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 /* free extra copy of the data pages */
392 int i
= r1_bio
->behind_page_count
;
394 safe_put_page(r1_bio
->behind_bvecs
[i
].bv_page
);
395 kfree(r1_bio
->behind_bvecs
);
396 r1_bio
->behind_bvecs
= NULL
;
398 /* clear the bitmap if all writes complete successfully */
399 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
401 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
402 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
403 md_write_end(r1_bio
->mddev
);
406 static void r1_bio_write_done(struct r1bio
*r1_bio
)
408 if (!atomic_dec_and_test(&r1_bio
->remaining
))
411 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
412 reschedule_retry(r1_bio
);
415 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
))
416 reschedule_retry(r1_bio
);
418 raid_end_bio_io(r1_bio
);
422 static void raid1_end_write_request(struct bio
*bio
)
424 struct r1bio
*r1_bio
= bio
->bi_private
;
425 int behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
426 struct r1conf
*conf
= r1_bio
->mddev
->private;
427 struct bio
*to_put
= NULL
;
428 int mirror
= find_bio_disk(r1_bio
, bio
);
429 struct md_rdev
*rdev
= conf
->mirrors
[mirror
].rdev
;
432 discard_error
= bio
->bi_error
&& bio_op(bio
) == REQ_OP_DISCARD
;
435 * 'one mirror IO has finished' event handler:
437 if (bio
->bi_error
&& !discard_error
) {
438 set_bit(WriteErrorSeen
, &rdev
->flags
);
439 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
440 set_bit(MD_RECOVERY_NEEDED
, &
441 conf
->mddev
->recovery
);
443 if (test_bit(FailFast
, &rdev
->flags
) &&
444 (bio
->bi_opf
& MD_FAILFAST
) &&
445 /* We never try FailFast to WriteMostly devices */
446 !test_bit(WriteMostly
, &rdev
->flags
)) {
447 md_error(r1_bio
->mddev
, rdev
);
448 if (!test_bit(Faulty
, &rdev
->flags
))
449 /* This is the only remaining device,
450 * We need to retry the write without
453 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
455 /* Finished with this branch */
456 r1_bio
->bios
[mirror
] = NULL
;
460 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
463 * Set R1BIO_Uptodate in our master bio, so that we
464 * will return a good error code for to the higher
465 * levels even if IO on some other mirrored buffer
468 * The 'master' represents the composite IO operation
469 * to user-side. So if something waits for IO, then it
470 * will wait for the 'master' bio.
475 r1_bio
->bios
[mirror
] = NULL
;
478 * Do not set R1BIO_Uptodate if the current device is
479 * rebuilding or Faulty. This is because we cannot use
480 * such device for properly reading the data back (we could
481 * potentially use it, if the current write would have felt
482 * before rdev->recovery_offset, but for simplicity we don't
485 if (test_bit(In_sync
, &rdev
->flags
) &&
486 !test_bit(Faulty
, &rdev
->flags
))
487 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
489 /* Maybe we can clear some bad blocks. */
490 if (is_badblock(rdev
, r1_bio
->sector
, r1_bio
->sectors
,
491 &first_bad
, &bad_sectors
) && !discard_error
) {
492 r1_bio
->bios
[mirror
] = IO_MADE_GOOD
;
493 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
498 if (test_bit(WriteMostly
, &rdev
->flags
))
499 atomic_dec(&r1_bio
->behind_remaining
);
502 * In behind mode, we ACK the master bio once the I/O
503 * has safely reached all non-writemostly
504 * disks. Setting the Returned bit ensures that this
505 * gets done only once -- we don't ever want to return
506 * -EIO here, instead we'll wait
508 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
509 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
510 /* Maybe we can return now */
511 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
512 struct bio
*mbio
= r1_bio
->master_bio
;
513 pr_debug("raid1: behind end write sectors"
515 (unsigned long long) mbio
->bi_iter
.bi_sector
,
516 (unsigned long long) bio_end_sector(mbio
) - 1);
517 call_bio_endio(r1_bio
);
521 if (r1_bio
->bios
[mirror
] == NULL
)
522 rdev_dec_pending(rdev
, conf
->mddev
);
525 * Let's see if all mirrored write operations have finished
528 r1_bio_write_done(r1_bio
);
534 static sector_t
align_to_barrier_unit_end(sector_t start_sector
,
539 WARN_ON(sectors
== 0);
541 * len is the number of sectors from start_sector to end of the
542 * barrier unit which start_sector belongs to.
544 len
= round_up(start_sector
+ 1, BARRIER_UNIT_SECTOR_SIZE
) -
554 * This routine returns the disk from which the requested read should
555 * be done. There is a per-array 'next expected sequential IO' sector
556 * number - if this matches on the next IO then we use the last disk.
557 * There is also a per-disk 'last know head position' sector that is
558 * maintained from IRQ contexts, both the normal and the resync IO
559 * completion handlers update this position correctly. If there is no
560 * perfect sequential match then we pick the disk whose head is closest.
562 * If there are 2 mirrors in the same 2 devices, performance degrades
563 * because position is mirror, not device based.
565 * The rdev for the device selected will have nr_pending incremented.
567 static int read_balance(struct r1conf
*conf
, struct r1bio
*r1_bio
, int *max_sectors
)
569 const sector_t this_sector
= r1_bio
->sector
;
571 int best_good_sectors
;
572 int best_disk
, best_dist_disk
, best_pending_disk
;
576 unsigned int min_pending
;
577 struct md_rdev
*rdev
;
579 int choose_next_idle
;
583 * Check if we can balance. We can balance on the whole
584 * device if no resync is going on, or below the resync window.
585 * We take the first readable disk when above the resync window.
588 sectors
= r1_bio
->sectors
;
591 best_dist
= MaxSector
;
592 best_pending_disk
= -1;
593 min_pending
= UINT_MAX
;
594 best_good_sectors
= 0;
596 choose_next_idle
= 0;
597 clear_bit(R1BIO_FailFast
, &r1_bio
->state
);
599 if ((conf
->mddev
->recovery_cp
< this_sector
+ sectors
) ||
600 (mddev_is_clustered(conf
->mddev
) &&
601 md_cluster_ops
->area_resyncing(conf
->mddev
, READ
, this_sector
,
602 this_sector
+ sectors
)))
607 for (disk
= 0 ; disk
< conf
->raid_disks
* 2 ; disk
++) {
611 unsigned int pending
;
614 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
615 if (r1_bio
->bios
[disk
] == IO_BLOCKED
617 || test_bit(Faulty
, &rdev
->flags
))
619 if (!test_bit(In_sync
, &rdev
->flags
) &&
620 rdev
->recovery_offset
< this_sector
+ sectors
)
622 if (test_bit(WriteMostly
, &rdev
->flags
)) {
623 /* Don't balance among write-mostly, just
624 * use the first as a last resort */
625 if (best_dist_disk
< 0) {
626 if (is_badblock(rdev
, this_sector
, sectors
,
627 &first_bad
, &bad_sectors
)) {
628 if (first_bad
<= this_sector
)
629 /* Cannot use this */
631 best_good_sectors
= first_bad
- this_sector
;
633 best_good_sectors
= sectors
;
634 best_dist_disk
= disk
;
635 best_pending_disk
= disk
;
639 /* This is a reasonable device to use. It might
642 if (is_badblock(rdev
, this_sector
, sectors
,
643 &first_bad
, &bad_sectors
)) {
644 if (best_dist
< MaxSector
)
645 /* already have a better device */
647 if (first_bad
<= this_sector
) {
648 /* cannot read here. If this is the 'primary'
649 * device, then we must not read beyond
650 * bad_sectors from another device..
652 bad_sectors
-= (this_sector
- first_bad
);
653 if (choose_first
&& sectors
> bad_sectors
)
654 sectors
= bad_sectors
;
655 if (best_good_sectors
> sectors
)
656 best_good_sectors
= sectors
;
659 sector_t good_sectors
= first_bad
- this_sector
;
660 if (good_sectors
> best_good_sectors
) {
661 best_good_sectors
= good_sectors
;
669 best_good_sectors
= sectors
;
672 /* At least two disks to choose from so failfast is OK */
673 set_bit(R1BIO_FailFast
, &r1_bio
->state
);
675 nonrot
= blk_queue_nonrot(bdev_get_queue(rdev
->bdev
));
676 has_nonrot_disk
|= nonrot
;
677 pending
= atomic_read(&rdev
->nr_pending
);
678 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
683 /* Don't change to another disk for sequential reads */
684 if (conf
->mirrors
[disk
].next_seq_sect
== this_sector
686 int opt_iosize
= bdev_io_opt(rdev
->bdev
) >> 9;
687 struct raid1_info
*mirror
= &conf
->mirrors
[disk
];
691 * If buffered sequential IO size exceeds optimal
692 * iosize, check if there is idle disk. If yes, choose
693 * the idle disk. read_balance could already choose an
694 * idle disk before noticing it's a sequential IO in
695 * this disk. This doesn't matter because this disk
696 * will idle, next time it will be utilized after the
697 * first disk has IO size exceeds optimal iosize. In
698 * this way, iosize of the first disk will be optimal
699 * iosize at least. iosize of the second disk might be
700 * small, but not a big deal since when the second disk
701 * starts IO, the first disk is likely still busy.
703 if (nonrot
&& opt_iosize
> 0 &&
704 mirror
->seq_start
!= MaxSector
&&
705 mirror
->next_seq_sect
> opt_iosize
&&
706 mirror
->next_seq_sect
- opt_iosize
>=
708 choose_next_idle
= 1;
714 if (choose_next_idle
)
717 if (min_pending
> pending
) {
718 min_pending
= pending
;
719 best_pending_disk
= disk
;
722 if (dist
< best_dist
) {
724 best_dist_disk
= disk
;
729 * If all disks are rotational, choose the closest disk. If any disk is
730 * non-rotational, choose the disk with less pending request even the
731 * disk is rotational, which might/might not be optimal for raids with
732 * mixed ratation/non-rotational disks depending on workload.
734 if (best_disk
== -1) {
735 if (has_nonrot_disk
|| min_pending
== 0)
736 best_disk
= best_pending_disk
;
738 best_disk
= best_dist_disk
;
741 if (best_disk
>= 0) {
742 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
745 atomic_inc(&rdev
->nr_pending
);
746 sectors
= best_good_sectors
;
748 if (conf
->mirrors
[best_disk
].next_seq_sect
!= this_sector
)
749 conf
->mirrors
[best_disk
].seq_start
= this_sector
;
751 conf
->mirrors
[best_disk
].next_seq_sect
= this_sector
+ sectors
;
754 *max_sectors
= sectors
;
759 static int raid1_congested(struct mddev
*mddev
, int bits
)
761 struct r1conf
*conf
= mddev
->private;
764 if ((bits
& (1 << WB_async_congested
)) &&
765 conf
->pending_count
>= max_queued_requests
)
769 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
770 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
771 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
772 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
776 /* Note the '|| 1' - when read_balance prefers
777 * non-congested targets, it can be removed
779 if ((bits
& (1 << WB_async_congested
)) || 1)
780 ret
|= bdi_congested(q
->backing_dev_info
, bits
);
782 ret
&= bdi_congested(q
->backing_dev_info
, bits
);
789 static void flush_pending_writes(struct r1conf
*conf
)
791 /* Any writes that have been queued but are awaiting
792 * bitmap updates get flushed here.
794 spin_lock_irq(&conf
->device_lock
);
796 if (conf
->pending_bio_list
.head
) {
798 bio
= bio_list_get(&conf
->pending_bio_list
);
799 conf
->pending_count
= 0;
800 spin_unlock_irq(&conf
->device_lock
);
801 /* flush any pending bitmap writes to
802 * disk before proceeding w/ I/O */
803 bitmap_unplug(conf
->mddev
->bitmap
);
804 wake_up(&conf
->wait_barrier
);
806 while (bio
) { /* submit pending writes */
807 struct bio
*next
= bio
->bi_next
;
808 struct md_rdev
*rdev
= (void*)bio
->bi_bdev
;
810 bio
->bi_bdev
= rdev
->bdev
;
811 if (test_bit(Faulty
, &rdev
->flags
)) {
812 bio
->bi_error
= -EIO
;
814 } else if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
815 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
819 generic_make_request(bio
);
823 spin_unlock_irq(&conf
->device_lock
);
827 * Sometimes we need to suspend IO while we do something else,
828 * either some resync/recovery, or reconfigure the array.
829 * To do this we raise a 'barrier'.
830 * The 'barrier' is a counter that can be raised multiple times
831 * to count how many activities are happening which preclude
833 * We can only raise the barrier if there is no pending IO.
834 * i.e. if nr_pending == 0.
835 * We choose only to raise the barrier if no-one is waiting for the
836 * barrier to go down. This means that as soon as an IO request
837 * is ready, no other operations which require a barrier will start
838 * until the IO request has had a chance.
840 * So: regular IO calls 'wait_barrier'. When that returns there
841 * is no backgroup IO happening, It must arrange to call
842 * allow_barrier when it has finished its IO.
843 * backgroup IO calls must call raise_barrier. Once that returns
844 * there is no normal IO happeing. It must arrange to call
845 * lower_barrier when the particular background IO completes.
847 static void raise_barrier(struct r1conf
*conf
, sector_t sector_nr
)
849 int idx
= sector_to_idx(sector_nr
);
851 spin_lock_irq(&conf
->resync_lock
);
853 /* Wait until no block IO is waiting */
854 wait_event_lock_irq(conf
->wait_barrier
,
855 !atomic_read(&conf
->nr_waiting
[idx
]),
858 /* block any new IO from starting */
859 atomic_inc(&conf
->barrier
[idx
]);
861 * In raise_barrier() we firstly increase conf->barrier[idx] then
862 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
863 * increase conf->nr_pending[idx] then check conf->barrier[idx].
864 * A memory barrier here to make sure conf->nr_pending[idx] won't
865 * be fetched before conf->barrier[idx] is increased. Otherwise
866 * there will be a race between raise_barrier() and _wait_barrier().
868 smp_mb__after_atomic();
870 /* For these conditions we must wait:
871 * A: while the array is in frozen state
872 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
873 * existing in corresponding I/O barrier bucket.
874 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
875 * max resync count which allowed on current I/O barrier bucket.
877 wait_event_lock_irq(conf
->wait_barrier
,
878 !conf
->array_frozen
&&
879 !atomic_read(&conf
->nr_pending
[idx
]) &&
880 atomic_read(&conf
->barrier
[idx
]) < RESYNC_DEPTH
,
883 atomic_inc(&conf
->nr_pending
[idx
]);
884 spin_unlock_irq(&conf
->resync_lock
);
887 static void lower_barrier(struct r1conf
*conf
, sector_t sector_nr
)
889 int idx
= sector_to_idx(sector_nr
);
891 BUG_ON(atomic_read(&conf
->barrier
[idx
]) <= 0);
893 atomic_dec(&conf
->barrier
[idx
]);
894 atomic_dec(&conf
->nr_pending
[idx
]);
895 wake_up(&conf
->wait_barrier
);
898 static void _wait_barrier(struct r1conf
*conf
, int idx
)
901 * We need to increase conf->nr_pending[idx] very early here,
902 * then raise_barrier() can be blocked when it waits for
903 * conf->nr_pending[idx] to be 0. Then we can avoid holding
904 * conf->resync_lock when there is no barrier raised in same
905 * barrier unit bucket. Also if the array is frozen, I/O
906 * should be blocked until array is unfrozen.
908 atomic_inc(&conf
->nr_pending
[idx
]);
910 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
911 * check conf->barrier[idx]. In raise_barrier() we firstly increase
912 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
913 * barrier is necessary here to make sure conf->barrier[idx] won't be
914 * fetched before conf->nr_pending[idx] is increased. Otherwise there
915 * will be a race between _wait_barrier() and raise_barrier().
917 smp_mb__after_atomic();
920 * Don't worry about checking two atomic_t variables at same time
921 * here. If during we check conf->barrier[idx], the array is
922 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
923 * 0, it is safe to return and make the I/O continue. Because the
924 * array is frozen, all I/O returned here will eventually complete
925 * or be queued, no race will happen. See code comment in
928 if (!READ_ONCE(conf
->array_frozen
) &&
929 !atomic_read(&conf
->barrier
[idx
]))
933 * After holding conf->resync_lock, conf->nr_pending[idx]
934 * should be decreased before waiting for barrier to drop.
935 * Otherwise, we may encounter a race condition because
936 * raise_barrer() might be waiting for conf->nr_pending[idx]
937 * to be 0 at same time.
939 spin_lock_irq(&conf
->resync_lock
);
940 atomic_inc(&conf
->nr_waiting
[idx
]);
941 atomic_dec(&conf
->nr_pending
[idx
]);
943 * In case freeze_array() is waiting for
944 * get_unqueued_pending() == extra
946 wake_up(&conf
->wait_barrier
);
947 /* Wait for the barrier in same barrier unit bucket to drop. */
948 wait_event_lock_irq(conf
->wait_barrier
,
949 !conf
->array_frozen
&&
950 !atomic_read(&conf
->barrier
[idx
]),
952 atomic_inc(&conf
->nr_pending
[idx
]);
953 atomic_dec(&conf
->nr_waiting
[idx
]);
954 spin_unlock_irq(&conf
->resync_lock
);
957 static void wait_read_barrier(struct r1conf
*conf
, sector_t sector_nr
)
959 int idx
= sector_to_idx(sector_nr
);
962 * Very similar to _wait_barrier(). The difference is, for read
963 * I/O we don't need wait for sync I/O, but if the whole array
964 * is frozen, the read I/O still has to wait until the array is
965 * unfrozen. Since there is no ordering requirement with
966 * conf->barrier[idx] here, memory barrier is unnecessary as well.
968 atomic_inc(&conf
->nr_pending
[idx
]);
970 if (!READ_ONCE(conf
->array_frozen
))
973 spin_lock_irq(&conf
->resync_lock
);
974 atomic_inc(&conf
->nr_waiting
[idx
]);
975 atomic_dec(&conf
->nr_pending
[idx
]);
977 * In case freeze_array() is waiting for
978 * get_unqueued_pending() == extra
980 wake_up(&conf
->wait_barrier
);
981 /* Wait for array to be unfrozen */
982 wait_event_lock_irq(conf
->wait_barrier
,
985 atomic_inc(&conf
->nr_pending
[idx
]);
986 atomic_dec(&conf
->nr_waiting
[idx
]);
987 spin_unlock_irq(&conf
->resync_lock
);
990 static void inc_pending(struct r1conf
*conf
, sector_t bi_sector
)
992 /* The current request requires multiple r1_bio, so
993 * we need to increment the pending count, and the corresponding
996 int idx
= sector_to_idx(bi_sector
);
997 atomic_inc(&conf
->nr_pending
[idx
]);
1000 static void wait_barrier(struct r1conf
*conf
, sector_t sector_nr
)
1002 int idx
= sector_to_idx(sector_nr
);
1004 _wait_barrier(conf
, idx
);
1007 static void wait_all_barriers(struct r1conf
*conf
)
1011 for (idx
= 0; idx
< BARRIER_BUCKETS_NR
; idx
++)
1012 _wait_barrier(conf
, idx
);
1015 static void _allow_barrier(struct r1conf
*conf
, int idx
)
1017 atomic_dec(&conf
->nr_pending
[idx
]);
1018 wake_up(&conf
->wait_barrier
);
1021 static void allow_barrier(struct r1conf
*conf
, sector_t sector_nr
)
1023 int idx
= sector_to_idx(sector_nr
);
1025 _allow_barrier(conf
, idx
);
1028 static void allow_all_barriers(struct r1conf
*conf
)
1032 for (idx
= 0; idx
< BARRIER_BUCKETS_NR
; idx
++)
1033 _allow_barrier(conf
, idx
);
1036 /* conf->resync_lock should be held */
1037 static int get_unqueued_pending(struct r1conf
*conf
)
1041 for (ret
= 0, idx
= 0; idx
< BARRIER_BUCKETS_NR
; idx
++)
1042 ret
+= atomic_read(&conf
->nr_pending
[idx
]) -
1043 atomic_read(&conf
->nr_queued
[idx
]);
1048 static void freeze_array(struct r1conf
*conf
, int extra
)
1050 /* Stop sync I/O and normal I/O and wait for everything to
1052 * This is called in two situations:
1053 * 1) management command handlers (reshape, remove disk, quiesce).
1054 * 2) one normal I/O request failed.
1056 * After array_frozen is set to 1, new sync IO will be blocked at
1057 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1058 * or wait_read_barrier(). The flying I/Os will either complete or be
1059 * queued. When everything goes quite, there are only queued I/Os left.
1061 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1062 * barrier bucket index which this I/O request hits. When all sync and
1063 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1064 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1065 * in handle_read_error(), we may call freeze_array() before trying to
1066 * fix the read error. In this case, the error read I/O is not queued,
1067 * so get_unqueued_pending() == 1.
1069 * Therefore before this function returns, we need to wait until
1070 * get_unqueued_pendings(conf) gets equal to extra. For
1071 * normal I/O context, extra is 1, in rested situations extra is 0.
1073 spin_lock_irq(&conf
->resync_lock
);
1074 conf
->array_frozen
= 1;
1075 raid1_log(conf
->mddev
, "wait freeze");
1076 wait_event_lock_irq_cmd(
1078 get_unqueued_pending(conf
) == extra
,
1080 flush_pending_writes(conf
));
1081 spin_unlock_irq(&conf
->resync_lock
);
1083 static void unfreeze_array(struct r1conf
*conf
)
1085 /* reverse the effect of the freeze */
1086 spin_lock_irq(&conf
->resync_lock
);
1087 conf
->array_frozen
= 0;
1088 spin_unlock_irq(&conf
->resync_lock
);
1089 wake_up(&conf
->wait_barrier
);
1092 /* duplicate the data pages for behind I/O
1094 static void alloc_behind_pages(struct bio
*bio
, struct r1bio
*r1_bio
)
1097 struct bio_vec
*bvec
;
1098 struct bio_vec
*bvecs
= kzalloc(bio
->bi_vcnt
* sizeof(struct bio_vec
),
1100 if (unlikely(!bvecs
))
1103 bio_for_each_segment_all(bvec
, bio
, i
) {
1105 bvecs
[i
].bv_page
= alloc_page(GFP_NOIO
);
1106 if (unlikely(!bvecs
[i
].bv_page
))
1108 memcpy(kmap(bvecs
[i
].bv_page
) + bvec
->bv_offset
,
1109 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
1110 kunmap(bvecs
[i
].bv_page
);
1111 kunmap(bvec
->bv_page
);
1113 r1_bio
->behind_bvecs
= bvecs
;
1114 r1_bio
->behind_page_count
= bio
->bi_vcnt
;
1115 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
1119 for (i
= 0; i
< bio
->bi_vcnt
; i
++)
1120 if (bvecs
[i
].bv_page
)
1121 put_page(bvecs
[i
].bv_page
);
1123 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1124 bio
->bi_iter
.bi_size
);
1127 struct raid1_plug_cb
{
1128 struct blk_plug_cb cb
;
1129 struct bio_list pending
;
1133 static void raid1_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1135 struct raid1_plug_cb
*plug
= container_of(cb
, struct raid1_plug_cb
,
1137 struct mddev
*mddev
= plug
->cb
.data
;
1138 struct r1conf
*conf
= mddev
->private;
1141 if (from_schedule
|| current
->bio_list
) {
1142 spin_lock_irq(&conf
->device_lock
);
1143 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1144 conf
->pending_count
+= plug
->pending_cnt
;
1145 spin_unlock_irq(&conf
->device_lock
);
1146 wake_up(&conf
->wait_barrier
);
1147 md_wakeup_thread(mddev
->thread
);
1152 /* we aren't scheduling, so we can do the write-out directly. */
1153 bio
= bio_list_get(&plug
->pending
);
1154 bitmap_unplug(mddev
->bitmap
);
1155 wake_up(&conf
->wait_barrier
);
1157 while (bio
) { /* submit pending writes */
1158 struct bio
*next
= bio
->bi_next
;
1159 struct md_rdev
*rdev
= (void*)bio
->bi_bdev
;
1160 bio
->bi_next
= NULL
;
1161 bio
->bi_bdev
= rdev
->bdev
;
1162 if (test_bit(Faulty
, &rdev
->flags
)) {
1163 bio
->bi_error
= -EIO
;
1165 } else if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
1166 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1167 /* Just ignore it */
1170 generic_make_request(bio
);
1176 static inline struct r1bio
*
1177 alloc_r1bio(struct mddev
*mddev
, struct bio
*bio
, sector_t sectors_handled
)
1179 struct r1conf
*conf
= mddev
->private;
1180 struct r1bio
*r1_bio
;
1182 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1184 r1_bio
->master_bio
= bio
;
1185 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1187 r1_bio
->mddev
= mddev
;
1188 r1_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1193 static void raid1_read_request(struct mddev
*mddev
, struct bio
*bio
)
1195 struct r1conf
*conf
= mddev
->private;
1196 struct raid1_info
*mirror
;
1197 struct r1bio
*r1_bio
;
1198 struct bio
*read_bio
;
1199 struct bitmap
*bitmap
= mddev
->bitmap
;
1200 const int op
= bio_op(bio
);
1201 const unsigned long do_sync
= (bio
->bi_opf
& REQ_SYNC
);
1202 int sectors_handled
;
1207 * Still need barrier for READ in case that whole
1210 wait_read_barrier(conf
, bio
->bi_iter
.bi_sector
);
1212 r1_bio
= alloc_r1bio(mddev
, bio
, 0);
1215 * make_request() can abort the operation when read-ahead is being
1216 * used and no empty request is available.
1219 rdisk
= read_balance(conf
, r1_bio
, &max_sectors
);
1222 /* couldn't find anywhere to read from */
1223 raid_end_bio_io(r1_bio
);
1226 mirror
= conf
->mirrors
+ rdisk
;
1228 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
1231 * Reading from a write-mostly device must take care not to
1232 * over-take any writes that are 'behind'
1234 raid1_log(mddev
, "wait behind writes");
1235 wait_event(bitmap
->behind_wait
,
1236 atomic_read(&bitmap
->behind_writes
) == 0);
1238 r1_bio
->read_disk
= rdisk
;
1240 read_bio
= bio_clone_fast(bio
, GFP_NOIO
, mddev
->bio_set
);
1241 bio_trim(read_bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
1244 r1_bio
->bios
[rdisk
] = read_bio
;
1246 read_bio
->bi_iter
.bi_sector
= r1_bio
->sector
+
1247 mirror
->rdev
->data_offset
;
1248 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
1249 read_bio
->bi_end_io
= raid1_end_read_request
;
1250 bio_set_op_attrs(read_bio
, op
, do_sync
);
1251 if (test_bit(FailFast
, &mirror
->rdev
->flags
) &&
1252 test_bit(R1BIO_FailFast
, &r1_bio
->state
))
1253 read_bio
->bi_opf
|= MD_FAILFAST
;
1254 read_bio
->bi_private
= r1_bio
;
1257 trace_block_bio_remap(bdev_get_queue(read_bio
->bi_bdev
),
1258 read_bio
, disk_devt(mddev
->gendisk
),
1261 if (max_sectors
< r1_bio
->sectors
) {
1263 * could not read all from this device, so we will need another
1266 sectors_handled
= (r1_bio
->sector
+ max_sectors
1267 - bio
->bi_iter
.bi_sector
);
1268 r1_bio
->sectors
= max_sectors
;
1269 bio_inc_remaining(bio
);
1272 * Cannot call generic_make_request directly as that will be
1273 * queued in __make_request and subsequent mempool_alloc might
1274 * block waiting for it. So hand bio over to raid1d.
1276 reschedule_retry(r1_bio
);
1278 r1_bio
= alloc_r1bio(mddev
, bio
, sectors_handled
);
1281 generic_make_request(read_bio
);
1284 static void raid1_write_request(struct mddev
*mddev
, struct bio
*bio
)
1286 struct r1conf
*conf
= mddev
->private;
1287 struct r1bio
*r1_bio
;
1289 struct bitmap
*bitmap
= mddev
->bitmap
;
1290 unsigned long flags
;
1291 struct md_rdev
*blocked_rdev
;
1292 struct blk_plug_cb
*cb
;
1293 struct raid1_plug_cb
*plug
= NULL
;
1295 int sectors_handled
;
1299 * Register the new request and wait if the reconstruction
1300 * thread has put up a bar for new requests.
1301 * Continue immediately if no resync is active currently.
1304 md_write_start(mddev
, bio
); /* wait on superblock update early */
1306 if ((bio_end_sector(bio
) > mddev
->suspend_lo
&&
1307 bio
->bi_iter
.bi_sector
< mddev
->suspend_hi
) ||
1308 (mddev_is_clustered(mddev
) &&
1309 md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1310 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
)))) {
1313 * As the suspend_* range is controlled by userspace, we want
1314 * an interruptible wait.
1318 flush_signals(current
);
1319 prepare_to_wait(&conf
->wait_barrier
,
1320 &w
, TASK_INTERRUPTIBLE
);
1321 if (bio_end_sector(bio
) <= mddev
->suspend_lo
||
1322 bio
->bi_iter
.bi_sector
>= mddev
->suspend_hi
||
1323 (mddev_is_clustered(mddev
) &&
1324 !md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1325 bio
->bi_iter
.bi_sector
,
1326 bio_end_sector(bio
))))
1330 finish_wait(&conf
->wait_barrier
, &w
);
1332 wait_barrier(conf
, bio
->bi_iter
.bi_sector
);
1334 r1_bio
= alloc_r1bio(mddev
, bio
, 0);
1336 if (conf
->pending_count
>= max_queued_requests
) {
1337 md_wakeup_thread(mddev
->thread
);
1338 raid1_log(mddev
, "wait queued");
1339 wait_event(conf
->wait_barrier
,
1340 conf
->pending_count
< max_queued_requests
);
1342 /* first select target devices under rcu_lock and
1343 * inc refcount on their rdev. Record them by setting
1345 * If there are known/acknowledged bad blocks on any device on
1346 * which we have seen a write error, we want to avoid writing those
1348 * This potentially requires several writes to write around
1349 * the bad blocks. Each set of writes gets it's own r1bio
1350 * with a set of bios attached.
1353 disks
= conf
->raid_disks
* 2;
1355 blocked_rdev
= NULL
;
1357 max_sectors
= r1_bio
->sectors
;
1358 for (i
= 0; i
< disks
; i
++) {
1359 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1360 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1361 atomic_inc(&rdev
->nr_pending
);
1362 blocked_rdev
= rdev
;
1365 r1_bio
->bios
[i
] = NULL
;
1366 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
1367 if (i
< conf
->raid_disks
)
1368 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1372 atomic_inc(&rdev
->nr_pending
);
1373 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1378 is_bad
= is_badblock(rdev
, r1_bio
->sector
, max_sectors
,
1379 &first_bad
, &bad_sectors
);
1381 /* mustn't write here until the bad block is
1383 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1384 blocked_rdev
= rdev
;
1387 if (is_bad
&& first_bad
<= r1_bio
->sector
) {
1388 /* Cannot write here at all */
1389 bad_sectors
-= (r1_bio
->sector
- first_bad
);
1390 if (bad_sectors
< max_sectors
)
1391 /* mustn't write more than bad_sectors
1392 * to other devices yet
1394 max_sectors
= bad_sectors
;
1395 rdev_dec_pending(rdev
, mddev
);
1396 /* We don't set R1BIO_Degraded as that
1397 * only applies if the disk is
1398 * missing, so it might be re-added,
1399 * and we want to know to recover this
1401 * In this case the device is here,
1402 * and the fact that this chunk is not
1403 * in-sync is recorded in the bad
1409 int good_sectors
= first_bad
- r1_bio
->sector
;
1410 if (good_sectors
< max_sectors
)
1411 max_sectors
= good_sectors
;
1414 r1_bio
->bios
[i
] = bio
;
1418 if (unlikely(blocked_rdev
)) {
1419 /* Wait for this device to become unblocked */
1422 for (j
= 0; j
< i
; j
++)
1423 if (r1_bio
->bios
[j
])
1424 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
1426 allow_barrier(conf
, bio
->bi_iter
.bi_sector
);
1427 raid1_log(mddev
, "wait rdev %d blocked", blocked_rdev
->raid_disk
);
1428 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1429 wait_barrier(conf
, bio
->bi_iter
.bi_sector
);
1433 if (max_sectors
< r1_bio
->sectors
)
1434 r1_bio
->sectors
= max_sectors
;
1436 sectors_handled
= r1_bio
->sector
+ max_sectors
- bio
->bi_iter
.bi_sector
;
1438 atomic_set(&r1_bio
->remaining
, 1);
1439 atomic_set(&r1_bio
->behind_remaining
, 0);
1442 for (i
= 0; i
< disks
; i
++) {
1443 struct bio
*mbio
= NULL
;
1445 if (!r1_bio
->bios
[i
])
1448 offset
= r1_bio
->sector
- bio
->bi_iter
.bi_sector
;
1452 * Not if there are too many, or cannot
1453 * allocate memory, or a reader on WriteMostly
1454 * is waiting for behind writes to flush */
1456 (atomic_read(&bitmap
->behind_writes
)
1457 < mddev
->bitmap_info
.max_write_behind
) &&
1458 !waitqueue_active(&bitmap
->behind_wait
)) {
1459 mbio
= bio_clone_bioset_partial(bio
, GFP_NOIO
,
1463 alloc_behind_pages(mbio
, r1_bio
);
1466 bitmap_startwrite(bitmap
, r1_bio
->sector
,
1468 test_bit(R1BIO_BehindIO
,
1474 if (r1_bio
->behind_bvecs
)
1475 mbio
= bio_clone_bioset_partial(bio
, GFP_NOIO
,
1480 mbio
= bio_clone_fast(bio
, GFP_NOIO
, mddev
->bio_set
);
1481 bio_trim(mbio
, offset
, max_sectors
);
1485 if (r1_bio
->behind_bvecs
) {
1486 struct bio_vec
*bvec
;
1490 * We trimmed the bio, so _all is legit
1492 bio_for_each_segment_all(bvec
, mbio
, j
)
1493 bvec
->bv_page
= r1_bio
->behind_bvecs
[j
].bv_page
;
1494 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
1495 atomic_inc(&r1_bio
->behind_remaining
);
1498 r1_bio
->bios
[i
] = mbio
;
1500 mbio
->bi_iter
.bi_sector
= (r1_bio
->sector
+
1501 conf
->mirrors
[i
].rdev
->data_offset
);
1502 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1503 mbio
->bi_end_io
= raid1_end_write_request
;
1504 mbio
->bi_opf
= bio_op(bio
) | (bio
->bi_opf
& (REQ_SYNC
| REQ_FUA
));
1505 if (test_bit(FailFast
, &conf
->mirrors
[i
].rdev
->flags
) &&
1506 !test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
) &&
1507 conf
->raid_disks
- mddev
->degraded
> 1)
1508 mbio
->bi_opf
|= MD_FAILFAST
;
1509 mbio
->bi_private
= r1_bio
;
1511 atomic_inc(&r1_bio
->remaining
);
1514 trace_block_bio_remap(bdev_get_queue(mbio
->bi_bdev
),
1515 mbio
, disk_devt(mddev
->gendisk
),
1517 /* flush_pending_writes() needs access to the rdev so...*/
1518 mbio
->bi_bdev
= (void*)conf
->mirrors
[i
].rdev
;
1520 cb
= blk_check_plugged(raid1_unplug
, mddev
, sizeof(*plug
));
1522 plug
= container_of(cb
, struct raid1_plug_cb
, cb
);
1525 spin_lock_irqsave(&conf
->device_lock
, flags
);
1527 bio_list_add(&plug
->pending
, mbio
);
1528 plug
->pending_cnt
++;
1530 bio_list_add(&conf
->pending_bio_list
, mbio
);
1531 conf
->pending_count
++;
1533 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1535 md_wakeup_thread(mddev
->thread
);
1537 /* Mustn't call r1_bio_write_done before this next test,
1538 * as it could result in the bio being freed.
1540 if (sectors_handled
< bio_sectors(bio
)) {
1541 /* We need another r1_bio, which must be counted */
1542 sector_t sect
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1544 inc_pending(conf
, sect
);
1545 bio_inc_remaining(bio
);
1546 r1_bio_write_done(r1_bio
);
1547 r1_bio
= alloc_r1bio(mddev
, bio
, sectors_handled
);
1551 r1_bio_write_done(r1_bio
);
1553 /* In case raid1d snuck in to freeze_array */
1554 wake_up(&conf
->wait_barrier
);
1557 static void raid1_make_request(struct mddev
*mddev
, struct bio
*bio
)
1562 if (unlikely(bio
->bi_opf
& REQ_PREFLUSH
)) {
1563 md_flush_request(mddev
, bio
);
1567 /* if bio exceeds barrier unit boundary, split it */
1569 sectors
= align_to_barrier_unit_end(
1570 bio
->bi_iter
.bi_sector
, bio_sectors(bio
));
1571 if (sectors
< bio_sectors(bio
)) {
1572 split
= bio_split(bio
, sectors
, GFP_NOIO
, fs_bio_set
);
1573 bio_chain(split
, bio
);
1578 if (bio_data_dir(split
) == READ
) {
1579 raid1_read_request(mddev
, split
);
1582 * If a bio is splitted, the first part of bio will
1583 * pass barrier but the bio is queued in
1584 * current->bio_list (see generic_make_request). If
1585 * there is a raise_barrier() called here, the second
1586 * part of bio can't pass barrier. But since the first
1587 * part bio isn't dispatched to underlaying disks yet,
1588 * the barrier is never released, hence raise_barrier
1589 * will alays wait. We have a deadlock.
1590 * Note, this only happens in read path. For write
1591 * path, the first part of bio is dispatched in a
1592 * schedule() call (because of blk plug) or offloaded
1594 * Quitting from the function immediately can change
1595 * the bio order queued in bio_list and avoid the deadlock.
1598 generic_make_request(bio
);
1602 raid1_write_request(mddev
, split
);
1603 } while (split
!= bio
);
1606 static void raid1_status(struct seq_file
*seq
, struct mddev
*mddev
)
1608 struct r1conf
*conf
= mddev
->private;
1611 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1612 conf
->raid_disks
- mddev
->degraded
);
1614 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1615 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1616 seq_printf(seq
, "%s",
1617 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1620 seq_printf(seq
, "]");
1623 static void raid1_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1625 char b
[BDEVNAME_SIZE
];
1626 struct r1conf
*conf
= mddev
->private;
1627 unsigned long flags
;
1630 * If it is not operational, then we have already marked it as dead
1631 * else if it is the last working disks, ignore the error, let the
1632 * next level up know.
1633 * else mark the drive as failed
1635 spin_lock_irqsave(&conf
->device_lock
, flags
);
1636 if (test_bit(In_sync
, &rdev
->flags
)
1637 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1639 * Don't fail the drive, act as though we were just a
1640 * normal single drive.
1641 * However don't try a recovery from this drive as
1642 * it is very likely to fail.
1644 conf
->recovery_disabled
= mddev
->recovery_disabled
;
1645 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1648 set_bit(Blocked
, &rdev
->flags
);
1649 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1651 set_bit(Faulty
, &rdev
->flags
);
1653 set_bit(Faulty
, &rdev
->flags
);
1654 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1656 * if recovery is running, make sure it aborts.
1658 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1659 set_mask_bits(&mddev
->sb_flags
, 0,
1660 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
1661 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1662 "md/raid1:%s: Operation continuing on %d devices.\n",
1663 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1664 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1667 static void print_conf(struct r1conf
*conf
)
1671 pr_debug("RAID1 conf printout:\n");
1673 pr_debug("(!conf)\n");
1676 pr_debug(" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1680 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1681 char b
[BDEVNAME_SIZE
];
1682 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1684 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1685 i
, !test_bit(In_sync
, &rdev
->flags
),
1686 !test_bit(Faulty
, &rdev
->flags
),
1687 bdevname(rdev
->bdev
,b
));
1692 static void close_sync(struct r1conf
*conf
)
1694 wait_all_barriers(conf
);
1695 allow_all_barriers(conf
);
1697 mempool_destroy(conf
->r1buf_pool
);
1698 conf
->r1buf_pool
= NULL
;
1701 static int raid1_spare_active(struct mddev
*mddev
)
1704 struct r1conf
*conf
= mddev
->private;
1706 unsigned long flags
;
1709 * Find all failed disks within the RAID1 configuration
1710 * and mark them readable.
1711 * Called under mddev lock, so rcu protection not needed.
1712 * device_lock used to avoid races with raid1_end_read_request
1713 * which expects 'In_sync' flags and ->degraded to be consistent.
1715 spin_lock_irqsave(&conf
->device_lock
, flags
);
1716 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1717 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1718 struct md_rdev
*repl
= conf
->mirrors
[conf
->raid_disks
+ i
].rdev
;
1720 && !test_bit(Candidate
, &repl
->flags
)
1721 && repl
->recovery_offset
== MaxSector
1722 && !test_bit(Faulty
, &repl
->flags
)
1723 && !test_and_set_bit(In_sync
, &repl
->flags
)) {
1724 /* replacement has just become active */
1726 !test_and_clear_bit(In_sync
, &rdev
->flags
))
1729 /* Replaced device not technically
1730 * faulty, but we need to be sure
1731 * it gets removed and never re-added
1733 set_bit(Faulty
, &rdev
->flags
);
1734 sysfs_notify_dirent_safe(
1739 && rdev
->recovery_offset
== MaxSector
1740 && !test_bit(Faulty
, &rdev
->flags
)
1741 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1743 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
1746 mddev
->degraded
-= count
;
1747 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1753 static int raid1_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1755 struct r1conf
*conf
= mddev
->private;
1758 struct raid1_info
*p
;
1760 int last
= conf
->raid_disks
- 1;
1762 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
1765 if (md_integrity_add_rdev(rdev
, mddev
))
1768 if (rdev
->raid_disk
>= 0)
1769 first
= last
= rdev
->raid_disk
;
1772 * find the disk ... but prefer rdev->saved_raid_disk
1775 if (rdev
->saved_raid_disk
>= 0 &&
1776 rdev
->saved_raid_disk
>= first
&&
1777 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1778 first
= last
= rdev
->saved_raid_disk
;
1780 for (mirror
= first
; mirror
<= last
; mirror
++) {
1781 p
= conf
->mirrors
+mirror
;
1785 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1786 rdev
->data_offset
<< 9);
1788 p
->head_position
= 0;
1789 rdev
->raid_disk
= mirror
;
1791 /* As all devices are equivalent, we don't need a full recovery
1792 * if this was recently any drive of the array
1794 if (rdev
->saved_raid_disk
< 0)
1796 rcu_assign_pointer(p
->rdev
, rdev
);
1799 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
1800 p
[conf
->raid_disks
].rdev
== NULL
) {
1801 /* Add this device as a replacement */
1802 clear_bit(In_sync
, &rdev
->flags
);
1803 set_bit(Replacement
, &rdev
->flags
);
1804 rdev
->raid_disk
= mirror
;
1807 rcu_assign_pointer(p
[conf
->raid_disks
].rdev
, rdev
);
1811 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1812 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1817 static int raid1_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1819 struct r1conf
*conf
= mddev
->private;
1821 int number
= rdev
->raid_disk
;
1822 struct raid1_info
*p
= conf
->mirrors
+ number
;
1824 if (rdev
!= p
->rdev
)
1825 p
= conf
->mirrors
+ conf
->raid_disks
+ number
;
1828 if (rdev
== p
->rdev
) {
1829 if (test_bit(In_sync
, &rdev
->flags
) ||
1830 atomic_read(&rdev
->nr_pending
)) {
1834 /* Only remove non-faulty devices if recovery
1837 if (!test_bit(Faulty
, &rdev
->flags
) &&
1838 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
1839 mddev
->degraded
< conf
->raid_disks
) {
1844 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
1846 if (atomic_read(&rdev
->nr_pending
)) {
1847 /* lost the race, try later */
1853 if (conf
->mirrors
[conf
->raid_disks
+ number
].rdev
) {
1854 /* We just removed a device that is being replaced.
1855 * Move down the replacement. We drain all IO before
1856 * doing this to avoid confusion.
1858 struct md_rdev
*repl
=
1859 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
;
1860 freeze_array(conf
, 0);
1861 clear_bit(Replacement
, &repl
->flags
);
1863 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
= NULL
;
1864 unfreeze_array(conf
);
1865 clear_bit(WantReplacement
, &rdev
->flags
);
1867 clear_bit(WantReplacement
, &rdev
->flags
);
1868 err
= md_integrity_register(mddev
);
1876 static void end_sync_read(struct bio
*bio
)
1878 struct r1bio
*r1_bio
= get_resync_r1bio(bio
);
1880 update_head_pos(r1_bio
->read_disk
, r1_bio
);
1883 * we have read a block, now it needs to be re-written,
1884 * or re-read if the read failed.
1885 * We don't do much here, just schedule handling by raid1d
1888 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1890 if (atomic_dec_and_test(&r1_bio
->remaining
))
1891 reschedule_retry(r1_bio
);
1894 static void end_sync_write(struct bio
*bio
)
1896 int uptodate
= !bio
->bi_error
;
1897 struct r1bio
*r1_bio
= get_resync_r1bio(bio
);
1898 struct mddev
*mddev
= r1_bio
->mddev
;
1899 struct r1conf
*conf
= mddev
->private;
1902 struct md_rdev
*rdev
= conf
->mirrors
[find_bio_disk(r1_bio
, bio
)].rdev
;
1905 sector_t sync_blocks
= 0;
1906 sector_t s
= r1_bio
->sector
;
1907 long sectors_to_go
= r1_bio
->sectors
;
1908 /* make sure these bits doesn't get cleared. */
1910 bitmap_end_sync(mddev
->bitmap
, s
,
1913 sectors_to_go
-= sync_blocks
;
1914 } while (sectors_to_go
> 0);
1915 set_bit(WriteErrorSeen
, &rdev
->flags
);
1916 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1917 set_bit(MD_RECOVERY_NEEDED
, &
1919 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
1920 } else if (is_badblock(rdev
, r1_bio
->sector
, r1_bio
->sectors
,
1921 &first_bad
, &bad_sectors
) &&
1922 !is_badblock(conf
->mirrors
[r1_bio
->read_disk
].rdev
,
1925 &first_bad
, &bad_sectors
)
1927 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
1929 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1930 int s
= r1_bio
->sectors
;
1931 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
1932 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1933 reschedule_retry(r1_bio
);
1936 md_done_sync(mddev
, s
, uptodate
);
1941 static int r1_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1942 int sectors
, struct page
*page
, int rw
)
1944 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, 0, false))
1948 set_bit(WriteErrorSeen
, &rdev
->flags
);
1949 if (!test_and_set_bit(WantReplacement
,
1951 set_bit(MD_RECOVERY_NEEDED
, &
1952 rdev
->mddev
->recovery
);
1954 /* need to record an error - either for the block or the device */
1955 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1956 md_error(rdev
->mddev
, rdev
);
1960 static int fix_sync_read_error(struct r1bio
*r1_bio
)
1962 /* Try some synchronous reads of other devices to get
1963 * good data, much like with normal read errors. Only
1964 * read into the pages we already have so we don't
1965 * need to re-issue the read request.
1966 * We don't need to freeze the array, because being in an
1967 * active sync request, there is no normal IO, and
1968 * no overlapping syncs.
1969 * We don't need to check is_badblock() again as we
1970 * made sure that anything with a bad block in range
1971 * will have bi_end_io clear.
1973 struct mddev
*mddev
= r1_bio
->mddev
;
1974 struct r1conf
*conf
= mddev
->private;
1975 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1976 struct page
**pages
= get_resync_pages(bio
)->pages
;
1977 sector_t sect
= r1_bio
->sector
;
1978 int sectors
= r1_bio
->sectors
;
1980 struct md_rdev
*rdev
;
1982 rdev
= conf
->mirrors
[r1_bio
->read_disk
].rdev
;
1983 if (test_bit(FailFast
, &rdev
->flags
)) {
1984 /* Don't try recovering from here - just fail it
1985 * ... unless it is the last working device of course */
1986 md_error(mddev
, rdev
);
1987 if (test_bit(Faulty
, &rdev
->flags
))
1988 /* Don't try to read from here, but make sure
1989 * put_buf does it's thing
1991 bio
->bi_end_io
= end_sync_write
;
1996 int d
= r1_bio
->read_disk
;
2000 if (s
> (PAGE_SIZE
>>9))
2003 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
2004 /* No rcu protection needed here devices
2005 * can only be removed when no resync is
2006 * active, and resync is currently active
2008 rdev
= conf
->mirrors
[d
].rdev
;
2009 if (sync_page_io(rdev
, sect
, s
<<9,
2011 REQ_OP_READ
, 0, false)) {
2017 if (d
== conf
->raid_disks
* 2)
2019 } while (!success
&& d
!= r1_bio
->read_disk
);
2022 char b
[BDEVNAME_SIZE
];
2024 /* Cannot read from anywhere, this block is lost.
2025 * Record a bad block on each device. If that doesn't
2026 * work just disable and interrupt the recovery.
2027 * Don't fail devices as that won't really help.
2029 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2031 bdevname(bio
->bi_bdev
, b
),
2032 (unsigned long long)r1_bio
->sector
);
2033 for (d
= 0; d
< conf
->raid_disks
* 2; d
++) {
2034 rdev
= conf
->mirrors
[d
].rdev
;
2035 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
2037 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
2041 conf
->recovery_disabled
=
2042 mddev
->recovery_disabled
;
2043 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2044 md_done_sync(mddev
, r1_bio
->sectors
, 0);
2056 /* write it back and re-read */
2057 while (d
!= r1_bio
->read_disk
) {
2059 d
= conf
->raid_disks
* 2;
2061 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
2063 rdev
= conf
->mirrors
[d
].rdev
;
2064 if (r1_sync_page_io(rdev
, sect
, s
,
2067 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
2068 rdev_dec_pending(rdev
, mddev
);
2072 while (d
!= r1_bio
->read_disk
) {
2074 d
= conf
->raid_disks
* 2;
2076 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
2078 rdev
= conf
->mirrors
[d
].rdev
;
2079 if (r1_sync_page_io(rdev
, sect
, s
,
2082 atomic_add(s
, &rdev
->corrected_errors
);
2088 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
2093 static void process_checks(struct r1bio
*r1_bio
)
2095 /* We have read all readable devices. If we haven't
2096 * got the block, then there is no hope left.
2097 * If we have, then we want to do a comparison
2098 * and skip the write if everything is the same.
2099 * If any blocks failed to read, then we need to
2100 * attempt an over-write
2102 struct mddev
*mddev
= r1_bio
->mddev
;
2103 struct r1conf
*conf
= mddev
->private;
2108 /* Fix variable parts of all bios */
2109 vcnt
= (r1_bio
->sectors
+ PAGE_SIZE
/ 512 - 1) >> (PAGE_SHIFT
- 9);
2110 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2115 struct bio
*b
= r1_bio
->bios
[i
];
2116 struct resync_pages
*rp
= get_resync_pages(b
);
2117 if (b
->bi_end_io
!= end_sync_read
)
2119 /* fixup the bio for reuse, but preserve errno */
2120 error
= b
->bi_error
;
2122 b
->bi_error
= error
;
2124 b
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
2125 b
->bi_iter
.bi_sector
= r1_bio
->sector
+
2126 conf
->mirrors
[i
].rdev
->data_offset
;
2127 b
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
2128 b
->bi_end_io
= end_sync_read
;
2129 rp
->raid_bio
= r1_bio
;
2132 size
= b
->bi_iter
.bi_size
;
2133 bio_for_each_segment_all(bi
, b
, j
) {
2135 if (size
> PAGE_SIZE
)
2136 bi
->bv_len
= PAGE_SIZE
;
2142 for (primary
= 0; primary
< conf
->raid_disks
* 2; primary
++)
2143 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
2144 !r1_bio
->bios
[primary
]->bi_error
) {
2145 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
2146 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
2149 r1_bio
->read_disk
= primary
;
2150 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2152 struct bio
*pbio
= r1_bio
->bios
[primary
];
2153 struct bio
*sbio
= r1_bio
->bios
[i
];
2154 int error
= sbio
->bi_error
;
2155 struct page
**ppages
= get_resync_pages(pbio
)->pages
;
2156 struct page
**spages
= get_resync_pages(sbio
)->pages
;
2158 int page_len
[RESYNC_PAGES
];
2160 if (sbio
->bi_end_io
!= end_sync_read
)
2162 /* Now we can 'fixup' the error value */
2165 bio_for_each_segment_all(bi
, sbio
, j
)
2166 page_len
[j
] = bi
->bv_len
;
2169 for (j
= vcnt
; j
-- ; ) {
2170 if (memcmp(page_address(ppages
[j
]),
2171 page_address(spages
[j
]),
2178 atomic64_add(r1_bio
->sectors
, &mddev
->resync_mismatches
);
2179 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
2181 /* No need to write to this device. */
2182 sbio
->bi_end_io
= NULL
;
2183 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
2187 bio_copy_data(sbio
, pbio
);
2191 static void sync_request_write(struct mddev
*mddev
, struct r1bio
*r1_bio
)
2193 struct r1conf
*conf
= mddev
->private;
2195 int disks
= conf
->raid_disks
* 2;
2196 struct bio
*bio
, *wbio
;
2198 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2200 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
2201 /* ouch - failed to read all of that. */
2202 if (!fix_sync_read_error(r1_bio
))
2205 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2206 process_checks(r1_bio
);
2211 atomic_set(&r1_bio
->remaining
, 1);
2212 for (i
= 0; i
< disks
; i
++) {
2213 wbio
= r1_bio
->bios
[i
];
2214 if (wbio
->bi_end_io
== NULL
||
2215 (wbio
->bi_end_io
== end_sync_read
&&
2216 (i
== r1_bio
->read_disk
||
2217 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
2220 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2221 if (test_bit(FailFast
, &conf
->mirrors
[i
].rdev
->flags
))
2222 wbio
->bi_opf
|= MD_FAILFAST
;
2224 wbio
->bi_end_io
= end_sync_write
;
2225 atomic_inc(&r1_bio
->remaining
);
2226 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, bio_sectors(wbio
));
2228 generic_make_request(wbio
);
2231 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
2232 /* if we're here, all write(s) have completed, so clean up */
2233 int s
= r1_bio
->sectors
;
2234 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2235 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2236 reschedule_retry(r1_bio
);
2239 md_done_sync(mddev
, s
, 1);
2245 * This is a kernel thread which:
2247 * 1. Retries failed read operations on working mirrors.
2248 * 2. Updates the raid superblock when problems encounter.
2249 * 3. Performs writes following reads for array synchronising.
2252 static void fix_read_error(struct r1conf
*conf
, int read_disk
,
2253 sector_t sect
, int sectors
)
2255 struct mddev
*mddev
= conf
->mddev
;
2261 struct md_rdev
*rdev
;
2263 if (s
> (PAGE_SIZE
>>9))
2271 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2273 (test_bit(In_sync
, &rdev
->flags
) ||
2274 (!test_bit(Faulty
, &rdev
->flags
) &&
2275 rdev
->recovery_offset
>= sect
+ s
)) &&
2276 is_badblock(rdev
, sect
, s
,
2277 &first_bad
, &bad_sectors
) == 0) {
2278 atomic_inc(&rdev
->nr_pending
);
2280 if (sync_page_io(rdev
, sect
, s
<<9,
2281 conf
->tmppage
, REQ_OP_READ
, 0, false))
2283 rdev_dec_pending(rdev
, mddev
);
2289 if (d
== conf
->raid_disks
* 2)
2291 } while (!success
&& d
!= read_disk
);
2294 /* Cannot read from anywhere - mark it bad */
2295 struct md_rdev
*rdev
= conf
->mirrors
[read_disk
].rdev
;
2296 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
2297 md_error(mddev
, rdev
);
2300 /* write it back and re-read */
2302 while (d
!= read_disk
) {
2304 d
= conf
->raid_disks
* 2;
2307 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2309 !test_bit(Faulty
, &rdev
->flags
)) {
2310 atomic_inc(&rdev
->nr_pending
);
2312 r1_sync_page_io(rdev
, sect
, s
,
2313 conf
->tmppage
, WRITE
);
2314 rdev_dec_pending(rdev
, mddev
);
2319 while (d
!= read_disk
) {
2320 char b
[BDEVNAME_SIZE
];
2322 d
= conf
->raid_disks
* 2;
2325 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2327 !test_bit(Faulty
, &rdev
->flags
)) {
2328 atomic_inc(&rdev
->nr_pending
);
2330 if (r1_sync_page_io(rdev
, sect
, s
,
2331 conf
->tmppage
, READ
)) {
2332 atomic_add(s
, &rdev
->corrected_errors
);
2333 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2335 (unsigned long long)(sect
+
2337 bdevname(rdev
->bdev
, b
));
2339 rdev_dec_pending(rdev
, mddev
);
2348 static int narrow_write_error(struct r1bio
*r1_bio
, int i
)
2350 struct mddev
*mddev
= r1_bio
->mddev
;
2351 struct r1conf
*conf
= mddev
->private;
2352 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2354 /* bio has the data to be written to device 'i' where
2355 * we just recently had a write error.
2356 * We repeatedly clone the bio and trim down to one block,
2357 * then try the write. Where the write fails we record
2359 * It is conceivable that the bio doesn't exactly align with
2360 * blocks. We must handle this somehow.
2362 * We currently own a reference on the rdev.
2368 int sect_to_write
= r1_bio
->sectors
;
2371 if (rdev
->badblocks
.shift
< 0)
2374 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2375 bdev_logical_block_size(rdev
->bdev
) >> 9);
2376 sector
= r1_bio
->sector
;
2377 sectors
= ((sector
+ block_sectors
)
2378 & ~(sector_t
)(block_sectors
- 1))
2381 while (sect_to_write
) {
2383 if (sectors
> sect_to_write
)
2384 sectors
= sect_to_write
;
2385 /* Write at 'sector' for 'sectors'*/
2387 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
2388 unsigned vcnt
= r1_bio
->behind_page_count
;
2389 struct bio_vec
*vec
= r1_bio
->behind_bvecs
;
2391 while (!vec
->bv_page
) {
2396 wbio
= bio_alloc_mddev(GFP_NOIO
, vcnt
, mddev
);
2397 memcpy(wbio
->bi_io_vec
, vec
, vcnt
* sizeof(struct bio_vec
));
2399 wbio
->bi_vcnt
= vcnt
;
2401 wbio
= bio_clone_fast(r1_bio
->master_bio
, GFP_NOIO
,
2405 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2406 wbio
->bi_iter
.bi_sector
= r1_bio
->sector
;
2407 wbio
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
2409 bio_trim(wbio
, sector
- r1_bio
->sector
, sectors
);
2410 wbio
->bi_iter
.bi_sector
+= rdev
->data_offset
;
2411 wbio
->bi_bdev
= rdev
->bdev
;
2413 if (submit_bio_wait(wbio
) < 0)
2415 ok
= rdev_set_badblocks(rdev
, sector
,
2420 sect_to_write
-= sectors
;
2422 sectors
= block_sectors
;
2427 static void handle_sync_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2430 int s
= r1_bio
->sectors
;
2431 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++) {
2432 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2433 struct bio
*bio
= r1_bio
->bios
[m
];
2434 if (bio
->bi_end_io
== NULL
)
2436 if (!bio
->bi_error
&&
2437 test_bit(R1BIO_MadeGood
, &r1_bio
->state
)) {
2438 rdev_clear_badblocks(rdev
, r1_bio
->sector
, s
, 0);
2440 if (bio
->bi_error
&&
2441 test_bit(R1BIO_WriteError
, &r1_bio
->state
)) {
2442 if (!rdev_set_badblocks(rdev
, r1_bio
->sector
, s
, 0))
2443 md_error(conf
->mddev
, rdev
);
2447 md_done_sync(conf
->mddev
, s
, 1);
2450 static void handle_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2455 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++)
2456 if (r1_bio
->bios
[m
] == IO_MADE_GOOD
) {
2457 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2458 rdev_clear_badblocks(rdev
,
2460 r1_bio
->sectors
, 0);
2461 rdev_dec_pending(rdev
, conf
->mddev
);
2462 } else if (r1_bio
->bios
[m
] != NULL
) {
2463 /* This drive got a write error. We need to
2464 * narrow down and record precise write
2468 if (!narrow_write_error(r1_bio
, m
)) {
2469 md_error(conf
->mddev
,
2470 conf
->mirrors
[m
].rdev
);
2471 /* an I/O failed, we can't clear the bitmap */
2472 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2474 rdev_dec_pending(conf
->mirrors
[m
].rdev
,
2478 spin_lock_irq(&conf
->device_lock
);
2479 list_add(&r1_bio
->retry_list
, &conf
->bio_end_io_list
);
2480 idx
= sector_to_idx(r1_bio
->sector
);
2481 atomic_inc(&conf
->nr_queued
[idx
]);
2482 spin_unlock_irq(&conf
->device_lock
);
2484 * In case freeze_array() is waiting for condition
2485 * get_unqueued_pending() == extra to be true.
2487 wake_up(&conf
->wait_barrier
);
2488 md_wakeup_thread(conf
->mddev
->thread
);
2490 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2491 close_write(r1_bio
);
2492 raid_end_bio_io(r1_bio
);
2496 static void handle_read_error(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2500 struct mddev
*mddev
= conf
->mddev
;
2502 char b
[BDEVNAME_SIZE
];
2503 struct md_rdev
*rdev
;
2505 sector_t bio_sector
;
2507 clear_bit(R1BIO_ReadError
, &r1_bio
->state
);
2508 /* we got a read error. Maybe the drive is bad. Maybe just
2509 * the block and we can fix it.
2510 * We freeze all other IO, and try reading the block from
2511 * other devices. When we find one, we re-write
2512 * and check it that fixes the read error.
2513 * This is all done synchronously while the array is
2517 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2518 bdevname(bio
->bi_bdev
, b
);
2519 bio_dev
= bio
->bi_bdev
->bd_dev
;
2520 bio_sector
= conf
->mirrors
[r1_bio
->read_disk
].rdev
->data_offset
+ r1_bio
->sector
;
2522 r1_bio
->bios
[r1_bio
->read_disk
] = NULL
;
2524 rdev
= conf
->mirrors
[r1_bio
->read_disk
].rdev
;
2526 && !test_bit(FailFast
, &rdev
->flags
)) {
2527 freeze_array(conf
, 1);
2528 fix_read_error(conf
, r1_bio
->read_disk
,
2529 r1_bio
->sector
, r1_bio
->sectors
);
2530 unfreeze_array(conf
);
2532 r1_bio
->bios
[r1_bio
->read_disk
] = IO_BLOCKED
;
2535 rdev_dec_pending(rdev
, conf
->mddev
);
2538 disk
= read_balance(conf
, r1_bio
, &max_sectors
);
2540 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2541 mdname(mddev
), b
, (unsigned long long)r1_bio
->sector
);
2542 raid_end_bio_io(r1_bio
);
2544 const unsigned long do_sync
2545 = r1_bio
->master_bio
->bi_opf
& REQ_SYNC
;
2546 r1_bio
->read_disk
= disk
;
2547 bio
= bio_clone_fast(r1_bio
->master_bio
, GFP_NOIO
,
2549 bio_trim(bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
2551 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
2552 rdev
= conf
->mirrors
[disk
].rdev
;
2553 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
2555 (unsigned long long)r1_bio
->sector
,
2556 bdevname(rdev
->bdev
, b
));
2557 bio
->bi_iter
.bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
2558 bio
->bi_bdev
= rdev
->bdev
;
2559 bio
->bi_end_io
= raid1_end_read_request
;
2560 bio_set_op_attrs(bio
, REQ_OP_READ
, do_sync
);
2561 if (test_bit(FailFast
, &rdev
->flags
) &&
2562 test_bit(R1BIO_FailFast
, &r1_bio
->state
))
2563 bio
->bi_opf
|= MD_FAILFAST
;
2564 bio
->bi_private
= r1_bio
;
2565 if (max_sectors
< r1_bio
->sectors
) {
2566 /* Drat - have to split this up more */
2567 struct bio
*mbio
= r1_bio
->master_bio
;
2568 int sectors_handled
= (r1_bio
->sector
+ max_sectors
2569 - mbio
->bi_iter
.bi_sector
);
2570 r1_bio
->sectors
= max_sectors
;
2571 bio_inc_remaining(mbio
);
2572 trace_block_bio_remap(bdev_get_queue(bio
->bi_bdev
),
2573 bio
, bio_dev
, bio_sector
);
2574 generic_make_request(bio
);
2577 r1_bio
= alloc_r1bio(mddev
, mbio
, sectors_handled
);
2578 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
2579 inc_pending(conf
, r1_bio
->sector
);
2583 trace_block_bio_remap(bdev_get_queue(bio
->bi_bdev
),
2584 bio
, bio_dev
, bio_sector
);
2585 generic_make_request(bio
);
2590 static void raid1d(struct md_thread
*thread
)
2592 struct mddev
*mddev
= thread
->mddev
;
2593 struct r1bio
*r1_bio
;
2594 unsigned long flags
;
2595 struct r1conf
*conf
= mddev
->private;
2596 struct list_head
*head
= &conf
->retry_list
;
2597 struct blk_plug plug
;
2600 md_check_recovery(mddev
);
2602 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2603 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
)) {
2605 spin_lock_irqsave(&conf
->device_lock
, flags
);
2606 if (!test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
))
2607 list_splice_init(&conf
->bio_end_io_list
, &tmp
);
2608 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2609 while (!list_empty(&tmp
)) {
2610 r1_bio
= list_first_entry(&tmp
, struct r1bio
,
2612 list_del(&r1_bio
->retry_list
);
2613 idx
= sector_to_idx(r1_bio
->sector
);
2614 atomic_dec(&conf
->nr_queued
[idx
]);
2615 if (mddev
->degraded
)
2616 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2617 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2618 close_write(r1_bio
);
2619 raid_end_bio_io(r1_bio
);
2623 blk_start_plug(&plug
);
2626 flush_pending_writes(conf
);
2628 spin_lock_irqsave(&conf
->device_lock
, flags
);
2629 if (list_empty(head
)) {
2630 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2633 r1_bio
= list_entry(head
->prev
, struct r1bio
, retry_list
);
2634 list_del(head
->prev
);
2635 idx
= sector_to_idx(r1_bio
->sector
);
2636 atomic_dec(&conf
->nr_queued
[idx
]);
2637 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2639 mddev
= r1_bio
->mddev
;
2640 conf
= mddev
->private;
2641 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
2642 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2643 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2644 handle_sync_write_finished(conf
, r1_bio
);
2646 sync_request_write(mddev
, r1_bio
);
2647 } else if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2648 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2649 handle_write_finished(conf
, r1_bio
);
2650 else if (test_bit(R1BIO_ReadError
, &r1_bio
->state
))
2651 handle_read_error(conf
, r1_bio
);
2653 /* just a partial read to be scheduled from separate
2656 generic_make_request(r1_bio
->bios
[r1_bio
->read_disk
]);
2659 if (mddev
->sb_flags
& ~(1<<MD_SB_CHANGE_PENDING
))
2660 md_check_recovery(mddev
);
2662 blk_finish_plug(&plug
);
2665 static int init_resync(struct r1conf
*conf
)
2669 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2670 BUG_ON(conf
->r1buf_pool
);
2671 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
2673 if (!conf
->r1buf_pool
)
2679 * perform a "sync" on one "block"
2681 * We need to make sure that no normal I/O request - particularly write
2682 * requests - conflict with active sync requests.
2684 * This is achieved by tracking pending requests and a 'barrier' concept
2685 * that can be installed to exclude normal IO requests.
2688 static sector_t
raid1_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2691 struct r1conf
*conf
= mddev
->private;
2692 struct r1bio
*r1_bio
;
2694 sector_t max_sector
, nr_sectors
;
2698 int write_targets
= 0, read_targets
= 0;
2699 sector_t sync_blocks
;
2700 int still_degraded
= 0;
2701 int good_sectors
= RESYNC_SECTORS
;
2702 int min_bad
= 0; /* number of sectors that are bad in all devices */
2703 int idx
= sector_to_idx(sector_nr
);
2705 if (!conf
->r1buf_pool
)
2706 if (init_resync(conf
))
2709 max_sector
= mddev
->dev_sectors
;
2710 if (sector_nr
>= max_sector
) {
2711 /* If we aborted, we need to abort the
2712 * sync on the 'current' bitmap chunk (there will
2713 * only be one in raid1 resync.
2714 * We can find the current addess in mddev->curr_resync
2716 if (mddev
->curr_resync
< max_sector
) /* aborted */
2717 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2719 else /* completed sync */
2722 bitmap_close_sync(mddev
->bitmap
);
2725 if (mddev_is_clustered(mddev
)) {
2726 conf
->cluster_sync_low
= 0;
2727 conf
->cluster_sync_high
= 0;
2732 if (mddev
->bitmap
== NULL
&&
2733 mddev
->recovery_cp
== MaxSector
&&
2734 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2735 conf
->fullsync
== 0) {
2737 return max_sector
- sector_nr
;
2739 /* before building a request, check if we can skip these blocks..
2740 * This call the bitmap_start_sync doesn't actually record anything
2742 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
2743 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2744 /* We can skip this block, and probably several more */
2750 * If there is non-resync activity waiting for a turn, then let it
2751 * though before starting on this new sync request.
2753 if (atomic_read(&conf
->nr_waiting
[idx
]))
2754 schedule_timeout_uninterruptible(1);
2756 /* we are incrementing sector_nr below. To be safe, we check against
2757 * sector_nr + two times RESYNC_SECTORS
2760 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
,
2761 mddev_is_clustered(mddev
) && (sector_nr
+ 2 * RESYNC_SECTORS
> conf
->cluster_sync_high
));
2762 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
2764 raise_barrier(conf
, sector_nr
);
2768 * If we get a correctably read error during resync or recovery,
2769 * we might want to read from a different device. So we
2770 * flag all drives that could conceivably be read from for READ,
2771 * and any others (which will be non-In_sync devices) for WRITE.
2772 * If a read fails, we try reading from something else for which READ
2776 r1_bio
->mddev
= mddev
;
2777 r1_bio
->sector
= sector_nr
;
2779 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
2780 /* make sure good_sectors won't go across barrier unit boundary */
2781 good_sectors
= align_to_barrier_unit_end(sector_nr
, good_sectors
);
2783 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2784 struct md_rdev
*rdev
;
2785 bio
= r1_bio
->bios
[i
];
2787 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
2789 test_bit(Faulty
, &rdev
->flags
)) {
2790 if (i
< conf
->raid_disks
)
2792 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
2793 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
2794 bio
->bi_end_io
= end_sync_write
;
2797 /* may need to read from here */
2798 sector_t first_bad
= MaxSector
;
2801 if (is_badblock(rdev
, sector_nr
, good_sectors
,
2802 &first_bad
, &bad_sectors
)) {
2803 if (first_bad
> sector_nr
)
2804 good_sectors
= first_bad
- sector_nr
;
2806 bad_sectors
-= (sector_nr
- first_bad
);
2808 min_bad
> bad_sectors
)
2809 min_bad
= bad_sectors
;
2812 if (sector_nr
< first_bad
) {
2813 if (test_bit(WriteMostly
, &rdev
->flags
)) {
2820 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
2821 bio
->bi_end_io
= end_sync_read
;
2823 } else if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
2824 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2825 !test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)) {
2827 * The device is suitable for reading (InSync),
2828 * but has bad block(s) here. Let's try to correct them,
2829 * if we are doing resync or repair. Otherwise, leave
2830 * this device alone for this sync request.
2832 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
2833 bio
->bi_end_io
= end_sync_write
;
2837 if (bio
->bi_end_io
) {
2838 atomic_inc(&rdev
->nr_pending
);
2839 bio
->bi_iter
.bi_sector
= sector_nr
+ rdev
->data_offset
;
2840 bio
->bi_bdev
= rdev
->bdev
;
2841 if (test_bit(FailFast
, &rdev
->flags
))
2842 bio
->bi_opf
|= MD_FAILFAST
;
2848 r1_bio
->read_disk
= disk
;
2850 if (read_targets
== 0 && min_bad
> 0) {
2851 /* These sectors are bad on all InSync devices, so we
2852 * need to mark them bad on all write targets
2855 for (i
= 0 ; i
< conf
->raid_disks
* 2 ; i
++)
2856 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_write
) {
2857 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2858 ok
= rdev_set_badblocks(rdev
, sector_nr
,
2862 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
2867 /* Cannot record the badblocks, so need to
2869 * If there are multiple read targets, could just
2870 * fail the really bad ones ???
2872 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2873 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2879 if (min_bad
> 0 && min_bad
< good_sectors
) {
2880 /* only resync enough to reach the next bad->good
2882 good_sectors
= min_bad
;
2885 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
2886 /* extra read targets are also write targets */
2887 write_targets
+= read_targets
-1;
2889 if (write_targets
== 0 || read_targets
== 0) {
2890 /* There is nowhere to write, so all non-sync
2891 * drives must be failed - so we are finished
2895 max_sector
= sector_nr
+ min_bad
;
2896 rv
= max_sector
- sector_nr
;
2902 if (max_sector
> mddev
->resync_max
)
2903 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2904 if (max_sector
> sector_nr
+ good_sectors
)
2905 max_sector
= sector_nr
+ good_sectors
;
2910 int len
= PAGE_SIZE
;
2911 if (sector_nr
+ (len
>>9) > max_sector
)
2912 len
= (max_sector
- sector_nr
) << 9;
2915 if (sync_blocks
== 0) {
2916 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2917 &sync_blocks
, still_degraded
) &&
2919 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2921 if ((len
>> 9) > sync_blocks
)
2922 len
= sync_blocks
<<9;
2925 for (i
= 0 ; i
< conf
->raid_disks
* 2; i
++) {
2926 struct resync_pages
*rp
;
2928 bio
= r1_bio
->bios
[i
];
2929 rp
= get_resync_pages(bio
);
2930 if (bio
->bi_end_io
) {
2931 page
= resync_fetch_page(rp
, rp
->idx
++);
2934 * won't fail because the vec table is big
2935 * enough to hold all these pages
2937 bio_add_page(bio
, page
, len
, 0);
2940 nr_sectors
+= len
>>9;
2941 sector_nr
+= len
>>9;
2942 sync_blocks
-= (len
>>9);
2943 } while (get_resync_pages(r1_bio
->bios
[disk
]->bi_private
)->idx
< RESYNC_PAGES
);
2945 r1_bio
->sectors
= nr_sectors
;
2947 if (mddev_is_clustered(mddev
) &&
2948 conf
->cluster_sync_high
< sector_nr
+ nr_sectors
) {
2949 conf
->cluster_sync_low
= mddev
->curr_resync_completed
;
2950 conf
->cluster_sync_high
= conf
->cluster_sync_low
+ CLUSTER_RESYNC_WINDOW_SECTORS
;
2951 /* Send resync message */
2952 md_cluster_ops
->resync_info_update(mddev
,
2953 conf
->cluster_sync_low
,
2954 conf
->cluster_sync_high
);
2957 /* For a user-requested sync, we read all readable devices and do a
2960 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2961 atomic_set(&r1_bio
->remaining
, read_targets
);
2962 for (i
= 0; i
< conf
->raid_disks
* 2 && read_targets
; i
++) {
2963 bio
= r1_bio
->bios
[i
];
2964 if (bio
->bi_end_io
== end_sync_read
) {
2966 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2967 if (read_targets
== 1)
2968 bio
->bi_opf
&= ~MD_FAILFAST
;
2969 generic_make_request(bio
);
2973 atomic_set(&r1_bio
->remaining
, 1);
2974 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2975 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2976 if (read_targets
== 1)
2977 bio
->bi_opf
&= ~MD_FAILFAST
;
2978 generic_make_request(bio
);
2984 static sector_t
raid1_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2989 return mddev
->dev_sectors
;
2992 static struct r1conf
*setup_conf(struct mddev
*mddev
)
2994 struct r1conf
*conf
;
2996 struct raid1_info
*disk
;
2997 struct md_rdev
*rdev
;
3000 conf
= kzalloc(sizeof(struct r1conf
), GFP_KERNEL
);
3004 conf
->nr_pending
= kcalloc(BARRIER_BUCKETS_NR
,
3005 sizeof(atomic_t
), GFP_KERNEL
);
3006 if (!conf
->nr_pending
)
3009 conf
->nr_waiting
= kcalloc(BARRIER_BUCKETS_NR
,
3010 sizeof(atomic_t
), GFP_KERNEL
);
3011 if (!conf
->nr_waiting
)
3014 conf
->nr_queued
= kcalloc(BARRIER_BUCKETS_NR
,
3015 sizeof(atomic_t
), GFP_KERNEL
);
3016 if (!conf
->nr_queued
)
3019 conf
->barrier
= kcalloc(BARRIER_BUCKETS_NR
,
3020 sizeof(atomic_t
), GFP_KERNEL
);
3024 conf
->mirrors
= kzalloc(sizeof(struct raid1_info
)
3025 * mddev
->raid_disks
* 2,
3030 conf
->tmppage
= alloc_page(GFP_KERNEL
);
3034 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
3035 if (!conf
->poolinfo
)
3037 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
* 2;
3038 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
3041 if (!conf
->r1bio_pool
)
3044 conf
->poolinfo
->mddev
= mddev
;
3047 spin_lock_init(&conf
->device_lock
);
3048 rdev_for_each(rdev
, mddev
) {
3049 struct request_queue
*q
;
3050 int disk_idx
= rdev
->raid_disk
;
3051 if (disk_idx
>= mddev
->raid_disks
3054 if (test_bit(Replacement
, &rdev
->flags
))
3055 disk
= conf
->mirrors
+ mddev
->raid_disks
+ disk_idx
;
3057 disk
= conf
->mirrors
+ disk_idx
;
3062 q
= bdev_get_queue(rdev
->bdev
);
3064 disk
->head_position
= 0;
3065 disk
->seq_start
= MaxSector
;
3067 conf
->raid_disks
= mddev
->raid_disks
;
3068 conf
->mddev
= mddev
;
3069 INIT_LIST_HEAD(&conf
->retry_list
);
3070 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
3072 spin_lock_init(&conf
->resync_lock
);
3073 init_waitqueue_head(&conf
->wait_barrier
);
3075 bio_list_init(&conf
->pending_bio_list
);
3076 conf
->pending_count
= 0;
3077 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
3080 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
3082 disk
= conf
->mirrors
+ i
;
3084 if (i
< conf
->raid_disks
&&
3085 disk
[conf
->raid_disks
].rdev
) {
3086 /* This slot has a replacement. */
3088 /* No original, just make the replacement
3089 * a recovering spare
3092 disk
[conf
->raid_disks
].rdev
;
3093 disk
[conf
->raid_disks
].rdev
= NULL
;
3094 } else if (!test_bit(In_sync
, &disk
->rdev
->flags
))
3095 /* Original is not in_sync - bad */
3100 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
3101 disk
->head_position
= 0;
3103 (disk
->rdev
->saved_raid_disk
< 0))
3109 conf
->thread
= md_register_thread(raid1d
, mddev
, "raid1");
3117 mempool_destroy(conf
->r1bio_pool
);
3118 kfree(conf
->mirrors
);
3119 safe_put_page(conf
->tmppage
);
3120 kfree(conf
->poolinfo
);
3121 kfree(conf
->nr_pending
);
3122 kfree(conf
->nr_waiting
);
3123 kfree(conf
->nr_queued
);
3124 kfree(conf
->barrier
);
3127 return ERR_PTR(err
);
3130 static void raid1_free(struct mddev
*mddev
, void *priv
);
3131 static int raid1_run(struct mddev
*mddev
)
3133 struct r1conf
*conf
;
3135 struct md_rdev
*rdev
;
3137 bool discard_supported
= false;
3139 if (mddev
->level
!= 1) {
3140 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3141 mdname(mddev
), mddev
->level
);
3144 if (mddev
->reshape_position
!= MaxSector
) {
3145 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3150 * copy the already verified devices into our private RAID1
3151 * bookkeeping area. [whatever we allocate in run(),
3152 * should be freed in raid1_free()]
3154 if (mddev
->private == NULL
)
3155 conf
= setup_conf(mddev
);
3157 conf
= mddev
->private;
3160 return PTR_ERR(conf
);
3163 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
3165 rdev_for_each(rdev
, mddev
) {
3166 if (!mddev
->gendisk
)
3168 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
3169 rdev
->data_offset
<< 9);
3170 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
3171 discard_supported
= true;
3174 mddev
->degraded
= 0;
3175 for (i
=0; i
< conf
->raid_disks
; i
++)
3176 if (conf
->mirrors
[i
].rdev
== NULL
||
3177 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
3178 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
3181 if (conf
->raid_disks
- mddev
->degraded
== 1)
3182 mddev
->recovery_cp
= MaxSector
;
3184 if (mddev
->recovery_cp
!= MaxSector
)
3185 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3187 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3188 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
3192 * Ok, everything is just fine now
3194 mddev
->thread
= conf
->thread
;
3195 conf
->thread
= NULL
;
3196 mddev
->private = conf
;
3197 set_bit(MD_FAILFAST_SUPPORTED
, &mddev
->flags
);
3199 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
3202 if (discard_supported
)
3203 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
3206 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
3210 ret
= md_integrity_register(mddev
);
3212 md_unregister_thread(&mddev
->thread
);
3213 raid1_free(mddev
, conf
);
3218 static void raid1_free(struct mddev
*mddev
, void *priv
)
3220 struct r1conf
*conf
= priv
;
3222 mempool_destroy(conf
->r1bio_pool
);
3223 kfree(conf
->mirrors
);
3224 safe_put_page(conf
->tmppage
);
3225 kfree(conf
->poolinfo
);
3226 kfree(conf
->nr_pending
);
3227 kfree(conf
->nr_waiting
);
3228 kfree(conf
->nr_queued
);
3229 kfree(conf
->barrier
);
3233 static int raid1_resize(struct mddev
*mddev
, sector_t sectors
)
3235 /* no resync is happening, and there is enough space
3236 * on all devices, so we can resize.
3237 * We need to make sure resync covers any new space.
3238 * If the array is shrinking we should possibly wait until
3239 * any io in the removed space completes, but it hardly seems
3242 sector_t newsize
= raid1_size(mddev
, sectors
, 0);
3243 if (mddev
->external_size
&&
3244 mddev
->array_sectors
> newsize
)
3246 if (mddev
->bitmap
) {
3247 int ret
= bitmap_resize(mddev
->bitmap
, newsize
, 0, 0);
3251 md_set_array_sectors(mddev
, newsize
);
3252 if (sectors
> mddev
->dev_sectors
&&
3253 mddev
->recovery_cp
> mddev
->dev_sectors
) {
3254 mddev
->recovery_cp
= mddev
->dev_sectors
;
3255 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3257 mddev
->dev_sectors
= sectors
;
3258 mddev
->resync_max_sectors
= sectors
;
3262 static int raid1_reshape(struct mddev
*mddev
)
3265 * 1/ resize the r1bio_pool
3266 * 2/ resize conf->mirrors
3268 * We allocate a new r1bio_pool if we can.
3269 * Then raise a device barrier and wait until all IO stops.
3270 * Then resize conf->mirrors and swap in the new r1bio pool.
3272 * At the same time, we "pack" the devices so that all the missing
3273 * devices have the higher raid_disk numbers.
3275 mempool_t
*newpool
, *oldpool
;
3276 struct pool_info
*newpoolinfo
;
3277 struct raid1_info
*newmirrors
;
3278 struct r1conf
*conf
= mddev
->private;
3279 int cnt
, raid_disks
;
3280 unsigned long flags
;
3283 /* Cannot change chunk_size, layout, or level */
3284 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
3285 mddev
->layout
!= mddev
->new_layout
||
3286 mddev
->level
!= mddev
->new_level
) {
3287 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3288 mddev
->new_layout
= mddev
->layout
;
3289 mddev
->new_level
= mddev
->level
;
3293 if (!mddev_is_clustered(mddev
)) {
3294 err
= md_allow_write(mddev
);
3299 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3301 if (raid_disks
< conf
->raid_disks
) {
3303 for (d
= 0; d
< conf
->raid_disks
; d
++)
3304 if (conf
->mirrors
[d
].rdev
)
3306 if (cnt
> raid_disks
)
3310 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
3313 newpoolinfo
->mddev
= mddev
;
3314 newpoolinfo
->raid_disks
= raid_disks
* 2;
3316 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
3317 r1bio_pool_free
, newpoolinfo
);
3322 newmirrors
= kzalloc(sizeof(struct raid1_info
) * raid_disks
* 2,
3326 mempool_destroy(newpool
);
3330 freeze_array(conf
, 0);
3332 /* ok, everything is stopped */
3333 oldpool
= conf
->r1bio_pool
;
3334 conf
->r1bio_pool
= newpool
;
3336 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
3337 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
3338 if (rdev
&& rdev
->raid_disk
!= d2
) {
3339 sysfs_unlink_rdev(mddev
, rdev
);
3340 rdev
->raid_disk
= d2
;
3341 sysfs_unlink_rdev(mddev
, rdev
);
3342 if (sysfs_link_rdev(mddev
, rdev
))
3343 pr_warn("md/raid1:%s: cannot register rd%d\n",
3344 mdname(mddev
), rdev
->raid_disk
);
3347 newmirrors
[d2
++].rdev
= rdev
;
3349 kfree(conf
->mirrors
);
3350 conf
->mirrors
= newmirrors
;
3351 kfree(conf
->poolinfo
);
3352 conf
->poolinfo
= newpoolinfo
;
3354 spin_lock_irqsave(&conf
->device_lock
, flags
);
3355 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
3356 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3357 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
3358 mddev
->delta_disks
= 0;
3360 unfreeze_array(conf
);
3362 set_bit(MD_RECOVERY_RECOVER
, &mddev
->recovery
);
3363 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3364 md_wakeup_thread(mddev
->thread
);
3366 mempool_destroy(oldpool
);
3370 static void raid1_quiesce(struct mddev
*mddev
, int state
)
3372 struct r1conf
*conf
= mddev
->private;
3375 case 2: /* wake for suspend */
3376 wake_up(&conf
->wait_barrier
);
3379 freeze_array(conf
, 0);
3382 unfreeze_array(conf
);
3387 static void *raid1_takeover(struct mddev
*mddev
)
3389 /* raid1 can take over:
3390 * raid5 with 2 devices, any layout or chunk size
3392 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
3393 struct r1conf
*conf
;
3394 mddev
->new_level
= 1;
3395 mddev
->new_layout
= 0;
3396 mddev
->new_chunk_sectors
= 0;
3397 conf
= setup_conf(mddev
);
3398 if (!IS_ERR(conf
)) {
3399 /* Array must appear to be quiesced */
3400 conf
->array_frozen
= 1;
3401 mddev_clear_unsupported_flags(mddev
,
3402 UNSUPPORTED_MDDEV_FLAGS
);
3406 return ERR_PTR(-EINVAL
);
3409 static struct md_personality raid1_personality
=
3413 .owner
= THIS_MODULE
,
3414 .make_request
= raid1_make_request
,
3417 .status
= raid1_status
,
3418 .error_handler
= raid1_error
,
3419 .hot_add_disk
= raid1_add_disk
,
3420 .hot_remove_disk
= raid1_remove_disk
,
3421 .spare_active
= raid1_spare_active
,
3422 .sync_request
= raid1_sync_request
,
3423 .resize
= raid1_resize
,
3425 .check_reshape
= raid1_reshape
,
3426 .quiesce
= raid1_quiesce
,
3427 .takeover
= raid1_takeover
,
3428 .congested
= raid1_congested
,
3431 static int __init
raid_init(void)
3433 return register_md_personality(&raid1_personality
);
3436 static void raid_exit(void)
3438 unregister_md_personality(&raid1_personality
);
3441 module_init(raid_init
);
3442 module_exit(raid_exit
);
3443 MODULE_LICENSE("GPL");
3444 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3445 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3446 MODULE_ALIAS("md-raid1");
3447 MODULE_ALIAS("md-level-1");
3449 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);