2 * raid1.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * RAID-1 management functions.
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
45 * Number of guaranteed r1bios in case of extreme VM load:
47 #define NR_RAID1_BIOS 256
49 /* when we get a read error on a read-only array, we redirect to another
50 * device without failing the first device, or trying to over-write to
51 * correct the read error. To keep track of bad blocks on a per-bio
52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
54 #define IO_BLOCKED ((struct bio *)1)
55 /* When we successfully write to a known bad-block, we need to remove the
56 * bad-block marking which must be done from process context. So we record
57 * the success by setting devs[n].bio to IO_MADE_GOOD
59 #define IO_MADE_GOOD ((struct bio *)2)
61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
63 /* When there are this many requests queue to be written by
64 * the raid1 thread, we become 'congested' to provide back-pressure
67 static int max_queued_requests = 1024;
69 static void allow_barrier(struct r1conf *conf);
70 static void lower_barrier(struct r1conf *conf);
72 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
74 struct pool_info *pi = data;
75 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
77 /* allocate a r1bio with room for raid_disks entries in the bios array */
78 return kzalloc(size, gfp_flags);
81 static void r1bio_pool_free(void *r1_bio, void *data)
86 #define RESYNC_BLOCK_SIZE (64*1024)
87 #define RESYNC_DEPTH 32
88 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
90 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
91 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
92 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
94 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
96 struct pool_info *pi = data;
101 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
106 * Allocate bios : 1 for reading, n-1 for writing
108 for (j = pi->raid_disks ; j-- ; ) {
109 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
112 r1_bio->bios[j] = bio;
115 * Allocate RESYNC_PAGES data pages and attach them to
117 * If this is a user-requested check/repair, allocate
118 * RESYNC_PAGES for each bio.
120 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
125 bio = r1_bio->bios[j];
126 bio->bi_vcnt = RESYNC_PAGES;
128 if (bio_alloc_pages(bio, gfp_flags))
131 /* If not user-requests, copy the page pointers to all bios */
132 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
133 for (i=0; i<RESYNC_PAGES ; i++)
134 for (j=1; j<pi->raid_disks; j++)
135 r1_bio->bios[j]->bi_io_vec[i].bv_page =
136 r1_bio->bios[0]->bi_io_vec[i].bv_page;
139 r1_bio->master_bio = NULL;
144 while (++j < pi->raid_disks)
145 bio_put(r1_bio->bios[j]);
146 r1bio_pool_free(r1_bio, data);
150 static void r1buf_pool_free(void *__r1_bio, void *data)
152 struct pool_info *pi = data;
154 struct r1bio *r1bio = __r1_bio;
156 for (i = 0; i < RESYNC_PAGES; i++)
157 for (j = pi->raid_disks; j-- ;) {
159 r1bio->bios[j]->bi_io_vec[i].bv_page !=
160 r1bio->bios[0]->bi_io_vec[i].bv_page)
161 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
163 for (i=0 ; i < pi->raid_disks; i++)
164 bio_put(r1bio->bios[i]);
166 r1bio_pool_free(r1bio, data);
169 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
173 for (i = 0; i < conf->raid_disks * 2; i++) {
174 struct bio **bio = r1_bio->bios + i;
175 if (!BIO_SPECIAL(*bio))
181 static void free_r1bio(struct r1bio *r1_bio)
183 struct r1conf *conf = r1_bio->mddev->private;
185 put_all_bios(conf, r1_bio);
186 mempool_free(r1_bio, conf->r1bio_pool);
189 static void put_buf(struct r1bio *r1_bio)
191 struct r1conf *conf = r1_bio->mddev->private;
194 for (i = 0; i < conf->raid_disks * 2; 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(struct r1bio *r1_bio)
208 struct mddev *mddev = r1_bio->mddev;
209 struct r1conf *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 call_bio_endio(struct r1bio *r1_bio)
227 struct bio *bio = r1_bio->master_bio;
229 struct r1conf *conf = r1_bio->mddev->private;
231 if (bio->bi_phys_segments) {
233 spin_lock_irqsave(&conf->device_lock, flags);
234 bio->bi_phys_segments--;
235 done = (bio->bi_phys_segments == 0);
236 spin_unlock_irqrestore(&conf->device_lock, flags);
240 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
241 clear_bit(BIO_UPTODATE, &bio->bi_flags);
245 * Wake up any possible resync thread that waits for the device
252 static void raid_end_bio_io(struct r1bio *r1_bio)
254 struct bio *bio = r1_bio->master_bio;
256 /* if nobody has done the final endio yet, do it now */
257 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
258 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
259 (bio_data_dir(bio) == WRITE) ? "write" : "read",
260 (unsigned long long) bio->bi_sector,
261 (unsigned long long) bio->bi_sector +
262 bio_sectors(bio) - 1);
264 call_bio_endio(r1_bio);
270 * Update disk head position estimator based on IRQ completion info.
272 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
274 struct r1conf *conf = r1_bio->mddev->private;
276 conf->mirrors[disk].head_position =
277 r1_bio->sector + (r1_bio->sectors);
281 * Find the disk number which triggered given bio
283 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
286 struct r1conf *conf = r1_bio->mddev->private;
287 int raid_disks = conf->raid_disks;
289 for (mirror = 0; mirror < raid_disks * 2; mirror++)
290 if (r1_bio->bios[mirror] == bio)
293 BUG_ON(mirror == raid_disks * 2);
294 update_head_pos(mirror, r1_bio);
299 static void raid1_end_read_request(struct bio *bio, int error)
301 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
302 struct r1bio *r1_bio = bio->bi_private;
304 struct r1conf *conf = r1_bio->mddev->private;
306 mirror = r1_bio->read_disk;
308 * this branch is our 'one mirror IO has finished' event handler:
310 update_head_pos(mirror, r1_bio);
313 set_bit(R1BIO_Uptodate, &r1_bio->state);
315 /* If all other devices have failed, we want to return
316 * the error upwards rather than fail the last device.
317 * Here we redefine "uptodate" to mean "Don't want to retry"
320 spin_lock_irqsave(&conf->device_lock, flags);
321 if (r1_bio->mddev->degraded == conf->raid_disks ||
322 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
323 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
325 spin_unlock_irqrestore(&conf->device_lock, flags);
329 raid_end_bio_io(r1_bio);
330 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
335 char b[BDEVNAME_SIZE];
337 KERN_ERR "md/raid1:%s: %s: "
338 "rescheduling sector %llu\n",
340 bdevname(conf->mirrors[mirror].rdev->bdev,
342 (unsigned long long)r1_bio->sector);
343 set_bit(R1BIO_ReadError, &r1_bio->state);
344 reschedule_retry(r1_bio);
345 /* don't drop the reference on read_disk yet */
349 static void close_write(struct r1bio *r1_bio)
351 /* it really is the end of this request */
352 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
353 /* free extra copy of the data pages */
354 int i = r1_bio->behind_page_count;
356 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
357 kfree(r1_bio->behind_bvecs);
358 r1_bio->behind_bvecs = NULL;
360 /* clear the bitmap if all writes complete successfully */
361 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
363 !test_bit(R1BIO_Degraded, &r1_bio->state),
364 test_bit(R1BIO_BehindIO, &r1_bio->state));
365 md_write_end(r1_bio->mddev);
368 static void r1_bio_write_done(struct r1bio *r1_bio)
370 if (!atomic_dec_and_test(&r1_bio->remaining))
373 if (test_bit(R1BIO_WriteError, &r1_bio->state))
374 reschedule_retry(r1_bio);
377 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
378 reschedule_retry(r1_bio);
380 raid_end_bio_io(r1_bio);
384 static void raid1_end_write_request(struct bio *bio, int error)
386 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
387 struct r1bio *r1_bio = bio->bi_private;
388 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
389 struct r1conf *conf = r1_bio->mddev->private;
390 struct bio *to_put = NULL;
392 mirror = find_bio_disk(r1_bio, bio);
395 * 'one mirror IO has finished' event handler:
398 set_bit(WriteErrorSeen,
399 &conf->mirrors[mirror].rdev->flags);
400 if (!test_and_set_bit(WantReplacement,
401 &conf->mirrors[mirror].rdev->flags))
402 set_bit(MD_RECOVERY_NEEDED, &
403 conf->mddev->recovery);
405 set_bit(R1BIO_WriteError, &r1_bio->state);
408 * Set R1BIO_Uptodate in our master bio, so that we
409 * will return a good error code for to the higher
410 * levels even if IO on some other mirrored buffer
413 * The 'master' represents the composite IO operation
414 * to user-side. So if something waits for IO, then it
415 * will wait for the 'master' bio.
420 r1_bio->bios[mirror] = NULL;
423 * Do not set R1BIO_Uptodate if the current device is
424 * rebuilding or Faulty. This is because we cannot use
425 * such device for properly reading the data back (we could
426 * potentially use it, if the current write would have felt
427 * before rdev->recovery_offset, but for simplicity we don't
430 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
431 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
432 set_bit(R1BIO_Uptodate, &r1_bio->state);
434 /* Maybe we can clear some bad blocks. */
435 if (is_badblock(conf->mirrors[mirror].rdev,
436 r1_bio->sector, r1_bio->sectors,
437 &first_bad, &bad_sectors)) {
438 r1_bio->bios[mirror] = IO_MADE_GOOD;
439 set_bit(R1BIO_MadeGood, &r1_bio->state);
444 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
445 atomic_dec(&r1_bio->behind_remaining);
448 * In behind mode, we ACK the master bio once the I/O
449 * has safely reached all non-writemostly
450 * disks. Setting the Returned bit ensures that this
451 * gets done only once -- we don't ever want to return
452 * -EIO here, instead we'll wait
454 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
455 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
456 /* Maybe we can return now */
457 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
458 struct bio *mbio = r1_bio->master_bio;
459 pr_debug("raid1: behind end write sectors"
461 (unsigned long long) mbio->bi_sector,
462 (unsigned long long) mbio->bi_sector +
463 bio_sectors(mbio) - 1);
464 call_bio_endio(r1_bio);
468 if (r1_bio->bios[mirror] == NULL)
469 rdev_dec_pending(conf->mirrors[mirror].rdev,
473 * Let's see if all mirrored write operations have finished
476 r1_bio_write_done(r1_bio);
484 * This routine returns the disk from which the requested read should
485 * be done. There is a per-array 'next expected sequential IO' sector
486 * number - if this matches on the next IO then we use the last disk.
487 * There is also a per-disk 'last know head position' sector that is
488 * maintained from IRQ contexts, both the normal and the resync IO
489 * completion handlers update this position correctly. If there is no
490 * perfect sequential match then we pick the disk whose head is closest.
492 * If there are 2 mirrors in the same 2 devices, performance degrades
493 * because position is mirror, not device based.
495 * The rdev for the device selected will have nr_pending incremented.
497 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
499 const sector_t this_sector = r1_bio->sector;
501 int best_good_sectors;
502 int best_disk, best_dist_disk, best_pending_disk;
506 unsigned int min_pending;
507 struct md_rdev *rdev;
509 int choose_next_idle;
513 * Check if we can balance. We can balance on the whole
514 * device if no resync is going on, or below the resync window.
515 * We take the first readable disk when above the resync window.
518 sectors = r1_bio->sectors;
521 best_dist = MaxSector;
522 best_pending_disk = -1;
523 min_pending = UINT_MAX;
524 best_good_sectors = 0;
526 choose_next_idle = 0;
528 if (conf->mddev->recovery_cp < MaxSector &&
529 (this_sector + sectors >= conf->next_resync))
534 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
538 unsigned int pending;
541 rdev = rcu_dereference(conf->mirrors[disk].rdev);
542 if (r1_bio->bios[disk] == IO_BLOCKED
544 || test_bit(Unmerged, &rdev->flags)
545 || test_bit(Faulty, &rdev->flags))
547 if (!test_bit(In_sync, &rdev->flags) &&
548 rdev->recovery_offset < this_sector + sectors)
550 if (test_bit(WriteMostly, &rdev->flags)) {
551 /* Don't balance among write-mostly, just
552 * use the first as a last resort */
554 if (is_badblock(rdev, this_sector, sectors,
555 &first_bad, &bad_sectors)) {
556 if (first_bad < this_sector)
557 /* Cannot use this */
559 best_good_sectors = first_bad - this_sector;
561 best_good_sectors = sectors;
566 /* This is a reasonable device to use. It might
569 if (is_badblock(rdev, this_sector, sectors,
570 &first_bad, &bad_sectors)) {
571 if (best_dist < MaxSector)
572 /* already have a better device */
574 if (first_bad <= this_sector) {
575 /* cannot read here. If this is the 'primary'
576 * device, then we must not read beyond
577 * bad_sectors from another device..
579 bad_sectors -= (this_sector - first_bad);
580 if (choose_first && sectors > bad_sectors)
581 sectors = bad_sectors;
582 if (best_good_sectors > sectors)
583 best_good_sectors = sectors;
586 sector_t good_sectors = first_bad - this_sector;
587 if (good_sectors > best_good_sectors) {
588 best_good_sectors = good_sectors;
596 best_good_sectors = sectors;
598 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
599 has_nonrot_disk |= nonrot;
600 pending = atomic_read(&rdev->nr_pending);
601 dist = abs(this_sector - conf->mirrors[disk].head_position);
606 /* Don't change to another disk for sequential reads */
607 if (conf->mirrors[disk].next_seq_sect == this_sector
609 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
610 struct raid1_info *mirror = &conf->mirrors[disk];
614 * If buffered sequential IO size exceeds optimal
615 * iosize, check if there is idle disk. If yes, choose
616 * the idle disk. read_balance could already choose an
617 * idle disk before noticing it's a sequential IO in
618 * this disk. This doesn't matter because this disk
619 * will idle, next time it will be utilized after the
620 * first disk has IO size exceeds optimal iosize. In
621 * this way, iosize of the first disk will be optimal
622 * iosize at least. iosize of the second disk might be
623 * small, but not a big deal since when the second disk
624 * starts IO, the first disk is likely still busy.
626 if (nonrot && opt_iosize > 0 &&
627 mirror->seq_start != MaxSector &&
628 mirror->next_seq_sect > opt_iosize &&
629 mirror->next_seq_sect - opt_iosize >=
631 choose_next_idle = 1;
636 /* If device is idle, use it */
642 if (choose_next_idle)
645 if (min_pending > pending) {
646 min_pending = pending;
647 best_pending_disk = disk;
650 if (dist < best_dist) {
652 best_dist_disk = disk;
657 * If all disks are rotational, choose the closest disk. If any disk is
658 * non-rotational, choose the disk with less pending request even the
659 * disk is rotational, which might/might not be optimal for raids with
660 * mixed ratation/non-rotational disks depending on workload.
662 if (best_disk == -1) {
664 best_disk = best_pending_disk;
666 best_disk = best_dist_disk;
669 if (best_disk >= 0) {
670 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
673 atomic_inc(&rdev->nr_pending);
674 if (test_bit(Faulty, &rdev->flags)) {
675 /* cannot risk returning a device that failed
676 * before we inc'ed nr_pending
678 rdev_dec_pending(rdev, conf->mddev);
681 sectors = best_good_sectors;
683 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
684 conf->mirrors[best_disk].seq_start = this_sector;
686 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
689 *max_sectors = sectors;
694 static int raid1_mergeable_bvec(struct request_queue *q,
695 struct bvec_merge_data *bvm,
696 struct bio_vec *biovec)
698 struct mddev *mddev = q->queuedata;
699 struct r1conf *conf = mddev->private;
700 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
701 int max = biovec->bv_len;
703 if (mddev->merge_check_needed) {
706 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
707 struct md_rdev *rdev = rcu_dereference(
708 conf->mirrors[disk].rdev);
709 if (rdev && !test_bit(Faulty, &rdev->flags)) {
710 struct request_queue *q =
711 bdev_get_queue(rdev->bdev);
712 if (q->merge_bvec_fn) {
713 bvm->bi_sector = sector +
715 bvm->bi_bdev = rdev->bdev;
716 max = min(max, q->merge_bvec_fn(
727 int md_raid1_congested(struct mddev *mddev, int bits)
729 struct r1conf *conf = mddev->private;
732 if ((bits & (1 << BDI_async_congested)) &&
733 conf->pending_count >= max_queued_requests)
737 for (i = 0; i < conf->raid_disks * 2; i++) {
738 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
739 if (rdev && !test_bit(Faulty, &rdev->flags)) {
740 struct request_queue *q = bdev_get_queue(rdev->bdev);
744 /* Note the '|| 1' - when read_balance prefers
745 * non-congested targets, it can be removed
747 if ((bits & (1<<BDI_async_congested)) || 1)
748 ret |= bdi_congested(&q->backing_dev_info, bits);
750 ret &= bdi_congested(&q->backing_dev_info, bits);
756 EXPORT_SYMBOL_GPL(md_raid1_congested);
758 static int raid1_congested(void *data, int bits)
760 struct mddev *mddev = data;
762 return mddev_congested(mddev, bits) ||
763 md_raid1_congested(mddev, bits);
766 static void flush_pending_writes(struct r1conf *conf)
768 /* Any writes that have been queued but are awaiting
769 * bitmap updates get flushed here.
771 spin_lock_irq(&conf->device_lock);
773 if (conf->pending_bio_list.head) {
775 bio = bio_list_get(&conf->pending_bio_list);
776 conf->pending_count = 0;
777 spin_unlock_irq(&conf->device_lock);
778 /* flush any pending bitmap writes to
779 * disk before proceeding w/ I/O */
780 bitmap_unplug(conf->mddev->bitmap);
781 wake_up(&conf->wait_barrier);
783 while (bio) { /* submit pending writes */
784 struct bio *next = bio->bi_next;
786 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
787 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
791 generic_make_request(bio);
795 spin_unlock_irq(&conf->device_lock);
799 * Sometimes we need to suspend IO while we do something else,
800 * either some resync/recovery, or reconfigure the array.
801 * To do this we raise a 'barrier'.
802 * The 'barrier' is a counter that can be raised multiple times
803 * to count how many activities are happening which preclude
805 * We can only raise the barrier if there is no pending IO.
806 * i.e. if nr_pending == 0.
807 * We choose only to raise the barrier if no-one is waiting for the
808 * barrier to go down. This means that as soon as an IO request
809 * is ready, no other operations which require a barrier will start
810 * until the IO request has had a chance.
812 * So: regular IO calls 'wait_barrier'. When that returns there
813 * is no backgroup IO happening, It must arrange to call
814 * allow_barrier when it has finished its IO.
815 * backgroup IO calls must call raise_barrier. Once that returns
816 * there is no normal IO happeing. It must arrange to call
817 * lower_barrier when the particular background IO completes.
819 static void raise_barrier(struct r1conf *conf)
821 spin_lock_irq(&conf->resync_lock);
823 /* Wait until no block IO is waiting */
824 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
827 /* block any new IO from starting */
830 /* Now wait for all pending IO to complete */
831 wait_event_lock_irq(conf->wait_barrier,
832 !conf->array_frozen &&
833 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
836 spin_unlock_irq(&conf->resync_lock);
839 static void lower_barrier(struct r1conf *conf)
842 BUG_ON(conf->barrier <= 0);
843 spin_lock_irqsave(&conf->resync_lock, flags);
845 spin_unlock_irqrestore(&conf->resync_lock, flags);
846 wake_up(&conf->wait_barrier);
849 static void wait_barrier(struct r1conf *conf)
851 spin_lock_irq(&conf->resync_lock);
854 /* Wait for the barrier to drop.
855 * However if there are already pending
856 * requests (preventing the barrier from
857 * rising completely), and the
858 * pre-process bio queue isn't empty,
859 * then don't wait, as we need to empty
860 * that queue to get the nr_pending
863 wait_event_lock_irq(conf->wait_barrier,
864 !conf->array_frozen &&
868 !bio_list_empty(current->bio_list))),
873 spin_unlock_irq(&conf->resync_lock);
876 static void allow_barrier(struct r1conf *conf)
879 spin_lock_irqsave(&conf->resync_lock, flags);
881 spin_unlock_irqrestore(&conf->resync_lock, flags);
882 wake_up(&conf->wait_barrier);
885 static void freeze_array(struct r1conf *conf, int extra)
887 /* stop syncio and normal IO and wait for everything to
889 * We wait until nr_pending match nr_queued+extra
890 * This is called in the context of one normal IO request
891 * that has failed. Thus any sync request that might be pending
892 * will be blocked by nr_pending, and we need to wait for
893 * pending IO requests to complete or be queued for re-try.
894 * Thus the number queued (nr_queued) plus this request (extra)
895 * must match the number of pending IOs (nr_pending) before
898 spin_lock_irq(&conf->resync_lock);
899 conf->array_frozen = 1;
900 wait_event_lock_irq_cmd(conf->wait_barrier,
901 conf->nr_pending == conf->nr_queued+extra,
903 flush_pending_writes(conf));
904 spin_unlock_irq(&conf->resync_lock);
906 static void unfreeze_array(struct r1conf *conf)
908 /* reverse the effect of the freeze */
909 spin_lock_irq(&conf->resync_lock);
910 conf->array_frozen = 0;
911 wake_up(&conf->wait_barrier);
912 spin_unlock_irq(&conf->resync_lock);
916 /* duplicate the data pages for behind I/O
918 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
921 struct bio_vec *bvec;
922 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
924 if (unlikely(!bvecs))
927 bio_for_each_segment_all(bvec, bio, i) {
929 bvecs[i].bv_page = alloc_page(GFP_NOIO);
930 if (unlikely(!bvecs[i].bv_page))
932 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
933 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
934 kunmap(bvecs[i].bv_page);
935 kunmap(bvec->bv_page);
937 r1_bio->behind_bvecs = bvecs;
938 r1_bio->behind_page_count = bio->bi_vcnt;
939 set_bit(R1BIO_BehindIO, &r1_bio->state);
943 for (i = 0; i < bio->bi_vcnt; i++)
944 if (bvecs[i].bv_page)
945 put_page(bvecs[i].bv_page);
947 pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
950 struct raid1_plug_cb {
951 struct blk_plug_cb cb;
952 struct bio_list pending;
956 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
958 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
960 struct mddev *mddev = plug->cb.data;
961 struct r1conf *conf = mddev->private;
964 if (from_schedule || current->bio_list) {
965 spin_lock_irq(&conf->device_lock);
966 bio_list_merge(&conf->pending_bio_list, &plug->pending);
967 conf->pending_count += plug->pending_cnt;
968 spin_unlock_irq(&conf->device_lock);
969 wake_up(&conf->wait_barrier);
970 md_wakeup_thread(mddev->thread);
975 /* we aren't scheduling, so we can do the write-out directly. */
976 bio = bio_list_get(&plug->pending);
977 bitmap_unplug(mddev->bitmap);
978 wake_up(&conf->wait_barrier);
980 while (bio) { /* submit pending writes */
981 struct bio *next = bio->bi_next;
983 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
984 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
988 generic_make_request(bio);
994 static void make_request(struct mddev *mddev, struct bio * bio)
996 struct r1conf *conf = mddev->private;
997 struct raid1_info *mirror;
998 struct r1bio *r1_bio;
999 struct bio *read_bio;
1001 struct bitmap *bitmap;
1002 unsigned long flags;
1003 const int rw = bio_data_dir(bio);
1004 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1005 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1006 const unsigned long do_discard = (bio->bi_rw
1007 & (REQ_DISCARD | REQ_SECURE));
1008 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1009 struct md_rdev *blocked_rdev;
1010 struct blk_plug_cb *cb;
1011 struct raid1_plug_cb *plug = NULL;
1013 int sectors_handled;
1017 * Register the new request and wait if the reconstruction
1018 * thread has put up a bar for new requests.
1019 * Continue immediately if no resync is active currently.
1022 md_write_start(mddev, bio); /* wait on superblock update early */
1024 if (bio_data_dir(bio) == WRITE &&
1025 bio_end_sector(bio) > mddev->suspend_lo &&
1026 bio->bi_sector < mddev->suspend_hi) {
1027 /* As the suspend_* range is controlled by
1028 * userspace, we want an interruptible
1033 flush_signals(current);
1034 prepare_to_wait(&conf->wait_barrier,
1035 &w, TASK_INTERRUPTIBLE);
1036 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1037 bio->bi_sector >= mddev->suspend_hi)
1041 finish_wait(&conf->wait_barrier, &w);
1046 bitmap = mddev->bitmap;
1049 * make_request() can abort the operation when READA is being
1050 * used and no empty request is available.
1053 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1055 r1_bio->master_bio = bio;
1056 r1_bio->sectors = bio_sectors(bio);
1058 r1_bio->mddev = mddev;
1059 r1_bio->sector = bio->bi_sector;
1061 /* We might need to issue multiple reads to different
1062 * devices if there are bad blocks around, so we keep
1063 * track of the number of reads in bio->bi_phys_segments.
1064 * If this is 0, there is only one r1_bio and no locking
1065 * will be needed when requests complete. If it is
1066 * non-zero, then it is the number of not-completed requests.
1068 bio->bi_phys_segments = 0;
1069 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1073 * read balancing logic:
1078 rdisk = read_balance(conf, r1_bio, &max_sectors);
1081 /* couldn't find anywhere to read from */
1082 raid_end_bio_io(r1_bio);
1085 mirror = conf->mirrors + rdisk;
1087 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1089 /* Reading from a write-mostly device must
1090 * take care not to over-take any writes
1093 wait_event(bitmap->behind_wait,
1094 atomic_read(&bitmap->behind_writes) == 0);
1096 r1_bio->read_disk = rdisk;
1098 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1099 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1102 r1_bio->bios[rdisk] = read_bio;
1104 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1105 read_bio->bi_bdev = mirror->rdev->bdev;
1106 read_bio->bi_end_io = raid1_end_read_request;
1107 read_bio->bi_rw = READ | do_sync;
1108 read_bio->bi_private = r1_bio;
1110 if (max_sectors < r1_bio->sectors) {
1111 /* could not read all from this device, so we will
1112 * need another r1_bio.
1115 sectors_handled = (r1_bio->sector + max_sectors
1117 r1_bio->sectors = max_sectors;
1118 spin_lock_irq(&conf->device_lock);
1119 if (bio->bi_phys_segments == 0)
1120 bio->bi_phys_segments = 2;
1122 bio->bi_phys_segments++;
1123 spin_unlock_irq(&conf->device_lock);
1124 /* Cannot call generic_make_request directly
1125 * as that will be queued in __make_request
1126 * and subsequent mempool_alloc might block waiting
1127 * for it. So hand bio over to raid1d.
1129 reschedule_retry(r1_bio);
1131 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1133 r1_bio->master_bio = bio;
1134 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1136 r1_bio->mddev = mddev;
1137 r1_bio->sector = bio->bi_sector + sectors_handled;
1140 generic_make_request(read_bio);
1147 if (conf->pending_count >= max_queued_requests) {
1148 md_wakeup_thread(mddev->thread);
1149 wait_event(conf->wait_barrier,
1150 conf->pending_count < max_queued_requests);
1152 /* first select target devices under rcu_lock and
1153 * inc refcount on their rdev. Record them by setting
1155 * If there are known/acknowledged bad blocks on any device on
1156 * which we have seen a write error, we want to avoid writing those
1158 * This potentially requires several writes to write around
1159 * the bad blocks. Each set of writes gets it's own r1bio
1160 * with a set of bios attached.
1163 disks = conf->raid_disks * 2;
1165 blocked_rdev = NULL;
1167 max_sectors = r1_bio->sectors;
1168 for (i = 0; i < disks; i++) {
1169 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1170 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1171 atomic_inc(&rdev->nr_pending);
1172 blocked_rdev = rdev;
1175 r1_bio->bios[i] = NULL;
1176 if (!rdev || test_bit(Faulty, &rdev->flags)
1177 || test_bit(Unmerged, &rdev->flags)) {
1178 if (i < conf->raid_disks)
1179 set_bit(R1BIO_Degraded, &r1_bio->state);
1183 atomic_inc(&rdev->nr_pending);
1184 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1189 is_bad = is_badblock(rdev, r1_bio->sector,
1191 &first_bad, &bad_sectors);
1193 /* mustn't write here until the bad block is
1195 set_bit(BlockedBadBlocks, &rdev->flags);
1196 blocked_rdev = rdev;
1199 if (is_bad && first_bad <= r1_bio->sector) {
1200 /* Cannot write here at all */
1201 bad_sectors -= (r1_bio->sector - first_bad);
1202 if (bad_sectors < max_sectors)
1203 /* mustn't write more than bad_sectors
1204 * to other devices yet
1206 max_sectors = bad_sectors;
1207 rdev_dec_pending(rdev, mddev);
1208 /* We don't set R1BIO_Degraded as that
1209 * only applies if the disk is
1210 * missing, so it might be re-added,
1211 * and we want to know to recover this
1213 * In this case the device is here,
1214 * and the fact that this chunk is not
1215 * in-sync is recorded in the bad
1221 int good_sectors = first_bad - r1_bio->sector;
1222 if (good_sectors < max_sectors)
1223 max_sectors = good_sectors;
1226 r1_bio->bios[i] = bio;
1230 if (unlikely(blocked_rdev)) {
1231 /* Wait for this device to become unblocked */
1234 for (j = 0; j < i; j++)
1235 if (r1_bio->bios[j])
1236 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1238 allow_barrier(conf);
1239 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1244 if (max_sectors < r1_bio->sectors) {
1245 /* We are splitting this write into multiple parts, so
1246 * we need to prepare for allocating another r1_bio.
1248 r1_bio->sectors = max_sectors;
1249 spin_lock_irq(&conf->device_lock);
1250 if (bio->bi_phys_segments == 0)
1251 bio->bi_phys_segments = 2;
1253 bio->bi_phys_segments++;
1254 spin_unlock_irq(&conf->device_lock);
1256 sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1258 atomic_set(&r1_bio->remaining, 1);
1259 atomic_set(&r1_bio->behind_remaining, 0);
1262 for (i = 0; i < disks; i++) {
1264 if (!r1_bio->bios[i])
1267 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1268 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1272 * Not if there are too many, or cannot
1273 * allocate memory, or a reader on WriteMostly
1274 * is waiting for behind writes to flush */
1276 (atomic_read(&bitmap->behind_writes)
1277 < mddev->bitmap_info.max_write_behind) &&
1278 !waitqueue_active(&bitmap->behind_wait))
1279 alloc_behind_pages(mbio, r1_bio);
1281 bitmap_startwrite(bitmap, r1_bio->sector,
1283 test_bit(R1BIO_BehindIO,
1287 if (r1_bio->behind_bvecs) {
1288 struct bio_vec *bvec;
1292 * We trimmed the bio, so _all is legit
1294 bio_for_each_segment_all(bvec, mbio, j)
1295 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1296 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1297 atomic_inc(&r1_bio->behind_remaining);
1300 r1_bio->bios[i] = mbio;
1302 mbio->bi_sector = (r1_bio->sector +
1303 conf->mirrors[i].rdev->data_offset);
1304 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1305 mbio->bi_end_io = raid1_end_write_request;
1307 WRITE | do_flush_fua | do_sync | do_discard | do_same;
1308 mbio->bi_private = r1_bio;
1310 atomic_inc(&r1_bio->remaining);
1312 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1314 plug = container_of(cb, struct raid1_plug_cb, cb);
1317 spin_lock_irqsave(&conf->device_lock, flags);
1319 bio_list_add(&plug->pending, mbio);
1320 plug->pending_cnt++;
1322 bio_list_add(&conf->pending_bio_list, mbio);
1323 conf->pending_count++;
1325 spin_unlock_irqrestore(&conf->device_lock, flags);
1327 md_wakeup_thread(mddev->thread);
1329 /* Mustn't call r1_bio_write_done before this next test,
1330 * as it could result in the bio being freed.
1332 if (sectors_handled < bio_sectors(bio)) {
1333 r1_bio_write_done(r1_bio);
1334 /* We need another r1_bio. It has already been counted
1335 * in bio->bi_phys_segments
1337 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1338 r1_bio->master_bio = bio;
1339 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1341 r1_bio->mddev = mddev;
1342 r1_bio->sector = bio->bi_sector + sectors_handled;
1346 r1_bio_write_done(r1_bio);
1348 /* In case raid1d snuck in to freeze_array */
1349 wake_up(&conf->wait_barrier);
1352 static void status(struct seq_file *seq, struct mddev *mddev)
1354 struct r1conf *conf = mddev->private;
1357 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1358 conf->raid_disks - mddev->degraded);
1360 for (i = 0; i < conf->raid_disks; i++) {
1361 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1362 seq_printf(seq, "%s",
1363 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1366 seq_printf(seq, "]");
1370 static void error(struct mddev *mddev, struct md_rdev *rdev)
1372 char b[BDEVNAME_SIZE];
1373 struct r1conf *conf = mddev->private;
1376 * If it is not operational, then we have already marked it as dead
1377 * else if it is the last working disks, ignore the error, let the
1378 * next level up know.
1379 * else mark the drive as failed
1381 if (test_bit(In_sync, &rdev->flags)
1382 && (conf->raid_disks - mddev->degraded) == 1) {
1384 * Don't fail the drive, act as though we were just a
1385 * normal single drive.
1386 * However don't try a recovery from this drive as
1387 * it is very likely to fail.
1389 conf->recovery_disabled = mddev->recovery_disabled;
1392 set_bit(Blocked, &rdev->flags);
1393 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1394 unsigned long flags;
1395 spin_lock_irqsave(&conf->device_lock, flags);
1397 set_bit(Faulty, &rdev->flags);
1398 spin_unlock_irqrestore(&conf->device_lock, flags);
1400 * if recovery is running, make sure it aborts.
1402 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1404 set_bit(Faulty, &rdev->flags);
1405 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1407 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1408 "md/raid1:%s: Operation continuing on %d devices.\n",
1409 mdname(mddev), bdevname(rdev->bdev, b),
1410 mdname(mddev), conf->raid_disks - mddev->degraded);
1413 static void print_conf(struct r1conf *conf)
1417 printk(KERN_DEBUG "RAID1 conf printout:\n");
1419 printk(KERN_DEBUG "(!conf)\n");
1422 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1426 for (i = 0; i < conf->raid_disks; i++) {
1427 char b[BDEVNAME_SIZE];
1428 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1430 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1431 i, !test_bit(In_sync, &rdev->flags),
1432 !test_bit(Faulty, &rdev->flags),
1433 bdevname(rdev->bdev,b));
1438 static void close_sync(struct r1conf *conf)
1441 allow_barrier(conf);
1443 mempool_destroy(conf->r1buf_pool);
1444 conf->r1buf_pool = NULL;
1447 static int raid1_spare_active(struct mddev *mddev)
1450 struct r1conf *conf = mddev->private;
1452 unsigned long flags;
1455 * Find all failed disks within the RAID1 configuration
1456 * and mark them readable.
1457 * Called under mddev lock, so rcu protection not needed.
1459 for (i = 0; i < conf->raid_disks; i++) {
1460 struct md_rdev *rdev = conf->mirrors[i].rdev;
1461 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1463 && repl->recovery_offset == MaxSector
1464 && !test_bit(Faulty, &repl->flags)
1465 && !test_and_set_bit(In_sync, &repl->flags)) {
1466 /* replacement has just become active */
1468 !test_and_clear_bit(In_sync, &rdev->flags))
1471 /* Replaced device not technically
1472 * faulty, but we need to be sure
1473 * it gets removed and never re-added
1475 set_bit(Faulty, &rdev->flags);
1476 sysfs_notify_dirent_safe(
1481 && rdev->recovery_offset == MaxSector
1482 && !test_bit(Faulty, &rdev->flags)
1483 && !test_and_set_bit(In_sync, &rdev->flags)) {
1485 sysfs_notify_dirent_safe(rdev->sysfs_state);
1488 spin_lock_irqsave(&conf->device_lock, flags);
1489 mddev->degraded -= count;
1490 spin_unlock_irqrestore(&conf->device_lock, flags);
1497 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1499 struct r1conf *conf = mddev->private;
1502 struct raid1_info *p;
1504 int last = conf->raid_disks - 1;
1505 struct request_queue *q = bdev_get_queue(rdev->bdev);
1507 if (mddev->recovery_disabled == conf->recovery_disabled)
1510 if (rdev->raid_disk >= 0)
1511 first = last = rdev->raid_disk;
1513 if (q->merge_bvec_fn) {
1514 set_bit(Unmerged, &rdev->flags);
1515 mddev->merge_check_needed = 1;
1518 for (mirror = first; mirror <= last; mirror++) {
1519 p = conf->mirrors+mirror;
1523 disk_stack_limits(mddev->gendisk, rdev->bdev,
1524 rdev->data_offset << 9);
1526 p->head_position = 0;
1527 rdev->raid_disk = mirror;
1529 /* As all devices are equivalent, we don't need a full recovery
1530 * if this was recently any drive of the array
1532 if (rdev->saved_raid_disk < 0)
1534 rcu_assign_pointer(p->rdev, rdev);
1537 if (test_bit(WantReplacement, &p->rdev->flags) &&
1538 p[conf->raid_disks].rdev == NULL) {
1539 /* Add this device as a replacement */
1540 clear_bit(In_sync, &rdev->flags);
1541 set_bit(Replacement, &rdev->flags);
1542 rdev->raid_disk = mirror;
1545 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1549 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1550 /* Some requests might not have seen this new
1551 * merge_bvec_fn. We must wait for them to complete
1552 * before merging the device fully.
1553 * First we make sure any code which has tested
1554 * our function has submitted the request, then
1555 * we wait for all outstanding requests to complete.
1557 synchronize_sched();
1558 freeze_array(conf, 0);
1559 unfreeze_array(conf);
1560 clear_bit(Unmerged, &rdev->flags);
1562 md_integrity_add_rdev(rdev, mddev);
1563 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1564 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1569 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1571 struct r1conf *conf = mddev->private;
1573 int number = rdev->raid_disk;
1574 struct raid1_info *p = conf->mirrors + number;
1576 if (rdev != p->rdev)
1577 p = conf->mirrors + conf->raid_disks + number;
1580 if (rdev == p->rdev) {
1581 if (test_bit(In_sync, &rdev->flags) ||
1582 atomic_read(&rdev->nr_pending)) {
1586 /* Only remove non-faulty devices if recovery
1589 if (!test_bit(Faulty, &rdev->flags) &&
1590 mddev->recovery_disabled != conf->recovery_disabled &&
1591 mddev->degraded < conf->raid_disks) {
1597 if (atomic_read(&rdev->nr_pending)) {
1598 /* lost the race, try later */
1602 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1603 /* We just removed a device that is being replaced.
1604 * Move down the replacement. We drain all IO before
1605 * doing this to avoid confusion.
1607 struct md_rdev *repl =
1608 conf->mirrors[conf->raid_disks + number].rdev;
1609 freeze_array(conf, 0);
1610 clear_bit(Replacement, &repl->flags);
1612 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1613 unfreeze_array(conf);
1614 clear_bit(WantReplacement, &rdev->flags);
1616 clear_bit(WantReplacement, &rdev->flags);
1617 err = md_integrity_register(mddev);
1626 static void end_sync_read(struct bio *bio, int error)
1628 struct r1bio *r1_bio = bio->bi_private;
1630 update_head_pos(r1_bio->read_disk, r1_bio);
1633 * we have read a block, now it needs to be re-written,
1634 * or re-read if the read failed.
1635 * We don't do much here, just schedule handling by raid1d
1637 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1638 set_bit(R1BIO_Uptodate, &r1_bio->state);
1640 if (atomic_dec_and_test(&r1_bio->remaining))
1641 reschedule_retry(r1_bio);
1644 static void end_sync_write(struct bio *bio, int error)
1646 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1647 struct r1bio *r1_bio = bio->bi_private;
1648 struct mddev *mddev = r1_bio->mddev;
1649 struct r1conf *conf = mddev->private;
1654 mirror = find_bio_disk(r1_bio, bio);
1657 sector_t sync_blocks = 0;
1658 sector_t s = r1_bio->sector;
1659 long sectors_to_go = r1_bio->sectors;
1660 /* make sure these bits doesn't get cleared. */
1662 bitmap_end_sync(mddev->bitmap, s,
1665 sectors_to_go -= sync_blocks;
1666 } while (sectors_to_go > 0);
1667 set_bit(WriteErrorSeen,
1668 &conf->mirrors[mirror].rdev->flags);
1669 if (!test_and_set_bit(WantReplacement,
1670 &conf->mirrors[mirror].rdev->flags))
1671 set_bit(MD_RECOVERY_NEEDED, &
1673 set_bit(R1BIO_WriteError, &r1_bio->state);
1674 } else if (is_badblock(conf->mirrors[mirror].rdev,
1677 &first_bad, &bad_sectors) &&
1678 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1681 &first_bad, &bad_sectors)
1683 set_bit(R1BIO_MadeGood, &r1_bio->state);
1685 if (atomic_dec_and_test(&r1_bio->remaining)) {
1686 int s = r1_bio->sectors;
1687 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1688 test_bit(R1BIO_WriteError, &r1_bio->state))
1689 reschedule_retry(r1_bio);
1692 md_done_sync(mddev, s, uptodate);
1697 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1698 int sectors, struct page *page, int rw)
1700 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1704 set_bit(WriteErrorSeen, &rdev->flags);
1705 if (!test_and_set_bit(WantReplacement,
1707 set_bit(MD_RECOVERY_NEEDED, &
1708 rdev->mddev->recovery);
1710 /* need to record an error - either for the block or the device */
1711 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1712 md_error(rdev->mddev, rdev);
1716 static int fix_sync_read_error(struct r1bio *r1_bio)
1718 /* Try some synchronous reads of other devices to get
1719 * good data, much like with normal read errors. Only
1720 * read into the pages we already have so we don't
1721 * need to re-issue the read request.
1722 * We don't need to freeze the array, because being in an
1723 * active sync request, there is no normal IO, and
1724 * no overlapping syncs.
1725 * We don't need to check is_badblock() again as we
1726 * made sure that anything with a bad block in range
1727 * will have bi_end_io clear.
1729 struct mddev *mddev = r1_bio->mddev;
1730 struct r1conf *conf = mddev->private;
1731 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1732 sector_t sect = r1_bio->sector;
1733 int sectors = r1_bio->sectors;
1738 int d = r1_bio->read_disk;
1740 struct md_rdev *rdev;
1743 if (s > (PAGE_SIZE>>9))
1746 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1747 /* No rcu protection needed here devices
1748 * can only be removed when no resync is
1749 * active, and resync is currently active
1751 rdev = conf->mirrors[d].rdev;
1752 if (sync_page_io(rdev, sect, s<<9,
1753 bio->bi_io_vec[idx].bv_page,
1760 if (d == conf->raid_disks * 2)
1762 } while (!success && d != r1_bio->read_disk);
1765 char b[BDEVNAME_SIZE];
1767 /* Cannot read from anywhere, this block is lost.
1768 * Record a bad block on each device. If that doesn't
1769 * work just disable and interrupt the recovery.
1770 * Don't fail devices as that won't really help.
1772 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1773 " for block %llu\n",
1775 bdevname(bio->bi_bdev, b),
1776 (unsigned long long)r1_bio->sector);
1777 for (d = 0; d < conf->raid_disks * 2; d++) {
1778 rdev = conf->mirrors[d].rdev;
1779 if (!rdev || test_bit(Faulty, &rdev->flags))
1781 if (!rdev_set_badblocks(rdev, sect, s, 0))
1785 conf->recovery_disabled =
1786 mddev->recovery_disabled;
1787 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1788 md_done_sync(mddev, r1_bio->sectors, 0);
1800 /* write it back and re-read */
1801 while (d != r1_bio->read_disk) {
1803 d = conf->raid_disks * 2;
1805 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1807 rdev = conf->mirrors[d].rdev;
1808 if (r1_sync_page_io(rdev, sect, s,
1809 bio->bi_io_vec[idx].bv_page,
1811 r1_bio->bios[d]->bi_end_io = NULL;
1812 rdev_dec_pending(rdev, mddev);
1816 while (d != r1_bio->read_disk) {
1818 d = conf->raid_disks * 2;
1820 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1822 rdev = conf->mirrors[d].rdev;
1823 if (r1_sync_page_io(rdev, sect, s,
1824 bio->bi_io_vec[idx].bv_page,
1826 atomic_add(s, &rdev->corrected_errors);
1832 set_bit(R1BIO_Uptodate, &r1_bio->state);
1833 set_bit(BIO_UPTODATE, &bio->bi_flags);
1837 static int process_checks(struct r1bio *r1_bio)
1839 /* We have read all readable devices. If we haven't
1840 * got the block, then there is no hope left.
1841 * If we have, then we want to do a comparison
1842 * and skip the write if everything is the same.
1843 * If any blocks failed to read, then we need to
1844 * attempt an over-write
1846 struct mddev *mddev = r1_bio->mddev;
1847 struct r1conf *conf = mddev->private;
1852 /* Fix variable parts of all bios */
1853 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1854 for (i = 0; i < conf->raid_disks * 2; i++) {
1857 struct bio *b = r1_bio->bios[i];
1858 if (b->bi_end_io != end_sync_read)
1860 /* fixup the bio for reuse */
1863 b->bi_size = r1_bio->sectors << 9;
1864 b->bi_sector = r1_bio->sector +
1865 conf->mirrors[i].rdev->data_offset;
1866 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1867 b->bi_end_io = end_sync_read;
1868 b->bi_private = r1_bio;
1871 for (j = 0; j < vcnt ; j++) {
1873 bi = &b->bi_io_vec[j];
1875 if (size > PAGE_SIZE)
1876 bi->bv_len = PAGE_SIZE;
1882 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1883 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1884 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1885 r1_bio->bios[primary]->bi_end_io = NULL;
1886 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1889 r1_bio->read_disk = primary;
1890 for (i = 0; i < conf->raid_disks * 2; i++) {
1892 struct bio *pbio = r1_bio->bios[primary];
1893 struct bio *sbio = r1_bio->bios[i];
1895 if (sbio->bi_end_io != end_sync_read)
1898 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1899 for (j = vcnt; j-- ; ) {
1901 p = pbio->bi_io_vec[j].bv_page;
1902 s = sbio->bi_io_vec[j].bv_page;
1903 if (memcmp(page_address(p),
1905 sbio->bi_io_vec[j].bv_len))
1911 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1912 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1913 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1914 /* No need to write to this device. */
1915 sbio->bi_end_io = NULL;
1916 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1920 bio_copy_data(sbio, pbio);
1925 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1927 struct r1conf *conf = mddev->private;
1929 int disks = conf->raid_disks * 2;
1930 struct bio *bio, *wbio;
1932 bio = r1_bio->bios[r1_bio->read_disk];
1934 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1935 /* ouch - failed to read all of that. */
1936 if (!fix_sync_read_error(r1_bio))
1939 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1940 if (process_checks(r1_bio) < 0)
1945 atomic_set(&r1_bio->remaining, 1);
1946 for (i = 0; i < disks ; i++) {
1947 wbio = r1_bio->bios[i];
1948 if (wbio->bi_end_io == NULL ||
1949 (wbio->bi_end_io == end_sync_read &&
1950 (i == r1_bio->read_disk ||
1951 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1954 wbio->bi_rw = WRITE;
1955 wbio->bi_end_io = end_sync_write;
1956 atomic_inc(&r1_bio->remaining);
1957 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
1959 generic_make_request(wbio);
1962 if (atomic_dec_and_test(&r1_bio->remaining)) {
1963 /* if we're here, all write(s) have completed, so clean up */
1964 int s = r1_bio->sectors;
1965 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1966 test_bit(R1BIO_WriteError, &r1_bio->state))
1967 reschedule_retry(r1_bio);
1970 md_done_sync(mddev, s, 1);
1976 * This is a kernel thread which:
1978 * 1. Retries failed read operations on working mirrors.
1979 * 2. Updates the raid superblock when problems encounter.
1980 * 3. Performs writes following reads for array synchronising.
1983 static void fix_read_error(struct r1conf *conf, int read_disk,
1984 sector_t sect, int sectors)
1986 struct mddev *mddev = conf->mddev;
1992 struct md_rdev *rdev;
1994 if (s > (PAGE_SIZE>>9))
1998 /* Note: no rcu protection needed here
1999 * as this is synchronous in the raid1d thread
2000 * which is the thread that might remove
2001 * a device. If raid1d ever becomes multi-threaded....
2006 rdev = conf->mirrors[d].rdev;
2008 (test_bit(In_sync, &rdev->flags) ||
2009 (!test_bit(Faulty, &rdev->flags) &&
2010 rdev->recovery_offset >= sect + s)) &&
2011 is_badblock(rdev, sect, s,
2012 &first_bad, &bad_sectors) == 0 &&
2013 sync_page_io(rdev, sect, s<<9,
2014 conf->tmppage, READ, false))
2018 if (d == conf->raid_disks * 2)
2021 } while (!success && d != read_disk);
2024 /* Cannot read from anywhere - mark it bad */
2025 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2026 if (!rdev_set_badblocks(rdev, sect, s, 0))
2027 md_error(mddev, rdev);
2030 /* write it back and re-read */
2032 while (d != read_disk) {
2034 d = conf->raid_disks * 2;
2036 rdev = conf->mirrors[d].rdev;
2038 test_bit(In_sync, &rdev->flags))
2039 r1_sync_page_io(rdev, sect, s,
2040 conf->tmppage, WRITE);
2043 while (d != read_disk) {
2044 char b[BDEVNAME_SIZE];
2046 d = conf->raid_disks * 2;
2048 rdev = conf->mirrors[d].rdev;
2050 test_bit(In_sync, &rdev->flags)) {
2051 if (r1_sync_page_io(rdev, sect, s,
2052 conf->tmppage, READ)) {
2053 atomic_add(s, &rdev->corrected_errors);
2055 "md/raid1:%s: read error corrected "
2056 "(%d sectors at %llu on %s)\n",
2058 (unsigned long long)(sect +
2060 bdevname(rdev->bdev, b));
2069 static int narrow_write_error(struct r1bio *r1_bio, int i)
2071 struct mddev *mddev = r1_bio->mddev;
2072 struct r1conf *conf = mddev->private;
2073 struct md_rdev *rdev = conf->mirrors[i].rdev;
2075 /* bio has the data to be written to device 'i' where
2076 * we just recently had a write error.
2077 * We repeatedly clone the bio and trim down to one block,
2078 * then try the write. Where the write fails we record
2080 * It is conceivable that the bio doesn't exactly align with
2081 * blocks. We must handle this somehow.
2083 * We currently own a reference on the rdev.
2089 int sect_to_write = r1_bio->sectors;
2092 if (rdev->badblocks.shift < 0)
2095 block_sectors = 1 << rdev->badblocks.shift;
2096 sector = r1_bio->sector;
2097 sectors = ((sector + block_sectors)
2098 & ~(sector_t)(block_sectors - 1))
2101 while (sect_to_write) {
2103 if (sectors > sect_to_write)
2104 sectors = sect_to_write;
2105 /* Write at 'sector' for 'sectors'*/
2107 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2108 unsigned vcnt = r1_bio->behind_page_count;
2109 struct bio_vec *vec = r1_bio->behind_bvecs;
2111 while (!vec->bv_page) {
2116 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2117 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2119 wbio->bi_vcnt = vcnt;
2121 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2124 wbio->bi_rw = WRITE;
2125 wbio->bi_sector = r1_bio->sector;
2126 wbio->bi_size = r1_bio->sectors << 9;
2128 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2129 wbio->bi_sector += rdev->data_offset;
2130 wbio->bi_bdev = rdev->bdev;
2131 if (submit_bio_wait(WRITE, wbio) == 0)
2133 ok = rdev_set_badblocks(rdev, sector,
2138 sect_to_write -= sectors;
2140 sectors = block_sectors;
2145 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2148 int s = r1_bio->sectors;
2149 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2150 struct md_rdev *rdev = conf->mirrors[m].rdev;
2151 struct bio *bio = r1_bio->bios[m];
2152 if (bio->bi_end_io == NULL)
2154 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2155 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2156 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2158 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2159 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2160 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2161 md_error(conf->mddev, rdev);
2165 md_done_sync(conf->mddev, s, 1);
2168 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2171 for (m = 0; m < conf->raid_disks * 2 ; m++)
2172 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2173 struct md_rdev *rdev = conf->mirrors[m].rdev;
2174 rdev_clear_badblocks(rdev,
2176 r1_bio->sectors, 0);
2177 rdev_dec_pending(rdev, conf->mddev);
2178 } else if (r1_bio->bios[m] != NULL) {
2179 /* This drive got a write error. We need to
2180 * narrow down and record precise write
2183 if (!narrow_write_error(r1_bio, m)) {
2184 md_error(conf->mddev,
2185 conf->mirrors[m].rdev);
2186 /* an I/O failed, we can't clear the bitmap */
2187 set_bit(R1BIO_Degraded, &r1_bio->state);
2189 rdev_dec_pending(conf->mirrors[m].rdev,
2192 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2193 close_write(r1_bio);
2194 raid_end_bio_io(r1_bio);
2197 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2201 struct mddev *mddev = conf->mddev;
2203 char b[BDEVNAME_SIZE];
2204 struct md_rdev *rdev;
2206 clear_bit(R1BIO_ReadError, &r1_bio->state);
2207 /* we got a read error. Maybe the drive is bad. Maybe just
2208 * the block and we can fix it.
2209 * We freeze all other IO, and try reading the block from
2210 * other devices. When we find one, we re-write
2211 * and check it that fixes the read error.
2212 * This is all done synchronously while the array is
2215 if (mddev->ro == 0) {
2216 freeze_array(conf, 1);
2217 fix_read_error(conf, r1_bio->read_disk,
2218 r1_bio->sector, r1_bio->sectors);
2219 unfreeze_array(conf);
2221 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2222 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2224 bio = r1_bio->bios[r1_bio->read_disk];
2225 bdevname(bio->bi_bdev, b);
2227 disk = read_balance(conf, r1_bio, &max_sectors);
2229 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2230 " read error for block %llu\n",
2231 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2232 raid_end_bio_io(r1_bio);
2234 const unsigned long do_sync
2235 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2237 r1_bio->bios[r1_bio->read_disk] =
2238 mddev->ro ? IO_BLOCKED : NULL;
2241 r1_bio->read_disk = disk;
2242 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2243 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2244 r1_bio->bios[r1_bio->read_disk] = bio;
2245 rdev = conf->mirrors[disk].rdev;
2246 printk_ratelimited(KERN_ERR
2247 "md/raid1:%s: redirecting sector %llu"
2248 " to other mirror: %s\n",
2250 (unsigned long long)r1_bio->sector,
2251 bdevname(rdev->bdev, b));
2252 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2253 bio->bi_bdev = rdev->bdev;
2254 bio->bi_end_io = raid1_end_read_request;
2255 bio->bi_rw = READ | do_sync;
2256 bio->bi_private = r1_bio;
2257 if (max_sectors < r1_bio->sectors) {
2258 /* Drat - have to split this up more */
2259 struct bio *mbio = r1_bio->master_bio;
2260 int sectors_handled = (r1_bio->sector + max_sectors
2262 r1_bio->sectors = max_sectors;
2263 spin_lock_irq(&conf->device_lock);
2264 if (mbio->bi_phys_segments == 0)
2265 mbio->bi_phys_segments = 2;
2267 mbio->bi_phys_segments++;
2268 spin_unlock_irq(&conf->device_lock);
2269 generic_make_request(bio);
2272 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2274 r1_bio->master_bio = mbio;
2275 r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2277 set_bit(R1BIO_ReadError, &r1_bio->state);
2278 r1_bio->mddev = mddev;
2279 r1_bio->sector = mbio->bi_sector + sectors_handled;
2283 generic_make_request(bio);
2287 static void raid1d(struct md_thread *thread)
2289 struct mddev *mddev = thread->mddev;
2290 struct r1bio *r1_bio;
2291 unsigned long flags;
2292 struct r1conf *conf = mddev->private;
2293 struct list_head *head = &conf->retry_list;
2294 struct blk_plug plug;
2296 md_check_recovery(mddev);
2298 blk_start_plug(&plug);
2301 flush_pending_writes(conf);
2303 spin_lock_irqsave(&conf->device_lock, flags);
2304 if (list_empty(head)) {
2305 spin_unlock_irqrestore(&conf->device_lock, flags);
2308 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2309 list_del(head->prev);
2311 spin_unlock_irqrestore(&conf->device_lock, flags);
2313 mddev = r1_bio->mddev;
2314 conf = mddev->private;
2315 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2316 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2317 test_bit(R1BIO_WriteError, &r1_bio->state))
2318 handle_sync_write_finished(conf, r1_bio);
2320 sync_request_write(mddev, r1_bio);
2321 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2322 test_bit(R1BIO_WriteError, &r1_bio->state))
2323 handle_write_finished(conf, r1_bio);
2324 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2325 handle_read_error(conf, r1_bio);
2327 /* just a partial read to be scheduled from separate
2330 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2333 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2334 md_check_recovery(mddev);
2336 blk_finish_plug(&plug);
2340 static int init_resync(struct r1conf *conf)
2344 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2345 BUG_ON(conf->r1buf_pool);
2346 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2348 if (!conf->r1buf_pool)
2350 conf->next_resync = 0;
2355 * perform a "sync" on one "block"
2357 * We need to make sure that no normal I/O request - particularly write
2358 * requests - conflict with active sync requests.
2360 * This is achieved by tracking pending requests and a 'barrier' concept
2361 * that can be installed to exclude normal IO requests.
2364 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2366 struct r1conf *conf = mddev->private;
2367 struct r1bio *r1_bio;
2369 sector_t max_sector, nr_sectors;
2373 int write_targets = 0, read_targets = 0;
2374 sector_t sync_blocks;
2375 int still_degraded = 0;
2376 int good_sectors = RESYNC_SECTORS;
2377 int min_bad = 0; /* number of sectors that are bad in all devices */
2379 if (!conf->r1buf_pool)
2380 if (init_resync(conf))
2383 max_sector = mddev->dev_sectors;
2384 if (sector_nr >= max_sector) {
2385 /* If we aborted, we need to abort the
2386 * sync on the 'current' bitmap chunk (there will
2387 * only be one in raid1 resync.
2388 * We can find the current addess in mddev->curr_resync
2390 if (mddev->curr_resync < max_sector) /* aborted */
2391 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2393 else /* completed sync */
2396 bitmap_close_sync(mddev->bitmap);
2401 if (mddev->bitmap == NULL &&
2402 mddev->recovery_cp == MaxSector &&
2403 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2404 conf->fullsync == 0) {
2406 return max_sector - sector_nr;
2408 /* before building a request, check if we can skip these blocks..
2409 * This call the bitmap_start_sync doesn't actually record anything
2411 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2412 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2413 /* We can skip this block, and probably several more */
2418 * If there is non-resync activity waiting for a turn,
2419 * and resync is going fast enough,
2420 * then let it though before starting on this new sync request.
2422 if (!go_faster && conf->nr_waiting)
2423 msleep_interruptible(1000);
2425 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2426 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2427 raise_barrier(conf);
2429 conf->next_resync = sector_nr;
2433 * If we get a correctably read error during resync or recovery,
2434 * we might want to read from a different device. So we
2435 * flag all drives that could conceivably be read from for READ,
2436 * and any others (which will be non-In_sync devices) for WRITE.
2437 * If a read fails, we try reading from something else for which READ
2441 r1_bio->mddev = mddev;
2442 r1_bio->sector = sector_nr;
2444 set_bit(R1BIO_IsSync, &r1_bio->state);
2446 for (i = 0; i < conf->raid_disks * 2; i++) {
2447 struct md_rdev *rdev;
2448 bio = r1_bio->bios[i];
2451 rdev = rcu_dereference(conf->mirrors[i].rdev);
2453 test_bit(Faulty, &rdev->flags)) {
2454 if (i < conf->raid_disks)
2456 } else if (!test_bit(In_sync, &rdev->flags)) {
2458 bio->bi_end_io = end_sync_write;
2461 /* may need to read from here */
2462 sector_t first_bad = MaxSector;
2465 if (is_badblock(rdev, sector_nr, good_sectors,
2466 &first_bad, &bad_sectors)) {
2467 if (first_bad > sector_nr)
2468 good_sectors = first_bad - sector_nr;
2470 bad_sectors -= (sector_nr - first_bad);
2472 min_bad > bad_sectors)
2473 min_bad = bad_sectors;
2476 if (sector_nr < first_bad) {
2477 if (test_bit(WriteMostly, &rdev->flags)) {
2485 bio->bi_end_io = end_sync_read;
2487 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2488 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2489 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2491 * The device is suitable for reading (InSync),
2492 * but has bad block(s) here. Let's try to correct them,
2493 * if we are doing resync or repair. Otherwise, leave
2494 * this device alone for this sync request.
2497 bio->bi_end_io = end_sync_write;
2501 if (bio->bi_end_io) {
2502 atomic_inc(&rdev->nr_pending);
2503 bio->bi_sector = sector_nr + rdev->data_offset;
2504 bio->bi_bdev = rdev->bdev;
2505 bio->bi_private = r1_bio;
2511 r1_bio->read_disk = disk;
2513 if (read_targets == 0 && min_bad > 0) {
2514 /* These sectors are bad on all InSync devices, so we
2515 * need to mark them bad on all write targets
2518 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2519 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2520 struct md_rdev *rdev = conf->mirrors[i].rdev;
2521 ok = rdev_set_badblocks(rdev, sector_nr,
2525 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2530 /* Cannot record the badblocks, so need to
2532 * If there are multiple read targets, could just
2533 * fail the really bad ones ???
2535 conf->recovery_disabled = mddev->recovery_disabled;
2536 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2542 if (min_bad > 0 && min_bad < good_sectors) {
2543 /* only resync enough to reach the next bad->good
2545 good_sectors = min_bad;
2548 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2549 /* extra read targets are also write targets */
2550 write_targets += read_targets-1;
2552 if (write_targets == 0 || read_targets == 0) {
2553 /* There is nowhere to write, so all non-sync
2554 * drives must be failed - so we are finished
2558 max_sector = sector_nr + min_bad;
2559 rv = max_sector - sector_nr;
2565 if (max_sector > mddev->resync_max)
2566 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2567 if (max_sector > sector_nr + good_sectors)
2568 max_sector = sector_nr + good_sectors;
2573 int len = PAGE_SIZE;
2574 if (sector_nr + (len>>9) > max_sector)
2575 len = (max_sector - sector_nr) << 9;
2578 if (sync_blocks == 0) {
2579 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2580 &sync_blocks, still_degraded) &&
2582 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2584 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2585 if ((len >> 9) > sync_blocks)
2586 len = sync_blocks<<9;
2589 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2590 bio = r1_bio->bios[i];
2591 if (bio->bi_end_io) {
2592 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2593 if (bio_add_page(bio, page, len, 0) == 0) {
2595 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2598 bio = r1_bio->bios[i];
2599 if (bio->bi_end_io==NULL)
2601 /* remove last page from this bio */
2603 bio->bi_size -= len;
2604 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2610 nr_sectors += len>>9;
2611 sector_nr += len>>9;
2612 sync_blocks -= (len>>9);
2613 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2615 r1_bio->sectors = nr_sectors;
2617 /* For a user-requested sync, we read all readable devices and do a
2620 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2621 atomic_set(&r1_bio->remaining, read_targets);
2622 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2623 bio = r1_bio->bios[i];
2624 if (bio->bi_end_io == end_sync_read) {
2626 md_sync_acct(bio->bi_bdev, nr_sectors);
2627 generic_make_request(bio);
2631 atomic_set(&r1_bio->remaining, 1);
2632 bio = r1_bio->bios[r1_bio->read_disk];
2633 md_sync_acct(bio->bi_bdev, nr_sectors);
2634 generic_make_request(bio);
2640 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2645 return mddev->dev_sectors;
2648 static struct r1conf *setup_conf(struct mddev *mddev)
2650 struct r1conf *conf;
2652 struct raid1_info *disk;
2653 struct md_rdev *rdev;
2656 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2660 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2661 * mddev->raid_disks * 2,
2666 conf->tmppage = alloc_page(GFP_KERNEL);
2670 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2671 if (!conf->poolinfo)
2673 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2674 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2677 if (!conf->r1bio_pool)
2680 conf->poolinfo->mddev = mddev;
2683 spin_lock_init(&conf->device_lock);
2684 rdev_for_each(rdev, mddev) {
2685 struct request_queue *q;
2686 int disk_idx = rdev->raid_disk;
2687 if (disk_idx >= mddev->raid_disks
2690 if (test_bit(Replacement, &rdev->flags))
2691 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2693 disk = conf->mirrors + disk_idx;
2698 q = bdev_get_queue(rdev->bdev);
2699 if (q->merge_bvec_fn)
2700 mddev->merge_check_needed = 1;
2702 disk->head_position = 0;
2703 disk->seq_start = MaxSector;
2705 conf->raid_disks = mddev->raid_disks;
2706 conf->mddev = mddev;
2707 INIT_LIST_HEAD(&conf->retry_list);
2709 spin_lock_init(&conf->resync_lock);
2710 init_waitqueue_head(&conf->wait_barrier);
2712 bio_list_init(&conf->pending_bio_list);
2713 conf->pending_count = 0;
2714 conf->recovery_disabled = mddev->recovery_disabled - 1;
2717 for (i = 0; i < conf->raid_disks * 2; i++) {
2719 disk = conf->mirrors + i;
2721 if (i < conf->raid_disks &&
2722 disk[conf->raid_disks].rdev) {
2723 /* This slot has a replacement. */
2725 /* No original, just make the replacement
2726 * a recovering spare
2729 disk[conf->raid_disks].rdev;
2730 disk[conf->raid_disks].rdev = NULL;
2731 } else if (!test_bit(In_sync, &disk->rdev->flags))
2732 /* Original is not in_sync - bad */
2737 !test_bit(In_sync, &disk->rdev->flags)) {
2738 disk->head_position = 0;
2740 (disk->rdev->saved_raid_disk < 0))
2746 conf->thread = md_register_thread(raid1d, mddev, "raid1");
2747 if (!conf->thread) {
2749 "md/raid1:%s: couldn't allocate thread\n",
2758 if (conf->r1bio_pool)
2759 mempool_destroy(conf->r1bio_pool);
2760 kfree(conf->mirrors);
2761 safe_put_page(conf->tmppage);
2762 kfree(conf->poolinfo);
2765 return ERR_PTR(err);
2768 static int stop(struct mddev *mddev);
2769 static int run(struct mddev *mddev)
2771 struct r1conf *conf;
2773 struct md_rdev *rdev;
2775 bool discard_supported = false;
2777 if (mddev->level != 1) {
2778 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2779 mdname(mddev), mddev->level);
2782 if (mddev->reshape_position != MaxSector) {
2783 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2788 * copy the already verified devices into our private RAID1
2789 * bookkeeping area. [whatever we allocate in run(),
2790 * should be freed in stop()]
2792 if (mddev->private == NULL)
2793 conf = setup_conf(mddev);
2795 conf = mddev->private;
2798 return PTR_ERR(conf);
2801 blk_queue_max_write_same_sectors(mddev->queue, 0);
2803 rdev_for_each(rdev, mddev) {
2804 if (!mddev->gendisk)
2806 disk_stack_limits(mddev->gendisk, rdev->bdev,
2807 rdev->data_offset << 9);
2808 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2809 discard_supported = true;
2812 mddev->degraded = 0;
2813 for (i=0; i < conf->raid_disks; i++)
2814 if (conf->mirrors[i].rdev == NULL ||
2815 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2816 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2819 if (conf->raid_disks - mddev->degraded == 1)
2820 mddev->recovery_cp = MaxSector;
2822 if (mddev->recovery_cp != MaxSector)
2823 printk(KERN_NOTICE "md/raid1:%s: not clean"
2824 " -- starting background reconstruction\n",
2827 "md/raid1:%s: active with %d out of %d mirrors\n",
2828 mdname(mddev), mddev->raid_disks - mddev->degraded,
2832 * Ok, everything is just fine now
2834 mddev->thread = conf->thread;
2835 conf->thread = NULL;
2836 mddev->private = conf;
2838 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2841 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2842 mddev->queue->backing_dev_info.congested_data = mddev;
2843 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2845 if (discard_supported)
2846 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2849 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2853 ret = md_integrity_register(mddev);
2859 static int stop(struct mddev *mddev)
2861 struct r1conf *conf = mddev->private;
2862 struct bitmap *bitmap = mddev->bitmap;
2864 /* wait for behind writes to complete */
2865 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2866 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2868 /* need to kick something here to make sure I/O goes? */
2869 wait_event(bitmap->behind_wait,
2870 atomic_read(&bitmap->behind_writes) == 0);
2873 freeze_array(conf, 0);
2874 unfreeze_array(conf);
2876 md_unregister_thread(&mddev->thread);
2877 if (conf->r1bio_pool)
2878 mempool_destroy(conf->r1bio_pool);
2879 kfree(conf->mirrors);
2880 safe_put_page(conf->tmppage);
2881 kfree(conf->poolinfo);
2883 mddev->private = NULL;
2887 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2889 /* no resync is happening, and there is enough space
2890 * on all devices, so we can resize.
2891 * We need to make sure resync covers any new space.
2892 * If the array is shrinking we should possibly wait until
2893 * any io in the removed space completes, but it hardly seems
2896 sector_t newsize = raid1_size(mddev, sectors, 0);
2897 if (mddev->external_size &&
2898 mddev->array_sectors > newsize)
2900 if (mddev->bitmap) {
2901 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2905 md_set_array_sectors(mddev, newsize);
2906 set_capacity(mddev->gendisk, mddev->array_sectors);
2907 revalidate_disk(mddev->gendisk);
2908 if (sectors > mddev->dev_sectors &&
2909 mddev->recovery_cp > mddev->dev_sectors) {
2910 mddev->recovery_cp = mddev->dev_sectors;
2911 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2913 mddev->dev_sectors = sectors;
2914 mddev->resync_max_sectors = sectors;
2918 static int raid1_reshape(struct mddev *mddev)
2921 * 1/ resize the r1bio_pool
2922 * 2/ resize conf->mirrors
2924 * We allocate a new r1bio_pool if we can.
2925 * Then raise a device barrier and wait until all IO stops.
2926 * Then resize conf->mirrors and swap in the new r1bio pool.
2928 * At the same time, we "pack" the devices so that all the missing
2929 * devices have the higher raid_disk numbers.
2931 mempool_t *newpool, *oldpool;
2932 struct pool_info *newpoolinfo;
2933 struct raid1_info *newmirrors;
2934 struct r1conf *conf = mddev->private;
2935 int cnt, raid_disks;
2936 unsigned long flags;
2939 /* Cannot change chunk_size, layout, or level */
2940 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2941 mddev->layout != mddev->new_layout ||
2942 mddev->level != mddev->new_level) {
2943 mddev->new_chunk_sectors = mddev->chunk_sectors;
2944 mddev->new_layout = mddev->layout;
2945 mddev->new_level = mddev->level;
2949 err = md_allow_write(mddev);
2953 raid_disks = mddev->raid_disks + mddev->delta_disks;
2955 if (raid_disks < conf->raid_disks) {
2957 for (d= 0; d < conf->raid_disks; d++)
2958 if (conf->mirrors[d].rdev)
2960 if (cnt > raid_disks)
2964 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2967 newpoolinfo->mddev = mddev;
2968 newpoolinfo->raid_disks = raid_disks * 2;
2970 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2971 r1bio_pool_free, newpoolinfo);
2976 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2980 mempool_destroy(newpool);
2984 freeze_array(conf, 0);
2986 /* ok, everything is stopped */
2987 oldpool = conf->r1bio_pool;
2988 conf->r1bio_pool = newpool;
2990 for (d = d2 = 0; d < conf->raid_disks; d++) {
2991 struct md_rdev *rdev = conf->mirrors[d].rdev;
2992 if (rdev && rdev->raid_disk != d2) {
2993 sysfs_unlink_rdev(mddev, rdev);
2994 rdev->raid_disk = d2;
2995 sysfs_unlink_rdev(mddev, rdev);
2996 if (sysfs_link_rdev(mddev, rdev))
2998 "md/raid1:%s: cannot register rd%d\n",
2999 mdname(mddev), rdev->raid_disk);
3002 newmirrors[d2++].rdev = rdev;
3004 kfree(conf->mirrors);
3005 conf->mirrors = newmirrors;
3006 kfree(conf->poolinfo);
3007 conf->poolinfo = newpoolinfo;
3009 spin_lock_irqsave(&conf->device_lock, flags);
3010 mddev->degraded += (raid_disks - conf->raid_disks);
3011 spin_unlock_irqrestore(&conf->device_lock, flags);
3012 conf->raid_disks = mddev->raid_disks = raid_disks;
3013 mddev->delta_disks = 0;
3015 unfreeze_array(conf);
3017 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3018 md_wakeup_thread(mddev->thread);
3020 mempool_destroy(oldpool);
3024 static void raid1_quiesce(struct mddev *mddev, int state)
3026 struct r1conf *conf = mddev->private;
3029 case 2: /* wake for suspend */
3030 wake_up(&conf->wait_barrier);
3033 freeze_array(conf, 0);
3036 unfreeze_array(conf);
3041 static void *raid1_takeover(struct mddev *mddev)
3043 /* raid1 can take over:
3044 * raid5 with 2 devices, any layout or chunk size
3046 if (mddev->level == 5 && mddev->raid_disks == 2) {
3047 struct r1conf *conf;
3048 mddev->new_level = 1;
3049 mddev->new_layout = 0;
3050 mddev->new_chunk_sectors = 0;
3051 conf = setup_conf(mddev);
3056 return ERR_PTR(-EINVAL);
3059 static struct md_personality raid1_personality =
3063 .owner = THIS_MODULE,
3064 .make_request = make_request,
3068 .error_handler = error,
3069 .hot_add_disk = raid1_add_disk,
3070 .hot_remove_disk= raid1_remove_disk,
3071 .spare_active = raid1_spare_active,
3072 .sync_request = sync_request,
3073 .resize = raid1_resize,
3075 .check_reshape = raid1_reshape,
3076 .quiesce = raid1_quiesce,
3077 .takeover = raid1_takeover,
3080 static int __init raid_init(void)
3082 return register_md_personality(&raid1_personality);
3085 static void raid_exit(void)
3087 unregister_md_personality(&raid1_personality);
3090 module_init(raid_init);
3091 module_exit(raid_exit);
3092 MODULE_LICENSE("GPL");
3093 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3094 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3095 MODULE_ALIAS("md-raid1");
3096 MODULE_ALIAS("md-level-1");
3098 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);