2 * raid1.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * RAID-1 management functions.
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
40 #include <linux/sched/signal.h>
42 #include <trace/events/block.h>
48 #define UNSUPPORTED_MDDEV_FLAGS \
49 ((1L << MD_HAS_JOURNAL) | \
50 (1L << MD_JOURNAL_CLEAN) | \
54 * Number of guaranteed r1bios in case of extreme VM load:
56 #define NR_RAID1_BIOS 256
58 /* when we get a read error on a read-only array, we redirect to another
59 * device without failing the first device, or trying to over-write to
60 * correct the read error. To keep track of bad blocks on a per-bio
61 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
63 #define IO_BLOCKED ((struct bio *)1)
64 /* When we successfully write to a known bad-block, we need to remove the
65 * bad-block marking which must be done from process context. So we record
66 * the success by setting devs[n].bio to IO_MADE_GOOD
68 #define IO_MADE_GOOD ((struct bio *)2)
70 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
72 /* When there are this many requests queue to be written by
73 * the raid1 thread, we become 'congested' to provide back-pressure
76 static int max_queued_requests = 1024;
78 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
79 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
81 #define raid1_log(md, fmt, args...) \
82 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
85 * 'strct resync_pages' stores actual pages used for doing the resync
86 * IO, and it is per-bio, so make .bi_private points to it.
88 static inline struct resync_pages *get_resync_pages(struct bio *bio)
90 return bio->bi_private;
94 * for resync bio, r1bio pointer can be retrieved from the per-bio
95 * 'struct resync_pages'.
97 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
99 return get_resync_pages(bio)->raid_bio;
102 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
104 struct pool_info *pi = data;
105 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
107 /* allocate a r1bio with room for raid_disks entries in the bios array */
108 return kzalloc(size, gfp_flags);
111 static void r1bio_pool_free(void *r1_bio, void *data)
116 #define RESYNC_DEPTH 32
117 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
118 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
119 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
120 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
121 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
123 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
125 struct pool_info *pi = data;
126 struct r1bio *r1_bio;
130 struct resync_pages *rps;
132 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
136 rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks,
142 * Allocate bios : 1 for reading, n-1 for writing
144 for (j = pi->raid_disks ; j-- ; ) {
145 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
148 r1_bio->bios[j] = bio;
151 * Allocate RESYNC_PAGES data pages and attach them to
153 * If this is a user-requested check/repair, allocate
154 * RESYNC_PAGES for each bio.
156 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
157 need_pages = pi->raid_disks;
160 for (j = 0; j < pi->raid_disks; j++) {
161 struct resync_pages *rp = &rps[j];
163 bio = r1_bio->bios[j];
165 if (j < need_pages) {
166 if (resync_alloc_pages(rp, gfp_flags))
169 memcpy(rp, &rps[0], sizeof(*rp));
170 resync_get_all_pages(rp);
174 rp->raid_bio = r1_bio;
175 bio->bi_private = rp;
178 r1_bio->master_bio = NULL;
184 resync_free_pages(&rps[j]);
187 while (++j < pi->raid_disks)
188 bio_put(r1_bio->bios[j]);
192 r1bio_pool_free(r1_bio, data);
196 static void r1buf_pool_free(void *__r1_bio, void *data)
198 struct pool_info *pi = data;
200 struct r1bio *r1bio = __r1_bio;
201 struct resync_pages *rp = NULL;
203 for (i = pi->raid_disks; i--; ) {
204 rp = get_resync_pages(r1bio->bios[i]);
205 resync_free_pages(rp);
206 bio_put(r1bio->bios[i]);
209 /* resync pages array stored in the 1st bio's .bi_private */
212 r1bio_pool_free(r1bio, data);
215 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
219 for (i = 0; i < conf->raid_disks * 2; i++) {
220 struct bio **bio = r1_bio->bios + i;
221 if (!BIO_SPECIAL(*bio))
227 static void free_r1bio(struct r1bio *r1_bio)
229 struct r1conf *conf = r1_bio->mddev->private;
231 put_all_bios(conf, r1_bio);
232 mempool_free(r1_bio, conf->r1bio_pool);
235 static void put_buf(struct r1bio *r1_bio)
237 struct r1conf *conf = r1_bio->mddev->private;
238 sector_t sect = r1_bio->sector;
241 for (i = 0; i < conf->raid_disks * 2; i++) {
242 struct bio *bio = r1_bio->bios[i];
244 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
247 mempool_free(r1_bio, conf->r1buf_pool);
249 lower_barrier(conf, sect);
252 static void reschedule_retry(struct r1bio *r1_bio)
255 struct mddev *mddev = r1_bio->mddev;
256 struct r1conf *conf = mddev->private;
259 idx = sector_to_idx(r1_bio->sector);
260 spin_lock_irqsave(&conf->device_lock, flags);
261 list_add(&r1_bio->retry_list, &conf->retry_list);
262 atomic_inc(&conf->nr_queued[idx]);
263 spin_unlock_irqrestore(&conf->device_lock, flags);
265 wake_up(&conf->wait_barrier);
266 md_wakeup_thread(mddev->thread);
270 * raid_end_bio_io() is called when we have finished servicing a mirrored
271 * operation and are ready to return a success/failure code to the buffer
274 static void call_bio_endio(struct r1bio *r1_bio)
276 struct bio *bio = r1_bio->master_bio;
277 struct r1conf *conf = r1_bio->mddev->private;
279 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
280 bio->bi_error = -EIO;
284 * Wake up any possible resync thread that waits for the device
287 allow_barrier(conf, r1_bio->sector);
290 static void raid_end_bio_io(struct r1bio *r1_bio)
292 struct bio *bio = r1_bio->master_bio;
294 /* if nobody has done the final endio yet, do it now */
295 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
296 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
297 (bio_data_dir(bio) == WRITE) ? "write" : "read",
298 (unsigned long long) bio->bi_iter.bi_sector,
299 (unsigned long long) bio_end_sector(bio) - 1);
301 call_bio_endio(r1_bio);
307 * Update disk head position estimator based on IRQ completion info.
309 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
311 struct r1conf *conf = r1_bio->mddev->private;
313 conf->mirrors[disk].head_position =
314 r1_bio->sector + (r1_bio->sectors);
318 * Find the disk number which triggered given bio
320 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
323 struct r1conf *conf = r1_bio->mddev->private;
324 int raid_disks = conf->raid_disks;
326 for (mirror = 0; mirror < raid_disks * 2; mirror++)
327 if (r1_bio->bios[mirror] == bio)
330 BUG_ON(mirror == raid_disks * 2);
331 update_head_pos(mirror, r1_bio);
336 static void raid1_end_read_request(struct bio *bio)
338 int uptodate = !bio->bi_error;
339 struct r1bio *r1_bio = bio->bi_private;
340 struct r1conf *conf = r1_bio->mddev->private;
341 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
344 * this branch is our 'one mirror IO has finished' event handler:
346 update_head_pos(r1_bio->read_disk, r1_bio);
349 set_bit(R1BIO_Uptodate, &r1_bio->state);
350 else if (test_bit(FailFast, &rdev->flags) &&
351 test_bit(R1BIO_FailFast, &r1_bio->state))
352 /* This was a fail-fast read so we definitely
356 /* If all other devices have failed, we want to return
357 * the error upwards rather than fail the last device.
358 * Here we redefine "uptodate" to mean "Don't want to retry"
361 spin_lock_irqsave(&conf->device_lock, flags);
362 if (r1_bio->mddev->degraded == conf->raid_disks ||
363 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
364 test_bit(In_sync, &rdev->flags)))
366 spin_unlock_irqrestore(&conf->device_lock, flags);
370 raid_end_bio_io(r1_bio);
371 rdev_dec_pending(rdev, conf->mddev);
376 char b[BDEVNAME_SIZE];
377 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
379 bdevname(rdev->bdev, b),
380 (unsigned long long)r1_bio->sector);
381 set_bit(R1BIO_ReadError, &r1_bio->state);
382 reschedule_retry(r1_bio);
383 /* don't drop the reference on read_disk yet */
387 static void close_write(struct r1bio *r1_bio)
389 /* it really is the end of this request */
390 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
391 bio_free_pages(r1_bio->behind_master_bio);
392 bio_put(r1_bio->behind_master_bio);
393 r1_bio->behind_master_bio = NULL;
395 /* clear the bitmap if all writes complete successfully */
396 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
398 !test_bit(R1BIO_Degraded, &r1_bio->state),
399 test_bit(R1BIO_BehindIO, &r1_bio->state));
400 md_write_end(r1_bio->mddev);
403 static void r1_bio_write_done(struct r1bio *r1_bio)
405 if (!atomic_dec_and_test(&r1_bio->remaining))
408 if (test_bit(R1BIO_WriteError, &r1_bio->state))
409 reschedule_retry(r1_bio);
412 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
413 reschedule_retry(r1_bio);
415 raid_end_bio_io(r1_bio);
419 static void raid1_end_write_request(struct bio *bio)
421 struct r1bio *r1_bio = bio->bi_private;
422 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
423 struct r1conf *conf = r1_bio->mddev->private;
424 struct bio *to_put = NULL;
425 int mirror = find_bio_disk(r1_bio, bio);
426 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
429 discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD;
432 * 'one mirror IO has finished' event handler:
434 if (bio->bi_error && !discard_error) {
435 set_bit(WriteErrorSeen, &rdev->flags);
436 if (!test_and_set_bit(WantReplacement, &rdev->flags))
437 set_bit(MD_RECOVERY_NEEDED, &
438 conf->mddev->recovery);
440 if (test_bit(FailFast, &rdev->flags) &&
441 (bio->bi_opf & MD_FAILFAST) &&
442 /* We never try FailFast to WriteMostly devices */
443 !test_bit(WriteMostly, &rdev->flags)) {
444 md_error(r1_bio->mddev, rdev);
445 if (!test_bit(Faulty, &rdev->flags))
446 /* This is the only remaining device,
447 * We need to retry the write without
450 set_bit(R1BIO_WriteError, &r1_bio->state);
452 /* Finished with this branch */
453 r1_bio->bios[mirror] = NULL;
457 set_bit(R1BIO_WriteError, &r1_bio->state);
460 * Set R1BIO_Uptodate in our master bio, so that we
461 * will return a good error code for to the higher
462 * levels even if IO on some other mirrored buffer
465 * The 'master' represents the composite IO operation
466 * to user-side. So if something waits for IO, then it
467 * will wait for the 'master' bio.
472 r1_bio->bios[mirror] = NULL;
475 * Do not set R1BIO_Uptodate if the current device is
476 * rebuilding or Faulty. This is because we cannot use
477 * such device for properly reading the data back (we could
478 * potentially use it, if the current write would have felt
479 * before rdev->recovery_offset, but for simplicity we don't
482 if (test_bit(In_sync, &rdev->flags) &&
483 !test_bit(Faulty, &rdev->flags))
484 set_bit(R1BIO_Uptodate, &r1_bio->state);
486 /* Maybe we can clear some bad blocks. */
487 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
488 &first_bad, &bad_sectors) && !discard_error) {
489 r1_bio->bios[mirror] = IO_MADE_GOOD;
490 set_bit(R1BIO_MadeGood, &r1_bio->state);
495 /* we release behind master bio when all write are done */
496 if (r1_bio->behind_master_bio == bio)
499 if (test_bit(WriteMostly, &rdev->flags))
500 atomic_dec(&r1_bio->behind_remaining);
503 * In behind mode, we ACK the master bio once the I/O
504 * has safely reached all non-writemostly
505 * disks. Setting the Returned bit ensures that this
506 * gets done only once -- we don't ever want to return
507 * -EIO here, instead we'll wait
509 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
510 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
511 /* Maybe we can return now */
512 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
513 struct bio *mbio = r1_bio->master_bio;
514 pr_debug("raid1: behind end write sectors"
516 (unsigned long long) mbio->bi_iter.bi_sector,
517 (unsigned long long) bio_end_sector(mbio) - 1);
518 call_bio_endio(r1_bio);
522 if (r1_bio->bios[mirror] == NULL)
523 rdev_dec_pending(rdev, conf->mddev);
526 * Let's see if all mirrored write operations have finished
529 r1_bio_write_done(r1_bio);
535 static sector_t align_to_barrier_unit_end(sector_t start_sector,
540 WARN_ON(sectors == 0);
542 * len is the number of sectors from start_sector to end of the
543 * barrier unit which start_sector belongs to.
545 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
555 * This routine returns the disk from which the requested read should
556 * be done. There is a per-array 'next expected sequential IO' sector
557 * number - if this matches on the next IO then we use the last disk.
558 * There is also a per-disk 'last know head position' sector that is
559 * maintained from IRQ contexts, both the normal and the resync IO
560 * completion handlers update this position correctly. If there is no
561 * perfect sequential match then we pick the disk whose head is closest.
563 * If there are 2 mirrors in the same 2 devices, performance degrades
564 * because position is mirror, not device based.
566 * The rdev for the device selected will have nr_pending incremented.
568 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
570 const sector_t this_sector = r1_bio->sector;
572 int best_good_sectors;
573 int best_disk, best_dist_disk, best_pending_disk;
577 unsigned int min_pending;
578 struct md_rdev *rdev;
580 int choose_next_idle;
584 * Check if we can balance. We can balance on the whole
585 * device if no resync is going on, or below the resync window.
586 * We take the first readable disk when above the resync window.
589 sectors = r1_bio->sectors;
592 best_dist = MaxSector;
593 best_pending_disk = -1;
594 min_pending = UINT_MAX;
595 best_good_sectors = 0;
597 choose_next_idle = 0;
598 clear_bit(R1BIO_FailFast, &r1_bio->state);
600 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
601 (mddev_is_clustered(conf->mddev) &&
602 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
603 this_sector + sectors)))
608 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
612 unsigned int pending;
615 rdev = rcu_dereference(conf->mirrors[disk].rdev);
616 if (r1_bio->bios[disk] == IO_BLOCKED
618 || test_bit(Faulty, &rdev->flags))
620 if (!test_bit(In_sync, &rdev->flags) &&
621 rdev->recovery_offset < this_sector + sectors)
623 if (test_bit(WriteMostly, &rdev->flags)) {
624 /* Don't balance among write-mostly, just
625 * use the first as a last resort */
626 if (best_dist_disk < 0) {
627 if (is_badblock(rdev, this_sector, sectors,
628 &first_bad, &bad_sectors)) {
629 if (first_bad <= this_sector)
630 /* Cannot use this */
632 best_good_sectors = first_bad - this_sector;
634 best_good_sectors = sectors;
635 best_dist_disk = disk;
636 best_pending_disk = disk;
640 /* This is a reasonable device to use. It might
643 if (is_badblock(rdev, this_sector, sectors,
644 &first_bad, &bad_sectors)) {
645 if (best_dist < MaxSector)
646 /* already have a better device */
648 if (first_bad <= this_sector) {
649 /* cannot read here. If this is the 'primary'
650 * device, then we must not read beyond
651 * bad_sectors from another device..
653 bad_sectors -= (this_sector - first_bad);
654 if (choose_first && sectors > bad_sectors)
655 sectors = bad_sectors;
656 if (best_good_sectors > sectors)
657 best_good_sectors = sectors;
660 sector_t good_sectors = first_bad - this_sector;
661 if (good_sectors > best_good_sectors) {
662 best_good_sectors = good_sectors;
670 best_good_sectors = sectors;
673 /* At least two disks to choose from so failfast is OK */
674 set_bit(R1BIO_FailFast, &r1_bio->state);
676 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
677 has_nonrot_disk |= nonrot;
678 pending = atomic_read(&rdev->nr_pending);
679 dist = abs(this_sector - conf->mirrors[disk].head_position);
684 /* Don't change to another disk for sequential reads */
685 if (conf->mirrors[disk].next_seq_sect == this_sector
687 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
688 struct raid1_info *mirror = &conf->mirrors[disk];
692 * If buffered sequential IO size exceeds optimal
693 * iosize, check if there is idle disk. If yes, choose
694 * the idle disk. read_balance could already choose an
695 * idle disk before noticing it's a sequential IO in
696 * this disk. This doesn't matter because this disk
697 * will idle, next time it will be utilized after the
698 * first disk has IO size exceeds optimal iosize. In
699 * this way, iosize of the first disk will be optimal
700 * iosize at least. iosize of the second disk might be
701 * small, but not a big deal since when the second disk
702 * starts IO, the first disk is likely still busy.
704 if (nonrot && opt_iosize > 0 &&
705 mirror->seq_start != MaxSector &&
706 mirror->next_seq_sect > opt_iosize &&
707 mirror->next_seq_sect - opt_iosize >=
709 choose_next_idle = 1;
715 if (choose_next_idle)
718 if (min_pending > pending) {
719 min_pending = pending;
720 best_pending_disk = disk;
723 if (dist < best_dist) {
725 best_dist_disk = disk;
730 * If all disks are rotational, choose the closest disk. If any disk is
731 * non-rotational, choose the disk with less pending request even the
732 * disk is rotational, which might/might not be optimal for raids with
733 * mixed ratation/non-rotational disks depending on workload.
735 if (best_disk == -1) {
736 if (has_nonrot_disk || min_pending == 0)
737 best_disk = best_pending_disk;
739 best_disk = best_dist_disk;
742 if (best_disk >= 0) {
743 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
746 atomic_inc(&rdev->nr_pending);
747 sectors = best_good_sectors;
749 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
750 conf->mirrors[best_disk].seq_start = this_sector;
752 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
755 *max_sectors = sectors;
760 static int raid1_congested(struct mddev *mddev, int bits)
762 struct r1conf *conf = mddev->private;
765 if ((bits & (1 << WB_async_congested)) &&
766 conf->pending_count >= max_queued_requests)
770 for (i = 0; i < conf->raid_disks * 2; i++) {
771 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
772 if (rdev && !test_bit(Faulty, &rdev->flags)) {
773 struct request_queue *q = bdev_get_queue(rdev->bdev);
777 /* Note the '|| 1' - when read_balance prefers
778 * non-congested targets, it can be removed
780 if ((bits & (1 << WB_async_congested)) || 1)
781 ret |= bdi_congested(q->backing_dev_info, bits);
783 ret &= bdi_congested(q->backing_dev_info, bits);
790 static void flush_pending_writes(struct r1conf *conf)
792 /* Any writes that have been queued but are awaiting
793 * bitmap updates get flushed here.
795 spin_lock_irq(&conf->device_lock);
797 if (conf->pending_bio_list.head) {
799 bio = bio_list_get(&conf->pending_bio_list);
800 conf->pending_count = 0;
801 spin_unlock_irq(&conf->device_lock);
802 /* flush any pending bitmap writes to
803 * disk before proceeding w/ I/O */
804 bitmap_unplug(conf->mddev->bitmap);
805 wake_up(&conf->wait_barrier);
807 while (bio) { /* submit pending writes */
808 struct bio *next = bio->bi_next;
809 struct md_rdev *rdev = (void*)bio->bi_bdev;
811 bio->bi_bdev = rdev->bdev;
812 if (test_bit(Faulty, &rdev->flags)) {
813 bio->bi_error = -EIO;
815 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
816 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
820 generic_make_request(bio);
824 spin_unlock_irq(&conf->device_lock);
828 * Sometimes we need to suspend IO while we do something else,
829 * either some resync/recovery, or reconfigure the array.
830 * To do this we raise a 'barrier'.
831 * The 'barrier' is a counter that can be raised multiple times
832 * to count how many activities are happening which preclude
834 * We can only raise the barrier if there is no pending IO.
835 * i.e. if nr_pending == 0.
836 * We choose only to raise the barrier if no-one is waiting for the
837 * barrier to go down. This means that as soon as an IO request
838 * is ready, no other operations which require a barrier will start
839 * until the IO request has had a chance.
841 * So: regular IO calls 'wait_barrier'. When that returns there
842 * is no backgroup IO happening, It must arrange to call
843 * allow_barrier when it has finished its IO.
844 * backgroup IO calls must call raise_barrier. Once that returns
845 * there is no normal IO happeing. It must arrange to call
846 * lower_barrier when the particular background IO completes.
848 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
850 int idx = sector_to_idx(sector_nr);
852 spin_lock_irq(&conf->resync_lock);
854 /* Wait until no block IO is waiting */
855 wait_event_lock_irq(conf->wait_barrier,
856 !atomic_read(&conf->nr_waiting[idx]),
859 /* block any new IO from starting */
860 atomic_inc(&conf->barrier[idx]);
862 * In raise_barrier() we firstly increase conf->barrier[idx] then
863 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
864 * increase conf->nr_pending[idx] then check conf->barrier[idx].
865 * A memory barrier here to make sure conf->nr_pending[idx] won't
866 * be fetched before conf->barrier[idx] is increased. Otherwise
867 * there will be a race between raise_barrier() and _wait_barrier().
869 smp_mb__after_atomic();
871 /* For these conditions we must wait:
872 * A: while the array is in frozen state
873 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
874 * existing in corresponding I/O barrier bucket.
875 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
876 * max resync count which allowed on current I/O barrier bucket.
878 wait_event_lock_irq(conf->wait_barrier,
879 !conf->array_frozen &&
880 !atomic_read(&conf->nr_pending[idx]) &&
881 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH,
884 atomic_inc(&conf->nr_pending[idx]);
885 spin_unlock_irq(&conf->resync_lock);
888 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
890 int idx = sector_to_idx(sector_nr);
892 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
894 atomic_dec(&conf->barrier[idx]);
895 atomic_dec(&conf->nr_pending[idx]);
896 wake_up(&conf->wait_barrier);
899 static void _wait_barrier(struct r1conf *conf, int idx)
902 * We need to increase conf->nr_pending[idx] very early here,
903 * then raise_barrier() can be blocked when it waits for
904 * conf->nr_pending[idx] to be 0. Then we can avoid holding
905 * conf->resync_lock when there is no barrier raised in same
906 * barrier unit bucket. Also if the array is frozen, I/O
907 * should be blocked until array is unfrozen.
909 atomic_inc(&conf->nr_pending[idx]);
911 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
912 * check conf->barrier[idx]. In raise_barrier() we firstly increase
913 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
914 * barrier is necessary here to make sure conf->barrier[idx] won't be
915 * fetched before conf->nr_pending[idx] is increased. Otherwise there
916 * will be a race between _wait_barrier() and raise_barrier().
918 smp_mb__after_atomic();
921 * Don't worry about checking two atomic_t variables at same time
922 * here. If during we check conf->barrier[idx], the array is
923 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
924 * 0, it is safe to return and make the I/O continue. Because the
925 * array is frozen, all I/O returned here will eventually complete
926 * or be queued, no race will happen. See code comment in
929 if (!READ_ONCE(conf->array_frozen) &&
930 !atomic_read(&conf->barrier[idx]))
934 * After holding conf->resync_lock, conf->nr_pending[idx]
935 * should be decreased before waiting for barrier to drop.
936 * Otherwise, we may encounter a race condition because
937 * raise_barrer() might be waiting for conf->nr_pending[idx]
938 * to be 0 at same time.
940 spin_lock_irq(&conf->resync_lock);
941 atomic_inc(&conf->nr_waiting[idx]);
942 atomic_dec(&conf->nr_pending[idx]);
944 * In case freeze_array() is waiting for
945 * get_unqueued_pending() == extra
947 wake_up(&conf->wait_barrier);
948 /* Wait for the barrier in same barrier unit bucket to drop. */
949 wait_event_lock_irq(conf->wait_barrier,
950 !conf->array_frozen &&
951 !atomic_read(&conf->barrier[idx]),
953 atomic_inc(&conf->nr_pending[idx]);
954 atomic_dec(&conf->nr_waiting[idx]);
955 spin_unlock_irq(&conf->resync_lock);
958 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
960 int idx = sector_to_idx(sector_nr);
963 * Very similar to _wait_barrier(). The difference is, for read
964 * I/O we don't need wait for sync I/O, but if the whole array
965 * is frozen, the read I/O still has to wait until the array is
966 * unfrozen. Since there is no ordering requirement with
967 * conf->barrier[idx] here, memory barrier is unnecessary as well.
969 atomic_inc(&conf->nr_pending[idx]);
971 if (!READ_ONCE(conf->array_frozen))
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 array to be unfrozen */
983 wait_event_lock_irq(conf->wait_barrier,
986 atomic_inc(&conf->nr_pending[idx]);
987 atomic_dec(&conf->nr_waiting[idx]);
988 spin_unlock_irq(&conf->resync_lock);
991 static void inc_pending(struct r1conf *conf, sector_t bi_sector)
993 /* The current request requires multiple r1_bio, so
994 * we need to increment the pending count, and the corresponding
997 int idx = sector_to_idx(bi_sector);
998 atomic_inc(&conf->nr_pending[idx]);
1001 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1003 int idx = sector_to_idx(sector_nr);
1005 _wait_barrier(conf, idx);
1008 static void wait_all_barriers(struct r1conf *conf)
1012 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1013 _wait_barrier(conf, idx);
1016 static void _allow_barrier(struct r1conf *conf, int idx)
1018 atomic_dec(&conf->nr_pending[idx]);
1019 wake_up(&conf->wait_barrier);
1022 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1024 int idx = sector_to_idx(sector_nr);
1026 _allow_barrier(conf, idx);
1029 static void allow_all_barriers(struct r1conf *conf)
1033 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1034 _allow_barrier(conf, idx);
1037 /* conf->resync_lock should be held */
1038 static int get_unqueued_pending(struct r1conf *conf)
1042 for (ret = 0, idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1043 ret += atomic_read(&conf->nr_pending[idx]) -
1044 atomic_read(&conf->nr_queued[idx]);
1049 static void freeze_array(struct r1conf *conf, int extra)
1051 /* Stop sync I/O and normal I/O and wait for everything to
1053 * This is called in two situations:
1054 * 1) management command handlers (reshape, remove disk, quiesce).
1055 * 2) one normal I/O request failed.
1057 * After array_frozen is set to 1, new sync IO will be blocked at
1058 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1059 * or wait_read_barrier(). The flying I/Os will either complete or be
1060 * queued. When everything goes quite, there are only queued I/Os left.
1062 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1063 * barrier bucket index which this I/O request hits. When all sync and
1064 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1065 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1066 * in handle_read_error(), we may call freeze_array() before trying to
1067 * fix the read error. In this case, the error read I/O is not queued,
1068 * so get_unqueued_pending() == 1.
1070 * Therefore before this function returns, we need to wait until
1071 * get_unqueued_pendings(conf) gets equal to extra. For
1072 * normal I/O context, extra is 1, in rested situations extra is 0.
1074 spin_lock_irq(&conf->resync_lock);
1075 conf->array_frozen = 1;
1076 raid1_log(conf->mddev, "wait freeze");
1077 wait_event_lock_irq_cmd(
1079 get_unqueued_pending(conf) == extra,
1081 flush_pending_writes(conf));
1082 spin_unlock_irq(&conf->resync_lock);
1084 static void unfreeze_array(struct r1conf *conf)
1086 /* reverse the effect of the freeze */
1087 spin_lock_irq(&conf->resync_lock);
1088 conf->array_frozen = 0;
1089 spin_unlock_irq(&conf->resync_lock);
1090 wake_up(&conf->wait_barrier);
1093 static struct bio *alloc_behind_master_bio(struct r1bio *r1_bio,
1095 int offset, int size)
1097 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1099 struct bio *behind_bio = NULL;
1101 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1105 /* discard op, we don't support writezero/writesame yet */
1106 if (!bio_has_data(bio))
1109 while (i < vcnt && size) {
1111 int len = min_t(int, PAGE_SIZE, size);
1113 page = alloc_page(GFP_NOIO);
1114 if (unlikely(!page))
1117 bio_add_page(behind_bio, page, len, 0);
1123 bio_copy_data_partial(behind_bio, bio, offset,
1124 behind_bio->bi_iter.bi_size);
1126 r1_bio->behind_master_bio = behind_bio;;
1127 set_bit(R1BIO_BehindIO, &r1_bio->state);
1132 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1133 bio->bi_iter.bi_size);
1134 bio_free_pages(behind_bio);
1139 struct raid1_plug_cb {
1140 struct blk_plug_cb cb;
1141 struct bio_list pending;
1145 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1147 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1149 struct mddev *mddev = plug->cb.data;
1150 struct r1conf *conf = mddev->private;
1153 if (from_schedule || current->bio_list) {
1154 spin_lock_irq(&conf->device_lock);
1155 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1156 conf->pending_count += plug->pending_cnt;
1157 spin_unlock_irq(&conf->device_lock);
1158 wake_up(&conf->wait_barrier);
1159 md_wakeup_thread(mddev->thread);
1164 /* we aren't scheduling, so we can do the write-out directly. */
1165 bio = bio_list_get(&plug->pending);
1166 bitmap_unplug(mddev->bitmap);
1167 wake_up(&conf->wait_barrier);
1169 while (bio) { /* submit pending writes */
1170 struct bio *next = bio->bi_next;
1171 struct md_rdev *rdev = (void*)bio->bi_bdev;
1172 bio->bi_next = NULL;
1173 bio->bi_bdev = rdev->bdev;
1174 if (test_bit(Faulty, &rdev->flags)) {
1175 bio->bi_error = -EIO;
1177 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1178 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1179 /* Just ignore it */
1182 generic_make_request(bio);
1188 static inline struct r1bio *
1189 alloc_r1bio(struct mddev *mddev, struct bio *bio, sector_t sectors_handled)
1191 struct r1conf *conf = mddev->private;
1192 struct r1bio *r1_bio;
1194 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1196 r1_bio->master_bio = bio;
1197 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1199 r1_bio->mddev = mddev;
1200 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1205 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1206 int max_read_sectors)
1208 struct r1conf *conf = mddev->private;
1209 struct raid1_info *mirror;
1210 struct r1bio *r1_bio;
1211 struct bio *read_bio;
1212 struct bitmap *bitmap = mddev->bitmap;
1213 const int op = bio_op(bio);
1214 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1219 * Still need barrier for READ in case that whole
1222 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1224 r1_bio = alloc_r1bio(mddev, bio, 0);
1225 r1_bio->sectors = max_read_sectors;
1228 * make_request() can abort the operation when read-ahead is being
1229 * used and no empty request is available.
1231 rdisk = read_balance(conf, r1_bio, &max_sectors);
1234 /* couldn't find anywhere to read from */
1235 raid_end_bio_io(r1_bio);
1238 mirror = conf->mirrors + rdisk;
1240 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1243 * Reading from a write-mostly device must take care not to
1244 * over-take any writes that are 'behind'
1246 raid1_log(mddev, "wait behind writes");
1247 wait_event(bitmap->behind_wait,
1248 atomic_read(&bitmap->behind_writes) == 0);
1251 if (max_sectors < bio_sectors(bio)) {
1252 struct bio *split = bio_split(bio, max_sectors,
1253 GFP_NOIO, conf->bio_split);
1254 bio_chain(split, bio);
1255 generic_make_request(bio);
1257 r1_bio->master_bio = bio;
1258 r1_bio->sectors = max_sectors;
1261 r1_bio->read_disk = rdisk;
1263 read_bio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1265 r1_bio->bios[rdisk] = read_bio;
1267 read_bio->bi_iter.bi_sector = r1_bio->sector +
1268 mirror->rdev->data_offset;
1269 read_bio->bi_bdev = mirror->rdev->bdev;
1270 read_bio->bi_end_io = raid1_end_read_request;
1271 bio_set_op_attrs(read_bio, op, do_sync);
1272 if (test_bit(FailFast, &mirror->rdev->flags) &&
1273 test_bit(R1BIO_FailFast, &r1_bio->state))
1274 read_bio->bi_opf |= MD_FAILFAST;
1275 read_bio->bi_private = r1_bio;
1278 trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1279 read_bio, disk_devt(mddev->gendisk),
1282 generic_make_request(read_bio);
1285 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1286 int max_write_sectors)
1288 struct r1conf *conf = mddev->private;
1289 struct r1bio *r1_bio;
1291 struct bitmap *bitmap = mddev->bitmap;
1292 unsigned long flags;
1293 struct md_rdev *blocked_rdev;
1294 struct blk_plug_cb *cb;
1295 struct raid1_plug_cb *plug = NULL;
1300 * Register the new request and wait if the reconstruction
1301 * thread has put up a bar for new requests.
1302 * Continue immediately if no resync is active currently.
1305 md_write_start(mddev, bio); /* wait on superblock update early */
1307 if ((bio_end_sector(bio) > mddev->suspend_lo &&
1308 bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1309 (mddev_is_clustered(mddev) &&
1310 md_cluster_ops->area_resyncing(mddev, WRITE,
1311 bio->bi_iter.bi_sector, bio_end_sector(bio)))) {
1314 * As the suspend_* range is controlled by userspace, we want
1315 * an interruptible wait.
1319 flush_signals(current);
1320 prepare_to_wait(&conf->wait_barrier,
1321 &w, TASK_INTERRUPTIBLE);
1322 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1323 bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1324 (mddev_is_clustered(mddev) &&
1325 !md_cluster_ops->area_resyncing(mddev, WRITE,
1326 bio->bi_iter.bi_sector,
1327 bio_end_sector(bio))))
1331 finish_wait(&conf->wait_barrier, &w);
1333 wait_barrier(conf, bio->bi_iter.bi_sector);
1335 r1_bio = alloc_r1bio(mddev, bio, 0);
1336 r1_bio->sectors = max_write_sectors;
1338 if (conf->pending_count >= max_queued_requests) {
1339 md_wakeup_thread(mddev->thread);
1340 raid1_log(mddev, "wait queued");
1341 wait_event(conf->wait_barrier,
1342 conf->pending_count < max_queued_requests);
1344 /* first select target devices under rcu_lock and
1345 * inc refcount on their rdev. Record them by setting
1347 * If there are known/acknowledged bad blocks on any device on
1348 * which we have seen a write error, we want to avoid writing those
1350 * This potentially requires several writes to write around
1351 * the bad blocks. Each set of writes gets it's own r1bio
1352 * with a set of bios attached.
1355 disks = conf->raid_disks * 2;
1357 blocked_rdev = NULL;
1359 max_sectors = r1_bio->sectors;
1360 for (i = 0; i < disks; i++) {
1361 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1362 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1363 atomic_inc(&rdev->nr_pending);
1364 blocked_rdev = rdev;
1367 r1_bio->bios[i] = NULL;
1368 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1369 if (i < conf->raid_disks)
1370 set_bit(R1BIO_Degraded, &r1_bio->state);
1374 atomic_inc(&rdev->nr_pending);
1375 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1380 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1381 &first_bad, &bad_sectors);
1383 /* mustn't write here until the bad block is
1385 set_bit(BlockedBadBlocks, &rdev->flags);
1386 blocked_rdev = rdev;
1389 if (is_bad && first_bad <= r1_bio->sector) {
1390 /* Cannot write here at all */
1391 bad_sectors -= (r1_bio->sector - first_bad);
1392 if (bad_sectors < max_sectors)
1393 /* mustn't write more than bad_sectors
1394 * to other devices yet
1396 max_sectors = bad_sectors;
1397 rdev_dec_pending(rdev, mddev);
1398 /* We don't set R1BIO_Degraded as that
1399 * only applies if the disk is
1400 * missing, so it might be re-added,
1401 * and we want to know to recover this
1403 * In this case the device is here,
1404 * and the fact that this chunk is not
1405 * in-sync is recorded in the bad
1411 int good_sectors = first_bad - r1_bio->sector;
1412 if (good_sectors < max_sectors)
1413 max_sectors = good_sectors;
1416 r1_bio->bios[i] = bio;
1420 if (unlikely(blocked_rdev)) {
1421 /* Wait for this device to become unblocked */
1424 for (j = 0; j < i; j++)
1425 if (r1_bio->bios[j])
1426 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1428 allow_barrier(conf, bio->bi_iter.bi_sector);
1429 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1430 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1431 wait_barrier(conf, bio->bi_iter.bi_sector);
1435 if (max_sectors < bio_sectors(bio)) {
1436 struct bio *split = bio_split(bio, max_sectors,
1437 GFP_NOIO, conf->bio_split);
1438 bio_chain(split, bio);
1439 generic_make_request(bio);
1441 r1_bio->master_bio = bio;
1442 r1_bio->sectors = max_sectors;
1445 atomic_set(&r1_bio->remaining, 1);
1446 atomic_set(&r1_bio->behind_remaining, 0);
1450 for (i = 0; i < disks; i++) {
1451 struct bio *mbio = NULL;
1452 if (!r1_bio->bios[i])
1458 * Not if there are too many, or cannot
1459 * allocate memory, or a reader on WriteMostly
1460 * is waiting for behind writes to flush */
1462 (atomic_read(&bitmap->behind_writes)
1463 < mddev->bitmap_info.max_write_behind) &&
1464 !waitqueue_active(&bitmap->behind_wait)) {
1465 mbio = alloc_behind_master_bio(r1_bio, bio,
1470 bitmap_startwrite(bitmap, r1_bio->sector,
1472 test_bit(R1BIO_BehindIO,
1478 if (r1_bio->behind_master_bio)
1479 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1483 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1486 if (r1_bio->behind_master_bio) {
1487 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1488 atomic_inc(&r1_bio->behind_remaining);
1491 r1_bio->bios[i] = mbio;
1493 mbio->bi_iter.bi_sector = (r1_bio->sector +
1494 conf->mirrors[i].rdev->data_offset);
1495 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1496 mbio->bi_end_io = raid1_end_write_request;
1497 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1498 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1499 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1500 conf->raid_disks - mddev->degraded > 1)
1501 mbio->bi_opf |= MD_FAILFAST;
1502 mbio->bi_private = r1_bio;
1504 atomic_inc(&r1_bio->remaining);
1507 trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1508 mbio, disk_devt(mddev->gendisk),
1510 /* flush_pending_writes() needs access to the rdev so...*/
1511 mbio->bi_bdev = (void*)conf->mirrors[i].rdev;
1513 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1515 plug = container_of(cb, struct raid1_plug_cb, cb);
1518 spin_lock_irqsave(&conf->device_lock, flags);
1520 bio_list_add(&plug->pending, mbio);
1521 plug->pending_cnt++;
1523 bio_list_add(&conf->pending_bio_list, mbio);
1524 conf->pending_count++;
1526 spin_unlock_irqrestore(&conf->device_lock, flags);
1528 md_wakeup_thread(mddev->thread);
1531 r1_bio_write_done(r1_bio);
1533 /* In case raid1d snuck in to freeze_array */
1534 wake_up(&conf->wait_barrier);
1537 static void raid1_make_request(struct mddev *mddev, struct bio *bio)
1541 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1542 md_flush_request(mddev, bio);
1547 * There is a limit to the maximum size, but
1548 * the read/write handler might find a lower limit
1549 * due to bad blocks. To avoid multiple splits,
1550 * we pass the maximum number of sectors down
1551 * and let the lower level perform the split.
1553 sectors = align_to_barrier_unit_end(
1554 bio->bi_iter.bi_sector, bio_sectors(bio));
1556 if (bio_data_dir(bio) == READ)
1557 raid1_read_request(mddev, bio, sectors);
1559 raid1_write_request(mddev, bio, sectors);
1562 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1564 struct r1conf *conf = mddev->private;
1567 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1568 conf->raid_disks - mddev->degraded);
1570 for (i = 0; i < conf->raid_disks; i++) {
1571 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1572 seq_printf(seq, "%s",
1573 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1576 seq_printf(seq, "]");
1579 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1581 char b[BDEVNAME_SIZE];
1582 struct r1conf *conf = mddev->private;
1583 unsigned long flags;
1586 * If it is not operational, then we have already marked it as dead
1587 * else if it is the last working disks, ignore the error, let the
1588 * next level up know.
1589 * else mark the drive as failed
1591 spin_lock_irqsave(&conf->device_lock, flags);
1592 if (test_bit(In_sync, &rdev->flags)
1593 && (conf->raid_disks - mddev->degraded) == 1) {
1595 * Don't fail the drive, act as though we were just a
1596 * normal single drive.
1597 * However don't try a recovery from this drive as
1598 * it is very likely to fail.
1600 conf->recovery_disabled = mddev->recovery_disabled;
1601 spin_unlock_irqrestore(&conf->device_lock, flags);
1604 set_bit(Blocked, &rdev->flags);
1605 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1607 set_bit(Faulty, &rdev->flags);
1609 set_bit(Faulty, &rdev->flags);
1610 spin_unlock_irqrestore(&conf->device_lock, flags);
1612 * if recovery is running, make sure it aborts.
1614 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1615 set_mask_bits(&mddev->sb_flags, 0,
1616 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1617 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1618 "md/raid1:%s: Operation continuing on %d devices.\n",
1619 mdname(mddev), bdevname(rdev->bdev, b),
1620 mdname(mddev), conf->raid_disks - mddev->degraded);
1623 static void print_conf(struct r1conf *conf)
1627 pr_debug("RAID1 conf printout:\n");
1629 pr_debug("(!conf)\n");
1632 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1636 for (i = 0; i < conf->raid_disks; i++) {
1637 char b[BDEVNAME_SIZE];
1638 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1640 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1641 i, !test_bit(In_sync, &rdev->flags),
1642 !test_bit(Faulty, &rdev->flags),
1643 bdevname(rdev->bdev,b));
1648 static void close_sync(struct r1conf *conf)
1650 wait_all_barriers(conf);
1651 allow_all_barriers(conf);
1653 mempool_destroy(conf->r1buf_pool);
1654 conf->r1buf_pool = NULL;
1657 static int raid1_spare_active(struct mddev *mddev)
1660 struct r1conf *conf = mddev->private;
1662 unsigned long flags;
1665 * Find all failed disks within the RAID1 configuration
1666 * and mark them readable.
1667 * Called under mddev lock, so rcu protection not needed.
1668 * device_lock used to avoid races with raid1_end_read_request
1669 * which expects 'In_sync' flags and ->degraded to be consistent.
1671 spin_lock_irqsave(&conf->device_lock, flags);
1672 for (i = 0; i < conf->raid_disks; i++) {
1673 struct md_rdev *rdev = conf->mirrors[i].rdev;
1674 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1676 && !test_bit(Candidate, &repl->flags)
1677 && repl->recovery_offset == MaxSector
1678 && !test_bit(Faulty, &repl->flags)
1679 && !test_and_set_bit(In_sync, &repl->flags)) {
1680 /* replacement has just become active */
1682 !test_and_clear_bit(In_sync, &rdev->flags))
1685 /* Replaced device not technically
1686 * faulty, but we need to be sure
1687 * it gets removed and never re-added
1689 set_bit(Faulty, &rdev->flags);
1690 sysfs_notify_dirent_safe(
1695 && rdev->recovery_offset == MaxSector
1696 && !test_bit(Faulty, &rdev->flags)
1697 && !test_and_set_bit(In_sync, &rdev->flags)) {
1699 sysfs_notify_dirent_safe(rdev->sysfs_state);
1702 mddev->degraded -= count;
1703 spin_unlock_irqrestore(&conf->device_lock, flags);
1709 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1711 struct r1conf *conf = mddev->private;
1714 struct raid1_info *p;
1716 int last = conf->raid_disks - 1;
1718 if (mddev->recovery_disabled == conf->recovery_disabled)
1721 if (md_integrity_add_rdev(rdev, mddev))
1724 if (rdev->raid_disk >= 0)
1725 first = last = rdev->raid_disk;
1728 * find the disk ... but prefer rdev->saved_raid_disk
1731 if (rdev->saved_raid_disk >= 0 &&
1732 rdev->saved_raid_disk >= first &&
1733 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1734 first = last = rdev->saved_raid_disk;
1736 for (mirror = first; mirror <= last; mirror++) {
1737 p = conf->mirrors+mirror;
1741 disk_stack_limits(mddev->gendisk, rdev->bdev,
1742 rdev->data_offset << 9);
1744 p->head_position = 0;
1745 rdev->raid_disk = mirror;
1747 /* As all devices are equivalent, we don't need a full recovery
1748 * if this was recently any drive of the array
1750 if (rdev->saved_raid_disk < 0)
1752 rcu_assign_pointer(p->rdev, rdev);
1755 if (test_bit(WantReplacement, &p->rdev->flags) &&
1756 p[conf->raid_disks].rdev == NULL) {
1757 /* Add this device as a replacement */
1758 clear_bit(In_sync, &rdev->flags);
1759 set_bit(Replacement, &rdev->flags);
1760 rdev->raid_disk = mirror;
1763 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1767 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1768 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1773 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1775 struct r1conf *conf = mddev->private;
1777 int number = rdev->raid_disk;
1778 struct raid1_info *p = conf->mirrors + number;
1780 if (rdev != p->rdev)
1781 p = conf->mirrors + conf->raid_disks + number;
1784 if (rdev == p->rdev) {
1785 if (test_bit(In_sync, &rdev->flags) ||
1786 atomic_read(&rdev->nr_pending)) {
1790 /* Only remove non-faulty devices if recovery
1793 if (!test_bit(Faulty, &rdev->flags) &&
1794 mddev->recovery_disabled != conf->recovery_disabled &&
1795 mddev->degraded < conf->raid_disks) {
1800 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1802 if (atomic_read(&rdev->nr_pending)) {
1803 /* lost the race, try later */
1809 if (conf->mirrors[conf->raid_disks + number].rdev) {
1810 /* We just removed a device that is being replaced.
1811 * Move down the replacement. We drain all IO before
1812 * doing this to avoid confusion.
1814 struct md_rdev *repl =
1815 conf->mirrors[conf->raid_disks + number].rdev;
1816 freeze_array(conf, 0);
1817 clear_bit(Replacement, &repl->flags);
1819 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1820 unfreeze_array(conf);
1821 clear_bit(WantReplacement, &rdev->flags);
1823 clear_bit(WantReplacement, &rdev->flags);
1824 err = md_integrity_register(mddev);
1832 static void end_sync_read(struct bio *bio)
1834 struct r1bio *r1_bio = get_resync_r1bio(bio);
1836 update_head_pos(r1_bio->read_disk, r1_bio);
1839 * we have read a block, now it needs to be re-written,
1840 * or re-read if the read failed.
1841 * We don't do much here, just schedule handling by raid1d
1844 set_bit(R1BIO_Uptodate, &r1_bio->state);
1846 if (atomic_dec_and_test(&r1_bio->remaining))
1847 reschedule_retry(r1_bio);
1850 static void end_sync_write(struct bio *bio)
1852 int uptodate = !bio->bi_error;
1853 struct r1bio *r1_bio = get_resync_r1bio(bio);
1854 struct mddev *mddev = r1_bio->mddev;
1855 struct r1conf *conf = mddev->private;
1858 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1861 sector_t sync_blocks = 0;
1862 sector_t s = r1_bio->sector;
1863 long sectors_to_go = r1_bio->sectors;
1864 /* make sure these bits doesn't get cleared. */
1866 bitmap_end_sync(mddev->bitmap, s,
1869 sectors_to_go -= sync_blocks;
1870 } while (sectors_to_go > 0);
1871 set_bit(WriteErrorSeen, &rdev->flags);
1872 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1873 set_bit(MD_RECOVERY_NEEDED, &
1875 set_bit(R1BIO_WriteError, &r1_bio->state);
1876 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1877 &first_bad, &bad_sectors) &&
1878 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1881 &first_bad, &bad_sectors)
1883 set_bit(R1BIO_MadeGood, &r1_bio->state);
1885 if (atomic_dec_and_test(&r1_bio->remaining)) {
1886 int s = r1_bio->sectors;
1887 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1888 test_bit(R1BIO_WriteError, &r1_bio->state))
1889 reschedule_retry(r1_bio);
1892 md_done_sync(mddev, s, uptodate);
1897 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1898 int sectors, struct page *page, int rw)
1900 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1904 set_bit(WriteErrorSeen, &rdev->flags);
1905 if (!test_and_set_bit(WantReplacement,
1907 set_bit(MD_RECOVERY_NEEDED, &
1908 rdev->mddev->recovery);
1910 /* need to record an error - either for the block or the device */
1911 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1912 md_error(rdev->mddev, rdev);
1916 static int fix_sync_read_error(struct r1bio *r1_bio)
1918 /* Try some synchronous reads of other devices to get
1919 * good data, much like with normal read errors. Only
1920 * read into the pages we already have so we don't
1921 * need to re-issue the read request.
1922 * We don't need to freeze the array, because being in an
1923 * active sync request, there is no normal IO, and
1924 * no overlapping syncs.
1925 * We don't need to check is_badblock() again as we
1926 * made sure that anything with a bad block in range
1927 * will have bi_end_io clear.
1929 struct mddev *mddev = r1_bio->mddev;
1930 struct r1conf *conf = mddev->private;
1931 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1932 struct page **pages = get_resync_pages(bio)->pages;
1933 sector_t sect = r1_bio->sector;
1934 int sectors = r1_bio->sectors;
1936 struct md_rdev *rdev;
1938 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1939 if (test_bit(FailFast, &rdev->flags)) {
1940 /* Don't try recovering from here - just fail it
1941 * ... unless it is the last working device of course */
1942 md_error(mddev, rdev);
1943 if (test_bit(Faulty, &rdev->flags))
1944 /* Don't try to read from here, but make sure
1945 * put_buf does it's thing
1947 bio->bi_end_io = end_sync_write;
1952 int d = r1_bio->read_disk;
1956 if (s > (PAGE_SIZE>>9))
1959 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1960 /* No rcu protection needed here devices
1961 * can only be removed when no resync is
1962 * active, and resync is currently active
1964 rdev = conf->mirrors[d].rdev;
1965 if (sync_page_io(rdev, sect, s<<9,
1967 REQ_OP_READ, 0, false)) {
1973 if (d == conf->raid_disks * 2)
1975 } while (!success && d != r1_bio->read_disk);
1978 char b[BDEVNAME_SIZE];
1980 /* Cannot read from anywhere, this block is lost.
1981 * Record a bad block on each device. If that doesn't
1982 * work just disable and interrupt the recovery.
1983 * Don't fail devices as that won't really help.
1985 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1987 bdevname(bio->bi_bdev, b),
1988 (unsigned long long)r1_bio->sector);
1989 for (d = 0; d < conf->raid_disks * 2; d++) {
1990 rdev = conf->mirrors[d].rdev;
1991 if (!rdev || test_bit(Faulty, &rdev->flags))
1993 if (!rdev_set_badblocks(rdev, sect, s, 0))
1997 conf->recovery_disabled =
1998 mddev->recovery_disabled;
1999 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2000 md_done_sync(mddev, r1_bio->sectors, 0);
2012 /* write it back and re-read */
2013 while (d != r1_bio->read_disk) {
2015 d = conf->raid_disks * 2;
2017 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2019 rdev = conf->mirrors[d].rdev;
2020 if (r1_sync_page_io(rdev, sect, s,
2023 r1_bio->bios[d]->bi_end_io = NULL;
2024 rdev_dec_pending(rdev, mddev);
2028 while (d != r1_bio->read_disk) {
2030 d = conf->raid_disks * 2;
2032 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2034 rdev = conf->mirrors[d].rdev;
2035 if (r1_sync_page_io(rdev, sect, s,
2038 atomic_add(s, &rdev->corrected_errors);
2044 set_bit(R1BIO_Uptodate, &r1_bio->state);
2049 static void process_checks(struct r1bio *r1_bio)
2051 /* We have read all readable devices. If we haven't
2052 * got the block, then there is no hope left.
2053 * If we have, then we want to do a comparison
2054 * and skip the write if everything is the same.
2055 * If any blocks failed to read, then we need to
2056 * attempt an over-write
2058 struct mddev *mddev = r1_bio->mddev;
2059 struct r1conf *conf = mddev->private;
2064 /* Fix variable parts of all bios */
2065 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2066 for (i = 0; i < conf->raid_disks * 2; i++) {
2071 struct bio *b = r1_bio->bios[i];
2072 struct resync_pages *rp = get_resync_pages(b);
2073 if (b->bi_end_io != end_sync_read)
2075 /* fixup the bio for reuse, but preserve errno */
2076 error = b->bi_error;
2078 b->bi_error = error;
2080 b->bi_iter.bi_size = r1_bio->sectors << 9;
2081 b->bi_iter.bi_sector = r1_bio->sector +
2082 conf->mirrors[i].rdev->data_offset;
2083 b->bi_bdev = conf->mirrors[i].rdev->bdev;
2084 b->bi_end_io = end_sync_read;
2085 rp->raid_bio = r1_bio;
2088 size = b->bi_iter.bi_size;
2089 bio_for_each_segment_all(bi, b, j) {
2091 if (size > PAGE_SIZE)
2092 bi->bv_len = PAGE_SIZE;
2098 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2099 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2100 !r1_bio->bios[primary]->bi_error) {
2101 r1_bio->bios[primary]->bi_end_io = NULL;
2102 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2105 r1_bio->read_disk = primary;
2106 for (i = 0; i < conf->raid_disks * 2; i++) {
2108 struct bio *pbio = r1_bio->bios[primary];
2109 struct bio *sbio = r1_bio->bios[i];
2110 int error = sbio->bi_error;
2111 struct page **ppages = get_resync_pages(pbio)->pages;
2112 struct page **spages = get_resync_pages(sbio)->pages;
2114 int page_len[RESYNC_PAGES] = { 0 };
2116 if (sbio->bi_end_io != end_sync_read)
2118 /* Now we can 'fixup' the error value */
2121 bio_for_each_segment_all(bi, sbio, j)
2122 page_len[j] = bi->bv_len;
2125 for (j = vcnt; j-- ; ) {
2126 if (memcmp(page_address(ppages[j]),
2127 page_address(spages[j]),
2134 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2135 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2137 /* No need to write to this device. */
2138 sbio->bi_end_io = NULL;
2139 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2143 bio_copy_data(sbio, pbio);
2147 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2149 struct r1conf *conf = mddev->private;
2151 int disks = conf->raid_disks * 2;
2152 struct bio *bio, *wbio;
2154 bio = r1_bio->bios[r1_bio->read_disk];
2156 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2157 /* ouch - failed to read all of that. */
2158 if (!fix_sync_read_error(r1_bio))
2161 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2162 process_checks(r1_bio);
2167 atomic_set(&r1_bio->remaining, 1);
2168 for (i = 0; i < disks ; i++) {
2169 wbio = r1_bio->bios[i];
2170 if (wbio->bi_end_io == NULL ||
2171 (wbio->bi_end_io == end_sync_read &&
2172 (i == r1_bio->read_disk ||
2173 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2175 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2178 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2179 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2180 wbio->bi_opf |= MD_FAILFAST;
2182 wbio->bi_end_io = end_sync_write;
2183 atomic_inc(&r1_bio->remaining);
2184 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2186 generic_make_request(wbio);
2189 if (atomic_dec_and_test(&r1_bio->remaining)) {
2190 /* if we're here, all write(s) have completed, so clean up */
2191 int s = r1_bio->sectors;
2192 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2193 test_bit(R1BIO_WriteError, &r1_bio->state))
2194 reschedule_retry(r1_bio);
2197 md_done_sync(mddev, s, 1);
2203 * This is a kernel thread which:
2205 * 1. Retries failed read operations on working mirrors.
2206 * 2. Updates the raid superblock when problems encounter.
2207 * 3. Performs writes following reads for array synchronising.
2210 static void fix_read_error(struct r1conf *conf, int read_disk,
2211 sector_t sect, int sectors)
2213 struct mddev *mddev = conf->mddev;
2219 struct md_rdev *rdev;
2221 if (s > (PAGE_SIZE>>9))
2229 rdev = rcu_dereference(conf->mirrors[d].rdev);
2231 (test_bit(In_sync, &rdev->flags) ||
2232 (!test_bit(Faulty, &rdev->flags) &&
2233 rdev->recovery_offset >= sect + s)) &&
2234 is_badblock(rdev, sect, s,
2235 &first_bad, &bad_sectors) == 0) {
2236 atomic_inc(&rdev->nr_pending);
2238 if (sync_page_io(rdev, sect, s<<9,
2239 conf->tmppage, REQ_OP_READ, 0, false))
2241 rdev_dec_pending(rdev, mddev);
2247 if (d == conf->raid_disks * 2)
2249 } while (!success && d != read_disk);
2252 /* Cannot read from anywhere - mark it bad */
2253 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2254 if (!rdev_set_badblocks(rdev, sect, s, 0))
2255 md_error(mddev, rdev);
2258 /* write it back and re-read */
2260 while (d != read_disk) {
2262 d = conf->raid_disks * 2;
2265 rdev = rcu_dereference(conf->mirrors[d].rdev);
2267 !test_bit(Faulty, &rdev->flags)) {
2268 atomic_inc(&rdev->nr_pending);
2270 r1_sync_page_io(rdev, sect, s,
2271 conf->tmppage, WRITE);
2272 rdev_dec_pending(rdev, mddev);
2277 while (d != read_disk) {
2278 char b[BDEVNAME_SIZE];
2280 d = conf->raid_disks * 2;
2283 rdev = rcu_dereference(conf->mirrors[d].rdev);
2285 !test_bit(Faulty, &rdev->flags)) {
2286 atomic_inc(&rdev->nr_pending);
2288 if (r1_sync_page_io(rdev, sect, s,
2289 conf->tmppage, READ)) {
2290 atomic_add(s, &rdev->corrected_errors);
2291 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2293 (unsigned long long)(sect +
2295 bdevname(rdev->bdev, b));
2297 rdev_dec_pending(rdev, mddev);
2306 static int narrow_write_error(struct r1bio *r1_bio, int i)
2308 struct mddev *mddev = r1_bio->mddev;
2309 struct r1conf *conf = mddev->private;
2310 struct md_rdev *rdev = conf->mirrors[i].rdev;
2312 /* bio has the data to be written to device 'i' where
2313 * we just recently had a write error.
2314 * We repeatedly clone the bio and trim down to one block,
2315 * then try the write. Where the write fails we record
2317 * It is conceivable that the bio doesn't exactly align with
2318 * blocks. We must handle this somehow.
2320 * We currently own a reference on the rdev.
2326 int sect_to_write = r1_bio->sectors;
2329 if (rdev->badblocks.shift < 0)
2332 block_sectors = roundup(1 << rdev->badblocks.shift,
2333 bdev_logical_block_size(rdev->bdev) >> 9);
2334 sector = r1_bio->sector;
2335 sectors = ((sector + block_sectors)
2336 & ~(sector_t)(block_sectors - 1))
2339 while (sect_to_write) {
2341 if (sectors > sect_to_write)
2342 sectors = sect_to_write;
2343 /* Write at 'sector' for 'sectors'*/
2345 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2346 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2349 /* We really need a _all clone */
2350 wbio->bi_iter = (struct bvec_iter){ 0 };
2352 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2356 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2357 wbio->bi_iter.bi_sector = r1_bio->sector;
2358 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2360 bio_trim(wbio, sector - r1_bio->sector, sectors);
2361 wbio->bi_iter.bi_sector += rdev->data_offset;
2362 wbio->bi_bdev = rdev->bdev;
2364 if (submit_bio_wait(wbio) < 0)
2366 ok = rdev_set_badblocks(rdev, sector,
2371 sect_to_write -= sectors;
2373 sectors = block_sectors;
2378 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2381 int s = r1_bio->sectors;
2382 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2383 struct md_rdev *rdev = conf->mirrors[m].rdev;
2384 struct bio *bio = r1_bio->bios[m];
2385 if (bio->bi_end_io == NULL)
2387 if (!bio->bi_error &&
2388 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2389 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2391 if (bio->bi_error &&
2392 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2393 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2394 md_error(conf->mddev, rdev);
2398 md_done_sync(conf->mddev, s, 1);
2401 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2406 for (m = 0; m < conf->raid_disks * 2 ; m++)
2407 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2408 struct md_rdev *rdev = conf->mirrors[m].rdev;
2409 rdev_clear_badblocks(rdev,
2411 r1_bio->sectors, 0);
2412 rdev_dec_pending(rdev, conf->mddev);
2413 } else if (r1_bio->bios[m] != NULL) {
2414 /* This drive got a write error. We need to
2415 * narrow down and record precise write
2419 if (!narrow_write_error(r1_bio, m)) {
2420 md_error(conf->mddev,
2421 conf->mirrors[m].rdev);
2422 /* an I/O failed, we can't clear the bitmap */
2423 set_bit(R1BIO_Degraded, &r1_bio->state);
2425 rdev_dec_pending(conf->mirrors[m].rdev,
2429 spin_lock_irq(&conf->device_lock);
2430 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2431 idx = sector_to_idx(r1_bio->sector);
2432 atomic_inc(&conf->nr_queued[idx]);
2433 spin_unlock_irq(&conf->device_lock);
2435 * In case freeze_array() is waiting for condition
2436 * get_unqueued_pending() == extra to be true.
2438 wake_up(&conf->wait_barrier);
2439 md_wakeup_thread(conf->mddev->thread);
2441 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2442 close_write(r1_bio);
2443 raid_end_bio_io(r1_bio);
2447 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2451 struct mddev *mddev = conf->mddev;
2453 char b[BDEVNAME_SIZE];
2454 struct md_rdev *rdev;
2456 sector_t bio_sector;
2458 clear_bit(R1BIO_ReadError, &r1_bio->state);
2459 /* we got a read error. Maybe the drive is bad. Maybe just
2460 * the block and we can fix it.
2461 * We freeze all other IO, and try reading the block from
2462 * other devices. When we find one, we re-write
2463 * and check it that fixes the read error.
2464 * This is all done synchronously while the array is
2468 bio = r1_bio->bios[r1_bio->read_disk];
2469 bdevname(bio->bi_bdev, b);
2470 bio_dev = bio->bi_bdev->bd_dev;
2471 bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2473 r1_bio->bios[r1_bio->read_disk] = NULL;
2475 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2477 && !test_bit(FailFast, &rdev->flags)) {
2478 freeze_array(conf, 1);
2479 fix_read_error(conf, r1_bio->read_disk,
2480 r1_bio->sector, r1_bio->sectors);
2481 unfreeze_array(conf);
2483 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2486 rdev_dec_pending(rdev, conf->mddev);
2489 disk = read_balance(conf, r1_bio, &max_sectors);
2491 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2492 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2493 raid_end_bio_io(r1_bio);
2495 const unsigned long do_sync
2496 = r1_bio->master_bio->bi_opf & REQ_SYNC;
2497 r1_bio->read_disk = disk;
2498 bio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2500 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2502 r1_bio->bios[r1_bio->read_disk] = bio;
2503 rdev = conf->mirrors[disk].rdev;
2504 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
2506 (unsigned long long)r1_bio->sector,
2507 bdevname(rdev->bdev, b));
2508 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2509 bio->bi_bdev = rdev->bdev;
2510 bio->bi_end_io = raid1_end_read_request;
2511 bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
2512 if (test_bit(FailFast, &rdev->flags) &&
2513 test_bit(R1BIO_FailFast, &r1_bio->state))
2514 bio->bi_opf |= MD_FAILFAST;
2515 bio->bi_private = r1_bio;
2516 if (max_sectors < r1_bio->sectors) {
2517 /* Drat - have to split this up more */
2518 struct bio *mbio = r1_bio->master_bio;
2519 int sectors_handled = (r1_bio->sector + max_sectors
2520 - mbio->bi_iter.bi_sector);
2521 r1_bio->sectors = max_sectors;
2522 bio_inc_remaining(mbio);
2523 trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
2524 bio, bio_dev, bio_sector);
2525 generic_make_request(bio);
2528 r1_bio = alloc_r1bio(mddev, mbio, sectors_handled);
2529 set_bit(R1BIO_ReadError, &r1_bio->state);
2530 inc_pending(conf, r1_bio->sector);
2534 trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
2535 bio, bio_dev, bio_sector);
2536 generic_make_request(bio);
2541 static void raid1d(struct md_thread *thread)
2543 struct mddev *mddev = thread->mddev;
2544 struct r1bio *r1_bio;
2545 unsigned long flags;
2546 struct r1conf *conf = mddev->private;
2547 struct list_head *head = &conf->retry_list;
2548 struct blk_plug plug;
2551 md_check_recovery(mddev);
2553 if (!list_empty_careful(&conf->bio_end_io_list) &&
2554 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2556 spin_lock_irqsave(&conf->device_lock, flags);
2557 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2558 list_splice_init(&conf->bio_end_io_list, &tmp);
2559 spin_unlock_irqrestore(&conf->device_lock, flags);
2560 while (!list_empty(&tmp)) {
2561 r1_bio = list_first_entry(&tmp, struct r1bio,
2563 list_del(&r1_bio->retry_list);
2564 idx = sector_to_idx(r1_bio->sector);
2565 atomic_dec(&conf->nr_queued[idx]);
2566 if (mddev->degraded)
2567 set_bit(R1BIO_Degraded, &r1_bio->state);
2568 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2569 close_write(r1_bio);
2570 raid_end_bio_io(r1_bio);
2574 blk_start_plug(&plug);
2577 flush_pending_writes(conf);
2579 spin_lock_irqsave(&conf->device_lock, flags);
2580 if (list_empty(head)) {
2581 spin_unlock_irqrestore(&conf->device_lock, flags);
2584 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2585 list_del(head->prev);
2586 idx = sector_to_idx(r1_bio->sector);
2587 atomic_dec(&conf->nr_queued[idx]);
2588 spin_unlock_irqrestore(&conf->device_lock, flags);
2590 mddev = r1_bio->mddev;
2591 conf = mddev->private;
2592 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2593 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2594 test_bit(R1BIO_WriteError, &r1_bio->state))
2595 handle_sync_write_finished(conf, r1_bio);
2597 sync_request_write(mddev, r1_bio);
2598 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2599 test_bit(R1BIO_WriteError, &r1_bio->state))
2600 handle_write_finished(conf, r1_bio);
2601 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2602 handle_read_error(conf, r1_bio);
2607 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2608 md_check_recovery(mddev);
2610 blk_finish_plug(&plug);
2613 static int init_resync(struct r1conf *conf)
2617 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2618 BUG_ON(conf->r1buf_pool);
2619 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2621 if (!conf->r1buf_pool)
2627 * perform a "sync" on one "block"
2629 * We need to make sure that no normal I/O request - particularly write
2630 * requests - conflict with active sync requests.
2632 * This is achieved by tracking pending requests and a 'barrier' concept
2633 * that can be installed to exclude normal IO requests.
2636 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2639 struct r1conf *conf = mddev->private;
2640 struct r1bio *r1_bio;
2642 sector_t max_sector, nr_sectors;
2646 int write_targets = 0, read_targets = 0;
2647 sector_t sync_blocks;
2648 int still_degraded = 0;
2649 int good_sectors = RESYNC_SECTORS;
2650 int min_bad = 0; /* number of sectors that are bad in all devices */
2651 int idx = sector_to_idx(sector_nr);
2653 if (!conf->r1buf_pool)
2654 if (init_resync(conf))
2657 max_sector = mddev->dev_sectors;
2658 if (sector_nr >= max_sector) {
2659 /* If we aborted, we need to abort the
2660 * sync on the 'current' bitmap chunk (there will
2661 * only be one in raid1 resync.
2662 * We can find the current addess in mddev->curr_resync
2664 if (mddev->curr_resync < max_sector) /* aborted */
2665 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2667 else /* completed sync */
2670 bitmap_close_sync(mddev->bitmap);
2673 if (mddev_is_clustered(mddev)) {
2674 conf->cluster_sync_low = 0;
2675 conf->cluster_sync_high = 0;
2680 if (mddev->bitmap == NULL &&
2681 mddev->recovery_cp == MaxSector &&
2682 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2683 conf->fullsync == 0) {
2685 return max_sector - sector_nr;
2687 /* before building a request, check if we can skip these blocks..
2688 * This call the bitmap_start_sync doesn't actually record anything
2690 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2691 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2692 /* We can skip this block, and probably several more */
2698 * If there is non-resync activity waiting for a turn, then let it
2699 * though before starting on this new sync request.
2701 if (atomic_read(&conf->nr_waiting[idx]))
2702 schedule_timeout_uninterruptible(1);
2704 /* we are incrementing sector_nr below. To be safe, we check against
2705 * sector_nr + two times RESYNC_SECTORS
2708 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2709 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2710 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2712 raise_barrier(conf, sector_nr);
2716 * If we get a correctably read error during resync or recovery,
2717 * we might want to read from a different device. So we
2718 * flag all drives that could conceivably be read from for READ,
2719 * and any others (which will be non-In_sync devices) for WRITE.
2720 * If a read fails, we try reading from something else for which READ
2724 r1_bio->mddev = mddev;
2725 r1_bio->sector = sector_nr;
2727 set_bit(R1BIO_IsSync, &r1_bio->state);
2728 /* make sure good_sectors won't go across barrier unit boundary */
2729 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2731 for (i = 0; i < conf->raid_disks * 2; i++) {
2732 struct md_rdev *rdev;
2733 bio = r1_bio->bios[i];
2735 rdev = rcu_dereference(conf->mirrors[i].rdev);
2737 test_bit(Faulty, &rdev->flags)) {
2738 if (i < conf->raid_disks)
2740 } else if (!test_bit(In_sync, &rdev->flags)) {
2741 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2742 bio->bi_end_io = end_sync_write;
2745 /* may need to read from here */
2746 sector_t first_bad = MaxSector;
2749 if (is_badblock(rdev, sector_nr, good_sectors,
2750 &first_bad, &bad_sectors)) {
2751 if (first_bad > sector_nr)
2752 good_sectors = first_bad - sector_nr;
2754 bad_sectors -= (sector_nr - first_bad);
2756 min_bad > bad_sectors)
2757 min_bad = bad_sectors;
2760 if (sector_nr < first_bad) {
2761 if (test_bit(WriteMostly, &rdev->flags)) {
2768 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2769 bio->bi_end_io = end_sync_read;
2771 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2772 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2773 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2775 * The device is suitable for reading (InSync),
2776 * but has bad block(s) here. Let's try to correct them,
2777 * if we are doing resync or repair. Otherwise, leave
2778 * this device alone for this sync request.
2780 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2781 bio->bi_end_io = end_sync_write;
2785 if (bio->bi_end_io) {
2786 atomic_inc(&rdev->nr_pending);
2787 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2788 bio->bi_bdev = rdev->bdev;
2789 if (test_bit(FailFast, &rdev->flags))
2790 bio->bi_opf |= MD_FAILFAST;
2796 r1_bio->read_disk = disk;
2798 if (read_targets == 0 && min_bad > 0) {
2799 /* These sectors are bad on all InSync devices, so we
2800 * need to mark them bad on all write targets
2803 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2804 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2805 struct md_rdev *rdev = conf->mirrors[i].rdev;
2806 ok = rdev_set_badblocks(rdev, sector_nr,
2810 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2815 /* Cannot record the badblocks, so need to
2817 * If there are multiple read targets, could just
2818 * fail the really bad ones ???
2820 conf->recovery_disabled = mddev->recovery_disabled;
2821 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2827 if (min_bad > 0 && min_bad < good_sectors) {
2828 /* only resync enough to reach the next bad->good
2830 good_sectors = min_bad;
2833 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2834 /* extra read targets are also write targets */
2835 write_targets += read_targets-1;
2837 if (write_targets == 0 || read_targets == 0) {
2838 /* There is nowhere to write, so all non-sync
2839 * drives must be failed - so we are finished
2843 max_sector = sector_nr + min_bad;
2844 rv = max_sector - sector_nr;
2850 if (max_sector > mddev->resync_max)
2851 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2852 if (max_sector > sector_nr + good_sectors)
2853 max_sector = sector_nr + good_sectors;
2858 int len = PAGE_SIZE;
2859 if (sector_nr + (len>>9) > max_sector)
2860 len = (max_sector - sector_nr) << 9;
2863 if (sync_blocks == 0) {
2864 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2865 &sync_blocks, still_degraded) &&
2867 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2869 if ((len >> 9) > sync_blocks)
2870 len = sync_blocks<<9;
2873 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2874 struct resync_pages *rp;
2876 bio = r1_bio->bios[i];
2877 rp = get_resync_pages(bio);
2878 if (bio->bi_end_io) {
2879 page = resync_fetch_page(rp, rp->idx++);
2882 * won't fail because the vec table is big
2883 * enough to hold all these pages
2885 bio_add_page(bio, page, len, 0);
2888 nr_sectors += len>>9;
2889 sector_nr += len>>9;
2890 sync_blocks -= (len>>9);
2891 } while (get_resync_pages(r1_bio->bios[disk]->bi_private)->idx < RESYNC_PAGES);
2893 r1_bio->sectors = nr_sectors;
2895 if (mddev_is_clustered(mddev) &&
2896 conf->cluster_sync_high < sector_nr + nr_sectors) {
2897 conf->cluster_sync_low = mddev->curr_resync_completed;
2898 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2899 /* Send resync message */
2900 md_cluster_ops->resync_info_update(mddev,
2901 conf->cluster_sync_low,
2902 conf->cluster_sync_high);
2905 /* For a user-requested sync, we read all readable devices and do a
2908 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2909 atomic_set(&r1_bio->remaining, read_targets);
2910 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2911 bio = r1_bio->bios[i];
2912 if (bio->bi_end_io == end_sync_read) {
2914 md_sync_acct(bio->bi_bdev, nr_sectors);
2915 if (read_targets == 1)
2916 bio->bi_opf &= ~MD_FAILFAST;
2917 generic_make_request(bio);
2921 atomic_set(&r1_bio->remaining, 1);
2922 bio = r1_bio->bios[r1_bio->read_disk];
2923 md_sync_acct(bio->bi_bdev, nr_sectors);
2924 if (read_targets == 1)
2925 bio->bi_opf &= ~MD_FAILFAST;
2926 generic_make_request(bio);
2932 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2937 return mddev->dev_sectors;
2940 static struct r1conf *setup_conf(struct mddev *mddev)
2942 struct r1conf *conf;
2944 struct raid1_info *disk;
2945 struct md_rdev *rdev;
2948 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2952 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2953 sizeof(atomic_t), GFP_KERNEL);
2954 if (!conf->nr_pending)
2957 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2958 sizeof(atomic_t), GFP_KERNEL);
2959 if (!conf->nr_waiting)
2962 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2963 sizeof(atomic_t), GFP_KERNEL);
2964 if (!conf->nr_queued)
2967 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2968 sizeof(atomic_t), GFP_KERNEL);
2972 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2973 * mddev->raid_disks * 2,
2978 conf->tmppage = alloc_page(GFP_KERNEL);
2982 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2983 if (!conf->poolinfo)
2985 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2986 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2989 if (!conf->r1bio_pool)
2992 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0);
2993 if (!conf->bio_split)
2996 conf->poolinfo->mddev = mddev;
2999 spin_lock_init(&conf->device_lock);
3000 rdev_for_each(rdev, mddev) {
3001 struct request_queue *q;
3002 int disk_idx = rdev->raid_disk;
3003 if (disk_idx >= mddev->raid_disks
3006 if (test_bit(Replacement, &rdev->flags))
3007 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3009 disk = conf->mirrors + disk_idx;
3014 q = bdev_get_queue(rdev->bdev);
3016 disk->head_position = 0;
3017 disk->seq_start = MaxSector;
3019 conf->raid_disks = mddev->raid_disks;
3020 conf->mddev = mddev;
3021 INIT_LIST_HEAD(&conf->retry_list);
3022 INIT_LIST_HEAD(&conf->bio_end_io_list);
3024 spin_lock_init(&conf->resync_lock);
3025 init_waitqueue_head(&conf->wait_barrier);
3027 bio_list_init(&conf->pending_bio_list);
3028 conf->pending_count = 0;
3029 conf->recovery_disabled = mddev->recovery_disabled - 1;
3032 for (i = 0; i < conf->raid_disks * 2; i++) {
3034 disk = conf->mirrors + i;
3036 if (i < conf->raid_disks &&
3037 disk[conf->raid_disks].rdev) {
3038 /* This slot has a replacement. */
3040 /* No original, just make the replacement
3041 * a recovering spare
3044 disk[conf->raid_disks].rdev;
3045 disk[conf->raid_disks].rdev = NULL;
3046 } else if (!test_bit(In_sync, &disk->rdev->flags))
3047 /* Original is not in_sync - bad */
3052 !test_bit(In_sync, &disk->rdev->flags)) {
3053 disk->head_position = 0;
3055 (disk->rdev->saved_raid_disk < 0))
3061 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3069 mempool_destroy(conf->r1bio_pool);
3070 kfree(conf->mirrors);
3071 safe_put_page(conf->tmppage);
3072 kfree(conf->poolinfo);
3073 kfree(conf->nr_pending);
3074 kfree(conf->nr_waiting);
3075 kfree(conf->nr_queued);
3076 kfree(conf->barrier);
3077 if (conf->bio_split)
3078 bioset_free(conf->bio_split);
3081 return ERR_PTR(err);
3084 static void raid1_free(struct mddev *mddev, void *priv);
3085 static int raid1_run(struct mddev *mddev)
3087 struct r1conf *conf;
3089 struct md_rdev *rdev;
3091 bool discard_supported = false;
3093 if (mddev->level != 1) {
3094 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3095 mdname(mddev), mddev->level);
3098 if (mddev->reshape_position != MaxSector) {
3099 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3104 * copy the already verified devices into our private RAID1
3105 * bookkeeping area. [whatever we allocate in run(),
3106 * should be freed in raid1_free()]
3108 if (mddev->private == NULL)
3109 conf = setup_conf(mddev);
3111 conf = mddev->private;
3114 return PTR_ERR(conf);
3117 blk_queue_max_write_same_sectors(mddev->queue, 0);
3119 rdev_for_each(rdev, mddev) {
3120 if (!mddev->gendisk)
3122 disk_stack_limits(mddev->gendisk, rdev->bdev,
3123 rdev->data_offset << 9);
3124 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3125 discard_supported = true;
3128 mddev->degraded = 0;
3129 for (i=0; i < conf->raid_disks; i++)
3130 if (conf->mirrors[i].rdev == NULL ||
3131 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3132 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3135 if (conf->raid_disks - mddev->degraded == 1)
3136 mddev->recovery_cp = MaxSector;
3138 if (mddev->recovery_cp != MaxSector)
3139 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3141 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3142 mdname(mddev), mddev->raid_disks - mddev->degraded,
3146 * Ok, everything is just fine now
3148 mddev->thread = conf->thread;
3149 conf->thread = NULL;
3150 mddev->private = conf;
3151 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3153 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3156 if (discard_supported)
3157 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3160 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3164 ret = md_integrity_register(mddev);
3166 md_unregister_thread(&mddev->thread);
3167 raid1_free(mddev, conf);
3172 static void raid1_free(struct mddev *mddev, void *priv)
3174 struct r1conf *conf = priv;
3176 mempool_destroy(conf->r1bio_pool);
3177 kfree(conf->mirrors);
3178 safe_put_page(conf->tmppage);
3179 kfree(conf->poolinfo);
3180 kfree(conf->nr_pending);
3181 kfree(conf->nr_waiting);
3182 kfree(conf->nr_queued);
3183 kfree(conf->barrier);
3184 if (conf->bio_split)
3185 bioset_free(conf->bio_split);
3189 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3191 /* no resync is happening, and there is enough space
3192 * on all devices, so we can resize.
3193 * We need to make sure resync covers any new space.
3194 * If the array is shrinking we should possibly wait until
3195 * any io in the removed space completes, but it hardly seems
3198 sector_t newsize = raid1_size(mddev, sectors, 0);
3199 if (mddev->external_size &&
3200 mddev->array_sectors > newsize)
3202 if (mddev->bitmap) {
3203 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3207 md_set_array_sectors(mddev, newsize);
3208 if (sectors > mddev->dev_sectors &&
3209 mddev->recovery_cp > mddev->dev_sectors) {
3210 mddev->recovery_cp = mddev->dev_sectors;
3211 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3213 mddev->dev_sectors = sectors;
3214 mddev->resync_max_sectors = sectors;
3218 static int raid1_reshape(struct mddev *mddev)
3221 * 1/ resize the r1bio_pool
3222 * 2/ resize conf->mirrors
3224 * We allocate a new r1bio_pool if we can.
3225 * Then raise a device barrier and wait until all IO stops.
3226 * Then resize conf->mirrors and swap in the new r1bio pool.
3228 * At the same time, we "pack" the devices so that all the missing
3229 * devices have the higher raid_disk numbers.
3231 mempool_t *newpool, *oldpool;
3232 struct pool_info *newpoolinfo;
3233 struct raid1_info *newmirrors;
3234 struct r1conf *conf = mddev->private;
3235 int cnt, raid_disks;
3236 unsigned long flags;
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 err = md_allow_write(mddev);
3255 raid_disks = mddev->raid_disks + mddev->delta_disks;
3257 if (raid_disks < conf->raid_disks) {
3259 for (d= 0; d < conf->raid_disks; d++)
3260 if (conf->mirrors[d].rdev)
3262 if (cnt > raid_disks)
3266 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3269 newpoolinfo->mddev = mddev;
3270 newpoolinfo->raid_disks = raid_disks * 2;
3272 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3273 r1bio_pool_free, newpoolinfo);
3278 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3282 mempool_destroy(newpool);
3286 freeze_array(conf, 0);
3288 /* ok, everything is stopped */
3289 oldpool = conf->r1bio_pool;
3290 conf->r1bio_pool = newpool;
3292 for (d = d2 = 0; d < conf->raid_disks; d++) {
3293 struct md_rdev *rdev = conf->mirrors[d].rdev;
3294 if (rdev && rdev->raid_disk != d2) {
3295 sysfs_unlink_rdev(mddev, rdev);
3296 rdev->raid_disk = d2;
3297 sysfs_unlink_rdev(mddev, rdev);
3298 if (sysfs_link_rdev(mddev, rdev))
3299 pr_warn("md/raid1:%s: cannot register rd%d\n",
3300 mdname(mddev), rdev->raid_disk);
3303 newmirrors[d2++].rdev = rdev;
3305 kfree(conf->mirrors);
3306 conf->mirrors = newmirrors;
3307 kfree(conf->poolinfo);
3308 conf->poolinfo = newpoolinfo;
3310 spin_lock_irqsave(&conf->device_lock, flags);
3311 mddev->degraded += (raid_disks - conf->raid_disks);
3312 spin_unlock_irqrestore(&conf->device_lock, flags);
3313 conf->raid_disks = mddev->raid_disks = raid_disks;
3314 mddev->delta_disks = 0;
3316 unfreeze_array(conf);
3318 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3319 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3320 md_wakeup_thread(mddev->thread);
3322 mempool_destroy(oldpool);
3326 static void raid1_quiesce(struct mddev *mddev, int state)
3328 struct r1conf *conf = mddev->private;
3331 case 2: /* wake for suspend */
3332 wake_up(&conf->wait_barrier);
3335 freeze_array(conf, 0);
3338 unfreeze_array(conf);
3343 static void *raid1_takeover(struct mddev *mddev)
3345 /* raid1 can take over:
3346 * raid5 with 2 devices, any layout or chunk size
3348 if (mddev->level == 5 && mddev->raid_disks == 2) {
3349 struct r1conf *conf;
3350 mddev->new_level = 1;
3351 mddev->new_layout = 0;
3352 mddev->new_chunk_sectors = 0;
3353 conf = setup_conf(mddev);
3354 if (!IS_ERR(conf)) {
3355 /* Array must appear to be quiesced */
3356 conf->array_frozen = 1;
3357 mddev_clear_unsupported_flags(mddev,
3358 UNSUPPORTED_MDDEV_FLAGS);
3362 return ERR_PTR(-EINVAL);
3365 static struct md_personality raid1_personality =
3369 .owner = THIS_MODULE,
3370 .make_request = raid1_make_request,
3373 .status = raid1_status,
3374 .error_handler = raid1_error,
3375 .hot_add_disk = raid1_add_disk,
3376 .hot_remove_disk= raid1_remove_disk,
3377 .spare_active = raid1_spare_active,
3378 .sync_request = raid1_sync_request,
3379 .resize = raid1_resize,
3381 .check_reshape = raid1_reshape,
3382 .quiesce = raid1_quiesce,
3383 .takeover = raid1_takeover,
3384 .congested = raid1_congested,
3387 static int __init raid_init(void)
3389 return register_md_personality(&raid1_personality);
3392 static void raid_exit(void)
3394 unregister_md_personality(&raid1_personality);
3397 module_init(raid_init);
3398 module_exit(raid_exit);
3399 MODULE_LICENSE("GPL");
3400 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3401 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3402 MODULE_ALIAS("md-raid1");
3403 MODULE_ALIAS("md-level-1");
3405 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);