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, sector_t start_next_window,
71 static void lower_barrier(struct r1conf *conf);
73 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
75 struct pool_info *pi = data;
76 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
78 /* allocate a r1bio with room for raid_disks entries in the bios array */
79 return kzalloc(size, gfp_flags);
82 static void r1bio_pool_free(void *r1_bio, void *data)
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_DEPTH 32
89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
93 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
95 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
97 struct pool_info *pi = data;
103 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
108 * Allocate bios : 1 for reading, n-1 for writing
110 for (j = pi->raid_disks ; j-- ; ) {
111 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
114 r1_bio->bios[j] = bio;
117 * Allocate RESYNC_PAGES data pages and attach them to
119 * If this is a user-requested check/repair, allocate
120 * RESYNC_PAGES for each bio.
122 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
123 need_pages = pi->raid_disks;
126 for (j = 0; j < need_pages; j++) {
127 bio = r1_bio->bios[j];
128 bio->bi_vcnt = RESYNC_PAGES;
130 if (bio_alloc_pages(bio, gfp_flags))
133 /* If not user-requests, copy the page pointers to all bios */
134 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
135 for (i=0; i<RESYNC_PAGES ; i++)
136 for (j=1; j<pi->raid_disks; j++)
137 r1_bio->bios[j]->bi_io_vec[i].bv_page =
138 r1_bio->bios[0]->bi_io_vec[i].bv_page;
141 r1_bio->master_bio = NULL;
149 bio_for_each_segment_all(bv, r1_bio->bios[j], i)
150 __free_page(bv->bv_page);
154 while (++j < pi->raid_disks)
155 bio_put(r1_bio->bios[j]);
156 r1bio_pool_free(r1_bio, data);
160 static void r1buf_pool_free(void *__r1_bio, void *data)
162 struct pool_info *pi = data;
164 struct r1bio *r1bio = __r1_bio;
166 for (i = 0; i < RESYNC_PAGES; i++)
167 for (j = pi->raid_disks; j-- ;) {
169 r1bio->bios[j]->bi_io_vec[i].bv_page !=
170 r1bio->bios[0]->bi_io_vec[i].bv_page)
171 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
173 for (i=0 ; i < pi->raid_disks; i++)
174 bio_put(r1bio->bios[i]);
176 r1bio_pool_free(r1bio, data);
179 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
183 for (i = 0; i < conf->raid_disks * 2; i++) {
184 struct bio **bio = r1_bio->bios + i;
185 if (!BIO_SPECIAL(*bio))
191 static void free_r1bio(struct r1bio *r1_bio)
193 struct r1conf *conf = r1_bio->mddev->private;
195 put_all_bios(conf, r1_bio);
196 mempool_free(r1_bio, conf->r1bio_pool);
199 static void put_buf(struct r1bio *r1_bio)
201 struct r1conf *conf = r1_bio->mddev->private;
204 for (i = 0; i < conf->raid_disks * 2; i++) {
205 struct bio *bio = r1_bio->bios[i];
207 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
210 mempool_free(r1_bio, conf->r1buf_pool);
215 static void reschedule_retry(struct r1bio *r1_bio)
218 struct mddev *mddev = r1_bio->mddev;
219 struct r1conf *conf = mddev->private;
221 spin_lock_irqsave(&conf->device_lock, flags);
222 list_add(&r1_bio->retry_list, &conf->retry_list);
224 spin_unlock_irqrestore(&conf->device_lock, flags);
226 wake_up(&conf->wait_barrier);
227 md_wakeup_thread(mddev->thread);
231 * raid_end_bio_io() is called when we have finished servicing a mirrored
232 * operation and are ready to return a success/failure code to the buffer
235 static void call_bio_endio(struct r1bio *r1_bio)
237 struct bio *bio = r1_bio->master_bio;
239 struct r1conf *conf = r1_bio->mddev->private;
240 sector_t start_next_window = r1_bio->start_next_window;
241 sector_t bi_sector = bio->bi_iter.bi_sector;
243 if (bio->bi_phys_segments) {
245 spin_lock_irqsave(&conf->device_lock, flags);
246 bio->bi_phys_segments--;
247 done = (bio->bi_phys_segments == 0);
248 spin_unlock_irqrestore(&conf->device_lock, flags);
250 * make_request() might be waiting for
251 * bi_phys_segments to decrease
253 wake_up(&conf->wait_barrier);
257 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
258 clear_bit(BIO_UPTODATE, &bio->bi_flags);
262 * Wake up any possible resync thread that waits for the device
265 allow_barrier(conf, start_next_window, bi_sector);
269 static void raid_end_bio_io(struct r1bio *r1_bio)
271 struct bio *bio = r1_bio->master_bio;
273 /* if nobody has done the final endio yet, do it now */
274 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
275 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
276 (bio_data_dir(bio) == WRITE) ? "write" : "read",
277 (unsigned long long) bio->bi_iter.bi_sector,
278 (unsigned long long) bio_end_sector(bio) - 1);
280 call_bio_endio(r1_bio);
286 * Update disk head position estimator based on IRQ completion info.
288 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
290 struct r1conf *conf = r1_bio->mddev->private;
292 conf->mirrors[disk].head_position =
293 r1_bio->sector + (r1_bio->sectors);
297 * Find the disk number which triggered given bio
299 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
302 struct r1conf *conf = r1_bio->mddev->private;
303 int raid_disks = conf->raid_disks;
305 for (mirror = 0; mirror < raid_disks * 2; mirror++)
306 if (r1_bio->bios[mirror] == bio)
309 BUG_ON(mirror == raid_disks * 2);
310 update_head_pos(mirror, r1_bio);
315 static void raid1_end_read_request(struct bio *bio, int error)
317 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
318 struct r1bio *r1_bio = bio->bi_private;
320 struct r1conf *conf = r1_bio->mddev->private;
322 mirror = r1_bio->read_disk;
324 * this branch is our 'one mirror IO has finished' event handler:
326 update_head_pos(mirror, r1_bio);
329 set_bit(R1BIO_Uptodate, &r1_bio->state);
331 /* If all other devices have failed, we want to return
332 * the error upwards rather than fail the last device.
333 * Here we redefine "uptodate" to mean "Don't want to retry"
336 spin_lock_irqsave(&conf->device_lock, flags);
337 if (r1_bio->mddev->degraded == conf->raid_disks ||
338 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
339 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
341 spin_unlock_irqrestore(&conf->device_lock, flags);
345 raid_end_bio_io(r1_bio);
346 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
351 char b[BDEVNAME_SIZE];
353 KERN_ERR "md/raid1:%s: %s: "
354 "rescheduling sector %llu\n",
356 bdevname(conf->mirrors[mirror].rdev->bdev,
358 (unsigned long long)r1_bio->sector);
359 set_bit(R1BIO_ReadError, &r1_bio->state);
360 reschedule_retry(r1_bio);
361 /* don't drop the reference on read_disk yet */
365 static void close_write(struct r1bio *r1_bio)
367 /* it really is the end of this request */
368 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
369 /* free extra copy of the data pages */
370 int i = r1_bio->behind_page_count;
372 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
373 kfree(r1_bio->behind_bvecs);
374 r1_bio->behind_bvecs = NULL;
376 /* clear the bitmap if all writes complete successfully */
377 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
379 !test_bit(R1BIO_Degraded, &r1_bio->state),
380 test_bit(R1BIO_BehindIO, &r1_bio->state));
381 md_write_end(r1_bio->mddev);
384 static void r1_bio_write_done(struct r1bio *r1_bio)
386 if (!atomic_dec_and_test(&r1_bio->remaining))
389 if (test_bit(R1BIO_WriteError, &r1_bio->state))
390 reschedule_retry(r1_bio);
393 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
394 reschedule_retry(r1_bio);
396 raid_end_bio_io(r1_bio);
400 static void raid1_end_write_request(struct bio *bio, int error)
402 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
403 struct r1bio *r1_bio = bio->bi_private;
404 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
405 struct r1conf *conf = r1_bio->mddev->private;
406 struct bio *to_put = NULL;
408 mirror = find_bio_disk(r1_bio, bio);
411 * 'one mirror IO has finished' event handler:
414 set_bit(WriteErrorSeen,
415 &conf->mirrors[mirror].rdev->flags);
416 if (!test_and_set_bit(WantReplacement,
417 &conf->mirrors[mirror].rdev->flags))
418 set_bit(MD_RECOVERY_NEEDED, &
419 conf->mddev->recovery);
421 set_bit(R1BIO_WriteError, &r1_bio->state);
424 * Set R1BIO_Uptodate in our master bio, so that we
425 * will return a good error code for to the higher
426 * levels even if IO on some other mirrored buffer
429 * The 'master' represents the composite IO operation
430 * to user-side. So if something waits for IO, then it
431 * will wait for the 'master' bio.
436 r1_bio->bios[mirror] = NULL;
439 * Do not set R1BIO_Uptodate if the current device is
440 * rebuilding or Faulty. This is because we cannot use
441 * such device for properly reading the data back (we could
442 * potentially use it, if the current write would have felt
443 * before rdev->recovery_offset, but for simplicity we don't
446 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
447 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
448 set_bit(R1BIO_Uptodate, &r1_bio->state);
450 /* Maybe we can clear some bad blocks. */
451 if (is_badblock(conf->mirrors[mirror].rdev,
452 r1_bio->sector, r1_bio->sectors,
453 &first_bad, &bad_sectors)) {
454 r1_bio->bios[mirror] = IO_MADE_GOOD;
455 set_bit(R1BIO_MadeGood, &r1_bio->state);
460 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
461 atomic_dec(&r1_bio->behind_remaining);
464 * In behind mode, we ACK the master bio once the I/O
465 * has safely reached all non-writemostly
466 * disks. Setting the Returned bit ensures that this
467 * gets done only once -- we don't ever want to return
468 * -EIO here, instead we'll wait
470 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
471 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
472 /* Maybe we can return now */
473 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
474 struct bio *mbio = r1_bio->master_bio;
475 pr_debug("raid1: behind end write sectors"
477 (unsigned long long) mbio->bi_iter.bi_sector,
478 (unsigned long long) bio_end_sector(mbio) - 1);
479 call_bio_endio(r1_bio);
483 if (r1_bio->bios[mirror] == NULL)
484 rdev_dec_pending(conf->mirrors[mirror].rdev,
488 * Let's see if all mirrored write operations have finished
491 r1_bio_write_done(r1_bio);
498 * This routine returns the disk from which the requested read should
499 * be done. There is a per-array 'next expected sequential IO' sector
500 * number - if this matches on the next IO then we use the last disk.
501 * There is also a per-disk 'last know head position' sector that is
502 * maintained from IRQ contexts, both the normal and the resync IO
503 * completion handlers update this position correctly. If there is no
504 * perfect sequential match then we pick the disk whose head is closest.
506 * If there are 2 mirrors in the same 2 devices, performance degrades
507 * because position is mirror, not device based.
509 * The rdev for the device selected will have nr_pending incremented.
511 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
513 const sector_t this_sector = r1_bio->sector;
515 int best_good_sectors;
516 int best_disk, best_dist_disk, best_pending_disk;
520 unsigned int min_pending;
521 struct md_rdev *rdev;
523 int choose_next_idle;
527 * Check if we can balance. We can balance on the whole
528 * device if no resync is going on, or below the resync window.
529 * We take the first readable disk when above the resync window.
532 sectors = r1_bio->sectors;
535 best_dist = MaxSector;
536 best_pending_disk = -1;
537 min_pending = UINT_MAX;
538 best_good_sectors = 0;
540 choose_next_idle = 0;
542 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
543 (mddev_is_clustered(conf->mddev) &&
544 md_cluster_ops->area_resyncing(conf->mddev, this_sector,
545 this_sector + sectors)))
550 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
554 unsigned int pending;
557 rdev = rcu_dereference(conf->mirrors[disk].rdev);
558 if (r1_bio->bios[disk] == IO_BLOCKED
560 || test_bit(Unmerged, &rdev->flags)
561 || test_bit(Faulty, &rdev->flags))
563 if (!test_bit(In_sync, &rdev->flags) &&
564 rdev->recovery_offset < this_sector + sectors)
566 if (test_bit(WriteMostly, &rdev->flags)) {
567 /* Don't balance among write-mostly, just
568 * use the first as a last resort */
570 if (is_badblock(rdev, this_sector, sectors,
571 &first_bad, &bad_sectors)) {
572 if (first_bad < this_sector)
573 /* Cannot use this */
575 best_good_sectors = first_bad - this_sector;
577 best_good_sectors = sectors;
582 /* This is a reasonable device to use. It might
585 if (is_badblock(rdev, this_sector, sectors,
586 &first_bad, &bad_sectors)) {
587 if (best_dist < MaxSector)
588 /* already have a better device */
590 if (first_bad <= this_sector) {
591 /* cannot read here. If this is the 'primary'
592 * device, then we must not read beyond
593 * bad_sectors from another device..
595 bad_sectors -= (this_sector - first_bad);
596 if (choose_first && sectors > bad_sectors)
597 sectors = bad_sectors;
598 if (best_good_sectors > sectors)
599 best_good_sectors = sectors;
602 sector_t good_sectors = first_bad - this_sector;
603 if (good_sectors > best_good_sectors) {
604 best_good_sectors = good_sectors;
612 best_good_sectors = sectors;
614 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
615 has_nonrot_disk |= nonrot;
616 pending = atomic_read(&rdev->nr_pending);
617 dist = abs(this_sector - conf->mirrors[disk].head_position);
622 /* Don't change to another disk for sequential reads */
623 if (conf->mirrors[disk].next_seq_sect == this_sector
625 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
626 struct raid1_info *mirror = &conf->mirrors[disk];
630 * If buffered sequential IO size exceeds optimal
631 * iosize, check if there is idle disk. If yes, choose
632 * the idle disk. read_balance could already choose an
633 * idle disk before noticing it's a sequential IO in
634 * this disk. This doesn't matter because this disk
635 * will idle, next time it will be utilized after the
636 * first disk has IO size exceeds optimal iosize. In
637 * this way, iosize of the first disk will be optimal
638 * iosize at least. iosize of the second disk might be
639 * small, but not a big deal since when the second disk
640 * starts IO, the first disk is likely still busy.
642 if (nonrot && opt_iosize > 0 &&
643 mirror->seq_start != MaxSector &&
644 mirror->next_seq_sect > opt_iosize &&
645 mirror->next_seq_sect - opt_iosize >=
647 choose_next_idle = 1;
652 /* If device is idle, use it */
658 if (choose_next_idle)
661 if (min_pending > pending) {
662 min_pending = pending;
663 best_pending_disk = disk;
666 if (dist < best_dist) {
668 best_dist_disk = disk;
673 * If all disks are rotational, choose the closest disk. If any disk is
674 * non-rotational, choose the disk with less pending request even the
675 * disk is rotational, which might/might not be optimal for raids with
676 * mixed ratation/non-rotational disks depending on workload.
678 if (best_disk == -1) {
680 best_disk = best_pending_disk;
682 best_disk = best_dist_disk;
685 if (best_disk >= 0) {
686 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
689 atomic_inc(&rdev->nr_pending);
690 if (test_bit(Faulty, &rdev->flags)) {
691 /* cannot risk returning a device that failed
692 * before we inc'ed nr_pending
694 rdev_dec_pending(rdev, conf->mddev);
697 sectors = best_good_sectors;
699 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
700 conf->mirrors[best_disk].seq_start = this_sector;
702 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
705 *max_sectors = sectors;
710 static int raid1_mergeable_bvec(struct mddev *mddev,
711 struct bvec_merge_data *bvm,
712 struct bio_vec *biovec)
714 struct r1conf *conf = mddev->private;
715 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
716 int max = biovec->bv_len;
718 if (mddev->merge_check_needed) {
721 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
722 struct md_rdev *rdev = rcu_dereference(
723 conf->mirrors[disk].rdev);
724 if (rdev && !test_bit(Faulty, &rdev->flags)) {
725 struct request_queue *q =
726 bdev_get_queue(rdev->bdev);
727 if (q->merge_bvec_fn) {
728 bvm->bi_sector = sector +
730 bvm->bi_bdev = rdev->bdev;
731 max = min(max, q->merge_bvec_fn(
742 static int raid1_congested(struct mddev *mddev, int bits)
744 struct r1conf *conf = mddev->private;
747 if ((bits & (1 << BDI_async_congested)) &&
748 conf->pending_count >= max_queued_requests)
752 for (i = 0; i < conf->raid_disks * 2; i++) {
753 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
754 if (rdev && !test_bit(Faulty, &rdev->flags)) {
755 struct request_queue *q = bdev_get_queue(rdev->bdev);
759 /* Note the '|| 1' - when read_balance prefers
760 * non-congested targets, it can be removed
762 if ((bits & (1<<BDI_async_congested)) || 1)
763 ret |= bdi_congested(&q->backing_dev_info, bits);
765 ret &= bdi_congested(&q->backing_dev_info, bits);
772 static void flush_pending_writes(struct r1conf *conf)
774 /* Any writes that have been queued but are awaiting
775 * bitmap updates get flushed here.
777 spin_lock_irq(&conf->device_lock);
779 if (conf->pending_bio_list.head) {
781 bio = bio_list_get(&conf->pending_bio_list);
782 conf->pending_count = 0;
783 spin_unlock_irq(&conf->device_lock);
784 /* flush any pending bitmap writes to
785 * disk before proceeding w/ I/O */
786 bitmap_unplug(conf->mddev->bitmap);
787 wake_up(&conf->wait_barrier);
789 while (bio) { /* submit pending writes */
790 struct bio *next = bio->bi_next;
792 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
793 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
797 generic_make_request(bio);
801 spin_unlock_irq(&conf->device_lock);
805 * Sometimes we need to suspend IO while we do something else,
806 * either some resync/recovery, or reconfigure the array.
807 * To do this we raise a 'barrier'.
808 * The 'barrier' is a counter that can be raised multiple times
809 * to count how many activities are happening which preclude
811 * We can only raise the barrier if there is no pending IO.
812 * i.e. if nr_pending == 0.
813 * We choose only to raise the barrier if no-one is waiting for the
814 * barrier to go down. This means that as soon as an IO request
815 * is ready, no other operations which require a barrier will start
816 * until the IO request has had a chance.
818 * So: regular IO calls 'wait_barrier'. When that returns there
819 * is no backgroup IO happening, It must arrange to call
820 * allow_barrier when it has finished its IO.
821 * backgroup IO calls must call raise_barrier. Once that returns
822 * there is no normal IO happeing. It must arrange to call
823 * lower_barrier when the particular background IO completes.
825 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
827 spin_lock_irq(&conf->resync_lock);
829 /* Wait until no block IO is waiting */
830 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
833 /* block any new IO from starting */
835 conf->next_resync = sector_nr;
837 /* For these conditions we must wait:
838 * A: while the array is in frozen state
839 * B: while barrier >= RESYNC_DEPTH, meaning resync reach
840 * the max count which allowed.
841 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
842 * next resync will reach to the window which normal bios are
844 * D: while there are any active requests in the current window.
846 wait_event_lock_irq(conf->wait_barrier,
847 !conf->array_frozen &&
848 conf->barrier < RESYNC_DEPTH &&
849 conf->current_window_requests == 0 &&
850 (conf->start_next_window >=
851 conf->next_resync + RESYNC_SECTORS),
855 spin_unlock_irq(&conf->resync_lock);
858 static void lower_barrier(struct r1conf *conf)
861 BUG_ON(conf->barrier <= 0);
862 spin_lock_irqsave(&conf->resync_lock, flags);
865 spin_unlock_irqrestore(&conf->resync_lock, flags);
866 wake_up(&conf->wait_barrier);
869 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
873 if (conf->array_frozen || !bio)
875 else if (conf->barrier && bio_data_dir(bio) == WRITE) {
876 if ((conf->mddev->curr_resync_completed
877 >= bio_end_sector(bio)) ||
878 (conf->next_resync + NEXT_NORMALIO_DISTANCE
879 <= bio->bi_iter.bi_sector))
888 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
892 spin_lock_irq(&conf->resync_lock);
893 if (need_to_wait_for_sync(conf, bio)) {
895 /* Wait for the barrier to drop.
896 * However if there are already pending
897 * requests (preventing the barrier from
898 * rising completely), and the
899 * per-process bio queue isn't empty,
900 * then don't wait, as we need to empty
901 * that queue to allow conf->start_next_window
904 wait_event_lock_irq(conf->wait_barrier,
905 !conf->array_frozen &&
907 ((conf->start_next_window <
908 conf->next_resync + RESYNC_SECTORS) &&
910 !bio_list_empty(current->bio_list))),
915 if (bio && bio_data_dir(bio) == WRITE) {
916 if (bio->bi_iter.bi_sector >=
917 conf->mddev->curr_resync_completed) {
918 if (conf->start_next_window == MaxSector)
919 conf->start_next_window =
921 NEXT_NORMALIO_DISTANCE;
923 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
924 <= bio->bi_iter.bi_sector)
925 conf->next_window_requests++;
927 conf->current_window_requests++;
928 sector = conf->start_next_window;
933 spin_unlock_irq(&conf->resync_lock);
937 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
942 spin_lock_irqsave(&conf->resync_lock, flags);
944 if (start_next_window) {
945 if (start_next_window == conf->start_next_window) {
946 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
948 conf->next_window_requests--;
950 conf->current_window_requests--;
952 conf->current_window_requests--;
954 if (!conf->current_window_requests) {
955 if (conf->next_window_requests) {
956 conf->current_window_requests =
957 conf->next_window_requests;
958 conf->next_window_requests = 0;
959 conf->start_next_window +=
960 NEXT_NORMALIO_DISTANCE;
962 conf->start_next_window = MaxSector;
965 spin_unlock_irqrestore(&conf->resync_lock, flags);
966 wake_up(&conf->wait_barrier);
969 static void freeze_array(struct r1conf *conf, int extra)
971 /* stop syncio and normal IO and wait for everything to
973 * We wait until nr_pending match nr_queued+extra
974 * This is called in the context of one normal IO request
975 * that has failed. Thus any sync request that might be pending
976 * will be blocked by nr_pending, and we need to wait for
977 * pending IO requests to complete or be queued for re-try.
978 * Thus the number queued (nr_queued) plus this request (extra)
979 * must match the number of pending IOs (nr_pending) before
982 spin_lock_irq(&conf->resync_lock);
983 conf->array_frozen = 1;
984 wait_event_lock_irq_cmd(conf->wait_barrier,
985 conf->nr_pending == conf->nr_queued+extra,
987 flush_pending_writes(conf));
988 spin_unlock_irq(&conf->resync_lock);
990 static void unfreeze_array(struct r1conf *conf)
992 /* reverse the effect of the freeze */
993 spin_lock_irq(&conf->resync_lock);
994 conf->array_frozen = 0;
995 wake_up(&conf->wait_barrier);
996 spin_unlock_irq(&conf->resync_lock);
999 /* duplicate the data pages for behind I/O
1001 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
1004 struct bio_vec *bvec;
1005 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
1007 if (unlikely(!bvecs))
1010 bio_for_each_segment_all(bvec, bio, i) {
1012 bvecs[i].bv_page = alloc_page(GFP_NOIO);
1013 if (unlikely(!bvecs[i].bv_page))
1015 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
1016 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
1017 kunmap(bvecs[i].bv_page);
1018 kunmap(bvec->bv_page);
1020 r1_bio->behind_bvecs = bvecs;
1021 r1_bio->behind_page_count = bio->bi_vcnt;
1022 set_bit(R1BIO_BehindIO, &r1_bio->state);
1026 for (i = 0; i < bio->bi_vcnt; i++)
1027 if (bvecs[i].bv_page)
1028 put_page(bvecs[i].bv_page);
1030 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1031 bio->bi_iter.bi_size);
1034 struct raid1_plug_cb {
1035 struct blk_plug_cb cb;
1036 struct bio_list pending;
1040 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1042 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1044 struct mddev *mddev = plug->cb.data;
1045 struct r1conf *conf = mddev->private;
1048 if (from_schedule || current->bio_list) {
1049 spin_lock_irq(&conf->device_lock);
1050 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1051 conf->pending_count += plug->pending_cnt;
1052 spin_unlock_irq(&conf->device_lock);
1053 wake_up(&conf->wait_barrier);
1054 md_wakeup_thread(mddev->thread);
1059 /* we aren't scheduling, so we can do the write-out directly. */
1060 bio = bio_list_get(&plug->pending);
1061 bitmap_unplug(mddev->bitmap);
1062 wake_up(&conf->wait_barrier);
1064 while (bio) { /* submit pending writes */
1065 struct bio *next = bio->bi_next;
1066 bio->bi_next = NULL;
1067 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1068 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1069 /* Just ignore it */
1072 generic_make_request(bio);
1078 static void make_request(struct mddev *mddev, struct bio * bio)
1080 struct r1conf *conf = mddev->private;
1081 struct raid1_info *mirror;
1082 struct r1bio *r1_bio;
1083 struct bio *read_bio;
1085 struct bitmap *bitmap;
1086 unsigned long flags;
1087 const int rw = bio_data_dir(bio);
1088 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1089 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1090 const unsigned long do_discard = (bio->bi_rw
1091 & (REQ_DISCARD | REQ_SECURE));
1092 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1093 struct md_rdev *blocked_rdev;
1094 struct blk_plug_cb *cb;
1095 struct raid1_plug_cb *plug = NULL;
1097 int sectors_handled;
1099 sector_t start_next_window;
1102 * Register the new request and wait if the reconstruction
1103 * thread has put up a bar for new requests.
1104 * Continue immediately if no resync is active currently.
1107 md_write_start(mddev, bio); /* wait on superblock update early */
1109 if (bio_data_dir(bio) == WRITE &&
1110 ((bio_end_sector(bio) > mddev->suspend_lo &&
1111 bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1112 (mddev_is_clustered(mddev) &&
1113 md_cluster_ops->area_resyncing(mddev, bio->bi_iter.bi_sector, bio_end_sector(bio))))) {
1114 /* As the suspend_* range is controlled by
1115 * userspace, we want an interruptible
1120 flush_signals(current);
1121 prepare_to_wait(&conf->wait_barrier,
1122 &w, TASK_INTERRUPTIBLE);
1123 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1124 bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1125 (mddev_is_clustered(mddev) &&
1126 !md_cluster_ops->area_resyncing(mddev,
1127 bio->bi_iter.bi_sector, bio_end_sector(bio))))
1131 finish_wait(&conf->wait_barrier, &w);
1134 start_next_window = wait_barrier(conf, bio);
1136 bitmap = mddev->bitmap;
1139 * make_request() can abort the operation when READA is being
1140 * used and no empty request is available.
1143 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1145 r1_bio->master_bio = bio;
1146 r1_bio->sectors = bio_sectors(bio);
1148 r1_bio->mddev = mddev;
1149 r1_bio->sector = bio->bi_iter.bi_sector;
1151 /* We might need to issue multiple reads to different
1152 * devices if there are bad blocks around, so we keep
1153 * track of the number of reads in bio->bi_phys_segments.
1154 * If this is 0, there is only one r1_bio and no locking
1155 * will be needed when requests complete. If it is
1156 * non-zero, then it is the number of not-completed requests.
1158 bio->bi_phys_segments = 0;
1159 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1163 * read balancing logic:
1168 rdisk = read_balance(conf, r1_bio, &max_sectors);
1171 /* couldn't find anywhere to read from */
1172 raid_end_bio_io(r1_bio);
1175 mirror = conf->mirrors + rdisk;
1177 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1179 /* Reading from a write-mostly device must
1180 * take care not to over-take any writes
1183 wait_event(bitmap->behind_wait,
1184 atomic_read(&bitmap->behind_writes) == 0);
1186 r1_bio->read_disk = rdisk;
1187 r1_bio->start_next_window = 0;
1189 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1190 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1193 r1_bio->bios[rdisk] = read_bio;
1195 read_bio->bi_iter.bi_sector = r1_bio->sector +
1196 mirror->rdev->data_offset;
1197 read_bio->bi_bdev = mirror->rdev->bdev;
1198 read_bio->bi_end_io = raid1_end_read_request;
1199 read_bio->bi_rw = READ | do_sync;
1200 read_bio->bi_private = r1_bio;
1202 if (max_sectors < r1_bio->sectors) {
1203 /* could not read all from this device, so we will
1204 * need another r1_bio.
1207 sectors_handled = (r1_bio->sector + max_sectors
1208 - bio->bi_iter.bi_sector);
1209 r1_bio->sectors = max_sectors;
1210 spin_lock_irq(&conf->device_lock);
1211 if (bio->bi_phys_segments == 0)
1212 bio->bi_phys_segments = 2;
1214 bio->bi_phys_segments++;
1215 spin_unlock_irq(&conf->device_lock);
1216 /* Cannot call generic_make_request directly
1217 * as that will be queued in __make_request
1218 * and subsequent mempool_alloc might block waiting
1219 * for it. So hand bio over to raid1d.
1221 reschedule_retry(r1_bio);
1223 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1225 r1_bio->master_bio = bio;
1226 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1228 r1_bio->mddev = mddev;
1229 r1_bio->sector = bio->bi_iter.bi_sector +
1233 generic_make_request(read_bio);
1240 if (conf->pending_count >= max_queued_requests) {
1241 md_wakeup_thread(mddev->thread);
1242 wait_event(conf->wait_barrier,
1243 conf->pending_count < max_queued_requests);
1245 /* first select target devices under rcu_lock and
1246 * inc refcount on their rdev. Record them by setting
1248 * If there are known/acknowledged bad blocks on any device on
1249 * which we have seen a write error, we want to avoid writing those
1251 * This potentially requires several writes to write around
1252 * the bad blocks. Each set of writes gets it's own r1bio
1253 * with a set of bios attached.
1256 disks = conf->raid_disks * 2;
1258 r1_bio->start_next_window = start_next_window;
1259 blocked_rdev = NULL;
1261 max_sectors = r1_bio->sectors;
1262 for (i = 0; i < disks; i++) {
1263 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1264 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1265 atomic_inc(&rdev->nr_pending);
1266 blocked_rdev = rdev;
1269 r1_bio->bios[i] = NULL;
1270 if (!rdev || test_bit(Faulty, &rdev->flags)
1271 || test_bit(Unmerged, &rdev->flags)) {
1272 if (i < conf->raid_disks)
1273 set_bit(R1BIO_Degraded, &r1_bio->state);
1277 atomic_inc(&rdev->nr_pending);
1278 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1283 is_bad = is_badblock(rdev, r1_bio->sector,
1285 &first_bad, &bad_sectors);
1287 /* mustn't write here until the bad block is
1289 set_bit(BlockedBadBlocks, &rdev->flags);
1290 blocked_rdev = rdev;
1293 if (is_bad && first_bad <= r1_bio->sector) {
1294 /* Cannot write here at all */
1295 bad_sectors -= (r1_bio->sector - first_bad);
1296 if (bad_sectors < max_sectors)
1297 /* mustn't write more than bad_sectors
1298 * to other devices yet
1300 max_sectors = bad_sectors;
1301 rdev_dec_pending(rdev, mddev);
1302 /* We don't set R1BIO_Degraded as that
1303 * only applies if the disk is
1304 * missing, so it might be re-added,
1305 * and we want to know to recover this
1307 * In this case the device is here,
1308 * and the fact that this chunk is not
1309 * in-sync is recorded in the bad
1315 int good_sectors = first_bad - r1_bio->sector;
1316 if (good_sectors < max_sectors)
1317 max_sectors = good_sectors;
1320 r1_bio->bios[i] = bio;
1324 if (unlikely(blocked_rdev)) {
1325 /* Wait for this device to become unblocked */
1327 sector_t old = start_next_window;
1329 for (j = 0; j < i; j++)
1330 if (r1_bio->bios[j])
1331 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1333 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1334 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1335 start_next_window = wait_barrier(conf, bio);
1337 * We must make sure the multi r1bios of bio have
1338 * the same value of bi_phys_segments
1340 if (bio->bi_phys_segments && old &&
1341 old != start_next_window)
1342 /* Wait for the former r1bio(s) to complete */
1343 wait_event(conf->wait_barrier,
1344 bio->bi_phys_segments == 1);
1348 if (max_sectors < r1_bio->sectors) {
1349 /* We are splitting this write into multiple parts, so
1350 * we need to prepare for allocating another r1_bio.
1352 r1_bio->sectors = max_sectors;
1353 spin_lock_irq(&conf->device_lock);
1354 if (bio->bi_phys_segments == 0)
1355 bio->bi_phys_segments = 2;
1357 bio->bi_phys_segments++;
1358 spin_unlock_irq(&conf->device_lock);
1360 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1362 atomic_set(&r1_bio->remaining, 1);
1363 atomic_set(&r1_bio->behind_remaining, 0);
1366 for (i = 0; i < disks; i++) {
1368 if (!r1_bio->bios[i])
1371 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1372 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1376 * Not if there are too many, or cannot
1377 * allocate memory, or a reader on WriteMostly
1378 * is waiting for behind writes to flush */
1380 (atomic_read(&bitmap->behind_writes)
1381 < mddev->bitmap_info.max_write_behind) &&
1382 !waitqueue_active(&bitmap->behind_wait))
1383 alloc_behind_pages(mbio, r1_bio);
1385 bitmap_startwrite(bitmap, r1_bio->sector,
1387 test_bit(R1BIO_BehindIO,
1391 if (r1_bio->behind_bvecs) {
1392 struct bio_vec *bvec;
1396 * We trimmed the bio, so _all is legit
1398 bio_for_each_segment_all(bvec, mbio, j)
1399 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1400 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1401 atomic_inc(&r1_bio->behind_remaining);
1404 r1_bio->bios[i] = mbio;
1406 mbio->bi_iter.bi_sector = (r1_bio->sector +
1407 conf->mirrors[i].rdev->data_offset);
1408 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1409 mbio->bi_end_io = raid1_end_write_request;
1411 WRITE | do_flush_fua | do_sync | do_discard | do_same;
1412 mbio->bi_private = r1_bio;
1414 atomic_inc(&r1_bio->remaining);
1416 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1418 plug = container_of(cb, struct raid1_plug_cb, cb);
1421 spin_lock_irqsave(&conf->device_lock, flags);
1423 bio_list_add(&plug->pending, mbio);
1424 plug->pending_cnt++;
1426 bio_list_add(&conf->pending_bio_list, mbio);
1427 conf->pending_count++;
1429 spin_unlock_irqrestore(&conf->device_lock, flags);
1431 md_wakeup_thread(mddev->thread);
1433 /* Mustn't call r1_bio_write_done before this next test,
1434 * as it could result in the bio being freed.
1436 if (sectors_handled < bio_sectors(bio)) {
1437 r1_bio_write_done(r1_bio);
1438 /* We need another r1_bio. It has already been counted
1439 * in bio->bi_phys_segments
1441 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1442 r1_bio->master_bio = bio;
1443 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1445 r1_bio->mddev = mddev;
1446 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1450 r1_bio_write_done(r1_bio);
1452 /* In case raid1d snuck in to freeze_array */
1453 wake_up(&conf->wait_barrier);
1456 static void status(struct seq_file *seq, struct mddev *mddev)
1458 struct r1conf *conf = mddev->private;
1461 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1462 conf->raid_disks - mddev->degraded);
1464 for (i = 0; i < conf->raid_disks; i++) {
1465 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1466 seq_printf(seq, "%s",
1467 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1470 seq_printf(seq, "]");
1473 static void error(struct mddev *mddev, struct md_rdev *rdev)
1475 char b[BDEVNAME_SIZE];
1476 struct r1conf *conf = mddev->private;
1479 * If it is not operational, then we have already marked it as dead
1480 * else if it is the last working disks, ignore the error, let the
1481 * next level up know.
1482 * else mark the drive as failed
1484 if (test_bit(In_sync, &rdev->flags)
1485 && (conf->raid_disks - mddev->degraded) == 1) {
1487 * Don't fail the drive, act as though we were just a
1488 * normal single drive.
1489 * However don't try a recovery from this drive as
1490 * it is very likely to fail.
1492 conf->recovery_disabled = mddev->recovery_disabled;
1495 set_bit(Blocked, &rdev->flags);
1496 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1497 unsigned long flags;
1498 spin_lock_irqsave(&conf->device_lock, flags);
1500 set_bit(Faulty, &rdev->flags);
1501 spin_unlock_irqrestore(&conf->device_lock, flags);
1503 set_bit(Faulty, &rdev->flags);
1505 * if recovery is running, make sure it aborts.
1507 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1508 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1510 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1511 "md/raid1:%s: Operation continuing on %d devices.\n",
1512 mdname(mddev), bdevname(rdev->bdev, b),
1513 mdname(mddev), conf->raid_disks - mddev->degraded);
1516 static void print_conf(struct r1conf *conf)
1520 printk(KERN_DEBUG "RAID1 conf printout:\n");
1522 printk(KERN_DEBUG "(!conf)\n");
1525 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1529 for (i = 0; i < conf->raid_disks; i++) {
1530 char b[BDEVNAME_SIZE];
1531 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1533 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1534 i, !test_bit(In_sync, &rdev->flags),
1535 !test_bit(Faulty, &rdev->flags),
1536 bdevname(rdev->bdev,b));
1541 static void close_sync(struct r1conf *conf)
1543 wait_barrier(conf, NULL);
1544 allow_barrier(conf, 0, 0);
1546 mempool_destroy(conf->r1buf_pool);
1547 conf->r1buf_pool = NULL;
1549 spin_lock_irq(&conf->resync_lock);
1550 conf->next_resync = 0;
1551 conf->start_next_window = MaxSector;
1552 conf->current_window_requests +=
1553 conf->next_window_requests;
1554 conf->next_window_requests = 0;
1555 spin_unlock_irq(&conf->resync_lock);
1558 static int raid1_spare_active(struct mddev *mddev)
1561 struct r1conf *conf = mddev->private;
1563 unsigned long flags;
1566 * Find all failed disks within the RAID1 configuration
1567 * and mark them readable.
1568 * Called under mddev lock, so rcu protection not needed.
1570 for (i = 0; i < conf->raid_disks; i++) {
1571 struct md_rdev *rdev = conf->mirrors[i].rdev;
1572 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1574 && !test_bit(Candidate, &repl->flags)
1575 && repl->recovery_offset == MaxSector
1576 && !test_bit(Faulty, &repl->flags)
1577 && !test_and_set_bit(In_sync, &repl->flags)) {
1578 /* replacement has just become active */
1580 !test_and_clear_bit(In_sync, &rdev->flags))
1583 /* Replaced device not technically
1584 * faulty, but we need to be sure
1585 * it gets removed and never re-added
1587 set_bit(Faulty, &rdev->flags);
1588 sysfs_notify_dirent_safe(
1593 && rdev->recovery_offset == MaxSector
1594 && !test_bit(Faulty, &rdev->flags)
1595 && !test_and_set_bit(In_sync, &rdev->flags)) {
1597 sysfs_notify_dirent_safe(rdev->sysfs_state);
1600 spin_lock_irqsave(&conf->device_lock, flags);
1601 mddev->degraded -= count;
1602 spin_unlock_irqrestore(&conf->device_lock, flags);
1608 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1610 struct r1conf *conf = mddev->private;
1613 struct raid1_info *p;
1615 int last = conf->raid_disks - 1;
1616 struct request_queue *q = bdev_get_queue(rdev->bdev);
1618 if (mddev->recovery_disabled == conf->recovery_disabled)
1621 if (rdev->raid_disk >= 0)
1622 first = last = rdev->raid_disk;
1624 if (q->merge_bvec_fn) {
1625 set_bit(Unmerged, &rdev->flags);
1626 mddev->merge_check_needed = 1;
1629 for (mirror = first; mirror <= last; mirror++) {
1630 p = conf->mirrors+mirror;
1634 disk_stack_limits(mddev->gendisk, rdev->bdev,
1635 rdev->data_offset << 9);
1637 p->head_position = 0;
1638 rdev->raid_disk = mirror;
1640 /* As all devices are equivalent, we don't need a full recovery
1641 * if this was recently any drive of the array
1643 if (rdev->saved_raid_disk < 0)
1645 rcu_assign_pointer(p->rdev, rdev);
1648 if (test_bit(WantReplacement, &p->rdev->flags) &&
1649 p[conf->raid_disks].rdev == NULL) {
1650 /* Add this device as a replacement */
1651 clear_bit(In_sync, &rdev->flags);
1652 set_bit(Replacement, &rdev->flags);
1653 rdev->raid_disk = mirror;
1656 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1660 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1661 /* Some requests might not have seen this new
1662 * merge_bvec_fn. We must wait for them to complete
1663 * before merging the device fully.
1664 * First we make sure any code which has tested
1665 * our function has submitted the request, then
1666 * we wait for all outstanding requests to complete.
1668 synchronize_sched();
1669 freeze_array(conf, 0);
1670 unfreeze_array(conf);
1671 clear_bit(Unmerged, &rdev->flags);
1673 md_integrity_add_rdev(rdev, mddev);
1674 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1675 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1680 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1682 struct r1conf *conf = mddev->private;
1684 int number = rdev->raid_disk;
1685 struct raid1_info *p = conf->mirrors + number;
1687 if (rdev != p->rdev)
1688 p = conf->mirrors + conf->raid_disks + number;
1691 if (rdev == p->rdev) {
1692 if (test_bit(In_sync, &rdev->flags) ||
1693 atomic_read(&rdev->nr_pending)) {
1697 /* Only remove non-faulty devices if recovery
1700 if (!test_bit(Faulty, &rdev->flags) &&
1701 mddev->recovery_disabled != conf->recovery_disabled &&
1702 mddev->degraded < conf->raid_disks) {
1708 if (atomic_read(&rdev->nr_pending)) {
1709 /* lost the race, try later */
1713 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1714 /* We just removed a device that is being replaced.
1715 * Move down the replacement. We drain all IO before
1716 * doing this to avoid confusion.
1718 struct md_rdev *repl =
1719 conf->mirrors[conf->raid_disks + number].rdev;
1720 freeze_array(conf, 0);
1721 clear_bit(Replacement, &repl->flags);
1723 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1724 unfreeze_array(conf);
1725 clear_bit(WantReplacement, &rdev->flags);
1727 clear_bit(WantReplacement, &rdev->flags);
1728 err = md_integrity_register(mddev);
1736 static void end_sync_read(struct bio *bio, int error)
1738 struct r1bio *r1_bio = bio->bi_private;
1740 update_head_pos(r1_bio->read_disk, r1_bio);
1743 * we have read a block, now it needs to be re-written,
1744 * or re-read if the read failed.
1745 * We don't do much here, just schedule handling by raid1d
1747 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1748 set_bit(R1BIO_Uptodate, &r1_bio->state);
1750 if (atomic_dec_and_test(&r1_bio->remaining))
1751 reschedule_retry(r1_bio);
1754 static void end_sync_write(struct bio *bio, int error)
1756 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1757 struct r1bio *r1_bio = bio->bi_private;
1758 struct mddev *mddev = r1_bio->mddev;
1759 struct r1conf *conf = mddev->private;
1764 mirror = find_bio_disk(r1_bio, bio);
1767 sector_t sync_blocks = 0;
1768 sector_t s = r1_bio->sector;
1769 long sectors_to_go = r1_bio->sectors;
1770 /* make sure these bits doesn't get cleared. */
1772 bitmap_end_sync(mddev->bitmap, s,
1775 sectors_to_go -= sync_blocks;
1776 } while (sectors_to_go > 0);
1777 set_bit(WriteErrorSeen,
1778 &conf->mirrors[mirror].rdev->flags);
1779 if (!test_and_set_bit(WantReplacement,
1780 &conf->mirrors[mirror].rdev->flags))
1781 set_bit(MD_RECOVERY_NEEDED, &
1783 set_bit(R1BIO_WriteError, &r1_bio->state);
1784 } else if (is_badblock(conf->mirrors[mirror].rdev,
1787 &first_bad, &bad_sectors) &&
1788 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1791 &first_bad, &bad_sectors)
1793 set_bit(R1BIO_MadeGood, &r1_bio->state);
1795 if (atomic_dec_and_test(&r1_bio->remaining)) {
1796 int s = r1_bio->sectors;
1797 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1798 test_bit(R1BIO_WriteError, &r1_bio->state))
1799 reschedule_retry(r1_bio);
1802 md_done_sync(mddev, s, uptodate);
1807 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1808 int sectors, struct page *page, int rw)
1810 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1814 set_bit(WriteErrorSeen, &rdev->flags);
1815 if (!test_and_set_bit(WantReplacement,
1817 set_bit(MD_RECOVERY_NEEDED, &
1818 rdev->mddev->recovery);
1820 /* need to record an error - either for the block or the device */
1821 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1822 md_error(rdev->mddev, rdev);
1826 static int fix_sync_read_error(struct r1bio *r1_bio)
1828 /* Try some synchronous reads of other devices to get
1829 * good data, much like with normal read errors. Only
1830 * read into the pages we already have so we don't
1831 * need to re-issue the read request.
1832 * We don't need to freeze the array, because being in an
1833 * active sync request, there is no normal IO, and
1834 * no overlapping syncs.
1835 * We don't need to check is_badblock() again as we
1836 * made sure that anything with a bad block in range
1837 * will have bi_end_io clear.
1839 struct mddev *mddev = r1_bio->mddev;
1840 struct r1conf *conf = mddev->private;
1841 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1842 sector_t sect = r1_bio->sector;
1843 int sectors = r1_bio->sectors;
1848 int d = r1_bio->read_disk;
1850 struct md_rdev *rdev;
1853 if (s > (PAGE_SIZE>>9))
1856 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1857 /* No rcu protection needed here devices
1858 * can only be removed when no resync is
1859 * active, and resync is currently active
1861 rdev = conf->mirrors[d].rdev;
1862 if (sync_page_io(rdev, sect, s<<9,
1863 bio->bi_io_vec[idx].bv_page,
1870 if (d == conf->raid_disks * 2)
1872 } while (!success && d != r1_bio->read_disk);
1875 char b[BDEVNAME_SIZE];
1877 /* Cannot read from anywhere, this block is lost.
1878 * Record a bad block on each device. If that doesn't
1879 * work just disable and interrupt the recovery.
1880 * Don't fail devices as that won't really help.
1882 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1883 " for block %llu\n",
1885 bdevname(bio->bi_bdev, b),
1886 (unsigned long long)r1_bio->sector);
1887 for (d = 0; d < conf->raid_disks * 2; d++) {
1888 rdev = conf->mirrors[d].rdev;
1889 if (!rdev || test_bit(Faulty, &rdev->flags))
1891 if (!rdev_set_badblocks(rdev, sect, s, 0))
1895 conf->recovery_disabled =
1896 mddev->recovery_disabled;
1897 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1898 md_done_sync(mddev, r1_bio->sectors, 0);
1910 /* write it back and re-read */
1911 while (d != r1_bio->read_disk) {
1913 d = conf->raid_disks * 2;
1915 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1917 rdev = conf->mirrors[d].rdev;
1918 if (r1_sync_page_io(rdev, sect, s,
1919 bio->bi_io_vec[idx].bv_page,
1921 r1_bio->bios[d]->bi_end_io = NULL;
1922 rdev_dec_pending(rdev, mddev);
1926 while (d != r1_bio->read_disk) {
1928 d = conf->raid_disks * 2;
1930 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1932 rdev = conf->mirrors[d].rdev;
1933 if (r1_sync_page_io(rdev, sect, s,
1934 bio->bi_io_vec[idx].bv_page,
1936 atomic_add(s, &rdev->corrected_errors);
1942 set_bit(R1BIO_Uptodate, &r1_bio->state);
1943 set_bit(BIO_UPTODATE, &bio->bi_flags);
1947 static void process_checks(struct r1bio *r1_bio)
1949 /* We have read all readable devices. If we haven't
1950 * got the block, then there is no hope left.
1951 * If we have, then we want to do a comparison
1952 * and skip the write if everything is the same.
1953 * If any blocks failed to read, then we need to
1954 * attempt an over-write
1956 struct mddev *mddev = r1_bio->mddev;
1957 struct r1conf *conf = mddev->private;
1962 /* Fix variable parts of all bios */
1963 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1964 for (i = 0; i < conf->raid_disks * 2; i++) {
1968 struct bio *b = r1_bio->bios[i];
1969 if (b->bi_end_io != end_sync_read)
1971 /* fixup the bio for reuse, but preserve BIO_UPTODATE */
1972 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags);
1975 clear_bit(BIO_UPTODATE, &b->bi_flags);
1977 b->bi_iter.bi_size = r1_bio->sectors << 9;
1978 b->bi_iter.bi_sector = r1_bio->sector +
1979 conf->mirrors[i].rdev->data_offset;
1980 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1981 b->bi_end_io = end_sync_read;
1982 b->bi_private = r1_bio;
1984 size = b->bi_iter.bi_size;
1985 for (j = 0; j < vcnt ; j++) {
1987 bi = &b->bi_io_vec[j];
1989 if (size > PAGE_SIZE)
1990 bi->bv_len = PAGE_SIZE;
1996 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1997 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1998 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1999 r1_bio->bios[primary]->bi_end_io = NULL;
2000 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2003 r1_bio->read_disk = primary;
2004 for (i = 0; i < conf->raid_disks * 2; i++) {
2006 struct bio *pbio = r1_bio->bios[primary];
2007 struct bio *sbio = r1_bio->bios[i];
2008 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags);
2010 if (sbio->bi_end_io != end_sync_read)
2012 /* Now we can 'fixup' the BIO_UPTODATE flag */
2013 set_bit(BIO_UPTODATE, &sbio->bi_flags);
2016 for (j = vcnt; j-- ; ) {
2018 p = pbio->bi_io_vec[j].bv_page;
2019 s = sbio->bi_io_vec[j].bv_page;
2020 if (memcmp(page_address(p),
2022 sbio->bi_io_vec[j].bv_len))
2028 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2029 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2031 /* No need to write to this device. */
2032 sbio->bi_end_io = NULL;
2033 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2037 bio_copy_data(sbio, pbio);
2041 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2043 struct r1conf *conf = mddev->private;
2045 int disks = conf->raid_disks * 2;
2046 struct bio *bio, *wbio;
2048 bio = r1_bio->bios[r1_bio->read_disk];
2050 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2051 /* ouch - failed to read all of that. */
2052 if (!fix_sync_read_error(r1_bio))
2055 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2056 process_checks(r1_bio);
2061 atomic_set(&r1_bio->remaining, 1);
2062 for (i = 0; i < disks ; i++) {
2063 wbio = r1_bio->bios[i];
2064 if (wbio->bi_end_io == NULL ||
2065 (wbio->bi_end_io == end_sync_read &&
2066 (i == r1_bio->read_disk ||
2067 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2070 wbio->bi_rw = WRITE;
2071 wbio->bi_end_io = end_sync_write;
2072 atomic_inc(&r1_bio->remaining);
2073 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2075 generic_make_request(wbio);
2078 if (atomic_dec_and_test(&r1_bio->remaining)) {
2079 /* if we're here, all write(s) have completed, so clean up */
2080 int s = r1_bio->sectors;
2081 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2082 test_bit(R1BIO_WriteError, &r1_bio->state))
2083 reschedule_retry(r1_bio);
2086 md_done_sync(mddev, s, 1);
2092 * This is a kernel thread which:
2094 * 1. Retries failed read operations on working mirrors.
2095 * 2. Updates the raid superblock when problems encounter.
2096 * 3. Performs writes following reads for array synchronising.
2099 static void fix_read_error(struct r1conf *conf, int read_disk,
2100 sector_t sect, int sectors)
2102 struct mddev *mddev = conf->mddev;
2108 struct md_rdev *rdev;
2110 if (s > (PAGE_SIZE>>9))
2114 /* Note: no rcu protection needed here
2115 * as this is synchronous in the raid1d thread
2116 * which is the thread that might remove
2117 * a device. If raid1d ever becomes multi-threaded....
2122 rdev = conf->mirrors[d].rdev;
2124 (test_bit(In_sync, &rdev->flags) ||
2125 (!test_bit(Faulty, &rdev->flags) &&
2126 rdev->recovery_offset >= sect + s)) &&
2127 is_badblock(rdev, sect, s,
2128 &first_bad, &bad_sectors) == 0 &&
2129 sync_page_io(rdev, sect, s<<9,
2130 conf->tmppage, READ, false))
2134 if (d == conf->raid_disks * 2)
2137 } while (!success && d != read_disk);
2140 /* Cannot read from anywhere - mark it bad */
2141 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2142 if (!rdev_set_badblocks(rdev, sect, s, 0))
2143 md_error(mddev, rdev);
2146 /* write it back and re-read */
2148 while (d != read_disk) {
2150 d = conf->raid_disks * 2;
2152 rdev = conf->mirrors[d].rdev;
2154 !test_bit(Faulty, &rdev->flags))
2155 r1_sync_page_io(rdev, sect, s,
2156 conf->tmppage, WRITE);
2159 while (d != read_disk) {
2160 char b[BDEVNAME_SIZE];
2162 d = conf->raid_disks * 2;
2164 rdev = conf->mirrors[d].rdev;
2166 !test_bit(Faulty, &rdev->flags)) {
2167 if (r1_sync_page_io(rdev, sect, s,
2168 conf->tmppage, READ)) {
2169 atomic_add(s, &rdev->corrected_errors);
2171 "md/raid1:%s: read error corrected "
2172 "(%d sectors at %llu on %s)\n",
2174 (unsigned long long)(sect +
2176 bdevname(rdev->bdev, b));
2185 static int narrow_write_error(struct r1bio *r1_bio, int i)
2187 struct mddev *mddev = r1_bio->mddev;
2188 struct r1conf *conf = mddev->private;
2189 struct md_rdev *rdev = conf->mirrors[i].rdev;
2191 /* bio has the data to be written to device 'i' where
2192 * we just recently had a write error.
2193 * We repeatedly clone the bio and trim down to one block,
2194 * then try the write. Where the write fails we record
2196 * It is conceivable that the bio doesn't exactly align with
2197 * blocks. We must handle this somehow.
2199 * We currently own a reference on the rdev.
2205 int sect_to_write = r1_bio->sectors;
2208 if (rdev->badblocks.shift < 0)
2211 block_sectors = roundup(1 << rdev->badblocks.shift,
2212 bdev_logical_block_size(rdev->bdev) >> 9);
2213 sector = r1_bio->sector;
2214 sectors = ((sector + block_sectors)
2215 & ~(sector_t)(block_sectors - 1))
2218 while (sect_to_write) {
2220 if (sectors > sect_to_write)
2221 sectors = sect_to_write;
2222 /* Write at 'sector' for 'sectors'*/
2224 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2225 unsigned vcnt = r1_bio->behind_page_count;
2226 struct bio_vec *vec = r1_bio->behind_bvecs;
2228 while (!vec->bv_page) {
2233 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2234 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2236 wbio->bi_vcnt = vcnt;
2238 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2241 wbio->bi_rw = WRITE;
2242 wbio->bi_iter.bi_sector = r1_bio->sector;
2243 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2245 bio_trim(wbio, sector - r1_bio->sector, sectors);
2246 wbio->bi_iter.bi_sector += rdev->data_offset;
2247 wbio->bi_bdev = rdev->bdev;
2248 if (submit_bio_wait(WRITE, wbio) == 0)
2250 ok = rdev_set_badblocks(rdev, sector,
2255 sect_to_write -= sectors;
2257 sectors = block_sectors;
2262 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2265 int s = r1_bio->sectors;
2266 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2267 struct md_rdev *rdev = conf->mirrors[m].rdev;
2268 struct bio *bio = r1_bio->bios[m];
2269 if (bio->bi_end_io == NULL)
2271 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2272 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2273 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2275 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2276 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2277 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2278 md_error(conf->mddev, rdev);
2282 md_done_sync(conf->mddev, s, 1);
2285 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2288 for (m = 0; m < conf->raid_disks * 2 ; m++)
2289 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2290 struct md_rdev *rdev = conf->mirrors[m].rdev;
2291 rdev_clear_badblocks(rdev,
2293 r1_bio->sectors, 0);
2294 rdev_dec_pending(rdev, conf->mddev);
2295 } else if (r1_bio->bios[m] != NULL) {
2296 /* This drive got a write error. We need to
2297 * narrow down and record precise write
2300 if (!narrow_write_error(r1_bio, m)) {
2301 md_error(conf->mddev,
2302 conf->mirrors[m].rdev);
2303 /* an I/O failed, we can't clear the bitmap */
2304 set_bit(R1BIO_Degraded, &r1_bio->state);
2306 rdev_dec_pending(conf->mirrors[m].rdev,
2309 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2310 close_write(r1_bio);
2311 raid_end_bio_io(r1_bio);
2314 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2318 struct mddev *mddev = conf->mddev;
2320 char b[BDEVNAME_SIZE];
2321 struct md_rdev *rdev;
2323 clear_bit(R1BIO_ReadError, &r1_bio->state);
2324 /* we got a read error. Maybe the drive is bad. Maybe just
2325 * the block and we can fix it.
2326 * We freeze all other IO, and try reading the block from
2327 * other devices. When we find one, we re-write
2328 * and check it that fixes the read error.
2329 * This is all done synchronously while the array is
2332 if (mddev->ro == 0) {
2333 freeze_array(conf, 1);
2334 fix_read_error(conf, r1_bio->read_disk,
2335 r1_bio->sector, r1_bio->sectors);
2336 unfreeze_array(conf);
2338 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2339 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2341 bio = r1_bio->bios[r1_bio->read_disk];
2342 bdevname(bio->bi_bdev, b);
2344 disk = read_balance(conf, r1_bio, &max_sectors);
2346 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2347 " read error for block %llu\n",
2348 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2349 raid_end_bio_io(r1_bio);
2351 const unsigned long do_sync
2352 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2354 r1_bio->bios[r1_bio->read_disk] =
2355 mddev->ro ? IO_BLOCKED : NULL;
2358 r1_bio->read_disk = disk;
2359 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2360 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2362 r1_bio->bios[r1_bio->read_disk] = bio;
2363 rdev = conf->mirrors[disk].rdev;
2364 printk_ratelimited(KERN_ERR
2365 "md/raid1:%s: redirecting sector %llu"
2366 " to other mirror: %s\n",
2368 (unsigned long long)r1_bio->sector,
2369 bdevname(rdev->bdev, b));
2370 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2371 bio->bi_bdev = rdev->bdev;
2372 bio->bi_end_io = raid1_end_read_request;
2373 bio->bi_rw = READ | do_sync;
2374 bio->bi_private = r1_bio;
2375 if (max_sectors < r1_bio->sectors) {
2376 /* Drat - have to split this up more */
2377 struct bio *mbio = r1_bio->master_bio;
2378 int sectors_handled = (r1_bio->sector + max_sectors
2379 - mbio->bi_iter.bi_sector);
2380 r1_bio->sectors = max_sectors;
2381 spin_lock_irq(&conf->device_lock);
2382 if (mbio->bi_phys_segments == 0)
2383 mbio->bi_phys_segments = 2;
2385 mbio->bi_phys_segments++;
2386 spin_unlock_irq(&conf->device_lock);
2387 generic_make_request(bio);
2390 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2392 r1_bio->master_bio = mbio;
2393 r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2395 set_bit(R1BIO_ReadError, &r1_bio->state);
2396 r1_bio->mddev = mddev;
2397 r1_bio->sector = mbio->bi_iter.bi_sector +
2402 generic_make_request(bio);
2406 static void raid1d(struct md_thread *thread)
2408 struct mddev *mddev = thread->mddev;
2409 struct r1bio *r1_bio;
2410 unsigned long flags;
2411 struct r1conf *conf = mddev->private;
2412 struct list_head *head = &conf->retry_list;
2413 struct blk_plug plug;
2415 md_check_recovery(mddev);
2417 blk_start_plug(&plug);
2420 flush_pending_writes(conf);
2422 spin_lock_irqsave(&conf->device_lock, flags);
2423 if (list_empty(head)) {
2424 spin_unlock_irqrestore(&conf->device_lock, flags);
2427 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2428 list_del(head->prev);
2430 spin_unlock_irqrestore(&conf->device_lock, flags);
2432 mddev = r1_bio->mddev;
2433 conf = mddev->private;
2434 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2435 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2436 test_bit(R1BIO_WriteError, &r1_bio->state))
2437 handle_sync_write_finished(conf, r1_bio);
2439 sync_request_write(mddev, r1_bio);
2440 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2441 test_bit(R1BIO_WriteError, &r1_bio->state))
2442 handle_write_finished(conf, r1_bio);
2443 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2444 handle_read_error(conf, r1_bio);
2446 /* just a partial read to be scheduled from separate
2449 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2452 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2453 md_check_recovery(mddev);
2455 blk_finish_plug(&plug);
2458 static int init_resync(struct r1conf *conf)
2462 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2463 BUG_ON(conf->r1buf_pool);
2464 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2466 if (!conf->r1buf_pool)
2468 conf->next_resync = 0;
2473 * perform a "sync" on one "block"
2475 * We need to make sure that no normal I/O request - particularly write
2476 * requests - conflict with active sync requests.
2478 * This is achieved by tracking pending requests and a 'barrier' concept
2479 * that can be installed to exclude normal IO requests.
2482 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2484 struct r1conf *conf = mddev->private;
2485 struct r1bio *r1_bio;
2487 sector_t max_sector, nr_sectors;
2491 int write_targets = 0, read_targets = 0;
2492 sector_t sync_blocks;
2493 int still_degraded = 0;
2494 int good_sectors = RESYNC_SECTORS;
2495 int min_bad = 0; /* number of sectors that are bad in all devices */
2497 if (!conf->r1buf_pool)
2498 if (init_resync(conf))
2501 max_sector = mddev->dev_sectors;
2502 if (sector_nr >= max_sector) {
2503 /* If we aborted, we need to abort the
2504 * sync on the 'current' bitmap chunk (there will
2505 * only be one in raid1 resync.
2506 * We can find the current addess in mddev->curr_resync
2508 if (mddev->curr_resync < max_sector) /* aborted */
2509 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2511 else /* completed sync */
2514 bitmap_close_sync(mddev->bitmap);
2519 if (mddev->bitmap == NULL &&
2520 mddev->recovery_cp == MaxSector &&
2521 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2522 conf->fullsync == 0) {
2524 return max_sector - sector_nr;
2526 /* before building a request, check if we can skip these blocks..
2527 * This call the bitmap_start_sync doesn't actually record anything
2529 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2530 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2531 /* We can skip this block, and probably several more */
2536 * If there is non-resync activity waiting for a turn,
2537 * and resync is going fast enough,
2538 * then let it though before starting on this new sync request.
2540 if (!go_faster && conf->nr_waiting)
2541 msleep_interruptible(1000);
2543 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2544 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2546 raise_barrier(conf, sector_nr);
2550 * If we get a correctably read error during resync or recovery,
2551 * we might want to read from a different device. So we
2552 * flag all drives that could conceivably be read from for READ,
2553 * and any others (which will be non-In_sync devices) for WRITE.
2554 * If a read fails, we try reading from something else for which READ
2558 r1_bio->mddev = mddev;
2559 r1_bio->sector = sector_nr;
2561 set_bit(R1BIO_IsSync, &r1_bio->state);
2563 for (i = 0; i < conf->raid_disks * 2; i++) {
2564 struct md_rdev *rdev;
2565 bio = r1_bio->bios[i];
2568 rdev = rcu_dereference(conf->mirrors[i].rdev);
2570 test_bit(Faulty, &rdev->flags)) {
2571 if (i < conf->raid_disks)
2573 } else if (!test_bit(In_sync, &rdev->flags)) {
2575 bio->bi_end_io = end_sync_write;
2578 /* may need to read from here */
2579 sector_t first_bad = MaxSector;
2582 if (is_badblock(rdev, sector_nr, good_sectors,
2583 &first_bad, &bad_sectors)) {
2584 if (first_bad > sector_nr)
2585 good_sectors = first_bad - sector_nr;
2587 bad_sectors -= (sector_nr - first_bad);
2589 min_bad > bad_sectors)
2590 min_bad = bad_sectors;
2593 if (sector_nr < first_bad) {
2594 if (test_bit(WriteMostly, &rdev->flags)) {
2602 bio->bi_end_io = end_sync_read;
2604 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2605 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2606 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2608 * The device is suitable for reading (InSync),
2609 * but has bad block(s) here. Let's try to correct them,
2610 * if we are doing resync or repair. Otherwise, leave
2611 * this device alone for this sync request.
2614 bio->bi_end_io = end_sync_write;
2618 if (bio->bi_end_io) {
2619 atomic_inc(&rdev->nr_pending);
2620 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2621 bio->bi_bdev = rdev->bdev;
2622 bio->bi_private = r1_bio;
2628 r1_bio->read_disk = disk;
2630 if (read_targets == 0 && min_bad > 0) {
2631 /* These sectors are bad on all InSync devices, so we
2632 * need to mark them bad on all write targets
2635 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2636 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2637 struct md_rdev *rdev = conf->mirrors[i].rdev;
2638 ok = rdev_set_badblocks(rdev, sector_nr,
2642 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2647 /* Cannot record the badblocks, so need to
2649 * If there are multiple read targets, could just
2650 * fail the really bad ones ???
2652 conf->recovery_disabled = mddev->recovery_disabled;
2653 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2659 if (min_bad > 0 && min_bad < good_sectors) {
2660 /* only resync enough to reach the next bad->good
2662 good_sectors = min_bad;
2665 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2666 /* extra read targets are also write targets */
2667 write_targets += read_targets-1;
2669 if (write_targets == 0 || read_targets == 0) {
2670 /* There is nowhere to write, so all non-sync
2671 * drives must be failed - so we are finished
2675 max_sector = sector_nr + min_bad;
2676 rv = max_sector - sector_nr;
2682 if (max_sector > mddev->resync_max)
2683 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2684 if (max_sector > sector_nr + good_sectors)
2685 max_sector = sector_nr + good_sectors;
2690 int len = PAGE_SIZE;
2691 if (sector_nr + (len>>9) > max_sector)
2692 len = (max_sector - sector_nr) << 9;
2695 if (sync_blocks == 0) {
2696 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2697 &sync_blocks, still_degraded) &&
2699 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2701 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2702 if ((len >> 9) > sync_blocks)
2703 len = sync_blocks<<9;
2706 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2707 bio = r1_bio->bios[i];
2708 if (bio->bi_end_io) {
2709 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2710 if (bio_add_page(bio, page, len, 0) == 0) {
2712 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2715 bio = r1_bio->bios[i];
2716 if (bio->bi_end_io==NULL)
2718 /* remove last page from this bio */
2720 bio->bi_iter.bi_size -= len;
2721 __clear_bit(BIO_SEG_VALID, &bio->bi_flags);
2727 nr_sectors += len>>9;
2728 sector_nr += len>>9;
2729 sync_blocks -= (len>>9);
2730 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2732 r1_bio->sectors = nr_sectors;
2734 /* For a user-requested sync, we read all readable devices and do a
2737 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2738 atomic_set(&r1_bio->remaining, read_targets);
2739 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2740 bio = r1_bio->bios[i];
2741 if (bio->bi_end_io == end_sync_read) {
2743 md_sync_acct(bio->bi_bdev, nr_sectors);
2744 generic_make_request(bio);
2748 atomic_set(&r1_bio->remaining, 1);
2749 bio = r1_bio->bios[r1_bio->read_disk];
2750 md_sync_acct(bio->bi_bdev, nr_sectors);
2751 generic_make_request(bio);
2757 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2762 return mddev->dev_sectors;
2765 static struct r1conf *setup_conf(struct mddev *mddev)
2767 struct r1conf *conf;
2769 struct raid1_info *disk;
2770 struct md_rdev *rdev;
2773 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2777 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2778 * mddev->raid_disks * 2,
2783 conf->tmppage = alloc_page(GFP_KERNEL);
2787 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2788 if (!conf->poolinfo)
2790 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2791 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2794 if (!conf->r1bio_pool)
2797 conf->poolinfo->mddev = mddev;
2800 spin_lock_init(&conf->device_lock);
2801 rdev_for_each(rdev, mddev) {
2802 struct request_queue *q;
2803 int disk_idx = rdev->raid_disk;
2804 if (disk_idx >= mddev->raid_disks
2807 if (test_bit(Replacement, &rdev->flags))
2808 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2810 disk = conf->mirrors + disk_idx;
2815 q = bdev_get_queue(rdev->bdev);
2816 if (q->merge_bvec_fn)
2817 mddev->merge_check_needed = 1;
2819 disk->head_position = 0;
2820 disk->seq_start = MaxSector;
2822 conf->raid_disks = mddev->raid_disks;
2823 conf->mddev = mddev;
2824 INIT_LIST_HEAD(&conf->retry_list);
2826 spin_lock_init(&conf->resync_lock);
2827 init_waitqueue_head(&conf->wait_barrier);
2829 bio_list_init(&conf->pending_bio_list);
2830 conf->pending_count = 0;
2831 conf->recovery_disabled = mddev->recovery_disabled - 1;
2833 conf->start_next_window = MaxSector;
2834 conf->current_window_requests = conf->next_window_requests = 0;
2837 for (i = 0; i < conf->raid_disks * 2; i++) {
2839 disk = conf->mirrors + i;
2841 if (i < conf->raid_disks &&
2842 disk[conf->raid_disks].rdev) {
2843 /* This slot has a replacement. */
2845 /* No original, just make the replacement
2846 * a recovering spare
2849 disk[conf->raid_disks].rdev;
2850 disk[conf->raid_disks].rdev = NULL;
2851 } else if (!test_bit(In_sync, &disk->rdev->flags))
2852 /* Original is not in_sync - bad */
2857 !test_bit(In_sync, &disk->rdev->flags)) {
2858 disk->head_position = 0;
2860 (disk->rdev->saved_raid_disk < 0))
2866 conf->thread = md_register_thread(raid1d, mddev, "raid1");
2867 if (!conf->thread) {
2869 "md/raid1:%s: couldn't allocate thread\n",
2878 if (conf->r1bio_pool)
2879 mempool_destroy(conf->r1bio_pool);
2880 kfree(conf->mirrors);
2881 safe_put_page(conf->tmppage);
2882 kfree(conf->poolinfo);
2885 return ERR_PTR(err);
2888 static void raid1_free(struct mddev *mddev, void *priv);
2889 static int run(struct mddev *mddev)
2891 struct r1conf *conf;
2893 struct md_rdev *rdev;
2895 bool discard_supported = false;
2897 if (mddev->level != 1) {
2898 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2899 mdname(mddev), mddev->level);
2902 if (mddev->reshape_position != MaxSector) {
2903 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2908 * copy the already verified devices into our private RAID1
2909 * bookkeeping area. [whatever we allocate in run(),
2910 * should be freed in raid1_free()]
2912 if (mddev->private == NULL)
2913 conf = setup_conf(mddev);
2915 conf = mddev->private;
2918 return PTR_ERR(conf);
2921 blk_queue_max_write_same_sectors(mddev->queue, 0);
2923 rdev_for_each(rdev, mddev) {
2924 if (!mddev->gendisk)
2926 disk_stack_limits(mddev->gendisk, rdev->bdev,
2927 rdev->data_offset << 9);
2928 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2929 discard_supported = true;
2932 mddev->degraded = 0;
2933 for (i=0; i < conf->raid_disks; i++)
2934 if (conf->mirrors[i].rdev == NULL ||
2935 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2936 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2939 if (conf->raid_disks - mddev->degraded == 1)
2940 mddev->recovery_cp = MaxSector;
2942 if (mddev->recovery_cp != MaxSector)
2943 printk(KERN_NOTICE "md/raid1:%s: not clean"
2944 " -- starting background reconstruction\n",
2947 "md/raid1:%s: active with %d out of %d mirrors\n",
2948 mdname(mddev), mddev->raid_disks - mddev->degraded,
2952 * Ok, everything is just fine now
2954 mddev->thread = conf->thread;
2955 conf->thread = NULL;
2956 mddev->private = conf;
2958 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2961 if (discard_supported)
2962 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2965 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2969 ret = md_integrity_register(mddev);
2971 md_unregister_thread(&mddev->thread);
2972 raid1_free(mddev, conf);
2977 static void raid1_free(struct mddev *mddev, void *priv)
2979 struct r1conf *conf = priv;
2981 if (conf->r1bio_pool)
2982 mempool_destroy(conf->r1bio_pool);
2983 kfree(conf->mirrors);
2984 safe_put_page(conf->tmppage);
2985 kfree(conf->poolinfo);
2989 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2991 /* no resync is happening, and there is enough space
2992 * on all devices, so we can resize.
2993 * We need to make sure resync covers any new space.
2994 * If the array is shrinking we should possibly wait until
2995 * any io in the removed space completes, but it hardly seems
2998 sector_t newsize = raid1_size(mddev, sectors, 0);
2999 if (mddev->external_size &&
3000 mddev->array_sectors > newsize)
3002 if (mddev->bitmap) {
3003 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3007 md_set_array_sectors(mddev, newsize);
3008 set_capacity(mddev->gendisk, mddev->array_sectors);
3009 revalidate_disk(mddev->gendisk);
3010 if (sectors > mddev->dev_sectors &&
3011 mddev->recovery_cp > mddev->dev_sectors) {
3012 mddev->recovery_cp = mddev->dev_sectors;
3013 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3015 mddev->dev_sectors = sectors;
3016 mddev->resync_max_sectors = sectors;
3020 static int raid1_reshape(struct mddev *mddev)
3023 * 1/ resize the r1bio_pool
3024 * 2/ resize conf->mirrors
3026 * We allocate a new r1bio_pool if we can.
3027 * Then raise a device barrier and wait until all IO stops.
3028 * Then resize conf->mirrors and swap in the new r1bio pool.
3030 * At the same time, we "pack" the devices so that all the missing
3031 * devices have the higher raid_disk numbers.
3033 mempool_t *newpool, *oldpool;
3034 struct pool_info *newpoolinfo;
3035 struct raid1_info *newmirrors;
3036 struct r1conf *conf = mddev->private;
3037 int cnt, raid_disks;
3038 unsigned long flags;
3041 /* Cannot change chunk_size, layout, or level */
3042 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3043 mddev->layout != mddev->new_layout ||
3044 mddev->level != mddev->new_level) {
3045 mddev->new_chunk_sectors = mddev->chunk_sectors;
3046 mddev->new_layout = mddev->layout;
3047 mddev->new_level = mddev->level;
3051 err = md_allow_write(mddev);
3055 raid_disks = mddev->raid_disks + mddev->delta_disks;
3057 if (raid_disks < conf->raid_disks) {
3059 for (d= 0; d < conf->raid_disks; d++)
3060 if (conf->mirrors[d].rdev)
3062 if (cnt > raid_disks)
3066 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3069 newpoolinfo->mddev = mddev;
3070 newpoolinfo->raid_disks = raid_disks * 2;
3072 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3073 r1bio_pool_free, newpoolinfo);
3078 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3082 mempool_destroy(newpool);
3086 freeze_array(conf, 0);
3088 /* ok, everything is stopped */
3089 oldpool = conf->r1bio_pool;
3090 conf->r1bio_pool = newpool;
3092 for (d = d2 = 0; d < conf->raid_disks; d++) {
3093 struct md_rdev *rdev = conf->mirrors[d].rdev;
3094 if (rdev && rdev->raid_disk != d2) {
3095 sysfs_unlink_rdev(mddev, rdev);
3096 rdev->raid_disk = d2;
3097 sysfs_unlink_rdev(mddev, rdev);
3098 if (sysfs_link_rdev(mddev, rdev))
3100 "md/raid1:%s: cannot register rd%d\n",
3101 mdname(mddev), rdev->raid_disk);
3104 newmirrors[d2++].rdev = rdev;
3106 kfree(conf->mirrors);
3107 conf->mirrors = newmirrors;
3108 kfree(conf->poolinfo);
3109 conf->poolinfo = newpoolinfo;
3111 spin_lock_irqsave(&conf->device_lock, flags);
3112 mddev->degraded += (raid_disks - conf->raid_disks);
3113 spin_unlock_irqrestore(&conf->device_lock, flags);
3114 conf->raid_disks = mddev->raid_disks = raid_disks;
3115 mddev->delta_disks = 0;
3117 unfreeze_array(conf);
3119 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3120 md_wakeup_thread(mddev->thread);
3122 mempool_destroy(oldpool);
3126 static void raid1_quiesce(struct mddev *mddev, int state)
3128 struct r1conf *conf = mddev->private;
3131 case 2: /* wake for suspend */
3132 wake_up(&conf->wait_barrier);
3135 freeze_array(conf, 0);
3138 unfreeze_array(conf);
3143 static void *raid1_takeover(struct mddev *mddev)
3145 /* raid1 can take over:
3146 * raid5 with 2 devices, any layout or chunk size
3148 if (mddev->level == 5 && mddev->raid_disks == 2) {
3149 struct r1conf *conf;
3150 mddev->new_level = 1;
3151 mddev->new_layout = 0;
3152 mddev->new_chunk_sectors = 0;
3153 conf = setup_conf(mddev);
3155 /* Array must appear to be quiesced */
3156 conf->array_frozen = 1;
3159 return ERR_PTR(-EINVAL);
3162 static struct md_personality raid1_personality =
3166 .owner = THIS_MODULE,
3167 .make_request = make_request,
3171 .error_handler = error,
3172 .hot_add_disk = raid1_add_disk,
3173 .hot_remove_disk= raid1_remove_disk,
3174 .spare_active = raid1_spare_active,
3175 .sync_request = sync_request,
3176 .resize = raid1_resize,
3178 .check_reshape = raid1_reshape,
3179 .quiesce = raid1_quiesce,
3180 .takeover = raid1_takeover,
3181 .congested = raid1_congested,
3182 .mergeable_bvec = raid1_mergeable_bvec,
3185 static int __init raid_init(void)
3187 return register_md_personality(&raid1_personality);
3190 static void raid_exit(void)
3192 unregister_md_personality(&raid1_personality);
3195 module_init(raid_init);
3196 module_exit(raid_exit);
3197 MODULE_LICENSE("GPL");
3198 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3199 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3200 MODULE_ALIAS("md-raid1");
3201 MODULE_ALIAS("md-level-1");
3203 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);