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))
45 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
46 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
48 #define raid1_log(md, fmt, args...) \
49 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
54 * for resync bio, r1bio pointer can be retrieved from the per-bio
55 * 'struct resync_pages'.
57 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
59 return get_resync_pages(bio)->raid_bio;
62 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
64 struct pool_info *pi = data;
65 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
67 /* allocate a r1bio with room for raid_disks entries in the bios array */
68 return kzalloc(size, gfp_flags);
71 static void r1bio_pool_free(void *r1_bio, void *data)
76 #define RESYNC_DEPTH 32
77 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
78 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
79 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
80 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
81 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
83 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
85 struct pool_info *pi = data;
90 struct resync_pages *rps;
92 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
96 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
102 * Allocate bios : 1 for reading, n-1 for writing
104 for (j = pi->raid_disks ; j-- ; ) {
105 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
108 r1_bio->bios[j] = bio;
111 * Allocate RESYNC_PAGES data pages and attach them to
113 * If this is a user-requested check/repair, allocate
114 * RESYNC_PAGES for each bio.
116 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
117 need_pages = pi->raid_disks;
120 for (j = 0; j < pi->raid_disks; j++) {
121 struct resync_pages *rp = &rps[j];
123 bio = r1_bio->bios[j];
125 if (j < need_pages) {
126 if (resync_alloc_pages(rp, gfp_flags))
129 memcpy(rp, &rps[0], sizeof(*rp));
130 resync_get_all_pages(rp);
133 rp->raid_bio = r1_bio;
134 bio->bi_private = rp;
137 r1_bio->master_bio = NULL;
143 resync_free_pages(&rps[j]);
146 while (++j < pi->raid_disks)
147 bio_put(r1_bio->bios[j]);
151 r1bio_pool_free(r1_bio, data);
155 static void r1buf_pool_free(void *__r1_bio, void *data)
157 struct pool_info *pi = data;
159 struct r1bio *r1bio = __r1_bio;
160 struct resync_pages *rp = NULL;
162 for (i = pi->raid_disks; i--; ) {
163 rp = get_resync_pages(r1bio->bios[i]);
164 resync_free_pages(rp);
165 bio_put(r1bio->bios[i]);
168 /* resync pages array stored in the 1st bio's .bi_private */
171 r1bio_pool_free(r1bio, data);
174 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
178 for (i = 0; i < conf->raid_disks * 2; i++) {
179 struct bio **bio = r1_bio->bios + i;
180 if (!BIO_SPECIAL(*bio))
186 static void free_r1bio(struct r1bio *r1_bio)
188 struct r1conf *conf = r1_bio->mddev->private;
190 put_all_bios(conf, r1_bio);
191 mempool_free(r1_bio, &conf->r1bio_pool);
194 static void put_buf(struct r1bio *r1_bio)
196 struct r1conf *conf = r1_bio->mddev->private;
197 sector_t sect = r1_bio->sector;
200 for (i = 0; i < conf->raid_disks * 2; i++) {
201 struct bio *bio = r1_bio->bios[i];
203 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
206 mempool_free(r1_bio, &conf->r1buf_pool);
208 lower_barrier(conf, sect);
211 static void reschedule_retry(struct r1bio *r1_bio)
214 struct mddev *mddev = r1_bio->mddev;
215 struct r1conf *conf = mddev->private;
218 idx = sector_to_idx(r1_bio->sector);
219 spin_lock_irqsave(&conf->device_lock, flags);
220 list_add(&r1_bio->retry_list, &conf->retry_list);
221 atomic_inc(&conf->nr_queued[idx]);
222 spin_unlock_irqrestore(&conf->device_lock, flags);
224 wake_up(&conf->wait_barrier);
225 md_wakeup_thread(mddev->thread);
229 * raid_end_bio_io() is called when we have finished servicing a mirrored
230 * operation and are ready to return a success/failure code to the buffer
233 static void call_bio_endio(struct r1bio *r1_bio)
235 struct bio *bio = r1_bio->master_bio;
236 struct r1conf *conf = r1_bio->mddev->private;
238 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
239 bio->bi_status = BLK_STS_IOERR;
243 * Wake up any possible resync thread that waits for the device
246 allow_barrier(conf, r1_bio->sector);
249 static void raid_end_bio_io(struct r1bio *r1_bio)
251 struct bio *bio = r1_bio->master_bio;
253 /* if nobody has done the final endio yet, do it now */
254 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
255 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
256 (bio_data_dir(bio) == WRITE) ? "write" : "read",
257 (unsigned long long) bio->bi_iter.bi_sector,
258 (unsigned long long) bio_end_sector(bio) - 1);
260 call_bio_endio(r1_bio);
266 * Update disk head position estimator based on IRQ completion info.
268 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
270 struct r1conf *conf = r1_bio->mddev->private;
272 conf->mirrors[disk].head_position =
273 r1_bio->sector + (r1_bio->sectors);
277 * Find the disk number which triggered given bio
279 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
282 struct r1conf *conf = r1_bio->mddev->private;
283 int raid_disks = conf->raid_disks;
285 for (mirror = 0; mirror < raid_disks * 2; mirror++)
286 if (r1_bio->bios[mirror] == bio)
289 BUG_ON(mirror == raid_disks * 2);
290 update_head_pos(mirror, r1_bio);
295 static void raid1_end_read_request(struct bio *bio)
297 int uptodate = !bio->bi_status;
298 struct r1bio *r1_bio = bio->bi_private;
299 struct r1conf *conf = r1_bio->mddev->private;
300 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
303 * this branch is our 'one mirror IO has finished' event handler:
305 update_head_pos(r1_bio->read_disk, r1_bio);
308 set_bit(R1BIO_Uptodate, &r1_bio->state);
309 else if (test_bit(FailFast, &rdev->flags) &&
310 test_bit(R1BIO_FailFast, &r1_bio->state))
311 /* This was a fail-fast read so we definitely
315 /* If all other devices have failed, we want to return
316 * the error upwards rather than fail the last device.
317 * Here we redefine "uptodate" to mean "Don't want to retry"
320 spin_lock_irqsave(&conf->device_lock, flags);
321 if (r1_bio->mddev->degraded == conf->raid_disks ||
322 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
323 test_bit(In_sync, &rdev->flags)))
325 spin_unlock_irqrestore(&conf->device_lock, flags);
329 raid_end_bio_io(r1_bio);
330 rdev_dec_pending(rdev, conf->mddev);
335 char b[BDEVNAME_SIZE];
336 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
338 bdevname(rdev->bdev, b),
339 (unsigned long long)r1_bio->sector);
340 set_bit(R1BIO_ReadError, &r1_bio->state);
341 reschedule_retry(r1_bio);
342 /* don't drop the reference on read_disk yet */
346 static void close_write(struct r1bio *r1_bio)
348 /* it really is the end of this request */
349 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
350 bio_free_pages(r1_bio->behind_master_bio);
351 bio_put(r1_bio->behind_master_bio);
352 r1_bio->behind_master_bio = NULL;
354 /* clear the bitmap if all writes complete successfully */
355 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
357 !test_bit(R1BIO_Degraded, &r1_bio->state),
358 test_bit(R1BIO_BehindIO, &r1_bio->state));
359 md_write_end(r1_bio->mddev);
362 static void r1_bio_write_done(struct r1bio *r1_bio)
364 if (!atomic_dec_and_test(&r1_bio->remaining))
367 if (test_bit(R1BIO_WriteError, &r1_bio->state))
368 reschedule_retry(r1_bio);
371 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
372 reschedule_retry(r1_bio);
374 raid_end_bio_io(r1_bio);
378 static void raid1_end_write_request(struct bio *bio)
380 struct r1bio *r1_bio = bio->bi_private;
381 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
382 struct r1conf *conf = r1_bio->mddev->private;
383 struct bio *to_put = NULL;
384 int mirror = find_bio_disk(r1_bio, bio);
385 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
388 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
391 * 'one mirror IO has finished' event handler:
393 if (bio->bi_status && !discard_error) {
394 set_bit(WriteErrorSeen, &rdev->flags);
395 if (!test_and_set_bit(WantReplacement, &rdev->flags))
396 set_bit(MD_RECOVERY_NEEDED, &
397 conf->mddev->recovery);
399 if (test_bit(FailFast, &rdev->flags) &&
400 (bio->bi_opf & MD_FAILFAST) &&
401 /* We never try FailFast to WriteMostly devices */
402 !test_bit(WriteMostly, &rdev->flags)) {
403 md_error(r1_bio->mddev, rdev);
404 if (!test_bit(Faulty, &rdev->flags))
405 /* This is the only remaining device,
406 * We need to retry the write without
409 set_bit(R1BIO_WriteError, &r1_bio->state);
411 /* Finished with this branch */
412 r1_bio->bios[mirror] = NULL;
416 set_bit(R1BIO_WriteError, &r1_bio->state);
419 * Set R1BIO_Uptodate in our master bio, so that we
420 * will return a good error code for to the higher
421 * levels even if IO on some other mirrored buffer
424 * The 'master' represents the composite IO operation
425 * to user-side. So if something waits for IO, then it
426 * will wait for the 'master' bio.
431 r1_bio->bios[mirror] = NULL;
434 * Do not set R1BIO_Uptodate if the current device is
435 * rebuilding or Faulty. This is because we cannot use
436 * such device for properly reading the data back (we could
437 * potentially use it, if the current write would have felt
438 * before rdev->recovery_offset, but for simplicity we don't
441 if (test_bit(In_sync, &rdev->flags) &&
442 !test_bit(Faulty, &rdev->flags))
443 set_bit(R1BIO_Uptodate, &r1_bio->state);
445 /* Maybe we can clear some bad blocks. */
446 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
447 &first_bad, &bad_sectors) && !discard_error) {
448 r1_bio->bios[mirror] = IO_MADE_GOOD;
449 set_bit(R1BIO_MadeGood, &r1_bio->state);
454 if (test_bit(WriteMostly, &rdev->flags))
455 atomic_dec(&r1_bio->behind_remaining);
458 * In behind mode, we ACK the master bio once the I/O
459 * has safely reached all non-writemostly
460 * disks. Setting the Returned bit ensures that this
461 * gets done only once -- we don't ever want to return
462 * -EIO here, instead we'll wait
464 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
465 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
466 /* Maybe we can return now */
467 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
468 struct bio *mbio = r1_bio->master_bio;
469 pr_debug("raid1: behind end write sectors"
471 (unsigned long long) mbio->bi_iter.bi_sector,
472 (unsigned long long) bio_end_sector(mbio) - 1);
473 call_bio_endio(r1_bio);
477 if (r1_bio->bios[mirror] == NULL)
478 rdev_dec_pending(rdev, conf->mddev);
481 * Let's see if all mirrored write operations have finished
484 r1_bio_write_done(r1_bio);
490 static sector_t align_to_barrier_unit_end(sector_t start_sector,
495 WARN_ON(sectors == 0);
497 * len is the number of sectors from start_sector to end of the
498 * barrier unit which start_sector belongs to.
500 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
510 * This routine returns the disk from which the requested read should
511 * be done. There is a per-array 'next expected sequential IO' sector
512 * number - if this matches on the next IO then we use the last disk.
513 * There is also a per-disk 'last know head position' sector that is
514 * maintained from IRQ contexts, both the normal and the resync IO
515 * completion handlers update this position correctly. If there is no
516 * perfect sequential match then we pick the disk whose head is closest.
518 * If there are 2 mirrors in the same 2 devices, performance degrades
519 * because position is mirror, not device based.
521 * The rdev for the device selected will have nr_pending incremented.
523 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
525 const sector_t this_sector = r1_bio->sector;
527 int best_good_sectors;
528 int best_disk, best_dist_disk, best_pending_disk;
532 unsigned int min_pending;
533 struct md_rdev *rdev;
535 int choose_next_idle;
539 * Check if we can balance. We can balance on the whole
540 * device if no resync is going on, or below the resync window.
541 * We take the first readable disk when above the resync window.
544 sectors = r1_bio->sectors;
547 best_dist = MaxSector;
548 best_pending_disk = -1;
549 min_pending = UINT_MAX;
550 best_good_sectors = 0;
552 choose_next_idle = 0;
553 clear_bit(R1BIO_FailFast, &r1_bio->state);
555 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
556 (mddev_is_clustered(conf->mddev) &&
557 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
558 this_sector + sectors)))
563 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
567 unsigned int pending;
570 rdev = rcu_dereference(conf->mirrors[disk].rdev);
571 if (r1_bio->bios[disk] == IO_BLOCKED
573 || test_bit(Faulty, &rdev->flags))
575 if (!test_bit(In_sync, &rdev->flags) &&
576 rdev->recovery_offset < this_sector + sectors)
578 if (test_bit(WriteMostly, &rdev->flags)) {
579 /* Don't balance among write-mostly, just
580 * use the first as a last resort */
581 if (best_dist_disk < 0) {
582 if (is_badblock(rdev, this_sector, sectors,
583 &first_bad, &bad_sectors)) {
584 if (first_bad <= this_sector)
585 /* Cannot use this */
587 best_good_sectors = first_bad - this_sector;
589 best_good_sectors = sectors;
590 best_dist_disk = disk;
591 best_pending_disk = disk;
595 /* This is a reasonable device to use. It might
598 if (is_badblock(rdev, this_sector, sectors,
599 &first_bad, &bad_sectors)) {
600 if (best_dist < MaxSector)
601 /* already have a better device */
603 if (first_bad <= this_sector) {
604 /* cannot read here. If this is the 'primary'
605 * device, then we must not read beyond
606 * bad_sectors from another device..
608 bad_sectors -= (this_sector - first_bad);
609 if (choose_first && sectors > bad_sectors)
610 sectors = bad_sectors;
611 if (best_good_sectors > sectors)
612 best_good_sectors = sectors;
615 sector_t good_sectors = first_bad - this_sector;
616 if (good_sectors > best_good_sectors) {
617 best_good_sectors = good_sectors;
625 if ((sectors > best_good_sectors) && (best_disk >= 0))
627 best_good_sectors = sectors;
631 /* At least two disks to choose from so failfast is OK */
632 set_bit(R1BIO_FailFast, &r1_bio->state);
634 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
635 has_nonrot_disk |= nonrot;
636 pending = atomic_read(&rdev->nr_pending);
637 dist = abs(this_sector - conf->mirrors[disk].head_position);
642 /* Don't change to another disk for sequential reads */
643 if (conf->mirrors[disk].next_seq_sect == this_sector
645 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
646 struct raid1_info *mirror = &conf->mirrors[disk];
650 * If buffered sequential IO size exceeds optimal
651 * iosize, check if there is idle disk. If yes, choose
652 * the idle disk. read_balance could already choose an
653 * idle disk before noticing it's a sequential IO in
654 * this disk. This doesn't matter because this disk
655 * will idle, next time it will be utilized after the
656 * first disk has IO size exceeds optimal iosize. In
657 * this way, iosize of the first disk will be optimal
658 * iosize at least. iosize of the second disk might be
659 * small, but not a big deal since when the second disk
660 * starts IO, the first disk is likely still busy.
662 if (nonrot && opt_iosize > 0 &&
663 mirror->seq_start != MaxSector &&
664 mirror->next_seq_sect > opt_iosize &&
665 mirror->next_seq_sect - opt_iosize >=
667 choose_next_idle = 1;
673 if (choose_next_idle)
676 if (min_pending > pending) {
677 min_pending = pending;
678 best_pending_disk = disk;
681 if (dist < best_dist) {
683 best_dist_disk = disk;
688 * If all disks are rotational, choose the closest disk. If any disk is
689 * non-rotational, choose the disk with less pending request even the
690 * disk is rotational, which might/might not be optimal for raids with
691 * mixed ratation/non-rotational disks depending on workload.
693 if (best_disk == -1) {
694 if (has_nonrot_disk || min_pending == 0)
695 best_disk = best_pending_disk;
697 best_disk = best_dist_disk;
700 if (best_disk >= 0) {
701 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
704 atomic_inc(&rdev->nr_pending);
705 sectors = best_good_sectors;
707 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
708 conf->mirrors[best_disk].seq_start = this_sector;
710 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
713 *max_sectors = sectors;
718 static int raid1_congested(struct mddev *mddev, int bits)
720 struct r1conf *conf = mddev->private;
723 if ((bits & (1 << WB_async_congested)) &&
724 conf->pending_count >= max_queued_requests)
728 for (i = 0; i < conf->raid_disks * 2; i++) {
729 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
730 if (rdev && !test_bit(Faulty, &rdev->flags)) {
731 struct request_queue *q = bdev_get_queue(rdev->bdev);
735 /* Note the '|| 1' - when read_balance prefers
736 * non-congested targets, it can be removed
738 if ((bits & (1 << WB_async_congested)) || 1)
739 ret |= bdi_congested(q->backing_dev_info, bits);
741 ret &= bdi_congested(q->backing_dev_info, bits);
748 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
750 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
751 md_bitmap_unplug(conf->mddev->bitmap);
752 wake_up(&conf->wait_barrier);
754 while (bio) { /* submit pending writes */
755 struct bio *next = bio->bi_next;
756 struct md_rdev *rdev = (void *)bio->bi_disk;
758 bio_set_dev(bio, rdev->bdev);
759 if (test_bit(Faulty, &rdev->flags)) {
761 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
762 !blk_queue_discard(bio->bi_disk->queue)))
766 generic_make_request(bio);
771 static void flush_pending_writes(struct r1conf *conf)
773 /* Any writes that have been queued but are awaiting
774 * bitmap updates get flushed here.
776 spin_lock_irq(&conf->device_lock);
778 if (conf->pending_bio_list.head) {
779 struct blk_plug plug;
782 bio = bio_list_get(&conf->pending_bio_list);
783 conf->pending_count = 0;
784 spin_unlock_irq(&conf->device_lock);
787 * As this is called in a wait_event() loop (see freeze_array),
788 * current->state might be TASK_UNINTERRUPTIBLE which will
789 * cause a warning when we prepare to wait again. As it is
790 * rare that this path is taken, it is perfectly safe to force
791 * us to go around the wait_event() loop again, so the warning
792 * is a false-positive. Silence the warning by resetting
795 __set_current_state(TASK_RUNNING);
796 blk_start_plug(&plug);
797 flush_bio_list(conf, bio);
798 blk_finish_plug(&plug);
800 spin_unlock_irq(&conf->device_lock);
804 * Sometimes we need to suspend IO while we do something else,
805 * either some resync/recovery, or reconfigure the array.
806 * To do this we raise a 'barrier'.
807 * The 'barrier' is a counter that can be raised multiple times
808 * to count how many activities are happening which preclude
810 * We can only raise the barrier if there is no pending IO.
811 * i.e. if nr_pending == 0.
812 * We choose only to raise the barrier if no-one is waiting for the
813 * barrier to go down. This means that as soon as an IO request
814 * is ready, no other operations which require a barrier will start
815 * until the IO request has had a chance.
817 * So: regular IO calls 'wait_barrier'. When that returns there
818 * is no backgroup IO happening, It must arrange to call
819 * allow_barrier when it has finished its IO.
820 * backgroup IO calls must call raise_barrier. Once that returns
821 * there is no normal IO happeing. It must arrange to call
822 * lower_barrier when the particular background IO completes.
824 static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
826 int idx = sector_to_idx(sector_nr);
828 spin_lock_irq(&conf->resync_lock);
830 /* Wait until no block IO is waiting */
831 wait_event_lock_irq(conf->wait_barrier,
832 !atomic_read(&conf->nr_waiting[idx]),
835 /* block any new IO from starting */
836 atomic_inc(&conf->barrier[idx]);
838 * In raise_barrier() we firstly increase conf->barrier[idx] then
839 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
840 * increase conf->nr_pending[idx] then check conf->barrier[idx].
841 * A memory barrier here to make sure conf->nr_pending[idx] won't
842 * be fetched before conf->barrier[idx] is increased. Otherwise
843 * there will be a race between raise_barrier() and _wait_barrier().
845 smp_mb__after_atomic();
847 /* For these conditions we must wait:
848 * A: while the array is in frozen state
849 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
850 * existing in corresponding I/O barrier bucket.
851 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
852 * max resync count which allowed on current I/O barrier bucket.
854 wait_event_lock_irq(conf->wait_barrier,
855 (!conf->array_frozen &&
856 !atomic_read(&conf->nr_pending[idx]) &&
857 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
858 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
861 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
862 atomic_dec(&conf->barrier[idx]);
863 spin_unlock_irq(&conf->resync_lock);
864 wake_up(&conf->wait_barrier);
868 atomic_inc(&conf->nr_sync_pending);
869 spin_unlock_irq(&conf->resync_lock);
874 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
876 int idx = sector_to_idx(sector_nr);
878 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
880 atomic_dec(&conf->barrier[idx]);
881 atomic_dec(&conf->nr_sync_pending);
882 wake_up(&conf->wait_barrier);
885 static void _wait_barrier(struct r1conf *conf, int idx)
888 * We need to increase conf->nr_pending[idx] very early here,
889 * then raise_barrier() can be blocked when it waits for
890 * conf->nr_pending[idx] to be 0. Then we can avoid holding
891 * conf->resync_lock when there is no barrier raised in same
892 * barrier unit bucket. Also if the array is frozen, I/O
893 * should be blocked until array is unfrozen.
895 atomic_inc(&conf->nr_pending[idx]);
897 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
898 * check conf->barrier[idx]. In raise_barrier() we firstly increase
899 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
900 * barrier is necessary here to make sure conf->barrier[idx] won't be
901 * fetched before conf->nr_pending[idx] is increased. Otherwise there
902 * will be a race between _wait_barrier() and raise_barrier().
904 smp_mb__after_atomic();
907 * Don't worry about checking two atomic_t variables at same time
908 * here. If during we check conf->barrier[idx], the array is
909 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
910 * 0, it is safe to return and make the I/O continue. Because the
911 * array is frozen, all I/O returned here will eventually complete
912 * or be queued, no race will happen. See code comment in
915 if (!READ_ONCE(conf->array_frozen) &&
916 !atomic_read(&conf->barrier[idx]))
920 * After holding conf->resync_lock, conf->nr_pending[idx]
921 * should be decreased before waiting for barrier to drop.
922 * Otherwise, we may encounter a race condition because
923 * raise_barrer() might be waiting for conf->nr_pending[idx]
924 * to be 0 at same time.
926 spin_lock_irq(&conf->resync_lock);
927 atomic_inc(&conf->nr_waiting[idx]);
928 atomic_dec(&conf->nr_pending[idx]);
930 * In case freeze_array() is waiting for
931 * get_unqueued_pending() == extra
933 wake_up(&conf->wait_barrier);
934 /* Wait for the barrier in same barrier unit bucket to drop. */
935 wait_event_lock_irq(conf->wait_barrier,
936 !conf->array_frozen &&
937 !atomic_read(&conf->barrier[idx]),
939 atomic_inc(&conf->nr_pending[idx]);
940 atomic_dec(&conf->nr_waiting[idx]);
941 spin_unlock_irq(&conf->resync_lock);
944 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
946 int idx = sector_to_idx(sector_nr);
949 * Very similar to _wait_barrier(). The difference is, for read
950 * I/O we don't need wait for sync I/O, but if the whole array
951 * is frozen, the read I/O still has to wait until the array is
952 * unfrozen. Since there is no ordering requirement with
953 * conf->barrier[idx] here, memory barrier is unnecessary as well.
955 atomic_inc(&conf->nr_pending[idx]);
957 if (!READ_ONCE(conf->array_frozen))
960 spin_lock_irq(&conf->resync_lock);
961 atomic_inc(&conf->nr_waiting[idx]);
962 atomic_dec(&conf->nr_pending[idx]);
964 * In case freeze_array() is waiting for
965 * get_unqueued_pending() == extra
967 wake_up(&conf->wait_barrier);
968 /* Wait for array to be unfrozen */
969 wait_event_lock_irq(conf->wait_barrier,
972 atomic_inc(&conf->nr_pending[idx]);
973 atomic_dec(&conf->nr_waiting[idx]);
974 spin_unlock_irq(&conf->resync_lock);
977 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
979 int idx = sector_to_idx(sector_nr);
981 _wait_barrier(conf, idx);
984 static void _allow_barrier(struct r1conf *conf, int idx)
986 atomic_dec(&conf->nr_pending[idx]);
987 wake_up(&conf->wait_barrier);
990 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
992 int idx = sector_to_idx(sector_nr);
994 _allow_barrier(conf, idx);
997 /* conf->resync_lock should be held */
998 static int get_unqueued_pending(struct r1conf *conf)
1002 ret = atomic_read(&conf->nr_sync_pending);
1003 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1004 ret += atomic_read(&conf->nr_pending[idx]) -
1005 atomic_read(&conf->nr_queued[idx]);
1010 static void freeze_array(struct r1conf *conf, int extra)
1012 /* Stop sync I/O and normal I/O and wait for everything to
1014 * This is called in two situations:
1015 * 1) management command handlers (reshape, remove disk, quiesce).
1016 * 2) one normal I/O request failed.
1018 * After array_frozen is set to 1, new sync IO will be blocked at
1019 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1020 * or wait_read_barrier(). The flying I/Os will either complete or be
1021 * queued. When everything goes quite, there are only queued I/Os left.
1023 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1024 * barrier bucket index which this I/O request hits. When all sync and
1025 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1026 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1027 * in handle_read_error(), we may call freeze_array() before trying to
1028 * fix the read error. In this case, the error read I/O is not queued,
1029 * so get_unqueued_pending() == 1.
1031 * Therefore before this function returns, we need to wait until
1032 * get_unqueued_pendings(conf) gets equal to extra. For
1033 * normal I/O context, extra is 1, in rested situations extra is 0.
1035 spin_lock_irq(&conf->resync_lock);
1036 conf->array_frozen = 1;
1037 raid1_log(conf->mddev, "wait freeze");
1038 wait_event_lock_irq_cmd(
1040 get_unqueued_pending(conf) == extra,
1042 flush_pending_writes(conf));
1043 spin_unlock_irq(&conf->resync_lock);
1045 static void unfreeze_array(struct r1conf *conf)
1047 /* reverse the effect of the freeze */
1048 spin_lock_irq(&conf->resync_lock);
1049 conf->array_frozen = 0;
1050 spin_unlock_irq(&conf->resync_lock);
1051 wake_up(&conf->wait_barrier);
1054 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1057 int size = bio->bi_iter.bi_size;
1058 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1060 struct bio *behind_bio = NULL;
1062 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1066 /* discard op, we don't support writezero/writesame yet */
1067 if (!bio_has_data(bio)) {
1068 behind_bio->bi_iter.bi_size = size;
1072 behind_bio->bi_write_hint = bio->bi_write_hint;
1074 while (i < vcnt && size) {
1076 int len = min_t(int, PAGE_SIZE, size);
1078 page = alloc_page(GFP_NOIO);
1079 if (unlikely(!page))
1082 bio_add_page(behind_bio, page, len, 0);
1088 bio_copy_data(behind_bio, bio);
1090 r1_bio->behind_master_bio = behind_bio;
1091 set_bit(R1BIO_BehindIO, &r1_bio->state);
1096 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1097 bio->bi_iter.bi_size);
1098 bio_free_pages(behind_bio);
1099 bio_put(behind_bio);
1102 struct raid1_plug_cb {
1103 struct blk_plug_cb cb;
1104 struct bio_list pending;
1108 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1110 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1112 struct mddev *mddev = plug->cb.data;
1113 struct r1conf *conf = mddev->private;
1116 if (from_schedule || current->bio_list) {
1117 spin_lock_irq(&conf->device_lock);
1118 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1119 conf->pending_count += plug->pending_cnt;
1120 spin_unlock_irq(&conf->device_lock);
1121 wake_up(&conf->wait_barrier);
1122 md_wakeup_thread(mddev->thread);
1127 /* we aren't scheduling, so we can do the write-out directly. */
1128 bio = bio_list_get(&plug->pending);
1129 flush_bio_list(conf, bio);
1133 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1135 r1_bio->master_bio = bio;
1136 r1_bio->sectors = bio_sectors(bio);
1138 r1_bio->mddev = mddev;
1139 r1_bio->sector = bio->bi_iter.bi_sector;
1142 static inline struct r1bio *
1143 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1145 struct r1conf *conf = mddev->private;
1146 struct r1bio *r1_bio;
1148 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1149 /* Ensure no bio records IO_BLOCKED */
1150 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1151 init_r1bio(r1_bio, mddev, bio);
1155 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1156 int max_read_sectors, struct r1bio *r1_bio)
1158 struct r1conf *conf = mddev->private;
1159 struct raid1_info *mirror;
1160 struct bio *read_bio;
1161 struct bitmap *bitmap = mddev->bitmap;
1162 const int op = bio_op(bio);
1163 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1166 bool print_msg = !!r1_bio;
1167 char b[BDEVNAME_SIZE];
1170 * If r1_bio is set, we are blocking the raid1d thread
1171 * so there is a tiny risk of deadlock. So ask for
1172 * emergency memory if needed.
1174 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1177 /* Need to get the block device name carefully */
1178 struct md_rdev *rdev;
1180 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1182 bdevname(rdev->bdev, b);
1189 * Still need barrier for READ in case that whole
1192 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1195 r1_bio = alloc_r1bio(mddev, bio);
1197 init_r1bio(r1_bio, mddev, bio);
1198 r1_bio->sectors = max_read_sectors;
1201 * make_request() can abort the operation when read-ahead is being
1202 * used and no empty request is available.
1204 rdisk = read_balance(conf, r1_bio, &max_sectors);
1207 /* couldn't find anywhere to read from */
1209 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1212 (unsigned long long)r1_bio->sector);
1214 raid_end_bio_io(r1_bio);
1217 mirror = conf->mirrors + rdisk;
1220 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1222 (unsigned long long)r1_bio->sector,
1223 bdevname(mirror->rdev->bdev, b));
1225 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1228 * Reading from a write-mostly device must take care not to
1229 * over-take any writes that are 'behind'
1231 raid1_log(mddev, "wait behind writes");
1232 wait_event(bitmap->behind_wait,
1233 atomic_read(&bitmap->behind_writes) == 0);
1236 if (max_sectors < bio_sectors(bio)) {
1237 struct bio *split = bio_split(bio, max_sectors,
1238 gfp, &conf->bio_split);
1239 bio_chain(split, bio);
1240 generic_make_request(bio);
1242 r1_bio->master_bio = bio;
1243 r1_bio->sectors = max_sectors;
1246 r1_bio->read_disk = rdisk;
1248 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1250 r1_bio->bios[rdisk] = read_bio;
1252 read_bio->bi_iter.bi_sector = r1_bio->sector +
1253 mirror->rdev->data_offset;
1254 bio_set_dev(read_bio, mirror->rdev->bdev);
1255 read_bio->bi_end_io = raid1_end_read_request;
1256 bio_set_op_attrs(read_bio, op, do_sync);
1257 if (test_bit(FailFast, &mirror->rdev->flags) &&
1258 test_bit(R1BIO_FailFast, &r1_bio->state))
1259 read_bio->bi_opf |= MD_FAILFAST;
1260 read_bio->bi_private = r1_bio;
1263 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1264 disk_devt(mddev->gendisk), r1_bio->sector);
1266 generic_make_request(read_bio);
1269 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1270 int max_write_sectors)
1272 struct r1conf *conf = mddev->private;
1273 struct r1bio *r1_bio;
1275 struct bitmap *bitmap = mddev->bitmap;
1276 unsigned long flags;
1277 struct md_rdev *blocked_rdev;
1278 struct blk_plug_cb *cb;
1279 struct raid1_plug_cb *plug = NULL;
1283 if (mddev_is_clustered(mddev) &&
1284 md_cluster_ops->area_resyncing(mddev, WRITE,
1285 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1289 prepare_to_wait(&conf->wait_barrier,
1291 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1292 bio->bi_iter.bi_sector,
1293 bio_end_sector(bio)))
1297 finish_wait(&conf->wait_barrier, &w);
1301 * Register the new request and wait if the reconstruction
1302 * thread has put up a bar for new requests.
1303 * Continue immediately if no resync is active currently.
1305 wait_barrier(conf, bio->bi_iter.bi_sector);
1307 r1_bio = alloc_r1bio(mddev, bio);
1308 r1_bio->sectors = max_write_sectors;
1310 if (conf->pending_count >= max_queued_requests) {
1311 md_wakeup_thread(mddev->thread);
1312 raid1_log(mddev, "wait queued");
1313 wait_event(conf->wait_barrier,
1314 conf->pending_count < max_queued_requests);
1316 /* first select target devices under rcu_lock and
1317 * inc refcount on their rdev. Record them by setting
1319 * If there are known/acknowledged bad blocks on any device on
1320 * which we have seen a write error, we want to avoid writing those
1322 * This potentially requires several writes to write around
1323 * the bad blocks. Each set of writes gets it's own r1bio
1324 * with a set of bios attached.
1327 disks = conf->raid_disks * 2;
1329 blocked_rdev = NULL;
1331 max_sectors = r1_bio->sectors;
1332 for (i = 0; i < disks; i++) {
1333 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1334 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1335 atomic_inc(&rdev->nr_pending);
1336 blocked_rdev = rdev;
1339 r1_bio->bios[i] = NULL;
1340 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1341 if (i < conf->raid_disks)
1342 set_bit(R1BIO_Degraded, &r1_bio->state);
1346 atomic_inc(&rdev->nr_pending);
1347 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1352 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1353 &first_bad, &bad_sectors);
1355 /* mustn't write here until the bad block is
1357 set_bit(BlockedBadBlocks, &rdev->flags);
1358 blocked_rdev = rdev;
1361 if (is_bad && first_bad <= r1_bio->sector) {
1362 /* Cannot write here at all */
1363 bad_sectors -= (r1_bio->sector - first_bad);
1364 if (bad_sectors < max_sectors)
1365 /* mustn't write more than bad_sectors
1366 * to other devices yet
1368 max_sectors = bad_sectors;
1369 rdev_dec_pending(rdev, mddev);
1370 /* We don't set R1BIO_Degraded as that
1371 * only applies if the disk is
1372 * missing, so it might be re-added,
1373 * and we want to know to recover this
1375 * In this case the device is here,
1376 * and the fact that this chunk is not
1377 * in-sync is recorded in the bad
1383 int good_sectors = first_bad - r1_bio->sector;
1384 if (good_sectors < max_sectors)
1385 max_sectors = good_sectors;
1388 r1_bio->bios[i] = bio;
1392 if (unlikely(blocked_rdev)) {
1393 /* Wait for this device to become unblocked */
1396 for (j = 0; j < i; j++)
1397 if (r1_bio->bios[j])
1398 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1400 allow_barrier(conf, bio->bi_iter.bi_sector);
1401 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1402 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1403 wait_barrier(conf, bio->bi_iter.bi_sector);
1407 if (max_sectors < bio_sectors(bio)) {
1408 struct bio *split = bio_split(bio, max_sectors,
1409 GFP_NOIO, &conf->bio_split);
1410 bio_chain(split, bio);
1411 generic_make_request(bio);
1413 r1_bio->master_bio = bio;
1414 r1_bio->sectors = max_sectors;
1417 atomic_set(&r1_bio->remaining, 1);
1418 atomic_set(&r1_bio->behind_remaining, 0);
1422 for (i = 0; i < disks; i++) {
1423 struct bio *mbio = NULL;
1424 if (!r1_bio->bios[i])
1429 * Not if there are too many, or cannot
1430 * allocate memory, or a reader on WriteMostly
1431 * is waiting for behind writes to flush */
1433 (atomic_read(&bitmap->behind_writes)
1434 < mddev->bitmap_info.max_write_behind) &&
1435 !waitqueue_active(&bitmap->behind_wait)) {
1436 alloc_behind_master_bio(r1_bio, bio);
1439 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1440 test_bit(R1BIO_BehindIO, &r1_bio->state));
1444 if (r1_bio->behind_master_bio)
1445 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1446 GFP_NOIO, &mddev->bio_set);
1448 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1450 if (r1_bio->behind_master_bio) {
1451 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1452 atomic_inc(&r1_bio->behind_remaining);
1455 r1_bio->bios[i] = mbio;
1457 mbio->bi_iter.bi_sector = (r1_bio->sector +
1458 conf->mirrors[i].rdev->data_offset);
1459 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1460 mbio->bi_end_io = raid1_end_write_request;
1461 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1462 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1463 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1464 conf->raid_disks - mddev->degraded > 1)
1465 mbio->bi_opf |= MD_FAILFAST;
1466 mbio->bi_private = r1_bio;
1468 atomic_inc(&r1_bio->remaining);
1471 trace_block_bio_remap(mbio->bi_disk->queue,
1472 mbio, disk_devt(mddev->gendisk),
1474 /* flush_pending_writes() needs access to the rdev so...*/
1475 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1477 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1479 plug = container_of(cb, struct raid1_plug_cb, cb);
1483 bio_list_add(&plug->pending, mbio);
1484 plug->pending_cnt++;
1486 spin_lock_irqsave(&conf->device_lock, flags);
1487 bio_list_add(&conf->pending_bio_list, mbio);
1488 conf->pending_count++;
1489 spin_unlock_irqrestore(&conf->device_lock, flags);
1490 md_wakeup_thread(mddev->thread);
1494 r1_bio_write_done(r1_bio);
1496 /* In case raid1d snuck in to freeze_array */
1497 wake_up(&conf->wait_barrier);
1500 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1504 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1505 md_flush_request(mddev, bio);
1510 * There is a limit to the maximum size, but
1511 * the read/write handler might find a lower limit
1512 * due to bad blocks. To avoid multiple splits,
1513 * we pass the maximum number of sectors down
1514 * and let the lower level perform the split.
1516 sectors = align_to_barrier_unit_end(
1517 bio->bi_iter.bi_sector, bio_sectors(bio));
1519 if (bio_data_dir(bio) == READ)
1520 raid1_read_request(mddev, bio, sectors, NULL);
1522 if (!md_write_start(mddev,bio))
1524 raid1_write_request(mddev, bio, sectors);
1529 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1531 struct r1conf *conf = mddev->private;
1534 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1535 conf->raid_disks - mddev->degraded);
1537 for (i = 0; i < conf->raid_disks; i++) {
1538 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1539 seq_printf(seq, "%s",
1540 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1543 seq_printf(seq, "]");
1546 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1548 char b[BDEVNAME_SIZE];
1549 struct r1conf *conf = mddev->private;
1550 unsigned long flags;
1553 * If it is not operational, then we have already marked it as dead
1554 * else if it is the last working disks, ignore the error, let the
1555 * next level up know.
1556 * else mark the drive as failed
1558 spin_lock_irqsave(&conf->device_lock, flags);
1559 if (test_bit(In_sync, &rdev->flags)
1560 && (conf->raid_disks - mddev->degraded) == 1) {
1562 * Don't fail the drive, act as though we were just a
1563 * normal single drive.
1564 * However don't try a recovery from this drive as
1565 * it is very likely to fail.
1567 conf->recovery_disabled = mddev->recovery_disabled;
1568 spin_unlock_irqrestore(&conf->device_lock, flags);
1571 set_bit(Blocked, &rdev->flags);
1572 if (test_and_clear_bit(In_sync, &rdev->flags))
1574 set_bit(Faulty, &rdev->flags);
1575 spin_unlock_irqrestore(&conf->device_lock, flags);
1577 * if recovery is running, make sure it aborts.
1579 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1580 set_mask_bits(&mddev->sb_flags, 0,
1581 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1582 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1583 "md/raid1:%s: Operation continuing on %d devices.\n",
1584 mdname(mddev), bdevname(rdev->bdev, b),
1585 mdname(mddev), conf->raid_disks - mddev->degraded);
1588 static void print_conf(struct r1conf *conf)
1592 pr_debug("RAID1 conf printout:\n");
1594 pr_debug("(!conf)\n");
1597 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1601 for (i = 0; i < conf->raid_disks; i++) {
1602 char b[BDEVNAME_SIZE];
1603 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1605 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1606 i, !test_bit(In_sync, &rdev->flags),
1607 !test_bit(Faulty, &rdev->flags),
1608 bdevname(rdev->bdev,b));
1613 static void close_sync(struct r1conf *conf)
1617 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1618 _wait_barrier(conf, idx);
1619 _allow_barrier(conf, idx);
1622 mempool_exit(&conf->r1buf_pool);
1625 static int raid1_spare_active(struct mddev *mddev)
1628 struct r1conf *conf = mddev->private;
1630 unsigned long flags;
1633 * Find all failed disks within the RAID1 configuration
1634 * and mark them readable.
1635 * Called under mddev lock, so rcu protection not needed.
1636 * device_lock used to avoid races with raid1_end_read_request
1637 * which expects 'In_sync' flags and ->degraded to be consistent.
1639 spin_lock_irqsave(&conf->device_lock, flags);
1640 for (i = 0; i < conf->raid_disks; i++) {
1641 struct md_rdev *rdev = conf->mirrors[i].rdev;
1642 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1644 && !test_bit(Candidate, &repl->flags)
1645 && repl->recovery_offset == MaxSector
1646 && !test_bit(Faulty, &repl->flags)
1647 && !test_and_set_bit(In_sync, &repl->flags)) {
1648 /* replacement has just become active */
1650 !test_and_clear_bit(In_sync, &rdev->flags))
1653 /* Replaced device not technically
1654 * faulty, but we need to be sure
1655 * it gets removed and never re-added
1657 set_bit(Faulty, &rdev->flags);
1658 sysfs_notify_dirent_safe(
1663 && rdev->recovery_offset == MaxSector
1664 && !test_bit(Faulty, &rdev->flags)
1665 && !test_and_set_bit(In_sync, &rdev->flags)) {
1667 sysfs_notify_dirent_safe(rdev->sysfs_state);
1670 mddev->degraded -= count;
1671 spin_unlock_irqrestore(&conf->device_lock, flags);
1677 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1679 struct r1conf *conf = mddev->private;
1682 struct raid1_info *p;
1684 int last = conf->raid_disks - 1;
1686 if (mddev->recovery_disabled == conf->recovery_disabled)
1689 if (md_integrity_add_rdev(rdev, mddev))
1692 if (rdev->raid_disk >= 0)
1693 first = last = rdev->raid_disk;
1696 * find the disk ... but prefer rdev->saved_raid_disk
1699 if (rdev->saved_raid_disk >= 0 &&
1700 rdev->saved_raid_disk >= first &&
1701 rdev->saved_raid_disk < conf->raid_disks &&
1702 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1703 first = last = rdev->saved_raid_disk;
1705 for (mirror = first; mirror <= last; mirror++) {
1706 p = conf->mirrors + mirror;
1709 disk_stack_limits(mddev->gendisk, rdev->bdev,
1710 rdev->data_offset << 9);
1712 p->head_position = 0;
1713 rdev->raid_disk = mirror;
1715 /* As all devices are equivalent, we don't need a full recovery
1716 * if this was recently any drive of the array
1718 if (rdev->saved_raid_disk < 0)
1720 rcu_assign_pointer(p->rdev, rdev);
1723 if (test_bit(WantReplacement, &p->rdev->flags) &&
1724 p[conf->raid_disks].rdev == NULL) {
1725 /* Add this device as a replacement */
1726 clear_bit(In_sync, &rdev->flags);
1727 set_bit(Replacement, &rdev->flags);
1728 rdev->raid_disk = mirror;
1731 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1735 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1736 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1741 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1743 struct r1conf *conf = mddev->private;
1745 int number = rdev->raid_disk;
1746 struct raid1_info *p = conf->mirrors + number;
1748 if (rdev != p->rdev)
1749 p = conf->mirrors + conf->raid_disks + number;
1752 if (rdev == p->rdev) {
1753 if (test_bit(In_sync, &rdev->flags) ||
1754 atomic_read(&rdev->nr_pending)) {
1758 /* Only remove non-faulty devices if recovery
1761 if (!test_bit(Faulty, &rdev->flags) &&
1762 mddev->recovery_disabled != conf->recovery_disabled &&
1763 mddev->degraded < conf->raid_disks) {
1768 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1770 if (atomic_read(&rdev->nr_pending)) {
1771 /* lost the race, try later */
1777 if (conf->mirrors[conf->raid_disks + number].rdev) {
1778 /* We just removed a device that is being replaced.
1779 * Move down the replacement. We drain all IO before
1780 * doing this to avoid confusion.
1782 struct md_rdev *repl =
1783 conf->mirrors[conf->raid_disks + number].rdev;
1784 freeze_array(conf, 0);
1785 if (atomic_read(&repl->nr_pending)) {
1786 /* It means that some queued IO of retry_list
1787 * hold repl. Thus, we cannot set replacement
1788 * as NULL, avoiding rdev NULL pointer
1789 * dereference in sync_request_write and
1790 * handle_write_finished.
1793 unfreeze_array(conf);
1796 clear_bit(Replacement, &repl->flags);
1798 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1799 unfreeze_array(conf);
1802 clear_bit(WantReplacement, &rdev->flags);
1803 err = md_integrity_register(mddev);
1811 static void end_sync_read(struct bio *bio)
1813 struct r1bio *r1_bio = get_resync_r1bio(bio);
1815 update_head_pos(r1_bio->read_disk, r1_bio);
1818 * we have read a block, now it needs to be re-written,
1819 * or re-read if the read failed.
1820 * We don't do much here, just schedule handling by raid1d
1822 if (!bio->bi_status)
1823 set_bit(R1BIO_Uptodate, &r1_bio->state);
1825 if (atomic_dec_and_test(&r1_bio->remaining))
1826 reschedule_retry(r1_bio);
1829 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1831 sector_t sync_blocks = 0;
1832 sector_t s = r1_bio->sector;
1833 long sectors_to_go = r1_bio->sectors;
1835 /* make sure these bits don't get cleared. */
1837 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1839 sectors_to_go -= sync_blocks;
1840 } while (sectors_to_go > 0);
1843 static void end_sync_write(struct bio *bio)
1845 int uptodate = !bio->bi_status;
1846 struct r1bio *r1_bio = get_resync_r1bio(bio);
1847 struct mddev *mddev = r1_bio->mddev;
1848 struct r1conf *conf = mddev->private;
1851 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1854 abort_sync_write(mddev, r1_bio);
1855 set_bit(WriteErrorSeen, &rdev->flags);
1856 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1857 set_bit(MD_RECOVERY_NEEDED, &
1859 set_bit(R1BIO_WriteError, &r1_bio->state);
1860 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1861 &first_bad, &bad_sectors) &&
1862 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1865 &first_bad, &bad_sectors)
1867 set_bit(R1BIO_MadeGood, &r1_bio->state);
1869 if (atomic_dec_and_test(&r1_bio->remaining)) {
1870 int s = r1_bio->sectors;
1871 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1872 test_bit(R1BIO_WriteError, &r1_bio->state))
1873 reschedule_retry(r1_bio);
1876 md_done_sync(mddev, s, uptodate);
1881 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1882 int sectors, struct page *page, int rw)
1884 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1888 set_bit(WriteErrorSeen, &rdev->flags);
1889 if (!test_and_set_bit(WantReplacement,
1891 set_bit(MD_RECOVERY_NEEDED, &
1892 rdev->mddev->recovery);
1894 /* need to record an error - either for the block or the device */
1895 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1896 md_error(rdev->mddev, rdev);
1900 static int fix_sync_read_error(struct r1bio *r1_bio)
1902 /* Try some synchronous reads of other devices to get
1903 * good data, much like with normal read errors. Only
1904 * read into the pages we already have so we don't
1905 * need to re-issue the read request.
1906 * We don't need to freeze the array, because being in an
1907 * active sync request, there is no normal IO, and
1908 * no overlapping syncs.
1909 * We don't need to check is_badblock() again as we
1910 * made sure that anything with a bad block in range
1911 * will have bi_end_io clear.
1913 struct mddev *mddev = r1_bio->mddev;
1914 struct r1conf *conf = mddev->private;
1915 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1916 struct page **pages = get_resync_pages(bio)->pages;
1917 sector_t sect = r1_bio->sector;
1918 int sectors = r1_bio->sectors;
1920 struct md_rdev *rdev;
1922 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1923 if (test_bit(FailFast, &rdev->flags)) {
1924 /* Don't try recovering from here - just fail it
1925 * ... unless it is the last working device of course */
1926 md_error(mddev, rdev);
1927 if (test_bit(Faulty, &rdev->flags))
1928 /* Don't try to read from here, but make sure
1929 * put_buf does it's thing
1931 bio->bi_end_io = end_sync_write;
1936 int d = r1_bio->read_disk;
1940 if (s > (PAGE_SIZE>>9))
1943 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1944 /* No rcu protection needed here devices
1945 * can only be removed when no resync is
1946 * active, and resync is currently active
1948 rdev = conf->mirrors[d].rdev;
1949 if (sync_page_io(rdev, sect, s<<9,
1951 REQ_OP_READ, 0, false)) {
1957 if (d == conf->raid_disks * 2)
1959 } while (!success && d != r1_bio->read_disk);
1962 char b[BDEVNAME_SIZE];
1964 /* Cannot read from anywhere, this block is lost.
1965 * Record a bad block on each device. If that doesn't
1966 * work just disable and interrupt the recovery.
1967 * Don't fail devices as that won't really help.
1969 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1970 mdname(mddev), bio_devname(bio, b),
1971 (unsigned long long)r1_bio->sector);
1972 for (d = 0; d < conf->raid_disks * 2; d++) {
1973 rdev = conf->mirrors[d].rdev;
1974 if (!rdev || test_bit(Faulty, &rdev->flags))
1976 if (!rdev_set_badblocks(rdev, sect, s, 0))
1980 conf->recovery_disabled =
1981 mddev->recovery_disabled;
1982 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1983 md_done_sync(mddev, r1_bio->sectors, 0);
1995 /* write it back and re-read */
1996 while (d != r1_bio->read_disk) {
1998 d = conf->raid_disks * 2;
2000 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2002 rdev = conf->mirrors[d].rdev;
2003 if (r1_sync_page_io(rdev, sect, s,
2006 r1_bio->bios[d]->bi_end_io = NULL;
2007 rdev_dec_pending(rdev, mddev);
2011 while (d != r1_bio->read_disk) {
2013 d = conf->raid_disks * 2;
2015 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2017 rdev = conf->mirrors[d].rdev;
2018 if (r1_sync_page_io(rdev, sect, s,
2021 atomic_add(s, &rdev->corrected_errors);
2027 set_bit(R1BIO_Uptodate, &r1_bio->state);
2032 static void process_checks(struct r1bio *r1_bio)
2034 /* We have read all readable devices. If we haven't
2035 * got the block, then there is no hope left.
2036 * If we have, then we want to do a comparison
2037 * and skip the write if everything is the same.
2038 * If any blocks failed to read, then we need to
2039 * attempt an over-write
2041 struct mddev *mddev = r1_bio->mddev;
2042 struct r1conf *conf = mddev->private;
2047 /* Fix variable parts of all bios */
2048 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2049 for (i = 0; i < conf->raid_disks * 2; i++) {
2050 blk_status_t status;
2051 struct bio *b = r1_bio->bios[i];
2052 struct resync_pages *rp = get_resync_pages(b);
2053 if (b->bi_end_io != end_sync_read)
2055 /* fixup the bio for reuse, but preserve errno */
2056 status = b->bi_status;
2058 b->bi_status = status;
2059 b->bi_iter.bi_sector = r1_bio->sector +
2060 conf->mirrors[i].rdev->data_offset;
2061 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2062 b->bi_end_io = end_sync_read;
2063 rp->raid_bio = r1_bio;
2066 /* initialize bvec table again */
2067 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2069 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2070 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2071 !r1_bio->bios[primary]->bi_status) {
2072 r1_bio->bios[primary]->bi_end_io = NULL;
2073 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2076 r1_bio->read_disk = primary;
2077 for (i = 0; i < conf->raid_disks * 2; i++) {
2079 struct bio *pbio = r1_bio->bios[primary];
2080 struct bio *sbio = r1_bio->bios[i];
2081 blk_status_t status = sbio->bi_status;
2082 struct page **ppages = get_resync_pages(pbio)->pages;
2083 struct page **spages = get_resync_pages(sbio)->pages;
2085 int page_len[RESYNC_PAGES] = { 0 };
2086 struct bvec_iter_all iter_all;
2088 if (sbio->bi_end_io != end_sync_read)
2090 /* Now we can 'fixup' the error value */
2091 sbio->bi_status = 0;
2093 bio_for_each_segment_all(bi, sbio, iter_all)
2094 page_len[j++] = bi->bv_len;
2097 for (j = vcnt; j-- ; ) {
2098 if (memcmp(page_address(ppages[j]),
2099 page_address(spages[j]),
2106 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2107 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2109 /* No need to write to this device. */
2110 sbio->bi_end_io = NULL;
2111 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2115 bio_copy_data(sbio, pbio);
2119 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2121 struct r1conf *conf = mddev->private;
2123 int disks = conf->raid_disks * 2;
2126 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2127 /* ouch - failed to read all of that. */
2128 if (!fix_sync_read_error(r1_bio))
2131 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2132 process_checks(r1_bio);
2137 atomic_set(&r1_bio->remaining, 1);
2138 for (i = 0; i < disks ; i++) {
2139 wbio = r1_bio->bios[i];
2140 if (wbio->bi_end_io == NULL ||
2141 (wbio->bi_end_io == end_sync_read &&
2142 (i == r1_bio->read_disk ||
2143 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2145 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2146 abort_sync_write(mddev, r1_bio);
2150 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2151 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2152 wbio->bi_opf |= MD_FAILFAST;
2154 wbio->bi_end_io = end_sync_write;
2155 atomic_inc(&r1_bio->remaining);
2156 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2158 generic_make_request(wbio);
2161 if (atomic_dec_and_test(&r1_bio->remaining)) {
2162 /* if we're here, all write(s) have completed, so clean up */
2163 int s = r1_bio->sectors;
2164 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2165 test_bit(R1BIO_WriteError, &r1_bio->state))
2166 reschedule_retry(r1_bio);
2169 md_done_sync(mddev, s, 1);
2175 * This is a kernel thread which:
2177 * 1. Retries failed read operations on working mirrors.
2178 * 2. Updates the raid superblock when problems encounter.
2179 * 3. Performs writes following reads for array synchronising.
2182 static void fix_read_error(struct r1conf *conf, int read_disk,
2183 sector_t sect, int sectors)
2185 struct mddev *mddev = conf->mddev;
2191 struct md_rdev *rdev;
2193 if (s > (PAGE_SIZE>>9))
2201 rdev = rcu_dereference(conf->mirrors[d].rdev);
2203 (test_bit(In_sync, &rdev->flags) ||
2204 (!test_bit(Faulty, &rdev->flags) &&
2205 rdev->recovery_offset >= sect + s)) &&
2206 is_badblock(rdev, sect, s,
2207 &first_bad, &bad_sectors) == 0) {
2208 atomic_inc(&rdev->nr_pending);
2210 if (sync_page_io(rdev, sect, s<<9,
2211 conf->tmppage, REQ_OP_READ, 0, false))
2213 rdev_dec_pending(rdev, mddev);
2219 if (d == conf->raid_disks * 2)
2221 } while (!success && d != read_disk);
2224 /* Cannot read from anywhere - mark it bad */
2225 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2226 if (!rdev_set_badblocks(rdev, sect, s, 0))
2227 md_error(mddev, rdev);
2230 /* write it back and re-read */
2232 while (d != read_disk) {
2234 d = conf->raid_disks * 2;
2237 rdev = rcu_dereference(conf->mirrors[d].rdev);
2239 !test_bit(Faulty, &rdev->flags)) {
2240 atomic_inc(&rdev->nr_pending);
2242 r1_sync_page_io(rdev, sect, s,
2243 conf->tmppage, WRITE);
2244 rdev_dec_pending(rdev, mddev);
2249 while (d != read_disk) {
2250 char b[BDEVNAME_SIZE];
2252 d = conf->raid_disks * 2;
2255 rdev = rcu_dereference(conf->mirrors[d].rdev);
2257 !test_bit(Faulty, &rdev->flags)) {
2258 atomic_inc(&rdev->nr_pending);
2260 if (r1_sync_page_io(rdev, sect, s,
2261 conf->tmppage, READ)) {
2262 atomic_add(s, &rdev->corrected_errors);
2263 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2265 (unsigned long long)(sect +
2267 bdevname(rdev->bdev, b));
2269 rdev_dec_pending(rdev, mddev);
2278 static int narrow_write_error(struct r1bio *r1_bio, int i)
2280 struct mddev *mddev = r1_bio->mddev;
2281 struct r1conf *conf = mddev->private;
2282 struct md_rdev *rdev = conf->mirrors[i].rdev;
2284 /* bio has the data to be written to device 'i' where
2285 * we just recently had a write error.
2286 * We repeatedly clone the bio and trim down to one block,
2287 * then try the write. Where the write fails we record
2289 * It is conceivable that the bio doesn't exactly align with
2290 * blocks. We must handle this somehow.
2292 * We currently own a reference on the rdev.
2298 int sect_to_write = r1_bio->sectors;
2301 if (rdev->badblocks.shift < 0)
2304 block_sectors = roundup(1 << rdev->badblocks.shift,
2305 bdev_logical_block_size(rdev->bdev) >> 9);
2306 sector = r1_bio->sector;
2307 sectors = ((sector + block_sectors)
2308 & ~(sector_t)(block_sectors - 1))
2311 while (sect_to_write) {
2313 if (sectors > sect_to_write)
2314 sectors = sect_to_write;
2315 /* Write at 'sector' for 'sectors'*/
2317 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2318 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2322 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2326 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2327 wbio->bi_iter.bi_sector = r1_bio->sector;
2328 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2330 bio_trim(wbio, sector - r1_bio->sector, sectors);
2331 wbio->bi_iter.bi_sector += rdev->data_offset;
2332 bio_set_dev(wbio, rdev->bdev);
2334 if (submit_bio_wait(wbio) < 0)
2336 ok = rdev_set_badblocks(rdev, sector,
2341 sect_to_write -= sectors;
2343 sectors = block_sectors;
2348 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2351 int s = r1_bio->sectors;
2352 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2353 struct md_rdev *rdev = conf->mirrors[m].rdev;
2354 struct bio *bio = r1_bio->bios[m];
2355 if (bio->bi_end_io == NULL)
2357 if (!bio->bi_status &&
2358 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2359 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2361 if (bio->bi_status &&
2362 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2363 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2364 md_error(conf->mddev, rdev);
2368 md_done_sync(conf->mddev, s, 1);
2371 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2376 for (m = 0; m < conf->raid_disks * 2 ; m++)
2377 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2378 struct md_rdev *rdev = conf->mirrors[m].rdev;
2379 rdev_clear_badblocks(rdev,
2381 r1_bio->sectors, 0);
2382 rdev_dec_pending(rdev, conf->mddev);
2383 } else if (r1_bio->bios[m] != NULL) {
2384 /* This drive got a write error. We need to
2385 * narrow down and record precise write
2389 if (!narrow_write_error(r1_bio, m)) {
2390 md_error(conf->mddev,
2391 conf->mirrors[m].rdev);
2392 /* an I/O failed, we can't clear the bitmap */
2393 set_bit(R1BIO_Degraded, &r1_bio->state);
2395 rdev_dec_pending(conf->mirrors[m].rdev,
2399 spin_lock_irq(&conf->device_lock);
2400 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2401 idx = sector_to_idx(r1_bio->sector);
2402 atomic_inc(&conf->nr_queued[idx]);
2403 spin_unlock_irq(&conf->device_lock);
2405 * In case freeze_array() is waiting for condition
2406 * get_unqueued_pending() == extra to be true.
2408 wake_up(&conf->wait_barrier);
2409 md_wakeup_thread(conf->mddev->thread);
2411 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2412 close_write(r1_bio);
2413 raid_end_bio_io(r1_bio);
2417 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2419 struct mddev *mddev = conf->mddev;
2421 struct md_rdev *rdev;
2423 clear_bit(R1BIO_ReadError, &r1_bio->state);
2424 /* we got a read error. Maybe the drive is bad. Maybe just
2425 * the block and we can fix it.
2426 * We freeze all other IO, and try reading the block from
2427 * other devices. When we find one, we re-write
2428 * and check it that fixes the read error.
2429 * This is all done synchronously while the array is
2433 bio = r1_bio->bios[r1_bio->read_disk];
2435 r1_bio->bios[r1_bio->read_disk] = NULL;
2437 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2439 && !test_bit(FailFast, &rdev->flags)) {
2440 freeze_array(conf, 1);
2441 fix_read_error(conf, r1_bio->read_disk,
2442 r1_bio->sector, r1_bio->sectors);
2443 unfreeze_array(conf);
2444 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2445 md_error(mddev, rdev);
2447 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2450 rdev_dec_pending(rdev, conf->mddev);
2451 allow_barrier(conf, r1_bio->sector);
2452 bio = r1_bio->master_bio;
2454 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2456 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2459 static void raid1d(struct md_thread *thread)
2461 struct mddev *mddev = thread->mddev;
2462 struct r1bio *r1_bio;
2463 unsigned long flags;
2464 struct r1conf *conf = mddev->private;
2465 struct list_head *head = &conf->retry_list;
2466 struct blk_plug plug;
2469 md_check_recovery(mddev);
2471 if (!list_empty_careful(&conf->bio_end_io_list) &&
2472 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2474 spin_lock_irqsave(&conf->device_lock, flags);
2475 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2476 list_splice_init(&conf->bio_end_io_list, &tmp);
2477 spin_unlock_irqrestore(&conf->device_lock, flags);
2478 while (!list_empty(&tmp)) {
2479 r1_bio = list_first_entry(&tmp, struct r1bio,
2481 list_del(&r1_bio->retry_list);
2482 idx = sector_to_idx(r1_bio->sector);
2483 atomic_dec(&conf->nr_queued[idx]);
2484 if (mddev->degraded)
2485 set_bit(R1BIO_Degraded, &r1_bio->state);
2486 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2487 close_write(r1_bio);
2488 raid_end_bio_io(r1_bio);
2492 blk_start_plug(&plug);
2495 flush_pending_writes(conf);
2497 spin_lock_irqsave(&conf->device_lock, flags);
2498 if (list_empty(head)) {
2499 spin_unlock_irqrestore(&conf->device_lock, flags);
2502 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2503 list_del(head->prev);
2504 idx = sector_to_idx(r1_bio->sector);
2505 atomic_dec(&conf->nr_queued[idx]);
2506 spin_unlock_irqrestore(&conf->device_lock, flags);
2508 mddev = r1_bio->mddev;
2509 conf = mddev->private;
2510 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2511 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2512 test_bit(R1BIO_WriteError, &r1_bio->state))
2513 handle_sync_write_finished(conf, r1_bio);
2515 sync_request_write(mddev, r1_bio);
2516 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2517 test_bit(R1BIO_WriteError, &r1_bio->state))
2518 handle_write_finished(conf, r1_bio);
2519 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2520 handle_read_error(conf, r1_bio);
2525 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2526 md_check_recovery(mddev);
2528 blk_finish_plug(&plug);
2531 static int init_resync(struct r1conf *conf)
2535 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2536 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2538 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2539 r1buf_pool_free, conf->poolinfo);
2542 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2544 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2545 struct resync_pages *rps;
2549 for (i = conf->poolinfo->raid_disks; i--; ) {
2550 bio = r1bio->bios[i];
2551 rps = bio->bi_private;
2553 bio->bi_private = rps;
2555 r1bio->master_bio = NULL;
2560 * perform a "sync" on one "block"
2562 * We need to make sure that no normal I/O request - particularly write
2563 * requests - conflict with active sync requests.
2565 * This is achieved by tracking pending requests and a 'barrier' concept
2566 * that can be installed to exclude normal IO requests.
2569 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2572 struct r1conf *conf = mddev->private;
2573 struct r1bio *r1_bio;
2575 sector_t max_sector, nr_sectors;
2579 int write_targets = 0, read_targets = 0;
2580 sector_t sync_blocks;
2581 int still_degraded = 0;
2582 int good_sectors = RESYNC_SECTORS;
2583 int min_bad = 0; /* number of sectors that are bad in all devices */
2584 int idx = sector_to_idx(sector_nr);
2587 if (!mempool_initialized(&conf->r1buf_pool))
2588 if (init_resync(conf))
2591 max_sector = mddev->dev_sectors;
2592 if (sector_nr >= max_sector) {
2593 /* If we aborted, we need to abort the
2594 * sync on the 'current' bitmap chunk (there will
2595 * only be one in raid1 resync.
2596 * We can find the current addess in mddev->curr_resync
2598 if (mddev->curr_resync < max_sector) /* aborted */
2599 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2601 else /* completed sync */
2604 md_bitmap_close_sync(mddev->bitmap);
2607 if (mddev_is_clustered(mddev)) {
2608 conf->cluster_sync_low = 0;
2609 conf->cluster_sync_high = 0;
2614 if (mddev->bitmap == NULL &&
2615 mddev->recovery_cp == MaxSector &&
2616 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2617 conf->fullsync == 0) {
2619 return max_sector - sector_nr;
2621 /* before building a request, check if we can skip these blocks..
2622 * This call the bitmap_start_sync doesn't actually record anything
2624 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2625 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2626 /* We can skip this block, and probably several more */
2632 * If there is non-resync activity waiting for a turn, then let it
2633 * though before starting on this new sync request.
2635 if (atomic_read(&conf->nr_waiting[idx]))
2636 schedule_timeout_uninterruptible(1);
2638 /* we are incrementing sector_nr below. To be safe, we check against
2639 * sector_nr + two times RESYNC_SECTORS
2642 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2643 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2646 if (raise_barrier(conf, sector_nr))
2649 r1_bio = raid1_alloc_init_r1buf(conf);
2653 * If we get a correctably read error during resync or recovery,
2654 * we might want to read from a different device. So we
2655 * flag all drives that could conceivably be read from for READ,
2656 * and any others (which will be non-In_sync devices) for WRITE.
2657 * If a read fails, we try reading from something else for which READ
2661 r1_bio->mddev = mddev;
2662 r1_bio->sector = sector_nr;
2664 set_bit(R1BIO_IsSync, &r1_bio->state);
2665 /* make sure good_sectors won't go across barrier unit boundary */
2666 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2668 for (i = 0; i < conf->raid_disks * 2; i++) {
2669 struct md_rdev *rdev;
2670 bio = r1_bio->bios[i];
2672 rdev = rcu_dereference(conf->mirrors[i].rdev);
2674 test_bit(Faulty, &rdev->flags)) {
2675 if (i < conf->raid_disks)
2677 } else if (!test_bit(In_sync, &rdev->flags)) {
2678 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2679 bio->bi_end_io = end_sync_write;
2682 /* may need to read from here */
2683 sector_t first_bad = MaxSector;
2686 if (is_badblock(rdev, sector_nr, good_sectors,
2687 &first_bad, &bad_sectors)) {
2688 if (first_bad > sector_nr)
2689 good_sectors = first_bad - sector_nr;
2691 bad_sectors -= (sector_nr - first_bad);
2693 min_bad > bad_sectors)
2694 min_bad = bad_sectors;
2697 if (sector_nr < first_bad) {
2698 if (test_bit(WriteMostly, &rdev->flags)) {
2705 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2706 bio->bi_end_io = end_sync_read;
2708 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2709 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2710 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2712 * The device is suitable for reading (InSync),
2713 * but has bad block(s) here. Let's try to correct them,
2714 * if we are doing resync or repair. Otherwise, leave
2715 * this device alone for this sync request.
2717 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2718 bio->bi_end_io = end_sync_write;
2722 if (bio->bi_end_io) {
2723 atomic_inc(&rdev->nr_pending);
2724 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2725 bio_set_dev(bio, rdev->bdev);
2726 if (test_bit(FailFast, &rdev->flags))
2727 bio->bi_opf |= MD_FAILFAST;
2733 r1_bio->read_disk = disk;
2735 if (read_targets == 0 && min_bad > 0) {
2736 /* These sectors are bad on all InSync devices, so we
2737 * need to mark them bad on all write targets
2740 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2741 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2742 struct md_rdev *rdev = conf->mirrors[i].rdev;
2743 ok = rdev_set_badblocks(rdev, sector_nr,
2747 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2752 /* Cannot record the badblocks, so need to
2754 * If there are multiple read targets, could just
2755 * fail the really bad ones ???
2757 conf->recovery_disabled = mddev->recovery_disabled;
2758 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2764 if (min_bad > 0 && min_bad < good_sectors) {
2765 /* only resync enough to reach the next bad->good
2767 good_sectors = min_bad;
2770 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2771 /* extra read targets are also write targets */
2772 write_targets += read_targets-1;
2774 if (write_targets == 0 || read_targets == 0) {
2775 /* There is nowhere to write, so all non-sync
2776 * drives must be failed - so we are finished
2780 max_sector = sector_nr + min_bad;
2781 rv = max_sector - sector_nr;
2787 if (max_sector > mddev->resync_max)
2788 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2789 if (max_sector > sector_nr + good_sectors)
2790 max_sector = sector_nr + good_sectors;
2795 int len = PAGE_SIZE;
2796 if (sector_nr + (len>>9) > max_sector)
2797 len = (max_sector - sector_nr) << 9;
2800 if (sync_blocks == 0) {
2801 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2802 &sync_blocks, still_degraded) &&
2804 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2806 if ((len >> 9) > sync_blocks)
2807 len = sync_blocks<<9;
2810 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2811 struct resync_pages *rp;
2813 bio = r1_bio->bios[i];
2814 rp = get_resync_pages(bio);
2815 if (bio->bi_end_io) {
2816 page = resync_fetch_page(rp, page_idx);
2819 * won't fail because the vec table is big
2820 * enough to hold all these pages
2822 bio_add_page(bio, page, len, 0);
2825 nr_sectors += len>>9;
2826 sector_nr += len>>9;
2827 sync_blocks -= (len>>9);
2828 } while (++page_idx < RESYNC_PAGES);
2830 r1_bio->sectors = nr_sectors;
2832 if (mddev_is_clustered(mddev) &&
2833 conf->cluster_sync_high < sector_nr + nr_sectors) {
2834 conf->cluster_sync_low = mddev->curr_resync_completed;
2835 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2836 /* Send resync message */
2837 md_cluster_ops->resync_info_update(mddev,
2838 conf->cluster_sync_low,
2839 conf->cluster_sync_high);
2842 /* For a user-requested sync, we read all readable devices and do a
2845 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2846 atomic_set(&r1_bio->remaining, read_targets);
2847 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2848 bio = r1_bio->bios[i];
2849 if (bio->bi_end_io == end_sync_read) {
2851 md_sync_acct_bio(bio, nr_sectors);
2852 if (read_targets == 1)
2853 bio->bi_opf &= ~MD_FAILFAST;
2854 generic_make_request(bio);
2858 atomic_set(&r1_bio->remaining, 1);
2859 bio = r1_bio->bios[r1_bio->read_disk];
2860 md_sync_acct_bio(bio, nr_sectors);
2861 if (read_targets == 1)
2862 bio->bi_opf &= ~MD_FAILFAST;
2863 generic_make_request(bio);
2868 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2873 return mddev->dev_sectors;
2876 static struct r1conf *setup_conf(struct mddev *mddev)
2878 struct r1conf *conf;
2880 struct raid1_info *disk;
2881 struct md_rdev *rdev;
2884 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2888 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2889 sizeof(atomic_t), GFP_KERNEL);
2890 if (!conf->nr_pending)
2893 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2894 sizeof(atomic_t), GFP_KERNEL);
2895 if (!conf->nr_waiting)
2898 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2899 sizeof(atomic_t), GFP_KERNEL);
2900 if (!conf->nr_queued)
2903 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2904 sizeof(atomic_t), GFP_KERNEL);
2908 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2909 mddev->raid_disks, 2),
2914 conf->tmppage = alloc_page(GFP_KERNEL);
2918 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2919 if (!conf->poolinfo)
2921 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2922 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2923 r1bio_pool_free, conf->poolinfo);
2927 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2931 conf->poolinfo->mddev = mddev;
2934 spin_lock_init(&conf->device_lock);
2935 rdev_for_each(rdev, mddev) {
2936 int disk_idx = rdev->raid_disk;
2937 if (disk_idx >= mddev->raid_disks
2940 if (test_bit(Replacement, &rdev->flags))
2941 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2943 disk = conf->mirrors + disk_idx;
2948 disk->head_position = 0;
2949 disk->seq_start = MaxSector;
2951 conf->raid_disks = mddev->raid_disks;
2952 conf->mddev = mddev;
2953 INIT_LIST_HEAD(&conf->retry_list);
2954 INIT_LIST_HEAD(&conf->bio_end_io_list);
2956 spin_lock_init(&conf->resync_lock);
2957 init_waitqueue_head(&conf->wait_barrier);
2959 bio_list_init(&conf->pending_bio_list);
2960 conf->pending_count = 0;
2961 conf->recovery_disabled = mddev->recovery_disabled - 1;
2964 for (i = 0; i < conf->raid_disks * 2; i++) {
2966 disk = conf->mirrors + i;
2968 if (i < conf->raid_disks &&
2969 disk[conf->raid_disks].rdev) {
2970 /* This slot has a replacement. */
2972 /* No original, just make the replacement
2973 * a recovering spare
2976 disk[conf->raid_disks].rdev;
2977 disk[conf->raid_disks].rdev = NULL;
2978 } else if (!test_bit(In_sync, &disk->rdev->flags))
2979 /* Original is not in_sync - bad */
2984 !test_bit(In_sync, &disk->rdev->flags)) {
2985 disk->head_position = 0;
2987 (disk->rdev->saved_raid_disk < 0))
2993 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3001 mempool_exit(&conf->r1bio_pool);
3002 kfree(conf->mirrors);
3003 safe_put_page(conf->tmppage);
3004 kfree(conf->poolinfo);
3005 kfree(conf->nr_pending);
3006 kfree(conf->nr_waiting);
3007 kfree(conf->nr_queued);
3008 kfree(conf->barrier);
3009 bioset_exit(&conf->bio_split);
3012 return ERR_PTR(err);
3015 static void raid1_free(struct mddev *mddev, void *priv);
3016 static int raid1_run(struct mddev *mddev)
3018 struct r1conf *conf;
3020 struct md_rdev *rdev;
3022 bool discard_supported = false;
3024 if (mddev->level != 1) {
3025 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3026 mdname(mddev), mddev->level);
3029 if (mddev->reshape_position != MaxSector) {
3030 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3034 if (mddev_init_writes_pending(mddev) < 0)
3037 * copy the already verified devices into our private RAID1
3038 * bookkeeping area. [whatever we allocate in run(),
3039 * should be freed in raid1_free()]
3041 if (mddev->private == NULL)
3042 conf = setup_conf(mddev);
3044 conf = mddev->private;
3047 return PTR_ERR(conf);
3050 blk_queue_max_write_same_sectors(mddev->queue, 0);
3051 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3054 rdev_for_each(rdev, mddev) {
3055 if (!mddev->gendisk)
3057 disk_stack_limits(mddev->gendisk, rdev->bdev,
3058 rdev->data_offset << 9);
3059 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3060 discard_supported = true;
3063 mddev->degraded = 0;
3064 for (i = 0; i < conf->raid_disks; i++)
3065 if (conf->mirrors[i].rdev == NULL ||
3066 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3067 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3070 if (conf->raid_disks - mddev->degraded == 1)
3071 mddev->recovery_cp = MaxSector;
3073 if (mddev->recovery_cp != MaxSector)
3074 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3076 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3077 mdname(mddev), mddev->raid_disks - mddev->degraded,
3081 * Ok, everything is just fine now
3083 mddev->thread = conf->thread;
3084 conf->thread = NULL;
3085 mddev->private = conf;
3086 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3088 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3091 if (discard_supported)
3092 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3095 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3099 ret = md_integrity_register(mddev);
3101 md_unregister_thread(&mddev->thread);
3102 raid1_free(mddev, conf);
3107 static void raid1_free(struct mddev *mddev, void *priv)
3109 struct r1conf *conf = priv;
3111 mempool_exit(&conf->r1bio_pool);
3112 kfree(conf->mirrors);
3113 safe_put_page(conf->tmppage);
3114 kfree(conf->poolinfo);
3115 kfree(conf->nr_pending);
3116 kfree(conf->nr_waiting);
3117 kfree(conf->nr_queued);
3118 kfree(conf->barrier);
3119 bioset_exit(&conf->bio_split);
3123 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3125 /* no resync is happening, and there is enough space
3126 * on all devices, so we can resize.
3127 * We need to make sure resync covers any new space.
3128 * If the array is shrinking we should possibly wait until
3129 * any io in the removed space completes, but it hardly seems
3132 sector_t newsize = raid1_size(mddev, sectors, 0);
3133 if (mddev->external_size &&
3134 mddev->array_sectors > newsize)
3136 if (mddev->bitmap) {
3137 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3141 md_set_array_sectors(mddev, newsize);
3142 if (sectors > mddev->dev_sectors &&
3143 mddev->recovery_cp > mddev->dev_sectors) {
3144 mddev->recovery_cp = mddev->dev_sectors;
3145 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3147 mddev->dev_sectors = sectors;
3148 mddev->resync_max_sectors = sectors;
3152 static int raid1_reshape(struct mddev *mddev)
3155 * 1/ resize the r1bio_pool
3156 * 2/ resize conf->mirrors
3158 * We allocate a new r1bio_pool if we can.
3159 * Then raise a device barrier and wait until all IO stops.
3160 * Then resize conf->mirrors and swap in the new r1bio pool.
3162 * At the same time, we "pack" the devices so that all the missing
3163 * devices have the higher raid_disk numbers.
3165 mempool_t newpool, oldpool;
3166 struct pool_info *newpoolinfo;
3167 struct raid1_info *newmirrors;
3168 struct r1conf *conf = mddev->private;
3169 int cnt, raid_disks;
3170 unsigned long flags;
3174 memset(&newpool, 0, sizeof(newpool));
3175 memset(&oldpool, 0, sizeof(oldpool));
3177 /* Cannot change chunk_size, layout, or level */
3178 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3179 mddev->layout != mddev->new_layout ||
3180 mddev->level != mddev->new_level) {
3181 mddev->new_chunk_sectors = mddev->chunk_sectors;
3182 mddev->new_layout = mddev->layout;
3183 mddev->new_level = mddev->level;
3187 if (!mddev_is_clustered(mddev))
3188 md_allow_write(mddev);
3190 raid_disks = mddev->raid_disks + mddev->delta_disks;
3192 if (raid_disks < conf->raid_disks) {
3194 for (d= 0; d < conf->raid_disks; d++)
3195 if (conf->mirrors[d].rdev)
3197 if (cnt > raid_disks)
3201 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3204 newpoolinfo->mddev = mddev;
3205 newpoolinfo->raid_disks = raid_disks * 2;
3207 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3208 r1bio_pool_free, newpoolinfo);
3213 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3218 mempool_exit(&newpool);
3222 freeze_array(conf, 0);
3224 /* ok, everything is stopped */
3225 oldpool = conf->r1bio_pool;
3226 conf->r1bio_pool = newpool;
3228 for (d = d2 = 0; d < conf->raid_disks; d++) {
3229 struct md_rdev *rdev = conf->mirrors[d].rdev;
3230 if (rdev && rdev->raid_disk != d2) {
3231 sysfs_unlink_rdev(mddev, rdev);
3232 rdev->raid_disk = d2;
3233 sysfs_unlink_rdev(mddev, rdev);
3234 if (sysfs_link_rdev(mddev, rdev))
3235 pr_warn("md/raid1:%s: cannot register rd%d\n",
3236 mdname(mddev), rdev->raid_disk);
3239 newmirrors[d2++].rdev = rdev;
3241 kfree(conf->mirrors);
3242 conf->mirrors = newmirrors;
3243 kfree(conf->poolinfo);
3244 conf->poolinfo = newpoolinfo;
3246 spin_lock_irqsave(&conf->device_lock, flags);
3247 mddev->degraded += (raid_disks - conf->raid_disks);
3248 spin_unlock_irqrestore(&conf->device_lock, flags);
3249 conf->raid_disks = mddev->raid_disks = raid_disks;
3250 mddev->delta_disks = 0;
3252 unfreeze_array(conf);
3254 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3255 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3256 md_wakeup_thread(mddev->thread);
3258 mempool_exit(&oldpool);
3262 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3264 struct r1conf *conf = mddev->private;
3267 freeze_array(conf, 0);
3269 unfreeze_array(conf);
3272 static void *raid1_takeover(struct mddev *mddev)
3274 /* raid1 can take over:
3275 * raid5 with 2 devices, any layout or chunk size
3277 if (mddev->level == 5 && mddev->raid_disks == 2) {
3278 struct r1conf *conf;
3279 mddev->new_level = 1;
3280 mddev->new_layout = 0;
3281 mddev->new_chunk_sectors = 0;
3282 conf = setup_conf(mddev);
3283 if (!IS_ERR(conf)) {
3284 /* Array must appear to be quiesced */
3285 conf->array_frozen = 1;
3286 mddev_clear_unsupported_flags(mddev,
3287 UNSUPPORTED_MDDEV_FLAGS);
3291 return ERR_PTR(-EINVAL);
3294 static struct md_personality raid1_personality =
3298 .owner = THIS_MODULE,
3299 .make_request = raid1_make_request,
3302 .status = raid1_status,
3303 .error_handler = raid1_error,
3304 .hot_add_disk = raid1_add_disk,
3305 .hot_remove_disk= raid1_remove_disk,
3306 .spare_active = raid1_spare_active,
3307 .sync_request = raid1_sync_request,
3308 .resize = raid1_resize,
3310 .check_reshape = raid1_reshape,
3311 .quiesce = raid1_quiesce,
3312 .takeover = raid1_takeover,
3313 .congested = raid1_congested,
3316 static int __init raid_init(void)
3318 return register_md_personality(&raid1_personality);
3321 static void raid_exit(void)
3323 unregister_md_personality(&raid1_personality);
3326 module_init(raid_init);
3327 module_exit(raid_exit);
3328 MODULE_LICENSE("GPL");
3329 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3330 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3331 MODULE_ALIAS("md-raid1");
3332 MODULE_ALIAS("md-level-1");
3334 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);