2 * raid1.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * RAID-1 management functions.
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
45 * Number of guaranteed r1bios in case of extreme VM load:
47 #define NR_RAID1_BIOS 256
49 /* when we get a read error on a read-only array, we redirect to another
50 * device without failing the first device, or trying to over-write to
51 * correct the read error. To keep track of bad blocks on a per-bio
52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
54 #define IO_BLOCKED ((struct bio *)1)
55 /* When we successfully write to a known bad-block, we need to remove the
56 * bad-block marking which must be done from process context. So we record
57 * the success by setting devs[n].bio to IO_MADE_GOOD
59 #define IO_MADE_GOOD ((struct bio *)2)
61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
63 /* When there are this many requests queue to be written by
64 * the raid1 thread, we become 'congested' to provide back-pressure
67 static int max_queued_requests
= 1024;
69 static void allow_barrier(struct r1conf
*conf
, sector_t start_next_window
,
71 static void lower_barrier(struct r1conf
*conf
);
73 static void * r1bio_pool_alloc(gfp_t gfp_flags
, void *data
)
75 struct pool_info
*pi
= data
;
76 int size
= offsetof(struct r1bio
, bios
[pi
->raid_disks
]);
78 /* allocate a r1bio with room for raid_disks entries in the bios array */
79 return kzalloc(size
, gfp_flags
);
82 static void r1bio_pool_free(void *r1_bio
, void *data
)
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_DEPTH 32
89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
93 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
94 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
95 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
97 static void * r1buf_pool_alloc(gfp_t gfp_flags
, void *data
)
99 struct pool_info
*pi
= data
;
100 struct r1bio
*r1_bio
;
105 r1_bio
= r1bio_pool_alloc(gfp_flags
, pi
);
110 * Allocate bios : 1 for reading, n-1 for writing
112 for (j
= pi
->raid_disks
; j
-- ; ) {
113 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
116 r1_bio
->bios
[j
] = bio
;
119 * Allocate RESYNC_PAGES data pages and attach them to
121 * If this is a user-requested check/repair, allocate
122 * RESYNC_PAGES for each bio.
124 if (test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
))
125 need_pages
= pi
->raid_disks
;
128 for (j
= 0; j
< need_pages
; j
++) {
129 bio
= r1_bio
->bios
[j
];
130 bio
->bi_vcnt
= RESYNC_PAGES
;
132 if (bio_alloc_pages(bio
, gfp_flags
))
135 /* If not user-requests, copy the page pointers to all bios */
136 if (!test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
)) {
137 for (i
=0; i
<RESYNC_PAGES
; i
++)
138 for (j
=1; j
<pi
->raid_disks
; j
++)
139 r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
=
140 r1_bio
->bios
[0]->bi_io_vec
[i
].bv_page
;
143 r1_bio
->master_bio
= NULL
;
149 bio_free_pages(r1_bio
->bios
[j
]);
152 while (++j
< pi
->raid_disks
)
153 bio_put(r1_bio
->bios
[j
]);
154 r1bio_pool_free(r1_bio
, data
);
158 static void r1buf_pool_free(void *__r1_bio
, void *data
)
160 struct pool_info
*pi
= data
;
162 struct r1bio
*r1bio
= __r1_bio
;
164 for (i
= 0; i
< RESYNC_PAGES
; i
++)
165 for (j
= pi
->raid_disks
; j
-- ;) {
167 r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
!=
168 r1bio
->bios
[0]->bi_io_vec
[i
].bv_page
)
169 safe_put_page(r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
171 for (i
=0 ; i
< pi
->raid_disks
; i
++)
172 bio_put(r1bio
->bios
[i
]);
174 r1bio_pool_free(r1bio
, data
);
177 static void put_all_bios(struct r1conf
*conf
, struct r1bio
*r1_bio
)
181 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
182 struct bio
**bio
= r1_bio
->bios
+ i
;
183 if (!BIO_SPECIAL(*bio
))
189 static void free_r1bio(struct r1bio
*r1_bio
)
191 struct r1conf
*conf
= r1_bio
->mddev
->private;
193 put_all_bios(conf
, r1_bio
);
194 mempool_free(r1_bio
, conf
->r1bio_pool
);
197 static void put_buf(struct r1bio
*r1_bio
)
199 struct r1conf
*conf
= r1_bio
->mddev
->private;
202 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
203 struct bio
*bio
= r1_bio
->bios
[i
];
205 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
208 mempool_free(r1_bio
, conf
->r1buf_pool
);
213 static void reschedule_retry(struct r1bio
*r1_bio
)
216 struct mddev
*mddev
= r1_bio
->mddev
;
217 struct r1conf
*conf
= mddev
->private;
219 spin_lock_irqsave(&conf
->device_lock
, flags
);
220 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
222 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
224 wake_up(&conf
->wait_barrier
);
225 md_wakeup_thread(mddev
->thread
);
229 * raid_end_bio_io() is called when we have finished servicing a mirrored
230 * operation and are ready to return a success/failure code to the buffer
233 static void call_bio_endio(struct r1bio
*r1_bio
)
235 struct bio
*bio
= r1_bio
->master_bio
;
237 struct r1conf
*conf
= r1_bio
->mddev
->private;
238 sector_t start_next_window
= r1_bio
->start_next_window
;
239 sector_t bi_sector
= bio
->bi_iter
.bi_sector
;
241 if (bio
->bi_phys_segments
) {
243 spin_lock_irqsave(&conf
->device_lock
, flags
);
244 bio
->bi_phys_segments
--;
245 done
= (bio
->bi_phys_segments
== 0);
246 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
248 * make_request() might be waiting for
249 * bi_phys_segments to decrease
251 wake_up(&conf
->wait_barrier
);
255 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
256 bio
->bi_error
= -EIO
;
261 * Wake up any possible resync thread that waits for the device
264 allow_barrier(conf
, start_next_window
, bi_sector
);
268 static void raid_end_bio_io(struct r1bio
*r1_bio
)
270 struct bio
*bio
= r1_bio
->master_bio
;
272 /* if nobody has done the final endio yet, do it now */
273 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
274 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
275 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
276 (unsigned long long) bio
->bi_iter
.bi_sector
,
277 (unsigned long long) bio_end_sector(bio
) - 1);
279 call_bio_endio(r1_bio
);
285 * Update disk head position estimator based on IRQ completion info.
287 static inline void update_head_pos(int disk
, struct r1bio
*r1_bio
)
289 struct r1conf
*conf
= r1_bio
->mddev
->private;
291 conf
->mirrors
[disk
].head_position
=
292 r1_bio
->sector
+ (r1_bio
->sectors
);
296 * Find the disk number which triggered given bio
298 static int find_bio_disk(struct r1bio
*r1_bio
, struct bio
*bio
)
301 struct r1conf
*conf
= r1_bio
->mddev
->private;
302 int raid_disks
= conf
->raid_disks
;
304 for (mirror
= 0; mirror
< raid_disks
* 2; mirror
++)
305 if (r1_bio
->bios
[mirror
] == bio
)
308 BUG_ON(mirror
== raid_disks
* 2);
309 update_head_pos(mirror
, r1_bio
);
314 static void raid1_end_read_request(struct bio
*bio
)
316 int uptodate
= !bio
->bi_error
;
317 struct r1bio
*r1_bio
= bio
->bi_private
;
318 struct r1conf
*conf
= r1_bio
->mddev
->private;
319 struct md_rdev
*rdev
= conf
->mirrors
[r1_bio
->read_disk
].rdev
;
322 * this branch is our 'one mirror IO has finished' event handler:
324 update_head_pos(r1_bio
->read_disk
, r1_bio
);
327 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
329 /* If all other devices have failed, we want to return
330 * the error upwards rather than fail the last device.
331 * Here we redefine "uptodate" to mean "Don't want to retry"
334 spin_lock_irqsave(&conf
->device_lock
, flags
);
335 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
336 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
337 test_bit(In_sync
, &rdev
->flags
)))
339 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
343 raid_end_bio_io(r1_bio
);
344 rdev_dec_pending(rdev
, conf
->mddev
);
349 char b
[BDEVNAME_SIZE
];
351 KERN_ERR
"md/raid1:%s: %s: "
352 "rescheduling sector %llu\n",
356 (unsigned long long)r1_bio
->sector
);
357 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
358 reschedule_retry(r1_bio
);
359 /* don't drop the reference on read_disk yet */
363 static void close_write(struct r1bio
*r1_bio
)
365 /* it really is the end of this request */
366 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
367 /* free extra copy of the data pages */
368 int i
= r1_bio
->behind_page_count
;
370 safe_put_page(r1_bio
->behind_bvecs
[i
].bv_page
);
371 kfree(r1_bio
->behind_bvecs
);
372 r1_bio
->behind_bvecs
= NULL
;
374 /* clear the bitmap if all writes complete successfully */
375 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
377 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
378 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
379 md_write_end(r1_bio
->mddev
);
382 static void r1_bio_write_done(struct r1bio
*r1_bio
)
384 if (!atomic_dec_and_test(&r1_bio
->remaining
))
387 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
388 reschedule_retry(r1_bio
);
391 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
))
392 reschedule_retry(r1_bio
);
394 raid_end_bio_io(r1_bio
);
398 static void raid1_end_write_request(struct bio
*bio
)
400 struct r1bio
*r1_bio
= bio
->bi_private
;
401 int behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
402 struct r1conf
*conf
= r1_bio
->mddev
->private;
403 struct bio
*to_put
= NULL
;
404 int mirror
= find_bio_disk(r1_bio
, bio
);
405 struct md_rdev
*rdev
= conf
->mirrors
[mirror
].rdev
;
408 discard_error
= bio
->bi_error
&& bio_op(bio
) == REQ_OP_DISCARD
;
411 * 'one mirror IO has finished' event handler:
413 if (bio
->bi_error
&& !discard_error
) {
414 set_bit(WriteErrorSeen
, &rdev
->flags
);
415 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
416 set_bit(MD_RECOVERY_NEEDED
, &
417 conf
->mddev
->recovery
);
419 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
422 * Set R1BIO_Uptodate in our master bio, so that we
423 * will return a good error code for to the higher
424 * levels even if IO on some other mirrored buffer
427 * The 'master' represents the composite IO operation
428 * to user-side. So if something waits for IO, then it
429 * will wait for the 'master' bio.
434 r1_bio
->bios
[mirror
] = NULL
;
437 * Do not set R1BIO_Uptodate if the current device is
438 * rebuilding or Faulty. This is because we cannot use
439 * such device for properly reading the data back (we could
440 * potentially use it, if the current write would have felt
441 * before rdev->recovery_offset, but for simplicity we don't
444 if (test_bit(In_sync
, &rdev
->flags
) &&
445 !test_bit(Faulty
, &rdev
->flags
))
446 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
448 /* Maybe we can clear some bad blocks. */
449 if (is_badblock(rdev
, r1_bio
->sector
, r1_bio
->sectors
,
450 &first_bad
, &bad_sectors
) && !discard_error
) {
451 r1_bio
->bios
[mirror
] = IO_MADE_GOOD
;
452 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
457 if (test_bit(WriteMostly
, &rdev
->flags
))
458 atomic_dec(&r1_bio
->behind_remaining
);
461 * In behind mode, we ACK the master bio once the I/O
462 * has safely reached all non-writemostly
463 * disks. Setting the Returned bit ensures that this
464 * gets done only once -- we don't ever want to return
465 * -EIO here, instead we'll wait
467 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
468 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
469 /* Maybe we can return now */
470 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
471 struct bio
*mbio
= r1_bio
->master_bio
;
472 pr_debug("raid1: behind end write sectors"
474 (unsigned long long) mbio
->bi_iter
.bi_sector
,
475 (unsigned long long) bio_end_sector(mbio
) - 1);
476 call_bio_endio(r1_bio
);
480 if (r1_bio
->bios
[mirror
] == NULL
)
481 rdev_dec_pending(rdev
, conf
->mddev
);
484 * Let's see if all mirrored write operations have finished
487 r1_bio_write_done(r1_bio
);
494 * This routine returns the disk from which the requested read should
495 * be done. There is a per-array 'next expected sequential IO' sector
496 * number - if this matches on the next IO then we use the last disk.
497 * There is also a per-disk 'last know head position' sector that is
498 * maintained from IRQ contexts, both the normal and the resync IO
499 * completion handlers update this position correctly. If there is no
500 * perfect sequential match then we pick the disk whose head is closest.
502 * If there are 2 mirrors in the same 2 devices, performance degrades
503 * because position is mirror, not device based.
505 * The rdev for the device selected will have nr_pending incremented.
507 static int read_balance(struct r1conf
*conf
, struct r1bio
*r1_bio
, int *max_sectors
)
509 const sector_t this_sector
= r1_bio
->sector
;
511 int best_good_sectors
;
512 int best_disk
, best_dist_disk
, best_pending_disk
;
516 unsigned int min_pending
;
517 struct md_rdev
*rdev
;
519 int choose_next_idle
;
523 * Check if we can balance. We can balance on the whole
524 * device if no resync is going on, or below the resync window.
525 * We take the first readable disk when above the resync window.
528 sectors
= r1_bio
->sectors
;
531 best_dist
= MaxSector
;
532 best_pending_disk
= -1;
533 min_pending
= UINT_MAX
;
534 best_good_sectors
= 0;
536 choose_next_idle
= 0;
538 if ((conf
->mddev
->recovery_cp
< this_sector
+ sectors
) ||
539 (mddev_is_clustered(conf
->mddev
) &&
540 md_cluster_ops
->area_resyncing(conf
->mddev
, READ
, this_sector
,
541 this_sector
+ sectors
)))
546 for (disk
= 0 ; disk
< conf
->raid_disks
* 2 ; disk
++) {
550 unsigned int pending
;
553 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
554 if (r1_bio
->bios
[disk
] == IO_BLOCKED
556 || test_bit(Faulty
, &rdev
->flags
))
558 if (!test_bit(In_sync
, &rdev
->flags
) &&
559 rdev
->recovery_offset
< this_sector
+ sectors
)
561 if (test_bit(WriteMostly
, &rdev
->flags
)) {
562 /* Don't balance among write-mostly, just
563 * use the first as a last resort */
564 if (best_dist_disk
< 0) {
565 if (is_badblock(rdev
, this_sector
, sectors
,
566 &first_bad
, &bad_sectors
)) {
567 if (first_bad
<= this_sector
)
568 /* Cannot use this */
570 best_good_sectors
= first_bad
- this_sector
;
572 best_good_sectors
= sectors
;
573 best_dist_disk
= disk
;
574 best_pending_disk
= disk
;
578 /* This is a reasonable device to use. It might
581 if (is_badblock(rdev
, this_sector
, sectors
,
582 &first_bad
, &bad_sectors
)) {
583 if (best_dist
< MaxSector
)
584 /* already have a better device */
586 if (first_bad
<= this_sector
) {
587 /* cannot read here. If this is the 'primary'
588 * device, then we must not read beyond
589 * bad_sectors from another device..
591 bad_sectors
-= (this_sector
- first_bad
);
592 if (choose_first
&& sectors
> bad_sectors
)
593 sectors
= bad_sectors
;
594 if (best_good_sectors
> sectors
)
595 best_good_sectors
= sectors
;
598 sector_t good_sectors
= first_bad
- this_sector
;
599 if (good_sectors
> best_good_sectors
) {
600 best_good_sectors
= good_sectors
;
608 best_good_sectors
= sectors
;
610 nonrot
= blk_queue_nonrot(bdev_get_queue(rdev
->bdev
));
611 has_nonrot_disk
|= nonrot
;
612 pending
= atomic_read(&rdev
->nr_pending
);
613 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
618 /* Don't change to another disk for sequential reads */
619 if (conf
->mirrors
[disk
].next_seq_sect
== this_sector
621 int opt_iosize
= bdev_io_opt(rdev
->bdev
) >> 9;
622 struct raid1_info
*mirror
= &conf
->mirrors
[disk
];
626 * If buffered sequential IO size exceeds optimal
627 * iosize, check if there is idle disk. If yes, choose
628 * the idle disk. read_balance could already choose an
629 * idle disk before noticing it's a sequential IO in
630 * this disk. This doesn't matter because this disk
631 * will idle, next time it will be utilized after the
632 * first disk has IO size exceeds optimal iosize. In
633 * this way, iosize of the first disk will be optimal
634 * iosize at least. iosize of the second disk might be
635 * small, but not a big deal since when the second disk
636 * starts IO, the first disk is likely still busy.
638 if (nonrot
&& opt_iosize
> 0 &&
639 mirror
->seq_start
!= MaxSector
&&
640 mirror
->next_seq_sect
> opt_iosize
&&
641 mirror
->next_seq_sect
- opt_iosize
>=
643 choose_next_idle
= 1;
648 /* If device is idle, use it */
654 if (choose_next_idle
)
657 if (min_pending
> pending
) {
658 min_pending
= pending
;
659 best_pending_disk
= disk
;
662 if (dist
< best_dist
) {
664 best_dist_disk
= disk
;
669 * If all disks are rotational, choose the closest disk. If any disk is
670 * non-rotational, choose the disk with less pending request even the
671 * disk is rotational, which might/might not be optimal for raids with
672 * mixed ratation/non-rotational disks depending on workload.
674 if (best_disk
== -1) {
676 best_disk
= best_pending_disk
;
678 best_disk
= best_dist_disk
;
681 if (best_disk
>= 0) {
682 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
685 atomic_inc(&rdev
->nr_pending
);
686 sectors
= best_good_sectors
;
688 if (conf
->mirrors
[best_disk
].next_seq_sect
!= this_sector
)
689 conf
->mirrors
[best_disk
].seq_start
= this_sector
;
691 conf
->mirrors
[best_disk
].next_seq_sect
= this_sector
+ sectors
;
694 *max_sectors
= sectors
;
699 static int raid1_congested(struct mddev
*mddev
, int bits
)
701 struct r1conf
*conf
= mddev
->private;
704 if ((bits
& (1 << WB_async_congested
)) &&
705 conf
->pending_count
>= max_queued_requests
)
709 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
710 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
711 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
712 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
716 /* Note the '|| 1' - when read_balance prefers
717 * non-congested targets, it can be removed
719 if ((bits
& (1 << WB_async_congested
)) || 1)
720 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
722 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
729 static void flush_pending_writes(struct r1conf
*conf
)
731 /* Any writes that have been queued but are awaiting
732 * bitmap updates get flushed here.
734 spin_lock_irq(&conf
->device_lock
);
736 if (conf
->pending_bio_list
.head
) {
738 bio
= bio_list_get(&conf
->pending_bio_list
);
739 conf
->pending_count
= 0;
740 spin_unlock_irq(&conf
->device_lock
);
741 /* flush any pending bitmap writes to
742 * disk before proceeding w/ I/O */
743 bitmap_unplug(conf
->mddev
->bitmap
);
744 wake_up(&conf
->wait_barrier
);
746 while (bio
) { /* submit pending writes */
747 struct bio
*next
= bio
->bi_next
;
749 if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
750 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
754 generic_make_request(bio
);
758 spin_unlock_irq(&conf
->device_lock
);
762 * Sometimes we need to suspend IO while we do something else,
763 * either some resync/recovery, or reconfigure the array.
764 * To do this we raise a 'barrier'.
765 * The 'barrier' is a counter that can be raised multiple times
766 * to count how many activities are happening which preclude
768 * We can only raise the barrier if there is no pending IO.
769 * i.e. if nr_pending == 0.
770 * We choose only to raise the barrier if no-one is waiting for the
771 * barrier to go down. This means that as soon as an IO request
772 * is ready, no other operations which require a barrier will start
773 * until the IO request has had a chance.
775 * So: regular IO calls 'wait_barrier'. When that returns there
776 * is no backgroup IO happening, It must arrange to call
777 * allow_barrier when it has finished its IO.
778 * backgroup IO calls must call raise_barrier. Once that returns
779 * there is no normal IO happeing. It must arrange to call
780 * lower_barrier when the particular background IO completes.
782 static void raise_barrier(struct r1conf
*conf
, sector_t sector_nr
)
784 spin_lock_irq(&conf
->resync_lock
);
786 /* Wait until no block IO is waiting */
787 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
790 /* block any new IO from starting */
792 conf
->next_resync
= sector_nr
;
794 /* For these conditions we must wait:
795 * A: while the array is in frozen state
796 * B: while barrier >= RESYNC_DEPTH, meaning resync reach
797 * the max count which allowed.
798 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
799 * next resync will reach to the window which normal bios are
801 * D: while there are any active requests in the current window.
803 wait_event_lock_irq(conf
->wait_barrier
,
804 !conf
->array_frozen
&&
805 conf
->barrier
< RESYNC_DEPTH
&&
806 conf
->current_window_requests
== 0 &&
807 (conf
->start_next_window
>=
808 conf
->next_resync
+ RESYNC_SECTORS
),
812 spin_unlock_irq(&conf
->resync_lock
);
815 static void lower_barrier(struct r1conf
*conf
)
818 BUG_ON(conf
->barrier
<= 0);
819 spin_lock_irqsave(&conf
->resync_lock
, flags
);
822 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
823 wake_up(&conf
->wait_barrier
);
826 static bool need_to_wait_for_sync(struct r1conf
*conf
, struct bio
*bio
)
830 if (conf
->array_frozen
|| !bio
)
832 else if (conf
->barrier
&& bio_data_dir(bio
) == WRITE
) {
833 if ((conf
->mddev
->curr_resync_completed
834 >= bio_end_sector(bio
)) ||
835 (conf
->next_resync
+ NEXT_NORMALIO_DISTANCE
836 <= bio
->bi_iter
.bi_sector
))
845 static sector_t
wait_barrier(struct r1conf
*conf
, struct bio
*bio
)
849 spin_lock_irq(&conf
->resync_lock
);
850 if (need_to_wait_for_sync(conf
, bio
)) {
852 /* Wait for the barrier to drop.
853 * However if there are already pending
854 * requests (preventing the barrier from
855 * rising completely), and the
856 * per-process bio queue isn't empty,
857 * then don't wait, as we need to empty
858 * that queue to allow conf->start_next_window
861 wait_event_lock_irq(conf
->wait_barrier
,
862 !conf
->array_frozen
&&
864 ((conf
->start_next_window
<
865 conf
->next_resync
+ RESYNC_SECTORS
) &&
867 !bio_list_empty(current
->bio_list
))),
872 if (bio
&& bio_data_dir(bio
) == WRITE
) {
873 if (bio
->bi_iter
.bi_sector
>= conf
->next_resync
) {
874 if (conf
->start_next_window
== MaxSector
)
875 conf
->start_next_window
=
877 NEXT_NORMALIO_DISTANCE
;
879 if ((conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
)
880 <= bio
->bi_iter
.bi_sector
)
881 conf
->next_window_requests
++;
883 conf
->current_window_requests
++;
884 sector
= conf
->start_next_window
;
889 spin_unlock_irq(&conf
->resync_lock
);
893 static void allow_barrier(struct r1conf
*conf
, sector_t start_next_window
,
898 spin_lock_irqsave(&conf
->resync_lock
, flags
);
900 if (start_next_window
) {
901 if (start_next_window
== conf
->start_next_window
) {
902 if (conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
904 conf
->next_window_requests
--;
906 conf
->current_window_requests
--;
908 conf
->current_window_requests
--;
910 if (!conf
->current_window_requests
) {
911 if (conf
->next_window_requests
) {
912 conf
->current_window_requests
=
913 conf
->next_window_requests
;
914 conf
->next_window_requests
= 0;
915 conf
->start_next_window
+=
916 NEXT_NORMALIO_DISTANCE
;
918 conf
->start_next_window
= MaxSector
;
921 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
922 wake_up(&conf
->wait_barrier
);
925 static void freeze_array(struct r1conf
*conf
, int extra
)
927 /* stop syncio and normal IO and wait for everything to
929 * We wait until nr_pending match nr_queued+extra
930 * This is called in the context of one normal IO request
931 * that has failed. Thus any sync request that might be pending
932 * will be blocked by nr_pending, and we need to wait for
933 * pending IO requests to complete or be queued for re-try.
934 * Thus the number queued (nr_queued) plus this request (extra)
935 * must match the number of pending IOs (nr_pending) before
938 spin_lock_irq(&conf
->resync_lock
);
939 conf
->array_frozen
= 1;
940 wait_event_lock_irq_cmd(conf
->wait_barrier
,
941 conf
->nr_pending
== conf
->nr_queued
+extra
,
943 flush_pending_writes(conf
));
944 spin_unlock_irq(&conf
->resync_lock
);
946 static void unfreeze_array(struct r1conf
*conf
)
948 /* reverse the effect of the freeze */
949 spin_lock_irq(&conf
->resync_lock
);
950 conf
->array_frozen
= 0;
951 wake_up(&conf
->wait_barrier
);
952 spin_unlock_irq(&conf
->resync_lock
);
955 /* duplicate the data pages for behind I/O
957 static void alloc_behind_pages(struct bio
*bio
, struct r1bio
*r1_bio
)
960 struct bio_vec
*bvec
;
961 struct bio_vec
*bvecs
= kzalloc(bio
->bi_vcnt
* sizeof(struct bio_vec
),
963 if (unlikely(!bvecs
))
966 bio_for_each_segment_all(bvec
, bio
, i
) {
968 bvecs
[i
].bv_page
= alloc_page(GFP_NOIO
);
969 if (unlikely(!bvecs
[i
].bv_page
))
971 memcpy(kmap(bvecs
[i
].bv_page
) + bvec
->bv_offset
,
972 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
973 kunmap(bvecs
[i
].bv_page
);
974 kunmap(bvec
->bv_page
);
976 r1_bio
->behind_bvecs
= bvecs
;
977 r1_bio
->behind_page_count
= bio
->bi_vcnt
;
978 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
982 for (i
= 0; i
< bio
->bi_vcnt
; i
++)
983 if (bvecs
[i
].bv_page
)
984 put_page(bvecs
[i
].bv_page
);
986 pr_debug("%dB behind alloc failed, doing sync I/O\n",
987 bio
->bi_iter
.bi_size
);
990 struct raid1_plug_cb
{
991 struct blk_plug_cb cb
;
992 struct bio_list pending
;
996 static void raid1_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
998 struct raid1_plug_cb
*plug
= container_of(cb
, struct raid1_plug_cb
,
1000 struct mddev
*mddev
= plug
->cb
.data
;
1001 struct r1conf
*conf
= mddev
->private;
1004 if (from_schedule
|| current
->bio_list
) {
1005 spin_lock_irq(&conf
->device_lock
);
1006 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1007 conf
->pending_count
+= plug
->pending_cnt
;
1008 spin_unlock_irq(&conf
->device_lock
);
1009 wake_up(&conf
->wait_barrier
);
1010 md_wakeup_thread(mddev
->thread
);
1015 /* we aren't scheduling, so we can do the write-out directly. */
1016 bio
= bio_list_get(&plug
->pending
);
1017 bitmap_unplug(mddev
->bitmap
);
1018 wake_up(&conf
->wait_barrier
);
1020 while (bio
) { /* submit pending writes */
1021 struct bio
*next
= bio
->bi_next
;
1022 bio
->bi_next
= NULL
;
1023 if (unlikely((bio_op(bio
) == REQ_OP_DISCARD
) &&
1024 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1025 /* Just ignore it */
1028 generic_make_request(bio
);
1034 static void raid1_make_request(struct mddev
*mddev
, struct bio
* bio
)
1036 struct r1conf
*conf
= mddev
->private;
1037 struct raid1_info
*mirror
;
1038 struct r1bio
*r1_bio
;
1039 struct bio
*read_bio
;
1041 struct bitmap
*bitmap
;
1042 unsigned long flags
;
1043 const int op
= bio_op(bio
);
1044 const int rw
= bio_data_dir(bio
);
1045 const unsigned long do_sync
= (bio
->bi_opf
& REQ_SYNC
);
1046 const unsigned long do_flush_fua
= (bio
->bi_opf
&
1047 (REQ_PREFLUSH
| REQ_FUA
));
1048 struct md_rdev
*blocked_rdev
;
1049 struct blk_plug_cb
*cb
;
1050 struct raid1_plug_cb
*plug
= NULL
;
1052 int sectors_handled
;
1054 sector_t start_next_window
;
1057 * Register the new request and wait if the reconstruction
1058 * thread has put up a bar for new requests.
1059 * Continue immediately if no resync is active currently.
1062 md_write_start(mddev
, bio
); /* wait on superblock update early */
1064 if (bio_data_dir(bio
) == WRITE
&&
1065 ((bio_end_sector(bio
) > mddev
->suspend_lo
&&
1066 bio
->bi_iter
.bi_sector
< mddev
->suspend_hi
) ||
1067 (mddev_is_clustered(mddev
) &&
1068 md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1069 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
))))) {
1070 /* As the suspend_* range is controlled by
1071 * userspace, we want an interruptible
1076 flush_signals(current
);
1077 prepare_to_wait(&conf
->wait_barrier
,
1078 &w
, TASK_INTERRUPTIBLE
);
1079 if (bio_end_sector(bio
) <= mddev
->suspend_lo
||
1080 bio
->bi_iter
.bi_sector
>= mddev
->suspend_hi
||
1081 (mddev_is_clustered(mddev
) &&
1082 !md_cluster_ops
->area_resyncing(mddev
, WRITE
,
1083 bio
->bi_iter
.bi_sector
, bio_end_sector(bio
))))
1087 finish_wait(&conf
->wait_barrier
, &w
);
1090 start_next_window
= wait_barrier(conf
, bio
);
1092 bitmap
= mddev
->bitmap
;
1095 * make_request() can abort the operation when read-ahead is being
1096 * used and no empty request is available.
1099 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1101 r1_bio
->master_bio
= bio
;
1102 r1_bio
->sectors
= bio_sectors(bio
);
1104 r1_bio
->mddev
= mddev
;
1105 r1_bio
->sector
= bio
->bi_iter
.bi_sector
;
1107 /* We might need to issue multiple reads to different
1108 * devices if there are bad blocks around, so we keep
1109 * track of the number of reads in bio->bi_phys_segments.
1110 * If this is 0, there is only one r1_bio and no locking
1111 * will be needed when requests complete. If it is
1112 * non-zero, then it is the number of not-completed requests.
1114 bio
->bi_phys_segments
= 0;
1115 bio_clear_flag(bio
, BIO_SEG_VALID
);
1119 * read balancing logic:
1124 rdisk
= read_balance(conf
, r1_bio
, &max_sectors
);
1127 /* couldn't find anywhere to read from */
1128 raid_end_bio_io(r1_bio
);
1131 mirror
= conf
->mirrors
+ rdisk
;
1133 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
1135 /* Reading from a write-mostly device must
1136 * take care not to over-take any writes
1139 wait_event(bitmap
->behind_wait
,
1140 atomic_read(&bitmap
->behind_writes
) == 0);
1142 r1_bio
->read_disk
= rdisk
;
1143 r1_bio
->start_next_window
= 0;
1145 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1146 bio_trim(read_bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
1149 r1_bio
->bios
[rdisk
] = read_bio
;
1151 read_bio
->bi_iter
.bi_sector
= r1_bio
->sector
+
1152 mirror
->rdev
->data_offset
;
1153 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
1154 read_bio
->bi_end_io
= raid1_end_read_request
;
1155 bio_set_op_attrs(read_bio
, op
, do_sync
);
1156 read_bio
->bi_private
= r1_bio
;
1158 if (max_sectors
< r1_bio
->sectors
) {
1159 /* could not read all from this device, so we will
1160 * need another r1_bio.
1163 sectors_handled
= (r1_bio
->sector
+ max_sectors
1164 - bio
->bi_iter
.bi_sector
);
1165 r1_bio
->sectors
= max_sectors
;
1166 spin_lock_irq(&conf
->device_lock
);
1167 if (bio
->bi_phys_segments
== 0)
1168 bio
->bi_phys_segments
= 2;
1170 bio
->bi_phys_segments
++;
1171 spin_unlock_irq(&conf
->device_lock
);
1172 /* Cannot call generic_make_request directly
1173 * as that will be queued in __make_request
1174 * and subsequent mempool_alloc might block waiting
1175 * for it. So hand bio over to raid1d.
1177 reschedule_retry(r1_bio
);
1179 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1181 r1_bio
->master_bio
= bio
;
1182 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1184 r1_bio
->mddev
= mddev
;
1185 r1_bio
->sector
= bio
->bi_iter
.bi_sector
+
1189 generic_make_request(read_bio
);
1196 if (conf
->pending_count
>= max_queued_requests
) {
1197 md_wakeup_thread(mddev
->thread
);
1198 wait_event(conf
->wait_barrier
,
1199 conf
->pending_count
< max_queued_requests
);
1201 /* first select target devices under rcu_lock and
1202 * inc refcount on their rdev. Record them by setting
1204 * If there are known/acknowledged bad blocks on any device on
1205 * which we have seen a write error, we want to avoid writing those
1207 * This potentially requires several writes to write around
1208 * the bad blocks. Each set of writes gets it's own r1bio
1209 * with a set of bios attached.
1212 disks
= conf
->raid_disks
* 2;
1214 r1_bio
->start_next_window
= start_next_window
;
1215 blocked_rdev
= NULL
;
1217 max_sectors
= r1_bio
->sectors
;
1218 for (i
= 0; i
< disks
; i
++) {
1219 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1220 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1221 atomic_inc(&rdev
->nr_pending
);
1222 blocked_rdev
= rdev
;
1225 r1_bio
->bios
[i
] = NULL
;
1226 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
1227 if (i
< conf
->raid_disks
)
1228 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1232 atomic_inc(&rdev
->nr_pending
);
1233 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1238 is_bad
= is_badblock(rdev
, r1_bio
->sector
,
1240 &first_bad
, &bad_sectors
);
1242 /* mustn't write here until the bad block is
1244 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1245 blocked_rdev
= rdev
;
1248 if (is_bad
&& first_bad
<= r1_bio
->sector
) {
1249 /* Cannot write here at all */
1250 bad_sectors
-= (r1_bio
->sector
- first_bad
);
1251 if (bad_sectors
< max_sectors
)
1252 /* mustn't write more than bad_sectors
1253 * to other devices yet
1255 max_sectors
= bad_sectors
;
1256 rdev_dec_pending(rdev
, mddev
);
1257 /* We don't set R1BIO_Degraded as that
1258 * only applies if the disk is
1259 * missing, so it might be re-added,
1260 * and we want to know to recover this
1262 * In this case the device is here,
1263 * and the fact that this chunk is not
1264 * in-sync is recorded in the bad
1270 int good_sectors
= first_bad
- r1_bio
->sector
;
1271 if (good_sectors
< max_sectors
)
1272 max_sectors
= good_sectors
;
1275 r1_bio
->bios
[i
] = bio
;
1279 if (unlikely(blocked_rdev
)) {
1280 /* Wait for this device to become unblocked */
1282 sector_t old
= start_next_window
;
1284 for (j
= 0; j
< i
; j
++)
1285 if (r1_bio
->bios
[j
])
1286 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
1288 allow_barrier(conf
, start_next_window
, bio
->bi_iter
.bi_sector
);
1289 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1290 start_next_window
= wait_barrier(conf
, bio
);
1292 * We must make sure the multi r1bios of bio have
1293 * the same value of bi_phys_segments
1295 if (bio
->bi_phys_segments
&& old
&&
1296 old
!= start_next_window
)
1297 /* Wait for the former r1bio(s) to complete */
1298 wait_event(conf
->wait_barrier
,
1299 bio
->bi_phys_segments
== 1);
1303 if (max_sectors
< r1_bio
->sectors
) {
1304 /* We are splitting this write into multiple parts, so
1305 * we need to prepare for allocating another r1_bio.
1307 r1_bio
->sectors
= max_sectors
;
1308 spin_lock_irq(&conf
->device_lock
);
1309 if (bio
->bi_phys_segments
== 0)
1310 bio
->bi_phys_segments
= 2;
1312 bio
->bi_phys_segments
++;
1313 spin_unlock_irq(&conf
->device_lock
);
1315 sectors_handled
= r1_bio
->sector
+ max_sectors
- bio
->bi_iter
.bi_sector
;
1317 atomic_set(&r1_bio
->remaining
, 1);
1318 atomic_set(&r1_bio
->behind_remaining
, 0);
1321 for (i
= 0; i
< disks
; i
++) {
1323 if (!r1_bio
->bios
[i
])
1326 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1327 bio_trim(mbio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
1331 * Not if there are too many, or cannot
1332 * allocate memory, or a reader on WriteMostly
1333 * is waiting for behind writes to flush */
1335 (atomic_read(&bitmap
->behind_writes
)
1336 < mddev
->bitmap_info
.max_write_behind
) &&
1337 !waitqueue_active(&bitmap
->behind_wait
))
1338 alloc_behind_pages(mbio
, r1_bio
);
1340 bitmap_startwrite(bitmap
, r1_bio
->sector
,
1342 test_bit(R1BIO_BehindIO
,
1346 if (r1_bio
->behind_bvecs
) {
1347 struct bio_vec
*bvec
;
1351 * We trimmed the bio, so _all is legit
1353 bio_for_each_segment_all(bvec
, mbio
, j
)
1354 bvec
->bv_page
= r1_bio
->behind_bvecs
[j
].bv_page
;
1355 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
1356 atomic_inc(&r1_bio
->behind_remaining
);
1359 r1_bio
->bios
[i
] = mbio
;
1361 mbio
->bi_iter
.bi_sector
= (r1_bio
->sector
+
1362 conf
->mirrors
[i
].rdev
->data_offset
);
1363 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1364 mbio
->bi_end_io
= raid1_end_write_request
;
1365 bio_set_op_attrs(mbio
, op
, do_flush_fua
| do_sync
);
1366 mbio
->bi_private
= r1_bio
;
1368 atomic_inc(&r1_bio
->remaining
);
1370 cb
= blk_check_plugged(raid1_unplug
, mddev
, sizeof(*plug
));
1372 plug
= container_of(cb
, struct raid1_plug_cb
, cb
);
1375 spin_lock_irqsave(&conf
->device_lock
, flags
);
1377 bio_list_add(&plug
->pending
, mbio
);
1378 plug
->pending_cnt
++;
1380 bio_list_add(&conf
->pending_bio_list
, mbio
);
1381 conf
->pending_count
++;
1383 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1385 md_wakeup_thread(mddev
->thread
);
1387 /* Mustn't call r1_bio_write_done before this next test,
1388 * as it could result in the bio being freed.
1390 if (sectors_handled
< bio_sectors(bio
)) {
1391 r1_bio_write_done(r1_bio
);
1392 /* We need another r1_bio. It has already been counted
1393 * in bio->bi_phys_segments
1395 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1396 r1_bio
->master_bio
= bio
;
1397 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1399 r1_bio
->mddev
= mddev
;
1400 r1_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1404 r1_bio_write_done(r1_bio
);
1406 /* In case raid1d snuck in to freeze_array */
1407 wake_up(&conf
->wait_barrier
);
1410 static void raid1_status(struct seq_file
*seq
, struct mddev
*mddev
)
1412 struct r1conf
*conf
= mddev
->private;
1415 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1416 conf
->raid_disks
- mddev
->degraded
);
1418 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1419 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1420 seq_printf(seq
, "%s",
1421 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1424 seq_printf(seq
, "]");
1427 static void raid1_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1429 char b
[BDEVNAME_SIZE
];
1430 struct r1conf
*conf
= mddev
->private;
1431 unsigned long flags
;
1434 * If it is not operational, then we have already marked it as dead
1435 * else if it is the last working disks, ignore the error, let the
1436 * next level up know.
1437 * else mark the drive as failed
1439 if (test_bit(In_sync
, &rdev
->flags
)
1440 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1442 * Don't fail the drive, act as though we were just a
1443 * normal single drive.
1444 * However don't try a recovery from this drive as
1445 * it is very likely to fail.
1447 conf
->recovery_disabled
= mddev
->recovery_disabled
;
1450 set_bit(Blocked
, &rdev
->flags
);
1451 spin_lock_irqsave(&conf
->device_lock
, flags
);
1452 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1454 set_bit(Faulty
, &rdev
->flags
);
1456 set_bit(Faulty
, &rdev
->flags
);
1457 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1459 * if recovery is running, make sure it aborts.
1461 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1462 set_mask_bits(&mddev
->flags
, 0,
1463 BIT(MD_CHANGE_DEVS
) | BIT(MD_CHANGE_PENDING
));
1465 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1466 "md/raid1:%s: Operation continuing on %d devices.\n",
1467 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1468 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1471 static void print_conf(struct r1conf
*conf
)
1475 printk(KERN_DEBUG
"RAID1 conf printout:\n");
1477 printk(KERN_DEBUG
"(!conf)\n");
1480 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1484 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1485 char b
[BDEVNAME_SIZE
];
1486 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1488 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1489 i
, !test_bit(In_sync
, &rdev
->flags
),
1490 !test_bit(Faulty
, &rdev
->flags
),
1491 bdevname(rdev
->bdev
,b
));
1496 static void close_sync(struct r1conf
*conf
)
1498 wait_barrier(conf
, NULL
);
1499 allow_barrier(conf
, 0, 0);
1501 mempool_destroy(conf
->r1buf_pool
);
1502 conf
->r1buf_pool
= NULL
;
1504 spin_lock_irq(&conf
->resync_lock
);
1505 conf
->next_resync
= MaxSector
- 2 * NEXT_NORMALIO_DISTANCE
;
1506 conf
->start_next_window
= MaxSector
;
1507 conf
->current_window_requests
+=
1508 conf
->next_window_requests
;
1509 conf
->next_window_requests
= 0;
1510 spin_unlock_irq(&conf
->resync_lock
);
1513 static int raid1_spare_active(struct mddev
*mddev
)
1516 struct r1conf
*conf
= mddev
->private;
1518 unsigned long flags
;
1521 * Find all failed disks within the RAID1 configuration
1522 * and mark them readable.
1523 * Called under mddev lock, so rcu protection not needed.
1524 * device_lock used to avoid races with raid1_end_read_request
1525 * which expects 'In_sync' flags and ->degraded to be consistent.
1527 spin_lock_irqsave(&conf
->device_lock
, flags
);
1528 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1529 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1530 struct md_rdev
*repl
= conf
->mirrors
[conf
->raid_disks
+ i
].rdev
;
1532 && !test_bit(Candidate
, &repl
->flags
)
1533 && repl
->recovery_offset
== MaxSector
1534 && !test_bit(Faulty
, &repl
->flags
)
1535 && !test_and_set_bit(In_sync
, &repl
->flags
)) {
1536 /* replacement has just become active */
1538 !test_and_clear_bit(In_sync
, &rdev
->flags
))
1541 /* Replaced device not technically
1542 * faulty, but we need to be sure
1543 * it gets removed and never re-added
1545 set_bit(Faulty
, &rdev
->flags
);
1546 sysfs_notify_dirent_safe(
1551 && rdev
->recovery_offset
== MaxSector
1552 && !test_bit(Faulty
, &rdev
->flags
)
1553 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1555 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
1558 mddev
->degraded
-= count
;
1559 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1565 static int raid1_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1567 struct r1conf
*conf
= mddev
->private;
1570 struct raid1_info
*p
;
1572 int last
= conf
->raid_disks
- 1;
1574 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
1577 if (md_integrity_add_rdev(rdev
, mddev
))
1580 if (rdev
->raid_disk
>= 0)
1581 first
= last
= rdev
->raid_disk
;
1584 * find the disk ... but prefer rdev->saved_raid_disk
1587 if (rdev
->saved_raid_disk
>= 0 &&
1588 rdev
->saved_raid_disk
>= first
&&
1589 conf
->mirrors
[rdev
->saved_raid_disk
].rdev
== NULL
)
1590 first
= last
= rdev
->saved_raid_disk
;
1592 for (mirror
= first
; mirror
<= last
; mirror
++) {
1593 p
= conf
->mirrors
+mirror
;
1597 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1598 rdev
->data_offset
<< 9);
1600 p
->head_position
= 0;
1601 rdev
->raid_disk
= mirror
;
1603 /* As all devices are equivalent, we don't need a full recovery
1604 * if this was recently any drive of the array
1606 if (rdev
->saved_raid_disk
< 0)
1608 rcu_assign_pointer(p
->rdev
, rdev
);
1611 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
1612 p
[conf
->raid_disks
].rdev
== NULL
) {
1613 /* Add this device as a replacement */
1614 clear_bit(In_sync
, &rdev
->flags
);
1615 set_bit(Replacement
, &rdev
->flags
);
1616 rdev
->raid_disk
= mirror
;
1619 rcu_assign_pointer(p
[conf
->raid_disks
].rdev
, rdev
);
1623 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1624 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1629 static int raid1_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1631 struct r1conf
*conf
= mddev
->private;
1633 int number
= rdev
->raid_disk
;
1634 struct raid1_info
*p
= conf
->mirrors
+ number
;
1636 if (rdev
!= p
->rdev
)
1637 p
= conf
->mirrors
+ conf
->raid_disks
+ number
;
1640 if (rdev
== p
->rdev
) {
1641 if (test_bit(In_sync
, &rdev
->flags
) ||
1642 atomic_read(&rdev
->nr_pending
)) {
1646 /* Only remove non-faulty devices if recovery
1649 if (!test_bit(Faulty
, &rdev
->flags
) &&
1650 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
1651 mddev
->degraded
< conf
->raid_disks
) {
1656 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
1658 if (atomic_read(&rdev
->nr_pending
)) {
1659 /* lost the race, try later */
1665 if (conf
->mirrors
[conf
->raid_disks
+ number
].rdev
) {
1666 /* We just removed a device that is being replaced.
1667 * Move down the replacement. We drain all IO before
1668 * doing this to avoid confusion.
1670 struct md_rdev
*repl
=
1671 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
;
1672 freeze_array(conf
, 0);
1673 clear_bit(Replacement
, &repl
->flags
);
1675 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
= NULL
;
1676 unfreeze_array(conf
);
1677 clear_bit(WantReplacement
, &rdev
->flags
);
1679 clear_bit(WantReplacement
, &rdev
->flags
);
1680 err
= md_integrity_register(mddev
);
1688 static void end_sync_read(struct bio
*bio
)
1690 struct r1bio
*r1_bio
= bio
->bi_private
;
1692 update_head_pos(r1_bio
->read_disk
, r1_bio
);
1695 * we have read a block, now it needs to be re-written,
1696 * or re-read if the read failed.
1697 * We don't do much here, just schedule handling by raid1d
1700 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1702 if (atomic_dec_and_test(&r1_bio
->remaining
))
1703 reschedule_retry(r1_bio
);
1706 static void end_sync_write(struct bio
*bio
)
1708 int uptodate
= !bio
->bi_error
;
1709 struct r1bio
*r1_bio
= bio
->bi_private
;
1710 struct mddev
*mddev
= r1_bio
->mddev
;
1711 struct r1conf
*conf
= mddev
->private;
1714 struct md_rdev
*rdev
= conf
->mirrors
[find_bio_disk(r1_bio
, bio
)].rdev
;
1717 sector_t sync_blocks
= 0;
1718 sector_t s
= r1_bio
->sector
;
1719 long sectors_to_go
= r1_bio
->sectors
;
1720 /* make sure these bits doesn't get cleared. */
1722 bitmap_end_sync(mddev
->bitmap
, s
,
1725 sectors_to_go
-= sync_blocks
;
1726 } while (sectors_to_go
> 0);
1727 set_bit(WriteErrorSeen
, &rdev
->flags
);
1728 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1729 set_bit(MD_RECOVERY_NEEDED
, &
1731 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
1732 } else if (is_badblock(rdev
, r1_bio
->sector
, r1_bio
->sectors
,
1733 &first_bad
, &bad_sectors
) &&
1734 !is_badblock(conf
->mirrors
[r1_bio
->read_disk
].rdev
,
1737 &first_bad
, &bad_sectors
)
1739 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
1741 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1742 int s
= r1_bio
->sectors
;
1743 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
1744 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1745 reschedule_retry(r1_bio
);
1748 md_done_sync(mddev
, s
, uptodate
);
1753 static int r1_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1754 int sectors
, struct page
*page
, int rw
)
1756 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, 0, false))
1760 set_bit(WriteErrorSeen
, &rdev
->flags
);
1761 if (!test_and_set_bit(WantReplacement
,
1763 set_bit(MD_RECOVERY_NEEDED
, &
1764 rdev
->mddev
->recovery
);
1766 /* need to record an error - either for the block or the device */
1767 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1768 md_error(rdev
->mddev
, rdev
);
1772 static int fix_sync_read_error(struct r1bio
*r1_bio
)
1774 /* Try some synchronous reads of other devices to get
1775 * good data, much like with normal read errors. Only
1776 * read into the pages we already have so we don't
1777 * need to re-issue the read request.
1778 * We don't need to freeze the array, because being in an
1779 * active sync request, there is no normal IO, and
1780 * no overlapping syncs.
1781 * We don't need to check is_badblock() again as we
1782 * made sure that anything with a bad block in range
1783 * will have bi_end_io clear.
1785 struct mddev
*mddev
= r1_bio
->mddev
;
1786 struct r1conf
*conf
= mddev
->private;
1787 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1788 sector_t sect
= r1_bio
->sector
;
1789 int sectors
= r1_bio
->sectors
;
1794 int d
= r1_bio
->read_disk
;
1796 struct md_rdev
*rdev
;
1799 if (s
> (PAGE_SIZE
>>9))
1802 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1803 /* No rcu protection needed here devices
1804 * can only be removed when no resync is
1805 * active, and resync is currently active
1807 rdev
= conf
->mirrors
[d
].rdev
;
1808 if (sync_page_io(rdev
, sect
, s
<<9,
1809 bio
->bi_io_vec
[idx
].bv_page
,
1810 REQ_OP_READ
, 0, false)) {
1816 if (d
== conf
->raid_disks
* 2)
1818 } while (!success
&& d
!= r1_bio
->read_disk
);
1821 char b
[BDEVNAME_SIZE
];
1823 /* Cannot read from anywhere, this block is lost.
1824 * Record a bad block on each device. If that doesn't
1825 * work just disable and interrupt the recovery.
1826 * Don't fail devices as that won't really help.
1828 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O read error"
1829 " for block %llu\n",
1831 bdevname(bio
->bi_bdev
, b
),
1832 (unsigned long long)r1_bio
->sector
);
1833 for (d
= 0; d
< conf
->raid_disks
* 2; d
++) {
1834 rdev
= conf
->mirrors
[d
].rdev
;
1835 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
1837 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
1841 conf
->recovery_disabled
=
1842 mddev
->recovery_disabled
;
1843 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1844 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1856 /* write it back and re-read */
1857 while (d
!= r1_bio
->read_disk
) {
1859 d
= conf
->raid_disks
* 2;
1861 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1863 rdev
= conf
->mirrors
[d
].rdev
;
1864 if (r1_sync_page_io(rdev
, sect
, s
,
1865 bio
->bi_io_vec
[idx
].bv_page
,
1867 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
1868 rdev_dec_pending(rdev
, mddev
);
1872 while (d
!= r1_bio
->read_disk
) {
1874 d
= conf
->raid_disks
* 2;
1876 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1878 rdev
= conf
->mirrors
[d
].rdev
;
1879 if (r1_sync_page_io(rdev
, sect
, s
,
1880 bio
->bi_io_vec
[idx
].bv_page
,
1882 atomic_add(s
, &rdev
->corrected_errors
);
1888 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1893 static void process_checks(struct r1bio
*r1_bio
)
1895 /* We have read all readable devices. If we haven't
1896 * got the block, then there is no hope left.
1897 * If we have, then we want to do a comparison
1898 * and skip the write if everything is the same.
1899 * If any blocks failed to read, then we need to
1900 * attempt an over-write
1902 struct mddev
*mddev
= r1_bio
->mddev
;
1903 struct r1conf
*conf
= mddev
->private;
1908 /* Fix variable parts of all bios */
1909 vcnt
= (r1_bio
->sectors
+ PAGE_SIZE
/ 512 - 1) >> (PAGE_SHIFT
- 9);
1910 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
1914 struct bio
*b
= r1_bio
->bios
[i
];
1915 if (b
->bi_end_io
!= end_sync_read
)
1917 /* fixup the bio for reuse, but preserve errno */
1918 error
= b
->bi_error
;
1920 b
->bi_error
= error
;
1922 b
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
1923 b
->bi_iter
.bi_sector
= r1_bio
->sector
+
1924 conf
->mirrors
[i
].rdev
->data_offset
;
1925 b
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1926 b
->bi_end_io
= end_sync_read
;
1927 b
->bi_private
= r1_bio
;
1929 size
= b
->bi_iter
.bi_size
;
1930 for (j
= 0; j
< vcnt
; j
++) {
1932 bi
= &b
->bi_io_vec
[j
];
1934 if (size
> PAGE_SIZE
)
1935 bi
->bv_len
= PAGE_SIZE
;
1941 for (primary
= 0; primary
< conf
->raid_disks
* 2; primary
++)
1942 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
1943 !r1_bio
->bios
[primary
]->bi_error
) {
1944 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
1945 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
1948 r1_bio
->read_disk
= primary
;
1949 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
1951 struct bio
*pbio
= r1_bio
->bios
[primary
];
1952 struct bio
*sbio
= r1_bio
->bios
[i
];
1953 int error
= sbio
->bi_error
;
1955 if (sbio
->bi_end_io
!= end_sync_read
)
1957 /* Now we can 'fixup' the error value */
1961 for (j
= vcnt
; j
-- ; ) {
1963 p
= pbio
->bi_io_vec
[j
].bv_page
;
1964 s
= sbio
->bi_io_vec
[j
].bv_page
;
1965 if (memcmp(page_address(p
),
1967 sbio
->bi_io_vec
[j
].bv_len
))
1973 atomic64_add(r1_bio
->sectors
, &mddev
->resync_mismatches
);
1974 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
1976 /* No need to write to this device. */
1977 sbio
->bi_end_io
= NULL
;
1978 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
1982 bio_copy_data(sbio
, pbio
);
1986 static void sync_request_write(struct mddev
*mddev
, struct r1bio
*r1_bio
)
1988 struct r1conf
*conf
= mddev
->private;
1990 int disks
= conf
->raid_disks
* 2;
1991 struct bio
*bio
, *wbio
;
1993 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1995 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
1996 /* ouch - failed to read all of that. */
1997 if (!fix_sync_read_error(r1_bio
))
2000 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2001 process_checks(r1_bio
);
2006 atomic_set(&r1_bio
->remaining
, 1);
2007 for (i
= 0; i
< disks
; i
++) {
2008 wbio
= r1_bio
->bios
[i
];
2009 if (wbio
->bi_end_io
== NULL
||
2010 (wbio
->bi_end_io
== end_sync_read
&&
2011 (i
== r1_bio
->read_disk
||
2012 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
2015 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2016 wbio
->bi_end_io
= end_sync_write
;
2017 atomic_inc(&r1_bio
->remaining
);
2018 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, bio_sectors(wbio
));
2020 generic_make_request(wbio
);
2023 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
2024 /* if we're here, all write(s) have completed, so clean up */
2025 int s
= r1_bio
->sectors
;
2026 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2027 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2028 reschedule_retry(r1_bio
);
2031 md_done_sync(mddev
, s
, 1);
2037 * This is a kernel thread which:
2039 * 1. Retries failed read operations on working mirrors.
2040 * 2. Updates the raid superblock when problems encounter.
2041 * 3. Performs writes following reads for array synchronising.
2044 static void fix_read_error(struct r1conf
*conf
, int read_disk
,
2045 sector_t sect
, int sectors
)
2047 struct mddev
*mddev
= conf
->mddev
;
2053 struct md_rdev
*rdev
;
2055 if (s
> (PAGE_SIZE
>>9))
2063 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2065 (test_bit(In_sync
, &rdev
->flags
) ||
2066 (!test_bit(Faulty
, &rdev
->flags
) &&
2067 rdev
->recovery_offset
>= sect
+ s
)) &&
2068 is_badblock(rdev
, sect
, s
,
2069 &first_bad
, &bad_sectors
) == 0) {
2070 atomic_inc(&rdev
->nr_pending
);
2072 if (sync_page_io(rdev
, sect
, s
<<9,
2073 conf
->tmppage
, REQ_OP_READ
, 0, false))
2075 rdev_dec_pending(rdev
, mddev
);
2081 if (d
== conf
->raid_disks
* 2)
2083 } while (!success
&& d
!= read_disk
);
2086 /* Cannot read from anywhere - mark it bad */
2087 struct md_rdev
*rdev
= conf
->mirrors
[read_disk
].rdev
;
2088 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
2089 md_error(mddev
, rdev
);
2092 /* write it back and re-read */
2094 while (d
!= read_disk
) {
2096 d
= conf
->raid_disks
* 2;
2099 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2101 !test_bit(Faulty
, &rdev
->flags
)) {
2102 atomic_inc(&rdev
->nr_pending
);
2104 r1_sync_page_io(rdev
, sect
, s
,
2105 conf
->tmppage
, WRITE
);
2106 rdev_dec_pending(rdev
, mddev
);
2111 while (d
!= read_disk
) {
2112 char b
[BDEVNAME_SIZE
];
2114 d
= conf
->raid_disks
* 2;
2117 rdev
= rcu_dereference(conf
->mirrors
[d
].rdev
);
2119 !test_bit(Faulty
, &rdev
->flags
)) {
2120 atomic_inc(&rdev
->nr_pending
);
2122 if (r1_sync_page_io(rdev
, sect
, s
,
2123 conf
->tmppage
, READ
)) {
2124 atomic_add(s
, &rdev
->corrected_errors
);
2126 "md/raid1:%s: read error corrected "
2127 "(%d sectors at %llu on %s)\n",
2129 (unsigned long long)(sect
+
2131 bdevname(rdev
->bdev
, b
));
2133 rdev_dec_pending(rdev
, mddev
);
2142 static int narrow_write_error(struct r1bio
*r1_bio
, int i
)
2144 struct mddev
*mddev
= r1_bio
->mddev
;
2145 struct r1conf
*conf
= mddev
->private;
2146 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2148 /* bio has the data to be written to device 'i' where
2149 * we just recently had a write error.
2150 * We repeatedly clone the bio and trim down to one block,
2151 * then try the write. Where the write fails we record
2153 * It is conceivable that the bio doesn't exactly align with
2154 * blocks. We must handle this somehow.
2156 * We currently own a reference on the rdev.
2162 int sect_to_write
= r1_bio
->sectors
;
2165 if (rdev
->badblocks
.shift
< 0)
2168 block_sectors
= roundup(1 << rdev
->badblocks
.shift
,
2169 bdev_logical_block_size(rdev
->bdev
) >> 9);
2170 sector
= r1_bio
->sector
;
2171 sectors
= ((sector
+ block_sectors
)
2172 & ~(sector_t
)(block_sectors
- 1))
2175 while (sect_to_write
) {
2177 if (sectors
> sect_to_write
)
2178 sectors
= sect_to_write
;
2179 /* Write at 'sector' for 'sectors'*/
2181 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
2182 unsigned vcnt
= r1_bio
->behind_page_count
;
2183 struct bio_vec
*vec
= r1_bio
->behind_bvecs
;
2185 while (!vec
->bv_page
) {
2190 wbio
= bio_alloc_mddev(GFP_NOIO
, vcnt
, mddev
);
2191 memcpy(wbio
->bi_io_vec
, vec
, vcnt
* sizeof(struct bio_vec
));
2193 wbio
->bi_vcnt
= vcnt
;
2195 wbio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2198 bio_set_op_attrs(wbio
, REQ_OP_WRITE
, 0);
2199 wbio
->bi_iter
.bi_sector
= r1_bio
->sector
;
2200 wbio
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
2202 bio_trim(wbio
, sector
- r1_bio
->sector
, sectors
);
2203 wbio
->bi_iter
.bi_sector
+= rdev
->data_offset
;
2204 wbio
->bi_bdev
= rdev
->bdev
;
2206 if (submit_bio_wait(wbio
) < 0)
2208 ok
= rdev_set_badblocks(rdev
, sector
,
2213 sect_to_write
-= sectors
;
2215 sectors
= block_sectors
;
2220 static void handle_sync_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2223 int s
= r1_bio
->sectors
;
2224 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++) {
2225 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2226 struct bio
*bio
= r1_bio
->bios
[m
];
2227 if (bio
->bi_end_io
== NULL
)
2229 if (!bio
->bi_error
&&
2230 test_bit(R1BIO_MadeGood
, &r1_bio
->state
)) {
2231 rdev_clear_badblocks(rdev
, r1_bio
->sector
, s
, 0);
2233 if (bio
->bi_error
&&
2234 test_bit(R1BIO_WriteError
, &r1_bio
->state
)) {
2235 if (!rdev_set_badblocks(rdev
, r1_bio
->sector
, s
, 0))
2236 md_error(conf
->mddev
, rdev
);
2240 md_done_sync(conf
->mddev
, s
, 1);
2243 static void handle_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2247 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++)
2248 if (r1_bio
->bios
[m
] == IO_MADE_GOOD
) {
2249 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2250 rdev_clear_badblocks(rdev
,
2252 r1_bio
->sectors
, 0);
2253 rdev_dec_pending(rdev
, conf
->mddev
);
2254 } else if (r1_bio
->bios
[m
] != NULL
) {
2255 /* This drive got a write error. We need to
2256 * narrow down and record precise write
2260 if (!narrow_write_error(r1_bio
, m
)) {
2261 md_error(conf
->mddev
,
2262 conf
->mirrors
[m
].rdev
);
2263 /* an I/O failed, we can't clear the bitmap */
2264 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2266 rdev_dec_pending(conf
->mirrors
[m
].rdev
,
2270 spin_lock_irq(&conf
->device_lock
);
2271 list_add(&r1_bio
->retry_list
, &conf
->bio_end_io_list
);
2273 spin_unlock_irq(&conf
->device_lock
);
2274 md_wakeup_thread(conf
->mddev
->thread
);
2276 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2277 close_write(r1_bio
);
2278 raid_end_bio_io(r1_bio
);
2282 static void handle_read_error(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2286 struct mddev
*mddev
= conf
->mddev
;
2288 char b
[BDEVNAME_SIZE
];
2289 struct md_rdev
*rdev
;
2291 clear_bit(R1BIO_ReadError
, &r1_bio
->state
);
2292 /* we got a read error. Maybe the drive is bad. Maybe just
2293 * the block and we can fix it.
2294 * We freeze all other IO, and try reading the block from
2295 * other devices. When we find one, we re-write
2296 * and check it that fixes the read error.
2297 * This is all done synchronously while the array is
2300 if (mddev
->ro
== 0) {
2301 freeze_array(conf
, 1);
2302 fix_read_error(conf
, r1_bio
->read_disk
,
2303 r1_bio
->sector
, r1_bio
->sectors
);
2304 unfreeze_array(conf
);
2306 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
2307 rdev_dec_pending(conf
->mirrors
[r1_bio
->read_disk
].rdev
, conf
->mddev
);
2309 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2310 bdevname(bio
->bi_bdev
, b
);
2312 disk
= read_balance(conf
, r1_bio
, &max_sectors
);
2314 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O"
2315 " read error for block %llu\n",
2316 mdname(mddev
), b
, (unsigned long long)r1_bio
->sector
);
2317 raid_end_bio_io(r1_bio
);
2319 const unsigned long do_sync
2320 = r1_bio
->master_bio
->bi_opf
& REQ_SYNC
;
2322 r1_bio
->bios
[r1_bio
->read_disk
] =
2323 mddev
->ro
? IO_BLOCKED
: NULL
;
2326 r1_bio
->read_disk
= disk
;
2327 bio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2328 bio_trim(bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
2330 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
2331 rdev
= conf
->mirrors
[disk
].rdev
;
2332 printk_ratelimited(KERN_ERR
2333 "md/raid1:%s: redirecting sector %llu"
2334 " to other mirror: %s\n",
2336 (unsigned long long)r1_bio
->sector
,
2337 bdevname(rdev
->bdev
, b
));
2338 bio
->bi_iter
.bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
2339 bio
->bi_bdev
= rdev
->bdev
;
2340 bio
->bi_end_io
= raid1_end_read_request
;
2341 bio_set_op_attrs(bio
, REQ_OP_READ
, do_sync
);
2342 bio
->bi_private
= r1_bio
;
2343 if (max_sectors
< r1_bio
->sectors
) {
2344 /* Drat - have to split this up more */
2345 struct bio
*mbio
= r1_bio
->master_bio
;
2346 int sectors_handled
= (r1_bio
->sector
+ max_sectors
2347 - mbio
->bi_iter
.bi_sector
);
2348 r1_bio
->sectors
= max_sectors
;
2349 spin_lock_irq(&conf
->device_lock
);
2350 if (mbio
->bi_phys_segments
== 0)
2351 mbio
->bi_phys_segments
= 2;
2353 mbio
->bi_phys_segments
++;
2354 spin_unlock_irq(&conf
->device_lock
);
2355 generic_make_request(bio
);
2358 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
2360 r1_bio
->master_bio
= mbio
;
2361 r1_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2363 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
2364 r1_bio
->mddev
= mddev
;
2365 r1_bio
->sector
= mbio
->bi_iter
.bi_sector
+
2370 generic_make_request(bio
);
2374 static void raid1d(struct md_thread
*thread
)
2376 struct mddev
*mddev
= thread
->mddev
;
2377 struct r1bio
*r1_bio
;
2378 unsigned long flags
;
2379 struct r1conf
*conf
= mddev
->private;
2380 struct list_head
*head
= &conf
->retry_list
;
2381 struct blk_plug plug
;
2383 md_check_recovery(mddev
);
2385 if (!list_empty_careful(&conf
->bio_end_io_list
) &&
2386 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2388 spin_lock_irqsave(&conf
->device_lock
, flags
);
2389 if (!test_bit(MD_CHANGE_PENDING
, &mddev
->flags
)) {
2390 while (!list_empty(&conf
->bio_end_io_list
)) {
2391 list_move(conf
->bio_end_io_list
.prev
, &tmp
);
2395 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2396 while (!list_empty(&tmp
)) {
2397 r1_bio
= list_first_entry(&tmp
, struct r1bio
,
2399 list_del(&r1_bio
->retry_list
);
2400 if (mddev
->degraded
)
2401 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2402 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2403 close_write(r1_bio
);
2404 raid_end_bio_io(r1_bio
);
2408 blk_start_plug(&plug
);
2411 flush_pending_writes(conf
);
2413 spin_lock_irqsave(&conf
->device_lock
, flags
);
2414 if (list_empty(head
)) {
2415 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2418 r1_bio
= list_entry(head
->prev
, struct r1bio
, retry_list
);
2419 list_del(head
->prev
);
2421 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2423 mddev
= r1_bio
->mddev
;
2424 conf
= mddev
->private;
2425 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
2426 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2427 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2428 handle_sync_write_finished(conf
, r1_bio
);
2430 sync_request_write(mddev
, r1_bio
);
2431 } else if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2432 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2433 handle_write_finished(conf
, r1_bio
);
2434 else if (test_bit(R1BIO_ReadError
, &r1_bio
->state
))
2435 handle_read_error(conf
, r1_bio
);
2437 /* just a partial read to be scheduled from separate
2440 generic_make_request(r1_bio
->bios
[r1_bio
->read_disk
]);
2443 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2444 md_check_recovery(mddev
);
2446 blk_finish_plug(&plug
);
2449 static int init_resync(struct r1conf
*conf
)
2453 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2454 BUG_ON(conf
->r1buf_pool
);
2455 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
2457 if (!conf
->r1buf_pool
)
2459 conf
->next_resync
= 0;
2464 * perform a "sync" on one "block"
2466 * We need to make sure that no normal I/O request - particularly write
2467 * requests - conflict with active sync requests.
2469 * This is achieved by tracking pending requests and a 'barrier' concept
2470 * that can be installed to exclude normal IO requests.
2473 static sector_t
raid1_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
2476 struct r1conf
*conf
= mddev
->private;
2477 struct r1bio
*r1_bio
;
2479 sector_t max_sector
, nr_sectors
;
2483 int write_targets
= 0, read_targets
= 0;
2484 sector_t sync_blocks
;
2485 int still_degraded
= 0;
2486 int good_sectors
= RESYNC_SECTORS
;
2487 int min_bad
= 0; /* number of sectors that are bad in all devices */
2489 if (!conf
->r1buf_pool
)
2490 if (init_resync(conf
))
2493 max_sector
= mddev
->dev_sectors
;
2494 if (sector_nr
>= max_sector
) {
2495 /* If we aborted, we need to abort the
2496 * sync on the 'current' bitmap chunk (there will
2497 * only be one in raid1 resync.
2498 * We can find the current addess in mddev->curr_resync
2500 if (mddev
->curr_resync
< max_sector
) /* aborted */
2501 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2503 else /* completed sync */
2506 bitmap_close_sync(mddev
->bitmap
);
2509 if (mddev_is_clustered(mddev
)) {
2510 conf
->cluster_sync_low
= 0;
2511 conf
->cluster_sync_high
= 0;
2516 if (mddev
->bitmap
== NULL
&&
2517 mddev
->recovery_cp
== MaxSector
&&
2518 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2519 conf
->fullsync
== 0) {
2521 return max_sector
- sector_nr
;
2523 /* before building a request, check if we can skip these blocks..
2524 * This call the bitmap_start_sync doesn't actually record anything
2526 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
2527 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2528 /* We can skip this block, and probably several more */
2534 * If there is non-resync activity waiting for a turn, then let it
2535 * though before starting on this new sync request.
2537 if (conf
->nr_waiting
)
2538 schedule_timeout_uninterruptible(1);
2540 /* we are incrementing sector_nr below. To be safe, we check against
2541 * sector_nr + two times RESYNC_SECTORS
2544 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
,
2545 mddev_is_clustered(mddev
) && (sector_nr
+ 2 * RESYNC_SECTORS
> conf
->cluster_sync_high
));
2546 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
2548 raise_barrier(conf
, sector_nr
);
2552 * If we get a correctably read error during resync or recovery,
2553 * we might want to read from a different device. So we
2554 * flag all drives that could conceivably be read from for READ,
2555 * and any others (which will be non-In_sync devices) for WRITE.
2556 * If a read fails, we try reading from something else for which READ
2560 r1_bio
->mddev
= mddev
;
2561 r1_bio
->sector
= sector_nr
;
2563 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
2565 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2566 struct md_rdev
*rdev
;
2567 bio
= r1_bio
->bios
[i
];
2570 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
2572 test_bit(Faulty
, &rdev
->flags
)) {
2573 if (i
< conf
->raid_disks
)
2575 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
2576 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
2577 bio
->bi_end_io
= end_sync_write
;
2580 /* may need to read from here */
2581 sector_t first_bad
= MaxSector
;
2584 if (is_badblock(rdev
, sector_nr
, good_sectors
,
2585 &first_bad
, &bad_sectors
)) {
2586 if (first_bad
> sector_nr
)
2587 good_sectors
= first_bad
- sector_nr
;
2589 bad_sectors
-= (sector_nr
- first_bad
);
2591 min_bad
> bad_sectors
)
2592 min_bad
= bad_sectors
;
2595 if (sector_nr
< first_bad
) {
2596 if (test_bit(WriteMostly
, &rdev
->flags
)) {
2603 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
2604 bio
->bi_end_io
= end_sync_read
;
2606 } else if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
2607 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) &&
2608 !test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)) {
2610 * The device is suitable for reading (InSync),
2611 * but has bad block(s) here. Let's try to correct them,
2612 * if we are doing resync or repair. Otherwise, leave
2613 * this device alone for this sync request.
2615 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
2616 bio
->bi_end_io
= end_sync_write
;
2620 if (bio
->bi_end_io
) {
2621 atomic_inc(&rdev
->nr_pending
);
2622 bio
->bi_iter
.bi_sector
= sector_nr
+ rdev
->data_offset
;
2623 bio
->bi_bdev
= rdev
->bdev
;
2624 bio
->bi_private
= r1_bio
;
2630 r1_bio
->read_disk
= disk
;
2632 if (read_targets
== 0 && min_bad
> 0) {
2633 /* These sectors are bad on all InSync devices, so we
2634 * need to mark them bad on all write targets
2637 for (i
= 0 ; i
< conf
->raid_disks
* 2 ; i
++)
2638 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_write
) {
2639 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2640 ok
= rdev_set_badblocks(rdev
, sector_nr
,
2644 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2649 /* Cannot record the badblocks, so need to
2651 * If there are multiple read targets, could just
2652 * fail the really bad ones ???
2654 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2655 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2661 if (min_bad
> 0 && min_bad
< good_sectors
) {
2662 /* only resync enough to reach the next bad->good
2664 good_sectors
= min_bad
;
2667 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
2668 /* extra read targets are also write targets */
2669 write_targets
+= read_targets
-1;
2671 if (write_targets
== 0 || read_targets
== 0) {
2672 /* There is nowhere to write, so all non-sync
2673 * drives must be failed - so we are finished
2677 max_sector
= sector_nr
+ min_bad
;
2678 rv
= max_sector
- sector_nr
;
2684 if (max_sector
> mddev
->resync_max
)
2685 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2686 if (max_sector
> sector_nr
+ good_sectors
)
2687 max_sector
= sector_nr
+ good_sectors
;
2692 int len
= PAGE_SIZE
;
2693 if (sector_nr
+ (len
>>9) > max_sector
)
2694 len
= (max_sector
- sector_nr
) << 9;
2697 if (sync_blocks
== 0) {
2698 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2699 &sync_blocks
, still_degraded
) &&
2701 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2703 if ((len
>> 9) > sync_blocks
)
2704 len
= sync_blocks
<<9;
2707 for (i
= 0 ; i
< conf
->raid_disks
* 2; i
++) {
2708 bio
= r1_bio
->bios
[i
];
2709 if (bio
->bi_end_io
) {
2710 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2711 if (bio_add_page(bio
, page
, len
, 0) == 0) {
2713 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2716 bio
= r1_bio
->bios
[i
];
2717 if (bio
->bi_end_io
==NULL
)
2719 /* remove last page from this bio */
2721 bio
->bi_iter
.bi_size
-= len
;
2722 bio_clear_flag(bio
, BIO_SEG_VALID
);
2728 nr_sectors
+= len
>>9;
2729 sector_nr
+= len
>>9;
2730 sync_blocks
-= (len
>>9);
2731 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
2733 r1_bio
->sectors
= nr_sectors
;
2735 if (mddev_is_clustered(mddev
) &&
2736 conf
->cluster_sync_high
< sector_nr
+ nr_sectors
) {
2737 conf
->cluster_sync_low
= mddev
->curr_resync_completed
;
2738 conf
->cluster_sync_high
= conf
->cluster_sync_low
+ CLUSTER_RESYNC_WINDOW_SECTORS
;
2739 /* Send resync message */
2740 md_cluster_ops
->resync_info_update(mddev
,
2741 conf
->cluster_sync_low
,
2742 conf
->cluster_sync_high
);
2745 /* For a user-requested sync, we read all readable devices and do a
2748 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2749 atomic_set(&r1_bio
->remaining
, read_targets
);
2750 for (i
= 0; i
< conf
->raid_disks
* 2 && read_targets
; i
++) {
2751 bio
= r1_bio
->bios
[i
];
2752 if (bio
->bi_end_io
== end_sync_read
) {
2754 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2755 generic_make_request(bio
);
2759 atomic_set(&r1_bio
->remaining
, 1);
2760 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2761 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2762 generic_make_request(bio
);
2768 static sector_t
raid1_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2773 return mddev
->dev_sectors
;
2776 static struct r1conf
*setup_conf(struct mddev
*mddev
)
2778 struct r1conf
*conf
;
2780 struct raid1_info
*disk
;
2781 struct md_rdev
*rdev
;
2784 conf
= kzalloc(sizeof(struct r1conf
), GFP_KERNEL
);
2788 conf
->mirrors
= kzalloc(sizeof(struct raid1_info
)
2789 * mddev
->raid_disks
* 2,
2794 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2798 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
2799 if (!conf
->poolinfo
)
2801 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
* 2;
2802 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2805 if (!conf
->r1bio_pool
)
2808 conf
->poolinfo
->mddev
= mddev
;
2811 spin_lock_init(&conf
->device_lock
);
2812 rdev_for_each(rdev
, mddev
) {
2813 struct request_queue
*q
;
2814 int disk_idx
= rdev
->raid_disk
;
2815 if (disk_idx
>= mddev
->raid_disks
2818 if (test_bit(Replacement
, &rdev
->flags
))
2819 disk
= conf
->mirrors
+ mddev
->raid_disks
+ disk_idx
;
2821 disk
= conf
->mirrors
+ disk_idx
;
2826 q
= bdev_get_queue(rdev
->bdev
);
2828 disk
->head_position
= 0;
2829 disk
->seq_start
= MaxSector
;
2831 conf
->raid_disks
= mddev
->raid_disks
;
2832 conf
->mddev
= mddev
;
2833 INIT_LIST_HEAD(&conf
->retry_list
);
2834 INIT_LIST_HEAD(&conf
->bio_end_io_list
);
2836 spin_lock_init(&conf
->resync_lock
);
2837 init_waitqueue_head(&conf
->wait_barrier
);
2839 bio_list_init(&conf
->pending_bio_list
);
2840 conf
->pending_count
= 0;
2841 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
2843 conf
->start_next_window
= MaxSector
;
2844 conf
->current_window_requests
= conf
->next_window_requests
= 0;
2847 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2849 disk
= conf
->mirrors
+ i
;
2851 if (i
< conf
->raid_disks
&&
2852 disk
[conf
->raid_disks
].rdev
) {
2853 /* This slot has a replacement. */
2855 /* No original, just make the replacement
2856 * a recovering spare
2859 disk
[conf
->raid_disks
].rdev
;
2860 disk
[conf
->raid_disks
].rdev
= NULL
;
2861 } else if (!test_bit(In_sync
, &disk
->rdev
->flags
))
2862 /* Original is not in_sync - bad */
2867 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2868 disk
->head_position
= 0;
2870 (disk
->rdev
->saved_raid_disk
< 0))
2876 conf
->thread
= md_register_thread(raid1d
, mddev
, "raid1");
2877 if (!conf
->thread
) {
2879 "md/raid1:%s: couldn't allocate thread\n",
2888 mempool_destroy(conf
->r1bio_pool
);
2889 kfree(conf
->mirrors
);
2890 safe_put_page(conf
->tmppage
);
2891 kfree(conf
->poolinfo
);
2894 return ERR_PTR(err
);
2897 static void raid1_free(struct mddev
*mddev
, void *priv
);
2898 static int raid1_run(struct mddev
*mddev
)
2900 struct r1conf
*conf
;
2902 struct md_rdev
*rdev
;
2904 bool discard_supported
= false;
2906 if (mddev
->level
!= 1) {
2907 printk(KERN_ERR
"md/raid1:%s: raid level not set to mirroring (%d)\n",
2908 mdname(mddev
), mddev
->level
);
2911 if (mddev
->reshape_position
!= MaxSector
) {
2912 printk(KERN_ERR
"md/raid1:%s: reshape_position set but not supported\n",
2917 * copy the already verified devices into our private RAID1
2918 * bookkeeping area. [whatever we allocate in run(),
2919 * should be freed in raid1_free()]
2921 if (mddev
->private == NULL
)
2922 conf
= setup_conf(mddev
);
2924 conf
= mddev
->private;
2927 return PTR_ERR(conf
);
2930 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
2932 rdev_for_each(rdev
, mddev
) {
2933 if (!mddev
->gendisk
)
2935 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2936 rdev
->data_offset
<< 9);
2937 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
2938 discard_supported
= true;
2941 mddev
->degraded
= 0;
2942 for (i
=0; i
< conf
->raid_disks
; i
++)
2943 if (conf
->mirrors
[i
].rdev
== NULL
||
2944 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
2945 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2948 if (conf
->raid_disks
- mddev
->degraded
== 1)
2949 mddev
->recovery_cp
= MaxSector
;
2951 if (mddev
->recovery_cp
!= MaxSector
)
2952 printk(KERN_NOTICE
"md/raid1:%s: not clean"
2953 " -- starting background reconstruction\n",
2956 "md/raid1:%s: active with %d out of %d mirrors\n",
2957 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2961 * Ok, everything is just fine now
2963 mddev
->thread
= conf
->thread
;
2964 conf
->thread
= NULL
;
2965 mddev
->private = conf
;
2967 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
2970 if (discard_supported
)
2971 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
2974 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
2978 ret
= md_integrity_register(mddev
);
2980 md_unregister_thread(&mddev
->thread
);
2981 raid1_free(mddev
, conf
);
2986 static void raid1_free(struct mddev
*mddev
, void *priv
)
2988 struct r1conf
*conf
= priv
;
2990 mempool_destroy(conf
->r1bio_pool
);
2991 kfree(conf
->mirrors
);
2992 safe_put_page(conf
->tmppage
);
2993 kfree(conf
->poolinfo
);
2997 static int raid1_resize(struct mddev
*mddev
, sector_t sectors
)
2999 /* no resync is happening, and there is enough space
3000 * on all devices, so we can resize.
3001 * We need to make sure resync covers any new space.
3002 * If the array is shrinking we should possibly wait until
3003 * any io in the removed space completes, but it hardly seems
3006 sector_t newsize
= raid1_size(mddev
, sectors
, 0);
3007 if (mddev
->external_size
&&
3008 mddev
->array_sectors
> newsize
)
3010 if (mddev
->bitmap
) {
3011 int ret
= bitmap_resize(mddev
->bitmap
, newsize
, 0, 0);
3015 md_set_array_sectors(mddev
, newsize
);
3016 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3017 revalidate_disk(mddev
->gendisk
);
3018 if (sectors
> mddev
->dev_sectors
&&
3019 mddev
->recovery_cp
> mddev
->dev_sectors
) {
3020 mddev
->recovery_cp
= mddev
->dev_sectors
;
3021 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3023 mddev
->dev_sectors
= sectors
;
3024 mddev
->resync_max_sectors
= sectors
;
3028 static int raid1_reshape(struct mddev
*mddev
)
3031 * 1/ resize the r1bio_pool
3032 * 2/ resize conf->mirrors
3034 * We allocate a new r1bio_pool if we can.
3035 * Then raise a device barrier and wait until all IO stops.
3036 * Then resize conf->mirrors and swap in the new r1bio pool.
3038 * At the same time, we "pack" the devices so that all the missing
3039 * devices have the higher raid_disk numbers.
3041 mempool_t
*newpool
, *oldpool
;
3042 struct pool_info
*newpoolinfo
;
3043 struct raid1_info
*newmirrors
;
3044 struct r1conf
*conf
= mddev
->private;
3045 int cnt
, raid_disks
;
3046 unsigned long flags
;
3049 /* Cannot change chunk_size, layout, or level */
3050 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
3051 mddev
->layout
!= mddev
->new_layout
||
3052 mddev
->level
!= mddev
->new_level
) {
3053 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3054 mddev
->new_layout
= mddev
->layout
;
3055 mddev
->new_level
= mddev
->level
;
3059 if (!mddev_is_clustered(mddev
)) {
3060 err
= md_allow_write(mddev
);
3065 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3067 if (raid_disks
< conf
->raid_disks
) {
3069 for (d
= 0; d
< conf
->raid_disks
; d
++)
3070 if (conf
->mirrors
[d
].rdev
)
3072 if (cnt
> raid_disks
)
3076 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
3079 newpoolinfo
->mddev
= mddev
;
3080 newpoolinfo
->raid_disks
= raid_disks
* 2;
3082 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
3083 r1bio_pool_free
, newpoolinfo
);
3088 newmirrors
= kzalloc(sizeof(struct raid1_info
) * raid_disks
* 2,
3092 mempool_destroy(newpool
);
3096 freeze_array(conf
, 0);
3098 /* ok, everything is stopped */
3099 oldpool
= conf
->r1bio_pool
;
3100 conf
->r1bio_pool
= newpool
;
3102 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
3103 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
3104 if (rdev
&& rdev
->raid_disk
!= d2
) {
3105 sysfs_unlink_rdev(mddev
, rdev
);
3106 rdev
->raid_disk
= d2
;
3107 sysfs_unlink_rdev(mddev
, rdev
);
3108 if (sysfs_link_rdev(mddev
, rdev
))
3110 "md/raid1:%s: cannot register rd%d\n",
3111 mdname(mddev
), rdev
->raid_disk
);
3114 newmirrors
[d2
++].rdev
= rdev
;
3116 kfree(conf
->mirrors
);
3117 conf
->mirrors
= newmirrors
;
3118 kfree(conf
->poolinfo
);
3119 conf
->poolinfo
= newpoolinfo
;
3121 spin_lock_irqsave(&conf
->device_lock
, flags
);
3122 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
3123 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3124 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
3125 mddev
->delta_disks
= 0;
3127 unfreeze_array(conf
);
3129 set_bit(MD_RECOVERY_RECOVER
, &mddev
->recovery
);
3130 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3131 md_wakeup_thread(mddev
->thread
);
3133 mempool_destroy(oldpool
);
3137 static void raid1_quiesce(struct mddev
*mddev
, int state
)
3139 struct r1conf
*conf
= mddev
->private;
3142 case 2: /* wake for suspend */
3143 wake_up(&conf
->wait_barrier
);
3146 freeze_array(conf
, 0);
3149 unfreeze_array(conf
);
3154 static void *raid1_takeover(struct mddev
*mddev
)
3156 /* raid1 can take over:
3157 * raid5 with 2 devices, any layout or chunk size
3159 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
3160 struct r1conf
*conf
;
3161 mddev
->new_level
= 1;
3162 mddev
->new_layout
= 0;
3163 mddev
->new_chunk_sectors
= 0;
3164 conf
= setup_conf(mddev
);
3166 /* Array must appear to be quiesced */
3167 conf
->array_frozen
= 1;
3170 return ERR_PTR(-EINVAL
);
3173 static struct md_personality raid1_personality
=
3177 .owner
= THIS_MODULE
,
3178 .make_request
= raid1_make_request
,
3181 .status
= raid1_status
,
3182 .error_handler
= raid1_error
,
3183 .hot_add_disk
= raid1_add_disk
,
3184 .hot_remove_disk
= raid1_remove_disk
,
3185 .spare_active
= raid1_spare_active
,
3186 .sync_request
= raid1_sync_request
,
3187 .resize
= raid1_resize
,
3189 .check_reshape
= raid1_reshape
,
3190 .quiesce
= raid1_quiesce
,
3191 .takeover
= raid1_takeover
,
3192 .congested
= raid1_congested
,
3195 static int __init
raid_init(void)
3197 return register_md_personality(&raid1_personality
);
3200 static void raid_exit(void)
3202 unregister_md_personality(&raid1_personality
);
3205 module_init(raid_init
);
3206 module_exit(raid_exit
);
3207 MODULE_LICENSE("GPL");
3208 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3209 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3210 MODULE_ALIAS("md-raid1");
3211 MODULE_ALIAS("md-level-1");
3213 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);