1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * raid1.c : Multiple Devices driver for Linux
5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
9 * RAID-1 management functions.
11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
17 * bitmapped intelligence in resync:
19 * - bitmap marked during normal i/o
20 * - bitmap used to skip nondirty blocks during sync
22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
23 * - persistent bitmap code
26 #include <linux/slab.h>
27 #include <linux/delay.h>
28 #include <linux/blkdev.h>
29 #include <linux/module.h>
30 #include <linux/seq_file.h>
31 #include <linux/ratelimit.h>
32 #include <linux/interval_tree_generic.h>
34 #include <trace/events/block.h>
38 #include "md-bitmap.h"
40 #define UNSUPPORTED_MDDEV_FLAGS \
41 ((1L << MD_HAS_JOURNAL) | \
42 (1L << MD_JOURNAL_CLEAN) | \
43 (1L << MD_HAS_PPL) | \
44 (1L << MD_HAS_MULTIPLE_PPLS))
46 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
47 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
49 #define raid1_log(md, fmt, args...) \
50 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
54 #define START(node) ((node)->start)
55 #define LAST(node) ((node)->last)
56 INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
57 START, LAST, static inline, raid1_rb);
59 static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
60 struct serial_info *si, int idx)
64 sector_t lo = r1_bio->sector;
65 sector_t hi = lo + r1_bio->sectors;
66 struct serial_in_rdev *serial = &rdev->serial[idx];
68 spin_lock_irqsave(&serial->serial_lock, flags);
69 /* collision happened */
70 if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
75 raid1_rb_insert(si, &serial->serial_rb);
77 spin_unlock_irqrestore(&serial->serial_lock, flags);
82 static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
84 struct mddev *mddev = rdev->mddev;
85 struct serial_info *si;
86 int idx = sector_to_idx(r1_bio->sector);
87 struct serial_in_rdev *serial = &rdev->serial[idx];
89 if (WARN_ON(!mddev->serial_info_pool))
91 si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
92 wait_event(serial->serial_io_wait,
93 check_and_add_serial(rdev, r1_bio, si, idx) == 0);
96 static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
98 struct serial_info *si;
101 struct mddev *mddev = rdev->mddev;
102 int idx = sector_to_idx(lo);
103 struct serial_in_rdev *serial = &rdev->serial[idx];
105 spin_lock_irqsave(&serial->serial_lock, flags);
106 for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
107 si; si = raid1_rb_iter_next(si, lo, hi)) {
108 if (si->start == lo && si->last == hi) {
109 raid1_rb_remove(si, &serial->serial_rb);
110 mempool_free(si, mddev->serial_info_pool);
116 WARN(1, "The write IO is not recorded for serialization\n");
117 spin_unlock_irqrestore(&serial->serial_lock, flags);
118 wake_up(&serial->serial_io_wait);
122 * for resync bio, r1bio pointer can be retrieved from the per-bio
123 * 'struct resync_pages'.
125 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
127 return get_resync_pages(bio)->raid_bio;
130 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
132 struct pool_info *pi = data;
133 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
135 /* allocate a r1bio with room for raid_disks entries in the bios array */
136 return kzalloc(size, gfp_flags);
139 #define RESYNC_DEPTH 32
140 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
141 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
142 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
143 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
144 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
146 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
148 struct pool_info *pi = data;
149 struct r1bio *r1_bio;
153 struct resync_pages *rps;
155 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
159 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
165 * Allocate bios : 1 for reading, n-1 for writing
167 for (j = pi->raid_disks ; j-- ; ) {
168 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
171 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
172 r1_bio->bios[j] = bio;
175 * Allocate RESYNC_PAGES data pages and attach them to
177 * If this is a user-requested check/repair, allocate
178 * RESYNC_PAGES for each bio.
180 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
181 need_pages = pi->raid_disks;
184 for (j = 0; j < pi->raid_disks; j++) {
185 struct resync_pages *rp = &rps[j];
187 bio = r1_bio->bios[j];
189 if (j < need_pages) {
190 if (resync_alloc_pages(rp, gfp_flags))
193 memcpy(rp, &rps[0], sizeof(*rp));
194 resync_get_all_pages(rp);
197 rp->raid_bio = r1_bio;
198 bio->bi_private = rp;
201 r1_bio->master_bio = NULL;
207 resync_free_pages(&rps[j]);
210 while (++j < pi->raid_disks) {
211 bio_uninit(r1_bio->bios[j]);
212 kfree(r1_bio->bios[j]);
217 rbio_pool_free(r1_bio, data);
221 static void r1buf_pool_free(void *__r1_bio, void *data)
223 struct pool_info *pi = data;
225 struct r1bio *r1bio = __r1_bio;
226 struct resync_pages *rp = NULL;
228 for (i = pi->raid_disks; i--; ) {
229 rp = get_resync_pages(r1bio->bios[i]);
230 resync_free_pages(rp);
231 bio_uninit(r1bio->bios[i]);
232 kfree(r1bio->bios[i]);
235 /* resync pages array stored in the 1st bio's .bi_private */
238 rbio_pool_free(r1bio, data);
241 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
245 for (i = 0; i < conf->raid_disks * 2; i++) {
246 struct bio **bio = r1_bio->bios + i;
247 if (!BIO_SPECIAL(*bio))
253 static void free_r1bio(struct r1bio *r1_bio)
255 struct r1conf *conf = r1_bio->mddev->private;
257 put_all_bios(conf, r1_bio);
258 mempool_free(r1_bio, &conf->r1bio_pool);
261 static void put_buf(struct r1bio *r1_bio)
263 struct r1conf *conf = r1_bio->mddev->private;
264 sector_t sect = r1_bio->sector;
267 for (i = 0; i < conf->raid_disks * 2; i++) {
268 struct bio *bio = r1_bio->bios[i];
270 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
273 mempool_free(r1_bio, &conf->r1buf_pool);
275 lower_barrier(conf, sect);
278 static void reschedule_retry(struct r1bio *r1_bio)
281 struct mddev *mddev = r1_bio->mddev;
282 struct r1conf *conf = mddev->private;
285 idx = sector_to_idx(r1_bio->sector);
286 spin_lock_irqsave(&conf->device_lock, flags);
287 list_add(&r1_bio->retry_list, &conf->retry_list);
288 atomic_inc(&conf->nr_queued[idx]);
289 spin_unlock_irqrestore(&conf->device_lock, flags);
291 wake_up(&conf->wait_barrier);
292 md_wakeup_thread(mddev->thread);
296 * raid_end_bio_io() is called when we have finished servicing a mirrored
297 * operation and are ready to return a success/failure code to the buffer
300 static void call_bio_endio(struct r1bio *r1_bio)
302 struct bio *bio = r1_bio->master_bio;
304 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
305 bio->bi_status = BLK_STS_IOERR;
307 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
308 bio_end_io_acct(bio, r1_bio->start_time);
312 static void raid_end_bio_io(struct r1bio *r1_bio)
314 struct bio *bio = r1_bio->master_bio;
315 struct r1conf *conf = r1_bio->mddev->private;
317 /* if nobody has done the final endio yet, do it now */
318 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
319 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
320 (bio_data_dir(bio) == WRITE) ? "write" : "read",
321 (unsigned long long) bio->bi_iter.bi_sector,
322 (unsigned long long) bio_end_sector(bio) - 1);
324 call_bio_endio(r1_bio);
327 * Wake up any possible resync thread that waits for the device
328 * to go idle. All I/Os, even write-behind writes, are done.
330 allow_barrier(conf, r1_bio->sector);
336 * Update disk head position estimator based on IRQ completion info.
338 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
340 struct r1conf *conf = r1_bio->mddev->private;
342 conf->mirrors[disk].head_position =
343 r1_bio->sector + (r1_bio->sectors);
347 * Find the disk number which triggered given bio
349 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
352 struct r1conf *conf = r1_bio->mddev->private;
353 int raid_disks = conf->raid_disks;
355 for (mirror = 0; mirror < raid_disks * 2; mirror++)
356 if (r1_bio->bios[mirror] == bio)
359 BUG_ON(mirror == raid_disks * 2);
360 update_head_pos(mirror, r1_bio);
365 static void raid1_end_read_request(struct bio *bio)
367 int uptodate = !bio->bi_status;
368 struct r1bio *r1_bio = bio->bi_private;
369 struct r1conf *conf = r1_bio->mddev->private;
370 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
373 * this branch is our 'one mirror IO has finished' event handler:
375 update_head_pos(r1_bio->read_disk, r1_bio);
378 set_bit(R1BIO_Uptodate, &r1_bio->state);
379 else if (test_bit(FailFast, &rdev->flags) &&
380 test_bit(R1BIO_FailFast, &r1_bio->state))
381 /* This was a fail-fast read so we definitely
385 /* If all other devices have failed, we want to return
386 * the error upwards rather than fail the last device.
387 * Here we redefine "uptodate" to mean "Don't want to retry"
390 spin_lock_irqsave(&conf->device_lock, flags);
391 if (r1_bio->mddev->degraded == conf->raid_disks ||
392 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
393 test_bit(In_sync, &rdev->flags)))
395 spin_unlock_irqrestore(&conf->device_lock, flags);
399 raid_end_bio_io(r1_bio);
400 rdev_dec_pending(rdev, conf->mddev);
405 pr_err_ratelimited("md/raid1:%s: %pg: rescheduling sector %llu\n",
408 (unsigned long long)r1_bio->sector);
409 set_bit(R1BIO_ReadError, &r1_bio->state);
410 reschedule_retry(r1_bio);
411 /* don't drop the reference on read_disk yet */
415 static void close_write(struct r1bio *r1_bio)
417 /* it really is the end of this request */
418 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
419 bio_free_pages(r1_bio->behind_master_bio);
420 bio_put(r1_bio->behind_master_bio);
421 r1_bio->behind_master_bio = NULL;
423 /* clear the bitmap if all writes complete successfully */
424 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
426 !test_bit(R1BIO_Degraded, &r1_bio->state),
427 test_bit(R1BIO_BehindIO, &r1_bio->state));
428 md_write_end(r1_bio->mddev);
431 static void r1_bio_write_done(struct r1bio *r1_bio)
433 if (!atomic_dec_and_test(&r1_bio->remaining))
436 if (test_bit(R1BIO_WriteError, &r1_bio->state))
437 reschedule_retry(r1_bio);
440 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
441 reschedule_retry(r1_bio);
443 raid_end_bio_io(r1_bio);
447 static void raid1_end_write_request(struct bio *bio)
449 struct r1bio *r1_bio = bio->bi_private;
450 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
451 struct r1conf *conf = r1_bio->mddev->private;
452 struct bio *to_put = NULL;
453 int mirror = find_bio_disk(r1_bio, bio);
454 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
456 sector_t lo = r1_bio->sector;
457 sector_t hi = r1_bio->sector + r1_bio->sectors;
459 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
462 * 'one mirror IO has finished' event handler:
464 if (bio->bi_status && !discard_error) {
465 set_bit(WriteErrorSeen, &rdev->flags);
466 if (!test_and_set_bit(WantReplacement, &rdev->flags))
467 set_bit(MD_RECOVERY_NEEDED, &
468 conf->mddev->recovery);
470 if (test_bit(FailFast, &rdev->flags) &&
471 (bio->bi_opf & MD_FAILFAST) &&
472 /* We never try FailFast to WriteMostly devices */
473 !test_bit(WriteMostly, &rdev->flags)) {
474 md_error(r1_bio->mddev, rdev);
478 * When the device is faulty, it is not necessary to
479 * handle write error.
481 if (!test_bit(Faulty, &rdev->flags))
482 set_bit(R1BIO_WriteError, &r1_bio->state);
484 /* Fail the request */
485 set_bit(R1BIO_Degraded, &r1_bio->state);
486 /* Finished with this branch */
487 r1_bio->bios[mirror] = NULL;
492 * Set R1BIO_Uptodate in our master bio, so that we
493 * will return a good error code for to the higher
494 * levels even if IO on some other mirrored buffer
497 * The 'master' represents the composite IO operation
498 * to user-side. So if something waits for IO, then it
499 * will wait for the 'master' bio.
504 r1_bio->bios[mirror] = NULL;
507 * Do not set R1BIO_Uptodate if the current device is
508 * rebuilding or Faulty. This is because we cannot use
509 * such device for properly reading the data back (we could
510 * potentially use it, if the current write would have felt
511 * before rdev->recovery_offset, but for simplicity we don't
514 if (test_bit(In_sync, &rdev->flags) &&
515 !test_bit(Faulty, &rdev->flags))
516 set_bit(R1BIO_Uptodate, &r1_bio->state);
518 /* Maybe we can clear some bad blocks. */
519 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
520 &first_bad, &bad_sectors) && !discard_error) {
521 r1_bio->bios[mirror] = IO_MADE_GOOD;
522 set_bit(R1BIO_MadeGood, &r1_bio->state);
527 if (test_bit(CollisionCheck, &rdev->flags))
528 remove_serial(rdev, lo, hi);
529 if (test_bit(WriteMostly, &rdev->flags))
530 atomic_dec(&r1_bio->behind_remaining);
533 * In behind mode, we ACK the master bio once the I/O
534 * has safely reached all non-writemostly
535 * disks. Setting the Returned bit ensures that this
536 * gets done only once -- we don't ever want to return
537 * -EIO here, instead we'll wait
539 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
540 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
541 /* Maybe we can return now */
542 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
543 struct bio *mbio = r1_bio->master_bio;
544 pr_debug("raid1: behind end write sectors"
546 (unsigned long long) mbio->bi_iter.bi_sector,
547 (unsigned long long) bio_end_sector(mbio) - 1);
548 call_bio_endio(r1_bio);
551 } else if (rdev->mddev->serialize_policy)
552 remove_serial(rdev, lo, hi);
553 if (r1_bio->bios[mirror] == NULL)
554 rdev_dec_pending(rdev, conf->mddev);
557 * Let's see if all mirrored write operations have finished
560 r1_bio_write_done(r1_bio);
566 static sector_t align_to_barrier_unit_end(sector_t start_sector,
571 WARN_ON(sectors == 0);
573 * len is the number of sectors from start_sector to end of the
574 * barrier unit which start_sector belongs to.
576 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
586 * This routine returns the disk from which the requested read should
587 * be done. There is a per-array 'next expected sequential IO' sector
588 * number - if this matches on the next IO then we use the last disk.
589 * There is also a per-disk 'last know head position' sector that is
590 * maintained from IRQ contexts, both the normal and the resync IO
591 * completion handlers update this position correctly. If there is no
592 * perfect sequential match then we pick the disk whose head is closest.
594 * If there are 2 mirrors in the same 2 devices, performance degrades
595 * because position is mirror, not device based.
597 * The rdev for the device selected will have nr_pending incremented.
599 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
601 const sector_t this_sector = r1_bio->sector;
603 int best_good_sectors;
604 int best_disk, best_dist_disk, best_pending_disk;
608 unsigned int min_pending;
609 struct md_rdev *rdev;
611 int choose_next_idle;
615 * Check if we can balance. We can balance on the whole
616 * device if no resync is going on, or below the resync window.
617 * We take the first readable disk when above the resync window.
620 sectors = r1_bio->sectors;
623 best_dist = MaxSector;
624 best_pending_disk = -1;
625 min_pending = UINT_MAX;
626 best_good_sectors = 0;
628 choose_next_idle = 0;
629 clear_bit(R1BIO_FailFast, &r1_bio->state);
631 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
632 (mddev_is_clustered(conf->mddev) &&
633 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
634 this_sector + sectors)))
639 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
643 unsigned int pending;
646 rdev = rcu_dereference(conf->mirrors[disk].rdev);
647 if (r1_bio->bios[disk] == IO_BLOCKED
649 || test_bit(Faulty, &rdev->flags))
651 if (!test_bit(In_sync, &rdev->flags) &&
652 rdev->recovery_offset < this_sector + sectors)
654 if (test_bit(WriteMostly, &rdev->flags)) {
655 /* Don't balance among write-mostly, just
656 * use the first as a last resort */
657 if (best_dist_disk < 0) {
658 if (is_badblock(rdev, this_sector, sectors,
659 &first_bad, &bad_sectors)) {
660 if (first_bad <= this_sector)
661 /* Cannot use this */
663 best_good_sectors = first_bad - this_sector;
665 best_good_sectors = sectors;
666 best_dist_disk = disk;
667 best_pending_disk = disk;
671 /* This is a reasonable device to use. It might
674 if (is_badblock(rdev, this_sector, sectors,
675 &first_bad, &bad_sectors)) {
676 if (best_dist < MaxSector)
677 /* already have a better device */
679 if (first_bad <= this_sector) {
680 /* cannot read here. If this is the 'primary'
681 * device, then we must not read beyond
682 * bad_sectors from another device..
684 bad_sectors -= (this_sector - first_bad);
685 if (choose_first && sectors > bad_sectors)
686 sectors = bad_sectors;
687 if (best_good_sectors > sectors)
688 best_good_sectors = sectors;
691 sector_t good_sectors = first_bad - this_sector;
692 if (good_sectors > best_good_sectors) {
693 best_good_sectors = good_sectors;
701 if ((sectors > best_good_sectors) && (best_disk >= 0))
703 best_good_sectors = sectors;
707 /* At least two disks to choose from so failfast is OK */
708 set_bit(R1BIO_FailFast, &r1_bio->state);
710 nonrot = bdev_nonrot(rdev->bdev);
711 has_nonrot_disk |= nonrot;
712 pending = atomic_read(&rdev->nr_pending);
713 dist = abs(this_sector - conf->mirrors[disk].head_position);
718 /* Don't change to another disk for sequential reads */
719 if (conf->mirrors[disk].next_seq_sect == this_sector
721 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
722 struct raid1_info *mirror = &conf->mirrors[disk];
726 * If buffered sequential IO size exceeds optimal
727 * iosize, check if there is idle disk. If yes, choose
728 * the idle disk. read_balance could already choose an
729 * idle disk before noticing it's a sequential IO in
730 * this disk. This doesn't matter because this disk
731 * will idle, next time it will be utilized after the
732 * first disk has IO size exceeds optimal iosize. In
733 * this way, iosize of the first disk will be optimal
734 * iosize at least. iosize of the second disk might be
735 * small, but not a big deal since when the second disk
736 * starts IO, the first disk is likely still busy.
738 if (nonrot && opt_iosize > 0 &&
739 mirror->seq_start != MaxSector &&
740 mirror->next_seq_sect > opt_iosize &&
741 mirror->next_seq_sect - opt_iosize >=
743 choose_next_idle = 1;
749 if (choose_next_idle)
752 if (min_pending > pending) {
753 min_pending = pending;
754 best_pending_disk = disk;
757 if (dist < best_dist) {
759 best_dist_disk = disk;
764 * If all disks are rotational, choose the closest disk. If any disk is
765 * non-rotational, choose the disk with less pending request even the
766 * disk is rotational, which might/might not be optimal for raids with
767 * mixed ratation/non-rotational disks depending on workload.
769 if (best_disk == -1) {
770 if (has_nonrot_disk || min_pending == 0)
771 best_disk = best_pending_disk;
773 best_disk = best_dist_disk;
776 if (best_disk >= 0) {
777 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
780 atomic_inc(&rdev->nr_pending);
781 sectors = best_good_sectors;
783 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
784 conf->mirrors[best_disk].seq_start = this_sector;
786 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
789 *max_sectors = sectors;
794 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
796 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
797 raid1_prepare_flush_writes(conf->mddev->bitmap);
798 wake_up(&conf->wait_barrier);
800 while (bio) { /* submit pending writes */
801 struct bio *next = bio->bi_next;
803 raid1_submit_write(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 spin_unlock_irq(&conf->device_lock);
824 * As this is called in a wait_event() loop (see freeze_array),
825 * current->state might be TASK_UNINTERRUPTIBLE which will
826 * cause a warning when we prepare to wait again. As it is
827 * rare that this path is taken, it is perfectly safe to force
828 * us to go around the wait_event() loop again, so the warning
829 * is a false-positive. Silence the warning by resetting
832 __set_current_state(TASK_RUNNING);
833 blk_start_plug(&plug);
834 flush_bio_list(conf, bio);
835 blk_finish_plug(&plug);
837 spin_unlock_irq(&conf->device_lock);
841 * Sometimes we need to suspend IO while we do something else,
842 * either some resync/recovery, or reconfigure the array.
843 * To do this we raise a 'barrier'.
844 * The 'barrier' is a counter that can be raised multiple times
845 * to count how many activities are happening which preclude
847 * We can only raise the barrier if there is no pending IO.
848 * i.e. if nr_pending == 0.
849 * We choose only to raise the barrier if no-one is waiting for the
850 * barrier to go down. This means that as soon as an IO request
851 * is ready, no other operations which require a barrier will start
852 * until the IO request has had a chance.
854 * So: regular IO calls 'wait_barrier'. When that returns there
855 * is no backgroup IO happening, It must arrange to call
856 * allow_barrier when it has finished its IO.
857 * backgroup IO calls must call raise_barrier. Once that returns
858 * there is no normal IO happeing. It must arrange to call
859 * lower_barrier when the particular background IO completes.
861 * If resync/recovery is interrupted, returns -EINTR;
862 * Otherwise, returns 0.
864 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
866 int idx = sector_to_idx(sector_nr);
868 spin_lock_irq(&conf->resync_lock);
870 /* Wait until no block IO is waiting */
871 wait_event_lock_irq(conf->wait_barrier,
872 !atomic_read(&conf->nr_waiting[idx]),
875 /* block any new IO from starting */
876 atomic_inc(&conf->barrier[idx]);
878 * In raise_barrier() we firstly increase conf->barrier[idx] then
879 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
880 * increase conf->nr_pending[idx] then check conf->barrier[idx].
881 * A memory barrier here to make sure conf->nr_pending[idx] won't
882 * be fetched before conf->barrier[idx] is increased. Otherwise
883 * there will be a race between raise_barrier() and _wait_barrier().
885 smp_mb__after_atomic();
887 /* For these conditions we must wait:
888 * A: while the array is in frozen state
889 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
890 * existing in corresponding I/O barrier bucket.
891 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
892 * max resync count which allowed on current I/O barrier bucket.
894 wait_event_lock_irq(conf->wait_barrier,
895 (!conf->array_frozen &&
896 !atomic_read(&conf->nr_pending[idx]) &&
897 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
898 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
901 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
902 atomic_dec(&conf->barrier[idx]);
903 spin_unlock_irq(&conf->resync_lock);
904 wake_up(&conf->wait_barrier);
908 atomic_inc(&conf->nr_sync_pending);
909 spin_unlock_irq(&conf->resync_lock);
914 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
916 int idx = sector_to_idx(sector_nr);
918 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
920 atomic_dec(&conf->barrier[idx]);
921 atomic_dec(&conf->nr_sync_pending);
922 wake_up(&conf->wait_barrier);
925 static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
930 * We need to increase conf->nr_pending[idx] very early here,
931 * then raise_barrier() can be blocked when it waits for
932 * conf->nr_pending[idx] to be 0. Then we can avoid holding
933 * conf->resync_lock when there is no barrier raised in same
934 * barrier unit bucket. Also if the array is frozen, I/O
935 * should be blocked until array is unfrozen.
937 atomic_inc(&conf->nr_pending[idx]);
939 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
940 * check conf->barrier[idx]. In raise_barrier() we firstly increase
941 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
942 * barrier is necessary here to make sure conf->barrier[idx] won't be
943 * fetched before conf->nr_pending[idx] is increased. Otherwise there
944 * will be a race between _wait_barrier() and raise_barrier().
946 smp_mb__after_atomic();
949 * Don't worry about checking two atomic_t variables at same time
950 * here. If during we check conf->barrier[idx], the array is
951 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
952 * 0, it is safe to return and make the I/O continue. Because the
953 * array is frozen, all I/O returned here will eventually complete
954 * or be queued, no race will happen. See code comment in
957 if (!READ_ONCE(conf->array_frozen) &&
958 !atomic_read(&conf->barrier[idx]))
962 * After holding conf->resync_lock, conf->nr_pending[idx]
963 * should be decreased before waiting for barrier to drop.
964 * Otherwise, we may encounter a race condition because
965 * raise_barrer() might be waiting for conf->nr_pending[idx]
966 * to be 0 at same time.
968 spin_lock_irq(&conf->resync_lock);
969 atomic_inc(&conf->nr_waiting[idx]);
970 atomic_dec(&conf->nr_pending[idx]);
972 * In case freeze_array() is waiting for
973 * get_unqueued_pending() == extra
975 wake_up(&conf->wait_barrier);
976 /* Wait for the barrier in same barrier unit bucket to drop. */
978 /* Return false when nowait flag is set */
982 wait_event_lock_irq(conf->wait_barrier,
983 !conf->array_frozen &&
984 !atomic_read(&conf->barrier[idx]),
986 atomic_inc(&conf->nr_pending[idx]);
989 atomic_dec(&conf->nr_waiting[idx]);
990 spin_unlock_irq(&conf->resync_lock);
994 static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
996 int idx = sector_to_idx(sector_nr);
1000 * Very similar to _wait_barrier(). The difference is, for read
1001 * I/O we don't need wait for sync I/O, but if the whole array
1002 * is frozen, the read I/O still has to wait until the array is
1003 * unfrozen. Since there is no ordering requirement with
1004 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1006 atomic_inc(&conf->nr_pending[idx]);
1008 if (!READ_ONCE(conf->array_frozen))
1011 spin_lock_irq(&conf->resync_lock);
1012 atomic_inc(&conf->nr_waiting[idx]);
1013 atomic_dec(&conf->nr_pending[idx]);
1015 * In case freeze_array() is waiting for
1016 * get_unqueued_pending() == extra
1018 wake_up(&conf->wait_barrier);
1019 /* Wait for array to be unfrozen */
1021 /* Return false when nowait flag is set */
1023 /* Return false when nowait flag is set */
1026 wait_event_lock_irq(conf->wait_barrier,
1027 !conf->array_frozen,
1029 atomic_inc(&conf->nr_pending[idx]);
1032 atomic_dec(&conf->nr_waiting[idx]);
1033 spin_unlock_irq(&conf->resync_lock);
1037 static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1039 int idx = sector_to_idx(sector_nr);
1041 return _wait_barrier(conf, idx, nowait);
1044 static void _allow_barrier(struct r1conf *conf, int idx)
1046 atomic_dec(&conf->nr_pending[idx]);
1047 wake_up(&conf->wait_barrier);
1050 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1052 int idx = sector_to_idx(sector_nr);
1054 _allow_barrier(conf, idx);
1057 /* conf->resync_lock should be held */
1058 static int get_unqueued_pending(struct r1conf *conf)
1062 ret = atomic_read(&conf->nr_sync_pending);
1063 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1064 ret += atomic_read(&conf->nr_pending[idx]) -
1065 atomic_read(&conf->nr_queued[idx]);
1070 static void freeze_array(struct r1conf *conf, int extra)
1072 /* Stop sync I/O and normal I/O and wait for everything to
1074 * This is called in two situations:
1075 * 1) management command handlers (reshape, remove disk, quiesce).
1076 * 2) one normal I/O request failed.
1078 * After array_frozen is set to 1, new sync IO will be blocked at
1079 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1080 * or wait_read_barrier(). The flying I/Os will either complete or be
1081 * queued. When everything goes quite, there are only queued I/Os left.
1083 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1084 * barrier bucket index which this I/O request hits. When all sync and
1085 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1086 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1087 * in handle_read_error(), we may call freeze_array() before trying to
1088 * fix the read error. In this case, the error read I/O is not queued,
1089 * so get_unqueued_pending() == 1.
1091 * Therefore before this function returns, we need to wait until
1092 * get_unqueued_pendings(conf) gets equal to extra. For
1093 * normal I/O context, extra is 1, in rested situations extra is 0.
1095 spin_lock_irq(&conf->resync_lock);
1096 conf->array_frozen = 1;
1097 raid1_log(conf->mddev, "wait freeze");
1098 wait_event_lock_irq_cmd(
1100 get_unqueued_pending(conf) == extra,
1102 flush_pending_writes(conf));
1103 spin_unlock_irq(&conf->resync_lock);
1105 static void unfreeze_array(struct r1conf *conf)
1107 /* reverse the effect of the freeze */
1108 spin_lock_irq(&conf->resync_lock);
1109 conf->array_frozen = 0;
1110 spin_unlock_irq(&conf->resync_lock);
1111 wake_up(&conf->wait_barrier);
1114 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1117 int size = bio->bi_iter.bi_size;
1118 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1120 struct bio *behind_bio = NULL;
1122 behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1123 &r1_bio->mddev->bio_set);
1127 /* discard op, we don't support writezero/writesame yet */
1128 if (!bio_has_data(bio)) {
1129 behind_bio->bi_iter.bi_size = size;
1133 while (i < vcnt && size) {
1135 int len = min_t(int, PAGE_SIZE, size);
1137 page = alloc_page(GFP_NOIO);
1138 if (unlikely(!page))
1141 if (!bio_add_page(behind_bio, page, len, 0)) {
1150 bio_copy_data(behind_bio, bio);
1152 r1_bio->behind_master_bio = behind_bio;
1153 set_bit(R1BIO_BehindIO, &r1_bio->state);
1158 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1159 bio->bi_iter.bi_size);
1160 bio_free_pages(behind_bio);
1161 bio_put(behind_bio);
1164 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1166 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1168 struct mddev *mddev = plug->cb.data;
1169 struct r1conf *conf = mddev->private;
1172 if (from_schedule) {
1173 spin_lock_irq(&conf->device_lock);
1174 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1175 spin_unlock_irq(&conf->device_lock);
1176 wake_up(&conf->wait_barrier);
1177 md_wakeup_thread(mddev->thread);
1182 /* we aren't scheduling, so we can do the write-out directly. */
1183 bio = bio_list_get(&plug->pending);
1184 flush_bio_list(conf, bio);
1188 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1190 r1_bio->master_bio = bio;
1191 r1_bio->sectors = bio_sectors(bio);
1193 r1_bio->mddev = mddev;
1194 r1_bio->sector = bio->bi_iter.bi_sector;
1197 static inline struct r1bio *
1198 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1200 struct r1conf *conf = mddev->private;
1201 struct r1bio *r1_bio;
1203 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1204 /* Ensure no bio records IO_BLOCKED */
1205 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1206 init_r1bio(r1_bio, mddev, bio);
1210 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1211 int max_read_sectors, struct r1bio *r1_bio)
1213 struct r1conf *conf = mddev->private;
1214 struct raid1_info *mirror;
1215 struct bio *read_bio;
1216 struct bitmap *bitmap = mddev->bitmap;
1217 const enum req_op op = bio_op(bio);
1218 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1221 bool r1bio_existed = !!r1_bio;
1222 char b[BDEVNAME_SIZE];
1225 * If r1_bio is set, we are blocking the raid1d thread
1226 * so there is a tiny risk of deadlock. So ask for
1227 * emergency memory if needed.
1229 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1231 if (r1bio_existed) {
1232 /* Need to get the block device name carefully */
1233 struct md_rdev *rdev;
1235 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1237 snprintf(b, sizeof(b), "%pg", rdev->bdev);
1244 * Still need barrier for READ in case that whole
1247 if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1248 bio->bi_opf & REQ_NOWAIT)) {
1249 bio_wouldblock_error(bio);
1254 r1_bio = alloc_r1bio(mddev, bio);
1256 init_r1bio(r1_bio, mddev, bio);
1257 r1_bio->sectors = max_read_sectors;
1260 * make_request() can abort the operation when read-ahead is being
1261 * used and no empty request is available.
1263 rdisk = read_balance(conf, r1_bio, &max_sectors);
1266 /* couldn't find anywhere to read from */
1267 if (r1bio_existed) {
1268 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1271 (unsigned long long)r1_bio->sector);
1273 raid_end_bio_io(r1_bio);
1276 mirror = conf->mirrors + rdisk;
1279 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %pg\n",
1281 (unsigned long long)r1_bio->sector,
1282 mirror->rdev->bdev);
1284 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1287 * Reading from a write-mostly device must take care not to
1288 * over-take any writes that are 'behind'
1290 raid1_log(mddev, "wait behind writes");
1291 wait_event(bitmap->behind_wait,
1292 atomic_read(&bitmap->behind_writes) == 0);
1295 if (max_sectors < bio_sectors(bio)) {
1296 struct bio *split = bio_split(bio, max_sectors,
1297 gfp, &conf->bio_split);
1298 bio_chain(split, bio);
1299 submit_bio_noacct(bio);
1301 r1_bio->master_bio = bio;
1302 r1_bio->sectors = max_sectors;
1305 r1_bio->read_disk = rdisk;
1307 if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1308 r1_bio->start_time = bio_start_io_acct(bio);
1310 read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1313 r1_bio->bios[rdisk] = read_bio;
1315 read_bio->bi_iter.bi_sector = r1_bio->sector +
1316 mirror->rdev->data_offset;
1317 read_bio->bi_end_io = raid1_end_read_request;
1318 read_bio->bi_opf = op | do_sync;
1319 if (test_bit(FailFast, &mirror->rdev->flags) &&
1320 test_bit(R1BIO_FailFast, &r1_bio->state))
1321 read_bio->bi_opf |= MD_FAILFAST;
1322 read_bio->bi_private = r1_bio;
1325 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1328 submit_bio_noacct(read_bio);
1331 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1332 int max_write_sectors)
1334 struct r1conf *conf = mddev->private;
1335 struct r1bio *r1_bio;
1337 struct bitmap *bitmap = mddev->bitmap;
1338 unsigned long flags;
1339 struct md_rdev *blocked_rdev;
1342 bool write_behind = false;
1344 if (mddev_is_clustered(mddev) &&
1345 md_cluster_ops->area_resyncing(mddev, WRITE,
1346 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1349 if (bio->bi_opf & REQ_NOWAIT) {
1350 bio_wouldblock_error(bio);
1354 prepare_to_wait(&conf->wait_barrier,
1356 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1357 bio->bi_iter.bi_sector,
1358 bio_end_sector(bio)))
1362 finish_wait(&conf->wait_barrier, &w);
1366 * Register the new request and wait if the reconstruction
1367 * thread has put up a bar for new requests.
1368 * Continue immediately if no resync is active currently.
1370 if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1371 bio->bi_opf & REQ_NOWAIT)) {
1372 bio_wouldblock_error(bio);
1376 r1_bio = alloc_r1bio(mddev, bio);
1377 r1_bio->sectors = max_write_sectors;
1379 /* first select target devices under rcu_lock and
1380 * inc refcount on their rdev. Record them by setting
1382 * If there are known/acknowledged bad blocks on any device on
1383 * which we have seen a write error, we want to avoid writing those
1385 * This potentially requires several writes to write around
1386 * the bad blocks. Each set of writes gets it's own r1bio
1387 * with a set of bios attached.
1390 disks = conf->raid_disks * 2;
1392 blocked_rdev = NULL;
1394 max_sectors = r1_bio->sectors;
1395 for (i = 0; i < disks; i++) {
1396 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1399 * The write-behind io is only attempted on drives marked as
1400 * write-mostly, which means we could allocate write behind
1403 if (rdev && test_bit(WriteMostly, &rdev->flags))
1404 write_behind = true;
1406 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1407 atomic_inc(&rdev->nr_pending);
1408 blocked_rdev = rdev;
1411 r1_bio->bios[i] = NULL;
1412 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1413 if (i < conf->raid_disks)
1414 set_bit(R1BIO_Degraded, &r1_bio->state);
1418 atomic_inc(&rdev->nr_pending);
1419 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1424 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1425 &first_bad, &bad_sectors);
1427 /* mustn't write here until the bad block is
1429 set_bit(BlockedBadBlocks, &rdev->flags);
1430 blocked_rdev = rdev;
1433 if (is_bad && first_bad <= r1_bio->sector) {
1434 /* Cannot write here at all */
1435 bad_sectors -= (r1_bio->sector - first_bad);
1436 if (bad_sectors < max_sectors)
1437 /* mustn't write more than bad_sectors
1438 * to other devices yet
1440 max_sectors = bad_sectors;
1441 rdev_dec_pending(rdev, mddev);
1442 /* We don't set R1BIO_Degraded as that
1443 * only applies if the disk is
1444 * missing, so it might be re-added,
1445 * and we want to know to recover this
1447 * In this case the device is here,
1448 * and the fact that this chunk is not
1449 * in-sync is recorded in the bad
1455 int good_sectors = first_bad - r1_bio->sector;
1456 if (good_sectors < max_sectors)
1457 max_sectors = good_sectors;
1460 r1_bio->bios[i] = bio;
1464 if (unlikely(blocked_rdev)) {
1465 /* Wait for this device to become unblocked */
1468 for (j = 0; j < i; j++)
1469 if (r1_bio->bios[j])
1470 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1472 allow_barrier(conf, bio->bi_iter.bi_sector);
1474 if (bio->bi_opf & REQ_NOWAIT) {
1475 bio_wouldblock_error(bio);
1478 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1479 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1480 wait_barrier(conf, bio->bi_iter.bi_sector, false);
1485 * When using a bitmap, we may call alloc_behind_master_bio below.
1486 * alloc_behind_master_bio allocates a copy of the data payload a page
1487 * at a time and thus needs a new bio that can fit the whole payload
1488 * this bio in page sized chunks.
1490 if (write_behind && bitmap)
1491 max_sectors = min_t(int, max_sectors,
1492 BIO_MAX_VECS * (PAGE_SIZE >> 9));
1493 if (max_sectors < bio_sectors(bio)) {
1494 struct bio *split = bio_split(bio, max_sectors,
1495 GFP_NOIO, &conf->bio_split);
1496 bio_chain(split, bio);
1497 submit_bio_noacct(bio);
1499 r1_bio->master_bio = bio;
1500 r1_bio->sectors = max_sectors;
1503 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1504 r1_bio->start_time = bio_start_io_acct(bio);
1505 atomic_set(&r1_bio->remaining, 1);
1506 atomic_set(&r1_bio->behind_remaining, 0);
1510 for (i = 0; i < disks; i++) {
1511 struct bio *mbio = NULL;
1512 struct md_rdev *rdev = conf->mirrors[i].rdev;
1513 if (!r1_bio->bios[i])
1518 * Not if there are too many, or cannot
1519 * allocate memory, or a reader on WriteMostly
1520 * is waiting for behind writes to flush */
1522 test_bit(WriteMostly, &rdev->flags) &&
1523 (atomic_read(&bitmap->behind_writes)
1524 < mddev->bitmap_info.max_write_behind) &&
1525 !waitqueue_active(&bitmap->behind_wait)) {
1526 alloc_behind_master_bio(r1_bio, bio);
1529 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1530 test_bit(R1BIO_BehindIO, &r1_bio->state));
1534 if (r1_bio->behind_master_bio) {
1535 mbio = bio_alloc_clone(rdev->bdev,
1536 r1_bio->behind_master_bio,
1537 GFP_NOIO, &mddev->bio_set);
1538 if (test_bit(CollisionCheck, &rdev->flags))
1539 wait_for_serialization(rdev, r1_bio);
1540 if (test_bit(WriteMostly, &rdev->flags))
1541 atomic_inc(&r1_bio->behind_remaining);
1543 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1546 if (mddev->serialize_policy)
1547 wait_for_serialization(rdev, r1_bio);
1550 r1_bio->bios[i] = mbio;
1552 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1553 mbio->bi_end_io = raid1_end_write_request;
1554 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1555 if (test_bit(FailFast, &rdev->flags) &&
1556 !test_bit(WriteMostly, &rdev->flags) &&
1557 conf->raid_disks - mddev->degraded > 1)
1558 mbio->bi_opf |= MD_FAILFAST;
1559 mbio->bi_private = r1_bio;
1561 atomic_inc(&r1_bio->remaining);
1564 trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1566 /* flush_pending_writes() needs access to the rdev so...*/
1567 mbio->bi_bdev = (void *)rdev;
1568 if (!raid1_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) {
1569 spin_lock_irqsave(&conf->device_lock, flags);
1570 bio_list_add(&conf->pending_bio_list, mbio);
1571 spin_unlock_irqrestore(&conf->device_lock, flags);
1572 md_wakeup_thread(mddev->thread);
1576 r1_bio_write_done(r1_bio);
1578 /* In case raid1d snuck in to freeze_array */
1579 wake_up(&conf->wait_barrier);
1582 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1586 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1587 && md_flush_request(mddev, bio))
1591 * There is a limit to the maximum size, but
1592 * the read/write handler might find a lower limit
1593 * due to bad blocks. To avoid multiple splits,
1594 * we pass the maximum number of sectors down
1595 * and let the lower level perform the split.
1597 sectors = align_to_barrier_unit_end(
1598 bio->bi_iter.bi_sector, bio_sectors(bio));
1600 if (bio_data_dir(bio) == READ)
1601 raid1_read_request(mddev, bio, sectors, NULL);
1603 if (!md_write_start(mddev,bio))
1605 raid1_write_request(mddev, bio, sectors);
1610 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1612 struct r1conf *conf = mddev->private;
1615 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1616 conf->raid_disks - mddev->degraded);
1618 for (i = 0; i < conf->raid_disks; i++) {
1619 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1620 seq_printf(seq, "%s",
1621 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1624 seq_printf(seq, "]");
1628 * raid1_error() - RAID1 error handler.
1629 * @mddev: affected md device.
1630 * @rdev: member device to fail.
1632 * The routine acknowledges &rdev failure and determines new @mddev state.
1633 * If it failed, then:
1634 * - &MD_BROKEN flag is set in &mddev->flags.
1635 * - recovery is disabled.
1636 * Otherwise, it must be degraded:
1637 * - recovery is interrupted.
1638 * - &mddev->degraded is bumped.
1640 * @rdev is marked as &Faulty excluding case when array is failed and
1641 * &mddev->fail_last_dev is off.
1643 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1645 struct r1conf *conf = mddev->private;
1646 unsigned long flags;
1648 spin_lock_irqsave(&conf->device_lock, flags);
1650 if (test_bit(In_sync, &rdev->flags) &&
1651 (conf->raid_disks - mddev->degraded) == 1) {
1652 set_bit(MD_BROKEN, &mddev->flags);
1654 if (!mddev->fail_last_dev) {
1655 conf->recovery_disabled = mddev->recovery_disabled;
1656 spin_unlock_irqrestore(&conf->device_lock, flags);
1660 set_bit(Blocked, &rdev->flags);
1661 if (test_and_clear_bit(In_sync, &rdev->flags))
1663 set_bit(Faulty, &rdev->flags);
1664 spin_unlock_irqrestore(&conf->device_lock, flags);
1666 * if recovery is running, make sure it aborts.
1668 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1669 set_mask_bits(&mddev->sb_flags, 0,
1670 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1671 pr_crit("md/raid1:%s: Disk failure on %pg, disabling device.\n"
1672 "md/raid1:%s: Operation continuing on %d devices.\n",
1673 mdname(mddev), rdev->bdev,
1674 mdname(mddev), conf->raid_disks - mddev->degraded);
1677 static void print_conf(struct r1conf *conf)
1681 pr_debug("RAID1 conf printout:\n");
1683 pr_debug("(!conf)\n");
1686 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1690 for (i = 0; i < conf->raid_disks; i++) {
1691 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1693 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
1694 i, !test_bit(In_sync, &rdev->flags),
1695 !test_bit(Faulty, &rdev->flags),
1701 static void close_sync(struct r1conf *conf)
1705 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1706 _wait_barrier(conf, idx, false);
1707 _allow_barrier(conf, idx);
1710 mempool_exit(&conf->r1buf_pool);
1713 static int raid1_spare_active(struct mddev *mddev)
1716 struct r1conf *conf = mddev->private;
1718 unsigned long flags;
1721 * Find all failed disks within the RAID1 configuration
1722 * and mark them readable.
1723 * Called under mddev lock, so rcu protection not needed.
1724 * device_lock used to avoid races with raid1_end_read_request
1725 * which expects 'In_sync' flags and ->degraded to be consistent.
1727 spin_lock_irqsave(&conf->device_lock, flags);
1728 for (i = 0; i < conf->raid_disks; i++) {
1729 struct md_rdev *rdev = conf->mirrors[i].rdev;
1730 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1732 && !test_bit(Candidate, &repl->flags)
1733 && repl->recovery_offset == MaxSector
1734 && !test_bit(Faulty, &repl->flags)
1735 && !test_and_set_bit(In_sync, &repl->flags)) {
1736 /* replacement has just become active */
1738 !test_and_clear_bit(In_sync, &rdev->flags))
1741 /* Replaced device not technically
1742 * faulty, but we need to be sure
1743 * it gets removed and never re-added
1745 set_bit(Faulty, &rdev->flags);
1746 sysfs_notify_dirent_safe(
1751 && rdev->recovery_offset == MaxSector
1752 && !test_bit(Faulty, &rdev->flags)
1753 && !test_and_set_bit(In_sync, &rdev->flags)) {
1755 sysfs_notify_dirent_safe(rdev->sysfs_state);
1758 mddev->degraded -= count;
1759 spin_unlock_irqrestore(&conf->device_lock, flags);
1765 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1767 struct r1conf *conf = mddev->private;
1770 struct raid1_info *p;
1772 int last = conf->raid_disks - 1;
1774 if (mddev->recovery_disabled == conf->recovery_disabled)
1777 if (md_integrity_add_rdev(rdev, mddev))
1780 if (rdev->raid_disk >= 0)
1781 first = last = rdev->raid_disk;
1784 * find the disk ... but prefer rdev->saved_raid_disk
1787 if (rdev->saved_raid_disk >= 0 &&
1788 rdev->saved_raid_disk >= first &&
1789 rdev->saved_raid_disk < conf->raid_disks &&
1790 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1791 first = last = rdev->saved_raid_disk;
1793 for (mirror = first; mirror <= last; mirror++) {
1794 p = conf->mirrors + mirror;
1797 disk_stack_limits(mddev->gendisk, rdev->bdev,
1798 rdev->data_offset << 9);
1800 p->head_position = 0;
1801 rdev->raid_disk = mirror;
1803 /* As all devices are equivalent, we don't need a full recovery
1804 * if this was recently any drive of the array
1806 if (rdev->saved_raid_disk < 0)
1808 rcu_assign_pointer(p->rdev, rdev);
1811 if (test_bit(WantReplacement, &p->rdev->flags) &&
1812 p[conf->raid_disks].rdev == NULL) {
1813 /* Add this device as a replacement */
1814 clear_bit(In_sync, &rdev->flags);
1815 set_bit(Replacement, &rdev->flags);
1816 rdev->raid_disk = mirror;
1819 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1827 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1829 struct r1conf *conf = mddev->private;
1831 int number = rdev->raid_disk;
1832 struct raid1_info *p = conf->mirrors + number;
1834 if (rdev != p->rdev)
1835 p = conf->mirrors + conf->raid_disks + number;
1838 if (rdev == p->rdev) {
1839 if (test_bit(In_sync, &rdev->flags) ||
1840 atomic_read(&rdev->nr_pending)) {
1844 /* Only remove non-faulty devices if recovery
1847 if (!test_bit(Faulty, &rdev->flags) &&
1848 mddev->recovery_disabled != conf->recovery_disabled &&
1849 mddev->degraded < conf->raid_disks) {
1854 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1856 if (atomic_read(&rdev->nr_pending)) {
1857 /* lost the race, try later */
1863 if (conf->mirrors[conf->raid_disks + number].rdev) {
1864 /* We just removed a device that is being replaced.
1865 * Move down the replacement. We drain all IO before
1866 * doing this to avoid confusion.
1868 struct md_rdev *repl =
1869 conf->mirrors[conf->raid_disks + number].rdev;
1870 freeze_array(conf, 0);
1871 if (atomic_read(&repl->nr_pending)) {
1872 /* It means that some queued IO of retry_list
1873 * hold repl. Thus, we cannot set replacement
1874 * as NULL, avoiding rdev NULL pointer
1875 * dereference in sync_request_write and
1876 * handle_write_finished.
1879 unfreeze_array(conf);
1882 clear_bit(Replacement, &repl->flags);
1884 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1885 unfreeze_array(conf);
1888 clear_bit(WantReplacement, &rdev->flags);
1889 err = md_integrity_register(mddev);
1897 static void end_sync_read(struct bio *bio)
1899 struct r1bio *r1_bio = get_resync_r1bio(bio);
1901 update_head_pos(r1_bio->read_disk, r1_bio);
1904 * we have read a block, now it needs to be re-written,
1905 * or re-read if the read failed.
1906 * We don't do much here, just schedule handling by raid1d
1908 if (!bio->bi_status)
1909 set_bit(R1BIO_Uptodate, &r1_bio->state);
1911 if (atomic_dec_and_test(&r1_bio->remaining))
1912 reschedule_retry(r1_bio);
1915 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1917 sector_t sync_blocks = 0;
1918 sector_t s = r1_bio->sector;
1919 long sectors_to_go = r1_bio->sectors;
1921 /* make sure these bits don't get cleared. */
1923 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1925 sectors_to_go -= sync_blocks;
1926 } while (sectors_to_go > 0);
1929 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1931 if (atomic_dec_and_test(&r1_bio->remaining)) {
1932 struct mddev *mddev = r1_bio->mddev;
1933 int s = r1_bio->sectors;
1935 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1936 test_bit(R1BIO_WriteError, &r1_bio->state))
1937 reschedule_retry(r1_bio);
1940 md_done_sync(mddev, s, uptodate);
1945 static void end_sync_write(struct bio *bio)
1947 int uptodate = !bio->bi_status;
1948 struct r1bio *r1_bio = get_resync_r1bio(bio);
1949 struct mddev *mddev = r1_bio->mddev;
1950 struct r1conf *conf = mddev->private;
1953 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1956 abort_sync_write(mddev, r1_bio);
1957 set_bit(WriteErrorSeen, &rdev->flags);
1958 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1959 set_bit(MD_RECOVERY_NEEDED, &
1961 set_bit(R1BIO_WriteError, &r1_bio->state);
1962 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1963 &first_bad, &bad_sectors) &&
1964 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1967 &first_bad, &bad_sectors)
1969 set_bit(R1BIO_MadeGood, &r1_bio->state);
1971 put_sync_write_buf(r1_bio, uptodate);
1974 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1975 int sectors, struct page *page, int rw)
1977 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1981 set_bit(WriteErrorSeen, &rdev->flags);
1982 if (!test_and_set_bit(WantReplacement,
1984 set_bit(MD_RECOVERY_NEEDED, &
1985 rdev->mddev->recovery);
1987 /* need to record an error - either for the block or the device */
1988 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1989 md_error(rdev->mddev, rdev);
1993 static int fix_sync_read_error(struct r1bio *r1_bio)
1995 /* Try some synchronous reads of other devices to get
1996 * good data, much like with normal read errors. Only
1997 * read into the pages we already have so we don't
1998 * need to re-issue the read request.
1999 * We don't need to freeze the array, because being in an
2000 * active sync request, there is no normal IO, and
2001 * no overlapping syncs.
2002 * We don't need to check is_badblock() again as we
2003 * made sure that anything with a bad block in range
2004 * will have bi_end_io clear.
2006 struct mddev *mddev = r1_bio->mddev;
2007 struct r1conf *conf = mddev->private;
2008 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2009 struct page **pages = get_resync_pages(bio)->pages;
2010 sector_t sect = r1_bio->sector;
2011 int sectors = r1_bio->sectors;
2013 struct md_rdev *rdev;
2015 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2016 if (test_bit(FailFast, &rdev->flags)) {
2017 /* Don't try recovering from here - just fail it
2018 * ... unless it is the last working device of course */
2019 md_error(mddev, rdev);
2020 if (test_bit(Faulty, &rdev->flags))
2021 /* Don't try to read from here, but make sure
2022 * put_buf does it's thing
2024 bio->bi_end_io = end_sync_write;
2029 int d = r1_bio->read_disk;
2033 if (s > (PAGE_SIZE>>9))
2036 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2037 /* No rcu protection needed here devices
2038 * can only be removed when no resync is
2039 * active, and resync is currently active
2041 rdev = conf->mirrors[d].rdev;
2042 if (sync_page_io(rdev, sect, s<<9,
2044 REQ_OP_READ, false)) {
2050 if (d == conf->raid_disks * 2)
2052 } while (!success && d != r1_bio->read_disk);
2056 /* Cannot read from anywhere, this block is lost.
2057 * Record a bad block on each device. If that doesn't
2058 * work just disable and interrupt the recovery.
2059 * Don't fail devices as that won't really help.
2061 pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2062 mdname(mddev), bio->bi_bdev,
2063 (unsigned long long)r1_bio->sector);
2064 for (d = 0; d < conf->raid_disks * 2; d++) {
2065 rdev = conf->mirrors[d].rdev;
2066 if (!rdev || test_bit(Faulty, &rdev->flags))
2068 if (!rdev_set_badblocks(rdev, sect, s, 0))
2072 conf->recovery_disabled =
2073 mddev->recovery_disabled;
2074 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2075 md_done_sync(mddev, r1_bio->sectors, 0);
2087 /* write it back and re-read */
2088 while (d != r1_bio->read_disk) {
2090 d = conf->raid_disks * 2;
2092 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2094 rdev = conf->mirrors[d].rdev;
2095 if (r1_sync_page_io(rdev, sect, s,
2098 r1_bio->bios[d]->bi_end_io = NULL;
2099 rdev_dec_pending(rdev, mddev);
2103 while (d != r1_bio->read_disk) {
2105 d = conf->raid_disks * 2;
2107 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2109 rdev = conf->mirrors[d].rdev;
2110 if (r1_sync_page_io(rdev, sect, s,
2113 atomic_add(s, &rdev->corrected_errors);
2119 set_bit(R1BIO_Uptodate, &r1_bio->state);
2124 static void process_checks(struct r1bio *r1_bio)
2126 /* We have read all readable devices. If we haven't
2127 * got the block, then there is no hope left.
2128 * If we have, then we want to do a comparison
2129 * and skip the write if everything is the same.
2130 * If any blocks failed to read, then we need to
2131 * attempt an over-write
2133 struct mddev *mddev = r1_bio->mddev;
2134 struct r1conf *conf = mddev->private;
2139 /* Fix variable parts of all bios */
2140 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2141 for (i = 0; i < conf->raid_disks * 2; i++) {
2142 blk_status_t status;
2143 struct bio *b = r1_bio->bios[i];
2144 struct resync_pages *rp = get_resync_pages(b);
2145 if (b->bi_end_io != end_sync_read)
2147 /* fixup the bio for reuse, but preserve errno */
2148 status = b->bi_status;
2149 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2150 b->bi_status = status;
2151 b->bi_iter.bi_sector = r1_bio->sector +
2152 conf->mirrors[i].rdev->data_offset;
2153 b->bi_end_io = end_sync_read;
2154 rp->raid_bio = r1_bio;
2157 /* initialize bvec table again */
2158 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2160 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2161 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2162 !r1_bio->bios[primary]->bi_status) {
2163 r1_bio->bios[primary]->bi_end_io = NULL;
2164 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2167 r1_bio->read_disk = primary;
2168 for (i = 0; i < conf->raid_disks * 2; i++) {
2170 struct bio *pbio = r1_bio->bios[primary];
2171 struct bio *sbio = r1_bio->bios[i];
2172 blk_status_t status = sbio->bi_status;
2173 struct page **ppages = get_resync_pages(pbio)->pages;
2174 struct page **spages = get_resync_pages(sbio)->pages;
2176 int page_len[RESYNC_PAGES] = { 0 };
2177 struct bvec_iter_all iter_all;
2179 if (sbio->bi_end_io != end_sync_read)
2181 /* Now we can 'fixup' the error value */
2182 sbio->bi_status = 0;
2184 bio_for_each_segment_all(bi, sbio, iter_all)
2185 page_len[j++] = bi->bv_len;
2188 for (j = vcnt; j-- ; ) {
2189 if (memcmp(page_address(ppages[j]),
2190 page_address(spages[j]),
2197 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2198 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2200 /* No need to write to this device. */
2201 sbio->bi_end_io = NULL;
2202 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2206 bio_copy_data(sbio, pbio);
2210 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2212 struct r1conf *conf = mddev->private;
2214 int disks = conf->raid_disks * 2;
2217 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2218 /* ouch - failed to read all of that. */
2219 if (!fix_sync_read_error(r1_bio))
2222 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2223 process_checks(r1_bio);
2228 atomic_set(&r1_bio->remaining, 1);
2229 for (i = 0; i < disks ; i++) {
2230 wbio = r1_bio->bios[i];
2231 if (wbio->bi_end_io == NULL ||
2232 (wbio->bi_end_io == end_sync_read &&
2233 (i == r1_bio->read_disk ||
2234 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2236 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2237 abort_sync_write(mddev, r1_bio);
2241 wbio->bi_opf = REQ_OP_WRITE;
2242 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2243 wbio->bi_opf |= MD_FAILFAST;
2245 wbio->bi_end_io = end_sync_write;
2246 atomic_inc(&r1_bio->remaining);
2247 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2249 submit_bio_noacct(wbio);
2252 put_sync_write_buf(r1_bio, 1);
2256 * This is a kernel thread which:
2258 * 1. Retries failed read operations on working mirrors.
2259 * 2. Updates the raid superblock when problems encounter.
2260 * 3. Performs writes following reads for array synchronising.
2263 static void fix_read_error(struct r1conf *conf, int read_disk,
2264 sector_t sect, int sectors)
2266 struct mddev *mddev = conf->mddev;
2272 struct md_rdev *rdev;
2274 if (s > (PAGE_SIZE>>9))
2282 rdev = rcu_dereference(conf->mirrors[d].rdev);
2284 (test_bit(In_sync, &rdev->flags) ||
2285 (!test_bit(Faulty, &rdev->flags) &&
2286 rdev->recovery_offset >= sect + s)) &&
2287 is_badblock(rdev, sect, s,
2288 &first_bad, &bad_sectors) == 0) {
2289 atomic_inc(&rdev->nr_pending);
2291 if (sync_page_io(rdev, sect, s<<9,
2292 conf->tmppage, REQ_OP_READ, false))
2294 rdev_dec_pending(rdev, mddev);
2300 if (d == conf->raid_disks * 2)
2302 } while (!success && d != read_disk);
2305 /* Cannot read from anywhere - mark it bad */
2306 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2307 if (!rdev_set_badblocks(rdev, sect, s, 0))
2308 md_error(mddev, rdev);
2311 /* write it back and re-read */
2313 while (d != read_disk) {
2315 d = conf->raid_disks * 2;
2318 rdev = rcu_dereference(conf->mirrors[d].rdev);
2320 !test_bit(Faulty, &rdev->flags)) {
2321 atomic_inc(&rdev->nr_pending);
2323 r1_sync_page_io(rdev, sect, s,
2324 conf->tmppage, WRITE);
2325 rdev_dec_pending(rdev, mddev);
2330 while (d != read_disk) {
2332 d = conf->raid_disks * 2;
2335 rdev = rcu_dereference(conf->mirrors[d].rdev);
2337 !test_bit(Faulty, &rdev->flags)) {
2338 atomic_inc(&rdev->nr_pending);
2340 if (r1_sync_page_io(rdev, sect, s,
2341 conf->tmppage, READ)) {
2342 atomic_add(s, &rdev->corrected_errors);
2343 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %pg)\n",
2345 (unsigned long long)(sect +
2349 rdev_dec_pending(rdev, mddev);
2358 static int narrow_write_error(struct r1bio *r1_bio, int i)
2360 struct mddev *mddev = r1_bio->mddev;
2361 struct r1conf *conf = mddev->private;
2362 struct md_rdev *rdev = conf->mirrors[i].rdev;
2364 /* bio has the data to be written to device 'i' where
2365 * we just recently had a write error.
2366 * We repeatedly clone the bio and trim down to one block,
2367 * then try the write. Where the write fails we record
2369 * It is conceivable that the bio doesn't exactly align with
2370 * blocks. We must handle this somehow.
2372 * We currently own a reference on the rdev.
2378 int sect_to_write = r1_bio->sectors;
2381 if (rdev->badblocks.shift < 0)
2384 block_sectors = roundup(1 << rdev->badblocks.shift,
2385 bdev_logical_block_size(rdev->bdev) >> 9);
2386 sector = r1_bio->sector;
2387 sectors = ((sector + block_sectors)
2388 & ~(sector_t)(block_sectors - 1))
2391 while (sect_to_write) {
2393 if (sectors > sect_to_write)
2394 sectors = sect_to_write;
2395 /* Write at 'sector' for 'sectors'*/
2397 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2398 wbio = bio_alloc_clone(rdev->bdev,
2399 r1_bio->behind_master_bio,
2400 GFP_NOIO, &mddev->bio_set);
2402 wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2403 GFP_NOIO, &mddev->bio_set);
2406 wbio->bi_opf = REQ_OP_WRITE;
2407 wbio->bi_iter.bi_sector = r1_bio->sector;
2408 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2410 bio_trim(wbio, sector - r1_bio->sector, sectors);
2411 wbio->bi_iter.bi_sector += rdev->data_offset;
2413 if (submit_bio_wait(wbio) < 0)
2415 ok = rdev_set_badblocks(rdev, sector,
2420 sect_to_write -= sectors;
2422 sectors = block_sectors;
2427 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2430 int s = r1_bio->sectors;
2431 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2432 struct md_rdev *rdev = conf->mirrors[m].rdev;
2433 struct bio *bio = r1_bio->bios[m];
2434 if (bio->bi_end_io == NULL)
2436 if (!bio->bi_status &&
2437 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2438 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2440 if (bio->bi_status &&
2441 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2442 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2443 md_error(conf->mddev, rdev);
2447 md_done_sync(conf->mddev, s, 1);
2450 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2455 for (m = 0; m < conf->raid_disks * 2 ; m++)
2456 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2457 struct md_rdev *rdev = conf->mirrors[m].rdev;
2458 rdev_clear_badblocks(rdev,
2460 r1_bio->sectors, 0);
2461 rdev_dec_pending(rdev, conf->mddev);
2462 } else if (r1_bio->bios[m] != NULL) {
2463 /* This drive got a write error. We need to
2464 * narrow down and record precise write
2468 if (!narrow_write_error(r1_bio, m)) {
2469 md_error(conf->mddev,
2470 conf->mirrors[m].rdev);
2471 /* an I/O failed, we can't clear the bitmap */
2472 set_bit(R1BIO_Degraded, &r1_bio->state);
2474 rdev_dec_pending(conf->mirrors[m].rdev,
2478 spin_lock_irq(&conf->device_lock);
2479 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2480 idx = sector_to_idx(r1_bio->sector);
2481 atomic_inc(&conf->nr_queued[idx]);
2482 spin_unlock_irq(&conf->device_lock);
2484 * In case freeze_array() is waiting for condition
2485 * get_unqueued_pending() == extra to be true.
2487 wake_up(&conf->wait_barrier);
2488 md_wakeup_thread(conf->mddev->thread);
2490 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2491 close_write(r1_bio);
2492 raid_end_bio_io(r1_bio);
2496 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2498 struct mddev *mddev = conf->mddev;
2500 struct md_rdev *rdev;
2502 clear_bit(R1BIO_ReadError, &r1_bio->state);
2503 /* we got a read error. Maybe the drive is bad. Maybe just
2504 * the block and we can fix it.
2505 * We freeze all other IO, and try reading the block from
2506 * other devices. When we find one, we re-write
2507 * and check it that fixes the read error.
2508 * This is all done synchronously while the array is
2512 bio = r1_bio->bios[r1_bio->read_disk];
2514 r1_bio->bios[r1_bio->read_disk] = NULL;
2516 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2518 && !test_bit(FailFast, &rdev->flags)) {
2519 freeze_array(conf, 1);
2520 fix_read_error(conf, r1_bio->read_disk,
2521 r1_bio->sector, r1_bio->sectors);
2522 unfreeze_array(conf);
2523 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2524 md_error(mddev, rdev);
2526 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2529 rdev_dec_pending(rdev, conf->mddev);
2530 allow_barrier(conf, r1_bio->sector);
2531 bio = r1_bio->master_bio;
2533 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2535 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2538 static void raid1d(struct md_thread *thread)
2540 struct mddev *mddev = thread->mddev;
2541 struct r1bio *r1_bio;
2542 unsigned long flags;
2543 struct r1conf *conf = mddev->private;
2544 struct list_head *head = &conf->retry_list;
2545 struct blk_plug plug;
2548 md_check_recovery(mddev);
2550 if (!list_empty_careful(&conf->bio_end_io_list) &&
2551 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2553 spin_lock_irqsave(&conf->device_lock, flags);
2554 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2555 list_splice_init(&conf->bio_end_io_list, &tmp);
2556 spin_unlock_irqrestore(&conf->device_lock, flags);
2557 while (!list_empty(&tmp)) {
2558 r1_bio = list_first_entry(&tmp, struct r1bio,
2560 list_del(&r1_bio->retry_list);
2561 idx = sector_to_idx(r1_bio->sector);
2562 atomic_dec(&conf->nr_queued[idx]);
2563 if (mddev->degraded)
2564 set_bit(R1BIO_Degraded, &r1_bio->state);
2565 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2566 close_write(r1_bio);
2567 raid_end_bio_io(r1_bio);
2571 blk_start_plug(&plug);
2574 flush_pending_writes(conf);
2576 spin_lock_irqsave(&conf->device_lock, flags);
2577 if (list_empty(head)) {
2578 spin_unlock_irqrestore(&conf->device_lock, flags);
2581 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2582 list_del(head->prev);
2583 idx = sector_to_idx(r1_bio->sector);
2584 atomic_dec(&conf->nr_queued[idx]);
2585 spin_unlock_irqrestore(&conf->device_lock, flags);
2587 mddev = r1_bio->mddev;
2588 conf = mddev->private;
2589 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2590 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2591 test_bit(R1BIO_WriteError, &r1_bio->state))
2592 handle_sync_write_finished(conf, r1_bio);
2594 sync_request_write(mddev, r1_bio);
2595 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2596 test_bit(R1BIO_WriteError, &r1_bio->state))
2597 handle_write_finished(conf, r1_bio);
2598 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2599 handle_read_error(conf, r1_bio);
2604 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2605 md_check_recovery(mddev);
2607 blk_finish_plug(&plug);
2610 static int init_resync(struct r1conf *conf)
2614 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2615 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2617 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2618 r1buf_pool_free, conf->poolinfo);
2621 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2623 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2624 struct resync_pages *rps;
2628 for (i = conf->poolinfo->raid_disks; i--; ) {
2629 bio = r1bio->bios[i];
2630 rps = bio->bi_private;
2631 bio_reset(bio, NULL, 0);
2632 bio->bi_private = rps;
2634 r1bio->master_bio = NULL;
2639 * perform a "sync" on one "block"
2641 * We need to make sure that no normal I/O request - particularly write
2642 * requests - conflict with active sync requests.
2644 * This is achieved by tracking pending requests and a 'barrier' concept
2645 * that can be installed to exclude normal IO requests.
2648 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2651 struct r1conf *conf = mddev->private;
2652 struct r1bio *r1_bio;
2654 sector_t max_sector, nr_sectors;
2658 int write_targets = 0, read_targets = 0;
2659 sector_t sync_blocks;
2660 int still_degraded = 0;
2661 int good_sectors = RESYNC_SECTORS;
2662 int min_bad = 0; /* number of sectors that are bad in all devices */
2663 int idx = sector_to_idx(sector_nr);
2666 if (!mempool_initialized(&conf->r1buf_pool))
2667 if (init_resync(conf))
2670 max_sector = mddev->dev_sectors;
2671 if (sector_nr >= max_sector) {
2672 /* If we aborted, we need to abort the
2673 * sync on the 'current' bitmap chunk (there will
2674 * only be one in raid1 resync.
2675 * We can find the current addess in mddev->curr_resync
2677 if (mddev->curr_resync < max_sector) /* aborted */
2678 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2680 else /* completed sync */
2683 md_bitmap_close_sync(mddev->bitmap);
2686 if (mddev_is_clustered(mddev)) {
2687 conf->cluster_sync_low = 0;
2688 conf->cluster_sync_high = 0;
2693 if (mddev->bitmap == NULL &&
2694 mddev->recovery_cp == MaxSector &&
2695 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2696 conf->fullsync == 0) {
2698 return max_sector - sector_nr;
2700 /* before building a request, check if we can skip these blocks..
2701 * This call the bitmap_start_sync doesn't actually record anything
2703 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2704 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2705 /* We can skip this block, and probably several more */
2711 * If there is non-resync activity waiting for a turn, then let it
2712 * though before starting on this new sync request.
2714 if (atomic_read(&conf->nr_waiting[idx]))
2715 schedule_timeout_uninterruptible(1);
2717 /* we are incrementing sector_nr below. To be safe, we check against
2718 * sector_nr + two times RESYNC_SECTORS
2721 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2722 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2725 if (raise_barrier(conf, sector_nr))
2728 r1_bio = raid1_alloc_init_r1buf(conf);
2732 * If we get a correctably read error during resync or recovery,
2733 * we might want to read from a different device. So we
2734 * flag all drives that could conceivably be read from for READ,
2735 * and any others (which will be non-In_sync devices) for WRITE.
2736 * If a read fails, we try reading from something else for which READ
2740 r1_bio->mddev = mddev;
2741 r1_bio->sector = sector_nr;
2743 set_bit(R1BIO_IsSync, &r1_bio->state);
2744 /* make sure good_sectors won't go across barrier unit boundary */
2745 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2747 for (i = 0; i < conf->raid_disks * 2; i++) {
2748 struct md_rdev *rdev;
2749 bio = r1_bio->bios[i];
2751 rdev = rcu_dereference(conf->mirrors[i].rdev);
2753 test_bit(Faulty, &rdev->flags)) {
2754 if (i < conf->raid_disks)
2756 } else if (!test_bit(In_sync, &rdev->flags)) {
2757 bio->bi_opf = REQ_OP_WRITE;
2758 bio->bi_end_io = end_sync_write;
2761 /* may need to read from here */
2762 sector_t first_bad = MaxSector;
2765 if (is_badblock(rdev, sector_nr, good_sectors,
2766 &first_bad, &bad_sectors)) {
2767 if (first_bad > sector_nr)
2768 good_sectors = first_bad - sector_nr;
2770 bad_sectors -= (sector_nr - first_bad);
2772 min_bad > bad_sectors)
2773 min_bad = bad_sectors;
2776 if (sector_nr < first_bad) {
2777 if (test_bit(WriteMostly, &rdev->flags)) {
2784 bio->bi_opf = REQ_OP_READ;
2785 bio->bi_end_io = end_sync_read;
2787 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2788 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2789 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2791 * The device is suitable for reading (InSync),
2792 * but has bad block(s) here. Let's try to correct them,
2793 * if we are doing resync or repair. Otherwise, leave
2794 * this device alone for this sync request.
2796 bio->bi_opf = REQ_OP_WRITE;
2797 bio->bi_end_io = end_sync_write;
2801 if (rdev && bio->bi_end_io) {
2802 atomic_inc(&rdev->nr_pending);
2803 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2804 bio_set_dev(bio, rdev->bdev);
2805 if (test_bit(FailFast, &rdev->flags))
2806 bio->bi_opf |= MD_FAILFAST;
2812 r1_bio->read_disk = disk;
2814 if (read_targets == 0 && min_bad > 0) {
2815 /* These sectors are bad on all InSync devices, so we
2816 * need to mark them bad on all write targets
2819 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2820 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2821 struct md_rdev *rdev = conf->mirrors[i].rdev;
2822 ok = rdev_set_badblocks(rdev, sector_nr,
2826 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2831 /* Cannot record the badblocks, so need to
2833 * If there are multiple read targets, could just
2834 * fail the really bad ones ???
2836 conf->recovery_disabled = mddev->recovery_disabled;
2837 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2843 if (min_bad > 0 && min_bad < good_sectors) {
2844 /* only resync enough to reach the next bad->good
2846 good_sectors = min_bad;
2849 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2850 /* extra read targets are also write targets */
2851 write_targets += read_targets-1;
2853 if (write_targets == 0 || read_targets == 0) {
2854 /* There is nowhere to write, so all non-sync
2855 * drives must be failed - so we are finished
2859 max_sector = sector_nr + min_bad;
2860 rv = max_sector - sector_nr;
2866 if (max_sector > mddev->resync_max)
2867 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2868 if (max_sector > sector_nr + good_sectors)
2869 max_sector = sector_nr + good_sectors;
2874 int len = PAGE_SIZE;
2875 if (sector_nr + (len>>9) > max_sector)
2876 len = (max_sector - sector_nr) << 9;
2879 if (sync_blocks == 0) {
2880 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2881 &sync_blocks, still_degraded) &&
2883 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2885 if ((len >> 9) > sync_blocks)
2886 len = sync_blocks<<9;
2889 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2890 struct resync_pages *rp;
2892 bio = r1_bio->bios[i];
2893 rp = get_resync_pages(bio);
2894 if (bio->bi_end_io) {
2895 page = resync_fetch_page(rp, page_idx);
2898 * won't fail because the vec table is big
2899 * enough to hold all these pages
2901 __bio_add_page(bio, page, len, 0);
2904 nr_sectors += len>>9;
2905 sector_nr += len>>9;
2906 sync_blocks -= (len>>9);
2907 } while (++page_idx < RESYNC_PAGES);
2909 r1_bio->sectors = nr_sectors;
2911 if (mddev_is_clustered(mddev) &&
2912 conf->cluster_sync_high < sector_nr + nr_sectors) {
2913 conf->cluster_sync_low = mddev->curr_resync_completed;
2914 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2915 /* Send resync message */
2916 md_cluster_ops->resync_info_update(mddev,
2917 conf->cluster_sync_low,
2918 conf->cluster_sync_high);
2921 /* For a user-requested sync, we read all readable devices and do a
2924 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2925 atomic_set(&r1_bio->remaining, read_targets);
2926 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2927 bio = r1_bio->bios[i];
2928 if (bio->bi_end_io == end_sync_read) {
2930 md_sync_acct_bio(bio, nr_sectors);
2931 if (read_targets == 1)
2932 bio->bi_opf &= ~MD_FAILFAST;
2933 submit_bio_noacct(bio);
2937 atomic_set(&r1_bio->remaining, 1);
2938 bio = r1_bio->bios[r1_bio->read_disk];
2939 md_sync_acct_bio(bio, nr_sectors);
2940 if (read_targets == 1)
2941 bio->bi_opf &= ~MD_FAILFAST;
2942 submit_bio_noacct(bio);
2947 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2952 return mddev->dev_sectors;
2955 static struct r1conf *setup_conf(struct mddev *mddev)
2957 struct r1conf *conf;
2959 struct raid1_info *disk;
2960 struct md_rdev *rdev;
2963 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2967 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2968 sizeof(atomic_t), GFP_KERNEL);
2969 if (!conf->nr_pending)
2972 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2973 sizeof(atomic_t), GFP_KERNEL);
2974 if (!conf->nr_waiting)
2977 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2978 sizeof(atomic_t), GFP_KERNEL);
2979 if (!conf->nr_queued)
2982 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2983 sizeof(atomic_t), GFP_KERNEL);
2987 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2988 mddev->raid_disks, 2),
2993 conf->tmppage = alloc_page(GFP_KERNEL);
2997 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2998 if (!conf->poolinfo)
3000 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3001 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3002 rbio_pool_free, conf->poolinfo);
3006 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3010 conf->poolinfo->mddev = mddev;
3013 spin_lock_init(&conf->device_lock);
3014 rdev_for_each(rdev, mddev) {
3015 int disk_idx = rdev->raid_disk;
3016 if (disk_idx >= mddev->raid_disks
3019 if (test_bit(Replacement, &rdev->flags))
3020 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3022 disk = conf->mirrors + disk_idx;
3027 disk->head_position = 0;
3028 disk->seq_start = MaxSector;
3030 conf->raid_disks = mddev->raid_disks;
3031 conf->mddev = mddev;
3032 INIT_LIST_HEAD(&conf->retry_list);
3033 INIT_LIST_HEAD(&conf->bio_end_io_list);
3035 spin_lock_init(&conf->resync_lock);
3036 init_waitqueue_head(&conf->wait_barrier);
3038 bio_list_init(&conf->pending_bio_list);
3039 conf->recovery_disabled = mddev->recovery_disabled - 1;
3042 for (i = 0; i < conf->raid_disks * 2; i++) {
3044 disk = conf->mirrors + i;
3046 if (i < conf->raid_disks &&
3047 disk[conf->raid_disks].rdev) {
3048 /* This slot has a replacement. */
3050 /* No original, just make the replacement
3051 * a recovering spare
3054 disk[conf->raid_disks].rdev;
3055 disk[conf->raid_disks].rdev = NULL;
3056 } else if (!test_bit(In_sync, &disk->rdev->flags))
3057 /* Original is not in_sync - bad */
3062 !test_bit(In_sync, &disk->rdev->flags)) {
3063 disk->head_position = 0;
3065 (disk->rdev->saved_raid_disk < 0))
3071 rcu_assign_pointer(conf->thread,
3072 md_register_thread(raid1d, mddev, "raid1"));
3080 mempool_exit(&conf->r1bio_pool);
3081 kfree(conf->mirrors);
3082 safe_put_page(conf->tmppage);
3083 kfree(conf->poolinfo);
3084 kfree(conf->nr_pending);
3085 kfree(conf->nr_waiting);
3086 kfree(conf->nr_queued);
3087 kfree(conf->barrier);
3088 bioset_exit(&conf->bio_split);
3091 return ERR_PTR(err);
3094 static void raid1_free(struct mddev *mddev, void *priv);
3095 static int raid1_run(struct mddev *mddev)
3097 struct r1conf *conf;
3099 struct md_rdev *rdev;
3102 if (mddev->level != 1) {
3103 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3104 mdname(mddev), mddev->level);
3107 if (mddev->reshape_position != MaxSector) {
3108 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3112 if (mddev_init_writes_pending(mddev) < 0)
3115 * copy the already verified devices into our private RAID1
3116 * bookkeeping area. [whatever we allocate in run(),
3117 * should be freed in raid1_free()]
3119 if (mddev->private == NULL)
3120 conf = setup_conf(mddev);
3122 conf = mddev->private;
3125 return PTR_ERR(conf);
3128 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3130 rdev_for_each(rdev, mddev) {
3131 if (!mddev->gendisk)
3133 disk_stack_limits(mddev->gendisk, rdev->bdev,
3134 rdev->data_offset << 9);
3137 mddev->degraded = 0;
3138 for (i = 0; i < conf->raid_disks; i++)
3139 if (conf->mirrors[i].rdev == NULL ||
3140 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3141 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3144 * RAID1 needs at least one disk in active
3146 if (conf->raid_disks - mddev->degraded < 1) {
3147 md_unregister_thread(&conf->thread);
3152 if (conf->raid_disks - mddev->degraded == 1)
3153 mddev->recovery_cp = MaxSector;
3155 if (mddev->recovery_cp != MaxSector)
3156 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3158 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3159 mdname(mddev), mddev->raid_disks - mddev->degraded,
3163 * Ok, everything is just fine now
3165 rcu_assign_pointer(mddev->thread, conf->thread);
3166 rcu_assign_pointer(conf->thread, NULL);
3167 mddev->private = conf;
3168 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3170 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3172 ret = md_integrity_register(mddev);
3174 md_unregister_thread(&mddev->thread);
3180 raid1_free(mddev, conf);
3184 static void raid1_free(struct mddev *mddev, void *priv)
3186 struct r1conf *conf = priv;
3188 mempool_exit(&conf->r1bio_pool);
3189 kfree(conf->mirrors);
3190 safe_put_page(conf->tmppage);
3191 kfree(conf->poolinfo);
3192 kfree(conf->nr_pending);
3193 kfree(conf->nr_waiting);
3194 kfree(conf->nr_queued);
3195 kfree(conf->barrier);
3196 bioset_exit(&conf->bio_split);
3200 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3202 /* no resync is happening, and there is enough space
3203 * on all devices, so we can resize.
3204 * We need to make sure resync covers any new space.
3205 * If the array is shrinking we should possibly wait until
3206 * any io in the removed space completes, but it hardly seems
3209 sector_t newsize = raid1_size(mddev, sectors, 0);
3210 if (mddev->external_size &&
3211 mddev->array_sectors > newsize)
3213 if (mddev->bitmap) {
3214 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3218 md_set_array_sectors(mddev, newsize);
3219 if (sectors > mddev->dev_sectors &&
3220 mddev->recovery_cp > mddev->dev_sectors) {
3221 mddev->recovery_cp = mddev->dev_sectors;
3222 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3224 mddev->dev_sectors = sectors;
3225 mddev->resync_max_sectors = sectors;
3229 static int raid1_reshape(struct mddev *mddev)
3232 * 1/ resize the r1bio_pool
3233 * 2/ resize conf->mirrors
3235 * We allocate a new r1bio_pool if we can.
3236 * Then raise a device barrier and wait until all IO stops.
3237 * Then resize conf->mirrors and swap in the new r1bio pool.
3239 * At the same time, we "pack" the devices so that all the missing
3240 * devices have the higher raid_disk numbers.
3242 mempool_t newpool, oldpool;
3243 struct pool_info *newpoolinfo;
3244 struct raid1_info *newmirrors;
3245 struct r1conf *conf = mddev->private;
3246 int cnt, raid_disks;
3247 unsigned long flags;
3251 memset(&newpool, 0, sizeof(newpool));
3252 memset(&oldpool, 0, sizeof(oldpool));
3254 /* Cannot change chunk_size, layout, or level */
3255 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3256 mddev->layout != mddev->new_layout ||
3257 mddev->level != mddev->new_level) {
3258 mddev->new_chunk_sectors = mddev->chunk_sectors;
3259 mddev->new_layout = mddev->layout;
3260 mddev->new_level = mddev->level;
3264 if (!mddev_is_clustered(mddev))
3265 md_allow_write(mddev);
3267 raid_disks = mddev->raid_disks + mddev->delta_disks;
3269 if (raid_disks < conf->raid_disks) {
3271 for (d= 0; d < conf->raid_disks; d++)
3272 if (conf->mirrors[d].rdev)
3274 if (cnt > raid_disks)
3278 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3281 newpoolinfo->mddev = mddev;
3282 newpoolinfo->raid_disks = raid_disks * 2;
3284 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3285 rbio_pool_free, newpoolinfo);
3290 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3295 mempool_exit(&newpool);
3299 freeze_array(conf, 0);
3301 /* ok, everything is stopped */
3302 oldpool = conf->r1bio_pool;
3303 conf->r1bio_pool = newpool;
3305 for (d = d2 = 0; d < conf->raid_disks; d++) {
3306 struct md_rdev *rdev = conf->mirrors[d].rdev;
3307 if (rdev && rdev->raid_disk != d2) {
3308 sysfs_unlink_rdev(mddev, rdev);
3309 rdev->raid_disk = d2;
3310 sysfs_unlink_rdev(mddev, rdev);
3311 if (sysfs_link_rdev(mddev, rdev))
3312 pr_warn("md/raid1:%s: cannot register rd%d\n",
3313 mdname(mddev), rdev->raid_disk);
3316 newmirrors[d2++].rdev = rdev;
3318 kfree(conf->mirrors);
3319 conf->mirrors = newmirrors;
3320 kfree(conf->poolinfo);
3321 conf->poolinfo = newpoolinfo;
3323 spin_lock_irqsave(&conf->device_lock, flags);
3324 mddev->degraded += (raid_disks - conf->raid_disks);
3325 spin_unlock_irqrestore(&conf->device_lock, flags);
3326 conf->raid_disks = mddev->raid_disks = raid_disks;
3327 mddev->delta_disks = 0;
3329 unfreeze_array(conf);
3331 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3332 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3333 md_wakeup_thread(mddev->thread);
3335 mempool_exit(&oldpool);
3339 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3341 struct r1conf *conf = mddev->private;
3344 freeze_array(conf, 0);
3346 unfreeze_array(conf);
3349 static void *raid1_takeover(struct mddev *mddev)
3351 /* raid1 can take over:
3352 * raid5 with 2 devices, any layout or chunk size
3354 if (mddev->level == 5 && mddev->raid_disks == 2) {
3355 struct r1conf *conf;
3356 mddev->new_level = 1;
3357 mddev->new_layout = 0;
3358 mddev->new_chunk_sectors = 0;
3359 conf = setup_conf(mddev);
3360 if (!IS_ERR(conf)) {
3361 /* Array must appear to be quiesced */
3362 conf->array_frozen = 1;
3363 mddev_clear_unsupported_flags(mddev,
3364 UNSUPPORTED_MDDEV_FLAGS);
3368 return ERR_PTR(-EINVAL);
3371 static struct md_personality raid1_personality =
3375 .owner = THIS_MODULE,
3376 .make_request = raid1_make_request,
3379 .status = raid1_status,
3380 .error_handler = raid1_error,
3381 .hot_add_disk = raid1_add_disk,
3382 .hot_remove_disk= raid1_remove_disk,
3383 .spare_active = raid1_spare_active,
3384 .sync_request = raid1_sync_request,
3385 .resize = raid1_resize,
3387 .check_reshape = raid1_reshape,
3388 .quiesce = raid1_quiesce,
3389 .takeover = raid1_takeover,
3392 static int __init raid_init(void)
3394 return register_md_personality(&raid1_personality);
3397 static void raid_exit(void)
3399 unregister_md_personality(&raid1_personality);
3402 module_init(raid_init);
3403 module_exit(raid_exit);
3404 MODULE_LICENSE("GPL");
3405 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3406 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3407 MODULE_ALIAS("md-raid1");
3408 MODULE_ALIAS("md-level-1");