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
)
302 if (atomic_dec_and_test(&r1_bio
->remaining
))
304 /* it really is the end of this request */
305 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
306 /* free extra copy of the data pages */
307 int i
= r1_bio
->behind_page_count
;
309 safe_put_page(r1_bio
->behind_pages
[i
]);
310 kfree(r1_bio
->behind_pages
);
311 r1_bio
->behind_pages
= NULL
;
313 /* clear the bitmap if all writes complete successfully */
314 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
316 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
317 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
318 md_write_end(r1_bio
->mddev
);
319 raid_end_bio_io(r1_bio
);
323 static void raid1_end_write_request(struct bio
*bio
, int error
)
325 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
326 r1bio_t
*r1_bio
= bio
->bi_private
;
327 int mirror
, behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
328 conf_t
*conf
= r1_bio
->mddev
->private;
329 struct bio
*to_put
= NULL
;
332 for (mirror
= 0; mirror
< conf
->raid_disks
; mirror
++)
333 if (r1_bio
->bios
[mirror
] == bio
)
337 * 'one mirror IO has finished' event handler:
339 r1_bio
->bios
[mirror
] = NULL
;
342 md_error(r1_bio
->mddev
, conf
->mirrors
[mirror
].rdev
);
343 /* an I/O failed, we can't clear the bitmap */
344 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
347 * Set R1BIO_Uptodate in our master bio, so that we
348 * will return a good error code for to the higher
349 * levels even if IO on some other mirrored buffer
352 * The 'master' represents the composite IO operation
353 * to user-side. So if something waits for IO, then it
354 * will wait for the 'master' bio.
356 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
358 update_head_pos(mirror
, r1_bio
);
361 if (test_bit(WriteMostly
, &conf
->mirrors
[mirror
].rdev
->flags
))
362 atomic_dec(&r1_bio
->behind_remaining
);
365 * In behind mode, we ACK the master bio once the I/O
366 * has safely reached all non-writemostly
367 * disks. Setting the Returned bit ensures that this
368 * gets done only once -- we don't ever want to return
369 * -EIO here, instead we'll wait
371 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
372 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
373 /* Maybe we can return now */
374 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
375 struct bio
*mbio
= r1_bio
->master_bio
;
376 PRINTK(KERN_DEBUG
"raid1: behind end write sectors %llu-%llu\n",
377 (unsigned long long) mbio
->bi_sector
,
378 (unsigned long long) mbio
->bi_sector
+
379 (mbio
->bi_size
>> 9) - 1);
384 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
387 * Let's see if all mirrored write operations have finished
390 r1_bio_write_done(r1_bio
);
398 * This routine returns the disk from which the requested read should
399 * be done. There is a per-array 'next expected sequential IO' sector
400 * number - if this matches on the next IO then we use the last disk.
401 * There is also a per-disk 'last know head position' sector that is
402 * maintained from IRQ contexts, both the normal and the resync IO
403 * completion handlers update this position correctly. If there is no
404 * perfect sequential match then we pick the disk whose head is closest.
406 * If there are 2 mirrors in the same 2 devices, performance degrades
407 * because position is mirror, not device based.
409 * The rdev for the device selected will have nr_pending incremented.
411 static int read_balance(conf_t
*conf
, r1bio_t
*r1_bio
)
413 const sector_t this_sector
= r1_bio
->sector
;
414 const int sectors
= r1_bio
->sectors
;
424 * Check if we can balance. We can balance on the whole
425 * device if no resync is going on, or below the resync window.
426 * We take the first readable disk when above the resync window.
430 best_dist
= MaxSector
;
431 if (conf
->mddev
->recovery_cp
< MaxSector
&&
432 (this_sector
+ sectors
>= conf
->next_resync
)) {
437 start_disk
= conf
->last_used
;
440 for (i
= 0 ; i
< conf
->raid_disks
; i
++) {
442 int disk
= start_disk
+ i
;
443 if (disk
>= conf
->raid_disks
)
444 disk
-= conf
->raid_disks
;
446 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
447 if (r1_bio
->bios
[disk
] == IO_BLOCKED
449 || test_bit(Faulty
, &rdev
->flags
))
451 if (!test_bit(In_sync
, &rdev
->flags
) &&
452 rdev
->recovery_offset
< this_sector
+ sectors
)
454 if (test_bit(WriteMostly
, &rdev
->flags
)) {
455 /* Don't balance among write-mostly, just
456 * use the first as a last resort */
461 /* This is a reasonable device to use. It might
464 dist
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
466 /* Don't change to another disk for sequential reads */
467 || conf
->next_seq_sect
== this_sector
469 /* If device is idle, use it */
470 || atomic_read(&rdev
->nr_pending
) == 0) {
474 if (dist
< best_dist
) {
480 if (best_disk
>= 0) {
481 rdev
= rcu_dereference(conf
->mirrors
[best_disk
].rdev
);
484 atomic_inc(&rdev
->nr_pending
);
485 if (test_bit(Faulty
, &rdev
->flags
)) {
486 /* cannot risk returning a device that failed
487 * before we inc'ed nr_pending
489 rdev_dec_pending(rdev
, conf
->mddev
);
492 conf
->next_seq_sect
= this_sector
+ sectors
;
493 conf
->last_used
= best_disk
;
500 int md_raid1_congested(mddev_t
*mddev
, int bits
)
502 conf_t
*conf
= mddev
->private;
506 for (i
= 0; i
< mddev
->raid_disks
; i
++) {
507 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
508 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
509 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
513 /* Note the '|| 1' - when read_balance prefers
514 * non-congested targets, it can be removed
516 if ((bits
& (1<<BDI_async_congested
)) || 1)
517 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
519 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
525 EXPORT_SYMBOL_GPL(md_raid1_congested
);
527 static int raid1_congested(void *data
, int bits
)
529 mddev_t
*mddev
= data
;
531 return mddev_congested(mddev
, bits
) ||
532 md_raid1_congested(mddev
, bits
);
535 static void flush_pending_writes(conf_t
*conf
)
537 /* Any writes that have been queued but are awaiting
538 * bitmap updates get flushed here.
540 spin_lock_irq(&conf
->device_lock
);
542 if (conf
->pending_bio_list
.head
) {
544 bio
= bio_list_get(&conf
->pending_bio_list
);
545 spin_unlock_irq(&conf
->device_lock
);
546 /* flush any pending bitmap writes to
547 * disk before proceeding w/ I/O */
548 bitmap_unplug(conf
->mddev
->bitmap
);
550 while (bio
) { /* submit pending writes */
551 struct bio
*next
= bio
->bi_next
;
553 generic_make_request(bio
);
557 spin_unlock_irq(&conf
->device_lock
);
561 * Sometimes we need to suspend IO while we do something else,
562 * either some resync/recovery, or reconfigure the array.
563 * To do this we raise a 'barrier'.
564 * The 'barrier' is a counter that can be raised multiple times
565 * to count how many activities are happening which preclude
567 * We can only raise the barrier if there is no pending IO.
568 * i.e. if nr_pending == 0.
569 * We choose only to raise the barrier if no-one is waiting for the
570 * barrier to go down. This means that as soon as an IO request
571 * is ready, no other operations which require a barrier will start
572 * until the IO request has had a chance.
574 * So: regular IO calls 'wait_barrier'. When that returns there
575 * is no backgroup IO happening, It must arrange to call
576 * allow_barrier when it has finished its IO.
577 * backgroup IO calls must call raise_barrier. Once that returns
578 * there is no normal IO happeing. It must arrange to call
579 * lower_barrier when the particular background IO completes.
581 #define RESYNC_DEPTH 32
583 static void raise_barrier(conf_t
*conf
)
585 spin_lock_irq(&conf
->resync_lock
);
587 /* Wait until no block IO is waiting */
588 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
589 conf
->resync_lock
, );
591 /* block any new IO from starting */
594 /* Now wait for all pending IO to complete */
595 wait_event_lock_irq(conf
->wait_barrier
,
596 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
597 conf
->resync_lock
, );
599 spin_unlock_irq(&conf
->resync_lock
);
602 static void lower_barrier(conf_t
*conf
)
605 BUG_ON(conf
->barrier
<= 0);
606 spin_lock_irqsave(&conf
->resync_lock
, flags
);
608 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
609 wake_up(&conf
->wait_barrier
);
612 static void wait_barrier(conf_t
*conf
)
614 spin_lock_irq(&conf
->resync_lock
);
617 /* Wait for the barrier to drop.
618 * However if there are already pending
619 * requests (preventing the barrier from
620 * rising completely), and the
621 * pre-process bio queue isn't empty,
622 * then don't wait, as we need to empty
623 * that queue to get the nr_pending
626 wait_event_lock_irq(conf
->wait_barrier
,
630 !bio_list_empty(current
->bio_list
)),
636 spin_unlock_irq(&conf
->resync_lock
);
639 static void allow_barrier(conf_t
*conf
)
642 spin_lock_irqsave(&conf
->resync_lock
, flags
);
644 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
645 wake_up(&conf
->wait_barrier
);
648 static void freeze_array(conf_t
*conf
)
650 /* stop syncio and normal IO and wait for everything to
652 * We increment barrier and nr_waiting, and then
653 * wait until nr_pending match nr_queued+1
654 * This is called in the context of one normal IO request
655 * that has failed. Thus any sync request that might be pending
656 * will be blocked by nr_pending, and we need to wait for
657 * pending IO requests to complete or be queued for re-try.
658 * Thus the number queued (nr_queued) plus this request (1)
659 * must match the number of pending IOs (nr_pending) before
662 spin_lock_irq(&conf
->resync_lock
);
665 wait_event_lock_irq(conf
->wait_barrier
,
666 conf
->nr_pending
== conf
->nr_queued
+1,
668 flush_pending_writes(conf
));
669 spin_unlock_irq(&conf
->resync_lock
);
671 static void unfreeze_array(conf_t
*conf
)
673 /* reverse the effect of the freeze */
674 spin_lock_irq(&conf
->resync_lock
);
677 wake_up(&conf
->wait_barrier
);
678 spin_unlock_irq(&conf
->resync_lock
);
682 /* duplicate the data pages for behind I/O
684 static void alloc_behind_pages(struct bio
*bio
, r1bio_t
*r1_bio
)
687 struct bio_vec
*bvec
;
688 struct page
**pages
= kzalloc(bio
->bi_vcnt
* sizeof(struct page
*),
690 if (unlikely(!pages
))
693 bio_for_each_segment(bvec
, bio
, i
) {
694 pages
[i
] = alloc_page(GFP_NOIO
);
695 if (unlikely(!pages
[i
]))
697 memcpy(kmap(pages
[i
]) + bvec
->bv_offset
,
698 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
700 kunmap(bvec
->bv_page
);
702 r1_bio
->behind_pages
= pages
;
703 r1_bio
->behind_page_count
= bio
->bi_vcnt
;
704 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
708 for (i
= 0; i
< bio
->bi_vcnt
; i
++)
712 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio
->bi_size
);
715 static int make_request(mddev_t
*mddev
, struct bio
* bio
)
717 conf_t
*conf
= mddev
->private;
718 mirror_info_t
*mirror
;
720 struct bio
*read_bio
;
721 int i
, targets
= 0, disks
;
722 struct bitmap
*bitmap
;
724 const int rw
= bio_data_dir(bio
);
725 const unsigned long do_sync
= (bio
->bi_rw
& REQ_SYNC
);
726 const unsigned long do_flush_fua
= (bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
));
727 mdk_rdev_t
*blocked_rdev
;
731 * Register the new request and wait if the reconstruction
732 * thread has put up a bar for new requests.
733 * Continue immediately if no resync is active currently.
736 md_write_start(mddev
, bio
); /* wait on superblock update early */
738 if (bio_data_dir(bio
) == WRITE
&&
739 bio
->bi_sector
+ bio
->bi_size
/512 > mddev
->suspend_lo
&&
740 bio
->bi_sector
< mddev
->suspend_hi
) {
741 /* As the suspend_* range is controlled by
742 * userspace, we want an interruptible
747 flush_signals(current
);
748 prepare_to_wait(&conf
->wait_barrier
,
749 &w
, TASK_INTERRUPTIBLE
);
750 if (bio
->bi_sector
+ bio
->bi_size
/512 <= mddev
->suspend_lo
||
751 bio
->bi_sector
>= mddev
->suspend_hi
)
755 finish_wait(&conf
->wait_barrier
, &w
);
760 bitmap
= mddev
->bitmap
;
763 * make_request() can abort the operation when READA is being
764 * used and no empty request is available.
767 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
769 r1_bio
->master_bio
= bio
;
770 r1_bio
->sectors
= bio
->bi_size
>> 9;
772 r1_bio
->mddev
= mddev
;
773 r1_bio
->sector
= bio
->bi_sector
;
777 * read balancing logic:
779 int rdisk
= read_balance(conf
, r1_bio
);
782 /* couldn't find anywhere to read from */
783 raid_end_bio_io(r1_bio
);
786 mirror
= conf
->mirrors
+ rdisk
;
788 if (test_bit(WriteMostly
, &mirror
->rdev
->flags
) &&
790 /* Reading from a write-mostly device must
791 * take care not to over-take any writes
794 wait_event(bitmap
->behind_wait
,
795 atomic_read(&bitmap
->behind_writes
) == 0);
797 r1_bio
->read_disk
= rdisk
;
799 read_bio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
801 r1_bio
->bios
[rdisk
] = read_bio
;
803 read_bio
->bi_sector
= r1_bio
->sector
+ mirror
->rdev
->data_offset
;
804 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
805 read_bio
->bi_end_io
= raid1_end_read_request
;
806 read_bio
->bi_rw
= READ
| do_sync
;
807 read_bio
->bi_private
= r1_bio
;
809 generic_make_request(read_bio
);
816 /* first select target devices under spinlock and
817 * inc refcount on their rdev. Record them by setting
820 plugged
= mddev_check_plugged(mddev
);
822 disks
= conf
->raid_disks
;
826 for (i
= 0; i
< disks
; i
++) {
827 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
828 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
829 atomic_inc(&rdev
->nr_pending
);
833 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
834 atomic_inc(&rdev
->nr_pending
);
835 if (test_bit(Faulty
, &rdev
->flags
)) {
836 rdev_dec_pending(rdev
, mddev
);
837 r1_bio
->bios
[i
] = NULL
;
839 r1_bio
->bios
[i
] = bio
;
843 r1_bio
->bios
[i
] = NULL
;
847 if (unlikely(blocked_rdev
)) {
848 /* Wait for this device to become unblocked */
851 for (j
= 0; j
< i
; j
++)
853 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
856 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
861 BUG_ON(targets
== 0); /* we never fail the last device */
863 if (targets
< conf
->raid_disks
) {
864 /* array is degraded, we will not clear the bitmap
865 * on I/O completion (see raid1_end_write_request) */
866 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
870 * Not if there are too many, or cannot allocate memory,
871 * or a reader on WriteMostly is waiting for behind writes
874 (atomic_read(&bitmap
->behind_writes
)
875 < mddev
->bitmap_info
.max_write_behind
) &&
876 !waitqueue_active(&bitmap
->behind_wait
))
877 alloc_behind_pages(bio
, r1_bio
);
879 atomic_set(&r1_bio
->remaining
, 1);
880 atomic_set(&r1_bio
->behind_remaining
, 0);
882 bitmap_startwrite(bitmap
, bio
->bi_sector
, r1_bio
->sectors
,
883 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
884 for (i
= 0; i
< disks
; i
++) {
886 if (!r1_bio
->bios
[i
])
889 mbio
= bio_clone_mddev(bio
, GFP_NOIO
, mddev
);
890 r1_bio
->bios
[i
] = mbio
;
892 mbio
->bi_sector
= r1_bio
->sector
+ conf
->mirrors
[i
].rdev
->data_offset
;
893 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
894 mbio
->bi_end_io
= raid1_end_write_request
;
895 mbio
->bi_rw
= WRITE
| do_flush_fua
| do_sync
;
896 mbio
->bi_private
= r1_bio
;
898 if (r1_bio
->behind_pages
) {
899 struct bio_vec
*bvec
;
902 /* Yes, I really want the '__' version so that
903 * we clear any unused pointer in the io_vec, rather
904 * than leave them unchanged. This is important
905 * because when we come to free the pages, we won't
906 * know the original bi_idx, so we just free
909 __bio_for_each_segment(bvec
, mbio
, j
, 0)
910 bvec
->bv_page
= r1_bio
->behind_pages
[j
];
911 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
912 atomic_inc(&r1_bio
->behind_remaining
);
915 atomic_inc(&r1_bio
->remaining
);
916 spin_lock_irqsave(&conf
->device_lock
, flags
);
917 bio_list_add(&conf
->pending_bio_list
, mbio
);
918 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
920 r1_bio_write_done(r1_bio
);
922 /* In case raid1d snuck in to freeze_array */
923 wake_up(&conf
->wait_barrier
);
925 if (do_sync
|| !bitmap
|| !plugged
)
926 md_wakeup_thread(mddev
->thread
);
931 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
933 conf_t
*conf
= mddev
->private;
936 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
937 conf
->raid_disks
- mddev
->degraded
);
939 for (i
= 0; i
< conf
->raid_disks
; i
++) {
940 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
941 seq_printf(seq
, "%s",
942 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
945 seq_printf(seq
, "]");
949 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
951 char b
[BDEVNAME_SIZE
];
952 conf_t
*conf
= mddev
->private;
955 * If it is not operational, then we have already marked it as dead
956 * else if it is the last working disks, ignore the error, let the
957 * next level up know.
958 * else mark the drive as failed
960 if (test_bit(In_sync
, &rdev
->flags
)
961 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
963 * Don't fail the drive, act as though we were just a
964 * normal single drive.
965 * However don't try a recovery from this drive as
966 * it is very likely to fail.
968 mddev
->recovery_disabled
= 1;
971 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
973 spin_lock_irqsave(&conf
->device_lock
, flags
);
975 set_bit(Faulty
, &rdev
->flags
);
976 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
978 * if recovery is running, make sure it aborts.
980 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
982 set_bit(Faulty
, &rdev
->flags
);
983 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
985 "md/raid1:%s: Disk failure on %s, disabling device.\n"
986 "md/raid1:%s: Operation continuing on %d devices.\n",
987 mdname(mddev
), bdevname(rdev
->bdev
, b
),
988 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
991 static void print_conf(conf_t
*conf
)
995 printk(KERN_DEBUG
"RAID1 conf printout:\n");
997 printk(KERN_DEBUG
"(!conf)\n");
1000 printk(KERN_DEBUG
" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1004 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1005 char b
[BDEVNAME_SIZE
];
1006 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1008 printk(KERN_DEBUG
" disk %d, wo:%d, o:%d, dev:%s\n",
1009 i
, !test_bit(In_sync
, &rdev
->flags
),
1010 !test_bit(Faulty
, &rdev
->flags
),
1011 bdevname(rdev
->bdev
,b
));
1016 static void close_sync(conf_t
*conf
)
1019 allow_barrier(conf
);
1021 mempool_destroy(conf
->r1buf_pool
);
1022 conf
->r1buf_pool
= NULL
;
1025 static int raid1_spare_active(mddev_t
*mddev
)
1028 conf_t
*conf
= mddev
->private;
1030 unsigned long flags
;
1033 * Find all failed disks within the RAID1 configuration
1034 * and mark them readable.
1035 * Called under mddev lock, so rcu protection not needed.
1037 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1038 mdk_rdev_t
*rdev
= conf
->mirrors
[i
].rdev
;
1040 && !test_bit(Faulty
, &rdev
->flags
)
1041 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1043 sysfs_notify_dirent(rdev
->sysfs_state
);
1046 spin_lock_irqsave(&conf
->device_lock
, flags
);
1047 mddev
->degraded
-= count
;
1048 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1055 static int raid1_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1057 conf_t
*conf
= mddev
->private;
1062 int last
= mddev
->raid_disks
- 1;
1064 if (rdev
->raid_disk
>= 0)
1065 first
= last
= rdev
->raid_disk
;
1067 for (mirror
= first
; mirror
<= last
; mirror
++)
1068 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1070 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1071 rdev
->data_offset
<< 9);
1072 /* as we don't honour merge_bvec_fn, we must
1073 * never risk violating it, so limit
1074 * ->max_segments to one lying with a single
1075 * page, as a one page request is never in
1078 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
1079 blk_queue_max_segments(mddev
->queue
, 1);
1080 blk_queue_segment_boundary(mddev
->queue
,
1081 PAGE_CACHE_SIZE
- 1);
1084 p
->head_position
= 0;
1085 rdev
->raid_disk
= mirror
;
1087 /* As all devices are equivalent, we don't need a full recovery
1088 * if this was recently any drive of the array
1090 if (rdev
->saved_raid_disk
< 0)
1092 rcu_assign_pointer(p
->rdev
, rdev
);
1095 md_integrity_add_rdev(rdev
, mddev
);
1100 static int raid1_remove_disk(mddev_t
*mddev
, int number
)
1102 conf_t
*conf
= mddev
->private;
1105 mirror_info_t
*p
= conf
->mirrors
+ number
;
1110 if (test_bit(In_sync
, &rdev
->flags
) ||
1111 atomic_read(&rdev
->nr_pending
)) {
1115 /* Only remove non-faulty devices if recovery
1118 if (!test_bit(Faulty
, &rdev
->flags
) &&
1119 !mddev
->recovery_disabled
&&
1120 mddev
->degraded
< conf
->raid_disks
) {
1126 if (atomic_read(&rdev
->nr_pending
)) {
1127 /* lost the race, try later */
1132 err
= md_integrity_register(mddev
);
1141 static void end_sync_read(struct bio
*bio
, int error
)
1143 r1bio_t
*r1_bio
= bio
->bi_private
;
1146 for (i
=r1_bio
->mddev
->raid_disks
; i
--; )
1147 if (r1_bio
->bios
[i
] == bio
)
1150 update_head_pos(i
, r1_bio
);
1152 * we have read a block, now it needs to be re-written,
1153 * or re-read if the read failed.
1154 * We don't do much here, just schedule handling by raid1d
1156 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1157 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1159 if (atomic_dec_and_test(&r1_bio
->remaining
))
1160 reschedule_retry(r1_bio
);
1163 static void end_sync_write(struct bio
*bio
, int error
)
1165 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1166 r1bio_t
*r1_bio
= bio
->bi_private
;
1167 mddev_t
*mddev
= r1_bio
->mddev
;
1168 conf_t
*conf
= mddev
->private;
1172 for (i
= 0; i
< conf
->raid_disks
; i
++)
1173 if (r1_bio
->bios
[i
] == bio
) {
1178 sector_t sync_blocks
= 0;
1179 sector_t s
= r1_bio
->sector
;
1180 long sectors_to_go
= r1_bio
->sectors
;
1181 /* make sure these bits doesn't get cleared. */
1183 bitmap_end_sync(mddev
->bitmap
, s
,
1186 sectors_to_go
-= sync_blocks
;
1187 } while (sectors_to_go
> 0);
1188 md_error(mddev
, conf
->mirrors
[mirror
].rdev
);
1191 update_head_pos(mirror
, r1_bio
);
1193 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1194 sector_t s
= r1_bio
->sectors
;
1196 md_done_sync(mddev
, s
, uptodate
);
1200 static int fix_sync_read_error(r1bio_t
*r1_bio
)
1202 /* Try some synchronous reads of other devices to get
1203 * good data, much like with normal read errors. Only
1204 * read into the pages we already have so we don't
1205 * need to re-issue the read request.
1206 * We don't need to freeze the array, because being in an
1207 * active sync request, there is no normal IO, and
1208 * no overlapping syncs.
1210 mddev_t
*mddev
= r1_bio
->mddev
;
1211 conf_t
*conf
= mddev
->private;
1212 struct bio
*bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1213 sector_t sect
= r1_bio
->sector
;
1214 int sectors
= r1_bio
->sectors
;
1219 int d
= r1_bio
->read_disk
;
1224 if (s
> (PAGE_SIZE
>>9))
1227 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1228 /* No rcu protection needed here devices
1229 * can only be removed when no resync is
1230 * active, and resync is currently active
1232 rdev
= conf
->mirrors
[d
].rdev
;
1233 if (sync_page_io(rdev
,
1236 bio
->bi_io_vec
[idx
].bv_page
,
1243 if (d
== conf
->raid_disks
)
1245 } while (!success
&& d
!= r1_bio
->read_disk
);
1248 char b
[BDEVNAME_SIZE
];
1249 /* Cannot read from anywhere, array is toast */
1250 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
1251 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O read error"
1252 " for block %llu\n",
1254 bdevname(bio
->bi_bdev
, b
),
1255 (unsigned long long)r1_bio
->sector
);
1256 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1262 /* write it back and re-read */
1263 while (d
!= r1_bio
->read_disk
) {
1265 d
= conf
->raid_disks
;
1267 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1269 rdev
= conf
->mirrors
[d
].rdev
;
1270 if (sync_page_io(rdev
,
1273 bio
->bi_io_vec
[idx
].bv_page
,
1274 WRITE
, false) == 0) {
1275 r1_bio
->bios
[d
]->bi_end_io
= NULL
;
1276 rdev_dec_pending(rdev
, mddev
);
1277 md_error(mddev
, rdev
);
1279 atomic_add(s
, &rdev
->corrected_errors
);
1282 while (d
!= r1_bio
->read_disk
) {
1284 d
= conf
->raid_disks
;
1286 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1288 rdev
= conf
->mirrors
[d
].rdev
;
1289 if (sync_page_io(rdev
,
1292 bio
->bi_io_vec
[idx
].bv_page
,
1294 md_error(mddev
, rdev
);
1300 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1301 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1305 static int process_checks(r1bio_t
*r1_bio
)
1307 /* We have read all readable devices. If we haven't
1308 * got the block, then there is no hope left.
1309 * If we have, then we want to do a comparison
1310 * and skip the write if everything is the same.
1311 * If any blocks failed to read, then we need to
1312 * attempt an over-write
1314 mddev_t
*mddev
= r1_bio
->mddev
;
1315 conf_t
*conf
= mddev
->private;
1319 for (primary
= 0; primary
< conf
->raid_disks
; primary
++)
1320 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
1321 test_bit(BIO_UPTODATE
, &r1_bio
->bios
[primary
]->bi_flags
)) {
1322 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
1323 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
1326 r1_bio
->read_disk
= primary
;
1327 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1329 int vcnt
= r1_bio
->sectors
>> (PAGE_SHIFT
- 9);
1330 struct bio
*pbio
= r1_bio
->bios
[primary
];
1331 struct bio
*sbio
= r1_bio
->bios
[i
];
1334 if (r1_bio
->bios
[i
]->bi_end_io
!= end_sync_read
)
1337 if (test_bit(BIO_UPTODATE
, &sbio
->bi_flags
)) {
1338 for (j
= vcnt
; j
-- ; ) {
1340 p
= pbio
->bi_io_vec
[j
].bv_page
;
1341 s
= sbio
->bi_io_vec
[j
].bv_page
;
1342 if (memcmp(page_address(p
),
1350 mddev
->resync_mismatches
+= r1_bio
->sectors
;
1351 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
1352 && test_bit(BIO_UPTODATE
, &sbio
->bi_flags
))) {
1353 /* No need to write to this device. */
1354 sbio
->bi_end_io
= NULL
;
1355 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
1358 /* fixup the bio for reuse */
1359 sbio
->bi_vcnt
= vcnt
;
1360 sbio
->bi_size
= r1_bio
->sectors
<< 9;
1362 sbio
->bi_phys_segments
= 0;
1363 sbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1364 sbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1365 sbio
->bi_next
= NULL
;
1366 sbio
->bi_sector
= r1_bio
->sector
+
1367 conf
->mirrors
[i
].rdev
->data_offset
;
1368 sbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1369 size
= sbio
->bi_size
;
1370 for (j
= 0; j
< vcnt
; j
++) {
1372 bi
= &sbio
->bi_io_vec
[j
];
1374 if (size
> PAGE_SIZE
)
1375 bi
->bv_len
= PAGE_SIZE
;
1379 memcpy(page_address(bi
->bv_page
),
1380 page_address(pbio
->bi_io_vec
[j
].bv_page
),
1387 static void sync_request_write(mddev_t
*mddev
, r1bio_t
*r1_bio
)
1389 conf_t
*conf
= mddev
->private;
1391 int disks
= conf
->raid_disks
;
1392 struct bio
*bio
, *wbio
;
1394 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1396 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
))
1397 /* ouch - failed to read all of that. */
1398 if (!fix_sync_read_error(r1_bio
))
1401 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
1402 if (process_checks(r1_bio
) < 0)
1407 atomic_set(&r1_bio
->remaining
, 1);
1408 for (i
= 0; i
< disks
; i
++) {
1409 wbio
= r1_bio
->bios
[i
];
1410 if (wbio
->bi_end_io
== NULL
||
1411 (wbio
->bi_end_io
== end_sync_read
&&
1412 (i
== r1_bio
->read_disk
||
1413 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
1416 wbio
->bi_rw
= WRITE
;
1417 wbio
->bi_end_io
= end_sync_write
;
1418 atomic_inc(&r1_bio
->remaining
);
1419 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, wbio
->bi_size
>> 9);
1421 generic_make_request(wbio
);
1424 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1425 /* if we're here, all write(s) have completed, so clean up */
1426 md_done_sync(mddev
, r1_bio
->sectors
, 1);
1432 * This is a kernel thread which:
1434 * 1. Retries failed read operations on working mirrors.
1435 * 2. Updates the raid superblock when problems encounter.
1436 * 3. Performs writes following reads for array syncronising.
1439 static void fix_read_error(conf_t
*conf
, int read_disk
,
1440 sector_t sect
, int sectors
)
1442 mddev_t
*mddev
= conf
->mddev
;
1450 if (s
> (PAGE_SIZE
>>9))
1454 /* Note: no rcu protection needed here
1455 * as this is synchronous in the raid1d thread
1456 * which is the thread that might remove
1457 * a device. If raid1d ever becomes multi-threaded....
1459 rdev
= conf
->mirrors
[d
].rdev
;
1461 test_bit(In_sync
, &rdev
->flags
) &&
1462 sync_page_io(rdev
, sect
, s
<<9,
1463 conf
->tmppage
, READ
, false))
1467 if (d
== conf
->raid_disks
)
1470 } while (!success
&& d
!= read_disk
);
1473 /* Cannot read from anywhere -- bye bye array */
1474 md_error(mddev
, conf
->mirrors
[read_disk
].rdev
);
1477 /* write it back and re-read */
1479 while (d
!= read_disk
) {
1481 d
= conf
->raid_disks
;
1483 rdev
= conf
->mirrors
[d
].rdev
;
1485 test_bit(In_sync
, &rdev
->flags
)) {
1486 if (sync_page_io(rdev
, sect
, s
<<9,
1487 conf
->tmppage
, WRITE
, false)
1489 /* Well, this device is dead */
1490 md_error(mddev
, rdev
);
1494 while (d
!= read_disk
) {
1495 char b
[BDEVNAME_SIZE
];
1497 d
= conf
->raid_disks
;
1499 rdev
= conf
->mirrors
[d
].rdev
;
1501 test_bit(In_sync
, &rdev
->flags
)) {
1502 if (sync_page_io(rdev
, sect
, s
<<9,
1503 conf
->tmppage
, READ
, false)
1505 /* Well, this device is dead */
1506 md_error(mddev
, rdev
);
1508 atomic_add(s
, &rdev
->corrected_errors
);
1510 "md/raid1:%s: read error corrected "
1511 "(%d sectors at %llu on %s)\n",
1513 (unsigned long long)(sect
+
1515 bdevname(rdev
->bdev
, b
));
1524 static void raid1d(mddev_t
*mddev
)
1528 unsigned long flags
;
1529 conf_t
*conf
= mddev
->private;
1530 struct list_head
*head
= &conf
->retry_list
;
1532 struct blk_plug plug
;
1534 md_check_recovery(mddev
);
1536 blk_start_plug(&plug
);
1538 char b
[BDEVNAME_SIZE
];
1540 if (atomic_read(&mddev
->plug_cnt
) == 0)
1541 flush_pending_writes(conf
);
1543 spin_lock_irqsave(&conf
->device_lock
, flags
);
1544 if (list_empty(head
)) {
1545 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1548 r1_bio
= list_entry(head
->prev
, r1bio_t
, retry_list
);
1549 list_del(head
->prev
);
1551 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1553 mddev
= r1_bio
->mddev
;
1554 conf
= mddev
->private;
1555 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
))
1556 sync_request_write(mddev
, r1_bio
);
1560 /* we got a read error. Maybe the drive is bad. Maybe just
1561 * the block and we can fix it.
1562 * We freeze all other IO, and try reading the block from
1563 * other devices. When we find one, we re-write
1564 * and check it that fixes the read error.
1565 * This is all done synchronously while the array is
1568 if (mddev
->ro
== 0) {
1570 fix_read_error(conf
, r1_bio
->read_disk
,
1573 unfreeze_array(conf
);
1576 conf
->mirrors
[r1_bio
->read_disk
].rdev
);
1578 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1579 if ((disk
=read_balance(conf
, r1_bio
)) == -1) {
1580 printk(KERN_ALERT
"md/raid1:%s: %s: unrecoverable I/O"
1581 " read error for block %llu\n",
1583 bdevname(bio
->bi_bdev
,b
),
1584 (unsigned long long)r1_bio
->sector
);
1585 raid_end_bio_io(r1_bio
);
1587 const unsigned long do_sync
= r1_bio
->master_bio
->bi_rw
& REQ_SYNC
;
1588 r1_bio
->bios
[r1_bio
->read_disk
] =
1589 mddev
->ro
? IO_BLOCKED
: NULL
;
1590 r1_bio
->read_disk
= disk
;
1592 bio
= bio_clone_mddev(r1_bio
->master_bio
,
1594 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
1595 rdev
= conf
->mirrors
[disk
].rdev
;
1596 if (printk_ratelimit())
1597 printk(KERN_ERR
"md/raid1:%s: redirecting sector %llu to"
1598 " other mirror: %s\n",
1600 (unsigned long long)r1_bio
->sector
,
1601 bdevname(rdev
->bdev
,b
));
1602 bio
->bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
1603 bio
->bi_bdev
= rdev
->bdev
;
1604 bio
->bi_end_io
= raid1_end_read_request
;
1605 bio
->bi_rw
= READ
| do_sync
;
1606 bio
->bi_private
= r1_bio
;
1607 generic_make_request(bio
);
1612 blk_finish_plug(&plug
);
1616 static int init_resync(conf_t
*conf
)
1620 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1621 BUG_ON(conf
->r1buf_pool
);
1622 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
1624 if (!conf
->r1buf_pool
)
1626 conf
->next_resync
= 0;
1631 * perform a "sync" on one "block"
1633 * We need to make sure that no normal I/O request - particularly write
1634 * requests - conflict with active sync requests.
1636 * This is achieved by tracking pending requests and a 'barrier' concept
1637 * that can be installed to exclude normal IO requests.
1640 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
1642 conf_t
*conf
= mddev
->private;
1645 sector_t max_sector
, nr_sectors
;
1649 int write_targets
= 0, read_targets
= 0;
1650 sector_t sync_blocks
;
1651 int still_degraded
= 0;
1653 if (!conf
->r1buf_pool
)
1654 if (init_resync(conf
))
1657 max_sector
= mddev
->dev_sectors
;
1658 if (sector_nr
>= max_sector
) {
1659 /* If we aborted, we need to abort the
1660 * sync on the 'current' bitmap chunk (there will
1661 * only be one in raid1 resync.
1662 * We can find the current addess in mddev->curr_resync
1664 if (mddev
->curr_resync
< max_sector
) /* aborted */
1665 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
1667 else /* completed sync */
1670 bitmap_close_sync(mddev
->bitmap
);
1675 if (mddev
->bitmap
== NULL
&&
1676 mddev
->recovery_cp
== MaxSector
&&
1677 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
1678 conf
->fullsync
== 0) {
1680 return max_sector
- sector_nr
;
1682 /* before building a request, check if we can skip these blocks..
1683 * This call the bitmap_start_sync doesn't actually record anything
1685 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
1686 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1687 /* We can skip this block, and probably several more */
1692 * If there is non-resync activity waiting for a turn,
1693 * and resync is going fast enough,
1694 * then let it though before starting on this new sync request.
1696 if (!go_faster
&& conf
->nr_waiting
)
1697 msleep_interruptible(1000);
1699 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
1700 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
1701 raise_barrier(conf
);
1703 conf
->next_resync
= sector_nr
;
1707 * If we get a correctably read error during resync or recovery,
1708 * we might want to read from a different device. So we
1709 * flag all drives that could conceivably be read from for READ,
1710 * and any others (which will be non-In_sync devices) for WRITE.
1711 * If a read fails, we try reading from something else for which READ
1715 r1_bio
->mddev
= mddev
;
1716 r1_bio
->sector
= sector_nr
;
1718 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
1720 for (i
=0; i
< conf
->raid_disks
; i
++) {
1722 bio
= r1_bio
->bios
[i
];
1724 /* take from bio_init */
1725 bio
->bi_next
= NULL
;
1726 bio
->bi_flags
&= ~(BIO_POOL_MASK
-1);
1727 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
1728 bio
->bi_comp_cpu
= -1;
1732 bio
->bi_phys_segments
= 0;
1734 bio
->bi_end_io
= NULL
;
1735 bio
->bi_private
= NULL
;
1737 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1739 test_bit(Faulty
, &rdev
->flags
)) {
1742 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
1744 bio
->bi_end_io
= end_sync_write
;
1747 /* may need to read from here */
1749 bio
->bi_end_io
= end_sync_read
;
1750 if (test_bit(WriteMostly
, &rdev
->flags
)) {
1759 atomic_inc(&rdev
->nr_pending
);
1760 bio
->bi_sector
= sector_nr
+ rdev
->data_offset
;
1761 bio
->bi_bdev
= rdev
->bdev
;
1762 bio
->bi_private
= r1_bio
;
1767 r1_bio
->read_disk
= disk
;
1769 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
1770 /* extra read targets are also write targets */
1771 write_targets
+= read_targets
-1;
1773 if (write_targets
== 0 || read_targets
== 0) {
1774 /* There is nowhere to write, so all non-sync
1775 * drives must be failed - so we are finished
1777 sector_t rv
= max_sector
- sector_nr
;
1783 if (max_sector
> mddev
->resync_max
)
1784 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
1789 int len
= PAGE_SIZE
;
1790 if (sector_nr
+ (len
>>9) > max_sector
)
1791 len
= (max_sector
- sector_nr
) << 9;
1794 if (sync_blocks
== 0) {
1795 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
1796 &sync_blocks
, still_degraded
) &&
1798 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
1800 BUG_ON(sync_blocks
< (PAGE_SIZE
>>9));
1801 if ((len
>> 9) > sync_blocks
)
1802 len
= sync_blocks
<<9;
1805 for (i
=0 ; i
< conf
->raid_disks
; i
++) {
1806 bio
= r1_bio
->bios
[i
];
1807 if (bio
->bi_end_io
) {
1808 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
1809 if (bio_add_page(bio
, page
, len
, 0) == 0) {
1811 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
1814 bio
= r1_bio
->bios
[i
];
1815 if (bio
->bi_end_io
==NULL
)
1817 /* remove last page from this bio */
1819 bio
->bi_size
-= len
;
1820 bio
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
1826 nr_sectors
+= len
>>9;
1827 sector_nr
+= len
>>9;
1828 sync_blocks
-= (len
>>9);
1829 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
1831 r1_bio
->sectors
= nr_sectors
;
1833 /* For a user-requested sync, we read all readable devices and do a
1836 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1837 atomic_set(&r1_bio
->remaining
, read_targets
);
1838 for (i
=0; i
<conf
->raid_disks
; i
++) {
1839 bio
= r1_bio
->bios
[i
];
1840 if (bio
->bi_end_io
== end_sync_read
) {
1841 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1842 generic_make_request(bio
);
1846 atomic_set(&r1_bio
->remaining
, 1);
1847 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1848 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1849 generic_make_request(bio
);
1855 static sector_t
raid1_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
1860 return mddev
->dev_sectors
;
1863 static conf_t
*setup_conf(mddev_t
*mddev
)
1867 mirror_info_t
*disk
;
1871 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
1875 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
1880 conf
->tmppage
= alloc_page(GFP_KERNEL
);
1884 conf
->poolinfo
= kzalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
1885 if (!conf
->poolinfo
)
1887 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
;
1888 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
1891 if (!conf
->r1bio_pool
)
1894 conf
->poolinfo
->mddev
= mddev
;
1896 spin_lock_init(&conf
->device_lock
);
1897 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
1898 int disk_idx
= rdev
->raid_disk
;
1899 if (disk_idx
>= mddev
->raid_disks
1902 disk
= conf
->mirrors
+ disk_idx
;
1906 disk
->head_position
= 0;
1908 conf
->raid_disks
= mddev
->raid_disks
;
1909 conf
->mddev
= mddev
;
1910 INIT_LIST_HEAD(&conf
->retry_list
);
1912 spin_lock_init(&conf
->resync_lock
);
1913 init_waitqueue_head(&conf
->wait_barrier
);
1915 bio_list_init(&conf
->pending_bio_list
);
1917 conf
->last_used
= -1;
1918 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1920 disk
= conf
->mirrors
+ i
;
1923 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
1924 disk
->head_position
= 0;
1927 } else if (conf
->last_used
< 0)
1929 * The first working device is used as a
1930 * starting point to read balancing.
1932 conf
->last_used
= i
;
1936 if (conf
->last_used
< 0) {
1937 printk(KERN_ERR
"md/raid1:%s: no operational mirrors\n",
1942 conf
->thread
= md_register_thread(raid1d
, mddev
, NULL
);
1943 if (!conf
->thread
) {
1945 "md/raid1:%s: couldn't allocate thread\n",
1954 if (conf
->r1bio_pool
)
1955 mempool_destroy(conf
->r1bio_pool
);
1956 kfree(conf
->mirrors
);
1957 safe_put_page(conf
->tmppage
);
1958 kfree(conf
->poolinfo
);
1961 return ERR_PTR(err
);
1964 static int run(mddev_t
*mddev
)
1970 if (mddev
->level
!= 1) {
1971 printk(KERN_ERR
"md/raid1:%s: raid level not set to mirroring (%d)\n",
1972 mdname(mddev
), mddev
->level
);
1975 if (mddev
->reshape_position
!= MaxSector
) {
1976 printk(KERN_ERR
"md/raid1:%s: reshape_position set but not supported\n",
1981 * copy the already verified devices into our private RAID1
1982 * bookkeeping area. [whatever we allocate in run(),
1983 * should be freed in stop()]
1985 if (mddev
->private == NULL
)
1986 conf
= setup_conf(mddev
);
1988 conf
= mddev
->private;
1991 return PTR_ERR(conf
);
1993 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
1994 if (!mddev
->gendisk
)
1996 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
1997 rdev
->data_offset
<< 9);
1998 /* as we don't honour merge_bvec_fn, we must never risk
1999 * violating it, so limit ->max_segments to 1 lying within
2000 * a single page, as a one page request is never in violation.
2002 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
) {
2003 blk_queue_max_segments(mddev
->queue
, 1);
2004 blk_queue_segment_boundary(mddev
->queue
,
2005 PAGE_CACHE_SIZE
- 1);
2009 mddev
->degraded
= 0;
2010 for (i
=0; i
< conf
->raid_disks
; i
++)
2011 if (conf
->mirrors
[i
].rdev
== NULL
||
2012 !test_bit(In_sync
, &conf
->mirrors
[i
].rdev
->flags
) ||
2013 test_bit(Faulty
, &conf
->mirrors
[i
].rdev
->flags
))
2016 if (conf
->raid_disks
- mddev
->degraded
== 1)
2017 mddev
->recovery_cp
= MaxSector
;
2019 if (mddev
->recovery_cp
!= MaxSector
)
2020 printk(KERN_NOTICE
"md/raid1:%s: not clean"
2021 " -- starting background reconstruction\n",
2024 "md/raid1:%s: active with %d out of %d mirrors\n",
2025 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2029 * Ok, everything is just fine now
2031 mddev
->thread
= conf
->thread
;
2032 conf
->thread
= NULL
;
2033 mddev
->private = conf
;
2035 md_set_array_sectors(mddev
, raid1_size(mddev
, 0, 0));
2038 mddev
->queue
->backing_dev_info
.congested_fn
= raid1_congested
;
2039 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2041 return md_integrity_register(mddev
);
2044 static int stop(mddev_t
*mddev
)
2046 conf_t
*conf
= mddev
->private;
2047 struct bitmap
*bitmap
= mddev
->bitmap
;
2049 /* wait for behind writes to complete */
2050 if (bitmap
&& atomic_read(&bitmap
->behind_writes
) > 0) {
2051 printk(KERN_INFO
"md/raid1:%s: behind writes in progress - waiting to stop.\n",
2053 /* need to kick something here to make sure I/O goes? */
2054 wait_event(bitmap
->behind_wait
,
2055 atomic_read(&bitmap
->behind_writes
) == 0);
2058 raise_barrier(conf
);
2059 lower_barrier(conf
);
2061 md_unregister_thread(&mddev
->thread
);
2062 if (conf
->r1bio_pool
)
2063 mempool_destroy(conf
->r1bio_pool
);
2064 kfree(conf
->mirrors
);
2065 kfree(conf
->poolinfo
);
2067 mddev
->private = NULL
;
2071 static int raid1_resize(mddev_t
*mddev
, sector_t sectors
)
2073 /* no resync is happening, and there is enough space
2074 * on all devices, so we can resize.
2075 * We need to make sure resync covers any new space.
2076 * If the array is shrinking we should possibly wait until
2077 * any io in the removed space completes, but it hardly seems
2080 md_set_array_sectors(mddev
, raid1_size(mddev
, sectors
, 0));
2081 if (mddev
->array_sectors
> raid1_size(mddev
, sectors
, 0))
2083 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
2084 revalidate_disk(mddev
->gendisk
);
2085 if (sectors
> mddev
->dev_sectors
&&
2086 mddev
->recovery_cp
> mddev
->dev_sectors
) {
2087 mddev
->recovery_cp
= mddev
->dev_sectors
;
2088 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2090 mddev
->dev_sectors
= sectors
;
2091 mddev
->resync_max_sectors
= sectors
;
2095 static int raid1_reshape(mddev_t
*mddev
)
2098 * 1/ resize the r1bio_pool
2099 * 2/ resize conf->mirrors
2101 * We allocate a new r1bio_pool if we can.
2102 * Then raise a device barrier and wait until all IO stops.
2103 * Then resize conf->mirrors and swap in the new r1bio pool.
2105 * At the same time, we "pack" the devices so that all the missing
2106 * devices have the higher raid_disk numbers.
2108 mempool_t
*newpool
, *oldpool
;
2109 struct pool_info
*newpoolinfo
;
2110 mirror_info_t
*newmirrors
;
2111 conf_t
*conf
= mddev
->private;
2112 int cnt
, raid_disks
;
2113 unsigned long flags
;
2116 /* Cannot change chunk_size, layout, or level */
2117 if (mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
||
2118 mddev
->layout
!= mddev
->new_layout
||
2119 mddev
->level
!= mddev
->new_level
) {
2120 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
2121 mddev
->new_layout
= mddev
->layout
;
2122 mddev
->new_level
= mddev
->level
;
2126 err
= md_allow_write(mddev
);
2130 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
2132 if (raid_disks
< conf
->raid_disks
) {
2134 for (d
= 0; d
< conf
->raid_disks
; d
++)
2135 if (conf
->mirrors
[d
].rdev
)
2137 if (cnt
> raid_disks
)
2141 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
2144 newpoolinfo
->mddev
= mddev
;
2145 newpoolinfo
->raid_disks
= raid_disks
;
2147 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2148 r1bio_pool_free
, newpoolinfo
);
2153 newmirrors
= kzalloc(sizeof(struct mirror_info
) * raid_disks
, GFP_KERNEL
);
2156 mempool_destroy(newpool
);
2160 raise_barrier(conf
);
2162 /* ok, everything is stopped */
2163 oldpool
= conf
->r1bio_pool
;
2164 conf
->r1bio_pool
= newpool
;
2166 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
2167 mdk_rdev_t
*rdev
= conf
->mirrors
[d
].rdev
;
2168 if (rdev
&& rdev
->raid_disk
!= d2
) {
2170 sprintf(nm
, "rd%d", rdev
->raid_disk
);
2171 sysfs_remove_link(&mddev
->kobj
, nm
);
2172 rdev
->raid_disk
= d2
;
2173 sprintf(nm
, "rd%d", rdev
->raid_disk
);
2174 sysfs_remove_link(&mddev
->kobj
, nm
);
2175 if (sysfs_create_link(&mddev
->kobj
,
2178 "md/raid1:%s: cannot register "
2183 newmirrors
[d2
++].rdev
= rdev
;
2185 kfree(conf
->mirrors
);
2186 conf
->mirrors
= newmirrors
;
2187 kfree(conf
->poolinfo
);
2188 conf
->poolinfo
= newpoolinfo
;
2190 spin_lock_irqsave(&conf
->device_lock
, flags
);
2191 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
2192 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2193 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
2194 mddev
->delta_disks
= 0;
2196 conf
->last_used
= 0; /* just make sure it is in-range */
2197 lower_barrier(conf
);
2199 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2200 md_wakeup_thread(mddev
->thread
);
2202 mempool_destroy(oldpool
);
2206 static void raid1_quiesce(mddev_t
*mddev
, int state
)
2208 conf_t
*conf
= mddev
->private;
2211 case 2: /* wake for suspend */
2212 wake_up(&conf
->wait_barrier
);
2215 raise_barrier(conf
);
2218 lower_barrier(conf
);
2223 static void *raid1_takeover(mddev_t
*mddev
)
2225 /* raid1 can take over:
2226 * raid5 with 2 devices, any layout or chunk size
2228 if (mddev
->level
== 5 && mddev
->raid_disks
== 2) {
2230 mddev
->new_level
= 1;
2231 mddev
->new_layout
= 0;
2232 mddev
->new_chunk_sectors
= 0;
2233 conf
= setup_conf(mddev
);
2238 return ERR_PTR(-EINVAL
);
2241 static struct mdk_personality raid1_personality
=
2245 .owner
= THIS_MODULE
,
2246 .make_request
= make_request
,
2250 .error_handler
= error
,
2251 .hot_add_disk
= raid1_add_disk
,
2252 .hot_remove_disk
= raid1_remove_disk
,
2253 .spare_active
= raid1_spare_active
,
2254 .sync_request
= sync_request
,
2255 .resize
= raid1_resize
,
2257 .check_reshape
= raid1_reshape
,
2258 .quiesce
= raid1_quiesce
,
2259 .takeover
= raid1_takeover
,
2262 static int __init
raid_init(void)
2264 return register_md_personality(&raid1_personality
);
2267 static void raid_exit(void)
2269 unregister_md_personality(&raid1_personality
);
2272 module_init(raid_init
);
2273 module_exit(raid_exit
);
2274 MODULE_LICENSE("GPL");
2275 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2276 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2277 MODULE_ALIAS("md-raid1");
2278 MODULE_ALIAS("md-level-1");