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>
40 #include <linux/sched/signal.h>
42 #include <trace/events/block.h>
48 #define UNSUPPORTED_MDDEV_FLAGS \
49 ((1L << MD_HAS_JOURNAL) | \
50 (1L << MD_JOURNAL_CLEAN) | \
51 (1L << MD_HAS_PPL) | \
52 (1L << MD_HAS_MULTIPLE_PPLS))
55 * Number of guaranteed r1bios in case of extreme VM load:
57 #define NR_RAID1_BIOS 256
59 /* when we get a read error on a read-only array, we redirect to another
60 * device without failing the first device, or trying to over-write to
61 * correct the read error. To keep track of bad blocks on a per-bio
62 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
64 #define IO_BLOCKED ((struct bio *)1)
65 /* When we successfully write to a known bad-block, we need to remove the
66 * bad-block marking which must be done from process context. So we record
67 * the success by setting devs[n].bio to IO_MADE_GOOD
69 #define IO_MADE_GOOD ((struct bio *)2)
71 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
73 /* When there are this many requests queue to be written by
74 * the raid1 thread, we become 'congested' to provide back-pressure
77 static int max_queued_requests = 1024;
79 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
80 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
82 #define raid1_log(md, fmt, args...) \
83 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
88 * for resync bio, r1bio pointer can be retrieved from the per-bio
89 * 'struct resync_pages'.
91 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
93 return get_resync_pages(bio)->raid_bio;
96 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
98 struct pool_info *pi = data;
99 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
101 /* allocate a r1bio with room for raid_disks entries in the bios array */
102 return kzalloc(size, gfp_flags);
105 static void r1bio_pool_free(void *r1_bio, void *data)
110 #define RESYNC_DEPTH 32
111 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
112 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
113 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
114 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
115 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
117 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
119 struct pool_info *pi = data;
120 struct r1bio *r1_bio;
124 struct resync_pages *rps;
126 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
130 rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks,
136 * Allocate bios : 1 for reading, n-1 for writing
138 for (j = pi->raid_disks ; j-- ; ) {
139 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
142 r1_bio->bios[j] = bio;
145 * Allocate RESYNC_PAGES data pages and attach them to
147 * If this is a user-requested check/repair, allocate
148 * RESYNC_PAGES for each bio.
150 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
151 need_pages = pi->raid_disks;
154 for (j = 0; j < pi->raid_disks; j++) {
155 struct resync_pages *rp = &rps[j];
157 bio = r1_bio->bios[j];
159 if (j < need_pages) {
160 if (resync_alloc_pages(rp, gfp_flags))
163 memcpy(rp, &rps[0], sizeof(*rp));
164 resync_get_all_pages(rp);
167 rp->raid_bio = r1_bio;
168 bio->bi_private = rp;
171 r1_bio->master_bio = NULL;
177 resync_free_pages(&rps[j]);
180 while (++j < pi->raid_disks)
181 bio_put(r1_bio->bios[j]);
185 r1bio_pool_free(r1_bio, data);
189 static void r1buf_pool_free(void *__r1_bio, void *data)
191 struct pool_info *pi = data;
193 struct r1bio *r1bio = __r1_bio;
194 struct resync_pages *rp = NULL;
196 for (i = pi->raid_disks; i--; ) {
197 rp = get_resync_pages(r1bio->bios[i]);
198 resync_free_pages(rp);
199 bio_put(r1bio->bios[i]);
202 /* resync pages array stored in the 1st bio's .bi_private */
205 r1bio_pool_free(r1bio, data);
208 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
212 for (i = 0; i < conf->raid_disks * 2; i++) {
213 struct bio **bio = r1_bio->bios + i;
214 if (!BIO_SPECIAL(*bio))
220 static void free_r1bio(struct r1bio *r1_bio)
222 struct r1conf *conf = r1_bio->mddev->private;
224 put_all_bios(conf, r1_bio);
225 mempool_free(r1_bio, conf->r1bio_pool);
228 static void put_buf(struct r1bio *r1_bio)
230 struct r1conf *conf = r1_bio->mddev->private;
231 sector_t sect = r1_bio->sector;
234 for (i = 0; i < conf->raid_disks * 2; i++) {
235 struct bio *bio = r1_bio->bios[i];
237 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
240 mempool_free(r1_bio, conf->r1buf_pool);
242 lower_barrier(conf, sect);
245 static void reschedule_retry(struct r1bio *r1_bio)
248 struct mddev *mddev = r1_bio->mddev;
249 struct r1conf *conf = mddev->private;
252 idx = sector_to_idx(r1_bio->sector);
253 spin_lock_irqsave(&conf->device_lock, flags);
254 list_add(&r1_bio->retry_list, &conf->retry_list);
255 atomic_inc(&conf->nr_queued[idx]);
256 spin_unlock_irqrestore(&conf->device_lock, flags);
258 wake_up(&conf->wait_barrier);
259 md_wakeup_thread(mddev->thread);
263 * raid_end_bio_io() is called when we have finished servicing a mirrored
264 * operation and are ready to return a success/failure code to the buffer
267 static void call_bio_endio(struct r1bio *r1_bio)
269 struct bio *bio = r1_bio->master_bio;
270 struct r1conf *conf = r1_bio->mddev->private;
272 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
273 bio->bi_status = BLK_STS_IOERR;
277 * Wake up any possible resync thread that waits for the device
280 allow_barrier(conf, r1_bio->sector);
283 static void raid_end_bio_io(struct r1bio *r1_bio)
285 struct bio *bio = r1_bio->master_bio;
287 /* if nobody has done the final endio yet, do it now */
288 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
289 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
290 (bio_data_dir(bio) == WRITE) ? "write" : "read",
291 (unsigned long long) bio->bi_iter.bi_sector,
292 (unsigned long long) bio_end_sector(bio) - 1);
294 call_bio_endio(r1_bio);
300 * Update disk head position estimator based on IRQ completion info.
302 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
304 struct r1conf *conf = r1_bio->mddev->private;
306 conf->mirrors[disk].head_position =
307 r1_bio->sector + (r1_bio->sectors);
311 * Find the disk number which triggered given bio
313 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
316 struct r1conf *conf = r1_bio->mddev->private;
317 int raid_disks = conf->raid_disks;
319 for (mirror = 0; mirror < raid_disks * 2; mirror++)
320 if (r1_bio->bios[mirror] == bio)
323 BUG_ON(mirror == raid_disks * 2);
324 update_head_pos(mirror, r1_bio);
329 static void raid1_end_read_request(struct bio *bio)
331 int uptodate = !bio->bi_status;
332 struct r1bio *r1_bio = bio->bi_private;
333 struct r1conf *conf = r1_bio->mddev->private;
334 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
337 * this branch is our 'one mirror IO has finished' event handler:
339 update_head_pos(r1_bio->read_disk, r1_bio);
342 set_bit(R1BIO_Uptodate, &r1_bio->state);
343 else if (test_bit(FailFast, &rdev->flags) &&
344 test_bit(R1BIO_FailFast, &r1_bio->state))
345 /* This was a fail-fast read so we definitely
349 /* If all other devices have failed, we want to return
350 * the error upwards rather than fail the last device.
351 * Here we redefine "uptodate" to mean "Don't want to retry"
354 spin_lock_irqsave(&conf->device_lock, flags);
355 if (r1_bio->mddev->degraded == conf->raid_disks ||
356 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
357 test_bit(In_sync, &rdev->flags)))
359 spin_unlock_irqrestore(&conf->device_lock, flags);
363 raid_end_bio_io(r1_bio);
364 rdev_dec_pending(rdev, conf->mddev);
369 char b[BDEVNAME_SIZE];
370 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
372 bdevname(rdev->bdev, b),
373 (unsigned long long)r1_bio->sector);
374 set_bit(R1BIO_ReadError, &r1_bio->state);
375 reschedule_retry(r1_bio);
376 /* don't drop the reference on read_disk yet */
380 static void close_write(struct r1bio *r1_bio)
382 /* it really is the end of this request */
383 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
384 bio_free_pages(r1_bio->behind_master_bio);
385 bio_put(r1_bio->behind_master_bio);
386 r1_bio->behind_master_bio = NULL;
388 /* clear the bitmap if all writes complete successfully */
389 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
391 !test_bit(R1BIO_Degraded, &r1_bio->state),
392 test_bit(R1BIO_BehindIO, &r1_bio->state));
393 md_write_end(r1_bio->mddev);
396 static void r1_bio_write_done(struct r1bio *r1_bio)
398 if (!atomic_dec_and_test(&r1_bio->remaining))
401 if (test_bit(R1BIO_WriteError, &r1_bio->state))
402 reschedule_retry(r1_bio);
405 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
406 reschedule_retry(r1_bio);
408 raid_end_bio_io(r1_bio);
412 static void raid1_end_write_request(struct bio *bio)
414 struct r1bio *r1_bio = bio->bi_private;
415 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
416 struct r1conf *conf = r1_bio->mddev->private;
417 struct bio *to_put = NULL;
418 int mirror = find_bio_disk(r1_bio, bio);
419 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
422 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
425 * 'one mirror IO has finished' event handler:
427 if (bio->bi_status && !discard_error) {
428 set_bit(WriteErrorSeen, &rdev->flags);
429 if (!test_and_set_bit(WantReplacement, &rdev->flags))
430 set_bit(MD_RECOVERY_NEEDED, &
431 conf->mddev->recovery);
433 if (test_bit(FailFast, &rdev->flags) &&
434 (bio->bi_opf & MD_FAILFAST) &&
435 /* We never try FailFast to WriteMostly devices */
436 !test_bit(WriteMostly, &rdev->flags)) {
437 md_error(r1_bio->mddev, rdev);
438 if (!test_bit(Faulty, &rdev->flags))
439 /* This is the only remaining device,
440 * We need to retry the write without
443 set_bit(R1BIO_WriteError, &r1_bio->state);
445 /* Finished with this branch */
446 r1_bio->bios[mirror] = NULL;
450 set_bit(R1BIO_WriteError, &r1_bio->state);
453 * Set R1BIO_Uptodate in our master bio, so that we
454 * will return a good error code for to the higher
455 * levels even if IO on some other mirrored buffer
458 * The 'master' represents the composite IO operation
459 * to user-side. So if something waits for IO, then it
460 * will wait for the 'master' bio.
465 r1_bio->bios[mirror] = NULL;
468 * Do not set R1BIO_Uptodate if the current device is
469 * rebuilding or Faulty. This is because we cannot use
470 * such device for properly reading the data back (we could
471 * potentially use it, if the current write would have felt
472 * before rdev->recovery_offset, but for simplicity we don't
475 if (test_bit(In_sync, &rdev->flags) &&
476 !test_bit(Faulty, &rdev->flags))
477 set_bit(R1BIO_Uptodate, &r1_bio->state);
479 /* Maybe we can clear some bad blocks. */
480 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
481 &first_bad, &bad_sectors) && !discard_error) {
482 r1_bio->bios[mirror] = IO_MADE_GOOD;
483 set_bit(R1BIO_MadeGood, &r1_bio->state);
488 if (test_bit(WriteMostly, &rdev->flags))
489 atomic_dec(&r1_bio->behind_remaining);
492 * In behind mode, we ACK the master bio once the I/O
493 * has safely reached all non-writemostly
494 * disks. Setting the Returned bit ensures that this
495 * gets done only once -- we don't ever want to return
496 * -EIO here, instead we'll wait
498 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
499 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
500 /* Maybe we can return now */
501 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
502 struct bio *mbio = r1_bio->master_bio;
503 pr_debug("raid1: behind end write sectors"
505 (unsigned long long) mbio->bi_iter.bi_sector,
506 (unsigned long long) bio_end_sector(mbio) - 1);
507 call_bio_endio(r1_bio);
511 if (r1_bio->bios[mirror] == NULL)
512 rdev_dec_pending(rdev, conf->mddev);
515 * Let's see if all mirrored write operations have finished
518 r1_bio_write_done(r1_bio);
524 static sector_t align_to_barrier_unit_end(sector_t start_sector,
529 WARN_ON(sectors == 0);
531 * len is the number of sectors from start_sector to end of the
532 * barrier unit which start_sector belongs to.
534 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
544 * This routine returns the disk from which the requested read should
545 * be done. There is a per-array 'next expected sequential IO' sector
546 * number - if this matches on the next IO then we use the last disk.
547 * There is also a per-disk 'last know head position' sector that is
548 * maintained from IRQ contexts, both the normal and the resync IO
549 * completion handlers update this position correctly. If there is no
550 * perfect sequential match then we pick the disk whose head is closest.
552 * If there are 2 mirrors in the same 2 devices, performance degrades
553 * because position is mirror, not device based.
555 * The rdev for the device selected will have nr_pending incremented.
557 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
559 const sector_t this_sector = r1_bio->sector;
561 int best_good_sectors;
562 int best_disk, best_dist_disk, best_pending_disk;
566 unsigned int min_pending;
567 struct md_rdev *rdev;
569 int choose_next_idle;
573 * Check if we can balance. We can balance on the whole
574 * device if no resync is going on, or below the resync window.
575 * We take the first readable disk when above the resync window.
578 sectors = r1_bio->sectors;
581 best_dist = MaxSector;
582 best_pending_disk = -1;
583 min_pending = UINT_MAX;
584 best_good_sectors = 0;
586 choose_next_idle = 0;
587 clear_bit(R1BIO_FailFast, &r1_bio->state);
589 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
590 (mddev_is_clustered(conf->mddev) &&
591 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
592 this_sector + sectors)))
597 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
601 unsigned int pending;
604 rdev = rcu_dereference(conf->mirrors[disk].rdev);
605 if (r1_bio->bios[disk] == IO_BLOCKED
607 || test_bit(Faulty, &rdev->flags))
609 if (!test_bit(In_sync, &rdev->flags) &&
610 rdev->recovery_offset < this_sector + sectors)
612 if (test_bit(WriteMostly, &rdev->flags)) {
613 /* Don't balance among write-mostly, just
614 * use the first as a last resort */
615 if (best_dist_disk < 0) {
616 if (is_badblock(rdev, this_sector, sectors,
617 &first_bad, &bad_sectors)) {
618 if (first_bad <= this_sector)
619 /* Cannot use this */
621 best_good_sectors = first_bad - this_sector;
623 best_good_sectors = sectors;
624 best_dist_disk = disk;
625 best_pending_disk = disk;
629 /* This is a reasonable device to use. It might
632 if (is_badblock(rdev, this_sector, sectors,
633 &first_bad, &bad_sectors)) {
634 if (best_dist < MaxSector)
635 /* already have a better device */
637 if (first_bad <= this_sector) {
638 /* cannot read here. If this is the 'primary'
639 * device, then we must not read beyond
640 * bad_sectors from another device..
642 bad_sectors -= (this_sector - first_bad);
643 if (choose_first && sectors > bad_sectors)
644 sectors = bad_sectors;
645 if (best_good_sectors > sectors)
646 best_good_sectors = sectors;
649 sector_t good_sectors = first_bad - this_sector;
650 if (good_sectors > best_good_sectors) {
651 best_good_sectors = good_sectors;
659 if ((sectors > best_good_sectors) && (best_disk >= 0))
661 best_good_sectors = sectors;
665 /* At least two disks to choose from so failfast is OK */
666 set_bit(R1BIO_FailFast, &r1_bio->state);
668 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
669 has_nonrot_disk |= nonrot;
670 pending = atomic_read(&rdev->nr_pending);
671 dist = abs(this_sector - conf->mirrors[disk].head_position);
676 /* Don't change to another disk for sequential reads */
677 if (conf->mirrors[disk].next_seq_sect == this_sector
679 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
680 struct raid1_info *mirror = &conf->mirrors[disk];
684 * If buffered sequential IO size exceeds optimal
685 * iosize, check if there is idle disk. If yes, choose
686 * the idle disk. read_balance could already choose an
687 * idle disk before noticing it's a sequential IO in
688 * this disk. This doesn't matter because this disk
689 * will idle, next time it will be utilized after the
690 * first disk has IO size exceeds optimal iosize. In
691 * this way, iosize of the first disk will be optimal
692 * iosize at least. iosize of the second disk might be
693 * small, but not a big deal since when the second disk
694 * starts IO, the first disk is likely still busy.
696 if (nonrot && opt_iosize > 0 &&
697 mirror->seq_start != MaxSector &&
698 mirror->next_seq_sect > opt_iosize &&
699 mirror->next_seq_sect - opt_iosize >=
701 choose_next_idle = 1;
707 if (choose_next_idle)
710 if (min_pending > pending) {
711 min_pending = pending;
712 best_pending_disk = disk;
715 if (dist < best_dist) {
717 best_dist_disk = disk;
722 * If all disks are rotational, choose the closest disk. If any disk is
723 * non-rotational, choose the disk with less pending request even the
724 * disk is rotational, which might/might not be optimal for raids with
725 * mixed ratation/non-rotational disks depending on workload.
727 if (best_disk == -1) {
728 if (has_nonrot_disk || min_pending == 0)
729 best_disk = best_pending_disk;
731 best_disk = best_dist_disk;
734 if (best_disk >= 0) {
735 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
738 atomic_inc(&rdev->nr_pending);
739 sectors = best_good_sectors;
741 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
742 conf->mirrors[best_disk].seq_start = this_sector;
744 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
747 *max_sectors = sectors;
752 static int raid1_congested(struct mddev *mddev, int bits)
754 struct r1conf *conf = mddev->private;
757 if ((bits & (1 << WB_async_congested)) &&
758 conf->pending_count >= max_queued_requests)
762 for (i = 0; i < conf->raid_disks * 2; i++) {
763 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
764 if (rdev && !test_bit(Faulty, &rdev->flags)) {
765 struct request_queue *q = bdev_get_queue(rdev->bdev);
769 /* Note the '|| 1' - when read_balance prefers
770 * non-congested targets, it can be removed
772 if ((bits & (1 << WB_async_congested)) || 1)
773 ret |= bdi_congested(q->backing_dev_info, bits);
775 ret &= bdi_congested(q->backing_dev_info, bits);
782 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
784 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
785 bitmap_unplug(conf->mddev->bitmap);
786 wake_up(&conf->wait_barrier);
788 while (bio) { /* submit pending writes */
789 struct bio *next = bio->bi_next;
790 struct md_rdev *rdev = (void *)bio->bi_disk;
792 bio_set_dev(bio, rdev->bdev);
793 if (test_bit(Faulty, &rdev->flags)) {
795 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
796 !blk_queue_discard(bio->bi_disk->queue)))
800 generic_make_request(bio);
805 static void flush_pending_writes(struct r1conf *conf)
807 /* Any writes that have been queued but are awaiting
808 * bitmap updates get flushed here.
810 spin_lock_irq(&conf->device_lock);
812 if (conf->pending_bio_list.head) {
814 bio = bio_list_get(&conf->pending_bio_list);
815 conf->pending_count = 0;
816 spin_unlock_irq(&conf->device_lock);
817 flush_bio_list(conf, bio);
819 spin_unlock_irq(&conf->device_lock);
823 * Sometimes we need to suspend IO while we do something else,
824 * either some resync/recovery, or reconfigure the array.
825 * To do this we raise a 'barrier'.
826 * The 'barrier' is a counter that can be raised multiple times
827 * to count how many activities are happening which preclude
829 * We can only raise the barrier if there is no pending IO.
830 * i.e. if nr_pending == 0.
831 * We choose only to raise the barrier if no-one is waiting for the
832 * barrier to go down. This means that as soon as an IO request
833 * is ready, no other operations which require a barrier will start
834 * until the IO request has had a chance.
836 * So: regular IO calls 'wait_barrier'. When that returns there
837 * is no backgroup IO happening, It must arrange to call
838 * allow_barrier when it has finished its IO.
839 * backgroup IO calls must call raise_barrier. Once that returns
840 * there is no normal IO happeing. It must arrange to call
841 * lower_barrier when the particular background IO completes.
843 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
845 int idx = sector_to_idx(sector_nr);
847 spin_lock_irq(&conf->resync_lock);
849 /* Wait until no block IO is waiting */
850 wait_event_lock_irq(conf->wait_barrier,
851 !atomic_read(&conf->nr_waiting[idx]),
854 /* block any new IO from starting */
855 atomic_inc(&conf->barrier[idx]);
857 * In raise_barrier() we firstly increase conf->barrier[idx] then
858 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
859 * increase conf->nr_pending[idx] then check conf->barrier[idx].
860 * A memory barrier here to make sure conf->nr_pending[idx] won't
861 * be fetched before conf->barrier[idx] is increased. Otherwise
862 * there will be a race between raise_barrier() and _wait_barrier().
864 smp_mb__after_atomic();
866 /* For these conditions we must wait:
867 * A: while the array is in frozen state
868 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
869 * existing in corresponding I/O barrier bucket.
870 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
871 * max resync count which allowed on current I/O barrier bucket.
873 wait_event_lock_irq(conf->wait_barrier,
874 !conf->array_frozen &&
875 !atomic_read(&conf->nr_pending[idx]) &&
876 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH,
879 atomic_inc(&conf->nr_sync_pending);
880 spin_unlock_irq(&conf->resync_lock);
883 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
885 int idx = sector_to_idx(sector_nr);
887 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
889 atomic_dec(&conf->barrier[idx]);
890 atomic_dec(&conf->nr_sync_pending);
891 wake_up(&conf->wait_barrier);
894 static void _wait_barrier(struct r1conf *conf, int idx)
897 * We need to increase conf->nr_pending[idx] very early here,
898 * then raise_barrier() can be blocked when it waits for
899 * conf->nr_pending[idx] to be 0. Then we can avoid holding
900 * conf->resync_lock when there is no barrier raised in same
901 * barrier unit bucket. Also if the array is frozen, I/O
902 * should be blocked until array is unfrozen.
904 atomic_inc(&conf->nr_pending[idx]);
906 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
907 * check conf->barrier[idx]. In raise_barrier() we firstly increase
908 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
909 * barrier is necessary here to make sure conf->barrier[idx] won't be
910 * fetched before conf->nr_pending[idx] is increased. Otherwise there
911 * will be a race between _wait_barrier() and raise_barrier().
913 smp_mb__after_atomic();
916 * Don't worry about checking two atomic_t variables at same time
917 * here. If during we check conf->barrier[idx], the array is
918 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
919 * 0, it is safe to return and make the I/O continue. Because the
920 * array is frozen, all I/O returned here will eventually complete
921 * or be queued, no race will happen. See code comment in
924 if (!READ_ONCE(conf->array_frozen) &&
925 !atomic_read(&conf->barrier[idx]))
929 * After holding conf->resync_lock, conf->nr_pending[idx]
930 * should be decreased before waiting for barrier to drop.
931 * Otherwise, we may encounter a race condition because
932 * raise_barrer() might be waiting for conf->nr_pending[idx]
933 * to be 0 at same time.
935 spin_lock_irq(&conf->resync_lock);
936 atomic_inc(&conf->nr_waiting[idx]);
937 atomic_dec(&conf->nr_pending[idx]);
939 * In case freeze_array() is waiting for
940 * get_unqueued_pending() == extra
942 wake_up(&conf->wait_barrier);
943 /* Wait for the barrier in same barrier unit bucket to drop. */
944 wait_event_lock_irq(conf->wait_barrier,
945 !conf->array_frozen &&
946 !atomic_read(&conf->barrier[idx]),
948 atomic_inc(&conf->nr_pending[idx]);
949 atomic_dec(&conf->nr_waiting[idx]);
950 spin_unlock_irq(&conf->resync_lock);
953 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
955 int idx = sector_to_idx(sector_nr);
958 * Very similar to _wait_barrier(). The difference is, for read
959 * I/O we don't need wait for sync I/O, but if the whole array
960 * is frozen, the read I/O still has to wait until the array is
961 * unfrozen. Since there is no ordering requirement with
962 * conf->barrier[idx] here, memory barrier is unnecessary as well.
964 atomic_inc(&conf->nr_pending[idx]);
966 if (!READ_ONCE(conf->array_frozen))
969 spin_lock_irq(&conf->resync_lock);
970 atomic_inc(&conf->nr_waiting[idx]);
971 atomic_dec(&conf->nr_pending[idx]);
973 * In case freeze_array() is waiting for
974 * get_unqueued_pending() == extra
976 wake_up(&conf->wait_barrier);
977 /* Wait for array to be unfrozen */
978 wait_event_lock_irq(conf->wait_barrier,
981 atomic_inc(&conf->nr_pending[idx]);
982 atomic_dec(&conf->nr_waiting[idx]);
983 spin_unlock_irq(&conf->resync_lock);
986 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
988 int idx = sector_to_idx(sector_nr);
990 _wait_barrier(conf, idx);
993 static void _allow_barrier(struct r1conf *conf, int idx)
995 atomic_dec(&conf->nr_pending[idx]);
996 wake_up(&conf->wait_barrier);
999 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1001 int idx = sector_to_idx(sector_nr);
1003 _allow_barrier(conf, idx);
1006 /* conf->resync_lock should be held */
1007 static int get_unqueued_pending(struct r1conf *conf)
1011 ret = atomic_read(&conf->nr_sync_pending);
1012 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1013 ret += atomic_read(&conf->nr_pending[idx]) -
1014 atomic_read(&conf->nr_queued[idx]);
1019 static void freeze_array(struct r1conf *conf, int extra)
1021 /* Stop sync I/O and normal I/O and wait for everything to
1023 * This is called in two situations:
1024 * 1) management command handlers (reshape, remove disk, quiesce).
1025 * 2) one normal I/O request failed.
1027 * After array_frozen is set to 1, new sync IO will be blocked at
1028 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1029 * or wait_read_barrier(). The flying I/Os will either complete or be
1030 * queued. When everything goes quite, there are only queued I/Os left.
1032 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1033 * barrier bucket index which this I/O request hits. When all sync and
1034 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1035 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1036 * in handle_read_error(), we may call freeze_array() before trying to
1037 * fix the read error. In this case, the error read I/O is not queued,
1038 * so get_unqueued_pending() == 1.
1040 * Therefore before this function returns, we need to wait until
1041 * get_unqueued_pendings(conf) gets equal to extra. For
1042 * normal I/O context, extra is 1, in rested situations extra is 0.
1044 spin_lock_irq(&conf->resync_lock);
1045 conf->array_frozen = 1;
1046 raid1_log(conf->mddev, "wait freeze");
1047 wait_event_lock_irq_cmd(
1049 get_unqueued_pending(conf) == extra,
1051 flush_pending_writes(conf));
1052 spin_unlock_irq(&conf->resync_lock);
1054 static void unfreeze_array(struct r1conf *conf)
1056 /* reverse the effect of the freeze */
1057 spin_lock_irq(&conf->resync_lock);
1058 conf->array_frozen = 0;
1059 spin_unlock_irq(&conf->resync_lock);
1060 wake_up(&conf->wait_barrier);
1063 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1066 int size = bio->bi_iter.bi_size;
1067 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1069 struct bio *behind_bio = NULL;
1071 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1075 /* discard op, we don't support writezero/writesame yet */
1076 if (!bio_has_data(bio)) {
1077 behind_bio->bi_iter.bi_size = size;
1081 while (i < vcnt && size) {
1083 int len = min_t(int, PAGE_SIZE, size);
1085 page = alloc_page(GFP_NOIO);
1086 if (unlikely(!page))
1089 bio_add_page(behind_bio, page, len, 0);
1095 bio_copy_data(behind_bio, bio);
1097 r1_bio->behind_master_bio = behind_bio;;
1098 set_bit(R1BIO_BehindIO, &r1_bio->state);
1103 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1104 bio->bi_iter.bi_size);
1105 bio_free_pages(behind_bio);
1106 bio_put(behind_bio);
1109 struct raid1_plug_cb {
1110 struct blk_plug_cb cb;
1111 struct bio_list pending;
1115 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1117 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1119 struct mddev *mddev = plug->cb.data;
1120 struct r1conf *conf = mddev->private;
1123 if (from_schedule || current->bio_list) {
1124 spin_lock_irq(&conf->device_lock);
1125 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1126 conf->pending_count += plug->pending_cnt;
1127 spin_unlock_irq(&conf->device_lock);
1128 wake_up(&conf->wait_barrier);
1129 md_wakeup_thread(mddev->thread);
1134 /* we aren't scheduling, so we can do the write-out directly. */
1135 bio = bio_list_get(&plug->pending);
1136 flush_bio_list(conf, bio);
1140 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1142 r1_bio->master_bio = bio;
1143 r1_bio->sectors = bio_sectors(bio);
1145 r1_bio->mddev = mddev;
1146 r1_bio->sector = bio->bi_iter.bi_sector;
1149 static inline struct r1bio *
1150 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1152 struct r1conf *conf = mddev->private;
1153 struct r1bio *r1_bio;
1155 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1156 /* Ensure no bio records IO_BLOCKED */
1157 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1158 init_r1bio(r1_bio, mddev, bio);
1162 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1163 int max_read_sectors, struct r1bio *r1_bio)
1165 struct r1conf *conf = mddev->private;
1166 struct raid1_info *mirror;
1167 struct bio *read_bio;
1168 struct bitmap *bitmap = mddev->bitmap;
1169 const int op = bio_op(bio);
1170 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1173 bool print_msg = !!r1_bio;
1174 char b[BDEVNAME_SIZE];
1177 * If r1_bio is set, we are blocking the raid1d thread
1178 * so there is a tiny risk of deadlock. So ask for
1179 * emergency memory if needed.
1181 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1184 /* Need to get the block device name carefully */
1185 struct md_rdev *rdev;
1187 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1189 bdevname(rdev->bdev, b);
1196 * Still need barrier for READ in case that whole
1199 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1202 r1_bio = alloc_r1bio(mddev, bio);
1204 init_r1bio(r1_bio, mddev, bio);
1205 r1_bio->sectors = max_read_sectors;
1208 * make_request() can abort the operation when read-ahead is being
1209 * used and no empty request is available.
1211 rdisk = read_balance(conf, r1_bio, &max_sectors);
1214 /* couldn't find anywhere to read from */
1216 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1219 (unsigned long long)r1_bio->sector);
1221 raid_end_bio_io(r1_bio);
1224 mirror = conf->mirrors + rdisk;
1227 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1229 (unsigned long long)r1_bio->sector,
1230 bdevname(mirror->rdev->bdev, b));
1232 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1235 * Reading from a write-mostly device must take care not to
1236 * over-take any writes that are 'behind'
1238 raid1_log(mddev, "wait behind writes");
1239 wait_event(bitmap->behind_wait,
1240 atomic_read(&bitmap->behind_writes) == 0);
1243 if (max_sectors < bio_sectors(bio)) {
1244 struct bio *split = bio_split(bio, max_sectors,
1245 gfp, conf->bio_split);
1246 bio_chain(split, bio);
1247 generic_make_request(bio);
1249 r1_bio->master_bio = bio;
1250 r1_bio->sectors = max_sectors;
1253 r1_bio->read_disk = rdisk;
1255 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1257 r1_bio->bios[rdisk] = read_bio;
1259 read_bio->bi_iter.bi_sector = r1_bio->sector +
1260 mirror->rdev->data_offset;
1261 bio_set_dev(read_bio, mirror->rdev->bdev);
1262 read_bio->bi_end_io = raid1_end_read_request;
1263 bio_set_op_attrs(read_bio, op, do_sync);
1264 if (test_bit(FailFast, &mirror->rdev->flags) &&
1265 test_bit(R1BIO_FailFast, &r1_bio->state))
1266 read_bio->bi_opf |= MD_FAILFAST;
1267 read_bio->bi_private = r1_bio;
1270 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1271 disk_devt(mddev->gendisk), r1_bio->sector);
1273 generic_make_request(read_bio);
1276 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1277 int max_write_sectors)
1279 struct r1conf *conf = mddev->private;
1280 struct r1bio *r1_bio;
1282 struct bitmap *bitmap = mddev->bitmap;
1283 unsigned long flags;
1284 struct md_rdev *blocked_rdev;
1285 struct blk_plug_cb *cb;
1286 struct raid1_plug_cb *plug = NULL;
1291 * Register the new request and wait if the reconstruction
1292 * thread has put up a bar for new requests.
1293 * Continue immediately if no resync is active currently.
1297 if ((bio_end_sector(bio) > mddev->suspend_lo &&
1298 bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1299 (mddev_is_clustered(mddev) &&
1300 md_cluster_ops->area_resyncing(mddev, WRITE,
1301 bio->bi_iter.bi_sector, bio_end_sector(bio)))) {
1304 * As the suspend_* range is controlled by userspace, we want
1305 * an interruptible wait.
1310 prepare_to_wait(&conf->wait_barrier,
1311 &w, TASK_INTERRUPTIBLE);
1312 if ((bio_end_sector(bio) <= mddev->suspend_lo ||
1313 bio->bi_iter.bi_sector >= mddev->suspend_hi) &&
1314 (!mddev_is_clustered(mddev) ||
1315 !md_cluster_ops->area_resyncing(mddev, WRITE,
1316 bio->bi_iter.bi_sector,
1317 bio_end_sector(bio))))
1320 sigprocmask(SIG_BLOCK, &full, &old);
1322 sigprocmask(SIG_SETMASK, &old, NULL);
1324 finish_wait(&conf->wait_barrier, &w);
1326 wait_barrier(conf, bio->bi_iter.bi_sector);
1328 r1_bio = alloc_r1bio(mddev, bio);
1329 r1_bio->sectors = max_write_sectors;
1331 if (conf->pending_count >= max_queued_requests) {
1332 md_wakeup_thread(mddev->thread);
1333 raid1_log(mddev, "wait queued");
1334 wait_event(conf->wait_barrier,
1335 conf->pending_count < max_queued_requests);
1337 /* first select target devices under rcu_lock and
1338 * inc refcount on their rdev. Record them by setting
1340 * If there are known/acknowledged bad blocks on any device on
1341 * which we have seen a write error, we want to avoid writing those
1343 * This potentially requires several writes to write around
1344 * the bad blocks. Each set of writes gets it's own r1bio
1345 * with a set of bios attached.
1348 disks = conf->raid_disks * 2;
1350 blocked_rdev = NULL;
1352 max_sectors = r1_bio->sectors;
1353 for (i = 0; i < disks; i++) {
1354 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1355 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1356 atomic_inc(&rdev->nr_pending);
1357 blocked_rdev = rdev;
1360 r1_bio->bios[i] = NULL;
1361 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1362 if (i < conf->raid_disks)
1363 set_bit(R1BIO_Degraded, &r1_bio->state);
1367 atomic_inc(&rdev->nr_pending);
1368 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1373 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1374 &first_bad, &bad_sectors);
1376 /* mustn't write here until the bad block is
1378 set_bit(BlockedBadBlocks, &rdev->flags);
1379 blocked_rdev = rdev;
1382 if (is_bad && first_bad <= r1_bio->sector) {
1383 /* Cannot write here at all */
1384 bad_sectors -= (r1_bio->sector - first_bad);
1385 if (bad_sectors < max_sectors)
1386 /* mustn't write more than bad_sectors
1387 * to other devices yet
1389 max_sectors = bad_sectors;
1390 rdev_dec_pending(rdev, mddev);
1391 /* We don't set R1BIO_Degraded as that
1392 * only applies if the disk is
1393 * missing, so it might be re-added,
1394 * and we want to know to recover this
1396 * In this case the device is here,
1397 * and the fact that this chunk is not
1398 * in-sync is recorded in the bad
1404 int good_sectors = first_bad - r1_bio->sector;
1405 if (good_sectors < max_sectors)
1406 max_sectors = good_sectors;
1409 r1_bio->bios[i] = bio;
1413 if (unlikely(blocked_rdev)) {
1414 /* Wait for this device to become unblocked */
1417 for (j = 0; j < i; j++)
1418 if (r1_bio->bios[j])
1419 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1421 allow_barrier(conf, bio->bi_iter.bi_sector);
1422 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1423 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1424 wait_barrier(conf, bio->bi_iter.bi_sector);
1428 if (max_sectors < bio_sectors(bio)) {
1429 struct bio *split = bio_split(bio, max_sectors,
1430 GFP_NOIO, conf->bio_split);
1431 bio_chain(split, bio);
1432 generic_make_request(bio);
1434 r1_bio->master_bio = bio;
1435 r1_bio->sectors = max_sectors;
1438 atomic_set(&r1_bio->remaining, 1);
1439 atomic_set(&r1_bio->behind_remaining, 0);
1443 for (i = 0; i < disks; i++) {
1444 struct bio *mbio = NULL;
1445 if (!r1_bio->bios[i])
1451 * Not if there are too many, or cannot
1452 * allocate memory, or a reader on WriteMostly
1453 * is waiting for behind writes to flush */
1455 (atomic_read(&bitmap->behind_writes)
1456 < mddev->bitmap_info.max_write_behind) &&
1457 !waitqueue_active(&bitmap->behind_wait)) {
1458 alloc_behind_master_bio(r1_bio, bio);
1461 bitmap_startwrite(bitmap, r1_bio->sector,
1463 test_bit(R1BIO_BehindIO,
1468 if (r1_bio->behind_master_bio)
1469 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1470 GFP_NOIO, mddev->bio_set);
1472 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1474 if (r1_bio->behind_master_bio) {
1475 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1476 atomic_inc(&r1_bio->behind_remaining);
1479 r1_bio->bios[i] = mbio;
1481 mbio->bi_iter.bi_sector = (r1_bio->sector +
1482 conf->mirrors[i].rdev->data_offset);
1483 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1484 mbio->bi_end_io = raid1_end_write_request;
1485 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1486 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1487 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1488 conf->raid_disks - mddev->degraded > 1)
1489 mbio->bi_opf |= MD_FAILFAST;
1490 mbio->bi_private = r1_bio;
1492 atomic_inc(&r1_bio->remaining);
1495 trace_block_bio_remap(mbio->bi_disk->queue,
1496 mbio, disk_devt(mddev->gendisk),
1498 /* flush_pending_writes() needs access to the rdev so...*/
1499 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1501 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1503 plug = container_of(cb, struct raid1_plug_cb, cb);
1507 bio_list_add(&plug->pending, mbio);
1508 plug->pending_cnt++;
1510 spin_lock_irqsave(&conf->device_lock, flags);
1511 bio_list_add(&conf->pending_bio_list, mbio);
1512 conf->pending_count++;
1513 spin_unlock_irqrestore(&conf->device_lock, flags);
1514 md_wakeup_thread(mddev->thread);
1518 r1_bio_write_done(r1_bio);
1520 /* In case raid1d snuck in to freeze_array */
1521 wake_up(&conf->wait_barrier);
1524 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1528 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1529 md_flush_request(mddev, bio);
1534 * There is a limit to the maximum size, but
1535 * the read/write handler might find a lower limit
1536 * due to bad blocks. To avoid multiple splits,
1537 * we pass the maximum number of sectors down
1538 * and let the lower level perform the split.
1540 sectors = align_to_barrier_unit_end(
1541 bio->bi_iter.bi_sector, bio_sectors(bio));
1543 if (bio_data_dir(bio) == READ)
1544 raid1_read_request(mddev, bio, sectors, NULL);
1546 if (!md_write_start(mddev,bio))
1548 raid1_write_request(mddev, bio, sectors);
1553 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1555 struct r1conf *conf = mddev->private;
1558 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1559 conf->raid_disks - mddev->degraded);
1561 for (i = 0; i < conf->raid_disks; i++) {
1562 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1563 seq_printf(seq, "%s",
1564 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1567 seq_printf(seq, "]");
1570 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1572 char b[BDEVNAME_SIZE];
1573 struct r1conf *conf = mddev->private;
1574 unsigned long flags;
1577 * If it is not operational, then we have already marked it as dead
1578 * else if it is the last working disks, ignore the error, let the
1579 * next level up know.
1580 * else mark the drive as failed
1582 spin_lock_irqsave(&conf->device_lock, flags);
1583 if (test_bit(In_sync, &rdev->flags)
1584 && (conf->raid_disks - mddev->degraded) == 1) {
1586 * Don't fail the drive, act as though we were just a
1587 * normal single drive.
1588 * However don't try a recovery from this drive as
1589 * it is very likely to fail.
1591 conf->recovery_disabled = mddev->recovery_disabled;
1592 spin_unlock_irqrestore(&conf->device_lock, flags);
1595 set_bit(Blocked, &rdev->flags);
1596 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1598 set_bit(Faulty, &rdev->flags);
1600 set_bit(Faulty, &rdev->flags);
1601 spin_unlock_irqrestore(&conf->device_lock, flags);
1603 * if recovery is running, make sure it aborts.
1605 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1606 set_mask_bits(&mddev->sb_flags, 0,
1607 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1608 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1609 "md/raid1:%s: Operation continuing on %d devices.\n",
1610 mdname(mddev), bdevname(rdev->bdev, b),
1611 mdname(mddev), conf->raid_disks - mddev->degraded);
1614 static void print_conf(struct r1conf *conf)
1618 pr_debug("RAID1 conf printout:\n");
1620 pr_debug("(!conf)\n");
1623 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1627 for (i = 0; i < conf->raid_disks; i++) {
1628 char b[BDEVNAME_SIZE];
1629 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1631 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1632 i, !test_bit(In_sync, &rdev->flags),
1633 !test_bit(Faulty, &rdev->flags),
1634 bdevname(rdev->bdev,b));
1639 static void close_sync(struct r1conf *conf)
1643 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1644 _wait_barrier(conf, idx);
1645 _allow_barrier(conf, idx);
1648 mempool_destroy(conf->r1buf_pool);
1649 conf->r1buf_pool = NULL;
1652 static int raid1_spare_active(struct mddev *mddev)
1655 struct r1conf *conf = mddev->private;
1657 unsigned long flags;
1660 * Find all failed disks within the RAID1 configuration
1661 * and mark them readable.
1662 * Called under mddev lock, so rcu protection not needed.
1663 * device_lock used to avoid races with raid1_end_read_request
1664 * which expects 'In_sync' flags and ->degraded to be consistent.
1666 spin_lock_irqsave(&conf->device_lock, flags);
1667 for (i = 0; i < conf->raid_disks; i++) {
1668 struct md_rdev *rdev = conf->mirrors[i].rdev;
1669 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1671 && !test_bit(Candidate, &repl->flags)
1672 && repl->recovery_offset == MaxSector
1673 && !test_bit(Faulty, &repl->flags)
1674 && !test_and_set_bit(In_sync, &repl->flags)) {
1675 /* replacement has just become active */
1677 !test_and_clear_bit(In_sync, &rdev->flags))
1680 /* Replaced device not technically
1681 * faulty, but we need to be sure
1682 * it gets removed and never re-added
1684 set_bit(Faulty, &rdev->flags);
1685 sysfs_notify_dirent_safe(
1690 && rdev->recovery_offset == MaxSector
1691 && !test_bit(Faulty, &rdev->flags)
1692 && !test_and_set_bit(In_sync, &rdev->flags)) {
1694 sysfs_notify_dirent_safe(rdev->sysfs_state);
1697 mddev->degraded -= count;
1698 spin_unlock_irqrestore(&conf->device_lock, flags);
1704 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1706 struct r1conf *conf = mddev->private;
1709 struct raid1_info *p;
1711 int last = conf->raid_disks - 1;
1713 if (mddev->recovery_disabled == conf->recovery_disabled)
1716 if (md_integrity_add_rdev(rdev, mddev))
1719 if (rdev->raid_disk >= 0)
1720 first = last = rdev->raid_disk;
1723 * find the disk ... but prefer rdev->saved_raid_disk
1726 if (rdev->saved_raid_disk >= 0 &&
1727 rdev->saved_raid_disk >= first &&
1728 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1729 first = last = rdev->saved_raid_disk;
1731 for (mirror = first; mirror <= last; mirror++) {
1732 p = conf->mirrors+mirror;
1736 disk_stack_limits(mddev->gendisk, rdev->bdev,
1737 rdev->data_offset << 9);
1739 p->head_position = 0;
1740 rdev->raid_disk = mirror;
1742 /* As all devices are equivalent, we don't need a full recovery
1743 * if this was recently any drive of the array
1745 if (rdev->saved_raid_disk < 0)
1747 rcu_assign_pointer(p->rdev, rdev);
1750 if (test_bit(WantReplacement, &p->rdev->flags) &&
1751 p[conf->raid_disks].rdev == NULL) {
1752 /* Add this device as a replacement */
1753 clear_bit(In_sync, &rdev->flags);
1754 set_bit(Replacement, &rdev->flags);
1755 rdev->raid_disk = mirror;
1758 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1762 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1763 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1768 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1770 struct r1conf *conf = mddev->private;
1772 int number = rdev->raid_disk;
1773 struct raid1_info *p = conf->mirrors + number;
1775 if (rdev != p->rdev)
1776 p = conf->mirrors + conf->raid_disks + number;
1779 if (rdev == p->rdev) {
1780 if (test_bit(In_sync, &rdev->flags) ||
1781 atomic_read(&rdev->nr_pending)) {
1785 /* Only remove non-faulty devices if recovery
1788 if (!test_bit(Faulty, &rdev->flags) &&
1789 mddev->recovery_disabled != conf->recovery_disabled &&
1790 mddev->degraded < conf->raid_disks) {
1795 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1797 if (atomic_read(&rdev->nr_pending)) {
1798 /* lost the race, try later */
1804 if (conf->mirrors[conf->raid_disks + number].rdev) {
1805 /* We just removed a device that is being replaced.
1806 * Move down the replacement. We drain all IO before
1807 * doing this to avoid confusion.
1809 struct md_rdev *repl =
1810 conf->mirrors[conf->raid_disks + number].rdev;
1811 freeze_array(conf, 0);
1812 clear_bit(Replacement, &repl->flags);
1814 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1815 unfreeze_array(conf);
1818 clear_bit(WantReplacement, &rdev->flags);
1819 err = md_integrity_register(mddev);
1827 static void end_sync_read(struct bio *bio)
1829 struct r1bio *r1_bio = get_resync_r1bio(bio);
1831 update_head_pos(r1_bio->read_disk, r1_bio);
1834 * we have read a block, now it needs to be re-written,
1835 * or re-read if the read failed.
1836 * We don't do much here, just schedule handling by raid1d
1838 if (!bio->bi_status)
1839 set_bit(R1BIO_Uptodate, &r1_bio->state);
1841 if (atomic_dec_and_test(&r1_bio->remaining))
1842 reschedule_retry(r1_bio);
1845 static void end_sync_write(struct bio *bio)
1847 int uptodate = !bio->bi_status;
1848 struct r1bio *r1_bio = get_resync_r1bio(bio);
1849 struct mddev *mddev = r1_bio->mddev;
1850 struct r1conf *conf = mddev->private;
1853 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1856 sector_t sync_blocks = 0;
1857 sector_t s = r1_bio->sector;
1858 long sectors_to_go = r1_bio->sectors;
1859 /* make sure these bits doesn't get cleared. */
1861 bitmap_end_sync(mddev->bitmap, s,
1864 sectors_to_go -= sync_blocks;
1865 } while (sectors_to_go > 0);
1866 set_bit(WriteErrorSeen, &rdev->flags);
1867 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1868 set_bit(MD_RECOVERY_NEEDED, &
1870 set_bit(R1BIO_WriteError, &r1_bio->state);
1871 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1872 &first_bad, &bad_sectors) &&
1873 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1876 &first_bad, &bad_sectors)
1878 set_bit(R1BIO_MadeGood, &r1_bio->state);
1880 if (atomic_dec_and_test(&r1_bio->remaining)) {
1881 int s = r1_bio->sectors;
1882 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1883 test_bit(R1BIO_WriteError, &r1_bio->state))
1884 reschedule_retry(r1_bio);
1887 md_done_sync(mddev, s, uptodate);
1892 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1893 int sectors, struct page *page, int rw)
1895 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1899 set_bit(WriteErrorSeen, &rdev->flags);
1900 if (!test_and_set_bit(WantReplacement,
1902 set_bit(MD_RECOVERY_NEEDED, &
1903 rdev->mddev->recovery);
1905 /* need to record an error - either for the block or the device */
1906 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1907 md_error(rdev->mddev, rdev);
1911 static int fix_sync_read_error(struct r1bio *r1_bio)
1913 /* Try some synchronous reads of other devices to get
1914 * good data, much like with normal read errors. Only
1915 * read into the pages we already have so we don't
1916 * need to re-issue the read request.
1917 * We don't need to freeze the array, because being in an
1918 * active sync request, there is no normal IO, and
1919 * no overlapping syncs.
1920 * We don't need to check is_badblock() again as we
1921 * made sure that anything with a bad block in range
1922 * will have bi_end_io clear.
1924 struct mddev *mddev = r1_bio->mddev;
1925 struct r1conf *conf = mddev->private;
1926 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1927 struct page **pages = get_resync_pages(bio)->pages;
1928 sector_t sect = r1_bio->sector;
1929 int sectors = r1_bio->sectors;
1931 struct md_rdev *rdev;
1933 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1934 if (test_bit(FailFast, &rdev->flags)) {
1935 /* Don't try recovering from here - just fail it
1936 * ... unless it is the last working device of course */
1937 md_error(mddev, rdev);
1938 if (test_bit(Faulty, &rdev->flags))
1939 /* Don't try to read from here, but make sure
1940 * put_buf does it's thing
1942 bio->bi_end_io = end_sync_write;
1947 int d = r1_bio->read_disk;
1951 if (s > (PAGE_SIZE>>9))
1954 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1955 /* No rcu protection needed here devices
1956 * can only be removed when no resync is
1957 * active, and resync is currently active
1959 rdev = conf->mirrors[d].rdev;
1960 if (sync_page_io(rdev, sect, s<<9,
1962 REQ_OP_READ, 0, false)) {
1968 if (d == conf->raid_disks * 2)
1970 } while (!success && d != r1_bio->read_disk);
1973 char b[BDEVNAME_SIZE];
1975 /* Cannot read from anywhere, this block is lost.
1976 * Record a bad block on each device. If that doesn't
1977 * work just disable and interrupt the recovery.
1978 * Don't fail devices as that won't really help.
1980 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1981 mdname(mddev), bio_devname(bio, b),
1982 (unsigned long long)r1_bio->sector);
1983 for (d = 0; d < conf->raid_disks * 2; d++) {
1984 rdev = conf->mirrors[d].rdev;
1985 if (!rdev || test_bit(Faulty, &rdev->flags))
1987 if (!rdev_set_badblocks(rdev, sect, s, 0))
1991 conf->recovery_disabled =
1992 mddev->recovery_disabled;
1993 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1994 md_done_sync(mddev, r1_bio->sectors, 0);
2006 /* write it back and re-read */
2007 while (d != r1_bio->read_disk) {
2009 d = conf->raid_disks * 2;
2011 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2013 rdev = conf->mirrors[d].rdev;
2014 if (r1_sync_page_io(rdev, sect, s,
2017 r1_bio->bios[d]->bi_end_io = NULL;
2018 rdev_dec_pending(rdev, mddev);
2022 while (d != r1_bio->read_disk) {
2024 d = conf->raid_disks * 2;
2026 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2028 rdev = conf->mirrors[d].rdev;
2029 if (r1_sync_page_io(rdev, sect, s,
2032 atomic_add(s, &rdev->corrected_errors);
2038 set_bit(R1BIO_Uptodate, &r1_bio->state);
2043 static void process_checks(struct r1bio *r1_bio)
2045 /* We have read all readable devices. If we haven't
2046 * got the block, then there is no hope left.
2047 * If we have, then we want to do a comparison
2048 * and skip the write if everything is the same.
2049 * If any blocks failed to read, then we need to
2050 * attempt an over-write
2052 struct mddev *mddev = r1_bio->mddev;
2053 struct r1conf *conf = mddev->private;
2058 /* Fix variable parts of all bios */
2059 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2060 for (i = 0; i < conf->raid_disks * 2; i++) {
2061 blk_status_t status;
2062 struct bio *b = r1_bio->bios[i];
2063 struct resync_pages *rp = get_resync_pages(b);
2064 if (b->bi_end_io != end_sync_read)
2066 /* fixup the bio for reuse, but preserve errno */
2067 status = b->bi_status;
2069 b->bi_status = status;
2070 b->bi_iter.bi_sector = r1_bio->sector +
2071 conf->mirrors[i].rdev->data_offset;
2072 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2073 b->bi_end_io = end_sync_read;
2074 rp->raid_bio = r1_bio;
2077 /* initialize bvec table again */
2078 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2080 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2081 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2082 !r1_bio->bios[primary]->bi_status) {
2083 r1_bio->bios[primary]->bi_end_io = NULL;
2084 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2087 r1_bio->read_disk = primary;
2088 for (i = 0; i < conf->raid_disks * 2; i++) {
2090 struct bio *pbio = r1_bio->bios[primary];
2091 struct bio *sbio = r1_bio->bios[i];
2092 blk_status_t status = sbio->bi_status;
2093 struct page **ppages = get_resync_pages(pbio)->pages;
2094 struct page **spages = get_resync_pages(sbio)->pages;
2096 int page_len[RESYNC_PAGES] = { 0 };
2098 if (sbio->bi_end_io != end_sync_read)
2100 /* Now we can 'fixup' the error value */
2101 sbio->bi_status = 0;
2103 bio_for_each_segment_all(bi, sbio, j)
2104 page_len[j] = bi->bv_len;
2107 for (j = vcnt; j-- ; ) {
2108 if (memcmp(page_address(ppages[j]),
2109 page_address(spages[j]),
2116 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2117 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2119 /* No need to write to this device. */
2120 sbio->bi_end_io = NULL;
2121 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2125 bio_copy_data(sbio, pbio);
2129 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2131 struct r1conf *conf = mddev->private;
2133 int disks = conf->raid_disks * 2;
2136 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2137 /* ouch - failed to read all of that. */
2138 if (!fix_sync_read_error(r1_bio))
2141 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2142 process_checks(r1_bio);
2147 atomic_set(&r1_bio->remaining, 1);
2148 for (i = 0; i < disks ; i++) {
2149 wbio = r1_bio->bios[i];
2150 if (wbio->bi_end_io == NULL ||
2151 (wbio->bi_end_io == end_sync_read &&
2152 (i == r1_bio->read_disk ||
2153 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2155 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2158 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2159 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2160 wbio->bi_opf |= MD_FAILFAST;
2162 wbio->bi_end_io = end_sync_write;
2163 atomic_inc(&r1_bio->remaining);
2164 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2166 generic_make_request(wbio);
2169 if (atomic_dec_and_test(&r1_bio->remaining)) {
2170 /* if we're here, all write(s) have completed, so clean up */
2171 int s = r1_bio->sectors;
2172 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2173 test_bit(R1BIO_WriteError, &r1_bio->state))
2174 reschedule_retry(r1_bio);
2177 md_done_sync(mddev, s, 1);
2183 * This is a kernel thread which:
2185 * 1. Retries failed read operations on working mirrors.
2186 * 2. Updates the raid superblock when problems encounter.
2187 * 3. Performs writes following reads for array synchronising.
2190 static void fix_read_error(struct r1conf *conf, int read_disk,
2191 sector_t sect, int sectors)
2193 struct mddev *mddev = conf->mddev;
2199 struct md_rdev *rdev;
2201 if (s > (PAGE_SIZE>>9))
2209 rdev = rcu_dereference(conf->mirrors[d].rdev);
2211 (test_bit(In_sync, &rdev->flags) ||
2212 (!test_bit(Faulty, &rdev->flags) &&
2213 rdev->recovery_offset >= sect + s)) &&
2214 is_badblock(rdev, sect, s,
2215 &first_bad, &bad_sectors) == 0) {
2216 atomic_inc(&rdev->nr_pending);
2218 if (sync_page_io(rdev, sect, s<<9,
2219 conf->tmppage, REQ_OP_READ, 0, false))
2221 rdev_dec_pending(rdev, mddev);
2227 if (d == conf->raid_disks * 2)
2229 } while (!success && d != read_disk);
2232 /* Cannot read from anywhere - mark it bad */
2233 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2234 if (!rdev_set_badblocks(rdev, sect, s, 0))
2235 md_error(mddev, rdev);
2238 /* write it back and re-read */
2240 while (d != read_disk) {
2242 d = conf->raid_disks * 2;
2245 rdev = rcu_dereference(conf->mirrors[d].rdev);
2247 !test_bit(Faulty, &rdev->flags)) {
2248 atomic_inc(&rdev->nr_pending);
2250 r1_sync_page_io(rdev, sect, s,
2251 conf->tmppage, WRITE);
2252 rdev_dec_pending(rdev, mddev);
2257 while (d != read_disk) {
2258 char b[BDEVNAME_SIZE];
2260 d = conf->raid_disks * 2;
2263 rdev = rcu_dereference(conf->mirrors[d].rdev);
2265 !test_bit(Faulty, &rdev->flags)) {
2266 atomic_inc(&rdev->nr_pending);
2268 if (r1_sync_page_io(rdev, sect, s,
2269 conf->tmppage, READ)) {
2270 atomic_add(s, &rdev->corrected_errors);
2271 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2273 (unsigned long long)(sect +
2275 bdevname(rdev->bdev, b));
2277 rdev_dec_pending(rdev, mddev);
2286 static int narrow_write_error(struct r1bio *r1_bio, int i)
2288 struct mddev *mddev = r1_bio->mddev;
2289 struct r1conf *conf = mddev->private;
2290 struct md_rdev *rdev = conf->mirrors[i].rdev;
2292 /* bio has the data to be written to device 'i' where
2293 * we just recently had a write error.
2294 * We repeatedly clone the bio and trim down to one block,
2295 * then try the write. Where the write fails we record
2297 * It is conceivable that the bio doesn't exactly align with
2298 * blocks. We must handle this somehow.
2300 * We currently own a reference on the rdev.
2306 int sect_to_write = r1_bio->sectors;
2309 if (rdev->badblocks.shift < 0)
2312 block_sectors = roundup(1 << rdev->badblocks.shift,
2313 bdev_logical_block_size(rdev->bdev) >> 9);
2314 sector = r1_bio->sector;
2315 sectors = ((sector + block_sectors)
2316 & ~(sector_t)(block_sectors - 1))
2319 while (sect_to_write) {
2321 if (sectors > sect_to_write)
2322 sectors = sect_to_write;
2323 /* Write at 'sector' for 'sectors'*/
2325 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2326 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2330 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2334 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2335 wbio->bi_iter.bi_sector = r1_bio->sector;
2336 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2338 bio_trim(wbio, sector - r1_bio->sector, sectors);
2339 wbio->bi_iter.bi_sector += rdev->data_offset;
2340 bio_set_dev(wbio, rdev->bdev);
2342 if (submit_bio_wait(wbio) < 0)
2344 ok = rdev_set_badblocks(rdev, sector,
2349 sect_to_write -= sectors;
2351 sectors = block_sectors;
2356 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2359 int s = r1_bio->sectors;
2360 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2361 struct md_rdev *rdev = conf->mirrors[m].rdev;
2362 struct bio *bio = r1_bio->bios[m];
2363 if (bio->bi_end_io == NULL)
2365 if (!bio->bi_status &&
2366 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2367 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2369 if (bio->bi_status &&
2370 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2371 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2372 md_error(conf->mddev, rdev);
2376 md_done_sync(conf->mddev, s, 1);
2379 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2384 for (m = 0; m < conf->raid_disks * 2 ; m++)
2385 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2386 struct md_rdev *rdev = conf->mirrors[m].rdev;
2387 rdev_clear_badblocks(rdev,
2389 r1_bio->sectors, 0);
2390 rdev_dec_pending(rdev, conf->mddev);
2391 } else if (r1_bio->bios[m] != NULL) {
2392 /* This drive got a write error. We need to
2393 * narrow down and record precise write
2397 if (!narrow_write_error(r1_bio, m)) {
2398 md_error(conf->mddev,
2399 conf->mirrors[m].rdev);
2400 /* an I/O failed, we can't clear the bitmap */
2401 set_bit(R1BIO_Degraded, &r1_bio->state);
2403 rdev_dec_pending(conf->mirrors[m].rdev,
2407 spin_lock_irq(&conf->device_lock);
2408 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2409 idx = sector_to_idx(r1_bio->sector);
2410 atomic_inc(&conf->nr_queued[idx]);
2411 spin_unlock_irq(&conf->device_lock);
2413 * In case freeze_array() is waiting for condition
2414 * get_unqueued_pending() == extra to be true.
2416 wake_up(&conf->wait_barrier);
2417 md_wakeup_thread(conf->mddev->thread);
2419 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2420 close_write(r1_bio);
2421 raid_end_bio_io(r1_bio);
2425 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2427 struct mddev *mddev = conf->mddev;
2429 struct md_rdev *rdev;
2430 sector_t bio_sector;
2432 clear_bit(R1BIO_ReadError, &r1_bio->state);
2433 /* we got a read error. Maybe the drive is bad. Maybe just
2434 * the block and we can fix it.
2435 * We freeze all other IO, and try reading the block from
2436 * other devices. When we find one, we re-write
2437 * and check it that fixes the read error.
2438 * This is all done synchronously while the array is
2442 bio = r1_bio->bios[r1_bio->read_disk];
2443 bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2445 r1_bio->bios[r1_bio->read_disk] = NULL;
2447 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2449 && !test_bit(FailFast, &rdev->flags)) {
2450 freeze_array(conf, 1);
2451 fix_read_error(conf, r1_bio->read_disk,
2452 r1_bio->sector, r1_bio->sectors);
2453 unfreeze_array(conf);
2455 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2458 rdev_dec_pending(rdev, conf->mddev);
2459 allow_barrier(conf, r1_bio->sector);
2460 bio = r1_bio->master_bio;
2462 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2464 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2467 static void raid1d(struct md_thread *thread)
2469 struct mddev *mddev = thread->mddev;
2470 struct r1bio *r1_bio;
2471 unsigned long flags;
2472 struct r1conf *conf = mddev->private;
2473 struct list_head *head = &conf->retry_list;
2474 struct blk_plug plug;
2477 md_check_recovery(mddev);
2479 if (!list_empty_careful(&conf->bio_end_io_list) &&
2480 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2482 spin_lock_irqsave(&conf->device_lock, flags);
2483 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2484 list_splice_init(&conf->bio_end_io_list, &tmp);
2485 spin_unlock_irqrestore(&conf->device_lock, flags);
2486 while (!list_empty(&tmp)) {
2487 r1_bio = list_first_entry(&tmp, struct r1bio,
2489 list_del(&r1_bio->retry_list);
2490 idx = sector_to_idx(r1_bio->sector);
2491 atomic_dec(&conf->nr_queued[idx]);
2492 if (mddev->degraded)
2493 set_bit(R1BIO_Degraded, &r1_bio->state);
2494 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2495 close_write(r1_bio);
2496 raid_end_bio_io(r1_bio);
2500 blk_start_plug(&plug);
2503 flush_pending_writes(conf);
2505 spin_lock_irqsave(&conf->device_lock, flags);
2506 if (list_empty(head)) {
2507 spin_unlock_irqrestore(&conf->device_lock, flags);
2510 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2511 list_del(head->prev);
2512 idx = sector_to_idx(r1_bio->sector);
2513 atomic_dec(&conf->nr_queued[idx]);
2514 spin_unlock_irqrestore(&conf->device_lock, flags);
2516 mddev = r1_bio->mddev;
2517 conf = mddev->private;
2518 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2519 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2520 test_bit(R1BIO_WriteError, &r1_bio->state))
2521 handle_sync_write_finished(conf, r1_bio);
2523 sync_request_write(mddev, r1_bio);
2524 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2525 test_bit(R1BIO_WriteError, &r1_bio->state))
2526 handle_write_finished(conf, r1_bio);
2527 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2528 handle_read_error(conf, r1_bio);
2533 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2534 md_check_recovery(mddev);
2536 blk_finish_plug(&plug);
2539 static int init_resync(struct r1conf *conf)
2543 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2544 BUG_ON(conf->r1buf_pool);
2545 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2547 if (!conf->r1buf_pool)
2552 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2554 struct r1bio *r1bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2555 struct resync_pages *rps;
2559 for (i = conf->poolinfo->raid_disks; i--; ) {
2560 bio = r1bio->bios[i];
2561 rps = bio->bi_private;
2563 bio->bi_private = rps;
2565 r1bio->master_bio = NULL;
2570 * perform a "sync" on one "block"
2572 * We need to make sure that no normal I/O request - particularly write
2573 * requests - conflict with active sync requests.
2575 * This is achieved by tracking pending requests and a 'barrier' concept
2576 * that can be installed to exclude normal IO requests.
2579 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2582 struct r1conf *conf = mddev->private;
2583 struct r1bio *r1_bio;
2585 sector_t max_sector, nr_sectors;
2589 int write_targets = 0, read_targets = 0;
2590 sector_t sync_blocks;
2591 int still_degraded = 0;
2592 int good_sectors = RESYNC_SECTORS;
2593 int min_bad = 0; /* number of sectors that are bad in all devices */
2594 int idx = sector_to_idx(sector_nr);
2597 if (!conf->r1buf_pool)
2598 if (init_resync(conf))
2601 max_sector = mddev->dev_sectors;
2602 if (sector_nr >= max_sector) {
2603 /* If we aborted, we need to abort the
2604 * sync on the 'current' bitmap chunk (there will
2605 * only be one in raid1 resync.
2606 * We can find the current addess in mddev->curr_resync
2608 if (mddev->curr_resync < max_sector) /* aborted */
2609 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2611 else /* completed sync */
2614 bitmap_close_sync(mddev->bitmap);
2617 if (mddev_is_clustered(mddev)) {
2618 conf->cluster_sync_low = 0;
2619 conf->cluster_sync_high = 0;
2624 if (mddev->bitmap == NULL &&
2625 mddev->recovery_cp == MaxSector &&
2626 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2627 conf->fullsync == 0) {
2629 return max_sector - sector_nr;
2631 /* before building a request, check if we can skip these blocks..
2632 * This call the bitmap_start_sync doesn't actually record anything
2634 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2635 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2636 /* We can skip this block, and probably several more */
2642 * If there is non-resync activity waiting for a turn, then let it
2643 * though before starting on this new sync request.
2645 if (atomic_read(&conf->nr_waiting[idx]))
2646 schedule_timeout_uninterruptible(1);
2648 /* we are incrementing sector_nr below. To be safe, we check against
2649 * sector_nr + two times RESYNC_SECTORS
2652 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2653 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2654 r1_bio = raid1_alloc_init_r1buf(conf);
2656 raise_barrier(conf, sector_nr);
2660 * If we get a correctably read error during resync or recovery,
2661 * we might want to read from a different device. So we
2662 * flag all drives that could conceivably be read from for READ,
2663 * and any others (which will be non-In_sync devices) for WRITE.
2664 * If a read fails, we try reading from something else for which READ
2668 r1_bio->mddev = mddev;
2669 r1_bio->sector = sector_nr;
2671 set_bit(R1BIO_IsSync, &r1_bio->state);
2672 /* make sure good_sectors won't go across barrier unit boundary */
2673 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2675 for (i = 0; i < conf->raid_disks * 2; i++) {
2676 struct md_rdev *rdev;
2677 bio = r1_bio->bios[i];
2679 rdev = rcu_dereference(conf->mirrors[i].rdev);
2681 test_bit(Faulty, &rdev->flags)) {
2682 if (i < conf->raid_disks)
2684 } else if (!test_bit(In_sync, &rdev->flags)) {
2685 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2686 bio->bi_end_io = end_sync_write;
2689 /* may need to read from here */
2690 sector_t first_bad = MaxSector;
2693 if (is_badblock(rdev, sector_nr, good_sectors,
2694 &first_bad, &bad_sectors)) {
2695 if (first_bad > sector_nr)
2696 good_sectors = first_bad - sector_nr;
2698 bad_sectors -= (sector_nr - first_bad);
2700 min_bad > bad_sectors)
2701 min_bad = bad_sectors;
2704 if (sector_nr < first_bad) {
2705 if (test_bit(WriteMostly, &rdev->flags)) {
2712 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2713 bio->bi_end_io = end_sync_read;
2715 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2716 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2717 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2719 * The device is suitable for reading (InSync),
2720 * but has bad block(s) here. Let's try to correct them,
2721 * if we are doing resync or repair. Otherwise, leave
2722 * this device alone for this sync request.
2724 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2725 bio->bi_end_io = end_sync_write;
2729 if (bio->bi_end_io) {
2730 atomic_inc(&rdev->nr_pending);
2731 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2732 bio_set_dev(bio, rdev->bdev);
2733 if (test_bit(FailFast, &rdev->flags))
2734 bio->bi_opf |= MD_FAILFAST;
2740 r1_bio->read_disk = disk;
2742 if (read_targets == 0 && min_bad > 0) {
2743 /* These sectors are bad on all InSync devices, so we
2744 * need to mark them bad on all write targets
2747 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2748 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2749 struct md_rdev *rdev = conf->mirrors[i].rdev;
2750 ok = rdev_set_badblocks(rdev, sector_nr,
2754 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2759 /* Cannot record the badblocks, so need to
2761 * If there are multiple read targets, could just
2762 * fail the really bad ones ???
2764 conf->recovery_disabled = mddev->recovery_disabled;
2765 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2771 if (min_bad > 0 && min_bad < good_sectors) {
2772 /* only resync enough to reach the next bad->good
2774 good_sectors = min_bad;
2777 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2778 /* extra read targets are also write targets */
2779 write_targets += read_targets-1;
2781 if (write_targets == 0 || read_targets == 0) {
2782 /* There is nowhere to write, so all non-sync
2783 * drives must be failed - so we are finished
2787 max_sector = sector_nr + min_bad;
2788 rv = max_sector - sector_nr;
2794 if (max_sector > mddev->resync_max)
2795 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2796 if (max_sector > sector_nr + good_sectors)
2797 max_sector = sector_nr + good_sectors;
2802 int len = PAGE_SIZE;
2803 if (sector_nr + (len>>9) > max_sector)
2804 len = (max_sector - sector_nr) << 9;
2807 if (sync_blocks == 0) {
2808 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2809 &sync_blocks, still_degraded) &&
2811 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2813 if ((len >> 9) > sync_blocks)
2814 len = sync_blocks<<9;
2817 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2818 struct resync_pages *rp;
2820 bio = r1_bio->bios[i];
2821 rp = get_resync_pages(bio);
2822 if (bio->bi_end_io) {
2823 page = resync_fetch_page(rp, page_idx);
2826 * won't fail because the vec table is big
2827 * enough to hold all these pages
2829 bio_add_page(bio, page, len, 0);
2832 nr_sectors += len>>9;
2833 sector_nr += len>>9;
2834 sync_blocks -= (len>>9);
2835 } while (++page_idx < RESYNC_PAGES);
2837 r1_bio->sectors = nr_sectors;
2839 if (mddev_is_clustered(mddev) &&
2840 conf->cluster_sync_high < sector_nr + nr_sectors) {
2841 conf->cluster_sync_low = mddev->curr_resync_completed;
2842 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2843 /* Send resync message */
2844 md_cluster_ops->resync_info_update(mddev,
2845 conf->cluster_sync_low,
2846 conf->cluster_sync_high);
2849 /* For a user-requested sync, we read all readable devices and do a
2852 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2853 atomic_set(&r1_bio->remaining, read_targets);
2854 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2855 bio = r1_bio->bios[i];
2856 if (bio->bi_end_io == end_sync_read) {
2858 md_sync_acct_bio(bio, nr_sectors);
2859 if (read_targets == 1)
2860 bio->bi_opf &= ~MD_FAILFAST;
2861 generic_make_request(bio);
2865 atomic_set(&r1_bio->remaining, 1);
2866 bio = r1_bio->bios[r1_bio->read_disk];
2867 md_sync_acct_bio(bio, nr_sectors);
2868 if (read_targets == 1)
2869 bio->bi_opf &= ~MD_FAILFAST;
2870 generic_make_request(bio);
2876 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2881 return mddev->dev_sectors;
2884 static struct r1conf *setup_conf(struct mddev *mddev)
2886 struct r1conf *conf;
2888 struct raid1_info *disk;
2889 struct md_rdev *rdev;
2892 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2896 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2897 sizeof(atomic_t), GFP_KERNEL);
2898 if (!conf->nr_pending)
2901 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2902 sizeof(atomic_t), GFP_KERNEL);
2903 if (!conf->nr_waiting)
2906 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2907 sizeof(atomic_t), GFP_KERNEL);
2908 if (!conf->nr_queued)
2911 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2912 sizeof(atomic_t), GFP_KERNEL);
2916 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2917 * mddev->raid_disks * 2,
2922 conf->tmppage = alloc_page(GFP_KERNEL);
2926 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2927 if (!conf->poolinfo)
2929 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2930 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2933 if (!conf->r1bio_pool)
2936 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
2937 if (!conf->bio_split)
2940 conf->poolinfo->mddev = mddev;
2943 spin_lock_init(&conf->device_lock);
2944 rdev_for_each(rdev, mddev) {
2945 int disk_idx = rdev->raid_disk;
2946 if (disk_idx >= mddev->raid_disks
2949 if (test_bit(Replacement, &rdev->flags))
2950 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2952 disk = conf->mirrors + disk_idx;
2957 disk->head_position = 0;
2958 disk->seq_start = MaxSector;
2960 conf->raid_disks = mddev->raid_disks;
2961 conf->mddev = mddev;
2962 INIT_LIST_HEAD(&conf->retry_list);
2963 INIT_LIST_HEAD(&conf->bio_end_io_list);
2965 spin_lock_init(&conf->resync_lock);
2966 init_waitqueue_head(&conf->wait_barrier);
2968 bio_list_init(&conf->pending_bio_list);
2969 conf->pending_count = 0;
2970 conf->recovery_disabled = mddev->recovery_disabled - 1;
2973 for (i = 0; i < conf->raid_disks * 2; i++) {
2975 disk = conf->mirrors + i;
2977 if (i < conf->raid_disks &&
2978 disk[conf->raid_disks].rdev) {
2979 /* This slot has a replacement. */
2981 /* No original, just make the replacement
2982 * a recovering spare
2985 disk[conf->raid_disks].rdev;
2986 disk[conf->raid_disks].rdev = NULL;
2987 } else if (!test_bit(In_sync, &disk->rdev->flags))
2988 /* Original is not in_sync - bad */
2993 !test_bit(In_sync, &disk->rdev->flags)) {
2994 disk->head_position = 0;
2996 (disk->rdev->saved_raid_disk < 0))
3002 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3010 mempool_destroy(conf->r1bio_pool);
3011 kfree(conf->mirrors);
3012 safe_put_page(conf->tmppage);
3013 kfree(conf->poolinfo);
3014 kfree(conf->nr_pending);
3015 kfree(conf->nr_waiting);
3016 kfree(conf->nr_queued);
3017 kfree(conf->barrier);
3018 if (conf->bio_split)
3019 bioset_free(conf->bio_split);
3022 return ERR_PTR(err);
3025 static void raid1_free(struct mddev *mddev, void *priv);
3026 static int raid1_run(struct mddev *mddev)
3028 struct r1conf *conf;
3030 struct md_rdev *rdev;
3032 bool discard_supported = false;
3034 if (mddev->level != 1) {
3035 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3036 mdname(mddev), mddev->level);
3039 if (mddev->reshape_position != MaxSector) {
3040 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3044 if (mddev_init_writes_pending(mddev) < 0)
3047 * copy the already verified devices into our private RAID1
3048 * bookkeeping area. [whatever we allocate in run(),
3049 * should be freed in raid1_free()]
3051 if (mddev->private == NULL)
3052 conf = setup_conf(mddev);
3054 conf = mddev->private;
3057 return PTR_ERR(conf);
3060 blk_queue_max_write_same_sectors(mddev->queue, 0);
3061 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3064 rdev_for_each(rdev, mddev) {
3065 if (!mddev->gendisk)
3067 disk_stack_limits(mddev->gendisk, rdev->bdev,
3068 rdev->data_offset << 9);
3069 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3070 discard_supported = true;
3073 mddev->degraded = 0;
3074 for (i=0; i < conf->raid_disks; i++)
3075 if (conf->mirrors[i].rdev == NULL ||
3076 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3077 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3080 if (conf->raid_disks - mddev->degraded == 1)
3081 mddev->recovery_cp = MaxSector;
3083 if (mddev->recovery_cp != MaxSector)
3084 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3086 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3087 mdname(mddev), mddev->raid_disks - mddev->degraded,
3091 * Ok, everything is just fine now
3093 mddev->thread = conf->thread;
3094 conf->thread = NULL;
3095 mddev->private = conf;
3096 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3098 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3101 if (discard_supported)
3102 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3105 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3109 ret = md_integrity_register(mddev);
3111 md_unregister_thread(&mddev->thread);
3112 raid1_free(mddev, conf);
3117 static void raid1_free(struct mddev *mddev, void *priv)
3119 struct r1conf *conf = priv;
3121 mempool_destroy(conf->r1bio_pool);
3122 kfree(conf->mirrors);
3123 safe_put_page(conf->tmppage);
3124 kfree(conf->poolinfo);
3125 kfree(conf->nr_pending);
3126 kfree(conf->nr_waiting);
3127 kfree(conf->nr_queued);
3128 kfree(conf->barrier);
3129 if (conf->bio_split)
3130 bioset_free(conf->bio_split);
3134 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3136 /* no resync is happening, and there is enough space
3137 * on all devices, so we can resize.
3138 * We need to make sure resync covers any new space.
3139 * If the array is shrinking we should possibly wait until
3140 * any io in the removed space completes, but it hardly seems
3143 sector_t newsize = raid1_size(mddev, sectors, 0);
3144 if (mddev->external_size &&
3145 mddev->array_sectors > newsize)
3147 if (mddev->bitmap) {
3148 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3152 md_set_array_sectors(mddev, newsize);
3153 if (sectors > mddev->dev_sectors &&
3154 mddev->recovery_cp > mddev->dev_sectors) {
3155 mddev->recovery_cp = mddev->dev_sectors;
3156 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3158 mddev->dev_sectors = sectors;
3159 mddev->resync_max_sectors = sectors;
3163 static int raid1_reshape(struct mddev *mddev)
3166 * 1/ resize the r1bio_pool
3167 * 2/ resize conf->mirrors
3169 * We allocate a new r1bio_pool if we can.
3170 * Then raise a device barrier and wait until all IO stops.
3171 * Then resize conf->mirrors and swap in the new r1bio pool.
3173 * At the same time, we "pack" the devices so that all the missing
3174 * devices have the higher raid_disk numbers.
3176 mempool_t *newpool, *oldpool;
3177 struct pool_info *newpoolinfo;
3178 struct raid1_info *newmirrors;
3179 struct r1conf *conf = mddev->private;
3180 int cnt, raid_disks;
3181 unsigned long flags;
3184 /* Cannot change chunk_size, layout, or level */
3185 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3186 mddev->layout != mddev->new_layout ||
3187 mddev->level != mddev->new_level) {
3188 mddev->new_chunk_sectors = mddev->chunk_sectors;
3189 mddev->new_layout = mddev->layout;
3190 mddev->new_level = mddev->level;
3194 if (!mddev_is_clustered(mddev))
3195 md_allow_write(mddev);
3197 raid_disks = mddev->raid_disks + mddev->delta_disks;
3199 if (raid_disks < conf->raid_disks) {
3201 for (d= 0; d < conf->raid_disks; d++)
3202 if (conf->mirrors[d].rdev)
3204 if (cnt > raid_disks)
3208 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3211 newpoolinfo->mddev = mddev;
3212 newpoolinfo->raid_disks = raid_disks * 2;
3214 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3215 r1bio_pool_free, newpoolinfo);
3220 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3224 mempool_destroy(newpool);
3228 freeze_array(conf, 0);
3230 /* ok, everything is stopped */
3231 oldpool = conf->r1bio_pool;
3232 conf->r1bio_pool = newpool;
3234 for (d = d2 = 0; d < conf->raid_disks; d++) {
3235 struct md_rdev *rdev = conf->mirrors[d].rdev;
3236 if (rdev && rdev->raid_disk != d2) {
3237 sysfs_unlink_rdev(mddev, rdev);
3238 rdev->raid_disk = d2;
3239 sysfs_unlink_rdev(mddev, rdev);
3240 if (sysfs_link_rdev(mddev, rdev))
3241 pr_warn("md/raid1:%s: cannot register rd%d\n",
3242 mdname(mddev), rdev->raid_disk);
3245 newmirrors[d2++].rdev = rdev;
3247 kfree(conf->mirrors);
3248 conf->mirrors = newmirrors;
3249 kfree(conf->poolinfo);
3250 conf->poolinfo = newpoolinfo;
3252 spin_lock_irqsave(&conf->device_lock, flags);
3253 mddev->degraded += (raid_disks - conf->raid_disks);
3254 spin_unlock_irqrestore(&conf->device_lock, flags);
3255 conf->raid_disks = mddev->raid_disks = raid_disks;
3256 mddev->delta_disks = 0;
3258 unfreeze_array(conf);
3260 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3261 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3262 md_wakeup_thread(mddev->thread);
3264 mempool_destroy(oldpool);
3268 static void raid1_quiesce(struct mddev *mddev, int state)
3270 struct r1conf *conf = mddev->private;
3273 case 2: /* wake for suspend */
3274 wake_up(&conf->wait_barrier);
3277 freeze_array(conf, 0);
3280 unfreeze_array(conf);
3285 static void *raid1_takeover(struct mddev *mddev)
3287 /* raid1 can take over:
3288 * raid5 with 2 devices, any layout or chunk size
3290 if (mddev->level == 5 && mddev->raid_disks == 2) {
3291 struct r1conf *conf;
3292 mddev->new_level = 1;
3293 mddev->new_layout = 0;
3294 mddev->new_chunk_sectors = 0;
3295 conf = setup_conf(mddev);
3296 if (!IS_ERR(conf)) {
3297 /* Array must appear to be quiesced */
3298 conf->array_frozen = 1;
3299 mddev_clear_unsupported_flags(mddev,
3300 UNSUPPORTED_MDDEV_FLAGS);
3304 return ERR_PTR(-EINVAL);
3307 static struct md_personality raid1_personality =
3311 .owner = THIS_MODULE,
3312 .make_request = raid1_make_request,
3315 .status = raid1_status,
3316 .error_handler = raid1_error,
3317 .hot_add_disk = raid1_add_disk,
3318 .hot_remove_disk= raid1_remove_disk,
3319 .spare_active = raid1_spare_active,
3320 .sync_request = raid1_sync_request,
3321 .resize = raid1_resize,
3323 .check_reshape = raid1_reshape,
3324 .quiesce = raid1_quiesce,
3325 .takeover = raid1_takeover,
3326 .congested = raid1_congested,
3329 static int __init raid_init(void)
3331 return register_md_personality(&raid1_personality);
3334 static void raid_exit(void)
3336 unregister_md_personality(&raid1_personality);
3339 module_init(raid_init);
3340 module_exit(raid_exit);
3341 MODULE_LICENSE("GPL");
3342 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3343 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3344 MODULE_ALIAS("md-raid1");
3345 MODULE_ALIAS("md-level-1");
3347 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);