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/seq_file.h>
38 #include <linux/ratelimit.h>
44 * Number of guaranteed r1bios in case of extreme VM load:
46 #define NR_RAID1_BIOS 256
48 /* When there are this many requests queue to be written by
49 * the raid1 thread, we become 'congested' to provide back-pressure
52 static int max_queued_requests = 1024;
54 static void allow_barrier(struct r1conf *conf);
55 static void lower_barrier(struct r1conf *conf);
57 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
59 struct pool_info *pi = data;
60 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
62 /* allocate a r1bio with room for raid_disks entries in the bios array */
63 return kzalloc(size, gfp_flags);
66 static void r1bio_pool_free(void *r1_bio, void *data)
71 #define RESYNC_BLOCK_SIZE (64*1024)
72 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
73 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
74 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
75 #define RESYNC_WINDOW (2048*1024)
77 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
79 struct pool_info *pi = data;
85 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
90 * Allocate bios : 1 for reading, n-1 for writing
92 for (j = pi->raid_disks ; j-- ; ) {
93 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
96 r1_bio->bios[j] = bio;
99 * Allocate RESYNC_PAGES data pages and attach them to
101 * If this is a user-requested check/repair, allocate
102 * RESYNC_PAGES for each bio.
104 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
109 bio = r1_bio->bios[j];
110 for (i = 0; i < RESYNC_PAGES; i++) {
111 page = alloc_page(gfp_flags);
115 bio->bi_io_vec[i].bv_page = page;
119 /* If not user-requests, copy the page pointers to all bios */
120 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
121 for (i=0; i<RESYNC_PAGES ; i++)
122 for (j=1; j<pi->raid_disks; j++)
123 r1_bio->bios[j]->bi_io_vec[i].bv_page =
124 r1_bio->bios[0]->bi_io_vec[i].bv_page;
127 r1_bio->master_bio = NULL;
132 for (j=0 ; j < pi->raid_disks; j++)
133 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
134 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
137 while ( ++j < pi->raid_disks )
138 bio_put(r1_bio->bios[j]);
139 r1bio_pool_free(r1_bio, data);
143 static void r1buf_pool_free(void *__r1_bio, void *data)
145 struct pool_info *pi = data;
147 struct r1bio *r1bio = __r1_bio;
149 for (i = 0; i < RESYNC_PAGES; i++)
150 for (j = pi->raid_disks; j-- ;) {
152 r1bio->bios[j]->bi_io_vec[i].bv_page !=
153 r1bio->bios[0]->bi_io_vec[i].bv_page)
154 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
156 for (i=0 ; i < pi->raid_disks; i++)
157 bio_put(r1bio->bios[i]);
159 r1bio_pool_free(r1bio, data);
162 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
166 for (i = 0; i < conf->raid_disks; i++) {
167 struct bio **bio = r1_bio->bios + i;
168 if (!BIO_SPECIAL(*bio))
174 static void free_r1bio(struct r1bio *r1_bio)
176 struct r1conf *conf = r1_bio->mddev->private;
178 put_all_bios(conf, r1_bio);
179 mempool_free(r1_bio, conf->r1bio_pool);
182 static void put_buf(struct r1bio *r1_bio)
184 struct r1conf *conf = r1_bio->mddev->private;
187 for (i=0; i<conf->raid_disks; i++) {
188 struct bio *bio = r1_bio->bios[i];
190 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
193 mempool_free(r1_bio, conf->r1buf_pool);
198 static void reschedule_retry(struct r1bio *r1_bio)
201 struct mddev *mddev = r1_bio->mddev;
202 struct r1conf *conf = mddev->private;
204 spin_lock_irqsave(&conf->device_lock, flags);
205 list_add(&r1_bio->retry_list, &conf->retry_list);
207 spin_unlock_irqrestore(&conf->device_lock, flags);
209 wake_up(&conf->wait_barrier);
210 md_wakeup_thread(mddev->thread);
214 * raid_end_bio_io() is called when we have finished servicing a mirrored
215 * operation and are ready to return a success/failure code to the buffer
218 static void call_bio_endio(struct r1bio *r1_bio)
220 struct bio *bio = r1_bio->master_bio;
222 struct r1conf *conf = r1_bio->mddev->private;
224 if (bio->bi_phys_segments) {
226 spin_lock_irqsave(&conf->device_lock, flags);
227 bio->bi_phys_segments--;
228 done = (bio->bi_phys_segments == 0);
229 spin_unlock_irqrestore(&conf->device_lock, flags);
233 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
234 clear_bit(BIO_UPTODATE, &bio->bi_flags);
238 * Wake up any possible resync thread that waits for the device
245 static void raid_end_bio_io(struct r1bio *r1_bio)
247 struct bio *bio = r1_bio->master_bio;
249 /* if nobody has done the final endio yet, do it now */
250 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
251 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
252 (bio_data_dir(bio) == WRITE) ? "write" : "read",
253 (unsigned long long) bio->bi_sector,
254 (unsigned long long) bio->bi_sector +
255 (bio->bi_size >> 9) - 1);
257 call_bio_endio(r1_bio);
263 * Update disk head position estimator based on IRQ completion info.
265 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
267 struct r1conf *conf = r1_bio->mddev->private;
269 conf->mirrors[disk].head_position =
270 r1_bio->sector + (r1_bio->sectors);
274 * Find the disk number which triggered given bio
276 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
279 int raid_disks = r1_bio->mddev->raid_disks;
281 for (mirror = 0; mirror < raid_disks; mirror++)
282 if (r1_bio->bios[mirror] == bio)
285 BUG_ON(mirror == raid_disks);
286 update_head_pos(mirror, r1_bio);
291 static void raid1_end_read_request(struct bio *bio, int error)
293 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
294 struct r1bio *r1_bio = bio->bi_private;
296 struct r1conf *conf = r1_bio->mddev->private;
298 mirror = r1_bio->read_disk;
300 * this branch is our 'one mirror IO has finished' event handler:
302 update_head_pos(mirror, r1_bio);
305 set_bit(R1BIO_Uptodate, &r1_bio->state);
307 /* If all other devices have failed, we want to return
308 * the error upwards rather than fail the last device.
309 * Here we redefine "uptodate" to mean "Don't want to retry"
312 spin_lock_irqsave(&conf->device_lock, flags);
313 if (r1_bio->mddev->degraded == conf->raid_disks ||
314 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
315 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
317 spin_unlock_irqrestore(&conf->device_lock, flags);
321 raid_end_bio_io(r1_bio);
326 char b[BDEVNAME_SIZE];
328 KERN_ERR "md/raid1:%s: %s: "
329 "rescheduling sector %llu\n",
331 bdevname(conf->mirrors[mirror].rdev->bdev,
333 (unsigned long long)r1_bio->sector);
334 set_bit(R1BIO_ReadError, &r1_bio->state);
335 reschedule_retry(r1_bio);
338 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
341 static void close_write(struct r1bio *r1_bio)
343 /* it really is the end of this request */
344 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
345 /* free extra copy of the data pages */
346 int i = r1_bio->behind_page_count;
348 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
349 kfree(r1_bio->behind_bvecs);
350 r1_bio->behind_bvecs = NULL;
352 /* clear the bitmap if all writes complete successfully */
353 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
355 !test_bit(R1BIO_Degraded, &r1_bio->state),
356 test_bit(R1BIO_BehindIO, &r1_bio->state));
357 md_write_end(r1_bio->mddev);
360 static void r1_bio_write_done(struct r1bio *r1_bio)
362 if (!atomic_dec_and_test(&r1_bio->remaining))
365 if (test_bit(R1BIO_WriteError, &r1_bio->state))
366 reschedule_retry(r1_bio);
369 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
370 reschedule_retry(r1_bio);
372 raid_end_bio_io(r1_bio);
376 static void raid1_end_write_request(struct bio *bio, int error)
378 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
379 struct r1bio *r1_bio = bio->bi_private;
380 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
381 struct r1conf *conf = r1_bio->mddev->private;
382 struct bio *to_put = NULL;
384 mirror = find_bio_disk(r1_bio, bio);
387 * 'one mirror IO has finished' event handler:
390 set_bit(WriteErrorSeen,
391 &conf->mirrors[mirror].rdev->flags);
392 set_bit(R1BIO_WriteError, &r1_bio->state);
395 * Set R1BIO_Uptodate in our master bio, so that we
396 * will return a good error code for to the higher
397 * levels even if IO on some other mirrored buffer
400 * The 'master' represents the composite IO operation
401 * to user-side. So if something waits for IO, then it
402 * will wait for the 'master' bio.
407 r1_bio->bios[mirror] = NULL;
409 set_bit(R1BIO_Uptodate, &r1_bio->state);
411 /* Maybe we can clear some bad blocks. */
412 if (is_badblock(conf->mirrors[mirror].rdev,
413 r1_bio->sector, r1_bio->sectors,
414 &first_bad, &bad_sectors)) {
415 r1_bio->bios[mirror] = IO_MADE_GOOD;
416 set_bit(R1BIO_MadeGood, &r1_bio->state);
421 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
422 atomic_dec(&r1_bio->behind_remaining);
425 * In behind mode, we ACK the master bio once the I/O
426 * has safely reached all non-writemostly
427 * disks. Setting the Returned bit ensures that this
428 * gets done only once -- we don't ever want to return
429 * -EIO here, instead we'll wait
431 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
432 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
433 /* Maybe we can return now */
434 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
435 struct bio *mbio = r1_bio->master_bio;
436 pr_debug("raid1: behind end write sectors"
438 (unsigned long long) mbio->bi_sector,
439 (unsigned long long) mbio->bi_sector +
440 (mbio->bi_size >> 9) - 1);
441 call_bio_endio(r1_bio);
445 if (r1_bio->bios[mirror] == NULL)
446 rdev_dec_pending(conf->mirrors[mirror].rdev,
450 * Let's see if all mirrored write operations have finished
453 r1_bio_write_done(r1_bio);
461 * This routine returns the disk from which the requested read should
462 * be done. There is a per-array 'next expected sequential IO' sector
463 * number - if this matches on the next IO then we use the last disk.
464 * There is also a per-disk 'last know head position' sector that is
465 * maintained from IRQ contexts, both the normal and the resync IO
466 * completion handlers update this position correctly. If there is no
467 * perfect sequential match then we pick the disk whose head is closest.
469 * If there are 2 mirrors in the same 2 devices, performance degrades
470 * because position is mirror, not device based.
472 * The rdev for the device selected will have nr_pending incremented.
474 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
476 const sector_t this_sector = r1_bio->sector;
478 int best_good_sectors;
483 struct md_rdev *rdev;
488 * Check if we can balance. We can balance on the whole
489 * device if no resync is going on, or below the resync window.
490 * We take the first readable disk when above the resync window.
493 sectors = r1_bio->sectors;
495 best_dist = MaxSector;
496 best_good_sectors = 0;
498 if (conf->mddev->recovery_cp < MaxSector &&
499 (this_sector + sectors >= conf->next_resync)) {
504 start_disk = conf->last_used;
507 for (i = 0 ; i < conf->raid_disks ; i++) {
512 int disk = start_disk + i;
513 if (disk >= conf->raid_disks)
514 disk -= conf->raid_disks;
516 rdev = rcu_dereference(conf->mirrors[disk].rdev);
517 if (r1_bio->bios[disk] == IO_BLOCKED
519 || test_bit(Faulty, &rdev->flags))
521 if (!test_bit(In_sync, &rdev->flags) &&
522 rdev->recovery_offset < this_sector + sectors)
524 if (test_bit(WriteMostly, &rdev->flags)) {
525 /* Don't balance among write-mostly, just
526 * use the first as a last resort */
531 /* This is a reasonable device to use. It might
534 if (is_badblock(rdev, this_sector, sectors,
535 &first_bad, &bad_sectors)) {
536 if (best_dist < MaxSector)
537 /* already have a better device */
539 if (first_bad <= this_sector) {
540 /* cannot read here. If this is the 'primary'
541 * device, then we must not read beyond
542 * bad_sectors from another device..
544 bad_sectors -= (this_sector - first_bad);
545 if (choose_first && sectors > bad_sectors)
546 sectors = bad_sectors;
547 if (best_good_sectors > sectors)
548 best_good_sectors = sectors;
551 sector_t good_sectors = first_bad - this_sector;
552 if (good_sectors > best_good_sectors) {
553 best_good_sectors = good_sectors;
561 best_good_sectors = sectors;
563 dist = abs(this_sector - conf->mirrors[disk].head_position);
565 /* Don't change to another disk for sequential reads */
566 || conf->next_seq_sect == this_sector
568 /* If device is idle, use it */
569 || atomic_read(&rdev->nr_pending) == 0) {
573 if (dist < best_dist) {
579 if (best_disk >= 0) {
580 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
583 atomic_inc(&rdev->nr_pending);
584 if (test_bit(Faulty, &rdev->flags)) {
585 /* cannot risk returning a device that failed
586 * before we inc'ed nr_pending
588 rdev_dec_pending(rdev, conf->mddev);
591 sectors = best_good_sectors;
592 conf->next_seq_sect = this_sector + sectors;
593 conf->last_used = best_disk;
596 *max_sectors = sectors;
601 int md_raid1_congested(struct mddev *mddev, int bits)
603 struct r1conf *conf = mddev->private;
606 if ((bits & (1 << BDI_async_congested)) &&
607 conf->pending_count >= max_queued_requests)
611 for (i = 0; i < mddev->raid_disks; i++) {
612 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
613 if (rdev && !test_bit(Faulty, &rdev->flags)) {
614 struct request_queue *q = bdev_get_queue(rdev->bdev);
618 /* Note the '|| 1' - when read_balance prefers
619 * non-congested targets, it can be removed
621 if ((bits & (1<<BDI_async_congested)) || 1)
622 ret |= bdi_congested(&q->backing_dev_info, bits);
624 ret &= bdi_congested(&q->backing_dev_info, bits);
630 EXPORT_SYMBOL_GPL(md_raid1_congested);
632 static int raid1_congested(void *data, int bits)
634 struct mddev *mddev = data;
636 return mddev_congested(mddev, bits) ||
637 md_raid1_congested(mddev, bits);
640 static void flush_pending_writes(struct r1conf *conf)
642 /* Any writes that have been queued but are awaiting
643 * bitmap updates get flushed here.
645 spin_lock_irq(&conf->device_lock);
647 if (conf->pending_bio_list.head) {
649 bio = bio_list_get(&conf->pending_bio_list);
650 conf->pending_count = 0;
651 spin_unlock_irq(&conf->device_lock);
652 /* flush any pending bitmap writes to
653 * disk before proceeding w/ I/O */
654 bitmap_unplug(conf->mddev->bitmap);
655 wake_up(&conf->wait_barrier);
657 while (bio) { /* submit pending writes */
658 struct bio *next = bio->bi_next;
660 generic_make_request(bio);
664 spin_unlock_irq(&conf->device_lock);
668 * Sometimes we need to suspend IO while we do something else,
669 * either some resync/recovery, or reconfigure the array.
670 * To do this we raise a 'barrier'.
671 * The 'barrier' is a counter that can be raised multiple times
672 * to count how many activities are happening which preclude
674 * We can only raise the barrier if there is no pending IO.
675 * i.e. if nr_pending == 0.
676 * We choose only to raise the barrier if no-one is waiting for the
677 * barrier to go down. This means that as soon as an IO request
678 * is ready, no other operations which require a barrier will start
679 * until the IO request has had a chance.
681 * So: regular IO calls 'wait_barrier'. When that returns there
682 * is no backgroup IO happening, It must arrange to call
683 * allow_barrier when it has finished its IO.
684 * backgroup IO calls must call raise_barrier. Once that returns
685 * there is no normal IO happeing. It must arrange to call
686 * lower_barrier when the particular background IO completes.
688 #define RESYNC_DEPTH 32
690 static void raise_barrier(struct r1conf *conf)
692 spin_lock_irq(&conf->resync_lock);
694 /* Wait until no block IO is waiting */
695 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
696 conf->resync_lock, );
698 /* block any new IO from starting */
701 /* Now wait for all pending IO to complete */
702 wait_event_lock_irq(conf->wait_barrier,
703 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
704 conf->resync_lock, );
706 spin_unlock_irq(&conf->resync_lock);
709 static void lower_barrier(struct r1conf *conf)
712 BUG_ON(conf->barrier <= 0);
713 spin_lock_irqsave(&conf->resync_lock, flags);
715 spin_unlock_irqrestore(&conf->resync_lock, flags);
716 wake_up(&conf->wait_barrier);
719 static void wait_barrier(struct r1conf *conf)
721 spin_lock_irq(&conf->resync_lock);
724 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
730 spin_unlock_irq(&conf->resync_lock);
733 static void allow_barrier(struct r1conf *conf)
736 spin_lock_irqsave(&conf->resync_lock, flags);
738 spin_unlock_irqrestore(&conf->resync_lock, flags);
739 wake_up(&conf->wait_barrier);
742 static void freeze_array(struct r1conf *conf)
744 /* stop syncio and normal IO and wait for everything to
746 * We increment barrier and nr_waiting, and then
747 * wait until nr_pending match nr_queued+1
748 * This is called in the context of one normal IO request
749 * that has failed. Thus any sync request that might be pending
750 * will be blocked by nr_pending, and we need to wait for
751 * pending IO requests to complete or be queued for re-try.
752 * Thus the number queued (nr_queued) plus this request (1)
753 * must match the number of pending IOs (nr_pending) before
756 spin_lock_irq(&conf->resync_lock);
759 wait_event_lock_irq(conf->wait_barrier,
760 conf->nr_pending == conf->nr_queued+1,
762 flush_pending_writes(conf));
763 spin_unlock_irq(&conf->resync_lock);
765 static void unfreeze_array(struct r1conf *conf)
767 /* reverse the effect of the freeze */
768 spin_lock_irq(&conf->resync_lock);
771 wake_up(&conf->wait_barrier);
772 spin_unlock_irq(&conf->resync_lock);
776 /* duplicate the data pages for behind I/O
778 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
781 struct bio_vec *bvec;
782 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
784 if (unlikely(!bvecs))
787 bio_for_each_segment(bvec, bio, i) {
789 bvecs[i].bv_page = alloc_page(GFP_NOIO);
790 if (unlikely(!bvecs[i].bv_page))
792 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
793 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
794 kunmap(bvecs[i].bv_page);
795 kunmap(bvec->bv_page);
797 r1_bio->behind_bvecs = bvecs;
798 r1_bio->behind_page_count = bio->bi_vcnt;
799 set_bit(R1BIO_BehindIO, &r1_bio->state);
803 for (i = 0; i < bio->bi_vcnt; i++)
804 if (bvecs[i].bv_page)
805 put_page(bvecs[i].bv_page);
807 pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
810 static int make_request(struct mddev *mddev, struct bio * bio)
812 struct r1conf *conf = mddev->private;
813 struct mirror_info *mirror;
814 struct r1bio *r1_bio;
815 struct bio *read_bio;
817 struct bitmap *bitmap;
819 const int rw = bio_data_dir(bio);
820 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
821 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
822 struct md_rdev *blocked_rdev;
829 * Register the new request and wait if the reconstruction
830 * thread has put up a bar for new requests.
831 * Continue immediately if no resync is active currently.
834 md_write_start(mddev, bio); /* wait on superblock update early */
836 if (bio_data_dir(bio) == WRITE &&
837 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
838 bio->bi_sector < mddev->suspend_hi) {
839 /* As the suspend_* range is controlled by
840 * userspace, we want an interruptible
845 flush_signals(current);
846 prepare_to_wait(&conf->wait_barrier,
847 &w, TASK_INTERRUPTIBLE);
848 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
849 bio->bi_sector >= mddev->suspend_hi)
853 finish_wait(&conf->wait_barrier, &w);
858 bitmap = mddev->bitmap;
861 * make_request() can abort the operation when READA is being
862 * used and no empty request is available.
865 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
867 r1_bio->master_bio = bio;
868 r1_bio->sectors = bio->bi_size >> 9;
870 r1_bio->mddev = mddev;
871 r1_bio->sector = bio->bi_sector;
873 /* We might need to issue multiple reads to different
874 * devices if there are bad blocks around, so we keep
875 * track of the number of reads in bio->bi_phys_segments.
876 * If this is 0, there is only one r1_bio and no locking
877 * will be needed when requests complete. If it is
878 * non-zero, then it is the number of not-completed requests.
880 bio->bi_phys_segments = 0;
881 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
885 * read balancing logic:
890 rdisk = read_balance(conf, r1_bio, &max_sectors);
893 /* couldn't find anywhere to read from */
894 raid_end_bio_io(r1_bio);
897 mirror = conf->mirrors + rdisk;
899 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
901 /* Reading from a write-mostly device must
902 * take care not to over-take any writes
905 wait_event(bitmap->behind_wait,
906 atomic_read(&bitmap->behind_writes) == 0);
908 r1_bio->read_disk = rdisk;
910 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
911 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
914 r1_bio->bios[rdisk] = read_bio;
916 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
917 read_bio->bi_bdev = mirror->rdev->bdev;
918 read_bio->bi_end_io = raid1_end_read_request;
919 read_bio->bi_rw = READ | do_sync;
920 read_bio->bi_private = r1_bio;
922 if (max_sectors < r1_bio->sectors) {
923 /* could not read all from this device, so we will
924 * need another r1_bio.
927 sectors_handled = (r1_bio->sector + max_sectors
929 r1_bio->sectors = max_sectors;
930 spin_lock_irq(&conf->device_lock);
931 if (bio->bi_phys_segments == 0)
932 bio->bi_phys_segments = 2;
934 bio->bi_phys_segments++;
935 spin_unlock_irq(&conf->device_lock);
936 /* Cannot call generic_make_request directly
937 * as that will be queued in __make_request
938 * and subsequent mempool_alloc might block waiting
939 * for it. So hand bio over to raid1d.
941 reschedule_retry(r1_bio);
943 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
945 r1_bio->master_bio = bio;
946 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
948 r1_bio->mddev = mddev;
949 r1_bio->sector = bio->bi_sector + sectors_handled;
952 generic_make_request(read_bio);
959 if (conf->pending_count >= max_queued_requests) {
960 md_wakeup_thread(mddev->thread);
961 wait_event(conf->wait_barrier,
962 conf->pending_count < max_queued_requests);
964 /* first select target devices under rcu_lock and
965 * inc refcount on their rdev. Record them by setting
967 * If there are known/acknowledged bad blocks on any device on
968 * which we have seen a write error, we want to avoid writing those
970 * This potentially requires several writes to write around
971 * the bad blocks. Each set of writes gets it's own r1bio
972 * with a set of bios attached.
974 plugged = mddev_check_plugged(mddev);
976 disks = conf->raid_disks;
980 max_sectors = r1_bio->sectors;
981 for (i = 0; i < disks; i++) {
982 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
983 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
984 atomic_inc(&rdev->nr_pending);
988 r1_bio->bios[i] = NULL;
989 if (!rdev || test_bit(Faulty, &rdev->flags)) {
990 set_bit(R1BIO_Degraded, &r1_bio->state);
994 atomic_inc(&rdev->nr_pending);
995 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1000 is_bad = is_badblock(rdev, r1_bio->sector,
1002 &first_bad, &bad_sectors);
1004 /* mustn't write here until the bad block is
1006 set_bit(BlockedBadBlocks, &rdev->flags);
1007 blocked_rdev = rdev;
1010 if (is_bad && first_bad <= r1_bio->sector) {
1011 /* Cannot write here at all */
1012 bad_sectors -= (r1_bio->sector - first_bad);
1013 if (bad_sectors < max_sectors)
1014 /* mustn't write more than bad_sectors
1015 * to other devices yet
1017 max_sectors = bad_sectors;
1018 rdev_dec_pending(rdev, mddev);
1019 /* We don't set R1BIO_Degraded as that
1020 * only applies if the disk is
1021 * missing, so it might be re-added,
1022 * and we want to know to recover this
1024 * In this case the device is here,
1025 * and the fact that this chunk is not
1026 * in-sync is recorded in the bad
1032 int good_sectors = first_bad - r1_bio->sector;
1033 if (good_sectors < max_sectors)
1034 max_sectors = good_sectors;
1037 r1_bio->bios[i] = bio;
1041 if (unlikely(blocked_rdev)) {
1042 /* Wait for this device to become unblocked */
1045 for (j = 0; j < i; j++)
1046 if (r1_bio->bios[j])
1047 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1049 allow_barrier(conf);
1050 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1055 if (max_sectors < r1_bio->sectors) {
1056 /* We are splitting this write into multiple parts, so
1057 * we need to prepare for allocating another r1_bio.
1059 r1_bio->sectors = max_sectors;
1060 spin_lock_irq(&conf->device_lock);
1061 if (bio->bi_phys_segments == 0)
1062 bio->bi_phys_segments = 2;
1064 bio->bi_phys_segments++;
1065 spin_unlock_irq(&conf->device_lock);
1067 sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1069 atomic_set(&r1_bio->remaining, 1);
1070 atomic_set(&r1_bio->behind_remaining, 0);
1073 for (i = 0; i < disks; i++) {
1075 if (!r1_bio->bios[i])
1078 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1079 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1083 * Not if there are too many, or cannot
1084 * allocate memory, or a reader on WriteMostly
1085 * is waiting for behind writes to flush */
1087 (atomic_read(&bitmap->behind_writes)
1088 < mddev->bitmap_info.max_write_behind) &&
1089 !waitqueue_active(&bitmap->behind_wait))
1090 alloc_behind_pages(mbio, r1_bio);
1092 bitmap_startwrite(bitmap, r1_bio->sector,
1094 test_bit(R1BIO_BehindIO,
1098 if (r1_bio->behind_bvecs) {
1099 struct bio_vec *bvec;
1102 /* Yes, I really want the '__' version so that
1103 * we clear any unused pointer in the io_vec, rather
1104 * than leave them unchanged. This is important
1105 * because when we come to free the pages, we won't
1106 * know the original bi_idx, so we just free
1109 __bio_for_each_segment(bvec, mbio, j, 0)
1110 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1111 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1112 atomic_inc(&r1_bio->behind_remaining);
1115 r1_bio->bios[i] = mbio;
1117 mbio->bi_sector = (r1_bio->sector +
1118 conf->mirrors[i].rdev->data_offset);
1119 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1120 mbio->bi_end_io = raid1_end_write_request;
1121 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1122 mbio->bi_private = r1_bio;
1124 atomic_inc(&r1_bio->remaining);
1125 spin_lock_irqsave(&conf->device_lock, flags);
1126 bio_list_add(&conf->pending_bio_list, mbio);
1127 conf->pending_count++;
1128 spin_unlock_irqrestore(&conf->device_lock, flags);
1130 /* Mustn't call r1_bio_write_done before this next test,
1131 * as it could result in the bio being freed.
1133 if (sectors_handled < (bio->bi_size >> 9)) {
1134 r1_bio_write_done(r1_bio);
1135 /* We need another r1_bio. It has already been counted
1136 * in bio->bi_phys_segments
1138 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1139 r1_bio->master_bio = bio;
1140 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1142 r1_bio->mddev = mddev;
1143 r1_bio->sector = bio->bi_sector + sectors_handled;
1147 r1_bio_write_done(r1_bio);
1149 /* In case raid1d snuck in to freeze_array */
1150 wake_up(&conf->wait_barrier);
1152 if (do_sync || !bitmap || !plugged)
1153 md_wakeup_thread(mddev->thread);
1158 static void status(struct seq_file *seq, struct mddev *mddev)
1160 struct r1conf *conf = mddev->private;
1163 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1164 conf->raid_disks - mddev->degraded);
1166 for (i = 0; i < conf->raid_disks; i++) {
1167 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1168 seq_printf(seq, "%s",
1169 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1172 seq_printf(seq, "]");
1176 static void error(struct mddev *mddev, struct md_rdev *rdev)
1178 char b[BDEVNAME_SIZE];
1179 struct r1conf *conf = mddev->private;
1182 * If it is not operational, then we have already marked it as dead
1183 * else if it is the last working disks, ignore the error, let the
1184 * next level up know.
1185 * else mark the drive as failed
1187 if (test_bit(In_sync, &rdev->flags)
1188 && (conf->raid_disks - mddev->degraded) == 1) {
1190 * Don't fail the drive, act as though we were just a
1191 * normal single drive.
1192 * However don't try a recovery from this drive as
1193 * it is very likely to fail.
1195 conf->recovery_disabled = mddev->recovery_disabled;
1198 set_bit(Blocked, &rdev->flags);
1199 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1200 unsigned long flags;
1201 spin_lock_irqsave(&conf->device_lock, flags);
1203 set_bit(Faulty, &rdev->flags);
1204 spin_unlock_irqrestore(&conf->device_lock, flags);
1206 * if recovery is running, make sure it aborts.
1208 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1210 set_bit(Faulty, &rdev->flags);
1211 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1213 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1214 "md/raid1:%s: Operation continuing on %d devices.\n",
1215 mdname(mddev), bdevname(rdev->bdev, b),
1216 mdname(mddev), conf->raid_disks - mddev->degraded);
1219 static void print_conf(struct r1conf *conf)
1223 printk(KERN_DEBUG "RAID1 conf printout:\n");
1225 printk(KERN_DEBUG "(!conf)\n");
1228 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1232 for (i = 0; i < conf->raid_disks; i++) {
1233 char b[BDEVNAME_SIZE];
1234 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1236 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1237 i, !test_bit(In_sync, &rdev->flags),
1238 !test_bit(Faulty, &rdev->flags),
1239 bdevname(rdev->bdev,b));
1244 static void close_sync(struct r1conf *conf)
1247 allow_barrier(conf);
1249 mempool_destroy(conf->r1buf_pool);
1250 conf->r1buf_pool = NULL;
1253 static int raid1_spare_active(struct mddev *mddev)
1256 struct r1conf *conf = mddev->private;
1258 unsigned long flags;
1261 * Find all failed disks within the RAID1 configuration
1262 * and mark them readable.
1263 * Called under mddev lock, so rcu protection not needed.
1265 for (i = 0; i < conf->raid_disks; i++) {
1266 struct md_rdev *rdev = conf->mirrors[i].rdev;
1268 && !test_bit(Faulty, &rdev->flags)
1269 && !test_and_set_bit(In_sync, &rdev->flags)) {
1271 sysfs_notify_dirent_safe(rdev->sysfs_state);
1274 spin_lock_irqsave(&conf->device_lock, flags);
1275 mddev->degraded -= count;
1276 spin_unlock_irqrestore(&conf->device_lock, flags);
1283 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1285 struct r1conf *conf = mddev->private;
1288 struct mirror_info *p;
1290 int last = mddev->raid_disks - 1;
1292 if (mddev->recovery_disabled == conf->recovery_disabled)
1295 if (rdev->raid_disk >= 0)
1296 first = last = rdev->raid_disk;
1298 for (mirror = first; mirror <= last; mirror++)
1299 if ( !(p=conf->mirrors+mirror)->rdev) {
1301 disk_stack_limits(mddev->gendisk, rdev->bdev,
1302 rdev->data_offset << 9);
1303 /* as we don't honour merge_bvec_fn, we must
1304 * never risk violating it, so limit
1305 * ->max_segments to one lying with a single
1306 * page, as a one page request is never in
1309 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1310 blk_queue_max_segments(mddev->queue, 1);
1311 blk_queue_segment_boundary(mddev->queue,
1312 PAGE_CACHE_SIZE - 1);
1315 p->head_position = 0;
1316 rdev->raid_disk = mirror;
1318 /* As all devices are equivalent, we don't need a full recovery
1319 * if this was recently any drive of the array
1321 if (rdev->saved_raid_disk < 0)
1323 rcu_assign_pointer(p->rdev, rdev);
1326 md_integrity_add_rdev(rdev, mddev);
1331 static int raid1_remove_disk(struct mddev *mddev, int number)
1333 struct r1conf *conf = mddev->private;
1335 struct md_rdev *rdev;
1336 struct mirror_info *p = conf->mirrors+ number;
1341 if (test_bit(In_sync, &rdev->flags) ||
1342 atomic_read(&rdev->nr_pending)) {
1346 /* Only remove non-faulty devices if recovery
1349 if (!test_bit(Faulty, &rdev->flags) &&
1350 mddev->recovery_disabled != conf->recovery_disabled &&
1351 mddev->degraded < conf->raid_disks) {
1357 if (atomic_read(&rdev->nr_pending)) {
1358 /* lost the race, try later */
1363 err = md_integrity_register(mddev);
1372 static void end_sync_read(struct bio *bio, int error)
1374 struct r1bio *r1_bio = bio->bi_private;
1376 update_head_pos(r1_bio->read_disk, r1_bio);
1379 * we have read a block, now it needs to be re-written,
1380 * or re-read if the read failed.
1381 * We don't do much here, just schedule handling by raid1d
1383 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1384 set_bit(R1BIO_Uptodate, &r1_bio->state);
1386 if (atomic_dec_and_test(&r1_bio->remaining))
1387 reschedule_retry(r1_bio);
1390 static void end_sync_write(struct bio *bio, int error)
1392 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1393 struct r1bio *r1_bio = bio->bi_private;
1394 struct mddev *mddev = r1_bio->mddev;
1395 struct r1conf *conf = mddev->private;
1400 mirror = find_bio_disk(r1_bio, bio);
1403 sector_t sync_blocks = 0;
1404 sector_t s = r1_bio->sector;
1405 long sectors_to_go = r1_bio->sectors;
1406 /* make sure these bits doesn't get cleared. */
1408 bitmap_end_sync(mddev->bitmap, s,
1411 sectors_to_go -= sync_blocks;
1412 } while (sectors_to_go > 0);
1413 set_bit(WriteErrorSeen,
1414 &conf->mirrors[mirror].rdev->flags);
1415 set_bit(R1BIO_WriteError, &r1_bio->state);
1416 } else if (is_badblock(conf->mirrors[mirror].rdev,
1419 &first_bad, &bad_sectors) &&
1420 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1423 &first_bad, &bad_sectors)
1425 set_bit(R1BIO_MadeGood, &r1_bio->state);
1427 if (atomic_dec_and_test(&r1_bio->remaining)) {
1428 int s = r1_bio->sectors;
1429 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1430 test_bit(R1BIO_WriteError, &r1_bio->state))
1431 reschedule_retry(r1_bio);
1434 md_done_sync(mddev, s, uptodate);
1439 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1440 int sectors, struct page *page, int rw)
1442 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1446 set_bit(WriteErrorSeen, &rdev->flags);
1447 /* need to record an error - either for the block or the device */
1448 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1449 md_error(rdev->mddev, rdev);
1453 static int fix_sync_read_error(struct r1bio *r1_bio)
1455 /* Try some synchronous reads of other devices to get
1456 * good data, much like with normal read errors. Only
1457 * read into the pages we already have so we don't
1458 * need to re-issue the read request.
1459 * We don't need to freeze the array, because being in an
1460 * active sync request, there is no normal IO, and
1461 * no overlapping syncs.
1462 * We don't need to check is_badblock() again as we
1463 * made sure that anything with a bad block in range
1464 * will have bi_end_io clear.
1466 struct mddev *mddev = r1_bio->mddev;
1467 struct r1conf *conf = mddev->private;
1468 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1469 sector_t sect = r1_bio->sector;
1470 int sectors = r1_bio->sectors;
1475 int d = r1_bio->read_disk;
1477 struct md_rdev *rdev;
1480 if (s > (PAGE_SIZE>>9))
1483 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1484 /* No rcu protection needed here devices
1485 * can only be removed when no resync is
1486 * active, and resync is currently active
1488 rdev = conf->mirrors[d].rdev;
1489 if (sync_page_io(rdev, sect, s<<9,
1490 bio->bi_io_vec[idx].bv_page,
1497 if (d == conf->raid_disks)
1499 } while (!success && d != r1_bio->read_disk);
1502 char b[BDEVNAME_SIZE];
1504 /* Cannot read from anywhere, this block is lost.
1505 * Record a bad block on each device. If that doesn't
1506 * work just disable and interrupt the recovery.
1507 * Don't fail devices as that won't really help.
1509 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1510 " for block %llu\n",
1512 bdevname(bio->bi_bdev, b),
1513 (unsigned long long)r1_bio->sector);
1514 for (d = 0; d < conf->raid_disks; d++) {
1515 rdev = conf->mirrors[d].rdev;
1516 if (!rdev || test_bit(Faulty, &rdev->flags))
1518 if (!rdev_set_badblocks(rdev, sect, s, 0))
1522 mddev->recovery_disabled = 1;
1523 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1524 md_done_sync(mddev, r1_bio->sectors, 0);
1536 /* write it back and re-read */
1537 while (d != r1_bio->read_disk) {
1539 d = conf->raid_disks;
1541 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1543 rdev = conf->mirrors[d].rdev;
1544 if (r1_sync_page_io(rdev, sect, s,
1545 bio->bi_io_vec[idx].bv_page,
1547 r1_bio->bios[d]->bi_end_io = NULL;
1548 rdev_dec_pending(rdev, mddev);
1552 while (d != r1_bio->read_disk) {
1554 d = conf->raid_disks;
1556 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1558 rdev = conf->mirrors[d].rdev;
1559 if (r1_sync_page_io(rdev, sect, s,
1560 bio->bi_io_vec[idx].bv_page,
1562 atomic_add(s, &rdev->corrected_errors);
1568 set_bit(R1BIO_Uptodate, &r1_bio->state);
1569 set_bit(BIO_UPTODATE, &bio->bi_flags);
1573 static int process_checks(struct r1bio *r1_bio)
1575 /* We have read all readable devices. If we haven't
1576 * got the block, then there is no hope left.
1577 * If we have, then we want to do a comparison
1578 * and skip the write if everything is the same.
1579 * If any blocks failed to read, then we need to
1580 * attempt an over-write
1582 struct mddev *mddev = r1_bio->mddev;
1583 struct r1conf *conf = mddev->private;
1587 for (primary = 0; primary < conf->raid_disks; primary++)
1588 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1589 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1590 r1_bio->bios[primary]->bi_end_io = NULL;
1591 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1594 r1_bio->read_disk = primary;
1595 for (i = 0; i < conf->raid_disks; i++) {
1597 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1598 struct bio *pbio = r1_bio->bios[primary];
1599 struct bio *sbio = r1_bio->bios[i];
1602 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1605 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1606 for (j = vcnt; j-- ; ) {
1608 p = pbio->bi_io_vec[j].bv_page;
1609 s = sbio->bi_io_vec[j].bv_page;
1610 if (memcmp(page_address(p),
1618 mddev->resync_mismatches += r1_bio->sectors;
1619 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1620 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1621 /* No need to write to this device. */
1622 sbio->bi_end_io = NULL;
1623 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1626 /* fixup the bio for reuse */
1627 sbio->bi_vcnt = vcnt;
1628 sbio->bi_size = r1_bio->sectors << 9;
1630 sbio->bi_phys_segments = 0;
1631 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1632 sbio->bi_flags |= 1 << BIO_UPTODATE;
1633 sbio->bi_next = NULL;
1634 sbio->bi_sector = r1_bio->sector +
1635 conf->mirrors[i].rdev->data_offset;
1636 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1637 size = sbio->bi_size;
1638 for (j = 0; j < vcnt ; j++) {
1640 bi = &sbio->bi_io_vec[j];
1642 if (size > PAGE_SIZE)
1643 bi->bv_len = PAGE_SIZE;
1647 memcpy(page_address(bi->bv_page),
1648 page_address(pbio->bi_io_vec[j].bv_page),
1655 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1657 struct r1conf *conf = mddev->private;
1659 int disks = conf->raid_disks;
1660 struct bio *bio, *wbio;
1662 bio = r1_bio->bios[r1_bio->read_disk];
1664 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1665 /* ouch - failed to read all of that. */
1666 if (!fix_sync_read_error(r1_bio))
1669 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1670 if (process_checks(r1_bio) < 0)
1675 atomic_set(&r1_bio->remaining, 1);
1676 for (i = 0; i < disks ; i++) {
1677 wbio = r1_bio->bios[i];
1678 if (wbio->bi_end_io == NULL ||
1679 (wbio->bi_end_io == end_sync_read &&
1680 (i == r1_bio->read_disk ||
1681 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1684 wbio->bi_rw = WRITE;
1685 wbio->bi_end_io = end_sync_write;
1686 atomic_inc(&r1_bio->remaining);
1687 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1689 generic_make_request(wbio);
1692 if (atomic_dec_and_test(&r1_bio->remaining)) {
1693 /* if we're here, all write(s) have completed, so clean up */
1694 md_done_sync(mddev, r1_bio->sectors, 1);
1700 * This is a kernel thread which:
1702 * 1. Retries failed read operations on working mirrors.
1703 * 2. Updates the raid superblock when problems encounter.
1704 * 3. Performs writes following reads for array synchronising.
1707 static void fix_read_error(struct r1conf *conf, int read_disk,
1708 sector_t sect, int sectors)
1710 struct mddev *mddev = conf->mddev;
1716 struct md_rdev *rdev;
1718 if (s > (PAGE_SIZE>>9))
1722 /* Note: no rcu protection needed here
1723 * as this is synchronous in the raid1d thread
1724 * which is the thread that might remove
1725 * a device. If raid1d ever becomes multi-threaded....
1730 rdev = conf->mirrors[d].rdev;
1732 test_bit(In_sync, &rdev->flags) &&
1733 is_badblock(rdev, sect, s,
1734 &first_bad, &bad_sectors) == 0 &&
1735 sync_page_io(rdev, sect, s<<9,
1736 conf->tmppage, READ, false))
1740 if (d == conf->raid_disks)
1743 } while (!success && d != read_disk);
1746 /* Cannot read from anywhere - mark it bad */
1747 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1748 if (!rdev_set_badblocks(rdev, sect, s, 0))
1749 md_error(mddev, rdev);
1752 /* write it back and re-read */
1754 while (d != read_disk) {
1756 d = conf->raid_disks;
1758 rdev = conf->mirrors[d].rdev;
1760 test_bit(In_sync, &rdev->flags))
1761 r1_sync_page_io(rdev, sect, s,
1762 conf->tmppage, WRITE);
1765 while (d != read_disk) {
1766 char b[BDEVNAME_SIZE];
1768 d = conf->raid_disks;
1770 rdev = conf->mirrors[d].rdev;
1772 test_bit(In_sync, &rdev->flags)) {
1773 if (r1_sync_page_io(rdev, sect, s,
1774 conf->tmppage, READ)) {
1775 atomic_add(s, &rdev->corrected_errors);
1777 "md/raid1:%s: read error corrected "
1778 "(%d sectors at %llu on %s)\n",
1780 (unsigned long long)(sect +
1782 bdevname(rdev->bdev, b));
1791 static void bi_complete(struct bio *bio, int error)
1793 complete((struct completion *)bio->bi_private);
1796 static int submit_bio_wait(int rw, struct bio *bio)
1798 struct completion event;
1801 init_completion(&event);
1802 bio->bi_private = &event;
1803 bio->bi_end_io = bi_complete;
1804 submit_bio(rw, bio);
1805 wait_for_completion(&event);
1807 return test_bit(BIO_UPTODATE, &bio->bi_flags);
1810 static int narrow_write_error(struct r1bio *r1_bio, int i)
1812 struct mddev *mddev = r1_bio->mddev;
1813 struct r1conf *conf = mddev->private;
1814 struct md_rdev *rdev = conf->mirrors[i].rdev;
1816 struct bio_vec *vec;
1818 /* bio has the data to be written to device 'i' where
1819 * we just recently had a write error.
1820 * We repeatedly clone the bio and trim down to one block,
1821 * then try the write. Where the write fails we record
1823 * It is conceivable that the bio doesn't exactly align with
1824 * blocks. We must handle this somehow.
1826 * We currently own a reference on the rdev.
1832 int sect_to_write = r1_bio->sectors;
1835 if (rdev->badblocks.shift < 0)
1838 block_sectors = 1 << rdev->badblocks.shift;
1839 sector = r1_bio->sector;
1840 sectors = ((sector + block_sectors)
1841 & ~(sector_t)(block_sectors - 1))
1844 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1845 vcnt = r1_bio->behind_page_count;
1846 vec = r1_bio->behind_bvecs;
1848 while (vec[idx].bv_page == NULL)
1851 vcnt = r1_bio->master_bio->bi_vcnt;
1852 vec = r1_bio->master_bio->bi_io_vec;
1853 idx = r1_bio->master_bio->bi_idx;
1855 while (sect_to_write) {
1857 if (sectors > sect_to_write)
1858 sectors = sect_to_write;
1859 /* Write at 'sector' for 'sectors'*/
1861 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1862 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1863 wbio->bi_sector = r1_bio->sector;
1864 wbio->bi_rw = WRITE;
1865 wbio->bi_vcnt = vcnt;
1866 wbio->bi_size = r1_bio->sectors << 9;
1869 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
1870 wbio->bi_sector += rdev->data_offset;
1871 wbio->bi_bdev = rdev->bdev;
1872 if (submit_bio_wait(WRITE, wbio) == 0)
1874 ok = rdev_set_badblocks(rdev, sector,
1879 sect_to_write -= sectors;
1881 sectors = block_sectors;
1886 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1889 int s = r1_bio->sectors;
1890 for (m = 0; m < conf->raid_disks ; m++) {
1891 struct md_rdev *rdev = conf->mirrors[m].rdev;
1892 struct bio *bio = r1_bio->bios[m];
1893 if (bio->bi_end_io == NULL)
1895 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1896 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1897 rdev_clear_badblocks(rdev, r1_bio->sector, s);
1899 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1900 test_bit(R1BIO_WriteError, &r1_bio->state)) {
1901 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
1902 md_error(conf->mddev, rdev);
1906 md_done_sync(conf->mddev, s, 1);
1909 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1912 for (m = 0; m < conf->raid_disks ; m++)
1913 if (r1_bio->bios[m] == IO_MADE_GOOD) {
1914 struct md_rdev *rdev = conf->mirrors[m].rdev;
1915 rdev_clear_badblocks(rdev,
1918 rdev_dec_pending(rdev, conf->mddev);
1919 } else if (r1_bio->bios[m] != NULL) {
1920 /* This drive got a write error. We need to
1921 * narrow down and record precise write
1924 if (!narrow_write_error(r1_bio, m)) {
1925 md_error(conf->mddev,
1926 conf->mirrors[m].rdev);
1927 /* an I/O failed, we can't clear the bitmap */
1928 set_bit(R1BIO_Degraded, &r1_bio->state);
1930 rdev_dec_pending(conf->mirrors[m].rdev,
1933 if (test_bit(R1BIO_WriteError, &r1_bio->state))
1934 close_write(r1_bio);
1935 raid_end_bio_io(r1_bio);
1938 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
1942 struct mddev *mddev = conf->mddev;
1944 char b[BDEVNAME_SIZE];
1945 struct md_rdev *rdev;
1947 clear_bit(R1BIO_ReadError, &r1_bio->state);
1948 /* we got a read error. Maybe the drive is bad. Maybe just
1949 * the block and we can fix it.
1950 * We freeze all other IO, and try reading the block from
1951 * other devices. When we find one, we re-write
1952 * and check it that fixes the read error.
1953 * This is all done synchronously while the array is
1956 if (mddev->ro == 0) {
1958 fix_read_error(conf, r1_bio->read_disk,
1959 r1_bio->sector, r1_bio->sectors);
1960 unfreeze_array(conf);
1962 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1964 bio = r1_bio->bios[r1_bio->read_disk];
1965 bdevname(bio->bi_bdev, b);
1967 disk = read_balance(conf, r1_bio, &max_sectors);
1969 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1970 " read error for block %llu\n",
1971 mdname(mddev), b, (unsigned long long)r1_bio->sector);
1972 raid_end_bio_io(r1_bio);
1974 const unsigned long do_sync
1975 = r1_bio->master_bio->bi_rw & REQ_SYNC;
1977 r1_bio->bios[r1_bio->read_disk] =
1978 mddev->ro ? IO_BLOCKED : NULL;
1981 r1_bio->read_disk = disk;
1982 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
1983 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
1984 r1_bio->bios[r1_bio->read_disk] = bio;
1985 rdev = conf->mirrors[disk].rdev;
1986 printk_ratelimited(KERN_ERR
1987 "md/raid1:%s: redirecting sector %llu"
1988 " to other mirror: %s\n",
1990 (unsigned long long)r1_bio->sector,
1991 bdevname(rdev->bdev, b));
1992 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1993 bio->bi_bdev = rdev->bdev;
1994 bio->bi_end_io = raid1_end_read_request;
1995 bio->bi_rw = READ | do_sync;
1996 bio->bi_private = r1_bio;
1997 if (max_sectors < r1_bio->sectors) {
1998 /* Drat - have to split this up more */
1999 struct bio *mbio = r1_bio->master_bio;
2000 int sectors_handled = (r1_bio->sector + max_sectors
2002 r1_bio->sectors = max_sectors;
2003 spin_lock_irq(&conf->device_lock);
2004 if (mbio->bi_phys_segments == 0)
2005 mbio->bi_phys_segments = 2;
2007 mbio->bi_phys_segments++;
2008 spin_unlock_irq(&conf->device_lock);
2009 generic_make_request(bio);
2012 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2014 r1_bio->master_bio = mbio;
2015 r1_bio->sectors = (mbio->bi_size >> 9)
2018 set_bit(R1BIO_ReadError, &r1_bio->state);
2019 r1_bio->mddev = mddev;
2020 r1_bio->sector = mbio->bi_sector + sectors_handled;
2024 generic_make_request(bio);
2028 static void raid1d(struct mddev *mddev)
2030 struct r1bio *r1_bio;
2031 unsigned long flags;
2032 struct r1conf *conf = mddev->private;
2033 struct list_head *head = &conf->retry_list;
2034 struct blk_plug plug;
2036 md_check_recovery(mddev);
2038 blk_start_plug(&plug);
2041 if (atomic_read(&mddev->plug_cnt) == 0)
2042 flush_pending_writes(conf);
2044 spin_lock_irqsave(&conf->device_lock, flags);
2045 if (list_empty(head)) {
2046 spin_unlock_irqrestore(&conf->device_lock, flags);
2049 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2050 list_del(head->prev);
2052 spin_unlock_irqrestore(&conf->device_lock, flags);
2054 mddev = r1_bio->mddev;
2055 conf = mddev->private;
2056 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2057 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2058 test_bit(R1BIO_WriteError, &r1_bio->state))
2059 handle_sync_write_finished(conf, r1_bio);
2061 sync_request_write(mddev, r1_bio);
2062 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2063 test_bit(R1BIO_WriteError, &r1_bio->state))
2064 handle_write_finished(conf, r1_bio);
2065 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2066 handle_read_error(conf, r1_bio);
2068 /* just a partial read to be scheduled from separate
2071 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2074 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2075 md_check_recovery(mddev);
2077 blk_finish_plug(&plug);
2081 static int init_resync(struct r1conf *conf)
2085 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2086 BUG_ON(conf->r1buf_pool);
2087 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2089 if (!conf->r1buf_pool)
2091 conf->next_resync = 0;
2096 * perform a "sync" on one "block"
2098 * We need to make sure that no normal I/O request - particularly write
2099 * requests - conflict with active sync requests.
2101 * This is achieved by tracking pending requests and a 'barrier' concept
2102 * that can be installed to exclude normal IO requests.
2105 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2107 struct r1conf *conf = mddev->private;
2108 struct r1bio *r1_bio;
2110 sector_t max_sector, nr_sectors;
2114 int write_targets = 0, read_targets = 0;
2115 sector_t sync_blocks;
2116 int still_degraded = 0;
2117 int good_sectors = RESYNC_SECTORS;
2118 int min_bad = 0; /* number of sectors that are bad in all devices */
2120 if (!conf->r1buf_pool)
2121 if (init_resync(conf))
2124 max_sector = mddev->dev_sectors;
2125 if (sector_nr >= max_sector) {
2126 /* If we aborted, we need to abort the
2127 * sync on the 'current' bitmap chunk (there will
2128 * only be one in raid1 resync.
2129 * We can find the current addess in mddev->curr_resync
2131 if (mddev->curr_resync < max_sector) /* aborted */
2132 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2134 else /* completed sync */
2137 bitmap_close_sync(mddev->bitmap);
2142 if (mddev->bitmap == NULL &&
2143 mddev->recovery_cp == MaxSector &&
2144 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2145 conf->fullsync == 0) {
2147 return max_sector - sector_nr;
2149 /* before building a request, check if we can skip these blocks..
2150 * This call the bitmap_start_sync doesn't actually record anything
2152 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2153 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2154 /* We can skip this block, and probably several more */
2159 * If there is non-resync activity waiting for a turn,
2160 * and resync is going fast enough,
2161 * then let it though before starting on this new sync request.
2163 if (!go_faster && conf->nr_waiting)
2164 msleep_interruptible(1000);
2166 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2167 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2168 raise_barrier(conf);
2170 conf->next_resync = sector_nr;
2174 * If we get a correctably read error during resync or recovery,
2175 * we might want to read from a different device. So we
2176 * flag all drives that could conceivably be read from for READ,
2177 * and any others (which will be non-In_sync devices) for WRITE.
2178 * If a read fails, we try reading from something else for which READ
2182 r1_bio->mddev = mddev;
2183 r1_bio->sector = sector_nr;
2185 set_bit(R1BIO_IsSync, &r1_bio->state);
2187 for (i=0; i < conf->raid_disks; i++) {
2188 struct md_rdev *rdev;
2189 bio = r1_bio->bios[i];
2191 /* take from bio_init */
2192 bio->bi_next = NULL;
2193 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2194 bio->bi_flags |= 1 << BIO_UPTODATE;
2195 bio->bi_comp_cpu = -1;
2199 bio->bi_phys_segments = 0;
2201 bio->bi_end_io = NULL;
2202 bio->bi_private = NULL;
2204 rdev = rcu_dereference(conf->mirrors[i].rdev);
2206 test_bit(Faulty, &rdev->flags)) {
2208 } else if (!test_bit(In_sync, &rdev->flags)) {
2210 bio->bi_end_io = end_sync_write;
2213 /* may need to read from here */
2214 sector_t first_bad = MaxSector;
2217 if (is_badblock(rdev, sector_nr, good_sectors,
2218 &first_bad, &bad_sectors)) {
2219 if (first_bad > sector_nr)
2220 good_sectors = first_bad - sector_nr;
2222 bad_sectors -= (sector_nr - first_bad);
2224 min_bad > bad_sectors)
2225 min_bad = bad_sectors;
2228 if (sector_nr < first_bad) {
2229 if (test_bit(WriteMostly, &rdev->flags)) {
2237 bio->bi_end_io = end_sync_read;
2241 if (bio->bi_end_io) {
2242 atomic_inc(&rdev->nr_pending);
2243 bio->bi_sector = sector_nr + rdev->data_offset;
2244 bio->bi_bdev = rdev->bdev;
2245 bio->bi_private = r1_bio;
2251 r1_bio->read_disk = disk;
2253 if (read_targets == 0 && min_bad > 0) {
2254 /* These sectors are bad on all InSync devices, so we
2255 * need to mark them bad on all write targets
2258 for (i = 0 ; i < conf->raid_disks ; i++)
2259 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2260 struct md_rdev *rdev =
2261 rcu_dereference(conf->mirrors[i].rdev);
2262 ok = rdev_set_badblocks(rdev, sector_nr,
2266 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2271 /* Cannot record the badblocks, so need to
2273 * If there are multiple read targets, could just
2274 * fail the really bad ones ???
2276 conf->recovery_disabled = mddev->recovery_disabled;
2277 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2283 if (min_bad > 0 && min_bad < good_sectors) {
2284 /* only resync enough to reach the next bad->good
2286 good_sectors = min_bad;
2289 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2290 /* extra read targets are also write targets */
2291 write_targets += read_targets-1;
2293 if (write_targets == 0 || read_targets == 0) {
2294 /* There is nowhere to write, so all non-sync
2295 * drives must be failed - so we are finished
2297 sector_t rv = max_sector - sector_nr;
2303 if (max_sector > mddev->resync_max)
2304 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2305 if (max_sector > sector_nr + good_sectors)
2306 max_sector = sector_nr + good_sectors;
2311 int len = PAGE_SIZE;
2312 if (sector_nr + (len>>9) > max_sector)
2313 len = (max_sector - sector_nr) << 9;
2316 if (sync_blocks == 0) {
2317 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2318 &sync_blocks, still_degraded) &&
2320 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2322 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2323 if ((len >> 9) > sync_blocks)
2324 len = sync_blocks<<9;
2327 for (i=0 ; i < conf->raid_disks; i++) {
2328 bio = r1_bio->bios[i];
2329 if (bio->bi_end_io) {
2330 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2331 if (bio_add_page(bio, page, len, 0) == 0) {
2333 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2336 bio = r1_bio->bios[i];
2337 if (bio->bi_end_io==NULL)
2339 /* remove last page from this bio */
2341 bio->bi_size -= len;
2342 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2348 nr_sectors += len>>9;
2349 sector_nr += len>>9;
2350 sync_blocks -= (len>>9);
2351 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2353 r1_bio->sectors = nr_sectors;
2355 /* For a user-requested sync, we read all readable devices and do a
2358 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2359 atomic_set(&r1_bio->remaining, read_targets);
2360 for (i=0; i<conf->raid_disks; i++) {
2361 bio = r1_bio->bios[i];
2362 if (bio->bi_end_io == end_sync_read) {
2363 md_sync_acct(bio->bi_bdev, nr_sectors);
2364 generic_make_request(bio);
2368 atomic_set(&r1_bio->remaining, 1);
2369 bio = r1_bio->bios[r1_bio->read_disk];
2370 md_sync_acct(bio->bi_bdev, nr_sectors);
2371 generic_make_request(bio);
2377 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2382 return mddev->dev_sectors;
2385 static struct r1conf *setup_conf(struct mddev *mddev)
2387 struct r1conf *conf;
2389 struct mirror_info *disk;
2390 struct md_rdev *rdev;
2393 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2397 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2402 conf->tmppage = alloc_page(GFP_KERNEL);
2406 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2407 if (!conf->poolinfo)
2409 conf->poolinfo->raid_disks = mddev->raid_disks;
2410 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2413 if (!conf->r1bio_pool)
2416 conf->poolinfo->mddev = mddev;
2418 spin_lock_init(&conf->device_lock);
2419 list_for_each_entry(rdev, &mddev->disks, same_set) {
2420 int disk_idx = rdev->raid_disk;
2421 if (disk_idx >= mddev->raid_disks
2424 disk = conf->mirrors + disk_idx;
2428 disk->head_position = 0;
2430 conf->raid_disks = mddev->raid_disks;
2431 conf->mddev = mddev;
2432 INIT_LIST_HEAD(&conf->retry_list);
2434 spin_lock_init(&conf->resync_lock);
2435 init_waitqueue_head(&conf->wait_barrier);
2437 bio_list_init(&conf->pending_bio_list);
2438 conf->pending_count = 0;
2440 conf->last_used = -1;
2441 for (i = 0; i < conf->raid_disks; i++) {
2443 disk = conf->mirrors + i;
2446 !test_bit(In_sync, &disk->rdev->flags)) {
2447 disk->head_position = 0;
2450 } else if (conf->last_used < 0)
2452 * The first working device is used as a
2453 * starting point to read balancing.
2455 conf->last_used = i;
2459 if (conf->last_used < 0) {
2460 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2465 conf->thread = md_register_thread(raid1d, mddev, NULL);
2466 if (!conf->thread) {
2468 "md/raid1:%s: couldn't allocate thread\n",
2477 if (conf->r1bio_pool)
2478 mempool_destroy(conf->r1bio_pool);
2479 kfree(conf->mirrors);
2480 safe_put_page(conf->tmppage);
2481 kfree(conf->poolinfo);
2484 return ERR_PTR(err);
2487 static int run(struct mddev *mddev)
2489 struct r1conf *conf;
2491 struct md_rdev *rdev;
2493 if (mddev->level != 1) {
2494 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2495 mdname(mddev), mddev->level);
2498 if (mddev->reshape_position != MaxSector) {
2499 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2504 * copy the already verified devices into our private RAID1
2505 * bookkeeping area. [whatever we allocate in run(),
2506 * should be freed in stop()]
2508 if (mddev->private == NULL)
2509 conf = setup_conf(mddev);
2511 conf = mddev->private;
2514 return PTR_ERR(conf);
2516 list_for_each_entry(rdev, &mddev->disks, same_set) {
2517 if (!mddev->gendisk)
2519 disk_stack_limits(mddev->gendisk, rdev->bdev,
2520 rdev->data_offset << 9);
2521 /* as we don't honour merge_bvec_fn, we must never risk
2522 * violating it, so limit ->max_segments to 1 lying within
2523 * a single page, as a one page request is never in violation.
2525 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2526 blk_queue_max_segments(mddev->queue, 1);
2527 blk_queue_segment_boundary(mddev->queue,
2528 PAGE_CACHE_SIZE - 1);
2532 mddev->degraded = 0;
2533 for (i=0; i < conf->raid_disks; i++)
2534 if (conf->mirrors[i].rdev == NULL ||
2535 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2536 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2539 if (conf->raid_disks - mddev->degraded == 1)
2540 mddev->recovery_cp = MaxSector;
2542 if (mddev->recovery_cp != MaxSector)
2543 printk(KERN_NOTICE "md/raid1:%s: not clean"
2544 " -- starting background reconstruction\n",
2547 "md/raid1:%s: active with %d out of %d mirrors\n",
2548 mdname(mddev), mddev->raid_disks - mddev->degraded,
2552 * Ok, everything is just fine now
2554 mddev->thread = conf->thread;
2555 conf->thread = NULL;
2556 mddev->private = conf;
2558 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2561 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2562 mddev->queue->backing_dev_info.congested_data = mddev;
2564 return md_integrity_register(mddev);
2567 static int stop(struct mddev *mddev)
2569 struct r1conf *conf = mddev->private;
2570 struct bitmap *bitmap = mddev->bitmap;
2572 /* wait for behind writes to complete */
2573 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2574 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2576 /* need to kick something here to make sure I/O goes? */
2577 wait_event(bitmap->behind_wait,
2578 atomic_read(&bitmap->behind_writes) == 0);
2581 raise_barrier(conf);
2582 lower_barrier(conf);
2584 md_unregister_thread(&mddev->thread);
2585 if (conf->r1bio_pool)
2586 mempool_destroy(conf->r1bio_pool);
2587 kfree(conf->mirrors);
2588 kfree(conf->poolinfo);
2590 mddev->private = NULL;
2594 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2596 /* no resync is happening, and there is enough space
2597 * on all devices, so we can resize.
2598 * We need to make sure resync covers any new space.
2599 * If the array is shrinking we should possibly wait until
2600 * any io in the removed space completes, but it hardly seems
2603 md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2604 if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2606 set_capacity(mddev->gendisk, mddev->array_sectors);
2607 revalidate_disk(mddev->gendisk);
2608 if (sectors > mddev->dev_sectors &&
2609 mddev->recovery_cp > mddev->dev_sectors) {
2610 mddev->recovery_cp = mddev->dev_sectors;
2611 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2613 mddev->dev_sectors = sectors;
2614 mddev->resync_max_sectors = sectors;
2618 static int raid1_reshape(struct mddev *mddev)
2621 * 1/ resize the r1bio_pool
2622 * 2/ resize conf->mirrors
2624 * We allocate a new r1bio_pool if we can.
2625 * Then raise a device barrier and wait until all IO stops.
2626 * Then resize conf->mirrors and swap in the new r1bio pool.
2628 * At the same time, we "pack" the devices so that all the missing
2629 * devices have the higher raid_disk numbers.
2631 mempool_t *newpool, *oldpool;
2632 struct pool_info *newpoolinfo;
2633 struct mirror_info *newmirrors;
2634 struct r1conf *conf = mddev->private;
2635 int cnt, raid_disks;
2636 unsigned long flags;
2639 /* Cannot change chunk_size, layout, or level */
2640 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2641 mddev->layout != mddev->new_layout ||
2642 mddev->level != mddev->new_level) {
2643 mddev->new_chunk_sectors = mddev->chunk_sectors;
2644 mddev->new_layout = mddev->layout;
2645 mddev->new_level = mddev->level;
2649 err = md_allow_write(mddev);
2653 raid_disks = mddev->raid_disks + mddev->delta_disks;
2655 if (raid_disks < conf->raid_disks) {
2657 for (d= 0; d < conf->raid_disks; d++)
2658 if (conf->mirrors[d].rdev)
2660 if (cnt > raid_disks)
2664 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2667 newpoolinfo->mddev = mddev;
2668 newpoolinfo->raid_disks = raid_disks;
2670 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2671 r1bio_pool_free, newpoolinfo);
2676 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2679 mempool_destroy(newpool);
2683 raise_barrier(conf);
2685 /* ok, everything is stopped */
2686 oldpool = conf->r1bio_pool;
2687 conf->r1bio_pool = newpool;
2689 for (d = d2 = 0; d < conf->raid_disks; d++) {
2690 struct md_rdev *rdev = conf->mirrors[d].rdev;
2691 if (rdev && rdev->raid_disk != d2) {
2692 sysfs_unlink_rdev(mddev, rdev);
2693 rdev->raid_disk = d2;
2694 sysfs_unlink_rdev(mddev, rdev);
2695 if (sysfs_link_rdev(mddev, rdev))
2697 "md/raid1:%s: cannot register rd%d\n",
2698 mdname(mddev), rdev->raid_disk);
2701 newmirrors[d2++].rdev = rdev;
2703 kfree(conf->mirrors);
2704 conf->mirrors = newmirrors;
2705 kfree(conf->poolinfo);
2706 conf->poolinfo = newpoolinfo;
2708 spin_lock_irqsave(&conf->device_lock, flags);
2709 mddev->degraded += (raid_disks - conf->raid_disks);
2710 spin_unlock_irqrestore(&conf->device_lock, flags);
2711 conf->raid_disks = mddev->raid_disks = raid_disks;
2712 mddev->delta_disks = 0;
2714 conf->last_used = 0; /* just make sure it is in-range */
2715 lower_barrier(conf);
2717 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2718 md_wakeup_thread(mddev->thread);
2720 mempool_destroy(oldpool);
2724 static void raid1_quiesce(struct mddev *mddev, int state)
2726 struct r1conf *conf = mddev->private;
2729 case 2: /* wake for suspend */
2730 wake_up(&conf->wait_barrier);
2733 raise_barrier(conf);
2736 lower_barrier(conf);
2741 static void *raid1_takeover(struct mddev *mddev)
2743 /* raid1 can take over:
2744 * raid5 with 2 devices, any layout or chunk size
2746 if (mddev->level == 5 && mddev->raid_disks == 2) {
2747 struct r1conf *conf;
2748 mddev->new_level = 1;
2749 mddev->new_layout = 0;
2750 mddev->new_chunk_sectors = 0;
2751 conf = setup_conf(mddev);
2756 return ERR_PTR(-EINVAL);
2759 static struct md_personality raid1_personality =
2763 .owner = THIS_MODULE,
2764 .make_request = make_request,
2768 .error_handler = error,
2769 .hot_add_disk = raid1_add_disk,
2770 .hot_remove_disk= raid1_remove_disk,
2771 .spare_active = raid1_spare_active,
2772 .sync_request = sync_request,
2773 .resize = raid1_resize,
2775 .check_reshape = raid1_reshape,
2776 .quiesce = raid1_quiesce,
2777 .takeover = raid1_takeover,
2780 static int __init raid_init(void)
2782 return register_md_personality(&raid1_personality);
2785 static void raid_exit(void)
2787 unregister_md_personality(&raid1_personality);
2790 module_init(raid_init);
2791 module_exit(raid_exit);
2792 MODULE_LICENSE("GPL");
2793 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2794 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2795 MODULE_ALIAS("md-raid1");
2796 MODULE_ALIAS("md-level-1");
2798 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);