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 >=
885 conf->mddev->curr_resync_completed) {
886 if (conf->start_next_window == MaxSector)
887 conf->start_next_window =
889 NEXT_NORMALIO_DISTANCE;
891 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
892 <= bio->bi_iter.bi_sector)
893 conf->next_window_requests++;
895 conf->current_window_requests++;
896 sector = conf->start_next_window;
901 spin_unlock_irq(&conf->resync_lock);
905 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
910 spin_lock_irqsave(&conf->resync_lock, flags);
912 if (start_next_window) {
913 if (start_next_window == conf->start_next_window) {
914 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
916 conf->next_window_requests--;
918 conf->current_window_requests--;
920 conf->current_window_requests--;
922 if (!conf->current_window_requests) {
923 if (conf->next_window_requests) {
924 conf->current_window_requests =
925 conf->next_window_requests;
926 conf->next_window_requests = 0;
927 conf->start_next_window +=
928 NEXT_NORMALIO_DISTANCE;
930 conf->start_next_window = MaxSector;
933 spin_unlock_irqrestore(&conf->resync_lock, flags);
934 wake_up(&conf->wait_barrier);
937 static void freeze_array(struct r1conf *conf, int extra)
939 /* stop syncio and normal IO and wait for everything to
941 * We wait until nr_pending match nr_queued+extra
942 * This is called in the context of one normal IO request
943 * that has failed. Thus any sync request that might be pending
944 * will be blocked by nr_pending, and we need to wait for
945 * pending IO requests to complete or be queued for re-try.
946 * Thus the number queued (nr_queued) plus this request (extra)
947 * must match the number of pending IOs (nr_pending) before
950 spin_lock_irq(&conf->resync_lock);
951 conf->array_frozen = 1;
952 wait_event_lock_irq_cmd(conf->wait_barrier,
953 conf->nr_pending == conf->nr_queued+extra,
955 flush_pending_writes(conf));
956 spin_unlock_irq(&conf->resync_lock);
958 static void unfreeze_array(struct r1conf *conf)
960 /* reverse the effect of the freeze */
961 spin_lock_irq(&conf->resync_lock);
962 conf->array_frozen = 0;
963 wake_up(&conf->wait_barrier);
964 spin_unlock_irq(&conf->resync_lock);
967 /* duplicate the data pages for behind I/O
969 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
972 struct bio_vec *bvec;
973 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
975 if (unlikely(!bvecs))
978 bio_for_each_segment_all(bvec, bio, i) {
980 bvecs[i].bv_page = alloc_page(GFP_NOIO);
981 if (unlikely(!bvecs[i].bv_page))
983 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
984 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
985 kunmap(bvecs[i].bv_page);
986 kunmap(bvec->bv_page);
988 r1_bio->behind_bvecs = bvecs;
989 r1_bio->behind_page_count = bio->bi_vcnt;
990 set_bit(R1BIO_BehindIO, &r1_bio->state);
994 for (i = 0; i < bio->bi_vcnt; i++)
995 if (bvecs[i].bv_page)
996 put_page(bvecs[i].bv_page);
998 pr_debug("%dB behind alloc failed, doing sync I/O\n",
999 bio->bi_iter.bi_size);
1002 struct raid1_plug_cb {
1003 struct blk_plug_cb cb;
1004 struct bio_list pending;
1008 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1010 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1012 struct mddev *mddev = plug->cb.data;
1013 struct r1conf *conf = mddev->private;
1016 if (from_schedule || current->bio_list) {
1017 spin_lock_irq(&conf->device_lock);
1018 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1019 conf->pending_count += plug->pending_cnt;
1020 spin_unlock_irq(&conf->device_lock);
1021 wake_up(&conf->wait_barrier);
1022 md_wakeup_thread(mddev->thread);
1027 /* we aren't scheduling, so we can do the write-out directly. */
1028 bio = bio_list_get(&plug->pending);
1029 bitmap_unplug(mddev->bitmap);
1030 wake_up(&conf->wait_barrier);
1032 while (bio) { /* submit pending writes */
1033 struct bio *next = bio->bi_next;
1034 bio->bi_next = NULL;
1035 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1036 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1037 /* Just ignore it */
1040 generic_make_request(bio);
1046 static void make_request(struct mddev *mddev, struct bio * bio)
1048 struct r1conf *conf = mddev->private;
1049 struct raid1_info *mirror;
1050 struct r1bio *r1_bio;
1051 struct bio *read_bio;
1053 struct bitmap *bitmap;
1054 unsigned long flags;
1055 const int rw = bio_data_dir(bio);
1056 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1057 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1058 const unsigned long do_discard = (bio->bi_rw
1059 & (REQ_DISCARD | REQ_SECURE));
1060 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1061 struct md_rdev *blocked_rdev;
1062 struct blk_plug_cb *cb;
1063 struct raid1_plug_cb *plug = NULL;
1065 int sectors_handled;
1067 sector_t start_next_window;
1070 * Register the new request and wait if the reconstruction
1071 * thread has put up a bar for new requests.
1072 * Continue immediately if no resync is active currently.
1075 md_write_start(mddev, bio); /* wait on superblock update early */
1077 if (bio_data_dir(bio) == WRITE &&
1078 ((bio_end_sector(bio) > mddev->suspend_lo &&
1079 bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1080 (mddev_is_clustered(mddev) &&
1081 md_cluster_ops->area_resyncing(mddev, WRITE,
1082 bio->bi_iter.bi_sector, bio_end_sector(bio))))) {
1083 /* As the suspend_* range is controlled by
1084 * userspace, we want an interruptible
1089 flush_signals(current);
1090 prepare_to_wait(&conf->wait_barrier,
1091 &w, TASK_INTERRUPTIBLE);
1092 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1093 bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1094 (mddev_is_clustered(mddev) &&
1095 !md_cluster_ops->area_resyncing(mddev, WRITE,
1096 bio->bi_iter.bi_sector, bio_end_sector(bio))))
1100 finish_wait(&conf->wait_barrier, &w);
1103 start_next_window = wait_barrier(conf, bio);
1105 bitmap = mddev->bitmap;
1108 * make_request() can abort the operation when READA is being
1109 * used and no empty request is available.
1112 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1114 r1_bio->master_bio = bio;
1115 r1_bio->sectors = bio_sectors(bio);
1117 r1_bio->mddev = mddev;
1118 r1_bio->sector = bio->bi_iter.bi_sector;
1120 /* We might need to issue multiple reads to different
1121 * devices if there are bad blocks around, so we keep
1122 * track of the number of reads in bio->bi_phys_segments.
1123 * If this is 0, there is only one r1_bio and no locking
1124 * will be needed when requests complete. If it is
1125 * non-zero, then it is the number of not-completed requests.
1127 bio->bi_phys_segments = 0;
1128 bio_clear_flag(bio, BIO_SEG_VALID);
1132 * read balancing logic:
1137 rdisk = read_balance(conf, r1_bio, &max_sectors);
1140 /* couldn't find anywhere to read from */
1141 raid_end_bio_io(r1_bio);
1144 mirror = conf->mirrors + rdisk;
1146 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1148 /* Reading from a write-mostly device must
1149 * take care not to over-take any writes
1152 wait_event(bitmap->behind_wait,
1153 atomic_read(&bitmap->behind_writes) == 0);
1155 r1_bio->read_disk = rdisk;
1156 r1_bio->start_next_window = 0;
1158 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1159 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1162 r1_bio->bios[rdisk] = read_bio;
1164 read_bio->bi_iter.bi_sector = r1_bio->sector +
1165 mirror->rdev->data_offset;
1166 read_bio->bi_bdev = mirror->rdev->bdev;
1167 read_bio->bi_end_io = raid1_end_read_request;
1168 read_bio->bi_rw = READ | do_sync;
1169 read_bio->bi_private = r1_bio;
1171 if (max_sectors < r1_bio->sectors) {
1172 /* could not read all from this device, so we will
1173 * need another r1_bio.
1176 sectors_handled = (r1_bio->sector + max_sectors
1177 - bio->bi_iter.bi_sector);
1178 r1_bio->sectors = max_sectors;
1179 spin_lock_irq(&conf->device_lock);
1180 if (bio->bi_phys_segments == 0)
1181 bio->bi_phys_segments = 2;
1183 bio->bi_phys_segments++;
1184 spin_unlock_irq(&conf->device_lock);
1185 /* Cannot call generic_make_request directly
1186 * as that will be queued in __make_request
1187 * and subsequent mempool_alloc might block waiting
1188 * for it. So hand bio over to raid1d.
1190 reschedule_retry(r1_bio);
1192 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1194 r1_bio->master_bio = bio;
1195 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1197 r1_bio->mddev = mddev;
1198 r1_bio->sector = bio->bi_iter.bi_sector +
1202 generic_make_request(read_bio);
1209 if (conf->pending_count >= max_queued_requests) {
1210 md_wakeup_thread(mddev->thread);
1211 wait_event(conf->wait_barrier,
1212 conf->pending_count < max_queued_requests);
1214 /* first select target devices under rcu_lock and
1215 * inc refcount on their rdev. Record them by setting
1217 * If there are known/acknowledged bad blocks on any device on
1218 * which we have seen a write error, we want to avoid writing those
1220 * This potentially requires several writes to write around
1221 * the bad blocks. Each set of writes gets it's own r1bio
1222 * with a set of bios attached.
1225 disks = conf->raid_disks * 2;
1227 r1_bio->start_next_window = start_next_window;
1228 blocked_rdev = NULL;
1230 max_sectors = r1_bio->sectors;
1231 for (i = 0; i < disks; i++) {
1232 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1233 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1234 atomic_inc(&rdev->nr_pending);
1235 blocked_rdev = rdev;
1238 r1_bio->bios[i] = NULL;
1239 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1240 if (i < conf->raid_disks)
1241 set_bit(R1BIO_Degraded, &r1_bio->state);
1245 atomic_inc(&rdev->nr_pending);
1246 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1251 is_bad = is_badblock(rdev, r1_bio->sector,
1253 &first_bad, &bad_sectors);
1255 /* mustn't write here until the bad block is
1257 set_bit(BlockedBadBlocks, &rdev->flags);
1258 blocked_rdev = rdev;
1261 if (is_bad && first_bad <= r1_bio->sector) {
1262 /* Cannot write here at all */
1263 bad_sectors -= (r1_bio->sector - first_bad);
1264 if (bad_sectors < max_sectors)
1265 /* mustn't write more than bad_sectors
1266 * to other devices yet
1268 max_sectors = bad_sectors;
1269 rdev_dec_pending(rdev, mddev);
1270 /* We don't set R1BIO_Degraded as that
1271 * only applies if the disk is
1272 * missing, so it might be re-added,
1273 * and we want to know to recover this
1275 * In this case the device is here,
1276 * and the fact that this chunk is not
1277 * in-sync is recorded in the bad
1283 int good_sectors = first_bad - r1_bio->sector;
1284 if (good_sectors < max_sectors)
1285 max_sectors = good_sectors;
1288 r1_bio->bios[i] = bio;
1292 if (unlikely(blocked_rdev)) {
1293 /* Wait for this device to become unblocked */
1295 sector_t old = start_next_window;
1297 for (j = 0; j < i; j++)
1298 if (r1_bio->bios[j])
1299 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1301 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1302 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1303 start_next_window = wait_barrier(conf, bio);
1305 * We must make sure the multi r1bios of bio have
1306 * the same value of bi_phys_segments
1308 if (bio->bi_phys_segments && old &&
1309 old != start_next_window)
1310 /* Wait for the former r1bio(s) to complete */
1311 wait_event(conf->wait_barrier,
1312 bio->bi_phys_segments == 1);
1316 if (max_sectors < r1_bio->sectors) {
1317 /* We are splitting this write into multiple parts, so
1318 * we need to prepare for allocating another r1_bio.
1320 r1_bio->sectors = max_sectors;
1321 spin_lock_irq(&conf->device_lock);
1322 if (bio->bi_phys_segments == 0)
1323 bio->bi_phys_segments = 2;
1325 bio->bi_phys_segments++;
1326 spin_unlock_irq(&conf->device_lock);
1328 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1330 atomic_set(&r1_bio->remaining, 1);
1331 atomic_set(&r1_bio->behind_remaining, 0);
1334 for (i = 0; i < disks; i++) {
1336 if (!r1_bio->bios[i])
1339 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1340 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1344 * Not if there are too many, or cannot
1345 * allocate memory, or a reader on WriteMostly
1346 * is waiting for behind writes to flush */
1348 (atomic_read(&bitmap->behind_writes)
1349 < mddev->bitmap_info.max_write_behind) &&
1350 !waitqueue_active(&bitmap->behind_wait))
1351 alloc_behind_pages(mbio, r1_bio);
1353 bitmap_startwrite(bitmap, r1_bio->sector,
1355 test_bit(R1BIO_BehindIO,
1359 if (r1_bio->behind_bvecs) {
1360 struct bio_vec *bvec;
1364 * We trimmed the bio, so _all is legit
1366 bio_for_each_segment_all(bvec, mbio, j)
1367 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1368 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1369 atomic_inc(&r1_bio->behind_remaining);
1372 r1_bio->bios[i] = mbio;
1374 mbio->bi_iter.bi_sector = (r1_bio->sector +
1375 conf->mirrors[i].rdev->data_offset);
1376 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1377 mbio->bi_end_io = raid1_end_write_request;
1379 WRITE | do_flush_fua | do_sync | do_discard | do_same;
1380 mbio->bi_private = r1_bio;
1382 atomic_inc(&r1_bio->remaining);
1384 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1386 plug = container_of(cb, struct raid1_plug_cb, cb);
1389 spin_lock_irqsave(&conf->device_lock, flags);
1391 bio_list_add(&plug->pending, mbio);
1392 plug->pending_cnt++;
1394 bio_list_add(&conf->pending_bio_list, mbio);
1395 conf->pending_count++;
1397 spin_unlock_irqrestore(&conf->device_lock, flags);
1399 md_wakeup_thread(mddev->thread);
1401 /* Mustn't call r1_bio_write_done before this next test,
1402 * as it could result in the bio being freed.
1404 if (sectors_handled < bio_sectors(bio)) {
1405 r1_bio_write_done(r1_bio);
1406 /* We need another r1_bio. It has already been counted
1407 * in bio->bi_phys_segments
1409 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1410 r1_bio->master_bio = bio;
1411 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1413 r1_bio->mddev = mddev;
1414 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1418 r1_bio_write_done(r1_bio);
1420 /* In case raid1d snuck in to freeze_array */
1421 wake_up(&conf->wait_barrier);
1424 static void status(struct seq_file *seq, struct mddev *mddev)
1426 struct r1conf *conf = mddev->private;
1429 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1430 conf->raid_disks - mddev->degraded);
1432 for (i = 0; i < conf->raid_disks; i++) {
1433 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1434 seq_printf(seq, "%s",
1435 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1438 seq_printf(seq, "]");
1441 static void error(struct mddev *mddev, struct md_rdev *rdev)
1443 char b[BDEVNAME_SIZE];
1444 struct r1conf *conf = mddev->private;
1445 unsigned long flags;
1448 * If it is not operational, then we have already marked it as dead
1449 * else if it is the last working disks, ignore the error, let the
1450 * next level up know.
1451 * else mark the drive as failed
1453 if (test_bit(In_sync, &rdev->flags)
1454 && (conf->raid_disks - mddev->degraded) == 1) {
1456 * Don't fail the drive, act as though we were just a
1457 * normal single drive.
1458 * However don't try a recovery from this drive as
1459 * it is very likely to fail.
1461 conf->recovery_disabled = mddev->recovery_disabled;
1464 set_bit(Blocked, &rdev->flags);
1465 spin_lock_irqsave(&conf->device_lock, flags);
1466 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1468 set_bit(Faulty, &rdev->flags);
1470 set_bit(Faulty, &rdev->flags);
1471 spin_unlock_irqrestore(&conf->device_lock, flags);
1473 * if recovery is running, make sure it aborts.
1475 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1476 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1477 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1479 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1480 "md/raid1:%s: Operation continuing on %d devices.\n",
1481 mdname(mddev), bdevname(rdev->bdev, b),
1482 mdname(mddev), conf->raid_disks - mddev->degraded);
1485 static void print_conf(struct r1conf *conf)
1489 printk(KERN_DEBUG "RAID1 conf printout:\n");
1491 printk(KERN_DEBUG "(!conf)\n");
1494 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1498 for (i = 0; i < conf->raid_disks; i++) {
1499 char b[BDEVNAME_SIZE];
1500 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1502 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1503 i, !test_bit(In_sync, &rdev->flags),
1504 !test_bit(Faulty, &rdev->flags),
1505 bdevname(rdev->bdev,b));
1510 static void close_sync(struct r1conf *conf)
1512 wait_barrier(conf, NULL);
1513 allow_barrier(conf, 0, 0);
1515 mempool_destroy(conf->r1buf_pool);
1516 conf->r1buf_pool = NULL;
1518 spin_lock_irq(&conf->resync_lock);
1519 conf->next_resync = 0;
1520 conf->start_next_window = MaxSector;
1521 conf->current_window_requests +=
1522 conf->next_window_requests;
1523 conf->next_window_requests = 0;
1524 spin_unlock_irq(&conf->resync_lock);
1527 static int raid1_spare_active(struct mddev *mddev)
1530 struct r1conf *conf = mddev->private;
1532 unsigned long flags;
1535 * Find all failed disks within the RAID1 configuration
1536 * and mark them readable.
1537 * Called under mddev lock, so rcu protection not needed.
1538 * device_lock used to avoid races with raid1_end_read_request
1539 * which expects 'In_sync' flags and ->degraded to be consistent.
1541 spin_lock_irqsave(&conf->device_lock, flags);
1542 for (i = 0; i < conf->raid_disks; i++) {
1543 struct md_rdev *rdev = conf->mirrors[i].rdev;
1544 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1546 && !test_bit(Candidate, &repl->flags)
1547 && repl->recovery_offset == MaxSector
1548 && !test_bit(Faulty, &repl->flags)
1549 && !test_and_set_bit(In_sync, &repl->flags)) {
1550 /* replacement has just become active */
1552 !test_and_clear_bit(In_sync, &rdev->flags))
1555 /* Replaced device not technically
1556 * faulty, but we need to be sure
1557 * it gets removed and never re-added
1559 set_bit(Faulty, &rdev->flags);
1560 sysfs_notify_dirent_safe(
1565 && rdev->recovery_offset == MaxSector
1566 && !test_bit(Faulty, &rdev->flags)
1567 && !test_and_set_bit(In_sync, &rdev->flags)) {
1569 sysfs_notify_dirent_safe(rdev->sysfs_state);
1572 mddev->degraded -= count;
1573 spin_unlock_irqrestore(&conf->device_lock, flags);
1579 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1581 struct r1conf *conf = mddev->private;
1584 struct raid1_info *p;
1586 int last = conf->raid_disks - 1;
1588 if (mddev->recovery_disabled == conf->recovery_disabled)
1591 if (rdev->raid_disk >= 0)
1592 first = last = rdev->raid_disk;
1594 for (mirror = first; mirror <= last; mirror++) {
1595 p = conf->mirrors+mirror;
1599 disk_stack_limits(mddev->gendisk, rdev->bdev,
1600 rdev->data_offset << 9);
1602 p->head_position = 0;
1603 rdev->raid_disk = mirror;
1605 /* As all devices are equivalent, we don't need a full recovery
1606 * if this was recently any drive of the array
1608 if (rdev->saved_raid_disk < 0)
1610 rcu_assign_pointer(p->rdev, rdev);
1613 if (test_bit(WantReplacement, &p->rdev->flags) &&
1614 p[conf->raid_disks].rdev == NULL) {
1615 /* Add this device as a replacement */
1616 clear_bit(In_sync, &rdev->flags);
1617 set_bit(Replacement, &rdev->flags);
1618 rdev->raid_disk = mirror;
1621 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1625 md_integrity_add_rdev(rdev, mddev);
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) {
1660 if (atomic_read(&rdev->nr_pending)) {
1661 /* lost the race, try later */
1665 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1666 /* We just removed a device that is being replaced.
1667 * Move down the replacement. We drain all IO before
1668 * doing this to avoid confusion.
1670 struct md_rdev *repl =
1671 conf->mirrors[conf->raid_disks + number].rdev;
1672 freeze_array(conf, 0);
1673 clear_bit(Replacement, &repl->flags);
1675 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1676 unfreeze_array(conf);
1677 clear_bit(WantReplacement, &rdev->flags);
1679 clear_bit(WantReplacement, &rdev->flags);
1680 err = md_integrity_register(mddev);
1688 static void end_sync_read(struct bio *bio)
1690 struct r1bio *r1_bio = bio->bi_private;
1692 update_head_pos(r1_bio->read_disk, r1_bio);
1695 * we have read a block, now it needs to be re-written,
1696 * or re-read if the read failed.
1697 * We don't do much here, just schedule handling by raid1d
1700 set_bit(R1BIO_Uptodate, &r1_bio->state);
1702 if (atomic_dec_and_test(&r1_bio->remaining))
1703 reschedule_retry(r1_bio);
1706 static void end_sync_write(struct bio *bio)
1708 int uptodate = !bio->bi_error;
1709 struct r1bio *r1_bio = bio->bi_private;
1710 struct mddev *mddev = r1_bio->mddev;
1711 struct r1conf *conf = mddev->private;
1716 mirror = find_bio_disk(r1_bio, bio);
1719 sector_t sync_blocks = 0;
1720 sector_t s = r1_bio->sector;
1721 long sectors_to_go = r1_bio->sectors;
1722 /* make sure these bits doesn't get cleared. */
1724 bitmap_end_sync(mddev->bitmap, s,
1727 sectors_to_go -= sync_blocks;
1728 } while (sectors_to_go > 0);
1729 set_bit(WriteErrorSeen,
1730 &conf->mirrors[mirror].rdev->flags);
1731 if (!test_and_set_bit(WantReplacement,
1732 &conf->mirrors[mirror].rdev->flags))
1733 set_bit(MD_RECOVERY_NEEDED, &
1735 set_bit(R1BIO_WriteError, &r1_bio->state);
1736 } else if (is_badblock(conf->mirrors[mirror].rdev,
1739 &first_bad, &bad_sectors) &&
1740 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1743 &first_bad, &bad_sectors)
1745 set_bit(R1BIO_MadeGood, &r1_bio->state);
1747 if (atomic_dec_and_test(&r1_bio->remaining)) {
1748 int s = r1_bio->sectors;
1749 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1750 test_bit(R1BIO_WriteError, &r1_bio->state))
1751 reschedule_retry(r1_bio);
1754 md_done_sync(mddev, s, uptodate);
1759 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1760 int sectors, struct page *page, int rw)
1762 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1766 set_bit(WriteErrorSeen, &rdev->flags);
1767 if (!test_and_set_bit(WantReplacement,
1769 set_bit(MD_RECOVERY_NEEDED, &
1770 rdev->mddev->recovery);
1772 /* need to record an error - either for the block or the device */
1773 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1774 md_error(rdev->mddev, rdev);
1778 static int fix_sync_read_error(struct r1bio *r1_bio)
1780 /* Try some synchronous reads of other devices to get
1781 * good data, much like with normal read errors. Only
1782 * read into the pages we already have so we don't
1783 * need to re-issue the read request.
1784 * We don't need to freeze the array, because being in an
1785 * active sync request, there is no normal IO, and
1786 * no overlapping syncs.
1787 * We don't need to check is_badblock() again as we
1788 * made sure that anything with a bad block in range
1789 * will have bi_end_io clear.
1791 struct mddev *mddev = r1_bio->mddev;
1792 struct r1conf *conf = mddev->private;
1793 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1794 sector_t sect = r1_bio->sector;
1795 int sectors = r1_bio->sectors;
1800 int d = r1_bio->read_disk;
1802 struct md_rdev *rdev;
1805 if (s > (PAGE_SIZE>>9))
1808 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1809 /* No rcu protection needed here devices
1810 * can only be removed when no resync is
1811 * active, and resync is currently active
1813 rdev = conf->mirrors[d].rdev;
1814 if (sync_page_io(rdev, sect, s<<9,
1815 bio->bi_io_vec[idx].bv_page,
1822 if (d == conf->raid_disks * 2)
1824 } while (!success && d != r1_bio->read_disk);
1827 char b[BDEVNAME_SIZE];
1829 /* Cannot read from anywhere, this block is lost.
1830 * Record a bad block on each device. If that doesn't
1831 * work just disable and interrupt the recovery.
1832 * Don't fail devices as that won't really help.
1834 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1835 " for block %llu\n",
1837 bdevname(bio->bi_bdev, b),
1838 (unsigned long long)r1_bio->sector);
1839 for (d = 0; d < conf->raid_disks * 2; d++) {
1840 rdev = conf->mirrors[d].rdev;
1841 if (!rdev || test_bit(Faulty, &rdev->flags))
1843 if (!rdev_set_badblocks(rdev, sect, s, 0))
1847 conf->recovery_disabled =
1848 mddev->recovery_disabled;
1849 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1850 md_done_sync(mddev, r1_bio->sectors, 0);
1862 /* write it back and re-read */
1863 while (d != r1_bio->read_disk) {
1865 d = conf->raid_disks * 2;
1867 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1869 rdev = conf->mirrors[d].rdev;
1870 if (r1_sync_page_io(rdev, sect, s,
1871 bio->bi_io_vec[idx].bv_page,
1873 r1_bio->bios[d]->bi_end_io = NULL;
1874 rdev_dec_pending(rdev, mddev);
1878 while (d != r1_bio->read_disk) {
1880 d = conf->raid_disks * 2;
1882 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1884 rdev = conf->mirrors[d].rdev;
1885 if (r1_sync_page_io(rdev, sect, s,
1886 bio->bi_io_vec[idx].bv_page,
1888 atomic_add(s, &rdev->corrected_errors);
1894 set_bit(R1BIO_Uptodate, &r1_bio->state);
1899 static void process_checks(struct r1bio *r1_bio)
1901 /* We have read all readable devices. If we haven't
1902 * got the block, then there is no hope left.
1903 * If we have, then we want to do a comparison
1904 * and skip the write if everything is the same.
1905 * If any blocks failed to read, then we need to
1906 * attempt an over-write
1908 struct mddev *mddev = r1_bio->mddev;
1909 struct r1conf *conf = mddev->private;
1914 /* Fix variable parts of all bios */
1915 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1916 for (i = 0; i < conf->raid_disks * 2; i++) {
1920 struct bio *b = r1_bio->bios[i];
1921 if (b->bi_end_io != end_sync_read)
1923 /* fixup the bio for reuse, but preserve errno */
1924 error = b->bi_error;
1926 b->bi_error = error;
1928 b->bi_iter.bi_size = r1_bio->sectors << 9;
1929 b->bi_iter.bi_sector = r1_bio->sector +
1930 conf->mirrors[i].rdev->data_offset;
1931 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1932 b->bi_end_io = end_sync_read;
1933 b->bi_private = r1_bio;
1935 size = b->bi_iter.bi_size;
1936 for (j = 0; j < vcnt ; j++) {
1938 bi = &b->bi_io_vec[j];
1940 if (size > PAGE_SIZE)
1941 bi->bv_len = PAGE_SIZE;
1947 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1948 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1949 !r1_bio->bios[primary]->bi_error) {
1950 r1_bio->bios[primary]->bi_end_io = NULL;
1951 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1954 r1_bio->read_disk = primary;
1955 for (i = 0; i < conf->raid_disks * 2; i++) {
1957 struct bio *pbio = r1_bio->bios[primary];
1958 struct bio *sbio = r1_bio->bios[i];
1959 int error = sbio->bi_error;
1961 if (sbio->bi_end_io != end_sync_read)
1963 /* Now we can 'fixup' the error value */
1967 for (j = vcnt; j-- ; ) {
1969 p = pbio->bi_io_vec[j].bv_page;
1970 s = sbio->bi_io_vec[j].bv_page;
1971 if (memcmp(page_address(p),
1973 sbio->bi_io_vec[j].bv_len))
1979 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1980 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1982 /* No need to write to this device. */
1983 sbio->bi_end_io = NULL;
1984 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1988 bio_copy_data(sbio, pbio);
1992 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1994 struct r1conf *conf = mddev->private;
1996 int disks = conf->raid_disks * 2;
1997 struct bio *bio, *wbio;
1999 bio = r1_bio->bios[r1_bio->read_disk];
2001 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2002 /* ouch - failed to read all of that. */
2003 if (!fix_sync_read_error(r1_bio))
2006 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2007 process_checks(r1_bio);
2012 atomic_set(&r1_bio->remaining, 1);
2013 for (i = 0; i < disks ; i++) {
2014 wbio = r1_bio->bios[i];
2015 if (wbio->bi_end_io == NULL ||
2016 (wbio->bi_end_io == end_sync_read &&
2017 (i == r1_bio->read_disk ||
2018 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2021 wbio->bi_rw = WRITE;
2022 wbio->bi_end_io = end_sync_write;
2023 atomic_inc(&r1_bio->remaining);
2024 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2026 generic_make_request(wbio);
2029 if (atomic_dec_and_test(&r1_bio->remaining)) {
2030 /* if we're here, all write(s) have completed, so clean up */
2031 int s = r1_bio->sectors;
2032 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2033 test_bit(R1BIO_WriteError, &r1_bio->state))
2034 reschedule_retry(r1_bio);
2037 md_done_sync(mddev, s, 1);
2043 * This is a kernel thread which:
2045 * 1. Retries failed read operations on working mirrors.
2046 * 2. Updates the raid superblock when problems encounter.
2047 * 3. Performs writes following reads for array synchronising.
2050 static void fix_read_error(struct r1conf *conf, int read_disk,
2051 sector_t sect, int sectors)
2053 struct mddev *mddev = conf->mddev;
2059 struct md_rdev *rdev;
2061 if (s > (PAGE_SIZE>>9))
2065 /* Note: no rcu protection needed here
2066 * as this is synchronous in the raid1d thread
2067 * which is the thread that might remove
2068 * a device. If raid1d ever becomes multi-threaded....
2073 rdev = conf->mirrors[d].rdev;
2075 (test_bit(In_sync, &rdev->flags) ||
2076 (!test_bit(Faulty, &rdev->flags) &&
2077 rdev->recovery_offset >= sect + s)) &&
2078 is_badblock(rdev, sect, s,
2079 &first_bad, &bad_sectors) == 0 &&
2080 sync_page_io(rdev, sect, s<<9,
2081 conf->tmppage, READ, false))
2085 if (d == conf->raid_disks * 2)
2088 } while (!success && d != read_disk);
2091 /* Cannot read from anywhere - mark it bad */
2092 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2093 if (!rdev_set_badblocks(rdev, sect, s, 0))
2094 md_error(mddev, rdev);
2097 /* write it back and re-read */
2099 while (d != read_disk) {
2101 d = conf->raid_disks * 2;
2103 rdev = conf->mirrors[d].rdev;
2105 !test_bit(Faulty, &rdev->flags))
2106 r1_sync_page_io(rdev, sect, s,
2107 conf->tmppage, WRITE);
2110 while (d != read_disk) {
2111 char b[BDEVNAME_SIZE];
2113 d = conf->raid_disks * 2;
2115 rdev = conf->mirrors[d].rdev;
2117 !test_bit(Faulty, &rdev->flags)) {
2118 if (r1_sync_page_io(rdev, sect, s,
2119 conf->tmppage, READ)) {
2120 atomic_add(s, &rdev->corrected_errors);
2122 "md/raid1:%s: read error corrected "
2123 "(%d sectors at %llu on %s)\n",
2125 (unsigned long long)(sect +
2127 bdevname(rdev->bdev, b));
2136 static int narrow_write_error(struct r1bio *r1_bio, int i)
2138 struct mddev *mddev = r1_bio->mddev;
2139 struct r1conf *conf = mddev->private;
2140 struct md_rdev *rdev = conf->mirrors[i].rdev;
2142 /* bio has the data to be written to device 'i' where
2143 * we just recently had a write error.
2144 * We repeatedly clone the bio and trim down to one block,
2145 * then try the write. Where the write fails we record
2147 * It is conceivable that the bio doesn't exactly align with
2148 * blocks. We must handle this somehow.
2150 * We currently own a reference on the rdev.
2156 int sect_to_write = r1_bio->sectors;
2159 if (rdev->badblocks.shift < 0)
2162 block_sectors = roundup(1 << rdev->badblocks.shift,
2163 bdev_logical_block_size(rdev->bdev) >> 9);
2164 sector = r1_bio->sector;
2165 sectors = ((sector + block_sectors)
2166 & ~(sector_t)(block_sectors - 1))
2169 while (sect_to_write) {
2171 if (sectors > sect_to_write)
2172 sectors = sect_to_write;
2173 /* Write at 'sector' for 'sectors'*/
2175 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2176 unsigned vcnt = r1_bio->behind_page_count;
2177 struct bio_vec *vec = r1_bio->behind_bvecs;
2179 while (!vec->bv_page) {
2184 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2185 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2187 wbio->bi_vcnt = vcnt;
2189 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2192 wbio->bi_rw = WRITE;
2193 wbio->bi_iter.bi_sector = r1_bio->sector;
2194 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2196 bio_trim(wbio, sector - r1_bio->sector, sectors);
2197 wbio->bi_iter.bi_sector += rdev->data_offset;
2198 wbio->bi_bdev = rdev->bdev;
2199 if (submit_bio_wait(WRITE, wbio) == 0)
2201 ok = rdev_set_badblocks(rdev, sector,
2206 sect_to_write -= sectors;
2208 sectors = block_sectors;
2213 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2216 int s = r1_bio->sectors;
2217 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2218 struct md_rdev *rdev = conf->mirrors[m].rdev;
2219 struct bio *bio = r1_bio->bios[m];
2220 if (bio->bi_end_io == NULL)
2222 if (!bio->bi_error &&
2223 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2224 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2226 if (bio->bi_error &&
2227 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2228 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2229 md_error(conf->mddev, rdev);
2233 md_done_sync(conf->mddev, s, 1);
2236 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2240 for (m = 0; m < conf->raid_disks * 2 ; m++)
2241 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2242 struct md_rdev *rdev = conf->mirrors[m].rdev;
2243 rdev_clear_badblocks(rdev,
2245 r1_bio->sectors, 0);
2246 rdev_dec_pending(rdev, conf->mddev);
2247 } else if (r1_bio->bios[m] != NULL) {
2248 /* This drive got a write error. We need to
2249 * narrow down and record precise write
2253 if (!narrow_write_error(r1_bio, m)) {
2254 md_error(conf->mddev,
2255 conf->mirrors[m].rdev);
2256 /* an I/O failed, we can't clear the bitmap */
2257 set_bit(R1BIO_Degraded, &r1_bio->state);
2259 rdev_dec_pending(conf->mirrors[m].rdev,
2262 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2263 close_write(r1_bio);
2265 spin_lock_irq(&conf->device_lock);
2266 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2267 spin_unlock_irq(&conf->device_lock);
2268 md_wakeup_thread(conf->mddev->thread);
2270 raid_end_bio_io(r1_bio);
2273 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2277 struct mddev *mddev = conf->mddev;
2279 char b[BDEVNAME_SIZE];
2280 struct md_rdev *rdev;
2282 clear_bit(R1BIO_ReadError, &r1_bio->state);
2283 /* we got a read error. Maybe the drive is bad. Maybe just
2284 * the block and we can fix it.
2285 * We freeze all other IO, and try reading the block from
2286 * other devices. When we find one, we re-write
2287 * and check it that fixes the read error.
2288 * This is all done synchronously while the array is
2291 if (mddev->ro == 0) {
2292 freeze_array(conf, 1);
2293 fix_read_error(conf, r1_bio->read_disk,
2294 r1_bio->sector, r1_bio->sectors);
2295 unfreeze_array(conf);
2297 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2298 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2300 bio = r1_bio->bios[r1_bio->read_disk];
2301 bdevname(bio->bi_bdev, b);
2303 disk = read_balance(conf, r1_bio, &max_sectors);
2305 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2306 " read error for block %llu\n",
2307 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2308 raid_end_bio_io(r1_bio);
2310 const unsigned long do_sync
2311 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2313 r1_bio->bios[r1_bio->read_disk] =
2314 mddev->ro ? IO_BLOCKED : NULL;
2317 r1_bio->read_disk = disk;
2318 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2319 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2321 r1_bio->bios[r1_bio->read_disk] = bio;
2322 rdev = conf->mirrors[disk].rdev;
2323 printk_ratelimited(KERN_ERR
2324 "md/raid1:%s: redirecting sector %llu"
2325 " to other mirror: %s\n",
2327 (unsigned long long)r1_bio->sector,
2328 bdevname(rdev->bdev, b));
2329 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2330 bio->bi_bdev = rdev->bdev;
2331 bio->bi_end_io = raid1_end_read_request;
2332 bio->bi_rw = READ | do_sync;
2333 bio->bi_private = r1_bio;
2334 if (max_sectors < r1_bio->sectors) {
2335 /* Drat - have to split this up more */
2336 struct bio *mbio = r1_bio->master_bio;
2337 int sectors_handled = (r1_bio->sector + max_sectors
2338 - mbio->bi_iter.bi_sector);
2339 r1_bio->sectors = max_sectors;
2340 spin_lock_irq(&conf->device_lock);
2341 if (mbio->bi_phys_segments == 0)
2342 mbio->bi_phys_segments = 2;
2344 mbio->bi_phys_segments++;
2345 spin_unlock_irq(&conf->device_lock);
2346 generic_make_request(bio);
2349 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2351 r1_bio->master_bio = mbio;
2352 r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2354 set_bit(R1BIO_ReadError, &r1_bio->state);
2355 r1_bio->mddev = mddev;
2356 r1_bio->sector = mbio->bi_iter.bi_sector +
2361 generic_make_request(bio);
2365 static void raid1d(struct md_thread *thread)
2367 struct mddev *mddev = thread->mddev;
2368 struct r1bio *r1_bio;
2369 unsigned long flags;
2370 struct r1conf *conf = mddev->private;
2371 struct list_head *head = &conf->retry_list;
2372 struct blk_plug plug;
2374 md_check_recovery(mddev);
2376 if (!list_empty_careful(&conf->bio_end_io_list) &&
2377 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2379 spin_lock_irqsave(&conf->device_lock, flags);
2380 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2381 list_add(&tmp, &conf->bio_end_io_list);
2382 list_del_init(&conf->bio_end_io_list);
2384 spin_unlock_irqrestore(&conf->device_lock, flags);
2385 while (!list_empty(&tmp)) {
2386 r1_bio = list_first_entry(&conf->bio_end_io_list,
2387 struct r1bio, retry_list);
2388 list_del(&r1_bio->retry_list);
2389 raid_end_bio_io(r1_bio);
2393 blk_start_plug(&plug);
2396 flush_pending_writes(conf);
2398 spin_lock_irqsave(&conf->device_lock, flags);
2399 if (list_empty(head)) {
2400 spin_unlock_irqrestore(&conf->device_lock, flags);
2403 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2404 list_del(head->prev);
2406 spin_unlock_irqrestore(&conf->device_lock, flags);
2408 mddev = r1_bio->mddev;
2409 conf = mddev->private;
2410 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2411 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2412 test_bit(R1BIO_WriteError, &r1_bio->state))
2413 handle_sync_write_finished(conf, r1_bio);
2415 sync_request_write(mddev, r1_bio);
2416 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2417 test_bit(R1BIO_WriteError, &r1_bio->state))
2418 handle_write_finished(conf, r1_bio);
2419 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2420 handle_read_error(conf, r1_bio);
2422 /* just a partial read to be scheduled from separate
2425 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2428 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2429 md_check_recovery(mddev);
2431 blk_finish_plug(&plug);
2434 static int init_resync(struct r1conf *conf)
2438 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2439 BUG_ON(conf->r1buf_pool);
2440 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2442 if (!conf->r1buf_pool)
2444 conf->next_resync = 0;
2449 * perform a "sync" on one "block"
2451 * We need to make sure that no normal I/O request - particularly write
2452 * requests - conflict with active sync requests.
2454 * This is achieved by tracking pending requests and a 'barrier' concept
2455 * that can be installed to exclude normal IO requests.
2458 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
2460 struct r1conf *conf = mddev->private;
2461 struct r1bio *r1_bio;
2463 sector_t max_sector, nr_sectors;
2467 int write_targets = 0, read_targets = 0;
2468 sector_t sync_blocks;
2469 int still_degraded = 0;
2470 int good_sectors = RESYNC_SECTORS;
2471 int min_bad = 0; /* number of sectors that are bad in all devices */
2473 if (!conf->r1buf_pool)
2474 if (init_resync(conf))
2477 max_sector = mddev->dev_sectors;
2478 if (sector_nr >= max_sector) {
2479 /* If we aborted, we need to abort the
2480 * sync on the 'current' bitmap chunk (there will
2481 * only be one in raid1 resync.
2482 * We can find the current addess in mddev->curr_resync
2484 if (mddev->curr_resync < max_sector) /* aborted */
2485 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2487 else /* completed sync */
2490 bitmap_close_sync(mddev->bitmap);
2495 if (mddev->bitmap == NULL &&
2496 mddev->recovery_cp == MaxSector &&
2497 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2498 conf->fullsync == 0) {
2500 return max_sector - sector_nr;
2502 /* before building a request, check if we can skip these blocks..
2503 * This call the bitmap_start_sync doesn't actually record anything
2505 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2506 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2507 /* We can skip this block, and probably several more */
2512 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2513 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2515 raise_barrier(conf, sector_nr);
2519 * If we get a correctably read error during resync or recovery,
2520 * we might want to read from a different device. So we
2521 * flag all drives that could conceivably be read from for READ,
2522 * and any others (which will be non-In_sync devices) for WRITE.
2523 * If a read fails, we try reading from something else for which READ
2527 r1_bio->mddev = mddev;
2528 r1_bio->sector = sector_nr;
2530 set_bit(R1BIO_IsSync, &r1_bio->state);
2532 for (i = 0; i < conf->raid_disks * 2; i++) {
2533 struct md_rdev *rdev;
2534 bio = r1_bio->bios[i];
2537 rdev = rcu_dereference(conf->mirrors[i].rdev);
2539 test_bit(Faulty, &rdev->flags)) {
2540 if (i < conf->raid_disks)
2542 } else if (!test_bit(In_sync, &rdev->flags)) {
2544 bio->bi_end_io = end_sync_write;
2547 /* may need to read from here */
2548 sector_t first_bad = MaxSector;
2551 if (is_badblock(rdev, sector_nr, good_sectors,
2552 &first_bad, &bad_sectors)) {
2553 if (first_bad > sector_nr)
2554 good_sectors = first_bad - sector_nr;
2556 bad_sectors -= (sector_nr - first_bad);
2558 min_bad > bad_sectors)
2559 min_bad = bad_sectors;
2562 if (sector_nr < first_bad) {
2563 if (test_bit(WriteMostly, &rdev->flags)) {
2571 bio->bi_end_io = end_sync_read;
2573 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2574 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2575 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2577 * The device is suitable for reading (InSync),
2578 * but has bad block(s) here. Let's try to correct them,
2579 * if we are doing resync or repair. Otherwise, leave
2580 * this device alone for this sync request.
2583 bio->bi_end_io = end_sync_write;
2587 if (bio->bi_end_io) {
2588 atomic_inc(&rdev->nr_pending);
2589 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2590 bio->bi_bdev = rdev->bdev;
2591 bio->bi_private = r1_bio;
2597 r1_bio->read_disk = disk;
2599 if (read_targets == 0 && min_bad > 0) {
2600 /* These sectors are bad on all InSync devices, so we
2601 * need to mark them bad on all write targets
2604 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2605 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2606 struct md_rdev *rdev = conf->mirrors[i].rdev;
2607 ok = rdev_set_badblocks(rdev, sector_nr,
2611 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2616 /* Cannot record the badblocks, so need to
2618 * If there are multiple read targets, could just
2619 * fail the really bad ones ???
2621 conf->recovery_disabled = mddev->recovery_disabled;
2622 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2628 if (min_bad > 0 && min_bad < good_sectors) {
2629 /* only resync enough to reach the next bad->good
2631 good_sectors = min_bad;
2634 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2635 /* extra read targets are also write targets */
2636 write_targets += read_targets-1;
2638 if (write_targets == 0 || read_targets == 0) {
2639 /* There is nowhere to write, so all non-sync
2640 * drives must be failed - so we are finished
2644 max_sector = sector_nr + min_bad;
2645 rv = max_sector - sector_nr;
2651 if (max_sector > mddev->resync_max)
2652 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2653 if (max_sector > sector_nr + good_sectors)
2654 max_sector = sector_nr + good_sectors;
2659 int len = PAGE_SIZE;
2660 if (sector_nr + (len>>9) > max_sector)
2661 len = (max_sector - sector_nr) << 9;
2664 if (sync_blocks == 0) {
2665 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2666 &sync_blocks, still_degraded) &&
2668 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2670 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2671 if ((len >> 9) > sync_blocks)
2672 len = sync_blocks<<9;
2675 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2676 bio = r1_bio->bios[i];
2677 if (bio->bi_end_io) {
2678 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2679 if (bio_add_page(bio, page, len, 0) == 0) {
2681 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2684 bio = r1_bio->bios[i];
2685 if (bio->bi_end_io==NULL)
2687 /* remove last page from this bio */
2689 bio->bi_iter.bi_size -= len;
2690 bio_clear_flag(bio, BIO_SEG_VALID);
2696 nr_sectors += len>>9;
2697 sector_nr += len>>9;
2698 sync_blocks -= (len>>9);
2699 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2701 r1_bio->sectors = nr_sectors;
2703 /* For a user-requested sync, we read all readable devices and do a
2706 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2707 atomic_set(&r1_bio->remaining, read_targets);
2708 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2709 bio = r1_bio->bios[i];
2710 if (bio->bi_end_io == end_sync_read) {
2712 md_sync_acct(bio->bi_bdev, nr_sectors);
2713 generic_make_request(bio);
2717 atomic_set(&r1_bio->remaining, 1);
2718 bio = r1_bio->bios[r1_bio->read_disk];
2719 md_sync_acct(bio->bi_bdev, nr_sectors);
2720 generic_make_request(bio);
2726 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2731 return mddev->dev_sectors;
2734 static struct r1conf *setup_conf(struct mddev *mddev)
2736 struct r1conf *conf;
2738 struct raid1_info *disk;
2739 struct md_rdev *rdev;
2742 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2746 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2747 * mddev->raid_disks * 2,
2752 conf->tmppage = alloc_page(GFP_KERNEL);
2756 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2757 if (!conf->poolinfo)
2759 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2760 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2763 if (!conf->r1bio_pool)
2766 conf->poolinfo->mddev = mddev;
2769 spin_lock_init(&conf->device_lock);
2770 rdev_for_each(rdev, mddev) {
2771 struct request_queue *q;
2772 int disk_idx = rdev->raid_disk;
2773 if (disk_idx >= mddev->raid_disks
2776 if (test_bit(Replacement, &rdev->flags))
2777 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2779 disk = conf->mirrors + disk_idx;
2784 q = bdev_get_queue(rdev->bdev);
2786 disk->head_position = 0;
2787 disk->seq_start = MaxSector;
2789 conf->raid_disks = mddev->raid_disks;
2790 conf->mddev = mddev;
2791 INIT_LIST_HEAD(&conf->retry_list);
2792 INIT_LIST_HEAD(&conf->bio_end_io_list);
2794 spin_lock_init(&conf->resync_lock);
2795 init_waitqueue_head(&conf->wait_barrier);
2797 bio_list_init(&conf->pending_bio_list);
2798 conf->pending_count = 0;
2799 conf->recovery_disabled = mddev->recovery_disabled - 1;
2801 conf->start_next_window = MaxSector;
2802 conf->current_window_requests = conf->next_window_requests = 0;
2805 for (i = 0; i < conf->raid_disks * 2; i++) {
2807 disk = conf->mirrors + i;
2809 if (i < conf->raid_disks &&
2810 disk[conf->raid_disks].rdev) {
2811 /* This slot has a replacement. */
2813 /* No original, just make the replacement
2814 * a recovering spare
2817 disk[conf->raid_disks].rdev;
2818 disk[conf->raid_disks].rdev = NULL;
2819 } else if (!test_bit(In_sync, &disk->rdev->flags))
2820 /* Original is not in_sync - bad */
2825 !test_bit(In_sync, &disk->rdev->flags)) {
2826 disk->head_position = 0;
2828 (disk->rdev->saved_raid_disk < 0))
2834 conf->thread = md_register_thread(raid1d, mddev, "raid1");
2835 if (!conf->thread) {
2837 "md/raid1:%s: couldn't allocate thread\n",
2846 if (conf->r1bio_pool)
2847 mempool_destroy(conf->r1bio_pool);
2848 kfree(conf->mirrors);
2849 safe_put_page(conf->tmppage);
2850 kfree(conf->poolinfo);
2853 return ERR_PTR(err);
2856 static void raid1_free(struct mddev *mddev, void *priv);
2857 static int run(struct mddev *mddev)
2859 struct r1conf *conf;
2861 struct md_rdev *rdev;
2863 bool discard_supported = false;
2865 if (mddev->level != 1) {
2866 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2867 mdname(mddev), mddev->level);
2870 if (mddev->reshape_position != MaxSector) {
2871 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2876 * copy the already verified devices into our private RAID1
2877 * bookkeeping area. [whatever we allocate in run(),
2878 * should be freed in raid1_free()]
2880 if (mddev->private == NULL)
2881 conf = setup_conf(mddev);
2883 conf = mddev->private;
2886 return PTR_ERR(conf);
2889 blk_queue_max_write_same_sectors(mddev->queue, 0);
2891 rdev_for_each(rdev, mddev) {
2892 if (!mddev->gendisk)
2894 disk_stack_limits(mddev->gendisk, rdev->bdev,
2895 rdev->data_offset << 9);
2896 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2897 discard_supported = true;
2900 mddev->degraded = 0;
2901 for (i=0; i < conf->raid_disks; i++)
2902 if (conf->mirrors[i].rdev == NULL ||
2903 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2904 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2907 if (conf->raid_disks - mddev->degraded == 1)
2908 mddev->recovery_cp = MaxSector;
2910 if (mddev->recovery_cp != MaxSector)
2911 printk(KERN_NOTICE "md/raid1:%s: not clean"
2912 " -- starting background reconstruction\n",
2915 "md/raid1:%s: active with %d out of %d mirrors\n",
2916 mdname(mddev), mddev->raid_disks - mddev->degraded,
2920 * Ok, everything is just fine now
2922 mddev->thread = conf->thread;
2923 conf->thread = NULL;
2924 mddev->private = conf;
2926 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2929 if (discard_supported)
2930 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2933 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2937 ret = md_integrity_register(mddev);
2939 md_unregister_thread(&mddev->thread);
2940 raid1_free(mddev, conf);
2945 static void raid1_free(struct mddev *mddev, void *priv)
2947 struct r1conf *conf = priv;
2949 if (conf->r1bio_pool)
2950 mempool_destroy(conf->r1bio_pool);
2951 kfree(conf->mirrors);
2952 safe_put_page(conf->tmppage);
2953 kfree(conf->poolinfo);
2957 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2959 /* no resync is happening, and there is enough space
2960 * on all devices, so we can resize.
2961 * We need to make sure resync covers any new space.
2962 * If the array is shrinking we should possibly wait until
2963 * any io in the removed space completes, but it hardly seems
2966 sector_t newsize = raid1_size(mddev, sectors, 0);
2967 if (mddev->external_size &&
2968 mddev->array_sectors > newsize)
2970 if (mddev->bitmap) {
2971 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2975 md_set_array_sectors(mddev, newsize);
2976 set_capacity(mddev->gendisk, mddev->array_sectors);
2977 revalidate_disk(mddev->gendisk);
2978 if (sectors > mddev->dev_sectors &&
2979 mddev->recovery_cp > mddev->dev_sectors) {
2980 mddev->recovery_cp = mddev->dev_sectors;
2981 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2983 mddev->dev_sectors = sectors;
2984 mddev->resync_max_sectors = sectors;
2988 static int raid1_reshape(struct mddev *mddev)
2991 * 1/ resize the r1bio_pool
2992 * 2/ resize conf->mirrors
2994 * We allocate a new r1bio_pool if we can.
2995 * Then raise a device barrier and wait until all IO stops.
2996 * Then resize conf->mirrors and swap in the new r1bio pool.
2998 * At the same time, we "pack" the devices so that all the missing
2999 * devices have the higher raid_disk numbers.
3001 mempool_t *newpool, *oldpool;
3002 struct pool_info *newpoolinfo;
3003 struct raid1_info *newmirrors;
3004 struct r1conf *conf = mddev->private;
3005 int cnt, raid_disks;
3006 unsigned long flags;
3009 /* Cannot change chunk_size, layout, or level */
3010 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3011 mddev->layout != mddev->new_layout ||
3012 mddev->level != mddev->new_level) {
3013 mddev->new_chunk_sectors = mddev->chunk_sectors;
3014 mddev->new_layout = mddev->layout;
3015 mddev->new_level = mddev->level;
3019 err = md_allow_write(mddev);
3023 raid_disks = mddev->raid_disks + mddev->delta_disks;
3025 if (raid_disks < conf->raid_disks) {
3027 for (d= 0; d < conf->raid_disks; d++)
3028 if (conf->mirrors[d].rdev)
3030 if (cnt > raid_disks)
3034 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3037 newpoolinfo->mddev = mddev;
3038 newpoolinfo->raid_disks = raid_disks * 2;
3040 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3041 r1bio_pool_free, newpoolinfo);
3046 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3050 mempool_destroy(newpool);
3054 freeze_array(conf, 0);
3056 /* ok, everything is stopped */
3057 oldpool = conf->r1bio_pool;
3058 conf->r1bio_pool = newpool;
3060 for (d = d2 = 0; d < conf->raid_disks; d++) {
3061 struct md_rdev *rdev = conf->mirrors[d].rdev;
3062 if (rdev && rdev->raid_disk != d2) {
3063 sysfs_unlink_rdev(mddev, rdev);
3064 rdev->raid_disk = d2;
3065 sysfs_unlink_rdev(mddev, rdev);
3066 if (sysfs_link_rdev(mddev, rdev))
3068 "md/raid1:%s: cannot register rd%d\n",
3069 mdname(mddev), rdev->raid_disk);
3072 newmirrors[d2++].rdev = rdev;
3074 kfree(conf->mirrors);
3075 conf->mirrors = newmirrors;
3076 kfree(conf->poolinfo);
3077 conf->poolinfo = newpoolinfo;
3079 spin_lock_irqsave(&conf->device_lock, flags);
3080 mddev->degraded += (raid_disks - conf->raid_disks);
3081 spin_unlock_irqrestore(&conf->device_lock, flags);
3082 conf->raid_disks = mddev->raid_disks = raid_disks;
3083 mddev->delta_disks = 0;
3085 unfreeze_array(conf);
3087 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3088 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3089 md_wakeup_thread(mddev->thread);
3091 mempool_destroy(oldpool);
3095 static void raid1_quiesce(struct mddev *mddev, int state)
3097 struct r1conf *conf = mddev->private;
3100 case 2: /* wake for suspend */
3101 wake_up(&conf->wait_barrier);
3104 freeze_array(conf, 0);
3107 unfreeze_array(conf);
3112 static void *raid1_takeover(struct mddev *mddev)
3114 /* raid1 can take over:
3115 * raid5 with 2 devices, any layout or chunk size
3117 if (mddev->level == 5 && mddev->raid_disks == 2) {
3118 struct r1conf *conf;
3119 mddev->new_level = 1;
3120 mddev->new_layout = 0;
3121 mddev->new_chunk_sectors = 0;
3122 conf = setup_conf(mddev);
3124 /* Array must appear to be quiesced */
3125 conf->array_frozen = 1;
3128 return ERR_PTR(-EINVAL);
3131 static struct md_personality raid1_personality =
3135 .owner = THIS_MODULE,
3136 .make_request = make_request,
3140 .error_handler = error,
3141 .hot_add_disk = raid1_add_disk,
3142 .hot_remove_disk= raid1_remove_disk,
3143 .spare_active = raid1_spare_active,
3144 .sync_request = sync_request,
3145 .resize = raid1_resize,
3147 .check_reshape = raid1_reshape,
3148 .quiesce = raid1_quiesce,
3149 .takeover = raid1_takeover,
3150 .congested = raid1_congested,
3153 static int __init raid_init(void)
3155 return register_md_personality(&raid1_personality);
3158 static void raid_exit(void)
3160 unregister_md_personality(&raid1_personality);
3163 module_init(raid_init);
3164 module_exit(raid_exit);
3165 MODULE_LICENSE("GPL");
3166 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3167 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3168 MODULE_ALIAS("md-raid1");
3169 MODULE_ALIAS("md-level-1");
3171 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);