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 previous, 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;
369 slot = r10_bio->read_slot;
370 dev = r10_bio->devs[slot].devnum;
371 rdev = r10_bio->devs[slot].rdev;
373 * this branch is our 'one mirror IO has finished' event handler:
375 update_head_pos(slot, r10_bio);
379 * Set R10BIO_Uptodate in our master bio, so that
380 * we will return a good error code to the higher
381 * levels even if IO on some other mirrored buffer fails.
383 * The 'master' represents the composite IO operation to
384 * user-side. So if something waits for IO, then it will
385 * wait for the 'master' bio.
387 set_bit(R10BIO_Uptodate, &r10_bio->state);
389 /* If all other devices that store this block have
390 * failed, we want to return the error upwards rather
391 * than fail the last device. Here we redefine
392 * "uptodate" to mean "Don't want to retry"
394 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
399 raid_end_bio_io(r10_bio);
400 rdev_dec_pending(rdev, conf->mddev);
403 * oops, read error - keep the refcount on the rdev
405 char b[BDEVNAME_SIZE];
406 printk_ratelimited(KERN_ERR
407 "md/raid10:%s: %s: rescheduling sector %llu\n",
409 bdevname(rdev->bdev, b),
410 (unsigned long long)r10_bio->sector);
411 set_bit(R10BIO_ReadError, &r10_bio->state);
412 reschedule_retry(r10_bio);
416 static void close_write(struct r10bio *r10_bio)
418 /* clear the bitmap if all writes complete successfully */
419 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
421 !test_bit(R10BIO_Degraded, &r10_bio->state),
423 md_write_end(r10_bio->mddev);
426 static void one_write_done(struct r10bio *r10_bio)
428 if (atomic_dec_and_test(&r10_bio->remaining)) {
429 if (test_bit(R10BIO_WriteError, &r10_bio->state))
430 reschedule_retry(r10_bio);
432 close_write(r10_bio);
433 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
434 reschedule_retry(r10_bio);
436 raid_end_bio_io(r10_bio);
441 static void raid10_end_write_request(struct bio *bio, int error)
443 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
444 struct r10bio *r10_bio = bio->bi_private;
447 struct r10conf *conf = r10_bio->mddev->private;
449 struct md_rdev *rdev = NULL;
451 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
454 rdev = conf->mirrors[dev].replacement;
458 rdev = conf->mirrors[dev].rdev;
461 * this branch is our 'one mirror IO has finished' event handler:
465 /* Never record new bad blocks to replacement,
468 md_error(rdev->mddev, rdev);
470 set_bit(WriteErrorSeen, &rdev->flags);
471 if (!test_and_set_bit(WantReplacement, &rdev->flags))
472 set_bit(MD_RECOVERY_NEEDED,
473 &rdev->mddev->recovery);
474 set_bit(R10BIO_WriteError, &r10_bio->state);
479 * Set R10BIO_Uptodate in our master bio, so that
480 * we will return a good error code for to the higher
481 * levels even if IO on some other mirrored buffer fails.
483 * The 'master' represents the composite IO operation to
484 * user-side. So if something waits for IO, then it will
485 * wait for the 'master' bio.
491 * Do not set R10BIO_Uptodate if the current device is
492 * rebuilding or Faulty. This is because we cannot use
493 * such device for properly reading the data back (we could
494 * potentially use it, if the current write would have felt
495 * before rdev->recovery_offset, but for simplicity we don't
498 if (test_bit(In_sync, &rdev->flags) &&
499 !test_bit(Faulty, &rdev->flags))
500 set_bit(R10BIO_Uptodate, &r10_bio->state);
502 /* Maybe we can clear some bad blocks. */
503 if (is_badblock(rdev,
504 r10_bio->devs[slot].addr,
506 &first_bad, &bad_sectors)) {
509 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
511 r10_bio->devs[slot].bio = IO_MADE_GOOD;
513 set_bit(R10BIO_MadeGood, &r10_bio->state);
519 * Let's see if all mirrored write operations have finished
522 one_write_done(r10_bio);
524 rdev_dec_pending(rdev, conf->mddev);
528 * RAID10 layout manager
529 * As well as the chunksize and raid_disks count, there are two
530 * parameters: near_copies and far_copies.
531 * near_copies * far_copies must be <= raid_disks.
532 * Normally one of these will be 1.
533 * If both are 1, we get raid0.
534 * If near_copies == raid_disks, we get raid1.
536 * Chunks are laid out in raid0 style with near_copies copies of the
537 * first chunk, followed by near_copies copies of the next chunk and
539 * If far_copies > 1, then after 1/far_copies of the array has been assigned
540 * as described above, we start again with a device offset of near_copies.
541 * So we effectively have another copy of the whole array further down all
542 * the drives, but with blocks on different drives.
543 * With this layout, and block is never stored twice on the one device.
545 * raid10_find_phys finds the sector offset of a given virtual sector
546 * on each device that it is on.
548 * raid10_find_virt does the reverse mapping, from a device and a
549 * sector offset to a virtual address
552 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
560 int last_far_set_start, last_far_set_size;
562 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
563 last_far_set_start *= geo->far_set_size;
565 last_far_set_size = geo->far_set_size;
566 last_far_set_size += (geo->raid_disks % geo->far_set_size);
568 /* now calculate first sector/dev */
569 chunk = r10bio->sector >> geo->chunk_shift;
570 sector = r10bio->sector & geo->chunk_mask;
572 chunk *= geo->near_copies;
574 dev = sector_div(stripe, geo->raid_disks);
576 stripe *= geo->far_copies;
578 sector += stripe << geo->chunk_shift;
580 /* and calculate all the others */
581 for (n = 0; n < geo->near_copies; n++) {
585 r10bio->devs[slot].devnum = d;
586 r10bio->devs[slot].addr = s;
589 for (f = 1; f < geo->far_copies; f++) {
590 set = d / geo->far_set_size;
591 d += geo->near_copies;
593 if ((geo->raid_disks % geo->far_set_size) &&
594 (d > last_far_set_start)) {
595 d -= last_far_set_start;
596 d %= last_far_set_size;
597 d += last_far_set_start;
599 d %= geo->far_set_size;
600 d += geo->far_set_size * set;
603 r10bio->devs[slot].devnum = d;
604 r10bio->devs[slot].addr = s;
608 if (dev >= geo->raid_disks) {
610 sector += (geo->chunk_mask + 1);
615 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
617 struct geom *geo = &conf->geo;
619 if (conf->reshape_progress != MaxSector &&
620 ((r10bio->sector >= conf->reshape_progress) !=
621 conf->mddev->reshape_backwards)) {
622 set_bit(R10BIO_Previous, &r10bio->state);
625 clear_bit(R10BIO_Previous, &r10bio->state);
627 __raid10_find_phys(geo, r10bio);
630 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
632 sector_t offset, chunk, vchunk;
633 /* Never use conf->prev as this is only called during resync
634 * or recovery, so reshape isn't happening
636 struct geom *geo = &conf->geo;
637 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
638 int far_set_size = geo->far_set_size;
639 int last_far_set_start;
641 if (geo->raid_disks % geo->far_set_size) {
642 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
643 last_far_set_start *= geo->far_set_size;
645 if (dev >= last_far_set_start) {
646 far_set_size = geo->far_set_size;
647 far_set_size += (geo->raid_disks % geo->far_set_size);
648 far_set_start = last_far_set_start;
652 offset = sector & geo->chunk_mask;
653 if (geo->far_offset) {
655 chunk = sector >> geo->chunk_shift;
656 fc = sector_div(chunk, geo->far_copies);
657 dev -= fc * geo->near_copies;
658 if (dev < far_set_start)
661 while (sector >= geo->stride) {
662 sector -= geo->stride;
663 if (dev < (geo->near_copies + far_set_start))
664 dev += far_set_size - geo->near_copies;
666 dev -= geo->near_copies;
668 chunk = sector >> geo->chunk_shift;
670 vchunk = chunk * geo->raid_disks + dev;
671 sector_div(vchunk, geo->near_copies);
672 return (vchunk << geo->chunk_shift) + offset;
676 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
677 * @mddev: the md device
678 * @bvm: properties of new bio
679 * @biovec: the request that could be merged to it.
681 * Return amount of bytes we can accept at this offset
682 * This requires checking for end-of-chunk if near_copies != raid_disks,
683 * and for subordinate merge_bvec_fns if merge_check_needed.
685 static int raid10_mergeable_bvec(struct mddev *mddev,
686 struct bvec_merge_data *bvm,
687 struct bio_vec *biovec)
689 struct r10conf *conf = mddev->private;
690 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
692 unsigned int chunk_sectors;
693 unsigned int bio_sectors = bvm->bi_size >> 9;
694 struct geom *geo = &conf->geo;
696 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
697 if (conf->reshape_progress != MaxSector &&
698 ((sector >= conf->reshape_progress) !=
699 conf->mddev->reshape_backwards))
702 if (geo->near_copies < geo->raid_disks) {
703 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
704 + bio_sectors)) << 9;
706 /* bio_add cannot handle a negative return */
708 if (max <= biovec->bv_len && bio_sectors == 0)
709 return biovec->bv_len;
711 max = biovec->bv_len;
713 if (mddev->merge_check_needed) {
715 struct r10bio r10_bio;
716 struct r10dev devs[conf->copies];
718 struct r10bio *r10_bio = &on_stack.r10_bio;
720 if (conf->reshape_progress != MaxSector) {
721 /* Cannot give any guidance during reshape */
722 if (max <= biovec->bv_len && bio_sectors == 0)
723 return biovec->bv_len;
726 r10_bio->sector = sector;
727 raid10_find_phys(conf, r10_bio);
729 for (s = 0; s < conf->copies; s++) {
730 int disk = r10_bio->devs[s].devnum;
731 struct md_rdev *rdev = rcu_dereference(
732 conf->mirrors[disk].rdev);
733 if (rdev && !test_bit(Faulty, &rdev->flags)) {
734 struct request_queue *q =
735 bdev_get_queue(rdev->bdev);
736 if (q->merge_bvec_fn) {
737 bvm->bi_sector = r10_bio->devs[s].addr
739 bvm->bi_bdev = rdev->bdev;
740 max = min(max, q->merge_bvec_fn(
744 rdev = rcu_dereference(conf->mirrors[disk].replacement);
745 if (rdev && !test_bit(Faulty, &rdev->flags)) {
746 struct request_queue *q =
747 bdev_get_queue(rdev->bdev);
748 if (q->merge_bvec_fn) {
749 bvm->bi_sector = r10_bio->devs[s].addr
751 bvm->bi_bdev = rdev->bdev;
752 max = min(max, q->merge_bvec_fn(
763 * This routine returns the disk from which the requested read should
764 * be done. There is a per-array 'next expected sequential IO' sector
765 * number - if this matches on the next IO then we use the last disk.
766 * There is also a per-disk 'last know head position' sector that is
767 * maintained from IRQ contexts, both the normal and the resync IO
768 * completion handlers update this position correctly. If there is no
769 * perfect sequential match then we pick the disk whose head is closest.
771 * If there are 2 mirrors in the same 2 devices, performance degrades
772 * because position is mirror, not device based.
774 * The rdev for the device selected will have nr_pending incremented.
778 * FIXME: possibly should rethink readbalancing and do it differently
779 * depending on near_copies / far_copies geometry.
781 static struct md_rdev *read_balance(struct r10conf *conf,
782 struct r10bio *r10_bio,
785 const sector_t this_sector = r10_bio->sector;
787 int sectors = r10_bio->sectors;
788 int best_good_sectors;
789 sector_t new_distance, best_dist;
790 struct md_rdev *best_rdev, *rdev = NULL;
793 struct geom *geo = &conf->geo;
795 raid10_find_phys(conf, r10_bio);
798 sectors = r10_bio->sectors;
801 best_dist = MaxSector;
802 best_good_sectors = 0;
805 * Check if we can balance. We can balance on the whole
806 * device if no resync is going on (recovery is ok), or below
807 * the resync window. We take the first readable disk when
808 * above the resync window.
810 if (conf->mddev->recovery_cp < MaxSector
811 && (this_sector + sectors >= conf->next_resync))
814 for (slot = 0; slot < conf->copies ; slot++) {
819 if (r10_bio->devs[slot].bio == IO_BLOCKED)
821 disk = r10_bio->devs[slot].devnum;
822 rdev = rcu_dereference(conf->mirrors[disk].replacement);
823 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
824 test_bit(Unmerged, &rdev->flags) ||
825 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
826 rdev = rcu_dereference(conf->mirrors[disk].rdev);
828 test_bit(Faulty, &rdev->flags) ||
829 test_bit(Unmerged, &rdev->flags))
831 if (!test_bit(In_sync, &rdev->flags) &&
832 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
835 dev_sector = r10_bio->devs[slot].addr;
836 if (is_badblock(rdev, dev_sector, sectors,
837 &first_bad, &bad_sectors)) {
838 if (best_dist < MaxSector)
839 /* Already have a better slot */
841 if (first_bad <= dev_sector) {
842 /* Cannot read here. If this is the
843 * 'primary' device, then we must not read
844 * beyond 'bad_sectors' from another device.
846 bad_sectors -= (dev_sector - first_bad);
847 if (!do_balance && sectors > bad_sectors)
848 sectors = bad_sectors;
849 if (best_good_sectors > sectors)
850 best_good_sectors = sectors;
852 sector_t good_sectors =
853 first_bad - dev_sector;
854 if (good_sectors > best_good_sectors) {
855 best_good_sectors = good_sectors;
860 /* Must read from here */
865 best_good_sectors = sectors;
870 /* This optimisation is debatable, and completely destroys
871 * sequential read speed for 'far copies' arrays. So only
872 * keep it for 'near' arrays, and review those later.
874 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
877 /* for far > 1 always use the lowest address */
878 if (geo->far_copies > 1)
879 new_distance = r10_bio->devs[slot].addr;
881 new_distance = abs(r10_bio->devs[slot].addr -
882 conf->mirrors[disk].head_position);
883 if (new_distance < best_dist) {
884 best_dist = new_distance;
889 if (slot >= conf->copies) {
895 atomic_inc(&rdev->nr_pending);
896 if (test_bit(Faulty, &rdev->flags)) {
897 /* Cannot risk returning a device that failed
898 * before we inc'ed nr_pending
900 rdev_dec_pending(rdev, conf->mddev);
903 r10_bio->read_slot = slot;
907 *max_sectors = best_good_sectors;
912 static int raid10_congested(struct mddev *mddev, int bits)
914 struct r10conf *conf = mddev->private;
917 if ((bits & (1 << BDI_async_congested)) &&
918 conf->pending_count >= max_queued_requests)
923 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
926 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
927 if (rdev && !test_bit(Faulty, &rdev->flags)) {
928 struct request_queue *q = bdev_get_queue(rdev->bdev);
930 ret |= bdi_congested(&q->backing_dev_info, bits);
937 static void flush_pending_writes(struct r10conf *conf)
939 /* Any writes that have been queued but are awaiting
940 * bitmap updates get flushed here.
942 spin_lock_irq(&conf->device_lock);
944 if (conf->pending_bio_list.head) {
946 bio = bio_list_get(&conf->pending_bio_list);
947 conf->pending_count = 0;
948 spin_unlock_irq(&conf->device_lock);
949 /* flush any pending bitmap writes to disk
950 * before proceeding w/ I/O */
951 bitmap_unplug(conf->mddev->bitmap);
952 wake_up(&conf->wait_barrier);
954 while (bio) { /* submit pending writes */
955 struct bio *next = bio->bi_next;
957 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
958 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
962 generic_make_request(bio);
966 spin_unlock_irq(&conf->device_lock);
970 * Sometimes we need to suspend IO while we do something else,
971 * either some resync/recovery, or reconfigure the array.
972 * To do this we raise a 'barrier'.
973 * The 'barrier' is a counter that can be raised multiple times
974 * to count how many activities are happening which preclude
976 * We can only raise the barrier if there is no pending IO.
977 * i.e. if nr_pending == 0.
978 * We choose only to raise the barrier if no-one is waiting for the
979 * barrier to go down. This means that as soon as an IO request
980 * is ready, no other operations which require a barrier will start
981 * until the IO request has had a chance.
983 * So: regular IO calls 'wait_barrier'. When that returns there
984 * is no backgroup IO happening, It must arrange to call
985 * allow_barrier when it has finished its IO.
986 * backgroup IO calls must call raise_barrier. Once that returns
987 * there is no normal IO happeing. It must arrange to call
988 * lower_barrier when the particular background IO completes.
991 static void raise_barrier(struct r10conf *conf, int force)
993 BUG_ON(force && !conf->barrier);
994 spin_lock_irq(&conf->resync_lock);
996 /* Wait until no block IO is waiting (unless 'force') */
997 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1000 /* block any new IO from starting */
1003 /* Now wait for all pending IO to complete */
1004 wait_event_lock_irq(conf->wait_barrier,
1005 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1008 spin_unlock_irq(&conf->resync_lock);
1011 static void lower_barrier(struct r10conf *conf)
1013 unsigned long flags;
1014 spin_lock_irqsave(&conf->resync_lock, flags);
1016 spin_unlock_irqrestore(&conf->resync_lock, flags);
1017 wake_up(&conf->wait_barrier);
1020 static void wait_barrier(struct r10conf *conf)
1022 spin_lock_irq(&conf->resync_lock);
1023 if (conf->barrier) {
1025 /* Wait for the barrier to drop.
1026 * However if there are already pending
1027 * requests (preventing the barrier from
1028 * rising completely), and the
1029 * pre-process bio queue isn't empty,
1030 * then don't wait, as we need to empty
1031 * that queue to get the nr_pending
1034 wait_event_lock_irq(conf->wait_barrier,
1036 (conf->nr_pending &&
1037 current->bio_list &&
1038 !bio_list_empty(current->bio_list)),
1043 spin_unlock_irq(&conf->resync_lock);
1046 static void allow_barrier(struct r10conf *conf)
1048 unsigned long flags;
1049 spin_lock_irqsave(&conf->resync_lock, flags);
1051 spin_unlock_irqrestore(&conf->resync_lock, flags);
1052 wake_up(&conf->wait_barrier);
1055 static void freeze_array(struct r10conf *conf, int extra)
1057 /* stop syncio and normal IO and wait for everything to
1059 * We increment barrier and nr_waiting, and then
1060 * wait until nr_pending match nr_queued+extra
1061 * This is called in the context of one normal IO request
1062 * that has failed. Thus any sync request that might be pending
1063 * will be blocked by nr_pending, and we need to wait for
1064 * pending IO requests to complete or be queued for re-try.
1065 * Thus the number queued (nr_queued) plus this request (extra)
1066 * must match the number of pending IOs (nr_pending) before
1069 spin_lock_irq(&conf->resync_lock);
1072 wait_event_lock_irq_cmd(conf->wait_barrier,
1073 conf->nr_pending == conf->nr_queued+extra,
1075 flush_pending_writes(conf));
1077 spin_unlock_irq(&conf->resync_lock);
1080 static void unfreeze_array(struct r10conf *conf)
1082 /* reverse the effect of the freeze */
1083 spin_lock_irq(&conf->resync_lock);
1086 wake_up(&conf->wait_barrier);
1087 spin_unlock_irq(&conf->resync_lock);
1090 static sector_t choose_data_offset(struct r10bio *r10_bio,
1091 struct md_rdev *rdev)
1093 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1094 test_bit(R10BIO_Previous, &r10_bio->state))
1095 return rdev->data_offset;
1097 return rdev->new_data_offset;
1100 struct raid10_plug_cb {
1101 struct blk_plug_cb cb;
1102 struct bio_list pending;
1106 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1108 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1110 struct mddev *mddev = plug->cb.data;
1111 struct r10conf *conf = mddev->private;
1114 if (from_schedule || current->bio_list) {
1115 spin_lock_irq(&conf->device_lock);
1116 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1117 conf->pending_count += plug->pending_cnt;
1118 spin_unlock_irq(&conf->device_lock);
1119 wake_up(&conf->wait_barrier);
1120 md_wakeup_thread(mddev->thread);
1125 /* we aren't scheduling, so we can do the write-out directly. */
1126 bio = bio_list_get(&plug->pending);
1127 bitmap_unplug(mddev->bitmap);
1128 wake_up(&conf->wait_barrier);
1130 while (bio) { /* submit pending writes */
1131 struct bio *next = bio->bi_next;
1132 bio->bi_next = NULL;
1133 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1134 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1135 /* Just ignore it */
1138 generic_make_request(bio);
1144 static void __make_request(struct mddev *mddev, struct bio *bio)
1146 struct r10conf *conf = mddev->private;
1147 struct r10bio *r10_bio;
1148 struct bio *read_bio;
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;
1165 * Register the new request and wait if the reconstruction
1166 * thread has put up a bar for new requests.
1167 * Continue immediately if no resync is active currently.
1171 sectors = bio_sectors(bio);
1172 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1173 bio->bi_iter.bi_sector < conf->reshape_progress &&
1174 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1175 /* IO spans the reshape position. Need to wait for
1178 allow_barrier(conf);
1179 wait_event(conf->wait_barrier,
1180 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1181 conf->reshape_progress >= bio->bi_iter.bi_sector +
1185 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1186 bio_data_dir(bio) == WRITE &&
1187 (mddev->reshape_backwards
1188 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1189 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1190 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1191 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1192 /* Need to update reshape_position in metadata */
1193 mddev->reshape_position = conf->reshape_progress;
1194 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1195 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1196 md_wakeup_thread(mddev->thread);
1197 wait_event(mddev->sb_wait,
1198 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1200 conf->reshape_safe = mddev->reshape_position;
1203 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1205 r10_bio->master_bio = bio;
1206 r10_bio->sectors = sectors;
1208 r10_bio->mddev = mddev;
1209 r10_bio->sector = bio->bi_iter.bi_sector;
1212 /* We might need to issue multiple reads to different
1213 * devices if there are bad blocks around, so we keep
1214 * track of the number of reads in bio->bi_phys_segments.
1215 * If this is 0, there is only one r10_bio and no locking
1216 * will be needed when the request completes. If it is
1217 * non-zero, then it is the number of not-completed requests.
1219 bio->bi_phys_segments = 0;
1220 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1224 * read balancing logic:
1226 struct md_rdev *rdev;
1230 rdev = read_balance(conf, r10_bio, &max_sectors);
1232 raid_end_bio_io(r10_bio);
1235 slot = r10_bio->read_slot;
1237 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1238 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1241 r10_bio->devs[slot].bio = read_bio;
1242 r10_bio->devs[slot].rdev = rdev;
1244 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1245 choose_data_offset(r10_bio, rdev);
1246 read_bio->bi_bdev = rdev->bdev;
1247 read_bio->bi_end_io = raid10_end_read_request;
1248 read_bio->bi_rw = READ | do_sync;
1249 read_bio->bi_private = r10_bio;
1251 if (max_sectors < r10_bio->sectors) {
1252 /* Could not read all from this device, so we will
1253 * need another r10_bio.
1255 sectors_handled = (r10_bio->sector + max_sectors
1256 - bio->bi_iter.bi_sector);
1257 r10_bio->sectors = max_sectors;
1258 spin_lock_irq(&conf->device_lock);
1259 if (bio->bi_phys_segments == 0)
1260 bio->bi_phys_segments = 2;
1262 bio->bi_phys_segments++;
1263 spin_unlock_irq(&conf->device_lock);
1264 /* Cannot call generic_make_request directly
1265 * as that will be queued in __generic_make_request
1266 * and subsequent mempool_alloc might block
1267 * waiting for it. so hand bio over to raid10d.
1269 reschedule_retry(r10_bio);
1271 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1273 r10_bio->master_bio = bio;
1274 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1276 r10_bio->mddev = mddev;
1277 r10_bio->sector = bio->bi_iter.bi_sector +
1281 generic_make_request(read_bio);
1288 if (conf->pending_count >= max_queued_requests) {
1289 md_wakeup_thread(mddev->thread);
1290 wait_event(conf->wait_barrier,
1291 conf->pending_count < max_queued_requests);
1293 /* first select target devices under rcu_lock and
1294 * inc refcount on their rdev. Record them by setting
1296 * If there are known/acknowledged bad blocks on any device
1297 * on which we have seen a write error, we want to avoid
1298 * writing to those blocks. This potentially requires several
1299 * writes to write around the bad blocks. Each set of writes
1300 * gets its own r10_bio with a set of bios attached. The number
1301 * of r10_bios is recored in bio->bi_phys_segments just as with
1305 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1306 raid10_find_phys(conf, r10_bio);
1308 blocked_rdev = NULL;
1310 max_sectors = r10_bio->sectors;
1312 for (i = 0; i < conf->copies; i++) {
1313 int d = r10_bio->devs[i].devnum;
1314 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1315 struct md_rdev *rrdev = rcu_dereference(
1316 conf->mirrors[d].replacement);
1319 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1320 atomic_inc(&rdev->nr_pending);
1321 blocked_rdev = rdev;
1324 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1325 atomic_inc(&rrdev->nr_pending);
1326 blocked_rdev = rrdev;
1329 if (rdev && (test_bit(Faulty, &rdev->flags)
1330 || test_bit(Unmerged, &rdev->flags)))
1332 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1333 || test_bit(Unmerged, &rrdev->flags)))
1336 r10_bio->devs[i].bio = NULL;
1337 r10_bio->devs[i].repl_bio = NULL;
1339 if (!rdev && !rrdev) {
1340 set_bit(R10BIO_Degraded, &r10_bio->state);
1343 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1345 sector_t dev_sector = r10_bio->devs[i].addr;
1349 is_bad = is_badblock(rdev, dev_sector,
1351 &first_bad, &bad_sectors);
1353 /* Mustn't write here until the bad block
1356 atomic_inc(&rdev->nr_pending);
1357 set_bit(BlockedBadBlocks, &rdev->flags);
1358 blocked_rdev = rdev;
1361 if (is_bad && first_bad <= dev_sector) {
1362 /* Cannot write here at all */
1363 bad_sectors -= (dev_sector - first_bad);
1364 if (bad_sectors < max_sectors)
1365 /* Mustn't write more than bad_sectors
1366 * to other devices yet
1368 max_sectors = bad_sectors;
1369 /* We don't set R10BIO_Degraded as that
1370 * only applies if the disk is missing,
1371 * so it might be re-added, and we want to
1372 * know to recover this chunk.
1373 * In this case the device is here, and the
1374 * fact that this chunk is not in-sync is
1375 * recorded in the bad block log.
1380 int good_sectors = first_bad - dev_sector;
1381 if (good_sectors < max_sectors)
1382 max_sectors = good_sectors;
1386 r10_bio->devs[i].bio = bio;
1387 atomic_inc(&rdev->nr_pending);
1390 r10_bio->devs[i].repl_bio = bio;
1391 atomic_inc(&rrdev->nr_pending);
1396 if (unlikely(blocked_rdev)) {
1397 /* Have to wait for this device to get unblocked, then retry */
1401 for (j = 0; j < i; j++) {
1402 if (r10_bio->devs[j].bio) {
1403 d = r10_bio->devs[j].devnum;
1404 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1406 if (r10_bio->devs[j].repl_bio) {
1407 struct md_rdev *rdev;
1408 d = r10_bio->devs[j].devnum;
1409 rdev = conf->mirrors[d].replacement;
1411 /* Race with remove_disk */
1413 rdev = conf->mirrors[d].rdev;
1415 rdev_dec_pending(rdev, mddev);
1418 allow_barrier(conf);
1419 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1424 if (max_sectors < r10_bio->sectors) {
1425 /* We are splitting this into multiple parts, so
1426 * we need to prepare for allocating another r10_bio.
1428 r10_bio->sectors = max_sectors;
1429 spin_lock_irq(&conf->device_lock);
1430 if (bio->bi_phys_segments == 0)
1431 bio->bi_phys_segments = 2;
1433 bio->bi_phys_segments++;
1434 spin_unlock_irq(&conf->device_lock);
1436 sectors_handled = r10_bio->sector + max_sectors -
1437 bio->bi_iter.bi_sector;
1439 atomic_set(&r10_bio->remaining, 1);
1440 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1442 for (i = 0; i < conf->copies; i++) {
1444 int d = r10_bio->devs[i].devnum;
1445 if (r10_bio->devs[i].bio) {
1446 struct md_rdev *rdev = conf->mirrors[d].rdev;
1447 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1448 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1450 r10_bio->devs[i].bio = mbio;
1452 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1453 choose_data_offset(r10_bio,
1455 mbio->bi_bdev = rdev->bdev;
1456 mbio->bi_end_io = raid10_end_write_request;
1458 WRITE | do_sync | do_fua | do_discard | do_same;
1459 mbio->bi_private = r10_bio;
1461 atomic_inc(&r10_bio->remaining);
1463 cb = blk_check_plugged(raid10_unplug, mddev,
1466 plug = container_of(cb, struct raid10_plug_cb,
1470 spin_lock_irqsave(&conf->device_lock, flags);
1472 bio_list_add(&plug->pending, mbio);
1473 plug->pending_cnt++;
1475 bio_list_add(&conf->pending_bio_list, mbio);
1476 conf->pending_count++;
1478 spin_unlock_irqrestore(&conf->device_lock, flags);
1480 md_wakeup_thread(mddev->thread);
1483 if (r10_bio->devs[i].repl_bio) {
1484 struct md_rdev *rdev = conf->mirrors[d].replacement;
1486 /* Replacement just got moved to main 'rdev' */
1488 rdev = conf->mirrors[d].rdev;
1490 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1491 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1493 r10_bio->devs[i].repl_bio = mbio;
1495 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1498 mbio->bi_bdev = rdev->bdev;
1499 mbio->bi_end_io = raid10_end_write_request;
1501 WRITE | do_sync | do_fua | do_discard | do_same;
1502 mbio->bi_private = r10_bio;
1504 atomic_inc(&r10_bio->remaining);
1505 spin_lock_irqsave(&conf->device_lock, flags);
1506 bio_list_add(&conf->pending_bio_list, mbio);
1507 conf->pending_count++;
1508 spin_unlock_irqrestore(&conf->device_lock, flags);
1509 if (!mddev_check_plugged(mddev))
1510 md_wakeup_thread(mddev->thread);
1514 /* Don't remove the bias on 'remaining' (one_write_done) until
1515 * after checking if we need to go around again.
1518 if (sectors_handled < bio_sectors(bio)) {
1519 one_write_done(r10_bio);
1520 /* We need another r10_bio. It has already been counted
1521 * in bio->bi_phys_segments.
1523 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1525 r10_bio->master_bio = bio;
1526 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1528 r10_bio->mddev = mddev;
1529 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1533 one_write_done(r10_bio);
1536 static void make_request(struct mddev *mddev, struct bio *bio)
1538 struct r10conf *conf = mddev->private;
1539 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1540 int chunk_sects = chunk_mask + 1;
1544 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1545 md_flush_request(mddev, bio);
1549 md_write_start(mddev, bio);
1554 * If this request crosses a chunk boundary, we need to split
1557 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1558 bio_sectors(bio) > chunk_sects
1559 && (conf->geo.near_copies < conf->geo.raid_disks
1560 || conf->prev.near_copies <
1561 conf->prev.raid_disks))) {
1562 split = bio_split(bio, chunk_sects -
1563 (bio->bi_iter.bi_sector &
1565 GFP_NOIO, fs_bio_set);
1566 bio_chain(split, bio);
1571 __make_request(mddev, split);
1572 } while (split != bio);
1574 /* In case raid10d snuck in to freeze_array */
1575 wake_up(&conf->wait_barrier);
1578 static void status(struct seq_file *seq, struct mddev *mddev)
1580 struct r10conf *conf = mddev->private;
1583 if (conf->geo.near_copies < conf->geo.raid_disks)
1584 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1585 if (conf->geo.near_copies > 1)
1586 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1587 if (conf->geo.far_copies > 1) {
1588 if (conf->geo.far_offset)
1589 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1591 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1593 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1594 conf->geo.raid_disks - mddev->degraded);
1595 for (i = 0; i < conf->geo.raid_disks; i++)
1596 seq_printf(seq, "%s",
1597 conf->mirrors[i].rdev &&
1598 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1599 seq_printf(seq, "]");
1602 /* check if there are enough drives for
1603 * every block to appear on atleast one.
1604 * Don't consider the device numbered 'ignore'
1605 * as we might be about to remove it.
1607 static int _enough(struct r10conf *conf, int previous, int ignore)
1613 disks = conf->prev.raid_disks;
1614 ncopies = conf->prev.near_copies;
1616 disks = conf->geo.raid_disks;
1617 ncopies = conf->geo.near_copies;
1622 int n = conf->copies;
1626 struct md_rdev *rdev;
1627 if (this != ignore &&
1628 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1629 test_bit(In_sync, &rdev->flags))
1631 this = (this+1) % disks;
1635 first = (first + ncopies) % disks;
1636 } while (first != 0);
1643 static int enough(struct r10conf *conf, int ignore)
1645 /* when calling 'enough', both 'prev' and 'geo' must
1647 * This is ensured if ->reconfig_mutex or ->device_lock
1650 return _enough(conf, 0, ignore) &&
1651 _enough(conf, 1, ignore);
1654 static void error(struct mddev *mddev, struct md_rdev *rdev)
1656 char b[BDEVNAME_SIZE];
1657 struct r10conf *conf = mddev->private;
1658 unsigned long flags;
1661 * If it is not operational, then we have already marked it as dead
1662 * else if it is the last working disks, ignore the error, let the
1663 * next level up know.
1664 * else mark the drive as failed
1666 spin_lock_irqsave(&conf->device_lock, flags);
1667 if (test_bit(In_sync, &rdev->flags)
1668 && !enough(conf, rdev->raid_disk)) {
1670 * Don't fail the drive, just return an IO error.
1672 spin_unlock_irqrestore(&conf->device_lock, flags);
1675 if (test_and_clear_bit(In_sync, &rdev->flags))
1678 * If recovery is running, make sure it aborts.
1680 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1681 set_bit(Blocked, &rdev->flags);
1682 set_bit(Faulty, &rdev->flags);
1683 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1684 spin_unlock_irqrestore(&conf->device_lock, flags);
1686 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1687 "md/raid10:%s: Operation continuing on %d devices.\n",
1688 mdname(mddev), bdevname(rdev->bdev, b),
1689 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1692 static void print_conf(struct r10conf *conf)
1695 struct raid10_info *tmp;
1697 printk(KERN_DEBUG "RAID10 conf printout:\n");
1699 printk(KERN_DEBUG "(!conf)\n");
1702 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1703 conf->geo.raid_disks);
1705 for (i = 0; i < conf->geo.raid_disks; i++) {
1706 char b[BDEVNAME_SIZE];
1707 tmp = conf->mirrors + i;
1709 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1710 i, !test_bit(In_sync, &tmp->rdev->flags),
1711 !test_bit(Faulty, &tmp->rdev->flags),
1712 bdevname(tmp->rdev->bdev,b));
1716 static void close_sync(struct r10conf *conf)
1719 allow_barrier(conf);
1721 mempool_destroy(conf->r10buf_pool);
1722 conf->r10buf_pool = NULL;
1725 static int raid10_spare_active(struct mddev *mddev)
1728 struct r10conf *conf = mddev->private;
1729 struct raid10_info *tmp;
1731 unsigned long flags;
1734 * Find all non-in_sync disks within the RAID10 configuration
1735 * and mark them in_sync
1737 for (i = 0; i < conf->geo.raid_disks; i++) {
1738 tmp = conf->mirrors + i;
1739 if (tmp->replacement
1740 && tmp->replacement->recovery_offset == MaxSector
1741 && !test_bit(Faulty, &tmp->replacement->flags)
1742 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1743 /* Replacement has just become active */
1745 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1748 /* Replaced device not technically faulty,
1749 * but we need to be sure it gets removed
1750 * and never re-added.
1752 set_bit(Faulty, &tmp->rdev->flags);
1753 sysfs_notify_dirent_safe(
1754 tmp->rdev->sysfs_state);
1756 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1757 } else if (tmp->rdev
1758 && tmp->rdev->recovery_offset == MaxSector
1759 && !test_bit(Faulty, &tmp->rdev->flags)
1760 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1762 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1765 spin_lock_irqsave(&conf->device_lock, flags);
1766 mddev->degraded -= count;
1767 spin_unlock_irqrestore(&conf->device_lock, flags);
1773 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1775 struct r10conf *conf = mddev->private;
1779 int last = conf->geo.raid_disks - 1;
1780 struct request_queue *q = bdev_get_queue(rdev->bdev);
1782 if (mddev->recovery_cp < MaxSector)
1783 /* only hot-add to in-sync arrays, as recovery is
1784 * very different from resync
1787 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1790 if (rdev->raid_disk >= 0)
1791 first = last = rdev->raid_disk;
1793 if (q->merge_bvec_fn) {
1794 set_bit(Unmerged, &rdev->flags);
1795 mddev->merge_check_needed = 1;
1798 if (rdev->saved_raid_disk >= first &&
1799 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1800 mirror = rdev->saved_raid_disk;
1803 for ( ; mirror <= last ; mirror++) {
1804 struct raid10_info *p = &conf->mirrors[mirror];
1805 if (p->recovery_disabled == mddev->recovery_disabled)
1808 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1809 p->replacement != NULL)
1811 clear_bit(In_sync, &rdev->flags);
1812 set_bit(Replacement, &rdev->flags);
1813 rdev->raid_disk = mirror;
1816 disk_stack_limits(mddev->gendisk, rdev->bdev,
1817 rdev->data_offset << 9);
1819 rcu_assign_pointer(p->replacement, rdev);
1824 disk_stack_limits(mddev->gendisk, rdev->bdev,
1825 rdev->data_offset << 9);
1827 p->head_position = 0;
1828 p->recovery_disabled = mddev->recovery_disabled - 1;
1829 rdev->raid_disk = mirror;
1831 if (rdev->saved_raid_disk != mirror)
1833 rcu_assign_pointer(p->rdev, rdev);
1836 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1837 /* Some requests might not have seen this new
1838 * merge_bvec_fn. We must wait for them to complete
1839 * before merging the device fully.
1840 * First we make sure any code which has tested
1841 * our function has submitted the request, then
1842 * we wait for all outstanding requests to complete.
1844 synchronize_sched();
1845 freeze_array(conf, 0);
1846 unfreeze_array(conf);
1847 clear_bit(Unmerged, &rdev->flags);
1849 md_integrity_add_rdev(rdev, mddev);
1850 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1851 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1857 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1859 struct r10conf *conf = mddev->private;
1861 int number = rdev->raid_disk;
1862 struct md_rdev **rdevp;
1863 struct raid10_info *p = conf->mirrors + number;
1866 if (rdev == p->rdev)
1868 else if (rdev == p->replacement)
1869 rdevp = &p->replacement;
1873 if (test_bit(In_sync, &rdev->flags) ||
1874 atomic_read(&rdev->nr_pending)) {
1878 /* Only remove faulty devices if recovery
1881 if (!test_bit(Faulty, &rdev->flags) &&
1882 mddev->recovery_disabled != p->recovery_disabled &&
1883 (!p->replacement || p->replacement == rdev) &&
1884 number < conf->geo.raid_disks &&
1891 if (atomic_read(&rdev->nr_pending)) {
1892 /* lost the race, try later */
1896 } else if (p->replacement) {
1897 /* We must have just cleared 'rdev' */
1898 p->rdev = p->replacement;
1899 clear_bit(Replacement, &p->replacement->flags);
1900 smp_mb(); /* Make sure other CPUs may see both as identical
1901 * but will never see neither -- if they are careful.
1903 p->replacement = NULL;
1904 clear_bit(WantReplacement, &rdev->flags);
1906 /* We might have just remove the Replacement as faulty
1907 * Clear the flag just in case
1909 clear_bit(WantReplacement, &rdev->flags);
1911 err = md_integrity_register(mddev);
1919 static void end_sync_read(struct bio *bio, int error)
1921 struct r10bio *r10_bio = bio->bi_private;
1922 struct r10conf *conf = r10_bio->mddev->private;
1925 if (bio == r10_bio->master_bio) {
1926 /* this is a reshape read */
1927 d = r10_bio->read_slot; /* really the read dev */
1929 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1931 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1932 set_bit(R10BIO_Uptodate, &r10_bio->state);
1934 /* The write handler will notice the lack of
1935 * R10BIO_Uptodate and record any errors etc
1937 atomic_add(r10_bio->sectors,
1938 &conf->mirrors[d].rdev->corrected_errors);
1940 /* for reconstruct, we always reschedule after a read.
1941 * for resync, only after all reads
1943 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1944 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1945 atomic_dec_and_test(&r10_bio->remaining)) {
1946 /* we have read all the blocks,
1947 * do the comparison in process context in raid10d
1949 reschedule_retry(r10_bio);
1953 static void end_sync_request(struct r10bio *r10_bio)
1955 struct mddev *mddev = r10_bio->mddev;
1957 while (atomic_dec_and_test(&r10_bio->remaining)) {
1958 if (r10_bio->master_bio == NULL) {
1959 /* the primary of several recovery bios */
1960 sector_t s = r10_bio->sectors;
1961 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1962 test_bit(R10BIO_WriteError, &r10_bio->state))
1963 reschedule_retry(r10_bio);
1966 md_done_sync(mddev, s, 1);
1969 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1970 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1971 test_bit(R10BIO_WriteError, &r10_bio->state))
1972 reschedule_retry(r10_bio);
1980 static void end_sync_write(struct bio *bio, int error)
1982 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1983 struct r10bio *r10_bio = bio->bi_private;
1984 struct mddev *mddev = r10_bio->mddev;
1985 struct r10conf *conf = mddev->private;
1991 struct md_rdev *rdev = NULL;
1993 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1995 rdev = conf->mirrors[d].replacement;
1997 rdev = conf->mirrors[d].rdev;
2001 md_error(mddev, rdev);
2003 set_bit(WriteErrorSeen, &rdev->flags);
2004 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2005 set_bit(MD_RECOVERY_NEEDED,
2006 &rdev->mddev->recovery);
2007 set_bit(R10BIO_WriteError, &r10_bio->state);
2009 } else if (is_badblock(rdev,
2010 r10_bio->devs[slot].addr,
2012 &first_bad, &bad_sectors))
2013 set_bit(R10BIO_MadeGood, &r10_bio->state);
2015 rdev_dec_pending(rdev, mddev);
2017 end_sync_request(r10_bio);
2021 * Note: sync and recover and handled very differently for raid10
2022 * This code is for resync.
2023 * For resync, we read through virtual addresses and read all blocks.
2024 * If there is any error, we schedule a write. The lowest numbered
2025 * drive is authoritative.
2026 * However requests come for physical address, so we need to map.
2027 * For every physical address there are raid_disks/copies virtual addresses,
2028 * which is always are least one, but is not necessarly an integer.
2029 * This means that a physical address can span multiple chunks, so we may
2030 * have to submit multiple io requests for a single sync request.
2033 * We check if all blocks are in-sync and only write to blocks that
2036 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2038 struct r10conf *conf = mddev->private;
2040 struct bio *tbio, *fbio;
2043 atomic_set(&r10_bio->remaining, 1);
2045 /* find the first device with a block */
2046 for (i=0; i<conf->copies; i++)
2047 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2050 if (i == conf->copies)
2054 fbio = r10_bio->devs[i].bio;
2056 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2057 /* now find blocks with errors */
2058 for (i=0 ; i < conf->copies ; i++) {
2061 tbio = r10_bio->devs[i].bio;
2063 if (tbio->bi_end_io != end_sync_read)
2067 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2068 /* We know that the bi_io_vec layout is the same for
2069 * both 'first' and 'i', so we just compare them.
2070 * All vec entries are PAGE_SIZE;
2072 int sectors = r10_bio->sectors;
2073 for (j = 0; j < vcnt; j++) {
2074 int len = PAGE_SIZE;
2075 if (sectors < (len / 512))
2076 len = sectors * 512;
2077 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2078 page_address(tbio->bi_io_vec[j].bv_page),
2085 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2086 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2087 /* Don't fix anything. */
2090 /* Ok, we need to write this bio, either to correct an
2091 * inconsistency or to correct an unreadable block.
2092 * First we need to fixup bv_offset, bv_len and
2093 * bi_vecs, as the read request might have corrupted these
2097 tbio->bi_vcnt = vcnt;
2098 tbio->bi_iter.bi_size = r10_bio->sectors << 9;
2099 tbio->bi_rw = WRITE;
2100 tbio->bi_private = r10_bio;
2101 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2103 for (j=0; j < vcnt ; j++) {
2104 tbio->bi_io_vec[j].bv_offset = 0;
2105 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2107 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2108 page_address(fbio->bi_io_vec[j].bv_page),
2111 tbio->bi_end_io = end_sync_write;
2113 d = r10_bio->devs[i].devnum;
2114 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2115 atomic_inc(&r10_bio->remaining);
2116 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2118 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2119 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2120 generic_make_request(tbio);
2123 /* Now write out to any replacement devices
2126 for (i = 0; i < conf->copies; i++) {
2129 tbio = r10_bio->devs[i].repl_bio;
2130 if (!tbio || !tbio->bi_end_io)
2132 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2133 && r10_bio->devs[i].bio != fbio)
2134 for (j = 0; j < vcnt; j++)
2135 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2136 page_address(fbio->bi_io_vec[j].bv_page),
2138 d = r10_bio->devs[i].devnum;
2139 atomic_inc(&r10_bio->remaining);
2140 md_sync_acct(conf->mirrors[d].replacement->bdev,
2142 generic_make_request(tbio);
2146 if (atomic_dec_and_test(&r10_bio->remaining)) {
2147 md_done_sync(mddev, r10_bio->sectors, 1);
2153 * Now for the recovery code.
2154 * Recovery happens across physical sectors.
2155 * We recover all non-is_sync drives by finding the virtual address of
2156 * each, and then choose a working drive that also has that virt address.
2157 * There is a separate r10_bio for each non-in_sync drive.
2158 * Only the first two slots are in use. The first for reading,
2159 * The second for writing.
2162 static void fix_recovery_read_error(struct r10bio *r10_bio)
2164 /* We got a read error during recovery.
2165 * We repeat the read in smaller page-sized sections.
2166 * If a read succeeds, write it to the new device or record
2167 * a bad block if we cannot.
2168 * If a read fails, record a bad block on both old and
2171 struct mddev *mddev = r10_bio->mddev;
2172 struct r10conf *conf = mddev->private;
2173 struct bio *bio = r10_bio->devs[0].bio;
2175 int sectors = r10_bio->sectors;
2177 int dr = r10_bio->devs[0].devnum;
2178 int dw = r10_bio->devs[1].devnum;
2182 struct md_rdev *rdev;
2186 if (s > (PAGE_SIZE>>9))
2189 rdev = conf->mirrors[dr].rdev;
2190 addr = r10_bio->devs[0].addr + sect,
2191 ok = sync_page_io(rdev,
2194 bio->bi_io_vec[idx].bv_page,
2197 rdev = conf->mirrors[dw].rdev;
2198 addr = r10_bio->devs[1].addr + sect;
2199 ok = sync_page_io(rdev,
2202 bio->bi_io_vec[idx].bv_page,
2205 set_bit(WriteErrorSeen, &rdev->flags);
2206 if (!test_and_set_bit(WantReplacement,
2208 set_bit(MD_RECOVERY_NEEDED,
2209 &rdev->mddev->recovery);
2213 /* We don't worry if we cannot set a bad block -
2214 * it really is bad so there is no loss in not
2217 rdev_set_badblocks(rdev, addr, s, 0);
2219 if (rdev != conf->mirrors[dw].rdev) {
2220 /* need bad block on destination too */
2221 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2222 addr = r10_bio->devs[1].addr + sect;
2223 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2225 /* just abort the recovery */
2227 "md/raid10:%s: recovery aborted"
2228 " due to read error\n",
2231 conf->mirrors[dw].recovery_disabled
2232 = mddev->recovery_disabled;
2233 set_bit(MD_RECOVERY_INTR,
2246 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2248 struct r10conf *conf = mddev->private;
2250 struct bio *wbio, *wbio2;
2252 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2253 fix_recovery_read_error(r10_bio);
2254 end_sync_request(r10_bio);
2259 * share the pages with the first bio
2260 * and submit the write request
2262 d = r10_bio->devs[1].devnum;
2263 wbio = r10_bio->devs[1].bio;
2264 wbio2 = r10_bio->devs[1].repl_bio;
2265 /* Need to test wbio2->bi_end_io before we call
2266 * generic_make_request as if the former is NULL,
2267 * the latter is free to free wbio2.
2269 if (wbio2 && !wbio2->bi_end_io)
2271 if (wbio->bi_end_io) {
2272 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2273 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2274 generic_make_request(wbio);
2277 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2278 md_sync_acct(conf->mirrors[d].replacement->bdev,
2279 bio_sectors(wbio2));
2280 generic_make_request(wbio2);
2285 * Used by fix_read_error() to decay the per rdev read_errors.
2286 * We halve the read error count for every hour that has elapsed
2287 * since the last recorded read error.
2290 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2292 struct timespec cur_time_mon;
2293 unsigned long hours_since_last;
2294 unsigned int read_errors = atomic_read(&rdev->read_errors);
2296 ktime_get_ts(&cur_time_mon);
2298 if (rdev->last_read_error.tv_sec == 0 &&
2299 rdev->last_read_error.tv_nsec == 0) {
2300 /* first time we've seen a read error */
2301 rdev->last_read_error = cur_time_mon;
2305 hours_since_last = (cur_time_mon.tv_sec -
2306 rdev->last_read_error.tv_sec) / 3600;
2308 rdev->last_read_error = cur_time_mon;
2311 * if hours_since_last is > the number of bits in read_errors
2312 * just set read errors to 0. We do this to avoid
2313 * overflowing the shift of read_errors by hours_since_last.
2315 if (hours_since_last >= 8 * sizeof(read_errors))
2316 atomic_set(&rdev->read_errors, 0);
2318 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2321 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2322 int sectors, struct page *page, int rw)
2327 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2328 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2330 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2334 set_bit(WriteErrorSeen, &rdev->flags);
2335 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2336 set_bit(MD_RECOVERY_NEEDED,
2337 &rdev->mddev->recovery);
2339 /* need to record an error - either for the block or the device */
2340 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2341 md_error(rdev->mddev, rdev);
2346 * This is a kernel thread which:
2348 * 1. Retries failed read operations on working mirrors.
2349 * 2. Updates the raid superblock when problems encounter.
2350 * 3. Performs writes following reads for array synchronising.
2353 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2355 int sect = 0; /* Offset from r10_bio->sector */
2356 int sectors = r10_bio->sectors;
2357 struct md_rdev*rdev;
2358 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2359 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2361 /* still own a reference to this rdev, so it cannot
2362 * have been cleared recently.
2364 rdev = conf->mirrors[d].rdev;
2366 if (test_bit(Faulty, &rdev->flags))
2367 /* drive has already been failed, just ignore any
2368 more fix_read_error() attempts */
2371 check_decay_read_errors(mddev, rdev);
2372 atomic_inc(&rdev->read_errors);
2373 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2374 char b[BDEVNAME_SIZE];
2375 bdevname(rdev->bdev, b);
2378 "md/raid10:%s: %s: Raid device exceeded "
2379 "read_error threshold [cur %d:max %d]\n",
2381 atomic_read(&rdev->read_errors), max_read_errors);
2383 "md/raid10:%s: %s: Failing raid device\n",
2385 md_error(mddev, conf->mirrors[d].rdev);
2386 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2392 int sl = r10_bio->read_slot;
2396 if (s > (PAGE_SIZE>>9))
2404 d = r10_bio->devs[sl].devnum;
2405 rdev = rcu_dereference(conf->mirrors[d].rdev);
2407 !test_bit(Unmerged, &rdev->flags) &&
2408 test_bit(In_sync, &rdev->flags) &&
2409 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2410 &first_bad, &bad_sectors) == 0) {
2411 atomic_inc(&rdev->nr_pending);
2413 success = sync_page_io(rdev,
2414 r10_bio->devs[sl].addr +
2417 conf->tmppage, READ, false);
2418 rdev_dec_pending(rdev, mddev);
2424 if (sl == conf->copies)
2426 } while (!success && sl != r10_bio->read_slot);
2430 /* Cannot read from anywhere, just mark the block
2431 * as bad on the first device to discourage future
2434 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2435 rdev = conf->mirrors[dn].rdev;
2437 if (!rdev_set_badblocks(
2439 r10_bio->devs[r10_bio->read_slot].addr
2442 md_error(mddev, rdev);
2443 r10_bio->devs[r10_bio->read_slot].bio
2450 /* write it back and re-read */
2452 while (sl != r10_bio->read_slot) {
2453 char b[BDEVNAME_SIZE];
2458 d = r10_bio->devs[sl].devnum;
2459 rdev = rcu_dereference(conf->mirrors[d].rdev);
2461 test_bit(Unmerged, &rdev->flags) ||
2462 !test_bit(In_sync, &rdev->flags))
2465 atomic_inc(&rdev->nr_pending);
2467 if (r10_sync_page_io(rdev,
2468 r10_bio->devs[sl].addr +
2470 s, conf->tmppage, WRITE)
2472 /* Well, this device is dead */
2474 "md/raid10:%s: read correction "
2476 " (%d sectors at %llu on %s)\n",
2478 (unsigned long long)(
2480 choose_data_offset(r10_bio,
2482 bdevname(rdev->bdev, b));
2483 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2486 bdevname(rdev->bdev, b));
2488 rdev_dec_pending(rdev, mddev);
2492 while (sl != r10_bio->read_slot) {
2493 char b[BDEVNAME_SIZE];
2498 d = r10_bio->devs[sl].devnum;
2499 rdev = rcu_dereference(conf->mirrors[d].rdev);
2501 !test_bit(In_sync, &rdev->flags))
2504 atomic_inc(&rdev->nr_pending);
2506 switch (r10_sync_page_io(rdev,
2507 r10_bio->devs[sl].addr +
2512 /* Well, this device is dead */
2514 "md/raid10:%s: unable to read back "
2516 " (%d sectors at %llu on %s)\n",
2518 (unsigned long long)(
2520 choose_data_offset(r10_bio, rdev)),
2521 bdevname(rdev->bdev, b));
2522 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2525 bdevname(rdev->bdev, b));
2529 "md/raid10:%s: read error corrected"
2530 " (%d sectors at %llu on %s)\n",
2532 (unsigned long long)(
2534 choose_data_offset(r10_bio, rdev)),
2535 bdevname(rdev->bdev, b));
2536 atomic_add(s, &rdev->corrected_errors);
2539 rdev_dec_pending(rdev, mddev);
2549 static int narrow_write_error(struct r10bio *r10_bio, int i)
2551 struct bio *bio = r10_bio->master_bio;
2552 struct mddev *mddev = r10_bio->mddev;
2553 struct r10conf *conf = mddev->private;
2554 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2555 /* bio has the data to be written to slot 'i' where
2556 * we just recently had a write error.
2557 * We repeatedly clone the bio and trim down to one block,
2558 * then try the write. Where the write fails we record
2560 * It is conceivable that the bio doesn't exactly align with
2561 * blocks. We must handle this.
2563 * We currently own a reference to the rdev.
2569 int sect_to_write = r10_bio->sectors;
2572 if (rdev->badblocks.shift < 0)
2575 block_sectors = roundup(1 << rdev->badblocks.shift,
2576 bdev_logical_block_size(rdev->bdev) >> 9);
2577 sector = r10_bio->sector;
2578 sectors = ((r10_bio->sector + block_sectors)
2579 & ~(sector_t)(block_sectors - 1))
2582 while (sect_to_write) {
2584 if (sectors > sect_to_write)
2585 sectors = sect_to_write;
2586 /* Write at 'sector' for 'sectors' */
2587 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2588 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2589 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2590 choose_data_offset(r10_bio, rdev) +
2591 (sector - r10_bio->sector));
2592 wbio->bi_bdev = rdev->bdev;
2593 if (submit_bio_wait(WRITE, wbio) == 0)
2595 ok = rdev_set_badblocks(rdev, sector,
2600 sect_to_write -= sectors;
2602 sectors = block_sectors;
2607 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2609 int slot = r10_bio->read_slot;
2611 struct r10conf *conf = mddev->private;
2612 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2613 char b[BDEVNAME_SIZE];
2614 unsigned long do_sync;
2617 /* we got a read error. Maybe the drive is bad. Maybe just
2618 * the block and we can fix it.
2619 * We freeze all other IO, and try reading the block from
2620 * other devices. When we find one, we re-write
2621 * and check it that fixes the read error.
2622 * This is all done synchronously while the array is
2625 bio = r10_bio->devs[slot].bio;
2626 bdevname(bio->bi_bdev, b);
2628 r10_bio->devs[slot].bio = NULL;
2630 if (mddev->ro == 0) {
2631 freeze_array(conf, 1);
2632 fix_read_error(conf, mddev, r10_bio);
2633 unfreeze_array(conf);
2635 r10_bio->devs[slot].bio = IO_BLOCKED;
2637 rdev_dec_pending(rdev, mddev);
2640 rdev = read_balance(conf, r10_bio, &max_sectors);
2642 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2643 " read error for block %llu\n",
2645 (unsigned long long)r10_bio->sector);
2646 raid_end_bio_io(r10_bio);
2650 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2651 slot = r10_bio->read_slot;
2654 "md/raid10:%s: %s: redirecting "
2655 "sector %llu to another mirror\n",
2657 bdevname(rdev->bdev, b),
2658 (unsigned long long)r10_bio->sector);
2659 bio = bio_clone_mddev(r10_bio->master_bio,
2661 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2662 r10_bio->devs[slot].bio = bio;
2663 r10_bio->devs[slot].rdev = rdev;
2664 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2665 + choose_data_offset(r10_bio, rdev);
2666 bio->bi_bdev = rdev->bdev;
2667 bio->bi_rw = READ | do_sync;
2668 bio->bi_private = r10_bio;
2669 bio->bi_end_io = raid10_end_read_request;
2670 if (max_sectors < r10_bio->sectors) {
2671 /* Drat - have to split this up more */
2672 struct bio *mbio = r10_bio->master_bio;
2673 int sectors_handled =
2674 r10_bio->sector + max_sectors
2675 - mbio->bi_iter.bi_sector;
2676 r10_bio->sectors = max_sectors;
2677 spin_lock_irq(&conf->device_lock);
2678 if (mbio->bi_phys_segments == 0)
2679 mbio->bi_phys_segments = 2;
2681 mbio->bi_phys_segments++;
2682 spin_unlock_irq(&conf->device_lock);
2683 generic_make_request(bio);
2685 r10_bio = mempool_alloc(conf->r10bio_pool,
2687 r10_bio->master_bio = mbio;
2688 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2690 set_bit(R10BIO_ReadError,
2692 r10_bio->mddev = mddev;
2693 r10_bio->sector = mbio->bi_iter.bi_sector
2698 generic_make_request(bio);
2701 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2703 /* Some sort of write request has finished and it
2704 * succeeded in writing where we thought there was a
2705 * bad block. So forget the bad block.
2706 * Or possibly if failed and we need to record
2710 struct md_rdev *rdev;
2712 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2713 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2714 for (m = 0; m < conf->copies; m++) {
2715 int dev = r10_bio->devs[m].devnum;
2716 rdev = conf->mirrors[dev].rdev;
2717 if (r10_bio->devs[m].bio == NULL)
2719 if (test_bit(BIO_UPTODATE,
2720 &r10_bio->devs[m].bio->bi_flags)) {
2721 rdev_clear_badblocks(
2723 r10_bio->devs[m].addr,
2724 r10_bio->sectors, 0);
2726 if (!rdev_set_badblocks(
2728 r10_bio->devs[m].addr,
2729 r10_bio->sectors, 0))
2730 md_error(conf->mddev, rdev);
2732 rdev = conf->mirrors[dev].replacement;
2733 if (r10_bio->devs[m].repl_bio == NULL)
2735 if (test_bit(BIO_UPTODATE,
2736 &r10_bio->devs[m].repl_bio->bi_flags)) {
2737 rdev_clear_badblocks(
2739 r10_bio->devs[m].addr,
2740 r10_bio->sectors, 0);
2742 if (!rdev_set_badblocks(
2744 r10_bio->devs[m].addr,
2745 r10_bio->sectors, 0))
2746 md_error(conf->mddev, rdev);
2751 for (m = 0; m < conf->copies; m++) {
2752 int dev = r10_bio->devs[m].devnum;
2753 struct bio *bio = r10_bio->devs[m].bio;
2754 rdev = conf->mirrors[dev].rdev;
2755 if (bio == IO_MADE_GOOD) {
2756 rdev_clear_badblocks(
2758 r10_bio->devs[m].addr,
2759 r10_bio->sectors, 0);
2760 rdev_dec_pending(rdev, conf->mddev);
2761 } else if (bio != NULL &&
2762 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2763 if (!narrow_write_error(r10_bio, m)) {
2764 md_error(conf->mddev, rdev);
2765 set_bit(R10BIO_Degraded,
2768 rdev_dec_pending(rdev, conf->mddev);
2770 bio = r10_bio->devs[m].repl_bio;
2771 rdev = conf->mirrors[dev].replacement;
2772 if (rdev && bio == IO_MADE_GOOD) {
2773 rdev_clear_badblocks(
2775 r10_bio->devs[m].addr,
2776 r10_bio->sectors, 0);
2777 rdev_dec_pending(rdev, conf->mddev);
2780 if (test_bit(R10BIO_WriteError,
2782 close_write(r10_bio);
2783 raid_end_bio_io(r10_bio);
2787 static void raid10d(struct md_thread *thread)
2789 struct mddev *mddev = thread->mddev;
2790 struct r10bio *r10_bio;
2791 unsigned long flags;
2792 struct r10conf *conf = mddev->private;
2793 struct list_head *head = &conf->retry_list;
2794 struct blk_plug plug;
2796 md_check_recovery(mddev);
2798 blk_start_plug(&plug);
2801 flush_pending_writes(conf);
2803 spin_lock_irqsave(&conf->device_lock, flags);
2804 if (list_empty(head)) {
2805 spin_unlock_irqrestore(&conf->device_lock, flags);
2808 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2809 list_del(head->prev);
2811 spin_unlock_irqrestore(&conf->device_lock, flags);
2813 mddev = r10_bio->mddev;
2814 conf = mddev->private;
2815 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2816 test_bit(R10BIO_WriteError, &r10_bio->state))
2817 handle_write_completed(conf, r10_bio);
2818 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2819 reshape_request_write(mddev, r10_bio);
2820 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2821 sync_request_write(mddev, r10_bio);
2822 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2823 recovery_request_write(mddev, r10_bio);
2824 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2825 handle_read_error(mddev, r10_bio);
2827 /* just a partial read to be scheduled from a
2830 int slot = r10_bio->read_slot;
2831 generic_make_request(r10_bio->devs[slot].bio);
2835 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2836 md_check_recovery(mddev);
2838 blk_finish_plug(&plug);
2841 static int init_resync(struct r10conf *conf)
2846 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2847 BUG_ON(conf->r10buf_pool);
2848 conf->have_replacement = 0;
2849 for (i = 0; i < conf->geo.raid_disks; i++)
2850 if (conf->mirrors[i].replacement)
2851 conf->have_replacement = 1;
2852 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2853 if (!conf->r10buf_pool)
2855 conf->next_resync = 0;
2860 * perform a "sync" on one "block"
2862 * We need to make sure that no normal I/O request - particularly write
2863 * requests - conflict with active sync requests.
2865 * This is achieved by tracking pending requests and a 'barrier' concept
2866 * that can be installed to exclude normal IO requests.
2868 * Resync and recovery are handled very differently.
2869 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2871 * For resync, we iterate over virtual addresses, read all copies,
2872 * and update if there are differences. If only one copy is live,
2874 * For recovery, we iterate over physical addresses, read a good
2875 * value for each non-in_sync drive, and over-write.
2877 * So, for recovery we may have several outstanding complex requests for a
2878 * given address, one for each out-of-sync device. We model this by allocating
2879 * a number of r10_bio structures, one for each out-of-sync device.
2880 * As we setup these structures, we collect all bio's together into a list
2881 * which we then process collectively to add pages, and then process again
2882 * to pass to generic_make_request.
2884 * The r10_bio structures are linked using a borrowed master_bio pointer.
2885 * This link is counted in ->remaining. When the r10_bio that points to NULL
2886 * has its remaining count decremented to 0, the whole complex operation
2891 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2894 struct r10conf *conf = mddev->private;
2895 struct r10bio *r10_bio;
2896 struct bio *biolist = NULL, *bio;
2897 sector_t max_sector, nr_sectors;
2900 sector_t sync_blocks;
2901 sector_t sectors_skipped = 0;
2902 int chunks_skipped = 0;
2903 sector_t chunk_mask = conf->geo.chunk_mask;
2905 if (!conf->r10buf_pool)
2906 if (init_resync(conf))
2910 * Allow skipping a full rebuild for incremental assembly
2911 * of a clean array, like RAID1 does.
2913 if (mddev->bitmap == NULL &&
2914 mddev->recovery_cp == MaxSector &&
2915 mddev->reshape_position == MaxSector &&
2916 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2917 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2918 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2919 conf->fullsync == 0) {
2921 return mddev->dev_sectors - sector_nr;
2925 max_sector = mddev->dev_sectors;
2926 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2927 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2928 max_sector = mddev->resync_max_sectors;
2929 if (sector_nr >= max_sector) {
2930 /* If we aborted, we need to abort the
2931 * sync on the 'current' bitmap chucks (there can
2932 * be several when recovering multiple devices).
2933 * as we may have started syncing it but not finished.
2934 * We can find the current address in
2935 * mddev->curr_resync, but for recovery,
2936 * we need to convert that to several
2937 * virtual addresses.
2939 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2945 if (mddev->curr_resync < max_sector) { /* aborted */
2946 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2947 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2949 else for (i = 0; i < conf->geo.raid_disks; i++) {
2951 raid10_find_virt(conf, mddev->curr_resync, i);
2952 bitmap_end_sync(mddev->bitmap, sect,
2956 /* completed sync */
2957 if ((!mddev->bitmap || conf->fullsync)
2958 && conf->have_replacement
2959 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2960 /* Completed a full sync so the replacements
2961 * are now fully recovered.
2963 for (i = 0; i < conf->geo.raid_disks; i++)
2964 if (conf->mirrors[i].replacement)
2965 conf->mirrors[i].replacement
2971 bitmap_close_sync(mddev->bitmap);
2974 return sectors_skipped;
2977 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2978 return reshape_request(mddev, sector_nr, skipped);
2980 if (chunks_skipped >= conf->geo.raid_disks) {
2981 /* if there has been nothing to do on any drive,
2982 * then there is nothing to do at all..
2985 return (max_sector - sector_nr) + sectors_skipped;
2988 if (max_sector > mddev->resync_max)
2989 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2991 /* make sure whole request will fit in a chunk - if chunks
2994 if (conf->geo.near_copies < conf->geo.raid_disks &&
2995 max_sector > (sector_nr | chunk_mask))
2996 max_sector = (sector_nr | chunk_mask) + 1;
2998 /* Again, very different code for resync and recovery.
2999 * Both must result in an r10bio with a list of bios that
3000 * have bi_end_io, bi_sector, bi_bdev set,
3001 * and bi_private set to the r10bio.
3002 * For recovery, we may actually create several r10bios
3003 * with 2 bios in each, that correspond to the bios in the main one.
3004 * In this case, the subordinate r10bios link back through a
3005 * borrowed master_bio pointer, and the counter in the master
3006 * includes a ref from each subordinate.
3008 /* First, we decide what to do and set ->bi_end_io
3009 * To end_sync_read if we want to read, and
3010 * end_sync_write if we will want to write.
3013 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3014 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3015 /* recovery... the complicated one */
3019 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3025 struct raid10_info *mirror = &conf->mirrors[i];
3027 if ((mirror->rdev == NULL ||
3028 test_bit(In_sync, &mirror->rdev->flags))
3030 (mirror->replacement == NULL ||
3032 &mirror->replacement->flags)))
3036 /* want to reconstruct this device */
3038 sect = raid10_find_virt(conf, sector_nr, i);
3039 if (sect >= mddev->resync_max_sectors) {
3040 /* last stripe is not complete - don't
3041 * try to recover this sector.
3045 /* Unless we are doing a full sync, or a replacement
3046 * we only need to recover the block if it is set in
3049 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3051 if (sync_blocks < max_sync)
3052 max_sync = sync_blocks;
3054 mirror->replacement == NULL &&
3056 /* yep, skip the sync_blocks here, but don't assume
3057 * that there will never be anything to do here
3059 chunks_skipped = -1;
3063 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3065 raise_barrier(conf, rb2 != NULL);
3066 atomic_set(&r10_bio->remaining, 0);
3068 r10_bio->master_bio = (struct bio*)rb2;
3070 atomic_inc(&rb2->remaining);
3071 r10_bio->mddev = mddev;
3072 set_bit(R10BIO_IsRecover, &r10_bio->state);
3073 r10_bio->sector = sect;
3075 raid10_find_phys(conf, r10_bio);
3077 /* Need to check if the array will still be
3080 for (j = 0; j < conf->geo.raid_disks; j++)
3081 if (conf->mirrors[j].rdev == NULL ||
3082 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3087 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3088 &sync_blocks, still_degraded);
3091 for (j=0; j<conf->copies;j++) {
3093 int d = r10_bio->devs[j].devnum;
3094 sector_t from_addr, to_addr;
3095 struct md_rdev *rdev;
3096 sector_t sector, first_bad;
3098 if (!conf->mirrors[d].rdev ||
3099 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3101 /* This is where we read from */
3103 rdev = conf->mirrors[d].rdev;
3104 sector = r10_bio->devs[j].addr;
3106 if (is_badblock(rdev, sector, max_sync,
3107 &first_bad, &bad_sectors)) {
3108 if (first_bad > sector)
3109 max_sync = first_bad - sector;
3111 bad_sectors -= (sector
3113 if (max_sync > bad_sectors)
3114 max_sync = bad_sectors;
3118 bio = r10_bio->devs[0].bio;
3120 bio->bi_next = biolist;
3122 bio->bi_private = r10_bio;
3123 bio->bi_end_io = end_sync_read;
3125 from_addr = r10_bio->devs[j].addr;
3126 bio->bi_iter.bi_sector = from_addr +
3128 bio->bi_bdev = rdev->bdev;
3129 atomic_inc(&rdev->nr_pending);
3130 /* and we write to 'i' (if not in_sync) */
3132 for (k=0; k<conf->copies; k++)
3133 if (r10_bio->devs[k].devnum == i)
3135 BUG_ON(k == conf->copies);
3136 to_addr = r10_bio->devs[k].addr;
3137 r10_bio->devs[0].devnum = d;
3138 r10_bio->devs[0].addr = from_addr;
3139 r10_bio->devs[1].devnum = i;
3140 r10_bio->devs[1].addr = to_addr;
3142 rdev = mirror->rdev;
3143 if (!test_bit(In_sync, &rdev->flags)) {
3144 bio = r10_bio->devs[1].bio;
3146 bio->bi_next = biolist;
3148 bio->bi_private = r10_bio;
3149 bio->bi_end_io = end_sync_write;
3151 bio->bi_iter.bi_sector = to_addr
3152 + rdev->data_offset;
3153 bio->bi_bdev = rdev->bdev;
3154 atomic_inc(&r10_bio->remaining);
3156 r10_bio->devs[1].bio->bi_end_io = NULL;
3158 /* and maybe write to replacement */
3159 bio = r10_bio->devs[1].repl_bio;
3161 bio->bi_end_io = NULL;
3162 rdev = mirror->replacement;
3163 /* Note: if rdev != NULL, then bio
3164 * cannot be NULL as r10buf_pool_alloc will
3165 * have allocated it.
3166 * So the second test here is pointless.
3167 * But it keeps semantic-checkers happy, and
3168 * this comment keeps human reviewers
3171 if (rdev == NULL || bio == NULL ||
3172 test_bit(Faulty, &rdev->flags))
3175 bio->bi_next = biolist;
3177 bio->bi_private = r10_bio;
3178 bio->bi_end_io = end_sync_write;
3180 bio->bi_iter.bi_sector = to_addr +
3182 bio->bi_bdev = rdev->bdev;
3183 atomic_inc(&r10_bio->remaining);
3186 if (j == conf->copies) {
3187 /* Cannot recover, so abort the recovery or
3188 * record a bad block */
3190 /* problem is that there are bad blocks
3191 * on other device(s)
3194 for (k = 0; k < conf->copies; k++)
3195 if (r10_bio->devs[k].devnum == i)
3197 if (!test_bit(In_sync,
3198 &mirror->rdev->flags)
3199 && !rdev_set_badblocks(
3201 r10_bio->devs[k].addr,
3204 if (mirror->replacement &&
3205 !rdev_set_badblocks(
3206 mirror->replacement,
3207 r10_bio->devs[k].addr,
3212 if (!test_and_set_bit(MD_RECOVERY_INTR,
3214 printk(KERN_INFO "md/raid10:%s: insufficient "
3215 "working devices for recovery.\n",
3217 mirror->recovery_disabled
3218 = mddev->recovery_disabled;
3222 atomic_dec(&rb2->remaining);
3227 if (biolist == NULL) {
3229 struct r10bio *rb2 = r10_bio;
3230 r10_bio = (struct r10bio*) rb2->master_bio;
3231 rb2->master_bio = NULL;
3237 /* resync. Schedule a read for every block at this virt offset */
3240 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3242 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3243 &sync_blocks, mddev->degraded) &&
3244 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3245 &mddev->recovery)) {
3246 /* We can skip this block */
3248 return sync_blocks + sectors_skipped;
3250 if (sync_blocks < max_sync)
3251 max_sync = sync_blocks;
3252 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3255 r10_bio->mddev = mddev;
3256 atomic_set(&r10_bio->remaining, 0);
3257 raise_barrier(conf, 0);
3258 conf->next_resync = sector_nr;
3260 r10_bio->master_bio = NULL;
3261 r10_bio->sector = sector_nr;
3262 set_bit(R10BIO_IsSync, &r10_bio->state);
3263 raid10_find_phys(conf, r10_bio);
3264 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3266 for (i = 0; i < conf->copies; i++) {
3267 int d = r10_bio->devs[i].devnum;
3268 sector_t first_bad, sector;
3271 if (r10_bio->devs[i].repl_bio)
3272 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3274 bio = r10_bio->devs[i].bio;
3276 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3277 if (conf->mirrors[d].rdev == NULL ||
3278 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3280 sector = r10_bio->devs[i].addr;
3281 if (is_badblock(conf->mirrors[d].rdev,
3283 &first_bad, &bad_sectors)) {
3284 if (first_bad > sector)
3285 max_sync = first_bad - sector;
3287 bad_sectors -= (sector - first_bad);
3288 if (max_sync > bad_sectors)
3289 max_sync = bad_sectors;
3293 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3294 atomic_inc(&r10_bio->remaining);
3295 bio->bi_next = biolist;
3297 bio->bi_private = r10_bio;
3298 bio->bi_end_io = end_sync_read;
3300 bio->bi_iter.bi_sector = sector +
3301 conf->mirrors[d].rdev->data_offset;
3302 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3305 if (conf->mirrors[d].replacement == NULL ||
3307 &conf->mirrors[d].replacement->flags))
3310 /* Need to set up for writing to the replacement */
3311 bio = r10_bio->devs[i].repl_bio;
3313 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3315 sector = r10_bio->devs[i].addr;
3316 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3317 bio->bi_next = biolist;
3319 bio->bi_private = r10_bio;
3320 bio->bi_end_io = end_sync_write;
3322 bio->bi_iter.bi_sector = sector +
3323 conf->mirrors[d].replacement->data_offset;
3324 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3329 for (i=0; i<conf->copies; i++) {
3330 int d = r10_bio->devs[i].devnum;
3331 if (r10_bio->devs[i].bio->bi_end_io)
3332 rdev_dec_pending(conf->mirrors[d].rdev,
3334 if (r10_bio->devs[i].repl_bio &&
3335 r10_bio->devs[i].repl_bio->bi_end_io)
3337 conf->mirrors[d].replacement,
3347 if (sector_nr + max_sync < max_sector)
3348 max_sector = sector_nr + max_sync;
3351 int len = PAGE_SIZE;
3352 if (sector_nr + (len>>9) > max_sector)
3353 len = (max_sector - sector_nr) << 9;
3356 for (bio= biolist ; bio ; bio=bio->bi_next) {
3358 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3359 if (bio_add_page(bio, page, len, 0))
3363 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3364 for (bio2 = biolist;
3365 bio2 && bio2 != bio;
3366 bio2 = bio2->bi_next) {
3367 /* remove last page from this bio */
3369 bio2->bi_iter.bi_size -= len;
3370 __clear_bit(BIO_SEG_VALID, &bio2->bi_flags);
3374 nr_sectors += len>>9;
3375 sector_nr += len>>9;
3376 } while (biolist->bi_vcnt < RESYNC_PAGES);
3378 r10_bio->sectors = nr_sectors;
3382 biolist = biolist->bi_next;
3384 bio->bi_next = NULL;
3385 r10_bio = bio->bi_private;
3386 r10_bio->sectors = nr_sectors;
3388 if (bio->bi_end_io == end_sync_read) {
3389 md_sync_acct(bio->bi_bdev, nr_sectors);
3390 set_bit(BIO_UPTODATE, &bio->bi_flags);
3391 generic_make_request(bio);
3395 if (sectors_skipped)
3396 /* pretend they weren't skipped, it makes
3397 * no important difference in this case
3399 md_done_sync(mddev, sectors_skipped, 1);
3401 return sectors_skipped + nr_sectors;
3403 /* There is nowhere to write, so all non-sync
3404 * drives must be failed or in resync, all drives
3405 * have a bad block, so try the next chunk...
3407 if (sector_nr + max_sync < max_sector)
3408 max_sector = sector_nr + max_sync;
3410 sectors_skipped += (max_sector - sector_nr);
3412 sector_nr = max_sector;
3417 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3420 struct r10conf *conf = mddev->private;
3423 raid_disks = min(conf->geo.raid_disks,
3424 conf->prev.raid_disks);
3426 sectors = conf->dev_sectors;
3428 size = sectors >> conf->geo.chunk_shift;
3429 sector_div(size, conf->geo.far_copies);
3430 size = size * raid_disks;
3431 sector_div(size, conf->geo.near_copies);
3433 return size << conf->geo.chunk_shift;
3436 static void calc_sectors(struct r10conf *conf, sector_t size)
3438 /* Calculate the number of sectors-per-device that will
3439 * actually be used, and set conf->dev_sectors and
3443 size = size >> conf->geo.chunk_shift;
3444 sector_div(size, conf->geo.far_copies);
3445 size = size * conf->geo.raid_disks;
3446 sector_div(size, conf->geo.near_copies);
3447 /* 'size' is now the number of chunks in the array */
3448 /* calculate "used chunks per device" */
3449 size = size * conf->copies;
3451 /* We need to round up when dividing by raid_disks to
3452 * get the stride size.
3454 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3456 conf->dev_sectors = size << conf->geo.chunk_shift;
3458 if (conf->geo.far_offset)
3459 conf->geo.stride = 1 << conf->geo.chunk_shift;
3461 sector_div(size, conf->geo.far_copies);
3462 conf->geo.stride = size << conf->geo.chunk_shift;
3466 enum geo_type {geo_new, geo_old, geo_start};
3467 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3470 int layout, chunk, disks;
3473 layout = mddev->layout;
3474 chunk = mddev->chunk_sectors;
3475 disks = mddev->raid_disks - mddev->delta_disks;
3478 layout = mddev->new_layout;
3479 chunk = mddev->new_chunk_sectors;
3480 disks = mddev->raid_disks;
3482 default: /* avoid 'may be unused' warnings */
3483 case geo_start: /* new when starting reshape - raid_disks not
3485 layout = mddev->new_layout;
3486 chunk = mddev->new_chunk_sectors;
3487 disks = mddev->raid_disks + mddev->delta_disks;
3492 if (chunk < (PAGE_SIZE >> 9) ||
3493 !is_power_of_2(chunk))
3496 fc = (layout >> 8) & 255;
3497 fo = layout & (1<<16);
3498 geo->raid_disks = disks;
3499 geo->near_copies = nc;
3500 geo->far_copies = fc;
3501 geo->far_offset = fo;
3502 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3503 geo->chunk_mask = chunk - 1;
3504 geo->chunk_shift = ffz(~chunk);
3508 static struct r10conf *setup_conf(struct mddev *mddev)
3510 struct r10conf *conf = NULL;
3515 copies = setup_geo(&geo, mddev, geo_new);
3518 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3519 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3520 mdname(mddev), PAGE_SIZE);
3524 if (copies < 2 || copies > mddev->raid_disks) {
3525 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3526 mdname(mddev), mddev->new_layout);
3531 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3535 /* FIXME calc properly */
3536 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3537 max(0,-mddev->delta_disks)),
3542 conf->tmppage = alloc_page(GFP_KERNEL);
3547 conf->copies = copies;
3548 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3549 r10bio_pool_free, conf);
3550 if (!conf->r10bio_pool)
3553 calc_sectors(conf, mddev->dev_sectors);
3554 if (mddev->reshape_position == MaxSector) {
3555 conf->prev = conf->geo;
3556 conf->reshape_progress = MaxSector;
3558 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3562 conf->reshape_progress = mddev->reshape_position;
3563 if (conf->prev.far_offset)
3564 conf->prev.stride = 1 << conf->prev.chunk_shift;
3566 /* far_copies must be 1 */
3567 conf->prev.stride = conf->dev_sectors;
3569 spin_lock_init(&conf->device_lock);
3570 INIT_LIST_HEAD(&conf->retry_list);
3572 spin_lock_init(&conf->resync_lock);
3573 init_waitqueue_head(&conf->wait_barrier);
3575 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3579 conf->mddev = mddev;
3584 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3587 if (conf->r10bio_pool)
3588 mempool_destroy(conf->r10bio_pool);
3589 kfree(conf->mirrors);
3590 safe_put_page(conf->tmppage);
3593 return ERR_PTR(err);
3596 static int run(struct mddev *mddev)
3598 struct r10conf *conf;
3599 int i, disk_idx, chunk_size;
3600 struct raid10_info *disk;
3601 struct md_rdev *rdev;
3603 sector_t min_offset_diff = 0;
3605 bool discard_supported = false;
3607 if (mddev->private == NULL) {
3608 conf = setup_conf(mddev);
3610 return PTR_ERR(conf);
3611 mddev->private = conf;
3613 conf = mddev->private;
3617 mddev->thread = conf->thread;
3618 conf->thread = NULL;
3620 chunk_size = mddev->chunk_sectors << 9;
3622 blk_queue_max_discard_sectors(mddev->queue,
3623 mddev->chunk_sectors);
3624 blk_queue_max_write_same_sectors(mddev->queue, 0);
3625 blk_queue_io_min(mddev->queue, chunk_size);
3626 if (conf->geo.raid_disks % conf->geo.near_copies)
3627 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3629 blk_queue_io_opt(mddev->queue, chunk_size *
3630 (conf->geo.raid_disks / conf->geo.near_copies));
3633 rdev_for_each(rdev, mddev) {
3635 struct request_queue *q;
3637 disk_idx = rdev->raid_disk;
3640 if (disk_idx >= conf->geo.raid_disks &&
3641 disk_idx >= conf->prev.raid_disks)
3643 disk = conf->mirrors + disk_idx;
3645 if (test_bit(Replacement, &rdev->flags)) {
3646 if (disk->replacement)
3648 disk->replacement = rdev;
3654 q = bdev_get_queue(rdev->bdev);
3655 if (q->merge_bvec_fn)
3656 mddev->merge_check_needed = 1;
3657 diff = (rdev->new_data_offset - rdev->data_offset);
3658 if (!mddev->reshape_backwards)
3662 if (first || diff < min_offset_diff)
3663 min_offset_diff = diff;
3666 disk_stack_limits(mddev->gendisk, rdev->bdev,
3667 rdev->data_offset << 9);
3669 disk->head_position = 0;
3671 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3672 discard_supported = true;
3676 if (discard_supported)
3677 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3680 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3683 /* need to check that every block has at least one working mirror */
3684 if (!enough(conf, -1)) {
3685 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3690 if (conf->reshape_progress != MaxSector) {
3691 /* must ensure that shape change is supported */
3692 if (conf->geo.far_copies != 1 &&
3693 conf->geo.far_offset == 0)
3695 if (conf->prev.far_copies != 1 &&
3696 conf->prev.far_offset == 0)
3700 mddev->degraded = 0;
3702 i < conf->geo.raid_disks
3703 || i < conf->prev.raid_disks;
3706 disk = conf->mirrors + i;
3708 if (!disk->rdev && disk->replacement) {
3709 /* The replacement is all we have - use it */
3710 disk->rdev = disk->replacement;
3711 disk->replacement = NULL;
3712 clear_bit(Replacement, &disk->rdev->flags);
3716 !test_bit(In_sync, &disk->rdev->flags)) {
3717 disk->head_position = 0;
3720 disk->rdev->saved_raid_disk < 0)
3723 disk->recovery_disabled = mddev->recovery_disabled - 1;
3726 if (mddev->recovery_cp != MaxSector)
3727 printk(KERN_NOTICE "md/raid10:%s: not clean"
3728 " -- starting background reconstruction\n",
3731 "md/raid10:%s: active with %d out of %d devices\n",
3732 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3733 conf->geo.raid_disks);
3735 * Ok, everything is just fine now
3737 mddev->dev_sectors = conf->dev_sectors;
3738 size = raid10_size(mddev, 0, 0);
3739 md_set_array_sectors(mddev, size);
3740 mddev->resync_max_sectors = size;
3743 int stripe = conf->geo.raid_disks *
3744 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3746 /* Calculate max read-ahead size.
3747 * We need to readahead at least twice a whole stripe....
3750 stripe /= conf->geo.near_copies;
3751 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3752 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3755 if (md_integrity_register(mddev))
3758 if (conf->reshape_progress != MaxSector) {
3759 unsigned long before_length, after_length;
3761 before_length = ((1 << conf->prev.chunk_shift) *
3762 conf->prev.far_copies);
3763 after_length = ((1 << conf->geo.chunk_shift) *
3764 conf->geo.far_copies);
3766 if (max(before_length, after_length) > min_offset_diff) {
3767 /* This cannot work */
3768 printk("md/raid10: offset difference not enough to continue reshape\n");
3771 conf->offset_diff = min_offset_diff;
3773 conf->reshape_safe = conf->reshape_progress;
3774 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3775 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3776 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3777 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3778 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3785 md_unregister_thread(&mddev->thread);
3786 if (conf->r10bio_pool)
3787 mempool_destroy(conf->r10bio_pool);
3788 safe_put_page(conf->tmppage);
3789 kfree(conf->mirrors);
3791 mddev->private = NULL;
3796 static void raid10_free(struct mddev *mddev, void *priv)
3798 struct r10conf *conf = priv;
3800 if (conf->r10bio_pool)
3801 mempool_destroy(conf->r10bio_pool);
3802 safe_put_page(conf->tmppage);
3803 kfree(conf->mirrors);
3804 kfree(conf->mirrors_old);
3805 kfree(conf->mirrors_new);
3809 static void raid10_quiesce(struct mddev *mddev, int state)
3811 struct r10conf *conf = mddev->private;
3815 raise_barrier(conf, 0);
3818 lower_barrier(conf);
3823 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3825 /* Resize of 'far' arrays is not supported.
3826 * For 'near' and 'offset' arrays we can set the
3827 * number of sectors used to be an appropriate multiple
3828 * of the chunk size.
3829 * For 'offset', this is far_copies*chunksize.
3830 * For 'near' the multiplier is the LCM of
3831 * near_copies and raid_disks.
3832 * So if far_copies > 1 && !far_offset, fail.
3833 * Else find LCM(raid_disks, near_copy)*far_copies and
3834 * multiply by chunk_size. Then round to this number.
3835 * This is mostly done by raid10_size()
3837 struct r10conf *conf = mddev->private;
3838 sector_t oldsize, size;
3840 if (mddev->reshape_position != MaxSector)
3843 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3846 oldsize = raid10_size(mddev, 0, 0);
3847 size = raid10_size(mddev, sectors, 0);
3848 if (mddev->external_size &&
3849 mddev->array_sectors > size)
3851 if (mddev->bitmap) {
3852 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3856 md_set_array_sectors(mddev, size);
3857 set_capacity(mddev->gendisk, mddev->array_sectors);
3858 revalidate_disk(mddev->gendisk);
3859 if (sectors > mddev->dev_sectors &&
3860 mddev->recovery_cp > oldsize) {
3861 mddev->recovery_cp = oldsize;
3862 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3864 calc_sectors(conf, sectors);
3865 mddev->dev_sectors = conf->dev_sectors;
3866 mddev->resync_max_sectors = size;
3870 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3872 struct md_rdev *rdev;
3873 struct r10conf *conf;
3875 if (mddev->degraded > 0) {
3876 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3878 return ERR_PTR(-EINVAL);
3880 sector_div(size, devs);
3882 /* Set new parameters */
3883 mddev->new_level = 10;
3884 /* new layout: far_copies = 1, near_copies = 2 */
3885 mddev->new_layout = (1<<8) + 2;
3886 mddev->new_chunk_sectors = mddev->chunk_sectors;
3887 mddev->delta_disks = mddev->raid_disks;
3888 mddev->raid_disks *= 2;
3889 /* make sure it will be not marked as dirty */
3890 mddev->recovery_cp = MaxSector;
3891 mddev->dev_sectors = size;
3893 conf = setup_conf(mddev);
3894 if (!IS_ERR(conf)) {
3895 rdev_for_each(rdev, mddev)
3896 if (rdev->raid_disk >= 0) {
3897 rdev->new_raid_disk = rdev->raid_disk * 2;
3898 rdev->sectors = size;
3906 static void *raid10_takeover(struct mddev *mddev)
3908 struct r0conf *raid0_conf;
3910 /* raid10 can take over:
3911 * raid0 - providing it has only two drives
3913 if (mddev->level == 0) {
3914 /* for raid0 takeover only one zone is supported */
3915 raid0_conf = mddev->private;
3916 if (raid0_conf->nr_strip_zones > 1) {
3917 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3918 " with more than one zone.\n",
3920 return ERR_PTR(-EINVAL);
3922 return raid10_takeover_raid0(mddev,
3923 raid0_conf->strip_zone->zone_end,
3924 raid0_conf->strip_zone->nb_dev);
3926 return ERR_PTR(-EINVAL);
3929 static int raid10_check_reshape(struct mddev *mddev)
3931 /* Called when there is a request to change
3932 * - layout (to ->new_layout)
3933 * - chunk size (to ->new_chunk_sectors)
3934 * - raid_disks (by delta_disks)
3935 * or when trying to restart a reshape that was ongoing.
3937 * We need to validate the request and possibly allocate
3938 * space if that might be an issue later.
3940 * Currently we reject any reshape of a 'far' mode array,
3941 * allow chunk size to change if new is generally acceptable,
3942 * allow raid_disks to increase, and allow
3943 * a switch between 'near' mode and 'offset' mode.
3945 struct r10conf *conf = mddev->private;
3948 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3951 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3952 /* mustn't change number of copies */
3954 if (geo.far_copies > 1 && !geo.far_offset)
3955 /* Cannot switch to 'far' mode */
3958 if (mddev->array_sectors & geo.chunk_mask)
3959 /* not factor of array size */
3962 if (!enough(conf, -1))
3965 kfree(conf->mirrors_new);
3966 conf->mirrors_new = NULL;
3967 if (mddev->delta_disks > 0) {
3968 /* allocate new 'mirrors' list */
3969 conf->mirrors_new = kzalloc(
3970 sizeof(struct raid10_info)
3971 *(mddev->raid_disks +
3972 mddev->delta_disks),
3974 if (!conf->mirrors_new)
3981 * Need to check if array has failed when deciding whether to:
3983 * - remove non-faulty devices
3986 * This determination is simple when no reshape is happening.
3987 * However if there is a reshape, we need to carefully check
3988 * both the before and after sections.
3989 * This is because some failed devices may only affect one
3990 * of the two sections, and some non-in_sync devices may
3991 * be insync in the section most affected by failed devices.
3993 static int calc_degraded(struct r10conf *conf)
3995 int degraded, degraded2;
4000 /* 'prev' section first */
4001 for (i = 0; i < conf->prev.raid_disks; i++) {
4002 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4003 if (!rdev || test_bit(Faulty, &rdev->flags))
4005 else if (!test_bit(In_sync, &rdev->flags))
4006 /* When we can reduce the number of devices in
4007 * an array, this might not contribute to
4008 * 'degraded'. It does now.
4013 if (conf->geo.raid_disks == conf->prev.raid_disks)
4017 for (i = 0; i < conf->geo.raid_disks; i++) {
4018 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4019 if (!rdev || test_bit(Faulty, &rdev->flags))
4021 else if (!test_bit(In_sync, &rdev->flags)) {
4022 /* If reshape is increasing the number of devices,
4023 * this section has already been recovered, so
4024 * it doesn't contribute to degraded.
4027 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4032 if (degraded2 > degraded)
4037 static int raid10_start_reshape(struct mddev *mddev)
4039 /* A 'reshape' has been requested. This commits
4040 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4041 * This also checks if there are enough spares and adds them
4043 * We currently require enough spares to make the final
4044 * array non-degraded. We also require that the difference
4045 * between old and new data_offset - on each device - is
4046 * enough that we never risk over-writing.
4049 unsigned long before_length, after_length;
4050 sector_t min_offset_diff = 0;
4053 struct r10conf *conf = mddev->private;
4054 struct md_rdev *rdev;
4058 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4061 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4064 before_length = ((1 << conf->prev.chunk_shift) *
4065 conf->prev.far_copies);
4066 after_length = ((1 << conf->geo.chunk_shift) *
4067 conf->geo.far_copies);
4069 rdev_for_each(rdev, mddev) {
4070 if (!test_bit(In_sync, &rdev->flags)
4071 && !test_bit(Faulty, &rdev->flags))
4073 if (rdev->raid_disk >= 0) {
4074 long long diff = (rdev->new_data_offset
4075 - rdev->data_offset);
4076 if (!mddev->reshape_backwards)
4080 if (first || diff < min_offset_diff)
4081 min_offset_diff = diff;
4085 if (max(before_length, after_length) > min_offset_diff)
4088 if (spares < mddev->delta_disks)
4091 conf->offset_diff = min_offset_diff;
4092 spin_lock_irq(&conf->device_lock);
4093 if (conf->mirrors_new) {
4094 memcpy(conf->mirrors_new, conf->mirrors,
4095 sizeof(struct raid10_info)*conf->prev.raid_disks);
4097 kfree(conf->mirrors_old);
4098 conf->mirrors_old = conf->mirrors;
4099 conf->mirrors = conf->mirrors_new;
4100 conf->mirrors_new = NULL;
4102 setup_geo(&conf->geo, mddev, geo_start);
4104 if (mddev->reshape_backwards) {
4105 sector_t size = raid10_size(mddev, 0, 0);
4106 if (size < mddev->array_sectors) {
4107 spin_unlock_irq(&conf->device_lock);
4108 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4112 mddev->resync_max_sectors = size;
4113 conf->reshape_progress = size;
4115 conf->reshape_progress = 0;
4116 spin_unlock_irq(&conf->device_lock);
4118 if (mddev->delta_disks && mddev->bitmap) {
4119 ret = bitmap_resize(mddev->bitmap,
4120 raid10_size(mddev, 0,
4121 conf->geo.raid_disks),
4126 if (mddev->delta_disks > 0) {
4127 rdev_for_each(rdev, mddev)
4128 if (rdev->raid_disk < 0 &&
4129 !test_bit(Faulty, &rdev->flags)) {
4130 if (raid10_add_disk(mddev, rdev) == 0) {
4131 if (rdev->raid_disk >=
4132 conf->prev.raid_disks)
4133 set_bit(In_sync, &rdev->flags);
4135 rdev->recovery_offset = 0;
4137 if (sysfs_link_rdev(mddev, rdev))
4138 /* Failure here is OK */;
4140 } else if (rdev->raid_disk >= conf->prev.raid_disks
4141 && !test_bit(Faulty, &rdev->flags)) {
4142 /* This is a spare that was manually added */
4143 set_bit(In_sync, &rdev->flags);
4146 /* When a reshape changes the number of devices,
4147 * ->degraded is measured against the larger of the
4148 * pre and post numbers.
4150 spin_lock_irq(&conf->device_lock);
4151 mddev->degraded = calc_degraded(conf);
4152 spin_unlock_irq(&conf->device_lock);
4153 mddev->raid_disks = conf->geo.raid_disks;
4154 mddev->reshape_position = conf->reshape_progress;
4155 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4157 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4158 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4159 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4160 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4161 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4163 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4165 if (!mddev->sync_thread) {
4169 conf->reshape_checkpoint = jiffies;
4170 md_wakeup_thread(mddev->sync_thread);
4171 md_new_event(mddev);
4175 mddev->recovery = 0;
4176 spin_lock_irq(&conf->device_lock);
4177 conf->geo = conf->prev;
4178 mddev->raid_disks = conf->geo.raid_disks;
4179 rdev_for_each(rdev, mddev)
4180 rdev->new_data_offset = rdev->data_offset;
4182 conf->reshape_progress = MaxSector;
4183 mddev->reshape_position = MaxSector;
4184 spin_unlock_irq(&conf->device_lock);
4188 /* Calculate the last device-address that could contain
4189 * any block from the chunk that includes the array-address 's'
4190 * and report the next address.
4191 * i.e. the address returned will be chunk-aligned and after
4192 * any data that is in the chunk containing 's'.
4194 static sector_t last_dev_address(sector_t s, struct geom *geo)
4196 s = (s | geo->chunk_mask) + 1;
4197 s >>= geo->chunk_shift;
4198 s *= geo->near_copies;
4199 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4200 s *= geo->far_copies;
4201 s <<= geo->chunk_shift;
4205 /* Calculate the first device-address that could contain
4206 * any block from the chunk that includes the array-address 's'.
4207 * This too will be the start of a chunk
4209 static sector_t first_dev_address(sector_t s, struct geom *geo)
4211 s >>= geo->chunk_shift;
4212 s *= geo->near_copies;
4213 sector_div(s, geo->raid_disks);
4214 s *= geo->far_copies;
4215 s <<= geo->chunk_shift;
4219 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4222 /* We simply copy at most one chunk (smallest of old and new)
4223 * at a time, possibly less if that exceeds RESYNC_PAGES,
4224 * or we hit a bad block or something.
4225 * This might mean we pause for normal IO in the middle of
4226 * a chunk, but that is not a problem was mddev->reshape_position
4227 * can record any location.
4229 * If we will want to write to a location that isn't
4230 * yet recorded as 'safe' (i.e. in metadata on disk) then
4231 * we need to flush all reshape requests and update the metadata.
4233 * When reshaping forwards (e.g. to more devices), we interpret
4234 * 'safe' as the earliest block which might not have been copied
4235 * down yet. We divide this by previous stripe size and multiply
4236 * by previous stripe length to get lowest device offset that we
4237 * cannot write to yet.
4238 * We interpret 'sector_nr' as an address that we want to write to.
4239 * From this we use last_device_address() to find where we might
4240 * write to, and first_device_address on the 'safe' position.
4241 * If this 'next' write position is after the 'safe' position,
4242 * we must update the metadata to increase the 'safe' position.
4244 * When reshaping backwards, we round in the opposite direction
4245 * and perform the reverse test: next write position must not be
4246 * less than current safe position.
4248 * In all this the minimum difference in data offsets
4249 * (conf->offset_diff - always positive) allows a bit of slack,
4250 * so next can be after 'safe', but not by more than offset_disk
4252 * We need to prepare all the bios here before we start any IO
4253 * to ensure the size we choose is acceptable to all devices.
4254 * The means one for each copy for write-out and an extra one for
4256 * We store the read-in bio in ->master_bio and the others in
4257 * ->devs[x].bio and ->devs[x].repl_bio.
4259 struct r10conf *conf = mddev->private;
4260 struct r10bio *r10_bio;
4261 sector_t next, safe, last;
4265 struct md_rdev *rdev;
4268 struct bio *bio, *read_bio;
4269 int sectors_done = 0;
4271 if (sector_nr == 0) {
4272 /* If restarting in the middle, skip the initial sectors */
4273 if (mddev->reshape_backwards &&
4274 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4275 sector_nr = (raid10_size(mddev, 0, 0)
4276 - conf->reshape_progress);
4277 } else if (!mddev->reshape_backwards &&
4278 conf->reshape_progress > 0)
4279 sector_nr = conf->reshape_progress;
4281 mddev->curr_resync_completed = sector_nr;
4282 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4288 /* We don't use sector_nr to track where we are up to
4289 * as that doesn't work well for ->reshape_backwards.
4290 * So just use ->reshape_progress.
4292 if (mddev->reshape_backwards) {
4293 /* 'next' is the earliest device address that we might
4294 * write to for this chunk in the new layout
4296 next = first_dev_address(conf->reshape_progress - 1,
4299 /* 'safe' is the last device address that we might read from
4300 * in the old layout after a restart
4302 safe = last_dev_address(conf->reshape_safe - 1,
4305 if (next + conf->offset_diff < safe)
4308 last = conf->reshape_progress - 1;
4309 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4310 & conf->prev.chunk_mask);
4311 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4312 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4314 /* 'next' is after the last device address that we
4315 * might write to for this chunk in the new layout
4317 next = last_dev_address(conf->reshape_progress, &conf->geo);
4319 /* 'safe' is the earliest device address that we might
4320 * read from in the old layout after a restart
4322 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4324 /* Need to update metadata if 'next' might be beyond 'safe'
4325 * as that would possibly corrupt data
4327 if (next > safe + conf->offset_diff)
4330 sector_nr = conf->reshape_progress;
4331 last = sector_nr | (conf->geo.chunk_mask
4332 & conf->prev.chunk_mask);
4334 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4335 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4339 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4340 /* Need to update reshape_position in metadata */
4342 mddev->reshape_position = conf->reshape_progress;
4343 if (mddev->reshape_backwards)
4344 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4345 - conf->reshape_progress;
4347 mddev->curr_resync_completed = conf->reshape_progress;
4348 conf->reshape_checkpoint = jiffies;
4349 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4350 md_wakeup_thread(mddev->thread);
4351 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4352 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4353 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4354 allow_barrier(conf);
4355 return sectors_done;
4357 conf->reshape_safe = mddev->reshape_position;
4358 allow_barrier(conf);
4362 /* Now schedule reads for blocks from sector_nr to last */
4363 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4365 raise_barrier(conf, sectors_done != 0);
4366 atomic_set(&r10_bio->remaining, 0);
4367 r10_bio->mddev = mddev;
4368 r10_bio->sector = sector_nr;
4369 set_bit(R10BIO_IsReshape, &r10_bio->state);
4370 r10_bio->sectors = last - sector_nr + 1;
4371 rdev = read_balance(conf, r10_bio, &max_sectors);
4372 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4375 /* Cannot read from here, so need to record bad blocks
4376 * on all the target devices.
4379 mempool_free(r10_bio, conf->r10buf_pool);
4380 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4381 return sectors_done;
4384 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4386 read_bio->bi_bdev = rdev->bdev;
4387 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4388 + rdev->data_offset);
4389 read_bio->bi_private = r10_bio;
4390 read_bio->bi_end_io = end_sync_read;
4391 read_bio->bi_rw = READ;
4392 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4393 __set_bit(BIO_UPTODATE, &read_bio->bi_flags);
4394 read_bio->bi_vcnt = 0;
4395 read_bio->bi_iter.bi_size = 0;
4396 r10_bio->master_bio = read_bio;
4397 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4399 /* Now find the locations in the new layout */
4400 __raid10_find_phys(&conf->geo, r10_bio);
4403 read_bio->bi_next = NULL;
4405 for (s = 0; s < conf->copies*2; s++) {
4407 int d = r10_bio->devs[s/2].devnum;
4408 struct md_rdev *rdev2;
4410 rdev2 = conf->mirrors[d].replacement;
4411 b = r10_bio->devs[s/2].repl_bio;
4413 rdev2 = conf->mirrors[d].rdev;
4414 b = r10_bio->devs[s/2].bio;
4416 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4420 b->bi_bdev = rdev2->bdev;
4421 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4422 rdev2->new_data_offset;
4423 b->bi_private = r10_bio;
4424 b->bi_end_io = end_reshape_write;
4430 /* Now add as many pages as possible to all of these bios. */
4433 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4434 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4435 int len = (max_sectors - s) << 9;
4436 if (len > PAGE_SIZE)
4438 for (bio = blist; bio ; bio = bio->bi_next) {
4440 if (bio_add_page(bio, page, len, 0))
4443 /* Didn't fit, must stop */
4445 bio2 && bio2 != bio;
4446 bio2 = bio2->bi_next) {
4447 /* Remove last page from this bio */
4449 bio2->bi_iter.bi_size -= len;
4450 __clear_bit(BIO_SEG_VALID, &bio2->bi_flags);
4454 sector_nr += len >> 9;
4455 nr_sectors += len >> 9;
4458 r10_bio->sectors = nr_sectors;
4460 /* Now submit the read */
4461 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4462 atomic_inc(&r10_bio->remaining);
4463 read_bio->bi_next = NULL;
4464 generic_make_request(read_bio);
4465 sector_nr += nr_sectors;
4466 sectors_done += nr_sectors;
4467 if (sector_nr <= last)
4470 /* Now that we have done the whole section we can
4471 * update reshape_progress
4473 if (mddev->reshape_backwards)
4474 conf->reshape_progress -= sectors_done;
4476 conf->reshape_progress += sectors_done;
4478 return sectors_done;
4481 static void end_reshape_request(struct r10bio *r10_bio);
4482 static int handle_reshape_read_error(struct mddev *mddev,
4483 struct r10bio *r10_bio);
4484 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4486 /* Reshape read completed. Hopefully we have a block
4488 * If we got a read error then we do sync 1-page reads from
4489 * elsewhere until we find the data - or give up.
4491 struct r10conf *conf = mddev->private;
4494 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4495 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4496 /* Reshape has been aborted */
4497 md_done_sync(mddev, r10_bio->sectors, 0);
4501 /* We definitely have the data in the pages, schedule the
4504 atomic_set(&r10_bio->remaining, 1);
4505 for (s = 0; s < conf->copies*2; s++) {
4507 int d = r10_bio->devs[s/2].devnum;
4508 struct md_rdev *rdev;
4510 rdev = conf->mirrors[d].replacement;
4511 b = r10_bio->devs[s/2].repl_bio;
4513 rdev = conf->mirrors[d].rdev;
4514 b = r10_bio->devs[s/2].bio;
4516 if (!rdev || test_bit(Faulty, &rdev->flags))
4518 atomic_inc(&rdev->nr_pending);
4519 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4520 atomic_inc(&r10_bio->remaining);
4522 generic_make_request(b);
4524 end_reshape_request(r10_bio);
4527 static void end_reshape(struct r10conf *conf)
4529 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4532 spin_lock_irq(&conf->device_lock);
4533 conf->prev = conf->geo;
4534 md_finish_reshape(conf->mddev);
4536 conf->reshape_progress = MaxSector;
4537 spin_unlock_irq(&conf->device_lock);
4539 /* read-ahead size must cover two whole stripes, which is
4540 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4542 if (conf->mddev->queue) {
4543 int stripe = conf->geo.raid_disks *
4544 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4545 stripe /= conf->geo.near_copies;
4546 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4547 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4552 static int handle_reshape_read_error(struct mddev *mddev,
4553 struct r10bio *r10_bio)
4555 /* Use sync reads to get the blocks from somewhere else */
4556 int sectors = r10_bio->sectors;
4557 struct r10conf *conf = mddev->private;
4559 struct r10bio r10_bio;
4560 struct r10dev devs[conf->copies];
4562 struct r10bio *r10b = &on_stack.r10_bio;
4565 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4567 r10b->sector = r10_bio->sector;
4568 __raid10_find_phys(&conf->prev, r10b);
4573 int first_slot = slot;
4575 if (s > (PAGE_SIZE >> 9))
4579 int d = r10b->devs[slot].devnum;
4580 struct md_rdev *rdev = conf->mirrors[d].rdev;
4583 test_bit(Faulty, &rdev->flags) ||
4584 !test_bit(In_sync, &rdev->flags))
4587 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4588 success = sync_page_io(rdev,
4597 if (slot >= conf->copies)
4599 if (slot == first_slot)
4603 /* couldn't read this block, must give up */
4604 set_bit(MD_RECOVERY_INTR,
4614 static void end_reshape_write(struct bio *bio, int error)
4616 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4617 struct r10bio *r10_bio = bio->bi_private;
4618 struct mddev *mddev = r10_bio->mddev;
4619 struct r10conf *conf = mddev->private;
4623 struct md_rdev *rdev = NULL;
4625 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4627 rdev = conf->mirrors[d].replacement;
4630 rdev = conf->mirrors[d].rdev;
4634 /* FIXME should record badblock */
4635 md_error(mddev, rdev);
4638 rdev_dec_pending(rdev, mddev);
4639 end_reshape_request(r10_bio);
4642 static void end_reshape_request(struct r10bio *r10_bio)
4644 if (!atomic_dec_and_test(&r10_bio->remaining))
4646 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4647 bio_put(r10_bio->master_bio);
4651 static void raid10_finish_reshape(struct mddev *mddev)
4653 struct r10conf *conf = mddev->private;
4655 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4658 if (mddev->delta_disks > 0) {
4659 sector_t size = raid10_size(mddev, 0, 0);
4660 md_set_array_sectors(mddev, size);
4661 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4662 mddev->recovery_cp = mddev->resync_max_sectors;
4663 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4665 mddev->resync_max_sectors = size;
4666 set_capacity(mddev->gendisk, mddev->array_sectors);
4667 revalidate_disk(mddev->gendisk);
4670 for (d = conf->geo.raid_disks ;
4671 d < conf->geo.raid_disks - mddev->delta_disks;
4673 struct md_rdev *rdev = conf->mirrors[d].rdev;
4675 clear_bit(In_sync, &rdev->flags);
4676 rdev = conf->mirrors[d].replacement;
4678 clear_bit(In_sync, &rdev->flags);
4681 mddev->layout = mddev->new_layout;
4682 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4683 mddev->reshape_position = MaxSector;
4684 mddev->delta_disks = 0;
4685 mddev->reshape_backwards = 0;
4688 static struct md_personality raid10_personality =
4692 .owner = THIS_MODULE,
4693 .make_request = make_request,
4695 .free = raid10_free,
4697 .error_handler = error,
4698 .hot_add_disk = raid10_add_disk,
4699 .hot_remove_disk= raid10_remove_disk,
4700 .spare_active = raid10_spare_active,
4701 .sync_request = sync_request,
4702 .quiesce = raid10_quiesce,
4703 .size = raid10_size,
4704 .resize = raid10_resize,
4705 .takeover = raid10_takeover,
4706 .check_reshape = raid10_check_reshape,
4707 .start_reshape = raid10_start_reshape,
4708 .finish_reshape = raid10_finish_reshape,
4709 .congested = raid10_congested,
4710 .mergeable_bvec = raid10_mergeable_bvec,
4713 static int __init raid_init(void)
4715 return register_md_personality(&raid10_personality);
4718 static void raid_exit(void)
4720 unregister_md_personality(&raid10_personality);
4723 module_init(raid_init);
4724 module_exit(raid_exit);
4725 MODULE_LICENSE("GPL");
4726 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4727 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4728 MODULE_ALIAS("md-raid10");
4729 MODULE_ALIAS("md-level-10");
4731 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);