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 "dm-bio-list.h"
35 #include <linux/delay.h>
36 #include <linux/raid/raid1.h>
37 #include <linux/raid/bitmap.h>
41 #define PRINTK(x...) printk(x)
47 * Number of guaranteed r1bios in case of extreme VM load:
49 #define NR_RAID1_BIOS 256
52 static void unplug_slaves(mddev_t
*mddev
);
54 static void allow_barrier(conf_t
*conf
);
55 static void lower_barrier(conf_t
*conf
);
57 static void * r1bio_pool_alloc(gfp_t gfp_flags
, void *data
)
59 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 r1_bio
= kzalloc(size
, gfp_flags
);
66 unplug_slaves(pi
->mddev
);
71 static void r1bio_pool_free(void *r1_bio
, void *data
)
76 #define RESYNC_BLOCK_SIZE (64*1024)
77 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
78 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
79 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
80 #define RESYNC_WINDOW (2048*1024)
82 static void * r1buf_pool_alloc(gfp_t gfp_flags
, void *data
)
84 struct pool_info
*pi
= data
;
90 r1_bio
= r1bio_pool_alloc(gfp_flags
, pi
);
92 unplug_slaves(pi
->mddev
);
97 * Allocate bios : 1 for reading, n-1 for writing
99 for (j
= pi
->raid_disks
; j
-- ; ) {
100 bio
= bio_alloc(gfp_flags
, RESYNC_PAGES
);
103 r1_bio
->bios
[j
] = bio
;
106 * Allocate RESYNC_PAGES data pages and attach them to
108 * If this is a user-requested check/repair, allocate
109 * RESYNC_PAGES for each bio.
111 if (test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
))
116 bio
= r1_bio
->bios
[j
];
117 for (i
= 0; i
< RESYNC_PAGES
; i
++) {
118 page
= alloc_page(gfp_flags
);
122 bio
->bi_io_vec
[i
].bv_page
= page
;
125 /* If not user-requests, copy the page pointers to all bios */
126 if (!test_bit(MD_RECOVERY_REQUESTED
, &pi
->mddev
->recovery
)) {
127 for (i
=0; i
<RESYNC_PAGES
; i
++)
128 for (j
=1; j
<pi
->raid_disks
; j
++)
129 r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
=
130 r1_bio
->bios
[0]->bi_io_vec
[i
].bv_page
;
133 r1_bio
->master_bio
= NULL
;
138 for (i
=0; i
< RESYNC_PAGES
; i
++)
139 for (j
=0 ; j
< pi
->raid_disks
; j
++)
140 safe_put_page(r1_bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
143 while ( ++j
< pi
->raid_disks
)
144 bio_put(r1_bio
->bios
[j
]);
145 r1bio_pool_free(r1_bio
, data
);
149 static void r1buf_pool_free(void *__r1_bio
, void *data
)
151 struct pool_info
*pi
= data
;
153 r1bio_t
*r1bio
= __r1_bio
;
155 for (i
= 0; i
< RESYNC_PAGES
; i
++)
156 for (j
= pi
->raid_disks
; j
-- ;) {
158 r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
!=
159 r1bio
->bios
[0]->bi_io_vec
[i
].bv_page
)
160 safe_put_page(r1bio
->bios
[j
]->bi_io_vec
[i
].bv_page
);
162 for (i
=0 ; i
< pi
->raid_disks
; i
++)
163 bio_put(r1bio
->bios
[i
]);
165 r1bio_pool_free(r1bio
, data
);
168 static void put_all_bios(conf_t
*conf
, r1bio_t
*r1_bio
)
172 for (i
= 0; i
< conf
->raid_disks
; i
++) {
173 struct bio
**bio
= r1_bio
->bios
+ i
;
174 if (*bio
&& *bio
!= IO_BLOCKED
)
180 static void free_r1bio(r1bio_t
*r1_bio
)
182 conf_t
*conf
= mddev_to_conf(r1_bio
->mddev
);
185 * Wake up any possible resync thread that waits for the device
190 put_all_bios(conf
, r1_bio
);
191 mempool_free(r1_bio
, conf
->r1bio_pool
);
194 static void put_buf(r1bio_t
*r1_bio
)
196 conf_t
*conf
= mddev_to_conf(r1_bio
->mddev
);
199 for (i
=0; i
<conf
->raid_disks
; i
++) {
200 struct bio
*bio
= r1_bio
->bios
[i
];
202 rdev_dec_pending(conf
->mirrors
[i
].rdev
, r1_bio
->mddev
);
205 mempool_free(r1_bio
, conf
->r1buf_pool
);
210 static void reschedule_retry(r1bio_t
*r1_bio
)
213 mddev_t
*mddev
= r1_bio
->mddev
;
214 conf_t
*conf
= mddev_to_conf(mddev
);
216 spin_lock_irqsave(&conf
->device_lock
, flags
);
217 list_add(&r1_bio
->retry_list
, &conf
->retry_list
);
219 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
221 wake_up(&conf
->wait_barrier
);
222 md_wakeup_thread(mddev
->thread
);
226 * raid_end_bio_io() is called when we have finished servicing a mirrored
227 * operation and are ready to return a success/failure code to the buffer
230 static void raid_end_bio_io(r1bio_t
*r1_bio
)
232 struct bio
*bio
= r1_bio
->master_bio
;
234 /* if nobody has done the final endio yet, do it now */
235 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
236 PRINTK(KERN_DEBUG
"raid1: sync end %s on sectors %llu-%llu\n",
237 (bio_data_dir(bio
) == WRITE
) ? "write" : "read",
238 (unsigned long long) bio
->bi_sector
,
239 (unsigned long long) bio
->bi_sector
+
240 (bio
->bi_size
>> 9) - 1);
243 test_bit(R1BIO_Uptodate
, &r1_bio
->state
) ? 0 : -EIO
);
249 * Update disk head position estimator based on IRQ completion info.
251 static inline void update_head_pos(int disk
, r1bio_t
*r1_bio
)
253 conf_t
*conf
= mddev_to_conf(r1_bio
->mddev
);
255 conf
->mirrors
[disk
].head_position
=
256 r1_bio
->sector
+ (r1_bio
->sectors
);
259 static void raid1_end_read_request(struct bio
*bio
, int error
)
261 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
262 r1bio_t
* r1_bio
= (r1bio_t
*)(bio
->bi_private
);
264 conf_t
*conf
= mddev_to_conf(r1_bio
->mddev
);
266 mirror
= r1_bio
->read_disk
;
268 * this branch is our 'one mirror IO has finished' event handler:
270 update_head_pos(mirror
, r1_bio
);
273 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
275 /* If all other devices have failed, we want to return
276 * the error upwards rather than fail the last device.
277 * Here we redefine "uptodate" to mean "Don't want to retry"
280 spin_lock_irqsave(&conf
->device_lock
, flags
);
281 if (r1_bio
->mddev
->degraded
== conf
->raid_disks
||
282 (r1_bio
->mddev
->degraded
== conf
->raid_disks
-1 &&
283 !test_bit(Faulty
, &conf
->mirrors
[mirror
].rdev
->flags
)))
285 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
289 raid_end_bio_io(r1_bio
);
294 char b
[BDEVNAME_SIZE
];
295 if (printk_ratelimit())
296 printk(KERN_ERR
"raid1: %s: rescheduling sector %llu\n",
297 bdevname(conf
->mirrors
[mirror
].rdev
->bdev
,b
), (unsigned long long)r1_bio
->sector
);
298 reschedule_retry(r1_bio
);
301 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
304 static void raid1_end_write_request(struct bio
*bio
, int error
)
306 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
307 r1bio_t
* r1_bio
= (r1bio_t
*)(bio
->bi_private
);
308 int mirror
, behind
= test_bit(R1BIO_BehindIO
, &r1_bio
->state
);
309 conf_t
*conf
= mddev_to_conf(r1_bio
->mddev
);
310 struct bio
*to_put
= NULL
;
313 for (mirror
= 0; mirror
< conf
->raid_disks
; mirror
++)
314 if (r1_bio
->bios
[mirror
] == bio
)
317 if (error
== -EOPNOTSUPP
&& test_bit(R1BIO_Barrier
, &r1_bio
->state
)) {
318 set_bit(BarriersNotsupp
, &conf
->mirrors
[mirror
].rdev
->flags
);
319 set_bit(R1BIO_BarrierRetry
, &r1_bio
->state
);
320 r1_bio
->mddev
->barriers_work
= 0;
321 /* Don't rdev_dec_pending in this branch - keep it for the retry */
324 * this branch is our 'one mirror IO has finished' event handler:
326 r1_bio
->bios
[mirror
] = NULL
;
329 md_error(r1_bio
->mddev
, conf
->mirrors
[mirror
].rdev
);
330 /* an I/O failed, we can't clear the bitmap */
331 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
334 * Set R1BIO_Uptodate in our master bio, so that
335 * we will return a good error code for to the higher
336 * levels even if IO on some other mirrored buffer fails.
338 * The 'master' represents the composite IO operation to
339 * user-side. So if something waits for IO, then it will
340 * wait for the 'master' bio.
342 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
344 update_head_pos(mirror
, r1_bio
);
347 if (test_bit(WriteMostly
, &conf
->mirrors
[mirror
].rdev
->flags
))
348 atomic_dec(&r1_bio
->behind_remaining
);
350 /* In behind mode, we ACK the master bio once the I/O has safely
351 * reached all non-writemostly disks. Setting the Returned bit
352 * ensures that this gets done only once -- we don't ever want to
353 * return -EIO here, instead we'll wait */
355 if (atomic_read(&r1_bio
->behind_remaining
) >= (atomic_read(&r1_bio
->remaining
)-1) &&
356 test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
357 /* Maybe we can return now */
358 if (!test_and_set_bit(R1BIO_Returned
, &r1_bio
->state
)) {
359 struct bio
*mbio
= r1_bio
->master_bio
;
360 PRINTK(KERN_DEBUG
"raid1: behind end write sectors %llu-%llu\n",
361 (unsigned long long) mbio
->bi_sector
,
362 (unsigned long long) mbio
->bi_sector
+
363 (mbio
->bi_size
>> 9) - 1);
368 rdev_dec_pending(conf
->mirrors
[mirror
].rdev
, conf
->mddev
);
372 * Let's see if all mirrored write operations have finished
375 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
376 if (test_bit(R1BIO_BarrierRetry
, &r1_bio
->state
))
377 reschedule_retry(r1_bio
);
379 /* it really is the end of this request */
380 if (test_bit(R1BIO_BehindIO
, &r1_bio
->state
)) {
381 /* free extra copy of the data pages */
382 int i
= bio
->bi_vcnt
;
384 safe_put_page(bio
->bi_io_vec
[i
].bv_page
);
386 /* clear the bitmap if all writes complete successfully */
387 bitmap_endwrite(r1_bio
->mddev
->bitmap
, r1_bio
->sector
,
389 !test_bit(R1BIO_Degraded
, &r1_bio
->state
),
391 md_write_end(r1_bio
->mddev
);
392 raid_end_bio_io(r1_bio
);
402 * This routine returns the disk from which the requested read should
403 * be done. There is a per-array 'next expected sequential IO' sector
404 * number - if this matches on the next IO then we use the last disk.
405 * There is also a per-disk 'last know head position' sector that is
406 * maintained from IRQ contexts, both the normal and the resync IO
407 * completion handlers update this position correctly. If there is no
408 * perfect sequential match then we pick the disk whose head is closest.
410 * If there are 2 mirrors in the same 2 devices, performance degrades
411 * because position is mirror, not device based.
413 * The rdev for the device selected will have nr_pending incremented.
415 static int read_balance(conf_t
*conf
, r1bio_t
*r1_bio
)
417 const unsigned long this_sector
= r1_bio
->sector
;
418 int new_disk
= conf
->last_used
, disk
= new_disk
;
420 const int sectors
= r1_bio
->sectors
;
421 sector_t new_distance
, current_distance
;
426 * Check if we can balance. We can balance on the whole
427 * device if no resync is going on, or below the resync window.
428 * We take the first readable disk when above the resync window.
431 if (conf
->mddev
->recovery_cp
< MaxSector
&&
432 (this_sector
+ sectors
>= conf
->next_resync
)) {
433 /* Choose the first operation device, for consistancy */
436 for (rdev
= rcu_dereference(conf
->mirrors
[new_disk
].rdev
);
437 r1_bio
->bios
[new_disk
] == IO_BLOCKED
||
438 !rdev
|| !test_bit(In_sync
, &rdev
->flags
)
439 || test_bit(WriteMostly
, &rdev
->flags
);
440 rdev
= rcu_dereference(conf
->mirrors
[++new_disk
].rdev
)) {
442 if (rdev
&& test_bit(In_sync
, &rdev
->flags
) &&
443 r1_bio
->bios
[new_disk
] != IO_BLOCKED
)
444 wonly_disk
= new_disk
;
446 if (new_disk
== conf
->raid_disks
- 1) {
447 new_disk
= wonly_disk
;
455 /* make sure the disk is operational */
456 for (rdev
= rcu_dereference(conf
->mirrors
[new_disk
].rdev
);
457 r1_bio
->bios
[new_disk
] == IO_BLOCKED
||
458 !rdev
|| !test_bit(In_sync
, &rdev
->flags
) ||
459 test_bit(WriteMostly
, &rdev
->flags
);
460 rdev
= rcu_dereference(conf
->mirrors
[new_disk
].rdev
)) {
462 if (rdev
&& test_bit(In_sync
, &rdev
->flags
) &&
463 r1_bio
->bios
[new_disk
] != IO_BLOCKED
)
464 wonly_disk
= new_disk
;
467 new_disk
= conf
->raid_disks
;
469 if (new_disk
== disk
) {
470 new_disk
= wonly_disk
;
479 /* now disk == new_disk == starting point for search */
482 * Don't change to another disk for sequential reads:
484 if (conf
->next_seq_sect
== this_sector
)
486 if (this_sector
== conf
->mirrors
[new_disk
].head_position
)
489 current_distance
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
491 /* Find the disk whose head is closest */
495 disk
= conf
->raid_disks
;
498 rdev
= rcu_dereference(conf
->mirrors
[disk
].rdev
);
500 if (!rdev
|| r1_bio
->bios
[disk
] == IO_BLOCKED
||
501 !test_bit(In_sync
, &rdev
->flags
) ||
502 test_bit(WriteMostly
, &rdev
->flags
))
505 if (!atomic_read(&rdev
->nr_pending
)) {
509 new_distance
= abs(this_sector
- conf
->mirrors
[disk
].head_position
);
510 if (new_distance
< current_distance
) {
511 current_distance
= new_distance
;
514 } while (disk
!= conf
->last_used
);
520 rdev
= rcu_dereference(conf
->mirrors
[new_disk
].rdev
);
523 atomic_inc(&rdev
->nr_pending
);
524 if (!test_bit(In_sync
, &rdev
->flags
)) {
525 /* cannot risk returning a device that failed
526 * before we inc'ed nr_pending
528 rdev_dec_pending(rdev
, conf
->mddev
);
531 conf
->next_seq_sect
= this_sector
+ sectors
;
532 conf
->last_used
= new_disk
;
539 static void unplug_slaves(mddev_t
*mddev
)
541 conf_t
*conf
= mddev_to_conf(mddev
);
545 for (i
=0; i
<mddev
->raid_disks
; i
++) {
546 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
547 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
548 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
550 atomic_inc(&rdev
->nr_pending
);
555 rdev_dec_pending(rdev
, mddev
);
562 static void raid1_unplug(struct request_queue
*q
)
564 mddev_t
*mddev
= q
->queuedata
;
566 unplug_slaves(mddev
);
567 md_wakeup_thread(mddev
->thread
);
570 static int raid1_congested(void *data
, int bits
)
572 mddev_t
*mddev
= data
;
573 conf_t
*conf
= mddev_to_conf(mddev
);
577 for (i
= 0; i
< mddev
->raid_disks
; i
++) {
578 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
579 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
580 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
582 /* Note the '|| 1' - when read_balance prefers
583 * non-congested targets, it can be removed
585 if ((bits
& (1<<BDI_write_congested
)) || 1)
586 ret
|= bdi_congested(&q
->backing_dev_info
, bits
);
588 ret
&= bdi_congested(&q
->backing_dev_info
, bits
);
596 static int flush_pending_writes(conf_t
*conf
)
598 /* Any writes that have been queued but are awaiting
599 * bitmap updates get flushed here.
600 * We return 1 if any requests were actually submitted.
604 spin_lock_irq(&conf
->device_lock
);
606 if (conf
->pending_bio_list
.head
) {
608 bio
= bio_list_get(&conf
->pending_bio_list
);
609 blk_remove_plug(conf
->mddev
->queue
);
610 spin_unlock_irq(&conf
->device_lock
);
611 /* flush any pending bitmap writes to
612 * disk before proceeding w/ I/O */
613 bitmap_unplug(conf
->mddev
->bitmap
);
615 while (bio
) { /* submit pending writes */
616 struct bio
*next
= bio
->bi_next
;
618 generic_make_request(bio
);
623 spin_unlock_irq(&conf
->device_lock
);
628 * Sometimes we need to suspend IO while we do something else,
629 * either some resync/recovery, or reconfigure the array.
630 * To do this we raise a 'barrier'.
631 * The 'barrier' is a counter that can be raised multiple times
632 * to count how many activities are happening which preclude
634 * We can only raise the barrier if there is no pending IO.
635 * i.e. if nr_pending == 0.
636 * We choose only to raise the barrier if no-one is waiting for the
637 * barrier to go down. This means that as soon as an IO request
638 * is ready, no other operations which require a barrier will start
639 * until the IO request has had a chance.
641 * So: regular IO calls 'wait_barrier'. When that returns there
642 * is no backgroup IO happening, It must arrange to call
643 * allow_barrier when it has finished its IO.
644 * backgroup IO calls must call raise_barrier. Once that returns
645 * there is no normal IO happeing. It must arrange to call
646 * lower_barrier when the particular background IO completes.
648 #define RESYNC_DEPTH 32
650 static void raise_barrier(conf_t
*conf
)
652 spin_lock_irq(&conf
->resync_lock
);
654 /* Wait until no block IO is waiting */
655 wait_event_lock_irq(conf
->wait_barrier
, !conf
->nr_waiting
,
657 raid1_unplug(conf
->mddev
->queue
));
659 /* block any new IO from starting */
662 /* No wait for all pending IO to complete */
663 wait_event_lock_irq(conf
->wait_barrier
,
664 !conf
->nr_pending
&& conf
->barrier
< RESYNC_DEPTH
,
666 raid1_unplug(conf
->mddev
->queue
));
668 spin_unlock_irq(&conf
->resync_lock
);
671 static void lower_barrier(conf_t
*conf
)
674 spin_lock_irqsave(&conf
->resync_lock
, flags
);
676 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
677 wake_up(&conf
->wait_barrier
);
680 static void wait_barrier(conf_t
*conf
)
682 spin_lock_irq(&conf
->resync_lock
);
685 wait_event_lock_irq(conf
->wait_barrier
, !conf
->barrier
,
687 raid1_unplug(conf
->mddev
->queue
));
691 spin_unlock_irq(&conf
->resync_lock
);
694 static void allow_barrier(conf_t
*conf
)
697 spin_lock_irqsave(&conf
->resync_lock
, flags
);
699 spin_unlock_irqrestore(&conf
->resync_lock
, flags
);
700 wake_up(&conf
->wait_barrier
);
703 static void freeze_array(conf_t
*conf
)
705 /* stop syncio and normal IO and wait for everything to
707 * We increment barrier and nr_waiting, and then
708 * wait until nr_pending match nr_queued+1
709 * This is called in the context of one normal IO request
710 * that has failed. Thus any sync request that might be pending
711 * will be blocked by nr_pending, and we need to wait for
712 * pending IO requests to complete or be queued for re-try.
713 * Thus the number queued (nr_queued) plus this request (1)
714 * must match the number of pending IOs (nr_pending) before
717 spin_lock_irq(&conf
->resync_lock
);
720 wait_event_lock_irq(conf
->wait_barrier
,
721 conf
->nr_pending
== conf
->nr_queued
+1,
723 ({ flush_pending_writes(conf
);
724 raid1_unplug(conf
->mddev
->queue
); }));
725 spin_unlock_irq(&conf
->resync_lock
);
727 static void unfreeze_array(conf_t
*conf
)
729 /* reverse the effect of the freeze */
730 spin_lock_irq(&conf
->resync_lock
);
733 wake_up(&conf
->wait_barrier
);
734 spin_unlock_irq(&conf
->resync_lock
);
738 /* duplicate the data pages for behind I/O */
739 static struct page
**alloc_behind_pages(struct bio
*bio
)
742 struct bio_vec
*bvec
;
743 struct page
**pages
= kzalloc(bio
->bi_vcnt
* sizeof(struct page
*),
745 if (unlikely(!pages
))
748 bio_for_each_segment(bvec
, bio
, i
) {
749 pages
[i
] = alloc_page(GFP_NOIO
);
750 if (unlikely(!pages
[i
]))
752 memcpy(kmap(pages
[i
]) + bvec
->bv_offset
,
753 kmap(bvec
->bv_page
) + bvec
->bv_offset
, bvec
->bv_len
);
755 kunmap(bvec
->bv_page
);
762 for (i
= 0; i
< bio
->bi_vcnt
&& pages
[i
]; i
++)
765 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio
->bi_size
);
769 static int make_request(struct request_queue
*q
, struct bio
* bio
)
771 mddev_t
*mddev
= q
->queuedata
;
772 conf_t
*conf
= mddev_to_conf(mddev
);
773 mirror_info_t
*mirror
;
775 struct bio
*read_bio
;
776 int i
, targets
= 0, disks
;
777 struct bitmap
*bitmap
;
780 struct page
**behind_pages
= NULL
;
781 const int rw
= bio_data_dir(bio
);
782 const int do_sync
= bio_sync(bio
);
783 int cpu
, do_barriers
;
784 mdk_rdev_t
*blocked_rdev
;
787 * Register the new request and wait if the reconstruction
788 * thread has put up a bar for new requests.
789 * Continue immediately if no resync is active currently.
790 * We test barriers_work *after* md_write_start as md_write_start
791 * may cause the first superblock write, and that will check out
795 md_write_start(mddev
, bio
); /* wait on superblock update early */
797 if (unlikely(!mddev
->barriers_work
&& bio_barrier(bio
))) {
800 bio_endio(bio
, -EOPNOTSUPP
);
806 bitmap
= mddev
->bitmap
;
808 cpu
= part_stat_lock();
809 part_stat_inc(cpu
, &mddev
->gendisk
->part0
, ios
[rw
]);
810 part_stat_add(cpu
, &mddev
->gendisk
->part0
, sectors
[rw
],
815 * make_request() can abort the operation when READA is being
816 * used and no empty request is available.
819 r1_bio
= mempool_alloc(conf
->r1bio_pool
, GFP_NOIO
);
821 r1_bio
->master_bio
= bio
;
822 r1_bio
->sectors
= bio
->bi_size
>> 9;
824 r1_bio
->mddev
= mddev
;
825 r1_bio
->sector
= bio
->bi_sector
;
829 * read balancing logic:
831 int rdisk
= read_balance(conf
, r1_bio
);
834 /* couldn't find anywhere to read from */
835 raid_end_bio_io(r1_bio
);
838 mirror
= conf
->mirrors
+ rdisk
;
840 r1_bio
->read_disk
= rdisk
;
842 read_bio
= bio_clone(bio
, GFP_NOIO
);
844 r1_bio
->bios
[rdisk
] = read_bio
;
846 read_bio
->bi_sector
= r1_bio
->sector
+ mirror
->rdev
->data_offset
;
847 read_bio
->bi_bdev
= mirror
->rdev
->bdev
;
848 read_bio
->bi_end_io
= raid1_end_read_request
;
849 read_bio
->bi_rw
= READ
| do_sync
;
850 read_bio
->bi_private
= r1_bio
;
852 generic_make_request(read_bio
);
859 /* first select target devices under spinlock and
860 * inc refcount on their rdev. Record them by setting
863 disks
= conf
->raid_disks
;
865 { static int first
=1;
866 if (first
) printk("First Write sector %llu disks %d\n",
867 (unsigned long long)r1_bio
->sector
, disks
);
874 for (i
= 0; i
< disks
; i
++) {
875 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
876 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
877 atomic_inc(&rdev
->nr_pending
);
881 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
)) {
882 atomic_inc(&rdev
->nr_pending
);
883 if (test_bit(Faulty
, &rdev
->flags
)) {
884 rdev_dec_pending(rdev
, mddev
);
885 r1_bio
->bios
[i
] = NULL
;
887 r1_bio
->bios
[i
] = bio
;
890 r1_bio
->bios
[i
] = NULL
;
894 if (unlikely(blocked_rdev
)) {
895 /* Wait for this device to become unblocked */
898 for (j
= 0; j
< i
; j
++)
900 rdev_dec_pending(conf
->mirrors
[j
].rdev
, mddev
);
903 md_wait_for_blocked_rdev(blocked_rdev
, mddev
);
908 BUG_ON(targets
== 0); /* we never fail the last device */
910 if (targets
< conf
->raid_disks
) {
911 /* array is degraded, we will not clear the bitmap
912 * on I/O completion (see raid1_end_write_request) */
913 set_bit(R1BIO_Degraded
, &r1_bio
->state
);
916 /* do behind I/O ? */
918 atomic_read(&bitmap
->behind_writes
) < bitmap
->max_write_behind
&&
919 (behind_pages
= alloc_behind_pages(bio
)) != NULL
)
920 set_bit(R1BIO_BehindIO
, &r1_bio
->state
);
922 atomic_set(&r1_bio
->remaining
, 0);
923 atomic_set(&r1_bio
->behind_remaining
, 0);
925 do_barriers
= bio_barrier(bio
);
927 set_bit(R1BIO_Barrier
, &r1_bio
->state
);
930 for (i
= 0; i
< disks
; i
++) {
932 if (!r1_bio
->bios
[i
])
935 mbio
= bio_clone(bio
, GFP_NOIO
);
936 r1_bio
->bios
[i
] = mbio
;
938 mbio
->bi_sector
= r1_bio
->sector
+ conf
->mirrors
[i
].rdev
->data_offset
;
939 mbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
940 mbio
->bi_end_io
= raid1_end_write_request
;
941 mbio
->bi_rw
= WRITE
| do_barriers
| do_sync
;
942 mbio
->bi_private
= r1_bio
;
945 struct bio_vec
*bvec
;
948 /* Yes, I really want the '__' version so that
949 * we clear any unused pointer in the io_vec, rather
950 * than leave them unchanged. This is important
951 * because when we come to free the pages, we won't
952 * know the originial bi_idx, so we just free
955 __bio_for_each_segment(bvec
, mbio
, j
, 0)
956 bvec
->bv_page
= behind_pages
[j
];
957 if (test_bit(WriteMostly
, &conf
->mirrors
[i
].rdev
->flags
))
958 atomic_inc(&r1_bio
->behind_remaining
);
961 atomic_inc(&r1_bio
->remaining
);
963 bio_list_add(&bl
, mbio
);
965 kfree(behind_pages
); /* the behind pages are attached to the bios now */
967 bitmap_startwrite(bitmap
, bio
->bi_sector
, r1_bio
->sectors
,
968 test_bit(R1BIO_BehindIO
, &r1_bio
->state
));
969 spin_lock_irqsave(&conf
->device_lock
, flags
);
970 bio_list_merge(&conf
->pending_bio_list
, &bl
);
973 blk_plug_device(mddev
->queue
);
974 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
976 /* In case raid1d snuck into freeze_array */
977 wake_up(&conf
->wait_barrier
);
980 md_wakeup_thread(mddev
->thread
);
982 while ((bio
= bio_list_pop(&bl
)) != NULL
)
983 generic_make_request(bio
);
989 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
991 conf_t
*conf
= mddev_to_conf(mddev
);
994 seq_printf(seq
, " [%d/%d] [", conf
->raid_disks
,
995 conf
->raid_disks
- mddev
->degraded
);
997 for (i
= 0; i
< conf
->raid_disks
; i
++) {
998 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
999 seq_printf(seq
, "%s",
1000 rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
1003 seq_printf(seq
, "]");
1007 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1009 char b
[BDEVNAME_SIZE
];
1010 conf_t
*conf
= mddev_to_conf(mddev
);
1013 * If it is not operational, then we have already marked it as dead
1014 * else if it is the last working disks, ignore the error, let the
1015 * next level up know.
1016 * else mark the drive as failed
1018 if (test_bit(In_sync
, &rdev
->flags
)
1019 && (conf
->raid_disks
- mddev
->degraded
) == 1) {
1021 * Don't fail the drive, act as though we were just a
1022 * normal single drive.
1023 * However don't try a recovery from this drive as
1024 * it is very likely to fail.
1026 mddev
->recovery_disabled
= 1;
1029 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1030 unsigned long flags
;
1031 spin_lock_irqsave(&conf
->device_lock
, flags
);
1033 set_bit(Faulty
, &rdev
->flags
);
1034 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1036 * if recovery is running, make sure it aborts.
1038 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1040 set_bit(Faulty
, &rdev
->flags
);
1041 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1042 printk(KERN_ALERT
"raid1: Disk failure on %s, disabling device.\n"
1043 "raid1: Operation continuing on %d devices.\n",
1044 bdevname(rdev
->bdev
,b
), conf
->raid_disks
- mddev
->degraded
);
1047 static void print_conf(conf_t
*conf
)
1051 printk("RAID1 conf printout:\n");
1053 printk("(!conf)\n");
1056 printk(" --- wd:%d rd:%d\n", conf
->raid_disks
- conf
->mddev
->degraded
,
1060 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1061 char b
[BDEVNAME_SIZE
];
1062 mdk_rdev_t
*rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1064 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1065 i
, !test_bit(In_sync
, &rdev
->flags
),
1066 !test_bit(Faulty
, &rdev
->flags
),
1067 bdevname(rdev
->bdev
,b
));
1072 static void close_sync(conf_t
*conf
)
1075 allow_barrier(conf
);
1077 mempool_destroy(conf
->r1buf_pool
);
1078 conf
->r1buf_pool
= NULL
;
1081 static int raid1_spare_active(mddev_t
*mddev
)
1084 conf_t
*conf
= mddev
->private;
1087 * Find all failed disks within the RAID1 configuration
1088 * and mark them readable.
1089 * Called under mddev lock, so rcu protection not needed.
1091 for (i
= 0; i
< conf
->raid_disks
; i
++) {
1092 mdk_rdev_t
*rdev
= conf
->mirrors
[i
].rdev
;
1094 && !test_bit(Faulty
, &rdev
->flags
)
1095 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
1096 unsigned long flags
;
1097 spin_lock_irqsave(&conf
->device_lock
, flags
);
1099 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1108 static int raid1_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1110 conf_t
*conf
= mddev
->private;
1115 int last
= mddev
->raid_disks
- 1;
1117 if (rdev
->raid_disk
>= 0)
1118 first
= last
= rdev
->raid_disk
;
1120 for (mirror
= first
; mirror
<= last
; mirror
++)
1121 if ( !(p
=conf
->mirrors
+mirror
)->rdev
) {
1123 blk_queue_stack_limits(mddev
->queue
,
1124 rdev
->bdev
->bd_disk
->queue
);
1125 /* as we don't honour merge_bvec_fn, we must never risk
1126 * violating it, so limit ->max_sector to one PAGE, as
1127 * a one page request is never in violation.
1129 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
1130 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
1131 blk_queue_max_sectors(mddev
->queue
, PAGE_SIZE
>>9);
1133 p
->head_position
= 0;
1134 rdev
->raid_disk
= mirror
;
1136 /* As all devices are equivalent, we don't need a full recovery
1137 * if this was recently any drive of the array
1139 if (rdev
->saved_raid_disk
< 0)
1141 rcu_assign_pointer(p
->rdev
, rdev
);
1149 static int raid1_remove_disk(mddev_t
*mddev
, int number
)
1151 conf_t
*conf
= mddev
->private;
1154 mirror_info_t
*p
= conf
->mirrors
+ number
;
1159 if (test_bit(In_sync
, &rdev
->flags
) ||
1160 atomic_read(&rdev
->nr_pending
)) {
1164 /* Only remove non-faulty devices is recovery
1167 if (!test_bit(Faulty
, &rdev
->flags
) &&
1168 mddev
->degraded
< conf
->raid_disks
) {
1174 if (atomic_read(&rdev
->nr_pending
)) {
1175 /* lost the race, try later */
1187 static void end_sync_read(struct bio
*bio
, int error
)
1189 r1bio_t
* r1_bio
= (r1bio_t
*)(bio
->bi_private
);
1192 for (i
=r1_bio
->mddev
->raid_disks
; i
--; )
1193 if (r1_bio
->bios
[i
] == bio
)
1196 update_head_pos(i
, r1_bio
);
1198 * we have read a block, now it needs to be re-written,
1199 * or re-read if the read failed.
1200 * We don't do much here, just schedule handling by raid1d
1202 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
1203 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1205 if (atomic_dec_and_test(&r1_bio
->remaining
))
1206 reschedule_retry(r1_bio
);
1209 static void end_sync_write(struct bio
*bio
, int error
)
1211 int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
1212 r1bio_t
* r1_bio
= (r1bio_t
*)(bio
->bi_private
);
1213 mddev_t
*mddev
= r1_bio
->mddev
;
1214 conf_t
*conf
= mddev_to_conf(mddev
);
1218 for (i
= 0; i
< conf
->raid_disks
; i
++)
1219 if (r1_bio
->bios
[i
] == bio
) {
1224 int sync_blocks
= 0;
1225 sector_t s
= r1_bio
->sector
;
1226 long sectors_to_go
= r1_bio
->sectors
;
1227 /* make sure these bits doesn't get cleared. */
1229 bitmap_end_sync(mddev
->bitmap
, s
,
1232 sectors_to_go
-= sync_blocks
;
1233 } while (sectors_to_go
> 0);
1234 md_error(mddev
, conf
->mirrors
[mirror
].rdev
);
1237 update_head_pos(mirror
, r1_bio
);
1239 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1240 sector_t s
= r1_bio
->sectors
;
1242 md_done_sync(mddev
, s
, uptodate
);
1246 static void sync_request_write(mddev_t
*mddev
, r1bio_t
*r1_bio
)
1248 conf_t
*conf
= mddev_to_conf(mddev
);
1250 int disks
= conf
->raid_disks
;
1251 struct bio
*bio
, *wbio
;
1253 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1256 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1257 /* We have read all readable devices. If we haven't
1258 * got the block, then there is no hope left.
1259 * If we have, then we want to do a comparison
1260 * and skip the write if everything is the same.
1261 * If any blocks failed to read, then we need to
1262 * attempt an over-write
1265 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
1266 for (i
=0; i
<mddev
->raid_disks
; i
++)
1267 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_read
)
1268 md_error(mddev
, conf
->mirrors
[i
].rdev
);
1270 md_done_sync(mddev
, r1_bio
->sectors
, 1);
1274 for (primary
=0; primary
<mddev
->raid_disks
; primary
++)
1275 if (r1_bio
->bios
[primary
]->bi_end_io
== end_sync_read
&&
1276 test_bit(BIO_UPTODATE
, &r1_bio
->bios
[primary
]->bi_flags
)) {
1277 r1_bio
->bios
[primary
]->bi_end_io
= NULL
;
1278 rdev_dec_pending(conf
->mirrors
[primary
].rdev
, mddev
);
1281 r1_bio
->read_disk
= primary
;
1282 for (i
=0; i
<mddev
->raid_disks
; i
++)
1283 if (r1_bio
->bios
[i
]->bi_end_io
== end_sync_read
) {
1285 int vcnt
= r1_bio
->sectors
>> (PAGE_SHIFT
- 9);
1286 struct bio
*pbio
= r1_bio
->bios
[primary
];
1287 struct bio
*sbio
= r1_bio
->bios
[i
];
1289 if (test_bit(BIO_UPTODATE
, &sbio
->bi_flags
)) {
1290 for (j
= vcnt
; j
-- ; ) {
1292 p
= pbio
->bi_io_vec
[j
].bv_page
;
1293 s
= sbio
->bi_io_vec
[j
].bv_page
;
1294 if (memcmp(page_address(p
),
1302 mddev
->resync_mismatches
+= r1_bio
->sectors
;
1303 if (j
< 0 || (test_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
)
1304 && test_bit(BIO_UPTODATE
, &sbio
->bi_flags
))) {
1305 sbio
->bi_end_io
= NULL
;
1306 rdev_dec_pending(conf
->mirrors
[i
].rdev
, mddev
);
1308 /* fixup the bio for reuse */
1310 sbio
->bi_vcnt
= vcnt
;
1311 sbio
->bi_size
= r1_bio
->sectors
<< 9;
1313 sbio
->bi_phys_segments
= 0;
1314 sbio
->bi_flags
&= ~(BIO_POOL_MASK
- 1);
1315 sbio
->bi_flags
|= 1 << BIO_UPTODATE
;
1316 sbio
->bi_next
= NULL
;
1317 sbio
->bi_sector
= r1_bio
->sector
+
1318 conf
->mirrors
[i
].rdev
->data_offset
;
1319 sbio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1320 size
= sbio
->bi_size
;
1321 for (j
= 0; j
< vcnt
; j
++) {
1323 bi
= &sbio
->bi_io_vec
[j
];
1325 if (size
> PAGE_SIZE
)
1326 bi
->bv_len
= PAGE_SIZE
;
1330 memcpy(page_address(bi
->bv_page
),
1331 page_address(pbio
->bi_io_vec
[j
].bv_page
),
1338 if (!test_bit(R1BIO_Uptodate
, &r1_bio
->state
)) {
1339 /* ouch - failed to read all of that.
1340 * Try some synchronous reads of other devices to get
1341 * good data, much like with normal read errors. Only
1342 * read into the pages we already have so we don't
1343 * need to re-issue the read request.
1344 * We don't need to freeze the array, because being in an
1345 * active sync request, there is no normal IO, and
1346 * no overlapping syncs.
1348 sector_t sect
= r1_bio
->sector
;
1349 int sectors
= r1_bio
->sectors
;
1354 int d
= r1_bio
->read_disk
;
1358 if (s
> (PAGE_SIZE
>>9))
1361 if (r1_bio
->bios
[d
]->bi_end_io
== end_sync_read
) {
1362 /* No rcu protection needed here devices
1363 * can only be removed when no resync is
1364 * active, and resync is currently active
1366 rdev
= conf
->mirrors
[d
].rdev
;
1367 if (sync_page_io(rdev
->bdev
,
1368 sect
+ rdev
->data_offset
,
1370 bio
->bi_io_vec
[idx
].bv_page
,
1377 if (d
== conf
->raid_disks
)
1379 } while (!success
&& d
!= r1_bio
->read_disk
);
1383 /* write it back and re-read */
1384 set_bit(R1BIO_Uptodate
, &r1_bio
->state
);
1385 while (d
!= r1_bio
->read_disk
) {
1387 d
= conf
->raid_disks
;
1389 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1391 rdev
= conf
->mirrors
[d
].rdev
;
1392 atomic_add(s
, &rdev
->corrected_errors
);
1393 if (sync_page_io(rdev
->bdev
,
1394 sect
+ rdev
->data_offset
,
1396 bio
->bi_io_vec
[idx
].bv_page
,
1398 md_error(mddev
, rdev
);
1401 while (d
!= r1_bio
->read_disk
) {
1403 d
= conf
->raid_disks
;
1405 if (r1_bio
->bios
[d
]->bi_end_io
!= end_sync_read
)
1407 rdev
= conf
->mirrors
[d
].rdev
;
1408 if (sync_page_io(rdev
->bdev
,
1409 sect
+ rdev
->data_offset
,
1411 bio
->bi_io_vec
[idx
].bv_page
,
1413 md_error(mddev
, rdev
);
1416 char b
[BDEVNAME_SIZE
];
1417 /* Cannot read from anywhere, array is toast */
1418 md_error(mddev
, conf
->mirrors
[r1_bio
->read_disk
].rdev
);
1419 printk(KERN_ALERT
"raid1: %s: unrecoverable I/O read error"
1420 " for block %llu\n",
1421 bdevname(bio
->bi_bdev
,b
),
1422 (unsigned long long)r1_bio
->sector
);
1423 md_done_sync(mddev
, r1_bio
->sectors
, 0);
1436 atomic_set(&r1_bio
->remaining
, 1);
1437 for (i
= 0; i
< disks
; i
++) {
1438 wbio
= r1_bio
->bios
[i
];
1439 if (wbio
->bi_end_io
== NULL
||
1440 (wbio
->bi_end_io
== end_sync_read
&&
1441 (i
== r1_bio
->read_disk
||
1442 !test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
))))
1445 wbio
->bi_rw
= WRITE
;
1446 wbio
->bi_end_io
= end_sync_write
;
1447 atomic_inc(&r1_bio
->remaining
);
1448 md_sync_acct(conf
->mirrors
[i
].rdev
->bdev
, wbio
->bi_size
>> 9);
1450 generic_make_request(wbio
);
1453 if (atomic_dec_and_test(&r1_bio
->remaining
)) {
1454 /* if we're here, all write(s) have completed, so clean up */
1455 md_done_sync(mddev
, r1_bio
->sectors
, 1);
1461 * This is a kernel thread which:
1463 * 1. Retries failed read operations on working mirrors.
1464 * 2. Updates the raid superblock when problems encounter.
1465 * 3. Performs writes following reads for array syncronising.
1468 static void fix_read_error(conf_t
*conf
, int read_disk
,
1469 sector_t sect
, int sectors
)
1471 mddev_t
*mddev
= conf
->mddev
;
1479 if (s
> (PAGE_SIZE
>>9))
1483 /* Note: no rcu protection needed here
1484 * as this is synchronous in the raid1d thread
1485 * which is the thread that might remove
1486 * a device. If raid1d ever becomes multi-threaded....
1488 rdev
= conf
->mirrors
[d
].rdev
;
1490 test_bit(In_sync
, &rdev
->flags
) &&
1491 sync_page_io(rdev
->bdev
,
1492 sect
+ rdev
->data_offset
,
1494 conf
->tmppage
, READ
))
1498 if (d
== conf
->raid_disks
)
1501 } while (!success
&& d
!= read_disk
);
1504 /* Cannot read from anywhere -- bye bye array */
1505 md_error(mddev
, conf
->mirrors
[read_disk
].rdev
);
1508 /* write it back and re-read */
1510 while (d
!= read_disk
) {
1512 d
= conf
->raid_disks
;
1514 rdev
= conf
->mirrors
[d
].rdev
;
1516 test_bit(In_sync
, &rdev
->flags
)) {
1517 if (sync_page_io(rdev
->bdev
,
1518 sect
+ rdev
->data_offset
,
1519 s
<<9, conf
->tmppage
, WRITE
)
1521 /* Well, this device is dead */
1522 md_error(mddev
, rdev
);
1526 while (d
!= read_disk
) {
1527 char b
[BDEVNAME_SIZE
];
1529 d
= conf
->raid_disks
;
1531 rdev
= conf
->mirrors
[d
].rdev
;
1533 test_bit(In_sync
, &rdev
->flags
)) {
1534 if (sync_page_io(rdev
->bdev
,
1535 sect
+ rdev
->data_offset
,
1536 s
<<9, conf
->tmppage
, READ
)
1538 /* Well, this device is dead */
1539 md_error(mddev
, rdev
);
1541 atomic_add(s
, &rdev
->corrected_errors
);
1543 "raid1:%s: read error corrected "
1544 "(%d sectors at %llu on %s)\n",
1546 (unsigned long long)(sect
+
1548 bdevname(rdev
->bdev
, b
));
1557 static void raid1d(mddev_t
*mddev
)
1561 unsigned long flags
;
1562 conf_t
*conf
= mddev_to_conf(mddev
);
1563 struct list_head
*head
= &conf
->retry_list
;
1567 md_check_recovery(mddev
);
1570 char b
[BDEVNAME_SIZE
];
1572 unplug
+= flush_pending_writes(conf
);
1574 spin_lock_irqsave(&conf
->device_lock
, flags
);
1575 if (list_empty(head
)) {
1576 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1579 r1_bio
= list_entry(head
->prev
, r1bio_t
, retry_list
);
1580 list_del(head
->prev
);
1582 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1584 mddev
= r1_bio
->mddev
;
1585 conf
= mddev_to_conf(mddev
);
1586 if (test_bit(R1BIO_IsSync
, &r1_bio
->state
)) {
1587 sync_request_write(mddev
, r1_bio
);
1589 } else if (test_bit(R1BIO_BarrierRetry
, &r1_bio
->state
)) {
1590 /* some requests in the r1bio were BIO_RW_BARRIER
1591 * requests which failed with -EOPNOTSUPP. Hohumm..
1592 * Better resubmit without the barrier.
1593 * We know which devices to resubmit for, because
1594 * all others have had their bios[] entry cleared.
1595 * We already have a nr_pending reference on these rdevs.
1598 const int do_sync
= bio_sync(r1_bio
->master_bio
);
1599 clear_bit(R1BIO_BarrierRetry
, &r1_bio
->state
);
1600 clear_bit(R1BIO_Barrier
, &r1_bio
->state
);
1601 for (i
=0; i
< conf
->raid_disks
; i
++)
1602 if (r1_bio
->bios
[i
])
1603 atomic_inc(&r1_bio
->remaining
);
1604 for (i
=0; i
< conf
->raid_disks
; i
++)
1605 if (r1_bio
->bios
[i
]) {
1606 struct bio_vec
*bvec
;
1609 bio
= bio_clone(r1_bio
->master_bio
, GFP_NOIO
);
1610 /* copy pages from the failed bio, as
1611 * this might be a write-behind device */
1612 __bio_for_each_segment(bvec
, bio
, j
, 0)
1613 bvec
->bv_page
= bio_iovec_idx(r1_bio
->bios
[i
], j
)->bv_page
;
1614 bio_put(r1_bio
->bios
[i
]);
1615 bio
->bi_sector
= r1_bio
->sector
+
1616 conf
->mirrors
[i
].rdev
->data_offset
;
1617 bio
->bi_bdev
= conf
->mirrors
[i
].rdev
->bdev
;
1618 bio
->bi_end_io
= raid1_end_write_request
;
1619 bio
->bi_rw
= WRITE
| do_sync
;
1620 bio
->bi_private
= r1_bio
;
1621 r1_bio
->bios
[i
] = bio
;
1622 generic_make_request(bio
);
1627 /* we got a read error. Maybe the drive is bad. Maybe just
1628 * the block and we can fix it.
1629 * We freeze all other IO, and try reading the block from
1630 * other devices. When we find one, we re-write
1631 * and check it that fixes the read error.
1632 * This is all done synchronously while the array is
1635 if (mddev
->ro
== 0) {
1637 fix_read_error(conf
, r1_bio
->read_disk
,
1640 unfreeze_array(conf
);
1643 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1644 if ((disk
=read_balance(conf
, r1_bio
)) == -1 ||
1645 disk
== r1_bio
->read_disk
) {
1646 printk(KERN_ALERT
"raid1: %s: unrecoverable I/O"
1647 " read error for block %llu\n",
1648 bdevname(bio
->bi_bdev
,b
),
1649 (unsigned long long)r1_bio
->sector
);
1650 raid_end_bio_io(r1_bio
);
1652 const int do_sync
= bio_sync(r1_bio
->master_bio
);
1653 r1_bio
->bios
[r1_bio
->read_disk
] =
1654 mddev
->ro
? IO_BLOCKED
: NULL
;
1655 r1_bio
->read_disk
= disk
;
1657 bio
= bio_clone(r1_bio
->master_bio
, GFP_NOIO
);
1658 r1_bio
->bios
[r1_bio
->read_disk
] = bio
;
1659 rdev
= conf
->mirrors
[disk
].rdev
;
1660 if (printk_ratelimit())
1661 printk(KERN_ERR
"raid1: %s: redirecting sector %llu to"
1662 " another mirror\n",
1663 bdevname(rdev
->bdev
,b
),
1664 (unsigned long long)r1_bio
->sector
);
1665 bio
->bi_sector
= r1_bio
->sector
+ rdev
->data_offset
;
1666 bio
->bi_bdev
= rdev
->bdev
;
1667 bio
->bi_end_io
= raid1_end_read_request
;
1668 bio
->bi_rw
= READ
| do_sync
;
1669 bio
->bi_private
= r1_bio
;
1671 generic_make_request(bio
);
1676 unplug_slaves(mddev
);
1680 static int init_resync(conf_t
*conf
)
1684 buffs
= RESYNC_WINDOW
/ RESYNC_BLOCK_SIZE
;
1685 BUG_ON(conf
->r1buf_pool
);
1686 conf
->r1buf_pool
= mempool_create(buffs
, r1buf_pool_alloc
, r1buf_pool_free
,
1688 if (!conf
->r1buf_pool
)
1690 conf
->next_resync
= 0;
1695 * perform a "sync" on one "block"
1697 * We need to make sure that no normal I/O request - particularly write
1698 * requests - conflict with active sync requests.
1700 * This is achieved by tracking pending requests and a 'barrier' concept
1701 * that can be installed to exclude normal IO requests.
1704 static sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
1706 conf_t
*conf
= mddev_to_conf(mddev
);
1709 sector_t max_sector
, nr_sectors
;
1713 int write_targets
= 0, read_targets
= 0;
1715 int still_degraded
= 0;
1717 if (!conf
->r1buf_pool
)
1720 printk("sync start - bitmap %p\n", mddev->bitmap);
1722 if (init_resync(conf
))
1726 max_sector
= mddev
->size
<< 1;
1727 if (sector_nr
>= max_sector
) {
1728 /* If we aborted, we need to abort the
1729 * sync on the 'current' bitmap chunk (there will
1730 * only be one in raid1 resync.
1731 * We can find the current addess in mddev->curr_resync
1733 if (mddev
->curr_resync
< max_sector
) /* aborted */
1734 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
1736 else /* completed sync */
1739 bitmap_close_sync(mddev
->bitmap
);
1744 if (mddev
->bitmap
== NULL
&&
1745 mddev
->recovery_cp
== MaxSector
&&
1746 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
1747 conf
->fullsync
== 0) {
1749 return max_sector
- sector_nr
;
1751 /* before building a request, check if we can skip these blocks..
1752 * This call the bitmap_start_sync doesn't actually record anything
1754 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
1755 !conf
->fullsync
&& !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1756 /* We can skip this block, and probably several more */
1761 * If there is non-resync activity waiting for a turn,
1762 * and resync is going fast enough,
1763 * then let it though before starting on this new sync request.
1765 if (!go_faster
&& conf
->nr_waiting
)
1766 msleep_interruptible(1000);
1768 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
1769 raise_barrier(conf
);
1771 conf
->next_resync
= sector_nr
;
1773 r1_bio
= mempool_alloc(conf
->r1buf_pool
, GFP_NOIO
);
1776 * If we get a correctably read error during resync or recovery,
1777 * we might want to read from a different device. So we
1778 * flag all drives that could conceivably be read from for READ,
1779 * and any others (which will be non-In_sync devices) for WRITE.
1780 * If a read fails, we try reading from something else for which READ
1784 r1_bio
->mddev
= mddev
;
1785 r1_bio
->sector
= sector_nr
;
1787 set_bit(R1BIO_IsSync
, &r1_bio
->state
);
1789 for (i
=0; i
< conf
->raid_disks
; i
++) {
1791 bio
= r1_bio
->bios
[i
];
1793 /* take from bio_init */
1794 bio
->bi_next
= NULL
;
1795 bio
->bi_flags
|= 1 << BIO_UPTODATE
;
1799 bio
->bi_phys_segments
= 0;
1801 bio
->bi_end_io
= NULL
;
1802 bio
->bi_private
= NULL
;
1804 rdev
= rcu_dereference(conf
->mirrors
[i
].rdev
);
1806 test_bit(Faulty
, &rdev
->flags
)) {
1809 } else if (!test_bit(In_sync
, &rdev
->flags
)) {
1811 bio
->bi_end_io
= end_sync_write
;
1814 /* may need to read from here */
1816 bio
->bi_end_io
= end_sync_read
;
1817 if (test_bit(WriteMostly
, &rdev
->flags
)) {
1826 atomic_inc(&rdev
->nr_pending
);
1827 bio
->bi_sector
= sector_nr
+ rdev
->data_offset
;
1828 bio
->bi_bdev
= rdev
->bdev
;
1829 bio
->bi_private
= r1_bio
;
1834 r1_bio
->read_disk
= disk
;
1836 if (test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
) && read_targets
> 0)
1837 /* extra read targets are also write targets */
1838 write_targets
+= read_targets
-1;
1840 if (write_targets
== 0 || read_targets
== 0) {
1841 /* There is nowhere to write, so all non-sync
1842 * drives must be failed - so we are finished
1844 sector_t rv
= max_sector
- sector_nr
;
1850 if (max_sector
> mddev
->resync_max
)
1851 max_sector
= mddev
->resync_max
; /* Don't do IO beyond here */
1856 int len
= PAGE_SIZE
;
1857 if (sector_nr
+ (len
>>9) > max_sector
)
1858 len
= (max_sector
- sector_nr
) << 9;
1861 if (sync_blocks
== 0) {
1862 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
,
1863 &sync_blocks
, still_degraded
) &&
1865 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
))
1867 BUG_ON(sync_blocks
< (PAGE_SIZE
>>9));
1868 if (len
> (sync_blocks
<<9))
1869 len
= sync_blocks
<<9;
1872 for (i
=0 ; i
< conf
->raid_disks
; i
++) {
1873 bio
= r1_bio
->bios
[i
];
1874 if (bio
->bi_end_io
) {
1875 page
= bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
;
1876 if (bio_add_page(bio
, page
, len
, 0) == 0) {
1878 bio
->bi_io_vec
[bio
->bi_vcnt
].bv_page
= page
;
1881 bio
= r1_bio
->bios
[i
];
1882 if (bio
->bi_end_io
==NULL
)
1884 /* remove last page from this bio */
1886 bio
->bi_size
-= len
;
1887 bio
->bi_flags
&= ~(1<< BIO_SEG_VALID
);
1893 nr_sectors
+= len
>>9;
1894 sector_nr
+= len
>>9;
1895 sync_blocks
-= (len
>>9);
1896 } while (r1_bio
->bios
[disk
]->bi_vcnt
< RESYNC_PAGES
);
1898 r1_bio
->sectors
= nr_sectors
;
1900 /* For a user-requested sync, we read all readable devices and do a
1903 if (test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
)) {
1904 atomic_set(&r1_bio
->remaining
, read_targets
);
1905 for (i
=0; i
<conf
->raid_disks
; i
++) {
1906 bio
= r1_bio
->bios
[i
];
1907 if (bio
->bi_end_io
== end_sync_read
) {
1908 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1909 generic_make_request(bio
);
1913 atomic_set(&r1_bio
->remaining
, 1);
1914 bio
= r1_bio
->bios
[r1_bio
->read_disk
];
1915 md_sync_acct(bio
->bi_bdev
, nr_sectors
);
1916 generic_make_request(bio
);
1922 static int run(mddev_t
*mddev
)
1926 mirror_info_t
*disk
;
1929 if (mddev
->level
!= 1) {
1930 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1931 mdname(mddev
), mddev
->level
);
1934 if (mddev
->reshape_position
!= MaxSector
) {
1935 printk("raid1: %s: reshape_position set but not supported\n",
1940 * copy the already verified devices into our private RAID1
1941 * bookkeeping area. [whatever we allocate in run(),
1942 * should be freed in stop()]
1944 conf
= kzalloc(sizeof(conf_t
), GFP_KERNEL
);
1945 mddev
->private = conf
;
1949 conf
->mirrors
= kzalloc(sizeof(struct mirror_info
)*mddev
->raid_disks
,
1954 conf
->tmppage
= alloc_page(GFP_KERNEL
);
1958 conf
->poolinfo
= kmalloc(sizeof(*conf
->poolinfo
), GFP_KERNEL
);
1959 if (!conf
->poolinfo
)
1961 conf
->poolinfo
->mddev
= mddev
;
1962 conf
->poolinfo
->raid_disks
= mddev
->raid_disks
;
1963 conf
->r1bio_pool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
1966 if (!conf
->r1bio_pool
)
1969 spin_lock_init(&conf
->device_lock
);
1970 mddev
->queue
->queue_lock
= &conf
->device_lock
;
1972 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
1973 disk_idx
= rdev
->raid_disk
;
1974 if (disk_idx
>= mddev
->raid_disks
1977 disk
= conf
->mirrors
+ disk_idx
;
1981 blk_queue_stack_limits(mddev
->queue
,
1982 rdev
->bdev
->bd_disk
->queue
);
1983 /* as we don't honour merge_bvec_fn, we must never risk
1984 * violating it, so limit ->max_sector to one PAGE, as
1985 * a one page request is never in violation.
1987 if (rdev
->bdev
->bd_disk
->queue
->merge_bvec_fn
&&
1988 mddev
->queue
->max_sectors
> (PAGE_SIZE
>>9))
1989 blk_queue_max_sectors(mddev
->queue
, PAGE_SIZE
>>9);
1991 disk
->head_position
= 0;
1993 conf
->raid_disks
= mddev
->raid_disks
;
1994 conf
->mddev
= mddev
;
1995 INIT_LIST_HEAD(&conf
->retry_list
);
1997 spin_lock_init(&conf
->resync_lock
);
1998 init_waitqueue_head(&conf
->wait_barrier
);
2000 bio_list_init(&conf
->pending_bio_list
);
2001 bio_list_init(&conf
->flushing_bio_list
);
2004 mddev
->degraded
= 0;
2005 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2007 disk
= conf
->mirrors
+ i
;
2010 !test_bit(In_sync
, &disk
->rdev
->flags
)) {
2011 disk
->head_position
= 0;
2017 if (mddev
->degraded
== conf
->raid_disks
) {
2018 printk(KERN_ERR
"raid1: no operational mirrors for %s\n",
2022 if (conf
->raid_disks
- mddev
->degraded
== 1)
2023 mddev
->recovery_cp
= MaxSector
;
2026 * find the first working one and use it as a starting point
2027 * to read balancing.
2029 for (j
= 0; j
< conf
->raid_disks
&&
2030 (!conf
->mirrors
[j
].rdev
||
2031 !test_bit(In_sync
, &conf
->mirrors
[j
].rdev
->flags
)) ; j
++)
2033 conf
->last_used
= j
;
2036 mddev
->thread
= md_register_thread(raid1d
, mddev
, "%s_raid1");
2037 if (!mddev
->thread
) {
2039 "raid1: couldn't allocate thread for %s\n",
2045 "raid1: raid set %s active with %d out of %d mirrors\n",
2046 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
2049 * Ok, everything is just fine now
2051 mddev
->array_sectors
= mddev
->size
* 2;
2053 mddev
->queue
->unplug_fn
= raid1_unplug
;
2054 mddev
->queue
->backing_dev_info
.congested_fn
= raid1_congested
;
2055 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
2060 printk(KERN_ERR
"raid1: couldn't allocate memory for %s\n",
2065 if (conf
->r1bio_pool
)
2066 mempool_destroy(conf
->r1bio_pool
);
2067 kfree(conf
->mirrors
);
2068 safe_put_page(conf
->tmppage
);
2069 kfree(conf
->poolinfo
);
2071 mddev
->private = NULL
;
2077 static int stop(mddev_t
*mddev
)
2079 conf_t
*conf
= mddev_to_conf(mddev
);
2080 struct bitmap
*bitmap
= mddev
->bitmap
;
2081 int behind_wait
= 0;
2083 /* wait for behind writes to complete */
2084 while (bitmap
&& atomic_read(&bitmap
->behind_writes
) > 0) {
2086 printk(KERN_INFO
"raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev
), behind_wait
);
2087 set_current_state(TASK_UNINTERRUPTIBLE
);
2088 schedule_timeout(HZ
); /* wait a second */
2089 /* need to kick something here to make sure I/O goes? */
2092 md_unregister_thread(mddev
->thread
);
2093 mddev
->thread
= NULL
;
2094 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
2095 if (conf
->r1bio_pool
)
2096 mempool_destroy(conf
->r1bio_pool
);
2097 kfree(conf
->mirrors
);
2098 kfree(conf
->poolinfo
);
2100 mddev
->private = NULL
;
2104 static int raid1_resize(mddev_t
*mddev
, sector_t sectors
)
2106 /* no resync is happening, and there is enough space
2107 * on all devices, so we can resize.
2108 * We need to make sure resync covers any new space.
2109 * If the array is shrinking we should possibly wait until
2110 * any io in the removed space completes, but it hardly seems
2113 mddev
->array_sectors
= sectors
;
2114 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
2116 if (mddev
->array_sectors
/ 2 > mddev
->size
&&
2117 mddev
->recovery_cp
== MaxSector
) {
2118 mddev
->recovery_cp
= mddev
->size
<< 1;
2119 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2121 mddev
->size
= mddev
->array_sectors
/ 2;
2122 mddev
->resync_max_sectors
= sectors
;
2126 static int raid1_reshape(mddev_t
*mddev
)
2129 * 1/ resize the r1bio_pool
2130 * 2/ resize conf->mirrors
2132 * We allocate a new r1bio_pool if we can.
2133 * Then raise a device barrier and wait until all IO stops.
2134 * Then resize conf->mirrors and swap in the new r1bio pool.
2136 * At the same time, we "pack" the devices so that all the missing
2137 * devices have the higher raid_disk numbers.
2139 mempool_t
*newpool
, *oldpool
;
2140 struct pool_info
*newpoolinfo
;
2141 mirror_info_t
*newmirrors
;
2142 conf_t
*conf
= mddev_to_conf(mddev
);
2143 int cnt
, raid_disks
;
2144 unsigned long flags
;
2147 /* Cannot change chunk_size, layout, or level */
2148 if (mddev
->chunk_size
!= mddev
->new_chunk
||
2149 mddev
->layout
!= mddev
->new_layout
||
2150 mddev
->level
!= mddev
->new_level
) {
2151 mddev
->new_chunk
= mddev
->chunk_size
;
2152 mddev
->new_layout
= mddev
->layout
;
2153 mddev
->new_level
= mddev
->level
;
2157 err
= md_allow_write(mddev
);
2161 raid_disks
= mddev
->raid_disks
+ mddev
->delta_disks
;
2163 if (raid_disks
< conf
->raid_disks
) {
2165 for (d
= 0; d
< conf
->raid_disks
; d
++)
2166 if (conf
->mirrors
[d
].rdev
)
2168 if (cnt
> raid_disks
)
2172 newpoolinfo
= kmalloc(sizeof(*newpoolinfo
), GFP_KERNEL
);
2175 newpoolinfo
->mddev
= mddev
;
2176 newpoolinfo
->raid_disks
= raid_disks
;
2178 newpool
= mempool_create(NR_RAID1_BIOS
, r1bio_pool_alloc
,
2179 r1bio_pool_free
, newpoolinfo
);
2184 newmirrors
= kzalloc(sizeof(struct mirror_info
) * raid_disks
, GFP_KERNEL
);
2187 mempool_destroy(newpool
);
2191 raise_barrier(conf
);
2193 /* ok, everything is stopped */
2194 oldpool
= conf
->r1bio_pool
;
2195 conf
->r1bio_pool
= newpool
;
2197 for (d
= d2
= 0; d
< conf
->raid_disks
; d
++) {
2198 mdk_rdev_t
*rdev
= conf
->mirrors
[d
].rdev
;
2199 if (rdev
&& rdev
->raid_disk
!= d2
) {
2201 sprintf(nm
, "rd%d", rdev
->raid_disk
);
2202 sysfs_remove_link(&mddev
->kobj
, nm
);
2203 rdev
->raid_disk
= d2
;
2204 sprintf(nm
, "rd%d", rdev
->raid_disk
);
2205 sysfs_remove_link(&mddev
->kobj
, nm
);
2206 if (sysfs_create_link(&mddev
->kobj
,
2209 "md/raid1: cannot register "
2214 newmirrors
[d2
++].rdev
= rdev
;
2216 kfree(conf
->mirrors
);
2217 conf
->mirrors
= newmirrors
;
2218 kfree(conf
->poolinfo
);
2219 conf
->poolinfo
= newpoolinfo
;
2221 spin_lock_irqsave(&conf
->device_lock
, flags
);
2222 mddev
->degraded
+= (raid_disks
- conf
->raid_disks
);
2223 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2224 conf
->raid_disks
= mddev
->raid_disks
= raid_disks
;
2225 mddev
->delta_disks
= 0;
2227 conf
->last_used
= 0; /* just make sure it is in-range */
2228 lower_barrier(conf
);
2230 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
2231 md_wakeup_thread(mddev
->thread
);
2233 mempool_destroy(oldpool
);
2237 static void raid1_quiesce(mddev_t
*mddev
, int state
)
2239 conf_t
*conf
= mddev_to_conf(mddev
);
2243 raise_barrier(conf
);
2246 lower_barrier(conf
);
2252 static struct mdk_personality raid1_personality
=
2256 .owner
= THIS_MODULE
,
2257 .make_request
= make_request
,
2261 .error_handler
= error
,
2262 .hot_add_disk
= raid1_add_disk
,
2263 .hot_remove_disk
= raid1_remove_disk
,
2264 .spare_active
= raid1_spare_active
,
2265 .sync_request
= sync_request
,
2266 .resize
= raid1_resize
,
2267 .check_reshape
= raid1_reshape
,
2268 .quiesce
= raid1_quiesce
,
2271 static int __init
raid_init(void)
2273 return register_md_personality(&raid1_personality
);
2276 static void raid_exit(void)
2278 unregister_md_personality(&raid1_personality
);
2281 module_init(raid_init
);
2282 module_exit(raid_exit
);
2283 MODULE_LICENSE("GPL");
2284 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2285 MODULE_ALIAS("md-raid1");
2286 MODULE_ALIAS("md-level-1");