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>
41 #include <trace/events/block.h>
45 #include "md-bitmap.h"
47 #define UNSUPPORTED_MDDEV_FLAGS \
48 ((1L << MD_HAS_JOURNAL) | \
49 (1L << MD_JOURNAL_CLEAN) | \
50 (1L << MD_HAS_PPL) | \
51 (1L << MD_HAS_MULTIPLE_PPLS))
54 * Number of guaranteed r1bios in case of extreme VM load:
56 #define NR_RAID1_BIOS 256
58 /* when we get a read error on a read-only array, we redirect to another
59 * device without failing the first device, or trying to over-write to
60 * correct the read error. To keep track of bad blocks on a per-bio
61 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
63 #define IO_BLOCKED ((struct bio *)1)
64 /* When we successfully write to a known bad-block, we need to remove the
65 * bad-block marking which must be done from process context. So we record
66 * the success by setting devs[n].bio to IO_MADE_GOOD
68 #define IO_MADE_GOOD ((struct bio *)2)
70 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
72 /* When there are this many requests queue to be written by
73 * the raid1 thread, we become 'congested' to provide back-pressure
76 static int max_queued_requests = 1024;
78 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
79 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
81 #define raid1_log(md, fmt, args...) \
82 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
87 * for resync bio, r1bio pointer can be retrieved from the per-bio
88 * 'struct resync_pages'.
90 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
92 return get_resync_pages(bio)->raid_bio;
95 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
97 struct pool_info *pi = data;
98 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
100 /* allocate a r1bio with room for raid_disks entries in the bios array */
101 return kzalloc(size, gfp_flags);
104 static void r1bio_pool_free(void *r1_bio, void *data)
109 #define RESYNC_DEPTH 32
110 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
111 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
112 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
113 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
114 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
116 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
118 struct pool_info *pi = data;
119 struct r1bio *r1_bio;
123 struct resync_pages *rps;
125 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
129 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
135 * Allocate bios : 1 for reading, n-1 for writing
137 for (j = pi->raid_disks ; j-- ; ) {
138 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
141 r1_bio->bios[j] = bio;
144 * Allocate RESYNC_PAGES data pages and attach them to
146 * If this is a user-requested check/repair, allocate
147 * RESYNC_PAGES for each bio.
149 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
150 need_pages = pi->raid_disks;
153 for (j = 0; j < pi->raid_disks; j++) {
154 struct resync_pages *rp = &rps[j];
156 bio = r1_bio->bios[j];
158 if (j < need_pages) {
159 if (resync_alloc_pages(rp, gfp_flags))
162 memcpy(rp, &rps[0], sizeof(*rp));
163 resync_get_all_pages(rp);
166 rp->raid_bio = r1_bio;
167 bio->bi_private = rp;
170 r1_bio->master_bio = NULL;
176 resync_free_pages(&rps[j]);
179 while (++j < pi->raid_disks)
180 bio_put(r1_bio->bios[j]);
184 r1bio_pool_free(r1_bio, data);
188 static void r1buf_pool_free(void *__r1_bio, void *data)
190 struct pool_info *pi = data;
192 struct r1bio *r1bio = __r1_bio;
193 struct resync_pages *rp = NULL;
195 for (i = pi->raid_disks; i--; ) {
196 rp = get_resync_pages(r1bio->bios[i]);
197 resync_free_pages(rp);
198 bio_put(r1bio->bios[i]);
201 /* resync pages array stored in the 1st bio's .bi_private */
204 r1bio_pool_free(r1bio, data);
207 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
211 for (i = 0; i < conf->raid_disks * 2; i++) {
212 struct bio **bio = r1_bio->bios + i;
213 if (!BIO_SPECIAL(*bio))
219 static void free_r1bio(struct r1bio *r1_bio)
221 struct r1conf *conf = r1_bio->mddev->private;
223 put_all_bios(conf, r1_bio);
224 mempool_free(r1_bio, &conf->r1bio_pool);
227 static void put_buf(struct r1bio *r1_bio)
229 struct r1conf *conf = r1_bio->mddev->private;
230 sector_t sect = r1_bio->sector;
233 for (i = 0; i < conf->raid_disks * 2; i++) {
234 struct bio *bio = r1_bio->bios[i];
236 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
239 mempool_free(r1_bio, &conf->r1buf_pool);
241 lower_barrier(conf, sect);
244 static void reschedule_retry(struct r1bio *r1_bio)
247 struct mddev *mddev = r1_bio->mddev;
248 struct r1conf *conf = mddev->private;
251 idx = sector_to_idx(r1_bio->sector);
252 spin_lock_irqsave(&conf->device_lock, flags);
253 list_add(&r1_bio->retry_list, &conf->retry_list);
254 atomic_inc(&conf->nr_queued[idx]);
255 spin_unlock_irqrestore(&conf->device_lock, flags);
257 wake_up(&conf->wait_barrier);
258 md_wakeup_thread(mddev->thread);
262 * raid_end_bio_io() is called when we have finished servicing a mirrored
263 * operation and are ready to return a success/failure code to the buffer
266 static void call_bio_endio(struct r1bio *r1_bio)
268 struct bio *bio = r1_bio->master_bio;
269 struct r1conf *conf = r1_bio->mddev->private;
271 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
272 bio->bi_status = BLK_STS_IOERR;
276 * Wake up any possible resync thread that waits for the device
279 allow_barrier(conf, r1_bio->sector);
282 static void raid_end_bio_io(struct r1bio *r1_bio)
284 struct bio *bio = r1_bio->master_bio;
286 /* if nobody has done the final endio yet, do it now */
287 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
288 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
289 (bio_data_dir(bio) == WRITE) ? "write" : "read",
290 (unsigned long long) bio->bi_iter.bi_sector,
291 (unsigned long long) bio_end_sector(bio) - 1);
293 call_bio_endio(r1_bio);
299 * Update disk head position estimator based on IRQ completion info.
301 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
303 struct r1conf *conf = r1_bio->mddev->private;
305 conf->mirrors[disk].head_position =
306 r1_bio->sector + (r1_bio->sectors);
310 * Find the disk number which triggered given bio
312 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
315 struct r1conf *conf = r1_bio->mddev->private;
316 int raid_disks = conf->raid_disks;
318 for (mirror = 0; mirror < raid_disks * 2; mirror++)
319 if (r1_bio->bios[mirror] == bio)
322 BUG_ON(mirror == raid_disks * 2);
323 update_head_pos(mirror, r1_bio);
328 static void raid1_end_read_request(struct bio *bio)
330 int uptodate = !bio->bi_status;
331 struct r1bio *r1_bio = bio->bi_private;
332 struct r1conf *conf = r1_bio->mddev->private;
333 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
336 * this branch is our 'one mirror IO has finished' event handler:
338 update_head_pos(r1_bio->read_disk, r1_bio);
341 set_bit(R1BIO_Uptodate, &r1_bio->state);
342 else if (test_bit(FailFast, &rdev->flags) &&
343 test_bit(R1BIO_FailFast, &r1_bio->state))
344 /* This was a fail-fast read so we definitely
348 /* If all other devices have failed, we want to return
349 * the error upwards rather than fail the last device.
350 * Here we redefine "uptodate" to mean "Don't want to retry"
353 spin_lock_irqsave(&conf->device_lock, flags);
354 if (r1_bio->mddev->degraded == conf->raid_disks ||
355 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
356 test_bit(In_sync, &rdev->flags)))
358 spin_unlock_irqrestore(&conf->device_lock, flags);
362 raid_end_bio_io(r1_bio);
363 rdev_dec_pending(rdev, conf->mddev);
368 char b[BDEVNAME_SIZE];
369 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
371 bdevname(rdev->bdev, b),
372 (unsigned long long)r1_bio->sector);
373 set_bit(R1BIO_ReadError, &r1_bio->state);
374 reschedule_retry(r1_bio);
375 /* don't drop the reference on read_disk yet */
379 static void close_write(struct r1bio *r1_bio)
381 /* it really is the end of this request */
382 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
383 bio_free_pages(r1_bio->behind_master_bio);
384 bio_put(r1_bio->behind_master_bio);
385 r1_bio->behind_master_bio = NULL;
387 /* clear the bitmap if all writes complete successfully */
388 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
390 !test_bit(R1BIO_Degraded, &r1_bio->state),
391 test_bit(R1BIO_BehindIO, &r1_bio->state));
392 md_write_end(r1_bio->mddev);
395 static void r1_bio_write_done(struct r1bio *r1_bio)
397 if (!atomic_dec_and_test(&r1_bio->remaining))
400 if (test_bit(R1BIO_WriteError, &r1_bio->state))
401 reschedule_retry(r1_bio);
404 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
405 reschedule_retry(r1_bio);
407 raid_end_bio_io(r1_bio);
411 static void raid1_end_write_request(struct bio *bio)
413 struct r1bio *r1_bio = bio->bi_private;
414 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
415 struct r1conf *conf = r1_bio->mddev->private;
416 struct bio *to_put = NULL;
417 int mirror = find_bio_disk(r1_bio, bio);
418 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
421 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
424 * 'one mirror IO has finished' event handler:
426 if (bio->bi_status && !discard_error) {
427 set_bit(WriteErrorSeen, &rdev->flags);
428 if (!test_and_set_bit(WantReplacement, &rdev->flags))
429 set_bit(MD_RECOVERY_NEEDED, &
430 conf->mddev->recovery);
432 if (test_bit(FailFast, &rdev->flags) &&
433 (bio->bi_opf & MD_FAILFAST) &&
434 /* We never try FailFast to WriteMostly devices */
435 !test_bit(WriteMostly, &rdev->flags)) {
436 md_error(r1_bio->mddev, rdev);
440 * When the device is faulty, it is not necessary to
441 * handle write error.
442 * For failfast, this is the only remaining device,
443 * We need to retry the write without FailFast.
445 if (!test_bit(Faulty, &rdev->flags))
446 set_bit(R1BIO_WriteError, &r1_bio->state);
448 /* Fail the request */
449 set_bit(R1BIO_Degraded, &r1_bio->state);
450 /* Finished with this branch */
451 r1_bio->bios[mirror] = NULL;
456 * Set R1BIO_Uptodate in our master bio, so that we
457 * will return a good error code for to the higher
458 * levels even if IO on some other mirrored buffer
461 * The 'master' represents the composite IO operation
462 * to user-side. So if something waits for IO, then it
463 * will wait for the 'master' bio.
468 r1_bio->bios[mirror] = NULL;
471 * Do not set R1BIO_Uptodate if the current device is
472 * rebuilding or Faulty. This is because we cannot use
473 * such device for properly reading the data back (we could
474 * potentially use it, if the current write would have felt
475 * before rdev->recovery_offset, but for simplicity we don't
478 if (test_bit(In_sync, &rdev->flags) &&
479 !test_bit(Faulty, &rdev->flags))
480 set_bit(R1BIO_Uptodate, &r1_bio->state);
482 /* Maybe we can clear some bad blocks. */
483 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
484 &first_bad, &bad_sectors) && !discard_error) {
485 r1_bio->bios[mirror] = IO_MADE_GOOD;
486 set_bit(R1BIO_MadeGood, &r1_bio->state);
491 if (test_bit(WriteMostly, &rdev->flags))
492 atomic_dec(&r1_bio->behind_remaining);
495 * In behind mode, we ACK the master bio once the I/O
496 * has safely reached all non-writemostly
497 * disks. Setting the Returned bit ensures that this
498 * gets done only once -- we don't ever want to return
499 * -EIO here, instead we'll wait
501 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
502 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
503 /* Maybe we can return now */
504 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
505 struct bio *mbio = r1_bio->master_bio;
506 pr_debug("raid1: behind end write sectors"
508 (unsigned long long) mbio->bi_iter.bi_sector,
509 (unsigned long long) bio_end_sector(mbio) - 1);
510 call_bio_endio(r1_bio);
514 if (r1_bio->bios[mirror] == NULL)
515 rdev_dec_pending(rdev, conf->mddev);
518 * Let's see if all mirrored write operations have finished
521 r1_bio_write_done(r1_bio);
527 static sector_t align_to_barrier_unit_end(sector_t start_sector,
532 WARN_ON(sectors == 0);
534 * len is the number of sectors from start_sector to end of the
535 * barrier unit which start_sector belongs to.
537 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
547 * This routine returns the disk from which the requested read should
548 * be done. There is a per-array 'next expected sequential IO' sector
549 * number - if this matches on the next IO then we use the last disk.
550 * There is also a per-disk 'last know head position' sector that is
551 * maintained from IRQ contexts, both the normal and the resync IO
552 * completion handlers update this position correctly. If there is no
553 * perfect sequential match then we pick the disk whose head is closest.
555 * If there are 2 mirrors in the same 2 devices, performance degrades
556 * because position is mirror, not device based.
558 * The rdev for the device selected will have nr_pending incremented.
560 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
562 const sector_t this_sector = r1_bio->sector;
564 int best_good_sectors;
565 int best_disk, best_dist_disk, best_pending_disk;
569 unsigned int min_pending;
570 struct md_rdev *rdev;
572 int choose_next_idle;
576 * Check if we can balance. We can balance on the whole
577 * device if no resync is going on, or below the resync window.
578 * We take the first readable disk when above the resync window.
581 sectors = r1_bio->sectors;
584 best_dist = MaxSector;
585 best_pending_disk = -1;
586 min_pending = UINT_MAX;
587 best_good_sectors = 0;
589 choose_next_idle = 0;
590 clear_bit(R1BIO_FailFast, &r1_bio->state);
592 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
593 (mddev_is_clustered(conf->mddev) &&
594 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
595 this_sector + sectors)))
600 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
604 unsigned int pending;
607 rdev = rcu_dereference(conf->mirrors[disk].rdev);
608 if (r1_bio->bios[disk] == IO_BLOCKED
610 || test_bit(Faulty, &rdev->flags))
612 if (!test_bit(In_sync, &rdev->flags) &&
613 rdev->recovery_offset < this_sector + sectors)
615 if (test_bit(WriteMostly, &rdev->flags)) {
616 /* Don't balance among write-mostly, just
617 * use the first as a last resort */
618 if (best_dist_disk < 0) {
619 if (is_badblock(rdev, this_sector, sectors,
620 &first_bad, &bad_sectors)) {
621 if (first_bad <= this_sector)
622 /* Cannot use this */
624 best_good_sectors = first_bad - this_sector;
626 best_good_sectors = sectors;
627 best_dist_disk = disk;
628 best_pending_disk = disk;
632 /* This is a reasonable device to use. It might
635 if (is_badblock(rdev, this_sector, sectors,
636 &first_bad, &bad_sectors)) {
637 if (best_dist < MaxSector)
638 /* already have a better device */
640 if (first_bad <= this_sector) {
641 /* cannot read here. If this is the 'primary'
642 * device, then we must not read beyond
643 * bad_sectors from another device..
645 bad_sectors -= (this_sector - first_bad);
646 if (choose_first && sectors > bad_sectors)
647 sectors = bad_sectors;
648 if (best_good_sectors > sectors)
649 best_good_sectors = sectors;
652 sector_t good_sectors = first_bad - this_sector;
653 if (good_sectors > best_good_sectors) {
654 best_good_sectors = good_sectors;
662 if ((sectors > best_good_sectors) && (best_disk >= 0))
664 best_good_sectors = sectors;
668 /* At least two disks to choose from so failfast is OK */
669 set_bit(R1BIO_FailFast, &r1_bio->state);
671 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
672 has_nonrot_disk |= nonrot;
673 pending = atomic_read(&rdev->nr_pending);
674 dist = abs(this_sector - conf->mirrors[disk].head_position);
679 /* Don't change to another disk for sequential reads */
680 if (conf->mirrors[disk].next_seq_sect == this_sector
682 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
683 struct raid1_info *mirror = &conf->mirrors[disk];
687 * If buffered sequential IO size exceeds optimal
688 * iosize, check if there is idle disk. If yes, choose
689 * the idle disk. read_balance could already choose an
690 * idle disk before noticing it's a sequential IO in
691 * this disk. This doesn't matter because this disk
692 * will idle, next time it will be utilized after the
693 * first disk has IO size exceeds optimal iosize. In
694 * this way, iosize of the first disk will be optimal
695 * iosize at least. iosize of the second disk might be
696 * small, but not a big deal since when the second disk
697 * starts IO, the first disk is likely still busy.
699 if (nonrot && opt_iosize > 0 &&
700 mirror->seq_start != MaxSector &&
701 mirror->next_seq_sect > opt_iosize &&
702 mirror->next_seq_sect - opt_iosize >=
704 choose_next_idle = 1;
710 if (choose_next_idle)
713 if (min_pending > pending) {
714 min_pending = pending;
715 best_pending_disk = disk;
718 if (dist < best_dist) {
720 best_dist_disk = disk;
725 * If all disks are rotational, choose the closest disk. If any disk is
726 * non-rotational, choose the disk with less pending request even the
727 * disk is rotational, which might/might not be optimal for raids with
728 * mixed ratation/non-rotational disks depending on workload.
730 if (best_disk == -1) {
731 if (has_nonrot_disk || min_pending == 0)
732 best_disk = best_pending_disk;
734 best_disk = best_dist_disk;
737 if (best_disk >= 0) {
738 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
741 atomic_inc(&rdev->nr_pending);
742 sectors = best_good_sectors;
744 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
745 conf->mirrors[best_disk].seq_start = this_sector;
747 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
750 *max_sectors = sectors;
755 static int raid1_congested(struct mddev *mddev, int bits)
757 struct r1conf *conf = mddev->private;
760 if ((bits & (1 << WB_async_congested)) &&
761 conf->pending_count >= max_queued_requests)
765 for (i = 0; i < conf->raid_disks * 2; i++) {
766 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
767 if (rdev && !test_bit(Faulty, &rdev->flags)) {
768 struct request_queue *q = bdev_get_queue(rdev->bdev);
772 /* Note the '|| 1' - when read_balance prefers
773 * non-congested targets, it can be removed
775 if ((bits & (1 << WB_async_congested)) || 1)
776 ret |= bdi_congested(q->backing_dev_info, bits);
778 ret &= bdi_congested(q->backing_dev_info, bits);
785 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
787 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
788 md_bitmap_unplug(conf->mddev->bitmap);
789 wake_up(&conf->wait_barrier);
791 while (bio) { /* submit pending writes */
792 struct bio *next = bio->bi_next;
793 struct md_rdev *rdev = (void *)bio->bi_disk;
795 bio_set_dev(bio, rdev->bdev);
796 if (test_bit(Faulty, &rdev->flags)) {
798 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
799 !blk_queue_discard(bio->bi_disk->queue)))
803 generic_make_request(bio);
808 static void flush_pending_writes(struct r1conf *conf)
810 /* Any writes that have been queued but are awaiting
811 * bitmap updates get flushed here.
813 spin_lock_irq(&conf->device_lock);
815 if (conf->pending_bio_list.head) {
816 struct blk_plug plug;
819 bio = bio_list_get(&conf->pending_bio_list);
820 conf->pending_count = 0;
821 spin_unlock_irq(&conf->device_lock);
824 * As this is called in a wait_event() loop (see freeze_array),
825 * current->state might be TASK_UNINTERRUPTIBLE which will
826 * cause a warning when we prepare to wait again. As it is
827 * rare that this path is taken, it is perfectly safe to force
828 * us to go around the wait_event() loop again, so the warning
829 * is a false-positive. Silence the warning by resetting
832 __set_current_state(TASK_RUNNING);
833 blk_start_plug(&plug);
834 flush_bio_list(conf, bio);
835 blk_finish_plug(&plug);
837 spin_unlock_irq(&conf->device_lock);
841 * Sometimes we need to suspend IO while we do something else,
842 * either some resync/recovery, or reconfigure the array.
843 * To do this we raise a 'barrier'.
844 * The 'barrier' is a counter that can be raised multiple times
845 * to count how many activities are happening which preclude
847 * We can only raise the barrier if there is no pending IO.
848 * i.e. if nr_pending == 0.
849 * We choose only to raise the barrier if no-one is waiting for the
850 * barrier to go down. This means that as soon as an IO request
851 * is ready, no other operations which require a barrier will start
852 * until the IO request has had a chance.
854 * So: regular IO calls 'wait_barrier'. When that returns there
855 * is no backgroup IO happening, It must arrange to call
856 * allow_barrier when it has finished its IO.
857 * backgroup IO calls must call raise_barrier. Once that returns
858 * there is no normal IO happeing. It must arrange to call
859 * lower_barrier when the particular background IO completes.
861 static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
863 int idx = sector_to_idx(sector_nr);
865 spin_lock_irq(&conf->resync_lock);
867 /* Wait until no block IO is waiting */
868 wait_event_lock_irq(conf->wait_barrier,
869 !atomic_read(&conf->nr_waiting[idx]),
872 /* block any new IO from starting */
873 atomic_inc(&conf->barrier[idx]);
875 * In raise_barrier() we firstly increase conf->barrier[idx] then
876 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
877 * increase conf->nr_pending[idx] then check conf->barrier[idx].
878 * A memory barrier here to make sure conf->nr_pending[idx] won't
879 * be fetched before conf->barrier[idx] is increased. Otherwise
880 * there will be a race between raise_barrier() and _wait_barrier().
882 smp_mb__after_atomic();
884 /* For these conditions we must wait:
885 * A: while the array is in frozen state
886 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
887 * existing in corresponding I/O barrier bucket.
888 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
889 * max resync count which allowed on current I/O barrier bucket.
891 wait_event_lock_irq(conf->wait_barrier,
892 (!conf->array_frozen &&
893 !atomic_read(&conf->nr_pending[idx]) &&
894 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
895 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
898 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
899 atomic_dec(&conf->barrier[idx]);
900 spin_unlock_irq(&conf->resync_lock);
901 wake_up(&conf->wait_barrier);
905 atomic_inc(&conf->nr_sync_pending);
906 spin_unlock_irq(&conf->resync_lock);
911 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
913 int idx = sector_to_idx(sector_nr);
915 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
917 atomic_dec(&conf->barrier[idx]);
918 atomic_dec(&conf->nr_sync_pending);
919 wake_up(&conf->wait_barrier);
922 static void _wait_barrier(struct r1conf *conf, int idx)
925 * We need to increase conf->nr_pending[idx] very early here,
926 * then raise_barrier() can be blocked when it waits for
927 * conf->nr_pending[idx] to be 0. Then we can avoid holding
928 * conf->resync_lock when there is no barrier raised in same
929 * barrier unit bucket. Also if the array is frozen, I/O
930 * should be blocked until array is unfrozen.
932 atomic_inc(&conf->nr_pending[idx]);
934 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
935 * check conf->barrier[idx]. In raise_barrier() we firstly increase
936 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
937 * barrier is necessary here to make sure conf->barrier[idx] won't be
938 * fetched before conf->nr_pending[idx] is increased. Otherwise there
939 * will be a race between _wait_barrier() and raise_barrier().
941 smp_mb__after_atomic();
944 * Don't worry about checking two atomic_t variables at same time
945 * here. If during we check conf->barrier[idx], the array is
946 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
947 * 0, it is safe to return and make the I/O continue. Because the
948 * array is frozen, all I/O returned here will eventually complete
949 * or be queued, no race will happen. See code comment in
952 if (!READ_ONCE(conf->array_frozen) &&
953 !atomic_read(&conf->barrier[idx]))
957 * After holding conf->resync_lock, conf->nr_pending[idx]
958 * should be decreased before waiting for barrier to drop.
959 * Otherwise, we may encounter a race condition because
960 * raise_barrer() might be waiting for conf->nr_pending[idx]
961 * to be 0 at same time.
963 spin_lock_irq(&conf->resync_lock);
964 atomic_inc(&conf->nr_waiting[idx]);
965 atomic_dec(&conf->nr_pending[idx]);
967 * In case freeze_array() is waiting for
968 * get_unqueued_pending() == extra
970 wake_up(&conf->wait_barrier);
971 /* Wait for the barrier in same barrier unit bucket to drop. */
972 wait_event_lock_irq(conf->wait_barrier,
973 !conf->array_frozen &&
974 !atomic_read(&conf->barrier[idx]),
976 atomic_inc(&conf->nr_pending[idx]);
977 atomic_dec(&conf->nr_waiting[idx]);
978 spin_unlock_irq(&conf->resync_lock);
981 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
983 int idx = sector_to_idx(sector_nr);
986 * Very similar to _wait_barrier(). The difference is, for read
987 * I/O we don't need wait for sync I/O, but if the whole array
988 * is frozen, the read I/O still has to wait until the array is
989 * unfrozen. Since there is no ordering requirement with
990 * conf->barrier[idx] here, memory barrier is unnecessary as well.
992 atomic_inc(&conf->nr_pending[idx]);
994 if (!READ_ONCE(conf->array_frozen))
997 spin_lock_irq(&conf->resync_lock);
998 atomic_inc(&conf->nr_waiting[idx]);
999 atomic_dec(&conf->nr_pending[idx]);
1001 * In case freeze_array() is waiting for
1002 * get_unqueued_pending() == extra
1004 wake_up(&conf->wait_barrier);
1005 /* Wait for array to be unfrozen */
1006 wait_event_lock_irq(conf->wait_barrier,
1007 !conf->array_frozen,
1009 atomic_inc(&conf->nr_pending[idx]);
1010 atomic_dec(&conf->nr_waiting[idx]);
1011 spin_unlock_irq(&conf->resync_lock);
1014 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1016 int idx = sector_to_idx(sector_nr);
1018 _wait_barrier(conf, idx);
1021 static void _allow_barrier(struct r1conf *conf, int idx)
1023 atomic_dec(&conf->nr_pending[idx]);
1024 wake_up(&conf->wait_barrier);
1027 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1029 int idx = sector_to_idx(sector_nr);
1031 _allow_barrier(conf, idx);
1034 /* conf->resync_lock should be held */
1035 static int get_unqueued_pending(struct r1conf *conf)
1039 ret = atomic_read(&conf->nr_sync_pending);
1040 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1041 ret += atomic_read(&conf->nr_pending[idx]) -
1042 atomic_read(&conf->nr_queued[idx]);
1047 static void freeze_array(struct r1conf *conf, int extra)
1049 /* Stop sync I/O and normal I/O and wait for everything to
1051 * This is called in two situations:
1052 * 1) management command handlers (reshape, remove disk, quiesce).
1053 * 2) one normal I/O request failed.
1055 * After array_frozen is set to 1, new sync IO will be blocked at
1056 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1057 * or wait_read_barrier(). The flying I/Os will either complete or be
1058 * queued. When everything goes quite, there are only queued I/Os left.
1060 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1061 * barrier bucket index which this I/O request hits. When all sync and
1062 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1063 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1064 * in handle_read_error(), we may call freeze_array() before trying to
1065 * fix the read error. In this case, the error read I/O is not queued,
1066 * so get_unqueued_pending() == 1.
1068 * Therefore before this function returns, we need to wait until
1069 * get_unqueued_pendings(conf) gets equal to extra. For
1070 * normal I/O context, extra is 1, in rested situations extra is 0.
1072 spin_lock_irq(&conf->resync_lock);
1073 conf->array_frozen = 1;
1074 raid1_log(conf->mddev, "wait freeze");
1075 wait_event_lock_irq_cmd(
1077 get_unqueued_pending(conf) == extra,
1079 flush_pending_writes(conf));
1080 spin_unlock_irq(&conf->resync_lock);
1082 static void unfreeze_array(struct r1conf *conf)
1084 /* reverse the effect of the freeze */
1085 spin_lock_irq(&conf->resync_lock);
1086 conf->array_frozen = 0;
1087 spin_unlock_irq(&conf->resync_lock);
1088 wake_up(&conf->wait_barrier);
1091 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1094 int size = bio->bi_iter.bi_size;
1095 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1097 struct bio *behind_bio = NULL;
1099 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1103 /* discard op, we don't support writezero/writesame yet */
1104 if (!bio_has_data(bio)) {
1105 behind_bio->bi_iter.bi_size = size;
1109 behind_bio->bi_write_hint = bio->bi_write_hint;
1111 while (i < vcnt && size) {
1113 int len = min_t(int, PAGE_SIZE, size);
1115 page = alloc_page(GFP_NOIO);
1116 if (unlikely(!page))
1119 bio_add_page(behind_bio, page, len, 0);
1125 bio_copy_data(behind_bio, bio);
1127 r1_bio->behind_master_bio = behind_bio;
1128 set_bit(R1BIO_BehindIO, &r1_bio->state);
1133 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1134 bio->bi_iter.bi_size);
1135 bio_free_pages(behind_bio);
1136 bio_put(behind_bio);
1139 struct raid1_plug_cb {
1140 struct blk_plug_cb cb;
1141 struct bio_list pending;
1145 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1147 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1149 struct mddev *mddev = plug->cb.data;
1150 struct r1conf *conf = mddev->private;
1153 if (from_schedule || current->bio_list) {
1154 spin_lock_irq(&conf->device_lock);
1155 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1156 conf->pending_count += plug->pending_cnt;
1157 spin_unlock_irq(&conf->device_lock);
1158 wake_up(&conf->wait_barrier);
1159 md_wakeup_thread(mddev->thread);
1164 /* we aren't scheduling, so we can do the write-out directly. */
1165 bio = bio_list_get(&plug->pending);
1166 flush_bio_list(conf, bio);
1170 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1172 r1_bio->master_bio = bio;
1173 r1_bio->sectors = bio_sectors(bio);
1175 r1_bio->mddev = mddev;
1176 r1_bio->sector = bio->bi_iter.bi_sector;
1179 static inline struct r1bio *
1180 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1182 struct r1conf *conf = mddev->private;
1183 struct r1bio *r1_bio;
1185 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1186 /* Ensure no bio records IO_BLOCKED */
1187 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1188 init_r1bio(r1_bio, mddev, bio);
1192 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1193 int max_read_sectors, struct r1bio *r1_bio)
1195 struct r1conf *conf = mddev->private;
1196 struct raid1_info *mirror;
1197 struct bio *read_bio;
1198 struct bitmap *bitmap = mddev->bitmap;
1199 const int op = bio_op(bio);
1200 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1203 bool print_msg = !!r1_bio;
1204 char b[BDEVNAME_SIZE];
1207 * If r1_bio is set, we are blocking the raid1d thread
1208 * so there is a tiny risk of deadlock. So ask for
1209 * emergency memory if needed.
1211 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1214 /* Need to get the block device name carefully */
1215 struct md_rdev *rdev;
1217 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1219 bdevname(rdev->bdev, b);
1226 * Still need barrier for READ in case that whole
1229 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1232 r1_bio = alloc_r1bio(mddev, bio);
1234 init_r1bio(r1_bio, mddev, bio);
1235 r1_bio->sectors = max_read_sectors;
1238 * make_request() can abort the operation when read-ahead is being
1239 * used and no empty request is available.
1241 rdisk = read_balance(conf, r1_bio, &max_sectors);
1244 /* couldn't find anywhere to read from */
1246 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1249 (unsigned long long)r1_bio->sector);
1251 raid_end_bio_io(r1_bio);
1254 mirror = conf->mirrors + rdisk;
1257 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1259 (unsigned long long)r1_bio->sector,
1260 bdevname(mirror->rdev->bdev, b));
1262 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1265 * Reading from a write-mostly device must take care not to
1266 * over-take any writes that are 'behind'
1268 raid1_log(mddev, "wait behind writes");
1269 wait_event(bitmap->behind_wait,
1270 atomic_read(&bitmap->behind_writes) == 0);
1273 if (max_sectors < bio_sectors(bio)) {
1274 struct bio *split = bio_split(bio, max_sectors,
1275 gfp, &conf->bio_split);
1276 bio_chain(split, bio);
1277 generic_make_request(bio);
1279 r1_bio->master_bio = bio;
1280 r1_bio->sectors = max_sectors;
1283 r1_bio->read_disk = rdisk;
1285 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1287 r1_bio->bios[rdisk] = read_bio;
1289 read_bio->bi_iter.bi_sector = r1_bio->sector +
1290 mirror->rdev->data_offset;
1291 bio_set_dev(read_bio, mirror->rdev->bdev);
1292 read_bio->bi_end_io = raid1_end_read_request;
1293 bio_set_op_attrs(read_bio, op, do_sync);
1294 if (test_bit(FailFast, &mirror->rdev->flags) &&
1295 test_bit(R1BIO_FailFast, &r1_bio->state))
1296 read_bio->bi_opf |= MD_FAILFAST;
1297 read_bio->bi_private = r1_bio;
1300 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1301 disk_devt(mddev->gendisk), r1_bio->sector);
1303 generic_make_request(read_bio);
1306 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1307 int max_write_sectors)
1309 struct r1conf *conf = mddev->private;
1310 struct r1bio *r1_bio;
1312 struct bitmap *bitmap = mddev->bitmap;
1313 unsigned long flags;
1314 struct md_rdev *blocked_rdev;
1315 struct blk_plug_cb *cb;
1316 struct raid1_plug_cb *plug = NULL;
1320 if (mddev_is_clustered(mddev) &&
1321 md_cluster_ops->area_resyncing(mddev, WRITE,
1322 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1326 prepare_to_wait(&conf->wait_barrier,
1328 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1329 bio->bi_iter.bi_sector,
1330 bio_end_sector(bio)))
1334 finish_wait(&conf->wait_barrier, &w);
1338 * Register the new request and wait if the reconstruction
1339 * thread has put up a bar for new requests.
1340 * Continue immediately if no resync is active currently.
1342 wait_barrier(conf, bio->bi_iter.bi_sector);
1344 r1_bio = alloc_r1bio(mddev, bio);
1345 r1_bio->sectors = max_write_sectors;
1347 if (conf->pending_count >= max_queued_requests) {
1348 md_wakeup_thread(mddev->thread);
1349 raid1_log(mddev, "wait queued");
1350 wait_event(conf->wait_barrier,
1351 conf->pending_count < max_queued_requests);
1353 /* first select target devices under rcu_lock and
1354 * inc refcount on their rdev. Record them by setting
1356 * If there are known/acknowledged bad blocks on any device on
1357 * which we have seen a write error, we want to avoid writing those
1359 * This potentially requires several writes to write around
1360 * the bad blocks. Each set of writes gets it's own r1bio
1361 * with a set of bios attached.
1364 disks = conf->raid_disks * 2;
1366 blocked_rdev = NULL;
1368 max_sectors = r1_bio->sectors;
1369 for (i = 0; i < disks; i++) {
1370 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1371 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1372 atomic_inc(&rdev->nr_pending);
1373 blocked_rdev = rdev;
1376 r1_bio->bios[i] = NULL;
1377 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1378 if (i < conf->raid_disks)
1379 set_bit(R1BIO_Degraded, &r1_bio->state);
1383 atomic_inc(&rdev->nr_pending);
1384 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1389 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1390 &first_bad, &bad_sectors);
1392 /* mustn't write here until the bad block is
1394 set_bit(BlockedBadBlocks, &rdev->flags);
1395 blocked_rdev = rdev;
1398 if (is_bad && first_bad <= r1_bio->sector) {
1399 /* Cannot write here at all */
1400 bad_sectors -= (r1_bio->sector - first_bad);
1401 if (bad_sectors < max_sectors)
1402 /* mustn't write more than bad_sectors
1403 * to other devices yet
1405 max_sectors = bad_sectors;
1406 rdev_dec_pending(rdev, mddev);
1407 /* We don't set R1BIO_Degraded as that
1408 * only applies if the disk is
1409 * missing, so it might be re-added,
1410 * and we want to know to recover this
1412 * In this case the device is here,
1413 * and the fact that this chunk is not
1414 * in-sync is recorded in the bad
1420 int good_sectors = first_bad - r1_bio->sector;
1421 if (good_sectors < max_sectors)
1422 max_sectors = good_sectors;
1425 r1_bio->bios[i] = bio;
1429 if (unlikely(blocked_rdev)) {
1430 /* Wait for this device to become unblocked */
1433 for (j = 0; j < i; j++)
1434 if (r1_bio->bios[j])
1435 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1437 allow_barrier(conf, bio->bi_iter.bi_sector);
1438 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1439 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1440 wait_barrier(conf, bio->bi_iter.bi_sector);
1444 if (max_sectors < bio_sectors(bio)) {
1445 struct bio *split = bio_split(bio, max_sectors,
1446 GFP_NOIO, &conf->bio_split);
1447 bio_chain(split, bio);
1448 generic_make_request(bio);
1450 r1_bio->master_bio = bio;
1451 r1_bio->sectors = max_sectors;
1454 atomic_set(&r1_bio->remaining, 1);
1455 atomic_set(&r1_bio->behind_remaining, 0);
1459 for (i = 0; i < disks; i++) {
1460 struct bio *mbio = NULL;
1461 if (!r1_bio->bios[i])
1467 * Not if there are too many, or cannot
1468 * allocate memory, or a reader on WriteMostly
1469 * is waiting for behind writes to flush */
1471 (atomic_read(&bitmap->behind_writes)
1472 < mddev->bitmap_info.max_write_behind) &&
1473 !waitqueue_active(&bitmap->behind_wait)) {
1474 alloc_behind_master_bio(r1_bio, bio);
1477 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1478 test_bit(R1BIO_BehindIO, &r1_bio->state));
1482 if (r1_bio->behind_master_bio)
1483 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1484 GFP_NOIO, &mddev->bio_set);
1486 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1488 if (r1_bio->behind_master_bio) {
1489 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1490 atomic_inc(&r1_bio->behind_remaining);
1493 r1_bio->bios[i] = mbio;
1495 mbio->bi_iter.bi_sector = (r1_bio->sector +
1496 conf->mirrors[i].rdev->data_offset);
1497 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1498 mbio->bi_end_io = raid1_end_write_request;
1499 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1500 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1501 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1502 conf->raid_disks - mddev->degraded > 1)
1503 mbio->bi_opf |= MD_FAILFAST;
1504 mbio->bi_private = r1_bio;
1506 atomic_inc(&r1_bio->remaining);
1509 trace_block_bio_remap(mbio->bi_disk->queue,
1510 mbio, disk_devt(mddev->gendisk),
1512 /* flush_pending_writes() needs access to the rdev so...*/
1513 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1515 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1517 plug = container_of(cb, struct raid1_plug_cb, cb);
1521 bio_list_add(&plug->pending, mbio);
1522 plug->pending_cnt++;
1524 spin_lock_irqsave(&conf->device_lock, flags);
1525 bio_list_add(&conf->pending_bio_list, mbio);
1526 conf->pending_count++;
1527 spin_unlock_irqrestore(&conf->device_lock, flags);
1528 md_wakeup_thread(mddev->thread);
1532 r1_bio_write_done(r1_bio);
1534 /* In case raid1d snuck in to freeze_array */
1535 wake_up(&conf->wait_barrier);
1538 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1542 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1543 && md_flush_request(mddev, bio))
1547 * There is a limit to the maximum size, but
1548 * the read/write handler might find a lower limit
1549 * due to bad blocks. To avoid multiple splits,
1550 * we pass the maximum number of sectors down
1551 * and let the lower level perform the split.
1553 sectors = align_to_barrier_unit_end(
1554 bio->bi_iter.bi_sector, bio_sectors(bio));
1556 if (bio_data_dir(bio) == READ)
1557 raid1_read_request(mddev, bio, sectors, NULL);
1559 if (!md_write_start(mddev,bio))
1561 raid1_write_request(mddev, bio, sectors);
1566 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1568 struct r1conf *conf = mddev->private;
1571 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1572 conf->raid_disks - mddev->degraded);
1574 for (i = 0; i < conf->raid_disks; i++) {
1575 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1576 seq_printf(seq, "%s",
1577 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1580 seq_printf(seq, "]");
1583 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1585 char b[BDEVNAME_SIZE];
1586 struct r1conf *conf = mddev->private;
1587 unsigned long flags;
1590 * If it is not operational, then we have already marked it as dead
1591 * else if it is the last working disks, ignore the error, let the
1592 * next level up know.
1593 * else mark the drive as failed
1595 spin_lock_irqsave(&conf->device_lock, flags);
1596 if (test_bit(In_sync, &rdev->flags)
1597 && (conf->raid_disks - mddev->degraded) == 1) {
1599 * Don't fail the drive, act as though we were just a
1600 * normal single drive.
1601 * However don't try a recovery from this drive as
1602 * it is very likely to fail.
1604 conf->recovery_disabled = mddev->recovery_disabled;
1605 spin_unlock_irqrestore(&conf->device_lock, flags);
1608 set_bit(Blocked, &rdev->flags);
1609 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1611 set_bit(Faulty, &rdev->flags);
1613 set_bit(Faulty, &rdev->flags);
1614 spin_unlock_irqrestore(&conf->device_lock, flags);
1616 * if recovery is running, make sure it aborts.
1618 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1619 set_mask_bits(&mddev->sb_flags, 0,
1620 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1621 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1622 "md/raid1:%s: Operation continuing on %d devices.\n",
1623 mdname(mddev), bdevname(rdev->bdev, b),
1624 mdname(mddev), conf->raid_disks - mddev->degraded);
1627 static void print_conf(struct r1conf *conf)
1631 pr_debug("RAID1 conf printout:\n");
1633 pr_debug("(!conf)\n");
1636 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1640 for (i = 0; i < conf->raid_disks; i++) {
1641 char b[BDEVNAME_SIZE];
1642 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1644 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1645 i, !test_bit(In_sync, &rdev->flags),
1646 !test_bit(Faulty, &rdev->flags),
1647 bdevname(rdev->bdev,b));
1652 static void close_sync(struct r1conf *conf)
1656 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1657 _wait_barrier(conf, idx);
1658 _allow_barrier(conf, idx);
1661 mempool_exit(&conf->r1buf_pool);
1664 static int raid1_spare_active(struct mddev *mddev)
1667 struct r1conf *conf = mddev->private;
1669 unsigned long flags;
1672 * Find all failed disks within the RAID1 configuration
1673 * and mark them readable.
1674 * Called under mddev lock, so rcu protection not needed.
1675 * device_lock used to avoid races with raid1_end_read_request
1676 * which expects 'In_sync' flags and ->degraded to be consistent.
1678 spin_lock_irqsave(&conf->device_lock, flags);
1679 for (i = 0; i < conf->raid_disks; i++) {
1680 struct md_rdev *rdev = conf->mirrors[i].rdev;
1681 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1683 && !test_bit(Candidate, &repl->flags)
1684 && repl->recovery_offset == MaxSector
1685 && !test_bit(Faulty, &repl->flags)
1686 && !test_and_set_bit(In_sync, &repl->flags)) {
1687 /* replacement has just become active */
1689 !test_and_clear_bit(In_sync, &rdev->flags))
1692 /* Replaced device not technically
1693 * faulty, but we need to be sure
1694 * it gets removed and never re-added
1696 set_bit(Faulty, &rdev->flags);
1697 sysfs_notify_dirent_safe(
1702 && rdev->recovery_offset == MaxSector
1703 && !test_bit(Faulty, &rdev->flags)
1704 && !test_and_set_bit(In_sync, &rdev->flags)) {
1706 sysfs_notify_dirent_safe(rdev->sysfs_state);
1709 mddev->degraded -= count;
1710 spin_unlock_irqrestore(&conf->device_lock, flags);
1716 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1718 struct r1conf *conf = mddev->private;
1721 struct raid1_info *p;
1723 int last = conf->raid_disks - 1;
1725 if (mddev->recovery_disabled == conf->recovery_disabled)
1728 if (md_integrity_add_rdev(rdev, mddev))
1731 if (rdev->raid_disk >= 0)
1732 first = last = rdev->raid_disk;
1735 * find the disk ... but prefer rdev->saved_raid_disk
1738 if (rdev->saved_raid_disk >= 0 &&
1739 rdev->saved_raid_disk >= first &&
1740 rdev->saved_raid_disk < conf->raid_disks &&
1741 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1742 first = last = rdev->saved_raid_disk;
1744 for (mirror = first; mirror <= last; mirror++) {
1745 p = conf->mirrors+mirror;
1749 disk_stack_limits(mddev->gendisk, rdev->bdev,
1750 rdev->data_offset << 9);
1752 p->head_position = 0;
1753 rdev->raid_disk = mirror;
1755 /* As all devices are equivalent, we don't need a full recovery
1756 * if this was recently any drive of the array
1758 if (rdev->saved_raid_disk < 0)
1760 rcu_assign_pointer(p->rdev, rdev);
1763 if (test_bit(WantReplacement, &p->rdev->flags) &&
1764 p[conf->raid_disks].rdev == NULL) {
1765 /* Add this device as a replacement */
1766 clear_bit(In_sync, &rdev->flags);
1767 set_bit(Replacement, &rdev->flags);
1768 rdev->raid_disk = mirror;
1771 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1775 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1776 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1781 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1783 struct r1conf *conf = mddev->private;
1785 int number = rdev->raid_disk;
1786 struct raid1_info *p = conf->mirrors + number;
1788 if (unlikely(number >= conf->raid_disks))
1791 if (rdev != p->rdev)
1792 p = conf->mirrors + conf->raid_disks + number;
1795 if (rdev == p->rdev) {
1796 if (test_bit(In_sync, &rdev->flags) ||
1797 atomic_read(&rdev->nr_pending)) {
1801 /* Only remove non-faulty devices if recovery
1804 if (!test_bit(Faulty, &rdev->flags) &&
1805 mddev->recovery_disabled != conf->recovery_disabled &&
1806 mddev->degraded < conf->raid_disks) {
1811 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1813 if (atomic_read(&rdev->nr_pending)) {
1814 /* lost the race, try later */
1820 if (conf->mirrors[conf->raid_disks + number].rdev) {
1821 /* We just removed a device that is being replaced.
1822 * Move down the replacement. We drain all IO before
1823 * doing this to avoid confusion.
1825 struct md_rdev *repl =
1826 conf->mirrors[conf->raid_disks + number].rdev;
1827 freeze_array(conf, 0);
1828 if (atomic_read(&repl->nr_pending)) {
1829 /* It means that some queued IO of retry_list
1830 * hold repl. Thus, we cannot set replacement
1831 * as NULL, avoiding rdev NULL pointer
1832 * dereference in sync_request_write and
1833 * handle_write_finished.
1836 unfreeze_array(conf);
1839 clear_bit(Replacement, &repl->flags);
1841 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1842 unfreeze_array(conf);
1845 clear_bit(WantReplacement, &rdev->flags);
1846 err = md_integrity_register(mddev);
1854 static void end_sync_read(struct bio *bio)
1856 struct r1bio *r1_bio = get_resync_r1bio(bio);
1858 update_head_pos(r1_bio->read_disk, r1_bio);
1861 * we have read a block, now it needs to be re-written,
1862 * or re-read if the read failed.
1863 * We don't do much here, just schedule handling by raid1d
1865 if (!bio->bi_status)
1866 set_bit(R1BIO_Uptodate, &r1_bio->state);
1868 if (atomic_dec_and_test(&r1_bio->remaining))
1869 reschedule_retry(r1_bio);
1872 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1874 sector_t sync_blocks = 0;
1875 sector_t s = r1_bio->sector;
1876 long sectors_to_go = r1_bio->sectors;
1878 /* make sure these bits don't get cleared. */
1880 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1882 sectors_to_go -= sync_blocks;
1883 } while (sectors_to_go > 0);
1886 static void end_sync_write(struct bio *bio)
1888 int uptodate = !bio->bi_status;
1889 struct r1bio *r1_bio = get_resync_r1bio(bio);
1890 struct mddev *mddev = r1_bio->mddev;
1891 struct r1conf *conf = mddev->private;
1894 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1897 abort_sync_write(mddev, r1_bio);
1898 set_bit(WriteErrorSeen, &rdev->flags);
1899 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1900 set_bit(MD_RECOVERY_NEEDED, &
1902 set_bit(R1BIO_WriteError, &r1_bio->state);
1903 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1904 &first_bad, &bad_sectors) &&
1905 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1908 &first_bad, &bad_sectors)
1910 set_bit(R1BIO_MadeGood, &r1_bio->state);
1912 if (atomic_dec_and_test(&r1_bio->remaining)) {
1913 int s = r1_bio->sectors;
1914 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1915 test_bit(R1BIO_WriteError, &r1_bio->state))
1916 reschedule_retry(r1_bio);
1919 md_done_sync(mddev, s, uptodate);
1924 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1925 int sectors, struct page *page, int rw)
1927 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1931 set_bit(WriteErrorSeen, &rdev->flags);
1932 if (!test_and_set_bit(WantReplacement,
1934 set_bit(MD_RECOVERY_NEEDED, &
1935 rdev->mddev->recovery);
1937 /* need to record an error - either for the block or the device */
1938 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1939 md_error(rdev->mddev, rdev);
1943 static int fix_sync_read_error(struct r1bio *r1_bio)
1945 /* Try some synchronous reads of other devices to get
1946 * good data, much like with normal read errors. Only
1947 * read into the pages we already have so we don't
1948 * need to re-issue the read request.
1949 * We don't need to freeze the array, because being in an
1950 * active sync request, there is no normal IO, and
1951 * no overlapping syncs.
1952 * We don't need to check is_badblock() again as we
1953 * made sure that anything with a bad block in range
1954 * will have bi_end_io clear.
1956 struct mddev *mddev = r1_bio->mddev;
1957 struct r1conf *conf = mddev->private;
1958 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1959 struct page **pages = get_resync_pages(bio)->pages;
1960 sector_t sect = r1_bio->sector;
1961 int sectors = r1_bio->sectors;
1963 struct md_rdev *rdev;
1965 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1966 if (test_bit(FailFast, &rdev->flags)) {
1967 /* Don't try recovering from here - just fail it
1968 * ... unless it is the last working device of course */
1969 md_error(mddev, rdev);
1970 if (test_bit(Faulty, &rdev->flags))
1971 /* Don't try to read from here, but make sure
1972 * put_buf does it's thing
1974 bio->bi_end_io = end_sync_write;
1979 int d = r1_bio->read_disk;
1983 if (s > (PAGE_SIZE>>9))
1986 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1987 /* No rcu protection needed here devices
1988 * can only be removed when no resync is
1989 * active, and resync is currently active
1991 rdev = conf->mirrors[d].rdev;
1992 if (sync_page_io(rdev, sect, s<<9,
1994 REQ_OP_READ, 0, false)) {
2000 if (d == conf->raid_disks * 2)
2002 } while (!success && d != r1_bio->read_disk);
2005 char b[BDEVNAME_SIZE];
2007 /* Cannot read from anywhere, this block is lost.
2008 * Record a bad block on each device. If that doesn't
2009 * work just disable and interrupt the recovery.
2010 * Don't fail devices as that won't really help.
2012 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2013 mdname(mddev), bio_devname(bio, b),
2014 (unsigned long long)r1_bio->sector);
2015 for (d = 0; d < conf->raid_disks * 2; d++) {
2016 rdev = conf->mirrors[d].rdev;
2017 if (!rdev || test_bit(Faulty, &rdev->flags))
2019 if (!rdev_set_badblocks(rdev, sect, s, 0))
2023 conf->recovery_disabled =
2024 mddev->recovery_disabled;
2025 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2026 md_done_sync(mddev, r1_bio->sectors, 0);
2038 /* write it back and re-read */
2039 while (d != r1_bio->read_disk) {
2041 d = conf->raid_disks * 2;
2043 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2045 rdev = conf->mirrors[d].rdev;
2046 if (r1_sync_page_io(rdev, sect, s,
2049 r1_bio->bios[d]->bi_end_io = NULL;
2050 rdev_dec_pending(rdev, mddev);
2054 while (d != r1_bio->read_disk) {
2056 d = conf->raid_disks * 2;
2058 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2060 rdev = conf->mirrors[d].rdev;
2061 if (r1_sync_page_io(rdev, sect, s,
2064 atomic_add(s, &rdev->corrected_errors);
2070 set_bit(R1BIO_Uptodate, &r1_bio->state);
2075 static void process_checks(struct r1bio *r1_bio)
2077 /* We have read all readable devices. If we haven't
2078 * got the block, then there is no hope left.
2079 * If we have, then we want to do a comparison
2080 * and skip the write if everything is the same.
2081 * If any blocks failed to read, then we need to
2082 * attempt an over-write
2084 struct mddev *mddev = r1_bio->mddev;
2085 struct r1conf *conf = mddev->private;
2090 /* Fix variable parts of all bios */
2091 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2092 for (i = 0; i < conf->raid_disks * 2; i++) {
2093 blk_status_t status;
2094 struct bio *b = r1_bio->bios[i];
2095 struct resync_pages *rp = get_resync_pages(b);
2096 if (b->bi_end_io != end_sync_read)
2098 /* fixup the bio for reuse, but preserve errno */
2099 status = b->bi_status;
2101 b->bi_status = status;
2102 b->bi_iter.bi_sector = r1_bio->sector +
2103 conf->mirrors[i].rdev->data_offset;
2104 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2105 b->bi_end_io = end_sync_read;
2106 rp->raid_bio = r1_bio;
2109 /* initialize bvec table again */
2110 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2112 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2113 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2114 !r1_bio->bios[primary]->bi_status) {
2115 r1_bio->bios[primary]->bi_end_io = NULL;
2116 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2119 r1_bio->read_disk = primary;
2120 for (i = 0; i < conf->raid_disks * 2; i++) {
2122 struct bio *pbio = r1_bio->bios[primary];
2123 struct bio *sbio = r1_bio->bios[i];
2124 blk_status_t status = sbio->bi_status;
2125 struct page **ppages = get_resync_pages(pbio)->pages;
2126 struct page **spages = get_resync_pages(sbio)->pages;
2128 int page_len[RESYNC_PAGES] = { 0 };
2130 if (sbio->bi_end_io != end_sync_read)
2132 /* Now we can 'fixup' the error value */
2133 sbio->bi_status = 0;
2135 bio_for_each_segment_all(bi, sbio, j)
2136 page_len[j] = bi->bv_len;
2139 for (j = vcnt; j-- ; ) {
2140 if (memcmp(page_address(ppages[j]),
2141 page_address(spages[j]),
2148 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2149 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2151 /* No need to write to this device. */
2152 sbio->bi_end_io = NULL;
2153 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2157 bio_copy_data(sbio, pbio);
2161 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2163 struct r1conf *conf = mddev->private;
2165 int disks = conf->raid_disks * 2;
2168 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2169 /* ouch - failed to read all of that. */
2170 if (!fix_sync_read_error(r1_bio))
2173 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2174 process_checks(r1_bio);
2179 atomic_set(&r1_bio->remaining, 1);
2180 for (i = 0; i < disks ; i++) {
2181 wbio = r1_bio->bios[i];
2182 if (wbio->bi_end_io == NULL ||
2183 (wbio->bi_end_io == end_sync_read &&
2184 (i == r1_bio->read_disk ||
2185 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2187 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2188 abort_sync_write(mddev, r1_bio);
2192 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2193 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2194 wbio->bi_opf |= MD_FAILFAST;
2196 wbio->bi_end_io = end_sync_write;
2197 atomic_inc(&r1_bio->remaining);
2198 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2200 generic_make_request(wbio);
2203 if (atomic_dec_and_test(&r1_bio->remaining)) {
2204 /* if we're here, all write(s) have completed, so clean up */
2205 int s = r1_bio->sectors;
2206 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2207 test_bit(R1BIO_WriteError, &r1_bio->state))
2208 reschedule_retry(r1_bio);
2211 md_done_sync(mddev, s, 1);
2217 * This is a kernel thread which:
2219 * 1. Retries failed read operations on working mirrors.
2220 * 2. Updates the raid superblock when problems encounter.
2221 * 3. Performs writes following reads for array synchronising.
2224 static void fix_read_error(struct r1conf *conf, int read_disk,
2225 sector_t sect, int sectors)
2227 struct mddev *mddev = conf->mddev;
2233 struct md_rdev *rdev;
2235 if (s > (PAGE_SIZE>>9))
2243 rdev = rcu_dereference(conf->mirrors[d].rdev);
2245 (test_bit(In_sync, &rdev->flags) ||
2246 (!test_bit(Faulty, &rdev->flags) &&
2247 rdev->recovery_offset >= sect + s)) &&
2248 is_badblock(rdev, sect, s,
2249 &first_bad, &bad_sectors) == 0) {
2250 atomic_inc(&rdev->nr_pending);
2252 if (sync_page_io(rdev, sect, s<<9,
2253 conf->tmppage, REQ_OP_READ, 0, false))
2255 rdev_dec_pending(rdev, mddev);
2261 if (d == conf->raid_disks * 2)
2263 } while (!success && d != read_disk);
2266 /* Cannot read from anywhere - mark it bad */
2267 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2268 if (!rdev_set_badblocks(rdev, sect, s, 0))
2269 md_error(mddev, rdev);
2272 /* write it back and re-read */
2274 while (d != read_disk) {
2276 d = conf->raid_disks * 2;
2279 rdev = rcu_dereference(conf->mirrors[d].rdev);
2281 !test_bit(Faulty, &rdev->flags)) {
2282 atomic_inc(&rdev->nr_pending);
2284 r1_sync_page_io(rdev, sect, s,
2285 conf->tmppage, WRITE);
2286 rdev_dec_pending(rdev, mddev);
2291 while (d != read_disk) {
2292 char b[BDEVNAME_SIZE];
2294 d = conf->raid_disks * 2;
2297 rdev = rcu_dereference(conf->mirrors[d].rdev);
2299 !test_bit(Faulty, &rdev->flags)) {
2300 atomic_inc(&rdev->nr_pending);
2302 if (r1_sync_page_io(rdev, sect, s,
2303 conf->tmppage, READ)) {
2304 atomic_add(s, &rdev->corrected_errors);
2305 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2307 (unsigned long long)(sect +
2309 bdevname(rdev->bdev, b));
2311 rdev_dec_pending(rdev, mddev);
2320 static int narrow_write_error(struct r1bio *r1_bio, int i)
2322 struct mddev *mddev = r1_bio->mddev;
2323 struct r1conf *conf = mddev->private;
2324 struct md_rdev *rdev = conf->mirrors[i].rdev;
2326 /* bio has the data to be written to device 'i' where
2327 * we just recently had a write error.
2328 * We repeatedly clone the bio and trim down to one block,
2329 * then try the write. Where the write fails we record
2331 * It is conceivable that the bio doesn't exactly align with
2332 * blocks. We must handle this somehow.
2334 * We currently own a reference on the rdev.
2340 int sect_to_write = r1_bio->sectors;
2343 if (rdev->badblocks.shift < 0)
2346 block_sectors = roundup(1 << rdev->badblocks.shift,
2347 bdev_logical_block_size(rdev->bdev) >> 9);
2348 sector = r1_bio->sector;
2349 sectors = ((sector + block_sectors)
2350 & ~(sector_t)(block_sectors - 1))
2353 while (sect_to_write) {
2355 if (sectors > sect_to_write)
2356 sectors = sect_to_write;
2357 /* Write at 'sector' for 'sectors'*/
2359 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2360 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2364 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2368 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2369 wbio->bi_iter.bi_sector = r1_bio->sector;
2370 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2372 bio_trim(wbio, sector - r1_bio->sector, sectors);
2373 wbio->bi_iter.bi_sector += rdev->data_offset;
2374 bio_set_dev(wbio, rdev->bdev);
2376 if (submit_bio_wait(wbio) < 0)
2378 ok = rdev_set_badblocks(rdev, sector,
2383 sect_to_write -= sectors;
2385 sectors = block_sectors;
2390 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2393 int s = r1_bio->sectors;
2394 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2395 struct md_rdev *rdev = conf->mirrors[m].rdev;
2396 struct bio *bio = r1_bio->bios[m];
2397 if (bio->bi_end_io == NULL)
2399 if (!bio->bi_status &&
2400 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2401 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2403 if (bio->bi_status &&
2404 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2405 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2406 md_error(conf->mddev, rdev);
2410 md_done_sync(conf->mddev, s, 1);
2413 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2418 for (m = 0; m < conf->raid_disks * 2 ; m++)
2419 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2420 struct md_rdev *rdev = conf->mirrors[m].rdev;
2421 rdev_clear_badblocks(rdev,
2423 r1_bio->sectors, 0);
2424 rdev_dec_pending(rdev, conf->mddev);
2425 } else if (r1_bio->bios[m] != NULL) {
2426 /* This drive got a write error. We need to
2427 * narrow down and record precise write
2431 if (!narrow_write_error(r1_bio, m)) {
2432 md_error(conf->mddev,
2433 conf->mirrors[m].rdev);
2434 /* an I/O failed, we can't clear the bitmap */
2435 set_bit(R1BIO_Degraded, &r1_bio->state);
2437 rdev_dec_pending(conf->mirrors[m].rdev,
2441 spin_lock_irq(&conf->device_lock);
2442 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2443 idx = sector_to_idx(r1_bio->sector);
2444 atomic_inc(&conf->nr_queued[idx]);
2445 spin_unlock_irq(&conf->device_lock);
2447 * In case freeze_array() is waiting for condition
2448 * get_unqueued_pending() == extra to be true.
2450 wake_up(&conf->wait_barrier);
2451 md_wakeup_thread(conf->mddev->thread);
2453 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2454 close_write(r1_bio);
2455 raid_end_bio_io(r1_bio);
2459 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2461 struct mddev *mddev = conf->mddev;
2463 struct md_rdev *rdev;
2465 clear_bit(R1BIO_ReadError, &r1_bio->state);
2466 /* we got a read error. Maybe the drive is bad. Maybe just
2467 * the block and we can fix it.
2468 * We freeze all other IO, and try reading the block from
2469 * other devices. When we find one, we re-write
2470 * and check it that fixes the read error.
2471 * This is all done synchronously while the array is
2475 bio = r1_bio->bios[r1_bio->read_disk];
2477 r1_bio->bios[r1_bio->read_disk] = NULL;
2479 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2481 && !test_bit(FailFast, &rdev->flags)) {
2482 freeze_array(conf, 1);
2483 fix_read_error(conf, r1_bio->read_disk,
2484 r1_bio->sector, r1_bio->sectors);
2485 unfreeze_array(conf);
2486 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2487 md_error(mddev, rdev);
2489 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2492 rdev_dec_pending(rdev, conf->mddev);
2493 allow_barrier(conf, r1_bio->sector);
2494 bio = r1_bio->master_bio;
2496 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2498 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2501 static void raid1d(struct md_thread *thread)
2503 struct mddev *mddev = thread->mddev;
2504 struct r1bio *r1_bio;
2505 unsigned long flags;
2506 struct r1conf *conf = mddev->private;
2507 struct list_head *head = &conf->retry_list;
2508 struct blk_plug plug;
2511 md_check_recovery(mddev);
2513 if (!list_empty_careful(&conf->bio_end_io_list) &&
2514 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2516 spin_lock_irqsave(&conf->device_lock, flags);
2517 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2518 list_splice_init(&conf->bio_end_io_list, &tmp);
2519 spin_unlock_irqrestore(&conf->device_lock, flags);
2520 while (!list_empty(&tmp)) {
2521 r1_bio = list_first_entry(&tmp, struct r1bio,
2523 list_del(&r1_bio->retry_list);
2524 idx = sector_to_idx(r1_bio->sector);
2525 atomic_dec(&conf->nr_queued[idx]);
2526 if (mddev->degraded)
2527 set_bit(R1BIO_Degraded, &r1_bio->state);
2528 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2529 close_write(r1_bio);
2530 raid_end_bio_io(r1_bio);
2534 blk_start_plug(&plug);
2537 flush_pending_writes(conf);
2539 spin_lock_irqsave(&conf->device_lock, flags);
2540 if (list_empty(head)) {
2541 spin_unlock_irqrestore(&conf->device_lock, flags);
2544 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2545 list_del(head->prev);
2546 idx = sector_to_idx(r1_bio->sector);
2547 atomic_dec(&conf->nr_queued[idx]);
2548 spin_unlock_irqrestore(&conf->device_lock, flags);
2550 mddev = r1_bio->mddev;
2551 conf = mddev->private;
2552 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2553 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2554 test_bit(R1BIO_WriteError, &r1_bio->state))
2555 handle_sync_write_finished(conf, r1_bio);
2557 sync_request_write(mddev, r1_bio);
2558 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2559 test_bit(R1BIO_WriteError, &r1_bio->state))
2560 handle_write_finished(conf, r1_bio);
2561 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2562 handle_read_error(conf, r1_bio);
2567 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2568 md_check_recovery(mddev);
2570 blk_finish_plug(&plug);
2573 static int init_resync(struct r1conf *conf)
2577 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2578 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2580 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2581 r1buf_pool_free, conf->poolinfo);
2584 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2586 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2587 struct resync_pages *rps;
2591 for (i = conf->poolinfo->raid_disks; i--; ) {
2592 bio = r1bio->bios[i];
2593 rps = bio->bi_private;
2595 bio->bi_private = rps;
2597 r1bio->master_bio = NULL;
2602 * perform a "sync" on one "block"
2604 * We need to make sure that no normal I/O request - particularly write
2605 * requests - conflict with active sync requests.
2607 * This is achieved by tracking pending requests and a 'barrier' concept
2608 * that can be installed to exclude normal IO requests.
2611 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2614 struct r1conf *conf = mddev->private;
2615 struct r1bio *r1_bio;
2617 sector_t max_sector, nr_sectors;
2621 int write_targets = 0, read_targets = 0;
2622 sector_t sync_blocks;
2623 int still_degraded = 0;
2624 int good_sectors = RESYNC_SECTORS;
2625 int min_bad = 0; /* number of sectors that are bad in all devices */
2626 int idx = sector_to_idx(sector_nr);
2629 if (!mempool_initialized(&conf->r1buf_pool))
2630 if (init_resync(conf))
2633 max_sector = mddev->dev_sectors;
2634 if (sector_nr >= max_sector) {
2635 /* If we aborted, we need to abort the
2636 * sync on the 'current' bitmap chunk (there will
2637 * only be one in raid1 resync.
2638 * We can find the current addess in mddev->curr_resync
2640 if (mddev->curr_resync < max_sector) /* aborted */
2641 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2643 else /* completed sync */
2646 md_bitmap_close_sync(mddev->bitmap);
2649 if (mddev_is_clustered(mddev)) {
2650 conf->cluster_sync_low = 0;
2651 conf->cluster_sync_high = 0;
2656 if (mddev->bitmap == NULL &&
2657 mddev->recovery_cp == MaxSector &&
2658 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2659 conf->fullsync == 0) {
2661 return max_sector - sector_nr;
2663 /* before building a request, check if we can skip these blocks..
2664 * This call the bitmap_start_sync doesn't actually record anything
2666 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2667 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2668 /* We can skip this block, and probably several more */
2674 * If there is non-resync activity waiting for a turn, then let it
2675 * though before starting on this new sync request.
2677 if (atomic_read(&conf->nr_waiting[idx]))
2678 schedule_timeout_uninterruptible(1);
2680 /* we are incrementing sector_nr below. To be safe, we check against
2681 * sector_nr + two times RESYNC_SECTORS
2684 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2685 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2688 if (raise_barrier(conf, sector_nr))
2691 r1_bio = raid1_alloc_init_r1buf(conf);
2695 * If we get a correctably read error during resync or recovery,
2696 * we might want to read from a different device. So we
2697 * flag all drives that could conceivably be read from for READ,
2698 * and any others (which will be non-In_sync devices) for WRITE.
2699 * If a read fails, we try reading from something else for which READ
2703 r1_bio->mddev = mddev;
2704 r1_bio->sector = sector_nr;
2706 set_bit(R1BIO_IsSync, &r1_bio->state);
2707 /* make sure good_sectors won't go across barrier unit boundary */
2708 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2710 for (i = 0; i < conf->raid_disks * 2; i++) {
2711 struct md_rdev *rdev;
2712 bio = r1_bio->bios[i];
2714 rdev = rcu_dereference(conf->mirrors[i].rdev);
2716 test_bit(Faulty, &rdev->flags)) {
2717 if (i < conf->raid_disks)
2719 } else if (!test_bit(In_sync, &rdev->flags)) {
2720 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2721 bio->bi_end_io = end_sync_write;
2724 /* may need to read from here */
2725 sector_t first_bad = MaxSector;
2728 if (is_badblock(rdev, sector_nr, good_sectors,
2729 &first_bad, &bad_sectors)) {
2730 if (first_bad > sector_nr)
2731 good_sectors = first_bad - sector_nr;
2733 bad_sectors -= (sector_nr - first_bad);
2735 min_bad > bad_sectors)
2736 min_bad = bad_sectors;
2739 if (sector_nr < first_bad) {
2740 if (test_bit(WriteMostly, &rdev->flags)) {
2747 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2748 bio->bi_end_io = end_sync_read;
2750 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2751 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2752 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2754 * The device is suitable for reading (InSync),
2755 * but has bad block(s) here. Let's try to correct them,
2756 * if we are doing resync or repair. Otherwise, leave
2757 * this device alone for this sync request.
2759 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2760 bio->bi_end_io = end_sync_write;
2764 if (rdev && bio->bi_end_io) {
2765 atomic_inc(&rdev->nr_pending);
2766 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2767 bio_set_dev(bio, rdev->bdev);
2768 if (test_bit(FailFast, &rdev->flags))
2769 bio->bi_opf |= MD_FAILFAST;
2775 r1_bio->read_disk = disk;
2777 if (read_targets == 0 && min_bad > 0) {
2778 /* These sectors are bad on all InSync devices, so we
2779 * need to mark them bad on all write targets
2782 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2783 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2784 struct md_rdev *rdev = conf->mirrors[i].rdev;
2785 ok = rdev_set_badblocks(rdev, sector_nr,
2789 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2794 /* Cannot record the badblocks, so need to
2796 * If there are multiple read targets, could just
2797 * fail the really bad ones ???
2799 conf->recovery_disabled = mddev->recovery_disabled;
2800 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2806 if (min_bad > 0 && min_bad < good_sectors) {
2807 /* only resync enough to reach the next bad->good
2809 good_sectors = min_bad;
2812 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2813 /* extra read targets are also write targets */
2814 write_targets += read_targets-1;
2816 if (write_targets == 0 || read_targets == 0) {
2817 /* There is nowhere to write, so all non-sync
2818 * drives must be failed - so we are finished
2822 max_sector = sector_nr + min_bad;
2823 rv = max_sector - sector_nr;
2829 if (max_sector > mddev->resync_max)
2830 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2831 if (max_sector > sector_nr + good_sectors)
2832 max_sector = sector_nr + good_sectors;
2837 int len = PAGE_SIZE;
2838 if (sector_nr + (len>>9) > max_sector)
2839 len = (max_sector - sector_nr) << 9;
2842 if (sync_blocks == 0) {
2843 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2844 &sync_blocks, still_degraded) &&
2846 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2848 if ((len >> 9) > sync_blocks)
2849 len = sync_blocks<<9;
2852 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2853 struct resync_pages *rp;
2855 bio = r1_bio->bios[i];
2856 rp = get_resync_pages(bio);
2857 if (bio->bi_end_io) {
2858 page = resync_fetch_page(rp, page_idx);
2861 * won't fail because the vec table is big
2862 * enough to hold all these pages
2864 bio_add_page(bio, page, len, 0);
2867 nr_sectors += len>>9;
2868 sector_nr += len>>9;
2869 sync_blocks -= (len>>9);
2870 } while (++page_idx < RESYNC_PAGES);
2872 r1_bio->sectors = nr_sectors;
2874 if (mddev_is_clustered(mddev) &&
2875 conf->cluster_sync_high < sector_nr + nr_sectors) {
2876 conf->cluster_sync_low = mddev->curr_resync_completed;
2877 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2878 /* Send resync message */
2879 md_cluster_ops->resync_info_update(mddev,
2880 conf->cluster_sync_low,
2881 conf->cluster_sync_high);
2884 /* For a user-requested sync, we read all readable devices and do a
2887 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2888 atomic_set(&r1_bio->remaining, read_targets);
2889 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2890 bio = r1_bio->bios[i];
2891 if (bio->bi_end_io == end_sync_read) {
2893 md_sync_acct_bio(bio, nr_sectors);
2894 if (read_targets == 1)
2895 bio->bi_opf &= ~MD_FAILFAST;
2896 generic_make_request(bio);
2900 atomic_set(&r1_bio->remaining, 1);
2901 bio = r1_bio->bios[r1_bio->read_disk];
2902 md_sync_acct_bio(bio, nr_sectors);
2903 if (read_targets == 1)
2904 bio->bi_opf &= ~MD_FAILFAST;
2905 generic_make_request(bio);
2911 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2916 return mddev->dev_sectors;
2919 static struct r1conf *setup_conf(struct mddev *mddev)
2921 struct r1conf *conf;
2923 struct raid1_info *disk;
2924 struct md_rdev *rdev;
2927 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2931 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2932 sizeof(atomic_t), GFP_KERNEL);
2933 if (!conf->nr_pending)
2936 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2937 sizeof(atomic_t), GFP_KERNEL);
2938 if (!conf->nr_waiting)
2941 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2942 sizeof(atomic_t), GFP_KERNEL);
2943 if (!conf->nr_queued)
2946 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2947 sizeof(atomic_t), GFP_KERNEL);
2951 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2952 mddev->raid_disks, 2),
2957 conf->tmppage = alloc_page(GFP_KERNEL);
2961 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2962 if (!conf->poolinfo)
2964 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2965 err = mempool_init(&conf->r1bio_pool, NR_RAID1_BIOS, r1bio_pool_alloc,
2966 r1bio_pool_free, conf->poolinfo);
2970 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2974 conf->poolinfo->mddev = mddev;
2977 spin_lock_init(&conf->device_lock);
2978 rdev_for_each(rdev, mddev) {
2979 int disk_idx = rdev->raid_disk;
2980 if (disk_idx >= mddev->raid_disks
2983 if (test_bit(Replacement, &rdev->flags))
2984 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2986 disk = conf->mirrors + disk_idx;
2991 disk->head_position = 0;
2992 disk->seq_start = MaxSector;
2994 conf->raid_disks = mddev->raid_disks;
2995 conf->mddev = mddev;
2996 INIT_LIST_HEAD(&conf->retry_list);
2997 INIT_LIST_HEAD(&conf->bio_end_io_list);
2999 spin_lock_init(&conf->resync_lock);
3000 init_waitqueue_head(&conf->wait_barrier);
3002 bio_list_init(&conf->pending_bio_list);
3003 conf->pending_count = 0;
3004 conf->recovery_disabled = mddev->recovery_disabled - 1;
3007 for (i = 0; i < conf->raid_disks * 2; i++) {
3009 disk = conf->mirrors + i;
3011 if (i < conf->raid_disks &&
3012 disk[conf->raid_disks].rdev) {
3013 /* This slot has a replacement. */
3015 /* No original, just make the replacement
3016 * a recovering spare
3019 disk[conf->raid_disks].rdev;
3020 disk[conf->raid_disks].rdev = NULL;
3021 } else if (!test_bit(In_sync, &disk->rdev->flags))
3022 /* Original is not in_sync - bad */
3027 !test_bit(In_sync, &disk->rdev->flags)) {
3028 disk->head_position = 0;
3030 (disk->rdev->saved_raid_disk < 0))
3036 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3044 mempool_exit(&conf->r1bio_pool);
3045 kfree(conf->mirrors);
3046 safe_put_page(conf->tmppage);
3047 kfree(conf->poolinfo);
3048 kfree(conf->nr_pending);
3049 kfree(conf->nr_waiting);
3050 kfree(conf->nr_queued);
3051 kfree(conf->barrier);
3052 bioset_exit(&conf->bio_split);
3055 return ERR_PTR(err);
3058 static void raid1_free(struct mddev *mddev, void *priv);
3059 static int raid1_run(struct mddev *mddev)
3061 struct r1conf *conf;
3063 struct md_rdev *rdev;
3065 bool discard_supported = false;
3067 if (mddev->level != 1) {
3068 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3069 mdname(mddev), mddev->level);
3072 if (mddev->reshape_position != MaxSector) {
3073 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3077 if (mddev_init_writes_pending(mddev) < 0)
3080 * copy the already verified devices into our private RAID1
3081 * bookkeeping area. [whatever we allocate in run(),
3082 * should be freed in raid1_free()]
3084 if (mddev->private == NULL)
3085 conf = setup_conf(mddev);
3087 conf = mddev->private;
3090 return PTR_ERR(conf);
3093 blk_queue_max_write_same_sectors(mddev->queue, 0);
3094 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3097 rdev_for_each(rdev, mddev) {
3098 if (!mddev->gendisk)
3100 disk_stack_limits(mddev->gendisk, rdev->bdev,
3101 rdev->data_offset << 9);
3102 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3103 discard_supported = true;
3106 mddev->degraded = 0;
3107 for (i=0; i < conf->raid_disks; i++)
3108 if (conf->mirrors[i].rdev == NULL ||
3109 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3110 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3113 * RAID1 needs at least one disk in active
3115 if (conf->raid_disks - mddev->degraded < 1) {
3116 md_unregister_thread(&conf->thread);
3121 if (conf->raid_disks - mddev->degraded == 1)
3122 mddev->recovery_cp = MaxSector;
3124 if (mddev->recovery_cp != MaxSector)
3125 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3127 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3128 mdname(mddev), mddev->raid_disks - mddev->degraded,
3132 * Ok, everything is just fine now
3134 mddev->thread = conf->thread;
3135 conf->thread = NULL;
3136 mddev->private = conf;
3137 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3139 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3142 if (discard_supported)
3143 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3146 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3150 ret = md_integrity_register(mddev);
3152 md_unregister_thread(&mddev->thread);
3158 raid1_free(mddev, conf);
3162 static void raid1_free(struct mddev *mddev, void *priv)
3164 struct r1conf *conf = priv;
3166 mempool_exit(&conf->r1bio_pool);
3167 kfree(conf->mirrors);
3168 safe_put_page(conf->tmppage);
3169 kfree(conf->poolinfo);
3170 kfree(conf->nr_pending);
3171 kfree(conf->nr_waiting);
3172 kfree(conf->nr_queued);
3173 kfree(conf->barrier);
3174 bioset_exit(&conf->bio_split);
3178 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3180 /* no resync is happening, and there is enough space
3181 * on all devices, so we can resize.
3182 * We need to make sure resync covers any new space.
3183 * If the array is shrinking we should possibly wait until
3184 * any io in the removed space completes, but it hardly seems
3187 sector_t newsize = raid1_size(mddev, sectors, 0);
3188 if (mddev->external_size &&
3189 mddev->array_sectors > newsize)
3191 if (mddev->bitmap) {
3192 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3196 md_set_array_sectors(mddev, newsize);
3197 if (sectors > mddev->dev_sectors &&
3198 mddev->recovery_cp > mddev->dev_sectors) {
3199 mddev->recovery_cp = mddev->dev_sectors;
3200 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3202 mddev->dev_sectors = sectors;
3203 mddev->resync_max_sectors = sectors;
3207 static int raid1_reshape(struct mddev *mddev)
3210 * 1/ resize the r1bio_pool
3211 * 2/ resize conf->mirrors
3213 * We allocate a new r1bio_pool if we can.
3214 * Then raise a device barrier and wait until all IO stops.
3215 * Then resize conf->mirrors and swap in the new r1bio pool.
3217 * At the same time, we "pack" the devices so that all the missing
3218 * devices have the higher raid_disk numbers.
3220 mempool_t newpool, oldpool;
3221 struct pool_info *newpoolinfo;
3222 struct raid1_info *newmirrors;
3223 struct r1conf *conf = mddev->private;
3224 int cnt, raid_disks;
3225 unsigned long flags;
3229 memset(&newpool, 0, sizeof(newpool));
3230 memset(&oldpool, 0, sizeof(oldpool));
3232 /* Cannot change chunk_size, layout, or level */
3233 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3234 mddev->layout != mddev->new_layout ||
3235 mddev->level != mddev->new_level) {
3236 mddev->new_chunk_sectors = mddev->chunk_sectors;
3237 mddev->new_layout = mddev->layout;
3238 mddev->new_level = mddev->level;
3242 if (!mddev_is_clustered(mddev))
3243 md_allow_write(mddev);
3245 raid_disks = mddev->raid_disks + mddev->delta_disks;
3247 if (raid_disks < conf->raid_disks) {
3249 for (d= 0; d < conf->raid_disks; d++)
3250 if (conf->mirrors[d].rdev)
3252 if (cnt > raid_disks)
3256 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3259 newpoolinfo->mddev = mddev;
3260 newpoolinfo->raid_disks = raid_disks * 2;
3262 ret = mempool_init(&newpool, NR_RAID1_BIOS, r1bio_pool_alloc,
3263 r1bio_pool_free, newpoolinfo);
3268 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3273 mempool_exit(&newpool);
3277 freeze_array(conf, 0);
3279 /* ok, everything is stopped */
3280 oldpool = conf->r1bio_pool;
3281 conf->r1bio_pool = newpool;
3283 for (d = d2 = 0; d < conf->raid_disks; d++) {
3284 struct md_rdev *rdev = conf->mirrors[d].rdev;
3285 if (rdev && rdev->raid_disk != d2) {
3286 sysfs_unlink_rdev(mddev, rdev);
3287 rdev->raid_disk = d2;
3288 sysfs_unlink_rdev(mddev, rdev);
3289 if (sysfs_link_rdev(mddev, rdev))
3290 pr_warn("md/raid1:%s: cannot register rd%d\n",
3291 mdname(mddev), rdev->raid_disk);
3294 newmirrors[d2++].rdev = rdev;
3296 kfree(conf->mirrors);
3297 conf->mirrors = newmirrors;
3298 kfree(conf->poolinfo);
3299 conf->poolinfo = newpoolinfo;
3301 spin_lock_irqsave(&conf->device_lock, flags);
3302 mddev->degraded += (raid_disks - conf->raid_disks);
3303 spin_unlock_irqrestore(&conf->device_lock, flags);
3304 conf->raid_disks = mddev->raid_disks = raid_disks;
3305 mddev->delta_disks = 0;
3307 unfreeze_array(conf);
3309 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3310 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3311 md_wakeup_thread(mddev->thread);
3313 mempool_exit(&oldpool);
3317 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3319 struct r1conf *conf = mddev->private;
3322 freeze_array(conf, 0);
3324 unfreeze_array(conf);
3327 static void *raid1_takeover(struct mddev *mddev)
3329 /* raid1 can take over:
3330 * raid5 with 2 devices, any layout or chunk size
3332 if (mddev->level == 5 && mddev->raid_disks == 2) {
3333 struct r1conf *conf;
3334 mddev->new_level = 1;
3335 mddev->new_layout = 0;
3336 mddev->new_chunk_sectors = 0;
3337 conf = setup_conf(mddev);
3338 if (!IS_ERR(conf)) {
3339 /* Array must appear to be quiesced */
3340 conf->array_frozen = 1;
3341 mddev_clear_unsupported_flags(mddev,
3342 UNSUPPORTED_MDDEV_FLAGS);
3346 return ERR_PTR(-EINVAL);
3349 static struct md_personality raid1_personality =
3353 .owner = THIS_MODULE,
3354 .make_request = raid1_make_request,
3357 .status = raid1_status,
3358 .error_handler = raid1_error,
3359 .hot_add_disk = raid1_add_disk,
3360 .hot_remove_disk= raid1_remove_disk,
3361 .spare_active = raid1_spare_active,
3362 .sync_request = raid1_sync_request,
3363 .resize = raid1_resize,
3365 .check_reshape = raid1_reshape,
3366 .quiesce = raid1_quiesce,
3367 .takeover = raid1_takeover,
3368 .congested = raid1_congested,
3371 static int __init raid_init(void)
3373 return register_md_personality(&raid1_personality);
3376 static void raid_exit(void)
3378 unregister_md_personality(&raid1_personality);
3381 module_init(raid_init);
3382 module_exit(raid_exit);
3383 MODULE_LICENSE("GPL");
3384 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3385 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3386 MODULE_ALIAS("md-raid1");
3387 MODULE_ALIAS("md-level-1");
3389 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);