2 * raid1.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * RAID-1 management functions.
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
45 * Number of guaranteed r1bios in case of extreme VM load:
47 #define NR_RAID1_BIOS 256
49 /* when we get a read error on a read-only array, we redirect to another
50 * device without failing the first device, or trying to over-write to
51 * correct the read error. To keep track of bad blocks on a per-bio
52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
54 #define IO_BLOCKED ((struct bio *)1)
55 /* When we successfully write to a known bad-block, we need to remove the
56 * bad-block marking which must be done from process context. So we record
57 * the success by setting devs[n].bio to IO_MADE_GOOD
59 #define IO_MADE_GOOD ((struct bio *)2)
61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
63 /* When there are this many requests queue to be written by
64 * the raid1 thread, we become 'congested' to provide back-pressure
67 static int max_queued_requests = 1024;
69 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
71 static void lower_barrier(struct r1conf *conf);
73 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
75 struct pool_info *pi = data;
76 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
78 /* allocate a r1bio with room for raid_disks entries in the bios array */
79 return kzalloc(size, gfp_flags);
82 static void r1bio_pool_free(void *r1_bio, void *data)
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_DEPTH 32
89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
93 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
94 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
95 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
97 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
99 struct pool_info *pi = data;
100 struct r1bio *r1_bio;
105 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
110 * Allocate bios : 1 for reading, n-1 for writing
112 for (j = pi->raid_disks ; j-- ; ) {
113 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
116 r1_bio->bios[j] = bio;
119 * Allocate RESYNC_PAGES data pages and attach them to
121 * If this is a user-requested check/repair, allocate
122 * RESYNC_PAGES for each bio.
124 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
125 need_pages = pi->raid_disks;
128 for (j = 0; j < need_pages; j++) {
129 bio = r1_bio->bios[j];
130 bio->bi_vcnt = RESYNC_PAGES;
132 if (bio_alloc_pages(bio, gfp_flags))
135 /* If not user-requests, copy the page pointers to all bios */
136 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
137 for (i=0; i<RESYNC_PAGES ; i++)
138 for (j=1; j<pi->raid_disks; j++)
139 r1_bio->bios[j]->bi_io_vec[i].bv_page =
140 r1_bio->bios[0]->bi_io_vec[i].bv_page;
143 r1_bio->master_bio = NULL;
151 bio_for_each_segment_all(bv, r1_bio->bios[j], i)
152 __free_page(bv->bv_page);
156 while (++j < pi->raid_disks)
157 bio_put(r1_bio->bios[j]);
158 r1bio_pool_free(r1_bio, data);
162 static void r1buf_pool_free(void *__r1_bio, void *data)
164 struct pool_info *pi = data;
166 struct r1bio *r1bio = __r1_bio;
168 for (i = 0; i < RESYNC_PAGES; i++)
169 for (j = pi->raid_disks; j-- ;) {
171 r1bio->bios[j]->bi_io_vec[i].bv_page !=
172 r1bio->bios[0]->bi_io_vec[i].bv_page)
173 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
175 for (i=0 ; i < pi->raid_disks; i++)
176 bio_put(r1bio->bios[i]);
178 r1bio_pool_free(r1bio, data);
181 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
185 for (i = 0; i < conf->raid_disks * 2; i++) {
186 struct bio **bio = r1_bio->bios + i;
187 if (!BIO_SPECIAL(*bio))
193 static void free_r1bio(struct r1bio *r1_bio)
195 struct r1conf *conf = r1_bio->mddev->private;
197 put_all_bios(conf, r1_bio);
198 mempool_free(r1_bio, conf->r1bio_pool);
201 static void put_buf(struct r1bio *r1_bio)
203 struct r1conf *conf = r1_bio->mddev->private;
206 for (i = 0; i < conf->raid_disks * 2; i++) {
207 struct bio *bio = r1_bio->bios[i];
209 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
212 mempool_free(r1_bio, conf->r1buf_pool);
217 static void reschedule_retry(struct r1bio *r1_bio)
220 struct mddev *mddev = r1_bio->mddev;
221 struct r1conf *conf = mddev->private;
223 spin_lock_irqsave(&conf->device_lock, flags);
224 list_add(&r1_bio->retry_list, &conf->retry_list);
226 spin_unlock_irqrestore(&conf->device_lock, flags);
228 wake_up(&conf->wait_barrier);
229 md_wakeup_thread(mddev->thread);
233 * raid_end_bio_io() is called when we have finished servicing a mirrored
234 * operation and are ready to return a success/failure code to the buffer
237 static void call_bio_endio(struct r1bio *r1_bio)
239 struct bio *bio = r1_bio->master_bio;
241 struct r1conf *conf = r1_bio->mddev->private;
242 sector_t start_next_window = r1_bio->start_next_window;
243 sector_t bi_sector = bio->bi_iter.bi_sector;
245 if (bio->bi_phys_segments) {
247 spin_lock_irqsave(&conf->device_lock, flags);
248 bio->bi_phys_segments--;
249 done = (bio->bi_phys_segments == 0);
250 spin_unlock_irqrestore(&conf->device_lock, flags);
252 * make_request() might be waiting for
253 * bi_phys_segments to decrease
255 wake_up(&conf->wait_barrier);
259 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
260 bio->bi_error = -EIO;
265 * Wake up any possible resync thread that waits for the device
268 allow_barrier(conf, start_next_window, bi_sector);
272 static void raid_end_bio_io(struct r1bio *r1_bio)
274 struct bio *bio = r1_bio->master_bio;
276 /* if nobody has done the final endio yet, do it now */
277 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
278 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
279 (bio_data_dir(bio) == WRITE) ? "write" : "read",
280 (unsigned long long) bio->bi_iter.bi_sector,
281 (unsigned long long) bio_end_sector(bio) - 1);
283 call_bio_endio(r1_bio);
289 * Update disk head position estimator based on IRQ completion info.
291 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
293 struct r1conf *conf = r1_bio->mddev->private;
295 conf->mirrors[disk].head_position =
296 r1_bio->sector + (r1_bio->sectors);
300 * Find the disk number which triggered given bio
302 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
305 struct r1conf *conf = r1_bio->mddev->private;
306 int raid_disks = conf->raid_disks;
308 for (mirror = 0; mirror < raid_disks * 2; mirror++)
309 if (r1_bio->bios[mirror] == bio)
312 BUG_ON(mirror == raid_disks * 2);
313 update_head_pos(mirror, r1_bio);
318 static void raid1_end_read_request(struct bio *bio)
320 int uptodate = !bio->bi_error;
321 struct r1bio *r1_bio = bio->bi_private;
322 struct r1conf *conf = r1_bio->mddev->private;
323 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
326 * this branch is our 'one mirror IO has finished' event handler:
328 update_head_pos(r1_bio->read_disk, r1_bio);
331 set_bit(R1BIO_Uptodate, &r1_bio->state);
333 /* If all other devices have failed, we want to return
334 * the error upwards rather than fail the last device.
335 * Here we redefine "uptodate" to mean "Don't want to retry"
338 spin_lock_irqsave(&conf->device_lock, flags);
339 if (r1_bio->mddev->degraded == conf->raid_disks ||
340 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
341 test_bit(In_sync, &rdev->flags)))
343 spin_unlock_irqrestore(&conf->device_lock, flags);
347 raid_end_bio_io(r1_bio);
348 rdev_dec_pending(rdev, conf->mddev);
353 char b[BDEVNAME_SIZE];
355 KERN_ERR "md/raid1:%s: %s: "
356 "rescheduling sector %llu\n",
360 (unsigned long long)r1_bio->sector);
361 set_bit(R1BIO_ReadError, &r1_bio->state);
362 reschedule_retry(r1_bio);
363 /* don't drop the reference on read_disk yet */
367 static void close_write(struct r1bio *r1_bio)
369 /* it really is the end of this request */
370 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
371 /* free extra copy of the data pages */
372 int i = r1_bio->behind_page_count;
374 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
375 kfree(r1_bio->behind_bvecs);
376 r1_bio->behind_bvecs = NULL;
378 /* clear the bitmap if all writes complete successfully */
379 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
381 !test_bit(R1BIO_Degraded, &r1_bio->state),
382 test_bit(R1BIO_BehindIO, &r1_bio->state));
383 md_write_end(r1_bio->mddev);
386 static void r1_bio_write_done(struct r1bio *r1_bio)
388 if (!atomic_dec_and_test(&r1_bio->remaining))
391 if (test_bit(R1BIO_WriteError, &r1_bio->state))
392 reschedule_retry(r1_bio);
395 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
396 reschedule_retry(r1_bio);
398 raid_end_bio_io(r1_bio);
402 static void raid1_end_write_request(struct bio *bio)
404 struct r1bio *r1_bio = bio->bi_private;
405 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
406 struct r1conf *conf = r1_bio->mddev->private;
407 struct bio *to_put = NULL;
408 int mirror = find_bio_disk(r1_bio, bio);
409 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
412 * 'one mirror IO has finished' event handler:
415 set_bit(WriteErrorSeen, &rdev->flags);
416 if (!test_and_set_bit(WantReplacement, &rdev->flags))
417 set_bit(MD_RECOVERY_NEEDED, &
418 conf->mddev->recovery);
420 set_bit(R1BIO_WriteError, &r1_bio->state);
423 * Set R1BIO_Uptodate in our master bio, so that we
424 * will return a good error code for to the higher
425 * levels even if IO on some other mirrored buffer
428 * The 'master' represents the composite IO operation
429 * to user-side. So if something waits for IO, then it
430 * will wait for the 'master' bio.
435 r1_bio->bios[mirror] = NULL;
438 * Do not set R1BIO_Uptodate if the current device is
439 * rebuilding or Faulty. This is because we cannot use
440 * such device for properly reading the data back (we could
441 * potentially use it, if the current write would have felt
442 * before rdev->recovery_offset, but for simplicity we don't
445 if (test_bit(In_sync, &rdev->flags) &&
446 !test_bit(Faulty, &rdev->flags))
447 set_bit(R1BIO_Uptodate, &r1_bio->state);
449 /* Maybe we can clear some bad blocks. */
450 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
451 &first_bad, &bad_sectors)) {
452 r1_bio->bios[mirror] = IO_MADE_GOOD;
453 set_bit(R1BIO_MadeGood, &r1_bio->state);
458 if (test_bit(WriteMostly, &rdev->flags))
459 atomic_dec(&r1_bio->behind_remaining);
462 * In behind mode, we ACK the master bio once the I/O
463 * has safely reached all non-writemostly
464 * disks. Setting the Returned bit ensures that this
465 * gets done only once -- we don't ever want to return
466 * -EIO here, instead we'll wait
468 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
469 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
470 /* Maybe we can return now */
471 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
472 struct bio *mbio = r1_bio->master_bio;
473 pr_debug("raid1: behind end write sectors"
475 (unsigned long long) mbio->bi_iter.bi_sector,
476 (unsigned long long) bio_end_sector(mbio) - 1);
477 call_bio_endio(r1_bio);
481 if (r1_bio->bios[mirror] == NULL)
482 rdev_dec_pending(rdev, conf->mddev);
485 * Let's see if all mirrored write operations have finished
488 r1_bio_write_done(r1_bio);
495 * This routine returns the disk from which the requested read should
496 * be done. There is a per-array 'next expected sequential IO' sector
497 * number - if this matches on the next IO then we use the last disk.
498 * There is also a per-disk 'last know head position' sector that is
499 * maintained from IRQ contexts, both the normal and the resync IO
500 * completion handlers update this position correctly. If there is no
501 * perfect sequential match then we pick the disk whose head is closest.
503 * If there are 2 mirrors in the same 2 devices, performance degrades
504 * because position is mirror, not device based.
506 * The rdev for the device selected will have nr_pending incremented.
508 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
510 const sector_t this_sector = r1_bio->sector;
512 int best_good_sectors;
513 int best_disk, best_dist_disk, best_pending_disk;
517 unsigned int min_pending;
518 struct md_rdev *rdev;
520 int choose_next_idle;
524 * Check if we can balance. We can balance on the whole
525 * device if no resync is going on, or below the resync window.
526 * We take the first readable disk when above the resync window.
529 sectors = r1_bio->sectors;
532 best_dist = MaxSector;
533 best_pending_disk = -1;
534 min_pending = UINT_MAX;
535 best_good_sectors = 0;
537 choose_next_idle = 0;
539 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
540 (mddev_is_clustered(conf->mddev) &&
541 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
542 this_sector + sectors)))
547 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
551 unsigned int pending;
554 rdev = rcu_dereference(conf->mirrors[disk].rdev);
555 if (r1_bio->bios[disk] == IO_BLOCKED
557 || test_bit(Faulty, &rdev->flags))
559 if (!test_bit(In_sync, &rdev->flags) &&
560 rdev->recovery_offset < this_sector + sectors)
562 if (test_bit(WriteMostly, &rdev->flags)) {
563 /* Don't balance among write-mostly, just
564 * use the first as a last resort */
565 if (best_dist_disk < 0) {
566 if (is_badblock(rdev, this_sector, sectors,
567 &first_bad, &bad_sectors)) {
568 if (first_bad <= this_sector)
569 /* Cannot use this */
571 best_good_sectors = first_bad - this_sector;
573 best_good_sectors = sectors;
574 best_dist_disk = disk;
575 best_pending_disk = disk;
579 /* This is a reasonable device to use. It might
582 if (is_badblock(rdev, this_sector, sectors,
583 &first_bad, &bad_sectors)) {
584 if (best_dist < MaxSector)
585 /* already have a better device */
587 if (first_bad <= this_sector) {
588 /* cannot read here. If this is the 'primary'
589 * device, then we must not read beyond
590 * bad_sectors from another device..
592 bad_sectors -= (this_sector - first_bad);
593 if (choose_first && sectors > bad_sectors)
594 sectors = bad_sectors;
595 if (best_good_sectors > sectors)
596 best_good_sectors = sectors;
599 sector_t good_sectors = first_bad - this_sector;
600 if (good_sectors > best_good_sectors) {
601 best_good_sectors = good_sectors;
609 best_good_sectors = sectors;
611 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
612 has_nonrot_disk |= nonrot;
613 pending = atomic_read(&rdev->nr_pending);
614 dist = abs(this_sector - conf->mirrors[disk].head_position);
619 /* Don't change to another disk for sequential reads */
620 if (conf->mirrors[disk].next_seq_sect == this_sector
622 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
623 struct raid1_info *mirror = &conf->mirrors[disk];
627 * If buffered sequential IO size exceeds optimal
628 * iosize, check if there is idle disk. If yes, choose
629 * the idle disk. read_balance could already choose an
630 * idle disk before noticing it's a sequential IO in
631 * this disk. This doesn't matter because this disk
632 * will idle, next time it will be utilized after the
633 * first disk has IO size exceeds optimal iosize. In
634 * this way, iosize of the first disk will be optimal
635 * iosize at least. iosize of the second disk might be
636 * small, but not a big deal since when the second disk
637 * starts IO, the first disk is likely still busy.
639 if (nonrot && opt_iosize > 0 &&
640 mirror->seq_start != MaxSector &&
641 mirror->next_seq_sect > opt_iosize &&
642 mirror->next_seq_sect - opt_iosize >=
644 choose_next_idle = 1;
649 /* If device is idle, use it */
655 if (choose_next_idle)
658 if (min_pending > pending) {
659 min_pending = pending;
660 best_pending_disk = disk;
663 if (dist < best_dist) {
665 best_dist_disk = disk;
670 * If all disks are rotational, choose the closest disk. If any disk is
671 * non-rotational, choose the disk with less pending request even the
672 * disk is rotational, which might/might not be optimal for raids with
673 * mixed ratation/non-rotational disks depending on workload.
675 if (best_disk == -1) {
677 best_disk = best_pending_disk;
679 best_disk = best_dist_disk;
682 if (best_disk >= 0) {
683 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
686 atomic_inc(&rdev->nr_pending);
687 sectors = best_good_sectors;
689 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
690 conf->mirrors[best_disk].seq_start = this_sector;
692 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
695 *max_sectors = sectors;
700 static int raid1_congested(struct mddev *mddev, int bits)
702 struct r1conf *conf = mddev->private;
705 if ((bits & (1 << WB_async_congested)) &&
706 conf->pending_count >= max_queued_requests)
710 for (i = 0; i < conf->raid_disks * 2; i++) {
711 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
712 if (rdev && !test_bit(Faulty, &rdev->flags)) {
713 struct request_queue *q = bdev_get_queue(rdev->bdev);
717 /* Note the '|| 1' - when read_balance prefers
718 * non-congested targets, it can be removed
720 if ((bits & (1 << WB_async_congested)) || 1)
721 ret |= bdi_congested(&q->backing_dev_info, bits);
723 ret &= bdi_congested(&q->backing_dev_info, bits);
730 static void flush_pending_writes(struct r1conf *conf)
732 /* Any writes that have been queued but are awaiting
733 * bitmap updates get flushed here.
735 spin_lock_irq(&conf->device_lock);
737 if (conf->pending_bio_list.head) {
739 bio = bio_list_get(&conf->pending_bio_list);
740 conf->pending_count = 0;
741 spin_unlock_irq(&conf->device_lock);
742 /* flush any pending bitmap writes to
743 * disk before proceeding w/ I/O */
744 bitmap_unplug(conf->mddev->bitmap);
745 wake_up(&conf->wait_barrier);
747 while (bio) { /* submit pending writes */
748 struct bio *next = bio->bi_next;
750 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
751 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
755 generic_make_request(bio);
759 spin_unlock_irq(&conf->device_lock);
763 * Sometimes we need to suspend IO while we do something else,
764 * either some resync/recovery, or reconfigure the array.
765 * To do this we raise a 'barrier'.
766 * The 'barrier' is a counter that can be raised multiple times
767 * to count how many activities are happening which preclude
769 * We can only raise the barrier if there is no pending IO.
770 * i.e. if nr_pending == 0.
771 * We choose only to raise the barrier if no-one is waiting for the
772 * barrier to go down. This means that as soon as an IO request
773 * is ready, no other operations which require a barrier will start
774 * until the IO request has had a chance.
776 * So: regular IO calls 'wait_barrier'. When that returns there
777 * is no backgroup IO happening, It must arrange to call
778 * allow_barrier when it has finished its IO.
779 * backgroup IO calls must call raise_barrier. Once that returns
780 * there is no normal IO happeing. It must arrange to call
781 * lower_barrier when the particular background IO completes.
783 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
785 spin_lock_irq(&conf->resync_lock);
787 /* Wait until no block IO is waiting */
788 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
791 /* block any new IO from starting */
793 conf->next_resync = sector_nr;
795 /* For these conditions we must wait:
796 * A: while the array is in frozen state
797 * B: while barrier >= RESYNC_DEPTH, meaning resync reach
798 * the max count which allowed.
799 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
800 * next resync will reach to the window which normal bios are
802 * D: while there are any active requests in the current window.
804 wait_event_lock_irq(conf->wait_barrier,
805 !conf->array_frozen &&
806 conf->barrier < RESYNC_DEPTH &&
807 conf->current_window_requests == 0 &&
808 (conf->start_next_window >=
809 conf->next_resync + RESYNC_SECTORS),
813 spin_unlock_irq(&conf->resync_lock);
816 static void lower_barrier(struct r1conf *conf)
819 BUG_ON(conf->barrier <= 0);
820 spin_lock_irqsave(&conf->resync_lock, flags);
823 spin_unlock_irqrestore(&conf->resync_lock, flags);
824 wake_up(&conf->wait_barrier);
827 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
831 if (conf->array_frozen || !bio)
833 else if (conf->barrier && bio_data_dir(bio) == WRITE) {
834 if ((conf->mddev->curr_resync_completed
835 >= bio_end_sector(bio)) ||
836 (conf->next_resync + NEXT_NORMALIO_DISTANCE
837 <= bio->bi_iter.bi_sector))
846 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
850 spin_lock_irq(&conf->resync_lock);
851 if (need_to_wait_for_sync(conf, bio)) {
853 /* Wait for the barrier to drop.
854 * However if there are already pending
855 * requests (preventing the barrier from
856 * rising completely), and the
857 * per-process bio queue isn't empty,
858 * then don't wait, as we need to empty
859 * that queue to allow conf->start_next_window
862 wait_event_lock_irq(conf->wait_barrier,
863 !conf->array_frozen &&
865 ((conf->start_next_window <
866 conf->next_resync + RESYNC_SECTORS) &&
868 !bio_list_empty(current->bio_list))),
873 if (bio && bio_data_dir(bio) == WRITE) {
874 if (bio->bi_iter.bi_sector >= conf->next_resync) {
875 if (conf->start_next_window == MaxSector)
876 conf->start_next_window =
878 NEXT_NORMALIO_DISTANCE;
880 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
881 <= bio->bi_iter.bi_sector)
882 conf->next_window_requests++;
884 conf->current_window_requests++;
885 sector = conf->start_next_window;
890 spin_unlock_irq(&conf->resync_lock);
894 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
899 spin_lock_irqsave(&conf->resync_lock, flags);
901 if (start_next_window) {
902 if (start_next_window == conf->start_next_window) {
903 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
905 conf->next_window_requests--;
907 conf->current_window_requests--;
909 conf->current_window_requests--;
911 if (!conf->current_window_requests) {
912 if (conf->next_window_requests) {
913 conf->current_window_requests =
914 conf->next_window_requests;
915 conf->next_window_requests = 0;
916 conf->start_next_window +=
917 NEXT_NORMALIO_DISTANCE;
919 conf->start_next_window = MaxSector;
922 spin_unlock_irqrestore(&conf->resync_lock, flags);
923 wake_up(&conf->wait_barrier);
926 static void freeze_array(struct r1conf *conf, int extra)
928 /* stop syncio and normal IO and wait for everything to
930 * We wait until nr_pending match nr_queued+extra
931 * This is called in the context of one normal IO request
932 * that has failed. Thus any sync request that might be pending
933 * will be blocked by nr_pending, and we need to wait for
934 * pending IO requests to complete or be queued for re-try.
935 * Thus the number queued (nr_queued) plus this request (extra)
936 * must match the number of pending IOs (nr_pending) before
939 spin_lock_irq(&conf->resync_lock);
940 conf->array_frozen = 1;
941 wait_event_lock_irq_cmd(conf->wait_barrier,
942 conf->nr_pending == conf->nr_queued+extra,
944 flush_pending_writes(conf));
945 spin_unlock_irq(&conf->resync_lock);
947 static void unfreeze_array(struct r1conf *conf)
949 /* reverse the effect of the freeze */
950 spin_lock_irq(&conf->resync_lock);
951 conf->array_frozen = 0;
952 wake_up(&conf->wait_barrier);
953 spin_unlock_irq(&conf->resync_lock);
956 /* duplicate the data pages for behind I/O
958 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
961 struct bio_vec *bvec;
962 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
964 if (unlikely(!bvecs))
967 bio_for_each_segment_all(bvec, bio, i) {
969 bvecs[i].bv_page = alloc_page(GFP_NOIO);
970 if (unlikely(!bvecs[i].bv_page))
972 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
973 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
974 kunmap(bvecs[i].bv_page);
975 kunmap(bvec->bv_page);
977 r1_bio->behind_bvecs = bvecs;
978 r1_bio->behind_page_count = bio->bi_vcnt;
979 set_bit(R1BIO_BehindIO, &r1_bio->state);
983 for (i = 0; i < bio->bi_vcnt; i++)
984 if (bvecs[i].bv_page)
985 put_page(bvecs[i].bv_page);
987 pr_debug("%dB behind alloc failed, doing sync I/O\n",
988 bio->bi_iter.bi_size);
991 struct raid1_plug_cb {
992 struct blk_plug_cb cb;
993 struct bio_list pending;
997 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
999 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1001 struct mddev *mddev = plug->cb.data;
1002 struct r1conf *conf = mddev->private;
1005 if (from_schedule || current->bio_list) {
1006 spin_lock_irq(&conf->device_lock);
1007 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1008 conf->pending_count += plug->pending_cnt;
1009 spin_unlock_irq(&conf->device_lock);
1010 wake_up(&conf->wait_barrier);
1011 md_wakeup_thread(mddev->thread);
1016 /* we aren't scheduling, so we can do the write-out directly. */
1017 bio = bio_list_get(&plug->pending);
1018 bitmap_unplug(mddev->bitmap);
1019 wake_up(&conf->wait_barrier);
1021 while (bio) { /* submit pending writes */
1022 struct bio *next = bio->bi_next;
1023 bio->bi_next = NULL;
1024 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1025 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1026 /* Just ignore it */
1029 generic_make_request(bio);
1035 static void raid1_make_request(struct mddev *mddev, struct bio * bio)
1037 struct r1conf *conf = mddev->private;
1038 struct raid1_info *mirror;
1039 struct r1bio *r1_bio;
1040 struct bio *read_bio;
1042 struct bitmap *bitmap;
1043 unsigned long flags;
1044 const int rw = bio_data_dir(bio);
1045 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1046 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1047 const unsigned long do_discard = (bio->bi_rw
1048 & (REQ_DISCARD | REQ_SECURE));
1049 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1050 struct md_rdev *blocked_rdev;
1051 struct blk_plug_cb *cb;
1052 struct raid1_plug_cb *plug = NULL;
1054 int sectors_handled;
1056 sector_t start_next_window;
1059 * Register the new request and wait if the reconstruction
1060 * thread has put up a bar for new requests.
1061 * Continue immediately if no resync is active currently.
1064 md_write_start(mddev, bio); /* wait on superblock update early */
1066 if (bio_data_dir(bio) == WRITE &&
1067 ((bio_end_sector(bio) > mddev->suspend_lo &&
1068 bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1069 (mddev_is_clustered(mddev) &&
1070 md_cluster_ops->area_resyncing(mddev, WRITE,
1071 bio->bi_iter.bi_sector, bio_end_sector(bio))))) {
1072 /* As the suspend_* range is controlled by
1073 * userspace, we want an interruptible
1078 flush_signals(current);
1079 prepare_to_wait(&conf->wait_barrier,
1080 &w, TASK_INTERRUPTIBLE);
1081 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1082 bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1083 (mddev_is_clustered(mddev) &&
1084 !md_cluster_ops->area_resyncing(mddev, WRITE,
1085 bio->bi_iter.bi_sector, bio_end_sector(bio))))
1089 finish_wait(&conf->wait_barrier, &w);
1092 start_next_window = wait_barrier(conf, bio);
1094 bitmap = mddev->bitmap;
1097 * make_request() can abort the operation when READA is being
1098 * used and no empty request is available.
1101 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1103 r1_bio->master_bio = bio;
1104 r1_bio->sectors = bio_sectors(bio);
1106 r1_bio->mddev = mddev;
1107 r1_bio->sector = bio->bi_iter.bi_sector;
1109 /* We might need to issue multiple reads to different
1110 * devices if there are bad blocks around, so we keep
1111 * track of the number of reads in bio->bi_phys_segments.
1112 * If this is 0, there is only one r1_bio and no locking
1113 * will be needed when requests complete. If it is
1114 * non-zero, then it is the number of not-completed requests.
1116 bio->bi_phys_segments = 0;
1117 bio_clear_flag(bio, BIO_SEG_VALID);
1121 * read balancing logic:
1126 rdisk = read_balance(conf, r1_bio, &max_sectors);
1129 /* couldn't find anywhere to read from */
1130 raid_end_bio_io(r1_bio);
1133 mirror = conf->mirrors + rdisk;
1135 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1137 /* Reading from a write-mostly device must
1138 * take care not to over-take any writes
1141 wait_event(bitmap->behind_wait,
1142 atomic_read(&bitmap->behind_writes) == 0);
1144 r1_bio->read_disk = rdisk;
1145 r1_bio->start_next_window = 0;
1147 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1148 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1151 r1_bio->bios[rdisk] = read_bio;
1153 read_bio->bi_iter.bi_sector = r1_bio->sector +
1154 mirror->rdev->data_offset;
1155 read_bio->bi_bdev = mirror->rdev->bdev;
1156 read_bio->bi_end_io = raid1_end_read_request;
1157 read_bio->bi_rw = READ | do_sync;
1158 read_bio->bi_private = r1_bio;
1160 if (max_sectors < r1_bio->sectors) {
1161 /* could not read all from this device, so we will
1162 * need another r1_bio.
1165 sectors_handled = (r1_bio->sector + max_sectors
1166 - bio->bi_iter.bi_sector);
1167 r1_bio->sectors = max_sectors;
1168 spin_lock_irq(&conf->device_lock);
1169 if (bio->bi_phys_segments == 0)
1170 bio->bi_phys_segments = 2;
1172 bio->bi_phys_segments++;
1173 spin_unlock_irq(&conf->device_lock);
1174 /* Cannot call generic_make_request directly
1175 * as that will be queued in __make_request
1176 * and subsequent mempool_alloc might block waiting
1177 * for it. So hand bio over to raid1d.
1179 reschedule_retry(r1_bio);
1181 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1183 r1_bio->master_bio = bio;
1184 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1186 r1_bio->mddev = mddev;
1187 r1_bio->sector = bio->bi_iter.bi_sector +
1191 generic_make_request(read_bio);
1198 if (conf->pending_count >= max_queued_requests) {
1199 md_wakeup_thread(mddev->thread);
1200 wait_event(conf->wait_barrier,
1201 conf->pending_count < max_queued_requests);
1203 /* first select target devices under rcu_lock and
1204 * inc refcount on their rdev. Record them by setting
1206 * If there are known/acknowledged bad blocks on any device on
1207 * which we have seen a write error, we want to avoid writing those
1209 * This potentially requires several writes to write around
1210 * the bad blocks. Each set of writes gets it's own r1bio
1211 * with a set of bios attached.
1214 disks = conf->raid_disks * 2;
1216 r1_bio->start_next_window = start_next_window;
1217 blocked_rdev = NULL;
1219 max_sectors = r1_bio->sectors;
1220 for (i = 0; i < disks; i++) {
1221 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1222 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1223 atomic_inc(&rdev->nr_pending);
1224 blocked_rdev = rdev;
1227 r1_bio->bios[i] = NULL;
1228 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1229 if (i < conf->raid_disks)
1230 set_bit(R1BIO_Degraded, &r1_bio->state);
1234 atomic_inc(&rdev->nr_pending);
1235 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1240 is_bad = is_badblock(rdev, r1_bio->sector,
1242 &first_bad, &bad_sectors);
1244 /* mustn't write here until the bad block is
1246 set_bit(BlockedBadBlocks, &rdev->flags);
1247 blocked_rdev = rdev;
1250 if (is_bad && first_bad <= r1_bio->sector) {
1251 /* Cannot write here at all */
1252 bad_sectors -= (r1_bio->sector - first_bad);
1253 if (bad_sectors < max_sectors)
1254 /* mustn't write more than bad_sectors
1255 * to other devices yet
1257 max_sectors = bad_sectors;
1258 rdev_dec_pending(rdev, mddev);
1259 /* We don't set R1BIO_Degraded as that
1260 * only applies if the disk is
1261 * missing, so it might be re-added,
1262 * and we want to know to recover this
1264 * In this case the device is here,
1265 * and the fact that this chunk is not
1266 * in-sync is recorded in the bad
1272 int good_sectors = first_bad - r1_bio->sector;
1273 if (good_sectors < max_sectors)
1274 max_sectors = good_sectors;
1277 r1_bio->bios[i] = bio;
1281 if (unlikely(blocked_rdev)) {
1282 /* Wait for this device to become unblocked */
1284 sector_t old = start_next_window;
1286 for (j = 0; j < i; j++)
1287 if (r1_bio->bios[j])
1288 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1290 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1291 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1292 start_next_window = wait_barrier(conf, bio);
1294 * We must make sure the multi r1bios of bio have
1295 * the same value of bi_phys_segments
1297 if (bio->bi_phys_segments && old &&
1298 old != start_next_window)
1299 /* Wait for the former r1bio(s) to complete */
1300 wait_event(conf->wait_barrier,
1301 bio->bi_phys_segments == 1);
1305 if (max_sectors < r1_bio->sectors) {
1306 /* We are splitting this write into multiple parts, so
1307 * we need to prepare for allocating another r1_bio.
1309 r1_bio->sectors = max_sectors;
1310 spin_lock_irq(&conf->device_lock);
1311 if (bio->bi_phys_segments == 0)
1312 bio->bi_phys_segments = 2;
1314 bio->bi_phys_segments++;
1315 spin_unlock_irq(&conf->device_lock);
1317 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1319 atomic_set(&r1_bio->remaining, 1);
1320 atomic_set(&r1_bio->behind_remaining, 0);
1323 for (i = 0; i < disks; i++) {
1325 if (!r1_bio->bios[i])
1328 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1329 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1333 * Not if there are too many, or cannot
1334 * allocate memory, or a reader on WriteMostly
1335 * is waiting for behind writes to flush */
1337 (atomic_read(&bitmap->behind_writes)
1338 < mddev->bitmap_info.max_write_behind) &&
1339 !waitqueue_active(&bitmap->behind_wait))
1340 alloc_behind_pages(mbio, r1_bio);
1342 bitmap_startwrite(bitmap, r1_bio->sector,
1344 test_bit(R1BIO_BehindIO,
1348 if (r1_bio->behind_bvecs) {
1349 struct bio_vec *bvec;
1353 * We trimmed the bio, so _all is legit
1355 bio_for_each_segment_all(bvec, mbio, j)
1356 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1357 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1358 atomic_inc(&r1_bio->behind_remaining);
1361 r1_bio->bios[i] = mbio;
1363 mbio->bi_iter.bi_sector = (r1_bio->sector +
1364 conf->mirrors[i].rdev->data_offset);
1365 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1366 mbio->bi_end_io = raid1_end_write_request;
1368 WRITE | do_flush_fua | do_sync | do_discard | do_same;
1369 mbio->bi_private = r1_bio;
1371 atomic_inc(&r1_bio->remaining);
1373 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1375 plug = container_of(cb, struct raid1_plug_cb, cb);
1378 spin_lock_irqsave(&conf->device_lock, flags);
1380 bio_list_add(&plug->pending, mbio);
1381 plug->pending_cnt++;
1383 bio_list_add(&conf->pending_bio_list, mbio);
1384 conf->pending_count++;
1386 spin_unlock_irqrestore(&conf->device_lock, flags);
1388 md_wakeup_thread(mddev->thread);
1390 /* Mustn't call r1_bio_write_done before this next test,
1391 * as it could result in the bio being freed.
1393 if (sectors_handled < bio_sectors(bio)) {
1394 r1_bio_write_done(r1_bio);
1395 /* We need another r1_bio. It has already been counted
1396 * in bio->bi_phys_segments
1398 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1399 r1_bio->master_bio = bio;
1400 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1402 r1_bio->mddev = mddev;
1403 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1407 r1_bio_write_done(r1_bio);
1409 /* In case raid1d snuck in to freeze_array */
1410 wake_up(&conf->wait_barrier);
1413 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1415 struct r1conf *conf = mddev->private;
1418 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1419 conf->raid_disks - mddev->degraded);
1421 for (i = 0; i < conf->raid_disks; i++) {
1422 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1423 seq_printf(seq, "%s",
1424 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1427 seq_printf(seq, "]");
1430 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1432 char b[BDEVNAME_SIZE];
1433 struct r1conf *conf = mddev->private;
1434 unsigned long flags;
1437 * If it is not operational, then we have already marked it as dead
1438 * else if it is the last working disks, ignore the error, let the
1439 * next level up know.
1440 * else mark the drive as failed
1442 if (test_bit(In_sync, &rdev->flags)
1443 && (conf->raid_disks - mddev->degraded) == 1) {
1445 * Don't fail the drive, act as though we were just a
1446 * normal single drive.
1447 * However don't try a recovery from this drive as
1448 * it is very likely to fail.
1450 conf->recovery_disabled = mddev->recovery_disabled;
1453 set_bit(Blocked, &rdev->flags);
1454 spin_lock_irqsave(&conf->device_lock, flags);
1455 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1457 set_bit(Faulty, &rdev->flags);
1459 set_bit(Faulty, &rdev->flags);
1460 spin_unlock_irqrestore(&conf->device_lock, flags);
1462 * if recovery is running, make sure it aborts.
1464 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1465 set_mask_bits(&mddev->flags, 0,
1466 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1468 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1469 "md/raid1:%s: Operation continuing on %d devices.\n",
1470 mdname(mddev), bdevname(rdev->bdev, b),
1471 mdname(mddev), conf->raid_disks - mddev->degraded);
1474 static void print_conf(struct r1conf *conf)
1478 printk(KERN_DEBUG "RAID1 conf printout:\n");
1480 printk(KERN_DEBUG "(!conf)\n");
1483 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1487 for (i = 0; i < conf->raid_disks; i++) {
1488 char b[BDEVNAME_SIZE];
1489 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1491 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1492 i, !test_bit(In_sync, &rdev->flags),
1493 !test_bit(Faulty, &rdev->flags),
1494 bdevname(rdev->bdev,b));
1499 static void close_sync(struct r1conf *conf)
1501 wait_barrier(conf, NULL);
1502 allow_barrier(conf, 0, 0);
1504 mempool_destroy(conf->r1buf_pool);
1505 conf->r1buf_pool = NULL;
1507 spin_lock_irq(&conf->resync_lock);
1508 conf->next_resync = MaxSector - 2 * NEXT_NORMALIO_DISTANCE;
1509 conf->start_next_window = MaxSector;
1510 conf->current_window_requests +=
1511 conf->next_window_requests;
1512 conf->next_window_requests = 0;
1513 spin_unlock_irq(&conf->resync_lock);
1516 static int raid1_spare_active(struct mddev *mddev)
1519 struct r1conf *conf = mddev->private;
1521 unsigned long flags;
1524 * Find all failed disks within the RAID1 configuration
1525 * and mark them readable.
1526 * Called under mddev lock, so rcu protection not needed.
1527 * device_lock used to avoid races with raid1_end_read_request
1528 * which expects 'In_sync' flags and ->degraded to be consistent.
1530 spin_lock_irqsave(&conf->device_lock, flags);
1531 for (i = 0; i < conf->raid_disks; i++) {
1532 struct md_rdev *rdev = conf->mirrors[i].rdev;
1533 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1535 && !test_bit(Candidate, &repl->flags)
1536 && repl->recovery_offset == MaxSector
1537 && !test_bit(Faulty, &repl->flags)
1538 && !test_and_set_bit(In_sync, &repl->flags)) {
1539 /* replacement has just become active */
1541 !test_and_clear_bit(In_sync, &rdev->flags))
1544 /* Replaced device not technically
1545 * faulty, but we need to be sure
1546 * it gets removed and never re-added
1548 set_bit(Faulty, &rdev->flags);
1549 sysfs_notify_dirent_safe(
1554 && rdev->recovery_offset == MaxSector
1555 && !test_bit(Faulty, &rdev->flags)
1556 && !test_and_set_bit(In_sync, &rdev->flags)) {
1558 sysfs_notify_dirent_safe(rdev->sysfs_state);
1561 mddev->degraded -= count;
1562 spin_unlock_irqrestore(&conf->device_lock, flags);
1568 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1570 struct r1conf *conf = mddev->private;
1573 struct raid1_info *p;
1575 int last = conf->raid_disks - 1;
1577 if (mddev->recovery_disabled == conf->recovery_disabled)
1580 if (md_integrity_add_rdev(rdev, mddev))
1583 if (rdev->raid_disk >= 0)
1584 first = last = rdev->raid_disk;
1587 * find the disk ... but prefer rdev->saved_raid_disk
1590 if (rdev->saved_raid_disk >= 0 &&
1591 rdev->saved_raid_disk >= first &&
1592 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1593 first = last = rdev->saved_raid_disk;
1595 for (mirror = first; mirror <= last; mirror++) {
1596 p = conf->mirrors+mirror;
1600 disk_stack_limits(mddev->gendisk, rdev->bdev,
1601 rdev->data_offset << 9);
1603 p->head_position = 0;
1604 rdev->raid_disk = mirror;
1606 /* As all devices are equivalent, we don't need a full recovery
1607 * if this was recently any drive of the array
1609 if (rdev->saved_raid_disk < 0)
1611 rcu_assign_pointer(p->rdev, rdev);
1614 if (test_bit(WantReplacement, &p->rdev->flags) &&
1615 p[conf->raid_disks].rdev == NULL) {
1616 /* Add this device as a replacement */
1617 clear_bit(In_sync, &rdev->flags);
1618 set_bit(Replacement, &rdev->flags);
1619 rdev->raid_disk = mirror;
1622 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1626 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1627 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1632 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1634 struct r1conf *conf = mddev->private;
1636 int number = rdev->raid_disk;
1637 struct raid1_info *p = conf->mirrors + number;
1639 if (rdev != p->rdev)
1640 p = conf->mirrors + conf->raid_disks + number;
1643 if (rdev == p->rdev) {
1644 if (test_bit(In_sync, &rdev->flags) ||
1645 atomic_read(&rdev->nr_pending)) {
1649 /* Only remove non-faulty devices if recovery
1652 if (!test_bit(Faulty, &rdev->flags) &&
1653 mddev->recovery_disabled != conf->recovery_disabled &&
1654 mddev->degraded < conf->raid_disks) {
1659 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1661 if (atomic_read(&rdev->nr_pending)) {
1662 /* lost the race, try later */
1668 if (conf->mirrors[conf->raid_disks + number].rdev) {
1669 /* We just removed a device that is being replaced.
1670 * Move down the replacement. We drain all IO before
1671 * doing this to avoid confusion.
1673 struct md_rdev *repl =
1674 conf->mirrors[conf->raid_disks + number].rdev;
1675 freeze_array(conf, 0);
1676 clear_bit(Replacement, &repl->flags);
1678 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1679 unfreeze_array(conf);
1680 clear_bit(WantReplacement, &rdev->flags);
1682 clear_bit(WantReplacement, &rdev->flags);
1683 err = md_integrity_register(mddev);
1691 static void end_sync_read(struct bio *bio)
1693 struct r1bio *r1_bio = bio->bi_private;
1695 update_head_pos(r1_bio->read_disk, r1_bio);
1698 * we have read a block, now it needs to be re-written,
1699 * or re-read if the read failed.
1700 * We don't do much here, just schedule handling by raid1d
1703 set_bit(R1BIO_Uptodate, &r1_bio->state);
1705 if (atomic_dec_and_test(&r1_bio->remaining))
1706 reschedule_retry(r1_bio);
1709 static void end_sync_write(struct bio *bio)
1711 int uptodate = !bio->bi_error;
1712 struct r1bio *r1_bio = bio->bi_private;
1713 struct mddev *mddev = r1_bio->mddev;
1714 struct r1conf *conf = mddev->private;
1717 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1720 sector_t sync_blocks = 0;
1721 sector_t s = r1_bio->sector;
1722 long sectors_to_go = r1_bio->sectors;
1723 /* make sure these bits doesn't get cleared. */
1725 bitmap_end_sync(mddev->bitmap, s,
1728 sectors_to_go -= sync_blocks;
1729 } while (sectors_to_go > 0);
1730 set_bit(WriteErrorSeen, &rdev->flags);
1731 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1732 set_bit(MD_RECOVERY_NEEDED, &
1734 set_bit(R1BIO_WriteError, &r1_bio->state);
1735 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1736 &first_bad, &bad_sectors) &&
1737 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1740 &first_bad, &bad_sectors)
1742 set_bit(R1BIO_MadeGood, &r1_bio->state);
1744 if (atomic_dec_and_test(&r1_bio->remaining)) {
1745 int s = r1_bio->sectors;
1746 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1747 test_bit(R1BIO_WriteError, &r1_bio->state))
1748 reschedule_retry(r1_bio);
1751 md_done_sync(mddev, s, uptodate);
1756 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1757 int sectors, struct page *page, int rw)
1759 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1763 set_bit(WriteErrorSeen, &rdev->flags);
1764 if (!test_and_set_bit(WantReplacement,
1766 set_bit(MD_RECOVERY_NEEDED, &
1767 rdev->mddev->recovery);
1769 /* need to record an error - either for the block or the device */
1770 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1771 md_error(rdev->mddev, rdev);
1775 static int fix_sync_read_error(struct r1bio *r1_bio)
1777 /* Try some synchronous reads of other devices to get
1778 * good data, much like with normal read errors. Only
1779 * read into the pages we already have so we don't
1780 * need to re-issue the read request.
1781 * We don't need to freeze the array, because being in an
1782 * active sync request, there is no normal IO, and
1783 * no overlapping syncs.
1784 * We don't need to check is_badblock() again as we
1785 * made sure that anything with a bad block in range
1786 * will have bi_end_io clear.
1788 struct mddev *mddev = r1_bio->mddev;
1789 struct r1conf *conf = mddev->private;
1790 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1791 sector_t sect = r1_bio->sector;
1792 int sectors = r1_bio->sectors;
1797 int d = r1_bio->read_disk;
1799 struct md_rdev *rdev;
1802 if (s > (PAGE_SIZE>>9))
1805 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1806 /* No rcu protection needed here devices
1807 * can only be removed when no resync is
1808 * active, and resync is currently active
1810 rdev = conf->mirrors[d].rdev;
1811 if (sync_page_io(rdev, sect, s<<9,
1812 bio->bi_io_vec[idx].bv_page,
1819 if (d == conf->raid_disks * 2)
1821 } while (!success && d != r1_bio->read_disk);
1824 char b[BDEVNAME_SIZE];
1826 /* Cannot read from anywhere, this block is lost.
1827 * Record a bad block on each device. If that doesn't
1828 * work just disable and interrupt the recovery.
1829 * Don't fail devices as that won't really help.
1831 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1832 " for block %llu\n",
1834 bdevname(bio->bi_bdev, b),
1835 (unsigned long long)r1_bio->sector);
1836 for (d = 0; d < conf->raid_disks * 2; d++) {
1837 rdev = conf->mirrors[d].rdev;
1838 if (!rdev || test_bit(Faulty, &rdev->flags))
1840 if (!rdev_set_badblocks(rdev, sect, s, 0))
1844 conf->recovery_disabled =
1845 mddev->recovery_disabled;
1846 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1847 md_done_sync(mddev, r1_bio->sectors, 0);
1859 /* write it back and re-read */
1860 while (d != r1_bio->read_disk) {
1862 d = conf->raid_disks * 2;
1864 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1866 rdev = conf->mirrors[d].rdev;
1867 if (r1_sync_page_io(rdev, sect, s,
1868 bio->bi_io_vec[idx].bv_page,
1870 r1_bio->bios[d]->bi_end_io = NULL;
1871 rdev_dec_pending(rdev, mddev);
1875 while (d != r1_bio->read_disk) {
1877 d = conf->raid_disks * 2;
1879 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1881 rdev = conf->mirrors[d].rdev;
1882 if (r1_sync_page_io(rdev, sect, s,
1883 bio->bi_io_vec[idx].bv_page,
1885 atomic_add(s, &rdev->corrected_errors);
1891 set_bit(R1BIO_Uptodate, &r1_bio->state);
1896 static void process_checks(struct r1bio *r1_bio)
1898 /* We have read all readable devices. If we haven't
1899 * got the block, then there is no hope left.
1900 * If we have, then we want to do a comparison
1901 * and skip the write if everything is the same.
1902 * If any blocks failed to read, then we need to
1903 * attempt an over-write
1905 struct mddev *mddev = r1_bio->mddev;
1906 struct r1conf *conf = mddev->private;
1911 /* Fix variable parts of all bios */
1912 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1913 for (i = 0; i < conf->raid_disks * 2; i++) {
1917 struct bio *b = r1_bio->bios[i];
1918 if (b->bi_end_io != end_sync_read)
1920 /* fixup the bio for reuse, but preserve errno */
1921 error = b->bi_error;
1923 b->bi_error = error;
1925 b->bi_iter.bi_size = r1_bio->sectors << 9;
1926 b->bi_iter.bi_sector = r1_bio->sector +
1927 conf->mirrors[i].rdev->data_offset;
1928 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1929 b->bi_end_io = end_sync_read;
1930 b->bi_private = r1_bio;
1932 size = b->bi_iter.bi_size;
1933 for (j = 0; j < vcnt ; j++) {
1935 bi = &b->bi_io_vec[j];
1937 if (size > PAGE_SIZE)
1938 bi->bv_len = PAGE_SIZE;
1944 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1945 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1946 !r1_bio->bios[primary]->bi_error) {
1947 r1_bio->bios[primary]->bi_end_io = NULL;
1948 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1951 r1_bio->read_disk = primary;
1952 for (i = 0; i < conf->raid_disks * 2; i++) {
1954 struct bio *pbio = r1_bio->bios[primary];
1955 struct bio *sbio = r1_bio->bios[i];
1956 int error = sbio->bi_error;
1958 if (sbio->bi_end_io != end_sync_read)
1960 /* Now we can 'fixup' the error value */
1964 for (j = vcnt; j-- ; ) {
1966 p = pbio->bi_io_vec[j].bv_page;
1967 s = sbio->bi_io_vec[j].bv_page;
1968 if (memcmp(page_address(p),
1970 sbio->bi_io_vec[j].bv_len))
1976 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1977 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1979 /* No need to write to this device. */
1980 sbio->bi_end_io = NULL;
1981 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1985 bio_copy_data(sbio, pbio);
1989 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1991 struct r1conf *conf = mddev->private;
1993 int disks = conf->raid_disks * 2;
1994 struct bio *bio, *wbio;
1996 bio = r1_bio->bios[r1_bio->read_disk];
1998 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1999 /* ouch - failed to read all of that. */
2000 if (!fix_sync_read_error(r1_bio))
2003 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2004 process_checks(r1_bio);
2009 atomic_set(&r1_bio->remaining, 1);
2010 for (i = 0; i < disks ; i++) {
2011 wbio = r1_bio->bios[i];
2012 if (wbio->bi_end_io == NULL ||
2013 (wbio->bi_end_io == end_sync_read &&
2014 (i == r1_bio->read_disk ||
2015 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2018 wbio->bi_rw = WRITE;
2019 wbio->bi_end_io = end_sync_write;
2020 atomic_inc(&r1_bio->remaining);
2021 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2023 generic_make_request(wbio);
2026 if (atomic_dec_and_test(&r1_bio->remaining)) {
2027 /* if we're here, all write(s) have completed, so clean up */
2028 int s = r1_bio->sectors;
2029 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2030 test_bit(R1BIO_WriteError, &r1_bio->state))
2031 reschedule_retry(r1_bio);
2034 md_done_sync(mddev, s, 1);
2040 * This is a kernel thread which:
2042 * 1. Retries failed read operations on working mirrors.
2043 * 2. Updates the raid superblock when problems encounter.
2044 * 3. Performs writes following reads for array synchronising.
2047 static void fix_read_error(struct r1conf *conf, int read_disk,
2048 sector_t sect, int sectors)
2050 struct mddev *mddev = conf->mddev;
2056 struct md_rdev *rdev;
2058 if (s > (PAGE_SIZE>>9))
2066 rdev = rcu_dereference(conf->mirrors[d].rdev);
2068 (test_bit(In_sync, &rdev->flags) ||
2069 (!test_bit(Faulty, &rdev->flags) &&
2070 rdev->recovery_offset >= sect + s)) &&
2071 is_badblock(rdev, sect, s,
2072 &first_bad, &bad_sectors) == 0) {
2073 atomic_inc(&rdev->nr_pending);
2075 if (sync_page_io(rdev, sect, s<<9,
2076 conf->tmppage, READ, false))
2078 rdev_dec_pending(rdev, mddev);
2084 if (d == conf->raid_disks * 2)
2086 } while (!success && d != read_disk);
2089 /* Cannot read from anywhere - mark it bad */
2090 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2091 if (!rdev_set_badblocks(rdev, sect, s, 0))
2092 md_error(mddev, rdev);
2095 /* write it back and re-read */
2097 while (d != read_disk) {
2099 d = conf->raid_disks * 2;
2102 rdev = rcu_dereference(conf->mirrors[d].rdev);
2104 !test_bit(Faulty, &rdev->flags)) {
2105 atomic_inc(&rdev->nr_pending);
2107 r1_sync_page_io(rdev, sect, s,
2108 conf->tmppage, WRITE);
2109 rdev_dec_pending(rdev, mddev);
2114 while (d != read_disk) {
2115 char b[BDEVNAME_SIZE];
2117 d = conf->raid_disks * 2;
2120 rdev = rcu_dereference(conf->mirrors[d].rdev);
2122 !test_bit(Faulty, &rdev->flags)) {
2123 atomic_inc(&rdev->nr_pending);
2125 if (r1_sync_page_io(rdev, sect, s,
2126 conf->tmppage, READ)) {
2127 atomic_add(s, &rdev->corrected_errors);
2129 "md/raid1:%s: read error corrected "
2130 "(%d sectors at %llu on %s)\n",
2132 (unsigned long long)(sect +
2134 bdevname(rdev->bdev, b));
2136 rdev_dec_pending(rdev, mddev);
2145 static int narrow_write_error(struct r1bio *r1_bio, int i)
2147 struct mddev *mddev = r1_bio->mddev;
2148 struct r1conf *conf = mddev->private;
2149 struct md_rdev *rdev = conf->mirrors[i].rdev;
2151 /* bio has the data to be written to device 'i' where
2152 * we just recently had a write error.
2153 * We repeatedly clone the bio and trim down to one block,
2154 * then try the write. Where the write fails we record
2156 * It is conceivable that the bio doesn't exactly align with
2157 * blocks. We must handle this somehow.
2159 * We currently own a reference on the rdev.
2165 int sect_to_write = r1_bio->sectors;
2168 if (rdev->badblocks.shift < 0)
2171 block_sectors = roundup(1 << rdev->badblocks.shift,
2172 bdev_logical_block_size(rdev->bdev) >> 9);
2173 sector = r1_bio->sector;
2174 sectors = ((sector + block_sectors)
2175 & ~(sector_t)(block_sectors - 1))
2178 while (sect_to_write) {
2180 if (sectors > sect_to_write)
2181 sectors = sect_to_write;
2182 /* Write at 'sector' for 'sectors'*/
2184 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2185 unsigned vcnt = r1_bio->behind_page_count;
2186 struct bio_vec *vec = r1_bio->behind_bvecs;
2188 while (!vec->bv_page) {
2193 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2194 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2196 wbio->bi_vcnt = vcnt;
2198 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2201 wbio->bi_rw = WRITE;
2202 wbio->bi_iter.bi_sector = r1_bio->sector;
2203 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2205 bio_trim(wbio, sector - r1_bio->sector, sectors);
2206 wbio->bi_iter.bi_sector += rdev->data_offset;
2207 wbio->bi_bdev = rdev->bdev;
2208 if (submit_bio_wait(WRITE, wbio) < 0)
2210 ok = rdev_set_badblocks(rdev, sector,
2215 sect_to_write -= sectors;
2217 sectors = block_sectors;
2222 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2225 int s = r1_bio->sectors;
2226 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2227 struct md_rdev *rdev = conf->mirrors[m].rdev;
2228 struct bio *bio = r1_bio->bios[m];
2229 if (bio->bi_end_io == NULL)
2231 if (!bio->bi_error &&
2232 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2233 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2235 if (bio->bi_error &&
2236 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2237 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2238 md_error(conf->mddev, rdev);
2242 md_done_sync(conf->mddev, s, 1);
2245 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2249 for (m = 0; m < conf->raid_disks * 2 ; m++)
2250 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2251 struct md_rdev *rdev = conf->mirrors[m].rdev;
2252 rdev_clear_badblocks(rdev,
2254 r1_bio->sectors, 0);
2255 rdev_dec_pending(rdev, conf->mddev);
2256 } else if (r1_bio->bios[m] != NULL) {
2257 /* This drive got a write error. We need to
2258 * narrow down and record precise write
2262 if (!narrow_write_error(r1_bio, m)) {
2263 md_error(conf->mddev,
2264 conf->mirrors[m].rdev);
2265 /* an I/O failed, we can't clear the bitmap */
2266 set_bit(R1BIO_Degraded, &r1_bio->state);
2268 rdev_dec_pending(conf->mirrors[m].rdev,
2272 spin_lock_irq(&conf->device_lock);
2273 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2275 spin_unlock_irq(&conf->device_lock);
2276 md_wakeup_thread(conf->mddev->thread);
2278 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2279 close_write(r1_bio);
2280 raid_end_bio_io(r1_bio);
2284 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2288 struct mddev *mddev = conf->mddev;
2290 char b[BDEVNAME_SIZE];
2291 struct md_rdev *rdev;
2293 clear_bit(R1BIO_ReadError, &r1_bio->state);
2294 /* we got a read error. Maybe the drive is bad. Maybe just
2295 * the block and we can fix it.
2296 * We freeze all other IO, and try reading the block from
2297 * other devices. When we find one, we re-write
2298 * and check it that fixes the read error.
2299 * This is all done synchronously while the array is
2302 if (mddev->ro == 0) {
2303 freeze_array(conf, 1);
2304 fix_read_error(conf, r1_bio->read_disk,
2305 r1_bio->sector, r1_bio->sectors);
2306 unfreeze_array(conf);
2308 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2309 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2311 bio = r1_bio->bios[r1_bio->read_disk];
2312 bdevname(bio->bi_bdev, b);
2314 disk = read_balance(conf, r1_bio, &max_sectors);
2316 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2317 " read error for block %llu\n",
2318 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2319 raid_end_bio_io(r1_bio);
2321 const unsigned long do_sync
2322 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2324 r1_bio->bios[r1_bio->read_disk] =
2325 mddev->ro ? IO_BLOCKED : NULL;
2328 r1_bio->read_disk = disk;
2329 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2330 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2332 r1_bio->bios[r1_bio->read_disk] = bio;
2333 rdev = conf->mirrors[disk].rdev;
2334 printk_ratelimited(KERN_ERR
2335 "md/raid1:%s: redirecting sector %llu"
2336 " to other mirror: %s\n",
2338 (unsigned long long)r1_bio->sector,
2339 bdevname(rdev->bdev, b));
2340 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2341 bio->bi_bdev = rdev->bdev;
2342 bio->bi_end_io = raid1_end_read_request;
2343 bio->bi_rw = READ | do_sync;
2344 bio->bi_private = r1_bio;
2345 if (max_sectors < r1_bio->sectors) {
2346 /* Drat - have to split this up more */
2347 struct bio *mbio = r1_bio->master_bio;
2348 int sectors_handled = (r1_bio->sector + max_sectors
2349 - mbio->bi_iter.bi_sector);
2350 r1_bio->sectors = max_sectors;
2351 spin_lock_irq(&conf->device_lock);
2352 if (mbio->bi_phys_segments == 0)
2353 mbio->bi_phys_segments = 2;
2355 mbio->bi_phys_segments++;
2356 spin_unlock_irq(&conf->device_lock);
2357 generic_make_request(bio);
2360 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2362 r1_bio->master_bio = mbio;
2363 r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2365 set_bit(R1BIO_ReadError, &r1_bio->state);
2366 r1_bio->mddev = mddev;
2367 r1_bio->sector = mbio->bi_iter.bi_sector +
2372 generic_make_request(bio);
2376 static void raid1d(struct md_thread *thread)
2378 struct mddev *mddev = thread->mddev;
2379 struct r1bio *r1_bio;
2380 unsigned long flags;
2381 struct r1conf *conf = mddev->private;
2382 struct list_head *head = &conf->retry_list;
2383 struct blk_plug plug;
2385 md_check_recovery(mddev);
2387 if (!list_empty_careful(&conf->bio_end_io_list) &&
2388 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2390 spin_lock_irqsave(&conf->device_lock, flags);
2391 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2392 while (!list_empty(&conf->bio_end_io_list)) {
2393 list_move(conf->bio_end_io_list.prev, &tmp);
2397 spin_unlock_irqrestore(&conf->device_lock, flags);
2398 while (!list_empty(&tmp)) {
2399 r1_bio = list_first_entry(&tmp, struct r1bio,
2401 list_del(&r1_bio->retry_list);
2402 if (mddev->degraded)
2403 set_bit(R1BIO_Degraded, &r1_bio->state);
2404 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2405 close_write(r1_bio);
2406 raid_end_bio_io(r1_bio);
2410 blk_start_plug(&plug);
2413 flush_pending_writes(conf);
2415 spin_lock_irqsave(&conf->device_lock, flags);
2416 if (list_empty(head)) {
2417 spin_unlock_irqrestore(&conf->device_lock, flags);
2420 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2421 list_del(head->prev);
2423 spin_unlock_irqrestore(&conf->device_lock, flags);
2425 mddev = r1_bio->mddev;
2426 conf = mddev->private;
2427 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2428 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2429 test_bit(R1BIO_WriteError, &r1_bio->state))
2430 handle_sync_write_finished(conf, r1_bio);
2432 sync_request_write(mddev, r1_bio);
2433 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2434 test_bit(R1BIO_WriteError, &r1_bio->state))
2435 handle_write_finished(conf, r1_bio);
2436 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2437 handle_read_error(conf, r1_bio);
2439 /* just a partial read to be scheduled from separate
2442 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2445 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2446 md_check_recovery(mddev);
2448 blk_finish_plug(&plug);
2451 static int init_resync(struct r1conf *conf)
2455 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2456 BUG_ON(conf->r1buf_pool);
2457 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2459 if (!conf->r1buf_pool)
2461 conf->next_resync = 0;
2466 * perform a "sync" on one "block"
2468 * We need to make sure that no normal I/O request - particularly write
2469 * requests - conflict with active sync requests.
2471 * This is achieved by tracking pending requests and a 'barrier' concept
2472 * that can be installed to exclude normal IO requests.
2475 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2478 struct r1conf *conf = mddev->private;
2479 struct r1bio *r1_bio;
2481 sector_t max_sector, nr_sectors;
2485 int write_targets = 0, read_targets = 0;
2486 sector_t sync_blocks;
2487 int still_degraded = 0;
2488 int good_sectors = RESYNC_SECTORS;
2489 int min_bad = 0; /* number of sectors that are bad in all devices */
2491 if (!conf->r1buf_pool)
2492 if (init_resync(conf))
2495 max_sector = mddev->dev_sectors;
2496 if (sector_nr >= max_sector) {
2497 /* If we aborted, we need to abort the
2498 * sync on the 'current' bitmap chunk (there will
2499 * only be one in raid1 resync.
2500 * We can find the current addess in mddev->curr_resync
2502 if (mddev->curr_resync < max_sector) /* aborted */
2503 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2505 else /* completed sync */
2508 bitmap_close_sync(mddev->bitmap);
2511 if (mddev_is_clustered(mddev)) {
2512 conf->cluster_sync_low = 0;
2513 conf->cluster_sync_high = 0;
2518 if (mddev->bitmap == NULL &&
2519 mddev->recovery_cp == MaxSector &&
2520 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2521 conf->fullsync == 0) {
2523 return max_sector - sector_nr;
2525 /* before building a request, check if we can skip these blocks..
2526 * This call the bitmap_start_sync doesn't actually record anything
2528 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2529 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2530 /* We can skip this block, and probably several more */
2536 * If there is non-resync activity waiting for a turn, then let it
2537 * though before starting on this new sync request.
2539 if (conf->nr_waiting)
2540 schedule_timeout_uninterruptible(1);
2542 /* we are incrementing sector_nr below. To be safe, we check against
2543 * sector_nr + two times RESYNC_SECTORS
2546 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2547 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2548 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2550 raise_barrier(conf, sector_nr);
2554 * If we get a correctably read error during resync or recovery,
2555 * we might want to read from a different device. So we
2556 * flag all drives that could conceivably be read from for READ,
2557 * and any others (which will be non-In_sync devices) for WRITE.
2558 * If a read fails, we try reading from something else for which READ
2562 r1_bio->mddev = mddev;
2563 r1_bio->sector = sector_nr;
2565 set_bit(R1BIO_IsSync, &r1_bio->state);
2567 for (i = 0; i < conf->raid_disks * 2; i++) {
2568 struct md_rdev *rdev;
2569 bio = r1_bio->bios[i];
2572 rdev = rcu_dereference(conf->mirrors[i].rdev);
2574 test_bit(Faulty, &rdev->flags)) {
2575 if (i < conf->raid_disks)
2577 } else if (!test_bit(In_sync, &rdev->flags)) {
2579 bio->bi_end_io = end_sync_write;
2582 /* may need to read from here */
2583 sector_t first_bad = MaxSector;
2586 if (is_badblock(rdev, sector_nr, good_sectors,
2587 &first_bad, &bad_sectors)) {
2588 if (first_bad > sector_nr)
2589 good_sectors = first_bad - sector_nr;
2591 bad_sectors -= (sector_nr - first_bad);
2593 min_bad > bad_sectors)
2594 min_bad = bad_sectors;
2597 if (sector_nr < first_bad) {
2598 if (test_bit(WriteMostly, &rdev->flags)) {
2606 bio->bi_end_io = end_sync_read;
2608 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2609 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2610 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2612 * The device is suitable for reading (InSync),
2613 * but has bad block(s) here. Let's try to correct them,
2614 * if we are doing resync or repair. Otherwise, leave
2615 * this device alone for this sync request.
2618 bio->bi_end_io = end_sync_write;
2622 if (bio->bi_end_io) {
2623 atomic_inc(&rdev->nr_pending);
2624 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2625 bio->bi_bdev = rdev->bdev;
2626 bio->bi_private = r1_bio;
2632 r1_bio->read_disk = disk;
2634 if (read_targets == 0 && min_bad > 0) {
2635 /* These sectors are bad on all InSync devices, so we
2636 * need to mark them bad on all write targets
2639 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2640 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2641 struct md_rdev *rdev = conf->mirrors[i].rdev;
2642 ok = rdev_set_badblocks(rdev, sector_nr,
2646 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2651 /* Cannot record the badblocks, so need to
2653 * If there are multiple read targets, could just
2654 * fail the really bad ones ???
2656 conf->recovery_disabled = mddev->recovery_disabled;
2657 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2663 if (min_bad > 0 && min_bad < good_sectors) {
2664 /* only resync enough to reach the next bad->good
2666 good_sectors = min_bad;
2669 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2670 /* extra read targets are also write targets */
2671 write_targets += read_targets-1;
2673 if (write_targets == 0 || read_targets == 0) {
2674 /* There is nowhere to write, so all non-sync
2675 * drives must be failed - so we are finished
2679 max_sector = sector_nr + min_bad;
2680 rv = max_sector - sector_nr;
2686 if (max_sector > mddev->resync_max)
2687 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2688 if (max_sector > sector_nr + good_sectors)
2689 max_sector = sector_nr + good_sectors;
2694 int len = PAGE_SIZE;
2695 if (sector_nr + (len>>9) > max_sector)
2696 len = (max_sector - sector_nr) << 9;
2699 if (sync_blocks == 0) {
2700 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2701 &sync_blocks, still_degraded) &&
2703 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2705 if ((len >> 9) > sync_blocks)
2706 len = sync_blocks<<9;
2709 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2710 bio = r1_bio->bios[i];
2711 if (bio->bi_end_io) {
2712 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2713 if (bio_add_page(bio, page, len, 0) == 0) {
2715 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2718 bio = r1_bio->bios[i];
2719 if (bio->bi_end_io==NULL)
2721 /* remove last page from this bio */
2723 bio->bi_iter.bi_size -= len;
2724 bio_clear_flag(bio, BIO_SEG_VALID);
2730 nr_sectors += len>>9;
2731 sector_nr += len>>9;
2732 sync_blocks -= (len>>9);
2733 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2735 r1_bio->sectors = nr_sectors;
2737 if (mddev_is_clustered(mddev) &&
2738 conf->cluster_sync_high < sector_nr + nr_sectors) {
2739 conf->cluster_sync_low = mddev->curr_resync_completed;
2740 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2741 /* Send resync message */
2742 md_cluster_ops->resync_info_update(mddev,
2743 conf->cluster_sync_low,
2744 conf->cluster_sync_high);
2747 /* For a user-requested sync, we read all readable devices and do a
2750 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2751 atomic_set(&r1_bio->remaining, read_targets);
2752 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2753 bio = r1_bio->bios[i];
2754 if (bio->bi_end_io == end_sync_read) {
2756 md_sync_acct(bio->bi_bdev, nr_sectors);
2757 generic_make_request(bio);
2761 atomic_set(&r1_bio->remaining, 1);
2762 bio = r1_bio->bios[r1_bio->read_disk];
2763 md_sync_acct(bio->bi_bdev, nr_sectors);
2764 generic_make_request(bio);
2770 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2775 return mddev->dev_sectors;
2778 static struct r1conf *setup_conf(struct mddev *mddev)
2780 struct r1conf *conf;
2782 struct raid1_info *disk;
2783 struct md_rdev *rdev;
2786 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2790 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2791 * mddev->raid_disks * 2,
2796 conf->tmppage = alloc_page(GFP_KERNEL);
2800 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2801 if (!conf->poolinfo)
2803 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2804 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2807 if (!conf->r1bio_pool)
2810 conf->poolinfo->mddev = mddev;
2813 spin_lock_init(&conf->device_lock);
2814 rdev_for_each(rdev, mddev) {
2815 struct request_queue *q;
2816 int disk_idx = rdev->raid_disk;
2817 if (disk_idx >= mddev->raid_disks
2820 if (test_bit(Replacement, &rdev->flags))
2821 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2823 disk = conf->mirrors + disk_idx;
2828 q = bdev_get_queue(rdev->bdev);
2830 disk->head_position = 0;
2831 disk->seq_start = MaxSector;
2833 conf->raid_disks = mddev->raid_disks;
2834 conf->mddev = mddev;
2835 INIT_LIST_HEAD(&conf->retry_list);
2836 INIT_LIST_HEAD(&conf->bio_end_io_list);
2838 spin_lock_init(&conf->resync_lock);
2839 init_waitqueue_head(&conf->wait_barrier);
2841 bio_list_init(&conf->pending_bio_list);
2842 conf->pending_count = 0;
2843 conf->recovery_disabled = mddev->recovery_disabled - 1;
2845 conf->start_next_window = MaxSector;
2846 conf->current_window_requests = conf->next_window_requests = 0;
2849 for (i = 0; i < conf->raid_disks * 2; i++) {
2851 disk = conf->mirrors + i;
2853 if (i < conf->raid_disks &&
2854 disk[conf->raid_disks].rdev) {
2855 /* This slot has a replacement. */
2857 /* No original, just make the replacement
2858 * a recovering spare
2861 disk[conf->raid_disks].rdev;
2862 disk[conf->raid_disks].rdev = NULL;
2863 } else if (!test_bit(In_sync, &disk->rdev->flags))
2864 /* Original is not in_sync - bad */
2869 !test_bit(In_sync, &disk->rdev->flags)) {
2870 disk->head_position = 0;
2872 (disk->rdev->saved_raid_disk < 0))
2878 conf->thread = md_register_thread(raid1d, mddev, "raid1");
2879 if (!conf->thread) {
2881 "md/raid1:%s: couldn't allocate thread\n",
2890 mempool_destroy(conf->r1bio_pool);
2891 kfree(conf->mirrors);
2892 safe_put_page(conf->tmppage);
2893 kfree(conf->poolinfo);
2896 return ERR_PTR(err);
2899 static void raid1_free(struct mddev *mddev, void *priv);
2900 static int raid1_run(struct mddev *mddev)
2902 struct r1conf *conf;
2904 struct md_rdev *rdev;
2906 bool discard_supported = false;
2908 if (mddev->level != 1) {
2909 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2910 mdname(mddev), mddev->level);
2913 if (mddev->reshape_position != MaxSector) {
2914 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2919 * copy the already verified devices into our private RAID1
2920 * bookkeeping area. [whatever we allocate in run(),
2921 * should be freed in raid1_free()]
2923 if (mddev->private == NULL)
2924 conf = setup_conf(mddev);
2926 conf = mddev->private;
2929 return PTR_ERR(conf);
2932 blk_queue_max_write_same_sectors(mddev->queue, 0);
2934 rdev_for_each(rdev, mddev) {
2935 if (!mddev->gendisk)
2937 disk_stack_limits(mddev->gendisk, rdev->bdev,
2938 rdev->data_offset << 9);
2939 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2940 discard_supported = true;
2943 mddev->degraded = 0;
2944 for (i=0; i < conf->raid_disks; i++)
2945 if (conf->mirrors[i].rdev == NULL ||
2946 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2947 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2950 if (conf->raid_disks - mddev->degraded == 1)
2951 mddev->recovery_cp = MaxSector;
2953 if (mddev->recovery_cp != MaxSector)
2954 printk(KERN_NOTICE "md/raid1:%s: not clean"
2955 " -- starting background reconstruction\n",
2958 "md/raid1:%s: active with %d out of %d mirrors\n",
2959 mdname(mddev), mddev->raid_disks - mddev->degraded,
2963 * Ok, everything is just fine now
2965 mddev->thread = conf->thread;
2966 conf->thread = NULL;
2967 mddev->private = conf;
2969 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2972 if (discard_supported)
2973 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2976 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2980 ret = md_integrity_register(mddev);
2982 md_unregister_thread(&mddev->thread);
2983 raid1_free(mddev, conf);
2988 static void raid1_free(struct mddev *mddev, void *priv)
2990 struct r1conf *conf = priv;
2992 mempool_destroy(conf->r1bio_pool);
2993 kfree(conf->mirrors);
2994 safe_put_page(conf->tmppage);
2995 kfree(conf->poolinfo);
2999 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3001 /* no resync is happening, and there is enough space
3002 * on all devices, so we can resize.
3003 * We need to make sure resync covers any new space.
3004 * If the array is shrinking we should possibly wait until
3005 * any io in the removed space completes, but it hardly seems
3008 sector_t newsize = raid1_size(mddev, sectors, 0);
3009 if (mddev->external_size &&
3010 mddev->array_sectors > newsize)
3012 if (mddev->bitmap) {
3013 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3017 md_set_array_sectors(mddev, newsize);
3018 set_capacity(mddev->gendisk, mddev->array_sectors);
3019 revalidate_disk(mddev->gendisk);
3020 if (sectors > mddev->dev_sectors &&
3021 mddev->recovery_cp > mddev->dev_sectors) {
3022 mddev->recovery_cp = mddev->dev_sectors;
3023 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3025 mddev->dev_sectors = sectors;
3026 mddev->resync_max_sectors = sectors;
3030 static int raid1_reshape(struct mddev *mddev)
3033 * 1/ resize the r1bio_pool
3034 * 2/ resize conf->mirrors
3036 * We allocate a new r1bio_pool if we can.
3037 * Then raise a device barrier and wait until all IO stops.
3038 * Then resize conf->mirrors and swap in the new r1bio pool.
3040 * At the same time, we "pack" the devices so that all the missing
3041 * devices have the higher raid_disk numbers.
3043 mempool_t *newpool, *oldpool;
3044 struct pool_info *newpoolinfo;
3045 struct raid1_info *newmirrors;
3046 struct r1conf *conf = mddev->private;
3047 int cnt, raid_disks;
3048 unsigned long flags;
3051 /* Cannot change chunk_size, layout, or level */
3052 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3053 mddev->layout != mddev->new_layout ||
3054 mddev->level != mddev->new_level) {
3055 mddev->new_chunk_sectors = mddev->chunk_sectors;
3056 mddev->new_layout = mddev->layout;
3057 mddev->new_level = mddev->level;
3061 if (!mddev_is_clustered(mddev)) {
3062 err = md_allow_write(mddev);
3067 raid_disks = mddev->raid_disks + mddev->delta_disks;
3069 if (raid_disks < conf->raid_disks) {
3071 for (d= 0; d < conf->raid_disks; d++)
3072 if (conf->mirrors[d].rdev)
3074 if (cnt > raid_disks)
3078 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3081 newpoolinfo->mddev = mddev;
3082 newpoolinfo->raid_disks = raid_disks * 2;
3084 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3085 r1bio_pool_free, newpoolinfo);
3090 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3094 mempool_destroy(newpool);
3098 freeze_array(conf, 0);
3100 /* ok, everything is stopped */
3101 oldpool = conf->r1bio_pool;
3102 conf->r1bio_pool = newpool;
3104 for (d = d2 = 0; d < conf->raid_disks; d++) {
3105 struct md_rdev *rdev = conf->mirrors[d].rdev;
3106 if (rdev && rdev->raid_disk != d2) {
3107 sysfs_unlink_rdev(mddev, rdev);
3108 rdev->raid_disk = d2;
3109 sysfs_unlink_rdev(mddev, rdev);
3110 if (sysfs_link_rdev(mddev, rdev))
3112 "md/raid1:%s: cannot register rd%d\n",
3113 mdname(mddev), rdev->raid_disk);
3116 newmirrors[d2++].rdev = rdev;
3118 kfree(conf->mirrors);
3119 conf->mirrors = newmirrors;
3120 kfree(conf->poolinfo);
3121 conf->poolinfo = newpoolinfo;
3123 spin_lock_irqsave(&conf->device_lock, flags);
3124 mddev->degraded += (raid_disks - conf->raid_disks);
3125 spin_unlock_irqrestore(&conf->device_lock, flags);
3126 conf->raid_disks = mddev->raid_disks = raid_disks;
3127 mddev->delta_disks = 0;
3129 unfreeze_array(conf);
3131 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3132 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3133 md_wakeup_thread(mddev->thread);
3135 mempool_destroy(oldpool);
3139 static void raid1_quiesce(struct mddev *mddev, int state)
3141 struct r1conf *conf = mddev->private;
3144 case 2: /* wake for suspend */
3145 wake_up(&conf->wait_barrier);
3148 freeze_array(conf, 0);
3151 unfreeze_array(conf);
3156 static void *raid1_takeover(struct mddev *mddev)
3158 /* raid1 can take over:
3159 * raid5 with 2 devices, any layout or chunk size
3161 if (mddev->level == 5 && mddev->raid_disks == 2) {
3162 struct r1conf *conf;
3163 mddev->new_level = 1;
3164 mddev->new_layout = 0;
3165 mddev->new_chunk_sectors = 0;
3166 conf = setup_conf(mddev);
3168 /* Array must appear to be quiesced */
3169 conf->array_frozen = 1;
3172 return ERR_PTR(-EINVAL);
3175 static struct md_personality raid1_personality =
3179 .owner = THIS_MODULE,
3180 .make_request = raid1_make_request,
3183 .status = raid1_status,
3184 .error_handler = raid1_error,
3185 .hot_add_disk = raid1_add_disk,
3186 .hot_remove_disk= raid1_remove_disk,
3187 .spare_active = raid1_spare_active,
3188 .sync_request = raid1_sync_request,
3189 .resize = raid1_resize,
3191 .check_reshape = raid1_reshape,
3192 .quiesce = raid1_quiesce,
3193 .takeover = raid1_takeover,
3194 .congested = raid1_congested,
3197 static int __init raid_init(void)
3199 return register_md_personality(&raid1_personality);
3202 static void raid_exit(void)
3204 unregister_md_personality(&raid1_personality);
3207 module_init(raid_init);
3208 module_exit(raid_exit);
3209 MODULE_LICENSE("GPL");
3210 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3211 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3212 MODULE_ALIAS("md-raid1");
3213 MODULE_ALIAS("md-level-1");
3215 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);