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;
303 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
304 bio_end_io_acct(bio, r1_bio->start_time);
308 static void raid_end_bio_io(struct r1bio *r1_bio)
310 struct bio *bio = r1_bio->master_bio;
311 struct r1conf *conf = r1_bio->mddev->private;
313 /* if nobody has done the final endio yet, do it now */
314 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
315 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
316 (bio_data_dir(bio) == WRITE) ? "write" : "read",
317 (unsigned long long) bio->bi_iter.bi_sector,
318 (unsigned long long) bio_end_sector(bio) - 1);
320 call_bio_endio(r1_bio);
323 * Wake up any possible resync thread that waits for the device
324 * to go idle. All I/Os, even write-behind writes, are done.
326 allow_barrier(conf, r1_bio->sector);
332 * Update disk head position estimator based on IRQ completion info.
334 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
336 struct r1conf *conf = r1_bio->mddev->private;
338 conf->mirrors[disk].head_position =
339 r1_bio->sector + (r1_bio->sectors);
343 * Find the disk number which triggered given bio
345 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
348 struct r1conf *conf = r1_bio->mddev->private;
349 int raid_disks = conf->raid_disks;
351 for (mirror = 0; mirror < raid_disks * 2; mirror++)
352 if (r1_bio->bios[mirror] == bio)
355 BUG_ON(mirror == raid_disks * 2);
356 update_head_pos(mirror, r1_bio);
361 static void raid1_end_read_request(struct bio *bio)
363 int uptodate = !bio->bi_status;
364 struct r1bio *r1_bio = bio->bi_private;
365 struct r1conf *conf = r1_bio->mddev->private;
366 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
369 * this branch is our 'one mirror IO has finished' event handler:
371 update_head_pos(r1_bio->read_disk, r1_bio);
374 set_bit(R1BIO_Uptodate, &r1_bio->state);
375 else if (test_bit(FailFast, &rdev->flags) &&
376 test_bit(R1BIO_FailFast, &r1_bio->state))
377 /* This was a fail-fast read so we definitely
381 /* If all other devices have failed, we want to return
382 * the error upwards rather than fail the last device.
383 * Here we redefine "uptodate" to mean "Don't want to retry"
386 spin_lock_irqsave(&conf->device_lock, flags);
387 if (r1_bio->mddev->degraded == conf->raid_disks ||
388 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
389 test_bit(In_sync, &rdev->flags)))
391 spin_unlock_irqrestore(&conf->device_lock, flags);
395 raid_end_bio_io(r1_bio);
396 rdev_dec_pending(rdev, conf->mddev);
401 char b[BDEVNAME_SIZE];
402 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
404 bdevname(rdev->bdev, b),
405 (unsigned long long)r1_bio->sector);
406 set_bit(R1BIO_ReadError, &r1_bio->state);
407 reschedule_retry(r1_bio);
408 /* don't drop the reference on read_disk yet */
412 static void close_write(struct r1bio *r1_bio)
414 /* it really is the end of this request */
415 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
416 bio_free_pages(r1_bio->behind_master_bio);
417 bio_put(r1_bio->behind_master_bio);
418 r1_bio->behind_master_bio = NULL;
420 /* clear the bitmap if all writes complete successfully */
421 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
423 !test_bit(R1BIO_Degraded, &r1_bio->state),
424 test_bit(R1BIO_BehindIO, &r1_bio->state));
425 md_write_end(r1_bio->mddev);
428 static void r1_bio_write_done(struct r1bio *r1_bio)
430 if (!atomic_dec_and_test(&r1_bio->remaining))
433 if (test_bit(R1BIO_WriteError, &r1_bio->state))
434 reschedule_retry(r1_bio);
437 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
438 reschedule_retry(r1_bio);
440 raid_end_bio_io(r1_bio);
444 static void raid1_end_write_request(struct bio *bio)
446 struct r1bio *r1_bio = bio->bi_private;
447 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
448 struct r1conf *conf = r1_bio->mddev->private;
449 struct bio *to_put = NULL;
450 int mirror = find_bio_disk(r1_bio, bio);
451 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
453 sector_t lo = r1_bio->sector;
454 sector_t hi = r1_bio->sector + r1_bio->sectors;
456 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
459 * 'one mirror IO has finished' event handler:
461 if (bio->bi_status && !discard_error) {
462 set_bit(WriteErrorSeen, &rdev->flags);
463 if (!test_and_set_bit(WantReplacement, &rdev->flags))
464 set_bit(MD_RECOVERY_NEEDED, &
465 conf->mddev->recovery);
467 if (test_bit(FailFast, &rdev->flags) &&
468 (bio->bi_opf & MD_FAILFAST) &&
469 /* We never try FailFast to WriteMostly devices */
470 !test_bit(WriteMostly, &rdev->flags)) {
471 md_error(r1_bio->mddev, rdev);
475 * When the device is faulty, it is not necessary to
476 * handle write error.
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 r1bio_existed = !!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;
1223 if (r1bio_existed) {
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 */
1255 if (r1bio_existed) {
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 if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1296 r1_bio->start_time = bio_start_io_acct(bio);
1298 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1300 r1_bio->bios[rdisk] = read_bio;
1302 read_bio->bi_iter.bi_sector = r1_bio->sector +
1303 mirror->rdev->data_offset;
1304 bio_set_dev(read_bio, mirror->rdev->bdev);
1305 read_bio->bi_end_io = raid1_end_read_request;
1306 bio_set_op_attrs(read_bio, op, do_sync);
1307 if (test_bit(FailFast, &mirror->rdev->flags) &&
1308 test_bit(R1BIO_FailFast, &r1_bio->state))
1309 read_bio->bi_opf |= MD_FAILFAST;
1310 read_bio->bi_private = r1_bio;
1313 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1316 submit_bio_noacct(read_bio);
1319 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1320 int max_write_sectors)
1322 struct r1conf *conf = mddev->private;
1323 struct r1bio *r1_bio;
1325 struct bitmap *bitmap = mddev->bitmap;
1326 unsigned long flags;
1327 struct md_rdev *blocked_rdev;
1328 struct blk_plug_cb *cb;
1329 struct raid1_plug_cb *plug = NULL;
1333 if (mddev_is_clustered(mddev) &&
1334 md_cluster_ops->area_resyncing(mddev, WRITE,
1335 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1339 prepare_to_wait(&conf->wait_barrier,
1341 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1342 bio->bi_iter.bi_sector,
1343 bio_end_sector(bio)))
1347 finish_wait(&conf->wait_barrier, &w);
1351 * Register the new request and wait if the reconstruction
1352 * thread has put up a bar for new requests.
1353 * Continue immediately if no resync is active currently.
1355 wait_barrier(conf, bio->bi_iter.bi_sector);
1357 r1_bio = alloc_r1bio(mddev, bio);
1358 r1_bio->sectors = max_write_sectors;
1360 if (conf->pending_count >= max_queued_requests) {
1361 md_wakeup_thread(mddev->thread);
1362 raid1_log(mddev, "wait queued");
1363 wait_event(conf->wait_barrier,
1364 conf->pending_count < max_queued_requests);
1366 /* first select target devices under rcu_lock and
1367 * inc refcount on their rdev. Record them by setting
1369 * If there are known/acknowledged bad blocks on any device on
1370 * which we have seen a write error, we want to avoid writing those
1372 * This potentially requires several writes to write around
1373 * the bad blocks. Each set of writes gets it's own r1bio
1374 * with a set of bios attached.
1377 disks = conf->raid_disks * 2;
1379 blocked_rdev = NULL;
1381 max_sectors = r1_bio->sectors;
1382 for (i = 0; i < disks; i++) {
1383 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1384 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1385 atomic_inc(&rdev->nr_pending);
1386 blocked_rdev = rdev;
1389 r1_bio->bios[i] = NULL;
1390 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1391 if (i < conf->raid_disks)
1392 set_bit(R1BIO_Degraded, &r1_bio->state);
1396 atomic_inc(&rdev->nr_pending);
1397 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1402 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1403 &first_bad, &bad_sectors);
1405 /* mustn't write here until the bad block is
1407 set_bit(BlockedBadBlocks, &rdev->flags);
1408 blocked_rdev = rdev;
1411 if (is_bad && first_bad <= r1_bio->sector) {
1412 /* Cannot write here at all */
1413 bad_sectors -= (r1_bio->sector - first_bad);
1414 if (bad_sectors < max_sectors)
1415 /* mustn't write more than bad_sectors
1416 * to other devices yet
1418 max_sectors = bad_sectors;
1419 rdev_dec_pending(rdev, mddev);
1420 /* We don't set R1BIO_Degraded as that
1421 * only applies if the disk is
1422 * missing, so it might be re-added,
1423 * and we want to know to recover this
1425 * In this case the device is here,
1426 * and the fact that this chunk is not
1427 * in-sync is recorded in the bad
1433 int good_sectors = first_bad - r1_bio->sector;
1434 if (good_sectors < max_sectors)
1435 max_sectors = good_sectors;
1438 r1_bio->bios[i] = bio;
1442 if (unlikely(blocked_rdev)) {
1443 /* Wait for this device to become unblocked */
1446 for (j = 0; j < i; j++)
1447 if (r1_bio->bios[j])
1448 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1450 allow_barrier(conf, bio->bi_iter.bi_sector);
1451 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1452 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1453 wait_barrier(conf, bio->bi_iter.bi_sector);
1457 if (max_sectors < bio_sectors(bio)) {
1458 struct bio *split = bio_split(bio, max_sectors,
1459 GFP_NOIO, &conf->bio_split);
1460 bio_chain(split, bio);
1461 submit_bio_noacct(bio);
1463 r1_bio->master_bio = bio;
1464 r1_bio->sectors = max_sectors;
1467 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1468 r1_bio->start_time = bio_start_io_acct(bio);
1469 atomic_set(&r1_bio->remaining, 1);
1470 atomic_set(&r1_bio->behind_remaining, 0);
1474 for (i = 0; i < disks; i++) {
1475 struct bio *mbio = NULL;
1476 struct md_rdev *rdev = conf->mirrors[i].rdev;
1477 if (!r1_bio->bios[i])
1482 * Not if there are too many, or cannot
1483 * allocate memory, or a reader on WriteMostly
1484 * is waiting for behind writes to flush */
1486 (atomic_read(&bitmap->behind_writes)
1487 < mddev->bitmap_info.max_write_behind) &&
1488 !waitqueue_active(&bitmap->behind_wait)) {
1489 alloc_behind_master_bio(r1_bio, bio);
1492 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1493 test_bit(R1BIO_BehindIO, &r1_bio->state));
1497 if (r1_bio->behind_master_bio)
1498 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1499 GFP_NOIO, &mddev->bio_set);
1501 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1503 if (r1_bio->behind_master_bio) {
1504 if (test_bit(CollisionCheck, &rdev->flags))
1505 wait_for_serialization(rdev, r1_bio);
1506 if (test_bit(WriteMostly, &rdev->flags))
1507 atomic_inc(&r1_bio->behind_remaining);
1508 } else if (mddev->serialize_policy)
1509 wait_for_serialization(rdev, r1_bio);
1511 r1_bio->bios[i] = mbio;
1513 mbio->bi_iter.bi_sector = (r1_bio->sector +
1514 conf->mirrors[i].rdev->data_offset);
1515 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1516 mbio->bi_end_io = raid1_end_write_request;
1517 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1518 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1519 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1520 conf->raid_disks - mddev->degraded > 1)
1521 mbio->bi_opf |= MD_FAILFAST;
1522 mbio->bi_private = r1_bio;
1524 atomic_inc(&r1_bio->remaining);
1527 trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1529 /* flush_pending_writes() needs access to the rdev so...*/
1530 mbio->bi_bdev = (void *)conf->mirrors[i].rdev;
1532 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1534 plug = container_of(cb, struct raid1_plug_cb, cb);
1538 bio_list_add(&plug->pending, mbio);
1539 plug->pending_cnt++;
1541 spin_lock_irqsave(&conf->device_lock, flags);
1542 bio_list_add(&conf->pending_bio_list, mbio);
1543 conf->pending_count++;
1544 spin_unlock_irqrestore(&conf->device_lock, flags);
1545 md_wakeup_thread(mddev->thread);
1549 r1_bio_write_done(r1_bio);
1551 /* In case raid1d snuck in to freeze_array */
1552 wake_up(&conf->wait_barrier);
1555 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1559 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1560 && md_flush_request(mddev, bio))
1564 * There is a limit to the maximum size, but
1565 * the read/write handler might find a lower limit
1566 * due to bad blocks. To avoid multiple splits,
1567 * we pass the maximum number of sectors down
1568 * and let the lower level perform the split.
1570 sectors = align_to_barrier_unit_end(
1571 bio->bi_iter.bi_sector, bio_sectors(bio));
1573 if (bio_data_dir(bio) == READ)
1574 raid1_read_request(mddev, bio, sectors, NULL);
1576 if (!md_write_start(mddev,bio))
1578 raid1_write_request(mddev, bio, sectors);
1583 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1585 struct r1conf *conf = mddev->private;
1588 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1589 conf->raid_disks - mddev->degraded);
1591 for (i = 0; i < conf->raid_disks; i++) {
1592 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1593 seq_printf(seq, "%s",
1594 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1597 seq_printf(seq, "]");
1600 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1602 char b[BDEVNAME_SIZE];
1603 struct r1conf *conf = mddev->private;
1604 unsigned long flags;
1607 * If it is not operational, then we have already marked it as dead
1608 * else if it is the last working disks with "fail_last_dev == false",
1609 * ignore the error, let the next level up know.
1610 * else mark the drive as failed
1612 spin_lock_irqsave(&conf->device_lock, flags);
1613 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1614 && (conf->raid_disks - mddev->degraded) == 1) {
1616 * Don't fail the drive, act as though we were just a
1617 * normal single drive.
1618 * However don't try a recovery from this drive as
1619 * it is very likely to fail.
1621 conf->recovery_disabled = mddev->recovery_disabled;
1622 spin_unlock_irqrestore(&conf->device_lock, flags);
1625 set_bit(Blocked, &rdev->flags);
1626 if (test_and_clear_bit(In_sync, &rdev->flags))
1628 set_bit(Faulty, &rdev->flags);
1629 spin_unlock_irqrestore(&conf->device_lock, flags);
1631 * if recovery is running, make sure it aborts.
1633 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1634 set_mask_bits(&mddev->sb_flags, 0,
1635 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1636 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1637 "md/raid1:%s: Operation continuing on %d devices.\n",
1638 mdname(mddev), bdevname(rdev->bdev, b),
1639 mdname(mddev), conf->raid_disks - mddev->degraded);
1642 static void print_conf(struct r1conf *conf)
1646 pr_debug("RAID1 conf printout:\n");
1648 pr_debug("(!conf)\n");
1651 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1655 for (i = 0; i < conf->raid_disks; i++) {
1656 char b[BDEVNAME_SIZE];
1657 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1659 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1660 i, !test_bit(In_sync, &rdev->flags),
1661 !test_bit(Faulty, &rdev->flags),
1662 bdevname(rdev->bdev,b));
1667 static void close_sync(struct r1conf *conf)
1671 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1672 _wait_barrier(conf, idx);
1673 _allow_barrier(conf, idx);
1676 mempool_exit(&conf->r1buf_pool);
1679 static int raid1_spare_active(struct mddev *mddev)
1682 struct r1conf *conf = mddev->private;
1684 unsigned long flags;
1687 * Find all failed disks within the RAID1 configuration
1688 * and mark them readable.
1689 * Called under mddev lock, so rcu protection not needed.
1690 * device_lock used to avoid races with raid1_end_read_request
1691 * which expects 'In_sync' flags and ->degraded to be consistent.
1693 spin_lock_irqsave(&conf->device_lock, flags);
1694 for (i = 0; i < conf->raid_disks; i++) {
1695 struct md_rdev *rdev = conf->mirrors[i].rdev;
1696 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1698 && !test_bit(Candidate, &repl->flags)
1699 && repl->recovery_offset == MaxSector
1700 && !test_bit(Faulty, &repl->flags)
1701 && !test_and_set_bit(In_sync, &repl->flags)) {
1702 /* replacement has just become active */
1704 !test_and_clear_bit(In_sync, &rdev->flags))
1707 /* Replaced device not technically
1708 * faulty, but we need to be sure
1709 * it gets removed and never re-added
1711 set_bit(Faulty, &rdev->flags);
1712 sysfs_notify_dirent_safe(
1717 && rdev->recovery_offset == MaxSector
1718 && !test_bit(Faulty, &rdev->flags)
1719 && !test_and_set_bit(In_sync, &rdev->flags)) {
1721 sysfs_notify_dirent_safe(rdev->sysfs_state);
1724 mddev->degraded -= count;
1725 spin_unlock_irqrestore(&conf->device_lock, flags);
1731 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1733 struct r1conf *conf = mddev->private;
1736 struct raid1_info *p;
1738 int last = conf->raid_disks - 1;
1740 if (mddev->recovery_disabled == conf->recovery_disabled)
1743 if (md_integrity_add_rdev(rdev, mddev))
1746 if (rdev->raid_disk >= 0)
1747 first = last = rdev->raid_disk;
1750 * find the disk ... but prefer rdev->saved_raid_disk
1753 if (rdev->saved_raid_disk >= 0 &&
1754 rdev->saved_raid_disk >= first &&
1755 rdev->saved_raid_disk < conf->raid_disks &&
1756 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1757 first = last = rdev->saved_raid_disk;
1759 for (mirror = first; mirror <= last; mirror++) {
1760 p = conf->mirrors + mirror;
1763 disk_stack_limits(mddev->gendisk, rdev->bdev,
1764 rdev->data_offset << 9);
1766 p->head_position = 0;
1767 rdev->raid_disk = mirror;
1769 /* As all devices are equivalent, we don't need a full recovery
1770 * if this was recently any drive of the array
1772 if (rdev->saved_raid_disk < 0)
1774 rcu_assign_pointer(p->rdev, rdev);
1777 if (test_bit(WantReplacement, &p->rdev->flags) &&
1778 p[conf->raid_disks].rdev == NULL) {
1779 /* Add this device as a replacement */
1780 clear_bit(In_sync, &rdev->flags);
1781 set_bit(Replacement, &rdev->flags);
1782 rdev->raid_disk = mirror;
1785 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1789 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1790 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1795 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1797 struct r1conf *conf = mddev->private;
1799 int number = rdev->raid_disk;
1800 struct raid1_info *p = conf->mirrors + number;
1802 if (rdev != p->rdev)
1803 p = conf->mirrors + conf->raid_disks + number;
1806 if (rdev == p->rdev) {
1807 if (test_bit(In_sync, &rdev->flags) ||
1808 atomic_read(&rdev->nr_pending)) {
1812 /* Only remove non-faulty devices if recovery
1815 if (!test_bit(Faulty, &rdev->flags) &&
1816 mddev->recovery_disabled != conf->recovery_disabled &&
1817 mddev->degraded < conf->raid_disks) {
1822 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1824 if (atomic_read(&rdev->nr_pending)) {
1825 /* lost the race, try later */
1831 if (conf->mirrors[conf->raid_disks + number].rdev) {
1832 /* We just removed a device that is being replaced.
1833 * Move down the replacement. We drain all IO before
1834 * doing this to avoid confusion.
1836 struct md_rdev *repl =
1837 conf->mirrors[conf->raid_disks + number].rdev;
1838 freeze_array(conf, 0);
1839 if (atomic_read(&repl->nr_pending)) {
1840 /* It means that some queued IO of retry_list
1841 * hold repl. Thus, we cannot set replacement
1842 * as NULL, avoiding rdev NULL pointer
1843 * dereference in sync_request_write and
1844 * handle_write_finished.
1847 unfreeze_array(conf);
1850 clear_bit(Replacement, &repl->flags);
1852 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1853 unfreeze_array(conf);
1856 clear_bit(WantReplacement, &rdev->flags);
1857 err = md_integrity_register(mddev);
1865 static void end_sync_read(struct bio *bio)
1867 struct r1bio *r1_bio = get_resync_r1bio(bio);
1869 update_head_pos(r1_bio->read_disk, r1_bio);
1872 * we have read a block, now it needs to be re-written,
1873 * or re-read if the read failed.
1874 * We don't do much here, just schedule handling by raid1d
1876 if (!bio->bi_status)
1877 set_bit(R1BIO_Uptodate, &r1_bio->state);
1879 if (atomic_dec_and_test(&r1_bio->remaining))
1880 reschedule_retry(r1_bio);
1883 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1885 sector_t sync_blocks = 0;
1886 sector_t s = r1_bio->sector;
1887 long sectors_to_go = r1_bio->sectors;
1889 /* make sure these bits don't get cleared. */
1891 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1893 sectors_to_go -= sync_blocks;
1894 } while (sectors_to_go > 0);
1897 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1899 if (atomic_dec_and_test(&r1_bio->remaining)) {
1900 struct mddev *mddev = r1_bio->mddev;
1901 int s = r1_bio->sectors;
1903 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1904 test_bit(R1BIO_WriteError, &r1_bio->state))
1905 reschedule_retry(r1_bio);
1908 md_done_sync(mddev, s, uptodate);
1913 static void end_sync_write(struct bio *bio)
1915 int uptodate = !bio->bi_status;
1916 struct r1bio *r1_bio = get_resync_r1bio(bio);
1917 struct mddev *mddev = r1_bio->mddev;
1918 struct r1conf *conf = mddev->private;
1921 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1924 abort_sync_write(mddev, r1_bio);
1925 set_bit(WriteErrorSeen, &rdev->flags);
1926 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1927 set_bit(MD_RECOVERY_NEEDED, &
1929 set_bit(R1BIO_WriteError, &r1_bio->state);
1930 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1931 &first_bad, &bad_sectors) &&
1932 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1935 &first_bad, &bad_sectors)
1937 set_bit(R1BIO_MadeGood, &r1_bio->state);
1939 put_sync_write_buf(r1_bio, uptodate);
1942 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1943 int sectors, struct page *page, int rw)
1945 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1949 set_bit(WriteErrorSeen, &rdev->flags);
1950 if (!test_and_set_bit(WantReplacement,
1952 set_bit(MD_RECOVERY_NEEDED, &
1953 rdev->mddev->recovery);
1955 /* need to record an error - either for the block or the device */
1956 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1957 md_error(rdev->mddev, rdev);
1961 static int fix_sync_read_error(struct r1bio *r1_bio)
1963 /* Try some synchronous reads of other devices to get
1964 * good data, much like with normal read errors. Only
1965 * read into the pages we already have so we don't
1966 * need to re-issue the read request.
1967 * We don't need to freeze the array, because being in an
1968 * active sync request, there is no normal IO, and
1969 * no overlapping syncs.
1970 * We don't need to check is_badblock() again as we
1971 * made sure that anything with a bad block in range
1972 * will have bi_end_io clear.
1974 struct mddev *mddev = r1_bio->mddev;
1975 struct r1conf *conf = mddev->private;
1976 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1977 struct page **pages = get_resync_pages(bio)->pages;
1978 sector_t sect = r1_bio->sector;
1979 int sectors = r1_bio->sectors;
1981 struct md_rdev *rdev;
1983 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1984 if (test_bit(FailFast, &rdev->flags)) {
1985 /* Don't try recovering from here - just fail it
1986 * ... unless it is the last working device of course */
1987 md_error(mddev, rdev);
1988 if (test_bit(Faulty, &rdev->flags))
1989 /* Don't try to read from here, but make sure
1990 * put_buf does it's thing
1992 bio->bi_end_io = end_sync_write;
1997 int d = r1_bio->read_disk;
2001 if (s > (PAGE_SIZE>>9))
2004 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2005 /* No rcu protection needed here devices
2006 * can only be removed when no resync is
2007 * active, and resync is currently active
2009 rdev = conf->mirrors[d].rdev;
2010 if (sync_page_io(rdev, sect, s<<9,
2012 REQ_OP_READ, 0, false)) {
2018 if (d == conf->raid_disks * 2)
2020 } while (!success && d != r1_bio->read_disk);
2023 char b[BDEVNAME_SIZE];
2025 /* Cannot read from anywhere, this block is lost.
2026 * Record a bad block on each device. If that doesn't
2027 * work just disable and interrupt the recovery.
2028 * Don't fail devices as that won't really help.
2030 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2031 mdname(mddev), bio_devname(bio, b),
2032 (unsigned long long)r1_bio->sector);
2033 for (d = 0; d < conf->raid_disks * 2; d++) {
2034 rdev = conf->mirrors[d].rdev;
2035 if (!rdev || test_bit(Faulty, &rdev->flags))
2037 if (!rdev_set_badblocks(rdev, sect, s, 0))
2041 conf->recovery_disabled =
2042 mddev->recovery_disabled;
2043 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2044 md_done_sync(mddev, r1_bio->sectors, 0);
2056 /* write it back and re-read */
2057 while (d != r1_bio->read_disk) {
2059 d = conf->raid_disks * 2;
2061 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2063 rdev = conf->mirrors[d].rdev;
2064 if (r1_sync_page_io(rdev, sect, s,
2067 r1_bio->bios[d]->bi_end_io = NULL;
2068 rdev_dec_pending(rdev, mddev);
2072 while (d != r1_bio->read_disk) {
2074 d = conf->raid_disks * 2;
2076 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2078 rdev = conf->mirrors[d].rdev;
2079 if (r1_sync_page_io(rdev, sect, s,
2082 atomic_add(s, &rdev->corrected_errors);
2088 set_bit(R1BIO_Uptodate, &r1_bio->state);
2093 static void process_checks(struct r1bio *r1_bio)
2095 /* We have read all readable devices. If we haven't
2096 * got the block, then there is no hope left.
2097 * If we have, then we want to do a comparison
2098 * and skip the write if everything is the same.
2099 * If any blocks failed to read, then we need to
2100 * attempt an over-write
2102 struct mddev *mddev = r1_bio->mddev;
2103 struct r1conf *conf = mddev->private;
2108 /* Fix variable parts of all bios */
2109 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2110 for (i = 0; i < conf->raid_disks * 2; i++) {
2111 blk_status_t status;
2112 struct bio *b = r1_bio->bios[i];
2113 struct resync_pages *rp = get_resync_pages(b);
2114 if (b->bi_end_io != end_sync_read)
2116 /* fixup the bio for reuse, but preserve errno */
2117 status = b->bi_status;
2119 b->bi_status = status;
2120 b->bi_iter.bi_sector = r1_bio->sector +
2121 conf->mirrors[i].rdev->data_offset;
2122 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2123 b->bi_end_io = end_sync_read;
2124 rp->raid_bio = r1_bio;
2127 /* initialize bvec table again */
2128 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2130 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2131 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2132 !r1_bio->bios[primary]->bi_status) {
2133 r1_bio->bios[primary]->bi_end_io = NULL;
2134 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2137 r1_bio->read_disk = primary;
2138 for (i = 0; i < conf->raid_disks * 2; i++) {
2140 struct bio *pbio = r1_bio->bios[primary];
2141 struct bio *sbio = r1_bio->bios[i];
2142 blk_status_t status = sbio->bi_status;
2143 struct page **ppages = get_resync_pages(pbio)->pages;
2144 struct page **spages = get_resync_pages(sbio)->pages;
2146 int page_len[RESYNC_PAGES] = { 0 };
2147 struct bvec_iter_all iter_all;
2149 if (sbio->bi_end_io != end_sync_read)
2151 /* Now we can 'fixup' the error value */
2152 sbio->bi_status = 0;
2154 bio_for_each_segment_all(bi, sbio, iter_all)
2155 page_len[j++] = bi->bv_len;
2158 for (j = vcnt; j-- ; ) {
2159 if (memcmp(page_address(ppages[j]),
2160 page_address(spages[j]),
2167 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2168 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2170 /* No need to write to this device. */
2171 sbio->bi_end_io = NULL;
2172 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2176 bio_copy_data(sbio, pbio);
2180 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2182 struct r1conf *conf = mddev->private;
2184 int disks = conf->raid_disks * 2;
2187 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2188 /* ouch - failed to read all of that. */
2189 if (!fix_sync_read_error(r1_bio))
2192 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2193 process_checks(r1_bio);
2198 atomic_set(&r1_bio->remaining, 1);
2199 for (i = 0; i < disks ; i++) {
2200 wbio = r1_bio->bios[i];
2201 if (wbio->bi_end_io == NULL ||
2202 (wbio->bi_end_io == end_sync_read &&
2203 (i == r1_bio->read_disk ||
2204 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2206 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2207 abort_sync_write(mddev, r1_bio);
2211 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2212 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2213 wbio->bi_opf |= MD_FAILFAST;
2215 wbio->bi_end_io = end_sync_write;
2216 atomic_inc(&r1_bio->remaining);
2217 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2219 submit_bio_noacct(wbio);
2222 put_sync_write_buf(r1_bio, 1);
2226 * This is a kernel thread which:
2228 * 1. Retries failed read operations on working mirrors.
2229 * 2. Updates the raid superblock when problems encounter.
2230 * 3. Performs writes following reads for array synchronising.
2233 static void fix_read_error(struct r1conf *conf, int read_disk,
2234 sector_t sect, int sectors)
2236 struct mddev *mddev = conf->mddev;
2242 struct md_rdev *rdev;
2244 if (s > (PAGE_SIZE>>9))
2252 rdev = rcu_dereference(conf->mirrors[d].rdev);
2254 (test_bit(In_sync, &rdev->flags) ||
2255 (!test_bit(Faulty, &rdev->flags) &&
2256 rdev->recovery_offset >= sect + s)) &&
2257 is_badblock(rdev, sect, s,
2258 &first_bad, &bad_sectors) == 0) {
2259 atomic_inc(&rdev->nr_pending);
2261 if (sync_page_io(rdev, sect, s<<9,
2262 conf->tmppage, REQ_OP_READ, 0, false))
2264 rdev_dec_pending(rdev, mddev);
2270 if (d == conf->raid_disks * 2)
2272 } while (!success && d != read_disk);
2275 /* Cannot read from anywhere - mark it bad */
2276 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2277 if (!rdev_set_badblocks(rdev, sect, s, 0))
2278 md_error(mddev, rdev);
2281 /* write it back and re-read */
2283 while (d != read_disk) {
2285 d = conf->raid_disks * 2;
2288 rdev = rcu_dereference(conf->mirrors[d].rdev);
2290 !test_bit(Faulty, &rdev->flags)) {
2291 atomic_inc(&rdev->nr_pending);
2293 r1_sync_page_io(rdev, sect, s,
2294 conf->tmppage, WRITE);
2295 rdev_dec_pending(rdev, mddev);
2300 while (d != read_disk) {
2301 char b[BDEVNAME_SIZE];
2303 d = conf->raid_disks * 2;
2306 rdev = rcu_dereference(conf->mirrors[d].rdev);
2308 !test_bit(Faulty, &rdev->flags)) {
2309 atomic_inc(&rdev->nr_pending);
2311 if (r1_sync_page_io(rdev, sect, s,
2312 conf->tmppage, READ)) {
2313 atomic_add(s, &rdev->corrected_errors);
2314 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2316 (unsigned long long)(sect +
2318 bdevname(rdev->bdev, b));
2320 rdev_dec_pending(rdev, mddev);
2329 static int narrow_write_error(struct r1bio *r1_bio, int i)
2331 struct mddev *mddev = r1_bio->mddev;
2332 struct r1conf *conf = mddev->private;
2333 struct md_rdev *rdev = conf->mirrors[i].rdev;
2335 /* bio has the data to be written to device 'i' where
2336 * we just recently had a write error.
2337 * We repeatedly clone the bio and trim down to one block,
2338 * then try the write. Where the write fails we record
2340 * It is conceivable that the bio doesn't exactly align with
2341 * blocks. We must handle this somehow.
2343 * We currently own a reference on the rdev.
2349 int sect_to_write = r1_bio->sectors;
2352 if (rdev->badblocks.shift < 0)
2355 block_sectors = roundup(1 << rdev->badblocks.shift,
2356 bdev_logical_block_size(rdev->bdev) >> 9);
2357 sector = r1_bio->sector;
2358 sectors = ((sector + block_sectors)
2359 & ~(sector_t)(block_sectors - 1))
2362 while (sect_to_write) {
2364 if (sectors > sect_to_write)
2365 sectors = sect_to_write;
2366 /* Write at 'sector' for 'sectors'*/
2368 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2369 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2373 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2377 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2378 wbio->bi_iter.bi_sector = r1_bio->sector;
2379 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2381 bio_trim(wbio, sector - r1_bio->sector, sectors);
2382 wbio->bi_iter.bi_sector += rdev->data_offset;
2383 bio_set_dev(wbio, rdev->bdev);
2385 if (submit_bio_wait(wbio) < 0)
2387 ok = rdev_set_badblocks(rdev, sector,
2392 sect_to_write -= sectors;
2394 sectors = block_sectors;
2399 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2402 int s = r1_bio->sectors;
2403 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2404 struct md_rdev *rdev = conf->mirrors[m].rdev;
2405 struct bio *bio = r1_bio->bios[m];
2406 if (bio->bi_end_io == NULL)
2408 if (!bio->bi_status &&
2409 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2410 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2412 if (bio->bi_status &&
2413 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2414 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2415 md_error(conf->mddev, rdev);
2419 md_done_sync(conf->mddev, s, 1);
2422 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2427 for (m = 0; m < conf->raid_disks * 2 ; m++)
2428 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2429 struct md_rdev *rdev = conf->mirrors[m].rdev;
2430 rdev_clear_badblocks(rdev,
2432 r1_bio->sectors, 0);
2433 rdev_dec_pending(rdev, conf->mddev);
2434 } else if (r1_bio->bios[m] != NULL) {
2435 /* This drive got a write error. We need to
2436 * narrow down and record precise write
2440 if (!narrow_write_error(r1_bio, m)) {
2441 md_error(conf->mddev,
2442 conf->mirrors[m].rdev);
2443 /* an I/O failed, we can't clear the bitmap */
2444 set_bit(R1BIO_Degraded, &r1_bio->state);
2446 rdev_dec_pending(conf->mirrors[m].rdev,
2450 spin_lock_irq(&conf->device_lock);
2451 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2452 idx = sector_to_idx(r1_bio->sector);
2453 atomic_inc(&conf->nr_queued[idx]);
2454 spin_unlock_irq(&conf->device_lock);
2456 * In case freeze_array() is waiting for condition
2457 * get_unqueued_pending() == extra to be true.
2459 wake_up(&conf->wait_barrier);
2460 md_wakeup_thread(conf->mddev->thread);
2462 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2463 close_write(r1_bio);
2464 raid_end_bio_io(r1_bio);
2468 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2470 struct mddev *mddev = conf->mddev;
2472 struct md_rdev *rdev;
2474 clear_bit(R1BIO_ReadError, &r1_bio->state);
2475 /* we got a read error. Maybe the drive is bad. Maybe just
2476 * the block and we can fix it.
2477 * We freeze all other IO, and try reading the block from
2478 * other devices. When we find one, we re-write
2479 * and check it that fixes the read error.
2480 * This is all done synchronously while the array is
2484 bio = r1_bio->bios[r1_bio->read_disk];
2486 r1_bio->bios[r1_bio->read_disk] = NULL;
2488 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2490 && !test_bit(FailFast, &rdev->flags)) {
2491 freeze_array(conf, 1);
2492 fix_read_error(conf, r1_bio->read_disk,
2493 r1_bio->sector, r1_bio->sectors);
2494 unfreeze_array(conf);
2495 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2496 md_error(mddev, rdev);
2498 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2501 rdev_dec_pending(rdev, conf->mddev);
2502 allow_barrier(conf, r1_bio->sector);
2503 bio = r1_bio->master_bio;
2505 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2507 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2510 static void raid1d(struct md_thread *thread)
2512 struct mddev *mddev = thread->mddev;
2513 struct r1bio *r1_bio;
2514 unsigned long flags;
2515 struct r1conf *conf = mddev->private;
2516 struct list_head *head = &conf->retry_list;
2517 struct blk_plug plug;
2520 md_check_recovery(mddev);
2522 if (!list_empty_careful(&conf->bio_end_io_list) &&
2523 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2525 spin_lock_irqsave(&conf->device_lock, flags);
2526 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2527 list_splice_init(&conf->bio_end_io_list, &tmp);
2528 spin_unlock_irqrestore(&conf->device_lock, flags);
2529 while (!list_empty(&tmp)) {
2530 r1_bio = list_first_entry(&tmp, struct r1bio,
2532 list_del(&r1_bio->retry_list);
2533 idx = sector_to_idx(r1_bio->sector);
2534 atomic_dec(&conf->nr_queued[idx]);
2535 if (mddev->degraded)
2536 set_bit(R1BIO_Degraded, &r1_bio->state);
2537 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2538 close_write(r1_bio);
2539 raid_end_bio_io(r1_bio);
2543 blk_start_plug(&plug);
2546 flush_pending_writes(conf);
2548 spin_lock_irqsave(&conf->device_lock, flags);
2549 if (list_empty(head)) {
2550 spin_unlock_irqrestore(&conf->device_lock, flags);
2553 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2554 list_del(head->prev);
2555 idx = sector_to_idx(r1_bio->sector);
2556 atomic_dec(&conf->nr_queued[idx]);
2557 spin_unlock_irqrestore(&conf->device_lock, flags);
2559 mddev = r1_bio->mddev;
2560 conf = mddev->private;
2561 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2562 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2563 test_bit(R1BIO_WriteError, &r1_bio->state))
2564 handle_sync_write_finished(conf, r1_bio);
2566 sync_request_write(mddev, r1_bio);
2567 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2568 test_bit(R1BIO_WriteError, &r1_bio->state))
2569 handle_write_finished(conf, r1_bio);
2570 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2571 handle_read_error(conf, r1_bio);
2576 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2577 md_check_recovery(mddev);
2579 blk_finish_plug(&plug);
2582 static int init_resync(struct r1conf *conf)
2586 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2587 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2589 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2590 r1buf_pool_free, conf->poolinfo);
2593 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2595 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2596 struct resync_pages *rps;
2600 for (i = conf->poolinfo->raid_disks; i--; ) {
2601 bio = r1bio->bios[i];
2602 rps = bio->bi_private;
2604 bio->bi_private = rps;
2606 r1bio->master_bio = NULL;
2611 * perform a "sync" on one "block"
2613 * We need to make sure that no normal I/O request - particularly write
2614 * requests - conflict with active sync requests.
2616 * This is achieved by tracking pending requests and a 'barrier' concept
2617 * that can be installed to exclude normal IO requests.
2620 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2623 struct r1conf *conf = mddev->private;
2624 struct r1bio *r1_bio;
2626 sector_t max_sector, nr_sectors;
2630 int write_targets = 0, read_targets = 0;
2631 sector_t sync_blocks;
2632 int still_degraded = 0;
2633 int good_sectors = RESYNC_SECTORS;
2634 int min_bad = 0; /* number of sectors that are bad in all devices */
2635 int idx = sector_to_idx(sector_nr);
2638 if (!mempool_initialized(&conf->r1buf_pool))
2639 if (init_resync(conf))
2642 max_sector = mddev->dev_sectors;
2643 if (sector_nr >= max_sector) {
2644 /* If we aborted, we need to abort the
2645 * sync on the 'current' bitmap chunk (there will
2646 * only be one in raid1 resync.
2647 * We can find the current addess in mddev->curr_resync
2649 if (mddev->curr_resync < max_sector) /* aborted */
2650 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2652 else /* completed sync */
2655 md_bitmap_close_sync(mddev->bitmap);
2658 if (mddev_is_clustered(mddev)) {
2659 conf->cluster_sync_low = 0;
2660 conf->cluster_sync_high = 0;
2665 if (mddev->bitmap == NULL &&
2666 mddev->recovery_cp == MaxSector &&
2667 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2668 conf->fullsync == 0) {
2670 return max_sector - sector_nr;
2672 /* before building a request, check if we can skip these blocks..
2673 * This call the bitmap_start_sync doesn't actually record anything
2675 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2676 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2677 /* We can skip this block, and probably several more */
2683 * If there is non-resync activity waiting for a turn, then let it
2684 * though before starting on this new sync request.
2686 if (atomic_read(&conf->nr_waiting[idx]))
2687 schedule_timeout_uninterruptible(1);
2689 /* we are incrementing sector_nr below. To be safe, we check against
2690 * sector_nr + two times RESYNC_SECTORS
2693 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2694 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2697 if (raise_barrier(conf, sector_nr))
2700 r1_bio = raid1_alloc_init_r1buf(conf);
2704 * If we get a correctably read error during resync or recovery,
2705 * we might want to read from a different device. So we
2706 * flag all drives that could conceivably be read from for READ,
2707 * and any others (which will be non-In_sync devices) for WRITE.
2708 * If a read fails, we try reading from something else for which READ
2712 r1_bio->mddev = mddev;
2713 r1_bio->sector = sector_nr;
2715 set_bit(R1BIO_IsSync, &r1_bio->state);
2716 /* make sure good_sectors won't go across barrier unit boundary */
2717 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2719 for (i = 0; i < conf->raid_disks * 2; i++) {
2720 struct md_rdev *rdev;
2721 bio = r1_bio->bios[i];
2723 rdev = rcu_dereference(conf->mirrors[i].rdev);
2725 test_bit(Faulty, &rdev->flags)) {
2726 if (i < conf->raid_disks)
2728 } else if (!test_bit(In_sync, &rdev->flags)) {
2729 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2730 bio->bi_end_io = end_sync_write;
2733 /* may need to read from here */
2734 sector_t first_bad = MaxSector;
2737 if (is_badblock(rdev, sector_nr, good_sectors,
2738 &first_bad, &bad_sectors)) {
2739 if (first_bad > sector_nr)
2740 good_sectors = first_bad - sector_nr;
2742 bad_sectors -= (sector_nr - first_bad);
2744 min_bad > bad_sectors)
2745 min_bad = bad_sectors;
2748 if (sector_nr < first_bad) {
2749 if (test_bit(WriteMostly, &rdev->flags)) {
2756 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2757 bio->bi_end_io = end_sync_read;
2759 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2760 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2761 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2763 * The device is suitable for reading (InSync),
2764 * but has bad block(s) here. Let's try to correct them,
2765 * if we are doing resync or repair. Otherwise, leave
2766 * this device alone for this sync request.
2768 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2769 bio->bi_end_io = end_sync_write;
2773 if (rdev && bio->bi_end_io) {
2774 atomic_inc(&rdev->nr_pending);
2775 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2776 bio_set_dev(bio, rdev->bdev);
2777 if (test_bit(FailFast, &rdev->flags))
2778 bio->bi_opf |= MD_FAILFAST;
2784 r1_bio->read_disk = disk;
2786 if (read_targets == 0 && min_bad > 0) {
2787 /* These sectors are bad on all InSync devices, so we
2788 * need to mark them bad on all write targets
2791 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2792 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2793 struct md_rdev *rdev = conf->mirrors[i].rdev;
2794 ok = rdev_set_badblocks(rdev, sector_nr,
2798 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2803 /* Cannot record the badblocks, so need to
2805 * If there are multiple read targets, could just
2806 * fail the really bad ones ???
2808 conf->recovery_disabled = mddev->recovery_disabled;
2809 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2815 if (min_bad > 0 && min_bad < good_sectors) {
2816 /* only resync enough to reach the next bad->good
2818 good_sectors = min_bad;
2821 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2822 /* extra read targets are also write targets */
2823 write_targets += read_targets-1;
2825 if (write_targets == 0 || read_targets == 0) {
2826 /* There is nowhere to write, so all non-sync
2827 * drives must be failed - so we are finished
2831 max_sector = sector_nr + min_bad;
2832 rv = max_sector - sector_nr;
2838 if (max_sector > mddev->resync_max)
2839 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2840 if (max_sector > sector_nr + good_sectors)
2841 max_sector = sector_nr + good_sectors;
2846 int len = PAGE_SIZE;
2847 if (sector_nr + (len>>9) > max_sector)
2848 len = (max_sector - sector_nr) << 9;
2851 if (sync_blocks == 0) {
2852 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2853 &sync_blocks, still_degraded) &&
2855 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2857 if ((len >> 9) > sync_blocks)
2858 len = sync_blocks<<9;
2861 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2862 struct resync_pages *rp;
2864 bio = r1_bio->bios[i];
2865 rp = get_resync_pages(bio);
2866 if (bio->bi_end_io) {
2867 page = resync_fetch_page(rp, page_idx);
2870 * won't fail because the vec table is big
2871 * enough to hold all these pages
2873 bio_add_page(bio, page, len, 0);
2876 nr_sectors += len>>9;
2877 sector_nr += len>>9;
2878 sync_blocks -= (len>>9);
2879 } while (++page_idx < RESYNC_PAGES);
2881 r1_bio->sectors = nr_sectors;
2883 if (mddev_is_clustered(mddev) &&
2884 conf->cluster_sync_high < sector_nr + nr_sectors) {
2885 conf->cluster_sync_low = mddev->curr_resync_completed;
2886 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2887 /* Send resync message */
2888 md_cluster_ops->resync_info_update(mddev,
2889 conf->cluster_sync_low,
2890 conf->cluster_sync_high);
2893 /* For a user-requested sync, we read all readable devices and do a
2896 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2897 atomic_set(&r1_bio->remaining, read_targets);
2898 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2899 bio = r1_bio->bios[i];
2900 if (bio->bi_end_io == end_sync_read) {
2902 md_sync_acct_bio(bio, nr_sectors);
2903 if (read_targets == 1)
2904 bio->bi_opf &= ~MD_FAILFAST;
2905 submit_bio_noacct(bio);
2909 atomic_set(&r1_bio->remaining, 1);
2910 bio = r1_bio->bios[r1_bio->read_disk];
2911 md_sync_acct_bio(bio, nr_sectors);
2912 if (read_targets == 1)
2913 bio->bi_opf &= ~MD_FAILFAST;
2914 submit_bio_noacct(bio);
2919 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2924 return mddev->dev_sectors;
2927 static struct r1conf *setup_conf(struct mddev *mddev)
2929 struct r1conf *conf;
2931 struct raid1_info *disk;
2932 struct md_rdev *rdev;
2935 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2939 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2940 sizeof(atomic_t), GFP_KERNEL);
2941 if (!conf->nr_pending)
2944 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2945 sizeof(atomic_t), GFP_KERNEL);
2946 if (!conf->nr_waiting)
2949 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2950 sizeof(atomic_t), GFP_KERNEL);
2951 if (!conf->nr_queued)
2954 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2955 sizeof(atomic_t), GFP_KERNEL);
2959 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2960 mddev->raid_disks, 2),
2965 conf->tmppage = alloc_page(GFP_KERNEL);
2969 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2970 if (!conf->poolinfo)
2972 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2973 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2974 rbio_pool_free, conf->poolinfo);
2978 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2982 conf->poolinfo->mddev = mddev;
2985 spin_lock_init(&conf->device_lock);
2986 rdev_for_each(rdev, mddev) {
2987 int disk_idx = rdev->raid_disk;
2988 if (disk_idx >= mddev->raid_disks
2991 if (test_bit(Replacement, &rdev->flags))
2992 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2994 disk = conf->mirrors + disk_idx;
2999 disk->head_position = 0;
3000 disk->seq_start = MaxSector;
3002 conf->raid_disks = mddev->raid_disks;
3003 conf->mddev = mddev;
3004 INIT_LIST_HEAD(&conf->retry_list);
3005 INIT_LIST_HEAD(&conf->bio_end_io_list);
3007 spin_lock_init(&conf->resync_lock);
3008 init_waitqueue_head(&conf->wait_barrier);
3010 bio_list_init(&conf->pending_bio_list);
3011 conf->pending_count = 0;
3012 conf->recovery_disabled = mddev->recovery_disabled - 1;
3015 for (i = 0; i < conf->raid_disks * 2; i++) {
3017 disk = conf->mirrors + i;
3019 if (i < conf->raid_disks &&
3020 disk[conf->raid_disks].rdev) {
3021 /* This slot has a replacement. */
3023 /* No original, just make the replacement
3024 * a recovering spare
3027 disk[conf->raid_disks].rdev;
3028 disk[conf->raid_disks].rdev = NULL;
3029 } else if (!test_bit(In_sync, &disk->rdev->flags))
3030 /* Original is not in_sync - bad */
3035 !test_bit(In_sync, &disk->rdev->flags)) {
3036 disk->head_position = 0;
3038 (disk->rdev->saved_raid_disk < 0))
3044 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3052 mempool_exit(&conf->r1bio_pool);
3053 kfree(conf->mirrors);
3054 safe_put_page(conf->tmppage);
3055 kfree(conf->poolinfo);
3056 kfree(conf->nr_pending);
3057 kfree(conf->nr_waiting);
3058 kfree(conf->nr_queued);
3059 kfree(conf->barrier);
3060 bioset_exit(&conf->bio_split);
3063 return ERR_PTR(err);
3066 static void raid1_free(struct mddev *mddev, void *priv);
3067 static int raid1_run(struct mddev *mddev)
3069 struct r1conf *conf;
3071 struct md_rdev *rdev;
3073 bool discard_supported = false;
3075 if (mddev->level != 1) {
3076 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3077 mdname(mddev), mddev->level);
3080 if (mddev->reshape_position != MaxSector) {
3081 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3085 if (mddev_init_writes_pending(mddev) < 0)
3088 * copy the already verified devices into our private RAID1
3089 * bookkeeping area. [whatever we allocate in run(),
3090 * should be freed in raid1_free()]
3092 if (mddev->private == NULL)
3093 conf = setup_conf(mddev);
3095 conf = mddev->private;
3098 return PTR_ERR(conf);
3101 blk_queue_max_write_same_sectors(mddev->queue, 0);
3102 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3105 rdev_for_each(rdev, mddev) {
3106 if (!mddev->gendisk)
3108 disk_stack_limits(mddev->gendisk, rdev->bdev,
3109 rdev->data_offset << 9);
3110 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3111 discard_supported = true;
3114 mddev->degraded = 0;
3115 for (i = 0; i < conf->raid_disks; i++)
3116 if (conf->mirrors[i].rdev == NULL ||
3117 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3118 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3121 * RAID1 needs at least one disk in active
3123 if (conf->raid_disks - mddev->degraded < 1) {
3128 if (conf->raid_disks - mddev->degraded == 1)
3129 mddev->recovery_cp = MaxSector;
3131 if (mddev->recovery_cp != MaxSector)
3132 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3134 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3135 mdname(mddev), mddev->raid_disks - mddev->degraded,
3139 * Ok, everything is just fine now
3141 mddev->thread = conf->thread;
3142 conf->thread = NULL;
3143 mddev->private = conf;
3144 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3146 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3149 if (discard_supported)
3150 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3153 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3157 ret = md_integrity_register(mddev);
3159 md_unregister_thread(&mddev->thread);
3165 raid1_free(mddev, conf);
3169 static void raid1_free(struct mddev *mddev, void *priv)
3171 struct r1conf *conf = priv;
3173 mempool_exit(&conf->r1bio_pool);
3174 kfree(conf->mirrors);
3175 safe_put_page(conf->tmppage);
3176 kfree(conf->poolinfo);
3177 kfree(conf->nr_pending);
3178 kfree(conf->nr_waiting);
3179 kfree(conf->nr_queued);
3180 kfree(conf->barrier);
3181 bioset_exit(&conf->bio_split);
3185 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3187 /* no resync is happening, and there is enough space
3188 * on all devices, so we can resize.
3189 * We need to make sure resync covers any new space.
3190 * If the array is shrinking we should possibly wait until
3191 * any io in the removed space completes, but it hardly seems
3194 sector_t newsize = raid1_size(mddev, sectors, 0);
3195 if (mddev->external_size &&
3196 mddev->array_sectors > newsize)
3198 if (mddev->bitmap) {
3199 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3203 md_set_array_sectors(mddev, newsize);
3204 if (sectors > mddev->dev_sectors &&
3205 mddev->recovery_cp > mddev->dev_sectors) {
3206 mddev->recovery_cp = mddev->dev_sectors;
3207 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3209 mddev->dev_sectors = sectors;
3210 mddev->resync_max_sectors = sectors;
3214 static int raid1_reshape(struct mddev *mddev)
3217 * 1/ resize the r1bio_pool
3218 * 2/ resize conf->mirrors
3220 * We allocate a new r1bio_pool if we can.
3221 * Then raise a device barrier and wait until all IO stops.
3222 * Then resize conf->mirrors and swap in the new r1bio pool.
3224 * At the same time, we "pack" the devices so that all the missing
3225 * devices have the higher raid_disk numbers.
3227 mempool_t newpool, oldpool;
3228 struct pool_info *newpoolinfo;
3229 struct raid1_info *newmirrors;
3230 struct r1conf *conf = mddev->private;
3231 int cnt, raid_disks;
3232 unsigned long flags;
3236 memset(&newpool, 0, sizeof(newpool));
3237 memset(&oldpool, 0, sizeof(oldpool));
3239 /* Cannot change chunk_size, layout, or level */
3240 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3241 mddev->layout != mddev->new_layout ||
3242 mddev->level != mddev->new_level) {
3243 mddev->new_chunk_sectors = mddev->chunk_sectors;
3244 mddev->new_layout = mddev->layout;
3245 mddev->new_level = mddev->level;
3249 if (!mddev_is_clustered(mddev))
3250 md_allow_write(mddev);
3252 raid_disks = mddev->raid_disks + mddev->delta_disks;
3254 if (raid_disks < conf->raid_disks) {
3256 for (d= 0; d < conf->raid_disks; d++)
3257 if (conf->mirrors[d].rdev)
3259 if (cnt > raid_disks)
3263 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3266 newpoolinfo->mddev = mddev;
3267 newpoolinfo->raid_disks = raid_disks * 2;
3269 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3270 rbio_pool_free, newpoolinfo);
3275 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3280 mempool_exit(&newpool);
3284 freeze_array(conf, 0);
3286 /* ok, everything is stopped */
3287 oldpool = conf->r1bio_pool;
3288 conf->r1bio_pool = newpool;
3290 for (d = d2 = 0; d < conf->raid_disks; d++) {
3291 struct md_rdev *rdev = conf->mirrors[d].rdev;
3292 if (rdev && rdev->raid_disk != d2) {
3293 sysfs_unlink_rdev(mddev, rdev);
3294 rdev->raid_disk = d2;
3295 sysfs_unlink_rdev(mddev, rdev);
3296 if (sysfs_link_rdev(mddev, rdev))
3297 pr_warn("md/raid1:%s: cannot register rd%d\n",
3298 mdname(mddev), rdev->raid_disk);
3301 newmirrors[d2++].rdev = rdev;
3303 kfree(conf->mirrors);
3304 conf->mirrors = newmirrors;
3305 kfree(conf->poolinfo);
3306 conf->poolinfo = newpoolinfo;
3308 spin_lock_irqsave(&conf->device_lock, flags);
3309 mddev->degraded += (raid_disks - conf->raid_disks);
3310 spin_unlock_irqrestore(&conf->device_lock, flags);
3311 conf->raid_disks = mddev->raid_disks = raid_disks;
3312 mddev->delta_disks = 0;
3314 unfreeze_array(conf);
3316 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3317 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3318 md_wakeup_thread(mddev->thread);
3320 mempool_exit(&oldpool);
3324 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3326 struct r1conf *conf = mddev->private;
3329 freeze_array(conf, 0);
3331 unfreeze_array(conf);
3334 static void *raid1_takeover(struct mddev *mddev)
3336 /* raid1 can take over:
3337 * raid5 with 2 devices, any layout or chunk size
3339 if (mddev->level == 5 && mddev->raid_disks == 2) {
3340 struct r1conf *conf;
3341 mddev->new_level = 1;
3342 mddev->new_layout = 0;
3343 mddev->new_chunk_sectors = 0;
3344 conf = setup_conf(mddev);
3345 if (!IS_ERR(conf)) {
3346 /* Array must appear to be quiesced */
3347 conf->array_frozen = 1;
3348 mddev_clear_unsupported_flags(mddev,
3349 UNSUPPORTED_MDDEV_FLAGS);
3353 return ERR_PTR(-EINVAL);
3356 static struct md_personality raid1_personality =
3360 .owner = THIS_MODULE,
3361 .make_request = raid1_make_request,
3364 .status = raid1_status,
3365 .error_handler = raid1_error,
3366 .hot_add_disk = raid1_add_disk,
3367 .hot_remove_disk= raid1_remove_disk,
3368 .spare_active = raid1_spare_active,
3369 .sync_request = raid1_sync_request,
3370 .resize = raid1_resize,
3372 .check_reshape = raid1_reshape,
3373 .quiesce = raid1_quiesce,
3374 .takeover = raid1_takeover,
3377 static int __init raid_init(void)
3379 return register_md_personality(&raid1_personality);
3382 static void raid_exit(void)
3384 unregister_md_personality(&raid1_personality);
3387 module_init(raid_init);
3388 module_exit(raid_exit);
3389 MODULE_LICENSE("GPL");
3390 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3391 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3392 MODULE_ALIAS("md-raid1");
3393 MODULE_ALIAS("md-level-1");
3395 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);