
;
2724 if (sector_nr
< first_bad
) {
2725 if (test_bit(WriteMostly
, &rdev
->flags
)) {
2732 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
2733 bio
->bi_end_io
= end_sync_read
;
2735 } else if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
2736 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2737 !test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)) {
2739 * The device is suitable for reading (InSync),
2740 * but has bad block(s) here. Let's try to correct them,
2741 * if we are doing resync or repair. Otherwise, leave
2742 * this device alone for this sync request.
2744 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
2745 bio
->bi_end_io
= end_sync_write
;
2749 if (bio
->bi_end_io
) {
2750 atomic_inc(&rdev
->nr_pending
);
2751 bio
->bi_iter
.bi_sector
= sector_nr
+ rdev
->data_offset
;
2752 bio
->bi_bdev
= rdev
->bdev
;
2753 if (test_bit(FailFast
, &rdev
->flags
))
2754 bio
->bi_opf
|= MD_FAILFAST
;
2760 r1_bio
->read_disk
= disk
;
2762 if (read_targets
== 0 && min_bad
> 0) {
2763 /* These sectors are bad on all InSync devices, so we
2764 * need to mark them bad on all write targets
2767 for (i
= 0 ; i
< conf
->raid_disks
* 2 ; i
++)
2768 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_write
) {
2769 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2770 ok
= rdev_set_badblocks(rdev
, sector_nr
,
2774 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
2779 /* Cannot record the badblocks, so need to
2781 * If there are multiple read targets, could just
2782 * fail the really bad ones ???
2784 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2785 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2791 if (min_bad
> 0 && min_bad
< good_sectors
) {
2792 /* only resync enough to reach the next bad->good
2794 good_sectors
= min_bad
;
2797 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
2798 /* extra read targets are also write targets */
2799 write_targets
+= read_targets
-1;
2801 if (write_targets
== 0 || read_targets
== 0) {
2802 /* There is nowhere to write, so all non-sync
2803 * drives must be failed - so we are finished
2807 max_sector
= sector_nr
+ min_bad
;
2808 rv
= max_sector
- sector_nr
;
2814 if (max_sector
> mddev
->resync_max
)
2815 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2816 if (max_sector
> sector_nr
+ good_sectors
)
2817 max_sector
= sector_nr
+ good_sectors
;
2822 int len
= PAGE_SIZE
;
2823 if (sector_nr
+ (len
>>9) > max_sector
)
2824 len
= (max_sector
- sector_nr
) << 9;
2827 if (sync_blocks
== 0) {
2828 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2829 &sync_blocks
, still_degraded
) &&
2831 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2833 if ((len
>> 9) > sync_blocks
)
2834 len
= sync_blocks
<<9;
2837 for (i
= 0 ; i
< conf
->raid_disks
* 2; i
++) {
2838 struct resync_pages
*rp
;
2840 bio
= r1_bio
->bios
[i
];
2841 rp
= get_resync_pages(bio
);
2842 if (bio
->bi_end_io
) {
2843 page
= resync_fetch_page(rp
, rp
->idx
++);
2846 * won't fail because the vec table is big
2847 * enough to hold all these pages
2849 bio_add_page(bio
, page
, len
, 0);
2852 nr_sectors
+= len
>>9;
2853 sector_nr
+= len
>>9;
2854 sync_blocks
-= (len
>>9);
2855 } while (get_resync_pages(r1_bio
->bios
[disk
]->bi_private
)->idx
< RESYNC_PAGES
);
2857 r1_bio
->sectors
= nr_sectors
;
2859 if (mddev_is_clustered(mddev
) &&
2860 conf
->cluster_sync_high
< sector_nr
+ nr_sectors
) {
2861 conf
->cluster_sync_low
= mddev
->curr_resync_completed
;
2862 conf
->cluster_sync_high
= conf
->cluster_sync_low
+ CLUSTER_RESYNC_WINDOW_SECTORS
;
2863 /* Send resync message */
2864 md_cluster_ops
->resync_info_update(mddev
,
2865 conf
->cluster_sync_low
,
2866 conf
->cluster_sync_high
);
2869 /* For a user-requested sync, we read all readable devices and do a
2872 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2873 atomic_set(&r1_bio
->remaining
, read_targets
);
2874 for (i
= 0; i
< conf
->raid_disks
* 2 && read_targets
; i
++) {
2875 bio
= r1_bio
->bios
[i
];
2876 if (bio
->bi_end_io
== end_sync_read
) {
2878 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2879 if (read_targets
== 1)
2880 bio
->bi_opf
&= ~MD_FAILFAST
;
2881 generic_make_request(bio
);
2885 atomic_set(&r1_bio
->remaining
, 1);
2886 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2887 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2888 if (read_targets
== 1)
2889 bio
->bi_opf
&= ~MD_FAILFAST
;
2890 generic_make_request(bio
);
2896 static sector_t
raid1_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2901 return mddev
->dev_sectors
;
2904 static struct r1conf
*setup_conf(struct mddev
*mddev
)
2906 struct r1conf
*conf
;
2908 struct raid1_info
*disk
;
2909 struct md_rdev
*rdev
;
2912 conf
= kzalloc(sizeof(struct r1conf
), GFP_KERNEL
);
2916 conf
->nr_pending
= kcalloc(BARRIER_BUCKETS_NR
,
2917 sizeof(atomic_t
), GFP_KERNEL
);
2918 if (!conf
->nr_pending
)
2921 conf
->nr_waiting
= kcalloc(BARRIER_BUCKETS_NR
,
2922 sizeof(atomic_t
), GFP_KERNEL
);
2923 if (!conf
->nr_waiting
)
2926 conf
->nr_queued
= kcalloc(BARRIER_BUCKETS_NR
,
2927 sizeof(atomic_t
), GFP_KERNEL
);
2928 if (!conf
->nr_queued
)
2931 conf
->barrier
= kcalloc(BARRIER_BUCKETS_NR
,
2932 sizeof(atomic_t
), GFP_KERNEL
);
2936 conf
->mirrors
= kzalloc(sizeof(struct raid1_info
)
2937 * mddev
->raid_disks
* 2,
2942 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2946 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
2947 if (!conf
->poolinfo
)
2949 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
* 2;
2950 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2953 if (!conf
->r1bio_pool
)
2956 conf
->bio_split
= bioset_create(BIO_POOL_SIZE
, 0);
2957 if (!conf
->bio_split
)
2960 conf
->poolinfo
->mddev
= mddev
;
2963 spin_lock_init(&conf
->device_lock
);
2964 rdev_for_each(rdev
, mddev
) {
2965 int disk_idx
= rdev
->raid_disk
;
2966 if (disk_idx
>= mddev
->raid_disks
2969 if (test_bit(Replacement
, &rdev
->flags
))
2970 disk
= conf
->mirrors
+ mddev
->raid_disks
+ disk_idx
;
2972 disk
= conf
->mirrors
+ disk_idx
;
2977 disk
->head_position
= 0;
2978 disk
->seq_start
= MaxSector
;
2980 conf
->raid_disks
= mddev
->raid_disks
;
2981 conf
->mddev
= mddev
;
2982 INIT_LIST_HEAD(&conf
->retry_list
);
2983 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
2985 spin_lock_init(&conf
->resync_lock
);
2986 init_waitqueue_head(&conf
->wait_barrier
);
2988 bio_list_init(&conf
->pending_bio_list
);
2989 conf
->pending_count
= 0;
2990 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
2993 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2995 disk
= conf
->mirrors
+ i
;
2997 if (i
< conf
->raid_disks
&&
2998 disk
[conf
->raid_disks
].rdev
) {
2999 /* This slot has a replacement. */
3001 /* No original, just make the replacement
3002 * a recovering spare
3005 disk
[conf
->raid_disks
].rdev
;
3006 disk
[conf
->raid_disks
].rdev
= NULL
;
3007 } else if (!test_bit(In_sync
, &disk
->rdev
->flags
))
3008 /* Original is not in_sync - bad */
3013 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3014 disk
->head_position
= 0;
3016 (disk
->rdev
->saved_raid_disk
< 0))
3022 conf
->thread
= md_register_thread(raid1d
, mddev
, "raid1");
3030 mempool_destroy(conf
->r1bio_pool
);
3031 kfree(conf
->mirrors
);
3032 safe_put_page(conf
->tmppage
);
3033 kfree(conf
->poolinfo
);
3034 kfree(conf
->nr_pending
);
3035 kfree(conf
->nr_waiting
);
3036 kfree(conf
->nr_queued
);
3037 kfree(conf
->barrier
);
3038 if (conf
->bio_split
)
3039 bioset_free(conf
->bio_split
);
3042 return ERR_PTR(err
);
3045 static void raid1_free(struct mddev
*mddev
, void *priv
);
3046 static int raid1_run(struct mddev
*mddev
)
3048 struct r1conf
*conf
;
3050 struct md_rdev
*rdev
;
3052 bool discard_supported
= false;
3054 if (mddev
->level
!= 1) {
3055 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3056 mdname(mddev
), mddev
->level
);
3059 if (mddev
->reshape_position
!= MaxSector
) {
3060 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3065 * copy the already verified devices into our private RAID1
3066 * bookkeeping area. [whatever we allocate in run(),
3067 * should be freed in raid1_free()]
3069 if (mddev
->private == NULL
)
3070 conf
= setup_conf(mddev
);
3072 conf
= mddev
->private;
3075 return PTR_ERR(conf
);
3078 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3079 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
3082 rdev_for_each(rdev
, mddev
) {
3083 if (!mddev
->gendisk
)
3085 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3086 rdev
->data_offset
<< 9);
3087 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3088 discard_supported
= true;
3091 mddev
->degraded
= 0;
3092 for (i
=0; i
< conf
->raid_disks
; i
++)
3093 if (conf
->mirrors
[i
].rdev
== NULL
||
3094 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
3095 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
3098 if (conf
->raid_disks
- mddev
->degraded
== 1)
3099 mddev
->recovery_cp
= MaxSector
;
3101 if (mddev
->recovery_cp
!= MaxSector
)
3102 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3104 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3105 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
3109 * Ok, everything is just fine now
3111 mddev
->thread
= conf
->thread
;
3112 conf
->thread
= NULL
;
3113 mddev
->private = conf
;
3114 set_bit(MD_FAILFAST_SUPPORTED
, &mddev
->flags
);
3116 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
3119 if (discard_supported
)
3120 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3123 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3127 ret
= md_integrity_register(mddev
);
3129 md_unregister_thread(&mddev
->thread
);
3130 raid1_free(mddev
, conf
);
3135 static void raid1_free(struct mddev
*mddev
, void *priv
)
3137 struct r1conf
*conf
= priv
;
3139 mempool_destroy(conf
->r1bio_pool
);
3140 kfree(conf
->mirrors
);
3141 safe_put_page(conf
->tmppage
);
3142 kfree(conf
->poolinfo
);
3143 kfree(conf
->nr_pending
);
3144 kfree(conf
->nr_waiting
);
3145 kfree(conf
->nr_queued
);
3146 kfree(conf
->barrier
);
3147 if (conf
->bio_split
)
3148 bioset_free(conf
->bio_split
);
3152 static int raid1_resize(struct mddev
*mddev
, sector_t sectors
)
3154 /* no resync is happening, and there is enough space
3155 * on all devices, so we can resize.
3156 * We need to make sure resync covers any new space.
3157 * If the array is shrinking we should possibly wait until
3158 * any io in the removed space completes, but it hardly seems
3161 sector_t newsize
= raid1_size(mddev
, sectors
, 0);
3162 if (mddev
->external_size
&&
3163 mddev
->array_sectors
> newsize
)
3165 if (mddev
->bitmap
) {
3166 int ret
= bitmap_resize(mddev
->bitmap
, newsize
, 0, 0);
3170 md_set_array_sectors(mddev
, newsize
);
3171 if (sectors
> mddev
->dev_sectors
&&
3172 mddev
->recovery_cp
> mddev
->dev_sectors
) {
3173 mddev
->recovery_cp
= mddev
->dev_sectors
;
3174 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3176 mddev
->dev_sectors
= sectors
;
3177 mddev
->resync_max_sectors
= sectors
;
3181 static int raid1_reshape(struct mddev
*mddev
)
3184 * 1/ resize the r1bio_pool
3185 * 2/ resize conf->mirrors
3187 * We allocate a new r1bio_pool if we can.
3188 * Then raise a device barrier and wait until all IO stops.
3189 * Then resize conf->mirrors and swap in the new r1bio pool.
3191 * At the same time, we "pack" the devices so that all the missing
3192 * devices have the higher raid_disk numbers.
3194 mempool_t
*newpool
, *oldpool
;
3195 struct pool_info
*newpoolinfo
;
3196 struct raid1_info
*newmirrors
;
3197 struct r1conf
*conf
= mddev
->private;
3198 int cnt
, raid_disks
;
3199 unsigned long flags
;
3202 /* Cannot change chunk_size, layout, or level */
3203 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
3204 mddev
->layout
!= mddev
->new_layout
||
3205 mddev
->level
!= mddev
->new_level
) {
3206 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3207 mddev
->new_layout
= mddev
->layout
;
3208 mddev
->new_level
= mddev
->level
;
3212 if (!mddev_is_clustered(mddev
)) {
3213 err
= md_allow_write(mddev
);
3218 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3220 if (raid_disks
< conf
->raid_disks
) {
3222 for (d
= 0; d
< conf
->raid_disks
; d
++)
3223 if (conf
->mirrors
[d
].rdev
)
3225 if (cnt
> raid_disks
)
3229 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
3232 newpoolinfo
->mddev
= mddev
;
3233 newpoolinfo
->raid_disks
= raid_disks
* 2;
3235 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
3236 r1bio_pool_free
, newpoolinfo
);
3241 newmirrors
= kzalloc(sizeof(struct raid1_info
) * raid_disks
* 2,
3245 mempool_destroy(newpool
);
3249 freeze_array(conf
, 0);
3251 /* ok, everything is stopped */
3252 oldpool
= conf
->r1bio_pool
;
3253 conf
->r1bio_pool
= newpool
;
3255 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
3256 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
3257 if (rdev
&& rdev
->raid_disk
!= d2
) {
3258 sysfs_unlink_rdev(mddev
, rdev
);
3259 rdev
->raid_disk
= d2
;
3260 sysfs_unlink_rdev(mddev
, rdev
);
3261 if (sysfs_link_rdev(mddev
, rdev
))
3262 pr_warn("md/raid1:%s: cannot register rd%d\n",
3263 mdname(mddev
), rdev
->raid_disk
);
3266 newmirrors
[d2
++].rdev
= rdev
;
3268 kfree(conf
->mirrors
);
3269 conf
->mirrors
= newmirrors
;
3270 kfree(conf
->poolinfo
);
3271 conf
->poolinfo
= newpoolinfo
;
3273 spin_lock_irqsave(&conf
->device_lock
, flags
);
3274 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
3275 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3276 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
3277 mddev
->delta_disks
= 0;
3279 unfreeze_array(conf
);
3281 set_bit(MD_RECOVERY_RECOVER
, &mddev
->recovery
);
3282 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3283 md_wakeup_thread(mddev
->thread
);
3285 mempool_destroy(oldpool
);
3289 static void raid1_quiesce(struct mddev
*mddev
, int state
)
3291 struct r1conf
*conf
= mddev
->private;
3294 case 2: /* wake for suspend */
3295 wake_up(&conf
->wait_barrier
);
3298 freeze_array(conf
, 0);
3301 unfreeze_array(conf
);
3306 static void *raid1_takeover(struct mddev
*mddev
)
3308 /* raid1 can take over:
3309 * raid5 with 2 devices, any layout or chunk size
3311 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
3312 struct r1conf
*conf
;
3313 mddev
->new_level
= 1;
3314 mddev
->new_layout
= 0;
3315 mddev
->new_chunk_sectors
= 0;
3316 conf
= setup_conf(mddev
);
3317 if (!IS_ERR(conf
)) {
3318 /* Array must appear to be quiesced */
3319 conf
->array_frozen
= 1;
3320 mddev_clear_unsupported_flags(mddev
,
3321 UNSUPPORTED_MDDEV_FLAGS
);
3325 return ERR_PTR(-EINVAL
);
3328 static struct md_personality raid1_personality
=
3332 .owner
= THIS_MODULE
,
3333 .make_request
= raid1_make_request
,
3336 .status
= raid1_status
,
3337 .error_handler
= raid1_error
,
3338 .hot_add_disk
= raid1_add_disk
,
3339 .hot_remove_disk
= raid1_remove_disk
,
3340 .spare_active
= raid1_spare_active
,
3341 .sync_request
= raid1_sync_request
,
3342 .resize
= raid1_resize
,
3344 .check_reshape
= raid1_reshape
,
3345 .quiesce
= raid1_quiesce
,
3346 .takeover
= raid1_takeover
,
3347 .congested
= raid1_congested
,
3350 static int __init
raid_init(void)
3352 return register_md_personality(&raid1_personality
);
3355 static void raid_exit(void)
3357 unregister_md_personality(&raid1_personality
);
3360 module_init(raid_init
);
3361 module_exit(raid_exit
);
3362 MODULE_LICENSE("GPL");
3363 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3364 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3365 MODULE_ALIAS("md-raid1");
3366 MODULE_ALIAS("md-level-1");
3368 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);