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/seq_file.h>
38 #include <linux/ratelimit.h>
44 * Number of guaranteed r1bios in case of extreme VM load:
46 #define NR_RAID1_BIOS 256
48 /* When there are this many requests queue to be written by
49 * the raid1 thread, we become 'congested' to provide back-pressure
52 static int max_queued_requests
= 1024;
54 static void allow_barrier(struct r1conf
*conf
);
55 static void lower_barrier(struct r1conf
*conf
);
57 static void * r1bio_pool_alloc(gfp_t gfp_flags
, void *data
)
59 struct pool_info
*pi
= data
;
60 int size
= offsetof(struct r1bio
, bios
[pi
->raid_disks
]);
62 /* allocate a r1bio with room for raid_disks entries in the bios array */
63 return kzalloc(size
, gfp_flags
);
66 static void r1bio_pool_free(void *r1_bio
, void *data
)
71 #define RESYNC_BLOCK_SIZE (64*1024)
72 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
73 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
74 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
75 #define RESYNC_WINDOW (2048*1024)
77 static void * r1buf_pool_alloc(gfp_t gfp_flags
, void *data
)
79 struct pool_info
*pi
= data
;
85 r1_bio
= r1bio_pool_alloc(gfp_flags
, pi
);
90 * Allocate bios : 1 for reading, n-1 for writing
92 for (j
= pi
->raid_disks
; j
-- ; ) {
93 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
96 r1_bio
->bios
[j
] = bio
;
99 * Allocate RESYNC_PAGES data pages and attach them to
101 * If this is a user-requested check/repair, allocate
102 * RESYNC_PAGES for each bio.
104 if (test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
))
109 bio
= r1_bio
->bios
[j
];
110 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
111 page
= alloc_page(gfp_flags
);
115 bio
->bi_io_vec
[i
].bv_page
= page
;
119 /* If not user-requests, copy the page pointers to all bios */
120 if (!test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
)) {
121 for (i
=0; i
<RESYNC_PAGES
; i
++)
122 for (j
=1; j
<pi
->raid_disks
; j
++)
123 r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
=
124 r1_bio
->bios
[0]->bi_io_vec
[i
].bv_page
;
127 r1_bio
->master_bio
= NULL
;
132 for (j
=0 ; j
< pi
->raid_disks
; j
++)
133 for (i
=0; i
< r1_bio
->bios
[j
]->bi_vcnt
; i
++)
134 put_page(r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
137 while ( ++j
< pi
->raid_disks
)
138 bio_put(r1_bio
->bios
[j
]);
139 r1bio_pool_free(r1_bio
, data
);
143 static void r1buf_pool_free(void *__r1_bio
, void *data
)
145 struct pool_info
*pi
= data
;
147 struct r1bio
*r1bio
= __r1_bio
;
149 for (i
= 0; i
< RESYNC_PAGES
; i
++)
150 for (j
= pi
->raid_disks
; j
-- ;) {
152 r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
!=
153 r1bio
->bios
[0]->bi_io_vec
[i
].bv_page
)
154 safe_put_page(r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
156 for (i
=0 ; i
< pi
->raid_disks
; i
++)
157 bio_put(r1bio
->bios
[i
]);
159 r1bio_pool_free(r1bio
, data
);
162 static void put_all_bios(struct r1conf
*conf
, struct r1bio
*r1_bio
)
166 for (i
= 0; i
< conf
->raid_disks
; i
++) {
167 struct bio
**bio
= r1_bio
->bios
+ i
;
168 if (!BIO_SPECIAL(*bio
))
174 static void free_r1bio(struct r1bio
*r1_bio
)
176 struct r1conf
*conf
= r1_bio
->mddev
->private;
178 put_all_bios(conf
, r1_bio
);
179 mempool_free(r1_bio
, conf
->r1bio_pool
);
182 static void put_buf(struct r1bio
*r1_bio
)
184 struct r1conf
*conf
= r1_bio
->mddev
->private;
187 for (i
=0; i
<conf
->raid_disks
; i
++) {
188 struct bio
*bio
= r1_bio
->bios
[i
];
190 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
193 mempool_free(r1_bio
, conf
->r1buf_pool
);
198 static void reschedule_retry(struct r1bio
*r1_bio
)
201 struct mddev
*mddev
= r1_bio
->mddev
;
202 struct r1conf
*conf
= mddev
->private;
204 spin_lock_irqsave(&conf
->device_lock
, flags
);
205 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
207 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
209 wake_up(&conf
->wait_barrier
);
210 md_wakeup_thread(mddev
->thread
);
214 * raid_end_bio_io() is called when we have finished servicing a mirrored
215 * operation and are ready to return a success/failure code to the buffer
218 static void call_bio_endio(struct r1bio
*r1_bio
)
220 struct bio
*bio
= r1_bio
->master_bio
;
222 struct r1conf
*conf
= r1_bio
->mddev
->private;
224 if (bio
->bi_phys_segments
) {
226 spin_lock_irqsave(&conf
->device_lock
, flags
);
227 bio
->bi_phys_segments
--;
228 done
= (bio
->bi_phys_segments
== 0);
229 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
233 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
234 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
238 * Wake up any possible resync thread that waits for the device
245 static void raid_end_bio_io(struct r1bio
*r1_bio
)
247 struct bio
*bio
= r1_bio
->master_bio
;
249 /* if nobody has done the final endio yet, do it now */
250 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
251 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
252 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
253 (unsigned long long) bio
->bi_sector
,
254 (unsigned long long) bio
->bi_sector
+
255 (bio
->bi_size
>> 9) - 1);
257 call_bio_endio(r1_bio
);
263 * Update disk head position estimator based on IRQ completion info.
265 static inline void update_head_pos(int disk
, struct r1bio
*r1_bio
)
267 struct r1conf
*conf
= r1_bio
->mddev
->private;
269 conf
->mirrors
[disk
].head_position
=
270 r1_bio
->sector
+ (r1_bio
->sectors
);
274 * Find the disk number which triggered given bio
276 static int find_bio_disk(struct r1bio
*r1_bio
, struct bio
*bio
)
279 int raid_disks
= r1_bio
->mddev
->raid_disks
;
281 for (mirror
= 0; mirror
< raid_disks
; mirror
++)
282 if (r1_bio
->bios
[mirror
] == bio
)
285 BUG_ON(mirror
== raid_disks
);
286 update_head_pos(mirror
, r1_bio
);
291 static void raid1_end_read_request(struct bio
*bio
, int error
)
293 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
294 struct r1bio
*r1_bio
= bio
->bi_private
;
296 struct r1conf
*conf
= r1_bio
->mddev
->private;
298 mirror
= r1_bio
->read_disk
;
300 * this branch is our 'one mirror IO has finished' event handler:
302 update_head_pos(mirror
, r1_bio
);
305 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
307 /* If all other devices have failed, we want to return
308 * the error upwards rather than fail the last device.
309 * Here we redefine "uptodate" to mean "Don't want to retry"
312 spin_lock_irqsave(&conf
->device_lock
, flags
);
313 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
314 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
315 !test_bit(Faulty
, &conf
->mirrors
[mirror
].rdev
->flags
)))
317 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
321 raid_end_bio_io(r1_bio
);
326 char b
[BDEVNAME_SIZE
];
328 KERN_ERR
"md/raid1:%s: %s: "
329 "rescheduling sector %llu\n",
331 bdevname(conf
->mirrors
[mirror
].rdev
->bdev
,
333 (unsigned long long)r1_bio
->sector
);
334 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
335 reschedule_retry(r1_bio
);
338 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
341 static void close_write(struct r1bio
*r1_bio
)
343 /* it really is the end of this request */
344 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
345 /* free extra copy of the data pages */
346 int i
= r1_bio
->behind_page_count
;
348 safe_put_page(r1_bio
->behind_bvecs
[i
].bv_page
);
349 kfree(r1_bio
->behind_bvecs
);
350 r1_bio
->behind_bvecs
= NULL
;
352 /* clear the bitmap if all writes complete successfully */
353 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
355 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
356 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
357 md_write_end(r1_bio
->mddev
);
360 static void r1_bio_write_done(struct r1bio
*r1_bio
)
362 if (!atomic_dec_and_test(&r1_bio
->remaining
))
365 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
366 reschedule_retry(r1_bio
);
369 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
))
370 reschedule_retry(r1_bio
);
372 raid_end_bio_io(r1_bio
);
376 static void raid1_end_write_request(struct bio
*bio
, int error
)
378 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
379 struct r1bio
*r1_bio
= bio
->bi_private
;
380 int mirror
, behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
381 struct r1conf
*conf
= r1_bio
->mddev
->private;
382 struct bio
*to_put
= NULL
;
384 mirror
= find_bio_disk(r1_bio
, bio
);
387 * 'one mirror IO has finished' event handler:
390 set_bit(WriteErrorSeen
,
391 &conf
->mirrors
[mirror
].rdev
->flags
);
392 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
395 * Set R1BIO_Uptodate in our master bio, so that we
396 * will return a good error code for to the higher
397 * levels even if IO on some other mirrored buffer
400 * The 'master' represents the composite IO operation
401 * to user-side. So if something waits for IO, then it
402 * will wait for the 'master' bio.
407 r1_bio
->bios
[mirror
] = NULL
;
409 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
411 /* Maybe we can clear some bad blocks. */
412 if (is_badblock(conf
->mirrors
[mirror
].rdev
,
413 r1_bio
->sector
, r1_bio
->sectors
,
414 &first_bad
, &bad_sectors
)) {
415 r1_bio
->bios
[mirror
] = IO_MADE_GOOD
;
416 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
421 if (test_bit(WriteMostly
, &conf
->mirrors
[mirror
].rdev
->flags
))
422 atomic_dec(&r1_bio
->behind_remaining
);
425 * In behind mode, we ACK the master bio once the I/O
426 * has safely reached all non-writemostly
427 * disks. Setting the Returned bit ensures that this
428 * gets done only once -- we don't ever want to return
429 * -EIO here, instead we'll wait
431 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
432 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
433 /* Maybe we can return now */
434 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
435 struct bio
*mbio
= r1_bio
->master_bio
;
436 pr_debug("raid1: behind end write sectors"
438 (unsigned long long) mbio
->bi_sector
,
439 (unsigned long long) mbio
->bi_sector
+
440 (mbio
->bi_size
>> 9) - 1);
441 call_bio_endio(r1_bio
);
445 if (r1_bio
->bios
[mirror
] == NULL
)
446 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
,
450 * Let's see if all mirrored write operations have finished
453 r1_bio_write_done(r1_bio
);
461 * This routine returns the disk from which the requested read should
462 * be done. There is a per-array 'next expected sequential IO' sector
463 * number - if this matches on the next IO then we use the last disk.
464 * There is also a per-disk 'last know head position' sector that is
465 * maintained from IRQ contexts, both the normal and the resync IO
466 * completion handlers update this position correctly. If there is no
467 * perfect sequential match then we pick the disk whose head is closest.
469 * If there are 2 mirrors in the same 2 devices, performance degrades
470 * because position is mirror, not device based.
472 * The rdev for the device selected will have nr_pending incremented.
474 static int read_balance(struct r1conf
*conf
, struct r1bio
*r1_bio
, int *max_sectors
)
476 const sector_t this_sector
= r1_bio
->sector
;
478 int best_good_sectors
;
483 struct md_rdev
*rdev
;
488 * Check if we can balance. We can balance on the whole
489 * device if no resync is going on, or below the resync window.
490 * We take the first readable disk when above the resync window.
493 sectors
= r1_bio
->sectors
;
495 best_dist
= MaxSector
;
496 best_good_sectors
= 0;
498 if (conf
->mddev
->recovery_cp
< MaxSector
&&
499 (this_sector
+ sectors
>= conf
->next_resync
)) {
504 start_disk
= conf
->last_used
;
507 for (i
= 0 ; i
< conf
->raid_disks
; i
++) {
512 int disk
= start_disk
+ i
;
513 if (disk
>= conf
->raid_disks
)
514 disk
-= conf
->raid_disks
;
516 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
517 if (r1_bio
->bios
[disk
] == IO_BLOCKED
519 || test_bit(Faulty
, &rdev
->flags
))
521 if (!test_bit(In_sync
, &rdev
->flags
) &&
522 rdev
->recovery_offset
< this_sector
+ sectors
)
524 if (test_bit(WriteMostly
, &rdev
->flags
)) {
525 /* Don't balance among write-mostly, just
526 * use the first as a last resort */
531 /* This is a reasonable device to use. It might
534 if (is_badblock(rdev
, this_sector
, sectors
,
535 &first_bad
, &bad_sectors
)) {
536 if (best_dist
< MaxSector
)
537 /* already have a better device */
539 if (first_bad
<= this_sector
) {
540 /* cannot read here. If this is the 'primary'
541 * device, then we must not read beyond
542 * bad_sectors from another device..
544 bad_sectors
-= (this_sector
- first_bad
);
545 if (choose_first
&& sectors
> bad_sectors
)
546 sectors
= bad_sectors
;
547 if (best_good_sectors
> sectors
)
548 best_good_sectors
= sectors
;
551 sector_t good_sectors
= first_bad
- this_sector
;
552 if (good_sectors
> best_good_sectors
) {
553 best_good_sectors
= good_sectors
;
561 best_good_sectors
= sectors
;
563 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
565 /* Don't change to another disk for sequential reads */
566 || conf
->next_seq_sect
== this_sector
568 /* If device is idle, use it */
569 || atomic_read(&rdev
->nr_pending
) == 0) {
573 if (dist
< best_dist
) {
579 if (best_disk
>= 0) {
580 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
583 atomic_inc(&rdev
->nr_pending
);
584 if (test_bit(Faulty
, &rdev
->flags
)) {
585 /* cannot risk returning a device that failed
586 * before we inc'ed nr_pending
588 rdev_dec_pending(rdev
, conf
->mddev
);
591 sectors
= best_good_sectors
;
592 conf
->next_seq_sect
= this_sector
+ sectors
;
593 conf
->last_used
= best_disk
;
596 *max_sectors
= sectors
;
601 int md_raid1_congested(struct mddev
*mddev
, int bits
)
603 struct r1conf
*conf
= mddev
->private;
606 if ((bits
& (1 << BDI_async_congested
)) &&
607 conf
->pending_count
>= max_queued_requests
)
611 for (i
= 0; i
< mddev
->raid_disks
; i
++) {
612 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
613 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
614 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
618 /* Note the '|| 1' - when read_balance prefers
619 * non-congested targets, it can be removed
621 if ((bits
& (1<<BDI_async_congested
)) || 1)
622 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
624 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
630 EXPORT_SYMBOL_GPL(md_raid1_congested
);
632 static int raid1_congested(void *data
, int bits
)
634 struct mddev
*mddev
= data
;
636 return mddev_congested(mddev
, bits
) ||
637 md_raid1_congested(mddev
, bits
);
640 static void flush_pending_writes(struct r1conf
*conf
)
642 /* Any writes that have been queued but are awaiting
643 * bitmap updates get flushed here.
645 spin_lock_irq(&conf
->device_lock
);
647 if (conf
->pending_bio_list
.head
) {
649 bio
= bio_list_get(&conf
->pending_bio_list
);
650 conf
->pending_count
= 0;
651 spin_unlock_irq(&conf
->device_lock
);
652 /* flush any pending bitmap writes to
653 * disk before proceeding w/ I/O */
654 bitmap_unplug(conf
->mddev
->bitmap
);
655 wake_up(&conf
->wait_barrier
);
657 while (bio
) { /* submit pending writes */
658 struct bio
*next
= bio
->bi_next
;
660 generic_make_request(bio
);
664 spin_unlock_irq(&conf
->device_lock
);
668 * Sometimes we need to suspend IO while we do something else,
669 * either some resync/recovery, or reconfigure the array.
670 * To do this we raise a 'barrier'.
671 * The 'barrier' is a counter that can be raised multiple times
672 * to count how many activities are happening which preclude
674 * We can only raise the barrier if there is no pending IO.
675 * i.e. if nr_pending == 0.
676 * We choose only to raise the barrier if no-one is waiting for the
677 * barrier to go down. This means that as soon as an IO request
678 * is ready, no other operations which require a barrier will start
679 * until the IO request has had a chance.
681 * So: regular IO calls 'wait_barrier'. When that returns there
682 * is no backgroup IO happening, It must arrange to call
683 * allow_barrier when it has finished its IO.
684 * backgroup IO calls must call raise_barrier. Once that returns
685 * there is no normal IO happeing. It must arrange to call
686 * lower_barrier when the particular background IO completes.
688 #define RESYNC_DEPTH 32
690 static void raise_barrier(struct r1conf
*conf
)
692 spin_lock_irq(&conf
->resync_lock
);
694 /* Wait until no block IO is waiting */
695 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
696 conf
->resync_lock
, );
698 /* block any new IO from starting */
701 /* Now wait for all pending IO to complete */
702 wait_event_lock_irq(conf
->wait_barrier
,
703 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
704 conf
->resync_lock
, );
706 spin_unlock_irq(&conf
->resync_lock
);
709 static void lower_barrier(struct r1conf
*conf
)
712 BUG_ON(conf
->barrier
<= 0);
713 spin_lock_irqsave(&conf
->resync_lock
, flags
);
715 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
716 wake_up(&conf
->wait_barrier
);
719 static void wait_barrier(struct r1conf
*conf
)
721 spin_lock_irq(&conf
->resync_lock
);
724 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
730 spin_unlock_irq(&conf
->resync_lock
);
733 static void allow_barrier(struct r1conf
*conf
)
736 spin_lock_irqsave(&conf
->resync_lock
, flags
);
738 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
739 wake_up(&conf
->wait_barrier
);
742 static void freeze_array(struct r1conf
*conf
)
744 /* stop syncio and normal IO and wait for everything to
746 * We increment barrier and nr_waiting, and then
747 * wait until nr_pending match nr_queued+1
748 * This is called in the context of one normal IO request
749 * that has failed. Thus any sync request that might be pending
750 * will be blocked by nr_pending, and we need to wait for
751 * pending IO requests to complete or be queued for re-try.
752 * Thus the number queued (nr_queued) plus this request (1)
753 * must match the number of pending IOs (nr_pending) before
756 spin_lock_irq(&conf
->resync_lock
);
759 wait_event_lock_irq(conf
->wait_barrier
,
760 conf
->nr_pending
== conf
->nr_queued
+1,
762 flush_pending_writes(conf
));
763 spin_unlock_irq(&conf
->resync_lock
);
765 static void unfreeze_array(struct r1conf
*conf
)
767 /* reverse the effect of the freeze */
768 spin_lock_irq(&conf
->resync_lock
);
771 wake_up(&conf
->wait_barrier
);
772 spin_unlock_irq(&conf
->resync_lock
);
776 /* duplicate the data pages for behind I/O
778 static void alloc_behind_pages(struct bio
*bio
, struct r1bio
*r1_bio
)
781 struct bio_vec
*bvec
;
782 struct bio_vec
*bvecs
= kzalloc(bio
->bi_vcnt
* sizeof(struct bio_vec
),
784 if (unlikely(!bvecs
))
787 bio_for_each_segment(bvec
, bio
, i
) {
789 bvecs
[i
].bv_page
= alloc_page(GFP_NOIO
);
790 if (unlikely(!bvecs
[i
].bv_page
))
792 memcpy(kmap(bvecs
[i
].bv_page
) + bvec
->bv_offset
,
793 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
794 kunmap(bvecs
[i
].bv_page
);
795 kunmap(bvec
->bv_page
);
797 r1_bio
->behind_bvecs
= bvecs
;
798 r1_bio
->behind_page_count
= bio
->bi_vcnt
;
799 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
803 for (i
= 0; i
< bio
->bi_vcnt
; i
++)
804 if (bvecs
[i
].bv_page
)
805 put_page(bvecs
[i
].bv_page
);
807 pr_debug("%dB behind alloc failed, doing sync I/O\n", bio
->bi_size
);
810 static int make_request(struct mddev
*mddev
, struct bio
* bio
)
812 struct r1conf
*conf
= mddev
->private;
813 struct mirror_info
*mirror
;
814 struct r1bio
*r1_bio
;
815 struct bio
*read_bio
;
817 struct bitmap
*bitmap
;
819 const int rw
= bio_data_dir(bio
);
820 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
821 const unsigned long do_flush_fua
= (bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
));
822 struct md_rdev
*blocked_rdev
;
829 * Register the new request and wait if the reconstruction
830 * thread has put up a bar for new requests.
831 * Continue immediately if no resync is active currently.
834 md_write_start(mddev
, bio
); /* wait on superblock update early */
836 if (bio_data_dir(bio
) == WRITE
&&
837 bio
->bi_sector
+ bio
->bi_size
/512 > mddev
->suspend_lo
&&
838 bio
->bi_sector
< mddev
->suspend_hi
) {
839 /* As the suspend_* range is controlled by
840 * userspace, we want an interruptible
845 flush_signals(current
);
846 prepare_to_wait(&conf
->wait_barrier
,
847 &w
, TASK_INTERRUPTIBLE
);
848 if (bio
->bi_sector
+ bio
->bi_size
/512 <= mddev
->suspend_lo
||
849 bio
->bi_sector
>= mddev
->suspend_hi
)
853 finish_wait(&conf
->wait_barrier
, &w
);
858 bitmap
= mddev
->bitmap
;
861 * make_request() can abort the operation when READA is being
862 * used and no empty request is available.
865 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
867 r1_bio
->master_bio
= bio
;
868 r1_bio
->sectors
= bio
->bi_size
>> 9;
870 r1_bio
->mddev
= mddev
;
871 r1_bio
->sector
= bio
->bi_sector
;
873 /* We might need to issue multiple reads to different
874 * devices if there are bad blocks around, so we keep
875 * track of the number of reads in bio->bi_phys_segments.
876 * If this is 0, there is only one r1_bio and no locking
877 * will be needed when requests complete. If it is
878 * non-zero, then it is the number of not-completed requests.
880 bio
->bi_phys_segments
= 0;
881 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
885 * read balancing logic:
890 rdisk
= read_balance(conf
, r1_bio
, &max_sectors
);
893 /* couldn't find anywhere to read from */
894 raid_end_bio_io(r1_bio
);
897 mirror
= conf
->mirrors
+ rdisk
;
899 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
901 /* Reading from a write-mostly device must
902 * take care not to over-take any writes
905 wait_event(bitmap
->behind_wait
,
906 atomic_read(&bitmap
->behind_writes
) == 0);
908 r1_bio
->read_disk
= rdisk
;
910 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
911 md_trim_bio(read_bio
, r1_bio
->sector
- bio
->bi_sector
,
914 r1_bio
->bios
[rdisk
] = read_bio
;
916 read_bio
->bi_sector
= r1_bio
->sector
+ mirror
->rdev
->data_offset
;
917 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
918 read_bio
->bi_end_io
= raid1_end_read_request
;
919 read_bio
->bi_rw
= READ
| do_sync
;
920 read_bio
->bi_private
= r1_bio
;
922 if (max_sectors
< r1_bio
->sectors
) {
923 /* could not read all from this device, so we will
924 * need another r1_bio.
927 sectors_handled
= (r1_bio
->sector
+ max_sectors
929 r1_bio
->sectors
= max_sectors
;
930 spin_lock_irq(&conf
->device_lock
);
931 if (bio
->bi_phys_segments
== 0)
932 bio
->bi_phys_segments
= 2;
934 bio
->bi_phys_segments
++;
935 spin_unlock_irq(&conf
->device_lock
);
936 /* Cannot call generic_make_request directly
937 * as that will be queued in __make_request
938 * and subsequent mempool_alloc might block waiting
939 * for it. So hand bio over to raid1d.
941 reschedule_retry(r1_bio
);
943 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
945 r1_bio
->master_bio
= bio
;
946 r1_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
948 r1_bio
->mddev
= mddev
;
949 r1_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
952 generic_make_request(read_bio
);
959 if (conf
->pending_count
>= max_queued_requests
) {
960 md_wakeup_thread(mddev
->thread
);
961 wait_event(conf
->wait_barrier
,
962 conf
->pending_count
< max_queued_requests
);
964 /* first select target devices under rcu_lock and
965 * inc refcount on their rdev. Record them by setting
967 * If there are known/acknowledged bad blocks on any device on
968 * which we have seen a write error, we want to avoid writing those
970 * This potentially requires several writes to write around
971 * the bad blocks. Each set of writes gets it's own r1bio
972 * with a set of bios attached.
974 plugged
= mddev_check_plugged(mddev
);
976 disks
= conf
->raid_disks
;
980 max_sectors
= r1_bio
->sectors
;
981 for (i
= 0; i
< disks
; i
++) {
982 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
983 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
984 atomic_inc(&rdev
->nr_pending
);
988 r1_bio
->bios
[i
] = NULL
;
989 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
990 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
994 atomic_inc(&rdev
->nr_pending
);
995 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1000 is_bad
= is_badblock(rdev
, r1_bio
->sector
,
1002 &first_bad
, &bad_sectors
);
1004 /* mustn't write here until the bad block is
1006 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1007 blocked_rdev
= rdev
;
1010 if (is_bad
&& first_bad
<= r1_bio
->sector
) {
1011 /* Cannot write here at all */
1012 bad_sectors
-= (r1_bio
->sector
- first_bad
);
1013 if (bad_sectors
< max_sectors
)
1014 /* mustn't write more than bad_sectors
1015 * to other devices yet
1017 max_sectors
= bad_sectors
;
1018 rdev_dec_pending(rdev
, mddev
);
1019 /* We don't set R1BIO_Degraded as that
1020 * only applies if the disk is
1021 * missing, so it might be re-added,
1022 * and we want to know to recover this
1024 * In this case the device is here,
1025 * and the fact that this chunk is not
1026 * in-sync is recorded in the bad
1032 int good_sectors
= first_bad
- r1_bio
->sector
;
1033 if (good_sectors
< max_sectors
)
1034 max_sectors
= good_sectors
;
1037 r1_bio
->bios
[i
] = bio
;
1041 if (unlikely(blocked_rdev
)) {
1042 /* Wait for this device to become unblocked */
1045 for (j
= 0; j
< i
; j
++)
1046 if (r1_bio
->bios
[j
])
1047 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
1049 allow_barrier(conf
);
1050 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1055 if (max_sectors
< r1_bio
->sectors
) {
1056 /* We are splitting this write into multiple parts, so
1057 * we need to prepare for allocating another r1_bio.
1059 r1_bio
->sectors
= max_sectors
;
1060 spin_lock_irq(&conf
->device_lock
);
1061 if (bio
->bi_phys_segments
== 0)
1062 bio
->bi_phys_segments
= 2;
1064 bio
->bi_phys_segments
++;
1065 spin_unlock_irq(&conf
->device_lock
);
1067 sectors_handled
= r1_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1069 atomic_set(&r1_bio
->remaining
, 1);
1070 atomic_set(&r1_bio
->behind_remaining
, 0);
1073 for (i
= 0; i
< disks
; i
++) {
1075 if (!r1_bio
->bios
[i
])
1078 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1079 md_trim_bio(mbio
, r1_bio
->sector
- bio
->bi_sector
, max_sectors
);
1083 * Not if there are too many, or cannot
1084 * allocate memory, or a reader on WriteMostly
1085 * is waiting for behind writes to flush */
1087 (atomic_read(&bitmap
->behind_writes
)
1088 < mddev
->bitmap_info
.max_write_behind
) &&
1089 !waitqueue_active(&bitmap
->behind_wait
))
1090 alloc_behind_pages(mbio
, r1_bio
);
1092 bitmap_startwrite(bitmap
, r1_bio
->sector
,
1094 test_bit(R1BIO_BehindIO
,
1098 if (r1_bio
->behind_bvecs
) {
1099 struct bio_vec
*bvec
;
1102 /* Yes, I really want the '__' version so that
1103 * we clear any unused pointer in the io_vec, rather
1104 * than leave them unchanged. This is important
1105 * because when we come to free the pages, we won't
1106 * know the original bi_idx, so we just free
1109 __bio_for_each_segment(bvec
, mbio
, j
, 0)
1110 bvec
->bv_page
= r1_bio
->behind_bvecs
[j
].bv_page
;
1111 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
1112 atomic_inc(&r1_bio
->behind_remaining
);
1115 r1_bio
->bios
[i
] = mbio
;
1117 mbio
->bi_sector
= (r1_bio
->sector
+
1118 conf
->mirrors
[i
].rdev
->data_offset
);
1119 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1120 mbio
->bi_end_io
= raid1_end_write_request
;
1121 mbio
->bi_rw
= WRITE
| do_flush_fua
| do_sync
;
1122 mbio
->bi_private
= r1_bio
;
1124 atomic_inc(&r1_bio
->remaining
);
1125 spin_lock_irqsave(&conf
->device_lock
, flags
);
1126 bio_list_add(&conf
->pending_bio_list
, mbio
);
1127 conf
->pending_count
++;
1128 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1130 /* Mustn't call r1_bio_write_done before this next test,
1131 * as it could result in the bio being freed.
1133 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1134 r1_bio_write_done(r1_bio
);
1135 /* We need another r1_bio. It has already been counted
1136 * in bio->bi_phys_segments
1138 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1139 r1_bio
->master_bio
= bio
;
1140 r1_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1142 r1_bio
->mddev
= mddev
;
1143 r1_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1147 r1_bio_write_done(r1_bio
);
1149 /* In case raid1d snuck in to freeze_array */
1150 wake_up(&conf
->wait_barrier
);
1152 if (do_sync
|| !bitmap
|| !plugged
)
1153 md_wakeup_thread(mddev
->thread
);
1158 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1160 struct r1conf
*conf
= mddev
->private;
1163 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1164 conf
->raid_disks
- mddev
->degraded
);
1166 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1167 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1168 seq_printf(seq
, "%s",
1169 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1172 seq_printf(seq
, "]");
1176 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1178 char b
[BDEVNAME_SIZE
];
1179 struct r1conf
*conf
= mddev
->private;
1182 * If it is not operational, then we have already marked it as dead
1183 * else if it is the last working disks, ignore the error, let the
1184 * next level up know.
1185 * else mark the drive as failed
1187 if (test_bit(In_sync
, &rdev
->flags
)
1188 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1190 * Don't fail the drive, act as though we were just a
1191 * normal single drive.
1192 * However don't try a recovery from this drive as
1193 * it is very likely to fail.
1195 conf
->recovery_disabled
= mddev
->recovery_disabled
;
1198 set_bit(Blocked
, &rdev
->flags
);
1199 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1200 unsigned long flags
;
1201 spin_lock_irqsave(&conf
->device_lock
, flags
);
1203 set_bit(Faulty
, &rdev
->flags
);
1204 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1206 * if recovery is running, make sure it aborts.
1208 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1210 set_bit(Faulty
, &rdev
->flags
);
1211 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1213 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1214 "md/raid1:%s: Operation continuing on %d devices.\n",
1215 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1216 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1219 static void print_conf(struct r1conf
*conf
)
1223 printk(KERN_DEBUG
"RAID1 conf printout:\n");
1225 printk(KERN_DEBUG
"(!conf)\n");
1228 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1232 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1233 char b
[BDEVNAME_SIZE
];
1234 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1236 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1237 i
, !test_bit(In_sync
, &rdev
->flags
),
1238 !test_bit(Faulty
, &rdev
->flags
),
1239 bdevname(rdev
->bdev
,b
));
1244 static void close_sync(struct r1conf
*conf
)
1247 allow_barrier(conf
);
1249 mempool_destroy(conf
->r1buf_pool
);
1250 conf
->r1buf_pool
= NULL
;
1253 static int raid1_spare_active(struct mddev
*mddev
)
1256 struct r1conf
*conf
= mddev
->private;
1258 unsigned long flags
;
1261 * Find all failed disks within the RAID1 configuration
1262 * and mark them readable.
1263 * Called under mddev lock, so rcu protection not needed.
1265 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1266 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1268 && !test_bit(Faulty
, &rdev
->flags
)
1269 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1271 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
1274 spin_lock_irqsave(&conf
->device_lock
, flags
);
1275 mddev
->degraded
-= count
;
1276 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1283 static int raid1_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1285 struct r1conf
*conf
= mddev
->private;
1288 struct mirror_info
*p
;
1290 int last
= mddev
->raid_disks
- 1;
1292 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
1295 if (rdev
->raid_disk
>= 0)
1296 first
= last
= rdev
->raid_disk
;
1298 for (mirror
= first
; mirror
<= last
; mirror
++)
1299 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1301 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1302 rdev
->data_offset
<< 9);
1303 /* as we don't honour merge_bvec_fn, we must
1304 * never risk violating it, so limit
1305 * ->max_segments to one lying with a single
1306 * page, as a one page request is never in
1309 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1310 blk_queue_max_segments(mddev
->queue
, 1);
1311 blk_queue_segment_boundary(mddev
->queue
,
1312 PAGE_CACHE_SIZE
- 1);
1315 p
->head_position
= 0;
1316 rdev
->raid_disk
= mirror
;
1318 /* As all devices are equivalent, we don't need a full recovery
1319 * if this was recently any drive of the array
1321 if (rdev
->saved_raid_disk
< 0)
1323 rcu_assign_pointer(p
->rdev
, rdev
);
1326 md_integrity_add_rdev(rdev
, mddev
);
1331 static int raid1_remove_disk(struct mddev
*mddev
, int number
)
1333 struct r1conf
*conf
= mddev
->private;
1335 struct md_rdev
*rdev
;
1336 struct mirror_info
*p
= conf
->mirrors
+ number
;
1341 if (test_bit(In_sync
, &rdev
->flags
) ||
1342 atomic_read(&rdev
->nr_pending
)) {
1346 /* Only remove non-faulty devices if recovery
1349 if (!test_bit(Faulty
, &rdev
->flags
) &&
1350 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
1351 mddev
->degraded
< conf
->raid_disks
) {
1357 if (atomic_read(&rdev
->nr_pending
)) {
1358 /* lost the race, try later */
1363 err
= md_integrity_register(mddev
);
1372 static void end_sync_read(struct bio
*bio
, int error
)
1374 struct r1bio
*r1_bio
= bio
->bi_private
;
1376 update_head_pos(r1_bio
->read_disk
, r1_bio
);
1379 * we have read a block, now it needs to be re-written,
1380 * or re-read if the read failed.
1381 * We don't do much here, just schedule handling by raid1d
1383 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1384 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1386 if (atomic_dec_and_test(&r1_bio
->remaining
))
1387 reschedule_retry(r1_bio
);
1390 static void end_sync_write(struct bio
*bio
, int error
)
1392 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1393 struct r1bio
*r1_bio
= bio
->bi_private
;
1394 struct mddev
*mddev
= r1_bio
->mddev
;
1395 struct r1conf
*conf
= mddev
->private;
1400 mirror
= find_bio_disk(r1_bio
, bio
);
1403 sector_t sync_blocks
= 0;
1404 sector_t s
= r1_bio
->sector
;
1405 long sectors_to_go
= r1_bio
->sectors
;
1406 /* make sure these bits doesn't get cleared. */
1408 bitmap_end_sync(mddev
->bitmap
, s
,
1411 sectors_to_go
-= sync_blocks
;
1412 } while (sectors_to_go
> 0);
1413 set_bit(WriteErrorSeen
,
1414 &conf
->mirrors
[mirror
].rdev
->flags
);
1415 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
1416 } else if (is_badblock(conf
->mirrors
[mirror
].rdev
,
1419 &first_bad
, &bad_sectors
) &&
1420 !is_badblock(conf
->mirrors
[r1_bio
->read_disk
].rdev
,
1423 &first_bad
, &bad_sectors
)
1425 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
1427 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1428 int s
= r1_bio
->sectors
;
1429 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
1430 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1431 reschedule_retry(r1_bio
);
1434 md_done_sync(mddev
, s
, uptodate
);
1439 static int r1_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1440 int sectors
, struct page
*page
, int rw
)
1442 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
1446 set_bit(WriteErrorSeen
, &rdev
->flags
);
1447 /* need to record an error - either for the block or the device */
1448 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1449 md_error(rdev
->mddev
, rdev
);
1453 static int fix_sync_read_error(struct r1bio
*r1_bio
)
1455 /* Try some synchronous reads of other devices to get
1456 * good data, much like with normal read errors. Only
1457 * read into the pages we already have so we don't
1458 * need to re-issue the read request.
1459 * We don't need to freeze the array, because being in an
1460 * active sync request, there is no normal IO, and
1461 * no overlapping syncs.
1462 * We don't need to check is_badblock() again as we
1463 * made sure that anything with a bad block in range
1464 * will have bi_end_io clear.
1466 struct mddev
*mddev
= r1_bio
->mddev
;
1467 struct r1conf
*conf
= mddev
->private;
1468 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1469 sector_t sect
= r1_bio
->sector
;
1470 int sectors
= r1_bio
->sectors
;
1475 int d
= r1_bio
->read_disk
;
1477 struct md_rdev
*rdev
;
1480 if (s
> (PAGE_SIZE
>>9))
1483 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1484 /* No rcu protection needed here devices
1485 * can only be removed when no resync is
1486 * active, and resync is currently active
1488 rdev
= conf
->mirrors
[d
].rdev
;
1489 if (sync_page_io(rdev
, sect
, s
<<9,
1490 bio
->bi_io_vec
[idx
].bv_page
,
1497 if (d
== conf
->raid_disks
)
1499 } while (!success
&& d
!= r1_bio
->read_disk
);
1502 char b
[BDEVNAME_SIZE
];
1504 /* Cannot read from anywhere, this block is lost.
1505 * Record a bad block on each device. If that doesn't
1506 * work just disable and interrupt the recovery.
1507 * Don't fail devices as that won't really help.
1509 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O read error"
1510 " for block %llu\n",
1512 bdevname(bio
->bi_bdev
, b
),
1513 (unsigned long long)r1_bio
->sector
);
1514 for (d
= 0; d
< conf
->raid_disks
; d
++) {
1515 rdev
= conf
->mirrors
[d
].rdev
;
1516 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
1518 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
1522 conf
->recovery_disabled
=
1523 mddev
->recovery_disabled
;
1524 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1525 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1537 /* write it back and re-read */
1538 while (d
!= r1_bio
->read_disk
) {
1540 d
= conf
->raid_disks
;
1542 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1544 rdev
= conf
->mirrors
[d
].rdev
;
1545 if (r1_sync_page_io(rdev
, sect
, s
,
1546 bio
->bi_io_vec
[idx
].bv_page
,
1548 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
1549 rdev_dec_pending(rdev
, mddev
);
1553 while (d
!= r1_bio
->read_disk
) {
1555 d
= conf
->raid_disks
;
1557 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1559 rdev
= conf
->mirrors
[d
].rdev
;
1560 if (r1_sync_page_io(rdev
, sect
, s
,
1561 bio
->bi_io_vec
[idx
].bv_page
,
1563 atomic_add(s
, &rdev
->corrected_errors
);
1569 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1570 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1574 static int process_checks(struct r1bio
*r1_bio
)
1576 /* We have read all readable devices. If we haven't
1577 * got the block, then there is no hope left.
1578 * If we have, then we want to do a comparison
1579 * and skip the write if everything is the same.
1580 * If any blocks failed to read, then we need to
1581 * attempt an over-write
1583 struct mddev
*mddev
= r1_bio
->mddev
;
1584 struct r1conf
*conf
= mddev
->private;
1588 for (primary
= 0; primary
< conf
->raid_disks
; primary
++)
1589 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
1590 test_bit(BIO_UPTODATE
, &r1_bio
->bios
[primary
]->bi_flags
)) {
1591 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
1592 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
1595 r1_bio
->read_disk
= primary
;
1596 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1598 int vcnt
= r1_bio
->sectors
>> (PAGE_SHIFT
- 9);
1599 struct bio
*pbio
= r1_bio
->bios
[primary
];
1600 struct bio
*sbio
= r1_bio
->bios
[i
];
1603 if (r1_bio
->bios
[i
]->bi_end_io
!= end_sync_read
)
1606 if (test_bit(BIO_UPTODATE
, &sbio
->bi_flags
)) {
1607 for (j
= vcnt
; j
-- ; ) {
1609 p
= pbio
->bi_io_vec
[j
].bv_page
;
1610 s
= sbio
->bi_io_vec
[j
].bv_page
;
1611 if (memcmp(page_address(p
),
1619 mddev
->resync_mismatches
+= r1_bio
->sectors
;
1620 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
1621 && test_bit(BIO_UPTODATE
, &sbio
->bi_flags
))) {
1622 /* No need to write to this device. */
1623 sbio
->bi_end_io
= NULL
;
1624 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
1627 /* fixup the bio for reuse */
1628 sbio
->bi_vcnt
= vcnt
;
1629 sbio
->bi_size
= r1_bio
->sectors
<< 9;
1631 sbio
->bi_phys_segments
= 0;
1632 sbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1633 sbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1634 sbio
->bi_next
= NULL
;
1635 sbio
->bi_sector
= r1_bio
->sector
+
1636 conf
->mirrors
[i
].rdev
->data_offset
;
1637 sbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1638 size
= sbio
->bi_size
;
1639 for (j
= 0; j
< vcnt
; j
++) {
1641 bi
= &sbio
->bi_io_vec
[j
];
1643 if (size
> PAGE_SIZE
)
1644 bi
->bv_len
= PAGE_SIZE
;
1648 memcpy(page_address(bi
->bv_page
),
1649 page_address(pbio
->bi_io_vec
[j
].bv_page
),
1656 static void sync_request_write(struct mddev
*mddev
, struct r1bio
*r1_bio
)
1658 struct r1conf
*conf
= mddev
->private;
1660 int disks
= conf
->raid_disks
;
1661 struct bio
*bio
, *wbio
;
1663 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1665 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
1666 /* ouch - failed to read all of that. */
1667 if (!fix_sync_read_error(r1_bio
))
1670 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
1671 if (process_checks(r1_bio
) < 0)
1676 atomic_set(&r1_bio
->remaining
, 1);
1677 for (i
= 0; i
< disks
; i
++) {
1678 wbio
= r1_bio
->bios
[i
];
1679 if (wbio
->bi_end_io
== NULL
||
1680 (wbio
->bi_end_io
== end_sync_read
&&
1681 (i
== r1_bio
->read_disk
||
1682 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
1685 wbio
->bi_rw
= WRITE
;
1686 wbio
->bi_end_io
= end_sync_write
;
1687 atomic_inc(&r1_bio
->remaining
);
1688 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, wbio
->bi_size
>> 9);
1690 generic_make_request(wbio
);
1693 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1694 /* if we're here, all write(s) have completed, so clean up */
1695 md_done_sync(mddev
, r1_bio
->sectors
, 1);
1701 * This is a kernel thread which:
1703 * 1. Retries failed read operations on working mirrors.
1704 * 2. Updates the raid superblock when problems encounter.
1705 * 3. Performs writes following reads for array synchronising.
1708 static void fix_read_error(struct r1conf
*conf
, int read_disk
,
1709 sector_t sect
, int sectors
)
1711 struct mddev
*mddev
= conf
->mddev
;
1717 struct md_rdev
*rdev
;
1719 if (s
> (PAGE_SIZE
>>9))
1723 /* Note: no rcu protection needed here
1724 * as this is synchronous in the raid1d thread
1725 * which is the thread that might remove
1726 * a device. If raid1d ever becomes multi-threaded....
1731 rdev
= conf
->mirrors
[d
].rdev
;
1733 test_bit(In_sync
, &rdev
->flags
) &&
1734 is_badblock(rdev
, sect
, s
,
1735 &first_bad
, &bad_sectors
) == 0 &&
1736 sync_page_io(rdev
, sect
, s
<<9,
1737 conf
->tmppage
, READ
, false))
1741 if (d
== conf
->raid_disks
)
1744 } while (!success
&& d
!= read_disk
);
1747 /* Cannot read from anywhere - mark it bad */
1748 struct md_rdev
*rdev
= conf
->mirrors
[read_disk
].rdev
;
1749 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
1750 md_error(mddev
, rdev
);
1753 /* write it back and re-read */
1755 while (d
!= read_disk
) {
1757 d
= conf
->raid_disks
;
1759 rdev
= conf
->mirrors
[d
].rdev
;
1761 test_bit(In_sync
, &rdev
->flags
))
1762 r1_sync_page_io(rdev
, sect
, s
,
1763 conf
->tmppage
, WRITE
);
1766 while (d
!= read_disk
) {
1767 char b
[BDEVNAME_SIZE
];
1769 d
= conf
->raid_disks
;
1771 rdev
= conf
->mirrors
[d
].rdev
;
1773 test_bit(In_sync
, &rdev
->flags
)) {
1774 if (r1_sync_page_io(rdev
, sect
, s
,
1775 conf
->tmppage
, READ
)) {
1776 atomic_add(s
, &rdev
->corrected_errors
);
1778 "md/raid1:%s: read error corrected "
1779 "(%d sectors at %llu on %s)\n",
1781 (unsigned long long)(sect
+
1783 bdevname(rdev
->bdev
, b
));
1792 static void bi_complete(struct bio
*bio
, int error
)
1794 complete((struct completion
*)bio
->bi_private
);
1797 static int submit_bio_wait(int rw
, struct bio
*bio
)
1799 struct completion event
;
1802 init_completion(&event
);
1803 bio
->bi_private
= &event
;
1804 bio
->bi_end_io
= bi_complete
;
1805 submit_bio(rw
, bio
);
1806 wait_for_completion(&event
);
1808 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1811 static int narrow_write_error(struct r1bio
*r1_bio
, int i
)
1813 struct mddev
*mddev
= r1_bio
->mddev
;
1814 struct r1conf
*conf
= mddev
->private;
1815 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1817 struct bio_vec
*vec
;
1819 /* bio has the data to be written to device 'i' where
1820 * we just recently had a write error.
1821 * We repeatedly clone the bio and trim down to one block,
1822 * then try the write. Where the write fails we record
1824 * It is conceivable that the bio doesn't exactly align with
1825 * blocks. We must handle this somehow.
1827 * We currently own a reference on the rdev.
1833 int sect_to_write
= r1_bio
->sectors
;
1836 if (rdev
->badblocks
.shift
< 0)
1839 block_sectors
= 1 << rdev
->badblocks
.shift
;
1840 sector
= r1_bio
->sector
;
1841 sectors
= ((sector
+ block_sectors
)
1842 & ~(sector_t
)(block_sectors
- 1))
1845 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
1846 vcnt
= r1_bio
->behind_page_count
;
1847 vec
= r1_bio
->behind_bvecs
;
1849 while (vec
[idx
].bv_page
== NULL
)
1852 vcnt
= r1_bio
->master_bio
->bi_vcnt
;
1853 vec
= r1_bio
->master_bio
->bi_io_vec
;
1854 idx
= r1_bio
->master_bio
->bi_idx
;
1856 while (sect_to_write
) {
1858 if (sectors
> sect_to_write
)
1859 sectors
= sect_to_write
;
1860 /* Write at 'sector' for 'sectors'*/
1862 wbio
= bio_alloc_mddev(GFP_NOIO
, vcnt
, mddev
);
1863 memcpy(wbio
->bi_io_vec
, vec
, vcnt
* sizeof(struct bio_vec
));
1864 wbio
->bi_sector
= r1_bio
->sector
;
1865 wbio
->bi_rw
= WRITE
;
1866 wbio
->bi_vcnt
= vcnt
;
1867 wbio
->bi_size
= r1_bio
->sectors
<< 9;
1870 md_trim_bio(wbio
, sector
- r1_bio
->sector
, sectors
);
1871 wbio
->bi_sector
+= rdev
->data_offset
;
1872 wbio
->bi_bdev
= rdev
->bdev
;
1873 if (submit_bio_wait(WRITE
, wbio
) == 0)
1875 ok
= rdev_set_badblocks(rdev
, sector
,
1880 sect_to_write
-= sectors
;
1882 sectors
= block_sectors
;
1887 static void handle_sync_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
1890 int s
= r1_bio
->sectors
;
1891 for (m
= 0; m
< conf
->raid_disks
; m
++) {
1892 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
1893 struct bio
*bio
= r1_bio
->bios
[m
];
1894 if (bio
->bi_end_io
== NULL
)
1896 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
1897 test_bit(R1BIO_MadeGood
, &r1_bio
->state
)) {
1898 rdev_clear_badblocks(rdev
, r1_bio
->sector
, s
);
1900 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
1901 test_bit(R1BIO_WriteError
, &r1_bio
->state
)) {
1902 if (!rdev_set_badblocks(rdev
, r1_bio
->sector
, s
, 0))
1903 md_error(conf
->mddev
, rdev
);
1907 md_done_sync(conf
->mddev
, s
, 1);
1910 static void handle_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
1913 for (m
= 0; m
< conf
->raid_disks
; m
++)
1914 if (r1_bio
->bios
[m
] == IO_MADE_GOOD
) {
1915 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
1916 rdev_clear_badblocks(rdev
,
1919 rdev_dec_pending(rdev
, conf
->mddev
);
1920 } else if (r1_bio
->bios
[m
] != NULL
) {
1921 /* This drive got a write error. We need to
1922 * narrow down and record precise write
1925 if (!narrow_write_error(r1_bio
, m
)) {
1926 md_error(conf
->mddev
,
1927 conf
->mirrors
[m
].rdev
);
1928 /* an I/O failed, we can't clear the bitmap */
1929 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1931 rdev_dec_pending(conf
->mirrors
[m
].rdev
,
1934 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1935 close_write(r1_bio
);
1936 raid_end_bio_io(r1_bio
);
1939 static void handle_read_error(struct r1conf
*conf
, struct r1bio
*r1_bio
)
1943 struct mddev
*mddev
= conf
->mddev
;
1945 char b
[BDEVNAME_SIZE
];
1946 struct md_rdev
*rdev
;
1948 clear_bit(R1BIO_ReadError
, &r1_bio
->state
);
1949 /* we got a read error. Maybe the drive is bad. Maybe just
1950 * the block and we can fix it.
1951 * We freeze all other IO, and try reading the block from
1952 * other devices. When we find one, we re-write
1953 * and check it that fixes the read error.
1954 * This is all done synchronously while the array is
1957 if (mddev
->ro
== 0) {
1959 fix_read_error(conf
, r1_bio
->read_disk
,
1960 r1_bio
->sector
, r1_bio
->sectors
);
1961 unfreeze_array(conf
);
1963 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
1965 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1966 bdevname(bio
->bi_bdev
, b
);
1968 disk
= read_balance(conf
, r1_bio
, &max_sectors
);
1970 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O"
1971 " read error for block %llu\n",
1972 mdname(mddev
), b
, (unsigned long long)r1_bio
->sector
);
1973 raid_end_bio_io(r1_bio
);
1975 const unsigned long do_sync
1976 = r1_bio
->master_bio
->bi_rw
& REQ_SYNC
;
1978 r1_bio
->bios
[r1_bio
->read_disk
] =
1979 mddev
->ro
? IO_BLOCKED
: NULL
;
1982 r1_bio
->read_disk
= disk
;
1983 bio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
1984 md_trim_bio(bio
, r1_bio
->sector
- bio
->bi_sector
, max_sectors
);
1985 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
1986 rdev
= conf
->mirrors
[disk
].rdev
;
1987 printk_ratelimited(KERN_ERR
1988 "md/raid1:%s: redirecting sector %llu"
1989 " to other mirror: %s\n",
1991 (unsigned long long)r1_bio
->sector
,
1992 bdevname(rdev
->bdev
, b
));
1993 bio
->bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
1994 bio
->bi_bdev
= rdev
->bdev
;
1995 bio
->bi_end_io
= raid1_end_read_request
;
1996 bio
->bi_rw
= READ
| do_sync
;
1997 bio
->bi_private
= r1_bio
;
1998 if (max_sectors
< r1_bio
->sectors
) {
1999 /* Drat - have to split this up more */
2000 struct bio
*mbio
= r1_bio
->master_bio
;
2001 int sectors_handled
= (r1_bio
->sector
+ max_sectors
2003 r1_bio
->sectors
= max_sectors
;
2004 spin_lock_irq(&conf
->device_lock
);
2005 if (mbio
->bi_phys_segments
== 0)
2006 mbio
->bi_phys_segments
= 2;
2008 mbio
->bi_phys_segments
++;
2009 spin_unlock_irq(&conf
->device_lock
);
2010 generic_make_request(bio
);
2013 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
2015 r1_bio
->master_bio
= mbio
;
2016 r1_bio
->sectors
= (mbio
->bi_size
>> 9)
2019 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
2020 r1_bio
->mddev
= mddev
;
2021 r1_bio
->sector
= mbio
->bi_sector
+ sectors_handled
;
2025 generic_make_request(bio
);
2029 static void raid1d(struct mddev
*mddev
)
2031 struct r1bio
*r1_bio
;
2032 unsigned long flags
;
2033 struct r1conf
*conf
= mddev
->private;
2034 struct list_head
*head
= &conf
->retry_list
;
2035 struct blk_plug plug
;
2037 md_check_recovery(mddev
);
2039 blk_start_plug(&plug
);
2042 if (atomic_read(&mddev
->plug_cnt
) == 0)
2043 flush_pending_writes(conf
);
2045 spin_lock_irqsave(&conf
->device_lock
, flags
);
2046 if (list_empty(head
)) {
2047 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2050 r1_bio
= list_entry(head
->prev
, struct r1bio
, retry_list
);
2051 list_del(head
->prev
);
2053 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2055 mddev
= r1_bio
->mddev
;
2056 conf
= mddev
->private;
2057 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
2058 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2059 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2060 handle_sync_write_finished(conf
, r1_bio
);
2062 sync_request_write(mddev
, r1_bio
);
2063 } else if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2064 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2065 handle_write_finished(conf
, r1_bio
);
2066 else if (test_bit(R1BIO_ReadError
, &r1_bio
->state
))
2067 handle_read_error(conf
, r1_bio
);
2069 /* just a partial read to be scheduled from separate
2072 generic_make_request(r1_bio
->bios
[r1_bio
->read_disk
]);
2075 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2076 md_check_recovery(mddev
);
2078 blk_finish_plug(&plug
);
2082 static int init_resync(struct r1conf
*conf
)
2086 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2087 BUG_ON(conf
->r1buf_pool
);
2088 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
2090 if (!conf
->r1buf_pool
)
2092 conf
->next_resync
= 0;
2097 * perform a "sync" on one "block"
2099 * We need to make sure that no normal I/O request - particularly write
2100 * requests - conflict with active sync requests.
2102 * This is achieved by tracking pending requests and a 'barrier' concept
2103 * that can be installed to exclude normal IO requests.
2106 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
2108 struct r1conf
*conf
= mddev
->private;
2109 struct r1bio
*r1_bio
;
2111 sector_t max_sector
, nr_sectors
;
2115 int write_targets
= 0, read_targets
= 0;
2116 sector_t sync_blocks
;
2117 int still_degraded
= 0;
2118 int good_sectors
= RESYNC_SECTORS
;
2119 int min_bad
= 0; /* number of sectors that are bad in all devices */
2121 if (!conf
->r1buf_pool
)
2122 if (init_resync(conf
))
2125 max_sector
= mddev
->dev_sectors
;
2126 if (sector_nr
>= max_sector
) {
2127 /* If we aborted, we need to abort the
2128 * sync on the 'current' bitmap chunk (there will
2129 * only be one in raid1 resync.
2130 * We can find the current addess in mddev->curr_resync
2132 if (mddev
->curr_resync
< max_sector
) /* aborted */
2133 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2135 else /* completed sync */
2138 bitmap_close_sync(mddev
->bitmap
);
2143 if (mddev
->bitmap
== NULL
&&
2144 mddev
->recovery_cp
== MaxSector
&&
2145 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2146 conf
->fullsync
== 0) {
2148 return max_sector
- sector_nr
;
2150 /* before building a request, check if we can skip these blocks..
2151 * This call the bitmap_start_sync doesn't actually record anything
2153 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
2154 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2155 /* We can skip this block, and probably several more */
2160 * If there is non-resync activity waiting for a turn,
2161 * and resync is going fast enough,
2162 * then let it though before starting on this new sync request.
2164 if (!go_faster
&& conf
->nr_waiting
)
2165 msleep_interruptible(1000);
2167 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
2168 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
2169 raise_barrier(conf
);
2171 conf
->next_resync
= sector_nr
;
2175 * If we get a correctably read error during resync or recovery,
2176 * we might want to read from a different device. So we
2177 * flag all drives that could conceivably be read from for READ,
2178 * and any others (which will be non-In_sync devices) for WRITE.
2179 * If a read fails, we try reading from something else for which READ
2183 r1_bio
->mddev
= mddev
;
2184 r1_bio
->sector
= sector_nr
;
2186 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
2188 for (i
=0; i
< conf
->raid_disks
; i
++) {
2189 struct md_rdev
*rdev
;
2190 bio
= r1_bio
->bios
[i
];
2192 /* take from bio_init */
2193 bio
->bi_next
= NULL
;
2194 bio
->bi_flags
&= ~(BIO_POOL_MASK
-1);
2195 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
2196 bio
->bi_comp_cpu
= -1;
2200 bio
->bi_phys_segments
= 0;
2202 bio
->bi_end_io
= NULL
;
2203 bio
->bi_private
= NULL
;
2205 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
2207 test_bit(Faulty
, &rdev
->flags
)) {
2209 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
2211 bio
->bi_end_io
= end_sync_write
;
2214 /* may need to read from here */
2215 sector_t first_bad
= MaxSector
;
2218 if (is_badblock(rdev
, sector_nr
, good_sectors
,
2219 &first_bad
, &bad_sectors
)) {
2220 if (first_bad
> sector_nr
)
2221 good_sectors
= first_bad
- sector_nr
;
2223 bad_sectors
-= (sector_nr
- first_bad
);
2225 min_bad
> bad_sectors
)
2226 min_bad
= bad_sectors
;
2229 if (sector_nr
< first_bad
) {
2230 if (test_bit(WriteMostly
, &rdev
->flags
)) {
2238 bio
->bi_end_io
= end_sync_read
;
2242 if (bio
->bi_end_io
) {
2243 atomic_inc(&rdev
->nr_pending
);
2244 bio
->bi_sector
= sector_nr
+ rdev
->data_offset
;
2245 bio
->bi_bdev
= rdev
->bdev
;
2246 bio
->bi_private
= r1_bio
;
2252 r1_bio
->read_disk
= disk
;
2254 if (read_targets
== 0 && min_bad
> 0) {
2255 /* These sectors are bad on all InSync devices, so we
2256 * need to mark them bad on all write targets
2259 for (i
= 0 ; i
< conf
->raid_disks
; i
++)
2260 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_write
) {
2261 struct md_rdev
*rdev
=
2262 rcu_dereference(conf
->mirrors
[i
].rdev
);
2263 ok
= rdev_set_badblocks(rdev
, sector_nr
,
2267 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2272 /* Cannot record the badblocks, so need to
2274 * If there are multiple read targets, could just
2275 * fail the really bad ones ???
2277 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2278 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2284 if (min_bad
> 0 && min_bad
< good_sectors
) {
2285 /* only resync enough to reach the next bad->good
2287 good_sectors
= min_bad
;
2290 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
2291 /* extra read targets are also write targets */
2292 write_targets
+= read_targets
-1;
2294 if (write_targets
== 0 || read_targets
== 0) {
2295 /* There is nowhere to write, so all non-sync
2296 * drives must be failed - so we are finished
2298 sector_t rv
= max_sector
- sector_nr
;
2304 if (max_sector
> mddev
->resync_max
)
2305 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2306 if (max_sector
> sector_nr
+ good_sectors
)
2307 max_sector
= sector_nr
+ good_sectors
;
2312 int len
= PAGE_SIZE
;
2313 if (sector_nr
+ (len
>>9) > max_sector
)
2314 len
= (max_sector
- sector_nr
) << 9;
2317 if (sync_blocks
== 0) {
2318 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2319 &sync_blocks
, still_degraded
) &&
2321 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2323 BUG_ON(sync_blocks
< (PAGE_SIZE
>>9));
2324 if ((len
>> 9) > sync_blocks
)
2325 len
= sync_blocks
<<9;
2328 for (i
=0 ; i
< conf
->raid_disks
; i
++) {
2329 bio
= r1_bio
->bios
[i
];
2330 if (bio
->bi_end_io
) {
2331 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2332 if (bio_add_page(bio
, page
, len
, 0) == 0) {
2334 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2337 bio
= r1_bio
->bios
[i
];
2338 if (bio
->bi_end_io
==NULL
)
2340 /* remove last page from this bio */
2342 bio
->bi_size
-= len
;
2343 bio
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
2349 nr_sectors
+= len
>>9;
2350 sector_nr
+= len
>>9;
2351 sync_blocks
-= (len
>>9);
2352 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
2354 r1_bio
->sectors
= nr_sectors
;
2356 /* For a user-requested sync, we read all readable devices and do a
2359 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2360 atomic_set(&r1_bio
->remaining
, read_targets
);
2361 for (i
=0; i
<conf
->raid_disks
; i
++) {
2362 bio
= r1_bio
->bios
[i
];
2363 if (bio
->bi_end_io
== end_sync_read
) {
2364 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2365 generic_make_request(bio
);
2369 atomic_set(&r1_bio
->remaining
, 1);
2370 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2371 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2372 generic_make_request(bio
);
2378 static sector_t
raid1_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2383 return mddev
->dev_sectors
;
2386 static struct r1conf
*setup_conf(struct mddev
*mddev
)
2388 struct r1conf
*conf
;
2390 struct mirror_info
*disk
;
2391 struct md_rdev
*rdev
;
2394 conf
= kzalloc(sizeof(struct r1conf
), GFP_KERNEL
);
2398 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
2403 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2407 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
2408 if (!conf
->poolinfo
)
2410 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
;
2411 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2414 if (!conf
->r1bio_pool
)
2417 conf
->poolinfo
->mddev
= mddev
;
2419 spin_lock_init(&conf
->device_lock
);
2420 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
2421 int disk_idx
= rdev
->raid_disk
;
2422 if (disk_idx
>= mddev
->raid_disks
2425 disk
= conf
->mirrors
+ disk_idx
;
2429 disk
->head_position
= 0;
2431 conf
->raid_disks
= mddev
->raid_disks
;
2432 conf
->mddev
= mddev
;
2433 INIT_LIST_HEAD(&conf
->retry_list
);
2435 spin_lock_init(&conf
->resync_lock
);
2436 init_waitqueue_head(&conf
->wait_barrier
);
2438 bio_list_init(&conf
->pending_bio_list
);
2439 conf
->pending_count
= 0;
2440 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
2442 conf
->last_used
= -1;
2443 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2445 disk
= conf
->mirrors
+ i
;
2448 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2449 disk
->head_position
= 0;
2452 } else if (conf
->last_used
< 0)
2454 * The first working device is used as a
2455 * starting point to read balancing.
2457 conf
->last_used
= i
;
2461 if (conf
->last_used
< 0) {
2462 printk(KERN_ERR
"md/raid1:%s: no operational mirrors\n",
2467 conf
->thread
= md_register_thread(raid1d
, mddev
, NULL
);
2468 if (!conf
->thread
) {
2470 "md/raid1:%s: couldn't allocate thread\n",
2479 if (conf
->r1bio_pool
)
2480 mempool_destroy(conf
->r1bio_pool
);
2481 kfree(conf
->mirrors
);
2482 safe_put_page(conf
->tmppage
);
2483 kfree(conf
->poolinfo
);
2486 return ERR_PTR(err
);
2489 static int run(struct mddev
*mddev
)
2491 struct r1conf
*conf
;
2493 struct md_rdev
*rdev
;
2495 if (mddev
->level
!= 1) {
2496 printk(KERN_ERR
"md/raid1:%s: raid level not set to mirroring (%d)\n",
2497 mdname(mddev
), mddev
->level
);
2500 if (mddev
->reshape_position
!= MaxSector
) {
2501 printk(KERN_ERR
"md/raid1:%s: reshape_position set but not supported\n",
2506 * copy the already verified devices into our private RAID1
2507 * bookkeeping area. [whatever we allocate in run(),
2508 * should be freed in stop()]
2510 if (mddev
->private == NULL
)
2511 conf
= setup_conf(mddev
);
2513 conf
= mddev
->private;
2516 return PTR_ERR(conf
);
2518 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
2519 if (!mddev
->gendisk
)
2521 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2522 rdev
->data_offset
<< 9);
2523 /* as we don't honour merge_bvec_fn, we must never risk
2524 * violating it, so limit ->max_segments to 1 lying within
2525 * a single page, as a one page request is never in violation.
2527 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
2528 blk_queue_max_segments(mddev
->queue
, 1);
2529 blk_queue_segment_boundary(mddev
->queue
,
2530 PAGE_CACHE_SIZE
- 1);
2534 mddev
->degraded
= 0;
2535 for (i
=0; i
< conf
->raid_disks
; i
++)
2536 if (conf
->mirrors
[i
].rdev
== NULL
||
2537 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
2538 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2541 if (conf
->raid_disks
- mddev
->degraded
== 1)
2542 mddev
->recovery_cp
= MaxSector
;
2544 if (mddev
->recovery_cp
!= MaxSector
)
2545 printk(KERN_NOTICE
"md/raid1:%s: not clean"
2546 " -- starting background reconstruction\n",
2549 "md/raid1:%s: active with %d out of %d mirrors\n",
2550 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2554 * Ok, everything is just fine now
2556 mddev
->thread
= conf
->thread
;
2557 conf
->thread
= NULL
;
2558 mddev
->private = conf
;
2560 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
2563 mddev
->queue
->backing_dev_info
.congested_fn
= raid1_congested
;
2564 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2566 return md_integrity_register(mddev
);
2569 static int stop(struct mddev
*mddev
)
2571 struct r1conf
*conf
= mddev
->private;
2572 struct bitmap
*bitmap
= mddev
->bitmap
;
2574 /* wait for behind writes to complete */
2575 if (bitmap
&& atomic_read(&bitmap
->behind_writes
) > 0) {
2576 printk(KERN_INFO
"md/raid1:%s: behind writes in progress - waiting to stop.\n",
2578 /* need to kick something here to make sure I/O goes? */
2579 wait_event(bitmap
->behind_wait
,
2580 atomic_read(&bitmap
->behind_writes
) == 0);
2583 raise_barrier(conf
);
2584 lower_barrier(conf
);
2586 md_unregister_thread(&mddev
->thread
);
2587 if (conf
->r1bio_pool
)
2588 mempool_destroy(conf
->r1bio_pool
);
2589 kfree(conf
->mirrors
);
2590 kfree(conf
->poolinfo
);
2592 mddev
->private = NULL
;
2596 static int raid1_resize(struct mddev
*mddev
, sector_t sectors
)
2598 /* no resync is happening, and there is enough space
2599 * on all devices, so we can resize.
2600 * We need to make sure resync covers any new space.
2601 * If the array is shrinking we should possibly wait until
2602 * any io in the removed space completes, but it hardly seems
2605 md_set_array_sectors(mddev
, raid1_size(mddev
, sectors
, 0));
2606 if (mddev
->array_sectors
> raid1_size(mddev
, sectors
, 0))
2608 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
2609 revalidate_disk(mddev
->gendisk
);
2610 if (sectors
> mddev
->dev_sectors
&&
2611 mddev
->recovery_cp
> mddev
->dev_sectors
) {
2612 mddev
->recovery_cp
= mddev
->dev_sectors
;
2613 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2615 mddev
->dev_sectors
= sectors
;
2616 mddev
->resync_max_sectors
= sectors
;
2620 static int raid1_reshape(struct mddev
*mddev
)
2623 * 1/ resize the r1bio_pool
2624 * 2/ resize conf->mirrors
2626 * We allocate a new r1bio_pool if we can.
2627 * Then raise a device barrier and wait until all IO stops.
2628 * Then resize conf->mirrors and swap in the new r1bio pool.
2630 * At the same time, we "pack" the devices so that all the missing
2631 * devices have the higher raid_disk numbers.
2633 mempool_t
*newpool
, *oldpool
;
2634 struct pool_info
*newpoolinfo
;
2635 struct mirror_info
*newmirrors
;
2636 struct r1conf
*conf
= mddev
->private;
2637 int cnt
, raid_disks
;
2638 unsigned long flags
;
2641 /* Cannot change chunk_size, layout, or level */
2642 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
2643 mddev
->layout
!= mddev
->new_layout
||
2644 mddev
->level
!= mddev
->new_level
) {
2645 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
2646 mddev
->new_layout
= mddev
->layout
;
2647 mddev
->new_level
= mddev
->level
;
2651 err
= md_allow_write(mddev
);
2655 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
2657 if (raid_disks
< conf
->raid_disks
) {
2659 for (d
= 0; d
< conf
->raid_disks
; d
++)
2660 if (conf
->mirrors
[d
].rdev
)
2662 if (cnt
> raid_disks
)
2666 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
2669 newpoolinfo
->mddev
= mddev
;
2670 newpoolinfo
->raid_disks
= raid_disks
;
2672 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2673 r1bio_pool_free
, newpoolinfo
);
2678 newmirrors
= kzalloc(sizeof(struct mirror_info
) * raid_disks
, GFP_KERNEL
);
2681 mempool_destroy(newpool
);
2685 raise_barrier(conf
);
2687 /* ok, everything is stopped */
2688 oldpool
= conf
->r1bio_pool
;
2689 conf
->r1bio_pool
= newpool
;
2691 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
2692 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
2693 if (rdev
&& rdev
->raid_disk
!= d2
) {
2694 sysfs_unlink_rdev(mddev
, rdev
);
2695 rdev
->raid_disk
= d2
;
2696 sysfs_unlink_rdev(mddev
, rdev
);
2697 if (sysfs_link_rdev(mddev
, rdev
))
2699 "md/raid1:%s: cannot register rd%d\n",
2700 mdname(mddev
), rdev
->raid_disk
);
2703 newmirrors
[d2
++].rdev
= rdev
;
2705 kfree(conf
->mirrors
);
2706 conf
->mirrors
= newmirrors
;
2707 kfree(conf
->poolinfo
);
2708 conf
->poolinfo
= newpoolinfo
;
2710 spin_lock_irqsave(&conf
->device_lock
, flags
);
2711 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
2712 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2713 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
2714 mddev
->delta_disks
= 0;
2716 conf
->last_used
= 0; /* just make sure it is in-range */
2717 lower_barrier(conf
);
2719 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2720 md_wakeup_thread(mddev
->thread
);
2722 mempool_destroy(oldpool
);
2726 static void raid1_quiesce(struct mddev
*mddev
, int state
)
2728 struct r1conf
*conf
= mddev
->private;
2731 case 2: /* wake for suspend */
2732 wake_up(&conf
->wait_barrier
);
2735 raise_barrier(conf
);
2738 lower_barrier(conf
);
2743 static void *raid1_takeover(struct mddev
*mddev
)
2745 /* raid1 can take over:
2746 * raid5 with 2 devices, any layout or chunk size
2748 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
2749 struct r1conf
*conf
;
2750 mddev
->new_level
= 1;
2751 mddev
->new_layout
= 0;
2752 mddev
->new_chunk_sectors
= 0;
2753 conf
= setup_conf(mddev
);
2758 return ERR_PTR(-EINVAL
);
2761 static struct md_personality raid1_personality
=
2765 .owner
= THIS_MODULE
,
2766 .make_request
= make_request
,
2770 .error_handler
= error
,
2771 .hot_add_disk
= raid1_add_disk
,
2772 .hot_remove_disk
= raid1_remove_disk
,
2773 .spare_active
= raid1_spare_active
,
2774 .sync_request
= sync_request
,
2775 .resize
= raid1_resize
,
2777 .check_reshape
= raid1_reshape
,
2778 .quiesce
= raid1_quiesce
,
2779 .takeover
= raid1_takeover
,
2782 static int __init
raid_init(void)
2784 return register_md_personality(&raid1_personality
);
2787 static void raid_exit(void)
2789 unregister_md_personality(&raid1_personality
);
2792 module_init(raid_init
);
2793 module_exit(raid_exit
);
2794 MODULE_LICENSE("GPL");
2795 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2796 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2797 MODULE_ALIAS("md-raid1");
2798 MODULE_ALIAS("md-level-1");
2800 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);