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 NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
95 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
97 struct pool_info *pi = data;
103 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
108 * Allocate bios : 1 for reading, n-1 for writing
110 for (j = pi->raid_disks ; j-- ; ) {
111 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
114 r1_bio->bios[j] = bio;
117 * Allocate RESYNC_PAGES data pages and attach them to
119 * If this is a user-requested check/repair, allocate
120 * RESYNC_PAGES for each bio.
122 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
123 need_pages = pi->raid_disks;
126 for (j = 0; j < need_pages; j++) {
127 bio = r1_bio->bios[j];
128 bio->bi_vcnt = RESYNC_PAGES;
130 if (bio_alloc_pages(bio, gfp_flags))
133 /* If not user-requests, copy the page pointers to all bios */
134 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
135 for (i=0; i<RESYNC_PAGES ; i++)
136 for (j=1; j<pi->raid_disks; j++)
137 r1_bio->bios[j]->bi_io_vec[i].bv_page =
138 r1_bio->bios[0]->bi_io_vec[i].bv_page;
141 r1_bio->master_bio = NULL;
149 bio_for_each_segment_all(bv, r1_bio->bios[j], i)
150 __free_page(bv->bv_page);
154 while (++j < pi->raid_disks)
155 bio_put(r1_bio->bios[j]);
156 r1bio_pool_free(r1_bio, data);
160 static void r1buf_pool_free(void *__r1_bio, void *data)
162 struct pool_info *pi = data;
164 struct r1bio *r1bio = __r1_bio;
166 for (i = 0; i < RESYNC_PAGES; i++)
167 for (j = pi->raid_disks; j-- ;) {
169 r1bio->bios[j]->bi_io_vec[i].bv_page !=
170 r1bio->bios[0]->bi_io_vec[i].bv_page)
171 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
173 for (i=0 ; i < pi->raid_disks; i++)
174 bio_put(r1bio->bios[i]);
176 r1bio_pool_free(r1bio, data);
179 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
183 for (i = 0; i < conf->raid_disks * 2; i++) {
184 struct bio **bio = r1_bio->bios + i;
185 if (!BIO_SPECIAL(*bio))
191 static void free_r1bio(struct r1bio *r1_bio)
193 struct r1conf *conf = r1_bio->mddev->private;
195 put_all_bios(conf, r1_bio);
196 mempool_free(r1_bio, conf->r1bio_pool);
199 static void put_buf(struct r1bio *r1_bio)
201 struct r1conf *conf = r1_bio->mddev->private;
204 for (i = 0; i < conf->raid_disks * 2; i++) {
205 struct bio *bio = r1_bio->bios[i];
207 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
210 mempool_free(r1_bio, conf->r1buf_pool);
215 static void reschedule_retry(struct r1bio *r1_bio)
218 struct mddev *mddev = r1_bio->mddev;
219 struct r1conf *conf = mddev->private;
221 spin_lock_irqsave(&conf->device_lock, flags);
222 list_add(&r1_bio->retry_list, &conf->retry_list);
224 spin_unlock_irqrestore(&conf->device_lock, flags);
226 wake_up(&conf->wait_barrier);
227 md_wakeup_thread(mddev->thread);
231 * raid_end_bio_io() is called when we have finished servicing a mirrored
232 * operation and are ready to return a success/failure code to the buffer
235 static void call_bio_endio(struct r1bio *r1_bio)
237 struct bio *bio = r1_bio->master_bio;
239 struct r1conf *conf = r1_bio->mddev->private;
240 sector_t start_next_window = r1_bio->start_next_window;
241 sector_t bi_sector = bio->bi_iter.bi_sector;
243 if (bio->bi_phys_segments) {
245 spin_lock_irqsave(&conf->device_lock, flags);
246 bio->bi_phys_segments--;
247 done = (bio->bi_phys_segments == 0);
248 spin_unlock_irqrestore(&conf->device_lock, flags);
250 * make_request() might be waiting for
251 * bi_phys_segments to decrease
253 wake_up(&conf->wait_barrier);
257 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
258 bio->bi_error = -EIO;
263 * Wake up any possible resync thread that waits for the device
266 allow_barrier(conf, start_next_window, bi_sector);
270 static void raid_end_bio_io(struct r1bio *r1_bio)
272 struct bio *bio = r1_bio->master_bio;
274 /* if nobody has done the final endio yet, do it now */
275 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
276 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
277 (bio_data_dir(bio) == WRITE) ? "write" : "read",
278 (unsigned long long) bio->bi_iter.bi_sector,
279 (unsigned long long) bio_end_sector(bio) - 1);
281 call_bio_endio(r1_bio);
287 * Update disk head position estimator based on IRQ completion info.
289 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
291 struct r1conf *conf = r1_bio->mddev->private;
293 conf->mirrors[disk].head_position =
294 r1_bio->sector + (r1_bio->sectors);
298 * Find the disk number which triggered given bio
300 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
303 struct r1conf *conf = r1_bio->mddev->private;
304 int raid_disks = conf->raid_disks;
306 for (mirror = 0; mirror < raid_disks * 2; mirror++)
307 if (r1_bio->bios[mirror] == bio)
310 BUG_ON(mirror == raid_disks * 2);
311 update_head_pos(mirror, r1_bio);
316 static void raid1_end_read_request(struct bio *bio)
318 int uptodate = !bio->bi_error;
319 struct r1bio *r1_bio = bio->bi_private;
321 struct r1conf *conf = r1_bio->mddev->private;
323 mirror = r1_bio->read_disk;
325 * this branch is our 'one mirror IO has finished' event handler:
327 update_head_pos(mirror, r1_bio);
330 set_bit(R1BIO_Uptodate, &r1_bio->state);
332 /* If all other devices have failed, we want to return
333 * the error upwards rather than fail the last device.
334 * Here we redefine "uptodate" to mean "Don't want to retry"
337 spin_lock_irqsave(&conf->device_lock, flags);
338 if (r1_bio->mddev->degraded == conf->raid_disks ||
339 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
340 test_bit(In_sync, &conf->mirrors[mirror].rdev->flags)))
342 spin_unlock_irqrestore(&conf->device_lock, flags);
346 raid_end_bio_io(r1_bio);
347 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
352 char b[BDEVNAME_SIZE];
354 KERN_ERR "md/raid1:%s: %s: "
355 "rescheduling sector %llu\n",
357 bdevname(conf->mirrors[mirror].rdev->bdev,
359 (unsigned long long)r1_bio->sector);
360 set_bit(R1BIO_ReadError, &r1_bio->state);
361 reschedule_retry(r1_bio);
362 /* don't drop the reference on read_disk yet */
366 static void close_write(struct r1bio *r1_bio)
368 /* it really is the end of this request */
369 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
370 /* free extra copy of the data pages */
371 int i = r1_bio->behind_page_count;
373 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
374 kfree(r1_bio->behind_bvecs);
375 r1_bio->behind_bvecs = NULL;
377 /* clear the bitmap if all writes complete successfully */
378 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
380 !test_bit(R1BIO_Degraded, &r1_bio->state),
381 test_bit(R1BIO_BehindIO, &r1_bio->state));
382 md_write_end(r1_bio->mddev);
385 static void r1_bio_write_done(struct r1bio *r1_bio)
387 if (!atomic_dec_and_test(&r1_bio->remaining))
390 if (test_bit(R1BIO_WriteError, &r1_bio->state))
391 reschedule_retry(r1_bio);
394 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
395 reschedule_retry(r1_bio);
397 raid_end_bio_io(r1_bio);
401 static void raid1_end_write_request(struct bio *bio)
403 struct r1bio *r1_bio = bio->bi_private;
404 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
405 struct r1conf *conf = r1_bio->mddev->private;
406 struct bio *to_put = NULL;
408 mirror = find_bio_disk(r1_bio, bio);
411 * 'one mirror IO has finished' event handler:
414 set_bit(WriteErrorSeen,
415 &conf->mirrors[mirror].rdev->flags);
416 if (!test_and_set_bit(WantReplacement,
417 &conf->mirrors[mirror].rdev->flags))
418 set_bit(MD_RECOVERY_NEEDED, &
419 conf->mddev->recovery);
421 set_bit(R1BIO_WriteError, &r1_bio->state);
424 * Set R1BIO_Uptodate in our master bio, so that we
425 * will return a good error code for to the higher
426 * levels even if IO on some other mirrored buffer
429 * The 'master' represents the composite IO operation
430 * to user-side. So if something waits for IO, then it
431 * will wait for the 'master' bio.
436 r1_bio->bios[mirror] = NULL;
439 * Do not set R1BIO_Uptodate if the current device is
440 * rebuilding or Faulty. This is because we cannot use
441 * such device for properly reading the data back (we could
442 * potentially use it, if the current write would have felt
443 * before rdev->recovery_offset, but for simplicity we don't
446 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
447 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
448 set_bit(R1BIO_Uptodate, &r1_bio->state);
450 /* Maybe we can clear some bad blocks. */
451 if (is_badblock(conf->mirrors[mirror].rdev,
452 r1_bio->sector, r1_bio->sectors,
453 &first_bad, &bad_sectors)) {
454 r1_bio->bios[mirror] = IO_MADE_GOOD;
455 set_bit(R1BIO_MadeGood, &r1_bio->state);
460 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
461 atomic_dec(&r1_bio->behind_remaining);
464 * In behind mode, we ACK the master bio once the I/O
465 * has safely reached all non-writemostly
466 * disks. Setting the Returned bit ensures that this
467 * gets done only once -- we don't ever want to return
468 * -EIO here, instead we'll wait
470 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
471 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
472 /* Maybe we can return now */
473 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
474 struct bio *mbio = r1_bio->master_bio;
475 pr_debug("raid1: behind end write sectors"
477 (unsigned long long) mbio->bi_iter.bi_sector,
478 (unsigned long long) bio_end_sector(mbio) - 1);
479 call_bio_endio(r1_bio);
483 if (r1_bio->bios[mirror] == NULL)
484 rdev_dec_pending(conf->mirrors[mirror].rdev,
488 * Let's see if all mirrored write operations have finished
491 r1_bio_write_done(r1_bio);
498 * This routine returns the disk from which the requested read should
499 * be done. There is a per-array 'next expected sequential IO' sector
500 * number - if this matches on the next IO then we use the last disk.
501 * There is also a per-disk 'last know head position' sector that is
502 * maintained from IRQ contexts, both the normal and the resync IO
503 * completion handlers update this position correctly. If there is no
504 * perfect sequential match then we pick the disk whose head is closest.
506 * If there are 2 mirrors in the same 2 devices, performance degrades
507 * because position is mirror, not device based.
509 * The rdev for the device selected will have nr_pending incremented.
511 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
513 const sector_t this_sector = r1_bio->sector;
515 int best_good_sectors;
516 int best_disk, best_dist_disk, best_pending_disk;
520 unsigned int min_pending;
521 struct md_rdev *rdev;
523 int choose_next_idle;
527 * Check if we can balance. We can balance on the whole
528 * device if no resync is going on, or below the resync window.
529 * We take the first readable disk when above the resync window.
532 sectors = r1_bio->sectors;
535 best_dist = MaxSector;
536 best_pending_disk = -1;
537 min_pending = UINT_MAX;
538 best_good_sectors = 0;
540 choose_next_idle = 0;
542 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
543 (mddev_is_clustered(conf->mddev) &&
544 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
545 this_sector + sectors)))
550 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
554 unsigned int pending;
557 rdev = rcu_dereference(conf->mirrors[disk].rdev);
558 if (r1_bio->bios[disk] == IO_BLOCKED
560 || test_bit(Faulty, &rdev->flags))
562 if (!test_bit(In_sync, &rdev->flags) &&
563 rdev->recovery_offset < this_sector + sectors)
565 if (test_bit(WriteMostly, &rdev->flags)) {
566 /* Don't balance among write-mostly, just
567 * use the first as a last resort */
568 if (best_dist_disk < 0) {
569 if (is_badblock(rdev, this_sector, sectors,
570 &first_bad, &bad_sectors)) {
571 if (first_bad < this_sector)
572 /* Cannot use this */
574 best_good_sectors = first_bad - this_sector;
576 best_good_sectors = sectors;
577 best_dist_disk = disk;
578 best_pending_disk = disk;
582 /* This is a reasonable device to use. It might
585 if (is_badblock(rdev, this_sector, sectors,
586 &first_bad, &bad_sectors)) {
587 if (best_dist < MaxSector)
588 /* already have a better device */
590 if (first_bad <= this_sector) {
591 /* cannot read here. If this is the 'primary'
592 * device, then we must not read beyond
593 * bad_sectors from another device..
595 bad_sectors -= (this_sector - first_bad);
596 if (choose_first && sectors > bad_sectors)
597 sectors = bad_sectors;
598 if (best_good_sectors > sectors)
599 best_good_sectors = sectors;
602 sector_t good_sectors = first_bad - this_sector;
603 if (good_sectors > best_good_sectors) {
604 best_good_sectors = good_sectors;
612 best_good_sectors = sectors;
614 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
615 has_nonrot_disk |= nonrot;
616 pending = atomic_read(&rdev->nr_pending);
617 dist = abs(this_sector - conf->mirrors[disk].head_position);
622 /* Don't change to another disk for sequential reads */
623 if (conf->mirrors[disk].next_seq_sect == this_sector
625 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
626 struct raid1_info *mirror = &conf->mirrors[disk];
630 * If buffered sequential IO size exceeds optimal
631 * iosize, check if there is idle disk. If yes, choose
632 * the idle disk. read_balance could already choose an
633 * idle disk before noticing it's a sequential IO in
634 * this disk. This doesn't matter because this disk
635 * will idle, next time it will be utilized after the
636 * first disk has IO size exceeds optimal iosize. In
637 * this way, iosize of the first disk will be optimal
638 * iosize at least. iosize of the second disk might be
639 * small, but not a big deal since when the second disk
640 * starts IO, the first disk is likely still busy.
642 if (nonrot && opt_iosize > 0 &&
643 mirror->seq_start != MaxSector &&
644 mirror->next_seq_sect > opt_iosize &&
645 mirror->next_seq_sect - opt_iosize >=
647 choose_next_idle = 1;
652 /* If device is idle, use it */
658 if (choose_next_idle)
661 if (min_pending > pending) {
662 min_pending = pending;
663 best_pending_disk = disk;
666 if (dist < best_dist) {
668 best_dist_disk = disk;
673 * If all disks are rotational, choose the closest disk. If any disk is
674 * non-rotational, choose the disk with less pending request even the
675 * disk is rotational, which might/might not be optimal for raids with
676 * mixed ratation/non-rotational disks depending on workload.
678 if (best_disk == -1) {
680 best_disk = best_pending_disk;
682 best_disk = best_dist_disk;
685 if (best_disk >= 0) {
686 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
689 atomic_inc(&rdev->nr_pending);
690 if (test_bit(Faulty, &rdev->flags)) {
691 /* cannot risk returning a device that failed
692 * before we inc'ed nr_pending
694 rdev_dec_pending(rdev, conf->mddev);
697 sectors = best_good_sectors;
699 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
700 conf->mirrors[best_disk].seq_start = this_sector;
702 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
705 *max_sectors = sectors;
710 static int raid1_congested(struct mddev *mddev, int bits)
712 struct r1conf *conf = mddev->private;
715 if ((bits & (1 << WB_async_congested)) &&
716 conf->pending_count >= max_queued_requests)
720 for (i = 0; i < conf->raid_disks * 2; i++) {
721 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
722 if (rdev && !test_bit(Faulty, &rdev->flags)) {
723 struct request_queue *q = bdev_get_queue(rdev->bdev);
727 /* Note the '|| 1' - when read_balance prefers
728 * non-congested targets, it can be removed
730 if ((bits & (1 << WB_async_congested)) || 1)
731 ret |= bdi_congested(&q->backing_dev_info, bits);
733 ret &= bdi_congested(&q->backing_dev_info, bits);
740 static void flush_pending_writes(struct r1conf *conf)
742 /* Any writes that have been queued but are awaiting
743 * bitmap updates get flushed here.
745 spin_lock_irq(&conf->device_lock);
747 if (conf->pending_bio_list.head) {
749 bio = bio_list_get(&conf->pending_bio_list);
750 conf->pending_count = 0;
751 spin_unlock_irq(&conf->device_lock);
752 /* flush any pending bitmap writes to
753 * disk before proceeding w/ I/O */
754 bitmap_unplug(conf->mddev->bitmap);
755 wake_up(&conf->wait_barrier);
757 while (bio) { /* submit pending writes */
758 struct bio *next = bio->bi_next;
760 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
761 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
765 generic_make_request(bio);
769 spin_unlock_irq(&conf->device_lock);
773 * Sometimes we need to suspend IO while we do something else,
774 * either some resync/recovery, or reconfigure the array.
775 * To do this we raise a 'barrier'.
776 * The 'barrier' is a counter that can be raised multiple times
777 * to count how many activities are happening which preclude
779 * We can only raise the barrier if there is no pending IO.
780 * i.e. if nr_pending == 0.
781 * We choose only to raise the barrier if no-one is waiting for the
782 * barrier to go down. This means that as soon as an IO request
783 * is ready, no other operations which require a barrier will start
784 * until the IO request has had a chance.
786 * So: regular IO calls 'wait_barrier'. When that returns there
787 * is no backgroup IO happening, It must arrange to call
788 * allow_barrier when it has finished its IO.
789 * backgroup IO calls must call raise_barrier. Once that returns
790 * there is no normal IO happeing. It must arrange to call
791 * lower_barrier when the particular background IO completes.
793 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
795 spin_lock_irq(&conf->resync_lock);
797 /* Wait until no block IO is waiting */
798 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
801 /* block any new IO from starting */
803 conf->next_resync = sector_nr;
805 /* For these conditions we must wait:
806 * A: while the array is in frozen state
807 * B: while barrier >= RESYNC_DEPTH, meaning resync reach
808 * the max count which allowed.
809 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
810 * next resync will reach to the window which normal bios are
812 * D: while there are any active requests in the current window.
814 wait_event_lock_irq(conf->wait_barrier,
815 !conf->array_frozen &&
816 conf->barrier < RESYNC_DEPTH &&
817 conf->current_window_requests == 0 &&
818 (conf->start_next_window >=
819 conf->next_resync + RESYNC_SECTORS),
823 spin_unlock_irq(&conf->resync_lock);
826 static void lower_barrier(struct r1conf *conf)
829 BUG_ON(conf->barrier <= 0);
830 spin_lock_irqsave(&conf->resync_lock, flags);
833 spin_unlock_irqrestore(&conf->resync_lock, flags);
834 wake_up(&conf->wait_barrier);
837 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
841 if (conf->array_frozen || !bio)
843 else if (conf->barrier && bio_data_dir(bio) == WRITE) {
844 if ((conf->mddev->curr_resync_completed
845 >= bio_end_sector(bio)) ||
846 (conf->next_resync + NEXT_NORMALIO_DISTANCE
847 <= bio->bi_iter.bi_sector))
856 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
860 spin_lock_irq(&conf->resync_lock);
861 if (need_to_wait_for_sync(conf, bio)) {
863 /* Wait for the barrier to drop.
864 * However if there are already pending
865 * requests (preventing the barrier from
866 * rising completely), and the
867 * per-process bio queue isn't empty,
868 * then don't wait, as we need to empty
869 * that queue to allow conf->start_next_window
872 wait_event_lock_irq(conf->wait_barrier,
873 !conf->array_frozen &&
875 ((conf->start_next_window <
876 conf->next_resync + RESYNC_SECTORS) &&
878 !bio_list_empty(current->bio_list))),
883 if (bio && bio_data_dir(bio) == WRITE) {
884 if (bio->bi_iter.bi_sector >= conf->next_resync) {
885 if (conf->start_next_window == MaxSector)
886 conf->start_next_window =
888 NEXT_NORMALIO_DISTANCE;
890 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
891 <= bio->bi_iter.bi_sector)
892 conf->next_window_requests++;
894 conf->current_window_requests++;
895 sector = conf->start_next_window;
900 spin_unlock_irq(&conf->resync_lock);
904 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
909 spin_lock_irqsave(&conf->resync_lock, flags);
911 if (start_next_window) {
912 if (start_next_window == conf->start_next_window) {
913 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
915 conf->next_window_requests--;
917 conf->current_window_requests--;
919 conf->current_window_requests--;
921 if (!conf->current_window_requests) {
922 if (conf->next_window_requests) {
923 conf->current_window_requests =
924 conf->next_window_requests;
925 conf->next_window_requests = 0;
926 conf->start_next_window +=
927 NEXT_NORMALIO_DISTANCE;
929 conf->start_next_window = MaxSector;
932 spin_unlock_irqrestore(&conf->resync_lock, flags);
933 wake_up(&conf->wait_barrier);
936 static void freeze_array(struct r1conf *conf, int extra)
938 /* stop syncio and normal IO and wait for everything to
940 * We wait until nr_pending match nr_queued+extra
941 * This is called in the context of one normal IO request
942 * that has failed. Thus any sync request that might be pending
943 * will be blocked by nr_pending, and we need to wait for
944 * pending IO requests to complete or be queued for re-try.
945 * Thus the number queued (nr_queued) plus this request (extra)
946 * must match the number of pending IOs (nr_pending) before
949 spin_lock_irq(&conf->resync_lock);
950 conf->array_frozen = 1;
951 wait_event_lock_irq_cmd(conf->wait_barrier,
952 conf->nr_pending == conf->nr_queued+extra,
954 flush_pending_writes(conf));
955 spin_unlock_irq(&conf->resync_lock);
957 static void unfreeze_array(struct r1conf *conf)
959 /* reverse the effect of the freeze */
960 spin_lock_irq(&conf->resync_lock);
961 conf->array_frozen = 0;
962 wake_up(&conf->wait_barrier);
963 spin_unlock_irq(&conf->resync_lock);
966 /* duplicate the data pages for behind I/O
968 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
971 struct bio_vec *bvec;
972 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
974 if (unlikely(!bvecs))
977 bio_for_each_segment_all(bvec, bio, i) {
979 bvecs[i].bv_page = alloc_page(GFP_NOIO);
980 if (unlikely(!bvecs[i].bv_page))
982 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
983 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
984 kunmap(bvecs[i].bv_page);
985 kunmap(bvec->bv_page);
987 r1_bio->behind_bvecs = bvecs;
988 r1_bio->behind_page_count = bio->bi_vcnt;
989 set_bit(R1BIO_BehindIO, &r1_bio->state);
993 for (i = 0; i < bio->bi_vcnt; i++)
994 if (bvecs[i].bv_page)
995 put_page(bvecs[i].bv_page);
997 pr_debug("%dB behind alloc failed, doing sync I/O\n",
998 bio->bi_iter.bi_size);
1001 struct raid1_plug_cb {
1002 struct blk_plug_cb cb;
1003 struct bio_list pending;
1007 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1009 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1011 struct mddev *mddev = plug->cb.data;
1012 struct r1conf *conf = mddev->private;
1015 if (from_schedule || current->bio_list) {
1016 spin_lock_irq(&conf->device_lock);
1017 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1018 conf->pending_count += plug->pending_cnt;
1019 spin_unlock_irq(&conf->device_lock);
1020 wake_up(&conf->wait_barrier);
1021 md_wakeup_thread(mddev->thread);
1026 /* we aren't scheduling, so we can do the write-out directly. */
1027 bio = bio_list_get(&plug->pending);
1028 bitmap_unplug(mddev->bitmap);
1029 wake_up(&conf->wait_barrier);
1031 while (bio) { /* submit pending writes */
1032 struct bio *next = bio->bi_next;
1033 bio->bi_next = NULL;
1034 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1035 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1036 /* Just ignore it */
1039 generic_make_request(bio);
1045 static void make_request(struct mddev *mddev, struct bio * bio)
1047 struct r1conf *conf = mddev->private;
1048 struct raid1_info *mirror;
1049 struct r1bio *r1_bio;
1050 struct bio *read_bio;
1052 struct bitmap *bitmap;
1053 unsigned long flags;
1054 const int rw = bio_data_dir(bio);
1055 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1056 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1057 const unsigned long do_discard = (bio->bi_rw
1058 & (REQ_DISCARD | REQ_SECURE));
1059 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1060 struct md_rdev *blocked_rdev;
1061 struct blk_plug_cb *cb;
1062 struct raid1_plug_cb *plug = NULL;
1064 int sectors_handled;
1066 sector_t start_next_window;
1069 * Register the new request and wait if the reconstruction
1070 * thread has put up a bar for new requests.
1071 * Continue immediately if no resync is active currently.
1074 md_write_start(mddev, bio); /* wait on superblock update early */
1076 if (bio_data_dir(bio) == WRITE &&
1077 ((bio_end_sector(bio) > mddev->suspend_lo &&
1078 bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1079 (mddev_is_clustered(mddev) &&
1080 md_cluster_ops->area_resyncing(mddev, WRITE,
1081 bio->bi_iter.bi_sector, bio_end_sector(bio))))) {
1082 /* As the suspend_* range is controlled by
1083 * userspace, we want an interruptible
1088 flush_signals(current);
1089 prepare_to_wait(&conf->wait_barrier,
1090 &w, TASK_INTERRUPTIBLE);
1091 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1092 bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1093 (mddev_is_clustered(mddev) &&
1094 !md_cluster_ops->area_resyncing(mddev, WRITE,
1095 bio->bi_iter.bi_sector, bio_end_sector(bio))))
1099 finish_wait(&conf->wait_barrier, &w);
1102 start_next_window = wait_barrier(conf, bio);
1104 bitmap = mddev->bitmap;
1107 * make_request() can abort the operation when READA is being
1108 * used and no empty request is available.
1111 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1113 r1_bio->master_bio = bio;
1114 r1_bio->sectors = bio_sectors(bio);
1116 r1_bio->mddev = mddev;
1117 r1_bio->sector = bio->bi_iter.bi_sector;
1119 /* We might need to issue multiple reads to different
1120 * devices if there are bad blocks around, so we keep
1121 * track of the number of reads in bio->bi_phys_segments.
1122 * If this is 0, there is only one r1_bio and no locking
1123 * will be needed when requests complete. If it is
1124 * non-zero, then it is the number of not-completed requests.
1126 bio->bi_phys_segments = 0;
1127 bio_clear_flag(bio, BIO_SEG_VALID);
1131 * read balancing logic:
1136 rdisk = read_balance(conf, r1_bio, &max_sectors);
1139 /* couldn't find anywhere to read from */
1140 raid_end_bio_io(r1_bio);
1143 mirror = conf->mirrors + rdisk;
1145 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1147 /* Reading from a write-mostly device must
1148 * take care not to over-take any writes
1151 wait_event(bitmap->behind_wait,
1152 atomic_read(&bitmap->behind_writes) == 0);
1154 r1_bio->read_disk = rdisk;
1155 r1_bio->start_next_window = 0;
1157 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1158 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1161 r1_bio->bios[rdisk] = read_bio;
1163 read_bio->bi_iter.bi_sector = r1_bio->sector +
1164 mirror->rdev->data_offset;
1165 read_bio->bi_bdev = mirror->rdev->bdev;
1166 read_bio->bi_end_io = raid1_end_read_request;
1167 read_bio->bi_rw = READ | do_sync;
1168 read_bio->bi_private = r1_bio;
1170 if (max_sectors < r1_bio->sectors) {
1171 /* could not read all from this device, so we will
1172 * need another r1_bio.
1175 sectors_handled = (r1_bio->sector + max_sectors
1176 - bio->bi_iter.bi_sector);
1177 r1_bio->sectors = max_sectors;
1178 spin_lock_irq(&conf->device_lock);
1179 if (bio->bi_phys_segments == 0)
1180 bio->bi_phys_segments = 2;
1182 bio->bi_phys_segments++;
1183 spin_unlock_irq(&conf->device_lock);
1184 /* Cannot call generic_make_request directly
1185 * as that will be queued in __make_request
1186 * and subsequent mempool_alloc might block waiting
1187 * for it. So hand bio over to raid1d.
1189 reschedule_retry(r1_bio);
1191 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1193 r1_bio->master_bio = bio;
1194 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1196 r1_bio->mddev = mddev;
1197 r1_bio->sector = bio->bi_iter.bi_sector +
1201 generic_make_request(read_bio);
1208 if (conf->pending_count >= max_queued_requests) {
1209 md_wakeup_thread(mddev->thread);
1210 wait_event(conf->wait_barrier,
1211 conf->pending_count < max_queued_requests);
1213 /* first select target devices under rcu_lock and
1214 * inc refcount on their rdev. Record them by setting
1216 * If there are known/acknowledged bad blocks on any device on
1217 * which we have seen a write error, we want to avoid writing those
1219 * This potentially requires several writes to write around
1220 * the bad blocks. Each set of writes gets it's own r1bio
1221 * with a set of bios attached.
1224 disks = conf->raid_disks * 2;
1226 r1_bio->start_next_window = start_next_window;
1227 blocked_rdev = NULL;
1229 max_sectors = r1_bio->sectors;
1230 for (i = 0; i < disks; i++) {
1231 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1232 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1233 atomic_inc(&rdev->nr_pending);
1234 blocked_rdev = rdev;
1237 r1_bio->bios[i] = NULL;
1238 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1239 if (i < conf->raid_disks)
1240 set_bit(R1BIO_Degraded, &r1_bio->state);
1244 atomic_inc(&rdev->nr_pending);
1245 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1250 is_bad = is_badblock(rdev, r1_bio->sector,
1252 &first_bad, &bad_sectors);
1254 /* mustn't write here until the bad block is
1256 set_bit(BlockedBadBlocks, &rdev->flags);
1257 blocked_rdev = rdev;
1260 if (is_bad && first_bad <= r1_bio->sector) {
1261 /* Cannot write here at all */
1262 bad_sectors -= (r1_bio->sector - first_bad);
1263 if (bad_sectors < max_sectors)
1264 /* mustn't write more than bad_sectors
1265 * to other devices yet
1267 max_sectors = bad_sectors;
1268 rdev_dec_pending(rdev, mddev);
1269 /* We don't set R1BIO_Degraded as that
1270 * only applies if the disk is
1271 * missing, so it might be re-added,
1272 * and we want to know to recover this
1274 * In this case the device is here,
1275 * and the fact that this chunk is not
1276 * in-sync is recorded in the bad
1282 int good_sectors = first_bad - r1_bio->sector;
1283 if (good_sectors < max_sectors)
1284 max_sectors = good_sectors;
1287 r1_bio->bios[i] = bio;
1291 if (unlikely(blocked_rdev)) {
1292 /* Wait for this device to become unblocked */
1294 sector_t old = start_next_window;
1296 for (j = 0; j < i; j++)
1297 if (r1_bio->bios[j])
1298 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1300 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1301 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1302 start_next_window = wait_barrier(conf, bio);
1304 * We must make sure the multi r1bios of bio have
1305 * the same value of bi_phys_segments
1307 if (bio->bi_phys_segments && old &&
1308 old != start_next_window)
1309 /* Wait for the former r1bio(s) to complete */
1310 wait_event(conf->wait_barrier,
1311 bio->bi_phys_segments == 1);
1315 if (max_sectors < r1_bio->sectors) {
1316 /* We are splitting this write into multiple parts, so
1317 * we need to prepare for allocating another r1_bio.
1319 r1_bio->sectors = max_sectors;
1320 spin_lock_irq(&conf->device_lock);
1321 if (bio->bi_phys_segments == 0)
1322 bio->bi_phys_segments = 2;
1324 bio->bi_phys_segments++;
1325 spin_unlock_irq(&conf->device_lock);
1327 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1329 atomic_set(&r1_bio->remaining, 1);
1330 atomic_set(&r1_bio->behind_remaining, 0);
1333 for (i = 0; i < disks; i++) {
1335 if (!r1_bio->bios[i])
1338 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1339 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1343 * Not if there are too many, or cannot
1344 * allocate memory, or a reader on WriteMostly
1345 * is waiting for behind writes to flush */
1347 (atomic_read(&bitmap->behind_writes)
1348 < mddev->bitmap_info.max_write_behind) &&
1349 !waitqueue_active(&bitmap->behind_wait))
1350 alloc_behind_pages(mbio, r1_bio);
1352 bitmap_startwrite(bitmap, r1_bio->sector,
1354 test_bit(R1BIO_BehindIO,
1358 if (r1_bio->behind_bvecs) {
1359 struct bio_vec *bvec;
1363 * We trimmed the bio, so _all is legit
1365 bio_for_each_segment_all(bvec, mbio, j)
1366 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1367 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1368 atomic_inc(&r1_bio->behind_remaining);
1371 r1_bio->bios[i] = mbio;
1373 mbio->bi_iter.bi_sector = (r1_bio->sector +
1374 conf->mirrors[i].rdev->data_offset);
1375 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1376 mbio->bi_end_io = raid1_end_write_request;
1378 WRITE | do_flush_fua | do_sync | do_discard | do_same;
1379 mbio->bi_private = r1_bio;
1381 atomic_inc(&r1_bio->remaining);
1383 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1385 plug = container_of(cb, struct raid1_plug_cb, cb);
1388 spin_lock_irqsave(&conf->device_lock, flags);
1390 bio_list_add(&plug->pending, mbio);
1391 plug->pending_cnt++;
1393 bio_list_add(&conf->pending_bio_list, mbio);
1394 conf->pending_count++;
1396 spin_unlock_irqrestore(&conf->device_lock, flags);
1398 md_wakeup_thread(mddev->thread);
1400 /* Mustn't call r1_bio_write_done before this next test,
1401 * as it could result in the bio being freed.
1403 if (sectors_handled < bio_sectors(bio)) {
1404 r1_bio_write_done(r1_bio);
1405 /* We need another r1_bio. It has already been counted
1406 * in bio->bi_phys_segments
1408 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1409 r1_bio->master_bio = bio;
1410 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1412 r1_bio->mddev = mddev;
1413 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1417 r1_bio_write_done(r1_bio);
1419 /* In case raid1d snuck in to freeze_array */
1420 wake_up(&conf->wait_barrier);
1423 static void status(struct seq_file *seq, struct mddev *mddev)
1425 struct r1conf *conf = mddev->private;
1428 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1429 conf->raid_disks - mddev->degraded);
1431 for (i = 0; i < conf->raid_disks; i++) {
1432 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1433 seq_printf(seq, "%s",
1434 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1437 seq_printf(seq, "]");
1440 static void error(struct mddev *mddev, struct md_rdev *rdev)
1442 char b[BDEVNAME_SIZE];
1443 struct r1conf *conf = mddev->private;
1444 unsigned long flags;
1447 * If it is not operational, then we have already marked it as dead
1448 * else if it is the last working disks, ignore the error, let the
1449 * next level up know.
1450 * else mark the drive as failed
1452 if (test_bit(In_sync, &rdev->flags)
1453 && (conf->raid_disks - mddev->degraded) == 1) {
1455 * Don't fail the drive, act as though we were just a
1456 * normal single drive.
1457 * However don't try a recovery from this drive as
1458 * it is very likely to fail.
1460 conf->recovery_disabled = mddev->recovery_disabled;
1463 set_bit(Blocked, &rdev->flags);
1464 spin_lock_irqsave(&conf->device_lock, flags);
1465 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1467 set_bit(Faulty, &rdev->flags);
1469 set_bit(Faulty, &rdev->flags);
1470 spin_unlock_irqrestore(&conf->device_lock, flags);
1472 * if recovery is running, make sure it aborts.
1474 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1475 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1476 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1478 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1479 "md/raid1:%s: Operation continuing on %d devices.\n",
1480 mdname(mddev), bdevname(rdev->bdev, b),
1481 mdname(mddev), conf->raid_disks - mddev->degraded);
1484 static void print_conf(struct r1conf *conf)
1488 printk(KERN_DEBUG "RAID1 conf printout:\n");
1490 printk(KERN_DEBUG "(!conf)\n");
1493 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1497 for (i = 0; i < conf->raid_disks; i++) {
1498 char b[BDEVNAME_SIZE];
1499 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1501 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1502 i, !test_bit(In_sync, &rdev->flags),
1503 !test_bit(Faulty, &rdev->flags),
1504 bdevname(rdev->bdev,b));
1509 static void close_sync(struct r1conf *conf)
1511 wait_barrier(conf, NULL);
1512 allow_barrier(conf, 0, 0);
1514 mempool_destroy(conf->r1buf_pool);
1515 conf->r1buf_pool = NULL;
1517 spin_lock_irq(&conf->resync_lock);
1518 conf->next_resync = MaxSector - 2 * NEXT_NORMALIO_DISTANCE;
1519 conf->start_next_window = MaxSector;
1520 conf->current_window_requests +=
1521 conf->next_window_requests;
1522 conf->next_window_requests = 0;
1523 spin_unlock_irq(&conf->resync_lock);
1526 static int raid1_spare_active(struct mddev *mddev)
1529 struct r1conf *conf = mddev->private;
1531 unsigned long flags;
1534 * Find all failed disks within the RAID1 configuration
1535 * and mark them readable.
1536 * Called under mddev lock, so rcu protection not needed.
1537 * device_lock used to avoid races with raid1_end_read_request
1538 * which expects 'In_sync' flags and ->degraded to be consistent.
1540 spin_lock_irqsave(&conf->device_lock, flags);
1541 for (i = 0; i < conf->raid_disks; i++) {
1542 struct md_rdev *rdev = conf->mirrors[i].rdev;
1543 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1545 && !test_bit(Candidate, &repl->flags)
1546 && repl->recovery_offset == MaxSector
1547 && !test_bit(Faulty, &repl->flags)
1548 && !test_and_set_bit(In_sync, &repl->flags)) {
1549 /* replacement has just become active */
1551 !test_and_clear_bit(In_sync, &rdev->flags))
1554 /* Replaced device not technically
1555 * faulty, but we need to be sure
1556 * it gets removed and never re-added
1558 set_bit(Faulty, &rdev->flags);
1559 sysfs_notify_dirent_safe(
1564 && rdev->recovery_offset == MaxSector
1565 && !test_bit(Faulty, &rdev->flags)
1566 && !test_and_set_bit(In_sync, &rdev->flags)) {
1568 sysfs_notify_dirent_safe(rdev->sysfs_state);
1571 mddev->degraded -= count;
1572 spin_unlock_irqrestore(&conf->device_lock, flags);
1578 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1580 struct r1conf *conf = mddev->private;
1583 struct raid1_info *p;
1585 int last = conf->raid_disks - 1;
1587 if (mddev->recovery_disabled == conf->recovery_disabled)
1590 if (rdev->raid_disk >= 0)
1591 first = last = rdev->raid_disk;
1593 for (mirror = first; mirror <= last; mirror++) {
1594 p = conf->mirrors+mirror;
1598 disk_stack_limits(mddev->gendisk, rdev->bdev,
1599 rdev->data_offset << 9);
1601 p->head_position = 0;
1602 rdev->raid_disk = mirror;
1604 /* As all devices are equivalent, we don't need a full recovery
1605 * if this was recently any drive of the array
1607 if (rdev->saved_raid_disk < 0)
1609 rcu_assign_pointer(p->rdev, rdev);
1612 if (test_bit(WantReplacement, &p->rdev->flags) &&
1613 p[conf->raid_disks].rdev == NULL) {
1614 /* Add this device as a replacement */
1615 clear_bit(In_sync, &rdev->flags);
1616 set_bit(Replacement, &rdev->flags);
1617 rdev->raid_disk = mirror;
1620 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1624 mddev_suspend(mddev);
1625 md_integrity_add_rdev(rdev, mddev);
1626 mddev_resume(mddev);
1627 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1628 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1633 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1635 struct r1conf *conf = mddev->private;
1637 int number = rdev->raid_disk;
1638 struct raid1_info *p = conf->mirrors + number;
1640 if (rdev != p->rdev)
1641 p = conf->mirrors + conf->raid_disks + number;
1644 if (rdev == p->rdev) {
1645 if (test_bit(In_sync, &rdev->flags) ||
1646 atomic_read(&rdev->nr_pending)) {
1650 /* Only remove non-faulty devices if recovery
1653 if (!test_bit(Faulty, &rdev->flags) &&
1654 mddev->recovery_disabled != conf->recovery_disabled &&
1655 mddev->degraded < conf->raid_disks) {
1661 if (atomic_read(&rdev->nr_pending)) {
1662 /* lost the race, try later */
1666 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1667 /* We just removed a device that is being replaced.
1668 * Move down the replacement. We drain all IO before
1669 * doing this to avoid confusion.
1671 struct md_rdev *repl =
1672 conf->mirrors[conf->raid_disks + number].rdev;
1673 freeze_array(conf, 0);
1674 clear_bit(Replacement, &repl->flags);
1676 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1677 unfreeze_array(conf);
1678 clear_bit(WantReplacement, &rdev->flags);
1680 clear_bit(WantReplacement, &rdev->flags);
1681 err = md_integrity_register(mddev);
1689 static void end_sync_read(struct bio *bio)
1691 struct r1bio *r1_bio = bio->bi_private;
1693 update_head_pos(r1_bio->read_disk, r1_bio);
1696 * we have read a block, now it needs to be re-written,
1697 * or re-read if the read failed.
1698 * We don't do much here, just schedule handling by raid1d
1701 set_bit(R1BIO_Uptodate, &r1_bio->state);
1703 if (atomic_dec_and_test(&r1_bio->remaining))
1704 reschedule_retry(r1_bio);
1707 static void end_sync_write(struct bio *bio)
1709 int uptodate = !bio->bi_error;
1710 struct r1bio *r1_bio = bio->bi_private;
1711 struct mddev *mddev = r1_bio->mddev;
1712 struct r1conf *conf = mddev->private;
1717 mirror = find_bio_disk(r1_bio, bio);
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,
1731 &conf->mirrors[mirror].rdev->flags);
1732 if (!test_and_set_bit(WantReplacement,
1733 &conf->mirrors[mirror].rdev->flags))
1734 set_bit(MD_RECOVERY_NEEDED, &
1736 set_bit(R1BIO_WriteError, &r1_bio->state);
1737 } else if (is_badblock(conf->mirrors[mirror].rdev,
1740 &first_bad, &bad_sectors) &&
1741 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1744 &first_bad, &bad_sectors)
1746 set_bit(R1BIO_MadeGood, &r1_bio->state);
1748 if (atomic_dec_and_test(&r1_bio->remaining)) {
1749 int s = r1_bio->sectors;
1750 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1751 test_bit(R1BIO_WriteError, &r1_bio->state))
1752 reschedule_retry(r1_bio);
1755 md_done_sync(mddev, s, uptodate);
1760 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1761 int sectors, struct page *page, int rw)
1763 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1767 set_bit(WriteErrorSeen, &rdev->flags);
1768 if (!test_and_set_bit(WantReplacement,
1770 set_bit(MD_RECOVERY_NEEDED, &
1771 rdev->mddev->recovery);
1773 /* need to record an error - either for the block or the device */
1774 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1775 md_error(rdev->mddev, rdev);
1779 static int fix_sync_read_error(struct r1bio *r1_bio)
1781 /* Try some synchronous reads of other devices to get
1782 * good data, much like with normal read errors. Only
1783 * read into the pages we already have so we don't
1784 * need to re-issue the read request.
1785 * We don't need to freeze the array, because being in an
1786 * active sync request, there is no normal IO, and
1787 * no overlapping syncs.
1788 * We don't need to check is_badblock() again as we
1789 * made sure that anything with a bad block in range
1790 * will have bi_end_io clear.
1792 struct mddev *mddev = r1_bio->mddev;
1793 struct r1conf *conf = mddev->private;
1794 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1795 sector_t sect = r1_bio->sector;
1796 int sectors = r1_bio->sectors;
1801 int d = r1_bio->read_disk;
1803 struct md_rdev *rdev;
1806 if (s > (PAGE_SIZE>>9))
1809 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1810 /* No rcu protection needed here devices
1811 * can only be removed when no resync is
1812 * active, and resync is currently active
1814 rdev = conf->mirrors[d].rdev;
1815 if (sync_page_io(rdev, sect, s<<9,
1816 bio->bi_io_vec[idx].bv_page,
1823 if (d == conf->raid_disks * 2)
1825 } while (!success && d != r1_bio->read_disk);
1828 char b[BDEVNAME_SIZE];
1830 /* Cannot read from anywhere, this block is lost.
1831 * Record a bad block on each device. If that doesn't
1832 * work just disable and interrupt the recovery.
1833 * Don't fail devices as that won't really help.
1835 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1836 " for block %llu\n",
1838 bdevname(bio->bi_bdev, b),
1839 (unsigned long long)r1_bio->sector);
1840 for (d = 0; d < conf->raid_disks * 2; d++) {
1841 rdev = conf->mirrors[d].rdev;
1842 if (!rdev || test_bit(Faulty, &rdev->flags))
1844 if (!rdev_set_badblocks(rdev, sect, s, 0))
1848 conf->recovery_disabled =
1849 mddev->recovery_disabled;
1850 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1851 md_done_sync(mddev, r1_bio->sectors, 0);
1863 /* write it back and re-read */
1864 while (d != r1_bio->read_disk) {
1866 d = conf->raid_disks * 2;
1868 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1870 rdev = conf->mirrors[d].rdev;
1871 if (r1_sync_page_io(rdev, sect, s,
1872 bio->bi_io_vec[idx].bv_page,
1874 r1_bio->bios[d]->bi_end_io = NULL;
1875 rdev_dec_pending(rdev, mddev);
1879 while (d != r1_bio->read_disk) {
1881 d = conf->raid_disks * 2;
1883 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1885 rdev = conf->mirrors[d].rdev;
1886 if (r1_sync_page_io(rdev, sect, s,
1887 bio->bi_io_vec[idx].bv_page,
1889 atomic_add(s, &rdev->corrected_errors);
1895 set_bit(R1BIO_Uptodate, &r1_bio->state);
1900 static void process_checks(struct r1bio *r1_bio)
1902 /* We have read all readable devices. If we haven't
1903 * got the block, then there is no hope left.
1904 * If we have, then we want to do a comparison
1905 * and skip the write if everything is the same.
1906 * If any blocks failed to read, then we need to
1907 * attempt an over-write
1909 struct mddev *mddev = r1_bio->mddev;
1910 struct r1conf *conf = mddev->private;
1915 /* Fix variable parts of all bios */
1916 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1917 for (i = 0; i < conf->raid_disks * 2; i++) {
1921 struct bio *b = r1_bio->bios[i];
1922 if (b->bi_end_io != end_sync_read)
1924 /* fixup the bio for reuse, but preserve errno */
1925 error = b->bi_error;
1927 b->bi_error = error;
1929 b->bi_iter.bi_size = r1_bio->sectors << 9;
1930 b->bi_iter.bi_sector = r1_bio->sector +
1931 conf->mirrors[i].rdev->data_offset;
1932 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1933 b->bi_end_io = end_sync_read;
1934 b->bi_private = r1_bio;
1936 size = b->bi_iter.bi_size;
1937 for (j = 0; j < vcnt ; j++) {
1939 bi = &b->bi_io_vec[j];
1941 if (size > PAGE_SIZE)
1942 bi->bv_len = PAGE_SIZE;
1948 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1949 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1950 !r1_bio->bios[primary]->bi_error) {
1951 r1_bio->bios[primary]->bi_end_io = NULL;
1952 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1955 r1_bio->read_disk = primary;
1956 for (i = 0; i < conf->raid_disks * 2; i++) {
1958 struct bio *pbio = r1_bio->bios[primary];
1959 struct bio *sbio = r1_bio->bios[i];
1960 int error = sbio->bi_error;
1962 if (sbio->bi_end_io != end_sync_read)
1964 /* Now we can 'fixup' the error value */
1968 for (j = vcnt; j-- ; ) {
1970 p = pbio->bi_io_vec[j].bv_page;
1971 s = sbio->bi_io_vec[j].bv_page;
1972 if (memcmp(page_address(p),
1974 sbio->bi_io_vec[j].bv_len))
1980 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1981 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1983 /* No need to write to this device. */
1984 sbio->bi_end_io = NULL;
1985 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1989 bio_copy_data(sbio, pbio);
1993 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1995 struct r1conf *conf = mddev->private;
1997 int disks = conf->raid_disks * 2;
1998 struct bio *bio, *wbio;
2000 bio = r1_bio->bios[r1_bio->read_disk];
2002 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2003 /* ouch - failed to read all of that. */
2004 if (!fix_sync_read_error(r1_bio))
2007 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2008 process_checks(r1_bio);
2013 atomic_set(&r1_bio->remaining, 1);
2014 for (i = 0; i < disks ; i++) {
2015 wbio = r1_bio->bios[i];
2016 if (wbio->bi_end_io == NULL ||
2017 (wbio->bi_end_io == end_sync_read &&
2018 (i == r1_bio->read_disk ||
2019 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2022 wbio->bi_rw = WRITE;
2023 wbio->bi_end_io = end_sync_write;
2024 atomic_inc(&r1_bio->remaining);
2025 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2027 generic_make_request(wbio);
2030 if (atomic_dec_and_test(&r1_bio->remaining)) {
2031 /* if we're here, all write(s) have completed, so clean up */
2032 int s = r1_bio->sectors;
2033 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2034 test_bit(R1BIO_WriteError, &r1_bio->state))
2035 reschedule_retry(r1_bio);
2038 md_done_sync(mddev, s, 1);
2044 * This is a kernel thread which:
2046 * 1. Retries failed read operations on working mirrors.
2047 * 2. Updates the raid superblock when problems encounter.
2048 * 3. Performs writes following reads for array synchronising.
2051 static void fix_read_error(struct r1conf *conf, int read_disk,
2052 sector_t sect, int sectors)
2054 struct mddev *mddev = conf->mddev;
2060 struct md_rdev *rdev;
2062 if (s > (PAGE_SIZE>>9))
2066 /* Note: no rcu protection needed here
2067 * as this is synchronous in the raid1d thread
2068 * which is the thread that might remove
2069 * a device. If raid1d ever becomes multi-threaded....
2074 rdev = conf->mirrors[d].rdev;
2076 (test_bit(In_sync, &rdev->flags) ||
2077 (!test_bit(Faulty, &rdev->flags) &&
2078 rdev->recovery_offset >= sect + s)) &&
2079 is_badblock(rdev, sect, s,
2080 &first_bad, &bad_sectors) == 0 &&
2081 sync_page_io(rdev, sect, s<<9,
2082 conf->tmppage, READ, false))
2086 if (d == conf->raid_disks * 2)
2089 } while (!success && d != read_disk);
2092 /* Cannot read from anywhere - mark it bad */
2093 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2094 if (!rdev_set_badblocks(rdev, sect, s, 0))
2095 md_error(mddev, rdev);
2098 /* write it back and re-read */
2100 while (d != read_disk) {
2102 d = conf->raid_disks * 2;
2104 rdev = conf->mirrors[d].rdev;
2106 !test_bit(Faulty, &rdev->flags))
2107 r1_sync_page_io(rdev, sect, s,
2108 conf->tmppage, WRITE);
2111 while (d != read_disk) {
2112 char b[BDEVNAME_SIZE];
2114 d = conf->raid_disks * 2;
2116 rdev = conf->mirrors[d].rdev;
2118 !test_bit(Faulty, &rdev->flags)) {
2119 if (r1_sync_page_io(rdev, sect, s,
2120 conf->tmppage, READ)) {
2121 atomic_add(s, &rdev->corrected_errors);
2123 "md/raid1:%s: read error corrected "
2124 "(%d sectors at %llu on %s)\n",
2126 (unsigned long long)(sect +
2128 bdevname(rdev->bdev, b));
2137 static int narrow_write_error(struct r1bio *r1_bio, int i)
2139 struct mddev *mddev = r1_bio->mddev;
2140 struct r1conf *conf = mddev->private;
2141 struct md_rdev *rdev = conf->mirrors[i].rdev;
2143 /* bio has the data to be written to device 'i' where
2144 * we just recently had a write error.
2145 * We repeatedly clone the bio and trim down to one block,
2146 * then try the write. Where the write fails we record
2148 * It is conceivable that the bio doesn't exactly align with
2149 * blocks. We must handle this somehow.
2151 * We currently own a reference on the rdev.
2157 int sect_to_write = r1_bio->sectors;
2160 if (rdev->badblocks.shift < 0)
2163 block_sectors = roundup(1 << rdev->badblocks.shift,
2164 bdev_logical_block_size(rdev->bdev) >> 9);
2165 sector = r1_bio->sector;
2166 sectors = ((sector + block_sectors)
2167 & ~(sector_t)(block_sectors - 1))
2170 while (sect_to_write) {
2172 if (sectors > sect_to_write)
2173 sectors = sect_to_write;
2174 /* Write at 'sector' for 'sectors'*/
2176 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2177 unsigned vcnt = r1_bio->behind_page_count;
2178 struct bio_vec *vec = r1_bio->behind_bvecs;
2180 while (!vec->bv_page) {
2185 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2186 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2188 wbio->bi_vcnt = vcnt;
2190 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2193 wbio->bi_rw = WRITE;
2194 wbio->bi_iter.bi_sector = r1_bio->sector;
2195 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2197 bio_trim(wbio, sector - r1_bio->sector, sectors);
2198 wbio->bi_iter.bi_sector += rdev->data_offset;
2199 wbio->bi_bdev = rdev->bdev;
2200 if (submit_bio_wait(WRITE, wbio) < 0)
2202 ok = rdev_set_badblocks(rdev, sector,
2207 sect_to_write -= sectors;
2209 sectors = block_sectors;
2214 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2217 int s = r1_bio->sectors;
2218 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2219 struct md_rdev *rdev = conf->mirrors[m].rdev;
2220 struct bio *bio = r1_bio->bios[m];
2221 if (bio->bi_end_io == NULL)
2223 if (!bio->bi_error &&
2224 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2225 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2227 if (bio->bi_error &&
2228 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2229 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2230 md_error(conf->mddev, rdev);
2234 md_done_sync(conf->mddev, s, 1);
2237 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2241 for (m = 0; m < conf->raid_disks * 2 ; m++)
2242 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2243 struct md_rdev *rdev = conf->mirrors[m].rdev;
2244 rdev_clear_badblocks(rdev,
2246 r1_bio->sectors, 0);
2247 rdev_dec_pending(rdev, conf->mddev);
2248 } else if (r1_bio->bios[m] != NULL) {
2249 /* This drive got a write error. We need to
2250 * narrow down and record precise write
2254 if (!narrow_write_error(r1_bio, m)) {
2255 md_error(conf->mddev,
2256 conf->mirrors[m].rdev);
2257 /* an I/O failed, we can't clear the bitmap */
2258 set_bit(R1BIO_Degraded, &r1_bio->state);
2260 rdev_dec_pending(conf->mirrors[m].rdev,
2264 spin_lock_irq(&conf->device_lock);
2265 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2266 spin_unlock_irq(&conf->device_lock);
2267 md_wakeup_thread(conf->mddev->thread);
2269 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2270 close_write(r1_bio);
2271 raid_end_bio_io(r1_bio);
2275 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2279 struct mddev *mddev = conf->mddev;
2281 char b[BDEVNAME_SIZE];
2282 struct md_rdev *rdev;
2284 clear_bit(R1BIO_ReadError, &r1_bio->state);
2285 /* we got a read error. Maybe the drive is bad. Maybe just
2286 * the block and we can fix it.
2287 * We freeze all other IO, and try reading the block from
2288 * other devices. When we find one, we re-write
2289 * and check it that fixes the read error.
2290 * This is all done synchronously while the array is
2293 if (mddev->ro == 0) {
2294 freeze_array(conf, 1);
2295 fix_read_error(conf, r1_bio->read_disk,
2296 r1_bio->sector, r1_bio->sectors);
2297 unfreeze_array(conf);
2299 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2300 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2302 bio = r1_bio->bios[r1_bio->read_disk];
2303 bdevname(bio->bi_bdev, b);
2305 disk = read_balance(conf, r1_bio, &max_sectors);
2307 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2308 " read error for block %llu\n",
2309 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2310 raid_end_bio_io(r1_bio);
2312 const unsigned long do_sync
2313 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2315 r1_bio->bios[r1_bio->read_disk] =
2316 mddev->ro ? IO_BLOCKED : NULL;
2319 r1_bio->read_disk = disk;
2320 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2321 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2323 r1_bio->bios[r1_bio->read_disk] = bio;
2324 rdev = conf->mirrors[disk].rdev;
2325 printk_ratelimited(KERN_ERR
2326 "md/raid1:%s: redirecting sector %llu"
2327 " to other mirror: %s\n",
2329 (unsigned long long)r1_bio->sector,
2330 bdevname(rdev->bdev, b));
2331 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2332 bio->bi_bdev = rdev->bdev;
2333 bio->bi_end_io = raid1_end_read_request;
2334 bio->bi_rw = READ | do_sync;
2335 bio->bi_private = r1_bio;
2336 if (max_sectors < r1_bio->sectors) {
2337 /* Drat - have to split this up more */
2338 struct bio *mbio = r1_bio->master_bio;
2339 int sectors_handled = (r1_bio->sector + max_sectors
2340 - mbio->bi_iter.bi_sector);
2341 r1_bio->sectors = max_sectors;
2342 spin_lock_irq(&conf->device_lock);
2343 if (mbio->bi_phys_segments == 0)
2344 mbio->bi_phys_segments = 2;
2346 mbio->bi_phys_segments++;
2347 spin_unlock_irq(&conf->device_lock);
2348 generic_make_request(bio);
2351 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2353 r1_bio->master_bio = mbio;
2354 r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2356 set_bit(R1BIO_ReadError, &r1_bio->state);
2357 r1_bio->mddev = mddev;
2358 r1_bio->sector = mbio->bi_iter.bi_sector +
2363 generic_make_request(bio);
2367 static void raid1d(struct md_thread *thread)
2369 struct mddev *mddev = thread->mddev;
2370 struct r1bio *r1_bio;
2371 unsigned long flags;
2372 struct r1conf *conf = mddev->private;
2373 struct list_head *head = &conf->retry_list;
2374 struct blk_plug plug;
2376 md_check_recovery(mddev);
2378 if (!list_empty_careful(&conf->bio_end_io_list) &&
2379 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2381 spin_lock_irqsave(&conf->device_lock, flags);
2382 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2383 list_add(&tmp, &conf->bio_end_io_list);
2384 list_del_init(&conf->bio_end_io_list);
2386 spin_unlock_irqrestore(&conf->device_lock, flags);
2387 while (!list_empty(&tmp)) {
2388 r1_bio = list_first_entry(&tmp, struct r1bio,
2390 list_del(&r1_bio->retry_list);
2391 if (mddev->degraded)
2392 set_bit(R1BIO_Degraded, &r1_bio->state);
2393 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2394 close_write(r1_bio);
2395 raid_end_bio_io(r1_bio);
2399 blk_start_plug(&plug);
2402 flush_pending_writes(conf);
2404 spin_lock_irqsave(&conf->device_lock, flags);
2405 if (list_empty(head)) {
2406 spin_unlock_irqrestore(&conf->device_lock, flags);
2409 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2410 list_del(head->prev);
2412 spin_unlock_irqrestore(&conf->device_lock, flags);
2414 mddev = r1_bio->mddev;
2415 conf = mddev->private;
2416 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2417 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2418 test_bit(R1BIO_WriteError, &r1_bio->state))
2419 handle_sync_write_finished(conf, r1_bio);
2421 sync_request_write(mddev, r1_bio);
2422 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2423 test_bit(R1BIO_WriteError, &r1_bio->state))
2424 handle_write_finished(conf, r1_bio);
2425 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2426 handle_read_error(conf, r1_bio);
2428 /* just a partial read to be scheduled from separate
2431 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2434 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2435 md_check_recovery(mddev);
2437 blk_finish_plug(&plug);
2440 static int init_resync(struct r1conf *conf)
2444 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2445 BUG_ON(conf->r1buf_pool);
2446 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2448 if (!conf->r1buf_pool)
2450 conf->next_resync = 0;
2455 * perform a "sync" on one "block"
2457 * We need to make sure that no normal I/O request - particularly write
2458 * requests - conflict with active sync requests.
2460 * This is achieved by tracking pending requests and a 'barrier' concept
2461 * that can be installed to exclude normal IO requests.
2464 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
2466 struct r1conf *conf = mddev->private;
2467 struct r1bio *r1_bio;
2469 sector_t max_sector, nr_sectors;
2473 int write_targets = 0, read_targets = 0;
2474 sector_t sync_blocks;
2475 int still_degraded = 0;
2476 int good_sectors = RESYNC_SECTORS;
2477 int min_bad = 0; /* number of sectors that are bad in all devices */
2479 if (!conf->r1buf_pool)
2480 if (init_resync(conf))
2483 max_sector = mddev->dev_sectors;
2484 if (sector_nr >= max_sector) {
2485 /* If we aborted, we need to abort the
2486 * sync on the 'current' bitmap chunk (there will
2487 * only be one in raid1 resync.
2488 * We can find the current addess in mddev->curr_resync
2490 if (mddev->curr_resync < max_sector) /* aborted */
2491 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2493 else /* completed sync */
2496 bitmap_close_sync(mddev->bitmap);
2501 if (mddev->bitmap == NULL &&
2502 mddev->recovery_cp == MaxSector &&
2503 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2504 conf->fullsync == 0) {
2506 return max_sector - sector_nr;
2508 /* before building a request, check if we can skip these blocks..
2509 * This call the bitmap_start_sync doesn't actually record anything
2511 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2512 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2513 /* We can skip this block, and probably several more */
2518 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2519 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2521 raise_barrier(conf, sector_nr);
2525 * If we get a correctably read error during resync or recovery,
2526 * we might want to read from a different device. So we
2527 * flag all drives that could conceivably be read from for READ,
2528 * and any others (which will be non-In_sync devices) for WRITE.
2529 * If a read fails, we try reading from something else for which READ
2533 r1_bio->mddev = mddev;
2534 r1_bio->sector = sector_nr;
2536 set_bit(R1BIO_IsSync, &r1_bio->state);
2538 for (i = 0; i < conf->raid_disks * 2; i++) {
2539 struct md_rdev *rdev;
2540 bio = r1_bio->bios[i];
2543 rdev = rcu_dereference(conf->mirrors[i].rdev);
2545 test_bit(Faulty, &rdev->flags)) {
2546 if (i < conf->raid_disks)
2548 } else if (!test_bit(In_sync, &rdev->flags)) {
2550 bio->bi_end_io = end_sync_write;
2553 /* may need to read from here */
2554 sector_t first_bad = MaxSector;
2557 if (is_badblock(rdev, sector_nr, good_sectors,
2558 &first_bad, &bad_sectors)) {
2559 if (first_bad > sector_nr)
2560 good_sectors = first_bad - sector_nr;
2562 bad_sectors -= (sector_nr - first_bad);
2564 min_bad > bad_sectors)
2565 min_bad = bad_sectors;
2568 if (sector_nr < first_bad) {
2569 if (test_bit(WriteMostly, &rdev->flags)) {
2577 bio->bi_end_io = end_sync_read;
2579 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2580 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2581 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2583 * The device is suitable for reading (InSync),
2584 * but has bad block(s) here. Let's try to correct them,
2585 * if we are doing resync or repair. Otherwise, leave
2586 * this device alone for this sync request.
2589 bio->bi_end_io = end_sync_write;
2593 if (bio->bi_end_io) {
2594 atomic_inc(&rdev->nr_pending);
2595 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2596 bio->bi_bdev = rdev->bdev;
2597 bio->bi_private = r1_bio;
2603 r1_bio->read_disk = disk;
2605 if (read_targets == 0 && min_bad > 0) {
2606 /* These sectors are bad on all InSync devices, so we
2607 * need to mark them bad on all write targets
2610 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2611 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2612 struct md_rdev *rdev = conf->mirrors[i].rdev;
2613 ok = rdev_set_badblocks(rdev, sector_nr,
2617 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2622 /* Cannot record the badblocks, so need to
2624 * If there are multiple read targets, could just
2625 * fail the really bad ones ???
2627 conf->recovery_disabled = mddev->recovery_disabled;
2628 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2634 if (min_bad > 0 && min_bad < good_sectors) {
2635 /* only resync enough to reach the next bad->good
2637 good_sectors = min_bad;
2640 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2641 /* extra read targets are also write targets */
2642 write_targets += read_targets-1;
2644 if (write_targets == 0 || read_targets == 0) {
2645 /* There is nowhere to write, so all non-sync
2646 * drives must be failed - so we are finished
2650 max_sector = sector_nr + min_bad;
2651 rv = max_sector - sector_nr;
2657 if (max_sector > mddev->resync_max)
2658 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2659 if (max_sector > sector_nr + good_sectors)
2660 max_sector = sector_nr + good_sectors;
2665 int len = PAGE_SIZE;
2666 if (sector_nr + (len>>9) > max_sector)
2667 len = (max_sector - sector_nr) << 9;
2670 if (sync_blocks == 0) {
2671 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2672 &sync_blocks, still_degraded) &&
2674 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2676 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2677 if ((len >> 9) > sync_blocks)
2678 len = sync_blocks<<9;
2681 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2682 bio = r1_bio->bios[i];
2683 if (bio->bi_end_io) {
2684 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2685 if (bio_add_page(bio, page, len, 0) == 0) {
2687 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2690 bio = r1_bio->bios[i];
2691 if (bio->bi_end_io==NULL)
2693 /* remove last page from this bio */
2695 bio->bi_iter.bi_size -= len;
2696 bio_clear_flag(bio, BIO_SEG_VALID);
2702 nr_sectors += len>>9;
2703 sector_nr += len>>9;
2704 sync_blocks -= (len>>9);
2705 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2707 r1_bio->sectors = nr_sectors;
2709 /* For a user-requested sync, we read all readable devices and do a
2712 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2713 atomic_set(&r1_bio->remaining, read_targets);
2714 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2715 bio = r1_bio->bios[i];
2716 if (bio->bi_end_io == end_sync_read) {
2718 md_sync_acct(bio->bi_bdev, nr_sectors);
2719 generic_make_request(bio);
2723 atomic_set(&r1_bio->remaining, 1);
2724 bio = r1_bio->bios[r1_bio->read_disk];
2725 md_sync_acct(bio->bi_bdev, nr_sectors);
2726 generic_make_request(bio);
2732 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2737 return mddev->dev_sectors;
2740 static struct r1conf *setup_conf(struct mddev *mddev)
2742 struct r1conf *conf;
2744 struct raid1_info *disk;
2745 struct md_rdev *rdev;
2748 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2752 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2753 * mddev->raid_disks * 2,
2758 conf->tmppage = alloc_page(GFP_KERNEL);
2762 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2763 if (!conf->poolinfo)
2765 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2766 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2769 if (!conf->r1bio_pool)
2772 conf->poolinfo->mddev = mddev;
2775 spin_lock_init(&conf->device_lock);
2776 rdev_for_each(rdev, mddev) {
2777 struct request_queue *q;
2778 int disk_idx = rdev->raid_disk;
2779 if (disk_idx >= mddev->raid_disks
2782 if (test_bit(Replacement, &rdev->flags))
2783 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2785 disk = conf->mirrors + disk_idx;
2790 q = bdev_get_queue(rdev->bdev);
2792 disk->head_position = 0;
2793 disk->seq_start = MaxSector;
2795 conf->raid_disks = mddev->raid_disks;
2796 conf->mddev = mddev;
2797 INIT_LIST_HEAD(&conf->retry_list);
2798 INIT_LIST_HEAD(&conf->bio_end_io_list);
2800 spin_lock_init(&conf->resync_lock);
2801 init_waitqueue_head(&conf->wait_barrier);
2803 bio_list_init(&conf->pending_bio_list);
2804 conf->pending_count = 0;
2805 conf->recovery_disabled = mddev->recovery_disabled - 1;
2807 conf->start_next_window = MaxSector;
2808 conf->current_window_requests = conf->next_window_requests = 0;
2811 for (i = 0; i < conf->raid_disks * 2; i++) {
2813 disk = conf->mirrors + i;
2815 if (i < conf->raid_disks &&
2816 disk[conf->raid_disks].rdev) {
2817 /* This slot has a replacement. */
2819 /* No original, just make the replacement
2820 * a recovering spare
2823 disk[conf->raid_disks].rdev;
2824 disk[conf->raid_disks].rdev = NULL;
2825 } else if (!test_bit(In_sync, &disk->rdev->flags))
2826 /* Original is not in_sync - bad */
2831 !test_bit(In_sync, &disk->rdev->flags)) {
2832 disk->head_position = 0;
2834 (disk->rdev->saved_raid_disk < 0))
2840 conf->thread = md_register_thread(raid1d, mddev, "raid1");
2841 if (!conf->thread) {
2843 "md/raid1:%s: couldn't allocate thread\n",
2852 mempool_destroy(conf->r1bio_pool);
2853 kfree(conf->mirrors);
2854 safe_put_page(conf->tmppage);
2855 kfree(conf->poolinfo);
2858 return ERR_PTR(err);
2861 static void raid1_free(struct mddev *mddev, void *priv);
2862 static int run(struct mddev *mddev)
2864 struct r1conf *conf;
2866 struct md_rdev *rdev;
2868 bool discard_supported = false;
2870 if (mddev->level != 1) {
2871 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2872 mdname(mddev), mddev->level);
2875 if (mddev->reshape_position != MaxSector) {
2876 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2881 * copy the already verified devices into our private RAID1
2882 * bookkeeping area. [whatever we allocate in run(),
2883 * should be freed in raid1_free()]
2885 if (mddev->private == NULL)
2886 conf = setup_conf(mddev);
2888 conf = mddev->private;
2891 return PTR_ERR(conf);
2894 blk_queue_max_write_same_sectors(mddev->queue, 0);
2896 rdev_for_each(rdev, mddev) {
2897 if (!mddev->gendisk)
2899 disk_stack_limits(mddev->gendisk, rdev->bdev,
2900 rdev->data_offset << 9);
2901 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2902 discard_supported = true;
2905 mddev->degraded = 0;
2906 for (i=0; i < conf->raid_disks; i++)
2907 if (conf->mirrors[i].rdev == NULL ||
2908 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2909 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2912 if (conf->raid_disks - mddev->degraded == 1)
2913 mddev->recovery_cp = MaxSector;
2915 if (mddev->recovery_cp != MaxSector)
2916 printk(KERN_NOTICE "md/raid1:%s: not clean"
2917 " -- starting background reconstruction\n",
2920 "md/raid1:%s: active with %d out of %d mirrors\n",
2921 mdname(mddev), mddev->raid_disks - mddev->degraded,
2925 * Ok, everything is just fine now
2927 mddev->thread = conf->thread;
2928 conf->thread = NULL;
2929 mddev->private = conf;
2931 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2934 if (discard_supported)
2935 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2938 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2942 ret = md_integrity_register(mddev);
2944 md_unregister_thread(&mddev->thread);
2945 raid1_free(mddev, conf);
2950 static void raid1_free(struct mddev *mddev, void *priv)
2952 struct r1conf *conf = priv;
2954 mempool_destroy(conf->r1bio_pool);
2955 kfree(conf->mirrors);
2956 safe_put_page(conf->tmppage);
2957 kfree(conf->poolinfo);
2961 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2963 /* no resync is happening, and there is enough space
2964 * on all devices, so we can resize.
2965 * We need to make sure resync covers any new space.
2966 * If the array is shrinking we should possibly wait until
2967 * any io in the removed space completes, but it hardly seems
2970 sector_t newsize = raid1_size(mddev, sectors, 0);
2971 if (mddev->external_size &&
2972 mddev->array_sectors > newsize)
2974 if (mddev->bitmap) {
2975 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2979 md_set_array_sectors(mddev, newsize);
2980 set_capacity(mddev->gendisk, mddev->array_sectors);
2981 revalidate_disk(mddev->gendisk);
2982 if (sectors > mddev->dev_sectors &&
2983 mddev->recovery_cp > mddev->dev_sectors) {
2984 mddev->recovery_cp = mddev->dev_sectors;
2985 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2987 mddev->dev_sectors = sectors;
2988 mddev->resync_max_sectors = sectors;
2992 static int raid1_reshape(struct mddev *mddev)
2995 * 1/ resize the r1bio_pool
2996 * 2/ resize conf->mirrors
2998 * We allocate a new r1bio_pool if we can.
2999 * Then raise a device barrier and wait until all IO stops.
3000 * Then resize conf->mirrors and swap in the new r1bio pool.
3002 * At the same time, we "pack" the devices so that all the missing
3003 * devices have the higher raid_disk numbers.
3005 mempool_t *newpool, *oldpool;
3006 struct pool_info *newpoolinfo;
3007 struct raid1_info *newmirrors;
3008 struct r1conf *conf = mddev->private;
3009 int cnt, raid_disks;
3010 unsigned long flags;
3013 /* Cannot change chunk_size, layout, or level */
3014 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3015 mddev->layout != mddev->new_layout ||
3016 mddev->level != mddev->new_level) {
3017 mddev->new_chunk_sectors = mddev->chunk_sectors;
3018 mddev->new_layout = mddev->layout;
3019 mddev->new_level = mddev->level;
3023 err = md_allow_write(mddev);
3027 raid_disks = mddev->raid_disks + mddev->delta_disks;
3029 if (raid_disks < conf->raid_disks) {
3031 for (d= 0; d < conf->raid_disks; d++)
3032 if (conf->mirrors[d].rdev)
3034 if (cnt > raid_disks)
3038 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3041 newpoolinfo->mddev = mddev;
3042 newpoolinfo->raid_disks = raid_disks * 2;
3044 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3045 r1bio_pool_free, newpoolinfo);
3050 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3054 mempool_destroy(newpool);
3058 freeze_array(conf, 0);
3060 /* ok, everything is stopped */
3061 oldpool = conf->r1bio_pool;
3062 conf->r1bio_pool = newpool;
3064 for (d = d2 = 0; d < conf->raid_disks; d++) {
3065 struct md_rdev *rdev = conf->mirrors[d].rdev;
3066 if (rdev && rdev->raid_disk != d2) {
3067 sysfs_unlink_rdev(mddev, rdev);
3068 rdev->raid_disk = d2;
3069 sysfs_unlink_rdev(mddev, rdev);
3070 if (sysfs_link_rdev(mddev, rdev))
3072 "md/raid1:%s: cannot register rd%d\n",
3073 mdname(mddev), rdev->raid_disk);
3076 newmirrors[d2++].rdev = rdev;
3078 kfree(conf->mirrors);
3079 conf->mirrors = newmirrors;
3080 kfree(conf->poolinfo);
3081 conf->poolinfo = newpoolinfo;
3083 spin_lock_irqsave(&conf->device_lock, flags);
3084 mddev->degraded += (raid_disks - conf->raid_disks);
3085 spin_unlock_irqrestore(&conf->device_lock, flags);
3086 conf->raid_disks = mddev->raid_disks = raid_disks;
3087 mddev->delta_disks = 0;
3089 unfreeze_array(conf);
3091 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3092 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3093 md_wakeup_thread(mddev->thread);
3095 mempool_destroy(oldpool);
3099 static void raid1_quiesce(struct mddev *mddev, int state)
3101 struct r1conf *conf = mddev->private;
3104 case 2: /* wake for suspend */
3105 wake_up(&conf->wait_barrier);
3108 freeze_array(conf, 0);
3111 unfreeze_array(conf);
3116 static void *raid1_takeover(struct mddev *mddev)
3118 /* raid1 can take over:
3119 * raid5 with 2 devices, any layout or chunk size
3121 if (mddev->level == 5 && mddev->raid_disks == 2) {
3122 struct r1conf *conf;
3123 mddev->new_level = 1;
3124 mddev->new_layout = 0;
3125 mddev->new_chunk_sectors = 0;
3126 conf = setup_conf(mddev);
3128 /* Array must appear to be quiesced */
3129 conf->array_frozen = 1;
3132 return ERR_PTR(-EINVAL);
3135 static struct md_personality raid1_personality =
3139 .owner = THIS_MODULE,
3140 .make_request = make_request,
3144 .error_handler = error,
3145 .hot_add_disk = raid1_add_disk,
3146 .hot_remove_disk= raid1_remove_disk,
3147 .spare_active = raid1_spare_active,
3148 .sync_request = sync_request,
3149 .resize = raid1_resize,
3151 .check_reshape = raid1_reshape,
3152 .quiesce = raid1_quiesce,
3153 .takeover = raid1_takeover,
3154 .congested = raid1_congested,
3157 static int __init raid_init(void)
3159 return register_md_personality(&raid1_personality);
3162 static void raid_exit(void)
3164 unregister_md_personality(&raid1_personality);
3167 module_init(raid_init);
3168 module_exit(raid_exit);
3169 MODULE_LICENSE("GPL");
3170 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3171 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3172 MODULE_ALIAS("md-raid1");
3173 MODULE_ALIAS("md-level-1");
3175 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);