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>
44 #define PRINTK(x...) printk(x)
50 * Number of guaranteed r1bios in case of extreme VM load:
52 #define NR_RAID1_BIOS 256
55 static void allow_barrier(conf_t
*conf
);
56 static void lower_barrier(conf_t
*conf
);
58 static void * r1bio_pool_alloc(gfp_t gfp_flags
, void *data
)
60 struct pool_info
*pi
= data
;
61 int size
= offsetof(r1bio_t
, 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 r1bio_t
*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(conf_t
*conf
, r1bio_t
*r1_bio
)
167 for (i
= 0; i
< conf
->raid_disks
; i
++) {
168 struct bio
**bio
= r1_bio
->bios
+ i
;
169 if (*bio
&& *bio
!= IO_BLOCKED
)
175 static void free_r1bio(r1bio_t
*r1_bio
)
177 conf_t
*conf
= r1_bio
->mddev
->private;
180 * Wake up any possible resync thread that waits for the device
185 put_all_bios(conf
, r1_bio
);
186 mempool_free(r1_bio
, conf
->r1bio_pool
);
189 static void put_buf(r1bio_t
*r1_bio
)
191 conf_t
*conf
= r1_bio
->mddev
->private;
194 for (i
=0; i
<conf
->raid_disks
; i
++) {
195 struct bio
*bio
= r1_bio
->bios
[i
];
197 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
200 mempool_free(r1_bio
, conf
->r1buf_pool
);
205 static void reschedule_retry(r1bio_t
*r1_bio
)
208 mddev_t
*mddev
= r1_bio
->mddev
;
209 conf_t
*conf
= mddev
->private;
211 spin_lock_irqsave(&conf
->device_lock
, flags
);
212 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
214 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
216 wake_up(&conf
->wait_barrier
);
217 md_wakeup_thread(mddev
->thread
);
221 * raid_end_bio_io() is called when we have finished servicing a mirrored
222 * operation and are ready to return a success/failure code to the buffer
225 static void raid_end_bio_io(r1bio_t
*r1_bio
)
227 struct bio
*bio
= r1_bio
->master_bio
;
229 /* if nobody has done the final endio yet, do it now */
230 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
231 PRINTK(KERN_DEBUG
"raid1: sync end %s on sectors %llu-%llu\n",
232 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
233 (unsigned long long) bio
->bi_sector
,
234 (unsigned long long) bio
->bi_sector
+
235 (bio
->bi_size
>> 9) - 1);
238 test_bit(R1BIO_Uptodate
, &r1_bio
->state
) ? 0 : -EIO
);
244 * Update disk head position estimator based on IRQ completion info.
246 static inline void update_head_pos(int disk
, r1bio_t
*r1_bio
)
248 conf_t
*conf
= r1_bio
->mddev
->private;
250 conf
->mirrors
[disk
].head_position
=
251 r1_bio
->sector
+ (r1_bio
->sectors
);
254 static void raid1_end_read_request(struct bio
*bio
, int error
)
256 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
257 r1bio_t
*r1_bio
= bio
->bi_private
;
259 conf_t
*conf
= r1_bio
->mddev
->private;
261 mirror
= r1_bio
->read_disk
;
263 * this branch is our 'one mirror IO has finished' event handler:
265 update_head_pos(mirror
, r1_bio
);
268 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
270 /* If all other devices have failed, we want to return
271 * the error upwards rather than fail the last device.
272 * Here we redefine "uptodate" to mean "Don't want to retry"
275 spin_lock_irqsave(&conf
->device_lock
, flags
);
276 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
277 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
278 !test_bit(Faulty
, &conf
->mirrors
[mirror
].rdev
->flags
)))
280 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
284 raid_end_bio_io(r1_bio
);
289 char b
[BDEVNAME_SIZE
];
290 if (printk_ratelimit())
291 printk(KERN_ERR
"md/raid1:%s: %s: rescheduling sector %llu\n",
293 bdevname(conf
->mirrors
[mirror
].rdev
->bdev
,b
), (unsigned long long)r1_bio
->sector
);
294 reschedule_retry(r1_bio
);
297 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
300 static void r1_bio_write_done(r1bio_t
*r1_bio
, int vcnt
, struct bio_vec
*bv
,
303 if (atomic_dec_and_test(&r1_bio
->remaining
))
305 /* it really is the end of this request */
306 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
307 /* free extra copy of the data pages */
310 safe_put_page(bv
[i
].bv_page
);
312 /* clear the bitmap if all writes complete successfully */
313 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
315 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
317 md_write_end(r1_bio
->mddev
);
318 raid_end_bio_io(r1_bio
);
322 static void raid1_end_write_request(struct bio
*bio
, int error
)
324 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
325 r1bio_t
*r1_bio
= bio
->bi_private
;
326 int mirror
, behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
327 conf_t
*conf
= r1_bio
->mddev
->private;
328 struct bio
*to_put
= NULL
;
331 for (mirror
= 0; mirror
< conf
->raid_disks
; mirror
++)
332 if (r1_bio
->bios
[mirror
] == bio
)
336 * 'one mirror IO has finished' event handler:
338 r1_bio
->bios
[mirror
] = NULL
;
341 md_error(r1_bio
->mddev
, conf
->mirrors
[mirror
].rdev
);
342 /* an I/O failed, we can't clear the bitmap */
343 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
346 * Set R1BIO_Uptodate in our master bio, so that we
347 * will return a good error code for to the higher
348 * levels even if IO on some other mirrored buffer
351 * The 'master' represents the composite IO operation
352 * to user-side. So if something waits for IO, then it
353 * will wait for the 'master' bio.
355 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
357 update_head_pos(mirror
, r1_bio
);
360 if (test_bit(WriteMostly
, &conf
->mirrors
[mirror
].rdev
->flags
))
361 atomic_dec(&r1_bio
->behind_remaining
);
364 * In behind mode, we ACK the master bio once the I/O
365 * has safely reached all non-writemostly
366 * disks. Setting the Returned bit ensures that this
367 * gets done only once -- we don't ever want to return
368 * -EIO here, instead we'll wait
370 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
371 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
372 /* Maybe we can return now */
373 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
374 struct bio
*mbio
= r1_bio
->master_bio
;
375 PRINTK(KERN_DEBUG
"raid1: behind end write sectors %llu-%llu\n",
376 (unsigned long long) mbio
->bi_sector
,
377 (unsigned long long) mbio
->bi_sector
+
378 (mbio
->bi_size
>> 9) - 1);
383 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
386 * Let's see if all mirrored write operations have finished
389 r1_bio_write_done(r1_bio
, bio
->bi_vcnt
, bio
->bi_io_vec
, behind
);
397 * This routine returns the disk from which the requested read should
398 * be done. There is a per-array 'next expected sequential IO' sector
399 * number - if this matches on the next IO then we use the last disk.
400 * There is also a per-disk 'last know head position' sector that is
401 * maintained from IRQ contexts, both the normal and the resync IO
402 * completion handlers update this position correctly. If there is no
403 * perfect sequential match then we pick the disk whose head is closest.
405 * If there are 2 mirrors in the same 2 devices, performance degrades
406 * because position is mirror, not device based.
408 * The rdev for the device selected will have nr_pending incremented.
410 static int read_balance(conf_t
*conf
, r1bio_t
*r1_bio
)
412 const sector_t this_sector
= r1_bio
->sector
;
413 const int sectors
= r1_bio
->sectors
;
423 * Check if we can balance. We can balance on the whole
424 * device if no resync is going on, or below the resync window.
425 * We take the first readable disk when above the resync window.
429 best_dist
= MaxSector
;
430 if (conf
->mddev
->recovery_cp
< MaxSector
&&
431 (this_sector
+ sectors
>= conf
->next_resync
)) {
436 start_disk
= conf
->last_used
;
439 for (i
= 0 ; i
< conf
->raid_disks
; i
++) {
441 int disk
= start_disk
+ i
;
442 if (disk
>= conf
->raid_disks
)
443 disk
-= conf
->raid_disks
;
445 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
446 if (r1_bio
->bios
[disk
] == IO_BLOCKED
448 || test_bit(Faulty
, &rdev
->flags
))
450 if (!test_bit(In_sync
, &rdev
->flags
) &&
451 rdev
->recovery_offset
< this_sector
+ sectors
)
453 if (test_bit(WriteMostly
, &rdev
->flags
)) {
454 /* Don't balance among write-mostly, just
455 * use the first as a last resort */
460 /* This is a reasonable device to use. It might
463 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
465 /* Don't change to another disk for sequential reads */
466 || conf
->next_seq_sect
== this_sector
468 /* If device is idle, use it */
469 || atomic_read(&rdev
->nr_pending
) == 0) {
473 if (dist
< best_dist
) {
479 if (best_disk
>= 0) {
480 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
483 atomic_inc(&rdev
->nr_pending
);
484 if (test_bit(Faulty
, &rdev
->flags
)) {
485 /* cannot risk returning a device that failed
486 * before we inc'ed nr_pending
488 rdev_dec_pending(rdev
, conf
->mddev
);
491 conf
->next_seq_sect
= this_sector
+ sectors
;
492 conf
->last_used
= best_disk
;
499 static int raid1_congested(void *data
, int bits
)
501 mddev_t
*mddev
= data
;
502 conf_t
*conf
= mddev
->private;
505 if (mddev_congested(mddev
, bits
))
509 for (i
= 0; i
< mddev
->raid_disks
; i
++) {
510 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
511 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
512 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
514 /* Note the '|| 1' - when read_balance prefers
515 * non-congested targets, it can be removed
517 if ((bits
& (1<<BDI_async_congested
)) || 1)
518 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
520 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
528 static void flush_pending_writes(conf_t
*conf
)
530 /* Any writes that have been queued but are awaiting
531 * bitmap updates get flushed here.
533 spin_lock_irq(&conf
->device_lock
);
535 if (conf
->pending_bio_list
.head
) {
537 bio
= bio_list_get(&conf
->pending_bio_list
);
538 spin_unlock_irq(&conf
->device_lock
);
539 /* flush any pending bitmap writes to
540 * disk before proceeding w/ I/O */
541 bitmap_unplug(conf
->mddev
->bitmap
);
543 while (bio
) { /* submit pending writes */
544 struct bio
*next
= bio
->bi_next
;
546 generic_make_request(bio
);
550 spin_unlock_irq(&conf
->device_lock
);
554 * Sometimes we need to suspend IO while we do something else,
555 * either some resync/recovery, or reconfigure the array.
556 * To do this we raise a 'barrier'.
557 * The 'barrier' is a counter that can be raised multiple times
558 * to count how many activities are happening which preclude
560 * We can only raise the barrier if there is no pending IO.
561 * i.e. if nr_pending == 0.
562 * We choose only to raise the barrier if no-one is waiting for the
563 * barrier to go down. This means that as soon as an IO request
564 * is ready, no other operations which require a barrier will start
565 * until the IO request has had a chance.
567 * So: regular IO calls 'wait_barrier'. When that returns there
568 * is no backgroup IO happening, It must arrange to call
569 * allow_barrier when it has finished its IO.
570 * backgroup IO calls must call raise_barrier. Once that returns
571 * there is no normal IO happeing. It must arrange to call
572 * lower_barrier when the particular background IO completes.
574 #define RESYNC_DEPTH 32
576 static void raise_barrier(conf_t
*conf
)
578 spin_lock_irq(&conf
->resync_lock
);
580 /* Wait until no block IO is waiting */
581 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
582 conf
->resync_lock
, );
584 /* block any new IO from starting */
587 /* Now wait for all pending IO to complete */
588 wait_event_lock_irq(conf
->wait_barrier
,
589 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
590 conf
->resync_lock
, );
592 spin_unlock_irq(&conf
->resync_lock
);
595 static void lower_barrier(conf_t
*conf
)
598 BUG_ON(conf
->barrier
<= 0);
599 spin_lock_irqsave(&conf
->resync_lock
, flags
);
601 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
602 wake_up(&conf
->wait_barrier
);
605 static void wait_barrier(conf_t
*conf
)
607 spin_lock_irq(&conf
->resync_lock
);
610 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
616 spin_unlock_irq(&conf
->resync_lock
);
619 static void allow_barrier(conf_t
*conf
)
622 spin_lock_irqsave(&conf
->resync_lock
, flags
);
624 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
625 wake_up(&conf
->wait_barrier
);
628 static void freeze_array(conf_t
*conf
)
630 /* stop syncio and normal IO and wait for everything to
632 * We increment barrier and nr_waiting, and then
633 * wait until nr_pending match nr_queued+1
634 * This is called in the context of one normal IO request
635 * that has failed. Thus any sync request that might be pending
636 * will be blocked by nr_pending, and we need to wait for
637 * pending IO requests to complete or be queued for re-try.
638 * Thus the number queued (nr_queued) plus this request (1)
639 * must match the number of pending IOs (nr_pending) before
642 spin_lock_irq(&conf
->resync_lock
);
645 wait_event_lock_irq(conf
->wait_barrier
,
646 conf
->nr_pending
== conf
->nr_queued
+1,
648 flush_pending_writes(conf
));
649 spin_unlock_irq(&conf
->resync_lock
);
651 static void unfreeze_array(conf_t
*conf
)
653 /* reverse the effect of the freeze */
654 spin_lock_irq(&conf
->resync_lock
);
657 wake_up(&conf
->wait_barrier
);
658 spin_unlock_irq(&conf
->resync_lock
);
662 /* duplicate the data pages for behind I/O
663 * We return a list of bio_vec rather than just page pointers
664 * as it makes freeing easier
666 static struct bio_vec
*alloc_behind_pages(struct bio
*bio
)
669 struct bio_vec
*bvec
;
670 struct bio_vec
*pages
= kzalloc(bio
->bi_vcnt
* sizeof(struct bio_vec
),
672 if (unlikely(!pages
))
675 bio_for_each_segment(bvec
, bio
, i
) {
676 pages
[i
].bv_page
= alloc_page(GFP_NOIO
);
677 if (unlikely(!pages
[i
].bv_page
))
679 memcpy(kmap(pages
[i
].bv_page
) + bvec
->bv_offset
,
680 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
681 kunmap(pages
[i
].bv_page
);
682 kunmap(bvec
->bv_page
);
689 for (i
= 0; i
< bio
->bi_vcnt
&& pages
[i
].bv_page
; i
++)
690 put_page(pages
[i
].bv_page
);
692 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio
->bi_size
);
696 static int make_request(mddev_t
*mddev
, struct bio
* bio
)
698 conf_t
*conf
= mddev
->private;
699 mirror_info_t
*mirror
;
701 struct bio
*read_bio
;
702 int i
, targets
= 0, disks
;
703 struct bitmap
*bitmap
;
705 struct bio_vec
*behind_pages
= NULL
;
706 const int rw
= bio_data_dir(bio
);
707 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
708 const unsigned long do_flush_fua
= (bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
));
709 mdk_rdev_t
*blocked_rdev
;
713 * Register the new request and wait if the reconstruction
714 * thread has put up a bar for new requests.
715 * Continue immediately if no resync is active currently.
718 md_write_start(mddev
, bio
); /* wait on superblock update early */
720 if (bio_data_dir(bio
) == WRITE
&&
721 bio
->bi_sector
+ bio
->bi_size
/512 > mddev
->suspend_lo
&&
722 bio
->bi_sector
< mddev
->suspend_hi
) {
723 /* As the suspend_* range is controlled by
724 * userspace, we want an interruptible
729 flush_signals(current
);
730 prepare_to_wait(&conf
->wait_barrier
,
731 &w
, TASK_INTERRUPTIBLE
);
732 if (bio
->bi_sector
+ bio
->bi_size
/512 <= mddev
->suspend_lo
||
733 bio
->bi_sector
>= mddev
->suspend_hi
)
737 finish_wait(&conf
->wait_barrier
, &w
);
742 bitmap
= mddev
->bitmap
;
745 * make_request() can abort the operation when READA is being
746 * used and no empty request is available.
749 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
751 r1_bio
->master_bio
= bio
;
752 r1_bio
->sectors
= bio
->bi_size
>> 9;
754 r1_bio
->mddev
= mddev
;
755 r1_bio
->sector
= bio
->bi_sector
;
759 * read balancing logic:
761 int rdisk
= read_balance(conf
, r1_bio
);
764 /* couldn't find anywhere to read from */
765 raid_end_bio_io(r1_bio
);
768 mirror
= conf
->mirrors
+ rdisk
;
770 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
772 /* Reading from a write-mostly device must
773 * take care not to over-take any writes
776 wait_event(bitmap
->behind_wait
,
777 atomic_read(&bitmap
->behind_writes
) == 0);
779 r1_bio
->read_disk
= rdisk
;
781 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
783 r1_bio
->bios
[rdisk
] = read_bio
;
785 read_bio
->bi_sector
= r1_bio
->sector
+ mirror
->rdev
->data_offset
;
786 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
787 read_bio
->bi_end_io
= raid1_end_read_request
;
788 read_bio
->bi_rw
= READ
| do_sync
;
789 read_bio
->bi_private
= r1_bio
;
791 generic_make_request(read_bio
);
798 /* first select target devices under spinlock and
799 * inc refcount on their rdev. Record them by setting
802 plugged
= mddev_check_plugged(mddev
);
804 disks
= conf
->raid_disks
;
808 for (i
= 0; i
< disks
; i
++) {
809 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
810 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
811 atomic_inc(&rdev
->nr_pending
);
815 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
816 atomic_inc(&rdev
->nr_pending
);
817 if (test_bit(Faulty
, &rdev
->flags
)) {
818 rdev_dec_pending(rdev
, mddev
);
819 r1_bio
->bios
[i
] = NULL
;
821 r1_bio
->bios
[i
] = bio
;
825 r1_bio
->bios
[i
] = NULL
;
829 if (unlikely(blocked_rdev
)) {
830 /* Wait for this device to become unblocked */
833 for (j
= 0; j
< i
; j
++)
835 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
838 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
843 BUG_ON(targets
== 0); /* we never fail the last device */
845 if (targets
< conf
->raid_disks
) {
846 /* array is degraded, we will not clear the bitmap
847 * on I/O completion (see raid1_end_write_request) */
848 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
852 * Not if there are too many, or cannot allocate memory,
853 * or a reader on WriteMostly is waiting for behind writes
856 (atomic_read(&bitmap
->behind_writes
)
857 < mddev
->bitmap_info
.max_write_behind
) &&
858 !waitqueue_active(&bitmap
->behind_wait
) &&
859 (behind_pages
= alloc_behind_pages(bio
)) != NULL
)
860 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
862 atomic_set(&r1_bio
->remaining
, 1);
863 atomic_set(&r1_bio
->behind_remaining
, 0);
865 bitmap_startwrite(bitmap
, bio
->bi_sector
, r1_bio
->sectors
,
866 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
867 for (i
= 0; i
< disks
; i
++) {
869 if (!r1_bio
->bios
[i
])
872 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
873 r1_bio
->bios
[i
] = mbio
;
875 mbio
->bi_sector
= r1_bio
->sector
+ conf
->mirrors
[i
].rdev
->data_offset
;
876 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
877 mbio
->bi_end_io
= raid1_end_write_request
;
878 mbio
->bi_rw
= WRITE
| do_flush_fua
| do_sync
;
879 mbio
->bi_private
= r1_bio
;
882 struct bio_vec
*bvec
;
885 /* Yes, I really want the '__' version so that
886 * we clear any unused pointer in the io_vec, rather
887 * than leave them unchanged. This is important
888 * because when we come to free the pages, we won't
889 * know the original bi_idx, so we just free
892 __bio_for_each_segment(bvec
, mbio
, j
, 0)
893 bvec
->bv_page
= behind_pages
[j
].bv_page
;
894 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
895 atomic_inc(&r1_bio
->behind_remaining
);
898 atomic_inc(&r1_bio
->remaining
);
899 spin_lock_irqsave(&conf
->device_lock
, flags
);
900 bio_list_add(&conf
->pending_bio_list
, mbio
);
901 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
903 r1_bio_write_done(r1_bio
, bio
->bi_vcnt
, behind_pages
, behind_pages
!= NULL
);
904 kfree(behind_pages
); /* the behind pages are attached to the bios now */
906 /* In case raid1d snuck in to freeze_array */
907 wake_up(&conf
->wait_barrier
);
909 if (do_sync
|| !bitmap
|| !plugged
)
910 md_wakeup_thread(mddev
->thread
);
915 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
917 conf_t
*conf
= mddev
->private;
920 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
921 conf
->raid_disks
- mddev
->degraded
);
923 for (i
= 0; i
< conf
->raid_disks
; i
++) {
924 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
925 seq_printf(seq
, "%s",
926 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
929 seq_printf(seq
, "]");
933 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
935 char b
[BDEVNAME_SIZE
];
936 conf_t
*conf
= mddev
->private;
939 * If it is not operational, then we have already marked it as dead
940 * else if it is the last working disks, ignore the error, let the
941 * next level up know.
942 * else mark the drive as failed
944 if (test_bit(In_sync
, &rdev
->flags
)
945 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
947 * Don't fail the drive, act as though we were just a
948 * normal single drive.
949 * However don't try a recovery from this drive as
950 * it is very likely to fail.
952 mddev
->recovery_disabled
= 1;
955 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
957 spin_lock_irqsave(&conf
->device_lock
, flags
);
959 set_bit(Faulty
, &rdev
->flags
);
960 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
962 * if recovery is running, make sure it aborts.
964 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
966 set_bit(Faulty
, &rdev
->flags
);
967 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
969 "md/raid1:%s: Disk failure on %s, disabling device.\n"
970 "md/raid1:%s: Operation continuing on %d devices.\n",
971 mdname(mddev
), bdevname(rdev
->bdev
, b
),
972 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
975 static void print_conf(conf_t
*conf
)
979 printk(KERN_DEBUG
"RAID1 conf printout:\n");
981 printk(KERN_DEBUG
"(!conf)\n");
984 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
988 for (i
= 0; i
< conf
->raid_disks
; i
++) {
989 char b
[BDEVNAME_SIZE
];
990 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
992 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
993 i
, !test_bit(In_sync
, &rdev
->flags
),
994 !test_bit(Faulty
, &rdev
->flags
),
995 bdevname(rdev
->bdev
,b
));
1000 static void close_sync(conf_t
*conf
)
1003 allow_barrier(conf
);
1005 mempool_destroy(conf
->r1buf_pool
);
1006 conf
->r1buf_pool
= NULL
;
1009 static int raid1_spare_active(mddev_t
*mddev
)
1012 conf_t
*conf
= mddev
->private;
1014 unsigned long flags
;
1017 * Find all failed disks within the RAID1 configuration
1018 * and mark them readable.
1019 * Called under mddev lock, so rcu protection not needed.
1021 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1022 mdk_rdev_t
*rdev
= conf
->mirrors
[i
].rdev
;
1024 && !test_bit(Faulty
, &rdev
->flags
)
1025 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1027 sysfs_notify_dirent(rdev
->sysfs_state
);
1030 spin_lock_irqsave(&conf
->device_lock
, flags
);
1031 mddev
->degraded
-= count
;
1032 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1039 static int raid1_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1041 conf_t
*conf
= mddev
->private;
1046 int last
= mddev
->raid_disks
- 1;
1048 if (rdev
->raid_disk
>= 0)
1049 first
= last
= rdev
->raid_disk
;
1051 for (mirror
= first
; mirror
<= last
; mirror
++)
1052 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1054 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1055 rdev
->data_offset
<< 9);
1056 /* as we don't honour merge_bvec_fn, we must
1057 * never risk violating it, so limit
1058 * ->max_segments to one lying with a single
1059 * page, as a one page request is never in
1062 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1063 blk_queue_max_segments(mddev
->queue
, 1);
1064 blk_queue_segment_boundary(mddev
->queue
,
1065 PAGE_CACHE_SIZE
- 1);
1068 p
->head_position
= 0;
1069 rdev
->raid_disk
= mirror
;
1071 /* As all devices are equivalent, we don't need a full recovery
1072 * if this was recently any drive of the array
1074 if (rdev
->saved_raid_disk
< 0)
1076 rcu_assign_pointer(p
->rdev
, rdev
);
1079 md_integrity_add_rdev(rdev
, mddev
);
1084 static int raid1_remove_disk(mddev_t
*mddev
, int number
)
1086 conf_t
*conf
= mddev
->private;
1089 mirror_info_t
*p
= conf
->mirrors
+ number
;
1094 if (test_bit(In_sync
, &rdev
->flags
) ||
1095 atomic_read(&rdev
->nr_pending
)) {
1099 /* Only remove non-faulty devices if recovery
1102 if (!test_bit(Faulty
, &rdev
->flags
) &&
1103 !mddev
->recovery_disabled
&&
1104 mddev
->degraded
< conf
->raid_disks
) {
1110 if (atomic_read(&rdev
->nr_pending
)) {
1111 /* lost the race, try later */
1116 err
= md_integrity_register(mddev
);
1125 static void end_sync_read(struct bio
*bio
, int error
)
1127 r1bio_t
*r1_bio
= bio
->bi_private
;
1130 for (i
=r1_bio
->mddev
->raid_disks
; i
--; )
1131 if (r1_bio
->bios
[i
] == bio
)
1134 update_head_pos(i
, r1_bio
);
1136 * we have read a block, now it needs to be re-written,
1137 * or re-read if the read failed.
1138 * We don't do much here, just schedule handling by raid1d
1140 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1141 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1143 if (atomic_dec_and_test(&r1_bio
->remaining
))
1144 reschedule_retry(r1_bio
);
1147 static void end_sync_write(struct bio
*bio
, int error
)
1149 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1150 r1bio_t
*r1_bio
= bio
->bi_private
;
1151 mddev_t
*mddev
= r1_bio
->mddev
;
1152 conf_t
*conf
= mddev
->private;
1156 for (i
= 0; i
< conf
->raid_disks
; i
++)
1157 if (r1_bio
->bios
[i
] == bio
) {
1162 sector_t sync_blocks
= 0;
1163 sector_t s
= r1_bio
->sector
;
1164 long sectors_to_go
= r1_bio
->sectors
;
1165 /* make sure these bits doesn't get cleared. */
1167 bitmap_end_sync(mddev
->bitmap
, s
,
1170 sectors_to_go
-= sync_blocks
;
1171 } while (sectors_to_go
> 0);
1172 md_error(mddev
, conf
->mirrors
[mirror
].rdev
);
1175 update_head_pos(mirror
, r1_bio
);
1177 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1178 sector_t s
= r1_bio
->sectors
;
1180 md_done_sync(mddev
, s
, uptodate
);
1184 static int fix_sync_read_error(r1bio_t
*r1_bio
)
1186 /* Try some synchronous reads of other devices to get
1187 * good data, much like with normal read errors. Only
1188 * read into the pages we already have so we don't
1189 * need to re-issue the read request.
1190 * We don't need to freeze the array, because being in an
1191 * active sync request, there is no normal IO, and
1192 * no overlapping syncs.
1194 mddev_t
*mddev
= r1_bio
->mddev
;
1195 conf_t
*conf
= mddev
->private;
1196 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1197 sector_t sect
= r1_bio
->sector
;
1198 int sectors
= r1_bio
->sectors
;
1203 int d
= r1_bio
->read_disk
;
1208 if (s
> (PAGE_SIZE
>>9))
1211 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1212 /* No rcu protection needed here devices
1213 * can only be removed when no resync is
1214 * active, and resync is currently active
1216 rdev
= conf
->mirrors
[d
].rdev
;
1217 if (sync_page_io(rdev
,
1220 bio
->bi_io_vec
[idx
].bv_page
,
1227 if (d
== conf
->raid_disks
)
1229 } while (!success
&& d
!= r1_bio
->read_disk
);
1232 char b
[BDEVNAME_SIZE
];
1233 /* Cannot read from anywhere, array is toast */
1234 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
1235 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O read error"
1236 " for block %llu\n",
1238 bdevname(bio
->bi_bdev
, b
),
1239 (unsigned long long)r1_bio
->sector
);
1240 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1246 /* write it back and re-read */
1247 while (d
!= r1_bio
->read_disk
) {
1249 d
= conf
->raid_disks
;
1251 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1253 rdev
= conf
->mirrors
[d
].rdev
;
1254 if (sync_page_io(rdev
,
1257 bio
->bi_io_vec
[idx
].bv_page
,
1258 WRITE
, false) == 0) {
1259 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
1260 rdev_dec_pending(rdev
, mddev
);
1261 md_error(mddev
, rdev
);
1263 atomic_add(s
, &rdev
->corrected_errors
);
1266 while (d
!= r1_bio
->read_disk
) {
1268 d
= conf
->raid_disks
;
1270 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1272 rdev
= conf
->mirrors
[d
].rdev
;
1273 if (sync_page_io(rdev
,
1276 bio
->bi_io_vec
[idx
].bv_page
,
1278 md_error(mddev
, rdev
);
1284 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1288 static int process_checks(r1bio_t
*r1_bio
)
1290 /* We have read all readable devices. If we haven't
1291 * got the block, then there is no hope left.
1292 * If we have, then we want to do a comparison
1293 * and skip the write if everything is the same.
1294 * If any blocks failed to read, then we need to
1295 * attempt an over-write
1297 mddev_t
*mddev
= r1_bio
->mddev
;
1298 conf_t
*conf
= mddev
->private;
1302 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
1303 for (i
=0; i
< conf
->raid_disks
; i
++)
1304 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_read
)
1305 md_error(mddev
, conf
->mirrors
[i
].rdev
);
1307 md_done_sync(mddev
, r1_bio
->sectors
, 1);
1311 for (primary
= 0; primary
< conf
->raid_disks
; primary
++)
1312 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
1313 test_bit(BIO_UPTODATE
, &r1_bio
->bios
[primary
]->bi_flags
)) {
1314 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
1315 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
1318 r1_bio
->read_disk
= primary
;
1319 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1321 int vcnt
= r1_bio
->sectors
>> (PAGE_SHIFT
- 9);
1322 struct bio
*pbio
= r1_bio
->bios
[primary
];
1323 struct bio
*sbio
= r1_bio
->bios
[i
];
1326 if (r1_bio
->bios
[i
]->bi_end_io
!= end_sync_read
)
1329 if (test_bit(BIO_UPTODATE
, &sbio
->bi_flags
)) {
1330 for (j
= vcnt
; j
-- ; ) {
1332 p
= pbio
->bi_io_vec
[j
].bv_page
;
1333 s
= sbio
->bi_io_vec
[j
].bv_page
;
1334 if (memcmp(page_address(p
),
1342 mddev
->resync_mismatches
+= r1_bio
->sectors
;
1343 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
1344 && test_bit(BIO_UPTODATE
, &sbio
->bi_flags
))) {
1345 /* No need to write to this device. */
1346 sbio
->bi_end_io
= NULL
;
1347 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
1350 /* fixup the bio for reuse */
1351 sbio
->bi_vcnt
= vcnt
;
1352 sbio
->bi_size
= r1_bio
->sectors
<< 9;
1354 sbio
->bi_phys_segments
= 0;
1355 sbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1356 sbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1357 sbio
->bi_next
= NULL
;
1358 sbio
->bi_sector
= r1_bio
->sector
+
1359 conf
->mirrors
[i
].rdev
->data_offset
;
1360 sbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1361 size
= sbio
->bi_size
;
1362 for (j
= 0; j
< vcnt
; j
++) {
1364 bi
= &sbio
->bi_io_vec
[j
];
1366 if (size
> PAGE_SIZE
)
1367 bi
->bv_len
= PAGE_SIZE
;
1371 memcpy(page_address(bi
->bv_page
),
1372 page_address(pbio
->bi_io_vec
[j
].bv_page
),
1379 static void sync_request_write(mddev_t
*mddev
, r1bio_t
*r1_bio
)
1381 conf_t
*conf
= mddev
->private;
1383 int disks
= conf
->raid_disks
;
1384 struct bio
*bio
, *wbio
;
1386 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1389 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
1390 if (process_checks(r1_bio
) < 0)
1393 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
1394 /* ouch - failed to read all of that. */
1395 if (!fix_sync_read_error(r1_bio
))
1400 atomic_set(&r1_bio
->remaining
, 1);
1401 for (i
= 0; i
< disks
; i
++) {
1402 wbio
= r1_bio
->bios
[i
];
1403 if (wbio
->bi_end_io
== NULL
||
1404 (wbio
->bi_end_io
== end_sync_read
&&
1405 (i
== r1_bio
->read_disk
||
1406 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
1409 wbio
->bi_rw
= WRITE
;
1410 wbio
->bi_end_io
= end_sync_write
;
1411 atomic_inc(&r1_bio
->remaining
);
1412 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, wbio
->bi_size
>> 9);
1414 generic_make_request(wbio
);
1417 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1418 /* if we're here, all write(s) have completed, so clean up */
1419 md_done_sync(mddev
, r1_bio
->sectors
, 1);
1425 * This is a kernel thread which:
1427 * 1. Retries failed read operations on working mirrors.
1428 * 2. Updates the raid superblock when problems encounter.
1429 * 3. Performs writes following reads for array syncronising.
1432 static void fix_read_error(conf_t
*conf
, int read_disk
,
1433 sector_t sect
, int sectors
)
1435 mddev_t
*mddev
= conf
->mddev
;
1443 if (s
> (PAGE_SIZE
>>9))
1447 /* Note: no rcu protection needed here
1448 * as this is synchronous in the raid1d thread
1449 * which is the thread that might remove
1450 * a device. If raid1d ever becomes multi-threaded....
1452 rdev
= conf
->mirrors
[d
].rdev
;
1454 test_bit(In_sync
, &rdev
->flags
) &&
1455 sync_page_io(rdev
, sect
, s
<<9,
1456 conf
->tmppage
, READ
, false))
1460 if (d
== conf
->raid_disks
)
1463 } while (!success
&& d
!= read_disk
);
1466 /* Cannot read from anywhere -- bye bye array */
1467 md_error(mddev
, conf
->mirrors
[read_disk
].rdev
);
1470 /* write it back and re-read */
1472 while (d
!= read_disk
) {
1474 d
= conf
->raid_disks
;
1476 rdev
= conf
->mirrors
[d
].rdev
;
1478 test_bit(In_sync
, &rdev
->flags
)) {
1479 if (sync_page_io(rdev
, sect
, s
<<9,
1480 conf
->tmppage
, WRITE
, false)
1482 /* Well, this device is dead */
1483 md_error(mddev
, rdev
);
1487 while (d
!= read_disk
) {
1488 char b
[BDEVNAME_SIZE
];
1490 d
= conf
->raid_disks
;
1492 rdev
= conf
->mirrors
[d
].rdev
;
1494 test_bit(In_sync
, &rdev
->flags
)) {
1495 if (sync_page_io(rdev
, sect
, s
<<9,
1496 conf
->tmppage
, READ
, false)
1498 /* Well, this device is dead */
1499 md_error(mddev
, rdev
);
1501 atomic_add(s
, &rdev
->corrected_errors
);
1503 "md/raid1:%s: read error corrected "
1504 "(%d sectors at %llu on %s)\n",
1506 (unsigned long long)(sect
+
1508 bdevname(rdev
->bdev
, b
));
1517 static void raid1d(mddev_t
*mddev
)
1521 unsigned long flags
;
1522 conf_t
*conf
= mddev
->private;
1523 struct list_head
*head
= &conf
->retry_list
;
1525 struct blk_plug plug
;
1527 md_check_recovery(mddev
);
1529 blk_start_plug(&plug
);
1531 char b
[BDEVNAME_SIZE
];
1533 if (atomic_read(&mddev
->plug_cnt
) == 0)
1534 flush_pending_writes(conf
);
1536 spin_lock_irqsave(&conf
->device_lock
, flags
);
1537 if (list_empty(head
)) {
1538 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1541 r1_bio
= list_entry(head
->prev
, r1bio_t
, retry_list
);
1542 list_del(head
->prev
);
1544 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1546 mddev
= r1_bio
->mddev
;
1547 conf
= mddev
->private;
1548 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
))
1549 sync_request_write(mddev
, r1_bio
);
1553 /* we got a read error. Maybe the drive is bad. Maybe just
1554 * the block and we can fix it.
1555 * We freeze all other IO, and try reading the block from
1556 * other devices. When we find one, we re-write
1557 * and check it that fixes the read error.
1558 * This is all done synchronously while the array is
1561 if (mddev
->ro
== 0) {
1563 fix_read_error(conf
, r1_bio
->read_disk
,
1566 unfreeze_array(conf
);
1569 conf
->mirrors
[r1_bio
->read_disk
].rdev
);
1571 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1572 if ((disk
=read_balance(conf
, r1_bio
)) == -1) {
1573 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O"
1574 " read error for block %llu\n",
1576 bdevname(bio
->bi_bdev
,b
),
1577 (unsigned long long)r1_bio
->sector
);
1578 raid_end_bio_io(r1_bio
);
1580 const unsigned long do_sync
= r1_bio
->master_bio
->bi_rw
& REQ_SYNC
;
1581 r1_bio
->bios
[r1_bio
->read_disk
] =
1582 mddev
->ro
? IO_BLOCKED
: NULL
;
1583 r1_bio
->read_disk
= disk
;
1585 bio
= bio_clone_mddev(r1_bio
->master_bio
,
1587 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
1588 rdev
= conf
->mirrors
[disk
].rdev
;
1589 if (printk_ratelimit())
1590 printk(KERN_ERR
"md/raid1:%s: redirecting sector %llu to"
1591 " other mirror: %s\n",
1593 (unsigned long long)r1_bio
->sector
,
1594 bdevname(rdev
->bdev
,b
));
1595 bio
->bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
1596 bio
->bi_bdev
= rdev
->bdev
;
1597 bio
->bi_end_io
= raid1_end_read_request
;
1598 bio
->bi_rw
= READ
| do_sync
;
1599 bio
->bi_private
= r1_bio
;
1600 generic_make_request(bio
);
1605 blk_finish_plug(&plug
);
1609 static int init_resync(conf_t
*conf
)
1613 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1614 BUG_ON(conf
->r1buf_pool
);
1615 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
1617 if (!conf
->r1buf_pool
)
1619 conf
->next_resync
= 0;
1624 * perform a "sync" on one "block"
1626 * We need to make sure that no normal I/O request - particularly write
1627 * requests - conflict with active sync requests.
1629 * This is achieved by tracking pending requests and a 'barrier' concept
1630 * that can be installed to exclude normal IO requests.
1633 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
1635 conf_t
*conf
= mddev
->private;
1638 sector_t max_sector
, nr_sectors
;
1642 int write_targets
= 0, read_targets
= 0;
1643 sector_t sync_blocks
;
1644 int still_degraded
= 0;
1646 if (!conf
->r1buf_pool
)
1647 if (init_resync(conf
))
1650 max_sector
= mddev
->dev_sectors
;
1651 if (sector_nr
>= max_sector
) {
1652 /* If we aborted, we need to abort the
1653 * sync on the 'current' bitmap chunk (there will
1654 * only be one in raid1 resync.
1655 * We can find the current addess in mddev->curr_resync
1657 if (mddev
->curr_resync
< max_sector
) /* aborted */
1658 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
1660 else /* completed sync */
1663 bitmap_close_sync(mddev
->bitmap
);
1668 if (mddev
->bitmap
== NULL
&&
1669 mddev
->recovery_cp
== MaxSector
&&
1670 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
1671 conf
->fullsync
== 0) {
1673 return max_sector
- sector_nr
;
1675 /* before building a request, check if we can skip these blocks..
1676 * This call the bitmap_start_sync doesn't actually record anything
1678 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
1679 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1680 /* We can skip this block, and probably several more */
1685 * If there is non-resync activity waiting for a turn,
1686 * and resync is going fast enough,
1687 * then let it though before starting on this new sync request.
1689 if (!go_faster
&& conf
->nr_waiting
)
1690 msleep_interruptible(1000);
1692 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
1693 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
1694 raise_barrier(conf
);
1696 conf
->next_resync
= sector_nr
;
1700 * If we get a correctably read error during resync or recovery,
1701 * we might want to read from a different device. So we
1702 * flag all drives that could conceivably be read from for READ,
1703 * and any others (which will be non-In_sync devices) for WRITE.
1704 * If a read fails, we try reading from something else for which READ
1708 r1_bio
->mddev
= mddev
;
1709 r1_bio
->sector
= sector_nr
;
1711 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
1713 for (i
=0; i
< conf
->raid_disks
; i
++) {
1715 bio
= r1_bio
->bios
[i
];
1717 /* take from bio_init */
1718 bio
->bi_next
= NULL
;
1719 bio
->bi_flags
&= ~(BIO_POOL_MASK
-1);
1720 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
1721 bio
->bi_comp_cpu
= -1;
1725 bio
->bi_phys_segments
= 0;
1727 bio
->bi_end_io
= NULL
;
1728 bio
->bi_private
= NULL
;
1730 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1732 test_bit(Faulty
, &rdev
->flags
)) {
1735 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
1737 bio
->bi_end_io
= end_sync_write
;
1740 /* may need to read from here */
1742 bio
->bi_end_io
= end_sync_read
;
1743 if (test_bit(WriteMostly
, &rdev
->flags
)) {
1752 atomic_inc(&rdev
->nr_pending
);
1753 bio
->bi_sector
= sector_nr
+ rdev
->data_offset
;
1754 bio
->bi_bdev
= rdev
->bdev
;
1755 bio
->bi_private
= r1_bio
;
1760 r1_bio
->read_disk
= disk
;
1762 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
1763 /* extra read targets are also write targets */
1764 write_targets
+= read_targets
-1;
1766 if (write_targets
== 0 || read_targets
== 0) {
1767 /* There is nowhere to write, so all non-sync
1768 * drives must be failed - so we are finished
1770 sector_t rv
= max_sector
- sector_nr
;
1776 if (max_sector
> mddev
->resync_max
)
1777 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
1782 int len
= PAGE_SIZE
;
1783 if (sector_nr
+ (len
>>9) > max_sector
)
1784 len
= (max_sector
- sector_nr
) << 9;
1787 if (sync_blocks
== 0) {
1788 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
1789 &sync_blocks
, still_degraded
) &&
1791 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
1793 BUG_ON(sync_blocks
< (PAGE_SIZE
>>9));
1794 if ((len
>> 9) > sync_blocks
)
1795 len
= sync_blocks
<<9;
1798 for (i
=0 ; i
< conf
->raid_disks
; i
++) {
1799 bio
= r1_bio
->bios
[i
];
1800 if (bio
->bi_end_io
) {
1801 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
1802 if (bio_add_page(bio
, page
, len
, 0) == 0) {
1804 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
1807 bio
= r1_bio
->bios
[i
];
1808 if (bio
->bi_end_io
==NULL
)
1810 /* remove last page from this bio */
1812 bio
->bi_size
-= len
;
1813 bio
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
1819 nr_sectors
+= len
>>9;
1820 sector_nr
+= len
>>9;
1821 sync_blocks
-= (len
>>9);
1822 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
1824 r1_bio
->sectors
= nr_sectors
;
1826 /* For a user-requested sync, we read all readable devices and do a
1829 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1830 atomic_set(&r1_bio
->remaining
, read_targets
);
1831 for (i
=0; i
<conf
->raid_disks
; i
++) {
1832 bio
= r1_bio
->bios
[i
];
1833 if (bio
->bi_end_io
== end_sync_read
) {
1834 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1835 generic_make_request(bio
);
1839 atomic_set(&r1_bio
->remaining
, 1);
1840 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1841 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1842 generic_make_request(bio
);
1848 static sector_t
raid1_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
1853 return mddev
->dev_sectors
;
1856 static conf_t
*setup_conf(mddev_t
*mddev
)
1860 mirror_info_t
*disk
;
1864 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
1868 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
1873 conf
->tmppage
= alloc_page(GFP_KERNEL
);
1877 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
1878 if (!conf
->poolinfo
)
1880 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
;
1881 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
1884 if (!conf
->r1bio_pool
)
1887 conf
->poolinfo
->mddev
= mddev
;
1889 spin_lock_init(&conf
->device_lock
);
1890 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
1891 int disk_idx
= rdev
->raid_disk
;
1892 if (disk_idx
>= mddev
->raid_disks
1895 disk
= conf
->mirrors
+ disk_idx
;
1899 disk
->head_position
= 0;
1901 conf
->raid_disks
= mddev
->raid_disks
;
1902 conf
->mddev
= mddev
;
1903 INIT_LIST_HEAD(&conf
->retry_list
);
1905 spin_lock_init(&conf
->resync_lock
);
1906 init_waitqueue_head(&conf
->wait_barrier
);
1908 bio_list_init(&conf
->pending_bio_list
);
1910 conf
->last_used
= -1;
1911 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1913 disk
= conf
->mirrors
+ i
;
1916 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
1917 disk
->head_position
= 0;
1920 } else if (conf
->last_used
< 0)
1922 * The first working device is used as a
1923 * starting point to read balancing.
1925 conf
->last_used
= i
;
1929 if (conf
->last_used
< 0) {
1930 printk(KERN_ERR
"md/raid1:%s: no operational mirrors\n",
1935 conf
->thread
= md_register_thread(raid1d
, mddev
, NULL
);
1936 if (!conf
->thread
) {
1938 "md/raid1:%s: couldn't allocate thread\n",
1947 if (conf
->r1bio_pool
)
1948 mempool_destroy(conf
->r1bio_pool
);
1949 kfree(conf
->mirrors
);
1950 safe_put_page(conf
->tmppage
);
1951 kfree(conf
->poolinfo
);
1954 return ERR_PTR(err
);
1957 static int run(mddev_t
*mddev
)
1963 if (mddev
->level
!= 1) {
1964 printk(KERN_ERR
"md/raid1:%s: raid level not set to mirroring (%d)\n",
1965 mdname(mddev
), mddev
->level
);
1968 if (mddev
->reshape_position
!= MaxSector
) {
1969 printk(KERN_ERR
"md/raid1:%s: reshape_position set but not supported\n",
1974 * copy the already verified devices into our private RAID1
1975 * bookkeeping area. [whatever we allocate in run(),
1976 * should be freed in stop()]
1978 if (mddev
->private == NULL
)
1979 conf
= setup_conf(mddev
);
1981 conf
= mddev
->private;
1984 return PTR_ERR(conf
);
1986 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
1987 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1988 rdev
->data_offset
<< 9);
1989 /* as we don't honour merge_bvec_fn, we must never risk
1990 * violating it, so limit ->max_segments to 1 lying within
1991 * a single page, as a one page request is never in violation.
1993 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1994 blk_queue_max_segments(mddev
->queue
, 1);
1995 blk_queue_segment_boundary(mddev
->queue
,
1996 PAGE_CACHE_SIZE
- 1);
2000 mddev
->degraded
= 0;
2001 for (i
=0; i
< conf
->raid_disks
; i
++)
2002 if (conf
->mirrors
[i
].rdev
== NULL
||
2003 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
2004 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2007 if (conf
->raid_disks
- mddev
->degraded
== 1)
2008 mddev
->recovery_cp
= MaxSector
;
2010 if (mddev
->recovery_cp
!= MaxSector
)
2011 printk(KERN_NOTICE
"md/raid1:%s: not clean"
2012 " -- starting background reconstruction\n",
2015 "md/raid1:%s: active with %d out of %d mirrors\n",
2016 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2020 * Ok, everything is just fine now
2022 mddev
->thread
= conf
->thread
;
2023 conf
->thread
= NULL
;
2024 mddev
->private = conf
;
2026 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
2028 mddev
->queue
->backing_dev_info
.congested_fn
= raid1_congested
;
2029 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2030 return md_integrity_register(mddev
);
2033 static int stop(mddev_t
*mddev
)
2035 conf_t
*conf
= mddev
->private;
2036 struct bitmap
*bitmap
= mddev
->bitmap
;
2038 /* wait for behind writes to complete */
2039 if (bitmap
&& atomic_read(&bitmap
->behind_writes
) > 0) {
2040 printk(KERN_INFO
"md/raid1:%s: behind writes in progress - waiting to stop.\n",
2042 /* need to kick something here to make sure I/O goes? */
2043 wait_event(bitmap
->behind_wait
,
2044 atomic_read(&bitmap
->behind_writes
) == 0);
2047 raise_barrier(conf
);
2048 lower_barrier(conf
);
2050 md_unregister_thread(mddev
->thread
);
2051 mddev
->thread
= NULL
;
2052 if (conf
->r1bio_pool
)
2053 mempool_destroy(conf
->r1bio_pool
);
2054 kfree(conf
->mirrors
);
2055 kfree(conf
->poolinfo
);
2057 mddev
->private = NULL
;
2061 static int raid1_resize(mddev_t
*mddev
, sector_t sectors
)
2063 /* no resync is happening, and there is enough space
2064 * on all devices, so we can resize.
2065 * We need to make sure resync covers any new space.
2066 * If the array is shrinking we should possibly wait until
2067 * any io in the removed space completes, but it hardly seems
2070 md_set_array_sectors(mddev
, raid1_size(mddev
, sectors
, 0));
2071 if (mddev
->array_sectors
> raid1_size(mddev
, sectors
, 0))
2073 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
2074 revalidate_disk(mddev
->gendisk
);
2075 if (sectors
> mddev
->dev_sectors
&&
2076 mddev
->recovery_cp
== MaxSector
) {
2077 mddev
->recovery_cp
= mddev
->dev_sectors
;
2078 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2080 mddev
->dev_sectors
= sectors
;
2081 mddev
->resync_max_sectors
= sectors
;
2085 static int raid1_reshape(mddev_t
*mddev
)
2088 * 1/ resize the r1bio_pool
2089 * 2/ resize conf->mirrors
2091 * We allocate a new r1bio_pool if we can.
2092 * Then raise a device barrier and wait until all IO stops.
2093 * Then resize conf->mirrors and swap in the new r1bio pool.
2095 * At the same time, we "pack" the devices so that all the missing
2096 * devices have the higher raid_disk numbers.
2098 mempool_t
*newpool
, *oldpool
;
2099 struct pool_info
*newpoolinfo
;
2100 mirror_info_t
*newmirrors
;
2101 conf_t
*conf
= mddev
->private;
2102 int cnt
, raid_disks
;
2103 unsigned long flags
;
2106 /* Cannot change chunk_size, layout, or level */
2107 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
2108 mddev
->layout
!= mddev
->new_layout
||
2109 mddev
->level
!= mddev
->new_level
) {
2110 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
2111 mddev
->new_layout
= mddev
->layout
;
2112 mddev
->new_level
= mddev
->level
;
2116 err
= md_allow_write(mddev
);
2120 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
2122 if (raid_disks
< conf
->raid_disks
) {
2124 for (d
= 0; d
< conf
->raid_disks
; d
++)
2125 if (conf
->mirrors
[d
].rdev
)
2127 if (cnt
> raid_disks
)
2131 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
2134 newpoolinfo
->mddev
= mddev
;
2135 newpoolinfo
->raid_disks
= raid_disks
;
2137 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2138 r1bio_pool_free
, newpoolinfo
);
2143 newmirrors
= kzalloc(sizeof(struct mirror_info
) * raid_disks
, GFP_KERNEL
);
2146 mempool_destroy(newpool
);
2150 raise_barrier(conf
);
2152 /* ok, everything is stopped */
2153 oldpool
= conf
->r1bio_pool
;
2154 conf
->r1bio_pool
= newpool
;
2156 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
2157 mdk_rdev_t
*rdev
= conf
->mirrors
[d
].rdev
;
2158 if (rdev
&& rdev
->raid_disk
!= d2
) {
2160 sprintf(nm
, "rd%d", rdev
->raid_disk
);
2161 sysfs_remove_link(&mddev
->kobj
, nm
);
2162 rdev
->raid_disk
= d2
;
2163 sprintf(nm
, "rd%d", rdev
->raid_disk
);
2164 sysfs_remove_link(&mddev
->kobj
, nm
);
2165 if (sysfs_create_link(&mddev
->kobj
,
2168 "md/raid1:%s: cannot register "
2173 newmirrors
[d2
++].rdev
= rdev
;
2175 kfree(conf
->mirrors
);
2176 conf
->mirrors
= newmirrors
;
2177 kfree(conf
->poolinfo
);
2178 conf
->poolinfo
= newpoolinfo
;
2180 spin_lock_irqsave(&conf
->device_lock
, flags
);
2181 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
2182 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2183 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
2184 mddev
->delta_disks
= 0;
2186 conf
->last_used
= 0; /* just make sure it is in-range */
2187 lower_barrier(conf
);
2189 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2190 md_wakeup_thread(mddev
->thread
);
2192 mempool_destroy(oldpool
);
2196 static void raid1_quiesce(mddev_t
*mddev
, int state
)
2198 conf_t
*conf
= mddev
->private;
2201 case 2: /* wake for suspend */
2202 wake_up(&conf
->wait_barrier
);
2205 raise_barrier(conf
);
2208 lower_barrier(conf
);
2213 static void *raid1_takeover(mddev_t
*mddev
)
2215 /* raid1 can take over:
2216 * raid5 with 2 devices, any layout or chunk size
2218 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
2220 mddev
->new_level
= 1;
2221 mddev
->new_layout
= 0;
2222 mddev
->new_chunk_sectors
= 0;
2223 conf
= setup_conf(mddev
);
2228 return ERR_PTR(-EINVAL
);
2231 static struct mdk_personality raid1_personality
=
2235 .owner
= THIS_MODULE
,
2236 .make_request
= make_request
,
2240 .error_handler
= error
,
2241 .hot_add_disk
= raid1_add_disk
,
2242 .hot_remove_disk
= raid1_remove_disk
,
2243 .spare_active
= raid1_spare_active
,
2244 .sync_request
= sync_request
,
2245 .resize
= raid1_resize
,
2247 .check_reshape
= raid1_reshape
,
2248 .quiesce
= raid1_quiesce
,
2249 .takeover
= raid1_takeover
,
2252 static int __init
raid_init(void)
2254 return register_md_personality(&raid1_personality
);
2257 static void raid_exit(void)
2259 unregister_md_personality(&raid1_personality
);
2262 module_init(raid_init
);
2263 module_exit(raid_exit
);
2264 MODULE_LICENSE("GPL");
2265 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2266 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2267 MODULE_ALIAS("md-raid1");
2268 MODULE_ALIAS("md-level-1");