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 /* Fail the request */
482 set_bit(R1BIO_Degraded, &r1_bio->state);
483 /* Finished with this branch */
484 r1_bio->bios[mirror] = NULL;
489 * Set R1BIO_Uptodate in our master bio, so that we
490 * will return a good error code for to the higher
491 * levels even if IO on some other mirrored buffer
494 * The 'master' represents the composite IO operation
495 * to user-side. So if something waits for IO, then it
496 * will wait for the 'master' bio.
501 r1_bio->bios[mirror] = NULL;
504 * Do not set R1BIO_Uptodate if the current device is
505 * rebuilding or Faulty. This is because we cannot use
506 * such device for properly reading the data back (we could
507 * potentially use it, if the current write would have felt
508 * before rdev->recovery_offset, but for simplicity we don't
511 if (test_bit(In_sync, &rdev->flags) &&
512 !test_bit(Faulty, &rdev->flags))
513 set_bit(R1BIO_Uptodate, &r1_bio->state);
515 /* Maybe we can clear some bad blocks. */
516 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
517 &first_bad, &bad_sectors) && !discard_error) {
518 r1_bio->bios[mirror] = IO_MADE_GOOD;
519 set_bit(R1BIO_MadeGood, &r1_bio->state);
524 if (test_bit(CollisionCheck, &rdev->flags))
525 remove_serial(rdev, lo, hi);
526 if (test_bit(WriteMostly, &rdev->flags))
527 atomic_dec(&r1_bio->behind_remaining);
530 * In behind mode, we ACK the master bio once the I/O
531 * has safely reached all non-writemostly
532 * disks. Setting the Returned bit ensures that this
533 * gets done only once -- we don't ever want to return
534 * -EIO here, instead we'll wait
536 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
537 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
538 /* Maybe we can return now */
539 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
540 struct bio *mbio = r1_bio->master_bio;
541 pr_debug("raid1: behind end write sectors"
543 (unsigned long long) mbio->bi_iter.bi_sector,
544 (unsigned long long) bio_end_sector(mbio) - 1);
545 call_bio_endio(r1_bio);
548 } else if (rdev->mddev->serialize_policy)
549 remove_serial(rdev, lo, hi);
550 if (r1_bio->bios[mirror] == NULL)
551 rdev_dec_pending(rdev, conf->mddev);
554 * Let's see if all mirrored write operations have finished
557 r1_bio_write_done(r1_bio);
563 static sector_t align_to_barrier_unit_end(sector_t start_sector,
568 WARN_ON(sectors == 0);
570 * len is the number of sectors from start_sector to end of the
571 * barrier unit which start_sector belongs to.
573 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
583 * This routine returns the disk from which the requested read should
584 * be done. There is a per-array 'next expected sequential IO' sector
585 * number - if this matches on the next IO then we use the last disk.
586 * There is also a per-disk 'last know head position' sector that is
587 * maintained from IRQ contexts, both the normal and the resync IO
588 * completion handlers update this position correctly. If there is no
589 * perfect sequential match then we pick the disk whose head is closest.
591 * If there are 2 mirrors in the same 2 devices, performance degrades
592 * because position is mirror, not device based.
594 * The rdev for the device selected will have nr_pending incremented.
596 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
598 const sector_t this_sector = r1_bio->sector;
600 int best_good_sectors;
601 int best_disk, best_dist_disk, best_pending_disk;
605 unsigned int min_pending;
606 struct md_rdev *rdev;
608 int choose_next_idle;
612 * Check if we can balance. We can balance on the whole
613 * device if no resync is going on, or below the resync window.
614 * We take the first readable disk when above the resync window.
617 sectors = r1_bio->sectors;
620 best_dist = MaxSector;
621 best_pending_disk = -1;
622 min_pending = UINT_MAX;
623 best_good_sectors = 0;
625 choose_next_idle = 0;
626 clear_bit(R1BIO_FailFast, &r1_bio->state);
628 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
629 (mddev_is_clustered(conf->mddev) &&
630 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
631 this_sector + sectors)))
636 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
640 unsigned int pending;
643 rdev = rcu_dereference(conf->mirrors[disk].rdev);
644 if (r1_bio->bios[disk] == IO_BLOCKED
646 || test_bit(Faulty, &rdev->flags))
648 if (!test_bit(In_sync, &rdev->flags) &&
649 rdev->recovery_offset < this_sector + sectors)
651 if (test_bit(WriteMostly, &rdev->flags)) {
652 /* Don't balance among write-mostly, just
653 * use the first as a last resort */
654 if (best_dist_disk < 0) {
655 if (is_badblock(rdev, this_sector, sectors,
656 &first_bad, &bad_sectors)) {
657 if (first_bad <= this_sector)
658 /* Cannot use this */
660 best_good_sectors = first_bad - this_sector;
662 best_good_sectors = sectors;
663 best_dist_disk = disk;
664 best_pending_disk = disk;
668 /* This is a reasonable device to use. It might
671 if (is_badblock(rdev, this_sector, sectors,
672 &first_bad, &bad_sectors)) {
673 if (best_dist < MaxSector)
674 /* already have a better device */
676 if (first_bad <= this_sector) {
677 /* cannot read here. If this is the 'primary'
678 * device, then we must not read beyond
679 * bad_sectors from another device..
681 bad_sectors -= (this_sector - first_bad);
682 if (choose_first && sectors > bad_sectors)
683 sectors = bad_sectors;
684 if (best_good_sectors > sectors)
685 best_good_sectors = sectors;
688 sector_t good_sectors = first_bad - this_sector;
689 if (good_sectors > best_good_sectors) {
690 best_good_sectors = good_sectors;
698 if ((sectors > best_good_sectors) && (best_disk >= 0))
700 best_good_sectors = sectors;
704 /* At least two disks to choose from so failfast is OK */
705 set_bit(R1BIO_FailFast, &r1_bio->state);
707 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
708 has_nonrot_disk |= nonrot;
709 pending = atomic_read(&rdev->nr_pending);
710 dist = abs(this_sector - conf->mirrors[disk].head_position);
715 /* Don't change to another disk for sequential reads */
716 if (conf->mirrors[disk].next_seq_sect == this_sector
718 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
719 struct raid1_info *mirror = &conf->mirrors[disk];
723 * If buffered sequential IO size exceeds optimal
724 * iosize, check if there is idle disk. If yes, choose
725 * the idle disk. read_balance could already choose an
726 * idle disk before noticing it's a sequential IO in
727 * this disk. This doesn't matter because this disk
728 * will idle, next time it will be utilized after the
729 * first disk has IO size exceeds optimal iosize. In
730 * this way, iosize of the first disk will be optimal
731 * iosize at least. iosize of the second disk might be
732 * small, but not a big deal since when the second disk
733 * starts IO, the first disk is likely still busy.
735 if (nonrot && opt_iosize > 0 &&
736 mirror->seq_start != MaxSector &&
737 mirror->next_seq_sect > opt_iosize &&
738 mirror->next_seq_sect - opt_iosize >=
740 choose_next_idle = 1;
746 if (choose_next_idle)
749 if (min_pending > pending) {
750 min_pending = pending;
751 best_pending_disk = disk;
754 if (dist < best_dist) {
756 best_dist_disk = disk;
761 * If all disks are rotational, choose the closest disk. If any disk is
762 * non-rotational, choose the disk with less pending request even the
763 * disk is rotational, which might/might not be optimal for raids with
764 * mixed ratation/non-rotational disks depending on workload.
766 if (best_disk == -1) {
767 if (has_nonrot_disk || min_pending == 0)
768 best_disk = best_pending_disk;
770 best_disk = best_dist_disk;
773 if (best_disk >= 0) {
774 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
777 atomic_inc(&rdev->nr_pending);
778 sectors = best_good_sectors;
780 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
781 conf->mirrors[best_disk].seq_start = this_sector;
783 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
786 *max_sectors = sectors;
791 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
793 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
794 md_bitmap_unplug(conf->mddev->bitmap);
795 wake_up(&conf->wait_barrier);
797 while (bio) { /* submit pending writes */
798 struct bio *next = bio->bi_next;
799 struct md_rdev *rdev = (void *)bio->bi_bdev;
801 bio_set_dev(bio, rdev->bdev);
802 if (test_bit(Faulty, &rdev->flags)) {
804 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
805 !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
809 submit_bio_noacct(bio);
815 static void flush_pending_writes(struct r1conf *conf)
817 /* Any writes that have been queued but are awaiting
818 * bitmap updates get flushed here.
820 spin_lock_irq(&conf->device_lock);
822 if (conf->pending_bio_list.head) {
823 struct blk_plug plug;
826 bio = bio_list_get(&conf->pending_bio_list);
827 conf->pending_count = 0;
828 spin_unlock_irq(&conf->device_lock);
831 * As this is called in a wait_event() loop (see freeze_array),
832 * current->state might be TASK_UNINTERRUPTIBLE which will
833 * cause a warning when we prepare to wait again. As it is
834 * rare that this path is taken, it is perfectly safe to force
835 * us to go around the wait_event() loop again, so the warning
836 * is a false-positive. Silence the warning by resetting
839 __set_current_state(TASK_RUNNING);
840 blk_start_plug(&plug);
841 flush_bio_list(conf, bio);
842 blk_finish_plug(&plug);
844 spin_unlock_irq(&conf->device_lock);
848 * Sometimes we need to suspend IO while we do something else,
849 * either some resync/recovery, or reconfigure the array.
850 * To do this we raise a 'barrier'.
851 * The 'barrier' is a counter that can be raised multiple times
852 * to count how many activities are happening which preclude
854 * We can only raise the barrier if there is no pending IO.
855 * i.e. if nr_pending == 0.
856 * We choose only to raise the barrier if no-one is waiting for the
857 * barrier to go down. This means that as soon as an IO request
858 * is ready, no other operations which require a barrier will start
859 * until the IO request has had a chance.
861 * So: regular IO calls 'wait_barrier'. When that returns there
862 * is no backgroup IO happening, It must arrange to call
863 * allow_barrier when it has finished its IO.
864 * backgroup IO calls must call raise_barrier. Once that returns
865 * there is no normal IO happeing. It must arrange to call
866 * lower_barrier when the particular background IO completes.
868 * If resync/recovery is interrupted, returns -EINTR;
869 * Otherwise, returns 0.
871 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
873 int idx = sector_to_idx(sector_nr);
875 spin_lock_irq(&conf->resync_lock);
877 /* Wait until no block IO is waiting */
878 wait_event_lock_irq(conf->wait_barrier,
879 !atomic_read(&conf->nr_waiting[idx]),
882 /* block any new IO from starting */
883 atomic_inc(&conf->barrier[idx]);
885 * In raise_barrier() we firstly increase conf->barrier[idx] then
886 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
887 * increase conf->nr_pending[idx] then check conf->barrier[idx].
888 * A memory barrier here to make sure conf->nr_pending[idx] won't
889 * be fetched before conf->barrier[idx] is increased. Otherwise
890 * there will be a race between raise_barrier() and _wait_barrier().
892 smp_mb__after_atomic();
894 /* For these conditions we must wait:
895 * A: while the array is in frozen state
896 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
897 * existing in corresponding I/O barrier bucket.
898 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
899 * max resync count which allowed on current I/O barrier bucket.
901 wait_event_lock_irq(conf->wait_barrier,
902 (!conf->array_frozen &&
903 !atomic_read(&conf->nr_pending[idx]) &&
904 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
905 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
908 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
909 atomic_dec(&conf->barrier[idx]);
910 spin_unlock_irq(&conf->resync_lock);
911 wake_up(&conf->wait_barrier);
915 atomic_inc(&conf->nr_sync_pending);
916 spin_unlock_irq(&conf->resync_lock);
921 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
923 int idx = sector_to_idx(sector_nr);
925 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
927 atomic_dec(&conf->barrier[idx]);
928 atomic_dec(&conf->nr_sync_pending);
929 wake_up(&conf->wait_barrier);
932 static void _wait_barrier(struct r1conf *conf, int idx)
935 * We need to increase conf->nr_pending[idx] very early here,
936 * then raise_barrier() can be blocked when it waits for
937 * conf->nr_pending[idx] to be 0. Then we can avoid holding
938 * conf->resync_lock when there is no barrier raised in same
939 * barrier unit bucket. Also if the array is frozen, I/O
940 * should be blocked until array is unfrozen.
942 atomic_inc(&conf->nr_pending[idx]);
944 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
945 * check conf->barrier[idx]. In raise_barrier() we firstly increase
946 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
947 * barrier is necessary here to make sure conf->barrier[idx] won't be
948 * fetched before conf->nr_pending[idx] is increased. Otherwise there
949 * will be a race between _wait_barrier() and raise_barrier().
951 smp_mb__after_atomic();
954 * Don't worry about checking two atomic_t variables at same time
955 * here. If during we check conf->barrier[idx], the array is
956 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
957 * 0, it is safe to return and make the I/O continue. Because the
958 * array is frozen, all I/O returned here will eventually complete
959 * or be queued, no race will happen. See code comment in
962 if (!READ_ONCE(conf->array_frozen) &&
963 !atomic_read(&conf->barrier[idx]))
967 * After holding conf->resync_lock, conf->nr_pending[idx]
968 * should be decreased before waiting for barrier to drop.
969 * Otherwise, we may encounter a race condition because
970 * raise_barrer() might be waiting for conf->nr_pending[idx]
971 * to be 0 at same time.
973 spin_lock_irq(&conf->resync_lock);
974 atomic_inc(&conf->nr_waiting[idx]);
975 atomic_dec(&conf->nr_pending[idx]);
977 * In case freeze_array() is waiting for
978 * get_unqueued_pending() == extra
980 wake_up(&conf->wait_barrier);
981 /* Wait for the barrier in same barrier unit bucket to drop. */
982 wait_event_lock_irq(conf->wait_barrier,
983 !conf->array_frozen &&
984 !atomic_read(&conf->barrier[idx]),
986 atomic_inc(&conf->nr_pending[idx]);
987 atomic_dec(&conf->nr_waiting[idx]);
988 spin_unlock_irq(&conf->resync_lock);
991 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
993 int idx = sector_to_idx(sector_nr);
996 * Very similar to _wait_barrier(). The difference is, for read
997 * I/O we don't need wait for sync I/O, but if the whole array
998 * is frozen, the read I/O still has to wait until the array is
999 * unfrozen. Since there is no ordering requirement with
1000 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1002 atomic_inc(&conf->nr_pending[idx]);
1004 if (!READ_ONCE(conf->array_frozen))
1007 spin_lock_irq(&conf->resync_lock);
1008 atomic_inc(&conf->nr_waiting[idx]);
1009 atomic_dec(&conf->nr_pending[idx]);
1011 * In case freeze_array() is waiting for
1012 * get_unqueued_pending() == extra
1014 wake_up(&conf->wait_barrier);
1015 /* Wait for array to be unfrozen */
1016 wait_event_lock_irq(conf->wait_barrier,
1017 !conf->array_frozen,
1019 atomic_inc(&conf->nr_pending[idx]);
1020 atomic_dec(&conf->nr_waiting[idx]);
1021 spin_unlock_irq(&conf->resync_lock);
1024 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1026 int idx = sector_to_idx(sector_nr);
1028 _wait_barrier(conf, idx);
1031 static void _allow_barrier(struct r1conf *conf, int idx)
1033 atomic_dec(&conf->nr_pending[idx]);
1034 wake_up(&conf->wait_barrier);
1037 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1039 int idx = sector_to_idx(sector_nr);
1041 _allow_barrier(conf, idx);
1044 /* conf->resync_lock should be held */
1045 static int get_unqueued_pending(struct r1conf *conf)
1049 ret = atomic_read(&conf->nr_sync_pending);
1050 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1051 ret += atomic_read(&conf->nr_pending[idx]) -
1052 atomic_read(&conf->nr_queued[idx]);
1057 static void freeze_array(struct r1conf *conf, int extra)
1059 /* Stop sync I/O and normal I/O and wait for everything to
1061 * This is called in two situations:
1062 * 1) management command handlers (reshape, remove disk, quiesce).
1063 * 2) one normal I/O request failed.
1065 * After array_frozen is set to 1, new sync IO will be blocked at
1066 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1067 * or wait_read_barrier(). The flying I/Os will either complete or be
1068 * queued. When everything goes quite, there are only queued I/Os left.
1070 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1071 * barrier bucket index which this I/O request hits. When all sync and
1072 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1073 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1074 * in handle_read_error(), we may call freeze_array() before trying to
1075 * fix the read error. In this case, the error read I/O is not queued,
1076 * so get_unqueued_pending() == 1.
1078 * Therefore before this function returns, we need to wait until
1079 * get_unqueued_pendings(conf) gets equal to extra. For
1080 * normal I/O context, extra is 1, in rested situations extra is 0.
1082 spin_lock_irq(&conf->resync_lock);
1083 conf->array_frozen = 1;
1084 raid1_log(conf->mddev, "wait freeze");
1085 wait_event_lock_irq_cmd(
1087 get_unqueued_pending(conf) == extra,
1089 flush_pending_writes(conf));
1090 spin_unlock_irq(&conf->resync_lock);
1092 static void unfreeze_array(struct r1conf *conf)
1094 /* reverse the effect of the freeze */
1095 spin_lock_irq(&conf->resync_lock);
1096 conf->array_frozen = 0;
1097 spin_unlock_irq(&conf->resync_lock);
1098 wake_up(&conf->wait_barrier);
1101 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1104 int size = bio->bi_iter.bi_size;
1105 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1107 struct bio *behind_bio = NULL;
1109 behind_bio = bio_alloc_bioset(GFP_NOIO, vcnt, &r1_bio->mddev->bio_set);
1113 /* discard op, we don't support writezero/writesame yet */
1114 if (!bio_has_data(bio)) {
1115 behind_bio->bi_iter.bi_size = size;
1119 behind_bio->bi_write_hint = bio->bi_write_hint;
1121 while (i < vcnt && size) {
1123 int len = min_t(int, PAGE_SIZE, size);
1125 page = alloc_page(GFP_NOIO);
1126 if (unlikely(!page))
1129 bio_add_page(behind_bio, page, len, 0);
1135 bio_copy_data(behind_bio, bio);
1137 r1_bio->behind_master_bio = behind_bio;
1138 set_bit(R1BIO_BehindIO, &r1_bio->state);
1143 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1144 bio->bi_iter.bi_size);
1145 bio_free_pages(behind_bio);
1146 bio_put(behind_bio);
1149 struct raid1_plug_cb {
1150 struct blk_plug_cb cb;
1151 struct bio_list pending;
1155 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1157 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1159 struct mddev *mddev = plug->cb.data;
1160 struct r1conf *conf = mddev->private;
1163 if (from_schedule || current->bio_list) {
1164 spin_lock_irq(&conf->device_lock);
1165 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1166 conf->pending_count += plug->pending_cnt;
1167 spin_unlock_irq(&conf->device_lock);
1168 wake_up(&conf->wait_barrier);
1169 md_wakeup_thread(mddev->thread);
1174 /* we aren't scheduling, so we can do the write-out directly. */
1175 bio = bio_list_get(&plug->pending);
1176 flush_bio_list(conf, bio);
1180 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1182 r1_bio->master_bio = bio;
1183 r1_bio->sectors = bio_sectors(bio);
1185 r1_bio->mddev = mddev;
1186 r1_bio->sector = bio->bi_iter.bi_sector;
1189 static inline struct r1bio *
1190 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1192 struct r1conf *conf = mddev->private;
1193 struct r1bio *r1_bio;
1195 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1196 /* Ensure no bio records IO_BLOCKED */
1197 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1198 init_r1bio(r1_bio, mddev, bio);
1202 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1203 int max_read_sectors, struct r1bio *r1_bio)
1205 struct r1conf *conf = mddev->private;
1206 struct raid1_info *mirror;
1207 struct bio *read_bio;
1208 struct bitmap *bitmap = mddev->bitmap;
1209 const int op = bio_op(bio);
1210 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1213 bool print_msg = !!r1_bio;
1214 char b[BDEVNAME_SIZE];
1217 * If r1_bio is set, we are blocking the raid1d thread
1218 * so there is a tiny risk of deadlock. So ask for
1219 * emergency memory if needed.
1221 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1224 /* Need to get the block device name carefully */
1225 struct md_rdev *rdev;
1227 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1229 bdevname(rdev->bdev, b);
1236 * Still need barrier for READ in case that whole
1239 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1242 r1_bio = alloc_r1bio(mddev, bio);
1244 init_r1bio(r1_bio, mddev, bio);
1245 r1_bio->sectors = max_read_sectors;
1248 * make_request() can abort the operation when read-ahead is being
1249 * used and no empty request is available.
1251 rdisk = read_balance(conf, r1_bio, &max_sectors);
1254 /* couldn't find anywhere to read from */
1256 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1259 (unsigned long long)r1_bio->sector);
1261 raid_end_bio_io(r1_bio);
1264 mirror = conf->mirrors + rdisk;
1267 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1269 (unsigned long long)r1_bio->sector,
1270 bdevname(mirror->rdev->bdev, b));
1272 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1275 * Reading from a write-mostly device must take care not to
1276 * over-take any writes that are 'behind'
1278 raid1_log(mddev, "wait behind writes");
1279 wait_event(bitmap->behind_wait,
1280 atomic_read(&bitmap->behind_writes) == 0);
1283 if (max_sectors < bio_sectors(bio)) {
1284 struct bio *split = bio_split(bio, max_sectors,
1285 gfp, &conf->bio_split);
1286 bio_chain(split, bio);
1287 submit_bio_noacct(bio);
1289 r1_bio->master_bio = bio;
1290 r1_bio->sectors = max_sectors;
1293 r1_bio->read_disk = rdisk;
1295 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1297 r1_bio->bios[rdisk] = read_bio;
1299 read_bio->bi_iter.bi_sector = r1_bio->sector +
1300 mirror->rdev->data_offset;
1301 bio_set_dev(read_bio, mirror->rdev->bdev);
1302 read_bio->bi_end_io = raid1_end_read_request;
1303 bio_set_op_attrs(read_bio, op, do_sync);
1304 if (test_bit(FailFast, &mirror->rdev->flags) &&
1305 test_bit(R1BIO_FailFast, &r1_bio->state))
1306 read_bio->bi_opf |= MD_FAILFAST;
1307 read_bio->bi_private = r1_bio;
1310 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1313 submit_bio_noacct(read_bio);
1316 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1317 int max_write_sectors)
1319 struct r1conf *conf = mddev->private;
1320 struct r1bio *r1_bio;
1322 struct bitmap *bitmap = mddev->bitmap;
1323 unsigned long flags;
1324 struct md_rdev *blocked_rdev;
1325 struct blk_plug_cb *cb;
1326 struct raid1_plug_cb *plug = NULL;
1330 if (mddev_is_clustered(mddev) &&
1331 md_cluster_ops->area_resyncing(mddev, WRITE,
1332 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1336 prepare_to_wait(&conf->wait_barrier,
1338 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1339 bio->bi_iter.bi_sector,
1340 bio_end_sector(bio)))
1344 finish_wait(&conf->wait_barrier, &w);
1348 * Register the new request and wait if the reconstruction
1349 * thread has put up a bar for new requests.
1350 * Continue immediately if no resync is active currently.
1352 wait_barrier(conf, bio->bi_iter.bi_sector);
1354 r1_bio = alloc_r1bio(mddev, bio);
1355 r1_bio->sectors = max_write_sectors;
1357 if (conf->pending_count >= max_queued_requests) {
1358 md_wakeup_thread(mddev->thread);
1359 raid1_log(mddev, "wait queued");
1360 wait_event(conf->wait_barrier,
1361 conf->pending_count < max_queued_requests);
1363 /* first select target devices under rcu_lock and
1364 * inc refcount on their rdev. Record them by setting
1366 * If there are known/acknowledged bad blocks on any device on
1367 * which we have seen a write error, we want to avoid writing those
1369 * This potentially requires several writes to write around
1370 * the bad blocks. Each set of writes gets it's own r1bio
1371 * with a set of bios attached.
1374 disks = conf->raid_disks * 2;
1376 blocked_rdev = NULL;
1378 max_sectors = r1_bio->sectors;
1379 for (i = 0; i < disks; i++) {
1380 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1381 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1382 atomic_inc(&rdev->nr_pending);
1383 blocked_rdev = rdev;
1386 r1_bio->bios[i] = NULL;
1387 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1388 if (i < conf->raid_disks)
1389 set_bit(R1BIO_Degraded, &r1_bio->state);
1393 atomic_inc(&rdev->nr_pending);
1394 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1399 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1400 &first_bad, &bad_sectors);
1402 /* mustn't write here until the bad block is
1404 set_bit(BlockedBadBlocks, &rdev->flags);
1405 blocked_rdev = rdev;
1408 if (is_bad && first_bad <= r1_bio->sector) {
1409 /* Cannot write here at all */
1410 bad_sectors -= (r1_bio->sector - first_bad);
1411 if (bad_sectors < max_sectors)
1412 /* mustn't write more than bad_sectors
1413 * to other devices yet
1415 max_sectors = bad_sectors;
1416 rdev_dec_pending(rdev, mddev);
1417 /* We don't set R1BIO_Degraded as that
1418 * only applies if the disk is
1419 * missing, so it might be re-added,
1420 * and we want to know to recover this
1422 * In this case the device is here,
1423 * and the fact that this chunk is not
1424 * in-sync is recorded in the bad
1430 int good_sectors = first_bad - r1_bio->sector;
1431 if (good_sectors < max_sectors)
1432 max_sectors = good_sectors;
1435 r1_bio->bios[i] = bio;
1439 if (unlikely(blocked_rdev)) {
1440 /* Wait for this device to become unblocked */
1443 for (j = 0; j < i; j++)
1444 if (r1_bio->bios[j])
1445 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1447 allow_barrier(conf, bio->bi_iter.bi_sector);
1448 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1449 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1450 wait_barrier(conf, bio->bi_iter.bi_sector);
1454 if (max_sectors < bio_sectors(bio)) {
1455 struct bio *split = bio_split(bio, max_sectors,
1456 GFP_NOIO, &conf->bio_split);
1457 bio_chain(split, bio);
1458 submit_bio_noacct(bio);
1460 r1_bio->master_bio = bio;
1461 r1_bio->sectors = max_sectors;
1464 atomic_set(&r1_bio->remaining, 1);
1465 atomic_set(&r1_bio->behind_remaining, 0);
1469 for (i = 0; i < disks; i++) {
1470 struct bio *mbio = NULL;
1471 struct md_rdev *rdev = conf->mirrors[i].rdev;
1472 if (!r1_bio->bios[i])
1477 * Not if there are too many, or cannot
1478 * allocate memory, or a reader on WriteMostly
1479 * is waiting for behind writes to flush */
1481 (atomic_read(&bitmap->behind_writes)
1482 < mddev->bitmap_info.max_write_behind) &&
1483 !waitqueue_active(&bitmap->behind_wait)) {
1484 alloc_behind_master_bio(r1_bio, bio);
1487 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1488 test_bit(R1BIO_BehindIO, &r1_bio->state));
1492 if (r1_bio->behind_master_bio)
1493 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1494 GFP_NOIO, &mddev->bio_set);
1496 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1498 if (r1_bio->behind_master_bio) {
1499 if (test_bit(CollisionCheck, &rdev->flags))
1500 wait_for_serialization(rdev, r1_bio);
1501 if (test_bit(WriteMostly, &rdev->flags))
1502 atomic_inc(&r1_bio->behind_remaining);
1503 } else if (mddev->serialize_policy)
1504 wait_for_serialization(rdev, r1_bio);
1506 r1_bio->bios[i] = mbio;
1508 mbio->bi_iter.bi_sector = (r1_bio->sector +
1509 conf->mirrors[i].rdev->data_offset);
1510 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1511 mbio->bi_end_io = raid1_end_write_request;
1512 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1513 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1514 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1515 conf->raid_disks - mddev->degraded > 1)
1516 mbio->bi_opf |= MD_FAILFAST;
1517 mbio->bi_private = r1_bio;
1519 atomic_inc(&r1_bio->remaining);
1522 trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1524 /* flush_pending_writes() needs access to the rdev so...*/
1525 mbio->bi_bdev = (void *)conf->mirrors[i].rdev;
1527 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1529 plug = container_of(cb, struct raid1_plug_cb, cb);
1533 bio_list_add(&plug->pending, mbio);
1534 plug->pending_cnt++;
1536 spin_lock_irqsave(&conf->device_lock, flags);
1537 bio_list_add(&conf->pending_bio_list, mbio);
1538 conf->pending_count++;
1539 spin_unlock_irqrestore(&conf->device_lock, flags);
1540 md_wakeup_thread(mddev->thread);
1544 r1_bio_write_done(r1_bio);
1546 /* In case raid1d snuck in to freeze_array */
1547 wake_up(&conf->wait_barrier);
1550 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1554 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1555 && md_flush_request(mddev, bio))
1559 * There is a limit to the maximum size, but
1560 * the read/write handler might find a lower limit
1561 * due to bad blocks. To avoid multiple splits,
1562 * we pass the maximum number of sectors down
1563 * and let the lower level perform the split.
1565 sectors = align_to_barrier_unit_end(
1566 bio->bi_iter.bi_sector, bio_sectors(bio));
1568 if (bio_data_dir(bio) == READ)
1569 raid1_read_request(mddev, bio, sectors, NULL);
1571 if (!md_write_start(mddev,bio))
1573 raid1_write_request(mddev, bio, sectors);
1578 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1580 struct r1conf *conf = mddev->private;
1583 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1584 conf->raid_disks - mddev->degraded);
1586 for (i = 0; i < conf->raid_disks; i++) {
1587 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1588 seq_printf(seq, "%s",
1589 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1592 seq_printf(seq, "]");
1595 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1597 char b[BDEVNAME_SIZE];
1598 struct r1conf *conf = mddev->private;
1599 unsigned long flags;
1602 * If it is not operational, then we have already marked it as dead
1603 * else if it is the last working disks with "fail_last_dev == false",
1604 * ignore the error, let the next level up know.
1605 * else mark the drive as failed
1607 spin_lock_irqsave(&conf->device_lock, flags);
1608 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1609 && (conf->raid_disks - mddev->degraded) == 1) {
1611 * Don't fail the drive, act as though we were just a
1612 * normal single drive.
1613 * However don't try a recovery from this drive as
1614 * it is very likely to fail.
1616 conf->recovery_disabled = mddev->recovery_disabled;
1617 spin_unlock_irqrestore(&conf->device_lock, flags);
1620 set_bit(Blocked, &rdev->flags);
1621 if (test_and_clear_bit(In_sync, &rdev->flags))
1623 set_bit(Faulty, &rdev->flags);
1624 spin_unlock_irqrestore(&conf->device_lock, flags);
1626 * if recovery is running, make sure it aborts.
1628 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1629 set_mask_bits(&mddev->sb_flags, 0,
1630 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1631 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1632 "md/raid1:%s: Operation continuing on %d devices.\n",
1633 mdname(mddev), bdevname(rdev->bdev, b),
1634 mdname(mddev), conf->raid_disks - mddev->degraded);
1637 static void print_conf(struct r1conf *conf)
1641 pr_debug("RAID1 conf printout:\n");
1643 pr_debug("(!conf)\n");
1646 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1650 for (i = 0; i < conf->raid_disks; i++) {
1651 char b[BDEVNAME_SIZE];
1652 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1654 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1655 i, !test_bit(In_sync, &rdev->flags),
1656 !test_bit(Faulty, &rdev->flags),
1657 bdevname(rdev->bdev,b));
1662 static void close_sync(struct r1conf *conf)
1666 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1667 _wait_barrier(conf, idx);
1668 _allow_barrier(conf, idx);
1671 mempool_exit(&conf->r1buf_pool);
1674 static int raid1_spare_active(struct mddev *mddev)
1677 struct r1conf *conf = mddev->private;
1679 unsigned long flags;
1682 * Find all failed disks within the RAID1 configuration
1683 * and mark them readable.
1684 * Called under mddev lock, so rcu protection not needed.
1685 * device_lock used to avoid races with raid1_end_read_request
1686 * which expects 'In_sync' flags and ->degraded to be consistent.
1688 spin_lock_irqsave(&conf->device_lock, flags);
1689 for (i = 0; i < conf->raid_disks; i++) {
1690 struct md_rdev *rdev = conf->mirrors[i].rdev;
1691 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1693 && !test_bit(Candidate, &repl->flags)
1694 && repl->recovery_offset == MaxSector
1695 && !test_bit(Faulty, &repl->flags)
1696 && !test_and_set_bit(In_sync, &repl->flags)) {
1697 /* replacement has just become active */
1699 !test_and_clear_bit(In_sync, &rdev->flags))
1702 /* Replaced device not technically
1703 * faulty, but we need to be sure
1704 * it gets removed and never re-added
1706 set_bit(Faulty, &rdev->flags);
1707 sysfs_notify_dirent_safe(
1712 && rdev->recovery_offset == MaxSector
1713 && !test_bit(Faulty, &rdev->flags)
1714 && !test_and_set_bit(In_sync, &rdev->flags)) {
1716 sysfs_notify_dirent_safe(rdev->sysfs_state);
1719 mddev->degraded -= count;
1720 spin_unlock_irqrestore(&conf->device_lock, flags);
1726 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1728 struct r1conf *conf = mddev->private;
1731 struct raid1_info *p;
1733 int last = conf->raid_disks - 1;
1735 if (mddev->recovery_disabled == conf->recovery_disabled)
1738 if (md_integrity_add_rdev(rdev, mddev))
1741 if (rdev->raid_disk >= 0)
1742 first = last = rdev->raid_disk;
1745 * find the disk ... but prefer rdev->saved_raid_disk
1748 if (rdev->saved_raid_disk >= 0 &&
1749 rdev->saved_raid_disk >= first &&
1750 rdev->saved_raid_disk < conf->raid_disks &&
1751 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1752 first = last = rdev->saved_raid_disk;
1754 for (mirror = first; mirror <= last; mirror++) {
1755 p = conf->mirrors + mirror;
1758 disk_stack_limits(mddev->gendisk, rdev->bdev,
1759 rdev->data_offset << 9);
1761 p->head_position = 0;
1762 rdev->raid_disk = mirror;
1764 /* As all devices are equivalent, we don't need a full recovery
1765 * if this was recently any drive of the array
1767 if (rdev->saved_raid_disk < 0)
1769 rcu_assign_pointer(p->rdev, rdev);
1772 if (test_bit(WantReplacement, &p->rdev->flags) &&
1773 p[conf->raid_disks].rdev == NULL) {
1774 /* Add this device as a replacement */
1775 clear_bit(In_sync, &rdev->flags);
1776 set_bit(Replacement, &rdev->flags);
1777 rdev->raid_disk = mirror;
1780 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1784 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1785 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1790 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1792 struct r1conf *conf = mddev->private;
1794 int number = rdev->raid_disk;
1795 struct raid1_info *p = conf->mirrors + number;
1797 if (rdev != p->rdev)
1798 p = conf->mirrors + conf->raid_disks + number;
1801 if (rdev == p->rdev) {
1802 if (test_bit(In_sync, &rdev->flags) ||
1803 atomic_read(&rdev->nr_pending)) {
1807 /* Only remove non-faulty devices if recovery
1810 if (!test_bit(Faulty, &rdev->flags) &&
1811 mddev->recovery_disabled != conf->recovery_disabled &&
1812 mddev->degraded < conf->raid_disks) {
1817 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1819 if (atomic_read(&rdev->nr_pending)) {
1820 /* lost the race, try later */
1826 if (conf->mirrors[conf->raid_disks + number].rdev) {
1827 /* We just removed a device that is being replaced.
1828 * Move down the replacement. We drain all IO before
1829 * doing this to avoid confusion.
1831 struct md_rdev *repl =
1832 conf->mirrors[conf->raid_disks + number].rdev;
1833 freeze_array(conf, 0);
1834 if (atomic_read(&repl->nr_pending)) {
1835 /* It means that some queued IO of retry_list
1836 * hold repl. Thus, we cannot set replacement
1837 * as NULL, avoiding rdev NULL pointer
1838 * dereference in sync_request_write and
1839 * handle_write_finished.
1842 unfreeze_array(conf);
1845 clear_bit(Replacement, &repl->flags);
1847 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1848 unfreeze_array(conf);
1851 clear_bit(WantReplacement, &rdev->flags);
1852 err = md_integrity_register(mddev);
1860 static void end_sync_read(struct bio *bio)
1862 struct r1bio *r1_bio = get_resync_r1bio(bio);
1864 update_head_pos(r1_bio->read_disk, r1_bio);
1867 * we have read a block, now it needs to be re-written,
1868 * or re-read if the read failed.
1869 * We don't do much here, just schedule handling by raid1d
1871 if (!bio->bi_status)
1872 set_bit(R1BIO_Uptodate, &r1_bio->state);
1874 if (atomic_dec_and_test(&r1_bio->remaining))
1875 reschedule_retry(r1_bio);
1878 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1880 sector_t sync_blocks = 0;
1881 sector_t s = r1_bio->sector;
1882 long sectors_to_go = r1_bio->sectors;
1884 /* make sure these bits don't get cleared. */
1886 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1888 sectors_to_go -= sync_blocks;
1889 } while (sectors_to_go > 0);
1892 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1894 if (atomic_dec_and_test(&r1_bio->remaining)) {
1895 struct mddev *mddev = r1_bio->mddev;
1896 int s = r1_bio->sectors;
1898 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1899 test_bit(R1BIO_WriteError, &r1_bio->state))
1900 reschedule_retry(r1_bio);
1903 md_done_sync(mddev, s, uptodate);
1908 static void end_sync_write(struct bio *bio)
1910 int uptodate = !bio->bi_status;
1911 struct r1bio *r1_bio = get_resync_r1bio(bio);
1912 struct mddev *mddev = r1_bio->mddev;
1913 struct r1conf *conf = mddev->private;
1916 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1919 abort_sync_write(mddev, r1_bio);
1920 set_bit(WriteErrorSeen, &rdev->flags);
1921 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1922 set_bit(MD_RECOVERY_NEEDED, &
1924 set_bit(R1BIO_WriteError, &r1_bio->state);
1925 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1926 &first_bad, &bad_sectors) &&
1927 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1930 &first_bad, &bad_sectors)
1932 set_bit(R1BIO_MadeGood, &r1_bio->state);
1934 put_sync_write_buf(r1_bio, uptodate);
1937 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1938 int sectors, struct page *page, int rw)
1940 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1944 set_bit(WriteErrorSeen, &rdev->flags);
1945 if (!test_and_set_bit(WantReplacement,
1947 set_bit(MD_RECOVERY_NEEDED, &
1948 rdev->mddev->recovery);
1950 /* need to record an error - either for the block or the device */
1951 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1952 md_error(rdev->mddev, rdev);
1956 static int fix_sync_read_error(struct r1bio *r1_bio)
1958 /* Try some synchronous reads of other devices to get
1959 * good data, much like with normal read errors. Only
1960 * read into the pages we already have so we don't
1961 * need to re-issue the read request.
1962 * We don't need to freeze the array, because being in an
1963 * active sync request, there is no normal IO, and
1964 * no overlapping syncs.
1965 * We don't need to check is_badblock() again as we
1966 * made sure that anything with a bad block in range
1967 * will have bi_end_io clear.
1969 struct mddev *mddev = r1_bio->mddev;
1970 struct r1conf *conf = mddev->private;
1971 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1972 struct page **pages = get_resync_pages(bio)->pages;
1973 sector_t sect = r1_bio->sector;
1974 int sectors = r1_bio->sectors;
1976 struct md_rdev *rdev;
1978 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1979 if (test_bit(FailFast, &rdev->flags)) {
1980 /* Don't try recovering from here - just fail it
1981 * ... unless it is the last working device of course */
1982 md_error(mddev, rdev);
1983 if (test_bit(Faulty, &rdev->flags))
1984 /* Don't try to read from here, but make sure
1985 * put_buf does it's thing
1987 bio->bi_end_io = end_sync_write;
1992 int d = r1_bio->read_disk;
1996 if (s > (PAGE_SIZE>>9))
1999 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2000 /* No rcu protection needed here devices
2001 * can only be removed when no resync is
2002 * active, and resync is currently active
2004 rdev = conf->mirrors[d].rdev;
2005 if (sync_page_io(rdev, sect, s<<9,
2007 REQ_OP_READ, 0, false)) {
2013 if (d == conf->raid_disks * 2)
2015 } while (!success && d != r1_bio->read_disk);
2018 char b[BDEVNAME_SIZE];
2020 /* Cannot read from anywhere, this block is lost.
2021 * Record a bad block on each device. If that doesn't
2022 * work just disable and interrupt the recovery.
2023 * Don't fail devices as that won't really help.
2025 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2026 mdname(mddev), bio_devname(bio, b),
2027 (unsigned long long)r1_bio->sector);
2028 for (d = 0; d < conf->raid_disks * 2; d++) {
2029 rdev = conf->mirrors[d].rdev;
2030 if (!rdev || test_bit(Faulty, &rdev->flags))
2032 if (!rdev_set_badblocks(rdev, sect, s, 0))
2036 conf->recovery_disabled =
2037 mddev->recovery_disabled;
2038 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2039 md_done_sync(mddev, r1_bio->sectors, 0);
2051 /* write it back and re-read */
2052 while (d != r1_bio->read_disk) {
2054 d = conf->raid_disks * 2;
2056 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2058 rdev = conf->mirrors[d].rdev;
2059 if (r1_sync_page_io(rdev, sect, s,
2062 r1_bio->bios[d]->bi_end_io = NULL;
2063 rdev_dec_pending(rdev, mddev);
2067 while (d != r1_bio->read_disk) {
2069 d = conf->raid_disks * 2;
2071 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2073 rdev = conf->mirrors[d].rdev;
2074 if (r1_sync_page_io(rdev, sect, s,
2077 atomic_add(s, &rdev->corrected_errors);
2083 set_bit(R1BIO_Uptodate, &r1_bio->state);
2088 static void process_checks(struct r1bio *r1_bio)
2090 /* We have read all readable devices. If we haven't
2091 * got the block, then there is no hope left.
2092 * If we have, then we want to do a comparison
2093 * and skip the write if everything is the same.
2094 * If any blocks failed to read, then we need to
2095 * attempt an over-write
2097 struct mddev *mddev = r1_bio->mddev;
2098 struct r1conf *conf = mddev->private;
2103 /* Fix variable parts of all bios */
2104 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2105 for (i = 0; i < conf->raid_disks * 2; i++) {
2106 blk_status_t status;
2107 struct bio *b = r1_bio->bios[i];
2108 struct resync_pages *rp = get_resync_pages(b);
2109 if (b->bi_end_io != end_sync_read)
2111 /* fixup the bio for reuse, but preserve errno */
2112 status = b->bi_status;
2114 b->bi_status = status;
2115 b->bi_iter.bi_sector = r1_bio->sector +
2116 conf->mirrors[i].rdev->data_offset;
2117 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2118 b->bi_end_io = end_sync_read;
2119 rp->raid_bio = r1_bio;
2122 /* initialize bvec table again */
2123 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2125 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2126 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2127 !r1_bio->bios[primary]->bi_status) {
2128 r1_bio->bios[primary]->bi_end_io = NULL;
2129 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2132 r1_bio->read_disk = primary;
2133 for (i = 0; i < conf->raid_disks * 2; i++) {
2135 struct bio *pbio = r1_bio->bios[primary];
2136 struct bio *sbio = r1_bio->bios[i];
2137 blk_status_t status = sbio->bi_status;
2138 struct page **ppages = get_resync_pages(pbio)->pages;
2139 struct page **spages = get_resync_pages(sbio)->pages;
2141 int page_len[RESYNC_PAGES] = { 0 };
2142 struct bvec_iter_all iter_all;
2144 if (sbio->bi_end_io != end_sync_read)
2146 /* Now we can 'fixup' the error value */
2147 sbio->bi_status = 0;
2149 bio_for_each_segment_all(bi, sbio, iter_all)
2150 page_len[j++] = bi->bv_len;
2153 for (j = vcnt; j-- ; ) {
2154 if (memcmp(page_address(ppages[j]),
2155 page_address(spages[j]),
2162 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2163 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2165 /* No need to write to this device. */
2166 sbio->bi_end_io = NULL;
2167 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2171 bio_copy_data(sbio, pbio);
2175 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2177 struct r1conf *conf = mddev->private;
2179 int disks = conf->raid_disks * 2;
2182 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2183 /* ouch - failed to read all of that. */
2184 if (!fix_sync_read_error(r1_bio))
2187 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2188 process_checks(r1_bio);
2193 atomic_set(&r1_bio->remaining, 1);
2194 for (i = 0; i < disks ; i++) {
2195 wbio = r1_bio->bios[i];
2196 if (wbio->bi_end_io == NULL ||
2197 (wbio->bi_end_io == end_sync_read &&
2198 (i == r1_bio->read_disk ||
2199 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2201 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2202 abort_sync_write(mddev, r1_bio);
2206 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2207 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2208 wbio->bi_opf |= MD_FAILFAST;
2210 wbio->bi_end_io = end_sync_write;
2211 atomic_inc(&r1_bio->remaining);
2212 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2214 submit_bio_noacct(wbio);
2217 put_sync_write_buf(r1_bio, 1);
2221 * This is a kernel thread which:
2223 * 1. Retries failed read operations on working mirrors.
2224 * 2. Updates the raid superblock when problems encounter.
2225 * 3. Performs writes following reads for array synchronising.
2228 static void fix_read_error(struct r1conf *conf, int read_disk,
2229 sector_t sect, int sectors)
2231 struct mddev *mddev = conf->mddev;
2237 struct md_rdev *rdev;
2239 if (s > (PAGE_SIZE>>9))
2247 rdev = rcu_dereference(conf->mirrors[d].rdev);
2249 (test_bit(In_sync, &rdev->flags) ||
2250 (!test_bit(Faulty, &rdev->flags) &&
2251 rdev->recovery_offset >= sect + s)) &&
2252 is_badblock(rdev, sect, s,
2253 &first_bad, &bad_sectors) == 0) {
2254 atomic_inc(&rdev->nr_pending);
2256 if (sync_page_io(rdev, sect, s<<9,
2257 conf->tmppage, REQ_OP_READ, 0, false))
2259 rdev_dec_pending(rdev, mddev);
2265 if (d == conf->raid_disks * 2)
2267 } while (!success && d != read_disk);
2270 /* Cannot read from anywhere - mark it bad */
2271 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2272 if (!rdev_set_badblocks(rdev, sect, s, 0))
2273 md_error(mddev, rdev);
2276 /* write it back and re-read */
2278 while (d != read_disk) {
2280 d = conf->raid_disks * 2;
2283 rdev = rcu_dereference(conf->mirrors[d].rdev);
2285 !test_bit(Faulty, &rdev->flags)) {
2286 atomic_inc(&rdev->nr_pending);
2288 r1_sync_page_io(rdev, sect, s,
2289 conf->tmppage, WRITE);
2290 rdev_dec_pending(rdev, mddev);
2295 while (d != read_disk) {
2296 char b[BDEVNAME_SIZE];
2298 d = conf->raid_disks * 2;
2301 rdev = rcu_dereference(conf->mirrors[d].rdev);
2303 !test_bit(Faulty, &rdev->flags)) {
2304 atomic_inc(&rdev->nr_pending);
2306 if (r1_sync_page_io(rdev, sect, s,
2307 conf->tmppage, READ)) {
2308 atomic_add(s, &rdev->corrected_errors);
2309 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2311 (unsigned long long)(sect +
2313 bdevname(rdev->bdev, b));
2315 rdev_dec_pending(rdev, mddev);
2324 static int narrow_write_error(struct r1bio *r1_bio, int i)
2326 struct mddev *mddev = r1_bio->mddev;
2327 struct r1conf *conf = mddev->private;
2328 struct md_rdev *rdev = conf->mirrors[i].rdev;
2330 /* bio has the data to be written to device 'i' where
2331 * we just recently had a write error.
2332 * We repeatedly clone the bio and trim down to one block,
2333 * then try the write. Where the write fails we record
2335 * It is conceivable that the bio doesn't exactly align with
2336 * blocks. We must handle this somehow.
2338 * We currently own a reference on the rdev.
2344 int sect_to_write = r1_bio->sectors;
2347 if (rdev->badblocks.shift < 0)
2350 block_sectors = roundup(1 << rdev->badblocks.shift,
2351 bdev_logical_block_size(rdev->bdev) >> 9);
2352 sector = r1_bio->sector;
2353 sectors = ((sector + block_sectors)
2354 & ~(sector_t)(block_sectors - 1))
2357 while (sect_to_write) {
2359 if (sectors > sect_to_write)
2360 sectors = sect_to_write;
2361 /* Write at 'sector' for 'sectors'*/
2363 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2364 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2368 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2372 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2373 wbio->bi_iter.bi_sector = r1_bio->sector;
2374 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2376 bio_trim(wbio, sector - r1_bio->sector, sectors);
2377 wbio->bi_iter.bi_sector += rdev->data_offset;
2378 bio_set_dev(wbio, rdev->bdev);
2380 if (submit_bio_wait(wbio) < 0)
2382 ok = rdev_set_badblocks(rdev, sector,
2387 sect_to_write -= sectors;
2389 sectors = block_sectors;
2394 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2397 int s = r1_bio->sectors;
2398 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2399 struct md_rdev *rdev = conf->mirrors[m].rdev;
2400 struct bio *bio = r1_bio->bios[m];
2401 if (bio->bi_end_io == NULL)
2403 if (!bio->bi_status &&
2404 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2405 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2407 if (bio->bi_status &&
2408 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2409 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2410 md_error(conf->mddev, rdev);
2414 md_done_sync(conf->mddev, s, 1);
2417 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2422 for (m = 0; m < conf->raid_disks * 2 ; m++)
2423 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2424 struct md_rdev *rdev = conf->mirrors[m].rdev;
2425 rdev_clear_badblocks(rdev,
2427 r1_bio->sectors, 0);
2428 rdev_dec_pending(rdev, conf->mddev);
2429 } else if (r1_bio->bios[m] != NULL) {
2430 /* This drive got a write error. We need to
2431 * narrow down and record precise write
2435 if (!narrow_write_error(r1_bio, m)) {
2436 md_error(conf->mddev,
2437 conf->mirrors[m].rdev);
2438 /* an I/O failed, we can't clear the bitmap */
2439 set_bit(R1BIO_Degraded, &r1_bio->state);
2441 rdev_dec_pending(conf->mirrors[m].rdev,
2445 spin_lock_irq(&conf->device_lock);
2446 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2447 idx = sector_to_idx(r1_bio->sector);
2448 atomic_inc(&conf->nr_queued[idx]);
2449 spin_unlock_irq(&conf->device_lock);
2451 * In case freeze_array() is waiting for condition
2452 * get_unqueued_pending() == extra to be true.
2454 wake_up(&conf->wait_barrier);
2455 md_wakeup_thread(conf->mddev->thread);
2457 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2458 close_write(r1_bio);
2459 raid_end_bio_io(r1_bio);
2463 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2465 struct mddev *mddev = conf->mddev;
2467 struct md_rdev *rdev;
2469 clear_bit(R1BIO_ReadError, &r1_bio->state);
2470 /* we got a read error. Maybe the drive is bad. Maybe just
2471 * the block and we can fix it.
2472 * We freeze all other IO, and try reading the block from
2473 * other devices. When we find one, we re-write
2474 * and check it that fixes the read error.
2475 * This is all done synchronously while the array is
2479 bio = r1_bio->bios[r1_bio->read_disk];
2481 r1_bio->bios[r1_bio->read_disk] = NULL;
2483 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2485 && !test_bit(FailFast, &rdev->flags)) {
2486 freeze_array(conf, 1);
2487 fix_read_error(conf, r1_bio->read_disk,
2488 r1_bio->sector, r1_bio->sectors);
2489 unfreeze_array(conf);
2490 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2491 md_error(mddev, rdev);
2493 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2496 rdev_dec_pending(rdev, conf->mddev);
2497 allow_barrier(conf, r1_bio->sector);
2498 bio = r1_bio->master_bio;
2500 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2502 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2505 static void raid1d(struct md_thread *thread)
2507 struct mddev *mddev = thread->mddev;
2508 struct r1bio *r1_bio;
2509 unsigned long flags;
2510 struct r1conf *conf = mddev->private;
2511 struct list_head *head = &conf->retry_list;
2512 struct blk_plug plug;
2515 md_check_recovery(mddev);
2517 if (!list_empty_careful(&conf->bio_end_io_list) &&
2518 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2520 spin_lock_irqsave(&conf->device_lock, flags);
2521 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2522 list_splice_init(&conf->bio_end_io_list, &tmp);
2523 spin_unlock_irqrestore(&conf->device_lock, flags);
2524 while (!list_empty(&tmp)) {
2525 r1_bio = list_first_entry(&tmp, struct r1bio,
2527 list_del(&r1_bio->retry_list);
2528 idx = sector_to_idx(r1_bio->sector);
2529 atomic_dec(&conf->nr_queued[idx]);
2530 if (mddev->degraded)
2531 set_bit(R1BIO_Degraded, &r1_bio->state);
2532 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2533 close_write(r1_bio);
2534 raid_end_bio_io(r1_bio);
2538 blk_start_plug(&plug);
2541 flush_pending_writes(conf);
2543 spin_lock_irqsave(&conf->device_lock, flags);
2544 if (list_empty(head)) {
2545 spin_unlock_irqrestore(&conf->device_lock, flags);
2548 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2549 list_del(head->prev);
2550 idx = sector_to_idx(r1_bio->sector);
2551 atomic_dec(&conf->nr_queued[idx]);
2552 spin_unlock_irqrestore(&conf->device_lock, flags);
2554 mddev = r1_bio->mddev;
2555 conf = mddev->private;
2556 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2557 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2558 test_bit(R1BIO_WriteError, &r1_bio->state))
2559 handle_sync_write_finished(conf, r1_bio);
2561 sync_request_write(mddev, r1_bio);
2562 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2563 test_bit(R1BIO_WriteError, &r1_bio->state))
2564 handle_write_finished(conf, r1_bio);
2565 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2566 handle_read_error(conf, r1_bio);
2571 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2572 md_check_recovery(mddev);
2574 blk_finish_plug(&plug);
2577 static int init_resync(struct r1conf *conf)
2581 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2582 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2584 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2585 r1buf_pool_free, conf->poolinfo);
2588 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2590 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2591 struct resync_pages *rps;
2595 for (i = conf->poolinfo->raid_disks; i--; ) {
2596 bio = r1bio->bios[i];
2597 rps = bio->bi_private;
2599 bio->bi_private = rps;
2601 r1bio->master_bio = NULL;
2606 * perform a "sync" on one "block"
2608 * We need to make sure that no normal I/O request - particularly write
2609 * requests - conflict with active sync requests.
2611 * This is achieved by tracking pending requests and a 'barrier' concept
2612 * that can be installed to exclude normal IO requests.
2615 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2618 struct r1conf *conf = mddev->private;
2619 struct r1bio *r1_bio;
2621 sector_t max_sector, nr_sectors;
2625 int write_targets = 0, read_targets = 0;
2626 sector_t sync_blocks;
2627 int still_degraded = 0;
2628 int good_sectors = RESYNC_SECTORS;
2629 int min_bad = 0; /* number of sectors that are bad in all devices */
2630 int idx = sector_to_idx(sector_nr);
2633 if (!mempool_initialized(&conf->r1buf_pool))
2634 if (init_resync(conf))
2637 max_sector = mddev->dev_sectors;
2638 if (sector_nr >= max_sector) {
2639 /* If we aborted, we need to abort the
2640 * sync on the 'current' bitmap chunk (there will
2641 * only be one in raid1 resync.
2642 * We can find the current addess in mddev->curr_resync
2644 if (mddev->curr_resync < max_sector) /* aborted */
2645 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2647 else /* completed sync */
2650 md_bitmap_close_sync(mddev->bitmap);
2653 if (mddev_is_clustered(mddev)) {
2654 conf->cluster_sync_low = 0;
2655 conf->cluster_sync_high = 0;
2660 if (mddev->bitmap == NULL &&
2661 mddev->recovery_cp == MaxSector &&
2662 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2663 conf->fullsync == 0) {
2665 return max_sector - sector_nr;
2667 /* before building a request, check if we can skip these blocks..
2668 * This call the bitmap_start_sync doesn't actually record anything
2670 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2671 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2672 /* We can skip this block, and probably several more */
2678 * If there is non-resync activity waiting for a turn, then let it
2679 * though before starting on this new sync request.
2681 if (atomic_read(&conf->nr_waiting[idx]))
2682 schedule_timeout_uninterruptible(1);
2684 /* we are incrementing sector_nr below. To be safe, we check against
2685 * sector_nr + two times RESYNC_SECTORS
2688 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2689 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2692 if (raise_barrier(conf, sector_nr))
2695 r1_bio = raid1_alloc_init_r1buf(conf);
2699 * If we get a correctably read error during resync or recovery,
2700 * we might want to read from a different device. So we
2701 * flag all drives that could conceivably be read from for READ,
2702 * and any others (which will be non-In_sync devices) for WRITE.
2703 * If a read fails, we try reading from something else for which READ
2707 r1_bio->mddev = mddev;
2708 r1_bio->sector = sector_nr;
2710 set_bit(R1BIO_IsSync, &r1_bio->state);
2711 /* make sure good_sectors won't go across barrier unit boundary */
2712 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2714 for (i = 0; i < conf->raid_disks * 2; i++) {
2715 struct md_rdev *rdev;
2716 bio = r1_bio->bios[i];
2718 rdev = rcu_dereference(conf->mirrors[i].rdev);
2720 test_bit(Faulty, &rdev->flags)) {
2721 if (i < conf->raid_disks)
2723 } else if (!test_bit(In_sync, &rdev->flags)) {
2724 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2725 bio->bi_end_io = end_sync_write;
2728 /* may need to read from here */
2729 sector_t first_bad = MaxSector;
2732 if (is_badblock(rdev, sector_nr, good_sectors,
2733 &first_bad, &bad_sectors)) {
2734 if (first_bad > sector_nr)
2735 good_sectors = first_bad - sector_nr;
2737 bad_sectors -= (sector_nr - first_bad);
2739 min_bad > bad_sectors)
2740 min_bad = bad_sectors;
2743 if (sector_nr < first_bad) {
2744 if (test_bit(WriteMostly, &rdev->flags)) {
2751 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2752 bio->bi_end_io = end_sync_read;
2754 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2755 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2756 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2758 * The device is suitable for reading (InSync),
2759 * but has bad block(s) here. Let's try to correct them,
2760 * if we are doing resync or repair. Otherwise, leave
2761 * this device alone for this sync request.
2763 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2764 bio->bi_end_io = end_sync_write;
2768 if (rdev && bio->bi_end_io) {
2769 atomic_inc(&rdev->nr_pending);
2770 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2771 bio_set_dev(bio, rdev->bdev);
2772 if (test_bit(FailFast, &rdev->flags))
2773 bio->bi_opf |= MD_FAILFAST;
2779 r1_bio->read_disk = disk;
2781 if (read_targets == 0 && min_bad > 0) {
2782 /* These sectors are bad on all InSync devices, so we
2783 * need to mark them bad on all write targets
2786 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2787 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2788 struct md_rdev *rdev = conf->mirrors[i].rdev;
2789 ok = rdev_set_badblocks(rdev, sector_nr,
2793 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2798 /* Cannot record the badblocks, so need to
2800 * If there are multiple read targets, could just
2801 * fail the really bad ones ???
2803 conf->recovery_disabled = mddev->recovery_disabled;
2804 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2810 if (min_bad > 0 && min_bad < good_sectors) {
2811 /* only resync enough to reach the next bad->good
2813 good_sectors = min_bad;
2816 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2817 /* extra read targets are also write targets */
2818 write_targets += read_targets-1;
2820 if (write_targets == 0 || read_targets == 0) {
2821 /* There is nowhere to write, so all non-sync
2822 * drives must be failed - so we are finished
2826 max_sector = sector_nr + min_bad;
2827 rv = max_sector - sector_nr;
2833 if (max_sector > mddev->resync_max)
2834 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2835 if (max_sector > sector_nr + good_sectors)
2836 max_sector = sector_nr + good_sectors;
2841 int len = PAGE_SIZE;
2842 if (sector_nr + (len>>9) > max_sector)
2843 len = (max_sector - sector_nr) << 9;
2846 if (sync_blocks == 0) {
2847 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2848 &sync_blocks, still_degraded) &&
2850 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2852 if ((len >> 9) > sync_blocks)
2853 len = sync_blocks<<9;
2856 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2857 struct resync_pages *rp;
2859 bio = r1_bio->bios[i];
2860 rp = get_resync_pages(bio);
2861 if (bio->bi_end_io) {
2862 page = resync_fetch_page(rp, page_idx);
2865 * won't fail because the vec table is big
2866 * enough to hold all these pages
2868 bio_add_page(bio, page, len, 0);
2871 nr_sectors += len>>9;
2872 sector_nr += len>>9;
2873 sync_blocks -= (len>>9);
2874 } while (++page_idx < RESYNC_PAGES);
2876 r1_bio->sectors = nr_sectors;
2878 if (mddev_is_clustered(mddev) &&
2879 conf->cluster_sync_high < sector_nr + nr_sectors) {
2880 conf->cluster_sync_low = mddev->curr_resync_completed;
2881 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2882 /* Send resync message */
2883 md_cluster_ops->resync_info_update(mddev,
2884 conf->cluster_sync_low,
2885 conf->cluster_sync_high);
2888 /* For a user-requested sync, we read all readable devices and do a
2891 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2892 atomic_set(&r1_bio->remaining, read_targets);
2893 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2894 bio = r1_bio->bios[i];
2895 if (bio->bi_end_io == end_sync_read) {
2897 md_sync_acct_bio(bio, nr_sectors);
2898 if (read_targets == 1)
2899 bio->bi_opf &= ~MD_FAILFAST;
2900 submit_bio_noacct(bio);
2904 atomic_set(&r1_bio->remaining, 1);
2905 bio = r1_bio->bios[r1_bio->read_disk];
2906 md_sync_acct_bio(bio, nr_sectors);
2907 if (read_targets == 1)
2908 bio->bi_opf &= ~MD_FAILFAST;
2909 submit_bio_noacct(bio);
2914 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2919 return mddev->dev_sectors;
2922 static struct r1conf *setup_conf(struct mddev *mddev)
2924 struct r1conf *conf;
2926 struct raid1_info *disk;
2927 struct md_rdev *rdev;
2930 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2934 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2935 sizeof(atomic_t), GFP_KERNEL);
2936 if (!conf->nr_pending)
2939 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2940 sizeof(atomic_t), GFP_KERNEL);
2941 if (!conf->nr_waiting)
2944 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2945 sizeof(atomic_t), GFP_KERNEL);
2946 if (!conf->nr_queued)
2949 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2950 sizeof(atomic_t), GFP_KERNEL);
2954 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2955 mddev->raid_disks, 2),
2960 conf->tmppage = alloc_page(GFP_KERNEL);
2964 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2965 if (!conf->poolinfo)
2967 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2968 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2969 rbio_pool_free, conf->poolinfo);
2973 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2977 conf->poolinfo->mddev = mddev;
2980 spin_lock_init(&conf->device_lock);
2981 rdev_for_each(rdev, mddev) {
2982 int disk_idx = rdev->raid_disk;
2983 if (disk_idx >= mddev->raid_disks
2986 if (test_bit(Replacement, &rdev->flags))
2987 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2989 disk = conf->mirrors + disk_idx;
2994 disk->head_position = 0;
2995 disk->seq_start = MaxSector;
2997 conf->raid_disks = mddev->raid_disks;
2998 conf->mddev = mddev;
2999 INIT_LIST_HEAD(&conf->retry_list);
3000 INIT_LIST_HEAD(&conf->bio_end_io_list);
3002 spin_lock_init(&conf->resync_lock);
3003 init_waitqueue_head(&conf->wait_barrier);
3005 bio_list_init(&conf->pending_bio_list);
3006 conf->pending_count = 0;
3007 conf->recovery_disabled = mddev->recovery_disabled - 1;
3010 for (i = 0; i < conf->raid_disks * 2; i++) {
3012 disk = conf->mirrors + i;
3014 if (i < conf->raid_disks &&
3015 disk[conf->raid_disks].rdev) {
3016 /* This slot has a replacement. */
3018 /* No original, just make the replacement
3019 * a recovering spare
3022 disk[conf->raid_disks].rdev;
3023 disk[conf->raid_disks].rdev = NULL;
3024 } else if (!test_bit(In_sync, &disk->rdev->flags))
3025 /* Original is not in_sync - bad */
3030 !test_bit(In_sync, &disk->rdev->flags)) {
3031 disk->head_position = 0;
3033 (disk->rdev->saved_raid_disk < 0))
3039 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3047 mempool_exit(&conf->r1bio_pool);
3048 kfree(conf->mirrors);
3049 safe_put_page(conf->tmppage);
3050 kfree(conf->poolinfo);
3051 kfree(conf->nr_pending);
3052 kfree(conf->nr_waiting);
3053 kfree(conf->nr_queued);
3054 kfree(conf->barrier);
3055 bioset_exit(&conf->bio_split);
3058 return ERR_PTR(err);
3061 static void raid1_free(struct mddev *mddev, void *priv);
3062 static int raid1_run(struct mddev *mddev)
3064 struct r1conf *conf;
3066 struct md_rdev *rdev;
3068 bool discard_supported = false;
3070 if (mddev->level != 1) {
3071 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3072 mdname(mddev), mddev->level);
3075 if (mddev->reshape_position != MaxSector) {
3076 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3080 if (mddev_init_writes_pending(mddev) < 0)
3083 * copy the already verified devices into our private RAID1
3084 * bookkeeping area. [whatever we allocate in run(),
3085 * should be freed in raid1_free()]
3087 if (mddev->private == NULL)
3088 conf = setup_conf(mddev);
3090 conf = mddev->private;
3093 return PTR_ERR(conf);
3096 blk_queue_max_write_same_sectors(mddev->queue, 0);
3097 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3100 rdev_for_each(rdev, mddev) {
3101 if (!mddev->gendisk)
3103 disk_stack_limits(mddev->gendisk, rdev->bdev,
3104 rdev->data_offset << 9);
3105 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3106 discard_supported = true;
3109 mddev->degraded = 0;
3110 for (i = 0; i < conf->raid_disks; i++)
3111 if (conf->mirrors[i].rdev == NULL ||
3112 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3113 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3116 * RAID1 needs at least one disk in active
3118 if (conf->raid_disks - mddev->degraded < 1) {
3123 if (conf->raid_disks - mddev->degraded == 1)
3124 mddev->recovery_cp = MaxSector;
3126 if (mddev->recovery_cp != MaxSector)
3127 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3129 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3130 mdname(mddev), mddev->raid_disks - mddev->degraded,
3134 * Ok, everything is just fine now
3136 mddev->thread = conf->thread;
3137 conf->thread = NULL;
3138 mddev->private = conf;
3139 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3141 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3144 if (discard_supported)
3145 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3148 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3152 ret = md_integrity_register(mddev);
3154 md_unregister_thread(&mddev->thread);
3160 raid1_free(mddev, conf);
3164 static void raid1_free(struct mddev *mddev, void *priv)
3166 struct r1conf *conf = priv;
3168 mempool_exit(&conf->r1bio_pool);
3169 kfree(conf->mirrors);
3170 safe_put_page(conf->tmppage);
3171 kfree(conf->poolinfo);
3172 kfree(conf->nr_pending);
3173 kfree(conf->nr_waiting);
3174 kfree(conf->nr_queued);
3175 kfree(conf->barrier);
3176 bioset_exit(&conf->bio_split);
3180 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3182 /* no resync is happening, and there is enough space
3183 * on all devices, so we can resize.
3184 * We need to make sure resync covers any new space.
3185 * If the array is shrinking we should possibly wait until
3186 * any io in the removed space completes, but it hardly seems
3189 sector_t newsize = raid1_size(mddev, sectors, 0);
3190 if (mddev->external_size &&
3191 mddev->array_sectors > newsize)
3193 if (mddev->bitmap) {
3194 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3198 md_set_array_sectors(mddev, newsize);
3199 if (sectors > mddev->dev_sectors &&
3200 mddev->recovery_cp > mddev->dev_sectors) {
3201 mddev->recovery_cp = mddev->dev_sectors;
3202 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3204 mddev->dev_sectors = sectors;
3205 mddev->resync_max_sectors = sectors;
3209 static int raid1_reshape(struct mddev *mddev)
3212 * 1/ resize the r1bio_pool
3213 * 2/ resize conf->mirrors
3215 * We allocate a new r1bio_pool if we can.
3216 * Then raise a device barrier and wait until all IO stops.
3217 * Then resize conf->mirrors and swap in the new r1bio pool.
3219 * At the same time, we "pack" the devices so that all the missing
3220 * devices have the higher raid_disk numbers.
3222 mempool_t newpool, oldpool;
3223 struct pool_info *newpoolinfo;
3224 struct raid1_info *newmirrors;
3225 struct r1conf *conf = mddev->private;
3226 int cnt, raid_disks;
3227 unsigned long flags;
3231 memset(&newpool, 0, sizeof(newpool));
3232 memset(&oldpool, 0, sizeof(oldpool));
3234 /* Cannot change chunk_size, layout, or level */
3235 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3236 mddev->layout != mddev->new_layout ||
3237 mddev->level != mddev->new_level) {
3238 mddev->new_chunk_sectors = mddev->chunk_sectors;
3239 mddev->new_layout = mddev->layout;
3240 mddev->new_level = mddev->level;
3244 if (!mddev_is_clustered(mddev))
3245 md_allow_write(mddev);
3247 raid_disks = mddev->raid_disks + mddev->delta_disks;
3249 if (raid_disks < conf->raid_disks) {
3251 for (d= 0; d < conf->raid_disks; d++)
3252 if (conf->mirrors[d].rdev)
3254 if (cnt > raid_disks)
3258 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3261 newpoolinfo->mddev = mddev;
3262 newpoolinfo->raid_disks = raid_disks * 2;
3264 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3265 rbio_pool_free, newpoolinfo);
3270 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3275 mempool_exit(&newpool);
3279 freeze_array(conf, 0);
3281 /* ok, everything is stopped */
3282 oldpool = conf->r1bio_pool;
3283 conf->r1bio_pool = newpool;
3285 for (d = d2 = 0; d < conf->raid_disks; d++) {
3286 struct md_rdev *rdev = conf->mirrors[d].rdev;
3287 if (rdev && rdev->raid_disk != d2) {
3288 sysfs_unlink_rdev(mddev, rdev);
3289 rdev->raid_disk = d2;
3290 sysfs_unlink_rdev(mddev, rdev);
3291 if (sysfs_link_rdev(mddev, rdev))
3292 pr_warn("md/raid1:%s: cannot register rd%d\n",
3293 mdname(mddev), rdev->raid_disk);
3296 newmirrors[d2++].rdev = rdev;
3298 kfree(conf->mirrors);
3299 conf->mirrors = newmirrors;
3300 kfree(conf->poolinfo);
3301 conf->poolinfo = newpoolinfo;
3303 spin_lock_irqsave(&conf->device_lock, flags);
3304 mddev->degraded += (raid_disks - conf->raid_disks);
3305 spin_unlock_irqrestore(&conf->device_lock, flags);
3306 conf->raid_disks = mddev->raid_disks = raid_disks;
3307 mddev->delta_disks = 0;
3309 unfreeze_array(conf);
3311 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3312 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3313 md_wakeup_thread(mddev->thread);
3315 mempool_exit(&oldpool);
3319 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3321 struct r1conf *conf = mddev->private;
3324 freeze_array(conf, 0);
3326 unfreeze_array(conf);
3329 static void *raid1_takeover(struct mddev *mddev)
3331 /* raid1 can take over:
3332 * raid5 with 2 devices, any layout or chunk size
3334 if (mddev->level == 5 && mddev->raid_disks == 2) {
3335 struct r1conf *conf;
3336 mddev->new_level = 1;
3337 mddev->new_layout = 0;
3338 mddev->new_chunk_sectors = 0;
3339 conf = setup_conf(mddev);
3340 if (!IS_ERR(conf)) {
3341 /* Array must appear to be quiesced */
3342 conf->array_frozen = 1;
3343 mddev_clear_unsupported_flags(mddev,
3344 UNSUPPORTED_MDDEV_FLAGS);
3348 return ERR_PTR(-EINVAL);
3351 static struct md_personality raid1_personality =
3355 .owner = THIS_MODULE,
3356 .make_request = raid1_make_request,
3359 .status = raid1_status,
3360 .error_handler = raid1_error,
3361 .hot_add_disk = raid1_add_disk,
3362 .hot_remove_disk= raid1_remove_disk,
3363 .spare_active = raid1_spare_active,
3364 .sync_request = raid1_sync_request,
3365 .resize = raid1_resize,
3367 .check_reshape = raid1_reshape,
3368 .quiesce = raid1_quiesce,
3369 .takeover = raid1_takeover,
3372 static int __init raid_init(void)
3374 return register_md_personality(&raid1_personality);
3377 static void raid_exit(void)
3379 unregister_md_personality(&raid1_personality);
3382 module_init(raid_init);
3383 module_exit(raid_exit);
3384 MODULE_LICENSE("GPL");
3385 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3386 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3387 MODULE_ALIAS("md-raid1");
3388 MODULE_ALIAS("md-level-1");
3390 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);