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;
100 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
105 * Allocate bios : 1 for reading, n-1 for writing
107 for (j = pi->raid_disks ; j-- ; ) {
108 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
111 r1_bio->bios[j] = bio;
114 * Allocate RESYNC_PAGES data pages and attach them to
116 * If this is a user-requested check/repair, allocate
117 * RESYNC_PAGES for each bio.
119 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
124 bio = r1_bio->bios[j];
125 for (i = 0; i < RESYNC_PAGES; i++) {
126 page = alloc_page(gfp_flags);
130 bio->bi_io_vec[i].bv_page = page;
134 /* If not user-requests, copy the page pointers to all bios */
135 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
136 for (i=0; i<RESYNC_PAGES ; i++)
137 for (j=1; j<pi->raid_disks; j++)
138 r1_bio->bios[j]->bi_io_vec[i].bv_page =
139 r1_bio->bios[0]->bi_io_vec[i].bv_page;
142 r1_bio->master_bio = NULL;
147 for (j=0 ; j < pi->raid_disks; j++)
148 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
149 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
152 while (++j < pi->raid_disks)
153 bio_put(r1_bio->bios[j]);
154 r1bio_pool_free(r1_bio, data);
158 static void r1buf_pool_free(void *__r1_bio, void *data)
160 struct pool_info *pi = data;
162 struct r1bio *r1bio = __r1_bio;
164 for (i = 0; i < RESYNC_PAGES; i++)
165 for (j = pi->raid_disks; j-- ;) {
167 r1bio->bios[j]->bi_io_vec[i].bv_page !=
168 r1bio->bios[0]->bi_io_vec[i].bv_page)
169 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
171 for (i=0 ; i < pi->raid_disks; i++)
172 bio_put(r1bio->bios[i]);
174 r1bio_pool_free(r1bio, data);
177 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
181 for (i = 0; i < conf->raid_disks * 2; i++) {
182 struct bio **bio = r1_bio->bios + i;
183 if (!BIO_SPECIAL(*bio))
189 static void free_r1bio(struct r1bio *r1_bio)
191 struct r1conf *conf = r1_bio->mddev->private;
193 put_all_bios(conf, r1_bio);
194 mempool_free(r1_bio, conf->r1bio_pool);
197 static void put_buf(struct r1bio *r1_bio)
199 struct r1conf *conf = r1_bio->mddev->private;
202 for (i = 0; i < conf->raid_disks * 2; i++) {
203 struct bio *bio = r1_bio->bios[i];
205 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
208 mempool_free(r1_bio, conf->r1buf_pool);
213 static void reschedule_retry(struct r1bio *r1_bio)
216 struct mddev *mddev = r1_bio->mddev;
217 struct r1conf *conf = mddev->private;
219 spin_lock_irqsave(&conf->device_lock, flags);
220 list_add(&r1_bio->retry_list, &conf->retry_list);
222 spin_unlock_irqrestore(&conf->device_lock, flags);
224 wake_up(&conf->wait_barrier);
225 md_wakeup_thread(mddev->thread);
229 * raid_end_bio_io() is called when we have finished servicing a mirrored
230 * operation and are ready to return a success/failure code to the buffer
233 static void call_bio_endio(struct r1bio *r1_bio)
235 struct bio *bio = r1_bio->master_bio;
237 struct r1conf *conf = r1_bio->mddev->private;
239 if (bio->bi_phys_segments) {
241 spin_lock_irqsave(&conf->device_lock, flags);
242 bio->bi_phys_segments--;
243 done = (bio->bi_phys_segments == 0);
244 spin_unlock_irqrestore(&conf->device_lock, flags);
248 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
249 clear_bit(BIO_UPTODATE, &bio->bi_flags);
253 * Wake up any possible resync thread that waits for the device
260 static void raid_end_bio_io(struct r1bio *r1_bio)
262 struct bio *bio = r1_bio->master_bio;
264 /* if nobody has done the final endio yet, do it now */
265 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
266 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
267 (bio_data_dir(bio) == WRITE) ? "write" : "read",
268 (unsigned long long) bio->bi_sector,
269 (unsigned long long) bio->bi_sector +
270 (bio->bi_size >> 9) - 1);
272 call_bio_endio(r1_bio);
278 * Update disk head position estimator based on IRQ completion info.
280 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
282 struct r1conf *conf = r1_bio->mddev->private;
284 conf->mirrors[disk].head_position =
285 r1_bio->sector + (r1_bio->sectors);
289 * Find the disk number which triggered given bio
291 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
294 struct r1conf *conf = r1_bio->mddev->private;
295 int raid_disks = conf->raid_disks;
297 for (mirror = 0; mirror < raid_disks * 2; mirror++)
298 if (r1_bio->bios[mirror] == bio)
301 BUG_ON(mirror == raid_disks * 2);
302 update_head_pos(mirror, r1_bio);
307 static void raid1_end_read_request(struct bio *bio, int error)
309 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310 struct r1bio *r1_bio = bio->bi_private;
312 struct r1conf *conf = r1_bio->mddev->private;
314 mirror = r1_bio->read_disk;
316 * this branch is our 'one mirror IO has finished' event handler:
318 update_head_pos(mirror, r1_bio);
321 set_bit(R1BIO_Uptodate, &r1_bio->state);
323 /* If all other devices have failed, we want to return
324 * the error upwards rather than fail the last device.
325 * Here we redefine "uptodate" to mean "Don't want to retry"
328 spin_lock_irqsave(&conf->device_lock, flags);
329 if (r1_bio->mddev->degraded == conf->raid_disks ||
330 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
331 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
333 spin_unlock_irqrestore(&conf->device_lock, flags);
337 raid_end_bio_io(r1_bio);
338 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
343 char b[BDEVNAME_SIZE];
345 KERN_ERR "md/raid1:%s: %s: "
346 "rescheduling sector %llu\n",
348 bdevname(conf->mirrors[mirror].rdev->bdev,
350 (unsigned long long)r1_bio->sector);
351 set_bit(R1BIO_ReadError, &r1_bio->state);
352 reschedule_retry(r1_bio);
353 /* don't drop the reference on read_disk yet */
357 static void close_write(struct r1bio *r1_bio)
359 /* it really is the end of this request */
360 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
361 /* free extra copy of the data pages */
362 int i = r1_bio->behind_page_count;
364 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
365 kfree(r1_bio->behind_bvecs);
366 r1_bio->behind_bvecs = NULL;
368 /* clear the bitmap if all writes complete successfully */
369 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
371 !test_bit(R1BIO_Degraded, &r1_bio->state),
372 test_bit(R1BIO_BehindIO, &r1_bio->state));
373 md_write_end(r1_bio->mddev);
376 static void r1_bio_write_done(struct r1bio *r1_bio)
378 if (!atomic_dec_and_test(&r1_bio->remaining))
381 if (test_bit(R1BIO_WriteError, &r1_bio->state))
382 reschedule_retry(r1_bio);
385 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
386 reschedule_retry(r1_bio);
388 raid_end_bio_io(r1_bio);
392 static void raid1_end_write_request(struct bio *bio, int error)
394 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
395 struct r1bio *r1_bio = bio->bi_private;
396 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
397 struct r1conf *conf = r1_bio->mddev->private;
398 struct bio *to_put = NULL;
400 mirror = find_bio_disk(r1_bio, bio);
403 * 'one mirror IO has finished' event handler:
406 set_bit(WriteErrorSeen,
407 &conf->mirrors[mirror].rdev->flags);
408 if (!test_and_set_bit(WantReplacement,
409 &conf->mirrors[mirror].rdev->flags))
410 set_bit(MD_RECOVERY_NEEDED, &
411 conf->mddev->recovery);
413 set_bit(R1BIO_WriteError, &r1_bio->state);
416 * Set R1BIO_Uptodate in our master bio, so that we
417 * will return a good error code for to the higher
418 * levels even if IO on some other mirrored buffer
421 * The 'master' represents the composite IO operation
422 * to user-side. So if something waits for IO, then it
423 * will wait for the 'master' bio.
428 r1_bio->bios[mirror] = NULL;
431 * Do not set R1BIO_Uptodate if the current device is
432 * rebuilding or Faulty. This is because we cannot use
433 * such device for properly reading the data back (we could
434 * potentially use it, if the current write would have felt
435 * before rdev->recovery_offset, but for simplicity we don't
438 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
439 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
440 set_bit(R1BIO_Uptodate, &r1_bio->state);
442 /* Maybe we can clear some bad blocks. */
443 if (is_badblock(conf->mirrors[mirror].rdev,
444 r1_bio->sector, r1_bio->sectors,
445 &first_bad, &bad_sectors)) {
446 r1_bio->bios[mirror] = IO_MADE_GOOD;
447 set_bit(R1BIO_MadeGood, &r1_bio->state);
452 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
453 atomic_dec(&r1_bio->behind_remaining);
456 * In behind mode, we ACK the master bio once the I/O
457 * has safely reached all non-writemostly
458 * disks. Setting the Returned bit ensures that this
459 * gets done only once -- we don't ever want to return
460 * -EIO here, instead we'll wait
462 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
463 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
464 /* Maybe we can return now */
465 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
466 struct bio *mbio = r1_bio->master_bio;
467 pr_debug("raid1: behind end write sectors"
469 (unsigned long long) mbio->bi_sector,
470 (unsigned long long) mbio->bi_sector +
471 (mbio->bi_size >> 9) - 1);
472 call_bio_endio(r1_bio);
476 if (r1_bio->bios[mirror] == NULL)
477 rdev_dec_pending(conf->mirrors[mirror].rdev,
481 * Let's see if all mirrored write operations have finished
484 r1_bio_write_done(r1_bio);
492 * This routine returns the disk from which the requested read should
493 * be done. There is a per-array 'next expected sequential IO' sector
494 * number - if this matches on the next IO then we use the last disk.
495 * There is also a per-disk 'last know head position' sector that is
496 * maintained from IRQ contexts, both the normal and the resync IO
497 * completion handlers update this position correctly. If there is no
498 * perfect sequential match then we pick the disk whose head is closest.
500 * If there are 2 mirrors in the same 2 devices, performance degrades
501 * because position is mirror, not device based.
503 * The rdev for the device selected will have nr_pending incremented.
505 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
507 const sector_t this_sector = r1_bio->sector;
509 int best_good_sectors;
510 int best_disk, best_dist_disk, best_pending_disk;
514 unsigned int min_pending;
515 struct md_rdev *rdev;
517 int choose_next_idle;
521 * Check if we can balance. We can balance on the whole
522 * device if no resync is going on, or below the resync window.
523 * We take the first readable disk when above the resync window.
526 sectors = r1_bio->sectors;
529 best_dist = MaxSector;
530 best_pending_disk = -1;
531 min_pending = UINT_MAX;
532 best_good_sectors = 0;
534 choose_next_idle = 0;
536 if (conf->mddev->recovery_cp < MaxSector &&
537 (this_sector + sectors >= conf->next_resync))
542 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
546 unsigned int pending;
549 rdev = rcu_dereference(conf->mirrors[disk].rdev);
550 if (r1_bio->bios[disk] == IO_BLOCKED
552 || test_bit(Unmerged, &rdev->flags)
553 || test_bit(Faulty, &rdev->flags))
555 if (!test_bit(In_sync, &rdev->flags) &&
556 rdev->recovery_offset < this_sector + sectors)
558 if (test_bit(WriteMostly, &rdev->flags)) {
559 /* Don't balance among write-mostly, just
560 * use the first as a last resort */
562 if (is_badblock(rdev, this_sector, sectors,
563 &first_bad, &bad_sectors)) {
564 if (first_bad < this_sector)
565 /* Cannot use this */
567 best_good_sectors = first_bad - this_sector;
569 best_good_sectors = sectors;
574 /* This is a reasonable device to use. It might
577 if (is_badblock(rdev, this_sector, sectors,
578 &first_bad, &bad_sectors)) {
579 if (best_dist < MaxSector)
580 /* already have a better device */
582 if (first_bad <= this_sector) {
583 /* cannot read here. If this is the 'primary'
584 * device, then we must not read beyond
585 * bad_sectors from another device..
587 bad_sectors -= (this_sector - first_bad);
588 if (choose_first && sectors > bad_sectors)
589 sectors = bad_sectors;
590 if (best_good_sectors > sectors)
591 best_good_sectors = sectors;
594 sector_t good_sectors = first_bad - this_sector;
595 if (good_sectors > best_good_sectors) {
596 best_good_sectors = good_sectors;
604 best_good_sectors = sectors;
606 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
607 has_nonrot_disk |= nonrot;
608 pending = atomic_read(&rdev->nr_pending);
609 dist = abs(this_sector - conf->mirrors[disk].head_position);
614 /* Don't change to another disk for sequential reads */
615 if (conf->mirrors[disk].next_seq_sect == this_sector
617 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
618 struct raid1_info *mirror = &conf->mirrors[disk];
622 * If buffered sequential IO size exceeds optimal
623 * iosize, check if there is idle disk. If yes, choose
624 * the idle disk. read_balance could already choose an
625 * idle disk before noticing it's a sequential IO in
626 * this disk. This doesn't matter because this disk
627 * will idle, next time it will be utilized after the
628 * first disk has IO size exceeds optimal iosize. In
629 * this way, iosize of the first disk will be optimal
630 * iosize at least. iosize of the second disk might be
631 * small, but not a big deal since when the second disk
632 * starts IO, the first disk is likely still busy.
634 if (nonrot && opt_iosize > 0 &&
635 mirror->seq_start != MaxSector &&
636 mirror->next_seq_sect > opt_iosize &&
637 mirror->next_seq_sect - opt_iosize >=
639 choose_next_idle = 1;
644 /* If device is idle, use it */
650 if (choose_next_idle)
653 if (min_pending > pending) {
654 min_pending = pending;
655 best_pending_disk = disk;
658 if (dist < best_dist) {
660 best_dist_disk = disk;
665 * If all disks are rotational, choose the closest disk. If any disk is
666 * non-rotational, choose the disk with less pending request even the
667 * disk is rotational, which might/might not be optimal for raids with
668 * mixed ratation/non-rotational disks depending on workload.
670 if (best_disk == -1) {
672 best_disk = best_pending_disk;
674 best_disk = best_dist_disk;
677 if (best_disk >= 0) {
678 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
681 atomic_inc(&rdev->nr_pending);
682 if (test_bit(Faulty, &rdev->flags)) {
683 /* cannot risk returning a device that failed
684 * before we inc'ed nr_pending
686 rdev_dec_pending(rdev, conf->mddev);
689 sectors = best_good_sectors;
691 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
692 conf->mirrors[best_disk].seq_start = this_sector;
694 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
697 *max_sectors = sectors;
702 static int raid1_mergeable_bvec(struct request_queue *q,
703 struct bvec_merge_data *bvm,
704 struct bio_vec *biovec)
706 struct mddev *mddev = q->queuedata;
707 struct r1conf *conf = mddev->private;
708 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
709 int max = biovec->bv_len;
711 if (mddev->merge_check_needed) {
714 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
715 struct md_rdev *rdev = rcu_dereference(
716 conf->mirrors[disk].rdev);
717 if (rdev && !test_bit(Faulty, &rdev->flags)) {
718 struct request_queue *q =
719 bdev_get_queue(rdev->bdev);
720 if (q->merge_bvec_fn) {
721 bvm->bi_sector = sector +
723 bvm->bi_bdev = rdev->bdev;
724 max = min(max, q->merge_bvec_fn(
735 int md_raid1_congested(struct mddev *mddev, int bits)
737 struct r1conf *conf = mddev->private;
740 if ((bits & (1 << BDI_async_congested)) &&
741 conf->pending_count >= max_queued_requests)
745 for (i = 0; i < conf->raid_disks * 2; i++) {
746 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
747 if (rdev && !test_bit(Faulty, &rdev->flags)) {
748 struct request_queue *q = bdev_get_queue(rdev->bdev);
752 /* Note the '|| 1' - when read_balance prefers
753 * non-congested targets, it can be removed
755 if ((bits & (1<<BDI_async_congested)) || 1)
756 ret |= bdi_congested(&q->backing_dev_info, bits);
758 ret &= bdi_congested(&q->backing_dev_info, bits);
764 EXPORT_SYMBOL_GPL(md_raid1_congested);
766 static int raid1_congested(void *data, int bits)
768 struct mddev *mddev = data;
770 return mddev_congested(mddev, bits) ||
771 md_raid1_congested(mddev, bits);
774 static void flush_pending_writes(struct r1conf *conf)
776 /* Any writes that have been queued but are awaiting
777 * bitmap updates get flushed here.
779 spin_lock_irq(&conf->device_lock);
781 if (conf->pending_bio_list.head) {
783 bio = bio_list_get(&conf->pending_bio_list);
784 conf->pending_count = 0;
785 spin_unlock_irq(&conf->device_lock);
786 /* flush any pending bitmap writes to
787 * disk before proceeding w/ I/O */
788 bitmap_unplug(conf->mddev->bitmap);
789 wake_up(&conf->wait_barrier);
791 while (bio) { /* submit pending writes */
792 struct bio *next = bio->bi_next;
794 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
795 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
799 generic_make_request(bio);
803 spin_unlock_irq(&conf->device_lock);
807 * Sometimes we need to suspend IO while we do something else,
808 * either some resync/recovery, or reconfigure the array.
809 * To do this we raise a 'barrier'.
810 * The 'barrier' is a counter that can be raised multiple times
811 * to count how many activities are happening which preclude
813 * We can only raise the barrier if there is no pending IO.
814 * i.e. if nr_pending == 0.
815 * We choose only to raise the barrier if no-one is waiting for the
816 * barrier to go down. This means that as soon as an IO request
817 * is ready, no other operations which require a barrier will start
818 * until the IO request has had a chance.
820 * So: regular IO calls 'wait_barrier'. When that returns there
821 * is no backgroup IO happening, It must arrange to call
822 * allow_barrier when it has finished its IO.
823 * backgroup IO calls must call raise_barrier. Once that returns
824 * there is no normal IO happeing. It must arrange to call
825 * lower_barrier when the particular background IO completes.
827 #define RESYNC_DEPTH 32
829 static void raise_barrier(struct r1conf *conf)
831 spin_lock_irq(&conf->resync_lock);
833 /* Wait until no block IO is waiting */
834 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
837 /* block any new IO from starting */
840 /* Now wait for all pending IO to complete */
841 wait_event_lock_irq(conf->wait_barrier,
842 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
845 spin_unlock_irq(&conf->resync_lock);
848 static void lower_barrier(struct r1conf *conf)
851 BUG_ON(conf->barrier <= 0);
852 spin_lock_irqsave(&conf->resync_lock, flags);
854 spin_unlock_irqrestore(&conf->resync_lock, flags);
855 wake_up(&conf->wait_barrier);
858 static void wait_barrier(struct r1conf *conf)
860 spin_lock_irq(&conf->resync_lock);
863 /* Wait for the barrier to drop.
864 * However if there are already pending
865 * requests (preventing the barrier from
866 * rising completely), and the
867 * pre-process bio queue isn't empty,
868 * then don't wait, as we need to empty
869 * that queue to get the nr_pending
872 wait_event_lock_irq(conf->wait_barrier,
876 !bio_list_empty(current->bio_list)),
881 spin_unlock_irq(&conf->resync_lock);
884 static void allow_barrier(struct r1conf *conf)
887 spin_lock_irqsave(&conf->resync_lock, flags);
889 spin_unlock_irqrestore(&conf->resync_lock, flags);
890 wake_up(&conf->wait_barrier);
893 static void freeze_array(struct r1conf *conf, int extra)
895 /* stop syncio and normal IO and wait for everything to
897 * We increment barrier and nr_waiting, and then
898 * wait until nr_pending match nr_queued+extra
899 * This is called in the context of one normal IO request
900 * that has failed. Thus any sync request that might be pending
901 * will be blocked by nr_pending, and we need to wait for
902 * pending IO requests to complete or be queued for re-try.
903 * Thus the number queued (nr_queued) plus this request (extra)
904 * must match the number of pending IOs (nr_pending) before
907 spin_lock_irq(&conf->resync_lock);
910 wait_event_lock_irq_cmd(conf->wait_barrier,
911 conf->nr_pending == conf->nr_queued+extra,
913 flush_pending_writes(conf));
914 spin_unlock_irq(&conf->resync_lock);
916 static void unfreeze_array(struct r1conf *conf)
918 /* reverse the effect of the freeze */
919 spin_lock_irq(&conf->resync_lock);
922 wake_up(&conf->wait_barrier);
923 spin_unlock_irq(&conf->resync_lock);
927 /* duplicate the data pages for behind I/O
929 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
932 struct bio_vec *bvec;
933 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
935 if (unlikely(!bvecs))
938 bio_for_each_segment(bvec, bio, i) {
940 bvecs[i].bv_page = alloc_page(GFP_NOIO);
941 if (unlikely(!bvecs[i].bv_page))
943 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
944 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
945 kunmap(bvecs[i].bv_page);
946 kunmap(bvec->bv_page);
948 r1_bio->behind_bvecs = bvecs;
949 r1_bio->behind_page_count = bio->bi_vcnt;
950 set_bit(R1BIO_BehindIO, &r1_bio->state);
954 for (i = 0; i < bio->bi_vcnt; i++)
955 if (bvecs[i].bv_page)
956 put_page(bvecs[i].bv_page);
958 pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
961 struct raid1_plug_cb {
962 struct blk_plug_cb cb;
963 struct bio_list pending;
967 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
969 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
971 struct mddev *mddev = plug->cb.data;
972 struct r1conf *conf = mddev->private;
975 if (from_schedule || current->bio_list) {
976 spin_lock_irq(&conf->device_lock);
977 bio_list_merge(&conf->pending_bio_list, &plug->pending);
978 conf->pending_count += plug->pending_cnt;
979 spin_unlock_irq(&conf->device_lock);
980 wake_up(&conf->wait_barrier);
981 md_wakeup_thread(mddev->thread);
986 /* we aren't scheduling, so we can do the write-out directly. */
987 bio = bio_list_get(&plug->pending);
988 bitmap_unplug(mddev->bitmap);
989 wake_up(&conf->wait_barrier);
991 while (bio) { /* submit pending writes */
992 struct bio *next = bio->bi_next;
994 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
995 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
999 generic_make_request(bio);
1005 static void make_request(struct mddev *mddev, struct bio * bio)
1007 struct r1conf *conf = mddev->private;
1008 struct raid1_info *mirror;
1009 struct r1bio *r1_bio;
1010 struct bio *read_bio;
1012 struct bitmap *bitmap;
1013 unsigned long flags;
1014 const int rw = bio_data_dir(bio);
1015 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1016 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1017 const unsigned long do_discard = (bio->bi_rw
1018 & (REQ_DISCARD | REQ_SECURE));
1019 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1020 struct md_rdev *blocked_rdev;
1021 struct blk_plug_cb *cb;
1022 struct raid1_plug_cb *plug = NULL;
1024 int sectors_handled;
1028 * Register the new request and wait if the reconstruction
1029 * thread has put up a bar for new requests.
1030 * Continue immediately if no resync is active currently.
1033 md_write_start(mddev, bio); /* wait on superblock update early */
1035 if (bio_data_dir(bio) == WRITE &&
1036 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
1037 bio->bi_sector < mddev->suspend_hi) {
1038 /* As the suspend_* range is controlled by
1039 * userspace, we want an interruptible
1044 flush_signals(current);
1045 prepare_to_wait(&conf->wait_barrier,
1046 &w, TASK_INTERRUPTIBLE);
1047 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
1048 bio->bi_sector >= mddev->suspend_hi)
1052 finish_wait(&conf->wait_barrier, &w);
1057 bitmap = mddev->bitmap;
1060 * make_request() can abort the operation when READA is being
1061 * used and no empty request is available.
1064 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1066 r1_bio->master_bio = bio;
1067 r1_bio->sectors = bio->bi_size >> 9;
1069 r1_bio->mddev = mddev;
1070 r1_bio->sector = bio->bi_sector;
1072 /* We might need to issue multiple reads to different
1073 * devices if there are bad blocks around, so we keep
1074 * track of the number of reads in bio->bi_phys_segments.
1075 * If this is 0, there is only one r1_bio and no locking
1076 * will be needed when requests complete. If it is
1077 * non-zero, then it is the number of not-completed requests.
1079 bio->bi_phys_segments = 0;
1080 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1084 * read balancing logic:
1089 rdisk = read_balance(conf, r1_bio, &max_sectors);
1092 /* couldn't find anywhere to read from */
1093 raid_end_bio_io(r1_bio);
1096 mirror = conf->mirrors + rdisk;
1098 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1100 /* Reading from a write-mostly device must
1101 * take care not to over-take any writes
1104 wait_event(bitmap->behind_wait,
1105 atomic_read(&bitmap->behind_writes) == 0);
1107 r1_bio->read_disk = rdisk;
1109 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1110 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1113 r1_bio->bios[rdisk] = read_bio;
1115 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1116 read_bio->bi_bdev = mirror->rdev->bdev;
1117 read_bio->bi_end_io = raid1_end_read_request;
1118 read_bio->bi_rw = READ | do_sync;
1119 read_bio->bi_private = r1_bio;
1121 if (max_sectors < r1_bio->sectors) {
1122 /* could not read all from this device, so we will
1123 * need another r1_bio.
1126 sectors_handled = (r1_bio->sector + max_sectors
1128 r1_bio->sectors = max_sectors;
1129 spin_lock_irq(&conf->device_lock);
1130 if (bio->bi_phys_segments == 0)
1131 bio->bi_phys_segments = 2;
1133 bio->bi_phys_segments++;
1134 spin_unlock_irq(&conf->device_lock);
1135 /* Cannot call generic_make_request directly
1136 * as that will be queued in __make_request
1137 * and subsequent mempool_alloc might block waiting
1138 * for it. So hand bio over to raid1d.
1140 reschedule_retry(r1_bio);
1142 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1144 r1_bio->master_bio = bio;
1145 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1147 r1_bio->mddev = mddev;
1148 r1_bio->sector = bio->bi_sector + sectors_handled;
1151 generic_make_request(read_bio);
1158 if (conf->pending_count >= max_queued_requests) {
1159 md_wakeup_thread(mddev->thread);
1160 wait_event(conf->wait_barrier,
1161 conf->pending_count < max_queued_requests);
1163 /* first select target devices under rcu_lock and
1164 * inc refcount on their rdev. Record them by setting
1166 * If there are known/acknowledged bad blocks on any device on
1167 * which we have seen a write error, we want to avoid writing those
1169 * This potentially requires several writes to write around
1170 * the bad blocks. Each set of writes gets it's own r1bio
1171 * with a set of bios attached.
1174 disks = conf->raid_disks * 2;
1176 blocked_rdev = NULL;
1178 max_sectors = r1_bio->sectors;
1179 for (i = 0; i < disks; i++) {
1180 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1181 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1182 atomic_inc(&rdev->nr_pending);
1183 blocked_rdev = rdev;
1186 r1_bio->bios[i] = NULL;
1187 if (!rdev || test_bit(Faulty, &rdev->flags)
1188 || test_bit(Unmerged, &rdev->flags)) {
1189 if (i < conf->raid_disks)
1190 set_bit(R1BIO_Degraded, &r1_bio->state);
1194 atomic_inc(&rdev->nr_pending);
1195 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1200 is_bad = is_badblock(rdev, r1_bio->sector,
1202 &first_bad, &bad_sectors);
1204 /* mustn't write here until the bad block is
1206 set_bit(BlockedBadBlocks, &rdev->flags);
1207 blocked_rdev = rdev;
1210 if (is_bad && first_bad <= r1_bio->sector) {
1211 /* Cannot write here at all */
1212 bad_sectors -= (r1_bio->sector - first_bad);
1213 if (bad_sectors < max_sectors)
1214 /* mustn't write more than bad_sectors
1215 * to other devices yet
1217 max_sectors = bad_sectors;
1218 rdev_dec_pending(rdev, mddev);
1219 /* We don't set R1BIO_Degraded as that
1220 * only applies if the disk is
1221 * missing, so it might be re-added,
1222 * and we want to know to recover this
1224 * In this case the device is here,
1225 * and the fact that this chunk is not
1226 * in-sync is recorded in the bad
1232 int good_sectors = first_bad - r1_bio->sector;
1233 if (good_sectors < max_sectors)
1234 max_sectors = good_sectors;
1237 r1_bio->bios[i] = bio;
1241 if (unlikely(blocked_rdev)) {
1242 /* Wait for this device to become unblocked */
1245 for (j = 0; j < i; j++)
1246 if (r1_bio->bios[j])
1247 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1249 allow_barrier(conf);
1250 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1255 if (max_sectors < r1_bio->sectors) {
1256 /* We are splitting this write into multiple parts, so
1257 * we need to prepare for allocating another r1_bio.
1259 r1_bio->sectors = max_sectors;
1260 spin_lock_irq(&conf->device_lock);
1261 if (bio->bi_phys_segments == 0)
1262 bio->bi_phys_segments = 2;
1264 bio->bi_phys_segments++;
1265 spin_unlock_irq(&conf->device_lock);
1267 sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1269 atomic_set(&r1_bio->remaining, 1);
1270 atomic_set(&r1_bio->behind_remaining, 0);
1273 for (i = 0; i < disks; i++) {
1275 if (!r1_bio->bios[i])
1278 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1279 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1283 * Not if there are too many, or cannot
1284 * allocate memory, or a reader on WriteMostly
1285 * is waiting for behind writes to flush */
1287 (atomic_read(&bitmap->behind_writes)
1288 < mddev->bitmap_info.max_write_behind) &&
1289 !waitqueue_active(&bitmap->behind_wait))
1290 alloc_behind_pages(mbio, r1_bio);
1292 bitmap_startwrite(bitmap, r1_bio->sector,
1294 test_bit(R1BIO_BehindIO,
1298 if (r1_bio->behind_bvecs) {
1299 struct bio_vec *bvec;
1302 /* Yes, I really want the '__' version so that
1303 * we clear any unused pointer in the io_vec, rather
1304 * than leave them unchanged. This is important
1305 * because when we come to free the pages, we won't
1306 * know the original bi_idx, so we just free
1309 __bio_for_each_segment(bvec, mbio, j, 0)
1310 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1311 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1312 atomic_inc(&r1_bio->behind_remaining);
1315 r1_bio->bios[i] = mbio;
1317 mbio->bi_sector = (r1_bio->sector +
1318 conf->mirrors[i].rdev->data_offset);
1319 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1320 mbio->bi_end_io = raid1_end_write_request;
1322 WRITE | do_flush_fua | do_sync | do_discard | do_same;
1323 mbio->bi_private = r1_bio;
1325 atomic_inc(&r1_bio->remaining);
1327 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1329 plug = container_of(cb, struct raid1_plug_cb, cb);
1332 spin_lock_irqsave(&conf->device_lock, flags);
1334 bio_list_add(&plug->pending, mbio);
1335 plug->pending_cnt++;
1337 bio_list_add(&conf->pending_bio_list, mbio);
1338 conf->pending_count++;
1340 spin_unlock_irqrestore(&conf->device_lock, flags);
1342 md_wakeup_thread(mddev->thread);
1344 /* Mustn't call r1_bio_write_done before this next test,
1345 * as it could result in the bio being freed.
1347 if (sectors_handled < (bio->bi_size >> 9)) {
1348 r1_bio_write_done(r1_bio);
1349 /* We need another r1_bio. It has already been counted
1350 * in bio->bi_phys_segments
1352 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1353 r1_bio->master_bio = bio;
1354 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1356 r1_bio->mddev = mddev;
1357 r1_bio->sector = bio->bi_sector + sectors_handled;
1361 r1_bio_write_done(r1_bio);
1363 /* In case raid1d snuck in to freeze_array */
1364 wake_up(&conf->wait_barrier);
1367 static void status(struct seq_file *seq, struct mddev *mddev)
1369 struct r1conf *conf = mddev->private;
1372 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1373 conf->raid_disks - mddev->degraded);
1375 for (i = 0; i < conf->raid_disks; i++) {
1376 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1377 seq_printf(seq, "%s",
1378 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1381 seq_printf(seq, "]");
1385 static void error(struct mddev *mddev, struct md_rdev *rdev)
1387 char b[BDEVNAME_SIZE];
1388 struct r1conf *conf = mddev->private;
1391 * If it is not operational, then we have already marked it as dead
1392 * else if it is the last working disks, ignore the error, let the
1393 * next level up know.
1394 * else mark the drive as failed
1396 if (test_bit(In_sync, &rdev->flags)
1397 && (conf->raid_disks - mddev->degraded) == 1) {
1399 * Don't fail the drive, act as though we were just a
1400 * normal single drive.
1401 * However don't try a recovery from this drive as
1402 * it is very likely to fail.
1404 conf->recovery_disabled = mddev->recovery_disabled;
1407 set_bit(Blocked, &rdev->flags);
1408 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1409 unsigned long flags;
1410 spin_lock_irqsave(&conf->device_lock, flags);
1412 set_bit(Faulty, &rdev->flags);
1413 spin_unlock_irqrestore(&conf->device_lock, flags);
1415 * if recovery is running, make sure it aborts.
1417 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1419 set_bit(Faulty, &rdev->flags);
1420 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1422 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1423 "md/raid1:%s: Operation continuing on %d devices.\n",
1424 mdname(mddev), bdevname(rdev->bdev, b),
1425 mdname(mddev), conf->raid_disks - mddev->degraded);
1428 static void print_conf(struct r1conf *conf)
1432 printk(KERN_DEBUG "RAID1 conf printout:\n");
1434 printk(KERN_DEBUG "(!conf)\n");
1437 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1441 for (i = 0; i < conf->raid_disks; i++) {
1442 char b[BDEVNAME_SIZE];
1443 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1445 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1446 i, !test_bit(In_sync, &rdev->flags),
1447 !test_bit(Faulty, &rdev->flags),
1448 bdevname(rdev->bdev,b));
1453 static void close_sync(struct r1conf *conf)
1456 allow_barrier(conf);
1458 mempool_destroy(conf->r1buf_pool);
1459 conf->r1buf_pool = NULL;
1462 static int raid1_spare_active(struct mddev *mddev)
1465 struct r1conf *conf = mddev->private;
1467 unsigned long flags;
1470 * Find all failed disks within the RAID1 configuration
1471 * and mark them readable.
1472 * Called under mddev lock, so rcu protection not needed.
1474 for (i = 0; i < conf->raid_disks; i++) {
1475 struct md_rdev *rdev = conf->mirrors[i].rdev;
1476 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1478 && repl->recovery_offset == MaxSector
1479 && !test_bit(Faulty, &repl->flags)
1480 && !test_and_set_bit(In_sync, &repl->flags)) {
1481 /* replacement has just become active */
1483 !test_and_clear_bit(In_sync, &rdev->flags))
1486 /* Replaced device not technically
1487 * faulty, but we need to be sure
1488 * it gets removed and never re-added
1490 set_bit(Faulty, &rdev->flags);
1491 sysfs_notify_dirent_safe(
1496 && !test_bit(Faulty, &rdev->flags)
1497 && !test_and_set_bit(In_sync, &rdev->flags)) {
1499 sysfs_notify_dirent_safe(rdev->sysfs_state);
1502 spin_lock_irqsave(&conf->device_lock, flags);
1503 mddev->degraded -= count;
1504 spin_unlock_irqrestore(&conf->device_lock, flags);
1511 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1513 struct r1conf *conf = mddev->private;
1516 struct raid1_info *p;
1518 int last = conf->raid_disks - 1;
1519 struct request_queue *q = bdev_get_queue(rdev->bdev);
1521 if (mddev->recovery_disabled == conf->recovery_disabled)
1524 if (rdev->raid_disk >= 0)
1525 first = last = rdev->raid_disk;
1527 if (q->merge_bvec_fn) {
1528 set_bit(Unmerged, &rdev->flags);
1529 mddev->merge_check_needed = 1;
1532 for (mirror = first; mirror <= last; mirror++) {
1533 p = conf->mirrors+mirror;
1536 disk_stack_limits(mddev->gendisk, rdev->bdev,
1537 rdev->data_offset << 9);
1539 p->head_position = 0;
1540 rdev->raid_disk = mirror;
1542 /* As all devices are equivalent, we don't need a full recovery
1543 * if this was recently any drive of the array
1545 if (rdev->saved_raid_disk < 0)
1547 rcu_assign_pointer(p->rdev, rdev);
1550 if (test_bit(WantReplacement, &p->rdev->flags) &&
1551 p[conf->raid_disks].rdev == NULL) {
1552 /* Add this device as a replacement */
1553 clear_bit(In_sync, &rdev->flags);
1554 set_bit(Replacement, &rdev->flags);
1555 rdev->raid_disk = mirror;
1558 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1562 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1563 /* Some requests might not have seen this new
1564 * merge_bvec_fn. We must wait for them to complete
1565 * before merging the device fully.
1566 * First we make sure any code which has tested
1567 * our function has submitted the request, then
1568 * we wait for all outstanding requests to complete.
1570 synchronize_sched();
1571 freeze_array(conf, 0);
1572 unfreeze_array(conf);
1573 clear_bit(Unmerged, &rdev->flags);
1575 md_integrity_add_rdev(rdev, mddev);
1576 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
1577 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1582 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1584 struct r1conf *conf = mddev->private;
1586 int number = rdev->raid_disk;
1587 struct raid1_info *p = conf->mirrors + number;
1589 if (rdev != p->rdev)
1590 p = conf->mirrors + conf->raid_disks + number;
1593 if (rdev == p->rdev) {
1594 if (test_bit(In_sync, &rdev->flags) ||
1595 atomic_read(&rdev->nr_pending)) {
1599 /* Only remove non-faulty devices if recovery
1602 if (!test_bit(Faulty, &rdev->flags) &&
1603 mddev->recovery_disabled != conf->recovery_disabled &&
1604 mddev->degraded < conf->raid_disks) {
1610 if (atomic_read(&rdev->nr_pending)) {
1611 /* lost the race, try later */
1615 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1616 /* We just removed a device that is being replaced.
1617 * Move down the replacement. We drain all IO before
1618 * doing this to avoid confusion.
1620 struct md_rdev *repl =
1621 conf->mirrors[conf->raid_disks + number].rdev;
1622 freeze_array(conf, 0);
1623 clear_bit(Replacement, &repl->flags);
1625 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1626 unfreeze_array(conf);
1627 clear_bit(WantReplacement, &rdev->flags);
1629 clear_bit(WantReplacement, &rdev->flags);
1630 err = md_integrity_register(mddev);
1639 static void end_sync_read(struct bio *bio, int error)
1641 struct r1bio *r1_bio = bio->bi_private;
1643 update_head_pos(r1_bio->read_disk, r1_bio);
1646 * we have read a block, now it needs to be re-written,
1647 * or re-read if the read failed.
1648 * We don't do much here, just schedule handling by raid1d
1650 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1651 set_bit(R1BIO_Uptodate, &r1_bio->state);
1653 if (atomic_dec_and_test(&r1_bio->remaining))
1654 reschedule_retry(r1_bio);
1657 static void end_sync_write(struct bio *bio, int error)
1659 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1660 struct r1bio *r1_bio = bio->bi_private;
1661 struct mddev *mddev = r1_bio->mddev;
1662 struct r1conf *conf = mddev->private;
1667 mirror = find_bio_disk(r1_bio, bio);
1670 sector_t sync_blocks = 0;
1671 sector_t s = r1_bio->sector;
1672 long sectors_to_go = r1_bio->sectors;
1673 /* make sure these bits doesn't get cleared. */
1675 bitmap_end_sync(mddev->bitmap, s,
1678 sectors_to_go -= sync_blocks;
1679 } while (sectors_to_go > 0);
1680 set_bit(WriteErrorSeen,
1681 &conf->mirrors[mirror].rdev->flags);
1682 if (!test_and_set_bit(WantReplacement,
1683 &conf->mirrors[mirror].rdev->flags))
1684 set_bit(MD_RECOVERY_NEEDED, &
1686 set_bit(R1BIO_WriteError, &r1_bio->state);
1687 } else if (is_badblock(conf->mirrors[mirror].rdev,
1690 &first_bad, &bad_sectors) &&
1691 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1694 &first_bad, &bad_sectors)
1696 set_bit(R1BIO_MadeGood, &r1_bio->state);
1698 if (atomic_dec_and_test(&r1_bio->remaining)) {
1699 int s = r1_bio->sectors;
1700 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1701 test_bit(R1BIO_WriteError, &r1_bio->state))
1702 reschedule_retry(r1_bio);
1705 md_done_sync(mddev, s, uptodate);
1710 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1711 int sectors, struct page *page, int rw)
1713 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1717 set_bit(WriteErrorSeen, &rdev->flags);
1718 if (!test_and_set_bit(WantReplacement,
1720 set_bit(MD_RECOVERY_NEEDED, &
1721 rdev->mddev->recovery);
1723 /* need to record an error - either for the block or the device */
1724 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1725 md_error(rdev->mddev, rdev);
1729 static int fix_sync_read_error(struct r1bio *r1_bio)
1731 /* Try some synchronous reads of other devices to get
1732 * good data, much like with normal read errors. Only
1733 * read into the pages we already have so we don't
1734 * need to re-issue the read request.
1735 * We don't need to freeze the array, because being in an
1736 * active sync request, there is no normal IO, and
1737 * no overlapping syncs.
1738 * We don't need to check is_badblock() again as we
1739 * made sure that anything with a bad block in range
1740 * will have bi_end_io clear.
1742 struct mddev *mddev = r1_bio->mddev;
1743 struct r1conf *conf = mddev->private;
1744 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1745 sector_t sect = r1_bio->sector;
1746 int sectors = r1_bio->sectors;
1751 int d = r1_bio->read_disk;
1753 struct md_rdev *rdev;
1756 if (s > (PAGE_SIZE>>9))
1759 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1760 /* No rcu protection needed here devices
1761 * can only be removed when no resync is
1762 * active, and resync is currently active
1764 rdev = conf->mirrors[d].rdev;
1765 if (sync_page_io(rdev, sect, s<<9,
1766 bio->bi_io_vec[idx].bv_page,
1773 if (d == conf->raid_disks * 2)
1775 } while (!success && d != r1_bio->read_disk);
1778 char b[BDEVNAME_SIZE];
1780 /* Cannot read from anywhere, this block is lost.
1781 * Record a bad block on each device. If that doesn't
1782 * work just disable and interrupt the recovery.
1783 * Don't fail devices as that won't really help.
1785 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1786 " for block %llu\n",
1788 bdevname(bio->bi_bdev, b),
1789 (unsigned long long)r1_bio->sector);
1790 for (d = 0; d < conf->raid_disks * 2; d++) {
1791 rdev = conf->mirrors[d].rdev;
1792 if (!rdev || test_bit(Faulty, &rdev->flags))
1794 if (!rdev_set_badblocks(rdev, sect, s, 0))
1798 conf->recovery_disabled =
1799 mddev->recovery_disabled;
1800 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1801 md_done_sync(mddev, r1_bio->sectors, 0);
1813 /* write it back and re-read */
1814 while (d != r1_bio->read_disk) {
1816 d = conf->raid_disks * 2;
1818 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1820 rdev = conf->mirrors[d].rdev;
1821 if (r1_sync_page_io(rdev, sect, s,
1822 bio->bi_io_vec[idx].bv_page,
1824 r1_bio->bios[d]->bi_end_io = NULL;
1825 rdev_dec_pending(rdev, mddev);
1829 while (d != r1_bio->read_disk) {
1831 d = conf->raid_disks * 2;
1833 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1835 rdev = conf->mirrors[d].rdev;
1836 if (r1_sync_page_io(rdev, sect, s,
1837 bio->bi_io_vec[idx].bv_page,
1839 atomic_add(s, &rdev->corrected_errors);
1845 set_bit(R1BIO_Uptodate, &r1_bio->state);
1846 set_bit(BIO_UPTODATE, &bio->bi_flags);
1850 static int process_checks(struct r1bio *r1_bio)
1852 /* We have read all readable devices. If we haven't
1853 * got the block, then there is no hope left.
1854 * If we have, then we want to do a comparison
1855 * and skip the write if everything is the same.
1856 * If any blocks failed to read, then we need to
1857 * attempt an over-write
1859 struct mddev *mddev = r1_bio->mddev;
1860 struct r1conf *conf = mddev->private;
1865 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1866 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1867 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1868 r1_bio->bios[primary]->bi_end_io = NULL;
1869 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1872 r1_bio->read_disk = primary;
1873 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1874 for (i = 0; i < conf->raid_disks * 2; i++) {
1876 struct bio *pbio = r1_bio->bios[primary];
1877 struct bio *sbio = r1_bio->bios[i];
1880 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1883 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1884 for (j = vcnt; j-- ; ) {
1886 p = pbio->bi_io_vec[j].bv_page;
1887 s = sbio->bi_io_vec[j].bv_page;
1888 if (memcmp(page_address(p),
1890 sbio->bi_io_vec[j].bv_len))
1896 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1897 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1898 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1899 /* No need to write to this device. */
1900 sbio->bi_end_io = NULL;
1901 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1904 /* fixup the bio for reuse */
1905 sbio->bi_vcnt = vcnt;
1906 sbio->bi_size = r1_bio->sectors << 9;
1908 sbio->bi_phys_segments = 0;
1909 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1910 sbio->bi_flags |= 1 << BIO_UPTODATE;
1911 sbio->bi_next = NULL;
1912 sbio->bi_sector = r1_bio->sector +
1913 conf->mirrors[i].rdev->data_offset;
1914 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1915 size = sbio->bi_size;
1916 for (j = 0; j < vcnt ; j++) {
1918 bi = &sbio->bi_io_vec[j];
1920 if (size > PAGE_SIZE)
1921 bi->bv_len = PAGE_SIZE;
1925 memcpy(page_address(bi->bv_page),
1926 page_address(pbio->bi_io_vec[j].bv_page),
1933 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1935 struct r1conf *conf = mddev->private;
1937 int disks = conf->raid_disks * 2;
1938 struct bio *bio, *wbio;
1940 bio = r1_bio->bios[r1_bio->read_disk];
1942 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1943 /* ouch - failed to read all of that. */
1944 if (!fix_sync_read_error(r1_bio))
1947 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1948 if (process_checks(r1_bio) < 0)
1953 atomic_set(&r1_bio->remaining, 1);
1954 for (i = 0; i < disks ; i++) {
1955 wbio = r1_bio->bios[i];
1956 if (wbio->bi_end_io == NULL ||
1957 (wbio->bi_end_io == end_sync_read &&
1958 (i == r1_bio->read_disk ||
1959 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1962 wbio->bi_rw = WRITE;
1963 wbio->bi_end_io = end_sync_write;
1964 atomic_inc(&r1_bio->remaining);
1965 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1967 generic_make_request(wbio);
1970 if (atomic_dec_and_test(&r1_bio->remaining)) {
1971 /* if we're here, all write(s) have completed, so clean up */
1972 int s = r1_bio->sectors;
1973 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1974 test_bit(R1BIO_WriteError, &r1_bio->state))
1975 reschedule_retry(r1_bio);
1978 md_done_sync(mddev, s, 1);
1984 * This is a kernel thread which:
1986 * 1. Retries failed read operations on working mirrors.
1987 * 2. Updates the raid superblock when problems encounter.
1988 * 3. Performs writes following reads for array synchronising.
1991 static void fix_read_error(struct r1conf *conf, int read_disk,
1992 sector_t sect, int sectors)
1994 struct mddev *mddev = conf->mddev;
2000 struct md_rdev *rdev;
2002 if (s > (PAGE_SIZE>>9))
2006 /* Note: no rcu protection needed here
2007 * as this is synchronous in the raid1d thread
2008 * which is the thread that might remove
2009 * a device. If raid1d ever becomes multi-threaded....
2014 rdev = conf->mirrors[d].rdev;
2016 (test_bit(In_sync, &rdev->flags) ||
2017 (!test_bit(Faulty, &rdev->flags) &&
2018 rdev->recovery_offset >= sect + s)) &&
2019 is_badblock(rdev, sect, s,
2020 &first_bad, &bad_sectors) == 0 &&
2021 sync_page_io(rdev, sect, s<<9,
2022 conf->tmppage, READ, false))
2026 if (d == conf->raid_disks * 2)
2029 } while (!success && d != read_disk);
2032 /* Cannot read from anywhere - mark it bad */
2033 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2034 if (!rdev_set_badblocks(rdev, sect, s, 0))
2035 md_error(mddev, rdev);
2038 /* write it back and re-read */
2040 while (d != read_disk) {
2042 d = conf->raid_disks * 2;
2044 rdev = conf->mirrors[d].rdev;
2046 test_bit(In_sync, &rdev->flags))
2047 r1_sync_page_io(rdev, sect, s,
2048 conf->tmppage, WRITE);
2051 while (d != read_disk) {
2052 char b[BDEVNAME_SIZE];
2054 d = conf->raid_disks * 2;
2056 rdev = conf->mirrors[d].rdev;
2058 test_bit(In_sync, &rdev->flags)) {
2059 if (r1_sync_page_io(rdev, sect, s,
2060 conf->tmppage, READ)) {
2061 atomic_add(s, &rdev->corrected_errors);
2063 "md/raid1:%s: read error corrected "
2064 "(%d sectors at %llu on %s)\n",
2066 (unsigned long long)(sect +
2068 bdevname(rdev->bdev, b));
2077 static void bi_complete(struct bio *bio, int error)
2079 complete((struct completion *)bio->bi_private);
2082 static int submit_bio_wait(int rw, struct bio *bio)
2084 struct completion event;
2087 init_completion(&event);
2088 bio->bi_private = &event;
2089 bio->bi_end_io = bi_complete;
2090 submit_bio(rw, bio);
2091 wait_for_completion(&event);
2093 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2096 static int narrow_write_error(struct r1bio *r1_bio, int i)
2098 struct mddev *mddev = r1_bio->mddev;
2099 struct r1conf *conf = mddev->private;
2100 struct md_rdev *rdev = conf->mirrors[i].rdev;
2102 struct bio_vec *vec;
2104 /* bio has the data to be written to device 'i' where
2105 * we just recently had a write error.
2106 * We repeatedly clone the bio and trim down to one block,
2107 * then try the write. Where the write fails we record
2109 * It is conceivable that the bio doesn't exactly align with
2110 * blocks. We must handle this somehow.
2112 * We currently own a reference on the rdev.
2118 int sect_to_write = r1_bio->sectors;
2121 if (rdev->badblocks.shift < 0)
2124 block_sectors = 1 << rdev->badblocks.shift;
2125 sector = r1_bio->sector;
2126 sectors = ((sector + block_sectors)
2127 & ~(sector_t)(block_sectors - 1))
2130 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2131 vcnt = r1_bio->behind_page_count;
2132 vec = r1_bio->behind_bvecs;
2134 while (vec[idx].bv_page == NULL)
2137 vcnt = r1_bio->master_bio->bi_vcnt;
2138 vec = r1_bio->master_bio->bi_io_vec;
2139 idx = r1_bio->master_bio->bi_idx;
2141 while (sect_to_write) {
2143 if (sectors > sect_to_write)
2144 sectors = sect_to_write;
2145 /* Write at 'sector' for 'sectors'*/
2147 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2148 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2149 wbio->bi_sector = r1_bio->sector;
2150 wbio->bi_rw = WRITE;
2151 wbio->bi_vcnt = vcnt;
2152 wbio->bi_size = r1_bio->sectors << 9;
2155 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2156 wbio->bi_sector += rdev->data_offset;
2157 wbio->bi_bdev = rdev->bdev;
2158 if (submit_bio_wait(WRITE, wbio) == 0)
2160 ok = rdev_set_badblocks(rdev, sector,
2165 sect_to_write -= sectors;
2167 sectors = block_sectors;
2172 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2175 int s = r1_bio->sectors;
2176 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2177 struct md_rdev *rdev = conf->mirrors[m].rdev;
2178 struct bio *bio = r1_bio->bios[m];
2179 if (bio->bi_end_io == NULL)
2181 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2182 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2183 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2185 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2186 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2187 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2188 md_error(conf->mddev, rdev);
2192 md_done_sync(conf->mddev, s, 1);
2195 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2198 for (m = 0; m < conf->raid_disks * 2 ; m++)
2199 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2200 struct md_rdev *rdev = conf->mirrors[m].rdev;
2201 rdev_clear_badblocks(rdev,
2203 r1_bio->sectors, 0);
2204 rdev_dec_pending(rdev, conf->mddev);
2205 } else if (r1_bio->bios[m] != NULL) {
2206 /* This drive got a write error. We need to
2207 * narrow down and record precise write
2210 if (!narrow_write_error(r1_bio, m)) {
2211 md_error(conf->mddev,
2212 conf->mirrors[m].rdev);
2213 /* an I/O failed, we can't clear the bitmap */
2214 set_bit(R1BIO_Degraded, &r1_bio->state);
2216 rdev_dec_pending(conf->mirrors[m].rdev,
2219 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2220 close_write(r1_bio);
2221 raid_end_bio_io(r1_bio);
2224 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2228 struct mddev *mddev = conf->mddev;
2230 char b[BDEVNAME_SIZE];
2231 struct md_rdev *rdev;
2233 clear_bit(R1BIO_ReadError, &r1_bio->state);
2234 /* we got a read error. Maybe the drive is bad. Maybe just
2235 * the block and we can fix it.
2236 * We freeze all other IO, and try reading the block from
2237 * other devices. When we find one, we re-write
2238 * and check it that fixes the read error.
2239 * This is all done synchronously while the array is
2242 if (mddev->ro == 0) {
2243 freeze_array(conf, 1);
2244 fix_read_error(conf, r1_bio->read_disk,
2245 r1_bio->sector, r1_bio->sectors);
2246 unfreeze_array(conf);
2248 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2249 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2251 bio = r1_bio->bios[r1_bio->read_disk];
2252 bdevname(bio->bi_bdev, b);
2254 disk = read_balance(conf, r1_bio, &max_sectors);
2256 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2257 " read error for block %llu\n",
2258 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2259 raid_end_bio_io(r1_bio);
2261 const unsigned long do_sync
2262 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2264 r1_bio->bios[r1_bio->read_disk] =
2265 mddev->ro ? IO_BLOCKED : NULL;
2268 r1_bio->read_disk = disk;
2269 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2270 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2271 r1_bio->bios[r1_bio->read_disk] = bio;
2272 rdev = conf->mirrors[disk].rdev;
2273 printk_ratelimited(KERN_ERR
2274 "md/raid1:%s: redirecting sector %llu"
2275 " to other mirror: %s\n",
2277 (unsigned long long)r1_bio->sector,
2278 bdevname(rdev->bdev, b));
2279 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2280 bio->bi_bdev = rdev->bdev;
2281 bio->bi_end_io = raid1_end_read_request;
2282 bio->bi_rw = READ | do_sync;
2283 bio->bi_private = r1_bio;
2284 if (max_sectors < r1_bio->sectors) {
2285 /* Drat - have to split this up more */
2286 struct bio *mbio = r1_bio->master_bio;
2287 int sectors_handled = (r1_bio->sector + max_sectors
2289 r1_bio->sectors = max_sectors;
2290 spin_lock_irq(&conf->device_lock);
2291 if (mbio->bi_phys_segments == 0)
2292 mbio->bi_phys_segments = 2;
2294 mbio->bi_phys_segments++;
2295 spin_unlock_irq(&conf->device_lock);
2296 generic_make_request(bio);
2299 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2301 r1_bio->master_bio = mbio;
2302 r1_bio->sectors = (mbio->bi_size >> 9)
2305 set_bit(R1BIO_ReadError, &r1_bio->state);
2306 r1_bio->mddev = mddev;
2307 r1_bio->sector = mbio->bi_sector + sectors_handled;
2311 generic_make_request(bio);
2315 static void raid1d(struct md_thread *thread)
2317 struct mddev *mddev = thread->mddev;
2318 struct r1bio *r1_bio;
2319 unsigned long flags;
2320 struct r1conf *conf = mddev->private;
2321 struct list_head *head = &conf->retry_list;
2322 struct blk_plug plug;
2324 md_check_recovery(mddev);
2326 blk_start_plug(&plug);
2329 flush_pending_writes(conf);
2331 spin_lock_irqsave(&conf->device_lock, flags);
2332 if (list_empty(head)) {
2333 spin_unlock_irqrestore(&conf->device_lock, flags);
2336 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2337 list_del(head->prev);
2339 spin_unlock_irqrestore(&conf->device_lock, flags);
2341 mddev = r1_bio->mddev;
2342 conf = mddev->private;
2343 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2344 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2345 test_bit(R1BIO_WriteError, &r1_bio->state))
2346 handle_sync_write_finished(conf, r1_bio);
2348 sync_request_write(mddev, r1_bio);
2349 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2350 test_bit(R1BIO_WriteError, &r1_bio->state))
2351 handle_write_finished(conf, r1_bio);
2352 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2353 handle_read_error(conf, r1_bio);
2355 /* just a partial read to be scheduled from separate
2358 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2361 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2362 md_check_recovery(mddev);
2364 blk_finish_plug(&plug);
2368 static int init_resync(struct r1conf *conf)
2372 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2373 BUG_ON(conf->r1buf_pool);
2374 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2376 if (!conf->r1buf_pool)
2378 conf->next_resync = 0;
2383 * perform a "sync" on one "block"
2385 * We need to make sure that no normal I/O request - particularly write
2386 * requests - conflict with active sync requests.
2388 * This is achieved by tracking pending requests and a 'barrier' concept
2389 * that can be installed to exclude normal IO requests.
2392 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2394 struct r1conf *conf = mddev->private;
2395 struct r1bio *r1_bio;
2397 sector_t max_sector, nr_sectors;
2401 int write_targets = 0, read_targets = 0;
2402 sector_t sync_blocks;
2403 int still_degraded = 0;
2404 int good_sectors = RESYNC_SECTORS;
2405 int min_bad = 0; /* number of sectors that are bad in all devices */
2407 if (!conf->r1buf_pool)
2408 if (init_resync(conf))
2411 max_sector = mddev->dev_sectors;
2412 if (sector_nr >= max_sector) {
2413 /* If we aborted, we need to abort the
2414 * sync on the 'current' bitmap chunk (there will
2415 * only be one in raid1 resync.
2416 * We can find the current addess in mddev->curr_resync
2418 if (mddev->curr_resync < max_sector) /* aborted */
2419 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2421 else /* completed sync */
2424 bitmap_close_sync(mddev->bitmap);
2429 if (mddev->bitmap == NULL &&
2430 mddev->recovery_cp == MaxSector &&
2431 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2432 conf->fullsync == 0) {
2434 return max_sector - sector_nr;
2436 /* before building a request, check if we can skip these blocks..
2437 * This call the bitmap_start_sync doesn't actually record anything
2439 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2440 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2441 /* We can skip this block, and probably several more */
2446 * If there is non-resync activity waiting for a turn,
2447 * and resync is going fast enough,
2448 * then let it though before starting on this new sync request.
2450 if (!go_faster && conf->nr_waiting)
2451 msleep_interruptible(1000);
2453 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2454 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2455 raise_barrier(conf);
2457 conf->next_resync = sector_nr;
2461 * If we get a correctably read error during resync or recovery,
2462 * we might want to read from a different device. So we
2463 * flag all drives that could conceivably be read from for READ,
2464 * and any others (which will be non-In_sync devices) for WRITE.
2465 * If a read fails, we try reading from something else for which READ
2469 r1_bio->mddev = mddev;
2470 r1_bio->sector = sector_nr;
2472 set_bit(R1BIO_IsSync, &r1_bio->state);
2474 for (i = 0; i < conf->raid_disks * 2; i++) {
2475 struct md_rdev *rdev;
2476 bio = r1_bio->bios[i];
2478 /* take from bio_init */
2479 bio->bi_next = NULL;
2480 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2481 bio->bi_flags |= 1 << BIO_UPTODATE;
2485 bio->bi_phys_segments = 0;
2487 bio->bi_end_io = NULL;
2488 bio->bi_private = NULL;
2490 rdev = rcu_dereference(conf->mirrors[i].rdev);
2492 test_bit(Faulty, &rdev->flags)) {
2493 if (i < conf->raid_disks)
2495 } else if (!test_bit(In_sync, &rdev->flags)) {
2497 bio->bi_end_io = end_sync_write;
2500 /* may need to read from here */
2501 sector_t first_bad = MaxSector;
2504 if (is_badblock(rdev, sector_nr, good_sectors,
2505 &first_bad, &bad_sectors)) {
2506 if (first_bad > sector_nr)
2507 good_sectors = first_bad - sector_nr;
2509 bad_sectors -= (sector_nr - first_bad);
2511 min_bad > bad_sectors)
2512 min_bad = bad_sectors;
2515 if (sector_nr < first_bad) {
2516 if (test_bit(WriteMostly, &rdev->flags)) {
2524 bio->bi_end_io = end_sync_read;
2526 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2527 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2528 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2530 * The device is suitable for reading (InSync),
2531 * but has bad block(s) here. Let's try to correct them,
2532 * if we are doing resync or repair. Otherwise, leave
2533 * this device alone for this sync request.
2536 bio->bi_end_io = end_sync_write;
2540 if (bio->bi_end_io) {
2541 atomic_inc(&rdev->nr_pending);
2542 bio->bi_sector = sector_nr + rdev->data_offset;
2543 bio->bi_bdev = rdev->bdev;
2544 bio->bi_private = r1_bio;
2550 r1_bio->read_disk = disk;
2552 if (read_targets == 0 && min_bad > 0) {
2553 /* These sectors are bad on all InSync devices, so we
2554 * need to mark them bad on all write targets
2557 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2558 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2559 struct md_rdev *rdev = conf->mirrors[i].rdev;
2560 ok = rdev_set_badblocks(rdev, sector_nr,
2564 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2569 /* Cannot record the badblocks, so need to
2571 * If there are multiple read targets, could just
2572 * fail the really bad ones ???
2574 conf->recovery_disabled = mddev->recovery_disabled;
2575 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2581 if (min_bad > 0 && min_bad < good_sectors) {
2582 /* only resync enough to reach the next bad->good
2584 good_sectors = min_bad;
2587 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2588 /* extra read targets are also write targets */
2589 write_targets += read_targets-1;
2591 if (write_targets == 0 || read_targets == 0) {
2592 /* There is nowhere to write, so all non-sync
2593 * drives must be failed - so we are finished
2597 max_sector = sector_nr + min_bad;
2598 rv = max_sector - sector_nr;
2604 if (max_sector > mddev->resync_max)
2605 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2606 if (max_sector > sector_nr + good_sectors)
2607 max_sector = sector_nr + good_sectors;
2612 int len = PAGE_SIZE;
2613 if (sector_nr + (len>>9) > max_sector)
2614 len = (max_sector - sector_nr) << 9;
2617 if (sync_blocks == 0) {
2618 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2619 &sync_blocks, still_degraded) &&
2621 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2623 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2624 if ((len >> 9) > sync_blocks)
2625 len = sync_blocks<<9;
2628 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2629 bio = r1_bio->bios[i];
2630 if (bio->bi_end_io) {
2631 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2632 if (bio_add_page(bio, page, len, 0) == 0) {
2634 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2637 bio = r1_bio->bios[i];
2638 if (bio->bi_end_io==NULL)
2640 /* remove last page from this bio */
2642 bio->bi_size -= len;
2643 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2649 nr_sectors += len>>9;
2650 sector_nr += len>>9;
2651 sync_blocks -= (len>>9);
2652 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2654 r1_bio->sectors = nr_sectors;
2656 /* For a user-requested sync, we read all readable devices and do a
2659 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2660 atomic_set(&r1_bio->remaining, read_targets);
2661 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2662 bio = r1_bio->bios[i];
2663 if (bio->bi_end_io == end_sync_read) {
2665 md_sync_acct(bio->bi_bdev, nr_sectors);
2666 generic_make_request(bio);
2670 atomic_set(&r1_bio->remaining, 1);
2671 bio = r1_bio->bios[r1_bio->read_disk];
2672 md_sync_acct(bio->bi_bdev, nr_sectors);
2673 generic_make_request(bio);
2679 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2684 return mddev->dev_sectors;
2687 static struct r1conf *setup_conf(struct mddev *mddev)
2689 struct r1conf *conf;
2691 struct raid1_info *disk;
2692 struct md_rdev *rdev;
2695 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2699 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2700 * mddev->raid_disks * 2,
2705 conf->tmppage = alloc_page(GFP_KERNEL);
2709 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2710 if (!conf->poolinfo)
2712 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2713 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2716 if (!conf->r1bio_pool)
2719 conf->poolinfo->mddev = mddev;
2722 spin_lock_init(&conf->device_lock);
2723 rdev_for_each(rdev, mddev) {
2724 struct request_queue *q;
2725 int disk_idx = rdev->raid_disk;
2726 if (disk_idx >= mddev->raid_disks
2729 if (test_bit(Replacement, &rdev->flags))
2730 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2732 disk = conf->mirrors + disk_idx;
2737 q = bdev_get_queue(rdev->bdev);
2738 if (q->merge_bvec_fn)
2739 mddev->merge_check_needed = 1;
2741 disk->head_position = 0;
2742 disk->seq_start = MaxSector;
2744 conf->raid_disks = mddev->raid_disks;
2745 conf->mddev = mddev;
2746 INIT_LIST_HEAD(&conf->retry_list);
2748 spin_lock_init(&conf->resync_lock);
2749 init_waitqueue_head(&conf->wait_barrier);
2751 bio_list_init(&conf->pending_bio_list);
2752 conf->pending_count = 0;
2753 conf->recovery_disabled = mddev->recovery_disabled - 1;
2756 for (i = 0; i < conf->raid_disks * 2; i++) {
2758 disk = conf->mirrors + i;
2760 if (i < conf->raid_disks &&
2761 disk[conf->raid_disks].rdev) {
2762 /* This slot has a replacement. */
2764 /* No original, just make the replacement
2765 * a recovering spare
2768 disk[conf->raid_disks].rdev;
2769 disk[conf->raid_disks].rdev = NULL;
2770 } else if (!test_bit(In_sync, &disk->rdev->flags))
2771 /* Original is not in_sync - bad */
2776 !test_bit(In_sync, &disk->rdev->flags)) {
2777 disk->head_position = 0;
2779 (disk->rdev->saved_raid_disk < 0))
2785 conf->thread = md_register_thread(raid1d, mddev, "raid1");
2786 if (!conf->thread) {
2788 "md/raid1:%s: couldn't allocate thread\n",
2797 if (conf->r1bio_pool)
2798 mempool_destroy(conf->r1bio_pool);
2799 kfree(conf->mirrors);
2800 safe_put_page(conf->tmppage);
2801 kfree(conf->poolinfo);
2804 return ERR_PTR(err);
2807 static int stop(struct mddev *mddev);
2808 static int run(struct mddev *mddev)
2810 struct r1conf *conf;
2812 struct md_rdev *rdev;
2814 bool discard_supported = false;
2816 if (mddev->level != 1) {
2817 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2818 mdname(mddev), mddev->level);
2821 if (mddev->reshape_position != MaxSector) {
2822 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2827 * copy the already verified devices into our private RAID1
2828 * bookkeeping area. [whatever we allocate in run(),
2829 * should be freed in stop()]
2831 if (mddev->private == NULL)
2832 conf = setup_conf(mddev);
2834 conf = mddev->private;
2837 return PTR_ERR(conf);
2840 blk_queue_max_write_same_sectors(mddev->queue,
2841 mddev->chunk_sectors);
2842 rdev_for_each(rdev, mddev) {
2843 if (!mddev->gendisk)
2845 disk_stack_limits(mddev->gendisk, rdev->bdev,
2846 rdev->data_offset << 9);
2847 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2848 discard_supported = true;
2851 mddev->degraded = 0;
2852 for (i=0; i < conf->raid_disks; i++)
2853 if (conf->mirrors[i].rdev == NULL ||
2854 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2855 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2858 if (conf->raid_disks - mddev->degraded == 1)
2859 mddev->recovery_cp = MaxSector;
2861 if (mddev->recovery_cp != MaxSector)
2862 printk(KERN_NOTICE "md/raid1:%s: not clean"
2863 " -- starting background reconstruction\n",
2866 "md/raid1:%s: active with %d out of %d mirrors\n",
2867 mdname(mddev), mddev->raid_disks - mddev->degraded,
2871 * Ok, everything is just fine now
2873 mddev->thread = conf->thread;
2874 conf->thread = NULL;
2875 mddev->private = conf;
2877 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2880 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2881 mddev->queue->backing_dev_info.congested_data = mddev;
2882 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2884 if (discard_supported)
2885 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2888 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2892 ret = md_integrity_register(mddev);
2898 static int stop(struct mddev *mddev)
2900 struct r1conf *conf = mddev->private;
2901 struct bitmap *bitmap = mddev->bitmap;
2903 /* wait for behind writes to complete */
2904 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2905 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2907 /* need to kick something here to make sure I/O goes? */
2908 wait_event(bitmap->behind_wait,
2909 atomic_read(&bitmap->behind_writes) == 0);
2912 raise_barrier(conf);
2913 lower_barrier(conf);
2915 md_unregister_thread(&mddev->thread);
2916 if (conf->r1bio_pool)
2917 mempool_destroy(conf->r1bio_pool);
2918 kfree(conf->mirrors);
2919 safe_put_page(conf->tmppage);
2920 kfree(conf->poolinfo);
2922 mddev->private = NULL;
2926 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2928 /* no resync is happening, and there is enough space
2929 * on all devices, so we can resize.
2930 * We need to make sure resync covers any new space.
2931 * If the array is shrinking we should possibly wait until
2932 * any io in the removed space completes, but it hardly seems
2935 sector_t newsize = raid1_size(mddev, sectors, 0);
2936 if (mddev->external_size &&
2937 mddev->array_sectors > newsize)
2939 if (mddev->bitmap) {
2940 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2944 md_set_array_sectors(mddev, newsize);
2945 set_capacity(mddev->gendisk, mddev->array_sectors);
2946 revalidate_disk(mddev->gendisk);
2947 if (sectors > mddev->dev_sectors &&
2948 mddev->recovery_cp > mddev->dev_sectors) {
2949 mddev->recovery_cp = mddev->dev_sectors;
2950 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2952 mddev->dev_sectors = sectors;
2953 mddev->resync_max_sectors = sectors;
2957 static int raid1_reshape(struct mddev *mddev)
2960 * 1/ resize the r1bio_pool
2961 * 2/ resize conf->mirrors
2963 * We allocate a new r1bio_pool if we can.
2964 * Then raise a device barrier and wait until all IO stops.
2965 * Then resize conf->mirrors and swap in the new r1bio pool.
2967 * At the same time, we "pack" the devices so that all the missing
2968 * devices have the higher raid_disk numbers.
2970 mempool_t *newpool, *oldpool;
2971 struct pool_info *newpoolinfo;
2972 struct raid1_info *newmirrors;
2973 struct r1conf *conf = mddev->private;
2974 int cnt, raid_disks;
2975 unsigned long flags;
2978 /* Cannot change chunk_size, layout, or level */
2979 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2980 mddev->layout != mddev->new_layout ||
2981 mddev->level != mddev->new_level) {
2982 mddev->new_chunk_sectors = mddev->chunk_sectors;
2983 mddev->new_layout = mddev->layout;
2984 mddev->new_level = mddev->level;
2988 err = md_allow_write(mddev);
2992 raid_disks = mddev->raid_disks + mddev->delta_disks;
2994 if (raid_disks < conf->raid_disks) {
2996 for (d= 0; d < conf->raid_disks; d++)
2997 if (conf->mirrors[d].rdev)
2999 if (cnt > raid_disks)
3003 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3006 newpoolinfo->mddev = mddev;
3007 newpoolinfo->raid_disks = raid_disks * 2;
3009 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3010 r1bio_pool_free, newpoolinfo);
3015 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3019 mempool_destroy(newpool);
3023 freeze_array(conf, 0);
3025 /* ok, everything is stopped */
3026 oldpool = conf->r1bio_pool;
3027 conf->r1bio_pool = newpool;
3029 for (d = d2 = 0; d < conf->raid_disks; d++) {
3030 struct md_rdev *rdev = conf->mirrors[d].rdev;
3031 if (rdev && rdev->raid_disk != d2) {
3032 sysfs_unlink_rdev(mddev, rdev);
3033 rdev->raid_disk = d2;
3034 sysfs_unlink_rdev(mddev, rdev);
3035 if (sysfs_link_rdev(mddev, rdev))
3037 "md/raid1:%s: cannot register rd%d\n",
3038 mdname(mddev), rdev->raid_disk);
3041 newmirrors[d2++].rdev = rdev;
3043 kfree(conf->mirrors);
3044 conf->mirrors = newmirrors;
3045 kfree(conf->poolinfo);
3046 conf->poolinfo = newpoolinfo;
3048 spin_lock_irqsave(&conf->device_lock, flags);
3049 mddev->degraded += (raid_disks - conf->raid_disks);
3050 spin_unlock_irqrestore(&conf->device_lock, flags);
3051 conf->raid_disks = mddev->raid_disks = raid_disks;
3052 mddev->delta_disks = 0;
3054 unfreeze_array(conf);
3056 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3057 md_wakeup_thread(mddev->thread);
3059 mempool_destroy(oldpool);
3063 static void raid1_quiesce(struct mddev *mddev, int state)
3065 struct r1conf *conf = mddev->private;
3068 case 2: /* wake for suspend */
3069 wake_up(&conf->wait_barrier);
3072 raise_barrier(conf);
3075 lower_barrier(conf);
3080 static void *raid1_takeover(struct mddev *mddev)
3082 /* raid1 can take over:
3083 * raid5 with 2 devices, any layout or chunk size
3085 if (mddev->level == 5 && mddev->raid_disks == 2) {
3086 struct r1conf *conf;
3087 mddev->new_level = 1;
3088 mddev->new_layout = 0;
3089 mddev->new_chunk_sectors = 0;
3090 conf = setup_conf(mddev);
3095 return ERR_PTR(-EINVAL);
3098 static struct md_personality raid1_personality =
3102 .owner = THIS_MODULE,
3103 .make_request = make_request,
3107 .error_handler = error,
3108 .hot_add_disk = raid1_add_disk,
3109 .hot_remove_disk= raid1_remove_disk,
3110 .spare_active = raid1_spare_active,
3111 .sync_request = sync_request,
3112 .resize = raid1_resize,
3114 .check_reshape = raid1_reshape,
3115 .quiesce = raid1_quiesce,
3116 .takeover = raid1_takeover,
3119 static int __init raid_init(void)
3121 return register_md_personality(&raid1_personality);
3124 static void raid_exit(void)
3126 unregister_md_personality(&raid1_personality);
3129 module_init(raid_init);
3130 module_exit(raid_exit);
3131 MODULE_LICENSE("GPL");
3132 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3133 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3134 MODULE_ALIAS("md-raid1");
3135 MODULE_ALIAS("md-level-1");
3137 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);