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 char b[BDEVNAME_SIZE];
406 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
408 bdevname(rdev->bdev, b),
409 (unsigned long long)r1_bio->sector);
410 set_bit(R1BIO_ReadError, &r1_bio->state);
411 reschedule_retry(r1_bio);
412 /* don't drop the reference on read_disk yet */
416 static void close_write(struct r1bio *r1_bio)
418 /* it really is the end of this request */
419 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
420 bio_free_pages(r1_bio->behind_master_bio);
421 bio_put(r1_bio->behind_master_bio);
422 r1_bio->behind_master_bio = NULL;
424 /* clear the bitmap if all writes complete successfully */
425 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
427 !test_bit(R1BIO_Degraded, &r1_bio->state),
428 test_bit(R1BIO_BehindIO, &r1_bio->state));
429 md_write_end(r1_bio->mddev);
432 static void r1_bio_write_done(struct r1bio *r1_bio)
434 if (!atomic_dec_and_test(&r1_bio->remaining))
437 if (test_bit(R1BIO_WriteError, &r1_bio->state))
438 reschedule_retry(r1_bio);
441 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
442 reschedule_retry(r1_bio);
444 raid_end_bio_io(r1_bio);
448 static void raid1_end_write_request(struct bio *bio)
450 struct r1bio *r1_bio = bio->bi_private;
451 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
452 struct r1conf *conf = r1_bio->mddev->private;
453 struct bio *to_put = NULL;
454 int mirror = find_bio_disk(r1_bio, bio);
455 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
457 sector_t lo = r1_bio->sector;
458 sector_t hi = r1_bio->sector + r1_bio->sectors;
460 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
463 * 'one mirror IO has finished' event handler:
465 if (bio->bi_status && !discard_error) {
466 set_bit(WriteErrorSeen, &rdev->flags);
467 if (!test_and_set_bit(WantReplacement, &rdev->flags))
468 set_bit(MD_RECOVERY_NEEDED, &
469 conf->mddev->recovery);
471 if (test_bit(FailFast, &rdev->flags) &&
472 (bio->bi_opf & MD_FAILFAST) &&
473 /* We never try FailFast to WriteMostly devices */
474 !test_bit(WriteMostly, &rdev->flags)) {
475 md_error(r1_bio->mddev, rdev);
479 * When the device is faulty, it is not necessary to
480 * handle write error.
482 if (!test_bit(Faulty, &rdev->flags))
483 set_bit(R1BIO_WriteError, &r1_bio->state);
485 /* Fail the request */
486 set_bit(R1BIO_Degraded, &r1_bio->state);
487 /* Finished with this branch */
488 r1_bio->bios[mirror] = NULL;
493 * Set R1BIO_Uptodate in our master bio, so that we
494 * will return a good error code for to the higher
495 * levels even if IO on some other mirrored buffer
498 * The 'master' represents the composite IO operation
499 * to user-side. So if something waits for IO, then it
500 * will wait for the 'master' bio.
505 r1_bio->bios[mirror] = NULL;
508 * Do not set R1BIO_Uptodate if the current device is
509 * rebuilding or Faulty. This is because we cannot use
510 * such device for properly reading the data back (we could
511 * potentially use it, if the current write would have felt
512 * before rdev->recovery_offset, but for simplicity we don't
515 if (test_bit(In_sync, &rdev->flags) &&
516 !test_bit(Faulty, &rdev->flags))
517 set_bit(R1BIO_Uptodate, &r1_bio->state);
519 /* Maybe we can clear some bad blocks. */
520 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
521 &first_bad, &bad_sectors) && !discard_error) {
522 r1_bio->bios[mirror] = IO_MADE_GOOD;
523 set_bit(R1BIO_MadeGood, &r1_bio->state);
528 if (test_bit(CollisionCheck, &rdev->flags))
529 remove_serial(rdev, lo, hi);
530 if (test_bit(WriteMostly, &rdev->flags))
531 atomic_dec(&r1_bio->behind_remaining);
534 * In behind mode, we ACK the master bio once the I/O
535 * has safely reached all non-writemostly
536 * disks. Setting the Returned bit ensures that this
537 * gets done only once -- we don't ever want to return
538 * -EIO here, instead we'll wait
540 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
541 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
542 /* Maybe we can return now */
543 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
544 struct bio *mbio = r1_bio->master_bio;
545 pr_debug("raid1: behind end write sectors"
547 (unsigned long long) mbio->bi_iter.bi_sector,
548 (unsigned long long) bio_end_sector(mbio) - 1);
549 call_bio_endio(r1_bio);
552 } else if (rdev->mddev->serialize_policy)
553 remove_serial(rdev, lo, hi);
554 if (r1_bio->bios[mirror] == NULL)
555 rdev_dec_pending(rdev, conf->mddev);
558 * Let's see if all mirrored write operations have finished
561 r1_bio_write_done(r1_bio);
567 static sector_t align_to_barrier_unit_end(sector_t start_sector,
572 WARN_ON(sectors == 0);
574 * len is the number of sectors from start_sector to end of the
575 * barrier unit which start_sector belongs to.
577 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
587 * This routine returns the disk from which the requested read should
588 * be done. There is a per-array 'next expected sequential IO' sector
589 * number - if this matches on the next IO then we use the last disk.
590 * There is also a per-disk 'last know head position' sector that is
591 * maintained from IRQ contexts, both the normal and the resync IO
592 * completion handlers update this position correctly. If there is no
593 * perfect sequential match then we pick the disk whose head is closest.
595 * If there are 2 mirrors in the same 2 devices, performance degrades
596 * because position is mirror, not device based.
598 * The rdev for the device selected will have nr_pending incremented.
600 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
602 const sector_t this_sector = r1_bio->sector;
604 int best_good_sectors;
605 int best_disk, best_dist_disk, best_pending_disk;
609 unsigned int min_pending;
610 struct md_rdev *rdev;
612 int choose_next_idle;
616 * Check if we can balance. We can balance on the whole
617 * device if no resync is going on, or below the resync window.
618 * We take the first readable disk when above the resync window.
621 sectors = r1_bio->sectors;
624 best_dist = MaxSector;
625 best_pending_disk = -1;
626 min_pending = UINT_MAX;
627 best_good_sectors = 0;
629 choose_next_idle = 0;
630 clear_bit(R1BIO_FailFast, &r1_bio->state);
632 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
633 (mddev_is_clustered(conf->mddev) &&
634 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
635 this_sector + sectors)))
640 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
644 unsigned int pending;
647 rdev = rcu_dereference(conf->mirrors[disk].rdev);
648 if (r1_bio->bios[disk] == IO_BLOCKED
650 || test_bit(Faulty, &rdev->flags))
652 if (!test_bit(In_sync, &rdev->flags) &&
653 rdev->recovery_offset < this_sector + sectors)
655 if (test_bit(WriteMostly, &rdev->flags)) {
656 /* Don't balance among write-mostly, just
657 * use the first as a last resort */
658 if (best_dist_disk < 0) {
659 if (is_badblock(rdev, this_sector, sectors,
660 &first_bad, &bad_sectors)) {
661 if (first_bad <= this_sector)
662 /* Cannot use this */
664 best_good_sectors = first_bad - this_sector;
666 best_good_sectors = sectors;
667 best_dist_disk = disk;
668 best_pending_disk = disk;
672 /* This is a reasonable device to use. It might
675 if (is_badblock(rdev, this_sector, sectors,
676 &first_bad, &bad_sectors)) {
677 if (best_dist < MaxSector)
678 /* already have a better device */
680 if (first_bad <= this_sector) {
681 /* cannot read here. If this is the 'primary'
682 * device, then we must not read beyond
683 * bad_sectors from another device..
685 bad_sectors -= (this_sector - first_bad);
686 if (choose_first && sectors > bad_sectors)
687 sectors = bad_sectors;
688 if (best_good_sectors > sectors)
689 best_good_sectors = sectors;
692 sector_t good_sectors = first_bad - this_sector;
693 if (good_sectors > best_good_sectors) {
694 best_good_sectors = good_sectors;
702 if ((sectors > best_good_sectors) && (best_disk >= 0))
704 best_good_sectors = sectors;
708 /* At least two disks to choose from so failfast is OK */
709 set_bit(R1BIO_FailFast, &r1_bio->state);
711 nonrot = bdev_nonrot(rdev->bdev);
712 has_nonrot_disk |= nonrot;
713 pending = atomic_read(&rdev->nr_pending);
714 dist = abs(this_sector - conf->mirrors[disk].head_position);
719 /* Don't change to another disk for sequential reads */
720 if (conf->mirrors[disk].next_seq_sect == this_sector
722 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
723 struct raid1_info *mirror = &conf->mirrors[disk];
727 * If buffered sequential IO size exceeds optimal
728 * iosize, check if there is idle disk. If yes, choose
729 * the idle disk. read_balance could already choose an
730 * idle disk before noticing it's a sequential IO in
731 * this disk. This doesn't matter because this disk
732 * will idle, next time it will be utilized after the
733 * first disk has IO size exceeds optimal iosize. In
734 * this way, iosize of the first disk will be optimal
735 * iosize at least. iosize of the second disk might be
736 * small, but not a big deal since when the second disk
737 * starts IO, the first disk is likely still busy.
739 if (nonrot && opt_iosize > 0 &&
740 mirror->seq_start != MaxSector &&
741 mirror->next_seq_sect > opt_iosize &&
742 mirror->next_seq_sect - opt_iosize >=
744 choose_next_idle = 1;
750 if (choose_next_idle)
753 if (min_pending > pending) {
754 min_pending = pending;
755 best_pending_disk = disk;
758 if (dist < best_dist) {
760 best_dist_disk = disk;
765 * If all disks are rotational, choose the closest disk. If any disk is
766 * non-rotational, choose the disk with less pending request even the
767 * disk is rotational, which might/might not be optimal for raids with
768 * mixed ratation/non-rotational disks depending on workload.
770 if (best_disk == -1) {
771 if (has_nonrot_disk || min_pending == 0)
772 best_disk = best_pending_disk;
774 best_disk = best_dist_disk;
777 if (best_disk >= 0) {
778 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
781 atomic_inc(&rdev->nr_pending);
782 sectors = best_good_sectors;
784 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
785 conf->mirrors[best_disk].seq_start = this_sector;
787 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
790 *max_sectors = sectors;
795 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
797 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
798 md_bitmap_unplug(conf->mddev->bitmap);
799 wake_up(&conf->wait_barrier);
801 while (bio) { /* submit pending writes */
802 struct bio *next = bio->bi_next;
803 struct md_rdev *rdev = (void *)bio->bi_bdev;
805 bio_set_dev(bio, rdev->bdev);
806 if (test_bit(Faulty, &rdev->flags)) {
808 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
809 !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
813 submit_bio_noacct(bio);
819 static void flush_pending_writes(struct r1conf *conf)
821 /* Any writes that have been queued but are awaiting
822 * bitmap updates get flushed here.
824 spin_lock_irq(&conf->device_lock);
826 if (conf->pending_bio_list.head) {
827 struct blk_plug plug;
830 bio = bio_list_get(&conf->pending_bio_list);
831 spin_unlock_irq(&conf->device_lock);
834 * As this is called in a wait_event() loop (see freeze_array),
835 * current->state might be TASK_UNINTERRUPTIBLE which will
836 * cause a warning when we prepare to wait again. As it is
837 * rare that this path is taken, it is perfectly safe to force
838 * us to go around the wait_event() loop again, so the warning
839 * is a false-positive. Silence the warning by resetting
842 __set_current_state(TASK_RUNNING);
843 blk_start_plug(&plug);
844 flush_bio_list(conf, bio);
845 blk_finish_plug(&plug);
847 spin_unlock_irq(&conf->device_lock);
851 * Sometimes we need to suspend IO while we do something else,
852 * either some resync/recovery, or reconfigure the array.
853 * To do this we raise a 'barrier'.
854 * The 'barrier' is a counter that can be raised multiple times
855 * to count how many activities are happening which preclude
857 * We can only raise the barrier if there is no pending IO.
858 * i.e. if nr_pending == 0.
859 * We choose only to raise the barrier if no-one is waiting for the
860 * barrier to go down. This means that as soon as an IO request
861 * is ready, no other operations which require a barrier will start
862 * until the IO request has had a chance.
864 * So: regular IO calls 'wait_barrier'. When that returns there
865 * is no backgroup IO happening, It must arrange to call
866 * allow_barrier when it has finished its IO.
867 * backgroup IO calls must call raise_barrier. Once that returns
868 * there is no normal IO happeing. It must arrange to call
869 * lower_barrier when the particular background IO completes.
871 * If resync/recovery is interrupted, returns -EINTR;
872 * Otherwise, returns 0.
874 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
876 int idx = sector_to_idx(sector_nr);
878 spin_lock_irq(&conf->resync_lock);
880 /* Wait until no block IO is waiting */
881 wait_event_lock_irq(conf->wait_barrier,
882 !atomic_read(&conf->nr_waiting[idx]),
885 /* block any new IO from starting */
886 atomic_inc(&conf->barrier[idx]);
888 * In raise_barrier() we firstly increase conf->barrier[idx] then
889 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
890 * increase conf->nr_pending[idx] then check conf->barrier[idx].
891 * A memory barrier here to make sure conf->nr_pending[idx] won't
892 * be fetched before conf->barrier[idx] is increased. Otherwise
893 * there will be a race between raise_barrier() and _wait_barrier().
895 smp_mb__after_atomic();
897 /* For these conditions we must wait:
898 * A: while the array is in frozen state
899 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
900 * existing in corresponding I/O barrier bucket.
901 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
902 * max resync count which allowed on current I/O barrier bucket.
904 wait_event_lock_irq(conf->wait_barrier,
905 (!conf->array_frozen &&
906 !atomic_read(&conf->nr_pending[idx]) &&
907 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
908 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
911 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
912 atomic_dec(&conf->barrier[idx]);
913 spin_unlock_irq(&conf->resync_lock);
914 wake_up(&conf->wait_barrier);
918 atomic_inc(&conf->nr_sync_pending);
919 spin_unlock_irq(&conf->resync_lock);
924 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
926 int idx = sector_to_idx(sector_nr);
928 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
930 atomic_dec(&conf->barrier[idx]);
931 atomic_dec(&conf->nr_sync_pending);
932 wake_up(&conf->wait_barrier);
935 static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
940 * We need to increase conf->nr_pending[idx] very early here,
941 * then raise_barrier() can be blocked when it waits for
942 * conf->nr_pending[idx] to be 0. Then we can avoid holding
943 * conf->resync_lock when there is no barrier raised in same
944 * barrier unit bucket. Also if the array is frozen, I/O
945 * should be blocked until array is unfrozen.
947 atomic_inc(&conf->nr_pending[idx]);
949 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
950 * check conf->barrier[idx]. In raise_barrier() we firstly increase
951 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
952 * barrier is necessary here to make sure conf->barrier[idx] won't be
953 * fetched before conf->nr_pending[idx] is increased. Otherwise there
954 * will be a race between _wait_barrier() and raise_barrier().
956 smp_mb__after_atomic();
959 * Don't worry about checking two atomic_t variables at same time
960 * here. If during we check conf->barrier[idx], the array is
961 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
962 * 0, it is safe to return and make the I/O continue. Because the
963 * array is frozen, all I/O returned here will eventually complete
964 * or be queued, no race will happen. See code comment in
967 if (!READ_ONCE(conf->array_frozen) &&
968 !atomic_read(&conf->barrier[idx]))
972 * After holding conf->resync_lock, conf->nr_pending[idx]
973 * should be decreased before waiting for barrier to drop.
974 * Otherwise, we may encounter a race condition because
975 * raise_barrer() might be waiting for conf->nr_pending[idx]
976 * to be 0 at same time.
978 spin_lock_irq(&conf->resync_lock);
979 atomic_inc(&conf->nr_waiting[idx]);
980 atomic_dec(&conf->nr_pending[idx]);
982 * In case freeze_array() is waiting for
983 * get_unqueued_pending() == extra
985 wake_up(&conf->wait_barrier);
986 /* Wait for the barrier in same barrier unit bucket to drop. */
988 /* Return false when nowait flag is set */
992 wait_event_lock_irq(conf->wait_barrier,
993 !conf->array_frozen &&
994 !atomic_read(&conf->barrier[idx]),
996 atomic_inc(&conf->nr_pending[idx]);
999 atomic_dec(&conf->nr_waiting[idx]);
1000 spin_unlock_irq(&conf->resync_lock);
1004 static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1006 int idx = sector_to_idx(sector_nr);
1010 * Very similar to _wait_barrier(). The difference is, for read
1011 * I/O we don't need wait for sync I/O, but if the whole array
1012 * is frozen, the read I/O still has to wait until the array is
1013 * unfrozen. Since there is no ordering requirement with
1014 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1016 atomic_inc(&conf->nr_pending[idx]);
1018 if (!READ_ONCE(conf->array_frozen))
1021 spin_lock_irq(&conf->resync_lock);
1022 atomic_inc(&conf->nr_waiting[idx]);
1023 atomic_dec(&conf->nr_pending[idx]);
1025 * In case freeze_array() is waiting for
1026 * get_unqueued_pending() == extra
1028 wake_up(&conf->wait_barrier);
1029 /* Wait for array to be unfrozen */
1031 /* Return false when nowait flag is set */
1033 /* Return false when nowait flag is set */
1036 wait_event_lock_irq(conf->wait_barrier,
1037 !conf->array_frozen,
1039 atomic_inc(&conf->nr_pending[idx]);
1042 atomic_dec(&conf->nr_waiting[idx]);
1043 spin_unlock_irq(&conf->resync_lock);
1047 static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1049 int idx = sector_to_idx(sector_nr);
1051 return _wait_barrier(conf, idx, nowait);
1054 static void _allow_barrier(struct r1conf *conf, int idx)
1056 atomic_dec(&conf->nr_pending[idx]);
1057 wake_up(&conf->wait_barrier);
1060 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1062 int idx = sector_to_idx(sector_nr);
1064 _allow_barrier(conf, idx);
1067 /* conf->resync_lock should be held */
1068 static int get_unqueued_pending(struct r1conf *conf)
1072 ret = atomic_read(&conf->nr_sync_pending);
1073 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1074 ret += atomic_read(&conf->nr_pending[idx]) -
1075 atomic_read(&conf->nr_queued[idx]);
1080 static void freeze_array(struct r1conf *conf, int extra)
1082 /* Stop sync I/O and normal I/O and wait for everything to
1084 * This is called in two situations:
1085 * 1) management command handlers (reshape, remove disk, quiesce).
1086 * 2) one normal I/O request failed.
1088 * After array_frozen is set to 1, new sync IO will be blocked at
1089 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1090 * or wait_read_barrier(). The flying I/Os will either complete or be
1091 * queued. When everything goes quite, there are only queued I/Os left.
1093 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1094 * barrier bucket index which this I/O request hits. When all sync and
1095 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1096 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1097 * in handle_read_error(), we may call freeze_array() before trying to
1098 * fix the read error. In this case, the error read I/O is not queued,
1099 * so get_unqueued_pending() == 1.
1101 * Therefore before this function returns, we need to wait until
1102 * get_unqueued_pendings(conf) gets equal to extra. For
1103 * normal I/O context, extra is 1, in rested situations extra is 0.
1105 spin_lock_irq(&conf->resync_lock);
1106 conf->array_frozen = 1;
1107 raid1_log(conf->mddev, "wait freeze");
1108 wait_event_lock_irq_cmd(
1110 get_unqueued_pending(conf) == extra,
1112 flush_pending_writes(conf));
1113 spin_unlock_irq(&conf->resync_lock);
1115 static void unfreeze_array(struct r1conf *conf)
1117 /* reverse the effect of the freeze */
1118 spin_lock_irq(&conf->resync_lock);
1119 conf->array_frozen = 0;
1120 spin_unlock_irq(&conf->resync_lock);
1121 wake_up(&conf->wait_barrier);
1124 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1127 int size = bio->bi_iter.bi_size;
1128 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1130 struct bio *behind_bio = NULL;
1132 behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1133 &r1_bio->mddev->bio_set);
1137 /* discard op, we don't support writezero/writesame yet */
1138 if (!bio_has_data(bio)) {
1139 behind_bio->bi_iter.bi_size = size;
1143 while (i < vcnt && size) {
1145 int len = min_t(int, PAGE_SIZE, size);
1147 page = alloc_page(GFP_NOIO);
1148 if (unlikely(!page))
1151 bio_add_page(behind_bio, page, len, 0);
1157 bio_copy_data(behind_bio, bio);
1159 r1_bio->behind_master_bio = behind_bio;
1160 set_bit(R1BIO_BehindIO, &r1_bio->state);
1165 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1166 bio->bi_iter.bi_size);
1167 bio_free_pages(behind_bio);
1168 bio_put(behind_bio);
1171 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1173 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1175 struct mddev *mddev = plug->cb.data;
1176 struct r1conf *conf = mddev->private;
1179 if (from_schedule || current->bio_list) {
1180 spin_lock_irq(&conf->device_lock);
1181 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1182 spin_unlock_irq(&conf->device_lock);
1183 wake_up(&conf->wait_barrier);
1184 md_wakeup_thread(mddev->thread);
1189 /* we aren't scheduling, so we can do the write-out directly. */
1190 bio = bio_list_get(&plug->pending);
1191 flush_bio_list(conf, bio);
1195 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1197 r1_bio->master_bio = bio;
1198 r1_bio->sectors = bio_sectors(bio);
1200 r1_bio->mddev = mddev;
1201 r1_bio->sector = bio->bi_iter.bi_sector;
1204 static inline struct r1bio *
1205 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1207 struct r1conf *conf = mddev->private;
1208 struct r1bio *r1_bio;
1210 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1211 /* Ensure no bio records IO_BLOCKED */
1212 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1213 init_r1bio(r1_bio, mddev, bio);
1217 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1218 int max_read_sectors, struct r1bio *r1_bio)
1220 struct r1conf *conf = mddev->private;
1221 struct raid1_info *mirror;
1222 struct bio *read_bio;
1223 struct bitmap *bitmap = mddev->bitmap;
1224 const int op = bio_op(bio);
1225 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1228 bool r1bio_existed = !!r1_bio;
1229 char b[BDEVNAME_SIZE];
1232 * If r1_bio is set, we are blocking the raid1d thread
1233 * so there is a tiny risk of deadlock. So ask for
1234 * emergency memory if needed.
1236 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1238 if (r1bio_existed) {
1239 /* Need to get the block device name carefully */
1240 struct md_rdev *rdev;
1242 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1244 bdevname(rdev->bdev, b);
1251 * Still need barrier for READ in case that whole
1254 if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1255 bio->bi_opf & REQ_NOWAIT)) {
1256 bio_wouldblock_error(bio);
1261 r1_bio = alloc_r1bio(mddev, bio);
1263 init_r1bio(r1_bio, mddev, bio);
1264 r1_bio->sectors = max_read_sectors;
1267 * make_request() can abort the operation when read-ahead is being
1268 * used and no empty request is available.
1270 rdisk = read_balance(conf, r1_bio, &max_sectors);
1273 /* couldn't find anywhere to read from */
1274 if (r1bio_existed) {
1275 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1278 (unsigned long long)r1_bio->sector);
1280 raid_end_bio_io(r1_bio);
1283 mirror = conf->mirrors + rdisk;
1286 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1288 (unsigned long long)r1_bio->sector,
1289 bdevname(mirror->rdev->bdev, b));
1291 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1294 * Reading from a write-mostly device must take care not to
1295 * over-take any writes that are 'behind'
1297 raid1_log(mddev, "wait behind writes");
1298 wait_event(bitmap->behind_wait,
1299 atomic_read(&bitmap->behind_writes) == 0);
1302 if (max_sectors < bio_sectors(bio)) {
1303 struct bio *split = bio_split(bio, max_sectors,
1304 gfp, &conf->bio_split);
1305 bio_chain(split, bio);
1306 submit_bio_noacct(bio);
1308 r1_bio->master_bio = bio;
1309 r1_bio->sectors = max_sectors;
1312 r1_bio->read_disk = rdisk;
1314 if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1315 r1_bio->start_time = bio_start_io_acct(bio);
1317 read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1320 r1_bio->bios[rdisk] = read_bio;
1322 read_bio->bi_iter.bi_sector = r1_bio->sector +
1323 mirror->rdev->data_offset;
1324 read_bio->bi_end_io = raid1_end_read_request;
1325 bio_set_op_attrs(read_bio, op, do_sync);
1326 if (test_bit(FailFast, &mirror->rdev->flags) &&
1327 test_bit(R1BIO_FailFast, &r1_bio->state))
1328 read_bio->bi_opf |= MD_FAILFAST;
1329 read_bio->bi_private = r1_bio;
1332 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1335 submit_bio_noacct(read_bio);
1338 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1339 int max_write_sectors)
1341 struct r1conf *conf = mddev->private;
1342 struct r1bio *r1_bio;
1344 struct bitmap *bitmap = mddev->bitmap;
1345 unsigned long flags;
1346 struct md_rdev *blocked_rdev;
1347 struct blk_plug_cb *cb;
1348 struct raid1_plug_cb *plug = NULL;
1351 bool write_behind = false;
1353 if (mddev_is_clustered(mddev) &&
1354 md_cluster_ops->area_resyncing(mddev, WRITE,
1355 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1358 if (bio->bi_opf & REQ_NOWAIT) {
1359 bio_wouldblock_error(bio);
1363 prepare_to_wait(&conf->wait_barrier,
1365 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1366 bio->bi_iter.bi_sector,
1367 bio_end_sector(bio)))
1371 finish_wait(&conf->wait_barrier, &w);
1375 * Register the new request and wait if the reconstruction
1376 * thread has put up a bar for new requests.
1377 * Continue immediately if no resync is active currently.
1379 if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1380 bio->bi_opf & REQ_NOWAIT)) {
1381 bio_wouldblock_error(bio);
1385 r1_bio = alloc_r1bio(mddev, bio);
1386 r1_bio->sectors = max_write_sectors;
1388 /* first select target devices under rcu_lock and
1389 * inc refcount on their rdev. Record them by setting
1391 * If there are known/acknowledged bad blocks on any device on
1392 * which we have seen a write error, we want to avoid writing those
1394 * This potentially requires several writes to write around
1395 * the bad blocks. Each set of writes gets it's own r1bio
1396 * with a set of bios attached.
1399 disks = conf->raid_disks * 2;
1401 blocked_rdev = NULL;
1403 max_sectors = r1_bio->sectors;
1404 for (i = 0; i < disks; i++) {
1405 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1408 * The write-behind io is only attempted on drives marked as
1409 * write-mostly, which means we could allocate write behind
1412 if (rdev && test_bit(WriteMostly, &rdev->flags))
1413 write_behind = true;
1415 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1416 atomic_inc(&rdev->nr_pending);
1417 blocked_rdev = rdev;
1420 r1_bio->bios[i] = NULL;
1421 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1422 if (i < conf->raid_disks)
1423 set_bit(R1BIO_Degraded, &r1_bio->state);
1427 atomic_inc(&rdev->nr_pending);
1428 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1433 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1434 &first_bad, &bad_sectors);
1436 /* mustn't write here until the bad block is
1438 set_bit(BlockedBadBlocks, &rdev->flags);
1439 blocked_rdev = rdev;
1442 if (is_bad && first_bad <= r1_bio->sector) {
1443 /* Cannot write here at all */
1444 bad_sectors -= (r1_bio->sector - first_bad);
1445 if (bad_sectors < max_sectors)
1446 /* mustn't write more than bad_sectors
1447 * to other devices yet
1449 max_sectors = bad_sectors;
1450 rdev_dec_pending(rdev, mddev);
1451 /* We don't set R1BIO_Degraded as that
1452 * only applies if the disk is
1453 * missing, so it might be re-added,
1454 * and we want to know to recover this
1456 * In this case the device is here,
1457 * and the fact that this chunk is not
1458 * in-sync is recorded in the bad
1464 int good_sectors = first_bad - r1_bio->sector;
1465 if (good_sectors < max_sectors)
1466 max_sectors = good_sectors;
1469 r1_bio->bios[i] = bio;
1473 if (unlikely(blocked_rdev)) {
1474 /* Wait for this device to become unblocked */
1477 for (j = 0; j < i; j++)
1478 if (r1_bio->bios[j])
1479 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1481 allow_barrier(conf, bio->bi_iter.bi_sector);
1483 if (bio->bi_opf & REQ_NOWAIT) {
1484 bio_wouldblock_error(bio);
1487 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1488 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1489 wait_barrier(conf, bio->bi_iter.bi_sector, false);
1494 * When using a bitmap, we may call alloc_behind_master_bio below.
1495 * alloc_behind_master_bio allocates a copy of the data payload a page
1496 * at a time and thus needs a new bio that can fit the whole payload
1497 * this bio in page sized chunks.
1499 if (write_behind && bitmap)
1500 max_sectors = min_t(int, max_sectors,
1501 BIO_MAX_VECS * (PAGE_SIZE >> 9));
1502 if (max_sectors < bio_sectors(bio)) {
1503 struct bio *split = bio_split(bio, max_sectors,
1504 GFP_NOIO, &conf->bio_split);
1505 bio_chain(split, bio);
1506 submit_bio_noacct(bio);
1508 r1_bio->master_bio = bio;
1509 r1_bio->sectors = max_sectors;
1512 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1513 r1_bio->start_time = bio_start_io_acct(bio);
1514 atomic_set(&r1_bio->remaining, 1);
1515 atomic_set(&r1_bio->behind_remaining, 0);
1519 for (i = 0; i < disks; i++) {
1520 struct bio *mbio = NULL;
1521 struct md_rdev *rdev = conf->mirrors[i].rdev;
1522 if (!r1_bio->bios[i])
1527 * Not if there are too many, or cannot
1528 * allocate memory, or a reader on WriteMostly
1529 * is waiting for behind writes to flush */
1531 test_bit(WriteMostly, &rdev->flags) &&
1532 (atomic_read(&bitmap->behind_writes)
1533 < mddev->bitmap_info.max_write_behind) &&
1534 !waitqueue_active(&bitmap->behind_wait)) {
1535 alloc_behind_master_bio(r1_bio, bio);
1538 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1539 test_bit(R1BIO_BehindIO, &r1_bio->state));
1543 if (r1_bio->behind_master_bio) {
1544 mbio = bio_alloc_clone(rdev->bdev,
1545 r1_bio->behind_master_bio,
1546 GFP_NOIO, &mddev->bio_set);
1547 if (test_bit(CollisionCheck, &rdev->flags))
1548 wait_for_serialization(rdev, r1_bio);
1549 if (test_bit(WriteMostly, &rdev->flags))
1550 atomic_inc(&r1_bio->behind_remaining);
1552 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1555 if (mddev->serialize_policy)
1556 wait_for_serialization(rdev, r1_bio);
1559 r1_bio->bios[i] = mbio;
1561 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1562 mbio->bi_end_io = raid1_end_write_request;
1563 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1564 if (test_bit(FailFast, &rdev->flags) &&
1565 !test_bit(WriteMostly, &rdev->flags) &&
1566 conf->raid_disks - mddev->degraded > 1)
1567 mbio->bi_opf |= MD_FAILFAST;
1568 mbio->bi_private = r1_bio;
1570 atomic_inc(&r1_bio->remaining);
1573 trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1575 /* flush_pending_writes() needs access to the rdev so...*/
1576 mbio->bi_bdev = (void *)rdev;
1578 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1580 plug = container_of(cb, struct raid1_plug_cb, cb);
1584 bio_list_add(&plug->pending, mbio);
1586 spin_lock_irqsave(&conf->device_lock, flags);
1587 bio_list_add(&conf->pending_bio_list, mbio);
1588 spin_unlock_irqrestore(&conf->device_lock, flags);
1589 md_wakeup_thread(mddev->thread);
1593 r1_bio_write_done(r1_bio);
1595 /* In case raid1d snuck in to freeze_array */
1596 wake_up(&conf->wait_barrier);
1599 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1603 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1604 && md_flush_request(mddev, bio))
1608 * There is a limit to the maximum size, but
1609 * the read/write handler might find a lower limit
1610 * due to bad blocks. To avoid multiple splits,
1611 * we pass the maximum number of sectors down
1612 * and let the lower level perform the split.
1614 sectors = align_to_barrier_unit_end(
1615 bio->bi_iter.bi_sector, bio_sectors(bio));
1617 if (bio_data_dir(bio) == READ)
1618 raid1_read_request(mddev, bio, sectors, NULL);
1620 if (!md_write_start(mddev,bio))
1622 raid1_write_request(mddev, bio, sectors);
1627 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1629 struct r1conf *conf = mddev->private;
1632 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1633 conf->raid_disks - mddev->degraded);
1635 for (i = 0; i < conf->raid_disks; i++) {
1636 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1637 seq_printf(seq, "%s",
1638 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1641 seq_printf(seq, "]");
1644 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1646 char b[BDEVNAME_SIZE];
1647 struct r1conf *conf = mddev->private;
1648 unsigned long flags;
1651 * If it is not operational, then we have already marked it as dead
1652 * else if it is the last working disks with "fail_last_dev == false",
1653 * ignore the error, let the next level up know.
1654 * else mark the drive as failed
1656 spin_lock_irqsave(&conf->device_lock, flags);
1657 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1658 && (conf->raid_disks - mddev->degraded) == 1) {
1660 * Don't fail the drive, act as though we were just a
1661 * normal single drive.
1662 * However don't try a recovery from this drive as
1663 * it is very likely to fail.
1665 conf->recovery_disabled = mddev->recovery_disabled;
1666 spin_unlock_irqrestore(&conf->device_lock, flags);
1669 set_bit(Blocked, &rdev->flags);
1670 if (test_and_clear_bit(In_sync, &rdev->flags))
1672 set_bit(Faulty, &rdev->flags);
1673 spin_unlock_irqrestore(&conf->device_lock, flags);
1675 * if recovery is running, make sure it aborts.
1677 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1678 set_mask_bits(&mddev->sb_flags, 0,
1679 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1680 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1681 "md/raid1:%s: Operation continuing on %d devices.\n",
1682 mdname(mddev), bdevname(rdev->bdev, b),
1683 mdname(mddev), conf->raid_disks - mddev->degraded);
1686 static void print_conf(struct r1conf *conf)
1690 pr_debug("RAID1 conf printout:\n");
1692 pr_debug("(!conf)\n");
1695 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1699 for (i = 0; i < conf->raid_disks; i++) {
1700 char b[BDEVNAME_SIZE];
1701 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1703 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1704 i, !test_bit(In_sync, &rdev->flags),
1705 !test_bit(Faulty, &rdev->flags),
1706 bdevname(rdev->bdev,b));
1711 static void close_sync(struct r1conf *conf)
1715 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1716 _wait_barrier(conf, idx, false);
1717 _allow_barrier(conf, idx);
1720 mempool_exit(&conf->r1buf_pool);
1723 static int raid1_spare_active(struct mddev *mddev)
1726 struct r1conf *conf = mddev->private;
1728 unsigned long flags;
1731 * Find all failed disks within the RAID1 configuration
1732 * and mark them readable.
1733 * Called under mddev lock, so rcu protection not needed.
1734 * device_lock used to avoid races with raid1_end_read_request
1735 * which expects 'In_sync' flags and ->degraded to be consistent.
1737 spin_lock_irqsave(&conf->device_lock, flags);
1738 for (i = 0; i < conf->raid_disks; i++) {
1739 struct md_rdev *rdev = conf->mirrors[i].rdev;
1740 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1742 && !test_bit(Candidate, &repl->flags)
1743 && repl->recovery_offset == MaxSector
1744 && !test_bit(Faulty, &repl->flags)
1745 && !test_and_set_bit(In_sync, &repl->flags)) {
1746 /* replacement has just become active */
1748 !test_and_clear_bit(In_sync, &rdev->flags))
1751 /* Replaced device not technically
1752 * faulty, but we need to be sure
1753 * it gets removed and never re-added
1755 set_bit(Faulty, &rdev->flags);
1756 sysfs_notify_dirent_safe(
1761 && rdev->recovery_offset == MaxSector
1762 && !test_bit(Faulty, &rdev->flags)
1763 && !test_and_set_bit(In_sync, &rdev->flags)) {
1765 sysfs_notify_dirent_safe(rdev->sysfs_state);
1768 mddev->degraded -= count;
1769 spin_unlock_irqrestore(&conf->device_lock, flags);
1775 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1777 struct r1conf *conf = mddev->private;
1780 struct raid1_info *p;
1782 int last = conf->raid_disks - 1;
1784 if (mddev->recovery_disabled == conf->recovery_disabled)
1787 if (md_integrity_add_rdev(rdev, mddev))
1790 if (rdev->raid_disk >= 0)
1791 first = last = rdev->raid_disk;
1794 * find the disk ... but prefer rdev->saved_raid_disk
1797 if (rdev->saved_raid_disk >= 0 &&
1798 rdev->saved_raid_disk >= first &&
1799 rdev->saved_raid_disk < conf->raid_disks &&
1800 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1801 first = last = rdev->saved_raid_disk;
1803 for (mirror = first; mirror <= last; mirror++) {
1804 p = conf->mirrors + mirror;
1807 disk_stack_limits(mddev->gendisk, rdev->bdev,
1808 rdev->data_offset << 9);
1810 p->head_position = 0;
1811 rdev->raid_disk = mirror;
1813 /* As all devices are equivalent, we don't need a full recovery
1814 * if this was recently any drive of the array
1816 if (rdev->saved_raid_disk < 0)
1818 rcu_assign_pointer(p->rdev, rdev);
1821 if (test_bit(WantReplacement, &p->rdev->flags) &&
1822 p[conf->raid_disks].rdev == NULL) {
1823 /* Add this device as a replacement */
1824 clear_bit(In_sync, &rdev->flags);
1825 set_bit(Replacement, &rdev->flags);
1826 rdev->raid_disk = mirror;
1829 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1833 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1834 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1839 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1841 struct r1conf *conf = mddev->private;
1843 int number = rdev->raid_disk;
1844 struct raid1_info *p = conf->mirrors + number;
1846 if (rdev != p->rdev)
1847 p = conf->mirrors + conf->raid_disks + number;
1850 if (rdev == p->rdev) {
1851 if (test_bit(In_sync, &rdev->flags) ||
1852 atomic_read(&rdev->nr_pending)) {
1856 /* Only remove non-faulty devices if recovery
1859 if (!test_bit(Faulty, &rdev->flags) &&
1860 mddev->recovery_disabled != conf->recovery_disabled &&
1861 mddev->degraded < conf->raid_disks) {
1866 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1868 if (atomic_read(&rdev->nr_pending)) {
1869 /* lost the race, try later */
1875 if (conf->mirrors[conf->raid_disks + number].rdev) {
1876 /* We just removed a device that is being replaced.
1877 * Move down the replacement. We drain all IO before
1878 * doing this to avoid confusion.
1880 struct md_rdev *repl =
1881 conf->mirrors[conf->raid_disks + number].rdev;
1882 freeze_array(conf, 0);
1883 if (atomic_read(&repl->nr_pending)) {
1884 /* It means that some queued IO of retry_list
1885 * hold repl. Thus, we cannot set replacement
1886 * as NULL, avoiding rdev NULL pointer
1887 * dereference in sync_request_write and
1888 * handle_write_finished.
1891 unfreeze_array(conf);
1894 clear_bit(Replacement, &repl->flags);
1896 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1897 unfreeze_array(conf);
1900 clear_bit(WantReplacement, &rdev->flags);
1901 err = md_integrity_register(mddev);
1909 static void end_sync_read(struct bio *bio)
1911 struct r1bio *r1_bio = get_resync_r1bio(bio);
1913 update_head_pos(r1_bio->read_disk, r1_bio);
1916 * we have read a block, now it needs to be re-written,
1917 * or re-read if the read failed.
1918 * We don't do much here, just schedule handling by raid1d
1920 if (!bio->bi_status)
1921 set_bit(R1BIO_Uptodate, &r1_bio->state);
1923 if (atomic_dec_and_test(&r1_bio->remaining))
1924 reschedule_retry(r1_bio);
1927 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1929 sector_t sync_blocks = 0;
1930 sector_t s = r1_bio->sector;
1931 long sectors_to_go = r1_bio->sectors;
1933 /* make sure these bits don't get cleared. */
1935 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1937 sectors_to_go -= sync_blocks;
1938 } while (sectors_to_go > 0);
1941 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1943 if (atomic_dec_and_test(&r1_bio->remaining)) {
1944 struct mddev *mddev = r1_bio->mddev;
1945 int s = r1_bio->sectors;
1947 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1948 test_bit(R1BIO_WriteError, &r1_bio->state))
1949 reschedule_retry(r1_bio);
1952 md_done_sync(mddev, s, uptodate);
1957 static void end_sync_write(struct bio *bio)
1959 int uptodate = !bio->bi_status;
1960 struct r1bio *r1_bio = get_resync_r1bio(bio);
1961 struct mddev *mddev = r1_bio->mddev;
1962 struct r1conf *conf = mddev->private;
1965 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1968 abort_sync_write(mddev, r1_bio);
1969 set_bit(WriteErrorSeen, &rdev->flags);
1970 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1971 set_bit(MD_RECOVERY_NEEDED, &
1973 set_bit(R1BIO_WriteError, &r1_bio->state);
1974 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1975 &first_bad, &bad_sectors) &&
1976 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1979 &first_bad, &bad_sectors)
1981 set_bit(R1BIO_MadeGood, &r1_bio->state);
1983 put_sync_write_buf(r1_bio, uptodate);
1986 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1987 int sectors, struct page *page, int rw)
1989 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1993 set_bit(WriteErrorSeen, &rdev->flags);
1994 if (!test_and_set_bit(WantReplacement,
1996 set_bit(MD_RECOVERY_NEEDED, &
1997 rdev->mddev->recovery);
1999 /* need to record an error - either for the block or the device */
2000 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2001 md_error(rdev->mddev, rdev);
2005 static int fix_sync_read_error(struct r1bio *r1_bio)
2007 /* Try some synchronous reads of other devices to get
2008 * good data, much like with normal read errors. Only
2009 * read into the pages we already have so we don't
2010 * need to re-issue the read request.
2011 * We don't need to freeze the array, because being in an
2012 * active sync request, there is no normal IO, and
2013 * no overlapping syncs.
2014 * We don't need to check is_badblock() again as we
2015 * made sure that anything with a bad block in range
2016 * will have bi_end_io clear.
2018 struct mddev *mddev = r1_bio->mddev;
2019 struct r1conf *conf = mddev->private;
2020 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2021 struct page **pages = get_resync_pages(bio)->pages;
2022 sector_t sect = r1_bio->sector;
2023 int sectors = r1_bio->sectors;
2025 struct md_rdev *rdev;
2027 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2028 if (test_bit(FailFast, &rdev->flags)) {
2029 /* Don't try recovering from here - just fail it
2030 * ... unless it is the last working device of course */
2031 md_error(mddev, rdev);
2032 if (test_bit(Faulty, &rdev->flags))
2033 /* Don't try to read from here, but make sure
2034 * put_buf does it's thing
2036 bio->bi_end_io = end_sync_write;
2041 int d = r1_bio->read_disk;
2045 if (s > (PAGE_SIZE>>9))
2048 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2049 /* No rcu protection needed here devices
2050 * can only be removed when no resync is
2051 * active, and resync is currently active
2053 rdev = conf->mirrors[d].rdev;
2054 if (sync_page_io(rdev, sect, s<<9,
2056 REQ_OP_READ, 0, false)) {
2062 if (d == conf->raid_disks * 2)
2064 } while (!success && d != r1_bio->read_disk);
2068 /* Cannot read from anywhere, this block is lost.
2069 * Record a bad block on each device. If that doesn't
2070 * work just disable and interrupt the recovery.
2071 * Don't fail devices as that won't really help.
2073 pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2074 mdname(mddev), bio->bi_bdev,
2075 (unsigned long long)r1_bio->sector);
2076 for (d = 0; d < conf->raid_disks * 2; d++) {
2077 rdev = conf->mirrors[d].rdev;
2078 if (!rdev || test_bit(Faulty, &rdev->flags))
2080 if (!rdev_set_badblocks(rdev, sect, s, 0))
2084 conf->recovery_disabled =
2085 mddev->recovery_disabled;
2086 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2087 md_done_sync(mddev, r1_bio->sectors, 0);
2099 /* write it back and re-read */
2100 while (d != r1_bio->read_disk) {
2102 d = conf->raid_disks * 2;
2104 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2106 rdev = conf->mirrors[d].rdev;
2107 if (r1_sync_page_io(rdev, sect, s,
2110 r1_bio->bios[d]->bi_end_io = NULL;
2111 rdev_dec_pending(rdev, mddev);
2115 while (d != r1_bio->read_disk) {
2117 d = conf->raid_disks * 2;
2119 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2121 rdev = conf->mirrors[d].rdev;
2122 if (r1_sync_page_io(rdev, sect, s,
2125 atomic_add(s, &rdev->corrected_errors);
2131 set_bit(R1BIO_Uptodate, &r1_bio->state);
2136 static void process_checks(struct r1bio *r1_bio)
2138 /* We have read all readable devices. If we haven't
2139 * got the block, then there is no hope left.
2140 * If we have, then we want to do a comparison
2141 * and skip the write if everything is the same.
2142 * If any blocks failed to read, then we need to
2143 * attempt an over-write
2145 struct mddev *mddev = r1_bio->mddev;
2146 struct r1conf *conf = mddev->private;
2151 /* Fix variable parts of all bios */
2152 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2153 for (i = 0; i < conf->raid_disks * 2; i++) {
2154 blk_status_t status;
2155 struct bio *b = r1_bio->bios[i];
2156 struct resync_pages *rp = get_resync_pages(b);
2157 if (b->bi_end_io != end_sync_read)
2159 /* fixup the bio for reuse, but preserve errno */
2160 status = b->bi_status;
2161 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2162 b->bi_status = status;
2163 b->bi_iter.bi_sector = r1_bio->sector +
2164 conf->mirrors[i].rdev->data_offset;
2165 b->bi_end_io = end_sync_read;
2166 rp->raid_bio = r1_bio;
2169 /* initialize bvec table again */
2170 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2172 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2173 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2174 !r1_bio->bios[primary]->bi_status) {
2175 r1_bio->bios[primary]->bi_end_io = NULL;
2176 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2179 r1_bio->read_disk = primary;
2180 for (i = 0; i < conf->raid_disks * 2; i++) {
2182 struct bio *pbio = r1_bio->bios[primary];
2183 struct bio *sbio = r1_bio->bios[i];
2184 blk_status_t status = sbio->bi_status;
2185 struct page **ppages = get_resync_pages(pbio)->pages;
2186 struct page **spages = get_resync_pages(sbio)->pages;
2188 int page_len[RESYNC_PAGES] = { 0 };
2189 struct bvec_iter_all iter_all;
2191 if (sbio->bi_end_io != end_sync_read)
2193 /* Now we can 'fixup' the error value */
2194 sbio->bi_status = 0;
2196 bio_for_each_segment_all(bi, sbio, iter_all)
2197 page_len[j++] = bi->bv_len;
2200 for (j = vcnt; j-- ; ) {
2201 if (memcmp(page_address(ppages[j]),
2202 page_address(spages[j]),
2209 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2210 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2212 /* No need to write to this device. */
2213 sbio->bi_end_io = NULL;
2214 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2218 bio_copy_data(sbio, pbio);
2222 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2224 struct r1conf *conf = mddev->private;
2226 int disks = conf->raid_disks * 2;
2229 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2230 /* ouch - failed to read all of that. */
2231 if (!fix_sync_read_error(r1_bio))
2234 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2235 process_checks(r1_bio);
2240 atomic_set(&r1_bio->remaining, 1);
2241 for (i = 0; i < disks ; i++) {
2242 wbio = r1_bio->bios[i];
2243 if (wbio->bi_end_io == NULL ||
2244 (wbio->bi_end_io == end_sync_read &&
2245 (i == r1_bio->read_disk ||
2246 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2248 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2249 abort_sync_write(mddev, r1_bio);
2253 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2254 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2255 wbio->bi_opf |= MD_FAILFAST;
2257 wbio->bi_end_io = end_sync_write;
2258 atomic_inc(&r1_bio->remaining);
2259 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2261 submit_bio_noacct(wbio);
2264 put_sync_write_buf(r1_bio, 1);
2268 * This is a kernel thread which:
2270 * 1. Retries failed read operations on working mirrors.
2271 * 2. Updates the raid superblock when problems encounter.
2272 * 3. Performs writes following reads for array synchronising.
2275 static void fix_read_error(struct r1conf *conf, int read_disk,
2276 sector_t sect, int sectors)
2278 struct mddev *mddev = conf->mddev;
2284 struct md_rdev *rdev;
2286 if (s > (PAGE_SIZE>>9))
2294 rdev = rcu_dereference(conf->mirrors[d].rdev);
2296 (test_bit(In_sync, &rdev->flags) ||
2297 (!test_bit(Faulty, &rdev->flags) &&
2298 rdev->recovery_offset >= sect + s)) &&
2299 is_badblock(rdev, sect, s,
2300 &first_bad, &bad_sectors) == 0) {
2301 atomic_inc(&rdev->nr_pending);
2303 if (sync_page_io(rdev, sect, s<<9,
2304 conf->tmppage, REQ_OP_READ, 0, false))
2306 rdev_dec_pending(rdev, mddev);
2312 if (d == conf->raid_disks * 2)
2314 } while (!success && d != read_disk);
2317 /* Cannot read from anywhere - mark it bad */
2318 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2319 if (!rdev_set_badblocks(rdev, sect, s, 0))
2320 md_error(mddev, rdev);
2323 /* write it back and re-read */
2325 while (d != read_disk) {
2327 d = conf->raid_disks * 2;
2330 rdev = rcu_dereference(conf->mirrors[d].rdev);
2332 !test_bit(Faulty, &rdev->flags)) {
2333 atomic_inc(&rdev->nr_pending);
2335 r1_sync_page_io(rdev, sect, s,
2336 conf->tmppage, WRITE);
2337 rdev_dec_pending(rdev, mddev);
2342 while (d != read_disk) {
2343 char b[BDEVNAME_SIZE];
2345 d = conf->raid_disks * 2;
2348 rdev = rcu_dereference(conf->mirrors[d].rdev);
2350 !test_bit(Faulty, &rdev->flags)) {
2351 atomic_inc(&rdev->nr_pending);
2353 if (r1_sync_page_io(rdev, sect, s,
2354 conf->tmppage, READ)) {
2355 atomic_add(s, &rdev->corrected_errors);
2356 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2358 (unsigned long long)(sect +
2360 bdevname(rdev->bdev, b));
2362 rdev_dec_pending(rdev, mddev);
2371 static int narrow_write_error(struct r1bio *r1_bio, int i)
2373 struct mddev *mddev = r1_bio->mddev;
2374 struct r1conf *conf = mddev->private;
2375 struct md_rdev *rdev = conf->mirrors[i].rdev;
2377 /* bio has the data to be written to device 'i' where
2378 * we just recently had a write error.
2379 * We repeatedly clone the bio and trim down to one block,
2380 * then try the write. Where the write fails we record
2382 * It is conceivable that the bio doesn't exactly align with
2383 * blocks. We must handle this somehow.
2385 * We currently own a reference on the rdev.
2391 int sect_to_write = r1_bio->sectors;
2394 if (rdev->badblocks.shift < 0)
2397 block_sectors = roundup(1 << rdev->badblocks.shift,
2398 bdev_logical_block_size(rdev->bdev) >> 9);
2399 sector = r1_bio->sector;
2400 sectors = ((sector + block_sectors)
2401 & ~(sector_t)(block_sectors - 1))
2404 while (sect_to_write) {
2406 if (sectors > sect_to_write)
2407 sectors = sect_to_write;
2408 /* Write at 'sector' for 'sectors'*/
2410 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2411 wbio = bio_alloc_clone(rdev->bdev,
2412 r1_bio->behind_master_bio,
2413 GFP_NOIO, &mddev->bio_set);
2415 wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2416 GFP_NOIO, &mddev->bio_set);
2419 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2420 wbio->bi_iter.bi_sector = r1_bio->sector;
2421 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2423 bio_trim(wbio, sector - r1_bio->sector, sectors);
2424 wbio->bi_iter.bi_sector += rdev->data_offset;
2426 if (submit_bio_wait(wbio) < 0)
2428 ok = rdev_set_badblocks(rdev, sector,
2433 sect_to_write -= sectors;
2435 sectors = block_sectors;
2440 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2443 int s = r1_bio->sectors;
2444 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2445 struct md_rdev *rdev = conf->mirrors[m].rdev;
2446 struct bio *bio = r1_bio->bios[m];
2447 if (bio->bi_end_io == NULL)
2449 if (!bio->bi_status &&
2450 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2451 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2453 if (bio->bi_status &&
2454 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2455 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2456 md_error(conf->mddev, rdev);
2460 md_done_sync(conf->mddev, s, 1);
2463 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2468 for (m = 0; m < conf->raid_disks * 2 ; m++)
2469 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2470 struct md_rdev *rdev = conf->mirrors[m].rdev;
2471 rdev_clear_badblocks(rdev,
2473 r1_bio->sectors, 0);
2474 rdev_dec_pending(rdev, conf->mddev);
2475 } else if (r1_bio->bios[m] != NULL) {
2476 /* This drive got a write error. We need to
2477 * narrow down and record precise write
2481 if (!narrow_write_error(r1_bio, m)) {
2482 md_error(conf->mddev,
2483 conf->mirrors[m].rdev);
2484 /* an I/O failed, we can't clear the bitmap */
2485 set_bit(R1BIO_Degraded, &r1_bio->state);
2487 rdev_dec_pending(conf->mirrors[m].rdev,
2491 spin_lock_irq(&conf->device_lock);
2492 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2493 idx = sector_to_idx(r1_bio->sector);
2494 atomic_inc(&conf->nr_queued[idx]);
2495 spin_unlock_irq(&conf->device_lock);
2497 * In case freeze_array() is waiting for condition
2498 * get_unqueued_pending() == extra to be true.
2500 wake_up(&conf->wait_barrier);
2501 md_wakeup_thread(conf->mddev->thread);
2503 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2504 close_write(r1_bio);
2505 raid_end_bio_io(r1_bio);
2509 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2511 struct mddev *mddev = conf->mddev;
2513 struct md_rdev *rdev;
2515 clear_bit(R1BIO_ReadError, &r1_bio->state);
2516 /* we got a read error. Maybe the drive is bad. Maybe just
2517 * the block and we can fix it.
2518 * We freeze all other IO, and try reading the block from
2519 * other devices. When we find one, we re-write
2520 * and check it that fixes the read error.
2521 * This is all done synchronously while the array is
2525 bio = r1_bio->bios[r1_bio->read_disk];
2527 r1_bio->bios[r1_bio->read_disk] = NULL;
2529 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2531 && !test_bit(FailFast, &rdev->flags)) {
2532 freeze_array(conf, 1);
2533 fix_read_error(conf, r1_bio->read_disk,
2534 r1_bio->sector, r1_bio->sectors);
2535 unfreeze_array(conf);
2536 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2537 md_error(mddev, rdev);
2539 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2542 rdev_dec_pending(rdev, conf->mddev);
2543 allow_barrier(conf, r1_bio->sector);
2544 bio = r1_bio->master_bio;
2546 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2548 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2551 static void raid1d(struct md_thread *thread)
2553 struct mddev *mddev = thread->mddev;
2554 struct r1bio *r1_bio;
2555 unsigned long flags;
2556 struct r1conf *conf = mddev->private;
2557 struct list_head *head = &conf->retry_list;
2558 struct blk_plug plug;
2561 md_check_recovery(mddev);
2563 if (!list_empty_careful(&conf->bio_end_io_list) &&
2564 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2566 spin_lock_irqsave(&conf->device_lock, flags);
2567 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2568 list_splice_init(&conf->bio_end_io_list, &tmp);
2569 spin_unlock_irqrestore(&conf->device_lock, flags);
2570 while (!list_empty(&tmp)) {
2571 r1_bio = list_first_entry(&tmp, struct r1bio,
2573 list_del(&r1_bio->retry_list);
2574 idx = sector_to_idx(r1_bio->sector);
2575 atomic_dec(&conf->nr_queued[idx]);
2576 if (mddev->degraded)
2577 set_bit(R1BIO_Degraded, &r1_bio->state);
2578 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2579 close_write(r1_bio);
2580 raid_end_bio_io(r1_bio);
2584 blk_start_plug(&plug);
2587 flush_pending_writes(conf);
2589 spin_lock_irqsave(&conf->device_lock, flags);
2590 if (list_empty(head)) {
2591 spin_unlock_irqrestore(&conf->device_lock, flags);
2594 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2595 list_del(head->prev);
2596 idx = sector_to_idx(r1_bio->sector);
2597 atomic_dec(&conf->nr_queued[idx]);
2598 spin_unlock_irqrestore(&conf->device_lock, flags);
2600 mddev = r1_bio->mddev;
2601 conf = mddev->private;
2602 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2603 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2604 test_bit(R1BIO_WriteError, &r1_bio->state))
2605 handle_sync_write_finished(conf, r1_bio);
2607 sync_request_write(mddev, r1_bio);
2608 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2609 test_bit(R1BIO_WriteError, &r1_bio->state))
2610 handle_write_finished(conf, r1_bio);
2611 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2612 handle_read_error(conf, r1_bio);
2617 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2618 md_check_recovery(mddev);
2620 blk_finish_plug(&plug);
2623 static int init_resync(struct r1conf *conf)
2627 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2628 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2630 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2631 r1buf_pool_free, conf->poolinfo);
2634 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2636 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2637 struct resync_pages *rps;
2641 for (i = conf->poolinfo->raid_disks; i--; ) {
2642 bio = r1bio->bios[i];
2643 rps = bio->bi_private;
2644 bio_reset(bio, NULL, 0);
2645 bio->bi_private = rps;
2647 r1bio->master_bio = NULL;
2652 * perform a "sync" on one "block"
2654 * We need to make sure that no normal I/O request - particularly write
2655 * requests - conflict with active sync requests.
2657 * This is achieved by tracking pending requests and a 'barrier' concept
2658 * that can be installed to exclude normal IO requests.
2661 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2664 struct r1conf *conf = mddev->private;
2665 struct r1bio *r1_bio;
2667 sector_t max_sector, nr_sectors;
2671 int write_targets = 0, read_targets = 0;
2672 sector_t sync_blocks;
2673 int still_degraded = 0;
2674 int good_sectors = RESYNC_SECTORS;
2675 int min_bad = 0; /* number of sectors that are bad in all devices */
2676 int idx = sector_to_idx(sector_nr);
2679 if (!mempool_initialized(&conf->r1buf_pool))
2680 if (init_resync(conf))
2683 max_sector = mddev->dev_sectors;
2684 if (sector_nr >= max_sector) {
2685 /* If we aborted, we need to abort the
2686 * sync on the 'current' bitmap chunk (there will
2687 * only be one in raid1 resync.
2688 * We can find the current addess in mddev->curr_resync
2690 if (mddev->curr_resync < max_sector) /* aborted */
2691 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2693 else /* completed sync */
2696 md_bitmap_close_sync(mddev->bitmap);
2699 if (mddev_is_clustered(mddev)) {
2700 conf->cluster_sync_low = 0;
2701 conf->cluster_sync_high = 0;
2706 if (mddev->bitmap == NULL &&
2707 mddev->recovery_cp == MaxSector &&
2708 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2709 conf->fullsync == 0) {
2711 return max_sector - sector_nr;
2713 /* before building a request, check if we can skip these blocks..
2714 * This call the bitmap_start_sync doesn't actually record anything
2716 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2717 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2718 /* We can skip this block, and probably several more */
2724 * If there is non-resync activity waiting for a turn, then let it
2725 * though before starting on this new sync request.
2727 if (atomic_read(&conf->nr_waiting[idx]))
2728 schedule_timeout_uninterruptible(1);
2730 /* we are incrementing sector_nr below. To be safe, we check against
2731 * sector_nr + two times RESYNC_SECTORS
2734 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2735 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2738 if (raise_barrier(conf, sector_nr))
2741 r1_bio = raid1_alloc_init_r1buf(conf);
2745 * If we get a correctably read error during resync or recovery,
2746 * we might want to read from a different device. So we
2747 * flag all drives that could conceivably be read from for READ,
2748 * and any others (which will be non-In_sync devices) for WRITE.
2749 * If a read fails, we try reading from something else for which READ
2753 r1_bio->mddev = mddev;
2754 r1_bio->sector = sector_nr;
2756 set_bit(R1BIO_IsSync, &r1_bio->state);
2757 /* make sure good_sectors won't go across barrier unit boundary */
2758 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2760 for (i = 0; i < conf->raid_disks * 2; i++) {
2761 struct md_rdev *rdev;
2762 bio = r1_bio->bios[i];
2764 rdev = rcu_dereference(conf->mirrors[i].rdev);
2766 test_bit(Faulty, &rdev->flags)) {
2767 if (i < conf->raid_disks)
2769 } else if (!test_bit(In_sync, &rdev->flags)) {
2770 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2771 bio->bi_end_io = end_sync_write;
2774 /* may need to read from here */
2775 sector_t first_bad = MaxSector;
2778 if (is_badblock(rdev, sector_nr, good_sectors,
2779 &first_bad, &bad_sectors)) {
2780 if (first_bad > sector_nr)
2781 good_sectors = first_bad - sector_nr;
2783 bad_sectors -= (sector_nr - first_bad);
2785 min_bad > bad_sectors)
2786 min_bad = bad_sectors;
2789 if (sector_nr < first_bad) {
2790 if (test_bit(WriteMostly, &rdev->flags)) {
2797 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2798 bio->bi_end_io = end_sync_read;
2800 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2801 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2802 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2804 * The device is suitable for reading (InSync),
2805 * but has bad block(s) here. Let's try to correct them,
2806 * if we are doing resync or repair. Otherwise, leave
2807 * this device alone for this sync request.
2809 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2810 bio->bi_end_io = end_sync_write;
2814 if (rdev && bio->bi_end_io) {
2815 atomic_inc(&rdev->nr_pending);
2816 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2817 bio_set_dev(bio, rdev->bdev);
2818 if (test_bit(FailFast, &rdev->flags))
2819 bio->bi_opf |= MD_FAILFAST;
2825 r1_bio->read_disk = disk;
2827 if (read_targets == 0 && min_bad > 0) {
2828 /* These sectors are bad on all InSync devices, so we
2829 * need to mark them bad on all write targets
2832 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2833 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2834 struct md_rdev *rdev = conf->mirrors[i].rdev;
2835 ok = rdev_set_badblocks(rdev, sector_nr,
2839 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2844 /* Cannot record the badblocks, so need to
2846 * If there are multiple read targets, could just
2847 * fail the really bad ones ???
2849 conf->recovery_disabled = mddev->recovery_disabled;
2850 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2856 if (min_bad > 0 && min_bad < good_sectors) {
2857 /* only resync enough to reach the next bad->good
2859 good_sectors = min_bad;
2862 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2863 /* extra read targets are also write targets */
2864 write_targets += read_targets-1;
2866 if (write_targets == 0 || read_targets == 0) {
2867 /* There is nowhere to write, so all non-sync
2868 * drives must be failed - so we are finished
2872 max_sector = sector_nr + min_bad;
2873 rv = max_sector - sector_nr;
2879 if (max_sector > mddev->resync_max)
2880 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2881 if (max_sector > sector_nr + good_sectors)
2882 max_sector = sector_nr + good_sectors;
2887 int len = PAGE_SIZE;
2888 if (sector_nr + (len>>9) > max_sector)
2889 len = (max_sector - sector_nr) << 9;
2892 if (sync_blocks == 0) {
2893 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2894 &sync_blocks, still_degraded) &&
2896 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2898 if ((len >> 9) > sync_blocks)
2899 len = sync_blocks<<9;
2902 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2903 struct resync_pages *rp;
2905 bio = r1_bio->bios[i];
2906 rp = get_resync_pages(bio);
2907 if (bio->bi_end_io) {
2908 page = resync_fetch_page(rp, page_idx);
2911 * won't fail because the vec table is big
2912 * enough to hold all these pages
2914 bio_add_page(bio, page, len, 0);
2917 nr_sectors += len>>9;
2918 sector_nr += len>>9;
2919 sync_blocks -= (len>>9);
2920 } while (++page_idx < RESYNC_PAGES);
2922 r1_bio->sectors = nr_sectors;
2924 if (mddev_is_clustered(mddev) &&
2925 conf->cluster_sync_high < sector_nr + nr_sectors) {
2926 conf->cluster_sync_low = mddev->curr_resync_completed;
2927 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2928 /* Send resync message */
2929 md_cluster_ops->resync_info_update(mddev,
2930 conf->cluster_sync_low,
2931 conf->cluster_sync_high);
2934 /* For a user-requested sync, we read all readable devices and do a
2937 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2938 atomic_set(&r1_bio->remaining, read_targets);
2939 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2940 bio = r1_bio->bios[i];
2941 if (bio->bi_end_io == end_sync_read) {
2943 md_sync_acct_bio(bio, nr_sectors);
2944 if (read_targets == 1)
2945 bio->bi_opf &= ~MD_FAILFAST;
2946 submit_bio_noacct(bio);
2950 atomic_set(&r1_bio->remaining, 1);
2951 bio = r1_bio->bios[r1_bio->read_disk];
2952 md_sync_acct_bio(bio, nr_sectors);
2953 if (read_targets == 1)
2954 bio->bi_opf &= ~MD_FAILFAST;
2955 submit_bio_noacct(bio);
2960 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2965 return mddev->dev_sectors;
2968 static struct r1conf *setup_conf(struct mddev *mddev)
2970 struct r1conf *conf;
2972 struct raid1_info *disk;
2973 struct md_rdev *rdev;
2976 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2980 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2981 sizeof(atomic_t), GFP_KERNEL);
2982 if (!conf->nr_pending)
2985 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2986 sizeof(atomic_t), GFP_KERNEL);
2987 if (!conf->nr_waiting)
2990 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2991 sizeof(atomic_t), GFP_KERNEL);
2992 if (!conf->nr_queued)
2995 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2996 sizeof(atomic_t), GFP_KERNEL);
3000 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3001 mddev->raid_disks, 2),
3006 conf->tmppage = alloc_page(GFP_KERNEL);
3010 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3011 if (!conf->poolinfo)
3013 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3014 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3015 rbio_pool_free, conf->poolinfo);
3019 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3023 conf->poolinfo->mddev = mddev;
3026 spin_lock_init(&conf->device_lock);
3027 rdev_for_each(rdev, mddev) {
3028 int disk_idx = rdev->raid_disk;
3029 if (disk_idx >= mddev->raid_disks
3032 if (test_bit(Replacement, &rdev->flags))
3033 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3035 disk = conf->mirrors + disk_idx;
3040 disk->head_position = 0;
3041 disk->seq_start = MaxSector;
3043 conf->raid_disks = mddev->raid_disks;
3044 conf->mddev = mddev;
3045 INIT_LIST_HEAD(&conf->retry_list);
3046 INIT_LIST_HEAD(&conf->bio_end_io_list);
3048 spin_lock_init(&conf->resync_lock);
3049 init_waitqueue_head(&conf->wait_barrier);
3051 bio_list_init(&conf->pending_bio_list);
3052 conf->recovery_disabled = mddev->recovery_disabled - 1;
3055 for (i = 0; i < conf->raid_disks * 2; i++) {
3057 disk = conf->mirrors + i;
3059 if (i < conf->raid_disks &&
3060 disk[conf->raid_disks].rdev) {
3061 /* This slot has a replacement. */
3063 /* No original, just make the replacement
3064 * a recovering spare
3067 disk[conf->raid_disks].rdev;
3068 disk[conf->raid_disks].rdev = NULL;
3069 } else if (!test_bit(In_sync, &disk->rdev->flags))
3070 /* Original is not in_sync - bad */
3075 !test_bit(In_sync, &disk->rdev->flags)) {
3076 disk->head_position = 0;
3078 (disk->rdev->saved_raid_disk < 0))
3084 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3092 mempool_exit(&conf->r1bio_pool);
3093 kfree(conf->mirrors);
3094 safe_put_page(conf->tmppage);
3095 kfree(conf->poolinfo);
3096 kfree(conf->nr_pending);
3097 kfree(conf->nr_waiting);
3098 kfree(conf->nr_queued);
3099 kfree(conf->barrier);
3100 bioset_exit(&conf->bio_split);
3103 return ERR_PTR(err);
3106 static void raid1_free(struct mddev *mddev, void *priv);
3107 static int raid1_run(struct mddev *mddev)
3109 struct r1conf *conf;
3111 struct md_rdev *rdev;
3113 bool discard_supported = false;
3115 if (mddev->level != 1) {
3116 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3117 mdname(mddev), mddev->level);
3120 if (mddev->reshape_position != MaxSector) {
3121 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3125 if (mddev_init_writes_pending(mddev) < 0)
3128 * copy the already verified devices into our private RAID1
3129 * bookkeeping area. [whatever we allocate in run(),
3130 * should be freed in raid1_free()]
3132 if (mddev->private == NULL)
3133 conf = setup_conf(mddev);
3135 conf = mddev->private;
3138 return PTR_ERR(conf);
3141 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3143 rdev_for_each(rdev, mddev) {
3144 if (!mddev->gendisk)
3146 disk_stack_limits(mddev->gendisk, rdev->bdev,
3147 rdev->data_offset << 9);
3148 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3149 discard_supported = true;
3152 mddev->degraded = 0;
3153 for (i = 0; i < conf->raid_disks; i++)
3154 if (conf->mirrors[i].rdev == NULL ||
3155 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3156 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3159 * RAID1 needs at least one disk in active
3161 if (conf->raid_disks - mddev->degraded < 1) {
3166 if (conf->raid_disks - mddev->degraded == 1)
3167 mddev->recovery_cp = MaxSector;
3169 if (mddev->recovery_cp != MaxSector)
3170 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3172 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3173 mdname(mddev), mddev->raid_disks - mddev->degraded,
3177 * Ok, everything is just fine now
3179 mddev->thread = conf->thread;
3180 conf->thread = NULL;
3181 mddev->private = conf;
3182 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3184 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3187 if (discard_supported)
3188 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3191 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3195 ret = md_integrity_register(mddev);
3197 md_unregister_thread(&mddev->thread);
3203 raid1_free(mddev, conf);
3207 static void raid1_free(struct mddev *mddev, void *priv)
3209 struct r1conf *conf = priv;
3211 mempool_exit(&conf->r1bio_pool);
3212 kfree(conf->mirrors);
3213 safe_put_page(conf->tmppage);
3214 kfree(conf->poolinfo);
3215 kfree(conf->nr_pending);
3216 kfree(conf->nr_waiting);
3217 kfree(conf->nr_queued);
3218 kfree(conf->barrier);
3219 bioset_exit(&conf->bio_split);
3223 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3225 /* no resync is happening, and there is enough space
3226 * on all devices, so we can resize.
3227 * We need to make sure resync covers any new space.
3228 * If the array is shrinking we should possibly wait until
3229 * any io in the removed space completes, but it hardly seems
3232 sector_t newsize = raid1_size(mddev, sectors, 0);
3233 if (mddev->external_size &&
3234 mddev->array_sectors > newsize)
3236 if (mddev->bitmap) {
3237 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3241 md_set_array_sectors(mddev, newsize);
3242 if (sectors > mddev->dev_sectors &&
3243 mddev->recovery_cp > mddev->dev_sectors) {
3244 mddev->recovery_cp = mddev->dev_sectors;
3245 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3247 mddev->dev_sectors = sectors;
3248 mddev->resync_max_sectors = sectors;
3252 static int raid1_reshape(struct mddev *mddev)
3255 * 1/ resize the r1bio_pool
3256 * 2/ resize conf->mirrors
3258 * We allocate a new r1bio_pool if we can.
3259 * Then raise a device barrier and wait until all IO stops.
3260 * Then resize conf->mirrors and swap in the new r1bio pool.
3262 * At the same time, we "pack" the devices so that all the missing
3263 * devices have the higher raid_disk numbers.
3265 mempool_t newpool, oldpool;
3266 struct pool_info *newpoolinfo;
3267 struct raid1_info *newmirrors;
3268 struct r1conf *conf = mddev->private;
3269 int cnt, raid_disks;
3270 unsigned long flags;
3274 memset(&newpool, 0, sizeof(newpool));
3275 memset(&oldpool, 0, sizeof(oldpool));
3277 /* Cannot change chunk_size, layout, or level */
3278 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3279 mddev->layout != mddev->new_layout ||
3280 mddev->level != mddev->new_level) {
3281 mddev->new_chunk_sectors = mddev->chunk_sectors;
3282 mddev->new_layout = mddev->layout;
3283 mddev->new_level = mddev->level;
3287 if (!mddev_is_clustered(mddev))
3288 md_allow_write(mddev);
3290 raid_disks = mddev->raid_disks + mddev->delta_disks;
3292 if (raid_disks < conf->raid_disks) {
3294 for (d= 0; d < conf->raid_disks; d++)
3295 if (conf->mirrors[d].rdev)
3297 if (cnt > raid_disks)
3301 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3304 newpoolinfo->mddev = mddev;
3305 newpoolinfo->raid_disks = raid_disks * 2;
3307 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3308 rbio_pool_free, newpoolinfo);
3313 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3318 mempool_exit(&newpool);
3322 freeze_array(conf, 0);
3324 /* ok, everything is stopped */
3325 oldpool = conf->r1bio_pool;
3326 conf->r1bio_pool = newpool;
3328 for (d = d2 = 0; d < conf->raid_disks; d++) {
3329 struct md_rdev *rdev = conf->mirrors[d].rdev;
3330 if (rdev && rdev->raid_disk != d2) {
3331 sysfs_unlink_rdev(mddev, rdev);
3332 rdev->raid_disk = d2;
3333 sysfs_unlink_rdev(mddev, rdev);
3334 if (sysfs_link_rdev(mddev, rdev))
3335 pr_warn("md/raid1:%s: cannot register rd%d\n",
3336 mdname(mddev), rdev->raid_disk);
3339 newmirrors[d2++].rdev = rdev;
3341 kfree(conf->mirrors);
3342 conf->mirrors = newmirrors;
3343 kfree(conf->poolinfo);
3344 conf->poolinfo = newpoolinfo;
3346 spin_lock_irqsave(&conf->device_lock, flags);
3347 mddev->degraded += (raid_disks - conf->raid_disks);
3348 spin_unlock_irqrestore(&conf->device_lock, flags);
3349 conf->raid_disks = mddev->raid_disks = raid_disks;
3350 mddev->delta_disks = 0;
3352 unfreeze_array(conf);
3354 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3355 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3356 md_wakeup_thread(mddev->thread);
3358 mempool_exit(&oldpool);
3362 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3364 struct r1conf *conf = mddev->private;
3367 freeze_array(conf, 0);
3369 unfreeze_array(conf);
3372 static void *raid1_takeover(struct mddev *mddev)
3374 /* raid1 can take over:
3375 * raid5 with 2 devices, any layout or chunk size
3377 if (mddev->level == 5 && mddev->raid_disks == 2) {
3378 struct r1conf *conf;
3379 mddev->new_level = 1;
3380 mddev->new_layout = 0;
3381 mddev->new_chunk_sectors = 0;
3382 conf = setup_conf(mddev);
3383 if (!IS_ERR(conf)) {
3384 /* Array must appear to be quiesced */
3385 conf->array_frozen = 1;
3386 mddev_clear_unsupported_flags(mddev,
3387 UNSUPPORTED_MDDEV_FLAGS);
3391 return ERR_PTR(-EINVAL);
3394 static struct md_personality raid1_personality =
3398 .owner = THIS_MODULE,
3399 .make_request = raid1_make_request,
3402 .status = raid1_status,
3403 .error_handler = raid1_error,
3404 .hot_add_disk = raid1_add_disk,
3405 .hot_remove_disk= raid1_remove_disk,
3406 .spare_active = raid1_spare_active,
3407 .sync_request = raid1_sync_request,
3408 .resize = raid1_resize,
3410 .check_reshape = raid1_reshape,
3411 .quiesce = raid1_quiesce,
3412 .takeover = raid1_takeover,
3415 static int __init raid_init(void)
3417 return register_md_personality(&raid1_personality);
3420 static void raid_exit(void)
3422 unregister_md_personality(&raid1_personality);
3425 module_init(raid_init);
3426 module_exit(raid_exit);
3427 MODULE_LICENSE("GPL");
3428 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3429 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3430 MODULE_ALIAS("md-raid1");
3431 MODULE_ALIAS("md-level-1");