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>
41 #include <trace/events/block.h>
45 #include "md-bitmap.h"
47 #define UNSUPPORTED_MDDEV_FLAGS \
48 ((1L << MD_HAS_JOURNAL) | \
49 (1L << MD_JOURNAL_CLEAN) | \
50 (1L << MD_HAS_PPL) | \
51 (1L << MD_HAS_MULTIPLE_PPLS))
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)
87 * for resync bio, r1bio pointer can be retrieved from the per-bio
88 * 'struct resync_pages'.
90 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
92 return get_resync_pages(bio)->raid_bio;
95 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
97 struct pool_info *pi = data;
98 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
100 /* allocate a r1bio with room for raid_disks entries in the bios array */
101 return kzalloc(size, gfp_flags);
104 static void r1bio_pool_free(void *r1_bio, void *data)
109 #define RESYNC_DEPTH 32
110 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
111 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
112 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
113 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
114 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
116 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
118 struct pool_info *pi = data;
119 struct r1bio *r1_bio;
123 struct resync_pages *rps;
125 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
129 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
135 * Allocate bios : 1 for reading, n-1 for writing
137 for (j = pi->raid_disks ; j-- ; ) {
138 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
141 r1_bio->bios[j] = bio;
144 * Allocate RESYNC_PAGES data pages and attach them to
146 * If this is a user-requested check/repair, allocate
147 * RESYNC_PAGES for each bio.
149 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
150 need_pages = pi->raid_disks;
153 for (j = 0; j < pi->raid_disks; j++) {
154 struct resync_pages *rp = &rps[j];
156 bio = r1_bio->bios[j];
158 if (j < need_pages) {
159 if (resync_alloc_pages(rp, gfp_flags))
162 memcpy(rp, &rps[0], sizeof(*rp));
163 resync_get_all_pages(rp);
166 rp->raid_bio = r1_bio;
167 bio->bi_private = rp;
170 r1_bio->master_bio = NULL;
176 resync_free_pages(&rps[j]);
179 while (++j < pi->raid_disks)
180 bio_put(r1_bio->bios[j]);
184 r1bio_pool_free(r1_bio, data);
188 static void r1buf_pool_free(void *__r1_bio, void *data)
190 struct pool_info *pi = data;
192 struct r1bio *r1bio = __r1_bio;
193 struct resync_pages *rp = NULL;
195 for (i = pi->raid_disks; i--; ) {
196 rp = get_resync_pages(r1bio->bios[i]);
197 resync_free_pages(rp);
198 bio_put(r1bio->bios[i]);
201 /* resync pages array stored in the 1st bio's .bi_private */
204 r1bio_pool_free(r1bio, data);
207 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
211 for (i = 0; i < conf->raid_disks * 2; i++) {
212 struct bio **bio = r1_bio->bios + i;
213 if (!BIO_SPECIAL(*bio))
219 static void free_r1bio(struct r1bio *r1_bio)
221 struct r1conf *conf = r1_bio->mddev->private;
223 put_all_bios(conf, r1_bio);
224 mempool_free(r1_bio, &conf->r1bio_pool);
227 static void put_buf(struct r1bio *r1_bio)
229 struct r1conf *conf = r1_bio->mddev->private;
230 sector_t sect = r1_bio->sector;
233 for (i = 0; i < conf->raid_disks * 2; i++) {
234 struct bio *bio = r1_bio->bios[i];
236 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
239 mempool_free(r1_bio, &conf->r1buf_pool);
241 lower_barrier(conf, sect);
244 static void reschedule_retry(struct r1bio *r1_bio)
247 struct mddev *mddev = r1_bio->mddev;
248 struct r1conf *conf = mddev->private;
251 idx = sector_to_idx(r1_bio->sector);
252 spin_lock_irqsave(&conf->device_lock, flags);
253 list_add(&r1_bio->retry_list, &conf->retry_list);
254 atomic_inc(&conf->nr_queued[idx]);
255 spin_unlock_irqrestore(&conf->device_lock, flags);
257 wake_up(&conf->wait_barrier);
258 md_wakeup_thread(mddev->thread);
262 * raid_end_bio_io() is called when we have finished servicing a mirrored
263 * operation and are ready to return a success/failure code to the buffer
266 static void call_bio_endio(struct r1bio *r1_bio)
268 struct bio *bio = r1_bio->master_bio;
269 struct r1conf *conf = r1_bio->mddev->private;
271 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
272 bio->bi_status = BLK_STS_IOERR;
276 * Wake up any possible resync thread that waits for the device
279 allow_barrier(conf, r1_bio->sector);
282 static void raid_end_bio_io(struct r1bio *r1_bio)
284 struct bio *bio = r1_bio->master_bio;
286 /* if nobody has done the final endio yet, do it now */
287 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
288 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
289 (bio_data_dir(bio) == WRITE) ? "write" : "read",
290 (unsigned long long) bio->bi_iter.bi_sector,
291 (unsigned long long) bio_end_sector(bio) - 1);
293 call_bio_endio(r1_bio);
299 * Update disk head position estimator based on IRQ completion info.
301 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
303 struct r1conf *conf = r1_bio->mddev->private;
305 conf->mirrors[disk].head_position =
306 r1_bio->sector + (r1_bio->sectors);
310 * Find the disk number which triggered given bio
312 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
315 struct r1conf *conf = r1_bio->mddev->private;
316 int raid_disks = conf->raid_disks;
318 for (mirror = 0; mirror < raid_disks * 2; mirror++)
319 if (r1_bio->bios[mirror] == bio)
322 BUG_ON(mirror == raid_disks * 2);
323 update_head_pos(mirror, r1_bio);
328 static void raid1_end_read_request(struct bio *bio)
330 int uptodate = !bio->bi_status;
331 struct r1bio *r1_bio = bio->bi_private;
332 struct r1conf *conf = r1_bio->mddev->private;
333 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
336 * this branch is our 'one mirror IO has finished' event handler:
338 update_head_pos(r1_bio->read_disk, r1_bio);
341 set_bit(R1BIO_Uptodate, &r1_bio->state);
342 else if (test_bit(FailFast, &rdev->flags) &&
343 test_bit(R1BIO_FailFast, &r1_bio->state))
344 /* This was a fail-fast read so we definitely
348 /* If all other devices have failed, we want to return
349 * the error upwards rather than fail the last device.
350 * Here we redefine "uptodate" to mean "Don't want to retry"
353 spin_lock_irqsave(&conf->device_lock, flags);
354 if (r1_bio->mddev->degraded == conf->raid_disks ||
355 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
356 test_bit(In_sync, &rdev->flags)))
358 spin_unlock_irqrestore(&conf->device_lock, flags);
362 raid_end_bio_io(r1_bio);
363 rdev_dec_pending(rdev, conf->mddev);
368 char b[BDEVNAME_SIZE];
369 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
371 bdevname(rdev->bdev, b),
372 (unsigned long long)r1_bio->sector);
373 set_bit(R1BIO_ReadError, &r1_bio->state);
374 reschedule_retry(r1_bio);
375 /* don't drop the reference on read_disk yet */
379 static void close_write(struct r1bio *r1_bio)
381 /* it really is the end of this request */
382 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
383 bio_free_pages(r1_bio->behind_master_bio);
384 bio_put(r1_bio->behind_master_bio);
385 r1_bio->behind_master_bio = NULL;
387 /* clear the bitmap if all writes complete successfully */
388 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
390 !test_bit(R1BIO_Degraded, &r1_bio->state),
391 test_bit(R1BIO_BehindIO, &r1_bio->state));
392 md_write_end(r1_bio->mddev);
395 static void r1_bio_write_done(struct r1bio *r1_bio)
397 if (!atomic_dec_and_test(&r1_bio->remaining))
400 if (test_bit(R1BIO_WriteError, &r1_bio->state))
401 reschedule_retry(r1_bio);
404 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
405 reschedule_retry(r1_bio);
407 raid_end_bio_io(r1_bio);
411 static void raid1_end_write_request(struct bio *bio)
413 struct r1bio *r1_bio = bio->bi_private;
414 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
415 struct r1conf *conf = r1_bio->mddev->private;
416 struct bio *to_put = NULL;
417 int mirror = find_bio_disk(r1_bio, bio);
418 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
421 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
424 * 'one mirror IO has finished' event handler:
426 if (bio->bi_status && !discard_error) {
427 set_bit(WriteErrorSeen, &rdev->flags);
428 if (!test_and_set_bit(WantReplacement, &rdev->flags))
429 set_bit(MD_RECOVERY_NEEDED, &
430 conf->mddev->recovery);
432 if (test_bit(FailFast, &rdev->flags) &&
433 (bio->bi_opf & MD_FAILFAST) &&
434 /* We never try FailFast to WriteMostly devices */
435 !test_bit(WriteMostly, &rdev->flags)) {
436 md_error(r1_bio->mddev, rdev);
437 if (!test_bit(Faulty, &rdev->flags))
438 /* This is the only remaining device,
439 * We need to retry the write without
442 set_bit(R1BIO_WriteError, &r1_bio->state);
444 /* Finished with this branch */
445 r1_bio->bios[mirror] = NULL;
449 set_bit(R1BIO_WriteError, &r1_bio->state);
452 * Set R1BIO_Uptodate in our master bio, so that we
453 * will return a good error code for to the higher
454 * levels even if IO on some other mirrored buffer
457 * The 'master' represents the composite IO operation
458 * to user-side. So if something waits for IO, then it
459 * will wait for the 'master' bio.
464 r1_bio->bios[mirror] = NULL;
467 * Do not set R1BIO_Uptodate if the current device is
468 * rebuilding or Faulty. This is because we cannot use
469 * such device for properly reading the data back (we could
470 * potentially use it, if the current write would have felt
471 * before rdev->recovery_offset, but for simplicity we don't
474 if (test_bit(In_sync, &rdev->flags) &&
475 !test_bit(Faulty, &rdev->flags))
476 set_bit(R1BIO_Uptodate, &r1_bio->state);
478 /* Maybe we can clear some bad blocks. */
479 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
480 &first_bad, &bad_sectors) && !discard_error) {
481 r1_bio->bios[mirror] = IO_MADE_GOOD;
482 set_bit(R1BIO_MadeGood, &r1_bio->state);
487 if (test_bit(WriteMostly, &rdev->flags))
488 atomic_dec(&r1_bio->behind_remaining);
491 * In behind mode, we ACK the master bio once the I/O
492 * has safely reached all non-writemostly
493 * disks. Setting the Returned bit ensures that this
494 * gets done only once -- we don't ever want to return
495 * -EIO here, instead we'll wait
497 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
498 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
499 /* Maybe we can return now */
500 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
501 struct bio *mbio = r1_bio->master_bio;
502 pr_debug("raid1: behind end write sectors"
504 (unsigned long long) mbio->bi_iter.bi_sector,
505 (unsigned long long) bio_end_sector(mbio) - 1);
506 call_bio_endio(r1_bio);
510 if (r1_bio->bios[mirror] == NULL)
511 rdev_dec_pending(rdev, conf->mddev);
514 * Let's see if all mirrored write operations have finished
517 r1_bio_write_done(r1_bio);
523 static sector_t align_to_barrier_unit_end(sector_t start_sector,
528 WARN_ON(sectors == 0);
530 * len is the number of sectors from start_sector to end of the
531 * barrier unit which start_sector belongs to.
533 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
543 * This routine returns the disk from which the requested read should
544 * be done. There is a per-array 'next expected sequential IO' sector
545 * number - if this matches on the next IO then we use the last disk.
546 * There is also a per-disk 'last know head position' sector that is
547 * maintained from IRQ contexts, both the normal and the resync IO
548 * completion handlers update this position correctly. If there is no
549 * perfect sequential match then we pick the disk whose head is closest.
551 * If there are 2 mirrors in the same 2 devices, performance degrades
552 * because position is mirror, not device based.
554 * The rdev for the device selected will have nr_pending incremented.
556 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
558 const sector_t this_sector = r1_bio->sector;
560 int best_good_sectors;
561 int best_disk, best_dist_disk, best_pending_disk;
565 unsigned int min_pending;
566 struct md_rdev *rdev;
568 int choose_next_idle;
572 * Check if we can balance. We can balance on the whole
573 * device if no resync is going on, or below the resync window.
574 * We take the first readable disk when above the resync window.
577 sectors = r1_bio->sectors;
580 best_dist = MaxSector;
581 best_pending_disk = -1;
582 min_pending = UINT_MAX;
583 best_good_sectors = 0;
585 choose_next_idle = 0;
586 clear_bit(R1BIO_FailFast, &r1_bio->state);
588 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
589 (mddev_is_clustered(conf->mddev) &&
590 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
591 this_sector + sectors)))
596 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
600 unsigned int pending;
603 rdev = rcu_dereference(conf->mirrors[disk].rdev);
604 if (r1_bio->bios[disk] == IO_BLOCKED
606 || test_bit(Faulty, &rdev->flags))
608 if (!test_bit(In_sync, &rdev->flags) &&
609 rdev->recovery_offset < this_sector + sectors)
611 if (test_bit(WriteMostly, &rdev->flags)) {
612 /* Don't balance among write-mostly, just
613 * use the first as a last resort */
614 if (best_dist_disk < 0) {
615 if (is_badblock(rdev, this_sector, sectors,
616 &first_bad, &bad_sectors)) {
617 if (first_bad <= this_sector)
618 /* Cannot use this */
620 best_good_sectors = first_bad - this_sector;
622 best_good_sectors = sectors;
623 best_dist_disk = disk;
624 best_pending_disk = disk;
628 /* This is a reasonable device to use. It might
631 if (is_badblock(rdev, this_sector, sectors,
632 &first_bad, &bad_sectors)) {
633 if (best_dist < MaxSector)
634 /* already have a better device */
636 if (first_bad <= this_sector) {
637 /* cannot read here. If this is the 'primary'
638 * device, then we must not read beyond
639 * bad_sectors from another device..
641 bad_sectors -= (this_sector - first_bad);
642 if (choose_first && sectors > bad_sectors)
643 sectors = bad_sectors;
644 if (best_good_sectors > sectors)
645 best_good_sectors = sectors;
648 sector_t good_sectors = first_bad - this_sector;
649 if (good_sectors > best_good_sectors) {
650 best_good_sectors = good_sectors;
658 if ((sectors > best_good_sectors) && (best_disk >= 0))
660 best_good_sectors = sectors;
664 /* At least two disks to choose from so failfast is OK */
665 set_bit(R1BIO_FailFast, &r1_bio->state);
667 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
668 has_nonrot_disk |= nonrot;
669 pending = atomic_read(&rdev->nr_pending);
670 dist = abs(this_sector - conf->mirrors[disk].head_position);
675 /* Don't change to another disk for sequential reads */
676 if (conf->mirrors[disk].next_seq_sect == this_sector
678 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
679 struct raid1_info *mirror = &conf->mirrors[disk];
683 * If buffered sequential IO size exceeds optimal
684 * iosize, check if there is idle disk. If yes, choose
685 * the idle disk. read_balance could already choose an
686 * idle disk before noticing it's a sequential IO in
687 * this disk. This doesn't matter because this disk
688 * will idle, next time it will be utilized after the
689 * first disk has IO size exceeds optimal iosize. In
690 * this way, iosize of the first disk will be optimal
691 * iosize at least. iosize of the second disk might be
692 * small, but not a big deal since when the second disk
693 * starts IO, the first disk is likely still busy.
695 if (nonrot && opt_iosize > 0 &&
696 mirror->seq_start != MaxSector &&
697 mirror->next_seq_sect > opt_iosize &&
698 mirror->next_seq_sect - opt_iosize >=
700 choose_next_idle = 1;
706 if (choose_next_idle)
709 if (min_pending > pending) {
710 min_pending = pending;
711 best_pending_disk = disk;
714 if (dist < best_dist) {
716 best_dist_disk = disk;
721 * If all disks are rotational, choose the closest disk. If any disk is
722 * non-rotational, choose the disk with less pending request even the
723 * disk is rotational, which might/might not be optimal for raids with
724 * mixed ratation/non-rotational disks depending on workload.
726 if (best_disk == -1) {
727 if (has_nonrot_disk || min_pending == 0)
728 best_disk = best_pending_disk;
730 best_disk = best_dist_disk;
733 if (best_disk >= 0) {
734 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
737 atomic_inc(&rdev->nr_pending);
738 sectors = best_good_sectors;
740 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
741 conf->mirrors[best_disk].seq_start = this_sector;
743 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
746 *max_sectors = sectors;
751 static int raid1_congested(struct mddev *mddev, int bits)
753 struct r1conf *conf = mddev->private;
756 if ((bits & (1 << WB_async_congested)) &&
757 conf->pending_count >= max_queued_requests)
761 for (i = 0; i < conf->raid_disks * 2; i++) {
762 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
763 if (rdev && !test_bit(Faulty, &rdev->flags)) {
764 struct request_queue *q = bdev_get_queue(rdev->bdev);
768 /* Note the '|| 1' - when read_balance prefers
769 * non-congested targets, it can be removed
771 if ((bits & (1 << WB_async_congested)) || 1)
772 ret |= bdi_congested(q->backing_dev_info, bits);
774 ret &= bdi_congested(q->backing_dev_info, bits);
781 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
783 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
784 md_bitmap_unplug(conf->mddev->bitmap);
785 wake_up(&conf->wait_barrier);
787 while (bio) { /* submit pending writes */
788 struct bio *next = bio->bi_next;
789 struct md_rdev *rdev = (void *)bio->bi_disk;
791 bio_set_dev(bio, rdev->bdev);
792 if (test_bit(Faulty, &rdev->flags)) {
794 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
795 !blk_queue_discard(bio->bi_disk->queue)))
799 generic_make_request(bio);
804 static void flush_pending_writes(struct r1conf *conf)
806 /* Any writes that have been queued but are awaiting
807 * bitmap updates get flushed here.
809 spin_lock_irq(&conf->device_lock);
811 if (conf->pending_bio_list.head) {
812 struct blk_plug plug;
815 bio = bio_list_get(&conf->pending_bio_list);
816 conf->pending_count = 0;
817 spin_unlock_irq(&conf->device_lock);
820 * As this is called in a wait_event() loop (see freeze_array),
821 * current->state might be TASK_UNINTERRUPTIBLE which will
822 * cause a warning when we prepare to wait again. As it is
823 * rare that this path is taken, it is perfectly safe to force
824 * us to go around the wait_event() loop again, so the warning
825 * is a false-positive. Silence the warning by resetting
828 __set_current_state(TASK_RUNNING);
829 blk_start_plug(&plug);
830 flush_bio_list(conf, bio);
831 blk_finish_plug(&plug);
833 spin_unlock_irq(&conf->device_lock);
837 * Sometimes we need to suspend IO while we do something else,
838 * either some resync/recovery, or reconfigure the array.
839 * To do this we raise a 'barrier'.
840 * The 'barrier' is a counter that can be raised multiple times
841 * to count how many activities are happening which preclude
843 * We can only raise the barrier if there is no pending IO.
844 * i.e. if nr_pending == 0.
845 * We choose only to raise the barrier if no-one is waiting for the
846 * barrier to go down. This means that as soon as an IO request
847 * is ready, no other operations which require a barrier will start
848 * until the IO request has had a chance.
850 * So: regular IO calls 'wait_barrier'. When that returns there
851 * is no backgroup IO happening, It must arrange to call
852 * allow_barrier when it has finished its IO.
853 * backgroup IO calls must call raise_barrier. Once that returns
854 * there is no normal IO happeing. It must arrange to call
855 * lower_barrier when the particular background IO completes.
857 static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
859 int idx = sector_to_idx(sector_nr);
861 spin_lock_irq(&conf->resync_lock);
863 /* Wait until no block IO is waiting */
864 wait_event_lock_irq(conf->wait_barrier,
865 !atomic_read(&conf->nr_waiting[idx]),
868 /* block any new IO from starting */
869 atomic_inc(&conf->barrier[idx]);
871 * In raise_barrier() we firstly increase conf->barrier[idx] then
872 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
873 * increase conf->nr_pending[idx] then check conf->barrier[idx].
874 * A memory barrier here to make sure conf->nr_pending[idx] won't
875 * be fetched before conf->barrier[idx] is increased. Otherwise
876 * there will be a race between raise_barrier() and _wait_barrier().
878 smp_mb__after_atomic();
880 /* For these conditions we must wait:
881 * A: while the array is in frozen state
882 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
883 * existing in corresponding I/O barrier bucket.
884 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
885 * max resync count which allowed on current I/O barrier bucket.
887 wait_event_lock_irq(conf->wait_barrier,
888 (!conf->array_frozen &&
889 !atomic_read(&conf->nr_pending[idx]) &&
890 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
891 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
894 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
895 atomic_dec(&conf->barrier[idx]);
896 spin_unlock_irq(&conf->resync_lock);
897 wake_up(&conf->wait_barrier);
901 atomic_inc(&conf->nr_sync_pending);
902 spin_unlock_irq(&conf->resync_lock);
907 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
909 int idx = sector_to_idx(sector_nr);
911 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
913 atomic_dec(&conf->barrier[idx]);
914 atomic_dec(&conf->nr_sync_pending);
915 wake_up(&conf->wait_barrier);
918 static void _wait_barrier(struct r1conf *conf, int idx)
921 * We need to increase conf->nr_pending[idx] very early here,
922 * then raise_barrier() can be blocked when it waits for
923 * conf->nr_pending[idx] to be 0. Then we can avoid holding
924 * conf->resync_lock when there is no barrier raised in same
925 * barrier unit bucket. Also if the array is frozen, I/O
926 * should be blocked until array is unfrozen.
928 atomic_inc(&conf->nr_pending[idx]);
930 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
931 * check conf->barrier[idx]. In raise_barrier() we firstly increase
932 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
933 * barrier is necessary here to make sure conf->barrier[idx] won't be
934 * fetched before conf->nr_pending[idx] is increased. Otherwise there
935 * will be a race between _wait_barrier() and raise_barrier().
937 smp_mb__after_atomic();
940 * Don't worry about checking two atomic_t variables at same time
941 * here. If during we check conf->barrier[idx], the array is
942 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
943 * 0, it is safe to return and make the I/O continue. Because the
944 * array is frozen, all I/O returned here will eventually complete
945 * or be queued, no race will happen. See code comment in
948 if (!READ_ONCE(conf->array_frozen) &&
949 !atomic_read(&conf->barrier[idx]))
953 * After holding conf->resync_lock, conf->nr_pending[idx]
954 * should be decreased before waiting for barrier to drop.
955 * Otherwise, we may encounter a race condition because
956 * raise_barrer() might be waiting for conf->nr_pending[idx]
957 * to be 0 at same time.
959 spin_lock_irq(&conf->resync_lock);
960 atomic_inc(&conf->nr_waiting[idx]);
961 atomic_dec(&conf->nr_pending[idx]);
963 * In case freeze_array() is waiting for
964 * get_unqueued_pending() == extra
966 wake_up(&conf->wait_barrier);
967 /* Wait for the barrier in same barrier unit bucket to drop. */
968 wait_event_lock_irq(conf->wait_barrier,
969 !conf->array_frozen &&
970 !atomic_read(&conf->barrier[idx]),
972 atomic_inc(&conf->nr_pending[idx]);
973 atomic_dec(&conf->nr_waiting[idx]);
974 spin_unlock_irq(&conf->resync_lock);
977 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
979 int idx = sector_to_idx(sector_nr);
982 * Very similar to _wait_barrier(). The difference is, for read
983 * I/O we don't need wait for sync I/O, but if the whole array
984 * is frozen, the read I/O still has to wait until the array is
985 * unfrozen. Since there is no ordering requirement with
986 * conf->barrier[idx] here, memory barrier is unnecessary as well.
988 atomic_inc(&conf->nr_pending[idx]);
990 if (!READ_ONCE(conf->array_frozen))
993 spin_lock_irq(&conf->resync_lock);
994 atomic_inc(&conf->nr_waiting[idx]);
995 atomic_dec(&conf->nr_pending[idx]);
997 * In case freeze_array() is waiting for
998 * get_unqueued_pending() == extra
1000 wake_up(&conf->wait_barrier);
1001 /* Wait for array to be unfrozen */
1002 wait_event_lock_irq(conf->wait_barrier,
1003 !conf->array_frozen,
1005 atomic_inc(&conf->nr_pending[idx]);
1006 atomic_dec(&conf->nr_waiting[idx]);
1007 spin_unlock_irq(&conf->resync_lock);
1010 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1012 int idx = sector_to_idx(sector_nr);
1014 _wait_barrier(conf, idx);
1017 static void _allow_barrier(struct r1conf *conf, int idx)
1019 atomic_dec(&conf->nr_pending[idx]);
1020 wake_up(&conf->wait_barrier);
1023 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1025 int idx = sector_to_idx(sector_nr);
1027 _allow_barrier(conf, idx);
1030 /* conf->resync_lock should be held */
1031 static int get_unqueued_pending(struct r1conf *conf)
1035 ret = atomic_read(&conf->nr_sync_pending);
1036 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1037 ret += atomic_read(&conf->nr_pending[idx]) -
1038 atomic_read(&conf->nr_queued[idx]);
1043 static void freeze_array(struct r1conf *conf, int extra)
1045 /* Stop sync I/O and normal I/O and wait for everything to
1047 * This is called in two situations:
1048 * 1) management command handlers (reshape, remove disk, quiesce).
1049 * 2) one normal I/O request failed.
1051 * After array_frozen is set to 1, new sync IO will be blocked at
1052 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1053 * or wait_read_barrier(). The flying I/Os will either complete or be
1054 * queued. When everything goes quite, there are only queued I/Os left.
1056 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1057 * barrier bucket index which this I/O request hits. When all sync and
1058 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1059 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1060 * in handle_read_error(), we may call freeze_array() before trying to
1061 * fix the read error. In this case, the error read I/O is not queued,
1062 * so get_unqueued_pending() == 1.
1064 * Therefore before this function returns, we need to wait until
1065 * get_unqueued_pendings(conf) gets equal to extra. For
1066 * normal I/O context, extra is 1, in rested situations extra is 0.
1068 spin_lock_irq(&conf->resync_lock);
1069 conf->array_frozen = 1;
1070 raid1_log(conf->mddev, "wait freeze");
1071 wait_event_lock_irq_cmd(
1073 get_unqueued_pending(conf) == extra,
1075 flush_pending_writes(conf));
1076 spin_unlock_irq(&conf->resync_lock);
1078 static void unfreeze_array(struct r1conf *conf)
1080 /* reverse the effect of the freeze */
1081 spin_lock_irq(&conf->resync_lock);
1082 conf->array_frozen = 0;
1083 spin_unlock_irq(&conf->resync_lock);
1084 wake_up(&conf->wait_barrier);
1087 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1090 int size = bio->bi_iter.bi_size;
1091 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1093 struct bio *behind_bio = NULL;
1095 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1099 /* discard op, we don't support writezero/writesame yet */
1100 if (!bio_has_data(bio)) {
1101 behind_bio->bi_iter.bi_size = size;
1105 behind_bio->bi_write_hint = bio->bi_write_hint;
1107 while (i < vcnt && size) {
1109 int len = min_t(int, PAGE_SIZE, size);
1111 page = alloc_page(GFP_NOIO);
1112 if (unlikely(!page))
1115 bio_add_page(behind_bio, page, len, 0);
1121 bio_copy_data(behind_bio, bio);
1123 r1_bio->behind_master_bio = behind_bio;
1124 set_bit(R1BIO_BehindIO, &r1_bio->state);
1129 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1130 bio->bi_iter.bi_size);
1131 bio_free_pages(behind_bio);
1132 bio_put(behind_bio);
1135 struct raid1_plug_cb {
1136 struct blk_plug_cb cb;
1137 struct bio_list pending;
1141 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1143 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1145 struct mddev *mddev = plug->cb.data;
1146 struct r1conf *conf = mddev->private;
1149 if (from_schedule || current->bio_list) {
1150 spin_lock_irq(&conf->device_lock);
1151 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1152 conf->pending_count += plug->pending_cnt;
1153 spin_unlock_irq(&conf->device_lock);
1154 wake_up(&conf->wait_barrier);
1155 md_wakeup_thread(mddev->thread);
1160 /* we aren't scheduling, so we can do the write-out directly. */
1161 bio = bio_list_get(&plug->pending);
1162 flush_bio_list(conf, bio);
1166 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1168 r1_bio->master_bio = bio;
1169 r1_bio->sectors = bio_sectors(bio);
1171 r1_bio->mddev = mddev;
1172 r1_bio->sector = bio->bi_iter.bi_sector;
1175 static inline struct r1bio *
1176 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1178 struct r1conf *conf = mddev->private;
1179 struct r1bio *r1_bio;
1181 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1182 /* Ensure no bio records IO_BLOCKED */
1183 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1184 init_r1bio(r1_bio, mddev, bio);
1188 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1189 int max_read_sectors, struct r1bio *r1_bio)
1191 struct r1conf *conf = mddev->private;
1192 struct raid1_info *mirror;
1193 struct bio *read_bio;
1194 struct bitmap *bitmap = mddev->bitmap;
1195 const int op = bio_op(bio);
1196 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1199 bool print_msg = !!r1_bio;
1200 char b[BDEVNAME_SIZE];
1203 * If r1_bio is set, we are blocking the raid1d thread
1204 * so there is a tiny risk of deadlock. So ask for
1205 * emergency memory if needed.
1207 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1210 /* Need to get the block device name carefully */
1211 struct md_rdev *rdev;
1213 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1215 bdevname(rdev->bdev, b);
1222 * Still need barrier for READ in case that whole
1225 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1228 r1_bio = alloc_r1bio(mddev, bio);
1230 init_r1bio(r1_bio, mddev, bio);
1231 r1_bio->sectors = max_read_sectors;
1234 * make_request() can abort the operation when read-ahead is being
1235 * used and no empty request is available.
1237 rdisk = read_balance(conf, r1_bio, &max_sectors);
1240 /* couldn't find anywhere to read from */
1242 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1245 (unsigned long long)r1_bio->sector);
1247 raid_end_bio_io(r1_bio);
1250 mirror = conf->mirrors + rdisk;
1253 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1255 (unsigned long long)r1_bio->sector,
1256 bdevname(mirror->rdev->bdev, b));
1258 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1261 * Reading from a write-mostly device must take care not to
1262 * over-take any writes that are 'behind'
1264 raid1_log(mddev, "wait behind writes");
1265 wait_event(bitmap->behind_wait,
1266 atomic_read(&bitmap->behind_writes) == 0);
1269 if (max_sectors < bio_sectors(bio)) {
1270 struct bio *split = bio_split(bio, max_sectors,
1271 gfp, &conf->bio_split);
1272 bio_chain(split, bio);
1273 generic_make_request(bio);
1275 r1_bio->master_bio = bio;
1276 r1_bio->sectors = max_sectors;
1279 r1_bio->read_disk = rdisk;
1281 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1283 r1_bio->bios[rdisk] = read_bio;
1285 read_bio->bi_iter.bi_sector = r1_bio->sector +
1286 mirror->rdev->data_offset;
1287 bio_set_dev(read_bio, mirror->rdev->bdev);
1288 read_bio->bi_end_io = raid1_end_read_request;
1289 bio_set_op_attrs(read_bio, op, do_sync);
1290 if (test_bit(FailFast, &mirror->rdev->flags) &&
1291 test_bit(R1BIO_FailFast, &r1_bio->state))
1292 read_bio->bi_opf |= MD_FAILFAST;
1293 read_bio->bi_private = r1_bio;
1296 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1297 disk_devt(mddev->gendisk), r1_bio->sector);
1299 generic_make_request(read_bio);
1302 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1303 int max_write_sectors)
1305 struct r1conf *conf = mddev->private;
1306 struct r1bio *r1_bio;
1308 struct bitmap *bitmap = mddev->bitmap;
1309 unsigned long flags;
1310 struct md_rdev *blocked_rdev;
1311 struct blk_plug_cb *cb;
1312 struct raid1_plug_cb *plug = NULL;
1316 if (mddev_is_clustered(mddev) &&
1317 md_cluster_ops->area_resyncing(mddev, WRITE,
1318 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1322 prepare_to_wait(&conf->wait_barrier,
1324 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1325 bio->bi_iter.bi_sector,
1326 bio_end_sector(bio)))
1330 finish_wait(&conf->wait_barrier, &w);
1334 * Register the new request and wait if the reconstruction
1335 * thread has put up a bar for new requests.
1336 * Continue immediately if no resync is active currently.
1338 wait_barrier(conf, bio->bi_iter.bi_sector);
1340 r1_bio = alloc_r1bio(mddev, bio);
1341 r1_bio->sectors = max_write_sectors;
1343 if (conf->pending_count >= max_queued_requests) {
1344 md_wakeup_thread(mddev->thread);
1345 raid1_log(mddev, "wait queued");
1346 wait_event(conf->wait_barrier,
1347 conf->pending_count < max_queued_requests);
1349 /* first select target devices under rcu_lock and
1350 * inc refcount on their rdev. Record them by setting
1352 * If there are known/acknowledged bad blocks on any device on
1353 * which we have seen a write error, we want to avoid writing those
1355 * This potentially requires several writes to write around
1356 * the bad blocks. Each set of writes gets it's own r1bio
1357 * with a set of bios attached.
1360 disks = conf->raid_disks * 2;
1362 blocked_rdev = NULL;
1364 max_sectors = r1_bio->sectors;
1365 for (i = 0; i < disks; i++) {
1366 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1367 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1368 atomic_inc(&rdev->nr_pending);
1369 blocked_rdev = rdev;
1372 r1_bio->bios[i] = NULL;
1373 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1374 if (i < conf->raid_disks)
1375 set_bit(R1BIO_Degraded, &r1_bio->state);
1379 atomic_inc(&rdev->nr_pending);
1380 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1385 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1386 &first_bad, &bad_sectors);
1388 /* mustn't write here until the bad block is
1390 set_bit(BlockedBadBlocks, &rdev->flags);
1391 blocked_rdev = rdev;
1394 if (is_bad && first_bad <= r1_bio->sector) {
1395 /* Cannot write here at all */
1396 bad_sectors -= (r1_bio->sector - first_bad);
1397 if (bad_sectors < max_sectors)
1398 /* mustn't write more than bad_sectors
1399 * to other devices yet
1401 max_sectors = bad_sectors;
1402 rdev_dec_pending(rdev, mddev);
1403 /* We don't set R1BIO_Degraded as that
1404 * only applies if the disk is
1405 * missing, so it might be re-added,
1406 * and we want to know to recover this
1408 * In this case the device is here,
1409 * and the fact that this chunk is not
1410 * in-sync is recorded in the bad
1416 int good_sectors = first_bad - r1_bio->sector;
1417 if (good_sectors < max_sectors)
1418 max_sectors = good_sectors;
1421 r1_bio->bios[i] = bio;
1425 if (unlikely(blocked_rdev)) {
1426 /* Wait for this device to become unblocked */
1429 for (j = 0; j < i; j++)
1430 if (r1_bio->bios[j])
1431 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1433 allow_barrier(conf, bio->bi_iter.bi_sector);
1434 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1435 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1436 wait_barrier(conf, bio->bi_iter.bi_sector);
1440 if (max_sectors < bio_sectors(bio)) {
1441 struct bio *split = bio_split(bio, max_sectors,
1442 GFP_NOIO, &conf->bio_split);
1443 bio_chain(split, bio);
1444 generic_make_request(bio);
1446 r1_bio->master_bio = bio;
1447 r1_bio->sectors = max_sectors;
1450 atomic_set(&r1_bio->remaining, 1);
1451 atomic_set(&r1_bio->behind_remaining, 0);
1455 for (i = 0; i < disks; i++) {
1456 struct bio *mbio = NULL;
1457 if (!r1_bio->bios[i])
1463 * Not if there are too many, or cannot
1464 * allocate memory, or a reader on WriteMostly
1465 * is waiting for behind writes to flush */
1467 (atomic_read(&bitmap->behind_writes)
1468 < mddev->bitmap_info.max_write_behind) &&
1469 !waitqueue_active(&bitmap->behind_wait)) {
1470 alloc_behind_master_bio(r1_bio, bio);
1473 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1474 test_bit(R1BIO_BehindIO, &r1_bio->state));
1478 if (r1_bio->behind_master_bio)
1479 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1480 GFP_NOIO, &mddev->bio_set);
1482 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1484 if (r1_bio->behind_master_bio) {
1485 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1486 atomic_inc(&r1_bio->behind_remaining);
1489 r1_bio->bios[i] = mbio;
1491 mbio->bi_iter.bi_sector = (r1_bio->sector +
1492 conf->mirrors[i].rdev->data_offset);
1493 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1494 mbio->bi_end_io = raid1_end_write_request;
1495 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1496 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1497 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1498 conf->raid_disks - mddev->degraded > 1)
1499 mbio->bi_opf |= MD_FAILFAST;
1500 mbio->bi_private = r1_bio;
1502 atomic_inc(&r1_bio->remaining);
1505 trace_block_bio_remap(mbio->bi_disk->queue,
1506 mbio, disk_devt(mddev->gendisk),
1508 /* flush_pending_writes() needs access to the rdev so...*/
1509 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1511 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1513 plug = container_of(cb, struct raid1_plug_cb, cb);
1517 bio_list_add(&plug->pending, mbio);
1518 plug->pending_cnt++;
1520 spin_lock_irqsave(&conf->device_lock, flags);
1521 bio_list_add(&conf->pending_bio_list, mbio);
1522 conf->pending_count++;
1523 spin_unlock_irqrestore(&conf->device_lock, flags);
1524 md_wakeup_thread(mddev->thread);
1528 r1_bio_write_done(r1_bio);
1530 /* In case raid1d snuck in to freeze_array */
1531 wake_up(&conf->wait_barrier);
1534 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1538 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1539 md_flush_request(mddev, bio);
1544 * There is a limit to the maximum size, but
1545 * the read/write handler might find a lower limit
1546 * due to bad blocks. To avoid multiple splits,
1547 * we pass the maximum number of sectors down
1548 * and let the lower level perform the split.
1550 sectors = align_to_barrier_unit_end(
1551 bio->bi_iter.bi_sector, bio_sectors(bio));
1553 if (bio_data_dir(bio) == READ)
1554 raid1_read_request(mddev, bio, sectors, NULL);
1556 if (!md_write_start(mddev,bio))
1558 raid1_write_request(mddev, bio, sectors);
1563 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1565 struct r1conf *conf = mddev->private;
1568 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1569 conf->raid_disks - mddev->degraded);
1571 for (i = 0; i < conf->raid_disks; i++) {
1572 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1573 seq_printf(seq, "%s",
1574 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1577 seq_printf(seq, "]");
1580 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1582 char b[BDEVNAME_SIZE];
1583 struct r1conf *conf = mddev->private;
1584 unsigned long flags;
1587 * If it is not operational, then we have already marked it as dead
1588 * else if it is the last working disks, ignore the error, let the
1589 * next level up know.
1590 * else mark the drive as failed
1592 spin_lock_irqsave(&conf->device_lock, flags);
1593 if (test_bit(In_sync, &rdev->flags)
1594 && (conf->raid_disks - mddev->degraded) == 1) {
1596 * Don't fail the drive, act as though we were just a
1597 * normal single drive.
1598 * However don't try a recovery from this drive as
1599 * it is very likely to fail.
1601 conf->recovery_disabled = mddev->recovery_disabled;
1602 spin_unlock_irqrestore(&conf->device_lock, flags);
1605 set_bit(Blocked, &rdev->flags);
1606 if (test_and_clear_bit(In_sync, &rdev->flags))
1608 set_bit(Faulty, &rdev->flags);
1609 spin_unlock_irqrestore(&conf->device_lock, flags);
1611 * if recovery is running, make sure it aborts.
1613 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1614 set_mask_bits(&mddev->sb_flags, 0,
1615 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1616 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1617 "md/raid1:%s: Operation continuing on %d devices.\n",
1618 mdname(mddev), bdevname(rdev->bdev, b),
1619 mdname(mddev), conf->raid_disks - mddev->degraded);
1622 static void print_conf(struct r1conf *conf)
1626 pr_debug("RAID1 conf printout:\n");
1628 pr_debug("(!conf)\n");
1631 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1635 for (i = 0; i < conf->raid_disks; i++) {
1636 char b[BDEVNAME_SIZE];
1637 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1639 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1640 i, !test_bit(In_sync, &rdev->flags),
1641 !test_bit(Faulty, &rdev->flags),
1642 bdevname(rdev->bdev,b));
1647 static void close_sync(struct r1conf *conf)
1651 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1652 _wait_barrier(conf, idx);
1653 _allow_barrier(conf, idx);
1656 mempool_exit(&conf->r1buf_pool);
1659 static int raid1_spare_active(struct mddev *mddev)
1662 struct r1conf *conf = mddev->private;
1664 unsigned long flags;
1667 * Find all failed disks within the RAID1 configuration
1668 * and mark them readable.
1669 * Called under mddev lock, so rcu protection not needed.
1670 * device_lock used to avoid races with raid1_end_read_request
1671 * which expects 'In_sync' flags and ->degraded to be consistent.
1673 spin_lock_irqsave(&conf->device_lock, flags);
1674 for (i = 0; i < conf->raid_disks; i++) {
1675 struct md_rdev *rdev = conf->mirrors[i].rdev;
1676 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1678 && !test_bit(Candidate, &repl->flags)
1679 && repl->recovery_offset == MaxSector
1680 && !test_bit(Faulty, &repl->flags)
1681 && !test_and_set_bit(In_sync, &repl->flags)) {
1682 /* replacement has just become active */
1684 !test_and_clear_bit(In_sync, &rdev->flags))
1687 /* Replaced device not technically
1688 * faulty, but we need to be sure
1689 * it gets removed and never re-added
1691 set_bit(Faulty, &rdev->flags);
1692 sysfs_notify_dirent_safe(
1697 && rdev->recovery_offset == MaxSector
1698 && !test_bit(Faulty, &rdev->flags)
1699 && !test_and_set_bit(In_sync, &rdev->flags)) {
1701 sysfs_notify_dirent_safe(rdev->sysfs_state);
1704 mddev->degraded -= count;
1705 spin_unlock_irqrestore(&conf->device_lock, flags);
1711 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1713 struct r1conf *conf = mddev->private;
1716 struct raid1_info *p;
1718 int last = conf->raid_disks - 1;
1720 if (mddev->recovery_disabled == conf->recovery_disabled)
1723 if (md_integrity_add_rdev(rdev, mddev))
1726 if (rdev->raid_disk >= 0)
1727 first = last = rdev->raid_disk;
1730 * find the disk ... but prefer rdev->saved_raid_disk
1733 if (rdev->saved_raid_disk >= 0 &&
1734 rdev->saved_raid_disk >= first &&
1735 rdev->saved_raid_disk < conf->raid_disks &&
1736 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1737 first = last = rdev->saved_raid_disk;
1739 for (mirror = first; mirror <= last; mirror++) {
1740 p = conf->mirrors+mirror;
1744 disk_stack_limits(mddev->gendisk, rdev->bdev,
1745 rdev->data_offset << 9);
1747 p->head_position = 0;
1748 rdev->raid_disk = mirror;
1750 /* As all devices are equivalent, we don't need a full recovery
1751 * if this was recently any drive of the array
1753 if (rdev->saved_raid_disk < 0)
1755 rcu_assign_pointer(p->rdev, rdev);
1758 if (test_bit(WantReplacement, &p->rdev->flags) &&
1759 p[conf->raid_disks].rdev == NULL) {
1760 /* Add this device as a replacement */
1761 clear_bit(In_sync, &rdev->flags);
1762 set_bit(Replacement, &rdev->flags);
1763 rdev->raid_disk = mirror;
1766 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1770 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1771 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1776 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1778 struct r1conf *conf = mddev->private;
1780 int number = rdev->raid_disk;
1781 struct raid1_info *p = conf->mirrors + number;
1783 if (rdev != p->rdev)
1784 p = conf->mirrors + conf->raid_disks + number;
1787 if (rdev == p->rdev) {
1788 if (test_bit(In_sync, &rdev->flags) ||
1789 atomic_read(&rdev->nr_pending)) {
1793 /* Only remove non-faulty devices if recovery
1796 if (!test_bit(Faulty, &rdev->flags) &&
1797 mddev->recovery_disabled != conf->recovery_disabled &&
1798 mddev->degraded < conf->raid_disks) {
1803 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1805 if (atomic_read(&rdev->nr_pending)) {
1806 /* lost the race, try later */
1812 if (conf->mirrors[conf->raid_disks + number].rdev) {
1813 /* We just removed a device that is being replaced.
1814 * Move down the replacement. We drain all IO before
1815 * doing this to avoid confusion.
1817 struct md_rdev *repl =
1818 conf->mirrors[conf->raid_disks + number].rdev;
1819 freeze_array(conf, 0);
1820 if (atomic_read(&repl->nr_pending)) {
1821 /* It means that some queued IO of retry_list
1822 * hold repl. Thus, we cannot set replacement
1823 * as NULL, avoiding rdev NULL pointer
1824 * dereference in sync_request_write and
1825 * handle_write_finished.
1828 unfreeze_array(conf);
1831 clear_bit(Replacement, &repl->flags);
1833 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1834 unfreeze_array(conf);
1837 clear_bit(WantReplacement, &rdev->flags);
1838 err = md_integrity_register(mddev);
1846 static void end_sync_read(struct bio *bio)
1848 struct r1bio *r1_bio = get_resync_r1bio(bio);
1850 update_head_pos(r1_bio->read_disk, r1_bio);
1853 * we have read a block, now it needs to be re-written,
1854 * or re-read if the read failed.
1855 * We don't do much here, just schedule handling by raid1d
1857 if (!bio->bi_status)
1858 set_bit(R1BIO_Uptodate, &r1_bio->state);
1860 if (atomic_dec_and_test(&r1_bio->remaining))
1861 reschedule_retry(r1_bio);
1864 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1866 sector_t sync_blocks = 0;
1867 sector_t s = r1_bio->sector;
1868 long sectors_to_go = r1_bio->sectors;
1870 /* make sure these bits don't get cleared. */
1872 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1874 sectors_to_go -= sync_blocks;
1875 } while (sectors_to_go > 0);
1878 static void end_sync_write(struct bio *bio)
1880 int uptodate = !bio->bi_status;
1881 struct r1bio *r1_bio = get_resync_r1bio(bio);
1882 struct mddev *mddev = r1_bio->mddev;
1883 struct r1conf *conf = mddev->private;
1886 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1889 abort_sync_write(mddev, r1_bio);
1890 set_bit(WriteErrorSeen, &rdev->flags);
1891 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1892 set_bit(MD_RECOVERY_NEEDED, &
1894 set_bit(R1BIO_WriteError, &r1_bio->state);
1895 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1896 &first_bad, &bad_sectors) &&
1897 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1900 &first_bad, &bad_sectors)
1902 set_bit(R1BIO_MadeGood, &r1_bio->state);
1904 if (atomic_dec_and_test(&r1_bio->remaining)) {
1905 int s = r1_bio->sectors;
1906 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1907 test_bit(R1BIO_WriteError, &r1_bio->state))
1908 reschedule_retry(r1_bio);
1911 md_done_sync(mddev, s, uptodate);
1916 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1917 int sectors, struct page *page, int rw)
1919 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1923 set_bit(WriteErrorSeen, &rdev->flags);
1924 if (!test_and_set_bit(WantReplacement,
1926 set_bit(MD_RECOVERY_NEEDED, &
1927 rdev->mddev->recovery);
1929 /* need to record an error - either for the block or the device */
1930 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1931 md_error(rdev->mddev, rdev);
1935 static int fix_sync_read_error(struct r1bio *r1_bio)
1937 /* Try some synchronous reads of other devices to get
1938 * good data, much like with normal read errors. Only
1939 * read into the pages we already have so we don't
1940 * need to re-issue the read request.
1941 * We don't need to freeze the array, because being in an
1942 * active sync request, there is no normal IO, and
1943 * no overlapping syncs.
1944 * We don't need to check is_badblock() again as we
1945 * made sure that anything with a bad block in range
1946 * will have bi_end_io clear.
1948 struct mddev *mddev = r1_bio->mddev;
1949 struct r1conf *conf = mddev->private;
1950 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1951 struct page **pages = get_resync_pages(bio)->pages;
1952 sector_t sect = r1_bio->sector;
1953 int sectors = r1_bio->sectors;
1955 struct md_rdev *rdev;
1957 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1958 if (test_bit(FailFast, &rdev->flags)) {
1959 /* Don't try recovering from here - just fail it
1960 * ... unless it is the last working device of course */
1961 md_error(mddev, rdev);
1962 if (test_bit(Faulty, &rdev->flags))
1963 /* Don't try to read from here, but make sure
1964 * put_buf does it's thing
1966 bio->bi_end_io = end_sync_write;
1971 int d = r1_bio->read_disk;
1975 if (s > (PAGE_SIZE>>9))
1978 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1979 /* No rcu protection needed here devices
1980 * can only be removed when no resync is
1981 * active, and resync is currently active
1983 rdev = conf->mirrors[d].rdev;
1984 if (sync_page_io(rdev, sect, s<<9,
1986 REQ_OP_READ, 0, false)) {
1992 if (d == conf->raid_disks * 2)
1994 } while (!success && d != r1_bio->read_disk);
1997 char b[BDEVNAME_SIZE];
1999 /* Cannot read from anywhere, this block is lost.
2000 * Record a bad block on each device. If that doesn't
2001 * work just disable and interrupt the recovery.
2002 * Don't fail devices as that won't really help.
2004 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2005 mdname(mddev), bio_devname(bio, b),
2006 (unsigned long long)r1_bio->sector);
2007 for (d = 0; d < conf->raid_disks * 2; d++) {
2008 rdev = conf->mirrors[d].rdev;
2009 if (!rdev || test_bit(Faulty, &rdev->flags))
2011 if (!rdev_set_badblocks(rdev, sect, s, 0))
2015 conf->recovery_disabled =
2016 mddev->recovery_disabled;
2017 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2018 md_done_sync(mddev, r1_bio->sectors, 0);
2030 /* write it back and re-read */
2031 while (d != r1_bio->read_disk) {
2033 d = conf->raid_disks * 2;
2035 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2037 rdev = conf->mirrors[d].rdev;
2038 if (r1_sync_page_io(rdev, sect, s,
2041 r1_bio->bios[d]->bi_end_io = NULL;
2042 rdev_dec_pending(rdev, mddev);
2046 while (d != r1_bio->read_disk) {
2048 d = conf->raid_disks * 2;
2050 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2052 rdev = conf->mirrors[d].rdev;
2053 if (r1_sync_page_io(rdev, sect, s,
2056 atomic_add(s, &rdev->corrected_errors);
2062 set_bit(R1BIO_Uptodate, &r1_bio->state);
2067 static void process_checks(struct r1bio *r1_bio)
2069 /* We have read all readable devices. If we haven't
2070 * got the block, then there is no hope left.
2071 * If we have, then we want to do a comparison
2072 * and skip the write if everything is the same.
2073 * If any blocks failed to read, then we need to
2074 * attempt an over-write
2076 struct mddev *mddev = r1_bio->mddev;
2077 struct r1conf *conf = mddev->private;
2082 /* Fix variable parts of all bios */
2083 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2084 for (i = 0; i < conf->raid_disks * 2; i++) {
2085 blk_status_t status;
2086 struct bio *b = r1_bio->bios[i];
2087 struct resync_pages *rp = get_resync_pages(b);
2088 if (b->bi_end_io != end_sync_read)
2090 /* fixup the bio for reuse, but preserve errno */
2091 status = b->bi_status;
2093 b->bi_status = status;
2094 b->bi_iter.bi_sector = r1_bio->sector +
2095 conf->mirrors[i].rdev->data_offset;
2096 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2097 b->bi_end_io = end_sync_read;
2098 rp->raid_bio = r1_bio;
2101 /* initialize bvec table again */
2102 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2104 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2105 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2106 !r1_bio->bios[primary]->bi_status) {
2107 r1_bio->bios[primary]->bi_end_io = NULL;
2108 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2111 r1_bio->read_disk = primary;
2112 for (i = 0; i < conf->raid_disks * 2; i++) {
2114 struct bio *pbio = r1_bio->bios[primary];
2115 struct bio *sbio = r1_bio->bios[i];
2116 blk_status_t status = sbio->bi_status;
2117 struct page **ppages = get_resync_pages(pbio)->pages;
2118 struct page **spages = get_resync_pages(sbio)->pages;
2120 int page_len[RESYNC_PAGES] = { 0 };
2121 struct bvec_iter_all iter_all;
2123 if (sbio->bi_end_io != end_sync_read)
2125 /* Now we can 'fixup' the error value */
2126 sbio->bi_status = 0;
2128 bio_for_each_segment_all(bi, sbio, iter_all)
2129 page_len[j++] = bi->bv_len;
2132 for (j = vcnt; j-- ; ) {
2133 if (memcmp(page_address(ppages[j]),
2134 page_address(spages[j]),
2141 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2142 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2144 /* No need to write to this device. */
2145 sbio->bi_end_io = NULL;
2146 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2150 bio_copy_data(sbio, pbio);
2154 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2156 struct r1conf *conf = mddev->private;
2158 int disks = conf->raid_disks * 2;
2161 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2162 /* ouch - failed to read all of that. */
2163 if (!fix_sync_read_error(r1_bio))
2166 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2167 process_checks(r1_bio);
2172 atomic_set(&r1_bio->remaining, 1);
2173 for (i = 0; i < disks ; i++) {
2174 wbio = r1_bio->bios[i];
2175 if (wbio->bi_end_io == NULL ||
2176 (wbio->bi_end_io == end_sync_read &&
2177 (i == r1_bio->read_disk ||
2178 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2180 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2181 abort_sync_write(mddev, r1_bio);
2185 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2186 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2187 wbio->bi_opf |= MD_FAILFAST;
2189 wbio->bi_end_io = end_sync_write;
2190 atomic_inc(&r1_bio->remaining);
2191 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2193 generic_make_request(wbio);
2196 if (atomic_dec_and_test(&r1_bio->remaining)) {
2197 /* if we're here, all write(s) have completed, so clean up */
2198 int s = r1_bio->sectors;
2199 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2200 test_bit(R1BIO_WriteError, &r1_bio->state))
2201 reschedule_retry(r1_bio);
2204 md_done_sync(mddev, s, 1);
2210 * This is a kernel thread which:
2212 * 1. Retries failed read operations on working mirrors.
2213 * 2. Updates the raid superblock when problems encounter.
2214 * 3. Performs writes following reads for array synchronising.
2217 static void fix_read_error(struct r1conf *conf, int read_disk,
2218 sector_t sect, int sectors)
2220 struct mddev *mddev = conf->mddev;
2226 struct md_rdev *rdev;
2228 if (s > (PAGE_SIZE>>9))
2236 rdev = rcu_dereference(conf->mirrors[d].rdev);
2238 (test_bit(In_sync, &rdev->flags) ||
2239 (!test_bit(Faulty, &rdev->flags) &&
2240 rdev->recovery_offset >= sect + s)) &&
2241 is_badblock(rdev, sect, s,
2242 &first_bad, &bad_sectors) == 0) {
2243 atomic_inc(&rdev->nr_pending);
2245 if (sync_page_io(rdev, sect, s<<9,
2246 conf->tmppage, REQ_OP_READ, 0, false))
2248 rdev_dec_pending(rdev, mddev);
2254 if (d == conf->raid_disks * 2)
2256 } while (!success && d != read_disk);
2259 /* Cannot read from anywhere - mark it bad */
2260 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2261 if (!rdev_set_badblocks(rdev, sect, s, 0))
2262 md_error(mddev, rdev);
2265 /* write it back and re-read */
2267 while (d != read_disk) {
2269 d = conf->raid_disks * 2;
2272 rdev = rcu_dereference(conf->mirrors[d].rdev);
2274 !test_bit(Faulty, &rdev->flags)) {
2275 atomic_inc(&rdev->nr_pending);
2277 r1_sync_page_io(rdev, sect, s,
2278 conf->tmppage, WRITE);
2279 rdev_dec_pending(rdev, mddev);
2284 while (d != read_disk) {
2285 char b[BDEVNAME_SIZE];
2287 d = conf->raid_disks * 2;
2290 rdev = rcu_dereference(conf->mirrors[d].rdev);
2292 !test_bit(Faulty, &rdev->flags)) {
2293 atomic_inc(&rdev->nr_pending);
2295 if (r1_sync_page_io(rdev, sect, s,
2296 conf->tmppage, READ)) {
2297 atomic_add(s, &rdev->corrected_errors);
2298 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2300 (unsigned long long)(sect +
2302 bdevname(rdev->bdev, b));
2304 rdev_dec_pending(rdev, mddev);
2313 static int narrow_write_error(struct r1bio *r1_bio, int i)
2315 struct mddev *mddev = r1_bio->mddev;
2316 struct r1conf *conf = mddev->private;
2317 struct md_rdev *rdev = conf->mirrors[i].rdev;
2319 /* bio has the data to be written to device 'i' where
2320 * we just recently had a write error.
2321 * We repeatedly clone the bio and trim down to one block,
2322 * then try the write. Where the write fails we record
2324 * It is conceivable that the bio doesn't exactly align with
2325 * blocks. We must handle this somehow.
2327 * We currently own a reference on the rdev.
2333 int sect_to_write = r1_bio->sectors;
2336 if (rdev->badblocks.shift < 0)
2339 block_sectors = roundup(1 << rdev->badblocks.shift,
2340 bdev_logical_block_size(rdev->bdev) >> 9);
2341 sector = r1_bio->sector;
2342 sectors = ((sector + block_sectors)
2343 & ~(sector_t)(block_sectors - 1))
2346 while (sect_to_write) {
2348 if (sectors > sect_to_write)
2349 sectors = sect_to_write;
2350 /* Write at 'sector' for 'sectors'*/
2352 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2353 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2357 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2361 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2362 wbio->bi_iter.bi_sector = r1_bio->sector;
2363 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2365 bio_trim(wbio, sector - r1_bio->sector, sectors);
2366 wbio->bi_iter.bi_sector += rdev->data_offset;
2367 bio_set_dev(wbio, rdev->bdev);
2369 if (submit_bio_wait(wbio) < 0)
2371 ok = rdev_set_badblocks(rdev, sector,
2376 sect_to_write -= sectors;
2378 sectors = block_sectors;
2383 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2386 int s = r1_bio->sectors;
2387 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2388 struct md_rdev *rdev = conf->mirrors[m].rdev;
2389 struct bio *bio = r1_bio->bios[m];
2390 if (bio->bi_end_io == NULL)
2392 if (!bio->bi_status &&
2393 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2394 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2396 if (bio->bi_status &&
2397 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2398 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2399 md_error(conf->mddev, rdev);
2403 md_done_sync(conf->mddev, s, 1);
2406 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2411 for (m = 0; m < conf->raid_disks * 2 ; m++)
2412 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2413 struct md_rdev *rdev = conf->mirrors[m].rdev;
2414 rdev_clear_badblocks(rdev,
2416 r1_bio->sectors, 0);
2417 rdev_dec_pending(rdev, conf->mddev);
2418 } else if (r1_bio->bios[m] != NULL) {
2419 /* This drive got a write error. We need to
2420 * narrow down and record precise write
2424 if (!narrow_write_error(r1_bio, m)) {
2425 md_error(conf->mddev,
2426 conf->mirrors[m].rdev);
2427 /* an I/O failed, we can't clear the bitmap */
2428 set_bit(R1BIO_Degraded, &r1_bio->state);
2430 rdev_dec_pending(conf->mirrors[m].rdev,
2434 spin_lock_irq(&conf->device_lock);
2435 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2436 idx = sector_to_idx(r1_bio->sector);
2437 atomic_inc(&conf->nr_queued[idx]);
2438 spin_unlock_irq(&conf->device_lock);
2440 * In case freeze_array() is waiting for condition
2441 * get_unqueued_pending() == extra to be true.
2443 wake_up(&conf->wait_barrier);
2444 md_wakeup_thread(conf->mddev->thread);
2446 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2447 close_write(r1_bio);
2448 raid_end_bio_io(r1_bio);
2452 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2454 struct mddev *mddev = conf->mddev;
2456 struct md_rdev *rdev;
2458 clear_bit(R1BIO_ReadError, &r1_bio->state);
2459 /* we got a read error. Maybe the drive is bad. Maybe just
2460 * the block and we can fix it.
2461 * We freeze all other IO, and try reading the block from
2462 * other devices. When we find one, we re-write
2463 * and check it that fixes the read error.
2464 * This is all done synchronously while the array is
2468 bio = r1_bio->bios[r1_bio->read_disk];
2470 r1_bio->bios[r1_bio->read_disk] = NULL;
2472 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2474 && !test_bit(FailFast, &rdev->flags)) {
2475 freeze_array(conf, 1);
2476 fix_read_error(conf, r1_bio->read_disk,
2477 r1_bio->sector, r1_bio->sectors);
2478 unfreeze_array(conf);
2479 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2480 md_error(mddev, rdev);
2482 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2485 rdev_dec_pending(rdev, conf->mddev);
2486 allow_barrier(conf, r1_bio->sector);
2487 bio = r1_bio->master_bio;
2489 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2491 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2494 static void raid1d(struct md_thread *thread)
2496 struct mddev *mddev = thread->mddev;
2497 struct r1bio *r1_bio;
2498 unsigned long flags;
2499 struct r1conf *conf = mddev->private;
2500 struct list_head *head = &conf->retry_list;
2501 struct blk_plug plug;
2504 md_check_recovery(mddev);
2506 if (!list_empty_careful(&conf->bio_end_io_list) &&
2507 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2509 spin_lock_irqsave(&conf->device_lock, flags);
2510 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2511 list_splice_init(&conf->bio_end_io_list, &tmp);
2512 spin_unlock_irqrestore(&conf->device_lock, flags);
2513 while (!list_empty(&tmp)) {
2514 r1_bio = list_first_entry(&tmp, struct r1bio,
2516 list_del(&r1_bio->retry_list);
2517 idx = sector_to_idx(r1_bio->sector);
2518 atomic_dec(&conf->nr_queued[idx]);
2519 if (mddev->degraded)
2520 set_bit(R1BIO_Degraded, &r1_bio->state);
2521 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2522 close_write(r1_bio);
2523 raid_end_bio_io(r1_bio);
2527 blk_start_plug(&plug);
2530 flush_pending_writes(conf);
2532 spin_lock_irqsave(&conf->device_lock, flags);
2533 if (list_empty(head)) {
2534 spin_unlock_irqrestore(&conf->device_lock, flags);
2537 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2538 list_del(head->prev);
2539 idx = sector_to_idx(r1_bio->sector);
2540 atomic_dec(&conf->nr_queued[idx]);
2541 spin_unlock_irqrestore(&conf->device_lock, flags);
2543 mddev = r1_bio->mddev;
2544 conf = mddev->private;
2545 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2546 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2547 test_bit(R1BIO_WriteError, &r1_bio->state))
2548 handle_sync_write_finished(conf, r1_bio);
2550 sync_request_write(mddev, r1_bio);
2551 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2552 test_bit(R1BIO_WriteError, &r1_bio->state))
2553 handle_write_finished(conf, r1_bio);
2554 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2555 handle_read_error(conf, r1_bio);
2560 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2561 md_check_recovery(mddev);
2563 blk_finish_plug(&plug);
2566 static int init_resync(struct r1conf *conf)
2570 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2571 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2573 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2574 r1buf_pool_free, conf->poolinfo);
2577 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2579 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2580 struct resync_pages *rps;
2584 for (i = conf->poolinfo->raid_disks; i--; ) {
2585 bio = r1bio->bios[i];
2586 rps = bio->bi_private;
2588 bio->bi_private = rps;
2590 r1bio->master_bio = NULL;
2595 * perform a "sync" on one "block"
2597 * We need to make sure that no normal I/O request - particularly write
2598 * requests - conflict with active sync requests.
2600 * This is achieved by tracking pending requests and a 'barrier' concept
2601 * that can be installed to exclude normal IO requests.
2604 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2607 struct r1conf *conf = mddev->private;
2608 struct r1bio *r1_bio;
2610 sector_t max_sector, nr_sectors;
2614 int write_targets = 0, read_targets = 0;
2615 sector_t sync_blocks;
2616 int still_degraded = 0;
2617 int good_sectors = RESYNC_SECTORS;
2618 int min_bad = 0; /* number of sectors that are bad in all devices */
2619 int idx = sector_to_idx(sector_nr);
2622 if (!mempool_initialized(&conf->r1buf_pool))
2623 if (init_resync(conf))
2626 max_sector = mddev->dev_sectors;
2627 if (sector_nr >= max_sector) {
2628 /* If we aborted, we need to abort the
2629 * sync on the 'current' bitmap chunk (there will
2630 * only be one in raid1 resync.
2631 * We can find the current addess in mddev->curr_resync
2633 if (mddev->curr_resync < max_sector) /* aborted */
2634 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2636 else /* completed sync */
2639 md_bitmap_close_sync(mddev->bitmap);
2642 if (mddev_is_clustered(mddev)) {
2643 conf->cluster_sync_low = 0;
2644 conf->cluster_sync_high = 0;
2649 if (mddev->bitmap == NULL &&
2650 mddev->recovery_cp == MaxSector &&
2651 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2652 conf->fullsync == 0) {
2654 return max_sector - sector_nr;
2656 /* before building a request, check if we can skip these blocks..
2657 * This call the bitmap_start_sync doesn't actually record anything
2659 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2660 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2661 /* We can skip this block, and probably several more */
2667 * If there is non-resync activity waiting for a turn, then let it
2668 * though before starting on this new sync request.
2670 if (atomic_read(&conf->nr_waiting[idx]))
2671 schedule_timeout_uninterruptible(1);
2673 /* we are incrementing sector_nr below. To be safe, we check against
2674 * sector_nr + two times RESYNC_SECTORS
2677 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2678 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2681 if (raise_barrier(conf, sector_nr))
2684 r1_bio = raid1_alloc_init_r1buf(conf);
2688 * If we get a correctably read error during resync or recovery,
2689 * we might want to read from a different device. So we
2690 * flag all drives that could conceivably be read from for READ,
2691 * and any others (which will be non-In_sync devices) for WRITE.
2692 * If a read fails, we try reading from something else for which READ
2696 r1_bio->mddev = mddev;
2697 r1_bio->sector = sector_nr;
2699 set_bit(R1BIO_IsSync, &r1_bio->state);
2700 /* make sure good_sectors won't go across barrier unit boundary */
2701 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2703 for (i = 0; i < conf->raid_disks * 2; i++) {
2704 struct md_rdev *rdev;
2705 bio = r1_bio->bios[i];
2707 rdev = rcu_dereference(conf->mirrors[i].rdev);
2709 test_bit(Faulty, &rdev->flags)) {
2710 if (i < conf->raid_disks)
2712 } else if (!test_bit(In_sync, &rdev->flags)) {
2713 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2714 bio->bi_end_io = end_sync_write;
2717 /* may need to read from here */
2718 sector_t first_bad = MaxSector;
2721 if (is_badblock(rdev, sector_nr, good_sectors,
2722 &first_bad, &bad_sectors)) {
2723 if (first_bad > sector_nr)
2724 good_sectors = first_bad - sector_nr;
2726 bad_sectors -= (sector_nr - first_bad);
2728 min_bad > bad_sectors)
2729 min_bad = bad_sectors;
2732 if (sector_nr < first_bad) {
2733 if (test_bit(WriteMostly, &rdev->flags)) {
2740 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2741 bio->bi_end_io = end_sync_read;
2743 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2744 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2745 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2747 * The device is suitable for reading (InSync),
2748 * but has bad block(s) here. Let's try to correct them,
2749 * if we are doing resync or repair. Otherwise, leave
2750 * this device alone for this sync request.
2752 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2753 bio->bi_end_io = end_sync_write;
2757 if (bio->bi_end_io) {
2758 atomic_inc(&rdev->nr_pending);
2759 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2760 bio_set_dev(bio, rdev->bdev);
2761 if (test_bit(FailFast, &rdev->flags))
2762 bio->bi_opf |= MD_FAILFAST;
2768 r1_bio->read_disk = disk;
2770 if (read_targets == 0 && min_bad > 0) {
2771 /* These sectors are bad on all InSync devices, so we
2772 * need to mark them bad on all write targets
2775 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2776 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2777 struct md_rdev *rdev = conf->mirrors[i].rdev;
2778 ok = rdev_set_badblocks(rdev, sector_nr,
2782 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2787 /* Cannot record the badblocks, so need to
2789 * If there are multiple read targets, could just
2790 * fail the really bad ones ???
2792 conf->recovery_disabled = mddev->recovery_disabled;
2793 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2799 if (min_bad > 0 && min_bad < good_sectors) {
2800 /* only resync enough to reach the next bad->good
2802 good_sectors = min_bad;
2805 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2806 /* extra read targets are also write targets */
2807 write_targets += read_targets-1;
2809 if (write_targets == 0 || read_targets == 0) {
2810 /* There is nowhere to write, so all non-sync
2811 * drives must be failed - so we are finished
2815 max_sector = sector_nr + min_bad;
2816 rv = max_sector - sector_nr;
2822 if (max_sector > mddev->resync_max)
2823 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2824 if (max_sector > sector_nr + good_sectors)
2825 max_sector = sector_nr + good_sectors;
2830 int len = PAGE_SIZE;
2831 if (sector_nr + (len>>9) > max_sector)
2832 len = (max_sector - sector_nr) << 9;
2835 if (sync_blocks == 0) {
2836 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2837 &sync_blocks, still_degraded) &&
2839 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2841 if ((len >> 9) > sync_blocks)
2842 len = sync_blocks<<9;
2845 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2846 struct resync_pages *rp;
2848 bio = r1_bio->bios[i];
2849 rp = get_resync_pages(bio);
2850 if (bio->bi_end_io) {
2851 page = resync_fetch_page(rp, page_idx);
2854 * won't fail because the vec table is big
2855 * enough to hold all these pages
2857 bio_add_page(bio, page, len, 0);
2860 nr_sectors += len>>9;
2861 sector_nr += len>>9;
2862 sync_blocks -= (len>>9);
2863 } while (++page_idx < RESYNC_PAGES);
2865 r1_bio->sectors = nr_sectors;
2867 if (mddev_is_clustered(mddev) &&
2868 conf->cluster_sync_high < sector_nr + nr_sectors) {
2869 conf->cluster_sync_low = mddev->curr_resync_completed;
2870 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2871 /* Send resync message */
2872 md_cluster_ops->resync_info_update(mddev,
2873 conf->cluster_sync_low,
2874 conf->cluster_sync_high);
2877 /* For a user-requested sync, we read all readable devices and do a
2880 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2881 atomic_set(&r1_bio->remaining, read_targets);
2882 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2883 bio = r1_bio->bios[i];
2884 if (bio->bi_end_io == end_sync_read) {
2886 md_sync_acct_bio(bio, nr_sectors);
2887 if (read_targets == 1)
2888 bio->bi_opf &= ~MD_FAILFAST;
2889 generic_make_request(bio);
2893 atomic_set(&r1_bio->remaining, 1);
2894 bio = r1_bio->bios[r1_bio->read_disk];
2895 md_sync_acct_bio(bio, nr_sectors);
2896 if (read_targets == 1)
2897 bio->bi_opf &= ~MD_FAILFAST;
2898 generic_make_request(bio);
2904 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2909 return mddev->dev_sectors;
2912 static struct r1conf *setup_conf(struct mddev *mddev)
2914 struct r1conf *conf;
2916 struct raid1_info *disk;
2917 struct md_rdev *rdev;
2920 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2924 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2925 sizeof(atomic_t), GFP_KERNEL);
2926 if (!conf->nr_pending)
2929 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2930 sizeof(atomic_t), GFP_KERNEL);
2931 if (!conf->nr_waiting)
2934 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2935 sizeof(atomic_t), GFP_KERNEL);
2936 if (!conf->nr_queued)
2939 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2940 sizeof(atomic_t), GFP_KERNEL);
2944 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2945 mddev->raid_disks, 2),
2950 conf->tmppage = alloc_page(GFP_KERNEL);
2954 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2955 if (!conf->poolinfo)
2957 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2958 err = mempool_init(&conf->r1bio_pool, NR_RAID1_BIOS, r1bio_pool_alloc,
2959 r1bio_pool_free, conf->poolinfo);
2963 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2967 conf->poolinfo->mddev = mddev;
2970 spin_lock_init(&conf->device_lock);
2971 rdev_for_each(rdev, mddev) {
2972 int disk_idx = rdev->raid_disk;
2973 if (disk_idx >= mddev->raid_disks
2976 if (test_bit(Replacement, &rdev->flags))
2977 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2979 disk = conf->mirrors + disk_idx;
2984 disk->head_position = 0;
2985 disk->seq_start = MaxSector;
2987 conf->raid_disks = mddev->raid_disks;
2988 conf->mddev = mddev;
2989 INIT_LIST_HEAD(&conf->retry_list);
2990 INIT_LIST_HEAD(&conf->bio_end_io_list);
2992 spin_lock_init(&conf->resync_lock);
2993 init_waitqueue_head(&conf->wait_barrier);
2995 bio_list_init(&conf->pending_bio_list);
2996 conf->pending_count = 0;
2997 conf->recovery_disabled = mddev->recovery_disabled - 1;
3000 for (i = 0; i < conf->raid_disks * 2; i++) {
3002 disk = conf->mirrors + i;
3004 if (i < conf->raid_disks &&
3005 disk[conf->raid_disks].rdev) {
3006 /* This slot has a replacement. */
3008 /* No original, just make the replacement
3009 * a recovering spare
3012 disk[conf->raid_disks].rdev;
3013 disk[conf->raid_disks].rdev = NULL;
3014 } else if (!test_bit(In_sync, &disk->rdev->flags))
3015 /* Original is not in_sync - bad */
3020 !test_bit(In_sync, &disk->rdev->flags)) {
3021 disk->head_position = 0;
3023 (disk->rdev->saved_raid_disk < 0))
3029 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3037 mempool_exit(&conf->r1bio_pool);
3038 kfree(conf->mirrors);
3039 safe_put_page(conf->tmppage);
3040 kfree(conf->poolinfo);
3041 kfree(conf->nr_pending);
3042 kfree(conf->nr_waiting);
3043 kfree(conf->nr_queued);
3044 kfree(conf->barrier);
3045 bioset_exit(&conf->bio_split);
3048 return ERR_PTR(err);
3051 static void raid1_free(struct mddev *mddev, void *priv);
3052 static int raid1_run(struct mddev *mddev)
3054 struct r1conf *conf;
3056 struct md_rdev *rdev;
3058 bool discard_supported = false;
3060 if (mddev->level != 1) {
3061 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3062 mdname(mddev), mddev->level);
3065 if (mddev->reshape_position != MaxSector) {
3066 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3070 if (mddev_init_writes_pending(mddev) < 0)
3073 * copy the already verified devices into our private RAID1
3074 * bookkeeping area. [whatever we allocate in run(),
3075 * should be freed in raid1_free()]
3077 if (mddev->private == NULL)
3078 conf = setup_conf(mddev);
3080 conf = mddev->private;
3083 return PTR_ERR(conf);
3086 blk_queue_max_write_same_sectors(mddev->queue, 0);
3087 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3090 rdev_for_each(rdev, mddev) {
3091 if (!mddev->gendisk)
3093 disk_stack_limits(mddev->gendisk, rdev->bdev,
3094 rdev->data_offset << 9);
3095 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3096 discard_supported = true;
3099 mddev->degraded = 0;
3100 for (i=0; i < conf->raid_disks; i++)
3101 if (conf->mirrors[i].rdev == NULL ||
3102 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3103 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3106 if (conf->raid_disks - mddev->degraded == 1)
3107 mddev->recovery_cp = MaxSector;
3109 if (mddev->recovery_cp != MaxSector)
3110 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3112 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3113 mdname(mddev), mddev->raid_disks - mddev->degraded,
3117 * Ok, everything is just fine now
3119 mddev->thread = conf->thread;
3120 conf->thread = NULL;
3121 mddev->private = conf;
3122 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3124 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3127 if (discard_supported)
3128 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3131 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3135 ret = md_integrity_register(mddev);
3137 md_unregister_thread(&mddev->thread);
3138 raid1_free(mddev, conf);
3143 static void raid1_free(struct mddev *mddev, void *priv)
3145 struct r1conf *conf = priv;
3147 mempool_exit(&conf->r1bio_pool);
3148 kfree(conf->mirrors);
3149 safe_put_page(conf->tmppage);
3150 kfree(conf->poolinfo);
3151 kfree(conf->nr_pending);
3152 kfree(conf->nr_waiting);
3153 kfree(conf->nr_queued);
3154 kfree(conf->barrier);
3155 bioset_exit(&conf->bio_split);
3159 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3161 /* no resync is happening, and there is enough space
3162 * on all devices, so we can resize.
3163 * We need to make sure resync covers any new space.
3164 * If the array is shrinking we should possibly wait until
3165 * any io in the removed space completes, but it hardly seems
3168 sector_t newsize = raid1_size(mddev, sectors, 0);
3169 if (mddev->external_size &&
3170 mddev->array_sectors > newsize)
3172 if (mddev->bitmap) {
3173 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3177 md_set_array_sectors(mddev, newsize);
3178 if (sectors > mddev->dev_sectors &&
3179 mddev->recovery_cp > mddev->dev_sectors) {
3180 mddev->recovery_cp = mddev->dev_sectors;
3181 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3183 mddev->dev_sectors = sectors;
3184 mddev->resync_max_sectors = sectors;
3188 static int raid1_reshape(struct mddev *mddev)
3191 * 1/ resize the r1bio_pool
3192 * 2/ resize conf->mirrors
3194 * We allocate a new r1bio_pool if we can.
3195 * Then raise a device barrier and wait until all IO stops.
3196 * Then resize conf->mirrors and swap in the new r1bio pool.
3198 * At the same time, we "pack" the devices so that all the missing
3199 * devices have the higher raid_disk numbers.
3201 mempool_t newpool, oldpool;
3202 struct pool_info *newpoolinfo;
3203 struct raid1_info *newmirrors;
3204 struct r1conf *conf = mddev->private;
3205 int cnt, raid_disks;
3206 unsigned long flags;
3210 memset(&newpool, 0, sizeof(newpool));
3211 memset(&oldpool, 0, sizeof(oldpool));
3213 /* Cannot change chunk_size, layout, or level */
3214 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3215 mddev->layout != mddev->new_layout ||
3216 mddev->level != mddev->new_level) {
3217 mddev->new_chunk_sectors = mddev->chunk_sectors;
3218 mddev->new_layout = mddev->layout;
3219 mddev->new_level = mddev->level;
3223 if (!mddev_is_clustered(mddev))
3224 md_allow_write(mddev);
3226 raid_disks = mddev->raid_disks + mddev->delta_disks;
3228 if (raid_disks < conf->raid_disks) {
3230 for (d= 0; d < conf->raid_disks; d++)
3231 if (conf->mirrors[d].rdev)
3233 if (cnt > raid_disks)
3237 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3240 newpoolinfo->mddev = mddev;
3241 newpoolinfo->raid_disks = raid_disks * 2;
3243 ret = mempool_init(&newpool, NR_RAID1_BIOS, r1bio_pool_alloc,
3244 r1bio_pool_free, newpoolinfo);
3249 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3254 mempool_exit(&newpool);
3258 freeze_array(conf, 0);
3260 /* ok, everything is stopped */
3261 oldpool = conf->r1bio_pool;
3262 conf->r1bio_pool = newpool;
3264 for (d = d2 = 0; d < conf->raid_disks; d++) {
3265 struct md_rdev *rdev = conf->mirrors[d].rdev;
3266 if (rdev && rdev->raid_disk != d2) {
3267 sysfs_unlink_rdev(mddev, rdev);
3268 rdev->raid_disk = d2;
3269 sysfs_unlink_rdev(mddev, rdev);
3270 if (sysfs_link_rdev(mddev, rdev))
3271 pr_warn("md/raid1:%s: cannot register rd%d\n",
3272 mdname(mddev), rdev->raid_disk);
3275 newmirrors[d2++].rdev = rdev;
3277 kfree(conf->mirrors);
3278 conf->mirrors = newmirrors;
3279 kfree(conf->poolinfo);
3280 conf->poolinfo = newpoolinfo;
3282 spin_lock_irqsave(&conf->device_lock, flags);
3283 mddev->degraded += (raid_disks - conf->raid_disks);
3284 spin_unlock_irqrestore(&conf->device_lock, flags);
3285 conf->raid_disks = mddev->raid_disks = raid_disks;
3286 mddev->delta_disks = 0;
3288 unfreeze_array(conf);
3290 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3291 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3292 md_wakeup_thread(mddev->thread);
3294 mempool_exit(&oldpool);
3298 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3300 struct r1conf *conf = mddev->private;
3303 freeze_array(conf, 0);
3305 unfreeze_array(conf);
3308 static void *raid1_takeover(struct mddev *mddev)
3310 /* raid1 can take over:
3311 * raid5 with 2 devices, any layout or chunk size
3313 if (mddev->level == 5 && mddev->raid_disks == 2) {
3314 struct r1conf *conf;
3315 mddev->new_level = 1;
3316 mddev->new_layout = 0;
3317 mddev->new_chunk_sectors = 0;
3318 conf = setup_conf(mddev);
3319 if (!IS_ERR(conf)) {
3320 /* Array must appear to be quiesced */
3321 conf->array_frozen = 1;
3322 mddev_clear_unsupported_flags(mddev,
3323 UNSUPPORTED_MDDEV_FLAGS);
3327 return ERR_PTR(-EINVAL);
3330 static struct md_personality raid1_personality =
3334 .owner = THIS_MODULE,
3335 .make_request = raid1_make_request,
3338 .status = raid1_status,
3339 .error_handler = raid1_error,
3340 .hot_add_disk = raid1_add_disk,
3341 .hot_remove_disk= raid1_remove_disk,
3342 .spare_active = raid1_spare_active,
3343 .sync_request = raid1_sync_request,
3344 .resize = raid1_resize,
3346 .check_reshape = raid1_reshape,
3347 .quiesce = raid1_quiesce,
3348 .takeover = raid1_takeover,
3349 .congested = raid1_congested,
3352 static int __init raid_init(void)
3354 return register_md_personality(&raid1_personality);
3357 static void raid_exit(void)
3359 unregister_md_personality(&raid1_personality);
3362 module_init(raid_init);
3363 module_exit(raid_exit);
3364 MODULE_LICENSE("GPL");
3365 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3366 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3367 MODULE_ALIAS("md-raid1");
3368 MODULE_ALIAS("md-level-1");
3370 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);