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_BLOCK_SIZE PAGE_SIZE
88 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
90 #define RESYNC_WINDOW (2048*1024)
92 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
94 struct pool_info *pi = data;
99 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
104 * Allocate bios : 1 for reading, n-1 for writing
106 for (j = pi->raid_disks ; j-- ; ) {
107 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
110 r1_bio->bios[j] = bio;
113 * Allocate RESYNC_PAGES data pages and attach them to
115 * If this is a user-requested check/repair, allocate
116 * RESYNC_PAGES for each bio.
118 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
123 bio = r1_bio->bios[j];
124 bio->bi_vcnt = RESYNC_PAGES;
126 if (bio_alloc_pages(bio, gfp_flags))
129 /* If not user-requests, copy the page pointers to all bios */
130 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
131 for (i=0; i<RESYNC_PAGES ; i++)
132 for (j=1; j<pi->raid_disks; j++)
133 r1_bio->bios[j]->bi_io_vec[i].bv_page =
134 r1_bio->bios[0]->bi_io_vec[i].bv_page;
137 r1_bio->master_bio = NULL;
142 while (++j < pi->raid_disks)
143 bio_put(r1_bio->bios[j]);
144 r1bio_pool_free(r1_bio, data);
148 static void r1buf_pool_free(void *__r1_bio, void *data)
150 struct pool_info *pi = data;
152 struct r1bio *r1bio = __r1_bio;
154 for (i = 0; i < RESYNC_PAGES; i++)
155 for (j = pi->raid_disks; j-- ;) {
157 r1bio->bios[j]->bi_io_vec[i].bv_page !=
158 r1bio->bios[0]->bi_io_vec[i].bv_page)
159 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
161 for (i=0 ; i < pi->raid_disks; i++)
162 bio_put(r1bio->bios[i]);
164 r1bio_pool_free(r1bio, data);
167 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
171 for (i = 0; i < conf->raid_disks * 2; i++) {
172 struct bio **bio = r1_bio->bios + i;
173 if (!BIO_SPECIAL(*bio))
179 static void free_r1bio(struct r1bio *r1_bio)
181 struct r1conf *conf = r1_bio->mddev->private;
183 put_all_bios(conf, r1_bio);
184 mempool_free(r1_bio, conf->r1bio_pool);
187 static void put_buf(struct r1bio *r1_bio)
189 struct r1conf *conf = r1_bio->mddev->private;
192 for (i = 0; i < conf->raid_disks * 2; i++) {
193 struct bio *bio = r1_bio->bios[i];
195 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
198 mempool_free(r1_bio, conf->r1buf_pool);
203 static void reschedule_retry(struct r1bio *r1_bio)
206 struct mddev *mddev = r1_bio->mddev;
207 struct r1conf *conf = mddev->private;
209 spin_lock_irqsave(&conf->device_lock, flags);
210 list_add(&r1_bio->retry_list, &conf->retry_list);
212 spin_unlock_irqrestore(&conf->device_lock, flags);
214 wake_up(&conf->wait_barrier);
215 md_wakeup_thread(mddev->thread);
219 * raid_end_bio_io() is called when we have finished servicing a mirrored
220 * operation and are ready to return a success/failure code to the buffer
223 static void call_bio_endio(struct r1bio *r1_bio)
225 struct bio *bio = r1_bio->master_bio;
227 struct r1conf *conf = r1_bio->mddev->private;
229 if (bio->bi_phys_segments) {
231 spin_lock_irqsave(&conf->device_lock, flags);
232 bio->bi_phys_segments--;
233 done = (bio->bi_phys_segments == 0);
234 spin_unlock_irqrestore(&conf->device_lock, flags);
238 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
239 clear_bit(BIO_UPTODATE, &bio->bi_flags);
243 * Wake up any possible resync thread that waits for the device
250 static void raid_end_bio_io(struct r1bio *r1_bio)
252 struct bio *bio = r1_bio->master_bio;
254 /* if nobody has done the final endio yet, do it now */
255 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
256 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
257 (bio_data_dir(bio) == WRITE) ? "write" : "read",
258 (unsigned long long) bio->bi_sector,
259 (unsigned long long) bio->bi_sector +
260 bio_sectors(bio) - 1);
262 call_bio_endio(r1_bio);
268 * Update disk head position estimator based on IRQ completion info.
270 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
272 struct r1conf *conf = r1_bio->mddev->private;
274 conf->mirrors[disk].head_position =
275 r1_bio->sector + (r1_bio->sectors);
279 * Find the disk number which triggered given bio
281 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
284 struct r1conf *conf = r1_bio->mddev->private;
285 int raid_disks = conf->raid_disks;
287 for (mirror = 0; mirror < raid_disks * 2; mirror++)
288 if (r1_bio->bios[mirror] == bio)
291 BUG_ON(mirror == raid_disks * 2);
292 update_head_pos(mirror, r1_bio);
297 static void raid1_end_read_request(struct bio *bio, int error)
299 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
300 struct r1bio *r1_bio = bio->bi_private;
302 struct r1conf *conf = r1_bio->mddev->private;
304 mirror = r1_bio->read_disk;
306 * this branch is our 'one mirror IO has finished' event handler:
308 update_head_pos(mirror, r1_bio);
311 set_bit(R1BIO_Uptodate, &r1_bio->state);
313 /* If all other devices have failed, we want to return
314 * the error upwards rather than fail the last device.
315 * Here we redefine "uptodate" to mean "Don't want to retry"
318 spin_lock_irqsave(&conf->device_lock, flags);
319 if (r1_bio->mddev->degraded == conf->raid_disks ||
320 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
321 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
323 spin_unlock_irqrestore(&conf->device_lock, flags);
327 raid_end_bio_io(r1_bio);
328 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
333 char b[BDEVNAME_SIZE];
335 KERN_ERR "md/raid1:%s: %s: "
336 "rescheduling sector %llu\n",
338 bdevname(conf->mirrors[mirror].rdev->bdev,
340 (unsigned long long)r1_bio->sector);
341 set_bit(R1BIO_ReadError, &r1_bio->state);
342 reschedule_retry(r1_bio);
343 /* don't drop the reference on read_disk yet */
347 static void close_write(struct r1bio *r1_bio)
349 /* it really is the end of this request */
350 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
351 /* free extra copy of the data pages */
352 int i = r1_bio->behind_page_count;
354 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
355 kfree(r1_bio->behind_bvecs);
356 r1_bio->behind_bvecs = NULL;
358 /* clear the bitmap if all writes complete successfully */
359 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
361 !test_bit(R1BIO_Degraded, &r1_bio->state),
362 test_bit(R1BIO_BehindIO, &r1_bio->state));
363 md_write_end(r1_bio->mddev);
366 static void r1_bio_write_done(struct r1bio *r1_bio)
368 if (!atomic_dec_and_test(&r1_bio->remaining))
371 if (test_bit(R1BIO_WriteError, &r1_bio->state))
372 reschedule_retry(r1_bio);
375 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
376 reschedule_retry(r1_bio);
378 raid_end_bio_io(r1_bio);
382 static void raid1_end_write_request(struct bio *bio, int error)
384 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
385 struct r1bio *r1_bio = bio->bi_private;
386 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
387 struct r1conf *conf = r1_bio->mddev->private;
388 struct bio *to_put = NULL;
390 mirror = find_bio_disk(r1_bio, bio);
393 * 'one mirror IO has finished' event handler:
396 set_bit(WriteErrorSeen,
397 &conf->mirrors[mirror].rdev->flags);
398 if (!test_and_set_bit(WantReplacement,
399 &conf->mirrors[mirror].rdev->flags))
400 set_bit(MD_RECOVERY_NEEDED, &
401 conf->mddev->recovery);
403 set_bit(R1BIO_WriteError, &r1_bio->state);
406 * Set R1BIO_Uptodate in our master bio, so that we
407 * will return a good error code for to the higher
408 * levels even if IO on some other mirrored buffer
411 * The 'master' represents the composite IO operation
412 * to user-side. So if something waits for IO, then it
413 * will wait for the 'master' bio.
418 r1_bio->bios[mirror] = NULL;
421 * Do not set R1BIO_Uptodate if the current device is
422 * rebuilding or Faulty. This is because we cannot use
423 * such device for properly reading the data back (we could
424 * potentially use it, if the current write would have felt
425 * before rdev->recovery_offset, but for simplicity we don't
428 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
429 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
430 set_bit(R1BIO_Uptodate, &r1_bio->state);
432 /* Maybe we can clear some bad blocks. */
433 if (is_badblock(conf->mirrors[mirror].rdev,
434 r1_bio->sector, r1_bio->sectors,
435 &first_bad, &bad_sectors)) {
436 r1_bio->bios[mirror] = IO_MADE_GOOD;
437 set_bit(R1BIO_MadeGood, &r1_bio->state);
442 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
443 atomic_dec(&r1_bio->behind_remaining);
446 * In behind mode, we ACK the master bio once the I/O
447 * has safely reached all non-writemostly
448 * disks. Setting the Returned bit ensures that this
449 * gets done only once -- we don't ever want to return
450 * -EIO here, instead we'll wait
452 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
453 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
454 /* Maybe we can return now */
455 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
456 struct bio *mbio = r1_bio->master_bio;
457 pr_debug("raid1: behind end write sectors"
459 (unsigned long long) mbio->bi_sector,
460 (unsigned long long) mbio->bi_sector +
461 bio_sectors(mbio) - 1);
462 call_bio_endio(r1_bio);
466 if (r1_bio->bios[mirror] == NULL)
467 rdev_dec_pending(conf->mirrors[mirror].rdev,
471 * Let's see if all mirrored write operations have finished
474 r1_bio_write_done(r1_bio);
482 * This routine returns the disk from which the requested read should
483 * be done. There is a per-array 'next expected sequential IO' sector
484 * number - if this matches on the next IO then we use the last disk.
485 * There is also a per-disk 'last know head position' sector that is
486 * maintained from IRQ contexts, both the normal and the resync IO
487 * completion handlers update this position correctly. If there is no
488 * perfect sequential match then we pick the disk whose head is closest.
490 * If there are 2 mirrors in the same 2 devices, performance degrades
491 * because position is mirror, not device based.
493 * The rdev for the device selected will have nr_pending incremented.
495 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
497 const sector_t this_sector = r1_bio->sector;
499 int best_good_sectors;
500 int best_disk, best_dist_disk, best_pending_disk;
504 unsigned int min_pending;
505 struct md_rdev *rdev;
507 int choose_next_idle;
511 * Check if we can balance. We can balance on the whole
512 * device if no resync is going on, or below the resync window.
513 * We take the first readable disk when above the resync window.
516 sectors = r1_bio->sectors;
519 best_dist = MaxSector;
520 best_pending_disk = -1;
521 min_pending = UINT_MAX;
522 best_good_sectors = 0;
524 choose_next_idle = 0;
526 if (conf->mddev->recovery_cp < MaxSector &&
527 (this_sector + sectors >= conf->next_resync))
532 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
536 unsigned int pending;
539 rdev = rcu_dereference(conf->mirrors[disk].rdev);
540 if (r1_bio->bios[disk] == IO_BLOCKED
542 || test_bit(Unmerged, &rdev->flags)
543 || test_bit(Faulty, &rdev->flags))
545 if (!test_bit(In_sync, &rdev->flags) &&
546 rdev->recovery_offset < this_sector + sectors)
548 if (test_bit(WriteMostly, &rdev->flags)) {
549 /* Don't balance among write-mostly, just
550 * use the first as a last resort */
552 if (is_badblock(rdev, this_sector, sectors,
553 &first_bad, &bad_sectors)) {
554 if (first_bad < this_sector)
555 /* Cannot use this */
557 best_good_sectors = first_bad - this_sector;
559 best_good_sectors = sectors;
564 /* This is a reasonable device to use. It might
567 if (is_badblock(rdev, this_sector, sectors,
568 &first_bad, &bad_sectors)) {
569 if (best_dist < MaxSector)
570 /* already have a better device */
572 if (first_bad <= this_sector) {
573 /* cannot read here. If this is the 'primary'
574 * device, then we must not read beyond
575 * bad_sectors from another device..
577 bad_sectors -= (this_sector - first_bad);
578 if (choose_first && sectors > bad_sectors)
579 sectors = bad_sectors;
580 if (best_good_sectors > sectors)
581 best_good_sectors = sectors;
584 sector_t good_sectors = first_bad - this_sector;
585 if (good_sectors > best_good_sectors) {
586 best_good_sectors = good_sectors;
594 best_good_sectors = sectors;
596 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
597 has_nonrot_disk |= nonrot;
598 pending = atomic_read(&rdev->nr_pending);
599 dist = abs(this_sector - conf->mirrors[disk].head_position);
604 /* Don't change to another disk for sequential reads */
605 if (conf->mirrors[disk].next_seq_sect == this_sector
607 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
608 struct raid1_info *mirror = &conf->mirrors[disk];
612 * If buffered sequential IO size exceeds optimal
613 * iosize, check if there is idle disk. If yes, choose
614 * the idle disk. read_balance could already choose an
615 * idle disk before noticing it's a sequential IO in
616 * this disk. This doesn't matter because this disk
617 * will idle, next time it will be utilized after the
618 * first disk has IO size exceeds optimal iosize. In
619 * this way, iosize of the first disk will be optimal
620 * iosize at least. iosize of the second disk might be
621 * small, but not a big deal since when the second disk
622 * starts IO, the first disk is likely still busy.
624 if (nonrot && opt_iosize > 0 &&
625 mirror->seq_start != MaxSector &&
626 mirror->next_seq_sect > opt_iosize &&
627 mirror->next_seq_sect - opt_iosize >=
629 choose_next_idle = 1;
634 /* If device is idle, use it */
640 if (choose_next_idle)
643 if (min_pending > pending) {
644 min_pending = pending;
645 best_pending_disk = disk;
648 if (dist < best_dist) {
650 best_dist_disk = disk;
655 * If all disks are rotational, choose the closest disk. If any disk is
656 * non-rotational, choose the disk with less pending request even the
657 * disk is rotational, which might/might not be optimal for raids with
658 * mixed ratation/non-rotational disks depending on workload.
660 if (best_disk == -1) {
662 best_disk = best_pending_disk;
664 best_disk = best_dist_disk;
667 if (best_disk >= 0) {
668 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
671 atomic_inc(&rdev->nr_pending);
672 if (test_bit(Faulty, &rdev->flags)) {
673 /* cannot risk returning a device that failed
674 * before we inc'ed nr_pending
676 rdev_dec_pending(rdev, conf->mddev);
679 sectors = best_good_sectors;
681 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
682 conf->mirrors[best_disk].seq_start = this_sector;
684 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
687 *max_sectors = sectors;
692 static int raid1_mergeable_bvec(struct request_queue *q,
693 struct bvec_merge_data *bvm,
694 struct bio_vec *biovec)
696 struct mddev *mddev = q->queuedata;
697 struct r1conf *conf = mddev->private;
698 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
699 int max = biovec->bv_len;
701 if (mddev->merge_check_needed) {
704 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
705 struct md_rdev *rdev = rcu_dereference(
706 conf->mirrors[disk].rdev);
707 if (rdev && !test_bit(Faulty, &rdev->flags)) {
708 struct request_queue *q =
709 bdev_get_queue(rdev->bdev);
710 if (q->merge_bvec_fn) {
711 bvm->bi_sector = sector +
713 bvm->bi_bdev = rdev->bdev;
714 max = min(max, q->merge_bvec_fn(
725 int md_raid1_congested(struct mddev *mddev, int bits)
727 struct r1conf *conf = mddev->private;
730 if ((bits & (1 << BDI_async_congested)) &&
731 conf->pending_count >= max_queued_requests)
735 for (i = 0; i < conf->raid_disks * 2; i++) {
736 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
737 if (rdev && !test_bit(Faulty, &rdev->flags)) {
738 struct request_queue *q = bdev_get_queue(rdev->bdev);
742 /* Note the '|| 1' - when read_balance prefers
743 * non-congested targets, it can be removed
745 if ((bits & (1<<BDI_async_congested)) || 1)
746 ret |= bdi_congested(&q->backing_dev_info, bits);
748 ret &= bdi_congested(&q->backing_dev_info, bits);
754 EXPORT_SYMBOL_GPL(md_raid1_congested);
756 static int raid1_congested(void *data, int bits)
758 struct mddev *mddev = data;
760 return mddev_congested(mddev, bits) ||
761 md_raid1_congested(mddev, bits);
764 static void flush_pending_writes(struct r1conf *conf)
766 /* Any writes that have been queued but are awaiting
767 * bitmap updates get flushed here.
769 spin_lock_irq(&conf->device_lock);
771 if (conf->pending_bio_list.head) {
773 bio = bio_list_get(&conf->pending_bio_list);
774 conf->pending_count = 0;
775 spin_unlock_irq(&conf->device_lock);
776 /* flush any pending bitmap writes to
777 * disk before proceeding w/ I/O */
778 bitmap_unplug(conf->mddev->bitmap);
779 wake_up(&conf->wait_barrier);
781 while (bio) { /* submit pending writes */
782 struct bio *next = bio->bi_next;
784 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
785 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
789 generic_make_request(bio);
793 spin_unlock_irq(&conf->device_lock);
797 * Sometimes we need to suspend IO while we do something else,
798 * either some resync/recovery, or reconfigure the array.
799 * To do this we raise a 'barrier'.
800 * The 'barrier' is a counter that can be raised multiple times
801 * to count how many activities are happening which preclude
803 * We can only raise the barrier if there is no pending IO.
804 * i.e. if nr_pending == 0.
805 * We choose only to raise the barrier if no-one is waiting for the
806 * barrier to go down. This means that as soon as an IO request
807 * is ready, no other operations which require a barrier will start
808 * until the IO request has had a chance.
810 * So: regular IO calls 'wait_barrier'. When that returns there
811 * is no backgroup IO happening, It must arrange to call
812 * allow_barrier when it has finished its IO.
813 * backgroup IO calls must call raise_barrier. Once that returns
814 * there is no normal IO happeing. It must arrange to call
815 * lower_barrier when the particular background IO completes.
817 #define RESYNC_DEPTH 32
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->nr_pending && conf->barrier < RESYNC_DEPTH,
835 spin_unlock_irq(&conf->resync_lock);
838 static void lower_barrier(struct r1conf *conf)
841 BUG_ON(conf->barrier <= 0);
842 spin_lock_irqsave(&conf->resync_lock, flags);
844 spin_unlock_irqrestore(&conf->resync_lock, flags);
845 wake_up(&conf->wait_barrier);
848 static void wait_barrier(struct r1conf *conf)
850 spin_lock_irq(&conf->resync_lock);
853 /* Wait for the barrier to drop.
854 * However if there are already pending
855 * requests (preventing the barrier from
856 * rising completely), and the
857 * pre-process bio queue isn't empty,
858 * then don't wait, as we need to empty
859 * that queue to get the nr_pending
862 wait_event_lock_irq(conf->wait_barrier,
866 !bio_list_empty(current->bio_list)),
871 spin_unlock_irq(&conf->resync_lock);
874 static void allow_barrier(struct r1conf *conf)
877 spin_lock_irqsave(&conf->resync_lock, flags);
879 spin_unlock_irqrestore(&conf->resync_lock, flags);
880 wake_up(&conf->wait_barrier);
883 static void freeze_array(struct r1conf *conf, int extra)
885 /* stop syncio and normal IO and wait for everything to
887 * We increment barrier and nr_waiting, and then
888 * wait until nr_pending match nr_queued+extra
889 * This is called in the context of one normal IO request
890 * that has failed. Thus any sync request that might be pending
891 * will be blocked by nr_pending, and we need to wait for
892 * pending IO requests to complete or be queued for re-try.
893 * Thus the number queued (nr_queued) plus this request (extra)
894 * must match the number of pending IOs (nr_pending) before
897 spin_lock_irq(&conf->resync_lock);
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);
912 wake_up(&conf->wait_barrier);
913 spin_unlock_irq(&conf->resync_lock);
917 /* duplicate the data pages for behind I/O
919 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
922 struct bio_vec *bvec;
923 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
925 if (unlikely(!bvecs))
928 bio_for_each_segment_all(bvec, bio, i) {
930 bvecs[i].bv_page = alloc_page(GFP_NOIO);
931 if (unlikely(!bvecs[i].bv_page))
933 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
934 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
935 kunmap(bvecs[i].bv_page);
936 kunmap(bvec->bv_page);
938 r1_bio->behind_bvecs = bvecs;
939 r1_bio->behind_page_count = bio->bi_vcnt;
940 set_bit(R1BIO_BehindIO, &r1_bio->state);
944 for (i = 0; i < bio->bi_vcnt; i++)
945 if (bvecs[i].bv_page)
946 put_page(bvecs[i].bv_page);
948 pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
951 struct raid1_plug_cb {
952 struct blk_plug_cb cb;
953 struct bio_list pending;
957 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
959 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
961 struct mddev *mddev = plug->cb.data;
962 struct r1conf *conf = mddev->private;
965 if (from_schedule || current->bio_list) {
966 spin_lock_irq(&conf->device_lock);
967 bio_list_merge(&conf->pending_bio_list, &plug->pending);
968 conf->pending_count += plug->pending_cnt;
969 spin_unlock_irq(&conf->device_lock);
970 wake_up(&conf->wait_barrier);
971 md_wakeup_thread(mddev->thread);
976 /* we aren't scheduling, so we can do the write-out directly. */
977 bio = bio_list_get(&plug->pending);
978 bitmap_unplug(mddev->bitmap);
979 wake_up(&conf->wait_barrier);
981 while (bio) { /* submit pending writes */
982 struct bio *next = bio->bi_next;
984 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
985 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
989 generic_make_request(bio);
995 static void make_request(struct mddev *mddev, struct bio * bio)
997 struct r1conf *conf = mddev->private;
998 struct raid1_info *mirror;
999 struct r1bio *r1_bio;
1000 struct bio *read_bio;
1002 struct bitmap *bitmap;
1003 unsigned long flags;
1004 const int rw = bio_data_dir(bio);
1005 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1006 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1007 const unsigned long do_discard = (bio->bi_rw
1008 & (REQ_DISCARD | REQ_SECURE));
1009 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1010 struct md_rdev *blocked_rdev;
1011 struct blk_plug_cb *cb;
1012 struct raid1_plug_cb *plug = NULL;
1014 int sectors_handled;
1018 * Register the new request and wait if the reconstruction
1019 * thread has put up a bar for new requests.
1020 * Continue immediately if no resync is active currently.
1023 md_write_start(mddev, bio); /* wait on superblock update early */
1025 if (bio_data_dir(bio) == WRITE &&
1026 bio_end_sector(bio) > mddev->suspend_lo &&
1027 bio->bi_sector < mddev->suspend_hi) {
1028 /* As the suspend_* range is controlled by
1029 * userspace, we want an interruptible
1034 flush_signals(current);
1035 prepare_to_wait(&conf->wait_barrier,
1036 &w, TASK_INTERRUPTIBLE);
1037 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1038 bio->bi_sector >= mddev->suspend_hi)
1042 finish_wait(&conf->wait_barrier, &w);
1047 bitmap = mddev->bitmap;
1050 * make_request() can abort the operation when READA is being
1051 * used and no empty request is available.
1054 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1056 r1_bio->master_bio = bio;
1057 r1_bio->sectors = bio_sectors(bio);
1059 r1_bio->mddev = mddev;
1060 r1_bio->sector = bio->bi_sector;
1062 /* We might need to issue multiple reads to different
1063 * devices if there are bad blocks around, so we keep
1064 * track of the number of reads in bio->bi_phys_segments.
1065 * If this is 0, there is only one r1_bio and no locking
1066 * will be needed when requests complete. If it is
1067 * non-zero, then it is the number of not-completed requests.
1069 bio->bi_phys_segments = 0;
1070 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1074 * read balancing logic:
1079 rdisk = read_balance(conf, r1_bio, &max_sectors);
1082 /* couldn't find anywhere to read from */
1083 raid_end_bio_io(r1_bio);
1086 mirror = conf->mirrors + rdisk;
1088 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1090 /* Reading from a write-mostly device must
1091 * take care not to over-take any writes
1094 wait_event(bitmap->behind_wait,
1095 atomic_read(&bitmap->behind_writes) == 0);
1097 r1_bio->read_disk = rdisk;
1099 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1100 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1103 r1_bio->bios[rdisk] = read_bio;
1105 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1106 read_bio->bi_bdev = mirror->rdev->bdev;
1107 read_bio->bi_end_io = raid1_end_read_request;
1108 read_bio->bi_rw = READ | do_sync;
1109 read_bio->bi_private = r1_bio;
1111 if (max_sectors < r1_bio->sectors) {
1112 /* could not read all from this device, so we will
1113 * need another r1_bio.
1116 sectors_handled = (r1_bio->sector + max_sectors
1118 r1_bio->sectors = max_sectors;
1119 spin_lock_irq(&conf->device_lock);
1120 if (bio->bi_phys_segments == 0)
1121 bio->bi_phys_segments = 2;
1123 bio->bi_phys_segments++;
1124 spin_unlock_irq(&conf->device_lock);
1125 /* Cannot call generic_make_request directly
1126 * as that will be queued in __make_request
1127 * and subsequent mempool_alloc might block waiting
1128 * for it. So hand bio over to raid1d.
1130 reschedule_retry(r1_bio);
1132 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1134 r1_bio->master_bio = bio;
1135 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1137 r1_bio->mddev = mddev;
1138 r1_bio->sector = bio->bi_sector + sectors_handled;
1141 generic_make_request(read_bio);
1148 if (conf->pending_count >= max_queued_requests) {
1149 md_wakeup_thread(mddev->thread);
1150 wait_event(conf->wait_barrier,
1151 conf->pending_count < max_queued_requests);
1153 /* first select target devices under rcu_lock and
1154 * inc refcount on their rdev. Record them by setting
1156 * If there are known/acknowledged bad blocks on any device on
1157 * which we have seen a write error, we want to avoid writing those
1159 * This potentially requires several writes to write around
1160 * the bad blocks. Each set of writes gets it's own r1bio
1161 * with a set of bios attached.
1164 disks = conf->raid_disks * 2;
1166 blocked_rdev = NULL;
1168 max_sectors = r1_bio->sectors;
1169 for (i = 0; i < disks; i++) {
1170 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1171 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1172 atomic_inc(&rdev->nr_pending);
1173 blocked_rdev = rdev;
1176 r1_bio->bios[i] = NULL;
1177 if (!rdev || test_bit(Faulty, &rdev->flags)
1178 || test_bit(Unmerged, &rdev->flags)) {
1179 if (i < conf->raid_disks)
1180 set_bit(R1BIO_Degraded, &r1_bio->state);
1184 atomic_inc(&rdev->nr_pending);
1185 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1190 is_bad = is_badblock(rdev, r1_bio->sector,
1192 &first_bad, &bad_sectors);
1194 /* mustn't write here until the bad block is
1196 set_bit(BlockedBadBlocks, &rdev->flags);
1197 blocked_rdev = rdev;
1200 if (is_bad && first_bad <= r1_bio->sector) {
1201 /* Cannot write here at all */
1202 bad_sectors -= (r1_bio->sector - first_bad);
1203 if (bad_sectors < max_sectors)
1204 /* mustn't write more than bad_sectors
1205 * to other devices yet
1207 max_sectors = bad_sectors;
1208 rdev_dec_pending(rdev, mddev);
1209 /* We don't set R1BIO_Degraded as that
1210 * only applies if the disk is
1211 * missing, so it might be re-added,
1212 * and we want to know to recover this
1214 * In this case the device is here,
1215 * and the fact that this chunk is not
1216 * in-sync is recorded in the bad
1222 int good_sectors = first_bad - r1_bio->sector;
1223 if (good_sectors < max_sectors)
1224 max_sectors = good_sectors;
1227 r1_bio->bios[i] = bio;
1231 if (unlikely(blocked_rdev)) {
1232 /* Wait for this device to become unblocked */
1235 for (j = 0; j < i; j++)
1236 if (r1_bio->bios[j])
1237 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1239 allow_barrier(conf);
1240 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1245 if (max_sectors < r1_bio->sectors) {
1246 /* We are splitting this write into multiple parts, so
1247 * we need to prepare for allocating another r1_bio.
1249 r1_bio->sectors = max_sectors;
1250 spin_lock_irq(&conf->device_lock);
1251 if (bio->bi_phys_segments == 0)
1252 bio->bi_phys_segments = 2;
1254 bio->bi_phys_segments++;
1255 spin_unlock_irq(&conf->device_lock);
1257 sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1259 atomic_set(&r1_bio->remaining, 1);
1260 atomic_set(&r1_bio->behind_remaining, 0);
1263 for (i = 0; i < disks; i++) {
1265 if (!r1_bio->bios[i])
1268 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1269 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1273 * Not if there are too many, or cannot
1274 * allocate memory, or a reader on WriteMostly
1275 * is waiting for behind writes to flush */
1277 (atomic_read(&bitmap->behind_writes)
1278 < mddev->bitmap_info.max_write_behind) &&
1279 !waitqueue_active(&bitmap->behind_wait))
1280 alloc_behind_pages(mbio, r1_bio);
1282 bitmap_startwrite(bitmap, r1_bio->sector,
1284 test_bit(R1BIO_BehindIO,
1288 if (r1_bio->behind_bvecs) {
1289 struct bio_vec *bvec;
1293 * We trimmed the bio, so _all is legit
1295 bio_for_each_segment_all(bvec, mbio, j)
1296 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1297 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1298 atomic_inc(&r1_bio->behind_remaining);
1301 r1_bio->bios[i] = mbio;
1303 mbio->bi_sector = (r1_bio->sector +
1304 conf->mirrors[i].rdev->data_offset);
1305 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1306 mbio->bi_end_io = raid1_end_write_request;
1308 WRITE | do_flush_fua | do_sync | do_discard | do_same;
1309 mbio->bi_private = r1_bio;
1311 atomic_inc(&r1_bio->remaining);
1313 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1315 plug = container_of(cb, struct raid1_plug_cb, cb);
1318 spin_lock_irqsave(&conf->device_lock, flags);
1320 bio_list_add(&plug->pending, mbio);
1321 plug->pending_cnt++;
1323 bio_list_add(&conf->pending_bio_list, mbio);
1324 conf->pending_count++;
1326 spin_unlock_irqrestore(&conf->device_lock, flags);
1328 md_wakeup_thread(mddev->thread);
1330 /* Mustn't call r1_bio_write_done before this next test,
1331 * as it could result in the bio being freed.
1333 if (sectors_handled < bio_sectors(bio)) {
1334 r1_bio_write_done(r1_bio);
1335 /* We need another r1_bio. It has already been counted
1336 * in bio->bi_phys_segments
1338 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1339 r1_bio->master_bio = bio;
1340 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1342 r1_bio->mddev = mddev;
1343 r1_bio->sector = bio->bi_sector + sectors_handled;
1347 r1_bio_write_done(r1_bio);
1349 /* In case raid1d snuck in to freeze_array */
1350 wake_up(&conf->wait_barrier);
1353 static void status(struct seq_file *seq, struct mddev *mddev)
1355 struct r1conf *conf = mddev->private;
1358 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1359 conf->raid_disks - mddev->degraded);
1361 for (i = 0; i < conf->raid_disks; i++) {
1362 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1363 seq_printf(seq, "%s",
1364 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1367 seq_printf(seq, "]");
1371 static void error(struct mddev *mddev, struct md_rdev *rdev)
1373 char b[BDEVNAME_SIZE];
1374 struct r1conf *conf = mddev->private;
1377 * If it is not operational, then we have already marked it as dead
1378 * else if it is the last working disks, ignore the error, let the
1379 * next level up know.
1380 * else mark the drive as failed
1382 if (test_bit(In_sync, &rdev->flags)
1383 && (conf->raid_disks - mddev->degraded) == 1) {
1385 * Don't fail the drive, act as though we were just a
1386 * normal single drive.
1387 * However don't try a recovery from this drive as
1388 * it is very likely to fail.
1390 conf->recovery_disabled = mddev->recovery_disabled;
1393 set_bit(Blocked, &rdev->flags);
1394 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1395 unsigned long flags;
1396 spin_lock_irqsave(&conf->device_lock, flags);
1398 set_bit(Faulty, &rdev->flags);
1399 spin_unlock_irqrestore(&conf->device_lock, flags);
1401 * if recovery is running, make sure it aborts.
1403 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1405 set_bit(Faulty, &rdev->flags);
1406 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1408 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1409 "md/raid1:%s: Operation continuing on %d devices.\n",
1410 mdname(mddev), bdevname(rdev->bdev, b),
1411 mdname(mddev), conf->raid_disks - mddev->degraded);
1414 static void print_conf(struct r1conf *conf)
1418 printk(KERN_DEBUG "RAID1 conf printout:\n");
1420 printk(KERN_DEBUG "(!conf)\n");
1423 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1427 for (i = 0; i < conf->raid_disks; i++) {
1428 char b[BDEVNAME_SIZE];
1429 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1431 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1432 i, !test_bit(In_sync, &rdev->flags),
1433 !test_bit(Faulty, &rdev->flags),
1434 bdevname(rdev->bdev,b));
1439 static void close_sync(struct r1conf *conf)
1442 allow_barrier(conf);
1444 mempool_destroy(conf->r1buf_pool);
1445 conf->r1buf_pool = NULL;
1448 static int raid1_spare_active(struct mddev *mddev)
1451 struct r1conf *conf = mddev->private;
1453 unsigned long flags;
1456 * Find all failed disks within the RAID1 configuration
1457 * and mark them readable.
1458 * Called under mddev lock, so rcu protection not needed.
1460 for (i = 0; i < conf->raid_disks; i++) {
1461 struct md_rdev *rdev = conf->mirrors[i].rdev;
1462 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1464 && repl->recovery_offset == MaxSector
1465 && !test_bit(Faulty, &repl->flags)
1466 && !test_and_set_bit(In_sync, &repl->flags)) {
1467 /* replacement has just become active */
1469 !test_and_clear_bit(In_sync, &rdev->flags))
1472 /* Replaced device not technically
1473 * faulty, but we need to be sure
1474 * it gets removed and never re-added
1476 set_bit(Faulty, &rdev->flags);
1477 sysfs_notify_dirent_safe(
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;
1522 disk_stack_limits(mddev->gendisk, rdev->bdev,
1523 rdev->data_offset << 9);
1525 p->head_position = 0;
1526 rdev->raid_disk = mirror;
1528 /* As all devices are equivalent, we don't need a full recovery
1529 * if this was recently any drive of the array
1531 if (rdev->saved_raid_disk < 0)
1533 rcu_assign_pointer(p->rdev, rdev);
1536 if (test_bit(WantReplacement, &p->rdev->flags) &&
1537 p[conf->raid_disks].rdev == NULL) {
1538 /* Add this device as a replacement */
1539 clear_bit(In_sync, &rdev->flags);
1540 set_bit(Replacement, &rdev->flags);
1541 rdev->raid_disk = mirror;
1544 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1548 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1549 /* Some requests might not have seen this new
1550 * merge_bvec_fn. We must wait for them to complete
1551 * before merging the device fully.
1552 * First we make sure any code which has tested
1553 * our function has submitted the request, then
1554 * we wait for all outstanding requests to complete.
1556 synchronize_sched();
1557 freeze_array(conf, 0);
1558 unfreeze_array(conf);
1559 clear_bit(Unmerged, &rdev->flags);
1561 md_integrity_add_rdev(rdev, mddev);
1562 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
1563 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1568 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1570 struct r1conf *conf = mddev->private;
1572 int number = rdev->raid_disk;
1573 struct raid1_info *p = conf->mirrors + number;
1575 if (rdev != p->rdev)
1576 p = conf->mirrors + conf->raid_disks + number;
1579 if (rdev == p->rdev) {
1580 if (test_bit(In_sync, &rdev->flags) ||
1581 atomic_read(&rdev->nr_pending)) {
1585 /* Only remove non-faulty devices if recovery
1588 if (!test_bit(Faulty, &rdev->flags) &&
1589 mddev->recovery_disabled != conf->recovery_disabled &&
1590 mddev->degraded < conf->raid_disks) {
1596 if (atomic_read(&rdev->nr_pending)) {
1597 /* lost the race, try later */
1601 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1602 /* We just removed a device that is being replaced.
1603 * Move down the replacement. We drain all IO before
1604 * doing this to avoid confusion.
1606 struct md_rdev *repl =
1607 conf->mirrors[conf->raid_disks + number].rdev;
1608 freeze_array(conf, 0);
1609 clear_bit(Replacement, &repl->flags);
1611 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1612 unfreeze_array(conf);
1613 clear_bit(WantReplacement, &rdev->flags);
1615 clear_bit(WantReplacement, &rdev->flags);
1616 err = md_integrity_register(mddev);
1625 static void end_sync_read(struct bio *bio, int error)
1627 struct r1bio *r1_bio = bio->bi_private;
1629 update_head_pos(r1_bio->read_disk, r1_bio);
1632 * we have read a block, now it needs to be re-written,
1633 * or re-read if the read failed.
1634 * We don't do much here, just schedule handling by raid1d
1636 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1637 set_bit(R1BIO_Uptodate, &r1_bio->state);
1639 if (atomic_dec_and_test(&r1_bio->remaining))
1640 reschedule_retry(r1_bio);
1643 static void end_sync_write(struct bio *bio, int error)
1645 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1646 struct r1bio *r1_bio = bio->bi_private;
1647 struct mddev *mddev = r1_bio->mddev;
1648 struct r1conf *conf = mddev->private;
1653 mirror = find_bio_disk(r1_bio, bio);
1656 sector_t sync_blocks = 0;
1657 sector_t s = r1_bio->sector;
1658 long sectors_to_go = r1_bio->sectors;
1659 /* make sure these bits doesn't get cleared. */
1661 bitmap_end_sync(mddev->bitmap, s,
1664 sectors_to_go -= sync_blocks;
1665 } while (sectors_to_go > 0);
1666 set_bit(WriteErrorSeen,
1667 &conf->mirrors[mirror].rdev->flags);
1668 if (!test_and_set_bit(WantReplacement,
1669 &conf->mirrors[mirror].rdev->flags))
1670 set_bit(MD_RECOVERY_NEEDED, &
1672 set_bit(R1BIO_WriteError, &r1_bio->state);
1673 } else if (is_badblock(conf->mirrors[mirror].rdev,
1676 &first_bad, &bad_sectors) &&
1677 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1680 &first_bad, &bad_sectors)
1682 set_bit(R1BIO_MadeGood, &r1_bio->state);
1684 if (atomic_dec_and_test(&r1_bio->remaining)) {
1685 int s = r1_bio->sectors;
1686 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1687 test_bit(R1BIO_WriteError, &r1_bio->state))
1688 reschedule_retry(r1_bio);
1691 md_done_sync(mddev, s, uptodate);
1696 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1697 int sectors, struct page *page, int rw)
1699 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1703 set_bit(WriteErrorSeen, &rdev->flags);
1704 if (!test_and_set_bit(WantReplacement,
1706 set_bit(MD_RECOVERY_NEEDED, &
1707 rdev->mddev->recovery);
1709 /* need to record an error - either for the block or the device */
1710 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1711 md_error(rdev->mddev, rdev);
1715 static int fix_sync_read_error(struct r1bio *r1_bio)
1717 /* Try some synchronous reads of other devices to get
1718 * good data, much like with normal read errors. Only
1719 * read into the pages we already have so we don't
1720 * need to re-issue the read request.
1721 * We don't need to freeze the array, because being in an
1722 * active sync request, there is no normal IO, and
1723 * no overlapping syncs.
1724 * We don't need to check is_badblock() again as we
1725 * made sure that anything with a bad block in range
1726 * will have bi_end_io clear.
1728 struct mddev *mddev = r1_bio->mddev;
1729 struct r1conf *conf = mddev->private;
1730 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1731 sector_t sect = r1_bio->sector;
1732 int sectors = r1_bio->sectors;
1737 int d = r1_bio->read_disk;
1739 struct md_rdev *rdev;
1742 if (s > (PAGE_SIZE>>9))
1745 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1746 /* No rcu protection needed here devices
1747 * can only be removed when no resync is
1748 * active, and resync is currently active
1750 rdev = conf->mirrors[d].rdev;
1751 if (sync_page_io(rdev, sect, s<<9,
1752 bio->bi_io_vec[idx].bv_page,
1759 if (d == conf->raid_disks * 2)
1761 } while (!success && d != r1_bio->read_disk);
1764 char b[BDEVNAME_SIZE];
1766 /* Cannot read from anywhere, this block is lost.
1767 * Record a bad block on each device. If that doesn't
1768 * work just disable and interrupt the recovery.
1769 * Don't fail devices as that won't really help.
1771 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1772 " for block %llu\n",
1774 bdevname(bio->bi_bdev, b),
1775 (unsigned long long)r1_bio->sector);
1776 for (d = 0; d < conf->raid_disks * 2; d++) {
1777 rdev = conf->mirrors[d].rdev;
1778 if (!rdev || test_bit(Faulty, &rdev->flags))
1780 if (!rdev_set_badblocks(rdev, sect, s, 0))
1784 conf->recovery_disabled =
1785 mddev->recovery_disabled;
1786 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1787 md_done_sync(mddev, r1_bio->sectors, 0);
1799 /* write it back and re-read */
1800 while (d != r1_bio->read_disk) {
1802 d = conf->raid_disks * 2;
1804 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1806 rdev = conf->mirrors[d].rdev;
1807 if (r1_sync_page_io(rdev, sect, s,
1808 bio->bi_io_vec[idx].bv_page,
1810 r1_bio->bios[d]->bi_end_io = NULL;
1811 rdev_dec_pending(rdev, mddev);
1815 while (d != r1_bio->read_disk) {
1817 d = conf->raid_disks * 2;
1819 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1821 rdev = conf->mirrors[d].rdev;
1822 if (r1_sync_page_io(rdev, sect, s,
1823 bio->bi_io_vec[idx].bv_page,
1825 atomic_add(s, &rdev->corrected_errors);
1831 set_bit(R1BIO_Uptodate, &r1_bio->state);
1832 set_bit(BIO_UPTODATE, &bio->bi_flags);
1836 static int process_checks(struct r1bio *r1_bio)
1838 /* We have read all readable devices. If we haven't
1839 * got the block, then there is no hope left.
1840 * If we have, then we want to do a comparison
1841 * and skip the write if everything is the same.
1842 * If any blocks failed to read, then we need to
1843 * attempt an over-write
1845 struct mddev *mddev = r1_bio->mddev;
1846 struct r1conf *conf = mddev->private;
1851 /* Fix variable parts of all bios */
1852 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1853 for (i = 0; i < conf->raid_disks * 2; i++) {
1856 struct bio *b = r1_bio->bios[i];
1857 if (b->bi_end_io != end_sync_read)
1859 /* fixup the bio for reuse */
1862 b->bi_size = r1_bio->sectors << 9;
1863 b->bi_sector = r1_bio->sector +
1864 conf->mirrors[i].rdev->data_offset;
1865 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1866 b->bi_end_io = end_sync_read;
1867 b->bi_private = r1_bio;
1870 for (j = 0; j < vcnt ; j++) {
1872 bi = &b->bi_io_vec[j];
1874 if (size > PAGE_SIZE)
1875 bi->bv_len = PAGE_SIZE;
1881 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1882 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1883 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1884 r1_bio->bios[primary]->bi_end_io = NULL;
1885 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1888 r1_bio->read_disk = primary;
1889 for (i = 0; i < conf->raid_disks * 2; i++) {
1891 struct bio *pbio = r1_bio->bios[primary];
1892 struct bio *sbio = r1_bio->bios[i];
1894 if (sbio->bi_end_io != end_sync_read)
1897 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1898 for (j = vcnt; j-- ; ) {
1900 p = pbio->bi_io_vec[j].bv_page;
1901 s = sbio->bi_io_vec[j].bv_page;
1902 if (memcmp(page_address(p),
1904 sbio->bi_io_vec[j].bv_len))
1910 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1911 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1912 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1913 /* No need to write to this device. */
1914 sbio->bi_end_io = NULL;
1915 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1919 bio_copy_data(sbio, pbio);
1924 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1926 struct r1conf *conf = mddev->private;
1928 int disks = conf->raid_disks * 2;
1929 struct bio *bio, *wbio;
1931 bio = r1_bio->bios[r1_bio->read_disk];
1933 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1934 /* ouch - failed to read all of that. */
1935 if (!fix_sync_read_error(r1_bio))
1938 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1939 if (process_checks(r1_bio) < 0)
1944 atomic_set(&r1_bio->remaining, 1);
1945 for (i = 0; i < disks ; i++) {
1946 wbio = r1_bio->bios[i];
1947 if (wbio->bi_end_io == NULL ||
1948 (wbio->bi_end_io == end_sync_read &&
1949 (i == r1_bio->read_disk ||
1950 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1953 wbio->bi_rw = WRITE;
1954 wbio->bi_end_io = end_sync_write;
1955 atomic_inc(&r1_bio->remaining);
1956 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
1958 generic_make_request(wbio);
1961 if (atomic_dec_and_test(&r1_bio->remaining)) {
1962 /* if we're here, all write(s) have completed, so clean up */
1963 int s = r1_bio->sectors;
1964 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1965 test_bit(R1BIO_WriteError, &r1_bio->state))
1966 reschedule_retry(r1_bio);
1969 md_done_sync(mddev, s, 1);
1975 * This is a kernel thread which:
1977 * 1. Retries failed read operations on working mirrors.
1978 * 2. Updates the raid superblock when problems encounter.
1979 * 3. Performs writes following reads for array synchronising.
1982 static void fix_read_error(struct r1conf *conf, int read_disk,
1983 sector_t sect, int sectors)
1985 struct mddev *mddev = conf->mddev;
1991 struct md_rdev *rdev;
1993 if (s > (PAGE_SIZE>>9))
1997 /* Note: no rcu protection needed here
1998 * as this is synchronous in the raid1d thread
1999 * which is the thread that might remove
2000 * a device. If raid1d ever becomes multi-threaded....
2005 rdev = conf->mirrors[d].rdev;
2007 (test_bit(In_sync, &rdev->flags) ||
2008 (!test_bit(Faulty, &rdev->flags) &&
2009 rdev->recovery_offset >= sect + s)) &&
2010 is_badblock(rdev, sect, s,
2011 &first_bad, &bad_sectors) == 0 &&
2012 sync_page_io(rdev, sect, s<<9,
2013 conf->tmppage, READ, false))
2017 if (d == conf->raid_disks * 2)
2020 } while (!success && d != read_disk);
2023 /* Cannot read from anywhere - mark it bad */
2024 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2025 if (!rdev_set_badblocks(rdev, sect, s, 0))
2026 md_error(mddev, rdev);
2029 /* write it back and re-read */
2031 while (d != read_disk) {
2033 d = conf->raid_disks * 2;
2035 rdev = conf->mirrors[d].rdev;
2037 test_bit(In_sync, &rdev->flags))
2038 r1_sync_page_io(rdev, sect, s,
2039 conf->tmppage, WRITE);
2042 while (d != read_disk) {
2043 char b[BDEVNAME_SIZE];
2045 d = conf->raid_disks * 2;
2047 rdev = conf->mirrors[d].rdev;
2049 test_bit(In_sync, &rdev->flags)) {
2050 if (r1_sync_page_io(rdev, sect, s,
2051 conf->tmppage, READ)) {
2052 atomic_add(s, &rdev->corrected_errors);
2054 "md/raid1:%s: read error corrected "
2055 "(%d sectors at %llu on %s)\n",
2057 (unsigned long long)(sect +
2059 bdevname(rdev->bdev, b));
2068 static int narrow_write_error(struct r1bio *r1_bio, int i)
2070 struct mddev *mddev = r1_bio->mddev;
2071 struct r1conf *conf = mddev->private;
2072 struct md_rdev *rdev = conf->mirrors[i].rdev;
2074 /* bio has the data to be written to device 'i' where
2075 * we just recently had a write error.
2076 * We repeatedly clone the bio and trim down to one block,
2077 * then try the write. Where the write fails we record
2079 * It is conceivable that the bio doesn't exactly align with
2080 * blocks. We must handle this somehow.
2082 * We currently own a reference on the rdev.
2088 int sect_to_write = r1_bio->sectors;
2091 if (rdev->badblocks.shift < 0)
2094 block_sectors = 1 << rdev->badblocks.shift;
2095 sector = r1_bio->sector;
2096 sectors = ((sector + block_sectors)
2097 & ~(sector_t)(block_sectors - 1))
2100 while (sect_to_write) {
2102 if (sectors > sect_to_write)
2103 sectors = sect_to_write;
2104 /* Write at 'sector' for 'sectors'*/
2106 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2107 unsigned vcnt = r1_bio->behind_page_count;
2108 struct bio_vec *vec = r1_bio->behind_bvecs;
2110 while (!vec->bv_page) {
2115 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2116 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2118 wbio->bi_vcnt = vcnt;
2120 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2123 wbio->bi_rw = WRITE;
2124 wbio->bi_sector = r1_bio->sector;
2125 wbio->bi_size = r1_bio->sectors << 9;
2127 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2128 wbio->bi_sector += rdev->data_offset;
2129 wbio->bi_bdev = rdev->bdev;
2130 if (submit_bio_wait(WRITE, wbio) == 0)
2132 ok = rdev_set_badblocks(rdev, sector,
2137 sect_to_write -= sectors;
2139 sectors = block_sectors;
2144 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2147 int s = r1_bio->sectors;
2148 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2149 struct md_rdev *rdev = conf->mirrors[m].rdev;
2150 struct bio *bio = r1_bio->bios[m];
2151 if (bio->bi_end_io == NULL)
2153 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2154 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2155 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2157 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2158 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2159 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2160 md_error(conf->mddev, rdev);
2164 md_done_sync(conf->mddev, s, 1);
2167 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2170 for (m = 0; m < conf->raid_disks * 2 ; m++)
2171 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2172 struct md_rdev *rdev = conf->mirrors[m].rdev;
2173 rdev_clear_badblocks(rdev,
2175 r1_bio->sectors, 0);
2176 rdev_dec_pending(rdev, conf->mddev);
2177 } else if (r1_bio->bios[m] != NULL) {
2178 /* This drive got a write error. We need to
2179 * narrow down and record precise write
2182 if (!narrow_write_error(r1_bio, m)) {
2183 md_error(conf->mddev,
2184 conf->mirrors[m].rdev);
2185 /* an I/O failed, we can't clear the bitmap */
2186 set_bit(R1BIO_Degraded, &r1_bio->state);
2188 rdev_dec_pending(conf->mirrors[m].rdev,
2191 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2192 close_write(r1_bio);
2193 raid_end_bio_io(r1_bio);
2196 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2200 struct mddev *mddev = conf->mddev;
2202 char b[BDEVNAME_SIZE];
2203 struct md_rdev *rdev;
2205 clear_bit(R1BIO_ReadError, &r1_bio->state);
2206 /* we got a read error. Maybe the drive is bad. Maybe just
2207 * the block and we can fix it.
2208 * We freeze all other IO, and try reading the block from
2209 * other devices. When we find one, we re-write
2210 * and check it that fixes the read error.
2211 * This is all done synchronously while the array is
2214 if (mddev->ro == 0) {
2215 freeze_array(conf, 1);
2216 fix_read_error(conf, r1_bio->read_disk,
2217 r1_bio->sector, r1_bio->sectors);
2218 unfreeze_array(conf);
2220 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2221 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2223 bio = r1_bio->bios[r1_bio->read_disk];
2224 bdevname(bio->bi_bdev, b);
2226 disk = read_balance(conf, r1_bio, &max_sectors);
2228 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2229 " read error for block %llu\n",
2230 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2231 raid_end_bio_io(r1_bio);
2233 const unsigned long do_sync
2234 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2236 r1_bio->bios[r1_bio->read_disk] =
2237 mddev->ro ? IO_BLOCKED : NULL;
2240 r1_bio->read_disk = disk;
2241 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2242 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2243 r1_bio->bios[r1_bio->read_disk] = bio;
2244 rdev = conf->mirrors[disk].rdev;
2245 printk_ratelimited(KERN_ERR
2246 "md/raid1:%s: redirecting sector %llu"
2247 " to other mirror: %s\n",
2249 (unsigned long long)r1_bio->sector,
2250 bdevname(rdev->bdev, b));
2251 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2252 bio->bi_bdev = rdev->bdev;
2253 bio->bi_end_io = raid1_end_read_request;
2254 bio->bi_rw = READ | do_sync;
2255 bio->bi_private = r1_bio;
2256 if (max_sectors < r1_bio->sectors) {
2257 /* Drat - have to split this up more */
2258 struct bio *mbio = r1_bio->master_bio;
2259 int sectors_handled = (r1_bio->sector + max_sectors
2261 r1_bio->sectors = max_sectors;
2262 spin_lock_irq(&conf->device_lock);
2263 if (mbio->bi_phys_segments == 0)
2264 mbio->bi_phys_segments = 2;
2266 mbio->bi_phys_segments++;
2267 spin_unlock_irq(&conf->device_lock);
2268 generic_make_request(bio);
2271 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2273 r1_bio->master_bio = mbio;
2274 r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2276 set_bit(R1BIO_ReadError, &r1_bio->state);
2277 r1_bio->mddev = mddev;
2278 r1_bio->sector = mbio->bi_sector + sectors_handled;
2282 generic_make_request(bio);
2286 static void raid1d(struct md_thread *thread)
2288 struct mddev *mddev = thread->mddev;
2289 struct r1bio *r1_bio;
2290 unsigned long flags;
2291 struct r1conf *conf = mddev->private;
2292 struct list_head *head = &conf->retry_list;
2293 struct blk_plug plug;
2295 md_check_recovery(mddev);
2297 blk_start_plug(&plug);
2300 flush_pending_writes(conf);
2302 spin_lock_irqsave(&conf->device_lock, flags);
2303 if (list_empty(head)) {
2304 spin_unlock_irqrestore(&conf->device_lock, flags);
2307 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2308 list_del(head->prev);
2310 spin_unlock_irqrestore(&conf->device_lock, flags);
2312 mddev = r1_bio->mddev;
2313 conf = mddev->private;
2314 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2315 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2316 test_bit(R1BIO_WriteError, &r1_bio->state))
2317 handle_sync_write_finished(conf, r1_bio);
2319 sync_request_write(mddev, r1_bio);
2320 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2321 test_bit(R1BIO_WriteError, &r1_bio->state))
2322 handle_write_finished(conf, r1_bio);
2323 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2324 handle_read_error(conf, r1_bio);
2326 /* just a partial read to be scheduled from separate
2329 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2332 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2333 md_check_recovery(mddev);
2335 blk_finish_plug(&plug);
2339 static int init_resync(struct r1conf *conf)
2343 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2344 BUG_ON(conf->r1buf_pool);
2345 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2347 if (!conf->r1buf_pool)
2349 conf->next_resync = 0;
2354 * perform a "sync" on one "block"
2356 * We need to make sure that no normal I/O request - particularly write
2357 * requests - conflict with active sync requests.
2359 * This is achieved by tracking pending requests and a 'barrier' concept
2360 * that can be installed to exclude normal IO requests.
2363 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2365 struct r1conf *conf = mddev->private;
2366 struct r1bio *r1_bio;
2368 sector_t max_sector, nr_sectors;
2372 int write_targets = 0, read_targets = 0;
2373 sector_t sync_blocks;
2374 int still_degraded = 0;
2375 int good_sectors = RESYNC_SECTORS;
2376 int min_bad = 0; /* number of sectors that are bad in all devices */
2378 if (!conf->r1buf_pool)
2379 if (init_resync(conf))
2382 max_sector = mddev->dev_sectors;
2383 if (sector_nr >= max_sector) {
2384 /* If we aborted, we need to abort the
2385 * sync on the 'current' bitmap chunk (there will
2386 * only be one in raid1 resync.
2387 * We can find the current addess in mddev->curr_resync
2389 if (mddev->curr_resync < max_sector) /* aborted */
2390 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2392 else /* completed sync */
2395 bitmap_close_sync(mddev->bitmap);
2400 if (mddev->bitmap == NULL &&
2401 mddev->recovery_cp == MaxSector &&
2402 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2403 conf->fullsync == 0) {
2405 return max_sector - sector_nr;
2407 /* before building a request, check if we can skip these blocks..
2408 * This call the bitmap_start_sync doesn't actually record anything
2410 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2411 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2412 /* We can skip this block, and probably several more */
2417 * If there is non-resync activity waiting for a turn,
2418 * and resync is going fast enough,
2419 * then let it though before starting on this new sync request.
2421 if (!go_faster && conf->nr_waiting)
2422 msleep_interruptible(1000);
2424 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2425 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2426 raise_barrier(conf);
2428 conf->next_resync = sector_nr;
2432 * If we get a correctably read error during resync or recovery,
2433 * we might want to read from a different device. So we
2434 * flag all drives that could conceivably be read from for READ,
2435 * and any others (which will be non-In_sync devices) for WRITE.
2436 * If a read fails, we try reading from something else for which READ
2440 r1_bio->mddev = mddev;
2441 r1_bio->sector = sector_nr;
2443 set_bit(R1BIO_IsSync, &r1_bio->state);
2445 for (i = 0; i < conf->raid_disks * 2; i++) {
2446 struct md_rdev *rdev;
2447 bio = r1_bio->bios[i];
2450 rdev = rcu_dereference(conf->mirrors[i].rdev);
2452 test_bit(Faulty, &rdev->flags)) {
2453 if (i < conf->raid_disks)
2455 } else if (!test_bit(In_sync, &rdev->flags)) {
2457 bio->bi_end_io = end_sync_write;
2460 /* may need to read from here */
2461 sector_t first_bad = MaxSector;
2464 if (is_badblock(rdev, sector_nr, good_sectors,
2465 &first_bad, &bad_sectors)) {
2466 if (first_bad > sector_nr)
2467 good_sectors = first_bad - sector_nr;
2469 bad_sectors -= (sector_nr - first_bad);
2471 min_bad > bad_sectors)
2472 min_bad = bad_sectors;
2475 if (sector_nr < first_bad) {
2476 if (test_bit(WriteMostly, &rdev->flags)) {
2484 bio->bi_end_io = end_sync_read;
2486 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2487 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2488 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2490 * The device is suitable for reading (InSync),
2491 * but has bad block(s) here. Let's try to correct them,
2492 * if we are doing resync or repair. Otherwise, leave
2493 * this device alone for this sync request.
2496 bio->bi_end_io = end_sync_write;
2500 if (bio->bi_end_io) {
2501 atomic_inc(&rdev->nr_pending);
2502 bio->bi_sector = sector_nr + rdev->data_offset;
2503 bio->bi_bdev = rdev->bdev;
2504 bio->bi_private = r1_bio;
2510 r1_bio->read_disk = disk;
2512 if (read_targets == 0 && min_bad > 0) {
2513 /* These sectors are bad on all InSync devices, so we
2514 * need to mark them bad on all write targets
2517 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2518 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2519 struct md_rdev *rdev = conf->mirrors[i].rdev;
2520 ok = rdev_set_badblocks(rdev, sector_nr,
2524 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2529 /* Cannot record the badblocks, so need to
2531 * If there are multiple read targets, could just
2532 * fail the really bad ones ???
2534 conf->recovery_disabled = mddev->recovery_disabled;
2535 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2541 if (min_bad > 0 && min_bad < good_sectors) {
2542 /* only resync enough to reach the next bad->good
2544 good_sectors = min_bad;
2547 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2548 /* extra read targets are also write targets */
2549 write_targets += read_targets-1;
2551 if (write_targets == 0 || read_targets == 0) {
2552 /* There is nowhere to write, so all non-sync
2553 * drives must be failed - so we are finished
2557 max_sector = sector_nr + min_bad;
2558 rv = max_sector - sector_nr;
2564 if (max_sector > mddev->resync_max)
2565 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2566 if (max_sector > sector_nr + good_sectors)
2567 max_sector = sector_nr + good_sectors;
2572 int len = PAGE_SIZE;
2573 if (sector_nr + (len>>9) > max_sector)
2574 len = (max_sector - sector_nr) << 9;
2577 if (sync_blocks == 0) {
2578 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2579 &sync_blocks, still_degraded) &&
2581 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2583 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2584 if ((len >> 9) > sync_blocks)
2585 len = sync_blocks<<9;
2588 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2589 bio = r1_bio->bios[i];
2590 if (bio->bi_end_io) {
2591 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2592 if (bio_add_page(bio, page, len, 0) == 0) {
2594 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2597 bio = r1_bio->bios[i];
2598 if (bio->bi_end_io==NULL)
2600 /* remove last page from this bio */
2602 bio->bi_size -= len;
2603 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2609 nr_sectors += len>>9;
2610 sector_nr += len>>9;
2611 sync_blocks -= (len>>9);
2612 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2614 r1_bio->sectors = nr_sectors;
2616 /* For a user-requested sync, we read all readable devices and do a
2619 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2620 atomic_set(&r1_bio->remaining, read_targets);
2621 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2622 bio = r1_bio->bios[i];
2623 if (bio->bi_end_io == end_sync_read) {
2625 md_sync_acct(bio->bi_bdev, nr_sectors);
2626 generic_make_request(bio);
2630 atomic_set(&r1_bio->remaining, 1);
2631 bio = r1_bio->bios[r1_bio->read_disk];
2632 md_sync_acct(bio->bi_bdev, nr_sectors);
2633 generic_make_request(bio);
2639 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2644 return mddev->dev_sectors;
2647 static struct r1conf *setup_conf(struct mddev *mddev)
2649 struct r1conf *conf;
2651 struct raid1_info *disk;
2652 struct md_rdev *rdev;
2655 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2659 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2660 * mddev->raid_disks * 2,
2665 conf->tmppage = alloc_page(GFP_KERNEL);
2669 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2670 if (!conf->poolinfo)
2672 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2673 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2676 if (!conf->r1bio_pool)
2679 conf->poolinfo->mddev = mddev;
2682 spin_lock_init(&conf->device_lock);
2683 rdev_for_each(rdev, mddev) {
2684 struct request_queue *q;
2685 int disk_idx = rdev->raid_disk;
2686 if (disk_idx >= mddev->raid_disks
2689 if (test_bit(Replacement, &rdev->flags))
2690 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2692 disk = conf->mirrors + disk_idx;
2697 q = bdev_get_queue(rdev->bdev);
2698 if (q->merge_bvec_fn)
2699 mddev->merge_check_needed = 1;
2701 disk->head_position = 0;
2702 disk->seq_start = MaxSector;
2704 conf->raid_disks = mddev->raid_disks;
2705 conf->mddev = mddev;
2706 INIT_LIST_HEAD(&conf->retry_list);
2708 spin_lock_init(&conf->resync_lock);
2709 init_waitqueue_head(&conf->wait_barrier);
2711 bio_list_init(&conf->pending_bio_list);
2712 conf->pending_count = 0;
2713 conf->recovery_disabled = mddev->recovery_disabled - 1;
2716 for (i = 0; i < conf->raid_disks * 2; i++) {
2718 disk = conf->mirrors + i;
2720 if (i < conf->raid_disks &&
2721 disk[conf->raid_disks].rdev) {
2722 /* This slot has a replacement. */
2724 /* No original, just make the replacement
2725 * a recovering spare
2728 disk[conf->raid_disks].rdev;
2729 disk[conf->raid_disks].rdev = NULL;
2730 } else if (!test_bit(In_sync, &disk->rdev->flags))
2731 /* Original is not in_sync - bad */
2736 !test_bit(In_sync, &disk->rdev->flags)) {
2737 disk->head_position = 0;
2739 (disk->rdev->saved_raid_disk < 0))
2745 conf->thread = md_register_thread(raid1d, mddev, "raid1");
2746 if (!conf->thread) {
2748 "md/raid1:%s: couldn't allocate thread\n",
2757 if (conf->r1bio_pool)
2758 mempool_destroy(conf->r1bio_pool);
2759 kfree(conf->mirrors);
2760 safe_put_page(conf->tmppage);
2761 kfree(conf->poolinfo);
2764 return ERR_PTR(err);
2767 static int stop(struct mddev *mddev);
2768 static int run(struct mddev *mddev)
2770 struct r1conf *conf;
2772 struct md_rdev *rdev;
2774 bool discard_supported = false;
2776 if (mddev->level != 1) {
2777 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2778 mdname(mddev), mddev->level);
2781 if (mddev->reshape_position != MaxSector) {
2782 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2787 * copy the already verified devices into our private RAID1
2788 * bookkeeping area. [whatever we allocate in run(),
2789 * should be freed in stop()]
2791 if (mddev->private == NULL)
2792 conf = setup_conf(mddev);
2794 conf = mddev->private;
2797 return PTR_ERR(conf);
2800 blk_queue_max_write_same_sectors(mddev->queue, 0);
2802 rdev_for_each(rdev, mddev) {
2803 if (!mddev->gendisk)
2805 disk_stack_limits(mddev->gendisk, rdev->bdev,
2806 rdev->data_offset << 9);
2807 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2808 discard_supported = true;
2811 mddev->degraded = 0;
2812 for (i=0; i < conf->raid_disks; i++)
2813 if (conf->mirrors[i].rdev == NULL ||
2814 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2815 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2818 if (conf->raid_disks - mddev->degraded == 1)
2819 mddev->recovery_cp = MaxSector;
2821 if (mddev->recovery_cp != MaxSector)
2822 printk(KERN_NOTICE "md/raid1:%s: not clean"
2823 " -- starting background reconstruction\n",
2826 "md/raid1:%s: active with %d out of %d mirrors\n",
2827 mdname(mddev), mddev->raid_disks - mddev->degraded,
2831 * Ok, everything is just fine now
2833 mddev->thread = conf->thread;
2834 conf->thread = NULL;
2835 mddev->private = conf;
2837 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2840 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2841 mddev->queue->backing_dev_info.congested_data = mddev;
2842 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2844 if (discard_supported)
2845 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2848 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2852 ret = md_integrity_register(mddev);
2858 static int stop(struct mddev *mddev)
2860 struct r1conf *conf = mddev->private;
2861 struct bitmap *bitmap = mddev->bitmap;
2863 /* wait for behind writes to complete */
2864 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2865 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2867 /* need to kick something here to make sure I/O goes? */
2868 wait_event(bitmap->behind_wait,
2869 atomic_read(&bitmap->behind_writes) == 0);
2872 raise_barrier(conf);
2873 lower_barrier(conf);
2875 md_unregister_thread(&mddev->thread);
2876 if (conf->r1bio_pool)
2877 mempool_destroy(conf->r1bio_pool);
2878 kfree(conf->mirrors);
2879 safe_put_page(conf->tmppage);
2880 kfree(conf->poolinfo);
2882 mddev->private = NULL;
2886 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2888 /* no resync is happening, and there is enough space
2889 * on all devices, so we can resize.
2890 * We need to make sure resync covers any new space.
2891 * If the array is shrinking we should possibly wait until
2892 * any io in the removed space completes, but it hardly seems
2895 sector_t newsize = raid1_size(mddev, sectors, 0);
2896 if (mddev->external_size &&
2897 mddev->array_sectors > newsize)
2899 if (mddev->bitmap) {
2900 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2904 md_set_array_sectors(mddev, newsize);
2905 set_capacity(mddev->gendisk, mddev->array_sectors);
2906 revalidate_disk(mddev->gendisk);
2907 if (sectors > mddev->dev_sectors &&
2908 mddev->recovery_cp > mddev->dev_sectors) {
2909 mddev->recovery_cp = mddev->dev_sectors;
2910 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2912 mddev->dev_sectors = sectors;
2913 mddev->resync_max_sectors = sectors;
2917 static int raid1_reshape(struct mddev *mddev)
2920 * 1/ resize the r1bio_pool
2921 * 2/ resize conf->mirrors
2923 * We allocate a new r1bio_pool if we can.
2924 * Then raise a device barrier and wait until all IO stops.
2925 * Then resize conf->mirrors and swap in the new r1bio pool.
2927 * At the same time, we "pack" the devices so that all the missing
2928 * devices have the higher raid_disk numbers.
2930 mempool_t *newpool, *oldpool;
2931 struct pool_info *newpoolinfo;
2932 struct raid1_info *newmirrors;
2933 struct r1conf *conf = mddev->private;
2934 int cnt, raid_disks;
2935 unsigned long flags;
2938 /* Cannot change chunk_size, layout, or level */
2939 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2940 mddev->layout != mddev->new_layout ||
2941 mddev->level != mddev->new_level) {
2942 mddev->new_chunk_sectors = mddev->chunk_sectors;
2943 mddev->new_layout = mddev->layout;
2944 mddev->new_level = mddev->level;
2948 err = md_allow_write(mddev);
2952 raid_disks = mddev->raid_disks + mddev->delta_disks;
2954 if (raid_disks < conf->raid_disks) {
2956 for (d= 0; d < conf->raid_disks; d++)
2957 if (conf->mirrors[d].rdev)
2959 if (cnt > raid_disks)
2963 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2966 newpoolinfo->mddev = mddev;
2967 newpoolinfo->raid_disks = raid_disks * 2;
2969 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2970 r1bio_pool_free, newpoolinfo);
2975 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2979 mempool_destroy(newpool);
2983 freeze_array(conf, 0);
2985 /* ok, everything is stopped */
2986 oldpool = conf->r1bio_pool;
2987 conf->r1bio_pool = newpool;
2989 for (d = d2 = 0; d < conf->raid_disks; d++) {
2990 struct md_rdev *rdev = conf->mirrors[d].rdev;
2991 if (rdev && rdev->raid_disk != d2) {
2992 sysfs_unlink_rdev(mddev, rdev);
2993 rdev->raid_disk = d2;
2994 sysfs_unlink_rdev(mddev, rdev);
2995 if (sysfs_link_rdev(mddev, rdev))
2997 "md/raid1:%s: cannot register rd%d\n",
2998 mdname(mddev), rdev->raid_disk);
3001 newmirrors[d2++].rdev = rdev;
3003 kfree(conf->mirrors);
3004 conf->mirrors = newmirrors;
3005 kfree(conf->poolinfo);
3006 conf->poolinfo = newpoolinfo;
3008 spin_lock_irqsave(&conf->device_lock, flags);
3009 mddev->degraded += (raid_disks - conf->raid_disks);
3010 spin_unlock_irqrestore(&conf->device_lock, flags);
3011 conf->raid_disks = mddev->raid_disks = raid_disks;
3012 mddev->delta_disks = 0;
3014 unfreeze_array(conf);
3016 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3017 md_wakeup_thread(mddev->thread);
3019 mempool_destroy(oldpool);
3023 static void raid1_quiesce(struct mddev *mddev, int state)
3025 struct r1conf *conf = mddev->private;
3028 case 2: /* wake for suspend */
3029 wake_up(&conf->wait_barrier);
3032 raise_barrier(conf);
3035 lower_barrier(conf);
3040 static void *raid1_takeover(struct mddev *mddev)
3042 /* raid1 can take over:
3043 * raid5 with 2 devices, any layout or chunk size
3045 if (mddev->level == 5 && mddev->raid_disks == 2) {
3046 struct r1conf *conf;
3047 mddev->new_level = 1;
3048 mddev->new_layout = 0;
3049 mddev->new_chunk_sectors = 0;
3050 conf = setup_conf(mddev);
3055 return ERR_PTR(-EINVAL);
3058 static struct md_personality raid1_personality =
3062 .owner = THIS_MODULE,
3063 .make_request = make_request,
3067 .error_handler = error,
3068 .hot_add_disk = raid1_add_disk,
3069 .hot_remove_disk= raid1_remove_disk,
3070 .spare_active = raid1_spare_active,
3071 .sync_request = sync_request,
3072 .resize = raid1_resize,
3074 .check_reshape = raid1_reshape,
3075 .quiesce = raid1_quiesce,
3076 .takeover = raid1_takeover,
3079 static int __init raid_init(void)
3081 return register_md_personality(&raid1_personality);
3084 static void raid_exit(void)
3086 unregister_md_personality(&raid1_personality);
3089 module_init(raid_init);
3090 module_exit(raid_exit);
3091 MODULE_LICENSE("GPL");
3092 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3093 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3094 MODULE_ALIAS("md-raid1");
3095 MODULE_ALIAS("md-level-1");
3097 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);