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)
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(sizeof(struct resync_pages) * pi->raid_disks,
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 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 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_bdev;
791 bio->bi_bdev = rdev->bdev;
792 if (test_bit(Faulty, &rdev->flags)) {
794 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
795 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
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) {
813 bio = bio_list_get(&conf->pending_bio_list);
814 conf->pending_count = 0;
815 spin_unlock_irq(&conf->device_lock);
816 flush_bio_list(conf, bio);
818 spin_unlock_irq(&conf->device_lock);
822 * Sometimes we need to suspend IO while we do something else,
823 * either some resync/recovery, or reconfigure the array.
824 * To do this we raise a 'barrier'.
825 * The 'barrier' is a counter that can be raised multiple times
826 * to count how many activities are happening which preclude
828 * We can only raise the barrier if there is no pending IO.
829 * i.e. if nr_pending == 0.
830 * We choose only to raise the barrier if no-one is waiting for the
831 * barrier to go down. This means that as soon as an IO request
832 * is ready, no other operations which require a barrier will start
833 * until the IO request has had a chance.
835 * So: regular IO calls 'wait_barrier'. When that returns there
836 * is no backgroup IO happening, It must arrange to call
837 * allow_barrier when it has finished its IO.
838 * backgroup IO calls must call raise_barrier. Once that returns
839 * there is no normal IO happeing. It must arrange to call
840 * lower_barrier when the particular background IO completes.
842 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
844 int idx = sector_to_idx(sector_nr);
846 spin_lock_irq(&conf->resync_lock);
848 /* Wait until no block IO is waiting */
849 wait_event_lock_irq(conf->wait_barrier,
850 !atomic_read(&conf->nr_waiting[idx]),
853 /* block any new IO from starting */
854 atomic_inc(&conf->barrier[idx]);
856 * In raise_barrier() we firstly increase conf->barrier[idx] then
857 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
858 * increase conf->nr_pending[idx] then check conf->barrier[idx].
859 * A memory barrier here to make sure conf->nr_pending[idx] won't
860 * be fetched before conf->barrier[idx] is increased. Otherwise
861 * there will be a race between raise_barrier() and _wait_barrier().
863 smp_mb__after_atomic();
865 /* For these conditions we must wait:
866 * A: while the array is in frozen state
867 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
868 * existing in corresponding I/O barrier bucket.
869 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
870 * max resync count which allowed on current I/O barrier bucket.
872 wait_event_lock_irq(conf->wait_barrier,
873 !conf->array_frozen &&
874 !atomic_read(&conf->nr_pending[idx]) &&
875 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH,
878 atomic_inc(&conf->nr_sync_pending);
879 spin_unlock_irq(&conf->resync_lock);
882 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
884 int idx = sector_to_idx(sector_nr);
886 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
888 atomic_dec(&conf->barrier[idx]);
889 atomic_dec(&conf->nr_sync_pending);
890 wake_up(&conf->wait_barrier);
893 static void _wait_barrier(struct r1conf *conf, int idx)
896 * We need to increase conf->nr_pending[idx] very early here,
897 * then raise_barrier() can be blocked when it waits for
898 * conf->nr_pending[idx] to be 0. Then we can avoid holding
899 * conf->resync_lock when there is no barrier raised in same
900 * barrier unit bucket. Also if the array is frozen, I/O
901 * should be blocked until array is unfrozen.
903 atomic_inc(&conf->nr_pending[idx]);
905 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
906 * check conf->barrier[idx]. In raise_barrier() we firstly increase
907 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
908 * barrier is necessary here to make sure conf->barrier[idx] won't be
909 * fetched before conf->nr_pending[idx] is increased. Otherwise there
910 * will be a race between _wait_barrier() and raise_barrier().
912 smp_mb__after_atomic();
915 * Don't worry about checking two atomic_t variables at same time
916 * here. If during we check conf->barrier[idx], the array is
917 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
918 * 0, it is safe to return and make the I/O continue. Because the
919 * array is frozen, all I/O returned here will eventually complete
920 * or be queued, no race will happen. See code comment in
923 if (!READ_ONCE(conf->array_frozen) &&
924 !atomic_read(&conf->barrier[idx]))
928 * After holding conf->resync_lock, conf->nr_pending[idx]
929 * should be decreased before waiting for barrier to drop.
930 * Otherwise, we may encounter a race condition because
931 * raise_barrer() might be waiting for conf->nr_pending[idx]
932 * to be 0 at same time.
934 spin_lock_irq(&conf->resync_lock);
935 atomic_inc(&conf->nr_waiting[idx]);
936 atomic_dec(&conf->nr_pending[idx]);
938 * In case freeze_array() is waiting for
939 * get_unqueued_pending() == extra
941 wake_up(&conf->wait_barrier);
942 /* Wait for the barrier in same barrier unit bucket to drop. */
943 wait_event_lock_irq(conf->wait_barrier,
944 !conf->array_frozen &&
945 !atomic_read(&conf->barrier[idx]),
947 atomic_inc(&conf->nr_pending[idx]);
948 atomic_dec(&conf->nr_waiting[idx]);
949 spin_unlock_irq(&conf->resync_lock);
952 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
954 int idx = sector_to_idx(sector_nr);
957 * Very similar to _wait_barrier(). The difference is, for read
958 * I/O we don't need wait for sync I/O, but if the whole array
959 * is frozen, the read I/O still has to wait until the array is
960 * unfrozen. Since there is no ordering requirement with
961 * conf->barrier[idx] here, memory barrier is unnecessary as well.
963 atomic_inc(&conf->nr_pending[idx]);
965 if (!READ_ONCE(conf->array_frozen))
968 spin_lock_irq(&conf->resync_lock);
969 atomic_inc(&conf->nr_waiting[idx]);
970 atomic_dec(&conf->nr_pending[idx]);
972 * In case freeze_array() is waiting for
973 * get_unqueued_pending() == extra
975 wake_up(&conf->wait_barrier);
976 /* Wait for array to be unfrozen */
977 wait_event_lock_irq(conf->wait_barrier,
980 atomic_inc(&conf->nr_pending[idx]);
981 atomic_dec(&conf->nr_waiting[idx]);
982 spin_unlock_irq(&conf->resync_lock);
985 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
987 int idx = sector_to_idx(sector_nr);
989 _wait_barrier(conf, idx);
992 static void wait_all_barriers(struct r1conf *conf)
996 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
997 _wait_barrier(conf, idx);
1000 static void _allow_barrier(struct r1conf *conf, int idx)
1002 atomic_dec(&conf->nr_pending[idx]);
1003 wake_up(&conf->wait_barrier);
1006 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1008 int idx = sector_to_idx(sector_nr);
1010 _allow_barrier(conf, idx);
1013 static void allow_all_barriers(struct r1conf *conf)
1017 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1018 _allow_barrier(conf, idx);
1021 /* conf->resync_lock should be held */
1022 static int get_unqueued_pending(struct r1conf *conf)
1026 ret = atomic_read(&conf->nr_sync_pending);
1027 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1028 ret += atomic_read(&conf->nr_pending[idx]) -
1029 atomic_read(&conf->nr_queued[idx]);
1034 static void freeze_array(struct r1conf *conf, int extra)
1036 /* Stop sync I/O and normal I/O and wait for everything to
1038 * This is called in two situations:
1039 * 1) management command handlers (reshape, remove disk, quiesce).
1040 * 2) one normal I/O request failed.
1042 * After array_frozen is set to 1, new sync IO will be blocked at
1043 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1044 * or wait_read_barrier(). The flying I/Os will either complete or be
1045 * queued. When everything goes quite, there are only queued I/Os left.
1047 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1048 * barrier bucket index which this I/O request hits. When all sync and
1049 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1050 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1051 * in handle_read_error(), we may call freeze_array() before trying to
1052 * fix the read error. In this case, the error read I/O is not queued,
1053 * so get_unqueued_pending() == 1.
1055 * Therefore before this function returns, we need to wait until
1056 * get_unqueued_pendings(conf) gets equal to extra. For
1057 * normal I/O context, extra is 1, in rested situations extra is 0.
1059 spin_lock_irq(&conf->resync_lock);
1060 conf->array_frozen = 1;
1061 raid1_log(conf->mddev, "wait freeze");
1062 wait_event_lock_irq_cmd(
1064 get_unqueued_pending(conf) == extra,
1066 flush_pending_writes(conf));
1067 spin_unlock_irq(&conf->resync_lock);
1069 static void unfreeze_array(struct r1conf *conf)
1071 /* reverse the effect of the freeze */
1072 spin_lock_irq(&conf->resync_lock);
1073 conf->array_frozen = 0;
1074 spin_unlock_irq(&conf->resync_lock);
1075 wake_up(&conf->wait_barrier);
1078 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1081 int size = bio->bi_iter.bi_size;
1082 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1084 struct bio *behind_bio = NULL;
1086 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1090 /* discard op, we don't support writezero/writesame yet */
1091 if (!bio_has_data(bio)) {
1092 behind_bio->bi_iter.bi_size = size;
1096 while (i < vcnt && size) {
1098 int len = min_t(int, PAGE_SIZE, size);
1100 page = alloc_page(GFP_NOIO);
1101 if (unlikely(!page))
1104 bio_add_page(behind_bio, page, len, 0);
1110 bio_copy_data(behind_bio, bio);
1112 r1_bio->behind_master_bio = behind_bio;;
1113 set_bit(R1BIO_BehindIO, &r1_bio->state);
1118 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1119 bio->bi_iter.bi_size);
1120 bio_free_pages(behind_bio);
1121 bio_put(behind_bio);
1124 struct raid1_plug_cb {
1125 struct blk_plug_cb cb;
1126 struct bio_list pending;
1130 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1132 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1134 struct mddev *mddev = plug->cb.data;
1135 struct r1conf *conf = mddev->private;
1138 if (from_schedule || current->bio_list) {
1139 spin_lock_irq(&conf->device_lock);
1140 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1141 conf->pending_count += plug->pending_cnt;
1142 spin_unlock_irq(&conf->device_lock);
1143 wake_up(&conf->wait_barrier);
1144 md_wakeup_thread(mddev->thread);
1149 /* we aren't scheduling, so we can do the write-out directly. */
1150 bio = bio_list_get(&plug->pending);
1151 flush_bio_list(conf, bio);
1155 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1157 r1_bio->master_bio = bio;
1158 r1_bio->sectors = bio_sectors(bio);
1160 r1_bio->mddev = mddev;
1161 r1_bio->sector = bio->bi_iter.bi_sector;
1164 static inline struct r1bio *
1165 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1167 struct r1conf *conf = mddev->private;
1168 struct r1bio *r1_bio;
1170 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1171 /* Ensure no bio records IO_BLOCKED */
1172 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1173 init_r1bio(r1_bio, mddev, bio);
1177 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1178 int max_read_sectors, struct r1bio *r1_bio)
1180 struct r1conf *conf = mddev->private;
1181 struct raid1_info *mirror;
1182 struct bio *read_bio;
1183 struct bitmap *bitmap = mddev->bitmap;
1184 const int op = bio_op(bio);
1185 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1188 bool print_msg = !!r1_bio;
1189 char b[BDEVNAME_SIZE];
1192 * If r1_bio is set, we are blocking the raid1d thread
1193 * so there is a tiny risk of deadlock. So ask for
1194 * emergency memory if needed.
1196 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1199 /* Need to get the block device name carefully */
1200 struct md_rdev *rdev;
1202 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1204 bdevname(rdev->bdev, b);
1211 * Still need barrier for READ in case that whole
1214 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1217 r1_bio = alloc_r1bio(mddev, bio);
1219 init_r1bio(r1_bio, mddev, bio);
1220 r1_bio->sectors = max_read_sectors;
1223 * make_request() can abort the operation when read-ahead is being
1224 * used and no empty request is available.
1226 rdisk = read_balance(conf, r1_bio, &max_sectors);
1229 /* couldn't find anywhere to read from */
1231 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1234 (unsigned long long)r1_bio->sector);
1236 raid_end_bio_io(r1_bio);
1239 mirror = conf->mirrors + rdisk;
1242 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1244 (unsigned long long)r1_bio->sector,
1245 bdevname(mirror->rdev->bdev, b));
1247 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1250 * Reading from a write-mostly device must take care not to
1251 * over-take any writes that are 'behind'
1253 raid1_log(mddev, "wait behind writes");
1254 wait_event(bitmap->behind_wait,
1255 atomic_read(&bitmap->behind_writes) == 0);
1258 if (max_sectors < bio_sectors(bio)) {
1259 struct bio *split = bio_split(bio, max_sectors,
1260 gfp, conf->bio_split);
1261 bio_chain(split, bio);
1262 generic_make_request(bio);
1264 r1_bio->master_bio = bio;
1265 r1_bio->sectors = max_sectors;
1268 r1_bio->read_disk = rdisk;
1270 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1272 r1_bio->bios[rdisk] = read_bio;
1274 read_bio->bi_iter.bi_sector = r1_bio->sector +
1275 mirror->rdev->data_offset;
1276 read_bio->bi_bdev = mirror->rdev->bdev;
1277 read_bio->bi_end_io = raid1_end_read_request;
1278 bio_set_op_attrs(read_bio, op, do_sync);
1279 if (test_bit(FailFast, &mirror->rdev->flags) &&
1280 test_bit(R1BIO_FailFast, &r1_bio->state))
1281 read_bio->bi_opf |= MD_FAILFAST;
1282 read_bio->bi_private = r1_bio;
1285 trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1286 read_bio, disk_devt(mddev->gendisk),
1289 generic_make_request(read_bio);
1292 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1293 int max_write_sectors)
1295 struct r1conf *conf = mddev->private;
1296 struct r1bio *r1_bio;
1298 struct bitmap *bitmap = mddev->bitmap;
1299 unsigned long flags;
1300 struct md_rdev *blocked_rdev;
1301 struct blk_plug_cb *cb;
1302 struct raid1_plug_cb *plug = NULL;
1307 * Register the new request and wait if the reconstruction
1308 * thread has put up a bar for new requests.
1309 * Continue immediately if no resync is active currently.
1313 if ((bio_end_sector(bio) > mddev->suspend_lo &&
1314 bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1315 (mddev_is_clustered(mddev) &&
1316 md_cluster_ops->area_resyncing(mddev, WRITE,
1317 bio->bi_iter.bi_sector, bio_end_sector(bio)))) {
1320 * As the suspend_* range is controlled by userspace, we want
1321 * an interruptible wait.
1326 prepare_to_wait(&conf->wait_barrier,
1327 &w, TASK_INTERRUPTIBLE);
1328 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1329 bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1330 (mddev_is_clustered(mddev) &&
1331 !md_cluster_ops->area_resyncing(mddev, WRITE,
1332 bio->bi_iter.bi_sector,
1333 bio_end_sector(bio))))
1336 sigprocmask(SIG_BLOCK, &full, &old);
1338 sigprocmask(SIG_SETMASK, &old, NULL);
1340 finish_wait(&conf->wait_barrier, &w);
1342 wait_barrier(conf, bio->bi_iter.bi_sector);
1344 r1_bio = alloc_r1bio(mddev, bio);
1345 r1_bio->sectors = max_write_sectors;
1347 if (conf->pending_count >= max_queued_requests) {
1348 md_wakeup_thread(mddev->thread);
1349 raid1_log(mddev, "wait queued");
1350 wait_event(conf->wait_barrier,
1351 conf->pending_count < max_queued_requests);
1353 /* first select target devices under rcu_lock and
1354 * inc refcount on their rdev. Record them by setting
1356 * If there are known/acknowledged bad blocks on any device on
1357 * which we have seen a write error, we want to avoid writing those
1359 * This potentially requires several writes to write around
1360 * the bad blocks. Each set of writes gets it's own r1bio
1361 * with a set of bios attached.
1364 disks = conf->raid_disks * 2;
1366 blocked_rdev = NULL;
1368 max_sectors = r1_bio->sectors;
1369 for (i = 0; i < disks; i++) {
1370 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1371 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1372 atomic_inc(&rdev->nr_pending);
1373 blocked_rdev = rdev;
1376 r1_bio->bios[i] = NULL;
1377 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1378 if (i < conf->raid_disks)
1379 set_bit(R1BIO_Degraded, &r1_bio->state);
1383 atomic_inc(&rdev->nr_pending);
1384 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1389 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1390 &first_bad, &bad_sectors);
1392 /* mustn't write here until the bad block is
1394 set_bit(BlockedBadBlocks, &rdev->flags);
1395 blocked_rdev = rdev;
1398 if (is_bad && first_bad <= r1_bio->sector) {
1399 /* Cannot write here at all */
1400 bad_sectors -= (r1_bio->sector - first_bad);
1401 if (bad_sectors < max_sectors)
1402 /* mustn't write more than bad_sectors
1403 * to other devices yet
1405 max_sectors = bad_sectors;
1406 rdev_dec_pending(rdev, mddev);
1407 /* We don't set R1BIO_Degraded as that
1408 * only applies if the disk is
1409 * missing, so it might be re-added,
1410 * and we want to know to recover this
1412 * In this case the device is here,
1413 * and the fact that this chunk is not
1414 * in-sync is recorded in the bad
1420 int good_sectors = first_bad - r1_bio->sector;
1421 if (good_sectors < max_sectors)
1422 max_sectors = good_sectors;
1425 r1_bio->bios[i] = bio;
1429 if (unlikely(blocked_rdev)) {
1430 /* Wait for this device to become unblocked */
1433 for (j = 0; j < i; j++)
1434 if (r1_bio->bios[j])
1435 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1437 allow_barrier(conf, bio->bi_iter.bi_sector);
1438 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1439 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1440 wait_barrier(conf, bio->bi_iter.bi_sector);
1444 if (max_sectors < bio_sectors(bio)) {
1445 struct bio *split = bio_split(bio, max_sectors,
1446 GFP_NOIO, conf->bio_split);
1447 bio_chain(split, bio);
1448 generic_make_request(bio);
1450 r1_bio->master_bio = bio;
1451 r1_bio->sectors = max_sectors;
1454 atomic_set(&r1_bio->remaining, 1);
1455 atomic_set(&r1_bio->behind_remaining, 0);
1459 for (i = 0; i < disks; i++) {
1460 struct bio *mbio = NULL;
1461 if (!r1_bio->bios[i])
1467 * Not if there are too many, or cannot
1468 * allocate memory, or a reader on WriteMostly
1469 * is waiting for behind writes to flush */
1471 (atomic_read(&bitmap->behind_writes)
1472 < mddev->bitmap_info.max_write_behind) &&
1473 !waitqueue_active(&bitmap->behind_wait)) {
1474 alloc_behind_master_bio(r1_bio, bio);
1477 bitmap_startwrite(bitmap, r1_bio->sector,
1479 test_bit(R1BIO_BehindIO,
1484 if (r1_bio->behind_master_bio)
1485 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1486 GFP_NOIO, mddev->bio_set);
1488 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1490 if (r1_bio->behind_master_bio) {
1491 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1492 atomic_inc(&r1_bio->behind_remaining);
1495 r1_bio->bios[i] = mbio;
1497 mbio->bi_iter.bi_sector = (r1_bio->sector +
1498 conf->mirrors[i].rdev->data_offset);
1499 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1500 mbio->bi_end_io = raid1_end_write_request;
1501 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1502 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1503 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1504 conf->raid_disks - mddev->degraded > 1)
1505 mbio->bi_opf |= MD_FAILFAST;
1506 mbio->bi_private = r1_bio;
1508 atomic_inc(&r1_bio->remaining);
1511 trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1512 mbio, disk_devt(mddev->gendisk),
1514 /* flush_pending_writes() needs access to the rdev so...*/
1515 mbio->bi_bdev = (void*)conf->mirrors[i].rdev;
1517 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1519 plug = container_of(cb, struct raid1_plug_cb, cb);
1523 bio_list_add(&plug->pending, mbio);
1524 plug->pending_cnt++;
1526 spin_lock_irqsave(&conf->device_lock, flags);
1527 bio_list_add(&conf->pending_bio_list, mbio);
1528 conf->pending_count++;
1529 spin_unlock_irqrestore(&conf->device_lock, flags);
1530 md_wakeup_thread(mddev->thread);
1534 r1_bio_write_done(r1_bio);
1536 /* In case raid1d snuck in to freeze_array */
1537 wake_up(&conf->wait_barrier);
1540 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1544 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1545 md_flush_request(mddev, bio);
1550 * There is a limit to the maximum size, but
1551 * the read/write handler might find a lower limit
1552 * due to bad blocks. To avoid multiple splits,
1553 * we pass the maximum number of sectors down
1554 * and let the lower level perform the split.
1556 sectors = align_to_barrier_unit_end(
1557 bio->bi_iter.bi_sector, bio_sectors(bio));
1559 if (bio_data_dir(bio) == READ)
1560 raid1_read_request(mddev, bio, sectors, NULL);
1562 if (!md_write_start(mddev,bio))
1564 raid1_write_request(mddev, bio, sectors);
1569 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1571 struct r1conf *conf = mddev->private;
1574 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1575 conf->raid_disks - mddev->degraded);
1577 for (i = 0; i < conf->raid_disks; i++) {
1578 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1579 seq_printf(seq, "%s",
1580 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1583 seq_printf(seq, "]");
1586 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1588 char b[BDEVNAME_SIZE];
1589 struct r1conf *conf = mddev->private;
1590 unsigned long flags;
1593 * If it is not operational, then we have already marked it as dead
1594 * else if it is the last working disks, ignore the error, let the
1595 * next level up know.
1596 * else mark the drive as failed
1598 spin_lock_irqsave(&conf->device_lock, flags);
1599 if (test_bit(In_sync, &rdev->flags)
1600 && (conf->raid_disks - mddev->degraded) == 1) {
1602 * Don't fail the drive, act as though we were just a
1603 * normal single drive.
1604 * However don't try a recovery from this drive as
1605 * it is very likely to fail.
1607 conf->recovery_disabled = mddev->recovery_disabled;
1608 spin_unlock_irqrestore(&conf->device_lock, flags);
1611 set_bit(Blocked, &rdev->flags);
1612 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1614 set_bit(Faulty, &rdev->flags);
1616 set_bit(Faulty, &rdev->flags);
1617 spin_unlock_irqrestore(&conf->device_lock, flags);
1619 * if recovery is running, make sure it aborts.
1621 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1622 set_mask_bits(&mddev->sb_flags, 0,
1623 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1624 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1625 "md/raid1:%s: Operation continuing on %d devices.\n",
1626 mdname(mddev), bdevname(rdev->bdev, b),
1627 mdname(mddev), conf->raid_disks - mddev->degraded);
1630 static void print_conf(struct r1conf *conf)
1634 pr_debug("RAID1 conf printout:\n");
1636 pr_debug("(!conf)\n");
1639 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1643 for (i = 0; i < conf->raid_disks; i++) {
1644 char b[BDEVNAME_SIZE];
1645 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1647 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1648 i, !test_bit(In_sync, &rdev->flags),
1649 !test_bit(Faulty, &rdev->flags),
1650 bdevname(rdev->bdev,b));
1655 static void close_sync(struct r1conf *conf)
1657 wait_all_barriers(conf);
1658 allow_all_barriers(conf);
1660 mempool_destroy(conf->r1buf_pool);
1661 conf->r1buf_pool = NULL;
1664 static int raid1_spare_active(struct mddev *mddev)
1667 struct r1conf *conf = mddev->private;
1669 unsigned long flags;
1672 * Find all failed disks within the RAID1 configuration
1673 * and mark them readable.
1674 * Called under mddev lock, so rcu protection not needed.
1675 * device_lock used to avoid races with raid1_end_read_request
1676 * which expects 'In_sync' flags and ->degraded to be consistent.
1678 spin_lock_irqsave(&conf->device_lock, flags);
1679 for (i = 0; i < conf->raid_disks; i++) {
1680 struct md_rdev *rdev = conf->mirrors[i].rdev;
1681 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1683 && !test_bit(Candidate, &repl->flags)
1684 && repl->recovery_offset == MaxSector
1685 && !test_bit(Faulty, &repl->flags)
1686 && !test_and_set_bit(In_sync, &repl->flags)) {
1687 /* replacement has just become active */
1689 !test_and_clear_bit(In_sync, &rdev->flags))
1692 /* Replaced device not technically
1693 * faulty, but we need to be sure
1694 * it gets removed and never re-added
1696 set_bit(Faulty, &rdev->flags);
1697 sysfs_notify_dirent_safe(
1702 && rdev->recovery_offset == MaxSector
1703 && !test_bit(Faulty, &rdev->flags)
1704 && !test_and_set_bit(In_sync, &rdev->flags)) {
1706 sysfs_notify_dirent_safe(rdev->sysfs_state);
1709 mddev->degraded -= count;
1710 spin_unlock_irqrestore(&conf->device_lock, flags);
1716 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1718 struct r1conf *conf = mddev->private;
1721 struct raid1_info *p;
1723 int last = conf->raid_disks - 1;
1725 if (mddev->recovery_disabled == conf->recovery_disabled)
1728 if (md_integrity_add_rdev(rdev, mddev))
1731 if (rdev->raid_disk >= 0)
1732 first = last = rdev->raid_disk;
1735 * find the disk ... but prefer rdev->saved_raid_disk
1738 if (rdev->saved_raid_disk >= 0 &&
1739 rdev->saved_raid_disk >= first &&
1740 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1741 first = last = rdev->saved_raid_disk;
1743 for (mirror = first; mirror <= last; mirror++) {
1744 p = conf->mirrors+mirror;
1748 disk_stack_limits(mddev->gendisk, rdev->bdev,
1749 rdev->data_offset << 9);
1751 p->head_position = 0;
1752 rdev->raid_disk = mirror;
1754 /* As all devices are equivalent, we don't need a full recovery
1755 * if this was recently any drive of the array
1757 if (rdev->saved_raid_disk < 0)
1759 rcu_assign_pointer(p->rdev, rdev);
1762 if (test_bit(WantReplacement, &p->rdev->flags) &&
1763 p[conf->raid_disks].rdev == NULL) {
1764 /* Add this device as a replacement */
1765 clear_bit(In_sync, &rdev->flags);
1766 set_bit(Replacement, &rdev->flags);
1767 rdev->raid_disk = mirror;
1770 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1774 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1775 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1780 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1782 struct r1conf *conf = mddev->private;
1784 int number = rdev->raid_disk;
1785 struct raid1_info *p = conf->mirrors + number;
1787 if (rdev != p->rdev)
1788 p = conf->mirrors + conf->raid_disks + number;
1791 if (rdev == p->rdev) {
1792 if (test_bit(In_sync, &rdev->flags) ||
1793 atomic_read(&rdev->nr_pending)) {
1797 /* Only remove non-faulty devices if recovery
1800 if (!test_bit(Faulty, &rdev->flags) &&
1801 mddev->recovery_disabled != conf->recovery_disabled &&
1802 mddev->degraded < conf->raid_disks) {
1807 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1809 if (atomic_read(&rdev->nr_pending)) {
1810 /* lost the race, try later */
1816 if (conf->mirrors[conf->raid_disks + number].rdev) {
1817 /* We just removed a device that is being replaced.
1818 * Move down the replacement. We drain all IO before
1819 * doing this to avoid confusion.
1821 struct md_rdev *repl =
1822 conf->mirrors[conf->raid_disks + number].rdev;
1823 freeze_array(conf, 0);
1824 clear_bit(Replacement, &repl->flags);
1826 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1827 unfreeze_array(conf);
1830 clear_bit(WantReplacement, &rdev->flags);
1831 err = md_integrity_register(mddev);
1839 static void end_sync_read(struct bio *bio)
1841 struct r1bio *r1_bio = get_resync_r1bio(bio);
1843 update_head_pos(r1_bio->read_disk, r1_bio);
1846 * we have read a block, now it needs to be re-written,
1847 * or re-read if the read failed.
1848 * We don't do much here, just schedule handling by raid1d
1850 if (!bio->bi_status)
1851 set_bit(R1BIO_Uptodate, &r1_bio->state);
1853 if (atomic_dec_and_test(&r1_bio->remaining))
1854 reschedule_retry(r1_bio);
1857 static void end_sync_write(struct bio *bio)
1859 int uptodate = !bio->bi_status;
1860 struct r1bio *r1_bio = get_resync_r1bio(bio);
1861 struct mddev *mddev = r1_bio->mddev;
1862 struct r1conf *conf = mddev->private;
1865 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1868 sector_t sync_blocks = 0;
1869 sector_t s = r1_bio->sector;
1870 long sectors_to_go = r1_bio->sectors;
1871 /* make sure these bits doesn't get cleared. */
1873 bitmap_end_sync(mddev->bitmap, s,
1876 sectors_to_go -= sync_blocks;
1877 } while (sectors_to_go > 0);
1878 set_bit(WriteErrorSeen, &rdev->flags);
1879 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1880 set_bit(MD_RECOVERY_NEEDED, &
1882 set_bit(R1BIO_WriteError, &r1_bio->state);
1883 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1884 &first_bad, &bad_sectors) &&
1885 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1888 &first_bad, &bad_sectors)
1890 set_bit(R1BIO_MadeGood, &r1_bio->state);
1892 if (atomic_dec_and_test(&r1_bio->remaining)) {
1893 int s = r1_bio->sectors;
1894 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1895 test_bit(R1BIO_WriteError, &r1_bio->state))
1896 reschedule_retry(r1_bio);
1899 md_done_sync(mddev, s, uptodate);
1904 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1905 int sectors, struct page *page, int rw)
1907 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1911 set_bit(WriteErrorSeen, &rdev->flags);
1912 if (!test_and_set_bit(WantReplacement,
1914 set_bit(MD_RECOVERY_NEEDED, &
1915 rdev->mddev->recovery);
1917 /* need to record an error - either for the block or the device */
1918 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1919 md_error(rdev->mddev, rdev);
1923 static int fix_sync_read_error(struct r1bio *r1_bio)
1925 /* Try some synchronous reads of other devices to get
1926 * good data, much like with normal read errors. Only
1927 * read into the pages we already have so we don't
1928 * need to re-issue the read request.
1929 * We don't need to freeze the array, because being in an
1930 * active sync request, there is no normal IO, and
1931 * no overlapping syncs.
1932 * We don't need to check is_badblock() again as we
1933 * made sure that anything with a bad block in range
1934 * will have bi_end_io clear.
1936 struct mddev *mddev = r1_bio->mddev;
1937 struct r1conf *conf = mddev->private;
1938 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1939 struct page **pages = get_resync_pages(bio)->pages;
1940 sector_t sect = r1_bio->sector;
1941 int sectors = r1_bio->sectors;
1943 struct md_rdev *rdev;
1945 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1946 if (test_bit(FailFast, &rdev->flags)) {
1947 /* Don't try recovering from here - just fail it
1948 * ... unless it is the last working device of course */
1949 md_error(mddev, rdev);
1950 if (test_bit(Faulty, &rdev->flags))
1951 /* Don't try to read from here, but make sure
1952 * put_buf does it's thing
1954 bio->bi_end_io = end_sync_write;
1959 int d = r1_bio->read_disk;
1963 if (s > (PAGE_SIZE>>9))
1966 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1967 /* No rcu protection needed here devices
1968 * can only be removed when no resync is
1969 * active, and resync is currently active
1971 rdev = conf->mirrors[d].rdev;
1972 if (sync_page_io(rdev, sect, s<<9,
1974 REQ_OP_READ, 0, false)) {
1980 if (d == conf->raid_disks * 2)
1982 } while (!success && d != r1_bio->read_disk);
1985 char b[BDEVNAME_SIZE];
1987 /* Cannot read from anywhere, this block is lost.
1988 * Record a bad block on each device. If that doesn't
1989 * work just disable and interrupt the recovery.
1990 * Don't fail devices as that won't really help.
1992 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1994 bdevname(bio->bi_bdev, b),
1995 (unsigned long long)r1_bio->sector);
1996 for (d = 0; d < conf->raid_disks * 2; d++) {
1997 rdev = conf->mirrors[d].rdev;
1998 if (!rdev || test_bit(Faulty, &rdev->flags))
2000 if (!rdev_set_badblocks(rdev, sect, s, 0))
2004 conf->recovery_disabled =
2005 mddev->recovery_disabled;
2006 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2007 md_done_sync(mddev, r1_bio->sectors, 0);
2019 /* write it back and re-read */
2020 while (d != r1_bio->read_disk) {
2022 d = conf->raid_disks * 2;
2024 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2026 rdev = conf->mirrors[d].rdev;
2027 if (r1_sync_page_io(rdev, sect, s,
2030 r1_bio->bios[d]->bi_end_io = NULL;
2031 rdev_dec_pending(rdev, mddev);
2035 while (d != r1_bio->read_disk) {
2037 d = conf->raid_disks * 2;
2039 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2041 rdev = conf->mirrors[d].rdev;
2042 if (r1_sync_page_io(rdev, sect, s,
2045 atomic_add(s, &rdev->corrected_errors);
2051 set_bit(R1BIO_Uptodate, &r1_bio->state);
2056 static void process_checks(struct r1bio *r1_bio)
2058 /* We have read all readable devices. If we haven't
2059 * got the block, then there is no hope left.
2060 * If we have, then we want to do a comparison
2061 * and skip the write if everything is the same.
2062 * If any blocks failed to read, then we need to
2063 * attempt an over-write
2065 struct mddev *mddev = r1_bio->mddev;
2066 struct r1conf *conf = mddev->private;
2071 /* Fix variable parts of all bios */
2072 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2073 for (i = 0; i < conf->raid_disks * 2; i++) {
2074 blk_status_t status;
2075 struct bio *b = r1_bio->bios[i];
2076 struct resync_pages *rp = get_resync_pages(b);
2077 if (b->bi_end_io != end_sync_read)
2079 /* fixup the bio for reuse, but preserve errno */
2080 status = b->bi_status;
2082 b->bi_status = status;
2083 b->bi_iter.bi_sector = r1_bio->sector +
2084 conf->mirrors[i].rdev->data_offset;
2085 b->bi_bdev = conf->mirrors[i].rdev->bdev;
2086 b->bi_end_io = end_sync_read;
2087 rp->raid_bio = r1_bio;
2090 /* initialize bvec table again */
2091 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2093 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2094 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2095 !r1_bio->bios[primary]->bi_status) {
2096 r1_bio->bios[primary]->bi_end_io = NULL;
2097 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2100 r1_bio->read_disk = primary;
2101 for (i = 0; i < conf->raid_disks * 2; i++) {
2103 struct bio *pbio = r1_bio->bios[primary];
2104 struct bio *sbio = r1_bio->bios[i];
2105 blk_status_t status = sbio->bi_status;
2106 struct page **ppages = get_resync_pages(pbio)->pages;
2107 struct page **spages = get_resync_pages(sbio)->pages;
2109 int page_len[RESYNC_PAGES] = { 0 };
2111 if (sbio->bi_end_io != end_sync_read)
2113 /* Now we can 'fixup' the error value */
2114 sbio->bi_status = 0;
2116 bio_for_each_segment_all(bi, sbio, j)
2117 page_len[j] = bi->bv_len;
2120 for (j = vcnt; j-- ; ) {
2121 if (memcmp(page_address(ppages[j]),
2122 page_address(spages[j]),
2129 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2130 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2132 /* No need to write to this device. */
2133 sbio->bi_end_io = NULL;
2134 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2138 bio_copy_data(sbio, pbio);
2142 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2144 struct r1conf *conf = mddev->private;
2146 int disks = conf->raid_disks * 2;
2149 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2150 /* ouch - failed to read all of that. */
2151 if (!fix_sync_read_error(r1_bio))
2154 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2155 process_checks(r1_bio);
2160 atomic_set(&r1_bio->remaining, 1);
2161 for (i = 0; i < disks ; i++) {
2162 wbio = r1_bio->bios[i];
2163 if (wbio->bi_end_io == NULL ||
2164 (wbio->bi_end_io == end_sync_read &&
2165 (i == r1_bio->read_disk ||
2166 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2168 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2171 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2172 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2173 wbio->bi_opf |= MD_FAILFAST;
2175 wbio->bi_end_io = end_sync_write;
2176 atomic_inc(&r1_bio->remaining);
2177 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2179 generic_make_request(wbio);
2182 if (atomic_dec_and_test(&r1_bio->remaining)) {
2183 /* if we're here, all write(s) have completed, so clean up */
2184 int s = r1_bio->sectors;
2185 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2186 test_bit(R1BIO_WriteError, &r1_bio->state))
2187 reschedule_retry(r1_bio);
2190 md_done_sync(mddev, s, 1);
2196 * This is a kernel thread which:
2198 * 1. Retries failed read operations on working mirrors.
2199 * 2. Updates the raid superblock when problems encounter.
2200 * 3. Performs writes following reads for array synchronising.
2203 static void fix_read_error(struct r1conf *conf, int read_disk,
2204 sector_t sect, int sectors)
2206 struct mddev *mddev = conf->mddev;
2212 struct md_rdev *rdev;
2214 if (s > (PAGE_SIZE>>9))
2222 rdev = rcu_dereference(conf->mirrors[d].rdev);
2224 (test_bit(In_sync, &rdev->flags) ||
2225 (!test_bit(Faulty, &rdev->flags) &&
2226 rdev->recovery_offset >= sect + s)) &&
2227 is_badblock(rdev, sect, s,
2228 &first_bad, &bad_sectors) == 0) {
2229 atomic_inc(&rdev->nr_pending);
2231 if (sync_page_io(rdev, sect, s<<9,
2232 conf->tmppage, REQ_OP_READ, 0, false))
2234 rdev_dec_pending(rdev, mddev);
2240 if (d == conf->raid_disks * 2)
2242 } while (!success && d != read_disk);
2245 /* Cannot read from anywhere - mark it bad */
2246 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2247 if (!rdev_set_badblocks(rdev, sect, s, 0))
2248 md_error(mddev, rdev);
2251 /* write it back and re-read */
2253 while (d != read_disk) {
2255 d = conf->raid_disks * 2;
2258 rdev = rcu_dereference(conf->mirrors[d].rdev);
2260 !test_bit(Faulty, &rdev->flags)) {
2261 atomic_inc(&rdev->nr_pending);
2263 r1_sync_page_io(rdev, sect, s,
2264 conf->tmppage, WRITE);
2265 rdev_dec_pending(rdev, mddev);
2270 while (d != read_disk) {
2271 char b[BDEVNAME_SIZE];
2273 d = conf->raid_disks * 2;
2276 rdev = rcu_dereference(conf->mirrors[d].rdev);
2278 !test_bit(Faulty, &rdev->flags)) {
2279 atomic_inc(&rdev->nr_pending);
2281 if (r1_sync_page_io(rdev, sect, s,
2282 conf->tmppage, READ)) {
2283 atomic_add(s, &rdev->corrected_errors);
2284 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2286 (unsigned long long)(sect +
2288 bdevname(rdev->bdev, b));
2290 rdev_dec_pending(rdev, mddev);
2299 static int narrow_write_error(struct r1bio *r1_bio, int i)
2301 struct mddev *mddev = r1_bio->mddev;
2302 struct r1conf *conf = mddev->private;
2303 struct md_rdev *rdev = conf->mirrors[i].rdev;
2305 /* bio has the data to be written to device 'i' where
2306 * we just recently had a write error.
2307 * We repeatedly clone the bio and trim down to one block,
2308 * then try the write. Where the write fails we record
2310 * It is conceivable that the bio doesn't exactly align with
2311 * blocks. We must handle this somehow.
2313 * We currently own a reference on the rdev.
2319 int sect_to_write = r1_bio->sectors;
2322 if (rdev->badblocks.shift < 0)
2325 block_sectors = roundup(1 << rdev->badblocks.shift,
2326 bdev_logical_block_size(rdev->bdev) >> 9);
2327 sector = r1_bio->sector;
2328 sectors = ((sector + block_sectors)
2329 & ~(sector_t)(block_sectors - 1))
2332 while (sect_to_write) {
2334 if (sectors > sect_to_write)
2335 sectors = sect_to_write;
2336 /* Write at 'sector' for 'sectors'*/
2338 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2339 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2343 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2347 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2348 wbio->bi_iter.bi_sector = r1_bio->sector;
2349 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2351 bio_trim(wbio, sector - r1_bio->sector, sectors);
2352 wbio->bi_iter.bi_sector += rdev->data_offset;
2353 wbio->bi_bdev = rdev->bdev;
2355 if (submit_bio_wait(wbio) < 0)
2357 ok = rdev_set_badblocks(rdev, sector,
2362 sect_to_write -= sectors;
2364 sectors = block_sectors;
2369 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2372 int s = r1_bio->sectors;
2373 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2374 struct md_rdev *rdev = conf->mirrors[m].rdev;
2375 struct bio *bio = r1_bio->bios[m];
2376 if (bio->bi_end_io == NULL)
2378 if (!bio->bi_status &&
2379 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2380 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2382 if (bio->bi_status &&
2383 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2384 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2385 md_error(conf->mddev, rdev);
2389 md_done_sync(conf->mddev, s, 1);
2392 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2397 for (m = 0; m < conf->raid_disks * 2 ; m++)
2398 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2399 struct md_rdev *rdev = conf->mirrors[m].rdev;
2400 rdev_clear_badblocks(rdev,
2402 r1_bio->sectors, 0);
2403 rdev_dec_pending(rdev, conf->mddev);
2404 } else if (r1_bio->bios[m] != NULL) {
2405 /* This drive got a write error. We need to
2406 * narrow down and record precise write
2410 if (!narrow_write_error(r1_bio, m)) {
2411 md_error(conf->mddev,
2412 conf->mirrors[m].rdev);
2413 /* an I/O failed, we can't clear the bitmap */
2414 set_bit(R1BIO_Degraded, &r1_bio->state);
2416 rdev_dec_pending(conf->mirrors[m].rdev,
2420 spin_lock_irq(&conf->device_lock);
2421 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2422 idx = sector_to_idx(r1_bio->sector);
2423 atomic_inc(&conf->nr_queued[idx]);
2424 spin_unlock_irq(&conf->device_lock);
2426 * In case freeze_array() is waiting for condition
2427 * get_unqueued_pending() == extra to be true.
2429 wake_up(&conf->wait_barrier);
2430 md_wakeup_thread(conf->mddev->thread);
2432 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2433 close_write(r1_bio);
2434 raid_end_bio_io(r1_bio);
2438 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2440 struct mddev *mddev = conf->mddev;
2442 struct md_rdev *rdev;
2444 sector_t bio_sector;
2446 clear_bit(R1BIO_ReadError, &r1_bio->state);
2447 /* we got a read error. Maybe the drive is bad. Maybe just
2448 * the block and we can fix it.
2449 * We freeze all other IO, and try reading the block from
2450 * other devices. When we find one, we re-write
2451 * and check it that fixes the read error.
2452 * This is all done synchronously while the array is
2456 bio = r1_bio->bios[r1_bio->read_disk];
2457 bio_dev = bio->bi_bdev->bd_dev;
2458 bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2460 r1_bio->bios[r1_bio->read_disk] = NULL;
2462 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2464 && !test_bit(FailFast, &rdev->flags)) {
2465 freeze_array(conf, 1);
2466 fix_read_error(conf, r1_bio->read_disk,
2467 r1_bio->sector, r1_bio->sectors);
2468 unfreeze_array(conf);
2470 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2473 rdev_dec_pending(rdev, conf->mddev);
2474 allow_barrier(conf, r1_bio->sector);
2475 bio = r1_bio->master_bio;
2477 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2479 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2482 static void raid1d(struct md_thread *thread)
2484 struct mddev *mddev = thread->mddev;
2485 struct r1bio *r1_bio;
2486 unsigned long flags;
2487 struct r1conf *conf = mddev->private;
2488 struct list_head *head = &conf->retry_list;
2489 struct blk_plug plug;
2492 md_check_recovery(mddev);
2494 if (!list_empty_careful(&conf->bio_end_io_list) &&
2495 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2497 spin_lock_irqsave(&conf->device_lock, flags);
2498 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2499 list_splice_init(&conf->bio_end_io_list, &tmp);
2500 spin_unlock_irqrestore(&conf->device_lock, flags);
2501 while (!list_empty(&tmp)) {
2502 r1_bio = list_first_entry(&tmp, struct r1bio,
2504 list_del(&r1_bio->retry_list);
2505 idx = sector_to_idx(r1_bio->sector);
2506 atomic_dec(&conf->nr_queued[idx]);
2507 if (mddev->degraded)
2508 set_bit(R1BIO_Degraded, &r1_bio->state);
2509 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2510 close_write(r1_bio);
2511 raid_end_bio_io(r1_bio);
2515 blk_start_plug(&plug);
2518 flush_pending_writes(conf);
2520 spin_lock_irqsave(&conf->device_lock, flags);
2521 if (list_empty(head)) {
2522 spin_unlock_irqrestore(&conf->device_lock, flags);
2525 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2526 list_del(head->prev);
2527 idx = sector_to_idx(r1_bio->sector);
2528 atomic_dec(&conf->nr_queued[idx]);
2529 spin_unlock_irqrestore(&conf->device_lock, flags);
2531 mddev = r1_bio->mddev;
2532 conf = mddev->private;
2533 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2534 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2535 test_bit(R1BIO_WriteError, &r1_bio->state))
2536 handle_sync_write_finished(conf, r1_bio);
2538 sync_request_write(mddev, r1_bio);
2539 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2540 test_bit(R1BIO_WriteError, &r1_bio->state))
2541 handle_write_finished(conf, r1_bio);
2542 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2543 handle_read_error(conf, r1_bio);
2548 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2549 md_check_recovery(mddev);
2551 blk_finish_plug(&plug);
2554 static int init_resync(struct r1conf *conf)
2558 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2559 BUG_ON(conf->r1buf_pool);
2560 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2562 if (!conf->r1buf_pool)
2568 * perform a "sync" on one "block"
2570 * We need to make sure that no normal I/O request - particularly write
2571 * requests - conflict with active sync requests.
2573 * This is achieved by tracking pending requests and a 'barrier' concept
2574 * that can be installed to exclude normal IO requests.
2577 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2580 struct r1conf *conf = mddev->private;
2581 struct r1bio *r1_bio;
2583 sector_t max_sector, nr_sectors;
2587 int write_targets = 0, read_targets = 0;
2588 sector_t sync_blocks;
2589 int still_degraded = 0;
2590 int good_sectors = RESYNC_SECTORS;
2591 int min_bad = 0; /* number of sectors that are bad in all devices */
2592 int idx = sector_to_idx(sector_nr);
2595 if (!conf->r1buf_pool)
2596 if (init_resync(conf))
2599 max_sector = mddev->dev_sectors;
2600 if (sector_nr >= max_sector) {
2601 /* If we aborted, we need to abort the
2602 * sync on the 'current' bitmap chunk (there will
2603 * only be one in raid1 resync.
2604 * We can find the current addess in mddev->curr_resync
2606 if (mddev->curr_resync < max_sector) /* aborted */
2607 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2609 else /* completed sync */
2612 bitmap_close_sync(mddev->bitmap);
2615 if (mddev_is_clustered(mddev)) {
2616 conf->cluster_sync_low = 0;
2617 conf->cluster_sync_high = 0;
2622 if (mddev->bitmap == NULL &&
2623 mddev->recovery_cp == MaxSector &&
2624 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2625 conf->fullsync == 0) {
2627 return max_sector - sector_nr;
2629 /* before building a request, check if we can skip these blocks..
2630 * This call the bitmap_start_sync doesn't actually record anything
2632 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2633 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2634 /* We can skip this block, and probably several more */
2640 * If there is non-resync activity waiting for a turn, then let it
2641 * though before starting on this new sync request.
2643 if (atomic_read(&conf->nr_waiting[idx]))
2644 schedule_timeout_uninterruptible(1);
2646 /* we are incrementing sector_nr below. To be safe, we check against
2647 * sector_nr + two times RESYNC_SECTORS
2650 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2651 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2652 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2654 raise_barrier(conf, sector_nr);
2658 * If we get a correctably read error during resync or recovery,
2659 * we might want to read from a different device. So we
2660 * flag all drives that could conceivably be read from for READ,
2661 * and any others (which will be non-In_sync devices) for WRITE.
2662 * If a read fails, we try reading from something else for which READ
2666 r1_bio->mddev = mddev;
2667 r1_bio->sector = sector_nr;
2669 set_bit(R1BIO_IsSync, &r1_bio->state);
2670 /* make sure good_sectors won't go across barrier unit boundary */
2671 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2673 for (i = 0; i < conf->raid_disks * 2; i++) {
2674 struct md_rdev *rdev;
2675 bio = r1_bio->bios[i];
2677 rdev = rcu_dereference(conf->mirrors[i].rdev);
2679 test_bit(Faulty, &rdev->flags)) {
2680 if (i < conf->raid_disks)
2682 } else if (!test_bit(In_sync, &rdev->flags)) {
2683 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2684 bio->bi_end_io = end_sync_write;
2687 /* may need to read from here */
2688 sector_t first_bad = MaxSector;
2691 if (is_badblock(rdev, sector_nr, good_sectors,
2692 &first_bad, &bad_sectors)) {
2693 if (first_bad > sector_nr)
2694 good_sectors = first_bad - sector_nr;
2696 bad_sectors -= (sector_nr - first_bad);
2698 min_bad > bad_sectors)
2699 min_bad = bad_sectors;
2702 if (sector_nr < first_bad) {
2703 if (test_bit(WriteMostly, &rdev->flags)) {
2710 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2711 bio->bi_end_io = end_sync_read;
2713 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2714 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2715 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2717 * The device is suitable for reading (InSync),
2718 * but has bad block(s) here. Let's try to correct them,
2719 * if we are doing resync or repair. Otherwise, leave
2720 * this device alone for this sync request.
2722 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2723 bio->bi_end_io = end_sync_write;
2727 if (bio->bi_end_io) {
2728 atomic_inc(&rdev->nr_pending);
2729 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2730 bio->bi_bdev = rdev->bdev;
2731 if (test_bit(FailFast, &rdev->flags))
2732 bio->bi_opf |= MD_FAILFAST;
2738 r1_bio->read_disk = disk;
2740 if (read_targets == 0 && min_bad > 0) {
2741 /* These sectors are bad on all InSync devices, so we
2742 * need to mark them bad on all write targets
2745 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2746 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2747 struct md_rdev *rdev = conf->mirrors[i].rdev;
2748 ok = rdev_set_badblocks(rdev, sector_nr,
2752 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2757 /* Cannot record the badblocks, so need to
2759 * If there are multiple read targets, could just
2760 * fail the really bad ones ???
2762 conf->recovery_disabled = mddev->recovery_disabled;
2763 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2769 if (min_bad > 0 && min_bad < good_sectors) {
2770 /* only resync enough to reach the next bad->good
2772 good_sectors = min_bad;
2775 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2776 /* extra read targets are also write targets */
2777 write_targets += read_targets-1;
2779 if (write_targets == 0 || read_targets == 0) {
2780 /* There is nowhere to write, so all non-sync
2781 * drives must be failed - so we are finished
2785 max_sector = sector_nr + min_bad;
2786 rv = max_sector - sector_nr;
2792 if (max_sector > mddev->resync_max)
2793 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2794 if (max_sector > sector_nr + good_sectors)
2795 max_sector = sector_nr + good_sectors;
2800 int len = PAGE_SIZE;
2801 if (sector_nr + (len>>9) > max_sector)
2802 len = (max_sector - sector_nr) << 9;
2805 if (sync_blocks == 0) {
2806 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2807 &sync_blocks, still_degraded) &&
2809 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2811 if ((len >> 9) > sync_blocks)
2812 len = sync_blocks<<9;
2815 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2816 struct resync_pages *rp;
2818 bio = r1_bio->bios[i];
2819 rp = get_resync_pages(bio);
2820 if (bio->bi_end_io) {
2821 page = resync_fetch_page(rp, page_idx);
2824 * won't fail because the vec table is big
2825 * enough to hold all these pages
2827 bio_add_page(bio, page, len, 0);
2830 nr_sectors += len>>9;
2831 sector_nr += len>>9;
2832 sync_blocks -= (len>>9);
2833 } while (++page_idx < RESYNC_PAGES);
2835 r1_bio->sectors = nr_sectors;
2837 if (mddev_is_clustered(mddev) &&
2838 conf->cluster_sync_high < sector_nr + nr_sectors) {
2839 conf->cluster_sync_low = mddev->curr_resync_completed;
2840 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2841 /* Send resync message */
2842 md_cluster_ops->resync_info_update(mddev,
2843 conf->cluster_sync_low,
2844 conf->cluster_sync_high);
2847 /* For a user-requested sync, we read all readable devices and do a
2850 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2851 atomic_set(&r1_bio->remaining, read_targets);
2852 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2853 bio = r1_bio->bios[i];
2854 if (bio->bi_end_io == end_sync_read) {
2856 md_sync_acct(bio->bi_bdev, nr_sectors);
2857 if (read_targets == 1)
2858 bio->bi_opf &= ~MD_FAILFAST;
2859 generic_make_request(bio);
2863 atomic_set(&r1_bio->remaining, 1);
2864 bio = r1_bio->bios[r1_bio->read_disk];
2865 md_sync_acct(bio->bi_bdev, nr_sectors);
2866 if (read_targets == 1)
2867 bio->bi_opf &= ~MD_FAILFAST;
2868 generic_make_request(bio);
2874 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2879 return mddev->dev_sectors;
2882 static struct r1conf *setup_conf(struct mddev *mddev)
2884 struct r1conf *conf;
2886 struct raid1_info *disk;
2887 struct md_rdev *rdev;
2890 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2894 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2895 sizeof(atomic_t), GFP_KERNEL);
2896 if (!conf->nr_pending)
2899 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2900 sizeof(atomic_t), GFP_KERNEL);
2901 if (!conf->nr_waiting)
2904 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2905 sizeof(atomic_t), GFP_KERNEL);
2906 if (!conf->nr_queued)
2909 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2910 sizeof(atomic_t), GFP_KERNEL);
2914 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2915 * mddev->raid_disks * 2,
2920 conf->tmppage = alloc_page(GFP_KERNEL);
2924 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2925 if (!conf->poolinfo)
2927 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2928 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2931 if (!conf->r1bio_pool)
2934 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
2935 if (!conf->bio_split)
2938 conf->poolinfo->mddev = mddev;
2941 spin_lock_init(&conf->device_lock);
2942 rdev_for_each(rdev, mddev) {
2943 int disk_idx = rdev->raid_disk;
2944 if (disk_idx >= mddev->raid_disks
2947 if (test_bit(Replacement, &rdev->flags))
2948 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2950 disk = conf->mirrors + disk_idx;
2955 disk->head_position = 0;
2956 disk->seq_start = MaxSector;
2958 conf->raid_disks = mddev->raid_disks;
2959 conf->mddev = mddev;
2960 INIT_LIST_HEAD(&conf->retry_list);
2961 INIT_LIST_HEAD(&conf->bio_end_io_list);
2963 spin_lock_init(&conf->resync_lock);
2964 init_waitqueue_head(&conf->wait_barrier);
2966 bio_list_init(&conf->pending_bio_list);
2967 conf->pending_count = 0;
2968 conf->recovery_disabled = mddev->recovery_disabled - 1;
2971 for (i = 0; i < conf->raid_disks * 2; i++) {
2973 disk = conf->mirrors + i;
2975 if (i < conf->raid_disks &&
2976 disk[conf->raid_disks].rdev) {
2977 /* This slot has a replacement. */
2979 /* No original, just make the replacement
2980 * a recovering spare
2983 disk[conf->raid_disks].rdev;
2984 disk[conf->raid_disks].rdev = NULL;
2985 } else if (!test_bit(In_sync, &disk->rdev->flags))
2986 /* Original is not in_sync - bad */
2991 !test_bit(In_sync, &disk->rdev->flags)) {
2992 disk->head_position = 0;
2994 (disk->rdev->saved_raid_disk < 0))
3000 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3008 mempool_destroy(conf->r1bio_pool);
3009 kfree(conf->mirrors);
3010 safe_put_page(conf->tmppage);
3011 kfree(conf->poolinfo);
3012 kfree(conf->nr_pending);
3013 kfree(conf->nr_waiting);
3014 kfree(conf->nr_queued);
3015 kfree(conf->barrier);
3016 if (conf->bio_split)
3017 bioset_free(conf->bio_split);
3020 return ERR_PTR(err);
3023 static void raid1_free(struct mddev *mddev, void *priv);
3024 static int raid1_run(struct mddev *mddev)
3026 struct r1conf *conf;
3028 struct md_rdev *rdev;
3030 bool discard_supported = false;
3032 if (mddev->level != 1) {
3033 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3034 mdname(mddev), mddev->level);
3037 if (mddev->reshape_position != MaxSector) {
3038 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3042 if (mddev_init_writes_pending(mddev) < 0)
3045 * copy the already verified devices into our private RAID1
3046 * bookkeeping area. [whatever we allocate in run(),
3047 * should be freed in raid1_free()]
3049 if (mddev->private == NULL)
3050 conf = setup_conf(mddev);
3052 conf = mddev->private;
3055 return PTR_ERR(conf);
3058 blk_queue_max_write_same_sectors(mddev->queue, 0);
3059 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3062 rdev_for_each(rdev, mddev) {
3063 if (!mddev->gendisk)
3065 disk_stack_limits(mddev->gendisk, rdev->bdev,
3066 rdev->data_offset << 9);
3067 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3068 discard_supported = true;
3071 mddev->degraded = 0;
3072 for (i=0; i < conf->raid_disks; i++)
3073 if (conf->mirrors[i].rdev == NULL ||
3074 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3075 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3078 if (conf->raid_disks - mddev->degraded == 1)
3079 mddev->recovery_cp = MaxSector;
3081 if (mddev->recovery_cp != MaxSector)
3082 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3084 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3085 mdname(mddev), mddev->raid_disks - mddev->degraded,
3089 * Ok, everything is just fine now
3091 mddev->thread = conf->thread;
3092 conf->thread = NULL;
3093 mddev->private = conf;
3094 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3096 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3099 if (discard_supported)
3100 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3103 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3107 ret = md_integrity_register(mddev);
3109 md_unregister_thread(&mddev->thread);
3110 raid1_free(mddev, conf);
3115 static void raid1_free(struct mddev *mddev, void *priv)
3117 struct r1conf *conf = priv;
3119 mempool_destroy(conf->r1bio_pool);
3120 kfree(conf->mirrors);
3121 safe_put_page(conf->tmppage);
3122 kfree(conf->poolinfo);
3123 kfree(conf->nr_pending);
3124 kfree(conf->nr_waiting);
3125 kfree(conf->nr_queued);
3126 kfree(conf->barrier);
3127 if (conf->bio_split)
3128 bioset_free(conf->bio_split);
3132 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3134 /* no resync is happening, and there is enough space
3135 * on all devices, so we can resize.
3136 * We need to make sure resync covers any new space.
3137 * If the array is shrinking we should possibly wait until
3138 * any io in the removed space completes, but it hardly seems
3141 sector_t newsize = raid1_size(mddev, sectors, 0);
3142 if (mddev->external_size &&
3143 mddev->array_sectors > newsize)
3145 if (mddev->bitmap) {
3146 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3150 md_set_array_sectors(mddev, newsize);
3151 if (sectors > mddev->dev_sectors &&
3152 mddev->recovery_cp > mddev->dev_sectors) {
3153 mddev->recovery_cp = mddev->dev_sectors;
3154 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3156 mddev->dev_sectors = sectors;
3157 mddev->resync_max_sectors = sectors;
3161 static int raid1_reshape(struct mddev *mddev)
3164 * 1/ resize the r1bio_pool
3165 * 2/ resize conf->mirrors
3167 * We allocate a new r1bio_pool if we can.
3168 * Then raise a device barrier and wait until all IO stops.
3169 * Then resize conf->mirrors and swap in the new r1bio pool.
3171 * At the same time, we "pack" the devices so that all the missing
3172 * devices have the higher raid_disk numbers.
3174 mempool_t *newpool, *oldpool;
3175 struct pool_info *newpoolinfo;
3176 struct raid1_info *newmirrors;
3177 struct r1conf *conf = mddev->private;
3178 int cnt, raid_disks;
3179 unsigned long flags;
3182 /* Cannot change chunk_size, layout, or level */
3183 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3184 mddev->layout != mddev->new_layout ||
3185 mddev->level != mddev->new_level) {
3186 mddev->new_chunk_sectors = mddev->chunk_sectors;
3187 mddev->new_layout = mddev->layout;
3188 mddev->new_level = mddev->level;
3192 if (!mddev_is_clustered(mddev))
3193 md_allow_write(mddev);
3195 raid_disks = mddev->raid_disks + mddev->delta_disks;
3197 if (raid_disks < conf->raid_disks) {
3199 for (d= 0; d < conf->raid_disks; d++)
3200 if (conf->mirrors[d].rdev)
3202 if (cnt > raid_disks)
3206 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3209 newpoolinfo->mddev = mddev;
3210 newpoolinfo->raid_disks = raid_disks * 2;
3212 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3213 r1bio_pool_free, newpoolinfo);
3218 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3222 mempool_destroy(newpool);
3226 freeze_array(conf, 0);
3228 /* ok, everything is stopped */
3229 oldpool = conf->r1bio_pool;
3230 conf->r1bio_pool = newpool;
3232 for (d = d2 = 0; d < conf->raid_disks; d++) {
3233 struct md_rdev *rdev = conf->mirrors[d].rdev;
3234 if (rdev && rdev->raid_disk != d2) {
3235 sysfs_unlink_rdev(mddev, rdev);
3236 rdev->raid_disk = d2;
3237 sysfs_unlink_rdev(mddev, rdev);
3238 if (sysfs_link_rdev(mddev, rdev))
3239 pr_warn("md/raid1:%s: cannot register rd%d\n",
3240 mdname(mddev), rdev->raid_disk);
3243 newmirrors[d2++].rdev = rdev;
3245 kfree(conf->mirrors);
3246 conf->mirrors = newmirrors;
3247 kfree(conf->poolinfo);
3248 conf->poolinfo = newpoolinfo;
3250 spin_lock_irqsave(&conf->device_lock, flags);
3251 mddev->degraded += (raid_disks - conf->raid_disks);
3252 spin_unlock_irqrestore(&conf->device_lock, flags);
3253 conf->raid_disks = mddev->raid_disks = raid_disks;
3254 mddev->delta_disks = 0;
3256 unfreeze_array(conf);
3258 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3259 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3260 md_wakeup_thread(mddev->thread);
3262 mempool_destroy(oldpool);
3266 static void raid1_quiesce(struct mddev *mddev, int state)
3268 struct r1conf *conf = mddev->private;
3271 case 2: /* wake for suspend */
3272 wake_up(&conf->wait_barrier);
3275 freeze_array(conf, 0);
3278 unfreeze_array(conf);
3283 static void *raid1_takeover(struct mddev *mddev)
3285 /* raid1 can take over:
3286 * raid5 with 2 devices, any layout or chunk size
3288 if (mddev->level == 5 && mddev->raid_disks == 2) {
3289 struct r1conf *conf;
3290 mddev->new_level = 1;
3291 mddev->new_layout = 0;
3292 mddev->new_chunk_sectors = 0;
3293 conf = setup_conf(mddev);
3294 if (!IS_ERR(conf)) {
3295 /* Array must appear to be quiesced */
3296 conf->array_frozen = 1;
3297 mddev_clear_unsupported_flags(mddev,
3298 UNSUPPORTED_MDDEV_FLAGS);
3302 return ERR_PTR(-EINVAL);
3305 static struct md_personality raid1_personality =
3309 .owner = THIS_MODULE,
3310 .make_request = raid1_make_request,
3313 .status = raid1_status,
3314 .error_handler = raid1_error,
3315 .hot_add_disk = raid1_add_disk,
3316 .hot_remove_disk= raid1_remove_disk,
3317 .spare_active = raid1_spare_active,
3318 .sync_request = raid1_sync_request,
3319 .resize = raid1_resize,
3321 .check_reshape = raid1_reshape,
3322 .quiesce = raid1_quiesce,
3323 .takeover = raid1_takeover,
3324 .congested = raid1_congested,
3327 static int __init raid_init(void)
3329 return register_md_personality(&raid1_personality);
3332 static void raid_exit(void)
3334 unregister_md_personality(&raid1_personality);
3337 module_init(raid_init);
3338 module_exit(raid_exit);
3339 MODULE_LICENSE("GPL");
3340 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3341 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3342 MODULE_ALIAS("md-raid1");
3343 MODULE_ALIAS("md-level-1");
3345 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);