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 )
41 * use_far_sets (stored in bit 17 of layout )
43 * The data to be stored is divided into chunks using chunksize. Each device
44 * is divided into far_copies sections. In each section, chunks are laid out
45 * in a style similar to raid0, but near_copies copies of each chunk is stored
46 * (each on a different drive). The starting device for each section is offset
47 * near_copies from the starting device of the previous section. Thus there
48 * are (near_copies * far_copies) of each chunk, and each is on a different
49 * drive. near_copies and far_copies must be at least one, and their product
50 * is at most raid_disks.
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of being very far
54 * apart on disk, there are adjacent stripes.
56 * The far and offset algorithms are handled slightly differently if
57 * 'use_far_sets' is true. In this case, the array's devices are grouped into
58 * sets that are (near_copies * far_copies) in size. The far copied stripes
59 * are still shifted by 'near_copies' devices, but this shifting stays confined
60 * to the set rather than the entire array. This is done to improve the number
61 * of device combinations that can fail without causing the array to fail.
62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68 * [A B] [C D] [A B] [C D E]
69 * |...| |...| |...| | ... |
70 * [B A] [D C] [B A] [E C D]
74 * Number of guaranteed r10bios in case of extreme VM load:
76 #define NR_RAID10_BIOS 256
78 /* when we get a read error on a read-only array, we redirect to another
79 * device without failing the first device, or trying to over-write to
80 * correct the read error. To keep track of bad blocks on a per-bio
81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83 #define IO_BLOCKED ((struct bio *)1)
84 /* When we successfully write to a known bad-block, we need to remove the
85 * bad-block marking which must be done from process context. So we record
86 * the success by setting devs[n].bio to IO_MADE_GOOD
88 #define IO_MADE_GOOD ((struct bio *)2)
90 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92 /* When there are this many requests queued to be written by
93 * the raid10 thread, we become 'congested' to provide back-pressure
96 static int max_queued_requests = 1024;
98 static void allow_barrier(struct r10conf *conf);
99 static void lower_barrier(struct r10conf *conf);
100 static int enough(struct r10conf *conf, int ignore);
101 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
103 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
104 static void end_reshape_write(struct bio *bio, int error);
105 static void end_reshape(struct r10conf *conf);
107 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
109 struct r10conf *conf = data;
110 int size = offsetof(struct r10bio, devs[conf->copies]);
112 /* allocate a r10bio with room for raid_disks entries in the
114 return kzalloc(size, gfp_flags);
117 static void r10bio_pool_free(void *r10_bio, void *data)
122 /* Maximum size of each resync request */
123 #define RESYNC_BLOCK_SIZE (64*1024)
124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
125 /* amount of memory to reserve for resync requests */
126 #define RESYNC_WINDOW (1024*1024)
127 /* maximum number of concurrent requests, memory permitting */
128 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
131 * When performing a resync, we need to read and compare, so
132 * we need as many pages are there are copies.
133 * When performing a recovery, we need 2 bios, one for read,
134 * one for write (we recover only one drive per r10buf)
137 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
139 struct r10conf *conf = data;
141 struct r10bio *r10_bio;
146 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
150 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
151 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
152 nalloc = conf->copies; /* resync */
154 nalloc = 2; /* recovery */
159 for (j = nalloc ; j-- ; ) {
160 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
163 r10_bio->devs[j].bio = bio;
164 if (!conf->have_replacement)
166 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
169 r10_bio->devs[j].repl_bio = bio;
172 * Allocate RESYNC_PAGES data pages and attach them
175 for (j = 0 ; j < nalloc; j++) {
176 struct bio *rbio = r10_bio->devs[j].repl_bio;
177 bio = r10_bio->devs[j].bio;
178 for (i = 0; i < RESYNC_PAGES; i++) {
179 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
180 &conf->mddev->recovery)) {
181 /* we can share bv_page's during recovery
183 struct bio *rbio = r10_bio->devs[0].bio;
184 page = rbio->bi_io_vec[i].bv_page;
187 page = alloc_page(gfp_flags);
191 bio->bi_io_vec[i].bv_page = page;
193 rbio->bi_io_vec[i].bv_page = page;
201 safe_put_page(bio->bi_io_vec[i-1].bv_page);
203 for (i = 0; i < RESYNC_PAGES ; i++)
204 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
207 for ( ; j < nalloc; j++) {
208 if (r10_bio->devs[j].bio)
209 bio_put(r10_bio->devs[j].bio);
210 if (r10_bio->devs[j].repl_bio)
211 bio_put(r10_bio->devs[j].repl_bio);
213 r10bio_pool_free(r10_bio, conf);
217 static void r10buf_pool_free(void *__r10_bio, void *data)
220 struct r10conf *conf = data;
221 struct r10bio *r10bio = __r10_bio;
224 for (j=0; j < conf->copies; j++) {
225 struct bio *bio = r10bio->devs[j].bio;
227 for (i = 0; i < RESYNC_PAGES; i++) {
228 safe_put_page(bio->bi_io_vec[i].bv_page);
229 bio->bi_io_vec[i].bv_page = NULL;
233 bio = r10bio->devs[j].repl_bio;
237 r10bio_pool_free(r10bio, conf);
240 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
244 for (i = 0; i < conf->copies; i++) {
245 struct bio **bio = & r10_bio->devs[i].bio;
246 if (!BIO_SPECIAL(*bio))
249 bio = &r10_bio->devs[i].repl_bio;
250 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
256 static void free_r10bio(struct r10bio *r10_bio)
258 struct r10conf *conf = r10_bio->mddev->private;
260 put_all_bios(conf, r10_bio);
261 mempool_free(r10_bio, conf->r10bio_pool);
264 static void put_buf(struct r10bio *r10_bio)
266 struct r10conf *conf = r10_bio->mddev->private;
268 mempool_free(r10_bio, conf->r10buf_pool);
273 static void reschedule_retry(struct r10bio *r10_bio)
276 struct mddev *mddev = r10_bio->mddev;
277 struct r10conf *conf = mddev->private;
279 spin_lock_irqsave(&conf->device_lock, flags);
280 list_add(&r10_bio->retry_list, &conf->retry_list);
282 spin_unlock_irqrestore(&conf->device_lock, flags);
284 /* wake up frozen array... */
285 wake_up(&conf->wait_barrier);
287 md_wakeup_thread(mddev->thread);
291 * raid_end_bio_io() is called when we have finished servicing a mirrored
292 * operation and are ready to return a success/failure code to the buffer
295 static void raid_end_bio_io(struct r10bio *r10_bio)
297 struct bio *bio = r10_bio->master_bio;
299 struct r10conf *conf = r10_bio->mddev->private;
301 if (bio->bi_phys_segments) {
303 spin_lock_irqsave(&conf->device_lock, flags);
304 bio->bi_phys_segments--;
305 done = (bio->bi_phys_segments == 0);
306 spin_unlock_irqrestore(&conf->device_lock, flags);
309 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
310 clear_bit(BIO_UPTODATE, &bio->bi_flags);
314 * Wake up any possible resync thread that waits for the device
319 free_r10bio(r10_bio);
323 * Update disk head position estimator based on IRQ completion info.
325 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
327 struct r10conf *conf = r10_bio->mddev->private;
329 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
330 r10_bio->devs[slot].addr + (r10_bio->sectors);
334 * Find the disk number which triggered given bio
336 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
337 struct bio *bio, int *slotp, int *replp)
342 for (slot = 0; slot < conf->copies; slot++) {
343 if (r10_bio->devs[slot].bio == bio)
345 if (r10_bio->devs[slot].repl_bio == bio) {
351 BUG_ON(slot == conf->copies);
352 update_head_pos(slot, r10_bio);
358 return r10_bio->devs[slot].devnum;
361 static void raid10_end_read_request(struct bio *bio, int error)
363 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
364 struct r10bio *r10_bio = bio->bi_private;
366 struct md_rdev *rdev;
367 struct r10conf *conf = r10_bio->mddev->private;
370 slot = r10_bio->read_slot;
371 dev = r10_bio->devs[slot].devnum;
372 rdev = r10_bio->devs[slot].rdev;
374 * this branch is our 'one mirror IO has finished' event handler:
376 update_head_pos(slot, r10_bio);
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
388 set_bit(R10BIO_Uptodate, &r10_bio->state);
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
396 spin_lock_irqsave(&conf->device_lock, flags);
397 if (!enough(conf, rdev->raid_disk))
399 spin_unlock_irqrestore(&conf->device_lock, flags);
402 raid_end_bio_io(r10_bio);
403 rdev_dec_pending(rdev, conf->mddev);
406 * oops, read error - keep the refcount on the rdev
408 char b[BDEVNAME_SIZE];
409 printk_ratelimited(KERN_ERR
410 "md/raid10:%s: %s: rescheduling sector %llu\n",
412 bdevname(rdev->bdev, b),
413 (unsigned long long)r10_bio->sector);
414 set_bit(R10BIO_ReadError, &r10_bio->state);
415 reschedule_retry(r10_bio);
419 static void close_write(struct r10bio *r10_bio)
421 /* clear the bitmap if all writes complete successfully */
422 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
424 !test_bit(R10BIO_Degraded, &r10_bio->state),
426 md_write_end(r10_bio->mddev);
429 static void one_write_done(struct r10bio *r10_bio)
431 if (atomic_dec_and_test(&r10_bio->remaining)) {
432 if (test_bit(R10BIO_WriteError, &r10_bio->state))
433 reschedule_retry(r10_bio);
435 close_write(r10_bio);
436 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
437 reschedule_retry(r10_bio);
439 raid_end_bio_io(r10_bio);
444 static void raid10_end_write_request(struct bio *bio, int error)
446 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
447 struct r10bio *r10_bio = bio->bi_private;
450 struct r10conf *conf = r10_bio->mddev->private;
452 struct md_rdev *rdev = NULL;
454 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
457 rdev = conf->mirrors[dev].replacement;
461 rdev = conf->mirrors[dev].rdev;
464 * this branch is our 'one mirror IO has finished' event handler:
468 /* Never record new bad blocks to replacement,
471 md_error(rdev->mddev, rdev);
473 set_bit(WriteErrorSeen, &rdev->flags);
474 if (!test_and_set_bit(WantReplacement, &rdev->flags))
475 set_bit(MD_RECOVERY_NEEDED,
476 &rdev->mddev->recovery);
477 set_bit(R10BIO_WriteError, &r10_bio->state);
482 * Set R10BIO_Uptodate in our master bio, so that
483 * we will return a good error code for to the higher
484 * levels even if IO on some other mirrored buffer fails.
486 * The 'master' represents the composite IO operation to
487 * user-side. So if something waits for IO, then it will
488 * wait for the 'master' bio.
493 set_bit(R10BIO_Uptodate, &r10_bio->state);
495 /* Maybe we can clear some bad blocks. */
496 if (is_badblock(rdev,
497 r10_bio->devs[slot].addr,
499 &first_bad, &bad_sectors)) {
502 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
504 r10_bio->devs[slot].bio = IO_MADE_GOOD;
506 set_bit(R10BIO_MadeGood, &r10_bio->state);
512 * Let's see if all mirrored write operations have finished
515 one_write_done(r10_bio);
517 rdev_dec_pending(rdev, conf->mddev);
521 * RAID10 layout manager
522 * As well as the chunksize and raid_disks count, there are two
523 * parameters: near_copies and far_copies.
524 * near_copies * far_copies must be <= raid_disks.
525 * Normally one of these will be 1.
526 * If both are 1, we get raid0.
527 * If near_copies == raid_disks, we get raid1.
529 * Chunks are laid out in raid0 style with near_copies copies of the
530 * first chunk, followed by near_copies copies of the next chunk and
532 * If far_copies > 1, then after 1/far_copies of the array has been assigned
533 * as described above, we start again with a device offset of near_copies.
534 * So we effectively have another copy of the whole array further down all
535 * the drives, but with blocks on different drives.
536 * With this layout, and block is never stored twice on the one device.
538 * raid10_find_phys finds the sector offset of a given virtual sector
539 * on each device that it is on.
541 * raid10_find_virt does the reverse mapping, from a device and a
542 * sector offset to a virtual address
545 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
553 int last_far_set_start, last_far_set_size;
555 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
556 last_far_set_start *= geo->far_set_size;
558 last_far_set_size = geo->far_set_size;
559 last_far_set_size += (geo->raid_disks % geo->far_set_size);
561 /* now calculate first sector/dev */
562 chunk = r10bio->sector >> geo->chunk_shift;
563 sector = r10bio->sector & geo->chunk_mask;
565 chunk *= geo->near_copies;
567 dev = sector_div(stripe, geo->raid_disks);
569 stripe *= geo->far_copies;
571 sector += stripe << geo->chunk_shift;
573 /* and calculate all the others */
574 for (n = 0; n < geo->near_copies; n++) {
578 r10bio->devs[slot].devnum = d;
579 r10bio->devs[slot].addr = s;
582 for (f = 1; f < geo->far_copies; f++) {
583 set = d / geo->far_set_size;
584 d += geo->near_copies;
586 if ((geo->raid_disks % geo->far_set_size) &&
587 (d > last_far_set_start)) {
588 d -= last_far_set_start;
589 d %= last_far_set_size;
590 d += last_far_set_start;
592 d %= geo->far_set_size;
593 d += geo->far_set_size * set;
596 r10bio->devs[slot].devnum = d;
597 r10bio->devs[slot].addr = s;
601 if (dev >= geo->raid_disks) {
603 sector += (geo->chunk_mask + 1);
608 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
610 struct geom *geo = &conf->geo;
612 if (conf->reshape_progress != MaxSector &&
613 ((r10bio->sector >= conf->reshape_progress) !=
614 conf->mddev->reshape_backwards)) {
615 set_bit(R10BIO_Previous, &r10bio->state);
618 clear_bit(R10BIO_Previous, &r10bio->state);
620 __raid10_find_phys(geo, r10bio);
623 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
625 sector_t offset, chunk, vchunk;
626 /* Never use conf->prev as this is only called during resync
627 * or recovery, so reshape isn't happening
629 struct geom *geo = &conf->geo;
630 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
631 int far_set_size = geo->far_set_size;
632 int last_far_set_start;
634 if (geo->raid_disks % geo->far_set_size) {
635 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
636 last_far_set_start *= geo->far_set_size;
638 if (dev >= last_far_set_start) {
639 far_set_size = geo->far_set_size;
640 far_set_size += (geo->raid_disks % geo->far_set_size);
641 far_set_start = last_far_set_start;
645 offset = sector & geo->chunk_mask;
646 if (geo->far_offset) {
648 chunk = sector >> geo->chunk_shift;
649 fc = sector_div(chunk, geo->far_copies);
650 dev -= fc * geo->near_copies;
651 if (dev < far_set_start)
654 while (sector >= geo->stride) {
655 sector -= geo->stride;
656 if (dev < (geo->near_copies + far_set_start))
657 dev += far_set_size - geo->near_copies;
659 dev -= geo->near_copies;
661 chunk = sector >> geo->chunk_shift;
663 vchunk = chunk * geo->raid_disks + dev;
664 sector_div(vchunk, geo->near_copies);
665 return (vchunk << geo->chunk_shift) + offset;
669 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
671 * @bvm: properties of new bio
672 * @biovec: the request that could be merged to it.
674 * Return amount of bytes we can accept at this offset
675 * This requires checking for end-of-chunk if near_copies != raid_disks,
676 * and for subordinate merge_bvec_fns if merge_check_needed.
678 static int raid10_mergeable_bvec(struct request_queue *q,
679 struct bvec_merge_data *bvm,
680 struct bio_vec *biovec)
682 struct mddev *mddev = q->queuedata;
683 struct r10conf *conf = mddev->private;
684 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
686 unsigned int chunk_sectors;
687 unsigned int bio_sectors = bvm->bi_size >> 9;
688 struct geom *geo = &conf->geo;
690 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
691 if (conf->reshape_progress != MaxSector &&
692 ((sector >= conf->reshape_progress) !=
693 conf->mddev->reshape_backwards))
696 if (geo->near_copies < geo->raid_disks) {
697 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
698 + bio_sectors)) << 9;
700 /* bio_add cannot handle a negative return */
702 if (max <= biovec->bv_len && bio_sectors == 0)
703 return biovec->bv_len;
705 max = biovec->bv_len;
707 if (mddev->merge_check_needed) {
709 struct r10bio r10_bio;
710 struct r10dev devs[conf->copies];
712 struct r10bio *r10_bio = &on_stack.r10_bio;
714 if (conf->reshape_progress != MaxSector) {
715 /* Cannot give any guidance during reshape */
716 if (max <= biovec->bv_len && bio_sectors == 0)
717 return biovec->bv_len;
720 r10_bio->sector = sector;
721 raid10_find_phys(conf, r10_bio);
723 for (s = 0; s < conf->copies; s++) {
724 int disk = r10_bio->devs[s].devnum;
725 struct md_rdev *rdev = rcu_dereference(
726 conf->mirrors[disk].rdev);
727 if (rdev && !test_bit(Faulty, &rdev->flags)) {
728 struct request_queue *q =
729 bdev_get_queue(rdev->bdev);
730 if (q->merge_bvec_fn) {
731 bvm->bi_sector = r10_bio->devs[s].addr
733 bvm->bi_bdev = rdev->bdev;
734 max = min(max, q->merge_bvec_fn(
738 rdev = rcu_dereference(conf->mirrors[disk].replacement);
739 if (rdev && !test_bit(Faulty, &rdev->flags)) {
740 struct request_queue *q =
741 bdev_get_queue(rdev->bdev);
742 if (q->merge_bvec_fn) {
743 bvm->bi_sector = r10_bio->devs[s].addr
745 bvm->bi_bdev = rdev->bdev;
746 max = min(max, q->merge_bvec_fn(
757 * This routine returns the disk from which the requested read should
758 * be done. There is a per-array 'next expected sequential IO' sector
759 * number - if this matches on the next IO then we use the last disk.
760 * There is also a per-disk 'last know head position' sector that is
761 * maintained from IRQ contexts, both the normal and the resync IO
762 * completion handlers update this position correctly. If there is no
763 * perfect sequential match then we pick the disk whose head is closest.
765 * If there are 2 mirrors in the same 2 devices, performance degrades
766 * because position is mirror, not device based.
768 * The rdev for the device selected will have nr_pending incremented.
772 * FIXME: possibly should rethink readbalancing and do it differently
773 * depending on near_copies / far_copies geometry.
775 static struct md_rdev *read_balance(struct r10conf *conf,
776 struct r10bio *r10_bio,
779 const sector_t this_sector = r10_bio->sector;
781 int sectors = r10_bio->sectors;
782 int best_good_sectors;
783 sector_t new_distance, best_dist;
784 struct md_rdev *best_rdev, *rdev = NULL;
787 struct geom *geo = &conf->geo;
789 raid10_find_phys(conf, r10_bio);
792 sectors = r10_bio->sectors;
795 best_dist = MaxSector;
796 best_good_sectors = 0;
799 * Check if we can balance. We can balance on the whole
800 * device if no resync is going on (recovery is ok), or below
801 * the resync window. We take the first readable disk when
802 * above the resync window.
804 if (conf->mddev->recovery_cp < MaxSector
805 && (this_sector + sectors >= conf->next_resync))
808 for (slot = 0; slot < conf->copies ; slot++) {
813 if (r10_bio->devs[slot].bio == IO_BLOCKED)
815 disk = r10_bio->devs[slot].devnum;
816 rdev = rcu_dereference(conf->mirrors[disk].replacement);
817 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
818 test_bit(Unmerged, &rdev->flags) ||
819 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
820 rdev = rcu_dereference(conf->mirrors[disk].rdev);
822 test_bit(Faulty, &rdev->flags) ||
823 test_bit(Unmerged, &rdev->flags))
825 if (!test_bit(In_sync, &rdev->flags) &&
826 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
829 dev_sector = r10_bio->devs[slot].addr;
830 if (is_badblock(rdev, dev_sector, sectors,
831 &first_bad, &bad_sectors)) {
832 if (best_dist < MaxSector)
833 /* Already have a better slot */
835 if (first_bad <= dev_sector) {
836 /* Cannot read here. If this is the
837 * 'primary' device, then we must not read
838 * beyond 'bad_sectors' from another device.
840 bad_sectors -= (dev_sector - first_bad);
841 if (!do_balance && sectors > bad_sectors)
842 sectors = bad_sectors;
843 if (best_good_sectors > sectors)
844 best_good_sectors = sectors;
846 sector_t good_sectors =
847 first_bad - dev_sector;
848 if (good_sectors > best_good_sectors) {
849 best_good_sectors = good_sectors;
854 /* Must read from here */
859 best_good_sectors = sectors;
864 /* This optimisation is debatable, and completely destroys
865 * sequential read speed for 'far copies' arrays. So only
866 * keep it for 'near' arrays, and review those later.
868 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
871 /* for far > 1 always use the lowest address */
872 if (geo->far_copies > 1)
873 new_distance = r10_bio->devs[slot].addr;
875 new_distance = abs(r10_bio->devs[slot].addr -
876 conf->mirrors[disk].head_position);
877 if (new_distance < best_dist) {
878 best_dist = new_distance;
883 if (slot >= conf->copies) {
889 atomic_inc(&rdev->nr_pending);
890 if (test_bit(Faulty, &rdev->flags)) {
891 /* Cannot risk returning a device that failed
892 * before we inc'ed nr_pending
894 rdev_dec_pending(rdev, conf->mddev);
897 r10_bio->read_slot = slot;
901 *max_sectors = best_good_sectors;
906 int md_raid10_congested(struct mddev *mddev, int bits)
908 struct r10conf *conf = mddev->private;
911 if ((bits & (1 << BDI_async_congested)) &&
912 conf->pending_count >= max_queued_requests)
917 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
920 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
921 if (rdev && !test_bit(Faulty, &rdev->flags)) {
922 struct request_queue *q = bdev_get_queue(rdev->bdev);
924 ret |= bdi_congested(&q->backing_dev_info, bits);
930 EXPORT_SYMBOL_GPL(md_raid10_congested);
932 static int raid10_congested(void *data, int bits)
934 struct mddev *mddev = data;
936 return mddev_congested(mddev, bits) ||
937 md_raid10_congested(mddev, bits);
940 static void flush_pending_writes(struct r10conf *conf)
942 /* Any writes that have been queued but are awaiting
943 * bitmap updates get flushed here.
945 spin_lock_irq(&conf->device_lock);
947 if (conf->pending_bio_list.head) {
949 bio = bio_list_get(&conf->pending_bio_list);
950 conf->pending_count = 0;
951 spin_unlock_irq(&conf->device_lock);
952 /* flush any pending bitmap writes to disk
953 * before proceeding w/ I/O */
954 bitmap_unplug(conf->mddev->bitmap);
955 wake_up(&conf->wait_barrier);
957 while (bio) { /* submit pending writes */
958 struct bio *next = bio->bi_next;
960 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
961 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
965 generic_make_request(bio);
969 spin_unlock_irq(&conf->device_lock);
973 * Sometimes we need to suspend IO while we do something else,
974 * either some resync/recovery, or reconfigure the array.
975 * To do this we raise a 'barrier'.
976 * The 'barrier' is a counter that can be raised multiple times
977 * to count how many activities are happening which preclude
979 * We can only raise the barrier if there is no pending IO.
980 * i.e. if nr_pending == 0.
981 * We choose only to raise the barrier if no-one is waiting for the
982 * barrier to go down. This means that as soon as an IO request
983 * is ready, no other operations which require a barrier will start
984 * until the IO request has had a chance.
986 * So: regular IO calls 'wait_barrier'. When that returns there
987 * is no backgroup IO happening, It must arrange to call
988 * allow_barrier when it has finished its IO.
989 * backgroup IO calls must call raise_barrier. Once that returns
990 * there is no normal IO happeing. It must arrange to call
991 * lower_barrier when the particular background IO completes.
994 static void raise_barrier(struct r10conf *conf, int force)
996 BUG_ON(force && !conf->barrier);
997 spin_lock_irq(&conf->resync_lock);
999 /* Wait until no block IO is waiting (unless 'force') */
1000 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1003 /* block any new IO from starting */
1006 /* Now wait for all pending IO to complete */
1007 wait_event_lock_irq(conf->wait_barrier,
1008 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1011 spin_unlock_irq(&conf->resync_lock);
1014 static void lower_barrier(struct r10conf *conf)
1016 unsigned long flags;
1017 spin_lock_irqsave(&conf->resync_lock, flags);
1019 spin_unlock_irqrestore(&conf->resync_lock, flags);
1020 wake_up(&conf->wait_barrier);
1023 static void wait_barrier(struct r10conf *conf)
1025 spin_lock_irq(&conf->resync_lock);
1026 if (conf->barrier) {
1028 /* Wait for the barrier to drop.
1029 * However if there are already pending
1030 * requests (preventing the barrier from
1031 * rising completely), and the
1032 * pre-process bio queue isn't empty,
1033 * then don't wait, as we need to empty
1034 * that queue to get the nr_pending
1037 wait_event_lock_irq(conf->wait_barrier,
1039 (conf->nr_pending &&
1040 current->bio_list &&
1041 !bio_list_empty(current->bio_list)),
1046 spin_unlock_irq(&conf->resync_lock);
1049 static void allow_barrier(struct r10conf *conf)
1051 unsigned long flags;
1052 spin_lock_irqsave(&conf->resync_lock, flags);
1054 spin_unlock_irqrestore(&conf->resync_lock, flags);
1055 wake_up(&conf->wait_barrier);
1058 static void freeze_array(struct r10conf *conf)
1060 /* stop syncio and normal IO and wait for everything to
1062 * We increment barrier and nr_waiting, and then
1063 * wait until nr_pending match nr_queued+1
1064 * This is called in the context of one normal IO request
1065 * that has failed. Thus any sync request that might be pending
1066 * will be blocked by nr_pending, and we need to wait for
1067 * pending IO requests to complete or be queued for re-try.
1068 * Thus the number queued (nr_queued) plus this request (1)
1069 * must match the number of pending IOs (nr_pending) before
1072 spin_lock_irq(&conf->resync_lock);
1075 wait_event_lock_irq_cmd(conf->wait_barrier,
1076 conf->nr_pending == conf->nr_queued+1,
1078 flush_pending_writes(conf));
1080 spin_unlock_irq(&conf->resync_lock);
1083 static void unfreeze_array(struct r10conf *conf)
1085 /* reverse the effect of the freeze */
1086 spin_lock_irq(&conf->resync_lock);
1089 wake_up(&conf->wait_barrier);
1090 spin_unlock_irq(&conf->resync_lock);
1093 static sector_t choose_data_offset(struct r10bio *r10_bio,
1094 struct md_rdev *rdev)
1096 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1097 test_bit(R10BIO_Previous, &r10_bio->state))
1098 return rdev->data_offset;
1100 return rdev->new_data_offset;
1103 struct raid10_plug_cb {
1104 struct blk_plug_cb cb;
1105 struct bio_list pending;
1109 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1111 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1113 struct mddev *mddev = plug->cb.data;
1114 struct r10conf *conf = mddev->private;
1117 if (from_schedule || current->bio_list) {
1118 spin_lock_irq(&conf->device_lock);
1119 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1120 conf->pending_count += plug->pending_cnt;
1121 spin_unlock_irq(&conf->device_lock);
1122 wake_up(&conf->wait_barrier);
1123 md_wakeup_thread(mddev->thread);
1128 /* we aren't scheduling, so we can do the write-out directly. */
1129 bio = bio_list_get(&plug->pending);
1130 bitmap_unplug(mddev->bitmap);
1131 wake_up(&conf->wait_barrier);
1133 while (bio) { /* submit pending writes */
1134 struct bio *next = bio->bi_next;
1135 bio->bi_next = NULL;
1136 generic_make_request(bio);
1142 static void make_request(struct mddev *mddev, struct bio * bio)
1144 struct r10conf *conf = mddev->private;
1145 struct r10bio *r10_bio;
1146 struct bio *read_bio;
1148 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1149 int chunk_sects = chunk_mask + 1;
1150 const int rw = bio_data_dir(bio);
1151 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1152 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1153 const unsigned long do_discard = (bio->bi_rw
1154 & (REQ_DISCARD | REQ_SECURE));
1155 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1156 unsigned long flags;
1157 struct md_rdev *blocked_rdev;
1158 struct blk_plug_cb *cb;
1159 struct raid10_plug_cb *plug = NULL;
1160 int sectors_handled;
1164 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1165 md_flush_request(mddev, bio);
1169 /* If this request crosses a chunk boundary, we need to
1170 * split it. This will only happen for 1 PAGE (or less) requests.
1172 if (unlikely((bio->bi_sector & chunk_mask) + bio_sectors(bio)
1174 && (conf->geo.near_copies < conf->geo.raid_disks
1175 || conf->prev.near_copies < conf->prev.raid_disks))) {
1176 struct bio_pair *bp;
1177 /* Sanity check -- queue functions should prevent this happening */
1178 if ((bio->bi_vcnt != 1 && bio->bi_vcnt != 0) ||
1181 /* This is a one page bio that upper layers
1182 * refuse to split for us, so we need to split it.
1185 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1187 /* Each of these 'make_request' calls will call 'wait_barrier'.
1188 * If the first succeeds but the second blocks due to the resync
1189 * thread raising the barrier, we will deadlock because the
1190 * IO to the underlying device will be queued in generic_make_request
1191 * and will never complete, so will never reduce nr_pending.
1192 * So increment nr_waiting here so no new raise_barriers will
1193 * succeed, and so the second wait_barrier cannot block.
1195 spin_lock_irq(&conf->resync_lock);
1197 spin_unlock_irq(&conf->resync_lock);
1199 make_request(mddev, &bp->bio1);
1200 make_request(mddev, &bp->bio2);
1202 spin_lock_irq(&conf->resync_lock);
1204 wake_up(&conf->wait_barrier);
1205 spin_unlock_irq(&conf->resync_lock);
1207 bio_pair_release(bp);
1210 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1211 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1212 (unsigned long long)bio->bi_sector, bio_sectors(bio) / 2);
1218 md_write_start(mddev, bio);
1221 * Register the new request and wait if the reconstruction
1222 * thread has put up a bar for new requests.
1223 * Continue immediately if no resync is active currently.
1227 sectors = bio_sectors(bio);
1228 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1229 bio->bi_sector < conf->reshape_progress &&
1230 bio->bi_sector + sectors > conf->reshape_progress) {
1231 /* IO spans the reshape position. Need to wait for
1234 allow_barrier(conf);
1235 wait_event(conf->wait_barrier,
1236 conf->reshape_progress <= bio->bi_sector ||
1237 conf->reshape_progress >= bio->bi_sector + sectors);
1240 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1241 bio_data_dir(bio) == WRITE &&
1242 (mddev->reshape_backwards
1243 ? (bio->bi_sector < conf->reshape_safe &&
1244 bio->bi_sector + sectors > conf->reshape_progress)
1245 : (bio->bi_sector + sectors > conf->reshape_safe &&
1246 bio->bi_sector < conf->reshape_progress))) {
1247 /* Need to update reshape_position in metadata */
1248 mddev->reshape_position = conf->reshape_progress;
1249 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1250 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1251 md_wakeup_thread(mddev->thread);
1252 wait_event(mddev->sb_wait,
1253 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1255 conf->reshape_safe = mddev->reshape_position;
1258 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1260 r10_bio->master_bio = bio;
1261 r10_bio->sectors = sectors;
1263 r10_bio->mddev = mddev;
1264 r10_bio->sector = bio->bi_sector;
1267 /* We might need to issue multiple reads to different
1268 * devices if there are bad blocks around, so we keep
1269 * track of the number of reads in bio->bi_phys_segments.
1270 * If this is 0, there is only one r10_bio and no locking
1271 * will be needed when the request completes. If it is
1272 * non-zero, then it is the number of not-completed requests.
1274 bio->bi_phys_segments = 0;
1275 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1279 * read balancing logic:
1281 struct md_rdev *rdev;
1285 rdev = read_balance(conf, r10_bio, &max_sectors);
1287 raid_end_bio_io(r10_bio);
1290 slot = r10_bio->read_slot;
1292 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1293 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1296 r10_bio->devs[slot].bio = read_bio;
1297 r10_bio->devs[slot].rdev = rdev;
1299 read_bio->bi_sector = r10_bio->devs[slot].addr +
1300 choose_data_offset(r10_bio, rdev);
1301 read_bio->bi_bdev = rdev->bdev;
1302 read_bio->bi_end_io = raid10_end_read_request;
1303 read_bio->bi_rw = READ | do_sync;
1304 read_bio->bi_private = r10_bio;
1306 if (max_sectors < r10_bio->sectors) {
1307 /* Could not read all from this device, so we will
1308 * need another r10_bio.
1310 sectors_handled = (r10_bio->sectors + max_sectors
1312 r10_bio->sectors = max_sectors;
1313 spin_lock_irq(&conf->device_lock);
1314 if (bio->bi_phys_segments == 0)
1315 bio->bi_phys_segments = 2;
1317 bio->bi_phys_segments++;
1318 spin_unlock(&conf->device_lock);
1319 /* Cannot call generic_make_request directly
1320 * as that will be queued in __generic_make_request
1321 * and subsequent mempool_alloc might block
1322 * waiting for it. so hand bio over to raid10d.
1324 reschedule_retry(r10_bio);
1326 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1328 r10_bio->master_bio = bio;
1329 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1331 r10_bio->mddev = mddev;
1332 r10_bio->sector = bio->bi_sector + sectors_handled;
1335 generic_make_request(read_bio);
1342 if (conf->pending_count >= max_queued_requests) {
1343 md_wakeup_thread(mddev->thread);
1344 wait_event(conf->wait_barrier,
1345 conf->pending_count < max_queued_requests);
1347 /* first select target devices under rcu_lock and
1348 * inc refcount on their rdev. Record them by setting
1350 * If there are known/acknowledged bad blocks on any device
1351 * on which we have seen a write error, we want to avoid
1352 * writing to those blocks. This potentially requires several
1353 * writes to write around the bad blocks. Each set of writes
1354 * gets its own r10_bio with a set of bios attached. The number
1355 * of r10_bios is recored in bio->bi_phys_segments just as with
1359 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1360 raid10_find_phys(conf, r10_bio);
1362 blocked_rdev = NULL;
1364 max_sectors = r10_bio->sectors;
1366 for (i = 0; i < conf->copies; i++) {
1367 int d = r10_bio->devs[i].devnum;
1368 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1369 struct md_rdev *rrdev = rcu_dereference(
1370 conf->mirrors[d].replacement);
1373 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1374 atomic_inc(&rdev->nr_pending);
1375 blocked_rdev = rdev;
1378 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1379 atomic_inc(&rrdev->nr_pending);
1380 blocked_rdev = rrdev;
1383 if (rdev && (test_bit(Faulty, &rdev->flags)
1384 || test_bit(Unmerged, &rdev->flags)))
1386 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1387 || test_bit(Unmerged, &rrdev->flags)))
1390 r10_bio->devs[i].bio = NULL;
1391 r10_bio->devs[i].repl_bio = NULL;
1393 if (!rdev && !rrdev) {
1394 set_bit(R10BIO_Degraded, &r10_bio->state);
1397 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1399 sector_t dev_sector = r10_bio->devs[i].addr;
1403 is_bad = is_badblock(rdev, dev_sector,
1405 &first_bad, &bad_sectors);
1407 /* Mustn't write here until the bad block
1410 atomic_inc(&rdev->nr_pending);
1411 set_bit(BlockedBadBlocks, &rdev->flags);
1412 blocked_rdev = rdev;
1415 if (is_bad && first_bad <= dev_sector) {
1416 /* Cannot write here at all */
1417 bad_sectors -= (dev_sector - first_bad);
1418 if (bad_sectors < max_sectors)
1419 /* Mustn't write more than bad_sectors
1420 * to other devices yet
1422 max_sectors = bad_sectors;
1423 /* We don't set R10BIO_Degraded as that
1424 * only applies if the disk is missing,
1425 * so it might be re-added, and we want to
1426 * know to recover this chunk.
1427 * In this case the device is here, and the
1428 * fact that this chunk is not in-sync is
1429 * recorded in the bad block log.
1434 int good_sectors = first_bad - dev_sector;
1435 if (good_sectors < max_sectors)
1436 max_sectors = good_sectors;
1440 r10_bio->devs[i].bio = bio;
1441 atomic_inc(&rdev->nr_pending);
1444 r10_bio->devs[i].repl_bio = bio;
1445 atomic_inc(&rrdev->nr_pending);
1450 if (unlikely(blocked_rdev)) {
1451 /* Have to wait for this device to get unblocked, then retry */
1455 for (j = 0; j < i; j++) {
1456 if (r10_bio->devs[j].bio) {
1457 d = r10_bio->devs[j].devnum;
1458 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1460 if (r10_bio->devs[j].repl_bio) {
1461 struct md_rdev *rdev;
1462 d = r10_bio->devs[j].devnum;
1463 rdev = conf->mirrors[d].replacement;
1465 /* Race with remove_disk */
1467 rdev = conf->mirrors[d].rdev;
1469 rdev_dec_pending(rdev, mddev);
1472 allow_barrier(conf);
1473 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1478 if (max_sectors < r10_bio->sectors) {
1479 /* We are splitting this into multiple parts, so
1480 * we need to prepare for allocating another r10_bio.
1482 r10_bio->sectors = max_sectors;
1483 spin_lock_irq(&conf->device_lock);
1484 if (bio->bi_phys_segments == 0)
1485 bio->bi_phys_segments = 2;
1487 bio->bi_phys_segments++;
1488 spin_unlock_irq(&conf->device_lock);
1490 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1492 atomic_set(&r10_bio->remaining, 1);
1493 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1495 for (i = 0; i < conf->copies; i++) {
1497 int d = r10_bio->devs[i].devnum;
1498 if (r10_bio->devs[i].bio) {
1499 struct md_rdev *rdev = conf->mirrors[d].rdev;
1500 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1501 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1503 r10_bio->devs[i].bio = mbio;
1505 mbio->bi_sector = (r10_bio->devs[i].addr+
1506 choose_data_offset(r10_bio,
1508 mbio->bi_bdev = rdev->bdev;
1509 mbio->bi_end_io = raid10_end_write_request;
1511 WRITE | do_sync | do_fua | do_discard | do_same;
1512 mbio->bi_private = r10_bio;
1514 atomic_inc(&r10_bio->remaining);
1516 cb = blk_check_plugged(raid10_unplug, mddev,
1519 plug = container_of(cb, struct raid10_plug_cb,
1523 spin_lock_irqsave(&conf->device_lock, flags);
1525 bio_list_add(&plug->pending, mbio);
1526 plug->pending_cnt++;
1528 bio_list_add(&conf->pending_bio_list, mbio);
1529 conf->pending_count++;
1531 spin_unlock_irqrestore(&conf->device_lock, flags);
1533 md_wakeup_thread(mddev->thread);
1536 if (r10_bio->devs[i].repl_bio) {
1537 struct md_rdev *rdev = conf->mirrors[d].replacement;
1539 /* Replacement just got moved to main 'rdev' */
1541 rdev = conf->mirrors[d].rdev;
1543 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1544 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1546 r10_bio->devs[i].repl_bio = mbio;
1548 mbio->bi_sector = (r10_bio->devs[i].addr +
1551 mbio->bi_bdev = rdev->bdev;
1552 mbio->bi_end_io = raid10_end_write_request;
1554 WRITE | do_sync | do_fua | do_discard | do_same;
1555 mbio->bi_private = r10_bio;
1557 atomic_inc(&r10_bio->remaining);
1558 spin_lock_irqsave(&conf->device_lock, flags);
1559 bio_list_add(&conf->pending_bio_list, mbio);
1560 conf->pending_count++;
1561 spin_unlock_irqrestore(&conf->device_lock, flags);
1562 if (!mddev_check_plugged(mddev))
1563 md_wakeup_thread(mddev->thread);
1567 /* Don't remove the bias on 'remaining' (one_write_done) until
1568 * after checking if we need to go around again.
1571 if (sectors_handled < bio_sectors(bio)) {
1572 one_write_done(r10_bio);
1573 /* We need another r10_bio. It has already been counted
1574 * in bio->bi_phys_segments.
1576 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1578 r10_bio->master_bio = bio;
1579 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1581 r10_bio->mddev = mddev;
1582 r10_bio->sector = bio->bi_sector + sectors_handled;
1586 one_write_done(r10_bio);
1588 /* In case raid10d snuck in to freeze_array */
1589 wake_up(&conf->wait_barrier);
1592 static void status(struct seq_file *seq, struct mddev *mddev)
1594 struct r10conf *conf = mddev->private;
1597 if (conf->geo.near_copies < conf->geo.raid_disks)
1598 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1599 if (conf->geo.near_copies > 1)
1600 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1601 if (conf->geo.far_copies > 1) {
1602 if (conf->geo.far_offset)
1603 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1605 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1607 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1608 conf->geo.raid_disks - mddev->degraded);
1609 for (i = 0; i < conf->geo.raid_disks; i++)
1610 seq_printf(seq, "%s",
1611 conf->mirrors[i].rdev &&
1612 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1613 seq_printf(seq, "]");
1616 /* check if there are enough drives for
1617 * every block to appear on atleast one.
1618 * Don't consider the device numbered 'ignore'
1619 * as we might be about to remove it.
1621 static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1626 int n = conf->copies;
1630 if (conf->mirrors[this].rdev &&
1633 this = (this+1) % geo->raid_disks;
1637 first = (first + geo->near_copies) % geo->raid_disks;
1638 } while (first != 0);
1642 static int enough(struct r10conf *conf, int ignore)
1644 return _enough(conf, &conf->geo, ignore) &&
1645 _enough(conf, &conf->prev, ignore);
1648 static void error(struct mddev *mddev, struct md_rdev *rdev)
1650 char b[BDEVNAME_SIZE];
1651 struct r10conf *conf = mddev->private;
1654 * If it is not operational, then we have already marked it as dead
1655 * else if it is the last working disks, ignore the error, let the
1656 * next level up know.
1657 * else mark the drive as failed
1659 if (test_bit(In_sync, &rdev->flags)
1660 && !enough(conf, rdev->raid_disk))
1662 * Don't fail the drive, just return an IO error.
1665 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1666 unsigned long flags;
1667 spin_lock_irqsave(&conf->device_lock, flags);
1669 spin_unlock_irqrestore(&conf->device_lock, flags);
1671 * if recovery is running, make sure it aborts.
1673 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1675 set_bit(Blocked, &rdev->flags);
1676 set_bit(Faulty, &rdev->flags);
1677 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1679 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1680 "md/raid10:%s: Operation continuing on %d devices.\n",
1681 mdname(mddev), bdevname(rdev->bdev, b),
1682 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1685 static void print_conf(struct r10conf *conf)
1688 struct raid10_info *tmp;
1690 printk(KERN_DEBUG "RAID10 conf printout:\n");
1692 printk(KERN_DEBUG "(!conf)\n");
1695 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1696 conf->geo.raid_disks);
1698 for (i = 0; i < conf->geo.raid_disks; i++) {
1699 char b[BDEVNAME_SIZE];
1700 tmp = conf->mirrors + i;
1702 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1703 i, !test_bit(In_sync, &tmp->rdev->flags),
1704 !test_bit(Faulty, &tmp->rdev->flags),
1705 bdevname(tmp->rdev->bdev,b));
1709 static void close_sync(struct r10conf *conf)
1712 allow_barrier(conf);
1714 mempool_destroy(conf->r10buf_pool);
1715 conf->r10buf_pool = NULL;
1718 static int raid10_spare_active(struct mddev *mddev)
1721 struct r10conf *conf = mddev->private;
1722 struct raid10_info *tmp;
1724 unsigned long flags;
1727 * Find all non-in_sync disks within the RAID10 configuration
1728 * and mark them in_sync
1730 for (i = 0; i < conf->geo.raid_disks; i++) {
1731 tmp = conf->mirrors + i;
1732 if (tmp->replacement
1733 && tmp->replacement->recovery_offset == MaxSector
1734 && !test_bit(Faulty, &tmp->replacement->flags)
1735 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1736 /* Replacement has just become active */
1738 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1741 /* Replaced device not technically faulty,
1742 * but we need to be sure it gets removed
1743 * and never re-added.
1745 set_bit(Faulty, &tmp->rdev->flags);
1746 sysfs_notify_dirent_safe(
1747 tmp->rdev->sysfs_state);
1749 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1750 } else if (tmp->rdev
1751 && !test_bit(Faulty, &tmp->rdev->flags)
1752 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1754 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1757 spin_lock_irqsave(&conf->device_lock, flags);
1758 mddev->degraded -= count;
1759 spin_unlock_irqrestore(&conf->device_lock, flags);
1766 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1768 struct r10conf *conf = mddev->private;
1772 int last = conf->geo.raid_disks - 1;
1773 struct request_queue *q = bdev_get_queue(rdev->bdev);
1775 if (mddev->recovery_cp < MaxSector)
1776 /* only hot-add to in-sync arrays, as recovery is
1777 * very different from resync
1780 if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1783 if (rdev->raid_disk >= 0)
1784 first = last = rdev->raid_disk;
1786 if (q->merge_bvec_fn) {
1787 set_bit(Unmerged, &rdev->flags);
1788 mddev->merge_check_needed = 1;
1791 if (rdev->saved_raid_disk >= first &&
1792 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1793 mirror = rdev->saved_raid_disk;
1796 for ( ; mirror <= last ; mirror++) {
1797 struct raid10_info *p = &conf->mirrors[mirror];
1798 if (p->recovery_disabled == mddev->recovery_disabled)
1801 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1802 p->replacement != NULL)
1804 clear_bit(In_sync, &rdev->flags);
1805 set_bit(Replacement, &rdev->flags);
1806 rdev->raid_disk = mirror;
1808 disk_stack_limits(mddev->gendisk, rdev->bdev,
1809 rdev->data_offset << 9);
1811 rcu_assign_pointer(p->replacement, rdev);
1815 disk_stack_limits(mddev->gendisk, rdev->bdev,
1816 rdev->data_offset << 9);
1818 p->head_position = 0;
1819 p->recovery_disabled = mddev->recovery_disabled - 1;
1820 rdev->raid_disk = mirror;
1822 if (rdev->saved_raid_disk != mirror)
1824 rcu_assign_pointer(p->rdev, rdev);
1827 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1828 /* Some requests might not have seen this new
1829 * merge_bvec_fn. We must wait for them to complete
1830 * before merging the device fully.
1831 * First we make sure any code which has tested
1832 * our function has submitted the request, then
1833 * we wait for all outstanding requests to complete.
1835 synchronize_sched();
1836 raise_barrier(conf, 0);
1837 lower_barrier(conf);
1838 clear_bit(Unmerged, &rdev->flags);
1840 md_integrity_add_rdev(rdev, mddev);
1841 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1842 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1848 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1850 struct r10conf *conf = mddev->private;
1852 int number = rdev->raid_disk;
1853 struct md_rdev **rdevp;
1854 struct raid10_info *p = conf->mirrors + number;
1857 if (rdev == p->rdev)
1859 else if (rdev == p->replacement)
1860 rdevp = &p->replacement;
1864 if (test_bit(In_sync, &rdev->flags) ||
1865 atomic_read(&rdev->nr_pending)) {
1869 /* Only remove faulty devices if recovery
1872 if (!test_bit(Faulty, &rdev->flags) &&
1873 mddev->recovery_disabled != p->recovery_disabled &&
1874 (!p->replacement || p->replacement == rdev) &&
1875 number < conf->geo.raid_disks &&
1882 if (atomic_read(&rdev->nr_pending)) {
1883 /* lost the race, try later */
1887 } else if (p->replacement) {
1888 /* We must have just cleared 'rdev' */
1889 p->rdev = p->replacement;
1890 clear_bit(Replacement, &p->replacement->flags);
1891 smp_mb(); /* Make sure other CPUs may see both as identical
1892 * but will never see neither -- if they are careful.
1894 p->replacement = NULL;
1895 clear_bit(WantReplacement, &rdev->flags);
1897 /* We might have just remove the Replacement as faulty
1898 * Clear the flag just in case
1900 clear_bit(WantReplacement, &rdev->flags);
1902 err = md_integrity_register(mddev);
1911 static void end_sync_read(struct bio *bio, int error)
1913 struct r10bio *r10_bio = bio->bi_private;
1914 struct r10conf *conf = r10_bio->mddev->private;
1917 if (bio == r10_bio->master_bio) {
1918 /* this is a reshape read */
1919 d = r10_bio->read_slot; /* really the read dev */
1921 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1923 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1924 set_bit(R10BIO_Uptodate, &r10_bio->state);
1926 /* The write handler will notice the lack of
1927 * R10BIO_Uptodate and record any errors etc
1929 atomic_add(r10_bio->sectors,
1930 &conf->mirrors[d].rdev->corrected_errors);
1932 /* for reconstruct, we always reschedule after a read.
1933 * for resync, only after all reads
1935 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1936 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1937 atomic_dec_and_test(&r10_bio->remaining)) {
1938 /* we have read all the blocks,
1939 * do the comparison in process context in raid10d
1941 reschedule_retry(r10_bio);
1945 static void end_sync_request(struct r10bio *r10_bio)
1947 struct mddev *mddev = r10_bio->mddev;
1949 while (atomic_dec_and_test(&r10_bio->remaining)) {
1950 if (r10_bio->master_bio == NULL) {
1951 /* the primary of several recovery bios */
1952 sector_t s = r10_bio->sectors;
1953 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1954 test_bit(R10BIO_WriteError, &r10_bio->state))
1955 reschedule_retry(r10_bio);
1958 md_done_sync(mddev, s, 1);
1961 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1962 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1963 test_bit(R10BIO_WriteError, &r10_bio->state))
1964 reschedule_retry(r10_bio);
1972 static void end_sync_write(struct bio *bio, int error)
1974 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1975 struct r10bio *r10_bio = bio->bi_private;
1976 struct mddev *mddev = r10_bio->mddev;
1977 struct r10conf *conf = mddev->private;
1983 struct md_rdev *rdev = NULL;
1985 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1987 rdev = conf->mirrors[d].replacement;
1989 rdev = conf->mirrors[d].rdev;
1993 md_error(mddev, rdev);
1995 set_bit(WriteErrorSeen, &rdev->flags);
1996 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1997 set_bit(MD_RECOVERY_NEEDED,
1998 &rdev->mddev->recovery);
1999 set_bit(R10BIO_WriteError, &r10_bio->state);
2001 } else if (is_badblock(rdev,
2002 r10_bio->devs[slot].addr,
2004 &first_bad, &bad_sectors))
2005 set_bit(R10BIO_MadeGood, &r10_bio->state);
2007 rdev_dec_pending(rdev, mddev);
2009 end_sync_request(r10_bio);
2013 * Note: sync and recover and handled very differently for raid10
2014 * This code is for resync.
2015 * For resync, we read through virtual addresses and read all blocks.
2016 * If there is any error, we schedule a write. The lowest numbered
2017 * drive is authoritative.
2018 * However requests come for physical address, so we need to map.
2019 * For every physical address there are raid_disks/copies virtual addresses,
2020 * which is always are least one, but is not necessarly an integer.
2021 * This means that a physical address can span multiple chunks, so we may
2022 * have to submit multiple io requests for a single sync request.
2025 * We check if all blocks are in-sync and only write to blocks that
2028 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2030 struct r10conf *conf = mddev->private;
2032 struct bio *tbio, *fbio;
2035 atomic_set(&r10_bio->remaining, 1);
2037 /* find the first device with a block */
2038 for (i=0; i<conf->copies; i++)
2039 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2042 if (i == conf->copies)
2046 fbio = r10_bio->devs[i].bio;
2048 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2049 /* now find blocks with errors */
2050 for (i=0 ; i < conf->copies ; i++) {
2053 tbio = r10_bio->devs[i].bio;
2055 if (tbio->bi_end_io != end_sync_read)
2059 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2060 /* We know that the bi_io_vec layout is the same for
2061 * both 'first' and 'i', so we just compare them.
2062 * All vec entries are PAGE_SIZE;
2064 for (j = 0; j < vcnt; j++)
2065 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2066 page_address(tbio->bi_io_vec[j].bv_page),
2067 fbio->bi_io_vec[j].bv_len))
2071 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2072 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2073 /* Don't fix anything. */
2076 /* Ok, we need to write this bio, either to correct an
2077 * inconsistency or to correct an unreadable block.
2078 * First we need to fixup bv_offset, bv_len and
2079 * bi_vecs, as the read request might have corrupted these
2081 tbio->bi_vcnt = vcnt;
2082 tbio->bi_size = r10_bio->sectors << 9;
2084 tbio->bi_phys_segments = 0;
2085 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
2086 tbio->bi_flags |= 1 << BIO_UPTODATE;
2087 tbio->bi_next = NULL;
2088 tbio->bi_rw = WRITE;
2089 tbio->bi_private = r10_bio;
2090 tbio->bi_sector = r10_bio->devs[i].addr;
2092 for (j=0; j < vcnt ; j++) {
2093 tbio->bi_io_vec[j].bv_offset = 0;
2094 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2096 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2097 page_address(fbio->bi_io_vec[j].bv_page),
2100 tbio->bi_end_io = end_sync_write;
2102 d = r10_bio->devs[i].devnum;
2103 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2104 atomic_inc(&r10_bio->remaining);
2105 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2107 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2108 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2109 generic_make_request(tbio);
2112 /* Now write out to any replacement devices
2115 for (i = 0; i < conf->copies; i++) {
2118 tbio = r10_bio->devs[i].repl_bio;
2119 if (!tbio || !tbio->bi_end_io)
2121 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2122 && r10_bio->devs[i].bio != fbio)
2123 for (j = 0; j < vcnt; j++)
2124 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2125 page_address(fbio->bi_io_vec[j].bv_page),
2127 d = r10_bio->devs[i].devnum;
2128 atomic_inc(&r10_bio->remaining);
2129 md_sync_acct(conf->mirrors[d].replacement->bdev,
2131 generic_make_request(tbio);
2135 if (atomic_dec_and_test(&r10_bio->remaining)) {
2136 md_done_sync(mddev, r10_bio->sectors, 1);
2142 * Now for the recovery code.
2143 * Recovery happens across physical sectors.
2144 * We recover all non-is_sync drives by finding the virtual address of
2145 * each, and then choose a working drive that also has that virt address.
2146 * There is a separate r10_bio for each non-in_sync drive.
2147 * Only the first two slots are in use. The first for reading,
2148 * The second for writing.
2151 static void fix_recovery_read_error(struct r10bio *r10_bio)
2153 /* We got a read error during recovery.
2154 * We repeat the read in smaller page-sized sections.
2155 * If a read succeeds, write it to the new device or record
2156 * a bad block if we cannot.
2157 * If a read fails, record a bad block on both old and
2160 struct mddev *mddev = r10_bio->mddev;
2161 struct r10conf *conf = mddev->private;
2162 struct bio *bio = r10_bio->devs[0].bio;
2164 int sectors = r10_bio->sectors;
2166 int dr = r10_bio->devs[0].devnum;
2167 int dw = r10_bio->devs[1].devnum;
2171 struct md_rdev *rdev;
2175 if (s > (PAGE_SIZE>>9))
2178 rdev = conf->mirrors[dr].rdev;
2179 addr = r10_bio->devs[0].addr + sect,
2180 ok = sync_page_io(rdev,
2183 bio->bi_io_vec[idx].bv_page,
2186 rdev = conf->mirrors[dw].rdev;
2187 addr = r10_bio->devs[1].addr + sect;
2188 ok = sync_page_io(rdev,
2191 bio->bi_io_vec[idx].bv_page,
2194 set_bit(WriteErrorSeen, &rdev->flags);
2195 if (!test_and_set_bit(WantReplacement,
2197 set_bit(MD_RECOVERY_NEEDED,
2198 &rdev->mddev->recovery);
2202 /* We don't worry if we cannot set a bad block -
2203 * it really is bad so there is no loss in not
2206 rdev_set_badblocks(rdev, addr, s, 0);
2208 if (rdev != conf->mirrors[dw].rdev) {
2209 /* need bad block on destination too */
2210 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2211 addr = r10_bio->devs[1].addr + sect;
2212 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2214 /* just abort the recovery */
2216 "md/raid10:%s: recovery aborted"
2217 " due to read error\n",
2220 conf->mirrors[dw].recovery_disabled
2221 = mddev->recovery_disabled;
2222 set_bit(MD_RECOVERY_INTR,
2235 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2237 struct r10conf *conf = mddev->private;
2239 struct bio *wbio, *wbio2;
2241 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2242 fix_recovery_read_error(r10_bio);
2243 end_sync_request(r10_bio);
2248 * share the pages with the first bio
2249 * and submit the write request
2251 d = r10_bio->devs[1].devnum;
2252 wbio = r10_bio->devs[1].bio;
2253 wbio2 = r10_bio->devs[1].repl_bio;
2254 if (wbio->bi_end_io) {
2255 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2256 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2257 generic_make_request(wbio);
2259 if (wbio2 && wbio2->bi_end_io) {
2260 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2261 md_sync_acct(conf->mirrors[d].replacement->bdev,
2262 bio_sectors(wbio2));
2263 generic_make_request(wbio2);
2269 * Used by fix_read_error() to decay the per rdev read_errors.
2270 * We halve the read error count for every hour that has elapsed
2271 * since the last recorded read error.
2274 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2276 struct timespec cur_time_mon;
2277 unsigned long hours_since_last;
2278 unsigned int read_errors = atomic_read(&rdev->read_errors);
2280 ktime_get_ts(&cur_time_mon);
2282 if (rdev->last_read_error.tv_sec == 0 &&
2283 rdev->last_read_error.tv_nsec == 0) {
2284 /* first time we've seen a read error */
2285 rdev->last_read_error = cur_time_mon;
2289 hours_since_last = (cur_time_mon.tv_sec -
2290 rdev->last_read_error.tv_sec) / 3600;
2292 rdev->last_read_error = cur_time_mon;
2295 * if hours_since_last is > the number of bits in read_errors
2296 * just set read errors to 0. We do this to avoid
2297 * overflowing the shift of read_errors by hours_since_last.
2299 if (hours_since_last >= 8 * sizeof(read_errors))
2300 atomic_set(&rdev->read_errors, 0);
2302 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2305 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2306 int sectors, struct page *page, int rw)
2311 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2312 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2314 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2318 set_bit(WriteErrorSeen, &rdev->flags);
2319 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2320 set_bit(MD_RECOVERY_NEEDED,
2321 &rdev->mddev->recovery);
2323 /* need to record an error - either for the block or the device */
2324 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2325 md_error(rdev->mddev, rdev);
2330 * This is a kernel thread which:
2332 * 1. Retries failed read operations on working mirrors.
2333 * 2. Updates the raid superblock when problems encounter.
2334 * 3. Performs writes following reads for array synchronising.
2337 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2339 int sect = 0; /* Offset from r10_bio->sector */
2340 int sectors = r10_bio->sectors;
2341 struct md_rdev*rdev;
2342 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2343 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2345 /* still own a reference to this rdev, so it cannot
2346 * have been cleared recently.
2348 rdev = conf->mirrors[d].rdev;
2350 if (test_bit(Faulty, &rdev->flags))
2351 /* drive has already been failed, just ignore any
2352 more fix_read_error() attempts */
2355 check_decay_read_errors(mddev, rdev);
2356 atomic_inc(&rdev->read_errors);
2357 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2358 char b[BDEVNAME_SIZE];
2359 bdevname(rdev->bdev, b);
2362 "md/raid10:%s: %s: Raid device exceeded "
2363 "read_error threshold [cur %d:max %d]\n",
2365 atomic_read(&rdev->read_errors), max_read_errors);
2367 "md/raid10:%s: %s: Failing raid device\n",
2369 md_error(mddev, conf->mirrors[d].rdev);
2370 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2376 int sl = r10_bio->read_slot;
2380 if (s > (PAGE_SIZE>>9))
2388 d = r10_bio->devs[sl].devnum;
2389 rdev = rcu_dereference(conf->mirrors[d].rdev);
2391 !test_bit(Unmerged, &rdev->flags) &&
2392 test_bit(In_sync, &rdev->flags) &&
2393 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2394 &first_bad, &bad_sectors) == 0) {
2395 atomic_inc(&rdev->nr_pending);
2397 success = sync_page_io(rdev,
2398 r10_bio->devs[sl].addr +
2401 conf->tmppage, READ, false);
2402 rdev_dec_pending(rdev, mddev);
2408 if (sl == conf->copies)
2410 } while (!success && sl != r10_bio->read_slot);
2414 /* Cannot read from anywhere, just mark the block
2415 * as bad on the first device to discourage future
2418 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2419 rdev = conf->mirrors[dn].rdev;
2421 if (!rdev_set_badblocks(
2423 r10_bio->devs[r10_bio->read_slot].addr
2426 md_error(mddev, rdev);
2427 r10_bio->devs[r10_bio->read_slot].bio
2434 /* write it back and re-read */
2436 while (sl != r10_bio->read_slot) {
2437 char b[BDEVNAME_SIZE];
2442 d = r10_bio->devs[sl].devnum;
2443 rdev = rcu_dereference(conf->mirrors[d].rdev);
2445 test_bit(Unmerged, &rdev->flags) ||
2446 !test_bit(In_sync, &rdev->flags))
2449 atomic_inc(&rdev->nr_pending);
2451 if (r10_sync_page_io(rdev,
2452 r10_bio->devs[sl].addr +
2454 s, conf->tmppage, WRITE)
2456 /* Well, this device is dead */
2458 "md/raid10:%s: read correction "
2460 " (%d sectors at %llu on %s)\n",
2462 (unsigned long long)(
2464 choose_data_offset(r10_bio,
2466 bdevname(rdev->bdev, b));
2467 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2470 bdevname(rdev->bdev, b));
2472 rdev_dec_pending(rdev, mddev);
2476 while (sl != r10_bio->read_slot) {
2477 char b[BDEVNAME_SIZE];
2482 d = r10_bio->devs[sl].devnum;
2483 rdev = rcu_dereference(conf->mirrors[d].rdev);
2485 !test_bit(In_sync, &rdev->flags))
2488 atomic_inc(&rdev->nr_pending);
2490 switch (r10_sync_page_io(rdev,
2491 r10_bio->devs[sl].addr +
2496 /* Well, this device is dead */
2498 "md/raid10:%s: unable to read back "
2500 " (%d sectors at %llu on %s)\n",
2502 (unsigned long long)(
2504 choose_data_offset(r10_bio, rdev)),
2505 bdevname(rdev->bdev, b));
2506 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2509 bdevname(rdev->bdev, b));
2513 "md/raid10:%s: read error corrected"
2514 " (%d sectors at %llu on %s)\n",
2516 (unsigned long long)(
2518 choose_data_offset(r10_bio, rdev)),
2519 bdevname(rdev->bdev, b));
2520 atomic_add(s, &rdev->corrected_errors);
2523 rdev_dec_pending(rdev, mddev);
2533 static void bi_complete(struct bio *bio, int error)
2535 complete((struct completion *)bio->bi_private);
2538 static int submit_bio_wait(int rw, struct bio *bio)
2540 struct completion event;
2543 init_completion(&event);
2544 bio->bi_private = &event;
2545 bio->bi_end_io = bi_complete;
2546 submit_bio(rw, bio);
2547 wait_for_completion(&event);
2549 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2552 static int narrow_write_error(struct r10bio *r10_bio, int i)
2554 struct bio *bio = r10_bio->master_bio;
2555 struct mddev *mddev = r10_bio->mddev;
2556 struct r10conf *conf = mddev->private;
2557 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2558 /* bio has the data to be written to slot 'i' where
2559 * we just recently had a write error.
2560 * We repeatedly clone the bio and trim down to one block,
2561 * then try the write. Where the write fails we record
2563 * It is conceivable that the bio doesn't exactly align with
2564 * blocks. We must handle this.
2566 * We currently own a reference to the rdev.
2572 int sect_to_write = r10_bio->sectors;
2575 if (rdev->badblocks.shift < 0)
2578 block_sectors = 1 << rdev->badblocks.shift;
2579 sector = r10_bio->sector;
2580 sectors = ((r10_bio->sector + block_sectors)
2581 & ~(sector_t)(block_sectors - 1))
2584 while (sect_to_write) {
2586 if (sectors > sect_to_write)
2587 sectors = sect_to_write;
2588 /* Write at 'sector' for 'sectors' */
2589 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2590 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2591 wbio->bi_sector = (r10_bio->devs[i].addr+
2592 choose_data_offset(r10_bio, rdev) +
2593 (sector - r10_bio->sector));
2594 wbio->bi_bdev = rdev->bdev;
2595 if (submit_bio_wait(WRITE, wbio) == 0)
2597 ok = rdev_set_badblocks(rdev, sector,
2602 sect_to_write -= sectors;
2604 sectors = block_sectors;
2609 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2611 int slot = r10_bio->read_slot;
2613 struct r10conf *conf = mddev->private;
2614 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2615 char b[BDEVNAME_SIZE];
2616 unsigned long do_sync;
2619 /* we got a read error. Maybe the drive is bad. Maybe just
2620 * the block and we can fix it.
2621 * We freeze all other IO, and try reading the block from
2622 * other devices. When we find one, we re-write
2623 * and check it that fixes the read error.
2624 * This is all done synchronously while the array is
2627 bio = r10_bio->devs[slot].bio;
2628 bdevname(bio->bi_bdev, b);
2630 r10_bio->devs[slot].bio = NULL;
2632 if (mddev->ro == 0) {
2634 fix_read_error(conf, mddev, r10_bio);
2635 unfreeze_array(conf);
2637 r10_bio->devs[slot].bio = IO_BLOCKED;
2639 rdev_dec_pending(rdev, mddev);
2642 rdev = read_balance(conf, r10_bio, &max_sectors);
2644 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2645 " read error for block %llu\n",
2647 (unsigned long long)r10_bio->sector);
2648 raid_end_bio_io(r10_bio);
2652 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2653 slot = r10_bio->read_slot;
2656 "md/raid10:%s: %s: redirecting "
2657 "sector %llu to another mirror\n",
2659 bdevname(rdev->bdev, b),
2660 (unsigned long long)r10_bio->sector);
2661 bio = bio_clone_mddev(r10_bio->master_bio,
2664 r10_bio->sector - bio->bi_sector,
2666 r10_bio->devs[slot].bio = bio;
2667 r10_bio->devs[slot].rdev = rdev;
2668 bio->bi_sector = r10_bio->devs[slot].addr
2669 + choose_data_offset(r10_bio, rdev);
2670 bio->bi_bdev = rdev->bdev;
2671 bio->bi_rw = READ | do_sync;
2672 bio->bi_private = r10_bio;
2673 bio->bi_end_io = raid10_end_read_request;
2674 if (max_sectors < r10_bio->sectors) {
2675 /* Drat - have to split this up more */
2676 struct bio *mbio = r10_bio->master_bio;
2677 int sectors_handled =
2678 r10_bio->sector + max_sectors
2680 r10_bio->sectors = max_sectors;
2681 spin_lock_irq(&conf->device_lock);
2682 if (mbio->bi_phys_segments == 0)
2683 mbio->bi_phys_segments = 2;
2685 mbio->bi_phys_segments++;
2686 spin_unlock_irq(&conf->device_lock);
2687 generic_make_request(bio);
2689 r10_bio = mempool_alloc(conf->r10bio_pool,
2691 r10_bio->master_bio = mbio;
2692 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2694 set_bit(R10BIO_ReadError,
2696 r10_bio->mddev = mddev;
2697 r10_bio->sector = mbio->bi_sector
2702 generic_make_request(bio);
2705 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2707 /* Some sort of write request has finished and it
2708 * succeeded in writing where we thought there was a
2709 * bad block. So forget the bad block.
2710 * Or possibly if failed and we need to record
2714 struct md_rdev *rdev;
2716 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2717 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2718 for (m = 0; m < conf->copies; m++) {
2719 int dev = r10_bio->devs[m].devnum;
2720 rdev = conf->mirrors[dev].rdev;
2721 if (r10_bio->devs[m].bio == NULL)
2723 if (test_bit(BIO_UPTODATE,
2724 &r10_bio->devs[m].bio->bi_flags)) {
2725 rdev_clear_badblocks(
2727 r10_bio->devs[m].addr,
2728 r10_bio->sectors, 0);
2730 if (!rdev_set_badblocks(
2732 r10_bio->devs[m].addr,
2733 r10_bio->sectors, 0))
2734 md_error(conf->mddev, rdev);
2736 rdev = conf->mirrors[dev].replacement;
2737 if (r10_bio->devs[m].repl_bio == NULL)
2739 if (test_bit(BIO_UPTODATE,
2740 &r10_bio->devs[m].repl_bio->bi_flags)) {
2741 rdev_clear_badblocks(
2743 r10_bio->devs[m].addr,
2744 r10_bio->sectors, 0);
2746 if (!rdev_set_badblocks(
2748 r10_bio->devs[m].addr,
2749 r10_bio->sectors, 0))
2750 md_error(conf->mddev, rdev);
2755 for (m = 0; m < conf->copies; m++) {
2756 int dev = r10_bio->devs[m].devnum;
2757 struct bio *bio = r10_bio->devs[m].bio;
2758 rdev = conf->mirrors[dev].rdev;
2759 if (bio == IO_MADE_GOOD) {
2760 rdev_clear_badblocks(
2762 r10_bio->devs[m].addr,
2763 r10_bio->sectors, 0);
2764 rdev_dec_pending(rdev, conf->mddev);
2765 } else if (bio != NULL &&
2766 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2767 if (!narrow_write_error(r10_bio, m)) {
2768 md_error(conf->mddev, rdev);
2769 set_bit(R10BIO_Degraded,
2772 rdev_dec_pending(rdev, conf->mddev);
2774 bio = r10_bio->devs[m].repl_bio;
2775 rdev = conf->mirrors[dev].replacement;
2776 if (rdev && bio == IO_MADE_GOOD) {
2777 rdev_clear_badblocks(
2779 r10_bio->devs[m].addr,
2780 r10_bio->sectors, 0);
2781 rdev_dec_pending(rdev, conf->mddev);
2784 if (test_bit(R10BIO_WriteError,
2786 close_write(r10_bio);
2787 raid_end_bio_io(r10_bio);
2791 static void raid10d(struct md_thread *thread)
2793 struct mddev *mddev = thread->mddev;
2794 struct r10bio *r10_bio;
2795 unsigned long flags;
2796 struct r10conf *conf = mddev->private;
2797 struct list_head *head = &conf->retry_list;
2798 struct blk_plug plug;
2800 md_check_recovery(mddev);
2802 blk_start_plug(&plug);
2805 flush_pending_writes(conf);
2807 spin_lock_irqsave(&conf->device_lock, flags);
2808 if (list_empty(head)) {
2809 spin_unlock_irqrestore(&conf->device_lock, flags);
2812 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2813 list_del(head->prev);
2815 spin_unlock_irqrestore(&conf->device_lock, flags);
2817 mddev = r10_bio->mddev;
2818 conf = mddev->private;
2819 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2820 test_bit(R10BIO_WriteError, &r10_bio->state))
2821 handle_write_completed(conf, r10_bio);
2822 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2823 reshape_request_write(mddev, r10_bio);
2824 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2825 sync_request_write(mddev, r10_bio);
2826 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2827 recovery_request_write(mddev, r10_bio);
2828 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2829 handle_read_error(mddev, r10_bio);
2831 /* just a partial read to be scheduled from a
2834 int slot = r10_bio->read_slot;
2835 generic_make_request(r10_bio->devs[slot].bio);
2839 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2840 md_check_recovery(mddev);
2842 blk_finish_plug(&plug);
2846 static int init_resync(struct r10conf *conf)
2851 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2852 BUG_ON(conf->r10buf_pool);
2853 conf->have_replacement = 0;
2854 for (i = 0; i < conf->geo.raid_disks; i++)
2855 if (conf->mirrors[i].replacement)
2856 conf->have_replacement = 1;
2857 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2858 if (!conf->r10buf_pool)
2860 conf->next_resync = 0;
2865 * perform a "sync" on one "block"
2867 * We need to make sure that no normal I/O request - particularly write
2868 * requests - conflict with active sync requests.
2870 * This is achieved by tracking pending requests and a 'barrier' concept
2871 * that can be installed to exclude normal IO requests.
2873 * Resync and recovery are handled very differently.
2874 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2876 * For resync, we iterate over virtual addresses, read all copies,
2877 * and update if there are differences. If only one copy is live,
2879 * For recovery, we iterate over physical addresses, read a good
2880 * value for each non-in_sync drive, and over-write.
2882 * So, for recovery we may have several outstanding complex requests for a
2883 * given address, one for each out-of-sync device. We model this by allocating
2884 * a number of r10_bio structures, one for each out-of-sync device.
2885 * As we setup these structures, we collect all bio's together into a list
2886 * which we then process collectively to add pages, and then process again
2887 * to pass to generic_make_request.
2889 * The r10_bio structures are linked using a borrowed master_bio pointer.
2890 * This link is counted in ->remaining. When the r10_bio that points to NULL
2891 * has its remaining count decremented to 0, the whole complex operation
2896 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2897 int *skipped, int go_faster)
2899 struct r10conf *conf = mddev->private;
2900 struct r10bio *r10_bio;
2901 struct bio *biolist = NULL, *bio;
2902 sector_t max_sector, nr_sectors;
2905 sector_t sync_blocks;
2906 sector_t sectors_skipped = 0;
2907 int chunks_skipped = 0;
2908 sector_t chunk_mask = conf->geo.chunk_mask;
2910 if (!conf->r10buf_pool)
2911 if (init_resync(conf))
2915 max_sector = mddev->dev_sectors;
2916 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2917 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2918 max_sector = mddev->resync_max_sectors;
2919 if (sector_nr >= max_sector) {
2920 /* If we aborted, we need to abort the
2921 * sync on the 'current' bitmap chucks (there can
2922 * be several when recovering multiple devices).
2923 * as we may have started syncing it but not finished.
2924 * We can find the current address in
2925 * mddev->curr_resync, but for recovery,
2926 * we need to convert that to several
2927 * virtual addresses.
2929 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2934 if (mddev->curr_resync < max_sector) { /* aborted */
2935 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2936 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2938 else for (i = 0; i < conf->geo.raid_disks; i++) {
2940 raid10_find_virt(conf, mddev->curr_resync, i);
2941 bitmap_end_sync(mddev->bitmap, sect,
2945 /* completed sync */
2946 if ((!mddev->bitmap || conf->fullsync)
2947 && conf->have_replacement
2948 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2949 /* Completed a full sync so the replacements
2950 * are now fully recovered.
2952 for (i = 0; i < conf->geo.raid_disks; i++)
2953 if (conf->mirrors[i].replacement)
2954 conf->mirrors[i].replacement
2960 bitmap_close_sync(mddev->bitmap);
2963 return sectors_skipped;
2966 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2967 return reshape_request(mddev, sector_nr, skipped);
2969 if (chunks_skipped >= conf->geo.raid_disks) {
2970 /* if there has been nothing to do on any drive,
2971 * then there is nothing to do at all..
2974 return (max_sector - sector_nr) + sectors_skipped;
2977 if (max_sector > mddev->resync_max)
2978 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2980 /* make sure whole request will fit in a chunk - if chunks
2983 if (conf->geo.near_copies < conf->geo.raid_disks &&
2984 max_sector > (sector_nr | chunk_mask))
2985 max_sector = (sector_nr | chunk_mask) + 1;
2987 * If there is non-resync activity waiting for us then
2988 * put in a delay to throttle resync.
2990 if (!go_faster && conf->nr_waiting)
2991 msleep_interruptible(1000);
2993 /* Again, very different code for resync and recovery.
2994 * Both must result in an r10bio with a list of bios that
2995 * have bi_end_io, bi_sector, bi_bdev set,
2996 * and bi_private set to the r10bio.
2997 * For recovery, we may actually create several r10bios
2998 * with 2 bios in each, that correspond to the bios in the main one.
2999 * In this case, the subordinate r10bios link back through a
3000 * borrowed master_bio pointer, and the counter in the master
3001 * includes a ref from each subordinate.
3003 /* First, we decide what to do and set ->bi_end_io
3004 * To end_sync_read if we want to read, and
3005 * end_sync_write if we will want to write.
3008 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3009 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3010 /* recovery... the complicated one */
3014 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3020 struct raid10_info *mirror = &conf->mirrors[i];
3022 if ((mirror->rdev == NULL ||
3023 test_bit(In_sync, &mirror->rdev->flags))
3025 (mirror->replacement == NULL ||
3027 &mirror->replacement->flags)))
3031 /* want to reconstruct this device */
3033 sect = raid10_find_virt(conf, sector_nr, i);
3034 if (sect >= mddev->resync_max_sectors) {
3035 /* last stripe is not complete - don't
3036 * try to recover this sector.
3040 /* Unless we are doing a full sync, or a replacement
3041 * we only need to recover the block if it is set in
3044 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3046 if (sync_blocks < max_sync)
3047 max_sync = sync_blocks;
3049 mirror->replacement == NULL &&
3051 /* yep, skip the sync_blocks here, but don't assume
3052 * that there will never be anything to do here
3054 chunks_skipped = -1;
3058 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3059 raise_barrier(conf, rb2 != NULL);
3060 atomic_set(&r10_bio->remaining, 0);
3062 r10_bio->master_bio = (struct bio*)rb2;
3064 atomic_inc(&rb2->remaining);
3065 r10_bio->mddev = mddev;
3066 set_bit(R10BIO_IsRecover, &r10_bio->state);
3067 r10_bio->sector = sect;
3069 raid10_find_phys(conf, r10_bio);
3071 /* Need to check if the array will still be
3074 for (j = 0; j < conf->geo.raid_disks; j++)
3075 if (conf->mirrors[j].rdev == NULL ||
3076 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3081 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3082 &sync_blocks, still_degraded);
3085 for (j=0; j<conf->copies;j++) {
3087 int d = r10_bio->devs[j].devnum;
3088 sector_t from_addr, to_addr;
3089 struct md_rdev *rdev;
3090 sector_t sector, first_bad;
3092 if (!conf->mirrors[d].rdev ||
3093 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3095 /* This is where we read from */
3097 rdev = conf->mirrors[d].rdev;
3098 sector = r10_bio->devs[j].addr;
3100 if (is_badblock(rdev, sector, max_sync,
3101 &first_bad, &bad_sectors)) {
3102 if (first_bad > sector)
3103 max_sync = first_bad - sector;
3105 bad_sectors -= (sector
3107 if (max_sync > bad_sectors)
3108 max_sync = bad_sectors;
3112 bio = r10_bio->devs[0].bio;
3113 bio->bi_next = biolist;
3115 bio->bi_private = r10_bio;
3116 bio->bi_end_io = end_sync_read;
3118 from_addr = r10_bio->devs[j].addr;
3119 bio->bi_sector = from_addr + rdev->data_offset;
3120 bio->bi_bdev = rdev->bdev;
3121 atomic_inc(&rdev->nr_pending);
3122 /* and we write to 'i' (if not in_sync) */
3124 for (k=0; k<conf->copies; k++)
3125 if (r10_bio->devs[k].devnum == i)
3127 BUG_ON(k == conf->copies);
3128 to_addr = r10_bio->devs[k].addr;
3129 r10_bio->devs[0].devnum = d;
3130 r10_bio->devs[0].addr = from_addr;
3131 r10_bio->devs[1].devnum = i;
3132 r10_bio->devs[1].addr = to_addr;
3134 rdev = mirror->rdev;
3135 if (!test_bit(In_sync, &rdev->flags)) {
3136 bio = r10_bio->devs[1].bio;
3137 bio->bi_next = biolist;
3139 bio->bi_private = r10_bio;
3140 bio->bi_end_io = end_sync_write;
3142 bio->bi_sector = to_addr
3143 + rdev->data_offset;
3144 bio->bi_bdev = rdev->bdev;
3145 atomic_inc(&r10_bio->remaining);
3147 r10_bio->devs[1].bio->bi_end_io = NULL;
3149 /* and maybe write to replacement */
3150 bio = r10_bio->devs[1].repl_bio;
3152 bio->bi_end_io = NULL;
3153 rdev = mirror->replacement;
3154 /* Note: if rdev != NULL, then bio
3155 * cannot be NULL as r10buf_pool_alloc will
3156 * have allocated it.
3157 * So the second test here is pointless.
3158 * But it keeps semantic-checkers happy, and
3159 * this comment keeps human reviewers
3162 if (rdev == NULL || bio == NULL ||
3163 test_bit(Faulty, &rdev->flags))
3165 bio->bi_next = biolist;
3167 bio->bi_private = r10_bio;
3168 bio->bi_end_io = end_sync_write;
3170 bio->bi_sector = to_addr + rdev->data_offset;
3171 bio->bi_bdev = rdev->bdev;
3172 atomic_inc(&r10_bio->remaining);
3175 if (j == conf->copies) {
3176 /* Cannot recover, so abort the recovery or
3177 * record a bad block */
3180 atomic_dec(&rb2->remaining);
3183 /* problem is that there are bad blocks
3184 * on other device(s)
3187 for (k = 0; k < conf->copies; k++)
3188 if (r10_bio->devs[k].devnum == i)
3190 if (!test_bit(In_sync,
3191 &mirror->rdev->flags)
3192 && !rdev_set_badblocks(
3194 r10_bio->devs[k].addr,
3197 if (mirror->replacement &&
3198 !rdev_set_badblocks(
3199 mirror->replacement,
3200 r10_bio->devs[k].addr,
3205 if (!test_and_set_bit(MD_RECOVERY_INTR,
3207 printk(KERN_INFO "md/raid10:%s: insufficient "
3208 "working devices for recovery.\n",
3210 mirror->recovery_disabled
3211 = mddev->recovery_disabled;
3216 if (biolist == NULL) {
3218 struct r10bio *rb2 = r10_bio;
3219 r10_bio = (struct r10bio*) rb2->master_bio;
3220 rb2->master_bio = NULL;
3226 /* resync. Schedule a read for every block at this virt offset */
3229 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3231 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3232 &sync_blocks, mddev->degraded) &&
3233 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3234 &mddev->recovery)) {
3235 /* We can skip this block */
3237 return sync_blocks + sectors_skipped;
3239 if (sync_blocks < max_sync)
3240 max_sync = sync_blocks;
3241 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3243 r10_bio->mddev = mddev;
3244 atomic_set(&r10_bio->remaining, 0);
3245 raise_barrier(conf, 0);
3246 conf->next_resync = sector_nr;
3248 r10_bio->master_bio = NULL;
3249 r10_bio->sector = sector_nr;
3250 set_bit(R10BIO_IsSync, &r10_bio->state);
3251 raid10_find_phys(conf, r10_bio);
3252 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3254 for (i = 0; i < conf->copies; i++) {
3255 int d = r10_bio->devs[i].devnum;
3256 sector_t first_bad, sector;
3259 if (r10_bio->devs[i].repl_bio)
3260 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3262 bio = r10_bio->devs[i].bio;
3263 bio->bi_end_io = NULL;
3264 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3265 if (conf->mirrors[d].rdev == NULL ||
3266 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3268 sector = r10_bio->devs[i].addr;
3269 if (is_badblock(conf->mirrors[d].rdev,
3271 &first_bad, &bad_sectors)) {
3272 if (first_bad > sector)
3273 max_sync = first_bad - sector;
3275 bad_sectors -= (sector - first_bad);
3276 if (max_sync > bad_sectors)
3277 max_sync = bad_sectors;
3281 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3282 atomic_inc(&r10_bio->remaining);
3283 bio->bi_next = biolist;
3285 bio->bi_private = r10_bio;
3286 bio->bi_end_io = end_sync_read;
3288 bio->bi_sector = sector +
3289 conf->mirrors[d].rdev->data_offset;
3290 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3293 if (conf->mirrors[d].replacement == NULL ||
3295 &conf->mirrors[d].replacement->flags))
3298 /* Need to set up for writing to the replacement */
3299 bio = r10_bio->devs[i].repl_bio;
3300 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3302 sector = r10_bio->devs[i].addr;
3303 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3304 bio->bi_next = biolist;
3306 bio->bi_private = r10_bio;
3307 bio->bi_end_io = end_sync_write;
3309 bio->bi_sector = sector +
3310 conf->mirrors[d].replacement->data_offset;
3311 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3316 for (i=0; i<conf->copies; i++) {
3317 int d = r10_bio->devs[i].devnum;
3318 if (r10_bio->devs[i].bio->bi_end_io)
3319 rdev_dec_pending(conf->mirrors[d].rdev,
3321 if (r10_bio->devs[i].repl_bio &&
3322 r10_bio->devs[i].repl_bio->bi_end_io)
3324 conf->mirrors[d].replacement,
3333 for (bio = biolist; bio ; bio=bio->bi_next) {
3335 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3337 bio->bi_flags |= 1 << BIO_UPTODATE;
3340 bio->bi_phys_segments = 0;
3345 if (sector_nr + max_sync < max_sector)
3346 max_sector = sector_nr + max_sync;
3349 int len = PAGE_SIZE;
3350 if (sector_nr + (len>>9) > max_sector)
3351 len = (max_sector - sector_nr) << 9;
3354 for (bio= biolist ; bio ; bio=bio->bi_next) {
3356 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3357 if (bio_add_page(bio, page, len, 0))
3361 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3362 for (bio2 = biolist;
3363 bio2 && bio2 != bio;
3364 bio2 = bio2->bi_next) {
3365 /* remove last page from this bio */
3367 bio2->bi_size -= len;
3368 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3372 nr_sectors += len>>9;
3373 sector_nr += len>>9;
3374 } while (biolist->bi_vcnt < RESYNC_PAGES);
3376 r10_bio->sectors = nr_sectors;
3380 biolist = biolist->bi_next;
3382 bio->bi_next = NULL;
3383 r10_bio = bio->bi_private;
3384 r10_bio->sectors = nr_sectors;
3386 if (bio->bi_end_io == end_sync_read) {
3387 md_sync_acct(bio->bi_bdev, nr_sectors);
3388 generic_make_request(bio);
3392 if (sectors_skipped)
3393 /* pretend they weren't skipped, it makes
3394 * no important difference in this case
3396 md_done_sync(mddev, sectors_skipped, 1);
3398 return sectors_skipped + nr_sectors;
3400 /* There is nowhere to write, so all non-sync
3401 * drives must be failed or in resync, all drives
3402 * have a bad block, so try the next chunk...
3404 if (sector_nr + max_sync < max_sector)
3405 max_sector = sector_nr + max_sync;
3407 sectors_skipped += (max_sector - sector_nr);
3409 sector_nr = max_sector;
3414 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3417 struct r10conf *conf = mddev->private;
3420 raid_disks = min(conf->geo.raid_disks,
3421 conf->prev.raid_disks);
3423 sectors = conf->dev_sectors;
3425 size = sectors >> conf->geo.chunk_shift;
3426 sector_div(size, conf->geo.far_copies);
3427 size = size * raid_disks;
3428 sector_div(size, conf->geo.near_copies);
3430 return size << conf->geo.chunk_shift;
3433 static void calc_sectors(struct r10conf *conf, sector_t size)
3435 /* Calculate the number of sectors-per-device that will
3436 * actually be used, and set conf->dev_sectors and
3440 size = size >> conf->geo.chunk_shift;
3441 sector_div(size, conf->geo.far_copies);
3442 size = size * conf->geo.raid_disks;
3443 sector_div(size, conf->geo.near_copies);
3444 /* 'size' is now the number of chunks in the array */
3445 /* calculate "used chunks per device" */
3446 size = size * conf->copies;
3448 /* We need to round up when dividing by raid_disks to
3449 * get the stride size.
3451 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3453 conf->dev_sectors = size << conf->geo.chunk_shift;
3455 if (conf->geo.far_offset)
3456 conf->geo.stride = 1 << conf->geo.chunk_shift;
3458 sector_div(size, conf->geo.far_copies);
3459 conf->geo.stride = size << conf->geo.chunk_shift;
3463 enum geo_type {geo_new, geo_old, geo_start};
3464 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3467 int layout, chunk, disks;
3470 layout = mddev->layout;
3471 chunk = mddev->chunk_sectors;
3472 disks = mddev->raid_disks - mddev->delta_disks;
3475 layout = mddev->new_layout;
3476 chunk = mddev->new_chunk_sectors;
3477 disks = mddev->raid_disks;
3479 default: /* avoid 'may be unused' warnings */
3480 case geo_start: /* new when starting reshape - raid_disks not
3482 layout = mddev->new_layout;
3483 chunk = mddev->new_chunk_sectors;
3484 disks = mddev->raid_disks + mddev->delta_disks;
3489 if (chunk < (PAGE_SIZE >> 9) ||
3490 !is_power_of_2(chunk))
3493 fc = (layout >> 8) & 255;
3494 fo = layout & (1<<16);
3495 geo->raid_disks = disks;
3496 geo->near_copies = nc;
3497 geo->far_copies = fc;
3498 geo->far_offset = fo;
3499 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3500 geo->chunk_mask = chunk - 1;
3501 geo->chunk_shift = ffz(~chunk);
3505 static struct r10conf *setup_conf(struct mddev *mddev)
3507 struct r10conf *conf = NULL;
3512 copies = setup_geo(&geo, mddev, geo_new);
3515 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3516 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3517 mdname(mddev), PAGE_SIZE);
3521 if (copies < 2 || copies > mddev->raid_disks) {
3522 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3523 mdname(mddev), mddev->new_layout);
3528 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3532 /* FIXME calc properly */
3533 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3534 max(0,mddev->delta_disks)),
3539 conf->tmppage = alloc_page(GFP_KERNEL);
3544 conf->copies = copies;
3545 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3546 r10bio_pool_free, conf);
3547 if (!conf->r10bio_pool)
3550 calc_sectors(conf, mddev->dev_sectors);
3551 if (mddev->reshape_position == MaxSector) {
3552 conf->prev = conf->geo;
3553 conf->reshape_progress = MaxSector;
3555 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3559 conf->reshape_progress = mddev->reshape_position;
3560 if (conf->prev.far_offset)
3561 conf->prev.stride = 1 << conf->prev.chunk_shift;
3563 /* far_copies must be 1 */
3564 conf->prev.stride = conf->dev_sectors;
3566 spin_lock_init(&conf->device_lock);
3567 INIT_LIST_HEAD(&conf->retry_list);
3569 spin_lock_init(&conf->resync_lock);
3570 init_waitqueue_head(&conf->wait_barrier);
3572 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3576 conf->mddev = mddev;
3581 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3584 if (conf->r10bio_pool)
3585 mempool_destroy(conf->r10bio_pool);
3586 kfree(conf->mirrors);
3587 safe_put_page(conf->tmppage);
3590 return ERR_PTR(err);
3593 static int run(struct mddev *mddev)
3595 struct r10conf *conf;
3596 int i, disk_idx, chunk_size;
3597 struct raid10_info *disk;
3598 struct md_rdev *rdev;
3600 sector_t min_offset_diff = 0;
3602 bool discard_supported = false;
3604 if (mddev->private == NULL) {
3605 conf = setup_conf(mddev);
3607 return PTR_ERR(conf);
3608 mddev->private = conf;
3610 conf = mddev->private;
3614 mddev->thread = conf->thread;
3615 conf->thread = NULL;
3617 chunk_size = mddev->chunk_sectors << 9;
3619 blk_queue_max_discard_sectors(mddev->queue,
3620 mddev->chunk_sectors);
3621 blk_queue_max_write_same_sectors(mddev->queue,
3622 mddev->chunk_sectors);
3623 blk_queue_io_min(mddev->queue, chunk_size);
3624 if (conf->geo.raid_disks % conf->geo.near_copies)
3625 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3627 blk_queue_io_opt(mddev->queue, chunk_size *
3628 (conf->geo.raid_disks / conf->geo.near_copies));
3631 rdev_for_each(rdev, mddev) {
3633 struct request_queue *q;
3635 disk_idx = rdev->raid_disk;
3638 if (disk_idx >= conf->geo.raid_disks &&
3639 disk_idx >= conf->prev.raid_disks)
3641 disk = conf->mirrors + disk_idx;
3643 if (test_bit(Replacement, &rdev->flags)) {
3644 if (disk->replacement)
3646 disk->replacement = rdev;
3652 q = bdev_get_queue(rdev->bdev);
3653 if (q->merge_bvec_fn)
3654 mddev->merge_check_needed = 1;
3655 diff = (rdev->new_data_offset - rdev->data_offset);
3656 if (!mddev->reshape_backwards)
3660 if (first || diff < min_offset_diff)
3661 min_offset_diff = diff;
3664 disk_stack_limits(mddev->gendisk, rdev->bdev,
3665 rdev->data_offset << 9);
3667 disk->head_position = 0;
3669 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3670 discard_supported = true;
3674 if (discard_supported)
3675 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3678 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3681 /* need to check that every block has at least one working mirror */
3682 if (!enough(conf, -1)) {
3683 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3688 if (conf->reshape_progress != MaxSector) {
3689 /* must ensure that shape change is supported */
3690 if (conf->geo.far_copies != 1 &&
3691 conf->geo.far_offset == 0)
3693 if (conf->prev.far_copies != 1 &&
3694 conf->geo.far_offset == 0)
3698 mddev->degraded = 0;
3700 i < conf->geo.raid_disks
3701 || i < conf->prev.raid_disks;
3704 disk = conf->mirrors + i;
3706 if (!disk->rdev && disk->replacement) {
3707 /* The replacement is all we have - use it */
3708 disk->rdev = disk->replacement;
3709 disk->replacement = NULL;
3710 clear_bit(Replacement, &disk->rdev->flags);
3714 !test_bit(In_sync, &disk->rdev->flags)) {
3715 disk->head_position = 0;
3720 disk->recovery_disabled = mddev->recovery_disabled - 1;
3723 if (mddev->recovery_cp != MaxSector)
3724 printk(KERN_NOTICE "md/raid10:%s: not clean"
3725 " -- starting background reconstruction\n",
3728 "md/raid10:%s: active with %d out of %d devices\n",
3729 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3730 conf->geo.raid_disks);
3732 * Ok, everything is just fine now
3734 mddev->dev_sectors = conf->dev_sectors;
3735 size = raid10_size(mddev, 0, 0);
3736 md_set_array_sectors(mddev, size);
3737 mddev->resync_max_sectors = size;
3740 int stripe = conf->geo.raid_disks *
3741 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3742 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3743 mddev->queue->backing_dev_info.congested_data = mddev;
3745 /* Calculate max read-ahead size.
3746 * We need to readahead at least twice a whole stripe....
3749 stripe /= conf->geo.near_copies;
3750 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3751 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3752 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3756 if (md_integrity_register(mddev))
3759 if (conf->reshape_progress != MaxSector) {
3760 unsigned long before_length, after_length;
3762 before_length = ((1 << conf->prev.chunk_shift) *
3763 conf->prev.far_copies);
3764 after_length = ((1 << conf->geo.chunk_shift) *
3765 conf->geo.far_copies);
3767 if (max(before_length, after_length) > min_offset_diff) {
3768 /* This cannot work */
3769 printk("md/raid10: offset difference not enough to continue reshape\n");
3772 conf->offset_diff = min_offset_diff;
3774 conf->reshape_safe = conf->reshape_progress;
3775 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3776 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3777 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3778 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3779 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3786 md_unregister_thread(&mddev->thread);
3787 if (conf->r10bio_pool)
3788 mempool_destroy(conf->r10bio_pool);
3789 safe_put_page(conf->tmppage);
3790 kfree(conf->mirrors);
3792 mddev->private = NULL;
3797 static int stop(struct mddev *mddev)
3799 struct r10conf *conf = mddev->private;
3801 raise_barrier(conf, 0);
3802 lower_barrier(conf);
3804 md_unregister_thread(&mddev->thread);
3806 /* the unplug fn references 'conf'*/
3807 blk_sync_queue(mddev->queue);
3809 if (conf->r10bio_pool)
3810 mempool_destroy(conf->r10bio_pool);
3811 kfree(conf->mirrors);
3813 mddev->private = NULL;
3817 static void raid10_quiesce(struct mddev *mddev, int state)
3819 struct r10conf *conf = mddev->private;
3823 raise_barrier(conf, 0);
3826 lower_barrier(conf);
3831 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3833 /* Resize of 'far' arrays is not supported.
3834 * For 'near' and 'offset' arrays we can set the
3835 * number of sectors used to be an appropriate multiple
3836 * of the chunk size.
3837 * For 'offset', this is far_copies*chunksize.
3838 * For 'near' the multiplier is the LCM of
3839 * near_copies and raid_disks.
3840 * So if far_copies > 1 && !far_offset, fail.
3841 * Else find LCM(raid_disks, near_copy)*far_copies and
3842 * multiply by chunk_size. Then round to this number.
3843 * This is mostly done by raid10_size()
3845 struct r10conf *conf = mddev->private;
3846 sector_t oldsize, size;
3848 if (mddev->reshape_position != MaxSector)
3851 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3854 oldsize = raid10_size(mddev, 0, 0);
3855 size = raid10_size(mddev, sectors, 0);
3856 if (mddev->external_size &&
3857 mddev->array_sectors > size)
3859 if (mddev->bitmap) {
3860 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3864 md_set_array_sectors(mddev, size);
3865 set_capacity(mddev->gendisk, mddev->array_sectors);
3866 revalidate_disk(mddev->gendisk);
3867 if (sectors > mddev->dev_sectors &&
3868 mddev->recovery_cp > oldsize) {
3869 mddev->recovery_cp = oldsize;
3870 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3872 calc_sectors(conf, sectors);
3873 mddev->dev_sectors = conf->dev_sectors;
3874 mddev->resync_max_sectors = size;
3878 static void *raid10_takeover_raid0(struct mddev *mddev)
3880 struct md_rdev *rdev;
3881 struct r10conf *conf;
3883 if (mddev->degraded > 0) {
3884 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3886 return ERR_PTR(-EINVAL);
3889 /* Set new parameters */
3890 mddev->new_level = 10;
3891 /* new layout: far_copies = 1, near_copies = 2 */
3892 mddev->new_layout = (1<<8) + 2;
3893 mddev->new_chunk_sectors = mddev->chunk_sectors;
3894 mddev->delta_disks = mddev->raid_disks;
3895 mddev->raid_disks *= 2;
3896 /* make sure it will be not marked as dirty */
3897 mddev->recovery_cp = MaxSector;
3899 conf = setup_conf(mddev);
3900 if (!IS_ERR(conf)) {
3901 rdev_for_each(rdev, mddev)
3902 if (rdev->raid_disk >= 0)
3903 rdev->new_raid_disk = rdev->raid_disk * 2;
3910 static void *raid10_takeover(struct mddev *mddev)
3912 struct r0conf *raid0_conf;
3914 /* raid10 can take over:
3915 * raid0 - providing it has only two drives
3917 if (mddev->level == 0) {
3918 /* for raid0 takeover only one zone is supported */
3919 raid0_conf = mddev->private;
3920 if (raid0_conf->nr_strip_zones > 1) {
3921 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3922 " with more than one zone.\n",
3924 return ERR_PTR(-EINVAL);
3926 return raid10_takeover_raid0(mddev);
3928 return ERR_PTR(-EINVAL);
3931 static int raid10_check_reshape(struct mddev *mddev)
3933 /* Called when there is a request to change
3934 * - layout (to ->new_layout)
3935 * - chunk size (to ->new_chunk_sectors)
3936 * - raid_disks (by delta_disks)
3937 * or when trying to restart a reshape that was ongoing.
3939 * We need to validate the request and possibly allocate
3940 * space if that might be an issue later.
3942 * Currently we reject any reshape of a 'far' mode array,
3943 * allow chunk size to change if new is generally acceptable,
3944 * allow raid_disks to increase, and allow
3945 * a switch between 'near' mode and 'offset' mode.
3947 struct r10conf *conf = mddev->private;
3950 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3953 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3954 /* mustn't change number of copies */
3956 if (geo.far_copies > 1 && !geo.far_offset)
3957 /* Cannot switch to 'far' mode */
3960 if (mddev->array_sectors & geo.chunk_mask)
3961 /* not factor of array size */
3964 if (!enough(conf, -1))
3967 kfree(conf->mirrors_new);
3968 conf->mirrors_new = NULL;
3969 if (mddev->delta_disks > 0) {
3970 /* allocate new 'mirrors' list */
3971 conf->mirrors_new = kzalloc(
3972 sizeof(struct raid10_info)
3973 *(mddev->raid_disks +
3974 mddev->delta_disks),
3976 if (!conf->mirrors_new)
3983 * Need to check if array has failed when deciding whether to:
3985 * - remove non-faulty devices
3988 * This determination is simple when no reshape is happening.
3989 * However if there is a reshape, we need to carefully check
3990 * both the before and after sections.
3991 * This is because some failed devices may only affect one
3992 * of the two sections, and some non-in_sync devices may
3993 * be insync in the section most affected by failed devices.
3995 static int calc_degraded(struct r10conf *conf)
3997 int degraded, degraded2;
4002 /* 'prev' section first */
4003 for (i = 0; i < conf->prev.raid_disks; i++) {
4004 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4005 if (!rdev || test_bit(Faulty, &rdev->flags))
4007 else if (!test_bit(In_sync, &rdev->flags))
4008 /* When we can reduce the number of devices in
4009 * an array, this might not contribute to
4010 * 'degraded'. It does now.
4015 if (conf->geo.raid_disks == conf->prev.raid_disks)
4019 for (i = 0; i < conf->geo.raid_disks; i++) {
4020 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4021 if (!rdev || test_bit(Faulty, &rdev->flags))
4023 else if (!test_bit(In_sync, &rdev->flags)) {
4024 /* If reshape is increasing the number of devices,
4025 * this section has already been recovered, so
4026 * it doesn't contribute to degraded.
4029 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4034 if (degraded2 > degraded)
4039 static int raid10_start_reshape(struct mddev *mddev)
4041 /* A 'reshape' has been requested. This commits
4042 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4043 * This also checks if there are enough spares and adds them
4045 * We currently require enough spares to make the final
4046 * array non-degraded. We also require that the difference
4047 * between old and new data_offset - on each device - is
4048 * enough that we never risk over-writing.
4051 unsigned long before_length, after_length;
4052 sector_t min_offset_diff = 0;
4055 struct r10conf *conf = mddev->private;
4056 struct md_rdev *rdev;
4060 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4063 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4066 before_length = ((1 << conf->prev.chunk_shift) *
4067 conf->prev.far_copies);
4068 after_length = ((1 << conf->geo.chunk_shift) *
4069 conf->geo.far_copies);
4071 rdev_for_each(rdev, mddev) {
4072 if (!test_bit(In_sync, &rdev->flags)
4073 && !test_bit(Faulty, &rdev->flags))
4075 if (rdev->raid_disk >= 0) {
4076 long long diff = (rdev->new_data_offset
4077 - rdev->data_offset);
4078 if (!mddev->reshape_backwards)
4082 if (first || diff < min_offset_diff)
4083 min_offset_diff = diff;
4087 if (max(before_length, after_length) > min_offset_diff)
4090 if (spares < mddev->delta_disks)
4093 conf->offset_diff = min_offset_diff;
4094 spin_lock_irq(&conf->device_lock);
4095 if (conf->mirrors_new) {
4096 memcpy(conf->mirrors_new, conf->mirrors,
4097 sizeof(struct raid10_info)*conf->prev.raid_disks);
4099 kfree(conf->mirrors_old); /* FIXME and elsewhere */
4100 conf->mirrors_old = conf->mirrors;
4101 conf->mirrors = conf->mirrors_new;
4102 conf->mirrors_new = NULL;
4104 setup_geo(&conf->geo, mddev, geo_start);
4106 if (mddev->reshape_backwards) {
4107 sector_t size = raid10_size(mddev, 0, 0);
4108 if (size < mddev->array_sectors) {
4109 spin_unlock_irq(&conf->device_lock);
4110 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4114 mddev->resync_max_sectors = size;
4115 conf->reshape_progress = size;
4117 conf->reshape_progress = 0;
4118 spin_unlock_irq(&conf->device_lock);
4120 if (mddev->delta_disks && mddev->bitmap) {
4121 ret = bitmap_resize(mddev->bitmap,
4122 raid10_size(mddev, 0,
4123 conf->geo.raid_disks),
4128 if (mddev->delta_disks > 0) {
4129 rdev_for_each(rdev, mddev)
4130 if (rdev->raid_disk < 0 &&
4131 !test_bit(Faulty, &rdev->flags)) {
4132 if (raid10_add_disk(mddev, rdev) == 0) {
4133 if (rdev->raid_disk >=
4134 conf->prev.raid_disks)
4135 set_bit(In_sync, &rdev->flags);
4137 rdev->recovery_offset = 0;
4139 if (sysfs_link_rdev(mddev, rdev))
4140 /* Failure here is OK */;
4142 } else if (rdev->raid_disk >= conf->prev.raid_disks
4143 && !test_bit(Faulty, &rdev->flags)) {
4144 /* This is a spare that was manually added */
4145 set_bit(In_sync, &rdev->flags);
4148 /* When a reshape changes the number of devices,
4149 * ->degraded is measured against the larger of the
4150 * pre and post numbers.
4152 spin_lock_irq(&conf->device_lock);
4153 mddev->degraded = calc_degraded(conf);
4154 spin_unlock_irq(&conf->device_lock);
4155 mddev->raid_disks = conf->geo.raid_disks;
4156 mddev->reshape_position = conf->reshape_progress;
4157 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4159 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4160 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4161 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4162 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4164 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4166 if (!mddev->sync_thread) {
4170 conf->reshape_checkpoint = jiffies;
4171 md_wakeup_thread(mddev->sync_thread);
4172 md_new_event(mddev);
4176 mddev->recovery = 0;
4177 spin_lock_irq(&conf->device_lock);
4178 conf->geo = conf->prev;
4179 mddev->raid_disks = conf->geo.raid_disks;
4180 rdev_for_each(rdev, mddev)
4181 rdev->new_data_offset = rdev->data_offset;
4183 conf->reshape_progress = MaxSector;
4184 mddev->reshape_position = MaxSector;
4185 spin_unlock_irq(&conf->device_lock);
4189 /* Calculate the last device-address that could contain
4190 * any block from the chunk that includes the array-address 's'
4191 * and report the next address.
4192 * i.e. the address returned will be chunk-aligned and after
4193 * any data that is in the chunk containing 's'.
4195 static sector_t last_dev_address(sector_t s, struct geom *geo)
4197 s = (s | geo->chunk_mask) + 1;
4198 s >>= geo->chunk_shift;
4199 s *= geo->near_copies;
4200 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4201 s *= geo->far_copies;
4202 s <<= geo->chunk_shift;
4206 /* Calculate the first device-address that could contain
4207 * any block from the chunk that includes the array-address 's'.
4208 * This too will be the start of a chunk
4210 static sector_t first_dev_address(sector_t s, struct geom *geo)
4212 s >>= geo->chunk_shift;
4213 s *= geo->near_copies;
4214 sector_div(s, geo->raid_disks);
4215 s *= geo->far_copies;
4216 s <<= geo->chunk_shift;
4220 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4223 /* We simply copy at most one chunk (smallest of old and new)
4224 * at a time, possibly less if that exceeds RESYNC_PAGES,
4225 * or we hit a bad block or something.
4226 * This might mean we pause for normal IO in the middle of
4227 * a chunk, but that is not a problem was mddev->reshape_position
4228 * can record any location.
4230 * If we will want to write to a location that isn't
4231 * yet recorded as 'safe' (i.e. in metadata on disk) then
4232 * we need to flush all reshape requests and update the metadata.
4234 * When reshaping forwards (e.g. to more devices), we interpret
4235 * 'safe' as the earliest block which might not have been copied
4236 * down yet. We divide this by previous stripe size and multiply
4237 * by previous stripe length to get lowest device offset that we
4238 * cannot write to yet.
4239 * We interpret 'sector_nr' as an address that we want to write to.
4240 * From this we use last_device_address() to find where we might
4241 * write to, and first_device_address on the 'safe' position.
4242 * If this 'next' write position is after the 'safe' position,
4243 * we must update the metadata to increase the 'safe' position.
4245 * When reshaping backwards, we round in the opposite direction
4246 * and perform the reverse test: next write position must not be
4247 * less than current safe position.
4249 * In all this the minimum difference in data offsets
4250 * (conf->offset_diff - always positive) allows a bit of slack,
4251 * so next can be after 'safe', but not by more than offset_disk
4253 * We need to prepare all the bios here before we start any IO
4254 * to ensure the size we choose is acceptable to all devices.
4255 * The means one for each copy for write-out and an extra one for
4257 * We store the read-in bio in ->master_bio and the others in
4258 * ->devs[x].bio and ->devs[x].repl_bio.
4260 struct r10conf *conf = mddev->private;
4261 struct r10bio *r10_bio;
4262 sector_t next, safe, last;
4266 struct md_rdev *rdev;
4269 struct bio *bio, *read_bio;
4270 int sectors_done = 0;
4272 if (sector_nr == 0) {
4273 /* If restarting in the middle, skip the initial sectors */
4274 if (mddev->reshape_backwards &&
4275 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4276 sector_nr = (raid10_size(mddev, 0, 0)
4277 - conf->reshape_progress);
4278 } else if (!mddev->reshape_backwards &&
4279 conf->reshape_progress > 0)
4280 sector_nr = conf->reshape_progress;
4282 mddev->curr_resync_completed = sector_nr;
4283 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4289 /* We don't use sector_nr to track where we are up to
4290 * as that doesn't work well for ->reshape_backwards.
4291 * So just use ->reshape_progress.
4293 if (mddev->reshape_backwards) {
4294 /* 'next' is the earliest device address that we might
4295 * write to for this chunk in the new layout
4297 next = first_dev_address(conf->reshape_progress - 1,
4300 /* 'safe' is the last device address that we might read from
4301 * in the old layout after a restart
4303 safe = last_dev_address(conf->reshape_safe - 1,
4306 if (next + conf->offset_diff < safe)
4309 last = conf->reshape_progress - 1;
4310 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4311 & conf->prev.chunk_mask);
4312 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4313 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4315 /* 'next' is after the last device address that we
4316 * might write to for this chunk in the new layout
4318 next = last_dev_address(conf->reshape_progress, &conf->geo);
4320 /* 'safe' is the earliest device address that we might
4321 * read from in the old layout after a restart
4323 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4325 /* Need to update metadata if 'next' might be beyond 'safe'
4326 * as that would possibly corrupt data
4328 if (next > safe + conf->offset_diff)
4331 sector_nr = conf->reshape_progress;
4332 last = sector_nr | (conf->geo.chunk_mask
4333 & conf->prev.chunk_mask);
4335 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4336 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4340 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4341 /* Need to update reshape_position in metadata */
4343 mddev->reshape_position = conf->reshape_progress;
4344 if (mddev->reshape_backwards)
4345 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4346 - conf->reshape_progress;
4348 mddev->curr_resync_completed = conf->reshape_progress;
4349 conf->reshape_checkpoint = jiffies;
4350 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4351 md_wakeup_thread(mddev->thread);
4352 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4353 kthread_should_stop());
4354 conf->reshape_safe = mddev->reshape_position;
4355 allow_barrier(conf);
4359 /* Now schedule reads for blocks from sector_nr to last */
4360 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4361 raise_barrier(conf, sectors_done != 0);
4362 atomic_set(&r10_bio->remaining, 0);
4363 r10_bio->mddev = mddev;
4364 r10_bio->sector = sector_nr;
4365 set_bit(R10BIO_IsReshape, &r10_bio->state);
4366 r10_bio->sectors = last - sector_nr + 1;
4367 rdev = read_balance(conf, r10_bio, &max_sectors);
4368 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4371 /* Cannot read from here, so need to record bad blocks
4372 * on all the target devices.
4375 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4376 return sectors_done;
4379 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4381 read_bio->bi_bdev = rdev->bdev;
4382 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4383 + rdev->data_offset);
4384 read_bio->bi_private = r10_bio;
4385 read_bio->bi_end_io = end_sync_read;
4386 read_bio->bi_rw = READ;
4387 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4388 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4389 read_bio->bi_vcnt = 0;
4390 read_bio->bi_idx = 0;
4391 read_bio->bi_size = 0;
4392 r10_bio->master_bio = read_bio;
4393 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4395 /* Now find the locations in the new layout */
4396 __raid10_find_phys(&conf->geo, r10_bio);
4399 read_bio->bi_next = NULL;
4401 for (s = 0; s < conf->copies*2; s++) {
4403 int d = r10_bio->devs[s/2].devnum;
4404 struct md_rdev *rdev2;
4406 rdev2 = conf->mirrors[d].replacement;
4407 b = r10_bio->devs[s/2].repl_bio;
4409 rdev2 = conf->mirrors[d].rdev;
4410 b = r10_bio->devs[s/2].bio;
4412 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4414 b->bi_bdev = rdev2->bdev;
4415 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4416 b->bi_private = r10_bio;
4417 b->bi_end_io = end_reshape_write;
4419 b->bi_flags &= ~(BIO_POOL_MASK - 1);
4420 b->bi_flags |= 1 << BIO_UPTODATE;
4428 /* Now add as many pages as possible to all of these bios. */
4431 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4432 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4433 int len = (max_sectors - s) << 9;
4434 if (len > PAGE_SIZE)
4436 for (bio = blist; bio ; bio = bio->bi_next) {
4438 if (bio_add_page(bio, page, len, 0))
4441 /* Didn't fit, must stop */
4443 bio2 && bio2 != bio;
4444 bio2 = bio2->bi_next) {
4445 /* Remove last page from this bio */
4447 bio2->bi_size -= len;
4448 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4452 sector_nr += len >> 9;
4453 nr_sectors += len >> 9;
4456 r10_bio->sectors = nr_sectors;
4458 /* Now submit the read */
4459 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4460 atomic_inc(&r10_bio->remaining);
4461 read_bio->bi_next = NULL;
4462 generic_make_request(read_bio);
4463 sector_nr += nr_sectors;
4464 sectors_done += nr_sectors;
4465 if (sector_nr <= last)
4468 /* Now that we have done the whole section we can
4469 * update reshape_progress
4471 if (mddev->reshape_backwards)
4472 conf->reshape_progress -= sectors_done;
4474 conf->reshape_progress += sectors_done;
4476 return sectors_done;
4479 static void end_reshape_request(struct r10bio *r10_bio);
4480 static int handle_reshape_read_error(struct mddev *mddev,
4481 struct r10bio *r10_bio);
4482 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4484 /* Reshape read completed. Hopefully we have a block
4486 * If we got a read error then we do sync 1-page reads from
4487 * elsewhere until we find the data - or give up.
4489 struct r10conf *conf = mddev->private;
4492 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4493 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4494 /* Reshape has been aborted */
4495 md_done_sync(mddev, r10_bio->sectors, 0);
4499 /* We definitely have the data in the pages, schedule the
4502 atomic_set(&r10_bio->remaining, 1);
4503 for (s = 0; s < conf->copies*2; s++) {
4505 int d = r10_bio->devs[s/2].devnum;
4506 struct md_rdev *rdev;
4508 rdev = conf->mirrors[d].replacement;
4509 b = r10_bio->devs[s/2].repl_bio;
4511 rdev = conf->mirrors[d].rdev;
4512 b = r10_bio->devs[s/2].bio;
4514 if (!rdev || test_bit(Faulty, &rdev->flags))
4516 atomic_inc(&rdev->nr_pending);
4517 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4518 atomic_inc(&r10_bio->remaining);
4520 generic_make_request(b);
4522 end_reshape_request(r10_bio);
4525 static void end_reshape(struct r10conf *conf)
4527 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4530 spin_lock_irq(&conf->device_lock);
4531 conf->prev = conf->geo;
4532 md_finish_reshape(conf->mddev);
4534 conf->reshape_progress = MaxSector;
4535 spin_unlock_irq(&conf->device_lock);
4537 /* read-ahead size must cover two whole stripes, which is
4538 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4540 if (conf->mddev->queue) {
4541 int stripe = conf->geo.raid_disks *
4542 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4543 stripe /= conf->geo.near_copies;
4544 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4545 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4551 static int handle_reshape_read_error(struct mddev *mddev,
4552 struct r10bio *r10_bio)
4554 /* Use sync reads to get the blocks from somewhere else */
4555 int sectors = r10_bio->sectors;
4556 struct r10conf *conf = mddev->private;
4558 struct r10bio r10_bio;
4559 struct r10dev devs[conf->copies];
4561 struct r10bio *r10b = &on_stack.r10_bio;
4564 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4566 r10b->sector = r10_bio->sector;
4567 __raid10_find_phys(&conf->prev, r10b);
4572 int first_slot = slot;
4574 if (s > (PAGE_SIZE >> 9))
4578 int d = r10b->devs[slot].devnum;
4579 struct md_rdev *rdev = conf->mirrors[d].rdev;
4582 test_bit(Faulty, &rdev->flags) ||
4583 !test_bit(In_sync, &rdev->flags))
4586 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4587 success = sync_page_io(rdev,
4596 if (slot >= conf->copies)
4598 if (slot == first_slot)
4602 /* couldn't read this block, must give up */
4603 set_bit(MD_RECOVERY_INTR,
4613 static void end_reshape_write(struct bio *bio, int error)
4615 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4616 struct r10bio *r10_bio = bio->bi_private;
4617 struct mddev *mddev = r10_bio->mddev;
4618 struct r10conf *conf = mddev->private;
4622 struct md_rdev *rdev = NULL;
4624 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4626 rdev = conf->mirrors[d].replacement;
4629 rdev = conf->mirrors[d].rdev;
4633 /* FIXME should record badblock */
4634 md_error(mddev, rdev);
4637 rdev_dec_pending(rdev, mddev);
4638 end_reshape_request(r10_bio);
4641 static void end_reshape_request(struct r10bio *r10_bio)
4643 if (!atomic_dec_and_test(&r10_bio->remaining))
4645 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4646 bio_put(r10_bio->master_bio);
4650 static void raid10_finish_reshape(struct mddev *mddev)
4652 struct r10conf *conf = mddev->private;
4654 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4657 if (mddev->delta_disks > 0) {
4658 sector_t size = raid10_size(mddev, 0, 0);
4659 md_set_array_sectors(mddev, size);
4660 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4661 mddev->recovery_cp = mddev->resync_max_sectors;
4662 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4664 mddev->resync_max_sectors = size;
4665 set_capacity(mddev->gendisk, mddev->array_sectors);
4666 revalidate_disk(mddev->gendisk);
4669 for (d = conf->geo.raid_disks ;
4670 d < conf->geo.raid_disks - mddev->delta_disks;
4672 struct md_rdev *rdev = conf->mirrors[d].rdev;
4674 clear_bit(In_sync, &rdev->flags);
4675 rdev = conf->mirrors[d].replacement;
4677 clear_bit(In_sync, &rdev->flags);
4680 mddev->layout = mddev->new_layout;
4681 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4682 mddev->reshape_position = MaxSector;
4683 mddev->delta_disks = 0;
4684 mddev->reshape_backwards = 0;
4687 static struct md_personality raid10_personality =
4691 .owner = THIS_MODULE,
4692 .make_request = make_request,
4696 .error_handler = error,
4697 .hot_add_disk = raid10_add_disk,
4698 .hot_remove_disk= raid10_remove_disk,
4699 .spare_active = raid10_spare_active,
4700 .sync_request = sync_request,
4701 .quiesce = raid10_quiesce,
4702 .size = raid10_size,
4703 .resize = raid10_resize,
4704 .takeover = raid10_takeover,
4705 .check_reshape = raid10_check_reshape,
4706 .start_reshape = raid10_start_reshape,
4707 .finish_reshape = raid10_finish_reshape,
4710 static int __init raid_init(void)
4712 return register_md_personality(&raid10_personality);
4715 static void raid_exit(void)
4717 unregister_md_personality(&raid10_personality);
4720 module_init(raid_init);
4721 module_exit(raid_exit);
4722 MODULE_LICENSE("GPL");
4723 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4724 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4725 MODULE_ALIAS("md-raid10");
4726 MODULE_ALIAS("md-level-10");
4728 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);