2 * raid1.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * RAID-1 management functions.
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
45 * Number of guaranteed r1bios in case of extreme VM load:
47 #define NR_RAID1_BIOS 256
49 /* when we get a read error on a read-only array, we redirect to another
50 * device without failing the first device, or trying to over-write to
51 * correct the read error. To keep track of bad blocks on a per-bio
52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
54 #define IO_BLOCKED ((struct bio *)1)
55 /* When we successfully write to a known bad-block, we need to remove the
56 * bad-block marking which must be done from process context. So we record
57 * the success by setting devs[n].bio to IO_MADE_GOOD
59 #define IO_MADE_GOOD ((struct bio *)2)
61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
63 /* When there are this many requests queue to be written by
64 * the raid1 thread, we become 'congested' to provide back-pressure
67 static int max_queued_requests
= 1024;
69 static void allow_barrier(struct r1conf
*conf
, sector_t start_next_window
,
71 static void lower_barrier(struct r1conf
*conf
);
73 static void * r1bio_pool_alloc(gfp_t gfp_flags
, void *data
)
75 struct pool_info
*pi
= data
;
76 int size
= offsetof(struct r1bio
, bios
[pi
->raid_disks
]);
78 /* allocate a r1bio with room for raid_disks entries in the bios array */
79 return kzalloc(size
, gfp_flags
);
82 static void r1bio_pool_free(void *r1_bio
, void *data
)
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_DEPTH 32
89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
93 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
95 static void * r1buf_pool_alloc(gfp_t gfp_flags
, void *data
)
97 struct pool_info
*pi
= data
;
103 r1_bio
= r1bio_pool_alloc(gfp_flags
, pi
);
108 * Allocate bios : 1 for reading, n-1 for writing
110 for (j
= pi
->raid_disks
; j
-- ; ) {
111 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
114 r1_bio
->bios
[j
] = bio
;
117 * Allocate RESYNC_PAGES data pages and attach them to
119 * If this is a user-requested check/repair, allocate
120 * RESYNC_PAGES for each bio.
122 if (test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
))
123 need_pages
= pi
->raid_disks
;
126 for (j
= 0; j
< need_pages
; j
++) {
127 bio
= r1_bio
->bios
[j
];
128 bio
->bi_vcnt
= RESYNC_PAGES
;
130 if (bio_alloc_pages(bio
, gfp_flags
))
133 /* If not user-requests, copy the page pointers to all bios */
134 if (!test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
)) {
135 for (i
=0; i
<RESYNC_PAGES
; i
++)
136 for (j
=1; j
<pi
->raid_disks
; j
++)
137 r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
=
138 r1_bio
->bios
[0]->bi_io_vec
[i
].bv_page
;
141 r1_bio
->master_bio
= NULL
;
149 bio_for_each_segment_all(bv
, r1_bio
->bios
[j
], i
)
150 __free_page(bv
->bv_page
);
154 while (++j
< pi
->raid_disks
)
155 bio_put(r1_bio
->bios
[j
]);
156 r1bio_pool_free(r1_bio
, data
);
160 static void r1buf_pool_free(void *__r1_bio
, void *data
)
162 struct pool_info
*pi
= data
;
164 struct r1bio
*r1bio
= __r1_bio
;
166 for (i
= 0; i
< RESYNC_PAGES
; i
++)
167 for (j
= pi
->raid_disks
; j
-- ;) {
169 r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
!=
170 r1bio
->bios
[0]->bi_io_vec
[i
].bv_page
)
171 safe_put_page(r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
173 for (i
=0 ; i
< pi
->raid_disks
; i
++)
174 bio_put(r1bio
->bios
[i
]);
176 r1bio_pool_free(r1bio
, data
);
179 static void put_all_bios(struct r1conf
*conf
, struct r1bio
*r1_bio
)
183 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
184 struct bio
**bio
= r1_bio
->bios
+ i
;
185 if (!BIO_SPECIAL(*bio
))
191 static void free_r1bio(struct r1bio
*r1_bio
)
193 struct r1conf
*conf
= r1_bio
->mddev
->private;
195 put_all_bios(conf
, r1_bio
);
196 mempool_free(r1_bio
, conf
->r1bio_pool
);
199 static void put_buf(struct r1bio
*r1_bio
)
201 struct r1conf
*conf
= r1_bio
->mddev
->private;
204 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
205 struct bio
*bio
= r1_bio
->bios
[i
];
207 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
210 mempool_free(r1_bio
, conf
->r1buf_pool
);
215 static void reschedule_retry(struct r1bio
*r1_bio
)
218 struct mddev
*mddev
= r1_bio
->mddev
;
219 struct r1conf
*conf
= mddev
->private;
221 spin_lock_irqsave(&conf
->device_lock
, flags
);
222 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
224 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
226 wake_up(&conf
->wait_barrier
);
227 md_wakeup_thread(mddev
->thread
);
231 * raid_end_bio_io() is called when we have finished servicing a mirrored
232 * operation and are ready to return a success/failure code to the buffer
235 static void call_bio_endio(struct r1bio
*r1_bio
)
237 struct bio
*bio
= r1_bio
->master_bio
;
239 struct r1conf
*conf
= r1_bio
->mddev
->private;
240 sector_t start_next_window
= r1_bio
->start_next_window
;
241 sector_t bi_sector
= bio
->bi_iter
.bi_sector
;
243 if (bio
->bi_phys_segments
) {
245 spin_lock_irqsave(&conf
->device_lock
, flags
);
246 bio
->bi_phys_segments
--;
247 done
= (bio
->bi_phys_segments
== 0);
248 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
250 * make_request() might be waiting for
251 * bi_phys_segments to decrease
253 wake_up(&conf
->wait_barrier
);
257 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
258 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
262 * Wake up any possible resync thread that waits for the device
265 allow_barrier(conf
, start_next_window
, bi_sector
);
269 static void raid_end_bio_io(struct r1bio
*r1_bio
)
271 struct bio
*bio
= r1_bio
->master_bio
;
273 /* if nobody has done the final endio yet, do it now */
274 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
275 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
276 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
277 (unsigned long long) bio
->bi_iter
.bi_sector
,
278 (unsigned long long) bio_end_sector(bio
) - 1);
280 call_bio_endio(r1_bio
);
286 * Update disk head position estimator based on IRQ completion info.
288 static inline void update_head_pos(int disk
, struct r1bio
*r1_bio
)
290 struct r1conf
*conf
= r1_bio
->mddev
->private;
292 conf
->mirrors
[disk
].head_position
=
293 r1_bio
->sector
+ (r1_bio
->sectors
);
297 * Find the disk number which triggered given bio
299 static int find_bio_disk(struct r1bio
*r1_bio
, struct bio
*bio
)
302 struct r1conf
*conf
= r1_bio
->mddev
->private;
303 int raid_disks
= conf
->raid_disks
;
305 for (mirror
= 0; mirror
< raid_disks
* 2; mirror
++)
306 if (r1_bio
->bios
[mirror
] == bio
)
309 BUG_ON(mirror
== raid_disks
* 2);
310 update_head_pos(mirror
, r1_bio
);
315 static void raid1_end_read_request(struct bio
*bio
, int error
)
317 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
318 struct r1bio
*r1_bio
= bio
->bi_private
;
320 struct r1conf
*conf
= r1_bio
->mddev
->private;
322 mirror
= r1_bio
->read_disk
;
324 * this branch is our 'one mirror IO has finished' event handler:
326 update_head_pos(mirror
, r1_bio
);
329 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
331 /* If all other devices have failed, we want to return
332 * the error upwards rather than fail the last device.
333 * Here we redefine "uptodate" to mean "Don't want to retry"
336 spin_lock_irqsave(&conf
->device_lock
, flags
);
337 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
338 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
339 !test_bit(Faulty
, &conf
->mirrors
[mirror
].rdev
->flags
)))
341 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
345 raid_end_bio_io(r1_bio
);
346 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
351 char b
[BDEVNAME_SIZE
];
353 KERN_ERR
"md/raid1:%s: %s: "
354 "rescheduling sector %llu\n",
356 bdevname(conf
->mirrors
[mirror
].rdev
->bdev
,
358 (unsigned long long)r1_bio
->sector
);
359 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
360 reschedule_retry(r1_bio
);
361 /* don't drop the reference on read_disk yet */
365 static void close_write(struct r1bio
*r1_bio
)
367 /* it really is the end of this request */
368 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
369 /* free extra copy of the data pages */
370 int i
= r1_bio
->behind_page_count
;
372 safe_put_page(r1_bio
->behind_bvecs
[i
].bv_page
);
373 kfree(r1_bio
->behind_bvecs
);
374 r1_bio
->behind_bvecs
= NULL
;
376 /* clear the bitmap if all writes complete successfully */
377 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
379 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
380 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
381 md_write_end(r1_bio
->mddev
);
384 static void r1_bio_write_done(struct r1bio
*r1_bio
)
386 if (!atomic_dec_and_test(&r1_bio
->remaining
))
389 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
390 reschedule_retry(r1_bio
);
393 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
))
394 reschedule_retry(r1_bio
);
396 raid_end_bio_io(r1_bio
);
400 static void raid1_end_write_request(struct bio
*bio
, int error
)
402 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
403 struct r1bio
*r1_bio
= bio
->bi_private
;
404 int mirror
, behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
405 struct r1conf
*conf
= r1_bio
->mddev
->private;
406 struct bio
*to_put
= NULL
;
408 mirror
= find_bio_disk(r1_bio
, bio
);
411 * 'one mirror IO has finished' event handler:
414 set_bit(WriteErrorSeen
,
415 &conf
->mirrors
[mirror
].rdev
->flags
);
416 if (!test_and_set_bit(WantReplacement
,
417 &conf
->mirrors
[mirror
].rdev
->flags
))
418 set_bit(MD_RECOVERY_NEEDED
, &
419 conf
->mddev
->recovery
);
421 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
424 * Set R1BIO_Uptodate in our master bio, so that we
425 * will return a good error code for to the higher
426 * levels even if IO on some other mirrored buffer
429 * The 'master' represents the composite IO operation
430 * to user-side. So if something waits for IO, then it
431 * will wait for the 'master' bio.
436 r1_bio
->bios
[mirror
] = NULL
;
439 * Do not set R1BIO_Uptodate if the current device is
440 * rebuilding or Faulty. This is because we cannot use
441 * such device for properly reading the data back (we could
442 * potentially use it, if the current write would have felt
443 * before rdev->recovery_offset, but for simplicity we don't
446 if (test_bit(In_sync
, &conf
->mirrors
[mirror
].rdev
->flags
) &&
447 !test_bit(Faulty
, &conf
->mirrors
[mirror
].rdev
->flags
))
448 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
450 /* Maybe we can clear some bad blocks. */
451 if (is_badblock(conf
->mirrors
[mirror
].rdev
,
452 r1_bio
->sector
, r1_bio
->sectors
,
453 &first_bad
, &bad_sectors
)) {
454 r1_bio
->bios
[mirror
] = IO_MADE_GOOD
;
455 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
460 if (test_bit(WriteMostly
, &conf
->mirrors
[mirror
].rdev
->flags
))
461 atomic_dec(&r1_bio
->behind_remaining
);
464 * In behind mode, we ACK the master bio once the I/O
465 * has safely reached all non-writemostly
466 * disks. Setting the Returned bit ensures that this
467 * gets done only once -- we don't ever want to return
468 * -EIO here, instead we'll wait
470 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
471 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
472 /* Maybe we can return now */
473 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
474 struct bio
*mbio
= r1_bio
->master_bio
;
475 pr_debug("raid1: behind end write sectors"
477 (unsigned long long) mbio
->bi_iter
.bi_sector
,
478 (unsigned long long) bio_end_sector(mbio
) - 1);
479 call_bio_endio(r1_bio
);
483 if (r1_bio
->bios
[mirror
] == NULL
)
484 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
,
488 * Let's see if all mirrored write operations have finished
491 r1_bio_write_done(r1_bio
);
499 * This routine returns the disk from which the requested read should
500 * be done. There is a per-array 'next expected sequential IO' sector
501 * number - if this matches on the next IO then we use the last disk.
502 * There is also a per-disk 'last know head position' sector that is
503 * maintained from IRQ contexts, both the normal and the resync IO
504 * completion handlers update this position correctly. If there is no
505 * perfect sequential match then we pick the disk whose head is closest.
507 * If there are 2 mirrors in the same 2 devices, performance degrades
508 * because position is mirror, not device based.
510 * The rdev for the device selected will have nr_pending incremented.
512 static int read_balance(struct r1conf
*conf
, struct r1bio
*r1_bio
, int *max_sectors
)
514 const sector_t this_sector
= r1_bio
->sector
;
516 int best_good_sectors
;
517 int best_disk
, best_dist_disk
, best_pending_disk
;
521 unsigned int min_pending
;
522 struct md_rdev
*rdev
;
524 int choose_next_idle
;
528 * Check if we can balance. We can balance on the whole
529 * device if no resync is going on, or below the resync window.
530 * We take the first readable disk when above the resync window.
533 sectors
= r1_bio
->sectors
;
536 best_dist
= MaxSector
;
537 best_pending_disk
= -1;
538 min_pending
= UINT_MAX
;
539 best_good_sectors
= 0;
541 choose_next_idle
= 0;
543 choose_first
= (conf
->mddev
->recovery_cp
< this_sector
+ sectors
);
545 for (disk
= 0 ; disk
< conf
->raid_disks
* 2 ; disk
++) {
549 unsigned int pending
;
552 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
553 if (r1_bio
->bios
[disk
] == IO_BLOCKED
555 || test_bit(Unmerged
, &rdev
->flags
)
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 */
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
;
577 /* This is a reasonable device to use. It might
580 if (is_badblock(rdev
, this_sector
, sectors
,
581 &first_bad
, &bad_sectors
)) {
582 if (best_dist
< MaxSector
)
583 /* already have a better device */
585 if (first_bad
<= this_sector
) {
586 /* cannot read here. If this is the 'primary'
587 * device, then we must not read beyond
588 * bad_sectors from another device..
590 bad_sectors
-= (this_sector
- first_bad
);
591 if (choose_first
&& sectors
> bad_sectors
)
592 sectors
= bad_sectors
;
593 if (best_good_sectors
> sectors
)
594 best_good_sectors
= sectors
;
597 sector_t good_sectors
= first_bad
- this_sector
;
598 if (good_sectors
> best_good_sectors
) {
599 best_good_sectors
= good_sectors
;
607 best_good_sectors
= sectors
;
609 nonrot
= blk_queue_nonrot(bdev_get_queue(rdev
->bdev
));
610 has_nonrot_disk
|= nonrot
;
611 pending
= atomic_read(&rdev
->nr_pending
);
612 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
617 /* Don't change to another disk for sequential reads */
618 if (conf
->mirrors
[disk
].next_seq_sect
== this_sector
620 int opt_iosize
= bdev_io_opt(rdev
->bdev
) >> 9;
621 struct raid1_info
*mirror
= &conf
->mirrors
[disk
];
625 * If buffered sequential IO size exceeds optimal
626 * iosize, check if there is idle disk. If yes, choose
627 * the idle disk. read_balance could already choose an
628 * idle disk before noticing it's a sequential IO in
629 * this disk. This doesn't matter because this disk
630 * will idle, next time it will be utilized after the
631 * first disk has IO size exceeds optimal iosize. In
632 * this way, iosize of the first disk will be optimal
633 * iosize at least. iosize of the second disk might be
634 * small, but not a big deal since when the second disk
635 * starts IO, the first disk is likely still busy.
637 if (nonrot
&& opt_iosize
> 0 &&
638 mirror
->seq_start
!= MaxSector
&&
639 mirror
->next_seq_sect
> opt_iosize
&&
640 mirror
->next_seq_sect
- opt_iosize
>=
642 choose_next_idle
= 1;
647 /* If device is idle, use it */
653 if (choose_next_idle
)
656 if (min_pending
> pending
) {
657 min_pending
= pending
;
658 best_pending_disk
= disk
;
661 if (dist
< best_dist
) {
663 best_dist_disk
= disk
;
668 * If all disks are rotational, choose the closest disk. If any disk is
669 * non-rotational, choose the disk with less pending request even the
670 * disk is rotational, which might/might not be optimal for raids with
671 * mixed ratation/non-rotational disks depending on workload.
673 if (best_disk
== -1) {
675 best_disk
= best_pending_disk
;
677 best_disk
= best_dist_disk
;
680 if (best_disk
>= 0) {
681 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
684 atomic_inc(&rdev
->nr_pending
);
685 if (test_bit(Faulty
, &rdev
->flags
)) {
686 /* cannot risk returning a device that failed
687 * before we inc'ed nr_pending
689 rdev_dec_pending(rdev
, conf
->mddev
);
692 sectors
= best_good_sectors
;
694 if (conf
->mirrors
[best_disk
].next_seq_sect
!= this_sector
)
695 conf
->mirrors
[best_disk
].seq_start
= this_sector
;
697 conf
->mirrors
[best_disk
].next_seq_sect
= this_sector
+ sectors
;
700 *max_sectors
= sectors
;
705 static int raid1_mergeable_bvec(struct request_queue
*q
,
706 struct bvec_merge_data
*bvm
,
707 struct bio_vec
*biovec
)
709 struct mddev
*mddev
= q
->queuedata
;
710 struct r1conf
*conf
= mddev
->private;
711 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
712 int max
= biovec
->bv_len
;
714 if (mddev
->merge_check_needed
) {
717 for (disk
= 0; disk
< conf
->raid_disks
* 2; disk
++) {
718 struct md_rdev
*rdev
= rcu_dereference(
719 conf
->mirrors
[disk
].rdev
);
720 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
721 struct request_queue
*q
=
722 bdev_get_queue(rdev
->bdev
);
723 if (q
->merge_bvec_fn
) {
724 bvm
->bi_sector
= sector
+
726 bvm
->bi_bdev
= rdev
->bdev
;
727 max
= min(max
, q
->merge_bvec_fn(
738 int md_raid1_congested(struct mddev
*mddev
, int bits
)
740 struct r1conf
*conf
= mddev
->private;
743 if ((bits
& (1 << BDI_async_congested
)) &&
744 conf
->pending_count
>= max_queued_requests
)
748 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
749 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
750 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
751 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
755 /* Note the '|| 1' - when read_balance prefers
756 * non-congested targets, it can be removed
758 if ((bits
& (1<<BDI_async_congested
)) || 1)
759 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
761 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
767 EXPORT_SYMBOL_GPL(md_raid1_congested
);
769 static int raid1_congested(void *data
, int bits
)
771 struct mddev
*mddev
= data
;
773 return mddev_congested(mddev
, bits
) ||
774 md_raid1_congested(mddev
, bits
);
777 static void flush_pending_writes(struct r1conf
*conf
)
779 /* Any writes that have been queued but are awaiting
780 * bitmap updates get flushed here.
782 spin_lock_irq(&conf
->device_lock
);
784 if (conf
->pending_bio_list
.head
) {
786 bio
= bio_list_get(&conf
->pending_bio_list
);
787 conf
->pending_count
= 0;
788 spin_unlock_irq(&conf
->device_lock
);
789 /* flush any pending bitmap writes to
790 * disk before proceeding w/ I/O */
791 bitmap_unplug(conf
->mddev
->bitmap
);
792 wake_up(&conf
->wait_barrier
);
794 while (bio
) { /* submit pending writes */
795 struct bio
*next
= bio
->bi_next
;
797 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
798 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
802 generic_make_request(bio
);
806 spin_unlock_irq(&conf
->device_lock
);
810 * Sometimes we need to suspend IO while we do something else,
811 * either some resync/recovery, or reconfigure the array.
812 * To do this we raise a 'barrier'.
813 * The 'barrier' is a counter that can be raised multiple times
814 * to count how many activities are happening which preclude
816 * We can only raise the barrier if there is no pending IO.
817 * i.e. if nr_pending == 0.
818 * We choose only to raise the barrier if no-one is waiting for the
819 * barrier to go down. This means that as soon as an IO request
820 * is ready, no other operations which require a barrier will start
821 * until the IO request has had a chance.
823 * So: regular IO calls 'wait_barrier'. When that returns there
824 * is no backgroup IO happening, It must arrange to call
825 * allow_barrier when it has finished its IO.
826 * backgroup IO calls must call raise_barrier. Once that returns
827 * there is no normal IO happeing. It must arrange to call
828 * lower_barrier when the particular background IO completes.
830 static void raise_barrier(struct r1conf
*conf
)
832 spin_lock_irq(&conf
->resync_lock
);
834 /* Wait until no block IO is waiting */
835 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
838 /* block any new IO from starting */
841 /* For these conditions we must wait:
842 * A: while the array is in frozen state
843 * B: while barrier >= RESYNC_DEPTH, meaning resync reach
844 * the max count which allowed.
845 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
846 * next resync will reach to the window which normal bios are
849 wait_event_lock_irq(conf
->wait_barrier
,
850 !conf
->array_frozen
&&
851 conf
->barrier
< RESYNC_DEPTH
&&
852 (conf
->start_next_window
>=
853 conf
->next_resync
+ RESYNC_SECTORS
),
856 spin_unlock_irq(&conf
->resync_lock
);
859 static void lower_barrier(struct r1conf
*conf
)
862 BUG_ON(conf
->barrier
<= 0);
863 spin_lock_irqsave(&conf
->resync_lock
, flags
);
865 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
866 wake_up(&conf
->wait_barrier
);
869 static bool need_to_wait_for_sync(struct r1conf
*conf
, struct bio
*bio
)
873 if (conf
->array_frozen
|| !bio
)
875 else if (conf
->barrier
&& bio_data_dir(bio
) == WRITE
) {
876 if (conf
->next_resync
< RESYNC_WINDOW_SECTORS
)
878 else if ((conf
->next_resync
- RESYNC_WINDOW_SECTORS
879 >= bio_end_sector(bio
)) ||
880 (conf
->next_resync
+ NEXT_NORMALIO_DISTANCE
881 <= bio
->bi_iter
.bi_sector
))
890 static sector_t
wait_barrier(struct r1conf
*conf
, struct bio
*bio
)
894 spin_lock_irq(&conf
->resync_lock
);
895 if (need_to_wait_for_sync(conf
, bio
)) {
897 /* Wait for the barrier to drop.
898 * However if there are already pending
899 * requests (preventing the barrier from
900 * rising completely), and the
901 * pre-process bio queue isn't empty,
902 * then don't wait, as we need to empty
903 * that queue to get the nr_pending
906 wait_event_lock_irq(conf
->wait_barrier
,
907 !conf
->array_frozen
&&
909 ((conf
->start_next_window
<
910 conf
->next_resync
+ RESYNC_SECTORS
) &&
912 !bio_list_empty(current
->bio_list
))),
917 if (bio
&& bio_data_dir(bio
) == WRITE
) {
918 if (conf
->next_resync
+ NEXT_NORMALIO_DISTANCE
919 <= bio
->bi_iter
.bi_sector
) {
920 if (conf
->start_next_window
== MaxSector
)
921 conf
->start_next_window
=
923 NEXT_NORMALIO_DISTANCE
;
925 if ((conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
)
926 <= bio
->bi_iter
.bi_sector
)
927 conf
->next_window_requests
++;
929 conf
->current_window_requests
++;
930 sector
= conf
->start_next_window
;
935 spin_unlock_irq(&conf
->resync_lock
);
939 static void allow_barrier(struct r1conf
*conf
, sector_t start_next_window
,
944 spin_lock_irqsave(&conf
->resync_lock
, flags
);
946 if (start_next_window
) {
947 if (start_next_window
== conf
->start_next_window
) {
948 if (conf
->start_next_window
+ NEXT_NORMALIO_DISTANCE
950 conf
->next_window_requests
--;
952 conf
->current_window_requests
--;
954 conf
->current_window_requests
--;
956 if (!conf
->current_window_requests
) {
957 if (conf
->next_window_requests
) {
958 conf
->current_window_requests
=
959 conf
->next_window_requests
;
960 conf
->next_window_requests
= 0;
961 conf
->start_next_window
+=
962 NEXT_NORMALIO_DISTANCE
;
964 conf
->start_next_window
= MaxSector
;
967 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
968 wake_up(&conf
->wait_barrier
);
971 static void freeze_array(struct r1conf
*conf
, int extra
)
973 /* stop syncio and normal IO and wait for everything to
975 * We wait until nr_pending match nr_queued+extra
976 * This is called in the context of one normal IO request
977 * that has failed. Thus any sync request that might be pending
978 * will be blocked by nr_pending, and we need to wait for
979 * pending IO requests to complete or be queued for re-try.
980 * Thus the number queued (nr_queued) plus this request (extra)
981 * must match the number of pending IOs (nr_pending) before
984 spin_lock_irq(&conf
->resync_lock
);
985 conf
->array_frozen
= 1;
986 wait_event_lock_irq_cmd(conf
->wait_barrier
,
987 conf
->nr_pending
== conf
->nr_queued
+extra
,
989 flush_pending_writes(conf
));
990 spin_unlock_irq(&conf
->resync_lock
);
992 static void unfreeze_array(struct r1conf
*conf
)
994 /* reverse the effect of the freeze */
995 spin_lock_irq(&conf
->resync_lock
);
996 conf
->array_frozen
= 0;
997 wake_up(&conf
->wait_barrier
);
998 spin_unlock_irq(&conf
->resync_lock
);
1002 /* duplicate the data pages for behind I/O
1004 static void alloc_behind_pages(struct bio
*bio
, struct r1bio
*r1_bio
)
1007 struct bio_vec
*bvec
;
1008 struct bio_vec
*bvecs
= kzalloc(bio
->bi_vcnt
* sizeof(struct bio_vec
),
1010 if (unlikely(!bvecs
))
1013 bio_for_each_segment_all(bvec
, bio
, i
) {
1015 bvecs
[i
].bv_page
= alloc_page(GFP_NOIO
);
1016 if (unlikely(!bvecs
[i
].bv_page
))
1018 memcpy(kmap(bvecs
[i
].bv_page
) + bvec
->bv_offset
,
1019 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
1020 kunmap(bvecs
[i
].bv_page
);
1021 kunmap(bvec
->bv_page
);
1023 r1_bio
->behind_bvecs
= bvecs
;
1024 r1_bio
->behind_page_count
= bio
->bi_vcnt
;
1025 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
1029 for (i
= 0; i
< bio
->bi_vcnt
; i
++)
1030 if (bvecs
[i
].bv_page
)
1031 put_page(bvecs
[i
].bv_page
);
1033 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1034 bio
->bi_iter
.bi_size
);
1037 struct raid1_plug_cb
{
1038 struct blk_plug_cb cb
;
1039 struct bio_list pending
;
1043 static void raid1_unplug(struct blk_plug_cb
*cb
, bool from_schedule
)
1045 struct raid1_plug_cb
*plug
= container_of(cb
, struct raid1_plug_cb
,
1047 struct mddev
*mddev
= plug
->cb
.data
;
1048 struct r1conf
*conf
= mddev
->private;
1051 if (from_schedule
|| current
->bio_list
) {
1052 spin_lock_irq(&conf
->device_lock
);
1053 bio_list_merge(&conf
->pending_bio_list
, &plug
->pending
);
1054 conf
->pending_count
+= plug
->pending_cnt
;
1055 spin_unlock_irq(&conf
->device_lock
);
1056 wake_up(&conf
->wait_barrier
);
1057 md_wakeup_thread(mddev
->thread
);
1062 /* we aren't scheduling, so we can do the write-out directly. */
1063 bio
= bio_list_get(&plug
->pending
);
1064 bitmap_unplug(mddev
->bitmap
);
1065 wake_up(&conf
->wait_barrier
);
1067 while (bio
) { /* submit pending writes */
1068 struct bio
*next
= bio
->bi_next
;
1069 bio
->bi_next
= NULL
;
1070 if (unlikely((bio
->bi_rw
& REQ_DISCARD
) &&
1071 !blk_queue_discard(bdev_get_queue(bio
->bi_bdev
))))
1072 /* Just ignore it */
1075 generic_make_request(bio
);
1081 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
1083 struct r1conf
*conf
= mddev
->private;
1084 struct raid1_info
*mirror
;
1085 struct r1bio
*r1_bio
;
1086 struct bio
*read_bio
;
1088 struct bitmap
*bitmap
;
1089 unsigned long flags
;
1090 const int rw
= bio_data_dir(bio
);
1091 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
1092 const unsigned long do_flush_fua
= (bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
));
1093 const unsigned long do_discard
= (bio
->bi_rw
1094 & (REQ_DISCARD
| REQ_SECURE
));
1095 const unsigned long do_same
= (bio
->bi_rw
& REQ_WRITE_SAME
);
1096 struct md_rdev
*blocked_rdev
;
1097 struct blk_plug_cb
*cb
;
1098 struct raid1_plug_cb
*plug
= NULL
;
1100 int sectors_handled
;
1102 sector_t start_next_window
;
1105 * Register the new request and wait if the reconstruction
1106 * thread has put up a bar for new requests.
1107 * Continue immediately if no resync is active currently.
1110 md_write_start(mddev
, bio
); /* wait on superblock update early */
1112 if (bio_data_dir(bio
) == WRITE
&&
1113 bio_end_sector(bio
) > mddev
->suspend_lo
&&
1114 bio
->bi_iter
.bi_sector
< mddev
->suspend_hi
) {
1115 /* As the suspend_* range is controlled by
1116 * userspace, we want an interruptible
1121 flush_signals(current
);
1122 prepare_to_wait(&conf
->wait_barrier
,
1123 &w
, TASK_INTERRUPTIBLE
);
1124 if (bio_end_sector(bio
) <= mddev
->suspend_lo
||
1125 bio
->bi_iter
.bi_sector
>= mddev
->suspend_hi
)
1129 finish_wait(&conf
->wait_barrier
, &w
);
1132 start_next_window
= wait_barrier(conf
, bio
);
1134 bitmap
= mddev
->bitmap
;
1137 * make_request() can abort the operation when READA is being
1138 * used and no empty request is available.
1141 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1143 r1_bio
->master_bio
= bio
;
1144 r1_bio
->sectors
= bio_sectors(bio
);
1146 r1_bio
->mddev
= mddev
;
1147 r1_bio
->sector
= bio
->bi_iter
.bi_sector
;
1149 /* We might need to issue multiple reads to different
1150 * devices if there are bad blocks around, so we keep
1151 * track of the number of reads in bio->bi_phys_segments.
1152 * If this is 0, there is only one r1_bio and no locking
1153 * will be needed when requests complete. If it is
1154 * non-zero, then it is the number of not-completed requests.
1156 bio
->bi_phys_segments
= 0;
1157 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
1161 * read balancing logic:
1166 rdisk
= read_balance(conf
, r1_bio
, &max_sectors
);
1169 /* couldn't find anywhere to read from */
1170 raid_end_bio_io(r1_bio
);
1173 mirror
= conf
->mirrors
+ rdisk
;
1175 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
1177 /* Reading from a write-mostly device must
1178 * take care not to over-take any writes
1181 wait_event(bitmap
->behind_wait
,
1182 atomic_read(&bitmap
->behind_writes
) == 0);
1184 r1_bio
->read_disk
= rdisk
;
1185 r1_bio
->start_next_window
= 0;
1187 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1188 bio_trim(read_bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
1191 r1_bio
->bios
[rdisk
] = read_bio
;
1193 read_bio
->bi_iter
.bi_sector
= r1_bio
->sector
+
1194 mirror
->rdev
->data_offset
;
1195 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
1196 read_bio
->bi_end_io
= raid1_end_read_request
;
1197 read_bio
->bi_rw
= READ
| do_sync
;
1198 read_bio
->bi_private
= r1_bio
;
1200 if (max_sectors
< r1_bio
->sectors
) {
1201 /* could not read all from this device, so we will
1202 * need another r1_bio.
1205 sectors_handled
= (r1_bio
->sector
+ max_sectors
1206 - bio
->bi_iter
.bi_sector
);
1207 r1_bio
->sectors
= max_sectors
;
1208 spin_lock_irq(&conf
->device_lock
);
1209 if (bio
->bi_phys_segments
== 0)
1210 bio
->bi_phys_segments
= 2;
1212 bio
->bi_phys_segments
++;
1213 spin_unlock_irq(&conf
->device_lock
);
1214 /* Cannot call generic_make_request directly
1215 * as that will be queued in __make_request
1216 * and subsequent mempool_alloc might block waiting
1217 * for it. So hand bio over to raid1d.
1219 reschedule_retry(r1_bio
);
1221 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1223 r1_bio
->master_bio
= bio
;
1224 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1226 r1_bio
->mddev
= mddev
;
1227 r1_bio
->sector
= bio
->bi_iter
.bi_sector
+
1231 generic_make_request(read_bio
);
1238 if (conf
->pending_count
>= max_queued_requests
) {
1239 md_wakeup_thread(mddev
->thread
);
1240 wait_event(conf
->wait_barrier
,
1241 conf
->pending_count
< max_queued_requests
);
1243 /* first select target devices under rcu_lock and
1244 * inc refcount on their rdev. Record them by setting
1246 * If there are known/acknowledged bad blocks on any device on
1247 * which we have seen a write error, we want to avoid writing those
1249 * This potentially requires several writes to write around
1250 * the bad blocks. Each set of writes gets it's own r1bio
1251 * with a set of bios attached.
1254 disks
= conf
->raid_disks
* 2;
1256 r1_bio
->start_next_window
= start_next_window
;
1257 blocked_rdev
= NULL
;
1259 max_sectors
= r1_bio
->sectors
;
1260 for (i
= 0; i
< disks
; i
++) {
1261 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1262 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
1263 atomic_inc(&rdev
->nr_pending
);
1264 blocked_rdev
= rdev
;
1267 r1_bio
->bios
[i
] = NULL
;
1268 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)
1269 || test_bit(Unmerged
, &rdev
->flags
)) {
1270 if (i
< conf
->raid_disks
)
1271 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1275 atomic_inc(&rdev
->nr_pending
);
1276 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1281 is_bad
= is_badblock(rdev
, r1_bio
->sector
,
1283 &first_bad
, &bad_sectors
);
1285 /* mustn't write here until the bad block is
1287 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1288 blocked_rdev
= rdev
;
1291 if (is_bad
&& first_bad
<= r1_bio
->sector
) {
1292 /* Cannot write here at all */
1293 bad_sectors
-= (r1_bio
->sector
- first_bad
);
1294 if (bad_sectors
< max_sectors
)
1295 /* mustn't write more than bad_sectors
1296 * to other devices yet
1298 max_sectors
= bad_sectors
;
1299 rdev_dec_pending(rdev
, mddev
);
1300 /* We don't set R1BIO_Degraded as that
1301 * only applies if the disk is
1302 * missing, so it might be re-added,
1303 * and we want to know to recover this
1305 * In this case the device is here,
1306 * and the fact that this chunk is not
1307 * in-sync is recorded in the bad
1313 int good_sectors
= first_bad
- r1_bio
->sector
;
1314 if (good_sectors
< max_sectors
)
1315 max_sectors
= good_sectors
;
1318 r1_bio
->bios
[i
] = bio
;
1322 if (unlikely(blocked_rdev
)) {
1323 /* Wait for this device to become unblocked */
1325 sector_t old
= start_next_window
;
1327 for (j
= 0; j
< i
; j
++)
1328 if (r1_bio
->bios
[j
])
1329 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
1331 allow_barrier(conf
, start_next_window
, bio
->bi_iter
.bi_sector
);
1332 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1333 start_next_window
= wait_barrier(conf
, bio
);
1335 * We must make sure the multi r1bios of bio have
1336 * the same value of bi_phys_segments
1338 if (bio
->bi_phys_segments
&& old
&&
1339 old
!= start_next_window
)
1340 /* Wait for the former r1bio(s) to complete */
1341 wait_event(conf
->wait_barrier
,
1342 bio
->bi_phys_segments
== 1);
1346 if (max_sectors
< r1_bio
->sectors
) {
1347 /* We are splitting this write into multiple parts, so
1348 * we need to prepare for allocating another r1_bio.
1350 r1_bio
->sectors
= max_sectors
;
1351 spin_lock_irq(&conf
->device_lock
);
1352 if (bio
->bi_phys_segments
== 0)
1353 bio
->bi_phys_segments
= 2;
1355 bio
->bi_phys_segments
++;
1356 spin_unlock_irq(&conf
->device_lock
);
1358 sectors_handled
= r1_bio
->sector
+ max_sectors
- bio
->bi_iter
.bi_sector
;
1360 atomic_set(&r1_bio
->remaining
, 1);
1361 atomic_set(&r1_bio
->behind_remaining
, 0);
1364 for (i
= 0; i
< disks
; i
++) {
1366 if (!r1_bio
->bios
[i
])
1369 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1370 bio_trim(mbio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
, max_sectors
);
1374 * Not if there are too many, or cannot
1375 * allocate memory, or a reader on WriteMostly
1376 * is waiting for behind writes to flush */
1378 (atomic_read(&bitmap
->behind_writes
)
1379 < mddev
->bitmap_info
.max_write_behind
) &&
1380 !waitqueue_active(&bitmap
->behind_wait
))
1381 alloc_behind_pages(mbio
, r1_bio
);
1383 bitmap_startwrite(bitmap
, r1_bio
->sector
,
1385 test_bit(R1BIO_BehindIO
,
1389 if (r1_bio
->behind_bvecs
) {
1390 struct bio_vec
*bvec
;
1394 * We trimmed the bio, so _all is legit
1396 bio_for_each_segment_all(bvec
, mbio
, j
)
1397 bvec
->bv_page
= r1_bio
->behind_bvecs
[j
].bv_page
;
1398 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
1399 atomic_inc(&r1_bio
->behind_remaining
);
1402 r1_bio
->bios
[i
] = mbio
;
1404 mbio
->bi_iter
.bi_sector
= (r1_bio
->sector
+
1405 conf
->mirrors
[i
].rdev
->data_offset
);
1406 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1407 mbio
->bi_end_io
= raid1_end_write_request
;
1409 WRITE
| do_flush_fua
| do_sync
| do_discard
| do_same
;
1410 mbio
->bi_private
= r1_bio
;
1412 atomic_inc(&r1_bio
->remaining
);
1414 cb
= blk_check_plugged(raid1_unplug
, mddev
, sizeof(*plug
));
1416 plug
= container_of(cb
, struct raid1_plug_cb
, cb
);
1419 spin_lock_irqsave(&conf
->device_lock
, flags
);
1421 bio_list_add(&plug
->pending
, mbio
);
1422 plug
->pending_cnt
++;
1424 bio_list_add(&conf
->pending_bio_list
, mbio
);
1425 conf
->pending_count
++;
1427 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1429 md_wakeup_thread(mddev
->thread
);
1431 /* Mustn't call r1_bio_write_done before this next test,
1432 * as it could result in the bio being freed.
1434 if (sectors_handled
< bio_sectors(bio
)) {
1435 r1_bio_write_done(r1_bio
);
1436 /* We need another r1_bio. It has already been counted
1437 * in bio->bi_phys_segments
1439 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1440 r1_bio
->master_bio
= bio
;
1441 r1_bio
->sectors
= bio_sectors(bio
) - sectors_handled
;
1443 r1_bio
->mddev
= mddev
;
1444 r1_bio
->sector
= bio
->bi_iter
.bi_sector
+ sectors_handled
;
1448 r1_bio_write_done(r1_bio
);
1450 /* In case raid1d snuck in to freeze_array */
1451 wake_up(&conf
->wait_barrier
);
1454 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1456 struct r1conf
*conf
= mddev
->private;
1459 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1460 conf
->raid_disks
- mddev
->degraded
);
1462 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1463 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1464 seq_printf(seq
, "%s",
1465 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1468 seq_printf(seq
, "]");
1472 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1474 char b
[BDEVNAME_SIZE
];
1475 struct r1conf
*conf
= mddev
->private;
1478 * If it is not operational, then we have already marked it as dead
1479 * else if it is the last working disks, ignore the error, let the
1480 * next level up know.
1481 * else mark the drive as failed
1483 if (test_bit(In_sync
, &rdev
->flags
)
1484 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1486 * Don't fail the drive, act as though we were just a
1487 * normal single drive.
1488 * However don't try a recovery from this drive as
1489 * it is very likely to fail.
1491 conf
->recovery_disabled
= mddev
->recovery_disabled
;
1494 set_bit(Blocked
, &rdev
->flags
);
1495 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1496 unsigned long flags
;
1497 spin_lock_irqsave(&conf
->device_lock
, flags
);
1499 set_bit(Faulty
, &rdev
->flags
);
1500 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1502 set_bit(Faulty
, &rdev
->flags
);
1504 * if recovery is running, make sure it aborts.
1506 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1507 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1509 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1510 "md/raid1:%s: Operation continuing on %d devices.\n",
1511 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1512 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1515 static void print_conf(struct r1conf
*conf
)
1519 printk(KERN_DEBUG
"RAID1 conf printout:\n");
1521 printk(KERN_DEBUG
"(!conf)\n");
1524 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1528 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1529 char b
[BDEVNAME_SIZE
];
1530 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1532 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1533 i
, !test_bit(In_sync
, &rdev
->flags
),
1534 !test_bit(Faulty
, &rdev
->flags
),
1535 bdevname(rdev
->bdev
,b
));
1540 static void close_sync(struct r1conf
*conf
)
1542 wait_barrier(conf
, NULL
);
1543 allow_barrier(conf
, 0, 0);
1545 mempool_destroy(conf
->r1buf_pool
);
1546 conf
->r1buf_pool
= NULL
;
1548 spin_lock_irq(&conf
->resync_lock
);
1549 conf
->next_resync
= 0;
1550 conf
->start_next_window
= MaxSector
;
1551 conf
->current_window_requests
+=
1552 conf
->next_window_requests
;
1553 conf
->next_window_requests
= 0;
1554 spin_unlock_irq(&conf
->resync_lock
);
1557 static int raid1_spare_active(struct mddev
*mddev
)
1560 struct r1conf
*conf
= mddev
->private;
1562 unsigned long flags
;
1565 * Find all failed disks within the RAID1 configuration
1566 * and mark them readable.
1567 * Called under mddev lock, so rcu protection not needed.
1569 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1570 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1571 struct md_rdev
*repl
= conf
->mirrors
[conf
->raid_disks
+ i
].rdev
;
1573 && repl
->recovery_offset
== MaxSector
1574 && !test_bit(Faulty
, &repl
->flags
)
1575 && !test_and_set_bit(In_sync
, &repl
->flags
)) {
1576 /* replacement has just become active */
1578 !test_and_clear_bit(In_sync
, &rdev
->flags
))
1581 /* Replaced device not technically
1582 * faulty, but we need to be sure
1583 * it gets removed and never re-added
1585 set_bit(Faulty
, &rdev
->flags
);
1586 sysfs_notify_dirent_safe(
1591 && rdev
->recovery_offset
== MaxSector
1592 && !test_bit(Faulty
, &rdev
->flags
)
1593 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1595 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
1598 spin_lock_irqsave(&conf
->device_lock
, flags
);
1599 mddev
->degraded
-= count
;
1600 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1607 static int raid1_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1609 struct r1conf
*conf
= mddev
->private;
1612 struct raid1_info
*p
;
1614 int last
= conf
->raid_disks
- 1;
1615 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1617 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
1620 if (rdev
->raid_disk
>= 0)
1621 first
= last
= rdev
->raid_disk
;
1623 if (q
->merge_bvec_fn
) {
1624 set_bit(Unmerged
, &rdev
->flags
);
1625 mddev
->merge_check_needed
= 1;
1628 for (mirror
= first
; mirror
<= last
; mirror
++) {
1629 p
= conf
->mirrors
+mirror
;
1633 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1634 rdev
->data_offset
<< 9);
1636 p
->head_position
= 0;
1637 rdev
->raid_disk
= mirror
;
1639 /* As all devices are equivalent, we don't need a full recovery
1640 * if this was recently any drive of the array
1642 if (rdev
->saved_raid_disk
< 0)
1644 rcu_assign_pointer(p
->rdev
, rdev
);
1647 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
1648 p
[conf
->raid_disks
].rdev
== NULL
) {
1649 /* Add this device as a replacement */
1650 clear_bit(In_sync
, &rdev
->flags
);
1651 set_bit(Replacement
, &rdev
->flags
);
1652 rdev
->raid_disk
= mirror
;
1655 rcu_assign_pointer(p
[conf
->raid_disks
].rdev
, rdev
);
1659 if (err
== 0 && test_bit(Unmerged
, &rdev
->flags
)) {
1660 /* Some requests might not have seen this new
1661 * merge_bvec_fn. We must wait for them to complete
1662 * before merging the device fully.
1663 * First we make sure any code which has tested
1664 * our function has submitted the request, then
1665 * we wait for all outstanding requests to complete.
1667 synchronize_sched();
1668 freeze_array(conf
, 0);
1669 unfreeze_array(conf
);
1670 clear_bit(Unmerged
, &rdev
->flags
);
1672 md_integrity_add_rdev(rdev
, mddev
);
1673 if (mddev
->queue
&& blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
1674 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
, mddev
->queue
);
1679 static int raid1_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1681 struct r1conf
*conf
= mddev
->private;
1683 int number
= rdev
->raid_disk
;
1684 struct raid1_info
*p
= conf
->mirrors
+ number
;
1686 if (rdev
!= p
->rdev
)
1687 p
= conf
->mirrors
+ conf
->raid_disks
+ number
;
1690 if (rdev
== p
->rdev
) {
1691 if (test_bit(In_sync
, &rdev
->flags
) ||
1692 atomic_read(&rdev
->nr_pending
)) {
1696 /* Only remove non-faulty devices if recovery
1699 if (!test_bit(Faulty
, &rdev
->flags
) &&
1700 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
1701 mddev
->degraded
< conf
->raid_disks
) {
1707 if (atomic_read(&rdev
->nr_pending
)) {
1708 /* lost the race, try later */
1712 } else if (conf
->mirrors
[conf
->raid_disks
+ number
].rdev
) {
1713 /* We just removed a device that is being replaced.
1714 * Move down the replacement. We drain all IO before
1715 * doing this to avoid confusion.
1717 struct md_rdev
*repl
=
1718 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
;
1719 freeze_array(conf
, 0);
1720 clear_bit(Replacement
, &repl
->flags
);
1722 conf
->mirrors
[conf
->raid_disks
+ number
].rdev
= NULL
;
1723 unfreeze_array(conf
);
1724 clear_bit(WantReplacement
, &rdev
->flags
);
1726 clear_bit(WantReplacement
, &rdev
->flags
);
1727 err
= md_integrity_register(mddev
);
1736 static void end_sync_read(struct bio
*bio
, int error
)
1738 struct r1bio
*r1_bio
= bio
->bi_private
;
1740 update_head_pos(r1_bio
->read_disk
, r1_bio
);
1743 * we have read a block, now it needs to be re-written,
1744 * or re-read if the read failed.
1745 * We don't do much here, just schedule handling by raid1d
1747 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1748 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1750 if (atomic_dec_and_test(&r1_bio
->remaining
))
1751 reschedule_retry(r1_bio
);
1754 static void end_sync_write(struct bio
*bio
, int error
)
1756 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1757 struct r1bio
*r1_bio
= bio
->bi_private
;
1758 struct mddev
*mddev
= r1_bio
->mddev
;
1759 struct r1conf
*conf
= mddev
->private;
1764 mirror
= find_bio_disk(r1_bio
, bio
);
1767 sector_t sync_blocks
= 0;
1768 sector_t s
= r1_bio
->sector
;
1769 long sectors_to_go
= r1_bio
->sectors
;
1770 /* make sure these bits doesn't get cleared. */
1772 bitmap_end_sync(mddev
->bitmap
, s
,
1775 sectors_to_go
-= sync_blocks
;
1776 } while (sectors_to_go
> 0);
1777 set_bit(WriteErrorSeen
,
1778 &conf
->mirrors
[mirror
].rdev
->flags
);
1779 if (!test_and_set_bit(WantReplacement
,
1780 &conf
->mirrors
[mirror
].rdev
->flags
))
1781 set_bit(MD_RECOVERY_NEEDED
, &
1783 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
1784 } else if (is_badblock(conf
->mirrors
[mirror
].rdev
,
1787 &first_bad
, &bad_sectors
) &&
1788 !is_badblock(conf
->mirrors
[r1_bio
->read_disk
].rdev
,
1791 &first_bad
, &bad_sectors
)
1793 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
1795 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1796 int s
= r1_bio
->sectors
;
1797 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
1798 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1799 reschedule_retry(r1_bio
);
1802 md_done_sync(mddev
, s
, uptodate
);
1807 static int r1_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1808 int sectors
, struct page
*page
, int rw
)
1810 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
1814 set_bit(WriteErrorSeen
, &rdev
->flags
);
1815 if (!test_and_set_bit(WantReplacement
,
1817 set_bit(MD_RECOVERY_NEEDED
, &
1818 rdev
->mddev
->recovery
);
1820 /* need to record an error - either for the block or the device */
1821 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1822 md_error(rdev
->mddev
, rdev
);
1826 static int fix_sync_read_error(struct r1bio
*r1_bio
)
1828 /* Try some synchronous reads of other devices to get
1829 * good data, much like with normal read errors. Only
1830 * read into the pages we already have so we don't
1831 * need to re-issue the read request.
1832 * We don't need to freeze the array, because being in an
1833 * active sync request, there is no normal IO, and
1834 * no overlapping syncs.
1835 * We don't need to check is_badblock() again as we
1836 * made sure that anything with a bad block in range
1837 * will have bi_end_io clear.
1839 struct mddev
*mddev
= r1_bio
->mddev
;
1840 struct r1conf
*conf
= mddev
->private;
1841 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1842 sector_t sect
= r1_bio
->sector
;
1843 int sectors
= r1_bio
->sectors
;
1848 int d
= r1_bio
->read_disk
;
1850 struct md_rdev
*rdev
;
1853 if (s
> (PAGE_SIZE
>>9))
1856 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1857 /* No rcu protection needed here devices
1858 * can only be removed when no resync is
1859 * active, and resync is currently active
1861 rdev
= conf
->mirrors
[d
].rdev
;
1862 if (sync_page_io(rdev
, sect
, s
<<9,
1863 bio
->bi_io_vec
[idx
].bv_page
,
1870 if (d
== conf
->raid_disks
* 2)
1872 } while (!success
&& d
!= r1_bio
->read_disk
);
1875 char b
[BDEVNAME_SIZE
];
1877 /* Cannot read from anywhere, this block is lost.
1878 * Record a bad block on each device. If that doesn't
1879 * work just disable and interrupt the recovery.
1880 * Don't fail devices as that won't really help.
1882 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O read error"
1883 " for block %llu\n",
1885 bdevname(bio
->bi_bdev
, b
),
1886 (unsigned long long)r1_bio
->sector
);
1887 for (d
= 0; d
< conf
->raid_disks
* 2; d
++) {
1888 rdev
= conf
->mirrors
[d
].rdev
;
1889 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
1891 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
1895 conf
->recovery_disabled
=
1896 mddev
->recovery_disabled
;
1897 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1898 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1910 /* write it back and re-read */
1911 while (d
!= r1_bio
->read_disk
) {
1913 d
= conf
->raid_disks
* 2;
1915 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1917 rdev
= conf
->mirrors
[d
].rdev
;
1918 if (r1_sync_page_io(rdev
, sect
, s
,
1919 bio
->bi_io_vec
[idx
].bv_page
,
1921 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
1922 rdev_dec_pending(rdev
, mddev
);
1926 while (d
!= r1_bio
->read_disk
) {
1928 d
= conf
->raid_disks
* 2;
1930 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1932 rdev
= conf
->mirrors
[d
].rdev
;
1933 if (r1_sync_page_io(rdev
, sect
, s
,
1934 bio
->bi_io_vec
[idx
].bv_page
,
1936 atomic_add(s
, &rdev
->corrected_errors
);
1942 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1943 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1947 static int process_checks(struct r1bio
*r1_bio
)
1949 /* We have read all readable devices. If we haven't
1950 * got the block, then there is no hope left.
1951 * If we have, then we want to do a comparison
1952 * and skip the write if everything is the same.
1953 * If any blocks failed to read, then we need to
1954 * attempt an over-write
1956 struct mddev
*mddev
= r1_bio
->mddev
;
1957 struct r1conf
*conf
= mddev
->private;
1962 /* Fix variable parts of all bios */
1963 vcnt
= (r1_bio
->sectors
+ PAGE_SIZE
/ 512 - 1) >> (PAGE_SHIFT
- 9);
1964 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
1968 struct bio
*b
= r1_bio
->bios
[i
];
1969 if (b
->bi_end_io
!= end_sync_read
)
1971 /* fixup the bio for reuse, but preserve BIO_UPTODATE */
1972 uptodate
= test_bit(BIO_UPTODATE
, &b
->bi_flags
);
1975 clear_bit(BIO_UPTODATE
, &b
->bi_flags
);
1977 b
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
1978 b
->bi_iter
.bi_sector
= r1_bio
->sector
+
1979 conf
->mirrors
[i
].rdev
->data_offset
;
1980 b
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1981 b
->bi_end_io
= end_sync_read
;
1982 b
->bi_private
= r1_bio
;
1984 size
= b
->bi_iter
.bi_size
;
1985 for (j
= 0; j
< vcnt
; j
++) {
1987 bi
= &b
->bi_io_vec
[j
];
1989 if (size
> PAGE_SIZE
)
1990 bi
->bv_len
= PAGE_SIZE
;
1996 for (primary
= 0; primary
< conf
->raid_disks
* 2; primary
++)
1997 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
1998 test_bit(BIO_UPTODATE
, &r1_bio
->bios
[primary
]->bi_flags
)) {
1999 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
2000 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
2003 r1_bio
->read_disk
= primary
;
2004 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2006 struct bio
*pbio
= r1_bio
->bios
[primary
];
2007 struct bio
*sbio
= r1_bio
->bios
[i
];
2008 int uptodate
= test_bit(BIO_UPTODATE
, &sbio
->bi_flags
);
2010 if (sbio
->bi_end_io
!= end_sync_read
)
2012 /* Now we can 'fixup' the BIO_UPTODATE flag */
2013 set_bit(BIO_UPTODATE
, &sbio
->bi_flags
);
2016 for (j
= vcnt
; j
-- ; ) {
2018 p
= pbio
->bi_io_vec
[j
].bv_page
;
2019 s
= sbio
->bi_io_vec
[j
].bv_page
;
2020 if (memcmp(page_address(p
),
2022 sbio
->bi_io_vec
[j
].bv_len
))
2028 atomic64_add(r1_bio
->sectors
, &mddev
->resync_mismatches
);
2029 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
2031 /* No need to write to this device. */
2032 sbio
->bi_end_io
= NULL
;
2033 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
2037 bio_copy_data(sbio
, pbio
);
2042 static void sync_request_write(struct mddev
*mddev
, struct r1bio
*r1_bio
)
2044 struct r1conf
*conf
= mddev
->private;
2046 int disks
= conf
->raid_disks
* 2;
2047 struct bio
*bio
, *wbio
;
2049 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2051 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
2052 /* ouch - failed to read all of that. */
2053 if (!fix_sync_read_error(r1_bio
))
2056 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2057 if (process_checks(r1_bio
) < 0)
2062 atomic_set(&r1_bio
->remaining
, 1);
2063 for (i
= 0; i
< disks
; i
++) {
2064 wbio
= r1_bio
->bios
[i
];
2065 if (wbio
->bi_end_io
== NULL
||
2066 (wbio
->bi_end_io
== end_sync_read
&&
2067 (i
== r1_bio
->read_disk
||
2068 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
2071 wbio
->bi_rw
= WRITE
;
2072 wbio
->bi_end_io
= end_sync_write
;
2073 atomic_inc(&r1_bio
->remaining
);
2074 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, bio_sectors(wbio
));
2076 generic_make_request(wbio
);
2079 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
2080 /* if we're here, all write(s) have completed, so clean up */
2081 int s
= r1_bio
->sectors
;
2082 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2083 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2084 reschedule_retry(r1_bio
);
2087 md_done_sync(mddev
, s
, 1);
2093 * This is a kernel thread which:
2095 * 1. Retries failed read operations on working mirrors.
2096 * 2. Updates the raid superblock when problems encounter.
2097 * 3. Performs writes following reads for array synchronising.
2100 static void fix_read_error(struct r1conf
*conf
, int read_disk
,
2101 sector_t sect
, int sectors
)
2103 struct mddev
*mddev
= conf
->mddev
;
2109 struct md_rdev
*rdev
;
2111 if (s
> (PAGE_SIZE
>>9))
2115 /* Note: no rcu protection needed here
2116 * as this is synchronous in the raid1d thread
2117 * which is the thread that might remove
2118 * a device. If raid1d ever becomes multi-threaded....
2123 rdev
= conf
->mirrors
[d
].rdev
;
2125 (test_bit(In_sync
, &rdev
->flags
) ||
2126 (!test_bit(Faulty
, &rdev
->flags
) &&
2127 rdev
->recovery_offset
>= sect
+ s
)) &&
2128 is_badblock(rdev
, sect
, s
,
2129 &first_bad
, &bad_sectors
) == 0 &&
2130 sync_page_io(rdev
, sect
, s
<<9,
2131 conf
->tmppage
, READ
, false))
2135 if (d
== conf
->raid_disks
* 2)
2138 } while (!success
&& d
!= read_disk
);
2141 /* Cannot read from anywhere - mark it bad */
2142 struct md_rdev
*rdev
= conf
->mirrors
[read_disk
].rdev
;
2143 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
2144 md_error(mddev
, rdev
);
2147 /* write it back and re-read */
2149 while (d
!= read_disk
) {
2151 d
= conf
->raid_disks
* 2;
2153 rdev
= conf
->mirrors
[d
].rdev
;
2155 test_bit(In_sync
, &rdev
->flags
))
2156 r1_sync_page_io(rdev
, sect
, s
,
2157 conf
->tmppage
, WRITE
);
2160 while (d
!= read_disk
) {
2161 char b
[BDEVNAME_SIZE
];
2163 d
= conf
->raid_disks
* 2;
2165 rdev
= conf
->mirrors
[d
].rdev
;
2167 test_bit(In_sync
, &rdev
->flags
)) {
2168 if (r1_sync_page_io(rdev
, sect
, s
,
2169 conf
->tmppage
, READ
)) {
2170 atomic_add(s
, &rdev
->corrected_errors
);
2172 "md/raid1:%s: read error corrected "
2173 "(%d sectors at %llu on %s)\n",
2175 (unsigned long long)(sect
+
2177 bdevname(rdev
->bdev
, b
));
2186 static int narrow_write_error(struct r1bio
*r1_bio
, int i
)
2188 struct mddev
*mddev
= r1_bio
->mddev
;
2189 struct r1conf
*conf
= mddev
->private;
2190 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
2192 /* bio has the data to be written to device 'i' where
2193 * we just recently had a write error.
2194 * We repeatedly clone the bio and trim down to one block,
2195 * then try the write. Where the write fails we record
2197 * It is conceivable that the bio doesn't exactly align with
2198 * blocks. We must handle this somehow.
2200 * We currently own a reference on the rdev.
2206 int sect_to_write
= r1_bio
->sectors
;
2209 if (rdev
->badblocks
.shift
< 0)
2212 block_sectors
= 1 << rdev
->badblocks
.shift
;
2213 sector
= r1_bio
->sector
;
2214 sectors
= ((sector
+ block_sectors
)
2215 & ~(sector_t
)(block_sectors
- 1))
2218 while (sect_to_write
) {
2220 if (sectors
> sect_to_write
)
2221 sectors
= sect_to_write
;
2222 /* Write at 'sector' for 'sectors'*/
2224 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
2225 unsigned vcnt
= r1_bio
->behind_page_count
;
2226 struct bio_vec
*vec
= r1_bio
->behind_bvecs
;
2228 while (!vec
->bv_page
) {
2233 wbio
= bio_alloc_mddev(GFP_NOIO
, vcnt
, mddev
);
2234 memcpy(wbio
->bi_io_vec
, vec
, vcnt
* sizeof(struct bio_vec
));
2236 wbio
->bi_vcnt
= vcnt
;
2238 wbio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2241 wbio
->bi_rw
= WRITE
;
2242 wbio
->bi_iter
.bi_sector
= r1_bio
->sector
;
2243 wbio
->bi_iter
.bi_size
= r1_bio
->sectors
<< 9;
2245 bio_trim(wbio
, sector
- r1_bio
->sector
, sectors
);
2246 wbio
->bi_iter
.bi_sector
+= rdev
->data_offset
;
2247 wbio
->bi_bdev
= rdev
->bdev
;
2248 if (submit_bio_wait(WRITE
, wbio
) == 0)
2250 ok
= rdev_set_badblocks(rdev
, sector
,
2255 sect_to_write
-= sectors
;
2257 sectors
= block_sectors
;
2262 static void handle_sync_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2265 int s
= r1_bio
->sectors
;
2266 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++) {
2267 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2268 struct bio
*bio
= r1_bio
->bios
[m
];
2269 if (bio
->bi_end_io
== NULL
)
2271 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
2272 test_bit(R1BIO_MadeGood
, &r1_bio
->state
)) {
2273 rdev_clear_badblocks(rdev
, r1_bio
->sector
, s
, 0);
2275 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
2276 test_bit(R1BIO_WriteError
, &r1_bio
->state
)) {
2277 if (!rdev_set_badblocks(rdev
, r1_bio
->sector
, s
, 0))
2278 md_error(conf
->mddev
, rdev
);
2282 md_done_sync(conf
->mddev
, s
, 1);
2285 static void handle_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2288 for (m
= 0; m
< conf
->raid_disks
* 2 ; m
++)
2289 if (r1_bio
->bios
[m
] == IO_MADE_GOOD
) {
2290 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
2291 rdev_clear_badblocks(rdev
,
2293 r1_bio
->sectors
, 0);
2294 rdev_dec_pending(rdev
, conf
->mddev
);
2295 } else if (r1_bio
->bios
[m
] != NULL
) {
2296 /* This drive got a write error. We need to
2297 * narrow down and record precise write
2300 if (!narrow_write_error(r1_bio
, m
)) {
2301 md_error(conf
->mddev
,
2302 conf
->mirrors
[m
].rdev
);
2303 /* an I/O failed, we can't clear the bitmap */
2304 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
2306 rdev_dec_pending(conf
->mirrors
[m
].rdev
,
2309 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2310 close_write(r1_bio
);
2311 raid_end_bio_io(r1_bio
);
2314 static void handle_read_error(struct r1conf
*conf
, struct r1bio
*r1_bio
)
2318 struct mddev
*mddev
= conf
->mddev
;
2320 char b
[BDEVNAME_SIZE
];
2321 struct md_rdev
*rdev
;
2323 clear_bit(R1BIO_ReadError
, &r1_bio
->state
);
2324 /* we got a read error. Maybe the drive is bad. Maybe just
2325 * the block and we can fix it.
2326 * We freeze all other IO, and try reading the block from
2327 * other devices. When we find one, we re-write
2328 * and check it that fixes the read error.
2329 * This is all done synchronously while the array is
2332 if (mddev
->ro
== 0) {
2333 freeze_array(conf
, 1);
2334 fix_read_error(conf
, r1_bio
->read_disk
,
2335 r1_bio
->sector
, r1_bio
->sectors
);
2336 unfreeze_array(conf
);
2338 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
2339 rdev_dec_pending(conf
->mirrors
[r1_bio
->read_disk
].rdev
, conf
->mddev
);
2341 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2342 bdevname(bio
->bi_bdev
, b
);
2344 disk
= read_balance(conf
, r1_bio
, &max_sectors
);
2346 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O"
2347 " read error for block %llu\n",
2348 mdname(mddev
), b
, (unsigned long long)r1_bio
->sector
);
2349 raid_end_bio_io(r1_bio
);
2351 const unsigned long do_sync
2352 = r1_bio
->master_bio
->bi_rw
& REQ_SYNC
;
2354 r1_bio
->bios
[r1_bio
->read_disk
] =
2355 mddev
->ro
? IO_BLOCKED
: NULL
;
2358 r1_bio
->read_disk
= disk
;
2359 bio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
2360 bio_trim(bio
, r1_bio
->sector
- bio
->bi_iter
.bi_sector
,
2362 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
2363 rdev
= conf
->mirrors
[disk
].rdev
;
2364 printk_ratelimited(KERN_ERR
2365 "md/raid1:%s: redirecting sector %llu"
2366 " to other mirror: %s\n",
2368 (unsigned long long)r1_bio
->sector
,
2369 bdevname(rdev
->bdev
, b
));
2370 bio
->bi_iter
.bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
2371 bio
->bi_bdev
= rdev
->bdev
;
2372 bio
->bi_end_io
= raid1_end_read_request
;
2373 bio
->bi_rw
= READ
| do_sync
;
2374 bio
->bi_private
= r1_bio
;
2375 if (max_sectors
< r1_bio
->sectors
) {
2376 /* Drat - have to split this up more */
2377 struct bio
*mbio
= r1_bio
->master_bio
;
2378 int sectors_handled
= (r1_bio
->sector
+ max_sectors
2379 - mbio
->bi_iter
.bi_sector
);
2380 r1_bio
->sectors
= max_sectors
;
2381 spin_lock_irq(&conf
->device_lock
);
2382 if (mbio
->bi_phys_segments
== 0)
2383 mbio
->bi_phys_segments
= 2;
2385 mbio
->bi_phys_segments
++;
2386 spin_unlock_irq(&conf
->device_lock
);
2387 generic_make_request(bio
);
2390 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
2392 r1_bio
->master_bio
= mbio
;
2393 r1_bio
->sectors
= bio_sectors(mbio
) - sectors_handled
;
2395 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
2396 r1_bio
->mddev
= mddev
;
2397 r1_bio
->sector
= mbio
->bi_iter
.bi_sector
+
2402 generic_make_request(bio
);
2406 static void raid1d(struct md_thread
*thread
)
2408 struct mddev
*mddev
= thread
->mddev
;
2409 struct r1bio
*r1_bio
;
2410 unsigned long flags
;
2411 struct r1conf
*conf
= mddev
->private;
2412 struct list_head
*head
= &conf
->retry_list
;
2413 struct blk_plug plug
;
2415 md_check_recovery(mddev
);
2417 blk_start_plug(&plug
);
2420 flush_pending_writes(conf
);
2422 spin_lock_irqsave(&conf
->device_lock
, flags
);
2423 if (list_empty(head
)) {
2424 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2427 r1_bio
= list_entry(head
->prev
, struct r1bio
, retry_list
);
2428 list_del(head
->prev
);
2430 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2432 mddev
= r1_bio
->mddev
;
2433 conf
= mddev
->private;
2434 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
2435 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2436 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2437 handle_sync_write_finished(conf
, r1_bio
);
2439 sync_request_write(mddev
, r1_bio
);
2440 } else if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2441 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2442 handle_write_finished(conf
, r1_bio
);
2443 else if (test_bit(R1BIO_ReadError
, &r1_bio
->state
))
2444 handle_read_error(conf
, r1_bio
);
2446 /* just a partial read to be scheduled from separate
2449 generic_make_request(r1_bio
->bios
[r1_bio
->read_disk
]);
2452 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2453 md_check_recovery(mddev
);
2455 blk_finish_plug(&plug
);
2459 static int init_resync(struct r1conf
*conf
)
2463 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2464 BUG_ON(conf
->r1buf_pool
);
2465 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
2467 if (!conf
->r1buf_pool
)
2469 conf
->next_resync
= 0;
2474 * perform a "sync" on one "block"
2476 * We need to make sure that no normal I/O request - particularly write
2477 * requests - conflict with active sync requests.
2479 * This is achieved by tracking pending requests and a 'barrier' concept
2480 * that can be installed to exclude normal IO requests.
2483 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
2485 struct r1conf
*conf
= mddev
->private;
2486 struct r1bio
*r1_bio
;
2488 sector_t max_sector
, nr_sectors
;
2492 int write_targets
= 0, read_targets
= 0;
2493 sector_t sync_blocks
;
2494 int still_degraded
= 0;
2495 int good_sectors
= RESYNC_SECTORS
;
2496 int min_bad
= 0; /* number of sectors that are bad in all devices */
2498 if (!conf
->r1buf_pool
)
2499 if (init_resync(conf
))
2502 max_sector
= mddev
->dev_sectors
;
2503 if (sector_nr
>= max_sector
) {
2504 /* If we aborted, we need to abort the
2505 * sync on the 'current' bitmap chunk (there will
2506 * only be one in raid1 resync.
2507 * We can find the current addess in mddev->curr_resync
2509 if (mddev
->curr_resync
< max_sector
) /* aborted */
2510 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2512 else /* completed sync */
2515 bitmap_close_sync(mddev
->bitmap
);
2520 if (mddev
->bitmap
== NULL
&&
2521 mddev
->recovery_cp
== MaxSector
&&
2522 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2523 conf
->fullsync
== 0) {
2525 return max_sector
- sector_nr
;
2527 /* before building a request, check if we can skip these blocks..
2528 * This call the bitmap_start_sync doesn't actually record anything
2530 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
2531 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2532 /* We can skip this block, and probably several more */
2537 * If there is non-resync activity waiting for a turn,
2538 * and resync is going fast enough,
2539 * then let it though before starting on this new sync request.
2541 if (!go_faster
&& conf
->nr_waiting
)
2542 msleep_interruptible(1000);
2544 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
2545 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
2546 raise_barrier(conf
);
2548 conf
->next_resync
= 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
)) {
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
)) {
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.
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 BUG_ON(sync_blocks
< (PAGE_SIZE
>>9));
2704 if ((len
>> 9) > sync_blocks
)
2705 len
= sync_blocks
<<9;
2708 for (i
= 0 ; i
< conf
->raid_disks
* 2; i
++) {
2709 bio
= r1_bio
->bios
[i
];
2710 if (bio
->bi_end_io
) {
2711 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2712 if (bio_add_page(bio
, page
, len
, 0) == 0) {
2714 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2717 bio
= r1_bio
->bios
[i
];
2718 if (bio
->bi_end_io
==NULL
)
2720 /* remove last page from this bio */
2722 bio
->bi_iter
.bi_size
-= len
;
2723 bio
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
2729 nr_sectors
+= len
>>9;
2730 sector_nr
+= len
>>9;
2731 sync_blocks
-= (len
>>9);
2732 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
2734 r1_bio
->sectors
= nr_sectors
;
2736 /* For a user-requested sync, we read all readable devices and do a
2739 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2740 atomic_set(&r1_bio
->remaining
, read_targets
);
2741 for (i
= 0; i
< conf
->raid_disks
* 2 && read_targets
; i
++) {
2742 bio
= r1_bio
->bios
[i
];
2743 if (bio
->bi_end_io
== end_sync_read
) {
2745 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2746 generic_make_request(bio
);
2750 atomic_set(&r1_bio
->remaining
, 1);
2751 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2752 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2753 generic_make_request(bio
);
2759 static sector_t
raid1_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2764 return mddev
->dev_sectors
;
2767 static struct r1conf
*setup_conf(struct mddev
*mddev
)
2769 struct r1conf
*conf
;
2771 struct raid1_info
*disk
;
2772 struct md_rdev
*rdev
;
2775 conf
= kzalloc(sizeof(struct r1conf
), GFP_KERNEL
);
2779 conf
->mirrors
= kzalloc(sizeof(struct raid1_info
)
2780 * mddev
->raid_disks
* 2,
2785 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2789 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
2790 if (!conf
->poolinfo
)
2792 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
* 2;
2793 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2796 if (!conf
->r1bio_pool
)
2799 conf
->poolinfo
->mddev
= mddev
;
2802 spin_lock_init(&conf
->device_lock
);
2803 rdev_for_each(rdev
, mddev
) {
2804 struct request_queue
*q
;
2805 int disk_idx
= rdev
->raid_disk
;
2806 if (disk_idx
>= mddev
->raid_disks
2809 if (test_bit(Replacement
, &rdev
->flags
))
2810 disk
= conf
->mirrors
+ mddev
->raid_disks
+ disk_idx
;
2812 disk
= conf
->mirrors
+ disk_idx
;
2817 q
= bdev_get_queue(rdev
->bdev
);
2818 if (q
->merge_bvec_fn
)
2819 mddev
->merge_check_needed
= 1;
2821 disk
->head_position
= 0;
2822 disk
->seq_start
= MaxSector
;
2824 conf
->raid_disks
= mddev
->raid_disks
;
2825 conf
->mddev
= mddev
;
2826 INIT_LIST_HEAD(&conf
->retry_list
);
2828 spin_lock_init(&conf
->resync_lock
);
2829 init_waitqueue_head(&conf
->wait_barrier
);
2831 bio_list_init(&conf
->pending_bio_list
);
2832 conf
->pending_count
= 0;
2833 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
2835 conf
->start_next_window
= MaxSector
;
2836 conf
->current_window_requests
= conf
->next_window_requests
= 0;
2839 for (i
= 0; i
< conf
->raid_disks
* 2; i
++) {
2841 disk
= conf
->mirrors
+ i
;
2843 if (i
< conf
->raid_disks
&&
2844 disk
[conf
->raid_disks
].rdev
) {
2845 /* This slot has a replacement. */
2847 /* No original, just make the replacement
2848 * a recovering spare
2851 disk
[conf
->raid_disks
].rdev
;
2852 disk
[conf
->raid_disks
].rdev
= NULL
;
2853 } else if (!test_bit(In_sync
, &disk
->rdev
->flags
))
2854 /* Original is not in_sync - bad */
2859 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2860 disk
->head_position
= 0;
2862 (disk
->rdev
->saved_raid_disk
< 0))
2868 conf
->thread
= md_register_thread(raid1d
, mddev
, "raid1");
2869 if (!conf
->thread
) {
2871 "md/raid1:%s: couldn't allocate thread\n",
2880 if (conf
->r1bio_pool
)
2881 mempool_destroy(conf
->r1bio_pool
);
2882 kfree(conf
->mirrors
);
2883 safe_put_page(conf
->tmppage
);
2884 kfree(conf
->poolinfo
);
2887 return ERR_PTR(err
);
2890 static int stop(struct mddev
*mddev
);
2891 static int run(struct mddev
*mddev
)
2893 struct r1conf
*conf
;
2895 struct md_rdev
*rdev
;
2897 bool discard_supported
= false;
2899 if (mddev
->level
!= 1) {
2900 printk(KERN_ERR
"md/raid1:%s: raid level not set to mirroring (%d)\n",
2901 mdname(mddev
), mddev
->level
);
2904 if (mddev
->reshape_position
!= MaxSector
) {
2905 printk(KERN_ERR
"md/raid1:%s: reshape_position set but not supported\n",
2910 * copy the already verified devices into our private RAID1
2911 * bookkeeping area. [whatever we allocate in run(),
2912 * should be freed in stop()]
2914 if (mddev
->private == NULL
)
2915 conf
= setup_conf(mddev
);
2917 conf
= mddev
->private;
2920 return PTR_ERR(conf
);
2923 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
2925 rdev_for_each(rdev
, mddev
) {
2926 if (!mddev
->gendisk
)
2928 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2929 rdev
->data_offset
<< 9);
2930 if (blk_queue_discard(bdev_get_queue(rdev
->bdev
)))
2931 discard_supported
= true;
2934 mddev
->degraded
= 0;
2935 for (i
=0; i
< conf
->raid_disks
; i
++)
2936 if (conf
->mirrors
[i
].rdev
== NULL
||
2937 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
2938 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2941 if (conf
->raid_disks
- mddev
->degraded
== 1)
2942 mddev
->recovery_cp
= MaxSector
;
2944 if (mddev
->recovery_cp
!= MaxSector
)
2945 printk(KERN_NOTICE
"md/raid1:%s: not clean"
2946 " -- starting background reconstruction\n",
2949 "md/raid1:%s: active with %d out of %d mirrors\n",
2950 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2954 * Ok, everything is just fine now
2956 mddev
->thread
= conf
->thread
;
2957 conf
->thread
= NULL
;
2958 mddev
->private = conf
;
2960 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
2963 mddev
->queue
->backing_dev_info
.congested_fn
= raid1_congested
;
2964 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2965 blk_queue_merge_bvec(mddev
->queue
, raid1_mergeable_bvec
);
2967 if (discard_supported
)
2968 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
2971 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
2975 ret
= md_integrity_register(mddev
);
2981 static int stop(struct mddev
*mddev
)
2983 struct r1conf
*conf
= mddev
->private;
2984 struct bitmap
*bitmap
= mddev
->bitmap
;
2986 /* wait for behind writes to complete */
2987 if (bitmap
&& atomic_read(&bitmap
->behind_writes
) > 0) {
2988 printk(KERN_INFO
"md/raid1:%s: behind writes in progress - waiting to stop.\n",
2990 /* need to kick something here to make sure I/O goes? */
2991 wait_event(bitmap
->behind_wait
,
2992 atomic_read(&bitmap
->behind_writes
) == 0);
2995 freeze_array(conf
, 0);
2996 unfreeze_array(conf
);
2998 md_unregister_thread(&mddev
->thread
);
2999 if (conf
->r1bio_pool
)
3000 mempool_destroy(conf
->r1bio_pool
);
3001 kfree(conf
->mirrors
);
3002 safe_put_page(conf
->tmppage
);
3003 kfree(conf
->poolinfo
);
3005 mddev
->private = NULL
;
3009 static int raid1_resize(struct mddev
*mddev
, sector_t sectors
)
3011 /* no resync is happening, and there is enough space
3012 * on all devices, so we can resize.
3013 * We need to make sure resync covers any new space.
3014 * If the array is shrinking we should possibly wait until
3015 * any io in the removed space completes, but it hardly seems
3018 sector_t newsize
= raid1_size(mddev
, sectors
, 0);
3019 if (mddev
->external_size
&&
3020 mddev
->array_sectors
> newsize
)
3022 if (mddev
->bitmap
) {
3023 int ret
= bitmap_resize(mddev
->bitmap
, newsize
, 0, 0);
3027 md_set_array_sectors(mddev
, newsize
);
3028 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
3029 revalidate_disk(mddev
->gendisk
);
3030 if (sectors
> mddev
->dev_sectors
&&
3031 mddev
->recovery_cp
> mddev
->dev_sectors
) {
3032 mddev
->recovery_cp
= mddev
->dev_sectors
;
3033 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3035 mddev
->dev_sectors
= sectors
;
3036 mddev
->resync_max_sectors
= sectors
;
3040 static int raid1_reshape(struct mddev
*mddev
)
3043 * 1/ resize the r1bio_pool
3044 * 2/ resize conf->mirrors
3046 * We allocate a new r1bio_pool if we can.
3047 * Then raise a device barrier and wait until all IO stops.
3048 * Then resize conf->mirrors and swap in the new r1bio pool.
3050 * At the same time, we "pack" the devices so that all the missing
3051 * devices have the higher raid_disk numbers.
3053 mempool_t
*newpool
, *oldpool
;
3054 struct pool_info
*newpoolinfo
;
3055 struct raid1_info
*newmirrors
;
3056 struct r1conf
*conf
= mddev
->private;
3057 int cnt
, raid_disks
;
3058 unsigned long flags
;
3061 /* Cannot change chunk_size, layout, or level */
3062 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
3063 mddev
->layout
!= mddev
->new_layout
||
3064 mddev
->level
!= mddev
->new_level
) {
3065 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
3066 mddev
->new_layout
= mddev
->layout
;
3067 mddev
->new_level
= mddev
->level
;
3071 err
= md_allow_write(mddev
);
3075 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
3077 if (raid_disks
< conf
->raid_disks
) {
3079 for (d
= 0; d
< conf
->raid_disks
; d
++)
3080 if (conf
->mirrors
[d
].rdev
)
3082 if (cnt
> raid_disks
)
3086 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
3089 newpoolinfo
->mddev
= mddev
;
3090 newpoolinfo
->raid_disks
= raid_disks
* 2;
3092 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
3093 r1bio_pool_free
, newpoolinfo
);
3098 newmirrors
= kzalloc(sizeof(struct raid1_info
) * raid_disks
* 2,
3102 mempool_destroy(newpool
);
3106 freeze_array(conf
, 0);
3108 /* ok, everything is stopped */
3109 oldpool
= conf
->r1bio_pool
;
3110 conf
->r1bio_pool
= newpool
;
3112 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
3113 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
3114 if (rdev
&& rdev
->raid_disk
!= d2
) {
3115 sysfs_unlink_rdev(mddev
, rdev
);
3116 rdev
->raid_disk
= d2
;
3117 sysfs_unlink_rdev(mddev
, rdev
);
3118 if (sysfs_link_rdev(mddev
, rdev
))
3120 "md/raid1:%s: cannot register rd%d\n",
3121 mdname(mddev
), rdev
->raid_disk
);
3124 newmirrors
[d2
++].rdev
= rdev
;
3126 kfree(conf
->mirrors
);
3127 conf
->mirrors
= newmirrors
;
3128 kfree(conf
->poolinfo
);
3129 conf
->poolinfo
= newpoolinfo
;
3131 spin_lock_irqsave(&conf
->device_lock
, flags
);
3132 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
3133 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3134 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
3135 mddev
->delta_disks
= 0;
3137 unfreeze_array(conf
);
3139 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
3140 md_wakeup_thread(mddev
->thread
);
3142 mempool_destroy(oldpool
);
3146 static void raid1_quiesce(struct mddev
*mddev
, int state
)
3148 struct r1conf
*conf
= mddev
->private;
3151 case 2: /* wake for suspend */
3152 wake_up(&conf
->wait_barrier
);
3155 freeze_array(conf
, 0);
3158 unfreeze_array(conf
);
3163 static void *raid1_takeover(struct mddev
*mddev
)
3165 /* raid1 can take over:
3166 * raid5 with 2 devices, any layout or chunk size
3168 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
3169 struct r1conf
*conf
;
3170 mddev
->new_level
= 1;
3171 mddev
->new_layout
= 0;
3172 mddev
->new_chunk_sectors
= 0;
3173 conf
= setup_conf(mddev
);
3175 /* Array must appear to be quiesced */
3176 conf
->array_frozen
= 1;
3179 return ERR_PTR(-EINVAL
);
3182 static struct md_personality raid1_personality
=
3186 .owner
= THIS_MODULE
,
3187 .make_request
= make_request
,
3191 .error_handler
= error
,
3192 .hot_add_disk
= raid1_add_disk
,
3193 .hot_remove_disk
= raid1_remove_disk
,
3194 .spare_active
= raid1_spare_active
,
3195 .sync_request
= sync_request
,
3196 .resize
= raid1_resize
,
3198 .check_reshape
= raid1_reshape
,
3199 .quiesce
= raid1_quiesce
,
3200 .takeover
= raid1_takeover
,
3203 static int __init
raid_init(void)
3205 return register_md_personality(&raid1_personality
);
3208 static void raid_exit(void)
3210 unregister_md_personality(&raid1_personality
);
3213 module_init(raid_init
);
3214 module_exit(raid_exit
);
3215 MODULE_LICENSE("GPL");
3216 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3217 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3218 MODULE_ALIAS("md-raid1");
3219 MODULE_ALIAS("md-level-1");
3221 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);