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(gfp_flags, RESYNC_PAGES);
171 r1_bio->bios[j] = bio;
174 * Allocate RESYNC_PAGES data pages and attach them to
176 * If this is a user-requested check/repair, allocate
177 * RESYNC_PAGES for each bio.
179 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
180 need_pages = pi->raid_disks;
183 for (j = 0; j < pi->raid_disks; j++) {
184 struct resync_pages *rp = &rps[j];
186 bio = r1_bio->bios[j];
188 if (j < need_pages) {
189 if (resync_alloc_pages(rp, gfp_flags))
192 memcpy(rp, &rps[0], sizeof(*rp));
193 resync_get_all_pages(rp);
196 rp->raid_bio = r1_bio;
197 bio->bi_private = rp;
200 r1_bio->master_bio = NULL;
206 resync_free_pages(&rps[j]);
209 while (++j < pi->raid_disks)
210 bio_put(r1_bio->bios[j]);
214 rbio_pool_free(r1_bio, data);
218 static void r1buf_pool_free(void *__r1_bio, void *data)
220 struct pool_info *pi = data;
222 struct r1bio *r1bio = __r1_bio;
223 struct resync_pages *rp = NULL;
225 for (i = pi->raid_disks; i--; ) {
226 rp = get_resync_pages(r1bio->bios[i]);
227 resync_free_pages(rp);
228 bio_put(r1bio->bios[i]);
231 /* resync pages array stored in the 1st bio's .bi_private */
234 rbio_pool_free(r1bio, data);
237 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
241 for (i = 0; i < conf->raid_disks * 2; i++) {
242 struct bio **bio = r1_bio->bios + i;
243 if (!BIO_SPECIAL(*bio))
249 static void free_r1bio(struct r1bio *r1_bio)
251 struct r1conf *conf = r1_bio->mddev->private;
253 put_all_bios(conf, r1_bio);
254 mempool_free(r1_bio, &conf->r1bio_pool);
257 static void put_buf(struct r1bio *r1_bio)
259 struct r1conf *conf = r1_bio->mddev->private;
260 sector_t sect = r1_bio->sector;
263 for (i = 0; i < conf->raid_disks * 2; i++) {
264 struct bio *bio = r1_bio->bios[i];
266 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
269 mempool_free(r1_bio, &conf->r1buf_pool);
271 lower_barrier(conf, sect);
274 static void reschedule_retry(struct r1bio *r1_bio)
277 struct mddev *mddev = r1_bio->mddev;
278 struct r1conf *conf = mddev->private;
281 idx = sector_to_idx(r1_bio->sector);
282 spin_lock_irqsave(&conf->device_lock, flags);
283 list_add(&r1_bio->retry_list, &conf->retry_list);
284 atomic_inc(&conf->nr_queued[idx]);
285 spin_unlock_irqrestore(&conf->device_lock, flags);
287 wake_up(&conf->wait_barrier);
288 md_wakeup_thread(mddev->thread);
292 * raid_end_bio_io() is called when we have finished servicing a mirrored
293 * operation and are ready to return a success/failure code to the buffer
296 static void call_bio_endio(struct r1bio *r1_bio)
298 struct bio *bio = r1_bio->master_bio;
300 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
301 bio->bi_status = BLK_STS_IOERR;
306 static void raid_end_bio_io(struct r1bio *r1_bio)
308 struct bio *bio = r1_bio->master_bio;
309 struct r1conf *conf = r1_bio->mddev->private;
311 /* if nobody has done the final endio yet, do it now */
312 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
313 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
314 (bio_data_dir(bio) == WRITE) ? "write" : "read",
315 (unsigned long long) bio->bi_iter.bi_sector,
316 (unsigned long long) bio_end_sector(bio) - 1);
318 call_bio_endio(r1_bio);
321 * Wake up any possible resync thread that waits for the device
322 * to go idle. All I/Os, even write-behind writes, are done.
324 allow_barrier(conf, r1_bio->sector);
330 * Update disk head position estimator based on IRQ completion info.
332 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
334 struct r1conf *conf = r1_bio->mddev->private;
336 conf->mirrors[disk].head_position =
337 r1_bio->sector + (r1_bio->sectors);
341 * Find the disk number which triggered given bio
343 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
346 struct r1conf *conf = r1_bio->mddev->private;
347 int raid_disks = conf->raid_disks;
349 for (mirror = 0; mirror < raid_disks * 2; mirror++)
350 if (r1_bio->bios[mirror] == bio)
353 BUG_ON(mirror == raid_disks * 2);
354 update_head_pos(mirror, r1_bio);
359 static void raid1_end_read_request(struct bio *bio)
361 int uptodate = !bio->bi_status;
362 struct r1bio *r1_bio = bio->bi_private;
363 struct r1conf *conf = r1_bio->mddev->private;
364 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
367 * this branch is our 'one mirror IO has finished' event handler:
369 update_head_pos(r1_bio->read_disk, r1_bio);
372 set_bit(R1BIO_Uptodate, &r1_bio->state);
373 else if (test_bit(FailFast, &rdev->flags) &&
374 test_bit(R1BIO_FailFast, &r1_bio->state))
375 /* This was a fail-fast read so we definitely
379 /* If all other devices have failed, we want to return
380 * the error upwards rather than fail the last device.
381 * Here we redefine "uptodate" to mean "Don't want to retry"
384 spin_lock_irqsave(&conf->device_lock, flags);
385 if (r1_bio->mddev->degraded == conf->raid_disks ||
386 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
387 test_bit(In_sync, &rdev->flags)))
389 spin_unlock_irqrestore(&conf->device_lock, flags);
393 raid_end_bio_io(r1_bio);
394 rdev_dec_pending(rdev, conf->mddev);
399 char b[BDEVNAME_SIZE];
400 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
402 bdevname(rdev->bdev, b),
403 (unsigned long long)r1_bio->sector);
404 set_bit(R1BIO_ReadError, &r1_bio->state);
405 reschedule_retry(r1_bio);
406 /* don't drop the reference on read_disk yet */
410 static void close_write(struct r1bio *r1_bio)
412 /* it really is the end of this request */
413 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
414 bio_free_pages(r1_bio->behind_master_bio);
415 bio_put(r1_bio->behind_master_bio);
416 r1_bio->behind_master_bio = NULL;
418 /* clear the bitmap if all writes complete successfully */
419 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
421 !test_bit(R1BIO_Degraded, &r1_bio->state),
422 test_bit(R1BIO_BehindIO, &r1_bio->state));
423 md_write_end(r1_bio->mddev);
426 static void r1_bio_write_done(struct r1bio *r1_bio)
428 if (!atomic_dec_and_test(&r1_bio->remaining))
431 if (test_bit(R1BIO_WriteError, &r1_bio->state))
432 reschedule_retry(r1_bio);
435 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
436 reschedule_retry(r1_bio);
438 raid_end_bio_io(r1_bio);
442 static void raid1_end_write_request(struct bio *bio)
444 struct r1bio *r1_bio = bio->bi_private;
445 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
446 struct r1conf *conf = r1_bio->mddev->private;
447 struct bio *to_put = NULL;
448 int mirror = find_bio_disk(r1_bio, bio);
449 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
451 sector_t lo = r1_bio->sector;
452 sector_t hi = r1_bio->sector + r1_bio->sectors;
454 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
457 * 'one mirror IO has finished' event handler:
459 if (bio->bi_status && !discard_error) {
460 set_bit(WriteErrorSeen, &rdev->flags);
461 if (!test_and_set_bit(WantReplacement, &rdev->flags))
462 set_bit(MD_RECOVERY_NEEDED, &
463 conf->mddev->recovery);
465 if (test_bit(FailFast, &rdev->flags) &&
466 (bio->bi_opf & MD_FAILFAST) &&
467 /* We never try FailFast to WriteMostly devices */
468 !test_bit(WriteMostly, &rdev->flags)) {
469 md_error(r1_bio->mddev, rdev);
473 * When the device is faulty, it is not necessary to
474 * handle write error.
475 * For failfast, this is the only remaining device,
476 * We need to retry the write without FailFast.
478 if (!test_bit(Faulty, &rdev->flags))
479 set_bit(R1BIO_WriteError, &r1_bio->state);
481 /* Finished with this branch */
482 r1_bio->bios[mirror] = NULL;
487 * Set R1BIO_Uptodate in our master bio, so that we
488 * will return a good error code for to the higher
489 * levels even if IO on some other mirrored buffer
492 * The 'master' represents the composite IO operation
493 * to user-side. So if something waits for IO, then it
494 * will wait for the 'master' bio.
499 r1_bio->bios[mirror] = NULL;
502 * Do not set R1BIO_Uptodate if the current device is
503 * rebuilding or Faulty. This is because we cannot use
504 * such device for properly reading the data back (we could
505 * potentially use it, if the current write would have felt
506 * before rdev->recovery_offset, but for simplicity we don't
509 if (test_bit(In_sync, &rdev->flags) &&
510 !test_bit(Faulty, &rdev->flags))
511 set_bit(R1BIO_Uptodate, &r1_bio->state);
513 /* Maybe we can clear some bad blocks. */
514 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
515 &first_bad, &bad_sectors) && !discard_error) {
516 r1_bio->bios[mirror] = IO_MADE_GOOD;
517 set_bit(R1BIO_MadeGood, &r1_bio->state);
522 if (test_bit(CollisionCheck, &rdev->flags))
523 remove_serial(rdev, lo, hi);
524 if (test_bit(WriteMostly, &rdev->flags))
525 atomic_dec(&r1_bio->behind_remaining);
528 * In behind mode, we ACK the master bio once the I/O
529 * has safely reached all non-writemostly
530 * disks. Setting the Returned bit ensures that this
531 * gets done only once -- we don't ever want to return
532 * -EIO here, instead we'll wait
534 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
535 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
536 /* Maybe we can return now */
537 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
538 struct bio *mbio = r1_bio->master_bio;
539 pr_debug("raid1: behind end write sectors"
541 (unsigned long long) mbio->bi_iter.bi_sector,
542 (unsigned long long) bio_end_sector(mbio) - 1);
543 call_bio_endio(r1_bio);
546 } else if (rdev->mddev->serialize_policy)
547 remove_serial(rdev, lo, hi);
548 if (r1_bio->bios[mirror] == NULL)
549 rdev_dec_pending(rdev, conf->mddev);
552 * Let's see if all mirrored write operations have finished
555 r1_bio_write_done(r1_bio);
561 static sector_t align_to_barrier_unit_end(sector_t start_sector,
566 WARN_ON(sectors == 0);
568 * len is the number of sectors from start_sector to end of the
569 * barrier unit which start_sector belongs to.
571 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
581 * This routine returns the disk from which the requested read should
582 * be done. There is a per-array 'next expected sequential IO' sector
583 * number - if this matches on the next IO then we use the last disk.
584 * There is also a per-disk 'last know head position' sector that is
585 * maintained from IRQ contexts, both the normal and the resync IO
586 * completion handlers update this position correctly. If there is no
587 * perfect sequential match then we pick the disk whose head is closest.
589 * If there are 2 mirrors in the same 2 devices, performance degrades
590 * because position is mirror, not device based.
592 * The rdev for the device selected will have nr_pending incremented.
594 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
596 const sector_t this_sector = r1_bio->sector;
598 int best_good_sectors;
599 int best_disk, best_dist_disk, best_pending_disk;
603 unsigned int min_pending;
604 struct md_rdev *rdev;
606 int choose_next_idle;
610 * Check if we can balance. We can balance on the whole
611 * device if no resync is going on, or below the resync window.
612 * We take the first readable disk when above the resync window.
615 sectors = r1_bio->sectors;
618 best_dist = MaxSector;
619 best_pending_disk = -1;
620 min_pending = UINT_MAX;
621 best_good_sectors = 0;
623 choose_next_idle = 0;
624 clear_bit(R1BIO_FailFast, &r1_bio->state);
626 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
627 (mddev_is_clustered(conf->mddev) &&
628 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
629 this_sector + sectors)))
634 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
638 unsigned int pending;
641 rdev = rcu_dereference(conf->mirrors[disk].rdev);
642 if (r1_bio->bios[disk] == IO_BLOCKED
644 || test_bit(Faulty, &rdev->flags))
646 if (!test_bit(In_sync, &rdev->flags) &&
647 rdev->recovery_offset < this_sector + sectors)
649 if (test_bit(WriteMostly, &rdev->flags)) {
650 /* Don't balance among write-mostly, just
651 * use the first as a last resort */
652 if (best_dist_disk < 0) {
653 if (is_badblock(rdev, this_sector, sectors,
654 &first_bad, &bad_sectors)) {
655 if (first_bad <= this_sector)
656 /* Cannot use this */
658 best_good_sectors = first_bad - this_sector;
660 best_good_sectors = sectors;
661 best_dist_disk = disk;
662 best_pending_disk = disk;
666 /* This is a reasonable device to use. It might
669 if (is_badblock(rdev, this_sector, sectors,
670 &first_bad, &bad_sectors)) {
671 if (best_dist < MaxSector)
672 /* already have a better device */
674 if (first_bad <= this_sector) {
675 /* cannot read here. If this is the 'primary'
676 * device, then we must not read beyond
677 * bad_sectors from another device..
679 bad_sectors -= (this_sector - first_bad);
680 if (choose_first && sectors > bad_sectors)
681 sectors = bad_sectors;
682 if (best_good_sectors > sectors)
683 best_good_sectors = sectors;
686 sector_t good_sectors = first_bad - this_sector;
687 if (good_sectors > best_good_sectors) {
688 best_good_sectors = good_sectors;
696 if ((sectors > best_good_sectors) && (best_disk >= 0))
698 best_good_sectors = sectors;
702 /* At least two disks to choose from so failfast is OK */
703 set_bit(R1BIO_FailFast, &r1_bio->state);
705 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
706 has_nonrot_disk |= nonrot;
707 pending = atomic_read(&rdev->nr_pending);
708 dist = abs(this_sector - conf->mirrors[disk].head_position);
713 /* Don't change to another disk for sequential reads */
714 if (conf->mirrors[disk].next_seq_sect == this_sector
716 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
717 struct raid1_info *mirror = &conf->mirrors[disk];
721 * If buffered sequential IO size exceeds optimal
722 * iosize, check if there is idle disk. If yes, choose
723 * the idle disk. read_balance could already choose an
724 * idle disk before noticing it's a sequential IO in
725 * this disk. This doesn't matter because this disk
726 * will idle, next time it will be utilized after the
727 * first disk has IO size exceeds optimal iosize. In
728 * this way, iosize of the first disk will be optimal
729 * iosize at least. iosize of the second disk might be
730 * small, but not a big deal since when the second disk
731 * starts IO, the first disk is likely still busy.
733 if (nonrot && opt_iosize > 0 &&
734 mirror->seq_start != MaxSector &&
735 mirror->next_seq_sect > opt_iosize &&
736 mirror->next_seq_sect - opt_iosize >=
738 choose_next_idle = 1;
744 if (choose_next_idle)
747 if (min_pending > pending) {
748 min_pending = pending;
749 best_pending_disk = disk;
752 if (dist < best_dist) {
754 best_dist_disk = disk;
759 * If all disks are rotational, choose the closest disk. If any disk is
760 * non-rotational, choose the disk with less pending request even the
761 * disk is rotational, which might/might not be optimal for raids with
762 * mixed ratation/non-rotational disks depending on workload.
764 if (best_disk == -1) {
765 if (has_nonrot_disk || min_pending == 0)
766 best_disk = best_pending_disk;
768 best_disk = best_dist_disk;
771 if (best_disk >= 0) {
772 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
775 atomic_inc(&rdev->nr_pending);
776 sectors = best_good_sectors;
778 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
779 conf->mirrors[best_disk].seq_start = this_sector;
781 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
784 *max_sectors = sectors;
789 static int raid1_congested(struct mddev *mddev, int bits)
791 struct r1conf *conf = mddev->private;
794 if ((bits & (1 << WB_async_congested)) &&
795 conf->pending_count >= max_queued_requests)
799 for (i = 0; i < conf->raid_disks * 2; i++) {
800 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
801 if (rdev && !test_bit(Faulty, &rdev->flags)) {
802 struct request_queue *q = bdev_get_queue(rdev->bdev);
806 /* Note the '|| 1' - when read_balance prefers
807 * non-congested targets, it can be removed
809 if ((bits & (1 << WB_async_congested)) || 1)
810 ret |= bdi_congested(q->backing_dev_info, bits);
812 ret &= bdi_congested(q->backing_dev_info, bits);
819 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
821 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
822 md_bitmap_unplug(conf->mddev->bitmap);
823 wake_up(&conf->wait_barrier);
825 while (bio) { /* submit pending writes */
826 struct bio *next = bio->bi_next;
827 struct md_rdev *rdev = (void *)bio->bi_disk;
829 bio_set_dev(bio, rdev->bdev);
830 if (test_bit(Faulty, &rdev->flags)) {
832 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
833 !blk_queue_discard(bio->bi_disk->queue)))
837 submit_bio_noacct(bio);
843 static void flush_pending_writes(struct r1conf *conf)
845 /* Any writes that have been queued but are awaiting
846 * bitmap updates get flushed here.
848 spin_lock_irq(&conf->device_lock);
850 if (conf->pending_bio_list.head) {
851 struct blk_plug plug;
854 bio = bio_list_get(&conf->pending_bio_list);
855 conf->pending_count = 0;
856 spin_unlock_irq(&conf->device_lock);
859 * As this is called in a wait_event() loop (see freeze_array),
860 * current->state might be TASK_UNINTERRUPTIBLE which will
861 * cause a warning when we prepare to wait again. As it is
862 * rare that this path is taken, it is perfectly safe to force
863 * us to go around the wait_event() loop again, so the warning
864 * is a false-positive. Silence the warning by resetting
867 __set_current_state(TASK_RUNNING);
868 blk_start_plug(&plug);
869 flush_bio_list(conf, bio);
870 blk_finish_plug(&plug);
872 spin_unlock_irq(&conf->device_lock);
876 * Sometimes we need to suspend IO while we do something else,
877 * either some resync/recovery, or reconfigure the array.
878 * To do this we raise a 'barrier'.
879 * The 'barrier' is a counter that can be raised multiple times
880 * to count how many activities are happening which preclude
882 * We can only raise the barrier if there is no pending IO.
883 * i.e. if nr_pending == 0.
884 * We choose only to raise the barrier if no-one is waiting for the
885 * barrier to go down. This means that as soon as an IO request
886 * is ready, no other operations which require a barrier will start
887 * until the IO request has had a chance.
889 * So: regular IO calls 'wait_barrier'. When that returns there
890 * is no backgroup IO happening, It must arrange to call
891 * allow_barrier when it has finished its IO.
892 * backgroup IO calls must call raise_barrier. Once that returns
893 * there is no normal IO happeing. It must arrange to call
894 * lower_barrier when the particular background IO completes.
896 * If resync/recovery is interrupted, returns -EINTR;
897 * Otherwise, returns 0.
899 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
901 int idx = sector_to_idx(sector_nr);
903 spin_lock_irq(&conf->resync_lock);
905 /* Wait until no block IO is waiting */
906 wait_event_lock_irq(conf->wait_barrier,
907 !atomic_read(&conf->nr_waiting[idx]),
910 /* block any new IO from starting */
911 atomic_inc(&conf->barrier[idx]);
913 * In raise_barrier() we firstly increase conf->barrier[idx] then
914 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
915 * increase conf->nr_pending[idx] then check conf->barrier[idx].
916 * A memory barrier here to make sure conf->nr_pending[idx] won't
917 * be fetched before conf->barrier[idx] is increased. Otherwise
918 * there will be a race between raise_barrier() and _wait_barrier().
920 smp_mb__after_atomic();
922 /* For these conditions we must wait:
923 * A: while the array is in frozen state
924 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
925 * existing in corresponding I/O barrier bucket.
926 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
927 * max resync count which allowed on current I/O barrier bucket.
929 wait_event_lock_irq(conf->wait_barrier,
930 (!conf->array_frozen &&
931 !atomic_read(&conf->nr_pending[idx]) &&
932 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
933 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
936 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
937 atomic_dec(&conf->barrier[idx]);
938 spin_unlock_irq(&conf->resync_lock);
939 wake_up(&conf->wait_barrier);
943 atomic_inc(&conf->nr_sync_pending);
944 spin_unlock_irq(&conf->resync_lock);
949 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
951 int idx = sector_to_idx(sector_nr);
953 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
955 atomic_dec(&conf->barrier[idx]);
956 atomic_dec(&conf->nr_sync_pending);
957 wake_up(&conf->wait_barrier);
960 static void _wait_barrier(struct r1conf *conf, int idx)
963 * We need to increase conf->nr_pending[idx] very early here,
964 * then raise_barrier() can be blocked when it waits for
965 * conf->nr_pending[idx] to be 0. Then we can avoid holding
966 * conf->resync_lock when there is no barrier raised in same
967 * barrier unit bucket. Also if the array is frozen, I/O
968 * should be blocked until array is unfrozen.
970 atomic_inc(&conf->nr_pending[idx]);
972 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
973 * check conf->barrier[idx]. In raise_barrier() we firstly increase
974 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
975 * barrier is necessary here to make sure conf->barrier[idx] won't be
976 * fetched before conf->nr_pending[idx] is increased. Otherwise there
977 * will be a race between _wait_barrier() and raise_barrier().
979 smp_mb__after_atomic();
982 * Don't worry about checking two atomic_t variables at same time
983 * here. If during we check conf->barrier[idx], the array is
984 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
985 * 0, it is safe to return and make the I/O continue. Because the
986 * array is frozen, all I/O returned here will eventually complete
987 * or be queued, no race will happen. See code comment in
990 if (!READ_ONCE(conf->array_frozen) &&
991 !atomic_read(&conf->barrier[idx]))
995 * After holding conf->resync_lock, conf->nr_pending[idx]
996 * should be decreased before waiting for barrier to drop.
997 * Otherwise, we may encounter a race condition because
998 * raise_barrer() might be waiting for conf->nr_pending[idx]
999 * to be 0 at same time.
1001 spin_lock_irq(&conf->resync_lock);
1002 atomic_inc(&conf->nr_waiting[idx]);
1003 atomic_dec(&conf->nr_pending[idx]);
1005 * In case freeze_array() is waiting for
1006 * get_unqueued_pending() == extra
1008 wake_up(&conf->wait_barrier);
1009 /* Wait for the barrier in same barrier unit bucket to drop. */
1010 wait_event_lock_irq(conf->wait_barrier,
1011 !conf->array_frozen &&
1012 !atomic_read(&conf->barrier[idx]),
1014 atomic_inc(&conf->nr_pending[idx]);
1015 atomic_dec(&conf->nr_waiting[idx]);
1016 spin_unlock_irq(&conf->resync_lock);
1019 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
1021 int idx = sector_to_idx(sector_nr);
1024 * Very similar to _wait_barrier(). The difference is, for read
1025 * I/O we don't need wait for sync I/O, but if the whole array
1026 * is frozen, the read I/O still has to wait until the array is
1027 * unfrozen. Since there is no ordering requirement with
1028 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1030 atomic_inc(&conf->nr_pending[idx]);
1032 if (!READ_ONCE(conf->array_frozen))
1035 spin_lock_irq(&conf->resync_lock);
1036 atomic_inc(&conf->nr_waiting[idx]);
1037 atomic_dec(&conf->nr_pending[idx]);
1039 * In case freeze_array() is waiting for
1040 * get_unqueued_pending() == extra
1042 wake_up(&conf->wait_barrier);
1043 /* Wait for array to be unfrozen */
1044 wait_event_lock_irq(conf->wait_barrier,
1045 !conf->array_frozen,
1047 atomic_inc(&conf->nr_pending[idx]);
1048 atomic_dec(&conf->nr_waiting[idx]);
1049 spin_unlock_irq(&conf->resync_lock);
1052 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1054 int idx = sector_to_idx(sector_nr);
1056 _wait_barrier(conf, idx);
1059 static void _allow_barrier(struct r1conf *conf, int idx)
1061 atomic_dec(&conf->nr_pending[idx]);
1062 wake_up(&conf->wait_barrier);
1065 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1067 int idx = sector_to_idx(sector_nr);
1069 _allow_barrier(conf, idx);
1072 /* conf->resync_lock should be held */
1073 static int get_unqueued_pending(struct r1conf *conf)
1077 ret = atomic_read(&conf->nr_sync_pending);
1078 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1079 ret += atomic_read(&conf->nr_pending[idx]) -
1080 atomic_read(&conf->nr_queued[idx]);
1085 static void freeze_array(struct r1conf *conf, int extra)
1087 /* Stop sync I/O and normal I/O and wait for everything to
1089 * This is called in two situations:
1090 * 1) management command handlers (reshape, remove disk, quiesce).
1091 * 2) one normal I/O request failed.
1093 * After array_frozen is set to 1, new sync IO will be blocked at
1094 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1095 * or wait_read_barrier(). The flying I/Os will either complete or be
1096 * queued. When everything goes quite, there are only queued I/Os left.
1098 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1099 * barrier bucket index which this I/O request hits. When all sync and
1100 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1101 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1102 * in handle_read_error(), we may call freeze_array() before trying to
1103 * fix the read error. In this case, the error read I/O is not queued,
1104 * so get_unqueued_pending() == 1.
1106 * Therefore before this function returns, we need to wait until
1107 * get_unqueued_pendings(conf) gets equal to extra. For
1108 * normal I/O context, extra is 1, in rested situations extra is 0.
1110 spin_lock_irq(&conf->resync_lock);
1111 conf->array_frozen = 1;
1112 raid1_log(conf->mddev, "wait freeze");
1113 wait_event_lock_irq_cmd(
1115 get_unqueued_pending(conf) == extra,
1117 flush_pending_writes(conf));
1118 spin_unlock_irq(&conf->resync_lock);
1120 static void unfreeze_array(struct r1conf *conf)
1122 /* reverse the effect of the freeze */
1123 spin_lock_irq(&conf->resync_lock);
1124 conf->array_frozen = 0;
1125 spin_unlock_irq(&conf->resync_lock);
1126 wake_up(&conf->wait_barrier);
1129 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1132 int size = bio->bi_iter.bi_size;
1133 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1135 struct bio *behind_bio = NULL;
1137 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1141 /* discard op, we don't support writezero/writesame yet */
1142 if (!bio_has_data(bio)) {
1143 behind_bio->bi_iter.bi_size = size;
1147 behind_bio->bi_write_hint = bio->bi_write_hint;
1149 while (i < vcnt && size) {
1151 int len = min_t(int, PAGE_SIZE, size);
1153 page = alloc_page(GFP_NOIO);
1154 if (unlikely(!page))
1157 bio_add_page(behind_bio, page, len, 0);
1163 bio_copy_data(behind_bio, bio);
1165 r1_bio->behind_master_bio = behind_bio;
1166 set_bit(R1BIO_BehindIO, &r1_bio->state);
1171 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1172 bio->bi_iter.bi_size);
1173 bio_free_pages(behind_bio);
1174 bio_put(behind_bio);
1177 struct raid1_plug_cb {
1178 struct blk_plug_cb cb;
1179 struct bio_list pending;
1183 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1185 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1187 struct mddev *mddev = plug->cb.data;
1188 struct r1conf *conf = mddev->private;
1191 if (from_schedule || current->bio_list) {
1192 spin_lock_irq(&conf->device_lock);
1193 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1194 conf->pending_count += plug->pending_cnt;
1195 spin_unlock_irq(&conf->device_lock);
1196 wake_up(&conf->wait_barrier);
1197 md_wakeup_thread(mddev->thread);
1202 /* we aren't scheduling, so we can do the write-out directly. */
1203 bio = bio_list_get(&plug->pending);
1204 flush_bio_list(conf, bio);
1208 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1210 r1_bio->master_bio = bio;
1211 r1_bio->sectors = bio_sectors(bio);
1213 r1_bio->mddev = mddev;
1214 r1_bio->sector = bio->bi_iter.bi_sector;
1217 static inline struct r1bio *
1218 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1220 struct r1conf *conf = mddev->private;
1221 struct r1bio *r1_bio;
1223 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1224 /* Ensure no bio records IO_BLOCKED */
1225 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1226 init_r1bio(r1_bio, mddev, bio);
1230 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1231 int max_read_sectors, struct r1bio *r1_bio)
1233 struct r1conf *conf = mddev->private;
1234 struct raid1_info *mirror;
1235 struct bio *read_bio;
1236 struct bitmap *bitmap = mddev->bitmap;
1237 const int op = bio_op(bio);
1238 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1241 bool print_msg = !!r1_bio;
1242 char b[BDEVNAME_SIZE];
1245 * If r1_bio is set, we are blocking the raid1d thread
1246 * so there is a tiny risk of deadlock. So ask for
1247 * emergency memory if needed.
1249 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1252 /* Need to get the block device name carefully */
1253 struct md_rdev *rdev;
1255 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1257 bdevname(rdev->bdev, b);
1264 * Still need barrier for READ in case that whole
1267 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1270 r1_bio = alloc_r1bio(mddev, bio);
1272 init_r1bio(r1_bio, mddev, bio);
1273 r1_bio->sectors = max_read_sectors;
1276 * make_request() can abort the operation when read-ahead is being
1277 * used and no empty request is available.
1279 rdisk = read_balance(conf, r1_bio, &max_sectors);
1282 /* couldn't find anywhere to read from */
1284 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1287 (unsigned long long)r1_bio->sector);
1289 raid_end_bio_io(r1_bio);
1292 mirror = conf->mirrors + rdisk;
1295 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1297 (unsigned long long)r1_bio->sector,
1298 bdevname(mirror->rdev->bdev, b));
1300 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1303 * Reading from a write-mostly device must take care not to
1304 * over-take any writes that are 'behind'
1306 raid1_log(mddev, "wait behind writes");
1307 wait_event(bitmap->behind_wait,
1308 atomic_read(&bitmap->behind_writes) == 0);
1311 if (max_sectors < bio_sectors(bio)) {
1312 struct bio *split = bio_split(bio, max_sectors,
1313 gfp, &conf->bio_split);
1314 bio_chain(split, bio);
1315 submit_bio_noacct(bio);
1317 r1_bio->master_bio = bio;
1318 r1_bio->sectors = max_sectors;
1321 r1_bio->read_disk = rdisk;
1323 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1325 r1_bio->bios[rdisk] = read_bio;
1327 read_bio->bi_iter.bi_sector = r1_bio->sector +
1328 mirror->rdev->data_offset;
1329 bio_set_dev(read_bio, mirror->rdev->bdev);
1330 read_bio->bi_end_io = raid1_end_read_request;
1331 bio_set_op_attrs(read_bio, op, do_sync);
1332 if (test_bit(FailFast, &mirror->rdev->flags) &&
1333 test_bit(R1BIO_FailFast, &r1_bio->state))
1334 read_bio->bi_opf |= MD_FAILFAST;
1335 read_bio->bi_private = r1_bio;
1338 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1339 disk_devt(mddev->gendisk), r1_bio->sector);
1341 submit_bio_noacct(read_bio);
1344 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1345 int max_write_sectors)
1347 struct r1conf *conf = mddev->private;
1348 struct r1bio *r1_bio;
1350 struct bitmap *bitmap = mddev->bitmap;
1351 unsigned long flags;
1352 struct md_rdev *blocked_rdev;
1353 struct blk_plug_cb *cb;
1354 struct raid1_plug_cb *plug = NULL;
1358 if (mddev_is_clustered(mddev) &&
1359 md_cluster_ops->area_resyncing(mddev, WRITE,
1360 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1364 prepare_to_wait(&conf->wait_barrier,
1366 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1367 bio->bi_iter.bi_sector,
1368 bio_end_sector(bio)))
1372 finish_wait(&conf->wait_barrier, &w);
1376 * Register the new request and wait if the reconstruction
1377 * thread has put up a bar for new requests.
1378 * Continue immediately if no resync is active currently.
1380 wait_barrier(conf, bio->bi_iter.bi_sector);
1382 r1_bio = alloc_r1bio(mddev, bio);
1383 r1_bio->sectors = max_write_sectors;
1385 if (conf->pending_count >= max_queued_requests) {
1386 md_wakeup_thread(mddev->thread);
1387 raid1_log(mddev, "wait queued");
1388 wait_event(conf->wait_barrier,
1389 conf->pending_count < max_queued_requests);
1391 /* first select target devices under rcu_lock and
1392 * inc refcount on their rdev. Record them by setting
1394 * If there are known/acknowledged bad blocks on any device on
1395 * which we have seen a write error, we want to avoid writing those
1397 * This potentially requires several writes to write around
1398 * the bad blocks. Each set of writes gets it's own r1bio
1399 * with a set of bios attached.
1402 disks = conf->raid_disks * 2;
1404 blocked_rdev = NULL;
1406 max_sectors = r1_bio->sectors;
1407 for (i = 0; i < disks; i++) {
1408 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1409 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1410 atomic_inc(&rdev->nr_pending);
1411 blocked_rdev = rdev;
1414 r1_bio->bios[i] = NULL;
1415 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1416 if (i < conf->raid_disks)
1417 set_bit(R1BIO_Degraded, &r1_bio->state);
1421 atomic_inc(&rdev->nr_pending);
1422 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1427 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1428 &first_bad, &bad_sectors);
1430 /* mustn't write here until the bad block is
1432 set_bit(BlockedBadBlocks, &rdev->flags);
1433 blocked_rdev = rdev;
1436 if (is_bad && first_bad <= r1_bio->sector) {
1437 /* Cannot write here at all */
1438 bad_sectors -= (r1_bio->sector - first_bad);
1439 if (bad_sectors < max_sectors)
1440 /* mustn't write more than bad_sectors
1441 * to other devices yet
1443 max_sectors = bad_sectors;
1444 rdev_dec_pending(rdev, mddev);
1445 /* We don't set R1BIO_Degraded as that
1446 * only applies if the disk is
1447 * missing, so it might be re-added,
1448 * and we want to know to recover this
1450 * In this case the device is here,
1451 * and the fact that this chunk is not
1452 * in-sync is recorded in the bad
1458 int good_sectors = first_bad - r1_bio->sector;
1459 if (good_sectors < max_sectors)
1460 max_sectors = good_sectors;
1463 r1_bio->bios[i] = bio;
1467 if (unlikely(blocked_rdev)) {
1468 /* Wait for this device to become unblocked */
1471 for (j = 0; j < i; j++)
1472 if (r1_bio->bios[j])
1473 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1475 allow_barrier(conf, bio->bi_iter.bi_sector);
1476 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1477 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1478 wait_barrier(conf, bio->bi_iter.bi_sector);
1482 if (max_sectors < bio_sectors(bio)) {
1483 struct bio *split = bio_split(bio, max_sectors,
1484 GFP_NOIO, &conf->bio_split);
1485 bio_chain(split, bio);
1486 submit_bio_noacct(bio);
1488 r1_bio->master_bio = bio;
1489 r1_bio->sectors = max_sectors;
1492 atomic_set(&r1_bio->remaining, 1);
1493 atomic_set(&r1_bio->behind_remaining, 0);
1497 for (i = 0; i < disks; i++) {
1498 struct bio *mbio = NULL;
1499 struct md_rdev *rdev = conf->mirrors[i].rdev;
1500 if (!r1_bio->bios[i])
1505 * Not if there are too many, or cannot
1506 * allocate memory, or a reader on WriteMostly
1507 * is waiting for behind writes to flush */
1509 (atomic_read(&bitmap->behind_writes)
1510 < mddev->bitmap_info.max_write_behind) &&
1511 !waitqueue_active(&bitmap->behind_wait)) {
1512 alloc_behind_master_bio(r1_bio, bio);
1515 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1516 test_bit(R1BIO_BehindIO, &r1_bio->state));
1520 if (r1_bio->behind_master_bio)
1521 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1522 GFP_NOIO, &mddev->bio_set);
1524 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1526 if (r1_bio->behind_master_bio) {
1527 if (test_bit(CollisionCheck, &rdev->flags))
1528 wait_for_serialization(rdev, r1_bio);
1529 if (test_bit(WriteMostly, &rdev->flags))
1530 atomic_inc(&r1_bio->behind_remaining);
1531 } else if (mddev->serialize_policy)
1532 wait_for_serialization(rdev, r1_bio);
1534 r1_bio->bios[i] = mbio;
1536 mbio->bi_iter.bi_sector = (r1_bio->sector +
1537 conf->mirrors[i].rdev->data_offset);
1538 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1539 mbio->bi_end_io = raid1_end_write_request;
1540 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1541 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1542 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1543 conf->raid_disks - mddev->degraded > 1)
1544 mbio->bi_opf |= MD_FAILFAST;
1545 mbio->bi_private = r1_bio;
1547 atomic_inc(&r1_bio->remaining);
1550 trace_block_bio_remap(mbio->bi_disk->queue,
1551 mbio, disk_devt(mddev->gendisk),
1553 /* flush_pending_writes() needs access to the rdev so...*/
1554 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1556 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1558 plug = container_of(cb, struct raid1_plug_cb, cb);
1562 bio_list_add(&plug->pending, mbio);
1563 plug->pending_cnt++;
1565 spin_lock_irqsave(&conf->device_lock, flags);
1566 bio_list_add(&conf->pending_bio_list, mbio);
1567 conf->pending_count++;
1568 spin_unlock_irqrestore(&conf->device_lock, flags);
1569 md_wakeup_thread(mddev->thread);
1573 r1_bio_write_done(r1_bio);
1575 /* In case raid1d snuck in to freeze_array */
1576 wake_up(&conf->wait_barrier);
1579 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1583 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1584 && md_flush_request(mddev, bio))
1588 * There is a limit to the maximum size, but
1589 * the read/write handler might find a lower limit
1590 * due to bad blocks. To avoid multiple splits,
1591 * we pass the maximum number of sectors down
1592 * and let the lower level perform the split.
1594 sectors = align_to_barrier_unit_end(
1595 bio->bi_iter.bi_sector, bio_sectors(bio));
1597 if (bio_data_dir(bio) == READ)
1598 raid1_read_request(mddev, bio, sectors, NULL);
1600 if (!md_write_start(mddev,bio))
1602 raid1_write_request(mddev, bio, sectors);
1607 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1609 struct r1conf *conf = mddev->private;
1612 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1613 conf->raid_disks - mddev->degraded);
1615 for (i = 0; i < conf->raid_disks; i++) {
1616 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1617 seq_printf(seq, "%s",
1618 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1621 seq_printf(seq, "]");
1624 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1626 char b[BDEVNAME_SIZE];
1627 struct r1conf *conf = mddev->private;
1628 unsigned long flags;
1631 * If it is not operational, then we have already marked it as dead
1632 * else if it is the last working disks with "fail_last_dev == false",
1633 * ignore the error, let the next level up know.
1634 * else mark the drive as failed
1636 spin_lock_irqsave(&conf->device_lock, flags);
1637 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1638 && (conf->raid_disks - mddev->degraded) == 1) {
1640 * Don't fail the drive, act as though we were just a
1641 * normal single drive.
1642 * However don't try a recovery from this drive as
1643 * it is very likely to fail.
1645 conf->recovery_disabled = mddev->recovery_disabled;
1646 spin_unlock_irqrestore(&conf->device_lock, flags);
1649 set_bit(Blocked, &rdev->flags);
1650 if (test_and_clear_bit(In_sync, &rdev->flags))
1652 set_bit(Faulty, &rdev->flags);
1653 spin_unlock_irqrestore(&conf->device_lock, flags);
1655 * if recovery is running, make sure it aborts.
1657 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1658 set_mask_bits(&mddev->sb_flags, 0,
1659 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1660 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1661 "md/raid1:%s: Operation continuing on %d devices.\n",
1662 mdname(mddev), bdevname(rdev->bdev, b),
1663 mdname(mddev), conf->raid_disks - mddev->degraded);
1666 static void print_conf(struct r1conf *conf)
1670 pr_debug("RAID1 conf printout:\n");
1672 pr_debug("(!conf)\n");
1675 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1679 for (i = 0; i < conf->raid_disks; i++) {
1680 char b[BDEVNAME_SIZE];
1681 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1683 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1684 i, !test_bit(In_sync, &rdev->flags),
1685 !test_bit(Faulty, &rdev->flags),
1686 bdevname(rdev->bdev,b));
1691 static void close_sync(struct r1conf *conf)
1695 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1696 _wait_barrier(conf, idx);
1697 _allow_barrier(conf, idx);
1700 mempool_exit(&conf->r1buf_pool);
1703 static int raid1_spare_active(struct mddev *mddev)
1706 struct r1conf *conf = mddev->private;
1708 unsigned long flags;
1711 * Find all failed disks within the RAID1 configuration
1712 * and mark them readable.
1713 * Called under mddev lock, so rcu protection not needed.
1714 * device_lock used to avoid races with raid1_end_read_request
1715 * which expects 'In_sync' flags and ->degraded to be consistent.
1717 spin_lock_irqsave(&conf->device_lock, flags);
1718 for (i = 0; i < conf->raid_disks; i++) {
1719 struct md_rdev *rdev = conf->mirrors[i].rdev;
1720 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1722 && !test_bit(Candidate, &repl->flags)
1723 && repl->recovery_offset == MaxSector
1724 && !test_bit(Faulty, &repl->flags)
1725 && !test_and_set_bit(In_sync, &repl->flags)) {
1726 /* replacement has just become active */
1728 !test_and_clear_bit(In_sync, &rdev->flags))
1731 /* Replaced device not technically
1732 * faulty, but we need to be sure
1733 * it gets removed and never re-added
1735 set_bit(Faulty, &rdev->flags);
1736 sysfs_notify_dirent_safe(
1741 && rdev->recovery_offset == MaxSector
1742 && !test_bit(Faulty, &rdev->flags)
1743 && !test_and_set_bit(In_sync, &rdev->flags)) {
1745 sysfs_notify_dirent_safe(rdev->sysfs_state);
1748 mddev->degraded -= count;
1749 spin_unlock_irqrestore(&conf->device_lock, flags);
1755 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1757 struct r1conf *conf = mddev->private;
1760 struct raid1_info *p;
1762 int last = conf->raid_disks - 1;
1764 if (mddev->recovery_disabled == conf->recovery_disabled)
1767 if (md_integrity_add_rdev(rdev, mddev))
1770 if (rdev->raid_disk >= 0)
1771 first = last = rdev->raid_disk;
1774 * find the disk ... but prefer rdev->saved_raid_disk
1777 if (rdev->saved_raid_disk >= 0 &&
1778 rdev->saved_raid_disk >= first &&
1779 rdev->saved_raid_disk < conf->raid_disks &&
1780 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1781 first = last = rdev->saved_raid_disk;
1783 for (mirror = first; mirror <= last; mirror++) {
1784 p = conf->mirrors + mirror;
1787 disk_stack_limits(mddev->gendisk, rdev->bdev,
1788 rdev->data_offset << 9);
1790 p->head_position = 0;
1791 rdev->raid_disk = mirror;
1793 /* As all devices are equivalent, we don't need a full recovery
1794 * if this was recently any drive of the array
1796 if (rdev->saved_raid_disk < 0)
1798 rcu_assign_pointer(p->rdev, rdev);
1801 if (test_bit(WantReplacement, &p->rdev->flags) &&
1802 p[conf->raid_disks].rdev == NULL) {
1803 /* Add this device as a replacement */
1804 clear_bit(In_sync, &rdev->flags);
1805 set_bit(Replacement, &rdev->flags);
1806 rdev->raid_disk = mirror;
1809 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1813 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1814 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1819 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1821 struct r1conf *conf = mddev->private;
1823 int number = rdev->raid_disk;
1824 struct raid1_info *p = conf->mirrors + number;
1826 if (rdev != p->rdev)
1827 p = conf->mirrors + conf->raid_disks + number;
1830 if (rdev == p->rdev) {
1831 if (test_bit(In_sync, &rdev->flags) ||
1832 atomic_read(&rdev->nr_pending)) {
1836 /* Only remove non-faulty devices if recovery
1839 if (!test_bit(Faulty, &rdev->flags) &&
1840 mddev->recovery_disabled != conf->recovery_disabled &&
1841 mddev->degraded < conf->raid_disks) {
1846 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1848 if (atomic_read(&rdev->nr_pending)) {
1849 /* lost the race, try later */
1855 if (conf->mirrors[conf->raid_disks + number].rdev) {
1856 /* We just removed a device that is being replaced.
1857 * Move down the replacement. We drain all IO before
1858 * doing this to avoid confusion.
1860 struct md_rdev *repl =
1861 conf->mirrors[conf->raid_disks + number].rdev;
1862 freeze_array(conf, 0);
1863 if (atomic_read(&repl->nr_pending)) {
1864 /* It means that some queued IO of retry_list
1865 * hold repl. Thus, we cannot set replacement
1866 * as NULL, avoiding rdev NULL pointer
1867 * dereference in sync_request_write and
1868 * handle_write_finished.
1871 unfreeze_array(conf);
1874 clear_bit(Replacement, &repl->flags);
1876 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1877 unfreeze_array(conf);
1880 clear_bit(WantReplacement, &rdev->flags);
1881 err = md_integrity_register(mddev);
1889 static void end_sync_read(struct bio *bio)
1891 struct r1bio *r1_bio = get_resync_r1bio(bio);
1893 update_head_pos(r1_bio->read_disk, r1_bio);
1896 * we have read a block, now it needs to be re-written,
1897 * or re-read if the read failed.
1898 * We don't do much here, just schedule handling by raid1d
1900 if (!bio->bi_status)
1901 set_bit(R1BIO_Uptodate, &r1_bio->state);
1903 if (atomic_dec_and_test(&r1_bio->remaining))
1904 reschedule_retry(r1_bio);
1907 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1909 sector_t sync_blocks = 0;
1910 sector_t s = r1_bio->sector;
1911 long sectors_to_go = r1_bio->sectors;
1913 /* make sure these bits don't get cleared. */
1915 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1917 sectors_to_go -= sync_blocks;
1918 } while (sectors_to_go > 0);
1921 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1923 if (atomic_dec_and_test(&r1_bio->remaining)) {
1924 struct mddev *mddev = r1_bio->mddev;
1925 int s = r1_bio->sectors;
1927 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1928 test_bit(R1BIO_WriteError, &r1_bio->state))
1929 reschedule_retry(r1_bio);
1932 md_done_sync(mddev, s, uptodate);
1937 static void end_sync_write(struct bio *bio)
1939 int uptodate = !bio->bi_status;
1940 struct r1bio *r1_bio = get_resync_r1bio(bio);
1941 struct mddev *mddev = r1_bio->mddev;
1942 struct r1conf *conf = mddev->private;
1945 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1948 abort_sync_write(mddev, r1_bio);
1949 set_bit(WriteErrorSeen, &rdev->flags);
1950 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1951 set_bit(MD_RECOVERY_NEEDED, &
1953 set_bit(R1BIO_WriteError, &r1_bio->state);
1954 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1955 &first_bad, &bad_sectors) &&
1956 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1959 &first_bad, &bad_sectors)
1961 set_bit(R1BIO_MadeGood, &r1_bio->state);
1963 put_sync_write_buf(r1_bio, uptodate);
1966 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1967 int sectors, struct page *page, int rw)
1969 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1973 set_bit(WriteErrorSeen, &rdev->flags);
1974 if (!test_and_set_bit(WantReplacement,
1976 set_bit(MD_RECOVERY_NEEDED, &
1977 rdev->mddev->recovery);
1979 /* need to record an error - either for the block or the device */
1980 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1981 md_error(rdev->mddev, rdev);
1985 static int fix_sync_read_error(struct r1bio *r1_bio)
1987 /* Try some synchronous reads of other devices to get
1988 * good data, much like with normal read errors. Only
1989 * read into the pages we already have so we don't
1990 * need to re-issue the read request.
1991 * We don't need to freeze the array, because being in an
1992 * active sync request, there is no normal IO, and
1993 * no overlapping syncs.
1994 * We don't need to check is_badblock() again as we
1995 * made sure that anything with a bad block in range
1996 * will have bi_end_io clear.
1998 struct mddev *mddev = r1_bio->mddev;
1999 struct r1conf *conf = mddev->private;
2000 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2001 struct page **pages = get_resync_pages(bio)->pages;
2002 sector_t sect = r1_bio->sector;
2003 int sectors = r1_bio->sectors;
2005 struct md_rdev *rdev;
2007 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2008 if (test_bit(FailFast, &rdev->flags)) {
2009 /* Don't try recovering from here - just fail it
2010 * ... unless it is the last working device of course */
2011 md_error(mddev, rdev);
2012 if (test_bit(Faulty, &rdev->flags))
2013 /* Don't try to read from here, but make sure
2014 * put_buf does it's thing
2016 bio->bi_end_io = end_sync_write;
2021 int d = r1_bio->read_disk;
2025 if (s > (PAGE_SIZE>>9))
2028 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2029 /* No rcu protection needed here devices
2030 * can only be removed when no resync is
2031 * active, and resync is currently active
2033 rdev = conf->mirrors[d].rdev;
2034 if (sync_page_io(rdev, sect, s<<9,
2036 REQ_OP_READ, 0, false)) {
2042 if (d == conf->raid_disks * 2)
2044 } while (!success && d != r1_bio->read_disk);
2047 char b[BDEVNAME_SIZE];
2049 /* Cannot read from anywhere, this block is lost.
2050 * Record a bad block on each device. If that doesn't
2051 * work just disable and interrupt the recovery.
2052 * Don't fail devices as that won't really help.
2054 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2055 mdname(mddev), bio_devname(bio, b),
2056 (unsigned long long)r1_bio->sector);
2057 for (d = 0; d < conf->raid_disks * 2; d++) {
2058 rdev = conf->mirrors[d].rdev;
2059 if (!rdev || test_bit(Faulty, &rdev->flags))
2061 if (!rdev_set_badblocks(rdev, sect, s, 0))
2065 conf->recovery_disabled =
2066 mddev->recovery_disabled;
2067 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2068 md_done_sync(mddev, r1_bio->sectors, 0);
2080 /* write it back and re-read */
2081 while (d != r1_bio->read_disk) {
2083 d = conf->raid_disks * 2;
2085 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2087 rdev = conf->mirrors[d].rdev;
2088 if (r1_sync_page_io(rdev, sect, s,
2091 r1_bio->bios[d]->bi_end_io = NULL;
2092 rdev_dec_pending(rdev, mddev);
2096 while (d != r1_bio->read_disk) {
2098 d = conf->raid_disks * 2;
2100 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2102 rdev = conf->mirrors[d].rdev;
2103 if (r1_sync_page_io(rdev, sect, s,
2106 atomic_add(s, &rdev->corrected_errors);
2112 set_bit(R1BIO_Uptodate, &r1_bio->state);
2117 static void process_checks(struct r1bio *r1_bio)
2119 /* We have read all readable devices. If we haven't
2120 * got the block, then there is no hope left.
2121 * If we have, then we want to do a comparison
2122 * and skip the write if everything is the same.
2123 * If any blocks failed to read, then we need to
2124 * attempt an over-write
2126 struct mddev *mddev = r1_bio->mddev;
2127 struct r1conf *conf = mddev->private;
2132 /* Fix variable parts of all bios */
2133 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2134 for (i = 0; i < conf->raid_disks * 2; i++) {
2135 blk_status_t status;
2136 struct bio *b = r1_bio->bios[i];
2137 struct resync_pages *rp = get_resync_pages(b);
2138 if (b->bi_end_io != end_sync_read)
2140 /* fixup the bio for reuse, but preserve errno */
2141 status = b->bi_status;
2143 b->bi_status = status;
2144 b->bi_iter.bi_sector = r1_bio->sector +
2145 conf->mirrors[i].rdev->data_offset;
2146 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2147 b->bi_end_io = end_sync_read;
2148 rp->raid_bio = r1_bio;
2151 /* initialize bvec table again */
2152 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2154 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2155 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2156 !r1_bio->bios[primary]->bi_status) {
2157 r1_bio->bios[primary]->bi_end_io = NULL;
2158 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2161 r1_bio->read_disk = primary;
2162 for (i = 0; i < conf->raid_disks * 2; i++) {
2164 struct bio *pbio = r1_bio->bios[primary];
2165 struct bio *sbio = r1_bio->bios[i];
2166 blk_status_t status = sbio->bi_status;
2167 struct page **ppages = get_resync_pages(pbio)->pages;
2168 struct page **spages = get_resync_pages(sbio)->pages;
2170 int page_len[RESYNC_PAGES] = { 0 };
2171 struct bvec_iter_all iter_all;
2173 if (sbio->bi_end_io != end_sync_read)
2175 /* Now we can 'fixup' the error value */
2176 sbio->bi_status = 0;
2178 bio_for_each_segment_all(bi, sbio, iter_all)
2179 page_len[j++] = bi->bv_len;
2182 for (j = vcnt; j-- ; ) {
2183 if (memcmp(page_address(ppages[j]),
2184 page_address(spages[j]),
2191 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2192 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2194 /* No need to write to this device. */
2195 sbio->bi_end_io = NULL;
2196 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2200 bio_copy_data(sbio, pbio);
2204 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2206 struct r1conf *conf = mddev->private;
2208 int disks = conf->raid_disks * 2;
2211 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2212 /* ouch - failed to read all of that. */
2213 if (!fix_sync_read_error(r1_bio))
2216 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2217 process_checks(r1_bio);
2222 atomic_set(&r1_bio->remaining, 1);
2223 for (i = 0; i < disks ; i++) {
2224 wbio = r1_bio->bios[i];
2225 if (wbio->bi_end_io == NULL ||
2226 (wbio->bi_end_io == end_sync_read &&
2227 (i == r1_bio->read_disk ||
2228 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2230 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2231 abort_sync_write(mddev, r1_bio);
2235 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2236 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2237 wbio->bi_opf |= MD_FAILFAST;
2239 wbio->bi_end_io = end_sync_write;
2240 atomic_inc(&r1_bio->remaining);
2241 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2243 submit_bio_noacct(wbio);
2246 put_sync_write_buf(r1_bio, 1);
2250 * This is a kernel thread which:
2252 * 1. Retries failed read operations on working mirrors.
2253 * 2. Updates the raid superblock when problems encounter.
2254 * 3. Performs writes following reads for array synchronising.
2257 static void fix_read_error(struct r1conf *conf, int read_disk,
2258 sector_t sect, int sectors)
2260 struct mddev *mddev = conf->mddev;
2266 struct md_rdev *rdev;
2268 if (s > (PAGE_SIZE>>9))
2276 rdev = rcu_dereference(conf->mirrors[d].rdev);
2278 (test_bit(In_sync, &rdev->flags) ||
2279 (!test_bit(Faulty, &rdev->flags) &&
2280 rdev->recovery_offset >= sect + s)) &&
2281 is_badblock(rdev, sect, s,
2282 &first_bad, &bad_sectors) == 0) {
2283 atomic_inc(&rdev->nr_pending);
2285 if (sync_page_io(rdev, sect, s<<9,
2286 conf->tmppage, REQ_OP_READ, 0, false))
2288 rdev_dec_pending(rdev, mddev);
2294 if (d == conf->raid_disks * 2)
2296 } while (!success && d != read_disk);
2299 /* Cannot read from anywhere - mark it bad */
2300 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2301 if (!rdev_set_badblocks(rdev, sect, s, 0))
2302 md_error(mddev, rdev);
2305 /* write it back and re-read */
2307 while (d != read_disk) {
2309 d = conf->raid_disks * 2;
2312 rdev = rcu_dereference(conf->mirrors[d].rdev);
2314 !test_bit(Faulty, &rdev->flags)) {
2315 atomic_inc(&rdev->nr_pending);
2317 r1_sync_page_io(rdev, sect, s,
2318 conf->tmppage, WRITE);
2319 rdev_dec_pending(rdev, mddev);
2324 while (d != read_disk) {
2325 char b[BDEVNAME_SIZE];
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 if (r1_sync_page_io(rdev, sect, s,
2336 conf->tmppage, READ)) {
2337 atomic_add(s, &rdev->corrected_errors);
2338 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2340 (unsigned long long)(sect +
2342 bdevname(rdev->bdev, b));
2344 rdev_dec_pending(rdev, mddev);
2353 static int narrow_write_error(struct r1bio *r1_bio, int i)
2355 struct mddev *mddev = r1_bio->mddev;
2356 struct r1conf *conf = mddev->private;
2357 struct md_rdev *rdev = conf->mirrors[i].rdev;
2359 /* bio has the data to be written to device 'i' where
2360 * we just recently had a write error.
2361 * We repeatedly clone the bio and trim down to one block,
2362 * then try the write. Where the write fails we record
2364 * It is conceivable that the bio doesn't exactly align with
2365 * blocks. We must handle this somehow.
2367 * We currently own a reference on the rdev.
2373 int sect_to_write = r1_bio->sectors;
2376 if (rdev->badblocks.shift < 0)
2379 block_sectors = roundup(1 << rdev->badblocks.shift,
2380 bdev_logical_block_size(rdev->bdev) >> 9);
2381 sector = r1_bio->sector;
2382 sectors = ((sector + block_sectors)
2383 & ~(sector_t)(block_sectors - 1))
2386 while (sect_to_write) {
2388 if (sectors > sect_to_write)
2389 sectors = sect_to_write;
2390 /* Write at 'sector' for 'sectors'*/
2392 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2393 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2397 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2401 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2402 wbio->bi_iter.bi_sector = r1_bio->sector;
2403 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2405 bio_trim(wbio, sector - r1_bio->sector, sectors);
2406 wbio->bi_iter.bi_sector += rdev->data_offset;
2407 bio_set_dev(wbio, rdev->bdev);
2409 if (submit_bio_wait(wbio) < 0)
2411 ok = rdev_set_badblocks(rdev, sector,
2416 sect_to_write -= sectors;
2418 sectors = block_sectors;
2423 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2426 int s = r1_bio->sectors;
2427 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2428 struct md_rdev *rdev = conf->mirrors[m].rdev;
2429 struct bio *bio = r1_bio->bios[m];
2430 if (bio->bi_end_io == NULL)
2432 if (!bio->bi_status &&
2433 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2434 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2436 if (bio->bi_status &&
2437 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2438 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2439 md_error(conf->mddev, rdev);
2443 md_done_sync(conf->mddev, s, 1);
2446 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2451 for (m = 0; m < conf->raid_disks * 2 ; m++)
2452 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2453 struct md_rdev *rdev = conf->mirrors[m].rdev;
2454 rdev_clear_badblocks(rdev,
2456 r1_bio->sectors, 0);
2457 rdev_dec_pending(rdev, conf->mddev);
2458 } else if (r1_bio->bios[m] != NULL) {
2459 /* This drive got a write error. We need to
2460 * narrow down and record precise write
2464 if (!narrow_write_error(r1_bio, m)) {
2465 md_error(conf->mddev,
2466 conf->mirrors[m].rdev);
2467 /* an I/O failed, we can't clear the bitmap */
2468 set_bit(R1BIO_Degraded, &r1_bio->state);
2470 rdev_dec_pending(conf->mirrors[m].rdev,
2474 spin_lock_irq(&conf->device_lock);
2475 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2476 idx = sector_to_idx(r1_bio->sector);
2477 atomic_inc(&conf->nr_queued[idx]);
2478 spin_unlock_irq(&conf->device_lock);
2480 * In case freeze_array() is waiting for condition
2481 * get_unqueued_pending() == extra to be true.
2483 wake_up(&conf->wait_barrier);
2484 md_wakeup_thread(conf->mddev->thread);
2486 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2487 close_write(r1_bio);
2488 raid_end_bio_io(r1_bio);
2492 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2494 struct mddev *mddev = conf->mddev;
2496 struct md_rdev *rdev;
2498 clear_bit(R1BIO_ReadError, &r1_bio->state);
2499 /* we got a read error. Maybe the drive is bad. Maybe just
2500 * the block and we can fix it.
2501 * We freeze all other IO, and try reading the block from
2502 * other devices. When we find one, we re-write
2503 * and check it that fixes the read error.
2504 * This is all done synchronously while the array is
2508 bio = r1_bio->bios[r1_bio->read_disk];
2510 r1_bio->bios[r1_bio->read_disk] = NULL;
2512 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2514 && !test_bit(FailFast, &rdev->flags)) {
2515 freeze_array(conf, 1);
2516 fix_read_error(conf, r1_bio->read_disk,
2517 r1_bio->sector, r1_bio->sectors);
2518 unfreeze_array(conf);
2519 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2520 md_error(mddev, rdev);
2522 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2525 rdev_dec_pending(rdev, conf->mddev);
2526 allow_barrier(conf, r1_bio->sector);
2527 bio = r1_bio->master_bio;
2529 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2531 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2534 static void raid1d(struct md_thread *thread)
2536 struct mddev *mddev = thread->mddev;
2537 struct r1bio *r1_bio;
2538 unsigned long flags;
2539 struct r1conf *conf = mddev->private;
2540 struct list_head *head = &conf->retry_list;
2541 struct blk_plug plug;
2544 md_check_recovery(mddev);
2546 if (!list_empty_careful(&conf->bio_end_io_list) &&
2547 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2549 spin_lock_irqsave(&conf->device_lock, flags);
2550 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2551 list_splice_init(&conf->bio_end_io_list, &tmp);
2552 spin_unlock_irqrestore(&conf->device_lock, flags);
2553 while (!list_empty(&tmp)) {
2554 r1_bio = list_first_entry(&tmp, struct r1bio,
2556 list_del(&r1_bio->retry_list);
2557 idx = sector_to_idx(r1_bio->sector);
2558 atomic_dec(&conf->nr_queued[idx]);
2559 if (mddev->degraded)
2560 set_bit(R1BIO_Degraded, &r1_bio->state);
2561 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2562 close_write(r1_bio);
2563 raid_end_bio_io(r1_bio);
2567 blk_start_plug(&plug);
2570 flush_pending_writes(conf);
2572 spin_lock_irqsave(&conf->device_lock, flags);
2573 if (list_empty(head)) {
2574 spin_unlock_irqrestore(&conf->device_lock, flags);
2577 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2578 list_del(head->prev);
2579 idx = sector_to_idx(r1_bio->sector);
2580 atomic_dec(&conf->nr_queued[idx]);
2581 spin_unlock_irqrestore(&conf->device_lock, flags);
2583 mddev = r1_bio->mddev;
2584 conf = mddev->private;
2585 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2586 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2587 test_bit(R1BIO_WriteError, &r1_bio->state))
2588 handle_sync_write_finished(conf, r1_bio);
2590 sync_request_write(mddev, r1_bio);
2591 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2592 test_bit(R1BIO_WriteError, &r1_bio->state))
2593 handle_write_finished(conf, r1_bio);
2594 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2595 handle_read_error(conf, r1_bio);
2600 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2601 md_check_recovery(mddev);
2603 blk_finish_plug(&plug);
2606 static int init_resync(struct r1conf *conf)
2610 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2611 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2613 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2614 r1buf_pool_free, conf->poolinfo);
2617 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2619 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2620 struct resync_pages *rps;
2624 for (i = conf->poolinfo->raid_disks; i--; ) {
2625 bio = r1bio->bios[i];
2626 rps = bio->bi_private;
2628 bio->bi_private = rps;
2630 r1bio->master_bio = NULL;
2635 * perform a "sync" on one "block"
2637 * We need to make sure that no normal I/O request - particularly write
2638 * requests - conflict with active sync requests.
2640 * This is achieved by tracking pending requests and a 'barrier' concept
2641 * that can be installed to exclude normal IO requests.
2644 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2647 struct r1conf *conf = mddev->private;
2648 struct r1bio *r1_bio;
2650 sector_t max_sector, nr_sectors;
2654 int write_targets = 0, read_targets = 0;
2655 sector_t sync_blocks;
2656 int still_degraded = 0;
2657 int good_sectors = RESYNC_SECTORS;
2658 int min_bad = 0; /* number of sectors that are bad in all devices */
2659 int idx = sector_to_idx(sector_nr);
2662 if (!mempool_initialized(&conf->r1buf_pool))
2663 if (init_resync(conf))
2666 max_sector = mddev->dev_sectors;
2667 if (sector_nr >= max_sector) {
2668 /* If we aborted, we need to abort the
2669 * sync on the 'current' bitmap chunk (there will
2670 * only be one in raid1 resync.
2671 * We can find the current addess in mddev->curr_resync
2673 if (mddev->curr_resync < max_sector) /* aborted */
2674 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2676 else /* completed sync */
2679 md_bitmap_close_sync(mddev->bitmap);
2682 if (mddev_is_clustered(mddev)) {
2683 conf->cluster_sync_low = 0;
2684 conf->cluster_sync_high = 0;
2689 if (mddev->bitmap == NULL &&
2690 mddev->recovery_cp == MaxSector &&
2691 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2692 conf->fullsync == 0) {
2694 return max_sector - sector_nr;
2696 /* before building a request, check if we can skip these blocks..
2697 * This call the bitmap_start_sync doesn't actually record anything
2699 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2700 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2701 /* We can skip this block, and probably several more */
2707 * If there is non-resync activity waiting for a turn, then let it
2708 * though before starting on this new sync request.
2710 if (atomic_read(&conf->nr_waiting[idx]))
2711 schedule_timeout_uninterruptible(1);
2713 /* we are incrementing sector_nr below. To be safe, we check against
2714 * sector_nr + two times RESYNC_SECTORS
2717 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2718 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2721 if (raise_barrier(conf, sector_nr))
2724 r1_bio = raid1_alloc_init_r1buf(conf);
2728 * If we get a correctably read error during resync or recovery,
2729 * we might want to read from a different device. So we
2730 * flag all drives that could conceivably be read from for READ,
2731 * and any others (which will be non-In_sync devices) for WRITE.
2732 * If a read fails, we try reading from something else for which READ
2736 r1_bio->mddev = mddev;
2737 r1_bio->sector = sector_nr;
2739 set_bit(R1BIO_IsSync, &r1_bio->state);
2740 /* make sure good_sectors won't go across barrier unit boundary */
2741 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2743 for (i = 0; i < conf->raid_disks * 2; i++) {
2744 struct md_rdev *rdev;
2745 bio = r1_bio->bios[i];
2747 rdev = rcu_dereference(conf->mirrors[i].rdev);
2749 test_bit(Faulty, &rdev->flags)) {
2750 if (i < conf->raid_disks)
2752 } else if (!test_bit(In_sync, &rdev->flags)) {
2753 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2754 bio->bi_end_io = end_sync_write;
2757 /* may need to read from here */
2758 sector_t first_bad = MaxSector;
2761 if (is_badblock(rdev, sector_nr, good_sectors,
2762 &first_bad, &bad_sectors)) {
2763 if (first_bad > sector_nr)
2764 good_sectors = first_bad - sector_nr;
2766 bad_sectors -= (sector_nr - first_bad);
2768 min_bad > bad_sectors)
2769 min_bad = bad_sectors;
2772 if (sector_nr < first_bad) {
2773 if (test_bit(WriteMostly, &rdev->flags)) {
2780 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2781 bio->bi_end_io = end_sync_read;
2783 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2784 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2785 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2787 * The device is suitable for reading (InSync),
2788 * but has bad block(s) here. Let's try to correct them,
2789 * if we are doing resync or repair. Otherwise, leave
2790 * this device alone for this sync request.
2792 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2793 bio->bi_end_io = end_sync_write;
2797 if (rdev && bio->bi_end_io) {
2798 atomic_inc(&rdev->nr_pending);
2799 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2800 bio_set_dev(bio, rdev->bdev);
2801 if (test_bit(FailFast, &rdev->flags))
2802 bio->bi_opf |= MD_FAILFAST;
2808 r1_bio->read_disk = disk;
2810 if (read_targets == 0 && min_bad > 0) {
2811 /* These sectors are bad on all InSync devices, so we
2812 * need to mark them bad on all write targets
2815 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2816 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2817 struct md_rdev *rdev = conf->mirrors[i].rdev;
2818 ok = rdev_set_badblocks(rdev, sector_nr,
2822 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2827 /* Cannot record the badblocks, so need to
2829 * If there are multiple read targets, could just
2830 * fail the really bad ones ???
2832 conf->recovery_disabled = mddev->recovery_disabled;
2833 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2839 if (min_bad > 0 && min_bad < good_sectors) {
2840 /* only resync enough to reach the next bad->good
2842 good_sectors = min_bad;
2845 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2846 /* extra read targets are also write targets */
2847 write_targets += read_targets-1;
2849 if (write_targets == 0 || read_targets == 0) {
2850 /* There is nowhere to write, so all non-sync
2851 * drives must be failed - so we are finished
2855 max_sector = sector_nr + min_bad;
2856 rv = max_sector - sector_nr;
2862 if (max_sector > mddev->resync_max)
2863 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2864 if (max_sector > sector_nr + good_sectors)
2865 max_sector = sector_nr + good_sectors;
2870 int len = PAGE_SIZE;
2871 if (sector_nr + (len>>9) > max_sector)
2872 len = (max_sector - sector_nr) << 9;
2875 if (sync_blocks == 0) {
2876 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2877 &sync_blocks, still_degraded) &&
2879 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2881 if ((len >> 9) > sync_blocks)
2882 len = sync_blocks<<9;
2885 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2886 struct resync_pages *rp;
2888 bio = r1_bio->bios[i];
2889 rp = get_resync_pages(bio);
2890 if (bio->bi_end_io) {
2891 page = resync_fetch_page(rp, page_idx);
2894 * won't fail because the vec table is big
2895 * enough to hold all these pages
2897 bio_add_page(bio, page, len, 0);
2900 nr_sectors += len>>9;
2901 sector_nr += len>>9;
2902 sync_blocks -= (len>>9);
2903 } while (++page_idx < RESYNC_PAGES);
2905 r1_bio->sectors = nr_sectors;
2907 if (mddev_is_clustered(mddev) &&
2908 conf->cluster_sync_high < sector_nr + nr_sectors) {
2909 conf->cluster_sync_low = mddev->curr_resync_completed;
2910 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2911 /* Send resync message */
2912 md_cluster_ops->resync_info_update(mddev,
2913 conf->cluster_sync_low,
2914 conf->cluster_sync_high);
2917 /* For a user-requested sync, we read all readable devices and do a
2920 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2921 atomic_set(&r1_bio->remaining, read_targets);
2922 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2923 bio = r1_bio->bios[i];
2924 if (bio->bi_end_io == end_sync_read) {
2926 md_sync_acct_bio(bio, nr_sectors);
2927 if (read_targets == 1)
2928 bio->bi_opf &= ~MD_FAILFAST;
2929 submit_bio_noacct(bio);
2933 atomic_set(&r1_bio->remaining, 1);
2934 bio = r1_bio->bios[r1_bio->read_disk];
2935 md_sync_acct_bio(bio, nr_sectors);
2936 if (read_targets == 1)
2937 bio->bi_opf &= ~MD_FAILFAST;
2938 submit_bio_noacct(bio);
2943 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2948 return mddev->dev_sectors;
2951 static struct r1conf *setup_conf(struct mddev *mddev)
2953 struct r1conf *conf;
2955 struct raid1_info *disk;
2956 struct md_rdev *rdev;
2959 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2963 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2964 sizeof(atomic_t), GFP_KERNEL);
2965 if (!conf->nr_pending)
2968 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2969 sizeof(atomic_t), GFP_KERNEL);
2970 if (!conf->nr_waiting)
2973 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2974 sizeof(atomic_t), GFP_KERNEL);
2975 if (!conf->nr_queued)
2978 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2979 sizeof(atomic_t), GFP_KERNEL);
2983 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2984 mddev->raid_disks, 2),
2989 conf->tmppage = alloc_page(GFP_KERNEL);
2993 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2994 if (!conf->poolinfo)
2996 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2997 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2998 rbio_pool_free, conf->poolinfo);
3002 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3006 conf->poolinfo->mddev = mddev;
3009 spin_lock_init(&conf->device_lock);
3010 rdev_for_each(rdev, mddev) {
3011 int disk_idx = rdev->raid_disk;
3012 if (disk_idx >= mddev->raid_disks
3015 if (test_bit(Replacement, &rdev->flags))
3016 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3018 disk = conf->mirrors + disk_idx;
3023 disk->head_position = 0;
3024 disk->seq_start = MaxSector;
3026 conf->raid_disks = mddev->raid_disks;
3027 conf->mddev = mddev;
3028 INIT_LIST_HEAD(&conf->retry_list);
3029 INIT_LIST_HEAD(&conf->bio_end_io_list);
3031 spin_lock_init(&conf->resync_lock);
3032 init_waitqueue_head(&conf->wait_barrier);
3034 bio_list_init(&conf->pending_bio_list);
3035 conf->pending_count = 0;
3036 conf->recovery_disabled = mddev->recovery_disabled - 1;
3039 for (i = 0; i < conf->raid_disks * 2; i++) {
3041 disk = conf->mirrors + i;
3043 if (i < conf->raid_disks &&
3044 disk[conf->raid_disks].rdev) {
3045 /* This slot has a replacement. */
3047 /* No original, just make the replacement
3048 * a recovering spare
3051 disk[conf->raid_disks].rdev;
3052 disk[conf->raid_disks].rdev = NULL;
3053 } else if (!test_bit(In_sync, &disk->rdev->flags))
3054 /* Original is not in_sync - bad */
3059 !test_bit(In_sync, &disk->rdev->flags)) {
3060 disk->head_position = 0;
3062 (disk->rdev->saved_raid_disk < 0))
3068 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3076 mempool_exit(&conf->r1bio_pool);
3077 kfree(conf->mirrors);
3078 safe_put_page(conf->tmppage);
3079 kfree(conf->poolinfo);
3080 kfree(conf->nr_pending);
3081 kfree(conf->nr_waiting);
3082 kfree(conf->nr_queued);
3083 kfree(conf->barrier);
3084 bioset_exit(&conf->bio_split);
3087 return ERR_PTR(err);
3090 static void raid1_free(struct mddev *mddev, void *priv);
3091 static int raid1_run(struct mddev *mddev)
3093 struct r1conf *conf;
3095 struct md_rdev *rdev;
3097 bool discard_supported = false;
3099 if (mddev->level != 1) {
3100 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3101 mdname(mddev), mddev->level);
3104 if (mddev->reshape_position != MaxSector) {
3105 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3109 if (mddev_init_writes_pending(mddev) < 0)
3112 * copy the already verified devices into our private RAID1
3113 * bookkeeping area. [whatever we allocate in run(),
3114 * should be freed in raid1_free()]
3116 if (mddev->private == NULL)
3117 conf = setup_conf(mddev);
3119 conf = mddev->private;
3122 return PTR_ERR(conf);
3125 blk_queue_max_write_same_sectors(mddev->queue, 0);
3126 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3129 rdev_for_each(rdev, mddev) {
3130 if (!mddev->gendisk)
3132 disk_stack_limits(mddev->gendisk, rdev->bdev,
3133 rdev->data_offset << 9);
3134 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3135 discard_supported = true;
3138 mddev->degraded = 0;
3139 for (i = 0; i < conf->raid_disks; i++)
3140 if (conf->mirrors[i].rdev == NULL ||
3141 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3142 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3145 * RAID1 needs at least one disk in active
3147 if (conf->raid_disks - mddev->degraded < 1) {
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 mddev->thread = conf->thread;
3166 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));
3173 if (discard_supported)
3174 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3177 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3181 ret = md_integrity_register(mddev);
3183 md_unregister_thread(&mddev->thread);
3189 raid1_free(mddev, conf);
3193 static void raid1_free(struct mddev *mddev, void *priv)
3195 struct r1conf *conf = priv;
3197 mempool_exit(&conf->r1bio_pool);
3198 kfree(conf->mirrors);
3199 safe_put_page(conf->tmppage);
3200 kfree(conf->poolinfo);
3201 kfree(conf->nr_pending);
3202 kfree(conf->nr_waiting);
3203 kfree(conf->nr_queued);
3204 kfree(conf->barrier);
3205 bioset_exit(&conf->bio_split);
3209 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3211 /* no resync is happening, and there is enough space
3212 * on all devices, so we can resize.
3213 * We need to make sure resync covers any new space.
3214 * If the array is shrinking we should possibly wait until
3215 * any io in the removed space completes, but it hardly seems
3218 sector_t newsize = raid1_size(mddev, sectors, 0);
3219 if (mddev->external_size &&
3220 mddev->array_sectors > newsize)
3222 if (mddev->bitmap) {
3223 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3227 md_set_array_sectors(mddev, newsize);
3228 if (sectors > mddev->dev_sectors &&
3229 mddev->recovery_cp > mddev->dev_sectors) {
3230 mddev->recovery_cp = mddev->dev_sectors;
3231 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3233 mddev->dev_sectors = sectors;
3234 mddev->resync_max_sectors = sectors;
3238 static int raid1_reshape(struct mddev *mddev)
3241 * 1/ resize the r1bio_pool
3242 * 2/ resize conf->mirrors
3244 * We allocate a new r1bio_pool if we can.
3245 * Then raise a device barrier and wait until all IO stops.
3246 * Then resize conf->mirrors and swap in the new r1bio pool.
3248 * At the same time, we "pack" the devices so that all the missing
3249 * devices have the higher raid_disk numbers.
3251 mempool_t newpool, oldpool;
3252 struct pool_info *newpoolinfo;
3253 struct raid1_info *newmirrors;
3254 struct r1conf *conf = mddev->private;
3255 int cnt, raid_disks;
3256 unsigned long flags;
3260 memset(&newpool, 0, sizeof(newpool));
3261 memset(&oldpool, 0, sizeof(oldpool));
3263 /* Cannot change chunk_size, layout, or level */
3264 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3265 mddev->layout != mddev->new_layout ||
3266 mddev->level != mddev->new_level) {
3267 mddev->new_chunk_sectors = mddev->chunk_sectors;
3268 mddev->new_layout = mddev->layout;
3269 mddev->new_level = mddev->level;
3273 if (!mddev_is_clustered(mddev))
3274 md_allow_write(mddev);
3276 raid_disks = mddev->raid_disks + mddev->delta_disks;
3278 if (raid_disks < conf->raid_disks) {
3280 for (d= 0; d < conf->raid_disks; d++)
3281 if (conf->mirrors[d].rdev)
3283 if (cnt > raid_disks)
3287 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3290 newpoolinfo->mddev = mddev;
3291 newpoolinfo->raid_disks = raid_disks * 2;
3293 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3294 rbio_pool_free, newpoolinfo);
3299 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3304 mempool_exit(&newpool);
3308 freeze_array(conf, 0);
3310 /* ok, everything is stopped */
3311 oldpool = conf->r1bio_pool;
3312 conf->r1bio_pool = newpool;
3314 for (d = d2 = 0; d < conf->raid_disks; d++) {
3315 struct md_rdev *rdev = conf->mirrors[d].rdev;
3316 if (rdev && rdev->raid_disk != d2) {
3317 sysfs_unlink_rdev(mddev, rdev);
3318 rdev->raid_disk = d2;
3319 sysfs_unlink_rdev(mddev, rdev);
3320 if (sysfs_link_rdev(mddev, rdev))
3321 pr_warn("md/raid1:%s: cannot register rd%d\n",
3322 mdname(mddev), rdev->raid_disk);
3325 newmirrors[d2++].rdev = rdev;
3327 kfree(conf->mirrors);
3328 conf->mirrors = newmirrors;
3329 kfree(conf->poolinfo);
3330 conf->poolinfo = newpoolinfo;
3332 spin_lock_irqsave(&conf->device_lock, flags);
3333 mddev->degraded += (raid_disks - conf->raid_disks);
3334 spin_unlock_irqrestore(&conf->device_lock, flags);
3335 conf->raid_disks = mddev->raid_disks = raid_disks;
3336 mddev->delta_disks = 0;
3338 unfreeze_array(conf);
3340 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3341 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3342 md_wakeup_thread(mddev->thread);
3344 mempool_exit(&oldpool);
3348 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3350 struct r1conf *conf = mddev->private;
3353 freeze_array(conf, 0);
3355 unfreeze_array(conf);
3358 static void *raid1_takeover(struct mddev *mddev)
3360 /* raid1 can take over:
3361 * raid5 with 2 devices, any layout or chunk size
3363 if (mddev->level == 5 && mddev->raid_disks == 2) {
3364 struct r1conf *conf;
3365 mddev->new_level = 1;
3366 mddev->new_layout = 0;
3367 mddev->new_chunk_sectors = 0;
3368 conf = setup_conf(mddev);
3369 if (!IS_ERR(conf)) {
3370 /* Array must appear to be quiesced */
3371 conf->array_frozen = 1;
3372 mddev_clear_unsupported_flags(mddev,
3373 UNSUPPORTED_MDDEV_FLAGS);
3377 return ERR_PTR(-EINVAL);
3380 static struct md_personality raid1_personality =
3384 .owner = THIS_MODULE,
3385 .make_request = raid1_make_request,
3388 .status = raid1_status,
3389 .error_handler = raid1_error,
3390 .hot_add_disk = raid1_add_disk,
3391 .hot_remove_disk= raid1_remove_disk,
3392 .spare_active = raid1_spare_active,
3393 .sync_request = raid1_sync_request,
3394 .resize = raid1_resize,
3396 .check_reshape = raid1_reshape,
3397 .quiesce = raid1_quiesce,
3398 .takeover = raid1_takeover,
3399 .congested = raid1_congested,
3402 static int __init raid_init(void)
3404 return register_md_personality(&raid1_personality);
3407 static void raid_exit(void)
3409 unregister_md_personality(&raid1_personality);
3412 module_init(raid_init);
3413 module_exit(raid_exit);
3414 MODULE_LICENSE("GPL");
3415 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3416 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3417 MODULE_ALIAS("md-raid1");
3418 MODULE_ALIAS("md-level-1");
3420 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);