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_bio_list(struct r1conf *conf, struct bio *bio)
792 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
793 bitmap_unplug(conf->mddev->bitmap);
794 wake_up(&conf->wait_barrier);
796 while (bio) { /* submit pending writes */
797 struct bio *next = bio->bi_next;
798 struct md_rdev *rdev = (void*)bio->bi_bdev;
800 bio->bi_bdev = rdev->bdev;
801 if (test_bit(Faulty, &rdev->flags)) {
802 bio->bi_error = -EIO;
804 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
805 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
809 generic_make_request(bio);
814 static void flush_pending_writes(struct r1conf *conf)
816 /* Any writes that have been queued but are awaiting
817 * bitmap updates get flushed here.
819 spin_lock_irq(&conf->device_lock);
821 if (conf->pending_bio_list.head) {
823 bio = bio_list_get(&conf->pending_bio_list);
824 conf->pending_count = 0;
825 spin_unlock_irq(&conf->device_lock);
826 flush_bio_list(conf, bio);
828 spin_unlock_irq(&conf->device_lock);
832 * Sometimes we need to suspend IO while we do something else,
833 * either some resync/recovery, or reconfigure the array.
834 * To do this we raise a 'barrier'.
835 * The 'barrier' is a counter that can be raised multiple times
836 * to count how many activities are happening which preclude
838 * We can only raise the barrier if there is no pending IO.
839 * i.e. if nr_pending == 0.
840 * We choose only to raise the barrier if no-one is waiting for the
841 * barrier to go down. This means that as soon as an IO request
842 * is ready, no other operations which require a barrier will start
843 * until the IO request has had a chance.
845 * So: regular IO calls 'wait_barrier'. When that returns there
846 * is no backgroup IO happening, It must arrange to call
847 * allow_barrier when it has finished its IO.
848 * backgroup IO calls must call raise_barrier. Once that returns
849 * there is no normal IO happeing. It must arrange to call
850 * lower_barrier when the particular background IO completes.
852 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
854 int idx = sector_to_idx(sector_nr);
856 spin_lock_irq(&conf->resync_lock);
858 /* Wait until no block IO is waiting */
859 wait_event_lock_irq(conf->wait_barrier,
860 !atomic_read(&conf->nr_waiting[idx]),
863 /* block any new IO from starting */
864 atomic_inc(&conf->barrier[idx]);
866 * In raise_barrier() we firstly increase conf->barrier[idx] then
867 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
868 * increase conf->nr_pending[idx] then check conf->barrier[idx].
869 * A memory barrier here to make sure conf->nr_pending[idx] won't
870 * be fetched before conf->barrier[idx] is increased. Otherwise
871 * there will be a race between raise_barrier() and _wait_barrier().
873 smp_mb__after_atomic();
875 /* For these conditions we must wait:
876 * A: while the array is in frozen state
877 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
878 * existing in corresponding I/O barrier bucket.
879 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
880 * max resync count which allowed on current I/O barrier bucket.
882 wait_event_lock_irq(conf->wait_barrier,
883 !conf->array_frozen &&
884 !atomic_read(&conf->nr_pending[idx]) &&
885 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH,
888 atomic_inc(&conf->nr_sync_pending);
889 spin_unlock_irq(&conf->resync_lock);
892 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
894 int idx = sector_to_idx(sector_nr);
896 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
898 atomic_dec(&conf->barrier[idx]);
899 atomic_dec(&conf->nr_sync_pending);
900 wake_up(&conf->wait_barrier);
903 static void _wait_barrier(struct r1conf *conf, int idx)
906 * We need to increase conf->nr_pending[idx] very early here,
907 * then raise_barrier() can be blocked when it waits for
908 * conf->nr_pending[idx] to be 0. Then we can avoid holding
909 * conf->resync_lock when there is no barrier raised in same
910 * barrier unit bucket. Also if the array is frozen, I/O
911 * should be blocked until array is unfrozen.
913 atomic_inc(&conf->nr_pending[idx]);
915 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
916 * check conf->barrier[idx]. In raise_barrier() we firstly increase
917 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
918 * barrier is necessary here to make sure conf->barrier[idx] won't be
919 * fetched before conf->nr_pending[idx] is increased. Otherwise there
920 * will be a race between _wait_barrier() and raise_barrier().
922 smp_mb__after_atomic();
925 * Don't worry about checking two atomic_t variables at same time
926 * here. If during we check conf->barrier[idx], the array is
927 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
928 * 0, it is safe to return and make the I/O continue. Because the
929 * array is frozen, all I/O returned here will eventually complete
930 * or be queued, no race will happen. See code comment in
933 if (!READ_ONCE(conf->array_frozen) &&
934 !atomic_read(&conf->barrier[idx]))
938 * After holding conf->resync_lock, conf->nr_pending[idx]
939 * should be decreased before waiting for barrier to drop.
940 * Otherwise, we may encounter a race condition because
941 * raise_barrer() might be waiting for conf->nr_pending[idx]
942 * to be 0 at same time.
944 spin_lock_irq(&conf->resync_lock);
945 atomic_inc(&conf->nr_waiting[idx]);
946 atomic_dec(&conf->nr_pending[idx]);
948 * In case freeze_array() is waiting for
949 * get_unqueued_pending() == extra
951 wake_up(&conf->wait_barrier);
952 /* Wait for the barrier in same barrier unit bucket to drop. */
953 wait_event_lock_irq(conf->wait_barrier,
954 !conf->array_frozen &&
955 !atomic_read(&conf->barrier[idx]),
957 atomic_inc(&conf->nr_pending[idx]);
958 atomic_dec(&conf->nr_waiting[idx]);
959 spin_unlock_irq(&conf->resync_lock);
962 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
964 int idx = sector_to_idx(sector_nr);
967 * Very similar to _wait_barrier(). The difference is, for read
968 * I/O we don't need wait for sync I/O, but if the whole array
969 * is frozen, the read I/O still has to wait until the array is
970 * unfrozen. Since there is no ordering requirement with
971 * conf->barrier[idx] here, memory barrier is unnecessary as well.
973 atomic_inc(&conf->nr_pending[idx]);
975 if (!READ_ONCE(conf->array_frozen))
978 spin_lock_irq(&conf->resync_lock);
979 atomic_inc(&conf->nr_waiting[idx]);
980 atomic_dec(&conf->nr_pending[idx]);
982 * In case freeze_array() is waiting for
983 * get_unqueued_pending() == extra
985 wake_up(&conf->wait_barrier);
986 /* Wait for array to be unfrozen */
987 wait_event_lock_irq(conf->wait_barrier,
990 atomic_inc(&conf->nr_pending[idx]);
991 atomic_dec(&conf->nr_waiting[idx]);
992 spin_unlock_irq(&conf->resync_lock);
995 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
997 int idx = sector_to_idx(sector_nr);
999 _wait_barrier(conf, idx);
1002 static void wait_all_barriers(struct r1conf *conf)
1006 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1007 _wait_barrier(conf, idx);
1010 static void _allow_barrier(struct r1conf *conf, int idx)
1012 atomic_dec(&conf->nr_pending[idx]);
1013 wake_up(&conf->wait_barrier);
1016 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1018 int idx = sector_to_idx(sector_nr);
1020 _allow_barrier(conf, idx);
1023 static void allow_all_barriers(struct r1conf *conf)
1027 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1028 _allow_barrier(conf, idx);
1031 /* conf->resync_lock should be held */
1032 static int get_unqueued_pending(struct r1conf *conf)
1036 ret = atomic_read(&conf->nr_sync_pending);
1037 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1038 ret += atomic_read(&conf->nr_pending[idx]) -
1039 atomic_read(&conf->nr_queued[idx]);
1044 static void freeze_array(struct r1conf *conf, int extra)
1046 /* Stop sync I/O and normal I/O and wait for everything to
1048 * This is called in two situations:
1049 * 1) management command handlers (reshape, remove disk, quiesce).
1050 * 2) one normal I/O request failed.
1052 * After array_frozen is set to 1, new sync IO will be blocked at
1053 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1054 * or wait_read_barrier(). The flying I/Os will either complete or be
1055 * queued. When everything goes quite, there are only queued I/Os left.
1057 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1058 * barrier bucket index which this I/O request hits. When all sync and
1059 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1060 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1061 * in handle_read_error(), we may call freeze_array() before trying to
1062 * fix the read error. In this case, the error read I/O is not queued,
1063 * so get_unqueued_pending() == 1.
1065 * Therefore before this function returns, we need to wait until
1066 * get_unqueued_pendings(conf) gets equal to extra. For
1067 * normal I/O context, extra is 1, in rested situations extra is 0.
1069 spin_lock_irq(&conf->resync_lock);
1070 conf->array_frozen = 1;
1071 raid1_log(conf->mddev, "wait freeze");
1072 wait_event_lock_irq_cmd(
1074 get_unqueued_pending(conf) == extra,
1076 flush_pending_writes(conf));
1077 spin_unlock_irq(&conf->resync_lock);
1079 static void unfreeze_array(struct r1conf *conf)
1081 /* reverse the effect of the freeze */
1082 spin_lock_irq(&conf->resync_lock);
1083 conf->array_frozen = 0;
1084 spin_unlock_irq(&conf->resync_lock);
1085 wake_up(&conf->wait_barrier);
1088 static struct bio *alloc_behind_master_bio(struct r1bio *r1_bio,
1091 int size = bio->bi_iter.bi_size;
1092 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1094 struct bio *behind_bio = NULL;
1096 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1100 /* discard op, we don't support writezero/writesame yet */
1101 if (!bio_has_data(bio))
1104 while (i < vcnt && size) {
1106 int len = min_t(int, PAGE_SIZE, size);
1108 page = alloc_page(GFP_NOIO);
1109 if (unlikely(!page))
1112 bio_add_page(behind_bio, page, len, 0);
1118 bio_copy_data(behind_bio, bio);
1120 r1_bio->behind_master_bio = behind_bio;;
1121 set_bit(R1BIO_BehindIO, &r1_bio->state);
1126 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1127 bio->bi_iter.bi_size);
1128 bio_free_pages(behind_bio);
1133 struct raid1_plug_cb {
1134 struct blk_plug_cb cb;
1135 struct bio_list pending;
1139 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1141 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1143 struct mddev *mddev = plug->cb.data;
1144 struct r1conf *conf = mddev->private;
1147 if (from_schedule || current->bio_list) {
1148 spin_lock_irq(&conf->device_lock);
1149 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1150 conf->pending_count += plug->pending_cnt;
1151 spin_unlock_irq(&conf->device_lock);
1152 wake_up(&conf->wait_barrier);
1153 md_wakeup_thread(mddev->thread);
1158 /* we aren't scheduling, so we can do the write-out directly. */
1159 bio = bio_list_get(&plug->pending);
1160 flush_bio_list(conf, bio);
1164 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1166 r1_bio->master_bio = bio;
1167 r1_bio->sectors = bio_sectors(bio);
1169 r1_bio->mddev = mddev;
1170 r1_bio->sector = bio->bi_iter.bi_sector;
1173 static inline struct r1bio *
1174 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1176 struct r1conf *conf = mddev->private;
1177 struct r1bio *r1_bio;
1179 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1180 /* Ensure no bio records IO_BLOCKED */
1181 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1182 init_r1bio(r1_bio, mddev, bio);
1186 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1187 int max_read_sectors, struct r1bio *r1_bio)
1189 struct r1conf *conf = mddev->private;
1190 struct raid1_info *mirror;
1191 struct bio *read_bio;
1192 struct bitmap *bitmap = mddev->bitmap;
1193 const int op = bio_op(bio);
1194 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1197 bool print_msg = !!r1_bio;
1198 char b[BDEVNAME_SIZE];
1201 * If r1_bio is set, we are blocking the raid1d thread
1202 * so there is a tiny risk of deadlock. So ask for
1203 * emergency memory if needed.
1205 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1208 /* Need to get the block device name carefully */
1209 struct md_rdev *rdev;
1211 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1213 bdevname(rdev->bdev, b);
1220 * Still need barrier for READ in case that whole
1223 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1226 r1_bio = alloc_r1bio(mddev, bio);
1228 init_r1bio(r1_bio, mddev, bio);
1229 r1_bio->sectors = max_read_sectors;
1232 * make_request() can abort the operation when read-ahead is being
1233 * used and no empty request is available.
1235 rdisk = read_balance(conf, r1_bio, &max_sectors);
1238 /* couldn't find anywhere to read from */
1240 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1243 (unsigned long long)r1_bio->sector);
1245 raid_end_bio_io(r1_bio);
1248 mirror = conf->mirrors + rdisk;
1251 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1253 (unsigned long long)r1_bio->sector,
1254 bdevname(mirror->rdev->bdev, b));
1256 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1259 * Reading from a write-mostly device must take care not to
1260 * over-take any writes that are 'behind'
1262 raid1_log(mddev, "wait behind writes");
1263 wait_event(bitmap->behind_wait,
1264 atomic_read(&bitmap->behind_writes) == 0);
1267 if (max_sectors < bio_sectors(bio)) {
1268 struct bio *split = bio_split(bio, max_sectors,
1269 gfp, conf->bio_split);
1270 bio_chain(split, bio);
1271 generic_make_request(bio);
1273 r1_bio->master_bio = bio;
1274 r1_bio->sectors = max_sectors;
1277 r1_bio->read_disk = rdisk;
1279 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1281 r1_bio->bios[rdisk] = read_bio;
1283 read_bio->bi_iter.bi_sector = r1_bio->sector +
1284 mirror->rdev->data_offset;
1285 read_bio->bi_bdev = mirror->rdev->bdev;
1286 read_bio->bi_end_io = raid1_end_read_request;
1287 bio_set_op_attrs(read_bio, op, do_sync);
1288 if (test_bit(FailFast, &mirror->rdev->flags) &&
1289 test_bit(R1BIO_FailFast, &r1_bio->state))
1290 read_bio->bi_opf |= MD_FAILFAST;
1291 read_bio->bi_private = r1_bio;
1294 trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1295 read_bio, disk_devt(mddev->gendisk),
1298 generic_make_request(read_bio);
1301 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1302 int max_write_sectors)
1304 struct r1conf *conf = mddev->private;
1305 struct r1bio *r1_bio;
1307 struct bitmap *bitmap = mddev->bitmap;
1308 unsigned long flags;
1309 struct md_rdev *blocked_rdev;
1310 struct blk_plug_cb *cb;
1311 struct raid1_plug_cb *plug = NULL;
1316 * Register the new request and wait if the reconstruction
1317 * thread has put up a bar for new requests.
1318 * Continue immediately if no resync is active currently.
1321 md_write_start(mddev, bio); /* wait on superblock update early */
1323 if ((bio_end_sector(bio) > mddev->suspend_lo &&
1324 bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1325 (mddev_is_clustered(mddev) &&
1326 md_cluster_ops->area_resyncing(mddev, WRITE,
1327 bio->bi_iter.bi_sector, bio_end_sector(bio)))) {
1330 * As the suspend_* range is controlled by userspace, we want
1331 * an interruptible wait.
1335 flush_signals(current);
1336 prepare_to_wait(&conf->wait_barrier,
1337 &w, TASK_INTERRUPTIBLE);
1338 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1339 bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1340 (mddev_is_clustered(mddev) &&
1341 !md_cluster_ops->area_resyncing(mddev, WRITE,
1342 bio->bi_iter.bi_sector,
1343 bio_end_sector(bio))))
1347 finish_wait(&conf->wait_barrier, &w);
1349 wait_barrier(conf, bio->bi_iter.bi_sector);
1351 r1_bio = alloc_r1bio(mddev, bio);
1352 r1_bio->sectors = max_write_sectors;
1354 if (conf->pending_count >= max_queued_requests) {
1355 md_wakeup_thread(mddev->thread);
1356 raid1_log(mddev, "wait queued");
1357 wait_event(conf->wait_barrier,
1358 conf->pending_count < max_queued_requests);
1360 /* first select target devices under rcu_lock and
1361 * inc refcount on their rdev. Record them by setting
1363 * If there are known/acknowledged bad blocks on any device on
1364 * which we have seen a write error, we want to avoid writing those
1366 * This potentially requires several writes to write around
1367 * the bad blocks. Each set of writes gets it's own r1bio
1368 * with a set of bios attached.
1371 disks = conf->raid_disks * 2;
1373 blocked_rdev = NULL;
1375 max_sectors = r1_bio->sectors;
1376 for (i = 0; i < disks; i++) {
1377 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1378 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1379 atomic_inc(&rdev->nr_pending);
1380 blocked_rdev = rdev;
1383 r1_bio->bios[i] = NULL;
1384 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1385 if (i < conf->raid_disks)
1386 set_bit(R1BIO_Degraded, &r1_bio->state);
1390 atomic_inc(&rdev->nr_pending);
1391 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1396 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1397 &first_bad, &bad_sectors);
1399 /* mustn't write here until the bad block is
1401 set_bit(BlockedBadBlocks, &rdev->flags);
1402 blocked_rdev = rdev;
1405 if (is_bad && first_bad <= r1_bio->sector) {
1406 /* Cannot write here at all */
1407 bad_sectors -= (r1_bio->sector - first_bad);
1408 if (bad_sectors < max_sectors)
1409 /* mustn't write more than bad_sectors
1410 * to other devices yet
1412 max_sectors = bad_sectors;
1413 rdev_dec_pending(rdev, mddev);
1414 /* We don't set R1BIO_Degraded as that
1415 * only applies if the disk is
1416 * missing, so it might be re-added,
1417 * and we want to know to recover this
1419 * In this case the device is here,
1420 * and the fact that this chunk is not
1421 * in-sync is recorded in the bad
1427 int good_sectors = first_bad - r1_bio->sector;
1428 if (good_sectors < max_sectors)
1429 max_sectors = good_sectors;
1432 r1_bio->bios[i] = bio;
1436 if (unlikely(blocked_rdev)) {
1437 /* Wait for this device to become unblocked */
1440 for (j = 0; j < i; j++)
1441 if (r1_bio->bios[j])
1442 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1444 allow_barrier(conf, bio->bi_iter.bi_sector);
1445 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1446 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1447 wait_barrier(conf, bio->bi_iter.bi_sector);
1451 if (max_sectors < bio_sectors(bio)) {
1452 struct bio *split = bio_split(bio, max_sectors,
1453 GFP_NOIO, conf->bio_split);
1454 bio_chain(split, bio);
1455 generic_make_request(bio);
1457 r1_bio->master_bio = bio;
1458 r1_bio->sectors = max_sectors;
1461 atomic_set(&r1_bio->remaining, 1);
1462 atomic_set(&r1_bio->behind_remaining, 0);
1466 for (i = 0; i < disks; i++) {
1467 struct bio *mbio = NULL;
1468 if (!r1_bio->bios[i])
1474 * Not if there are too many, or cannot
1475 * allocate memory, or a reader on WriteMostly
1476 * is waiting for behind writes to flush */
1478 (atomic_read(&bitmap->behind_writes)
1479 < mddev->bitmap_info.max_write_behind) &&
1480 !waitqueue_active(&bitmap->behind_wait)) {
1481 mbio = alloc_behind_master_bio(r1_bio, bio);
1484 bitmap_startwrite(bitmap, r1_bio->sector,
1486 test_bit(R1BIO_BehindIO,
1492 if (r1_bio->behind_master_bio)
1493 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1497 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1500 if (r1_bio->behind_master_bio) {
1501 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1502 atomic_inc(&r1_bio->behind_remaining);
1505 r1_bio->bios[i] = mbio;
1507 mbio->bi_iter.bi_sector = (r1_bio->sector +
1508 conf->mirrors[i].rdev->data_offset);
1509 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1510 mbio->bi_end_io = raid1_end_write_request;
1511 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1512 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1513 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1514 conf->raid_disks - mddev->degraded > 1)
1515 mbio->bi_opf |= MD_FAILFAST;
1516 mbio->bi_private = r1_bio;
1518 atomic_inc(&r1_bio->remaining);
1521 trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1522 mbio, disk_devt(mddev->gendisk),
1524 /* flush_pending_writes() needs access to the rdev so...*/
1525 mbio->bi_bdev = (void*)conf->mirrors[i].rdev;
1527 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1529 plug = container_of(cb, struct raid1_plug_cb, cb);
1533 bio_list_add(&plug->pending, mbio);
1534 plug->pending_cnt++;
1536 spin_lock_irqsave(&conf->device_lock, flags);
1537 bio_list_add(&conf->pending_bio_list, mbio);
1538 conf->pending_count++;
1539 spin_unlock_irqrestore(&conf->device_lock, flags);
1540 md_wakeup_thread(mddev->thread);
1544 r1_bio_write_done(r1_bio);
1546 /* In case raid1d snuck in to freeze_array */
1547 wake_up(&conf->wait_barrier);
1550 static void raid1_make_request(struct mddev *mddev, struct bio *bio)
1554 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1555 md_flush_request(mddev, bio);
1560 * There is a limit to the maximum size, but
1561 * the read/write handler might find a lower limit
1562 * due to bad blocks. To avoid multiple splits,
1563 * we pass the maximum number of sectors down
1564 * and let the lower level perform the split.
1566 sectors = align_to_barrier_unit_end(
1567 bio->bi_iter.bi_sector, bio_sectors(bio));
1569 if (bio_data_dir(bio) == READ)
1570 raid1_read_request(mddev, bio, sectors, NULL);
1572 raid1_write_request(mddev, bio, sectors);
1575 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1577 struct r1conf *conf = mddev->private;
1580 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1581 conf->raid_disks - mddev->degraded);
1583 for (i = 0; i < conf->raid_disks; i++) {
1584 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1585 seq_printf(seq, "%s",
1586 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1589 seq_printf(seq, "]");
1592 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1594 char b[BDEVNAME_SIZE];
1595 struct r1conf *conf = mddev->private;
1596 unsigned long flags;
1599 * If it is not operational, then we have already marked it as dead
1600 * else if it is the last working disks, ignore the error, let the
1601 * next level up know.
1602 * else mark the drive as failed
1604 spin_lock_irqsave(&conf->device_lock, flags);
1605 if (test_bit(In_sync, &rdev->flags)
1606 && (conf->raid_disks - mddev->degraded) == 1) {
1608 * Don't fail the drive, act as though we were just a
1609 * normal single drive.
1610 * However don't try a recovery from this drive as
1611 * it is very likely to fail.
1613 conf->recovery_disabled = mddev->recovery_disabled;
1614 spin_unlock_irqrestore(&conf->device_lock, flags);
1617 set_bit(Blocked, &rdev->flags);
1618 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1620 set_bit(Faulty, &rdev->flags);
1622 set_bit(Faulty, &rdev->flags);
1623 spin_unlock_irqrestore(&conf->device_lock, flags);
1625 * if recovery is running, make sure it aborts.
1627 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1628 set_mask_bits(&mddev->sb_flags, 0,
1629 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1630 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1631 "md/raid1:%s: Operation continuing on %d devices.\n",
1632 mdname(mddev), bdevname(rdev->bdev, b),
1633 mdname(mddev), conf->raid_disks - mddev->degraded);
1636 static void print_conf(struct r1conf *conf)
1640 pr_debug("RAID1 conf printout:\n");
1642 pr_debug("(!conf)\n");
1645 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1649 for (i = 0; i < conf->raid_disks; i++) {
1650 char b[BDEVNAME_SIZE];
1651 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1653 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1654 i, !test_bit(In_sync, &rdev->flags),
1655 !test_bit(Faulty, &rdev->flags),
1656 bdevname(rdev->bdev,b));
1661 static void close_sync(struct r1conf *conf)
1663 wait_all_barriers(conf);
1664 allow_all_barriers(conf);
1666 mempool_destroy(conf->r1buf_pool);
1667 conf->r1buf_pool = NULL;
1670 static int raid1_spare_active(struct mddev *mddev)
1673 struct r1conf *conf = mddev->private;
1675 unsigned long flags;
1678 * Find all failed disks within the RAID1 configuration
1679 * and mark them readable.
1680 * Called under mddev lock, so rcu protection not needed.
1681 * device_lock used to avoid races with raid1_end_read_request
1682 * which expects 'In_sync' flags and ->degraded to be consistent.
1684 spin_lock_irqsave(&conf->device_lock, flags);
1685 for (i = 0; i < conf->raid_disks; i++) {
1686 struct md_rdev *rdev = conf->mirrors[i].rdev;
1687 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1689 && !test_bit(Candidate, &repl->flags)
1690 && repl->recovery_offset == MaxSector
1691 && !test_bit(Faulty, &repl->flags)
1692 && !test_and_set_bit(In_sync, &repl->flags)) {
1693 /* replacement has just become active */
1695 !test_and_clear_bit(In_sync, &rdev->flags))
1698 /* Replaced device not technically
1699 * faulty, but we need to be sure
1700 * it gets removed and never re-added
1702 set_bit(Faulty, &rdev->flags);
1703 sysfs_notify_dirent_safe(
1708 && rdev->recovery_offset == MaxSector
1709 && !test_bit(Faulty, &rdev->flags)
1710 && !test_and_set_bit(In_sync, &rdev->flags)) {
1712 sysfs_notify_dirent_safe(rdev->sysfs_state);
1715 mddev->degraded -= count;
1716 spin_unlock_irqrestore(&conf->device_lock, flags);
1722 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1724 struct r1conf *conf = mddev->private;
1727 struct raid1_info *p;
1729 int last = conf->raid_disks - 1;
1731 if (mddev->recovery_disabled == conf->recovery_disabled)
1734 if (md_integrity_add_rdev(rdev, mddev))
1737 if (rdev->raid_disk >= 0)
1738 first = last = rdev->raid_disk;
1741 * find the disk ... but prefer rdev->saved_raid_disk
1744 if (rdev->saved_raid_disk >= 0 &&
1745 rdev->saved_raid_disk >= first &&
1746 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1747 first = last = rdev->saved_raid_disk;
1749 for (mirror = first; mirror <= last; mirror++) {
1750 p = conf->mirrors+mirror;
1754 disk_stack_limits(mddev->gendisk, rdev->bdev,
1755 rdev->data_offset << 9);
1757 p->head_position = 0;
1758 rdev->raid_disk = mirror;
1760 /* As all devices are equivalent, we don't need a full recovery
1761 * if this was recently any drive of the array
1763 if (rdev->saved_raid_disk < 0)
1765 rcu_assign_pointer(p->rdev, rdev);
1768 if (test_bit(WantReplacement, &p->rdev->flags) &&
1769 p[conf->raid_disks].rdev == NULL) {
1770 /* Add this device as a replacement */
1771 clear_bit(In_sync, &rdev->flags);
1772 set_bit(Replacement, &rdev->flags);
1773 rdev->raid_disk = mirror;
1776 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1780 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1781 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1786 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1788 struct r1conf *conf = mddev->private;
1790 int number = rdev->raid_disk;
1791 struct raid1_info *p = conf->mirrors + number;
1793 if (rdev != p->rdev)
1794 p = conf->mirrors + conf->raid_disks + number;
1797 if (rdev == p->rdev) {
1798 if (test_bit(In_sync, &rdev->flags) ||
1799 atomic_read(&rdev->nr_pending)) {
1803 /* Only remove non-faulty devices if recovery
1806 if (!test_bit(Faulty, &rdev->flags) &&
1807 mddev->recovery_disabled != conf->recovery_disabled &&
1808 mddev->degraded < conf->raid_disks) {
1813 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1815 if (atomic_read(&rdev->nr_pending)) {
1816 /* lost the race, try later */
1822 if (conf->mirrors[conf->raid_disks + number].rdev) {
1823 /* We just removed a device that is being replaced.
1824 * Move down the replacement. We drain all IO before
1825 * doing this to avoid confusion.
1827 struct md_rdev *repl =
1828 conf->mirrors[conf->raid_disks + number].rdev;
1829 freeze_array(conf, 0);
1830 clear_bit(Replacement, &repl->flags);
1832 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1833 unfreeze_array(conf);
1836 clear_bit(WantReplacement, &rdev->flags);
1837 err = md_integrity_register(mddev);
1845 static void end_sync_read(struct bio *bio)
1847 struct r1bio *r1_bio = get_resync_r1bio(bio);
1849 update_head_pos(r1_bio->read_disk, r1_bio);
1852 * we have read a block, now it needs to be re-written,
1853 * or re-read if the read failed.
1854 * We don't do much here, just schedule handling by raid1d
1857 set_bit(R1BIO_Uptodate, &r1_bio->state);
1859 if (atomic_dec_and_test(&r1_bio->remaining))
1860 reschedule_retry(r1_bio);
1863 static void end_sync_write(struct bio *bio)
1865 int uptodate = !bio->bi_error;
1866 struct r1bio *r1_bio = get_resync_r1bio(bio);
1867 struct mddev *mddev = r1_bio->mddev;
1868 struct r1conf *conf = mddev->private;
1871 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1874 sector_t sync_blocks = 0;
1875 sector_t s = r1_bio->sector;
1876 long sectors_to_go = r1_bio->sectors;
1877 /* make sure these bits doesn't get cleared. */
1879 bitmap_end_sync(mddev->bitmap, s,
1882 sectors_to_go -= sync_blocks;
1883 } while (sectors_to_go > 0);
1884 set_bit(WriteErrorSeen, &rdev->flags);
1885 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1886 set_bit(MD_RECOVERY_NEEDED, &
1888 set_bit(R1BIO_WriteError, &r1_bio->state);
1889 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1890 &first_bad, &bad_sectors) &&
1891 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1894 &first_bad, &bad_sectors)
1896 set_bit(R1BIO_MadeGood, &r1_bio->state);
1898 if (atomic_dec_and_test(&r1_bio->remaining)) {
1899 int s = r1_bio->sectors;
1900 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1901 test_bit(R1BIO_WriteError, &r1_bio->state))
1902 reschedule_retry(r1_bio);
1905 md_done_sync(mddev, s, uptodate);
1910 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1911 int sectors, struct page *page, int rw)
1913 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1917 set_bit(WriteErrorSeen, &rdev->flags);
1918 if (!test_and_set_bit(WantReplacement,
1920 set_bit(MD_RECOVERY_NEEDED, &
1921 rdev->mddev->recovery);
1923 /* need to record an error - either for the block or the device */
1924 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1925 md_error(rdev->mddev, rdev);
1929 static int fix_sync_read_error(struct r1bio *r1_bio)
1931 /* Try some synchronous reads of other devices to get
1932 * good data, much like with normal read errors. Only
1933 * read into the pages we already have so we don't
1934 * need to re-issue the read request.
1935 * We don't need to freeze the array, because being in an
1936 * active sync request, there is no normal IO, and
1937 * no overlapping syncs.
1938 * We don't need to check is_badblock() again as we
1939 * made sure that anything with a bad block in range
1940 * will have bi_end_io clear.
1942 struct mddev *mddev = r1_bio->mddev;
1943 struct r1conf *conf = mddev->private;
1944 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1945 struct page **pages = get_resync_pages(bio)->pages;
1946 sector_t sect = r1_bio->sector;
1947 int sectors = r1_bio->sectors;
1949 struct md_rdev *rdev;
1951 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1952 if (test_bit(FailFast, &rdev->flags)) {
1953 /* Don't try recovering from here - just fail it
1954 * ... unless it is the last working device of course */
1955 md_error(mddev, rdev);
1956 if (test_bit(Faulty, &rdev->flags))
1957 /* Don't try to read from here, but make sure
1958 * put_buf does it's thing
1960 bio->bi_end_io = end_sync_write;
1965 int d = r1_bio->read_disk;
1969 if (s > (PAGE_SIZE>>9))
1972 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1973 /* No rcu protection needed here devices
1974 * can only be removed when no resync is
1975 * active, and resync is currently active
1977 rdev = conf->mirrors[d].rdev;
1978 if (sync_page_io(rdev, sect, s<<9,
1980 REQ_OP_READ, 0, false)) {
1986 if (d == conf->raid_disks * 2)
1988 } while (!success && d != r1_bio->read_disk);
1991 char b[BDEVNAME_SIZE];
1993 /* Cannot read from anywhere, this block is lost.
1994 * Record a bad block on each device. If that doesn't
1995 * work just disable and interrupt the recovery.
1996 * Don't fail devices as that won't really help.
1998 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2000 bdevname(bio->bi_bdev, b),
2001 (unsigned long long)r1_bio->sector);
2002 for (d = 0; d < conf->raid_disks * 2; d++) {
2003 rdev = conf->mirrors[d].rdev;
2004 if (!rdev || test_bit(Faulty, &rdev->flags))
2006 if (!rdev_set_badblocks(rdev, sect, s, 0))
2010 conf->recovery_disabled =
2011 mddev->recovery_disabled;
2012 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2013 md_done_sync(mddev, r1_bio->sectors, 0);
2025 /* write it back and re-read */
2026 while (d != r1_bio->read_disk) {
2028 d = conf->raid_disks * 2;
2030 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2032 rdev = conf->mirrors[d].rdev;
2033 if (r1_sync_page_io(rdev, sect, s,
2036 r1_bio->bios[d]->bi_end_io = NULL;
2037 rdev_dec_pending(rdev, mddev);
2041 while (d != r1_bio->read_disk) {
2043 d = conf->raid_disks * 2;
2045 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2047 rdev = conf->mirrors[d].rdev;
2048 if (r1_sync_page_io(rdev, sect, s,
2051 atomic_add(s, &rdev->corrected_errors);
2057 set_bit(R1BIO_Uptodate, &r1_bio->state);
2062 static void process_checks(struct r1bio *r1_bio)
2064 /* We have read all readable devices. If we haven't
2065 * got the block, then there is no hope left.
2066 * If we have, then we want to do a comparison
2067 * and skip the write if everything is the same.
2068 * If any blocks failed to read, then we need to
2069 * attempt an over-write
2071 struct mddev *mddev = r1_bio->mddev;
2072 struct r1conf *conf = mddev->private;
2077 /* Fix variable parts of all bios */
2078 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2079 for (i = 0; i < conf->raid_disks * 2; i++) {
2084 struct bio *b = r1_bio->bios[i];
2085 struct resync_pages *rp = get_resync_pages(b);
2086 if (b->bi_end_io != end_sync_read)
2088 /* fixup the bio for reuse, but preserve errno */
2089 error = b->bi_error;
2091 b->bi_error = error;
2093 b->bi_iter.bi_size = r1_bio->sectors << 9;
2094 b->bi_iter.bi_sector = r1_bio->sector +
2095 conf->mirrors[i].rdev->data_offset;
2096 b->bi_bdev = conf->mirrors[i].rdev->bdev;
2097 b->bi_end_io = end_sync_read;
2098 rp->raid_bio = r1_bio;
2101 size = b->bi_iter.bi_size;
2102 bio_for_each_segment_all(bi, b, j) {
2104 if (size > PAGE_SIZE)
2105 bi->bv_len = PAGE_SIZE;
2111 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2112 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2113 !r1_bio->bios[primary]->bi_error) {
2114 r1_bio->bios[primary]->bi_end_io = NULL;
2115 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2118 r1_bio->read_disk = primary;
2119 for (i = 0; i < conf->raid_disks * 2; i++) {
2121 struct bio *pbio = r1_bio->bios[primary];
2122 struct bio *sbio = r1_bio->bios[i];
2123 int error = sbio->bi_error;
2124 struct page **ppages = get_resync_pages(pbio)->pages;
2125 struct page **spages = get_resync_pages(sbio)->pages;
2127 int page_len[RESYNC_PAGES] = { 0 };
2129 if (sbio->bi_end_io != end_sync_read)
2131 /* Now we can 'fixup' the error value */
2134 bio_for_each_segment_all(bi, sbio, j)
2135 page_len[j] = bi->bv_len;
2138 for (j = vcnt; j-- ; ) {
2139 if (memcmp(page_address(ppages[j]),
2140 page_address(spages[j]),
2147 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2148 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2150 /* No need to write to this device. */
2151 sbio->bi_end_io = NULL;
2152 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2156 bio_copy_data(sbio, pbio);
2160 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2162 struct r1conf *conf = mddev->private;
2164 int disks = conf->raid_disks * 2;
2165 struct bio *bio, *wbio;
2167 bio = r1_bio->bios[r1_bio->read_disk];
2169 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2170 /* ouch - failed to read all of that. */
2171 if (!fix_sync_read_error(r1_bio))
2174 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2175 process_checks(r1_bio);
2180 atomic_set(&r1_bio->remaining, 1);
2181 for (i = 0; i < disks ; i++) {
2182 wbio = r1_bio->bios[i];
2183 if (wbio->bi_end_io == NULL ||
2184 (wbio->bi_end_io == end_sync_read &&
2185 (i == r1_bio->read_disk ||
2186 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2188 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2191 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2192 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2193 wbio->bi_opf |= MD_FAILFAST;
2195 wbio->bi_end_io = end_sync_write;
2196 atomic_inc(&r1_bio->remaining);
2197 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2199 generic_make_request(wbio);
2202 if (atomic_dec_and_test(&r1_bio->remaining)) {
2203 /* if we're here, all write(s) have completed, so clean up */
2204 int s = r1_bio->sectors;
2205 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2206 test_bit(R1BIO_WriteError, &r1_bio->state))
2207 reschedule_retry(r1_bio);
2210 md_done_sync(mddev, s, 1);
2216 * This is a kernel thread which:
2218 * 1. Retries failed read operations on working mirrors.
2219 * 2. Updates the raid superblock when problems encounter.
2220 * 3. Performs writes following reads for array synchronising.
2223 static void fix_read_error(struct r1conf *conf, int read_disk,
2224 sector_t sect, int sectors)
2226 struct mddev *mddev = conf->mddev;
2232 struct md_rdev *rdev;
2234 if (s > (PAGE_SIZE>>9))
2242 rdev = rcu_dereference(conf->mirrors[d].rdev);
2244 (test_bit(In_sync, &rdev->flags) ||
2245 (!test_bit(Faulty, &rdev->flags) &&
2246 rdev->recovery_offset >= sect + s)) &&
2247 is_badblock(rdev, sect, s,
2248 &first_bad, &bad_sectors) == 0) {
2249 atomic_inc(&rdev->nr_pending);
2251 if (sync_page_io(rdev, sect, s<<9,
2252 conf->tmppage, REQ_OP_READ, 0, false))
2254 rdev_dec_pending(rdev, mddev);
2260 if (d == conf->raid_disks * 2)
2262 } while (!success && d != read_disk);
2265 /* Cannot read from anywhere - mark it bad */
2266 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2267 if (!rdev_set_badblocks(rdev, sect, s, 0))
2268 md_error(mddev, rdev);
2271 /* write it back and re-read */
2273 while (d != read_disk) {
2275 d = conf->raid_disks * 2;
2278 rdev = rcu_dereference(conf->mirrors[d].rdev);
2280 !test_bit(Faulty, &rdev->flags)) {
2281 atomic_inc(&rdev->nr_pending);
2283 r1_sync_page_io(rdev, sect, s,
2284 conf->tmppage, WRITE);
2285 rdev_dec_pending(rdev, mddev);
2290 while (d != read_disk) {
2291 char b[BDEVNAME_SIZE];
2293 d = conf->raid_disks * 2;
2296 rdev = rcu_dereference(conf->mirrors[d].rdev);
2298 !test_bit(Faulty, &rdev->flags)) {
2299 atomic_inc(&rdev->nr_pending);
2301 if (r1_sync_page_io(rdev, sect, s,
2302 conf->tmppage, READ)) {
2303 atomic_add(s, &rdev->corrected_errors);
2304 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2306 (unsigned long long)(sect +
2308 bdevname(rdev->bdev, b));
2310 rdev_dec_pending(rdev, mddev);
2319 static int narrow_write_error(struct r1bio *r1_bio, int i)
2321 struct mddev *mddev = r1_bio->mddev;
2322 struct r1conf *conf = mddev->private;
2323 struct md_rdev *rdev = conf->mirrors[i].rdev;
2325 /* bio has the data to be written to device 'i' where
2326 * we just recently had a write error.
2327 * We repeatedly clone the bio and trim down to one block,
2328 * then try the write. Where the write fails we record
2330 * It is conceivable that the bio doesn't exactly align with
2331 * blocks. We must handle this somehow.
2333 * We currently own a reference on the rdev.
2339 int sect_to_write = r1_bio->sectors;
2342 if (rdev->badblocks.shift < 0)
2345 block_sectors = roundup(1 << rdev->badblocks.shift,
2346 bdev_logical_block_size(rdev->bdev) >> 9);
2347 sector = r1_bio->sector;
2348 sectors = ((sector + block_sectors)
2349 & ~(sector_t)(block_sectors - 1))
2352 while (sect_to_write) {
2354 if (sectors > sect_to_write)
2355 sectors = sect_to_write;
2356 /* Write at 'sector' for 'sectors'*/
2358 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2359 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2362 /* We really need a _all clone */
2363 wbio->bi_iter = (struct bvec_iter){ 0 };
2365 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2369 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2370 wbio->bi_iter.bi_sector = r1_bio->sector;
2371 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2373 bio_trim(wbio, sector - r1_bio->sector, sectors);
2374 wbio->bi_iter.bi_sector += rdev->data_offset;
2375 wbio->bi_bdev = rdev->bdev;
2377 if (submit_bio_wait(wbio) < 0)
2379 ok = rdev_set_badblocks(rdev, sector,
2384 sect_to_write -= sectors;
2386 sectors = block_sectors;
2391 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2394 int s = r1_bio->sectors;
2395 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2396 struct md_rdev *rdev = conf->mirrors[m].rdev;
2397 struct bio *bio = r1_bio->bios[m];
2398 if (bio->bi_end_io == NULL)
2400 if (!bio->bi_error &&
2401 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2402 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2404 if (bio->bi_error &&
2405 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2406 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2407 md_error(conf->mddev, rdev);
2411 md_done_sync(conf->mddev, s, 1);
2414 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2419 for (m = 0; m < conf->raid_disks * 2 ; m++)
2420 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2421 struct md_rdev *rdev = conf->mirrors[m].rdev;
2422 rdev_clear_badblocks(rdev,
2424 r1_bio->sectors, 0);
2425 rdev_dec_pending(rdev, conf->mddev);
2426 } else if (r1_bio->bios[m] != NULL) {
2427 /* This drive got a write error. We need to
2428 * narrow down and record precise write
2432 if (!narrow_write_error(r1_bio, m)) {
2433 md_error(conf->mddev,
2434 conf->mirrors[m].rdev);
2435 /* an I/O failed, we can't clear the bitmap */
2436 set_bit(R1BIO_Degraded, &r1_bio->state);
2438 rdev_dec_pending(conf->mirrors[m].rdev,
2442 spin_lock_irq(&conf->device_lock);
2443 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2444 idx = sector_to_idx(r1_bio->sector);
2445 atomic_inc(&conf->nr_queued[idx]);
2446 spin_unlock_irq(&conf->device_lock);
2448 * In case freeze_array() is waiting for condition
2449 * get_unqueued_pending() == extra to be true.
2451 wake_up(&conf->wait_barrier);
2452 md_wakeup_thread(conf->mddev->thread);
2454 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2455 close_write(r1_bio);
2456 raid_end_bio_io(r1_bio);
2460 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2462 struct mddev *mddev = conf->mddev;
2464 struct md_rdev *rdev;
2466 sector_t bio_sector;
2468 clear_bit(R1BIO_ReadError, &r1_bio->state);
2469 /* we got a read error. Maybe the drive is bad. Maybe just
2470 * the block and we can fix it.
2471 * We freeze all other IO, and try reading the block from
2472 * other devices. When we find one, we re-write
2473 * and check it that fixes the read error.
2474 * This is all done synchronously while the array is
2478 bio = r1_bio->bios[r1_bio->read_disk];
2479 bio_dev = bio->bi_bdev->bd_dev;
2480 bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2482 r1_bio->bios[r1_bio->read_disk] = NULL;
2484 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2486 && !test_bit(FailFast, &rdev->flags)) {
2487 freeze_array(conf, 1);
2488 fix_read_error(conf, r1_bio->read_disk,
2489 r1_bio->sector, r1_bio->sectors);
2490 unfreeze_array(conf);
2492 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2495 rdev_dec_pending(rdev, conf->mddev);
2496 allow_barrier(conf, r1_bio->sector);
2497 bio = r1_bio->master_bio;
2499 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2501 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2504 static void raid1d(struct md_thread *thread)
2506 struct mddev *mddev = thread->mddev;
2507 struct r1bio *r1_bio;
2508 unsigned long flags;
2509 struct r1conf *conf = mddev->private;
2510 struct list_head *head = &conf->retry_list;
2511 struct blk_plug plug;
2514 md_check_recovery(mddev);
2516 if (!list_empty_careful(&conf->bio_end_io_list) &&
2517 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2519 spin_lock_irqsave(&conf->device_lock, flags);
2520 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2521 list_splice_init(&conf->bio_end_io_list, &tmp);
2522 spin_unlock_irqrestore(&conf->device_lock, flags);
2523 while (!list_empty(&tmp)) {
2524 r1_bio = list_first_entry(&tmp, struct r1bio,
2526 list_del(&r1_bio->retry_list);
2527 idx = sector_to_idx(r1_bio->sector);
2528 atomic_dec(&conf->nr_queued[idx]);
2529 if (mddev->degraded)
2530 set_bit(R1BIO_Degraded, &r1_bio->state);
2531 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2532 close_write(r1_bio);
2533 raid_end_bio_io(r1_bio);
2537 blk_start_plug(&plug);
2540 flush_pending_writes(conf);
2542 spin_lock_irqsave(&conf->device_lock, flags);
2543 if (list_empty(head)) {
2544 spin_unlock_irqrestore(&conf->device_lock, flags);
2547 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2548 list_del(head->prev);
2549 idx = sector_to_idx(r1_bio->sector);
2550 atomic_dec(&conf->nr_queued[idx]);
2551 spin_unlock_irqrestore(&conf->device_lock, flags);
2553 mddev = r1_bio->mddev;
2554 conf = mddev->private;
2555 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2556 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2557 test_bit(R1BIO_WriteError, &r1_bio->state))
2558 handle_sync_write_finished(conf, r1_bio);
2560 sync_request_write(mddev, r1_bio);
2561 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2562 test_bit(R1BIO_WriteError, &r1_bio->state))
2563 handle_write_finished(conf, r1_bio);
2564 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2565 handle_read_error(conf, r1_bio);
2570 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2571 md_check_recovery(mddev);
2573 blk_finish_plug(&plug);
2576 static int init_resync(struct r1conf *conf)
2580 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2581 BUG_ON(conf->r1buf_pool);
2582 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2584 if (!conf->r1buf_pool)
2590 * perform a "sync" on one "block"
2592 * We need to make sure that no normal I/O request - particularly write
2593 * requests - conflict with active sync requests.
2595 * This is achieved by tracking pending requests and a 'barrier' concept
2596 * that can be installed to exclude normal IO requests.
2599 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2602 struct r1conf *conf = mddev->private;
2603 struct r1bio *r1_bio;
2605 sector_t max_sector, nr_sectors;
2609 int write_targets = 0, read_targets = 0;
2610 sector_t sync_blocks;
2611 int still_degraded = 0;
2612 int good_sectors = RESYNC_SECTORS;
2613 int min_bad = 0; /* number of sectors that are bad in all devices */
2614 int idx = sector_to_idx(sector_nr);
2616 if (!conf->r1buf_pool)
2617 if (init_resync(conf))
2620 max_sector = mddev->dev_sectors;
2621 if (sector_nr >= max_sector) {
2622 /* If we aborted, we need to abort the
2623 * sync on the 'current' bitmap chunk (there will
2624 * only be one in raid1 resync.
2625 * We can find the current addess in mddev->curr_resync
2627 if (mddev->curr_resync < max_sector) /* aborted */
2628 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2630 else /* completed sync */
2633 bitmap_close_sync(mddev->bitmap);
2636 if (mddev_is_clustered(mddev)) {
2637 conf->cluster_sync_low = 0;
2638 conf->cluster_sync_high = 0;
2643 if (mddev->bitmap == NULL &&
2644 mddev->recovery_cp == MaxSector &&
2645 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2646 conf->fullsync == 0) {
2648 return max_sector - sector_nr;
2650 /* before building a request, check if we can skip these blocks..
2651 * This call the bitmap_start_sync doesn't actually record anything
2653 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2654 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2655 /* We can skip this block, and probably several more */
2661 * If there is non-resync activity waiting for a turn, then let it
2662 * though before starting on this new sync request.
2664 if (atomic_read(&conf->nr_waiting[idx]))
2665 schedule_timeout_uninterruptible(1);
2667 /* we are incrementing sector_nr below. To be safe, we check against
2668 * sector_nr + two times RESYNC_SECTORS
2671 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2672 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2673 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2675 raise_barrier(conf, sector_nr);
2679 * If we get a correctably read error during resync or recovery,
2680 * we might want to read from a different device. So we
2681 * flag all drives that could conceivably be read from for READ,
2682 * and any others (which will be non-In_sync devices) for WRITE.
2683 * If a read fails, we try reading from something else for which READ
2687 r1_bio->mddev = mddev;
2688 r1_bio->sector = sector_nr;
2690 set_bit(R1BIO_IsSync, &r1_bio->state);
2691 /* make sure good_sectors won't go across barrier unit boundary */
2692 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2694 for (i = 0; i < conf->raid_disks * 2; i++) {
2695 struct md_rdev *rdev;
2696 bio = r1_bio->bios[i];
2698 rdev = rcu_dereference(conf->mirrors[i].rdev);
2700 test_bit(Faulty, &rdev->flags)) {
2701 if (i < conf->raid_disks)
2703 } else if (!test_bit(In_sync, &rdev->flags)) {
2704 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2705 bio->bi_end_io = end_sync_write;
2708 /* may need to read from here */
2709 sector_t first_bad = MaxSector;
2712 if (is_badblock(rdev, sector_nr, good_sectors,
2713 &first_bad, &bad_sectors)) {
2714 if (first_bad > sector_nr)
2715 good_sectors = first_bad - sector_nr;
2717 bad_sectors -= (sector_nr - first_bad);
2719 min_bad > bad_sectors)
2720 min_bad = bad_sectors;
2723 if (sector_nr < first_bad) {
2724 if (test_bit(WriteMostly, &rdev->flags)) {
2731 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2732 bio->bi_end_io = end_sync_read;
2734 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2735 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2736 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2738 * The device is suitable for reading (InSync),
2739 * but has bad block(s) here. Let's try to correct them,
2740 * if we are doing resync or repair. Otherwise, leave
2741 * this device alone for this sync request.
2743 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2744 bio->bi_end_io = end_sync_write;
2748 if (bio->bi_end_io) {
2749 atomic_inc(&rdev->nr_pending);
2750 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2751 bio->bi_bdev = rdev->bdev;
2752 if (test_bit(FailFast, &rdev->flags))
2753 bio->bi_opf |= MD_FAILFAST;
2759 r1_bio->read_disk = disk;
2761 if (read_targets == 0 && min_bad > 0) {
2762 /* These sectors are bad on all InSync devices, so we
2763 * need to mark them bad on all write targets
2766 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2767 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2768 struct md_rdev *rdev = conf->mirrors[i].rdev;
2769 ok = rdev_set_badblocks(rdev, sector_nr,
2773 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2778 /* Cannot record the badblocks, so need to
2780 * If there are multiple read targets, could just
2781 * fail the really bad ones ???
2783 conf->recovery_disabled = mddev->recovery_disabled;
2784 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2790 if (min_bad > 0 && min_bad < good_sectors) {
2791 /* only resync enough to reach the next bad->good
2793 good_sectors = min_bad;
2796 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2797 /* extra read targets are also write targets */
2798 write_targets += read_targets-1;
2800 if (write_targets == 0 || read_targets == 0) {
2801 /* There is nowhere to write, so all non-sync
2802 * drives must be failed - so we are finished
2806 max_sector = sector_nr + min_bad;
2807 rv = max_sector - sector_nr;
2813 if (max_sector > mddev->resync_max)
2814 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2815 if (max_sector > sector_nr + good_sectors)
2816 max_sector = sector_nr + good_sectors;
2821 int len = PAGE_SIZE;
2822 if (sector_nr + (len>>9) > max_sector)
2823 len = (max_sector - sector_nr) << 9;
2826 if (sync_blocks == 0) {
2827 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2828 &sync_blocks, still_degraded) &&
2830 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2832 if ((len >> 9) > sync_blocks)
2833 len = sync_blocks<<9;
2836 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2837 struct resync_pages *rp;
2839 bio = r1_bio->bios[i];
2840 rp = get_resync_pages(bio);
2841 if (bio->bi_end_io) {
2842 page = resync_fetch_page(rp, rp->idx++);
2845 * won't fail because the vec table is big
2846 * enough to hold all these pages
2848 bio_add_page(bio, page, len, 0);
2851 nr_sectors += len>>9;
2852 sector_nr += len>>9;
2853 sync_blocks -= (len>>9);
2854 } while (get_resync_pages(r1_bio->bios[disk]->bi_private)->idx < RESYNC_PAGES);
2856 r1_bio->sectors = nr_sectors;
2858 if (mddev_is_clustered(mddev) &&
2859 conf->cluster_sync_high < sector_nr + nr_sectors) {
2860 conf->cluster_sync_low = mddev->curr_resync_completed;
2861 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2862 /* Send resync message */
2863 md_cluster_ops->resync_info_update(mddev,
2864 conf->cluster_sync_low,
2865 conf->cluster_sync_high);
2868 /* For a user-requested sync, we read all readable devices and do a
2871 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2872 atomic_set(&r1_bio->remaining, read_targets);
2873 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2874 bio = r1_bio->bios[i];
2875 if (bio->bi_end_io == end_sync_read) {
2877 md_sync_acct(bio->bi_bdev, nr_sectors);
2878 if (read_targets == 1)
2879 bio->bi_opf &= ~MD_FAILFAST;
2880 generic_make_request(bio);
2884 atomic_set(&r1_bio->remaining, 1);
2885 bio = r1_bio->bios[r1_bio->read_disk];
2886 md_sync_acct(bio->bi_bdev, nr_sectors);
2887 if (read_targets == 1)
2888 bio->bi_opf &= ~MD_FAILFAST;
2889 generic_make_request(bio);
2895 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2900 return mddev->dev_sectors;
2903 static struct r1conf *setup_conf(struct mddev *mddev)
2905 struct r1conf *conf;
2907 struct raid1_info *disk;
2908 struct md_rdev *rdev;
2911 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2915 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2916 sizeof(atomic_t), GFP_KERNEL);
2917 if (!conf->nr_pending)
2920 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2921 sizeof(atomic_t), GFP_KERNEL);
2922 if (!conf->nr_waiting)
2925 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2926 sizeof(atomic_t), GFP_KERNEL);
2927 if (!conf->nr_queued)
2930 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2931 sizeof(atomic_t), GFP_KERNEL);
2935 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2936 * mddev->raid_disks * 2,
2941 conf->tmppage = alloc_page(GFP_KERNEL);
2945 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2946 if (!conf->poolinfo)
2948 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2949 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2952 if (!conf->r1bio_pool)
2955 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0);
2956 if (!conf->bio_split)
2959 conf->poolinfo->mddev = mddev;
2962 spin_lock_init(&conf->device_lock);
2963 rdev_for_each(rdev, mddev) {
2964 int disk_idx = rdev->raid_disk;
2965 if (disk_idx >= mddev->raid_disks
2968 if (test_bit(Replacement, &rdev->flags))
2969 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2971 disk = conf->mirrors + disk_idx;
2976 disk->head_position = 0;
2977 disk->seq_start = MaxSector;
2979 conf->raid_disks = mddev->raid_disks;
2980 conf->mddev = mddev;
2981 INIT_LIST_HEAD(&conf->retry_list);
2982 INIT_LIST_HEAD(&conf->bio_end_io_list);
2984 spin_lock_init(&conf->resync_lock);
2985 init_waitqueue_head(&conf->wait_barrier);
2987 bio_list_init(&conf->pending_bio_list);
2988 conf->pending_count = 0;
2989 conf->recovery_disabled = mddev->recovery_disabled - 1;
2992 for (i = 0; i < conf->raid_disks * 2; i++) {
2994 disk = conf->mirrors + i;
2996 if (i < conf->raid_disks &&
2997 disk[conf->raid_disks].rdev) {
2998 /* This slot has a replacement. */
3000 /* No original, just make the replacement
3001 * a recovering spare
3004 disk[conf->raid_disks].rdev;
3005 disk[conf->raid_disks].rdev = NULL;
3006 } else if (!test_bit(In_sync, &disk->rdev->flags))
3007 /* Original is not in_sync - bad */
3012 !test_bit(In_sync, &disk->rdev->flags)) {
3013 disk->head_position = 0;
3015 (disk->rdev->saved_raid_disk < 0))
3021 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3029 mempool_destroy(conf->r1bio_pool);
3030 kfree(conf->mirrors);
3031 safe_put_page(conf->tmppage);
3032 kfree(conf->poolinfo);
3033 kfree(conf->nr_pending);
3034 kfree(conf->nr_waiting);
3035 kfree(conf->nr_queued);
3036 kfree(conf->barrier);
3037 if (conf->bio_split)
3038 bioset_free(conf->bio_split);
3041 return ERR_PTR(err);
3044 static void raid1_free(struct mddev *mddev, void *priv);
3045 static int raid1_run(struct mddev *mddev)
3047 struct r1conf *conf;
3049 struct md_rdev *rdev;
3051 bool discard_supported = false;
3053 if (mddev->level != 1) {
3054 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3055 mdname(mddev), mddev->level);
3058 if (mddev->reshape_position != MaxSector) {
3059 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3064 * copy the already verified devices into our private RAID1
3065 * bookkeeping area. [whatever we allocate in run(),
3066 * should be freed in raid1_free()]
3068 if (mddev->private == NULL)
3069 conf = setup_conf(mddev);
3071 conf = mddev->private;
3074 return PTR_ERR(conf);
3077 blk_queue_max_write_same_sectors(mddev->queue, 0);
3078 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3081 rdev_for_each(rdev, mddev) {
3082 if (!mddev->gendisk)
3084 disk_stack_limits(mddev->gendisk, rdev->bdev,
3085 rdev->data_offset << 9);
3086 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3087 discard_supported = true;
3090 mddev->degraded = 0;
3091 for (i=0; i < conf->raid_disks; i++)
3092 if (conf->mirrors[i].rdev == NULL ||
3093 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3094 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3097 if (conf->raid_disks - mddev->degraded == 1)
3098 mddev->recovery_cp = MaxSector;
3100 if (mddev->recovery_cp != MaxSector)
3101 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3103 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3104 mdname(mddev), mddev->raid_disks - mddev->degraded,
3108 * Ok, everything is just fine now
3110 mddev->thread = conf->thread;
3111 conf->thread = NULL;
3112 mddev->private = conf;
3113 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3115 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3118 if (discard_supported)
3119 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3122 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3126 ret = md_integrity_register(mddev);
3128 md_unregister_thread(&mddev->thread);
3129 raid1_free(mddev, conf);
3134 static void raid1_free(struct mddev *mddev, void *priv)
3136 struct r1conf *conf = priv;
3138 mempool_destroy(conf->r1bio_pool);
3139 kfree(conf->mirrors);
3140 safe_put_page(conf->tmppage);
3141 kfree(conf->poolinfo);
3142 kfree(conf->nr_pending);
3143 kfree(conf->nr_waiting);
3144 kfree(conf->nr_queued);
3145 kfree(conf->barrier);
3146 if (conf->bio_split)
3147 bioset_free(conf->bio_split);
3151 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3153 /* no resync is happening, and there is enough space
3154 * on all devices, so we can resize.
3155 * We need to make sure resync covers any new space.
3156 * If the array is shrinking we should possibly wait until
3157 * any io in the removed space completes, but it hardly seems
3160 sector_t newsize = raid1_size(mddev, sectors, 0);
3161 if (mddev->external_size &&
3162 mddev->array_sectors > newsize)
3164 if (mddev->bitmap) {
3165 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3169 md_set_array_sectors(mddev, newsize);
3170 if (sectors > mddev->dev_sectors &&
3171 mddev->recovery_cp > mddev->dev_sectors) {
3172 mddev->recovery_cp = mddev->dev_sectors;
3173 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3175 mddev->dev_sectors = sectors;
3176 mddev->resync_max_sectors = sectors;
3180 static int raid1_reshape(struct mddev *mddev)
3183 * 1/ resize the r1bio_pool
3184 * 2/ resize conf->mirrors
3186 * We allocate a new r1bio_pool if we can.
3187 * Then raise a device barrier and wait until all IO stops.
3188 * Then resize conf->mirrors and swap in the new r1bio pool.
3190 * At the same time, we "pack" the devices so that all the missing
3191 * devices have the higher raid_disk numbers.
3193 mempool_t *newpool, *oldpool;
3194 struct pool_info *newpoolinfo;
3195 struct raid1_info *newmirrors;
3196 struct r1conf *conf = mddev->private;
3197 int cnt, raid_disks;
3198 unsigned long flags;
3201 /* Cannot change chunk_size, layout, or level */
3202 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3203 mddev->layout != mddev->new_layout ||
3204 mddev->level != mddev->new_level) {
3205 mddev->new_chunk_sectors = mddev->chunk_sectors;
3206 mddev->new_layout = mddev->layout;
3207 mddev->new_level = mddev->level;
3211 if (!mddev_is_clustered(mddev))
3212 md_allow_write(mddev);
3214 raid_disks = mddev->raid_disks + mddev->delta_disks;
3216 if (raid_disks < conf->raid_disks) {
3218 for (d= 0; d < conf->raid_disks; d++)
3219 if (conf->mirrors[d].rdev)
3221 if (cnt > raid_disks)
3225 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3228 newpoolinfo->mddev = mddev;
3229 newpoolinfo->raid_disks = raid_disks * 2;
3231 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3232 r1bio_pool_free, newpoolinfo);
3237 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3241 mempool_destroy(newpool);
3245 freeze_array(conf, 0);
3247 /* ok, everything is stopped */
3248 oldpool = conf->r1bio_pool;
3249 conf->r1bio_pool = newpool;
3251 for (d = d2 = 0; d < conf->raid_disks; d++) {
3252 struct md_rdev *rdev = conf->mirrors[d].rdev;
3253 if (rdev && rdev->raid_disk != d2) {
3254 sysfs_unlink_rdev(mddev, rdev);
3255 rdev->raid_disk = d2;
3256 sysfs_unlink_rdev(mddev, rdev);
3257 if (sysfs_link_rdev(mddev, rdev))
3258 pr_warn("md/raid1:%s: cannot register rd%d\n",
3259 mdname(mddev), rdev->raid_disk);
3262 newmirrors[d2++].rdev = rdev;
3264 kfree(conf->mirrors);
3265 conf->mirrors = newmirrors;
3266 kfree(conf->poolinfo);
3267 conf->poolinfo = newpoolinfo;
3269 spin_lock_irqsave(&conf->device_lock, flags);
3270 mddev->degraded += (raid_disks - conf->raid_disks);
3271 spin_unlock_irqrestore(&conf->device_lock, flags);
3272 conf->raid_disks = mddev->raid_disks = raid_disks;
3273 mddev->delta_disks = 0;
3275 unfreeze_array(conf);
3277 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3278 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3279 md_wakeup_thread(mddev->thread);
3281 mempool_destroy(oldpool);
3285 static void raid1_quiesce(struct mddev *mddev, int state)
3287 struct r1conf *conf = mddev->private;
3290 case 2: /* wake for suspend */
3291 wake_up(&conf->wait_barrier);
3294 freeze_array(conf, 0);
3297 unfreeze_array(conf);
3302 static void *raid1_takeover(struct mddev *mddev)
3304 /* raid1 can take over:
3305 * raid5 with 2 devices, any layout or chunk size
3307 if (mddev->level == 5 && mddev->raid_disks == 2) {
3308 struct r1conf *conf;
3309 mddev->new_level = 1;
3310 mddev->new_layout = 0;
3311 mddev->new_chunk_sectors = 0;
3312 conf = setup_conf(mddev);
3313 if (!IS_ERR(conf)) {
3314 /* Array must appear to be quiesced */
3315 conf->array_frozen = 1;
3316 mddev_clear_unsupported_flags(mddev,
3317 UNSUPPORTED_MDDEV_FLAGS);
3321 return ERR_PTR(-EINVAL);
3324 static struct md_personality raid1_personality =
3328 .owner = THIS_MODULE,
3329 .make_request = raid1_make_request,
3332 .status = raid1_status,
3333 .error_handler = raid1_error,
3334 .hot_add_disk = raid1_add_disk,
3335 .hot_remove_disk= raid1_remove_disk,
3336 .spare_active = raid1_spare_active,
3337 .sync_request = raid1_sync_request,
3338 .resize = raid1_resize,
3340 .check_reshape = raid1_reshape,
3341 .quiesce = raid1_quiesce,
3342 .takeover = raid1_takeover,
3343 .congested = raid1_congested,
3346 static int __init raid_init(void)
3348 return register_md_personality(&raid1_personality);
3351 static void raid_exit(void)
3353 unregister_md_personality(&raid1_personality);
3356 module_init(raid_init);
3357 module_exit(raid_exit);
3358 MODULE_LICENSE("GPL");
3359 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3360 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3361 MODULE_ALIAS("md-raid1");
3362 MODULE_ALIAS("md-level-1");
3364 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);