2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
42 * The data to be stored is divided into chunks using chunksize.
43 * Each device is divided into far_copies sections.
44 * In each section, chunks are laid out in a style similar to raid0, but
45 * near_copies copies of each chunk is stored (each on a different drive).
46 * The starting device for each section is offset near_copies from the starting
47 * device of the previous section.
48 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
50 * near_copies and far_copies must be at least one, and their product is at most
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of be very far apart
55 * on disk, there are adjacent stripes.
59 * Number of guaranteed r10bios in case of extreme VM load:
61 #define NR_RAID10_BIOS 256
63 /* When there are this many requests queue to be written by
64 * the raid10 thread, we become 'congested' to provide back-pressure
67 static int max_queued_requests = 1024;
69 static void allow_barrier(struct r10conf *conf);
70 static void lower_barrier(struct r10conf *conf);
71 static int enough(struct r10conf *conf, int ignore);
72 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
74 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
75 static void end_reshape_write(struct bio *bio, int error);
76 static void end_reshape(struct r10conf *conf);
78 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
80 struct r10conf *conf = data;
81 int size = offsetof(struct r10bio, devs[conf->copies]);
83 /* allocate a r10bio with room for raid_disks entries in the
85 return kzalloc(size, gfp_flags);
88 static void r10bio_pool_free(void *r10_bio, void *data)
93 /* Maximum size of each resync request */
94 #define RESYNC_BLOCK_SIZE (64*1024)
95 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
96 /* amount of memory to reserve for resync requests */
97 #define RESYNC_WINDOW (1024*1024)
98 /* maximum number of concurrent requests, memory permitting */
99 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
102 * When performing a resync, we need to read and compare, so
103 * we need as many pages are there are copies.
104 * When performing a recovery, we need 2 bios, one for read,
105 * one for write (we recover only one drive per r10buf)
108 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
110 struct r10conf *conf = data;
112 struct r10bio *r10_bio;
117 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
121 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
122 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
123 nalloc = conf->copies; /* resync */
125 nalloc = 2; /* recovery */
130 for (j = nalloc ; j-- ; ) {
131 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
134 r10_bio->devs[j].bio = bio;
135 if (!conf->have_replacement)
137 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
140 r10_bio->devs[j].repl_bio = bio;
143 * Allocate RESYNC_PAGES data pages and attach them
146 for (j = 0 ; j < nalloc; j++) {
147 struct bio *rbio = r10_bio->devs[j].repl_bio;
148 bio = r10_bio->devs[j].bio;
149 for (i = 0; i < RESYNC_PAGES; i++) {
150 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
151 &conf->mddev->recovery)) {
152 /* we can share bv_page's during recovery
154 struct bio *rbio = r10_bio->devs[0].bio;
155 page = rbio->bi_io_vec[i].bv_page;
158 page = alloc_page(gfp_flags);
162 bio->bi_io_vec[i].bv_page = page;
164 rbio->bi_io_vec[i].bv_page = page;
172 safe_put_page(bio->bi_io_vec[i-1].bv_page);
174 for (i = 0; i < RESYNC_PAGES ; i++)
175 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
178 for ( ; j < nalloc; j++) {
179 if (r10_bio->devs[j].bio)
180 bio_put(r10_bio->devs[j].bio);
181 if (r10_bio->devs[j].repl_bio)
182 bio_put(r10_bio->devs[j].repl_bio);
184 r10bio_pool_free(r10_bio, conf);
188 static void r10buf_pool_free(void *__r10_bio, void *data)
191 struct r10conf *conf = data;
192 struct r10bio *r10bio = __r10_bio;
195 for (j=0; j < conf->copies; j++) {
196 struct bio *bio = r10bio->devs[j].bio;
198 for (i = 0; i < RESYNC_PAGES; i++) {
199 safe_put_page(bio->bi_io_vec[i].bv_page);
200 bio->bi_io_vec[i].bv_page = NULL;
204 bio = r10bio->devs[j].repl_bio;
208 r10bio_pool_free(r10bio, conf);
211 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
215 for (i = 0; i < conf->copies; i++) {
216 struct bio **bio = & r10_bio->devs[i].bio;
217 if (!BIO_SPECIAL(*bio))
220 bio = &r10_bio->devs[i].repl_bio;
221 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
227 static void free_r10bio(struct r10bio *r10_bio)
229 struct r10conf *conf = r10_bio->mddev->private;
231 put_all_bios(conf, r10_bio);
232 mempool_free(r10_bio, conf->r10bio_pool);
235 static void put_buf(struct r10bio *r10_bio)
237 struct r10conf *conf = r10_bio->mddev->private;
239 mempool_free(r10_bio, conf->r10buf_pool);
244 static void reschedule_retry(struct r10bio *r10_bio)
247 struct mddev *mddev = r10_bio->mddev;
248 struct r10conf *conf = mddev->private;
250 spin_lock_irqsave(&conf->device_lock, flags);
251 list_add(&r10_bio->retry_list, &conf->retry_list);
253 spin_unlock_irqrestore(&conf->device_lock, flags);
255 /* wake up frozen array... */
256 wake_up(&conf->wait_barrier);
258 md_wakeup_thread(mddev->thread);
262 * raid_end_bio_io() is called when we have finished servicing a mirrored
263 * operation and are ready to return a success/failure code to the buffer
266 static void raid_end_bio_io(struct r10bio *r10_bio)
268 struct bio *bio = r10_bio->master_bio;
270 struct r10conf *conf = r10_bio->mddev->private;
272 if (bio->bi_phys_segments) {
274 spin_lock_irqsave(&conf->device_lock, flags);
275 bio->bi_phys_segments--;
276 done = (bio->bi_phys_segments == 0);
277 spin_unlock_irqrestore(&conf->device_lock, flags);
280 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
281 clear_bit(BIO_UPTODATE, &bio->bi_flags);
285 * Wake up any possible resync thread that waits for the device
290 free_r10bio(r10_bio);
294 * Update disk head position estimator based on IRQ completion info.
296 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
298 struct r10conf *conf = r10_bio->mddev->private;
300 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
301 r10_bio->devs[slot].addr + (r10_bio->sectors);
305 * Find the disk number which triggered given bio
307 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
308 struct bio *bio, int *slotp, int *replp)
313 for (slot = 0; slot < conf->copies; slot++) {
314 if (r10_bio->devs[slot].bio == bio)
316 if (r10_bio->devs[slot].repl_bio == bio) {
322 BUG_ON(slot == conf->copies);
323 update_head_pos(slot, r10_bio);
329 return r10_bio->devs[slot].devnum;
332 static void raid10_end_read_request(struct bio *bio, int error)
334 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
335 struct r10bio *r10_bio = bio->bi_private;
337 struct md_rdev *rdev;
338 struct r10conf *conf = r10_bio->mddev->private;
341 slot = r10_bio->read_slot;
342 dev = r10_bio->devs[slot].devnum;
343 rdev = r10_bio->devs[slot].rdev;
345 * this branch is our 'one mirror IO has finished' event handler:
347 update_head_pos(slot, r10_bio);
351 * Set R10BIO_Uptodate in our master bio, so that
352 * we will return a good error code to the higher
353 * levels even if IO on some other mirrored buffer fails.
355 * The 'master' represents the composite IO operation to
356 * user-side. So if something waits for IO, then it will
357 * wait for the 'master' bio.
359 set_bit(R10BIO_Uptodate, &r10_bio->state);
361 /* If all other devices that store this block have
362 * failed, we want to return the error upwards rather
363 * than fail the last device. Here we redefine
364 * "uptodate" to mean "Don't want to retry"
367 spin_lock_irqsave(&conf->device_lock, flags);
368 if (!enough(conf, rdev->raid_disk))
370 spin_unlock_irqrestore(&conf->device_lock, flags);
373 raid_end_bio_io(r10_bio);
374 rdev_dec_pending(rdev, conf->mddev);
377 * oops, read error - keep the refcount on the rdev
379 char b[BDEVNAME_SIZE];
380 printk_ratelimited(KERN_ERR
381 "md/raid10:%s: %s: rescheduling sector %llu\n",
383 bdevname(rdev->bdev, b),
384 (unsigned long long)r10_bio->sector);
385 set_bit(R10BIO_ReadError, &r10_bio->state);
386 reschedule_retry(r10_bio);
390 static void close_write(struct r10bio *r10_bio)
392 /* clear the bitmap if all writes complete successfully */
393 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
395 !test_bit(R10BIO_Degraded, &r10_bio->state),
397 md_write_end(r10_bio->mddev);
400 static void one_write_done(struct r10bio *r10_bio)
402 if (atomic_dec_and_test(&r10_bio->remaining)) {
403 if (test_bit(R10BIO_WriteError, &r10_bio->state))
404 reschedule_retry(r10_bio);
406 close_write(r10_bio);
407 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
408 reschedule_retry(r10_bio);
410 raid_end_bio_io(r10_bio);
415 static void raid10_end_write_request(struct bio *bio, int error)
417 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
418 struct r10bio *r10_bio = bio->bi_private;
421 struct r10conf *conf = r10_bio->mddev->private;
423 struct md_rdev *rdev = NULL;
425 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
428 rdev = conf->mirrors[dev].replacement;
432 rdev = conf->mirrors[dev].rdev;
435 * this branch is our 'one mirror IO has finished' event handler:
439 /* Never record new bad blocks to replacement,
442 md_error(rdev->mddev, rdev);
444 set_bit(WriteErrorSeen, &rdev->flags);
445 if (!test_and_set_bit(WantReplacement, &rdev->flags))
446 set_bit(MD_RECOVERY_NEEDED,
447 &rdev->mddev->recovery);
448 set_bit(R10BIO_WriteError, &r10_bio->state);
453 * Set R10BIO_Uptodate in our master bio, so that
454 * we will return a good error code for to the higher
455 * levels even if IO on some other mirrored buffer fails.
457 * The 'master' represents the composite IO operation to
458 * user-side. So if something waits for IO, then it will
459 * wait for the 'master' bio.
464 set_bit(R10BIO_Uptodate, &r10_bio->state);
466 /* Maybe we can clear some bad blocks. */
467 if (is_badblock(rdev,
468 r10_bio->devs[slot].addr,
470 &first_bad, &bad_sectors)) {
473 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
475 r10_bio->devs[slot].bio = IO_MADE_GOOD;
477 set_bit(R10BIO_MadeGood, &r10_bio->state);
483 * Let's see if all mirrored write operations have finished
486 one_write_done(r10_bio);
488 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
492 * RAID10 layout manager
493 * As well as the chunksize and raid_disks count, there are two
494 * parameters: near_copies and far_copies.
495 * near_copies * far_copies must be <= raid_disks.
496 * Normally one of these will be 1.
497 * If both are 1, we get raid0.
498 * If near_copies == raid_disks, we get raid1.
500 * Chunks are laid out in raid0 style with near_copies copies of the
501 * first chunk, followed by near_copies copies of the next chunk and
503 * If far_copies > 1, then after 1/far_copies of the array has been assigned
504 * as described above, we start again with a device offset of near_copies.
505 * So we effectively have another copy of the whole array further down all
506 * the drives, but with blocks on different drives.
507 * With this layout, and block is never stored twice on the one device.
509 * raid10_find_phys finds the sector offset of a given virtual sector
510 * on each device that it is on.
512 * raid10_find_virt does the reverse mapping, from a device and a
513 * sector offset to a virtual address
516 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
525 /* now calculate first sector/dev */
526 chunk = r10bio->sector >> geo->chunk_shift;
527 sector = r10bio->sector & geo->chunk_mask;
529 chunk *= geo->near_copies;
531 dev = sector_div(stripe, geo->raid_disks);
533 stripe *= geo->far_copies;
535 sector += stripe << geo->chunk_shift;
537 /* and calculate all the others */
538 for (n = 0; n < geo->near_copies; n++) {
541 r10bio->devs[slot].addr = sector;
542 r10bio->devs[slot].devnum = d;
545 for (f = 1; f < geo->far_copies; f++) {
546 d += geo->near_copies;
547 if (d >= geo->raid_disks)
548 d -= geo->raid_disks;
550 r10bio->devs[slot].devnum = d;
551 r10bio->devs[slot].addr = s;
555 if (dev >= geo->raid_disks) {
557 sector += (geo->chunk_mask + 1);
562 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
564 struct geom *geo = &conf->geo;
566 if (conf->reshape_progress != MaxSector &&
567 ((r10bio->sector >= conf->reshape_progress) !=
568 conf->mddev->reshape_backwards)) {
569 set_bit(R10BIO_Previous, &r10bio->state);
572 clear_bit(R10BIO_Previous, &r10bio->state);
574 __raid10_find_phys(geo, r10bio);
577 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
579 sector_t offset, chunk, vchunk;
580 /* Never use conf->prev as this is only called during resync
581 * or recovery, so reshape isn't happening
583 struct geom *geo = &conf->geo;
585 offset = sector & geo->chunk_mask;
586 if (geo->far_offset) {
588 chunk = sector >> geo->chunk_shift;
589 fc = sector_div(chunk, geo->far_copies);
590 dev -= fc * geo->near_copies;
592 dev += geo->raid_disks;
594 while (sector >= geo->stride) {
595 sector -= geo->stride;
596 if (dev < geo->near_copies)
597 dev += geo->raid_disks - geo->near_copies;
599 dev -= geo->near_copies;
601 chunk = sector >> geo->chunk_shift;
603 vchunk = chunk * geo->raid_disks + dev;
604 sector_div(vchunk, geo->near_copies);
605 return (vchunk << geo->chunk_shift) + offset;
609 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
611 * @bvm: properties of new bio
612 * @biovec: the request that could be merged to it.
614 * Return amount of bytes we can accept at this offset
615 * This requires checking for end-of-chunk if near_copies != raid_disks,
616 * and for subordinate merge_bvec_fns if merge_check_needed.
618 static int raid10_mergeable_bvec(struct request_queue *q,
619 struct bvec_merge_data *bvm,
620 struct bio_vec *biovec)
622 struct mddev *mddev = q->queuedata;
623 struct r10conf *conf = mddev->private;
624 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
626 unsigned int chunk_sectors;
627 unsigned int bio_sectors = bvm->bi_size >> 9;
628 struct geom *geo = &conf->geo;
630 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
631 if (conf->reshape_progress != MaxSector &&
632 ((sector >= conf->reshape_progress) !=
633 conf->mddev->reshape_backwards))
636 if (geo->near_copies < geo->raid_disks) {
637 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
638 + bio_sectors)) << 9;
640 /* bio_add cannot handle a negative return */
642 if (max <= biovec->bv_len && bio_sectors == 0)
643 return biovec->bv_len;
645 max = biovec->bv_len;
647 if (mddev->merge_check_needed) {
648 struct r10bio r10_bio;
650 if (conf->reshape_progress != MaxSector) {
651 /* Cannot give any guidance during reshape */
652 if (max <= biovec->bv_len && bio_sectors == 0)
653 return biovec->bv_len;
656 r10_bio.sector = sector;
657 raid10_find_phys(conf, &r10_bio);
659 for (s = 0; s < conf->copies; s++) {
660 int disk = r10_bio.devs[s].devnum;
661 struct md_rdev *rdev = rcu_dereference(
662 conf->mirrors[disk].rdev);
663 if (rdev && !test_bit(Faulty, &rdev->flags)) {
664 struct request_queue *q =
665 bdev_get_queue(rdev->bdev);
666 if (q->merge_bvec_fn) {
667 bvm->bi_sector = r10_bio.devs[s].addr
669 bvm->bi_bdev = rdev->bdev;
670 max = min(max, q->merge_bvec_fn(
674 rdev = rcu_dereference(conf->mirrors[disk].replacement);
675 if (rdev && !test_bit(Faulty, &rdev->flags)) {
676 struct request_queue *q =
677 bdev_get_queue(rdev->bdev);
678 if (q->merge_bvec_fn) {
679 bvm->bi_sector = r10_bio.devs[s].addr
681 bvm->bi_bdev = rdev->bdev;
682 max = min(max, q->merge_bvec_fn(
693 * This routine returns the disk from which the requested read should
694 * be done. There is a per-array 'next expected sequential IO' sector
695 * number - if this matches on the next IO then we use the last disk.
696 * There is also a per-disk 'last know head position' sector that is
697 * maintained from IRQ contexts, both the normal and the resync IO
698 * completion handlers update this position correctly. If there is no
699 * perfect sequential match then we pick the disk whose head is closest.
701 * If there are 2 mirrors in the same 2 devices, performance degrades
702 * because position is mirror, not device based.
704 * The rdev for the device selected will have nr_pending incremented.
708 * FIXME: possibly should rethink readbalancing and do it differently
709 * depending on near_copies / far_copies geometry.
711 static struct md_rdev *read_balance(struct r10conf *conf,
712 struct r10bio *r10_bio,
715 const sector_t this_sector = r10_bio->sector;
717 int sectors = r10_bio->sectors;
718 int best_good_sectors;
719 sector_t new_distance, best_dist;
720 struct md_rdev *rdev, *best_rdev;
723 struct geom *geo = &conf->geo;
725 raid10_find_phys(conf, r10_bio);
728 sectors = r10_bio->sectors;
731 best_dist = MaxSector;
732 best_good_sectors = 0;
735 * Check if we can balance. We can balance on the whole
736 * device if no resync is going on (recovery is ok), or below
737 * the resync window. We take the first readable disk when
738 * above the resync window.
740 if (conf->mddev->recovery_cp < MaxSector
741 && (this_sector + sectors >= conf->next_resync))
744 for (slot = 0; slot < conf->copies ; slot++) {
749 if (r10_bio->devs[slot].bio == IO_BLOCKED)
751 disk = r10_bio->devs[slot].devnum;
752 rdev = rcu_dereference(conf->mirrors[disk].replacement);
753 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
754 test_bit(Unmerged, &rdev->flags) ||
755 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
756 rdev = rcu_dereference(conf->mirrors[disk].rdev);
758 test_bit(Faulty, &rdev->flags) ||
759 test_bit(Unmerged, &rdev->flags))
761 if (!test_bit(In_sync, &rdev->flags) &&
762 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
765 dev_sector = r10_bio->devs[slot].addr;
766 if (is_badblock(rdev, dev_sector, sectors,
767 &first_bad, &bad_sectors)) {
768 if (best_dist < MaxSector)
769 /* Already have a better slot */
771 if (first_bad <= dev_sector) {
772 /* Cannot read here. If this is the
773 * 'primary' device, then we must not read
774 * beyond 'bad_sectors' from another device.
776 bad_sectors -= (dev_sector - first_bad);
777 if (!do_balance && sectors > bad_sectors)
778 sectors = bad_sectors;
779 if (best_good_sectors > sectors)
780 best_good_sectors = sectors;
782 sector_t good_sectors =
783 first_bad - dev_sector;
784 if (good_sectors > best_good_sectors) {
785 best_good_sectors = good_sectors;
790 /* Must read from here */
795 best_good_sectors = sectors;
800 /* This optimisation is debatable, and completely destroys
801 * sequential read speed for 'far copies' arrays. So only
802 * keep it for 'near' arrays, and review those later.
804 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
807 /* for far > 1 always use the lowest address */
808 if (geo->far_copies > 1)
809 new_distance = r10_bio->devs[slot].addr;
811 new_distance = abs(r10_bio->devs[slot].addr -
812 conf->mirrors[disk].head_position);
813 if (new_distance < best_dist) {
814 best_dist = new_distance;
819 if (slot >= conf->copies) {
825 atomic_inc(&rdev->nr_pending);
826 if (test_bit(Faulty, &rdev->flags)) {
827 /* Cannot risk returning a device that failed
828 * before we inc'ed nr_pending
830 rdev_dec_pending(rdev, conf->mddev);
833 r10_bio->read_slot = slot;
837 *max_sectors = best_good_sectors;
842 static int raid10_congested(void *data, int bits)
844 struct mddev *mddev = data;
845 struct r10conf *conf = mddev->private;
848 if ((bits & (1 << BDI_async_congested)) &&
849 conf->pending_count >= max_queued_requests)
852 if (mddev_congested(mddev, bits))
856 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
859 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
860 if (rdev && !test_bit(Faulty, &rdev->flags)) {
861 struct request_queue *q = bdev_get_queue(rdev->bdev);
863 ret |= bdi_congested(&q->backing_dev_info, bits);
870 static void flush_pending_writes(struct r10conf *conf)
872 /* Any writes that have been queued but are awaiting
873 * bitmap updates get flushed here.
875 spin_lock_irq(&conf->device_lock);
877 if (conf->pending_bio_list.head) {
879 bio = bio_list_get(&conf->pending_bio_list);
880 conf->pending_count = 0;
881 spin_unlock_irq(&conf->device_lock);
882 /* flush any pending bitmap writes to disk
883 * before proceeding w/ I/O */
884 bitmap_unplug(conf->mddev->bitmap);
885 wake_up(&conf->wait_barrier);
887 while (bio) { /* submit pending writes */
888 struct bio *next = bio->bi_next;
890 generic_make_request(bio);
894 spin_unlock_irq(&conf->device_lock);
898 * Sometimes we need to suspend IO while we do something else,
899 * either some resync/recovery, or reconfigure the array.
900 * To do this we raise a 'barrier'.
901 * The 'barrier' is a counter that can be raised multiple times
902 * to count how many activities are happening which preclude
904 * We can only raise the barrier if there is no pending IO.
905 * i.e. if nr_pending == 0.
906 * We choose only to raise the barrier if no-one is waiting for the
907 * barrier to go down. This means that as soon as an IO request
908 * is ready, no other operations which require a barrier will start
909 * until the IO request has had a chance.
911 * So: regular IO calls 'wait_barrier'. When that returns there
912 * is no backgroup IO happening, It must arrange to call
913 * allow_barrier when it has finished its IO.
914 * backgroup IO calls must call raise_barrier. Once that returns
915 * there is no normal IO happeing. It must arrange to call
916 * lower_barrier when the particular background IO completes.
919 static void raise_barrier(struct r10conf *conf, int force)
921 BUG_ON(force && !conf->barrier);
922 spin_lock_irq(&conf->resync_lock);
924 /* Wait until no block IO is waiting (unless 'force') */
925 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
926 conf->resync_lock, );
928 /* block any new IO from starting */
931 /* Now wait for all pending IO to complete */
932 wait_event_lock_irq(conf->wait_barrier,
933 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
934 conf->resync_lock, );
936 spin_unlock_irq(&conf->resync_lock);
939 static void lower_barrier(struct r10conf *conf)
942 spin_lock_irqsave(&conf->resync_lock, flags);
944 spin_unlock_irqrestore(&conf->resync_lock, flags);
945 wake_up(&conf->wait_barrier);
948 static void wait_barrier(struct r10conf *conf)
950 spin_lock_irq(&conf->resync_lock);
953 /* Wait for the barrier to drop.
954 * However if there are already pending
955 * requests (preventing the barrier from
956 * rising completely), and the
957 * pre-process bio queue isn't empty,
958 * then don't wait, as we need to empty
959 * that queue to get the nr_pending
962 wait_event_lock_irq(conf->wait_barrier,
966 !bio_list_empty(current->bio_list)),
972 spin_unlock_irq(&conf->resync_lock);
975 static void allow_barrier(struct r10conf *conf)
978 spin_lock_irqsave(&conf->resync_lock, flags);
980 spin_unlock_irqrestore(&conf->resync_lock, flags);
981 wake_up(&conf->wait_barrier);
984 static void freeze_array(struct r10conf *conf)
986 /* stop syncio and normal IO and wait for everything to
988 * We increment barrier and nr_waiting, and then
989 * wait until nr_pending match nr_queued+1
990 * This is called in the context of one normal IO request
991 * that has failed. Thus any sync request that might be pending
992 * will be blocked by nr_pending, and we need to wait for
993 * pending IO requests to complete or be queued for re-try.
994 * Thus the number queued (nr_queued) plus this request (1)
995 * must match the number of pending IOs (nr_pending) before
998 spin_lock_irq(&conf->resync_lock);
1001 wait_event_lock_irq(conf->wait_barrier,
1002 conf->nr_pending == conf->nr_queued+1,
1004 flush_pending_writes(conf));
1006 spin_unlock_irq(&conf->resync_lock);
1009 static void unfreeze_array(struct r10conf *conf)
1011 /* reverse the effect of the freeze */
1012 spin_lock_irq(&conf->resync_lock);
1015 wake_up(&conf->wait_barrier);
1016 spin_unlock_irq(&conf->resync_lock);
1019 static sector_t choose_data_offset(struct r10bio *r10_bio,
1020 struct md_rdev *rdev)
1022 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1023 test_bit(R10BIO_Previous, &r10_bio->state))
1024 return rdev->data_offset;
1026 return rdev->new_data_offset;
1029 static void make_request(struct mddev *mddev, struct bio * bio)
1031 struct r10conf *conf = mddev->private;
1032 struct r10bio *r10_bio;
1033 struct bio *read_bio;
1035 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1036 int chunk_sects = chunk_mask + 1;
1037 const int rw = bio_data_dir(bio);
1038 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1039 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1040 unsigned long flags;
1041 struct md_rdev *blocked_rdev;
1043 int sectors_handled;
1047 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1048 md_flush_request(mddev, bio);
1052 /* If this request crosses a chunk boundary, we need to
1053 * split it. This will only happen for 1 PAGE (or less) requests.
1055 if (unlikely((bio->bi_sector & chunk_mask) + (bio->bi_size >> 9)
1057 && (conf->geo.near_copies < conf->geo.raid_disks
1058 || conf->prev.near_copies < conf->prev.raid_disks))) {
1059 struct bio_pair *bp;
1060 /* Sanity check -- queue functions should prevent this happening */
1061 if (bio->bi_vcnt != 1 ||
1064 /* This is a one page bio that upper layers
1065 * refuse to split for us, so we need to split it.
1068 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1070 /* Each of these 'make_request' calls will call 'wait_barrier'.
1071 * If the first succeeds but the second blocks due to the resync
1072 * thread raising the barrier, we will deadlock because the
1073 * IO to the underlying device will be queued in generic_make_request
1074 * and will never complete, so will never reduce nr_pending.
1075 * So increment nr_waiting here so no new raise_barriers will
1076 * succeed, and so the second wait_barrier cannot block.
1078 spin_lock_irq(&conf->resync_lock);
1080 spin_unlock_irq(&conf->resync_lock);
1082 make_request(mddev, &bp->bio1);
1083 make_request(mddev, &bp->bio2);
1085 spin_lock_irq(&conf->resync_lock);
1087 wake_up(&conf->wait_barrier);
1088 spin_unlock_irq(&conf->resync_lock);
1090 bio_pair_release(bp);
1093 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1094 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1095 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
1101 md_write_start(mddev, bio);
1104 * Register the new request and wait if the reconstruction
1105 * thread has put up a bar for new requests.
1106 * Continue immediately if no resync is active currently.
1110 sectors = bio->bi_size >> 9;
1111 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1112 bio->bi_sector < conf->reshape_progress &&
1113 bio->bi_sector + sectors > conf->reshape_progress) {
1114 /* IO spans the reshape position. Need to wait for
1117 allow_barrier(conf);
1118 wait_event(conf->wait_barrier,
1119 conf->reshape_progress <= bio->bi_sector ||
1120 conf->reshape_progress >= bio->bi_sector + sectors);
1123 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1124 bio_data_dir(bio) == WRITE &&
1125 (mddev->reshape_backwards
1126 ? (bio->bi_sector < conf->reshape_safe &&
1127 bio->bi_sector + sectors > conf->reshape_progress)
1128 : (bio->bi_sector + sectors > conf->reshape_safe &&
1129 bio->bi_sector < conf->reshape_progress))) {
1130 /* Need to update reshape_position in metadata */
1131 mddev->reshape_position = conf->reshape_progress;
1132 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1133 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1134 md_wakeup_thread(mddev->thread);
1135 wait_event(mddev->sb_wait,
1136 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1138 conf->reshape_safe = mddev->reshape_position;
1141 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1143 r10_bio->master_bio = bio;
1144 r10_bio->sectors = sectors;
1146 r10_bio->mddev = mddev;
1147 r10_bio->sector = bio->bi_sector;
1150 /* We might need to issue multiple reads to different
1151 * devices if there are bad blocks around, so we keep
1152 * track of the number of reads in bio->bi_phys_segments.
1153 * If this is 0, there is only one r10_bio and no locking
1154 * will be needed when the request completes. If it is
1155 * non-zero, then it is the number of not-completed requests.
1157 bio->bi_phys_segments = 0;
1158 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1162 * read balancing logic:
1164 struct md_rdev *rdev;
1168 rdev = read_balance(conf, r10_bio, &max_sectors);
1170 raid_end_bio_io(r10_bio);
1173 slot = r10_bio->read_slot;
1175 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1176 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1179 r10_bio->devs[slot].bio = read_bio;
1180 r10_bio->devs[slot].rdev = rdev;
1182 read_bio->bi_sector = r10_bio->devs[slot].addr +
1183 choose_data_offset(r10_bio, rdev);
1184 read_bio->bi_bdev = rdev->bdev;
1185 read_bio->bi_end_io = raid10_end_read_request;
1186 read_bio->bi_rw = READ | do_sync;
1187 read_bio->bi_private = r10_bio;
1189 if (max_sectors < r10_bio->sectors) {
1190 /* Could not read all from this device, so we will
1191 * need another r10_bio.
1193 sectors_handled = (r10_bio->sectors + max_sectors
1195 r10_bio->sectors = max_sectors;
1196 spin_lock_irq(&conf->device_lock);
1197 if (bio->bi_phys_segments == 0)
1198 bio->bi_phys_segments = 2;
1200 bio->bi_phys_segments++;
1201 spin_unlock(&conf->device_lock);
1202 /* Cannot call generic_make_request directly
1203 * as that will be queued in __generic_make_request
1204 * and subsequent mempool_alloc might block
1205 * waiting for it. so hand bio over to raid10d.
1207 reschedule_retry(r10_bio);
1209 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1211 r10_bio->master_bio = bio;
1212 r10_bio->sectors = ((bio->bi_size >> 9)
1215 r10_bio->mddev = mddev;
1216 r10_bio->sector = bio->bi_sector + sectors_handled;
1219 generic_make_request(read_bio);
1226 if (conf->pending_count >= max_queued_requests) {
1227 md_wakeup_thread(mddev->thread);
1228 wait_event(conf->wait_barrier,
1229 conf->pending_count < max_queued_requests);
1231 /* first select target devices under rcu_lock and
1232 * inc refcount on their rdev. Record them by setting
1234 * If there are known/acknowledged bad blocks on any device
1235 * on which we have seen a write error, we want to avoid
1236 * writing to those blocks. This potentially requires several
1237 * writes to write around the bad blocks. Each set of writes
1238 * gets its own r10_bio with a set of bios attached. The number
1239 * of r10_bios is recored in bio->bi_phys_segments just as with
1242 plugged = mddev_check_plugged(mddev);
1244 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1245 raid10_find_phys(conf, r10_bio);
1247 blocked_rdev = NULL;
1249 max_sectors = r10_bio->sectors;
1251 for (i = 0; i < conf->copies; i++) {
1252 int d = r10_bio->devs[i].devnum;
1253 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1254 struct md_rdev *rrdev = rcu_dereference(
1255 conf->mirrors[d].replacement);
1258 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1259 atomic_inc(&rdev->nr_pending);
1260 blocked_rdev = rdev;
1263 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1264 atomic_inc(&rrdev->nr_pending);
1265 blocked_rdev = rrdev;
1268 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1269 || test_bit(Unmerged, &rrdev->flags)))
1272 r10_bio->devs[i].bio = NULL;
1273 r10_bio->devs[i].repl_bio = NULL;
1274 if (!rdev || test_bit(Faulty, &rdev->flags) ||
1275 test_bit(Unmerged, &rdev->flags)) {
1276 set_bit(R10BIO_Degraded, &r10_bio->state);
1279 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1281 sector_t dev_sector = r10_bio->devs[i].addr;
1285 is_bad = is_badblock(rdev, dev_sector,
1287 &first_bad, &bad_sectors);
1289 /* Mustn't write here until the bad block
1292 atomic_inc(&rdev->nr_pending);
1293 set_bit(BlockedBadBlocks, &rdev->flags);
1294 blocked_rdev = rdev;
1297 if (is_bad && first_bad <= dev_sector) {
1298 /* Cannot write here at all */
1299 bad_sectors -= (dev_sector - first_bad);
1300 if (bad_sectors < max_sectors)
1301 /* Mustn't write more than bad_sectors
1302 * to other devices yet
1304 max_sectors = bad_sectors;
1305 /* We don't set R10BIO_Degraded as that
1306 * only applies if the disk is missing,
1307 * so it might be re-added, and we want to
1308 * know to recover this chunk.
1309 * In this case the device is here, and the
1310 * fact that this chunk is not in-sync is
1311 * recorded in the bad block log.
1316 int good_sectors = first_bad - dev_sector;
1317 if (good_sectors < max_sectors)
1318 max_sectors = good_sectors;
1321 r10_bio->devs[i].bio = bio;
1322 atomic_inc(&rdev->nr_pending);
1324 r10_bio->devs[i].repl_bio = bio;
1325 atomic_inc(&rrdev->nr_pending);
1330 if (unlikely(blocked_rdev)) {
1331 /* Have to wait for this device to get unblocked, then retry */
1335 for (j = 0; j < i; j++) {
1336 if (r10_bio->devs[j].bio) {
1337 d = r10_bio->devs[j].devnum;
1338 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1340 if (r10_bio->devs[j].repl_bio) {
1341 struct md_rdev *rdev;
1342 d = r10_bio->devs[j].devnum;
1343 rdev = conf->mirrors[d].replacement;
1345 /* Race with remove_disk */
1347 rdev = conf->mirrors[d].rdev;
1349 rdev_dec_pending(rdev, mddev);
1352 allow_barrier(conf);
1353 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1358 if (max_sectors < r10_bio->sectors) {
1359 /* We are splitting this into multiple parts, so
1360 * we need to prepare for allocating another r10_bio.
1362 r10_bio->sectors = max_sectors;
1363 spin_lock_irq(&conf->device_lock);
1364 if (bio->bi_phys_segments == 0)
1365 bio->bi_phys_segments = 2;
1367 bio->bi_phys_segments++;
1368 spin_unlock_irq(&conf->device_lock);
1370 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1372 atomic_set(&r10_bio->remaining, 1);
1373 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1375 for (i = 0; i < conf->copies; i++) {
1377 int d = r10_bio->devs[i].devnum;
1378 if (!r10_bio->devs[i].bio)
1381 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1382 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1384 r10_bio->devs[i].bio = mbio;
1386 mbio->bi_sector = (r10_bio->devs[i].addr+
1387 choose_data_offset(r10_bio,
1388 conf->mirrors[d].rdev));
1389 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1390 mbio->bi_end_io = raid10_end_write_request;
1391 mbio->bi_rw = WRITE | do_sync | do_fua;
1392 mbio->bi_private = r10_bio;
1394 atomic_inc(&r10_bio->remaining);
1395 spin_lock_irqsave(&conf->device_lock, flags);
1396 bio_list_add(&conf->pending_bio_list, mbio);
1397 conf->pending_count++;
1398 spin_unlock_irqrestore(&conf->device_lock, flags);
1400 if (!r10_bio->devs[i].repl_bio)
1403 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1404 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1406 r10_bio->devs[i].repl_bio = mbio;
1408 /* We are actively writing to the original device
1409 * so it cannot disappear, so the replacement cannot
1412 mbio->bi_sector = (r10_bio->devs[i].addr +
1415 conf->mirrors[d].replacement));
1416 mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
1417 mbio->bi_end_io = raid10_end_write_request;
1418 mbio->bi_rw = WRITE | do_sync | do_fua;
1419 mbio->bi_private = r10_bio;
1421 atomic_inc(&r10_bio->remaining);
1422 spin_lock_irqsave(&conf->device_lock, flags);
1423 bio_list_add(&conf->pending_bio_list, mbio);
1424 conf->pending_count++;
1425 spin_unlock_irqrestore(&conf->device_lock, flags);
1428 /* Don't remove the bias on 'remaining' (one_write_done) until
1429 * after checking if we need to go around again.
1432 if (sectors_handled < (bio->bi_size >> 9)) {
1433 one_write_done(r10_bio);
1434 /* We need another r10_bio. It has already been counted
1435 * in bio->bi_phys_segments.
1437 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1439 r10_bio->master_bio = bio;
1440 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1442 r10_bio->mddev = mddev;
1443 r10_bio->sector = bio->bi_sector + sectors_handled;
1447 one_write_done(r10_bio);
1449 /* In case raid10d snuck in to freeze_array */
1450 wake_up(&conf->wait_barrier);
1452 if (do_sync || !mddev->bitmap || !plugged)
1453 md_wakeup_thread(mddev->thread);
1456 static void status(struct seq_file *seq, struct mddev *mddev)
1458 struct r10conf *conf = mddev->private;
1461 if (conf->geo.near_copies < conf->geo.raid_disks)
1462 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1463 if (conf->geo.near_copies > 1)
1464 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1465 if (conf->geo.far_copies > 1) {
1466 if (conf->geo.far_offset)
1467 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1469 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1471 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1472 conf->geo.raid_disks - mddev->degraded);
1473 for (i = 0; i < conf->geo.raid_disks; i++)
1474 seq_printf(seq, "%s",
1475 conf->mirrors[i].rdev &&
1476 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1477 seq_printf(seq, "]");
1480 /* check if there are enough drives for
1481 * every block to appear on atleast one.
1482 * Don't consider the device numbered 'ignore'
1483 * as we might be about to remove it.
1485 static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1490 int n = conf->copies;
1493 if (conf->mirrors[first].rdev &&
1496 first = (first+1) % geo->raid_disks;
1500 } while (first != 0);
1504 static int enough(struct r10conf *conf, int ignore)
1506 return _enough(conf, &conf->geo, ignore) &&
1507 _enough(conf, &conf->prev, ignore);
1510 static void error(struct mddev *mddev, struct md_rdev *rdev)
1512 char b[BDEVNAME_SIZE];
1513 struct r10conf *conf = mddev->private;
1516 * If it is not operational, then we have already marked it as dead
1517 * else if it is the last working disks, ignore the error, let the
1518 * next level up know.
1519 * else mark the drive as failed
1521 if (test_bit(In_sync, &rdev->flags)
1522 && !enough(conf, rdev->raid_disk))
1524 * Don't fail the drive, just return an IO error.
1527 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1528 unsigned long flags;
1529 spin_lock_irqsave(&conf->device_lock, flags);
1531 spin_unlock_irqrestore(&conf->device_lock, flags);
1533 * if recovery is running, make sure it aborts.
1535 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1537 set_bit(Blocked, &rdev->flags);
1538 set_bit(Faulty, &rdev->flags);
1539 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1541 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1542 "md/raid10:%s: Operation continuing on %d devices.\n",
1543 mdname(mddev), bdevname(rdev->bdev, b),
1544 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1547 static void print_conf(struct r10conf *conf)
1550 struct mirror_info *tmp;
1552 printk(KERN_DEBUG "RAID10 conf printout:\n");
1554 printk(KERN_DEBUG "(!conf)\n");
1557 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1558 conf->geo.raid_disks);
1560 for (i = 0; i < conf->geo.raid_disks; i++) {
1561 char b[BDEVNAME_SIZE];
1562 tmp = conf->mirrors + i;
1564 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1565 i, !test_bit(In_sync, &tmp->rdev->flags),
1566 !test_bit(Faulty, &tmp->rdev->flags),
1567 bdevname(tmp->rdev->bdev,b));
1571 static void close_sync(struct r10conf *conf)
1574 allow_barrier(conf);
1576 mempool_destroy(conf->r10buf_pool);
1577 conf->r10buf_pool = NULL;
1580 static int raid10_spare_active(struct mddev *mddev)
1583 struct r10conf *conf = mddev->private;
1584 struct mirror_info *tmp;
1586 unsigned long flags;
1589 * Find all non-in_sync disks within the RAID10 configuration
1590 * and mark them in_sync
1592 for (i = 0; i < conf->geo.raid_disks; i++) {
1593 tmp = conf->mirrors + i;
1594 if (tmp->replacement
1595 && tmp->replacement->recovery_offset == MaxSector
1596 && !test_bit(Faulty, &tmp->replacement->flags)
1597 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1598 /* Replacement has just become active */
1600 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1603 /* Replaced device not technically faulty,
1604 * but we need to be sure it gets removed
1605 * and never re-added.
1607 set_bit(Faulty, &tmp->rdev->flags);
1608 sysfs_notify_dirent_safe(
1609 tmp->rdev->sysfs_state);
1611 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1612 } else if (tmp->rdev
1613 && !test_bit(Faulty, &tmp->rdev->flags)
1614 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1616 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1619 spin_lock_irqsave(&conf->device_lock, flags);
1620 mddev->degraded -= count;
1621 spin_unlock_irqrestore(&conf->device_lock, flags);
1628 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1630 struct r10conf *conf = mddev->private;
1634 int last = conf->geo.raid_disks - 1;
1635 struct request_queue *q = bdev_get_queue(rdev->bdev);
1637 if (mddev->recovery_cp < MaxSector)
1638 /* only hot-add to in-sync arrays, as recovery is
1639 * very different from resync
1642 if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1645 if (rdev->raid_disk >= 0)
1646 first = last = rdev->raid_disk;
1648 if (q->merge_bvec_fn) {
1649 set_bit(Unmerged, &rdev->flags);
1650 mddev->merge_check_needed = 1;
1653 if (rdev->saved_raid_disk >= first &&
1654 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1655 mirror = rdev->saved_raid_disk;
1658 for ( ; mirror <= last ; mirror++) {
1659 struct mirror_info *p = &conf->mirrors[mirror];
1660 if (p->recovery_disabled == mddev->recovery_disabled)
1663 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1664 p->replacement != NULL)
1666 clear_bit(In_sync, &rdev->flags);
1667 set_bit(Replacement, &rdev->flags);
1668 rdev->raid_disk = mirror;
1670 disk_stack_limits(mddev->gendisk, rdev->bdev,
1671 rdev->data_offset << 9);
1673 rcu_assign_pointer(p->replacement, rdev);
1677 disk_stack_limits(mddev->gendisk, rdev->bdev,
1678 rdev->data_offset << 9);
1680 p->head_position = 0;
1681 p->recovery_disabled = mddev->recovery_disabled - 1;
1682 rdev->raid_disk = mirror;
1684 if (rdev->saved_raid_disk != mirror)
1686 rcu_assign_pointer(p->rdev, rdev);
1689 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1690 /* Some requests might not have seen this new
1691 * merge_bvec_fn. We must wait for them to complete
1692 * before merging the device fully.
1693 * First we make sure any code which has tested
1694 * our function has submitted the request, then
1695 * we wait for all outstanding requests to complete.
1697 synchronize_sched();
1698 raise_barrier(conf, 0);
1699 lower_barrier(conf);
1700 clear_bit(Unmerged, &rdev->flags);
1702 md_integrity_add_rdev(rdev, mddev);
1707 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1709 struct r10conf *conf = mddev->private;
1711 int number = rdev->raid_disk;
1712 struct md_rdev **rdevp;
1713 struct mirror_info *p = conf->mirrors + number;
1716 if (rdev == p->rdev)
1718 else if (rdev == p->replacement)
1719 rdevp = &p->replacement;
1723 if (test_bit(In_sync, &rdev->flags) ||
1724 atomic_read(&rdev->nr_pending)) {
1728 /* Only remove faulty devices if recovery
1731 if (!test_bit(Faulty, &rdev->flags) &&
1732 mddev->recovery_disabled != p->recovery_disabled &&
1733 (!p->replacement || p->replacement == rdev) &&
1734 number < conf->geo.raid_disks &&
1741 if (atomic_read(&rdev->nr_pending)) {
1742 /* lost the race, try later */
1746 } else if (p->replacement) {
1747 /* We must have just cleared 'rdev' */
1748 p->rdev = p->replacement;
1749 clear_bit(Replacement, &p->replacement->flags);
1750 smp_mb(); /* Make sure other CPUs may see both as identical
1751 * but will never see neither -- if they are careful.
1753 p->replacement = NULL;
1754 clear_bit(WantReplacement, &rdev->flags);
1756 /* We might have just remove the Replacement as faulty
1757 * Clear the flag just in case
1759 clear_bit(WantReplacement, &rdev->flags);
1761 err = md_integrity_register(mddev);
1770 static void end_sync_read(struct bio *bio, int error)
1772 struct r10bio *r10_bio = bio->bi_private;
1773 struct r10conf *conf = r10_bio->mddev->private;
1776 if (bio == r10_bio->master_bio) {
1777 /* this is a reshape read */
1778 d = r10_bio->read_slot; /* really the read dev */
1780 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1782 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1783 set_bit(R10BIO_Uptodate, &r10_bio->state);
1785 /* The write handler will notice the lack of
1786 * R10BIO_Uptodate and record any errors etc
1788 atomic_add(r10_bio->sectors,
1789 &conf->mirrors[d].rdev->corrected_errors);
1791 /* for reconstruct, we always reschedule after a read.
1792 * for resync, only after all reads
1794 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1795 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1796 atomic_dec_and_test(&r10_bio->remaining)) {
1797 /* we have read all the blocks,
1798 * do the comparison in process context in raid10d
1800 reschedule_retry(r10_bio);
1804 static void end_sync_request(struct r10bio *r10_bio)
1806 struct mddev *mddev = r10_bio->mddev;
1808 while (atomic_dec_and_test(&r10_bio->remaining)) {
1809 if (r10_bio->master_bio == NULL) {
1810 /* the primary of several recovery bios */
1811 sector_t s = r10_bio->sectors;
1812 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1813 test_bit(R10BIO_WriteError, &r10_bio->state))
1814 reschedule_retry(r10_bio);
1817 md_done_sync(mddev, s, 1);
1820 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1821 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1822 test_bit(R10BIO_WriteError, &r10_bio->state))
1823 reschedule_retry(r10_bio);
1831 static void end_sync_write(struct bio *bio, int error)
1833 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1834 struct r10bio *r10_bio = bio->bi_private;
1835 struct mddev *mddev = r10_bio->mddev;
1836 struct r10conf *conf = mddev->private;
1842 struct md_rdev *rdev = NULL;
1844 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1846 rdev = conf->mirrors[d].replacement;
1848 rdev = conf->mirrors[d].rdev;
1852 md_error(mddev, rdev);
1854 set_bit(WriteErrorSeen, &rdev->flags);
1855 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1856 set_bit(MD_RECOVERY_NEEDED,
1857 &rdev->mddev->recovery);
1858 set_bit(R10BIO_WriteError, &r10_bio->state);
1860 } else if (is_badblock(rdev,
1861 r10_bio->devs[slot].addr,
1863 &first_bad, &bad_sectors))
1864 set_bit(R10BIO_MadeGood, &r10_bio->state);
1866 rdev_dec_pending(rdev, mddev);
1868 end_sync_request(r10_bio);
1872 * Note: sync and recover and handled very differently for raid10
1873 * This code is for resync.
1874 * For resync, we read through virtual addresses and read all blocks.
1875 * If there is any error, we schedule a write. The lowest numbered
1876 * drive is authoritative.
1877 * However requests come for physical address, so we need to map.
1878 * For every physical address there are raid_disks/copies virtual addresses,
1879 * which is always are least one, but is not necessarly an integer.
1880 * This means that a physical address can span multiple chunks, so we may
1881 * have to submit multiple io requests for a single sync request.
1884 * We check if all blocks are in-sync and only write to blocks that
1887 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1889 struct r10conf *conf = mddev->private;
1891 struct bio *tbio, *fbio;
1894 atomic_set(&r10_bio->remaining, 1);
1896 /* find the first device with a block */
1897 for (i=0; i<conf->copies; i++)
1898 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1901 if (i == conf->copies)
1905 fbio = r10_bio->devs[i].bio;
1907 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1908 /* now find blocks with errors */
1909 for (i=0 ; i < conf->copies ; i++) {
1912 tbio = r10_bio->devs[i].bio;
1914 if (tbio->bi_end_io != end_sync_read)
1918 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1919 /* We know that the bi_io_vec layout is the same for
1920 * both 'first' and 'i', so we just compare them.
1921 * All vec entries are PAGE_SIZE;
1923 for (j = 0; j < vcnt; j++)
1924 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1925 page_address(tbio->bi_io_vec[j].bv_page),
1926 fbio->bi_io_vec[j].bv_len))
1930 mddev->resync_mismatches += r10_bio->sectors;
1931 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1932 /* Don't fix anything. */
1935 /* Ok, we need to write this bio, either to correct an
1936 * inconsistency or to correct an unreadable block.
1937 * First we need to fixup bv_offset, bv_len and
1938 * bi_vecs, as the read request might have corrupted these
1940 tbio->bi_vcnt = vcnt;
1941 tbio->bi_size = r10_bio->sectors << 9;
1943 tbio->bi_phys_segments = 0;
1944 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1945 tbio->bi_flags |= 1 << BIO_UPTODATE;
1946 tbio->bi_next = NULL;
1947 tbio->bi_rw = WRITE;
1948 tbio->bi_private = r10_bio;
1949 tbio->bi_sector = r10_bio->devs[i].addr;
1951 for (j=0; j < vcnt ; j++) {
1952 tbio->bi_io_vec[j].bv_offset = 0;
1953 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1955 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1956 page_address(fbio->bi_io_vec[j].bv_page),
1959 tbio->bi_end_io = end_sync_write;
1961 d = r10_bio->devs[i].devnum;
1962 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1963 atomic_inc(&r10_bio->remaining);
1964 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1966 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1967 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1968 generic_make_request(tbio);
1971 /* Now write out to any replacement devices
1974 for (i = 0; i < conf->copies; i++) {
1977 tbio = r10_bio->devs[i].repl_bio;
1978 if (!tbio || !tbio->bi_end_io)
1980 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
1981 && r10_bio->devs[i].bio != fbio)
1982 for (j = 0; j < vcnt; j++)
1983 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1984 page_address(fbio->bi_io_vec[j].bv_page),
1986 d = r10_bio->devs[i].devnum;
1987 atomic_inc(&r10_bio->remaining);
1988 md_sync_acct(conf->mirrors[d].replacement->bdev,
1989 tbio->bi_size >> 9);
1990 generic_make_request(tbio);
1994 if (atomic_dec_and_test(&r10_bio->remaining)) {
1995 md_done_sync(mddev, r10_bio->sectors, 1);
2001 * Now for the recovery code.
2002 * Recovery happens across physical sectors.
2003 * We recover all non-is_sync drives by finding the virtual address of
2004 * each, and then choose a working drive that also has that virt address.
2005 * There is a separate r10_bio for each non-in_sync drive.
2006 * Only the first two slots are in use. The first for reading,
2007 * The second for writing.
2010 static void fix_recovery_read_error(struct r10bio *r10_bio)
2012 /* We got a read error during recovery.
2013 * We repeat the read in smaller page-sized sections.
2014 * If a read succeeds, write it to the new device or record
2015 * a bad block if we cannot.
2016 * If a read fails, record a bad block on both old and
2019 struct mddev *mddev = r10_bio->mddev;
2020 struct r10conf *conf = mddev->private;
2021 struct bio *bio = r10_bio->devs[0].bio;
2023 int sectors = r10_bio->sectors;
2025 int dr = r10_bio->devs[0].devnum;
2026 int dw = r10_bio->devs[1].devnum;
2030 struct md_rdev *rdev;
2034 if (s > (PAGE_SIZE>>9))
2037 rdev = conf->mirrors[dr].rdev;
2038 addr = r10_bio->devs[0].addr + sect,
2039 ok = sync_page_io(rdev,
2042 bio->bi_io_vec[idx].bv_page,
2045 rdev = conf->mirrors[dw].rdev;
2046 addr = r10_bio->devs[1].addr + sect;
2047 ok = sync_page_io(rdev,
2050 bio->bi_io_vec[idx].bv_page,
2053 set_bit(WriteErrorSeen, &rdev->flags);
2054 if (!test_and_set_bit(WantReplacement,
2056 set_bit(MD_RECOVERY_NEEDED,
2057 &rdev->mddev->recovery);
2061 /* We don't worry if we cannot set a bad block -
2062 * it really is bad so there is no loss in not
2065 rdev_set_badblocks(rdev, addr, s, 0);
2067 if (rdev != conf->mirrors[dw].rdev) {
2068 /* need bad block on destination too */
2069 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2070 addr = r10_bio->devs[1].addr + sect;
2071 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2073 /* just abort the recovery */
2075 "md/raid10:%s: recovery aborted"
2076 " due to read error\n",
2079 conf->mirrors[dw].recovery_disabled
2080 = mddev->recovery_disabled;
2081 set_bit(MD_RECOVERY_INTR,
2094 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2096 struct r10conf *conf = mddev->private;
2098 struct bio *wbio, *wbio2;
2100 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2101 fix_recovery_read_error(r10_bio);
2102 end_sync_request(r10_bio);
2107 * share the pages with the first bio
2108 * and submit the write request
2110 d = r10_bio->devs[1].devnum;
2111 wbio = r10_bio->devs[1].bio;
2112 wbio2 = r10_bio->devs[1].repl_bio;
2113 if (wbio->bi_end_io) {
2114 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2115 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
2116 generic_make_request(wbio);
2118 if (wbio2 && wbio2->bi_end_io) {
2119 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2120 md_sync_acct(conf->mirrors[d].replacement->bdev,
2121 wbio2->bi_size >> 9);
2122 generic_make_request(wbio2);
2128 * Used by fix_read_error() to decay the per rdev read_errors.
2129 * We halve the read error count for every hour that has elapsed
2130 * since the last recorded read error.
2133 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2135 struct timespec cur_time_mon;
2136 unsigned long hours_since_last;
2137 unsigned int read_errors = atomic_read(&rdev->read_errors);
2139 ktime_get_ts(&cur_time_mon);
2141 if (rdev->last_read_error.tv_sec == 0 &&
2142 rdev->last_read_error.tv_nsec == 0) {
2143 /* first time we've seen a read error */
2144 rdev->last_read_error = cur_time_mon;
2148 hours_since_last = (cur_time_mon.tv_sec -
2149 rdev->last_read_error.tv_sec) / 3600;
2151 rdev->last_read_error = cur_time_mon;
2154 * if hours_since_last is > the number of bits in read_errors
2155 * just set read errors to 0. We do this to avoid
2156 * overflowing the shift of read_errors by hours_since_last.
2158 if (hours_since_last >= 8 * sizeof(read_errors))
2159 atomic_set(&rdev->read_errors, 0);
2161 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2164 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2165 int sectors, struct page *page, int rw)
2170 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2171 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2173 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2177 set_bit(WriteErrorSeen, &rdev->flags);
2178 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2179 set_bit(MD_RECOVERY_NEEDED,
2180 &rdev->mddev->recovery);
2182 /* need to record an error - either for the block or the device */
2183 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2184 md_error(rdev->mddev, rdev);
2189 * This is a kernel thread which:
2191 * 1. Retries failed read operations on working mirrors.
2192 * 2. Updates the raid superblock when problems encounter.
2193 * 3. Performs writes following reads for array synchronising.
2196 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2198 int sect = 0; /* Offset from r10_bio->sector */
2199 int sectors = r10_bio->sectors;
2200 struct md_rdev*rdev;
2201 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2202 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2204 /* still own a reference to this rdev, so it cannot
2205 * have been cleared recently.
2207 rdev = conf->mirrors[d].rdev;
2209 if (test_bit(Faulty, &rdev->flags))
2210 /* drive has already been failed, just ignore any
2211 more fix_read_error() attempts */
2214 check_decay_read_errors(mddev, rdev);
2215 atomic_inc(&rdev->read_errors);
2216 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2217 char b[BDEVNAME_SIZE];
2218 bdevname(rdev->bdev, b);
2221 "md/raid10:%s: %s: Raid device exceeded "
2222 "read_error threshold [cur %d:max %d]\n",
2224 atomic_read(&rdev->read_errors), max_read_errors);
2226 "md/raid10:%s: %s: Failing raid device\n",
2228 md_error(mddev, conf->mirrors[d].rdev);
2229 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2235 int sl = r10_bio->read_slot;
2239 if (s > (PAGE_SIZE>>9))
2247 d = r10_bio->devs[sl].devnum;
2248 rdev = rcu_dereference(conf->mirrors[d].rdev);
2250 !test_bit(Unmerged, &rdev->flags) &&
2251 test_bit(In_sync, &rdev->flags) &&
2252 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2253 &first_bad, &bad_sectors) == 0) {
2254 atomic_inc(&rdev->nr_pending);
2256 success = sync_page_io(rdev,
2257 r10_bio->devs[sl].addr +
2260 conf->tmppage, READ, false);
2261 rdev_dec_pending(rdev, mddev);
2267 if (sl == conf->copies)
2269 } while (!success && sl != r10_bio->read_slot);
2273 /* Cannot read from anywhere, just mark the block
2274 * as bad on the first device to discourage future
2277 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2278 rdev = conf->mirrors[dn].rdev;
2280 if (!rdev_set_badblocks(
2282 r10_bio->devs[r10_bio->read_slot].addr
2285 md_error(mddev, rdev);
2286 r10_bio->devs[r10_bio->read_slot].bio
2293 /* write it back and re-read */
2295 while (sl != r10_bio->read_slot) {
2296 char b[BDEVNAME_SIZE];
2301 d = r10_bio->devs[sl].devnum;
2302 rdev = rcu_dereference(conf->mirrors[d].rdev);
2304 test_bit(Unmerged, &rdev->flags) ||
2305 !test_bit(In_sync, &rdev->flags))
2308 atomic_inc(&rdev->nr_pending);
2310 if (r10_sync_page_io(rdev,
2311 r10_bio->devs[sl].addr +
2313 s<<9, conf->tmppage, WRITE)
2315 /* Well, this device is dead */
2317 "md/raid10:%s: read correction "
2319 " (%d sectors at %llu on %s)\n",
2321 (unsigned long long)(
2323 choose_data_offset(r10_bio,
2325 bdevname(rdev->bdev, b));
2326 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2329 bdevname(rdev->bdev, b));
2331 rdev_dec_pending(rdev, mddev);
2335 while (sl != r10_bio->read_slot) {
2336 char b[BDEVNAME_SIZE];
2341 d = r10_bio->devs[sl].devnum;
2342 rdev = rcu_dereference(conf->mirrors[d].rdev);
2344 !test_bit(In_sync, &rdev->flags))
2347 atomic_inc(&rdev->nr_pending);
2349 switch (r10_sync_page_io(rdev,
2350 r10_bio->devs[sl].addr +
2352 s<<9, conf->tmppage,
2355 /* Well, this device is dead */
2357 "md/raid10:%s: unable to read back "
2359 " (%d sectors at %llu on %s)\n",
2361 (unsigned long long)(
2363 choose_data_offset(r10_bio, rdev)),
2364 bdevname(rdev->bdev, b));
2365 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2368 bdevname(rdev->bdev, b));
2372 "md/raid10:%s: read error corrected"
2373 " (%d sectors at %llu on %s)\n",
2375 (unsigned long long)(
2377 choose_data_offset(r10_bio, rdev)),
2378 bdevname(rdev->bdev, b));
2379 atomic_add(s, &rdev->corrected_errors);
2382 rdev_dec_pending(rdev, mddev);
2392 static void bi_complete(struct bio *bio, int error)
2394 complete((struct completion *)bio->bi_private);
2397 static int submit_bio_wait(int rw, struct bio *bio)
2399 struct completion event;
2402 init_completion(&event);
2403 bio->bi_private = &event;
2404 bio->bi_end_io = bi_complete;
2405 submit_bio(rw, bio);
2406 wait_for_completion(&event);
2408 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2411 static int narrow_write_error(struct r10bio *r10_bio, int i)
2413 struct bio *bio = r10_bio->master_bio;
2414 struct mddev *mddev = r10_bio->mddev;
2415 struct r10conf *conf = mddev->private;
2416 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2417 /* bio has the data to be written to slot 'i' where
2418 * we just recently had a write error.
2419 * We repeatedly clone the bio and trim down to one block,
2420 * then try the write. Where the write fails we record
2422 * It is conceivable that the bio doesn't exactly align with
2423 * blocks. We must handle this.
2425 * We currently own a reference to the rdev.
2431 int sect_to_write = r10_bio->sectors;
2434 if (rdev->badblocks.shift < 0)
2437 block_sectors = 1 << rdev->badblocks.shift;
2438 sector = r10_bio->sector;
2439 sectors = ((r10_bio->sector + block_sectors)
2440 & ~(sector_t)(block_sectors - 1))
2443 while (sect_to_write) {
2445 if (sectors > sect_to_write)
2446 sectors = sect_to_write;
2447 /* Write at 'sector' for 'sectors' */
2448 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2449 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2450 wbio->bi_sector = (r10_bio->devs[i].addr+
2451 choose_data_offset(r10_bio, rdev) +
2452 (sector - r10_bio->sector));
2453 wbio->bi_bdev = rdev->bdev;
2454 if (submit_bio_wait(WRITE, wbio) == 0)
2456 ok = rdev_set_badblocks(rdev, sector,
2461 sect_to_write -= sectors;
2463 sectors = block_sectors;
2468 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2470 int slot = r10_bio->read_slot;
2472 struct r10conf *conf = mddev->private;
2473 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2474 char b[BDEVNAME_SIZE];
2475 unsigned long do_sync;
2478 /* we got a read error. Maybe the drive is bad. Maybe just
2479 * the block and we can fix it.
2480 * We freeze all other IO, and try reading the block from
2481 * other devices. When we find one, we re-write
2482 * and check it that fixes the read error.
2483 * This is all done synchronously while the array is
2486 bio = r10_bio->devs[slot].bio;
2487 bdevname(bio->bi_bdev, b);
2489 r10_bio->devs[slot].bio = NULL;
2491 if (mddev->ro == 0) {
2493 fix_read_error(conf, mddev, r10_bio);
2494 unfreeze_array(conf);
2496 r10_bio->devs[slot].bio = IO_BLOCKED;
2498 rdev_dec_pending(rdev, mddev);
2501 rdev = read_balance(conf, r10_bio, &max_sectors);
2503 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2504 " read error for block %llu\n",
2506 (unsigned long long)r10_bio->sector);
2507 raid_end_bio_io(r10_bio);
2511 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2512 slot = r10_bio->read_slot;
2515 "md/raid10:%s: %s: redirecting"
2516 "sector %llu to another mirror\n",
2518 bdevname(rdev->bdev, b),
2519 (unsigned long long)r10_bio->sector);
2520 bio = bio_clone_mddev(r10_bio->master_bio,
2523 r10_bio->sector - bio->bi_sector,
2525 r10_bio->devs[slot].bio = bio;
2526 r10_bio->devs[slot].rdev = rdev;
2527 bio->bi_sector = r10_bio->devs[slot].addr
2528 + choose_data_offset(r10_bio, rdev);
2529 bio->bi_bdev = rdev->bdev;
2530 bio->bi_rw = READ | do_sync;
2531 bio->bi_private = r10_bio;
2532 bio->bi_end_io = raid10_end_read_request;
2533 if (max_sectors < r10_bio->sectors) {
2534 /* Drat - have to split this up more */
2535 struct bio *mbio = r10_bio->master_bio;
2536 int sectors_handled =
2537 r10_bio->sector + max_sectors
2539 r10_bio->sectors = max_sectors;
2540 spin_lock_irq(&conf->device_lock);
2541 if (mbio->bi_phys_segments == 0)
2542 mbio->bi_phys_segments = 2;
2544 mbio->bi_phys_segments++;
2545 spin_unlock_irq(&conf->device_lock);
2546 generic_make_request(bio);
2548 r10_bio = mempool_alloc(conf->r10bio_pool,
2550 r10_bio->master_bio = mbio;
2551 r10_bio->sectors = (mbio->bi_size >> 9)
2554 set_bit(R10BIO_ReadError,
2556 r10_bio->mddev = mddev;
2557 r10_bio->sector = mbio->bi_sector
2562 generic_make_request(bio);
2565 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2567 /* Some sort of write request has finished and it
2568 * succeeded in writing where we thought there was a
2569 * bad block. So forget the bad block.
2570 * Or possibly if failed and we need to record
2574 struct md_rdev *rdev;
2576 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2577 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2578 for (m = 0; m < conf->copies; m++) {
2579 int dev = r10_bio->devs[m].devnum;
2580 rdev = conf->mirrors[dev].rdev;
2581 if (r10_bio->devs[m].bio == NULL)
2583 if (test_bit(BIO_UPTODATE,
2584 &r10_bio->devs[m].bio->bi_flags)) {
2585 rdev_clear_badblocks(
2587 r10_bio->devs[m].addr,
2588 r10_bio->sectors, 0);
2590 if (!rdev_set_badblocks(
2592 r10_bio->devs[m].addr,
2593 r10_bio->sectors, 0))
2594 md_error(conf->mddev, rdev);
2596 rdev = conf->mirrors[dev].replacement;
2597 if (r10_bio->devs[m].repl_bio == NULL)
2599 if (test_bit(BIO_UPTODATE,
2600 &r10_bio->devs[m].repl_bio->bi_flags)) {
2601 rdev_clear_badblocks(
2603 r10_bio->devs[m].addr,
2604 r10_bio->sectors, 0);
2606 if (!rdev_set_badblocks(
2608 r10_bio->devs[m].addr,
2609 r10_bio->sectors, 0))
2610 md_error(conf->mddev, rdev);
2615 for (m = 0; m < conf->copies; m++) {
2616 int dev = r10_bio->devs[m].devnum;
2617 struct bio *bio = r10_bio->devs[m].bio;
2618 rdev = conf->mirrors[dev].rdev;
2619 if (bio == IO_MADE_GOOD) {
2620 rdev_clear_badblocks(
2622 r10_bio->devs[m].addr,
2623 r10_bio->sectors, 0);
2624 rdev_dec_pending(rdev, conf->mddev);
2625 } else if (bio != NULL &&
2626 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2627 if (!narrow_write_error(r10_bio, m)) {
2628 md_error(conf->mddev, rdev);
2629 set_bit(R10BIO_Degraded,
2632 rdev_dec_pending(rdev, conf->mddev);
2634 bio = r10_bio->devs[m].repl_bio;
2635 rdev = conf->mirrors[dev].replacement;
2636 if (rdev && bio == IO_MADE_GOOD) {
2637 rdev_clear_badblocks(
2639 r10_bio->devs[m].addr,
2640 r10_bio->sectors, 0);
2641 rdev_dec_pending(rdev, conf->mddev);
2644 if (test_bit(R10BIO_WriteError,
2646 close_write(r10_bio);
2647 raid_end_bio_io(r10_bio);
2651 static void raid10d(struct mddev *mddev)
2653 struct r10bio *r10_bio;
2654 unsigned long flags;
2655 struct r10conf *conf = mddev->private;
2656 struct list_head *head = &conf->retry_list;
2657 struct blk_plug plug;
2659 md_check_recovery(mddev);
2661 blk_start_plug(&plug);
2664 flush_pending_writes(conf);
2666 spin_lock_irqsave(&conf->device_lock, flags);
2667 if (list_empty(head)) {
2668 spin_unlock_irqrestore(&conf->device_lock, flags);
2671 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2672 list_del(head->prev);
2674 spin_unlock_irqrestore(&conf->device_lock, flags);
2676 mddev = r10_bio->mddev;
2677 conf = mddev->private;
2678 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2679 test_bit(R10BIO_WriteError, &r10_bio->state))
2680 handle_write_completed(conf, r10_bio);
2681 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2682 reshape_request_write(mddev, r10_bio);
2683 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2684 sync_request_write(mddev, r10_bio);
2685 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2686 recovery_request_write(mddev, r10_bio);
2687 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2688 handle_read_error(mddev, r10_bio);
2690 /* just a partial read to be scheduled from a
2693 int slot = r10_bio->read_slot;
2694 generic_make_request(r10_bio->devs[slot].bio);
2698 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2699 md_check_recovery(mddev);
2701 blk_finish_plug(&plug);
2705 static int init_resync(struct r10conf *conf)
2710 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2711 BUG_ON(conf->r10buf_pool);
2712 conf->have_replacement = 0;
2713 for (i = 0; i < conf->geo.raid_disks; i++)
2714 if (conf->mirrors[i].replacement)
2715 conf->have_replacement = 1;
2716 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2717 if (!conf->r10buf_pool)
2719 conf->next_resync = 0;
2724 * perform a "sync" on one "block"
2726 * We need to make sure that no normal I/O request - particularly write
2727 * requests - conflict with active sync requests.
2729 * This is achieved by tracking pending requests and a 'barrier' concept
2730 * that can be installed to exclude normal IO requests.
2732 * Resync and recovery are handled very differently.
2733 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2735 * For resync, we iterate over virtual addresses, read all copies,
2736 * and update if there are differences. If only one copy is live,
2738 * For recovery, we iterate over physical addresses, read a good
2739 * value for each non-in_sync drive, and over-write.
2741 * So, for recovery we may have several outstanding complex requests for a
2742 * given address, one for each out-of-sync device. We model this by allocating
2743 * a number of r10_bio structures, one for each out-of-sync device.
2744 * As we setup these structures, we collect all bio's together into a list
2745 * which we then process collectively to add pages, and then process again
2746 * to pass to generic_make_request.
2748 * The r10_bio structures are linked using a borrowed master_bio pointer.
2749 * This link is counted in ->remaining. When the r10_bio that points to NULL
2750 * has its remaining count decremented to 0, the whole complex operation
2755 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2756 int *skipped, int go_faster)
2758 struct r10conf *conf = mddev->private;
2759 struct r10bio *r10_bio;
2760 struct bio *biolist = NULL, *bio;
2761 sector_t max_sector, nr_sectors;
2764 sector_t sync_blocks;
2765 sector_t sectors_skipped = 0;
2766 int chunks_skipped = 0;
2767 sector_t chunk_mask = conf->geo.chunk_mask;
2769 if (!conf->r10buf_pool)
2770 if (init_resync(conf))
2774 max_sector = mddev->dev_sectors;
2775 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2776 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2777 max_sector = mddev->resync_max_sectors;
2778 if (sector_nr >= max_sector) {
2779 /* If we aborted, we need to abort the
2780 * sync on the 'current' bitmap chucks (there can
2781 * be several when recovering multiple devices).
2782 * as we may have started syncing it but not finished.
2783 * We can find the current address in
2784 * mddev->curr_resync, but for recovery,
2785 * we need to convert that to several
2786 * virtual addresses.
2788 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2793 if (mddev->curr_resync < max_sector) { /* aborted */
2794 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2795 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2797 else for (i = 0; i < conf->geo.raid_disks; i++) {
2799 raid10_find_virt(conf, mddev->curr_resync, i);
2800 bitmap_end_sync(mddev->bitmap, sect,
2804 /* completed sync */
2805 if ((!mddev->bitmap || conf->fullsync)
2806 && conf->have_replacement
2807 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2808 /* Completed a full sync so the replacements
2809 * are now fully recovered.
2811 for (i = 0; i < conf->geo.raid_disks; i++)
2812 if (conf->mirrors[i].replacement)
2813 conf->mirrors[i].replacement
2819 bitmap_close_sync(mddev->bitmap);
2822 return sectors_skipped;
2825 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2826 return reshape_request(mddev, sector_nr, skipped);
2828 if (chunks_skipped >= conf->geo.raid_disks) {
2829 /* if there has been nothing to do on any drive,
2830 * then there is nothing to do at all..
2833 return (max_sector - sector_nr) + sectors_skipped;
2836 if (max_sector > mddev->resync_max)
2837 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2839 /* make sure whole request will fit in a chunk - if chunks
2842 if (conf->geo.near_copies < conf->geo.raid_disks &&
2843 max_sector > (sector_nr | chunk_mask))
2844 max_sector = (sector_nr | chunk_mask) + 1;
2846 * If there is non-resync activity waiting for us then
2847 * put in a delay to throttle resync.
2849 if (!go_faster && conf->nr_waiting)
2850 msleep_interruptible(1000);
2852 /* Again, very different code for resync and recovery.
2853 * Both must result in an r10bio with a list of bios that
2854 * have bi_end_io, bi_sector, bi_bdev set,
2855 * and bi_private set to the r10bio.
2856 * For recovery, we may actually create several r10bios
2857 * with 2 bios in each, that correspond to the bios in the main one.
2858 * In this case, the subordinate r10bios link back through a
2859 * borrowed master_bio pointer, and the counter in the master
2860 * includes a ref from each subordinate.
2862 /* First, we decide what to do and set ->bi_end_io
2863 * To end_sync_read if we want to read, and
2864 * end_sync_write if we will want to write.
2867 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2868 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2869 /* recovery... the complicated one */
2873 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2879 struct mirror_info *mirror = &conf->mirrors[i];
2881 if ((mirror->rdev == NULL ||
2882 test_bit(In_sync, &mirror->rdev->flags))
2884 (mirror->replacement == NULL ||
2886 &mirror->replacement->flags)))
2890 /* want to reconstruct this device */
2892 sect = raid10_find_virt(conf, sector_nr, i);
2893 if (sect >= mddev->resync_max_sectors) {
2894 /* last stripe is not complete - don't
2895 * try to recover this sector.
2899 /* Unless we are doing a full sync, or a replacement
2900 * we only need to recover the block if it is set in
2903 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2905 if (sync_blocks < max_sync)
2906 max_sync = sync_blocks;
2908 mirror->replacement == NULL &&
2910 /* yep, skip the sync_blocks here, but don't assume
2911 * that there will never be anything to do here
2913 chunks_skipped = -1;
2917 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2918 raise_barrier(conf, rb2 != NULL);
2919 atomic_set(&r10_bio->remaining, 0);
2921 r10_bio->master_bio = (struct bio*)rb2;
2923 atomic_inc(&rb2->remaining);
2924 r10_bio->mddev = mddev;
2925 set_bit(R10BIO_IsRecover, &r10_bio->state);
2926 r10_bio->sector = sect;
2928 raid10_find_phys(conf, r10_bio);
2930 /* Need to check if the array will still be
2933 for (j = 0; j < conf->geo.raid_disks; j++)
2934 if (conf->mirrors[j].rdev == NULL ||
2935 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2940 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2941 &sync_blocks, still_degraded);
2944 for (j=0; j<conf->copies;j++) {
2946 int d = r10_bio->devs[j].devnum;
2947 sector_t from_addr, to_addr;
2948 struct md_rdev *rdev;
2949 sector_t sector, first_bad;
2951 if (!conf->mirrors[d].rdev ||
2952 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2954 /* This is where we read from */
2956 rdev = conf->mirrors[d].rdev;
2957 sector = r10_bio->devs[j].addr;
2959 if (is_badblock(rdev, sector, max_sync,
2960 &first_bad, &bad_sectors)) {
2961 if (first_bad > sector)
2962 max_sync = first_bad - sector;
2964 bad_sectors -= (sector
2966 if (max_sync > bad_sectors)
2967 max_sync = bad_sectors;
2971 bio = r10_bio->devs[0].bio;
2972 bio->bi_next = biolist;
2974 bio->bi_private = r10_bio;
2975 bio->bi_end_io = end_sync_read;
2977 from_addr = r10_bio->devs[j].addr;
2978 bio->bi_sector = from_addr + rdev->data_offset;
2979 bio->bi_bdev = rdev->bdev;
2980 atomic_inc(&rdev->nr_pending);
2981 /* and we write to 'i' (if not in_sync) */
2983 for (k=0; k<conf->copies; k++)
2984 if (r10_bio->devs[k].devnum == i)
2986 BUG_ON(k == conf->copies);
2987 to_addr = r10_bio->devs[k].addr;
2988 r10_bio->devs[0].devnum = d;
2989 r10_bio->devs[0].addr = from_addr;
2990 r10_bio->devs[1].devnum = i;
2991 r10_bio->devs[1].addr = to_addr;
2993 rdev = mirror->rdev;
2994 if (!test_bit(In_sync, &rdev->flags)) {
2995 bio = r10_bio->devs[1].bio;
2996 bio->bi_next = biolist;
2998 bio->bi_private = r10_bio;
2999 bio->bi_end_io = end_sync_write;
3001 bio->bi_sector = to_addr
3002 + rdev->data_offset;
3003 bio->bi_bdev = rdev->bdev;
3004 atomic_inc(&r10_bio->remaining);
3006 r10_bio->devs[1].bio->bi_end_io = NULL;
3008 /* and maybe write to replacement */
3009 bio = r10_bio->devs[1].repl_bio;
3011 bio->bi_end_io = NULL;
3012 rdev = mirror->replacement;
3013 /* Note: if rdev != NULL, then bio
3014 * cannot be NULL as r10buf_pool_alloc will
3015 * have allocated it.
3016 * So the second test here is pointless.
3017 * But it keeps semantic-checkers happy, and
3018 * this comment keeps human reviewers
3021 if (rdev == NULL || bio == NULL ||
3022 test_bit(Faulty, &rdev->flags))
3024 bio->bi_next = biolist;
3026 bio->bi_private = r10_bio;
3027 bio->bi_end_io = end_sync_write;
3029 bio->bi_sector = to_addr + rdev->data_offset;
3030 bio->bi_bdev = rdev->bdev;
3031 atomic_inc(&r10_bio->remaining);
3034 if (j == conf->copies) {
3035 /* Cannot recover, so abort the recovery or
3036 * record a bad block */
3039 atomic_dec(&rb2->remaining);
3042 /* problem is that there are bad blocks
3043 * on other device(s)
3046 for (k = 0; k < conf->copies; k++)
3047 if (r10_bio->devs[k].devnum == i)
3049 if (!test_bit(In_sync,
3050 &mirror->rdev->flags)
3051 && !rdev_set_badblocks(
3053 r10_bio->devs[k].addr,
3056 if (mirror->replacement &&
3057 !rdev_set_badblocks(
3058 mirror->replacement,
3059 r10_bio->devs[k].addr,
3064 if (!test_and_set_bit(MD_RECOVERY_INTR,
3066 printk(KERN_INFO "md/raid10:%s: insufficient "
3067 "working devices for recovery.\n",
3069 mirror->recovery_disabled
3070 = mddev->recovery_disabled;
3075 if (biolist == NULL) {
3077 struct r10bio *rb2 = r10_bio;
3078 r10_bio = (struct r10bio*) rb2->master_bio;
3079 rb2->master_bio = NULL;
3085 /* resync. Schedule a read for every block at this virt offset */
3088 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3090 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3091 &sync_blocks, mddev->degraded) &&
3092 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3093 &mddev->recovery)) {
3094 /* We can skip this block */
3096 return sync_blocks + sectors_skipped;
3098 if (sync_blocks < max_sync)
3099 max_sync = sync_blocks;
3100 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3102 r10_bio->mddev = mddev;
3103 atomic_set(&r10_bio->remaining, 0);
3104 raise_barrier(conf, 0);
3105 conf->next_resync = sector_nr;
3107 r10_bio->master_bio = NULL;
3108 r10_bio->sector = sector_nr;
3109 set_bit(R10BIO_IsSync, &r10_bio->state);
3110 raid10_find_phys(conf, r10_bio);
3111 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3113 for (i = 0; i < conf->copies; i++) {
3114 int d = r10_bio->devs[i].devnum;
3115 sector_t first_bad, sector;
3118 if (r10_bio->devs[i].repl_bio)
3119 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3121 bio = r10_bio->devs[i].bio;
3122 bio->bi_end_io = NULL;
3123 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3124 if (conf->mirrors[d].rdev == NULL ||
3125 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3127 sector = r10_bio->devs[i].addr;
3128 if (is_badblock(conf->mirrors[d].rdev,
3130 &first_bad, &bad_sectors)) {
3131 if (first_bad > sector)
3132 max_sync = first_bad - sector;
3134 bad_sectors -= (sector - first_bad);
3135 if (max_sync > bad_sectors)
3136 max_sync = max_sync;
3140 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3141 atomic_inc(&r10_bio->remaining);
3142 bio->bi_next = biolist;
3144 bio->bi_private = r10_bio;
3145 bio->bi_end_io = end_sync_read;
3147 bio->bi_sector = sector +
3148 conf->mirrors[d].rdev->data_offset;
3149 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3152 if (conf->mirrors[d].replacement == NULL ||
3154 &conf->mirrors[d].replacement->flags))
3157 /* Need to set up for writing to the replacement */
3158 bio = r10_bio->devs[i].repl_bio;
3159 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3161 sector = r10_bio->devs[i].addr;
3162 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3163 bio->bi_next = biolist;
3165 bio->bi_private = r10_bio;
3166 bio->bi_end_io = end_sync_write;
3168 bio->bi_sector = sector +
3169 conf->mirrors[d].replacement->data_offset;
3170 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3175 for (i=0; i<conf->copies; i++) {
3176 int d = r10_bio->devs[i].devnum;
3177 if (r10_bio->devs[i].bio->bi_end_io)
3178 rdev_dec_pending(conf->mirrors[d].rdev,
3180 if (r10_bio->devs[i].repl_bio &&
3181 r10_bio->devs[i].repl_bio->bi_end_io)
3183 conf->mirrors[d].replacement,
3192 for (bio = biolist; bio ; bio=bio->bi_next) {
3194 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3196 bio->bi_flags |= 1 << BIO_UPTODATE;
3199 bio->bi_phys_segments = 0;
3204 if (sector_nr + max_sync < max_sector)
3205 max_sector = sector_nr + max_sync;
3208 int len = PAGE_SIZE;
3209 if (sector_nr + (len>>9) > max_sector)
3210 len = (max_sector - sector_nr) << 9;
3213 for (bio= biolist ; bio ; bio=bio->bi_next) {
3215 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3216 if (bio_add_page(bio, page, len, 0))
3220 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3221 for (bio2 = biolist;
3222 bio2 && bio2 != bio;
3223 bio2 = bio2->bi_next) {
3224 /* remove last page from this bio */
3226 bio2->bi_size -= len;
3227 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3231 nr_sectors += len>>9;
3232 sector_nr += len>>9;
3233 } while (biolist->bi_vcnt < RESYNC_PAGES);
3235 r10_bio->sectors = nr_sectors;
3239 biolist = biolist->bi_next;
3241 bio->bi_next = NULL;
3242 r10_bio = bio->bi_private;
3243 r10_bio->sectors = nr_sectors;
3245 if (bio->bi_end_io == end_sync_read) {
3246 md_sync_acct(bio->bi_bdev, nr_sectors);
3247 generic_make_request(bio);
3251 if (sectors_skipped)
3252 /* pretend they weren't skipped, it makes
3253 * no important difference in this case
3255 md_done_sync(mddev, sectors_skipped, 1);
3257 return sectors_skipped + nr_sectors;
3259 /* There is nowhere to write, so all non-sync
3260 * drives must be failed or in resync, all drives
3261 * have a bad block, so try the next chunk...
3263 if (sector_nr + max_sync < max_sector)
3264 max_sector = sector_nr + max_sync;
3266 sectors_skipped += (max_sector - sector_nr);
3268 sector_nr = max_sector;
3273 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3276 struct r10conf *conf = mddev->private;
3279 raid_disks = min(conf->geo.raid_disks,
3280 conf->prev.raid_disks);
3282 sectors = conf->dev_sectors;
3284 size = sectors >> conf->geo.chunk_shift;
3285 sector_div(size, conf->geo.far_copies);
3286 size = size * raid_disks;
3287 sector_div(size, conf->geo.near_copies);
3289 return size << conf->geo.chunk_shift;
3292 static void calc_sectors(struct r10conf *conf, sector_t size)
3294 /* Calculate the number of sectors-per-device that will
3295 * actually be used, and set conf->dev_sectors and
3299 size = size >> conf->geo.chunk_shift;
3300 sector_div(size, conf->geo.far_copies);
3301 size = size * conf->geo.raid_disks;
3302 sector_div(size, conf->geo.near_copies);
3303 /* 'size' is now the number of chunks in the array */
3304 /* calculate "used chunks per device" */
3305 size = size * conf->copies;
3307 /* We need to round up when dividing by raid_disks to
3308 * get the stride size.
3310 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3312 conf->dev_sectors = size << conf->geo.chunk_shift;
3314 if (conf->geo.far_offset)
3315 conf->geo.stride = 1 << conf->geo.chunk_shift;
3317 sector_div(size, conf->geo.far_copies);
3318 conf->geo.stride = size << conf->geo.chunk_shift;
3322 enum geo_type {geo_new, geo_old, geo_start};
3323 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3326 int layout, chunk, disks;
3329 layout = mddev->layout;
3330 chunk = mddev->chunk_sectors;
3331 disks = mddev->raid_disks - mddev->delta_disks;
3334 layout = mddev->new_layout;
3335 chunk = mddev->new_chunk_sectors;
3336 disks = mddev->raid_disks;
3338 default: /* avoid 'may be unused' warnings */
3339 case geo_start: /* new when starting reshape - raid_disks not
3341 layout = mddev->new_layout;
3342 chunk = mddev->new_chunk_sectors;
3343 disks = mddev->raid_disks + mddev->delta_disks;
3348 if (chunk < (PAGE_SIZE >> 9) ||
3349 !is_power_of_2(chunk))
3352 fc = (layout >> 8) & 255;
3353 fo = layout & (1<<16);
3354 geo->raid_disks = disks;
3355 geo->near_copies = nc;
3356 geo->far_copies = fc;
3357 geo->far_offset = fo;
3358 geo->chunk_mask = chunk - 1;
3359 geo->chunk_shift = ffz(~chunk);
3363 static struct r10conf *setup_conf(struct mddev *mddev)
3365 struct r10conf *conf = NULL;
3370 copies = setup_geo(&geo, mddev, geo_new);
3373 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3374 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3375 mdname(mddev), PAGE_SIZE);
3379 if (copies < 2 || copies > mddev->raid_disks) {
3380 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3381 mdname(mddev), mddev->new_layout);
3386 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3390 /* FIXME calc properly */
3391 conf->mirrors = kzalloc(sizeof(struct mirror_info)*(mddev->raid_disks +
3392 max(0,mddev->delta_disks)),
3397 conf->tmppage = alloc_page(GFP_KERNEL);
3402 conf->copies = copies;
3403 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3404 r10bio_pool_free, conf);
3405 if (!conf->r10bio_pool)
3408 calc_sectors(conf, mddev->dev_sectors);
3409 if (mddev->reshape_position == MaxSector) {
3410 conf->prev = conf->geo;
3411 conf->reshape_progress = MaxSector;
3413 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3417 conf->reshape_progress = mddev->reshape_position;
3418 if (conf->prev.far_offset)
3419 conf->prev.stride = 1 << conf->prev.chunk_shift;
3421 /* far_copies must be 1 */
3422 conf->prev.stride = conf->dev_sectors;
3424 spin_lock_init(&conf->device_lock);
3425 INIT_LIST_HEAD(&conf->retry_list);
3427 spin_lock_init(&conf->resync_lock);
3428 init_waitqueue_head(&conf->wait_barrier);
3430 conf->thread = md_register_thread(raid10d, mddev, NULL);
3434 conf->mddev = mddev;
3439 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3442 if (conf->r10bio_pool)
3443 mempool_destroy(conf->r10bio_pool);
3444 kfree(conf->mirrors);
3445 safe_put_page(conf->tmppage);
3448 return ERR_PTR(err);
3451 static int run(struct mddev *mddev)
3453 struct r10conf *conf;
3454 int i, disk_idx, chunk_size;
3455 struct mirror_info *disk;
3456 struct md_rdev *rdev;
3458 sector_t min_offset_diff = 0;
3461 if (mddev->private == NULL) {
3462 conf = setup_conf(mddev);
3464 return PTR_ERR(conf);
3465 mddev->private = conf;
3467 conf = mddev->private;
3471 mddev->thread = conf->thread;
3472 conf->thread = NULL;
3474 chunk_size = mddev->chunk_sectors << 9;
3475 blk_queue_io_min(mddev->queue, chunk_size);
3476 if (conf->geo.raid_disks % conf->geo.near_copies)
3477 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3479 blk_queue_io_opt(mddev->queue, chunk_size *
3480 (conf->geo.raid_disks / conf->geo.near_copies));
3482 rdev_for_each(rdev, mddev) {
3484 struct request_queue *q;
3486 disk_idx = rdev->raid_disk;
3489 if (disk_idx >= conf->geo.raid_disks &&
3490 disk_idx >= conf->prev.raid_disks)
3492 disk = conf->mirrors + disk_idx;
3494 if (test_bit(Replacement, &rdev->flags)) {
3495 if (disk->replacement)
3497 disk->replacement = rdev;
3503 q = bdev_get_queue(rdev->bdev);
3504 if (q->merge_bvec_fn)
3505 mddev->merge_check_needed = 1;
3506 diff = (rdev->new_data_offset - rdev->data_offset);
3507 if (!mddev->reshape_backwards)
3511 if (first || diff < min_offset_diff)
3512 min_offset_diff = diff;
3514 disk_stack_limits(mddev->gendisk, rdev->bdev,
3515 rdev->data_offset << 9);
3517 disk->head_position = 0;
3520 /* need to check that every block has at least one working mirror */
3521 if (!enough(conf, -1)) {
3522 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3527 if (conf->reshape_progress != MaxSector) {
3528 /* must ensure that shape change is supported */
3529 if (conf->geo.far_copies != 1 &&
3530 conf->geo.far_offset == 0)
3532 if (conf->prev.far_copies != 1 &&
3533 conf->geo.far_offset == 0)
3537 mddev->degraded = 0;
3539 i < conf->geo.raid_disks
3540 || i < conf->prev.raid_disks;
3543 disk = conf->mirrors + i;
3545 if (!disk->rdev && disk->replacement) {
3546 /* The replacement is all we have - use it */
3547 disk->rdev = disk->replacement;
3548 disk->replacement = NULL;
3549 clear_bit(Replacement, &disk->rdev->flags);
3553 !test_bit(In_sync, &disk->rdev->flags)) {
3554 disk->head_position = 0;
3559 disk->recovery_disabled = mddev->recovery_disabled - 1;
3562 if (mddev->recovery_cp != MaxSector)
3563 printk(KERN_NOTICE "md/raid10:%s: not clean"
3564 " -- starting background reconstruction\n",
3567 "md/raid10:%s: active with %d out of %d devices\n",
3568 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3569 conf->geo.raid_disks);
3571 * Ok, everything is just fine now
3573 mddev->dev_sectors = conf->dev_sectors;
3574 size = raid10_size(mddev, 0, 0);
3575 md_set_array_sectors(mddev, size);
3576 mddev->resync_max_sectors = size;
3578 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3579 mddev->queue->backing_dev_info.congested_data = mddev;
3581 /* Calculate max read-ahead size.
3582 * We need to readahead at least twice a whole stripe....
3586 int stripe = conf->geo.raid_disks *
3587 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3588 stripe /= conf->geo.near_copies;
3589 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3590 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3593 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3595 if (md_integrity_register(mddev))
3598 if (conf->reshape_progress != MaxSector) {
3599 unsigned long before_length, after_length;
3601 before_length = ((1 << conf->prev.chunk_shift) *
3602 conf->prev.far_copies);
3603 after_length = ((1 << conf->geo.chunk_shift) *
3604 conf->geo.far_copies);
3606 if (max(before_length, after_length) > min_offset_diff) {
3607 /* This cannot work */
3608 printk("md/raid10: offset difference not enough to continue reshape\n");
3611 conf->offset_diff = min_offset_diff;
3613 conf->reshape_safe = conf->reshape_progress;
3614 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3615 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3616 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3617 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3618 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3625 md_unregister_thread(&mddev->thread);
3626 if (conf->r10bio_pool)
3627 mempool_destroy(conf->r10bio_pool);
3628 safe_put_page(conf->tmppage);
3629 kfree(conf->mirrors);
3631 mddev->private = NULL;
3636 static int stop(struct mddev *mddev)
3638 struct r10conf *conf = mddev->private;
3640 raise_barrier(conf, 0);
3641 lower_barrier(conf);
3643 md_unregister_thread(&mddev->thread);
3644 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3645 if (conf->r10bio_pool)
3646 mempool_destroy(conf->r10bio_pool);
3647 kfree(conf->mirrors);
3649 mddev->private = NULL;
3653 static void raid10_quiesce(struct mddev *mddev, int state)
3655 struct r10conf *conf = mddev->private;
3659 raise_barrier(conf, 0);
3662 lower_barrier(conf);
3667 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3669 /* Resize of 'far' arrays is not supported.
3670 * For 'near' and 'offset' arrays we can set the
3671 * number of sectors used to be an appropriate multiple
3672 * of the chunk size.
3673 * For 'offset', this is far_copies*chunksize.
3674 * For 'near' the multiplier is the LCM of
3675 * near_copies and raid_disks.
3676 * So if far_copies > 1 && !far_offset, fail.
3677 * Else find LCM(raid_disks, near_copy)*far_copies and
3678 * multiply by chunk_size. Then round to this number.
3679 * This is mostly done by raid10_size()
3681 struct r10conf *conf = mddev->private;
3682 sector_t oldsize, size;
3684 if (mddev->reshape_position != MaxSector)
3687 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3690 oldsize = raid10_size(mddev, 0, 0);
3691 size = raid10_size(mddev, sectors, 0);
3692 if (mddev->external_size &&
3693 mddev->array_sectors > size)
3695 if (mddev->bitmap) {
3696 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3700 md_set_array_sectors(mddev, size);
3701 set_capacity(mddev->gendisk, mddev->array_sectors);
3702 revalidate_disk(mddev->gendisk);
3703 if (sectors > mddev->dev_sectors &&
3704 mddev->recovery_cp > oldsize) {
3705 mddev->recovery_cp = oldsize;
3706 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3708 calc_sectors(conf, sectors);
3709 mddev->dev_sectors = conf->dev_sectors;
3710 mddev->resync_max_sectors = size;
3714 static void *raid10_takeover_raid0(struct mddev *mddev)
3716 struct md_rdev *rdev;
3717 struct r10conf *conf;
3719 if (mddev->degraded > 0) {
3720 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3722 return ERR_PTR(-EINVAL);
3725 /* Set new parameters */
3726 mddev->new_level = 10;
3727 /* new layout: far_copies = 1, near_copies = 2 */
3728 mddev->new_layout = (1<<8) + 2;
3729 mddev->new_chunk_sectors = mddev->chunk_sectors;
3730 mddev->delta_disks = mddev->raid_disks;
3731 mddev->raid_disks *= 2;
3732 /* make sure it will be not marked as dirty */
3733 mddev->recovery_cp = MaxSector;
3735 conf = setup_conf(mddev);
3736 if (!IS_ERR(conf)) {
3737 rdev_for_each(rdev, mddev)
3738 if (rdev->raid_disk >= 0)
3739 rdev->new_raid_disk = rdev->raid_disk * 2;
3746 static void *raid10_takeover(struct mddev *mddev)
3748 struct r0conf *raid0_conf;
3750 /* raid10 can take over:
3751 * raid0 - providing it has only two drives
3753 if (mddev->level == 0) {
3754 /* for raid0 takeover only one zone is supported */
3755 raid0_conf = mddev->private;
3756 if (raid0_conf->nr_strip_zones > 1) {
3757 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3758 " with more than one zone.\n",
3760 return ERR_PTR(-EINVAL);
3762 return raid10_takeover_raid0(mddev);
3764 return ERR_PTR(-EINVAL);
3767 static int raid10_check_reshape(struct mddev *mddev)
3769 /* Called when there is a request to change
3770 * - layout (to ->new_layout)
3771 * - chunk size (to ->new_chunk_sectors)
3772 * - raid_disks (by delta_disks)
3773 * or when trying to restart a reshape that was ongoing.
3775 * We need to validate the request and possibly allocate
3776 * space if that might be an issue later.
3778 * Currently we reject any reshape of a 'far' mode array,
3779 * allow chunk size to change if new is generally acceptable,
3780 * allow raid_disks to increase, and allow
3781 * a switch between 'near' mode and 'offset' mode.
3783 struct r10conf *conf = mddev->private;
3786 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3789 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3790 /* mustn't change number of copies */
3792 if (geo.far_copies > 1 && !geo.far_offset)
3793 /* Cannot switch to 'far' mode */
3796 if (mddev->array_sectors & geo.chunk_mask)
3797 /* not factor of array size */
3800 if (!enough(conf, -1))
3803 kfree(conf->mirrors_new);
3804 conf->mirrors_new = NULL;
3805 if (mddev->delta_disks > 0) {
3806 /* allocate new 'mirrors' list */
3807 conf->mirrors_new = kzalloc(
3808 sizeof(struct mirror_info)
3809 *(mddev->raid_disks +
3810 mddev->delta_disks),
3812 if (!conf->mirrors_new)
3819 * Need to check if array has failed when deciding whether to:
3821 * - remove non-faulty devices
3824 * This determination is simple when no reshape is happening.
3825 * However if there is a reshape, we need to carefully check
3826 * both the before and after sections.
3827 * This is because some failed devices may only affect one
3828 * of the two sections, and some non-in_sync devices may
3829 * be insync in the section most affected by failed devices.
3831 static int calc_degraded(struct r10conf *conf)
3833 int degraded, degraded2;
3838 /* 'prev' section first */
3839 for (i = 0; i < conf->prev.raid_disks; i++) {
3840 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3841 if (!rdev || test_bit(Faulty, &rdev->flags))
3843 else if (!test_bit(In_sync, &rdev->flags))
3844 /* When we can reduce the number of devices in
3845 * an array, this might not contribute to
3846 * 'degraded'. It does now.
3851 if (conf->geo.raid_disks == conf->prev.raid_disks)
3855 for (i = 0; i < conf->geo.raid_disks; i++) {
3856 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3857 if (!rdev || test_bit(Faulty, &rdev->flags))
3859 else if (!test_bit(In_sync, &rdev->flags)) {
3860 /* If reshape is increasing the number of devices,
3861 * this section has already been recovered, so
3862 * it doesn't contribute to degraded.
3865 if (conf->geo.raid_disks <= conf->prev.raid_disks)
3870 if (degraded2 > degraded)
3875 static int raid10_start_reshape(struct mddev *mddev)
3877 /* A 'reshape' has been requested. This commits
3878 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3879 * This also checks if there are enough spares and adds them
3881 * We currently require enough spares to make the final
3882 * array non-degraded. We also require that the difference
3883 * between old and new data_offset - on each device - is
3884 * enough that we never risk over-writing.
3887 unsigned long before_length, after_length;
3888 sector_t min_offset_diff = 0;
3891 struct r10conf *conf = mddev->private;
3892 struct md_rdev *rdev;
3896 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3899 if (setup_geo(&new, mddev, geo_start) != conf->copies)
3902 before_length = ((1 << conf->prev.chunk_shift) *
3903 conf->prev.far_copies);
3904 after_length = ((1 << conf->geo.chunk_shift) *
3905 conf->geo.far_copies);
3907 rdev_for_each(rdev, mddev) {
3908 if (!test_bit(In_sync, &rdev->flags)
3909 && !test_bit(Faulty, &rdev->flags))
3911 if (rdev->raid_disk >= 0) {
3912 long long diff = (rdev->new_data_offset
3913 - rdev->data_offset);
3914 if (!mddev->reshape_backwards)
3918 if (first || diff < min_offset_diff)
3919 min_offset_diff = diff;
3923 if (max(before_length, after_length) > min_offset_diff)
3926 if (spares < mddev->delta_disks)
3929 conf->offset_diff = min_offset_diff;
3930 spin_lock_irq(&conf->device_lock);
3931 if (conf->mirrors_new) {
3932 memcpy(conf->mirrors_new, conf->mirrors,
3933 sizeof(struct mirror_info)*conf->prev.raid_disks);
3935 kfree(conf->mirrors_old); /* FIXME and elsewhere */
3936 conf->mirrors_old = conf->mirrors;
3937 conf->mirrors = conf->mirrors_new;
3938 conf->mirrors_new = NULL;
3940 setup_geo(&conf->geo, mddev, geo_start);
3942 if (mddev->reshape_backwards) {
3943 sector_t size = raid10_size(mddev, 0, 0);
3944 if (size < mddev->array_sectors) {
3945 spin_unlock_irq(&conf->device_lock);
3946 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
3950 mddev->resync_max_sectors = size;
3951 conf->reshape_progress = size;
3953 conf->reshape_progress = 0;
3954 spin_unlock_irq(&conf->device_lock);
3956 if (mddev->delta_disks && mddev->bitmap) {
3957 ret = bitmap_resize(mddev->bitmap,
3958 raid10_size(mddev, 0,
3959 conf->geo.raid_disks),
3964 if (mddev->delta_disks > 0) {
3965 rdev_for_each(rdev, mddev)
3966 if (rdev->raid_disk < 0 &&
3967 !test_bit(Faulty, &rdev->flags)) {
3968 if (raid10_add_disk(mddev, rdev) == 0) {
3969 if (rdev->raid_disk >=
3970 conf->prev.raid_disks)
3971 set_bit(In_sync, &rdev->flags);
3973 rdev->recovery_offset = 0;
3975 if (sysfs_link_rdev(mddev, rdev))
3976 /* Failure here is OK */;
3978 } else if (rdev->raid_disk >= conf->prev.raid_disks
3979 && !test_bit(Faulty, &rdev->flags)) {
3980 /* This is a spare that was manually added */
3981 set_bit(In_sync, &rdev->flags);
3984 /* When a reshape changes the number of devices,
3985 * ->degraded is measured against the larger of the
3986 * pre and post numbers.
3988 spin_lock_irq(&conf->device_lock);
3989 mddev->degraded = calc_degraded(conf);
3990 spin_unlock_irq(&conf->device_lock);
3991 mddev->raid_disks = conf->geo.raid_disks;
3992 mddev->reshape_position = conf->reshape_progress;
3993 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3995 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3996 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3997 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3998 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4000 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4002 if (!mddev->sync_thread) {
4006 conf->reshape_checkpoint = jiffies;
4007 md_wakeup_thread(mddev->sync_thread);
4008 md_new_event(mddev);
4012 mddev->recovery = 0;
4013 spin_lock_irq(&conf->device_lock);
4014 conf->geo = conf->prev;
4015 mddev->raid_disks = conf->geo.raid_disks;
4016 rdev_for_each(rdev, mddev)
4017 rdev->new_data_offset = rdev->data_offset;
4019 conf->reshape_progress = MaxSector;
4020 mddev->reshape_position = MaxSector;
4021 spin_unlock_irq(&conf->device_lock);
4025 /* Calculate the last device-address that could contain
4026 * any block from the chunk that includes the array-address 's'
4027 * and report the next address.
4028 * i.e. the address returned will be chunk-aligned and after
4029 * any data that is in the chunk containing 's'.
4031 static sector_t last_dev_address(sector_t s, struct geom *geo)
4033 s = (s | geo->chunk_mask) + 1;
4034 s >>= geo->chunk_shift;
4035 s *= geo->near_copies;
4036 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4037 s *= geo->far_copies;
4038 s <<= geo->chunk_shift;
4042 /* Calculate the first device-address that could contain
4043 * any block from the chunk that includes the array-address 's'.
4044 * This too will be the start of a chunk
4046 static sector_t first_dev_address(sector_t s, struct geom *geo)
4048 s >>= geo->chunk_shift;
4049 s *= geo->near_copies;
4050 sector_div(s, geo->raid_disks);
4051 s *= geo->far_copies;
4052 s <<= geo->chunk_shift;
4056 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4059 /* We simply copy at most one chunk (smallest of old and new)
4060 * at a time, possibly less if that exceeds RESYNC_PAGES,
4061 * or we hit a bad block or something.
4062 * This might mean we pause for normal IO in the middle of
4063 * a chunk, but that is not a problem was mddev->reshape_position
4064 * can record any location.
4066 * If we will want to write to a location that isn't
4067 * yet recorded as 'safe' (i.e. in metadata on disk) then
4068 * we need to flush all reshape requests and update the metadata.
4070 * When reshaping forwards (e.g. to more devices), we interpret
4071 * 'safe' as the earliest block which might not have been copied
4072 * down yet. We divide this by previous stripe size and multiply
4073 * by previous stripe length to get lowest device offset that we
4074 * cannot write to yet.
4075 * We interpret 'sector_nr' as an address that we want to write to.
4076 * From this we use last_device_address() to find where we might
4077 * write to, and first_device_address on the 'safe' position.
4078 * If this 'next' write position is after the 'safe' position,
4079 * we must update the metadata to increase the 'safe' position.
4081 * When reshaping backwards, we round in the opposite direction
4082 * and perform the reverse test: next write position must not be
4083 * less than current safe position.
4085 * In all this the minimum difference in data offsets
4086 * (conf->offset_diff - always positive) allows a bit of slack,
4087 * so next can be after 'safe', but not by more than offset_disk
4089 * We need to prepare all the bios here before we start any IO
4090 * to ensure the size we choose is acceptable to all devices.
4091 * The means one for each copy for write-out and an extra one for
4093 * We store the read-in bio in ->master_bio and the others in
4094 * ->devs[x].bio and ->devs[x].repl_bio.
4096 struct r10conf *conf = mddev->private;
4097 struct r10bio *r10_bio;
4098 sector_t next, safe, last;
4102 struct md_rdev *rdev;
4105 struct bio *bio, *read_bio;
4106 int sectors_done = 0;
4108 if (sector_nr == 0) {
4109 /* If restarting in the middle, skip the initial sectors */
4110 if (mddev->reshape_backwards &&
4111 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4112 sector_nr = (raid10_size(mddev, 0, 0)
4113 - conf->reshape_progress);
4114 } else if (!mddev->reshape_backwards &&
4115 conf->reshape_progress > 0)
4116 sector_nr = conf->reshape_progress;
4118 mddev->curr_resync_completed = sector_nr;
4119 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4125 /* We don't use sector_nr to track where we are up to
4126 * as that doesn't work well for ->reshape_backwards.
4127 * So just use ->reshape_progress.
4129 if (mddev->reshape_backwards) {
4130 /* 'next' is the earliest device address that we might
4131 * write to for this chunk in the new layout
4133 next = first_dev_address(conf->reshape_progress - 1,
4136 /* 'safe' is the last device address that we might read from
4137 * in the old layout after a restart
4139 safe = last_dev_address(conf->reshape_safe - 1,
4142 if (next + conf->offset_diff < safe)
4145 last = conf->reshape_progress - 1;
4146 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4147 & conf->prev.chunk_mask);
4148 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4149 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4151 /* 'next' is after the last device address that we
4152 * might write to for this chunk in the new layout
4154 next = last_dev_address(conf->reshape_progress, &conf->geo);
4156 /* 'safe' is the earliest device address that we might
4157 * read from in the old layout after a restart
4159 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4161 /* Need to update metadata if 'next' might be beyond 'safe'
4162 * as that would possibly corrupt data
4164 if (next > safe + conf->offset_diff)
4167 sector_nr = conf->reshape_progress;
4168 last = sector_nr | (conf->geo.chunk_mask
4169 & conf->prev.chunk_mask);
4171 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4172 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4176 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4177 /* Need to update reshape_position in metadata */
4179 mddev->reshape_position = conf->reshape_progress;
4180 if (mddev->reshape_backwards)
4181 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4182 - conf->reshape_progress;
4184 mddev->curr_resync_completed = conf->reshape_progress;
4185 conf->reshape_checkpoint = jiffies;
4186 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4187 md_wakeup_thread(mddev->thread);
4188 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4189 kthread_should_stop());
4190 conf->reshape_safe = mddev->reshape_position;
4191 allow_barrier(conf);
4195 /* Now schedule reads for blocks from sector_nr to last */
4196 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4197 raise_barrier(conf, sectors_done != 0);
4198 atomic_set(&r10_bio->remaining, 0);
4199 r10_bio->mddev = mddev;
4200 r10_bio->sector = sector_nr;
4201 set_bit(R10BIO_IsReshape, &r10_bio->state);
4202 r10_bio->sectors = last - sector_nr + 1;
4203 rdev = read_balance(conf, r10_bio, &max_sectors);
4204 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4207 /* Cannot read from here, so need to record bad blocks
4208 * on all the target devices.
4211 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4212 return sectors_done;
4215 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4217 read_bio->bi_bdev = rdev->bdev;
4218 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4219 + rdev->data_offset);
4220 read_bio->bi_private = r10_bio;
4221 read_bio->bi_end_io = end_sync_read;
4222 read_bio->bi_rw = READ;
4223 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4224 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4225 read_bio->bi_vcnt = 0;
4226 read_bio->bi_idx = 0;
4227 read_bio->bi_size = 0;
4228 r10_bio->master_bio = read_bio;
4229 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4231 /* Now find the locations in the new layout */
4232 __raid10_find_phys(&conf->geo, r10_bio);
4235 read_bio->bi_next = NULL;
4237 for (s = 0; s < conf->copies*2; s++) {
4239 int d = r10_bio->devs[s/2].devnum;
4240 struct md_rdev *rdev2;
4242 rdev2 = conf->mirrors[d].replacement;
4243 b = r10_bio->devs[s/2].repl_bio;
4245 rdev2 = conf->mirrors[d].rdev;
4246 b = r10_bio->devs[s/2].bio;
4248 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4250 b->bi_bdev = rdev2->bdev;
4251 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4252 b->bi_private = r10_bio;
4253 b->bi_end_io = end_reshape_write;
4255 b->bi_flags &= ~(BIO_POOL_MASK - 1);
4256 b->bi_flags |= 1 << BIO_UPTODATE;
4264 /* Now add as many pages as possible to all of these bios. */
4267 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4268 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4269 int len = (max_sectors - s) << 9;
4270 if (len > PAGE_SIZE)
4272 for (bio = blist; bio ; bio = bio->bi_next) {
4274 if (bio_add_page(bio, page, len, 0))
4277 /* Didn't fit, must stop */
4279 bio2 && bio2 != bio;
4280 bio2 = bio2->bi_next) {
4281 /* Remove last page from this bio */
4283 bio2->bi_size -= len;
4284 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4288 sector_nr += len >> 9;
4289 nr_sectors += len >> 9;
4292 r10_bio->sectors = nr_sectors;
4294 /* Now submit the read */
4295 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4296 atomic_inc(&r10_bio->remaining);
4297 read_bio->bi_next = NULL;
4298 generic_make_request(read_bio);
4299 sector_nr += nr_sectors;
4300 sectors_done += nr_sectors;
4301 if (sector_nr <= last)
4304 /* Now that we have done the whole section we can
4305 * update reshape_progress
4307 if (mddev->reshape_backwards)
4308 conf->reshape_progress -= sectors_done;
4310 conf->reshape_progress += sectors_done;
4312 return sectors_done;
4315 static void end_reshape_request(struct r10bio *r10_bio);
4316 static int handle_reshape_read_error(struct mddev *mddev,
4317 struct r10bio *r10_bio);
4318 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4320 /* Reshape read completed. Hopefully we have a block
4322 * If we got a read error then we do sync 1-page reads from
4323 * elsewhere until we find the data - or give up.
4325 struct r10conf *conf = mddev->private;
4328 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4329 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4330 /* Reshape has been aborted */
4331 md_done_sync(mddev, r10_bio->sectors, 0);
4335 /* We definitely have the data in the pages, schedule the
4338 atomic_set(&r10_bio->remaining, 1);
4339 for (s = 0; s < conf->copies*2; s++) {
4341 int d = r10_bio->devs[s/2].devnum;
4342 struct md_rdev *rdev;
4344 rdev = conf->mirrors[d].replacement;
4345 b = r10_bio->devs[s/2].repl_bio;
4347 rdev = conf->mirrors[d].rdev;
4348 b = r10_bio->devs[s/2].bio;
4350 if (!rdev || test_bit(Faulty, &rdev->flags))
4352 atomic_inc(&rdev->nr_pending);
4353 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4354 atomic_inc(&r10_bio->remaining);
4356 generic_make_request(b);
4358 end_reshape_request(r10_bio);
4361 static void end_reshape(struct r10conf *conf)
4363 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4366 spin_lock_irq(&conf->device_lock);
4367 conf->prev = conf->geo;
4368 md_finish_reshape(conf->mddev);
4370 conf->reshape_progress = MaxSector;
4371 spin_unlock_irq(&conf->device_lock);
4373 /* read-ahead size must cover two whole stripes, which is
4374 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4376 if (conf->mddev->queue) {
4377 int stripe = conf->geo.raid_disks *
4378 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4379 stripe /= conf->geo.near_copies;
4380 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4381 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4387 static int handle_reshape_read_error(struct mddev *mddev,
4388 struct r10bio *r10_bio)
4390 /* Use sync reads to get the blocks from somewhere else */
4391 int sectors = r10_bio->sectors;
4393 struct r10conf *conf = mddev->private;
4396 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4398 r10b.sector = r10_bio->sector;
4399 __raid10_find_phys(&conf->prev, &r10b);
4404 int first_slot = slot;
4406 if (s > (PAGE_SIZE >> 9))
4410 int d = r10b.devs[slot].devnum;
4411 struct md_rdev *rdev = conf->mirrors[d].rdev;
4414 test_bit(Faulty, &rdev->flags) ||
4415 !test_bit(In_sync, &rdev->flags))
4418 addr = r10b.devs[slot].addr + idx * PAGE_SIZE;
4419 success = sync_page_io(rdev,
4428 if (slot >= conf->copies)
4430 if (slot == first_slot)
4434 /* couldn't read this block, must give up */
4435 set_bit(MD_RECOVERY_INTR,
4445 static void end_reshape_write(struct bio *bio, int error)
4447 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4448 struct r10bio *r10_bio = bio->bi_private;
4449 struct mddev *mddev = r10_bio->mddev;
4450 struct r10conf *conf = mddev->private;
4454 struct md_rdev *rdev = NULL;
4456 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4458 rdev = conf->mirrors[d].replacement;
4461 rdev = conf->mirrors[d].rdev;
4465 /* FIXME should record badblock */
4466 md_error(mddev, rdev);
4469 rdev_dec_pending(rdev, mddev);
4470 end_reshape_request(r10_bio);
4473 static void end_reshape_request(struct r10bio *r10_bio)
4475 if (!atomic_dec_and_test(&r10_bio->remaining))
4477 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4478 bio_put(r10_bio->master_bio);
4482 static void raid10_finish_reshape(struct mddev *mddev)
4484 struct r10conf *conf = mddev->private;
4486 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4489 if (mddev->delta_disks > 0) {
4490 sector_t size = raid10_size(mddev, 0, 0);
4491 md_set_array_sectors(mddev, size);
4492 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4493 mddev->recovery_cp = mddev->resync_max_sectors;
4494 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4496 mddev->resync_max_sectors = size;
4497 set_capacity(mddev->gendisk, mddev->array_sectors);
4498 revalidate_disk(mddev->gendisk);
4501 for (d = conf->geo.raid_disks ;
4502 d < conf->geo.raid_disks - mddev->delta_disks;
4504 struct md_rdev *rdev = conf->mirrors[d].rdev;
4506 clear_bit(In_sync, &rdev->flags);
4507 rdev = conf->mirrors[d].replacement;
4509 clear_bit(In_sync, &rdev->flags);
4512 mddev->layout = mddev->new_layout;
4513 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4514 mddev->reshape_position = MaxSector;
4515 mddev->delta_disks = 0;
4516 mddev->reshape_backwards = 0;
4519 static struct md_personality raid10_personality =
4523 .owner = THIS_MODULE,
4524 .make_request = make_request,
4528 .error_handler = error,
4529 .hot_add_disk = raid10_add_disk,
4530 .hot_remove_disk= raid10_remove_disk,
4531 .spare_active = raid10_spare_active,
4532 .sync_request = sync_request,
4533 .quiesce = raid10_quiesce,
4534 .size = raid10_size,
4535 .resize = raid10_resize,
4536 .takeover = raid10_takeover,
4537 .check_reshape = raid10_check_reshape,
4538 .start_reshape = raid10_start_reshape,
4539 .finish_reshape = raid10_finish_reshape,
4542 static int __init raid_init(void)
4544 return register_md_personality(&raid10_personality);
4547 static void raid_exit(void)
4549 unregister_md_personality(&raid10_personality);
4552 module_init(raid_init);
4553 module_exit(raid_exit);
4554 MODULE_LICENSE("GPL");
4555 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4556 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4557 MODULE_ALIAS("md-raid10");
4558 MODULE_ALIAS("md-level-10");
4560 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);