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 there are this many requests queue to be written by
50 * the raid1 thread, we become 'congested' to provide back-pressure
53 static int max_queued_requests
= 1024;
55 static void allow_barrier(struct r1conf
*conf
);
56 static void lower_barrier(struct r1conf
*conf
);
58 static void * r1bio_pool_alloc(gfp_t gfp_flags
, void *data
)
60 struct pool_info
*pi
= data
;
61 int size
= offsetof(struct r1bio
, bios
[pi
->raid_disks
]);
63 /* allocate a r1bio with room for raid_disks entries in the bios array */
64 return kzalloc(size
, gfp_flags
);
67 static void r1bio_pool_free(void *r1_bio
, void *data
)
72 #define RESYNC_BLOCK_SIZE (64*1024)
73 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
74 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
75 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
76 #define RESYNC_WINDOW (2048*1024)
78 static void * r1buf_pool_alloc(gfp_t gfp_flags
, void *data
)
80 struct pool_info
*pi
= data
;
86 r1_bio
= r1bio_pool_alloc(gfp_flags
, pi
);
91 * Allocate bios : 1 for reading, n-1 for writing
93 for (j
= pi
->raid_disks
; j
-- ; ) {
94 bio
= bio_kmalloc(gfp_flags
, RESYNC_PAGES
);
97 r1_bio
->bios
[j
] = bio
;
100 * Allocate RESYNC_PAGES data pages and attach them to
102 * If this is a user-requested check/repair, allocate
103 * RESYNC_PAGES for each bio.
105 if (test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
))
110 bio
= r1_bio
->bios
[j
];
111 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
112 page
= alloc_page(gfp_flags
);
116 bio
->bi_io_vec
[i
].bv_page
= page
;
120 /* If not user-requests, copy the page pointers to all bios */
121 if (!test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
)) {
122 for (i
=0; i
<RESYNC_PAGES
; i
++)
123 for (j
=1; j
<pi
->raid_disks
; j
++)
124 r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
=
125 r1_bio
->bios
[0]->bi_io_vec
[i
].bv_page
;
128 r1_bio
->master_bio
= NULL
;
133 for (j
=0 ; j
< pi
->raid_disks
; j
++)
134 for (i
=0; i
< r1_bio
->bios
[j
]->bi_vcnt
; i
++)
135 put_page(r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
138 while ( ++j
< pi
->raid_disks
)
139 bio_put(r1_bio
->bios
[j
]);
140 r1bio_pool_free(r1_bio
, data
);
144 static void r1buf_pool_free(void *__r1_bio
, void *data
)
146 struct pool_info
*pi
= data
;
148 struct r1bio
*r1bio
= __r1_bio
;
150 for (i
= 0; i
< RESYNC_PAGES
; i
++)
151 for (j
= pi
->raid_disks
; j
-- ;) {
153 r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
!=
154 r1bio
->bios
[0]->bi_io_vec
[i
].bv_page
)
155 safe_put_page(r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
157 for (i
=0 ; i
< pi
->raid_disks
; i
++)
158 bio_put(r1bio
->bios
[i
]);
160 r1bio_pool_free(r1bio
, data
);
163 static void put_all_bios(struct r1conf
*conf
, struct r1bio
*r1_bio
)
167 for (i
= 0; i
< conf
->raid_disks
; i
++) {
168 struct bio
**bio
= r1_bio
->bios
+ i
;
169 if (!BIO_SPECIAL(*bio
))
175 static void free_r1bio(struct r1bio
*r1_bio
)
177 struct r1conf
*conf
= r1_bio
->mddev
->private;
179 put_all_bios(conf
, r1_bio
);
180 mempool_free(r1_bio
, conf
->r1bio_pool
);
183 static void put_buf(struct r1bio
*r1_bio
)
185 struct r1conf
*conf
= r1_bio
->mddev
->private;
188 for (i
=0; i
<conf
->raid_disks
; i
++) {
189 struct bio
*bio
= r1_bio
->bios
[i
];
191 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
194 mempool_free(r1_bio
, conf
->r1buf_pool
);
199 static void reschedule_retry(struct r1bio
*r1_bio
)
202 struct mddev
*mddev
= r1_bio
->mddev
;
203 struct r1conf
*conf
= mddev
->private;
205 spin_lock_irqsave(&conf
->device_lock
, flags
);
206 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
208 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
210 wake_up(&conf
->wait_barrier
);
211 md_wakeup_thread(mddev
->thread
);
215 * raid_end_bio_io() is called when we have finished servicing a mirrored
216 * operation and are ready to return a success/failure code to the buffer
219 static void call_bio_endio(struct r1bio
*r1_bio
)
221 struct bio
*bio
= r1_bio
->master_bio
;
223 struct r1conf
*conf
= r1_bio
->mddev
->private;
225 if (bio
->bi_phys_segments
) {
227 spin_lock_irqsave(&conf
->device_lock
, flags
);
228 bio
->bi_phys_segments
--;
229 done
= (bio
->bi_phys_segments
== 0);
230 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
234 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
235 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
239 * Wake up any possible resync thread that waits for the device
246 static void raid_end_bio_io(struct r1bio
*r1_bio
)
248 struct bio
*bio
= r1_bio
->master_bio
;
250 /* if nobody has done the final endio yet, do it now */
251 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
252 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
253 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
254 (unsigned long long) bio
->bi_sector
,
255 (unsigned long long) bio
->bi_sector
+
256 (bio
->bi_size
>> 9) - 1);
258 call_bio_endio(r1_bio
);
264 * Update disk head position estimator based on IRQ completion info.
266 static inline void update_head_pos(int disk
, struct r1bio
*r1_bio
)
268 struct r1conf
*conf
= r1_bio
->mddev
->private;
270 conf
->mirrors
[disk
].head_position
=
271 r1_bio
->sector
+ (r1_bio
->sectors
);
275 * Find the disk number which triggered given bio
277 static int find_bio_disk(struct r1bio
*r1_bio
, struct bio
*bio
)
280 int raid_disks
= r1_bio
->mddev
->raid_disks
;
282 for (mirror
= 0; mirror
< raid_disks
; mirror
++)
283 if (r1_bio
->bios
[mirror
] == bio
)
286 BUG_ON(mirror
== raid_disks
);
287 update_head_pos(mirror
, r1_bio
);
292 static void raid1_end_read_request(struct bio
*bio
, int error
)
294 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
295 struct r1bio
*r1_bio
= bio
->bi_private
;
297 struct r1conf
*conf
= r1_bio
->mddev
->private;
299 mirror
= r1_bio
->read_disk
;
301 * this branch is our 'one mirror IO has finished' event handler:
303 update_head_pos(mirror
, r1_bio
);
306 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
308 /* If all other devices have failed, we want to return
309 * the error upwards rather than fail the last device.
310 * Here we redefine "uptodate" to mean "Don't want to retry"
313 spin_lock_irqsave(&conf
->device_lock
, flags
);
314 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
315 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
316 !test_bit(Faulty
, &conf
->mirrors
[mirror
].rdev
->flags
)))
318 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
322 raid_end_bio_io(r1_bio
);
327 char b
[BDEVNAME_SIZE
];
329 KERN_ERR
"md/raid1:%s: %s: "
330 "rescheduling sector %llu\n",
332 bdevname(conf
->mirrors
[mirror
].rdev
->bdev
,
334 (unsigned long long)r1_bio
->sector
);
335 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
336 reschedule_retry(r1_bio
);
339 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
342 static void close_write(struct r1bio
*r1_bio
)
344 /* it really is the end of this request */
345 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
346 /* free extra copy of the data pages */
347 int i
= r1_bio
->behind_page_count
;
349 safe_put_page(r1_bio
->behind_bvecs
[i
].bv_page
);
350 kfree(r1_bio
->behind_bvecs
);
351 r1_bio
->behind_bvecs
= NULL
;
353 /* clear the bitmap if all writes complete successfully */
354 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
356 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
357 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
358 md_write_end(r1_bio
->mddev
);
361 static void r1_bio_write_done(struct r1bio
*r1_bio
)
363 if (!atomic_dec_and_test(&r1_bio
->remaining
))
366 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
367 reschedule_retry(r1_bio
);
370 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
))
371 reschedule_retry(r1_bio
);
373 raid_end_bio_io(r1_bio
);
377 static void raid1_end_write_request(struct bio
*bio
, int error
)
379 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
380 struct r1bio
*r1_bio
= bio
->bi_private
;
381 int mirror
, behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
382 struct r1conf
*conf
= r1_bio
->mddev
->private;
383 struct bio
*to_put
= NULL
;
385 mirror
= find_bio_disk(r1_bio
, bio
);
388 * 'one mirror IO has finished' event handler:
391 set_bit(WriteErrorSeen
,
392 &conf
->mirrors
[mirror
].rdev
->flags
);
393 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
396 * Set R1BIO_Uptodate in our master bio, so that we
397 * will return a good error code for to the higher
398 * levels even if IO on some other mirrored buffer
401 * The 'master' represents the composite IO operation
402 * to user-side. So if something waits for IO, then it
403 * will wait for the 'master' bio.
408 r1_bio
->bios
[mirror
] = NULL
;
410 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
412 /* Maybe we can clear some bad blocks. */
413 if (is_badblock(conf
->mirrors
[mirror
].rdev
,
414 r1_bio
->sector
, r1_bio
->sectors
,
415 &first_bad
, &bad_sectors
)) {
416 r1_bio
->bios
[mirror
] = IO_MADE_GOOD
;
417 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
422 if (test_bit(WriteMostly
, &conf
->mirrors
[mirror
].rdev
->flags
))
423 atomic_dec(&r1_bio
->behind_remaining
);
426 * In behind mode, we ACK the master bio once the I/O
427 * has safely reached all non-writemostly
428 * disks. Setting the Returned bit ensures that this
429 * gets done only once -- we don't ever want to return
430 * -EIO here, instead we'll wait
432 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
433 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
434 /* Maybe we can return now */
435 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
436 struct bio
*mbio
= r1_bio
->master_bio
;
437 pr_debug("raid1: behind end write sectors"
439 (unsigned long long) mbio
->bi_sector
,
440 (unsigned long long) mbio
->bi_sector
+
441 (mbio
->bi_size
>> 9) - 1);
442 call_bio_endio(r1_bio
);
446 if (r1_bio
->bios
[mirror
] == NULL
)
447 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
,
451 * Let's see if all mirrored write operations have finished
454 r1_bio_write_done(r1_bio
);
462 * This routine returns the disk from which the requested read should
463 * be done. There is a per-array 'next expected sequential IO' sector
464 * number - if this matches on the next IO then we use the last disk.
465 * There is also a per-disk 'last know head position' sector that is
466 * maintained from IRQ contexts, both the normal and the resync IO
467 * completion handlers update this position correctly. If there is no
468 * perfect sequential match then we pick the disk whose head is closest.
470 * If there are 2 mirrors in the same 2 devices, performance degrades
471 * because position is mirror, not device based.
473 * The rdev for the device selected will have nr_pending incremented.
475 static int read_balance(struct r1conf
*conf
, struct r1bio
*r1_bio
, int *max_sectors
)
477 const sector_t this_sector
= r1_bio
->sector
;
479 int best_good_sectors
;
484 struct md_rdev
*rdev
;
489 * Check if we can balance. We can balance on the whole
490 * device if no resync is going on, or below the resync window.
491 * We take the first readable disk when above the resync window.
494 sectors
= r1_bio
->sectors
;
496 best_dist
= MaxSector
;
497 best_good_sectors
= 0;
499 if (conf
->mddev
->recovery_cp
< MaxSector
&&
500 (this_sector
+ sectors
>= conf
->next_resync
)) {
505 start_disk
= conf
->last_used
;
508 for (i
= 0 ; i
< conf
->raid_disks
; i
++) {
513 int disk
= start_disk
+ i
;
514 if (disk
>= conf
->raid_disks
)
515 disk
-= conf
->raid_disks
;
517 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
518 if (r1_bio
->bios
[disk
] == IO_BLOCKED
520 || test_bit(Faulty
, &rdev
->flags
))
522 if (!test_bit(In_sync
, &rdev
->flags
) &&
523 rdev
->recovery_offset
< this_sector
+ sectors
)
525 if (test_bit(WriteMostly
, &rdev
->flags
)) {
526 /* Don't balance among write-mostly, just
527 * use the first as a last resort */
529 if (is_badblock(rdev
, this_sector
, sectors
,
530 &first_bad
, &bad_sectors
)) {
531 if (first_bad
< this_sector
)
532 /* Cannot use this */
534 best_good_sectors
= first_bad
- this_sector
;
536 best_good_sectors
= sectors
;
541 /* This is a reasonable device to use. It might
544 if (is_badblock(rdev
, this_sector
, sectors
,
545 &first_bad
, &bad_sectors
)) {
546 if (best_dist
< MaxSector
)
547 /* already have a better device */
549 if (first_bad
<= this_sector
) {
550 /* cannot read here. If this is the 'primary'
551 * device, then we must not read beyond
552 * bad_sectors from another device..
554 bad_sectors
-= (this_sector
- first_bad
);
555 if (choose_first
&& sectors
> bad_sectors
)
556 sectors
= bad_sectors
;
557 if (best_good_sectors
> sectors
)
558 best_good_sectors
= sectors
;
561 sector_t good_sectors
= first_bad
- this_sector
;
562 if (good_sectors
> best_good_sectors
) {
563 best_good_sectors
= good_sectors
;
571 best_good_sectors
= sectors
;
573 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
575 /* Don't change to another disk for sequential reads */
576 || conf
->next_seq_sect
== this_sector
578 /* If device is idle, use it */
579 || atomic_read(&rdev
->nr_pending
) == 0) {
583 if (dist
< best_dist
) {
589 if (best_disk
>= 0) {
590 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
593 atomic_inc(&rdev
->nr_pending
);
594 if (test_bit(Faulty
, &rdev
->flags
)) {
595 /* cannot risk returning a device that failed
596 * before we inc'ed nr_pending
598 rdev_dec_pending(rdev
, conf
->mddev
);
601 sectors
= best_good_sectors
;
602 conf
->next_seq_sect
= this_sector
+ sectors
;
603 conf
->last_used
= best_disk
;
606 *max_sectors
= sectors
;
611 int md_raid1_congested(struct mddev
*mddev
, int bits
)
613 struct r1conf
*conf
= mddev
->private;
616 if ((bits
& (1 << BDI_async_congested
)) &&
617 conf
->pending_count
>= max_queued_requests
)
621 for (i
= 0; i
< mddev
->raid_disks
; i
++) {
622 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
623 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
624 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
628 /* Note the '|| 1' - when read_balance prefers
629 * non-congested targets, it can be removed
631 if ((bits
& (1<<BDI_async_congested
)) || 1)
632 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
634 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
640 EXPORT_SYMBOL_GPL(md_raid1_congested
);
642 static int raid1_congested(void *data
, int bits
)
644 struct mddev
*mddev
= data
;
646 return mddev_congested(mddev
, bits
) ||
647 md_raid1_congested(mddev
, bits
);
650 static void flush_pending_writes(struct r1conf
*conf
)
652 /* Any writes that have been queued but are awaiting
653 * bitmap updates get flushed here.
655 spin_lock_irq(&conf
->device_lock
);
657 if (conf
->pending_bio_list
.head
) {
659 bio
= bio_list_get(&conf
->pending_bio_list
);
660 conf
->pending_count
= 0;
661 spin_unlock_irq(&conf
->device_lock
);
662 /* flush any pending bitmap writes to
663 * disk before proceeding w/ I/O */
664 bitmap_unplug(conf
->mddev
->bitmap
);
665 wake_up(&conf
->wait_barrier
);
667 while (bio
) { /* submit pending writes */
668 struct bio
*next
= bio
->bi_next
;
670 generic_make_request(bio
);
674 spin_unlock_irq(&conf
->device_lock
);
678 * Sometimes we need to suspend IO while we do something else,
679 * either some resync/recovery, or reconfigure the array.
680 * To do this we raise a 'barrier'.
681 * The 'barrier' is a counter that can be raised multiple times
682 * to count how many activities are happening which preclude
684 * We can only raise the barrier if there is no pending IO.
685 * i.e. if nr_pending == 0.
686 * We choose only to raise the barrier if no-one is waiting for the
687 * barrier to go down. This means that as soon as an IO request
688 * is ready, no other operations which require a barrier will start
689 * until the IO request has had a chance.
691 * So: regular IO calls 'wait_barrier'. When that returns there
692 * is no backgroup IO happening, It must arrange to call
693 * allow_barrier when it has finished its IO.
694 * backgroup IO calls must call raise_barrier. Once that returns
695 * there is no normal IO happeing. It must arrange to call
696 * lower_barrier when the particular background IO completes.
698 #define RESYNC_DEPTH 32
700 static void raise_barrier(struct r1conf
*conf
)
702 spin_lock_irq(&conf
->resync_lock
);
704 /* Wait until no block IO is waiting */
705 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
706 conf
->resync_lock
, );
708 /* block any new IO from starting */
711 /* Now wait for all pending IO to complete */
712 wait_event_lock_irq(conf
->wait_barrier
,
713 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
714 conf
->resync_lock
, );
716 spin_unlock_irq(&conf
->resync_lock
);
719 static void lower_barrier(struct r1conf
*conf
)
722 BUG_ON(conf
->barrier
<= 0);
723 spin_lock_irqsave(&conf
->resync_lock
, flags
);
725 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
726 wake_up(&conf
->wait_barrier
);
729 static void wait_barrier(struct r1conf
*conf
)
731 spin_lock_irq(&conf
->resync_lock
);
734 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
740 spin_unlock_irq(&conf
->resync_lock
);
743 static void allow_barrier(struct r1conf
*conf
)
746 spin_lock_irqsave(&conf
->resync_lock
, flags
);
748 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
749 wake_up(&conf
->wait_barrier
);
752 static void freeze_array(struct r1conf
*conf
)
754 /* stop syncio and normal IO and wait for everything to
756 * We increment barrier and nr_waiting, and then
757 * wait until nr_pending match nr_queued+1
758 * This is called in the context of one normal IO request
759 * that has failed. Thus any sync request that might be pending
760 * will be blocked by nr_pending, and we need to wait for
761 * pending IO requests to complete or be queued for re-try.
762 * Thus the number queued (nr_queued) plus this request (1)
763 * must match the number of pending IOs (nr_pending) before
766 spin_lock_irq(&conf
->resync_lock
);
769 wait_event_lock_irq(conf
->wait_barrier
,
770 conf
->nr_pending
== conf
->nr_queued
+1,
772 flush_pending_writes(conf
));
773 spin_unlock_irq(&conf
->resync_lock
);
775 static void unfreeze_array(struct r1conf
*conf
)
777 /* reverse the effect of the freeze */
778 spin_lock_irq(&conf
->resync_lock
);
781 wake_up(&conf
->wait_barrier
);
782 spin_unlock_irq(&conf
->resync_lock
);
786 /* duplicate the data pages for behind I/O
788 static void alloc_behind_pages(struct bio
*bio
, struct r1bio
*r1_bio
)
791 struct bio_vec
*bvec
;
792 struct bio_vec
*bvecs
= kzalloc(bio
->bi_vcnt
* sizeof(struct bio_vec
),
794 if (unlikely(!bvecs
))
797 bio_for_each_segment(bvec
, bio
, i
) {
799 bvecs
[i
].bv_page
= alloc_page(GFP_NOIO
);
800 if (unlikely(!bvecs
[i
].bv_page
))
802 memcpy(kmap(bvecs
[i
].bv_page
) + bvec
->bv_offset
,
803 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
804 kunmap(bvecs
[i
].bv_page
);
805 kunmap(bvec
->bv_page
);
807 r1_bio
->behind_bvecs
= bvecs
;
808 r1_bio
->behind_page_count
= bio
->bi_vcnt
;
809 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
813 for (i
= 0; i
< bio
->bi_vcnt
; i
++)
814 if (bvecs
[i
].bv_page
)
815 put_page(bvecs
[i
].bv_page
);
817 pr_debug("%dB behind alloc failed, doing sync I/O\n", bio
->bi_size
);
820 static void make_request(struct mddev
*mddev
, struct bio
* bio
)
822 struct r1conf
*conf
= mddev
->private;
823 struct mirror_info
*mirror
;
824 struct r1bio
*r1_bio
;
825 struct bio
*read_bio
;
827 struct bitmap
*bitmap
;
829 const int rw
= bio_data_dir(bio
);
830 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
831 const unsigned long do_flush_fua
= (bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
));
832 struct md_rdev
*blocked_rdev
;
839 * Register the new request and wait if the reconstruction
840 * thread has put up a bar for new requests.
841 * Continue immediately if no resync is active currently.
844 md_write_start(mddev
, bio
); /* wait on superblock update early */
846 if (bio_data_dir(bio
) == WRITE
&&
847 bio
->bi_sector
+ bio
->bi_size
/512 > mddev
->suspend_lo
&&
848 bio
->bi_sector
< mddev
->suspend_hi
) {
849 /* As the suspend_* range is controlled by
850 * userspace, we want an interruptible
855 flush_signals(current
);
856 prepare_to_wait(&conf
->wait_barrier
,
857 &w
, TASK_INTERRUPTIBLE
);
858 if (bio
->bi_sector
+ bio
->bi_size
/512 <= mddev
->suspend_lo
||
859 bio
->bi_sector
>= mddev
->suspend_hi
)
863 finish_wait(&conf
->wait_barrier
, &w
);
868 bitmap
= mddev
->bitmap
;
871 * make_request() can abort the operation when READA is being
872 * used and no empty request is available.
875 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
877 r1_bio
->master_bio
= bio
;
878 r1_bio
->sectors
= bio
->bi_size
>> 9;
880 r1_bio
->mddev
= mddev
;
881 r1_bio
->sector
= bio
->bi_sector
;
883 /* We might need to issue multiple reads to different
884 * devices if there are bad blocks around, so we keep
885 * track of the number of reads in bio->bi_phys_segments.
886 * If this is 0, there is only one r1_bio and no locking
887 * will be needed when requests complete. If it is
888 * non-zero, then it is the number of not-completed requests.
890 bio
->bi_phys_segments
= 0;
891 clear_bit(BIO_SEG_VALID
, &bio
->bi_flags
);
895 * read balancing logic:
900 rdisk
= read_balance(conf
, r1_bio
, &max_sectors
);
903 /* couldn't find anywhere to read from */
904 raid_end_bio_io(r1_bio
);
907 mirror
= conf
->mirrors
+ rdisk
;
909 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
911 /* Reading from a write-mostly device must
912 * take care not to over-take any writes
915 wait_event(bitmap
->behind_wait
,
916 atomic_read(&bitmap
->behind_writes
) == 0);
918 r1_bio
->read_disk
= rdisk
;
920 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
921 md_trim_bio(read_bio
, r1_bio
->sector
- bio
->bi_sector
,
924 r1_bio
->bios
[rdisk
] = read_bio
;
926 read_bio
->bi_sector
= r1_bio
->sector
+ mirror
->rdev
->data_offset
;
927 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
928 read_bio
->bi_end_io
= raid1_end_read_request
;
929 read_bio
->bi_rw
= READ
| do_sync
;
930 read_bio
->bi_private
= r1_bio
;
932 if (max_sectors
< r1_bio
->sectors
) {
933 /* could not read all from this device, so we will
934 * need another r1_bio.
937 sectors_handled
= (r1_bio
->sector
+ max_sectors
939 r1_bio
->sectors
= max_sectors
;
940 spin_lock_irq(&conf
->device_lock
);
941 if (bio
->bi_phys_segments
== 0)
942 bio
->bi_phys_segments
= 2;
944 bio
->bi_phys_segments
++;
945 spin_unlock_irq(&conf
->device_lock
);
946 /* Cannot call generic_make_request directly
947 * as that will be queued in __make_request
948 * and subsequent mempool_alloc might block waiting
949 * for it. So hand bio over to raid1d.
951 reschedule_retry(r1_bio
);
953 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
955 r1_bio
->master_bio
= bio
;
956 r1_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
958 r1_bio
->mddev
= mddev
;
959 r1_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
962 generic_make_request(read_bio
);
969 if (conf
->pending_count
>= max_queued_requests
) {
970 md_wakeup_thread(mddev
->thread
);
971 wait_event(conf
->wait_barrier
,
972 conf
->pending_count
< max_queued_requests
);
974 /* first select target devices under rcu_lock and
975 * inc refcount on their rdev. Record them by setting
977 * If there are known/acknowledged bad blocks on any device on
978 * which we have seen a write error, we want to avoid writing those
980 * This potentially requires several writes to write around
981 * the bad blocks. Each set of writes gets it's own r1bio
982 * with a set of bios attached.
984 plugged
= mddev_check_plugged(mddev
);
986 disks
= conf
->raid_disks
;
990 max_sectors
= r1_bio
->sectors
;
991 for (i
= 0; i
< disks
; i
++) {
992 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
993 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
994 atomic_inc(&rdev
->nr_pending
);
998 r1_bio
->bios
[i
] = NULL
;
999 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
)) {
1000 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1004 atomic_inc(&rdev
->nr_pending
);
1005 if (test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1010 is_bad
= is_badblock(rdev
, r1_bio
->sector
,
1012 &first_bad
, &bad_sectors
);
1014 /* mustn't write here until the bad block is
1016 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1017 blocked_rdev
= rdev
;
1020 if (is_bad
&& first_bad
<= r1_bio
->sector
) {
1021 /* Cannot write here at all */
1022 bad_sectors
-= (r1_bio
->sector
- first_bad
);
1023 if (bad_sectors
< max_sectors
)
1024 /* mustn't write more than bad_sectors
1025 * to other devices yet
1027 max_sectors
= bad_sectors
;
1028 rdev_dec_pending(rdev
, mddev
);
1029 /* We don't set R1BIO_Degraded as that
1030 * only applies if the disk is
1031 * missing, so it might be re-added,
1032 * and we want to know to recover this
1034 * In this case the device is here,
1035 * and the fact that this chunk is not
1036 * in-sync is recorded in the bad
1042 int good_sectors
= first_bad
- r1_bio
->sector
;
1043 if (good_sectors
< max_sectors
)
1044 max_sectors
= good_sectors
;
1047 r1_bio
->bios
[i
] = bio
;
1051 if (unlikely(blocked_rdev
)) {
1052 /* Wait for this device to become unblocked */
1055 for (j
= 0; j
< i
; j
++)
1056 if (r1_bio
->bios
[j
])
1057 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
1059 allow_barrier(conf
);
1060 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
1065 if (max_sectors
< r1_bio
->sectors
) {
1066 /* We are splitting this write into multiple parts, so
1067 * we need to prepare for allocating another r1_bio.
1069 r1_bio
->sectors
= max_sectors
;
1070 spin_lock_irq(&conf
->device_lock
);
1071 if (bio
->bi_phys_segments
== 0)
1072 bio
->bi_phys_segments
= 2;
1074 bio
->bi_phys_segments
++;
1075 spin_unlock_irq(&conf
->device_lock
);
1077 sectors_handled
= r1_bio
->sector
+ max_sectors
- bio
->bi_sector
;
1079 atomic_set(&r1_bio
->remaining
, 1);
1080 atomic_set(&r1_bio
->behind_remaining
, 0);
1083 for (i
= 0; i
< disks
; i
++) {
1085 if (!r1_bio
->bios
[i
])
1088 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
1089 md_trim_bio(mbio
, r1_bio
->sector
- bio
->bi_sector
, max_sectors
);
1093 * Not if there are too many, or cannot
1094 * allocate memory, or a reader on WriteMostly
1095 * is waiting for behind writes to flush */
1097 (atomic_read(&bitmap
->behind_writes
)
1098 < mddev
->bitmap_info
.max_write_behind
) &&
1099 !waitqueue_active(&bitmap
->behind_wait
))
1100 alloc_behind_pages(mbio
, r1_bio
);
1102 bitmap_startwrite(bitmap
, r1_bio
->sector
,
1104 test_bit(R1BIO_BehindIO
,
1108 if (r1_bio
->behind_bvecs
) {
1109 struct bio_vec
*bvec
;
1112 /* Yes, I really want the '__' version so that
1113 * we clear any unused pointer in the io_vec, rather
1114 * than leave them unchanged. This is important
1115 * because when we come to free the pages, we won't
1116 * know the original bi_idx, so we just free
1119 __bio_for_each_segment(bvec
, mbio
, j
, 0)
1120 bvec
->bv_page
= r1_bio
->behind_bvecs
[j
].bv_page
;
1121 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
1122 atomic_inc(&r1_bio
->behind_remaining
);
1125 r1_bio
->bios
[i
] = mbio
;
1127 mbio
->bi_sector
= (r1_bio
->sector
+
1128 conf
->mirrors
[i
].rdev
->data_offset
);
1129 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1130 mbio
->bi_end_io
= raid1_end_write_request
;
1131 mbio
->bi_rw
= WRITE
| do_flush_fua
| do_sync
;
1132 mbio
->bi_private
= r1_bio
;
1134 atomic_inc(&r1_bio
->remaining
);
1135 spin_lock_irqsave(&conf
->device_lock
, flags
);
1136 bio_list_add(&conf
->pending_bio_list
, mbio
);
1137 conf
->pending_count
++;
1138 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1140 /* Mustn't call r1_bio_write_done before this next test,
1141 * as it could result in the bio being freed.
1143 if (sectors_handled
< (bio
->bi_size
>> 9)) {
1144 r1_bio_write_done(r1_bio
);
1145 /* We need another r1_bio. It has already been counted
1146 * in bio->bi_phys_segments
1148 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
1149 r1_bio
->master_bio
= bio
;
1150 r1_bio
->sectors
= (bio
->bi_size
>> 9) - sectors_handled
;
1152 r1_bio
->mddev
= mddev
;
1153 r1_bio
->sector
= bio
->bi_sector
+ sectors_handled
;
1157 r1_bio_write_done(r1_bio
);
1159 /* In case raid1d snuck in to freeze_array */
1160 wake_up(&conf
->wait_barrier
);
1162 if (do_sync
|| !bitmap
|| !plugged
)
1163 md_wakeup_thread(mddev
->thread
);
1166 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
1168 struct r1conf
*conf
= mddev
->private;
1171 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
1172 conf
->raid_disks
- mddev
->degraded
);
1174 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1175 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1176 seq_printf(seq
, "%s",
1177 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1180 seq_printf(seq
, "]");
1184 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1186 char b
[BDEVNAME_SIZE
];
1187 struct r1conf
*conf
= mddev
->private;
1190 * If it is not operational, then we have already marked it as dead
1191 * else if it is the last working disks, ignore the error, let the
1192 * next level up know.
1193 * else mark the drive as failed
1195 if (test_bit(In_sync
, &rdev
->flags
)
1196 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1198 * Don't fail the drive, act as though we were just a
1199 * normal single drive.
1200 * However don't try a recovery from this drive as
1201 * it is very likely to fail.
1203 conf
->recovery_disabled
= mddev
->recovery_disabled
;
1206 set_bit(Blocked
, &rdev
->flags
);
1207 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1208 unsigned long flags
;
1209 spin_lock_irqsave(&conf
->device_lock
, flags
);
1211 set_bit(Faulty
, &rdev
->flags
);
1212 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1214 * if recovery is running, make sure it aborts.
1216 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1218 set_bit(Faulty
, &rdev
->flags
);
1219 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1221 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1222 "md/raid1:%s: Operation continuing on %d devices.\n",
1223 mdname(mddev
), bdevname(rdev
->bdev
, b
),
1224 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
1227 static void print_conf(struct r1conf
*conf
)
1231 printk(KERN_DEBUG
"RAID1 conf printout:\n");
1233 printk(KERN_DEBUG
"(!conf)\n");
1236 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1240 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1241 char b
[BDEVNAME_SIZE
];
1242 struct md_rdev
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1244 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1245 i
, !test_bit(In_sync
, &rdev
->flags
),
1246 !test_bit(Faulty
, &rdev
->flags
),
1247 bdevname(rdev
->bdev
,b
));
1252 static void close_sync(struct r1conf
*conf
)
1255 allow_barrier(conf
);
1257 mempool_destroy(conf
->r1buf_pool
);
1258 conf
->r1buf_pool
= NULL
;
1261 static int raid1_spare_active(struct mddev
*mddev
)
1264 struct r1conf
*conf
= mddev
->private;
1266 unsigned long flags
;
1269 * Find all failed disks within the RAID1 configuration
1270 * and mark them readable.
1271 * Called under mddev lock, so rcu protection not needed.
1273 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1274 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1276 && !test_bit(Faulty
, &rdev
->flags
)
1277 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1279 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
1282 spin_lock_irqsave(&conf
->device_lock
, flags
);
1283 mddev
->degraded
-= count
;
1284 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1291 static int raid1_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
1293 struct r1conf
*conf
= mddev
->private;
1296 struct mirror_info
*p
;
1298 int last
= mddev
->raid_disks
- 1;
1300 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
1303 if (rdev
->raid_disk
>= 0)
1304 first
= last
= rdev
->raid_disk
;
1306 for (mirror
= first
; mirror
<= last
; mirror
++)
1307 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1309 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1310 rdev
->data_offset
<< 9);
1311 /* as we don't honour merge_bvec_fn, we must
1312 * never risk violating it, so limit
1313 * ->max_segments to one lying with a single
1314 * page, as a one page request is never in
1317 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1318 blk_queue_max_segments(mddev
->queue
, 1);
1319 blk_queue_segment_boundary(mddev
->queue
,
1320 PAGE_CACHE_SIZE
- 1);
1323 p
->head_position
= 0;
1324 rdev
->raid_disk
= mirror
;
1326 /* As all devices are equivalent, we don't need a full recovery
1327 * if this was recently any drive of the array
1329 if (rdev
->saved_raid_disk
< 0)
1331 rcu_assign_pointer(p
->rdev
, rdev
);
1334 md_integrity_add_rdev(rdev
, mddev
);
1339 static int raid1_remove_disk(struct mddev
*mddev
, int number
)
1341 struct r1conf
*conf
= mddev
->private;
1343 struct md_rdev
*rdev
;
1344 struct mirror_info
*p
= conf
->mirrors
+ number
;
1349 if (test_bit(In_sync
, &rdev
->flags
) ||
1350 atomic_read(&rdev
->nr_pending
)) {
1354 /* Only remove non-faulty devices if recovery
1357 if (!test_bit(Faulty
, &rdev
->flags
) &&
1358 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
1359 mddev
->degraded
< conf
->raid_disks
) {
1365 if (atomic_read(&rdev
->nr_pending
)) {
1366 /* lost the race, try later */
1371 err
= md_integrity_register(mddev
);
1380 static void end_sync_read(struct bio
*bio
, int error
)
1382 struct r1bio
*r1_bio
= bio
->bi_private
;
1384 update_head_pos(r1_bio
->read_disk
, r1_bio
);
1387 * we have read a block, now it needs to be re-written,
1388 * or re-read if the read failed.
1389 * We don't do much here, just schedule handling by raid1d
1391 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1392 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1394 if (atomic_dec_and_test(&r1_bio
->remaining
))
1395 reschedule_retry(r1_bio
);
1398 static void end_sync_write(struct bio
*bio
, int error
)
1400 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1401 struct r1bio
*r1_bio
= bio
->bi_private
;
1402 struct mddev
*mddev
= r1_bio
->mddev
;
1403 struct r1conf
*conf
= mddev
->private;
1408 mirror
= find_bio_disk(r1_bio
, bio
);
1411 sector_t sync_blocks
= 0;
1412 sector_t s
= r1_bio
->sector
;
1413 long sectors_to_go
= r1_bio
->sectors
;
1414 /* make sure these bits doesn't get cleared. */
1416 bitmap_end_sync(mddev
->bitmap
, s
,
1419 sectors_to_go
-= sync_blocks
;
1420 } while (sectors_to_go
> 0);
1421 set_bit(WriteErrorSeen
,
1422 &conf
->mirrors
[mirror
].rdev
->flags
);
1423 set_bit(R1BIO_WriteError
, &r1_bio
->state
);
1424 } else if (is_badblock(conf
->mirrors
[mirror
].rdev
,
1427 &first_bad
, &bad_sectors
) &&
1428 !is_badblock(conf
->mirrors
[r1_bio
->read_disk
].rdev
,
1431 &first_bad
, &bad_sectors
)
1433 set_bit(R1BIO_MadeGood
, &r1_bio
->state
);
1435 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1436 int s
= r1_bio
->sectors
;
1437 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
1438 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1439 reschedule_retry(r1_bio
);
1442 md_done_sync(mddev
, s
, uptodate
);
1447 static int r1_sync_page_io(struct md_rdev
*rdev
, sector_t sector
,
1448 int sectors
, struct page
*page
, int rw
)
1450 if (sync_page_io(rdev
, sector
, sectors
<< 9, page
, rw
, false))
1454 set_bit(WriteErrorSeen
, &rdev
->flags
);
1455 /* need to record an error - either for the block or the device */
1456 if (!rdev_set_badblocks(rdev
, sector
, sectors
, 0))
1457 md_error(rdev
->mddev
, rdev
);
1461 static int fix_sync_read_error(struct r1bio
*r1_bio
)
1463 /* Try some synchronous reads of other devices to get
1464 * good data, much like with normal read errors. Only
1465 * read into the pages we already have so we don't
1466 * need to re-issue the read request.
1467 * We don't need to freeze the array, because being in an
1468 * active sync request, there is no normal IO, and
1469 * no overlapping syncs.
1470 * We don't need to check is_badblock() again as we
1471 * made sure that anything with a bad block in range
1472 * will have bi_end_io clear.
1474 struct mddev
*mddev
= r1_bio
->mddev
;
1475 struct r1conf
*conf
= mddev
->private;
1476 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1477 sector_t sect
= r1_bio
->sector
;
1478 int sectors
= r1_bio
->sectors
;
1483 int d
= r1_bio
->read_disk
;
1485 struct md_rdev
*rdev
;
1488 if (s
> (PAGE_SIZE
>>9))
1491 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1492 /* No rcu protection needed here devices
1493 * can only be removed when no resync is
1494 * active, and resync is currently active
1496 rdev
= conf
->mirrors
[d
].rdev
;
1497 if (sync_page_io(rdev
, sect
, s
<<9,
1498 bio
->bi_io_vec
[idx
].bv_page
,
1505 if (d
== conf
->raid_disks
)
1507 } while (!success
&& d
!= r1_bio
->read_disk
);
1510 char b
[BDEVNAME_SIZE
];
1512 /* Cannot read from anywhere, this block is lost.
1513 * Record a bad block on each device. If that doesn't
1514 * work just disable and interrupt the recovery.
1515 * Don't fail devices as that won't really help.
1517 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O read error"
1518 " for block %llu\n",
1520 bdevname(bio
->bi_bdev
, b
),
1521 (unsigned long long)r1_bio
->sector
);
1522 for (d
= 0; d
< conf
->raid_disks
; d
++) {
1523 rdev
= conf
->mirrors
[d
].rdev
;
1524 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
1526 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
1530 conf
->recovery_disabled
=
1531 mddev
->recovery_disabled
;
1532 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1533 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1545 /* write it back and re-read */
1546 while (d
!= r1_bio
->read_disk
) {
1548 d
= conf
->raid_disks
;
1550 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1552 rdev
= conf
->mirrors
[d
].rdev
;
1553 if (r1_sync_page_io(rdev
, sect
, s
,
1554 bio
->bi_io_vec
[idx
].bv_page
,
1556 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
1557 rdev_dec_pending(rdev
, mddev
);
1561 while (d
!= r1_bio
->read_disk
) {
1563 d
= conf
->raid_disks
;
1565 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1567 rdev
= conf
->mirrors
[d
].rdev
;
1568 if (r1_sync_page_io(rdev
, sect
, s
,
1569 bio
->bi_io_vec
[idx
].bv_page
,
1571 atomic_add(s
, &rdev
->corrected_errors
);
1577 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1578 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1582 static int process_checks(struct r1bio
*r1_bio
)
1584 /* We have read all readable devices. If we haven't
1585 * got the block, then there is no hope left.
1586 * If we have, then we want to do a comparison
1587 * and skip the write if everything is the same.
1588 * If any blocks failed to read, then we need to
1589 * attempt an over-write
1591 struct mddev
*mddev
= r1_bio
->mddev
;
1592 struct r1conf
*conf
= mddev
->private;
1596 for (primary
= 0; primary
< conf
->raid_disks
; primary
++)
1597 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
1598 test_bit(BIO_UPTODATE
, &r1_bio
->bios
[primary
]->bi_flags
)) {
1599 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
1600 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
1603 r1_bio
->read_disk
= primary
;
1604 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1606 int vcnt
= r1_bio
->sectors
>> (PAGE_SHIFT
- 9);
1607 struct bio
*pbio
= r1_bio
->bios
[primary
];
1608 struct bio
*sbio
= r1_bio
->bios
[i
];
1611 if (r1_bio
->bios
[i
]->bi_end_io
!= end_sync_read
)
1614 if (test_bit(BIO_UPTODATE
, &sbio
->bi_flags
)) {
1615 for (j
= vcnt
; j
-- ; ) {
1617 p
= pbio
->bi_io_vec
[j
].bv_page
;
1618 s
= sbio
->bi_io_vec
[j
].bv_page
;
1619 if (memcmp(page_address(p
),
1627 mddev
->resync_mismatches
+= r1_bio
->sectors
;
1628 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
1629 && test_bit(BIO_UPTODATE
, &sbio
->bi_flags
))) {
1630 /* No need to write to this device. */
1631 sbio
->bi_end_io
= NULL
;
1632 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
1635 /* fixup the bio for reuse */
1636 sbio
->bi_vcnt
= vcnt
;
1637 sbio
->bi_size
= r1_bio
->sectors
<< 9;
1639 sbio
->bi_phys_segments
= 0;
1640 sbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1641 sbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1642 sbio
->bi_next
= NULL
;
1643 sbio
->bi_sector
= r1_bio
->sector
+
1644 conf
->mirrors
[i
].rdev
->data_offset
;
1645 sbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1646 size
= sbio
->bi_size
;
1647 for (j
= 0; j
< vcnt
; j
++) {
1649 bi
= &sbio
->bi_io_vec
[j
];
1651 if (size
> PAGE_SIZE
)
1652 bi
->bv_len
= PAGE_SIZE
;
1656 memcpy(page_address(bi
->bv_page
),
1657 page_address(pbio
->bi_io_vec
[j
].bv_page
),
1664 static void sync_request_write(struct mddev
*mddev
, struct r1bio
*r1_bio
)
1666 struct r1conf
*conf
= mddev
->private;
1668 int disks
= conf
->raid_disks
;
1669 struct bio
*bio
, *wbio
;
1671 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1673 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
1674 /* ouch - failed to read all of that. */
1675 if (!fix_sync_read_error(r1_bio
))
1678 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
1679 if (process_checks(r1_bio
) < 0)
1684 atomic_set(&r1_bio
->remaining
, 1);
1685 for (i
= 0; i
< disks
; i
++) {
1686 wbio
= r1_bio
->bios
[i
];
1687 if (wbio
->bi_end_io
== NULL
||
1688 (wbio
->bi_end_io
== end_sync_read
&&
1689 (i
== r1_bio
->read_disk
||
1690 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
1693 wbio
->bi_rw
= WRITE
;
1694 wbio
->bi_end_io
= end_sync_write
;
1695 atomic_inc(&r1_bio
->remaining
);
1696 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, wbio
->bi_size
>> 9);
1698 generic_make_request(wbio
);
1701 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1702 /* if we're here, all write(s) have completed, so clean up */
1703 md_done_sync(mddev
, r1_bio
->sectors
, 1);
1709 * This is a kernel thread which:
1711 * 1. Retries failed read operations on working mirrors.
1712 * 2. Updates the raid superblock when problems encounter.
1713 * 3. Performs writes following reads for array synchronising.
1716 static void fix_read_error(struct r1conf
*conf
, int read_disk
,
1717 sector_t sect
, int sectors
)
1719 struct mddev
*mddev
= conf
->mddev
;
1725 struct md_rdev
*rdev
;
1727 if (s
> (PAGE_SIZE
>>9))
1731 /* Note: no rcu protection needed here
1732 * as this is synchronous in the raid1d thread
1733 * which is the thread that might remove
1734 * a device. If raid1d ever becomes multi-threaded....
1739 rdev
= conf
->mirrors
[d
].rdev
;
1741 test_bit(In_sync
, &rdev
->flags
) &&
1742 is_badblock(rdev
, sect
, s
,
1743 &first_bad
, &bad_sectors
) == 0 &&
1744 sync_page_io(rdev
, sect
, s
<<9,
1745 conf
->tmppage
, READ
, false))
1749 if (d
== conf
->raid_disks
)
1752 } while (!success
&& d
!= read_disk
);
1755 /* Cannot read from anywhere - mark it bad */
1756 struct md_rdev
*rdev
= conf
->mirrors
[read_disk
].rdev
;
1757 if (!rdev_set_badblocks(rdev
, sect
, s
, 0))
1758 md_error(mddev
, rdev
);
1761 /* write it back and re-read */
1763 while (d
!= read_disk
) {
1765 d
= conf
->raid_disks
;
1767 rdev
= conf
->mirrors
[d
].rdev
;
1769 test_bit(In_sync
, &rdev
->flags
))
1770 r1_sync_page_io(rdev
, sect
, s
,
1771 conf
->tmppage
, WRITE
);
1774 while (d
!= read_disk
) {
1775 char b
[BDEVNAME_SIZE
];
1777 d
= conf
->raid_disks
;
1779 rdev
= conf
->mirrors
[d
].rdev
;
1781 test_bit(In_sync
, &rdev
->flags
)) {
1782 if (r1_sync_page_io(rdev
, sect
, s
,
1783 conf
->tmppage
, READ
)) {
1784 atomic_add(s
, &rdev
->corrected_errors
);
1786 "md/raid1:%s: read error corrected "
1787 "(%d sectors at %llu on %s)\n",
1789 (unsigned long long)(sect
+
1791 bdevname(rdev
->bdev
, b
));
1800 static void bi_complete(struct bio
*bio
, int error
)
1802 complete((struct completion
*)bio
->bi_private
);
1805 static int submit_bio_wait(int rw
, struct bio
*bio
)
1807 struct completion event
;
1810 init_completion(&event
);
1811 bio
->bi_private
= &event
;
1812 bio
->bi_end_io
= bi_complete
;
1813 submit_bio(rw
, bio
);
1814 wait_for_completion(&event
);
1816 return test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1819 static int narrow_write_error(struct r1bio
*r1_bio
, int i
)
1821 struct mddev
*mddev
= r1_bio
->mddev
;
1822 struct r1conf
*conf
= mddev
->private;
1823 struct md_rdev
*rdev
= conf
->mirrors
[i
].rdev
;
1825 struct bio_vec
*vec
;
1827 /* bio has the data to be written to device 'i' where
1828 * we just recently had a write error.
1829 * We repeatedly clone the bio and trim down to one block,
1830 * then try the write. Where the write fails we record
1832 * It is conceivable that the bio doesn't exactly align with
1833 * blocks. We must handle this somehow.
1835 * We currently own a reference on the rdev.
1841 int sect_to_write
= r1_bio
->sectors
;
1844 if (rdev
->badblocks
.shift
< 0)
1847 block_sectors
= 1 << rdev
->badblocks
.shift
;
1848 sector
= r1_bio
->sector
;
1849 sectors
= ((sector
+ block_sectors
)
1850 & ~(sector_t
)(block_sectors
- 1))
1853 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
1854 vcnt
= r1_bio
->behind_page_count
;
1855 vec
= r1_bio
->behind_bvecs
;
1857 while (vec
[idx
].bv_page
== NULL
)
1860 vcnt
= r1_bio
->master_bio
->bi_vcnt
;
1861 vec
= r1_bio
->master_bio
->bi_io_vec
;
1862 idx
= r1_bio
->master_bio
->bi_idx
;
1864 while (sect_to_write
) {
1866 if (sectors
> sect_to_write
)
1867 sectors
= sect_to_write
;
1868 /* Write at 'sector' for 'sectors'*/
1870 wbio
= bio_alloc_mddev(GFP_NOIO
, vcnt
, mddev
);
1871 memcpy(wbio
->bi_io_vec
, vec
, vcnt
* sizeof(struct bio_vec
));
1872 wbio
->bi_sector
= r1_bio
->sector
;
1873 wbio
->bi_rw
= WRITE
;
1874 wbio
->bi_vcnt
= vcnt
;
1875 wbio
->bi_size
= r1_bio
->sectors
<< 9;
1878 md_trim_bio(wbio
, sector
- r1_bio
->sector
, sectors
);
1879 wbio
->bi_sector
+= rdev
->data_offset
;
1880 wbio
->bi_bdev
= rdev
->bdev
;
1881 if (submit_bio_wait(WRITE
, wbio
) == 0)
1883 ok
= rdev_set_badblocks(rdev
, sector
,
1888 sect_to_write
-= sectors
;
1890 sectors
= block_sectors
;
1895 static void handle_sync_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
1898 int s
= r1_bio
->sectors
;
1899 for (m
= 0; m
< conf
->raid_disks
; m
++) {
1900 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
1901 struct bio
*bio
= r1_bio
->bios
[m
];
1902 if (bio
->bi_end_io
== NULL
)
1904 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
1905 test_bit(R1BIO_MadeGood
, &r1_bio
->state
)) {
1906 rdev_clear_badblocks(rdev
, r1_bio
->sector
, s
);
1908 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
) &&
1909 test_bit(R1BIO_WriteError
, &r1_bio
->state
)) {
1910 if (!rdev_set_badblocks(rdev
, r1_bio
->sector
, s
, 0))
1911 md_error(conf
->mddev
, rdev
);
1915 md_done_sync(conf
->mddev
, s
, 1);
1918 static void handle_write_finished(struct r1conf
*conf
, struct r1bio
*r1_bio
)
1921 for (m
= 0; m
< conf
->raid_disks
; m
++)
1922 if (r1_bio
->bios
[m
] == IO_MADE_GOOD
) {
1923 struct md_rdev
*rdev
= conf
->mirrors
[m
].rdev
;
1924 rdev_clear_badblocks(rdev
,
1927 rdev_dec_pending(rdev
, conf
->mddev
);
1928 } else if (r1_bio
->bios
[m
] != NULL
) {
1929 /* This drive got a write error. We need to
1930 * narrow down and record precise write
1933 if (!narrow_write_error(r1_bio
, m
)) {
1934 md_error(conf
->mddev
,
1935 conf
->mirrors
[m
].rdev
);
1936 /* an I/O failed, we can't clear the bitmap */
1937 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
1939 rdev_dec_pending(conf
->mirrors
[m
].rdev
,
1942 if (test_bit(R1BIO_WriteError
, &r1_bio
->state
))
1943 close_write(r1_bio
);
1944 raid_end_bio_io(r1_bio
);
1947 static void handle_read_error(struct r1conf
*conf
, struct r1bio
*r1_bio
)
1951 struct mddev
*mddev
= conf
->mddev
;
1953 char b
[BDEVNAME_SIZE
];
1954 struct md_rdev
*rdev
;
1956 clear_bit(R1BIO_ReadError
, &r1_bio
->state
);
1957 /* we got a read error. Maybe the drive is bad. Maybe just
1958 * the block and we can fix it.
1959 * We freeze all other IO, and try reading the block from
1960 * other devices. When we find one, we re-write
1961 * and check it that fixes the read error.
1962 * This is all done synchronously while the array is
1965 if (mddev
->ro
== 0) {
1967 fix_read_error(conf
, r1_bio
->read_disk
,
1968 r1_bio
->sector
, r1_bio
->sectors
);
1969 unfreeze_array(conf
);
1971 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
1973 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1974 bdevname(bio
->bi_bdev
, b
);
1976 disk
= read_balance(conf
, r1_bio
, &max_sectors
);
1978 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O"
1979 " read error for block %llu\n",
1980 mdname(mddev
), b
, (unsigned long long)r1_bio
->sector
);
1981 raid_end_bio_io(r1_bio
);
1983 const unsigned long do_sync
1984 = r1_bio
->master_bio
->bi_rw
& REQ_SYNC
;
1986 r1_bio
->bios
[r1_bio
->read_disk
] =
1987 mddev
->ro
? IO_BLOCKED
: NULL
;
1990 r1_bio
->read_disk
= disk
;
1991 bio
= bio_clone_mddev(r1_bio
->master_bio
, GFP_NOIO
, mddev
);
1992 md_trim_bio(bio
, r1_bio
->sector
- bio
->bi_sector
, max_sectors
);
1993 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
1994 rdev
= conf
->mirrors
[disk
].rdev
;
1995 printk_ratelimited(KERN_ERR
1996 "md/raid1:%s: redirecting sector %llu"
1997 " to other mirror: %s\n",
1999 (unsigned long long)r1_bio
->sector
,
2000 bdevname(rdev
->bdev
, b
));
2001 bio
->bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
2002 bio
->bi_bdev
= rdev
->bdev
;
2003 bio
->bi_end_io
= raid1_end_read_request
;
2004 bio
->bi_rw
= READ
| do_sync
;
2005 bio
->bi_private
= r1_bio
;
2006 if (max_sectors
< r1_bio
->sectors
) {
2007 /* Drat - have to split this up more */
2008 struct bio
*mbio
= r1_bio
->master_bio
;
2009 int sectors_handled
= (r1_bio
->sector
+ max_sectors
2011 r1_bio
->sectors
= max_sectors
;
2012 spin_lock_irq(&conf
->device_lock
);
2013 if (mbio
->bi_phys_segments
== 0)
2014 mbio
->bi_phys_segments
= 2;
2016 mbio
->bi_phys_segments
++;
2017 spin_unlock_irq(&conf
->device_lock
);
2018 generic_make_request(bio
);
2021 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
2023 r1_bio
->master_bio
= mbio
;
2024 r1_bio
->sectors
= (mbio
->bi_size
>> 9)
2027 set_bit(R1BIO_ReadError
, &r1_bio
->state
);
2028 r1_bio
->mddev
= mddev
;
2029 r1_bio
->sector
= mbio
->bi_sector
+ sectors_handled
;
2033 generic_make_request(bio
);
2037 static void raid1d(struct mddev
*mddev
)
2039 struct r1bio
*r1_bio
;
2040 unsigned long flags
;
2041 struct r1conf
*conf
= mddev
->private;
2042 struct list_head
*head
= &conf
->retry_list
;
2043 struct blk_plug plug
;
2045 md_check_recovery(mddev
);
2047 blk_start_plug(&plug
);
2050 if (atomic_read(&mddev
->plug_cnt
) == 0)
2051 flush_pending_writes(conf
);
2053 spin_lock_irqsave(&conf
->device_lock
, flags
);
2054 if (list_empty(head
)) {
2055 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2058 r1_bio
= list_entry(head
->prev
, struct r1bio
, retry_list
);
2059 list_del(head
->prev
);
2061 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2063 mddev
= r1_bio
->mddev
;
2064 conf
= mddev
->private;
2065 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
2066 if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2067 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2068 handle_sync_write_finished(conf
, r1_bio
);
2070 sync_request_write(mddev
, r1_bio
);
2071 } else if (test_bit(R1BIO_MadeGood
, &r1_bio
->state
) ||
2072 test_bit(R1BIO_WriteError
, &r1_bio
->state
))
2073 handle_write_finished(conf
, r1_bio
);
2074 else if (test_bit(R1BIO_ReadError
, &r1_bio
->state
))
2075 handle_read_error(conf
, r1_bio
);
2077 /* just a partial read to be scheduled from separate
2080 generic_make_request(r1_bio
->bios
[r1_bio
->read_disk
]);
2083 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
2084 md_check_recovery(mddev
);
2086 blk_finish_plug(&plug
);
2090 static int init_resync(struct r1conf
*conf
)
2094 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
2095 BUG_ON(conf
->r1buf_pool
);
2096 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
2098 if (!conf
->r1buf_pool
)
2100 conf
->next_resync
= 0;
2105 * perform a "sync" on one "block"
2107 * We need to make sure that no normal I/O request - particularly write
2108 * requests - conflict with active sync requests.
2110 * This is achieved by tracking pending requests and a 'barrier' concept
2111 * that can be installed to exclude normal IO requests.
2114 static sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
2116 struct r1conf
*conf
= mddev
->private;
2117 struct r1bio
*r1_bio
;
2119 sector_t max_sector
, nr_sectors
;
2123 int write_targets
= 0, read_targets
= 0;
2124 sector_t sync_blocks
;
2125 int still_degraded
= 0;
2126 int good_sectors
= RESYNC_SECTORS
;
2127 int min_bad
= 0; /* number of sectors that are bad in all devices */
2129 if (!conf
->r1buf_pool
)
2130 if (init_resync(conf
))
2133 max_sector
= mddev
->dev_sectors
;
2134 if (sector_nr
>= max_sector
) {
2135 /* If we aborted, we need to abort the
2136 * sync on the 'current' bitmap chunk (there will
2137 * only be one in raid1 resync.
2138 * We can find the current addess in mddev->curr_resync
2140 if (mddev
->curr_resync
< max_sector
) /* aborted */
2141 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
2143 else /* completed sync */
2146 bitmap_close_sync(mddev
->bitmap
);
2151 if (mddev
->bitmap
== NULL
&&
2152 mddev
->recovery_cp
== MaxSector
&&
2153 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
2154 conf
->fullsync
== 0) {
2156 return max_sector
- sector_nr
;
2158 /* before building a request, check if we can skip these blocks..
2159 * This call the bitmap_start_sync doesn't actually record anything
2161 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
2162 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2163 /* We can skip this block, and probably several more */
2168 * If there is non-resync activity waiting for a turn,
2169 * and resync is going fast enough,
2170 * then let it though before starting on this new sync request.
2172 if (!go_faster
&& conf
->nr_waiting
)
2173 msleep_interruptible(1000);
2175 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
2176 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
2177 raise_barrier(conf
);
2179 conf
->next_resync
= sector_nr
;
2183 * If we get a correctably read error during resync or recovery,
2184 * we might want to read from a different device. So we
2185 * flag all drives that could conceivably be read from for READ,
2186 * and any others (which will be non-In_sync devices) for WRITE.
2187 * If a read fails, we try reading from something else for which READ
2191 r1_bio
->mddev
= mddev
;
2192 r1_bio
->sector
= sector_nr
;
2194 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
2196 for (i
=0; i
< conf
->raid_disks
; i
++) {
2197 struct md_rdev
*rdev
;
2198 bio
= r1_bio
->bios
[i
];
2200 /* take from bio_init */
2201 bio
->bi_next
= NULL
;
2202 bio
->bi_flags
&= ~(BIO_POOL_MASK
-1);
2203 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
2207 bio
->bi_phys_segments
= 0;
2209 bio
->bi_end_io
= NULL
;
2210 bio
->bi_private
= NULL
;
2212 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
2214 test_bit(Faulty
, &rdev
->flags
)) {
2216 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
2218 bio
->bi_end_io
= end_sync_write
;
2221 /* may need to read from here */
2222 sector_t first_bad
= MaxSector
;
2225 if (is_badblock(rdev
, sector_nr
, good_sectors
,
2226 &first_bad
, &bad_sectors
)) {
2227 if (first_bad
> sector_nr
)
2228 good_sectors
= first_bad
- sector_nr
;
2230 bad_sectors
-= (sector_nr
- first_bad
);
2232 min_bad
> bad_sectors
)
2233 min_bad
= bad_sectors
;
2236 if (sector_nr
< first_bad
) {
2237 if (test_bit(WriteMostly
, &rdev
->flags
)) {
2245 bio
->bi_end_io
= end_sync_read
;
2249 if (bio
->bi_end_io
) {
2250 atomic_inc(&rdev
->nr_pending
);
2251 bio
->bi_sector
= sector_nr
+ rdev
->data_offset
;
2252 bio
->bi_bdev
= rdev
->bdev
;
2253 bio
->bi_private
= r1_bio
;
2259 r1_bio
->read_disk
= disk
;
2261 if (read_targets
== 0 && min_bad
> 0) {
2262 /* These sectors are bad on all InSync devices, so we
2263 * need to mark them bad on all write targets
2266 for (i
= 0 ; i
< conf
->raid_disks
; i
++)
2267 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_write
) {
2268 struct md_rdev
*rdev
=
2269 rcu_dereference(conf
->mirrors
[i
].rdev
);
2270 ok
= rdev_set_badblocks(rdev
, sector_nr
,
2274 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
2279 /* Cannot record the badblocks, so need to
2281 * If there are multiple read targets, could just
2282 * fail the really bad ones ???
2284 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2285 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2291 if (min_bad
> 0 && min_bad
< good_sectors
) {
2292 /* only resync enough to reach the next bad->good
2294 good_sectors
= min_bad
;
2297 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
2298 /* extra read targets are also write targets */
2299 write_targets
+= read_targets
-1;
2301 if (write_targets
== 0 || read_targets
== 0) {
2302 /* There is nowhere to write, so all non-sync
2303 * drives must be failed - so we are finished
2305 sector_t rv
= max_sector
- sector_nr
;
2311 if (max_sector
> mddev
->resync_max
)
2312 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
2313 if (max_sector
> sector_nr
+ good_sectors
)
2314 max_sector
= sector_nr
+ good_sectors
;
2319 int len
= PAGE_SIZE
;
2320 if (sector_nr
+ (len
>>9) > max_sector
)
2321 len
= (max_sector
- sector_nr
) << 9;
2324 if (sync_blocks
== 0) {
2325 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
2326 &sync_blocks
, still_degraded
) &&
2328 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
2330 BUG_ON(sync_blocks
< (PAGE_SIZE
>>9));
2331 if ((len
>> 9) > sync_blocks
)
2332 len
= sync_blocks
<<9;
2335 for (i
=0 ; i
< conf
->raid_disks
; i
++) {
2336 bio
= r1_bio
->bios
[i
];
2337 if (bio
->bi_end_io
) {
2338 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
2339 if (bio_add_page(bio
, page
, len
, 0) == 0) {
2341 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
2344 bio
= r1_bio
->bios
[i
];
2345 if (bio
->bi_end_io
==NULL
)
2347 /* remove last page from this bio */
2349 bio
->bi_size
-= len
;
2350 bio
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
2356 nr_sectors
+= len
>>9;
2357 sector_nr
+= len
>>9;
2358 sync_blocks
-= (len
>>9);
2359 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
2361 r1_bio
->sectors
= nr_sectors
;
2363 /* For a user-requested sync, we read all readable devices and do a
2366 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
2367 atomic_set(&r1_bio
->remaining
, read_targets
);
2368 for (i
=0; i
<conf
->raid_disks
; i
++) {
2369 bio
= r1_bio
->bios
[i
];
2370 if (bio
->bi_end_io
== end_sync_read
) {
2371 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2372 generic_make_request(bio
);
2376 atomic_set(&r1_bio
->remaining
, 1);
2377 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
2378 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
2379 generic_make_request(bio
);
2385 static sector_t
raid1_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
2390 return mddev
->dev_sectors
;
2393 static struct r1conf
*setup_conf(struct mddev
*mddev
)
2395 struct r1conf
*conf
;
2397 struct mirror_info
*disk
;
2398 struct md_rdev
*rdev
;
2401 conf
= kzalloc(sizeof(struct r1conf
), GFP_KERNEL
);
2405 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
2410 conf
->tmppage
= alloc_page(GFP_KERNEL
);
2414 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
2415 if (!conf
->poolinfo
)
2417 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
;
2418 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2421 if (!conf
->r1bio_pool
)
2424 conf
->poolinfo
->mddev
= mddev
;
2426 spin_lock_init(&conf
->device_lock
);
2427 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
2428 int disk_idx
= rdev
->raid_disk
;
2429 if (disk_idx
>= mddev
->raid_disks
2432 disk
= conf
->mirrors
+ disk_idx
;
2436 disk
->head_position
= 0;
2438 conf
->raid_disks
= mddev
->raid_disks
;
2439 conf
->mddev
= mddev
;
2440 INIT_LIST_HEAD(&conf
->retry_list
);
2442 spin_lock_init(&conf
->resync_lock
);
2443 init_waitqueue_head(&conf
->wait_barrier
);
2445 bio_list_init(&conf
->pending_bio_list
);
2446 conf
->pending_count
= 0;
2447 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
2449 conf
->last_used
= -1;
2450 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2452 disk
= conf
->mirrors
+ i
;
2455 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2456 disk
->head_position
= 0;
2459 } else if (conf
->last_used
< 0)
2461 * The first working device is used as a
2462 * starting point to read balancing.
2464 conf
->last_used
= i
;
2468 if (conf
->last_used
< 0) {
2469 printk(KERN_ERR
"md/raid1:%s: no operational mirrors\n",
2474 conf
->thread
= md_register_thread(raid1d
, mddev
, NULL
);
2475 if (!conf
->thread
) {
2477 "md/raid1:%s: couldn't allocate thread\n",
2486 if (conf
->r1bio_pool
)
2487 mempool_destroy(conf
->r1bio_pool
);
2488 kfree(conf
->mirrors
);
2489 safe_put_page(conf
->tmppage
);
2490 kfree(conf
->poolinfo
);
2493 return ERR_PTR(err
);
2496 static int run(struct mddev
*mddev
)
2498 struct r1conf
*conf
;
2500 struct md_rdev
*rdev
;
2502 if (mddev
->level
!= 1) {
2503 printk(KERN_ERR
"md/raid1:%s: raid level not set to mirroring (%d)\n",
2504 mdname(mddev
), mddev
->level
);
2507 if (mddev
->reshape_position
!= MaxSector
) {
2508 printk(KERN_ERR
"md/raid1:%s: reshape_position set but not supported\n",
2513 * copy the already verified devices into our private RAID1
2514 * bookkeeping area. [whatever we allocate in run(),
2515 * should be freed in stop()]
2517 if (mddev
->private == NULL
)
2518 conf
= setup_conf(mddev
);
2520 conf
= mddev
->private;
2523 return PTR_ERR(conf
);
2525 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
2526 if (!mddev
->gendisk
)
2528 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
2529 rdev
->data_offset
<< 9);
2530 /* as we don't honour merge_bvec_fn, we must never risk
2531 * violating it, so limit ->max_segments to 1 lying within
2532 * a single page, as a one page request is never in violation.
2534 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
2535 blk_queue_max_segments(mddev
->queue
, 1);
2536 blk_queue_segment_boundary(mddev
->queue
,
2537 PAGE_CACHE_SIZE
- 1);
2541 mddev
->degraded
= 0;
2542 for (i
=0; i
< conf
->raid_disks
; i
++)
2543 if (conf
->mirrors
[i
].rdev
== NULL
||
2544 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
2545 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2548 if (conf
->raid_disks
- mddev
->degraded
== 1)
2549 mddev
->recovery_cp
= MaxSector
;
2551 if (mddev
->recovery_cp
!= MaxSector
)
2552 printk(KERN_NOTICE
"md/raid1:%s: not clean"
2553 " -- starting background reconstruction\n",
2556 "md/raid1:%s: active with %d out of %d mirrors\n",
2557 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2561 * Ok, everything is just fine now
2563 mddev
->thread
= conf
->thread
;
2564 conf
->thread
= NULL
;
2565 mddev
->private = conf
;
2567 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
2570 mddev
->queue
->backing_dev_info
.congested_fn
= raid1_congested
;
2571 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2573 return md_integrity_register(mddev
);
2576 static int stop(struct mddev
*mddev
)
2578 struct r1conf
*conf
= mddev
->private;
2579 struct bitmap
*bitmap
= mddev
->bitmap
;
2581 /* wait for behind writes to complete */
2582 if (bitmap
&& atomic_read(&bitmap
->behind_writes
) > 0) {
2583 printk(KERN_INFO
"md/raid1:%s: behind writes in progress - waiting to stop.\n",
2585 /* need to kick something here to make sure I/O goes? */
2586 wait_event(bitmap
->behind_wait
,
2587 atomic_read(&bitmap
->behind_writes
) == 0);
2590 raise_barrier(conf
);
2591 lower_barrier(conf
);
2593 md_unregister_thread(&mddev
->thread
);
2594 if (conf
->r1bio_pool
)
2595 mempool_destroy(conf
->r1bio_pool
);
2596 kfree(conf
->mirrors
);
2597 kfree(conf
->poolinfo
);
2599 mddev
->private = NULL
;
2603 static int raid1_resize(struct mddev
*mddev
, sector_t sectors
)
2605 /* no resync is happening, and there is enough space
2606 * on all devices, so we can resize.
2607 * We need to make sure resync covers any new space.
2608 * If the array is shrinking we should possibly wait until
2609 * any io in the removed space completes, but it hardly seems
2612 md_set_array_sectors(mddev
, raid1_size(mddev
, sectors
, 0));
2613 if (mddev
->array_sectors
> raid1_size(mddev
, sectors
, 0))
2615 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
2616 revalidate_disk(mddev
->gendisk
);
2617 if (sectors
> mddev
->dev_sectors
&&
2618 mddev
->recovery_cp
> mddev
->dev_sectors
) {
2619 mddev
->recovery_cp
= mddev
->dev_sectors
;
2620 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2622 mddev
->dev_sectors
= sectors
;
2623 mddev
->resync_max_sectors
= sectors
;
2627 static int raid1_reshape(struct mddev
*mddev
)
2630 * 1/ resize the r1bio_pool
2631 * 2/ resize conf->mirrors
2633 * We allocate a new r1bio_pool if we can.
2634 * Then raise a device barrier and wait until all IO stops.
2635 * Then resize conf->mirrors and swap in the new r1bio pool.
2637 * At the same time, we "pack" the devices so that all the missing
2638 * devices have the higher raid_disk numbers.
2640 mempool_t
*newpool
, *oldpool
;
2641 struct pool_info
*newpoolinfo
;
2642 struct mirror_info
*newmirrors
;
2643 struct r1conf
*conf
= mddev
->private;
2644 int cnt
, raid_disks
;
2645 unsigned long flags
;
2648 /* Cannot change chunk_size, layout, or level */
2649 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
2650 mddev
->layout
!= mddev
->new_layout
||
2651 mddev
->level
!= mddev
->new_level
) {
2652 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
2653 mddev
->new_layout
= mddev
->layout
;
2654 mddev
->new_level
= mddev
->level
;
2658 err
= md_allow_write(mddev
);
2662 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
2664 if (raid_disks
< conf
->raid_disks
) {
2666 for (d
= 0; d
< conf
->raid_disks
; d
++)
2667 if (conf
->mirrors
[d
].rdev
)
2669 if (cnt
> raid_disks
)
2673 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
2676 newpoolinfo
->mddev
= mddev
;
2677 newpoolinfo
->raid_disks
= raid_disks
;
2679 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2680 r1bio_pool_free
, newpoolinfo
);
2685 newmirrors
= kzalloc(sizeof(struct mirror_info
) * raid_disks
, GFP_KERNEL
);
2688 mempool_destroy(newpool
);
2692 raise_barrier(conf
);
2694 /* ok, everything is stopped */
2695 oldpool
= conf
->r1bio_pool
;
2696 conf
->r1bio_pool
= newpool
;
2698 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
2699 struct md_rdev
*rdev
= conf
->mirrors
[d
].rdev
;
2700 if (rdev
&& rdev
->raid_disk
!= d2
) {
2701 sysfs_unlink_rdev(mddev
, rdev
);
2702 rdev
->raid_disk
= d2
;
2703 sysfs_unlink_rdev(mddev
, rdev
);
2704 if (sysfs_link_rdev(mddev
, rdev
))
2706 "md/raid1:%s: cannot register rd%d\n",
2707 mdname(mddev
), rdev
->raid_disk
);
2710 newmirrors
[d2
++].rdev
= rdev
;
2712 kfree(conf
->mirrors
);
2713 conf
->mirrors
= newmirrors
;
2714 kfree(conf
->poolinfo
);
2715 conf
->poolinfo
= newpoolinfo
;
2717 spin_lock_irqsave(&conf
->device_lock
, flags
);
2718 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
2719 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2720 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
2721 mddev
->delta_disks
= 0;
2723 conf
->last_used
= 0; /* just make sure it is in-range */
2724 lower_barrier(conf
);
2726 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2727 md_wakeup_thread(mddev
->thread
);
2729 mempool_destroy(oldpool
);
2733 static void raid1_quiesce(struct mddev
*mddev
, int state
)
2735 struct r1conf
*conf
= mddev
->private;
2738 case 2: /* wake for suspend */
2739 wake_up(&conf
->wait_barrier
);
2742 raise_barrier(conf
);
2745 lower_barrier(conf
);
2750 static void *raid1_takeover(struct mddev
*mddev
)
2752 /* raid1 can take over:
2753 * raid5 with 2 devices, any layout or chunk size
2755 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
2756 struct r1conf
*conf
;
2757 mddev
->new_level
= 1;
2758 mddev
->new_layout
= 0;
2759 mddev
->new_chunk_sectors
= 0;
2760 conf
= setup_conf(mddev
);
2765 return ERR_PTR(-EINVAL
);
2768 static struct md_personality raid1_personality
=
2772 .owner
= THIS_MODULE
,
2773 .make_request
= make_request
,
2777 .error_handler
= error
,
2778 .hot_add_disk
= raid1_add_disk
,
2779 .hot_remove_disk
= raid1_remove_disk
,
2780 .spare_active
= raid1_spare_active
,
2781 .sync_request
= sync_request
,
2782 .resize
= raid1_resize
,
2784 .check_reshape
= raid1_reshape
,
2785 .quiesce
= raid1_quiesce
,
2786 .takeover
= raid1_takeover
,
2789 static int __init
raid_init(void)
2791 return register_md_personality(&raid1_personality
);
2794 static void raid_exit(void)
2796 unregister_md_personality(&raid1_personality
);
2799 module_init(raid_init
);
2800 module_exit(raid_exit
);
2801 MODULE_LICENSE("GPL");
2802 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2803 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2804 MODULE_ALIAS("md-raid1");
2805 MODULE_ALIAS("md-level-1");
2807 module_param(max_queued_requests
, int, S_IRUGO
|S_IWUSR
);