1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * raid1.c : Multiple Devices driver for Linux
5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
9 * RAID-1 management functions.
11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
17 * bitmapped intelligence in resync:
19 * - bitmap marked during normal i/o
20 * - bitmap used to skip nondirty blocks during sync
22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
23 * - persistent bitmap code
26 #include <linux/slab.h>
27 #include <linux/delay.h>
28 #include <linux/blkdev.h>
29 #include <linux/module.h>
30 #include <linux/seq_file.h>
31 #include <linux/ratelimit.h>
33 #include <trace/events/block.h>
37 #include "md-bitmap.h"
39 #define UNSUPPORTED_MDDEV_FLAGS \
40 ((1L << MD_HAS_JOURNAL) | \
41 (1L << MD_JOURNAL_CLEAN) | \
42 (1L << MD_HAS_PPL) | \
43 (1L << MD_HAS_MULTIPLE_PPLS))
46 * Number of guaranteed r1bios in case of extreme VM load:
48 #define NR_RAID1_BIOS 256
50 /* when we get a read error on a read-only array, we redirect to another
51 * device without failing the first device, or trying to over-write to
52 * correct the read error. To keep track of bad blocks on a per-bio
53 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
55 #define IO_BLOCKED ((struct bio *)1)
56 /* When we successfully write to a known bad-block, we need to remove the
57 * bad-block marking which must be done from process context. So we record
58 * the success by setting devs[n].bio to IO_MADE_GOOD
60 #define IO_MADE_GOOD ((struct bio *)2)
62 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
64 /* When there are this many requests queue to be written by
65 * the raid1 thread, we become 'congested' to provide back-pressure
68 static int max_queued_requests = 1024;
70 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
71 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
73 #define raid1_log(md, fmt, args...) \
74 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
79 * for resync bio, r1bio pointer can be retrieved from the per-bio
80 * 'struct resync_pages'.
82 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
84 return get_resync_pages(bio)->raid_bio;
87 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
89 struct pool_info *pi = data;
90 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
92 /* allocate a r1bio with room for raid_disks entries in the bios array */
93 return kzalloc(size, gfp_flags);
96 static void r1bio_pool_free(void *r1_bio, void *data)
101 #define RESYNC_DEPTH 32
102 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
103 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
104 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
105 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
106 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
108 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
110 struct pool_info *pi = data;
111 struct r1bio *r1_bio;
115 struct resync_pages *rps;
117 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
121 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
127 * Allocate bios : 1 for reading, n-1 for writing
129 for (j = pi->raid_disks ; j-- ; ) {
130 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
133 r1_bio->bios[j] = bio;
136 * Allocate RESYNC_PAGES data pages and attach them to
138 * If this is a user-requested check/repair, allocate
139 * RESYNC_PAGES for each bio.
141 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
142 need_pages = pi->raid_disks;
145 for (j = 0; j < pi->raid_disks; j++) {
146 struct resync_pages *rp = &rps[j];
148 bio = r1_bio->bios[j];
150 if (j < need_pages) {
151 if (resync_alloc_pages(rp, gfp_flags))
154 memcpy(rp, &rps[0], sizeof(*rp));
155 resync_get_all_pages(rp);
158 rp->raid_bio = r1_bio;
159 bio->bi_private = rp;
162 r1_bio->master_bio = NULL;
168 resync_free_pages(&rps[j]);
171 while (++j < pi->raid_disks)
172 bio_put(r1_bio->bios[j]);
176 r1bio_pool_free(r1_bio, data);
180 static void r1buf_pool_free(void *__r1_bio, void *data)
182 struct pool_info *pi = data;
184 struct r1bio *r1bio = __r1_bio;
185 struct resync_pages *rp = NULL;
187 for (i = pi->raid_disks; i--; ) {
188 rp = get_resync_pages(r1bio->bios[i]);
189 resync_free_pages(rp);
190 bio_put(r1bio->bios[i]);
193 /* resync pages array stored in the 1st bio's .bi_private */
196 r1bio_pool_free(r1bio, data);
199 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
203 for (i = 0; i < conf->raid_disks * 2; i++) {
204 struct bio **bio = r1_bio->bios + i;
205 if (!BIO_SPECIAL(*bio))
211 static void free_r1bio(struct r1bio *r1_bio)
213 struct r1conf *conf = r1_bio->mddev->private;
215 put_all_bios(conf, r1_bio);
216 mempool_free(r1_bio, &conf->r1bio_pool);
219 static void put_buf(struct r1bio *r1_bio)
221 struct r1conf *conf = r1_bio->mddev->private;
222 sector_t sect = r1_bio->sector;
225 for (i = 0; i < conf->raid_disks * 2; i++) {
226 struct bio *bio = r1_bio->bios[i];
228 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
231 mempool_free(r1_bio, &conf->r1buf_pool);
233 lower_barrier(conf, sect);
236 static void reschedule_retry(struct r1bio *r1_bio)
239 struct mddev *mddev = r1_bio->mddev;
240 struct r1conf *conf = mddev->private;
243 idx = sector_to_idx(r1_bio->sector);
244 spin_lock_irqsave(&conf->device_lock, flags);
245 list_add(&r1_bio->retry_list, &conf->retry_list);
246 atomic_inc(&conf->nr_queued[idx]);
247 spin_unlock_irqrestore(&conf->device_lock, flags);
249 wake_up(&conf->wait_barrier);
250 md_wakeup_thread(mddev->thread);
254 * raid_end_bio_io() is called when we have finished servicing a mirrored
255 * operation and are ready to return a success/failure code to the buffer
258 static void call_bio_endio(struct r1bio *r1_bio)
260 struct bio *bio = r1_bio->master_bio;
261 struct r1conf *conf = r1_bio->mddev->private;
263 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
264 bio->bi_status = BLK_STS_IOERR;
268 * Wake up any possible resync thread that waits for the device
271 allow_barrier(conf, r1_bio->sector);
274 static void raid_end_bio_io(struct r1bio *r1_bio)
276 struct bio *bio = r1_bio->master_bio;
278 /* if nobody has done the final endio yet, do it now */
279 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
280 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
281 (bio_data_dir(bio) == WRITE) ? "write" : "read",
282 (unsigned long long) bio->bi_iter.bi_sector,
283 (unsigned long long) bio_end_sector(bio) - 1);
285 call_bio_endio(r1_bio);
291 * Update disk head position estimator based on IRQ completion info.
293 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
295 struct r1conf *conf = r1_bio->mddev->private;
297 conf->mirrors[disk].head_position =
298 r1_bio->sector + (r1_bio->sectors);
302 * Find the disk number which triggered given bio
304 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
307 struct r1conf *conf = r1_bio->mddev->private;
308 int raid_disks = conf->raid_disks;
310 for (mirror = 0; mirror < raid_disks * 2; mirror++)
311 if (r1_bio->bios[mirror] == bio)
314 BUG_ON(mirror == raid_disks * 2);
315 update_head_pos(mirror, r1_bio);
320 static void raid1_end_read_request(struct bio *bio)
322 int uptodate = !bio->bi_status;
323 struct r1bio *r1_bio = bio->bi_private;
324 struct r1conf *conf = r1_bio->mddev->private;
325 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
328 * this branch is our 'one mirror IO has finished' event handler:
330 update_head_pos(r1_bio->read_disk, r1_bio);
333 set_bit(R1BIO_Uptodate, &r1_bio->state);
334 else if (test_bit(FailFast, &rdev->flags) &&
335 test_bit(R1BIO_FailFast, &r1_bio->state))
336 /* This was a fail-fast read so we definitely
340 /* If all other devices have failed, we want to return
341 * the error upwards rather than fail the last device.
342 * Here we redefine "uptodate" to mean "Don't want to retry"
345 spin_lock_irqsave(&conf->device_lock, flags);
346 if (r1_bio->mddev->degraded == conf->raid_disks ||
347 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
348 test_bit(In_sync, &rdev->flags)))
350 spin_unlock_irqrestore(&conf->device_lock, flags);
354 raid_end_bio_io(r1_bio);
355 rdev_dec_pending(rdev, conf->mddev);
360 char b[BDEVNAME_SIZE];
361 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
363 bdevname(rdev->bdev, b),
364 (unsigned long long)r1_bio->sector);
365 set_bit(R1BIO_ReadError, &r1_bio->state);
366 reschedule_retry(r1_bio);
367 /* don't drop the reference on read_disk yet */
371 static void close_write(struct r1bio *r1_bio)
373 /* it really is the end of this request */
374 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
375 bio_free_pages(r1_bio->behind_master_bio);
376 bio_put(r1_bio->behind_master_bio);
377 r1_bio->behind_master_bio = NULL;
379 /* clear the bitmap if all writes complete successfully */
380 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
382 !test_bit(R1BIO_Degraded, &r1_bio->state),
383 test_bit(R1BIO_BehindIO, &r1_bio->state));
384 md_write_end(r1_bio->mddev);
387 static void r1_bio_write_done(struct r1bio *r1_bio)
389 if (!atomic_dec_and_test(&r1_bio->remaining))
392 if (test_bit(R1BIO_WriteError, &r1_bio->state))
393 reschedule_retry(r1_bio);
396 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
397 reschedule_retry(r1_bio);
399 raid_end_bio_io(r1_bio);
403 static void raid1_end_write_request(struct bio *bio)
405 struct r1bio *r1_bio = bio->bi_private;
406 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
407 struct r1conf *conf = r1_bio->mddev->private;
408 struct bio *to_put = NULL;
409 int mirror = find_bio_disk(r1_bio, bio);
410 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
413 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
416 * 'one mirror IO has finished' event handler:
418 if (bio->bi_status && !discard_error) {
419 set_bit(WriteErrorSeen, &rdev->flags);
420 if (!test_and_set_bit(WantReplacement, &rdev->flags))
421 set_bit(MD_RECOVERY_NEEDED, &
422 conf->mddev->recovery);
424 if (test_bit(FailFast, &rdev->flags) &&
425 (bio->bi_opf & MD_FAILFAST) &&
426 /* We never try FailFast to WriteMostly devices */
427 !test_bit(WriteMostly, &rdev->flags)) {
428 md_error(r1_bio->mddev, rdev);
429 if (!test_bit(Faulty, &rdev->flags))
430 /* This is the only remaining device,
431 * We need to retry the write without
434 set_bit(R1BIO_WriteError, &r1_bio->state);
436 /* Finished with this branch */
437 r1_bio->bios[mirror] = NULL;
441 set_bit(R1BIO_WriteError, &r1_bio->state);
444 * Set R1BIO_Uptodate in our master bio, so that we
445 * will return a good error code for to the higher
446 * levels even if IO on some other mirrored buffer
449 * The 'master' represents the composite IO operation
450 * to user-side. So if something waits for IO, then it
451 * will wait for the 'master' bio.
456 r1_bio->bios[mirror] = NULL;
459 * Do not set R1BIO_Uptodate if the current device is
460 * rebuilding or Faulty. This is because we cannot use
461 * such device for properly reading the data back (we could
462 * potentially use it, if the current write would have felt
463 * before rdev->recovery_offset, but for simplicity we don't
466 if (test_bit(In_sync, &rdev->flags) &&
467 !test_bit(Faulty, &rdev->flags))
468 set_bit(R1BIO_Uptodate, &r1_bio->state);
470 /* Maybe we can clear some bad blocks. */
471 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
472 &first_bad, &bad_sectors) && !discard_error) {
473 r1_bio->bios[mirror] = IO_MADE_GOOD;
474 set_bit(R1BIO_MadeGood, &r1_bio->state);
479 if (test_bit(WriteMostly, &rdev->flags))
480 atomic_dec(&r1_bio->behind_remaining);
483 * In behind mode, we ACK the master bio once the I/O
484 * has safely reached all non-writemostly
485 * disks. Setting the Returned bit ensures that this
486 * gets done only once -- we don't ever want to return
487 * -EIO here, instead we'll wait
489 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
490 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
491 /* Maybe we can return now */
492 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
493 struct bio *mbio = r1_bio->master_bio;
494 pr_debug("raid1: behind end write sectors"
496 (unsigned long long) mbio->bi_iter.bi_sector,
497 (unsigned long long) bio_end_sector(mbio) - 1);
498 call_bio_endio(r1_bio);
502 if (r1_bio->bios[mirror] == NULL)
503 rdev_dec_pending(rdev, conf->mddev);
506 * Let's see if all mirrored write operations have finished
509 r1_bio_write_done(r1_bio);
515 static sector_t align_to_barrier_unit_end(sector_t start_sector,
520 WARN_ON(sectors == 0);
522 * len is the number of sectors from start_sector to end of the
523 * barrier unit which start_sector belongs to.
525 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
535 * This routine returns the disk from which the requested read should
536 * be done. There is a per-array 'next expected sequential IO' sector
537 * number - if this matches on the next IO then we use the last disk.
538 * There is also a per-disk 'last know head position' sector that is
539 * maintained from IRQ contexts, both the normal and the resync IO
540 * completion handlers update this position correctly. If there is no
541 * perfect sequential match then we pick the disk whose head is closest.
543 * If there are 2 mirrors in the same 2 devices, performance degrades
544 * because position is mirror, not device based.
546 * The rdev for the device selected will have nr_pending incremented.
548 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
550 const sector_t this_sector = r1_bio->sector;
552 int best_good_sectors;
553 int best_disk, best_dist_disk, best_pending_disk;
557 unsigned int min_pending;
558 struct md_rdev *rdev;
560 int choose_next_idle;
564 * Check if we can balance. We can balance on the whole
565 * device if no resync is going on, or below the resync window.
566 * We take the first readable disk when above the resync window.
569 sectors = r1_bio->sectors;
572 best_dist = MaxSector;
573 best_pending_disk = -1;
574 min_pending = UINT_MAX;
575 best_good_sectors = 0;
577 choose_next_idle = 0;
578 clear_bit(R1BIO_FailFast, &r1_bio->state);
580 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
581 (mddev_is_clustered(conf->mddev) &&
582 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
583 this_sector + sectors)))
588 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
592 unsigned int pending;
595 rdev = rcu_dereference(conf->mirrors[disk].rdev);
596 if (r1_bio->bios[disk] == IO_BLOCKED
598 || test_bit(Faulty, &rdev->flags))
600 if (!test_bit(In_sync, &rdev->flags) &&
601 rdev->recovery_offset < this_sector + sectors)
603 if (test_bit(WriteMostly, &rdev->flags)) {
604 /* Don't balance among write-mostly, just
605 * use the first as a last resort */
606 if (best_dist_disk < 0) {
607 if (is_badblock(rdev, this_sector, sectors,
608 &first_bad, &bad_sectors)) {
609 if (first_bad <= this_sector)
610 /* Cannot use this */
612 best_good_sectors = first_bad - this_sector;
614 best_good_sectors = sectors;
615 best_dist_disk = disk;
616 best_pending_disk = disk;
620 /* This is a reasonable device to use. It might
623 if (is_badblock(rdev, this_sector, sectors,
624 &first_bad, &bad_sectors)) {
625 if (best_dist < MaxSector)
626 /* already have a better device */
628 if (first_bad <= this_sector) {
629 /* cannot read here. If this is the 'primary'
630 * device, then we must not read beyond
631 * bad_sectors from another device..
633 bad_sectors -= (this_sector - first_bad);
634 if (choose_first && sectors > bad_sectors)
635 sectors = bad_sectors;
636 if (best_good_sectors > sectors)
637 best_good_sectors = sectors;
640 sector_t good_sectors = first_bad - this_sector;
641 if (good_sectors > best_good_sectors) {
642 best_good_sectors = good_sectors;
650 if ((sectors > best_good_sectors) && (best_disk >= 0))
652 best_good_sectors = sectors;
656 /* At least two disks to choose from so failfast is OK */
657 set_bit(R1BIO_FailFast, &r1_bio->state);
659 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
660 has_nonrot_disk |= nonrot;
661 pending = atomic_read(&rdev->nr_pending);
662 dist = abs(this_sector - conf->mirrors[disk].head_position);
667 /* Don't change to another disk for sequential reads */
668 if (conf->mirrors[disk].next_seq_sect == this_sector
670 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
671 struct raid1_info *mirror = &conf->mirrors[disk];
675 * If buffered sequential IO size exceeds optimal
676 * iosize, check if there is idle disk. If yes, choose
677 * the idle disk. read_balance could already choose an
678 * idle disk before noticing it's a sequential IO in
679 * this disk. This doesn't matter because this disk
680 * will idle, next time it will be utilized after the
681 * first disk has IO size exceeds optimal iosize. In
682 * this way, iosize of the first disk will be optimal
683 * iosize at least. iosize of the second disk might be
684 * small, but not a big deal since when the second disk
685 * starts IO, the first disk is likely still busy.
687 if (nonrot && opt_iosize > 0 &&
688 mirror->seq_start != MaxSector &&
689 mirror->next_seq_sect > opt_iosize &&
690 mirror->next_seq_sect - opt_iosize >=
692 choose_next_idle = 1;
698 if (choose_next_idle)
701 if (min_pending > pending) {
702 min_pending = pending;
703 best_pending_disk = disk;
706 if (dist < best_dist) {
708 best_dist_disk = disk;
713 * If all disks are rotational, choose the closest disk. If any disk is
714 * non-rotational, choose the disk with less pending request even the
715 * disk is rotational, which might/might not be optimal for raids with
716 * mixed ratation/non-rotational disks depending on workload.
718 if (best_disk == -1) {
719 if (has_nonrot_disk || min_pending == 0)
720 best_disk = best_pending_disk;
722 best_disk = best_dist_disk;
725 if (best_disk >= 0) {
726 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
729 atomic_inc(&rdev->nr_pending);
730 sectors = best_good_sectors;
732 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
733 conf->mirrors[best_disk].seq_start = this_sector;
735 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
738 *max_sectors = sectors;
743 static int raid1_congested(struct mddev *mddev, int bits)
745 struct r1conf *conf = mddev->private;
748 if ((bits & (1 << WB_async_congested)) &&
749 conf->pending_count >= max_queued_requests)
753 for (i = 0; i < conf->raid_disks * 2; i++) {
754 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
755 if (rdev && !test_bit(Faulty, &rdev->flags)) {
756 struct request_queue *q = bdev_get_queue(rdev->bdev);
760 /* Note the '|| 1' - when read_balance prefers
761 * non-congested targets, it can be removed
763 if ((bits & (1 << WB_async_congested)) || 1)
764 ret |= bdi_congested(q->backing_dev_info, bits);
766 ret &= bdi_congested(q->backing_dev_info, bits);
773 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
775 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
776 md_bitmap_unplug(conf->mddev->bitmap);
777 wake_up(&conf->wait_barrier);
779 while (bio) { /* submit pending writes */
780 struct bio *next = bio->bi_next;
781 struct md_rdev *rdev = (void *)bio->bi_disk;
783 bio_set_dev(bio, rdev->bdev);
784 if (test_bit(Faulty, &rdev->flags)) {
786 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
787 !blk_queue_discard(bio->bi_disk->queue)))
791 generic_make_request(bio);
796 static void flush_pending_writes(struct r1conf *conf)
798 /* Any writes that have been queued but are awaiting
799 * bitmap updates get flushed here.
801 spin_lock_irq(&conf->device_lock);
803 if (conf->pending_bio_list.head) {
804 struct blk_plug plug;
807 bio = bio_list_get(&conf->pending_bio_list);
808 conf->pending_count = 0;
809 spin_unlock_irq(&conf->device_lock);
812 * As this is called in a wait_event() loop (see freeze_array),
813 * current->state might be TASK_UNINTERRUPTIBLE which will
814 * cause a warning when we prepare to wait again. As it is
815 * rare that this path is taken, it is perfectly safe to force
816 * us to go around the wait_event() loop again, so the warning
817 * is a false-positive. Silence the warning by resetting
820 __set_current_state(TASK_RUNNING);
821 blk_start_plug(&plug);
822 flush_bio_list(conf, bio);
823 blk_finish_plug(&plug);
825 spin_unlock_irq(&conf->device_lock);
829 * Sometimes we need to suspend IO while we do something else,
830 * either some resync/recovery, or reconfigure the array.
831 * To do this we raise a 'barrier'.
832 * The 'barrier' is a counter that can be raised multiple times
833 * to count how many activities are happening which preclude
835 * We can only raise the barrier if there is no pending IO.
836 * i.e. if nr_pending == 0.
837 * We choose only to raise the barrier if no-one is waiting for the
838 * barrier to go down. This means that as soon as an IO request
839 * is ready, no other operations which require a barrier will start
840 * until the IO request has had a chance.
842 * So: regular IO calls 'wait_barrier'. When that returns there
843 * is no backgroup IO happening, It must arrange to call
844 * allow_barrier when it has finished its IO.
845 * backgroup IO calls must call raise_barrier. Once that returns
846 * there is no normal IO happeing. It must arrange to call
847 * lower_barrier when the particular background IO completes.
849 static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
851 int idx = sector_to_idx(sector_nr);
853 spin_lock_irq(&conf->resync_lock);
855 /* Wait until no block IO is waiting */
856 wait_event_lock_irq(conf->wait_barrier,
857 !atomic_read(&conf->nr_waiting[idx]),
860 /* block any new IO from starting */
861 atomic_inc(&conf->barrier[idx]);
863 * In raise_barrier() we firstly increase conf->barrier[idx] then
864 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
865 * increase conf->nr_pending[idx] then check conf->barrier[idx].
866 * A memory barrier here to make sure conf->nr_pending[idx] won't
867 * be fetched before conf->barrier[idx] is increased. Otherwise
868 * there will be a race between raise_barrier() and _wait_barrier().
870 smp_mb__after_atomic();
872 /* For these conditions we must wait:
873 * A: while the array is in frozen state
874 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
875 * existing in corresponding I/O barrier bucket.
876 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
877 * max resync count which allowed on current I/O barrier bucket.
879 wait_event_lock_irq(conf->wait_barrier,
880 (!conf->array_frozen &&
881 !atomic_read(&conf->nr_pending[idx]) &&
882 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
883 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
886 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
887 atomic_dec(&conf->barrier[idx]);
888 spin_unlock_irq(&conf->resync_lock);
889 wake_up(&conf->wait_barrier);
893 atomic_inc(&conf->nr_sync_pending);
894 spin_unlock_irq(&conf->resync_lock);
899 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
901 int idx = sector_to_idx(sector_nr);
903 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
905 atomic_dec(&conf->barrier[idx]);
906 atomic_dec(&conf->nr_sync_pending);
907 wake_up(&conf->wait_barrier);
910 static void _wait_barrier(struct r1conf *conf, int idx)
913 * We need to increase conf->nr_pending[idx] very early here,
914 * then raise_barrier() can be blocked when it waits for
915 * conf->nr_pending[idx] to be 0. Then we can avoid holding
916 * conf->resync_lock when there is no barrier raised in same
917 * barrier unit bucket. Also if the array is frozen, I/O
918 * should be blocked until array is unfrozen.
920 atomic_inc(&conf->nr_pending[idx]);
922 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
923 * check conf->barrier[idx]. In raise_barrier() we firstly increase
924 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
925 * barrier is necessary here to make sure conf->barrier[idx] won't be
926 * fetched before conf->nr_pending[idx] is increased. Otherwise there
927 * will be a race between _wait_barrier() and raise_barrier().
929 smp_mb__after_atomic();
932 * Don't worry about checking two atomic_t variables at same time
933 * here. If during we check conf->barrier[idx], the array is
934 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
935 * 0, it is safe to return and make the I/O continue. Because the
936 * array is frozen, all I/O returned here will eventually complete
937 * or be queued, no race will happen. See code comment in
940 if (!READ_ONCE(conf->array_frozen) &&
941 !atomic_read(&conf->barrier[idx]))
945 * After holding conf->resync_lock, conf->nr_pending[idx]
946 * should be decreased before waiting for barrier to drop.
947 * Otherwise, we may encounter a race condition because
948 * raise_barrer() might be waiting for conf->nr_pending[idx]
949 * to be 0 at same time.
951 spin_lock_irq(&conf->resync_lock);
952 atomic_inc(&conf->nr_waiting[idx]);
953 atomic_dec(&conf->nr_pending[idx]);
955 * In case freeze_array() is waiting for
956 * get_unqueued_pending() == extra
958 wake_up(&conf->wait_barrier);
959 /* Wait for the barrier in same barrier unit bucket to drop. */
960 wait_event_lock_irq(conf->wait_barrier,
961 !conf->array_frozen &&
962 !atomic_read(&conf->barrier[idx]),
964 atomic_inc(&conf->nr_pending[idx]);
965 atomic_dec(&conf->nr_waiting[idx]);
966 spin_unlock_irq(&conf->resync_lock);
969 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
971 int idx = sector_to_idx(sector_nr);
974 * Very similar to _wait_barrier(). The difference is, for read
975 * I/O we don't need wait for sync I/O, but if the whole array
976 * is frozen, the read I/O still has to wait until the array is
977 * unfrozen. Since there is no ordering requirement with
978 * conf->barrier[idx] here, memory barrier is unnecessary as well.
980 atomic_inc(&conf->nr_pending[idx]);
982 if (!READ_ONCE(conf->array_frozen))
985 spin_lock_irq(&conf->resync_lock);
986 atomic_inc(&conf->nr_waiting[idx]);
987 atomic_dec(&conf->nr_pending[idx]);
989 * In case freeze_array() is waiting for
990 * get_unqueued_pending() == extra
992 wake_up(&conf->wait_barrier);
993 /* Wait for array to be unfrozen */
994 wait_event_lock_irq(conf->wait_barrier,
997 atomic_inc(&conf->nr_pending[idx]);
998 atomic_dec(&conf->nr_waiting[idx]);
999 spin_unlock_irq(&conf->resync_lock);
1002 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1004 int idx = sector_to_idx(sector_nr);
1006 _wait_barrier(conf, idx);
1009 static void _allow_barrier(struct r1conf *conf, int idx)
1011 atomic_dec(&conf->nr_pending[idx]);
1012 wake_up(&conf->wait_barrier);
1015 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1017 int idx = sector_to_idx(sector_nr);
1019 _allow_barrier(conf, idx);
1022 /* conf->resync_lock should be held */
1023 static int get_unqueued_pending(struct r1conf *conf)
1027 ret = atomic_read(&conf->nr_sync_pending);
1028 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1029 ret += atomic_read(&conf->nr_pending[idx]) -
1030 atomic_read(&conf->nr_queued[idx]);
1035 static void freeze_array(struct r1conf *conf, int extra)
1037 /* Stop sync I/O and normal I/O and wait for everything to
1039 * This is called in two situations:
1040 * 1) management command handlers (reshape, remove disk, quiesce).
1041 * 2) one normal I/O request failed.
1043 * After array_frozen is set to 1, new sync IO will be blocked at
1044 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1045 * or wait_read_barrier(). The flying I/Os will either complete or be
1046 * queued. When everything goes quite, there are only queued I/Os left.
1048 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1049 * barrier bucket index which this I/O request hits. When all sync and
1050 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1051 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1052 * in handle_read_error(), we may call freeze_array() before trying to
1053 * fix the read error. In this case, the error read I/O is not queued,
1054 * so get_unqueued_pending() == 1.
1056 * Therefore before this function returns, we need to wait until
1057 * get_unqueued_pendings(conf) gets equal to extra. For
1058 * normal I/O context, extra is 1, in rested situations extra is 0.
1060 spin_lock_irq(&conf->resync_lock);
1061 conf->array_frozen = 1;
1062 raid1_log(conf->mddev, "wait freeze");
1063 wait_event_lock_irq_cmd(
1065 get_unqueued_pending(conf) == extra,
1067 flush_pending_writes(conf));
1068 spin_unlock_irq(&conf->resync_lock);
1070 static void unfreeze_array(struct r1conf *conf)
1072 /* reverse the effect of the freeze */
1073 spin_lock_irq(&conf->resync_lock);
1074 conf->array_frozen = 0;
1075 spin_unlock_irq(&conf->resync_lock);
1076 wake_up(&conf->wait_barrier);
1079 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1082 int size = bio->bi_iter.bi_size;
1083 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1085 struct bio *behind_bio = NULL;
1087 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1091 /* discard op, we don't support writezero/writesame yet */
1092 if (!bio_has_data(bio)) {
1093 behind_bio->bi_iter.bi_size = size;
1097 behind_bio->bi_write_hint = bio->bi_write_hint;
1099 while (i < vcnt && size) {
1101 int len = min_t(int, PAGE_SIZE, size);
1103 page = alloc_page(GFP_NOIO);
1104 if (unlikely(!page))
1107 bio_add_page(behind_bio, page, len, 0);
1113 bio_copy_data(behind_bio, bio);
1115 r1_bio->behind_master_bio = behind_bio;
1116 set_bit(R1BIO_BehindIO, &r1_bio->state);
1121 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1122 bio->bi_iter.bi_size);
1123 bio_free_pages(behind_bio);
1124 bio_put(behind_bio);
1127 struct raid1_plug_cb {
1128 struct blk_plug_cb cb;
1129 struct bio_list pending;
1133 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1135 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1137 struct mddev *mddev = plug->cb.data;
1138 struct r1conf *conf = mddev->private;
1141 if (from_schedule || current->bio_list) {
1142 spin_lock_irq(&conf->device_lock);
1143 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1144 conf->pending_count += plug->pending_cnt;
1145 spin_unlock_irq(&conf->device_lock);
1146 wake_up(&conf->wait_barrier);
1147 md_wakeup_thread(mddev->thread);
1152 /* we aren't scheduling, so we can do the write-out directly. */
1153 bio = bio_list_get(&plug->pending);
1154 flush_bio_list(conf, bio);
1158 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1160 r1_bio->master_bio = bio;
1161 r1_bio->sectors = bio_sectors(bio);
1163 r1_bio->mddev = mddev;
1164 r1_bio->sector = bio->bi_iter.bi_sector;
1167 static inline struct r1bio *
1168 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1170 struct r1conf *conf = mddev->private;
1171 struct r1bio *r1_bio;
1173 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1174 /* Ensure no bio records IO_BLOCKED */
1175 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1176 init_r1bio(r1_bio, mddev, bio);
1180 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1181 int max_read_sectors, struct r1bio *r1_bio)
1183 struct r1conf *conf = mddev->private;
1184 struct raid1_info *mirror;
1185 struct bio *read_bio;
1186 struct bitmap *bitmap = mddev->bitmap;
1187 const int op = bio_op(bio);
1188 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1191 bool print_msg = !!r1_bio;
1192 char b[BDEVNAME_SIZE];
1195 * If r1_bio is set, we are blocking the raid1d thread
1196 * so there is a tiny risk of deadlock. So ask for
1197 * emergency memory if needed.
1199 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1202 /* Need to get the block device name carefully */
1203 struct md_rdev *rdev;
1205 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1207 bdevname(rdev->bdev, b);
1214 * Still need barrier for READ in case that whole
1217 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1220 r1_bio = alloc_r1bio(mddev, bio);
1222 init_r1bio(r1_bio, mddev, bio);
1223 r1_bio->sectors = max_read_sectors;
1226 * make_request() can abort the operation when read-ahead is being
1227 * used and no empty request is available.
1229 rdisk = read_balance(conf, r1_bio, &max_sectors);
1232 /* couldn't find anywhere to read from */
1234 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1237 (unsigned long long)r1_bio->sector);
1239 raid_end_bio_io(r1_bio);
1242 mirror = conf->mirrors + rdisk;
1245 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1247 (unsigned long long)r1_bio->sector,
1248 bdevname(mirror->rdev->bdev, b));
1250 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1253 * Reading from a write-mostly device must take care not to
1254 * over-take any writes that are 'behind'
1256 raid1_log(mddev, "wait behind writes");
1257 wait_event(bitmap->behind_wait,
1258 atomic_read(&bitmap->behind_writes) == 0);
1261 if (max_sectors < bio_sectors(bio)) {
1262 struct bio *split = bio_split(bio, max_sectors,
1263 gfp, &conf->bio_split);
1264 bio_chain(split, bio);
1265 generic_make_request(bio);
1267 r1_bio->master_bio = bio;
1268 r1_bio->sectors = max_sectors;
1271 r1_bio->read_disk = rdisk;
1273 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1275 r1_bio->bios[rdisk] = read_bio;
1277 read_bio->bi_iter.bi_sector = r1_bio->sector +
1278 mirror->rdev->data_offset;
1279 bio_set_dev(read_bio, mirror->rdev->bdev);
1280 read_bio->bi_end_io = raid1_end_read_request;
1281 bio_set_op_attrs(read_bio, op, do_sync);
1282 if (test_bit(FailFast, &mirror->rdev->flags) &&
1283 test_bit(R1BIO_FailFast, &r1_bio->state))
1284 read_bio->bi_opf |= MD_FAILFAST;
1285 read_bio->bi_private = r1_bio;
1288 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1289 disk_devt(mddev->gendisk), r1_bio->sector);
1291 generic_make_request(read_bio);
1294 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1295 int max_write_sectors)
1297 struct r1conf *conf = mddev->private;
1298 struct r1bio *r1_bio;
1300 struct bitmap *bitmap = mddev->bitmap;
1301 unsigned long flags;
1302 struct md_rdev *blocked_rdev;
1303 struct blk_plug_cb *cb;
1304 struct raid1_plug_cb *plug = NULL;
1308 if (mddev_is_clustered(mddev) &&
1309 md_cluster_ops->area_resyncing(mddev, WRITE,
1310 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1314 prepare_to_wait(&conf->wait_barrier,
1316 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1317 bio->bi_iter.bi_sector,
1318 bio_end_sector(bio)))
1322 finish_wait(&conf->wait_barrier, &w);
1326 * Register the new request and wait if the reconstruction
1327 * thread has put up a bar for new requests.
1328 * Continue immediately if no resync is active currently.
1330 wait_barrier(conf, bio->bi_iter.bi_sector);
1332 r1_bio = alloc_r1bio(mddev, bio);
1333 r1_bio->sectors = max_write_sectors;
1335 if (conf->pending_count >= max_queued_requests) {
1336 md_wakeup_thread(mddev->thread);
1337 raid1_log(mddev, "wait queued");
1338 wait_event(conf->wait_barrier,
1339 conf->pending_count < max_queued_requests);
1341 /* first select target devices under rcu_lock and
1342 * inc refcount on their rdev. Record them by setting
1344 * If there are known/acknowledged bad blocks on any device on
1345 * which we have seen a write error, we want to avoid writing those
1347 * This potentially requires several writes to write around
1348 * the bad blocks. Each set of writes gets it's own r1bio
1349 * with a set of bios attached.
1352 disks = conf->raid_disks * 2;
1354 blocked_rdev = NULL;
1356 max_sectors = r1_bio->sectors;
1357 for (i = 0; i < disks; i++) {
1358 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1359 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1360 atomic_inc(&rdev->nr_pending);
1361 blocked_rdev = rdev;
1364 r1_bio->bios[i] = NULL;
1365 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1366 if (i < conf->raid_disks)
1367 set_bit(R1BIO_Degraded, &r1_bio->state);
1371 atomic_inc(&rdev->nr_pending);
1372 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1377 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1378 &first_bad, &bad_sectors);
1380 /* mustn't write here until the bad block is
1382 set_bit(BlockedBadBlocks, &rdev->flags);
1383 blocked_rdev = rdev;
1386 if (is_bad && first_bad <= r1_bio->sector) {
1387 /* Cannot write here at all */
1388 bad_sectors -= (r1_bio->sector - first_bad);
1389 if (bad_sectors < max_sectors)
1390 /* mustn't write more than bad_sectors
1391 * to other devices yet
1393 max_sectors = bad_sectors;
1394 rdev_dec_pending(rdev, mddev);
1395 /* We don't set R1BIO_Degraded as that
1396 * only applies if the disk is
1397 * missing, so it might be re-added,
1398 * and we want to know to recover this
1400 * In this case the device is here,
1401 * and the fact that this chunk is not
1402 * in-sync is recorded in the bad
1408 int good_sectors = first_bad - r1_bio->sector;
1409 if (good_sectors < max_sectors)
1410 max_sectors = good_sectors;
1413 r1_bio->bios[i] = bio;
1417 if (unlikely(blocked_rdev)) {
1418 /* Wait for this device to become unblocked */
1421 for (j = 0; j < i; j++)
1422 if (r1_bio->bios[j])
1423 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1425 allow_barrier(conf, bio->bi_iter.bi_sector);
1426 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1427 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1428 wait_barrier(conf, bio->bi_iter.bi_sector);
1432 if (max_sectors < bio_sectors(bio)) {
1433 struct bio *split = bio_split(bio, max_sectors,
1434 GFP_NOIO, &conf->bio_split);
1435 bio_chain(split, bio);
1436 generic_make_request(bio);
1438 r1_bio->master_bio = bio;
1439 r1_bio->sectors = max_sectors;
1442 atomic_set(&r1_bio->remaining, 1);
1443 atomic_set(&r1_bio->behind_remaining, 0);
1447 for (i = 0; i < disks; i++) {
1448 struct bio *mbio = NULL;
1449 if (!r1_bio->bios[i])
1455 * Not if there are too many, or cannot
1456 * allocate memory, or a reader on WriteMostly
1457 * is waiting for behind writes to flush */
1459 (atomic_read(&bitmap->behind_writes)
1460 < mddev->bitmap_info.max_write_behind) &&
1461 !waitqueue_active(&bitmap->behind_wait)) {
1462 alloc_behind_master_bio(r1_bio, bio);
1465 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1466 test_bit(R1BIO_BehindIO, &r1_bio->state));
1470 if (r1_bio->behind_master_bio)
1471 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1472 GFP_NOIO, &mddev->bio_set);
1474 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1476 if (r1_bio->behind_master_bio) {
1477 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1478 atomic_inc(&r1_bio->behind_remaining);
1481 r1_bio->bios[i] = mbio;
1483 mbio->bi_iter.bi_sector = (r1_bio->sector +
1484 conf->mirrors[i].rdev->data_offset);
1485 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1486 mbio->bi_end_io = raid1_end_write_request;
1487 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1488 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1489 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1490 conf->raid_disks - mddev->degraded > 1)
1491 mbio->bi_opf |= MD_FAILFAST;
1492 mbio->bi_private = r1_bio;
1494 atomic_inc(&r1_bio->remaining);
1497 trace_block_bio_remap(mbio->bi_disk->queue,
1498 mbio, disk_devt(mddev->gendisk),
1500 /* flush_pending_writes() needs access to the rdev so...*/
1501 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1503 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1505 plug = container_of(cb, struct raid1_plug_cb, cb);
1509 bio_list_add(&plug->pending, mbio);
1510 plug->pending_cnt++;
1512 spin_lock_irqsave(&conf->device_lock, flags);
1513 bio_list_add(&conf->pending_bio_list, mbio);
1514 conf->pending_count++;
1515 spin_unlock_irqrestore(&conf->device_lock, flags);
1516 md_wakeup_thread(mddev->thread);
1520 r1_bio_write_done(r1_bio);
1522 /* In case raid1d snuck in to freeze_array */
1523 wake_up(&conf->wait_barrier);
1526 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1530 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1531 md_flush_request(mddev, bio);
1536 * There is a limit to the maximum size, but
1537 * the read/write handler might find a lower limit
1538 * due to bad blocks. To avoid multiple splits,
1539 * we pass the maximum number of sectors down
1540 * and let the lower level perform the split.
1542 sectors = align_to_barrier_unit_end(
1543 bio->bi_iter.bi_sector, bio_sectors(bio));
1545 if (bio_data_dir(bio) == READ)
1546 raid1_read_request(mddev, bio, sectors, NULL);
1548 if (!md_write_start(mddev,bio))
1550 raid1_write_request(mddev, bio, sectors);
1555 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1557 struct r1conf *conf = mddev->private;
1560 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1561 conf->raid_disks - mddev->degraded);
1563 for (i = 0; i < conf->raid_disks; i++) {
1564 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1565 seq_printf(seq, "%s",
1566 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1569 seq_printf(seq, "]");
1572 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1574 char b[BDEVNAME_SIZE];
1575 struct r1conf *conf = mddev->private;
1576 unsigned long flags;
1579 * If it is not operational, then we have already marked it as dead
1580 * else if it is the last working disks, ignore the error, let the
1581 * next level up know.
1582 * else mark the drive as failed
1584 spin_lock_irqsave(&conf->device_lock, flags);
1585 if (test_bit(In_sync, &rdev->flags)
1586 && (conf->raid_disks - mddev->degraded) == 1) {
1588 * Don't fail the drive, act as though we were just a
1589 * normal single drive.
1590 * However don't try a recovery from this drive as
1591 * it is very likely to fail.
1593 conf->recovery_disabled = mddev->recovery_disabled;
1594 spin_unlock_irqrestore(&conf->device_lock, flags);
1597 set_bit(Blocked, &rdev->flags);
1598 if (test_and_clear_bit(In_sync, &rdev->flags))
1600 set_bit(Faulty, &rdev->flags);
1601 spin_unlock_irqrestore(&conf->device_lock, flags);
1603 * if recovery is running, make sure it aborts.
1605 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1606 set_mask_bits(&mddev->sb_flags, 0,
1607 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1608 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1609 "md/raid1:%s: Operation continuing on %d devices.\n",
1610 mdname(mddev), bdevname(rdev->bdev, b),
1611 mdname(mddev), conf->raid_disks - mddev->degraded);
1614 static void print_conf(struct r1conf *conf)
1618 pr_debug("RAID1 conf printout:\n");
1620 pr_debug("(!conf)\n");
1623 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1627 for (i = 0; i < conf->raid_disks; i++) {
1628 char b[BDEVNAME_SIZE];
1629 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1631 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1632 i, !test_bit(In_sync, &rdev->flags),
1633 !test_bit(Faulty, &rdev->flags),
1634 bdevname(rdev->bdev,b));
1639 static void close_sync(struct r1conf *conf)
1643 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1644 _wait_barrier(conf, idx);
1645 _allow_barrier(conf, idx);
1648 mempool_exit(&conf->r1buf_pool);
1651 static int raid1_spare_active(struct mddev *mddev)
1654 struct r1conf *conf = mddev->private;
1656 unsigned long flags;
1659 * Find all failed disks within the RAID1 configuration
1660 * and mark them readable.
1661 * Called under mddev lock, so rcu protection not needed.
1662 * device_lock used to avoid races with raid1_end_read_request
1663 * which expects 'In_sync' flags and ->degraded to be consistent.
1665 spin_lock_irqsave(&conf->device_lock, flags);
1666 for (i = 0; i < conf->raid_disks; i++) {
1667 struct md_rdev *rdev = conf->mirrors[i].rdev;
1668 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1670 && !test_bit(Candidate, &repl->flags)
1671 && repl->recovery_offset == MaxSector
1672 && !test_bit(Faulty, &repl->flags)
1673 && !test_and_set_bit(In_sync, &repl->flags)) {
1674 /* replacement has just become active */
1676 !test_and_clear_bit(In_sync, &rdev->flags))
1679 /* Replaced device not technically
1680 * faulty, but we need to be sure
1681 * it gets removed and never re-added
1683 set_bit(Faulty, &rdev->flags);
1684 sysfs_notify_dirent_safe(
1689 && rdev->recovery_offset == MaxSector
1690 && !test_bit(Faulty, &rdev->flags)
1691 && !test_and_set_bit(In_sync, &rdev->flags)) {
1693 sysfs_notify_dirent_safe(rdev->sysfs_state);
1696 mddev->degraded -= count;
1697 spin_unlock_irqrestore(&conf->device_lock, flags);
1703 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1705 struct r1conf *conf = mddev->private;
1708 struct raid1_info *p;
1710 int last = conf->raid_disks - 1;
1712 if (mddev->recovery_disabled == conf->recovery_disabled)
1715 if (md_integrity_add_rdev(rdev, mddev))
1718 if (rdev->raid_disk >= 0)
1719 first = last = rdev->raid_disk;
1722 * find the disk ... but prefer rdev->saved_raid_disk
1725 if (rdev->saved_raid_disk >= 0 &&
1726 rdev->saved_raid_disk >= first &&
1727 rdev->saved_raid_disk < conf->raid_disks &&
1728 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1729 first = last = rdev->saved_raid_disk;
1731 for (mirror = first; mirror <= last; mirror++) {
1732 p = conf->mirrors+mirror;
1736 disk_stack_limits(mddev->gendisk, rdev->bdev,
1737 rdev->data_offset << 9);
1739 p->head_position = 0;
1740 rdev->raid_disk = mirror;
1742 /* As all devices are equivalent, we don't need a full recovery
1743 * if this was recently any drive of the array
1745 if (rdev->saved_raid_disk < 0)
1747 rcu_assign_pointer(p->rdev, rdev);
1750 if (test_bit(WantReplacement, &p->rdev->flags) &&
1751 p[conf->raid_disks].rdev == NULL) {
1752 /* Add this device as a replacement */
1753 clear_bit(In_sync, &rdev->flags);
1754 set_bit(Replacement, &rdev->flags);
1755 rdev->raid_disk = mirror;
1758 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1762 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1763 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1768 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1770 struct r1conf *conf = mddev->private;
1772 int number = rdev->raid_disk;
1773 struct raid1_info *p = conf->mirrors + number;
1775 if (rdev != p->rdev)
1776 p = conf->mirrors + conf->raid_disks + number;
1779 if (rdev == p->rdev) {
1780 if (test_bit(In_sync, &rdev->flags) ||
1781 atomic_read(&rdev->nr_pending)) {
1785 /* Only remove non-faulty devices if recovery
1788 if (!test_bit(Faulty, &rdev->flags) &&
1789 mddev->recovery_disabled != conf->recovery_disabled &&
1790 mddev->degraded < conf->raid_disks) {
1795 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1797 if (atomic_read(&rdev->nr_pending)) {
1798 /* lost the race, try later */
1804 if (conf->mirrors[conf->raid_disks + number].rdev) {
1805 /* We just removed a device that is being replaced.
1806 * Move down the replacement. We drain all IO before
1807 * doing this to avoid confusion.
1809 struct md_rdev *repl =
1810 conf->mirrors[conf->raid_disks + number].rdev;
1811 freeze_array(conf, 0);
1812 if (atomic_read(&repl->nr_pending)) {
1813 /* It means that some queued IO of retry_list
1814 * hold repl. Thus, we cannot set replacement
1815 * as NULL, avoiding rdev NULL pointer
1816 * dereference in sync_request_write and
1817 * handle_write_finished.
1820 unfreeze_array(conf);
1823 clear_bit(Replacement, &repl->flags);
1825 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1826 unfreeze_array(conf);
1829 clear_bit(WantReplacement, &rdev->flags);
1830 err = md_integrity_register(mddev);
1838 static void end_sync_read(struct bio *bio)
1840 struct r1bio *r1_bio = get_resync_r1bio(bio);
1842 update_head_pos(r1_bio->read_disk, r1_bio);
1845 * we have read a block, now it needs to be re-written,
1846 * or re-read if the read failed.
1847 * We don't do much here, just schedule handling by raid1d
1849 if (!bio->bi_status)
1850 set_bit(R1BIO_Uptodate, &r1_bio->state);
1852 if (atomic_dec_and_test(&r1_bio->remaining))
1853 reschedule_retry(r1_bio);
1856 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1858 sector_t sync_blocks = 0;
1859 sector_t s = r1_bio->sector;
1860 long sectors_to_go = r1_bio->sectors;
1862 /* make sure these bits don't get cleared. */
1864 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1866 sectors_to_go -= sync_blocks;
1867 } while (sectors_to_go > 0);
1870 static void end_sync_write(struct bio *bio)
1872 int uptodate = !bio->bi_status;
1873 struct r1bio *r1_bio = get_resync_r1bio(bio);
1874 struct mddev *mddev = r1_bio->mddev;
1875 struct r1conf *conf = mddev->private;
1878 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1881 abort_sync_write(mddev, r1_bio);
1882 set_bit(WriteErrorSeen, &rdev->flags);
1883 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1884 set_bit(MD_RECOVERY_NEEDED, &
1886 set_bit(R1BIO_WriteError, &r1_bio->state);
1887 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1888 &first_bad, &bad_sectors) &&
1889 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1892 &first_bad, &bad_sectors)
1894 set_bit(R1BIO_MadeGood, &r1_bio->state);
1896 if (atomic_dec_and_test(&r1_bio->remaining)) {
1897 int s = r1_bio->sectors;
1898 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1899 test_bit(R1BIO_WriteError, &r1_bio->state))
1900 reschedule_retry(r1_bio);
1903 md_done_sync(mddev, s, uptodate);
1908 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1909 int sectors, struct page *page, int rw)
1911 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1915 set_bit(WriteErrorSeen, &rdev->flags);
1916 if (!test_and_set_bit(WantReplacement,
1918 set_bit(MD_RECOVERY_NEEDED, &
1919 rdev->mddev->recovery);
1921 /* need to record an error - either for the block or the device */
1922 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1923 md_error(rdev->mddev, rdev);
1927 static int fix_sync_read_error(struct r1bio *r1_bio)
1929 /* Try some synchronous reads of other devices to get
1930 * good data, much like with normal read errors. Only
1931 * read into the pages we already have so we don't
1932 * need to re-issue the read request.
1933 * We don't need to freeze the array, because being in an
1934 * active sync request, there is no normal IO, and
1935 * no overlapping syncs.
1936 * We don't need to check is_badblock() again as we
1937 * made sure that anything with a bad block in range
1938 * will have bi_end_io clear.
1940 struct mddev *mddev = r1_bio->mddev;
1941 struct r1conf *conf = mddev->private;
1942 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1943 struct page **pages = get_resync_pages(bio)->pages;
1944 sector_t sect = r1_bio->sector;
1945 int sectors = r1_bio->sectors;
1947 struct md_rdev *rdev;
1949 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1950 if (test_bit(FailFast, &rdev->flags)) {
1951 /* Don't try recovering from here - just fail it
1952 * ... unless it is the last working device of course */
1953 md_error(mddev, rdev);
1954 if (test_bit(Faulty, &rdev->flags))
1955 /* Don't try to read from here, but make sure
1956 * put_buf does it's thing
1958 bio->bi_end_io = end_sync_write;
1963 int d = r1_bio->read_disk;
1967 if (s > (PAGE_SIZE>>9))
1970 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1971 /* No rcu protection needed here devices
1972 * can only be removed when no resync is
1973 * active, and resync is currently active
1975 rdev = conf->mirrors[d].rdev;
1976 if (sync_page_io(rdev, sect, s<<9,
1978 REQ_OP_READ, 0, false)) {
1984 if (d == conf->raid_disks * 2)
1986 } while (!success && d != r1_bio->read_disk);
1989 char b[BDEVNAME_SIZE];
1991 /* Cannot read from anywhere, this block is lost.
1992 * Record a bad block on each device. If that doesn't
1993 * work just disable and interrupt the recovery.
1994 * Don't fail devices as that won't really help.
1996 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1997 mdname(mddev), bio_devname(bio, b),
1998 (unsigned long long)r1_bio->sector);
1999 for (d = 0; d < conf->raid_disks * 2; d++) {
2000 rdev = conf->mirrors[d].rdev;
2001 if (!rdev || test_bit(Faulty, &rdev->flags))
2003 if (!rdev_set_badblocks(rdev, sect, s, 0))
2007 conf->recovery_disabled =
2008 mddev->recovery_disabled;
2009 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2010 md_done_sync(mddev, r1_bio->sectors, 0);
2022 /* write it back and re-read */
2023 while (d != r1_bio->read_disk) {
2025 d = conf->raid_disks * 2;
2027 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2029 rdev = conf->mirrors[d].rdev;
2030 if (r1_sync_page_io(rdev, sect, s,
2033 r1_bio->bios[d]->bi_end_io = NULL;
2034 rdev_dec_pending(rdev, mddev);
2038 while (d != r1_bio->read_disk) {
2040 d = conf->raid_disks * 2;
2042 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2044 rdev = conf->mirrors[d].rdev;
2045 if (r1_sync_page_io(rdev, sect, s,
2048 atomic_add(s, &rdev->corrected_errors);
2054 set_bit(R1BIO_Uptodate, &r1_bio->state);
2059 static void process_checks(struct r1bio *r1_bio)
2061 /* We have read all readable devices. If we haven't
2062 * got the block, then there is no hope left.
2063 * If we have, then we want to do a comparison
2064 * and skip the write if everything is the same.
2065 * If any blocks failed to read, then we need to
2066 * attempt an over-write
2068 struct mddev *mddev = r1_bio->mddev;
2069 struct r1conf *conf = mddev->private;
2074 /* Fix variable parts of all bios */
2075 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2076 for (i = 0; i < conf->raid_disks * 2; i++) {
2077 blk_status_t status;
2078 struct bio *b = r1_bio->bios[i];
2079 struct resync_pages *rp = get_resync_pages(b);
2080 if (b->bi_end_io != end_sync_read)
2082 /* fixup the bio for reuse, but preserve errno */
2083 status = b->bi_status;
2085 b->bi_status = status;
2086 b->bi_iter.bi_sector = r1_bio->sector +
2087 conf->mirrors[i].rdev->data_offset;
2088 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2089 b->bi_end_io = end_sync_read;
2090 rp->raid_bio = r1_bio;
2093 /* initialize bvec table again */
2094 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2096 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2097 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2098 !r1_bio->bios[primary]->bi_status) {
2099 r1_bio->bios[primary]->bi_end_io = NULL;
2100 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2103 r1_bio->read_disk = primary;
2104 for (i = 0; i < conf->raid_disks * 2; i++) {
2106 struct bio *pbio = r1_bio->bios[primary];
2107 struct bio *sbio = r1_bio->bios[i];
2108 blk_status_t status = sbio->bi_status;
2109 struct page **ppages = get_resync_pages(pbio)->pages;
2110 struct page **spages = get_resync_pages(sbio)->pages;
2112 int page_len[RESYNC_PAGES] = { 0 };
2113 struct bvec_iter_all iter_all;
2115 if (sbio->bi_end_io != end_sync_read)
2117 /* Now we can 'fixup' the error value */
2118 sbio->bi_status = 0;
2120 bio_for_each_segment_all(bi, sbio, iter_all)
2121 page_len[j++] = bi->bv_len;
2124 for (j = vcnt; j-- ; ) {
2125 if (memcmp(page_address(ppages[j]),
2126 page_address(spages[j]),
2133 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2134 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2136 /* No need to write to this device. */
2137 sbio->bi_end_io = NULL;
2138 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2142 bio_copy_data(sbio, pbio);
2146 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2148 struct r1conf *conf = mddev->private;
2150 int disks = conf->raid_disks * 2;
2153 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2154 /* ouch - failed to read all of that. */
2155 if (!fix_sync_read_error(r1_bio))
2158 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2159 process_checks(r1_bio);
2164 atomic_set(&r1_bio->remaining, 1);
2165 for (i = 0; i < disks ; i++) {
2166 wbio = r1_bio->bios[i];
2167 if (wbio->bi_end_io == NULL ||
2168 (wbio->bi_end_io == end_sync_read &&
2169 (i == r1_bio->read_disk ||
2170 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2172 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2173 abort_sync_write(mddev, r1_bio);
2177 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2178 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2179 wbio->bi_opf |= MD_FAILFAST;
2181 wbio->bi_end_io = end_sync_write;
2182 atomic_inc(&r1_bio->remaining);
2183 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2185 generic_make_request(wbio);
2188 if (atomic_dec_and_test(&r1_bio->remaining)) {
2189 /* if we're here, all write(s) have completed, so clean up */
2190 int s = r1_bio->sectors;
2191 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2192 test_bit(R1BIO_WriteError, &r1_bio->state))
2193 reschedule_retry(r1_bio);
2196 md_done_sync(mddev, s, 1);
2202 * This is a kernel thread which:
2204 * 1. Retries failed read operations on working mirrors.
2205 * 2. Updates the raid superblock when problems encounter.
2206 * 3. Performs writes following reads for array synchronising.
2209 static void fix_read_error(struct r1conf *conf, int read_disk,
2210 sector_t sect, int sectors)
2212 struct mddev *mddev = conf->mddev;
2218 struct md_rdev *rdev;
2220 if (s > (PAGE_SIZE>>9))
2228 rdev = rcu_dereference(conf->mirrors[d].rdev);
2230 (test_bit(In_sync, &rdev->flags) ||
2231 (!test_bit(Faulty, &rdev->flags) &&
2232 rdev->recovery_offset >= sect + s)) &&
2233 is_badblock(rdev, sect, s,
2234 &first_bad, &bad_sectors) == 0) {
2235 atomic_inc(&rdev->nr_pending);
2237 if (sync_page_io(rdev, sect, s<<9,
2238 conf->tmppage, REQ_OP_READ, 0, false))
2240 rdev_dec_pending(rdev, mddev);
2246 if (d == conf->raid_disks * 2)
2248 } while (!success && d != read_disk);
2251 /* Cannot read from anywhere - mark it bad */
2252 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2253 if (!rdev_set_badblocks(rdev, sect, s, 0))
2254 md_error(mddev, rdev);
2257 /* write it back and re-read */
2259 while (d != read_disk) {
2261 d = conf->raid_disks * 2;
2264 rdev = rcu_dereference(conf->mirrors[d].rdev);
2266 !test_bit(Faulty, &rdev->flags)) {
2267 atomic_inc(&rdev->nr_pending);
2269 r1_sync_page_io(rdev, sect, s,
2270 conf->tmppage, WRITE);
2271 rdev_dec_pending(rdev, mddev);
2276 while (d != read_disk) {
2277 char b[BDEVNAME_SIZE];
2279 d = conf->raid_disks * 2;
2282 rdev = rcu_dereference(conf->mirrors[d].rdev);
2284 !test_bit(Faulty, &rdev->flags)) {
2285 atomic_inc(&rdev->nr_pending);
2287 if (r1_sync_page_io(rdev, sect, s,
2288 conf->tmppage, READ)) {
2289 atomic_add(s, &rdev->corrected_errors);
2290 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2292 (unsigned long long)(sect +
2294 bdevname(rdev->bdev, b));
2296 rdev_dec_pending(rdev, mddev);
2305 static int narrow_write_error(struct r1bio *r1_bio, int i)
2307 struct mddev *mddev = r1_bio->mddev;
2308 struct r1conf *conf = mddev->private;
2309 struct md_rdev *rdev = conf->mirrors[i].rdev;
2311 /* bio has the data to be written to device 'i' where
2312 * we just recently had a write error.
2313 * We repeatedly clone the bio and trim down to one block,
2314 * then try the write. Where the write fails we record
2316 * It is conceivable that the bio doesn't exactly align with
2317 * blocks. We must handle this somehow.
2319 * We currently own a reference on the rdev.
2325 int sect_to_write = r1_bio->sectors;
2328 if (rdev->badblocks.shift < 0)
2331 block_sectors = roundup(1 << rdev->badblocks.shift,
2332 bdev_logical_block_size(rdev->bdev) >> 9);
2333 sector = r1_bio->sector;
2334 sectors = ((sector + block_sectors)
2335 & ~(sector_t)(block_sectors - 1))
2338 while (sect_to_write) {
2340 if (sectors > sect_to_write)
2341 sectors = sect_to_write;
2342 /* Write at 'sector' for 'sectors'*/
2344 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2345 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2349 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2353 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2354 wbio->bi_iter.bi_sector = r1_bio->sector;
2355 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2357 bio_trim(wbio, sector - r1_bio->sector, sectors);
2358 wbio->bi_iter.bi_sector += rdev->data_offset;
2359 bio_set_dev(wbio, rdev->bdev);
2361 if (submit_bio_wait(wbio) < 0)
2363 ok = rdev_set_badblocks(rdev, sector,
2368 sect_to_write -= sectors;
2370 sectors = block_sectors;
2375 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2378 int s = r1_bio->sectors;
2379 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2380 struct md_rdev *rdev = conf->mirrors[m].rdev;
2381 struct bio *bio = r1_bio->bios[m];
2382 if (bio->bi_end_io == NULL)
2384 if (!bio->bi_status &&
2385 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2386 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2388 if (bio->bi_status &&
2389 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2390 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2391 md_error(conf->mddev, rdev);
2395 md_done_sync(conf->mddev, s, 1);
2398 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2403 for (m = 0; m < conf->raid_disks * 2 ; m++)
2404 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2405 struct md_rdev *rdev = conf->mirrors[m].rdev;
2406 rdev_clear_badblocks(rdev,
2408 r1_bio->sectors, 0);
2409 rdev_dec_pending(rdev, conf->mddev);
2410 } else if (r1_bio->bios[m] != NULL) {
2411 /* This drive got a write error. We need to
2412 * narrow down and record precise write
2416 if (!narrow_write_error(r1_bio, m)) {
2417 md_error(conf->mddev,
2418 conf->mirrors[m].rdev);
2419 /* an I/O failed, we can't clear the bitmap */
2420 set_bit(R1BIO_Degraded, &r1_bio->state);
2422 rdev_dec_pending(conf->mirrors[m].rdev,
2426 spin_lock_irq(&conf->device_lock);
2427 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2428 idx = sector_to_idx(r1_bio->sector);
2429 atomic_inc(&conf->nr_queued[idx]);
2430 spin_unlock_irq(&conf->device_lock);
2432 * In case freeze_array() is waiting for condition
2433 * get_unqueued_pending() == extra to be true.
2435 wake_up(&conf->wait_barrier);
2436 md_wakeup_thread(conf->mddev->thread);
2438 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2439 close_write(r1_bio);
2440 raid_end_bio_io(r1_bio);
2444 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2446 struct mddev *mddev = conf->mddev;
2448 struct md_rdev *rdev;
2450 clear_bit(R1BIO_ReadError, &r1_bio->state);
2451 /* we got a read error. Maybe the drive is bad. Maybe just
2452 * the block and we can fix it.
2453 * We freeze all other IO, and try reading the block from
2454 * other devices. When we find one, we re-write
2455 * and check it that fixes the read error.
2456 * This is all done synchronously while the array is
2460 bio = r1_bio->bios[r1_bio->read_disk];
2462 r1_bio->bios[r1_bio->read_disk] = NULL;
2464 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2466 && !test_bit(FailFast, &rdev->flags)) {
2467 freeze_array(conf, 1);
2468 fix_read_error(conf, r1_bio->read_disk,
2469 r1_bio->sector, r1_bio->sectors);
2470 unfreeze_array(conf);
2471 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2472 md_error(mddev, rdev);
2474 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2477 rdev_dec_pending(rdev, conf->mddev);
2478 allow_barrier(conf, r1_bio->sector);
2479 bio = r1_bio->master_bio;
2481 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2483 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2486 static void raid1d(struct md_thread *thread)
2488 struct mddev *mddev = thread->mddev;
2489 struct r1bio *r1_bio;
2490 unsigned long flags;
2491 struct r1conf *conf = mddev->private;
2492 struct list_head *head = &conf->retry_list;
2493 struct blk_plug plug;
2496 md_check_recovery(mddev);
2498 if (!list_empty_careful(&conf->bio_end_io_list) &&
2499 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2501 spin_lock_irqsave(&conf->device_lock, flags);
2502 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2503 list_splice_init(&conf->bio_end_io_list, &tmp);
2504 spin_unlock_irqrestore(&conf->device_lock, flags);
2505 while (!list_empty(&tmp)) {
2506 r1_bio = list_first_entry(&tmp, struct r1bio,
2508 list_del(&r1_bio->retry_list);
2509 idx = sector_to_idx(r1_bio->sector);
2510 atomic_dec(&conf->nr_queued[idx]);
2511 if (mddev->degraded)
2512 set_bit(R1BIO_Degraded, &r1_bio->state);
2513 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2514 close_write(r1_bio);
2515 raid_end_bio_io(r1_bio);
2519 blk_start_plug(&plug);
2522 flush_pending_writes(conf);
2524 spin_lock_irqsave(&conf->device_lock, flags);
2525 if (list_empty(head)) {
2526 spin_unlock_irqrestore(&conf->device_lock, flags);
2529 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2530 list_del(head->prev);
2531 idx = sector_to_idx(r1_bio->sector);
2532 atomic_dec(&conf->nr_queued[idx]);
2533 spin_unlock_irqrestore(&conf->device_lock, flags);
2535 mddev = r1_bio->mddev;
2536 conf = mddev->private;
2537 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2538 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2539 test_bit(R1BIO_WriteError, &r1_bio->state))
2540 handle_sync_write_finished(conf, r1_bio);
2542 sync_request_write(mddev, r1_bio);
2543 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2544 test_bit(R1BIO_WriteError, &r1_bio->state))
2545 handle_write_finished(conf, r1_bio);
2546 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2547 handle_read_error(conf, r1_bio);
2552 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2553 md_check_recovery(mddev);
2555 blk_finish_plug(&plug);
2558 static int init_resync(struct r1conf *conf)
2562 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2563 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2565 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2566 r1buf_pool_free, conf->poolinfo);
2569 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2571 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2572 struct resync_pages *rps;
2576 for (i = conf->poolinfo->raid_disks; i--; ) {
2577 bio = r1bio->bios[i];
2578 rps = bio->bi_private;
2580 bio->bi_private = rps;
2582 r1bio->master_bio = NULL;
2587 * perform a "sync" on one "block"
2589 * We need to make sure that no normal I/O request - particularly write
2590 * requests - conflict with active sync requests.
2592 * This is achieved by tracking pending requests and a 'barrier' concept
2593 * that can be installed to exclude normal IO requests.
2596 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2599 struct r1conf *conf = mddev->private;
2600 struct r1bio *r1_bio;
2602 sector_t max_sector, nr_sectors;
2606 int write_targets = 0, read_targets = 0;
2607 sector_t sync_blocks;
2608 int still_degraded = 0;
2609 int good_sectors = RESYNC_SECTORS;
2610 int min_bad = 0; /* number of sectors that are bad in all devices */
2611 int idx = sector_to_idx(sector_nr);
2614 if (!mempool_initialized(&conf->r1buf_pool))
2615 if (init_resync(conf))
2618 max_sector = mddev->dev_sectors;
2619 if (sector_nr >= max_sector) {
2620 /* If we aborted, we need to abort the
2621 * sync on the 'current' bitmap chunk (there will
2622 * only be one in raid1 resync.
2623 * We can find the current addess in mddev->curr_resync
2625 if (mddev->curr_resync < max_sector) /* aborted */
2626 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2628 else /* completed sync */
2631 md_bitmap_close_sync(mddev->bitmap);
2634 if (mddev_is_clustered(mddev)) {
2635 conf->cluster_sync_low = 0;
2636 conf->cluster_sync_high = 0;
2641 if (mddev->bitmap == NULL &&
2642 mddev->recovery_cp == MaxSector &&
2643 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2644 conf->fullsync == 0) {
2646 return max_sector - sector_nr;
2648 /* before building a request, check if we can skip these blocks..
2649 * This call the bitmap_start_sync doesn't actually record anything
2651 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2652 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2653 /* We can skip this block, and probably several more */
2659 * If there is non-resync activity waiting for a turn, then let it
2660 * though before starting on this new sync request.
2662 if (atomic_read(&conf->nr_waiting[idx]))
2663 schedule_timeout_uninterruptible(1);
2665 /* we are incrementing sector_nr below. To be safe, we check against
2666 * sector_nr + two times RESYNC_SECTORS
2669 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2670 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2673 if (raise_barrier(conf, sector_nr))
2676 r1_bio = raid1_alloc_init_r1buf(conf);
2680 * If we get a correctably read error during resync or recovery,
2681 * we might want to read from a different device. So we
2682 * flag all drives that could conceivably be read from for READ,
2683 * and any others (which will be non-In_sync devices) for WRITE.
2684 * If a read fails, we try reading from something else for which READ
2688 r1_bio->mddev = mddev;
2689 r1_bio->sector = sector_nr;
2691 set_bit(R1BIO_IsSync, &r1_bio->state);
2692 /* make sure good_sectors won't go across barrier unit boundary */
2693 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2695 for (i = 0; i < conf->raid_disks * 2; i++) {
2696 struct md_rdev *rdev;
2697 bio = r1_bio->bios[i];
2699 rdev = rcu_dereference(conf->mirrors[i].rdev);
2701 test_bit(Faulty, &rdev->flags)) {
2702 if (i < conf->raid_disks)
2704 } else if (!test_bit(In_sync, &rdev->flags)) {
2705 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2706 bio->bi_end_io = end_sync_write;
2709 /* may need to read from here */
2710 sector_t first_bad = MaxSector;
2713 if (is_badblock(rdev, sector_nr, good_sectors,
2714 &first_bad, &bad_sectors)) {
2715 if (first_bad > sector_nr)
2716 good_sectors = first_bad - sector_nr;
2718 bad_sectors -= (sector_nr - first_bad);
2720 min_bad > bad_sectors)
2721 min_bad = bad_sectors;
2724 if (sector_nr < first_bad) {
2725 if (test_bit(WriteMostly, &rdev->flags)) {
2732 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2733 bio->bi_end_io = end_sync_read;
2735 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2736 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2737 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2739 * The device is suitable for reading (InSync),
2740 * but has bad block(s) here. Let's try to correct them,
2741 * if we are doing resync or repair. Otherwise, leave
2742 * this device alone for this sync request.
2744 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2745 bio->bi_end_io = end_sync_write;
2749 if (bio->bi_end_io) {
2750 atomic_inc(&rdev->nr_pending);
2751 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2752 bio_set_dev(bio, rdev->bdev);
2753 if (test_bit(FailFast, &rdev->flags))
2754 bio->bi_opf |= MD_FAILFAST;
2760 r1_bio->read_disk = disk;
2762 if (read_targets == 0 && min_bad > 0) {
2763 /* These sectors are bad on all InSync devices, so we
2764 * need to mark them bad on all write targets
2767 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2768 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2769 struct md_rdev *rdev = conf->mirrors[i].rdev;
2770 ok = rdev_set_badblocks(rdev, sector_nr,
2774 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2779 /* Cannot record the badblocks, so need to
2781 * If there are multiple read targets, could just
2782 * fail the really bad ones ???
2784 conf->recovery_disabled = mddev->recovery_disabled;
2785 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2791 if (min_bad > 0 && min_bad < good_sectors) {
2792 /* only resync enough to reach the next bad->good
2794 good_sectors = min_bad;
2797 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2798 /* extra read targets are also write targets */
2799 write_targets += read_targets-1;
2801 if (write_targets == 0 || read_targets == 0) {
2802 /* There is nowhere to write, so all non-sync
2803 * drives must be failed - so we are finished
2807 max_sector = sector_nr + min_bad;
2808 rv = max_sector - sector_nr;
2814 if (max_sector > mddev->resync_max)
2815 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2816 if (max_sector > sector_nr + good_sectors)
2817 max_sector = sector_nr + good_sectors;
2822 int len = PAGE_SIZE;
2823 if (sector_nr + (len>>9) > max_sector)
2824 len = (max_sector - sector_nr) << 9;
2827 if (sync_blocks == 0) {
2828 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2829 &sync_blocks, still_degraded) &&
2831 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2833 if ((len >> 9) > sync_blocks)
2834 len = sync_blocks<<9;
2837 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2838 struct resync_pages *rp;
2840 bio = r1_bio->bios[i];
2841 rp = get_resync_pages(bio);
2842 if (bio->bi_end_io) {
2843 page = resync_fetch_page(rp, page_idx);
2846 * won't fail because the vec table is big
2847 * enough to hold all these pages
2849 bio_add_page(bio, page, len, 0);
2852 nr_sectors += len>>9;
2853 sector_nr += len>>9;
2854 sync_blocks -= (len>>9);
2855 } while (++page_idx < RESYNC_PAGES);
2857 r1_bio->sectors = nr_sectors;
2859 if (mddev_is_clustered(mddev) &&
2860 conf->cluster_sync_high < sector_nr + nr_sectors) {
2861 conf->cluster_sync_low = mddev->curr_resync_completed;
2862 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2863 /* Send resync message */
2864 md_cluster_ops->resync_info_update(mddev,
2865 conf->cluster_sync_low,
2866 conf->cluster_sync_high);
2869 /* For a user-requested sync, we read all readable devices and do a
2872 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2873 atomic_set(&r1_bio->remaining, read_targets);
2874 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2875 bio = r1_bio->bios[i];
2876 if (bio->bi_end_io == end_sync_read) {
2878 md_sync_acct_bio(bio, nr_sectors);
2879 if (read_targets == 1)
2880 bio->bi_opf &= ~MD_FAILFAST;
2881 generic_make_request(bio);
2885 atomic_set(&r1_bio->remaining, 1);
2886 bio = r1_bio->bios[r1_bio->read_disk];
2887 md_sync_acct_bio(bio, nr_sectors);
2888 if (read_targets == 1)
2889 bio->bi_opf &= ~MD_FAILFAST;
2890 generic_make_request(bio);
2896 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2901 return mddev->dev_sectors;
2904 static struct r1conf *setup_conf(struct mddev *mddev)
2906 struct r1conf *conf;
2908 struct raid1_info *disk;
2909 struct md_rdev *rdev;
2912 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2916 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2917 sizeof(atomic_t), GFP_KERNEL);
2918 if (!conf->nr_pending)
2921 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2922 sizeof(atomic_t), GFP_KERNEL);
2923 if (!conf->nr_waiting)
2926 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2927 sizeof(atomic_t), GFP_KERNEL);
2928 if (!conf->nr_queued)
2931 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2932 sizeof(atomic_t), GFP_KERNEL);
2936 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2937 mddev->raid_disks, 2),
2942 conf->tmppage = alloc_page(GFP_KERNEL);
2946 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2947 if (!conf->poolinfo)
2949 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2950 err = mempool_init(&conf->r1bio_pool, NR_RAID1_BIOS, r1bio_pool_alloc,
2951 r1bio_pool_free, conf->poolinfo);
2955 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2959 conf->poolinfo->mddev = mddev;
2962 spin_lock_init(&conf->device_lock);
2963 rdev_for_each(rdev, mddev) {
2964 int disk_idx = rdev->raid_disk;
2965 if (disk_idx >= mddev->raid_disks
2968 if (test_bit(Replacement, &rdev->flags))
2969 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2971 disk = conf->mirrors + disk_idx;
2976 disk->head_position = 0;
2977 disk->seq_start = MaxSector;
2979 conf->raid_disks = mddev->raid_disks;
2980 conf->mddev = mddev;
2981 INIT_LIST_HEAD(&conf->retry_list);
2982 INIT_LIST_HEAD(&conf->bio_end_io_list);
2984 spin_lock_init(&conf->resync_lock);
2985 init_waitqueue_head(&conf->wait_barrier);
2987 bio_list_init(&conf->pending_bio_list);
2988 conf->pending_count = 0;
2989 conf->recovery_disabled = mddev->recovery_disabled - 1;
2992 for (i = 0; i < conf->raid_disks * 2; i++) {
2994 disk = conf->mirrors + i;
2996 if (i < conf->raid_disks &&
2997 disk[conf->raid_disks].rdev) {
2998 /* This slot has a replacement. */
3000 /* No original, just make the replacement
3001 * a recovering spare
3004 disk[conf->raid_disks].rdev;
3005 disk[conf->raid_disks].rdev = NULL;
3006 } else if (!test_bit(In_sync, &disk->rdev->flags))
3007 /* Original is not in_sync - bad */
3012 !test_bit(In_sync, &disk->rdev->flags)) {
3013 disk->head_position = 0;
3015 (disk->rdev->saved_raid_disk < 0))
3021 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3029 mempool_exit(&conf->r1bio_pool);
3030 kfree(conf->mirrors);
3031 safe_put_page(conf->tmppage);
3032 kfree(conf->poolinfo);
3033 kfree(conf->nr_pending);
3034 kfree(conf->nr_waiting);
3035 kfree(conf->nr_queued);
3036 kfree(conf->barrier);
3037 bioset_exit(&conf->bio_split);
3040 return ERR_PTR(err);
3043 static void raid1_free(struct mddev *mddev, void *priv);
3044 static int raid1_run(struct mddev *mddev)
3046 struct r1conf *conf;
3048 struct md_rdev *rdev;
3050 bool discard_supported = false;
3052 if (mddev->level != 1) {
3053 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3054 mdname(mddev), mddev->level);
3057 if (mddev->reshape_position != MaxSector) {
3058 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3062 if (mddev_init_writes_pending(mddev) < 0)
3065 * copy the already verified devices into our private RAID1
3066 * bookkeeping area. [whatever we allocate in run(),
3067 * should be freed in raid1_free()]
3069 if (mddev->private == NULL)
3070 conf = setup_conf(mddev);
3072 conf = mddev->private;
3075 return PTR_ERR(conf);
3078 blk_queue_max_write_same_sectors(mddev->queue, 0);
3079 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3082 rdev_for_each(rdev, mddev) {
3083 if (!mddev->gendisk)
3085 disk_stack_limits(mddev->gendisk, rdev->bdev,
3086 rdev->data_offset << 9);
3087 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3088 discard_supported = true;
3091 mddev->degraded = 0;
3092 for (i=0; i < conf->raid_disks; i++)
3093 if (conf->mirrors[i].rdev == NULL ||
3094 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3095 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3098 if (conf->raid_disks - mddev->degraded == 1)
3099 mddev->recovery_cp = MaxSector;
3101 if (mddev->recovery_cp != MaxSector)
3102 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3104 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3105 mdname(mddev), mddev->raid_disks - mddev->degraded,
3109 * Ok, everything is just fine now
3111 mddev->thread = conf->thread;
3112 conf->thread = NULL;
3113 mddev->private = conf;
3114 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3116 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3119 if (discard_supported)
3120 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3123 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3127 ret = md_integrity_register(mddev);
3129 md_unregister_thread(&mddev->thread);
3130 raid1_free(mddev, conf);
3135 static void raid1_free(struct mddev *mddev, void *priv)
3137 struct r1conf *conf = priv;
3139 mempool_exit(&conf->r1bio_pool);
3140 kfree(conf->mirrors);
3141 safe_put_page(conf->tmppage);
3142 kfree(conf->poolinfo);
3143 kfree(conf->nr_pending);
3144 kfree(conf->nr_waiting);
3145 kfree(conf->nr_queued);
3146 kfree(conf->barrier);
3147 bioset_exit(&conf->bio_split);
3151 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3153 /* no resync is happening, and there is enough space
3154 * on all devices, so we can resize.
3155 * We need to make sure resync covers any new space.
3156 * If the array is shrinking we should possibly wait until
3157 * any io in the removed space completes, but it hardly seems
3160 sector_t newsize = raid1_size(mddev, sectors, 0);
3161 if (mddev->external_size &&
3162 mddev->array_sectors > newsize)
3164 if (mddev->bitmap) {
3165 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3169 md_set_array_sectors(mddev, newsize);
3170 if (sectors > mddev->dev_sectors &&
3171 mddev->recovery_cp > mddev->dev_sectors) {
3172 mddev->recovery_cp = mddev->dev_sectors;
3173 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3175 mddev->dev_sectors = sectors;
3176 mddev->resync_max_sectors = sectors;
3180 static int raid1_reshape(struct mddev *mddev)
3183 * 1/ resize the r1bio_pool
3184 * 2/ resize conf->mirrors
3186 * We allocate a new r1bio_pool if we can.
3187 * Then raise a device barrier and wait until all IO stops.
3188 * Then resize conf->mirrors and swap in the new r1bio pool.
3190 * At the same time, we "pack" the devices so that all the missing
3191 * devices have the higher raid_disk numbers.
3193 mempool_t newpool, oldpool;
3194 struct pool_info *newpoolinfo;
3195 struct raid1_info *newmirrors;
3196 struct r1conf *conf = mddev->private;
3197 int cnt, raid_disks;
3198 unsigned long flags;
3202 memset(&newpool, 0, sizeof(newpool));
3203 memset(&oldpool, 0, sizeof(oldpool));
3205 /* Cannot change chunk_size, layout, or level */
3206 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3207 mddev->layout != mddev->new_layout ||
3208 mddev->level != mddev->new_level) {
3209 mddev->new_chunk_sectors = mddev->chunk_sectors;
3210 mddev->new_layout = mddev->layout;
3211 mddev->new_level = mddev->level;
3215 if (!mddev_is_clustered(mddev))
3216 md_allow_write(mddev);
3218 raid_disks = mddev->raid_disks + mddev->delta_disks;
3220 if (raid_disks < conf->raid_disks) {
3222 for (d= 0; d < conf->raid_disks; d++)
3223 if (conf->mirrors[d].rdev)
3225 if (cnt > raid_disks)
3229 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3232 newpoolinfo->mddev = mddev;
3233 newpoolinfo->raid_disks = raid_disks * 2;
3235 ret = mempool_init(&newpool, NR_RAID1_BIOS, r1bio_pool_alloc,
3236 r1bio_pool_free, newpoolinfo);
3241 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3246 mempool_exit(&newpool);
3250 freeze_array(conf, 0);
3252 /* ok, everything is stopped */
3253 oldpool = conf->r1bio_pool;
3254 conf->r1bio_pool = newpool;
3256 for (d = d2 = 0; d < conf->raid_disks; d++) {
3257 struct md_rdev *rdev = conf->mirrors[d].rdev;
3258 if (rdev && rdev->raid_disk != d2) {
3259 sysfs_unlink_rdev(mddev, rdev);
3260 rdev->raid_disk = d2;
3261 sysfs_unlink_rdev(mddev, rdev);
3262 if (sysfs_link_rdev(mddev, rdev))
3263 pr_warn("md/raid1:%s: cannot register rd%d\n",
3264 mdname(mddev), rdev->raid_disk);
3267 newmirrors[d2++].rdev = rdev;
3269 kfree(conf->mirrors);
3270 conf->mirrors = newmirrors;
3271 kfree(conf->poolinfo);
3272 conf->poolinfo = newpoolinfo;
3274 spin_lock_irqsave(&conf->device_lock, flags);
3275 mddev->degraded += (raid_disks - conf->raid_disks);
3276 spin_unlock_irqrestore(&conf->device_lock, flags);
3277 conf->raid_disks = mddev->raid_disks = raid_disks;
3278 mddev->delta_disks = 0;
3280 unfreeze_array(conf);
3282 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3283 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3284 md_wakeup_thread(mddev->thread);
3286 mempool_exit(&oldpool);
3290 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3292 struct r1conf *conf = mddev->private;
3295 freeze_array(conf, 0);
3297 unfreeze_array(conf);
3300 static void *raid1_takeover(struct mddev *mddev)
3302 /* raid1 can take over:
3303 * raid5 with 2 devices, any layout or chunk size
3305 if (mddev->level == 5 && mddev->raid_disks == 2) {
3306 struct r1conf *conf;
3307 mddev->new_level = 1;
3308 mddev->new_layout = 0;
3309 mddev->new_chunk_sectors = 0;
3310 conf = setup_conf(mddev);
3311 if (!IS_ERR(conf)) {
3312 /* Array must appear to be quiesced */
3313 conf->array_frozen = 1;
3314 mddev_clear_unsupported_flags(mddev,
3315 UNSUPPORTED_MDDEV_FLAGS);
3319 return ERR_PTR(-EINVAL);
3322 static struct md_personality raid1_personality =
3326 .owner = THIS_MODULE,
3327 .make_request = raid1_make_request,
3330 .status = raid1_status,
3331 .error_handler = raid1_error,
3332 .hot_add_disk = raid1_add_disk,
3333 .hot_remove_disk= raid1_remove_disk,
3334 .spare_active = raid1_spare_active,
3335 .sync_request = raid1_sync_request,
3336 .resize = raid1_resize,
3338 .check_reshape = raid1_reshape,
3339 .quiesce = raid1_quiesce,
3340 .takeover = raid1_takeover,
3341 .congested = raid1_congested,
3344 static int __init raid_init(void)
3346 return register_md_personality(&raid1_personality);
3349 static void raid_exit(void)
3351 unregister_md_personality(&raid1_personality);
3354 module_init(raid_init);
3355 module_exit(raid_exit);
3356 MODULE_LICENSE("GPL");
3357 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3358 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3359 MODULE_ALIAS("md-raid1");
3360 MODULE_ALIAS("md-level-1");
3362 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);