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)
1207 struct r1conf *conf = mddev->private;
1208 struct raid1_info *mirror;
1209 struct r1bio *r1_bio;
1210 struct bio *read_bio;
1211 struct bitmap *bitmap = mddev->bitmap;
1212 const int op = bio_op(bio);
1213 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1214 int sectors_handled;
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);
1227 * make_request() can abort the operation when read-ahead is being
1228 * 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);
1250 r1_bio->read_disk = rdisk;
1252 read_bio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1253 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1256 r1_bio->bios[rdisk] = read_bio;
1258 read_bio->bi_iter.bi_sector = r1_bio->sector +
1259 mirror->rdev->data_offset;
1260 read_bio->bi_bdev = mirror->rdev->bdev;
1261 read_bio->bi_end_io = raid1_end_read_request;
1262 bio_set_op_attrs(read_bio, op, do_sync);
1263 if (test_bit(FailFast, &mirror->rdev->flags) &&
1264 test_bit(R1BIO_FailFast, &r1_bio->state))
1265 read_bio->bi_opf |= MD_FAILFAST;
1266 read_bio->bi_private = r1_bio;
1269 trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1270 read_bio, disk_devt(mddev->gendisk),
1273 if (max_sectors < r1_bio->sectors) {
1275 * could not read all from this device, so we will need another
1278 sectors_handled = (r1_bio->sector + max_sectors
1279 - bio->bi_iter.bi_sector);
1280 r1_bio->sectors = max_sectors;
1281 bio_inc_remaining(bio);
1284 * Cannot call generic_make_request directly as that will be
1285 * queued in __make_request and subsequent mempool_alloc might
1286 * block waiting for it. So hand bio over to raid1d.
1288 reschedule_retry(r1_bio);
1290 r1_bio = alloc_r1bio(mddev, bio, sectors_handled);
1293 generic_make_request(read_bio);
1296 static void raid1_write_request(struct mddev *mddev, struct bio *bio)
1298 struct r1conf *conf = mddev->private;
1299 struct r1bio *r1_bio;
1301 struct bitmap *bitmap = mddev->bitmap;
1302 unsigned long flags;
1303 struct md_rdev *blocked_rdev;
1304 struct blk_plug_cb *cb;
1305 struct raid1_plug_cb *plug = NULL;
1307 int sectors_handled;
1312 * Register the new request and wait if the reconstruction
1313 * thread has put up a bar for new requests.
1314 * Continue immediately if no resync is active currently.
1317 md_write_start(mddev, bio); /* wait on superblock update early */
1319 if ((bio_end_sector(bio) > mddev->suspend_lo &&
1320 bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1321 (mddev_is_clustered(mddev) &&
1322 md_cluster_ops->area_resyncing(mddev, WRITE,
1323 bio->bi_iter.bi_sector, bio_end_sector(bio)))) {
1326 * As the suspend_* range is controlled by userspace, we want
1327 * an interruptible wait.
1331 flush_signals(current);
1332 prepare_to_wait(&conf->wait_barrier,
1333 &w, TASK_INTERRUPTIBLE);
1334 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1335 bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1336 (mddev_is_clustered(mddev) &&
1337 !md_cluster_ops->area_resyncing(mddev, WRITE,
1338 bio->bi_iter.bi_sector,
1339 bio_end_sector(bio))))
1343 finish_wait(&conf->wait_barrier, &w);
1345 wait_barrier(conf, bio->bi_iter.bi_sector);
1347 r1_bio = alloc_r1bio(mddev, bio, 0);
1349 if (conf->pending_count >= max_queued_requests) {
1350 md_wakeup_thread(mddev->thread);
1351 raid1_log(mddev, "wait queued");
1352 wait_event(conf->wait_barrier,
1353 conf->pending_count < max_queued_requests);
1355 /* first select target devices under rcu_lock and
1356 * inc refcount on their rdev. Record them by setting
1358 * If there are known/acknowledged bad blocks on any device on
1359 * which we have seen a write error, we want to avoid writing those
1361 * This potentially requires several writes to write around
1362 * the bad blocks. Each set of writes gets it's own r1bio
1363 * with a set of bios attached.
1366 disks = conf->raid_disks * 2;
1368 blocked_rdev = NULL;
1370 max_sectors = r1_bio->sectors;
1371 for (i = 0; i < disks; i++) {
1372 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1373 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1374 atomic_inc(&rdev->nr_pending);
1375 blocked_rdev = rdev;
1378 r1_bio->bios[i] = NULL;
1379 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1380 if (i < conf->raid_disks)
1381 set_bit(R1BIO_Degraded, &r1_bio->state);
1385 atomic_inc(&rdev->nr_pending);
1386 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1391 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1392 &first_bad, &bad_sectors);
1394 /* mustn't write here until the bad block is
1396 set_bit(BlockedBadBlocks, &rdev->flags);
1397 blocked_rdev = rdev;
1400 if (is_bad && first_bad <= r1_bio->sector) {
1401 /* Cannot write here at all */
1402 bad_sectors -= (r1_bio->sector - first_bad);
1403 if (bad_sectors < max_sectors)
1404 /* mustn't write more than bad_sectors
1405 * to other devices yet
1407 max_sectors = bad_sectors;
1408 rdev_dec_pending(rdev, mddev);
1409 /* We don't set R1BIO_Degraded as that
1410 * only applies if the disk is
1411 * missing, so it might be re-added,
1412 * and we want to know to recover this
1414 * In this case the device is here,
1415 * and the fact that this chunk is not
1416 * in-sync is recorded in the bad
1422 int good_sectors = first_bad - r1_bio->sector;
1423 if (good_sectors < max_sectors)
1424 max_sectors = good_sectors;
1427 r1_bio->bios[i] = bio;
1431 if (unlikely(blocked_rdev)) {
1432 /* Wait for this device to become unblocked */
1435 for (j = 0; j < i; j++)
1436 if (r1_bio->bios[j])
1437 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1439 allow_barrier(conf, bio->bi_iter.bi_sector);
1440 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1441 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1442 wait_barrier(conf, bio->bi_iter.bi_sector);
1446 if (max_sectors < r1_bio->sectors)
1447 r1_bio->sectors = max_sectors;
1449 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1451 atomic_set(&r1_bio->remaining, 1);
1452 atomic_set(&r1_bio->behind_remaining, 0);
1456 offset = r1_bio->sector - bio->bi_iter.bi_sector;
1457 for (i = 0; i < disks; i++) {
1458 struct bio *mbio = NULL;
1459 if (!r1_bio->bios[i])
1465 * Not if there are too many, or cannot
1466 * allocate memory, or a reader on WriteMostly
1467 * is waiting for behind writes to flush */
1469 (atomic_read(&bitmap->behind_writes)
1470 < mddev->bitmap_info.max_write_behind) &&
1471 !waitqueue_active(&bitmap->behind_wait)) {
1472 mbio = alloc_behind_master_bio(r1_bio, bio,
1477 bitmap_startwrite(bitmap, r1_bio->sector,
1479 test_bit(R1BIO_BehindIO,
1485 if (r1_bio->behind_master_bio)
1486 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1490 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1491 bio_trim(mbio, offset, max_sectors);
1495 if (r1_bio->behind_master_bio) {
1496 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1497 atomic_inc(&r1_bio->behind_remaining);
1500 r1_bio->bios[i] = mbio;
1502 mbio->bi_iter.bi_sector = (r1_bio->sector +
1503 conf->mirrors[i].rdev->data_offset);
1504 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1505 mbio->bi_end_io = raid1_end_write_request;
1506 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1507 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1508 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1509 conf->raid_disks - mddev->degraded > 1)
1510 mbio->bi_opf |= MD_FAILFAST;
1511 mbio->bi_private = r1_bio;
1513 atomic_inc(&r1_bio->remaining);
1516 trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1517 mbio, disk_devt(mddev->gendisk),
1519 /* flush_pending_writes() needs access to the rdev so...*/
1520 mbio->bi_bdev = (void*)conf->mirrors[i].rdev;
1522 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1524 plug = container_of(cb, struct raid1_plug_cb, cb);
1527 spin_lock_irqsave(&conf->device_lock, flags);
1529 bio_list_add(&plug->pending, mbio);
1530 plug->pending_cnt++;
1532 bio_list_add(&conf->pending_bio_list, mbio);
1533 conf->pending_count++;
1535 spin_unlock_irqrestore(&conf->device_lock, flags);
1537 md_wakeup_thread(mddev->thread);
1539 /* Mustn't call r1_bio_write_done before this next test,
1540 * as it could result in the bio being freed.
1542 if (sectors_handled < bio_sectors(bio)) {
1543 /* We need another r1_bio, which must be counted */
1544 sector_t sect = bio->bi_iter.bi_sector + sectors_handled;
1546 inc_pending(conf, sect);
1547 bio_inc_remaining(bio);
1548 r1_bio_write_done(r1_bio);
1549 r1_bio = alloc_r1bio(mddev, bio, sectors_handled);
1553 r1_bio_write_done(r1_bio);
1555 /* In case raid1d snuck in to freeze_array */
1556 wake_up(&conf->wait_barrier);
1559 static void raid1_make_request(struct mddev *mddev, struct bio *bio)
1564 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1565 md_flush_request(mddev, bio);
1569 /* if bio exceeds barrier unit boundary, split it */
1571 sectors = align_to_barrier_unit_end(
1572 bio->bi_iter.bi_sector, bio_sectors(bio));
1573 if (sectors < bio_sectors(bio)) {
1574 split = bio_split(bio, sectors, GFP_NOIO, fs_bio_set);
1575 bio_chain(split, bio);
1580 if (bio_data_dir(split) == READ) {
1581 raid1_read_request(mddev, split);
1584 * If a bio is splitted, the first part of bio will
1585 * pass barrier but the bio is queued in
1586 * current->bio_list (see generic_make_request). If
1587 * there is a raise_barrier() called here, the second
1588 * part of bio can't pass barrier. But since the first
1589 * part bio isn't dispatched to underlaying disks yet,
1590 * the barrier is never released, hence raise_barrier
1591 * will alays wait. We have a deadlock.
1592 * Note, this only happens in read path. For write
1593 * path, the first part of bio is dispatched in a
1594 * schedule() call (because of blk plug) or offloaded
1596 * Quitting from the function immediately can change
1597 * the bio order queued in bio_list and avoid the deadlock.
1600 generic_make_request(bio);
1604 raid1_write_request(mddev, split);
1605 } while (split != bio);
1608 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1610 struct r1conf *conf = mddev->private;
1613 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1614 conf->raid_disks - mddev->degraded);
1616 for (i = 0; i < conf->raid_disks; i++) {
1617 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1618 seq_printf(seq, "%s",
1619 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1622 seq_printf(seq, "]");
1625 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1627 char b[BDEVNAME_SIZE];
1628 struct r1conf *conf = mddev->private;
1629 unsigned long flags;
1632 * If it is not operational, then we have already marked it as dead
1633 * else if it is the last working disks, ignore the error, let the
1634 * next level up know.
1635 * else mark the drive as failed
1637 spin_lock_irqsave(&conf->device_lock, flags);
1638 if (test_bit(In_sync, &rdev->flags)
1639 && (conf->raid_disks - mddev->degraded) == 1) {
1641 * Don't fail the drive, act as though we were just a
1642 * normal single drive.
1643 * However don't try a recovery from this drive as
1644 * it is very likely to fail.
1646 conf->recovery_disabled = mddev->recovery_disabled;
1647 spin_unlock_irqrestore(&conf->device_lock, flags);
1650 set_bit(Blocked, &rdev->flags);
1651 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1653 set_bit(Faulty, &rdev->flags);
1655 set_bit(Faulty, &rdev->flags);
1656 spin_unlock_irqrestore(&conf->device_lock, flags);
1658 * if recovery is running, make sure it aborts.
1660 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1661 set_mask_bits(&mddev->sb_flags, 0,
1662 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1663 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1664 "md/raid1:%s: Operation continuing on %d devices.\n",
1665 mdname(mddev), bdevname(rdev->bdev, b),
1666 mdname(mddev), conf->raid_disks - mddev->degraded);
1669 static void print_conf(struct r1conf *conf)
1673 pr_debug("RAID1 conf printout:\n");
1675 pr_debug("(!conf)\n");
1678 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1682 for (i = 0; i < conf->raid_disks; i++) {
1683 char b[BDEVNAME_SIZE];
1684 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1686 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1687 i, !test_bit(In_sync, &rdev->flags),
1688 !test_bit(Faulty, &rdev->flags),
1689 bdevname(rdev->bdev,b));
1694 static void close_sync(struct r1conf *conf)
1696 wait_all_barriers(conf);
1697 allow_all_barriers(conf);
1699 mempool_destroy(conf->r1buf_pool);
1700 conf->r1buf_pool = NULL;
1703 static int raid1_spare_active(struct mddev *mddev)
1706 struct r1conf *conf = mddev->private;
1708 unsigned long flags;
1711 * Find all failed disks within the RAID1 configuration
1712 * and mark them readable.
1713 * Called under mddev lock, so rcu protection not needed.
1714 * device_lock used to avoid races with raid1_end_read_request
1715 * which expects 'In_sync' flags and ->degraded to be consistent.
1717 spin_lock_irqsave(&conf->device_lock, flags);
1718 for (i = 0; i < conf->raid_disks; i++) {
1719 struct md_rdev *rdev = conf->mirrors[i].rdev;
1720 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1722 && !test_bit(Candidate, &repl->flags)
1723 && repl->recovery_offset == MaxSector
1724 && !test_bit(Faulty, &repl->flags)
1725 && !test_and_set_bit(In_sync, &repl->flags)) {
1726 /* replacement has just become active */
1728 !test_and_clear_bit(In_sync, &rdev->flags))
1731 /* Replaced device not technically
1732 * faulty, but we need to be sure
1733 * it gets removed and never re-added
1735 set_bit(Faulty, &rdev->flags);
1736 sysfs_notify_dirent_safe(
1741 && rdev->recovery_offset == MaxSector
1742 && !test_bit(Faulty, &rdev->flags)
1743 && !test_and_set_bit(In_sync, &rdev->flags)) {
1745 sysfs_notify_dirent_safe(rdev->sysfs_state);
1748 mddev->degraded -= count;
1749 spin_unlock_irqrestore(&conf->device_lock, flags);
1755 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1757 struct r1conf *conf = mddev->private;
1760 struct raid1_info *p;
1762 int last = conf->raid_disks - 1;
1764 if (mddev->recovery_disabled == conf->recovery_disabled)
1767 if (md_integrity_add_rdev(rdev, mddev))
1770 if (rdev->raid_disk >= 0)
1771 first = last = rdev->raid_disk;
1774 * find the disk ... but prefer rdev->saved_raid_disk
1777 if (rdev->saved_raid_disk >= 0 &&
1778 rdev->saved_raid_disk >= first &&
1779 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1780 first = last = rdev->saved_raid_disk;
1782 for (mirror = first; mirror <= last; mirror++) {
1783 p = conf->mirrors+mirror;
1787 disk_stack_limits(mddev->gendisk, rdev->bdev,
1788 rdev->data_offset << 9);
1790 p->head_position = 0;
1791 rdev->raid_disk = mirror;
1793 /* As all devices are equivalent, we don't need a full recovery
1794 * if this was recently any drive of the array
1796 if (rdev->saved_raid_disk < 0)
1798 rcu_assign_pointer(p->rdev, rdev);
1801 if (test_bit(WantReplacement, &p->rdev->flags) &&
1802 p[conf->raid_disks].rdev == NULL) {
1803 /* Add this device as a replacement */
1804 clear_bit(In_sync, &rdev->flags);
1805 set_bit(Replacement, &rdev->flags);
1806 rdev->raid_disk = mirror;
1809 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1813 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1814 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1819 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1821 struct r1conf *conf = mddev->private;
1823 int number = rdev->raid_disk;
1824 struct raid1_info *p = conf->mirrors + number;
1826 if (rdev != p->rdev)
1827 p = conf->mirrors + conf->raid_disks + number;
1830 if (rdev == p->rdev) {
1831 if (test_bit(In_sync, &rdev->flags) ||
1832 atomic_read(&rdev->nr_pending)) {
1836 /* Only remove non-faulty devices if recovery
1839 if (!test_bit(Faulty, &rdev->flags) &&
1840 mddev->recovery_disabled != conf->recovery_disabled &&
1841 mddev->degraded < conf->raid_disks) {
1846 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1848 if (atomic_read(&rdev->nr_pending)) {
1849 /* lost the race, try later */
1855 if (conf->mirrors[conf->raid_disks + number].rdev) {
1856 /* We just removed a device that is being replaced.
1857 * Move down the replacement. We drain all IO before
1858 * doing this to avoid confusion.
1860 struct md_rdev *repl =
1861 conf->mirrors[conf->raid_disks + number].rdev;
1862 freeze_array(conf, 0);
1863 clear_bit(Replacement, &repl->flags);
1865 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1866 unfreeze_array(conf);
1867 clear_bit(WantReplacement, &rdev->flags);
1869 clear_bit(WantReplacement, &rdev->flags);
1870 err = md_integrity_register(mddev);
1878 static void end_sync_read(struct bio *bio)
1880 struct r1bio *r1_bio = get_resync_r1bio(bio);
1882 update_head_pos(r1_bio->read_disk, r1_bio);
1885 * we have read a block, now it needs to be re-written,
1886 * or re-read if the read failed.
1887 * We don't do much here, just schedule handling by raid1d
1890 set_bit(R1BIO_Uptodate, &r1_bio->state);
1892 if (atomic_dec_and_test(&r1_bio->remaining))
1893 reschedule_retry(r1_bio);
1896 static void end_sync_write(struct bio *bio)
1898 int uptodate = !bio->bi_error;
1899 struct r1bio *r1_bio = get_resync_r1bio(bio);
1900 struct mddev *mddev = r1_bio->mddev;
1901 struct r1conf *conf = mddev->private;
1904 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1907 sector_t sync_blocks = 0;
1908 sector_t s = r1_bio->sector;
1909 long sectors_to_go = r1_bio->sectors;
1910 /* make sure these bits doesn't get cleared. */
1912 bitmap_end_sync(mddev->bitmap, s,
1915 sectors_to_go -= sync_blocks;
1916 } while (sectors_to_go > 0);
1917 set_bit(WriteErrorSeen, &rdev->flags);
1918 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1919 set_bit(MD_RECOVERY_NEEDED, &
1921 set_bit(R1BIO_WriteError, &r1_bio->state);
1922 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1923 &first_bad, &bad_sectors) &&
1924 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1927 &first_bad, &bad_sectors)
1929 set_bit(R1BIO_MadeGood, &r1_bio->state);
1931 if (atomic_dec_and_test(&r1_bio->remaining)) {
1932 int s = r1_bio->sectors;
1933 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1934 test_bit(R1BIO_WriteError, &r1_bio->state))
1935 reschedule_retry(r1_bio);
1938 md_done_sync(mddev, s, uptodate);
1943 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1944 int sectors, struct page *page, int rw)
1946 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1950 set_bit(WriteErrorSeen, &rdev->flags);
1951 if (!test_and_set_bit(WantReplacement,
1953 set_bit(MD_RECOVERY_NEEDED, &
1954 rdev->mddev->recovery);
1956 /* need to record an error - either for the block or the device */
1957 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1958 md_error(rdev->mddev, rdev);
1962 static int fix_sync_read_error(struct r1bio *r1_bio)
1964 /* Try some synchronous reads of other devices to get
1965 * good data, much like with normal read errors. Only
1966 * read into the pages we already have so we don't
1967 * need to re-issue the read request.
1968 * We don't need to freeze the array, because being in an
1969 * active sync request, there is no normal IO, and
1970 * no overlapping syncs.
1971 * We don't need to check is_badblock() again as we
1972 * made sure that anything with a bad block in range
1973 * will have bi_end_io clear.
1975 struct mddev *mddev = r1_bio->mddev;
1976 struct r1conf *conf = mddev->private;
1977 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1978 struct page **pages = get_resync_pages(bio)->pages;
1979 sector_t sect = r1_bio->sector;
1980 int sectors = r1_bio->sectors;
1982 struct md_rdev *rdev;
1984 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1985 if (test_bit(FailFast, &rdev->flags)) {
1986 /* Don't try recovering from here - just fail it
1987 * ... unless it is the last working device of course */
1988 md_error(mddev, rdev);
1989 if (test_bit(Faulty, &rdev->flags))
1990 /* Don't try to read from here, but make sure
1991 * put_buf does it's thing
1993 bio->bi_end_io = end_sync_write;
1998 int d = r1_bio->read_disk;
2002 if (s > (PAGE_SIZE>>9))
2005 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2006 /* No rcu protection needed here devices
2007 * can only be removed when no resync is
2008 * active, and resync is currently active
2010 rdev = conf->mirrors[d].rdev;
2011 if (sync_page_io(rdev, sect, s<<9,
2013 REQ_OP_READ, 0, false)) {
2019 if (d == conf->raid_disks * 2)
2021 } while (!success && d != r1_bio->read_disk);
2024 char b[BDEVNAME_SIZE];
2026 /* Cannot read from anywhere, this block is lost.
2027 * Record a bad block on each device. If that doesn't
2028 * work just disable and interrupt the recovery.
2029 * Don't fail devices as that won't really help.
2031 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2033 bdevname(bio->bi_bdev, b),
2034 (unsigned long long)r1_bio->sector);
2035 for (d = 0; d < conf->raid_disks * 2; d++) {
2036 rdev = conf->mirrors[d].rdev;
2037 if (!rdev || test_bit(Faulty, &rdev->flags))
2039 if (!rdev_set_badblocks(rdev, sect, s, 0))
2043 conf->recovery_disabled =
2044 mddev->recovery_disabled;
2045 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2046 md_done_sync(mddev, r1_bio->sectors, 0);
2058 /* write it back and re-read */
2059 while (d != r1_bio->read_disk) {
2061 d = conf->raid_disks * 2;
2063 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2065 rdev = conf->mirrors[d].rdev;
2066 if (r1_sync_page_io(rdev, sect, s,
2069 r1_bio->bios[d]->bi_end_io = NULL;
2070 rdev_dec_pending(rdev, mddev);
2074 while (d != r1_bio->read_disk) {
2076 d = conf->raid_disks * 2;
2078 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2080 rdev = conf->mirrors[d].rdev;
2081 if (r1_sync_page_io(rdev, sect, s,
2084 atomic_add(s, &rdev->corrected_errors);
2090 set_bit(R1BIO_Uptodate, &r1_bio->state);
2095 static void process_checks(struct r1bio *r1_bio)
2097 /* We have read all readable devices. If we haven't
2098 * got the block, then there is no hope left.
2099 * If we have, then we want to do a comparison
2100 * and skip the write if everything is the same.
2101 * If any blocks failed to read, then we need to
2102 * attempt an over-write
2104 struct mddev *mddev = r1_bio->mddev;
2105 struct r1conf *conf = mddev->private;
2110 /* Fix variable parts of all bios */
2111 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2112 for (i = 0; i < conf->raid_disks * 2; i++) {
2117 struct bio *b = r1_bio->bios[i];
2118 struct resync_pages *rp = get_resync_pages(b);
2119 if (b->bi_end_io != end_sync_read)
2121 /* fixup the bio for reuse, but preserve errno */
2122 error = b->bi_error;
2124 b->bi_error = error;
2126 b->bi_iter.bi_size = r1_bio->sectors << 9;
2127 b->bi_iter.bi_sector = r1_bio->sector +
2128 conf->mirrors[i].rdev->data_offset;
2129 b->bi_bdev = conf->mirrors[i].rdev->bdev;
2130 b->bi_end_io = end_sync_read;
2131 rp->raid_bio = r1_bio;
2134 size = b->bi_iter.bi_size;
2135 bio_for_each_segment_all(bi, b, j) {
2137 if (size > PAGE_SIZE)
2138 bi->bv_len = PAGE_SIZE;
2144 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2145 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2146 !r1_bio->bios[primary]->bi_error) {
2147 r1_bio->bios[primary]->bi_end_io = NULL;
2148 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2151 r1_bio->read_disk = primary;
2152 for (i = 0; i < conf->raid_disks * 2; i++) {
2154 struct bio *pbio = r1_bio->bios[primary];
2155 struct bio *sbio = r1_bio->bios[i];
2156 int error = sbio->bi_error;
2157 struct page **ppages = get_resync_pages(pbio)->pages;
2158 struct page **spages = get_resync_pages(sbio)->pages;
2160 int page_len[RESYNC_PAGES] = { 0 };
2162 if (sbio->bi_end_io != end_sync_read)
2164 /* Now we can 'fixup' the error value */
2167 bio_for_each_segment_all(bi, sbio, j)
2168 page_len[j] = bi->bv_len;
2171 for (j = vcnt; j-- ; ) {
2172 if (memcmp(page_address(ppages[j]),
2173 page_address(spages[j]),
2180 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2181 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2183 /* No need to write to this device. */
2184 sbio->bi_end_io = NULL;
2185 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2189 bio_copy_data(sbio, pbio);
2193 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2195 struct r1conf *conf = mddev->private;
2197 int disks = conf->raid_disks * 2;
2198 struct bio *bio, *wbio;
2200 bio = r1_bio->bios[r1_bio->read_disk];
2202 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2203 /* ouch - failed to read all of that. */
2204 if (!fix_sync_read_error(r1_bio))
2207 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2208 process_checks(r1_bio);
2213 atomic_set(&r1_bio->remaining, 1);
2214 for (i = 0; i < disks ; i++) {
2215 wbio = r1_bio->bios[i];
2216 if (wbio->bi_end_io == NULL ||
2217 (wbio->bi_end_io == end_sync_read &&
2218 (i == r1_bio->read_disk ||
2219 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2221 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2224 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2225 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2226 wbio->bi_opf |= MD_FAILFAST;
2228 wbio->bi_end_io = end_sync_write;
2229 atomic_inc(&r1_bio->remaining);
2230 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2232 generic_make_request(wbio);
2235 if (atomic_dec_and_test(&r1_bio->remaining)) {
2236 /* if we're here, all write(s) have completed, so clean up */
2237 int s = r1_bio->sectors;
2238 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2239 test_bit(R1BIO_WriteError, &r1_bio->state))
2240 reschedule_retry(r1_bio);
2243 md_done_sync(mddev, s, 1);
2249 * This is a kernel thread which:
2251 * 1. Retries failed read operations on working mirrors.
2252 * 2. Updates the raid superblock when problems encounter.
2253 * 3. Performs writes following reads for array synchronising.
2256 static void fix_read_error(struct r1conf *conf, int read_disk,
2257 sector_t sect, int sectors)
2259 struct mddev *mddev = conf->mddev;
2265 struct md_rdev *rdev;
2267 if (s > (PAGE_SIZE>>9))
2275 rdev = rcu_dereference(conf->mirrors[d].rdev);
2277 (test_bit(In_sync, &rdev->flags) ||
2278 (!test_bit(Faulty, &rdev->flags) &&
2279 rdev->recovery_offset >= sect + s)) &&
2280 is_badblock(rdev, sect, s,
2281 &first_bad, &bad_sectors) == 0) {
2282 atomic_inc(&rdev->nr_pending);
2284 if (sync_page_io(rdev, sect, s<<9,
2285 conf->tmppage, REQ_OP_READ, 0, false))
2287 rdev_dec_pending(rdev, mddev);
2293 if (d == conf->raid_disks * 2)
2295 } while (!success && d != read_disk);
2298 /* Cannot read from anywhere - mark it bad */
2299 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2300 if (!rdev_set_badblocks(rdev, sect, s, 0))
2301 md_error(mddev, rdev);
2304 /* write it back and re-read */
2306 while (d != read_disk) {
2308 d = conf->raid_disks * 2;
2311 rdev = rcu_dereference(conf->mirrors[d].rdev);
2313 !test_bit(Faulty, &rdev->flags)) {
2314 atomic_inc(&rdev->nr_pending);
2316 r1_sync_page_io(rdev, sect, s,
2317 conf->tmppage, WRITE);
2318 rdev_dec_pending(rdev, mddev);
2323 while (d != read_disk) {
2324 char b[BDEVNAME_SIZE];
2326 d = conf->raid_disks * 2;
2329 rdev = rcu_dereference(conf->mirrors[d].rdev);
2331 !test_bit(Faulty, &rdev->flags)) {
2332 atomic_inc(&rdev->nr_pending);
2334 if (r1_sync_page_io(rdev, sect, s,
2335 conf->tmppage, READ)) {
2336 atomic_add(s, &rdev->corrected_errors);
2337 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2339 (unsigned long long)(sect +
2341 bdevname(rdev->bdev, b));
2343 rdev_dec_pending(rdev, mddev);
2352 static int narrow_write_error(struct r1bio *r1_bio, int i)
2354 struct mddev *mddev = r1_bio->mddev;
2355 struct r1conf *conf = mddev->private;
2356 struct md_rdev *rdev = conf->mirrors[i].rdev;
2358 /* bio has the data to be written to device 'i' where
2359 * we just recently had a write error.
2360 * We repeatedly clone the bio and trim down to one block,
2361 * then try the write. Where the write fails we record
2363 * It is conceivable that the bio doesn't exactly align with
2364 * blocks. We must handle this somehow.
2366 * We currently own a reference on the rdev.
2372 int sect_to_write = r1_bio->sectors;
2375 if (rdev->badblocks.shift < 0)
2378 block_sectors = roundup(1 << rdev->badblocks.shift,
2379 bdev_logical_block_size(rdev->bdev) >> 9);
2380 sector = r1_bio->sector;
2381 sectors = ((sector + block_sectors)
2382 & ~(sector_t)(block_sectors - 1))
2385 while (sect_to_write) {
2387 if (sectors > sect_to_write)
2388 sectors = sect_to_write;
2389 /* Write at 'sector' for 'sectors'*/
2391 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2392 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2395 /* We really need a _all clone */
2396 wbio->bi_iter = (struct bvec_iter){ 0 };
2398 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2402 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2403 wbio->bi_iter.bi_sector = r1_bio->sector;
2404 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2406 bio_trim(wbio, sector - r1_bio->sector, sectors);
2407 wbio->bi_iter.bi_sector += rdev->data_offset;
2408 wbio->bi_bdev = rdev->bdev;
2410 if (submit_bio_wait(wbio) < 0)
2412 ok = rdev_set_badblocks(rdev, sector,
2417 sect_to_write -= sectors;
2419 sectors = block_sectors;
2424 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2427 int s = r1_bio->sectors;
2428 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2429 struct md_rdev *rdev = conf->mirrors[m].rdev;
2430 struct bio *bio = r1_bio->bios[m];
2431 if (bio->bi_end_io == NULL)
2433 if (!bio->bi_error &&
2434 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2435 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2437 if (bio->bi_error &&
2438 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2439 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2440 md_error(conf->mddev, rdev);
2444 md_done_sync(conf->mddev, s, 1);
2447 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2452 for (m = 0; m < conf->raid_disks * 2 ; m++)
2453 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2454 struct md_rdev *rdev = conf->mirrors[m].rdev;
2455 rdev_clear_badblocks(rdev,
2457 r1_bio->sectors, 0);
2458 rdev_dec_pending(rdev, conf->mddev);
2459 } else if (r1_bio->bios[m] != NULL) {
2460 /* This drive got a write error. We need to
2461 * narrow down and record precise write
2465 if (!narrow_write_error(r1_bio, m)) {
2466 md_error(conf->mddev,
2467 conf->mirrors[m].rdev);
2468 /* an I/O failed, we can't clear the bitmap */
2469 set_bit(R1BIO_Degraded, &r1_bio->state);
2471 rdev_dec_pending(conf->mirrors[m].rdev,
2475 spin_lock_irq(&conf->device_lock);
2476 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2477 idx = sector_to_idx(r1_bio->sector);
2478 atomic_inc(&conf->nr_queued[idx]);
2479 spin_unlock_irq(&conf->device_lock);
2481 * In case freeze_array() is waiting for condition
2482 * get_unqueued_pending() == extra to be true.
2484 wake_up(&conf->wait_barrier);
2485 md_wakeup_thread(conf->mddev->thread);
2487 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2488 close_write(r1_bio);
2489 raid_end_bio_io(r1_bio);
2493 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2497 struct mddev *mddev = conf->mddev;
2499 char b[BDEVNAME_SIZE];
2500 struct md_rdev *rdev;
2502 sector_t bio_sector;
2504 clear_bit(R1BIO_ReadError, &r1_bio->state);
2505 /* we got a read error. Maybe the drive is bad. Maybe just
2506 * the block and we can fix it.
2507 * We freeze all other IO, and try reading the block from
2508 * other devices. When we find one, we re-write
2509 * and check it that fixes the read error.
2510 * This is all done synchronously while the array is
2514 bio = r1_bio->bios[r1_bio->read_disk];
2515 bdevname(bio->bi_bdev, b);
2516 bio_dev = bio->bi_bdev->bd_dev;
2517 bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2519 r1_bio->bios[r1_bio->read_disk] = NULL;
2521 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2523 && !test_bit(FailFast, &rdev->flags)) {
2524 freeze_array(conf, 1);
2525 fix_read_error(conf, r1_bio->read_disk,
2526 r1_bio->sector, r1_bio->sectors);
2527 unfreeze_array(conf);
2529 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2532 rdev_dec_pending(rdev, conf->mddev);
2535 disk = read_balance(conf, r1_bio, &max_sectors);
2537 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2538 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2539 raid_end_bio_io(r1_bio);
2541 const unsigned long do_sync
2542 = r1_bio->master_bio->bi_opf & REQ_SYNC;
2543 r1_bio->read_disk = disk;
2544 bio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2546 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2548 r1_bio->bios[r1_bio->read_disk] = bio;
2549 rdev = conf->mirrors[disk].rdev;
2550 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
2552 (unsigned long long)r1_bio->sector,
2553 bdevname(rdev->bdev, b));
2554 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2555 bio->bi_bdev = rdev->bdev;
2556 bio->bi_end_io = raid1_end_read_request;
2557 bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
2558 if (test_bit(FailFast, &rdev->flags) &&
2559 test_bit(R1BIO_FailFast, &r1_bio->state))
2560 bio->bi_opf |= MD_FAILFAST;
2561 bio->bi_private = r1_bio;
2562 if (max_sectors < r1_bio->sectors) {
2563 /* Drat - have to split this up more */
2564 struct bio *mbio = r1_bio->master_bio;
2565 int sectors_handled = (r1_bio->sector + max_sectors
2566 - mbio->bi_iter.bi_sector);
2567 r1_bio->sectors = max_sectors;
2568 bio_inc_remaining(mbio);
2569 trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
2570 bio, bio_dev, bio_sector);
2571 generic_make_request(bio);
2574 r1_bio = alloc_r1bio(mddev, mbio, sectors_handled);
2575 set_bit(R1BIO_ReadError, &r1_bio->state);
2576 inc_pending(conf, r1_bio->sector);
2580 trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
2581 bio, bio_dev, bio_sector);
2582 generic_make_request(bio);
2587 static void raid1d(struct md_thread *thread)
2589 struct mddev *mddev = thread->mddev;
2590 struct r1bio *r1_bio;
2591 unsigned long flags;
2592 struct r1conf *conf = mddev->private;
2593 struct list_head *head = &conf->retry_list;
2594 struct blk_plug plug;
2597 md_check_recovery(mddev);
2599 if (!list_empty_careful(&conf->bio_end_io_list) &&
2600 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2602 spin_lock_irqsave(&conf->device_lock, flags);
2603 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2604 list_splice_init(&conf->bio_end_io_list, &tmp);
2605 spin_unlock_irqrestore(&conf->device_lock, flags);
2606 while (!list_empty(&tmp)) {
2607 r1_bio = list_first_entry(&tmp, struct r1bio,
2609 list_del(&r1_bio->retry_list);
2610 idx = sector_to_idx(r1_bio->sector);
2611 atomic_dec(&conf->nr_queued[idx]);
2612 if (mddev->degraded)
2613 set_bit(R1BIO_Degraded, &r1_bio->state);
2614 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2615 close_write(r1_bio);
2616 raid_end_bio_io(r1_bio);
2620 blk_start_plug(&plug);
2623 flush_pending_writes(conf);
2625 spin_lock_irqsave(&conf->device_lock, flags);
2626 if (list_empty(head)) {
2627 spin_unlock_irqrestore(&conf->device_lock, flags);
2630 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2631 list_del(head->prev);
2632 idx = sector_to_idx(r1_bio->sector);
2633 atomic_dec(&conf->nr_queued[idx]);
2634 spin_unlock_irqrestore(&conf->device_lock, flags);
2636 mddev = r1_bio->mddev;
2637 conf = mddev->private;
2638 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2639 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2640 test_bit(R1BIO_WriteError, &r1_bio->state))
2641 handle_sync_write_finished(conf, r1_bio);
2643 sync_request_write(mddev, r1_bio);
2644 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2645 test_bit(R1BIO_WriteError, &r1_bio->state))
2646 handle_write_finished(conf, r1_bio);
2647 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2648 handle_read_error(conf, r1_bio);
2650 /* just a partial read to be scheduled from separate
2653 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2656 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2657 md_check_recovery(mddev);
2659 blk_finish_plug(&plug);
2662 static int init_resync(struct r1conf *conf)
2666 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2667 BUG_ON(conf->r1buf_pool);
2668 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2670 if (!conf->r1buf_pool)
2676 * perform a "sync" on one "block"
2678 * We need to make sure that no normal I/O request - particularly write
2679 * requests - conflict with active sync requests.
2681 * This is achieved by tracking pending requests and a 'barrier' concept
2682 * that can be installed to exclude normal IO requests.
2685 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2688 struct r1conf *conf = mddev->private;
2689 struct r1bio *r1_bio;
2691 sector_t max_sector, nr_sectors;
2695 int write_targets = 0, read_targets = 0;
2696 sector_t sync_blocks;
2697 int still_degraded = 0;
2698 int good_sectors = RESYNC_SECTORS;
2699 int min_bad = 0; /* number of sectors that are bad in all devices */
2700 int idx = sector_to_idx(sector_nr);
2702 if (!conf->r1buf_pool)
2703 if (init_resync(conf))
2706 max_sector = mddev->dev_sectors;
2707 if (sector_nr >= max_sector) {
2708 /* If we aborted, we need to abort the
2709 * sync on the 'current' bitmap chunk (there will
2710 * only be one in raid1 resync.
2711 * We can find the current addess in mddev->curr_resync
2713 if (mddev->curr_resync < max_sector) /* aborted */
2714 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2716 else /* completed sync */
2719 bitmap_close_sync(mddev->bitmap);
2722 if (mddev_is_clustered(mddev)) {
2723 conf->cluster_sync_low = 0;
2724 conf->cluster_sync_high = 0;
2729 if (mddev->bitmap == NULL &&
2730 mddev->recovery_cp == MaxSector &&
2731 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2732 conf->fullsync == 0) {
2734 return max_sector - sector_nr;
2736 /* before building a request, check if we can skip these blocks..
2737 * This call the bitmap_start_sync doesn't actually record anything
2739 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2740 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2741 /* We can skip this block, and probably several more */
2747 * If there is non-resync activity waiting for a turn, then let it
2748 * though before starting on this new sync request.
2750 if (atomic_read(&conf->nr_waiting[idx]))
2751 schedule_timeout_uninterruptible(1);
2753 /* we are incrementing sector_nr below. To be safe, we check against
2754 * sector_nr + two times RESYNC_SECTORS
2757 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2758 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2759 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2761 raise_barrier(conf, sector_nr);
2765 * If we get a correctably read error during resync or recovery,
2766 * we might want to read from a different device. So we
2767 * flag all drives that could conceivably be read from for READ,
2768 * and any others (which will be non-In_sync devices) for WRITE.
2769 * If a read fails, we try reading from something else for which READ
2773 r1_bio->mddev = mddev;
2774 r1_bio->sector = sector_nr;
2776 set_bit(R1BIO_IsSync, &r1_bio->state);
2777 /* make sure good_sectors won't go across barrier unit boundary */
2778 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2780 for (i = 0; i < conf->raid_disks * 2; i++) {
2781 struct md_rdev *rdev;
2782 bio = r1_bio->bios[i];
2784 rdev = rcu_dereference(conf->mirrors[i].rdev);
2786 test_bit(Faulty, &rdev->flags)) {
2787 if (i < conf->raid_disks)
2789 } else if (!test_bit(In_sync, &rdev->flags)) {
2790 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2791 bio->bi_end_io = end_sync_write;
2794 /* may need to read from here */
2795 sector_t first_bad = MaxSector;
2798 if (is_badblock(rdev, sector_nr, good_sectors,
2799 &first_bad, &bad_sectors)) {
2800 if (first_bad > sector_nr)
2801 good_sectors = first_bad - sector_nr;
2803 bad_sectors -= (sector_nr - first_bad);
2805 min_bad > bad_sectors)
2806 min_bad = bad_sectors;
2809 if (sector_nr < first_bad) {
2810 if (test_bit(WriteMostly, &rdev->flags)) {
2817 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2818 bio->bi_end_io = end_sync_read;
2820 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2821 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2822 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2824 * The device is suitable for reading (InSync),
2825 * but has bad block(s) here. Let's try to correct them,
2826 * if we are doing resync or repair. Otherwise, leave
2827 * this device alone for this sync request.
2829 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2830 bio->bi_end_io = end_sync_write;
2834 if (bio->bi_end_io) {
2835 atomic_inc(&rdev->nr_pending);
2836 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2837 bio->bi_bdev = rdev->bdev;
2838 if (test_bit(FailFast, &rdev->flags))
2839 bio->bi_opf |= MD_FAILFAST;
2845 r1_bio->read_disk = disk;
2847 if (read_targets == 0 && min_bad > 0) {
2848 /* These sectors are bad on all InSync devices, so we
2849 * need to mark them bad on all write targets
2852 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2853 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2854 struct md_rdev *rdev = conf->mirrors[i].rdev;
2855 ok = rdev_set_badblocks(rdev, sector_nr,
2859 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2864 /* Cannot record the badblocks, so need to
2866 * If there are multiple read targets, could just
2867 * fail the really bad ones ???
2869 conf->recovery_disabled = mddev->recovery_disabled;
2870 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2876 if (min_bad > 0 && min_bad < good_sectors) {
2877 /* only resync enough to reach the next bad->good
2879 good_sectors = min_bad;
2882 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2883 /* extra read targets are also write targets */
2884 write_targets += read_targets-1;
2886 if (write_targets == 0 || read_targets == 0) {
2887 /* There is nowhere to write, so all non-sync
2888 * drives must be failed - so we are finished
2892 max_sector = sector_nr + min_bad;
2893 rv = max_sector - sector_nr;
2899 if (max_sector > mddev->resync_max)
2900 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2901 if (max_sector > sector_nr + good_sectors)
2902 max_sector = sector_nr + good_sectors;
2907 int len = PAGE_SIZE;
2908 if (sector_nr + (len>>9) > max_sector)
2909 len = (max_sector - sector_nr) << 9;
2912 if (sync_blocks == 0) {
2913 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2914 &sync_blocks, still_degraded) &&
2916 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2918 if ((len >> 9) > sync_blocks)
2919 len = sync_blocks<<9;
2922 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2923 struct resync_pages *rp;
2925 bio = r1_bio->bios[i];
2926 rp = get_resync_pages(bio);
2927 if (bio->bi_end_io) {
2928 page = resync_fetch_page(rp, rp->idx++);
2931 * won't fail because the vec table is big
2932 * enough to hold all these pages
2934 bio_add_page(bio, page, len, 0);
2937 nr_sectors += len>>9;
2938 sector_nr += len>>9;
2939 sync_blocks -= (len>>9);
2940 } while (get_resync_pages(r1_bio->bios[disk]->bi_private)->idx < RESYNC_PAGES);
2942 r1_bio->sectors = nr_sectors;
2944 if (mddev_is_clustered(mddev) &&
2945 conf->cluster_sync_high < sector_nr + nr_sectors) {
2946 conf->cluster_sync_low = mddev->curr_resync_completed;
2947 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2948 /* Send resync message */
2949 md_cluster_ops->resync_info_update(mddev,
2950 conf->cluster_sync_low,
2951 conf->cluster_sync_high);
2954 /* For a user-requested sync, we read all readable devices and do a
2957 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2958 atomic_set(&r1_bio->remaining, read_targets);
2959 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2960 bio = r1_bio->bios[i];
2961 if (bio->bi_end_io == end_sync_read) {
2963 md_sync_acct(bio->bi_bdev, nr_sectors);
2964 if (read_targets == 1)
2965 bio->bi_opf &= ~MD_FAILFAST;
2966 generic_make_request(bio);
2970 atomic_set(&r1_bio->remaining, 1);
2971 bio = r1_bio->bios[r1_bio->read_disk];
2972 md_sync_acct(bio->bi_bdev, nr_sectors);
2973 if (read_targets == 1)
2974 bio->bi_opf &= ~MD_FAILFAST;
2975 generic_make_request(bio);
2981 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2986 return mddev->dev_sectors;
2989 static struct r1conf *setup_conf(struct mddev *mddev)
2991 struct r1conf *conf;
2993 struct raid1_info *disk;
2994 struct md_rdev *rdev;
2997 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
3001 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
3002 sizeof(atomic_t), GFP_KERNEL);
3003 if (!conf->nr_pending)
3006 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
3007 sizeof(atomic_t), GFP_KERNEL);
3008 if (!conf->nr_waiting)
3011 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
3012 sizeof(atomic_t), GFP_KERNEL);
3013 if (!conf->nr_queued)
3016 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
3017 sizeof(atomic_t), GFP_KERNEL);
3021 conf->mirrors = kzalloc(sizeof(struct raid1_info)
3022 * mddev->raid_disks * 2,
3027 conf->tmppage = alloc_page(GFP_KERNEL);
3031 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3032 if (!conf->poolinfo)
3034 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3035 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3038 if (!conf->r1bio_pool)
3041 conf->poolinfo->mddev = mddev;
3044 spin_lock_init(&conf->device_lock);
3045 rdev_for_each(rdev, mddev) {
3046 struct request_queue *q;
3047 int disk_idx = rdev->raid_disk;
3048 if (disk_idx >= mddev->raid_disks
3051 if (test_bit(Replacement, &rdev->flags))
3052 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3054 disk = conf->mirrors + disk_idx;
3059 q = bdev_get_queue(rdev->bdev);
3061 disk->head_position = 0;
3062 disk->seq_start = MaxSector;
3064 conf->raid_disks = mddev->raid_disks;
3065 conf->mddev = mddev;
3066 INIT_LIST_HEAD(&conf->retry_list);
3067 INIT_LIST_HEAD(&conf->bio_end_io_list);
3069 spin_lock_init(&conf->resync_lock);
3070 init_waitqueue_head(&conf->wait_barrier);
3072 bio_list_init(&conf->pending_bio_list);
3073 conf->pending_count = 0;
3074 conf->recovery_disabled = mddev->recovery_disabled - 1;
3077 for (i = 0; i < conf->raid_disks * 2; i++) {
3079 disk = conf->mirrors + i;
3081 if (i < conf->raid_disks &&
3082 disk[conf->raid_disks].rdev) {
3083 /* This slot has a replacement. */
3085 /* No original, just make the replacement
3086 * a recovering spare
3089 disk[conf->raid_disks].rdev;
3090 disk[conf->raid_disks].rdev = NULL;
3091 } else if (!test_bit(In_sync, &disk->rdev->flags))
3092 /* Original is not in_sync - bad */
3097 !test_bit(In_sync, &disk->rdev->flags)) {
3098 disk->head_position = 0;
3100 (disk->rdev->saved_raid_disk < 0))
3106 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3114 mempool_destroy(conf->r1bio_pool);
3115 kfree(conf->mirrors);
3116 safe_put_page(conf->tmppage);
3117 kfree(conf->poolinfo);
3118 kfree(conf->nr_pending);
3119 kfree(conf->nr_waiting);
3120 kfree(conf->nr_queued);
3121 kfree(conf->barrier);
3124 return ERR_PTR(err);
3127 static void raid1_free(struct mddev *mddev, void *priv);
3128 static int raid1_run(struct mddev *mddev)
3130 struct r1conf *conf;
3132 struct md_rdev *rdev;
3134 bool discard_supported = false;
3136 if (mddev->level != 1) {
3137 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3138 mdname(mddev), mddev->level);
3141 if (mddev->reshape_position != MaxSector) {
3142 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3147 * copy the already verified devices into our private RAID1
3148 * bookkeeping area. [whatever we allocate in run(),
3149 * should be freed in raid1_free()]
3151 if (mddev->private == NULL)
3152 conf = setup_conf(mddev);
3154 conf = mddev->private;
3157 return PTR_ERR(conf);
3160 blk_queue_max_write_same_sectors(mddev->queue, 0);
3162 rdev_for_each(rdev, mddev) {
3163 if (!mddev->gendisk)
3165 disk_stack_limits(mddev->gendisk, rdev->bdev,
3166 rdev->data_offset << 9);
3167 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3168 discard_supported = true;
3171 mddev->degraded = 0;
3172 for (i=0; i < conf->raid_disks; i++)
3173 if (conf->mirrors[i].rdev == NULL ||
3174 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3175 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3178 if (conf->raid_disks - mddev->degraded == 1)
3179 mddev->recovery_cp = MaxSector;
3181 if (mddev->recovery_cp != MaxSector)
3182 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3184 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3185 mdname(mddev), mddev->raid_disks - mddev->degraded,
3189 * Ok, everything is just fine now
3191 mddev->thread = conf->thread;
3192 conf->thread = NULL;
3193 mddev->private = conf;
3194 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3196 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3199 if (discard_supported)
3200 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3203 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3207 ret = md_integrity_register(mddev);
3209 md_unregister_thread(&mddev->thread);
3210 raid1_free(mddev, conf);
3215 static void raid1_free(struct mddev *mddev, void *priv)
3217 struct r1conf *conf = priv;
3219 mempool_destroy(conf->r1bio_pool);
3220 kfree(conf->mirrors);
3221 safe_put_page(conf->tmppage);
3222 kfree(conf->poolinfo);
3223 kfree(conf->nr_pending);
3224 kfree(conf->nr_waiting);
3225 kfree(conf->nr_queued);
3226 kfree(conf->barrier);
3230 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3232 /* no resync is happening, and there is enough space
3233 * on all devices, so we can resize.
3234 * We need to make sure resync covers any new space.
3235 * If the array is shrinking we should possibly wait until
3236 * any io in the removed space completes, but it hardly seems
3239 sector_t newsize = raid1_size(mddev, sectors, 0);
3240 if (mddev->external_size &&
3241 mddev->array_sectors > newsize)
3243 if (mddev->bitmap) {
3244 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3248 md_set_array_sectors(mddev, newsize);
3249 if (sectors > mddev->dev_sectors &&
3250 mddev->recovery_cp > mddev->dev_sectors) {
3251 mddev->recovery_cp = mddev->dev_sectors;
3252 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3254 mddev->dev_sectors = sectors;
3255 mddev->resync_max_sectors = sectors;
3259 static int raid1_reshape(struct mddev *mddev)
3262 * 1/ resize the r1bio_pool
3263 * 2/ resize conf->mirrors
3265 * We allocate a new r1bio_pool if we can.
3266 * Then raise a device barrier and wait until all IO stops.
3267 * Then resize conf->mirrors and swap in the new r1bio pool.
3269 * At the same time, we "pack" the devices so that all the missing
3270 * devices have the higher raid_disk numbers.
3272 mempool_t *newpool, *oldpool;
3273 struct pool_info *newpoolinfo;
3274 struct raid1_info *newmirrors;
3275 struct r1conf *conf = mddev->private;
3276 int cnt, raid_disks;
3277 unsigned long flags;
3280 /* Cannot change chunk_size, layout, or level */
3281 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3282 mddev->layout != mddev->new_layout ||
3283 mddev->level != mddev->new_level) {
3284 mddev->new_chunk_sectors = mddev->chunk_sectors;
3285 mddev->new_layout = mddev->layout;
3286 mddev->new_level = mddev->level;
3290 if (!mddev_is_clustered(mddev)) {
3291 err = md_allow_write(mddev);
3296 raid_disks = mddev->raid_disks + mddev->delta_disks;
3298 if (raid_disks < conf->raid_disks) {
3300 for (d= 0; d < conf->raid_disks; d++)
3301 if (conf->mirrors[d].rdev)
3303 if (cnt > raid_disks)
3307 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3310 newpoolinfo->mddev = mddev;
3311 newpoolinfo->raid_disks = raid_disks * 2;
3313 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3314 r1bio_pool_free, newpoolinfo);
3319 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3323 mempool_destroy(newpool);
3327 freeze_array(conf, 0);
3329 /* ok, everything is stopped */
3330 oldpool = conf->r1bio_pool;
3331 conf->r1bio_pool = newpool;
3333 for (d = d2 = 0; d < conf->raid_disks; d++) {
3334 struct md_rdev *rdev = conf->mirrors[d].rdev;
3335 if (rdev && rdev->raid_disk != d2) {
3336 sysfs_unlink_rdev(mddev, rdev);
3337 rdev->raid_disk = d2;
3338 sysfs_unlink_rdev(mddev, rdev);
3339 if (sysfs_link_rdev(mddev, rdev))
3340 pr_warn("md/raid1:%s: cannot register rd%d\n",
3341 mdname(mddev), rdev->raid_disk);
3344 newmirrors[d2++].rdev = rdev;
3346 kfree(conf->mirrors);
3347 conf->mirrors = newmirrors;
3348 kfree(conf->poolinfo);
3349 conf->poolinfo = newpoolinfo;
3351 spin_lock_irqsave(&conf->device_lock, flags);
3352 mddev->degraded += (raid_disks - conf->raid_disks);
3353 spin_unlock_irqrestore(&conf->device_lock, flags);
3354 conf->raid_disks = mddev->raid_disks = raid_disks;
3355 mddev->delta_disks = 0;
3357 unfreeze_array(conf);
3359 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3360 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3361 md_wakeup_thread(mddev->thread);
3363 mempool_destroy(oldpool);
3367 static void raid1_quiesce(struct mddev *mddev, int state)
3369 struct r1conf *conf = mddev->private;
3372 case 2: /* wake for suspend */
3373 wake_up(&conf->wait_barrier);
3376 freeze_array(conf, 0);
3379 unfreeze_array(conf);
3384 static void *raid1_takeover(struct mddev *mddev)
3386 /* raid1 can take over:
3387 * raid5 with 2 devices, any layout or chunk size
3389 if (mddev->level == 5 && mddev->raid_disks == 2) {
3390 struct r1conf *conf;
3391 mddev->new_level = 1;
3392 mddev->new_layout = 0;
3393 mddev->new_chunk_sectors = 0;
3394 conf = setup_conf(mddev);
3395 if (!IS_ERR(conf)) {
3396 /* Array must appear to be quiesced */
3397 conf->array_frozen = 1;
3398 mddev_clear_unsupported_flags(mddev,
3399 UNSUPPORTED_MDDEV_FLAGS);
3403 return ERR_PTR(-EINVAL);
3406 static struct md_personality raid1_personality =
3410 .owner = THIS_MODULE,
3411 .make_request = raid1_make_request,
3414 .status = raid1_status,
3415 .error_handler = raid1_error,
3416 .hot_add_disk = raid1_add_disk,
3417 .hot_remove_disk= raid1_remove_disk,
3418 .spare_active = raid1_spare_active,
3419 .sync_request = raid1_sync_request,
3420 .resize = raid1_resize,
3422 .check_reshape = raid1_reshape,
3423 .quiesce = raid1_quiesce,
3424 .takeover = raid1_takeover,
3425 .congested = raid1_congested,
3428 static int __init raid_init(void)
3430 return register_md_personality(&raid1_personality);
3433 static void raid_exit(void)
3435 unregister_md_personality(&raid1_personality);
3438 module_init(raid_init);
3439 module_exit(raid_exit);
3440 MODULE_LICENSE("GPL");
3441 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3442 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3443 MODULE_ALIAS("md-raid1");
3444 MODULE_ALIAS("md-level-1");
3446 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);