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);
441 * When the device is faulty, it is not necessary to
442 * handle write error.
443 * For failfast, this is the only remaining device,
444 * We need to retry the write without FailFast.
446 if (!test_bit(Faulty, &rdev->flags))
447 set_bit(R1BIO_WriteError, &r1_bio->state);
449 /* Fail the request */
450 set_bit(R1BIO_Degraded, &r1_bio->state);
451 /* Finished with this branch */
452 r1_bio->bios[mirror] = NULL;
457 * Set R1BIO_Uptodate in our master bio, so that we
458 * will return a good error code for to the higher
459 * levels even if IO on some other mirrored buffer
462 * The 'master' represents the composite IO operation
463 * to user-side. So if something waits for IO, then it
464 * will wait for the 'master' bio.
469 r1_bio->bios[mirror] = NULL;
472 * Do not set R1BIO_Uptodate if the current device is
473 * rebuilding or Faulty. This is because we cannot use
474 * such device for properly reading the data back (we could
475 * potentially use it, if the current write would have felt
476 * before rdev->recovery_offset, but for simplicity we don't
479 if (test_bit(In_sync, &rdev->flags) &&
480 !test_bit(Faulty, &rdev->flags))
481 set_bit(R1BIO_Uptodate, &r1_bio->state);
483 /* Maybe we can clear some bad blocks. */
484 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
485 &first_bad, &bad_sectors) && !discard_error) {
486 r1_bio->bios[mirror] = IO_MADE_GOOD;
487 set_bit(R1BIO_MadeGood, &r1_bio->state);
492 if (test_bit(WriteMostly, &rdev->flags))
493 atomic_dec(&r1_bio->behind_remaining);
496 * In behind mode, we ACK the master bio once the I/O
497 * has safely reached all non-writemostly
498 * disks. Setting the Returned bit ensures that this
499 * gets done only once -- we don't ever want to return
500 * -EIO here, instead we'll wait
502 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
503 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
504 /* Maybe we can return now */
505 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
506 struct bio *mbio = r1_bio->master_bio;
507 pr_debug("raid1: behind end write sectors"
509 (unsigned long long) mbio->bi_iter.bi_sector,
510 (unsigned long long) bio_end_sector(mbio) - 1);
511 call_bio_endio(r1_bio);
515 if (r1_bio->bios[mirror] == NULL)
516 rdev_dec_pending(rdev, conf->mddev);
519 * Let's see if all mirrored write operations have finished
522 r1_bio_write_done(r1_bio);
528 static sector_t align_to_barrier_unit_end(sector_t start_sector,
533 WARN_ON(sectors == 0);
535 * len is the number of sectors from start_sector to end of the
536 * barrier unit which start_sector belongs to.
538 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
548 * This routine returns the disk from which the requested read should
549 * be done. There is a per-array 'next expected sequential IO' sector
550 * number - if this matches on the next IO then we use the last disk.
551 * There is also a per-disk 'last know head position' sector that is
552 * maintained from IRQ contexts, both the normal and the resync IO
553 * completion handlers update this position correctly. If there is no
554 * perfect sequential match then we pick the disk whose head is closest.
556 * If there are 2 mirrors in the same 2 devices, performance degrades
557 * because position is mirror, not device based.
559 * The rdev for the device selected will have nr_pending incremented.
561 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
563 const sector_t this_sector = r1_bio->sector;
565 int best_good_sectors;
566 int best_disk, best_dist_disk, best_pending_disk;
570 unsigned int min_pending;
571 struct md_rdev *rdev;
573 int choose_next_idle;
577 * Check if we can balance. We can balance on the whole
578 * device if no resync is going on, or below the resync window.
579 * We take the first readable disk when above the resync window.
582 sectors = r1_bio->sectors;
585 best_dist = MaxSector;
586 best_pending_disk = -1;
587 min_pending = UINT_MAX;
588 best_good_sectors = 0;
590 choose_next_idle = 0;
591 clear_bit(R1BIO_FailFast, &r1_bio->state);
593 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
594 (mddev_is_clustered(conf->mddev) &&
595 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
596 this_sector + sectors)))
601 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
605 unsigned int pending;
608 rdev = rcu_dereference(conf->mirrors[disk].rdev);
609 if (r1_bio->bios[disk] == IO_BLOCKED
611 || test_bit(Faulty, &rdev->flags))
613 if (!test_bit(In_sync, &rdev->flags) &&
614 rdev->recovery_offset < this_sector + sectors)
616 if (test_bit(WriteMostly, &rdev->flags)) {
617 /* Don't balance among write-mostly, just
618 * use the first as a last resort */
619 if (best_dist_disk < 0) {
620 if (is_badblock(rdev, this_sector, sectors,
621 &first_bad, &bad_sectors)) {
622 if (first_bad <= this_sector)
623 /* Cannot use this */
625 best_good_sectors = first_bad - this_sector;
627 best_good_sectors = sectors;
628 best_dist_disk = disk;
629 best_pending_disk = disk;
633 /* This is a reasonable device to use. It might
636 if (is_badblock(rdev, this_sector, sectors,
637 &first_bad, &bad_sectors)) {
638 if (best_dist < MaxSector)
639 /* already have a better device */
641 if (first_bad <= this_sector) {
642 /* cannot read here. If this is the 'primary'
643 * device, then we must not read beyond
644 * bad_sectors from another device..
646 bad_sectors -= (this_sector - first_bad);
647 if (choose_first && sectors > bad_sectors)
648 sectors = bad_sectors;
649 if (best_good_sectors > sectors)
650 best_good_sectors = sectors;
653 sector_t good_sectors = first_bad - this_sector;
654 if (good_sectors > best_good_sectors) {
655 best_good_sectors = good_sectors;
663 if ((sectors > best_good_sectors) && (best_disk >= 0))
665 best_good_sectors = sectors;
669 /* At least two disks to choose from so failfast is OK */
670 set_bit(R1BIO_FailFast, &r1_bio->state);
672 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
673 has_nonrot_disk |= nonrot;
674 pending = atomic_read(&rdev->nr_pending);
675 dist = abs(this_sector - conf->mirrors[disk].head_position);
680 /* Don't change to another disk for sequential reads */
681 if (conf->mirrors[disk].next_seq_sect == this_sector
683 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
684 struct raid1_info *mirror = &conf->mirrors[disk];
688 * If buffered sequential IO size exceeds optimal
689 * iosize, check if there is idle disk. If yes, choose
690 * the idle disk. read_balance could already choose an
691 * idle disk before noticing it's a sequential IO in
692 * this disk. This doesn't matter because this disk
693 * will idle, next time it will be utilized after the
694 * first disk has IO size exceeds optimal iosize. In
695 * this way, iosize of the first disk will be optimal
696 * iosize at least. iosize of the second disk might be
697 * small, but not a big deal since when the second disk
698 * starts IO, the first disk is likely still busy.
700 if (nonrot && opt_iosize > 0 &&
701 mirror->seq_start != MaxSector &&
702 mirror->next_seq_sect > opt_iosize &&
703 mirror->next_seq_sect - opt_iosize >=
705 choose_next_idle = 1;
711 if (choose_next_idle)
714 if (min_pending > pending) {
715 min_pending = pending;
716 best_pending_disk = disk;
719 if (dist < best_dist) {
721 best_dist_disk = disk;
726 * If all disks are rotational, choose the closest disk. If any disk is
727 * non-rotational, choose the disk with less pending request even the
728 * disk is rotational, which might/might not be optimal for raids with
729 * mixed ratation/non-rotational disks depending on workload.
731 if (best_disk == -1) {
732 if (has_nonrot_disk || min_pending == 0)
733 best_disk = best_pending_disk;
735 best_disk = best_dist_disk;
738 if (best_disk >= 0) {
739 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
742 atomic_inc(&rdev->nr_pending);
743 sectors = best_good_sectors;
745 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
746 conf->mirrors[best_disk].seq_start = this_sector;
748 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
751 *max_sectors = sectors;
756 static int raid1_congested(struct mddev *mddev, int bits)
758 struct r1conf *conf = mddev->private;
761 if ((bits & (1 << WB_async_congested)) &&
762 conf->pending_count >= max_queued_requests)
766 for (i = 0; i < conf->raid_disks * 2; i++) {
767 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
768 if (rdev && !test_bit(Faulty, &rdev->flags)) {
769 struct request_queue *q = bdev_get_queue(rdev->bdev);
773 /* Note the '|| 1' - when read_balance prefers
774 * non-congested targets, it can be removed
776 if ((bits & (1 << WB_async_congested)) || 1)
777 ret |= bdi_congested(q->backing_dev_info, bits);
779 ret &= bdi_congested(q->backing_dev_info, bits);
786 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
788 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
789 bitmap_unplug(conf->mddev->bitmap);
790 wake_up(&conf->wait_barrier);
792 while (bio) { /* submit pending writes */
793 struct bio *next = bio->bi_next;
794 struct md_rdev *rdev = (void *)bio->bi_disk;
796 bio_set_dev(bio, rdev->bdev);
797 if (test_bit(Faulty, &rdev->flags)) {
799 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
800 !blk_queue_discard(bio->bi_disk->queue)))
804 generic_make_request(bio);
809 static void flush_pending_writes(struct r1conf *conf)
811 /* Any writes that have been queued but are awaiting
812 * bitmap updates get flushed here.
814 spin_lock_irq(&conf->device_lock);
816 if (conf->pending_bio_list.head) {
817 struct blk_plug plug;
820 bio = bio_list_get(&conf->pending_bio_list);
821 conf->pending_count = 0;
822 spin_unlock_irq(&conf->device_lock);
823 blk_start_plug(&plug);
824 flush_bio_list(conf, bio);
825 blk_finish_plug(&plug);
827 spin_unlock_irq(&conf->device_lock);
831 * Sometimes we need to suspend IO while we do something else,
832 * either some resync/recovery, or reconfigure the array.
833 * To do this we raise a 'barrier'.
834 * The 'barrier' is a counter that can be raised multiple times
835 * to count how many activities are happening which preclude
837 * We can only raise the barrier if there is no pending IO.
838 * i.e. if nr_pending == 0.
839 * We choose only to raise the barrier if no-one is waiting for the
840 * barrier to go down. This means that as soon as an IO request
841 * is ready, no other operations which require a barrier will start
842 * until the IO request has had a chance.
844 * So: regular IO calls 'wait_barrier'. When that returns there
845 * is no backgroup IO happening, It must arrange to call
846 * allow_barrier when it has finished its IO.
847 * backgroup IO calls must call raise_barrier. Once that returns
848 * there is no normal IO happeing. It must arrange to call
849 * lower_barrier when the particular background IO completes.
851 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
853 int idx = sector_to_idx(sector_nr);
855 spin_lock_irq(&conf->resync_lock);
857 /* Wait until no block IO is waiting */
858 wait_event_lock_irq(conf->wait_barrier,
859 !atomic_read(&conf->nr_waiting[idx]),
862 /* block any new IO from starting */
863 atomic_inc(&conf->barrier[idx]);
865 * In raise_barrier() we firstly increase conf->barrier[idx] then
866 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
867 * increase conf->nr_pending[idx] then check conf->barrier[idx].
868 * A memory barrier here to make sure conf->nr_pending[idx] won't
869 * be fetched before conf->barrier[idx] is increased. Otherwise
870 * there will be a race between raise_barrier() and _wait_barrier().
872 smp_mb__after_atomic();
874 /* For these conditions we must wait:
875 * A: while the array is in frozen state
876 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
877 * existing in corresponding I/O barrier bucket.
878 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
879 * max resync count which allowed on current I/O barrier bucket.
881 wait_event_lock_irq(conf->wait_barrier,
882 !conf->array_frozen &&
883 !atomic_read(&conf->nr_pending[idx]) &&
884 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH,
887 atomic_inc(&conf->nr_sync_pending);
888 spin_unlock_irq(&conf->resync_lock);
891 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
893 int idx = sector_to_idx(sector_nr);
895 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
897 atomic_dec(&conf->barrier[idx]);
898 atomic_dec(&conf->nr_sync_pending);
899 wake_up(&conf->wait_barrier);
902 static void _wait_barrier(struct r1conf *conf, int idx)
905 * We need to increase conf->nr_pending[idx] very early here,
906 * then raise_barrier() can be blocked when it waits for
907 * conf->nr_pending[idx] to be 0. Then we can avoid holding
908 * conf->resync_lock when there is no barrier raised in same
909 * barrier unit bucket. Also if the array is frozen, I/O
910 * should be blocked until array is unfrozen.
912 atomic_inc(&conf->nr_pending[idx]);
914 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
915 * check conf->barrier[idx]. In raise_barrier() we firstly increase
916 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
917 * barrier is necessary here to make sure conf->barrier[idx] won't be
918 * fetched before conf->nr_pending[idx] is increased. Otherwise there
919 * will be a race between _wait_barrier() and raise_barrier().
921 smp_mb__after_atomic();
924 * Don't worry about checking two atomic_t variables at same time
925 * here. If during we check conf->barrier[idx], the array is
926 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
927 * 0, it is safe to return and make the I/O continue. Because the
928 * array is frozen, all I/O returned here will eventually complete
929 * or be queued, no race will happen. See code comment in
932 if (!READ_ONCE(conf->array_frozen) &&
933 !atomic_read(&conf->barrier[idx]))
937 * After holding conf->resync_lock, conf->nr_pending[idx]
938 * should be decreased before waiting for barrier to drop.
939 * Otherwise, we may encounter a race condition because
940 * raise_barrer() might be waiting for conf->nr_pending[idx]
941 * to be 0 at same time.
943 spin_lock_irq(&conf->resync_lock);
944 atomic_inc(&conf->nr_waiting[idx]);
945 atomic_dec(&conf->nr_pending[idx]);
947 * In case freeze_array() is waiting for
948 * get_unqueued_pending() == extra
950 wake_up(&conf->wait_barrier);
951 /* Wait for the barrier in same barrier unit bucket to drop. */
952 wait_event_lock_irq(conf->wait_barrier,
953 !conf->array_frozen &&
954 !atomic_read(&conf->barrier[idx]),
956 atomic_inc(&conf->nr_pending[idx]);
957 atomic_dec(&conf->nr_waiting[idx]);
958 spin_unlock_irq(&conf->resync_lock);
961 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
963 int idx = sector_to_idx(sector_nr);
966 * Very similar to _wait_barrier(). The difference is, for read
967 * I/O we don't need wait for sync I/O, but if the whole array
968 * is frozen, the read I/O still has to wait until the array is
969 * unfrozen. Since there is no ordering requirement with
970 * conf->barrier[idx] here, memory barrier is unnecessary as well.
972 atomic_inc(&conf->nr_pending[idx]);
974 if (!READ_ONCE(conf->array_frozen))
977 spin_lock_irq(&conf->resync_lock);
978 atomic_inc(&conf->nr_waiting[idx]);
979 atomic_dec(&conf->nr_pending[idx]);
981 * In case freeze_array() is waiting for
982 * get_unqueued_pending() == extra
984 wake_up(&conf->wait_barrier);
985 /* Wait for array to be unfrozen */
986 wait_event_lock_irq(conf->wait_barrier,
989 atomic_inc(&conf->nr_pending[idx]);
990 atomic_dec(&conf->nr_waiting[idx]);
991 spin_unlock_irq(&conf->resync_lock);
994 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
996 int idx = sector_to_idx(sector_nr);
998 _wait_barrier(conf, idx);
1001 static void _allow_barrier(struct r1conf *conf, int idx)
1003 atomic_dec(&conf->nr_pending[idx]);
1004 wake_up(&conf->wait_barrier);
1007 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1009 int idx = sector_to_idx(sector_nr);
1011 _allow_barrier(conf, idx);
1014 /* conf->resync_lock should be held */
1015 static int get_unqueued_pending(struct r1conf *conf)
1019 ret = atomic_read(&conf->nr_sync_pending);
1020 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1021 ret += atomic_read(&conf->nr_pending[idx]) -
1022 atomic_read(&conf->nr_queued[idx]);
1027 static void freeze_array(struct r1conf *conf, int extra)
1029 /* Stop sync I/O and normal I/O and wait for everything to
1031 * This is called in two situations:
1032 * 1) management command handlers (reshape, remove disk, quiesce).
1033 * 2) one normal I/O request failed.
1035 * After array_frozen is set to 1, new sync IO will be blocked at
1036 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1037 * or wait_read_barrier(). The flying I/Os will either complete or be
1038 * queued. When everything goes quite, there are only queued I/Os left.
1040 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1041 * barrier bucket index which this I/O request hits. When all sync and
1042 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1043 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1044 * in handle_read_error(), we may call freeze_array() before trying to
1045 * fix the read error. In this case, the error read I/O is not queued,
1046 * so get_unqueued_pending() == 1.
1048 * Therefore before this function returns, we need to wait until
1049 * get_unqueued_pendings(conf) gets equal to extra. For
1050 * normal I/O context, extra is 1, in rested situations extra is 0.
1052 spin_lock_irq(&conf->resync_lock);
1053 conf->array_frozen = 1;
1054 raid1_log(conf->mddev, "wait freeze");
1055 wait_event_lock_irq_cmd(
1057 get_unqueued_pending(conf) == extra,
1059 flush_pending_writes(conf));
1060 spin_unlock_irq(&conf->resync_lock);
1062 static void unfreeze_array(struct r1conf *conf)
1064 /* reverse the effect of the freeze */
1065 spin_lock_irq(&conf->resync_lock);
1066 conf->array_frozen = 0;
1067 spin_unlock_irq(&conf->resync_lock);
1068 wake_up(&conf->wait_barrier);
1071 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1074 int size = bio->bi_iter.bi_size;
1075 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1077 struct bio *behind_bio = NULL;
1079 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1083 /* discard op, we don't support writezero/writesame yet */
1084 if (!bio_has_data(bio)) {
1085 behind_bio->bi_iter.bi_size = size;
1089 while (i < vcnt && size) {
1091 int len = min_t(int, PAGE_SIZE, size);
1093 page = alloc_page(GFP_NOIO);
1094 if (unlikely(!page))
1097 bio_add_page(behind_bio, page, len, 0);
1103 bio_copy_data(behind_bio, bio);
1105 r1_bio->behind_master_bio = behind_bio;;
1106 set_bit(R1BIO_BehindIO, &r1_bio->state);
1111 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1112 bio->bi_iter.bi_size);
1113 bio_free_pages(behind_bio);
1114 bio_put(behind_bio);
1117 struct raid1_plug_cb {
1118 struct blk_plug_cb cb;
1119 struct bio_list pending;
1123 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1125 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1127 struct mddev *mddev = plug->cb.data;
1128 struct r1conf *conf = mddev->private;
1131 if (from_schedule || current->bio_list) {
1132 spin_lock_irq(&conf->device_lock);
1133 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1134 conf->pending_count += plug->pending_cnt;
1135 spin_unlock_irq(&conf->device_lock);
1136 wake_up(&conf->wait_barrier);
1137 md_wakeup_thread(mddev->thread);
1142 /* we aren't scheduling, so we can do the write-out directly. */
1143 bio = bio_list_get(&plug->pending);
1144 flush_bio_list(conf, bio);
1148 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1150 r1_bio->master_bio = bio;
1151 r1_bio->sectors = bio_sectors(bio);
1153 r1_bio->mddev = mddev;
1154 r1_bio->sector = bio->bi_iter.bi_sector;
1157 static inline struct r1bio *
1158 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1160 struct r1conf *conf = mddev->private;
1161 struct r1bio *r1_bio;
1163 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1164 /* Ensure no bio records IO_BLOCKED */
1165 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1166 init_r1bio(r1_bio, mddev, bio);
1170 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1171 int max_read_sectors, struct r1bio *r1_bio)
1173 struct r1conf *conf = mddev->private;
1174 struct raid1_info *mirror;
1175 struct bio *read_bio;
1176 struct bitmap *bitmap = mddev->bitmap;
1177 const int op = bio_op(bio);
1178 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1181 bool print_msg = !!r1_bio;
1182 char b[BDEVNAME_SIZE];
1185 * If r1_bio is set, we are blocking the raid1d thread
1186 * so there is a tiny risk of deadlock. So ask for
1187 * emergency memory if needed.
1189 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1192 /* Need to get the block device name carefully */
1193 struct md_rdev *rdev;
1195 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1197 bdevname(rdev->bdev, b);
1204 * Still need barrier for READ in case that whole
1207 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1210 r1_bio = alloc_r1bio(mddev, bio);
1212 init_r1bio(r1_bio, mddev, bio);
1213 r1_bio->sectors = max_read_sectors;
1216 * make_request() can abort the operation when read-ahead is being
1217 * used and no empty request is available.
1219 rdisk = read_balance(conf, r1_bio, &max_sectors);
1222 /* couldn't find anywhere to read from */
1224 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1227 (unsigned long long)r1_bio->sector);
1229 raid_end_bio_io(r1_bio);
1232 mirror = conf->mirrors + rdisk;
1235 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1237 (unsigned long long)r1_bio->sector,
1238 bdevname(mirror->rdev->bdev, b));
1240 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1243 * Reading from a write-mostly device must take care not to
1244 * over-take any writes that are 'behind'
1246 raid1_log(mddev, "wait behind writes");
1247 wait_event(bitmap->behind_wait,
1248 atomic_read(&bitmap->behind_writes) == 0);
1251 if (max_sectors < bio_sectors(bio)) {
1252 struct bio *split = bio_split(bio, max_sectors,
1253 gfp, conf->bio_split);
1254 bio_chain(split, bio);
1255 generic_make_request(bio);
1257 r1_bio->master_bio = bio;
1258 r1_bio->sectors = max_sectors;
1261 r1_bio->read_disk = rdisk;
1263 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1265 r1_bio->bios[rdisk] = read_bio;
1267 read_bio->bi_iter.bi_sector = r1_bio->sector +
1268 mirror->rdev->data_offset;
1269 bio_set_dev(read_bio, mirror->rdev->bdev);
1270 read_bio->bi_end_io = raid1_end_read_request;
1271 bio_set_op_attrs(read_bio, op, do_sync);
1272 if (test_bit(FailFast, &mirror->rdev->flags) &&
1273 test_bit(R1BIO_FailFast, &r1_bio->state))
1274 read_bio->bi_opf |= MD_FAILFAST;
1275 read_bio->bi_private = r1_bio;
1278 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1279 disk_devt(mddev->gendisk), r1_bio->sector);
1281 generic_make_request(read_bio);
1284 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1285 int max_write_sectors)
1287 struct r1conf *conf = mddev->private;
1288 struct r1bio *r1_bio;
1290 struct bitmap *bitmap = mddev->bitmap;
1291 unsigned long flags;
1292 struct md_rdev *blocked_rdev;
1293 struct blk_plug_cb *cb;
1294 struct raid1_plug_cb *plug = NULL;
1299 * Register the new request and wait if the reconstruction
1300 * thread has put up a bar for new requests.
1301 * Continue immediately if no resync is active currently.
1305 if (mddev_is_clustered(mddev) &&
1306 md_cluster_ops->area_resyncing(mddev, WRITE,
1307 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1310 * As the suspend_* range is controlled by userspace, we want
1311 * an interruptible wait.
1316 prepare_to_wait(&conf->wait_barrier,
1317 &w, TASK_INTERRUPTIBLE);
1318 if (!mddev_is_clustered(mddev) ||
1319 !md_cluster_ops->area_resyncing(mddev, WRITE,
1320 bio->bi_iter.bi_sector,
1321 bio_end_sector(bio)))
1324 sigprocmask(SIG_BLOCK, &full, &old);
1326 sigprocmask(SIG_SETMASK, &old, NULL);
1328 finish_wait(&conf->wait_barrier, &w);
1330 wait_barrier(conf, bio->bi_iter.bi_sector);
1332 r1_bio = alloc_r1bio(mddev, bio);
1333 r1_bio->sectors = max_write_sectors;
1335 if (conf->pending_count >= max_queued_requests) {
1336 md_wakeup_thread(mddev->thread);
1337 raid1_log(mddev, "wait queued");
1338 wait_event(conf->wait_barrier,
1339 conf->pending_count < max_queued_requests);
1341 /* first select target devices under rcu_lock and
1342 * inc refcount on their rdev. Record them by setting
1344 * If there are known/acknowledged bad blocks on any device on
1345 * which we have seen a write error, we want to avoid writing those
1347 * This potentially requires several writes to write around
1348 * the bad blocks. Each set of writes gets it's own r1bio
1349 * with a set of bios attached.
1352 disks = conf->raid_disks * 2;
1354 blocked_rdev = NULL;
1356 max_sectors = r1_bio->sectors;
1357 for (i = 0; i < disks; i++) {
1358 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1359 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1360 atomic_inc(&rdev->nr_pending);
1361 blocked_rdev = rdev;
1364 r1_bio->bios[i] = NULL;
1365 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1366 if (i < conf->raid_disks)
1367 set_bit(R1BIO_Degraded, &r1_bio->state);
1371 atomic_inc(&rdev->nr_pending);
1372 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1377 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1378 &first_bad, &bad_sectors);
1380 /* mustn't write here until the bad block is
1382 set_bit(BlockedBadBlocks, &rdev->flags);
1383 blocked_rdev = rdev;
1386 if (is_bad && first_bad <= r1_bio->sector) {
1387 /* Cannot write here at all */
1388 bad_sectors -= (r1_bio->sector - first_bad);
1389 if (bad_sectors < max_sectors)
1390 /* mustn't write more than bad_sectors
1391 * to other devices yet
1393 max_sectors = bad_sectors;
1394 rdev_dec_pending(rdev, mddev);
1395 /* We don't set R1BIO_Degraded as that
1396 * only applies if the disk is
1397 * missing, so it might be re-added,
1398 * and we want to know to recover this
1400 * In this case the device is here,
1401 * and the fact that this chunk is not
1402 * in-sync is recorded in the bad
1408 int good_sectors = first_bad - r1_bio->sector;
1409 if (good_sectors < max_sectors)
1410 max_sectors = good_sectors;
1413 r1_bio->bios[i] = bio;
1417 if (unlikely(blocked_rdev)) {
1418 /* Wait for this device to become unblocked */
1421 for (j = 0; j < i; j++)
1422 if (r1_bio->bios[j])
1423 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1425 allow_barrier(conf, bio->bi_iter.bi_sector);
1426 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1427 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1428 wait_barrier(conf, bio->bi_iter.bi_sector);
1432 if (max_sectors < bio_sectors(bio)) {
1433 struct bio *split = bio_split(bio, max_sectors,
1434 GFP_NOIO, conf->bio_split);
1435 bio_chain(split, bio);
1436 generic_make_request(bio);
1438 r1_bio->master_bio = bio;
1439 r1_bio->sectors = max_sectors;
1442 atomic_set(&r1_bio->remaining, 1);
1443 atomic_set(&r1_bio->behind_remaining, 0);
1447 for (i = 0; i < disks; i++) {
1448 struct bio *mbio = NULL;
1449 if (!r1_bio->bios[i])
1455 * Not if there are too many, or cannot
1456 * allocate memory, or a reader on WriteMostly
1457 * is waiting for behind writes to flush */
1459 (atomic_read(&bitmap->behind_writes)
1460 < mddev->bitmap_info.max_write_behind) &&
1461 !waitqueue_active(&bitmap->behind_wait)) {
1462 alloc_behind_master_bio(r1_bio, bio);
1465 bitmap_startwrite(bitmap, r1_bio->sector,
1467 test_bit(R1BIO_BehindIO,
1472 if (r1_bio->behind_master_bio)
1473 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1474 GFP_NOIO, mddev->bio_set);
1476 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1478 if (r1_bio->behind_master_bio) {
1479 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1480 atomic_inc(&r1_bio->behind_remaining);
1483 r1_bio->bios[i] = mbio;
1485 mbio->bi_iter.bi_sector = (r1_bio->sector +
1486 conf->mirrors[i].rdev->data_offset);
1487 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1488 mbio->bi_end_io = raid1_end_write_request;
1489 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1490 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1491 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1492 conf->raid_disks - mddev->degraded > 1)
1493 mbio->bi_opf |= MD_FAILFAST;
1494 mbio->bi_private = r1_bio;
1496 atomic_inc(&r1_bio->remaining);
1499 trace_block_bio_remap(mbio->bi_disk->queue,
1500 mbio, disk_devt(mddev->gendisk),
1502 /* flush_pending_writes() needs access to the rdev so...*/
1503 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1505 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1507 plug = container_of(cb, struct raid1_plug_cb, cb);
1511 bio_list_add(&plug->pending, mbio);
1512 plug->pending_cnt++;
1514 spin_lock_irqsave(&conf->device_lock, flags);
1515 bio_list_add(&conf->pending_bio_list, mbio);
1516 conf->pending_count++;
1517 spin_unlock_irqrestore(&conf->device_lock, flags);
1518 md_wakeup_thread(mddev->thread);
1522 r1_bio_write_done(r1_bio);
1524 /* In case raid1d snuck in to freeze_array */
1525 wake_up(&conf->wait_barrier);
1528 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1532 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1533 md_flush_request(mddev, bio);
1538 * There is a limit to the maximum size, but
1539 * the read/write handler might find a lower limit
1540 * due to bad blocks. To avoid multiple splits,
1541 * we pass the maximum number of sectors down
1542 * and let the lower level perform the split.
1544 sectors = align_to_barrier_unit_end(
1545 bio->bi_iter.bi_sector, bio_sectors(bio));
1547 if (bio_data_dir(bio) == READ)
1548 raid1_read_request(mddev, bio, sectors, NULL);
1550 if (!md_write_start(mddev,bio))
1552 raid1_write_request(mddev, bio, sectors);
1557 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1559 struct r1conf *conf = mddev->private;
1562 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1563 conf->raid_disks - mddev->degraded);
1565 for (i = 0; i < conf->raid_disks; i++) {
1566 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1567 seq_printf(seq, "%s",
1568 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1571 seq_printf(seq, "]");
1574 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1576 char b[BDEVNAME_SIZE];
1577 struct r1conf *conf = mddev->private;
1578 unsigned long flags;
1581 * If it is not operational, then we have already marked it as dead
1582 * else if it is the last working disks, ignore the error, let the
1583 * next level up know.
1584 * else mark the drive as failed
1586 spin_lock_irqsave(&conf->device_lock, flags);
1587 if (test_bit(In_sync, &rdev->flags)
1588 && (conf->raid_disks - mddev->degraded) == 1) {
1590 * Don't fail the drive, act as though we were just a
1591 * normal single drive.
1592 * However don't try a recovery from this drive as
1593 * it is very likely to fail.
1595 conf->recovery_disabled = mddev->recovery_disabled;
1596 spin_unlock_irqrestore(&conf->device_lock, flags);
1599 set_bit(Blocked, &rdev->flags);
1600 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1602 set_bit(Faulty, &rdev->flags);
1604 set_bit(Faulty, &rdev->flags);
1605 spin_unlock_irqrestore(&conf->device_lock, flags);
1607 * if recovery is running, make sure it aborts.
1609 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1610 set_mask_bits(&mddev->sb_flags, 0,
1611 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1612 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1613 "md/raid1:%s: Operation continuing on %d devices.\n",
1614 mdname(mddev), bdevname(rdev->bdev, b),
1615 mdname(mddev), conf->raid_disks - mddev->degraded);
1618 static void print_conf(struct r1conf *conf)
1622 pr_debug("RAID1 conf printout:\n");
1624 pr_debug("(!conf)\n");
1627 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1631 for (i = 0; i < conf->raid_disks; i++) {
1632 char b[BDEVNAME_SIZE];
1633 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1635 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1636 i, !test_bit(In_sync, &rdev->flags),
1637 !test_bit(Faulty, &rdev->flags),
1638 bdevname(rdev->bdev,b));
1643 static void close_sync(struct r1conf *conf)
1647 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1648 _wait_barrier(conf, idx);
1649 _allow_barrier(conf, idx);
1652 mempool_destroy(conf->r1buf_pool);
1653 conf->r1buf_pool = NULL;
1656 static int raid1_spare_active(struct mddev *mddev)
1659 struct r1conf *conf = mddev->private;
1661 unsigned long flags;
1664 * Find all failed disks within the RAID1 configuration
1665 * and mark them readable.
1666 * Called under mddev lock, so rcu protection not needed.
1667 * device_lock used to avoid races with raid1_end_read_request
1668 * which expects 'In_sync' flags and ->degraded to be consistent.
1670 spin_lock_irqsave(&conf->device_lock, flags);
1671 for (i = 0; i < conf->raid_disks; i++) {
1672 struct md_rdev *rdev = conf->mirrors[i].rdev;
1673 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1675 && !test_bit(Candidate, &repl->flags)
1676 && repl->recovery_offset == MaxSector
1677 && !test_bit(Faulty, &repl->flags)
1678 && !test_and_set_bit(In_sync, &repl->flags)) {
1679 /* replacement has just become active */
1681 !test_and_clear_bit(In_sync, &rdev->flags))
1684 /* Replaced device not technically
1685 * faulty, but we need to be sure
1686 * it gets removed and never re-added
1688 set_bit(Faulty, &rdev->flags);
1689 sysfs_notify_dirent_safe(
1694 && rdev->recovery_offset == MaxSector
1695 && !test_bit(Faulty, &rdev->flags)
1696 && !test_and_set_bit(In_sync, &rdev->flags)) {
1698 sysfs_notify_dirent_safe(rdev->sysfs_state);
1701 mddev->degraded -= count;
1702 spin_unlock_irqrestore(&conf->device_lock, flags);
1708 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1710 struct r1conf *conf = mddev->private;
1713 struct raid1_info *p;
1715 int last = conf->raid_disks - 1;
1717 if (mddev->recovery_disabled == conf->recovery_disabled)
1720 if (md_integrity_add_rdev(rdev, mddev))
1723 if (rdev->raid_disk >= 0)
1724 first = last = rdev->raid_disk;
1727 * find the disk ... but prefer rdev->saved_raid_disk
1730 if (rdev->saved_raid_disk >= 0 &&
1731 rdev->saved_raid_disk >= first &&
1732 rdev->saved_raid_disk < conf->raid_disks &&
1733 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1734 first = last = rdev->saved_raid_disk;
1736 for (mirror = first; mirror <= last; mirror++) {
1737 p = conf->mirrors+mirror;
1741 disk_stack_limits(mddev->gendisk, rdev->bdev,
1742 rdev->data_offset << 9);
1744 p->head_position = 0;
1745 rdev->raid_disk = mirror;
1747 /* As all devices are equivalent, we don't need a full recovery
1748 * if this was recently any drive of the array
1750 if (rdev->saved_raid_disk < 0)
1752 rcu_assign_pointer(p->rdev, rdev);
1755 if (test_bit(WantReplacement, &p->rdev->flags) &&
1756 p[conf->raid_disks].rdev == NULL) {
1757 /* Add this device as a replacement */
1758 clear_bit(In_sync, &rdev->flags);
1759 set_bit(Replacement, &rdev->flags);
1760 rdev->raid_disk = mirror;
1763 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1767 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1768 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1773 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1775 struct r1conf *conf = mddev->private;
1777 int number = rdev->raid_disk;
1778 struct raid1_info *p = conf->mirrors + number;
1780 if (unlikely(number >= conf->raid_disks))
1783 if (rdev != p->rdev)
1784 p = conf->mirrors + conf->raid_disks + number;
1787 if (rdev == p->rdev) {
1788 if (test_bit(In_sync, &rdev->flags) ||
1789 atomic_read(&rdev->nr_pending)) {
1793 /* Only remove non-faulty devices if recovery
1796 if (!test_bit(Faulty, &rdev->flags) &&
1797 mddev->recovery_disabled != conf->recovery_disabled &&
1798 mddev->degraded < conf->raid_disks) {
1803 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1805 if (atomic_read(&rdev->nr_pending)) {
1806 /* lost the race, try later */
1812 if (conf->mirrors[conf->raid_disks + number].rdev) {
1813 /* We just removed a device that is being replaced.
1814 * Move down the replacement. We drain all IO before
1815 * doing this to avoid confusion.
1817 struct md_rdev *repl =
1818 conf->mirrors[conf->raid_disks + number].rdev;
1819 freeze_array(conf, 0);
1820 if (atomic_read(&repl->nr_pending)) {
1821 /* It means that some queued IO of retry_list
1822 * hold repl. Thus, we cannot set replacement
1823 * as NULL, avoiding rdev NULL pointer
1824 * dereference in sync_request_write and
1825 * handle_write_finished.
1828 unfreeze_array(conf);
1831 clear_bit(Replacement, &repl->flags);
1833 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1834 unfreeze_array(conf);
1837 clear_bit(WantReplacement, &rdev->flags);
1838 err = md_integrity_register(mddev);
1846 static void end_sync_read(struct bio *bio)
1848 struct r1bio *r1_bio = get_resync_r1bio(bio);
1850 update_head_pos(r1_bio->read_disk, r1_bio);
1853 * we have read a block, now it needs to be re-written,
1854 * or re-read if the read failed.
1855 * We don't do much here, just schedule handling by raid1d
1857 if (!bio->bi_status)
1858 set_bit(R1BIO_Uptodate, &r1_bio->state);
1860 if (atomic_dec_and_test(&r1_bio->remaining))
1861 reschedule_retry(r1_bio);
1864 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1866 sector_t sync_blocks = 0;
1867 sector_t s = r1_bio->sector;
1868 long sectors_to_go = r1_bio->sectors;
1870 /* make sure these bits don't get cleared. */
1872 bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1874 sectors_to_go -= sync_blocks;
1875 } while (sectors_to_go > 0);
1878 static void end_sync_write(struct bio *bio)
1880 int uptodate = !bio->bi_status;
1881 struct r1bio *r1_bio = get_resync_r1bio(bio);
1882 struct mddev *mddev = r1_bio->mddev;
1883 struct r1conf *conf = mddev->private;
1886 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1889 abort_sync_write(mddev, r1_bio);
1890 set_bit(WriteErrorSeen, &rdev->flags);
1891 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1892 set_bit(MD_RECOVERY_NEEDED, &
1894 set_bit(R1BIO_WriteError, &r1_bio->state);
1895 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1896 &first_bad, &bad_sectors) &&
1897 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1900 &first_bad, &bad_sectors)
1902 set_bit(R1BIO_MadeGood, &r1_bio->state);
1904 if (atomic_dec_and_test(&r1_bio->remaining)) {
1905 int s = r1_bio->sectors;
1906 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1907 test_bit(R1BIO_WriteError, &r1_bio->state))
1908 reschedule_retry(r1_bio);
1911 md_done_sync(mddev, s, uptodate);
1916 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1917 int sectors, struct page *page, int rw)
1919 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1923 set_bit(WriteErrorSeen, &rdev->flags);
1924 if (!test_and_set_bit(WantReplacement,
1926 set_bit(MD_RECOVERY_NEEDED, &
1927 rdev->mddev->recovery);
1929 /* need to record an error - either for the block or the device */
1930 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1931 md_error(rdev->mddev, rdev);
1935 static int fix_sync_read_error(struct r1bio *r1_bio)
1937 /* Try some synchronous reads of other devices to get
1938 * good data, much like with normal read errors. Only
1939 * read into the pages we already have so we don't
1940 * need to re-issue the read request.
1941 * We don't need to freeze the array, because being in an
1942 * active sync request, there is no normal IO, and
1943 * no overlapping syncs.
1944 * We don't need to check is_badblock() again as we
1945 * made sure that anything with a bad block in range
1946 * will have bi_end_io clear.
1948 struct mddev *mddev = r1_bio->mddev;
1949 struct r1conf *conf = mddev->private;
1950 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1951 struct page **pages = get_resync_pages(bio)->pages;
1952 sector_t sect = r1_bio->sector;
1953 int sectors = r1_bio->sectors;
1955 struct md_rdev *rdev;
1957 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1958 if (test_bit(FailFast, &rdev->flags)) {
1959 /* Don't try recovering from here - just fail it
1960 * ... unless it is the last working device of course */
1961 md_error(mddev, rdev);
1962 if (test_bit(Faulty, &rdev->flags))
1963 /* Don't try to read from here, but make sure
1964 * put_buf does it's thing
1966 bio->bi_end_io = end_sync_write;
1971 int d = r1_bio->read_disk;
1975 if (s > (PAGE_SIZE>>9))
1978 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1979 /* No rcu protection needed here devices
1980 * can only be removed when no resync is
1981 * active, and resync is currently active
1983 rdev = conf->mirrors[d].rdev;
1984 if (sync_page_io(rdev, sect, s<<9,
1986 REQ_OP_READ, 0, false)) {
1992 if (d == conf->raid_disks * 2)
1994 } while (!success && d != r1_bio->read_disk);
1997 char b[BDEVNAME_SIZE];
1999 /* Cannot read from anywhere, this block is lost.
2000 * Record a bad block on each device. If that doesn't
2001 * work just disable and interrupt the recovery.
2002 * Don't fail devices as that won't really help.
2004 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2005 mdname(mddev), bio_devname(bio, b),
2006 (unsigned long long)r1_bio->sector);
2007 for (d = 0; d < conf->raid_disks * 2; d++) {
2008 rdev = conf->mirrors[d].rdev;
2009 if (!rdev || test_bit(Faulty, &rdev->flags))
2011 if (!rdev_set_badblocks(rdev, sect, s, 0))
2015 conf->recovery_disabled =
2016 mddev->recovery_disabled;
2017 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2018 md_done_sync(mddev, r1_bio->sectors, 0);
2030 /* write it back and re-read */
2031 while (d != r1_bio->read_disk) {
2033 d = conf->raid_disks * 2;
2035 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2037 rdev = conf->mirrors[d].rdev;
2038 if (r1_sync_page_io(rdev, sect, s,
2041 r1_bio->bios[d]->bi_end_io = NULL;
2042 rdev_dec_pending(rdev, mddev);
2046 while (d != r1_bio->read_disk) {
2048 d = conf->raid_disks * 2;
2050 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2052 rdev = conf->mirrors[d].rdev;
2053 if (r1_sync_page_io(rdev, sect, s,
2056 atomic_add(s, &rdev->corrected_errors);
2062 set_bit(R1BIO_Uptodate, &r1_bio->state);
2067 static void process_checks(struct r1bio *r1_bio)
2069 /* We have read all readable devices. If we haven't
2070 * got the block, then there is no hope left.
2071 * If we have, then we want to do a comparison
2072 * and skip the write if everything is the same.
2073 * If any blocks failed to read, then we need to
2074 * attempt an over-write
2076 struct mddev *mddev = r1_bio->mddev;
2077 struct r1conf *conf = mddev->private;
2082 /* Fix variable parts of all bios */
2083 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2084 for (i = 0; i < conf->raid_disks * 2; i++) {
2085 blk_status_t status;
2086 struct bio *b = r1_bio->bios[i];
2087 struct resync_pages *rp = get_resync_pages(b);
2088 if (b->bi_end_io != end_sync_read)
2090 /* fixup the bio for reuse, but preserve errno */
2091 status = b->bi_status;
2093 b->bi_status = status;
2094 b->bi_iter.bi_sector = r1_bio->sector +
2095 conf->mirrors[i].rdev->data_offset;
2096 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2097 b->bi_end_io = end_sync_read;
2098 rp->raid_bio = r1_bio;
2101 /* initialize bvec table again */
2102 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2104 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2105 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2106 !r1_bio->bios[primary]->bi_status) {
2107 r1_bio->bios[primary]->bi_end_io = NULL;
2108 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2111 r1_bio->read_disk = primary;
2112 for (i = 0; i < conf->raid_disks * 2; i++) {
2114 struct bio *pbio = r1_bio->bios[primary];
2115 struct bio *sbio = r1_bio->bios[i];
2116 blk_status_t status = sbio->bi_status;
2117 struct page **ppages = get_resync_pages(pbio)->pages;
2118 struct page **spages = get_resync_pages(sbio)->pages;
2120 int page_len[RESYNC_PAGES] = { 0 };
2122 if (sbio->bi_end_io != end_sync_read)
2124 /* Now we can 'fixup' the error value */
2125 sbio->bi_status = 0;
2127 bio_for_each_segment_all(bi, sbio, j)
2128 page_len[j] = bi->bv_len;
2131 for (j = vcnt; j-- ; ) {
2132 if (memcmp(page_address(ppages[j]),
2133 page_address(spages[j]),
2140 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2141 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2143 /* No need to write to this device. */
2144 sbio->bi_end_io = NULL;
2145 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2149 bio_copy_data(sbio, pbio);
2153 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2155 struct r1conf *conf = mddev->private;
2157 int disks = conf->raid_disks * 2;
2160 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2161 /* ouch - failed to read all of that. */
2162 if (!fix_sync_read_error(r1_bio))
2165 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2166 process_checks(r1_bio);
2171 atomic_set(&r1_bio->remaining, 1);
2172 for (i = 0; i < disks ; i++) {
2173 wbio = r1_bio->bios[i];
2174 if (wbio->bi_end_io == NULL ||
2175 (wbio->bi_end_io == end_sync_read &&
2176 (i == r1_bio->read_disk ||
2177 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2179 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2180 abort_sync_write(mddev, r1_bio);
2184 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2185 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2186 wbio->bi_opf |= MD_FAILFAST;
2188 wbio->bi_end_io = end_sync_write;
2189 atomic_inc(&r1_bio->remaining);
2190 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2192 generic_make_request(wbio);
2195 if (atomic_dec_and_test(&r1_bio->remaining)) {
2196 /* if we're here, all write(s) have completed, so clean up */
2197 int s = r1_bio->sectors;
2198 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2199 test_bit(R1BIO_WriteError, &r1_bio->state))
2200 reschedule_retry(r1_bio);
2203 md_done_sync(mddev, s, 1);
2209 * This is a kernel thread which:
2211 * 1. Retries failed read operations on working mirrors.
2212 * 2. Updates the raid superblock when problems encounter.
2213 * 3. Performs writes following reads for array synchronising.
2216 static void fix_read_error(struct r1conf *conf, int read_disk,
2217 sector_t sect, int sectors)
2219 struct mddev *mddev = conf->mddev;
2225 struct md_rdev *rdev;
2227 if (s > (PAGE_SIZE>>9))
2235 rdev = rcu_dereference(conf->mirrors[d].rdev);
2237 (test_bit(In_sync, &rdev->flags) ||
2238 (!test_bit(Faulty, &rdev->flags) &&
2239 rdev->recovery_offset >= sect + s)) &&
2240 is_badblock(rdev, sect, s,
2241 &first_bad, &bad_sectors) == 0) {
2242 atomic_inc(&rdev->nr_pending);
2244 if (sync_page_io(rdev, sect, s<<9,
2245 conf->tmppage, REQ_OP_READ, 0, false))
2247 rdev_dec_pending(rdev, mddev);
2253 if (d == conf->raid_disks * 2)
2255 } while (!success && d != read_disk);
2258 /* Cannot read from anywhere - mark it bad */
2259 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2260 if (!rdev_set_badblocks(rdev, sect, s, 0))
2261 md_error(mddev, rdev);
2264 /* write it back and re-read */
2266 while (d != read_disk) {
2268 d = conf->raid_disks * 2;
2271 rdev = rcu_dereference(conf->mirrors[d].rdev);
2273 !test_bit(Faulty, &rdev->flags)) {
2274 atomic_inc(&rdev->nr_pending);
2276 r1_sync_page_io(rdev, sect, s,
2277 conf->tmppage, WRITE);
2278 rdev_dec_pending(rdev, mddev);
2283 while (d != read_disk) {
2284 char b[BDEVNAME_SIZE];
2286 d = conf->raid_disks * 2;
2289 rdev = rcu_dereference(conf->mirrors[d].rdev);
2291 !test_bit(Faulty, &rdev->flags)) {
2292 atomic_inc(&rdev->nr_pending);
2294 if (r1_sync_page_io(rdev, sect, s,
2295 conf->tmppage, READ)) {
2296 atomic_add(s, &rdev->corrected_errors);
2297 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2299 (unsigned long long)(sect +
2301 bdevname(rdev->bdev, b));
2303 rdev_dec_pending(rdev, mddev);
2312 static int narrow_write_error(struct r1bio *r1_bio, int i)
2314 struct mddev *mddev = r1_bio->mddev;
2315 struct r1conf *conf = mddev->private;
2316 struct md_rdev *rdev = conf->mirrors[i].rdev;
2318 /* bio has the data to be written to device 'i' where
2319 * we just recently had a write error.
2320 * We repeatedly clone the bio and trim down to one block,
2321 * then try the write. Where the write fails we record
2323 * It is conceivable that the bio doesn't exactly align with
2324 * blocks. We must handle this somehow.
2326 * We currently own a reference on the rdev.
2332 int sect_to_write = r1_bio->sectors;
2335 if (rdev->badblocks.shift < 0)
2338 block_sectors = roundup(1 << rdev->badblocks.shift,
2339 bdev_logical_block_size(rdev->bdev) >> 9);
2340 sector = r1_bio->sector;
2341 sectors = ((sector + block_sectors)
2342 & ~(sector_t)(block_sectors - 1))
2345 while (sect_to_write) {
2347 if (sectors > sect_to_write)
2348 sectors = sect_to_write;
2349 /* Write at 'sector' for 'sectors'*/
2351 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2352 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2356 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2360 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2361 wbio->bi_iter.bi_sector = r1_bio->sector;
2362 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2364 bio_trim(wbio, sector - r1_bio->sector, sectors);
2365 wbio->bi_iter.bi_sector += rdev->data_offset;
2366 bio_set_dev(wbio, rdev->bdev);
2368 if (submit_bio_wait(wbio) < 0)
2370 ok = rdev_set_badblocks(rdev, sector,
2375 sect_to_write -= sectors;
2377 sectors = block_sectors;
2382 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2385 int s = r1_bio->sectors;
2386 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2387 struct md_rdev *rdev = conf->mirrors[m].rdev;
2388 struct bio *bio = r1_bio->bios[m];
2389 if (bio->bi_end_io == NULL)
2391 if (!bio->bi_status &&
2392 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2393 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2395 if (bio->bi_status &&
2396 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2397 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2398 md_error(conf->mddev, rdev);
2402 md_done_sync(conf->mddev, s, 1);
2405 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2410 for (m = 0; m < conf->raid_disks * 2 ; m++)
2411 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2412 struct md_rdev *rdev = conf->mirrors[m].rdev;
2413 rdev_clear_badblocks(rdev,
2415 r1_bio->sectors, 0);
2416 rdev_dec_pending(rdev, conf->mddev);
2417 } else if (r1_bio->bios[m] != NULL) {
2418 /* This drive got a write error. We need to
2419 * narrow down and record precise write
2423 if (!narrow_write_error(r1_bio, m)) {
2424 md_error(conf->mddev,
2425 conf->mirrors[m].rdev);
2426 /* an I/O failed, we can't clear the bitmap */
2427 set_bit(R1BIO_Degraded, &r1_bio->state);
2429 rdev_dec_pending(conf->mirrors[m].rdev,
2433 spin_lock_irq(&conf->device_lock);
2434 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2435 idx = sector_to_idx(r1_bio->sector);
2436 atomic_inc(&conf->nr_queued[idx]);
2437 spin_unlock_irq(&conf->device_lock);
2439 * In case freeze_array() is waiting for condition
2440 * get_unqueued_pending() == extra to be true.
2442 wake_up(&conf->wait_barrier);
2443 md_wakeup_thread(conf->mddev->thread);
2445 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2446 close_write(r1_bio);
2447 raid_end_bio_io(r1_bio);
2451 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2453 struct mddev *mddev = conf->mddev;
2455 struct md_rdev *rdev;
2456 sector_t bio_sector;
2458 clear_bit(R1BIO_ReadError, &r1_bio->state);
2459 /* we got a read error. Maybe the drive is bad. Maybe just
2460 * the block and we can fix it.
2461 * We freeze all other IO, and try reading the block from
2462 * other devices. When we find one, we re-write
2463 * and check it that fixes the read error.
2464 * This is all done synchronously while the array is
2468 bio = r1_bio->bios[r1_bio->read_disk];
2469 bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2471 r1_bio->bios[r1_bio->read_disk] = NULL;
2473 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2475 && !test_bit(FailFast, &rdev->flags)) {
2476 freeze_array(conf, 1);
2477 fix_read_error(conf, r1_bio->read_disk,
2478 r1_bio->sector, r1_bio->sectors);
2479 unfreeze_array(conf);
2480 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2481 md_error(mddev, rdev);
2483 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2486 rdev_dec_pending(rdev, conf->mddev);
2487 allow_barrier(conf, r1_bio->sector);
2488 bio = r1_bio->master_bio;
2490 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2492 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2495 static void raid1d(struct md_thread *thread)
2497 struct mddev *mddev = thread->mddev;
2498 struct r1bio *r1_bio;
2499 unsigned long flags;
2500 struct r1conf *conf = mddev->private;
2501 struct list_head *head = &conf->retry_list;
2502 struct blk_plug plug;
2505 md_check_recovery(mddev);
2507 if (!list_empty_careful(&conf->bio_end_io_list) &&
2508 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2510 spin_lock_irqsave(&conf->device_lock, flags);
2511 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2512 list_splice_init(&conf->bio_end_io_list, &tmp);
2513 spin_unlock_irqrestore(&conf->device_lock, flags);
2514 while (!list_empty(&tmp)) {
2515 r1_bio = list_first_entry(&tmp, struct r1bio,
2517 list_del(&r1_bio->retry_list);
2518 idx = sector_to_idx(r1_bio->sector);
2519 atomic_dec(&conf->nr_queued[idx]);
2520 if (mddev->degraded)
2521 set_bit(R1BIO_Degraded, &r1_bio->state);
2522 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2523 close_write(r1_bio);
2524 raid_end_bio_io(r1_bio);
2528 blk_start_plug(&plug);
2531 flush_pending_writes(conf);
2533 spin_lock_irqsave(&conf->device_lock, flags);
2534 if (list_empty(head)) {
2535 spin_unlock_irqrestore(&conf->device_lock, flags);
2538 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2539 list_del(head->prev);
2540 idx = sector_to_idx(r1_bio->sector);
2541 atomic_dec(&conf->nr_queued[idx]);
2542 spin_unlock_irqrestore(&conf->device_lock, flags);
2544 mddev = r1_bio->mddev;
2545 conf = mddev->private;
2546 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2547 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2548 test_bit(R1BIO_WriteError, &r1_bio->state))
2549 handle_sync_write_finished(conf, r1_bio);
2551 sync_request_write(mddev, r1_bio);
2552 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2553 test_bit(R1BIO_WriteError, &r1_bio->state))
2554 handle_write_finished(conf, r1_bio);
2555 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2556 handle_read_error(conf, r1_bio);
2561 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2562 md_check_recovery(mddev);
2564 blk_finish_plug(&plug);
2567 static int init_resync(struct r1conf *conf)
2571 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2572 BUG_ON(conf->r1buf_pool);
2573 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2575 if (!conf->r1buf_pool)
2580 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2582 struct r1bio *r1bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2583 struct resync_pages *rps;
2587 for (i = conf->poolinfo->raid_disks; i--; ) {
2588 bio = r1bio->bios[i];
2589 rps = bio->bi_private;
2591 bio->bi_private = rps;
2593 r1bio->master_bio = NULL;
2598 * perform a "sync" on one "block"
2600 * We need to make sure that no normal I/O request - particularly write
2601 * requests - conflict with active sync requests.
2603 * This is achieved by tracking pending requests and a 'barrier' concept
2604 * that can be installed to exclude normal IO requests.
2607 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2610 struct r1conf *conf = mddev->private;
2611 struct r1bio *r1_bio;
2613 sector_t max_sector, nr_sectors;
2617 int write_targets = 0, read_targets = 0;
2618 sector_t sync_blocks;
2619 int still_degraded = 0;
2620 int good_sectors = RESYNC_SECTORS;
2621 int min_bad = 0; /* number of sectors that are bad in all devices */
2622 int idx = sector_to_idx(sector_nr);
2625 if (!conf->r1buf_pool)
2626 if (init_resync(conf))
2629 max_sector = mddev->dev_sectors;
2630 if (sector_nr >= max_sector) {
2631 /* If we aborted, we need to abort the
2632 * sync on the 'current' bitmap chunk (there will
2633 * only be one in raid1 resync.
2634 * We can find the current addess in mddev->curr_resync
2636 if (mddev->curr_resync < max_sector) /* aborted */
2637 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2639 else /* completed sync */
2642 bitmap_close_sync(mddev->bitmap);
2645 if (mddev_is_clustered(mddev)) {
2646 conf->cluster_sync_low = 0;
2647 conf->cluster_sync_high = 0;
2652 if (mddev->bitmap == NULL &&
2653 mddev->recovery_cp == MaxSector &&
2654 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2655 conf->fullsync == 0) {
2657 return max_sector - sector_nr;
2659 /* before building a request, check if we can skip these blocks..
2660 * This call the bitmap_start_sync doesn't actually record anything
2662 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2663 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2664 /* We can skip this block, and probably several more */
2670 * If there is non-resync activity waiting for a turn, then let it
2671 * though before starting on this new sync request.
2673 if (atomic_read(&conf->nr_waiting[idx]))
2674 schedule_timeout_uninterruptible(1);
2676 /* we are incrementing sector_nr below. To be safe, we check against
2677 * sector_nr + two times RESYNC_SECTORS
2680 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2681 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2682 r1_bio = raid1_alloc_init_r1buf(conf);
2684 raise_barrier(conf, sector_nr);
2688 * If we get a correctably read error during resync or recovery,
2689 * we might want to read from a different device. So we
2690 * flag all drives that could conceivably be read from for READ,
2691 * and any others (which will be non-In_sync devices) for WRITE.
2692 * If a read fails, we try reading from something else for which READ
2696 r1_bio->mddev = mddev;
2697 r1_bio->sector = sector_nr;
2699 set_bit(R1BIO_IsSync, &r1_bio->state);
2700 /* make sure good_sectors won't go across barrier unit boundary */
2701 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2703 for (i = 0; i < conf->raid_disks * 2; i++) {
2704 struct md_rdev *rdev;
2705 bio = r1_bio->bios[i];
2707 rdev = rcu_dereference(conf->mirrors[i].rdev);
2709 test_bit(Faulty, &rdev->flags)) {
2710 if (i < conf->raid_disks)
2712 } else if (!test_bit(In_sync, &rdev->flags)) {
2713 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2714 bio->bi_end_io = end_sync_write;
2717 /* may need to read from here */
2718 sector_t first_bad = MaxSector;
2721 if (is_badblock(rdev, sector_nr, good_sectors,
2722 &first_bad, &bad_sectors)) {
2723 if (first_bad > sector_nr)
2724 good_sectors = first_bad - sector_nr;
2726 bad_sectors -= (sector_nr - first_bad);
2728 min_bad > bad_sectors)
2729 min_bad = bad_sectors;
2732 if (sector_nr < first_bad) {
2733 if (test_bit(WriteMostly, &rdev->flags)) {
2740 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2741 bio->bi_end_io = end_sync_read;
2743 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2744 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2745 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2747 * The device is suitable for reading (InSync),
2748 * but has bad block(s) here. Let's try to correct them,
2749 * if we are doing resync or repair. Otherwise, leave
2750 * this device alone for this sync request.
2752 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2753 bio->bi_end_io = end_sync_write;
2757 if (rdev && bio->bi_end_io) {
2758 atomic_inc(&rdev->nr_pending);
2759 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2760 bio_set_dev(bio, rdev->bdev);
2761 if (test_bit(FailFast, &rdev->flags))
2762 bio->bi_opf |= MD_FAILFAST;
2768 r1_bio->read_disk = disk;
2770 if (read_targets == 0 && min_bad > 0) {
2771 /* These sectors are bad on all InSync devices, so we
2772 * need to mark them bad on all write targets
2775 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2776 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2777 struct md_rdev *rdev = conf->mirrors[i].rdev;
2778 ok = rdev_set_badblocks(rdev, sector_nr,
2782 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2787 /* Cannot record the badblocks, so need to
2789 * If there are multiple read targets, could just
2790 * fail the really bad ones ???
2792 conf->recovery_disabled = mddev->recovery_disabled;
2793 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2799 if (min_bad > 0 && min_bad < good_sectors) {
2800 /* only resync enough to reach the next bad->good
2802 good_sectors = min_bad;
2805 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2806 /* extra read targets are also write targets */
2807 write_targets += read_targets-1;
2809 if (write_targets == 0 || read_targets == 0) {
2810 /* There is nowhere to write, so all non-sync
2811 * drives must be failed - so we are finished
2815 max_sector = sector_nr + min_bad;
2816 rv = max_sector - sector_nr;
2822 if (max_sector > mddev->resync_max)
2823 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2824 if (max_sector > sector_nr + good_sectors)
2825 max_sector = sector_nr + good_sectors;
2830 int len = PAGE_SIZE;
2831 if (sector_nr + (len>>9) > max_sector)
2832 len = (max_sector - sector_nr) << 9;
2835 if (sync_blocks == 0) {
2836 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2837 &sync_blocks, still_degraded) &&
2839 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2841 if ((len >> 9) > sync_blocks)
2842 len = sync_blocks<<9;
2845 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2846 struct resync_pages *rp;
2848 bio = r1_bio->bios[i];
2849 rp = get_resync_pages(bio);
2850 if (bio->bi_end_io) {
2851 page = resync_fetch_page(rp, page_idx);
2854 * won't fail because the vec table is big
2855 * enough to hold all these pages
2857 bio_add_page(bio, page, len, 0);
2860 nr_sectors += len>>9;
2861 sector_nr += len>>9;
2862 sync_blocks -= (len>>9);
2863 } while (++page_idx < RESYNC_PAGES);
2865 r1_bio->sectors = nr_sectors;
2867 if (mddev_is_clustered(mddev) &&
2868 conf->cluster_sync_high < sector_nr + nr_sectors) {
2869 conf->cluster_sync_low = mddev->curr_resync_completed;
2870 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2871 /* Send resync message */
2872 md_cluster_ops->resync_info_update(mddev,
2873 conf->cluster_sync_low,
2874 conf->cluster_sync_high);
2877 /* For a user-requested sync, we read all readable devices and do a
2880 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2881 atomic_set(&r1_bio->remaining, read_targets);
2882 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2883 bio = r1_bio->bios[i];
2884 if (bio->bi_end_io == end_sync_read) {
2886 md_sync_acct_bio(bio, nr_sectors);
2887 if (read_targets == 1)
2888 bio->bi_opf &= ~MD_FAILFAST;
2889 generic_make_request(bio);
2893 atomic_set(&r1_bio->remaining, 1);
2894 bio = r1_bio->bios[r1_bio->read_disk];
2895 md_sync_acct_bio(bio, nr_sectors);
2896 if (read_targets == 1)
2897 bio->bi_opf &= ~MD_FAILFAST;
2898 generic_make_request(bio);
2904 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2909 return mddev->dev_sectors;
2912 static struct r1conf *setup_conf(struct mddev *mddev)
2914 struct r1conf *conf;
2916 struct raid1_info *disk;
2917 struct md_rdev *rdev;
2920 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2924 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2925 sizeof(atomic_t), GFP_KERNEL);
2926 if (!conf->nr_pending)
2929 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2930 sizeof(atomic_t), GFP_KERNEL);
2931 if (!conf->nr_waiting)
2934 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2935 sizeof(atomic_t), GFP_KERNEL);
2936 if (!conf->nr_queued)
2939 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2940 sizeof(atomic_t), GFP_KERNEL);
2944 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2945 * mddev->raid_disks * 2,
2950 conf->tmppage = alloc_page(GFP_KERNEL);
2954 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2955 if (!conf->poolinfo)
2957 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2958 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2961 if (!conf->r1bio_pool)
2964 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
2965 if (!conf->bio_split)
2968 conf->poolinfo->mddev = mddev;
2971 spin_lock_init(&conf->device_lock);
2972 rdev_for_each(rdev, mddev) {
2973 int disk_idx = rdev->raid_disk;
2974 if (disk_idx >= mddev->raid_disks
2977 if (test_bit(Replacement, &rdev->flags))
2978 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2980 disk = conf->mirrors + disk_idx;
2985 disk->head_position = 0;
2986 disk->seq_start = MaxSector;
2988 conf->raid_disks = mddev->raid_disks;
2989 conf->mddev = mddev;
2990 INIT_LIST_HEAD(&conf->retry_list);
2991 INIT_LIST_HEAD(&conf->bio_end_io_list);
2993 spin_lock_init(&conf->resync_lock);
2994 init_waitqueue_head(&conf->wait_barrier);
2996 bio_list_init(&conf->pending_bio_list);
2997 conf->pending_count = 0;
2998 conf->recovery_disabled = mddev->recovery_disabled - 1;
3001 for (i = 0; i < conf->raid_disks * 2; i++) {
3003 disk = conf->mirrors + i;
3005 if (i < conf->raid_disks &&
3006 disk[conf->raid_disks].rdev) {
3007 /* This slot has a replacement. */
3009 /* No original, just make the replacement
3010 * a recovering spare
3013 disk[conf->raid_disks].rdev;
3014 disk[conf->raid_disks].rdev = NULL;
3015 } else if (!test_bit(In_sync, &disk->rdev->flags))
3016 /* Original is not in_sync - bad */
3021 !test_bit(In_sync, &disk->rdev->flags)) {
3022 disk->head_position = 0;
3024 (disk->rdev->saved_raid_disk < 0))
3030 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3038 mempool_destroy(conf->r1bio_pool);
3039 kfree(conf->mirrors);
3040 safe_put_page(conf->tmppage);
3041 kfree(conf->poolinfo);
3042 kfree(conf->nr_pending);
3043 kfree(conf->nr_waiting);
3044 kfree(conf->nr_queued);
3045 kfree(conf->barrier);
3046 if (conf->bio_split)
3047 bioset_free(conf->bio_split);
3050 return ERR_PTR(err);
3053 static void raid1_free(struct mddev *mddev, void *priv);
3054 static int raid1_run(struct mddev *mddev)
3056 struct r1conf *conf;
3058 struct md_rdev *rdev;
3060 bool discard_supported = false;
3062 if (mddev->level != 1) {
3063 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3064 mdname(mddev), mddev->level);
3067 if (mddev->reshape_position != MaxSector) {
3068 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3072 if (mddev_init_writes_pending(mddev) < 0)
3075 * copy the already verified devices into our private RAID1
3076 * bookkeeping area. [whatever we allocate in run(),
3077 * should be freed in raid1_free()]
3079 if (mddev->private == NULL)
3080 conf = setup_conf(mddev);
3082 conf = mddev->private;
3085 return PTR_ERR(conf);
3088 blk_queue_max_write_same_sectors(mddev->queue, 0);
3089 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3092 rdev_for_each(rdev, mddev) {
3093 if (!mddev->gendisk)
3095 disk_stack_limits(mddev->gendisk, rdev->bdev,
3096 rdev->data_offset << 9);
3097 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3098 discard_supported = true;
3101 mddev->degraded = 0;
3102 for (i=0; i < conf->raid_disks; i++)
3103 if (conf->mirrors[i].rdev == NULL ||
3104 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3105 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3108 * RAID1 needs at least one disk in active
3110 if (conf->raid_disks - mddev->degraded < 1) {
3111 md_unregister_thread(&conf->thread);
3116 if (conf->raid_disks - mddev->degraded == 1)
3117 mddev->recovery_cp = MaxSector;
3119 if (mddev->recovery_cp != MaxSector)
3120 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3122 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3123 mdname(mddev), mddev->raid_disks - mddev->degraded,
3127 * Ok, everything is just fine now
3129 mddev->thread = conf->thread;
3130 conf->thread = NULL;
3131 mddev->private = conf;
3132 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3134 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3137 if (discard_supported)
3138 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3141 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3145 ret = md_integrity_register(mddev);
3147 md_unregister_thread(&mddev->thread);
3153 raid1_free(mddev, conf);
3157 static void raid1_free(struct mddev *mddev, void *priv)
3159 struct r1conf *conf = priv;
3161 mempool_destroy(conf->r1bio_pool);
3162 kfree(conf->mirrors);
3163 safe_put_page(conf->tmppage);
3164 kfree(conf->poolinfo);
3165 kfree(conf->nr_pending);
3166 kfree(conf->nr_waiting);
3167 kfree(conf->nr_queued);
3168 kfree(conf->barrier);
3169 if (conf->bio_split)
3170 bioset_free(conf->bio_split);
3174 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3176 /* no resync is happening, and there is enough space
3177 * on all devices, so we can resize.
3178 * We need to make sure resync covers any new space.
3179 * If the array is shrinking we should possibly wait until
3180 * any io in the removed space completes, but it hardly seems
3183 sector_t newsize = raid1_size(mddev, sectors, 0);
3184 if (mddev->external_size &&
3185 mddev->array_sectors > newsize)
3187 if (mddev->bitmap) {
3188 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3192 md_set_array_sectors(mddev, newsize);
3193 if (sectors > mddev->dev_sectors &&
3194 mddev->recovery_cp > mddev->dev_sectors) {
3195 mddev->recovery_cp = mddev->dev_sectors;
3196 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3198 mddev->dev_sectors = sectors;
3199 mddev->resync_max_sectors = sectors;
3203 static int raid1_reshape(struct mddev *mddev)
3206 * 1/ resize the r1bio_pool
3207 * 2/ resize conf->mirrors
3209 * We allocate a new r1bio_pool if we can.
3210 * Then raise a device barrier and wait until all IO stops.
3211 * Then resize conf->mirrors and swap in the new r1bio pool.
3213 * At the same time, we "pack" the devices so that all the missing
3214 * devices have the higher raid_disk numbers.
3216 mempool_t *newpool, *oldpool;
3217 struct pool_info *newpoolinfo;
3218 struct raid1_info *newmirrors;
3219 struct r1conf *conf = mddev->private;
3220 int cnt, raid_disks;
3221 unsigned long flags;
3224 /* Cannot change chunk_size, layout, or level */
3225 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3226 mddev->layout != mddev->new_layout ||
3227 mddev->level != mddev->new_level) {
3228 mddev->new_chunk_sectors = mddev->chunk_sectors;
3229 mddev->new_layout = mddev->layout;
3230 mddev->new_level = mddev->level;
3234 if (!mddev_is_clustered(mddev))
3235 md_allow_write(mddev);
3237 raid_disks = mddev->raid_disks + mddev->delta_disks;
3239 if (raid_disks < conf->raid_disks) {
3241 for (d= 0; d < conf->raid_disks; d++)
3242 if (conf->mirrors[d].rdev)
3244 if (cnt > raid_disks)
3248 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3251 newpoolinfo->mddev = mddev;
3252 newpoolinfo->raid_disks = raid_disks * 2;
3254 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3255 r1bio_pool_free, newpoolinfo);
3260 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3264 mempool_destroy(newpool);
3268 freeze_array(conf, 0);
3270 /* ok, everything is stopped */
3271 oldpool = conf->r1bio_pool;
3272 conf->r1bio_pool = newpool;
3274 for (d = d2 = 0; d < conf->raid_disks; d++) {
3275 struct md_rdev *rdev = conf->mirrors[d].rdev;
3276 if (rdev && rdev->raid_disk != d2) {
3277 sysfs_unlink_rdev(mddev, rdev);
3278 rdev->raid_disk = d2;
3279 sysfs_unlink_rdev(mddev, rdev);
3280 if (sysfs_link_rdev(mddev, rdev))
3281 pr_warn("md/raid1:%s: cannot register rd%d\n",
3282 mdname(mddev), rdev->raid_disk);
3285 newmirrors[d2++].rdev = rdev;
3287 kfree(conf->mirrors);
3288 conf->mirrors = newmirrors;
3289 kfree(conf->poolinfo);
3290 conf->poolinfo = newpoolinfo;
3292 spin_lock_irqsave(&conf->device_lock, flags);
3293 mddev->degraded += (raid_disks - conf->raid_disks);
3294 spin_unlock_irqrestore(&conf->device_lock, flags);
3295 conf->raid_disks = mddev->raid_disks = raid_disks;
3296 mddev->delta_disks = 0;
3298 unfreeze_array(conf);
3300 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3301 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3302 md_wakeup_thread(mddev->thread);
3304 mempool_destroy(oldpool);
3308 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3310 struct r1conf *conf = mddev->private;
3313 freeze_array(conf, 0);
3315 unfreeze_array(conf);
3318 static void *raid1_takeover(struct mddev *mddev)
3320 /* raid1 can take over:
3321 * raid5 with 2 devices, any layout or chunk size
3323 if (mddev->level == 5 && mddev->raid_disks == 2) {
3324 struct r1conf *conf;
3325 mddev->new_level = 1;
3326 mddev->new_layout = 0;
3327 mddev->new_chunk_sectors = 0;
3328 conf = setup_conf(mddev);
3329 if (!IS_ERR(conf)) {
3330 /* Array must appear to be quiesced */
3331 conf->array_frozen = 1;
3332 mddev_clear_unsupported_flags(mddev,
3333 UNSUPPORTED_MDDEV_FLAGS);
3337 return ERR_PTR(-EINVAL);
3340 static struct md_personality raid1_personality =
3344 .owner = THIS_MODULE,
3345 .make_request = raid1_make_request,
3348 .status = raid1_status,
3349 .error_handler = raid1_error,
3350 .hot_add_disk = raid1_add_disk,
3351 .hot_remove_disk= raid1_remove_disk,
3352 .spare_active = raid1_spare_active,
3353 .sync_request = raid1_sync_request,
3354 .resize = raid1_resize,
3356 .check_reshape = raid1_reshape,
3357 .quiesce = raid1_quiesce,
3358 .takeover = raid1_takeover,
3359 .congested = raid1_congested,
3362 static int __init raid_init(void)
3364 return register_md_personality(&raid1_personality);
3367 static void raid_exit(void)
3369 unregister_md_personality(&raid1_personality);
3372 module_init(raid_init);
3373 module_exit(raid_exit);
3374 MODULE_LICENSE("GPL");
3375 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3376 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3377 MODULE_ALIAS("md-raid1");
3378 MODULE_ALIAS("md-level-1");
3380 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);