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 "dm-bio-list.h"
35 #include <linux/raid/raid1.h>
36 #include <linux/raid/bitmap.h>
40 #define PRINTK(x...) printk(x)
46 * Number of guaranteed r1bios in case of extreme VM load:
48 #define NR_RAID1_BIOS 256
50 static mdk_personality_t raid1_personality;
52 static void unplug_slaves(mddev_t *mddev);
55 static void * r1bio_pool_alloc(unsigned int __nocast gfp_flags, void *data)
57 struct pool_info *pi = data;
59 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
61 /* allocate a r1bio with room for raid_disks entries in the bios array */
62 r1_bio = kmalloc(size, gfp_flags);
64 memset(r1_bio, 0, size);
66 unplug_slaves(pi->mddev);
71 static void r1bio_pool_free(void *r1_bio, void *data)
76 #define RESYNC_BLOCK_SIZE (64*1024)
77 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
78 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
79 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
80 #define RESYNC_WINDOW (2048*1024)
82 static void * r1buf_pool_alloc(unsigned int __nocast gfp_flags, void *data)
84 struct pool_info *pi = data;
90 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
92 unplug_slaves(pi->mddev);
97 * Allocate bios : 1 for reading, n-1 for writing
99 for (j = pi->raid_disks ; j-- ; ) {
100 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
103 r1_bio->bios[j] = bio;
106 * Allocate RESYNC_PAGES data pages and attach them to
109 bio = r1_bio->bios[0];
110 for (i = 0; i < RESYNC_PAGES; i++) {
111 page = alloc_page(gfp_flags);
115 bio->bi_io_vec[i].bv_page = page;
118 r1_bio->master_bio = NULL;
124 __free_page(bio->bi_io_vec[i-1].bv_page);
126 while ( ++j < pi->raid_disks )
127 bio_put(r1_bio->bios[j]);
128 r1bio_pool_free(r1_bio, data);
132 static void r1buf_pool_free(void *__r1_bio, void *data)
134 struct pool_info *pi = data;
136 r1bio_t *r1bio = __r1_bio;
137 struct bio *bio = r1bio->bios[0];
139 for (i = 0; i < RESYNC_PAGES; i++) {
140 __free_page(bio->bi_io_vec[i].bv_page);
141 bio->bi_io_vec[i].bv_page = NULL;
143 for (i=0 ; i < pi->raid_disks; i++)
144 bio_put(r1bio->bios[i]);
146 r1bio_pool_free(r1bio, data);
149 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
153 for (i = 0; i < conf->raid_disks; i++) {
154 struct bio **bio = r1_bio->bios + i;
161 static inline void free_r1bio(r1bio_t *r1_bio)
165 conf_t *conf = mddev_to_conf(r1_bio->mddev);
168 * Wake up any possible resync thread that waits for the device
171 spin_lock_irqsave(&conf->resync_lock, flags);
172 if (!--conf->nr_pending) {
173 wake_up(&conf->wait_idle);
174 wake_up(&conf->wait_resume);
176 spin_unlock_irqrestore(&conf->resync_lock, flags);
178 put_all_bios(conf, r1_bio);
179 mempool_free(r1_bio, conf->r1bio_pool);
182 static inline void put_buf(r1bio_t *r1_bio)
184 conf_t *conf = mddev_to_conf(r1_bio->mddev);
187 mempool_free(r1_bio, conf->r1buf_pool);
189 spin_lock_irqsave(&conf->resync_lock, flags);
193 wake_up(&conf->wait_resume);
194 wake_up(&conf->wait_idle);
196 if (!--conf->nr_pending) {
197 wake_up(&conf->wait_idle);
198 wake_up(&conf->wait_resume);
200 spin_unlock_irqrestore(&conf->resync_lock, flags);
203 static void reschedule_retry(r1bio_t *r1_bio)
206 mddev_t *mddev = r1_bio->mddev;
207 conf_t *conf = mddev_to_conf(mddev);
209 spin_lock_irqsave(&conf->device_lock, flags);
210 list_add(&r1_bio->retry_list, &conf->retry_list);
211 spin_unlock_irqrestore(&conf->device_lock, flags);
213 md_wakeup_thread(mddev->thread);
217 * raid_end_bio_io() is called when we have finished servicing a mirrored
218 * operation and are ready to return a success/failure code to the buffer
221 static void raid_end_bio_io(r1bio_t *r1_bio)
223 struct bio *bio = r1_bio->master_bio;
225 bio_endio(bio, bio->bi_size,
226 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
231 * Update disk head position estimator based on IRQ completion info.
233 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
235 conf_t *conf = mddev_to_conf(r1_bio->mddev);
237 conf->mirrors[disk].head_position =
238 r1_bio->sector + (r1_bio->sectors);
241 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
243 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
244 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
246 conf_t *conf = mddev_to_conf(r1_bio->mddev);
251 mirror = r1_bio->read_disk;
253 * this branch is our 'one mirror IO has finished' event handler:
256 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
259 * Set R1BIO_Uptodate in our master bio, so that
260 * we will return a good error code for to the higher
261 * levels even if IO on some other mirrored buffer fails.
263 * The 'master' represents the composite IO operation to
264 * user-side. So if something waits for IO, then it will
265 * wait for the 'master' bio.
267 set_bit(R1BIO_Uptodate, &r1_bio->state);
269 update_head_pos(mirror, r1_bio);
272 * we have only one bio on the read side
275 raid_end_bio_io(r1_bio);
280 char b[BDEVNAME_SIZE];
281 if (printk_ratelimit())
282 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
283 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
284 reschedule_retry(r1_bio);
287 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
291 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
293 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
294 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
296 conf_t *conf = mddev_to_conf(r1_bio->mddev);
301 for (mirror = 0; mirror < conf->raid_disks; mirror++)
302 if (r1_bio->bios[mirror] == bio)
306 * this branch is our 'one mirror IO has finished' event handler:
309 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
310 /* an I/O failed, we can't clear the bitmap */
311 set_bit(R1BIO_Degraded, &r1_bio->state);
314 * Set R1BIO_Uptodate in our master bio, so that
315 * we will return a good error code for to the higher
316 * levels even if IO on some other mirrored buffer fails.
318 * The 'master' represents the composite IO operation to
319 * user-side. So if something waits for IO, then it will
320 * wait for the 'master' bio.
322 set_bit(R1BIO_Uptodate, &r1_bio->state);
324 update_head_pos(mirror, r1_bio);
328 * Let's see if all mirrored write operations have finished
331 if (atomic_dec_and_test(&r1_bio->remaining)) {
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
335 !test_bit(R1BIO_Degraded, &r1_bio->state));
336 md_write_end(r1_bio->mddev);
337 raid_end_bio_io(r1_bio);
340 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
346 * This routine returns the disk from which the requested read should
347 * be done. There is a per-array 'next expected sequential IO' sector
348 * number - if this matches on the next IO then we use the last disk.
349 * There is also a per-disk 'last know head position' sector that is
350 * maintained from IRQ contexts, both the normal and the resync IO
351 * completion handlers update this position correctly. If there is no
352 * perfect sequential match then we pick the disk whose head is closest.
354 * If there are 2 mirrors in the same 2 devices, performance degrades
355 * because position is mirror, not device based.
357 * The rdev for the device selected will have nr_pending incremented.
359 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
361 const unsigned long this_sector = r1_bio->sector;
362 int new_disk = conf->last_used, disk = new_disk;
363 const int sectors = r1_bio->sectors;
364 sector_t new_distance, current_distance;
365 mdk_rdev_t *new_rdev, *rdev;
369 * Check if it if we can balance. We can balance on the whole
370 * device if no resync is going on, or below the resync window.
371 * We take the first readable disk when above the resync window.
374 if (conf->mddev->recovery_cp < MaxSector &&
375 (this_sector + sectors >= conf->next_resync)) {
376 /* Choose the first operation device, for consistancy */
379 while ((new_rdev=conf->mirrors[new_disk].rdev) == NULL ||
380 !new_rdev->in_sync) {
382 if (new_disk == conf->raid_disks) {
391 /* make sure the disk is operational */
392 while ((new_rdev=conf->mirrors[new_disk].rdev) == NULL ||
393 !new_rdev->in_sync) {
395 new_disk = conf->raid_disks;
397 if (new_disk == disk) {
403 /* now disk == new_disk == starting point for search */
406 * Don't change to another disk for sequential reads:
408 if (conf->next_seq_sect == this_sector)
410 if (this_sector == conf->mirrors[new_disk].head_position)
413 current_distance = abs(this_sector - conf->mirrors[disk].head_position);
415 /* Find the disk whose head is closest */
419 disk = conf->raid_disks;
422 if ((rdev=conf->mirrors[disk].rdev) == NULL ||
426 if (!atomic_read(&rdev->nr_pending)) {
431 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
432 if (new_distance < current_distance) {
433 current_distance = new_distance;
437 } while (disk != conf->last_used);
443 conf->next_seq_sect = this_sector + sectors;
444 conf->last_used = new_disk;
445 atomic_inc(&new_rdev->nr_pending);
446 if (!new_rdev->in_sync) {
447 /* cannot risk returning a device that failed
448 * before we inc'ed nr_pending
450 atomic_dec(&new_rdev->nr_pending);
459 static void unplug_slaves(mddev_t *mddev)
461 conf_t *conf = mddev_to_conf(mddev);
465 for (i=0; i<mddev->raid_disks; i++) {
466 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
467 if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
468 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
470 atomic_inc(&rdev->nr_pending);
473 if (r_queue->unplug_fn)
474 r_queue->unplug_fn(r_queue);
476 rdev_dec_pending(rdev, mddev);
483 static void raid1_unplug(request_queue_t *q)
485 mddev_t *mddev = q->queuedata;
487 unplug_slaves(mddev);
488 md_wakeup_thread(mddev->thread);
491 static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
492 sector_t *error_sector)
494 mddev_t *mddev = q->queuedata;
495 conf_t *conf = mddev_to_conf(mddev);
499 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
500 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
501 if (rdev && !rdev->faulty) {
502 struct block_device *bdev = rdev->bdev;
503 request_queue_t *r_queue = bdev_get_queue(bdev);
505 if (!r_queue->issue_flush_fn)
508 atomic_inc(&rdev->nr_pending);
510 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
512 rdev_dec_pending(rdev, mddev);
522 * Throttle resync depth, so that we can both get proper overlapping of
523 * requests, but are still able to handle normal requests quickly.
525 #define RESYNC_DEPTH 32
527 static void device_barrier(conf_t *conf, sector_t sect)
529 spin_lock_irq(&conf->resync_lock);
530 wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
531 conf->resync_lock, raid1_unplug(conf->mddev->queue));
533 if (!conf->barrier++) {
534 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
535 conf->resync_lock, raid1_unplug(conf->mddev->queue));
536 if (conf->nr_pending)
539 wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
540 conf->resync_lock, raid1_unplug(conf->mddev->queue));
541 conf->next_resync = sect;
542 spin_unlock_irq(&conf->resync_lock);
545 static int make_request(request_queue_t *q, struct bio * bio)
547 mddev_t *mddev = q->queuedata;
548 conf_t *conf = mddev_to_conf(mddev);
549 mirror_info_t *mirror;
551 struct bio *read_bio;
552 int i, targets = 0, disks;
554 struct bitmap *bitmap = mddev->bitmap;
558 if (unlikely(bio_barrier(bio))) {
559 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
564 * Register the new request and wait if the reconstruction
565 * thread has put up a bar for new requests.
566 * Continue immediately if no resync is active currently.
568 md_write_start(mddev, bio); /* wait on superblock update early */
570 spin_lock_irq(&conf->resync_lock);
571 wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
573 spin_unlock_irq(&conf->resync_lock);
575 if (bio_data_dir(bio)==WRITE) {
576 disk_stat_inc(mddev->gendisk, writes);
577 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
579 disk_stat_inc(mddev->gendisk, reads);
580 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
584 * make_request() can abort the operation when READA is being
585 * used and no empty request is available.
588 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
590 r1_bio->master_bio = bio;
591 r1_bio->sectors = bio->bi_size >> 9;
593 r1_bio->mddev = mddev;
594 r1_bio->sector = bio->bi_sector;
598 if (bio_data_dir(bio) == READ) {
600 * read balancing logic:
602 int rdisk = read_balance(conf, r1_bio);
605 /* couldn't find anywhere to read from */
606 raid_end_bio_io(r1_bio);
609 mirror = conf->mirrors + rdisk;
611 r1_bio->read_disk = rdisk;
613 read_bio = bio_clone(bio, GFP_NOIO);
615 r1_bio->bios[rdisk] = read_bio;
617 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
618 read_bio->bi_bdev = mirror->rdev->bdev;
619 read_bio->bi_end_io = raid1_end_read_request;
620 read_bio->bi_rw = READ;
621 read_bio->bi_private = r1_bio;
623 generic_make_request(read_bio);
630 /* first select target devices under spinlock and
631 * inc refcount on their rdev. Record them by setting
634 disks = conf->raid_disks;
636 { static int first=1;
637 if (first) printk("First Write sector %llu disks %d\n",
638 (unsigned long long)r1_bio->sector, disks);
643 for (i = 0; i < disks; i++) {
644 if ((rdev=conf->mirrors[i].rdev) != NULL &&
646 atomic_inc(&rdev->nr_pending);
648 atomic_dec(&rdev->nr_pending);
649 r1_bio->bios[i] = NULL;
651 r1_bio->bios[i] = bio;
654 r1_bio->bios[i] = NULL;
658 if (targets < conf->raid_disks) {
659 /* array is degraded, we will not clear the bitmap
660 * on I/O completion (see raid1_end_write_request) */
661 set_bit(R1BIO_Degraded, &r1_bio->state);
664 atomic_set(&r1_bio->remaining, 0);
667 for (i = 0; i < disks; i++) {
669 if (!r1_bio->bios[i])
672 mbio = bio_clone(bio, GFP_NOIO);
673 r1_bio->bios[i] = mbio;
675 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
676 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
677 mbio->bi_end_io = raid1_end_write_request;
679 mbio->bi_private = r1_bio;
681 atomic_inc(&r1_bio->remaining);
683 bio_list_add(&bl, mbio);
686 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors);
687 spin_lock_irqsave(&conf->device_lock, flags);
688 bio_list_merge(&conf->pending_bio_list, &bl);
691 blk_plug_device(mddev->queue);
692 spin_unlock_irqrestore(&conf->device_lock, flags);
695 while ((bio = bio_list_pop(&bl)) != NULL)
696 generic_make_request(bio);
702 static void status(struct seq_file *seq, mddev_t *mddev)
704 conf_t *conf = mddev_to_conf(mddev);
707 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
708 conf->working_disks);
709 for (i = 0; i < conf->raid_disks; i++)
710 seq_printf(seq, "%s",
711 conf->mirrors[i].rdev &&
712 conf->mirrors[i].rdev->in_sync ? "U" : "_");
713 seq_printf(seq, "]");
717 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
719 char b[BDEVNAME_SIZE];
720 conf_t *conf = mddev_to_conf(mddev);
723 * If it is not operational, then we have already marked it as dead
724 * else if it is the last working disks, ignore the error, let the
725 * next level up know.
726 * else mark the drive as failed
729 && conf->working_disks == 1)
731 * Don't fail the drive, act as though we were just a
732 * normal single drive
737 conf->working_disks--;
739 * if recovery is running, make sure it aborts.
741 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
746 printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
747 " Operation continuing on %d devices\n",
748 bdevname(rdev->bdev,b), conf->working_disks);
751 static void print_conf(conf_t *conf)
756 printk("RAID1 conf printout:\n");
761 printk(" --- wd:%d rd:%d\n", conf->working_disks,
764 for (i = 0; i < conf->raid_disks; i++) {
765 char b[BDEVNAME_SIZE];
766 tmp = conf->mirrors + i;
768 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
769 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
770 bdevname(tmp->rdev->bdev,b));
774 static void close_sync(conf_t *conf)
776 spin_lock_irq(&conf->resync_lock);
777 wait_event_lock_irq(conf->wait_resume, !conf->barrier,
778 conf->resync_lock, raid1_unplug(conf->mddev->queue));
779 spin_unlock_irq(&conf->resync_lock);
781 if (conf->barrier) BUG();
782 if (waitqueue_active(&conf->wait_idle)) BUG();
784 mempool_destroy(conf->r1buf_pool);
785 conf->r1buf_pool = NULL;
788 static int raid1_spare_active(mddev_t *mddev)
791 conf_t *conf = mddev->private;
795 * Find all failed disks within the RAID1 configuration
796 * and mark them readable
798 for (i = 0; i < conf->raid_disks; i++) {
799 tmp = conf->mirrors + i;
801 && !tmp->rdev->faulty
802 && !tmp->rdev->in_sync) {
803 conf->working_disks++;
805 tmp->rdev->in_sync = 1;
814 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
816 conf_t *conf = mddev->private;
821 if (rdev->saved_raid_disk >= 0 &&
822 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
823 mirror = rdev->saved_raid_disk;
824 for (mirror=0; mirror < mddev->raid_disks; mirror++)
825 if ( !(p=conf->mirrors+mirror)->rdev) {
827 blk_queue_stack_limits(mddev->queue,
828 rdev->bdev->bd_disk->queue);
829 /* as we don't honour merge_bvec_fn, we must never risk
830 * violating it, so limit ->max_sector to one PAGE, as
831 * a one page request is never in violation.
833 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
834 mddev->queue->max_sectors > (PAGE_SIZE>>9))
835 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
837 p->head_position = 0;
838 rdev->raid_disk = mirror;
840 if (rdev->saved_raid_disk != mirror)
850 static int raid1_remove_disk(mddev_t *mddev, int number)
852 conf_t *conf = mddev->private;
855 mirror_info_t *p = conf->mirrors+ number;
861 atomic_read(&rdev->nr_pending)) {
867 if (atomic_read(&rdev->nr_pending)) {
868 /* lost the race, try later */
880 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
882 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
883 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
884 conf_t *conf = mddev_to_conf(r1_bio->mddev);
889 if (r1_bio->bios[r1_bio->read_disk] != bio)
891 update_head_pos(r1_bio->read_disk, r1_bio);
893 * we have read a block, now it needs to be re-written,
894 * or re-read if the read failed.
895 * We don't do much here, just schedule handling by raid1d
898 md_error(r1_bio->mddev,
899 conf->mirrors[r1_bio->read_disk].rdev);
901 set_bit(R1BIO_Uptodate, &r1_bio->state);
902 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
903 reschedule_retry(r1_bio);
907 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
909 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
910 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
911 mddev_t *mddev = r1_bio->mddev;
912 conf_t *conf = mddev_to_conf(mddev);
919 for (i = 0; i < conf->raid_disks; i++)
920 if (r1_bio->bios[i] == bio) {
925 md_error(mddev, conf->mirrors[mirror].rdev);
927 update_head_pos(mirror, r1_bio);
929 if (atomic_dec_and_test(&r1_bio->remaining)) {
930 md_done_sync(mddev, r1_bio->sectors, uptodate);
933 rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
937 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
939 conf_t *conf = mddev_to_conf(mddev);
941 int disks = conf->raid_disks;
942 struct bio *bio, *wbio;
944 bio = r1_bio->bios[r1_bio->read_disk];
947 if (r1_bio->sector == 0) printk("First sync write startss\n");
952 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
954 * There is no point trying a read-for-reconstruct as
955 * reconstruct is about to be aborted
957 char b[BDEVNAME_SIZE];
958 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
960 bdevname(bio->bi_bdev,b),
961 (unsigned long long)r1_bio->sector);
962 md_done_sync(mddev, r1_bio->sectors, 0);
967 atomic_set(&r1_bio->remaining, 1);
968 for (i = 0; i < disks ; i++) {
969 wbio = r1_bio->bios[i];
970 if (wbio->bi_end_io != end_sync_write)
973 atomic_inc(&conf->mirrors[i].rdev->nr_pending);
974 atomic_inc(&r1_bio->remaining);
975 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
977 generic_make_request(wbio);
980 if (atomic_dec_and_test(&r1_bio->remaining)) {
981 /* if we're here, all write(s) have completed, so clean up */
982 md_done_sync(mddev, r1_bio->sectors, 1);
988 * This is a kernel thread which:
990 * 1. Retries failed read operations on working mirrors.
991 * 2. Updates the raid superblock when problems encounter.
992 * 3. Performs writes following reads for array syncronising.
995 static void raid1d(mddev_t *mddev)
1000 conf_t *conf = mddev_to_conf(mddev);
1001 struct list_head *head = &conf->retry_list;
1005 md_check_recovery(mddev);
1008 char b[BDEVNAME_SIZE];
1009 spin_lock_irqsave(&conf->device_lock, flags);
1011 if (conf->pending_bio_list.head) {
1012 bio = bio_list_get(&conf->pending_bio_list);
1013 blk_remove_plug(mddev->queue);
1014 spin_unlock_irqrestore(&conf->device_lock, flags);
1015 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1016 if (bitmap_unplug(mddev->bitmap) != 0)
1017 printk("%s: bitmap file write failed!\n", mdname(mddev));
1019 while (bio) { /* submit pending writes */
1020 struct bio *next = bio->bi_next;
1021 bio->bi_next = NULL;
1022 generic_make_request(bio);
1030 if (list_empty(head))
1032 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1033 list_del(head->prev);
1034 spin_unlock_irqrestore(&conf->device_lock, flags);
1036 mddev = r1_bio->mddev;
1037 conf = mddev_to_conf(mddev);
1038 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1039 sync_request_write(mddev, r1_bio);
1043 bio = r1_bio->bios[r1_bio->read_disk];
1044 if ((disk=read_balance(conf, r1_bio)) == -1) {
1045 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1046 " read error for block %llu\n",
1047 bdevname(bio->bi_bdev,b),
1048 (unsigned long long)r1_bio->sector);
1049 raid_end_bio_io(r1_bio);
1051 r1_bio->bios[r1_bio->read_disk] = NULL;
1052 r1_bio->read_disk = disk;
1054 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1055 r1_bio->bios[r1_bio->read_disk] = bio;
1056 rdev = conf->mirrors[disk].rdev;
1057 if (printk_ratelimit())
1058 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1059 " another mirror\n",
1060 bdevname(rdev->bdev,b),
1061 (unsigned long long)r1_bio->sector);
1062 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1063 bio->bi_bdev = rdev->bdev;
1064 bio->bi_end_io = raid1_end_read_request;
1066 bio->bi_private = r1_bio;
1068 generic_make_request(bio);
1072 spin_unlock_irqrestore(&conf->device_lock, flags);
1074 unplug_slaves(mddev);
1078 static int init_resync(conf_t *conf)
1082 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1083 if (conf->r1buf_pool)
1085 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1087 if (!conf->r1buf_pool)
1089 conf->next_resync = 0;
1094 * perform a "sync" on one "block"
1096 * We need to make sure that no normal I/O request - particularly write
1097 * requests - conflict with active sync requests.
1099 * This is achieved by tracking pending requests and a 'barrier' concept
1100 * that can be installed to exclude normal IO requests.
1103 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1105 conf_t *conf = mddev_to_conf(mddev);
1106 mirror_info_t *mirror;
1109 sector_t max_sector, nr_sectors;
1112 int write_targets = 0;
1114 int still_degraded = 0;
1116 if (!conf->r1buf_pool)
1119 printk("sync start - bitmap %p\n", mddev->bitmap);
1121 if (init_resync(conf))
1125 max_sector = mddev->size << 1;
1126 if (sector_nr >= max_sector) {
1127 /* If we aborted, we need to abort the
1128 * sync on the 'current' bitmap chunk (there will
1129 * only be one in raid1 resync.
1130 * We can find the current addess in mddev->curr_resync
1132 if (mddev->curr_resync < max_sector) /* aborted */
1133 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1135 else /* completed sync */
1138 bitmap_close_sync(mddev->bitmap);
1143 /* before building a request, check if we can skip these blocks..
1144 * This call the bitmap_start_sync doesn't actually record anything
1146 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1148 /* We can skip this block, and probably several more */
1153 * If there is non-resync activity waiting for us then
1154 * put in a delay to throttle resync.
1156 if (!go_faster && waitqueue_active(&conf->wait_resume))
1157 msleep_interruptible(1000);
1158 device_barrier(conf, sector_nr + RESYNC_SECTORS);
1161 * If reconstructing, and >1 working disc,
1162 * could dedicate one to rebuild and others to
1163 * service read requests ..
1165 disk = conf->last_used;
1166 /* make sure disk is operational */
1168 while (conf->mirrors[disk].rdev == NULL ||
1169 !conf->mirrors[disk].rdev->in_sync) {
1171 disk = conf->raid_disks;
1173 if (disk == conf->last_used)
1176 conf->last_used = disk;
1177 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
1180 mirror = conf->mirrors + disk;
1182 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1184 spin_lock_irq(&conf->resync_lock);
1186 spin_unlock_irq(&conf->resync_lock);
1188 r1_bio->mddev = mddev;
1189 r1_bio->sector = sector_nr;
1191 set_bit(R1BIO_IsSync, &r1_bio->state);
1192 r1_bio->read_disk = disk;
1194 for (i=0; i < conf->raid_disks; i++) {
1195 bio = r1_bio->bios[i];
1197 /* take from bio_init */
1198 bio->bi_next = NULL;
1199 bio->bi_flags |= 1 << BIO_UPTODATE;
1203 bio->bi_phys_segments = 0;
1204 bio->bi_hw_segments = 0;
1206 bio->bi_end_io = NULL;
1207 bio->bi_private = NULL;
1211 bio->bi_end_io = end_sync_read;
1212 } else if (conf->mirrors[i].rdev == NULL ||
1213 conf->mirrors[i].rdev->faulty) {
1216 } else if (!conf->mirrors[i].rdev->in_sync ||
1217 sector_nr + RESYNC_SECTORS > mddev->recovery_cp) {
1219 bio->bi_end_io = end_sync_write;
1222 /* no need to read or write here */
1224 bio->bi_sector = sector_nr + conf->mirrors[i].rdev->data_offset;
1225 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1226 bio->bi_private = r1_bio;
1229 if (write_targets == 0) {
1230 /* There is nowhere to write, so all non-sync
1231 * drives must be failed - so we are finished
1233 sector_t rv = max_sector - sector_nr;
1236 rdev_dec_pending(conf->mirrors[disk].rdev, mddev);
1244 int len = PAGE_SIZE;
1245 if (sector_nr + (len>>9) > max_sector)
1246 len = (max_sector - sector_nr) << 9;
1249 if (sync_blocks == 0) {
1250 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1251 &sync_blocks, still_degraded) &&
1254 if (sync_blocks < (PAGE_SIZE>>9))
1256 if (len > (sync_blocks<<9))
1257 len = sync_blocks<<9;
1260 for (i=0 ; i < conf->raid_disks; i++) {
1261 bio = r1_bio->bios[i];
1262 if (bio->bi_end_io) {
1263 page = r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page;
1264 if (bio_add_page(bio, page, len, 0) == 0) {
1266 r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page = page;
1269 bio = r1_bio->bios[i];
1270 if (bio->bi_end_io==NULL)
1272 /* remove last page from this bio */
1274 bio->bi_size -= len;
1275 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1281 nr_sectors += len>>9;
1282 sector_nr += len>>9;
1283 sync_blocks -= (len>>9);
1284 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1286 bio = r1_bio->bios[disk];
1287 r1_bio->sectors = nr_sectors;
1289 md_sync_acct(mirror->rdev->bdev, nr_sectors);
1291 generic_make_request(bio);
1296 static int run(mddev_t *mddev)
1300 mirror_info_t *disk;
1302 struct list_head *tmp;
1304 if (mddev->level != 1) {
1305 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1306 mdname(mddev), mddev->level);
1310 * copy the already verified devices into our private RAID1
1311 * bookkeeping area. [whatever we allocate in run(),
1312 * should be freed in stop()]
1314 conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1315 mddev->private = conf;
1319 memset(conf, 0, sizeof(*conf));
1320 conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1325 memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1327 conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1328 if (!conf->poolinfo)
1330 conf->poolinfo->mddev = mddev;
1331 conf->poolinfo->raid_disks = mddev->raid_disks;
1332 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1335 if (!conf->r1bio_pool)
1338 ITERATE_RDEV(mddev, rdev, tmp) {
1339 disk_idx = rdev->raid_disk;
1340 if (disk_idx >= mddev->raid_disks
1343 disk = conf->mirrors + disk_idx;
1347 blk_queue_stack_limits(mddev->queue,
1348 rdev->bdev->bd_disk->queue);
1349 /* as we don't honour merge_bvec_fn, we must never risk
1350 * violating it, so limit ->max_sector to one PAGE, as
1351 * a one page request is never in violation.
1353 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1354 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1355 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1357 disk->head_position = 0;
1358 if (!rdev->faulty && rdev->in_sync)
1359 conf->working_disks++;
1361 conf->raid_disks = mddev->raid_disks;
1362 conf->mddev = mddev;
1363 spin_lock_init(&conf->device_lock);
1364 INIT_LIST_HEAD(&conf->retry_list);
1365 if (conf->working_disks == 1)
1366 mddev->recovery_cp = MaxSector;
1368 spin_lock_init(&conf->resync_lock);
1369 init_waitqueue_head(&conf->wait_idle);
1370 init_waitqueue_head(&conf->wait_resume);
1372 bio_list_init(&conf->pending_bio_list);
1373 bio_list_init(&conf->flushing_bio_list);
1375 if (!conf->working_disks) {
1376 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1381 mddev->degraded = 0;
1382 for (i = 0; i < conf->raid_disks; i++) {
1384 disk = conf->mirrors + i;
1387 disk->head_position = 0;
1393 * find the first working one and use it as a starting point
1394 * to read balancing.
1396 for (j = 0; j < conf->raid_disks &&
1397 (!conf->mirrors[j].rdev ||
1398 !conf->mirrors[j].rdev->in_sync) ; j++)
1400 conf->last_used = j;
1403 mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1404 if (!mddev->thread) {
1406 "raid1: couldn't allocate thread for %s\n",
1410 if (mddev->bitmap) mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1413 "raid1: raid set %s active with %d out of %d mirrors\n",
1414 mdname(mddev), mddev->raid_disks - mddev->degraded,
1417 * Ok, everything is just fine now
1419 mddev->array_size = mddev->size;
1421 mddev->queue->unplug_fn = raid1_unplug;
1422 mddev->queue->issue_flush_fn = raid1_issue_flush;
1427 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1432 if (conf->r1bio_pool)
1433 mempool_destroy(conf->r1bio_pool);
1434 kfree(conf->mirrors);
1435 kfree(conf->poolinfo);
1437 mddev->private = NULL;
1443 static int stop(mddev_t *mddev)
1445 conf_t *conf = mddev_to_conf(mddev);
1447 md_unregister_thread(mddev->thread);
1448 mddev->thread = NULL;
1449 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1450 if (conf->r1bio_pool)
1451 mempool_destroy(conf->r1bio_pool);
1452 kfree(conf->mirrors);
1453 kfree(conf->poolinfo);
1455 mddev->private = NULL;
1459 static int raid1_resize(mddev_t *mddev, sector_t sectors)
1461 /* no resync is happening, and there is enough space
1462 * on all devices, so we can resize.
1463 * We need to make sure resync covers any new space.
1464 * If the array is shrinking we should possibly wait until
1465 * any io in the removed space completes, but it hardly seems
1468 mddev->array_size = sectors>>1;
1469 set_capacity(mddev->gendisk, mddev->array_size << 1);
1471 if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
1472 mddev->recovery_cp = mddev->size << 1;
1473 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1475 mddev->size = mddev->array_size;
1476 mddev->resync_max_sectors = sectors;
1480 static int raid1_reshape(mddev_t *mddev, int raid_disks)
1483 * 1/ resize the r1bio_pool
1484 * 2/ resize conf->mirrors
1486 * We allocate a new r1bio_pool if we can.
1487 * Then raise a device barrier and wait until all IO stops.
1488 * Then resize conf->mirrors and swap in the new r1bio pool.
1490 * At the same time, we "pack" the devices so that all the missing
1491 * devices have the higher raid_disk numbers.
1493 mempool_t *newpool, *oldpool;
1494 struct pool_info *newpoolinfo;
1495 mirror_info_t *newmirrors;
1496 conf_t *conf = mddev_to_conf(mddev);
1501 if (raid_disks < conf->raid_disks) {
1503 for (d= 0; d < conf->raid_disks; d++)
1504 if (conf->mirrors[d].rdev)
1506 if (cnt > raid_disks)
1510 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
1513 newpoolinfo->mddev = mddev;
1514 newpoolinfo->raid_disks = raid_disks;
1516 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1517 r1bio_pool_free, newpoolinfo);
1522 newmirrors = kmalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
1525 mempool_destroy(newpool);
1528 memset(newmirrors, 0, sizeof(struct mirror_info)*raid_disks);
1530 spin_lock_irq(&conf->resync_lock);
1532 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
1533 conf->resync_lock, raid1_unplug(mddev->queue));
1534 spin_unlock_irq(&conf->resync_lock);
1536 /* ok, everything is stopped */
1537 oldpool = conf->r1bio_pool;
1538 conf->r1bio_pool = newpool;
1540 for (d=d2=0; d < conf->raid_disks; d++)
1541 if (conf->mirrors[d].rdev) {
1542 conf->mirrors[d].rdev->raid_disk = d2;
1543 newmirrors[d2++].rdev = conf->mirrors[d].rdev;
1545 kfree(conf->mirrors);
1546 conf->mirrors = newmirrors;
1547 kfree(conf->poolinfo);
1548 conf->poolinfo = newpoolinfo;
1550 mddev->degraded += (raid_disks - conf->raid_disks);
1551 conf->raid_disks = mddev->raid_disks = raid_disks;
1553 conf->last_used = 0; /* just make sure it is in-range */
1554 spin_lock_irq(&conf->resync_lock);
1556 spin_unlock_irq(&conf->resync_lock);
1557 wake_up(&conf->wait_resume);
1558 wake_up(&conf->wait_idle);
1561 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1562 md_wakeup_thread(mddev->thread);
1564 mempool_destroy(oldpool);
1569 static mdk_personality_t raid1_personality =
1572 .owner = THIS_MODULE,
1573 .make_request = make_request,
1577 .error_handler = error,
1578 .hot_add_disk = raid1_add_disk,
1579 .hot_remove_disk= raid1_remove_disk,
1580 .spare_active = raid1_spare_active,
1581 .sync_request = sync_request,
1582 .resize = raid1_resize,
1583 .reshape = raid1_reshape,
1586 static int __init raid_init(void)
1588 return register_md_personality(RAID1, &raid1_personality);
1591 static void raid_exit(void)
1593 unregister_md_personality(RAID1);
1596 module_init(raid_init);
1597 module_exit(raid_exit);
1598 MODULE_LICENSE("GPL");
1599 MODULE_ALIAS("md-personality-3"); /* RAID1 */