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 spin_unlock_irq(&conf->device_lock);
830 * As this is called in a wait_event() loop (see freeze_array),
831 * current->state might be TASK_UNINTERRUPTIBLE which will
832 * cause a warning when we prepare to wait again. As it is
833 * rare that this path is taken, it is perfectly safe to force
834 * us to go around the wait_event() loop again, so the warning
835 * is a false-positive. Silence the warning by resetting
838 __set_current_state(TASK_RUNNING);
839 blk_start_plug(&plug);
840 flush_bio_list(conf, bio);
841 blk_finish_plug(&plug);
843 spin_unlock_irq(&conf->device_lock);
847 * Sometimes we need to suspend IO while we do something else,
848 * either some resync/recovery, or reconfigure the array.
849 * To do this we raise a 'barrier'.
850 * The 'barrier' is a counter that can be raised multiple times
851 * to count how many activities are happening which preclude
853 * We can only raise the barrier if there is no pending IO.
854 * i.e. if nr_pending == 0.
855 * We choose only to raise the barrier if no-one is waiting for the
856 * barrier to go down. This means that as soon as an IO request
857 * is ready, no other operations which require a barrier will start
858 * until the IO request has had a chance.
860 * So: regular IO calls 'wait_barrier'. When that returns there
861 * is no backgroup IO happening, It must arrange to call
862 * allow_barrier when it has finished its IO.
863 * backgroup IO calls must call raise_barrier. Once that returns
864 * there is no normal IO happeing. It must arrange to call
865 * lower_barrier when the particular background IO completes.
867 * If resync/recovery is interrupted, returns -EINTR;
868 * Otherwise, returns 0.
870 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
872 int idx = sector_to_idx(sector_nr);
874 spin_lock_irq(&conf->resync_lock);
876 /* Wait until no block IO is waiting */
877 wait_event_lock_irq(conf->wait_barrier,
878 !atomic_read(&conf->nr_waiting[idx]),
881 /* block any new IO from starting */
882 atomic_inc(&conf->barrier[idx]);
884 * In raise_barrier() we firstly increase conf->barrier[idx] then
885 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
886 * increase conf->nr_pending[idx] then check conf->barrier[idx].
887 * A memory barrier here to make sure conf->nr_pending[idx] won't
888 * be fetched before conf->barrier[idx] is increased. Otherwise
889 * there will be a race between raise_barrier() and _wait_barrier().
891 smp_mb__after_atomic();
893 /* For these conditions we must wait:
894 * A: while the array is in frozen state
895 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
896 * existing in corresponding I/O barrier bucket.
897 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
898 * max resync count which allowed on current I/O barrier bucket.
900 wait_event_lock_irq(conf->wait_barrier,
901 (!conf->array_frozen &&
902 !atomic_read(&conf->nr_pending[idx]) &&
903 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
904 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
907 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
908 atomic_dec(&conf->barrier[idx]);
909 spin_unlock_irq(&conf->resync_lock);
910 wake_up(&conf->wait_barrier);
914 atomic_inc(&conf->nr_sync_pending);
915 spin_unlock_irq(&conf->resync_lock);
920 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
922 int idx = sector_to_idx(sector_nr);
924 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
926 atomic_dec(&conf->barrier[idx]);
927 atomic_dec(&conf->nr_sync_pending);
928 wake_up(&conf->wait_barrier);
931 static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
936 * We need to increase conf->nr_pending[idx] very early here,
937 * then raise_barrier() can be blocked when it waits for
938 * conf->nr_pending[idx] to be 0. Then we can avoid holding
939 * conf->resync_lock when there is no barrier raised in same
940 * barrier unit bucket. Also if the array is frozen, I/O
941 * should be blocked until array is unfrozen.
943 atomic_inc(&conf->nr_pending[idx]);
945 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
946 * check conf->barrier[idx]. In raise_barrier() we firstly increase
947 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
948 * barrier is necessary here to make sure conf->barrier[idx] won't be
949 * fetched before conf->nr_pending[idx] is increased. Otherwise there
950 * will be a race between _wait_barrier() and raise_barrier().
952 smp_mb__after_atomic();
955 * Don't worry about checking two atomic_t variables at same time
956 * here. If during we check conf->barrier[idx], the array is
957 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
958 * 0, it is safe to return and make the I/O continue. Because the
959 * array is frozen, all I/O returned here will eventually complete
960 * or be queued, no race will happen. See code comment in
963 if (!READ_ONCE(conf->array_frozen) &&
964 !atomic_read(&conf->barrier[idx]))
968 * After holding conf->resync_lock, conf->nr_pending[idx]
969 * should be decreased before waiting for barrier to drop.
970 * Otherwise, we may encounter a race condition because
971 * raise_barrer() might be waiting for conf->nr_pending[idx]
972 * to be 0 at same time.
974 spin_lock_irq(&conf->resync_lock);
975 atomic_inc(&conf->nr_waiting[idx]);
976 atomic_dec(&conf->nr_pending[idx]);
978 * In case freeze_array() is waiting for
979 * get_unqueued_pending() == extra
981 wake_up(&conf->wait_barrier);
982 /* Wait for the barrier in same barrier unit bucket to drop. */
984 /* Return false when nowait flag is set */
988 wait_event_lock_irq(conf->wait_barrier,
989 !conf->array_frozen &&
990 !atomic_read(&conf->barrier[idx]),
992 atomic_inc(&conf->nr_pending[idx]);
995 atomic_dec(&conf->nr_waiting[idx]);
996 spin_unlock_irq(&conf->resync_lock);
1000 static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1002 int idx = sector_to_idx(sector_nr);
1006 * Very similar to _wait_barrier(). The difference is, for read
1007 * I/O we don't need wait for sync I/O, but if the whole array
1008 * is frozen, the read I/O still has to wait until the array is
1009 * unfrozen. Since there is no ordering requirement with
1010 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1012 atomic_inc(&conf->nr_pending[idx]);
1014 if (!READ_ONCE(conf->array_frozen))
1017 spin_lock_irq(&conf->resync_lock);
1018 atomic_inc(&conf->nr_waiting[idx]);
1019 atomic_dec(&conf->nr_pending[idx]);
1021 * In case freeze_array() is waiting for
1022 * get_unqueued_pending() == extra
1024 wake_up(&conf->wait_barrier);
1025 /* Wait for array to be unfrozen */
1027 /* Return false when nowait flag is set */
1029 /* Return false when nowait flag is set */
1032 wait_event_lock_irq(conf->wait_barrier,
1033 !conf->array_frozen,
1035 atomic_inc(&conf->nr_pending[idx]);
1038 atomic_dec(&conf->nr_waiting[idx]);
1039 spin_unlock_irq(&conf->resync_lock);
1043 static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1045 int idx = sector_to_idx(sector_nr);
1047 return _wait_barrier(conf, idx, nowait);
1050 static void _allow_barrier(struct r1conf *conf, int idx)
1052 atomic_dec(&conf->nr_pending[idx]);
1053 wake_up(&conf->wait_barrier);
1056 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1058 int idx = sector_to_idx(sector_nr);
1060 _allow_barrier(conf, idx);
1063 /* conf->resync_lock should be held */
1064 static int get_unqueued_pending(struct r1conf *conf)
1068 ret = atomic_read(&conf->nr_sync_pending);
1069 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1070 ret += atomic_read(&conf->nr_pending[idx]) -
1071 atomic_read(&conf->nr_queued[idx]);
1076 static void freeze_array(struct r1conf *conf, int extra)
1078 /* Stop sync I/O and normal I/O and wait for everything to
1080 * This is called in two situations:
1081 * 1) management command handlers (reshape, remove disk, quiesce).
1082 * 2) one normal I/O request failed.
1084 * After array_frozen is set to 1, new sync IO will be blocked at
1085 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1086 * or wait_read_barrier(). The flying I/Os will either complete or be
1087 * queued. When everything goes quite, there are only queued I/Os left.
1089 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1090 * barrier bucket index which this I/O request hits. When all sync and
1091 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1092 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1093 * in handle_read_error(), we may call freeze_array() before trying to
1094 * fix the read error. In this case, the error read I/O is not queued,
1095 * so get_unqueued_pending() == 1.
1097 * Therefore before this function returns, we need to wait until
1098 * get_unqueued_pendings(conf) gets equal to extra. For
1099 * normal I/O context, extra is 1, in rested situations extra is 0.
1101 spin_lock_irq(&conf->resync_lock);
1102 conf->array_frozen = 1;
1103 raid1_log(conf->mddev, "wait freeze");
1104 wait_event_lock_irq_cmd(
1106 get_unqueued_pending(conf) == extra,
1108 flush_pending_writes(conf));
1109 spin_unlock_irq(&conf->resync_lock);
1111 static void unfreeze_array(struct r1conf *conf)
1113 /* reverse the effect of the freeze */
1114 spin_lock_irq(&conf->resync_lock);
1115 conf->array_frozen = 0;
1116 spin_unlock_irq(&conf->resync_lock);
1117 wake_up(&conf->wait_barrier);
1120 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1123 int size = bio->bi_iter.bi_size;
1124 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1126 struct bio *behind_bio = NULL;
1128 behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1129 &r1_bio->mddev->bio_set);
1133 /* discard op, we don't support writezero/writesame yet */
1134 if (!bio_has_data(bio)) {
1135 behind_bio->bi_iter.bi_size = size;
1139 behind_bio->bi_write_hint = bio->bi_write_hint;
1141 while (i < vcnt && size) {
1143 int len = min_t(int, PAGE_SIZE, size);
1145 page = alloc_page(GFP_NOIO);
1146 if (unlikely(!page))
1149 bio_add_page(behind_bio, page, len, 0);
1155 bio_copy_data(behind_bio, bio);
1157 r1_bio->behind_master_bio = behind_bio;
1158 set_bit(R1BIO_BehindIO, &r1_bio->state);
1163 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1164 bio->bi_iter.bi_size);
1165 bio_free_pages(behind_bio);
1166 bio_put(behind_bio);
1169 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1171 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1173 struct mddev *mddev = plug->cb.data;
1174 struct r1conf *conf = mddev->private;
1177 if (from_schedule || current->bio_list) {
1178 spin_lock_irq(&conf->device_lock);
1179 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1180 spin_unlock_irq(&conf->device_lock);
1181 wake_up(&conf->wait_barrier);
1182 md_wakeup_thread(mddev->thread);
1187 /* we aren't scheduling, so we can do the write-out directly. */
1188 bio = bio_list_get(&plug->pending);
1189 flush_bio_list(conf, bio);
1193 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1195 r1_bio->master_bio = bio;
1196 r1_bio->sectors = bio_sectors(bio);
1198 r1_bio->mddev = mddev;
1199 r1_bio->sector = bio->bi_iter.bi_sector;
1202 static inline struct r1bio *
1203 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1205 struct r1conf *conf = mddev->private;
1206 struct r1bio *r1_bio;
1208 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1209 /* Ensure no bio records IO_BLOCKED */
1210 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1211 init_r1bio(r1_bio, mddev, bio);
1215 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1216 int max_read_sectors, struct r1bio *r1_bio)
1218 struct r1conf *conf = mddev->private;
1219 struct raid1_info *mirror;
1220 struct bio *read_bio;
1221 struct bitmap *bitmap = mddev->bitmap;
1222 const int op = bio_op(bio);
1223 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1226 bool r1bio_existed = !!r1_bio;
1227 char b[BDEVNAME_SIZE];
1230 * If r1_bio is set, we are blocking the raid1d thread
1231 * so there is a tiny risk of deadlock. So ask for
1232 * emergency memory if needed.
1234 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1236 if (r1bio_existed) {
1237 /* Need to get the block device name carefully */
1238 struct md_rdev *rdev;
1240 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1242 bdevname(rdev->bdev, b);
1249 * Still need barrier for READ in case that whole
1252 if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1253 bio->bi_opf & REQ_NOWAIT)) {
1254 bio_wouldblock_error(bio);
1259 r1_bio = alloc_r1bio(mddev, bio);
1261 init_r1bio(r1_bio, mddev, bio);
1262 r1_bio->sectors = max_read_sectors;
1265 * make_request() can abort the operation when read-ahead is being
1266 * used and no empty request is available.
1268 rdisk = read_balance(conf, r1_bio, &max_sectors);
1271 /* couldn't find anywhere to read from */
1272 if (r1bio_existed) {
1273 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1276 (unsigned long long)r1_bio->sector);
1278 raid_end_bio_io(r1_bio);
1281 mirror = conf->mirrors + rdisk;
1284 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1286 (unsigned long long)r1_bio->sector,
1287 bdevname(mirror->rdev->bdev, b));
1289 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1292 * Reading from a write-mostly device must take care not to
1293 * over-take any writes that are 'behind'
1295 raid1_log(mddev, "wait behind writes");
1296 wait_event(bitmap->behind_wait,
1297 atomic_read(&bitmap->behind_writes) == 0);
1300 if (max_sectors < bio_sectors(bio)) {
1301 struct bio *split = bio_split(bio, max_sectors,
1302 gfp, &conf->bio_split);
1303 bio_chain(split, bio);
1304 submit_bio_noacct(bio);
1306 r1_bio->master_bio = bio;
1307 r1_bio->sectors = max_sectors;
1310 r1_bio->read_disk = rdisk;
1312 if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1313 r1_bio->start_time = bio_start_io_acct(bio);
1315 read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1318 r1_bio->bios[rdisk] = read_bio;
1320 read_bio->bi_iter.bi_sector = r1_bio->sector +
1321 mirror->rdev->data_offset;
1322 read_bio->bi_end_io = raid1_end_read_request;
1323 bio_set_op_attrs(read_bio, op, do_sync);
1324 if (test_bit(FailFast, &mirror->rdev->flags) &&
1325 test_bit(R1BIO_FailFast, &r1_bio->state))
1326 read_bio->bi_opf |= MD_FAILFAST;
1327 read_bio->bi_private = r1_bio;
1330 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1333 submit_bio_noacct(read_bio);
1336 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1337 int max_write_sectors)
1339 struct r1conf *conf = mddev->private;
1340 struct r1bio *r1_bio;
1342 struct bitmap *bitmap = mddev->bitmap;
1343 unsigned long flags;
1344 struct md_rdev *blocked_rdev;
1345 struct blk_plug_cb *cb;
1346 struct raid1_plug_cb *plug = NULL;
1349 bool write_behind = false;
1351 if (mddev_is_clustered(mddev) &&
1352 md_cluster_ops->area_resyncing(mddev, WRITE,
1353 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1356 if (bio->bi_opf & REQ_NOWAIT) {
1357 bio_wouldblock_error(bio);
1361 prepare_to_wait(&conf->wait_barrier,
1363 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1364 bio->bi_iter.bi_sector,
1365 bio_end_sector(bio)))
1369 finish_wait(&conf->wait_barrier, &w);
1373 * Register the new request and wait if the reconstruction
1374 * thread has put up a bar for new requests.
1375 * Continue immediately if no resync is active currently.
1377 if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1378 bio->bi_opf & REQ_NOWAIT)) {
1379 bio_wouldblock_error(bio);
1383 r1_bio = alloc_r1bio(mddev, bio);
1384 r1_bio->sectors = max_write_sectors;
1386 /* first select target devices under rcu_lock and
1387 * inc refcount on their rdev. Record them by setting
1389 * If there are known/acknowledged bad blocks on any device on
1390 * which we have seen a write error, we want to avoid writing those
1392 * This potentially requires several writes to write around
1393 * the bad blocks. Each set of writes gets it's own r1bio
1394 * with a set of bios attached.
1397 disks = conf->raid_disks * 2;
1399 blocked_rdev = NULL;
1401 max_sectors = r1_bio->sectors;
1402 for (i = 0; i < disks; i++) {
1403 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1406 * The write-behind io is only attempted on drives marked as
1407 * write-mostly, which means we could allocate write behind
1410 if (rdev && test_bit(WriteMostly, &rdev->flags))
1411 write_behind = true;
1413 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1414 atomic_inc(&rdev->nr_pending);
1415 blocked_rdev = rdev;
1418 r1_bio->bios[i] = NULL;
1419 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1420 if (i < conf->raid_disks)
1421 set_bit(R1BIO_Degraded, &r1_bio->state);
1425 atomic_inc(&rdev->nr_pending);
1426 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1431 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1432 &first_bad, &bad_sectors);
1434 /* mustn't write here until the bad block is
1436 set_bit(BlockedBadBlocks, &rdev->flags);
1437 blocked_rdev = rdev;
1440 if (is_bad && first_bad <= r1_bio->sector) {
1441 /* Cannot write here at all */
1442 bad_sectors -= (r1_bio->sector - first_bad);
1443 if (bad_sectors < max_sectors)
1444 /* mustn't write more than bad_sectors
1445 * to other devices yet
1447 max_sectors = bad_sectors;
1448 rdev_dec_pending(rdev, mddev);
1449 /* We don't set R1BIO_Degraded as that
1450 * only applies if the disk is
1451 * missing, so it might be re-added,
1452 * and we want to know to recover this
1454 * In this case the device is here,
1455 * and the fact that this chunk is not
1456 * in-sync is recorded in the bad
1462 int good_sectors = first_bad - r1_bio->sector;
1463 if (good_sectors < max_sectors)
1464 max_sectors = good_sectors;
1467 r1_bio->bios[i] = bio;
1471 if (unlikely(blocked_rdev)) {
1472 /* Wait for this device to become unblocked */
1475 for (j = 0; j < i; j++)
1476 if (r1_bio->bios[j])
1477 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1479 allow_barrier(conf, bio->bi_iter.bi_sector);
1481 if (bio->bi_opf & REQ_NOWAIT) {
1482 bio_wouldblock_error(bio);
1485 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1486 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1487 wait_barrier(conf, bio->bi_iter.bi_sector, false);
1492 * When using a bitmap, we may call alloc_behind_master_bio below.
1493 * alloc_behind_master_bio allocates a copy of the data payload a page
1494 * at a time and thus needs a new bio that can fit the whole payload
1495 * this bio in page sized chunks.
1497 if (write_behind && bitmap)
1498 max_sectors = min_t(int, max_sectors,
1499 BIO_MAX_VECS * (PAGE_SIZE >> 9));
1500 if (max_sectors < bio_sectors(bio)) {
1501 struct bio *split = bio_split(bio, max_sectors,
1502 GFP_NOIO, &conf->bio_split);
1503 bio_chain(split, bio);
1504 submit_bio_noacct(bio);
1506 r1_bio->master_bio = bio;
1507 r1_bio->sectors = max_sectors;
1510 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1511 r1_bio->start_time = bio_start_io_acct(bio);
1512 atomic_set(&r1_bio->remaining, 1);
1513 atomic_set(&r1_bio->behind_remaining, 0);
1517 for (i = 0; i < disks; i++) {
1518 struct bio *mbio = NULL;
1519 struct md_rdev *rdev = conf->mirrors[i].rdev;
1520 if (!r1_bio->bios[i])
1525 * Not if there are too many, or cannot
1526 * allocate memory, or a reader on WriteMostly
1527 * is waiting for behind writes to flush */
1529 test_bit(WriteMostly, &rdev->flags) &&
1530 (atomic_read(&bitmap->behind_writes)
1531 < mddev->bitmap_info.max_write_behind) &&
1532 !waitqueue_active(&bitmap->behind_wait)) {
1533 alloc_behind_master_bio(r1_bio, bio);
1536 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1537 test_bit(R1BIO_BehindIO, &r1_bio->state));
1541 if (r1_bio->behind_master_bio) {
1542 mbio = bio_alloc_clone(rdev->bdev,
1543 r1_bio->behind_master_bio,
1544 GFP_NOIO, &mddev->bio_set);
1545 if (test_bit(CollisionCheck, &rdev->flags))
1546 wait_for_serialization(rdev, r1_bio);
1547 if (test_bit(WriteMostly, &rdev->flags))
1548 atomic_inc(&r1_bio->behind_remaining);
1550 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1553 if (mddev->serialize_policy)
1554 wait_for_serialization(rdev, r1_bio);
1557 r1_bio->bios[i] = mbio;
1559 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1560 mbio->bi_end_io = raid1_end_write_request;
1561 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1562 if (test_bit(FailFast, &rdev->flags) &&
1563 !test_bit(WriteMostly, &rdev->flags) &&
1564 conf->raid_disks - mddev->degraded > 1)
1565 mbio->bi_opf |= MD_FAILFAST;
1566 mbio->bi_private = r1_bio;
1568 atomic_inc(&r1_bio->remaining);
1571 trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1573 /* flush_pending_writes() needs access to the rdev so...*/
1574 mbio->bi_bdev = (void *)rdev;
1576 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1578 plug = container_of(cb, struct raid1_plug_cb, cb);
1582 bio_list_add(&plug->pending, mbio);
1584 spin_lock_irqsave(&conf->device_lock, flags);
1585 bio_list_add(&conf->pending_bio_list, mbio);
1586 spin_unlock_irqrestore(&conf->device_lock, flags);
1587 md_wakeup_thread(mddev->thread);
1591 r1_bio_write_done(r1_bio);
1593 /* In case raid1d snuck in to freeze_array */
1594 wake_up(&conf->wait_barrier);
1597 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1601 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1602 && md_flush_request(mddev, bio))
1606 * There is a limit to the maximum size, but
1607 * the read/write handler might find a lower limit
1608 * due to bad blocks. To avoid multiple splits,
1609 * we pass the maximum number of sectors down
1610 * and let the lower level perform the split.
1612 sectors = align_to_barrier_unit_end(
1613 bio->bi_iter.bi_sector, bio_sectors(bio));
1615 if (bio_data_dir(bio) == READ)
1616 raid1_read_request(mddev, bio, sectors, NULL);
1618 if (!md_write_start(mddev,bio))
1620 raid1_write_request(mddev, bio, sectors);
1625 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1627 struct r1conf *conf = mddev->private;
1630 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1631 conf->raid_disks - mddev->degraded);
1633 for (i = 0; i < conf->raid_disks; i++) {
1634 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1635 seq_printf(seq, "%s",
1636 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1639 seq_printf(seq, "]");
1642 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1644 char b[BDEVNAME_SIZE];
1645 struct r1conf *conf = mddev->private;
1646 unsigned long flags;
1649 * If it is not operational, then we have already marked it as dead
1650 * else if it is the last working disks with "fail_last_dev == false",
1651 * ignore the error, let the next level up know.
1652 * else mark the drive as failed
1654 spin_lock_irqsave(&conf->device_lock, flags);
1655 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1656 && (conf->raid_disks - mddev->degraded) == 1) {
1658 * Don't fail the drive, act as though we were just a
1659 * normal single drive.
1660 * However don't try a recovery from this drive as
1661 * it is very likely to fail.
1663 conf->recovery_disabled = mddev->recovery_disabled;
1664 spin_unlock_irqrestore(&conf->device_lock, flags);
1667 set_bit(Blocked, &rdev->flags);
1668 if (test_and_clear_bit(In_sync, &rdev->flags))
1670 set_bit(Faulty, &rdev->flags);
1671 spin_unlock_irqrestore(&conf->device_lock, flags);
1673 * if recovery is running, make sure it aborts.
1675 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1676 set_mask_bits(&mddev->sb_flags, 0,
1677 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1678 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1679 "md/raid1:%s: Operation continuing on %d devices.\n",
1680 mdname(mddev), bdevname(rdev->bdev, b),
1681 mdname(mddev), conf->raid_disks - mddev->degraded);
1684 static void print_conf(struct r1conf *conf)
1688 pr_debug("RAID1 conf printout:\n");
1690 pr_debug("(!conf)\n");
1693 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1697 for (i = 0; i < conf->raid_disks; i++) {
1698 char b[BDEVNAME_SIZE];
1699 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1701 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1702 i, !test_bit(In_sync, &rdev->flags),
1703 !test_bit(Faulty, &rdev->flags),
1704 bdevname(rdev->bdev,b));
1709 static void close_sync(struct r1conf *conf)
1713 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1714 _wait_barrier(conf, idx, false);
1715 _allow_barrier(conf, idx);
1718 mempool_exit(&conf->r1buf_pool);
1721 static int raid1_spare_active(struct mddev *mddev)
1724 struct r1conf *conf = mddev->private;
1726 unsigned long flags;
1729 * Find all failed disks within the RAID1 configuration
1730 * and mark them readable.
1731 * Called under mddev lock, so rcu protection not needed.
1732 * device_lock used to avoid races with raid1_end_read_request
1733 * which expects 'In_sync' flags and ->degraded to be consistent.
1735 spin_lock_irqsave(&conf->device_lock, flags);
1736 for (i = 0; i < conf->raid_disks; i++) {
1737 struct md_rdev *rdev = conf->mirrors[i].rdev;
1738 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1740 && !test_bit(Candidate, &repl->flags)
1741 && repl->recovery_offset == MaxSector
1742 && !test_bit(Faulty, &repl->flags)
1743 && !test_and_set_bit(In_sync, &repl->flags)) {
1744 /* replacement has just become active */
1746 !test_and_clear_bit(In_sync, &rdev->flags))
1749 /* Replaced device not technically
1750 * faulty, but we need to be sure
1751 * it gets removed and never re-added
1753 set_bit(Faulty, &rdev->flags);
1754 sysfs_notify_dirent_safe(
1759 && rdev->recovery_offset == MaxSector
1760 && !test_bit(Faulty, &rdev->flags)
1761 && !test_and_set_bit(In_sync, &rdev->flags)) {
1763 sysfs_notify_dirent_safe(rdev->sysfs_state);
1766 mddev->degraded -= count;
1767 spin_unlock_irqrestore(&conf->device_lock, flags);
1773 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1775 struct r1conf *conf = mddev->private;
1778 struct raid1_info *p;
1780 int last = conf->raid_disks - 1;
1782 if (mddev->recovery_disabled == conf->recovery_disabled)
1785 if (md_integrity_add_rdev(rdev, mddev))
1788 if (rdev->raid_disk >= 0)
1789 first = last = rdev->raid_disk;
1792 * find the disk ... but prefer rdev->saved_raid_disk
1795 if (rdev->saved_raid_disk >= 0 &&
1796 rdev->saved_raid_disk >= first &&
1797 rdev->saved_raid_disk < conf->raid_disks &&
1798 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1799 first = last = rdev->saved_raid_disk;
1801 for (mirror = first; mirror <= last; mirror++) {
1802 p = conf->mirrors + mirror;
1805 disk_stack_limits(mddev->gendisk, rdev->bdev,
1806 rdev->data_offset << 9);
1808 p->head_position = 0;
1809 rdev->raid_disk = mirror;
1811 /* As all devices are equivalent, we don't need a full recovery
1812 * if this was recently any drive of the array
1814 if (rdev->saved_raid_disk < 0)
1816 rcu_assign_pointer(p->rdev, rdev);
1819 if (test_bit(WantReplacement, &p->rdev->flags) &&
1820 p[conf->raid_disks].rdev == NULL) {
1821 /* Add this device as a replacement */
1822 clear_bit(In_sync, &rdev->flags);
1823 set_bit(Replacement, &rdev->flags);
1824 rdev->raid_disk = mirror;
1827 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1831 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1832 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1837 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1839 struct r1conf *conf = mddev->private;
1841 int number = rdev->raid_disk;
1842 struct raid1_info *p = conf->mirrors + number;
1844 if (rdev != p->rdev)
1845 p = conf->mirrors + conf->raid_disks + number;
1848 if (rdev == p->rdev) {
1849 if (test_bit(In_sync, &rdev->flags) ||
1850 atomic_read(&rdev->nr_pending)) {
1854 /* Only remove non-faulty devices if recovery
1857 if (!test_bit(Faulty, &rdev->flags) &&
1858 mddev->recovery_disabled != conf->recovery_disabled &&
1859 mddev->degraded < conf->raid_disks) {
1864 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1866 if (atomic_read(&rdev->nr_pending)) {
1867 /* lost the race, try later */
1873 if (conf->mirrors[conf->raid_disks + number].rdev) {
1874 /* We just removed a device that is being replaced.
1875 * Move down the replacement. We drain all IO before
1876 * doing this to avoid confusion.
1878 struct md_rdev *repl =
1879 conf->mirrors[conf->raid_disks + number].rdev;
1880 freeze_array(conf, 0);
1881 if (atomic_read(&repl->nr_pending)) {
1882 /* It means that some queued IO of retry_list
1883 * hold repl. Thus, we cannot set replacement
1884 * as NULL, avoiding rdev NULL pointer
1885 * dereference in sync_request_write and
1886 * handle_write_finished.
1889 unfreeze_array(conf);
1892 clear_bit(Replacement, &repl->flags);
1894 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1895 unfreeze_array(conf);
1898 clear_bit(WantReplacement, &rdev->flags);
1899 err = md_integrity_register(mddev);
1907 static void end_sync_read(struct bio *bio)
1909 struct r1bio *r1_bio = get_resync_r1bio(bio);
1911 update_head_pos(r1_bio->read_disk, r1_bio);
1914 * we have read a block, now it needs to be re-written,
1915 * or re-read if the read failed.
1916 * We don't do much here, just schedule handling by raid1d
1918 if (!bio->bi_status)
1919 set_bit(R1BIO_Uptodate, &r1_bio->state);
1921 if (atomic_dec_and_test(&r1_bio->remaining))
1922 reschedule_retry(r1_bio);
1925 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1927 sector_t sync_blocks = 0;
1928 sector_t s = r1_bio->sector;
1929 long sectors_to_go = r1_bio->sectors;
1931 /* make sure these bits don't get cleared. */
1933 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1935 sectors_to_go -= sync_blocks;
1936 } while (sectors_to_go > 0);
1939 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1941 if (atomic_dec_and_test(&r1_bio->remaining)) {
1942 struct mddev *mddev = r1_bio->mddev;
1943 int s = r1_bio->sectors;
1945 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1946 test_bit(R1BIO_WriteError, &r1_bio->state))
1947 reschedule_retry(r1_bio);
1950 md_done_sync(mddev, s, uptodate);
1955 static void end_sync_write(struct bio *bio)
1957 int uptodate = !bio->bi_status;
1958 struct r1bio *r1_bio = get_resync_r1bio(bio);
1959 struct mddev *mddev = r1_bio->mddev;
1960 struct r1conf *conf = mddev->private;
1963 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1966 abort_sync_write(mddev, r1_bio);
1967 set_bit(WriteErrorSeen, &rdev->flags);
1968 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1969 set_bit(MD_RECOVERY_NEEDED, &
1971 set_bit(R1BIO_WriteError, &r1_bio->state);
1972 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1973 &first_bad, &bad_sectors) &&
1974 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1977 &first_bad, &bad_sectors)
1979 set_bit(R1BIO_MadeGood, &r1_bio->state);
1981 put_sync_write_buf(r1_bio, uptodate);
1984 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1985 int sectors, struct page *page, int rw)
1987 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1991 set_bit(WriteErrorSeen, &rdev->flags);
1992 if (!test_and_set_bit(WantReplacement,
1994 set_bit(MD_RECOVERY_NEEDED, &
1995 rdev->mddev->recovery);
1997 /* need to record an error - either for the block or the device */
1998 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1999 md_error(rdev->mddev, rdev);
2003 static int fix_sync_read_error(struct r1bio *r1_bio)
2005 /* Try some synchronous reads of other devices to get
2006 * good data, much like with normal read errors. Only
2007 * read into the pages we already have so we don't
2008 * need to re-issue the read request.
2009 * We don't need to freeze the array, because being in an
2010 * active sync request, there is no normal IO, and
2011 * no overlapping syncs.
2012 * We don't need to check is_badblock() again as we
2013 * made sure that anything with a bad block in range
2014 * will have bi_end_io clear.
2016 struct mddev *mddev = r1_bio->mddev;
2017 struct r1conf *conf = mddev->private;
2018 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2019 struct page **pages = get_resync_pages(bio)->pages;
2020 sector_t sect = r1_bio->sector;
2021 int sectors = r1_bio->sectors;
2023 struct md_rdev *rdev;
2025 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2026 if (test_bit(FailFast, &rdev->flags)) {
2027 /* Don't try recovering from here - just fail it
2028 * ... unless it is the last working device of course */
2029 md_error(mddev, rdev);
2030 if (test_bit(Faulty, &rdev->flags))
2031 /* Don't try to read from here, but make sure
2032 * put_buf does it's thing
2034 bio->bi_end_io = end_sync_write;
2039 int d = r1_bio->read_disk;
2043 if (s > (PAGE_SIZE>>9))
2046 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2047 /* No rcu protection needed here devices
2048 * can only be removed when no resync is
2049 * active, and resync is currently active
2051 rdev = conf->mirrors[d].rdev;
2052 if (sync_page_io(rdev, sect, s<<9,
2054 REQ_OP_READ, 0, false)) {
2060 if (d == conf->raid_disks * 2)
2062 } while (!success && d != r1_bio->read_disk);
2066 /* Cannot read from anywhere, this block is lost.
2067 * Record a bad block on each device. If that doesn't
2068 * work just disable and interrupt the recovery.
2069 * Don't fail devices as that won't really help.
2071 pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2072 mdname(mddev), bio->bi_bdev,
2073 (unsigned long long)r1_bio->sector);
2074 for (d = 0; d < conf->raid_disks * 2; d++) {
2075 rdev = conf->mirrors[d].rdev;
2076 if (!rdev || test_bit(Faulty, &rdev->flags))
2078 if (!rdev_set_badblocks(rdev, sect, s, 0))
2082 conf->recovery_disabled =
2083 mddev->recovery_disabled;
2084 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2085 md_done_sync(mddev, r1_bio->sectors, 0);
2097 /* write it back and re-read */
2098 while (d != r1_bio->read_disk) {
2100 d = conf->raid_disks * 2;
2102 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2104 rdev = conf->mirrors[d].rdev;
2105 if (r1_sync_page_io(rdev, sect, s,
2108 r1_bio->bios[d]->bi_end_io = NULL;
2109 rdev_dec_pending(rdev, mddev);
2113 while (d != r1_bio->read_disk) {
2115 d = conf->raid_disks * 2;
2117 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2119 rdev = conf->mirrors[d].rdev;
2120 if (r1_sync_page_io(rdev, sect, s,
2123 atomic_add(s, &rdev->corrected_errors);
2129 set_bit(R1BIO_Uptodate, &r1_bio->state);
2134 static void process_checks(struct r1bio *r1_bio)
2136 /* We have read all readable devices. If we haven't
2137 * got the block, then there is no hope left.
2138 * If we have, then we want to do a comparison
2139 * and skip the write if everything is the same.
2140 * If any blocks failed to read, then we need to
2141 * attempt an over-write
2143 struct mddev *mddev = r1_bio->mddev;
2144 struct r1conf *conf = mddev->private;
2149 /* Fix variable parts of all bios */
2150 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2151 for (i = 0; i < conf->raid_disks * 2; i++) {
2152 blk_status_t status;
2153 struct bio *b = r1_bio->bios[i];
2154 struct resync_pages *rp = get_resync_pages(b);
2155 if (b->bi_end_io != end_sync_read)
2157 /* fixup the bio for reuse, but preserve errno */
2158 status = b->bi_status;
2159 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2160 b->bi_status = status;
2161 b->bi_iter.bi_sector = r1_bio->sector +
2162 conf->mirrors[i].rdev->data_offset;
2163 b->bi_end_io = end_sync_read;
2164 rp->raid_bio = r1_bio;
2167 /* initialize bvec table again */
2168 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2170 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2171 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2172 !r1_bio->bios[primary]->bi_status) {
2173 r1_bio->bios[primary]->bi_end_io = NULL;
2174 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2177 r1_bio->read_disk = primary;
2178 for (i = 0; i < conf->raid_disks * 2; i++) {
2180 struct bio *pbio = r1_bio->bios[primary];
2181 struct bio *sbio = r1_bio->bios[i];
2182 blk_status_t status = sbio->bi_status;
2183 struct page **ppages = get_resync_pages(pbio)->pages;
2184 struct page **spages = get_resync_pages(sbio)->pages;
2186 int page_len[RESYNC_PAGES] = { 0 };
2187 struct bvec_iter_all iter_all;
2189 if (sbio->bi_end_io != end_sync_read)
2191 /* Now we can 'fixup' the error value */
2192 sbio->bi_status = 0;
2194 bio_for_each_segment_all(bi, sbio, iter_all)
2195 page_len[j++] = bi->bv_len;
2198 for (j = vcnt; j-- ; ) {
2199 if (memcmp(page_address(ppages[j]),
2200 page_address(spages[j]),
2207 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2208 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2210 /* No need to write to this device. */
2211 sbio->bi_end_io = NULL;
2212 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2216 bio_copy_data(sbio, pbio);
2220 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2222 struct r1conf *conf = mddev->private;
2224 int disks = conf->raid_disks * 2;
2227 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2228 /* ouch - failed to read all of that. */
2229 if (!fix_sync_read_error(r1_bio))
2232 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2233 process_checks(r1_bio);
2238 atomic_set(&r1_bio->remaining, 1);
2239 for (i = 0; i < disks ; i++) {
2240 wbio = r1_bio->bios[i];
2241 if (wbio->bi_end_io == NULL ||
2242 (wbio->bi_end_io == end_sync_read &&
2243 (i == r1_bio->read_disk ||
2244 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2246 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2247 abort_sync_write(mddev, r1_bio);
2251 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2252 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2253 wbio->bi_opf |= MD_FAILFAST;
2255 wbio->bi_end_io = end_sync_write;
2256 atomic_inc(&r1_bio->remaining);
2257 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2259 submit_bio_noacct(wbio);
2262 put_sync_write_buf(r1_bio, 1);
2266 * This is a kernel thread which:
2268 * 1. Retries failed read operations on working mirrors.
2269 * 2. Updates the raid superblock when problems encounter.
2270 * 3. Performs writes following reads for array synchronising.
2273 static void fix_read_error(struct r1conf *conf, int read_disk,
2274 sector_t sect, int sectors)
2276 struct mddev *mddev = conf->mddev;
2282 struct md_rdev *rdev;
2284 if (s > (PAGE_SIZE>>9))
2292 rdev = rcu_dereference(conf->mirrors[d].rdev);
2294 (test_bit(In_sync, &rdev->flags) ||
2295 (!test_bit(Faulty, &rdev->flags) &&
2296 rdev->recovery_offset >= sect + s)) &&
2297 is_badblock(rdev, sect, s,
2298 &first_bad, &bad_sectors) == 0) {
2299 atomic_inc(&rdev->nr_pending);
2301 if (sync_page_io(rdev, sect, s<<9,
2302 conf->tmppage, REQ_OP_READ, 0, false))
2304 rdev_dec_pending(rdev, mddev);
2310 if (d == conf->raid_disks * 2)
2312 } while (!success && d != read_disk);
2315 /* Cannot read from anywhere - mark it bad */
2316 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2317 if (!rdev_set_badblocks(rdev, sect, s, 0))
2318 md_error(mddev, rdev);
2321 /* write it back and re-read */
2323 while (d != read_disk) {
2325 d = conf->raid_disks * 2;
2328 rdev = rcu_dereference(conf->mirrors[d].rdev);
2330 !test_bit(Faulty, &rdev->flags)) {
2331 atomic_inc(&rdev->nr_pending);
2333 r1_sync_page_io(rdev, sect, s,
2334 conf->tmppage, WRITE);
2335 rdev_dec_pending(rdev, mddev);
2340 while (d != read_disk) {
2341 char b[BDEVNAME_SIZE];
2343 d = conf->raid_disks * 2;
2346 rdev = rcu_dereference(conf->mirrors[d].rdev);
2348 !test_bit(Faulty, &rdev->flags)) {
2349 atomic_inc(&rdev->nr_pending);
2351 if (r1_sync_page_io(rdev, sect, s,
2352 conf->tmppage, READ)) {
2353 atomic_add(s, &rdev->corrected_errors);
2354 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2356 (unsigned long long)(sect +
2358 bdevname(rdev->bdev, b));
2360 rdev_dec_pending(rdev, mddev);
2369 static int narrow_write_error(struct r1bio *r1_bio, int i)
2371 struct mddev *mddev = r1_bio->mddev;
2372 struct r1conf *conf = mddev->private;
2373 struct md_rdev *rdev = conf->mirrors[i].rdev;
2375 /* bio has the data to be written to device 'i' where
2376 * we just recently had a write error.
2377 * We repeatedly clone the bio and trim down to one block,
2378 * then try the write. Where the write fails we record
2380 * It is conceivable that the bio doesn't exactly align with
2381 * blocks. We must handle this somehow.
2383 * We currently own a reference on the rdev.
2389 int sect_to_write = r1_bio->sectors;
2392 if (rdev->badblocks.shift < 0)
2395 block_sectors = roundup(1 << rdev->badblocks.shift,
2396 bdev_logical_block_size(rdev->bdev) >> 9);
2397 sector = r1_bio->sector;
2398 sectors = ((sector + block_sectors)
2399 & ~(sector_t)(block_sectors - 1))
2402 while (sect_to_write) {
2404 if (sectors > sect_to_write)
2405 sectors = sect_to_write;
2406 /* Write at 'sector' for 'sectors'*/
2408 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2409 wbio = bio_alloc_clone(rdev->bdev,
2410 r1_bio->behind_master_bio,
2411 GFP_NOIO, &mddev->bio_set);
2413 wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2414 GFP_NOIO, &mddev->bio_set);
2417 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2418 wbio->bi_iter.bi_sector = r1_bio->sector;
2419 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2421 bio_trim(wbio, sector - r1_bio->sector, sectors);
2422 wbio->bi_iter.bi_sector += rdev->data_offset;
2424 if (submit_bio_wait(wbio) < 0)
2426 ok = rdev_set_badblocks(rdev, sector,
2431 sect_to_write -= sectors;
2433 sectors = block_sectors;
2438 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2441 int s = r1_bio->sectors;
2442 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2443 struct md_rdev *rdev = conf->mirrors[m].rdev;
2444 struct bio *bio = r1_bio->bios[m];
2445 if (bio->bi_end_io == NULL)
2447 if (!bio->bi_status &&
2448 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2449 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2451 if (bio->bi_status &&
2452 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2453 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2454 md_error(conf->mddev, rdev);
2458 md_done_sync(conf->mddev, s, 1);
2461 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2466 for (m = 0; m < conf->raid_disks * 2 ; m++)
2467 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2468 struct md_rdev *rdev = conf->mirrors[m].rdev;
2469 rdev_clear_badblocks(rdev,
2471 r1_bio->sectors, 0);
2472 rdev_dec_pending(rdev, conf->mddev);
2473 } else if (r1_bio->bios[m] != NULL) {
2474 /* This drive got a write error. We need to
2475 * narrow down and record precise write
2479 if (!narrow_write_error(r1_bio, m)) {
2480 md_error(conf->mddev,
2481 conf->mirrors[m].rdev);
2482 /* an I/O failed, we can't clear the bitmap */
2483 set_bit(R1BIO_Degraded, &r1_bio->state);
2485 rdev_dec_pending(conf->mirrors[m].rdev,
2489 spin_lock_irq(&conf->device_lock);
2490 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2491 idx = sector_to_idx(r1_bio->sector);
2492 atomic_inc(&conf->nr_queued[idx]);
2493 spin_unlock_irq(&conf->device_lock);
2495 * In case freeze_array() is waiting for condition
2496 * get_unqueued_pending() == extra to be true.
2498 wake_up(&conf->wait_barrier);
2499 md_wakeup_thread(conf->mddev->thread);
2501 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2502 close_write(r1_bio);
2503 raid_end_bio_io(r1_bio);
2507 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2509 struct mddev *mddev = conf->mddev;
2511 struct md_rdev *rdev;
2513 clear_bit(R1BIO_ReadError, &r1_bio->state);
2514 /* we got a read error. Maybe the drive is bad. Maybe just
2515 * the block and we can fix it.
2516 * We freeze all other IO, and try reading the block from
2517 * other devices. When we find one, we re-write
2518 * and check it that fixes the read error.
2519 * This is all done synchronously while the array is
2523 bio = r1_bio->bios[r1_bio->read_disk];
2525 r1_bio->bios[r1_bio->read_disk] = NULL;
2527 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2529 && !test_bit(FailFast, &rdev->flags)) {
2530 freeze_array(conf, 1);
2531 fix_read_error(conf, r1_bio->read_disk,
2532 r1_bio->sector, r1_bio->sectors);
2533 unfreeze_array(conf);
2534 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2535 md_error(mddev, rdev);
2537 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2540 rdev_dec_pending(rdev, conf->mddev);
2541 allow_barrier(conf, r1_bio->sector);
2542 bio = r1_bio->master_bio;
2544 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2546 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2549 static void raid1d(struct md_thread *thread)
2551 struct mddev *mddev = thread->mddev;
2552 struct r1bio *r1_bio;
2553 unsigned long flags;
2554 struct r1conf *conf = mddev->private;
2555 struct list_head *head = &conf->retry_list;
2556 struct blk_plug plug;
2559 md_check_recovery(mddev);
2561 if (!list_empty_careful(&conf->bio_end_io_list) &&
2562 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2564 spin_lock_irqsave(&conf->device_lock, flags);
2565 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2566 list_splice_init(&conf->bio_end_io_list, &tmp);
2567 spin_unlock_irqrestore(&conf->device_lock, flags);
2568 while (!list_empty(&tmp)) {
2569 r1_bio = list_first_entry(&tmp, struct r1bio,
2571 list_del(&r1_bio->retry_list);
2572 idx = sector_to_idx(r1_bio->sector);
2573 atomic_dec(&conf->nr_queued[idx]);
2574 if (mddev->degraded)
2575 set_bit(R1BIO_Degraded, &r1_bio->state);
2576 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2577 close_write(r1_bio);
2578 raid_end_bio_io(r1_bio);
2582 blk_start_plug(&plug);
2585 flush_pending_writes(conf);
2587 spin_lock_irqsave(&conf->device_lock, flags);
2588 if (list_empty(head)) {
2589 spin_unlock_irqrestore(&conf->device_lock, flags);
2592 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2593 list_del(head->prev);
2594 idx = sector_to_idx(r1_bio->sector);
2595 atomic_dec(&conf->nr_queued[idx]);
2596 spin_unlock_irqrestore(&conf->device_lock, flags);
2598 mddev = r1_bio->mddev;
2599 conf = mddev->private;
2600 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2601 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2602 test_bit(R1BIO_WriteError, &r1_bio->state))
2603 handle_sync_write_finished(conf, r1_bio);
2605 sync_request_write(mddev, r1_bio);
2606 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2607 test_bit(R1BIO_WriteError, &r1_bio->state))
2608 handle_write_finished(conf, r1_bio);
2609 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2610 handle_read_error(conf, r1_bio);
2615 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2616 md_check_recovery(mddev);
2618 blk_finish_plug(&plug);
2621 static int init_resync(struct r1conf *conf)
2625 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2626 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2628 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2629 r1buf_pool_free, conf->poolinfo);
2632 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2634 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2635 struct resync_pages *rps;
2639 for (i = conf->poolinfo->raid_disks; i--; ) {
2640 bio = r1bio->bios[i];
2641 rps = bio->bi_private;
2642 bio_reset(bio, NULL, 0);
2643 bio->bi_private = rps;
2645 r1bio->master_bio = NULL;
2650 * perform a "sync" on one "block"
2652 * We need to make sure that no normal I/O request - particularly write
2653 * requests - conflict with active sync requests.
2655 * This is achieved by tracking pending requests and a 'barrier' concept
2656 * that can be installed to exclude normal IO requests.
2659 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2662 struct r1conf *conf = mddev->private;
2663 struct r1bio *r1_bio;
2665 sector_t max_sector, nr_sectors;
2669 int write_targets = 0, read_targets = 0;
2670 sector_t sync_blocks;
2671 int still_degraded = 0;
2672 int good_sectors = RESYNC_SECTORS;
2673 int min_bad = 0; /* number of sectors that are bad in all devices */
2674 int idx = sector_to_idx(sector_nr);
2677 if (!mempool_initialized(&conf->r1buf_pool))
2678 if (init_resync(conf))
2681 max_sector = mddev->dev_sectors;
2682 if (sector_nr >= max_sector) {
2683 /* If we aborted, we need to abort the
2684 * sync on the 'current' bitmap chunk (there will
2685 * only be one in raid1 resync.
2686 * We can find the current addess in mddev->curr_resync
2688 if (mddev->curr_resync < max_sector) /* aborted */
2689 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2691 else /* completed sync */
2694 md_bitmap_close_sync(mddev->bitmap);
2697 if (mddev_is_clustered(mddev)) {
2698 conf->cluster_sync_low = 0;
2699 conf->cluster_sync_high = 0;
2704 if (mddev->bitmap == NULL &&
2705 mddev->recovery_cp == MaxSector &&
2706 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2707 conf->fullsync == 0) {
2709 return max_sector - sector_nr;
2711 /* before building a request, check if we can skip these blocks..
2712 * This call the bitmap_start_sync doesn't actually record anything
2714 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2715 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2716 /* We can skip this block, and probably several more */
2722 * If there is non-resync activity waiting for a turn, then let it
2723 * though before starting on this new sync request.
2725 if (atomic_read(&conf->nr_waiting[idx]))
2726 schedule_timeout_uninterruptible(1);
2728 /* we are incrementing sector_nr below. To be safe, we check against
2729 * sector_nr + two times RESYNC_SECTORS
2732 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2733 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2736 if (raise_barrier(conf, sector_nr))
2739 r1_bio = raid1_alloc_init_r1buf(conf);
2743 * If we get a correctably read error during resync or recovery,
2744 * we might want to read from a different device. So we
2745 * flag all drives that could conceivably be read from for READ,
2746 * and any others (which will be non-In_sync devices) for WRITE.
2747 * If a read fails, we try reading from something else for which READ
2751 r1_bio->mddev = mddev;
2752 r1_bio->sector = sector_nr;
2754 set_bit(R1BIO_IsSync, &r1_bio->state);
2755 /* make sure good_sectors won't go across barrier unit boundary */
2756 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2758 for (i = 0; i < conf->raid_disks * 2; i++) {
2759 struct md_rdev *rdev;
2760 bio = r1_bio->bios[i];
2762 rdev = rcu_dereference(conf->mirrors[i].rdev);
2764 test_bit(Faulty, &rdev->flags)) {
2765 if (i < conf->raid_disks)
2767 } else if (!test_bit(In_sync, &rdev->flags)) {
2768 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2769 bio->bi_end_io = end_sync_write;
2772 /* may need to read from here */
2773 sector_t first_bad = MaxSector;
2776 if (is_badblock(rdev, sector_nr, good_sectors,
2777 &first_bad, &bad_sectors)) {
2778 if (first_bad > sector_nr)
2779 good_sectors = first_bad - sector_nr;
2781 bad_sectors -= (sector_nr - first_bad);
2783 min_bad > bad_sectors)
2784 min_bad = bad_sectors;
2787 if (sector_nr < first_bad) {
2788 if (test_bit(WriteMostly, &rdev->flags)) {
2795 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2796 bio->bi_end_io = end_sync_read;
2798 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2799 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2800 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2802 * The device is suitable for reading (InSync),
2803 * but has bad block(s) here. Let's try to correct them,
2804 * if we are doing resync or repair. Otherwise, leave
2805 * this device alone for this sync request.
2807 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2808 bio->bi_end_io = end_sync_write;
2812 if (rdev && bio->bi_end_io) {
2813 atomic_inc(&rdev->nr_pending);
2814 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2815 bio_set_dev(bio, rdev->bdev);
2816 if (test_bit(FailFast, &rdev->flags))
2817 bio->bi_opf |= MD_FAILFAST;
2823 r1_bio->read_disk = disk;
2825 if (read_targets == 0 && min_bad > 0) {
2826 /* These sectors are bad on all InSync devices, so we
2827 * need to mark them bad on all write targets
2830 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2831 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2832 struct md_rdev *rdev = conf->mirrors[i].rdev;
2833 ok = rdev_set_badblocks(rdev, sector_nr,
2837 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2842 /* Cannot record the badblocks, so need to
2844 * If there are multiple read targets, could just
2845 * fail the really bad ones ???
2847 conf->recovery_disabled = mddev->recovery_disabled;
2848 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2854 if (min_bad > 0 && min_bad < good_sectors) {
2855 /* only resync enough to reach the next bad->good
2857 good_sectors = min_bad;
2860 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2861 /* extra read targets are also write targets */
2862 write_targets += read_targets-1;
2864 if (write_targets == 0 || read_targets == 0) {
2865 /* There is nowhere to write, so all non-sync
2866 * drives must be failed - so we are finished
2870 max_sector = sector_nr + min_bad;
2871 rv = max_sector - sector_nr;
2877 if (max_sector > mddev->resync_max)
2878 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2879 if (max_sector > sector_nr + good_sectors)
2880 max_sector = sector_nr + good_sectors;
2885 int len = PAGE_SIZE;
2886 if (sector_nr + (len>>9) > max_sector)
2887 len = (max_sector - sector_nr) << 9;
2890 if (sync_blocks == 0) {
2891 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2892 &sync_blocks, still_degraded) &&
2894 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2896 if ((len >> 9) > sync_blocks)
2897 len = sync_blocks<<9;
2900 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2901 struct resync_pages *rp;
2903 bio = r1_bio->bios[i];
2904 rp = get_resync_pages(bio);
2905 if (bio->bi_end_io) {
2906 page = resync_fetch_page(rp, page_idx);
2909 * won't fail because the vec table is big
2910 * enough to hold all these pages
2912 bio_add_page(bio, page, len, 0);
2915 nr_sectors += len>>9;
2916 sector_nr += len>>9;
2917 sync_blocks -= (len>>9);
2918 } while (++page_idx < RESYNC_PAGES);
2920 r1_bio->sectors = nr_sectors;
2922 if (mddev_is_clustered(mddev) &&
2923 conf->cluster_sync_high < sector_nr + nr_sectors) {
2924 conf->cluster_sync_low = mddev->curr_resync_completed;
2925 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2926 /* Send resync message */
2927 md_cluster_ops->resync_info_update(mddev,
2928 conf->cluster_sync_low,
2929 conf->cluster_sync_high);
2932 /* For a user-requested sync, we read all readable devices and do a
2935 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2936 atomic_set(&r1_bio->remaining, read_targets);
2937 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2938 bio = r1_bio->bios[i];
2939 if (bio->bi_end_io == end_sync_read) {
2941 md_sync_acct_bio(bio, nr_sectors);
2942 if (read_targets == 1)
2943 bio->bi_opf &= ~MD_FAILFAST;
2944 submit_bio_noacct(bio);
2948 atomic_set(&r1_bio->remaining, 1);
2949 bio = r1_bio->bios[r1_bio->read_disk];
2950 md_sync_acct_bio(bio, nr_sectors);
2951 if (read_targets == 1)
2952 bio->bi_opf &= ~MD_FAILFAST;
2953 submit_bio_noacct(bio);
2958 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2963 return mddev->dev_sectors;
2966 static struct r1conf *setup_conf(struct mddev *mddev)
2968 struct r1conf *conf;
2970 struct raid1_info *disk;
2971 struct md_rdev *rdev;
2974 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2978 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2979 sizeof(atomic_t), GFP_KERNEL);
2980 if (!conf->nr_pending)
2983 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2984 sizeof(atomic_t), GFP_KERNEL);
2985 if (!conf->nr_waiting)
2988 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2989 sizeof(atomic_t), GFP_KERNEL);
2990 if (!conf->nr_queued)
2993 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2994 sizeof(atomic_t), GFP_KERNEL);
2998 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2999 mddev->raid_disks, 2),
3004 conf->tmppage = alloc_page(GFP_KERNEL);
3008 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3009 if (!conf->poolinfo)
3011 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3012 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3013 rbio_pool_free, conf->poolinfo);
3017 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3021 conf->poolinfo->mddev = mddev;
3024 spin_lock_init(&conf->device_lock);
3025 rdev_for_each(rdev, mddev) {
3026 int disk_idx = rdev->raid_disk;
3027 if (disk_idx >= mddev->raid_disks
3030 if (test_bit(Replacement, &rdev->flags))
3031 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3033 disk = conf->mirrors + disk_idx;
3038 disk->head_position = 0;
3039 disk->seq_start = MaxSector;
3041 conf->raid_disks = mddev->raid_disks;
3042 conf->mddev = mddev;
3043 INIT_LIST_HEAD(&conf->retry_list);
3044 INIT_LIST_HEAD(&conf->bio_end_io_list);
3046 spin_lock_init(&conf->resync_lock);
3047 init_waitqueue_head(&conf->wait_barrier);
3049 bio_list_init(&conf->pending_bio_list);
3050 conf->recovery_disabled = mddev->recovery_disabled - 1;
3053 for (i = 0; i < conf->raid_disks * 2; i++) {
3055 disk = conf->mirrors + i;
3057 if (i < conf->raid_disks &&
3058 disk[conf->raid_disks].rdev) {
3059 /* This slot has a replacement. */
3061 /* No original, just make the replacement
3062 * a recovering spare
3065 disk[conf->raid_disks].rdev;
3066 disk[conf->raid_disks].rdev = NULL;
3067 } else if (!test_bit(In_sync, &disk->rdev->flags))
3068 /* Original is not in_sync - bad */
3073 !test_bit(In_sync, &disk->rdev->flags)) {
3074 disk->head_position = 0;
3076 (disk->rdev->saved_raid_disk < 0))
3082 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3090 mempool_exit(&conf->r1bio_pool);
3091 kfree(conf->mirrors);
3092 safe_put_page(conf->tmppage);
3093 kfree(conf->poolinfo);
3094 kfree(conf->nr_pending);
3095 kfree(conf->nr_waiting);
3096 kfree(conf->nr_queued);
3097 kfree(conf->barrier);
3098 bioset_exit(&conf->bio_split);
3101 return ERR_PTR(err);
3104 static void raid1_free(struct mddev *mddev, void *priv);
3105 static int raid1_run(struct mddev *mddev)
3107 struct r1conf *conf;
3109 struct md_rdev *rdev;
3111 bool discard_supported = false;
3113 if (mddev->level != 1) {
3114 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3115 mdname(mddev), mddev->level);
3118 if (mddev->reshape_position != MaxSector) {
3119 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3123 if (mddev_init_writes_pending(mddev) < 0)
3126 * copy the already verified devices into our private RAID1
3127 * bookkeeping area. [whatever we allocate in run(),
3128 * should be freed in raid1_free()]
3130 if (mddev->private == NULL)
3131 conf = setup_conf(mddev);
3133 conf = mddev->private;
3136 return PTR_ERR(conf);
3139 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3141 rdev_for_each(rdev, mddev) {
3142 if (!mddev->gendisk)
3144 disk_stack_limits(mddev->gendisk, rdev->bdev,
3145 rdev->data_offset << 9);
3146 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3147 discard_supported = true;
3150 mddev->degraded = 0;
3151 for (i = 0; i < conf->raid_disks; i++)
3152 if (conf->mirrors[i].rdev == NULL ||
3153 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3154 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3157 * RAID1 needs at least one disk in active
3159 if (conf->raid_disks - mddev->degraded < 1) {
3164 if (conf->raid_disks - mddev->degraded == 1)
3165 mddev->recovery_cp = MaxSector;
3167 if (mddev->recovery_cp != MaxSector)
3168 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3170 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3171 mdname(mddev), mddev->raid_disks - mddev->degraded,
3175 * Ok, everything is just fine now
3177 mddev->thread = conf->thread;
3178 conf->thread = NULL;
3179 mddev->private = conf;
3180 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3182 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3185 if (discard_supported)
3186 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3189 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3193 ret = md_integrity_register(mddev);
3195 md_unregister_thread(&mddev->thread);
3201 raid1_free(mddev, conf);
3205 static void raid1_free(struct mddev *mddev, void *priv)
3207 struct r1conf *conf = priv;
3209 mempool_exit(&conf->r1bio_pool);
3210 kfree(conf->mirrors);
3211 safe_put_page(conf->tmppage);
3212 kfree(conf->poolinfo);
3213 kfree(conf->nr_pending);
3214 kfree(conf->nr_waiting);
3215 kfree(conf->nr_queued);
3216 kfree(conf->barrier);
3217 bioset_exit(&conf->bio_split);
3221 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3223 /* no resync is happening, and there is enough space
3224 * on all devices, so we can resize.
3225 * We need to make sure resync covers any new space.
3226 * If the array is shrinking we should possibly wait until
3227 * any io in the removed space completes, but it hardly seems
3230 sector_t newsize = raid1_size(mddev, sectors, 0);
3231 if (mddev->external_size &&
3232 mddev->array_sectors > newsize)
3234 if (mddev->bitmap) {
3235 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3239 md_set_array_sectors(mddev, newsize);
3240 if (sectors > mddev->dev_sectors &&
3241 mddev->recovery_cp > mddev->dev_sectors) {
3242 mddev->recovery_cp = mddev->dev_sectors;
3243 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3245 mddev->dev_sectors = sectors;
3246 mddev->resync_max_sectors = sectors;
3250 static int raid1_reshape(struct mddev *mddev)
3253 * 1/ resize the r1bio_pool
3254 * 2/ resize conf->mirrors
3256 * We allocate a new r1bio_pool if we can.
3257 * Then raise a device barrier and wait until all IO stops.
3258 * Then resize conf->mirrors and swap in the new r1bio pool.
3260 * At the same time, we "pack" the devices so that all the missing
3261 * devices have the higher raid_disk numbers.
3263 mempool_t newpool, oldpool;
3264 struct pool_info *newpoolinfo;
3265 struct raid1_info *newmirrors;
3266 struct r1conf *conf = mddev->private;
3267 int cnt, raid_disks;
3268 unsigned long flags;
3272 memset(&newpool, 0, sizeof(newpool));
3273 memset(&oldpool, 0, sizeof(oldpool));
3275 /* Cannot change chunk_size, layout, or level */
3276 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3277 mddev->layout != mddev->new_layout ||
3278 mddev->level != mddev->new_level) {
3279 mddev->new_chunk_sectors = mddev->chunk_sectors;
3280 mddev->new_layout = mddev->layout;
3281 mddev->new_level = mddev->level;
3285 if (!mddev_is_clustered(mddev))
3286 md_allow_write(mddev);
3288 raid_disks = mddev->raid_disks + mddev->delta_disks;
3290 if (raid_disks < conf->raid_disks) {
3292 for (d= 0; d < conf->raid_disks; d++)
3293 if (conf->mirrors[d].rdev)
3295 if (cnt > raid_disks)
3299 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3302 newpoolinfo->mddev = mddev;
3303 newpoolinfo->raid_disks = raid_disks * 2;
3305 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3306 rbio_pool_free, newpoolinfo);
3311 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3316 mempool_exit(&newpool);
3320 freeze_array(conf, 0);
3322 /* ok, everything is stopped */
3323 oldpool = conf->r1bio_pool;
3324 conf->r1bio_pool = newpool;
3326 for (d = d2 = 0; d < conf->raid_disks; d++) {
3327 struct md_rdev *rdev = conf->mirrors[d].rdev;
3328 if (rdev && rdev->raid_disk != d2) {
3329 sysfs_unlink_rdev(mddev, rdev);
3330 rdev->raid_disk = d2;
3331 sysfs_unlink_rdev(mddev, rdev);
3332 if (sysfs_link_rdev(mddev, rdev))
3333 pr_warn("md/raid1:%s: cannot register rd%d\n",
3334 mdname(mddev), rdev->raid_disk);
3337 newmirrors[d2++].rdev = rdev;
3339 kfree(conf->mirrors);
3340 conf->mirrors = newmirrors;
3341 kfree(conf->poolinfo);
3342 conf->poolinfo = newpoolinfo;
3344 spin_lock_irqsave(&conf->device_lock, flags);
3345 mddev->degraded += (raid_disks - conf->raid_disks);
3346 spin_unlock_irqrestore(&conf->device_lock, flags);
3347 conf->raid_disks = mddev->raid_disks = raid_disks;
3348 mddev->delta_disks = 0;
3350 unfreeze_array(conf);
3352 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3353 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3354 md_wakeup_thread(mddev->thread);
3356 mempool_exit(&oldpool);
3360 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3362 struct r1conf *conf = mddev->private;
3365 freeze_array(conf, 0);
3367 unfreeze_array(conf);
3370 static void *raid1_takeover(struct mddev *mddev)
3372 /* raid1 can take over:
3373 * raid5 with 2 devices, any layout or chunk size
3375 if (mddev->level == 5 && mddev->raid_disks == 2) {
3376 struct r1conf *conf;
3377 mddev->new_level = 1;
3378 mddev->new_layout = 0;
3379 mddev->new_chunk_sectors = 0;
3380 conf = setup_conf(mddev);
3381 if (!IS_ERR(conf)) {
3382 /* Array must appear to be quiesced */
3383 conf->array_frozen = 1;
3384 mddev_clear_unsupported_flags(mddev,
3385 UNSUPPORTED_MDDEV_FLAGS);
3389 return ERR_PTR(-EINVAL);
3392 static struct md_personality raid1_personality =
3396 .owner = THIS_MODULE,
3397 .make_request = raid1_make_request,
3400 .status = raid1_status,
3401 .error_handler = raid1_error,
3402 .hot_add_disk = raid1_add_disk,
3403 .hot_remove_disk= raid1_remove_disk,
3404 .spare_active = raid1_spare_active,
3405 .sync_request = raid1_sync_request,
3406 .resize = raid1_resize,
3408 .check_reshape = raid1_reshape,
3409 .quiesce = raid1_quiesce,
3410 .takeover = raid1_takeover,
3413 static int __init raid_init(void)
3415 return register_md_personality(&raid1_personality);
3418 static void raid_exit(void)
3420 unregister_md_personality(&raid1_personality);
3423 module_init(raid_init);
3424 module_exit(raid_exit);
3425 MODULE_LICENSE("GPL");
3426 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3427 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3428 MODULE_ALIAS("md-raid1");
3429 MODULE_ALIAS("md-level-1");